[Perf] Do FP4 quant before All gather on flashinfer trtllmgen MOE (#30014)

Signed-off-by: jiahanc <173873397+jiahanc@users.noreply.github.com>
This commit is contained in:
jiahanc
2025-12-16 13:01:48 -08:00
committed by GitHub
parent f21f5ea38c
commit 254a7f8fd6
9 changed files with 165 additions and 31 deletions
@@ -71,6 +71,18 @@ class FusedMoEMethodBase(QuantizeMethodBase):
"implementation based on the prepare_finalize"
)
def prepare_dp_allgather_tensor(
self,
layer: "FusedMoE", # type: ignore[name-defined] # noqa: F821
hidden_states: torch.Tensor,
router_logits: torch.Tensor,
) -> tuple[torch.Tensor, list[torch.Tensor]]:
"""Hook to prepare tensors and extra tensors for DP allgather + EP dispatch."""
raise NotImplementedError(
"Method 'prepare_dp_allgather_tensor' is not implemented in "
f"{self.__class__.__name__}."
)
@abstractmethod
def get_fused_moe_quant_config(
self, layer: torch.nn.Module
+39 -2
View File
@@ -44,6 +44,7 @@ from vllm.model_executor.layers.quantization.utils.flashinfer_utils import (
is_flashinfer_supporting_global_sf,
)
from vllm.platforms import current_platform
from vllm.utils.flashinfer import has_flashinfer_trtllm_fused_moe
from vllm.utils.math_utils import cdiv, round_up
from vllm.utils.torch_utils import (
aux_stream,
@@ -1933,10 +1934,46 @@ class FusedMoE(CustomOp):
)
with sp_ctx:
extra_tensors = None
if do_naive_dispatch_combine:
hidden_states_combined, router_logits = get_ep_group().dispatch(
hidden_states, router_logits, self.is_sequence_parallel
# Avoid circular import
from vllm.model_executor.layers.quantization.modelopt import (
ModelOptNvFp4FusedMoE,
)
post_quant_allgather = (
has_flashinfer_trtllm_fused_moe()
and self.quant_method is not None
and self.dp_size > 1
and self.use_ep
and isinstance(self.quant_method, ModelOptNvFp4FusedMoE)
)
if post_quant_allgather:
hidden_states_to_dispatch, extra_tensors = (
self.quant_method.prepare_dp_allgather_tensor(
self, hidden_states, router_logits
)
)
else:
hidden_states_to_dispatch = hidden_states
dispatch_res = get_ep_group().dispatch(
hidden_states_to_dispatch,
router_logits,
self.is_sequence_parallel,
extra_tensors=extra_tensors,
)
if extra_tensors is not None:
hidden_states_combined, router_logits, extra_tensors_combined = (
dispatch_res
)
hidden_states_combined = (
hidden_states_combined,
extra_tensors_combined[0],
)
else:
hidden_states_combined, router_logits = dispatch_res
# Run shared experts before matrix multiply.
# because matrix multiply maybe modify the hidden_states.
if has_separate_shared_experts and not use_shared_experts_stream:
@@ -1522,6 +1522,24 @@ class ModelOptNvFp4FusedMoE(FusedMoEMethodBase):
w2_blockscale_swizzled, requires_grad=False
)
def prepare_dp_allgather_tensor(
self,
layer: FusedMoE,
hidden_states: torch.Tensor,
router_logits: torch.Tensor,
) -> tuple[torch.Tensor, list[torch.Tensor]]:
"""Optionally prepare extra tensors to carry through DP allgather/EP."""
import flashinfer
a1_gscale = layer.w13_input_scale_quant
hidden_states_fp4, hidden_states_sf = flashinfer.fp4_quantize(
hidden_states,
a1_gscale,
is_sf_swizzled_layout=False,
)
extra_tensors: list[torch.Tensor] = [hidden_states_sf]
return hidden_states_fp4, extra_tensors
def get_fused_moe_quant_config(
self, layer: torch.nn.Module
) -> FusedMoEQuantConfig | None:
@@ -1576,8 +1594,13 @@ class ModelOptNvFp4FusedMoE(FusedMoEMethodBase):
e_score_correction_bias=layer.e_score_correction_bias,
)
# Hidden_states in select_experts is only used to extract metadata
if isinstance(x, tuple):
x_routing, _ = x
else:
x_routing = x
topk_weights, topk_ids, _ = layer.select_experts(
hidden_states=x,
hidden_states=x_routing,
router_logits=router_logits,
)
@@ -238,7 +238,7 @@ def prepare_static_weights_for_trtllm_fp4_moe(
def flashinfer_trtllm_fp4_moe(
layer: torch.nn.Module,
x: torch.Tensor,
x: torch.Tensor | tuple[torch.Tensor, torch.Tensor],
router_logits: torch.Tensor,
top_k: int,
global_num_experts: int,
@@ -269,12 +269,16 @@ def flashinfer_trtllm_fp4_moe(
from vllm.model_executor.models.llama4 import Llama4MoE
# Quantize input to FP4
a1_gscale = layer.w13_input_scale_quant
(hidden_states_fp4, hidden_states_scale_linear_fp4) = flashinfer.fp4_quantize(
x,
a1_gscale,
is_sf_swizzled_layout=False,
)
if isinstance(x, tuple):
hidden_states_fp4, hidden_states_scale_linear_fp4 = x
else:
# hidden_states is the already quantized
a1_gscale = layer.w13_input_scale_quant
(hidden_states_fp4, hidden_states_scale_linear_fp4) = flashinfer.fp4_quantize(
x,
a1_gscale,
is_sf_swizzled_layout=False,
)
# Determine routing method type
use_llama4_routing = custom_routing_function is Llama4MoE.custom_routing_function
@@ -360,13 +364,17 @@ def flashinfer_trtllm_fp4_routed_moe(
torch.bfloat16
).view(torch.int16)
# Quantize input to FP4
a1_gscale = layer.w13_input_scale_quant
(hidden_states_fp4, hidden_states_scale_linear_fp4) = flashinfer.fp4_quantize(
x,
a1_gscale,
is_sf_swizzled_layout=False,
)
if isinstance(x, tuple):
# Hidden_states is the already quantized
hidden_states_fp4, hidden_states_scale_linear_fp4 = x
else:
# Quantize input to FP4
a1_gscale = layer.w13_input_scale_quant
(hidden_states_fp4, hidden_states_scale_linear_fp4) = flashinfer.fp4_quantize(
x,
a1_gscale,
is_sf_swizzled_layout=False,
)
# Call TRT-LLM FP4 block-scale MoE kernel
out = flashinfer.fused_moe.trtllm_fp4_block_scale_routed_moe(