[fal.ai] LoRA loading fails with Windows local path sent to Linux cloud worker

[livepeer-tessa]

Summary

When a user configures LoRA models on a Windows machine, the local Windows file paths (e.g. C:\Users\RONDO\.daydream-scope\models\lora\...) are being passed directly to the Linux fal.ai cloud worker, causing pipeline load failure.

Error

2026-03-31 00:54:40,475 - scope.server.pipeline_manager - ERROR - [e605847d] Failed to load pipeline longlive: LongLivePipeline.__init__: LoRA loading failed. File not found: C:\Users\RONDO\.daydream-scope\models\lora\Wan2.1-1.3b-lora-highresfix-v1_new.safetensors. Ensure the file exists in the models/lora/ directory.. If this error persists, consider removing the models directory '/data/models' and re-downloading models.
2026-03-31 00:54:40,476 - scope.server.pipeline_manager - ERROR - [e605847d] Failed to load pipeline: longlive
2026-03-31 00:54:40,980 - scope.server.pipeline_manager - ERROR - [e605847d] Some pipelines failed to load

The initial load params sent to the worker confirm the Windows path is being forwarded verbatim:

'loras': [
  {'path': 'C:\\Users\\RONDO\\.daydream-scope\\models\\lora\\Wan2.1-1.3b-lora-highresfix-v1_new.safetensors', 'scale': 0.7, 'merge_mode': 'permanent_merge'},
  {'path': 'C:\\Users\\RONDO\\.daydream-scope\\models\\lora\\daydream-scope-dissolve.safetensors', 'scale': 1.5, 'merge_mode': 'permanent_merge'}
]

App: github_f1lhgmk5v76a0ev1w0u378by-scope-app--prod (prod)
Session: e605847d
Timestamp: 2026-03-31 ~00:54 UTC

Root Cause

The client is sending absolute local file paths for LoRAs directly to the cloud runner. On Windows these are C:\... paths that don't exist on the Linux worker. Only LoRAs uploaded to or available in the cloud model directory (/data/models/lora/) should be referenced.

Expected Behavior

The server/cloud path should either:

1. Reject local absolute paths (especially Windows paths) with a user-friendly error before attempting to load, or
2. Translate them to relative model names that the cloud worker can resolve, or
3. The frontend should not send local-only paths to the cloud runner at all — only cloud-available model references should be forwarded

Impact

• Severity: Medium — pipeline fails to load entirely, user gets no AI output
• Frequency: 2 occurrences in this 12h window (same session)
• Affects Windows users with locally-configured LoRA models who try to use remote inference

[livepeer-tessa]

PR #773 is open with the fix: #773

Fix summary: Added PipelineManager._sanitize_lora_paths() which detects platform-foreign absolute paths (Windows C:\... or Unix paths outside lora_dir) and rewrites them to just the filename resolved under the cloud worker's configured LoRA directory (/data/models/lora/).

Just pushed a follow-up commit to fix a ruff formatting issue and a Windows CI test failure (path separator mismatch — using as_posix() now so paths always use forward slashes regardless of the test runner OS).