silx-ai/raven

Repo-local Raven dependency and Quasar-Preview runner used with silx-ai/Quasar-Preview revision 48d9dc1d2dcdb2eeb087c7b79c89300d3e46a70b.

This packages the Raven implementation from goombalab/raven under raven/, the fixed Quasar FLA folder under fla/, plus a simple script that wires them for Quasar greedy generation.

Run

python scripts/generate_quasar.py \
  --model-dir /workspace/quasar_preview_run/Quasar-Preview \
  --raven-dir /workspace/quasar_preview_run/goombalab-raven \
  --prompt "Mo Salah is" \
  --max-new-tokens 24

The runner handles:

• direct tokenizer.json loading because the checkpoint tokenizer config references TokenizersBackend
• symlinking raven/ into the model dir and Transformers dynamic-module cache
• restoring experts_w3 tensors that Transformers may leave as meta tensors
• running direct model.generate()

The Raven/FLA generation-cache fix is included here under fla/ops/gsa/fused_recurrent.py and has also been pushed to silx-ai/Quasar-Preview.

Simpler script

Use scripts/generate_quasar.py for direct greedy text output. It hides the checkpoint workarounds and prints only generated text.