[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
@@ -16,7 +16,7 @@ from vllm.model_executor.layers.quantization.base_config import (
from vllm.model_executor.layers.quantization.utils.marlin_utils_fp8 import (
apply_fp8_marlin_linear, prepare_fp8_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, maybe_create_device_identity, normalize_e4m3fn_to_e4m3fnuz)
from vllm.model_executor.parameter import (ChannelQuantScaleParameter,
@@ -37,7 +37,6 @@ class FBGEMMFp8Config(QuantizationConfig):
# For GPUs that lack FP8 hardware support, we can leverage the Marlin
# kernel for fast weight-only FP8 quantization
self.use_marlin = not current_platform.has_device_capability(89)
self.fp8_linear = Fp8LinearOp()
@classmethod
def get_name(cls) -> QuantizationMethods:
@@ -76,7 +75,8 @@ class FBGEMMFp8LinearMethod(LinearMethodBase):
def __init__(self, quant_config: FBGEMMFp8Config):
self.quant_config = quant_config
self.fp8_linear = Fp8LinearOp(use_per_token_if_dynamic=True)
self.fp8_linear = Fp8LinearOp(
act_quant_static=False, act_quant_group_shape=GroupShape.PER_TOKEN)
self.out_dtype = torch.get_default_dtype()
def create_weights(