[Performance] Move apply_w8a8_block_fp8_linear to an op class (#24666)

Signed-off-by: ElizaWszola <ewszola@redhat.com>
Signed-off-by: ElizaWszola <elizaw.9289@gmail.com>
Signed-off-by: Luka Govedič <lgovedic@redhat.com>
Signed-off-by: Luka Govedič <ProExpertProg@users.noreply.github.com>
Co-authored-by: Luka Govedič <ProExpertProg@users.noreply.github.com>
Co-authored-by: Michael Goin <mgoin64@gmail.com>
Co-authored-by: Luka Govedič <lgovedic@redhat.com>
This commit is contained in:
ElizaWszola
2025-09-23 12:03:10 -07:00
committed by GitHub
co-authored by Luka Govedič Michael Goin Luka Govedič
parent 8c1c81a3de
commit 63400259d0
14 changed files with 341 additions and 201 deletions
+8 -4
View File
@@ -9,7 +9,7 @@ from __future__ import annotations
import functools
import importlib
import os
from typing import Any, Callable, NoReturn
from typing import Any, Callable, NoReturn, Optional
import torch
@@ -172,9 +172,13 @@ def calc_diff(x: torch.Tensor, y: torch.Tensor):
return 1 - sim
def should_use_deepgemm_for_fp8_linear(output_dtype: torch.dtype,
weight: torch.Tensor):
return (is_deep_gemm_supported() and output_dtype == torch.bfloat16
def should_use_deepgemm_for_fp8_linear(
output_dtype: torch.dtype,
weight: torch.Tensor,
supports_deep_gemm: Optional[bool] = None):
if supports_deep_gemm is None:
supports_deep_gemm = is_deep_gemm_supported()
return (supports_deep_gemm and output_dtype == torch.bfloat16
and weight.shape[0] % 128 == 0 and weight.shape[1] % 128 == 0)