Add GB10 FP8 fused MoE Triton config (#25665)

python/sglang/srt/layers/moe/moe_runner/triton_utils/configs/triton_3_6_0/E=128,N=768,device_name=NVIDIA_GB10,dtype=fp8_w8a8.json

@@ -0,0 +1,146 @@
+{
+  "1": {
+    "BLOCK_SIZE_M": 32,
+    "BLOCK_SIZE_N": 32,
+    "BLOCK_SIZE_K": 256,
+    "GROUP_SIZE_M": 1,
+    "num_warps": 8,
+    "num_stages": 5
+  },
+  "2": {
+    "BLOCK_SIZE_M": 16,
+    "BLOCK_SIZE_N": 128,
+    "BLOCK_SIZE_K": 256,
+    "GROUP_SIZE_M": 64,
+    "num_warps": 8,
+    "num_stages": 2
+  },
+  "4": {
+    "BLOCK_SIZE_M": 32,
+    "BLOCK_SIZE_N": 32,
+    "BLOCK_SIZE_K": 256,
+    "GROUP_SIZE_M": 32,
+    "num_warps": 8,
+    "num_stages": 4
+  },
+  "8": {
+    "BLOCK_SIZE_M": 32,
+    "BLOCK_SIZE_N": 64,
+    "BLOCK_SIZE_K": 256,
+    "GROUP_SIZE_M": 32,
+    "num_warps": 8,
+    "num_stages": 4
+  },
+  "16": {
+    "BLOCK_SIZE_M": 32,
+    "BLOCK_SIZE_N": 64,
+    "BLOCK_SIZE_K": 256,
+    "GROUP_SIZE_M": 32,
+    "num_warps": 8,
+    "num_stages": 4
+  },
+  "24": {
+    "BLOCK_SIZE_M": 16,
+    "BLOCK_SIZE_N": 64,
+    "BLOCK_SIZE_K": 256,
+    "GROUP_SIZE_M": 1,
+    "num_warps": 4,
+    "num_stages": 3
+  },
+  "32": {
+    "BLOCK_SIZE_M": 16,
+    "BLOCK_SIZE_N": 64,
+    "BLOCK_SIZE_K": 256,
+    "GROUP_SIZE_M": 64,
+    "num_warps": 4,
+    "num_stages": 5
+  },
+  "48": {
+    "BLOCK_SIZE_M": 16,
+    "BLOCK_SIZE_N": 64,
+    "BLOCK_SIZE_K": 256,
+    "GROUP_SIZE_M": 64,
+    "num_warps": 4,
+    "num_stages": 5
+  },
+  "64": {
+    "BLOCK_SIZE_M": 16,
+    "BLOCK_SIZE_N": 256,
+    "BLOCK_SIZE_K": 256,
+    "GROUP_SIZE_M": 1,
+    "num_warps": 8,
+    "num_stages": 2
+  },
+  "96": {
+    "BLOCK_SIZE_M": 16,
+    "BLOCK_SIZE_N": 32,
+    "BLOCK_SIZE_K": 256,
+    "GROUP_SIZE_M": 1,
+    "num_warps": 4,
+    "num_stages": 5
+  },
+  "128": {
+    "BLOCK_SIZE_M": 32,
+    "BLOCK_SIZE_N": 64,
+    "BLOCK_SIZE_K": 256,
+    "GROUP_SIZE_M": 1,
+    "num_warps": 4,
+    "num_stages": 4
+  },
+  "256": {
+    "BLOCK_SIZE_M": 64,
+    "BLOCK_SIZE_N": 32,
+    "BLOCK_SIZE_K": 256,
+    "GROUP_SIZE_M": 1,
+    "num_warps": 8,
+    "num_stages": 3
+  },
+  "512": {
+    "BLOCK_SIZE_M": 16,
+    "BLOCK_SIZE_N": 64,
+    "BLOCK_SIZE_K": 256,
+    "GROUP_SIZE_M": 64,
+    "num_warps": 4,
+    "num_stages": 4
+  },
+  "1024": {
+    "BLOCK_SIZE_M": 128,
+    "BLOCK_SIZE_N": 128,
+    "BLOCK_SIZE_K": 256,
+    "GROUP_SIZE_M": 32,
+    "num_warps": 8,
+    "num_stages": 2
+  },
+  "1536": {
+    "BLOCK_SIZE_M": 64,
+    "BLOCK_SIZE_N": 128,
+    "BLOCK_SIZE_K": 256,
+    "GROUP_SIZE_M": 64,
+    "num_warps": 4,
+    "num_stages": 3
+  },
+  "2048": {
+    "BLOCK_SIZE_M": 64,
+    "BLOCK_SIZE_N": 128,
+    "BLOCK_SIZE_K": 256,
+    "GROUP_SIZE_M": 64,
+    "num_warps": 4,
+    "num_stages": 2
+  },
+  "3072": {
+    "BLOCK_SIZE_M": 128,
+    "BLOCK_SIZE_N": 64,
+    "BLOCK_SIZE_K": 128,
+    "GROUP_SIZE_M": 64,
+    "num_warps": 4,
+    "num_stages": 2
+  },
+  "4096": {
+    "BLOCK_SIZE_M": 128,
+    "BLOCK_SIZE_N": 64,
+    "BLOCK_SIZE_K": 128,
+    "GROUP_SIZE_M": 64,
+    "num_warps": 4,
+    "num_stages": 3
+  }
+}

python/sglang/srt/layers/moe/moe_runner/triton_utils/fused_moe_triton_config.py

@@ -14,6 +14,23 @@ from sglang.srt.utils import get_device_name, is_hip
 
 logger = logging.getLogger(__name__)
 _is_hip = is_hip()
+_LOW_SMEM_FP8_DEFAULT_CUTOFF_BYTES = 128 * 1024
+
+
+@functools.lru_cache(maxsize=None)
+def _get_cuda_shared_memory_per_block_optin() -> Optional[int]:
+    if _is_hip or not torch.cuda.is_available():
+        return None
+    try:
+        props = torch.cuda.get_device_properties(torch.cuda.current_device())
+    except (AssertionError, RuntimeError):
+        return None
+    return getattr(props, "shared_memory_per_block_optin", None)
+
+
+def _use_low_smem_fp8_default() -> bool:
+    smem_limit = _get_cuda_shared_memory_per_block_optin()
+    return smem_limit is not None and smem_limit < _LOW_SMEM_FP8_DEFAULT_CUTOFF_BYTES
 
 
 def get_config_file_name(
@@ -163,6 +180,25 @@ def get_default_config(
         return config
     if dtype == "fp8_w8a8":
         if block_shape is None:
+            if _use_low_smem_fp8_default():
+                config = {
+                    "BLOCK_SIZE_M": 32,
+                    "BLOCK_SIZE_N": 64,
+                    "BLOCK_SIZE_K": 256,
+                    "GROUP_SIZE_M": 1,
+                    "num_warps": 4,
+                    "num_stages": 4,
+                }
+                if M > E:
+                    config = {
+                        "BLOCK_SIZE_M": 64,
+                        "BLOCK_SIZE_N": 128,
+                        "BLOCK_SIZE_K": 256,
+                        "GROUP_SIZE_M": 64,
+                        "num_warps": 4,
+                        "num_stages": 2,
+                    }
+            else:
                 config = {
                     "BLOCK_SIZE_M": 128,
                     "BLOCK_SIZE_N": 256,