[ROCm] Use vLLM's fp8 quant max in AITER hipBLASLt accuracy test (#46176)
Signed-off-by: Djordje Ramic <djoramic@amd.com>
djramic committed
Jun 19, 2026 at 20:24 UTC
dec860fb19fcd8a39c62a2204c5939feb4781f14
1 file changed
+2
-1
tests/rocm/aiter/test_aiter_hipb_mm_linear_kernel.py
+2
-1
@@ -18,6 +18,7 @@ from vllm.model_executor.kernels.linear.scaled_mm.ScaledMMLinearKernel import (
18
FP8ScaledMMLinearLayerConfig,
19
)
20
from vllm.model_executor.layers.quantization.utils.quant_utils import (
21
+ get_fp8_min_max,
22
kFp8DynamicTokenSym,
23
kFp8StaticChannelSym,
24
kFp8StaticTensorSym,
@@ -309,7 +310,7 @@ def test_hipb_mm_kernel_forward_accuracy(enable_hipb_mm_kernel):
310
_check_bpreshuffle_runtime_support(weight_shape, num_tokens=num_tokens)
311
312
fp8_dtype = current_platform.fp8_dtype()
312
- fp8_max = torch.finfo(fp8_dtype).max
313
+ fp8_max = get_fp8_min_max()[1]
314
device = torch.device("cuda")
315
316
# Build a bf16 weight and quantize per output channel (one scale per row).