Compare commits

...
529 Commits
Author SHA1 Message Date
zicai.pu ee9170e7bb fix(mooncake): bind IntraNode batches to the bootstrap CUDA device
Sender and executor threads can enter synchronous IntraNode transfers with
another current CUDA device. Keep the explicit bootstrap ordinal, check it
against the published model device, and scope the native batch call to that
device while restoring the calling thread on success or failure.

Ordinary transports, force-TCP and non-CUDA devices keep their existing
path. Missing or inconsistent CUDA identity is rejected before native IO.

Add 13 CPU regressions importing the production wrapper, including real
constructor calls with a mocked SDK, concurrent thread-local device states,
error restoration and transport/device compatibility. All 13 pass; two
selected tests fail on the unchanged wrapper at behavior assertions. Native
and CUDA boundaries are mocked, so this does not certify physical CUDA IPC
or attribute the separate container-PID collision to this change.
2026-10-30 11:42:19 +08:00
zicai.pu 17505e6cf6 hicache: align prefetch and host restore boundaries
Use the request radix-key cap for L3 prefetch so bounded replay cannot
split a queried trailing-state checkpoint at the later rematch. Reject
Host-only SWA-gap candidates before they mask a complete ancestor, while
preserving device matching and the final load-back completeness guard.

Add CPU regressions that import the real scheduler, cache and TreeCore
methods. All 13 test methods pass; the same tests against the unmodified
scheduler/SWA code fail six methods (ten assertions including subtests).
The checks use synthetic cache state and controller boundaries, not GPU
or live Store execution.
2026-10-30 11:42:19 +08:00
zicai.pu 81f7398855 fix(hicache): reject unproven trailing restore endpoints
Do not invent shorter checkpoints when a trailing pool has no restorable-prefix inventory. Reject partial allocation and vote no common endpoint while preserving the collective shape and plain-KV contiguous behavior.

Add negative controls for missing inventory and retain pure-KV coverage. Validation: 33 cache/PP CPU tests passed with an external Mac native-hash import adapter using the repository Python hash oracle; repository pre-commit checks passed.
2026-10-30 11:42:19 +08:00
zicai.pu 0a56a5f6e7 fix(hicache): preserve sparse restore checkpoints under host pressure
Choose a common validated trailing-state endpoint before partial host allocation, retain that endpoint set across query synchronization, and size auxiliary staging from the selected length. Assumed trailing hits prove only their original endpoint.

Add controller/cache CPU regressions and a PP checkpoint-vote control. Validation: 14 targeted CPU tests and repository pre-commit checks passed.
2026-10-30 11:35:29 +08:00
minke.yu 7f6fe22445 mem_cache: fail-soft on SWA load-back gaps instead of asserting
With long prompts and L3 under pressure, a matched prefix can contain a node
whose SWA component has neither a device nor a host value (host tombstone
evicted between match and load-back). build_hicache_transfers asserted on
this and killed the prefill scheduler.

Add SWAComponent.has_load_back_gap() and have build_load_back_spec return an
empty transfer spec when a gap is found, so the caller backs off and the
request is recomputed. Measured on peak15-v6 (1M, cap 64): 0 scheduler
exceptions, hit 59.4%->94.0%, TTFT p95 47.0s->3.4s with the SWA prefix-match
gate disabled.
2026-10-26 13:23:17 +00:00
minke.yu f6fc6ac3e7 mem_cache: allow disabling the DSv4.1 SWA prefix-match gate via env
The unified-cache SWA component gates prefix matching on a live SWA window,
so a long prefix only matches up to the last sliding window and the tail is
recomputed. Add SGLANG_DSV4_SWA_GATES_PREFIX_MATCH (default True) so the gate
can be turned off for measurement; the SWA window is then rebuilt/recovered
downstream as for the req-ring paths. Measured on 4xB300 DP4 at 160k prefix:
c32 hit tokens 40.6% -> ~100% (with HiCache L2/L3 enabled), needle test
(cold vs hit) identical.
2026-10-26 04:06:25 +00:00
minke.yu 0a3c370451 mooncake: probe pointers via cuda.bindings instead of ctypes libcudart
The ctypes-based probing loaded a second libcudart copy into the process,
which corrupted CUDA/runtime state and segfaulted the scheduler processes
shortly after registration (reproducible on prefill ranks, intermittent on
decode). cuda.bindings.runtime.cudaPointerGetAttributes is a properly
typed binding and boots cleanly on both roles.
2026-10-25 14:04:56 +00:00
minke.yu b1b303c9f2 mooncake: support host-memory payloads with intra-node NVLink transport
IntraNodeNvlinkTransport can only register/reach device memory. sglang's
register list includes host regions (aux buffers, some state components);
one host region failed the whole registerLocalMemoryBatch and the engine
rolled back every region, leaving the segment descriptor empty so all KV
transfers failed with 'Requested address ... not found'.

When the intra-node NVLink transport is active (MC_INTRANODE_NVLINK etc.):
- register only device-memory regions (probe via cudaPointerGetAttributes)
- route transfer blocks whose local source is host memory over the ordered
  per-endpoint zmq channel (new STATE_DATA message) instead of the engine;
  the receiver validates and writes the block into its buffer, relying on
  the same ordering guarantee as the aux TCP path
- force aux over the TCP socket path automatically
2026-10-25 14:04:52 +00:00
Liangsheng Yin 7d996300b9 [Perf] dsv41: Use the fused ratio-2 compressor in the prefill graph (#46230) 2026-10-23 21:33:42 -07:00
Liangsheng Yin 54df642fe6 [Perf] dsv41: Translate SWA locations once per prefill graph step (#46229) 2026-10-23 21:33:15 -07:00
Liangsheng Yin 564aec61c8 [Perf] dsv41: Feed wo_b the fused MXFP8 attention output in the prefill graph (#46231) 2026-10-23 21:31:33 -07:00
Yihao WangandClaude Opus 5.5 a7183a8246 [Feat] spec: Let out-of-tree platforms provide EAGLE draft graph runners (#45145)
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-23 21:28:56 -07:00
Yihao WangandClaude Opus 5.5 af3f9e4c8f [Feat] spec: Let out-of-tree platforms provide their own speculative worker (#45144)
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-23 21:24:58 -07:00
MickandMick Qian f09cec65e9 [Fix] diffusion/storage: Join uploads before request cleanup (#46068)
Co-authored-by: Mick Qian <mickqian@users.noreply.github.com>
2026-10-24 12:19:10 +08:00
MickandMick Qian f4ecaae888 [Fix] diffusion/mesh: Guard asynchronous job publication and defaults (#46081)
Co-authored-by: Mick Qian <mickqian@users.noreply.github.com>
2026-10-24 12:18:25 +08:00
MickandMick Qian cdab6c1bcc [Fix] diffusion/loader: Preserve checkpoints on read failures (#46095)
Co-authored-by: Mick Qian <mickqian@users.noreply.github.com>
2026-10-24 12:17:46 +08:00
f2c634961b [Perf] diffusion: Project LTX-2 prompt K/V once per denoising loop (#46105)
Co-authored-by: Mick Qian <mickqian@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-24 12:17:04 +08:00
Liangsheng Yin 3b95120f00 [Fix] dsv41: Revert MegaGate routing in prefill (#46242) 2026-10-23 21:15:38 -07:00
55f23a8b59 [Fix] [AMD] dsv4.1: Fix decode graph capture with quantized low-ratio activations (#46151)
Co-authored-by: Kevin Mi <kevin.mi@radixark.ai>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-23 20:55:13 -07:00
Liangsheng Yin 3c6a6c8068 [Perf] dsv41: Run MegaGate routing in prefill (#46232) 2026-10-23 20:31:50 -07:00
Liangsheng Yin 17d92ea671 [Feat] openai: Parse DeepSeek text for the native endpoints through deepseek-recipe (#46219) 2026-10-23 20:17:37 -07:00
Richard WangandMick Qian 7cefa60765 [Perf] diffusion: Wait on the socket when the scheduler is idle (#46007)
Co-authored-by: Mick Qian <mickqian@users.noreply.github.com>
2026-10-24 11:02:24 +08:00
25ba24e9a7 [Perf] diffusion: Build one LTX-2.3 AdaLN row per sample when every token shares the sigma (#46109)
Co-authored-by: Mick Qian <mickqian@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-24 11:00:14 +08:00
42ac81c655 [Perf] diffusion: Route dense attention with a sub-tile side to FA4 on Blackwell (#46107)
Co-authored-by: Mick Qian <mickqian@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-24 10:59:25 +08:00
jambow0320andShangming Cai de7055f596 [Refactor] pd: Consolidate sender and receiver implementations (#41140)
Co-authored-by: Shangming Cai <csmthu@gmail.com>
2026-10-24 10:50:10 +08:00
jacky.cheng 88ac7c02a6 [Fix] [AMD] diffusion: Fix the first-step denoise regression from compiling ENABLE_CK=0 BF16 attention (#45838) 2026-10-23 19:42:54 -07:00
Jackey Huaandhnyls2002 a77998e157 [Feat] openai: parse DeepSeek output through deepseek-recipe (#45907)
Co-authored-by: hnyls2002 <lsyincs@gmail.com>
2026-10-23 19:23:22 -07:00
Liangsheng Yinandkpham-sgl 79701a2549 [Perf] dsv41: Read Engram rows from peer device shards (#46180)
Co-authored-by: kpham-sgl <khoa.pham@radixark.ai>
2026-10-23 19:01:26 -07:00
Yihao WangandClaude Opus 5 327c906764 [Perf] [XPU] diffusion: Avoid QKV copies in flash attention (#41944)
Co-authored-by: Claude Opus 5 <noreply@anthropic.com>
2026-10-24 09:33:26 +08:00
Yihao WangandClaude Opus 5 5bef70c0a2 [Perf] flux2: Build the single-block output projection without a cat (#41943)
Co-authored-by: Claude Opus 5 <noreply@anthropic.com>
2026-10-24 09:33:26 +08:00
MickandMick Qian 1d4dcc81cc [Perf] diffusion/registry: Reduce model discovery work at startup (#46043)
Co-authored-by: Mick Qian <mickqian@users.noreply.github.com>
2026-10-24 09:30:04 +08:00
47c0791ad1 [Perf] diffusion: Write uncompressed PNG responses without Pillow's row filters (#46099)
Co-authored-by: Mick Qian <mickqian@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-24 09:28:55 +08:00
bbc77e7db4 [Perf] diffusion: Cut LTX-2.3 per-block elementwise work (contiguous AdaLN values, gate-folded masks, vectorized head gate) (#46110)
Co-authored-by: Mick Qian <mickqian@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-24 09:27:39 +08:00
4e2d5b17ad [Perf] diffusion: Keep every component resident in auto mode when memory allows (#45978)
Co-authored-by: Mick Qian <mickqian@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-24 09:26:13 +08:00
be8600d02d [Fix] diffusion: Step MOVA latents outside the serial-CFG branch (#45993)
Co-authored-by: Mick Qian <mickqian@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-24 09:23:54 +08:00
a08bb0ec9e [Perf] diffusion: Run LTX-2 stage 2 on a premerged resident DiT (#46108)
Co-authored-by: Mick Qian <mickqian@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-24 09:23:08 +08:00
e66c9e981d [Perf] diffusion: Turn on breakable CUDA graphs by default for LTX-2/LTX-2.3 two-stage serving (#45977)
Co-authored-by: Mick Qian <mickqian@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-24 09:22:06 +08:00
2e6d87e2ca [Perf] diffusion: Serve FA4 head sizes on Blackwell instead of falling back to sm80 SDPA (#46102)
Co-authored-by: Mick Qian <mickqian@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-24 09:20:32 +08:00
386fa9ecfd [Perf] diffusion: Drop LTX-2's all-valid prompt mask under lossless quality (#46106)
Co-authored-by: Mick Qian <mickqian@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-24 09:19:12 +08:00
5b8955e70f [Perf] diffusion: Shard LTX-2 stage 2 over the CFG ranks as Ulysses (#45965)
Co-authored-by: Mick Qian <mickqian@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-24 09:17:41 +08:00
738e4535ea [Perf] diffusion/qwen-image-2.1: Keep an image prefix's K/V in the packed buffer across steps (#46101)
Co-authored-by: Mick Qian <mickqian@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-24 09:14:10 +08:00
3712b5944d [Perf] diffusion: Patchify Wan latents straight into tokens (#46148)
Co-authored-by: Mick Qian <mickqian@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-24 09:12:49 +08:00
MickandMick Qian 2d5ad3243f [Fix] diffusion/http: Allow bounded graceful shutdown (#46169)
Co-authored-by: Mick Qian <mickqian@users.noreply.github.com>
2026-10-24 09:11:15 +08:00
Jackey Huaandhnyls2002 e22a530859 [Feat] openai: Render DeepSeek-V4/V4.1 chat prompts with deepseek-recipe (#45315)
Co-authored-by: hnyls2002 <lsyincs@gmail.com>
2026-10-23 18:05:35 -07:00
weireweireandweireweire 49a2fbf817 [Fix] docker: Pin sgl-nvidia-mega to 3333e4c so the megakernel starts on main (#46030)
Co-authored-by: weireweire <20922698+weireweire@users.noreply.github.com>
2026-10-23 17:58:04 -07:00
Khoa PhamandClaude Fable 5.1 ee2dd727a3 [Perf] dsv41: Let Mega-mHC write the MegaMoE input under MegaGate (#46163)
Co-authored-by: Claude Fable 5.1 <noreply@anthropic.com>
2026-10-23 17:54:03 -07:00
Liangsheng Yin 79e5c4bb9b [Perf] dsv41: Trim per-step DSpark decode overhead (#46176) 2026-10-23 17:24:06 -07:00
Qiaolin Yu 5994c5597d [Perf] spec-decoding: fuse simulated speculative acceptance tensor preparation (#41169) 2026-10-23 17:15:21 -07:00
Baizhou Zhangandsid-rp 6a319e7b34 [Fix] muse: Support Responses namespace tools and defer_loading (#44685)
Co-authored-by: sid-rp <sidrp@meta.com>
2026-10-23 16:48:08 -07:00
0d592c475b [Fix] rust-server: Keep a transfer backend's own bootstrap port on prefill (#46066)
Co-authored-by: metamergebot <metamergebot@users.noreply.github.com>
Co-authored-by: shodoco <shodoco@users.noreply.github.com>
2026-10-23 16:47:14 -07:00
Cheng Wan 855c092deb [Fix] inkling: Keep BF16 LM-head operands with FP32 output (#45335) 2026-10-23 15:51:44 -07:00
Ziang Li 197be2203a [Doc] quantization: Link W4A16 design from environment reference (#46170) 2026-10-23 15:49:18 -07:00
Liangsheng Yin f05ad5ec01 [Perf] logits: Cast bf16 logits into the fp32 buffer with a Triton kernel (#46167) 2026-10-23 15:06:09 -07:00
Rohan VarmaandJun Wan 15dc838f9b [Fix] http_server: Apply SGLANG_TIMEOUT_KEEP_ALIVE to Granian HTTP/1.1 connections (#45760)
Signed-off-by: Rohan Varma <rohanv@nvidia.com>
Co-authored-by: Jun Wan <8323392+junwan01@users.noreply.github.com>
2026-10-23 14:43:54 -07:00
Baizhou Zhang ea226cf782 [Feat] glm5.3flash: Support padded CP BCG and prefix replay with PD/DFlash2 (#43774) 2026-10-23 14:21:31 -07:00
Yuhao YangandClaude Opus 5.5 5f7542d762 [Perf] dsv41: Build the MegaGate row masks once per forward (#46071)
Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
2026-10-23 14:15:01 -07:00
27ac614f05 [Feat] parser: Add checkpoint-driven response-template adapters (#40479)
Signed-off-by: Yoni Gozlan <yonigozlan@users.noreply.github.com>
Co-authored-by: Yoni Gozlan <yonigozlan@users.noreply.github.com>
Co-authored-by: Xinyuan Tong <xinyuantong.cs@gmail.com>
Co-authored-by: Jimmy Shong <69131491+Jiminator@users.noreply.github.com>
2026-10-23 14:05:54 -07:00
em-dreyer 4542fc8f59 [Fix] mistral3: Subclass LlavaForConditionalGeneration instead of wrapping it (#45656) 2026-10-23 14:05:32 -07:00
Michael 194abb8e54 [CI] [AMD] diffusion: Drain the Cosmos Dreams decode stream after freeing its outputs (#45983) 2026-10-23 13:26:56 -07:00
1d605b7ece [CI] vlm: Harden fixture cleanup after setup and shutdown failures (#45461)
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: Alison Shao <a.shao@wustl.edu>
2026-10-23 13:20:06 -07:00
005333d029 [Refactor] radix-tree: Make each cache action a typed CacheAction implementation (#45444)
Co-authored-by: metamergebot <metamergebot@users.noreply.github.com>
Co-authored-by: Jialin <Jialin@users.noreply.github.com>
2026-10-23 12:49:30 -07:00
Cheng Wan d0022e1df3 [Fix] lora: Preserve MoE projection identity and routing lifetimes (#45681) 2026-10-23 12:32:55 -07:00
Cheng Wan ac3db731b3 [Fix] moe: Preserve CUTLASS routing metadata and execution budgets (#45917) 2026-10-23 12:32:05 -07:00
Cheng Wan 8b9b992e41 [Perf] moe: Preserve Marlin workspace and route-count budgets (#45462) 2026-10-23 12:31:17 -07:00
0931dfedc9 [Fix] [AMD] dsa: Fix AITER prefill with an FP8 KV cache (#39083)
Co-authored-by: Pawel Kowalski <pawel.kowalski@silo.ai>
Co-authored-by: Thomas Wang <thomawan@amd.com>
2026-10-23 12:30:42 -07:00
Cheng Wan e6451c1a95 [Fix] gguf: Preserve MoE routing, weighting and allocation budgets (#45916) 2026-10-23 12:29:51 -07:00
Cheng Wan 1340044e97 [Fix] deepseek-v4: Preserve decoder-tail token-count ownership (#45915) 2026-10-23 12:28:43 -07:00
9d7ad308bd [Fix] [AMD] Quark: fix Quark block-FP8 loading in DeepSeek loader (#40776)
Co-authored-by: clintg6 <7388379+clintg6@users.noreply.github.com>
Co-authored-by: Thomas Wang <thomawan@amd.com>
Co-authored-by: Bowen Bao <bowenbao@amd.com>
Co-authored-by: Hubert Lu <Hubert.Lu@amd.com>
2026-10-23 12:11:21 -07:00
Alex NailsandClaude Opus 5.5 baf80c61b7 [Feat] platforms: Let out-of-tree platforms bind group devices and custom ops (#45897)
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-23 12:02:42 -07:00
Shangming Cai 1e894eb92d [CI] disaggregation: Run the SWA decode radix cache test on Mooncake (#46126) 2026-10-24 01:41:39 +08:00
Peng WuandClaude Opus 5.5 16aa816060 [Fix] disaggregation: Restart a lost prefill's heartbeat miss count (#45922)
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-24 01:26:10 +08:00
74c08f3dc5 [Feat] rust-server: Let a multimodal processor finalize item layout after caller hashes (#45914)
Co-authored-by: metamergebot <metamergebot@users.noreply.github.com>
Co-authored-by: Lu Fang <30275821+houseroad@users.noreply.github.com>
2026-10-23 09:26:52 -07:00
30dc2f9430 [Feat] benchmark: Add --input-len-list to bench_one_batch_server for per-request input lengths (#45982)
Co-authored-by: metamergebot <metamergebot@users.noreply.github.com>
Co-authored-by: wpc <wpc@users.noreply.github.com>
2026-10-23 09:04:59 -07:00
5a86d6ee17 [Fix] rust-server: Deliver scheduler-side aborts to Rust frontend clients (#45899)
Co-authored-by: metamergebot <metamergebot@users.noreply.github.com>
Co-authored-by: shodoco <shodoco@users.noreply.github.com>
2026-10-23 09:04:54 -07:00
eb8ef80842 [Feat] rust-server: Serve /abort_request from the Rust frontend (#45888)
Co-authored-by: metamergebot <metamergebot@users.noreply.github.com>
Co-authored-by: shodoco <shodoco@users.noreply.github.com>
2026-10-23 09:04:49 -07:00
Avaya Aggarwal efd2ce8a9d [Feat] comfyui: video-to-video input and prompt enhancement for server-mode nodes (#43673) 2026-10-23 23:09:30 +08:00
c677171c1d [Feat] hicache: Support MLA DCP with HiCache L3(file/mooncake) (#41187)
Co-authored-by: Codex <codex@openai.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: alphabetc1 <2508695655@qq.com>
Co-authored-by: Venkat Raman <venkat.raman@verda.com>
2026-10-23 23:09:04 +08:00
02b983603d [Fix] diffusion: Keep dynamic LoRA deltas on fused linear-GELU sites (#45943)
Co-authored-by: Mick Qian <mickqian@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-23 21:19:53 +08:00
0e34962884 [Perf] diffusion: Cut LTX-2 denoising-step host overhead (#45966)
Co-authored-by: Mick Qian <mickqian@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-23 21:18:57 +08:00
82ff16dad0 [Perf] diffusion: Skip the residency calibration probe in auto mode until a planner consumes it (#45979)
Co-authored-by: Mick Qian <mickqian@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-23 21:18:06 +08:00
3e5f5505ed [Perf] diffusion: Serve image responses from in-memory encoded bytes (#45987)
Co-authored-by: Mick Qian <mickqian@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-23 21:16:44 +08:00
2fbd094a8c [Fix] diffusion: Offset LTX-2.3 BCG video RoPE by the SP shard start (#45995)
Co-authored-by: Mick Qian <mickqian@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-23 21:14:37 +08:00
pllimax 9b46f27090 [Fix] [NPU] tests: Restore evalscope GSM8K baseline via NPUGSM8KMixin backend (#45732) 2026-10-23 21:06:21 +08:00
shun8686 2a46e32fbc [CI] [NPU] precise-test: skip recommend-tests-from-coverage when no NPU paths changed (#46055) 2026-10-23 21:01:42 +08:00
86933b1f33 [Fix] deps: Pin python-dateutil<3 to keep botocore resolvable (#45857)
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: Po-Han Huang <pohanh@nvidia.com>
2026-10-23 20:57:44 +08:00
shun8686 0347b2bb4e [CI] [NPU] precise-test: gate precision on PR judgeability, not bucket count (#46042) 2026-10-23 20:41:22 +08:00
Jimmy Shong 4b1941e889 [CI] diffusion: Run the FLUX.2 XPU norm-fusion test on the XPU suite only (#46040) 2026-10-23 19:54:03 +08:00
chuyehandCursor f967418a35 [Doc] cookbook: Qwen3.5 MI355X FP8 agentic adds scheduler/tokenizer tuning from the InferenceX recipe (#46032)
Co-authored-by: Cursor <cursoragent@cursor.com>
2026-10-23 19:12:05 +08:00
shun8686 535aeae2a7 [CI] [NPU] precise-test: run only the recommended multimodal_gen cases (#46014) 2026-10-23 18:06:58 +08:00
chuyehandCursor 69f830756e [Doc] cookbook: Qwen3.5 MI355X FP8 agentic uses HiCache write_through on the v0.5.22 image (#46017)
Co-authored-by: Cursor <cursoragent@cursor.com>
2026-10-23 17:47:52 +08:00
Sage 611a703061 [Chore] api: add fields to generate request (#45164)
Signed-off-by: Sage Ahrac <sagiahrak@gmail.com>
2026-10-23 17:10:23 +08:00
494eeecb6c [Fix] diffusion: Keep the linear-GELU verdict cache out of torch.compile traces (#45944)
Co-authored-by: Mick Qian <mickqian@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-23 17:01:51 +08:00
f9129915a3 [Fix] diffusion: Give every SP rank the same Wan TI2V timestep (#45994)
Co-authored-by: Mick Qian <mickqian@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-23 17:00:40 +08:00
Xinyuan Tong 8b9f72417f [Feat] Inkling: render and parse chat via tml-renderers (#45340) 2026-10-23 16:43:38 +08:00
Michael bb72b47daf [Test] [AMD] deepseek-v4.1: Give the encoder-replay test's request window a zero_row (#45728) 2026-10-23 01:38:54 -07:00
Shuwen Wang 77bc8cab18 [Fix] hicache/pd: Promise prefill only the L3 span the decode prefetch takes (#43548) 2026-10-23 16:32:41 +08:00
Michael 7d1a16d3eb [CI] [AMD] diffusion: Skip the LTX-2.5 diffusion decoder case on the 2-GPU lane (#45861) 2026-10-23 01:28:12 -07:00
2b8e4cde95 [Doc] [CPU] Cookbook: Update to the cookbook for models supported by Intel Xeon CPUs (#40092)
Co-authored-by: Niu Ziru <niuziru@a4bf018d3341.jf.intel.com>
Co-authored-by: Niu Ziru <niuziru@gnr799217.jf.intel.com>
2026-10-23 16:22:50 +08:00
Avaya Aggarwal 4b273537de [Fix] diffusion: runtime utils bug fixes (#44887) 2026-10-23 16:08:13 +08:00
Po-Han Huang (NVIDIA)andClaude Opus 5.5 02d2445138 [Doc] cookbook: Refresh SemiAnalysis AgentX launches with seven open InferenceX PRs (#46005)
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-23 16:04:01 +08:00
shun8686 472208fec0 [CI] [NPU] precise-test: recommend multimodal_gen cases at test-case granularity (#45955) 2026-10-23 15:57:31 +08:00
Yao MatrixandAlex Nails 6d6c9cd4ae [Fix] offloader: Preserve cached weights during CPU offload (#44139)
Co-authored-by: Alex Nails <alex.nails@radixark.ai>
2026-10-23 00:50:24 -07:00
Yihao WangandClaude Opus 5 6f466f655a [Perf] [XPU] diffusion/flux2: Fuse bit-exact LayerNorm and gated residual norm (#41942)
Co-authored-by: Claude Opus 5 <noreply@anthropic.com>
2026-10-23 15:38:12 +08:00
8708d2ee99 [Fix] diffusion/ltx: Keep tiled VAE decode off the sharded convs under parallel decode (#45942)
Co-authored-by: Mick Qian <mickqian@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-23 15:32:03 +08:00
4f47c48b33 [Fix] diffusion: Keep the N-rank IPC all-to-all's flags and staging out of expandable segments (#45948)
Co-authored-by: Mick Qian <mickqian@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-23 15:31:19 +08:00
1cd0ebb3d3 [Perf] diffusion: Remove per-step and per-request host syncs in Cosmos3 and Qwen-Image-2.1 (#45967)
Co-authored-by: Mick Qian <mickqian@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-23 15:30:09 +08:00
Avaya Aggarwal 4da6f585f3 [Fix] diffusion: OpenAI API hardening (#43585) 2026-10-23 15:22:00 +08:00
ShunkangzandZhangheng 79ab8f6d7c [Feat] KVShard: Support kv cache sharding for MTP (#40929)
Co-authored-by: Zhangheng <hzh0425@apache.org>
2026-10-23 15:20:19 +08:00
Banruo LiuandClaude Opus 5.5 6e667c4310 [Fix] disaggregation: Poison the deferred abort ack when a NIXL or Mori write fails (#45973)
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-23 14:47:59 +08:00
Liangsheng Yin 6c844eb12e [Perf] dsv4: Replay the V4.1 prefill graph past 16K contexts on SM90 (#45941) 2026-10-22 23:47:17 -07:00
ashwini rathi 6dd0dfabf0 [CI] [Intel] docker/xpu: Tag nightly images with main-whl and main-main kernel variant (#45854) 2026-10-23 14:44:55 +08:00
Jimmy Shong 2be6577ba9 [Doc] beam: Offer Python and Docker commands in the cookbook (#45991) 2026-10-22 23:39:10 -07:00
Chetan Kumar VermaandClaude Opus 5.5 76c382e80f [Perf] [XPU] diffusion: Use the AOT SYCL fused residual-norm-scale-shift kernel (#41401)
Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
2026-10-23 14:32:25 +08:00
ZiruandNiu Ziru 904691e0b5 [Feat] [CPU] model: Support Cosmos3 series models on CPU (#45478)
Co-authored-by: Niu Ziru <niuziru@gnr799217.jf.intel.com>
2026-10-23 14:24:28 +08:00
vikram singh shekhawat 3f36fc0833 [Fix] [XPU] runtime: Fix XPU rotary embeddings and tokenizer padding (#41213) 2026-10-23 14:09:59 +08:00
Liangsheng Yin 0e338777b0 [Perf] dsv4: Let the fused MXFP8 producers run inside the breakable prefill graph (#45980) 2026-10-22 23:00:10 -07:00
Pinjie Xu 7c680a0d45 [Perf] diffusion/cosmos3: optimize v2v preprocess performance (#45398)
Signed-off-by: pinjie <pinjiex@nvidia.com>
2026-10-23 13:57:38 +08:00
Michael 7baacd1d1b [CI] [AMD] docker: Add a build-only dispatch option to the ROCm nightly image workflows (#45882) 2026-10-23 13:41:02 +08:00
Khoa Pham c7994e7076 [Fix] dsv41: Keep MegaGate out of prefill graphs by gating on the forward mode (#45969) 2026-10-22 22:11:18 -07:00
Khoa Phamandhnyls2002 5bdbec0d48 [Perf] dsv41: Drop the MegaGate padded-row copy on sgl-deep-gemm 0.2.3.post2 (#45933)
Co-authored-by: hnyls2002 <lsyincs@gmail.com>
2026-10-22 22:10:57 -07:00
e83f585ae6 [Perf] dsv41: Serve token-interleaved V4.1 KV caches with sgl-flash-mla (#45893)
Co-authored-by: kpham-sgl <khoa.pham@radixark.ai>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-22 22:02:49 -07:00
Liangsheng Yin 49933325cc [Perf] dspark: Combine the draft's mHC streams from bf16 without an fp32 copy (#45964) 2026-10-22 21:44:01 -07:00
Mick f4a2df3dac [CI] diffusion: Calibrate single-GPU shard scheduling from CI (#45831) 2026-10-23 12:03:24 +08:00
Avaya Aggarwalandniehen6174 fa628c5359 [Fix] comfyui: Fix GGUF detection, import guard and VRAM accounting in SGLDiffusion plugin (#43536)
Co-authored-by: niehen6174 <1639206518@qq.com>
2026-10-23 11:58:23 +08:00
em-dreyerandClaude Opus 5.5 79eb10e182 [CI] [NVIDIA] gb300: Restore the aarch64 rust-ext cache on ARM runners (#45920)
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-22 20:01:42 -07:00
Wookje Han f2873f1613 [Perf] [NVIDIA] muse-spark-open: Fuse pre-expert RMSNorm with static-scale NVFP4 quantization (#44554) 2026-10-22 20:00:37 -07:00
Baizhou Zhang 698651acd0 [Feat] HiCache: Support fused DFLASH KV in unified hybrid SWA pools (#45700) 2026-10-22 20:00:28 -07:00
chenyang08056032 ce2261863f [CI] [NPU] precise-test: add empty-recommendation skip-all rule, slim baseline artifacts, and move setup-covstub off the NPU pool (#45571) 2026-10-23 10:52:14 +08:00
Liangsheng Yin a4838160b3 [Perf] dsv4: Write V4.1 K rows with one warp per token in the fused k-norm-rope store (#45934) 2026-10-22 19:50:50 -07:00
Liangsheng Yin c4fa35b829 [Perf] dsv4: Map V4.1 prefill index selections to pool slots in one kernel (#45938) 2026-10-22 19:50:31 -07:00
Liangsheng Yin 2da6c0c237 [Perf] dsv4: Combine the final mHC streams from bf16 without an fp32 copy (#45940) 2026-10-22 19:43:46 -07:00
MickandMick Qian 363e6c6b4d [Feat] diffusion/disaggregation: split FLUX.3 prompt and image encoders (#45309)
Co-authored-by: Mick Qian <mickqian@users.noreply.github.com>
2026-10-23 10:35:41 +08:00
Lianmin Zheng 5daa1352ba [CI] tests: Retire test/srt and reject legacy test paths (#42567) 2026-10-22 19:15:55 -07:00
Khoa PhamandClaude Opus 5.5 e5dc0c363a [Perf] dsv41: Integrate DeepGEMM MegaGate routing (#43262)
Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
2026-10-22 19:08:39 -07:00
YAMYandOasis-Git c82f7141ad [Perf] [NVIDIA] dspark: Move the host work between DSv4.1 decode graphs into the graphs for the megakernel path (#45450)
Co-authored-by: Oasis-Git <ayw.sirius19@gmail.com>
2026-10-22 19:08:06 -07:00
metamergebotandJialin 94e4611eac [Refactor] unified-cache: Factor match validators and match completion into helpers (#45714)
Co-authored-by: Jialin <Jialin@users.noreply.github.com>
2026-10-22 18:58:15 -07:00
Xinyuan Tong 92e57d5de7 [Perf] request_logger: skip per-token watermark redaction walk (#44692) 2026-10-22 18:57:19 -07:00
Liangsheng Yin b588aa39e9 [Perf] dsv4: Optionally stage V4.1 engram rows from the shared host table on the scheduler (#45921) 2026-10-22 18:35:55 -07:00
zijiexiaandClaude Opus 5.5 f36ee5de93 [Doc] Cookbook: add MiniMax-M3.1 (#45924)
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-23 09:07:35 +08:00
26396bf566 [Fix] [NVIDIA] deepseek-v4.1: Fix the breakable prefill CUDA graph in the late-layer tail on SM90 (#45407)
Co-authored-by: JohnZhao01 <325926370+JohnZhao01@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Co-authored-by: hnyls2002 <lsyincs@gmail.com>
2026-10-22 17:57:16 -07:00
Michael a95ae5f1c5 [CI] [AMD] docker: Start the ROCm 10 nightly at 10:00 UTC and ROCm 7.2 at 11:00 UTC (#45848) 2026-10-23 08:52:06 +08:00
Liangsheng Yin 97b79f0796 [Fix] prefill-graph: Keep every DP rank eager when the replay tail captures behind rank barriers (#45881) 2026-10-22 17:36:00 -07:00
Liangsheng Yin 9c7dd3c45f [Perf] dsv4: Dequantize each shared sparse-prefill KV source once per forward (#45909) 2026-10-22 17:08:40 -07:00
Liangsheng Yin ae4258715c [Perf] dsv4: Keep V4.1 image rows inside the Engram gate kernel (#45904) 2026-10-22 17:05:01 -07:00
HEJIAN SANGandClaude Opus 5.5 77f90ade97 [Fix] weight_checker: Make the checksum device-agnostic (#45918)
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-22 16:57:49 -07:00
bb9e9d5632 [Perf] mlp: Fuse dense SwiGLU with block-FP8 quantization on decode (#45219)
Co-authored-by: Jayden Yu <jaydenyu@fb.com>
Co-authored-by: Xiaoyu Zhang <1182563586@qq.com>
2026-10-23 07:49:39 +08:00
Chunan ZengandClaude Opus 5.5 5849ff5337 [Fix] [NVIDIA] jit: Build sm_100f on CC 10.7 until the bundled CUTLASS supports SM107 (#45911)
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-22 16:42:00 -07:00
Baizhou Zhang 8d47d515e4 [CI] deepgemm: Remove obsolete ref_mhc preflight check (#45910) 2026-10-22 16:27:40 -07:00
Michael 035ed73b9d [Fix] [AMD] kernels: Import the HIP launch helpers from jit.utils.hipcc in smallm_fp8_blockscale_gfx950 (#45676) 2026-10-22 16:19:34 -07:00
Alex Nails c394a90c2b [Perf] gemma4: Capture single-step frozen-KV drafts (#45771) 2026-10-22 15:37:19 -07:00
Liangsheng Yin f9b7225f2e [Perf] dsv4: Use the fused MXFP8 producers in eager prefill (#45903) 2026-10-22 15:29:09 -07:00
Liangsheng Yin 9c0f6d6a61 [Perf] dsv4: Write the MegaAttention output straight into the prefill graph buffer (#45901) 2026-10-22 14:56:08 -07:00
Liangsheng Yin 818d7932d7 [Perf] dsv4: Drop local references that keep residual-stream tensors alive through the prefill step (#45900) 2026-10-22 14:55:49 -07:00
3b0003dac1 [Fix] HiCache: stale host KV from write-through copies racing the in-flight forward (#42701)
Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Co-authored-by: Zhiqiang Xie <xiezhq@stanford.edu>
2026-10-22 14:46:51 -07:00
Alison Shao a904e4d31a [Fix] ci: Reduce runner utilization scan overhead and bound API retries (#45873) 2026-10-22 14:34:28 -07:00
Liangsheng Yin 840a6313c9 [Perf] dsv4: Score the V4.1 index logits in bf16 on SM100 (#45880) 2026-10-22 14:29:08 -07:00
25ad569ff9 [Feat] speculative: Seed rejection-sampling draft proposals for deterministic inference (#33395)
Co-authored-by: Ajit Mistry <ajmistry@nvidia.com>
Co-authored-by: Ajit Mistry <ajmistry@amistry.nvidia.com>
Co-authored-by: Po-Han Huang (NVIDIA) <53919306+nvpohanh@users.noreply.github.com>
Co-authored-by: Po-Han Huang <pohanh@nvidia.com>
2026-10-22 14:04:32 -07:00
AlexandJimmy Shong e5db41b329 [Fix] nemotron_vl: Pad visual spans using processor offsets (#44581)
Co-authored-by: Jimmy Shong <69131491+Jiminator@users.noreply.github.com>
2026-10-22 13:58:47 -07:00
786cdfc040 [Fix] nccl: Disable graph buffer registration whenever a captured graph can hold a collective (#43216)
Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
Co-authored-by: Jimmy Shong <69131491+Jiminator@users.noreply.github.com>
2026-10-22 13:56:19 -07:00
Yanbin Jiang 13999673ef [Feat] lora: Add plan-driven dense and MoE LoRA engines (#45892) 2026-10-22 13:40:56 -07:00
Shunkangz 154e29c1ec [Perf] CP: QSA Context Parallel Optimization (#43792) 2026-10-22 13:33:04 -07:00
Liangsheng Yin ba83415eea [Perf] engram: Read host-table rows in address order on GPUs without host page tables (#45879) 2026-10-22 13:13:43 -07:00
bf0e193bfd [Refactor] radix-tree: Move insert overlap handling into a resumable OVERLAP phase (#45702)
Co-authored-by: metamergebot <metamergebot@users.noreply.github.com>
Co-authored-by: Jialin <Jialin@users.noreply.github.com>
2026-10-22 13:12:37 -07:00
aae2145eb2 [Feat] radix-tree: Report the last visited node and a shorter FULL device match in match results (#45703)
Co-authored-by: metamergebot <metamergebot@users.noreply.github.com>
Co-authored-by: Jialin <Jialin@users.noreply.github.com>
2026-10-22 13:11:33 -07:00
357323e912 [Refactor] hicache: Route buffer-mode backup admission keys and FULL transfers through pipeline hooks (#45701)
Co-authored-by: metamergebot <metamergebot@users.noreply.github.com>
Co-authored-by: Jialin <Jialin@users.noreply.github.com>
2026-10-22 13:10:18 -07:00
Liangsheng Yin 4f96807ea8 [CI] radix-cache: Build unified radix cache test states through production paths (#45334) 2026-10-22 13:04:23 -07:00
YAMY d67d1ae02c [Fix] logits: Preserve lm_head precision with shared buffers (#45855) 2026-10-22 12:46:27 -07:00
Liangsheng Yin 0d6b283d7e [Refactor] dsv4: Keep only the two-level V4.1 indexer in the prefill graph (#45783) 2026-10-22 12:26:43 -07:00
WenhaoZhang 8d8916815b [Fix] ci: Add pytest entry point for Muse Spark router tests (#45843) 2026-10-22 11:39:34 -07:00
Zhangheng 913e91ed51 [Fix] CI: add missing entry point to Muse Spark router test (#45845) 2026-10-23 00:45:41 +08:00
Aarne Talman c80285ca9a [Fix] [AMD] deterministic: Refuse RCCL all-reduce in deterministic mode above TP2 (#43809)
Signed-off-by: Aarne Talman <aarne.talman@amd.com>
2026-10-22 09:31:25 -07:00
ShunkangzandClaude Opus 5.5 cc811ff0c3 [Feat] KVShard: Add attention and indexer support for page-sharded DSA caches (#44280)
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-23 00:16:35 +08:00
Shuwen Wang 8cd3c494a3 [Fix] mem_cache: Skip linker L3 lookup on PD decode (#45538) 2026-10-22 22:36:32 +08:00
Wenbo Ji 嵇文博 afd0d7b47b [Fix] diffusion/flux: use true_cfg_scale as the FLUX CFG weight (#44260) 2026-10-22 21:30:45 +08:00
Mick 0f7d3a0451 [CI] diffusion: target thirty-minute CI shards (#45761) 2026-10-22 21:19:42 +08:00
SunandMick dcf05724f7 [Perf] diffusion: accumulate chunked LoRA merges with addmm (#37620)
Co-authored-by: Mick <mickjagger19@icloud.com>
2026-10-22 21:18:39 +08:00
trebladev 698fa949ec [Perf] diffusion: fuse FLUX3 SwiGLU with rowwise FP8 quantization (#42758) 2026-10-22 21:17:56 +08:00
Song Bianandkernel-design-agents 635836129b [Perf] [KDA] muse_spark: use BLOCK_M 240 for BF16 MegaMoE prefill shards (#45637)
Co-authored-by: kernel-design-agents <kernel-design-agents@users.noreply.github.com>
2026-10-22 20:57:30 +08:00
gjsheu a9902f47e3 [Fix] [NPU] qwen: adapt pcp-v2 (#41564) 2026-10-22 20:54:11 +08:00
Song Bianandkernel-design-agents 851419717a [Perf] [KDA] muse_spark: Run the one-token router GEMV on CuTe DSL (#45508)
Co-authored-by: kernel-design-agents <kernel-design-agents@users.noreply.github.com>
2026-10-22 20:49:29 +08:00
Trevor Morris 91fc46a017 [Fix] [NVIDIA] docker: Disable torch 2.15 nccl2 in CUDA 13.4 container to fix EPLB rebalance crash (#45456) 2026-10-22 20:46:32 +08:00
Song Bianandkernel-design-agents 0ec3fee429 [Perf] [KDA] cookbook: enable custom all-reduce for GB300 BF16 Muse Spark (#45429)
Co-authored-by: kernel-design-agents <kernel-design-agents@users.noreply.github.com>
2026-10-22 20:46:28 +08:00
Song Bianandkernel-design-agents e43f6128a0 [Fix] [KDA] cookbook: add MoE runner backend to Muse Spark BF16 recipes (#45422)
Co-authored-by: kernel-design-agents <kernel-design-agents@users.noreply.github.com>
2026-10-22 20:38:35 +08:00
Jzz1943 f4ccfd4183 [Perf] moe: Add H200 block-FP8 Triton configurations (#45591) 2026-10-22 20:37:54 +08:00
MonishverandXiaoyu Zhang 178febef5d [Perf] qknorm: run head_dim-128 rows two per warp with 16B lanes, bit-identical (#45214)
Signed-off-by: Monishver Chandrasekaran <monishverchandrasekaran@gmail.com>
Co-authored-by: Xiaoyu Zhang <1182563586@qq.com>
2026-10-22 20:34:16 +08:00
81841d6bdc [Fix] hicache: Fail fast on a null pointer from the Mooncake host allocator (#40565)
Co-authored-by: huangtingwei <141888744+huangtingwei9988@users.noreply.github.com>
Co-authored-by: Teng Ma <sima.mt@alibaba-inc.com>
2026-10-22 20:12:25 +08:00
Jzz1943 8a6770d566 [Fix] NIXL: Preserve explicit UCX network device selection (#45722) 2026-10-22 20:07:27 +08:00
weireweireandweireweire 96f013b452 [Doc] Cookbook: update GB300 DSV4.1 disagg AgentX launches from InferenceX #4303 (#45495)
Co-authored-by: weireweire <20922698+weireweire@users.noreply.github.com>
2026-10-22 19:18:55 +08:00
Alex NailsandClaude Opus 5.5 d0a7565436 [Fix] diffusion: Drop stale mismatch_msg from FLUX 3 video VAE gate call (#45778)
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-22 01:31:54 -07:00
Liangsheng Yin 3f6c313389 [Fix] diffusion: Restore the kernel registry import after #41894 (#45775) 2026-10-22 01:25:13 -07:00
ashwini rathi 5c2cb60790 [Fix] [Intel] diffusion: Restore copy offload for FLUX.2-dev nightly (#45725) 2026-10-22 16:24:41 +08:00
Michael 0a87f163dc [Test] [AMD] deepseek-v4.1: Fix the HIP breakable-graph unit tests for the default prefill context limit and the DSpark draft ring (#45507) 2026-10-22 01:14:14 -07:00
kangwangamd 1c967f5043 [Fix] [AMD] test: Recalibrate MiniMax-M2.7 GSM8K threshold 0.93->0.91 for MI300 runners (#34241) 2026-10-22 01:08:54 -07:00
jacky.cheng b5137693a5 [Fix] [AMD] diffusion: Enable --quantization fp8 on gfx1250 (torch.compile lowering + per-channel weight scales) (#45304) 2026-10-22 01:05:36 -07:00
WenhaoZhangandMick Qian 24e45961c6 [Perf] diffusion: fuse QKV norm + RoPE, cached RoPE tables for FLUX 3 video VAE (#41894)
Co-authored-by: Mick Qian <mickqian@users.noreply.github.com>
2026-10-22 15:44:54 +08:00
James XuandCursor c64886f9e3 [Perf] diffusion: fuse Wan VAE DupUp3D shortcut add (#32860)
Co-authored-by: Cursor <cursoragent@cursor.com>
2026-10-22 15:38:51 +08:00
Ankit DhallandAnkit Dhall f727076ed0 [Feat] diffusion: support wan-animate-2 distilled model (#44085)
Co-authored-by: Ankit Dhall <8938083+ankitdhall@users.noreply.github.com>
2026-10-22 15:37:08 +08:00
Shuwen Wang 9d5386fd13 [Perf] dsv41: Fuse request-window layout construction (#45726) 2026-10-22 00:13:33 -07:00
shun8686 33a0734ca1 [CI] [NPU] precise-test: collect multimodal_gen coverage per test case (#45553) 2026-10-22 15:11:47 +08:00
黄孝君 569dda1e61 [CI] [NPU] npu: Update SGLang kernel tag to 2026.9.0.post10 (#45288) 2026-10-22 14:55:56 +08:00
Baizhou Zhang 779adb4b82 [Perf] cuda-graph: Raise MLA prefill capture default to chunked size capped at 8192 (#45565) 2026-10-21 23:49:25 -07:00
f68e82e7bc [Fix] epd: Accept tokens-in requests on scheduler-side transfer backends (#41376)
Signed-off-by: Krishnan Prashanth <kprashanth@nvidia.com>
Co-authored-by: Shangming Cai <csmthu@gmail.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-22 14:45:38 +08:00
Liangsheng Yin 82f7249ce6 [Perf] dsv4: Gather V4.1 index-K slots once per prefill graph replay (#45741) 2026-10-21 23:34:17 -07:00
Ziyi XuandClaude Opus 5.5 a56024c267 [CI] flash-mla: Add the sgl-flash-mla release workflow (#45651)
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-21 23:34:14 -07:00
Shangming Cai 9569a36561 [Feat] scheduler: Gate --prefill-decode-interval on running batch size (#44001) 2026-10-22 14:30:12 +08:00
Richard Wang ce66b3ab59 [Feat] systemone: serve Kev pointer head decision models (#45413) 2026-10-22 14:16:52 +08:00
Po-Han Huang (NVIDIA)andClaude Opus 5.5 12370da115 [Doc] cookbook: Refresh SemiAnalysis AgentX launches with five open InferenceX PRs (#45749)
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-22 14:15:19 +08:00
Shangming Cai a6a6beffac [Refactor] disaggregation: Share the flag all-reduce and host-destination predicate in the decode transfer queue (#45641) 2026-10-22 13:42:55 +08:00
WenhaoZhangandMick d9ce511646 [Feat] diffusion: support FastH3 Omni (Ref2VA 8-step PDD) (#45530)
Co-authored-by: Mick <mickjagger19@icloud.com>
2026-10-22 13:07:13 +08:00
MickandMick Qian 3472e16f96 [Refactor] diffusion: remove the deprecated kitchen_int8 alias (#45706)
Co-authored-by: Mick Qian <mickqian@users.noreply.github.com>
2026-10-22 13:06:07 +08:00
Shuwen Wang 80263b5580 [Fix] sgl-router: Do not abort forwarded requests on reqwest's default 30 s TCP_USER_TIMEOUT (#45696) 2026-10-22 12:43:54 +08:00
zijieandZijie Chen f91b81b82a [Doc] [AMD] cookbook: Add Muse Spark Open MI355X AgentX recipe (#45689)
Co-authored-by: Zijie Chen <300606707+zijiecode@users.noreply.github.com>
2026-10-21 21:12:52 -07:00
ApeiriaandRain Jiang 8d7df07a56 [Feat] rust-server: Add HiCache storage management endpoints (#44726)
Signed-off-by: 0130w <1157374677@qq.com>
Co-authored-by: Rain Jiang <96632942+rainj-me@users.noreply.github.com>
2026-10-22 04:01:35 +00:00
Liangsheng Yin 17babd67b4 [Fix] dsv4: Never read attention metadata for idle SWA stores (#45699) 2026-10-21 20:56:40 -07:00
WenhaoZhang 45de4adde9 [Feat] diffusion: Add worker-side ComfyUI compatibility patches (#44841) 2026-10-21 20:41:51 -07:00
Thomas Wang 9ee9858805 [Doc] [AMD] dsv41: Update amd cookbook 1021 (#45693) 2026-10-21 20:33:34 -07:00
Khoa Pham f1e26dc658 [Perf] logits: Write lm_head logits straight into the fp32 logits buffer (#45631) 2026-10-21 20:32:16 -07:00
WenhaoZhang 1962cbeebe [Perf] diffusion: convrot_int8 comfy_kitchen issues one GEMM per layer by default (#44179) 2026-10-22 11:30:32 +08:00
Viraat ChandraandClaude Opus 5.5 5f62e23d30 [Doc] cookbook: Add Muse Spark Open GB300 Dynamo + SGLang deployments measured with AA AgentPerf (#45629)
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-22 11:24:58 +08:00
Shuwen Wang 6929afb528 [Fix] hisparse: Keep PD destination host pages until the decode KV release (#45108) 2026-10-22 11:19:21 +08:00
8dc7f98005 [Perf] pd: Send prefill KV chunks without a host sync on the scheduler thread (#45492)
Co-authored-by: alphabetc1 <2508695655@qq.com>
Co-authored-by: Shuwen Wang <47200617+alphabetc1@users.noreply.github.com>
2026-10-21 20:11:16 -07:00
3468a65b51 [CI] labeler: Add a mooncake PR label (#45578)
Co-authored-by: Cursor Agent <cursoragent@cursor.com>
Co-authored-by: Teng Ma <stmatengss@users.noreply.github.com>
2026-10-22 11:09:48 +08:00
MickandMick Qian 780a5b1acc [Doc] diffusion: document prompt enhancer tasks and link Cosmos recipes (#45705)
Co-authored-by: Mick Qian <mickqian@users.noreply.github.com>
2026-10-22 10:28:58 +08:00
Tamir BaydasovandClaude Opus 5.5 b057d5b3f1 [Fix] [NPU] mla-pool: DCP allocator page and index-K bound, short-pool warning (#42976)
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-22 05:24:14 +03:00
b453996b7d [Doc] diffusion/minimax-h3: Refresh VDN-H3 B200 benchmarks, fix 8-step bench presets and popular-models snippet (#45267)
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: zijiexia <37504505+zijiexia@users.noreply.github.com>
2026-10-22 10:09:13 +08:00
em-dreyerandClaude Opus 5.5 3d508d614f [CI] [NVIDIA] gb300: Run DSV4.1-Flash tests on 2 GPUs with engram host tables (#45294)
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-21 18:53:24 -07:00
wutao 67bbfcee3c [Perf] [NPU] scheduler: Remove blocking seq_lens D2H from graph replay (#45170) 2026-10-22 09:50:04 +08:00
Liangsheng Yin d61f618276 [Perf] spec: Let DSPARK prefill publish read-done by placing its draft KV inject before the target forward (#45682) 2026-10-21 18:40:44 -07:00
4049941a82 [Fix] [NPU] scope: fix dtype error in sampling backend ascend (#39609)
Co-authored-by: ronnie_zheng <zl19940307@163.com>
Co-authored-by: Артем Савкин <58187114+OrangeRedeng@users.noreply.github.com>
2026-10-22 04:34:29 +03:00
Khoa PhamandClaude Opus 5.5 aa10927271 [Perf] dsv41: Mega-mHC and the wo_a einsum write MXFP8 for wqkv_a and wo_b (#45632)
Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
2026-10-21 18:27:21 -07:00
Baizhou Zhang 59149553b9 [Refactor] attention: centralize linear and draft TP group selection (#45684) 2026-10-21 18:25:40 -07:00
Rohan VarmaandYangmin Li 9ec74a53cc [Fix] [NVIDIA] docker: Build sgl-deep-gemm 0.2.3 for cu134 to match the sglang pin (#45521)
Co-authored-by: Yangmin Li <yangminl@nvidia.com>
2026-10-22 09:01:05 +08:00
Jimmy Shong 9301ca37a0 [Fix] modelopt: Correct NVFP4 TRT-LLM clamps for untied gate/up scales (#45386) 2026-10-21 17:37:07 -07:00
Michael 5fff58a169 [CI] [AMD] miles: Fix the ROCm 10 nightly suites that never finish (#45469) 2026-10-21 17:24:51 -07:00
eacd0a07d7 [Feat] bench: support --tokenize-prompt for generated-shared-prefix (#40900)
Co-authored-by: Yihao Wang <42559837+AgainstEntropy@users.noreply.github.com>
Co-authored-by: chaoyujiang <chaoyujiang@hygon.com>
Co-authored-by: Shangming Cai <csmthu@gmail.com>
2026-10-22 08:17:15 +08:00
Khoa Pham ed89693c4a [Perf] dsv41: Fuse the q_norm MXFP8 quantize for wq_b at every row count (#45633) 2026-10-21 16:56:45 -07:00
Liangsheng Yin da7b97a737 [Perf] dsv4: Publish prefill read-done under the request window and on V4.1 (#45582) 2026-10-21 16:29:28 -07:00
Liangsheng Yin 3438ed660d [Test] cpu: prune mirror and tautology cases from the per-commit CPU suite (#45672) 2026-10-21 16:02:47 -07:00
ed1f56fead [Fix] [Apple] mlx: Give the stub model a shared-expert layout (#45419)
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: R0CKSTAR <yeahdongcn@gmail.com>
2026-10-21 15:54:49 -07:00
Baizhou Zhang f4035cab0b [Feat] cp: Enable GLM-5.3 Flash interleave BCG with batched prefill and DFlash2 (#43182) 2026-10-21 15:51:28 -07:00
42db2db4a8 [Fix] rust-server: Check only a client-sent max_new_tokens against the context window (#45460)
Co-authored-by: metamergebot <metamergebot@users.noreply.github.com>
Co-authored-by: shodoco <shodoco@users.noreply.github.com>
2026-10-21 15:50:13 -07:00
991b846227 [Perf] speculative: Parallelize EAGLE spec-decode KV index building over token blocks (flashinfer draft decode; verify/draft-extend for flashinfer, flashinfer_mla, aiter) (#42803)
Co-authored-by: Claude Fable 5.1 <noreply@anthropic.com>
Co-authored-by: YAMY1234 <74099316+YAMY1234@users.noreply.github.com>
Co-authored-by: Po-Han Huang (NVIDIA) <53919306+nvpohanh@users.noreply.github.com>
Co-authored-by: Po-Han Huang <pohanh@nvidia.com>
2026-10-21 15:26:35 -07:00
maocheng23 4b9936b5e3 [Fix] deepseek_v4: Fix mHC prewarm when chunked prefill is disabled (#43448) 2026-10-21 15:21:31 -07:00
Liangsheng Yin 99cf403c9e [Fix] dsv4: Write idle SWA stores to an idle request-window layout (#45586) 2026-10-21 15:18:34 -07:00
em-dreyer aa5880e2b8 [Fix] bailing: Renormalize top-k routing weights by default for Ling-2.5/2.6 (#45662) 2026-10-21 14:53:58 -07:00
Michael 51baa1f6e9 [Test] [AMD] debug-utils: Opt TestCrashDump into expect_server_exit (#45408) 2026-10-21 14:49:53 -07:00
Liangsheng Yin f3f0f7b68c [Perf] dsv4: Let the decoder tail leave the prefill graph under DP attention with MegaMoE (#45581) 2026-10-21 14:44:21 -07:00
Liangsheng Yin f5eca05b05 [Fix] [SM100] quantization: Honor deep_gemm for 32x32 block-FP8 linears (#45585) 2026-10-21 14:42:50 -07:00
29e78045fb [Fix] dsv4: Skip DP-attention pad rows in the engram hasher and low-ratio sources (#40424)
Co-authored-by: jiashuai <jiashuai@baidu.com>
Co-authored-by: hnyls2002 <lsyincs@gmail.com>
2026-10-21 14:37:55 -07:00
Liangsheng Yin 68623cae92 [Fix] test: Gather Mamba track indices from track_position in the mixed-batch track test (#45670) 2026-10-21 14:33:48 -07:00
Liangsheng Yin f7d4f06242 [Perf] dsv4: Compute request-window history rows once per batch (#45579) 2026-10-21 14:01:19 -07:00
Liangsheng Yin bc412213c1 [Perf] dsv4: Gather V4.1 index-K slots once per forward and cache the prefill index plan (#45577) 2026-10-21 14:00:35 -07:00
Liangsheng Yin 620718c8ea [Fix] moe: Keep the DeepSeek V4.1 vision router off the packed-ids path under routing simulation (#45575) 2026-10-21 13:58:44 -07:00
Liangsheng Yin 97945831f9 [Perf] dsv4: Restore sparse prefill for request-window rows (#45573) 2026-10-21 13:54:32 -07:00
Liangsheng Yinandlixiufei-leo 2c7a9e3b50 [Fix] radix-cache: Order linker offloads behind the forward that writes the pages (#45559)
Co-authored-by: lixiufei-leo <294029388+lixiufei-leo@users.noreply.github.com>
2026-10-21 13:53:37 -07:00
Khoa Pham 028de2c5d4 [Perf] dsv41: Use tuned MXFP8 tactics at decode sizes (#45630) 2026-10-21 13:35:56 -07:00
Yoni GozlanandJimmy Shong a2e4ffbcea [Fix] parser: Accept whole-number floats for integer tool arguments (#43941)
Co-authored-by: Jimmy Shong <69131491+Jiminator@users.noreply.github.com>
2026-10-21 13:04:15 -07:00
Michael 7e0e259435 [CI] [AMD] disaggregation: Register Mooncake KV buffers without HIP dmabuf (#45519) 2026-10-21 12:47:42 -07:00
a3d00eff58 [Perf] srt: Record on the host whether an extend batch completes a prompt (#45487)
Co-authored-by: Lu Fang <116399278+luccafong@users.noreply.github.com>
Co-authored-by: shodoco <shodoco@users.noreply.github.com>
2026-10-21 11:58:25 -07:00
7e5256d1ba [Feat] srt: Balance uvicorn tokenizer workers with per-worker SO_REUSEPORT listeners (#45482)
Co-authored-by: luccafong <116399278+luccafong@users.noreply.github.com>
Co-authored-by: joshuadeng <joshuadeng@users.noreply.github.com>
2026-10-21 11:57:26 -07:00
Baizhou Zhang f06d1bb1f4 [Perf] cuda-graph: Capture FlashInfer KDA prefill (#45473) 2026-10-21 11:57:23 -07:00
740d5f9f9e [Perf] scheduler: Plan hybrid Mamba extend batches without host-device syncs (#45472)
Co-authored-by: Lu Fang <116399278+luccafong@users.noreply.github.com>
Co-authored-by: Shiyan Deng <20773414+842974287@users.noreply.github.com>
2026-10-21 11:56:50 -07:00
5e79333c95 [Refactor] radix-tree: Reuse the insert walk, component hooks and bindings for custom tree cores (#45439)
Co-authored-by: metamergebot <metamergebot@users.noreply.github.com>
Co-authored-by: Jialin <Jialin@users.noreply.github.com>
2026-10-21 10:53:26 -07:00
paulzhang-tm b57b5e2286 [Fix] disaggregation: Synchronize deferred decode KV release across TP ranks (#45453) 2026-10-22 01:37:28 +08:00
Khoa Pham ed268dd320 [Fix] moe: Apply SGLANG_SIMULATE_* routing on DeepSeek-V4.1 vision-biased gates (#45628) 2026-10-21 10:00:15 -07:00
vorapolsiloai 6e98c84c60 [Perf] [AMD] quark: Add opt-in AITER ASM FP4 GEMM for Quark dense linear (#39598) 2026-10-21 09:37:47 -07:00
Michael 4b1a44bc77 [Test] [AMD] diffusion: Avoid Python 3.11-only enterContext in test_usp_ring_replicated (#45409) 2026-10-22 00:11:08 +08:00
zijiexia 75878f27c8 [Doc] cookbook: Add Recursive Flint cookbook (#45625) 2026-10-21 08:58:16 -07:00
Shiyang Lyu cdbd478422 [Doc] [NPU] cookbook: add tested A2 deployment and A3 reference for MiniMax-M2.5 (#40120) 2026-10-21 22:59:37 +08:00
Huang LongandShangming Cai af49ada16d [Fix] pd: Allow runtime role switching with DP and DP attention (#44215)
Co-authored-by: Shangming Cai <csmthu@gmail.com>
2026-10-21 22:58:29 +08:00
Peng Xingchen 7163e97462 [Doc] [NPU] GLM: Add Ascend cookbook deployment recipes (#45597) 2026-10-21 22:44:08 +08:00
8e3e2980db [Feat] [Intel] disaggregation: Support XPU for staging-buffer KV transfer (#40861)
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: Shangming Cai <csmthu@gmail.com>
2026-10-21 22:09:04 +08:00
pllimax 7e5d3780ef [Fix] [NPU] tests: Run the HiCache Mamba case with bfloat16 SSM kernels (#45593) 2026-10-21 21:48:12 +08:00
Shawn Yang dd9ea06c00 [Fix] rust: Format health probe handle setup (#45606) 2026-10-21 21:40:08 +08:00
unclezhou e109bc1832 [Doc] [NPU] deepseek-v4: Add A3/950 single-node cookbook recipes (#45602) 2026-10-21 21:16:49 +08:00
amote-i 2815eed49c [Fix] [NPU] pre_warm_nccl: init npu backend before the parallel runtime touches a device (#43039) 2026-10-21 19:44:30 +08:00
amote-i fda84f80b7 [Doc] [NPU] docs: reference released Qwen3.8-Max W4A8 checkpoint in tutorial (#45574) 2026-10-21 19:38:17 +08:00
Shangming Cai a175e39b84 [CI] cuda: Rebuild workspace crates when the shared Cargo target dir came from another rust/ tree (#45557) 2026-10-21 19:32:22 +08:00
Avaya Aggarwalandniehen6174 62b4127668 [Feat] comfyui: Apply ControlNet residuals in Flux integrated mode (#43698)
Co-authored-by: niehen6174 <1639206518@qq.com>
2026-10-21 19:29:43 +08:00
Shangming Cai 43a3f956cb [Fix] sgl-router: Wait for the detached PD prefill in routing tests (#45502) 2026-10-21 19:04:26 +08:00
pllimax 43728298fe [Fix] [NPU] qwen3_6: Set accuracy cases to greedy decoding and tune retries (#40889) 2026-10-21 18:43:30 +08:00
ca72dad953 [Fix] rust-server: Answer health probes a busy scheduler skipped on the response ring (#45475)
Co-authored-by: Lu Fang <116399278+luccafong@users.noreply.github.com>
Co-authored-by: shodoco <shodoco@users.noreply.github.com>
2026-10-21 03:41:36 -07:00
Jimmy Shong 5bf10d4c79 [Fix] nemotron: Correct weight-export failures and GEMV dispatch (#45082) 2026-10-21 18:38:35 +08:00
Jzz1943 2e7c9d851c [Feat] NIXL: Support per-GPU UCX device selection (#43482) 2026-10-21 17:54:11 +08:00
bz f 9c10e0dd61 [Doc] [NPU] cookbook: add NPU (A3 series) deployment scenarios for Qwen3.6 (#38678) 2026-10-21 17:37:24 +08:00
Shangming Cai 7554be5a69 [CI] codeowners: Add owners for the top-level sgl-router (#45564) 2026-10-21 17:28:21 +08:00
kangwangamd 131c42ba33 [Perf] [AMD] aiter: Update ROCm AITER pin to 285579b62 (gfx942/gfx950 only) (#45498) 2026-10-21 02:25:02 -07:00
Shangming Cai be65391abd [Feat] mooncake: Enable the failed-session probe by default with backoff and idle pruning (#45338) 2026-10-21 17:16:19 +08:00
Thomas Wang 1278087336 [Fix] [AMD] mono-kernel: Update sgl-amd-mono version (#45546) 2026-10-21 02:02:03 -07:00
2c9d919071 [Fix] srt: Pin the source address of TCP ZMQ connects for the peer allowlist (#45437)
Co-authored-by: Lu Fang <116399278+luccafong@users.noreply.github.com>
Co-authored-by: fei-xx <fei-xx@users.noreply.github.com>
2026-10-21 01:59:03 -07:00
21f775cd97 [Feat] rust-server: Let the launcher set the default max_new_tokens for /generate (#45431)
Co-authored-by: metamergebot <metamergebot@users.noreply.github.com>
Co-authored-by: shodoco <shodoco@users.noreply.github.com>
2026-10-21 01:58:34 -07:00
Aurick QiaoandXiaoyu Zhang 7a632cf998 [Fix] quantization: Preserve BF16 NaNs in per-token FP8 quantization (#42689)
Co-authored-by: Xiaoyu Zhang <1182563586@qq.com>
2026-10-21 16:52:28 +08:00
8bd6aa9a6c [Feat] parser: Enrich chat_parsing streaming events (#40478)
Co-authored-by: Yoni Gozlan <yonigozlan@users.noreply.github.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
Co-authored-by: Xinyuan Tong <xinyuantong.cs@gmail.com>
Co-authored-by: Jimmy Shong <69131491+Jiminator@users.noreply.github.com>
2026-10-21 01:46:26 -07:00
AlexandJimmy Shong dbe0a9a7c9 [Fix] nemotron_vl: Preserve tokenized prompts during media expansion (#44184)
Co-authored-by: Jimmy Shong <69131491+Jiminator@users.noreply.github.com>
2026-10-21 01:40:51 -07:00
Liangsheng Yin 6efdd33d07 [Refactor] spec: Move platform dispatch and Triton kernels into kernels/ops (#45504) 2026-10-21 01:22:16 -07:00
6cac4dc3fe [Refactor] diffusion: Rollout controller as a scheduler-step plugin instead of mixins (#44603)
Signed-off-by: rockdu <kangrdu@gmail.com>
Co-authored-by: WenhaoZhang <42087078+niehen6174@users.noreply.github.com>
Co-authored-by: JingwenGu0829 <75733630+JingwenGu0829@users.noreply.github.com>
2026-10-21 16:04:33 +08:00
Avaya AggarwalandWenhaoZhang fee32a5c1f [Feat] ComfyUI: Add online INT8 quantization for BF16 Qwen-Image-2.1 single files (#44865)
Co-authored-by: WenhaoZhang <42087078+niehen6174@users.noreply.github.com>
2026-10-21 16:01:18 +08:00
Jzz1943 a361f0c410 [Fix] kda: Align graph tracking branch dtype with the recurrent state pool (#44886) 2026-10-21 15:58:34 +08:00
f7b8c6fe40 [Fix] glm-5.2: Record side-stream work right before its join to keep CUDA-graph replay on one stream (#45349)
Co-authored-by: Claude Sonnet 5.5 <noreply@anthropic.com>
Co-authored-by: Po-Han Huang (NVIDIA) <53919306+nvpohanh@users.noreply.github.com>
2026-10-21 15:36:54 +08:00
Hongkuan Zhouandishandhanani 8acfabbd67 [Fix] observability: Correct speculative FPM timing and decode lengths (#38711)
Signed-off-by: hongkuanz <hongkuanz@nvidia.com>
Co-authored-by: ishandhanani <82981111+ishandhanani@users.noreply.github.com>
2026-10-21 15:31:18 +08:00
Shangming Cai cd61db49c7 [Fix] sgl-router: Bump rustls-webpki and webpki-roots for emSign advisories (#45526) 2026-10-21 15:24:11 +08:00
YC Yen-Ching Tseng 132836b369 [Fix] [AMD] docker: Build Triton from source for ROCm 10.1 gfx1250 (#45329) 2026-10-21 15:14:24 +08:00
Peng Xingchenandpengxingchen0810 4449d7ecb0 [Doc] [NPU] GLM: Add GLM-5.3 and GLM-5.3-Flash deployment tutorials (#43372)
Co-authored-by: pengxingchen0810 <pengxingchen0810@noreply.gitcode.com>
2026-10-21 15:02:58 +08:00
Junlin Wu 844d2b5dec [Fix] [NPU] qwen3.5: Fix dense serving on Ascend 950 (#32745) 2026-10-21 09:50:38 +03:00
Jimmy Shong 8599b7251f [Chore] deps: Bump the OpenAI SDK to 2.8.0 for Responses shell tools (2/8) (#45291) 2026-10-20 23:49:30 -07:00
wutao fb6e6a09ce [Fix] [NPU] disaggregation: Fix host-pool backup for decode retraction (#43814) 2026-10-21 14:48:01 +08:00
chenyang08056032 c3b2b55bbb [CI] [NPU] precise-test: add recommendation summary & run artifacts, fix NPU-only coverage view, and quiet benign teardown noise (#45374) 2026-10-21 14:40:50 +08:00
Khoa PhamandAMD-yanfeiwang da7e3f1c12 [Fix] cache: Accept packed draft layer mappings in the UMBP linker (#44929)
Co-authored-by: AMD-yanfeiwang <yanfei.wang@amd.com>
2026-10-20 23:27:22 -07:00
Thomas Wang bd5a445bf7 [Fix] [AMD] dsv41: Set moe act to fp8 by default (#45480) 2026-10-20 23:09:27 -07:00
Shangming Cai 574f9cdb2e [CI] slash-command: Add /run-cpu-ci to run only the CPU unit suite (#45501) 2026-10-21 13:59:16 +08:00
956b9a24ee [Perf] [AMD] qwen3.5: Fuse activation quant into small-M FP8 block-scale GEMM on gfx950 (#42229)
Co-authored-by: Cursor <cursoragent@cursor.com>
Co-authored-by: jacky.cheng <yichiche@amd.com>
2026-10-20 22:32:30 -07:00
Po-Han Huang (NVIDIA)andClaude Opus 5.5 1689cb2d52 [Doc] cookbook: Add GLM-5.3 AgentX section and refresh with eight open InferenceX PRs (#45513)
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-21 13:22:38 +08:00
Aarne Talman 4b5c9c4b5e [Perf] [AMD] attention: Restore the gfx950 extend tile for head_dim <= 128 (#45099) 2026-10-20 22:02:50 -07:00
32c26788f3 [Fix] mamba: Keep prefill rows' checkpoint tracking through mixed chunked prefill (#45435)
Co-authored-by: metamergebot <metamergebot@users.noreply.github.com>
Co-authored-by: YavorGIvanov <YavorGIvanov@users.noreply.github.com>
2026-10-20 21:56:50 -07:00
Liangsheng Yin 573b5716f5 [Fix] spec: Take post-verify KV move targets from the verify window (#45491) 2026-10-20 21:12:41 -07:00
1c2f42470c [Fix] disaggregation: Keep a requeued optimistic prefill request's multimodal inputs (#45440)
Co-authored-by: metamergebot <metamergebot@users.noreply.github.com>
Co-authored-by: libertyeagle <30348494+libertyeagle@users.noreply.github.com>
Co-authored-by: fei-xx <135589532+fei-xx@users.noreply.github.com>
2026-10-20 20:27:34 -07:00
metamergebotandshodoco 6bd1c76735 [Fix] rust-server: Drop the fixed 64-item multimodal cap in favor of --limit-mm-data-per-request (#45432)
Co-authored-by: shodoco <shodoco@users.noreply.github.com>
2026-10-20 20:25:22 -07:00
Kan WuandClaude Opus 5.5 aba892fbdf [Refactor] sgl-router: Move experimental/sgl-router to top-level sgl-router (#45414)
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-21 11:09:13 +08:00
eb91a41f38 [Feat] unified_cache: Support FULL + MAMBA MLA models (Kimi-K3) in the direct external linker (#44977)
Co-authored-by: Yutong Wu <yutong.wu@amd.com>
Co-authored-by: Zhangheng <hzh0425@apache.org>
2026-10-21 10:47:17 +08:00
metamergebot f1ba723c64 [Test] [AMD] moe: Expect gfx1250 W4A8 layers to build with expert parallelism (#45476) 2026-10-21 10:45:31 +08:00
Chunan Zeng 93088901fb [Chore] attention: Remove the tokenspeed_mla backend and dependency (#44803) 2026-10-20 18:50:05 -07:00
unclezhou c0d18c0254 [Fix] [NPU] dspark: Fix attn_tp > 1 draft wo_b all_reduce group (#40399) 2026-10-21 09:20:15 +08:00
c1c287a429 [Doc] cookbook: add TPU7x recipe for GLM-5.3-Flash (#45326)
Co-authored-by: Qinghan Chen <qinghan.chen@radixark.net>
Co-authored-by: zijiexia <37504505+zijiexia@users.noreply.github.com>
2026-10-20 17:10:18 -07:00
b326e8349d [Perf] mamba: Copy tracked decode states at HBM bandwidth in one persistent launch (#45213)
Co-authored-by: metamergebot <metamergebot@users.noreply.github.com>
Co-authored-by: YavorGIvanov <YavorGIvanov@users.noreply.github.com>
2026-10-20 16:52:15 -07:00
Baizhou Zhang 5b1dbc2c80 [Fix] [NVIDIA] modelopt-nvfp4: Keep MoE global scales in gate/up order (#45289) 2026-10-20 16:49:13 -07:00
kkandwunhuang 9686edcf80 [Perf] [AMD] dsv41: faster mHC boundary Triton tile for 96+ rows, bitwise unchanged (#45124)
Co-authored-by: wunhuang <wunhuang@amd.com>
2026-10-20 16:39:34 -07:00
Cheng Wan 027208ae02 [Doc] layer-boundary/moe: Correct guides and comments that drifted from the code (#44896) 2026-10-20 16:26:40 -07:00
Cheng WanandAditya Suresh 299a75ac5d [Refactor] moe: Derive routing inputs from the layer and active forward (#44994)
Co-authored-by: Aditya Suresh <152752311+adi-suresh01@users.noreply.github.com>
2026-10-20 16:26:15 -07:00
kkandwunhuang 533020b500 [Perf] [AMD] dsv41: run gfx950 MXFP8 dense GEMMs on aiter's preshuffled FlyDSL kernel (#45123)
Co-authored-by: wunhuang <wunhuang@amd.com>
2026-10-20 16:19:32 -07:00
Cheng WanandAditya Suresh 4f8f06faf3 [Refactor] moe: Derive layer contracts from the selected quant method (#44993)
Co-authored-by: Aditya Suresh <152752311+adi-suresh01@users.noreply.github.com>
2026-10-20 16:18:45 -07:00
zydoandCheng Wan 7b1d5cd02d [Doc] agent-skills: Fix stale paths and deleted guard tests in skills and rules (#43666)
Co-authored-by: Cheng Wan <cheng.wan@radixark.ai>
2026-10-20 16:15:10 -07:00
bf145750a6 [Perf] prefill-cuda-graph: Make prefill shared-read completion explicit and default-on (#37958)
Co-authored-by: weireweire <20922698+weireweire@users.noreply.github.com>
Co-authored-by: Po-Han Huang (NVIDIA) <53919306+nvpohanh@users.noreply.github.com>
Co-authored-by: Po-Han Huang <pohanh@nvidia.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: hnyls2002 <lsyincs@gmail.com>
2026-10-20 16:12:55 -07:00
Aarne Talman d5e83ea4f6 [Perf] [AMD] deterministic: Faster extend attention launch on gfx950 (#45098) 2026-10-20 16:11:01 -07:00
averyjones4andaveryjones4 c1e43b95f4 [Doc] cookbook: Add DeepReinforce Ornith-1.5 (#45455)
Co-authored-by: averyjones4 <averyjones4@users.noreply.github.com>
2026-10-20 16:08:53 -07:00
Cheng WanandAditya Suresh f1eec1232d [Fix] config: Preserve ServerArgs positional constructor slots (#44991)
Co-authored-by: Aditya Suresh <152752311+adi-suresh01@users.noreply.github.com>
2026-10-20 16:08:28 -07:00
Aditya SureshandCheng Wan e2e1a97fc3 [Fix] moe: Map expert ids for layers whose runner needs local ids under EP (#44480)
Co-authored-by: Cheng Wan <cheng.wan@radixark.ai>
2026-10-20 16:07:48 -07:00
Baizhou Zhang b008ec3de6 [CI] muse_spark: Move unified memory test to B300 (#45454) 2026-10-20 15:56:49 -07:00
Liangsheng Yin 1dd0f361b8 [Fix] moe: Run the shared-slot routing test without a published server config (#45448) 2026-10-20 15:03:47 -07:00
Baizhou Zhang 583cfc4d55 [CI] muse_spark: Use complete local checkpoints in model tests (#45428) 2026-10-20 14:57:01 -07:00
Yuwei An f0b6590722 [Fix] DSpark: Disable simulated acceptance shortcuts by default (#45433) 2026-10-20 14:10:59 -07:00
YAMY 7d9bc10493 [Perf] [NVIDIA] DSpark: Enable the simulated-acceptance fast paths on CUDA (#45421) 2026-10-20 13:36:33 -07:00
cc7ae372ac [Refactor] radix-cache: Route pending-backup split bookkeeping through one component action (#45242)
Co-authored-by: metamergebot <metamergebot@users.noreply.github.com>
Co-authored-by: Jialin <Jialin@users.noreply.github.com>
2026-10-20 13:24:49 -07:00
0b536ca112 [Fix] radix-cache: Separate logical and reclaimable value lengths in tree components (#45238)
Co-authored-by: metamergebot <metamergebot@users.noreply.github.com>
Co-authored-by: Jialin <Jialin@users.noreply.github.com>
2026-10-20 13:24:20 -07:00
metamergebotandjoshuadeng 468e26f253 [Perf] detokenizer: Add SGLANG_BATCH_TOKENIZER_OUTPUTS to opt out of grouped return sends (#45280)
Co-authored-by: joshuadeng <joshuadeng@users.noreply.github.com>
2026-10-20 13:24:14 -07:00
bab4a808d8 [Perf] unified-memory: Allocate free pages from each sub-pool's own end first (#45222)
Co-authored-by: metamergebot <metamergebot@users.noreply.github.com>
Co-authored-by: ZYHowell <ZYHowell@users.noreply.github.com>
2026-10-20 13:24:09 -07:00
0d7638b035 [Fix] scheduler: Exclude accepted storage-tier hits from the waiting prefill load (#45158)
Co-authored-by: metamergebot <metamergebot@users.noreply.github.com>
Co-authored-by: Jialin <Jialin@users.noreply.github.com>
2026-10-20 13:24:04 -07:00
dc62c3effd [Feat] radix-tree: Report each path node's child count in collect_path_info (#45241)
Co-authored-by: metamergebot <metamergebot@users.noreply.github.com>
Co-authored-by: Jialin <Jialin@users.noreply.github.com>
2026-10-20 13:23:59 -07:00
metamergebotanddaniel-salib 194876bea0 [Fix] srt: Admit only allowlisted peers on TCP-bound ZMQ IPC sockets (#45250)
Co-authored-by: daniel-salib <daniel-salib@users.noreply.github.com>
2026-10-20 13:23:53 -07:00
Yuwei An afe3e1df31 [Fix] DSpark: Restore weight-loading CPU tests (#45427) 2026-10-20 13:13:39 -07:00
Jimmy Shong d69966f9c0 [Chore] docker: Add Beam CUDA image (1/8) (#45253) 2026-10-20 13:00:46 -07:00
Baizhou Zhang cfec1b91bb [Fix] muse_spark: support unified KV memory with TRT-LLM (#45356) 2026-10-20 12:54:53 -07:00
Aarne Talman 29274b41c6 [Fix] [AMD] activation: Run ReLU2 natively on ROCm (#44198)
Signed-off-by: Aarne Talman <aarne.talman@amd.com>
2026-10-20 12:05:14 -07:00
Thomas Wang 82ac83a8d0 [Perf] [AMD] dsv41: Skip unneeded layer 20 computation under decoder-bounded-replay (#45339) 2026-10-20 11:49:05 -07:00
Michaelandjacky.cheng f82465718b [Test] [AMD] diffusion: Skip MOSS-Image mismatched-operand test on ROCm (#44596)
Co-authored-by: jacky.cheng <yichiche@amd.com>
2026-10-20 11:37:32 -07:00
Khoa Pham c2e5742464 [Perf] deepseek-v4: DSpark draft stages take the TF32 mHC statistics path from 32 rows on SM100 (#45260) 2026-10-20 10:12:06 -07:00
sogalin_codegen 89860fbd2e [Fix] [AMD] gpt-oss: support EP for gfx1250 W4A8 MXFP4 MoE (#45365) 2026-10-20 10:08:39 -07:00
Michael 7591118794 [Fix] [AMD] grok: accept allow_patterns_overrides in presharded _prepare_weights patch (#45217) 2026-10-20 16:12:06 +00:00
eb4b167d27 [Feat] KVShard: PD transfer and e2e test (#38473)
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: Zhangheng <hzh0425@apache.org>
2026-10-20 22:48:43 +08:00
7656071309 [Feat] platform: Add initial T-HEAD SAIL platform support [2/N] (#43826)
Co-authored-by: Zhangheng <hzh0425@apache.org>
Co-authored-by: yzshi <163985193+Yunzshi310@users.noreply.github.com>
2026-10-20 22:19:07 +08:00
pllimax e262b44eef [CI] [NPU] tests: Replace DeepSeek-V2-Lite-W8A8 cases with Moonlight (#43063) 2026-10-20 21:35:49 +08:00
Shangming Cai 952c63b1bb [CI] sgl-router: Gate on live run-ci label so /tag-and-rerun-ci can start the tiers (#45383) 2026-10-20 21:33:42 +08:00
Jensenandronnie_zheng 10d5293546 [Perf] [NPU] glm5.2: Support dynamic INT8 MLAProlog for attention (#43469)
Co-authored-by: ronnie_zheng <zl19940307@163.com>
2026-10-20 16:33:08 +03:00
Jensen 3bf4cbc6cf [Perf] [NPU] mem_cache: Increase GLM-5.2-W4A8C8 KV cache capacity with compact DSA indexer allocation (#44436) 2026-10-20 16:30:01 +03:00
gjsheu aa91108f64 [Fix] [NPU] glm5.2: support pcp v2 with interleave & zigzag (#40165) 2026-10-20 21:26:30 +08:00
gjsheu ee2f95a522 [Fix] [NPU] glm5.2-mtp: L2 cache adaptation (#41642) 2026-10-20 21:25:37 +08:00
df13d5dce3 [Feat] sgl-router: Route prefill buckets by uncached tokens (#45103)
Co-authored-by: Xiufei Li <294029388+lixiufei-leo@users.noreply.github.com>
Co-authored-by: Shangming Cai <csmthu@gmail.com>
2026-10-20 20:56:14 +08:00
Po-Han Huang (NVIDIA)andClaude Opus 5.5 eae134d61f [Test] kernels: Import silu_and_mul_contig_post_quant from its new home (#45368)
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-20 19:48:59 +08:00
Kurkur 07eb671b6a [Perf] [NPU] attn: add not needs_cpu_seq_lens for dsa modules (#38461) 2026-10-20 19:19:34 +08:00
Shijin ZhangandXinyuan Tong ca473ef3e4 [Feat] glm-5.3: Render GLM-5.3 and GLM-5.3-Flash chat prompts with a Python chat template (#45154)
Signed-off-by: Shijin Zhang <75300765+Dovis01@users.noreply.github.com>
Co-authored-by: Xinyuan Tong <xinyuantong.cs@gmail.com>
2026-10-20 19:11:40 +08:00
Ho-Ren (Jack) Chuang 800e60e345 [Fix] moe: Publish a server config in the CUTLASS runner dispatch-format and placement tests (#45376) 2026-10-20 19:05:17 +08:00
Yuxuan Zhou 306ccf6f57 [Perf] spec: Let the FlashAttention backend capture the draft-extend CUDA graph (#40762) 2026-10-20 03:05:42 -07:00
Sihan Chen 9db58ab547 [Fix] eplb: Fix Qwen3 MoE output corruption after EPLB rebalance (#39580)
Signed-off-by: Spycsh <sihan.chen@intel.com>
2026-10-20 03:00:39 -07:00
a70b1b73d3 [Fix] tracing: Fix OTel tracing context lost in multi-tokenizer router processes (#38211)
Co-authored-by: Ma <T9FOPWP@cn.vwgroup.com>
Co-authored-by: Shangming Cai <csmthu@gmail.com>
2026-10-20 17:57:16 +08:00
16eb87c06f [Fix] openai: distinguish s3 served-model-name from LoRA adapter syntax (#44450)
Signed-off-by: Muqi Li <muqi1029@gmail.com>
Signed-off-by: Xinyuan Tong <xinyuantong.cs@gmail.com>
Co-authored-by: Xinyuan Tong <xinyuantong.cs@gmail.com>
Co-authored-by: Xinyuan Tong <115166877+JustinTong0323@users.noreply.github.com>
2026-10-20 17:11:06 +08:00
Xiaoyu ZhangandClaude Opus 5.5 2229345dff [Doc] kda_kernels: Point the SM121 QSA kernel provenance at its SGLang PR (#45359)
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-20 16:50:58 +08:00
Ho-Ren (Jack) Chuang f8fd41c36e [Perf] moe: Extend FlashInfer CUTLASS MoE autotune coverage to prefill chunks (#43363) 2026-10-20 01:41:16 -07:00
Ma MingfeiandCopilot 06314dc2e8 [CI] [Intel] tests: Remove CPU monitor and crash-dump tests from Xeon CI (#45281)
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
2026-10-20 16:40:13 +08:00
Xiaoyu ZhangandClaude Opus 5.5 0788f1fdd0 [Refactor] kernels: Register agent-generated kernels under their KDA / Cake backend (#45277)
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-20 16:35:11 +08:00
Xinyuan Tong fdd89d182a [Fix] moe: add uses_expert_location to MegaMoE test fixtures (#45350)
Signed-off-by: Xinyuan Tong <xinyuantong.cs@gmail.com>
2026-10-20 16:20:46 +08:00
Mohammad Miadh AngkadandMohammad Angkad 78185f4b3f [Fix] [NVIDIA] modelopt-nvfp4: Read w13 global scales in the TRT-LLM [up, gate] order (#44858)
Co-authored-by: Mohammad Angkad <mohammad.angkad@radixark.ai>
2026-10-20 07:38:45 +00:00
XXXXRT666andAlex Nails 1739d962a4 [Refactor] [Apple] MLX: Use Native Extension Builder for Metal Kernels (#44367)
Co-authored-by: Alex Nails <alex.nails@radixark.ai>
2026-10-20 00:35:29 -07:00
Shuwen Wang 15e8e687db [Chore] environ: Remove retired unified radix tree environment variable (#45090) 2026-10-20 07:02:43 +00:00
Po-Han Huang (NVIDIA)andClaude Opus 5.5 9e23107772 [Doc] cookbook: Refresh SemiAnalysis AgentX launches with five open InferenceX PRs (#45311)
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-20 14:48:48 +08:00
1ec902e00f [Fix] cache: Support DeepSeek-V4.1 external linker pools (#44623)
Co-authored-by: Zhangheng <hzh0425@apache.org>
Co-authored-by: Shuwen Wang <47200617+alphabetc1@users.noreply.github.com>
2026-10-19 23:40:25 -07:00
Michael 407bfc94f9 [Test] [AMD] diffusion: Make multimodal_gen unit tests ROCm-aware (#44599) 2026-10-20 14:24:37 +08:00
Xinyuan Tong 4b83f215f4 [Feat] entrypoints: add a chat encoding adapter registry (#45300) 2026-10-20 14:10:01 +08:00
lpdinkandZhangheng 4a651146de [Feat] hicache: Notify storage backends when the initial host-pool registration burst completes (#44030)
Co-authored-by: Zhangheng <hzh0425@apache.org>
2026-10-20 14:09:13 +08:00
Michael 36978d6c4d [Test] [AMD] diffusion: Skip the Kandinsky6 FA dispatch test on ROCm (#44421) 2026-10-20 13:54:23 +08:00
Duyi-Wang 50fa3727eb [Doc] [AMD] cookbook: Fix DeepSeek-V4 PD image and c1 capacity (#45297) 2026-10-19 22:45:50 -07:00
Liangsheng Yin 472f0d540b [Fix] deepseek-v4.1: Keep the padded query buffer inside decoder tail graphs (#45301) 2026-10-19 22:41:53 -07:00
Khoa PhamandClaude Fable 5.1 476fe86a12 [Perf] deepseek-v4.1: Score a request's verify rows as one request in the full-scan indexer (#45251)
Co-authored-by: Claude Fable 5.1 <noreply@anthropic.com>
2026-10-19 22:25:04 -07:00
35c873cff4 [Fix] observability: Log scheduler stats once per attention CP group (#45319)
Co-authored-by: Hun-ho Kim <hunho.kim@samsung.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-19 22:22:31 -07:00
weireweireandweireweire c7687a0e7c [Fix] deepseek-v4: Keep the draft MoE out of the target's expert placement (#45299)
Co-authored-by: weireweire <20922698+weireweire@users.noreply.github.com>
2026-10-19 22:20:26 -07:00
602c92ccd8 [Perf] DeepSeek-V4.1: Raise the SM100 swap-AB decode cap to 40 rows (#44632)
Co-authored-by: weireweire <20922698+weireweire@users.noreply.github.com>
Co-authored-by: nbommi19 <227267412+nbommi19@users.noreply.github.com>
2026-10-19 22:20:10 -07:00
553836c164 [Feat] simulator: Add baseline collection and ML latency predictor training tools (#44299)
Co-authored-by: shijunyu.sjy <shijunyu.sjy@alibaba-inc.com>
Co-authored-by: baoyunong.byn <baoyunong.byn@alibaba-inc.com>
Co-authored-by: wangyi.wy <aaronwangyi0201@gmail.com>
2026-10-20 13:02:33 +08:00
parkeychenandparkeychen 43a896933e [Fix] deepseek-v4: Refresh attn_sink after loading weights (#41630)
Co-authored-by: parkeychen <parkeychen@tencent.com>
2026-10-19 21:59:32 -07:00
ashwini rathiandClaude Opus 5.5 d1c3346f74 [Fix] [Intel] spec: Fix XPU nightly dflash stop_regex and DSpark capture check (#45033)
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-20 12:27:36 +08:00
72d886efb8 [Perf] DeepSeek-V4.1: Run the small-batch decode fusions at TP2 (#43838)
Co-authored-by: weireweire <20922698+weireweire@users.noreply.github.com>
Co-authored-by: nbommi19 <227267412+nbommi19@users.noreply.github.com>
Co-authored-by: Po-Han Huang <pohanh@nvidia.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: Po-Han Huang (NVIDIA) <53919306+nvpohanh@users.noreply.github.com>
2026-10-19 21:17:05 -07:00
Mohammad Miadh AngkadandMohammad Angkad fdd53418a0 [CI] [NVIDIA] gb300: Drop Kimi-K2.5 and Qwen3.5 nightly tests (#45308)
Co-authored-by: Mohammad Angkad <mohammad.angkad@radixark.ai>
2026-10-19 21:13:00 -07:00
Shuwen Wang c53b98d1fe [Fix] test-utils: Allow expected server exits in canary tests (#45172) 2026-10-19 21:07:03 -07:00
Mohammad Miadh AngkadandMohammad Angkad 5b683d6b72 [Fix] test: Add logical_forward_mode to the KDA prefill-graph stand-in batch (#45218)
Co-authored-by: Mohammad Angkad <mohammad.angkad@radixark.ai>
2026-10-19 21:06:10 -07:00
Mingyi 4a0c6f2fe0 docs: promote SGLang Summit 2026 in README (#45271) 2026-10-20 03:36:56 +00:00
wxzhoucsandXiaoyu Zhang e3205080ce [Fix] qwen3.8: Fix batch-invariant kernels for Qwen3.8 Flash-Next (#42087)
Co-authored-by: Xiaoyu Zhang <1182563586@qq.com>
2026-10-20 11:32:40 +08:00
1140a8dfd9 [Fix] [AMD] aiter: Fix garbage decode output for non-sliding-window models on the vectorized_5d KV layout (#43335)
Co-authored-by: Cursor <cursoragent@cursor.com>
Co-authored-by: Xiaoyu Zhang <1182563586@qq.com>
2026-10-20 11:31:20 +08:00
Cao EandMa Mingfei 3c07ff2654 [Perf] [XPU] fp8: Optimize FP8 scale layouts (#40052)
Co-authored-by: Ma Mingfei <mingfei.ma@intel.com>
2026-10-20 11:24:16 +08:00
huangtingwei 3eb4bd598f [Perf] hicache: Unify PP prefetch hit queries and buffer allocation (#43344) 2026-10-20 11:05:00 +08:00
Trevor Morris ec1fdc22a9 [Fix] [NVIDIA] docker: Fix CUDA 13.4 container by updating torch nightly (#45252) 2026-10-20 10:52:42 +08:00
kpjeeja 8683f8ad4a [Feat] [Intel] hicache: Enable L1<->L2 support on XPU (#32503) 2026-10-20 10:42:21 +08:00
Liangsheng Yin 5e630fa571 [Perf] deepseek-v4.1: Drop host syncs from encoder-replay folding and tail metadata (#45275) 2026-10-19 19:39:39 -07:00
Mohammad Miadh AngkadandMohammad Angkad af856717c2 [Fix] mamba: Skip zero-width conv pools in the MTP verify scatter (#45212)
Co-authored-by: Mohammad Angkad <mohammad.angkad@radixark.ai>
Co-authored-by: Mohammad Angkad <mohammad.angkad@radixark.ai>
2026-10-19 19:23:17 -07:00
Vladimir Serovandronnie_zheng 55b2aef7ad [Fix] [NPU] LoRA: Update for follow-up with changes in sgl-kernel-npu (#40908)
Co-authored-by: ronnie_zheng <zl19940307@163.com>
2026-10-20 05:16:05 +03:00
wutao 2c00b5d129 [Test] constrained: Cover JSON whitespace limits (#43981) 2026-10-20 10:12:22 +08:00
Gabriel Wuandnumber-eleven11 67d9ee156c [Fix] [SM120] quantization: Requantize block-FP8 MoE weights to UE8M0 for DeepGEMM (#43556)
Signed-off-by: Zihua Wu <13583761+lucifer1004@users.noreply.github.com>
Co-authored-by: number-eleven11 <1635219023@qq.com>
2026-10-19 19:06:21 -07:00
Liangsheng Yin ce329c85fa [Perf] deepseek-v4.1: Skip the request-window history check when invariant checks are off (#45276) 2026-10-19 18:38:44 -07:00
79ed13fc2c [Fix] glm47: Disable tool grammar for tool_choice="none" (#41233)
Co-authored-by: Po-Han Huang (NVIDIA) <53919306+nvpohanh@users.noreply.github.com>
Co-authored-by: Xinyuan Tong <115166877+JustinTong0323@users.noreply.github.com>
Co-authored-by: Po-Han Huang <pohanh@nvidia.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-20 09:36:06 +08:00
Lee Nau 1897575ce1 [Fix] runtime: Version FlashInfer caches for default-tactic tuning. (#44514) 2026-10-19 18:33:49 -07:00
Anihcandsunguozhong ebc71e650e [Perf] [NPU] mem_cache: Add optional ScatterPa KV cache writes (#40535)
Co-authored-by: sunguozhong <sun_gz2772@163.com>
2026-10-20 09:33:34 +08:00
71ee2dc950 [Fix] inkling: Return 400 instead of 500 for unfetchable/undecodable multimodal input (#40899)
Co-authored-by: apex-mochen <265981728+apex-mochen@users.noreply.github.com>
Co-authored-by: Po-Han Huang (NVIDIA) <53919306+nvpohanh@users.noreply.github.com>
Co-authored-by: apex-mochen <2756823972@qq.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: Po-Han Huang <pohanh@nvidia.com>
2026-10-20 09:18:18 +08:00
Xiaoyu ZhangandClaude Opus 5.5 693d333306 [Refactor] kernels: Regroup operators by computation and dissolve ops/muse_spark (#45112)
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-20 08:47:57 +08:00
Liangsheng Yin ab580b2e06 [CI] radix-cache: Split unified radix cache unit tests into topic files and a shared kit (#45236) 2026-10-19 17:00:42 -07:00
Cheng Wan 6b90e5716d [Refactor] moe: Scope shared-expert configuration to each model (#44913) 2026-10-19 16:59:37 -07:00
Cheng Wan 964ab2b280 [Refactor] moe: Split the two meanings of num_fused_shared_experts (#44912) 2026-10-19 16:57:15 -07:00
Cheng Wan 48ea486928 [Fix] moe: Renormalize per-rank shared-slot routing over routed experts only (#44904) 2026-10-19 16:56:21 -07:00
Cheng Wan 3a059de94c [Fix] moe: Count routed_scaling_factor once when top-k folds it (#44901) 2026-10-19 16:54:27 -07:00
Qiaolin Yuandzijiexia 28448a86d4 [Doc] cookbook: Default Qwen3.8-Flash-Next datacenter recipes to FP8 KV cache and QSA indexer (#45245)
Co-authored-by: zijiexia <37504505+zijiexia@users.noreply.github.com>
2026-10-19 15:57:22 -07:00
Liangsheng Yin 1238edd954 [Perf] deepseek-v4.1: Reset request-window tags without a host sync (#45247) 2026-10-19 15:45:17 -07:00
Mingyi 9a949403f1 [Doc] tpu: Introduce SGLang-TorchTPU in README and TPU guide (#45246) 2026-10-19 15:34:47 -07:00
ab258b6453 [Fix] symm-mem: Name the window registration in the scheduler watchdog dump (#44681)
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
Co-authored-by: hnyls2002 <lsyincs@gmail.com>
2026-10-19 15:29:35 -07:00
Gabriel Wu 05d7f7a166 [Fix] attention: Honor selected FlashAttention implementation (#33007) 2026-10-19 15:27:52 -07:00
Jiuqiang Tang 49766a74f9 [Fix] attn_residual: int64 program ids in the score/combine kernels (int32 overflow at 37,451 tokens) (#43438) 2026-10-19 15:12:12 -07:00
0z5aandRichard Wang 80ea5ee638 [Fix] dLLM: Preserve prompt tails and literal mask tokens across prefix reuse (#39592)
Signed-off-by: 0z5a <dezhen.lu@student.uni-tuebingen.de>
Co-authored-by: Richard Wang <wangrichard08@gmail.com>
2026-10-19 14:48:37 -07:00
Liangsheng Yin fdc2dd0d05 [Perf] dspark: Make the prefill SWA valid-from update sync-free (#45234) 2026-10-19 14:44:01 -07:00
Liu Keandhnyls2002 81ac15391c [Fix] metrics: record HTTP request metrics in tokenizer workers (#44663)
Co-authored-by: hnyls2002 <lsyincs@gmail.com>
2026-10-19 14:36:16 -07:00
Mohammad Miadh AngkadandMohammad Angkad 7656ea0e74 [CI] python: Move the remaining CUDA build paths off Python 3.10 (#44393)
Co-authored-by: Mohammad Angkad <mohammad.angkad@radixark.ai>
2026-10-19 14:29:00 -07:00
Liangsheng Yin b20cb3665d [CI] radix-cache: Drop redundant unified radix cache unit tests (#45232) 2026-10-19 14:20:32 -07:00
Baizhou Zhang 642a671c30 [CI] tests: Re-enable muse spark B300 PD coverage (#44975) 2026-10-19 13:40:02 -07:00
Liangsheng Yin 8b1d1e0d74 [Perf] deepseek-v4.1: Check request-window SWA history without per-layer host syncs (#45223) 2026-10-19 13:34:52 -07:00
Rain Jiang aaa9388c91 [Fix] kda: Keep FlashInfer prefill off its grouped kernel's bf16/flat-dt_bias contract (#45018) 2026-10-19 13:16:54 -07:00
Liangsheng Yin f1a2ee763d [Perf] deepseek-v4.1: Run decoder replay tail graphs in prefill-graph steps and under DP attention (#45107) 2026-10-19 13:02:07 -07:00
Qiaolin Yu 5385866108 [Fix] moe: gated NVFP4 MoE intermediate alignment for FlashInfer TRTLLM (#44138) 2026-10-19 12:52:36 -07:00
Qiaolin Yu 9ab74b8517 [Fix] Qwen3.8next: Stop retaining per-size PLE outputs in breakable prefill CUDA graphs (#44150) 2026-10-19 12:41:39 -07:00
0ce41c5a06 [Fix] PD: retraction restores racing the in-flight decode forward (#42702)
Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Co-authored-by: Zhiqiang Xie <xiezhq@stanford.edu>
2026-10-19 11:55:04 -07:00
YAMY 722669e606 [Feat] [NVIDIA] megakernels: Add --enable-nvidia-megakernel and a layer for NVIDIA megakernel plugins (#44945) 2026-10-19 11:54:44 -07:00
AMD-yanfeiwang 678913a0f0 [Feat] sgl-model-gateway: Add --decode-dispatch-limit to pick PD decode workers near KV arrival (#44689) 2026-10-20 02:16:21 +08:00
jacky.cheng 80b86dbecb [Fix] [AMD] diffusion: Accept residual and skip_bias in the ROCm Conv3d-to-Conv2d VAE patch (#44672) 2026-10-19 11:07:37 -07:00
jacky.cheng 5ed306bd69 [Fix] [AMD] diffusion: Keep gfx1250 FlyDSL BF16 attention out of torch.compile (#44671) 2026-10-19 10:36:28 -07:00
lixiufei-leoandXiufei Li 4ef9765e76 [Feat] sgl-router: Add --pd-stream-after-prefill (#45102)
Co-authored-by: Xiufei Li <294029388+lixiufei-leo@users.noreply.github.com>
2026-10-20 01:13:25 +08:00
chenyang08056032 716c9151f2 [CI] [NPU] npu-coverage: run only coverage-recommended tests on PRs (precise test selection) (#44019) 2026-10-20 00:57:27 +08:00
Jan SchalkampandXiaoyu Zhang 8e7601ae4a [Feat] paged-experts: Serve MoE models whose weights do not fit VRAM (#42557)
Co-authored-by: Xiaoyu Zhang <1182563586@qq.com>
2026-10-19 23:10:04 +08:00
Monishver 806e7423af [Perf] deepseek_v4: size the masked silu-mul-quant grid to the GPU, not the padded capacity (#44618)
Signed-off-by: Monishver Chandrasekaran <monishverchandrasekaran@gmail.com>
2026-10-19 23:07:12 +08:00
Avaya Aggarwal c04f9a61f0 [Fix] diffusion: Fix VMoBA chunk selection bugs (#44286) 2026-10-19 23:04:26 +08:00
Siliang ChenandCursor 7c4cc5a51c [Perf] [AMD] attention: Spread uneven Triton decode batches across XCDs (#41229)
Co-authored-by: Cursor <cursoragent@cursor.com>
2026-10-19 23:03:27 +08:00
Sam Shleifer 96ba108072 [Fix] dsa-eagle-topk: Copy token keys and scales when relocating index cache (#44564) 2026-10-19 22:54:08 +08:00
Sam Shleifer 417606527d [Fix] nvfp4: Own streamed tensors while waiting for paired weights (#44500) 2026-10-19 22:41:53 +08:00
pllimax 8c4680cb65 [CI] npu: Move llada2-mini FDFO test to full test (#45153) 2026-10-19 22:29:30 +08:00
lixiufei-leoandXiufei Li 62cee1b15c [Fix] sgl-router: Keep the kv-indexer connection through bursts of cancelled prefix queries (#45100)
Co-authored-by: Xiufei Li <294029388+lixiufei-leo@users.noreply.github.com>
2026-10-19 21:41:03 +08:00
zhaozx-cn ed0b8f433f [Fix] [NPU] kimi-k3: skip fused gmm1 for situ activation in Ascend runner (#43837) 2026-10-19 19:30:41 +08:00
Shangming Cai 0025f93944 [Fix] sgl-router: Refuse prefill workers that cannot compute before bootstrap under --decode-dispatch-limit (#45120) 2026-10-19 18:24:17 +08:00
Quanli Liand全力 e45cbce599 [Fix] diffusion: Prevent int32 overflow in Qwen-Image-2.1-Turbo VAE RMSNorm (#45118)
Co-authored-by: 全力 <liquanli.lql@antgroup.com>
2026-10-19 18:01:02 +08:00
Niko Maandinkcherry a90e9c5dd8 [Fix] [AMD] mhc: Bound the Triton fused post+pre buffer cache (#44443)
Co-authored-by: inkcherry <mingzhi.liu@amd.com>
2026-10-19 01:37:50 -07:00
8111841103 [Perf] deepseek-v4.1: Replay the breakable prefill CUDA graph at any context length (#43733)
Co-authored-by: Oasis-Git <ayw.sirius19@gmail.com>
Co-authored-by: hnyls2002 <lsyincs@gmail.com>
2026-10-19 01:29:14 -07:00
33dd8f6a2e [Fix] hybrid-SSM: DP attention failures at low concurrency (#34535)
Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
Co-authored-by: Jimmy Shong <69131491+Jiminator@users.noreply.github.com>
Co-authored-by: Po-Han Huang (NVIDIA) <53919306+nvpohanh@users.noreply.github.com>
2026-10-19 16:24:41 +08:00
Thomas Wang 5d9c956f8e [Doc] [AMD] dsv4: Update amd cookbook 1016 (#45070) 2026-10-19 16:11:11 +08:00
Po-Han Huang (NVIDIA)andClaude Opus 5.5 af9bb2a32d [Doc] cookbook: Refresh SemiAnalysis AgentX launches with twelve open InferenceX PRs (#45111)
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-19 15:24:08 +08:00
Bingxu Chen 3d588779e1 [CI] [AMD] test: Run UMBP HiCache E2E through standalone DRAM server (#40674) 2026-10-19 00:23:52 -07:00
AMD-yanfeiwang f86a04a331 [Feat] sgl-router: Add --decode-dispatch-limit to pick decode DP ranks near KV arrival (#44402) 2026-10-19 15:11:40 +08:00
a5117beb11 [Perf] mhc: Parallelize small-batch post across hidden tiles (#44906)
Co-authored-by: leonardHONG <178642616+leonardHONG@users.noreply.github.com>
Co-authored-by: Xiaoyu Zhang <1182563586@qq.com>
2026-10-19 15:07:13 +08:00
ashwini rathiandClaude Opus 5.5 b6d3b7e163 [CI] [Intel] xpu: Disable L0 copy offload to fix first-H2D-copy hang (#45013)
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-19 14:48:45 +08:00
97857f9f96 [Doc] cookbook: GB300 full-set accuracy and tool follow-up output for Ling-3.1-flash (#44307)
Co-authored-by: andyluo7 <andy.luo@amd.com>
Co-authored-by: zijiexia <37504505+zijiexia@users.noreply.github.com>
2026-10-18 23:29:09 -07:00
desmond-intel dc81f17b90 [Fix] [XPU] gemma3: Enable bidirectional image attention (#44225) 2026-10-19 14:08:24 +08:00
Liangsheng Yin a3330237db [Perf] deepseek-v4.1: Run decoder replay tail layers from CUDA graphs on eager prefill steps (#45059) 2026-10-18 23:07:50 -07:00
Trevor Morris 04c8c9703f [Fix] docker: Install numactl in runtime images - needed for SGLANG_NUMA_BIND_V2 (#43883) 2026-10-18 22:55:58 -07:00
7c9a78efbe [Feat] nemotron: Port Nemotron 3.5 Super serving changes (#44518)
Co-authored-by: BBuf <1182563586@qq.com>
Co-authored-by: elvischenv <219235043+elvischenv@users.noreply.github.com>
Co-authored-by: Jimmy Shong <jimmy.shong@mac.attlocal.net>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-19 13:27:22 +08:00
Ziru f726461287 [Feat] [CPU] model: Support LTX and MOVA models on CPU (#36632) 2026-10-19 13:25:13 +08:00
ZiruandNiu Ziru 9a93a17c08 [Feat] [CPU] model: Support Muse models on CPU (#44624)
Co-authored-by: Niu Ziru <niuziru@gnr799217.jf.intel.com>
2026-10-19 13:12:32 +08:00
MickandMick Qian 2b75048e2a [CI] diffusion: tighten H100 baselines and enable B200 performance checks (#44996)
Co-authored-by: Mick Qian <mickqian@users.noreply.github.com>
2026-10-19 12:54:54 +08:00
Avaya Aggarwal 58ad0748a6 [Fix] diffusion: Unmerge LoRA layers the next adapter does not cover (#44770) 2026-10-19 12:33:40 +08:00
Xinyuan Tongand文贤 266550048a [Fix] bailing: MoE gate dtype follows config.router_dtype (#44715)
Co-authored-by: 文贤 <luoyuan.luo@antgroup.com>
2026-10-19 12:21:29 +08:00
Liangsheng Yin 82c9a8b037 [Perf] deepseek-v4.1: Default PD to MegaAttention, with the KV layout in the PD handshake (#45075) 2026-10-18 21:16:45 -07:00
Yuwei An 834bfe4ef0 [Fix] attention: Opt FlashMLA scheduler into dynamic shared memory (#44580) 2026-10-18 20:31:39 -07:00
Duyi-Wang 8d3eee46a7 [Perf] [AMD] deepseek-v4: Skip top-k refinement for proven positive-zero ties (#44927) 2026-10-19 11:27:47 +08:00
06342acad2 [Perf] deepseek-v4.1: Stage late-layer DP row counts through pinned memory (#45062)
Co-authored-by: weireweire <20922698+weireweire@users.noreply.github.com>
Co-authored-by: kpham-sgl <khoa.pham@radixark.ai>
Co-authored-by: Claude Fable 5.1 <noreply@anthropic.com>
2026-10-18 20:19:26 -07:00
f02458fff1 [Fix] [CPU] memory: Honor cgroup memory.max when sizing CPU memory (#40066)
Co-authored-by: Zhang, Mingxu <mingxu.zhang@intel.com>
Co-authored-by: MingxuZh <109504044+MingxuZh@users.noreply.github.com>
2026-10-19 11:11:44 +08:00
Alexandr 617c8b67b1 [Fix] Z-Image: Synchronize device after warmup request (#44306) 2026-10-19 05:36:16 +03:00
Ankur SinghandClaude Opus 5.5 7b1710324c [Doc] cookbook: Refresh SemiAnalysis AgentX launches for Qwen3.8-Flash-Next H200 (v0.5.22, TP2 points) (#45039)
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>

Force merge to unblock SA AgentX submission
2026-10-19 10:31:51 +08:00
Kotthagattu Meher Sai 7f96dc6874 [Perf] [XPU] server_args: Enable fused QK-norm + RoPE by default (#43345) 2026-10-19 10:19:52 +08:00
642f43ce59 [Fix] diffusion: size the in-context slot from both input grids for wan-animate-2(#44090)
Co-authored-by: Ankit Dhall <8938083+ankitdhall@users.noreply.github.com>
Co-authored-by: Mick <mickjagger19@icloud.com>
2026-10-19 10:11:58 +08:00
Khoa Pham a7fd2e1158 [Feat] dsv4.1: Return prompt logprobs under encoder SWA bounded replay (#45050) 2026-10-18 19:08:20 -07:00
Wuhen DuanandEven Zhou 69076b9f9e [CI] [NPU] tests: Reduce HiCache sizes and enable expandable pinned memory in focused tests (#41675)
Co-authored-by: Even Zhou <even.y.zhou@outlook.com>
2026-10-19 10:06:35 +08:00
fea48754de [Feat] rust/grpc: Add ListModels, Detokenize, and Abort to api.v1 (#42118)
Signed-off-by: jain-ria <riajain@NVIDIA.com>
Co-authored-by: ishandhanani <82981111+ishandhanani@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: Ishan Dhanani <ishandhanani@gmail.com>
2026-10-19 09:53:56 +08:00
Mohammad Miadh AngkadandMohammad Angkad 377cd867fd [Perf] dsa: Route the k-pool indexer top-k through the top-k v2 kernel (#43775)
Co-authored-by: Mohammad Angkad <mohammad.angkad@radixark.ai>
2026-10-19 09:49:39 +08:00
Yuwei An df078cc729 [Fix] DSv4.1: Support CUDA image requests with encoder bounded replay and CI Test (#45048) 2026-10-18 18:47:17 -07:00
Emre AlbayrakandXiaoyu Zhang 10ec53ac2f [Perf] [NVIDIA] fp8_fa_sm120: Use the block-scaled MMA encoding (#43390)
Co-authored-by: Xiaoyu Zhang <1182563586@qq.com>
2026-10-19 09:36:21 +08:00
Jimmy Shong 0dc6de2e74 [Doc] nemotron-3.5-super: update H200 and Spark wizard recipes (#44587) 2026-10-18 18:22:05 -07:00
A&W cff88d6eac [Fix] diffusion: Invalidate merged LoRA cache after adapter rewrites (#44204) 2026-10-19 01:08:57 +00:00
Liangsheng Yin 299da60131 [CI] test-utils: Abort the test run when its server dies inside a CustomTestCase method (#45053) 2026-10-18 18:00:06 -07:00
Xiaoyu ZhangandClaude Opus 5.5 329072d99a [Refactor] kernels/diffusion/tests: Prune dead code and fix tests that never ran (#44690)
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-19 08:55:47 +08:00
Liangsheng Yin 59fec011dd [Fix] deepseek-v4.1: Reject encoder replay on a chunk too small for a hit or on other models (#45051) 2026-10-18 17:55:35 -07:00
Liangsheng Yin 3c560089e1 [Refactor] scheduler: Bill encoder replay rows in _update_prefill_budget (#45047) 2026-10-18 17:17:38 -07:00
Liangsheng Yin 1637a4e305 [Fix] scheduler: Clear stale encoder_swa_reset in convert_decode_to_extend (#45042) 2026-10-18 17:09:19 -07:00
Liangsheng Yin 6708426728 [Fix] spec: Remove --speculative-skip-dp-mlp-sync (#45043) 2026-10-18 16:41:21 -07:00
Liangsheng Yin 3f13f5d966 [Fix] scheduler: Drop prefill embedding overrides in prepare_for_decode (#45041) 2026-10-18 15:54:20 -07:00
Liangsheng Yin 1461152cb0 [Fix] deepseek-v4.1: Leave room for encoder replay rows when sizing a prefill chunk (#45040) 2026-10-18 15:43:08 -07:00
Liangsheng Yin fe89a7c890 [Fix] deepseek-v4.1: Start the PD decode draft SWA window at the prefill draft valid-from (#45037) 2026-10-18 15:06:20 -07:00
Liangsheng Yin da414adda7 [Fix] engram: Clamp multimodal pad ids in host-built history rows (#45038) 2026-10-18 12:24:05 -07:00
Liangsheng Yin 0f4aedfecc [Perf] deepseek-v4.1: Default decoder SWA bounded replay on for both PD roles on CUDA (#45035) 2026-10-18 11:53:35 -07:00
Ke Bao 4ec19ca27f [Refactor] memory-pool: Build KV host pools from buffer info (#40917) 2026-10-18 23:21:16 +08:00
Ke Bao bb1d4cddd9 [Refactor] hicache: Route Mamba host transfers from pool info (#44968) 2026-10-18 23:02:46 +08:00
Xinyuan Tong d4e3c163c9 [CI] radix-cache: Register the Inkling cache linker KL test to extra-b (#45004) 2026-10-18 22:23:34 +08:00
e9b22a0818 [Perf] [AMD] monokernels: Let a plugin skip per-step attention metadata (#44601)
Co-authored-by: Kevin Mi <kevin.mi@radixark.ai>
Co-authored-by: kkHuang-amd <43161300+kkHuang-amd@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-18 14:04:57 +00:00
Cheng Wan 57db5089c2 [Perf] attention: Capture translated FA4 verify metadata for DSPARK (#44973) 2026-10-18 05:42:16 -07:00
Cheng Wan 2e9b2cacda [Perf] speculative: Fuse draft write-window gathering and KV translation (#44972) 2026-10-18 05:41:34 -07:00
Cheng Wan 2a83471676 [Perf] mamba: Fuse strided state updates and replay index translation (#44971) 2026-10-18 05:40:57 -07:00
Cheng Wan d318a546da [Perf] unified-memory: Reduce allocation and SWA retirement overhead (#44970) 2026-10-18 05:40:25 -07:00
Cheng Wan a5b21145e1 [Refactor] attention: Separate decode metadata from pool ownership (#44969) 2026-10-18 05:39:44 -07:00
3edb156dec [Feat] unified-memory: Fuse drafts with sliding-window and recurrent-state layers (#43326)
Co-authored-by: Caihua Li <caihua.li@bytedance.com>
Co-authored-by: Cheng Wan <cheng.wan@radixark.ai>
2026-10-18 05:39:09 -07:00
caihuali95 19cb24f1e1 [Feat] unified-memory: Support asymmetric KV rows (#43325) 2026-10-18 05:37:50 -07:00
caihuali95andCaihua Li 1d17066bad [Perf] gdn: Select FlashInfer prefill for page-major pools (#43171)
Co-authored-by: Caihua Li <caihua.li@bytedance.com>
2026-10-18 05:37:24 -07:00
caihuali95andCaihua Li ba50ea3d77 [Perf] unified-memory: Free SWA segments without a host sync (#43170)
Co-authored-by: Caihua Li <caihua.li@bytedance.com>
2026-10-18 05:36:56 -07:00
caihuali95andCaihua Li 47eeddcd18 [Fix] swa: Retain a whole window behind insertion pages (#43168)
Co-authored-by: Caihua Li <caihua.li@bytedance.com>
2026-10-18 05:36:02 -07:00
lixiufei-leoandXiufei Li fe1851c3c2 [Fix] sgl-router: Receive kv-indexer-bridge event batches outside tokio's coop budget (#44305)
Co-authored-by: Xiufei Li <294029388+lixiufei-leo@users.noreply.github.com>
2026-10-18 20:35:50 +08:00
WenhaoZhang e17fed6ea7 [Perf] diffusion: fused gate/up + SwiGLU MLP on comfy_kitchen INT8 for Qwen-Image 2.1 (#44196) 2026-10-18 20:33:31 +08:00
LindaandMick 680a2ff27e [Perf] [NVIDIA] diffusion/FLUX.3: Run dynamic LoRA through PrimsTS GEMMs (#44761)
Co-authored-by: Mick <mickjagger19@icloud.com>
2026-10-18 20:32:19 +08:00
WenhaoZhangandMick 1fd46ddd56 [Perf] diffusion: fused indexed modulation for FP32 AdaLN tables (#44180)
Co-authored-by: Mick <mickjagger19@icloud.com>
2026-10-18 20:31:36 +08:00
Shuwen Wang 8dee3f6bdb [Fix] [H200] ci: Split GLM-5.3-Flash serving recipes into separate tests (#44892) 2026-10-18 19:26:40 +08:00
Shuwen Wang f8e32706fc [CI] radix-cache: Re-enable Inkling unified cache linker KL test (#44938) 2026-10-18 19:26:22 +08:00
AMD-yanfeiwang a3abea3560 [Fix] sgl-router: Resend on a closed pooled connection; add --disable-circuit-breaker (#44075) 2026-10-18 18:23:18 +08:00
3b932cde26 [Fix] VLM: harden Kimi image cache identity across preprocessing and radix hashing (#43043)
Co-authored-by: Rahul Chalamala <22563365+rchalamala@users.noreply.github.com>
Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
Co-authored-by: Mick Qian <mickqian@users.noreply.github.com>
Co-authored-by: Liangsheng Yin <hnyls2002@gmail.com>
2026-10-18 17:39:37 +08:00
5a930cfd88 [Fix] glm4v: Clamp request video budgets to the deployment limit (#43871)
Co-authored-by: cipheraxat <cipheraxat@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-18 17:31:33 +08:00
2937 changed files with 206556 additions and 93537 deletions
+1 -1
View File
@@ -742,7 +742,7 @@ cd test && python3 run_suite.py --hw cuda --suite base-b-kernel-benchmark-test-1
- `docs/docs/developer_guide/development_jit_kernel_guide.mdx`
- `test/run_suite.py` — suite names, discovery of `test/registered/`, execution entrypoint for CI
- `python/sglang/test/ci/ci_register.py` — `register_cuda_ci` and AST registration rules
- `python/sglang/kernels/jit/utils/compile.py` — `load_jit`, `make_cpp_args`
- `python/sglang/kernels/jit/utils/compile/` — `load_jit` (`loader.py`), `make_cpp_args` (`cpp_args.py`)
- `python/sglang/kernels/jit/utils/common.py` — `cache_once`, `should_run_full_tests`, `get_ci_test_range`
- `python/sglang/kernels/jit/include/sgl_kernel/tensor.h` — `TensorMatcher`, `SymbolicSize/DType/Device`, `is_type`
- `python/sglang/kernels/jit/include/sgl_kernel/ffi.h` — `ffi::empty`, `ffi::empty_like`, `ffi::from_blob`
@@ -398,6 +398,7 @@ group: pr-test-{event_name}-{branch}-{pr_sha}-{stage}
| `/run-extra-ci` | Adds both labels + reruns **only** `PR Test Extra`, leaving baseline runs alone. Alias: `/rerun-extra-ci` |
| `/rerun-test <test-file> [<test-file> ...]` | Reruns specific test file(s) via `rerun-test.yml`. A file arg containing a glob metacharacter (`*`, `?`, `[...]`) expands against `test/registered/` and the multimodal test dir to every matching `test_*.py` (e.g. `/rerun-test test_*backend*.py` — wrap in backticks so GitHub doesn't italicize the `*`); matches are deduped, grouped by dispatch shape, and can't carry a `::test` selector. No match → single ⛔ reply, nothing dispatched. |
| `/rerun-group <group> [<group> ...]` | Expands registered test groups, then reuses `/rerun-test` |
| `/run-cpu-ci` | Dispatches `run-cpu-ci.yml` on main: only the `base-a-test-cpu` suite, at the PR head SHA, no caches saved. Same gate as `/rerun-test`. Alias: `/rerun-cpu-ci` |
Handled by `scripts/ci/utils/slash_command_handler.py` → `.github/workflows/slash-command-handler.yml`.
@@ -16,7 +16,7 @@ These are investigation hints, not a list of changes to apply.
| `No platform detected. Using base SRTPlatform with defaults.` | Platform detection | Decide whether fallback is expected on this machine or indicates a missing plugin. |
| `Unexpected error during package walk` in CUTE_DSL | Installed cutlass package, its handler, root propagation, and `CUTE_DSL_LOG_LEVEL` | Inspect the actual exception and duplicate-handler path; do not assume the current log level or hide all warnings. |
| Repeated dtype, hybrid-model, or tokenizer fallback messages | Model-config and tokenizer construction in launcher, scheduler, and detokenizer | Multiple constructors may explain repetition. Retain meaningful per-process information and distinguish these from handler duplication. |
| Repeated template detection messages | `managers/template_detection.py` and `managers/template_manager.py` | Check whether a summary already contains the same information. |
| Repeated template detection messages | `parser/template_detection.py` and `parser/template_manager.py` | Check whether a summary already contains the same information. |
| KV cache dtype logged separately from allocation | Model runner and memory pool | Check whether the allocation summary already carries dtype; retain useful allocation and SWA details. |
| CUTLASS backend disabled during graph capture | Attention backend selection | Expected fallback may be useful information; retain the reason and selected behavior. |
| `Ignore import error when loading ...` | Model, multimodal, or algorithm registries | Distinguish an unused optional model dependency from failure to load the requested model. |
+1 -1
View File
@@ -205,7 +205,7 @@ than restating.
`python/sglang/srt/function_call/function_call_parser.py` (ToolCallParserEnum) —
prose naming a near-miss id (e.g. the reasoning id where the tool id differs) is a
factual bug. `--…-parser auto` is acceptable ONLY if the template-detection rules
(`python/sglang/srt/managers/template_detection.py`) actually resolve THIS model's
(`python/sglang/srt/parser/template_detection.py`) actually resolve THIS model's
chat template to the right parser — no rule match means auto silently disables the
parser; when in doubt require explicit ids (the DSv4 page pins explicit ids).
+36 -13
View File
@@ -17,25 +17,47 @@ logical grouping and remove the old `_jit_` filename prefix.
The `sgl_kernel` wheel retains its own public API and packaging tests.
- Group by computation, not model or GPU: GEMM and GEMV in `gemm`, expert
routing in `moe`, attention index selection in `attention`, sampling in
`sampling`, normalization in `layernorm`. A quantized GEMM is still a GEMM;
quantization alone belongs in `quantization`.
`sampling`, normalization in `layernorm`.
- Classify a fused operator by its complete contract; do not split a fused
kernel into separate launches just to separate its parts. Split unrelated
public entry points that happen to share a source file. Decided rules:
- Any cross-rank collective or peer-memory access (all-reduce, all-gather,
reduce-scatter, a2a, multicast/NVLS) puts the op in `communication`,
whatever compute it fuses.
- Producer + output quantization stays with the producer (norm + quant in
`layernorm`, activation + quant in `activation`, GEMM with a quant epilogue
in `gemm`, MoE dispatch/permute + quant in `moe`). `quantization` holds only
pure format conversion: quant, dequant, scale-layout conversion, repack.
- RoPE stays in `attention` (standalone RoPE, position ids, QK-norm + RoPE).
- `kvcache` owns ops that read/write KV-cache contents; `attention` owns the
index/shape/schedule tables (page tables, `cu_seqlens`, `kv_indices`, split
metadata) and index selection. A fused compute + store stays with its
compute group.
- Model-specific files/subpackages and tuning data are allowed **inside** a
logical group. Do not add a model bundle or an implementation file directly
under `ops/`. Propose a new logical group only for a distinct responsibility,
and update `ops._GROUPS`, documentation, and tests together.
- Classify a fused operator by its complete contract. Do not split a fused
kernel into separate launches just to separate norm, RoPE, or quantization.
Split unrelated public entry points that happen to share a source file.
- Keep shared CUDA build/runtime infrastructure in `kernels/jit`; operator
wrappers call it from their logical group. Keep process groups, communicator
state, model dispatch, and buffer ownership in `srt`. K3-specific adapters in
`srt/layers/communication/` need not pretend to be generic interfaces.
and update `ops._GROUPS`, the README group list, and tests together
(`test_kernels_namespace.py` checks the README against `_GROUPS` and the
README layout against the directories under `kernels/`).
- Agent-generated implementations live in `kernels/kda_kernels/` (Kernel Design
Agents, `KernelBackend.KDA`) and `kernels/cake_kernels/` (Cake adapters over
FlashInfer). They are backends, not groups: register their public ops in the
logical group with a `KernelSpec` whose target points into them, and do not
reorganize those directories.
- Keep shared build/runtime infrastructure (including the Helion runtime
helper) in `kernels/jit`; operator wrappers call it from their logical group.
Keep process groups, communicator state, workspace classes, model dispatch,
and buffer ownership in `srt` (`srt/layers/communication/` for the K3 and Muse
Spark adapters); kernels must not import srt runtime modules to inspect
model objects, so pass flags in. Adapters there need not pretend to be generic
interfaces.
## Register the public entry point
Add lazy `KernelSpec` metadata in the owning group's `__init__.py`, or use the
existing `BaseFusedOp` registration when the operation has interchangeable
backends. Group imports must not import GPU implementations or compile kernels.
backends (a `BaseFusedOp` built per layer registers `module:Class.forward_<backend>`
targets). Group imports must not import GPU implementations or compile kernels.
Use `<group>.<name>` for the op id and `module:callable` for the target. Preserve
existing backend dispatch and describe actual device/architecture restrictions
@@ -53,8 +75,9 @@ factories, reference helpers, and runtime classes are not separate kernel APIs.
runtime module. A CUDA allocation does not make a runtime test a kernel test.
- Non-CI smoke scripts and benchmarks: `test/manual/kernels/`; do not leave
standalone test/benchmark entry points in production operator modules.
- Shared test helpers: `sglang.test.kernels`. Preserve vendored upstream trees
and AOT wheel packaging boundaries instead of reorganizing them incidentally.
- Shared test helpers and reference oracles: `sglang.test.kernels`. Preserve
vendored upstream trees and AOT wheel packaging boundaries instead of
reorganizing them incidentally.
For a move/split, preserve assertions, parametrization, fixtures, platform skips,
execution entry points, and CI stages/runners. Apportion existing time estimates
@@ -49,15 +49,15 @@ in-flight row as shipped.
| FlashInfer unified `allreduce_fusion` | `cross_device_reduce_1stage*`<br>`all_reduce`<br>`FusedAddRMSNormKernel`<br>`rmsnorm*` | `python/sglang/srt/layers/flashinfer_comm_fusion.py`<br>`python/sglang/srt/layers/layernorm.py::forward_with_allreduce_fusion`<br>`python/sglang/srt/layers/layer_boundary/residual/add_norm.py::apply_flashinfer_allreduce_fusion` | FlashInfer workspace creation plus `allreduce_fusion(..., pattern=AllReduceFusionPattern.kARResidualRMSNorm, ...)` | First suspect missing / disabled / unsupported FlashInfer allreduce fusion, not a brand new TP fusion idea. |
| AITER allreduce fusion | ROCm all-reduce plus RMSNorm still split | `python/sglang/srt/layers/layernorm.py::forward_with_allreduce_fusion`<br>`python/sglang/srt/distributed/communication_op.py::tensor_model_parallel_fused_allreduce_rmsnorm`<br>`python/sglang/srt/layers/layer_boundary/residual/add_norm.py::apply_aiter_all_reduce_fusion` | ROCm-side fused TP all-reduce + RMSNorm with fallback to plain all-reduce plus norm | On AMD, rule out existing AITER fusion before proposing a new communication fusion. |
| Fused activation-and-mul (`SwiGLU` / `GeGLU`) | `silu_and_mul`<br>`gelu_and_mul`<br>`npu_swiglu` | `python/sglang/srt/layers/activation.py` | Single op covers activation plus elementwise multiply across CUDA / CPU / NPU / XPU backends | Treat separate activation + mul on packed MLP outputs as missing existing fusion. |
| Fused dual residual RMSNorm | residual add plus two RMSNorm-like kernels around Grok blocks | `python/sglang/srt/layers/elementwise.py::fused_dual_residual_rmsnorm`<br>`python/sglang/srt/models/grok.py` | One Triton kernel computes intermediate residual update and next RMSNorm output together | On Grok-like residual layouts, treat split residual + norm as missing existing fusion. |
| Fused dual residual RMSNorm | residual add plus two RMSNorm-like kernels around Grok blocks | `python/sglang/kernels/ops/layernorm/fused_rmsnorm.py::fused_dual_residual_rmsnorm`<br>`python/sglang/srt/models/grok.py` | One Triton kernel computes intermediate residual update and next RMSNorm output together | On Grok-like residual layouts, treat split residual + norm as missing existing fusion. |
| In-place QK RMSNorm | split `q_norm` / `k_norm` kernels | `python/sglang/srt/models/utils.py::apply_qk_norm`<br>`python/sglang/kernels/ops/layernorm/norm.py::fused_inplace_qknorm` | In-place JIT QK norm plus optional `alt_stream` overlap for K | Check shape, dtype, deterministic mode, and in-place legality before proposing a new QK fuse. |
| TorchInductor horizontal Q/K norm combo-kernels | `combo_kernels`<br>`benchmark_combo_kernel`<br>`q_norm`<br>`k_norm`<br>`split_with_sizes` | `torch._inductor.config.combo_kernels` | TorchInductor can horizontally fuse sibling Q-norm and K-norm kernels in compiled traces, often deleting `split_with_sizes` / `clone` ladders | Treat separate Q/K norm ladders in compile-heavy traces as an existing compiler-fusion family first. |
| MiniMax TP fused QK RMSNorm | `MiniMaxM2RMSNormTP`<br>`rms_sumsq_serial`<br>`rms_apply_serial`<br>`forward_qk` | `python/sglang/srt/models/minimax_m2.py` | Triton kernels compute Q / K sumsq together, TP all-reduces shared stats, then apply both RMSNorms together | On MiniMax traces, separate Q norm and K norm are usually a missed model-specific Triton fusion. |
| Fused QK RMSNorm + RoPE | `qknorm*` + `rope*` + `rotary*` as separate steps | `python/sglang/kernels/ops/attention/fused_qknorm_rope.py`<br>`python/sglang/srt/models/qwen3_moe.py` | One JIT kernel applies QK RMSNorm and RoPE in-place on packed QKV | For compatible LLMs, classify split QK norm + RoPE as a missing existing fusion. |
| Fused QK RoPE reshape + KV cache write | `fused_qk_rope_reshape_and_cache*`<br>RoPE followed by reshape / cache DtoD | `python/sglang/srt/layers/attention/utils.py::fused_qk_rope_reshape_and_cache` | One Triton kernel applies RoPE to Q / K, reshapes cache layout, and writes K / V directly to paged cache | Treat separate RoPE + reshape + cache-write ladders as an existing attention-prep fusion family. |
| Fused QK RoPE reshape + KV cache write | `fused_qk_rope_reshape_and_cache*`<br>RoPE followed by reshape / cache DtoD | `python/sglang/kernels/ops/attention/fused_qk_rope_reshape_and_cache.py::fused_qk_rope_reshape_and_cache` | One Triton kernel applies RoPE to Q / K, reshapes cache layout, and writes K / V directly to paged cache | Treat separate RoPE + reshape + cache-write ladders as an existing attention-prep fusion family. |
| Fused RoPE + KV cache store | `fused_set_kv_buffer`<br>RoPE followed by KV-store, DtoD, or cache-write kernels | `python/sglang/kernels/ops/attention/rope.py`<br>`python/sglang/srt/models/utils.py::enable_fused_set_kv_buffer` | Shared entrypoints can route to fused RoPE + KV-store or model-side `fused_set_kv_buffer` fast paths | Compare against the fused cache-store path before proposing a new KV rewrite. |
| Fused decode metadata setup | `normal_decode_set_metadata`<br>`cache_seqlens_int32`<br>`cu_seqlens_k`<br>`page_table`<br>`swa_page_table` | `python/sglang/srt/layers/attention/flashattention_backend.py::normal_decode_set_metadata` | Triton decode path fuses seq-len cast/add, prefix-sum, req-to-token gather, page-table divide, and optional SWA metadata build into 1-2 kernels | If decode exposes multiple tiny metadata kernels before attention, first compare against this existing fused metadata-prep path. |
| NSA fused metadata copy for graph replay | `fused_metadata_copy`<br>`fused_metadata_copy_multi`<br>`fused_nsa_cache_seqlens`<br>`fused_flashmla_metadata` | `python/sglang/kernels/ops/attention/fused_metadata_copy.py` | CUDA graph replay path fuses multiple metadata copies into one kernel or one multi-destination kernel | Treat bursts of tiny metadata-copy kernels around NSA replay as a missed existing replay fusion. |
| NSA fused metadata copy for graph replay | `fused_metadata_copy`<br>`fused_metadata_copy_multi`<br>`fused_nsa_cache_seqlens`<br>`fused_flashmla_metadata` | `python/sglang/kernels/ops/attention/dsa/fused_metadata_copy.py` | CUDA graph replay path fuses multiple metadata copies into one kernel or one multi-destination kernel | Treat bursts of tiny metadata-copy kernels around NSA replay as a missed existing replay fusion. |
| DeepSeek MLA fused projection + norm + RoPE | `qkv_proj_with_rope_fused_weight`<br>`fused_qkv_a_proj_with_mqa`<br>`forward_absorb_fused_mla_rope*` | `python/sglang/srt/models/deepseek_common/attention_forward_methods/forward_mla_fused_rope_cpu.py`<br>`python/sglang/srt/models/deepseek_common/attention_forward_methods/forward_mla_fused_rope_rocm.py`<br>`python/sglang/srt/models/deepseek_v2.py` | CPU / ROCm paths fuse DeepSeek MLA projection packing with q / k norm, RoPE, and cache-oriented MLA prep | For DeepSeek MLA, split proj / norm / rope prep is usually an existing backend-specific fuse that did not fire. |
| Fused QK RoPE concat + MLA cache write | `fused_qk_rope_cat_and_cache_mla`<br>`set_mla_kv_buffer` | `python/sglang/srt/layers/rocm_linear_utils.py`<br>`python/sglang/srt/models/deepseek_common/attention_forward_methods/forward_mla.py` | ROCm MLA path can fuse Q / K RoPE packing, concat, and MLA cache write in one backend-specific op | On DeepSeek / MLA traces, separate RoPE-cat-cache steps are not automatically novel. |
| Qwen3 decode fused QK norm + 3D mRoPE + KV cache write | `fused_qk_norm_mrope_3d_cache_pts_quant_shuffle`<br>`mrope`<br>decode cache write | `python/sglang/srt/models/qwen3.py` | ROCm / AITER decode path fuses QK norm, 3D mRoPE, and paged KV cache write | On Qwen3-style decode, separate norm + mRoPE + cache-store kernels are not a novel opportunity. |
@@ -69,18 +69,18 @@ in-flight row as shipped.
| Qwen-style shared-expert append into routed top-k output | `_append_shared_to_topk_output`<br>`fused_append_shared_experts_with_weights`<br>`num_fused_shared_experts` | `python/sglang/srt/models/qwen2_moe.py`<br>`python/sglang/srt/layers/moe/moe_runner/triton_utils/fused_moe_triton_kernels.py` | Qwen-style MoE paths can append shared-expert ids and sigmoid gate weights to routed top-k output in one Triton kernel so the shared experts execute inside the fused MoE path | Treat routed top-k plus shared-expert pad / concat ladders as an existing MoE-prep fusion family first. |
| Fused MoE dispatch / permute / combine | token permutation<br>dispatch / combine<br>grouped top-k<br>many small MoE support kernels | `python/sglang/srt/layers/moe/fused_moe_triton/layer.py`<br>`python/sglang/srt/layers/moe/fused_moe_triton/fused_moe.py` | `FusedMoE` plus DeepEP / FlashInfer / FuseEP / standard dispatch backends and `permute_fusion=True` | First ask whether the model is missing an existing `FusedMoE`-style path or backend-specific dispatcher path. |
| Fused MoE sum + all-reduce | routed MoE followed by explicit sum-reduce kernels | `python/sglang/srt/layers/moe/fused_moe_triton/fused_moe.py`<br>`python/sglang/srt/layers/moe/fused_moe_triton/fused_moe_triton_kernels.py` | `fuse_sum_all_reduce=True` path in the second MoE GEMM | Before inventing a new MoE reduction fuse, check whether `enable_fused_moe_sum_all_reduce` is simply off or the quant path is incompatible. |
| Fused MoE activation + quant / re-quant | `silu_and_mul_*quant*`<br>`npu_dequant_swiglu_quant`<br>`swiglu_quant` | `python/sglang/srt/layers/moe/ep_moe/kernels.py`<br>`python/sglang/kernels/ops/quantization/nvfp4_gemm_swiglu_nvfp4_quant.py`<br>`python/sglang/srt/layers/moe/cutlass_w4a8_moe.py`<br>`python/sglang/srt/hardware_backend/npu/quantization/moe_methods.py` | Quantized MoE backends fuse SwiGLU / SiLU-and-mul with FP8 / FP4 / NPU re-quant before the second expert GEMM | If MoE traces show standalone activation then quant kernels, first check whether the quantized fused path is missing. |
| Fused MoE activation + quant / re-quant | `silu_and_mul_*quant*`<br>`npu_dequant_swiglu_quant`<br>`swiglu_quant` | `python/sglang/srt/layers/moe/ep_moe/kernels.py`<br>`python/sglang/kernels/ops/gemm/nvfp4_gemm_swiglu_nvfp4_quant.py`<br>`python/sglang/srt/layers/moe/cutlass_w4a8_moe.py`<br>`python/sglang/srt/hardware_backend/npu/quantization/moe_methods.py` | Quantized MoE backends fuse SwiGLU / SiLU-and-mul with FP8 / FP4 / NPU re-quant before the second expert GEMM | If MoE traces show standalone activation then quant kernels, first check whether the quantized fused path is missing. |
| DeepSeek comm-prep fused RMSNorm + quant / flatten-quant | `fused_rms_fp8_group_quant`<br>`fused_rms_mxfp4_quant`<br>`fused_flatten_fp8_group_quant`<br>`fused_flatten_mxfp4_quant` | `python/sglang/srt/layers/layer_boundary/residual/add_norm.py`<br>`python/sglang/srt/models/deepseek_common/attention_forward_methods/forward_mla.py`<br>`python/sglang/srt/models/deepseek_common/attention_forward_methods/forward_mha.py` | DeepSeek MLA / MHA ROCm paths fuse RMSNorm or flatten with FP8 / MXFP4 quantization for comm / attention prep | On DeepSeek quant traces, split norm + quant or flatten + quant is an existing family, not a new idea. |
| NSA fused top-k transform / page-table build | `fast_topk_transform_fused`<br>`fast_topk_transform_ragged_fused` | `python/sglang/srt/layers/attention/dsa_backend.py` | NSA can fuse top-k selection with paged / ragged index transform instead of separate top-k plus metadata scatter | If NSA top-k metadata work is split, check `SGLANG_DSA_FUSE_TOPK` and backend support first. |
| NSA fused quantize + indexed K-cache store | `fused_store_index_k_cache`<br>`act_quant`<br>`index_k_with_scale_buffer` | `python/sglang/kernels/ops/attention/fused_store_index_cache.py`<br>`python/sglang/srt/layers/attention/nsa/nsa_indexer.py` | Single JIT kernel quantizes bf16 K to fp8 + scale and writes directly into NSA index cache | Treat split `act_quant` + buffer-store on CUDA as missing an existing fused store path. |
| NSA fused quantize + indexed K-cache store | `fused_store_index_k_cache`<br>`act_quant`<br>`index_k_with_scale_buffer` | `python/sglang/kernels/ops/kvcache/dsa_fused_store_index_cache.py`<br>`python/sglang/srt/layers/attention/nsa/nsa_indexer.py` | Single JIT kernel quantizes bf16 K to fp8 + scale and writes directly into NSA index cache | Treat split `act_quant` + buffer-store on CUDA as missing an existing fused store path. |
| Fused sampling temperature + softmax | `fused_temperature_softmax*` | `python/sglang/srt/layers/fused_sampling.py`<br>`python/sglang/srt/layers/sampler.py` | Triton single-pass / multi-pass kernels fuse temperature scaling and softmax during decode | Separate temp-divide + softmax at decode batch sizes is often a missed existing fusion. |
| Fused logit softcap | `fused_softcap`<br>`final_logit_softcapping` | `python/sglang/srt/layers/elementwise.py`<br>`python/sglang/srt/layers/logits_processor.py` | Triton kernels fuse cast-to-float and softcap / tanh math for logits or generic elementwise softcapping | Treat exposed cast + softcap ladders as an existing Triton fuse family. |
| Linear-attention packed projection reshuffle | `fused_qkvzba_split_reshape_cat*`<br>`qkvz_proj`<br>`ba_proj`<br>`qkvabz_proj`<br>`fused_qkvbfg_a_proj` | `python/sglang/kernels/ops/attention/triton_gdn_fused_proj.py`<br>`python/sglang/srt/models/qwen3_next.py`<br>`python/sglang/srt/models/qwen3_5.py`<br>`python/sglang/srt/models/kimi_linear.py`<br>`python/sglang/srt/models/jet_nemotron.py` | GDN / Kimi / Jet-style linear-attn models pack multiple projections, then fuse split / reshape / cat into one kernel | Treat split reshape / transpose / cat ladders as an existing linear-attention fusion family. |
| Fused GDN gating prep | `fused_gdn_gating`<br>`softplus`<br>`beta_output` | `python/sglang/srt/layers/attention/fla/fused_gdn_gating.py` | Triton kernel computes GDN gate preparation such as `-exp(A_log) * softplus(...)` and `sigmoid(b)` together | On GDN traces, treat split gate-prep elementwise kernels as missing existing fusion first. |
| Fused GDN gating prep | `fused_gdn_gating`<br>`softplus`<br>`beta_output` | `python/sglang/kernels/ops/attention/linear/fused_gdn_gating.py` | Triton kernel computes GDN gate preparation such as `-exp(A_log) * softplus(...)` and `sigmoid(b)` together | On GDN traces, treat split gate-prep elementwise kernels as missing existing fusion first. |
| Fused RMSNorm-gated linear-attention output | `FusedRMSNormGated`<br>`layer_norm_gated_fwd` | `python/sglang/srt/layers/attention/fla/fused_norm_gate.py`<br>`python/sglang/srt/models/qwen3_next.py`<br>`python/sglang/srt/models/kimi_linear.py` | One Triton op covers residual-aware (RMS)Norm plus sigmoid / swish gating | If norm and output gate appear as separate kernels in GDN / Kimi-like blocks, first suspect a missing existing fusion. |
| Fused gated RMSNorm / LayerNorm | `rms_norm_gated`<br>`layer_norm_gated` | `python/sglang/srt/layers/attention/mamba/ops/layernorm_gated.py` | Mamba-derived kernels can fuse normalization with the gating branch `z * sigmoid(z)` | Treat split norm and gate post-processing on Mamba-style blocks as an existing fusion family. |
| Fused linear-attention chunk KKT + solve_tril | `chunk_gated_delta_rule_fwd_kkt_solve_kernel`<br>`scaled_dot_kkt`<br>`solve_tril`<br>`recompute_w_u` | `python/sglang/srt/layers/attention/fla/chunk_fwd.py`<br>`python/sglang/srt/layers/attention/fla/kda.py` | GDN / KDA chunk forward fuses `scaled_dot_kkt + solve_tril` in the prefill / intra-chunk path, then finishes `recompute_w_u` as the next step | Treat split KKT + triangular-solve ladders as an existing linear-attention fusion family first. |
| Fused linear-attention recurrent / KDA update | `fused_sigmoid_gating_delta_rule_update`<br>`fused_recurrent_gated_delta_rule_update`<br>`fused_kda_gate` | `python/sglang/srt/layers/attention/fla/fused_sigmoid_gating_recurrent.py`<br>`python/sglang/srt/layers/attention/fla/fused_recurrent.py`<br>`python/sglang/srt/models/kimi_linear.py`<br>`python/sglang/srt/models/jet_nemotron.py` | Triton / CuTeDSL kernels fuse gating math, optional QK l2norm, recurrent state update, and output generation | Treat split gating + recurrent-update chains as existing linear-attention fusion, not a novel opportunity. |
| Fused linear-attention recurrent / KDA update | `fused_sigmoid_gating_delta_rule_update`<br>`fused_recurrent_gated_delta_rule_update`<br>`fused_kda_gate` | `python/sglang/kernels/ops/attention/linear/fused_sigmoid_gating_recurrent.py`<br>`python/sglang/kernels/ops/attention/fla/fused_recurrent.py`<br>`python/sglang/srt/models/kimi_linear.py`<br>`python/sglang/srt/models/jet_nemotron.py` | Triton / CuTeDSL kernels fuse gating math, optional QK l2norm, recurrent state update, and output generation | Treat split gating + recurrent-update chains as existing linear-attention fusion, not a novel opportunity. |
| Fused Mamba state gather/scatter with mask | `fused_mamba_state_scatter_with_mask`<br>`index_elementwise_kernel` | `python/sglang/srt/layers/attention/mamba/mamba_state_scatter_triton.py` | Triton kernel replaces multiple masked gather / scatter index kernels with one fused update | If Mamba verify/update shows many tiny index kernels, first compare against this existing fused path. |
| Staging-buffer fused gather / scatter | `_fused_gather_to_staging_kernel`<br>`_fused_scatter_from_staging_kernel` | `python/sglang/srt/disaggregation/common/staging_buffer.py` | Triton kernels gather scattered KV slices into contiguous staging memory and scatter them back into KV cache on decode | Treat ladders of tiny gather/scatter/copy kernels in heterogeneous TP staging as missing an existing Triton fusion. |
@@ -146,7 +146,7 @@ Stable entries should be folded into the mainline family rows above.
| PR `#22005` fused add + RMSNorm + per-token FP8 quant | `fused_add_rmsnorm_per_token_quant`<br>`per_token_quant_fp8` | `PR #22005`<br>`python/sglang/kernels/jit/csrc/elementwise/fused_add_rmsnorm_per_token_quant.cuh`<br>`python/sglang/kernels/jit/fused_add_rmsnorm_per_token_quant.py` | CUDA JIT kernel keeps normed values in registers and emits BF16 + FP8 outputs plus per-token scales | If FP8 online-quant traces show add+norm followed by per-token quant, treat this as an in-flight upstream CUDA fuse family. |
| PR `#20667` Qwen3.5 fused QK norm + RoPE + KV cache write | `fused_qk_norm_rope_cache_pts_quant_shuffle`<br>`fused_qk_norm_mrope_3d_cache_pts_quant_shuffle`<br>`rotary_dim` | `PR #20667`<br>`python/sglang/srt/models/qwen3_5.py`<br>`python/sglang/srt/models/utils.py` | ROCm / AITER path fuses Q / K RMSNorm, partial or 3D RoPE, and direct KV cache write for Qwen3.5 attention | Treat split QK-norm + RoPE + cache-store on Qwen3.5 as a concrete in-flight upstream family, not a novel idea. |
| PR `#22392` CUTLASS FP8 GEMM replacing nvjet | `cutlass_scaled_mm`<br>`fp8_scaled_mm`<br>`nvjet`<br>`cudaMemsetAsync` | `PR #22392`<br>`python/sglang/kernels/aot/python/sgl_kernel/gemm.py`<br>`python/sglang/srt/layers/quantization/fp8_utils.py` | Runtime replacement swaps nvjet FP8 GEMMs for CUTLASS kernels, removing per-launch memset bubbles and extra output-copy kernels | Treat nvjet GEMM + memset bubble ladders as an in-flight SGLang linear-kernel family before calling them novel. |
| PR `#18612` NVFP4 CUTLASS MoE fused SiLU+Mul+quant | `silu_and_mul_scaled_nvfp4`<br>`nvfp4 expert quant`<br>`cutlass moe` | `PR #18612`<br>`python/sglang/srt/layers/moe/cutlass_w4a8_moe.py`<br>`python/sglang/kernels/ops/quantization/nvfp4_gemm_swiglu_nvfp4_quant.py` | Fuses MoE activation epilogue and NVFP4 expert quantization before the CUTLASS MoE second GEMM | Treat split SiLU+Mul then NVFP4 expert quant in CUTLASS MoE traces as an in-flight upstream SGLang family. |
| PR `#18612` NVFP4 CUTLASS MoE fused SiLU+Mul+quant | `silu_and_mul_scaled_nvfp4`<br>`nvfp4 expert quant`<br>`cutlass moe` | `PR #18612`<br>`python/sglang/srt/layers/moe/cutlass_w4a8_moe.py`<br>`python/sglang/kernels/ops/gemm/nvfp4_gemm_swiglu_nvfp4_quant.py` | Fuses MoE activation epilogue and NVFP4 expert quantization before the CUTLASS MoE second GEMM | Treat split SiLU+Mul then NVFP4 expert quant in CUTLASS MoE traces as an in-flight upstream SGLang family. |
| PR `#22918` FlashInfer per-token NVFP4 MoE | `per_token_nvfp4`<br>`trtllm_fp4_block_scale_moe`<br>`FlashInfer MoE` | `PR #22918`<br>`python/sglang/srt/layers/moe/fused_moe_triton/fused_moe.py` | Adds FlashInfer-backed per-token NVFP4 MoE execution so expert quant/dequant work can move into the fused MoE backend | Treat standalone per-token NVFP4 MoE support kernels as a candidate missing backend-selection path, not an automatically novel kernel idea. |
| PR `#22851` NSA top-k backend and FlashInfer / PyTorch top-k split | `nsa topk`<br>`flashinfer_topk`<br>`pytorch_topk`<br>`fast_topk_transform` | `PR #22851`<br>`python/sglang/srt/layers/attention/dsa_backend.py` | Makes NSA top-k backend selection explicit and aligns fused top-k transform with FlashInfer / PyTorch fallbacks | When NSA top-k dominates decode, first classify it as backend selection or fused-transform eligibility work. |
| PR `#24125` GLM5 NSA decode CatArrayBatchedCopy removal | `CatArrayBatchedCopy`<br>`GLM-5`<br>`NSA`<br>`TileLang decode` | `PR #24125`<br>`python/sglang/srt/layers/attention/dsa_backend.py` | Skips redundant cat/copy work in the GLM5 NSA TileLang decode path | Treat cat/copy bursts in GLM5 NSA decode as a concrete in-flight cleanup opportunity. |
@@ -18,7 +18,7 @@ Use these upstream files when the workflow or behavior needs to be justified fro
## Scheduler-side trace writing
- `python/sglang/srt/managers/scheduler_profiler_mixin.py`
- `python/sglang/srt/managers/scheduler_components/profiler_manager.py`
- actual trace start/stop behavior
- filename pattern for `TP/DP/PP/EP` and optional stage suffixes
- `CUDA_PROFILER` and torch profiler handling
@@ -37,6 +37,3 @@ Use these upstream files when the workflow or behavior needs to be justified fro
- `test/registered/profiling/test_start_profile.py`
- validates `/start_profile` behavior, including `CUDA_PROFILER`
- `test/registered/profiling/test_profile_v2.py`
- validates stage-scoped trace outputs under `SGLANG_PROFILE_V2`
@@ -482,7 +482,9 @@ FUSION_PATTERN_REGISTRY: Tuple[FusionPatternSpec, ...] = (
),
FusionPatternSpec(
pattern="Fused QK RoPE reshape + KV cache write",
candidate_path="python/sglang/srt/layers/attention/utils.py",
candidate_path=(
"python/sglang/kernels/ops/attention/fused_qk_rope_reshape_and_cache.py"
),
active_keywords=("fused_qk_rope_reshape_and_cache",),
split_groups=(
("rotary", "rope", "mrope"),
@@ -533,7 +535,7 @@ FUSION_PATTERN_REGISTRY: Tuple[FusionPatternSpec, ...] = (
),
FusionPatternSpec(
pattern="NSA fused metadata copy for graph replay",
candidate_path="python/sglang/kernels/ops/attention/fused_metadata_copy.py",
candidate_path="python/sglang/kernels/ops/attention/dsa/fused_metadata_copy.py",
active_keywords=(
"fused_metadata_copy",
"fused_metadata_copy_multi",
@@ -700,7 +702,7 @@ FUSION_PATTERN_REGISTRY: Tuple[FusionPatternSpec, ...] = (
pattern="Fused MoE activation + quant / re-quant",
candidate_path=(
"python/sglang/srt/layers/moe/ep_moe/kernels.py"
"<br>python/sglang/kernels/ops/quantization/nvfp4_gemm_swiglu_nvfp4_quant.py"
"<br>python/sglang/kernels/ops/gemm/nvfp4_gemm_swiglu_nvfp4_quant.py"
"<br>python/sglang/srt/layers/moe/cutlass_w4a8_moe.py"
),
active_keywords=(
@@ -761,7 +763,7 @@ FUSION_PATTERN_REGISTRY: Tuple[FusionPatternSpec, ...] = (
FusionPatternSpec(
pattern="NSA fused quantize + indexed K-cache store",
candidate_path=(
"python/sglang/kernels/ops/attention/fused_store_index_cache.py"
"python/sglang/kernels/ops/kvcache/dsa_fused_store_index_cache.py"
"<br>python/sglang/srt/layers/attention/nsa/nsa_indexer.py"
),
active_keywords=("fused_store_index_k_cache",),
+87 -78
View File
@@ -15,7 +15,7 @@ One container owns process-static runtime state: `sglang.srt.runtime_context.Run
| runtime flags | `get_flags()` | state that is *not* a pure function of config: `capture` (cuda-graph lifecycle), `moe` (ACTIVE backends, swappable), `dp` (DP-attention runtime flags) | materialized at subsystem init; groups offer `override()` for tests |
| resources | `get_resources()`, `get_stream(name)`, `get_buffer(name, factory)` | process-level handles: graph pools, EPLB state, EP dispatcher state, named side streams, workspace buffers | lazy; cleared by `reset_context()` |
| per-forward | `get_forward()` | forward-scoped flags (multi-stream switch, MoE output buffer, attn-TP inputs, extend-in-batch) | contextvar-backed; `scoped(**kw)` restores on exit; new threads see defaults |
| parallel | `get_parallel()` | one spelling per name: ranks and group handles are the live topology (`@property`, read-through); every other name, sizes included, is a leaf of the parallel config bag | ranks/groups: after dist init; leaves: after publish |
| parallel | `get_parallel()` | one spelling per name: scoped override, then permanent stamp, then published parallel config leaf | ranks: stamped at publish; group handles: stamped at distributed init; configured and derived widths: after publish |
`reset_context()` (unit-test teardown) drops the published config and installs fresh
flags/resources/forward tiers.
@@ -87,13 +87,12 @@ with what the operator typed, not with what resolution decided.**
`ModelRunner.update_model_fields` overrides `model_path` / `load_format` for
the target runner.
- **Late launcher-stage resolution (pre-publish)**: a few rules cannot run inside
`__post_init__` — LoRA normalization, and the auto-parser detection that needs a
`ServerArgs.resolve_once` — LoRA normalization, and the auto-parser detection that needs a
tokenizer/chat-template load. They are resolution, not mutation, and they
**declare** via `arg_groups.overrides.declare_resolution(server_args, source,
**fields)`, the same call the rest of the pipeline makes; there is no
`declare_late_resolution` any more. *When* a declaration is made is not
something the code marks — the guardrails that used to read that marker
cover these sites through the ordinary keyword scan instead. The declaration lands
something the code marks. The declaration lands
in the stash on that very object, so every holder of it carries the decision —
the HTTP server, the multi-tokenizer workers it is serialized for, the
schedulers it forks — and each of them publishes bags projected from it. The
@@ -135,16 +134,23 @@ bag to override at all.
instance: **constructor arguments** (`ModelRunner(draft_attention_backend=...)`,
`MMEncoder(gpu_id=...)`) and **runner attributes holding the resolved value**
(`model_runner.kv_cache_dtype_str`, `prefill_attention_backend_str`,
`num_fused_shared_experts`, `linear_attn_backends`) — threaded to consumers as
`linear_attn_backends`) — threaded to consumers as
arguments, never backfilled onto a shared object. A per-runner choice also stays
*out* of the bags: recording it there is how a second runner inherits the first
one's answer, which is exactly the bug `linear_attn_backends` replaced. The one sanctioned bend in that rule is
*scoped*: `ModelRunner._load_format_scope` exposes the draft's
`--speculative-draft-load-format` through `get_model().override(load_format=...)`
for exactly the duration of the draft build, because model construction
reads that bag leaf — the override restores on exit, so nothing outlives
the scope. When there is a runner in hand, read its
stamp; that is a different rule from "read the instance".
one's answer, which is exactly the bug `linear_attn_backends` replaced. The two
sanctioned bends in that rule are *scoped*. `ModelRunner._load_format_scope`
exposes the draft's `--speculative-draft-load-format` through
`get_model().override(load_format=...)` for exactly the duration of the draft
build, because model construction reads that bag leaf — the override restores
on exit, so nothing outlives the scope. The shared-expert layout is another scoped view:
`get_model().shared_experts` selects the immutable `SharedExpertsConfig` owned
by the loaded model (`model.shared_experts_config`). The loader binds it for
construction, and runners bind it for execution, warmup and graph capture.
Both the fused count and ordinary TopK slot count belong to this object;
neither is a separately mutable flag. A scope selects a built layout and
restores the previous reference on exit, including exceptions. It cannot
resize weights, buffers or a captured graph. When the model is in hand, read
its owned config directly (as the routed-experts capturer does).
- **Per-instance boundaries** — the tokenizer-manager family, everything under
`entrypoints/`, and the tokenizer-process multimodal processors read the bags.
The old justification for keeping them on `self.server_args` ("several
@@ -178,8 +184,8 @@ bag to override at all.
### Four ways a config sweep breaks something no test runs
Each of these shipped in a review round and cost a real defect; each now has a
guard, named here so the next sweep checks the same four things by hand first.
Each of these shipped in a review round and cost a real defect. Where a test
guards one it is named; the next sweep checks all four by hand first.
1. **The other implementations of an interface.** Dropping a parameter means
auditing implementers, not just callers: `CustomSpecAlgo` is the plugin
@@ -191,7 +197,8 @@ guard, named here so the next sweep checks the same four things by hand first.
`publish` says nothing about whether a given read runs before it. Spawned
workers (`MMEncoder` for encoder DP/TP, the Ray scheduler actor) start with
an empty context, so a bag read above the publish raises only there.
Guard: `test_publish_precedes_bag_reads.py`.
Guarded on the launch path only (`test_pre_publish_readers.py`); other
entries, spawned workers included, are checked by review.
3. **The role namespace a process publishes under.** `ROLE_NAMESPACE_SETS`
narrows what each role may read; the DP controller is audited for `exec`
alone. A helper that reaches for another namespace passes every default-mode
@@ -201,7 +208,9 @@ guard, named here so the next sweep checks the same four things by hand first.
means enumerating the others: HTTP, gRPC and in-process `Engine` each have
their own server-info and model-info, and each passes its own tests while
its users lose the field.
Guard: `test_effective_state_surfaces.py`.
Guarded for server info across HTTP, both gRPC roles and the Rust frontend
(`TestServerInfoTransportParity` in `entrypoints/test_server_info.py`);
`Engine.get_server_info` and model info are checked by review.
A fifth, from the same rounds: the accessor **name itself**. Called as an
object member (`manager.get_disagg()`), or shadowed by a same-named import
@@ -223,12 +232,13 @@ slack (`>= len(...) - 1`) or compared key names instead of value sources.
### `get_parallel()`: one spelling per name
**There is no `.config` hop.** Ranks and group handles are `@property`
read-through over the canonical getters, so they answer with the live process
groups. Everything else — `tp_size`, `pp_size`, `attn_cp_size`, `dcp_size`,
`moe_dp_size` included, alongside config-only leaves like `nccl_port`,
`dp_size`, `ep_size`, `dwdp_size` — is answered from the
published `parallel` bag. Reading a leaf before publish raises a `ValueError`
**There is no `.config` hop.** `ParallelContext._read` answers every name in
one order: a scoped override, then a permanent stamp, then the published
`parallel` bag leaf. `publish` stamps ranks from the launcher; distributed init
stamps group handles and validates their widths. Configured widths such as
`tp_size`, `pp_size`, `attn_cp_size`, `dcp_size` and `moe_dp_size`, and config-only
leaves such as `nccl_port`, `dp_size`, `ep_size` and `dwdp_size`, fall back to the
published bag. Reading a config leaf before publish raises a `ValueError`
naming the namespace; an unknown name is an `AttributeError`.
A size reads from the configuration because the groups are built at exactly the
@@ -245,38 +255,33 @@ configured widths — checked at every assignment to `_TP` / `_PP` / `_ATTN_CP`
coordinators keep their construction width. Those are different names, not two
answers to one name.
DCP keeps its own pair: `get_parallel().attn_dcp_size` / `.dcp_enabled` answer the
*effective* topology (`1` / `False` with no group installed), while `dcp_size` is
DCP keeps its own pair: `get_parallel().attn_dcp_size` / `.dcp_enabled` are derived
from the published configuration unless explicitly overridden; `dcp_size` is
what the launch requested.
A process-global seed field-read of one of these sizes
(`get_server_args().tp_size`, or an alias of it) is a read-ratchet failure. A
`server_args` the object was *handed* is a different thing and not a ratchet
matter — see "Reads that legitimately stay on a ServerArgs instance".
Fail-loud is narrower: before dist init, a live *rank/group* read raises. The six
Fail-loud is narrower: an unstamped rank or group handle with no declared default
raises; ranks are available after publish, before dist init. The six
parallel quotients are not live reads at all — `attn_tp_size`, `attn_dp_size`,
`attn_dcp_size`, `moe_ep_size`, `moe_tp_size`, `dcp_enabled` are a function of the
configured leaves, computed once at publish into bag leaves, and answered
override → stamp → published leaf. So `dcp_enabled` means "the launch configured
DCP" (`dcp_size > 1`), not "a DCP group is installed here"; in a scheduler the
stamp makes the two identical, in a process that publishes without dist init they
differ. `test_a_topology_is_stated_by_naming_the_width` and its neighbours in
`test_runtime_context.py` pin this; they replaced
DCP" (`dcp_size > 1`), not "a DCP group is installed here". Normal distributed
initialization builds the groups to match the configured widths; it does not
restamp these DCP leaves. `test_a_topology_is_stated_by_naming_the_width` and its
neighbours in `test_runtime_context.py` pin this; they replaced
`test_attn_dcp_defaults_when_group_is_uninitialized`. One consequence for tests:
overriding a leaf no longer moves its quotient — state a topology by publishing a
config, or by naming the width. After init,
only the DCP group is optional (`_DCP` exists only when `dcp_size > 1`; attn-CP and
moe-DP always install, as size-1 aliases if unused). The `config` hop is
deliberately dynamo-traceable (a plain property over a slot, no
`object.__getattribute__`); gate helpers like `enable_moe_dense_fully_dp()` run inside
moe-DP always install, as size-1 aliases if unused). Bare parallel leaf reads are
dynamo-traceable; gate helpers like `enable_moe_dense_fully_dp()` run inside
compiled model forwards (`test_parallel_config_leaves_trace_under_torch_compile` pins
this).
A third surface carries the same names: `ParallelState` (`self.ps` / `mr.ps`), the
frozen per-process snapshot built once in `Scheduler.__init__` from these configured
sizes plus this process's ranks, and handed down (draft runners included). Prefer it
where an object was handed one; it is not a global accessor.
### Reading config: the seed is off limits
`get_server_args().field` in business code is a ratchet failure. Read:
@@ -322,12 +327,13 @@ where an object was handed one; it is not a global accessor.
follows a post-publish override. Two questions are *not* that, and have their
own spelling: the width of the MoE communicator you are about to collectively
operate on is `get_moe_cp_size()` (the `_MOE_DP = _ATTN_CP` alias makes it
differ), and the effective DCP topology is `get_parallel().attn_dcp_size` /
`.dcp_enabled` (`1` / `False` when no group is installed), which does not need
dist init to answer.
differ), and DCP's derived width and enablement are
`get_parallel().attn_dcp_size` / `.dcp_enabled` (published configuration
unless explicitly overridden).
- **this runner's resolved value** → the runner
(`prefill_attention_backend_str`, `kv_cache_dtype_str`,
`draft_attention_backend`, `num_fused_shared_experts` on the model).
`draft_attention_backend`); for what the model's layers bake in at
construction, the layer (`num_fused_shared_experts` on a MoE block).
`self.server_args.field` is still right for handed per-instance config (see
"Reads that legitimately stay on a ServerArgs instance" above for the full set —
@@ -335,10 +341,9 @@ per-instance boundaries and whole-object passes; there are no per-runner config
copies to read any more). The allow-list is `GrammarManager` and `MMEncoder`;
what sits beside it is residue, not a family — and not for one single reason:
- the tokenizer-manager family and `entrypoints/` **read the bags**; what is
left of them in the exposure ratchet is a handful of individually-dispositioned
pairs, not a family awaiting conversion. Read the ratchet for the current set
rather than assuming a directory is off-limits;
- the tokenizer-manager family and `entrypoints/` **read the bags**; review a new
instance read there against the raw-input contract rather than assuming a
directory is off-limits;
- `GrammarManager` is a handed instance for its residual `self.server_args`
reads, but backend selection is **not** on the instance any more:
`create_grammar_backend` reads `get_exec().kernel.grammar_backend`, and
@@ -384,15 +389,15 @@ if you do it, say so in the test.
### Mid-resolution reads (inside the pipeline only)
Resolution runs in `__post_init__` and **writes nothing onto the record**: a
handler declares (`self._declare` / `declare_resolution`), the declaration goes
Resolution runs in `ServerArgs.resolve_once` and **writes nothing onto the record**: a
handler declares (`declare_resolution`), the declaration goes
into the stash, and the fields keep what the caller passed. So a mid-resolution
read of a field answers with the *raw input* — every reader in the pipeline goes
through a view instead:
- `resolving_view(server_args)` / `self._resolved()` — the live view (walks the
- `resolving_view(server_args)` — the live view (walks the
stash per read). This is what handlers and hooks bind, conventionally as
`cfg = resolving_view(self)` at the top of the handler.
`cfg = resolving_view(server_args)` at the top of the handler.
- `resolved_view(server_args)` — snapshots the overlay when built, which is what
a post-process pass wants: it reads the state at *its* slot.
@@ -451,13 +456,18 @@ Never assign `server_args` fields from model code. Declare instead
its checkpoint drives (shared-experts fusion) is a question the *loader* asks
the model class — `shared_experts_fusion_disable_reason(hf_config,
quant_config)`, a classmethod answering without an instance — at the single
model-instantiation point, and
`install_shared_experts_fusion_decision` writes the answer to the ACTIVE moe
flag before that model's layers build and read it
(`is_shared_experts_fusion_disabled`, config-intent fallback).
`draft_model_build_scope` brackets every draft build and routes the draft's
answer to the speculative leaf, so a draft's decision never overwrites the
target's. A process-level load-time fact (the sm80 dtype fallback —
model-instantiation point, together with how many shared experts it fuses
(`num_fusable_shared_experts(hf_config)`; an entry class without both hooks
fuses none). `resolve_shared_experts_config` returns one immutable config;
the loader scopes it through `get_model().override(shared_experts=config)`
while building and retains it on the resulting model. `num_shared_topk_slots`
derives from the fused count and frozen Waterfill policy, so ordinary routing
and Waterfill cannot both claim the shared slots. `TopK`/`HashTopK` retain the
same config; `top_k` counts only routed experts at construction.
`draft_model_build_scope` still selects draft boundary/MMA preferences; it
does not save shared-expert flags. Target and draft each own their layout.
These values have model lifetime, so they do not belong in `get_forward()`.
A process-level load-time fact (the sm80 dtype fallback —
device-driven, identical for every runner) records directly via
`get_context().override`.
@@ -466,12 +476,12 @@ dataclass. A declaration against a non-whitelisted field fails at its slot.
### Load-time vs resolution-time (critical)
`__post_init__` runs in the launcher process before any model/platform import. Logic that
`ServerArgs.resolve_once` runs in the launcher process before any model/platform import. Logic that
consults an **extensible registry** (e.g. out-of-tree platforms registering attention
backends in `init_backend()`, which runs at `model_runner` import) must stay at load time
(ModelRunner init), writing through `get_context().override()`. Before moving any
load-time logic into resolution, verify everything it reads is already complete at
construction time.
resolution time.
## Runtime flags (`get_flags()`)
@@ -499,8 +509,8 @@ Named slots + two keyed-lazy registries:
side stream is `"lora_side"`. Two call sites may share a name only if their work belongs
on one stream — sharing across roles serializes intended overlap.
- `get_buffer(name, factory)` — get-or-create a named persistent buffer. Grow-only or
per-device semantics manage their `resources.buffers` entries directly (see tokenspeed /
SM120 split / Marlin workspace). Buffer names are per-backend today; do not silently
per-device semantics manage their `resources.buffers` entries directly (see SM120
split / Marlin workspace). Buffer names are per-backend today; do not silently
share.
- Singletons with manager semantics (EP dispatcher buffers, EPLB recorder/metadata, graph
memory pool) keep their owning accessors/classes as facades; only the *state* lives in a
@@ -543,8 +553,9 @@ ONE thread — do not design for TBO threads that don't exist.
them explicitly on the mock; `MagicMock(spec=...)` raises on attributes that only
exist post-`__init__`, which is the fastest way to find a missed stub.
- `reset_context()` in teardown when a test publishes outside a scoped override.
- `ServerArgs(model_path="dummy")` early-returns the pipeline (few declarations, no
strict guard) — fine for lightweight fixtures.
- `ServerArgs(model_path="dummy")` resolves through the pipeline's early return (few
declarations) — fine for lightweight fixtures. The strict guard still applies once
`resolve_once` has run, as it does at `publish`.
- **Asserting what resolution decided reads `resolution_result(sa, "field")`**, not
`sa.field`: the field is the raw input. Assert the field only when the point of
the case *is* that the record stayed pristine (the FA4 page-size and waterfill
@@ -553,7 +564,7 @@ ONE thread — do not design for TBO threads that don't exist.
pytest run lets a context published by an earlier file mask a missing-publish bug in a
later one.
## Guardrails (these fail CI; what to do when they fire)
## Guardrails and review rules
1. **Strict mutation guard** (always on, and with no exception): bare
`server_args.x = ...` after resolution raises unconditionally in
@@ -563,10 +574,9 @@ ONE thread — do not design for TBO threads that don't exist.
resolved config with `get_context().override`; hand a per-runner value to its
runner as a constructor argument. Projected bags are sealed the same way (leaf
assignment raises).
2. **Mutation ratchet** (`test_server_args_mutation_ratchet.py`, exact pin 0 over the whole
package minus the pipeline / multimodal_gen): textual scan for assignment forms. Never
raise the baseline.
3. **No-copy contract** (`test_server_args_no_instance_mutation_entry.py`): neither
2. **No textual mutation ratchet**: the strict guard above is the only mutation
check.
3. **No-copy contract** (review only, no automated guard): neither
`ServerArgs.override` nor `ServerArgs.derive` exists, and nothing in the package
calls either form. Rerouting a writer to the bags means flipping **all its readers
in the same commit** (no transitional dual-write).
@@ -577,23 +587,23 @@ ONE thread — do not design for TBO threads that don't exist.
runner's stamp — not `get_server_args().field`, which the read ratchet below pins
at zero. `runtime_context.get_server_args()` is only for the
whole-object shapes (dumps, provenance, a hand-off to a callee that takes a config).
5. **Global config read ratchet** (`test_global_config_read_ratchet.py`): baselines are
**0** for both the direct `get_server_args().field` and the alias form (function-local
5. **Global config read check** (`test_global_config_read_ratchet.py`): no business code
reads a field off `get_server_args()`, directly or through an alias (function-local
— including local copies of an alias, `cfg = sa` — module-level, or parked on an
instance attribute, plus the `getattr(..., "field")` spelling of each; a name
computed at runtime or indirection deeper than a local name copy is census-tool
territory, per the test's docstring). The scanner matches `get_server_args` by its
computed at runtime or indirection deeper than a local name copy is outside the
check, per the test's docstring). The scanner matches `get_server_args` by its
literal name — bare or module-qualified (`ctx.get_server_args()`) — and
`TestNoRenamedAccessorImports` in the same file *bans* `import ... as` renames of it,
which is what makes literal-name matching sound. Exempt by owner
module only (`runtime_context.py`, `server_args.py`, `arg_groups/`). Two classes,
no more: `TestGlobalConfigReadRatchet` holds the two baselines and
no more: `TestGlobalConfigReads` holds the zero check and
`TestNoRenamedAccessorImports` holds the ban. There is no configured-size registry
here any longer — `get_parallel()` has one spelling per name, so a size read is not a
choice between two answers and nothing needs registering.
6. **Module-state ratchet** (`test_module_state_ratchet.py`): `global` statements in the
flag-owning layers are pinned by name. A new module-level runtime global belongs on a
flags group / resources slot instead; migrating a pinned survivor must shrink the pin.
6. **Module-state rule** (review only, no automated guard): a new module-level
runtime global in the flag-owning layers belongs on a flags group / resources slot
instead.
7. **Namespace coverage** (`test_server_args_namespaces.py`,
`test_runtime_context_config_bags.py`): every `ServerArgs` field resolves to a
namespace — from the `arg_groups/fields/` class that declares it — and the
@@ -649,8 +659,7 @@ Key source files: `python/sglang/srt/runtime_context.py` (the container, every t
`python/sglang/srt/arg_groups/overrides.py` (override registry, passes,
`declare_resolution` and the spellings around it), `python/sglang/srt/server_args.py` (`NS` metadata,
`Arg(..., resolvable=True)`, `__setattr__` strict guard), and the guardrail tests under
`test/registered/unit/` (`test_server_args_mutation_ratchet.py`,
`test_global_config_read_ratchet.py`,
`test_module_state_ratchet.py`, `test_server_args_namespaces.py`,
`test/registered/unit/` (`test_global_config_read_ratchet.py`,
`test_server_args_namespaces.py`,
`test_runtime_context.py` — the last one doubles
as executable documentation of every tier's semantics).
+1 -1
View File
@@ -36,4 +36,4 @@ a real `AttributeError` if the field is ever renamed:
revision=getattr(server_args, "revision", None), # BAD
revision=server_args.revision, # GOOD
```
(see `python/sglang/srt/managers/template_detection.py`)
(see `python/sglang/srt/parser/template_detection.py`)
+2 -2
View File
@@ -27,8 +27,8 @@ A new unit test case must fall into one of these categories:
3. **Critical-path bookkeeping.** Defends conventions that are easy to break by
forgetting to sync -- registry completeness, field lifecycle, serialization
compatibility. Enumerating assertions are fine here; the guarded failure
mode is "someone extended X without updating Y". Example: the ratchet tests
(`test/registered/unit/test_module_state_ratchet.py`).
mode is "someone extended X without updating Y". Example:
`test/registered/unit/test_global_config_read_ratchet.py`.
Not admissible:
+8
View File
@@ -1,12 +1,14 @@
.github @merrymercy @Fridge003 @ispobock @Kangyan-Zhou @HaiShaw @bingxche
/docker @Fridge003 @ispobock @HaiShaw @ishandhanani @yctseng0211 @sogalin
/docker/npu.Dockerfile @ping1jing2 @iforgetmyname @whybeyoung
/docker/sgl-router.Dockerfile @sherlockwu @Kangyan-Zhou @ShangmingCai
/docs @wisclmy0611 @zijiexia @sogalin
/docs @wisclmy0611 @zijiexia @Richardczl98 @JustinTong0323 @sogalin
/python/pyproject.toml @merrymercy @Fridge003 @ispobock
/python/sglang/kernels @DarkSharpness @HaiShaw @BBuf @celve @HydraQYH @yuan-luo
/python/sglang/kernels/ops/diffusion @yingluosanqian @BBuf @mickqian @triple-mu @niehen6174
/python/sglang/kernels/ops/attention/fla @yizhang2077 @hebiao064 @yuan-luo
/python/sglang/kernels/ops/attention/linear @yizhang2077 @hebiao064 @yuan-luo
/python/sglang/kernels/ops/lora @Ying1123 @Fridge003 @lifuhuang @yushengsu-thu @jybsuper
/python/sglang/kernels/jit/csrc/lora @Ying1123 @Fridge003 @lifuhuang @yushengsu-thu @jybsuper
/python/sglang/multimodal_gen @mickqian @ping1jing2 @HaiShaw @yichiche @AgainstEntropy @BBuf @triple-mu @kevin-mii @niehen6174
@@ -81,6 +83,11 @@
/rust/sglang-processor @sagearc @sherlockwu
/rust/sglang-radix-tree @Jialin @alphabetc1
/rust/sglang-server @rainj-me @merrymercy
/sgl-router @sherlockwu @Kangyan-Zhou @ShangmingCai @hzh0425
/sgl-router/sgl-kv-indexer @sherlockwu @lixiufei-leo @ShangmingCai @wuyl1 @hzh0425
/sgl-router/src/policies @sherlockwu @ShangmingCai @Bo-Vincent
/sgl-router/src/workers @sherlockwu @AMD-yanfeiwang
/sgl-router/src/protocol @sherlockwu @whybeyoung
/sgl-model-gateway @slin1237 @CatherineSue
/sgl-model-gateway/benches @slin1237
/sgl-model-gateway/bindings/python @CatherineSue @key4ng @slin1237
@@ -107,6 +114,7 @@
/python/sglang/srt/speculative/cpp_ngram @hnyls2002 @Qiaolin-Yu @kpham-sgl
/python/sglang/srt/speculative/frozen_kv_mtp_*.py @hnyls2002 @Qiaolin-Yu @kpham-sgl @pyc96
/python/sglang/kernels/ops/speculative/ngram_*.py @hnyls2002 @Qiaolin-Yu @kpham-sgl
/python/sglang/kernels/ops/embeddings/ngram_embedding.py @hnyls2002 @Qiaolin-Yu @kpham-sgl
/python/sglang/kernels/jit/csrc/ngram_corpus @hnyls2002 @Qiaolin-Yu @kpham-sgl
/python/sglang/srt/models/laguna.py @Jiminator
/python/sglang/srt/configs/laguna.py @Jiminator
+2 -2
View File
@@ -14,9 +14,9 @@ Recognized permission keys:
| --- | --- |
| `can_tag_run_ci_label` | `/tag-run-ci-label`, `/tag-and-rerun-ci`, `/run-full-ci`, `/run-extra-ci` |
| `can_rerun_failed_ci` | `/rerun-failed-ci`, `/tag-and-rerun-ci`, `/run-full-ci`, `/run-extra-ci` |
| `cooldown_interval_minutes` | rate limit in `pr-gate.yml`; `0` also grants `/rerun-test`, `/rerun-group` |
| `cooldown_interval_minutes` | rate limit in `pr-gate.yml`; `0` also grants `/rerun-test`, `/rerun-group`, `/run-cpu-ci` |
`/rerun-test` and `/rerun-group` are gated on the commenter alone: either
`/rerun-test`, `/rerun-group` and `/run-cpu-ci` are gated on the commenter alone: either
`cooldown_interval_minutes: 0`, or `write`/`admin` permission on the repo. Where
the PR comes from makes no difference, and authoring it grants nothing.
+2 -2
View File
@@ -11,7 +11,7 @@ inputs:
required: false
default: ''
cache_key_prefix:
description: 'Must match what _pr-test-rust-ext-build.yml saves under.'
description: 'Must match what _pr-test-rust-ext-build.yml saves under. ARM64 runners always restore rust-ext-aarch64-cp312.'
required: false
default: rust-ext-x86_64-cp312
@@ -53,7 +53,7 @@ runs:
path: |
python/sglang/srt/rust_extensions/_*.so
python/sglang/srt/mem_cache/rust_tree_core/mem_cache*.so
key: ${{ inputs.cache_key_prefix }}-${{ hashFiles('rust/**', 'proto/**', 'python/setup.py', 'python/pyproject.toml', 'python/sglang/srt/rust_extensions/torch_build.py', '.github/workflows/_pr-test-rust-ext-build.yml', 'scripts/ci/utils/stage_rust_ext_modules.sh') }}
key: ${{ runner.arch == 'ARM64' && 'rust-ext-aarch64-cp312' || inputs.cache_key_prefix }}-${{ hashFiles('rust/**', 'proto/**', 'python/setup.py', 'python/pyproject.toml', 'python/sglang/srt/rust_extensions/torch_build.py', '.github/workflows/_pr-test-rust-ext-build.yml', 'scripts/ci/utils/stage_rust_ext_modules.sh') }}
# Job-wide, but only setup.py reads it, and only while building.
# Whether the modules suit this interpreter is not decided here:
+22
View File
@@ -135,6 +135,28 @@ hicache:
- 'python/sglang/srt/mem_cache/**/*storage*'
- 'python/sglang/srt/mem_cache/l2_transfer.py'
# A *mooncake* filename glob also matches benchmark/datasets/mooncake.py.
mooncake:
- changed-files:
- any-glob-to-any-file:
# PD KV transfer.
- 'python/sglang/srt/disaggregation/mooncake/**/*'
- 'python/sglang/srt/distributed/device_communicators/mooncake_transfer_engine.py'
- 'python/sglang/srt/observability/mooncake_trace.py'
- 'test/registered/unit/disaggregation/**/*mooncake*'
- 'test/manual/kv_transfer/**/*mooncake*'
# HiCache store.
- 'python/sglang/srt/mem_cache/storage/mooncake_store/**/*'
- 'python/sglang/test/mooncake_utils.py'
- 'test/registered/unit/mem_cache/**/*mooncake*'
- 'test/registered/hicache/**/*mooncake*'
# MoE EP. elastic_ep.py also implements nixl.
- 'python/sglang/srt/layers/moe/token_dispatcher/mooncake.py'
- 'python/sglang/srt/elastic_ep/expert_backup_client.py'
- 'python/sglang/srt/elastic_ep/expert_backup_manager.py'
- 'test/registered/ep/**/*mooncake*'
- 'test/manual/ep/**/*mooncake*'
# Deterministic
deterministic:
- changed-files:
-133
View File
@@ -1,133 +0,0 @@
# Reusable workflow: cross-reference CI test failures with coverage-based
# test recommendations. Called by pr-test-npu.yml's analyze-failure-report job.
name: Analyze Failure Report
on:
workflow_call:
inputs:
log_artifact_pattern:
type: string
required: true
description: 'Glob pattern for test log artifacts (e.g. selected-test-logs-*)'
diffusion_artifact_pattern:
type: string
required: false
default: 'diffusion-failures-*'
description: 'Glob pattern for diffusion failure artifacts (consistency failures)'
recommendations_content:
type: string
required: false
default: ''
description: 'Content of recommendations from upstream job output'
defaults:
run:
shell: bash -el {0}
jobs:
analyze:
name: cross-reference failures with recommendations
continue-on-error: true # Non-blocking on parse/read failure.
runs-on: ubuntu-latest
steps:
- name: Checkout sglang repo
uses: actions/checkout@v4
# ----- Step 1: Locate recommendations -----
- name: Locate recommendations file
id: locate-recs
run: |
mkdir -p ./test-logs
# Always create the file so the analyzer can run on diffusion-only failures.
touch ./test-logs/_recommended.txt
echo "RECS_FILE=./test-logs/_recommended.txt" >> $GITHUB_ENV
if [ -n "${{ inputs.recommendations_content }}" ]; then
echo "has_recommendations=true" >> $GITHUB_OUTPUT
echo "Using workflow output recommendations"
echo "${{ inputs.recommendations_content }}" > ./test-logs/_recommended.txt
echo "RECS_SOURCE=output" >> $GITHUB_ENV
COUNT=$(wc -l < ./test-logs/_recommended.txt)
echo "Recommendations count: ${COUNT}"
else
echo "has_recommendations=false" >> $GITHUB_OUTPUT
echo "RECS_SOURCE=none" >> $GITHUB_ENV
echo "::notice::No recommended test cases detected; analyzing log and diffusion failures only."
fi
# ----- Step 2: Download test logs -----
# continue-on-error is a step keyword, not a download-artifact input.
- name: Download test logs
uses: actions/download-artifact@v7
with:
pattern: ${{ inputs.log_artifact_pattern }}
path: ./test-logs
merge-multiple: true
continue-on-error: true
# ----- Step 2b: Download diffusion failure artifacts -----
# Keep merge-multiple off so each partition's consistency_failures/summary.json
# lands in its own <artifact-name>/ directory instead of clobbering the others.
- name: Download diffusion failure artifacts
uses: actions/download-artifact@v7
with:
pattern: ${{ inputs.diffusion_artifact_pattern }}
path: ./diffusion-failures
continue-on-error: true
- name: Check for diffusion failure artifacts
id: check-diffusion
run: |
if compgen -G "./diffusion-failures/*/consistency_failures/summary.json" > /dev/null; then
echo "found=true" >> $GITHUB_OUTPUT
echo "Diffusion consistency failure artifacts found"
else
echo "found=false" >> $GITHUB_OUTPUT
echo "No diffusion consistency failure artifacts"
fi
- name: Show downloaded artifacts structure
run: |
echo "Downloaded log files:"
find ./test-logs -type f | head -50 || echo " (no logs found)"
echo "Downloaded diffusion failure artifacts:"
if compgen -G "./diffusion-failures/*" > /dev/null; then
find ./diffusion-failures -type f | head -50
else
echo " (no diffusion failure artifacts found)"
fi
# ----- Step 3: Run analysis -----
- name: Ensure regex is available
if: ${{ steps.locate-recs.outputs.has_recommendations != 'false' || steps.check-diffusion.outputs.found == 'true' }}
run: |
set -euo pipefail
if ! python3 -c "import regex" >/dev/null 2>&1; then
echo "regex not found for $(python3 -V); bootstrapping pip and installing regex"
curl -fsSL https://bootstrap.pypa.io/get-pip.py -o get-pip.py
python3 get-pip.py --break-system-packages
python3 -m pip install --break-system-packages regex
fi
python3 -c "import regex; print('regex ok:', regex.__file__)"
- name: Run failure analysis
if: ${{ steps.locate-recs.outputs.has_recommendations != 'false' || steps.check-diffusion.outputs.found == 'true' }}
id: analysis
continue-on-error: true
run: |
python3 scripts/ci/npu/precise-test/analyze_failure_report.py \
--log-dir ./test-logs \
--diffusion-dir ./diffusion-failures \
--recommendations-file "${RECS_FILE}" \
--recommendations-source "${RECS_SOURCE:-none}" \
--output ./test-logs/failure_report.md
# ----- Step 4: Upload report -----
- name: Upload failure analysis report
if: always()
uses: actions/upload-artifact@v7
with:
name: failure-analysis-report
path: ./test-logs/failure_report.md
if-no-files-found: ignore
retention-days: 14
+17 -14
View File
@@ -50,10 +50,10 @@ on:
description: 'If true, delegate test execution to scripts/ci/npu/precise-test/run_tests_with_coverage.sh instead of running via run_suite.py directly'
type: boolean
default: false
upload_test_logs:
description: 'If true, upload the test logs as a GitHub Actions artifact (only enabled by pr-test-npu)'
type: boolean
default: false
only_tests:
description: 'Coverage-recommended test files for this suite (one repo-relative path per line, from recommend-tests-from-coverage). Empty = run the full suite.'
type: string
default: ''
github-token:
description: 'GitHub token for API calls'
type: string
@@ -70,6 +70,7 @@ env:
jobs:
run:
name: ${{ inputs.use_coverage_runner && inputs.self_name || 'run' }} (${{ matrix.partition }})
runs-on: ${{ inputs.runner_config }}
container:
image: ${{ inputs.image }}
@@ -241,6 +242,7 @@ jobs:
env:
CONTINUE_ON_ERROR_FLAG: ${{ inputs.continue_on_error == 'true' && '--continue-on-error' || '' }}
TEST_LOG_DIR: /tmp/test-logs
ONLY_TESTS: ${{ inputs.only_tests }}
shell: bash
run: |
# Fail fast on any command error, undefined variable, or pipe failure.
@@ -284,6 +286,16 @@ jobs:
NIGHTLY_FLAG="--nightly --continue-on-error"
fi
# Precision-test mode: run only the coverage-recommended files for
# this suite. The bucketed list comes from recommend-tests-from-
# coverage; empty input keeps the full-suite behavior.
ONLY_FILE_ARGS=""
if [ -n "${ONLY_TESTS}" ]; then
printf '%s\n' "${ONLY_TESTS}" > /tmp/precision_tests.txt
ONLY_FILE_ARGS="--only-file-list /tmp/precision_tests.txt"
echo "Precision mode: running $(wc -l < /tmp/precision_tests.txt) recommended file(s)"
fi
# Optional: delegate execution to scripts/ci/npu/precise-test/run_tests_with_coverage.sh when the
# caller sets use_coverage_runner=true. list_tests.py (same directory) enumerates the
# selected test files; the original run_suite.py execution below is preserved
@@ -308,13 +320,4 @@ jobs:
--auto-partition-id ${{ matrix.partition }} \
--auto-partition-size ${{ fromJson(inputs.partitions).size }} \
${{ inputs.timeout_per_file && format('--timeout-per-file {0}', inputs.timeout_per_file) || '' }} \
$NIGHTLY_FLAG $CONTINUE_ON_ERROR_FLAG 2>&1 | tee "${LOG_FILE}"
- name: Upload test logs
if: always() && inputs.use_coverage_runner != true && inputs.upload_test_logs == true
uses: actions/upload-artifact@v7
with:
name: selected-test-logs-${{ inputs.self_name }}-part${{ matrix.partition }}
path: /tmp/test-logs/
if-no-files-found: ignore
retention-days: 14
$ONLY_FILE_ARGS $NIGHTLY_FLAG $CONTINUE_ON_ERROR_FLAG 2>&1 | tee "${LOG_FILE}"
@@ -71,10 +71,10 @@ on:
description: 'If true, delegate test execution to scripts/ci/npu/precise-test/run_tests_with_coverage.sh instead of running via run_suite.py directly'
type: boolean
default: false
upload_test_logs:
description: 'If true, upload the test logs as a GitHub Actions artifact (only enabled by pr-test-npu)'
type: boolean
default: false
only_tests:
description: 'Coverage-recommended test files for this suite (one repo-relative path per line, from recommend-tests-from-coverage). Empty = run the full suite.'
type: string
default: ''
env:
SKIP_PR_TEST_HEALTH_CHECK: ${{ inputs.skip_pr_test_health_check }}
@@ -85,7 +85,7 @@ concurrency:
jobs:
e2e:
name: ${{ inputs.test_suite }}
name: ${{ inputs.test_suite }}${{ inputs.use_coverage_runner && format(' ({0})', matrix.partition) || '' }}
strategy:
fail-fast: false
matrix:
@@ -238,6 +238,7 @@ jobs:
SGLANG_IS_IN_CI: true
TRANSFORMERS_VERBOSITY: "error"
GDN_ATTN_BACKEND_TRITON: 1
ONLY_TESTS: ${{ inputs.only_tests }}
shell: bash
run: |
# Fail fast on any command error, undefined variable, or pipe failure.
@@ -401,6 +402,16 @@ jobs:
PARTITION_ARGS="--auto-partition-id ${PARTITION_ID} --auto-partition-size ${PARTITION_SIZE}"
fi
# Precision-test mode: run only the coverage-recommended files for
# this suite. The bucketed list comes from recommend-tests-from-
# coverage; empty input keeps the full-suite behavior.
ONLY_FILE_ARGS=""
if [ -n "${ONLY_TESTS}" ]; then
printf '%s\n' "${ONLY_TESTS}" > /tmp/precision_tests.txt
ONLY_FILE_ARGS="--only-file-list /tmp/precision_tests.txt"
echo "Precision mode: running $(wc -l < /tmp/precision_tests.txt) recommended file(s)"
fi
# Append `--timeout-from-est‑time` when the job belongs to the nightly pipeline.
# per‑case timeout is derived from each case's est_time.
TIMEOUT_FLAG=""
@@ -435,7 +446,7 @@ jobs:
# Run NPU test suite with run_suite.py.
# Capture mode (--enable-retry --max-attempts 1) keeps test subprocesses off the
# tee pipe, so leftover server processes cannot block the pipeline on exit.
${PYTHON_FOR_SGLANG} -u run_suite.py --hw npu --suite ${test_suite} ${NIGHTLY_FLAG} ${PARTITION_ARGS} ${TIMEOUT_FLAG} --enable-retry --max-attempts 1 --timeout-per-file 7200 2>&1 | tee "${LOG_TEE_TARGETS[@]}" || test_exit_code=$?
${PYTHON_FOR_SGLANG} -u run_suite.py --hw npu --suite ${test_suite} ${NIGHTLY_FLAG} ${PARTITION_ARGS} ${ONLY_FILE_ARGS} ${TIMEOUT_FLAG} --enable-retry --max-attempts 1 --timeout-per-file 7200 2>&1 | tee "${LOG_TEE_TARGETS[@]}" || test_exit_code=$?
echo "Finished test: ${tc_name}"
@@ -467,12 +478,3 @@ jobs:
echo "" >> $GITHUB_STEP_SUMMARY
fi
exit ${test_exit_code}
- name: Upload test logs
if: always() && inputs.use_coverage_runner != true && inputs.upload_test_logs == true
uses: actions/upload-artifact@v7
with:
name: selected-test-logs-${{ inputs.test_suite }}-part${{ matrix.partition }}
path: /tmp/test-logs/
if-no-files-found: ignore
retention-days: 14
+2 -1
View File
@@ -95,7 +95,7 @@ jobs:
- "python/setup.py"
- "python/sglang/!(multimodal_gen|kernels)/**/!(*.md)"
# Diffusion-only kernels and tests are left to multimodal_gen and
# jit_kernel; ops/diffusion/common stays since srt imports it via kda_kernels.
# jit_kernel; ops/diffusion/common stays since kda_kernels (an srt path) imports it.
- "python/sglang/kernels/{,ops/,jit/,jit/csrc/}!(*.md)"
- "python/sglang/kernels/!(aot|ops|jit)/**/!(*.md)"
- "python/sglang/kernels/jit/!(csrc)/**/!(*.md)"
@@ -130,6 +130,7 @@ jobs:
simulator:
- ".github/workflows/pr-test-extra.yml"
- ".github/workflows/_pr-test-simulator-cpu.yml"
- ".github/workflows/_pr-test-simulator-gpu.yml"
- "benchmark/simulator/**/!(*.md)"
- "tools/sglang-simulator/**/!(*.md)"
multimodal_gen:
@@ -50,7 +50,7 @@ jobs:
strategy:
matrix:
include:
- python-version: "3.10"
- python-version: "3.11"
cuda-version: "13.0"
name: ${{ inputs.job_display_name }}
steps:
+2 -1
View File
@@ -70,7 +70,7 @@ jobs:
run: |
uv pip install -e "python" --index-strategy unsafe-best-match --prerelease allow
uv pip install pytest
uv pip install -e "tools/sglang-simulator[aic]"
uv pip install "./tools/sglang-simulator[aic]"
- name: Prebuild HiCache native hash extension
timeout-minutes: 5
@@ -91,6 +91,7 @@ jobs:
PYTEST_ADDOPTS: "-s"
TORCH_EXTENSIONS_DIR: ${{ runner.temp }}/torch-extensions
run: |
OMP_NUM_THREADS=2 python3 -m pytest -q tools/sglang-simulator/test/test_train_latency_model.py
python3 -m pytest -q tools/sglang-simulator/test/test_simulation_sglang_runner.py
python3 -m pytest -q tools/sglang-simulator/test/test_simulation_sglang_serving.py
python3 -m pytest -q tools/sglang-simulator/test/test_simulation_cache_hit_ratio.py
@@ -0,0 +1,68 @@
name: PR Test SGLang Simulator (GPU)
on:
workflow_call:
inputs:
check_changes:
description: 'toJson(needs.check-changes.outputs).'
type: string
required: true
caller_inputs:
description: 'toJson(inputs) from the caller workflow.'
type: string
required: true
rust_ext_artifact:
description: 'Artifact of prebuilt Rust extension modules.'
type: string
default: ''
env:
SGLANG_IS_IN_CI: true
SKIP_PR_TEST_HEALTH_CHECK: ${{ (fromJson(inputs.caller_inputs).skip_pr_test_health_check || fromJson(inputs.caller_inputs).test_parallel_dispatch || fromJson(inputs.caller_inputs).run_all_tests) && 'true' || 'false' }}
PR_TEST_BYPASS_MAINTENANCE_ON_MAIN: ${{ github.ref == 'refs/heads/main' && 'true' || 'false' }}
USE_VENV: "1"
jobs:
run:
name: simulator-test-gpu
if: fromJson(inputs.check_changes).main_package == 'true' || fromJson(inputs.check_changes).simulator == 'true' || fromJson(inputs.check_changes).sgl_kernel == 'true'
runs-on: ${{ fromJson(fromJson(inputs.check_changes).runs_on_map)['1-gpu-small'] }}
timeout-minutes: 40
steps:
- uses: actions/checkout@v4
with:
ref: ${{ fromJson(inputs.caller_inputs).git_ref || github.sha }}
- uses: ./.github/actions/check-pr-test-health
- uses: ./.github/actions/check-maintenance
- name: Download kernel wheels
if: fromJson(inputs.check_changes).sgl_kernel == 'true'
uses: actions/download-artifact@v4
with:
path: python/sglang/kernels/aot/dist/
merge-multiple: true
pattern: wheel-python3.11-cuda*
- uses: ./.github/actions/download-rust-ext
with:
artifact_name: ${{ inputs.rust_ext_artifact }}
- name: Install dependencies
timeout-minutes: 20
env:
CUSTOM_BUILD_SGL_KERNEL: ${{ fromJson(inputs.check_changes).sgl_kernel }}
run: bash scripts/ci/cuda/ci_install_dependency.sh
- name: Test baseline collection
timeout-minutes: 10
env:
PYTEST_DISABLE_PLUGIN_AUTOLOAD: "1"
run: |
uv pip install "./tools/sglang-simulator"
python3 -m pytest -q tools/sglang-simulator/test/test_baseline_collection_sglang.py
- name: Cleanup venv
if: always()
run: bash scripts/ci/cuda/ci_cleanup_venv.sh
+6 -3
View File
@@ -17,7 +17,7 @@ on:
type: string
required: true
caller_inputs:
description: 'toJson(inputs) from pr-test.yml. Read via fromJson(...).git_ref / skip_pr_test_health_check / test_parallel_dispatch / run_all_tests.'
description: 'toJson(inputs) from pr-test.yml. Read via fromJson(...).git_ref / skip_pr_test_health_check / test_parallel_dispatch / run_all_tests / skip_cache_save.'
type: string
required: true
partitions:
@@ -110,7 +110,7 @@ jobs:
with:
workspaces: rust
shared-key: "sglang-grpc-cpu"
save-if: ${{ matrix.partition == 0 }}
save-if: ${{ matrix.partition == 0 && !fromJson(inputs.caller_inputs).skip_cache_save }}
# uv pip targets a venv by default; setup-python has no venv — install into that interpreter (see UV_SYSTEM_PYTHON in https://docs.astral.sh/uv/guides/integration/github/)
- name: Install dependencies
@@ -156,9 +156,12 @@ jobs:
python3 run_suite.py --hw cpu --suite ${{ inputs.self_name }} --auto-partition-id ${{ matrix.partition }} --auto-partition-size ${{ fromJson(inputs.partitions)[inputs.self_name].size }} --partition-model-file /tmp/partition-model.json --fork-worker-batch-size 20 $CONTINUE_ON_ERROR_FLAG
# PRs save only what the base lacks; the implicit if: success() keeps a
# half-downloaded cache unpublished.
# half-downloaded cache unpublished. skip_cache_save is for callers that
# dispatch on main but test PR code (run-cpu-ci.yml): a save there would
# publish PR-written files under main's cache scope.
- name: Decide whether to save the HF cache
id: hf_cache_save
if: ${{ !fromJson(inputs.caller_inputs).skip_cache_save }}
run: |
size=$(du -sb "$HF_HOME" 2>/dev/null | cut -f1)
after=${size:-0}
+3 -3
View File
@@ -147,7 +147,7 @@ jobs:
with:
path: python/sglang/kernels/aot/dist/
merge-multiple: true
pattern: wheel-python3.10-cuda*
pattern: wheel-python3.11-cuda*
- name: Download artifacts (v6)
if: ${{ fromJson(inputs.check_changes).sgl_kernel == 'true' && steps.rc.outputs.artifact_version == 'v6' }}
@@ -155,7 +155,7 @@ jobs:
with:
path: python/sglang/kernels/aot/dist/
merge-multiple: true
pattern: wheel-python3.10-cuda*
pattern: wheel-python3.11-cuda*
- uses: ./.github/actions/download-rust-ext
if: ${{ !inputs.skip_prebuilt_rust_ext }}
@@ -216,7 +216,7 @@ jobs:
CONTINUE_ON_ERROR_FLAG: ${{ fromJson(inputs.check_changes).continue_on_error == 'true' && '--continue-on-error' || '' }}
RUNNER_LABELS: ${{ fromJson(fromJson(inputs.check_changes).runs_on_map)[inputs.runner_config] }}
GPU_CONFIG: ${{ inputs.runner_config }}
HF_MUSE_SPARK_TOKEN: ${{ inputs.self_name == 'base-c-test-8-gpu-b300' && secrets.HF_MUSE_SPARK_TOKEN || '' }}
HF_MUSE_SPARK_TOKEN: ${{ (inputs.self_name == 'base-c-test-8-gpu-b300' || inputs.self_name == 'extra-b-test-8-gpu-b300') && secrets.HF_MUSE_SPARK_TOKEN || '' }}
# Read only by test_runai_model_loader.py (nightly 1-gpu); inert elsewhere.
# Left unconditional because an `env:` key cannot be omitted, and an empty
# value would reach the runai streamer as a malformed limit.
+1 -5
View File
@@ -58,11 +58,7 @@ jobs:
- name: Set up Python
uses: actions/setup-python@v5
with:
python-version: '3.10'
- name: Install Python dependencies
run: |
pip install tomli
python-version: '3.12'
- name: Configure Git and branch
run: |
@@ -24,11 +24,7 @@ jobs:
- name: Set up Python
uses: actions/setup-python@v5
with:
python-version: '3.10'
- name: Install Python dependencies
run: |
pip install tomli
python-version: '3.12'
- name: Configure Git and branch
run: |
@@ -22,11 +22,7 @@ jobs:
- name: Set up Python
uses: actions/setup-python@v5
with:
python-version: '3.10'
- name: Install Python dependencies
run: |
pip install tomli
python-version: '3.12'
- name: Check if sync is needed
id: check_sync
@@ -24,11 +24,7 @@ jobs:
- name: Set up Python
uses: actions/setup-python@v5
with:
python-version: '3.10'
- name: Install Python dependencies
run: |
pip install tomli
python-version: '3.12'
- name: Configure Git and branch
run: |
@@ -26,11 +26,7 @@ jobs:
- name: Set up Python
uses: actions/setup-python@v5
with:
python-version: '3.10'
- name: Install Python dependencies
run: |
pip install tomli
python-version: '3.12'
- name: Configure Git and branch
id: set_output
+5 -5
View File
@@ -25,7 +25,7 @@ jobs:
- name: Set up Python
uses: actions/setup-python@v5
with:
python-version: '3.10'
python-version: '3.12'
- name: Generate Summary Report
run: |
@@ -41,7 +41,7 @@ jobs:
- name: Set up Python
uses: actions/setup-python@v5
with:
python-version: '3.10'
python-version: '3.12'
- name: Generate Tests by Folder Report
run: |
@@ -57,7 +57,7 @@ jobs:
- name: Set up Python
uses: actions/setup-python@v5
with:
python-version: '3.10'
python-version: '3.12'
- name: Generate Tests by Suite Report
run: |
@@ -73,7 +73,7 @@ jobs:
- name: Set up Python
uses: actions/setup-python@v5
with:
python-version: '3.10'
python-version: '3.12'
- name: Generate Tests by GPU Count Report
run: |
@@ -184,7 +184,7 @@ jobs:
- name: Set up Python
uses: actions/setup-python@v5
with:
python-version: '3.10'
python-version: '3.12'
- name: Generate JSON Report
run: |
+13 -11
View File
@@ -1,15 +1,17 @@
# Coverage data collection pipeline for the NPU precision test selection.
# Coverage baseline collection for the NPU precision test selection.
# Delegates to pr-test-npu.yml with coverage_mode=true, which switches:
# - Runners to the coverage pool (shared /root/.cache with setup-covstub)
# - use_coverage_runner=true on all test jobs
# - Skips pr-gate, multimodal-gen, recommend, analyze
# - Runs setup-covstub to assemble coverage data and publish it to the
# shared disk at a fixed path (/root/.cache/tests/precise-test/coverage)
# pr-test-npu-finish is the final gate in the called workflow.
#
# Manually triggered only (workflow_dispatch): PRs run the standard
# pr-test-npu.yml instead; this pipeline builds the full coverage baseline
# on demand.
# - Test jobs to the coverage pool with use_coverage_runner=true
# - base-a to a hosted runner (base-a-test-npu-cov, data merged via artifact)
# - Runs the multimodal-gen suites too. Their runner is asked for this
# shard's targets (case-level pytest node ids) and each target is run in
# its own coverage invocation, so diffusion coverage is attributed per
# test case instead of per shard.
# - Skips pr-gate, recommend
# - setup-covstub assembles and publishes the baseline to the shared disk
# at /root/.cache/tests/precise-test/coverage
# pr-test-npu-finish is the final gate in the called workflow. Manually
# triggered only (workflow_dispatch): PRs run the standard pr-test-npu.yml
# instead; this pipeline builds the full coverage baseline on demand.
name: Coverage Collection (NPU)
on:
+7 -7
View File
@@ -24,7 +24,7 @@ on:
default: ''
type: string
kernel_artifact_run_id:
description: "Optional sgl-kernel wheel source: GitHub Actions run ID of a pr-test.yml run on the SAME commit as 'ref' that uploaded wheel-python3.10-cuda13.0. Use only when this branch's torch/ABI change makes the PyPI sglang-kernel wheel incompatible. Find it under Actions > pr-test; artifacts expire after 90 days."
description: "Optional sgl-kernel wheel source: GitHub Actions run ID of a pr-test.yml run on the SAME commit as 'ref' that uploaded wheel-python3.11-cuda13.0. Use only when this branch's torch/ABI change makes the PyPI sglang-kernel wheel incompatible. Find it under Actions > pr-test; artifacts expire after 90 days."
required: false
default: ''
type: string
@@ -204,7 +204,7 @@ jobs:
with:
path: python/sglang/kernels/aot/dist/
merge-multiple: true
name: wheel-python3.10-cuda13.0
name: wheel-python3.11-cuda13.0
run-id: ${{ inputs.kernel_artifact_run_id }}
github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -351,7 +351,7 @@ jobs:
- name: Set up Python
uses: actions/setup-python@v5
with:
python-version: '3.10'
python-version: '3.12'
- name: Verify write access to sgl-project/ci-data-diffusion
if: ${{ inputs.publish }}
@@ -403,7 +403,7 @@ jobs:
with:
path: python/sglang/kernels/aot/dist/
merge-multiple: true
name: wheel-python3.10-cuda13.0
name: wheel-python3.11-cuda13.0
run-id: ${{ inputs.kernel_artifact_run_id }}
github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -479,7 +479,7 @@ jobs:
with:
path: python/sglang/kernels/aot/dist/
merge-multiple: true
name: wheel-python3.10-cuda13.0
name: wheel-python3.11-cuda13.0
run-id: ${{ inputs.kernel_artifact_run_id }}
github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -557,7 +557,7 @@ jobs:
with:
path: python/sglang/kernels/aot/dist/
merge-multiple: true
name: wheel-python3.10-cuda13.0
name: wheel-python3.11-cuda13.0
run-id: ${{ inputs.kernel_artifact_run_id }}
github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -630,7 +630,7 @@ jobs:
with:
path: python/sglang/kernels/aot/dist/
merge-multiple: true
name: wheel-python3.10-cuda13.0
name: wheel-python3.11-cuda13.0
run-id: ${{ inputs.kernel_artifact_run_id }}
github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -1,4 +1,4 @@
name: Nightly Experimental sgl-router Docker Image
name: Nightly sgl-router Docker Image
on:
schedule:
@@ -34,15 +34,18 @@ jobs:
- suite: nightly-stage-c-2-gpu-mi350
runner: linux-mi300-2gpu-sglang
timeout: 180
# Single-request decode of Qwen3.8-27B-FP8 runs at ~20 tok/s on MI300X, so long
# generations outlast the test client's 180 s request timeout.
skip_tests: tests/e2e/sglang/test_session_server_multi_role/test_qwen38small.py
# test_qwen38small: single-request decode of Qwen3.8-27B-FP8 runs at ~20 tok/s on
# MI300X, so long generations outlast the test client's 180 s request timeout.
# test_glm5_2_744b_a40b_5layer_ci: its tilelang indexer kernel needs 98 KiB of LDS
# per block; gfx942 has 64 KiB.
skip_tests: >-
tests/e2e/sglang/test_session_server_multi_role/test_qwen38small.py
tests/e2e/megatron/model_scripts/test_glm5_2_744b_a40b_5layer_ci.py
# The 4- and 8-GPU suites run past 180 min on MI300X; they take ~155 and ~112 min on MI355X.
- suite: nightly-stage-c-4-gpu-mi350
runner: linux-mi300-4gpu-sglang
timeout: 300
# Its tilelang indexer kernel needs 98 KiB of LDS per block; gfx942 has 64 KiB.
skip_tests: tests/e2e/megatron/model_scripts/test_glm5_2_744b_a40b_5layer_ci.py
skip_tests: ""
- suite: nightly-stage-c-8-gpu-mi350
runner: linux-mi300-8gpu-sglang
timeout: 300
@@ -24,15 +24,28 @@ concurrency:
jobs:
test-miles-mi35x:
if: github.repository == 'sgl-project/sglang'
name: test-miles-mi35x (${{ matrix.suite }}, ${{ matrix.runner }}${{ matrix.partition_size && format(', part {0}', matrix.part) || '' }})
strategy:
fail-fast: false
matrix:
# All suites use the 8-GPU MI35X pool. The 2-GPU suite includes the
# GLM-5.2 5-layer test, which takes the linux-mi35x-gpu-2 runner down
# ("runner lost communication") but passes on an 8-GPU runner.
include:
- suite: nightly-stage-c-2-gpu-mi350
runner: linux-mi35x-gpu-2
# Use 8-GPU MI35X pool for 4-GPU suite.
runner: linux-mi35x-gpu-8
# The 4-GPU suite is ~6 h of estimated test time; split it so each
# half fits in the step timeout.
- suite: nightly-stage-c-4-gpu-mi350
runner: linux-mi35x-gpu-8
partition_id: 0
partition_size: 2
part: 1/2
- suite: nightly-stage-c-4-gpu-mi350
runner: linux-mi35x-gpu-8
partition_id: 1
partition_size: 2
part: 2/2
- suite: nightly-stage-c-8-gpu-mi350
runner: linux-mi35x-gpu-8
runs-on: ${{ matrix.runner }}
@@ -88,6 +101,7 @@ jobs:
-e PYTHONPATH=/root/miles:/opt/tilelang \
-e GITHUB_STEP_SUMMARY=/sglang-checkout/github_summary.md \
python3 tests/ci/run_suite.py --hw rocm --suite ${{ matrix.suite }} --cadence nightly \
${{ matrix.partition_size && format('--auto-partition-id {0} --auto-partition-size {1}', matrix.partition_id, matrix.partition_size) || '' }} \
${{ (github.event_name != 'workflow_dispatch' || inputs.continue_on_error) && '--continue-on-error' || '' }} || TEST_EXIT_CODE=$?
cat "${GITHUB_WORKSPACE}/github_summary.md" >> "$GITHUB_STEP_SUMMARY" || true
exit ${TEST_EXIT_CODE}
+10
View File
@@ -1000,6 +1000,11 @@ jobs:
# HF download env vars:
# - HF_HUB_ENABLE_HF_TRANSFER=1: Use faster hf_transfer for downloads (if available)
# - HF_HUB_DISABLE_SYMLINKS_WARNING=1: Suppress symlink warnings
#
# Skips ltx_2_5_diffusion_decoder_2gpus: on mi300 its server has never
# finished warmup inside the 1200s startup deadline, and the harness
# retries that TimeoutError six times, running the shard into the
# step timeout.
docker exec \
-e HF_TOKEN \
-e HUGGING_FACE_HUB_TOKEN \
@@ -1019,6 +1024,7 @@ jobs:
--suite 2-gpu \
--partition-id ${{ matrix.part }} \
--total-partitions 2 \
-k "not ltx_2_5_diffusion_decoder" \
${{ needs.check-changes.outputs.continue_on_error == 'true' && '--continue-on-error' || '' }}
# Post-test diagnostics
@@ -1456,9 +1462,13 @@ jobs:
run: |
docker exec ci_sglang rocm-smi --showtopotype || true
docker exec ci_sglang bash -c 'strings $(python3 -c "import mooncake,os;print(os.path.dirname(mooncake.__file__))")/engine*.so | grep -iE "rdma,hip|Unsupported multi-protocol"' || true
# With Mooncake v0.3.14-rc1, registering the KV cache through HIP dmabuf
# leaves the server's next HIP stream creation failing (out of memory on
# rocm10, segfault on rocm720). Register through ibv_reg_mr instead.
bash scripts/ci/amd/amd_ci_exec.sh \
-e SGLANG_TEST_RDMA_DEVICE="${{ env.SGLANG_TEST_RDMA_DEVICE }}" \
-e MC_LOG_LEVEL=TRACE \
-e MOONCAKE_DISABLE_HIP_DMABUF=1 \
-e GLOG_logtostderr=1 \
-w "/sglang-checkout/test" python3 run_suite.py --hw amd --suite stage-b-test-large-8-gpu-mi35x-disaggregation-amd --timeout-per-file 1800 ${{ needs.check-changes.outputs.continue_on_error == 'true' && '--continue-on-error' || '' }}
+10
View File
@@ -150,6 +150,15 @@ jobs:
rust_ext_artifact: ${{ needs.rust-ext-build.outputs.artifact_name }}
# No `secrets: inherit`: this hosted CPU job has no secret consumer.
simulator-test-gpu:
needs: [check-changes, call-gate, sgl-kernel-build-wheels, rust-ext-build]
if: ${{ !failure() && !cancelled() && needs.check-changes.result == 'success' && (needs.call-gate.result == 'success' || needs.call-gate.result == 'skipped') }}
uses: ./.github/workflows/_pr-test-simulator-gpu.yml
with:
check_changes: ${{ toJson(needs.check-changes.outputs) }}
caller_inputs: ${{ toJson(inputs) }}
rust_ext_artifact: ${{ needs.rust-ext-build.outputs.artifact_name }}
# =============================================== extra-a / extra-b ===============================================
# `name:` keeps the job names wait-for-jobs and check-pr-test-health match on.
extra-test:
@@ -191,6 +200,7 @@ jobs:
sgl-kernel-build-wheels,
rust-ext-build,
simulator-test-cpu,
simulator-test-gpu,
extra-test,
]
if: always()
+1 -1
View File
@@ -99,7 +99,7 @@ jobs:
with:
path: python/sglang/kernels/aot/dist/
merge-multiple: true
pattern: wheel-python3.10-cuda13.0
pattern: wheel-python3.11-cuda13.0
- uses: ./.github/actions/download-rust-ext
with:
+38 -9
View File
@@ -65,7 +65,7 @@ jobs:
- name: Set up Python
uses: actions/setup-python@v5
with:
python-version: '3.10'
python-version: '3.12'
- name: Compute partitions
id: compute
@@ -101,7 +101,7 @@ jobs:
with:
path: python/sglang/kernels/aot/dist/
merge-multiple: true
pattern: wheel-python3.10-cuda*
pattern: wheel-python3.11-cuda*
- name: Install dependencies
timeout-minutes: 20
@@ -168,13 +168,19 @@ jobs:
with:
path: python/sglang/kernels/aot/dist/
merge-multiple: true
pattern: wheel-python3.10-cuda*
pattern: wheel-python3.11-cuda*
- name: Install dependencies
timeout-minutes: 20
run: |
CUSTOM_BUILD_SGL_KERNEL=${{inputs.sgl_kernel}} bash scripts/ci/cuda/ci_install_dependency.sh diffusion
- name: Run SM120-only diffusion unit tests
timeout-minutes: 10
working-directory: python
run: |
python3 -m pytest -v sglang/multimodal_gen/test/unit/test_fp8_fa_sm120_attn.py
- name: Run 5090 diffusion canary tests
timeout-minutes: 120
env:
@@ -233,7 +239,7 @@ jobs:
with:
path: python/sglang/kernels/aot/dist/
merge-multiple: true
pattern: wheel-python3.10-cuda*
pattern: wheel-python3.11-cuda*
- name: Install dependencies
timeout-minutes: 20
@@ -298,7 +304,7 @@ jobs:
with:
path: python/sglang/kernels/aot/dist/
merge-multiple: true
pattern: wheel-python3.10-cuda*
pattern: wheel-python3.11-cuda*
- name: Install dependencies
timeout-minutes: 20
@@ -368,7 +374,7 @@ jobs:
with:
path: python/sglang/kernels/aot/dist/
merge-multiple: true
pattern: wheel-python3.10-cuda*
pattern: wheel-python3.11-cuda*
- name: Install dependencies
timeout-minutes: 20
@@ -413,7 +419,7 @@ jobs:
with:
path: python/sglang/kernels/aot/dist/
merge-multiple: true
pattern: wheel-python3.10-cuda*
pattern: wheel-python3.11-cuda*
- name: Install dependencies
timeout-minutes: 20
@@ -427,6 +433,22 @@ jobs:
run: |
python3 -m pytest -v python/sglang/multimodal_gen/test/single_test_file/test_wan_vae_parallel_groups_4_gpu.py
- name: Run Blackwell-only diffusion unit tests
timeout-minutes: 10
working-directory: python
run: |
python3 -m pytest -v \
sglang/multimodal_gen/test/unit/test_mxfp8_online_gemm.py \
sglang/multimodal_gen/test/unit/test_wan_gelu_mlp.py
- name: Run 4-GPU diffusion unit tests
timeout-minutes: 15
working-directory: python
run: |
python3 -m pytest -v \
sglang/multimodal_gen/test/unit/test_kandinsky6_sequence_parallel.py::test_joint_video_audio_tensor_sequence_parallel \
sglang/multimodal_gen/test/unit/test_kandinsky6_sr_sequence_parallel.py::test_sr_four_gpus
- name: Run diffusion server tests
timeout-minutes: 240
env:
@@ -483,13 +505,20 @@ jobs:
with:
path: python/sglang/kernels/aot/dist/
merge-multiple: true
pattern: wheel-python3.10-cuda*
pattern: wheel-python3.11-cuda*
- name: Install dependencies
timeout-minutes: 20
run: |
CUSTOM_BUILD_SGL_KERNEL=${{inputs.sgl_kernel}} bash scripts/ci/cuda/ci_install_dependency.sh diffusion
- name: Install Quant-VideoGen
timeout-minutes: 5
run: |
# quant-videogen pins an incompatible torch, so install it without deps
# plus the import-time deps of the diffusion-qvg extra.
uv pip install --no-deps quant-videogen==0.1.0 "loguru>=0.7" "termcolor>=2.3"
- name: Run diffusion unit tests
timeout-minutes: 60
run: |
@@ -519,7 +548,7 @@ jobs:
- name: Set up Python
uses: actions/setup-python@v5
with:
python-version: '3.10'
python-version: '3.12'
- name: Download all execution reports
uses: actions/download-artifact@v4
+596 -108
View File
@@ -1,11 +1,13 @@
name: PR Test (NPU)
on:
push:
branches: [ main ]
schedule:
- cron: '0 12 * * *' # Run daily at 12:00 UTC
pull_request:
branches:
- main
- 'npu/**'
- 'npu-*'
workflow_dispatch:
inputs:
coverage_mode:
@@ -169,9 +171,84 @@ jobs:
cd test
python3 run_suite.py --hw npu --suite base-a-test-npu
# Coverage-mode-only counterpart of the base-a PR gate above. Stays on
# ubuntu-latest like the gate (its UTs mock the NPU stack and need a CPU
# triton; the NPU image's triton-ascend breaks those mocks), so the shared
# template (unconditional CANN container:) cannot be used and coverage
# data reaches setup-covstub via artifact instead of the shared disk.
base-a-test-npu-cov:
needs: [set-image-config]
if: ${{ inputs.coverage_mode == true }}
runs-on: ubuntu-latest
timeout-minutes: 45
env:
SGLANG_IS_IN_CI: true
steps:
- name: Checkout code
uses: actions/checkout@v4
with:
# Pin to the event SHA so this job tests the same commit as the rest of the run.
ref: ${{ inputs.ref || github.sha }}
- name: Set up Python
uses: actions/setup-python@v5
with:
python-version: '3.12'
- name: Install uv
uses: astral-sh/setup-uv@v5
with:
enable-cache: true
cache-dependency-glob: 'python/pyproject_npu.toml'
- name: Install dependencies
timeout-minutes: 30
env:
# uv pip targets a venv by default; setup-python has no venv.
UV_SYSTEM_PYTHON: "1"
run: |
# Same dependency set as the base-a PR gate above (CPU torch + CPU
# triton, no Ascend wheels); keep the two blocks in sync.
rm -f python/pyproject.toml
mv python/pyproject_npu.toml python/pyproject.toml
uv pip install torch==2.10.0 torchvision==0.25.0 torchaudio==2.10.0 triton \
--index-url https://download.pytorch.org/whl/cpu \
--extra-index-url https://pypi.org/simple/ \
--index-strategy unsafe-best-match
uv pip install -e "python[dev_npu]" --index-strategy unsafe-best-match
pip install coverage==7.8.*
- name: Run test with coverage
timeout-minutes: 15
run: |
# Hosted runners cannot write the shared disk, so stage locally;
# run_tests_with_coverage.sh derives RUN_DIR/date tag from run-level
# env vars, so the artifact merges 1:1 into the shared-disk layout.
export PRECISE_TEST_CACHE_ROOT=/tmp/precise-test
cd test
python3 ../scripts/ci/npu/precise-test/list_tests.py --hw npu --suite base-a-test-npu \
-o /tmp/selected_tests.txt
echo "Selected test files:"
cat /tmp/selected_tests.txt
if [ -s /tmp/selected_tests.txt ]; then
bash ../scripts/ci/npu/precise-test/run_tests_with_coverage.sh $(cat /tmp/selected_tests.txt)
fi
- name: Upload coverage outputs
if: always()
uses: actions/upload-artifact@v4
with:
name: coverage-outputs-base-a-test-npu
path: /tmp/precise-test/
retention-days: 1
if-no-files-found: warn
base-b-test-1-npu-a3:
needs: [check-changes, pr-gate, set-image-config, base-a-test-npu]
if: ${{ !failure() && !cancelled() && (inputs.coverage_mode == true || needs.check-changes.outputs.main_package == 'true') }}
needs: [check-changes, pr-gate, set-image-config, base-a-test-npu, base-a-test-npu-cov, recommend-tests-from-coverage]
if: >-
${{ !failure() && !cancelled() &&
(inputs.coverage_mode == true || needs.check-changes.outputs.main_package == 'true') &&
needs.recommend-tests-from-coverage.outputs.should_run_base_b_test_1_npu_a3 != 'false' }}
uses: ./.github/workflows/_npu-pr-test-stage.yml
with:
self_name: base-b-test-1-npu-a3
@@ -179,13 +256,16 @@ jobs:
image: ${{ needs.set-image-config.outputs.CANN_image_a3 }}
run_timeout_minutes: '60'
timeout_per_file: '3600'
only_tests: ${{ needs.recommend-tests-from-coverage.outputs.only_tests_base_b_test_1_npu_a3 }}
use_coverage_runner: ${{ inputs.coverage_mode == true }}
upload_test_logs: ${{ inputs.coverage_mode != true }}
secrets: inherit
base-b-test-2-npu-a3:
needs: [check-changes, pr-gate, set-image-config, base-a-test-npu]
if: ${{ !failure() && !cancelled() && (inputs.coverage_mode == true || needs.check-changes.outputs.main_package == 'true') }}
needs: [check-changes, pr-gate, set-image-config, base-a-test-npu, base-a-test-npu-cov, recommend-tests-from-coverage]
if: >-
${{ !failure() && !cancelled() &&
(inputs.coverage_mode == true || needs.check-changes.outputs.main_package == 'true') &&
needs.recommend-tests-from-coverage.outputs.should_run_base_b_test_2_npu_a3 != 'false' }}
uses: ./.github/workflows/_npu-pr-test-stage.yml
with:
self_name: base-b-test-2-npu-a3
@@ -193,13 +273,16 @@ jobs:
image: ${{ needs.set-image-config.outputs.CANN_image_a3 }}
run_timeout_minutes: '60'
timeout_per_file: '3600'
only_tests: ${{ needs.recommend-tests-from-coverage.outputs.only_tests_base_b_test_2_npu_a3 }}
use_coverage_runner: ${{ inputs.coverage_mode == true }}
upload_test_logs: ${{ inputs.coverage_mode != true }}
secrets: inherit
base-b-test-4-npu-a3:
needs: [check-changes, pr-gate, set-image-config, base-a-test-npu]
if: ${{ !failure() && !cancelled() && (inputs.coverage_mode == true || needs.check-changes.outputs.main_package == 'true') }}
needs: [check-changes, pr-gate, set-image-config, base-a-test-npu, base-a-test-npu-cov, recommend-tests-from-coverage]
if: >-
${{ !failure() && !cancelled() &&
(inputs.coverage_mode == true || needs.check-changes.outputs.main_package == 'true') &&
needs.recommend-tests-from-coverage.outputs.should_run_base_b_test_4_npu_a3 != 'false' }}
uses: ./.github/workflows/_npu-pr-test-stage.yml
with:
self_name: base-b-test-4-npu-a3
@@ -208,13 +291,16 @@ jobs:
run_timeout_minutes: '120'
timeout_per_file: '3600'
partitions: '{"size":2,"arr":[0, 1]}'
only_tests: ${{ needs.recommend-tests-from-coverage.outputs.only_tests_base_b_test_4_npu_a3 }}
use_coverage_runner: ${{ inputs.coverage_mode == true }}
upload_test_logs: ${{ inputs.coverage_mode != true }}
secrets: inherit
base-b-test-8-npu-a3:
needs: [check-changes, pr-gate, set-image-config, base-a-test-npu]
if: ${{ !failure() && !cancelled() && (inputs.coverage_mode == true || needs.check-changes.outputs.main_package == 'true') }}
needs: [check-changes, pr-gate, set-image-config, base-a-test-npu, base-a-test-npu-cov, recommend-tests-from-coverage]
if: >-
${{ !failure() && !cancelled() &&
(inputs.coverage_mode == true || needs.check-changes.outputs.main_package == 'true') &&
needs.recommend-tests-from-coverage.outputs.should_run_base_b_test_8_npu_a3 != 'false' }}
uses: ./.github/workflows/_npu-pr-test-stage.yml
with:
self_name: base-b-test-8-npu-a3
@@ -222,13 +308,16 @@ jobs:
image: ${{ needs.set-image-config.outputs.CANN_image_a3 }}
run_timeout_minutes: '60'
timeout_per_file: '3600'
only_tests: ${{ needs.recommend-tests-from-coverage.outputs.only_tests_base_b_test_8_npu_a3 }}
use_coverage_runner: ${{ inputs.coverage_mode == true }}
upload_test_logs: ${{ inputs.coverage_mode != true }}
secrets: inherit
base-b-test-16-npu-a3:
needs: [check-changes, pr-gate, set-image-config, base-a-test-npu]
if: ${{ !failure() && !cancelled() && (inputs.coverage_mode == true || needs.check-changes.outputs.main_package == 'true') }}
needs: [check-changes, pr-gate, set-image-config, base-a-test-npu, base-a-test-npu-cov, recommend-tests-from-coverage]
if: >-
${{ !failure() && !cancelled() &&
(inputs.coverage_mode == true || needs.check-changes.outputs.main_package == 'true') &&
needs.recommend-tests-from-coverage.outputs.should_run_base_b_test_16_npu_a3 != 'false' }}
uses: ./.github/workflows/_npu-pr-test-stage.yml
with:
self_name: base-b-test-16-npu-a3
@@ -236,13 +325,24 @@ jobs:
image: ${{ needs.set-image-config.outputs.CANN_image_a3 }}
run_timeout_minutes: '120'
timeout_per_file: '3600'
only_tests: ${{ needs.recommend-tests-from-coverage.outputs.only_tests_base_b_test_16_npu_a3 }}
use_coverage_runner: ${{ inputs.coverage_mode == true }}
upload_test_logs: ${{ inputs.coverage_mode != true }}
secrets: inherit
multimodal-gen-test-1-npu-a3:
needs: [check-changes, pr-gate, set-image-config, base-a-test-npu]
if: ${{ !failure() && !cancelled() && inputs.coverage_mode != true && needs.check-changes.outputs.multimodal_gen == 'true' }}
needs:
[
check-changes,
pr-gate,
set-image-config,
base-a-test-npu,
base-a-test-npu-cov,
recommend-tests-from-coverage,
]
if: >-
${{ !failure() && !cancelled() &&
(inputs.coverage_mode == true || needs.check-changes.outputs.multimodal_gen == 'true') &&
needs.recommend-tests-from-coverage.outputs.should_run_multimodal_gen_test_1_npu_a3 != 'false' }}
runs-on: linux-aarch64-a3-800t-2
strategy:
fail-fast: false
@@ -290,7 +390,8 @@ jobs:
run: python3 scripts/ci/utils/diffusion/prefetch_input_image.py
- name: Run test
timeout-minutes: 60
timeout-minutes: ${{ inputs.coverage_mode == true && 150 || 60 }}
shell: bash
env:
SGLANG_USE_MODELSCOPE: true
SGLANG_IS_IN_CI: true
@@ -299,12 +400,75 @@ jobs:
PYTORCH_NPU_ALLOC_CONF: "expandable_segments:True"
STREAMS_PER_DEVICE: 32
SGLANG_DIFFUSION_ARTIFACT_DIR: ${{ github.workspace }}/diffusion-failures
ONLY_TESTS: ${{ needs.recommend-tests-from-coverage.outputs.only_tests_multimodal_gen_test_1_npu_a3 }}
run: |
# Coverage mode: enumerate this shard's targets with an NPU-only CI
# helper (it reuses the shared diffusion runner read-only) and run
# each target in its own `coverage run -m pytest` invocation. The
# runner normally batches a whole shard into one pytest call, which
# cannot be attributed per test case by the precise-test selector.
if [[ "${{ inputs.coverage_mode }}" == "true" ]]; then
set -euo pipefail
pip install coverage==7.8.*
# Coverage tracing slows down the Python glue, which shows up in
# wall-clock assertions. Relax only the timing tolerances for the
# collection run; PR runs (no coverage) keep the strict profile and
# the memory/VRAM checks are untouched.
export SGLANG_E2E_TOLERANCE=1.0
export SGLANG_STAGE_TIME_TOLERANCE=1.0
export SGLANG_NON_DENOISE_STAGE_TIME_TOLERANCE=2.0
export SGLANG_DENOISE_STEP_TOLERANCE=1.0
export SGLANG_DENOISE_AGG_TOLERANCE=1.0
python3 scripts/ci/npu/precise-test/list_multimodal_tests.py \
--suite 1-npu \
--partition-id ${{ matrix.part }} \
--total-partitions 2 \
-o /tmp/selected_tests.txt
echo "Selected test targets:"
cat /tmp/selected_tests.txt
if [ -s /tmp/selected_tests.txt ]; then
bash scripts/ci/npu/precise-test/run_tests_with_coverage.sh $(cat /tmp/selected_tests.txt)
fi
exit 0
fi
# Precision mode: ONLY_TESTS holds one pytest -k token per line
# ('[case_id]' for parametrized cases, a file stem for standalone
# files). It is empty when nothing was recommended, precision mode
# is off, or a safety gate forced the full suite; then no filter is
# applied and the shard runs its whole assignment.
FILTER=""
while IFS= read -r line; do
if [ -z "${line}" ]; then
continue
fi
# Every line must be a -k token ('[case_id]' or a standalone file
# stem). A path means the bucket contract drifted: with a path the
# -k expression matches nothing, pytest exits 5, and the shard
# would report success after running zero tests.
if [[ "${line}" == */* || "${line}" == *.py ]]; then
echo "::error::Unexpected only_tests entry '${line}' (expected a pytest -k token)"
exit 1
fi
if [ -n "${FILTER}" ]; then
FILTER="${FILTER} or "
fi
FILTER="${FILTER}${line}"
done <<< "${ONLY_TESTS}"
FILTER_ARGS=()
if [ -n "${FILTER}" ]; then
FILTER_ARGS+=(-k "${FILTER}")
echo "Precision mode: filtered to ${FILTER}"
else
echo "Precision mode: no filter (full suite)"
fi
cd python
python3 sglang/multimodal_gen/test/run_suite.py \
--suite 1-npu \
--partition-id ${{ matrix.part }} \
--total-partitions 2
--total-partitions 2 \
"${FILTER_ARGS[@]}"
- name: Upload diffusion failure artifacts
if: always()
@@ -316,8 +480,19 @@ jobs:
retention-days: 7
multimodal-gen-test-4-npu-a3:
needs: [check-changes, pr-gate, set-image-config, base-a-test-npu]
if: ${{ !failure() && !cancelled() && inputs.coverage_mode != true && needs.check-changes.outputs.multimodal_gen == 'true' }}
needs:
[
check-changes,
pr-gate,
set-image-config,
base-a-test-npu,
base-a-test-npu-cov,
recommend-tests-from-coverage,
]
if: >-
${{ !failure() && !cancelled() &&
(inputs.coverage_mode == true || needs.check-changes.outputs.multimodal_gen == 'true') &&
needs.recommend-tests-from-coverage.outputs.should_run_multimodal_gen_test_4_npu_a3 != 'false' }}
runs-on: linux-aarch64-a3-800t-4
strategy:
fail-fast: false
@@ -360,7 +535,8 @@ jobs:
python3 scripts/ci/utils/diffusion/prefetch_consistency_gt.py
- name: Run test
timeout-minutes: 60
timeout-minutes: ${{ inputs.coverage_mode == true && 150 || 60 }}
shell: bash
env:
SGLANG_USE_MODELSCOPE: true
SGLANG_IS_IN_CI: true
@@ -369,12 +545,75 @@ jobs:
PYTORCH_NPU_ALLOC_CONF: "expandable_segments:True"
STREAMS_PER_DEVICE: 32
SGLANG_DIFFUSION_ARTIFACT_DIR: ${{ github.workspace }}/diffusion-failures
ONLY_TESTS: ${{ needs.recommend-tests-from-coverage.outputs.only_tests_multimodal_gen_test_4_npu_a3 }}
run: |
# Coverage mode: enumerate this shard's targets with an NPU-only CI
# helper (it reuses the shared diffusion runner read-only) and run
# each target in its own `coverage run -m pytest` invocation. The
# runner normally batches a whole shard into one pytest call, which
# cannot be attributed per test case by the precise-test selector.
if [[ "${{ inputs.coverage_mode }}" == "true" ]]; then
set -euo pipefail
pip install coverage==7.8.*
# Coverage tracing slows down the Python glue, which shows up in
# wall-clock assertions. Relax only the timing tolerances for the
# collection run; PR runs (no coverage) keep the strict profile and
# the memory/VRAM checks are untouched.
export SGLANG_E2E_TOLERANCE=1.0
export SGLANG_STAGE_TIME_TOLERANCE=1.0
export SGLANG_NON_DENOISE_STAGE_TIME_TOLERANCE=2.0
export SGLANG_DENOISE_STEP_TOLERANCE=1.0
export SGLANG_DENOISE_AGG_TOLERANCE=1.0
python3 scripts/ci/npu/precise-test/list_multimodal_tests.py \
--suite 2-npu \
--partition-id ${{ matrix.part }} \
--total-partitions 2 \
-o /tmp/selected_tests.txt
echo "Selected test targets:"
cat /tmp/selected_tests.txt
if [ -s /tmp/selected_tests.txt ]; then
bash scripts/ci/npu/precise-test/run_tests_with_coverage.sh $(cat /tmp/selected_tests.txt)
fi
exit 0
fi
# Precision mode: ONLY_TESTS holds one pytest -k token per line
# ('[case_id]' for parametrized cases, a file stem for standalone
# files). It is empty when nothing was recommended, precision mode
# is off, or a safety gate forced the full suite; then no filter is
# applied and the shard runs its whole assignment.
FILTER=""
while IFS= read -r line; do
if [ -z "${line}" ]; then
continue
fi
# Every line must be a -k token ('[case_id]' or a standalone file
# stem). A path means the bucket contract drifted: with a path the
# -k expression matches nothing, pytest exits 5, and the shard
# would report success after running zero tests.
if [[ "${line}" == */* || "${line}" == *.py ]]; then
echo "::error::Unexpected only_tests entry '${line}' (expected a pytest -k token)"
exit 1
fi
if [ -n "${FILTER}" ]; then
FILTER="${FILTER} or "
fi
FILTER="${FILTER}${line}"
done <<< "${ONLY_TESTS}"
FILTER_ARGS=()
if [ -n "${FILTER}" ]; then
FILTER_ARGS+=(-k "${FILTER}")
echo "Precision mode: filtered to ${FILTER}"
else
echo "Precision mode: no filter (full suite)"
fi
cd python
python3 sglang/multimodal_gen/test/run_suite.py \
--suite 2-npu \
--partition-id ${{ matrix.part }} \
--total-partitions 2
--total-partitions 2 \
"${FILTER_ARGS[@]}"
- name: Upload diffusion failure artifacts
if: always()
@@ -387,8 +626,11 @@ jobs:
base-c-test-acc-2-npu-a3:
name: base-c-test-acc-2-npu-a3
needs: [ check-changes, pr-gate, set-image-config, base-b-test-1-npu-a3, base-b-test-2-npu-a3, base-b-test-4-npu-a3, base-b-test-8-npu-a3, base-b-test-16-npu-a3 ]
if: ${{ !failure() && !cancelled() && (inputs.coverage_mode == true || needs.check-changes.outputs.main_package == 'true') }}
needs: [ check-changes, pr-gate, set-image-config, base-b-test-1-npu-a3, base-b-test-2-npu-a3, base-b-test-4-npu-a3, base-b-test-8-npu-a3, base-b-test-16-npu-a3, recommend-tests-from-coverage ]
if: >-
${{ !failure() && !cancelled() &&
(inputs.coverage_mode == true || needs.check-changes.outputs.main_package == 'true') &&
needs.recommend-tests-from-coverage.outputs.should_run_base_c_test_acc_2_npu_a3 != 'false' }}
uses: ./.github/workflows/_npu-single-node-test-stage.yml
with:
runner: ${{ inputs.coverage_mode == true && 'linux-aarch64-a3-800t-2' || 'linux-aarch64-a3-2-' }}
@@ -398,13 +640,16 @@ jobs:
install_sglang_deps: true
device_type_for_deps: 'a3'
partitions: '{"size":2,"arr":[0,1]}'
only_tests: ${{ needs.recommend-tests-from-coverage.outputs.only_tests_base_c_test_acc_2_npu_a3 }}
use_coverage_runner: ${{ inputs.coverage_mode == true }}
upload_test_logs: ${{ inputs.coverage_mode != true }}
base-c-test-acc-16-npu-a3:
name: base-c-test-acc-16-npu-a3
needs: [ check-changes, pr-gate, set-image-config, base-b-test-1-npu-a3, base-b-test-2-npu-a3, base-b-test-4-npu-a3, base-b-test-8-npu-a3, base-b-test-16-npu-a3 ]
if: ${{ !failure() && !cancelled() && (inputs.coverage_mode == true || needs.check-changes.outputs.main_package == 'true') }}
needs: [ check-changes, pr-gate, set-image-config, base-b-test-1-npu-a3, base-b-test-2-npu-a3, base-b-test-4-npu-a3, base-b-test-8-npu-a3, base-b-test-16-npu-a3, recommend-tests-from-coverage ]
if: >-
${{ !failure() && !cancelled() &&
(inputs.coverage_mode == true || needs.check-changes.outputs.main_package == 'true') &&
needs.recommend-tests-from-coverage.outputs.should_run_base_c_test_acc_16_npu_a3 != 'false' }}
uses: ./.github/workflows/_npu-single-node-test-stage.yml
with:
runner: ${{ inputs.coverage_mode == true && 'linux-aarch64-a3-800t-16' || 'linux-aarch64-a3-16-' }}
@@ -413,13 +658,16 @@ jobs:
image: ${{ needs.set-image-config.outputs.CANN_image_a3 }}
install_sglang_deps: true
device_type_for_deps: 'a3'
only_tests: ${{ needs.recommend-tests-from-coverage.outputs.only_tests_base_c_test_acc_16_npu_a3 }}
use_coverage_runner: ${{ inputs.coverage_mode == true }}
upload_test_logs: ${{ inputs.coverage_mode != true }}
base-c-test-perf-2-npu-a3:
name: base-c-test-perf-2-npu-a3
needs: [ check-changes, pr-gate, set-image-config, base-c-test-acc-2-npu-a3, base-c-test-acc-16-npu-a3 ]
if: ${{ !failure() && !cancelled() && (inputs.coverage_mode == true || needs.check-changes.outputs.main_package == 'true') }}
needs: [ check-changes, pr-gate, set-image-config, base-c-test-acc-2-npu-a3, base-c-test-acc-16-npu-a3, recommend-tests-from-coverage ]
if: >-
${{ !failure() && !cancelled() &&
(inputs.coverage_mode == true || needs.check-changes.outputs.main_package == 'true') &&
needs.recommend-tests-from-coverage.outputs.should_run_base_c_test_perf_2_npu_a3 != 'false' }}
uses: ./.github/workflows/_npu-single-node-test-stage.yml
with:
runner: linux-aarch64-a3-800t-2
@@ -428,13 +676,16 @@ jobs:
image: ${{ needs.set-image-config.outputs.CANN_image_a3 }}
install_sglang_deps: true
device_type_for_deps: 'a3'
only_tests: ${{ needs.recommend-tests-from-coverage.outputs.only_tests_base_c_test_perf_2_npu_a3 }}
use_coverage_runner: ${{ inputs.coverage_mode == true }}
upload_test_logs: ${{ inputs.coverage_mode != true }}
base-c-test-perf-16-npu-a3:
name: base-c-test-perf-16-npu-a3
needs: [ check-changes, pr-gate, set-image-config, base-c-test-acc-2-npu-a3, base-c-test-acc-16-npu-a3 ]
if: ${{ !failure() && !cancelled() && (inputs.coverage_mode == true || needs.check-changes.outputs.main_package == 'true') }}
needs: [ check-changes, pr-gate, set-image-config, base-c-test-acc-2-npu-a3, base-c-test-acc-16-npu-a3, recommend-tests-from-coverage ]
if: >-
${{ !failure() && !cancelled() &&
(inputs.coverage_mode == true || needs.check-changes.outputs.main_package == 'true') &&
needs.recommend-tests-from-coverage.outputs.should_run_base_c_test_perf_16_npu_a3 != 'false' }}
uses: ./.github/workflows/_npu-single-node-test-stage.yml
with:
runner: linux-aarch64-a3-800t-16
@@ -444,8 +695,8 @@ jobs:
install_sglang_deps: true
device_type_for_deps: 'a3'
test_timeout_minutes: '180'
only_tests: ${{ needs.recommend-tests-from-coverage.outputs.only_tests_base_c_test_perf_16_npu_a3 }}
use_coverage_runner: ${{ inputs.coverage_mode == true }}
upload_test_logs: ${{ inputs.coverage_mode != true }}
# ==================== Recommend Tests from Coverage ==================== #
@@ -453,25 +704,63 @@ jobs:
# continue-on-error so it never blocks the PR gate.
recommend-tests-from-coverage:
name: Recommend tests from coverage
needs: [check-changes, pr-gate]
runs-on: ubuntu-latest
continue-on-error: true
# TEMP-DISABLED: OBS coverage baseline returns 403; re-enable once fixed.
if: false
# if: ${{ !cancelled() && github.event_name == 'pull_request' && inputs.coverage_mode != true }}
if: >-
${{ !failure() && !cancelled() &&
github.event_name == 'pull_request' &&
inputs.coverage_mode != true &&
needs.check-changes.outputs.changes_exist == 'true' }}
outputs:
coverage_paths: ${{ steps.export-paths.outputs.coverage_paths }}
should_run_base_b_test_1_npu_a3: ${{ steps.bucket-tests.outputs.should_run_base_b_test_1_npu_a3 }}
only_tests_base_b_test_1_npu_a3: ${{ steps.bucket-tests.outputs.only_tests_base_b_test_1_npu_a3 }}
should_run_base_b_test_2_npu_a3: ${{ steps.bucket-tests.outputs.should_run_base_b_test_2_npu_a3 }}
only_tests_base_b_test_2_npu_a3: ${{ steps.bucket-tests.outputs.only_tests_base_b_test_2_npu_a3 }}
should_run_base_b_test_4_npu_a3: ${{ steps.bucket-tests.outputs.should_run_base_b_test_4_npu_a3 }}
only_tests_base_b_test_4_npu_a3: ${{ steps.bucket-tests.outputs.only_tests_base_b_test_4_npu_a3 }}
should_run_base_b_test_8_npu_a3: ${{ steps.bucket-tests.outputs.should_run_base_b_test_8_npu_a3 }}
only_tests_base_b_test_8_npu_a3: ${{ steps.bucket-tests.outputs.only_tests_base_b_test_8_npu_a3 }}
should_run_base_b_test_16_npu_a3: ${{ steps.bucket-tests.outputs.should_run_base_b_test_16_npu_a3 }}
only_tests_base_b_test_16_npu_a3: ${{ steps.bucket-tests.outputs.only_tests_base_b_test_16_npu_a3 }}
should_run_base_c_test_acc_2_npu_a3: ${{ steps.bucket-tests.outputs.should_run_base_c_test_acc_2_npu_a3 }}
only_tests_base_c_test_acc_2_npu_a3: ${{ steps.bucket-tests.outputs.only_tests_base_c_test_acc_2_npu_a3 }}
should_run_base_c_test_acc_16_npu_a3: ${{ steps.bucket-tests.outputs.should_run_base_c_test_acc_16_npu_a3 }}
only_tests_base_c_test_acc_16_npu_a3: ${{ steps.bucket-tests.outputs.only_tests_base_c_test_acc_16_npu_a3 }}
should_run_base_c_test_perf_2_npu_a3: ${{ steps.bucket-tests.outputs.should_run_base_c_test_perf_2_npu_a3 }}
only_tests_base_c_test_perf_2_npu_a3: ${{ steps.bucket-tests.outputs.only_tests_base_c_test_perf_2_npu_a3 }}
should_run_base_c_test_perf_16_npu_a3: ${{ steps.bucket-tests.outputs.should_run_base_c_test_perf_16_npu_a3 }}
only_tests_base_c_test_perf_16_npu_a3: ${{ steps.bucket-tests.outputs.only_tests_base_c_test_perf_16_npu_a3 }}
should_run_multimodal_gen_test_1_npu_a3: ${{ steps.bucket-tests.outputs.should_run_multimodal_gen_test_1_npu_a3 }}
only_tests_multimodal_gen_test_1_npu_a3: ${{ steps.bucket-tests.outputs.only_tests_multimodal_gen_test_1_npu_a3 }}
should_run_multimodal_gen_test_4_npu_a3: ${{ steps.bucket-tests.outputs.should_run_multimodal_gen_test_4_npu_a3 }}
only_tests_multimodal_gen_test_4_npu_a3: ${{ steps.bucket-tests.outputs.only_tests_multimodal_gen_test_4_npu_a3 }}
env:
GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
steps:
# Every step carries step-level continue-on-error so the job conclusion
# is ALWAYS 'success' — the fast-fail health check in other jobs filters
# by job conclusion, and job-level continue-on-error still reports
# 'failure' via the API. Failed steps stay visible (red ✗ + annotation)
# but never cascade to test jobs.
- name: Checkout sglang scripts
continue-on-error: true
uses: actions/checkout@v4
with:
# test/registered + python/sglang/test/ci are needed by
# list_tests.py to enumerate each suite's test files for bucketing.
# The diffusion case config is read (AST only) by
# bucket_recommended_tests.py to enumerate multimodal cases.
sparse-checkout: |
scripts
test/registered
python/sglang/test/ci
python/sglang/multimodal_gen/test/server/ascend/testcase_configs_npu.py
sparse-checkout-cone-mode: false
- name: Download test case map
continue-on-error: true
run: |
set -euo pipefail
curl -fsSL \
@@ -485,6 +774,7 @@ jobs:
du -h test_case_map.json
- name: Download coverage package and extract covstub
continue-on-error: true
run: |
set -euo pipefail
curl -fsSL \
@@ -502,6 +792,7 @@ jobs:
find outputs/covstub -maxdepth 2 -type d | head -20 || true
- name: Recommend tests for current PR
continue-on-error: true
run: |
set -euo pipefail
if ! python3 -c "import regex" >/dev/null 2>&1; then
@@ -520,52 +811,42 @@ jobs:
--source-dir "$PWD/outputs/covstub" \
--map-file "$PWD/test_case_map.json"
- name: Export coverage paths output
id: export-paths
# Bucket the recommended tests per suite. Downstream jobs consume two
# outputs per suite: should_run_<suite> ('false' -> skip the job, e.g.
# empty bucket in precision mode) and only_tests_<suite> (non-empty only
# in precision mode, gated inside the script). On failure outputs stay
# empty -> callers run full suites.
- name: Bucket recommended tests per suite
id: bucket-tests
continue-on-error: true
if: always()
env:
# Precision-test switch: 'true' = the a3 suites run only the
# recommended subset, 'false' = full suites.
PRECISION_MODE: 'true'
run: |
FILE="scripts/ci/npu/precise-test/recommended_pytest_paths.txt"
if [ -f "$FILE" ]; then
{
echo "coverage_paths<<EOF"
cat "$FILE"
echo "EOF"
} >> "$GITHUB_OUTPUT"
echo "Exported $(wc -l < "$FILE") recommended paths"
else
echo "coverage_paths=" >> "$GITHUB_OUTPUT"
echo "::notice::recommended_pytest_paths.txt not found"
fi
set -euo pipefail
python3 scripts/ci/npu/precise-test/bucket_recommended_tests.py \
--recommended scripts/ci/npu/precise-test/recommended_pytest_paths.txt \
--changed-files scripts/ci/npu/precise-test/pr_changed_files.txt \
--suites "base-b-test-1-npu-a3,base-b-test-2-npu-a3,base-b-test-4-npu-a3,base-b-test-8-npu-a3,base-b-test-16-npu-a3,base-c-test-acc-2-npu-a3,base-c-test-acc-16-npu-a3,base-c-test-perf-2-npu-a3,base-c-test-perf-16-npu-a3" \
--diffusion-suites "multimodal-gen-test-1-npu-a3:1-npu,multimodal-gen-test-4-npu-a3:2-npu" \
--case-map "$PWD/test_case_map.json" \
--precision-mode "${PRECISION_MODE}" \
--github-output "${GITHUB_OUTPUT}" \
--github-summary "${GITHUB_STEP_SUMMARY:-}" \
--report-file recommended-tests-report.md
# ==================== Analyze Failure Report ==================== #
# Cross-references test failures with coverage recommendations.
# !cancelled() only: skipped jobs in needs poison success().
analyze-failure-report:
name: Analyze failure report
needs:
[
base-a-test-npu,
base-b-test-1-npu-a3,
base-b-test-2-npu-a3,
base-b-test-4-npu-a3,
base-b-test-8-npu-a3,
base-b-test-16-npu-a3,
multimodal-gen-test-1-npu-a3,
multimodal-gen-test-4-npu-a3,
base-c-test-acc-2-npu-a3,
base-c-test-acc-16-npu-a3,
base-c-test-perf-2-npu-a3,
base-c-test-perf-16-npu-a3,
recommend-tests-from-coverage,
]
if: ${{ !cancelled() && inputs.coverage_mode != true }}
uses: ./.github/workflows/_npu-analyze-failure.yml
with:
log_artifact_pattern: selected-test-logs-*
recommendations_content: ${{ needs.recommend-tests-from-coverage.outputs.coverage_paths || '' }}
# Same report as the step summary, downloadable from the run page.
- name: Upload recommended tests report
continue-on-error: true
if: always()
uses: actions/upload-artifact@v4
with:
name: recommended-tests-report
path: recommended-tests-report.md
if-no-files-found: warn
retention-days: 7
# ==================== Coverage Assembly & Publish ==================== #
# Assembles coverage data and publishes the baseline to the shared disk.
@@ -573,7 +854,7 @@ jobs:
setup-covstub:
needs:
[
set-image-config,
base-a-test-npu-cov,
base-b-test-1-npu-a3,
base-b-test-2-npu-a3,
base-b-test-4-npu-a3,
@@ -583,11 +864,13 @@ jobs:
base-c-test-acc-16-npu-a3,
base-c-test-perf-2-npu-a3,
base-c-test-perf-16-npu-a3,
multimodal-gen-test-1-npu-a3,
multimodal-gen-test-4-npu-a3,
]
if: ${{ !failure() && !cancelled() && inputs.coverage_mode == true }}
runs-on: linux-aarch64-a3-800t-2
runs-on: linux-amd64-cpu-4
container:
image: ${{ needs.set-image-config.outputs.CANN_image_a3 }}
image: swr.ap-southeast-1.myhuaweicloud.com/base_image/ascend-ci/sglang:main-x86
steps:
- name: Checkout code
uses: actions/checkout@v4
@@ -596,12 +879,31 @@ jobs:
# checks out the same commit even if the branch advances mid-run.
ref: ${{ inputs.ref || github.sha }}
- name: Collect base-a coverage artifact
# Runs on ubuntu-latest (no shared disk) — hence the artifact relay.
uses: actions/download-artifact@v4
with:
name: coverage-outputs-base-a-test-npu
path: /tmp/base-a-coverage
- name: Merge base-a coverage into shared run dir
run: |
# Artifact stages {RUN_DIR}/outputs/... — merges 1:1 into the shared layout.
CACHE_ROOT="/root/.cache/tests/precise-test"
mkdir -p "${CACHE_ROOT}"
cp -a /tmp/base-a-coverage/. "${CACHE_ROOT}/"
echo "Merged base-a coverage into ${CACHE_ROOT}:"
find "${CACHE_ROOT}" -maxdepth 3 -type d | head -20
- name: Copy sglang source to covstub
run: |
RUN_DIR="/root/.cache/tests/precise-test/${{ github.run_id }}-attempt-${{ github.run_attempt }}"
COVSTUB="${RUN_DIR}/outputs/covstub"
mkdir -p "${COVSTUB}"
cp -r python/sglang "${COVSTUB}/"
# _version.py is build-generated (git-ignored); the placeholder keeps
# coverage html from failing on a missing recorded source file.
echo '__version__ = "0.0.0"' > "${COVSTUB}/sglang/_version.py"
- name: Build test case map
shell: bash
@@ -609,11 +911,9 @@ jobs:
set -euo pipefail
RUN_DIR="/root/.cache/tests/precise-test/${{ github.run_id }}-attempt-${{ github.run_attempt }}"
# Date tag derivation must match scripts/ci/npu/precise-test/run_tests_with_coverage.sh.
# The custom k8s runner only injects explicitly-set env vars, so
# GITHUB_RUN_STARTED_AT may be absent (set -u would abort); fall back
# to local date exactly like run_tests_with_coverage.sh does, so both
# sides derive the same tag.
# Must match run_tests_with_coverage.sh: GITHUB_RUN_STARTED_AT may be
# unset on the k8s runner (only explicitly-set envs are injected), so
# fall back to local date the same way.
if [ -n "${GITHUB_RUN_STARTED_AT:-}" ]; then
COV_DATE_TAG="${GITHUB_RUN_STARTED_AT:0:10}"
COV_DATE_TAG="${COV_DATE_TAG//-/}"
@@ -622,14 +922,14 @@ jobs:
fi
COV_ROOT="${RUN_DIR}/outputs/sglang@${COV_DATE_TAG}"
COVSTUB="${RUN_DIR}/outputs/covstub"
# Placed outside outputs/ so it is NOT included in outputs.tar.gz.
# Standalone top-level deliverable, never packed inside outputs/.
MAP_FILE="${RUN_DIR}/test_case_map.json"
echo "Coverage dir: ${COV_ROOT}"
echo "Source dir: ${COVSTUB}"
# Coverage data may be absent when all test jobs were skipped
# (e.g. by the changes filter). Skip map building in that case.
# Coverage data may be absent when no test targets were enumerated
# (run_tests_with_coverage.sh never ran). Skip map building then.
if [ ! -d "${COV_ROOT}" ]; then
echo "::warning::Coverage dir not found: ${COV_ROOT}, skip building test case map"
exit 0
@@ -643,11 +943,11 @@ jobs:
fi
if ! python3 -c "import regex" 2>/dev/null; then
pip install regex \
python3 -m pip install regex \
-i http://cache-service.nginx-pypi-cache.svc.cluster.local/pypi/simple \
--trusted-host cache-service.nginx-pypi-cache.svc.cluster.local \
--retries 3 --timeout 60 \
|| pip install regex -i https://mirrors.tuna.tsinghua.edu.cn/pypi/web/simple
|| python3 -m pip install regex -i https://mirrors.tuna.tsinghua.edu.cn/pypi/web/simple
fi
python3 scripts/ci/npu/precise-test/test_selector.py \
@@ -658,10 +958,182 @@ jobs:
test -s "${MAP_FILE}"
du -h "${MAP_FILE}"
- name: Generate HTML coverage report
# Best-effort add-on: continue-on-error keeps the job green and later
# steps running, while a failed generation still shows a red ✗ in the UI.
continue-on-error: true
shell: bash
run: |
set -euo pipefail
RUN_DIR="/root/.cache/tests/precise-test/${{ github.run_id }}-attempt-${{ github.run_attempt }}"
# Date tag derivation must match run_tests_with_coverage.sh.
if [ -n "${GITHUB_RUN_STARTED_AT:-}" ]; then
COV_DATE_TAG="${GITHUB_RUN_STARTED_AT:0:10}"
COV_DATE_TAG="${COV_DATE_TAG//-/}"
else
COV_DATE_TAG="$(date +%Y%m%d)"
fi
COV_ROOT="${RUN_DIR}/outputs/sglang@${COV_DATE_TAG}"
COVSTUB="${RUN_DIR}/outputs/covstub"
# Standalone top-level dir, packaged into htmlcov.tar.gz at publish.
REPORT_DIR="${RUN_DIR}/htmlcov"
# Same skip conditions as map building: no data -> nothing to report.
if [ ! -d "${COV_ROOT}" ] \
|| [ "$(find "${COV_ROOT}" -name 'coverage.*' -type f | wc -l)" -eq 0 ]; then
echo "::warning::No coverage data found, skip HTML report"
exit 0
fi
# Failures below exit non-zero (errexit) — red ✗, job stays green.
# Do NOT convert failures to exit 0 behind an `if ! ( set -e; ... )`
# guard: bash ignores -e inside a subshell in a condition context,
# which silently swallowed three failures in run 38713909169.
# Pin matches the test jobs' coverage==7.8.* so combine can read
# their SQLite data.
if ! python3 -c "import coverage" 2>/dev/null; then
python3 -m pip install coverage==7.8.* \
-i http://cache-service.nginx-pypi-cache.svc.cluster.local/pypi/simple \
--trusted-host cache-service.nginx-pypi-cache.svc.cluster.local \
--retries 3 --timeout 60 \
|| python3 -m pip install coverage==7.8.* -i https://mirrors.tuna.tsinghua.edu.cn/pypi/web/simple
fi
# Alias runner-specific source paths onto covstub; first entry must
# be a real local path (no wildcards).
{
echo "[run]"
echo "branch = True"
echo "[paths]"
echo "sglang ="
echo " ${COVSTUB}/sglang/"
echo " */python/sglang/"
} > /tmp/report-rc
# --keep preserves the per-test raw coverage files: coverage-data.tar.gz
# archives them after this step for manual map rebuilds.
COVERAGE_FILE=/tmp/.coverage-all \
python3 -m coverage combine --rcfile=/tmp/report-rc --keep \
$(find "${COV_ROOT}" -name 'coverage.*' -type f)
rm -rf "${REPORT_DIR}"
COVERAGE_FILE=/tmp/.coverage-all \
python3 -m coverage html --rcfile=/tmp/report-rc \
--directory="${REPORT_DIR}"
# NPU-only view: [run] source does not filter at report time (a
# source-based rcfile yields the full report), so instead touch the
# unexecuted npu .py files into a data copy (os.walk, so the PEP 420
# namespace tree counts) and filter with --include, which works.
cp /tmp/.coverage-all /tmp/.coverage-npu
python3 - "${COVSTUB}" <<'PYEOF'
import os
import sys
from coverage import CoverageData
npu_root = os.path.join(sys.argv[1], "sglang", "srt", "hardware_backend", "npu")
data = CoverageData(basename="/tmp/.coverage-npu")
data.read()
measured = set(data.measured_files())
added = 0
for root, _, files in os.walk(npu_root):
for f in files:
if f.endswith(".py"):
p = os.path.join(root, f)
if p not in measured:
data.touch_file(p)
added += 1
data.write()
print(f"NPU view: {len(measured)} measured, touched {added} unexecuted files")
PYEOF
rm -rf "${REPORT_DIR}-npu"
COVERAGE_FILE=/tmp/.coverage-npu \
python3 -m coverage html --rcfile=/tmp/report-rc \
--include='*/sglang/srt/hardware_backend/npu/*' \
--directory="${REPORT_DIR}-npu"
# Surface the totals directly on the run page.
# GITHUB_STEP_SUMMARY may be unset on the k8s runner pool.
if [ -n "${GITHUB_STEP_SUMMARY:-}" ]; then
{
echo "### Coverage"
echo '```'
echo "python/sglang (measured files only):"
COVERAGE_FILE=/tmp/.coverage-all \
python3 -m coverage report --rcfile=/tmp/report-rc | tail -n 1
echo "python/sglang/srt/hardware_backend/npu (full-tree denominator):"
COVERAGE_FILE=/tmp/.coverage-npu \
python3 -m coverage report --rcfile=/tmp/report-rc \
--include='*/sglang/srt/hardware_backend/npu/*' | tail -n 1
echo '```'
echo "Full line-by-line report: download the \`coverage-html-report\` artifact and open \`htmlcov/index.html\` (NPU-only view: \`htmlcov-npu/index.html\`)."
} >> "${GITHUB_STEP_SUMMARY}"
fi
echo "HTML coverage report: ${REPORT_DIR}"
echo "NPU-only HTML coverage report: ${REPORT_DIR}-npu"
- name: Upload HTML coverage report
# Report upload must never block the baseline publish pipeline.
continue-on-error: true
uses: actions/upload-artifact@v4
with:
name: coverage-html-report
# Two paths -> artifact root becomes the run dir, so the archive
# holds htmlcov/ and htmlcov-npu/ side by side.
path: |
/root/.cache/tests/precise-test/${{ github.run_id }}-attempt-${{ github.run_attempt }}/htmlcov
/root/.cache/tests/precise-test/${{ github.run_id }}-attempt-${{ github.run_attempt }}/htmlcov-npu
if-no-files-found: ignore
retention-days: 14
- name: Archive outputs
run: |
RUN_DIR="/root/.cache/tests/precise-test/${{ github.run_id }}-attempt-${{ github.run_attempt }}"
tar -czf "${RUN_DIR}/outputs.tar.gz" -C "${RUN_DIR}" outputs
# Source-only baseline: recommend mode needs covstub only, and it
# shrinks the package from ~300MB to a few MB.
tar -czf "${RUN_DIR}/outputs.tar.gz" -C "${RUN_DIR}" outputs/covstub
# Raw coverage data for manual map rebuilds: extract both archives
# under one root to reconstruct outputs/. Absent when no data.
# Date tag derivation must match run_tests_with_coverage.sh.
if [ -n "${GITHUB_RUN_STARTED_AT:-}" ]; then
COV_DATE_TAG="${GITHUB_RUN_STARTED_AT:0:10}"
COV_DATE_TAG="${COV_DATE_TAG//-/}"
else
COV_DATE_TAG="$(date +%Y%m%d)"
fi
if [ -d "${RUN_DIR}/outputs/sglang@${COV_DATE_TAG}" ]; then
tar -czf "${RUN_DIR}/coverage-data.tar.gz" -C "${RUN_DIR}" "outputs/sglang@${COV_DATE_TAG}"
fi
- name: Upload outputs and test case map
# Never blocks the publish pipeline; must run before publish mv's
# the files to the fixed PUBLISH_DIR (races with concurrent runs).
continue-on-error: true
uses: actions/upload-artifact@v4
with:
name: precise-test-outputs
# Two paths -> artifact root becomes RUN_DIR, so the archive
# holds outputs.tar.gz and test_case_map.json side by side.
path: |
/root/.cache/tests/precise-test/${{ github.run_id }}-attempt-${{ github.run_attempt }}/outputs.tar.gz
/root/.cache/tests/precise-test/${{ github.run_id }}-attempt-${{ github.run_attempt }}/test_case_map.json
if-no-files-found: ignore
retention-days: 14
- name: Upload raw coverage data
# Separate artifact keeps the ~300MB bulk out of the baseline package;
# must run before publish mv's the file to the fixed PUBLISH_DIR.
continue-on-error: true
uses: actions/upload-artifact@v4
with:
name: precise-test-coverage-data
path: /root/.cache/tests/precise-test/${{ github.run_id }}-attempt-${{ github.run_attempt }}/coverage-data.tar.gz
# Absent when all test jobs were skipped (no coverage data).
if-no-files-found: ignore
retention-days: 14
- name: Publish outputs and map to shared disk
shell: bash
@@ -669,14 +1141,14 @@ jobs:
set -euo pipefail
RUN_DIR="/root/.cache/tests/precise-test/${{ github.run_id }}-attempt-${{ github.run_attempt }}"
# Fixed publish dir with no RUN_DIR layer: other pipelines have a
# different GITHUB_RUN_ID and must be able to read the baseline.
# Must match PUBLISH_DIR in the recommend-tests-from-coverage job.
# Fixed-name copy for manual use; the recommend job's baseline lives
# in OBS (upload outputs.tar.gz + test_case_map.json manually after
# each run). coverage-data.tar.gz has no automatic consumer — no OBS
# copy, only this dir and the precise-test-coverage-data artifact.
PUBLISH_DIR="/root/.cache/tests/precise-test/coverage"
# Skip publish when no coverage data was collected this run (map
# build was skipped), so we never overwrite the historical
# baseline on the shared disk with an incomplete one.
# Skip publish when no coverage data this run — never overwrite
# the historical baseline with an incomplete one.
if [ ! -s "${RUN_DIR}/test_case_map.json" ]; then
echo "::warning::test_case_map.json missing (no coverage data this run), skip publish"
exit 0
@@ -684,14 +1156,29 @@ jobs:
mkdir -p "${PUBLISH_DIR}"
# mv (rename) within the same shared filesystem is atomic, so a
# concurrent reader sees either the old or the new baseline,
# never a partial file.
# Same-filesystem mv is atomic per file, so readers never see a
# partial file (files from two concurrent publishes can interleave).
mv "${RUN_DIR}/outputs.tar.gz" "${PUBLISH_DIR}/outputs.tar.gz"
mv "${RUN_DIR}/test_case_map.json" "${PUBLISH_DIR}/test_case_map.json"
# Guarded: absent in the unlikely case this run collected no
# coverage data while still producing a map.
if [ -f "${RUN_DIR}/coverage-data.tar.gz" ]; then
mv "${RUN_DIR}/coverage-data.tar.gz" "${PUBLISH_DIR}/coverage-data.tar.gz"
fi
# HTML reports too, fixed names — latest wins, like the baseline.
# [ -d ] guards the continue-on-error generation step. A tar failure
# aborts red after the baseline mv completed; Cleanup stays skipped,
# keeping RUN_DIR for recovery.
if [ -d "${RUN_DIR}/htmlcov" ]; then
tar -czf "${PUBLISH_DIR}/htmlcov.tar.gz" -C "${RUN_DIR}" htmlcov
fi
if [ -d "${RUN_DIR}/htmlcov-npu" ]; then
tar -czf "${PUBLISH_DIR}/htmlcov-npu.tar.gz" -C "${RUN_DIR}" htmlcov-npu
fi
echo "Published baseline to ${PUBLISH_DIR}"
du -h "${PUBLISH_DIR}/outputs.tar.gz" "${PUBLISH_DIR}/test_case_map.json"
ls -lh "${PUBLISH_DIR}" | tail -n +2
- name: Clean up per-run directory
# Skipped on publish failure, keeping RUN_DIR for manual recovery.
@@ -708,6 +1195,7 @@ jobs:
check-changes,
base-a-test-npu,
base-a-test-npu-cov,
base-b-test-1-npu-a3,
base-b-test-2-npu-a3,
base-b-test-4-npu-a3,
+167 -53
View File
@@ -9,14 +9,14 @@ on:
workflow_dispatch:
inputs:
run_all_tests:
description: "Run the preflight even if the last commit touched no PPU path"
description: "Build and run both sanity files even without PPU path changes"
required: false
type: boolean
default: true
workflow_call:
inputs:
ref:
description: 'Git ref (branch, tag, or SHA) to test. If not provided, uses the default branch.'
description: "Git ref to test; defaults to the triggering event SHA"
required: false
type: string
default: ''
@@ -27,8 +27,7 @@ on:
default: false
concurrency:
# Newer runs cancel older runs for the same ref; unrelated label events
# use isolated groups to avoid interfering with active runs.
# Unrelated label events must not cancel or replace the active PPU checks.
group: >-
pr-test-ppu-${{
(github.event_name == 'pull_request' &&
@@ -40,11 +39,15 @@ concurrency:
}}
cancel-in-progress: true
defaults:
run:
shell: bash
env:
GIT_CONFIG_NOSYSTEM: "1"
jobs:
# ==================== Check Changes ==================== #
check-changes:
# Give skipped checks from unrelated labels distinct names so they cannot
# replace the corresponding checks from an active run.
name: >-
${{ (github.event_name == 'pull_request' && github.event.action == 'labeled' &&
github.event.label.name != 'run-ci' && github.event.label.name != 'run-ci-ppu') &&
@@ -59,25 +62,28 @@ jobs:
contents: read
pull-requests: read
outputs:
# Scope only; the shared pr-gate checks both required labels.
ppu: ${{ steps.filter.outputs.ppu == 'true' || steps.run-mode.outputs.run_all_tests == 'true' }}
source-sha: ${{ steps.source.outputs.sha }}
steps:
- name: Checkout code
uses: actions/checkout@v4
with:
# github.sha (not github.ref) so a runner that dequeues after main
# advances still validates the SHA that triggered the run.
ref: ${{ inputs.ref || github.sha }}
persist-credentials: false
- name: Pin the source revision for every job
id: source
run: printf 'sha=%s\n' "$(git rev-parse HEAD)" >> "$GITHUB_OUTPUT"
- name: Determine run mode
id: run-mode
env:
RUN_ALL_TESTS: ${{ inputs.run_all_tests }}
run: |
if [[ "${{ inputs.run_all_tests }}" == "true" ]]; then
if [[ "$RUN_ALL_TESTS" == "true" ]]; then
echo "run_all_tests=true" >> "$GITHUB_OUTPUT"
echo "Run mode: ALL TESTS"
else
echo "run_all_tests=false" >> "$GITHUB_OUTPUT"
echo "Run mode: FILTERED (triggered by ${{ github.event_name }})"
fi
- name: Detect file changes
@@ -85,16 +91,18 @@ jobs:
uses: dorny/paths-filter@v3
if: steps.run-mode.outputs.run_all_tests != 'true'
with:
# PPU-specific paths only.
filters: |
ppu:
- ".github/workflows/pr-test-ppu.yml"
- "scripts/ci/ppu/**"
- "python/sglang/srt/platforms/ppu.py"
- "python/sglang/srt/hardware_backend/ppu/**"
- "python/sglang/kernels/aot/csrc/ppu/**"
- "python/sglang/srt/platforms/sail.py"
- "python/sglang/kernels/aot/setup_sail.py"
- "python/sglang/kernels/aot/pyproject_sail.toml"
- "python/sglang/kernels/aot/csrc/common_extension_sail.cc"
- "python/sglang/kernels/jit/utils/compile/sailify.py"
- "test/registered/core/test_basic_sanity_eagle3.py"
- "test/registered/core/test_basic_sanity_dflash.py"
# ==================== PR Gate ==================== #
pr-gate:
name: >-
${{ (github.event_name == 'pull_request' && github.event.action == 'labeled' &&
@@ -108,62 +116,168 @@ jobs:
require-label: run-ci-ppu
secrets: inherit
# ==================== PPU Runner Preflight ==================== #
# Validates the runner contract only; the SRT platform and AOT kernels land
# in later PRs and build their first real test on top of this job.
ppu-preflight:
ppu-build:
name: >-
${{ (github.event_name == 'pull_request' && github.event.action == 'labeled' &&
github.event.label.name != 'run-ci' && github.event.label.name != 'run-ci-ppu') &&
'ignored-label-ppu-preflight' || 'ppu-preflight' }}
'ignored-label-ppu-build' || 'ppu-build' }}
needs: [check-changes, pr-gate]
if: needs.check-changes.outputs.ppu == 'true'
runs-on: 1-ppu-m890p
timeout-minutes: 60
runs-on: ubuntu-24.04
timeout-minutes: 120
permissions:
contents: read
container:
image: download.t-head.cn/docker_release/pytorch:sdk2.2.0-2.13.0-ubuntu24.04-hggcrt3-py312-20261005
env:
SGLANG_IS_IN_CI: true
PPU_CI_SOURCE_SHA: ${{ needs.check-changes.outputs.source-sha }}
MAX_JOBS: "2"
SGL_KERNEL_COMPILE_THREADS: "2"
outputs:
wheel-artifact-id: ${{ steps.wheels.outputs.artifact-id }}
steps:
- name: Checkout code
- name: Initialize private build directories and Git configuration
run: |
work="$(mktemp -d "${RUNNER_TEMP}/ppu-build.XXXXXX")"
touch "$work/gitconfig"
mkdir "$work/wheels" "$work/artifacts"
{
printf 'GIT_CONFIG_GLOBAL=%s\n' "$work/gitconfig"
printf 'PPU_CI_WHEEL_DIR=%s\n' "$work/wheels"
printf 'PPU_CI_ARTIFACT_DIR=%s\n' "$work/artifacts"
} >> "$GITHUB_ENV"
# Same ref as check-changes. build_and_test.sh rejects any HEAD other than
# PPU_CI_SOURCE_SHA, so a workflow_call branch that moves after
# check-changes fails the run instead of building different code.
- name: Checkout the triggering source revision
uses: actions/checkout@v4
with:
ref: ${{ inputs.ref || github.sha }}
persist-credentials: false
- name: Build native SAIL wheels
run: |
mkdir -p "$PPU_CI_WHEEL_DIR"
bash scripts/ci/ppu/build_and_test.sh build
- name: Upload PPU wheels
id: wheels
uses: actions/upload-artifact@v4
with:
name: ppu-wheels-${{ github.run_id }}-${{ github.run_attempt }}
path: ${{ env.PPU_CI_WHEEL_DIR }}
if-no-files-found: error
- name: Upload build diagnostics
if: always()
uses: actions/upload-artifact@v4
with:
name: ppu-build-logs-${{ github.run_id }}-${{ github.run_attempt }}
path: ${{ env.PPU_CI_ARTIFACT_DIR }}
if-no-files-found: warn
ppu-sanity:
name: >-
${{ (github.event_name == 'pull_request' && github.event.action == 'labeled' &&
github.event.label.name != 'run-ci' && github.event.label.name != 'run-ci-ppu') &&
format('ignored-label-ppu-sanity ({0})', matrix.runner) ||
format('ppu-sanity ({0})', matrix.runner) }}
needs: [check-changes, pr-gate, ppu-build]
if: needs.check-changes.outputs.ppu == 'true'
# Both PPU generations reuse the same wheels; each host supplies its own
# runtime image and device, while checkpoints come from the shared NAS.
strategy:
# A failure on one generation must not hide the other generation's result.
fail-fast: false
matrix:
runner: [8-ppu-m890p, 16-ppu-zw810e]
runs-on: ${{ matrix.runner }}
timeout-minutes: 60
permissions:
contents: read
actions: read
env:
PPU_CI_SOURCE_SHA: ${{ needs.check-changes.outputs.source-sha }}
steps:
- name: Initialize private runtime directories and Git configuration
run: |
work="$(mktemp -d "${RUNNER_TEMP}/ppu-runtime.XXXXXX")"
touch "$work/gitconfig"
mkdir "$work/wheels" "$work/artifacts"
{
printf 'GIT_CONFIG_GLOBAL=%s\n' "$work/gitconfig"
printf 'PPU_CI_WHEEL_DIR=%s\n' "$work/wheels"
printf 'PPU_CI_ARTIFACT_DIR=%s\n' "$work/artifacts"
} >> "$GITHUB_ENV"
# Same ref as check-changes; run_test.sh rejects any HEAD other than
# PPU_CI_SOURCE_SHA.
- name: Checkout the triggering source revision
timeout-minutes: 10
uses: actions/checkout@v4
with:
ref: ${{ inputs.ref || github.sha }}
persist-credentials: false
- name: Validate PPU runner contract
timeout-minutes: 10
run: python3 scripts/ci/ppu/check_ppu_environment.py
- name: Download the matching PPU wheel artifact
uses: actions/download-artifact@v4
with:
artifact-ids: ${{ needs.ppu-build.outputs.wheel-artifact-id }}
merge-multiple: true
path: ${{ env.PPU_CI_WHEEL_DIR }}
- name: Run both speculative sanity files in isolated Docker
env:
CONFIG_PPU_CI_DEVICE: ${{ vars.PPU_CI_DEVICE }}
CONFIG_PPU_CI_TARGET_MODEL_PATH: ${{ vars.PPU_CI_TARGET_MODEL_PATH }}
CONFIG_PPU_CI_EAGLE3_MODEL_PATH: ${{ vars.PPU_CI_EAGLE3_MODEL_PATH }}
CONFIG_PPU_CI_DFLASH_MODEL_PATH: ${{ vars.PPU_CI_DFLASH_MODEL_PATH }}
CONFIG_PPU_CI_EVAL_CACHE: ${{ vars.PPU_CI_EVAL_CACHE }}
run: |
for key in PPU_CI_DEVICE PPU_CI_TARGET_MODEL_PATH PPU_CI_EAGLE3_MODEL_PATH PPU_CI_DFLASH_MODEL_PATH PPU_CI_EVAL_CACHE; do
override="CONFIG_$key"
if [[ -n "${!override}" ]]; then
export "$key=${!override}"
fi
done
exec bash scripts/ci/ppu/run_test.sh
- name: Upload runtime diagnostics
if: always()
uses: actions/upload-artifact@v4
with:
name: ppu-runtime-logs-${{ matrix.runner }}-${{ github.run_id }}-${{ github.run_attempt }}
path: ${{ env.PPU_CI_ARTIFACT_DIR }}
if-no-files-found: warn
# ==================== Finish ==================== #
pr-test-ppu-finish:
name: >-
${{ (github.event_name == 'pull_request' && github.event.action == 'labeled' &&
github.event.label.name != 'run-ci' && github.event.label.name != 'run-ci-ppu') &&
'ignored-label-pr-test-ppu-finish' || 'pr-test-ppu-finish' }}
needs: [pr-gate, check-changes, ppu-preflight]
needs: [pr-gate, check-changes, ppu-build, ppu-sanity]
if: always() && needs.check-changes.result != 'skipped'
runs-on: ubuntu-latest
permissions: {}
steps:
- name: Check all dependent job statuses
- name: Require successful native build and both PPU sanity files when requested
env:
NEEDS_JSON: ${{ toJson(needs) }}
PPU_IN_SCOPE: ${{ needs.check-changes.outputs.ppu }}
BUILD_RESULT: ${{ needs.ppu-build.result }}
SANITY_RESULT: ${{ needs.ppu-sanity.result }}
run: |
json_needs='${{ toJson(needs) }}'
for job in $(echo "$json_needs" | jq -r 'keys_unsorted[]'); do
result=$(echo "$json_needs" | jq -r --arg j "$job" '.[$j].result')
echo "$job: $result"
if [[ "$result" == "failure" || "$result" == "cancelled" ]]; then
echo "The above jobs failed."
exit 1
fi
done
echo "All jobs completed successfully"
- name: Require the preflight when PPU is in scope
run: |
in_scope='${{ needs.check-changes.outputs.ppu }}'
preflight='${{ needs.ppu-preflight.result }}'
echo "PPU in scope: '$in_scope' / ppu-preflight: '$preflight'"
if [[ "$in_scope" == "true" && "$preflight" != "success" ]]; then
echo "PPU was in scope but ppu-preflight did not succeed."
if jq -e 'any(.[]; .result == "failure" or .result == "cancelled")' <<< "$NEEDS_JSON" >/dev/null; then
printf '%s\n' '::error::A PPU CI dependency failed or was cancelled.'
exit 1
fi
echo "Preflight status is consistent with the requested scope."
if [[ "$PPU_IN_SCOPE" == "true" ]]; then
if [[ "$BUILD_RESULT" != "success" || "$SANITY_RESULT" != "success" ]]; then
printf '%s\n' '::error::PPU scope requires a successful native build and both PPU sanity files.'
exit 1
fi
printf '%s\n' 'Native build and both PPU sanity files passed.'
else
printf '%s\n' 'PPU is outside this run scope; no PPU tests were requested.'
fi
+4 -4
View File
@@ -56,7 +56,7 @@ jobs:
with:
path: python/sglang/kernels/aot/dist/
merge-multiple: true
pattern: wheel-python3.10-cuda*
pattern: wheel-python3.11-cuda*
- uses: ./.github/actions/download-rust-ext
with:
@@ -95,7 +95,7 @@ jobs:
with:
path: python/sglang/kernels/aot/dist/
merge-multiple: true
pattern: wheel-python3.10-cuda*
pattern: wheel-python3.11-cuda*
- uses: ./.github/actions/download-rust-ext
with:
@@ -146,7 +146,7 @@ jobs:
with:
path: python/sglang/kernels/aot/dist/
merge-multiple: true
pattern: wheel-python3.10-cuda*
pattern: wheel-python3.11-cuda*
- uses: ./.github/actions/download-rust-ext
with:
@@ -181,7 +181,7 @@ jobs:
# with:
# path: python/sglang/kernels/aot/dist/
# merge-multiple: true
# pattern: wheel-python3.10-cuda*
# pattern: wheel-python3.11-cuda*
# - name: Install dependencies
# run: |
+108 -85
View File
@@ -6,10 +6,10 @@ name: PR Test (sgl-router)
# fires on every PR synchronize / push, so the run record always
# appears as a check on the PR (no more "workflow silently didn't
# fire" mystery debugging). Path-based skip decisions are made
# inside the `sgl-router-gate` job below, where we can log the
# inside the `sgl-router-changes` job below, where we can log the
# reason. Same for the `run-ci` label gate — moved off the
# workflow trigger and into the gate job so its outcome is
# observable.
# observable, and now fails (rather than skips) when it is missing.
#
# * `push` on `main` keeps firing so post-merge runs still record
# against the default branch.
@@ -29,21 +29,17 @@ env:
SGLANG_IS_IN_CI: true
jobs:
# Stage 0 — decide whether to run the heavy tiers. Always runs
# (cheap, ubuntu-latest), and emits a single `should_run` output
# consumed by every tier below. Reasons it might evict downstream
# work:
# * `pull_request` event without the `run-ci` label (budget gate)
# * No files matching the sgl-router paths changed in this PR
# Either outcome is logged on the gate job's page, so operators can
# see *why* the tiers were skipped instead of guessing.
sgl-router-gate:
name: gate
# Stage 0a — path detection. Always runs (cheap, ubuntu-latest) and
# emits a single `should_run` output consumed by every tier below.
# It only answers "does this PR touch sgl-router?"; the `run-ci`
# label is checked by the separate gate job so that a missing label
# fails that job instead of silently skipping the tiers (see below).
sgl-router-changes:
name: changes
runs-on: ubuntu-latest
timeout-minutes: 5
outputs:
should_run: ${{ steps.decide.outputs.should_run }}
paths_changed: ${{ steps.paths.outputs.sgl_router }}
has_run_ci_label: ${{ steps.label.outputs.has_run_ci }}
steps:
- uses: actions/checkout@v4
# `dorny/paths-filter` computes the diff against the PR base
@@ -53,53 +49,62 @@ jobs:
# the gate is observable.
- name: Detect sgl-router path changes
id: paths
if: github.event_name != 'workflow_dispatch'
uses: dorny/paths-filter@v3
with:
filters: |
sgl_router:
- 'experimental/sgl-router/**'
- 'sgl-router/**'
- 'rust/Cargo.toml'
- 'rust/sglang-processor/**'
- 'rust/sglang-api-types/**'
- 'proto/**'
- '.github/workflows/pr-test-sgl-router.yml'
- 'docker/sgl-router.Dockerfile'
- 'scripts/ci/cuda/ci_install_dependency.sh'
# `run-ci` label gate. Cheap workflow runs (e.g. docs-only PRs
# that happen to touch the sgl-router dir) still need the
# opt-in label before consuming the H100 / kind tiers below.
# `push` events on main and `workflow_dispatch` skip this gate.
- name: Check run-ci label
id: label
run: |
if [[ "${{ github.event_name }}" != "pull_request" ]]; then
echo "has_run_ci=true" >> "$GITHUB_OUTPUT"
echo "Non-PR event (${{ github.event_name }}); label gate bypassed."
exit 0
fi
if [[ "${{ contains(github.event.pull_request.labels.*.name, 'run-ci') }}" == "true" ]]; then
echo "has_run_ci=true" >> "$GITHUB_OUTPUT"
echo "PR has run-ci label; downstream tiers will run."
else
echo "has_run_ci=false" >> "$GITHUB_OUTPUT"
echo "::warning::PR is missing the 'run-ci' label; skipping sgl-router tiers. Apply the label to opt in."
fi
- name: Decide
id: decide
env:
EVENT_NAME: ${{ github.event_name }}
PATHS_CHANGED: ${{ steps.paths.outputs.sgl_router }}
run: |
paths='${{ steps.paths.outputs.sgl_router }}'
label='${{ steps.label.outputs.has_run_ci }}'
if [[ "$paths" == "true" && "$label" == "true" ]]; then
if [[ "$EVENT_NAME" == "workflow_dispatch" ]]; then
echo "should_run=true" >> "$GITHUB_OUTPUT"
echo "Both path filter and run-ci label match; running tiers."
else
echo "Manual dispatch; running sgl-router tiers."
elif [[ "$PATHS_CHANGED" != "true" ]]; then
echo "should_run=false" >> "$GITHUB_OUTPUT"
echo "Skipping tiers (paths_changed=$paths, has_run_ci=$label)."
echo "No sgl-router-related changes; skipping tiers."
else
echo "should_run=true" >> "$GITHUB_OUTPUT"
echo "sgl-router-related changes detected; running tiers."
fi
# Stage 0b — `run-ci` label gate, shared with `pr-test.yml`. Cheap
# workflow runs (e.g. docs-only PRs that happen to touch the
# sgl-router dir) still need the opt-in label before consuming the
# H100 / kind tiers below. `push` on main and `workflow_dispatch`
# pass straight through.
#
# `pr-gate.yml` fetches the PR's labels live and FAILS when `run-ci`
# is missing. Both matter for `/tag-and-rerun-ci`: the slash handler
# adds the label with GITHUB_TOKEN, which fires no `labeled` event,
# then reruns the PR's unsuccessful runs. A rerun replays the original
# event payload, so a gate reading `github.event.pull_request.labels`
# would skip again, and a run that concluded `success` with every
# tier skipped is never rerun at all. Failing here on the live label
# set makes the rerun both happen and pick the label up.
sgl-router-gate:
name: gate
needs: sgl-router-changes
if: needs.sgl-router-changes.outputs.should_run == 'true'
uses: ./.github/workflows/pr-gate.yml
with:
require-run-ci: true
sgl-router-lint:
name: tier-1 — lint
needs: sgl-router-gate
if: needs.sgl-router-gate.outputs.should_run == 'true'
needs: [sgl-router-changes, sgl-router-gate]
if: needs.sgl-router-changes.outputs.should_run == 'true'
runs-on: ubuntu-latest
env:
# Scoped per-job (not workflow-wide) to avoid SMG's documented breakage
@@ -133,22 +138,22 @@ jobs:
- name: Rust cache
uses: Swatinem/rust-cache@v2
with:
workspaces: experimental/sgl-router
workspaces: sgl-router
shared-key: "sgl-router-cache"
cache-all-crates: true
cache-on-failure: true
save-if: ${{ github.ref == 'refs/heads/main' }}
- name: cargo check
working-directory: experimental/sgl-router
working-directory: sgl-router
run: cargo check --locked --workspace --all-targets
- name: cargo clippy
working-directory: experimental/sgl-router
working-directory: sgl-router
run: cargo clippy --locked --workspace --all-targets -- -D warnings
- name: cargo fmt
working-directory: experimental/sgl-router
working-directory: sgl-router
run: |
rustup toolchain install nightly --profile minimal --component rustfmt
cargo +nightly fmt --all -- --check
@@ -158,7 +163,7 @@ jobs:
run: cargo install --locked cargo-deny@0.19.6
- name: cargo deny check
working-directory: experimental/sgl-router
working-directory: sgl-router
run: cargo deny check licenses bans sources advisories
- name: Show sccache stats
@@ -167,8 +172,8 @@ jobs:
sgl-router-build-and-test:
name: tier-2 — build + test
needs: [sgl-router-gate, sgl-router-lint]
if: needs.sgl-router-gate.outputs.should_run == 'true'
needs: [sgl-router-changes, sgl-router-gate, sgl-router-lint]
if: needs.sgl-router-changes.outputs.should_run == 'true'
runs-on: ubuntu-latest
env:
RUSTC_WRAPPER: sccache
@@ -206,14 +211,14 @@ jobs:
- name: Rust cache
uses: Swatinem/rust-cache@v2
with:
workspaces: experimental/sgl-router
workspaces: sgl-router
shared-key: "sgl-router-cache"
cache-all-crates: true
cache-on-failure: true
save-if: ${{ github.ref == 'refs/heads/main' }}
- name: cargo build (record wall time)
working-directory: experimental/sgl-router
working-directory: sgl-router
run: |
START=$(date +%s)
cargo build --locked --release --all-targets
@@ -223,7 +228,7 @@ jobs:
echo "::notice title=Tier-2 build duration::${BUILD_SECONDS}s"
- name: cargo test
working-directory: experimental/sgl-router
working-directory: sgl-router
# Skip the tokenizer parity_matrix test here: ubuntu-latest has
# no HuggingFace cache, so the matrix would hard-fail (see the
# test docstring). The e2e job runs it after pytest populates
@@ -245,13 +250,13 @@ jobs:
# through to a no-diff "pass" when no fixture was actually
# regenerated.
- name: kv_events hash parity fixture (drift check)
working-directory: experimental/sgl-router
working-directory: sgl-router
run: |
set -euo pipefail
python3 tests/scripts/generate_kv_events_hash_parity.py
if ! git diff --exit-code -- tests/fixtures/kv_events_hash_parity.json; then
echo "::error::kv_events hash parity fixture drifted." \
"Re-run experimental/sgl-router/tests/scripts/generate_kv_events_hash_parity.py" \
"Re-run sgl-router/tests/scripts/generate_kv_events_hash_parity.py" \
"and commit the updated tests/fixtures/kv_events_hash_parity.json."
exit 1
fi
@@ -268,8 +273,8 @@ jobs:
sgl-router-docker-build-test:
name: tier-3 — docker (placeholder)
needs: [sgl-router-gate, sgl-router-build-and-test]
if: needs.sgl-router-gate.outputs.should_run == 'true'
needs: [sgl-router-changes, sgl-router-gate, sgl-router-build-and-test]
if: needs.sgl-router-changes.outputs.should_run == 'true'
runs-on: ubuntu-24.04
timeout-minutes: 15
steps:
@@ -286,8 +291,8 @@ jobs:
# the GPU/cluster runners from spinning up.
sgl-router-k8s-integration:
name: tier-3 — k8s integration
needs: [sgl-router-gate, sgl-router-build-and-test]
if: needs.sgl-router-gate.outputs.should_run == 'true'
needs: [sgl-router-changes, sgl-router-gate, sgl-router-build-and-test]
if: needs.sgl-router-changes.outputs.should_run == 'true'
runs-on: ubuntu-22.04
timeout-minutes: 30
steps:
@@ -303,20 +308,20 @@ jobs:
curl -Lo /tmp/kubectl https://dl.k8s.io/release/v1.31.0/bin/linux/amd64/kubectl
chmod +x /tmp/kubectl && sudo mv /tmp/kubectl /usr/local/bin/kubectl
- name: Build router image
run: docker build -t sgl-router:e2e -f experimental/sgl-router/tests/e2e/k8s_integration/Dockerfile.router .
run: docker build -t sgl-router:e2e -f sgl-router/tests/e2e/k8s_integration/Dockerfile.router .
- name: Build fake_worker image
run: |
docker build -t sgl-router-fake-worker:e2e \
-f experimental/sgl-router/tests/e2e/k8s_integration/Dockerfile.fake_worker \
experimental/sgl-router/tests/e2e/k8s_integration/
-f sgl-router/tests/e2e/k8s_integration/Dockerfile.fake_worker \
sgl-router/tests/e2e/k8s_integration/
- name: Bootstrap kind + deploy
run: bash experimental/sgl-router/tests/e2e/k8s_integration/setup.sh
run: bash sgl-router/tests/e2e/k8s_integration/setup.sh
- name: Set up Python
run: |
python3 -m venv /tmp/e2e-venv
/tmp/e2e-venv/bin/pip install -r experimental/sgl-router/tests/e2e/k8s_integration/requirements.txt
/tmp/e2e-venv/bin/pip install -r sgl-router/tests/e2e/k8s_integration/requirements.txt
- name: Run E2E
run: /tmp/e2e-venv/bin/pytest experimental/sgl-router/tests/e2e/k8s_integration/ -v --tb=short
run: /tmp/e2e-venv/bin/pytest sgl-router/tests/e2e/k8s_integration/ -v --tb=short
- name: Dump cluster + router diagnostics on failure
if: failure()
run: |
@@ -333,8 +338,8 @@ jobs:
sgl-router-e2e:
name: tier-3 — e2e
needs: [sgl-router-gate, sgl-router-build-and-test]
if: needs.sgl-router-gate.outputs.should_run == 'true'
needs: [sgl-router-changes, sgl-router-gate, sgl-router-build-and-test]
if: needs.sgl-router-changes.outputs.should_run == 'true'
runs-on: 2-gpu-h100
timeout-minutes: 45
steps:
@@ -344,10 +349,10 @@ jobs:
- name: Rust cache
uses: Swatinem/rust-cache@v2
with:
workspaces: experimental/sgl-router
workspaces: sgl-router
shared-key: "sgl-router-cache"
- name: cargo build (release)
working-directory: experimental/sgl-router
working-directory: sgl-router
run: |
source "$HOME/.cargo/env"
cargo build --locked --release --workspace
@@ -365,7 +370,7 @@ jobs:
- name: Install e2e test dependencies
run: |
python3 -m pip install -r experimental/sgl-router/tests/e2e/requirements.txt
python3 -m pip install -r sgl-router/tests/e2e/requirements.txt
# IMPORTANT: --ignore=tests/e2e/k8s_integration. The k8s integration
# suite is owned by the `sgl-router-k8s-integration` job which has
@@ -376,7 +381,7 @@ jobs:
# image pull) must not surface as e2e failures on H100, and e2e
# flakes (model load, HF auth) must not surface as k8s failures.
- name: Run e2e
working-directory: experimental/sgl-router
working-directory: sgl-router
env:
HF_TOKEN: ${{ secrets.HF_TOKEN }}
run: |
@@ -384,7 +389,7 @@ jobs:
--ignore=tests/e2e/k8s_integration
- name: Tokenizer parity matrix (uses HF cache populated by e2e)
working-directory: experimental/sgl-router
working-directory: sgl-router
env:
HF_TOKEN: ${{ secrets.HF_TOKEN }}
run: |
@@ -394,25 +399,43 @@ jobs:
sgl-router-finish:
name: finish
needs:
- sgl-router-changes
- sgl-router-gate
- sgl-router-lint
- sgl-router-build-and-test
- sgl-router-docker-build-test
- sgl-router-k8s-integration
- sgl-router-e2e
# `always()` lets `finish` run after upstream `skipped` outcomes
# (when the gate decided to skip). The downstream-skipped path is
# an explicit success — the workflow was correctly bypassed for a
# PR that didn't need full CI. The downstream-failure /
# downstream-cancelled path is a real fail. We also accept the
# gate's `skipped` (impossible today, kept for symmetry) and
# require the gate itself to have succeeded.
if: |
always() &&
needs.sgl-router-gate.result == 'success' &&
!contains(needs.*.result, 'failure') &&
!contains(needs.*.result, 'cancelled')
# `always()` lets `finish` run after upstream `skipped` outcomes.
# Two outcomes are a success: every tier passed, or the PR touches
# no sgl-router path and every tier was skipped. Anything else,
# including a gate that failed for a missing `run-ci` label, fails
# the run, so `/tag-and-rerun-ci` sees an unsuccessful run to rerun.
if: always()
runs-on: ubuntu-latest
timeout-minutes: 5
steps:
- name: All required checks completed
run: echo "OK"
- name: Check required results
env:
CHANGES_RESULT: ${{ needs.sgl-router-changes.result }}
SHOULD_RUN: ${{ needs.sgl-router-changes.outputs.should_run }}
GATE_RESULT: ${{ needs.sgl-router-gate.result }}
LINT_RESULT: ${{ needs.sgl-router-lint.result }}
BUILD_RESULT: ${{ needs.sgl-router-build-and-test.result }}
DOCKER_RESULT: ${{ needs.sgl-router-docker-build-test.result }}
K8S_RESULT: ${{ needs.sgl-router-k8s-integration.result }}
E2E_RESULT: ${{ needs.sgl-router-e2e.result }}
run: |
if [[ "$CHANGES_RESULT" != "success" ]]; then
echo "::error::sgl-router path detection did not succeed ($CHANGES_RESULT)."
exit 1
fi
tiers="$GATE_RESULT $LINT_RESULT $BUILD_RESULT $DOCKER_RESULT $K8S_RESULT $E2E_RESULT"
if [[ "$SHOULD_RUN" == "true" && "$tiers" == "success success success success success success" ]]; then
echo "All sgl-router checks passed."
elif [[ "$SHOULD_RUN" == "false" && "$tiers" == "skipped skipped skipped skipped skipped skipped" ]]; then
echo "No sgl-router-related changes; tiers skipped."
else
echo "::error::Unexpected sgl-router results: should_run=$SHOULD_RUN, gate=$GATE_RESULT, lint=$LINT_RESULT, build=$BUILD_RESULT, docker=$DOCKER_RESULT, k8s=$K8S_RESULT, e2e=$E2E_RESULT."
exit 1
fi
+66 -43
View File
@@ -93,31 +93,27 @@ jobs:
suite: "stage-a-tp-test-cpu-intel"
name: gnr
# Three SPR boxes each run ONE full-machine container taking one
# partition of a 3-way split over the merged non-TP pool
# (base-b non-TP + base-c). Single container per box (no socket
# split): sglang auto-computes mem_fraction_static from host memory
# and does not honor cgroup limits, so two parallel containers on
# one box exhaust host RAM and OOM (tracked upstream). One container
# per box keeps each server within the box's memory budget.
# Each SPR box (2 sockets / 2 NUMA, 32 physical cores each) runs TWO
# socket-pinned containers in parallel, each taking one partition of a
# 4-way split over the merged non-TP pool. Each container is capped
# with --memory so sglang (which reads cgroup memory.max, PR #38260 +
# the empty-NUMA-node fix) sizes its KV pool against one socket's
# budget, not the whole host -- so two parallel servers no longer
# exhaust host RAM. 2 boxes x 2 sockets = 4 partitions.
- runner: xeon-spr
role: spr
suite: "stage-a-test-cpu-intel,stage-b-test-cpu-intel"
name: spr1
part_id: 0
part_size: 3
part_lo: 0
part_hi: 1
part_size: 4
- runner: xeon-spr
role: spr
suite: "stage-a-test-cpu-intel,stage-b-test-cpu-intel"
name: spr2
part_id: 1
part_size: 3
- runner: xeon-spr
role: spr
suite: "stage-a-test-cpu-intel,stage-b-test-cpu-intel"
name: spr3
part_id: 2
part_size: 3
part_lo: 2
part_hi: 3
part_size: 4
steps:
- name: Cleanup workspace
run: |
@@ -169,43 +165,70 @@ jobs:
timeout-minutes: 120
run: |
docker exec -w /sglang-checkout/ ci_sglang_${{ matrix.name }} \
bash -c "source /opt/.venv/bin/activate && cd ./test && python3 run_suite.py --hw cpu --suite ${{ matrix.suite }} --timeout-from-est-time"
bash -c "source /opt/.venv/bin/activate && hf auth login --token \$HF_TOKEN && cd ./test && python3 run_suite.py --hw cpu --suite ${{ matrix.suite }} --timeout-from-est-time"
# ============ SPR: single full-machine container, one partition ====== #
- name: Run container (SPR)
if: matrix.role == 'spr'
run: |
docker rm -f ci_sglang_${{ matrix.name }} || true
docker run -dt \
-v ${{ github.workspace }}:/sglang-checkout/ --shm-size=64g \
-v $HOME/.cache/huggingface:/root/.cache/huggingface \
-v $HOME/.cache/sgl_eval:/root/.cache/sgl_eval \
-e HF_TOKEN="$(cat ~/huggingface_token.txt)" \
--name ci_sglang_${{ matrix.name }} \
sglang_xeon
- name: Check AMX support (SPR)
if: matrix.role == 'spr'
timeout-minutes: 5
run: |
docker exec -w /sglang-checkout/ ci_sglang_${{ matrix.name }} \
bash -c "source /opt/.venv/bin/activate && python3 -c 'import torch; import sgl_kernel; assert torch.cpu._is_amx_tile_supported(); assert hasattr(torch.ops.sgl_kernel, \"convert_weight_packed\"); '"
- name: Run unit tests (SPR)
# ========= SPR: two socket-pinned containers, run in parallel ======== #
- name: Run unit tests (SPR, two sockets in parallel)
if: matrix.role == 'spr'
timeout-minutes: 120
run: |
docker exec -w /sglang-checkout/ ci_sglang_${{ matrix.name }} \
bash -c "source /opt/.venv/bin/activate && cd ./test && python3 run_suite.py --hw cpu --suite ${{ matrix.suite }} --auto-partition-id ${{ matrix.part_id }} --auto-partition-size ${{ matrix.part_size }} --timeout-from-est-time"
set -uo pipefail
HF_TOKEN_VAL="$(cat ~/huggingface_token.txt)"
# One socket-pinned container. Physical cores only (0-31 = socket0,
# 32-63 = socket1); logical/HT siblings (64-127) are never bound.
# --memory caps the cgroup so sglang sizes KV against one socket, and
# --cpuset-mems keeps allocations on the local memory bank.
# $1 socket id, $2 cpuset-cpus, $3 cpuset-mems, $4 partition id
run_socket () {
local sid=$1 cpus=$2 mems=$3 pid=$4
local cname=ci_sglang_${{ matrix.name }}_s${sid}
docker rm -f "$cname" >/dev/null 2>&1 || true
docker run -dt \
--cpuset-cpus="$cpus" --cpuset-mems="$mems" --memory=240g \
-v ${{ github.workspace }}:/sglang-checkout/ --shm-size=32g \
-v $HOME/.cache/huggingface:/root/.cache/huggingface \
-v $HOME/.cache/sgl_eval:/root/.cache/sgl_eval \
-e HF_TOKEN="$HF_TOKEN_VAL" \
-e SGLANG_CPU_OMP_THREADS_BIND="$cpus" \
--name "$cname" \
sglang_xeon
docker exec -w /sglang-checkout/ "$cname" bash -c \
"source /opt/.venv/bin/activate && python3 -c 'import torch; import sgl_kernel; assert torch.cpu._is_amx_tile_supported(); assert hasattr(torch.ops.sgl_kernel, \"convert_weight_packed\")' && \
hf auth login --token \$HF_TOKEN && \
cd ./test && python3 run_suite.py --hw cpu --suite ${{ matrix.suite }} --auto-partition-id $pid --auto-partition-size ${{ matrix.part_size }} --timeout-from-est-time"
}
# socket0 -> part_lo, socket1 -> part_hi; run concurrently, wait both.
run_socket 0 "0-31" 0 ${{ matrix.part_lo }} > s0.log 2>&1 &
P0=$!
run_socket 1 "32-63" 1 ${{ matrix.part_hi }} > s1.log 2>&1 &
P1=$!
rc=0
wait $P0 || rc=1
wait $P1 || rc=1
echo "==================== socket0 (partition ${{ matrix.part_lo }}) ===================="
cat s0.log || true
echo "==================== socket1 (partition ${{ matrix.part_hi }}) ===================="
cat s1.log || true
exit $rc
# ============================ Common ================================= #
- name: Change permission
if: always()
timeout-minutes: 2
run: |
docker exec -u root ci_sglang_${{ matrix.name }} bash -c "rm -rf /tmp/ci-home && chown -R $(id -u):$(id -g) /sglang-checkout/ 2>/dev/null || true" 2>/dev/null || true
# GNR has one container (ci_sglang_<name>); SPR has two socket
# containers (ci_sglang_<name>_s0/_s1). Chown from whichever exist.
for c in ci_sglang_${{ matrix.name }} ci_sglang_${{ matrix.name }}_s0 ci_sglang_${{ matrix.name }}_s1; do
docker exec -u root "$c" bash -c "rm -rf /tmp/ci-home && chown -R $(id -u):$(id -g) /sglang-checkout/ 2>/dev/null || true" 2>/dev/null || true
done
- name: Cleanup container
if: always()
run: |
docker rm -f ci_sglang_${{ matrix.name }} || true
for c in ci_sglang_${{ matrix.name }} ci_sglang_${{ matrix.name }}_s0 ci_sglang_${{ matrix.name }}_s1; do
docker rm -f "$c" || true
done
+1
View File
@@ -446,6 +446,7 @@ jobs:
fail-fast: false
matrix:
include:
- { runner_config: 2-gpu-gb300, run_timeout: '30', timeout_per_file: '1800' }
- { runner_config: 4-gpu-gb300, run_timeout: '30', timeout_per_file: '1800' }
uses: ./.github/workflows/_pr-test-stage.yml
with:
@@ -2,10 +2,20 @@ name: Release Docker Images Nightly gfx1151 (AMD)
on:
workflow_dispatch:
inputs:
push:
description: "Push the image to Docker Hub (uncheck to only check that the branch builds)"
required: false
type: boolean
default: true
schedule:
# Stagger this build from the CDNA ROCm nightlies at 12:00 UTC.
- cron: "0 14 * * *"
# Scheduled runs always push; a manual dispatch can turn the push off.
env:
PUSH_IMAGES: ${{ github.event_name != 'workflow_dispatch' || inputs.push }}
concurrency:
group: ${{ github.workflow }}-${{ github.ref }}
cancel-in-progress: true
@@ -59,5 +69,9 @@ jobs:
--no-cache \
--progress=plain \
-t "rocm/sgl-dev:${IMAGE_TAG}"
docker push "rocm/sgl-dev:${IMAGE_TAG}"
echo "Published rocm/sgl-dev:${IMAGE_TAG}" >> "$GITHUB_STEP_SUMMARY"
if [ "${PUSH_IMAGES}" = "true" ]; then
docker push "rocm/sgl-dev:${IMAGE_TAG}"
echo "Published rocm/sgl-dev:${IMAGE_TAG}" >> "$GITHUB_STEP_SUMMARY"
else
echo "Build-only run: not pushing rocm/sgl-dev:${IMAGE_TAG}" >> "$GITHUB_STEP_SUMMARY"
fi
@@ -21,8 +21,17 @@ on:
- 'all'
- gfx942-rocm1000
- gfx950-rocm1000
push:
description: 'Push the images to Docker Hub (uncheck to only check that the branch builds)'
required: false
type: boolean
default: true
schedule:
- cron: '0 12 * * *'
- cron: '0 10 * * *' # before the ROCm 7.2 build, so it is not queued behind it
# Scheduled runs always push; a manual dispatch can turn the push off.
env:
PUSH_IMAGES: ${{ github.event_name != 'workflow_dispatch' || inputs.push }}
concurrency:
group: ${{ github.workflow }}-${{ github.event.number || github.sha }}
@@ -106,15 +115,21 @@ jobs:
--build-arg UBUNTU_MIRROR=https://archive.ubuntu.com \
-t "rocm/sgl-dev:${image_tag}" \
--no-cache
docker push "rocm/sgl-dev:${image_tag}"
if [ "${PUSH_IMAGES}" = "true" ]; then
docker push "rocm/sgl-dev:${image_tag}"
else
echo "Build-only run: not pushing rocm/sgl-dev:${image_tag}"
fi
- name: Login to Docker Hub (lmsys)
if: env.PUSH_IMAGES == 'true'
uses: docker/login-action@v3
with:
username: ${{ secrets.DOCKERHUB_USERNAME }}
password: ${{ secrets.DOCKERHUB_TOKEN }}
- name: Push to lmsysorg/sglang-rocm
if: env.PUSH_IMAGES == 'true'
run: |
docker tag "rocm/sgl-dev:${IMAGE_TAG}" "lmsysorg/sglang-rocm:${IMAGE_TAG}"
docker push "lmsysorg/sglang-rocm:${IMAGE_TAG}"
@@ -178,15 +193,21 @@ jobs:
--build-arg UBUNTU_MIRROR=https://archive.ubuntu.com \
-t "rocm/sgl-dev:${image_tag}" \
--no-cache
docker push "rocm/sgl-dev:${image_tag}"
if [ "${PUSH_IMAGES}" = "true" ]; then
docker push "rocm/sgl-dev:${image_tag}"
else
echo "Build-only run: not pushing rocm/sgl-dev:${image_tag}"
fi
- name: Login to Docker Hub (lmsys)
if: env.PUSH_IMAGES == 'true'
uses: docker/login-action@v3
with:
username: ${{ secrets.DOCKERHUB_USERNAME }}
password: ${{ secrets.DOCKERHUB_TOKEN }}
- name: Push to lmsysorg/sglang-rocm
if: env.PUSH_IMAGES == 'true'
run: |
docker tag "rocm/sgl-dev:${IMAGE_TAG}" "lmsysorg/sglang-rocm:${IMAGE_TAG}"
docker push "lmsysorg/sglang-rocm:${IMAGE_TAG}"
@@ -253,15 +274,21 @@ jobs:
--build-arg UBUNTU_MIRROR=https://archive.ubuntu.com \
-t "rocm/sgl-dev:${image_tag}" \
--no-cache
docker push "rocm/sgl-dev:${image_tag}"
if [ "${PUSH_IMAGES}" = "true" ]; then
docker push "rocm/sgl-dev:${image_tag}"
else
echo "Build-only run: not pushing rocm/sgl-dev:${image_tag}"
fi
- name: Login to Docker Hub (lmsys)
if: env.PUSH_IMAGES == 'true'
uses: docker/login-action@v3
with:
username: ${{ secrets.DOCKERHUB_USERNAME }}
password: ${{ secrets.DOCKERHUB_TOKEN }}
- name: Push to lmsysorg/sglang-rocm
if: env.PUSH_IMAGES == 'true'
run: |
docker tag "rocm/sgl-dev:${IMAGE_TAG}" "lmsysorg/sglang-rocm:${IMAGE_TAG}"
docker push "lmsysorg/sglang-rocm:${IMAGE_TAG}"
@@ -21,8 +21,17 @@ on:
- gfx950-rocm724
- gfx942-rocm720
- gfx950-rocm720
push:
description: 'Push the images to Docker Hub (uncheck to only check that the branch builds)'
required: false
type: boolean
default: true
schedule:
- cron: '0 12 * * *'
- cron: '0 11 * * *' # an hour after ROCm 10, ahead of the 12:00 UTC MILES builds
# Scheduled runs always push; a manual dispatch can turn the push off.
env:
PUSH_IMAGES: ${{ github.event_name != 'workflow_dispatch' || inputs.push }}
concurrency:
# A PR number if a pull request and otherwise the commit hash. This cancels
@@ -112,18 +121,24 @@ jobs:
# UBUNTU_MIRROR forces apt over HTTPS to dodge port-80 reachability flakes
# to Canonical's archive.ubuntu.com mirror IPs from the amd-docker-scale runner.
docker build . -f docker/rocm.Dockerfile --build-arg SGL_BRANCH=${{ github.ref_name }} --build-arg BUILD_TYPE=${{ matrix.build_type }} --build-arg GPU_ARCH=${{ matrix.gpu_arch }} --build-arg ENABLE_MORI=1 --build-arg ENABLE_NIXL=1 --build-arg SETUPTOOLS_SCM_PRETEND_VERSION=${pretend_version} --build-arg UBUNTU_MIRROR=https://archive.ubuntu.com -t rocm/sgl-dev:${tag}-${{ env.DATE }} --no-cache
docker push rocm/sgl-dev:${tag}-${{ env.DATE }}
if [ "${PUSH_IMAGES}" = "true" ]; then
docker push rocm/sgl-dev:${tag}-${{ env.DATE }}
else
echo "Build-only run: not pushing rocm/sgl-dev:${tag}-${{ env.DATE }}"
fi
# Persist the tag right after rocm/sgl-dev push succeeds so the local
# registry mirror can run even if a later step in this job (lmsys push)
# fails. By default this step only runs when the previous step succeeded,
# so the artifact only exists when an image actually landed on Docker Hub.
- name: Save published image tag
if: env.PUSH_IMAGES == 'true'
run: |
mkdir -p image-tag
echo "${{ env.IMAGE_TAG }}" > "image-tag/${{ matrix.gpu_arch }}.txt"
- name: Upload image tag artifact
if: env.PUSH_IMAGES == 'true'
uses: actions/upload-artifact@v4
with:
name: image-tag-${{ matrix.gpu_arch }}
@@ -131,12 +146,14 @@ jobs:
retention-days: 1
- name: Login to Docker Hub (lmsys)
if: env.PUSH_IMAGES == 'true'
uses: docker/login-action@v2
with:
username: ${{ secrets.DOCKERHUB_USERNAME }}
password: ${{ secrets.DOCKERHUB_TOKEN }}
- name: Push to lmsysorg/sglang-rocm
if: env.PUSH_IMAGES == 'true'
run: |
docker tag rocm/sgl-dev:${{ env.IMAGE_TAG }} lmsysorg/sglang-rocm:${{ env.IMAGE_TAG }}
docker push lmsysorg/sglang-rocm:${{ env.IMAGE_TAG }}
@@ -57,6 +57,7 @@ jobs:
kernel_source=wheel
kernel_ref=main
latest_tag=latest
image_tag="${image_tag}-main-whl"
else
# Pin the kernel commit so the tag names exactly what was built.
kernel_ref=$(git ls-remote https://github.com/sgl-project/sgl-kernel-xpu.git refs/heads/main | cut -f1)
@@ -66,7 +67,7 @@ jobs:
fi
kernel_source=source
latest_tag=latest-kernel-main
image_tag="${image_tag}-kernel-main-${kernel_ref:0:7}"
image_tag="${image_tag}-main-main-${kernel_ref:0:7}"
fi
echo "DATE=${date_tag}" >> $GITHUB_ENV
echo "commit_hash=${commit_hash}" >> $GITHUB_OUTPUT
@@ -83,7 +83,7 @@ jobs:
provenance: false
sbom: false
build-args: |
SGLANG_KERNEL_NPU_TAG=2026.9.0.post9
SGLANG_KERNEL_NPU_TAG=2026.9.0.post10
CANN_VERSION=${{ matrix.cann_version }}
DEVICE_TYPE=${{ matrix.device_type }}
+1 -1
View File
@@ -87,7 +87,7 @@ jobs:
push: ${{ github.repository == 'sgl-project/sglang' && github.event_name != 'pull_request' }}
provenance: false
build-args: |
SGLANG_KERNEL_NPU_TAG=2026.9.0.post9
SGLANG_KERNEL_NPU_TAG=2026.9.0.post10
CANN_VERSION=${{ matrix.cann_version }}
DEVICE_TYPE=${{ matrix.device_type }}
SGLANG_TAG=${{ steps.version.outputs.version }}
+1 -1
View File
@@ -161,7 +161,7 @@ jobs:
- name: Set up Python
uses: actions/setup-python@v5
with:
python-version: "3.10"
python-version: "3.12"
- name: Update wheel index
run: |
+1 -1
View File
@@ -170,7 +170,7 @@ jobs:
- name: Set up Python
uses: actions/setup-python@v5
with:
python-version: "3.10"
python-version: "3.12"
- name: Update wheel index
run: |
+1 -1
View File
@@ -104,7 +104,7 @@ jobs:
- name: Set up Python
uses: actions/setup-python@v5
with:
python-version: "3.10"
python-version: "3.12"
- name: Upload to pypi
run: |
+1 -1
View File
@@ -26,7 +26,7 @@ jobs:
strategy:
fail-fast: false
matrix:
python-version: ["3.10", "3.11", "3.12", "3.13"]
python-version: ["3.11", "3.12", "3.13"]
arch: [x86_64, aarch64]
include:
- arch: x86_64
@@ -168,7 +168,6 @@ jobs:
# so bind eagerly to fail here on any missing symbol.
LD_BIND_NOW=1 python3 - <<'PY'
import ctypes
import sys
import torch
import deep_ep
import deep_gemm
@@ -179,9 +178,6 @@ jobs:
assert ctypes.CDLL("libnccl.so.2").ncclGetVersion(ctypes.byref(nccl_version)) == 0
assert nccl_version.value == 23007, nccl_version.value
assert hasattr(deep_ep, "ElasticBuffer"), "DeepEP ElasticBuffer is required"
sys.path.insert(0, "DeepGEMM/third-party")
from tilelang_ops import ref_mhc
assert ref_mhc.has_baseline(), "Mega mHC reference operators are unavailable"
print("DeepGEMM test dependencies ready; NCCL:", nccl_version.value)
PY
+252
View File
@@ -0,0 +1,252 @@
name: Release sgl-flash-mla
on:
workflow_dispatch:
inputs:
version:
description: "Wheel version (e.g. 1.0.0, 1.0.1rc0)"
type: string
required: true
branch:
description: "FlashMLA branch to build from (default: sgl)"
type: string
required: false
default: 'sgl'
validation_only:
description: "Build and test without publishing wheels"
type: boolean
required: false
default: false
skip_tests:
description: "Skip the GPU test suite (publishes untested wheels)"
type: boolean
required: false
default: false
concurrency:
group: release-sgl-flash-mla-${{ github.ref }}
cancel-in-progress: true
env:
# The wheel pins this; it is also what the test jobs install.
# The extensions use PyTorch's stable C++ ABI and would load against any 2.10+,
# but a release is tested against one version.
TORCH_VER: "2.14.1"
jobs:
build:
if: github.repository == 'sgl-project/sglang'
strategy:
matrix:
# One wheel per architecture. No Python matrix: the wheel is abi3
# (cp310-abi3) and serves CPython 3.10+, so 3.12 only drives the build.
python-version: ["3.12"]
cuda-version: ["13.4"]
arch: [x86_64, aarch64]
include:
- arch: x86_64
runner: x64-kernel-build-node
- arch: aarch64
runner: arm-kernel-build-node
runs-on: ${{ matrix.runner }}
steps:
- name: Clean workspace (remove root-owned files from prior runs)
run: |
docker run --rm -v "${{ github.workspace }}:/workspace" alpine:3 \
sh -c 'rm -rf /workspace/..?* /workspace/.[!.]* /workspace/*' || true
- uses: actions/checkout@v4
- name: Checkout FlashMLA
uses: actions/checkout@v4
with:
repository: sgl-project/FlashMLA
ref: ${{ inputs.branch || 'sgl' }}
path: FlashMLA
submodules: recursive
- name: Set up Python ${{ matrix.python-version }}
uses: actions/setup-python@v5
with:
python-version: ${{ matrix.python-version }}
- name: Set wheel version
run: |
echo -n "${{ inputs.version }}" > FlashMLA/flash_mla/VERSION
cat FlashMLA/flash_mla/VERSION
- name: Build wheel
run: |
chmod +x ./scripts/build_sgl_flash_mla.sh ./scripts/rename_sgl_flash_mla_whl.sh
./scripts/build_sgl_flash_mla.sh "${{ matrix.python-version }}" "${{ matrix.cuda-version }}" "${{ github.workspace }}/FlashMLA" "${{ matrix.arch }}"
- name: Upload artifacts
uses: actions/upload-artifact@v4
with:
name: flashmla-wheel-${{ matrix.arch }}
path: FlashMLA/dist/*.whl
- name: Upload PyPI artifacts
uses: actions/upload-artifact@v4
with:
name: flashmla-pypi-${{ matrix.arch }}
path: FlashMLA/dist-pypi/*.whl
test:
needs: build
if: ${{ !inputs.skip_tests }}
strategy:
fail-fast: false
matrix:
include:
# SM90 runs only the frozen kernels: the current ones are sm_100a/sm_103a.
# The frozen ones are built for sm_90a/sm_100a/sm_103a, so SM100 runs both.
- arch_label: sm90
runner: 8-gpu-h200
wheel_arch: x86_64
tests: legacy
timeout: 90
- arch_label: sm100
runner: 8-gpu-b200
wheel_arch: x86_64
tests: current legacy
timeout: 180
- arch_label: sm103-aarch64
runner: 4-gpu-gb300
wheel_arch: aarch64
tests: current legacy
timeout: 180
runs-on: ${{ matrix.runner }}
timeout-minutes: ${{ matrix.timeout }}
steps:
- uses: actions/checkout@v4
- name: Checkout FlashMLA
uses: actions/checkout@v4
with:
repository: sgl-project/FlashMLA
ref: ${{ inputs.branch || 'sgl' }}
path: FlashMLA
submodules: recursive
# The tests need only tests/. The source package would shadow the installed
# wheel (its directory has no built extensions), so it is removed.
- name: Remove the FlashMLA source package
run: rm -rf FlashMLA/flash_mla
- name: Download wheel
uses: actions/download-artifact@v4
with:
path: dist/
merge-multiple: true
pattern: flashmla-wheel-${{ matrix.wheel_arch }}
- name: Install wheel and deps
run: |
python3 -m pip install --upgrade pip
python3 -m pip install "torch==${TORCH_VER}" --index-url https://download.pytorch.org/whl/cu130 --force-reinstall
python3 -m pip install numpy einops rich pytest
# --force-reinstall: a base image may already carry this version, in
# which case pip skips the wheel and the job tests a stale extension.
python3 -m pip install dist/*.whl --force-reinstall --no-deps
# Both extensions register through STABLE_TORCH_LIBRARY at dlopen time,
# so a missing symbol surfaces here rather than at the first launch.
python3 -c "
import flash_mla
from flash_mla._loader import get_ops
print('flash_mla:', flash_mla.__file__)
import os
assert not flash_mla.__file__.startswith(os.environ['GITHUB_WORKSPACE']), 'not the installed wheel'
print('version:', flash_mla.__version__)
print('has_legacy:', flash_mla.has_legacy())
assert get_ops(use_legacy=False).sparse_decode_fwd is not None
assert get_ops(use_legacy=True).dense_decode_fwd is not None
"
- name: Run tests (current kernels)
if: contains(matrix.tests, 'current')
working-directory: FlashMLA
env:
PYTHONPATH: ${{ github.workspace }}/FlashMLA/tests
run: |
python3 -c "import os, flash_mla; assert not flash_mla.__file__.startswith(os.environ['GITHUB_WORKSPACE']), flash_mla.__file__"
python3 tests/test-sparse-prefill.py
python3 tests/test-sparse-decode.py
- name: Run tests (frozen kernels)
if: contains(matrix.tests, 'legacy')
working-directory: FlashMLA
env:
# conftest.py only fixes sys.path under pytest; these are scripts.
PYTHONPATH: ${{ github.workspace }}/FlashMLA/tests
run: |
python3 -c "import os, flash_mla; assert not flash_mla.__file__.startswith(os.environ['GITHUB_WORKSPACE']), flash_mla.__file__"
python3 -m pytest tests/legacy/ -q
# Dense MLA decoding is SM90-only.
if [ "${{ matrix.arch_label }}" = sm90 ]; then
python3 tests/legacy/test_flash_mla_dense_decoding.py
fi
python3 tests/legacy/test_flash_mla_sparse_prefill.py
python3 tests/legacy/test_flash_mla_sparse_decoding.py
python3 tests/legacy/test_flash_mla_sparse_decoding_v32_bit_equal.py
release:
if: |
!cancelled() && !inputs.validation_only &&
needs.build.result == 'success' &&
(needs.test.result == 'success' || needs.test.result == 'skipped')
needs: [build, test]
runs-on: ubuntu-latest
steps:
- uses: actions/checkout@v4
- name: Download artifacts
uses: actions/download-artifact@v4
with:
path: dist/
merge-multiple: true
pattern: flashmla-wheel-*
- name: Download PyPI artifacts
uses: actions/download-artifact@v4
with:
path: dist-pypi/
merge-multiple: true
pattern: flashmla-pypi-*
- name: Record that the GPU test suite was skipped
if: ${{ inputs.skip_tests }}
run: |
echo "Releasing v${{ inputs.version }} with the GPU test suite skipped." \
>> "$GITHUB_STEP_SUMMARY"
- name: Upload to PyPI
run: |
pip install --upgrade twine "packaging>=24.2"
python3 -m twine upload --skip-existing dist-pypi/* -u __token__ -p ${{ secrets.SGL_FLASH_MLA_PYPI_TOKEN }}
- name: Release
uses: softprops/action-gh-release@v2
with:
tag_name: v${{ inputs.version }}
repository: sgl-project/whl
token: ${{ secrets.GH_PAT_FOR_WHL_RELEASE }}
files: |
dist/*
- name: Clone wheel index
run: git clone https://oauth2:${WHL_TOKEN}@github.com/sgl-project/whl.git sgl-whl
env:
WHL_TOKEN: ${{ secrets.GH_PAT_FOR_WHL_RELEASE }}
- name: Update wheel index
run: python3 scripts/update_flashmla_whl_index.py --cuda 130
- name: Push wheel index
run: |
cd sgl-whl
git config --local user.name "sglang-bot"
git config --local user.email "sglangbot@gmail.com"
git add -A
git commit -m "update sgl-flash-mla whl index for v${{ inputs.version }}${{ inputs.skip_tests && ' (GPU tests skipped)' || '' }}"
git push
+4 -3
View File
@@ -39,7 +39,7 @@ jobs:
(github.event_name == 'push' || github.event.inputs.target == 'all' || github.event.inputs.target == 'cu130')
strategy:
matrix:
python-version: ["3.10"]
python-version: ["3.11"]
cuda-version: ["13.0"]
arch: [x86_64, aarch64]
include:
@@ -174,8 +174,9 @@ jobs:
matrix:
python-version: ["3.10"]
# ROCm 10.0.0 builds inside a Python 3.12 container; the wheel is still
# tagged cp310-abi3 (wheel.py-api in pyproject.toml), so it stays under
# the same python-version leg as the 7.x wheels.
# tagged cp310-abi3 (build_rocm.sh sets wheel.py-api to cp310 for every
# ROCm flavor), so it stays under the same python-version leg as the 7.x
# wheels.
#
# A dispatch that names one flavor builds only that flavor. Every extra
# leg is another image build on the shared AMD node, and its release job
+1 -1
View File
@@ -183,7 +183,7 @@ jobs:
env:
TEST_COMMAND: ${{ matrix.test_command }}
REQUESTED_TEST_COMMAND: ${{ inputs.test_command }}
HF_MUSE_SPARK_TOKEN: ${{ inputs.runs_on == '8-gpu-b300' && inputs.pr_head_sha == '' && contains(matrix.test_command, 'registered/e2e/models/test_muse_spark_12_b300_pd.py') && secrets.HF_MUSE_SPARK_TOKEN || '' }}
HF_MUSE_SPARK_TOKEN: ${{ inputs.pr_head_sha == '' && inputs.runs_on == '8-gpu-b300' && (contains(matrix.test_command, 'registered/e2e/models/test_muse_spark_12_b300_pd.py') || contains(matrix.test_command, 'registered/e2e/models/test_muse_spark_unified_memory.py')) && secrets.HF_MUSE_SPARK_TOKEN || '' }}
SGLANG_PRECISION_HF_TOKEN: ${{ inputs.refresh_precision_baseline && inputs.mode == 'cuda' && inputs.runs_on == '8-gpu-h200' && inputs.test_command == 'registered/debug_utils/test_nightly_precision_regression.py' && inputs.pr_head_sha == '' && secrets.HF_TOKEN_PRECISION_STORE || '' }}
run: |
if [[ "${{ inputs.refresh_precision_baseline }}" == "true" ]]; then
+69
View File
@@ -0,0 +1,69 @@
# Runs only the per-commit CPU unit suite (base-a-test-cpu) against a PR head,
# dispatched by the `/run-cpu-ci` slash command. A cheap check for small PRs:
# pair it with `/rerun-group <module>` instead of a full `run-ci` round.
#
# Always dispatched on main, so the workflow definition and every reusable
# workflow it calls are trusted; only the code under test comes from the PR.
# That code runs on hosted runners with a read-only token and no secrets
# (nothing here inherits them), the same exposure as `/rerun-test` cpu mode.
name: Run CPU CI
run-name: "[run-cpu-ci] ${{ inputs.git_ref }}"
on:
workflow_dispatch:
inputs:
git_ref:
description: "PR head SHA to test"
required: true
type: string
pr_number:
description: "PR number (informational)"
required: false
type: string
default: ""
# Read by _pr-test-stage-cpu.yml via caller_inputs. Its fail-fast health
# check inspects sibling jobs of a pull_request run, which this is not.
skip_pr_test_health_check:
description: "Skip the PR test health check fail-fast"
required: false
type: boolean
default: true
# Also read via caller_inputs. This run is on main's ref, so a cache it
# saved would be restored by main and every PR, yet hold PR-written files.
skip_cache_save:
description: "Do not save the HF / Rust caches"
required: false
type: boolean
default: true
concurrency:
group: run-cpu-ci-${{ inputs.git_ref }}
cancel-in-progress: true
permissions:
contents: read
issues: read
pull-requests: read
jobs:
# pr_test_yml stays pr-test.yml: compute_partitions.py reads it from the PR
# checkout, which may predate this file, and sizing shards off the per-commit
# stage's run_timeout_minutes keeps the split identical to PR CI.
check-changes:
uses: ./.github/workflows/_pr-test-check-changes.yml
with:
git_ref: ${{ inputs.git_ref }}
run_all_tests: true
# self_name must be the suite name: the stage passes it to run_suite.py.
# No rust_ext_artifact: its compile job runs on a self-hosted build node, so
# the stage falls back to its cache and compiles on a miss.
base-a-test-cpu:
needs: check-changes
uses: ./.github/workflows/_pr-test-stage-cpu.yml
with:
self_name: base-a-test-cpu
check_changes: ${{ toJson(needs.check-changes.outputs) }}
caller_inputs: ${{ toJson(inputs) }}
partitions: ${{ needs.check-changes.outputs.partitions }}
run_timeout_minutes: '15'
+3 -2
View File
@@ -7,6 +7,7 @@ on:
paths:
- '.github/workflows/runner-utilization.yml'
- 'scripts/ci/utils/runner_utilization_report.py'
- 'scripts/ci/utils/test_runner_utilization_report.py'
workflow_dispatch:
inputs:
hours:
@@ -50,7 +51,7 @@ jobs:
- name: Set up Python
uses: actions/setup-python@v5
with:
python-version: '3.10'
python-version: '3.12'
- name: Unit-test report logic
# Pure-logic tests (stdlib only, no API calls) — guards the
@@ -70,7 +71,7 @@ jobs:
run: |
# PR-trigger is just a script smoke check; scan a 20-minute window
# (0.34h) to bound API cost. schedule / dispatch run the full report.
python scripts/ci/utils/runner_utilization_report.py \
python -u scripts/ci/utils/runner_utilization_report.py \
--repo ${{ github.repository }} \
--hours ${{ (github.event_name == 'pull_request' && '0.34') || inputs.hours || '24' }} \
${{ inputs.filter && format('--filter {0}', inputs.filter) || '' }} \
+3 -1
View File
@@ -27,6 +27,8 @@ jobs:
contains(github.event.comment.body, '/run-extra-ci') ||
contains(github.event.comment.body, '/rerun-full-ci') ||
contains(github.event.comment.body, '/rerun-extra-ci') ||
contains(github.event.comment.body, '/run-cpu-ci') ||
contains(github.event.comment.body, '/rerun-cpu-ci') ||
contains(github.event.comment.body, '/rerun-group') ||
contains(github.event.comment.body, '/rerun-test') ||
startsWith(github.event.comment.body, '/tag-') ||
@@ -92,7 +94,7 @@ jobs:
- name: Set up Python
uses: actions/setup-python@v5
with:
python-version: '3.10'
python-version: '3.12'
- name: Install dependencies
run: |
+1 -1
View File
@@ -21,7 +21,7 @@ jobs:
- name: Set up Python
uses: actions/setup-python@v5
with:
python-version: '3.10'
python-version: '3.12'
- name: Update est_time values
env:
+3 -2
View File
@@ -73,7 +73,8 @@ instance/
# PyBuilder
.pybuilder/
target/
# `**/` because .dockerignore links here and Docker anchors bare patterns at the context root.
**/target/
# Jupyter Notebook
.ipynb_checkpoints
@@ -232,7 +233,7 @@ work_dirs/
Cargo.lock
!rust/Cargo.lock
!rust/sglang-radix-tree/Cargo.lock
!experimental/sgl-router/Cargo.lock
!sgl-router/Cargo.lock
# Other repos
lmms-eval
+15 -7
View File
@@ -11,10 +11,6 @@ repos:
- id: end-of-file-fixer
- id: check-yaml
args: [--allow-multiple-documents]
# Helm chart templates embed Go template syntax ({{- ... -}})
# that is not valid YAML on its own; the rendered output is
# validated by `helm template` / `helm lint`.
exclude: ^experimental/sgl-router/helm/.*/templates/.*\.(yaml|tpl)$
- id: check-toml
- id: check-ast
- id: check-added-large-files
@@ -74,6 +70,12 @@ repos:
- '--extra-keys=metadata.kernelspec metadata.language_info.version'
- repo: local
hooks:
- id: forbid-test-srt-path
name: forbid test/srt/ (use test/registered/ or test/manual/)
entry: >-
test/srt/ is retired. Move these tests under test/registered/ or test/manual/ instead.
language: fail
files: ^test/srt/
- id: forbid-docs-new-path
name: forbid docs_new/ (renamed to docs/)
# docs_new/ was renamed to docs/ in #32123. `language: fail` fires only
@@ -110,6 +112,12 @@ repos:
language: system
files: ^(\.github/actions/download-rust-ext/action\.yml|\.github/workflows/_pr-test-rust-ext-build\.yml)$
pass_filenames: false
- id: check-dockerfile-wheel-pins
name: check Dockerfile sgl-deep-gemm/sgl-deep-ep versions match python/pyproject.toml
entry: python3 scripts/lint/check_dockerfile_wheel_pins.py
language: system
files: ^(python/pyproject\.toml|docker/.*Dockerfile.*|scripts/lint/check_dockerfile_wheel_pins\.py)$
pass_filenames: false
- id: check-no-bare-pytest-main
name: reject bare pytest.main calls in __main__ blocks
entry: python3 scripts/lint/check_no_bare_pytest_main.py
@@ -134,10 +142,10 @@ repos:
files: ^sgl-model-gateway/.*\.rs$
pass_filenames: false
- id: rustfmt-sgl-router
name: rustfmt experimental/sgl-router
entry: bash -c 'cd experimental/sgl-router && cargo fmt --all -- --check'
name: rustfmt sgl-router
entry: bash -c 'cd sgl-router && cargo fmt --all -- --check'
language: system
files: ^experimental/sgl-router/.*\.rs$
files: ^sgl-router/.*\.rs$
pass_filenames: false
- id: clippy-rust-workspace
name: clippy rust/ workspace (auto-fix)
+15 -1
View File
@@ -56,6 +56,19 @@ if [[ "${ROCM_VERSION}" == "1000" ]]; then
WHEEL_ROCM_VERSION="1000"
fi
# The build reuses the CUDA pyproject.toml, which requires Python 3.11 and tags
# its wheel cp311-abi3. The ROCm 7.2.0 image and amd-sglang still run on Python
# 3.10, so the ROCm wheels keep the Python 3.10 floor and the cp310-abi3 tag.
sed -i \
-e 's/^requires-python = ">=3.11"$/requires-python = ">=3.10"/' \
-e 's/^wheel.py-api = "cp311"$/wheel.py-api = "cp310"/' \
pyproject.toml
if ! grep -qx 'requires-python = ">=3.10"' pyproject.toml \
|| ! grep -qx 'wheel.py-api = "cp310"' pyproject.toml; then
echo "ERROR: could not set the ROCm Python 3.10 / cp310 metadata in pyproject.toml." >&2
exit 1
fi
# Pull and run the latest image
echo "Pulling Docker image: ${IMAGE}"
docker pull "${IMAGE}"
@@ -86,5 +99,6 @@ ${ROCM_SETUP}
rm -rf CMakeLists.txt && mv CMakeLists_rocm.txt CMakeLists.txt && \
${PYTHON_ROOT_PATH}/python rocm_hipify.py && \
${PYTHON_ROOT_PATH}/python -m uv build --wheel -Cbuild-dir=build . --color=always --no-build-isolation && \
./rename_wheels_rocm.sh ${WHEEL_ROCM_VERSION}
./rename_wheels_rocm.sh ${WHEEL_ROCM_VERSION} && \
${PYTHON_ROOT_PATH}/pip install --dry-run --no-deps dist/*.whl
"
+8 -1
View File
@@ -22,6 +22,11 @@ SGLang is an open-source inference framework for large language, vision-language
👋 Get started below, or meet the community at [SGLang Events](https://www.sglang.io/events), including meetups, developer meetings, workshops, and office hours.
> **[SGLang Summit 2026: Meet the builders of open AI](https://www.sglang.io/summit)**<br>
> **Our flagship annual gathering · November 12–13 · San Francisco**
>
> Go deep on inference, training, and AI hardware through technical talks, live demos, and hands-on workshops. Bring your deployment challenges and meet the engineers and researchers building the stack.
## Get Started
Pull the Docker image, which includes SGLang and its dependencies:
@@ -49,7 +54,7 @@ SGLang supports a wide range of GPUs, TPUs, NPUs, CPUs, and Apple Silicon platfo
| --- | --- |
| [NVIDIA](https://docs.sglang.io/docs/hardware-platforms/nvidia-gpus) | B200/B300/GB200/GB300; H100/H200/H800/H20; A100; RTX 30/40/50 series, RTX 6000 Ada / PRO 6000; DGX Spark, Jetson Orin |
| [AMD](https://docs.sglang.io/docs/hardware-platforms/amd_gpu) | Instinct MI300X, MI325X, MI350X, MI355X |
| [Google TPU](https://docs.sglang.io/docs/hardware-platforms/tpu) | v6e, v7; [SGL-JAX](https://github.com/sgl-project/sglang-jax) / [SGL-torchtpu](https://lmsys.org/blog/2026-07-30-sglang-google-tpu/) |
| [Google TPU](https://docs.sglang.io/docs/hardware-platforms/tpu) | v6e, v7; [SGLang-TorchTPU](https://github.com/sgl-project/sglang-torchtpu) / [SGL-JAX](https://github.com/sgl-project/sglang-jax) |
| Intel ([GPU](https://docs.sglang.io/docs/hardware-platforms/xpu) / [CPU](https://docs.sglang.io/docs/hardware-platforms/cpu_server)) | Arc / Arc Pro B-Series GPUs, Xeon CPUs |
| [Apple Silicon](https://docs.sglang.io/docs/hardware-platforms/apple_metal) | Macs via Metal / MLX |
| [Huawei Ascend](https://docs.sglang.io/docs/hardware-platforms/ascend-npus/getting-started/installation) | A2, A3, 950PR/DT NPUs |
@@ -57,6 +62,8 @@ SGLang supports a wide range of GPUs, TPUs, NPUs, CPUs, and Apple Silicon platfo
Integrations in progress: AWS Trainium, [Alibaba T-Head PPU](https://github.com/sgl-project/sglang/issues/37519), [Cambricon MLU](https://github.com/sgl-project/sglang/pull/26898), Qualcomm QAIC, MetaX, Hygon HCU/DCU, Iluvatar CoreX, and more.
SGLang-TorchTPU brings the SGLang serving engine to Google TPUs through PyTorch, torch_tpu, and Pallas kernels. It supports LLM and vision-language serving on TPU7x (Ironwood), including tensor and expert parallelism across hosts. SGL-JAX provides the JAX-based TPU serving path.
See the [Cookbook](https://cookbook.sglang.io/) and platform guides for model compatibility and setup.
## SGL Ecosystem
@@ -20,10 +20,10 @@ import triton
from sglang.kernels.ops.attention.cutedsl_kda import (
cutedsl_fused_sigmoid_gating_kda_update,
)
from sglang.kernels.ops.attention.fla.fused_sigmoid_gating_recurrent import (
from sglang.kernels.ops.attention.fla.kda import chunk_kda
from sglang.kernels.ops.attention.linear.fused_sigmoid_gating_recurrent import (
fused_sigmoid_gating_delta_rule_update,
)
from sglang.kernels.ops.attention.fla.kda import chunk_kda
def make_inputs(
@@ -32,7 +32,7 @@ import triton
from sglang.kernels.ops.attention.fla.fused_recurrent import (
fused_recurrent_gated_delta_rule_packed_decode,
)
from sglang.kernels.ops.attention.fla.fused_sigmoid_gating_recurrent import (
from sglang.kernels.ops.attention.linear.fused_sigmoid_gating_recurrent import (
fused_sigmoid_gating_delta_rule_update,
)
@@ -26,12 +26,12 @@ import torch
from sglang.kernels.ops.attention.fla.fused_recurrent import (
fused_recurrent_kda_packed_decode,
)
from sglang.kernels.ops.attention.fla.fused_sigmoid_gating_recurrent import (
fused_sigmoid_gating_delta_rule_update,
)
from sglang.kernels.ops.attention.helion.kda_decode import (
helion_fused_recurrent_kda_packed_decode,
)
from sglang.kernels.ops.attention.linear.fused_sigmoid_gating_recurrent import (
fused_sigmoid_gating_delta_rule_update,
)
def make_inputs(
@@ -37,7 +37,7 @@ import triton
from sglang.kernels.ops.attention.fla.fused_recurrent import (
fused_recurrent_gated_delta_rule_packed_decode,
)
from sglang.kernels.ops.attention.fla.fused_recurrent_linear_replayssm import (
from sglang.kernels.ops.attention.linear.fused_recurrent_linear_replayssm import (
fused_recurrent_gdn_replayssm_decode,
)
+3 -3
View File
@@ -22,11 +22,11 @@ import argparse
import torch
import sglang.kernels.ops.attention.fla.fused_kda_conv_recurrent_verify as fused_mod
from sglang.kernels.ops.attention.fla.fused_kda_conv_recurrent_verify import (
import sglang.kernels.ops.attention.linear.fused_kda_conv_recurrent_verify as fused_mod
from sglang.kernels.ops.attention.linear.fused_kda_conv_recurrent_verify import (
fused_kda_conv_gating_verify,
)
from sglang.kernels.ops.attention.fla.fused_sigmoid_gating_recurrent import (
from sglang.kernels.ops.attention.linear.fused_sigmoid_gating_recurrent import (
fused_sigmoid_gating_delta_rule_update,
)
from sglang.kernels.ops.mamba.causal_conv1d_triton import causal_conv1d_update
@@ -3,7 +3,7 @@
Compares the legacy gather (``gather_dequant_requant_fp8_paged_legacy``:
fresh ``torch.zeros`` destination + one program per (token, 128-elem
slice)) against the new gather (``gather_dequant_requant_fp8_paged``:
slice)) against the new gather (``gather_dequant_requant_fp8_paged_dsa``:
no pre-zeroing needed, fused pad-row zero-fill, TOKENS_PER_PROG tokens
per program with 16B-vectorized access), in three flavors:
@@ -43,6 +43,8 @@ import sys
from pathlib import Path
import torch
import triton
import triton.language as tl
_REPO_ROOT = Path(__file__).resolve().parents[3]
_MOD_PATH = _REPO_ROOT / "python/sglang/kernels/ops/attention/dsa/dequant_k_cache.py"
@@ -54,8 +56,127 @@ _spec = importlib.util.spec_from_file_location("dequant_k_cache", _MOD_PATH)
_mod = importlib.util.module_from_spec(_spec)
_spec.loader.exec_module(_mod)
gather_new = _mod.gather_dequant_requant_fp8_paged
gather_legacy = _mod.gather_dequant_requant_fp8_paged_legacy
gather_new = _mod.gather_dequant_requant_fp8_paged_dsa
_infer_dsa_dims = _mod._infer_dsa_dims
def gather_dequant_requant_fp8_paged_legacy(
quant_k_cache: torch.Tensor,
page_table_1_flattened: torch.Tensor,
group_size: int = 128,
extra_rows: int = 0,
) -> torch.Tensor:
"""Legacy (pre-vectorization) gather + dequant + requant.
Kept as the bit-exactness / performance reference for
``gather_dequant_requant_fp8_paged_dsa`` (see
``benchmark/kernels/deepseek/benchmark_q8kv8_kv_gather.py``). Allocates and zero-fills the
full destination each call, then launches one program per
(token, 128-elem slice).
"""
dim_quant = quant_k_cache.shape[-1]
dim_nope, dim_rope = _infer_dsa_dims(dim_quant)
quant_k_cache = quant_k_cache.view((-1, dim_quant))
num_tokens = page_table_1_flattened.shape[0]
assert quant_k_cache.dtype == torch.float8_e4m3fn
num_tiles = dim_nope // group_size
out_dim = dim_nope + dim_rope
assert num_tiles * group_size == dim_nope
total_rows = num_tokens + extra_rows
# Allocate a fresh zero-filled buffer. The extra landing-pad rows at
# the tail must read as zeros (the kernel may over-read past
# num_tokens for masked indices).
output = torch.zeros(
(total_rows, 1, out_dim),
dtype=torch.float8_e4m3fn,
device=quant_k_cache.device,
)
num_blocks_per_token = num_tiles + triton.cdiv(dim_rope, group_size)
input_nope_q = quant_k_cache[:, :dim_nope]
input_nope_s = quant_k_cache[:, dim_nope : dim_nope + num_tiles * 4].view(
torch.float32
)
input_rope = quant_k_cache[:, dim_nope + num_tiles * 4 :].view(torch.bfloat16)
_gather_dequant_requant_fp8_paged_kernel[(num_tokens, num_blocks_per_token)](
output,
input_nope_q,
input_nope_s,
input_rope,
page_table_1_flattened,
output.stride(0),
input_nope_q.stride(0),
input_nope_s.stride(0),
input_rope.stride(0),
NUM_NOPE_BLOCKS=num_tiles,
GROUP_SIZE=group_size,
DIM_NOPE=dim_nope,
DIM_ROPE=dim_rope,
)
return output
@triton.jit
def _gather_dequant_requant_fp8_paged_kernel(
output_ptr,
input_nope_q_ptr,
input_nope_s_ptr,
input_rope_ptr,
page_table_1_ptr,
output_stride_0: int,
input_nope_q_stride_0: int,
input_nope_s_stride_0: int,
input_rope_stride_0: int,
NUM_NOPE_BLOCKS: tl.constexpr,
GROUP_SIZE: tl.constexpr,
DIM_NOPE: tl.constexpr,
DIM_ROPE: tl.constexpr,
):
"""Fused gather + dequant(per-group) + requant(per-tensor) -> fp8."""
token_id = tl.program_id(0)
token_id_paged = tl.load(page_table_1_ptr + token_id).to(tl.int32)
raw_block_id = tl.program_id(1)
if raw_block_id < NUM_NOPE_BLOCKS:
# nope: read fp8, mul group scale -> f32, cast to fp8_e4m3fn
effective_block_id = raw_block_id
offs_q = effective_block_id * GROUP_SIZE + tl.arange(0, GROUP_SIZE)
mask = offs_q < DIM_NOPE
ptr_q = input_nope_q_ptr + token_id_paged * input_nope_q_stride_0 + offs_q
ptr_s = (
input_nope_s_ptr
+ token_id_paged * input_nope_s_stride_0
+ effective_block_id
)
y_q = tl.load(ptr_q, mask=mask, other=0.0).to(tl.float32)
y_s = tl.load(ptr_s)
# dequant -> f32 -> requant to fp8
y = (y_q * y_s).to(tl.float8e4nv)
dst_ptr = output_ptr + token_id * output_stride_0 + offs_q
tl.store(dst_ptr, y, mask=mask)
else:
# rope: read bf16, cast to fp8
effective_block_id = raw_block_id - NUM_NOPE_BLOCKS
offs = effective_block_id * GROUP_SIZE + tl.arange(0, GROUP_SIZE)
mask = offs < DIM_ROPE
src_ptr = input_rope_ptr + token_id_paged * input_rope_stride_0 + offs
dst_ptr = output_ptr + token_id * output_stride_0 + DIM_NOPE + offs
data = tl.load(src_ptr, mask=mask).to(tl.float8e4nv)
tl.store(dst_ptr, data, mask=mask)
gather_legacy = gather_dequant_requant_fp8_paged_legacy
PAGE_SIZE = 64
DIM_QUANT = 656 # 512 nope fp8 + 16 scale bytes + 128 rope bytes
@@ -54,7 +54,7 @@ import argparse
import torch
from sglang.kernels.ops.kvcache.cache_ops import (
from sglang.kernels.ops.attention.mla_qk_prep import (
absorbed_bmm_concat_cast_q_fp8,
concat_and_cast_q_fp8_pad,
)
@@ -1,5 +1,5 @@
"""Tune the gfx950 native MXFP8 dense route for one (N, K) and write its rows into
python/sglang/kernels/ops/quantization/mxfp8_gemv_gfx95_configs.json.
python/sglang/kernels/ops/gemm/mxfp8_gemv_gfx95_configs.json.
Two tables: the scaled-MFMA GEMV config per M bucket (1 ... 32) and the dot_scaled tile per
M bucket (64 ... 16384). Each candidate is timed by HIP-graph replay rotating over weight
@@ -18,7 +18,7 @@ import torch
from triton.runtime.errors import OutOfResources
from sglang.kernels.jit.utils import empty_sentinel
from sglang.kernels.ops.quantization import mxfp8_native_amd_gfx95 as native
from sglang.kernels.ops.gemm import mxfp8_native_amd_gfx95 as native
from sglang.kernels.ops.quantization.mxfp8_amd_gfx95 import fp8_grid_quantize
GRAPH_CALLS = 20
@@ -0,0 +1,89 @@
"""Benchmark seeded versus unseeded EAGLE draft Gumbel sampling.
Example:
python benchmark/kernels/sampling/bench_deterministic_fast_sample.py \
--batch-sizes 1 8 32 128 --vocab-sizes 32000 128256 163840
"""
import argparse
import torch
import triton
from sglang.srt.speculative.spec_utils import fast_sample
def parse_args():
parser = argparse.ArgumentParser()
parser.add_argument("--batch-sizes", type=int, nargs="+", default=[1, 8, 32, 128])
parser.add_argument(
"--vocab-sizes", type=int, nargs="+", default=[32000, 128256, 163840]
)
parser.add_argument("--warmup-ms", type=int, default=100)
parser.add_argument("--rep-ms", type=int, default=500)
parser.add_argument(
"--modes", nargs="+", choices=("eager", "graph"), default=["graph"]
)
return parser.parse_args()
def capture(fn):
fn()
torch.cuda.synchronize()
graph = torch.cuda.CUDAGraph()
with torch.cuda.graph(graph):
fn()
return graph.replay
def main():
args = parse_args()
device = torch.device("cuda")
print(f"device={torch.cuda.get_device_name(device)}")
print("mode\tbatch\tvocab\tunseeded_us\tseeded_us\tslowdown")
for vocab_size in args.vocab_sizes:
for batch_size in args.batch_sizes:
logits = torch.randn(
(batch_size, vocab_size), device=device, dtype=torch.float32
)
probs = torch.softmax(logits, dim=-1)
seeds = torch.arange(batch_size, device=device, dtype=torch.int64) + 12345
positions = (
torch.arange(batch_size, device=device, dtype=torch.int64) + 4096
)
def unseeded():
return fast_sample(probs)
def seeded():
return fast_sample(
probs,
sampling_seed=seeds,
positions=positions,
draft_step=1,
)
# Compile/JIT and populate allocator caches before timing.
unseeded()
seeded()
torch.cuda.synchronize()
for mode in args.modes:
unseeded_fn = capture(unseeded) if mode == "graph" else unseeded
seeded_fn = capture(seeded) if mode == "graph" else seeded
unseeded_ms = triton.testing.do_bench(
unseeded_fn, warmup=args.warmup_ms, rep=args.rep_ms
)
seeded_ms = triton.testing.do_bench(
seeded_fn, warmup=args.warmup_ms, rep=args.rep_ms
)
print(
f"{mode}\t{batch_size}\t{vocab_size}\t"
f"{unseeded_ms * 1000:.1f}\t{seeded_ms * 1000:.1f}\t"
f"{seeded_ms / unseeded_ms:.2f}x"
)
if __name__ == "__main__":
main()
+21
View File
@@ -626,6 +626,26 @@ RUN --mount=type=cache,target=/root/.cache/uv \
uv pip install --reinstall /tmp/gateway_wheels/*.whl \
&& rm -rf /tmp/gateway_wheels
# sgl-nvidia-mega: the NVIDIA megakernel plugin, inert unless a server passes
# --enable-nvidia-megakernel. Installed with CUDA >= 12.9 only (its kernels are
# SM103, compiled at runtime); disable with --build-arg ENABLE_SGL_NVIDIA_MEGA=0.
ARG ENABLE_SGL_NVIDIA_MEGA=1
ARG SGL_NVIDIA_MEGA_REPO="https://github.com/sgl-project/sgl-nvidia-mega.git"
ARG SGL_NVIDIA_MEGA_COMMIT="3333e4cfac0f2b15603565d702b4c67538f1628f"
RUN --mount=type=cache,target=/root/.cache/uv \
if [ "${ENABLE_SGL_NVIDIA_MEGA}" != "1" ]; then \
echo "[sgl-nvidia-mega] skip (ENABLE_SGL_NVIDIA_MEGA=${ENABLE_SGL_NVIDIA_MEGA})"; \
elif ! printf '%s\n' 12.9 "${CUDA_VERSION}" | sort -V -C; then \
echo "[sgl-nvidia-mega] skip (CUDA_VERSION=${CUDA_VERSION} < 12.9)"; \
else \
git clone "${SGL_NVIDIA_MEGA_REPO}" /sgl-workspace/sgl-nvidia-mega \
&& cd /sgl-workspace/sgl-nvidia-mega \
&& git checkout -f "${SGL_NVIDIA_MEGA_COMMIT}" \
&& uv pip install --no-deps -e . \
&& cd / \
&& python3 -c "from importlib.metadata import entry_points; assert any(e.name == 'nvidia_megakernel' for e in entry_points(group='sglang.srt.plugins')), 'sgl-nvidia-mega plugin not registered'"; \
fi
# Set workspace directory
WORKDIR /sgl-workspace/sglang
@@ -691,6 +711,7 @@ RUN --mount=type=cache,target=/var/cache/apt,id=runtime-apt \
# Runtime libraries
libopenmpi3 \
libnuma1 \
numactl \
libibverbs1 \
libibumad3 \
librdmacm1 \
+32 -5
View File
@@ -1,13 +1,13 @@
# CUDA 13.4 only args
# Torch nighly args
ARG TORCH_NIGHTLY_INDEX=https://download.pytorch.org/whl/nightly/cu134
ARG TORCH_NIGHTLY_VERSION=2.15.0.dev20260818+cu134
ARG TORCHVISION_NIGHTLY_VERSION=0.30.0.dev20260819+cu134
ARG TORCHAUDIO_NIGHTLY_VERSION=2.11.0.dev20260818+cu134
ARG TORCH_NIGHTLY_VERSION=2.15.0.dev20261005+cu134
ARG TORCHVISION_NIGHTLY_VERSION=0.30.0.dev20261005+cu134
ARG TORCHAUDIO_NIGHTLY_VERSION=2.11.0.dev20261004+cu134
# Wheel build args
ARG MANYLINUX_IMAGE=pytorch/manylinuxaarch64-builder:cuda13.4
ARG DEEPGEMM_REF=v0.2.1.post1
ARG SGL_DEEP_GEMM_VERSION=0.2.1.post1
ARG DEEPGEMM_REF=v0.2.3.post2
ARG SGL_DEEP_GEMM_VERSION=0.2.3.post2
ARG DEEPEP_SOURCE_REF=sgl-deepep
ARG SGL_DEEP_EP_VERSION=0.1.2.post1
@@ -39,6 +39,9 @@ ENV DEBIAN_FRONTEND=noninteractive \
ENV TRITON_PTXAS_PATH=/usr/local/cuda-13.4/bin/ptxas \
TRITON_PTXAS_BLACKWELL_PATH=/usr/local/cuda-13.4/bin/ptxas
# Turn off torch.distributed's NCCL2 path in the torch nightly.
ENV TORCH_DIST_USE_NCCL2=0
# Add GKE default lib and bin locations
ENV PATH="${PATH}:/usr/local/nvidia/bin" \
LD_LIBRARY_PATH="${LD_LIBRARY_PATH}:/usr/local/nvidia/lib:/usr/local/nvidia/lib64"
@@ -877,6 +880,26 @@ RUN --mount=type=cache,target=/root/.cache/pip \
python3 -m pip install --force-reinstall /tmp/gateway_wheels/*.whl \
&& rm -rf /tmp/gateway_wheels
# sgl-nvidia-mega: the NVIDIA megakernel plugin, inert unless a server passes
# --enable-nvidia-megakernel. Installed with CUDA >= 12.9 only (its kernels are
# SM103, compiled at runtime); disable with --build-arg ENABLE_SGL_NVIDIA_MEGA=0.
ARG ENABLE_SGL_NVIDIA_MEGA=1
ARG SGL_NVIDIA_MEGA_REPO="https://github.com/sgl-project/sgl-nvidia-mega.git"
ARG SGL_NVIDIA_MEGA_COMMIT="3333e4cfac0f2b15603565d702b4c67538f1628f"
RUN --mount=type=cache,target=/root/.cache/pip \
if [ "${ENABLE_SGL_NVIDIA_MEGA}" != "1" ]; then \
echo "[sgl-nvidia-mega] skip (ENABLE_SGL_NVIDIA_MEGA=${ENABLE_SGL_NVIDIA_MEGA})"; \
elif ! printf '%s\n' 12.9 "${CUDA_VERSION}" | sort -V -C; then \
echo "[sgl-nvidia-mega] skip (CUDA_VERSION=${CUDA_VERSION} < 12.9)"; \
else \
git clone "${SGL_NVIDIA_MEGA_REPO}" /sgl-workspace/sgl-nvidia-mega \
&& cd /sgl-workspace/sgl-nvidia-mega \
&& git checkout -f "${SGL_NVIDIA_MEGA_COMMIT}" \
&& python3 -m pip install --no-deps -e . \
&& cd / \
&& python3 -c "from importlib.metadata import entry_points; assert any(e.name == 'nvidia_megakernel' for e in entry_points(group='sglang.srt.plugins')), 'sgl-nvidia-mega plugin not registered'"; \
fi
# Set workspace directory
WORKDIR /sgl-workspace/sglang
@@ -927,6 +950,9 @@ ENV DEBIAN_FRONTEND=noninteractive \
ENV TRITON_PTXAS_PATH=/usr/local/cuda-13.4/bin/ptxas \
TRITON_PTXAS_BLACKWELL_PATH=/usr/local/cuda-13.4/bin/ptxas
# Turn off torch.distributed's NCCL2 path in the torch nightly.
ENV TORCH_DIST_USE_NCCL2=0
# Add GKE default lib and bin locations + CUDA compiler paths for FlashInfer JIT
ENV PATH="${PATH}:/usr/local/nvidia/bin:/usr/local/cuda/bin:/usr/local/cuda/nvvm/bin" \
LD_LIBRARY_PATH="${LD_LIBRARY_PATH}:/usr/local/nvidia/lib:/usr/local/nvidia/lib64"
@@ -947,6 +973,7 @@ RUN --mount=type=cache,target=/var/cache/apt,id=runtime-apt \
# Runtime libraries
libopenmpi3 \
libnuma1 \
numactl \
libibverbs1 \
libibumad3 \
librdmacm1 \
+31
View File
@@ -0,0 +1,31 @@
# syntax=docker/dockerfile:1
# Beam serving image (CUDA). Build from the repository root; application source
# is cloned from upstream:
# docker buildx build --platform linux/amd64 --load \
# -f docker/beam/beam.Dockerfile -t beam:local-test .
FROM lmsysorg/sglang:v0.5.22@sha256:e3a4f06001a29542fc4a8638536d9ffc5eb9c9942b08603c09b19a56eae4a1ae AS base
SHELL ["/bin/bash", "-o", "pipefail", "-c"]
# Replace the stock SGLang with the upstream beam branch. Pass a commit SHA as
# BEAM_REF to pin the source of a published image.
ARG BEAM_REF=beam
WORKDIR /sgl-workspace
RUN git init -q /sgl-workspace/sglang-beam \
&& git -C /sgl-workspace/sglang-beam fetch -q --depth 1 \
https://github.com/sgl-project/sglang.git "${BEAM_REF}" \
&& git -C /sgl-workspace/sglang-beam checkout -q --detach FETCH_HEAD \
&& uv pip uninstall sglang \
&& uv pip install --no-deps -e /sgl-workspace/sglang-beam/python
# Pins the beam branch adds on top of the v0.5.22 release: OpenAI 2.8.0 for the
# Responses shell types and the cuTile compiler for the Beam router GEMM. Record
# the new versions so later installs keep them.
RUN sed -i -E '/^(openai|nvidia-cuda-tileiras)==/d' /sgl-workspace/constraints.txt \
&& uv pip install -c /sgl-workspace/constraints.txt \
openai==2.8.0 nvidia-cuda-tileiras==13.4.92 \
&& uv pip freeze | grep -E '^(openai|nvidia-cuda-tileiras)==' \
>> /sgl-workspace/constraints.txt \
&& python3 -c "from importlib.metadata import version; assert version('openai') == '2.8.0'; assert version('nvidia-cuda-tileiras') == '13.4.92'"
WORKDIR /sgl-workspace/sglang-beam
+1 -1
View File
@@ -20,7 +20,7 @@ ARG TORCH_NPU_VERSION="2.10.0.post6"
ARG TORCH_NPU_INDEX_URL="https://ascend.devcloud.huaweicloud.com/pypi/simple/"
ARG SGLANG_TAG=main
ARG ASCEND_CANN_PATH=/usr/local/Ascend/ascend-toolkit
ARG SGLANG_KERNEL_NPU_TAG=2026.9.0.post9
ARG SGLANG_KERNEL_NPU_TAG=2026.9.0.post10
ARG PIP_INSTALL="python3 -m pip install --no-cache-dir"
ARG DEVICE_TYPE
ARG MODELSCOPE_VERSION=""
+31 -13
View File
@@ -65,7 +65,7 @@ ENV BUILD_TRITON="0"
ENV BUILD_LLVM="0"
ENV BUILD_AITER_ALL="1"
ENV BUILD_MOONCAKE="1"
ENV AITER_COMMIT_DEFAULT="2bb549102ba2d9e51da8ef21107361827075d8b6"
ENV AITER_COMMIT_DEFAULT="285579b6211b0da30f35b91a890c29c823597cd1"
# ===============================
# Base image 942 with rocm720 and args
@@ -75,7 +75,7 @@ ENV BUILD_TRITON="1"
ENV BUILD_LLVM="0"
ENV BUILD_AITER_ALL="1"
ENV BUILD_MOONCAKE="1"
ENV AITER_COMMIT_DEFAULT="2bb549102ba2d9e51da8ef21107361827075d8b6"
ENV AITER_COMMIT_DEFAULT="285579b6211b0da30f35b91a890c29c823597cd1"
ENV TRITON_COMMIT_DEFAULT="42270451990532c67e69d753fbd026f28fcc4840"
# ===============================
@@ -86,7 +86,7 @@ ENV BUILD_TRITON="1"
ENV BUILD_LLVM="0"
ENV BUILD_AITER_ALL="1"
ENV BUILD_MOONCAKE="1"
ENV AITER_COMMIT_DEFAULT="2bb549102ba2d9e51da8ef21107361827075d8b6"
ENV AITER_COMMIT_DEFAULT="285579b6211b0da30f35b91a890c29c823597cd1"
# Pin the ROCm torch stack for every pip invocation in this flavor. The file is
# filled in after the torch 2.11 upgrade below; it must already exist (empty is
# valid) because pip reads PIP_CONSTRAINT from the first pip call onwards.
@@ -110,7 +110,7 @@ ENV BUILD_TRITON="0"
ENV BUILD_LLVM="0"
ENV BUILD_AITER_ALL="1"
ENV BUILD_MOONCAKE="1"
ENV AITER_COMMIT_DEFAULT="2bb549102ba2d9e51da8ef21107361827075d8b6"
ENV AITER_COMMIT_DEFAULT="285579b6211b0da30f35b91a890c29c823597cd1"
# ===============================
# Base image 950 with rocm720 and args
@@ -120,7 +120,7 @@ ENV BUILD_TRITON="1"
ENV BUILD_LLVM="0"
ENV BUILD_AITER_ALL="1"
ENV BUILD_MOONCAKE="1"
ENV AITER_COMMIT_DEFAULT="2bb549102ba2d9e51da8ef21107361827075d8b6"
ENV AITER_COMMIT_DEFAULT="285579b6211b0da30f35b91a890c29c823597cd1"
ENV TRITON_COMMIT_DEFAULT="42270451990532c67e69d753fbd026f28fcc4840"
# ===============================
@@ -131,7 +131,7 @@ ENV BUILD_TRITON="1"
ENV BUILD_LLVM="0"
ENV BUILD_AITER_ALL="1"
ENV BUILD_MOONCAKE="1"
ENV AITER_COMMIT_DEFAULT="2bb549102ba2d9e51da8ef21107361827075d8b6"
ENV AITER_COMMIT_DEFAULT="285579b6211b0da30f35b91a890c29c823597cd1"
# Pin the ROCm torch stack for every pip invocation in this flavor. The file is
# filled in after the torch 2.11 upgrade below; it must already exist (empty is
# valid) because pip reads PIP_CONSTRAINT from the first pip call onwards.
@@ -308,7 +308,7 @@ ENV BUILD_TRITON="0"
ENV BUILD_LLVM="0"
ENV BUILD_AITER_ALL="1"
ENV BUILD_MOONCAKE="1"
ENV AITER_COMMIT_DEFAULT="2bb549102ba2d9e51da8ef21107361827075d8b6"
ENV AITER_COMMIT_DEFAULT="285579b6211b0da30f35b91a890c29c823597cd1"
# Same reasoning as the rocm724 stages: keep pip from resolving the image's
# ROCm torch away to a PyPI CUDA build. Populated after the stack is in place.
ENV PIP_CONSTRAINT="/etc/sglang/constraints/torch-rocm.txt"
@@ -322,7 +322,7 @@ ENV BUILD_TRITON="0"
ENV BUILD_LLVM="0"
ENV BUILD_AITER_ALL="1"
ENV BUILD_MOONCAKE="1"
ENV AITER_COMMIT_DEFAULT="2bb549102ba2d9e51da8ef21107361827075d8b6"
ENV AITER_COMMIT_DEFAULT="285579b6211b0da30f35b91a890c29c823597cd1"
ENV PIP_CONSTRAINT="/etc/sglang/constraints/torch-rocm.txt"
RUN mkdir -p /etc/sglang/constraints && : > /etc/sglang/constraints/torch-rocm.txt
@@ -356,7 +356,12 @@ ENV BUILD_VLLM="0"
# Keep the Triton that comes with the SDK. The rocm1000 flavor builds a newer
# Triton from source, and that one no longer has block pointers or
# triton_kernels (triton-lang/triton#10833), which sglang still uses.
ENV BUILD_TRITON="0"
ENV BUILD_TRITON="1"
ENV TRITON_COMMIT_DEFAULT="76940ad348795521b3dc9f6c79acd7309ff924e3"
# triton-lang/triton#10833 removed block pointers, which sglang's FLA kernels
# (fla/, used by Qwen3-Next, MiniMax and Kimi Linear) still call. The revert is
# pure Python: #9668 had already moved block pointers out of C++/MLIR.
ENV TRITON_REVERT_COMMITS_DEFAULT="9debefe44117a44c274b05033cfe9e4858883b86"
ENV BUILD_LLVM="0"
ENV BUILD_AITER_ALL="1"
ENV BUILD_MOONCAKE="1"
@@ -399,6 +404,8 @@ ARG SETUPTOOLS_SCM_PRETEND_VERSION=""
ARG TRITON_REPO="https://github.com/triton-lang/triton.git"
ENV TRITON_COMMIT="${TRITON_COMMIT:-${TRITON_COMMIT_DEFAULT}}"
ARG TRITON_REVERT_COMMITS=""
ENV TRITON_REVERT_COMMITS="${TRITON_REVERT_COMMITS:-${TRITON_REVERT_COMMITS_DEFAULT}}"
# ROCm 7.2 Triton (BUILD_TRITON=1 stages only). Both wheels are the same
# upstream revision, triton-lang/triton@89002410. AITER only requires
@@ -440,7 +447,8 @@ ARG FHT_COMMIT="46efb7d776d38638fc39f3c803eaee3dd7016bd1"
# gfx950); disable with --build-arg ENABLE_SGL_AMD_MONO=0.
ARG ENABLE_SGL_AMD_MONO=1
ARG SGL_AMD_MONO_REPO="https://github.com/sgl-project/sgl-amd-mono.git"
ARG SGL_AMD_MONO_COMMIT="815b104f483d180027054a8f44bad0c2a3f895fd"
# empty: the commit pip installed through the monokernels_hip extra
ARG SGL_AMD_MONO_COMMIT=""
ARG ENABLE_MORI=0
ARG NIC_BACKEND=none
@@ -895,9 +903,11 @@ RUN /bin/bash -lc 'set -euo pipefail; \
RUN /bin/bash -lc 'set -euo pipefail; \
[ "${ENABLE_SGL_AMD_MONO}" = "1" ] || { echo "[sgl-amd-mono] skip (ENABLE_SGL_AMD_MONO=${ENABLE_SGL_AMD_MONO})"; exit 0; }; \
case "${GPU_ARCH}" in gfx950*) ;; *) echo "[sgl-amd-mono] skip (GPU_ARCH=${GPU_ARCH})"; exit 0 ;; esac; \
python3 -m pip show sgl-amd-mono > /dev/null || { echo "[sgl-amd-mono] skip (this SGLang does not install it)"; exit 0; }; \
commit="${SGL_AMD_MONO_COMMIT:-$(python3 -c "import json, importlib.metadata as md; print(json.loads(md.distribution(\"sgl-amd-mono\").read_text(\"direct_url.json\"))[\"vcs_info\"][\"commit_id\"])")}"; \
git clone "${SGL_AMD_MONO_REPO}" /sgl-workspace/sgl-amd-mono; \
cd /sgl-workspace/sgl-amd-mono; \
git checkout -f "${SGL_AMD_MONO_COMMIT}"; \
git checkout -f "${commit}"; \
python3 -m pip install --no-cache-dir --no-deps -e .; \
python3 -c "from importlib.metadata import entry_points; \
assert any(e.name == \"sgl_amd_mono\" for e in entry_points(group=\"sglang.srt.plugins\")), \"sgl-amd-mono plugin not registered\""'
@@ -1165,13 +1175,21 @@ RUN cd /tmp/whl \
# torch 2.11 names this `triton-rocm`; uninstall it so the pin is the only Triton.
RUN if [ "$BUILD_TRITON" = "1" ]; then \
case "${GPU_ARCH}" in \
gfx1250-rocm1000) \
gfx1250-rocm1000|gfx1250-rocm1010) \
echo "[Triton] gfx1250: building ${TRITON_COMMIT} from source"; \
pip uninstall -y triton triton-rocm || true \
&& apt-get update && apt-get install -y --no-install-recommends cmake && rm -rf /var/lib/apt/lists/* \
&& git clone ${TRITON_REPO} triton-custom \
&& cd triton-custom \
&& git checkout ${TRITON_COMMIT} \
&& for rev in ${TRITON_REVERT_COMMITS}; do \
echo "[Triton] reverting $rev"; \
git -c user.email=build@sglang -c user.name=sglang-build \
revert --no-commit "$rev" \
|| git checkout --ours -- Makefile; \
git -c user.email=build@sglang -c user.name=sglang-build \
commit -q -am "Revert $rev" || exit 1; \
done \
&& pip install -r python/requirements.txt \
&& pip install -e . \
&& if [ -d python/triton_kernels ]; then pip install -e python/triton_kernels --no-deps; fi; \
@@ -1187,7 +1205,7 @@ RUN if [ "$BUILD_TRITON" = "1" ]; then \
fi
# torch 2.11 still Requires-Dist: triton-rocm after the swap above.
RUN case "${GPU_ARCH}" in *-rocm724) python3 -c "import pathlib,re,importlib.metadata as m; p=pathlib.Path(m.distribution('torch')._path)/'METADATA'; v=m.version('triton'); t,n=re.subn(r'^Requires-Dist: (?:triton|triton-rocm)==[^ ;]+', 'Requires-Dist: triton=='+v, p.read_text(), count=1, flags=re.M); assert n==1, n; p.write_text(t)" ;; esac
RUN case "${GPU_ARCH}" in *-rocm724|gfx1250-rocm1010) python3 -c "import pathlib,re,importlib.metadata as m; p=pathlib.Path(m.distribution('torch')._path)/'METADATA'; v=m.version('triton'); t,n=re.subn(r'^Requires-Dist: (?:triton|triton-rocm)==[^\s;]+', 'Requires-Dist: triton=='+v, p.read_text(), count=1, flags=re.M); assert n==1, n; p.write_text(t)" ;; esac
# -----------------------
# Performance environment variable.
+37
View File
@@ -0,0 +1,37 @@
ARG BASE_IMG=pytorch/manylinux2_28-builder
# FlashMLA's kernels use `cvt` with `.scaled::n2::ue8m0`, which needs nvcc >= 13.1,
# and older toolkits emit suboptimal instructions for them. Build with 13.4.
ARG CUDA_VERSION=13.4
FROM ${BASE_IMG}:cuda${CUDA_VERSION}
ARG ARCH=x86_64
ARG CUDA_VERSION=13.4
ARG PYTHON_VERSION=3.12
ARG PYTHON_TAG=cp312-cp312
ARG TORCH_VER=2.14.1
ARG PIP_DEFAULT_INDEX=https://pypi.python.org/simple
ARG PYTORCH_MIRROR=download.pytorch.org
ENV PYTHON_ROOT_PATH=/opt/python/${PYTHON_TAG}
ENV PATH=${PYTHON_ROOT_PATH}/bin:${PATH}
RUN yum install -y --nogpgcheck git wget tar gcc gcc-c++ make \
&& yum clean all && rm -rf /var/cache/yum
# The build links -lcuda, which on a machine with no driver resolves against the
# toolkit's stub.
RUN set -eux; \
if [ "${ARCH}" = "aarch64" ]; then _LIB=sbsa; else _LIB="${ARCH}"; fi; \
mkdir -p /usr/lib/${ARCH}-linux-gnu/; \
ln -sf /usr/local/cuda-${CUDA_VERSION}/targets/${_LIB}-linux/lib/stubs/libcuda.so /usr/lib/${ARCH}-linux-gnu/libcuda.so
# torch is needed at build time (setup.py imports it) and is what the wheel pins.
# The CUDA minor version of this torch does not have to match CUDA_VERSION above:
# the extensions link libcudart.so.13, whose SONAME carries only the major, and
# use no API newer than 13.0.
RUN --mount=type=cache,id=sgl-flash-mla-pip,target=/root/.cache/pip \
set -eux; \
${PYTHON_ROOT_PATH}/bin/pip install torch==${TORCH_VER} --index-url https://${PYTORCH_MIRROR}/whl/cu130; \
${PYTHON_ROOT_PATH}/bin/pip install --index-url ${PIP_DEFAULT_INDEX} \
ninja setuptools wheel build numpy
+14 -14
View File
@@ -12,7 +12,7 @@
# manifests → cargo fetch → copy src" approach caches only the fetched
# registry; every source change still recompiles every dep.
#
# `experimental/sgl-router/Cargo.lock` is committed, and every cargo step
# `sgl-router/Cargo.lock` is committed, and every cargo step
# below runs with `--locked`, so the image ships exactly the dependency
# graph CI checked, tested, and license-audited.
#
@@ -32,9 +32,9 @@ ARG DEBIAN_VERSION=bookworm
######################## STAGE 1 — chef recipe ##########################
FROM rust:${RUST_VERSION}-${DEBIAN_VERSION} AS chef
RUN cargo install cargo-chef --locked --version ^0.1
WORKDIR /work/experimental/sgl-router
COPY experimental/sgl-router/Cargo.toml experimental/sgl-router/Cargo.lock ./
COPY experimental/sgl-router/sgl-kv-indexer/Cargo.toml sgl-kv-indexer/Cargo.toml
WORKDIR /work/sgl-router
COPY sgl-router/Cargo.toml sgl-router/Cargo.lock ./
COPY sgl-router/sgl-kv-indexer/Cargo.toml sgl-kv-indexer/Cargo.toml
# sglang-processor is a path dependency that inherits from the rust/ workspace manifest.
COPY rust/Cargo.toml /work/rust/Cargo.toml
COPY rust/sglang-api-types /work/rust/sglang-api-types
@@ -57,7 +57,7 @@ RUN apt-get update \
&& apt-get install -y --no-install-recommends protobuf-compiler \
&& rm -rf /var/lib/apt/lists/* \
&& cargo install cargo-chef --locked --version ^0.1
WORKDIR /work/experimental/sgl-router
WORKDIR /work/sgl-router
# `dynamo-tokenizers` pulls in `pcre2-sys`, whose build.rs links the SYSTEM
# libpcre2-8 whenever pkg-config finds it (it does here — the rust:bookworm
@@ -67,8 +67,8 @@ WORKDIR /work/experimental/sgl-router
# it statically, keeping the runtime self-contained.
ENV PCRE2_SYS_STATIC=1
COPY --from=chef /work/experimental/sgl-router/recipe.json ./recipe.json
COPY experimental/sgl-router/sgl-kv-indexer/Cargo.toml sgl-kv-indexer/Cargo.toml
COPY --from=chef /work/sgl-router/recipe.json ./recipe.json
COPY sgl-router/sgl-kv-indexer/Cargo.toml sgl-kv-indexer/Cargo.toml
# The recipe omits path dependencies outside this workspace, so cook against a
# stub of sglang-processor; its real sources arrive with the router's below.
COPY rust/Cargo.toml /work/rust/Cargo.toml
@@ -82,12 +82,12 @@ RUN mkdir -p /work/rust/sglang-processor/src && touch /work/rust/sglang-processo
RUN cargo chef cook --locked --release --recipe-path recipe.json
# Now bring in the real sources and the manifests they need.
COPY experimental/sgl-router/Cargo.toml experimental/sgl-router/Cargo.lock ./
COPY experimental/sgl-router/src ./src
COPY experimental/sgl-router/sgl-kv-indexer/Cargo.toml sgl-kv-indexer/Cargo.toml
COPY experimental/sgl-router/sgl-kv-indexer/build.rs sgl-kv-indexer/build.rs
COPY experimental/sgl-router/sgl-kv-indexer/proto sgl-kv-indexer/proto
COPY experimental/sgl-router/sgl-kv-indexer/src sgl-kv-indexer/src
COPY sgl-router/Cargo.toml sgl-router/Cargo.lock ./
COPY sgl-router/src ./src
COPY sgl-router/sgl-kv-indexer/Cargo.toml sgl-kv-indexer/Cargo.toml
COPY sgl-router/sgl-kv-indexer/build.rs sgl-kv-indexer/build.rs
COPY sgl-router/sgl-kv-indexer/proto sgl-kv-indexer/proto
COPY sgl-router/sgl-kv-indexer/src sgl-kv-indexer/src
COPY rust/sglang-processor/src /work/rust/sglang-processor/src
RUN touch sgl-kv-indexer/build.rs /work/rust/sglang-processor/src/lib.rs \
@@ -97,7 +97,7 @@ RUN touch sgl-kv-indexer/build.rs /work/rust/sglang-processor/src/lib.rs \
######################## STAGE 3 — runtime ##############################
FROM gcr.io/distroless/cc-debian12:nonroot AS runtime
COPY --from=builder /work/experimental/sgl-router/target/release/sgl-router /usr/local/bin/sgl-router
COPY --from=builder /work/sgl-router/target/release/sgl-router /usr/local/bin/sgl-router
# Default config path; mount your own via `-v <host-path>:/etc/sgl-router`.
ENV SGL_ROUTER_CONFIG=/etc/sgl-router/sgl-router.yaml
Binary file not shown.

After

Width:  |  Height:  |  Size: 11 KiB

@@ -1,7 +1,6 @@
---
title: Ornith-1.0
description: "Deploy DeepReinforce Ornith-1.0 with SGLang - a self-improving agentic-coding model family with 397B, 35B, and 9B checkpoints plus FP8 and GGUF variants."
tag: NEW
---
## Deployment
@@ -0,0 +1,172 @@
---
title: Ornith-1.5
description: "Deploy Ornith-1.5 with SGLang on NVIDIA H100 and H200 GPUs, with recipes for 9B, 35B-A3B, 122B, and 397B models, FP8, reasoning, and tool calling."
tag: NEW
---
import { Deployment } from "/src/snippets/_deployment.jsx";
import { config } from "/src/snippets/configs/ornith-ai/ornith-1.5.jsx";
## Deployment
These recipes target a single Linux node with NVIDIA CUDA GPUs. The Ornith-1.5 model cards recommend SGLang `>=0.5.9`. Both parsers are included in every launch command: `qwen3` separates reasoning from the answer, and `qwen3_coder` parses tool calls.
<Accordion title="Install SGLang">
<Tabs>
<Tab title="Python">
```bash Command
pip install --upgrade "sglang>=0.5.9"
```
Run the Python command from the deployment panel in this environment. See the [installation guide](/docs/get-started/install) for platform-specific requirements.
</Tab>
<Tab title="Docker">
```bash Command
docker pull lmsysorg/sglang:latest
```
Select Docker in the deployment panel to generate the container command.
</Tab>
</Tabs>
</Accordion>
Choose your hardware, model size, and checkpoint format. The 9B recipe uses one GPU, 35B-A3B uses two, and 122B and 397B use eight. BF16 397B requires H200 in this matrix; its FP8 checkpoint is also available on H100. Select **BF16** for a BF16 checkpoint or **FP8** for an FP8 checkpoint. All commands use Hugging Face repository IDs and default to port **30000**.
<Deployment config={config} />
## 1. Model introduction
Ornith-1.5 is an agentic model family published by [Ornith AI](https://huggingface.co/ornith-ai). The 9B, 35B-A3B, 122B, and 397B checkpoints use Qwen3.5 architectures. The family extends Ornith-1.0's self-improvement loop by jointly optimizing task generation, scaffold construction, and solution rollouts through reinforcement learning. The model cards describe applications in coding, reasoning, and tool-using agents.
The recipes expose an OpenAI-compatible API with separate reasoning output, parsed tool calls, and a 262,144-token context window. The table lists the checkpoints covered by this recipe.
<table style={{width: "100%", borderCollapse: "collapse"}}>
<thead>
<tr style={{borderBottom: "2px solid #d55816"}}>
<th style={{textAlign: "left", padding: "10px 12px"}}>Checkpoint</th>
<th style={{textAlign: "left", padding: "10px 12px"}}>Format</th>
<th style={{textAlign: "left", padding: "10px 12px"}}>Recipe</th>
</tr>
</thead>
<tbody>
<tr>
<td style={{padding: "9px 12px"}}><a href="https://huggingface.co/ornith-ai/Ornith-1.5-9B">Ornith-1.5-9B</a></td>
<td style={{padding: "9px 12px"}}>BF16</td>
<td style={{padding: "9px 12px"}}>Dense; 1× H100 or H200</td>
</tr>
<tr>
<td style={{padding: "9px 12px"}}><a href="https://huggingface.co/ornith-ai/Ornith-1.5-35B-A3B">Ornith-1.5-35B-A3B</a></td>
<td style={{padding: "9px 12px"}}>BF16</td>
<td style={{padding: "9px 12px"}}>MoE, ~3B active; 2× H100 or H200</td>
</tr>
<tr>
<td style={{padding: "9px 12px"}}><a href="https://huggingface.co/ornith-ai/Ornith-1.5-35B-A3B-FP8">Ornith-1.5-35B-A3B-FP8</a></td>
<td style={{padding: "9px 12px"}}>FP8</td>
<td style={{padding: "9px 12px"}}>2× H100 or H200</td>
</tr>
<tr>
<td style={{padding: "9px 12px"}}><a href="https://huggingface.co/ornith-ai/Ornith-1.5-122B">Ornith-1.5-122B</a></td>
<td style={{padding: "9px 12px"}}>BF16</td>
<td style={{padding: "9px 12px"}}>MoE, 256 experts with 8 routed per token; 8× H100 or H200</td>
</tr>
<tr>
<td style={{padding: "9px 12px"}}><a href="https://huggingface.co/ornith-ai/Ornith-1.5-397B">Ornith-1.5-397B</a></td>
<td style={{padding: "9px 12px"}}>BF16</td>
<td style={{padding: "9px 12px"}}>MoE; 8× H200</td>
</tr>
<tr>
<td style={{padding: "9px 12px"}}><a href="https://huggingface.co/ornith-ai/Ornith-1.5-397B-FP8">Ornith-1.5-397B-FP8</a></td>
<td style={{padding: "9px 12px"}}>FP8</td>
<td style={{padding: "9px 12px"}}>8× H100 or H200</td>
</tr>
</tbody>
</table>
The official [9B](https://huggingface.co/ornith-ai/Ornith-1.5-9B), [35B-A3B](https://huggingface.co/ornith-ai/Ornith-1.5-35B-A3B), [122B](https://huggingface.co/ornith-ai/Ornith-1.5-122B), and [397B](https://huggingface.co/ornith-ai/Ornith-1.5-397B) model cards declare the MIT license.
**Resources:** [Ornith-1.5 model repositories](https://huggingface.co/ornith-ai/models?search=Ornith-1.5) · [Ornith-1.5 blog](https://ornith.ai/ornith_1_5.html)
## 2. Configuration tips
- **Model IDs:** use the `ornith-ai` namespace. The 35B repository includes the `-A3B` suffix. FP8 selections use the actual `-FP8` repository rather than the BF16 repository shown in their model-card quickstarts.
- **Context and memory:** every recipe uses `--context-length 262144`. The 9B, 35B-A3B, and 397B recipes use `--mem-fraction-static 0.85`; the 122B recipe uses `0.80`. Lower the context length if you need more memory headroom.
- **Reasoning:** `--reasoning-parser qwen3` exposes reasoning as `message.reasoning_content`, with the answer in `message.content`.
- **Tool calling:** `--tool-call-parser qwen3_coder` exposes parsed calls in `message.tool_calls`. Your application executes the requested function and submits its result in a follow-up tool message.
- **Sampling:** 35B-A3B and 397B recommend `temperature=0.6`, `top_p=0.95`, and `top_k=20`. For 9B and 122B, the model cards recommend `temperature=1.0` and `presence_penalty=1.5` for general tasks, or `temperature=0.6` and `presence_penalty=0.0` for precise coding; both use `top_p=0.95`, `top_k=20`, `min_p=0.0`, and `repetition_penalty=1.0`. These settings are distinct from benchmark-specific settings. The Python examples below use the precise-coding sampling settings.
## 3. Usage examples
These requests use the precise-coding sampling settings shared by 9B and 122B. Set `model` to the deployment panel's `--served-model-name`, such as `Ornith-1.5-9B` or `Ornith-1.5-122B`. The examples use the default port; match the client port if you change the launch port.
<Accordion title="Chat completion with reasoning (Python)">
```python Example
from openai import OpenAI
client = OpenAI(base_url="http://localhost:30000/v1", api_key="EMPTY")
response = client.chat.completions.create(
model="Ornith-1.5-9B",
messages=[{"role": "user", "content": "Write a compact Python is_prime function."}],
temperature=0.6,
top_p=0.95,
presence_penalty=0.0,
max_tokens=1024,
extra_body={"top_k": 20, "min_p": 0.0, "repetition_penalty": 1.0},
)
message = response.choices[0].message
print("Reasoning:", getattr(message, "reasoning_content", None))
print("Answer:", message.content)
```
</Accordion>
<Accordion title="Tool-call request (Python)">
```python Example
from openai import OpenAI
client = OpenAI(base_url="http://localhost:30000/v1", api_key="EMPTY")
response = client.chat.completions.create(
model="Ornith-1.5-9B",
messages=[{"role": "user", "content": "Run the parser module's unit tests."}],
tools=[{
"type": "function",
"function": {
"name": "run_tests",
"description": "Run the project's test suite.",
"parameters": {
"type": "object",
"properties": {"target": {"type": "string"}},
"required": ["target"],
},
},
}],
tool_choice="auto",
temperature=0.6,
top_p=0.95,
presence_penalty=0.0,
max_tokens=2048,
extra_body={"top_k": 20, "min_p": 0.0, "repetition_penalty": 1.0},
)
message = response.choices[0].message
print("Reasoning:", getattr(message, "reasoning_content", None))
print("Tool calls:", message.tool_calls)
```
</Accordion>
### Agent integration
Configure your agent or coding CLI to use the same OpenAI-compatible endpoint:
- **API base URL:** `http://localhost:30000/v1`.
- **API key:** `EMPTY` for a local server without API-key authentication.
- **Model:** the selected checkpoint's `--served-model-name`, such as `Ornith-1.5-122B`.
Use the served model name rather than the Hugging Face repository ID when the two differ. Keep both parsers enabled so the agent receives separate reasoning and structured tool calls. Your agent executes tools and returns their results to the server in subsequent messages.
@@ -26,8 +26,6 @@ For more details, please refer to the [official DeepSeek-OCR repository](https:/
Please refer to the [official SGLang installation guide](../../../docs/get-started/install) for installation instructions.
For SGLang CPU installation, please refer to the [CPU version installation guide](../../../docs/hardware-platforms/cpu_server#installation).
## 3. Model Deployment
This section provides deployment configurations optimized for different hardware platforms and use cases.
@@ -37,8 +37,6 @@ For more details, please refer to the [official DeepSeek-R1 repository](https://
Please refer to the [official SGLang installation guide](../../../docs/get-started/install) for installation instructions.
For SGLang CPU installation, please refer to the [CPU version installation guide](../../../docs/hardware-platforms/cpu_server#installation).
## 3. Model Deployment
This section provides deployment configurations optimized for different hardware platforms and use cases.
@@ -22,8 +22,6 @@ SGLang offers multiple installation methods. You can choose the most suitable in
Please refer to the [official SGLang installation guide](../../../docs/get-started/install) for installation instructions.
For SGLang CPU installation, please refer to the [CPU version installation guide](../../../docs/hardware-platforms/cpu_server#installation).
## 3. Model Deployment
This section provides a progressive guide from quick deployment to performance optimization, suitable for users at different levels.
@@ -18,8 +18,6 @@ SGLang offers multiple installation methods. You can choose the most suitable in
Please refer to the [official SGLang installation guide](../../../docs/get-started/install) for installation instructions.
For SGLang CPU installation, please refer to the [CPU version installation guide](../../../docs/hardware-platforms/cpu_server#installation).
## 3. Model Deployment
This section provides a progressive guide from quick deployment to performance optimization, suitable for users at different levels.
@@ -51,11 +51,11 @@ docker run --gpus all \
AMD uses the daily-updated `lmsysorg/sglang-rocm` images. You can find the latest images on [Docker Hub](https://hub.docker.com/r/lmsysorg/sglang-rocm/tags). We recommend the ROCm 7.2 version.
For example:
- **MI355X** → `lmsysorg/sglang-rocm:v0.5.21-rocm720-mi35x-20261007`
- **MI300X** → `lmsysorg/sglang-rocm:v0.5.21-rocm720-mi30x-20261007`
- **MI355X** → `lmsysorg/sglang-rocm:v0.5.22-rocm720-mi35x-20261016`
- **MI300X** → `lmsysorg/sglang-rocm:v0.5.22-rocm720-mi30x-20261016`
```bash Command
docker pull lmsysorg/sglang-rocm:v0.5.21-rocm720-{mi35x,mi30x}-20261007
docker pull lmsysorg/sglang-rocm:v0.5.22-rocm720-{mi35x,mi30x}-20261016
docker run \
--device=/dev/kfd --device=/dev/dri \
@@ -65,7 +65,7 @@ docker run \
-p 30000:30000 \
-v ~/.cache/huggingface:/root/.cache/huggingface \
--env "HF_TOKEN=<your-hf-token>" \
lmsysorg/sglang-rocm:v0.5.21-rocm720-{mi35x,mi30x}-20261007 \
lmsysorg/sglang-rocm:v0.5.22-rocm720-{mi35x,mi30x}-20261016 \
sglang serve <use args below>
```
@@ -77,7 +77,7 @@ docker run \
For **MI355X → Pro / Pro Official → FP4 → Multi-Nodes**, the panel generates a **1P1D deployment**: one prefill node and one decode node, each running an independent worker. Select **Router**, **Prefill**, or **Decode** to copy its command. Start the [UMBP tier](#3-8-unified-cache-external-linker-umbp) on the prefill node, then both workers, then the router. Send client requests to the router on port `8000`.
The MI355X multi-node strategies follow the [InferenceX AgentX submission](https://github.com/SemiAnalysisAI/InferenceX/pull/3664): **Low-Latency**, **Balanced**, and **High-Throughput** use fixed concurrency presets of **16**, **48**, and **256**, respectively. **Pro** uses EAGLE/MTP; **Pro Official (0813)** uses DSpark, matching the InferenceX checkpoint. Both enable UMBP on prefill by default. The Playground inherits the selected worker, including its speculative decoder and cache settings.
The MI355X multi-node strategies **Low-Latency**, **Balanced**, and **High-Throughput** use fixed concurrency presets of **16**, **48**, and **256**, respectively. **Pro** uses EAGLE/MTP; **Pro Official (0813)** uses DSpark. Both enable UMBP on prefill by default. The Playground inherits the selected worker, including its speculative decoder and cache settings.
Run these recipes' **Python** commands inside the ROCm image shown in the **Docker** tab, where Aiter is already enabled. Add explicit parser and network-device settings through the Playground's **Parsers** and **PD Disagg → IB Device** controls.
@@ -773,13 +773,13 @@ Four adjustments are not in the commands below. Add them when they apply:
- `--cuda-graph-backend-prefill breakable` at concurrency 1–4.
- `--enable-prefill-cp --cp-strategy interleave --enable-cp-decode-attn-tp` on the TP-only commands at concurrency 16 and above. Prefill context parallelism splits each prompt across all eight ranks (CP size = TP size).
- `SGLANG_DSV4_GROUPED_ASM=1` on the TP-only commands at concurrency 16 and above — the kernel needs at least 16 requests per draft token.
- `--prefill-decode-interval 15` on the MegaMoE command at concurrency 384, replacing the `20` it ships with.
- At concurrency 384 on the MegaMoE command: `--prefill-decode-interval 20`, `SGLANG_AMD_FLYDSL_MEGA_MOE_DECODE_MTPR=1024`, and `AITER_MEGA_FIXED_SLOT_MAX_MTPR=2047`.
**TP8, concurrency 1–16:**
```bash Command
SGLANG_OPT_UNIFIED_CACHE_FREE_OUT_OF_WINDOW_SLOTS=1 \
SGLANG_ENABLE_UNIFIED_RADIX_TREE=1 \
SGLANG_DSV4_UNIFIED_KV_FP8=1 \
SGLANG_HIP_DSV4_C4_INDEXER_ROWGROUP=1 \
SGLANG_USE_ROCM700A=0 \
TORCH_BLAS_PREFER_HIPBLASLT=1 \
SGLANG_HACK_FLASHMLA_BACKEND=unified_kv_triton \
@@ -819,8 +819,8 @@ python3 -m sglang.launch_server \
**DP8 (DP Attention), concurrency 128:**
```bash Command
SGLANG_OPT_UNIFIED_CACHE_FREE_OUT_OF_WINDOW_SLOTS=1 \
SGLANG_ENABLE_UNIFIED_RADIX_TREE=1 \
SGLANG_DSV4_UNIFIED_KV_FP8=1 \
SGLANG_HIP_DSV4_C4_INDEXER_ROWGROUP=1 \
SGLANG_USE_ROCM700A=0 \
TORCH_BLAS_PREFER_HIPBLASLT=1 \
SGLANG_HACK_FLASHMLA_BACKEND=unified_kv_triton \
@@ -851,7 +851,7 @@ python3 -m sglang.launch_server \
--kv-cache-dtype fp8_e4m3 \
--enforce-shared-experts-fusion \
--mem-fraction-static 0.92 \
--chunked-prefill-size 65536 \
--chunked-prefill-size 131072 \
--tool-call-parser deepseekv4 \
--reasoning-parser deepseek-v4 \
--speculative-algorithm DSPARK \
@@ -864,13 +864,13 @@ python3 -m sglang.launch_server \
--watchdog-timeout 3600
```
`--chunked-prefill-size` is a global budget divided by `--attn-dp-size`, so 65536 keeps 8192 tokens per rank. `--enable-dp-lm-head` is required for DSpark under DP attention. Front the DP ranks with `sglang_router --policy cache_aware`, so a multi-turn session lands on the rank already holding its prefix; above concurrency 160 add `--balance-abs-threshold 32`. Both apply to the MegaMoE command below as well.
`--chunked-prefill-size` is a global budget divided by `--attn-dp-size`, so 131072 keeps 16384 tokens per rank. `--enable-dp-lm-head` is required for DSpark under DP attention. Front the DP ranks with `sglang_router --policy cache_aware`, so a multi-turn session lands on the rank already holding its prefix; above concurrency 160 add `--balance-abs-threshold 32`. Both apply to the MegaMoE command below as well.
**DP8 EP8 MegaMoE, concurrency 128–384:**
```bash Command
SGLANG_OPT_UNIFIED_CACHE_FREE_OUT_OF_WINDOW_SLOTS=1 \
SGLANG_ENABLE_UNIFIED_RADIX_TREE=1 \
SGLANG_DSV4_UNIFIED_KV_FP8=1 \
SGLANG_HIP_DSV4_C4_INDEXER_ROWGROUP=1 \
SGLANG_USE_ROCM700A=0 \
TORCH_BLAS_PREFER_HIPBLASLT=1 \
SGLANG_HACK_FLASHMLA_BACKEND=unified_kv_triton \
@@ -878,12 +878,15 @@ AITER_BF16_FP8_MOE_BOUND=0 \
SGLANG_OPT_USE_AITER_BATCHED_GEMM=1 \
SGLANG_OPT_FP8_WO_A_FUSED_INVROPE=1 \
SGLANG_AMD_USE_FLYDSL_MEGA_MOE=1 \
SGLANG_AMD_FLYDSL_MEGA_MOE_MTPR=8192 \
SGLANG_AMD_FLYDSL_MEGA_MOE_MTPR=16384 \
SGLANG_AMD_FLYDSL_MEGA_MOE_DECODE_MTPR=512 \
AITER_MEGA_FIXED_SLOT_MAX_MTPR=1023 \
SGLANG_AMD_FLYDSL_MEGA_QUANT=a8w4 \
MORI_SHMEM_HEAP_SIZE=17179869184 \
SGLANG_AITER_MEGA_RANK_SYNC=1 \
SGLANG_AITER_MEGA_EPLB_PREFILL_ONLY=1 \
SGLANG_AITER_MEGA_EPLB_FUSED_MAP_RECORD=1 \
SGLANG_EPLB_P2P_BATCH_CHUNK_SIZE=7 \
SGLANG_DP_USE_REDUCE_SCATTER=0 \
GPU_MAX_HW_QUEUES=5 \
python3 -m sglang.launch_server \
@@ -897,7 +900,7 @@ python3 -m sglang.launch_server \
--enable-dp-attention-local-control-broadcast \
--tokenizer-worker-num 8 \
--stream-interval 20 \
--prefill-decode-interval 20 \
--prefill-decode-interval 40 \
--prefill-delayer-token-usage-low-watermark 0.7 \
--load-balance-method total_requests \
--attention-backend dsv4 \
@@ -912,7 +915,7 @@ python3 -m sglang.launch_server \
--eplb-rebalance-num-iterations 200 \
--expert-distribution-recorder-mode stat \
--mem-fraction-static 0.85 \
--chunked-prefill-size 65536 \
--chunked-prefill-size 131072 \
--tool-call-parser deepseekv4 \
--reasoning-parser deepseek-v4 \
--speculative-algorithm DSPARK \
@@ -925,7 +928,7 @@ python3 -m sglang.launch_server \
--watchdog-timeout 3600
```
This command adds expert parallelism on top of the DP8 one. [§3.7](#3-7-pd-disaggregation-on-mi355x-mori-io) explains `SGLANG_AMD_USE_FLYDSL_MEGA_MOE`, `SGLANG_AMD_FLYDSL_MEGA_MOE_MTPR` and the MORI symmetric heap; the heap is 16 GiB here rather than 8 GiB because `--mem-fraction-static 0.85` leaves more GPU memory free. `SGLANG_DP_USE_REDUCE_SCATTER=0` turns off the reduce-scatter path that MegaMoE's all-to-all replaces. The three `SGLANG_AITER_MEGA_*` variables and `--enable-eplb` enable expert-load balancing, recorded during prefill only.
This command adds expert parallelism on top of the DP8 one. [§3.7](#3-7-pd-disaggregation-on-mi355x-mori-io) explains `SGLANG_AMD_USE_FLYDSL_MEGA_MOE`, `SGLANG_AMD_FLYDSL_MEGA_MOE_MTPR` and the MORI symmetric heap; the heap is 16 GiB here rather than 8 GiB because `--mem-fraction-static 0.85` leaves more GPU memory free. `SGLANG_DP_USE_REDUCE_SCATTER=0` turns off the reduce-scatter path that MegaMoE's all-to-all replaces. The three `SGLANG_AITER_MEGA_*` variables and `--enable-eplb` enable expert-load balancing, recorded during prefill only. `SGLANG_EPLB_P2P_BATCH_CHUNK_SIZE=7` keeps each RCCL P2P group under 31 ops during EPLB rebalance.
### 3.7 PD Disaggregation on MI355X (MORI-IO)
@@ -944,17 +947,17 @@ docker run \
--ulimit nofile=1048576:1048576 \
--network host --ipc=host --shm-size 32g \
-v ~/.cache/huggingface:/root/.cache/huggingface \
lmsysorg/sglang-rocm:v0.5.21-rocm724-mi35x-20261001 \
lmsysorg/sglang-rocm:v0.5.22-rocm720-mi35x-20261019 \
python3 -m sglang.launch_server <role args below>
```
The Deployment and Playground Docker outputs retain this `v0.5.21` base image for the MI355X Pro and Pro Official FP4 multi-node recipes. `--network host` replaces the single-node `-p` mapping because the two roles run on different hosts.
Use this image for the **MI355X Pro / Pro Official FP4 multi-node** workers and router, including the additional concurrency points below. The Deployment and Playground Docker outputs select it for all three PD roles. It supports the documented `SGLANG_DSV4_DENSE_MXSCALE` and `SGLANG_ROUTER_P2C_LOAD_METRIC` settings. `--network host` replaces the single-node `-p` mapping because the two roles run on different hosts.
For the **MI355X Pro / Pro Official FP4** presets, **High-Throughput** sets `MORI_IO_QP_MAX_SEND_WR=32767` and `MORI_IO_QP_MAX_CQE=32768` on both MORI roles. Low-Latency and Balanced use the transport's default queue sizes. The send-queue backoff timeout uses MORI's default of 5 seconds.
#### Role-based configuration
Select **MI355X → Pro Official → FP4 → Multi-Nodes**, then a strategy and role. The commands use the [submitted recipe at `df5e433239d7`](https://github.com/SemiAnalysisAI/InferenceX/blob/df5e433239d72106529ece4dd97cc36e4c3adf4c/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/dsv4/sglang/mi355x-fp4/agentx/disagg-umbp-dspark.yaml), with cluster-specific paths replaced by local placeholders and benchmark acceptance simulation removed. The Pro Official commands additionally apply the kernel and routing updates described below.
Select **MI355X → Pro Official → FP4 → Multi-Nodes**, then a strategy and role. Replace the host and model-path placeholders with your deployment's values. The Pro Official commands use the kernel and routing settings described below.
| Strategy | Concurrency | Prefill → Decode | `--max-running-requests` | DSpark block size |
|---|---|---|---|---|
@@ -975,7 +978,7 @@ Selecting a TP-only role forces **DP-Attention** off and greys the control. `--m
</Note>
<Warning>
**Speculative decoding changes how you size `--max-running-requests`.** For a target concurrency of N, set it to `N*2` on **both** roles; at N the served batch stays below the concurrency you are aiming for. This holds for DSpark and EAGLE/MTP alike. The table above follows it: 32 serves a concurrency of 16, 96 serves 48, and 512 serves the high-throughput concurrency of 256.
**Speculative decoding changes how you size `--max-running-requests`.** For the strategy presets above, a target concurrency of N uses `N*2` on **both** roles; at N the served batch stays below the concurrency you are aiming for. This holds for DSpark and EAGLE/MTP alike. The table above follows it: 32 serves a concurrency of 16, 96 serves 48, and 512 serves the high-throughput concurrency of 256. The Pro Official c1 override below is an exception: retain 8 request slots on each role even though client concurrency is 1.
The decode role's `--cuda-graph-bs-decode` then has to cover `N*2 / attn_dp_size`, the per-rank batch. The strategy presets apply both rules: decode graphs cover 1–32, 1–96, and 1–64, respectively.
</Warning>
@@ -1137,7 +1140,7 @@ The multi-node router presets retain the default request timeout. The standalone
The panel keeps its c16 / c48 / c256 strategy presets. Apply the following overrides to the **Multi-Nodes** commands to reproduce the optimized recipe's endpoints. Here `cN` means a target concurrency of N.
**c1 (replaces c4):** start from Low-Latency. Keep prefill TP4, set decode to `--tp 4`, and set `--max-running-requests 2` on both roles. Replace decode's graph list with `--cuda-graph-bs-decode 1 2`. Keep DSpark block size 6 (7 draft tokens), `MORI_MAX_DISPATCH_TOKENS_DECODE=448`, and both kernel switches on both roles. The source also sets `SGLANG_MLA_HCA_KV_SPLITS=8` on both roles; the FP8 KV path does not read this setting. Benchmark with concurrency 1.
**c1 (replaces c4):** start from Low-Latency. Keep prefill TP4, set decode to `--tp 4`, and set `--max-running-requests 8` on both roles. Replace decode's graph list with `--cuda-graph-bs-decode 1 2 3 4 5 6 7 8`. Keep DSpark block size 6 (7 draft tokens), `MORI_MAX_DISPATCH_TOKENS_DECODE=448`, and both kernel switches on both roles. The source also sets `SGLANG_MLA_HCA_KV_SPLITS=8` on both roles; the FP8 KV path does not read this setting. Keep these 8 server slots and graph sizes when benchmarking with client concurrency 1; do not reduce them to 2 using the strategy presets' `N*2` rule.
**c32:** start from Low-Latency, keep TP4 → TP8 and block size 6, set both admission limits to 64, and capture decode graph sizes 1–64. Both roles use FP8 KV and mxscale, as at c16 and c48.
@@ -42,7 +42,7 @@ docker run --gpus all \
**AMD GPUs (ROCm)** — AMD uses the daily-updated `lmsysorg/sglang-rocm` images. You can find the latest images on [Docker Hub](https://hub.docker.com/r/lmsysorg/sglang-rocm/tags). We recommend the ROCm 10 version.
```bash Command
docker pull lmsysorg/sglang-rocm:v0.5.22-rocm10-mi35x-20261016
docker pull lmsysorg/sglang-rocm:v0.5.22-rocm10-mi35x-20261021
docker run \
--device=/dev/kfd --device=/dev/dri \
@@ -52,7 +52,7 @@ docker run \
-p 30000:30000 \
-v ~/.cache/huggingface:/root/.cache/huggingface \
--env "HF_TOKEN=<your-hf-token>" \
lmsysorg/sglang-rocm:v0.5.22-rocm10-mi35x-20261016 \
lmsysorg/sglang-rocm:v0.5.22-rocm10-mi35x-20261021 \
sglang serve <use args below>
```
@@ -105,7 +105,7 @@ import { agentx } from "/src/snippets/agentx/deepseek-ai/deepseek-v4_1.jsx";
**Engram.** An additive n-gram hash memory at two layers. Token ids are normalized *before* hashing, so `" The"`, `"the"` and `"THE"` cannot fork into separate rows. Its two fp8 tables are the largest single block of weight in the checkpoint, and by default they load row-sharded across the TP group, which costs an all-reduce per engram layer. `SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE=1` (**opt-in**) moves them to host memory instead: the freed HBM goes to the KV pool and output is bitwise unchanged — at the cost of host RAM, a longer load, and needing huge-page backing to keep the gather cheap. `SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT` trades the two all-reduces for page size: `shared` (the default) maps a single copy all ranks read and needs no reduce; `per_rank` splits that same table into one shard per rank and keeps the reduce, but is the only one the kernel reliably backs with 2 MiB pages.
**SWA bounded replay.** Decoder SWA bounded replay runs the early layers over the whole extend and the late layers over only each request’s last 128 tokens, sharing late-layer KV from source layer 20. It defaults on for supported standalone CUDA/HIP configurations without context parallelism. Encoder SWA replay can be combined with decoder replay, subject to encoder-specific validation. Eager and breakable prefill are supported; full prefill graphs keep it disabled. With DP attention, automatic enablement requires attention TP 1 and the `none` or `megamoe` MoE backend on CUDA, or `none` on HIP. Use `--enable-decoder-swa-bounded-replay` to explicitly enable replay, subject to compatibility validation. Bounded replay is not numerically equivalent to full prefill. A request for prompt logprobs or for every hidden state runs its whole prompt through the late layers with full sliding windows; other requests, including their output-token logprobs, use the bounded computation. Use `--no-enable-decoder-swa-bounded-replay` to retain full prefill, for example for RL rollouts whose logprobs must match the trainer.
**SWA bounded replay.** Decoder SWA bounded replay runs the early layers over the whole extend and the late layers over only each request’s last 128 tokens, sharing late-layer KV from source layer 20. It defaults on for supported CUDA/HIP configurations without context parallelism, including both PD roles on CUDA. Encoder SWA replay can be combined with decoder replay, subject to encoder-specific validation. Eager and breakable prefill are supported; full prefill graphs keep it disabled. With DP attention, automatic enablement requires attention TP 1 and the `none` or `megamoe` MoE backend on CUDA, or `none` on HIP. Use `--enable-decoder-swa-bounded-replay` to explicitly enable replay, subject to compatibility validation. Bounded replay is not numerically equivalent to full prefill. A request for prompt logprobs or for every hidden state runs its whole prompt through the late layers with full sliding windows; other requests, including their output-token logprobs, use the bounded computation. Use `--no-enable-decoder-swa-bounded-replay` to retain full prefill, for example for RL rollouts whose logprobs must match the trainer.
**Kernels.** At decode token counts these kernels are launch-bound rather than bandwidth-bound, so the work is fusion and overlap: the compressor projection, ratio-2 pooling, RoPE with fp4 quantization and indexer packing, paged fp4 indexer scoring with top-k resolved straight to cache addresses, and the mHC statistics and Sinkhorn — with compression, indexing and mHC overlapped against attention and FFN on dedicated streams. All of it is selected automatically; the environment switches it once sat behind were removed, so there is nothing to turn on. One caveat outranks the fusions: **output is not bitwise stable across batch composition today** — two default kernels are shape-guarded to a single token — and `--enable-deterministic-inference` is refused on this backend.
@@ -202,6 +202,8 @@ What to expect:
Prefill/decode disaggregation is validated token-identical against a single server across 32 greedy prompts, with GSM8K matching through the router. Use the **PD Disaggregation** card in the Playground to generate the prefill role, the decode role and the router command.
Both roles run MegaAttention (`--dsv4-attn-backend flashmla_mega`) by default on SM100/SM103, as a single server does. Prefill and decode must store KV in the same layout, so they must agree on `--dsv4-attn-backend` and `SGLANG_DSV4_KV_LAYOUT`; a decode server paired with a prefill whose layout differs fails its first request with a `KV layout mismatch` error. A decode on a GPU other than SM100/SM103 therefore needs `--dsv4-attn-backend flashmla` on the prefill too.
One deployment note: Mooncake needs the RDMA fabric visible inside the container, so launch with `--device /dev/infiniband:/dev/infiniband --cap-add IPC_LOCK --ulimit memlock=-1`. Without it Mooncake selects its NVLink transport, which only serves buffers from its own allocator and fails with `Requested address ... not found`. If you hit that, force TCP with `MOONCAKE_PROTOCOL=tcp` and `MC_FORCE_TCP=1` — the Playground's Mooncake option sets both.
PD runs with DSpark: on a DSpark recipe both Playground roles keep the speculative flags, and the prefill and decode servers must both pass them. The server accepts this with Mooncake or NIXL transfer, static verification (`SGLANG_RAGGED_VERIFY_MODE`, `static` by default), and no context parallelism on decode (CP=1, DCP=1); it rejects any other combination at startup.
@@ -222,14 +224,14 @@ Only the L2 tier is exposed here. For the storage (L3) tier and the canonical fl
Agentic replay is a different workload: long-context prompts, multi-turn sessions that reuse their own prefix across turns, and subagent branches that fan out within one session. Concurrency below counts live sessions, not requests. Every command keeps the engram tables on the host, and none adds a HiCache tier. The sessions fit the device pool.
Two commands follow, one per concurrency range. TP2 covers concurrency 1 to 32 and TP4 covers 1 to 16; at 32 and 64 the same two GPUs also run DP attention with TP MoE.
TP2 covers concurrency 1 to 32 and TP4 covers 1 to 16; at 32 and 64 the same two GPUs also run DP attention with TP MoE.
**TP2 or TP4, concurrency 1–32:**
**TP2 or TP4, concurrency 1–16:**
- TP2 raises `--mem-fraction-static` from concurrency 16 and continues through 32; TP4 stops at 16 and holds 0.70, with a quarter of the checkpoint per rank leaving room to spare.
- TP2 raises `--mem-fraction-static` at concurrency 16; TP4 stops at 16 and holds 0.70, with a quarter of the checkpoint per rank leaving room to spare.
- A longer `--prefill-decode-interval` runs more decode rounds between prefill batches, trading TTFT for P90 inter-token latency.
- Up to concurrency 8 it pairs with `--schedule-policy lpm`, which admits the requests with the longest cached prefix first.
- Concurrency ≤ 16 uses the same low-latency path as §3's Low-Latency card (`--monokernel-backend amd` and `--disable-shared-experts-fusion`, as in the command below); at concurrency 32 swap those two flags for `--enforce-shared-experts-fusion`.
- These points set `--monokernel-backend amd`. Do not set `SGLANG_HIP_DSV41_INDEX_K_PAGE8`: the monokernel refuses that layout.
The braced values in the command follow concurrency:
@@ -250,13 +252,11 @@ SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT=per_rank \
SGLANG_USE_AITER=1 \
SGLANG_USE_ROCM700A=0 \
TORCH_BLAS_PREFER_HIPBLASLT=1 \
AITER_BF16_FP8_MOE_BOUND=0 \
TRITON_HIP_USE_ASYNC_COPY=0 \
ROCM_QUICK_REDUCE_QUANTIZATION=INT8 \
SGLANG_OPT_HIP_OPUS_SPARSE_PREFILL=true \
SGLANG_OPT_HIP_FLYDSL_MXFP8=1 \
SGLANG_HIP_FP4_INDEXER_ROWGROUP=1 \
SGLANG_HIP_DSV41_INDEX_K_PAGE8=1 \
HSA_NO_SCRATCH_RECLAIM=0 \
GPU_MAX_HW_QUEUES=2 \
sglang serve \
@@ -267,8 +267,9 @@ sglang serve \
--speculative-algorithm DSPARK \
--speculative-dspark-block-size 5 \
--monokernel-backend amd \
--disable-shared-experts-fusion \
--enforce-shared-experts-fusion \
--fp8-gemm-backend aiter \
--enable-decoder-swa-bounded-replay \
--cuda-graph-backend-prefill disabled \
--mem-fraction-static {mem-fraction-static} \
--chunked-prefill-size 16384 \
@@ -282,6 +283,8 @@ sglang serve \
--watchdog-timeout 3600
```
**TP2, concurrency 32:** the command above with the concurrency-32 column, without `--monokernel-backend amd`, and with `SGLANG_HIP_DSV41_INDEX_K_PAGE8=1`.
**DP2, concurrency 32–64:** two DP attention ranks, one per GPU, with the MoE still tensor parallel across both. The braced values follow concurrency:
| Concurrency | 32 | 64 |
@@ -298,7 +301,6 @@ SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT=per_rank \
SGLANG_USE_AITER=1 \
SGLANG_USE_ROCM700A=0 \
TORCH_BLAS_PREFER_HIPBLASLT=1 \
AITER_BF16_FP8_MOE_BOUND=0 \
TRITON_HIP_USE_ASYNC_COPY=0 \
ROCM_QUICK_REDUCE_QUANTIZATION=INT8 \
SGLANG_OPT_HIP_OPUS_SPARSE_PREFILL=true \
@@ -327,6 +329,7 @@ sglang serve \
--speculative-dspark-block-size 5 \
--enforce-shared-experts-fusion \
--fp8-gemm-backend aiter \
--enable-decoder-swa-bounded-replay \
--cuda-graph-backend-prefill disabled \
--mem-fraction-static {mem-fraction-static} \
--chunked-prefill-size 32768 \
@@ -15,7 +15,7 @@ ERNIE-4.5 delivers advanced features as below:
- **Vision Encoder**: Dedicated adaptive-resolution ViT with 2D RoPE and image packing; for video, adaptive frame sampling and timestamp rendering, supporting both shared and modality-specific visual processing.
- **Adapter**: Shared modality-bridging module with spatial and temporal compression to align vision to text embedding space, enabling cross-modal understanding without compromising text representations.
- **Multimodal Position Embedding**: Unified 3D RoPE (temporal, height, width) for vision and 1D RoPE for text in a single embedding space, supporting parameter sharing while encoding modality-specific positions.
- **Hardware Optimization**: Specifically tuned for AMD MI300X, MI325X, and MI355X GPUs.
- **Hardware Optimization**: Specifically tuned for AMD MI300X, MI325X, MI355X GPUs and Intel Xeon CPUs.
## 2. SGLang Installation
+50 -1
View File
@@ -1,6 +1,6 @@
---
title: GLM-5.2
description: "Deploy GLM-5.2 with SGLang — Z.ai's DeepSeek-Sparse-Attention (DSA) Mixture-of-Experts model with MTP speculative decoding and 1M context, on H200, B200, B300, GB300, and AMD MI300X/MI325X/MI355X."
description: "Deploy GLM-5.2 with SGLang — Z.ai's DeepSeek-Sparse-Attention (DSA) Mixture-of-Experts model with MTP speculative decoding and 1M context, on H200, B200, B300, GB300, and AMD MI300X/MI325X/MI355X, plus Ascend NPU recipes."
---
## Deployment
@@ -39,6 +39,55 @@ For how to launch the image, see [Install → Method 3: Using Docker](../../../d
</Accordion>
<Accordion title="Install and prepare Ascend NPUs">
Use a Linux host with the Ascend driver and a compatible CANN / TorchNPU environment. Follow the
[Ascend installation guide](/docs/hardware-platforms/ascend-npus/getting-started/installation) and the
[GLM-5.2 deployment tutorial](/docs/hardware-platforms/ascend-npus/model-deployment/tutorials/glm_5_2#installation) to start a container with the NPU devices, driver mounts, and checkpoint directory. Run the setup and the generated **Python** command inside that container.
<Tabs>
<Tab title="A3">
```bash Command
docker pull quay.io/ascend/sglang:cann9.0.0-a3-v0.5.16
```
Run the following setup before copying the generated command (CANN 9.0.0). CPU tuning requires host privileges.
```bash Command
unset http_proxy https_proxy HTTP_PROXY HTTPS_PROXY all_proxy ALL_PROXY ASCEND_LAUNCH_BLOCKING
echo performance | tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor
sysctl -w vm.swappiness=0
sysctl -w kernel.numa_balancing=0
sysctl -w kernel.sched_migration_cost_ns=50000
source /usr/local/Ascend/ascend-toolkit/set_env.sh
source /usr/local/Ascend/nnal/atb/set_env.sh
```
</Tab>
<Tab title="950PR/DT series">
Use the compatible 950PR/DT series SGLang checkout, CANN environment, and custom operators used for your deployment. Set `SGLANG_ROOT` to that checkout; these recipes do not prescribe an image tag.
```bash Command
unset http_proxy https_proxy HTTP_PROXY HTTPS_PROXY all_proxy ALL_PROXY ASCEND_LAUNCH_BLOCKING
unset HCCL_IF_IP HCCL_SOCKET_FAMILY RANK_TABLE_FILE
SGLANG_ROOT=/path/to/sglang
export PYTHONPATH="${SGLANG_ROOT}/python:${PYTHONPATH:-}"
```
</Tab>
</Tabs>
Replace the model path in the generated command with your mounted ModelSlim checkpoint directory. Include its configuration, tokenizer, quantization metadata, and MTP weights.
For A3, use the locally prepared **GLM-5.2-w8a8** checkpoint described in the [model-weight prerequisites](/docs/hardware-platforms/ascend-npus/model-deployment/tutorials/glm_5_2#model-weights).
For 950PR/DT series, use your locally prepared GLM-5.2 **W8A8C8_MXFP8** checkpoint, including MTP weights, and replace the model path in the generated command with its mounted path.
Replace network placeholders such as `<network-interface>`, `<prefill-head-ip>`, and `<decode-head-ip>` in the generated NPU command with your cluster settings.
</Accordion>
Pick your hardware + recipe to generate the launch command. The three serving strategies cover the common operating points:
- **Low-Latency** — fastest reply for a single user. Pick for chat.
@@ -1,6 +1,6 @@
---
title: GLM-5.3-Flash
description: "Deploy GLM-5.3-Flash with SGLang using recipes for H100, H200, B200, B300, GB200, GB300, and MI355X, with MTP and multimodal serving."
description: "Deploy GLM-5.3-Flash with SGLang using recipes for H100, H200, B200, B300, GB200, GB300, MI355X, and TPU7x, with MTP and multimodal serving, plus Ascend A3 and 950PR/DT series NPU recipes."
tag: NEW
---
@@ -20,6 +20,67 @@ For the listed NVIDIA hardware, the deployment panel can render a complete `dock
For MI355X, follow the [AMD GPU installation guide](/docs/hardware-platforms/amd_gpu) and run the generated **Python** command on four GPUs. Select FP8 or [MXFP4](https://huggingface.co/OneNexus/GLM-5.3-Flash-MXFP4); SGLang downloads the model and tokenizer automatically.
For TPU7x, install [sglang-torchtpu](https://github.com/sgl-project/sglang-torchtpu) from source (its `main` branch; the `v0.1.0-tpu` image predates the GLM-5.3-Flash MTP fixes) and run the generated **Python** command on one TPU7x host with 4 chips:
```bash Command
git clone https://github.com/sgl-project/sglang-torchtpu.git
cd sglang-torchtpu
cp python/pyproject_tpu.toml python/pyproject.toml
pip install --extra-index-url https://download.pytorch.org/whl/cpu -e python/
pip install --no-deps -e .
```
</Accordion>
<Accordion title="Install and prepare Ascend NPUs">
Use a Linux host with the Ascend driver and a compatible CANN / TorchNPU environment. Follow the
[Ascend installation guide](/docs/hardware-platforms/ascend-npus/getting-started/installation) and the
[GLM-5.3-Flash deployment tutorial](/docs/hardware-platforms/ascend-npus/model-deployment/tutorials/glm_5_3_flash#installation) to start a container with the NPU devices, driver mounts, and checkpoint directory. Run the setup and the generated **Python** command inside that container.
<Tabs>
<Tab title="A3">
```bash Command
docker pull swr.cn-southwest-2.myhuaweicloud.com/base_image/dockerhub/lmsysorg/sglang:cann9.1.0-a3-glm5.3-flash
```
Run the following setup before copying the generated command (CANN 9.1.0). CPU tuning requires host privileges.
```bash Command
unset http_proxy https_proxy HTTP_PROXY HTTPS_PROXY all_proxy ALL_PROXY ASCEND_LAUNCH_BLOCKING
source /usr/local/Ascend/ascend-toolkit/set_env.sh
source /usr/local/Ascend/nnal/atb/set_env.sh
source /usr/local/Ascend/cann-9.1.0/opp/vendors/customize/bin/set_env.bash
```
</Tab>
<Tab title="950PR/DT series">
Use a CANN 9.1.0 SGLang image with the custom operators and MemFabric runtime required by the tutorial; PD mode also requires DP-attention support. The tutorial does not specify a 950PR/DT series image tag.
```bash Command
NPU_IMAGE="<your-compatible-950pr-dt-series-image>"
docker pull "$NPU_IMAGE"
```
```bash Command
unset http_proxy https_proxy HTTP_PROXY HTTPS_PROXY all_proxy ALL_PROXY ASCEND_LAUNCH_BLOCKING
source /usr/local/Ascend/ascend-toolkit/set_env.sh
source /usr/local/Ascend/ascend-toolkit/latest/opp/vendors/customize/bin/set_env.bash
source /usr/local/Ascend/ascend-toolkit/latest/opp/vendors/custom_transformer/bin/set_env.bash
source /usr/local/memfabric_hybrid/set_env.sh
```
</Tab>
</Tabs>
Replace the model path in the generated command with your mounted ModelSlim checkpoint directory. Include its configuration, tokenizer, quantization metadata, and MTP weights.
Use [W8A8](https://www.modelscope.cn/models/Eco-Tech/GLM-5.3-Flash-w8a8) for A3 and [W8A8_MXFP8](https://www.modelscope.cn/models/Eco-Tech/GLM-5.3-Flash-w8a8-mxfp8) for 950PR/DT series.
Replace network placeholders such as `<network-interface>`, `<prefill-head-ip>`, and `<decode-head-ip>` in the generated NPU command with your cluster settings.
</Accordion>
The cookbook opens on **B300**. Choose your hardware, then choose the operating point that matches your workload:
@@ -27,7 +88,7 @@ The cookbook opens on **B300**. Choose your hardware, then choose the operating
- **Low Latency** starts with MTP 5/1/6 speculative decoding and tensor parallelism to shorten interactive responses.
- **High Throughput** starts with speculative decoding off, which avoids draft-and-verify overhead under sustained batches.
Every listed hardware platform exposes both strategies. A **Verified** badge means that exact hardware and command were tested. **Final Verification In Progress** means the recipe runs and is queued for measurement on the final weights. **Not Verified** means the command is a supported starting point that still needs workload validation. A choice is disabled only when the underlying runtime combination is known to be unsupported.
Every listed GPU platform exposes both strategies; TPU7x has a Low Latency recipe only. A **Verified** badge means that exact hardware and command were tested. **Final Verification In Progress** means the recipe runs and is queued for measurement on the final weights. **Not Verified** means the command is a supported starting point that still needs workload validation. A choice is disabled only when the underlying runtime combination is known to be unsupported.
The recommended selection is only a starting point. The same panel also lets you override the KV/DSA pairing, multimodal feature transport, Breakable Cuda Graph, and HiCache tiers. The MI355X commands are marked **Not Verified** because the selector defaults to HiCache off, whereas the measured AgentX configuration adds the host tier described in [Extend the cache hierarchy](#extend-the-cache-hierarchy). Do not treat the CUDA benchmark rows as MI355X measurements.
@@ -110,6 +171,10 @@ The **Speculative** card in the Playground changes the algorithm without leaving
Neither algorithm runs with DP-Attention; the card disables the affected chips and names the reason. DFlash2's FA4 draft path is also disabled on MI355X.
### Serve on TPU7x
The TPU7x recipe serves the language model with DeepSeek Sparse Attention on the Pallas sparse MLA kernels and MTP 3/1/4 on the experimental chain speculative worker (`SGLANG_TPU_ENABLE_EXPERIMENTAL_SPEC_WORKER=1`). On TPU7x, 3/1/4 decodes faster than 5/1/6 at the same accept length, because each extra draft row widens the target verify. It is sized for one request with 8K input and 1K output: `--context-length 10240`, `--max-running-requests 1`, and a 20,480-token KV pool. Raise `--context-length`, `--max-total-tokens`, `--max-running-requests` and `--max-mamba-cache-size` together for longer or concurrent requests; those settings are outside the verified command. The first start compiles every decode and prefill shape listed in `--tpu-compile-bs` and `--tpu-compile-tokens`, so it takes longer than later starts that reuse the compilation cache.
### Size both memory pools
GLM-5.3-Flash maintains a paged KV pool for attention and a separate KDA state pool. The KDA state pool can limit concurrency before the KV pool is full. If startup reduces `max_running_requests` because of KDA state capacity, increase `--mamba-full-memory-ratio` or set `--max-mamba-cache-size` for the expected concurrency, then tune `--max-running-requests` to the workload.
+69 -2
View File
@@ -1,6 +1,6 @@
---
title: GLM-5.3
description: "Deploy GLM-5.3 with SGLang — Z.ai's DeepSeek-Sparse-Attention (DSA) Mixture-of-Experts model with MTP speculative decoding and 1M context, on H200, B200, B300, GB300, and AMD MI300X/MI325X/MI355X."
description: "Deploy GLM-5.3 with SGLang — Z.ai's DeepSeek-Sparse-Attention (DSA) Mixture-of-Experts model with MTP speculative decoding and 1M context, on H200, B200, B300, GB300, and AMD MI300X/MI325X/MI355X, plus Ascend NPU recipes."
tag: NEW
---
@@ -40,6 +40,60 @@ For how to launch the image, see [Install → Method 3: Using Docker](../../../d
</Accordion>
<Accordion title="Install and prepare Ascend NPUs">
Use a Linux host with the Ascend driver and a compatible CANN / TorchNPU environment. Follow the
[Ascend installation guide](/docs/hardware-platforms/ascend-npus/getting-started/installation) and the
[GLM-5.3 deployment tutorial](/docs/hardware-platforms/ascend-npus/model-deployment/tutorials/glm_5_3#installation) to start a container with the NPU devices, driver mounts, and checkpoint directory. Run the setup and the generated **Python** command inside that container.
<Tabs>
<Tab title="A3">
```bash Command
docker pull quay.io/ascend/sglang:cann9.1.0-a3-B080
```
This CANN 9.1.0 image requires a compatible SGLang build, TorchNPU, ATB, and custom operators before serving. Set `SGLANG_ROOT` to that source checkout.
Run the following setup before copying the generated command (CANN 9.1.0). CPU tuning requires host privileges.
```bash Command
unset http_proxy https_proxy HTTP_PROXY HTTPS_PROXY all_proxy ALL_PROXY ASCEND_LAUNCH_BLOCKING
echo performance | tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor
sysctl -w vm.swappiness=0
sysctl -w kernel.numa_balancing=0
sysctl -w kernel.sched_migration_cost_ns=50000
source /usr/local/Ascend/ascend-toolkit/set_env.sh
source /usr/local/Ascend/nnal/atb/set_env.sh
source /usr/local/Ascend/cann-9.1.0/opp/vendors/customize/bin/set_env.bash
SGLANG_ROOT=/path/to/sglang
export PYTHONPATH="${SGLANG_ROOT}/python:${PYTHONPATH:-}"
export LD_LIBRARY_PATH="/usr/local/Ascend/ascend-toolkit/latest/opp/vendors/customize/op_api/lib:${LD_LIBRARY_PATH:-}"
```
</Tab>
<Tab title="950PR/DT series">
Use the compatible 950PR/DT series SGLang checkout, CANN environment, and custom operators used for your deployment. Set `SGLANG_ROOT` to that checkout; these recipes do not prescribe an image tag.
```bash Command
unset http_proxy https_proxy HTTP_PROXY HTTPS_PROXY all_proxy ALL_PROXY ASCEND_LAUNCH_BLOCKING
unset HCCL_IF_IP HCCL_SOCKET_FAMILY RANK_TABLE_FILE
SGLANG_ROOT=/path/to/sglang
export PYTHONPATH="${SGLANG_ROOT}/python:${PYTHONPATH:-}"
```
</Tab>
</Tabs>
Replace the model path in the generated command with your mounted ModelSlim checkpoint directory. Include its configuration, tokenizer, quantization metadata, and MTP weights.
Use a locally prepared **W8A8** checkpoint for A3. On 950PR/DT series, both unified serving and PD disaggregation use the same **W8A8C8_MXFP8** checkpoint. Both 950PR/DT series recipes explicitly use FP8 E4M3 KV cache and BF16 draft KV cache.
Replace network placeholders such as `<network-interface>`, `<prefill-head-ip>`, and `<decode-head-ip>` in the generated NPU command with your cluster settings.
</Accordion>
Pick your hardware + recipe to generate the launch command. The three serving strategies cover the common operating points:
- **Low-Latency** — fastest reply for a single user. Pick for chat.
@@ -50,7 +104,7 @@ import { Deployment } from "/src/snippets/_deployment.jsx";
import { config } from "/src/snippets/configs/zai-org/glm-5.3.jsx";
import { benchmarks } from "/src/snippets/configs/zai-org/glm-5.3-benchmarks.jsx";
<Deployment config={config} benchmarks={benchmarks} />
<Deployment config={config} benchmarks={benchmarks} agenticLink="agentic-long-context-deployment" />
<Warning>
All recipes here run the DSA indexer top-k on the default `--dsa-topk-backend sgl-kernel`. Other top-k backend choices have not been fully validated on GLM-5.3.
@@ -64,6 +118,19 @@ import { Playground } from "/src/snippets/_playground.jsx";
<Playground config={config} />
## Agentic Long-Context Deployment
Agentic workloads, such as coding agents and other multi-turn, tool-using assistants, send long prompts whose prefixes grow and repeat from turn to turn. Serving them well depends on reusing the KV cache across turns, offloading it beyond GPU memory, and keeping each session on the workers that already hold its prefix. This section gives complete launch commands for such deployments, each benchmarked in [SemiAnalysis InferenceX](https://inferencex.semianalysis.com/) and linked to the run that measured it. Pick a hardware platform, checkpoint, deployment shape, concurrency, KV cache offloading mechanism and router to get every process of the deployment as a raw command, one tab per process in launch order.
- **Router.** Each configuration was submitted behind one front door: Dynamo (`dynamo.frontend` with `dynamo.sglang` workers) or SGLang (`sglang.launch_server`, behind `sglang_router` when there are several workers or DP-attention ranks).
- **KV offload.** Choose among these KV cache offloading mechanisms: none (GPU KV only), HiCache (host DRAM), or a Mooncake store attached through the external linker. The Mooncake option needs an image whose SGLang build includes `--enable-unified-cache-external-linker`.
- **Container image.** Run the commands inside the image listed for each configuration. Several are SGLang nightly images, so flag names follow that image.
import { AgentX } from "/src/snippets/_agentx.jsx";
import { agentx } from "/src/snippets/agentx/zai-org/glm-5.3.jsx";
<AgentX data={agentx} />
## 1. Model Introduction
**GLM-5.3** is Z.ai's flagship Mixture-of-Experts model built on **DeepSeek Sparse Attention (DSA)**: a lightning indexer selects a sparse set of key tokens per query (top-2048), so attention cost stays near-constant as context grows. It keeps the same base architecture while updating post-training for complex coding and long-horizon tasks. It ships in two precisions — **FP8** (`zai-org/GLM-5.3`) and full **BF16** (`zai-org/GLM-5.3-BF16`) — both with **78 transformer layers**, **256 routed experts** (8 active per token), a **1M-token context window**, and a single **MTP (Multi-Token Prediction)** layer for built-in EAGLE-style speculative decoding. FP8 is the recommended deployment; BF16 (~1.5 TB) needs an 8×B300 node or a multi-node setup. For Blackwell, RadixArk publishes an **experimental NVFP4** build (`RadixArk/GLM-5.3-NVFP4`, Model Optimizer) that quantizes only the routed experts' linear weights and activations to 4-bit (attention, shared experts, dense layers, MTP, embeddings, and the LM head stay unquantized), cutting the weight footprint to ~0.45 TB so a 4-GPU GB300 node can serve it with TP4. **The NVFP4 cells are experimental** — benchmark data (accuracy / throughput) is still pending; treat them as provisional until the benchmark cards fill in.
@@ -97,7 +97,7 @@ For other installation methods, please refer to the [official SGLang installatio
### 3.1 Basic Configuration
The Gemma 4 series offers models in various sizes and architectures, optimized for different hardware platforms including NVIDIA GPUs, AMD GPUs, and Intel Arc Pro B-Series GPUs(codename: BMG (Battlemage)). The recommended launch configurations vary by hardware and model size.
The Gemma 4 series offers models in various sizes and architectures, optimized for different hardware platforms including NVIDIA GPUs, AMD GPUs, and Intel Arc Pro B-Series GPUs(codename: BMG (Battlemage)), Xeon CPUs. The recommended launch configurations vary by hardware and model size.
**Interactive Command Generator**: Use the configuration selector below to automatically generate the appropriate deployment command for your hardware platform and model variant.
@@ -22,11 +22,23 @@ Select **Docker** in the command panel to launch the matching image. The **Pytho
The first launch in a fresh container spends about five extra minutes compiling JIT kernels for the speculative-decoding graphs (about 4–5 minutes once, then about 30 seconds). Mount a persistent cache volume, for example `-v ~/.cache:/root/.cache`, to skip it on later starts.
</Note>
<Note>
**AMD MI325X / MI350X:** the published ROCm images predate the ROCm gated-MLA fix that Ling-3.1-flash needs; without it the server fails while capturing decode CUDA graphs. Until a ROCm image includes the fix, build a thin image that switches the nightly's editable SGLang source to the Ling-3.1-flash branch (use the `mi35x` base for MI350X), then substitute it for the image in the generated Docker command:
```dockerfile Dockerfile
FROM lmsysorg/sglang-rocm:v0.5.21-rocm724-mi30x-20261013
RUN cd /sgl-workspace/sglang \
&& git fetch https://github.com/sgl-project/sglang.git bailing-v3-day0 \
&& git checkout FETCH_HEAD
```
</Note>
</Accordion>
Pick your hardware to generate the launch command:
- **Low-Latency** — single-user chat and interactive requests, with NEXTN speculative decoding. Verified on H200 and GB300; H20-3e, H800, H100, and B200 recipes are provided unverified.
- **Low-Latency** — TP-only (EP=1) with NEXTN speculative decoding for single-user chat and interactive requests. Verified on H200.
- **High-Throughput** — expert-parallel (EP) serving for batch and multi-user traffic: TP8 + EP8 on H200/H800/H100/B200-class nodes, TP4 + EP4 on GB300, TP8 + EP8 with AITER on AMD MI300X/MI325X/MI350X/MI355X (without speculative decoding). Verified on H200 and GB300; other platforms are unverified. AMD MI325X and MI350X recipes (TP8 + EP8 with AITER, without speculative decoding) are provided unverified; the MI325X accuracy in the benchmark card was measured by AMD.
For prefix-heavy multi-turn workloads, enable **HiCache** in the Playground after starting the Mooncake services in §3.3. The combination is verified: prefix hits return from Mooncake L3 on every TP rank.
@@ -60,6 +72,8 @@ The model has about 560B total parameters (512 routed experts with 8 active plus
- On H200, enable `SGLANG_ENABLE_FP8_BLOCK_GEMV=1` and `SGLANG_ENABLE_FUSED_KDA_OUTPUT_QUANT=1`. Both are included in the generated commands; they are no-ops on GB300 (SM10x).
- Leave `--mem-fraction-static` at its default. SGLang auto-selects the `ling3` reasoning and tool-call parsers for this architecture; you can change the parser settings in the Playground.
- **NEXTN:** add `--speculative-algorithm NEXTN` for built-in MTP speculative decoding. Measured acceptance length is 2.61 on random tokens and 2.13 on real-text probes.
- **Long context:** the checkpoint supports 262,144 tokens natively. For up to 1,048,576 tokens, add the YaRN override from §3.4.
- **Topology:** the Low-Latency recipe runs TP-only (no `--ep-size`) and is clearly faster at batch size 1 (about 9-13% lower TPOT on H200, verified on both 2048/512 and 8192/1024 workloads). At concurrency 16-32 the two recipes deliver the same throughput within ~2%. TP-only relies on the loader's TP-requant support; it fails loudly at startup on builds without it.
- For prefix-heavy, multi-turn workloads, select **Cache fork and tail** in the Playground. It combines `--mamba-radix-cache-strategy extra_buffer`, `--chunked-prefill-size 4096`, and `SGLANG_SPLIT_PREFILL_AT_MAMBA_FORK=1`. GPU checks confirmed both fork and tail states are cached, saving about 576 tokens per subsequent turn in the measured workload.
### Playground support
@@ -71,7 +85,7 @@ The model has about 560B total parameters (512 routed experts with 8 active plus
### Performance
On H200 TP8 + EP8 with NEXTN and the decode gates enabled, batch-size-1 output throughput measures **251.2 tokens/s**. GB300 NEXTN batch-size-1 result: `180.0 tok/s (TPOT 5.14 ms, accept 2.59)`.
On H200 TP8 with NEXTN and the decode gates enabled, batch-size-1 output throughput measures **269.2 tokens/s** (TP8 + EP8: 239.5 tokens/s). GB300 TP4 with NEXTN measures **175.3 tokens/s** (`TPOT 5.22 ms, accept 2.60`; TP4 + EP4: 160.0 tokens/s).
Acceptance length is the mean number of tokens accepted per speculative verification step, including the target bonus token:
@@ -108,7 +122,7 @@ python3 -m sglang.bench_serving \
--flush-cache --seed 11
```
For the concurrency-32 row in the benchmark widget:
For the saturation row in the benchmark widget, launch the H200 **High-Throughput** recipe instead and run:
```bash Command
python3 -m sglang.bench_serving \
@@ -116,9 +130,9 @@ python3 -m sglang.bench_serving \
--host localhost --port 30000 \
--model inclusionAI/Ling-3.1-flash \
--dataset-name random \
--random-input-len 4096 --random-output-len 512 \
--random-input-len 8192 --random-output-len 1024 \
--random-range-ratio 1 \
--num-prompts 96 --max-concurrency 32 \
--num-prompts 256 --max-concurrency 48 \
--flush-cache --seed 11
```
@@ -133,7 +147,7 @@ sgl-eval run gsm8k \
--num-threads 32
```
Evaluate all 1,319 questions and inspect completion reasons as well as accuracy. For the H200 NEXTN recipe, compare the stop rate with the measured 99.85%; accuracy alone can hide generations that reach the token limit.
Evaluate all 1,319 questions and inspect completion reasons as well as accuracy. For the H200 NEXTN recipes, compare the stop rate with the measured 99.77% (Low-Latency) or 99.92% (High-Throughput); accuracy alone can hide generations that reach the token limit.
</Accordion>
@@ -277,30 +291,48 @@ print(follow_up.choices[0].message.content)
```json Output
{
"message": {
"role": "assistant",
"content": "",
"reasoning_content": "The user wants to know the current weather in Hangzhou. I'll call the get_weather function with city \"Hangzhou\". The unit parameter is optional - I'll use celsius as default since Hangzhou is in China.",
"tool_calls": [
{
"id": "call_2bae8419df9d44f2bb54c22f",
"index": 0,
"type": "function",
"function": {
"name": "get_weather",
"arguments": "{\"city\": \"Hangzhou\", \"unit\": \"celsius\"}"
}
}
]
},
"finish_reason": "tool_calls"
"content": "",
"role": "assistant",
"tool_calls": [
{
"id": "call_ecd286d7dccb4c70a8c9c586",
"function": {
"arguments": "{\"city\": \"Hangzhou\", \"unit\": \"celsius\"}",
"name": "get_weather"
},
"type": "function",
"index": 0
}
],
"reasoning_content": "The user is asking about the current weather in Hangzhou. I need to call the get_weather function with the city parameter set to \"Hangzhou\". The unit parameter is optional - I'll default to celsius since that's commonly used in China."
}
```
The follow-up assistant turn answers with both `reasoning_content` and `content` populated.
</Accordion>
<Accordion title="Example Output (follow-up)">
```text Output
The user asked about the weather in Hangzhou and I got the result. Let me provide a natural response.
The weather in Hangzhou right now is sunny with a temperature of 22°C. It's a pleasant day!
```
The first line is `reasoning_content`; the second is `content`. `finish_reason` is `stop`.
</Accordion>
### 3.4 Long context with YaRN (up to 1M tokens)
The model's native context is 262,144 tokens. Verified on H200 and GB300, YaRN extrapolation serves up to **1,048,576 tokens**:
```bash Command
--json-model-override-args '{"rope_scaling": {"rope_type": "yarn", "factor": 4.0, "original_max_position_embeddings": 262144, "rope_theta": 6000000, "partial_rotary_factor": 0.5}, "max_position_embeddings": 1048576}'
```
Add it to the Low-Latency serve command. All four keys are required: the override replaces the checkpoint's RoPE parameters, so omitting `rope_theta` or `partial_rotary_factor` silently changes the RoPE base, and omitting `max_position_embeddings` keeps the 262,144 limit. Only enable it for prompts beyond 256K.
Verification on the current checkpoint: a fact planted at token 700,023 of a 900,022-token prompt is retrieved exactly, NEXTN acceptance stays about 2.1 at 900K, and a 300,101-token prompt is answered correctly (H200 TTFT 15.4 s, GB300 12.6 s). One 1M-token request fits alongside normal serving traffic on both platforms.
### 3.3 HiCache with Mooncake
With HiCache enabled, prefix state is written through to Mooncake L3 and later requests prefetch it back: measured 5952 of 6000 tokens restored from storage on every TP rank. Mooncake runs as L3 prefix storage over TCP; start the metadata server, master, and storage client before launching SGLang. Run these services on the same host as the SGLang container:
+78
View File
@@ -0,0 +1,78 @@
---
title: Kev
description: "Deploy the Kev decision models with SGLang. Kev returns calibrated probabilities for typed questions about a document, in sizes from 0.8B to 27B, each on one NVIDIA H200 in BF16."
tag: NEW
---
## Deployment
Every Kev size runs on one GPU in BF16 with TP=1 and needs no extra flag. SGLang reads the checkpoint's `head.pt` and configures the server.
```bash Command
sglang serve --model-path jaredpalmer/kev-27b --host 0.0.0.0 --port 30000
```
Replace the model path with any size in the table below. Each was validated on one H200.
| Model | Hugging Face ID | Checkpoint | Backbone | GPU memory for weights |
| --- | --- | --- | --- | --- |
| Kev-27B | `jaredpalmer/kev-27b` | full weights | Qwen3.8-27B | 47.8 GB |
| Kev-9B | `jaredpalmer/kev-9b` | LoRA adapter | Qwen3.5-9B-Base | 14.9 GB |
| Kev-4B | `jaredpalmer/kev-4b` | LoRA adapter | Qwen3.5-4B-Base | 8.0 GB |
| Kev-0.8B | `jaredpalmer/kev-0.8b` | LoRA adapter | Qwen3.5-0.8B-Base | 1.5 GB |
For a LoRA adapter, SGLang loads the base revision that the adapter's `head.pt` pins and merges the adapter into its weights at startup, in FP32 with one rounding to BF16, as the reference code merges it. The base is downloaded on first use.
## 1. Model introduction
[Kev](https://huggingface.co/collections/jaredpalmer/kev) is a family of decision models published by Jared Palmer under the Apache-2.0 license, with [reference code](https://github.com/jaredpalmer/kev). Each model keeps a Qwen3.5 or Qwen3.8 backbone and replaces the LM head with a pointer head. Each question runs as its own prompt after the shared state. The head compares the final hidden state at each option's closing token with the one at the prompt's last token and returns one logit per option, with no text generation. A temperature stored in `head.pt` calibrates the probabilities.
Pass a text string, JSON object, or JSON array as the state. `/v1/systemone` returns one answer per question: a `choice` over allowed options, a `noul` probability of true, or a `score` over ordered levels. JSON states are flattened into labeled text as the reference code does.
## 2. Configuration tips
- **Route.** Send requests to `/v1/systemone`, or set your System One client's base URL to the server. This includes TypeSafe SDK clients. See the [API reference](/docs/supported-models/decision_models#system-one-compatible-api) for request and response fields. Kev refuses `/v1/decisions`, chat, generation, and `/encode` requests.
- **One prefill per question.** The model reads the state once for every question, as the reference does, so a request with many questions over a long state costs that many prefills. SGLang enables embedding mode and disables radix caching, chunked prefill, and CUDA graphs. Keep TP=1 and PP=1.
- **Answers.** `probabilities` are the softmax of the head's logits divided by the calibrated temperature. `confidence` uses the System One adapter formulas, as Kev's own server does. SGLang does not round the values to four decimals. Answers omit `x_label_mass` because the head scores only the allowed options.
- **Long states.** The reference server accepts states of up to 65,536 tokens. In SGLang each question's prompt must also fit one prefill, 16,384 tokens by default, so longer prompts are refused with an error that gives both lengths. Raising `--max-prefill-tokens` is not supported for this model yet, because the server does not reserve memory for longer unchunked prefills and can run out of memory under load.
- **Inputs.** Requests are text only, so `images` and `chat_template_kwargs` are refused. A `score` question takes up to 10 levels and a `choice` question up to 255 options.
- **Not supported.** Kev-27B v1, the `v1-lora` revision of `jaredpalmer/kev-27b`, is an adapter trained on a BF16 backbone that the reference applies unmerged, so SGLang refuses it and serves the current full weight release instead. Kev-0.5B is a prototype on a Qwen2.5 base and is refused too.
## 3. Advanced usage
### 3.1 TypeSafe SDK
```bash Command
pip install typesafe-sdk
```
<Accordion title="Text decision example (Python)">
```python Example
from typesafe_sdk import Choice, Noul, TypeSafeClient
client = TypeSafeClient(api_key="local", base_url="http://localhost:30000", model="kev-latest")
response = client.system_one(
state="I was charged twice for order 1182. Please refund one of the charges.",
questions={
"team": Choice(
instructions="Which team should handle this?",
criteria={"billing": "Charges and refunds", "shipping": "Deliveries", "returns": "Exchanges"},
),
"urgent": Noul(instructions="Does this need a reply today?"),
},
)
print(response.choices["team"].choice, response.nouls["urgent"].noul)
```
</Accordion>
<Accordion title="Example output">
Captured with Kev-27B on one H200. Probabilities can vary with the model size, hardware, and numerical backend.
```text Output
billing 0.7591662905989526
```
</Accordion>
@@ -48,6 +48,445 @@ import { agentx } from "/src/snippets/agentx/meta-models/muse-spark-open.jsx";
<AgentX data={agentx} />
### GB300 deployments measured with Artificial Analysis AgentPerf
Five aggregated Dynamo + SGLang deployments on GB300, measured with the [Artificial Analysis AgentPerf](https://github.com/ArtificialAnalysis/agentperf-client) coding-agent workload. Together they span the trade-off from the lowest per-user latency (concurrency 1) to the highest throughput per GPU (concurrency 160), ordered from low latency to high throughput. Each one starts from the Artificial Analysis SGLang submission for this model and changes only flags, environment variables and parallelism.
- **Container image.** `lmsysorg/sglang:dev-muse-spark-open@sha256:8cfc5a72cb296771f5f8a4a4688cb6d05685610a830cbb616fa4ab0eb7bb2b0a` with the Dynamo wheels added: `pip install --no-deps ai-dynamo==1.6.0.dev20261019 ai-dynamo-runtime==1.6.0.dev20261019`.
- **Router.** One Dynamo frontend with the KV-aware router. Session affinity keeps requests that carry the same `X-Dynamo-Session-ID` header on one worker; the AgentPerf client sends one ID per agent session.
- **Checkpoints.** `meta-models/Muse-Spark-Open-816B-A42B-NVFP4` with the DFlash2 draft `RadixArk/Muse-Spark-Open-DFlash2` (8 draft tokens, FP8 KV cache).
- **HiCache.** `--hicache-size` is the host tier in GB per rank; scale it to your host memory.
- **Two-node deployments.** The TP8·EP8 worker spans two GB300 trays (4 GPUs each) over multi-node NVLink: run the worker command on both nodes, with `NODE_RANK` 0 on the first.
<AccordionGroup>
<Accordion title="Agg TP8·EP8 (2 nodes), lowest latency · 8 GPUs · concurrency 1">
<Tabs sync={false}>
<Tab title="Env Vars">
```bash
export ETCD_IP=<node running etcd>
export AGG_HEAD_IP=<first node of the worker>
export NODE_RANK=<0 on the first node, 1 on the second>
```
</Tab>
<Tab title="etcd">
```bash
etcd \
--data-dir /tmp/etcd \
--listen-client-urls http://0.0.0.0:2379 \
--advertise-client-urls http://$ETCD_IP:2379
```
</Tab>
<Tab title="Dynamo frontend">
```bash
export DYN_REQUEST_PLANE=tcp
export DYN_ROUTER_QUEUE_THRESHOLD=None
export DYN_ROUTER_TEMPERATURE=0
export DYN_TOKENIZER=fastokens
export DYN_TOKENIZER_CACHE=1
export DYN_TOKENIZER_CACHE_BYTES=8000000000
export ETCD_ENDPOINTS=http://$ETCD_IP:2379
python3 -m dynamo.frontend \
--http-port 8000 \
--router-mode kv \
--router-session-affinity-ttl-secs 14400 \
--active-decode-blocks-threshold None \
--active-prefill-tokens-threshold None \
--active-prefill-tokens-threshold-frac None
```
</Tab>
<Tab title="SGLang worker (run on each of the 2 nodes)">
```bash
export DYN_REQUEST_PLANE=tcp
export ETCD_ENDPOINTS=http://$ETCD_IP:2379
export NCCL_CUMEM_ENABLE=1
export NCCL_MNNVL_ENABLE=1
export PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True
export SGLANG_TIMEOUT_KEEP_ALIVE=900
python3 -m dynamo.sglang \
--model-path meta-models/Muse-Spark-Open-816B-A42B-NVFP4 \
--served-model-name meta-models/Muse-Spark-Open-816B-A42B-NVFP4 \
--host 0.0.0.0 \
--port 30000 \
--dist-init-addr $AGG_HEAD_IP:20000 \
--nnodes 2 \
--node-rank $NODE_RANK \
--kv-events-config '{"publisher":"zmq","topic":"kv-events","endpoint":"tcp://*:5557"}' \
--request-plane tcp \
--dyn-reasoning-parser muse \
--dyn-tool-call-parser muse \
--expert-parallel-size 8 \
--kv-cache-dtype fp8_e4m3 \
--max-running-requests 32 \
--mem-fraction-static 0.85 \
--mm-feature-transport cpu \
--moe-runner-backend flashinfer_trtllm \
--quantization modelopt_fp4 \
--reasoning-parser muse \
--speculative-algorithm DFLASH \
--speculative-draft-attention-backend fa4 \
--speculative-draft-model-path RadixArk/Muse-Spark-Open-DFlash2 \
--speculative-num-draft-tokens 8 \
--swa-full-tokens-ratio 0.1 \
--tensor-parallel-size 8 \
--trust-remote-code \
--watchdog-timeout 3600
```
</Tab>
</Tabs>
</Accordion>
<Accordion title="Agg TP8·EP8 (2 nodes), low latency · 8 GPUs · concurrency 4">
<Tabs sync={false}>
<Tab title="Env Vars">
```bash
export ETCD_IP=<node running etcd>
export AGG_HEAD_IP=<first node of the worker>
export NODE_RANK=<0 on the first node, 1 on the second>
```
</Tab>
<Tab title="etcd">
```bash
etcd \
--data-dir /tmp/etcd \
--listen-client-urls http://0.0.0.0:2379 \
--advertise-client-urls http://$ETCD_IP:2379
```
</Tab>
<Tab title="Dynamo frontend">
```bash
export DYN_REQUEST_PLANE=tcp
export DYN_ROUTER_QUEUE_THRESHOLD=None
export DYN_ROUTER_TEMPERATURE=0
export DYN_TOKENIZER=fastokens
export DYN_TOKENIZER_CACHE=1
export DYN_TOKENIZER_CACHE_BYTES=8000000000
export ETCD_ENDPOINTS=http://$ETCD_IP:2379
python3 -m dynamo.frontend \
--http-port 8000 \
--router-mode kv \
--router-session-affinity-ttl-secs 14400 \
--active-decode-blocks-threshold None \
--active-prefill-tokens-threshold None \
--active-prefill-tokens-threshold-frac None
```
</Tab>
<Tab title="SGLang worker (run on each of the 2 nodes)">
```bash
export DYN_REQUEST_PLANE=tcp
export ETCD_ENDPOINTS=http://$ETCD_IP:2379
export NCCL_CUMEM_ENABLE=1
export NCCL_MNNVL_ENABLE=1
export PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True
export SGLANG_TIMEOUT_KEEP_ALIVE=900
python3 -m dynamo.sglang \
--model-path meta-models/Muse-Spark-Open-816B-A42B-NVFP4 \
--served-model-name meta-models/Muse-Spark-Open-816B-A42B-NVFP4 \
--host 0.0.0.0 \
--port 30000 \
--dist-init-addr $AGG_HEAD_IP:20000 \
--nnodes 2 \
--node-rank $NODE_RANK \
--kv-events-config '{"publisher":"zmq","topic":"kv-events","endpoint":"tcp://*:5557"}' \
--request-plane tcp \
--dyn-reasoning-parser muse \
--dyn-tool-call-parser muse \
--expert-parallel-size 8 \
--kv-cache-dtype fp8_e4m3 \
--max-running-requests 32 \
--mem-fraction-static 0.85 \
--mm-feature-transport cpu \
--moe-runner-backend flashinfer_trtllm \
--quantization modelopt_fp4 \
--reasoning-parser muse \
--speculative-algorithm DFLASH \
--speculative-draft-attention-backend fa4 \
--speculative-draft-model-path RadixArk/Muse-Spark-Open-DFlash2 \
--speculative-num-draft-tokens 8 \
--swa-full-tokens-ratio 0.1 \
--tensor-parallel-size 8 \
--trust-remote-code \
--watchdog-timeout 3600
```
</Tab>
</Tabs>
</Accordion>
<Accordion title="Agg TP4·EP4, balanced · 4 GPUs · concurrency 16">
<Tabs sync={false}>
<Tab title="Env Vars">
```bash
export ETCD_IP=<node running etcd>
export AGG_HEAD_IP=<first node of the worker>
```
</Tab>
<Tab title="etcd">
```bash
etcd \
--data-dir /tmp/etcd \
--listen-client-urls http://0.0.0.0:2379 \
--advertise-client-urls http://$ETCD_IP:2379
```
</Tab>
<Tab title="Dynamo frontend">
```bash
export DYN_REQUEST_PLANE=tcp
export DYN_ROUTER_QUEUE_THRESHOLD=None
export DYN_ROUTER_TEMPERATURE=0
export DYN_TOKENIZER=fastokens
export DYN_TOKENIZER_CACHE=1
export DYN_TOKENIZER_CACHE_BYTES=8000000000
export ETCD_ENDPOINTS=http://$ETCD_IP:2379
python3 -m dynamo.frontend \
--http-port 8000 \
--router-mode kv \
--router-session-affinity-ttl-secs 14400 \
--active-decode-blocks-threshold None \
--active-prefill-tokens-threshold None \
--active-prefill-tokens-threshold-frac None
```
</Tab>
<Tab title="SGLang worker">
```bash
export DYN_REQUEST_PLANE=tcp
export ETCD_ENDPOINTS=http://$ETCD_IP:2379
export NCCL_CUMEM_ENABLE=1
export SGLANG_TIMEOUT_KEEP_ALIVE=900
python3 -m dynamo.sglang \
--model-path meta-models/Muse-Spark-Open-816B-A42B-NVFP4 \
--served-model-name meta-models/Muse-Spark-Open-816B-A42B-NVFP4 \
--host 0.0.0.0 \
--port 30000 \
--kv-events-config '{"publisher":"zmq","topic":"kv-events","endpoint":"tcp://*:5557"}' \
--request-plane tcp \
--dyn-reasoning-parser muse \
--dyn-tool-call-parser muse \
--enable-layernorm-sp \
--enable-megamoe-for-layernorm-sp \
--expert-parallel-size 4 \
--kv-cache-dtype fp8_e4m3 \
--max-running-requests 32 \
--mem-fraction-static 0.85 \
--mm-feature-transport cpu \
--moe-runner-backend flashinfer_trtllm \
--prefill-decode-interval 64 \
--quantization modelopt_fp4 \
--reasoning-parser muse \
--speculative-algorithm DFLASH \
--speculative-draft-attention-backend fa4 \
--speculative-draft-model-path RadixArk/Muse-Spark-Open-DFlash2 \
--speculative-num-draft-tokens 8 \
--swa-full-tokens-ratio 0.1 \
--tensor-parallel-size 4 \
--trust-remote-code \
--watchdog-timeout 3600
```
</Tab>
</Tabs>
</Accordion>
<Accordion title="Agg TP4·EP4, high throughput · 4 GPUs · concurrency 64">
<Tabs sync={false}>
<Tab title="Env Vars">
```bash
export ETCD_IP=<node running etcd>
export AGG_HEAD_IP=<first node of the worker>
```
</Tab>
<Tab title="etcd">
```bash
etcd \
--data-dir /tmp/etcd \
--listen-client-urls http://0.0.0.0:2379 \
--advertise-client-urls http://$ETCD_IP:2379
```
</Tab>
<Tab title="Dynamo frontend">
```bash
export DYN_REQUEST_PLANE=tcp
export DYN_ROUTER_QUEUE_THRESHOLD=None
export DYN_ROUTER_TEMPERATURE=0
export DYN_TOKENIZER=fastokens
export DYN_TOKENIZER_CACHE=1
export DYN_TOKENIZER_CACHE_BYTES=8000000000
export ETCD_ENDPOINTS=http://$ETCD_IP:2379
python3 -m dynamo.frontend \
--http-port 8000 \
--router-mode kv \
--router-session-affinity-ttl-secs 14400 \
--active-decode-blocks-threshold None \
--active-prefill-tokens-threshold None \
--active-prefill-tokens-threshold-frac None
```
</Tab>
<Tab title="SGLang worker">
```bash
export DYN_REQUEST_PLANE=tcp
export ETCD_ENDPOINTS=http://$ETCD_IP:2379
export NCCL_CUMEM_ENABLE=1
export SGLANG_OPT_UNIFIED_CACHE_FREE_OUT_OF_WINDOW_SLOTS=1
export SGLANG_TIMEOUT_KEEP_ALIVE=900
python3 -m dynamo.sglang \
--model-path meta-models/Muse-Spark-Open-816B-A42B-NVFP4 \
--served-model-name meta-models/Muse-Spark-Open-816B-A42B-NVFP4 \
--host 0.0.0.0 \
--port 30000 \
--kv-events-config '{"publisher":"zmq","topic":"kv-events","endpoint":"tcp://*:5557"}' \
--request-plane tcp \
--cuda-graph-max-bs-decode 64 \
--dyn-reasoning-parser muse \
--dyn-tool-call-parser muse \
--enable-layernorm-sp \
--enable-megamoe-for-layernorm-sp \
--expert-parallel-size 4 \
--kv-cache-dtype fp8_e4m3 \
--max-running-requests 64 \
--mem-fraction-static 0.85 \
--mm-feature-transport cpu \
--moe-runner-backend flashinfer_trtllm \
--prefill-decode-interval 24 \
--quantization modelopt_fp4 \
--reasoning-parser muse \
--speculative-algorithm DFLASH \
--speculative-draft-attention-backend fa4 \
--speculative-draft-model-path RadixArk/Muse-Spark-Open-DFlash2 \
--speculative-num-draft-tokens 8 \
--swa-full-tokens-ratio 0.15 \
--tensor-parallel-size 4 \
--trust-remote-code \
--watchdog-timeout 3600
```
</Tab>
</Tabs>
</Accordion>
<Accordion title="Agg TP4·EP4 + HiCache, max throughput · 4 GPUs · concurrency 160">
<Tabs sync={false}>
<Tab title="Env Vars">
```bash
export ETCD_IP=<node running etcd>
export AGG_HEAD_IP=<first node of the worker>
```
</Tab>
<Tab title="etcd">
```bash
etcd \
--data-dir /tmp/etcd \
--listen-client-urls http://0.0.0.0:2379 \
--advertise-client-urls http://$ETCD_IP:2379
```
</Tab>
<Tab title="Dynamo frontend">
```bash
export DYN_REQUEST_PLANE=tcp
export DYN_ROUTER_QUEUE_THRESHOLD=None
export DYN_ROUTER_TEMPERATURE=0
export DYN_TOKENIZER=fastokens
export DYN_TOKENIZER_CACHE=1
export DYN_TOKENIZER_CACHE_BYTES=8000000000
export ETCD_ENDPOINTS=http://$ETCD_IP:2379
python3 -m dynamo.frontend \
--http-port 8000 \
--router-mode kv \
--router-session-affinity-ttl-secs 14400 \
--active-decode-blocks-threshold None \
--active-prefill-tokens-threshold None \
--active-prefill-tokens-threshold-frac None
```
</Tab>
<Tab title="SGLang worker">
```bash
export DYN_REQUEST_PLANE=tcp
export ETCD_ENDPOINTS=http://$ETCD_IP:2379
export NCCL_CUMEM_ENABLE=1
export SGLANG_OPT_UNIFIED_CACHE_FREE_OUT_OF_WINDOW_SLOTS=1
export SGLANG_TIMEOUT_KEEP_ALIVE=900
python3 -m dynamo.sglang \
--model-path meta-models/Muse-Spark-Open-816B-A42B-NVFP4 \
--served-model-name meta-models/Muse-Spark-Open-816B-A42B-NVFP4 \
--host 0.0.0.0 \
--port 30000 \
--kv-events-config '{"publisher":"zmq","topic":"kv-events","endpoint":"tcp://*:5557"}' \
--request-plane tcp \
--cuda-graph-max-bs-decode 176 \
--dyn-reasoning-parser muse \
--dyn-tool-call-parser muse \
--enable-hierarchical-cache \
--enable-layernorm-sp \
--enable-megamoe-for-layernorm-sp \
--expert-parallel-size 4 \
--hicache-size 120 \
--kv-cache-dtype fp8_e4m3 \
--max-running-requests 176 \
--mem-fraction-static 0.9 \
--mm-feature-transport cpu \
--moe-runner-backend flashinfer_trtllm \
--prefill-decode-interval 6 \
--quantization modelopt_fp4 \
--reasoning-parser muse \
--speculative-algorithm DFLASH \
--speculative-draft-attention-backend fa4 \
--speculative-draft-model-path RadixArk/Muse-Spark-Open-DFlash2 \
--speculative-num-draft-tokens 8 \
--swa-full-tokens-ratio 0.15 \
--tensor-parallel-size 4 \
--trust-remote-code \
--watchdog-timeout 3600
```
</Tab>
</Tabs>
</Accordion>
</AccordionGroup>
### MI355X deployment measured with SemiAnalysis InferenceX
TP8·EP8 with DFlash V2 on 8× MI355X, measured in [SemiAnalysis InferenceX](https://inferencex.semianalysis.com/) at concurrency 1–64.
- **Container image.** `lmsysorg/sglang-rocm:muse-spark-open-1020`.
- **HiCache.** At concurrency 16 and above, add the HiCache flags below. `--hicache-size` is the host tier in GB per rank, so 253 GB needs about 2 TB of host memory across the eight ranks.
```bash Command
SGLANG_USE_AITER=1 \
SGLANG_USE_AITER_UNIFIED_ATTN=1 \
MUSE_MOE_BACKEND=flydsl \
python -m sglang.launch_server \
--model-path amd/Muse-Spark-Open-816B-A42B-Quark-MXFP4 \
--tp-size 8 \
--ep-size 8 \
--attention-backend aiter \
--page-size 16 \
--kv-cache-dtype fp8_e4m3 \
--mem-fraction-static 0.85 \
--max-running-requests 128 \
--swa-full-tokens-ratio 0.2 \
--prefill-decode-interval 20 \
--scheduler-recv-interval 20 \
--tokenizer-worker-num 4 \
--speculative-algorithm DFLASH \
--speculative-draft-model-path RadixArk/Muse-Spark-Open-DFlash2 \
--speculative-draft-attention-backend triton \
--speculative-num-draft-tokens 8 \
--speculative-draft-window-size 2048 \
--reasoning-parser muse \
--tool-call-parser muse \
--watchdog-timeout 1800 \
--host 0.0.0.0 \
--port 30000
```
```bash HiCache (concurrency 16 and above)
--enable-hierarchical-cache \
--hicache-size 253 \
--hicache-write-policy write_through \
--hicache-io-backend kernel \
--hicache-mem-layout page_first
```
## 1. Model introduction
This cookbook provides deployment recipes for the Muse Spark Open 816B-A42B checkpoints:

Some files were not shown because too many files have changed in this diff Show More