[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:
@@ -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(
|
||||
|
||||
Reference in New Issue
Block a user