Follow
Atmic Intelligence Project- Sri Ramanasramam 4
Sep 20 28 days ago 05:04 Add DPO LoRA adapter (unmerged, PEFT) and handler that applies it; DPO_ADAPTER=off restores plain v1 incl. original defaults Jun 17 03:41 Update handler.py repetition penalty updated. May 29 10:15 Update handler.py May 28 21:01 Update handler.py repetition penalty 1.05 - Double Sri Sri before Bhagavan's name with 1.0 May 26 16:43 Update handler.py repetition penalty removal. 11:27 Update handler.py bos token to be pushed to prompt builder 11:00 Update handler.py bos token added May 25 14:06 Update handler.py Based on HF repo files of AtmicIntelv1 12:29 Update handler.py 12:15 Update handler.py 11:52 Update modeling_phi3.py 11:38 Update modeling_phi3.py 11:25 Update modeling_phi3.py rant correction 10:33 Update config.json revert to original 10:20 Update config.json reverting original 10:09 Update handler.py 10:07 Update handler.py 10:06 Update handler.py 09:58 Update handler.py aligning chat template 08:53 Update handler.py updated based on detailed analysis of training and evaluation phases. 07:15 Update handler.py aligned with sft dataset which was created using claude haiku 06:58 Update handler.py parameters same as evaluation stage. May 24 08:50 Update handler.py matches the unified requirements.txt 08:12 Update handler.py 08:06 Update handler.py hallucination correction which was due to LoRA weight skipping 07:51 Update handler.py self.model.config.attn_implementation = "eager" 07:49 Update handler.py patching handler 07:47 Update config.json dynamic cache implementation 07:41 Update requirements.txt latest transformers 07:40 Update handler.py The shim adds back the single missing method (get_max_length) that transformers ≥ 4.38 removed, returning None which is safe — the internal code that calls it only uses the value for optional length checks, not critical logic.
Without cache (old): token 1 → recompute all attention for N tokens
token 2 → recompute all attention for N+1 tokens
token K → recompute all attention for N+K tokens ← O(n²)
With cache (new): token 1 → compute, store KV in DynamicCache
token 2 → only compute for new token, reuse cache ← O(n) 07:28 Update requirements.txt hope transformers version downgrade works 🥺 07:19 Upload 2 files I hope this commit works 😆 Aligned to Atmic-Prod-v1 and AtmicIntelv1 07:14 Update requirements.txt requirements aligned with model transformer version. AtmicIntelv1 06:29 Update handler.py aligned handler.py to Atmic-Prod-v1 06:13 Update handler.py 05:08 Update handler.py Aligned with Atmic-Prod-v1 May 23 20:08 Update handler.py 19:41 Upload 2 files Endpoint commits May 1 02:51 Add model card 02:51 Upload tokenizer 02:51 Upload Phi3ForCausalLM 02:50 initial commit 
