[W8A8 Block Linear Refactor][2/N] Remove W8A8Fp8BlockLinearOp and adopt Fp8 block linear kernel selections. (#33892)

Signed-off-by: maral <maralbahari.98@gmail.com>
Signed-off-by: Maral <maralbahari.98@gmail.com>
This commit is contained in:
Maral
2026-04-09 08:50:39 +08:00
committed by GitHub
parent 8332078cfd
commit 2e9034c998
35 changed files with 1710 additions and 904 deletions
@@ -7,6 +7,7 @@ import torch
from torch.nn import Module
from torch.nn.parameter import Parameter
from vllm.config import get_current_vllm_config
from vllm.logger import init_logger
from vllm.model_executor.kernels.linear import (
init_fp8_linear_kernel,
@@ -93,12 +94,7 @@ class FBGEMMFp8LinearMethod(LinearMethodBase):
def __init__(self, quant_config: FBGEMMFp8Config):
self.quant_config = quant_config
self.out_dtype = torch.get_default_dtype()
self.fp8_linear = init_fp8_linear_kernel(
activation_quant_key=kFp8DynamicTokenSym,
weight_quant_key=kFp8StaticTokenSym,
out_dtype=torch.get_default_dtype(),
module_name=self.__class__.__name__,
)
self.input_dtype = get_current_vllm_config().model_config.dtype
def create_weights(
self,
@@ -149,6 +145,15 @@ class FBGEMMFp8LinearMethod(LinearMethodBase):
)
layer.input_scale_ub = input_scale_ub
self.fp8_linear = init_fp8_linear_kernel(
activation_quant_key=kFp8DynamicTokenSym,
weight_quant_key=kFp8StaticTokenSym,
weight_shape=layer.weight.shape,
input_dtype=self.input_dtype,
out_dtype=self.out_dtype,
module_name=self.__class__.__name__,
)
def process_weights_after_loading(self, layer: Module) -> None:
# required by torch.compile
layer.weight_scale = Parameter(layer.weight_scale.data, requires_grad=False)