README.md

metadata

title: InternViT Fixed BFloat16 Type
emoji: 🚀
colorFrom: green
colorTo: blue
sdk: docker
pinned: false

InternViT-6B with Fixed BFloat16 Type Conversion

This Space demonstrates loading and using the InternViT-6B model with proper BFloat16 type handling.

Fixes Applied

1. Fixed the data type mismatch between input tensors (float32) and model parameters (bfloat16)
2. Always converts input tensors to bfloat16 before processing
3. Uses a dummy flash-attn implementation to avoid compilation
4. Provides detailed diagnostics on startup

Benefits

• Much faster build time
• No data type mismatches or conversion errors
• Still loads the full 6B parameter model
• Demonstrates basic functionality

Instructions

1. Click the "Test Model Loading" button
2. Wait for the model to load and run the test
3. Check the results for success or errors