Spaces:
Runtime error
Issues
📋 FinLLaVA Installation Issue - Detailed Summary
Problem Overview:
FinLLaVA model fails to load on macOS M5 with error: "data did not match any variant of untagged enum ModelWrapper"
Root Cause:
The FinLLaVA model on HuggingFace has a tokenizer.json file format incompatibility with current tokenizers library versions. The model's tokenizer was created with an older/different version of the Rust-based tokenizers library, and the serialization format doesn't match what the current libraries expect to parse.
What We Tried:
✅ Initial setup: Conda, Python 3.10, PyTorch with MPS (Metal) support
✅ Model downloaded successfully (3.5GB+, 111 files)
✅ HuggingFace authentication and access granted
✅ Installed LLaVA framework (llava 1.2.2.post1)
✅ Tried multiple tokenizers versions:
0.22.1 (from transformers 4.57.1)
0.19.1 (attempted fix)
0.15.1 (LLaVA requirement) ✓ Final version
✅ Installed transformers from source (5.0.0.dev0)
✅ Re-downloaded model files (fresh cache)
❌ Model still fails to load despite all attempts
Technical Details:
Environment: macOS with M5 chip, 24GB RAM
Python: 3.10
Key Libraries:
transformers 4.37.2
tokenizers 0.15.1
llava 1.2.2.post1
torch 2.1.2 (MPS support)
Error Location: Line 277156 in tokenizer.json (enum parsing failure)
Conclusion:
FinLLaVA's tokenizer.json file has a format incompatibility that prevents local deployment with current library versions. The model may work with very specific legacy versions or requires modifications to the tokenizer files that aren't publicly documented.
Recommendation for TheFinAI Team:
Update tokenizer.json to current tokenizers library format
Provide explicit version requirements in model card
Or provide a conversion script for tokenizer compatibility