[fal.ai] ltx2: i2v_image FileNotFoundError on Linux /tmp path — 2500 chunk errors per session (race condition)

[livepeer-tessa]

Summary

The ltx2 pipeline crashes with a FileNotFoundError on every processed chunk when i2v_image references a Linux /tmp/.daydream-scope/assets/ path that doesn't exist on the worker. This is the Linux variant of the Windows path issue (#826) — in this case the file is expected to exist on the server at a /tmp/ temp path, but either hasn't been downloaded yet (race condition) or has been deleted before inference starts.

2500+ chunk errors in a single session — the most frequent error in the last 12h window.

Error Frequency (2026-04-11 18:10–06:10 UTC, last 12h)

• ~2500 occurrences in session 99a5095b (20:54–20:57 UTC)
• File: /tmp/.daydream-scope/assets/hakoniwa_3D_animationgirlcute_kawaiiindoorsanime_flat_toon_Re_e6855abc-c8d7-4ab6-ac0b-1b7982e322ab_2.png
• Session started: 2026-04-11 20:52:52 UTC — connected user user_3CE4VzRXzh1pTU1qAW6DvYqaVzm
• Pipeline loaded successfully at 20:53:37
• Initial params received at 20:53:50 (camera mode, no i2v reference in initial params)
• Errors begin at 20:54:00

Error Messages

2026-04-11 20:54:54,861 - scope.server.pipeline_processor - ERROR - [99a5095b] Error processing chunk for ltx2: [Errno 2] No such file or directory: '/tmp/.daydream-scope/assets/hakoniwa_3D_animationgirlcute_kawaiiindoorsanime_flat_toon_Re_e6855abc-c8d7-4ab6-ac0b-1b7982e322ab_2.png'

Pattern repeats at ~10–50ms intervals for the entire session duration. The pipeline continues running (generating seeds, detecting "Image-to-video conditioning") but fails on every chunk because the image file is missing.

Session Timeline

20:52:52 - WebSocket connection accepted [99a5095b]
20:53:05 - Loading pipeline ltx2
20:53:37 - Pipeline ltx2 loaded successfully
20:53:50 - Initial params received: {input_mode: "text", prompts: [...], pipeline_ids: ["ltx2"]}
20:53:50 - PipelineProcessor started
20:54:00 - libvpx decode error (single occurrence)
20:54:04 - First chunk processed normally (no i2v at this point)
~20:54:54 - ERROR flood begins: FileNotFoundError on /tmp path

Analysis

This is distinct from #826 (Windows paths like C:\Users\...):

• This error is on a Linux path (/tmp/.daydream-scope/assets/...)
• The file path contains a user-uploaded asset (AI-generated image name from Hakoniwa), suggesting the client sent an i2v_image parameter update referencing a file that either:
  1. Was not yet downloaded to the worker's /tmp when the pipeline tried to use it (race condition between CDN download and inference start)
  2. Was deleted from /tmp during a session reload or cleanup

Unlike #826, where the pipeline eventually recovers using the CDN copy, this session generates errors for the entire observable window without recovery.

Root Cause Hypothesis

The scope_ltx_2 pipeline's _load_image_tensor is called on every chunk when i2v_image is set. If the file is referenced by its /tmp/.daydream-scope/assets/ local path (set during a param update) but the download is still in-flight or the file was evicted, every chunk fails.

The fix should ensure either:

1. The pipeline waits for the file to be available before setting it as active (use CDN URL as fallback)
2. OR the pipeline processor catches the FileNotFoundError and falls back to CDN re-download instead of propagating as an error

See also: #826 (Windows path variant), #770 (Windows LoRA paths)

Environment

• Service: fal.ai (app: github_f1lhgmk5v76a0ev1w0u378by-scope-app--prod)
• Session: 99a5095b
• User: user_3CE4VzRXzh1pTU1qAW6DvYqaVzm
• Timestamp: 2026-04-11 20:54–20:57 UTC