Skip to content

GitLab

Commit ef27eec0 authored by 王敏's avatar 王敏
Browse files

[feat]添加BW deepseek_v3 fused_moe configs

parent 87fb7e7f
Show whitespace changes
Inline Side-by-side
Showing with 525 additions and 33 deletions
vllm/model_executor/layers/fused_moe/configs/E=16,N=1024,device_name=BW200_nn.json 0 → 100644
View file @ ef27eec0
{
"1": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"2": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"4": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"8": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"16": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"24": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"32": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"48": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 2,
"num_ldmatrixes": 1
},
"64": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"96": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"128": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"256": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"512": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 2,
"num_ldmatrixes": 1
},
"1024": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 16,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"1536": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 16,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"2048": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 16,
"num_warps": 8,
"num_stages": 2,
"num_ldmatrixes": 1
},
"3072": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"4096": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
}
}
vllm/model_executor/layers/fused_moe/configs/E=256,N=64,device_name=BW200_nn.json
View file @ ef27eec0
{ {
"1": { "1": {
"BLOCK_SIZE_M": 32, "BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128, "BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 64, "BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 16, "GROUP_SIZE_M": 1,
"num_warps": 4, "num_warps": 4,
"num_stages": 2, "num_stages": 2,
"num_ldmatrixes": 1 "num_ldmatrixes": 1
...@@ -30,7 +30,7 @@ ...@@ -30,7 +30,7 @@
"BLOCK_SIZE_M": 16, "BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128, "BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64, "BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 64, "GROUP_SIZE_M": 1,
"num_warps": 4, "num_warps": 4,
"num_stages": 2, "num_stages": 2,
"num_ldmatrixes": 1 "num_ldmatrixes": 1
...@@ -39,7 +39,7 @@ ...@@ -39,7 +39,7 @@
"BLOCK_SIZE_M": 16, "BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128, "BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64, "BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 32, "GROUP_SIZE_M": 1,
"num_warps": 4, "num_warps": 4,
"num_stages": 2, "num_stages": 2,
"num_ldmatrixes": 1 "num_ldmatrixes": 1
...@@ -48,7 +48,7 @@ ...@@ -48,7 +48,7 @@
"BLOCK_SIZE_M": 16, "BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128, "BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64, "BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 32, "GROUP_SIZE_M": 1,
"num_warps": 4, "num_warps": 4,
"num_stages": 2, "num_stages": 2,
"num_ldmatrixes": 1 "num_ldmatrixes": 1
...@@ -56,17 +56,17 @@ ...@@ -56,17 +56,17 @@
"32": { "32": {
"BLOCK_SIZE_M": 16, "BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128, "BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64, "BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 16, "GROUP_SIZE_M": 1,
"num_warps": 4, "num_warps": 4,
"num_stages": 2, "num_stages": 2,
"num_ldmatrixes": 1 "num_ldmatrixes": 1
}, },
"48": { "48": {
"BLOCK_SIZE_M": 16, "BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128, "BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64, "BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 64, "GROUP_SIZE_M": 1,
"num_warps": 4, "num_warps": 4,
"num_stages": 2, "num_stages": 2,
"num_ldmatrixes": 1 "num_ldmatrixes": 1
...@@ -74,7 +74,7 @@ ...@@ -74,7 +74,7 @@
"64": { "64": {
"BLOCK_SIZE_M": 16, "BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128, "BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64, "BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 1, "GROUP_SIZE_M": 1,
"num_warps": 4, "num_warps": 4,
"num_stages": 2, "num_stages": 2,
...@@ -83,8 +83,8 @@ ...@@ -83,8 +83,8 @@
"96": { "96": {
"BLOCK_SIZE_M": 16, "BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128, "BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64, "BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 64, "GROUP_SIZE_M": 1,
"num_warps": 4, "num_warps": 4,
"num_stages": 2, "num_stages": 2,
"num_ldmatrixes": 1 "num_ldmatrixes": 1
...@@ -92,17 +92,17 @@ ...@@ -92,17 +92,17 @@
"128": { "128": {
"BLOCK_SIZE_M": 16, "BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128, "BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64, "BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 32, "GROUP_SIZE_M": 1,
"num_warps": 4, "num_warps": 4,
"num_stages": 2, "num_stages": 2,
"num_ldmatrixes": 1 "num_ldmatrixes": 1
}, },
"256": { "256": {
"BLOCK_SIZE_M": 16, "BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128, "BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64, "BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 64, "GROUP_SIZE_M": 1,
"num_warps": 4, "num_warps": 4,
"num_stages": 2, "num_stages": 2,
"num_ldmatrixes": 1 "num_ldmatrixes": 1
...@@ -110,34 +110,34 @@ ...@@ -110,34 +110,34 @@
"512": { "512": {
"BLOCK_SIZE_M": 32, "BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128, "BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64, "BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 1, "GROUP_SIZE_M": 1,
"num_warps": 8, "num_warps": 4,
"num_stages": 2, "num_stages": 2,
"num_ldmatrixes": 1 "num_ldmatrixes": 1
}, },
"1024": { "1024": {
"BLOCK_SIZE_M": 64, "BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128, "BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64, "BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 1, "GROUP_SIZE_M": 1,
"num_warps": 8, "num_warps": 4,
"num_stages": 2, "num_stages": 2,
"num_ldmatrixes": 1 "num_ldmatrixes": 1
}, },
"1536": { "1536": {
"BLOCK_SIZE_M": 64, "BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128, "BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64, "BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 1, "GROUP_SIZE_M": 1,
"num_warps": 8, "num_warps": 4,
"num_stages": 2, "num_stages": 2,
"num_ldmatrixes": 1 "num_ldmatrixes": 1
}, },
"2048": { "2048": {
"BLOCK_SIZE_M": 32, "BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128, "BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64, "BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 1, "GROUP_SIZE_M": 1,
"num_warps": 4, "num_warps": 4,
"num_stages": 2, "num_stages": 2,
...@@ -145,8 +145,8 @@ ...@@ -145,8 +145,8 @@
}, },
"3072": { "3072": {
"BLOCK_SIZE_M": 128, "BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128, "BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64, "BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 1, "GROUP_SIZE_M": 1,
"num_warps": 4, "num_warps": 4,
"num_stages": 2, "num_stages": 2,
...@@ -154,10 +154,10 @@ ...@@ -154,10 +154,10 @@
}, },
"4096": { "4096": {
"BLOCK_SIZE_M": 64, "BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128, "BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64, "BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 1, "GROUP_SIZE_M": 1,
"num_warps": 4, "num_warps": 8,
"num_stages": 2, "num_stages": 2,
"num_ldmatrixes": 1 "num_ldmatrixes": 1
} }
......
vllm/model_executor/layers/fused_moe/configs/E=32,N=512,device_name=BW200_nn.json 0 → 100644
View file @ ef27eec0
{
"1": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 64,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"2": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"4": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"8": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"16": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 2,
"num_ldmatrixes": 1
},
"24": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 2,
"num_ldmatrixes": 1
},
"32": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 2,
"num_ldmatrixes": 1
},
"48": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"64": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 2,
"num_ldmatrixes": 1
},
"96": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 2,
"num_ldmatrixes": 1
},
"128": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 2,
"num_ldmatrixes": 1
},
"256": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"512": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"1024": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"1536": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"2048": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"3072": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"4096": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
}
}
vllm/model_executor/layers/fused_moe/configs/E=64,N=256,device_name=BW200_nn.json 0 → 100644
View file @ ef27eec0
{
"1": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 32,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"2": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"4": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"8": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"16": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 128,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 2,
"num_ldmatrixes": 1
},
"24": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 2,
"num_ldmatrixes": 1
},
"32": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 2,
"num_ldmatrixes": 1
},
"48": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 2,
"num_ldmatrixes": 1
},
"64": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 2,
"num_ldmatrixes": 1
},
"96": {
"BLOCK_SIZE_M": 16,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 1,
"num_warps": 8,
"num_stages": 2,
"num_ldmatrixes": 1
},
"128": {
"BLOCK_SIZE_M": 32,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"256": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"512": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"1024": {
"BLOCK_SIZE_M": 64,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"1536": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 128,
"BLOCK_SIZE_K": 64,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"2048": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"3072": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
},
"4096": {
"BLOCK_SIZE_M": 128,
"BLOCK_SIZE_N": 256,
"BLOCK_SIZE_K": 32,
"GROUP_SIZE_M": 1,
"num_warps": 4,
"num_stages": 2,
"num_ldmatrixes": 1
}
}
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