Summary
The longlive pipeline fails to load with "LoRA file not found" for [flux.2.klein]pixelart_redmond-000032.safetensors on the first pipeline load attempt. The error resolves itself on the next attempt once the LoRA download from Civitai completes. This is a race condition between asynchronous LoRA asset download and synchronous pipeline initialisation.
This is a distinct issue from:
cc @mjh1 @emranemran
Error Messages
2026-04-13 23:38:53,747 - scope.core.pipelines.wan2_1.lora.strategies.peft_lora - ERROR - LongLivePipeline.__init__: LoRA file not found: /tmp/.daydream-scope/assets/lora/[flux.2.klein]pixelart_redmond-000032.safetensors
2026-04-13 23:38:53,747 - scope.server.pipeline_manager - ERROR - Failed to load pipeline longlive: LongLivePipeline.__init__: LoRA loading failed. File not found: /tmp/.daydream-scope/assets/lora/[flux.2.klein]pixelart_redmond-000032.safetensors. If this error persists, consider removing the models directory '/data/models' and re-downloading models.
2026-04-13 23:38:53,957 - scope.server.pipeline_manager - ERROR - Some pipelines failed to load
Timeline (2026-04-13 ~23:35–23:40 UTC)
23:35:19 - httpx: HTTP Request: GET https://b2.civitai.com/.../5bflux2Klein5dpixelart.u4AF.safetensors
(Civitai LoRA download begins)
23:36:06 - ✅ Pipeline yolo_mask loaded successfully (first job — LoRA already downloaded)
Initial load params: loras=[{path: /tmp/.daydream-scope/assets/lora/[flux.2.klein]pixelart_redmond-000032.safetensors}]
23:36:34 - ✅ LongLivePipeline: Loaded LoRA '[flux.2.klein]pixelart_redmond-000032.safetensors' (first attempt, succeeds)
23:38:36 - Session resets (yolo_mask + longlive unloaded)
23:38:36 - pipeline_manager: Loading 2 pipeline(s): [yolo_mask, longlive]
23:38:53 - httpx: HTTP Request: GET https://b2.civitai.com/.../5bflux2Klein5dpixelart.u4AF.safetensors
(Civitai download starts AFTER pipeline_manager already tries to load longlive)
23:38:53 - ❌ LoRA file not found: /tmp/.daydream-scope/assets/lora/[flux.2.klein]pixelart_redmond-000032.safetensors
(file not yet downloaded from Civitai)
23:38:54 - scope_yolo_mask plugin installed
23:39:42 - ✅ LongLivePipeline: Loaded LoRA '[flux.2.klein]pixelart_redmond-000032.safetensors' (retry, succeeds)
Job IDs:
ea7914a1-6a2f-4472-ad7b-e6fc0983cde2 (no app label, grpc-run)a8a03ca5-6fce-4cdd-8bca-580e8fbafeeb (scope-app--prod)
LoRA file: [flux.2.klein]pixelart_redmond-000032.safetensors (source: Civitai)
Frequency (last 12h): 1 session (23:36–23:40 UTC 2026-04-13)
Root Cause
When a session is set up with a Civitai-hosted LoRA, the file download is initiated but the pipeline's __init__ proceeds synchronously to load the LoRA before the download has completed. This causes a spurious FileNotFoundError on session reinitialisation (the second job).
On the first job, the LoRA download had already completed from a prior session, so it was available. On the second job restart, the tmp/ state was partially cleared or the download re-triggered, causing the race.
Impact
Suggested Fix
Ensure LoRA assets are fully downloaded to /tmp/.daydream-scope/assets/lora/ before LongLivePipeline.__init__ attempts to load them. Options:
- Await all asset downloads before invoking pipeline
__init__
- Add a polling retry in
peft_lora for the file to appear (with timeout)
- Pre-seed the download on session setup so it's ready when the pipeline loads
Summary
The
longlivepipeline fails to load with "LoRA file not found" for[flux.2.klein]pixelart_redmond-000032.safetensorson the first pipeline load attempt. The error resolves itself on the next attempt once the LoRA download from Civitai completes. This is a race condition between asynchronous LoRA asset download and synchronous pipeline initialisation.This is a distinct issue from:
/tmp/cleared between jobs)cc @mjh1 @emranemran
Error Messages
Timeline (2026-04-13 ~23:35–23:40 UTC)
Job IDs:
ea7914a1-6a2f-4472-ad7b-e6fc0983cde2(no app label, grpc-run)a8a03ca5-6fce-4cdd-8bca-580e8fbafeeb(scope-app--prod)LoRA file:
[flux.2.klein]pixelart_redmond-000032.safetensors(source: Civitai)Frequency (last 12h): 1 session (23:36–23:40 UTC 2026-04-13)
Root Cause
When a session is set up with a Civitai-hosted LoRA, the file download is initiated but the pipeline's
__init__proceeds synchronously to load the LoRA before the download has completed. This causes a spuriousFileNotFoundErroron session reinitialisation (the second job).On the first job, the LoRA download had already completed from a prior session, so it was available. On the second job restart, the
tmp/state was partially cleared or the download re-triggered, causing the race.Impact
Suggested Fix
Ensure LoRA assets are fully downloaded to
/tmp/.daydream-scope/assets/lora/beforeLongLivePipeline.__init__attempts to load them. Options:__init__peft_lorafor the file to appear (with timeout)