chenchenghao/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 11 Packages Projects Releases Wiki Activity

Update Triton configs for block fp8 kernels (#2641)

Browse Source
This commit is contained in:
HandH1998
2024-12-29 22:53:47 +08:00
committed by GitHub
parent 30828e7192
commit afa0341e57
43 changed files with 2321 additions and 1448 deletions
Download Patch File Download Diff File Expand all files Collapse all files

5
python/sglang/srt/layers/quantization/fp8_kernel.py
View File

@@ -309,15 +309,14 @@ def w8a8_block_fp8_matmul(
config = configs[min(configs.keys(), key=lambda x: abs(x - M))]
else:
# Default config
# Block-wise quant: BLOCK_SIZE_N must be divisable by block_size[0]
# BLOCK_SIZE_K must be divisable by block_size[1]
# Block-wise quant: BLOCK_SIZE_K must be divisable by block_size[1]
config = {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": block_size[0],
"BLOCK_SIZE_K": block_size[1],
"GROUP_SIZE_M": 32,
"num_warps": 4,
"num_stages": 2,
"num_stages": 3,
}
def grid(META):