Upload README.md with huggingface_hub
Browse files
README.md
CHANGED
@@ -1,10 +1,38 @@
|
|
1 |
---
|
2 |
-
title:
|
3 |
-
emoji:
|
4 |
-
colorFrom:
|
5 |
-
colorTo:
|
6 |
sdk: docker
|
7 |
pinned: false
|
8 |
---
|
9 |
|
10 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
---
|
2 |
+
title: InternViT No Flash-Attn Test
|
3 |
+
emoji: π
|
4 |
+
colorFrom: yellow
|
5 |
+
colorTo: red
|
6 |
sdk: docker
|
7 |
pinned: false
|
8 |
---
|
9 |
|
10 |
+
# InternViT-6B without Flash Attention
|
11 |
+
|
12 |
+
This Space demonstrates loading and using the InternViT-6B model without the flash-attn dependency.
|
13 |
+
|
14 |
+
## Approach
|
15 |
+
|
16 |
+
Instead of trying to compile flash-attn (which times out during build), this version:
|
17 |
+
|
18 |
+
1. Uses a mock/dummy implementation of flash-attn
|
19 |
+
2. Only includes required dependencies (einops)
|
20 |
+
3. Uses a runtime PyTorch image instead of development image
|
21 |
+
4. Monkey patches the imports to avoid errors
|
22 |
+
|
23 |
+
## Benefits
|
24 |
+
|
25 |
+
- Much faster build time
|
26 |
+
- No compilation errors or timeouts
|
27 |
+
- Still loads the full 6B parameter model
|
28 |
+
- Demonstrates basic functionality
|
29 |
+
|
30 |
+
## Note
|
31 |
+
|
32 |
+
Some functionality might be reduced without the flash-attn optimization, but the model should still work for basic inference.
|
33 |
+
|
34 |
+
## Instructions
|
35 |
+
|
36 |
+
1. Click the "Test Model Loading" button
|
37 |
+
2. Wait for the model to load and run the test
|
38 |
+
3. Check the results for success or errors
|