biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

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

Browse Source 
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 21:35:13 +02:00
committed by GitHub
parent 3d5f1c8640
commit 502640c3f9
13 changed files with 412 additions and 200 deletions
Download Patch File Download Diff File Expand all files Collapse all files

5
tests/kernels/quantization/test_block_fp8.py
View File

@@ -11,7 +11,7 @@ from tests.kernels.quant_utils import (native_per_token_group_quant_fp8,
native_w8a8_block_matmul)
from vllm.config import VllmConfig
from vllm.model_executor.layers.quantization.utils.fp8_utils import (
cutlass_scaled_mm, per_token_group_quant_fp8, w8a8_block_fp8_matmul)
cutlass_scaled_mm, per_token_group_quant_fp8, w8a8_triton_block_scaled_mm)
from vllm.platforms import current_platform
from vllm.utils import has_deep_gemm
from vllm.utils.deep_gemm import (fp8_gemm_nt,
@@ -91,7 +91,8 @@ def test_w8a8_block_fp8_matmul(M, N, K, block_size, out_dtype, seed):
ref_out = native_w8a8_block_matmul(A_fp8, B_fp8, As, Bs, block_size,
out_dtype)
out = w8a8_block_fp8_matmul(A_fp8, B_fp8, As, Bs, block_size, out_dtype)
out = w8a8_triton_block_scaled_mm(A_fp8, B_fp8, As, Bs, block_size,
out_dtype)
rel_diff = (torch.mean(
torch.abs(out.to(torch.float32) - ref_out.to(torch.float32))) /