freesolo-flash 0.2.12

Flash — managed LoRA post-training (SFT/GRPO) for verifiers environments, driven by the `flash` CLI

Flash

Managed LoRA post-training service: SFT and GRPO on managed GPUs across multiple providers — RunPod Flash (serverless queue; RTX 4090/5090 classes) and Vast.ai (rented verified-datacenter instances; L40S / RTX Pro 4000 / A100 classes). The allocator picks the cheapest GPU class that fits the run across both providers.

Scope

• flash train <cfg.toml> / control-plane POST /runs — submit a training job; one dedicated GPU per run, supervised server-side (stall watchdog, bounded auto-retry resuming from the last streamed checkpoint, endpoint GC).
• flash deploy (scale-to-zero or always-on), flash chat — serving for trained adapters.
• Verifiers-only environments. Every run names a Prime Intellect verifiers environment by its published Hub slug ([environment] id = "owner/name"). Scaffold a local env, publish it with flash env push, then reference it by id. The worker wraps it via flash/envs/adapter.py. There are no built-in task environments and no freesolo bridge. Single-turn environments are fully supported (SFT/GRPO/eval).

Layout

• flash/catalog.py — curated model catalog (Qwen3 dense supported tier; Qwen3.5/3.6 experimental tier) + model_policy = "allow" VRAM-fit check + each model's thinking capability (opt-in reasoning mode thinking = true)
• flash/schema.py, flash/spec.py — TOML → JobSpec
• flash/runner.py — server-side run supervisor (durable job handle, retries, cost guard, endpoint GC)
• flash/providers/ — RunPod Flash + Vast.ai provider subtrees (pricing, gpus, durable submit/poll, preflight) behind one base.Provider protocol, with a cross-provider allocator.py that picks the cheapest fitting class
• flash/engine/ — the on-GPU worker (TRL + colocated vLLM rollouts) and the shared recipe; SFT targets and RL rewards route through the active environment (task-specific grading lives with its example, not in the engine)
• flash/envs/ — environment machinery: registry and the adapter that wraps Prime Intellect / Hub verifiers environments onto the worker's interface
• flash lab setup / flash env init — scaffold a starter local verifiers env and a ready-to-run config to start from
• flash/serve/, flash/server/ — adapter serving and the FastAPI control plane (run operator-side via the separate flash-server command)
• flash/mcp/ — stdio MCP bridge for coding agents
• Dockerfile — the control-plane image (used by the repo docker-compose)
• tests/ — pytest suite (CPU-only; offline-by-default, no GPU/network)

Local commands

cd flash
uv sync --extra server
uv run pytest                           # CPU tests (offline-by-default, no GPU/network)
uv run ruff check . && uv run ruff format .
uv run flash --help
uv run flash-server                      # control plane (operator-side, run once)

The control plane owns provider credentials: RUNPOD_API_KEY is always required (RunPod is the default substrate), VAST_API_KEY is opt-in (only checked when set), plus the shared HF_TOKEN. The artifact repo is per-run (the run TOML's [train] hf_repo), not an operator-wide env var. Clients authenticate with their freesolo API key (flash login).