FlashInfer NVFP4 MoE with EP & 2-stream shared expert (#7327)

Co-authored-by: JieXin Liang <Alcanderian@users.noreply.github.com>
Co-authored-by: alcanderian <alcanderian@gmail.com>

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

@@ -1295,6 +1295,9 @@ class DeepEPMoE(EPMoE):
 def get_moe_impl_class():
     if global_server_args_dict["enable_deepep_moe"]:
         return DeepEPMoE
+    if global_server_args_dict["enable_flashinfer_moe"]:
+        # Must come before EPMoE because FusedMoE also supports enable_ep_moe
+        return FusedMoE
     if global_server_args_dict["enable_ep_moe"]:
         return EPMoE
     return FusedMoE

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

@@ -314,6 +314,8 @@ class FusedMoE(torch.nn.Module):
         inplace: bool = True,
         no_combine: bool = False,
         routed_scaling_factor: Optional[float] = None,
+        enable_flashinfer_moe: Optional[bool] = False,
+        enable_ep_moe: Optional[bool] = False,
     ):
         super().__init__()
 
@@ -324,9 +326,34 @@ class FusedMoE(torch.nn.Module):
         self.tp_size = (
             tp_size if tp_size is not None else get_tensor_model_parallel_world_size()
         )
+        self.tp_rank = get_tensor_model_parallel_rank()
+        self.num_experts = num_experts
+        self.expert_map = None
+        self.enable_flashinfer_moe = enable_flashinfer_moe
+        if enable_ep_moe:
+            assert (
+                self.enable_flashinfer_moe
+            ), "FusedMoE only supports EP with --enable-flashinfer-moe"
+            self.ep_size = self.tp_size
+            self.ep_rank = self.tp_rank
+            self.tp_size = 1
+            self.tp_rank = 0
+            # Create a tensor of size num_experts filled with -1
+            self.expert_map = torch.full((self.num_experts,), -1, dtype=torch.int32)
+            # Create a expert map for the local experts
+            assert num_experts % self.ep_size == 0
+            self.local_num_experts = num_experts // self.ep_size
+            self.expert_map[
+                self.ep_rank
+                * self.local_num_experts : (self.ep_rank + 1)
+                * self.local_num_experts
+            ] = torch.arange(0, self.local_num_experts, dtype=torch.int32, device="cpu")
+        else:
+            self.ep_size = 1
+            self.ep_rank = 0
+            self.local_num_experts = num_experts
         self.routed_scaling_factor = routed_scaling_factor
         self.top_k = top_k
-        self.num_experts = num_experts
         assert intermediate_size % self.tp_size == 0
         self.intermediate_size_per_partition = intermediate_size // self.tp_size
         self.reduce_results = reduce_results
@@ -344,7 +371,6 @@ class FusedMoE(torch.nn.Module):
         self.use_presharded_weights = use_presharded_weights
         self.inplace = inplace
         self.no_combine = no_combine
-        self.local_num_experts = num_experts
 
         if quant_config is None:
             self.quant_method: Optional[QuantizeMethodBase] = (
@@ -352,11 +378,13 @@ class FusedMoE(torch.nn.Module):
             )
         else:
             self.quant_method = quant_config.get_quant_method(self, prefix)
+            if self.quant_method.__class__.__name__ == "ModelOptNvFp4FusedMoEMethod":
+                self.quant_method.enable_flashinfer_moe = self.enable_flashinfer_moe
         assert self.quant_method is not None
 
         self.quant_method.create_weights(
             layer=self,
-            num_experts=num_experts,
+            num_experts=self.local_num_experts,
             hidden_size=hidden_size,
             # FIXME: figure out which intermediate_size to use
             intermediate_size=self.intermediate_size_per_partition,
@@ -450,12 +478,15 @@ class FusedMoE(torch.nn.Module):
 
         # Narrow parameter and load.
         # w1, gate_proj: Load into first logical weight of w13.
-        if shard_id == "w1":
-            expert_data = expert_data.narrow(shard_dim, 0, shard_size)
         # w3, up_proj: Load into second logical weight of w13.
+        # trtllm cutlass kernel assumes differently
+        assert shard_id in ("w1", "w3")
+        switch_w13 = getattr(self.quant_method, "load_up_proj_weight_first", False)
+        if (switch_w13 and shard_id == "w1") or (not switch_w13 and shard_id == "w3"):
+            start = shard_size
         else:
-            assert shard_id == "w3"
-            expert_data = expert_data.narrow(shard_dim, shard_size, shard_size)
+            start = 0
+        expert_data = expert_data.narrow(shard_dim, start, shard_size)
         expert_data.copy_(loaded_weight)
 
     def _load_w2(
@@ -509,6 +540,11 @@ class FusedMoE(torch.nn.Module):
             assert shard_id in ("w1", "w3")
             expert_data.copy_(loaded_weight)
 
+    def _map_global_expert_id_to_local_expert_id(self, expert_id: int) -> int:
+        if self.expert_map is None:
+            return expert_id
+        return self.expert_map[expert_id].item()
+
     def weight_loader(
         self,
         param: torch.nn.Parameter,
@@ -517,6 +553,13 @@ class FusedMoE(torch.nn.Module):
         shard_id: str,
         expert_id: int,
     ) -> None:
+        expert_id = self._map_global_expert_id_to_local_expert_id(expert_id)
+        if expert_id == -1:
+            return
+
+        # TP rank is set to 0 if EP is enabled
+        tp_rank = 0 if self.ep_size > 1 else get_tensor_model_parallel_rank()
+
         # compressed-tensors checkpoints with packed weights are stored flipped
         # TODO (mgoin): check self.quant_method.quant_config.quant_format
         # against known CompressionFormat enum values that have this quality
@@ -541,7 +584,6 @@ class FusedMoE(torch.nn.Module):
         SHARD_ID_TO_SHARDED_DIM = {"w1": 0, "w2": 1, "w3": 0}
 
         expert_data = param.data[expert_id]
-        tp_rank = get_tensor_model_parallel_rank()
 
         # is_transposed: if the dim to shard the weight
         # should be flipped. Required by GPTQ, compressed-tensors
@@ -549,7 +591,7 @@ class FusedMoE(torch.nn.Module):
         is_transposed = getattr(param, "is_transposed", False)
         shard_dim = SHARD_ID_TO_SHARDED_DIM[shard_id]
         if is_transposed:
-            shard_dim = ~shard_dim
+            shard_dim = int(not shard_dim)
 
         # Case input scale: input_scale loading is only supported for fp8
         if "input_scale" in weight_name:
@@ -690,9 +732,19 @@ class FusedMoE(torch.nn.Module):
             activation=self.activation,
             apply_router_weight_on_input=self.apply_router_weight_on_input,
             routed_scaling_factor=self.routed_scaling_factor,
+            **(
+                dict(
+                    tp_rank=self.tp_rank,
+                    tp_size=self.tp_size,
+                    ep_rank=self.ep_rank,
+                    ep_size=self.ep_size,
+                )
+                if self.quant_method.__class__.__name__ == "ModelOptNvFp4FusedMoEMethod"
+                else {}
+            ),
         )
 
-        if self.reduce_results and self.tp_size > 1:
+        if self.reduce_results and (self.tp_size > 1 or self.ep_size > 1):
             final_hidden_states = tensor_model_parallel_all_reduce(final_hidden_states)
 
         return final_hidden_states

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

@@ -29,11 +29,17 @@ from sglang.srt.layers.quantization.utils import (
     requantize_with_max_scale,
 )
 from sglang.srt.layers.radix_attention import RadixAttention
-from sglang.srt.utils import is_cuda
+from sglang.srt.utils import is_cuda, next_power_of_2
 
 if is_cuda():
     from sgl_kernel import cutlass_scaled_fp4_mm, scaled_fp4_quant
 
+try:
+    from flashinfer import fp4_quantize as fp4_quantize
+    from flashinfer.fused_moe import cutlass_fused_moe as flashinfer_cutlass_fused_moe
+except ImportError:
+    flashinfer_cutlass_fused_moe = None
+
 # Initialize logger for the module
 logger = logging.getLogger(__name__)
 
@@ -429,6 +435,9 @@ class ModelOptFp4LinearMethod(LinearMethodBase):
         layer.alpha = Parameter(
             layer.input_scale * layer.weight_scale_2, requires_grad=False
         )
+        layer.input_scale_inv = Parameter(
+            (1 / input_scale_2).to(torch.float32), requires_grad=False
+        )
 
         # Pad and blockwise interleave weight_scale
         scales = layer.weight_scale
@@ -467,7 +476,7 @@ class ModelOptFp4LinearMethod(LinearMethodBase):
         output_shape = [x_m, w_n]
 
         # Quantize BF16 or FP16 to (FP4 and interleaved block scale)
-        x_fp4, x_scale_interleaved = scaled_fp4_quant(x, 1 / layer.input_scale)
+        x_fp4, x_scale_interleaved = scaled_fp4_quant(x, layer.input_scale_inv)
 
         assert x_fp4.dtype == torch.uint8
         assert x_scale_interleaved.dtype == torch.float8_e4m3fn
@@ -521,6 +530,7 @@ class ModelOptNvFp4FusedMoEMethod:
                 " quantization. Please use Blackwell and"
                 " above."
             )
+        self.enable_flashinfer_moe = False
 
     def create_weights(
         self,
@@ -674,6 +684,9 @@ class ModelOptNvFp4FusedMoEMethod:
         w13_weight_scale_2 = layer.w13_weight_scale_2[:, 0]
         layer.w13_weight_scale_2 = Parameter(w13_weight_scale_2, requires_grad=False)
 
+        if self.enable_flashinfer_moe:
+            w13_input_scale = layer.w13_input_scale.max().to(torch.float32)
+        else:
             w13_input_scale = layer.w13_input_scale.max(dim=1).values.to(torch.float32)
         layer.g1_alphas = Parameter(
             (w13_input_scale * w13_weight_scale_2).to(torch.float32),
@@ -700,14 +713,19 @@ class ModelOptNvFp4FusedMoEMethod:
         layer.w13_weight = Parameter(layer.w13_weight.data, requires_grad=False)
 
         # GEMM 2
+        if self.enable_flashinfer_moe:
+            w2_input_scale = layer.w2_input_scale.max().to(torch.float32)
+        else:
+            w2_input_scale = layer.w2_input_scale
+
         layer.g2_alphas = Parameter(
-            (layer.w2_input_scale * layer.w2_weight_scale_2).to(torch.float32),
+            (w2_input_scale * layer.w2_weight_scale_2).to(torch.float32),
             requires_grad=False,
         )
 
         # This is for quantization, so we need to invert it.
         layer.w2_input_scale_quant = Parameter(
-            (1 / layer.w2_input_scale).to(torch.float32), requires_grad=False
+            (1 / w2_input_scale).to(torch.float32), requires_grad=False
         )
 
         assert (
@@ -727,11 +745,16 @@ class ModelOptNvFp4FusedMoEMethod:
         layer.cutlass_moe_params = CutlassMoEParams(
             CutlassMoEType.BlockscaledFP4,
             device,
-            num_experts=layer.num_experts,
+            num_experts=layer.num_experts,  # global num experts
             intermediate_size_per_partition=layer.w2_weight.shape[2] * 2,  # n
             hidden_size=layer.w13_weight.shape[2] * 2,
         )  # k
 
+    @property
+    def load_up_proj_weight_first(self) -> bool:
+        # FlashInfer CUTLASS kernel assumes [Up, Gate] Proj as W13
+        return self.enable_flashinfer_moe
+
     def apply(
         self,
         layer: torch.nn.Module,
@@ -750,11 +773,13 @@ class ModelOptNvFp4FusedMoEMethod:
         inplace: bool = True,
         no_combine: bool = False,
         routed_scaling_factor: Optional[float] = None,
+        ep_rank: Optional[int] = None,
+        ep_size: Optional[int] = None,
+        tp_rank: Optional[int] = None,
+        tp_size: Optional[int] = None,
     ) -> torch.Tensor:
 
         assert activation == "silu", "Only SiLU activation is supported."
-
-        from sglang.srt.layers.moe.fused_moe_triton.fused_moe import fused_experts
         from sglang.srt.layers.moe.topk import select_experts
 
         topk_weights, topk_ids = select_experts(
@@ -771,6 +796,35 @@ class ModelOptNvFp4FusedMoEMethod:
             routed_scaling_factor=routed_scaling_factor,
         )
 
+        if self.enable_flashinfer_moe:
+            assert (
+                not apply_router_weight_on_input
+            ), "apply_router_weight_on_input is not supported for Flashinfer"
+            # TRTLLM Cutlass moe takes in activations in BF16/Half/nvfp4 precision
+            # and fp4 quantized weights loaded from the checkpoint
+            output = flashinfer_cutlass_fused_moe(
+                x,
+                topk_ids.to(torch.int),
+                topk_weights,
+                layer.w13_weight.view(torch.long),
+                layer.w2_weight.view(torch.long),
+                x.dtype,
+                quant_scales=[
+                    layer.w13_input_scale_quant,
+                    layer.w13_blockscale_swizzled.view(torch.int32),
+                    layer.g1_alphas,
+                    layer.w2_input_scale_quant,
+                    layer.w2_blockscale_swizzled.view(torch.int32),
+                    layer.g2_alphas,
+                ],
+                ep_size=ep_size,
+                ep_rank=ep_rank,
+                tp_size=tp_size,
+                tp_rank=tp_rank,
+                tune_max_num_tokens=next_power_of_2(x.shape[0]),
+            )
+            return output[0]
+
         from sglang.srt.layers.moe.cutlass_moe import cutlass_moe_fp4
 
         return cutlass_moe_fp4(

python/sglang/srt/managers/schedule_batch.py

@@ -86,6 +86,7 @@ GLOBAL_SERVER_ARGS_KEYS = [
     "enable_deepep_moe",
     "deepep_mode",
     "enable_ep_moe",
+    "enable_flashinfer_moe",
     "moe_dense_tp_size",
     "ep_dispatch_algorithm",
     "deepep_config",

python/sglang/srt/models/deepseek_v2.py

@@ -226,6 +226,7 @@ class DeepseekV2MoE(nn.Module):
         layer_id: int,
         quant_config: Optional[QuantizationConfig] = None,
         prefix: str = "",
+        alt_stream: Optional[torch.cuda.Stream] = None,
     ):
         super().__init__()
         self.tp_size = get_tensor_model_parallel_world_size()
@@ -238,6 +239,7 @@ class DeepseekV2MoE(nn.Module):
         )
         self.config = config
         self.layer_id = layer_id
+        self.alt_stream = alt_stream
 
         if self.tp_size > config.n_routed_experts:
             raise ValueError(
@@ -275,6 +277,15 @@ class DeepseekV2MoE(nn.Module):
                 if global_server_args_dict["enable_deepep_moe"]
                 else {}
             ),
+            # Additional args for FusedMoE
+            **(
+                dict(
+                    enable_flashinfer_moe=True,
+                    enable_ep_moe=global_server_args_dict["enable_ep_moe"],
+                )
+                if global_server_args_dict["enable_flashinfer_moe"]
+                else {}
+            ),
         )
 
         if config.n_shared_experts is not None and self.num_fused_shared_experts == 0:
@@ -338,10 +349,36 @@ class DeepseekV2MoE(nn.Module):
         self, hidden_states: torch.Tensor, forward_batch: Optional[ForwardBatch] = None
     ) -> torch.Tensor:
         if not self._enable_deepep_moe:
+            DUAL_STREAM_TOKEN_THRESHOLD = 1024
+            if (
+                self.alt_stream is not None
+                and self.num_fused_shared_experts == 0
+                and hidden_states.shape[0] <= DUAL_STREAM_TOKEN_THRESHOLD
+            ):
+                return self.forward_normal_dual_stream(hidden_states)
+            else:
                 return self.forward_normal(hidden_states)
         else:
             return self.forward_deepep(hidden_states, forward_batch)
 
+    def forward_normal_dual_stream(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        current_stream = torch.cuda.current_stream()
+        self.alt_stream.wait_stream(current_stream)
+        shared_output = self._forward_shared_experts(hidden_states)
+        with torch.cuda.stream(self.alt_stream):
+            # router_logits: (num_tokens, n_experts)
+            router_logits = self.gate(hidden_states)
+            final_hidden_states = self.experts(
+                hidden_states=hidden_states, router_logits=router_logits
+            )
+            if not _is_cuda:
+                final_hidden_states *= self.routed_scaling_factor
+        current_stream.wait_stream(self.alt_stream)
+        final_hidden_states = final_hidden_states + shared_output
+        if self.tp_size > 1:
+            final_hidden_states = tensor_model_parallel_all_reduce(final_hidden_states)
+        return final_hidden_states
+
     def forward_normal(self, hidden_states: torch.Tensor) -> torch.Tensor:
         shared_output = self._forward_shared_experts(hidden_states)
         # router_logits: (num_tokens, n_experts)
@@ -1446,6 +1483,7 @@ class DeepseekV2DecoderLayer(nn.Module):
                 quant_config=quant_config,
                 prefix=add_prefix("mlp", prefix),
                 layer_id=self.layer_id,
+                alt_stream=alt_stream,
             )
         else:
             if enable_moe_dense_fully_dp():

python/sglang/srt/server_args.py

@@ -152,6 +152,7 @@ class ServerArgs:
     ep_size: int = 1
     enable_ep_moe: bool = False
     enable_deepep_moe: bool = False
+    enable_flashinfer_moe: bool = False
     deepep_mode: Optional[Literal["auto", "normal", "low_latency"]] = "auto"
     ep_num_redundant_experts: int = 0
     ep_dispatch_algorithm: Optional[Literal["static", "dynamic", "fake"]] = None
@@ -244,7 +245,15 @@ class ServerArgs:
             logger.warning(
                 f"EP MoE is enabled. The expert parallel size is adjusted to be the same as the tensor parallel size[{self.tp_size}]."
             )
-
+        if self.enable_flashinfer_moe:
+            assert (
+                self.quantization == "modelopt_fp4"
+            ), "modelopt_fp4 quantization is required for Flashinfer MOE"
+            os.environ["TRTLLM_ENABLE_PDL"] = "1"
+            self.disable_shared_experts_fusion = True
+            logger.warning(
+                f"Flashinfer MoE is enabled. Shared expert fusion is disabled."
+            )
         # Set missing default values
         if self.tokenizer_path is None:
             self.tokenizer_path = self.model_path
@@ -1162,6 +1171,11 @@ class ServerArgs:
             action="store_true",
             help="Enabling expert parallelism for moe. The ep size is equal to the tp size.",
         )
+        parser.add_argument(
+            "--enable-flashinfer-moe",
+            action="store_true",
+            help="Enable FlashInfer CUTLASS MoE backend for modelopt_fp4 quant on Blackwell. Supports MoE-EP with --enable-ep-moe",
+        )
         parser.add_argument(
             "--enable-deepep-moe",
             action="store_true",