Commit Graph
100 Commits
Author SHA1 Message Date
Liangsheng Yin 17afd8421f [Spec] Share the grammar mask build and verify-tree staging across spec workers (#32393) 2026-07-25 03:40:21 -07:00
Liangsheng Yin 3c5bf1f6d2 [Spec] Derive NGRAM grammar tree links on the host instead of reading back retrive_next_token (#32380) 2026-07-25 02:25:16 -07:00
Liangsheng Yin cff20a2fbb [Spec] Consolidate the grammar sync decision into ScheduleBatch.grammar_needs_sync (#32353) 2026-07-24 20:42:38 -07:00
Liangsheng Yin a31542ebd9 [Feature] Add leveled invariant-check primitive for nan/inf/oob validity checks (#32308) 2026-07-24 04:27:20 -07:00
Liangsheng Yin f4f15162bc [Fix] Fail fast when a safetensors index references missing shard files (#32279) 2026-07-24 02:18:56 -07:00
Liangsheng Yin d059b0f56e [Fix] Clamp degenerate all-sentinel draft rows to token 0 in dspark _online_combine_kernel (#32277) 2026-07-24 00:53:56 -07:00
Liangsheng Yin 99b29bf188 [Fix] Support ENCODER_ONLY target-verify in the trtllm_mha backend (#32178) 2026-07-23 20:33:39 -07:00
Liangsheng Yin 59ef3b15cc [Fix] Reserve the mamba pool's +1 padding slot in the memory budget solve (#32184) 2026-07-23 14:55:35 -07:00
Liangsheng Yin f35411ee81 [Spec] Enable grammar overlap scheduling for STANDALONE speculative decoding (#32110) 2026-07-23 01:47:27 -07:00
Liangsheng Yin 9b853e6832 [Scheduler] Enable decode retraction ordering under speculative decoding (#32023) 2026-07-23 00:42:56 -07:00
Liangsheng Yin 1b63155efe [Perf] Skip blocks past per-request live length in full-width Triton kernels (#32109) 2026-07-22 22:00:25 -07:00
Liangsheng Yin 98cc8d91cf [Fix] Evict only the KV shortfall in evict_from_tree_cache (#32016) 2026-07-22 12:41:06 -07:00
Liangsheng Yin 5c6a29b3c3 [Fix] Unify pinned host pool release on graceful shutdown (#32029) 2026-07-22 12:29:34 -07:00
Liangsheng Yin 024639a372 [Perf] Stack dspark dense draft per-layer ctx KV projection into one GEMM (#31986) 2026-07-21 17:46:13 -07:00
Liangsheng Yin 8ae0eb83fc [Perf] Fold dspark dense draft embedding into the draft graph via forward_embed (#31985) 2026-07-21 17:00:42 -07:00
Liangsheng Yin 08cb081ed0 [Perf] Skip page-table columns past kv length in DSA draft-extend metadata kernel (#31981) 2026-07-21 16:20:14 -07:00
Liangsheng Yin cb324cc958 Add wangfakang to CI_PERMISSIONS.json (#31980) 2026-07-21 14:03:00 -07:00
Liangsheng Yin 7fe9ad25ac [Spec] Extract DFlash compact draft-cache rebuild helpers (#31677) 2026-07-20 15:07:55 -07:00
Liangsheng Yin ff6c755952 [Refactor] Unify logprob results into a single LogprobResult and rename chunk env vars (#31733) 2026-07-20 12:44:27 -07:00
Liangsheng Yin b15a83983c [Fix] Release hierarchical cache host pool on graceful shutdown (#31746) 2026-07-19 22:27:29 -07:00
Liangsheng Yin b3570a4531 [Refactor] Unify input logprob processing on a single chunked path (#31655) 2026-07-19 16:03:23 -07:00
Liangsheng Yin cce5fe7696 [Scheduler] Move the WAR barrier to right after each run_batch launch (#31687) 2026-07-18 20:38:34 -07:00
Liangsheng Yin 10908a6793 [Fix] Respect cache_protected_len in ChunkCache and disabled-radix release paths (#31662) 2026-07-18 12:04:25 -07:00
Liangsheng Yin 639261f7b2 [Refactor] Move output logprob processing into the logprob_processor layer (#31624) 2026-07-17 22:33:52 -07:00
Liangsheng Yin 19c53c44a0 [Fix] Account zero-logprob sequences correctly in chunked logprob stitching (#31639) 2026-07-17 22:33:27 -07:00
Liangsheng Yin 2c856abbe3 [Fix] Enable chunked input-logprob processing by default to cap peak memory (#31498) 2026-07-17 13:19:03 -07:00
Liangsheng Yin 19f4859b30 [CI] Exclude current process memory from GPU idle check (#31571) 2026-07-17 02:11:41 -07:00
Liangsheng Yin 6e3be088a9 [Spec] Allocate the verify tree-mask scratch on the target backend only (#31527) 2026-07-16 23:23:21 -07:00
Liangsheng Yin 45824c69ca fa3: build the topk>1 verify replay page table on-device (#31381) 2026-07-16 21:21:01 -07:00
Liangsheng Yin 27ad9d11b1 [CI] Wait for GPU memory release before each test class setUpClass (#31509) 2026-07-16 21:06:29 -07:00
Liangsheng Yin dfa6278370 [Fix] Publish idle scheduler metrics immediately when the running-reqs gauge is stale (#31517) 2026-07-16 21:04:10 -07:00
Liangsheng Yin 9910ef8167 flashmla: sync-free spec via device-side draft-extend (#31090) 2026-07-16 19:38:12 -07:00
Liangsheng Yin 640101da13 [Fix] Deflake score engine determinism tests (#31494) 2026-07-16 15:42:47 -07:00
Liangsheng Yin 9a4d640244 [Perf] Cache uniform ragged-verify layout for DSpark verify-all compact (#31434) 2026-07-16 15:00:41 -07:00
Liangsheng Yin fc1e3797b7 [Spec] Split the capture width from num_tokens_per_req and gate replay on it (#31255) 2026-07-16 14:59:11 -07:00
Liangsheng Yin 44f4ea917c fa3: sync-free eagle spec via fixed-window draft-extend metadata (#31364) 2026-07-16 12:31:50 -07:00
Liangsheng Yin b55228cfdb [Spec] Consolidate the verify step into eagle_worker_common.run_eagle_verify (#31380) 2026-07-16 00:56:54 -07:00
Liangsheng Yin 0b372d03de [CI] Guard partition consumers against failed check-changes and degenerate fits (#31416) 2026-07-15 22:09:40 -07:00
Liangsheng Yin b871a509e6 [Fix] Wire the detokenizer soft watchdog into the multi-http-worker event loop (#31392) 2026-07-15 21:55:46 -07:00
Liangsheng Yin 1cc9493747 [Spec] Converge DP-attention spec width scaling onto num_tokens_per_req (#31244) 2026-07-15 17:57:08 -07:00
Liangsheng Yin b0b2dfbda1 [Spec] Extract the shared draft() tail into build_eagle_verify_input (#31375) 2026-07-15 15:59:31 -07:00
Liangsheng Yin e76cc75cfa [CI] Remove nightly registrations redundant with scheduled stage runs (#31371) 2026-07-15 14:12:07 -07:00
Liangsheng Yin dc078ddd2a [Spec] Extract stateless draft prepare helpers into eagle_worker_common (#31257) 2026-07-15 10:54:31 -07:00
Liangsheng Yin 76dc427806 [Spec] Single-source num_tokens_per_req derivation and access (#31013) 2026-07-14 18:41:08 -07:00
Liangsheng Yin ca0ee3f1a8 [Spec] Consolidate spec-worker weight updates into BaseSpecWorker via draft_runners (#31078) 2026-07-14 18:39:46 -07:00
Liangsheng Yin 50d1edaa7f [misc] Move SchedulerRecvSkipper into scheduler_components (#31222) 2026-07-14 16:35:04 -07:00
Liangsheng Yin e2728ac504 [Spec] Remove dead padded_static_len and stale SGLANG_ENABLE_SPEC_V2 references (#30998) 2026-07-13 15:30:31 -05:00
Liangsheng Yinandronhuafeng 2ab531cfcf fa3/fa4: sync-free for all backends and phases (#29589)
Co-authored-by: ronhuafeng <ronhuafeng@users.noreply.github.com>
2026-07-13 15:09:57 -05:00
Liangsheng Yin 48fff1f2bd [Spec] Deduplicate spec-v2 worker lifecycle boilerplate into BaseSpecWorker (#31008) 2026-07-13 13:48:40 -05:00
Liangsheng Yin c0f1f7e062 [Spec] Rename num_tokens_per_bs to num_tokens_per_req (#30977) 2026-07-13 13:47:53 -05:00
Liangsheng Yin 7a82178277 [Fix] Disable FlashInfer allreduce fusion in Nemotron-3-Nano lm-eval test (#30945) 2026-07-12 15:39:59 -05:00
Liangsheng Yin 5ba3c5147e [Spec] Add kill-switch env for draft-extend CUDA graph capture (#30944) 2026-07-12 14:54:36 -05:00
Liangsheng Yin 4884f6fbee [Fix] Unify ForwardBatch extend lens cpu fields to their declared list type (#30896) 2026-07-11 17:30:53 -05:00
Liangsheng Yin d8ef76682e [Spec] Extract shared draft worker construction and generalize draft sampler capture (#30857) 2026-07-11 14:34:09 -05:00
Liangsheng Yin 348e6fd29b [Fix] Guard kernel OOB accesses and harden runtime edge cases (#30847) 2026-07-11 14:22:13 -05:00
Liangsheng Yin 7bac9c8cdb [Spec] Enable draft extend cuda graph for DeepSeek-V4 attention backend (#30853) 2026-07-11 04:29:49 -05:00
Liangsheng Yin fef5eda4fb [Refactor] Split DeepSeek-V4 MQALayer into a reusable attention base (#30711) 2026-07-10 01:14:03 -07:00
Liangsheng Yin 23390589f7 [misc] Remove unit test cases that fail the admission criteria (round 3) (#30713) 2026-07-09 19:42:51 -07:00
Liangsheng Yin 7b9b2e4798 [style] Extract init-static values in memory-cache path (#30710) 2026-07-09 19:38:15 -07:00
Liangsheng Yin ccd2028def [style] Extract init-static values in tokenizer + multimodal path (#30709) 2026-07-09 19:36:34 -07:00
Liangsheng Yin dda61b476e [style] Extract init-static values in forward path (#30708) 2026-07-09 19:35:00 -07:00
Liangsheng Yin b5e75b9423 [style] Extract init-static values in scheduler hot path (#30707) 2026-07-09 19:33:02 -07:00
Liangsheng Yin 504570f425 [Refactor] Make DeepSeek-V4 attention backend tolerate an absent CPU seq_lens mirror (#30695) 2026-07-09 16:39:47 -07:00
Liangsheng Yin a36c873147 [misc] Remove unit test cases that fail the admission criteria (round 2) (#30703) 2026-07-09 16:35:59 -07:00
Liangsheng Yin c53559ba10 [misc] Remove unit test cases that fail the admission criteria (#30690) 2026-07-09 15:31:28 -07:00
Liangsheng Yin c46b291c77 [misc] Add init-static value extraction to the general code style rule (#30701) 2026-07-09 15:25:53 -07:00
Liangsheng Yin 078f06fbf4 [Refactor] Share chat encoding dispatch between serving and offline tools (#30623) 2026-07-09 02:45:42 -07:00
Liangsheng Yin bc5d376c2c [Bench] Add fixed-prompt mode and per-request spec accept length metrics (#30615) 2026-07-09 02:06:04 -07:00
Liangsheng Yin 61602b95fb [Fix] Prevent silent VLM server crash when /dev/shm is exhausted during multimodal feature transport (#30602) 2026-07-09 00:44:42 -07:00
Liangsheng Yin 64e2a73c80 [Fix] Serialize FanOutCommunicator queueing calls with a FIFO-fair asyncio.Lock (#30606) 2026-07-09 00:04:43 -07:00
Liangsheng Yin 9e4483d725 [misc] Add unit test admission criteria to agent rules (#30608) 2026-07-08 23:41:48 -07:00
Liangsheng Yin 8f9307736a [misc] Add CI-only guards for the FutureMap seq_lens relay (#30471) 2026-07-08 15:29:53 -07:00
Liangsheng Yin 10e7f2925f [Fix] Chain the seq_lens publish event records so prebuilt seeding keeps the forward fence (#30435) 2026-07-08 13:43:09 -07:00
Liangsheng Yin c0d45cda29 [Spec] Add DFLASH basic sanity CI test (#29338) 2026-06-28 23:45:05 -07:00
Liangsheng Yin be17475a1b [Spec] Publish DFLASH verify read-done event for fine-grained WAR barrier (#29541) 2026-06-28 18:51:38 -07:00
Liangsheng Yin 909123ddb8 [misc] Use --cuda-graph-max-bs-decode in tests, examples, and docs (#29591) 2026-06-28 18:38:28 -07:00
Liangsheng Yin abaee4664e [dflash] fa3/fa4: device-side page table; drop seq_lens_cpu D2H sync (#29343) 2026-06-28 18:06:33 -07:00
Liangsheng Yin 6582fc1e5d dflash: drop verify_done barrier; rely on scheduler WAR fallback (#29556) 2026-06-28 01:47:17 -07:00
Liangsheng Yin 593f5ba68f [Spec] Replace shared-infra dflash special-cases with capabilities (WAR barrier + seq_lens_cpu) (#29232) 2026-06-28 00:02:32 -07:00
Liangsheng Yin 4e6ce8e4ab [Spec] Capture DFLASH draft greedy sampling inside the draft decode cuda graph (#29395) 2026-06-27 19:34:26 -07:00
Liangsheng Yin cd6dedf972 [router] Count requests/responses at the HTTP edge for true intake (#29467) 2026-06-26 21:21:41 -07:00
Liangsheng Yin b6ebdcc92e [test] Split perf table out of fwd occupancy kit report (#29348) 2026-06-25 20:23:42 -07:00
Liangsheng Yin e4696ed62d [test] Report token tps in fwd occupancy kit and force ignore_eos (#29332) 2026-06-25 15:08:33 -07:00
Liangsheng Yin 2812a3c93a [Spec] Unify spec/non-spec decode result handling and overlap relay-payload gating (#29225) 2026-06-25 00:44:38 -07:00
Liangsheng Yin 8314247d9d [Spec] Merge dflash triton kernels into a single dflash.py (#29228) 2026-06-24 18:57:15 -07:00
Liangsheng Yin c7734e6871 [Spec] Dissolve EagleDraftInputV2Mixin so spec-info dataclasses hold data only (#29220) 2026-06-24 18:08:30 -07:00
Liangsheng Yin 85a3f1f75c [Spec] Unify the overlap stash relay behind a RelayPayload dataclass (#29124) 2026-06-24 15:36:54 -07:00
Liangsheng Yin c6822f81b1 [Spec] Make the overlap bonus-token relay unconditional (#29122) 2026-06-24 15:34:02 -07:00
Liangsheng Yin d3dd184e3e [Spec] Fold DFlash verified_id into the shared bonus_tokens relay channel (#29118) 2026-06-24 15:30:29 -07:00
Liangsheng Yin 8e1988b746 [Perf] Overlap result D2H copy with the next forward step (#29075) 2026-06-24 01:34:02 -07:00
Liangsheng Yin e3f1fa9d8e [misc] Unify benchmark deprecation shims and one_batch_server CLI entrypoint (#29108) 2026-06-23 21:10:25 -07:00
Liangsheng Yin f76c6c95be [misc] Add sglang.bench_offline_throughput deprecation shim (#29085) 2026-06-23 18:39:58 -07:00
Liangsheng Yin b448b08401 [misc] Move bench_offline_throughput into sglang/benchmark/ with a back-compat shim (#28747) 2026-06-23 18:37:52 -07:00
Liangsheng Yin 9ef1830701 [Scheduler] Extract DFlash prefill refill into a standalone MinFreeSlotsDelayer (#29089) 2026-06-23 17:04:55 -07:00
Liangsheng Yin 11e7c9e0e6 [misc] Add sglang.bench_one_batch deprecation shim (#29082) 2026-06-23 14:57:35 -07:00
Liangsheng Yin c864c8d9c2 [misc] Move bench_one_batch into sglang/benchmark/ with a back-compat shim (#28687) 2026-06-23 14:48:35 -07:00
Liangsheng Yin 6c5f466023 [server_args] Reland FA4 page_size auto-force for combined --attention-backend fa4 (#28976) 2026-06-23 14:21:59 -07:00
Liangsheng Yin ed0a62e4dd [Mem] Add KV-page double-free checks to the invariant checker (#27731) 2026-06-23 02:29:40 -07:00
Liangsheng Yin 854c688121 [Spec] Unify decode KV-commit bookkeeping across spec-v2 workers (#28754) 2026-06-23 00:53:05 -07:00
Liangsheng Yin 4740f23e1f Revert "[server_args] compute mem_fraction_static after dp chunked-prefill division" (#28991) 2026-06-22 20:48:16 -07:00