"""Vast.ai provider: verified-datacenter single-GPU instances (REST only). The Vast substrate rents a single-GPU instance from a verified-datacenter offer, ships a self-contained bootstrap through the onstart script, and detects completion purely from the worker's HF artifacts (no inbound network, no serverless queue). It implements the SAME ``base.Provider`` interface behind the SAME module layout as RunPod, so the orchestrator/allocator treat the two interchangeably. ``PROVIDER`` is the ``base.Provider`` implementation the registry hands out. """ from __future__ import annotations import os from typing import Any from flash.providers.base import GpuClass, JobHandle, PollResult, Provider class VastProvider: """``base.Provider`` for the Vast.ai verified-datacenter substrate.""" name = "vast" def is_configured(self) -> bool: from flash.providers.vast.auth import load_api_key # Vast needs its operator key AND a live network path: it is a live-market # substrate (offer search), so FLASH_SKIP_NET (offline/CI) disables Vast # entirely; offline allocation then degrades deterministically to RunPod's # static catalog. if os.environ.get("FLASH_SKIP_NET"): return False return load_api_key() is not None def preflight(self, require_hf: bool = True) -> list[str]: from flash.providers.vast.preflight import missing_credentials return missing_credentials(require_hf=require_hf) def gpu_classes(self) -> list[GpuClass]: from flash.providers.vast.gpus import gpu_classes return gpu_classes() def hourly_rate(self, gpu: str) -> float: from flash.providers.vast.pricing import hourly_rate return hourly_rate(gpu) def submit_run( self, spec, seed: int, *, log: Any = None, on_handle: Any = None, attempt: int = 0, offers: Any = None, exclude_machine_ids: Any = frozenset(), ) -> PollResult: from flash.providers.vast.jobs import submit_run_vast return submit_run_vast( spec, seed, log=log, on_handle=on_handle, attempt=attempt, offers=offers, exclude_machine_ids=exclude_machine_ids, ) def poll(self, handle: JobHandle, spec, seed: int, *, log: Any = None) -> PollResult: from flash.providers.runpod.jobs import make_hf_heartbeat_reader from flash.providers.vast.jobs import VastJobHandle, poll_vast_job hf_repo = spec.train.hf_repo prefix = f"{spec.phase}/{spec.run_id}/seed{seed}" reader = make_hf_heartbeat_reader(hf_repo, prefix) if hf_repo else None vh = VastJobHandle.from_dict(handle.to_dict()) if log is not None: print(f"attaching: vast instance={vh.instance_id}", file=log, flush=True) # Reattach must apply the SAME stall tuning + wall-cap deadline as submit_run_vast # (see jobs.py), mirroring RunPod's reattach (runpod/__init__.py). Vast has no # server-side execution timeout, so a recovered run that dropped the client-side # deadline could bill unbounded. stall = 1500.0 deadline = max(60, int(spec.gpu.max_wall_seconds)) + 1800 return poll_vast_job( vh, spec, seed, log=log, heartbeat_reader=reader, stall_after_s=stall, deadline_s=deadline, ) def cancel(self, handle: JobHandle) -> None: from flash.providers.vast.jobs import cancel cancel(handle.to_dict()) def destroy(self, handle: JobHandle) -> None: from flash.providers.vast import api as vast_api d = handle.to_dict() if d.get("instance_id"): vast_api.destroy_instance(int(d["instance_id"])) def gc(self, spec) -> None: from flash.providers.vast.jobs import destroy_run_instances destroy_run_instances(spec.run_id) def sweep_orphans(self, active_labels: set[str] | None = None) -> list[int]: """Vast-only crash-recovery sweep (called via the provider object at startup).""" from flash.providers.vast.jobs import sweep_orphans return sweep_orphans(active_labels=active_labels) PROVIDER: Provider = VastProvider()