Commit History

  1. Update README.md

    Verified signature
    ada3e5a
  2. Update README.md

    Verified signature
    281242b

  1. fix: has_no_defaults_at_init=True per evitare RecursionError in to_diff_dict

    Verified signature
    c54f415
  2. Update README.md

    Verified signature
    86ea512
  3. Update README.md

    Verified signature
    6b0d704
  4. feat: repetition penalty in generate_text

    Verified signature
    3a388e1
  5. fix: inv_freq calcolato runtime, non buffer (evita corruzione meta-device)

    Verified signature
    38feb38
  6. fix: RotaryEmbedding lazy cache build (evita garbage da meta-device init)

    Verified signature
    4e1e026
  7. fix: safetensors ricostruito solo da named_parameters

    Verified signature
    176de1f
  8. fix: inv_freq persistent=False, safetensors solo named_parameters

    Verified signature
    257acf0

  1. fix: cast float32 prima di RoPE per evitare overflow

    Verified signature
    9716697
  2. fix: SDPA in float32 per stabilità numerica

    Verified signature
    23eadd7
  3. fix: cast q,k a dtype di v dopo RoPE — identico a train.py

    Verified signature
    434cb12
  4. fix: aggiunto inv_freq nel safetensors

    Verified signature
    674494c
  5. fix: inv_freq persistent=True + incluso nel safetensors

    Verified signature
    5f72285
  6. fix: weight tying via embed_tokens.weight.T, rimuove lm_head

    Verified signature
    d660488
  7. fix: safetensors da named_parameters() con verifica NaN

    Verified signature
    4d92421
  8. fix: cast q,k,v a dtype originale dopo RoPE

    Verified signature
    b0ed64c
  9. fix: v.to(q.dtype) prima di SDPA

    Verified signature
    85e1731
  10. fix: copia esatta architettura da train.py, RoPE senza cast dtype

    Verified signature
    5c0671e
  11. fix: tie_word_embeddings false

    Verified signature
    3c77b79
  12. fix: lm_head.weight esplicito nel safetensors

    Verified signature
    3db9f1c
  13. fix: NaN guard in generate_text + _keys_to_ignore_on_load_missing

    Verified signature
    91da9fd
  14. fix: tie_weights(**kwargs) per compatibilità transformers 4.40+

    Verified signature
    e0d9139
  15. fix: dtype cast in SDPA + tie_weights() esplicito

    Verified signature
    cd6716a
  16. Export Quark Instruct checkpoint

    Verified signature
    92a7d9d
  17. initial commit

    Verified signature
    3c3cbe3