[ROCm][CI] fix fp8 range in vit_fp8_quant (#46410)
Signed-off-by: Divakar Verma <divakar.verma@amd.com> Signed-off-by: Divakar Verma <137818590+divakar-amd@users.noreply.github.com> Co-authored-by: Andreas Karatzas <akaratza@amd.com> Co-authored-by: Wentao Ye <44945378+yewentao256@users.noreply.github.com>
Divakar Verma committed
Jun 23, 2026 at 12:34 UTC
84586c9acc0236fcee93fbbbb796ef50655d9484
1 file changed
+4
-2
tests/kernels/core/test_vit_fp8_quant.py
+4
-2
@@ -5,6 +5,9 @@
5
import pytest
6
import torch
7
8
+from vllm.model_executor.layers.quantization.utils.quant_utils import (
9
+ get_fp8_min_max,
10
+)
11
from vllm.platforms import current_platform
12
from vllm.triton_utils import HAS_TRITON
13
@@ -24,8 +27,7 @@ def _naive_fp8_quantize(
27
) -> torch.Tensor:
28
"""Reference FP8 quantization in PyTorch."""
29
fp8_dtype = current_platform.fp8_dtype()
27
- fp8_max = torch.finfo(fp8_dtype).max
28
- fp8_min = -fp8_max
30
+ fp8_min, fp8_max = get_fp8_min_max()
31
32
x = tensor.float()
33
if not skip_scale: