[Perf][fp8] Use CustomOp abstraction for fp8 quant for better perf (#19830)

Signed-off-by: Luka Govedic <lgovedic@redhat.com>
Co-authored-by: mgoin <mgoin64@gmail.com>
This commit is contained in:
Luka Govedič
2025-07-11 04:56:28 +00:00
committed by GitHub
co-authored by mgoin
parent 35514b682a
commit 31d5c1797f
18 changed files with 368 additions and 104 deletions
@@ -22,7 +22,7 @@ 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)
from vllm.model_executor.layers.quantization.utils.quant_utils import (
is_layer_skipped)
GroupShape, is_layer_skipped)
from vllm.model_executor.layers.quantization.utils.w8a8_utils import (
Fp8LinearOp, requantize_with_max_scale)
from vllm.model_executor.parameter import (ModelWeightParameter,
@@ -102,7 +102,8 @@ class ModelOptFp8LinearMethod(LinearMethodBase):
def __init__(self, quant_config: ModelOptFp8Config):
self.quant_config = quant_config
self.fp8_linear = Fp8LinearOp()
self.fp8_linear = Fp8LinearOp(
act_quant_static=True, act_quant_group_shape=GroupShape.PER_TENSOR)
def create_weights(
self,