From 3c2685645ed602e3db8010cd62d88d3130115f06 Mon Sep 17 00:00:00 2001 From: rasmith Date: Wed, 14 Jan 2026 23:04:34 -0600 Subject: [PATCH] [CI][AMD][Quantization][BugFix] Fix fp8 max in quant_utils.py and update test_fp8_quant.::test_static_fp8_quant_group_2d to use correct fp8 dtype and adjust atol/rtol (#32201) Signed-off-by: Randall Smith --- tests/kernels/quantization/test_fp8_quant.py | 4 ++-- vllm/model_executor/layers/quantization/utils/quant_utils.py | 3 ++- 2 files changed, 4 insertions(+), 3 deletions(-) diff --git a/tests/kernels/quantization/test_fp8_quant.py b/tests/kernels/quantization/test_fp8_quant.py index 4572c28de5..325665c488 100644 --- a/tests/kernels/quantization/test_fp8_quant.py +++ b/tests/kernels/quantization/test_fp8_quant.py @@ -178,12 +178,12 @@ def test_static_fp8_quant_group_2d( x = torch.rand(num_tokens, hidden_size, dtype=dtype, device="cuda") ref_out, scale = scaled_quantize( - x, group_shape, FP8_DTYPE, compute_dtype=torch.float32 + x, group_shape, current_platform.fp8_dtype(), compute_dtype=torch.float32 ) ops_out, ops_scale = ops.scaled_fp8_quant(x, scale=scale, group_shape=group_shape) torch.testing.assert_close(scale, ops_scale) - torch.testing.assert_close(ref_out.float(), ops_out.float(), rtol=0.12, atol=0.0) + torch.testing.assert_close(ref_out.float(), ops_out.float(), rtol=1.2e-1, atol=1e-5) opcheck_fp8_quant(ops_out, x, scale=scale) diff --git a/vllm/model_executor/layers/quantization/utils/quant_utils.py b/vllm/model_executor/layers/quantization/utils/quant_utils.py index 08262ed1a3..c625768e97 100644 --- a/vllm/model_executor/layers/quantization/utils/quant_utils.py +++ b/vllm/model_executor/layers/quantization/utils/quant_utils.py @@ -221,7 +221,8 @@ def scaled_quantize( # Compute scales min_val, max_val = x_blkd_permd.aminmax(dim=-1) amax = torch.maximum(min_val.abs(), max_val.abs()).clamp(min=1e-12) - scale = finfo.max / amax + _, fp8_max = get_fp8_min_max() + scale = fp8_max / amax # Apply scale and convert form: # (BLK_M, BLK_N, BLOCK_SIZE_M * BLOCK_SIZE_N) to (M, N)