From dec860fb19fcd8a39c62a2204c5939feb4781f14 Mon Sep 17 00:00:00 2001 From: djramic Date: Fri, 19 Jun 2026 20:24:02 +0200 Subject: [PATCH] [ROCm] Use vLLM's fp8 quant max in AITER hipBLASLt accuracy test (#46176) Signed-off-by: Djordje Ramic --- tests/rocm/aiter/test_aiter_hipb_mm_linear_kernel.py | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) diff --git a/tests/rocm/aiter/test_aiter_hipb_mm_linear_kernel.py b/tests/rocm/aiter/test_aiter_hipb_mm_linear_kernel.py index 92017e95cb..c855d0e819 100644 --- a/tests/rocm/aiter/test_aiter_hipb_mm_linear_kernel.py +++ b/tests/rocm/aiter/test_aiter_hipb_mm_linear_kernel.py @@ -18,6 +18,7 @@ from vllm.model_executor.kernels.linear.scaled_mm.ScaledMMLinearKernel import ( FP8ScaledMMLinearLayerConfig, ) from vllm.model_executor.layers.quantization.utils.quant_utils import ( + get_fp8_min_max, kFp8DynamicTokenSym, kFp8StaticChannelSym, kFp8StaticTensorSym, @@ -309,7 +310,7 @@ def test_hipb_mm_kernel_forward_accuracy(enable_hipb_mm_kernel): _check_bpreshuffle_runtime_support(weight_shape, num_tokens=num_tokens) fp8_dtype = current_platform.fp8_dtype() - fp8_max = torch.finfo(fp8_dtype).max + fp8_max = get_fp8_min_max()[1] device = torch.device("cuda") # Build a bf16 weight and quantize per output channel (one scale per row).