biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[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)

Browse Source 
Signed-off-by: Randall Smith <ransmith@amd.com>
This commit is contained in:
rasmith
2026-01-14 23:04:34 -06:00
committed by GitHub
parent 773d7073ae
commit 3c2685645e
2 changed files with 4 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
tests/kernels/quantization/test_fp8_quant.py
View File

@@ -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)

3
vllm/model_executor/layers/quantization/utils/quant_utils.py
View File

@@ -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)