Upload README.md with huggingface_hub
Browse files
README.md
CHANGED
|
@@ -1,10 +1,32 @@
|
|
| 1 |
---
|
| 2 |
-
title:
|
| 3 |
-
emoji:
|
| 4 |
-
colorFrom:
|
| 5 |
-
colorTo:
|
| 6 |
sdk: docker
|
| 7 |
pinned: false
|
| 8 |
---
|
| 9 |
|
| 10 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
---
|
| 2 |
+
title: InternViT Fixed BFloat16 Type
|
| 3 |
+
emoji: 🚀
|
| 4 |
+
colorFrom: green
|
| 5 |
+
colorTo: blue
|
| 6 |
sdk: docker
|
| 7 |
pinned: false
|
| 8 |
---
|
| 9 |
|
| 10 |
+
# InternViT-6B with Fixed BFloat16 Type Conversion
|
| 11 |
+
|
| 12 |
+
This Space demonstrates loading and using the InternViT-6B model with proper BFloat16 type handling.
|
| 13 |
+
|
| 14 |
+
## Fixes Applied
|
| 15 |
+
|
| 16 |
+
1. Fixed the data type mismatch between input tensors (float32) and model parameters (bfloat16)
|
| 17 |
+
2. Always converts input tensors to bfloat16 before processing
|
| 18 |
+
3. Uses a dummy flash-attn implementation to avoid compilation
|
| 19 |
+
4. Provides detailed diagnostics on startup
|
| 20 |
+
|
| 21 |
+
## Benefits
|
| 22 |
+
|
| 23 |
+
- Much faster build time
|
| 24 |
+
- No data type mismatches or conversion errors
|
| 25 |
+
- Still loads the full 6B parameter model
|
| 26 |
+
- Demonstrates basic functionality
|
| 27 |
+
|
| 28 |
+
## Instructions
|
| 29 |
+
|
| 30 |
+
1. Click the "Test Model Loading" button
|
| 31 |
+
2. Wait for the model to load and run the test
|
| 32 |
+
3. Check the results for success or errors
|