bump: version 2.0.0b31 → 2.0.0b32

Author: github-actions[bot]

CHANGELOG.md

@@ -1,3 +1,36 @@
+## v2.0.0b32 (2026-04-17)
+
+
+- Merge pull request #217 from OpenMOSS/dev
+- Release: refactor TransformerLensLanguageModel to be inherited from HookedTransformer
+- fix(backend): cast encoding['input_ids'] to Tensor for basedpyright
+- fix(server): basedpyright/ruff errors from backend refactor
+- - Add TokenizerOnlyLanguageModel to __all__ so ruff sees the export.
+- Assert tokenizer is not None before chat-template paths in circuit
+  preview + generate (model.tokenizer is Optional on the new API).
+- Cast apply_chat_template result to str (tokenize=False is known by
+  us but not inferable from the overload union).
+- fix(server): offload sync @distributed calls to thread in host-execution mode
+- Under num_workers=0, the @distributed wrapper ran the target function
+synchronously on the asyncio event loop. Long-running torch work (circuit
+attribution, model forwards) blocked the loop, so parallel frontend
+requests (progress polling, sub-resource loads) hung until the heavy call
+finished. Route sync functions through asyncio.to_thread so the loop
+stays responsive; async functions are awaited directly.
+- refactor(backend): TransformerLensLanguageModel via multiple inheritance
+- Switch from composition (`self.model = HookedTransformer(...)`) to
+multiple inheritance (`class TransformerLensLanguageModel(HookedTransformer,
+LanguageModel)`), exposing the full TL API directly without per-method proxies.
+- - Rename our `self.cfg` → `self.lm_cfg` to avoid clashing with
+  HookedTransformerConfig.
+- Drop `use_flash_attn`, `load_ckpt`, `tokenizer_only` config fields.
+- Replace the `tokenizer_only=True` flag with a dedicated
+  `TokenizerOnlyLanguageModel` backend (`backend="tokenizer_only"`).
+- Add `from_hooked_transformer` classmethod to upgrade bare HookedTransformer
+  instances via zero-copy __class__ swap.
+- Update all call sites (`model.model.blocks` → `model.blocks`, etc.) across
+  circuits, initializer, analysis, server, and CLI.
+
 ## v2.0.0b31 (2026-04-17)
 
 

README.md

@@ -35,13 +35,13 @@
 Use [pip](https://pypi.org/project/pip/) to install Language-Model-SAEs:
 
 ```bash
-pip install lm-saes==2.0.0b31
+pip install lm-saes==2.0.0b32
 ```
 
 We also highly recommend using [uv](https://docs.astral.sh/uv/) to manage your own project dependencies. You can use
 
 ```bash
-uv add lm-saes==2.0.0b31
+uv add lm-saes==2.0.0b32
 ```
 
 to add Language-Model-SAEs as your project dependency.

docs/index.md

@@ -24,7 +24,7 @@ This library provides:
     To add our library as a project dependency, run:
 
     ```bash
-    uv add lm-saes==2.0.0b31
+    uv add lm-saes==2.0.0b32
     ```
 
     We also support [Ascend NPU](https://github.com/Ascend/pytorch) as an accelerator backend. To add our library as a project dependency with NPU dependency constraints, run:
@@ -38,7 +38,7 @@ This library provides:
     Of course, you can also directly use [pip](https://pypi.org/project/pip/) to install our library. To install our library with pip, run:
 
     ```bash
-    pip install lm-saes==2.0.0b31
+    pip install lm-saes==2.0.0b32
     ```
 
     We also support [Ascend NPU](https://github.com/Ascend/pytorch) as an accelerator backend. To install our library with NPU dependency constraints, run:

pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "lm-saes"
-version = "2.0.0b31"
+version = "2.0.0b32"
 description = "For OpenMOSS Mechanistic Interpretability Team's Sparse Autoencoder (SAE) research. Open-sourced and constantly updated."
 dependencies = [
     "transformer-lens>=2.16.2",