1. Export route: stop_training() only signals the subprocess — wait up to
30s for it to actually exit before loading the export checkpoint, avoiding
a GPU memory race.
2. Training reset: clear _should_stop so /api/train/status returns phase=idle
instead of staying stuck on phase=stopped after a user-triggered stop.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Two overlapping /chat/completions requests could both read from the shared
resp_queue, consuming and dropping each other's token events. Replace the
request_id filtering (which silently dropped non-matching messages) with a
threading.Lock that serializes generation — correct for single-GPU inference.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Models like GLM-4.7-Flash have architectures (glm4_moe_lite) that
AutoConfig in the main process (transformers 4.57.x) can't recognize.
Instead of a raw config.json workaround, run the AutoConfig check in
a subprocess with .venv_t5/ activated — same pattern as training and
inference workers. This is more robust and consistent.
AutoConfig.from_pretrained() fails for models needing transformers 5.x
(e.g. glm4_moe_lite) when running with 4.57.x. Add a raw config.json
fallback that bypasses AutoConfig's architecture registry — fetches
config.json directly from local path or HuggingFace Hub and checks
for vision indicators without needing the architecture to be registered.
All version switching now uses .venv_t5/ (pre-installed by setup.sh).
The old .venv_overlay/ with runtime pip installs is removed.
ensure_transformers_version() (used only by export) now does a
lightweight sys.path swap instead of pip installing at runtime.
Reusing a subprocess after unsloth patches torch internals causes
inspect.getsource() failures when loading a different model type.
Each load now gets a clean Python interpreter.
Replaces cmd_queue-based cancel polling with a shared mp.Event.
Fixes two issues:
- Loading a new model while generating no longer hangs (cancel is instant)
- Subprocess shuts down cleanly after explicit stop generation
Inference now runs in a persistent subprocess, solving the same
transformers version-switching problem that was fixed for training.
The subprocess stays alive between requests (model in GPU memory)
and is only restarted when switching transformers versions.
New files:
- core/inference/worker.py: subprocess entry point with command loop
- core/inference/orchestrator.py: parent-side proxy with same API
Modified:
- core/inference/__init__.py: exports orchestrator as default backend
- routes/inference.py: removed in-process ensure_transformers_version()
