EngineX-Hygon/sglang
5
0
Fork 0
Code Issues Pull Requests Actions 7 Projects Releases Wiki Activity

Support serving DeepSeek-R1-Channel-INT8 with 32 L40S. (#4418)

Browse Source
This commit is contained in:
Wenbo Yang
2025-03-17 15:03:43 +08:00
committed by GitHub
parent 0f52fb55ec
commit 75b656488a
7 changed files with 489 additions and 11 deletions
Download Patch File Download Diff File Expand all files Collapse all files

19
python/sglang/srt/layers/attention/triton_ops/extend_attention.py
View File

@@ -341,12 +341,21 @@ def extend_attention_fwd(
else:
BLOCK_M, BLOCK_N = (32, 64)
elif is_cuda_available and CUDA_CAPABILITY[0] >= 8:
if Lq <= 128:
BLOCK_M, BLOCK_N = (128, 128)
elif Lq <= 256:
BLOCK_M, BLOCK_N = (64, 64)
# 8.9 has a much smaller shared memory size (100K) than 8.0 (160K)
if CUDA_CAPABILITY[1] == 9:
if Lq <= 128:
BLOCK_M, BLOCK_N = (64, 128)
elif Lq <= 256:
BLOCK_M, BLOCK_N = (64, 64)
else:
BLOCK_M, BLOCK_N = (32, 32)
else:
BLOCK_M, BLOCK_N = (32, 64)
if Lq <= 128:
BLOCK_M, BLOCK_N = (128, 128)
elif Lq <= 256:
BLOCK_M, BLOCK_N = (64, 64)
else:
BLOCK_M, BLOCK_N = (32, 64)
else:
BLOCK_M, BLOCK_N = (64, 64) if Lq <= 128 else (32, 32)

146
python/sglang/srt/layers/moe/fused_moe_triton/configs/E=256,N=64,device_name=NVIDIA_L20,dtype=int8_w8a8.json Normal file
View File

@@ -0,0 +1,146 @@
{
"1": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 32,
"num_warps": 4,
"num_stages": 2
},
"2": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 64,
"num_warps": 4,
"num_stages": 2
},
"4": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 64,
"num_warps": 4,
"num_stages": 2
},
"8": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 16,
"num_warps": 4,
"num_stages": 2
},
"16": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 64,
"num_warps": 4,
"num_stages": 2
},
"24": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 32,
"num_warps": 4,
"num_stages": 2
},
"32": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 64,
"num_warps": 8,
"num_stages": 2
},
"48": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 16,
"num_warps": 4,
"num_stages": 3
},
"64": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 64,
"num_warps": 4,
"num_stages": 3
},
"96": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 64,
"num_warps": 4,
"num_stages": 3
},
"128": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 16,
"num_warps": 4,
"num_stages": 3
},
"256": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 32,
"num_warps": 8,
"num_stages": 2
},
"512": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 64,
"num_warps": 4,
"num_stages": 2
},
"1024": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 32,
"num_warps": 4,
"num_stages": 2
},
"1536": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 32,
"num_warps": 4,
"num_stages": 2
},
"2048": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 16,
"num_warps": 4,
"num_stages": 2
},
"3072": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 16,
"num_warps": 4,
"num_stages": 2
},
"4096": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 16,
"num_warps": 4,
"num_stages": 2
}
}

146
python/sglang/srt/layers/moe/fused_moe_triton/configs/E=256,N=64,device_name=NVIDIA_L40S,dtype=int8_w8a8.json Normal file
View File

@@ -0,0 +1,146 @@
{
"1": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 64,
"num_warps": 4,
"num_stages": 3
},
"2": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 64,
"num_warps": 4,
"num_stages": 3
},
"4": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 64,
"num_warps": 4,
"num_stages": 3
},
"8": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 32,
"num_warps": 4,
"num_stages": 3
},
"16": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 16,
"num_warps": 4,
"num_stages": 2
},
"24": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 64,
"num_warps": 8,
"num_stages": 2
},
"32": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 32,
"num_warps": 8,
"num_stages": 2
},
"48": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2
},
"64": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 64,
"num_warps": 8,
"num_stages": 2
},
"96": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 2
},
"128": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 2
},
"256": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 2
},
"512": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 64,
"num_warps": 4,
"num_stages": 2
},
"1024": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 32,
"num_warps": 4,
"num_stages": 2
},
"1536": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 16,
"num_warps": 8,
"num_stages": 3
},
"2048": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 32,
"num_warps": 4,
"num_stages": 2
},
"3072": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 256,
"GROUP_SIZE_M": 16,
"num_warps": 8,
"num_stages": 2
},
"4096": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 16,
"num_warps": 8,
"num_stages": 4
}
}