[ROCm][Test] Fix test_per_token_group_quant_fp8 tolerance for 1-ULP FP8 rounding on gfx950 (#46944)
Signed-off-by: Spandan Tiwari <sptiwari@amd.com>
This commit is contained in:
@@ -11,6 +11,7 @@ from tests.kernels.quant_utils import (
|
||||
native_per_token_group_quant_fp8,
|
||||
native_w8a8_block_matmul,
|
||||
)
|
||||
from tests.kernels.utils import fp8_ulp_distance
|
||||
from vllm.config import VllmConfig
|
||||
from vllm.model_executor.kernels.linear.scaled_mm.cutlass import cutlass_scaled_mm
|
||||
from vllm.model_executor.layers.quantization.utils.fp8_utils import (
|
||||
@@ -93,7 +94,24 @@ def test_per_token_group_quant_fp8(
|
||||
tma_aligned_scales=tma_aligned_scales,
|
||||
)
|
||||
|
||||
assert torch.allclose(out.to(torch.float32), ref_out.to(torch.float32), rtol=0.15)
|
||||
if current_platform.is_rocm():
|
||||
# On gfx950 the Triton and PyTorch FP8 kernels can round in opposite
|
||||
# directions when an element lands at the midpoint between two adjacent
|
||||
# e4m3fn values (1-ULP tie-breaking). Verify: (1) no element is more
|
||||
# than 1 FP8 ULP away, and (2) fewer than 0.05% of elements have any
|
||||
# mismatch. Observed worst case across all parameter combos: 0.049%,
|
||||
# max ULP = 1.
|
||||
ulp = fp8_ulp_distance(out, ref_out)
|
||||
assert (ulp <= 1).all(), (
|
||||
f"FP8 mismatch > 1 ULP: {int((ulp > 1).sum())} elements"
|
||||
)
|
||||
assert float((ulp > 0).float().mean()) < 5e-4, (
|
||||
f"Too many 1-ULP mismatches: {int((ulp > 0).sum())}/{ulp.numel()}"
|
||||
)
|
||||
else:
|
||||
assert torch.allclose(
|
||||
out.to(torch.float32), ref_out.to(torch.float32), rtol=0.15
|
||||
)
|
||||
assert torch.allclose(scale, ref_scale)
|
||||
|
||||
if column_major_scales:
|
||||
|
||||
Reference in New Issue
Block a user