ATOM v0.1.5 — paired with AITER v0.1.16.
Paired container
rocm/atom-dev:atom0.1.5-aiter0.1.16 (ROCm 7.2.4 base, gfx942;gfx950, aiter @ 0730b33f)
Key fixes (vs v0.1.4)
- MiniMax-M2.5 torch.compile: restore the
qkv.shape[0] <= 256shape guard on the fused qknorm-allreduce path (#1299). Without it the fused custom collective runs unconditionally at TP>1 inside the compiled attention forward, forcing a top-level graph break that re-enters VllmBackend and hard-fails engine init (VllmBackend can only be called once). - Compile-cache frozen-path guard:
backends.pynow skips any non-existent traced path (<frozen posixpath>, etc.) instead of only<string>/<frozen os>(#1299). - fastapi cap
>=0.115,<0.137to avoid prometheus-fastapi-instrumentator crash on serve startup (fastapi >=0.137 is now on PyPI) (#1221).
Validation
mi355-gpu-15, GSM8K 3-shot flexible-extract on the released container:
- MiniMax-M2.5 0.9325, DeepSeek-R1-0528 0.9484 (smoke)
- Full gate (pilot, same fixes): DSR1 0.952 / GLM5 0.947 / Kimi 0.940 / MiniMax 3x mean 0.929 / Qwen3 3x mean 0.871