Pull requests / #2914

#2914 fix(setup AMD): put the engine's lib_dirs on LD_LIBRARY_PATH on Linux (the PATH-only #2372 env made strata --families exit 127, so setup falsely refused Qwen3.6)

open · @xxxliqu1dxxx · 0 Kommentare · Auf GitHub

Setup & installServer & APIMulti-GPUAMD / HIPNVIDIA / CUDAWindowsLinux

Beschreibung

   ## Title

   Issue: (no open issue — follow-up to #2372's fix, commit 093108c)

   ## Summary

   Setup's own engine probes (`strata --families`, `strata-device --list-devices`) now get the
   engine's ROCm `lib_dirs` on the dynamic-library path the loader actually uses on the current
   platform: `PATH` on Windows (unchanged from #2372), `LD_LIBRARY_PATH` elsewhere. Fixes the
   AMD Linux false positive where Step 4 of `./setup.sh --backend hip` said "this engine does not
   run Qwen3.6-35B-A3B yet" while the local HIP engine does run it.

   ## What changed

   - `setup.py` — `engine_env` only: on POSIX, `lib_dirs` are prepended to `LD_LIBRARY_PATH`
     (preserving its existing value) instead of `PATH`; on Windows the `PATH` prepend is
     byte-identical to before. Mirrors `serve/server.py`'s existing
     `var = "PATH" if os.name == "nt" else "LD_LIBRARY_PATH"` choice.
   - `tools/test_setup_amd.py` — `EngineLaunchEnv` only:
     - `test_engine_env_puts_rocm_bin_first` → `..._on_windows`, pinned to `os.name = "nt"`.
     - New `test_engine_env_puts_rocm_bin_first_on_ld_library_path`: lib dir first on
       `LD_LIBRARY_PATH` (existing value preserved), `PATH` untouched.
     - `test_families_probe_runs_with_that_env` and `test_device_probe_runs_with_that_env`
       pinned to `os.name = "nt"` (they assert the Windows `PATH` behavior).

   Left alone: the engine binary, the build, the server launch env, the `engine_families`
   fallback, `hip_lib_dirs`, prebuilt Windows engine handling.

   ## Extra Notes

   - How to check: `python tools/test_setup_amd.py` (no GPU needed) — 41 tests OK. Live check on
     an AMD Linux box: `env = setup.engine_env(Path('engine'))` + `engine/strata --families`
     → rc 0, prints `qwen4exp qwen35moe` (previously exit 127).
   - Repro hardware: AMD Radeon RX 9060 series (gfx1200), 16 GB VRAM, 31 GB RAM, local HIP build
     of engine 0.1.45 (BUILD.json source "local-hip").
   - Not tested: multi-GPU, Windows (no Windows box here), NVIDIA/CUDA paths (untouched).

Mehr auf der Site

Links zu Install, Modellen, Releases.