From 273a04964ed9d1e389c4a7cb9d660e8294782f8d Mon Sep 17 00:00:00 2001 From: josh-schell-amd Date: Wed, 14 Oct 2026 18:50:24 -0500 Subject: [PATCH] [ROCm] Bump AITER to v0.1.25 and relax the DSv4.1 mono MoE tolerance (#61697) Signed-off-by: josh schell Co-authored-by: Claude Opus 5.5 (1M context) --- docker/Dockerfile.rock_base | 2 +- docker/Dockerfile.rocm_base | 2 +- tests/models/test_dsv41_mono_numerics.py | 8 ++++++-- 3 files changed, 8 insertions(+), 4 deletions(-) diff --git a/docker/Dockerfile.rock_base b/docker/Dockerfile.rock_base index 40100a1bbf..57bf6ae6fe 100644 --- a/docker/Dockerfile.rock_base +++ b/docker/Dockerfile.rock_base @@ -8,7 +8,7 @@ ARG TRITON_VERSION=3.8.0+git669b31ac ARG TRITON_REPO="https://github.com/ROCm/triton.git" ARG FA_BRANCH="0e60e394" ARG FA_REPO="https://github.com/Dao-AILab/flash-attention.git" -ARG AITER_BRANCH="v0.1.24.post1" +ARG AITER_BRANCH="v0.1.25" ARG AITER_REPO="https://github.com/ROCm/aiter.git" ARG MORI_BRANCH="v1.2.3.post1" ARG MORI_REPO="https://github.com/ROCm/mori.git" diff --git a/docker/Dockerfile.rocm_base b/docker/Dockerfile.rocm_base index bd8099dd15..33dd0f6b10 100644 --- a/docker/Dockerfile.rocm_base +++ b/docker/Dockerfile.rocm_base @@ -9,7 +9,7 @@ ARG PYTORCH_AUDIO_BRANCH="v2.11.0" ARG PYTORCH_AUDIO_REPO="https://github.com/pytorch/audio.git" ARG FA_BRANCH="0e60e394" ARG FA_REPO="https://github.com/Dao-AILab/flash-attention.git" -ARG AITER_BRANCH="v0.1.24.post1" +ARG AITER_BRANCH="v0.1.25" ARG AITER_REPO="https://github.com/ROCm/aiter.git" ARG MORI_BRANCH="v1.2.3.post1" ARG MORI_REPO="https://github.com/ROCm/mori.git" diff --git a/tests/models/test_dsv41_mono_numerics.py b/tests/models/test_dsv41_mono_numerics.py index f7cd877caf..83308d46b0 100644 --- a/tests/models/test_dsv41_mono_numerics.py +++ b/tests/models/test_dsv41_mono_numerics.py @@ -468,7 +468,10 @@ def _check_layers(rank: int, tp: int, device) -> None: # relative errors about 3x what the kernels show (0.3-0.5%) assert _cos(o[1], res2) > 0.9999 and _rel(o[1], res2) < 0.01, tag assert _cos(normed, xf) > 0.999 and _rel(normed, xf) < 0.012, tag - assert _cos(o[0], same_in) > 0.9999 and _rel(o[0], same_in) < 0.015, tag + # TODO: back to cos > 0.9999, rel < 0.015 once + # https://github.com/ROCm/aiter/issues/6548 is fixed: aiter's fused-FP8 + # stage-1 MoE kernels use another MXFP8 scale rule (1.53% at TP4) + assert _cos(o[0], same_in) > 0.9998 and _rel(o[0], same_in) < 0.02, tag # per token, where both FFN inputs pick the same experts same = (top6(xf) == top6(normed)).all(1) assert same.float().mean() >= 0.5, tag @@ -479,7 +482,8 @@ def _check_layers(rank: int, tp: int, device) -> None: f = runner.ffn(mw, part, res, post, comb, attn_pre) torch.testing.assert_close(f[1], res2, rtol=2**-8, atol=1e-6) torch.testing.assert_close(normed, xf, rtol=2**-7, atol=1e-6) - assert _cos(f[0], out) > 0.9999 and _rel(f[0], out) < 0.015, tag + # TODO: back to 0.9999 / 0.015 with the one above (ROCm/aiter#6548) + assert _cos(f[0], out) > 0.9998 and _rel(f[0], out) < 0.02, tag torch.accelerator.synchronize() whole, ffn = runner._outs(M, res0), runner._outs(M, res)