iHore/vLLM
8
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[XPU] Add tuned Mamba SSU configs for Intel Arc Pro B60 (#56765)

Browse Source 
Signed-off-by: pmanczak <pawel.manczak@intel.com>
Co-authored-by: Kunshang Ji <kunshang.ji@intel.com>
This commit is contained in:
Paweł Mańczak
2026-10-09 14:29:29 +08:00
committed by GitHub
co-authored by Kunshang Ji
parent ad73a4740d
commit f0a5f111f2
8 changed files with 480 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files
vllm/model_executor/layers/mamba/ops/configs/selective_state_update/headdim=128,dstate=256,device_name=Intel(R)_Arc(TM)_Pro_B60_Graphics,cache_dtype=float16.json
+51
View File
@@ -0,0 +1,51 @@
{
"triton_version": "3.8.0",
"8": {
"BLOCK_SIZE_M": 4,
"num_warps": 1
},
"16": {
"BLOCK_SIZE_M": 4,
"num_warps": 1
},
"32": {
"BLOCK_SIZE_M": 4,
"num_warps": 1
},
"64": {
"BLOCK_SIZE_M": 4,
"num_warps": 1
},
"128": {
"BLOCK_SIZE_M": 4,
"num_warps": 1
},
"256": {
"BLOCK_SIZE_M": 4,
"num_warps": 1
},
"512": {
"BLOCK_SIZE_M": 4,
"num_warps": 1
},
"1024": {
"BLOCK_SIZE_M": 4,
"num_warps": 1
},
"2048": {
"BLOCK_SIZE_M": 4,
"num_warps": 1
},
"4096": {
"BLOCK_SIZE_M": 4,
"num_warps": 1
},
"8192": {
"BLOCK_SIZE_M": 4,
"num_warps": 1
},
"16384": {
"BLOCK_SIZE_M": 4,
"num_warps": 1
}
}
vllm/model_executor/layers/mamba/ops/configs/selective_state_update/headdim=128,dstate=256,device_name=Intel(R)_Arc(TM)_Pro_B60_Graphics,cache_dtype=float32.json
+51
View File
@@ -0,0 +1,51 @@
{
"triton_version": "3.8.0",
"8": {
"BLOCK_SIZE_M": 4,
"num_warps": 2
},
"16": {
"BLOCK_SIZE_M": 4,
"num_warps": 2
},
"32": {
"BLOCK_SIZE_M": 8,
"num_warps": 2
},
"64": {
"BLOCK_SIZE_M": 4,
"num_warps": 2
},
"128": {
"BLOCK_SIZE_M": 4,
"num_warps": 2
},
"256": {
"BLOCK_SIZE_M": 4,
"num_warps": 1
},
"512": {
"BLOCK_SIZE_M": 8,
"num_warps": 2
},
"1024": {
"BLOCK_SIZE_M": 8,
"num_warps": 2
},
"2048": {
"BLOCK_SIZE_M": 4,
"num_warps": 1
},
"4096": {
"BLOCK_SIZE_M": 4,
"num_warps": 1
},
"8192": {
"BLOCK_SIZE_M": 4,
"num_warps": 1
},
"16384": {
"BLOCK_SIZE_M": 4,
"num_warps": 1
}
}
vllm/model_executor/layers/mamba/ops/configs/selective_state_update/headdim=64,dstate=128,device_name=Intel(R)_Arc(TM)_Pro_B60_Graphics,cache_dtype=float16.json
+63
View File
@@ -0,0 +1,63 @@
{
"triton_version": "3.8.0",
"8": {
"BLOCK_SIZE_M": 16,
"num_warps": 8
},
"16": {
"BLOCK_SIZE_M": 64,
"num_warps": 8
},
"32": {
"BLOCK_SIZE_M": 64,
"num_warps": 8
},
"64": {
"BLOCK_SIZE_M": 64,
"num_warps": 8
},
"128": {
"BLOCK_SIZE_M": 64,
"num_warps": 8
},
"256": {
"BLOCK_SIZE_M": 64,
"num_warps": 8
},
"512": {
"BLOCK_SIZE_M": 4,
"num_warps": 1
},
"1024": {
"BLOCK_SIZE_M": 64,
"num_warps": 8
},
"2048": {
"BLOCK_SIZE_M": 4,
"num_warps": 1
},
"4096": {
"BLOCK_SIZE_M": 64,
"num_warps": 8
},
"8192": {
"BLOCK_SIZE_M": 64,
"num_warps": 8
},
"16384": {
"BLOCK_SIZE_M": 64,
"num_warps": 8
},
"32768": {
"BLOCK_SIZE_M": 64,
"num_warps": 8
},
"65536": {
"BLOCK_SIZE_M": 4,
"num_warps": 1
},
"131072": {
"BLOCK_SIZE_M": 4,
"num_warps": 1
}
}
vllm/model_executor/layers/mamba/ops/configs/selective_state_update/headdim=64,dstate=128,device_name=Intel(R)_Arc(TM)_Pro_B60_Graphics,cache_dtype=float32.json
+63
View File
@@ -0,0 +1,63 @@
{
"triton_version": "3.8.0",
"8": {
"BLOCK_SIZE_M": 8,
"num_warps": 8
},
"16": {
"BLOCK_SIZE_M": 8,
"num_warps": 1
},
"32": {
"BLOCK_SIZE_M": 4,
"num_warps": 1
},
"64": {
"BLOCK_SIZE_M": 4,
"num_warps": 1
},
"128": {
"BLOCK_SIZE_M": 4,
"num_warps": 1
},
"256": {
"BLOCK_SIZE_M": 4,
"num_warps": 1
},
"512": {
"BLOCK_SIZE_M": 4,
"num_warps": 1
},
"1024": {
"BLOCK_SIZE_M": 4,
"num_warps": 1
},
"2048": {
"BLOCK_SIZE_M": 4,
"num_warps": 1
},
"4096": {
"BLOCK_SIZE_M": 4,
"num_warps": 1
},
"8192": {
"BLOCK_SIZE_M": 8,
"num_warps": 1
},
"16384": {
"BLOCK_SIZE_M": 8,
"num_warps": 1
},
"32768": {
"BLOCK_SIZE_M": 8,
"num_warps": 1
},
"65536": {
"BLOCK_SIZE_M": 8,
"num_warps": 1
},
"131072": {
"BLOCK_SIZE_M": 8,
"num_warps": 1
}
}
vllm/model_executor/layers/mamba/ops/configs/selective_state_update/headdim=64,dstate=256,device_name=Intel(R)_Arc(TM)_Pro_B60_Graphics,cache_dtype=float16.json
+63
View File
@@ -0,0 +1,63 @@
{
"triton_version": "3.8.0",
"8": {
"BLOCK_SIZE_M": 4,
"num_warps": 1
},
"16": {
"BLOCK_SIZE_M": 4,
"num_warps": 1
},
"32": {
"BLOCK_SIZE_M": 4,
"num_warps": 1
},
"64": {
"BLOCK_SIZE_M": 4,
"num_warps": 1
},
"128": {
"BLOCK_SIZE_M": 4,
"num_warps": 1
},
"256": {
"BLOCK_SIZE_M": 4,
"num_warps": 1
},
"512": {
"BLOCK_SIZE_M": 4,
"num_warps": 1
},
"1024": {
"BLOCK_SIZE_M": 4,
"num_warps": 1
},
"2048": {
"BLOCK_SIZE_M": 4,
"num_warps": 1
},
"4096": {
"BLOCK_SIZE_M": 4,
"num_warps": 1
},
"8192": {
"BLOCK_SIZE_M": 4,
"num_warps": 1
},
"16384": {
"BLOCK_SIZE_M": 4,
"num_warps": 1
},
"32768": {
"BLOCK_SIZE_M": 4,
"num_warps": 1
},
"65536": {
"BLOCK_SIZE_M": 4,
"num_warps": 1
},
"131072": {
"BLOCK_SIZE_M": 4,
"num_warps": 1
}
}
vllm/model_executor/layers/mamba/ops/configs/selective_state_update/headdim=64,dstate=256,device_name=Intel(R)_Arc(TM)_Pro_B60_Graphics,cache_dtype=float32.json
+63
View File
@@ -0,0 +1,63 @@
{
"triton_version": "3.8.0",
"8": {
"BLOCK_SIZE_M": 4,
"num_warps": 2
},
"16": {
"BLOCK_SIZE_M": 4,
"num_warps": 2
},
"32": {
"BLOCK_SIZE_M": 8,
"num_warps": 2
},
"64": {
"BLOCK_SIZE_M": 8,
"num_warps": 2
},
"128": {
"BLOCK_SIZE_M": 4,
"num_warps": 2
},
"256": {
"BLOCK_SIZE_M": 8,
"num_warps": 2
},
"512": {
"BLOCK_SIZE_M": 8,
"num_warps": 2
},
"1024": {
"BLOCK_SIZE_M": 8,
"num_warps": 2
},
"2048": {
"BLOCK_SIZE_M": 4,
"num_warps": 1
},
"4096": {
"BLOCK_SIZE_M": 8,
"num_warps": 2
},
"8192": {
"BLOCK_SIZE_M": 8,
"num_warps": 2
},
"16384": {
"BLOCK_SIZE_M": 8,
"num_warps": 2
},
"32768": {
"BLOCK_SIZE_M": 4,
"num_warps": 1
},
"65536": {
"BLOCK_SIZE_M": 4,
"num_warps": 1
},
"131072": {
"BLOCK_SIZE_M": 4,
"num_warps": 1
}
}
vllm/model_executor/layers/mamba/ops/configs/selective_state_update/headdim=64,dstate=64,device_name=Intel(R)_Arc(TM)_Pro_B60_Graphics,cache_dtype=float16.json
+63
View File
@@ -0,0 +1,63 @@
{
"triton_version": "3.8.0",
"8": {
"BLOCK_SIZE_M": 16,
"num_warps": 4
},
"16": {
"BLOCK_SIZE_M": 16,
"num_warps": 4
},
"32": {
"BLOCK_SIZE_M": 16,
"num_warps": 4
},
"64": {
"BLOCK_SIZE_M": 4,
"num_warps": 1
},
"128": {
"BLOCK_SIZE_M": 64,
"num_warps": 8
},
"256": {
"BLOCK_SIZE_M": 8,
"num_warps": 1
},
"512": {
"BLOCK_SIZE_M": 8,
"num_warps": 1
},
"1024": {
"BLOCK_SIZE_M": 8,
"num_warps": 1
},
"2048": {
"BLOCK_SIZE_M": 8,
"num_warps": 1
},
"4096": {
"BLOCK_SIZE_M": 4,
"num_warps": 1
},
"8192": {
"BLOCK_SIZE_M": 8,
"num_warps": 1
},
"16384": {
"BLOCK_SIZE_M": 8,
"num_warps": 1
},
"32768": {
"BLOCK_SIZE_M": 8,
"num_warps": 1
},
"65536": {
"BLOCK_SIZE_M": 8,
"num_warps": 1
},
"131072": {
"BLOCK_SIZE_M": 8,
"num_warps": 1
}
}
vllm/model_executor/layers/mamba/ops/configs/selective_state_update/headdim=64,dstate=64,device_name=Intel(R)_Arc(TM)_Pro_B60_Graphics,cache_dtype=float32.json
+63
View File
@@ -0,0 +1,63 @@
{
"triton_version": "3.8.0",
"8": {
"BLOCK_SIZE_M": 4,
"num_warps": 1
},
"16": {
"BLOCK_SIZE_M": 16,
"num_warps": 8
},
"32": {
"BLOCK_SIZE_M": 4,
"num_warps": 1
},
"64": {
"BLOCK_SIZE_M": 4,
"num_warps": 1
},
"128": {
"BLOCK_SIZE_M": 4,
"num_warps": 1
},
"256": {
"BLOCK_SIZE_M": 4,
"num_warps": 1
},
"512": {
"BLOCK_SIZE_M": 4,
"num_warps": 1
},
"1024": {
"BLOCK_SIZE_M": 4,
"num_warps": 1
},
"2048": {
"BLOCK_SIZE_M": 4,
"num_warps": 1
},
"4096": {
"BLOCK_SIZE_M": 4,
"num_warps": 1
},
"8192": {
"BLOCK_SIZE_M": 4,
"num_warps": 1
},
"16384": {
"BLOCK_SIZE_M": 4,
"num_warps": 1
},
"32768": {
"BLOCK_SIZE_M": 4,
"num_warps": 1
},
"65536": {
"BLOCK_SIZE_M": 4,
"num_warps": 1
},
"131072": {
"BLOCK_SIZE_M": 4,
"num_warps": 1
}
}