vastai-workflow（瓦泰工作流程）

Vast.ai Workflow

Overview

Use this skill for provider-level Vast.ai automation only. Keep it project-agnostic and parameter-driven.

Scope Boundaries

• Handle only Vast API workflow logic: offers, create, poll, SSH attach, lifecycle, billing.
• Do not embed project-specific repo paths, branch names, labels, or training commands.
• If the user asks for workload-specific execution (for example Reparo training), switch to the corresponding workload skill after infrastructure provisioning.

Dialog-First Required Fields

Before any create/order call, confirm required runtime fields in dialog. If the user does not provide them, propose defaults and ask for confirmation.

Required fields and default suggestions:

• api_key_source: default VAST_API_KEY environment variable.

- If missing, suggest loading from a local file path the user confirms (for example keys/.vast_env).

• instance_type (offer filter): default gpu_name="RTX 4090".

- Ask whether to lock by exact GPU, VRAM minimum, region, max price, and reliability.

• image: default pytorch/pytorch:latest.
• disk_gb: default 64.
• count: default 1.
• label: do not ask by default.

- Derive from Vast account identity when possible (nickname or email local-part). - Fallback label: vast-user.

Suggested minimal question set:

1. Which action now (offers, create, status, ssh, destroy, billing)?
2. Confirm runtime fields (image, disk_gb, instance filter).
3. Confirm API key source (VAST_API_KEY env vs file path).

If the user approves defaults, proceed without extra questions.

Workflow

1) Preflight

• Check active instances first.
• If any instances are unintentionally active, ask whether to stop/destroy before creating a new one.
• Ensure API key is loaded and never echo it.

2) Find offers

• Query offers (/bundles is commonly reliable in practice).
• Apply user filters plus mandatory rentable=true and rented=false.
• Sort by user objective (price, performance, reliability).
• Use user policy for selection:

- Default: choose cheapest valid offer. - Optional conservative mode: choose second cheapest.

3) Create instance

• Create from a single selected offer (PUT /asks/{id}).
• Use confirmed image, resolved label, and disk_gb.
• Parse and store the returned instance ID.
• If create fails (no_such_ask or already taken), re-query and retry once with next candidate.

Label resolution order:

1. Load profile/account endpoint data and use nickname when present.
2. If nickname missing, use the email local-part when available.
3. If neither field is available, use vast-user.

4) Readiness and access

• Poll instance status until running.
• Add SSH key to account and attach to instance.
• Retry SSH readiness for up to 2 minutes before declaring failure.

5) Lifecycle and billing

• Support start/stop/restart when available.
• Use destroy promptly when requested to prevent costs.
• Verify final state and check usage/invoices when needed.

6) Runtime preflight (optional, when user plans to run workloads)

• Validate runtime inside the instance instead of trusting image names/tags.
• Confirm GPU visibility and basic CUDA readiness before starting long jobs.
• Verify required Python package versions explicitly and report them in logs.
• If workload depends on optional CUDA extensions (for example flash_attn or xformers), check availability first and fail fast on mismatch.
• Prefer explicit environment checks over silent fallbacks to slower backends.

Request Templates

Use explicit, reproducible requests and validate JSON before chaining calls.

curl -sS -L -G "https://console.vast.ai/api/v0/<endpoint>/" \
  --data-urlencode "api_key=$VAST_API_KEY" \
  --data-urlencode "<param>=<value>"

curl -sS -L -X PUT "https://console.vast.ai/api/v0/asks/$OFFER_ID/?api_key=$VAST_API_KEY" \
  -H "Content-Type: application/json" \
  -d "{\"image\":\"$IMAGE\",\"disk\":$DISK_GB,\"label\":\"$LABEL\"}"

Error Handling

• 401/403: API key missing, invalid, or not authorized for requested org/action.
• 429: rate limit; retry with backoff.
• 4xx: invalid endpoint or params; re-check request shape and required fields.
• 5xx: provider-side issue; retry with backoff and re-validate state.
• Empty/parse failures: retry once, then save response to a temp file and parse from file.

Resources

• references/api.md: concise endpoint map and safe calling checklist.
• Treat references/api.md as the working source for endpoint details and refresh it against official docs regularly.