minke.yu/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 10 Packages Projects Releases Wiki Activity
17,598 Commits 2 Branches 0 Tags
f4c17fed0767626cd81ada078ed2fda5d3187f25
Commit Graph
100 Commits
This Branch
This Branch
All Branches
Author SHA1 Message Date
Xiaoyu ZhangandCursor f4c17fed07 [Diffusion] Fuse FLUX.2 NVFP4 FC1, SwiGLU, and FC2 quantization (#37096) 
Co-authored-by: Cursor <cursoragent@cursor.com>
 2026-09-02 08:21:14 +08:00
Xiaoyu ZhangSong BianCursor
 		c593527f33 [Kernel] Add KDA NVFP4 GEMM for Qwen3.x on SM120 (#36865) 
Co-authored-by: Song Bian <biansonghz@gmail.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
 2026-09-02 08:16:54 +08:00
Xiaoyu ZhangSong BianCursor
 		ee462b5899 [Kernel] Add tuned LFM2.5 Triton MoE configs on B300 (#37158) 
Co-authored-by: Song Bian <biansonghz@gmail.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
 2026-09-01 22:55:38 +08:00
Xiaoyu Zhang 5993f91f84 [Kernel] Register merged diffusion agent kernels with KDA backend (#37385) 2026-09-01 22:51:10 +08:00
Xiaoyu Zhang 4c7ff0d906 [CI] Double JIT kernel unit test timeout (#37435) 2026-09-01 22:37:37 +08:00
Xiaoyu ZhangandCursor 1c3ad92438 [Diffusion] Fuse FLUX.2 ModelOpt FP8 producers and QKV packing (#37162) 
Co-authored-by: Cursor <cursoragent@cursor.com>
 2026-09-01 16:14:29 +08:00
Xiaoyu ZhangMickCursor
 		71cee04ebe [Diffusion] Optimize Qwen-Image TP collectives and attention (#36680) 
Co-authored-by: Mick <mickjagger19@icloud.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
 2026-09-01 10:28:37 +08:00
Xiaoyu ZhangandCursor 175973d834 [Diffusion] Fuse Qwen-Image residual norm and NVFP4 quantization (#37129) 
Co-authored-by: Cursor <cursoragent@cursor.com>
 2026-09-01 08:54:10 +08:00
Xiaoyu ZhangandCursor 079afaffb1 [Diffusion] Fuse FLUX.2 gated residual normalization on Blackwell (#37112) 
Co-authored-by: Cursor <cursoragent@cursor.com>
 2026-09-01 08:38:33 +08:00
Xiaoyu ZhangandCursor 6715debb2a [Diffusion] Fuse Qwen-Image FP8 QKV projection and Blackwell epilogue (#37123) 
Co-authored-by: Cursor <cursoragent@cursor.com>
 2026-09-01 08:34:41 +08:00
Xiaoyu Zhang 52e1c24744 [Diffusion] Fuse FLUX.2 token concatenation and NVFP4 quantization (#37141) 2026-08-31 21:33:02 +08:00
Xiaoyu ZhangandSong Bian d60d658f5f [Kernel] Add GB300 Triton MoE configs for GLM-4.5 FP8 (#37159) 
Co-authored-by: Song Bian <biansonghz@gmail.com>
 2026-08-31 21:31:04 +08:00
Xiaoyu Zhang 771e613d96 [Diffusion] Fuse Qwen-Image final adaptive LayerNorm (#37144) 2026-08-31 18:22:20 +08:00
Xiaoyu ZhangandCursor 1ed9bfac2c [Diffusion] Fuse Qwen-Image FP8 norm and activation quantization (#37156) 
Co-authored-by: Cursor <cursoragent@cursor.com>
 2026-08-31 18:19:42 +08:00
Xiaoyu Zhang bb5e619860 [diffusion] perf: absorb Qwen-Image output projection biases (#37116) 2026-08-31 08:25:37 +08:00
Xiaoyu Zhang 5ab97c4f44 [Diffusion] Cache Qwen-Image modulation across serial CFG branches (#37090) 2026-08-31 01:40:29 +08:00
Xiaoyu Zhang 8c28cdd116 [Diffusion][Kernel] Fuse Wan2.2 NVFP4 bias + GELU on Blackwell (#37075) 2026-08-31 01:37:35 +08:00
Xiaoyu Zhang 0e1146d04f [diffusion] optimization: optimize Pi0.5 inference and bounded graph serving (#34599) 2026-08-29 14:47:32 +08:00
Xiaoyu Zhang db6f0a9d53 Refactor JIT kernel and expert-pack directory layout (#36704) 2026-08-29 07:41:25 +08:00
Xiaoyu Zhang 50bc1a3767 [diffusion] Keep Cosmos3 Nano resident on 96 GB GPUs (#36641) 2026-08-29 07:40:46 +08:00
Xiaoyu Zhang 96a4dcdde8 [CI] Slim JIT kernel unit tests (#36887) 2026-08-29 07:26:49 +08:00
Xiaoyu Zhang eebb99c049 [diffusion][kernel] avoid 4D scale-shift autotuning (#36521) 2026-08-28 16:58:19 +08:00
Xiaoyu Zhang 45424d8434 [diffusion][kernel] support transposed residual-gate add (#36504) 2026-08-28 16:54:37 +08:00
Xiaoyu Zhang 96b31770f9 [diffusion] fuse Helios paired transposed RoPE (#36502) 2026-08-28 08:57:37 +08:00
Xiaoyu Zhang 024a7a1031 [diffusion] Fix native LingBot-Video text encoding (#36542) 2026-08-27 22:58:22 +08:00
Xiaoyu Zhang a7e3f590ca [Diffusion] Fuse LongCat residual gate updates (#36577) 2026-08-27 21:07:04 +08:00
Xiaoyu Zhang e061dd1b47 [Diffusion][Kernel] Fuse Wan FFN GELU epilogue (#36592) 2026-08-27 21:06:28 +08:00
Xiaoyu Zhang 1af95ffded [diffusion] Fuse Cosmos3 Nano T2I attention on Hopper (#36571) 2026-08-27 20:59:15 +08:00
Xiaoyu Zhang db4125bb56 [diffusion] Accept mesh benchmark artifacts (#36553) 2026-08-27 20:53:25 +08:00
Xiaoyu Zhang d42fa5e10a [diffusion] align video BCG warmup frame count (#36485) 2026-08-27 20:52:00 +08:00
Xiaoyu Zhang ad911a5ec0 [kernel] Tune LingBot-Video MoE TMA configs for H100 (#36543) 2026-08-27 02:47:32 -07:00
Xiaoyu Zhang 702de26310 [diffusion] make benchmark caches seedable and cover missing native families (#36463) 2026-08-26 19:04:46 +08:00
Xiaoyu Zhang fa3ac61661 [Diffusion] Bound reusable Ulysses A2A staging buffers across shapes (#36327) 2026-08-26 10:57:29 +08:00
Xiaoyu Zhang 46b92b22e2 [diffusion] Accelerate LingBot Video RMSNorm in quality=high (#35969) 2026-08-24 18:02:00 +08:00
Xiaoyu Zhang 9866fe910b [diffusion] Speed up LingBot high-quality VAE decode (#36024) 2026-08-24 14:13:03 +08:00
Xiaoyu Zhang cc74aba330 [diffusion] Honor XDG cache for model overlays (#36019) 2026-08-24 14:11:26 +08:00
Xiaoyu Zhang 6d40b8aebf [diffusion] Fix Hunyuan QKV pack indexing at production video shapes (#36009) 2026-08-24 13:39:26 +08:00
Xiaoyu Zhang b43931e878 [diffusion] Refresh quality and BCG benchmark skills (#36016) 
Signed-off-by: BBuf <1182563586@qq.com>
 2026-08-24 13:37:45 +08:00
Xiaoyu Zhang 344613c159 [diffusion] Default Hunyuan VAE to tiled decode (#36012) 2026-08-24 13:18:06 +08:00
Xiaoyu Zhang 8dcfb3b5e7 [diffusion] Fuse LongCat-Image QKNorm and interleaved RoPE (#35995) 2026-08-24 12:07:26 +08:00
Xiaoyu Zhang 09592f5889 [diffusion] Keep LongLive2 components resident on large GPUs (#35993) 2026-08-24 12:06:52 +08:00
Xiaoyu Zhang e129fe21e5 [diffusion] Flatten Wan VAE RMSNorm row addressing (#35981) 2026-08-24 08:57:54 +08:00
Xiaoyu Zhang b2eb0fa51e [diffusion] Keep Cosmos3 Nano resident on high-memory GPUs (#36000) 2026-08-24 08:51:58 +08:00
Xiaoyu Zhang 447048dba2 [diffusion] Reject unsafe quality=high BCG replay (#36008) 2026-08-24 08:50:26 +08:00
Xiaoyu Zhang f4448e677f [diffusion] Reuse SANA fast paths in SANA-Video BCG (#35961) 2026-08-24 08:47:04 +08:00
Xiaoyu Zhang 96bfd2476c [diffusion] Enable SANA-Video breakable CUDA graphs (#35729) 2026-08-22 12:57:09 +08:00
Xiaoyu Zhang 83e9ece672 [diffusion] Fuse SANA-Video interleaved RoPE (#35695) 2026-08-22 12:56:45 +08:00
Xiaoyu Zhang 39d4d65a51 [diffusion] Accelerate SANA-Video linear attention in quality=high (#35728) 2026-08-21 18:05:43 +08:00
Xiaoyu Zhang a5c52a9358 [diffusion] Enable LongCat breakable CUDA graphs (#35724) 2026-08-21 17:59:16 +08:00
Xiaoyu Zhang 7e80e889a2 [diffusion] Fuse LTX-2.5 decoder 3D RoPE (#35698) 2026-08-21 10:13:09 +08:00
Xiaoyu Zhang 04444ee352 [diffusion] Refresh eager optimization skills and benchmark safeguards (#35679) 2026-08-20 22:03:57 +08:00
Xiaoyu ZhangandClaude Opus 5 9113fc6d93 [docs] Add a fused-kernels page for SGLang Diffusion (#35436) 
Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
 2026-08-19 16:32:42 +08:00
Xiaoyu ZhangandClaude Opus 5 ae6945e112 [kernels] Reorganize ops/diffusion by operator domain behind a lazy facade (#35114) 
Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
 2026-08-18 20:37:43 +08:00
Xiaoyu Zhang 0aa09ab40d [diffusion] Reuse bit-exact modulation fast path for LTX-2.3 (#34930) 2026-08-17 09:04:10 +08:00
Xiaoyu Zhang 41abbb0d32 [diffusion] Accelerate Cosmos3 T2I QKNorm+RoPE (#34932) 2026-08-16 20:15:32 +08:00
Xiaoyu Zhang 095ec6c997 [diffusion][kernel] Accelerate Sana BCG with bit-exact conv post-processing (#34928) 2026-08-16 20:05:57 +08:00
Xiaoyu Zhang 0761d3f3a4 [diffusion] Accelerate lossless Ideogram norm post-processing (#34931) 2026-08-16 17:20:09 +08:00
Xiaoyu Zhang b752f1e533 [diffusion] Enable breakable CUDA graphs for LTX-2.3 (#34929) 2026-08-16 17:18:43 +08:00
Xiaoyu Zhang 0c072235f4 [diffusion] Bound overlong weight lock filenames (#34825) 2026-08-15 17:20:30 +08:00
Xiaoyu Zhang 9c9a3273be [diffusion] Fix Helios denoising profiler stepping (#34826) 2026-08-14 23:21:54 +08:00
Xiaoyu Zhang 5f2a6d6422 [diffusion] Fix symbolic replicated-mode counting under torch.compile (#34824) 2026-08-14 23:20:29 +08:00
Xiaoyu ZhangandClaude Fable 5 704e512836 [GDN] Honor configured linear-attn verify backend in the kernel dispatcher (#34592) 
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
 2026-08-14 08:59:52 +08:00
Xiaoyu Zhang c255fbc4fe [Diffusion] Add @triple-mu as a code owner (#34748) 2026-08-14 00:37:49 +08:00
Xiaoyu Zhang ebca0bbde4 [Diffusion][ERNIE] Fuse QKNorm with full-width RoPE (#34620) 2026-08-13 23:23:21 +08:00
Xiaoyu Zhang 82f7afb881 [Diffusion] Make auto residency decisions component-scoped (#34615) 2026-08-13 23:20:55 +08:00
Xiaoyu Zhang 74c0322342 [Diffusion][FLUX.2] Fuse eager AdaLN and packed SwiGLU (#34616) 2026-08-13 19:55:53 +08:00
Xiaoyu Zhang 3c1791a7df [Diffusion][HunyuanVideo] Fuse eager QKV packing and high-quality QKNorm (#34617) 2026-08-13 19:54:00 +08:00
Xiaoyu ZhangandClaude Fable 5 a23670ddbf [diffusion] Wan2.2-TI2V: fuse per-token adaLN table add into contiguous slices + hoist rope cache (denoise -13.1% H100 / -12.6% H200, bit-exact; eager beats compile) (#34584) 
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
 2026-08-13 17:26:46 +08:00
Xiaoyu Zhang 26579d893e [Diffusion][GLM-Image] Retune QK head LayerNorm for SM103 (#34619) 2026-08-13 10:49:13 +08:00
Xiaoyu Zhang 1f008dc226 [Diffusion][LTX-2] Allocate AdaLN outputs from one contiguous slab (#34508) 2026-08-12 16:28:27 +08:00
Xiaoyu Zhang daae3acb36 [Diffusion][Z-Image] Tune native QK RMSNorm launch for SM103 (#34507) 2026-08-12 16:27:44 +08:00
Xiaoyu Zhang 4827061247 [Diffusion] Make weight-only FP8 dequant cache torch.compile-safe (#34506) 2026-08-12 16:26:31 +08:00
Xiaoyu Zhang 84ce7502cf [Diffusion][MiniMax H3] Extend exact QKNorm+RoPE rounding to SM103 (#34505) 2026-08-12 16:25:24 +08:00
Xiaoyu Zhang 45f7063335 [Diffusion] Tune QK head LayerNorm for SM103 (#34503) 2026-08-12 16:24:18 +08:00
Xiaoyu Zhang 22e4b3a81f [Diffusion] Avoid slow cuBLASLt GELU epilogue on SM120 (#34350) 2026-08-12 12:12:00 +08:00
Xiaoyu Zhang 4aff4b1822 [Diffusion] Improve bit-exact fusion fallback diagnostics (#34412) 2026-08-12 10:42:48 +08:00
Xiaoyu Zhang 3f9d184833 [Diffusion] Tune QK head LayerNorm for SM120 (#34349) 2026-08-12 10:41:54 +08:00
Xiaoyu Zhang a53d3636ce [diffusion][model] Add native SANA-Video T2V support (#32921) 2026-08-12 10:07:24 +08:00
Xiaoyu ZhangandClaude Fable 5 37c631ef23 [diffusion] Ideogram-4: fuse Qwen3-style RoPE and SwiGLU silu-mul (denoise -5.1% H100 / -4.7% H200, bit-exact) (#34314) 
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
 2026-08-12 09:19:35 +08:00
Xiaoyu Zhang b1b8ce715b [Diffusion][MiniMax H3] Fix SM120 QKNorm+RoPE rounding (#34347) 2026-08-12 09:18:16 +08:00
Xiaoyu ZhangandClaude Fable 5 546965fc72 [diffusion] LTX-2: mount the bit-exact fused modulate at the 8 bare adaLN sites (ltx23-one-stage denoise -2.8% H100 / -2.6% H200) (#34315) 
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
 2026-08-11 18:23:08 +08:00
Xiaoyu ZhangandClaude Fable 5 071f0f1e9d [diffusion] ERNIE-Image: fuse rotate-half RoPE + GELU-mul and hoist rope cos/sin (denoise -16.2% H100 / -12.7% H200, bit-exact) (#34306) 
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
 2026-08-11 18:18:37 +08:00
Xiaoyu ZhangandClaude Fable 5 ba3dc16401 [diffusion] weight-only FP8: dequantize linear weights once at first use (Ideogram-4 denoise -18.8% H200 / -7.8% H100, bit-exact) (#34305) 
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
 2026-08-11 18:15:26 +08:00
Xiaoyu ZhangandClaude Fable 5 f5f0c3ee7a [diffusion] Z-Image single-GPU BCG: fix the replay crash and make output bit-exact vs eager (#34183) (#34210) 
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
 2026-08-10 22:18:32 +08:00
Xiaoyu ZhangCursorClaude Fable 5
 		fd3036523a [diffusion] Clean up shared bitexact gates, helpers, and stale naming (#34180) 
Co-authored-by: Cursor <cursoragent@cursor.com>
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
 2026-08-10 22:14:09 +08:00
Xiaoyu ZhangandClaude Opus 4.8 441910f926 [diffusion] LTX-2 quality=high fused RMSNorm+modulate + FFN GELU epilogue (H200 ltx23-one-stage denoise 45.85->43.24 s, ~matches torch.compile) (#34172) 
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
 2026-08-10 09:46:17 +08:00
Xiaoyu ZhangandClaude Opus 4.8 56ef810cad [diffusion] BCG: auto-capture the default warmup resolution instead of hard-requiring --warmup-resolutions (H200 SANA denoise 0.73->0.457 s with a single flag) (#34174) 
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
 2026-08-10 08:37:32 +08:00
Xiaoyu Zhang f6cbdc1dd1 docs(diffusion): refresh skills for latest runtime (#34143) 2026-08-09 10:30:46 +08:00
Xiaoyu Zhang 38c007dfe5 [diffusion] FLUX.1: route the adaLN LN+modulate sites through the bit-exact fused LayerNorm+modulate kernel (H200 1024^2 lossless denoise -1.2%, e2e wall -2.9%) (#34126) 2026-08-09 09:52:37 +08:00
Xiaoyu ZhangandClaude Fable 5 6424fec326 [diffusion] Bit-exact data-movement elimination for the Wan causal VAE decoder (H200 LongLive2 704x1280x61f: decode 2.80->2.32 s lossless / 2.12->1.67 s quality=high, e2e -10.7%) (#34125) 
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
 2026-08-09 09:50:56 +08:00
Xiaoyu ZhangandClaude Fable 5 33ed5d4413 [diffusion] perf_logger: SYNC_STAGE_PROFILING must drain the GPU queue for stage records too (fixes 2-3x inflated DecodingStage readings) (#34124) 
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
 2026-08-09 09:49:52 +08:00
Xiaoyu Zhang dc9624deb2 [diffusion] Clean up kernels and shared fast paths (#34085) 2026-08-09 00:37:00 +08:00
Xiaoyu ZhangandClaude Fable 5 891445676c [diffusion] Sana: bit-exact fused aten LayerNorm+modulate under BCG (H200 denoise -4.8%) (#34015) 
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
 2026-08-08 16:05:25 +08:00
Xiaoyu ZhangandClaude Fable 5 5dffa06fe1 [diffusion] GLM-Image bit-exact fused aten LayerNorm+modulate / qk-LN (H200 30-step denoise -8.1%) (#34008) 
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
 2026-08-08 13:26:11 +08:00
Xiaoyu Zhang 148f15b0af [diffusion] FLUX.1 fused adaLN modulate (bit-exact) + RoPE cache hoist, LN-affine folding behind quality=high (H200 e2e -3.5% lossless / -6.9% high) (#34004) 2026-08-08 13:07:42 +08:00
Xiaoyu ZhangandClaude Fable 5 6c7498113f [diffusion] Enable breakable CUDA graph for SANA (H200 1024px e2e -26%, bit-exact) (#33989) 
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
 2026-08-07 23:54:39 +08:00
Xiaoyu Zhang d4be483efb [diffusion] Enable breakable CUDA graph for LTX-2 (H200 two-stage e2e 10.75 s -> 6.90 s, 1.56x) (#33885) 2026-08-07 22:29:14 +08:00
Xiaoyu ZhangandClaude Fable 5 572434e2f6 [diffusion] Z-Image bit-exact fused qk-norm (H200 Turbo 1024px e2e -6.4%) (#33886) 
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
 2026-08-07 17:03:28 +08:00
Xiaoyu ZhangandClaude Fable 5 591cfb0881 [diffusion] FLUX.2 bit-exact residual-gate fast path (H200 klein-4B 50-step denoise -1.2%) (#33823) 
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
 2026-08-06 22:54:40 +08:00
Xiaoyu ZhangandClaude Fable 5 dd98c9572a [diffusion] Generalize the FLUX.2 VAE decoder fast path to AutoencoderKL (Z-Image / FLUX.1) behind quality=high (#33818) 
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
 2026-08-06 22:53:15 +08:00