Skip to content

[Fix] [AMD] docker: Repair the AITER build step after the 2bb549102 pin bump #44534

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
HaiShaw merged 2 commits into from
Oct 15, 2026
Merged

[Fix] [AMD] docker: Repair the AITER build step after the 2bb549102 pin bump #44534

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
f213435
[AMD] [Docker] Fix dangling line continuation after the AITER pip ins…
michaelzhang-ai Oct 15, 2026
1d3000d
[AMD] [Docker] Drop the AITER patches already in the 2bb549102 pin
michaelzhang-ai Oct 15, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
20 changes: 1 addition & 19 deletions docker/rocm.Dockerfile
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -622,32 +622,14 @@ RUN pip uninstall -y aiter
# block AITER_COMMIT overrides that predate that rule. The working tree was just
# produced by a fresh `git clone` above, so there are no real user changes to
# preserve.
# cherry pick ROCm/aiter#6042 to unpeel the 64-bit sparse MLA loop below Triton 3.8, may be removed in next aiter upgrade
# apply the DSV4.1 TP2 fp8/fp4 FMoE tuning CSVs from ROCm/aiter#5967, may be removed in next aiter upgrade
# apply fix for v4 fp4 indexer, may be removed in next aiter upgrade
# backport ROCm/aiter#6007 (masked E8M0 scales default to 0) to the group32 GEMM: with other=127 the
# packed small-M kernel reads stale LDS as the masked K-tail scale, and a 0xFF byte turns the tile into NaN;
# may be removed in next aiter upgrade
# The #6042, #5967 and #6007 patches are written against the gfx942/gfx950 AITER pin and
# do not apply to the older gfx1250 one (a conflict in pa_decode_sparse.py, and
# the dsv41 CSVs and the group32 GEMM do not exist there yet), so gfx1250 skips them.
RUN git clone ${AITER_REPO} \
&& cd aiter \
&& git checkout -f ${AITER_COMMIT} \
&& if [ "${GPU_ARCH_LIST}" != "gfx1250" ]; then \
git fetch origin pull/6042/head \
&& git cherry-pick --no-commit 042a289183e5b47e193df6c10160915a6e8b69f5 \
&& git fetch origin pull/5967/head \
&& git diff -U0 89a47b84ac4b576339a50047e290c383c2377389 2bd147293d887c91fd10da10b1dbbf4690844270 -- \
aiter/configs/model_configs/dsv41_fp4_tuned_fmoe.csv \
aiter/configs/model_configs/dsv41_fp4_untuned_fmoe.csv \
| git apply --unidiff-zero \
&& sed -i 's/other=127/other=0/' aiter/ops/triton/_triton_kernels/gemm/basic/gemm_a8w8_blockscale_group32.py; \
fi \
&& sed -i 's/from functools import lru_cache/from functools import cache/' aiter/ops/flydsl/kernels/mqa_logits/pa_mqa_logits_fp4_prefill.py \
&& sed -i 's/@lru_cache(maxsize=32)/@cache/' aiter/ops/flydsl/kernels/mqa_logits/pa_mqa_logits_fp4_prefill.py \
&& git submodule update --init --recursive \
&& pip install -r requirements.txt \
&& pip install -r requirements.txt

# ROCm/aiter#5195's .co with a mask assembled in between the q rows sharing a
# tile, which the grouped dsv4 decode needs; bit-identical at max_seqlen_q=1.
Expand Down
Loading