A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 8, 12:30 UTC
Threshold fix wvSplitk for occasional CI fails (#34013)
Signed-off-by: Hashem Hashemi <hashem.hashemi@amd.com>
Hashem Hashemi committed
Feb 10, 2026 at 19:59 UTC
1b3540e6c6d3833118d448c3246434de1a60e558
1 file changed
+5
-2
tests/kernels/quantization/test_rocm_skinny_gemms.py
+5
-2
@@ -270,6 +270,9 @@ def test_rocm_wvsplitk_fp8_kernel(
270
out = ops.wvSplitKQ(B, A, dtype, scale_a, scale_b, get_cu_count(), BIAS)
271
272
if xnorm:
273
- assert torch.allclose(out, ref_out, atol=1e-3, rtol=1e-8)
273
+ torch.testing.assert_close(out, ref_out, atol=1e-3, rtol=1e-8)
274
+ elif k >= 32 * 1024:
275
+ # wider pytrch thresh for large-K & no xnorm
276
+ torch.testing.assert_close(out, ref_out, atol=0.07, rtol=5e-2)
277
else:
275
- assert torch.allclose(out, ref_out, 0.01)
278
+ torch.testing.assert_close(out, ref_out, atol=0.01, rtol=0.01)