[Perf] Fix and reapply move apply w8a8 block fp8 linear to class (#25696)

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
2025-10-02 19:35:13 +00:00
committed by GitHub
co-authored by Luka Govedič Michael Goin Luka Govedič
parent 3d5f1c8640
commit 502640c3f9
13 changed files with 412 additions and 200 deletions
@@ -17,7 +17,7 @@ from weight_shapes import WEIGHT_SHAPES
from vllm import _custom_ops as ops
from vllm.model_executor.layers.quantization.utils.fp8_utils import (
w8a8_block_fp8_matmul,
w8a8_triton_block_scaled_mm,
)
from vllm.utils import FlexibleArgumentParser, cdiv
@@ -158,7 +158,7 @@ def bench_fp8(
"cutlass_fp8_fp8_fp16_scaled_mm_bias": lambda: ops.cutlass_scaled_mm(
a, b, scale_a, scale_b, torch.float16, bias.to(dtype=torch.float16)
),
"triton_fp8_fp8_fp16_scaled_mm_blockwise": lambda: w8a8_block_fp8_matmul(
"triton_fp8_fp8_fp16_scaled_mm_blockwise": lambda: w8a8_triton_block_scaled_mm(
a_cont, b.t(), block_scale_a, block_scale_b.t(), (128, 128)
),
"cutlass_fp8_fp8_fp16_scaled_mm_blockwise": lambda: ops.cutlass_scaled_mm(
@@ -9,7 +9,7 @@ import torch
from vllm import _custom_ops as ops
from vllm.model_executor.layers.quantization.utils.fp8_utils import (
per_token_group_quant_fp8,
w8a8_block_fp8_matmul,
w8a8_triton_block_scaled_mm,
)
from vllm.triton_utils import triton
from vllm.utils.deep_gemm import (
@@ -63,7 +63,7 @@ def benchmark_shape(m: int,
# === vLLM Triton Implementation ===
def vllm_triton_gemm():
return w8a8_block_fp8_matmul(A_vllm,
return w8a8_triton_block_scaled_mm(A_vllm,
B_vllm,
A_scale_vllm,
B_scale_vllm,