Skip to content

Aiter MHC fix and keep DSv4 ATOM conc1 #1202

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
Oseltamivir merged 3 commits into from
Apr 28, 2026
+331 −41
Merged

Aiter MHC fix and keep DSv4 ATOM conc1 #1202

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
952c923
fix: use aiter mhc device fix for dsv4 atom
Oseltamivir Apr 27, 2026
0d94067
fix: restore DSv4 ATOM aiter mhc + perf stack to CI-proven state
Oseltamivir Apr 28, 2026
ce87aa8
Merge branch 'main' into codex/dsv4-atom-aiter-mhc-fix
Oseltamivir Apr 28, 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
18 changes: 6 additions & 12 deletions .github/configs/amd-master.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1515,12 +1515,12 @@ dsv4-fp8-mi355x-vllm:
- { tp: 8, conc-start: 1, conc-end: 1 }

# Day-0 single-sequence marker for DeepSeek-V4 on ATOM (ROCm/ATOM#650).
# PR1 of the ATOM DSv4 series — single-sequence only (kv_cache[:1,...]
# hardcode), --enforce-eager required, ATOM_USE_TRITON_MOE=1 required on
# gfx950. Image is the standard atom0.1.2.post MI355X base (matching
# qwen3.5-fp8-mi355x-atom); the DSv4 PR is overlaid at runtime by
# benchmarks/single_node/dsv4_fp4_mi355x_atom.sh at a pinned SHA. Sweep
# will expand once ATOM PR3 (multi-request) and PR4 (CUDAGraph) land.
# PR1 of the ATOM DSv4 series still uses torch sparse-attention fallbacks
# that OOM once warmup/prefill batches multiple requests; keep CONC=1 until
# the AITER sparse-attention kernel / multi-request path lands upstream.
# --enforce-eager and ATOM_USE_TRITON_MOE=1 are required on gfx950. Image is
# the standard atom0.1.2.post MI355X base (matching qwen3.5-fp8-mi355x-atom);
# the DSv4 PR is overlaid at runtime by dsv4_fp4_mi355x_atom.sh at a pinned SHA.
dsv4-fp4-mi355x-atom:
image: rocm/atom:rocm7.2.2_ubuntu24.04_py3.12_pytorch_release_2.10.0_atom0.1.2.post
model: deepseek-ai/DeepSeek-V4-Pro
Expand All @@ -1534,13 +1534,7 @@ dsv4-fp4-mi355x-atom:
osl: 1024
search-space:
- { tp: 8, ep: 1, conc-start: 1, conc-end: 1 }
- { tp: 8, ep: 1, conc-start: 4, conc-end: 4 }
- { tp: 8, ep: 1, conc-start: 16, conc-end: 16 }
- { tp: 8, ep: 1, conc-start: 32, conc-end: 32 }
- isl: 8192
osl: 1024
search-space:
- { tp: 8, ep: 1, conc-start: 1, conc-end: 1 }
- { tp: 8, ep: 1, conc-start: 4, conc-end: 4 }
- { tp: 8, ep: 1, conc-start: 16, conc-end: 16 }
- { tp: 8, ep: 1, conc-start: 32, conc-end: 32 }
Loading
Loading