Update Unsloth llama.cpp prebuilt to v11541 #19

Merged

brian merged 1 commit from renovate/unslothai-llama.cpp-11541.x into main 2026-10-11 15:31:20 +00:00

Conversation 0 Commits 1 Files changed 1 +1 -1

brian commented 2026-10-11 02:06:08 +00:00

Owner

Copy link

This PR contains the following updates:

Package	Update	Change
unslothai/llama.cpp	major	b11160-mix-a6922cc → b11541-mix-2500203

---

Release Notes

unslothai/llama.cpp

vb11541-mix-2500203: llama.cpp prebuilt b11541-mix-2500203

Compare Source

Automated Unsloth llama.cpp CUDA + ROCm + Vulkan + macOS + CPU prebuild for upstream b11541, merged with:

• DiffusionGemma (#​24423, commit 6431891)
• Add TML Inkling architecture (#​25731, commit 68cd3f6)
• 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 02a015c)
• sampling: index penalties by token id instead of scanning every candidate (unslothai/llama.cpp#95, commit 3db8cb5)
• 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)
• 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 e2519eb)
• Load GLM-5-Next GGUFs converted with the earlier glm5next arch name (unslothai/llama.cpp#239, commit bec2164)
• qwen4exp: run MTP heads that share the target's token_embd and output (unslothai/llama.cpp#240, commit 36175a8)
• Carry what the dropped GLM-5-Next, Qwen MTP and readahead pins had beyond upstream (unslothai/llama.cpp#241, commit 20dd977)
• glm5-next : keep the indexer compute buffer bounded at long contexts (unslothai/llama.cpp#243, commit f67cec2)
• common : add --moe-cache-mib auto to size the MoE expert cache from the fit (unslothai/llama.cpp#251, commit 13f4c71)

vb11539-mix-2500203: llama.cpp prebuilt b11539-mix-2500203

Compare Source

Automated Unsloth llama.cpp CUDA + ROCm + Vulkan + macOS + CPU prebuild for upstream b11539, merged with:

• DiffusionGemma (#​24423, commit 6431891)
• Add TML Inkling architecture (#​25731, commit 68cd3f6)
• 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 02a015c)
• sampling: index penalties by token id instead of scanning every candidate (unslothai/llama.cpp#95, commit 3db8cb5)
• 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)
• 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 e2519eb)
• Load GLM-5-Next GGUFs converted with the earlier glm5next arch name (unslothai/llama.cpp#239, commit bec2164)
• qwen4exp: run MTP heads that share the target's token_embd and output (unslothai/llama.cpp#240, commit 36175a8)
• Carry what the dropped GLM-5-Next, Qwen MTP and readahead pins had beyond upstream (unslothai/llama.cpp#241, commit 20dd977)
• glm5-next : keep the indexer compute buffer bounded at long contexts (unslothai/llama.cpp#243, commit f67cec2)
• common : add --moe-cache-mib auto to size the MoE expert cache from the fit (unslothai/llama.cpp#251, commit 13f4c71)

vb11505-mix-5209c9f: llama.cpp prebuilt b11505-mix-5209c9f

Compare Source

Automated Unsloth llama.cpp CUDA + ROCm + Vulkan + macOS + CPU prebuild for upstream b11505, merged with:

• DiffusionGemma (#​24423, commit 6431891)
• Add TML Inkling architecture (#​25731, commit fddee40)
• 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 ef45f21)
• sampling: index penalties by token id instead of scanning every candidate (unslothai/llama.cpp#95, commit 3db8cb5)
• 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)
• 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 e2519eb)
• Load GLM-5-Next GGUFs converted with the earlier glm5next arch name (unslothai/llama.cpp#239, commit bec2164)
• qwen4exp: run MTP heads that share the target's token_embd and output (unslothai/llama.cpp#240, commit 36175a8)
• ggml-cuda: drop the CUDA graph cache count cap, as upstream does (unslothai/llama.cpp#253, commit 20dd977)
• glm5-next : keep the indexer compute buffer bounded at long contexts (unslothai/llama.cpp#243, commit f67cec2)
• common : add --moe-cache-mib auto to size the MoE expert cache from the fit (unslothai/llama.cpp#251, commit 13f4c71)

vb11443-mix-d65395f: llama.cpp prebuilt b11443-mix-d65395f

Compare Source

Automated Unsloth llama.cpp CUDA + ROCm + Vulkan + macOS + CPU prebuild for upstream b11443, merged with:

• DiffusionGemma (#​24423, commit 6431891)
• Add TML Inkling architecture (#​25731, commit efd2b13)
• 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 ef45f21)
• sampling: index penalties by token id instead of scanning every candidate (unslothai/llama.cpp#95, commit 3db8cb5)
• 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)
• 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 e2519eb)
• Load GLM-5-Next GGUFs converted with the earlier glm5next arch name (unslothai/llama.cpp#239, commit bec2164)
• qwen4exp: run MTP heads that share the target's token_embd and output (unslothai/llama.cpp#240, commit 36175a8)
• Carry what the dropped GLM-5-Next, Qwen MTP and readahead pins had beyond upstream (unslothai/llama.cpp#241, commit b96a713)
• glm5-next : keep the indexer compute buffer bounded at long contexts (unslothai/llama.cpp#243, commit aba4a4c)
• EmbeddingGemma-2 support (unslothai/llama.cpp#247, commit 73c2f73)

vb11438-mix-d65395f: llama.cpp prebuilt b11438-mix-d65395f

Compare Source

Automated Unsloth llama.cpp CUDA + ROCm + Vulkan + macOS + CPU prebuild for upstream b11438, merged with:

• DiffusionGemma (#​24423, commit 6431891)
• Add TML Inkling architecture (#​25731, commit efd2b13)
• 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 ef45f21)
• sampling: index penalties by token id instead of scanning every candidate (unslothai/llama.cpp#95, commit 3db8cb5)
• 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)
• 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 e2519eb)
• Load GLM-5-Next GGUFs converted with the earlier glm5next arch name (unslothai/llama.cpp#239, commit bec2164)
• qwen4exp: run MTP heads that share the target's token_embd and output (unslothai/llama.cpp#240, commit 36175a8)
• Carry what the dropped GLM-5-Next, Qwen MTP and readahead pins had beyond upstream (unslothai/llama.cpp#241, commit b96a713)
• glm5-next : keep the indexer compute buffer bounded at long contexts (unslothai/llama.cpp#243, commit aba4a4c)
• EmbeddingGemma-2 support (unslothai/llama.cpp#247, commit 73c2f73)

vb11408-mix-1e24fc5: llama.cpp prebuilt b11408-mix-1e24fc5

Compare Source

Automated Unsloth llama.cpp CUDA + ROCm + Vulkan + macOS + CPU prebuild for upstream b11408, merged with:

• DiffusionGemma (#​24423, commit 6431891)
• Add TML Inkling architecture (#​25731, commit efd2b13)
• 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 ef45f21)
• sampling: index penalties by token id instead of scanning every candidate (unslothai/llama.cpp#95, commit 3db8cb5)
• 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)
• 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 e2519eb)
• Load GLM-5-Next GGUFs converted with the earlier glm5next arch name (unslothai/llama.cpp#239, commit bec2164)
• qwen4exp: run MTP heads that share the target's token_embd and output (unslothai/llama.cpp#240, commit 36175a8)
• Carry what the dropped GLM-5-Next, Qwen MTP and readahead pins had beyond upstream (unslothai/llama.cpp#241, commit b96a713)

---

Configuration

📅 Schedule: (UTC)

• Branch creation
  • At any time (no schedule defined)
• Automerge
  • At any time (no schedule defined)

🚦 Automerge: Disabled by config. Please merge this manually once you are satisfied.

♻ Rebasing: Whenever PR becomes conflicted, or you tick the rebase/retry checkbox.

🔕 Ignore: Close this PR and you won't be reminded about this update again.

---

• If you want to rebase/retry this PR, check this box

---

This PR has been generated by Mend Renovate CLI.

This PR contains the following updates: | Package | Update | Change | |---|---|---| | [unslothai/llama.cpp](https://github.com/unslothai/llama.cpp) | major | `b11160-mix-a6922cc` → `b11541-mix-2500203` | --- ### Release Notes <details> <summary>unslothai/llama.cpp</summary> ### [`vb11541-mix-2500203`](https://github.com/unslothai/llama.cpp/releases/tag/b11541-mix-2500203): llama.cpp prebuilt b11541-mix-2500203 [Compare Source](https://github.com/unslothai/llama.cpp/compare/b11539-mix-2500203...b11541-mix-2500203) Automated Unsloth llama.cpp CUDA + ROCm + Vulkan + macOS + CPU prebuild for upstream [b11541](https://github.com/ggml-org/llama.cpp/releases/tag/b11541), merged with: - DiffusionGemma ([#&#8203;24423](https://github.com/ggml-org/llama.cpp/pull/24423), commit [6431891](https://github.com/ggml-org/llama.cpp/pull/24423/commits/643189144c98fba431e8d2da169210b7b6d5d930)) - Add TML Inkling architecture ([#&#8203;25731](https://github.com/ggml-org/llama.cpp/pull/25731), commit [68cd3f6](https://github.com/ggml-org/llama.cpp/pull/25731/commits/68cd3f6cb06b1e425ea7221af03a93099aa6ab84)) - kimi-k3 : the MoonViT-3d vision tower and full-size loading fixes ([unslothai/llama.cpp#70](https://github.com/unslothai/llama.cpp/pull/70), commit [883f2c9](https://github.com/unslothai/llama.cpp/pull/70/commits/883f2c9ba78f3847148454adf025da29385fff3e)) - IQ1_XS, IQ1_XXS, IQ1_XXXS: three quant types below IQ1_S ([unslothai/llama.cpp#61](https://github.com/unslothai/llama.cpp/pull/61), commit [02a015c](https://github.com/unslothai/llama.cpp/pull/61/commits/02a015c22ad0f9a6a2b1d4361900b45aa212f936)) - sampling: index penalties by token id instead of scanning every candidate ([unslothai/llama.cpp#95](https://github.com/unslothai/llama.cpp/pull/95), commit [3db8cb5](https://github.com/unslothai/llama.cpp/pull/95/commits/3db8cb5b2e9bf291057b9f19960e8601a162da81)) - ggml-cuda: avoid direct ROCm_Host compute on HIP integrated GPUs (port of [ggml-org#25863](https://github.com/ggml-org/llama.cpp/issues/25863)) ([unslothai/llama.cpp#158](https://github.com/unslothai/llama.cpp/pull/158), commit [abfc45b](https://github.com/unslothai/llama.cpp/pull/158/commits/abfc45b9cb21eae4848cb82196e659f42c9a8341)) - ggml-cuda: use cudaMemcpyDefault in the ggml_cuda_cpy 2D fast path ([unslothai/llama.cpp#157](https://github.com/unslothai/llama.cpp/pull/157), commit [6c6da89](https://github.com/unslothai/llama.cpp/pull/157/commits/6c6da89266ba7839d825c9997782af4f4d26b81b)) - ggml-cuda: make GGML_CUDA_ENABLE_UNIFIED_MEMORY=0 actually disable it, and say so on HIP ([unslothai/llama.cpp#149](https://github.com/unslothai/llama.cpp/pull/149), commit [b65a2dc](https://github.com/unslothai/llama.cpp/pull/149/commits/b65a2dce12c14a489e19a059cb6ee59112f1b733)) - llama: map each contiguous run of a context's tensors, not one span over all of them ([unslothai/llama.cpp#152](https://github.com/unslothai/llama.cpp/pull/152), commit [b2b5ed9](https://github.com/unslothai/llama.cpp/pull/152/commits/b2b5ed9ff86427a530b762a45d3fdbd453bcd4e8)) - mtmd: test that every projector is registered and uniquely named ([unslothai/llama.cpp#176](https://github.com/unslothai/llama.cpp/pull/176), commit [e2519eb](https://github.com/unslothai/llama.cpp/pull/176/commits/e2519ebc09468b477847a7529834a8da998ed2da)) - Load GLM-5-Next GGUFs converted with the earlier glm5next arch name ([unslothai/llama.cpp#239](https://github.com/unslothai/llama.cpp/pull/239), commit [bec2164](https://github.com/unslothai/llama.cpp/pull/239/commits/bec216444552a20bc39bf2156fa5396154970608)) - qwen4exp: run MTP heads that share the target's token_embd and output ([unslothai/llama.cpp#240](https://github.com/unslothai/llama.cpp/pull/240), commit [36175a8](https://github.com/unslothai/llama.cpp/pull/240/commits/36175a8325e24303704cfbab1b7d8a0bd30500a5)) - Carry what the dropped GLM-5-Next, Qwen MTP and readahead pins had beyond upstream ([unslothai/llama.cpp#241](https://github.com/unslothai/llama.cpp/pull/241), commit [20dd977](https://github.com/unslothai/llama.cpp/pull/241/commits/20dd977eae5e1a3b0633bba85f03d8b09692379f)) - glm5-next : keep the indexer compute buffer bounded at long contexts ([unslothai/llama.cpp#243](https://github.com/unslothai/llama.cpp/pull/243), commit [f67cec2](https://github.com/unslothai/llama.cpp/pull/243/commits/f67cec2bd6806d5294e2abc5d62a16405dd3eb74)) - common : add --moe-cache-mib auto to size the MoE expert cache from the fit ([unslothai/llama.cpp#251](https://github.com/unslothai/llama.cpp/pull/251), commit [13f4c71](https://github.com/unslothai/llama.cpp/pull/251/commits/13f4c714dff3bf7aa1a83be7490102b7706244ee)) ### [`vb11539-mix-2500203`](https://github.com/unslothai/llama.cpp/releases/tag/b11539-mix-2500203): llama.cpp prebuilt b11539-mix-2500203 [Compare Source](https://github.com/unslothai/llama.cpp/compare/b11505-mix-5209c9f...b11539-mix-2500203) Automated Unsloth llama.cpp CUDA + ROCm + Vulkan + macOS + CPU prebuild for upstream [b11539](https://github.com/ggml-org/llama.cpp/releases/tag/b11539), merged with: - DiffusionGemma ([#&#8203;24423](https://github.com/ggml-org/llama.cpp/pull/24423), commit [6431891](https://github.com/ggml-org/llama.cpp/pull/24423/commits/643189144c98fba431e8d2da169210b7b6d5d930)) - Add TML Inkling architecture ([#&#8203;25731](https://github.com/ggml-org/llama.cpp/pull/25731), commit [68cd3f6](https://github.com/ggml-org/llama.cpp/pull/25731/commits/68cd3f6cb06b1e425ea7221af03a93099aa6ab84)) - kimi-k3 : the MoonViT-3d vision tower and full-size loading fixes ([unslothai/llama.cpp#70](https://github.com/unslothai/llama.cpp/pull/70), commit [883f2c9](https://github.com/unslothai/llama.cpp/pull/70/commits/883f2c9ba78f3847148454adf025da29385fff3e)) - IQ1_XS, IQ1_XXS, IQ1_XXXS: three quant types below IQ1_S ([unslothai/llama.cpp#61](https://github.com/unslothai/llama.cpp/pull/61), commit [02a015c](https://github.com/unslothai/llama.cpp/pull/61/commits/02a015c22ad0f9a6a2b1d4361900b45aa212f936)) - sampling: index penalties by token id instead of scanning every candidate ([unslothai/llama.cpp#95](https://github.com/unslothai/llama.cpp/pull/95), commit [3db8cb5](https://github.com/unslothai/llama.cpp/pull/95/commits/3db8cb5b2e9bf291057b9f19960e8601a162da81)) - ggml-cuda: avoid direct ROCm_Host compute on HIP integrated GPUs (port of [ggml-org#25863](https://github.com/ggml-org/llama.cpp/issues/25863)) ([unslothai/llama.cpp#158](https://github.com/unslothai/llama.cpp/pull/158), commit [abfc45b](https://github.com/unslothai/llama.cpp/pull/158/commits/abfc45b9cb21eae4848cb82196e659f42c9a8341)) - ggml-cuda: use cudaMemcpyDefault in the ggml_cuda_cpy 2D fast path ([unslothai/llama.cpp#157](https://github.com/unslothai/llama.cpp/pull/157), commit [6c6da89](https://github.com/unslothai/llama.cpp/pull/157/commits/6c6da89266ba7839d825c9997782af4f4d26b81b)) - ggml-cuda: make GGML_CUDA_ENABLE_UNIFIED_MEMORY=0 actually disable it, and say so on HIP ([unslothai/llama.cpp#149](https://github.com/unslothai/llama.cpp/pull/149), commit [b65a2dc](https://github.com/unslothai/llama.cpp/pull/149/commits/b65a2dce12c14a489e19a059cb6ee59112f1b733)) - llama: map each contiguous run of a context's tensors, not one span over all of them ([unslothai/llama.cpp#152](https://github.com/unslothai/llama.cpp/pull/152), commit [b2b5ed9](https://github.com/unslothai/llama.cpp/pull/152/commits/b2b5ed9ff86427a530b762a45d3fdbd453bcd4e8)) - mtmd: test that every projector is registered and uniquely named ([unslothai/llama.cpp#176](https://github.com/unslothai/llama.cpp/pull/176), commit [e2519eb](https://github.com/unslothai/llama.cpp/pull/176/commits/e2519ebc09468b477847a7529834a8da998ed2da)) - Load GLM-5-Next GGUFs converted with the earlier glm5next arch name ([unslothai/llama.cpp#239](https://github.com/unslothai/llama.cpp/pull/239), commit [bec2164](https://github.com/unslothai/llama.cpp/pull/239/commits/bec216444552a20bc39bf2156fa5396154970608)) - qwen4exp: run MTP heads that share the target's token_embd and output ([unslothai/llama.cpp#240](https://github.com/unslothai/llama.cpp/pull/240), commit [36175a8](https://github.com/unslothai/llama.cpp/pull/240/commits/36175a8325e24303704cfbab1b7d8a0bd30500a5)) - Carry what the dropped GLM-5-Next, Qwen MTP and readahead pins had beyond upstream ([unslothai/llama.cpp#241](https://github.com/unslothai/llama.cpp/pull/241), commit [20dd977](https://github.com/unslothai/llama.cpp/pull/241/commits/20dd977eae5e1a3b0633bba85f03d8b09692379f)) - glm5-next : keep the indexer compute buffer bounded at long contexts ([unslothai/llama.cpp#243](https://github.com/unslothai/llama.cpp/pull/243), commit [f67cec2](https://github.com/unslothai/llama.cpp/pull/243/commits/f67cec2bd6806d5294e2abc5d62a16405dd3eb74)) - common : add --moe-cache-mib auto to size the MoE expert cache from the fit ([unslothai/llama.cpp#251](https://github.com/unslothai/llama.cpp/pull/251), commit [13f4c71](https://github.com/unslothai/llama.cpp/pull/251/commits/13f4c714dff3bf7aa1a83be7490102b7706244ee)) ### [`vb11505-mix-5209c9f`](https://github.com/unslothai/llama.cpp/releases/tag/b11505-mix-5209c9f): llama.cpp prebuilt b11505-mix-5209c9f [Compare Source](https://github.com/unslothai/llama.cpp/compare/b11443-mix-d65395f...b11505-mix-5209c9f) Automated Unsloth llama.cpp CUDA + ROCm + Vulkan + macOS + CPU prebuild for upstream [b11505](https://github.com/ggml-org/llama.cpp/releases/tag/b11505), merged with: - DiffusionGemma ([#&#8203;24423](https://github.com/ggml-org/llama.cpp/pull/24423), commit [6431891](https://github.com/ggml-org/llama.cpp/pull/24423/commits/643189144c98fba431e8d2da169210b7b6d5d930)) - Add TML Inkling architecture ([#&#8203;25731](https://github.com/ggml-org/llama.cpp/pull/25731), commit [fddee40](https://github.com/ggml-org/llama.cpp/pull/25731/commits/fddee40253015ff5f7e578fa941bd074d591cf33)) - kimi-k3 : the MoonViT-3d vision tower and full-size loading fixes ([unslothai/llama.cpp#70](https://github.com/unslothai/llama.cpp/pull/70), commit [883f2c9](https://github.com/unslothai/llama.cpp/pull/70/commits/883f2c9ba78f3847148454adf025da29385fff3e)) - IQ1_XS, IQ1_XXS, IQ1_XXXS: three quant types below IQ1_S ([unslothai/llama.cpp#61](https://github.com/unslothai/llama.cpp/pull/61), commit [ef45f21](https://github.com/unslothai/llama.cpp/pull/61/commits/ef45f212c90d5769e0dd8d482f99fd66760312e7)) - sampling: index penalties by token id instead of scanning every candidate ([unslothai/llama.cpp#95](https://github.com/unslothai/llama.cpp/pull/95), commit [3db8cb5](https://github.com/unslothai/llama.cpp/pull/95/commits/3db8cb5b2e9bf291057b9f19960e8601a162da81)) - ggml-cuda: avoid direct ROCm_Host compute on HIP integrated GPUs (port of [ggml-org#25863](https://github.com/ggml-org/llama.cpp/issues/25863)) ([unslothai/llama.cpp#158](https://github.com/unslothai/llama.cpp/pull/158), commit [abfc45b](https://github.com/unslothai/llama.cpp/pull/158/commits/abfc45b9cb21eae4848cb82196e659f42c9a8341)) - ggml-cuda: use cudaMemcpyDefault in the ggml_cuda_cpy 2D fast path ([unslothai/llama.cpp#157](https://github.com/unslothai/llama.cpp/pull/157), commit [6c6da89](https://github.com/unslothai/llama.cpp/pull/157/commits/6c6da89266ba7839d825c9997782af4f4d26b81b)) - ggml-cuda: make GGML_CUDA_ENABLE_UNIFIED_MEMORY=0 actually disable it, and say so on HIP ([unslothai/llama.cpp#149](https://github.com/unslothai/llama.cpp/pull/149), commit [b65a2dc](https://github.com/unslothai/llama.cpp/pull/149/commits/b65a2dce12c14a489e19a059cb6ee59112f1b733)) - llama: map each contiguous run of a context's tensors, not one span over all of them ([unslothai/llama.cpp#152](https://github.com/unslothai/llama.cpp/pull/152), commit [b2b5ed9](https://github.com/unslothai/llama.cpp/pull/152/commits/b2b5ed9ff86427a530b762a45d3fdbd453bcd4e8)) - mtmd: test that every projector is registered and uniquely named ([unslothai/llama.cpp#176](https://github.com/unslothai/llama.cpp/pull/176), commit [e2519eb](https://github.com/unslothai/llama.cpp/pull/176/commits/e2519ebc09468b477847a7529834a8da998ed2da)) - Load GLM-5-Next GGUFs converted with the earlier glm5next arch name ([unslothai/llama.cpp#239](https://github.com/unslothai/llama.cpp/pull/239), commit [bec2164](https://github.com/unslothai/llama.cpp/pull/239/commits/bec216444552a20bc39bf2156fa5396154970608)) - qwen4exp: run MTP heads that share the target's token_embd and output ([unslothai/llama.cpp#240](https://github.com/unslothai/llama.cpp/pull/240), commit [36175a8](https://github.com/unslothai/llama.cpp/pull/240/commits/36175a8325e24303704cfbab1b7d8a0bd30500a5)) - ggml-cuda: drop the CUDA graph cache count cap, as upstream does ([unslothai/llama.cpp#253](https://github.com/unslothai/llama.cpp/pull/253), commit [20dd977](https://github.com/unslothai/llama.cpp/pull/253/commits/20dd977eae5e1a3b0633bba85f03d8b09692379f)) - glm5-next : keep the indexer compute buffer bounded at long contexts ([unslothai/llama.cpp#243](https://github.com/unslothai/llama.cpp/pull/243), commit [f67cec2](https://github.com/unslothai/llama.cpp/pull/243/commits/f67cec2bd6806d5294e2abc5d62a16405dd3eb74)) - common : add --moe-cache-mib auto to size the MoE expert cache from the fit ([unslothai/llama.cpp#251](https://github.com/unslothai/llama.cpp/pull/251), commit [13f4c71](https://github.com/unslothai/llama.cpp/pull/251/commits/13f4c714dff3bf7aa1a83be7490102b7706244ee)) ### [`vb11443-mix-d65395f`](https://github.com/unslothai/llama.cpp/releases/tag/b11443-mix-d65395f): llama.cpp prebuilt b11443-mix-d65395f [Compare Source](https://github.com/unslothai/llama.cpp/compare/b11438-mix-d65395f...b11443-mix-d65395f) Automated Unsloth llama.cpp CUDA + ROCm + Vulkan + macOS + CPU prebuild for upstream [b11443](https://github.com/ggml-org/llama.cpp/releases/tag/b11443), merged with: - DiffusionGemma ([#&#8203;24423](https://github.com/ggml-org/llama.cpp/pull/24423), commit [6431891](https://github.com/ggml-org/llama.cpp/pull/24423/commits/643189144c98fba431e8d2da169210b7b6d5d930)) - Add TML Inkling architecture ([#&#8203;25731](https://github.com/ggml-org/llama.cpp/pull/25731), commit [efd2b13](https://github.com/ggml-org/llama.cpp/pull/25731/commits/efd2b13f80c68df745f0134a2a0f56a434a2d33d)) - kimi-k3 : the MoonViT-3d vision tower and full-size loading fixes ([unslothai/llama.cpp#70](https://github.com/unslothai/llama.cpp/pull/70), commit [883f2c9](https://github.com/unslothai/llama.cpp/pull/70/commits/883f2c9ba78f3847148454adf025da29385fff3e)) - IQ1_XS, IQ1_XXS, IQ1_XXXS: three quant types below IQ1_S ([unslothai/llama.cpp#61](https://github.com/unslothai/llama.cpp/pull/61), commit [ef45f21](https://github.com/unslothai/llama.cpp/pull/61/commits/ef45f212c90d5769e0dd8d482f99fd66760312e7)) - sampling: index penalties by token id instead of scanning every candidate ([unslothai/llama.cpp#95](https://github.com/unslothai/llama.cpp/pull/95), commit [3db8cb5](https://github.com/unslothai/llama.cpp/pull/95/commits/3db8cb5b2e9bf291057b9f19960e8601a162da81)) - ggml-cuda: avoid direct ROCm_Host compute on HIP integrated GPUs (port of [ggml-org#25863](https://github.com/ggml-org/llama.cpp/issues/25863)) ([unslothai/llama.cpp#158](https://github.com/unslothai/llama.cpp/pull/158), commit [abfc45b](https://github.com/unslothai/llama.cpp/pull/158/commits/abfc45b9cb21eae4848cb82196e659f42c9a8341)) - ggml-cuda: use cudaMemcpyDefault in the ggml_cuda_cpy 2D fast path ([unslothai/llama.cpp#157](https://github.com/unslothai/llama.cpp/pull/157), commit [6c6da89](https://github.com/unslothai/llama.cpp/pull/157/commits/6c6da89266ba7839d825c9997782af4f4d26b81b)) - ggml-cuda: make GGML_CUDA_ENABLE_UNIFIED_MEMORY=0 actually disable it, and say so on HIP ([unslothai/llama.cpp#149](https://github.com/unslothai/llama.cpp/pull/149), commit [b65a2dc](https://github.com/unslothai/llama.cpp/pull/149/commits/b65a2dce12c14a489e19a059cb6ee59112f1b733)) - llama: map each contiguous run of a context's tensors, not one span over all of them ([unslothai/llama.cpp#152](https://github.com/unslothai/llama.cpp/pull/152), commit [b2b5ed9](https://github.com/unslothai/llama.cpp/pull/152/commits/b2b5ed9ff86427a530b762a45d3fdbd453bcd4e8)) - mtmd: test that every projector is registered and uniquely named ([unslothai/llama.cpp#176](https://github.com/unslothai/llama.cpp/pull/176), commit [e2519eb](https://github.com/unslothai/llama.cpp/pull/176/commits/e2519ebc09468b477847a7529834a8da998ed2da)) - Load GLM-5-Next GGUFs converted with the earlier glm5next arch name ([unslothai/llama.cpp#239](https://github.com/unslothai/llama.cpp/pull/239), commit [bec2164](https://github.com/unslothai/llama.cpp/pull/239/commits/bec216444552a20bc39bf2156fa5396154970608)) - qwen4exp: run MTP heads that share the target's token_embd and output ([unslothai/llama.cpp#240](https://github.com/unslothai/llama.cpp/pull/240), commit [36175a8](https://github.com/unslothai/llama.cpp/pull/240/commits/36175a8325e24303704cfbab1b7d8a0bd30500a5)) - Carry what the dropped GLM-5-Next, Qwen MTP and readahead pins had beyond upstream ([unslothai/llama.cpp#241](https://github.com/unslothai/llama.cpp/pull/241), commit [b96a713](https://github.com/unslothai/llama.cpp/pull/241/commits/b96a713a48313a32c92fd6f77b367ea9d803417b)) - glm5-next : keep the indexer compute buffer bounded at long contexts ([unslothai/llama.cpp#243](https://github.com/unslothai/llama.cpp/pull/243), commit [aba4a4c](https://github.com/unslothai/llama.cpp/pull/243/commits/aba4a4c1db96d56846e86d1169dd35241bafd55d)) - EmbeddingGemma-2 support ([unslothai/llama.cpp#247](https://github.com/unslothai/llama.cpp/pull/247), commit [73c2f73](https://github.com/unslothai/llama.cpp/pull/247/commits/73c2f733e8beef7dd104de747442eb9e21f2bb96)) ### [`vb11438-mix-d65395f`](https://github.com/unslothai/llama.cpp/releases/tag/b11438-mix-d65395f): llama.cpp prebuilt b11438-mix-d65395f [Compare Source](https://github.com/unslothai/llama.cpp/compare/b11408-mix-1e24fc5...b11438-mix-d65395f) Automated Unsloth llama.cpp CUDA + ROCm + Vulkan + macOS + CPU prebuild for upstream [b11438](https://github.com/ggml-org/llama.cpp/releases/tag/b11438), merged with: - DiffusionGemma ([#&#8203;24423](https://github.com/ggml-org/llama.cpp/pull/24423), commit [6431891](https://github.com/ggml-org/llama.cpp/pull/24423/commits/643189144c98fba431e8d2da169210b7b6d5d930)) - Add TML Inkling architecture ([#&#8203;25731](https://github.com/ggml-org/llama.cpp/pull/25731), commit [efd2b13](https://github.com/ggml-org/llama.cpp/pull/25731/commits/efd2b13f80c68df745f0134a2a0f56a434a2d33d)) - kimi-k3 : the MoonViT-3d vision tower and full-size loading fixes ([unslothai/llama.cpp#70](https://github.com/unslothai/llama.cpp/pull/70), commit [883f2c9](https://github.com/unslothai/llama.cpp/pull/70/commits/883f2c9ba78f3847148454adf025da29385fff3e)) - IQ1_XS, IQ1_XXS, IQ1_XXXS: three quant types below IQ1_S ([unslothai/llama.cpp#61](https://github.com/unslothai/llama.cpp/pull/61), commit [ef45f21](https://github.com/unslothai/llama.cpp/pull/61/commits/ef45f212c90d5769e0dd8d482f99fd66760312e7)) - sampling: index penalties by token id instead of scanning every candidate ([unslothai/llama.cpp#95](https://github.com/unslothai/llama.cpp/pull/95), commit [3db8cb5](https://github.com/unslothai/llama.cpp/pull/95/commits/3db8cb5b2e9bf291057b9f19960e8601a162da81)) - ggml-cuda: avoid direct ROCm_Host compute on HIP integrated GPUs (port of [ggml-org#25863](https://github.com/ggml-org/llama.cpp/issues/25863)) ([unslothai/llama.cpp#158](https://github.com/unslothai/llama.cpp/pull/158), commit [abfc45b](https://github.com/unslothai/llama.cpp/pull/158/commits/abfc45b9cb21eae4848cb82196e659f42c9a8341)) - ggml-cuda: use cudaMemcpyDefault in the ggml_cuda_cpy 2D fast path ([unslothai/llama.cpp#157](https://github.com/unslothai/llama.cpp/pull/157), commit [6c6da89](https://github.com/unslothai/llama.cpp/pull/157/commits/6c6da89266ba7839d825c9997782af4f4d26b81b)) - ggml-cuda: make GGML_CUDA_ENABLE_UNIFIED_MEMORY=0 actually disable it, and say so on HIP ([unslothai/llama.cpp#149](https://github.com/unslothai/llama.cpp/pull/149), commit [b65a2dc](https://github.com/unslothai/llama.cpp/pull/149/commits/b65a2dce12c14a489e19a059cb6ee59112f1b733)) - llama: map each contiguous run of a context's tensors, not one span over all of them ([unslothai/llama.cpp#152](https://github.com/unslothai/llama.cpp/pull/152), commit [b2b5ed9](https://github.com/unslothai/llama.cpp/pull/152/commits/b2b5ed9ff86427a530b762a45d3fdbd453bcd4e8)) - mtmd: test that every projector is registered and uniquely named ([unslothai/llama.cpp#176](https://github.com/unslothai/llama.cpp/pull/176), commit [e2519eb](https://github.com/unslothai/llama.cpp/pull/176/commits/e2519ebc09468b477847a7529834a8da998ed2da)) - Load GLM-5-Next GGUFs converted with the earlier glm5next arch name ([unslothai/llama.cpp#239](https://github.com/unslothai/llama.cpp/pull/239), commit [bec2164](https://github.com/unslothai/llama.cpp/pull/239/commits/bec216444552a20bc39bf2156fa5396154970608)) - qwen4exp: run MTP heads that share the target's token_embd and output ([unslothai/llama.cpp#240](https://github.com/unslothai/llama.cpp/pull/240), commit [36175a8](https://github.com/unslothai/llama.cpp/pull/240/commits/36175a8325e24303704cfbab1b7d8a0bd30500a5)) - Carry what the dropped GLM-5-Next, Qwen MTP and readahead pins had beyond upstream ([unslothai/llama.cpp#241](https://github.com/unslothai/llama.cpp/pull/241), commit [b96a713](https://github.com/unslothai/llama.cpp/pull/241/commits/b96a713a48313a32c92fd6f77b367ea9d803417b)) - glm5-next : keep the indexer compute buffer bounded at long contexts ([unslothai/llama.cpp#243](https://github.com/unslothai/llama.cpp/pull/243), commit [aba4a4c](https://github.com/unslothai/llama.cpp/pull/243/commits/aba4a4c1db96d56846e86d1169dd35241bafd55d)) - EmbeddingGemma-2 support ([unslothai/llama.cpp#247](https://github.com/unslothai/llama.cpp/pull/247), commit [73c2f73](https://github.com/unslothai/llama.cpp/pull/247/commits/73c2f733e8beef7dd104de747442eb9e21f2bb96)) ### [`vb11408-mix-1e24fc5`](https://github.com/unslothai/llama.cpp/releases/tag/b11408-mix-1e24fc5): llama.cpp prebuilt b11408-mix-1e24fc5 [Compare Source](https://github.com/unslothai/llama.cpp/compare/b11160-mix-a6922cc...b11408-mix-1e24fc5) Automated Unsloth llama.cpp CUDA + ROCm + Vulkan + macOS + CPU prebuild for upstream [b11408](https://github.com/ggml-org/llama.cpp/releases/tag/b11408), merged with: - DiffusionGemma ([#&#8203;24423](https://github.com/ggml-org/llama.cpp/pull/24423), commit [6431891](https://github.com/ggml-org/llama.cpp/pull/24423/commits/643189144c98fba431e8d2da169210b7b6d5d930)) - Add TML Inkling architecture ([#&#8203;25731](https://github.com/ggml-org/llama.cpp/pull/25731), commit [efd2b13](https://github.com/ggml-org/llama.cpp/pull/25731/commits/efd2b13f80c68df745f0134a2a0f56a434a2d33d)) - kimi-k3 : the MoonViT-3d vision tower and full-size loading fixes ([unslothai/llama.cpp#70](https://github.com/unslothai/llama.cpp/pull/70), commit [883f2c9](https://github.com/unslothai/llama.cpp/pull/70/commits/883f2c9ba78f3847148454adf025da29385fff3e)) - IQ1_XS, IQ1_XXS, IQ1_XXXS: three quant types below IQ1_S ([unslothai/llama.cpp#61](https://github.com/unslothai/llama.cpp/pull/61), commit [ef45f21](https://github.com/unslothai/llama.cpp/pull/61/commits/ef45f212c90d5769e0dd8d482f99fd66760312e7)) - sampling: index penalties by token id instead of scanning every candidate ([unslothai/llama.cpp#95](https://github.com/unslothai/llama.cpp/pull/95), commit [3db8cb5](https://github.com/unslothai/llama.cpp/pull/95/commits/3db8cb5b2e9bf291057b9f19960e8601a162da81)) - ggml-cuda: avoid direct ROCm_Host compute on HIP integrated GPUs (port of [ggml-org#25863](https://github.com/ggml-org/llama.cpp/issues/25863)) ([unslothai/llama.cpp#158](https://github.com/unslothai/llama.cpp/pull/158), commit [abfc45b](https://github.com/unslothai/llama.cpp/pull/158/commits/abfc45b9cb21eae4848cb82196e659f42c9a8341)) - ggml-cuda: use cudaMemcpyDefault in the ggml_cuda_cpy 2D fast path ([unslothai/llama.cpp#157](https://github.com/unslothai/llama.cpp/pull/157), commit [6c6da89](https://github.com/unslothai/llama.cpp/pull/157/commits/6c6da89266ba7839d825c9997782af4f4d26b81b)) - ggml-cuda: make GGML_CUDA_ENABLE_UNIFIED_MEMORY=0 actually disable it, and say so on HIP ([unslothai/llama.cpp#149](https://github.com/unslothai/llama.cpp/pull/149), commit [b65a2dc](https://github.com/unslothai/llama.cpp/pull/149/commits/b65a2dce12c14a489e19a059cb6ee59112f1b733)) - llama: map each contiguous run of a context's tensors, not one span over all of them ([unslothai/llama.cpp#152](https://github.com/unslothai/llama.cpp/pull/152), commit [b2b5ed9](https://github.com/unslothai/llama.cpp/pull/152/commits/b2b5ed9ff86427a530b762a45d3fdbd453bcd4e8)) - mtmd: test that every projector is registered and uniquely named ([unslothai/llama.cpp#176](https://github.com/unslothai/llama.cpp/pull/176), commit [e2519eb](https://github.com/unslothai/llama.cpp/pull/176/commits/e2519ebc09468b477847a7529834a8da998ed2da)) - Load GLM-5-Next GGUFs converted with the earlier glm5next arch name ([unslothai/llama.cpp#239](https://github.com/unslothai/llama.cpp/pull/239), commit [bec2164](https://github.com/unslothai/llama.cpp/pull/239/commits/bec216444552a20bc39bf2156fa5396154970608)) - qwen4exp: run MTP heads that share the target's token_embd and output ([unslothai/llama.cpp#240](https://github.com/unslothai/llama.cpp/pull/240), commit [36175a8](https://github.com/unslothai/llama.cpp/pull/240/commits/36175a8325e24303704cfbab1b7d8a0bd30500a5)) - Carry what the dropped GLM-5-Next, Qwen MTP and readahead pins had beyond upstream ([unslothai/llama.cpp#241](https://github.com/unslothai/llama.cpp/pull/241), commit [b96a713](https://github.com/unslothai/llama.cpp/pull/241/commits/b96a713a48313a32c92fd6f77b367ea9d803417b)) </details> --- ### Configuration 📅 **Schedule**: (UTC) - Branch creation - At any time (no schedule defined) - Automerge - At any time (no schedule defined) 🚦 **Automerge**: Disabled by config. Please merge this manually once you are satisfied. ♻ **Rebasing**: Whenever PR becomes conflicted, or you tick the rebase/retry checkbox. 🔕 **Ignore**: Close this PR and you won't be reminded about this update again. --- - [ ] <!-- rebase-check -->If you want to rebase/retry this PR, check this box --- This PR has been generated by [Mend Renovate CLI](https://github.com/renovatebot/renovate). <!--renovate-debug:eyJjcmVhdGVkSW5WZXIiOiI0NC4xNDkuMCIsInVwZGF0ZWRJblZlciI6IjQ0LjE0OS4wIiwidGFyZ2V0QnJhbmNoIjoibWFpbiIsImxhYmVscyI6WyJyZW5vdmF0ZSJdfQ==-->

brian added 1 commit 2026-10-11 02:06:08 +00:00

Update Unsloth llama.cpp prebuilt to v11541 1d4fe1066c

brian merged commit 198f50320b into main 2026-10-11 15:31:20 +00:00

brian referenced this pull request from a commit 2026-10-11 15:31:20 +00:00

Merge pull request 'Update Unsloth llama.cpp prebuilt to v11541' (#19) from renovate/unslothai-llama.cpp-11541.x into main

Sign in to join this conversation.

Reviewers

No reviewers

Labels

Clear labels

No items

No labels

Milestone

Clear milestone

No items

No milestone

Assignees

Clear assignees

No assignees

1 participant

Notifications

Subscribe

Due date

The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set

Reference

brian/llama-unsloth-container!19

No description provided.