Akiragpu / modules /local_gpu_llm.py

Commit History

  1. local-gpu: Qwen2.5-7B-Instruct 4-bit (~5GB, faster load, smarter)

    Isaac Quarenta
    dc575e5
  2. local-llm: swap 8B 16GB for Qwen2.5-3B (~6GB) — loads in ~10s on ZeroGPU FREE

    Isaac Quarenta
    89bd0bd

  1. fix(generate): BatchEncoding->input_ids/attention_mask explicitos (fim do 500/AttributeError shape)

    Isaac Quarenta
    902b1f1
  2. feat(gpu-ext): fallback GPU externa 4-bit (Kaggle T4) + AKIRA_QUANT toggle + kaggle_backend.py

    Isaac Quarenta
    7c8bb54
  3. refactor(gpu): padrao brain - spaces primeiro, bf16+.to(cuda) sem 4-bit, warmup processo principal

    Isaac Quarenta
    0e55a7e
  4. fix(hf): snapshot_download autenticado via rotacao de tokens + hf-transfer (fim do 429) + fix Gradio JSON tuple

    Isaac Quarenta
    ce42495
  5. feat(status): device via nvidia-smi (NVML, sem contexto CUDA) p/ identificar a GPU actual no Status

    Isaac Quarenta
    298bf42
  6. fix(status): sondagem de GPU/bitsandbytes apenas dentro do contexto spaces.GPU (fora pendurava o arranque)

    Isaac Quarenta
    daa766f
  7. feat(status): reporta dispositivo ZeroGPU actual (device/VRAM/capability) + versoes torch/bnb p/ confirmar suporte Blackwell

    Isaac Quarenta
    729353d
  8. fix(gpu): cache explicito em /data (evita download dentro da janela GPU) + deteccao de VRAM libertada + cooldown apos quota esgotada⁠

    Isaac Quarenta
    a57148b
  9. feat(gpu): GPU ativo por omissao (40s) + modelo local 8B com prefetch para /data⁠

    Isaac Quarenta
    69df760

  1. fix(chat): chat nunca parte no Gradio (fallback GPU->sem GPU em 429 ZeroGPU) + NLP/classificadores forcados a CPU + probes de CUDA reais

    Isaac Quarenta
    e48cf8f
  2. feat(chat): default local LLM Qwen2.5-14B-Instruct + theme movido para launch() (Gradio 6)

    Isaac Quarenta
    ab42ce2
  3. fix(build): transformers 5.x + hub 1.33 (resolve ResolutionImpossible vs gradio 6.29) + QLoRA passa a usar peft_config

    Isaac Quarenta
    6989ead
  4. feat(gpu): migracao para Space Gradio GPU (Akiragpu/L4) - modelo local 4-bit em 1 prioridade da chain, QLoRA real, app.py Chat/Treino/Status

    Isaac Quarenta
    bfcd5c0