llm-research-notes

Notes & lightweight experiments on LLMs, open-weight models, multimodal systems Ghi chú & thử nghiệm nhẹ về LLM, mô hình open-weight, hệ thống đa phương thức

Goals / Mục tiêu

• Track learning & decisions clearly (theory → practice).
• Build a minimal but solid stack for open-weight models (download, storage, inference)
• Explore “core block / living core” integration as a state layer on top of base models

Scope / Phạm vi

• Model: gpt-oss-120B, Llama 3/4, Mixtral (gguf), vLLM/Transformers/Ollama.
• Infra: Vertex AI endpoints, Hugging Face Hub, Cloud Storage (GCS/S3), basic MLOps
• Techniques: prompting, RAG, external agent state, light SFT/LoRA.

Progress Log / Tiến độ

• [YYYY-MM-DD] Init repo & plan structure.
• [YYYY-MM-DD] Notes on downloading open-weight models & storage strategy.
• [YYYY-MM-DD] Vertex AI endpoint reading + cost model (per-token vs per-hour).

Update this section as you go / Cập nhật mục này khi bạn tiến hành.

Experiments / Thử nghiệm

• Minimal inference with vLLM (local or cloud)
• Compare memory/storage footprints (Dense vs MoE, quantization)
• State management layer for “core block” (external orchestration)

References / Tham khảo

• Hugging Face Hub docs, vLLM, Transformers
• Vertex AI Docs (Model Garden, Endpoints)
• Llama model cards & licenses

#LLA Research Notes

License

MIT