Skip to content

CHANGELOG.md

Latest commit

 

History

History
133 lines (93 loc) · 6.21 KB

CHANGELOG.md

File metadata and controls

133 lines (93 loc) · 6.21 KB
Raw

Changelog

All notable changes to this project will be documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

[0.5.2] - 2026-05-15

Fixed

  • Type safety: resolve all mypy errors across 70 source files
    • Remove invalid # type: ignore[Any] comments (Any is not a valid mypy error code)
    • Replace any/callable builtins with proper Any/Callable type annotations
    • Add @overload signatures for tokenizer methods with union return types
    • Add explicit None checks to narrow union types for mypy
  • Code quality: resolve all ruff lint issues
    • Replace assert with explicit if + raise ValueError (S101 bandit rule)
    • Fix Callable import from typing instead of collections.abc (UP035)
    • Fix module-level import not at top of file (E402)
    • Fix dummy variable _plot_dir accessed after prefix (RUF052)
  • Bug fixes discovered during type checking:
    • Fix run_noise_tunnel() call missing required base_method argument
    • Fix DNAInference.__init__ not storing self.config
    • Fix CLI config key mismatch (training_args -> finetune)
    • Fix Mutagenesis.evaluate() return type (list[dict] -> dict[str, Any])
    • Fix LoggingContext.__exit__ crash when original_level is None
  • Ruff formatting: reformat 4 files after line-length changes

[0.5.1] - 2026-05-09

Added

  • MCP: Streamable HTTP transport support (client + server) — MCP Streamable HTTP migration (Phase 5)
  • MCP: transport="streamable-http" support in DNALLMMCPClient using mcp.client.streamable_http.streamablehttp_client() (Phase 5)
  • MCP: StreamableHTTPConfig configuration block with host, port, and path fields (Phase 5)
  • MCP: Streamable HTTP integration tests in test_streamable_http_client.py (Phase 5)
  • MCP: Client SDK unit tests for transport="streamable-http" initialization and connection (Phase 5)

Changed

  • MCP: DNALLMMCPClient transport type expanded from Literal["sse", "stdio"] to Literal["streamable-http", "sse", "stdio"] (Phase 5)
  • MCP: Server docstrings and CLI help text now recommend streamable-http as the primary remote transport (Phase 5)
  • README: MCP Server section now shows streamable-http as the primary example with SSE noted as legacy (Phase 5)
  • Recommended remote transport changed from SSE to Streamable HTTP per MCP spec 2025-11-25

Deprecated

  • SSE transport marked as legacy; still supported for backward compatibility

[0.5.0] - 2026-05-08

Overview

Version 0.5.0 is a major release that transforms DNALLM from a basic fine-tuning toolkit into a production-ready platform for DNA language model research. This release introduces five major capability areas:

  1. Quality Assurance Infrastructure (Phase 1) — Restored CI/CD, unified testing, and modern tooling
  2. Advanced Training Features (Phase 2) — Early stopping, hyperparameter search, QLoRA, and visualization
  3. MCP Server & Client SDK (Phase 3) — Full Model Context Protocol integration with 13 tools
  4. Dependency Modernization & UX (Phase 4) — Updated dependencies and Click-based CLI

Added

Training & Fine-tuning (Phase 2)

  • Fine-tuning: Early stopping callback in DNATrainer with configurable patience and threshold
  • Fine-tuning: Optuna hyperparameter search via YAML-configurable search space with auto-inferred distributions
  • Fine-tuning: Training visualization with loss curves and learning rate schedule plotting utilities
  • Fine-tuning: QLoRA 4-bit quantization support via bitsandbytes with automatic fix for improperly quantized layers
  • Fine-tuning: Configurable gradient clipping integrated into TrainingConfig

MCP Server & Client (Phase 3)

  • MCP: dna_mutagenesis tool exposing Mutagenesis class with 5 mutation types
  • MCP: dna_interpret tool exposing DNAInterpret class with 8 Captum attribution methods
  • MCP: Python client SDK (DNALLMMCPClient) with dual sync/async API for all 13 server tools
  • MCP: Request timeout handling on all 13 tools with configurable tool_timeout_seconds
  • MCP: Structured JSON/text dual-format logging for production observability

CLI & User Experience (Phase 4)

  • CLI: dnallm-mutagenesis standalone command for in-silico mutation analysis
  • CLI: dnallm-train migrated from sys.argv to Click framework with typed options and help text
  • CLI: dnallm-inference migrated from sys.argv to Click framework with typed options and help text

Infrastructure & Tooling (Phase 1)

  • QA: CI test execution restored with unified pytest configuration
  • QA: Shared fixtures centralized in tests/conftest.py
  • QA: GPU/CUDA tests re-enabled in CI
  • QA: Mamba tests restored
  • QA: Pre-commit hooks with ruff and mypy
  • QA: Dependabot config for pip and GitHub Actions
  • CHANGELOG.md initialized

Changed

  • Dependencies: numpy constraint relaxed to >=1.26.0 (supports 1.x and 2.x)
  • Dependencies: transformers upper bound removed (>=4.49.0)
  • Dependencies: torch upper bound relaxed from <=2.7 to <2.12 for RTX 5090 support
  • README: Installation instructions moved to prominent position near the top

Fixed

  • Architecture: FocalLoss moved to module level for proper importability
  • Architecture: .flake8 removed, fully migrated to ruff for linting
  • Architecture: mypy ignore list pruned for stub-covered packages
  • Type safety: load_model_and_tokenizer return type changed from tuple[Any, Any] to tuple[PreTrainedModel, PreTrainedTokenizer]

Removed

  • .flake8 configuration file

[0.4.0] - 2025-12-01

Overview

Last stable release before the 0.5.x development cycle. Provided core fine-tuning and inference capabilities for DNA language models with support for multiple model architectures (DNABERT2, Nucleotide Transformer, GPN, HyenaDNA, etc.).

Added

  • Core fine-tuning pipeline with DNATrainer supporting classification, regression, and masked language modeling
  • Multi-model architecture support via AutoModel integration
  • Basic inference engine with batch processing
  • Model zoo with 20+ pre-trained DNA language models

Changed

  • Initial project structure and package layout

Fixed

  • Various stability improvements for model loading and tokenization