[GLM5][MoE] perf: Write FlashInfer TRT-LLM MoE output directly (#28416)

python/sglang/srt/layers/moe/flashinfer_trtllm_moe.py

@@ -5,7 +5,7 @@ import torch
 from sglang.srt.utils.custom_op import register_custom_op
 
 
-def _fake_fp8_block_scale_moe(
+def _fake_fp8_block_scale_moe_out(
     routing_logits: torch.Tensor,
     routing_bias: Optional[torch.Tensor],
     hidden_states: torch.Tensor,
@@ -14,6 +14,7 @@ def _fake_fp8_block_scale_moe(
     gemm1_weights_scale: torch.Tensor,
     gemm2_weights: torch.Tensor,
     gemm2_weights_scale: torch.Tensor,
+    output: torch.Tensor,
     num_experts: int,
     top_k: int,
     n_group: Optional[int],
@@ -29,14 +30,15 @@ def _fake_fp8_block_scale_moe(
     tune_max_num_tokens: int = 8192,
     fp8_quantization_type: Optional[int] = None,
     activation_type: Optional[int] = None,
-) -> torch.Tensor:
-    return torch.empty(
-        hidden_states.shape, dtype=torch.bfloat16, device=hidden_states.device
-    )
+) -> None:
+    return None
 
 
-@register_custom_op(fake_impl=_fake_fp8_block_scale_moe)
-def trtllm_fp8_block_scale_moe_wrapper(
+@register_custom_op(
+    fake_impl=_fake_fp8_block_scale_moe_out,
+    mutates_args=["output"],
+)
+def trtllm_fp8_block_scale_moe_out_wrapper(
     routing_logits: torch.Tensor,
     routing_bias: Optional[torch.Tensor],
     hidden_states: torch.Tensor,
@@ -45,6 +47,7 @@ def trtllm_fp8_block_scale_moe_wrapper(
     gemm1_weights_scale: torch.Tensor,
     gemm2_weights: torch.Tensor,
     gemm2_weights_scale: torch.Tensor,
+    output: torch.Tensor,
     num_experts: int,
     top_k: int,
     n_group: Optional[int],
@@ -60,7 +63,7 @@ def trtllm_fp8_block_scale_moe_wrapper(
     tune_max_num_tokens: int = 8192,
     fp8_quantization_type: Optional[int] = None,
     activation_type: Optional[int] = None,
-) -> torch.Tensor:
+) -> None:
     try:
         from flashinfer.fused_moe import trtllm_fp8_block_scale_moe
     except ImportError as e:
@@ -68,6 +71,7 @@ def trtllm_fp8_block_scale_moe_wrapper(
             "Can't import trtllm_fp8_block_scale_moe from flashinfer. "
             "Please check flashinfer version."
         ) from e
+
     kwargs = {
         "routing_logits": routing_logits,
         "routing_bias": routing_bias,
@@ -77,6 +81,7 @@ def trtllm_fp8_block_scale_moe_wrapper(
         "gemm1_weights_scale": gemm1_weights_scale,
         "gemm2_weights": gemm2_weights,
         "gemm2_weights_scale": gemm2_weights_scale,
+        "output": output,
         "num_experts": num_experts,
         "top_k": top_k,
         "n_group": n_group,
@@ -101,10 +106,10 @@ def trtllm_fp8_block_scale_moe_wrapper(
 
         kwargs["activation_type"] = ActivationType(activation_type)
 
-    return trtllm_fp8_block_scale_moe(**kwargs)
+    trtllm_fp8_block_scale_moe(**kwargs)
 
 
-def _fake_fp8_block_scale_routed_moe(
+def _fake_fp8_block_scale_routed_moe_out(
     topk_ids: torch.Tensor,
     routing_bias: Optional[torch.Tensor],
     hidden_states: torch.Tensor,
@@ -121,6 +126,7 @@ def _fake_fp8_block_scale_routed_moe(
     local_expert_offset: int,
     local_num_experts: int,
     routed_scaling_factor: Optional[float],
+    output: torch.Tensor,
     routing_method_type: int = 0,
     use_shuffled_weight: bool = False,
     weight_layout: int = 0,
@@ -128,14 +134,15 @@ def _fake_fp8_block_scale_routed_moe(
     tune_max_num_tokens: int = 8192,
     fp8_quantization_type: Optional[int] = None,
     activation_type: Optional[int] = None,
-) -> torch.Tensor:
-    return torch.empty(
-        hidden_states.shape, dtype=torch.bfloat16, device=hidden_states.device
-    )
+) -> None:
+    return None
 
 
-@register_custom_op(fake_impl=_fake_fp8_block_scale_routed_moe)
-def trtllm_fp8_block_scale_routed_moe_wrapper(
+@register_custom_op(
+    fake_impl=_fake_fp8_block_scale_routed_moe_out,
+    mutates_args=["output"],
+)
+def trtllm_fp8_block_scale_routed_moe_out_wrapper(
     topk_ids: torch.Tensor,
     routing_bias: Optional[torch.Tensor],
     hidden_states: torch.Tensor,
@@ -152,6 +159,7 @@ def trtllm_fp8_block_scale_routed_moe_wrapper(
     local_expert_offset: int,
     local_num_experts: int,
     routed_scaling_factor: Optional[float],
+    output: torch.Tensor,
     routing_method_type: int = 0,
     use_shuffled_weight: bool = False,
     weight_layout: int = 0,
@@ -159,7 +167,7 @@ def trtllm_fp8_block_scale_routed_moe_wrapper(
     tune_max_num_tokens: int = 8192,
     fp8_quantization_type: Optional[int] = None,
     activation_type: Optional[int] = None,
-) -> torch.Tensor:
+) -> None:
     try:
         from flashinfer.fused_moe import trtllm_fp8_block_scale_routed_moe
     except ImportError as e:
@@ -167,6 +175,7 @@ def trtllm_fp8_block_scale_routed_moe_wrapper(
             "Can't import trtllm_fp8_block_scale_routed_moe from flashinfer. "
             "Please check flashinfer version."
         ) from e
+
     kwargs = {
         "topk_ids": topk_ids,
         "routing_bias": routing_bias,
@@ -176,6 +185,7 @@ def trtllm_fp8_block_scale_routed_moe_wrapper(
         "gemm1_weights_scale": gemm1_weights_scale,
         "gemm2_weights": gemm2_weights,
         "gemm2_weights_scale": gemm2_weights_scale,
+        "output": output,
         "num_experts": num_experts,
         "top_k": top_k,
         "n_group": n_group,
@@ -200,7 +210,7 @@ def trtllm_fp8_block_scale_routed_moe_wrapper(
 
         kwargs["activation_type"] = ActivationType(activation_type)
 
-    return trtllm_fp8_block_scale_routed_moe(**kwargs)
+    trtllm_fp8_block_scale_routed_moe(**kwargs)
 
 
 def _fake_fp8_per_tensor_scale_moe(

python/sglang/srt/layers/moe/moe_runner/flashinfer_trtllm.py

@@ -25,8 +25,8 @@ from sglang.srt.distributed.device_communicators.pynccl_allocator import (
 from sglang.srt.environ import envs
 from sglang.srt.layers.dp_attention import is_allocation_symmetric
 from sglang.srt.layers.moe.flashinfer_trtllm_moe import (
-    trtllm_fp8_block_scale_moe_wrapper,
-    trtllm_fp8_block_scale_routed_moe_wrapper,
+    trtllm_fp8_block_scale_moe_out_wrapper,
+    trtllm_fp8_block_scale_routed_moe_out_wrapper,
     trtllm_fp8_per_tensor_scale_moe_wrapper,
 )
 from sglang.srt.layers.moe.moe_runner.base import (
@@ -735,7 +735,7 @@ def fused_experts_none_to_flashinfer_trtllm_fp8(
             ), "runner_config.top_k is required for flashinfer_trtllm_routed."
             packed_topk_ids = _get_packed_topk_ids_for_flashinfer_routed(topk_output)
 
-            output = trtllm_fp8_block_scale_routed_moe_wrapper(
+            trtllm_fp8_block_scale_routed_moe_out_wrapper(
                 topk_ids=packed_topk_ids,
                 routing_bias=None,
                 hidden_states=a_q,
@@ -762,6 +762,7 @@ def fused_experts_none_to_flashinfer_trtllm_fp8(
                     else routing_method_type
                 ),
                 use_shuffled_weight=use_shuffled_weight,
+                output=symm_output,
                 tune_max_num_tokens=next_power_of_2(a_q.shape[0]),
                 fp8_quantization_type=int(fp8_quantization_type),
                 activation_type=quant_info.activation_type,
@@ -769,7 +770,7 @@ def fused_experts_none_to_flashinfer_trtllm_fp8(
         else:
             assert TopKOutputChecker.format_is_bypassed(topk_output)
 
-            output = trtllm_fp8_block_scale_moe_wrapper(
+            trtllm_fp8_block_scale_moe_out_wrapper(
                 routing_logits=router_logits,
                 routing_bias=correction_bias,
                 hidden_states=a_q,
@@ -778,6 +779,7 @@ def fused_experts_none_to_flashinfer_trtllm_fp8(
                 gemm1_weights_scale=quant_info.w13_weight_scale_inv,
                 gemm2_weights=quant_info.w2_weight,
                 gemm2_weights_scale=quant_info.w2_weight_scale_inv,
+                output=symm_output,
                 num_experts=quant_info.global_num_experts,
                 top_k=topk_config.top_k,
                 n_group=topk_config.num_expert_group,
@@ -796,8 +798,6 @@ def fused_experts_none_to_flashinfer_trtllm_fp8(
                 fp8_quantization_type=int(fp8_quantization_type),
                 activation_type=quant_info.activation_type,
             )
-        # TODO: Once https://github.com/flashinfer-ai/flashinfer/issues/2703 is fixed, pass output to moe kernel and remove this copy.
-        symm_output.copy_(output)
         output = symm_output
     else:
         assert TopKOutputChecker.format_is_bypassed(topk_output)