xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2025-12-27 08:35:16 +08:00
Code Issues Packages Projects Releases Wiki Activity

[CI/Build][Kernel][BugFix][AMD] Fix per_token_group_quant_fp8 to use correct fp8 min/max values and update atol/rtol in test_quantfp8_group_functionality (#30292)

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 17:41:56 -06:00 committed by GitHub
parent b4039c08b5
commit 08f8a5627e
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
2 changed files with 7 additions and 4 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
tests/kernels/quantization/test_fp8_quant_group.py
View File

@ -62,7 +62,7 @@ def test_quantfp8_group_functionality(
assert scales_col.stride(1) == batch_size
# Test column-major scales consistency
assert torch.allclose(scales_col, scales_native, rtol=1e-9, atol=1e-8)
torch.testing.assert_close(scales_col, scales_native, rtol=1e-9, atol=1e-8)
# 3. Test CUDA implementation (only for divisible dimensions)
if is_divisible:
@ -71,7 +71,7 @@ def test_quantfp8_group_functionality(
assert scales_cuda.shape == (batch_size, expected_num_groups)
# Verify CUDA/native consistency
assert torch.allclose(scales_cuda, scales_native, rtol=1e-9, atol=1e-8)
torch.testing.assert_close(scales_cuda, scales_native, rtol=2e-7, atol=2e-8)
# Quantized values should mostly match
diff_count = (x_quant_cuda != x_quant_native).sum().item()

7
vllm/model_executor/layers/quantization/utils/fp8_utils.py
View File

@ -762,9 +762,12 @@ def per_token_group_quant_fp8(
)
assert x.stride(-1) == 1, "`x` groups must be contiguous"
# Using the default value (240.0) from pytorch will cause accuracy
# issue on dynamic quantization models. Here use 224.0 for fnuz on ROCm
# platforms that use the torch.float8_e4mefnuz dtype.
finfo = torch.finfo(dtype)
fp8_min = finfo.min
fp8_max = finfo.max
fp8_min = -224.0 if current_platform.is_fp8_fnuz() else finfo.min
fp8_max = 224.0 if current_platform.is_fp8_fnuz() else finfo.max
assert out_q is None or out_q.shape == x.shape
x_q = out_q