Search: “context”

The search results do not contain specific details about the b9494 release. Based on the context from the llama.cpp GitHub releases page, b9494 is an intermediate build release of llama.cpp, a C/C++ p

The search results do not contain specific information about release b9496. Based on the available context, b9496 is a build/release version of llama.cpp, a C/C++ implementation for efficient LLM infe

The search results show recent llama.cpp releases but do not contain specific information about release b9467. Based on the context of llama.cpp releases and the GitHub repository structure, b9467 is

The search results don't contain specific information about release b9470. Based on available context, b9470 is a build release from llama.cpp, a C/C++ implementation that enables LLM inference with m

The search results do not contain specific details about release b9354. Based on the context from the llama.cpp project, b9354 is a build release of llama.cpp, a C/C++ implementation for running large

The search didn't provide specific details about release b9365. Based on the GitHub repository pattern and context, b9365 is a build release of llama.cpp, which is a C/C++ implementation for enabling

The search results show information about llama.cpp releases but don't contain specific details about build b9297. Based on the context and numbering (b9297 is a recent build after b9294), b9297 is li

Release b9260 of llama.cpp includes OpenCL backend refactoring that improves initialization, GPU identification, and performance by caching global memory size in device context. llama.cpp enables LLM

I was unable to find specific details about the b9213 release in the search results. Based on the context from the llama.cpp project, b9213 is an intermediate build release of llama.cpp, the C/C++ imp

The search did not return specific information about the b9140 release. Based on the context, b9140 is a specific build number from the llama.cpp project releases. llama.cpp is an LLM inference implem

The search results don't contain specific information about release b9087. Based on the context, b9087 is a release from llama.cpp, a C/C++ implementation for LLM inference . This release likely conta

I was unable to find specific details about the b9058 release. Based on the context, b9058 is a build version from the llama.cpp project, which is an open source software library that performs inferen

The search did not return specific details about release b8982. Based on the llama.cpp project context from the search results, b8982 is a build/release version from the llama.cpp repository that foll

The search results do not contain specific information about release b8948 itself. Based on the context from llama.cpp releases, b8948 is one of the continuously updated build releases from the llama.

The search results don't contain specific details about the b8950 release. Based on the naming convention and context, b8950 is an intermediate build release of llama.cpp, a C/C++ implementation of LL

The search results don't contain specific details about release b8918. Based on the context from the source and other release information found, b8918 is an intermediate build version of llama.cpp, th

The search results did not return the specific changelog details for build b8756. Based on what is available and the general context of llama.cpp's rolling release model, here is a factual summary ...

llama : report compute graph reallocations during inference (#30329) Add an optional ggml_backend_sched_realloc_callback that user code registers with ggml_backend_sched_set_realloc_callback to be not

add GLM-5.3-Flash (GLM5-Next) support (#27773) Rebase GLM-Next support onto master, and migrate to llama-memory-hybrid-idx Add initial MTP support Merge branch optimizations. Reduce allocated compute

vulkan: int8 coopmat1 matmul implementation for AMD RDNA3 and RDNA4 (#27952) vulkan: add int8 coopmat quantized matmul shader apply scales inline use scalar sums probe and directly access coopmat valu

The decode loop releases MLX's pool of freed buffers every 256 generated tokens, which is also how often the KV cache grows and drops its previous, smaller buffers. The check fires only when the token

OpenVINO: optimize stateful decode and GPU MoE inference (#28638) exclude GPU/NPU failing POOL_2D case Fix pool case ggml-openvino: fix stateful decode for Gemma-4 per-layer-type head sizes ggml-openv

sycl: rfc: Use radix select for top_k (#28670) sycl: GPU-resident TOP_K for large k, parallelised over the device The SYCL backend refused GGML_OP_TOP_K above k = 32 and let it fall back to the CPU, a

qwen4exp: sum the indexer heads by slices (#28023) qwen4exp: sum the indexer heads by slices The head reduction went through a transpose and a sum_rows over ne[1], which left sum_rows with ne0 = 4, on

common : skip device_info loop if it's not going to be printed (#26692) The device_info loop iterates over the discovered devices and gets the available and total memory counts. With the CUDA backend

model: add Kimi-K3 text model (#26185) model: add Kimi-K3 text model Hybrid KDA (linear) + MLA (full) attention as in Kimi-Linear-48B, plus five things that architecture does not have: cross-layer res

OpenVINO: Qwen3.5, memory optimization, and test-recurrent-state-rollback (#26952) OpenVINO backend: 1) enable gpt-oss moe on OV bk; 2) enable mxfp4 support OpenVINO backend: disable TOPK_MOE op test

Release b9273 of llama.cpp introduces support for the HybridDNATokenizer with new pre-type and dispatched tokenization logic, alongside pure helper functions for DNA k-mer processing and conversion ut

hexagon: add IM2COL_3D (#30305) hexagon: add IM2COL_3D support Assisted-by: OpenCode hexagon: add DMA paths for IM2COL_3D Assisted-by: OpenCode hexagon: avoid scalar fp16 conversions in IM2COL Assiste

server: leave a busy slot untouched when a request pins it (#30295) A request asking for a busy id_slot still ran the prompt cache update on that slot before being deferred. When the RAM cache held a

meta : handle host views (#30217) meta : handle views of tensors allocated on the host A view shares the memory of its view_src, so ggml-alloc never allocates a view in the buffer of the split it land

metal : fix MUL_MAT+ADD fusion when the residual is itself a MUL_MAT (#30100) metal : fix MUL_MAT+ADD fusion when the residual is itself a MUL_MAT ggml_metal_op_mul_mat_mma picks the residual of a fus

models : consolidate nextn row cropping into shared helpers (#30017) mimo2 : always emit h_nextn the other nextn-capable models set it unconditionally Assisted-by: pi:llama.cpp/MiMo-V2.6-Flash-MOPD mo

qwen4exp : halve the indexer score memory (#29825) qwen4exp : halve the indexer score memory The indexer scored all heads in one product and rectified a copy of it, so two [n_pool, n_idx_h, n_tokens]

llama : clamp kpool re-pool bound to existing pools (#29805) tests : simplify function signature llama : clamp kpool re-pool bound to existing pools The n_tokens/kpool + n_seqs_unq bound on n_new_g ov

ggml : add alloc_buffer_n to buffer type interface (#23671) ggml : add alloc_buffer_n to buffer type interface Add alloc_buffer_n method to ggml_backend_buffer_type_i interface, with a public API ggml

Decision models Ollama now supports decision models through /v1/systemone, based on TypeSafe’s Jev API. Decision models return choices, probabilities, and scores instead of text. Use them for tasks su

server : allow RANK pooling batch splitting for causal LLM rerankers (ie. Qwen3 and Qwen3-VL) (#28876) server : allow splitting RANK pooling for causal LLM rerankers Rerank models fall into two catego

CUDA: enable sparse-fa for dsv4 prefill (again) (#29298) CUDA: enable sparse-fa for dsv4 prefill (again) CUDA: unroll the query loop of the sparse mask scan The query loop of flash_attn_mask_to_sparse

spec : support DFlash for HunyuanOCR (#28890) model : add DFlash layer-input taps for HunyuanVL DFlash speculative decoding needs the target graph to expose the residual stream entering each layer (re

hexagon: support for multi-device model split (aka row-split) (#28589) hex-row-split: add support for multi-device row spliting Co-authored-by: Max Krasnyansky maxk@qti.qualcomm.com hex-mdev: add work

kv-cells: look up the n-gram history in the sequence position index (#28040) get_prev_tokens() rebuilt a (seq, pos) -> token hash map on every ubatch by walking all used cells, while llama_kv_cells al

kv-cache : optimize restoring non-contiguous cells (#27991) kv cache : batch state restore scatter reads per contiguous run When restoring state into non-contiguous destination cells (e.g. a prompt-ca

CUDA: use the fast mm_ids_helper path for any n_expert_used (#27978) The optimized path grouped warp lanes by token and required warp_size % n_expert_used == 0, with a single hardcoded exception paddi

sycl: bind the f16 KV cache in place for the oneDNN SDPA path (#27468) Measured at a live KV length of 34816 (32768 depth plus one 2048 ubatch), on Qwen3.8 27B Q4_K_S: per tensor 4 * 34816 * 256 * 2 B

OpenVINO: Update OV to 2026.3.1, whisper.cpp support, Qwen3.5 on NPU, and new ops (#27843) OpenVINO Backend: Fuse IM2COL + MatMul convolution into OpenVINO convolution ci:ggml-ov: Skip recurrent state