[hijohnnylin]

Summary

New features for every backend: the lens in the engine, steering in generation and capture, reads of an activation through a LoRA adapter, and a single-model HTTP server. This is a minor release. Nothing is removed. The old APIs that changed remain as deprecated shims.

Added

• Lens
  • generate_with_lens() on every backend.
  • Named J_bar sets: two Jacobian lenses (for example J and J++) can be loaded together.
  • A faster lens read-out, and a live residual feed.
• Steering
  • Steer at more than one point with a list of specs (Steering).
  • New ops: NormScaledAddSpec, AblateSpec, SwapSpec.
  • steer() blocks can skip positions, and they reach generation and capture on vLLM too.
  • SteerMethod, steering_from_directions, and the TransformerLens steer mapping.
• LoRA reads: a PEFT adapter reads one residual vector out as text, on eager and vLLM. The text for each layer can stream.
• Protocol additions (every backend)
  • describe().
  • A per-token generation stream, and the engine's own tokenize.
  • unembed_rows.
  • generate_steps_from_embeds, and sample_from_embeds (n completions of one embeds prompt in one request).
  • A check for whether the model serves a point.
  • The sampling recommendation from generation_config.json.
• Capture: capture() is the new name, and run_with_cache stays as an alias. Also new:
  • capture(rows=...);
  • correct positions for padded rows;
  • last-token capture;
  • capture on the speculative engine and on compiled decode steps.
• HTTP server: python -m interp_engine.server serves one model, with /v1/project and /v1/lens.
• API manifest: api/engine.yaml generates interp_engine/api.py, with conformance fixtures.

Fixed

• A vLLM worker that runs out of memory now fails only that request, not the engine.
• release() frees the engines.
• select_backend names the backend. It refuses a backend that the machine cannot run before it loads anything.
• Clearer eager refusals for quantized checkpoints, and for FP8 on transformers 5.11 to 5.16.
• Tokenize span fixes, and HubMirror read-through.

Deprecated

These still work and give a DeprecationWarning. They will be removed in a later major release.

• VLLMModel.generate_from_embeds(prompt_embeds, sampling_params): use generate_steps_from_embeds or sample_from_embeds.
• lens_intervention= on capture_generation, capture_generation_stream and lens_capture_readout_stream: use a steer() block with NormScaledAddSpec / AblateSpec / SwapSpec.

Tooling

• Validator: Qwen3.5-0.8B and gemma-4-E2B added. The result headings show the engine version.
• GPU sizer: a speculative drafter field.
• Visualizer: a point card.
• The server extra is in the lock file.

Testing

• CPU: 2641 passed. Validator: 541 passed. ruff, pyright, the API manifest check and the visualizer checks pass.
• GPU (A100 80GB): the vLLM GPU tests pass, except for 2 failures that also occur before this change. On a production-config pod, tokenize, J and J++ lens, LoRA read and steering gave the same answers as the current production engine.