feat: Add Triton fallback option and SM120 MoE configs for FP8 models (#9251)

7cd2ee06 · Martin Vit · GitHub · eb19ccad · 7cd2ee06 · 7cd2ee06
Unverified Commit 7cd2ee06 authored Aug 21, 2025 by Martin Vit Committed by GitHub Aug 20, 2025
3 changed files
--- a/python/sglang/srt/layers/moe/fused_moe_triton/configs/triton_3_4_0/E=129,N=352,device_name=NVIDIA_RTX_PRO_6000_Blackwell_Max-Q_Workstation_Edition,dtype=fp8_w8a8.json
+++ b/python/sglang/srt/layers/moe/fused_moe_triton/configs/triton_3_4_0/E=129,N=352,device_name=NVIDIA_RTX_PRO_6000_Blackwell_Max-Q_Workstation_Edition,dtype=fp8_w8a8.json
+{
+    "1": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2
+    },
+    "2": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 2
+    },
+    "4": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 2
+    },
+    "8": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "16": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 8,
+        "num_stages": 2
+    },
+    "24": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "32": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "48": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 8,
+        "num_stages": 3
+    },
+    "64": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 8,
+        "num_stages": 2
+    },
+    "96": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "128": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2
+    },
+    "256": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 2
+    },
+    "512": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 2
+    },
+    "1024": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 2
+    },
+    "1536": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 2
+    },
+    "2048": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "3072": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 256,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 8,
+        "num_stages": 3
+    },
+    "4096": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 256,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 8,
+        "num_stages": 3
+    }
+}
--- a/python/sglang/srt/layers/moe/fused_moe_triton/configs/triton_3_4_0/E=161,N=384,device_name=NVIDIA_RTX_PRO_6000_Blackwell_Max-Q_Workstation_Edition,dtype=fp8_w8a8.json
+++ b/python/sglang/srt/layers/moe/fused_moe_triton/configs/triton_3_4_0/E=161,N=384,device_name=NVIDIA_RTX_PRO_6000_Blackwell_Max-Q_Workstation_Edition,dtype=fp8_w8a8.json
+{
+    "1": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2
+    },
+    "2": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 2
+    },
+    "4": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 2
+    },
+    "8": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "16": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 8,
+        "num_stages": 2
+    },
+    "24": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "32": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 4
+    },
+    "48": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 8,
+        "num_stages": 3
+    },
+    "64": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 8,
+        "num_stages": 2
+    },
+    "96": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "128": {
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 4,
+        "num_stages": 2
+    },
+    "256": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 2
+    },
+    "512": {
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 64,
+        "num_warps": 4,
+        "num_stages": 2
+    },
+    "1024": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 2
+    },
+    "1536": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 16,
+        "num_warps": 4,
+        "num_stages": 2
+    },
+    "2048": {
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 4,
+        "num_stages": 3
+    },
+    "3072": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 256,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
+        "num_warps": 8,
+        "num_stages": 3
+    },
+    "4096": {
+        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_N": 256,
+        "BLOCK_SIZE_K": 64,
+        "GROUP_SIZE_M": 32,
+        "num_warps": 8,
+        "num_stages": 3
+    }
+}
--- a/python/sglang/srt/layers/quantization/fp8_utils.py
+++ b/python/sglang/srt/layers/quantization/fp8_utils.py
@@ -53,6 +53,7 @@ if _is_cuda:
    from sgl_kernel import fp8_blockwise_scaled_mm, fp8_scaled_mm

 use_vllm_cutlass_w8a8_fp8_kernel = get_bool_env_var("USE_VLLM_CUTLASS_W8A8_FP8_KERNEL")
+use_triton_w8a8_fp8_kernel = get_bool_env_var("USE_TRITON_W8A8_FP8_KERNEL")

 # Input scaling factors are no longer optional in _scaled_mm starting
 # from pytorch 2.5. Allocating a dummy tensor to pass as input_scale
@@ -592,7 +593,7 @@ def apply_fp8_linear(
                cutlass_compatible_b = (
                    weight.shape[0] % 16 == 0 and weight.shape[1] % 16 == 0
                )
-                if not cutlass_compatible_b:
+                if not cutlass_compatible_b or use_triton_w8a8_fp8_kernel:
                    # Massage the input to be 2D
                    qinput = qinput.view(-1, qinput.shape[-1])
                    output = triton_scaled_mm(
@@ -735,14 +736,25 @@ def apply_fp8_linear(
                    assert (
                        weight_scale.numel() == weight.shape[1]
                    ), "cutlass w8a8 fp8 sgl-kernel only supports per-channel scale"
-                    output = fp8_scaled_mm(
-                        qinput,
-                        weight,
-                        x_scale,
-                        weight_scale,
-                        out_dtype=input.dtype,
-                        bias=bias,
+
+                    cutlass_compatible_b = (
+                        weight.shape[0] % 16 == 0 and weight.shape[1] % 16 == 0
                    )
+                    if not cutlass_compatible_b or use_triton_w8a8_fp8_kernel:
+                        # Massage the input to be 2D
+                        qinput = qinput.view(-1, qinput.shape[-1])
+                        output = triton_scaled_mm(
+                            qinput, weight, x_scale, weight_scale, input.dtype, bias
+                        )
+                    else:
+                        output = fp8_scaled_mm(
+                            qinput,
+                            weight,
+                            x_scale,
+                            weight_scale,
+                            out_dtype=input.dtype,
+                            bias=bias,
+                        )
                return output.view(*output_shape)
            except (ImportError, NameError, AttributeError):
                pass