[NVIDIA] Add Low Latency NVFP4 decode kernels from Flashinfer (#8552)

Co-authored-by: Cheng Wan <cwan@x.ai>

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

@@ -14,13 +14,9 @@ from sglang.srt.layers.moe.ep_moe.kernels import (
     silu_and_mul_masked_post_quant_fwd,
     tma_align_input_scale,
 )
-from sglang.srt.layers.moe.fused_moe_triton.layer import (
-    FlashInferFusedMoE,
-    FusedMoE,
-    should_use_flashinfer_trtllm_moe,
-)
+from sglang.srt.layers.moe.fused_moe_triton.layer import FlashInferFusedMoE, FusedMoE
 from sglang.srt.layers.moe.topk import TopKOutput
-from sglang.srt.layers.moe.utils import DeepEPMode
+from sglang.srt.layers.moe.utils import DeepEPMode, should_use_flashinfer_trtllm_moe
 from sglang.srt.layers.quantization import deep_gemm_wrapper
 from sglang.srt.layers.quantization.base_config import QuantizationConfig
 from sglang.srt.layers.quantization.fp8 import (
@@ -48,7 +44,6 @@ _is_npu = is_npu()
 _is_fp8_fnuz = is_fp8_fnuz()
 _use_aiter = get_bool_env_var("SGLANG_USE_AITER") and _is_hip
 
-
 if not (_is_npu or _is_hip):
     from sgl_kernel import silu_and_mul
 
@@ -741,6 +736,22 @@ class FlashInferEPMoE(EPMoE):
 def get_moe_impl_class():
     if global_server_args_dict["moe_a2a_backend"].is_deepep():
         return DeepEPMoE
+
+    # NEW: Direct FP4 detection (bypasses EP requirements)
+    # Check for FP4 quantization with TRTLLM flag, regardless of EP
+    if global_server_args_dict.get("enable_flashinfer_trtllm_moe", False):
+        try:
+            # Check the quantization argument directly
+            quantization = global_server_args_dict.get("quantization")
+            if quantization == "modelopt_fp4":
+                from sglang.srt.layers.moe.fused_moe_triton.layer import (
+                    FlashInferFP4MoE,
+                )
+
+                return FlashInferFP4MoE
+        except:
+            pass
+
     if global_server_args_dict["enable_flashinfer_cutlass_moe"]:
         return FusedMoE
     if get_moe_expert_parallel_world_size() > 1:

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

 # Adapted from https://github.com/vllm-project/vllm/blob/a6221a144af772fd1a68fe7e627935dc53e81738/vllm/model_executor/layers/fused_moe/layer.py
 
-import importlib.util
+import datetime
+import glob
 import logging
+import os
+import sys
 from enum import Enum
-from functools import lru_cache
 from typing import List, Optional, Tuple
 
 import torch
-from packaging import version as pkg_version
 
 from sglang.srt.distributed import (
     get_moe_expert_parallel_rank,
@@ -22,6 +23,7 @@ from sglang.srt.distributed.device_communicators.pynccl_allocator import (
 )
 from sglang.srt.eplb.expert_location import get_global_expert_location_metadata
 from sglang.srt.layers.moe.topk import StandardTopKOutput
+from sglang.srt.layers.moe.utils import should_use_flashinfer_trtllm_moe
 from sglang.srt.layers.quantization.base_config import (
     QuantizationConfig,
     QuantizeMethodBase,
@@ -29,22 +31,58 @@ from sglang.srt.layers.quantization.base_config import (
 from sglang.srt.layers.quantization.unquant import UnquantizedFusedMoEMethod
 from sglang.srt.managers.schedule_batch import global_server_args_dict
 from sglang.srt.model_loader.weight_utils import narrow_padded_param_and_loaded_weight
-from sglang.srt.utils import cpu_has_amx_support, get_bool_env_var, is_cpu, is_hip
+from sglang.srt.utils import (
+    cpu_has_amx_support,
+    get_bool_env_var,
+    is_cpu,
+    is_flashinfer_available,
+    is_hip,
+    next_power_of_2,
+)
+
+if is_flashinfer_available():
+    from flashinfer import (
+        RoutingMethodType,
+        fp4_quantize,
+        reorder_rows_for_gated_act_gemm,
+        shuffle_matrix_a,
+        shuffle_matrix_sf_a,
+    )
 
 _is_hip = is_hip()
 _is_cpu_amx_available = cpu_has_amx_support()
 _is_cpu = is_cpu()
 
+
+# Try to import FP4 TRTLLM function if flashinfer is available
+trtllm_fp4_block_scale_moe = None
+if should_use_flashinfer_trtllm_moe():
+    try:
+        from flashinfer.fused_moe import trtllm_fp4_block_scale_moe
+    except ImportError:
+        trtllm_fp4_block_scale_moe = None
+
 logger = logging.getLogger(__name__)
 
 
-@lru_cache(maxsize=1)
-def should_use_flashinfer_trtllm_moe():
-    return global_server_args_dict["enable_flashinfer_trtllm_moe"] and (
-        not importlib.util.find_spec("flashinfer")
-        or pkg_version.parse(__import__("flashinfer").__version__)
-        >= pkg_version.parse("0.2.9rc1")
-    )
+def _is_fp4_quantization_enabled():
+    """Check if ModelOpt FP4 quantization is enabled."""
+    try:
+        # Use the same simple check that works for class selection
+        quantization = global_server_args_dict.get("quantization")
+        return quantization == "modelopt_fp4"
+    except:
+        return False
+
+
+def _get_tile_tokens_dim(num_tokens, top_k, num_experts):
+    # Guess tokens per expert assuming perfect expert distribution first.
+    num_tokens_per_expert = (num_tokens * top_k) // num_experts
+    # And pad the number to the next power of 2.
+    tile_tokens_dim = next_power_of_2(num_tokens_per_expert)
+    # Cap to 8-64 tokens per CTA tile as it's the range supported by the kernel.
+    tile_tokens_dim = min(max(tile_tokens_dim, 8), 64)
+    return tile_tokens_dim
 
 
 class FusedMoeWeightScaleSupported(Enum):
@@ -157,10 +195,6 @@ 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_cutlass_moe = (
-                    self.enable_flashinfer_cutlass_moe
-                )
         assert self.quant_method is not None
 
         self.quant_config = quant_config
@@ -747,7 +781,130 @@ class FlashInferFusedMoE(FusedMoE):
             routed_scaling_factor=self.routed_scaling_factor,
         )
 
-        if self.reduce_results and (self.tp_size > 1 or self.ep_size > 1):
+        if self.reduce_results and (self.moe_tp_size > 1 or self.moe_ep_size > 1):
             final_hidden_states = tensor_model_parallel_all_reduce(final_hidden_states)
 
         return final_hidden_states
+
+
+class FlashInferFP4MoE(FusedMoE):
+    """FP4 TRTLLM MoE implementation using FlashInfer."""
+
+    def __init__(self, *args, **kwargs):
+        # Extract DeepSeek-specific parameters
+        renormalize = kwargs.pop("renormalize", True)
+        num_fused_shared_experts = kwargs.pop("num_fused_shared_experts", 0)
+        use_grouped_topk = kwargs.pop("use_grouped_topk", False)
+        num_expert_group = kwargs.pop("num_expert_group", None)
+        topk_group = kwargs.pop("topk_group", None)
+        correction_bias = kwargs.pop("correction_bias", None)
+
+        # Extract additional TopK parameters that were previously extracted in forward
+        routed_scaling_factor = kwargs.pop("routed_scaling_factor", None)
+
+        super().__init__(*args, **kwargs)
+
+        # Store DeepSeek parameters
+        self.renormalize = renormalize
+        self.num_fused_shared_experts = num_fused_shared_experts
+        self.use_grouped_topk = use_grouped_topk
+        self.num_expert_group = num_expert_group
+        self.topk_group = topk_group
+        self.correction_bias = correction_bias
+        self.routed_scaling_factor = routed_scaling_factor
+
+    # ---------------------------------------------------------------------
+    # Helper: quantize hidden states to FP4 each forward pass
+    # ---------------------------------------------------------------------
+    def _quantize_hidden_states_fp4(self, hidden_states: torch.Tensor):
+        """
+        Quantize hidden states using global scale factor from quantization method.
+
+        Global scale factor is set by ModelOptNvFp4FusedMoEMethod during weight loading.
+        Only block scales are computed at runtime for efficiency.
+
+        Returns (packed_fp4_uint8, scale_float8_e4m3fn_runtime, global_scale_float32)
+        """
+
+        # flashinfer.fp4_quantize returns (packed_uint8, scale_fp8)
+        # Only the block scales are computed at runtime
+        hs_fp4_bytes, hs_sf_bytes = fp4_quantize(
+            hidden_states,
+            self.w13_input_scale_quant,
+            16,  # sf_vec_size
+            False,  # use_ue8m0
+            False,  # is_sf_swizzled_layout
+        )
+
+        hs_fp4 = hs_fp4_bytes.reshape(
+            hidden_states.shape[0], hidden_states.shape[1] // 2
+        )
+        hs_sf = hs_sf_bytes.view(torch.float8_e4m3fn).reshape(-1)
+
+        return hs_fp4, hs_sf
+
+    def forward(self, hidden_states: torch.Tensor, topk_output):
+        """Forward pass using FP4 TRTLLM kernel.
+
+        Args:
+            hidden_states: Input tensor
+            topk_output: Should be tuple of (TopK_config, router_logits) for TRTLLM mode
+        """
+
+        # TRTLLM mode expects (TopK_config, router_logits) tuple
+        if not isinstance(topk_output, tuple) or len(topk_output) != 2:
+            raise ValueError(
+                f"FlashInferFP4MoE expects (TopK_config, router_logits) tuple, got {type(topk_output)}"
+            )
+
+        _, router_logits = topk_output
+
+        hs_fp4, hs_scale_linear = self._quantize_hidden_states_fp4(hidden_states)
+
+        router_logits = router_logits.to(torch.float32)
+
+        result = trtllm_fp4_block_scale_moe(
+            routing_logits=router_logits,
+            routing_bias=self.correction_bias.to(hidden_states.dtype),
+            hidden_states=hs_fp4,
+            hidden_states_scale=hs_scale_linear.view(torch.float8_e4m3fn).flatten(),
+            gemm1_weights=self.gemm1_weights_fp4_shuffled.data,
+            gemm1_weights_scale=self.gemm1_scales_fp4_shuffled.data.view(
+                torch.float8_e4m3fn
+            ),
+            gemm2_weights=self.gemm2_weights_fp4_shuffled.data,
+            gemm2_weights_scale=self.gemm2_scales_fp4_shuffled.data.view(
+                torch.float8_e4m3fn
+            ),
+            output1_scale_scalar=self.g1_scale_c.data,
+            output1_scale_gate_scalar=self.g1_alphas.data,
+            output2_scale_scalar=self.g2_alphas.data,
+            num_experts=self.num_experts,
+            top_k=self.top_k,
+            n_group=self.num_expert_group,
+            topk_group=self.topk_group,
+            intermediate_size=self.intermediate_size_per_partition,
+            local_expert_offset=self.moe_ep_rank * self.num_local_experts,
+            local_num_experts=self.num_local_experts,
+            routed_scaling_factor=self.routed_scaling_factor,
+            tile_tokens_dim=_get_tile_tokens_dim(
+                hidden_states.shape[0], self.top_k, self.num_local_experts
+            ),
+            routing_method_type=RoutingMethodType.DeepSeekV3,
+            do_finalize=True,
+        )[0]
+
+        return result
+
+
+def get_fused_moe_impl_class():
+    """Factory function to get the appropriate FusedMoE implementation class."""
+    if should_use_flashinfer_trtllm_moe() and _is_fp4_quantization_enabled():
+        # Use FP4 variant when FP4 quantization is enabled
+        return FlashInferFP4MoE
+    elif should_use_flashinfer_trtllm_moe():
+        # Use regular FlashInfer variant for non-FP4 FlashInfer cases
+        return FlashInferFusedMoE
+    else:
+        # Default case
+        return FusedMoE

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

+import importlib.util
 from enum import Enum
+from functools import lru_cache
+
+from packaging import version as pkg_version
+
+from sglang.srt.managers.schedule_batch import global_server_args_dict
+
+
+@lru_cache(maxsize=1)
+def should_use_flashinfer_trtllm_moe():
+    result = global_server_args_dict["enable_flashinfer_trtllm_moe"] and (
+        not importlib.util.find_spec("flashinfer")
+        or pkg_version.parse(__import__("flashinfer").__version__)
+        >= pkg_version.parse("0.2.9rc1")
+    )
+    return result
 
 
 class MoeA2ABackend(Enum):

python/sglang/srt/managers/schedule_batch.py

@@ -51,7 +51,6 @@ from sglang.srt.disaggregation.decode_schedule_batch_mixin import (
     ScheduleBatchDisaggregationDecodeMixin,
 )
 from sglang.srt.distributed.parallel_state import get_tensor_model_parallel_rank
-from sglang.srt.layers.moe.utils import DeepEPMode, MoeA2ABackend
 from sglang.srt.mem_cache.allocator import (
     BaseTokenToKVPoolAllocator,
     SWATokenToKVPoolAllocator,
@@ -109,6 +108,7 @@ GLOBAL_SERVER_ARGS_KEYS = [
     "enable_triton_kernel_moe",
     "enable_multimodal",
     "enable_symm_mem",
+    "quantization",
 ]
 
 # Put some global args for easy access

python/sglang/srt/models/deepseek_v2.py

@@ -60,12 +60,9 @@ from sglang.srt.layers.linear import (
     RowParallelLinear,
 )
 from sglang.srt.layers.logits_processor import LogitsProcessor
-from sglang.srt.layers.moe.ep_moe.layer import (
-    DeepEPMoE,
-    get_moe_impl_class,
-    should_use_flashinfer_trtllm_moe,
-)
+from sglang.srt.layers.moe.ep_moe.layer import DeepEPMoE, get_moe_impl_class
 from sglang.srt.layers.moe.topk import TopK
+from sglang.srt.layers.moe.utils import should_use_flashinfer_trtllm_moe
 from sglang.srt.layers.quantization import deep_gemm_wrapper
 from sglang.srt.layers.quantization.base_config import QuantizationConfig
 from sglang.srt.layers.quantization.fp8_kernel import (
@@ -307,19 +304,15 @@ class DeepseekV2MoE(nn.Module):
             config=config, prefix=add_prefix("gate", prefix), is_nextn=is_nextn
         )
 
-        self.topk = (
-            TopK(
-                top_k=config.num_experts_per_tok + self.num_fused_shared_experts,
-                renormalize=config.norm_topk_prob,
-                use_grouped_topk=True,
-                num_expert_group=config.n_group,
-                num_fused_shared_experts=self.num_fused_shared_experts,
-                topk_group=config.topk_group,
-                correction_bias=self.gate.e_score_correction_bias,
-                routed_scaling_factor=self.routed_scaling_factor,
-            )
-            if not should_use_flashinfer_trtllm_moe()
-            else None
+        self.topk = TopK(
+            top_k=config.num_experts_per_tok + self.num_fused_shared_experts,
+            renormalize=config.norm_topk_prob,
+            use_grouped_topk=True,
+            num_expert_group=config.n_group,
+            num_fused_shared_experts=self.num_fused_shared_experts,
+            topk_group=config.topk_group,
+            correction_bias=self.gate.e_score_correction_bias,
+            routed_scaling_factor=self.routed_scaling_factor,
         )
 
         self.experts = get_moe_impl_class()(
@@ -476,10 +469,14 @@ class DeepseekV2MoE(nn.Module):
             # router_logits: (num_tokens, n_experts)
             router_logits = self.gate(hidden_states)
             kwargs = {"hidden_states": hidden_states}
-            if self.topk is not None:
-                kwargs["topk_output"] = self.topk(hidden_states, router_logits)
+
+            # FlashInferFP4MoE (TRTLLM path) expects (TopK, router_logits) tuple
+            # Regular FusedMoE (CUTLASS path) expects StandardTopKOutput
+            if should_use_flashinfer_trtllm_moe():
+                kwargs["topk_output"] = (self.topk, router_logits)
             else:
-                kwargs["router_logits"] = router_logits
+                kwargs["topk_output"] = self.topk(hidden_states, router_logits)
+
             final_hidden_states = self.experts(**kwargs)
             if not _is_cuda:
                 final_hidden_states *= self.routed_scaling_factor
@@ -505,10 +502,14 @@ class DeepseekV2MoE(nn.Module):
         # router_logits: (num_tokens, n_experts)
         router_logits = self.gate(hidden_states)
         kwargs = {"hidden_states": hidden_states}
-        if self.topk is not None:
-            kwargs["topk_output"] = self.topk(hidden_states, router_logits)
+
+        # FlashInferFP4MoE (TRTLLM path) expects (TopK, router_logits) tuple
+        # Regular FusedMoE (CUTLASS path) expects StandardTopKOutput
+        if should_use_flashinfer_trtllm_moe():
+            kwargs["topk_output"] = (self.topk, router_logits)
         else:
-            kwargs["router_logits"] = router_logits
+            kwargs["topk_output"] = self.topk(hidden_states, router_logits)
+
         final_hidden_states = self.experts(**kwargs)
         if not _is_cuda and not _use_aiter:
             # fused in biased_grouped_topk so we can skip here

python/sglang/srt/models/glm4_moe.py

@@ -50,11 +50,9 @@ from sglang.srt.layers.linear import (
     RowParallelLinear,
 )
 from sglang.srt.layers.logits_processor import LogitsProcessor
-from sglang.srt.layers.moe.ep_moe.layer import (
-    get_moe_impl_class,
-    should_use_flashinfer_trtllm_moe,
-)
+from sglang.srt.layers.moe.ep_moe.layer import get_moe_impl_class
 from sglang.srt.layers.moe.topk import TopK
+from sglang.srt.layers.moe.utils import should_use_flashinfer_trtllm_moe
 from sglang.srt.layers.quantization.base_config import QuantizationConfig
 from sglang.srt.layers.quantization.fp8_kernel import (
     is_fp8_fnuz,

python/sglang/srt/server_args.py

@@ -481,6 +481,13 @@ class ServerArgs:
                 self.tp_size,
             ], "The expert parallel size must be 1 or the same as the tensor parallel size"
 
+        if self.enable_flashinfer_trtllm_moe:
+            if not self.disable_shared_experts_fusion:
+                self.disable_shared_experts_fusion = True
+                logger.warning(
+                    "FlashInfer TRTLLM MoE is enabled. --disable-shared-experts-fusion is automatically set."
+                )
+
         # DeepEP MoE
         if self.moe_a2a_backend == "deepep":
             if self.deepep_mode == "normal":