AITER v0.1.20.dev0

AITER v0.1.20.dev0 — gfx1250 / ROCm 7.14 (pre-release)

Development wheel of amd-aiter for gfx1250 on ROCm 7.14.
Cross-compiled, validated on real gfx1250 hardware. Marked pre-release.

What this is

A .dev0 snapshot pinned to aiter commit d9e5ef7 (PR #4406), the last
commit that runs DeepSeek-V4 on gfx1250 without the FlyDSL lowering crash
(raw.ptr.buffer.load.lds → LLVM ERROR: Do not know how to expand this operator's operand). Later commits re-introduce that crash on gfx1250.

Asset

file	sha256
amd_aiter-0.1.20.dev0+rocm7.14.0.gfx1250-cp312-cp312-linux_x86_64.whl	8484202f16d13b267ff17d78db6c9936683fdce2f86c1840346f52308ce92b6f

• size 88 MB · Python 3.12 · torch 2.11.0+rocm7.14.0 · flydsl 0.2.4

Validation (real gfx1250)

• SGLang · DeepSeek-V4-Flash · tp1 · attention-backend dsv4 · kv fp8_e4m3
• gsm8k 0.925 (Invalid 0.000, server exit 0); server log flydsl count = 0

Install

pip install --force-reinstall --no-deps amd_aiter-0.1.20.dev0+rocm7.14.0.gfx1250-cp312-cp312-linux_x86_64.whl
pip install --force-reinstall --no-deps flydsl==0.2.4
# ensure no source-tree aiter (/app/aiter, /sgl-workspace/aiter) shadows the wheel

Caveats

• Built on glibc 2.35 → not manylinux_2_28; for gfx1250 self-use, not the
  official 6-wheel matrix (which is gfx942;gfx950).
• torch ABI must match the target container's torch exactly.