flash

Runpod Flash

Write code locally, iterate with flash dev — it runs your functions on remote Runpod GPUs/CPUs with hot-reload and live worker logs — then flash deploy to ship. Endpoint handles provisioning.

Load on demand — this skill keeps the mental model + gotchas inline; details live in reference/:

Need	Read
Install, auth, flash init, and the full flash command list	reference/setup-and-cli.md
Endpoint(...) constructor params, NetworkVolume/PodTemplate/EndpointJob, GPU & CPU enum tables	reference/api.md
Worked patterns — choosing a model, warm-worker model loading, CPU→GPU pipeline, parallel calls	reference/patterns.md

Quick start: uv tool install runpod-flash → flash login (or export RUNPOD_API_KEY=...) → flash init my-project → flash dev. Details in reference/setup-and-cli.md.

Dev vs Deploy

• flash dev — iterate. Local server at :8888, but your decorated functions execute on remote GPU/CPU workers. Hot-reloads on save and streams the worker's logs live to the terminal. No build/upload/deploy wait — use this the whole time you develop.
• flash deploy — ship. Builds an artifact and deploys a stable endpoint. Slow (build + upload + provision); only do this once the code works under flash dev.

flash dev ships only the function body to the worker, so a NameError for a module-level name surfaces immediately here. flash deploy imports the whole module and can mask that bug (see Gotcha #1). Develop against flash dev and you catch it first.

Autonomous Dev Loop

flash dev is a long-running server. Three rules:

• Run it in the background — don't block on it.
• Capture its output to a log file.
• Drive it over HTTP.

The captured log is the remote worker's live stream (cold start, model load, prints, tracebacks) — read it to debug.

flash dev > /tmp/flash-dev.log 2>&1 &                          # background; never run it blocking
for i in $(seq 1 60); do grep -q "flash dev  localhost:" /tmp/flash-dev.log && break; sleep 2; done  # bounded ~2min; if it never appears, check the log for errors
URL=$(grep -o "localhost:[0-9]*" /tmp/flash-dev.log | head -1)               # actual port (8888 bumps if taken)
curl -s "$URL/main/predict" -d '{"data": {...}}'               # dispatches to the remote worker

• Read the real URL from the log — flash auto-bumps the port if 8888 is in use, and prints ✓ flash dev localhost:<port> plus the route table.
• Routes are namespaced by file: main.py's /predict is served at /main/predict.
• Two route shapes, two body shapes (mismatch → 422 naming the missing field in loc):
  • Load-balanced (@api.post("/predict")) → POST /main/predict, body is the arg at top level: a handler def predict(data: dict) wants {"data": {...}} (not the bare object).
  • Queue-based (bare @Endpoint decorator) → POST /main/runsync (the local dev server only generates /runsync; production also exposes /run), body is double-wrapped in input: a handler def synthesize(data: dict) wants {"input": {"data": {...}}}. The outer input is the queue envelope; the inner key is the handler's param name.
• Edit a handler and save — hot-reload re-syncs the body; just re-send the request, no redeploy. Add --auto-provision to skip the first-call cold start. kill %1 when done.

Endpoint: Three Modes

Full constructor params and the GPU/CPU enum tables are in reference/api.md.

Mode 1: Your Code (Queue-Based Decorator)

One function = one endpoint with its own workers.

from runpod_flash import Endpoint, GpuGroup

@Endpoint(name="my-worker", gpu=GpuGroup.AMPERE_80, workers=5, dependencies=["torch"])
async def compute(data):
    import torch  # MUST import inside function (cloudpickle)
    return {"sum": torch.tensor(data, device="cuda").sum().item()}

result = await compute([1, 2, 3])

Mode 2: Your Code (Load-Balanced Routes)

Multiple HTTP routes share one pool of workers.

from runpod_flash import Endpoint, GpuGroup

api = Endpoint(name="my-api", gpu=GpuGroup.ADA_24, workers=(1, 5), dependencies=["torch"])

@api.post("/predict")
async def predict(data: list[float]):
    import torch
    return {"result": torch.tensor(data, device="cuda").sum().item()}

@api.get("/health")
async def health():
    return {"status": "ok"}

Mode 3: External Image (Client)

Deploy a pre-built Docker image and call it via HTTP.

from runpod_flash import Endpoint, GpuGroup, PodTemplate

server = Endpoint(
    name="my-server",
    image="my-org/my-image:latest",
    gpu=GpuGroup.AMPERE_80,
    workers=1,
    env={"HF_TOKEN": "xxx"},
    template=PodTemplate(containerDiskInGb=100),
)

# LB-style
result = await server.post("/v1/completions", {"prompt": "hello"})
models = await server.get("/v1/models")

# QB-style
job = await server.run({"prompt": "hello"})        # optional: webhook="https://..." for completion callback
await job.wait()
print(job.output)

Connect to an existing endpoint by ID (no provisioning):

ep = Endpoint(id="abc123")
job = await ep.runsync({"prompt": "hello"})  # runsync wraps this as {"input": {"prompt": "hello"}}
print(job.output)

How Mode Is Determined

Parameters	Mode
name= only	Decorator (your code)
image= set	Client (deploys image, then HTTP calls)
id= set	Client (connects to existing, no provisioning)

The table above is how the mode is picked from params. When to reach for image=:

When to use image= (custom container) vs your own code

Default to writing Python (decorator / routes) — it runs arbitrary code with dependencies=[...]/system_dependencies=[...] and needs no Dockerfile. Even large HuggingFace models stay in decorator mode (weights stream at runtime — see reference/patterns.md → Loading ML models). Reach for image= only when you need:

• a pre-built inference server — vLLM, TensorRT-LLM (image="vllm/vllm-openai:latest", or runpod/worker-vllm, runpod/worker-comfy)
• system-level deps not pip-installable — a specific CUDA/cuDNN, OS libraries
• models baked into the image — to skip the runtime download entirely
• an existing Runpod Serverless worker — you already have a working image

Trade-off: image= mode can't run arbitrary Python (the image owns all logic) and the image must implement a Runpod Serverless handler. Full list + examples: https://docs.runpod.io/flash/custom-docker-images

Gotchas

1. Only the function body ships to the worker -- most common error. Put imports and any module-level constants/helpers the function uses inside the decorated body. flash deploy imports the whole module so module globals happen to work; flash dev ships just the body, so a module-level name raises NameError. A handler that works deployed can break under dev — fix it by moving everything inside.
2. Forgetting await -- all decorated functions and client methods need await.
3. Missing dependencies -- must list in dependencies=[].
4. gpu/cpu are exclusive -- pick one per Endpoint.
5. idle_timeout is seconds -- default 60s, not minutes.
6. 10MB payload limit -- pass URLs, not large objects. Return binary (audio/images/files) as base64 in the JSON ({"audio_b64": ...}) and decode client-side; for larger outputs write to a NetworkVolume or upload to storage and return a URL.
7. Client vs decorator -- image=/id= = client. Otherwise = decorator.
8. Auto GPU switching requires workers >= 5 -- pass a list of GPU types (e.g. gpu=[GpuGroup.ADA_24, GpuGroup.AMPERE_80]) and set workers=5 or higher. The platform only auto-switches GPU types based on supply when max workers is at least 5.
9. runsync timeout is 60s -- cold starts can exceed 60s. Use ep.runsync(data, timeout=120) for first requests or use ep.run() + job.wait() instead.
10. Request body shape (raw/external HTTP callers only) -- match the request shape to the endpoint type:
    • LB routes (@api.post(...)): send the handler arg at the top level — {"data": {...}}.
    • QB endpoints (bare @Endpoint, hit via .../run or .../runsync): the worker calls handler(**job_input), so the request's input keys must match the handler's parameter names — def transcribe(input_data: dict) wants {"input": {"input_data": {...}}}, and def read(input: dict) wants {"input": {"input": {...}}}. A mismatch fails with got an unexpected keyword argument …. Use **kwargs if the handler ignores the payload.
    • Never send an empty input. A QB request with {"input": {}} is rejected by the worker SDK as Job has missing field(s): id or input — always include at least one key.
    • Context: the flash client (ep.runsync(x), api.post(...)) hides the spreading, so this only bites raw HTTP/external callers (mismatch behavior verified 2026-07-10 via worker logs). See Autonomous Dev Loop.
11. Load a model once per worker (not per call) -- for real inference use a class @Endpoint whose __init__ loads the model once per worker (see reference/patterns.md → Loading ML models). In function-form, reconcile with #1 by caching in a module global inside the body so it works under both flash dev and deploy:

    global _MODEL
    try: _MODEL
    except NameError: _MODEL = load_model()   # runs once per worker, reused across calls
    

12. Native CUDA libs go in dependencies=[] too -- e.g. CTranslate2/faster-whisper needs nvidia-cublas-cu12 + nvidia-cudnn-cu12 or it silently falls back to CPU. Add them alongside the Python package.
13. Silent 401 auth failure -- a set RUNPOD_API_KEY env var overrides the flash login token, so a bad/expired key wins. The failure is quiet: provisioning logs GraphQL request failed: 401, but flash dev still prints its normal ready line ("failed endpoints deploy on-demand"), so it looks healthy. When endpoints fail to provision:
    1. Check the provisioning log for GraphQL request failed: 401.
    2. Verify the current key independently: curl -s -o /dev/null -w '%{http_code}' https://rest.runpod.io/v1/endpoints -H "Authorization: Bearer $RUNPOD_API_KEY" (200 = good, 401 = bad).
    3. Fix it: unset RUNPOD_API_KEY to fall back to the flash login token, or export a valid key.
14. system_dependencies= adds to cold start -- apt packages (e.g. ["ffmpeg", "espeak-ng"]) install on the worker before first use, so the initial call is slower (on top of any model download); warm calls are unaffected.
15. Teardown a deployed app with flash app delete <app> -- flash undeploy list may show "no endpoints" for an app that is deployed and serving; flash app delete (or runpodctl serverless delete <id>) reliably removes it.

Resources

• Setup & CLI: reference/setup-and-cli.md · API & compute enums: reference/api.md · Patterns: reference/patterns.md
• Flash source: https://github.com/runpod/flash
• Runnable examples: https://github.com/runpod/flash-examples — clone and adapt the closest one
• Package (PyPI): https://pypi.org/project/runpod-flash/
• Docs: https://docs.runpod.io/flash/overview
  • Custom Docker images (when + how): https://docs.runpod.io/flash/custom-docker-images
  • Storage / network volumes: https://docs.runpod.io/flash/configuration/storage