What's Changed
- support Wan models for video generation by @zhongkaifu in #133
- fix a vulkan build break by @zhongkaifu in #135
- fix ggml_metal build break by @zhongkaifu in #136
- Feature/support muse glimmer model inference by @zhongkaifu in #137
- update documents for muse glimmer model by @zhongkaifu in #138
- optimize wan performance in ggml_metal backend by @zhongkaifu in #139
- optimize muse glimmer model performance by @zhongkaifu in #140
- Fix ggml master build break, mixed-quant UD GGUFs, and a ~27x Vulkan slowdown on small-BAR AMD GPUs by @craig-b in #141
- Fix out-bool marshalling of mlx_array_is_row_contiguous by @craig-b in #143
- Marshal native-bound strings as UTF-8 instead of CharSet.Ansi by @craig-b in #144
- Update .gitignore to avoid extra files added to the commit. by @vvdb-architecture in #145
- Confine /api/chat attachment paths to the upload directory by @craig-b in #146
- Clamp generation reserve to remaining context by @craig-b in #147
- Map prompt-context overflow to HTTP 400 instead of 500 by @craig-b in #148
- Feature/fix qwen38 issues by @zhongkaifu in #149
- Restrict uploads to an extension allow-list with safe serve content types by @craig-b in #150
- Stop exposing host filesystem paths via /api/models and the 404 fallback by @craig-b in #151
- Confine image-edit and video imagePath to the upload directory correctly by @craig-b in #152
- fix json parser by @zhongkaifu in #155
- Trait-based test lanes: filterable inner loop, and PR CI on a stock runner by @craig-b in #156
- optimize wan performance by @zhongkaifu in #154
- Report environment-gated tests as skipped instead of silently passing by @craig-b in #157
- Keep a serving model when a runtime reload fails by @craig-b in #158
- Reference uploads by server filename instead of absolute host path by @craig-b in #159
- Add --no-webui / TS_NO_WEBUI to run the server API-only by @craig-b in #160
- support glm model by @zhongkaifu in #163
- Add opt-in storage limits for the upload directory by @craig-b in #161
- support GLM MTP by @zhongkaifu in #165
- Migrate DllImport P/Invokes to LibraryImport by @craig-b in #162
- optimize wan performance by @zhongkaifu in #167
- Feature/optimize wan by @zhongkaifu in #168
- optimize GLM MTP by @zhongkaifu in #169
- fix build errors by @zhongkaifu in #170
- refactor specultive decoding by @zhongkaifu in #171
- This corrects a problem of resource cleanup (file deletion) in unit tests. by @vvdb-architecture in #172
- Feature/support dflash2 by @zhongkaifu in #175
- refactoring by @zhongkaifu in #176
New Contributors
- @craig-b made their first contribution in #141
- @vvdb-architecture made their first contribution in #145
Full Changelog: v3.2.1.0...v3.3.0.0
What's Changed
- support Wan models for video generation by @zhongkaifu in #133
- fix a vulkan build break by @zhongkaifu in #135
- fix ggml_metal build break by @zhongkaifu in #136
- Feature/support muse glimmer model inference by @zhongkaifu in #137
- update documents for muse glimmer model by @zhongkaifu in #138
- optimize wan performance in ggml_metal backend by @zhongkaifu in #139
- optimize muse glimmer model performance by @zhongkaifu in #140
- Fix ggml master build break, mixed-quant UD GGUFs, and a ~27x Vulkan slowdown on small-BAR AMD GPUs by @craig-b in #141
- Fix out-bool marshalling of mlx_array_is_row_contiguous by @craig-b in #143
- Marshal native-bound strings as UTF-8 instead of CharSet.Ansi by @craig-b in #144
- Update .gitignore to avoid extra files added to the commit. by @vvdb-architecture in #145
- Confine /api/chat attachment paths to the upload directory by @craig-b in #146
- Clamp generation reserve to remaining context by @craig-b in #147
- Map prompt-context overflow to HTTP 400 instead of 500 by @craig-b in #148
- Feature/fix qwen38 issues by @zhongkaifu in #149
- Restrict uploads to an extension allow-list with safe serve content types by @craig-b in #150
- Stop exposing host filesystem paths via /api/models and the 404 fallback by @craig-b in #151
- Confine image-edit and video imagePath to the upload directory correctly by @craig-b in #152
- fix json parser by @zhongkaifu in #155
- Trait-based test lanes: filterable inner loop, and PR CI on a stock runner by @craig-b in #156
- optimize wan performance by @zhongkaifu in #154
- Report environment-gated tests as skipped instead of silently passing by @craig-b in #157
- Keep a serving model when a runtime reload fails by @craig-b in #158
- Reference uploads by server filename instead of absolute host path by @craig-b in #159
- Add --no-webui / TS_NO_WEBUI to run the server API-only by @craig-b in #160
- support glm model by @zhongkaifu in #163
- Add opt-in storage limits for the upload directory by @craig-b in #161
- support GLM MTP by @zhongkaifu in #165
- Migrate DllImport P/Invokes to LibraryImport by @craig-b in #162
- optimize wan performance by @zhongkaifu in #167
- Feature/optimize wan by @zhongkaifu in #168
- optimize GLM MTP by @zhongkaifu in #169
- fix build errors by @zhongkaifu in #170
- refactor specultive decoding by @zhongkaifu in #171
- This corrects a problem of resource cleanup (file deletion) in unit tests. by @vvdb-architecture in #172
- Feature/support dflash2 by @zhongkaifu in #175
- refactoring by @zhongkaifu in #176
New Contributors
- @craig-b made their first contribution in #141
- @vvdb-architecture made their first contribution in #145
Full Changelog: v3.2.1.0...v3.3.0.0
Contributors
craig-b, zhongkaifu, and vvdb-architecture
Assets 12
0
Join discussion