Skip to content

GitLab

"vscode:/vscode.git/clone" did not exist on "addf20e48b7f712761d467e7413fdace517f9e3f"
Unverified Commit afa0341e authored by HandH1998's avatar HandH1998 Committed by GitHub
Browse files

Update Triton configs for block fp8 kernels (#2641)

parent 30828e71
Hide whitespace changes
Inline Side-by-side
Showing with 187 additions and 42 deletions
python/sglang/srt/layers/quantization/configs/N=7168,K=2304,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json
View file @ afa0341e
{ {
"1": { "1": {
"BLOCK_SIZE_M": 16, "BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32, "BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128, "BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 32, "GROUP_SIZE_M": 1,
"num_warps": 4, "num_warps": 4,
"num_stages": 2 "num_stages": 3
}, },
"2": { "2": {
"BLOCK_SIZE_M": 64, "BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64, "BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128, "BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 16, "GROUP_SIZE_M": 32,
"num_warps": 4, "num_warps": 4,
"num_stages": 2 "num_stages": 4
}, },
"4": { "4": {
"BLOCK_SIZE_M": 64, "BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64, "BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128, "BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1, "GROUP_SIZE_M": 32,
"num_warps": 4, "num_warps": 4,
"num_stages": 2 "num_stages": 4
}, },
"8": { "8": {
"BLOCK_SIZE_M": 64, "BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64, "BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128, "BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 16, "GROUP_SIZE_M": 32,
"num_warps": 4, "num_warps": 4,
"num_stages": 2 "num_stages": 4
}, },
"16": { "16": {
"BLOCK_SIZE_M": 64, "BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 32, "BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 128, "BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 16, "GROUP_SIZE_M": 64,
"num_warps": 4, "num_warps": 4,
"num_stages": 2 "num_stages": 4
}, },
"24": { "24": {
"BLOCK_SIZE_M": 64, "BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 32, "BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 128, "BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1, "GROUP_SIZE_M": 64,
"num_warps": 4, "num_warps": 4,
"num_stages": 2 "num_stages": 4
}, },
"32": { "32": {
"BLOCK_SIZE_M": 64, "BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 32, "BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 128, "BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 32, "GROUP_SIZE_M": 1,
"num_warps": 4, "num_warps": 4,
"num_stages": 2 "num_stages": 4
}, },
"48": { "48": {
"BLOCK_SIZE_M": 64, "BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 32, "BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 128, "BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1, "GROUP_SIZE_M": 32,
"num_warps": 4, "num_warps": 4,
"num_stages": 2 "num_stages": 4
}, },
"64": { "64": {
"BLOCK_SIZE_M": 64, "BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 32, "BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 128, "BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 64, "GROUP_SIZE_M": 1,
"num_warps": 4, "num_warps": 4,
"num_stages": 2 "num_stages": 4
}, },
"96": { "96": {
"BLOCK_SIZE_M": 64, "BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 32, "BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128, "BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1, "GROUP_SIZE_M": 32,
"num_warps": 4, "num_warps": 4,
"num_stages": 2 "num_stages": 4
}, },
"128": { "128": {
"BLOCK_SIZE_M": 64, "BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 32, "BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128, "BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1, "GROUP_SIZE_M": 64,
"num_warps": 4, "num_warps": 4,
"num_stages": 2 "num_stages": 4
}, },
"256": { "256": {
"BLOCK_SIZE_M": 64, "BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128, "BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128, "BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1, "GROUP_SIZE_M": 16,
"num_warps": 4, "num_warps": 4,
"num_stages": 2 "num_stages": 3
}, },
"512": { "512": {
"BLOCK_SIZE_M": 64, "BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128, "BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128, "BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1, "GROUP_SIZE_M": 64,
"num_warps": 4, "num_warps": 4,
"num_stages": 2 "num_stages": 3
}, },
"1024": { "1024": {
"BLOCK_SIZE_M": 64, "BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64, "BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128, "BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1, "GROUP_SIZE_M": 16,
"num_warps": 4, "num_warps": 4,
"num_stages": 2 "num_stages": 3
}, },
"1536": { "1536": {
"BLOCK_SIZE_M": 64, "BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64, "BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128, "BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1, "GROUP_SIZE_M": 32,
"num_warps": 4, "num_warps": 4,
"num_stages": 2 "num_stages": 3
}, },
"2048": { "2048": {
"BLOCK_SIZE_M": 64, "BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128, "BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128, "BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1, "GROUP_SIZE_M": 64,
"num_warps": 4, "num_warps": 4,
"num_stages": 2 "num_stages": 3
}, },
"3072": { "3072": {
"BLOCK_SIZE_M": 64, "BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64, "BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128, "BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1, "GROUP_SIZE_M": 64,
"num_warps": 4, "num_warps": 4,
"num_stages": 2 "num_stages": 3
}, },
"4096": { "4096": {
"BLOCK_SIZE_M": 64, "BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128, "BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128, "BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 64, "GROUP_SIZE_M": 1,
"num_warps": 4, "num_warps": 4,
"num_stages": 2 "num_stages": 3
} }
} }
python/sglang/srt/layers/quantization/configs/N=7168,K=256,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json 0 → 100644
View file @ afa0341e
{
"1": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 16,
"num_warps": 4,
"num_stages": 3
},
"2": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 16,
"num_warps": 4,
"num_stages": 3
},
"4": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 3
},
"8": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 16,
"num_warps": 4,
"num_stages": 3
},
"16": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 3
},
"24": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 16,
"num_warps": 4,
"num_stages": 3
},
"32": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 3
},
"48": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 3
},
"64": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 32,
"num_warps": 4,
"num_stages": 4
},
"96": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 16,
"num_warps": 4,
"num_stages": 5
},
"128": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 32,
"num_warps": 4,
"num_stages": 5
},
"256": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 64,
"num_warps": 4,
"num_stages": 4
},
"512": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 16,
"num_warps": 4,
"num_stages": 3
},
"1024": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 3
},
"1536": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2
},
"2048": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2
},
"3072": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2
},
"4096": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2
}
}
python/sglang/srt/layers/quantization/fp8_kernel.py
View file @ afa0341e
...@@ -309,15 +309,14 @@ def w8a8_block_fp8_matmul( ...@@ -309,15 +309,14 @@ def w8a8_block_fp8_matmul(
config = configs[min(configs.keys(), key=lambda x: abs(x - M))] config = configs[min(configs.keys(), key=lambda x: abs(x - M))]
else: else:
# Default config # Default config
# Block-wise quant: BLOCK_SIZE_N must be divisable by block_size[0] # Block-wise quant: BLOCK_SIZE_K must be divisable by block_size[1]
# BLOCK_SIZE_K must be divisable by block_size[1]
config = { config = {
"BLOCK_SIZE_M": 64, "BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": block_size[0], "BLOCK_SIZE_N": block_size[0],
"BLOCK_SIZE_K": block_size[1], "BLOCK_SIZE_K": block_size[1],
"GROUP_SIZE_M": 32, "GROUP_SIZE_M": 32,
"num_warps": 4, "num_warps": 4,
"num_stages": 2, "num_stages": 3,
} }
def grid(META): def grid(META):
......
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment