llama.cpp project releases multiple pre-release updates with new model support and optimizations
ByPulseAugur Editorial·[74 sources]·
The llama.cpp project has released several pre-release updates, including b11559 which adds support for Gemma-4 GPU decoding and optimizes decoding for various configurations. Release b11558 addresses image limits for Q4_K dense bin kernels and RMS norm calculations in OpenCL. Other updates include improvements to the RPC backend logging, support for new models like Prism Bonsai 2 27B and MiniCPM-V 4.7, and fixes for CPU and Vulkan functionalities. These releases also continue to expand compatibility across various operating systems and hardware architectures.
AI
IMPACT
These updates enhance the performance and compatibility of the llama.cpp framework, potentially improving the efficiency of running various LLMs on diverse hardware.
RANK_REASON
The cluster consists of multiple pre-release updates to the llama.cpp software, detailing bug fixes and new model support, rather than a core AI model release or significant industry event.
Composite score across the factors below. Higher = stronger signal that this story matters right now.
Newsworthiness bucket
Tool
The cluster consists of multiple pre-release updates to the llama.cpp software, detailing bug fixes and new model support, rather than a core AI model release or significant industry event.
Source corroboration
74 independent sources
Strong cross-source corroboration — multiple independent publishers covered this within the clustering window.
Topics
model release, product
Editorial topic classification. Feeds into how the story surfaces on /topic/<slug> hub pages and into the per-entity coverage mix.
AI-industry relevance
High
Clearly on-topic for AI-industry coverage.
Story freshness
6 days old
Aged out of breaking-news scoring windows; ranking reflects the durable signal from the full source set.
Coverage growth since scoring
+29 source(s) since last score
New sources have picked up this story since our last re-score. Score will update on the next scoring pass.
<details open=""> <p>graph: skip the LM head for pooled embeddings (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/30445">#30445</a>)</p> <ul> <li>graph: skip the LM head for pooled embeddings</li> </ul> <p>A context that extracts pooled embe…
<details open=""> <p>tests : consolidate context state tests into test-llama-context (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/30023">#30023</a>)</p> <ul> <li>tests : consolidate the context state tests into test-llama-context</li> </ul…
<p>graph : avoid reallocations in the "no cache" path (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/30439">#30439</a>)</p>
<details open=""> <p>graph: pad the KV dimension of the no-cache attention to 256 (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/30416">#30416</a>)</p> <p>The no-cache path builds a [n_tokens, n_tokens] mask and passes K and V<br /> unpadded…
<details open=""> <p>tests : link each test against only the libraries it uses (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/30299">#30299</a>)</p> <p>Give llama_build() and llama_build_and_test() a LINK keyword (default:<br /> llama-common…
<details open=""> <p>vulkan: fix buffer logging on 32-bit platforms after source split (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/30326">#30326</a>)</p> <p>Move the vk::Buffer stream operator to the shared header so it is visible to ggml…
<details open=""> <p>vulkan: handle mul_mat_id duplicates in the prepass rather than looping (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/29998">#29998</a>)</p> <ul> <li> <p>vulkan: compute every row of an expert in mul_mm_id when ids repe…
<details open=""> <p>server: leave a busy slot untouched when a request pins it (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/30295">#30295</a>)</p> <p>A request asking for a busy id_slot still ran the prompt cache update<br /> on that slot…
<p>cuda: always use MMVQ for MUL_MAT_ID on sm_60 (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/27828">#27828</a>)</p>
<details open=""> <p>CUDA: Remove redundant CUDA copies after SSM_SCAN (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/29807">#29807</a>)</p> <ul> <li> <p>CUDA: fuse copy of updated state snapshots into recurrent cache with ssm_scan</p> </li>…
<details open=""> <p>ggml: fix OOB write in ggml_acc with negative offset (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/30135">#30135</a>)</p> <p>ggml_acc_impl narrowed a size_t offset to int32_t without checking that it<br /> fits, so a la…
<details open=""> <p>webgpu: use 2D workgroup dispatch for all the ops which use 1D dispatch (e.g., rms_norm) (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/30219">#30219</a>)</p> <ul> <li> <p>remove 1D workgroups dispatching</p> </li> <li> …
<details open=""> <p>musa: drop mp_21 from the default architectures (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/30203">#30203</a>)</p> <ul> <li>musa: drop mp_21 from the default architectures</li> </ul> <p>Signed-off-by: Xiaodong Ye <a h…
<details open=""> <p>ci: fix Models Backend webgpu by supporting GGML_OP_DUP (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/30216">#30216</a>)</p> <p>The mixed batch path of PR-29622 writes the token rows with set_rows<br /> into a dup of th…
<details open=""> <p>server: use port 9931 by default (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/30159">#30159</a>)</p> <ul> <li> <p>server: use port 9931 by default</p> </li> <li> <p>revert unrelated changes</p> </li> </ul> </details> <…
<details open=""> <p>CUDA : looped PAD kernel for more than 65535 rows or slices (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/30147">#30147</a>)</p> </details> <p><strong>Website:</strong></p> <ul> <li><a href="https://llama.app" rel="nofo…
<details open=""> <p>CUDA: fix CCCL version guard breaking on major version rollover (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/29453">#29453</a>)</p> <ul> <li>CUDA: fix CCCL version guard breaking on major version rollover</li> </ul> <p…
<details open=""> <p>vulkan : fix TOP_K for +inf/NaN inputs and k = 1 on negative values (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/30107">#30107</a>)</p> <p>The bucket search in topk_nary_search.comp started from the range<br /> [0, 0xF…
<p>vulkan: extend sparse FA support to coopmat2 (<a class="issue-link js-issue-link" href="https://github.com/ggml-org/llama.cpp/pull/30003">#30003</a>)</p>
