[streamdiffusion] _load_model fails for 'stabilityai/sd-turbo' — all loading methods exhausted (22 occurrences, 2026-03-24)

[livepeer-tessa]

Summary

The _load_model function in wrapper.py fails when a client requests model_id: 'stabilityai/sd-turbo' (non-SDXL). All three loading strategies are attempted and all fail. The pipeline falls back to the previously loaded model. 22 occurrences across 3 test stream sessions in the last 12 hours (staging).

Error Details

level=ERROR location=wrapper.py:1169:_load_model
message=_load_model: All loading methods failed for model 'stabilityai/sd-turbo'.
Last error: Invalid `pretrained_model_name_or_path` provided. Please set it to a valid URL.

Followed by:

level=ERROR location=pipeline.py:142:_reload_pipeline
message=[update_params] Error reloading pipeline, falling back to previous params:
  _load_model: All loading methods failed for model 'stabilityai/sd-turbo'.
  Last error: Invalid `pretrained_model_name_or_path` provided. Please set it to a valid URL.

Loading Sequence (from logs)

_load_model: Attempting to load with AutoPipeline from_pretrained...
  ⚠ WARNING: stabilityai/sd-turbo does not appear to have a file named model_index.json.
_load_model: Attempting to load with SD from_single_file...
  ⚠ WARNING: SD from_single_file failed: Invalid pretrained_model_name_or_path. Please set it to a valid URL.
_load_model: Attempting to load with SDXL from_single_file...
  ⚠ WARNING: SDXL from_single_file failed: Invalid pretrained_model_name_or_path. Please set it to a valid URL.
ERROR: All loading methods failed

Stack Trace

File "/app/live/streamdiffusion/.venv/lib/python3.11/site-packages/streamdiffusion/wrapper.py", line 1174, in _load_model
RuntimeError: _load_model: All loading methods failed for model 'stabilityai/sd-turbo'.
Last error: Invalid `pretrained_model_name_or_path` provided. Please set it to a valid URL.

Context

• Affected streams: aiJobTesterStream-1774367307005720297, aiJobTesterStream-1774365236335977691, aiJobTesterStream-1774357901550169619 (AI job tester, not real users)
• Node: ai.nightnode.net
• Pipeline: streamdiffusion-sdxl
• First seen: ~13:11 UTC 2026-03-24; Last seen: ~15:48 UTC 2026-03-24
• Frequency: 22 errors across 3 stream sessions

Root Cause

stabilityai/sd-turbo (the non-SDXL variant) is being passed as model_id. It lacks a model_index.json file on HuggingFace Hub so AutoPipeline.from_pretrained() fails. The from_single_file() fallback also fails because HF hub paths aren't supported for that loader.

The system supports stabilityai/sdxl-turbo correctly (which loads via AutoPipeline). This is likely a client/tester misconfiguration.

Suggested Fix

1. Add validation in _reload_pipeline / param update path: check if model_id == 'stabilityai/sd-turbo' and either remap to stabilityai/sdxl-turbo or return a clear validation error before attempting load.
2. Or: add sd-turbo to a blocklist/allowlist at the API layer with a helpful error message pointing to the correct model ID.
3. Improve the error message emitted to the monitoring event to be more actionable.

Observed Impact

• Non-fatal: pipeline falls back to previous model. User-visible effect is that the requested model change silently fails.
• Monitoring event of type error is emitted to the stream, so the client can detect it.

Grafana Reference

fal.ai Logs Dashboard — 2026-03-24, 13:00–16:00 UTC

[livepeer-tessa]

Fix implemented in daydreamlive/StreamDiffusion — PR opened:

daydreamlive/StreamDiffusion#3 — add from_pretrained fallbacks for HF hub model IDs in _load_model
daydreamlive/StreamDiffusion#3

Root cause: The fallback chain for non-.safetensors paths only tried from_single_file(), which requires a local file path and always fails for HF hub IDs like stabilityai/sd-turbo. Correct fix is to use StableDiffusionPipeline.from_pretrained for hub IDs.

What the PR does:

• Detects HF Hub IDs and routes them to from_pretrained-based loaders only
• Adds logger.warning when stabilityai/sd-turbo is used (non-SDXL variant) suggesting stabilityai/sdxl-turbo
• Fixes the SDXL type-check retry block to also use from_pretrained for hub IDs

Mitigation: Once the updated StreamDiffusion package is deployed to ai.nightnode.net, both sd-turbo and other hub-only SD 2.x models will load correctly. The client-side misconfiguration (sd-turbo instead of sdxl-turbo) will now produce a clear warning rather than 22 silent fallbacks.