Commit Graph

50 Commits

Author SHA1 Message Date
hlu1
1e85589dc5 Make fp4_quantize kernels work on sm103 (#9807)
Signed-off-by: Hao Lu <14827759+hlu1@users.noreply.github.com>
2025-08-29 21:15:08 -07:00
Kaixi Hou
5c34b4f1c7 [NVIDIA] [2/N] Optimize silu_and_mul_scaled_fp4_grouped_quant perf (#9556) 2025-08-29 17:17:03 -07:00
hlu1
7a16db9bd9 Make sm100 fp8 kernels available on sm103 (#9789)
Signed-off-by: Hao Lu <14827759+hlu1@users.noreply.github.com>
2025-08-28 23:47:29 -07:00
Kaixi Hou
e5638573c1 [NVIDA] [1/N] Nvfp4 Masked Gemm: Add quant op for the flashinfer grouped gemm (#9200) 2025-08-22 12:19:45 -07:00
Hubert Lu
c6c379ab31 [AMD] Reorganize hip-related header files in sgl-kernel (#9320) 2025-08-18 16:53:44 -07:00
jy-song-hub
4fc09e0df0 Fp4 MOE quant kernel optimization (#8777)
Co-authored-by: Rain Jiang <96632942+rainj-me@users.noreply.github.com>
2025-08-15 01:46:16 -07:00
Yuan Luo
432f2053dd [sgl-kernel] 1/N Refactor sglang cutlass 3x - gemm fp8 blockwise sm90 (#8913)
Co-authored-by: luoyuan.luo <luoyuan.luo@antgroup.com>
2025-08-14 10:55:54 -07:00
Peng Zhang
5aa1ebd242 [2/n]decouple quantization implementation from vLLM dependency (#8112)
Co-authored-by: walker-ai <yiyun.wyt@antgroup.com>
Co-authored-by: leoneo <1320612015@qq.com>
2025-08-14 03:19:03 -07:00
henryg
841810f227 [Perf] Tunings for SM100 FP8 CUTLASS kernel (#8818) 2025-08-13 21:59:22 -07:00
triple-mu
444013585d Fix typos and unify size(s)/stride(s) API calls (#8799) 2025-08-08 00:18:08 -07:00
Xiaoyu Zhang
f57d2dc162 [sgl-kernel] avoid per_token_quant_fp8.cu hardcode sm_count (#8738) 2025-08-04 12:55:57 +08:00
Stefan He
db7343c992 fix per token cuda kernel hidden dim cannot divide by 16 (#8543) 2025-08-01 09:27:18 -07:00
Peter Pan
6bdd27861b [Kimi K2] dsv3_router_gemm supports NUM_EXPERTS == 384 (#8013) 2025-08-01 22:01:24 +08:00
Xiaoyu Zhang
7a4309cc8a [sgl-kernel performace] fix fp8 quant kernels dispatch __nv_fp8_e4m3 bug to improve performance 10%-20% (#8499)
Co-authored-by: Ke Bao <ispobaoke@gmail.com>
2025-07-29 23:31:54 +08:00
strgrb
fb4ce17de6 Fix per_token_group_quant_8bit when hidden_dim // group_size is not divided by 4. (#8449)
Co-authored-by: Zhang Kaihong <zhangkaihong.zkh@alibaba-inc.com>
2025-07-28 01:32:46 -07:00
Yuan Luo
0c8dab9e67 [sgl-kernel] Opt per_token_quant_fp8 with warp reduce (#8130)
Co-authored-by: luoyuan.luo <luoyuan.luo@antgroup.com>
2025-07-23 21:22:59 +08:00
Baizhou Zhang
282eb59ff3 Add bf16 output option for dsv3_router_gemm kernel (#7999) 2025-07-20 09:49:37 +08:00
Qi Yuhang
6e92da8fca [Fix][Ready]Fix register spilling in cutlass nvfp4 gemm kernel on Blackwell (#8127) 2025-07-17 20:49:36 -07:00
likesen-alibaba
4a0d19198b Fix bug of deepseek-v3 under DP+EP mode with large batchsize/seqlen (#6449) 2025-07-10 01:19:56 -07:00
Baizhou Zhang
7248272ccc Add dsv3 router gemm kernel (#7627) 2025-06-29 23:31:55 -07:00
Ke Bao
04b35190e2 Add dsv3 fused a gemm to sgl-kernel (#7630) 2025-06-29 02:52:24 -07:00
AniZpZ
3eb4a800e8 Fix AWQ Dequant and Weight Loading of deepseek v2 (#6842) 2025-06-17 13:45:10 -07:00
fzyzcjy
5c66c4424f Support new DeepGEMM format in per token group quant (#7146) 2025-06-13 02:00:22 -07:00
Pavani Majety
eb38c7d1ca [1/2] Add Kernel support for Cutlass based Fused FP4 MoE (#6093)
Signed-off-by: Pavani Majety <pmajety@nvidia.com>
2025-06-02 13:48:03 -07:00
HandH1998
4d643f6c7a [1/2] Support Qserve (#6457)
Co-authored-by: yych0745 <1398089567@qq.com>
Co-authored-by: sleepcoo <sleepcoo@gmail.com>
2025-05-21 19:48:59 -07:00
Elfie Guo
c23a7072b6 Upgrade CUTLASS 4.0 (#6336)
Co-authored-by: zhyncs <me@zhyncs.com>
2025-05-15 17:42:23 -07:00
Yineng Zhang
6f56614734 chore: upgrade cutlass 3.9.2 (#6004)
Co-authored-by: yizhang2077 <1109276519@qq.com>
2025-05-06 13:34:08 -07:00
Xiaoyu Zhang
5bb0accbcf cutlass 3.9 supported to improve fp8_blockwise_gemm (#5820) 2025-04-28 21:52:36 -07:00
Yineng Zhang
136b8e6afb fix: remove cublas_grouped_gemm (#5307) 2025-04-11 16:22:37 -07:00
Yi Zhang
ebf495f013 sgl-kernel use cutlass latest version for fp8 blockwise gemm (#5207) 2025-04-09 11:47:04 -07:00
Xiaoyu Zhang
2c8fd99363 [sgl-kernel] per token group quant support COLUMN MAJOR (#4817) 2025-04-02 18:29:59 -07:00
Yuhong Guo
ee47a6c1c3 [Build] Fix cuda12.8 build error in nvfp4_scaled_mm_kernels.cu (#4953) 2025-03-31 12:00:34 -07:00
Yineng Zhang
ec3ee0289d fix sgl-kernel cu118 build (#4872) 2025-03-28 17:23:51 -07:00
Yineng Zhang
8bf6d7f406 support cmake for sgl-kernel (#4706)
Co-authored-by: hebiao064 <hebiaobuaa@gmail.com>
Co-authored-by: yinfan98 <1106310035@qq.com>
2025-03-27 01:42:28 -07:00
Yi Pan
45fdf1f7f3 Fix shared memory OOM on sm86 GPUs. (#4797) 2025-03-26 10:41:53 -07:00
Trevor Morris
e9f8e42318 Support FP4 gemm (1/2) (#3899) 2025-03-24 19:50:23 -07:00
Chunan Zeng
65c24c28f9 [Quant Kernel] refactored per token group quant fp8 to support int8 up-to 2x faster (#4396) 2025-03-23 23:44:17 -07:00
AniZpZ
321ab756bc [1/3] fix dsv3 awq issue (#4556)
Co-authored-by: leoneo <1320612015@qq.com>
2025-03-22 01:07:17 -07:00
Chunan Zeng
6a384d5c01 Speed up per token and per tensor quant by 15% (#4639) 2025-03-22 00:37:57 -07:00
Shu Wang
ad4e58bf67 Support fp8 gemm for blackwell (#4558) 2025-03-20 12:40:28 -07:00
Wenbo Yang
75b656488a Support serving DeepSeek-R1-Channel-INT8 with 32 L40S. (#4418) 2025-03-17 00:03:43 -07:00
Yineng Zhang
2937387a50 fix accuracy issue (#4376) 2025-03-13 02:06:22 -07:00
Qingquan Song
4068e01292 Fix per token fp8 quant precision (#4362) 2025-03-12 21:19:05 -07:00
Elfie Guo
7c86671131 Support Blackwell Block Scale FP8 Gemm (#4278) 2025-03-12 14:17:11 -07:00
Rex
07f944631e Add awq dequantize kernel to sgl with 1x to 3x speedup (#4104) 2025-03-12 00:10:02 -07:00
Stefan He
e0917e6bd0 Remove vllm ops scaled fp8 quant and accelerate per token quant by 20-28% (#4215)
Co-authored-by: Stefan He <bhe@linkedin.com>
2025-03-12 00:08:03 -07:00
Xiaoyu Zhang
23308a9032 fix per_token_group_quant_fp8 illegal memory when num_groups % 16 != 0 (#4231) 2025-03-10 01:42:58 -07:00
Lianmin Zheng
aa957102a9 Simplify tests & Fix trtllm custom allreduce registration (#4252) 2025-03-10 01:24:22 -07:00
Lianmin Zheng
eb06dbcbf8 Move rope and bmm into sgl-kernel (#4241) 2025-03-09 18:38:15 -07:00
Lianmin Zheng
8abf74e3c9 Rename files in sgl kernel to avoid nested folder structure (#4213)
Co-authored-by: zhyncs <me@zhyncs.com>
2025-03-08 22:54:51 -08:00