v0.0.5

Added

• SFT (Supervised Fine-tuning):

  • SFTDataset for instruction tuning with input masking
  • SFTDataModule for data loading
  • SFTTask registered as --task sft in CLI
  • Tests for all SFT components

• DPO (Direct Preference Optimization):

  • DPODataset handling chosen/rejected pairs
  • DPODataModule for preference data loading
  • DPOTask with reference model management and DPO loss
  • Registered as --task dpo in CLI
  • Tests for all DPO components

• Continuous Batching Engine (Serving):

  • src/llm/serving/engine.py with ContinuousBatchingEngine class
  • Iteration-level scheduling via Scheduler and SlotAllocator
  • Pre-allocated KV cache pool for efficient memory management
  • Supports mixed prefill/decode batching with automatic padding
  • Clean API: requires model and tokenizer instances upfront
  • src/llm/serving/scheduler.py with FCFS scheduling logic

• LoRA (Low-Rank Adaptation):

  • src/llm/core/lora.py with LoRALinear class for parameter-efficient fine-tuning
  • apply_lora(), merge_lora(), get_lora_parameters() helper functions
  • Device/dtype handling for CUDA compatibility
  • 17 tests covering training and weight merging

• QLoRA (Quantized LoRA):

  • src/llm/core/qlora.py with QLoRALinear class
  • NF4 4-bit quantization for base weights (~4x memory reduction)
  • LoRA adapters remain in fp16/bf16 for training stability
  • apply_qlora() and get_qlora_parameters() helpers

• RoPE (Rotary Position Embedding):

  • src/llm/core/rope.py with RotaryPositionEmbedding class
  • Linear, dynamic, and NTK-aware scaling methods for extended context
  • apply_rotary_pos_emb(), get_rope_scaling_factor() utilities
  • 15 tests

• ALiBi (Attention with Linear Biases):

  • src/llm/core/alibi.py with ALiBiPositionBias class
  • get_alibi_slopes(), build_alibi_bias() functions
  • Cached bias computation for efficiency
  • 13 tests

• Sliding Window Attention:

  • window_size parameter in scaled_dot_product_attention
  • Propagated through MultiHeadAttention, TransformerBlock, DecoderModel
  • Reduces memory for long sequences by limiting attention scope
  • 10 tests

• KV Cache Optimization:

  • src/llm/core/kv_cache.py with KVCache class for pre-allocated cache buffers
  • In-place updates during autoregressive generation (avoids O(n²) memory operations)
  • Integrated into MHA, TransformerBlock, DecoderModel
  • Factory method KVCache.from_model_config() for easy instantiation
  • Backward compatible: legacy past_key_value tuple format still works

• E2E Testing Infrastructure:

  • tests/e2e/ directory with comprehensive pipeline tests
  • test_training.py, test_sft.py, test_dpo.py
  • test_gradient_accumulation.py, test_resume_training.py
  • Advanced inference and callback tests

• Documentation:

  • notebooks/quick_start.ipynb interactive tutorial
  • Covers model building, training, inference, and advanced features

Changed

• SDPA Refactoring:

  • Consolidated scaled_dot_product_attention wrapper into src/llm/core/attn/sdpa.py
  • Refactored MultiHeadAttention and MultiLatentAttention to use common sdpa wrapper
  • Archived custom implementation to _learning/03_lab/experiments/custom_sdpa.py

• Test Suite Refactoring:

  • Organized test files into subdirectories (tests/training/, tests/inference/, etc.)
  • Converted to functional testing style (real components over mocks)
  • Added shared fixtures in tests/conftest.py
  • Test count: 385 → 432

• TrainingEngine:

  • Support for dictionary batches in training/validation loops
  • Gradient accumulation implementation

• DPO Reference Model:

  • Use model reconstruction instead of deepcopy for ref_model creation

• Documentation:

  • Added docs/README.md as documentation entry point
  • Added MkDocs Material configuration (mkdocs.yml) for documentation site
  • Added GitHub Actions workflow for automatic GitHub Pages deployment
  • Added guide-finetuning.md (LoRA/QLoRA) and guide-inference.md (KVCache/GQA/Continuous Batching)
  • Enhanced architecture.md with detailed component diagrams and data flow analysis
  • Updated ROADMAP Phase 10.2 (Continuous Batching complete)