modal

name	modal
description	Build serverless AI pipelines with Modal for cloud-based compute and inference

Modal

When to activate

• User is deploying ML/AI workloads to Modal (modal.com)
• Code imports modal or references @app.function, @app.cls, Stub, App
• User needs serverless GPU compute for inference, fine-tuning, or batch ML jobs
• User asks about cold starts, GPU selection, model weight caching, or streaming inference on Modal
• User is building a web endpoint or scheduled job backed by GPU compute

When NOT to use

• User is deploying to a different serverless GPU platform (RunPod, Lambda Labs, Replicate, Banana)
• Task is CPU-only and has no ML workload — use standard serverless (Lambda, Cloud Run)
• User already has a managed inference endpoint (OpenAI, Anthropic, Together) and just needs an API call
• The question is about ML model architecture, not infrastructure

Instructions

App and Image setup

Modal's entry point is an App (formerly Stub). Every deployed function belongs to one.

import modal

app = modal.App("my-inference-app")

# Build a custom image — do this once, it is cached layer by layer
image = (
    modal.Image.debian_slim(python_version="3.11")
    .pip_install(["torch", "transformers", "accelerate", "vllm"])
    .env({"HF_HUB_ENABLE_HF_TRANSFER": "1"})
)

Build heavy dependencies into the image, not at function startup. Each .pip_install() call is a separate Docker layer and is cached independently.

GPU selection

GPU	VRAM	Use when
T4	16 GB	Dev/testing, small models (≤7B at fp16), cost-sensitive
A10G	24 GB	Production inference, 7–13B models, best price/performance
A100-40GB	40 GB	13–30B models, fine-tuning ≤13B, batch throughput
A100-80GB	80 GB	30–70B models, full fine-tuning of 13B+
H100	80 GB	Maximum throughput, largest models, multi-node training

@app.function(gpu="A10G", image=image)
def run_inference(prompt: str) -> str:
    ...

# Request a specific count
@app.function(gpu=modal.gpu.A100(count=2))
def multi_gpu_job():
    ...

Cold start optimization

Cold starts on GPU instances are dominated by model weight loading, not container spin-up.

# keep_warm holds N containers alive — eliminates cold starts, billed continuously
@app.function(gpu="A10G", keep_warm=1, image=image)
def fast_inference(prompt: str) -> str:
    ...

# container_idle_timeout keeps a container alive N seconds after last request
# Good middle ground: reduces cold starts without keep_warm cost
@app.function(gpu="A10G", container_idle_timeout=300, image=image)
def balanced_inference(prompt: str) -> str:
    ...

Use @app.cls with @modal.enter() to load models once per container lifecycle:

@app.cls(gpu="A10G", image=image, container_idle_timeout=300)
class InferenceModel:
    @modal.enter()
    def load(self):
        from transformers import pipeline
        self.pipe = pipeline("text-generation", model="mistralai/Mistral-7B-v0.1")

    @modal.method()
    def generate(self, prompt: str) -> str:
        return self.pipe(prompt, max_new_tokens=256)[0]["generated_text"]

Volumes for model weights

Never download model weights at function runtime — store them in a Modal Volume.

model_volume = modal.Volume.from_name("model-weights", create_if_missing=True)

@app.function(
    gpu="A10G",
    image=image,
    volumes={"/models": model_volume},
)
def download_model():
    from huggingface_hub import snapshot_download
    snapshot_download("mistralai/Mistral-7B-v0.1", local_dir="/models/mistral-7b")
    model_volume.commit()  # persist writes

@app.function(gpu="A10G", image=image, volumes={"/models": model_volume})
def inference(prompt: str) -> str:
    from transformers import pipeline
    pipe = pipeline("text-generation", model="/models/mistral-7b")
    return pipe(prompt)[0]["generated_text"]

Secrets management

# Create secret in Modal dashboard, reference by name
hf_secret = modal.Secret.from_name("huggingface-token")

@app.function(gpu="A10G", image=image, secrets=[hf_secret])
def download_gated_model():
    import os
    token = os.environ["HF_TOKEN"]  # key set in Modal secret
    ...

Scheduled functions

from modal import Period

@app.function(schedule=Period(hours=6))
def refresh_embeddings():
    # Runs every 6 hours — no cron syntax needed
    ...

# Cron syntax is also supported
@app.function(schedule=modal.Cron("0 2 * * *"))
def nightly_batch():
    ...

Web endpoints

@app.function(gpu="A10G", image=image, allow_concurrent_inputs=10)
@modal.web_endpoint(method="POST")
def serve(request: dict) -> dict:
    result = model.generate(request["prompt"])
    return {"text": result}

allow_concurrent_inputs lets one container handle N simultaneous requests — critical for throughput on long-running inference. Without it, each request gets its own container.

For full ASGI apps:

@app.function(gpu="A10G", image=image)
@modal.asgi_app()
def fastapi_app():
    from api import app as fastapi_app  # your FastAPI app
    return fastapi_app

Mounting local code

local_mount = modal.Mount.from_local_dir(
    "./src",
    remote_path="/root/src",
)

@app.function(mounts=[local_mount], image=image)
def run_local_code():
    import sys
    sys.path.insert(0, "/root/src")
    from my_module import process
    process()

Streaming inference

Use yield to stream tokens back to the caller:

@app.function(gpu="A10G", image=image)
def stream_tokens(prompt: str):
    from vllm import LLM, SamplingParams
    llm = LLM(model="/models/mistral-7b")
    params = SamplingParams(temperature=0.7, max_tokens=512)
    for output in llm.generate([prompt], params, use_tqdm=False):
        for token in output.outputs[0].text:
            yield token

# Caller iterates the generator
for chunk in stream_tokens.remote_gen("Tell me about GPUs"):
    print(chunk, end="", flush=True)

Fine-tuning pattern

@app.function(
    gpu=modal.gpu.A100(memory=80),
    image=image,
    volumes={"/models": model_volume, "/checkpoints": checkpoint_volume},
    timeout=3600 * 8,  # 8-hour max for long training runs
)
def finetune(config: dict):
    from transformers import Trainer, TrainingArguments
    # ... load dataset, model, run Trainer
    trainer.train()
    trainer.save_model("/checkpoints/finetuned")
    checkpoint_volume.commit()

Set timeout explicitly for long jobs — the default is 300 seconds.

Cost model

• Billed per second of GPU time — no idle cost when containers are stopped
• Free tier: $30/month of compute credit
• A10G: ~$1.10/hr; A100-40GB: ~$3.00/hr; H100: ~$7.00/hr (approximate, check modal.com/pricing)
• keep_warm=1 on an A10G costs ~$26/day — only use for latency-critical prod endpoints
• Prefer container_idle_timeout (pay only when traffic exists) over keep_warm for most workloads

Example

Deploying a vLLM inference endpoint on Modal with streaming and web API:

import modal

app = modal.App("vllm-endpoint")

model_volume = modal.Volume.from_name("vllm-weights", create_if_missing=True)
hf_secret = modal.Secret.from_name("huggingface-token")

image = (
    modal.Image.debian_slim(python_version="3.11")
    .pip_install(["vllm==0.4.0", "huggingface_hub[hf_transfer]", "fastapi"])
    .env({"HF_HUB_ENABLE_HF_TRANSFER": "1"})
)

MODEL_ID = "mistralai/Mistral-7B-Instruct-v0.2"
MODEL_DIR = "/models/mistral-7b-instruct"


@app.function(
    image=image,
    gpu="A10G",
    volumes={"/models": model_volume},
    secrets=[hf_secret],
)
def download_weights():
    from huggingface_hub import snapshot_download
    import os
    snapshot_download(
        MODEL_ID,
        local_dir=MODEL_DIR,
        token=os.environ["HF_TOKEN"],
    )
    model_volume.commit()


@app.cls(
    gpu="A10G",
    image=image,
    volumes={"/models": model_volume},
    container_idle_timeout=300,
    allow_concurrent_inputs=8,
)
class VLLMEndpoint:
    @modal.enter()
    def load(self):
        from vllm import LLM, SamplingParams
        self.llm = LLM(model=MODEL_DIR, max_model_len=4096)
        self.sampling_params = SamplingParams(temperature=0.7, max_tokens=)


     () -> :
        outputs = .llm.generate([prompt], .sampling_params)
         outputs[].outputs[].text


     () -> :
         {: .generate(request[])}

Run modal deploy modal_app.py to publish the endpoint. Modal returns a stable HTTPS URL for the web endpoint.