iHore/vLLM
8
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[CI/Build][AMD] Fix ref_dynamic_per_token_quant reference implementation on ROCm. (#30291)

Browse Source 
Signed-off-by: Randall Smith <ransmith@amd.com>
Co-authored-by: Randall Smith <ransmith@amd.com>
This commit is contained in:
rasmith
2025-12-12 09:30:23 +00:00
committed by GitHub
co-authored by Randall Smith
parent 8f8fda261a
commit 302b2c1eb9
1 changed files with 4 additions and 9 deletions
Download Patch File Download Diff File Expand all files Collapse all files
tests/kernels/quant_utils.py
+4 -9
View File
@@ -30,16 +30,11 @@ def ref_dynamic_per_token_quant(
if quant_dtype == torch.int8
else torch.finfo(quant_dtype)
)
qtype_traits_max = (
ROCM_FP8FNUZ_MAX
if current_platform.is_rocm() and current_platform.is_fp8_fnuz()
else qtype_traits.max
)
qtype_traits_min = (
-ROCM_FP8FNUZ_MAX
if current_platform.is_rocm() and current_platform.is_fp8_fnuz()
else qtype_traits.min
use_fp8fnuz = (
current_platform.is_fp8_fnuz() and quant_dtype == current_platform.fp8_dtype()
)
qtype_traits_max = ROCM_FP8FNUZ_MAX if use_fp8fnuz else qtype_traits.max
qtype_traits_min = -ROCM_FP8FNUZ_MAX if use_fp8fnuz else qtype_traits.min
qtype_max = as_float32_tensor(qtype_traits_max)
s_1 = as_float32_tensor(1.0)
s_512 = as_float32_tensor(512.0)