refactor: Change scaling factors calculation for flashinfer FusedMoE (#22812)
Signed-off-by: Amir Klein <203507526+amirkl94@users.noreply.github.com> Co-authored-by: Michael Goin <mgoin64@gmail.com>
This commit is contained in:
@@ -25,8 +25,8 @@ from vllm.model_executor.layers.quantization.utils.flashinfer_fp4_moe import (
|
||||
build_flashinfer_fp4_cutlass_moe_kernel,
|
||||
flashinfer_fp4_cutlass_moe_forward, reorder_w1w3_to_w3w1)
|
||||
from vllm.model_executor.layers.quantization.utils.flashinfer_utils import (
|
||||
apply_flashinfer_per_tensor_scale_fp8, rotate_flashinfer_fp8_moe_weights,
|
||||
swap_w13_to_w31)
|
||||
apply_flashinfer_per_tensor_scale_fp8, register_moe_scaling_factors,
|
||||
rotate_flashinfer_fp8_moe_weights, swap_w13_to_w31)
|
||||
from vllm.model_executor.layers.quantization.utils.marlin_utils_fp4 import (
|
||||
apply_fp4_marlin_linear, is_fp4_marlin_supported,
|
||||
prepare_fp4_layer_for_marlin, prepare_moe_fp4_layer_for_marlin)
|
||||
@@ -430,6 +430,7 @@ class ModelOptFp8MoEMethod(FusedMoEMethodBase):
|
||||
layer.w13_weight.data = swap_w13_to_w31(layer.w13_weight.data)
|
||||
rotate_flashinfer_fp8_moe_weights(layer.w13_weight,
|
||||
layer.w2_weight)
|
||||
register_moe_scaling_factors(layer)
|
||||
|
||||
def apply(
|
||||
self,
|
||||
|
||||
Reference in New Issue
Block a user