 Jia Guo
|
4ed698a491
|
fix(fa3): no NaN embeddings with fa_skip_kv_cache under piecewise CUDA graph (#27343)
|
2026-06-15 13:46:22 -07:00 |
|
|
Jia Guo
|
da0eeb82f2
|
perf: add --prefill-only-disable-kv-cache to skip KV pool allocation (#23675)
|
2026-05-11 13:10:24 -07:00 |
|
|
Jia Guo
|
fd636410a2
|
Restrict fa_skip_kv_cache to non-MLA backends (#24097)
|
2026-05-09 09:25:02 +00:00 |
|
|
Jia GuoandClaude Opus 4.7
|
bead2e3470
|
perf: optimize PCG inductor path for FP8 models (redo of #21734) (#23227)
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
|
2026-04-26 20:34:27 -07:00 |
|
|
Jia GuoandClaude Opus 4.6
|
587fd15bd2
|
perf: eliminate attention DtoD copy by passing pre-allocated output to FA (#21985)
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
|
2026-04-24 12:05:16 -07:00 |
|
|
Jia Guo
|
6428392b6f
|
ci: fix cu129 wheel tagging + pipefail-abort in install script (follow-up to #23497) (#23587)
|
2026-04-23 14:52:58 -07:00 |
|
|
Jia Guo
|
b3e6cf60aa
|
ci: build sgl-kernel wheels for both cu129 and cu130 (#23497)
|
2026-04-22 18:08:36 -07:00 |
|
|
Jia GuoandClaude Opus 4.6
|
286fba2073
|
ci: use rerun_failed_jobs for skipped workflows in /rerun-failed-ci (#23008)
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
|
2026-04-21 23:59:15 -07:00 |
|
|
Jia GuoandClaude Opus 4.6
|
6da3aba6a5
|
perf: optimize PCG inductor path for FP8 models (#21734)
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
|
2026-04-14 17:51:27 +08:00 |
|
|
Jia GuoandClaude Opus 4.6
|
bc16130a17
|
ci: skip full rerun when sgl-kernel wheel already built (#22534)
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
|
2026-04-13 20:32:55 -07:00 |
|
|
Jia GuoandClaude Opus 4.6
|
a2b5111962
|
perf: skip KV cache in FA backend for embedding mode (#21971)
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
|
2026-04-13 16:27:52 -07:00 |
|
|
Jia GuoandClaude Opus 4.6
|
5cb4ea1d4d
|
perf: enable inductor combo_kernels for horizontal fusion (#21977)
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
|
2026-04-11 01:01:14 +08:00 |
|
|
Jia GuoandClaude Opus 4.6
|
ec01ef9092
|
Fix torch.compile/dynamo crash with Qwen3 QK-norm in piecewise CUDA g… (#19818)
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
|
2026-03-20 18:05:09 +08:00 |
|
|
Jia Guo
|
87549f8f0b
|
perf(mamba): use Triton conv1d for non-contiguous input to avoid .contiguous() copy (#20469)
|
2026-03-19 19:38:46 -07:00 |
|