Upload docs/HOW_FRACTUS_IS_TRAINED.md with huggingface_hub
Browse files
docs/HOW_FRACTUS_IS_TRAINED.md
CHANGED
|
@@ -170,3 +170,8 @@ Fractus is trained as eight parallel continuous-thought engines on sharded token
|
|
| 170 |
---
|
| 171 |
|
| 172 |
Machine notes from the live 8x5090 recovery run. Update when the recipe changes.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 170 |
---
|
| 171 |
|
| 172 |
Machine notes from the live 8x5090 recovery run. Update when the recipe changes.
|
| 173 |
+
|
| 174 |
+
|
| 175 |
+
## CPU mini-Fractus and merges
|
| 176 |
+
|
| 177 |
+
See **docs/CPU_MINI_MERGE_AND_DIMENSIONS.md** — shape rules, what CPU does and does not advance, mini→1B is not free mean-merge.
|