Support FP8 block quant for CompressedTensorsW8A16Fp8 (#33280)

Signed-off-by: mgoin <mgoin64@gmail.com>
This commit is contained in:
Michael Goin
2026-01-30 11:15:20 -05:00
committed by GitHub
parent f857a03f6b
commit fd0e377244
4 changed files with 74 additions and 64 deletions
@@ -29,7 +29,7 @@ from vllm.model_executor.parameter import (
ChannelQuantScaleParameter,
PerTensorScaleParameter,
)
from vllm.model_executor.utils import replace_parameter
from vllm.model_executor.utils import replace_parameter, set_weight_attrs
from vllm.platforms import current_platform
from vllm.triton_utils import tl, triton
from vllm.utils.deep_gemm import (
@@ -1520,6 +1520,7 @@ def create_fp8_scale_parameter(
raise ValueError(f"Unknown parameter type: {parameter_type}")
scale[:] = torch.finfo(torch.float32).min
set_weight_attrs(scale, {"scale_type": "weight_scale"})
return scale