 Daniel Stokes
|
8ee200972e
|
[fix] Add support for flashinfer MOE A2A to Qwen3 BF16 model path (#26255)
|
2026-07-01 01:59:54 -07:00 |
|
|
ashwini rathiandSinghal, Shubham
|
5134dcdcab
|
[Intel XPU] Initially add nightly GSM8K accuracy tests for Llama-3.1-8B (TP=2) and Qwen3-32B (TP=4) (#28908)
Co-authored-by: Singhal, Shubham <shubham.singhal@intel.com>
|
2026-07-01 16:27:24 +08:00 |
|
|
Bingxu Chen
|
81a472efcc
|
[AMD] Update ROCm AITER pin to 9127c94 (#29816)
|
2026-07-01 01:09:11 -07:00 |
|
|
huangtingwei
|
5e1ccd9320
|
[HiCache] Optimize HiCache hash generation with bulk token byte conversion (#28287)
|
2026-07-01 15:44:23 +08:00 |
|
|
Xiaoyu Zhang
|
8205aa3603
|
chore: clean diffusion dead code (#29789)
|
2026-07-01 15:42:16 +08:00 |
|
|
Shijin Zhang
|
00e12cebb4
|
[Fix]: Defer DSA MLA CP KV gather for fp8 trtllm prefill in PD mode (#29161)
Signed-off-by: Shijin Zhang <75300765+Dovis01@users.noreply.github.com>
|
2026-07-01 00:20:12 -07:00 |
|
|
zhaozx-cn
|
721350656d
|
[NPU] Fix glm 4.6v (#29381)
|
2026-07-01 15:13:48 +08:00 |
|
|
Jialin Ouyang
|
40594bd381
|
[passthrough] engine: zstd request-body decompression + header overrides (#29684)
|
2026-06-30 23:48:56 -07:00 |
|
|
Xiaoyu Zhang
|
47ae1241d3
|
[CI] Relax ModelOpt NVFP4 diffusion consistency thresholds (#29767)
|
2026-07-01 14:45:08 +08:00 |
|
|
Xiaoyu Zhang
|
fcb9f229b3
|
[KDA-Pilot] Add LTX2 QKNorm split-RoPE CUDA fast path (#29708)
|
2026-07-01 14:42:07 +08:00 |
|
|
Chang Min Bark
|
41779d56fd
|
fix: make write_token dynamic (#29271)
|
2026-06-30 22:55:57 -07:00 |
|
|
Tai An
|
546d6e23ad
|
fix(mlx): default prefill_aware_swa=False on MlxModelRunnerStub (#29681)
|
2026-06-30 22:55:43 -07:00 |
|
|
Lijuan Tang
|
3cdc2415b1
|
[MLX] Fix step-bounded profiling for bench tools on Apple Silicon (#29217)
|
2026-06-30 22:55:29 -07:00 |
|
|
|
5de66d7fbc
|
[EPD][BugFix] Fix encoder health check with global cache TP (#29296)
Co-authored-by: 晟海 <huangtingwei.htw@antgroup.com>
Co-authored-by: siyu <liusy58@linux.alibaba.com>
Co-authored-by: Shangming Cai <csmthu@gmail.com>
|
2026-07-01 11:04:03 +08:00 |
|
|
zhangxiaolei
|
56f22cd520
|
Optimize C128 state pool allocation using request state pool (#28612)
|
2026-06-30 19:11:29 -07:00 |
|
|
Shijin Zhang
|
5b76f55d90
|
[Fix]: Inline H2D during CUDA graph capture to avoid stream isolation in Offloader (#29166)
Signed-off-by: Shijin Zhang <75300765+Dovis01@users.noreply.github.com>
|
2026-06-30 17:49:58 -07:00 |
|
|
weireweireandweireweire
|
a01afdd526
|
Support real draft tokens to simulated acceptance (#29645)
Co-authored-by: weireweire <20922698+weireweire@users.noreply.github.com>
|
2026-06-30 17:22:05 -07:00 |
|
|
|
53c61bd5e6
|
[bug2] skip swa recovery on locked full kv (#29352)
Co-authored-by: Zhangheng <hzh0425@apache.org>
Co-authored-by: Hanming Lu <69857889+hanming-lu@users.noreply.github.com>
|
2026-06-30 16:07:06 -07:00 |
|
|
Yuzhen Zhou
|
99b8f36cb1
|
Skip custom all-reduce v2 CUDA graph capture with torch memory saver. (#27948)
|
2026-06-30 14:53:29 -07:00 |
|
|
TzZtzt
|
e721259079
|
typo-fix: correct mooncake package name to "mooncake-transfer-engine" (#25153)
|
2026-06-30 11:54:36 -07:00 |
|
|
Feng Yaoandispobock
|
1c2523b719
|
[bug1] keep full kv when swa skips leaf data (#29351)
Co-authored-by: ispobock <ispobaoke@gmail.com>
|
2026-07-01 00:51:18 +08:00 |
|
|
|
cf36dca6d4
|
[JIT Kernel] Triton moe fused gate (#25835)
Co-authored-by: Claude <noreply@anthropic.com>
Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>
Co-authored-by: ziyi.xu <ziyi.xu@radixark.ai>
Co-authored-by: Xiaoyu Zhang <1182563586@qq.com>
|
2026-06-30 23:26:18 +08:00 |
|
|
Yihao WangandClaude Opus 4.8
|
a531d81c19
|
[diffusion][cache-dit] support Krea-2 + run-driven has_separate_cfg (#29688)
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
|
2026-06-30 08:22:05 -07:00 |
|
|
Estrella-xx
|
c6a7c98ae4
|
[NPU]GLM-4.7-Flash optimize with fused kernels (#29509)
|
2026-06-30 19:22:19 +08:00 |
|
|
nvjullin
|
2f730e299f
|
Disable dsr1 prefill cudagraphs by default (#28053)
|
2026-06-30 16:34:44 +08:00 |
|
|
Stanley Winata
|
a5e6dd3767
|
[AMD] Implement QuarkW4A8MXFp4MoE to support amd/gpt-oss-120b-w-mxfp4-a-fp8 (#27204)
Signed-off-by: Stanley Winata <stanley.winata@amd.com>
|
2026-06-30 01:24:22 -07:00 |
|
|
|
89620b9169
|
[NPU] Support DeepSeek V4 Flash MTP on Ascend (#28980)
Co-authored-by: Kurkur <mccllm@qq.com>
Co-authored-by: gjsheu <gjsheu@163.com>
Co-authored-by: root <root@localhost.localdomain>
Co-authored-by: khalil2ji3mp6 <khalilzhk@gmail.com>
|
2026-06-30 16:22:13 +08:00 |
|
|
cctryandcctry
|
4b4b4af583
|
Budget EAGLE/STANDALONE draft KV pool in SWA pool configurators (#29622)
Co-authored-by: cctry <cctry@fb.com>
|
2026-06-30 00:58:33 -07:00 |
|
|
silencejade
|
ff51acd67b
|
[NPU] Qwen3-VL-8B use split_qkv_rmsnorm_rope for extend (#29627)
|
2026-06-30 15:55:50 +08:00 |
|
|
Mohammad Miadh Angkad
|
bae78a44da
|
[Deps] Bump transformers to 5.12.1 (#29393)
|
2026-06-30 00:54:01 -07:00 |
|
|
Thomas Wang
|
f920a37da4
|
[AMD] Copy decode result on forward_stream instead of copy_stream (#29642)
|
2026-06-29 23:53:48 -07:00 |
|
|
Mohammad Miadh Angkad
|
b6fceaa789
|
[DSA] Use cos_sin_cache for DSA indexer fusion (#29613)
|
2026-06-29 23:49:00 -07:00 |
|
|
|
54e71506b3
|
[AMD][DSV4] Remove per-batch D2H syncs in MTP to avoid bubbles between 2 batches (#29420)
Co-authored-by: amd-danli103 <amd-danli103@users.noreply.github.com>
Co-authored-by: Thomas Wang <thomawan@amd.com>
|
2026-06-29 22:39:59 -07:00 |
|
|
yifei wuandXinyuan Tong
|
f41d455f37
|
[Model] Add HrmTextForCausalLM (Hierarchical Reasoning Model - Text) (#27887)
Co-authored-by: Xinyuan Tong <115166877+JustinTong0323@users.noreply.github.com>
|
2026-06-30 13:32:32 +08:00 |
|
|
Mick
|
2068ae7eec
|
[diffusion] chore: document VAE decode parallel group axes (#29364)
|
2026-06-30 12:14:55 +08:00 |
|
|
Xiaoyu ZhangandClaude Opus 4.8
|
3add35e26d
|
[Diffusion] Reuse shared AlignedVector and tidy jit_kernel/diffusion (#29664)
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
|
2026-06-30 11:38:22 +08:00 |
|
|
Mohammad Miadh Angkad
|
3a72d02415
|
[DSA] Optimize DSA CUDA graph replay metadata generation (#29499)
|
2026-06-29 19:53:23 -07:00 |
|
|
Mick
|
25b6051c70
|
[diffusion] warmup: default to model sampling resolution (declare Z-Image default) (#29519)
|
2026-06-30 10:32:11 +08:00 |
|
|
Ho-Ren (Jack) Chuang
|
bc8b3ab1f5
|
[Kernel] Add SM90 Q8KV8 FP8 Sparse MLA Prefill JIT Kernel with Tests and Benchmark (#25751)
|
2026-06-30 09:00:11 +08:00 |
|
|
Ho-Ren (Jack) Chuang
|
a6bc432fd8
|
[Perf][Kernel] Fuse SiLU+Mul into NVFP4 Expert Quantization for CUTLASS MoE (#18612)
|
2026-06-29 16:51:01 -07:00 |
|
|
eeechoandAliceChenyy
|
f0bf96390b
|
[SM120] Add FlashInfer sparse MLA decode for DSv4-Flash (#27455)
Co-authored-by: AliceChenyy <alicechenyy@users.noreply.github.com>
|
2026-06-29 16:27:16 -07:00 |
|
|
Trevor Morris
|
b8c25bfaa7
|
[NVIDIA] Support flashinfer a2a with flashinfer_trtllm_routed moe (#22394)
|
2026-06-29 16:23:58 -07:00 |
|
|
Khoa PhamandClaude Opus 4.7
|
6bdecb8206
|
[DeepSeek V3] Reland: run routed experts on main stream in dual-stream MoE (#29463)
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
|
2026-06-29 15:05:01 -07:00 |
|
|
Xinyuan Tong
|
45314a9fcb
|
[spec] Fix index_share_for_mtp_iteration being a no-op in EAGLE MTP draft (#29654)
|
2026-06-29 14:52:23 -07:00 |
|
|
Cheng Wanandlch1475369
|
fc96edd297
|
feat(mem_cache): page-major (layer-major within a page) KV/state layout (#29533)
Co-authored-by: lch1475369 <lch1475369@gmail.com>
|
2026-06-29 14:49:54 -07:00 |
|
|
Mohammad Miadh Angkad
|
6c018eb4d1
|
Fix bounded checkpoint prefetching and buffered drop-cache handling (#29156)
|
2026-06-29 21:49:16 +00:00 |
|
|
|
5556631789
|
[Speculative Decoding] Validate vocabulary compatibility in STANDALONE mode (#23838)
Co-authored-by: kpham-sgl <khoa.pham@radixark.ai>
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
|
2026-06-29 14:45:12 -07:00 |
|
|
Brayden Zhong
|
98d0e702c3
|
[GLM-5] Tune the threshold of router GEMM (#29470)
|
2026-06-29 14:19:24 -07:00 |
|
|
cctryandcctry
|
a5c45a12bb
|
CUDA graph executable dedup via cudaGraphExecUpdate (#29625)
Co-authored-by: cctry <cctry@fb.com>
|
2026-06-29 13:56:33 -07:00 |
|
|
|
f480c5f1f9
|
[Spec] Frozen-KV MTP: delay target KV binding to pool init + reset stale draft out_cache_loc (#29616)
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Co-authored-by: bef0rewind <490285+ronhuafeng@users.noreply.github.com>
Co-authored-by: AlejandroParedesLT <99832715+AlejandroParedesLT@users.noreply.github.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
|
2026-06-29 11:57:10 -07:00 |
|