## cactus-compute/cactus — v2.0.1…v2.1.0

_98 commits._

### Features
- **New run operations with runtime plans** (aa06bf6)
- **New run operations with runtime plans** (3b0b9cf)
- **Further LFM support** (ddcbcc8)
- **Minor modifications to converter and input processing to support MOE models** (80243d3)
- **Modifications to ModelProfiles classes to add additional cache and input meta-data** (26e53ba)
- **add probe during gemma-4-e2b-it conversion step** (1ace6d7)
- **add tool calling support to run env (#766)** (df4482a)
- **add needle support (#765)** (0fd8a93)
- **add wider cq support (#763)** (49e1256)
- **Reset architecture to support harder fusions** (6961ed4)
- **Added a new field to ValueNodes to keep track of value users** (7a5c750)

### Fixes
- **fix: use per-matrix CQ4 reciprocals in Metal batch transform (#784)** (22b86a6)
- **Fix chunk retargeting of attention broadcasts** (2b95286)
- **Fix formatting in README.md** (785cd06)

### Backend
- **Bucket the Parakeet audio encoder (#791)** (ba2ee75)
- **Include <iomanip> directly for std::setw/setfill/setprecision usage (#782)** (ea7948c)
- **Merge pull request #778 from cactus-compute/custom-transpiler** (c90a1af)
- **Visual chagnes** (8226cfe)
- **Code cleanup** (e9773b1)
- **Code simplifications** (4abf717)
- **Tune CQ GEMV decode scheduling** (dc1a58d)
- **Optimize contiguous last-axis padding** (9e69e03)
- **Merge main into custom-transpiler** (1ac9d72)
- **Code reduction** (ba1e98b)
- **Kernel optmizations** (ee89a96)
- **code reduction** (e4f0d3e)
- **Working optmized transpiler and runtime + advancements to shared kv caches and states.** (9a6b4dc)
- **Memory optmizations** (449363f)
- **Working main models** (1d90b40)
- **Performance upgrades** (fb6176c)
- **Code reformatting** (4c6733b)
- **Performace upgrades** (909da19)
- **performance upgrades** (e78aeca)
- **Performance upgrades** (b65b667)
- **Added runtime plan** (e07c158)
- **LFM convs** (1d1837a)
- **Additional conv-state matching logic** (fb5963e)
- **Added cactus ops lowering with generator** (6db22b2)
- **optimize attention and gemv kernels (#776)** (8e0eec1)
- **Added simplification of IR step (uses fusion matchers to determine fusion)** (3a51ea5)
- **Added simplification of IR step (uses fusion matchers to determine fusion)** (f97178c)
- **LFM with cache (prefill + decode)** (9083766)
- **Additional matchers for lfm MOE** (0cdd3d6)
- **Added LFM MOE model profile** (ae567c6)
- **Additional fusions + nodes for MOE ops** (8aa7672)
- **Additional extra matchers for RoPE, linear projs, and LSTMs (yet to be confirmed)** (610d3b4)
- **Added extra matchers for extraneous constraints for fusions** (841bdae)
- **Added fusion definition matchers** (aa231af)
- **Simplifications to fusion matcher functions** (344edc9)
- **Added graph matchers** (53c759e)
- **Simplification of logic** (1e5297b)
- **Converter testing** (338b5a5)
- **Matching nodes** (46ecc0d)
- **Future instructions** (9c7ed82)
- **Added fusion results object** (04b9057)
- **Initial graph classes** (8852af1)
- **Added fusion patterns** (de8514e)
- **Removed unecessary code** (72e6874)
- **Converter input_utils removal + input loading code simplification** (1c4a1f0)
- **Code refactoring** (198f46c)
- **Converter reorganization** (052a2ba)
- **Updated required files for Gemma** (af0797f)
- **Reorganizing + optimizing converter** (0bf9dc7)
- **Cleaned out convert.py** (c886f1c)
- **Streamlining convertert logic + code; still need to clean up** (21fa95a)
- **Brute-force code for converted jsons** (c0ff8a0)
- **Added input loading utilities** (1b6f44d)
- **Load files function** (a0530e1)
- **Load files function** (8d86c94)
- **Progress to convert** (5898e7e)
- **Added supported modalities to model profiles class** (391b6fd)
- **Removed redundant Files class + modified existing model profiles** (a6311c5)
- **Removed redundant objects and refactored class structure** (3333bb1)
- **Added converter logic** (7fd1cfc)
- **Added qwen2.5-0.5b profile** (d54bda5)
- **Added additional profiles** (4d0e6a6)
- **Model profiles** (a74d3fa)
- **Folder restructuring** (386ff84)
- **File restructure** (802b326)
- **M3 pro bench results (#762)** (0969edb)
- **Revise model performance metrics in README** (1e121d3)
- **Reset** (a4c7af6)
- **Changes to transpiler converter** (edd39a5)
- **Added matchers which will check for synthetic graph and main graph equality (utilized to detect valid fusions)** (e6baa92)
- **DAG based fusion: Initial models refactoring** (c9876a8)
- **Added simplified graph representation + topo sort on simplified graph** (ea0207a)
- **Added final fused graph clss** (6b2069b)
- **Added FusionResult object for fusions** (e73b35c)
- **Updated fusion matching logic** (5d66cdd)
- **Fusion matching logic + attributions cleaning** (ecdfdaf)
- **Updates to fusions base object** (88d9446)
- **improving ir** (c23a00c)
- **Completed computation graph object** (f8b7ae3)
- **Inital IR simplification logic** (09cb668)
- **Initial converter complete: convers text-to-text model graphs to JSON** (cfaa176)
- **Initial converter progress** (a9fcf3f)
- **Removed existing transpiler** (a7f75f1)

### Docs
- **docs: fix "HuggigFace" typo in README (#746)** (1c2c91c)

_Recap by [Repo Wrapped](https://repowrapped.com/gh/cactus-compute/cactus?utm_source=github-action)._