typed-lm 0.1.1

0.1.1

Bug Fixes

• serve the model as typed-lm, never a jev- alias
• docker: make the container images runnable

Continuous Integration

• generate release notes from commits with git-cliff
• install cublas, curand and nvrtc for the CUDA build
• install the full CUDA toolkit for the CUDA build
• add the documentation deploy workflow

Documentation

• redesign the README with the book theme
• add the sk-hero banner to the README
• highlight training and add the Docker quickstart

Other Changes

• use crates.io-valid category slugs
• release 0.1.1

typed-lm 0.1.0

0.1.0

Bug Fixes

• use public checkpoint for live test and default model
• common: resolve clippy warnings in MXFP4 quantization
• preserve FP4 shape/element count and map unknown model to 404
• adapt existing config constructors and trainer dispatch to the extended dense config
• route the trainer through the unified dense loader

Continuous Integration

• add release automation, packaging metadata and refresh dependencies

Documentation

• explain HF_TOKEN requirement for gated models
• write project README
• add example.http REST client calls

Features

• add Candle evaluator with Rig-orchestrated prompts
• expose Jev HTTP API via actix-web scope handlers
• wire serve command with model, context and examples
• route HF token through CLI with env fallback
• add environment fallback for all CLI options
• make memory context optional
• multi-format Llama/Qwen2 inference with session-cache reuse
• common: W1 01-tokenizer
• common: W1 02-device
• common: W1 03-checkpoint
• common: W1 04-quantization
• common: W1 05-resolver treats FP8/FP4 checkpoints as loadable
• common: extend architecture enum and config to all dense families

Miscellaneous

• ignore .worktrees/ and keep temporary/ in the gitignore
• baseline for multi-architecture and from-scratch plan
• add the toml dependency for configuration files

Other Changes

• devcontainer: install the CUDA toolkit and request the GPU
• w0-workspace-foundation
• wave 0: workspace foundation (typed-lm-common/serve/trainer)
• task/01-01-tokenizer-tokenizer
• task/01-02-device-device
• task/01-03-checkpoint-checkpoint
• task/01-04-quantization-quantization
• task/01-05-resolver-resolver
• wave 1: typed-lm-common new modules (tokenizer, precision, fp8/fp4, resolver)
• wave 2: serve dequant-on-load for FP8/FP4 and accept low-precision checkpoints
• wave 3: trainer dataset discovery, records, loader and decision-position collation
• wave 4: trainer trainable model (precision, LoRA, Llama/Qwen2 forward, frozen weight loading)
• wave 5: trainer losses, AdamW schedule/clip/accumulation, adapter checkpoints and epoch loop
• wave 6: trainer CLI, train/quantize dispatch, PTQ export and end-to-end tests
• wave 7: workspace docs, devcontainer paths, VS Code launch config and trainer README
• wave 8: final validation and remove abbreviated identifiers in tests
• wave 9: harden FP4 artifact detection and add serving artifact loader tests
• extend trainer CLI with full/from-scratch methods, geometry, seed and config file
• extend trainer CLI (full/from-scratch, geometry, seed, configuration-file)
• add typed configuration-file and initialization error variants
• typed configuration-file and initialization error variants
• map every dense model_type and reject MoE/MLA actionably
• map every dense model_type and reject MoE/MLA actionably
• parameterize the dense serving forward for all families
• parameterize the dense serving forward for all families
• parameterize the differentiable dense forward for all families
• parameterize the differentiable dense forward for all families
• dispatch dense serving across all supported families
• dispatch dense serving across all supported families
• unified dense trainable loader replacing the Qwen2 alias
• unified dense trainable loader
• deterministic seeded weight initializer for from-scratch training
• deterministic seeded weight initializer
• fully-trainable linear layer for full-parameter training
• fully-trainable linear layer
• fully-trainable RMSNorm with Gemma unit offset
• fully-trainable RMSNorm
• save and load a complete dense checkpoint with config and tokenizer
• full checkpoint save/load
• wire from-scratch initialization and trainable primitives into the model tree
• TOML configuration file schema, parsing and typed errors
• TOML configuration file schema, parsing and typed errors
• resolve configuration with CLI over TOML over default precedence
• resolve configuration with CLI over TOML over default precedence
• integrate the configuration file into the CLI resolution
• wire full-parameter and from-scratch training into the trainer entry point
• cover every dense architecture's weight names in the quantized export
• cover every dense architecture's weight names in the quantized export
• from-scratch to quantize to serve end-to-end over HTTP
• from-scratch to quantize to serve end-to-end over HTTP
• document the supported dense architectures and MoE/MLA rejection
• docs (readme)
• document full/from-scratch training and the TOML configuration reference
• docs (docs-config)
• document the new trainer methods/flags and servable from-scratch artifact
• docs (trainer-readme)
• update AGENTS.md for the dense families, training methods and TOML config
• docs (agents)
• workspace fmt/clippy cleanup for the multi-architecture work
• model the Gemma/Qwen3 forward capabilities in the shared config
• correct the vendored serving forward for Gemma/Qwen3
• correct the trainable forwards and initializer for Gemma/Qwen3
• emit serveable configs and honor the TOML initializers
• add per-family forward parity tests and stabilize the live cache test
• route all logging through tracing and require devcontainer GPU runs
• document the new geometry flags and optional dataset
• remove the misplaced from-scratch plan and migrate its facts into docs
• enforce English-only repository content and drop the Sciencekit mention
• translate the remaining Portuguese documentation to English

Performance

• add Intel MKL support with an F16 GEMM shim

Refactoring

• split api module into scoped submodules
• move serve wiring out of main
• w0: split manaca-typed into typed-lm workspace

Tests

• exercise the FP4 *_shape metadata in the serving artifact test

Full Changelog: https://github.com/neurono-ml/typed-lm/commits/0.1.0