Add extraction script: README.md
Browse files- extraction_scripts/README.md +27 -0
extraction_scripts/README.md
ADDED
|
@@ -0,0 +1,27 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# LoRA Extraction Scripts
|
| 2 |
+
|
| 3 |
+
Scripts used to extract LoRA adapters from the Biomni-R0-32B model.
|
| 4 |
+
|
| 5 |
+
## Files
|
| 6 |
+
|
| 7 |
+
- `extract_lora_256.sh` - SLURM batch script for LoRA extraction using MergeKit
|
| 8 |
+
- `sanitize_lora.py` - Script to sanitize LoRA adapters for vLLM compatibility
|
| 9 |
+
|
| 10 |
+
## Usage
|
| 11 |
+
|
| 12 |
+
### Extraction
|
| 13 |
+
|
| 14 |
+
```bash
|
| 15 |
+
sbatch extract_lora_256.sh
|
| 16 |
+
```
|
| 17 |
+
|
| 18 |
+
### Sanitization (for vLLM)
|
| 19 |
+
|
| 20 |
+
```bash
|
| 21 |
+
python sanitize_lora.py
|
| 22 |
+
```
|
| 23 |
+
|
| 24 |
+
The sanitization script:
|
| 25 |
+
1. Removes unsupported layers (lm_head, embed_tokens, norms)
|
| 26 |
+
2. Fixes adapter_config.json for vLLM compatibility
|
| 27 |
+
3. Creates backups before making changes
|