biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[Perf][Kernel] Optimize FP4 quantization kernels (SM100F) (#32520)

Browse Source 
Signed-off-by: LopezCastroRoberto <rocastro@redhat.com>
This commit is contained in:
Roberto L. Castro
2026-01-25 02:45:27 +01:00
committed by GitHub
parent 1ebdff412a
commit fcb9df99bd
18 changed files with 508 additions and 151 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
csrc/quantization/fp4/nvfp4_experts_quant.cu
View File

@@ -140,8 +140,8 @@ __global__ void __launch_bounds__(512, VLLM_BLOCKS_PER_SM(512))
CVT_FP4_NUM_THREADS_PER_SF>(
rowIdx_in_expert, colIdx, numKTiles, SFout_in_expert);
out_pos =
cvt_warp_fp16_to_fp4<Type, UE8M0_SF>(quant_input, SFScaleVal, sf_out);
out_pos = cvt_warp_fp16_to_fp4<Type, CVT_FP4_NUM_THREADS_PER_SF, UE8M0_SF>(
quant_input, SFScaleVal, sf_out);
}
}
@@ -246,8 +246,8 @@ __global__ void __launch_bounds__(1024, VLLM_BLOCKS_PER_SM(1024))
CVT_FP4_NUM_THREADS_PER_SF>(
rowIdx_in_expert, colIdx, numKTiles, SFout_in_expert);
out_pos =
cvt_warp_fp16_to_fp4<Type, UE8M0_SF>(quant_input, SFScaleVal, sf_out);
out_pos = cvt_warp_fp16_to_fp4<Type, CVT_FP4_NUM_THREADS_PER_SF, UE8M0_SF>(
quant_input, SFScaleVal, sf_out);
}
}