tag:github.com,2008:https://github.com/Overworldai/world_engine/releases 2026-05-20T15:59:31Z tag:github.com,2008:Repository/1100875802/1.5.6 2026-05-20T15:59:31Z lapp0 tag:github.com,2008:Repository/1100875802/1.5.5 2026-04-30T20:48:32Z lapp0 tag:github.com,2008:Repository/1100875802/1.5.4 2026-04-23T17:53:44Z <p>MoE Support (<a class="issue-link js-issue-link" href="https://github.com/Overworldai/world_engine/pull/47">#47</a>)</p> <p>* implement state loading / saving</p> <p>* moe + fbgemm optimization</p> <p>* wp-1.5 staging</p> <p>* clean up and fix ae</p> <p>* fix temporal compression rope bugs</p> <p>* vae reset in world_engine.reset</p> <p>* reduce peak memory</p> <p>* Implements the orthorope angles computation instead of precomputing (<a class="issue-link js-issue-link" href="https://github.com/Overworldai/world_engine/pull/25">#25</a>)</p> <p>* fix: uv sync issue with python version 3.9</p> <p>* fix: VRAM explosion</p> <p>* refactor: init on gpu device directly</p> <p>* fix: don't use fbgemm on windows for now</p> <p>* feat: orthoropeangles</p> <p>* fix: NoCastModule OrthoRoPEAngles</p> <p>* fix: remove pos_ids from args</p> <p>* fix: remove old src rope replacement patch</p> <p>* fix: remove out of scope ae changes</p> <p>* fix: remove out of scope text encoder changes</p> <p>* fix: patch_model pos_ids</p> <p>---------</p> <p>Co-authored-by: Philpax &lt;me@philpax.me&gt;</p> <p>* test: revert direct device init (<a class="issue-link js-issue-link" href="https://github.com/Overworldai/world_engine/pull/28">#28</a>)</p> <p>* feat: use built triton-windows fork to fix long-path issue</p> <p>* update gen_sample</p> <p>* better quant</p> <p>* avoid warning when creating mouse / scroll tensors</p> <p>* disable unimportant compile options</p> <p>* clean up model loading</p> <p>* remove unnecessary push_to_hub</p> <p>* remove unnecessary save_pretrained</p> <p>* reduce cpu memory</p> <p>* pass device</p> <p>* moe wip</p> <p>* default quant=None for benchmarking</p> <p>* moe support</p> <p>* remove deprecated benchmark_moe</p> <p>* remove blocking path benchmark</p> <p>* remove flashinfer path</p> <p>* fix moe loading, expert weighing, refactor MoE class</p> <p>---------</p> <p>Co-authored-by: Clydingus &lt;40514241+Clydingus@users.noreply.github.com&gt; <br />Co-authored-by: Philpax &lt;me@philpax.me&gt;</p> lapp0 tag:github.com,2008:Repository/1100875802/1.5.3 2026-04-09T01:04:54Z lapp0 tag:github.com,2008:Repository/1100875802/1.5.2 2026-04-09T00:58:37Z lapp0 tag:github.com,2008:Repository/1100875802/1.5.1 2026-04-08T19:35:35Z <p>Excluding layers from quantization (<a class="issue-link js-issue-link" href="https://github.com/Overworldai/world_engine/pull/36">#36</a>)</p> <p>* exclude layers from quant using fqn</p> <p>* strict checks</p> anm-ol tag:github.com,2008:Repository/1100875802/1.5.0 2026-04-08T14:37:03Z <p>Adds support for int8 w8a8_gemlite quantization (<a class="issue-link js-issue-link" href="https://github.com/Overworldai/world_engine/pull/34">#34</a>)</p> <p>* add torchao quantize_</p> <p>* testing</p> <p>* testing yes</p> <p>* use taehv overide</p> <p>* yuh</p> <p>* add apply qat</p> <p>* yuh</p> <p>* uh</p> <p>* enable int4 benchmarking and inference</p> <p>* apply quantize_model w8a8</p> <p>* add int8 ptq</p> <p>* quant none</p> <p>* int8 gemlite implementation</p> <p>* clean up, remove torchao quantization</p> <p>* add gemlite to requirements</p> <p>* remove unused quant kernels and imports</p> <p>* restore gen_sample.py, more cleanup</p> <p>* update readme with Quantization docs</p> <p>* fixed requirements gemlite</p> <p>* Clean up pyproject.toml and add config defaults to base_model.py</p> <p>* Add gemlite warmup+cache, and update gemlite version</p> <p>* cleanup pyproject.toml and resize in examples</p> anm-ol