Fuse routed scaling factor in deepseek (#6970)

benchmark/kernels/fused_moe_triton/benchmark_sum_scale.py

@@ -0,0 +1,199 @@
+import torch
+import triton
+import triton.language as tl
+from triton.testing import do_bench
+
+
+# _moe_sum_reduce_kernel kernel modified from https://github.com/ModelTC/lightllm/blob/main/lightllm/common/fused_moe/moe_sum_reduce.py
+@triton.jit
+def _moe_sum_reduce_kernel(
+    input_ptr,
+    input_stride_0,
+    input_stride_1,
+    input_stride_2,
+    output_ptr,
+    output_stride_0,
+    output_stride_1,
+    token_num: int,
+    topk_num: int,
+    hidden_dim: int,
+    routed_scaling_factor: tl.constexpr,
+    BLOCK_M: tl.constexpr,
+    BLOCK_DIM: tl.constexpr,
+    NUM_STAGE: tl.constexpr,
+):
+    input_stride_0 = tl.cast(input_stride_0, dtype=tl.int64)
+    input_stride_1 = tl.cast(input_stride_1, dtype=tl.int64)
+    output_stride_0 = tl.cast(output_stride_0, dtype=tl.int64)
+
+    token_block_id = tl.program_id(0)
+    dim_block_id = tl.program_id(1)
+
+    token_start = token_block_id * BLOCK_M
+    token_end = min((token_block_id + 1) * BLOCK_M, token_num)
+
+    dim_start = dim_block_id * BLOCK_DIM
+    dim_end = min((dim_block_id + 1) * BLOCK_DIM, hidden_dim)
+
+    offs_dim = dim_start + tl.arange(0, BLOCK_DIM)
+
+    for token_index in range(token_start, token_end):
+        accumulator = tl.zeros((BLOCK_DIM,), dtype=tl.float32)
+        input_t_ptr = input_ptr + token_index * input_stride_0 + offs_dim
+        for i in tl.range(0, topk_num, num_stages=NUM_STAGE):
+            tmp = tl.load(
+                input_t_ptr + i * input_stride_1, mask=offs_dim < dim_end, other=0.0
+            )
+            accumulator += tmp
+        accumulator = accumulator * routed_scaling_factor
+        store_t_ptr = output_ptr + token_index * output_stride_0 + offs_dim
+        tl.store(
+            store_t_ptr,
+            accumulator.to(input_ptr.dtype.element_ty),
+            mask=offs_dim < dim_end,
+        )
+
+
+def moe_sum_reduce(
+    input: torch.Tensor, output: torch.Tensor, routed_scaling_factor: float
+):
+    assert input.is_contiguous()
+    assert output.is_contiguous()
+
+    token_num, topk_num, hidden_dim = input.shape
+    assert output.shape[0] == token_num and output.shape[1] == hidden_dim
+
+    BLOCK_M = 1
+    BLOCK_DIM = 2048
+    NUM_STAGE = 1
+    num_warps = 8
+
+    grid = (
+        triton.cdiv(token_num, BLOCK_M),
+        triton.cdiv(hidden_dim, BLOCK_DIM),
+    )
+
+    _moe_sum_reduce_kernel[grid](
+        input,
+        *input.stride(),
+        output,
+        *output.stride(),
+        token_num=token_num,
+        topk_num=topk_num,
+        hidden_dim=hidden_dim,
+        routed_scaling_factor=routed_scaling_factor,
+        BLOCK_M=BLOCK_M,
+        BLOCK_DIM=BLOCK_DIM,
+        NUM_STAGE=NUM_STAGE,
+        num_warps=num_warps,
+    )
+    return
+
+
+def compute_sum_scaled_baseline(
+    x: torch.Tensor, out: torch.Tensor, routed_scaling_factor: float
+) -> torch.Tensor:
+    torch.sum(x, dim=1, out=out)
+    out.mul_(routed_scaling_factor)
+    return out
+
+
+@torch.compile
+def compute_sum_scaled_compiled(
+    x: torch.Tensor, out: torch.Tensor, routed_scaling_factor: float
+) -> torch.Tensor:
+    torch.sum(x * routed_scaling_factor, dim=1, out=out)
+    return out
+
+
+def get_benchmark():
+    num_tokens_range = [2**i for i in range(0, 13)]
+
+    @triton.testing.perf_report(
+        triton.testing.Benchmark(
+            x_names=["num_tokens"],
+            x_vals=num_tokens_range,
+            line_arg="version",
+            line_vals=["baseline", "compiled", "triton"],
+            line_names=["Original", "TorchCompile", "TritonKernel"],
+            styles=[("blue", "-"), ("green", "-"), ("red", "-")],
+            ylabel="us",
+            plot_name="sum_scaled_performance",
+            args={},
+        )
+    )
+    def benchmark(num_tokens, version):
+        topk = 9
+        hidden_size = 4096
+        dtype = torch.bfloat16
+        scaling_factor = 0.3
+
+        x = torch.randn(num_tokens, topk, hidden_size, dtype=dtype, device="cuda")
+        out = torch.empty(num_tokens, hidden_size, dtype=dtype, device="cuda")
+
+        # Warmup
+        for _ in range(3):
+            if version == "baseline":
+                compute_sum_scaled_baseline(x, out, scaling_factor)
+            elif version == "compiled":
+                compute_sum_scaled_compiled(x, out, scaling_factor)
+            else:
+                moe_sum_reduce(x, out, scaling_factor)
+
+        # Benchmark
+        quantiles = [0.5, 0.2, 0.8]
+        if version == "baseline":
+            ms, min_ms, max_ms = do_bench(
+                lambda: compute_sum_scaled_baseline(x, out, scaling_factor),
+                quantiles=quantiles,
+            )
+        elif version == "compiled":
+            ms, min_ms, max_ms = do_bench(
+                lambda: compute_sum_scaled_compiled(x, out, scaling_factor),
+                quantiles=quantiles,
+            )
+        else:
+            ms, min_ms, max_ms = do_bench(
+                lambda: moe_sum_reduce(x, out, scaling_factor), quantiles=quantiles
+            )
+
+        return 1000 * ms, 1000 * max_ms, 1000 * min_ms
+
+    return benchmark
+
+
+def verify_correctness(num_tokens=1024):
+    x = torch.randn(num_tokens, 9, 4096, device="cuda", dtype=torch.bfloat16)
+    scaling_factor = 0.3
+
+    out_baseline = torch.empty_like(x[:, 0])
+    compute_sum_scaled_baseline(x, out_baseline, scaling_factor)
+
+    out_compiled = torch.empty_like(out_baseline)
+    compute_sum_scaled_compiled(x, out_compiled, scaling_factor)
+
+    out_triton = torch.empty_like(out_baseline)
+    moe_sum_reduce(x, out_triton, scaling_factor)
+
+    if torch.allclose(
+        out_baseline, out_compiled, atol=1e-2, rtol=1e-2
+    ) and torch.allclose(out_baseline, out_triton, atol=1e-2, rtol=1e-2):
+        print("✅ All implementations match")
+    else:
+        print("❌ Implementations differ")
+        print(
+            f"Baseline vs Compiled: {(out_baseline - out_compiled).abs().max().item()}"
+        )
+        print(f"Baseline vs Triton: {(out_baseline - out_triton).abs().max().item()}")
+
+
+if __name__ == "__main__":
+    print("Running correctness verification...")
+    verify_correctness()
+
+    print("\nRunning performance benchmark...")
+    benchmark = get_benchmark()
+    benchmark.run(
+        print_data=True,
+        # save_path="./configs/benchmark_ops/sum_scaled/"
+    )

python/sglang/srt/layers/moe/fused_moe_triton/fused_moe.py

@@ -1155,6 +1155,7 @@ def inplace_fused_experts(
     a1_scale: Optional[torch.Tensor] = None,
     a2_scale: Optional[torch.Tensor] = None,
     block_shape: Optional[List[int]] = None,
+    routed_scaling_factor: Optional[float] = None,
 ) -> None:
     fused_experts_impl(
         hidden_states,
@@ -1177,6 +1178,8 @@ def inplace_fused_experts(
         a1_scale,
         a2_scale,
         block_shape,
+        False,
+        routed_scaling_factor,
     )
 
 
@@ -1200,6 +1203,7 @@ def inplace_fused_experts_fake(
     a1_scale: Optional[torch.Tensor] = None,
     a2_scale: Optional[torch.Tensor] = None,
     block_shape: Optional[List[int]] = None,
+    routed_scaling_factor: Optional[float] = None,
 ) -> None:
     pass
 
@@ -1233,6 +1237,7 @@ def outplace_fused_experts(
     a2_scale: Optional[torch.Tensor] = None,
     block_shape: Optional[List[int]] = None,
     no_combine: bool = False,
+    routed_scaling_factor: Optional[float] = None,
 ) -> torch.Tensor:
     return fused_experts_impl(
         hidden_states,
@@ -1256,6 +1261,7 @@ def outplace_fused_experts(
         a2_scale,
         block_shape,
         no_combine=no_combine,
+        routed_scaling_factor=routed_scaling_factor,
     )
 
 
@@ -1280,6 +1286,7 @@ def outplace_fused_experts_fake(
     a2_scale: Optional[torch.Tensor] = None,
     block_shape: Optional[List[int]] = None,
     no_combine: bool = False,
+    routed_scaling_factor: Optional[float] = None,
 ) -> torch.Tensor:
     return torch.empty_like(hidden_states)
 
@@ -1314,7 +1321,9 @@ def fused_experts(
     a2_scale: Optional[torch.Tensor] = None,
     block_shape: Optional[List[int]] = None,
     no_combine: bool = False,
+    routed_scaling_factor: Optional[float] = None,
 ):
+
     if inplace:
         assert not no_combine, "no combine + inplace makes no sense"
         torch.ops.sglang.inplace_fused_experts(
@@ -1337,6 +1346,7 @@ def fused_experts(
             a1_scale,
             a2_scale,
             block_shape,
+            routed_scaling_factor,
         )
         return hidden_states
     else:
@@ -1361,9 +1371,102 @@ def fused_experts(
             a2_scale,
             block_shape,
             no_combine=no_combine,
+            routed_scaling_factor=routed_scaling_factor,
         )
 
 
+# _moe_sum_reduce_kernel kernel modified from https://github.com/ModelTC/lightllm/blob/main/lightllm/common/fused_moe/moe_sum_reduce.py
+@triton.jit
+def _moe_sum_reduce_kernel(
+    input_ptr,
+    input_stride_0,
+    input_stride_1,
+    input_stride_2,
+    output_ptr,
+    output_stride_0,
+    output_stride_1,
+    token_num: int,
+    topk_num: int,
+    hidden_dim: int,
+    routed_scaling_factor: tl.constexpr,
+    BLOCK_M: tl.constexpr,
+    BLOCK_DIM: tl.constexpr,
+    NUM_STAGE: tl.constexpr,
+):
+    input_stride_0 = tl.cast(input_stride_0, dtype=tl.int64)
+    input_stride_1 = tl.cast(input_stride_1, dtype=tl.int64)
+    output_stride_0 = tl.cast(output_stride_0, dtype=tl.int64)
+
+    token_block_id = tl.program_id(0)
+    dim_block_id = tl.program_id(1)
+
+    token_start = token_block_id * BLOCK_M
+    token_end = min((token_block_id + 1) * BLOCK_M, token_num)
+
+    dim_start = dim_block_id * BLOCK_DIM
+    dim_end = min((dim_block_id + 1) * BLOCK_DIM, hidden_dim)
+
+    offs_dim = dim_start + tl.arange(0, BLOCK_DIM)
+
+    for token_index in range(token_start, token_end):
+        accumulator = tl.zeros((BLOCK_DIM,), dtype=tl.float32)
+        input_t_ptr = input_ptr + token_index * input_stride_0 + offs_dim
+        for i in tl.range(0, topk_num, num_stages=NUM_STAGE):
+            tmp = tl.load(
+                input_t_ptr + i * input_stride_1, mask=offs_dim < dim_end, other=0.0
+            )
+            accumulator += tmp
+        accumulator = accumulator * routed_scaling_factor
+        store_t_ptr = output_ptr + token_index * output_stride_0 + offs_dim
+        tl.store(
+            store_t_ptr,
+            accumulator.to(input_ptr.dtype.element_ty),
+            mask=offs_dim < dim_end,
+        )
+
+
+def moe_sum_reduce_triton(
+    input: torch.Tensor, output: torch.Tensor, routed_scaling_factor: float
+):
+    assert input.is_contiguous()
+    assert output.is_contiguous()
+
+    token_num, topk_num, hidden_dim = input.shape
+    assert output.shape[0] == token_num and output.shape[1] == hidden_dim
+
+    BLOCK_M = 1
+    BLOCK_DIM = 2048
+    NUM_STAGE = 1
+    num_warps = 8
+
+    grid = (
+        triton.cdiv(token_num, BLOCK_M),
+        triton.cdiv(hidden_dim, BLOCK_DIM),
+    )
+
+    _moe_sum_reduce_kernel[grid](
+        input,
+        *input.stride(),
+        output,
+        *output.stride(),
+        token_num=token_num,
+        topk_num=topk_num,
+        hidden_dim=hidden_dim,
+        routed_scaling_factor=routed_scaling_factor,
+        BLOCK_M=BLOCK_M,
+        BLOCK_DIM=BLOCK_DIM,
+        NUM_STAGE=NUM_STAGE,
+        num_warps=num_warps,
+    )
+    return
+
+
+@torch.compile
+def moe_sum_reduce_torch_compile(x, out, routed_scaling_factor):
+    torch.sum(x, dim=1, out=out)
+    out.mul_(routed_scaling_factor)
+
+
 def fused_experts_impl(
     hidden_states: torch.Tensor,
     w1: torch.Tensor,
@@ -1386,6 +1489,7 @@ def fused_experts_impl(
     a2_scale: Optional[torch.Tensor] = None,
     block_shape: Optional[List[int]] = None,
     no_combine: bool = False,
+    routed_scaling_factor: Optional[float] = None,
 ):
     padded_size = padding_size
     if (
@@ -1562,28 +1666,39 @@ def fused_experts_impl(
             block_shape=block_shape,
         )
 
+        if routed_scaling_factor is None:
+            routed_scaling_factor = 1.0
+
         if no_combine:
             pass
-        elif _is_hip:
+        elif _is_cuda:
-            vllm_ops.moe_sum(
+            if topk_ids.shape[1] == 1 and routed_scaling_factor == 1.0:
-                intermediate_cache3.view(*intermediate_cache3.shape),
-                out_hidden_states[begin_chunk_idx:end_chunk_idx],
-            )
-        else:
-            if topk_ids.shape[1] == 1:
                 pass  # we write directly into out_hidden_states
-            elif topk_ids.shape[1] == 2:
+            elif topk_ids.shape[1] == 2 and routed_scaling_factor == 1.0:
                 torch.add(
                     intermediate_cache3[:, 0],
                     intermediate_cache3[:, 1],
                     out=out_hidden_states[begin_chunk_idx:end_chunk_idx],
                 ).squeeze(dim=1)
-            elif topk_ids.shape[1] > 2:
+            else:
-                torch.sum(
+                # According to micro benchmark results, torch.compile can get better performance for small token.
-                    intermediate_cache3.view(*intermediate_cache3.shape),
+                if tokens_in_chunk <= 32:
-                    dim=1,
+                    moe_sum_reduce_torch_compile(
-                    out=out_hidden_states[begin_chunk_idx:end_chunk_idx],
+                        intermediate_cache3.view(*intermediate_cache3.shape),
-                )
+                        out_hidden_states[begin_chunk_idx:end_chunk_idx],
+                        routed_scaling_factor,
+                    )
+                else:
+                    moe_sum_reduce_triton(
+                        intermediate_cache3.view(*intermediate_cache3.shape),
+                        out_hidden_states[begin_chunk_idx:end_chunk_idx],
+                        routed_scaling_factor,
+                    )
+        else:
+            vllm_ops.moe_sum(
+                intermediate_cache3.view(*intermediate_cache3.shape),
+                out_hidden_states[begin_chunk_idx:end_chunk_idx],
+            )
 
     return out_hidden_states
 
@@ -1695,4 +1810,5 @@ def fused_moe(
         a2_scale=a2_scale,
         block_shape=block_shape,
         no_combine=no_combine,
+        routed_scaling_factor=routed_scaling_factor,
     )

python/sglang/srt/layers/moe/fused_moe_triton/layer.py

@@ -225,6 +225,7 @@ class UnquantizedFusedMoEMethod(FusedMoEMethodBase, CustomOp):
                 activation=activation,
                 apply_router_weight_on_input=apply_router_weight_on_input,
                 no_combine=no_combine,
+                routed_scaling_factor=routed_scaling_factor,
             )
 
     def forward_cpu(

python/sglang/srt/layers/quantization/blockwise_int8.py

@@ -411,4 +411,5 @@ class BlockInt8MoEMethod:
             a2_scale=layer.w2_input_scale,
             block_shape=self.quant_config.weight_block_size,
             no_combine=no_combine,
+            routed_scaling_factor=routed_scaling_factor,
         )

python/sglang/srt/layers/quantization/compressed_tensors/compressed_tensors_moe.py

@@ -317,6 +317,7 @@ class CompressedTensorsW8A8Fp8MoEMethod(CompressedTensorsMoEMethod):
             a1_scale=layer.w13_input_scale,
             a2_scale=layer.w2_input_scale,
             apply_router_weight_on_input=apply_router_weight_on_input,
+            routed_scaling_factor=routed_scaling_factor,
         )
 
 

python/sglang/srt/layers/quantization/fp8.py

@@ -1030,6 +1030,7 @@ class Fp8MoEMethod:
             a2_scale=layer.w2_input_scale,
             block_shape=self.quant_config.weight_block_size,
             no_combine=no_combine,
+            routed_scaling_factor=routed_scaling_factor,
         )
 
     def maybe_apply_hip_fused_experts(

python/sglang/srt/layers/quantization/moe_wna16.py

@@ -388,6 +388,7 @@ class MoeWNA16Method:
             w2_zp=layer.w2_qzeros if has_zp else None,
             block_shape=[0, layer.group_size],
             no_combine=no_combine,
+            routed_scaling_factor=routed_scaling_factor,
         )
 
     @staticmethod

python/sglang/srt/layers/quantization/w8a8_fp8.py

@@ -328,4 +328,5 @@ class W8A8FP8MoEMethod:
             a1_scale=layer.w13_input_scale,
             a2_scale=layer.w2_input_scale,
             no_combine=no_combine,
+            routed_scaling_factor=routed_scaling_factor,
         )

python/sglang/srt/layers/quantization/w8a8_int8.py

@@ -268,4 +268,5 @@ class W8A8Int8MoEMethod:
             a1_scale=layer.w13_input_scale,
             a2_scale=layer.w2_input_scale,
             no_combine=no_combine,
+            routed_scaling_factor=routed_scaling_factor,
         )

python/sglang/srt/models/deepseek_v2.py

@@ -346,7 +346,8 @@ class DeepseekV2MoE(nn.Module):
         final_hidden_states = self.experts(
             hidden_states=hidden_states, router_logits=router_logits
         )
-        final_hidden_states *= self.routed_scaling_factor
+        if not _is_cuda:
+            final_hidden_states *= self.routed_scaling_factor
         if shared_output is not None:
             final_hidden_states = final_hidden_states + shared_output
         if self.tp_size > 1: