llama.cpp prebuilt b11160-mix-a6922cc

Automated Unsloth llama.cpp CUDA + ROCm + Vulkan + macOS + CPU prebuild for upstream b11160, merged with:

• DiffusionGemma (#24423, commit 3dac51d)
• Add TML Inkling architecture (#25731, commit 3870aa4)
• kimi-k3 : the MoonViT-3d vision tower and full-size loading fixes (unslothai/llama.cpp#70, commit 883f2c9)
• IQ1_XS, IQ1_XXS, IQ1_XXXS: three quant types below IQ1_S (unslothai/llama.cpp#61, commit 46cbf0e)
• 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 86ebfef)
• 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)
• MTP for Qwen3.8-Flash-Next (unslothai/llama.cpp#144, commit ca14269)
• llama: map each contiguous run of a context's tensors, not one span over all of them (unslothai/llama.cpp#152, commit b2b5ed9)
• mtmd: test that every projector is registered and uniquely named (unslothai/llama.cpp#176, commit 09ce1a4)

llama.cpp prebuilt b11139-mix-a6922cc

Automated Unsloth llama.cpp CUDA + ROCm + Vulkan + macOS + CPU prebuild for upstream b11139, merged with:

• DiffusionGemma (#24423, commit 3dac51d)
• Add TML Inkling architecture (#25731, commit 3870aa4)
• kimi-k3 : the MoonViT-3d vision tower and full-size loading fixes (unslothai/llama.cpp#70, commit 883f2c9)
• IQ1_XS, IQ1_XXS, IQ1_XXXS: three quant types below IQ1_S (unslothai/llama.cpp#61, commit 46cbf0e)
• 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 86ebfef)
• 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)
• MTP for Qwen3.8-Flash-Next (unslothai/llama.cpp#144, commit ca14269)
• llama: map each contiguous run of a context's tensors, not one span over all of them (unslothai/llama.cpp#152, commit b2b5ed9)
• mtmd: test that every projector is registered and uniquely named (unslothai/llama.cpp#176, commit 09ce1a4)

llama.cpp prebuilt b11115-mix-a6922cc

Automated Unsloth llama.cpp CUDA + ROCm + Vulkan + macOS + CPU prebuild for upstream b11115, merged with:

• DiffusionGemma (#24423, commit 3dac51d)
• Add TML Inkling architecture (#25731, commit 3870aa4)
• kimi-k3 : the MoonViT-3d vision tower and full-size loading fixes (unslothai/llama.cpp#70, commit 883f2c9)
• IQ1_XS, IQ1_XXS, IQ1_XXXS: three quant types below IQ1_S (unslothai/llama.cpp#61, commit 46cbf0e)
• 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 86ebfef)
• 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)
• MTP for Qwen3.8-Flash-Next (unslothai/llama.cpp#144, commit ca14269)
• llama: map each contiguous run of a context's tensors, not one span over all of them (unslothai/llama.cpp#152, commit b2b5ed9)
• mtmd: test that every projector is registered and uniquely named (unslothai/llama.cpp#176, commit 09ce1a4)

llama.cpp prebuilt b11030-mix-5ff778e

Automated Unsloth llama.cpp CUDA + ROCm + Vulkan + macOS + CPU prebuild for upstream b11030, merged with:

• Qwen3.8-Flash-Next MTP Fix (2x faster now)
• DiffusionGemma (#24423, commit 12e0a96)
• Add TML Inkling architecture (#25731, commit 946fc11)
• kimi-k3 : the MoonViT-3d vision tower and full-size loading fixes (unslothai/llama.cpp#70, commit 883f2c9)
• IQ1_XS, IQ1_XXS, IQ1_XXXS: three quant types below IQ1_S (unslothai/llama.cpp#61, commit 46cbf0e)
• 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 86ebfef)
• 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)
• MTP for Qwen3.8-Flash-Next (unslothai/llama.cpp#144, commit ca14269)
• llama: map each contiguous run of a context's tensors, not one span over all of them (unslothai/llama.cpp#152, commit b2b5ed9)
• mtmd: test that every projector is registered and uniquely named (unslothai/llama.cpp#176, commit 09ce1a4)

llama.cpp prebuilt b11027-mix-3e83366

Automated Unsloth llama.cpp CUDA + ROCm + Vulkan + macOS + CPU prebuild for upstream b11027, merged with:

• DiffusionGemma (#24423, commit 12e0a96)
• Add TML Inkling architecture (#25731, commit 946fc11)
• kimi-k3 : the MoonViT-3d vision tower and full-size loading fixes (unslothai/llama.cpp#70, commit 883f2c9)
• IQ1_XS, IQ1_XXS, IQ1_XXXS: three quant types below IQ1_S (unslothai/llama.cpp#61, commit 46cbf0e)
• 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 86ebfef)
• 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)
• MTP for Qwen3.8-Flash-Next (unslothai/llama.cpp#144, commit f40f64a)
• llama: map each contiguous run of a context's tensors, not one span over all of them (unslothai/llama.cpp#152, commit b2b5ed9)
• mtmd: test that every projector is registered and uniquely named (unslothai/llama.cpp#176, commit 09ce1a4)

llama.cpp prebuilt b11007-mix-3e83366

Automated Unsloth llama.cpp CUDA + ROCm + Vulkan + macOS + CPU prebuild for upstream b11007, merged with:

• DiffusionGemma (#24423, commit 12e0a96)
• Add TML Inkling architecture (#25731, commit 946fc11)
• kimi-k3 : the MoonViT-3d vision tower and full-size loading fixes (unslothai/llama.cpp#70, commit 883f2c9)
• IQ1_XS, IQ1_XXS, IQ1_XXXS: three quant types below IQ1_S (unslothai/llama.cpp#61, commit 46cbf0e)
• 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 86ebfef)
• 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)
• MTP for Qwen3.8-Flash-Next (unslothai/llama.cpp#144, commit f40f64a)
• llama: map each contiguous run of a context's tensors, not one span over all of them (unslothai/llama.cpp#152, commit b2b5ed9)
• mtmd: test that every projector is registered and uniquely named (unslothai/llama.cpp#176, commit 09ce1a4)

llama.cpp prebuilt b10995-mix-3e83366

Automated Unsloth llama.cpp CUDA + ROCm + Vulkan + macOS + CPU prebuild for upstream b10995, merged with:

• DiffusionGemma (#24423, commit 12e0a96)
• Add TML Inkling architecture (#25731, commit 946fc11)
• kimi-k3 : the MoonViT-3d vision tower and full-size loading fixes (unslothai/llama.cpp#70, commit 883f2c9)
• IQ1_XS, IQ1_XXS, IQ1_XXXS: three quant types below IQ1_S (unslothai/llama.cpp#61, commit 46cbf0e)
• 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 86ebfef)
• 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)
• MTP for Qwen3.8-Flash-Next (unslothai/llama.cpp#144, commit f40f64a)
• llama: map each contiguous run of a context's tensors, not one span over all of them (unslothai/llama.cpp#152, commit b2b5ed9)
• mtmd: test that every projector is registered and uniquely named (unslothai/llama.cpp#176, commit 09ce1a4)

llama.cpp prebuilt b10909-mix-bea84f7

Automated Unsloth llama.cpp CUDA + ROCm + Vulkan + macOS + CPU prebuild for upstream b10909, merged with:

• DiffusionGemma (#24423, commit 168b21a)
• Add TML Inkling architecture (#25731, commit 946fc11)
• kimi-k3 : the MoonViT-3d vision tower and full-size loading fixes (unslothai/llama.cpp#70, commit 883f2c9)
• IQ1_XS, IQ1_XXS, IQ1_XXXS: three quant types below IQ1_S (unslothai/llama.cpp#61, commit 46cbf0e)
• 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 d94f44e)
• 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)
• MTP for Qwen3.8-Flash-Next (unslothai/llama.cpp#144, commit a9e9c3c)
• llama: map each contiguous run of a context's tensors, not one span over all of them (unslothai/llama.cpp#152, commit b2b5ed9)
• mtmd: test that every projector is registered and uniquely named (unslothai/llama.cpp#176, commit 09ce1a4)

llama.cpp prebuilt b10840-mix-d5c17a0

Automated Unsloth llama.cpp CUDA + ROCm + Vulkan + macOS + CPU prebuild for upstream b10840, merged with:

• DiffusionGemma (#24423, commit 168b21a)
• Add TML Inkling architecture (#25731, commit 1066edc)
• kimi-k3 : the MoonViT-3d vision tower and full-size loading fixes (unslothai/llama.cpp#70, commit 883f2c9)
• IQ1_XS, IQ1_XXS, IQ1_XXXS: three quant types below IQ1_S (unslothai/llama.cpp#61, commit 46cbf0e)
• 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 b9b8207)
• 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)
• MTP for Qwen3.8-Flash-Next (unslothai/llama.cpp#144, commit a9e9c3c)
• llama: map each contiguous run of a context's tensors, not one span over all of them (unslothai/llama.cpp#152, commit b2b5ed9)
• mtmd: test that every projector is registered and uniquely named (unslothai/llama.cpp#176, commit 09ce1a4)

llama.cpp prebuilt b10830-mix-d5c17a0

Automated Unsloth llama.cpp CUDA + ROCm + Vulkan + macOS + CPU prebuild for upstream b10830, merged with:

• DiffusionGemma (#24423, commit 168b21a)
• Add TML Inkling architecture (#25731, commit 1066edc)
• kimi-k3 : the MoonViT-3d vision tower and full-size loading fixes (unslothai/llama.cpp#70, commit 883f2c9)
• IQ1_XS, IQ1_XXS, IQ1_XXXS: three quant types below IQ1_S (unslothai/llama.cpp#61, commit 46cbf0e)
• 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 b9b8207)
• 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)
• MTP for Qwen3.8-Flash-Next (unslothai/llama.cpp#144, commit a9e9c3c)
• llama: map each contiguous run of a context's tensors, not one span over all of them (unslothai/llama.cpp#152, commit b2b5ed9)
• mtmd: test that every projector is registered and uniquely named (unslothai/llama.cpp#176, commit 09ce1a4)