The checkpoint stored both tok_embeddings.weight and a separate
lm_head.weight, each [24576, 576] = 14,155,776 params, even though
config.json sets tie_word_embeddings: true. At runtime the head is
tied to the embedding, so lm_head.weight is never read.
This removes that one tensor. Verified:
- 201 -> 200 tensors; stored elements 96,017,472 -> 81,861,696
(exactly the card's count)
- file 192,054,704 -> 163,743,061 bytes (-28.3 MB)
- loaded both versions with the repo's own PicoLMV2ForCausalLM:
identical logits on fixed inputs (max |diff| = 0.0), same 81,861,696
unique params, head tied to embedding in both.
Requested by aethertp in the discussion (Option 1).
