Spaces:
Running
Running
Commit ·
72da90f
1
Parent(s): 7fb343c
Add RunPod resume flags to README
Browse files
README.md
CHANGED
|
@@ -237,6 +237,26 @@ If a run is interrupted, rerun with `--skip-census` and/or
|
|
| 237 |
`--skip-existing-analysis` only after confirming the expected `.npz` files and
|
| 238 |
analysis outputs exist.
|
| 239 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 240 |
### Combine outputs into the atlas
|
| 241 |
|
| 242 |
Use these commands to build a final combined atlas from existing MLP and
|
|
|
|
| 237 |
`--skip-existing-analysis` only after confirming the expected `.npz` files and
|
| 238 |
analysis outputs exist.
|
| 239 |
|
| 240 |
+
When resuming an existing RunPod atlas root, this is the flag set we actually
|
| 241 |
+
used:
|
| 242 |
+
|
| 243 |
+
```bash
|
| 244 |
+
python app.py --model meta-llama/Llama-3.1-8B-Instruct \
|
| 245 |
+
--corpus prompts/prompts_balanced.jsonl \
|
| 246 |
+
--positive prompts/authentic.jsonl --negative prompts/corporate.jsonl \
|
| 247 |
+
--outdir outputs/llama8b-attn --atlas atlas/llama8b-attn \
|
| 248 |
+
--components attn,heads,q,k,v --max-length 128 --batch-size 64 \
|
| 249 |
+
--dtype float16 --attn-implementation flash_attention_2 \
|
| 250 |
+
--layers all --timing-every 10 \
|
| 251 |
+
--skip-census --skip-existing-analysis \
|
| 252 |
+
--sub-zero --logit-lens \
|
| 253 |
+
--wandb-project qwip-atlas --wandb-run-name llama3.1-8b-attn-qkv \
|
| 254 |
+
--persist-login
|
| 255 |
+
```
|
| 256 |
+
|
| 257 |
+
If census already exists but you want to rerun later stages only, drop
|
| 258 |
+
`--skip-census` and keep `--skip-existing-analysis`.
|
| 259 |
+
|
| 260 |
### Combine outputs into the atlas
|
| 261 |
|
| 262 |
Use these commands to build a final combined atlas from existing MLP and
|