·
3 commits
to master
since this release
Automated Unsloth llama.cpp CUDA + ROCm + Vulkan + macOS + CPU prebuild for upstream b10715, merged with:
- Carry ggml-org#24423 (DiffusionGemma) merged onto b10630 (unslothai/llama.cpp#107, commit 74acc40)
- Add TML Inkling architecture (#25731, commit 44eb88e)
- kimi-k3 : the MoonViT-3d vision tower and full-size loading fixes (unslothai/llama.cpp#70, commit edfd4c1)
- IQ1_XS, IQ1_XXS, IQ1_XXXS on an upstream base so the nightly can pin them (unslothai/llama.cpp#91, commit c86ed26)
- sampling: index penalties by token id instead of scanning every candidate (unslothai/llama.cpp#95, commit 3db8cb5)
- model: add GLM-5-Next (GLM-5.3-Flash) (#27754, commit 949f7ef)
- llama: batched readahead for lazily read gather tables (unslothai/llama.cpp#137, commit 4e1865e)
- ggml-cuda: avoid direct ROCm_Host compute on HIP integrated GPUs (port of ggml-org#25863) (unslothai/llama.cpp#158, commit abfc45b)
- ggml-cuda: use cudaMemcpyDefault in the ggml_cuda_cpy 2D fast path (unslothai/llama.cpp#157, commit 6c6da89)
- ggml-cuda: make GGML_CUDA_ENABLE_UNIFIED_MEMORY=0 actually disable it, and say so on HIP (unslothai/llama.cpp#149, commit b65a2dc)
- qwen4exp: follow up fixes (#27941, commit 6b2b85c)
- MTP for Qwen3.8-Flash-Next (unslothai/llama.cpp#144, commit 586b15e)
- llama: map each contiguous run of a context's tensors, not one span over all of them (unslothai/llama.cpp#152, commit 258345e)
- kv-cells: hoist the sequence list out of the for_each_token_in cell loop (unslothai/llama.cpp#154, commit 31e432e)