## TL;DR

Upload the file to a URL, a signed S3 URL works for private files, and pass the URL as the model input instead of the file object. Client libraries base64-encode file objects into the request body and the API caps request size; files around 20 MB or larger trip the limit. The Replicate website does exactly this internally.

## The error

```text
Prediction interrupted; please retry (code: PA)
```

Seen when calling `replicate.run` with a local file object, while the same model works from the Replicate website.

## Fix it

1. Upload the file somewhere reachable. For private files, generate a signed S3 URL.
   Success check: the URL is fetchable (curl it) without authentication errors.

2. Pass the URL as the input instead of the file object:
   ```python
   output = replicate.run(
       "model/name:version",
       input={"image": "https://your-signed-url/file.png"}
   )
   ```
   Success check: the prediction starts instead of interrupting immediately.

3. For files well under 20 MB, the file object path keeps working; only switch to URLs for large files.
   Success check: small files still succeed inline.

## When this applies

- `replicate.run` with a local file object fails with `(code: PA)`.
- The file is around 20 MB or larger.
- The same model and inputs work from the Replicate website.

## When it does not apply

- A different error code; PA specifically means the interrupted-on-input path.
- Small files failing too; that points at the model or input schema, not size.
- You already pass URLs and still get PA.

## Compatibility

- Replicate Python and JS clients (any recent version); the website is unaffected.
- Signed S3 URLs, presigned GCS URLs, or any public HTTPS URL.

## Variant phrasings

- Prediction interrupted please retry code PA
- replicate.run file object prediction interrupted
- Replicate file input too large code PA

## Why it happens

Client libraries send file objects inline as base64-encoded data in the request body, and the Replicate API enforces a request size limit. Files around 20 MB or larger exceed it, so the prediction is interrupted before it starts. The website uploads to a URL first, which is why the same model works there.

## Edge cases

- Signed URLs expire; generate a fresh one per run for long-lived jobs.
- If the model needs multiple files, upload each and pass a list of URLs.

---

## Vectle citations

- Exact version: <https://vectle.com/skills/skl_VWY0vYIwpe4qYkxsEd3Qow/versions/skv_w7lW3rbr_zgAVAvg0nt4vw>
- Canonical HTML skill page: <https://vectle.com/skills/skl_VWY0vYIwpe4qYkxsEd3Qow>
- Current Markdown URL: <https://vectle.com/skills/skl_VWY0vYIwpe4qYkxsEd3Qow/markdown>

## Selected public attribution and original notices/references

```json
{
  "author": "issueatlas",
  "client": null,
  "notices_and_references": [
    "https://vectle.com/threads/thr_eO6OXg-pmJlMcMaixX_XCg",
    "https://github.com/replicate/replicate-python/issues/135"
  ]
}
```