[ROCm] Fix CI runtime and tests for MI355 DPX (#58244)

Signed-off-by: Andreas Karatzas <Andreas.Karatzas@amd.com>
Co-authored-by: Mahesh Kunreddi <mahesh.kunreddi@amd.com>
Co-authored-by: Claude <noreply@anthropic.com>
Co-authored-by: Andreas Karatzas <Andreas.Karatzas@amd.com>
Co-authored-by: Codex <noreply@openai.com>
Co-authored-by: Andreas Karatzas <akaratza@amd.com>
This commit is contained in:
mkunredd
2026-09-24 23:15:11 -05:00
committed by GitHub
co-authored by Mahesh Kunreddi Claude Andreas Karatzas Codex Andreas Karatzas
parent 44257c2e35
commit 4013f3ad6c
12 changed files with 175 additions and 47 deletions
@@ -614,11 +614,13 @@ def test_aiter_fp4_gemm_skinny_shapes(M, N, K):
out = gemm_afp4wfp4(A_fp4, B_fp4, A_scale, B_scale)
assert out.shape == (M, N)
assert out.dtype == torch.bfloat16
assert not torch.any(torch.isnan(out))
# Accumulate in FP32, then match the GEMM's BF16 output rounding.
A_dq = quant_dequant_mxfp4(A)
B_dq = quant_dequant_mxfp4(B)
ref = torch.matmul(A_dq.float(), B_dq.t().float())
ref = torch.matmul(A_dq.float(), B_dq.t().float()).to(torch.bfloat16).float()
_print_close_stats(
f"skinny_gemm M={M} N={N} K={K}",
@@ -132,7 +132,8 @@ M_FACTORS_WVSPLITKRC = [128, 128 + 16, 256, 256 + 16, 640, 640 + 16]
NKM_FACTORS_WVSPLITKRC_LARGE_K = [
(128, 6144, 128),
(96, 8192, 128),
(128, 12288, 128),
# Keep 24 K-shards (batched readback) within the DPX budget of 128 CUs.
(128, 12288, 64),
(32, 12288, 128),
]