From f0a5f111f205b5c73bb34fbd41f8f0d9936b543f Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Pawe=C5=82=20Ma=C5=84czak?= <pawel.manczak@intel.com>
Date: Fri, 9 Oct 2026 08:29:29 +0200
Subject: [PATCH] [XPU] Add tuned Mamba SSU configs for Intel Arc Pro B60
 (#56765)

Signed-off-by: pmanczak <pawel.manczak@intel.com>
Co-authored-by: Kunshang Ji <kunshang.ji@intel.com>
---
 ..._Pro_B60_Graphics,cache_dtype=float16.json | 51 +++++++++++++++
 ..._Pro_B60_Graphics,cache_dtype=float32.json | 51 +++++++++++++++
 ..._Pro_B60_Graphics,cache_dtype=float16.json | 63 +++++++++++++++++++
 ..._Pro_B60_Graphics,cache_dtype=float32.json | 63 +++++++++++++++++++
 ..._Pro_B60_Graphics,cache_dtype=float16.json | 63 +++++++++++++++++++
 ..._Pro_B60_Graphics,cache_dtype=float32.json | 63 +++++++++++++++++++
 ..._Pro_B60_Graphics,cache_dtype=float16.json | 63 +++++++++++++++++++
 ..._Pro_B60_Graphics,cache_dtype=float32.json | 63 +++++++++++++++++++
 8 files changed, 480 insertions(+)
 create mode 100644 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
 create mode 100644 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
 create mode 100644 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
 create mode 100644 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
 create mode 100644 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
 create mode 100644 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
 create mode 100644 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
 create mode 100644 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

diff --git a/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 b/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
new file mode 100644
index 0000000000..64087ce74c
--- /dev/null
+++ b/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
@@ -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
+    }
+}
diff --git a/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 b/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
new file mode 100644
index 0000000000..0a378592c5
--- /dev/null
+++ b/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
@@ -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
+    }
+}
diff --git a/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 b/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
new file mode 100644
index 0000000000..91c6e28fa7
--- /dev/null
+++ b/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
@@ -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
+    }
+}
diff --git a/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 b/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
new file mode 100644
index 0000000000..c30784ce77
--- /dev/null
+++ b/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
@@ -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
+    }
+}
diff --git a/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 b/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
new file mode 100644
index 0000000000..5f29bcd76c
--- /dev/null
+++ b/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
@@ -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
+    }
+}
diff --git a/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 b/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
new file mode 100644
index 0000000000..8f815a1b33
--- /dev/null
+++ b/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
@@ -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
+    }
+}
diff --git a/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 b/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
new file mode 100644
index 0000000000..2df2525541
--- /dev/null
+++ b/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
@@ -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
+    }
+}
diff --git a/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 b/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
new file mode 100644
index 0000000000..f943b523c4
--- /dev/null
+++ b/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
@@ -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
+    }
+}