A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 14, 05:30 UTC
[Benchmarks] Fix imports in FP8 tuning script (#26407)
Signed-off-by: Lukas Geiger <lukas.geiger94@gmail.com>
Lukas Geiger committed
Oct 8, 2025 at 17:31 UTC
6273fe8d3d72d26b7a60e0d60d5c378273156e9e
1 file changed
+2
-2
benchmarks/kernels/benchmark_w8a8_block_fp8.py
+2
-2
@@ -14,7 +14,7 @@ import torch
14
from tqdm import tqdm
15
16
from vllm.model_executor.layers.quantization.utils.fp8_utils import (
17
- _w8a8_block_fp8_matmul,
17
+ _w8a8_triton_block_scaled_mm,
18
)
19
from vllm.platforms import current_platform
20
from vllm.triton_utils import triton
@@ -83,7 +83,7 @@ def w8a8_block_matmul(
83
)
84
85
if A.dtype == torch.float8_e4m3fn:
86
- kernel = _w8a8_block_fp8_matmul
86
+ kernel = _w8a8_triton_block_scaled_mm
87
else:
88
raise RuntimeError("Currently, only support tune w8a8 block fp8 kernel.")
89