b10313
server: (router) add LRU scheduler (#26572) add lru_sched handle coalescing (req leaves waiting queue) add tests fix stream case address review comments Website: https://llama.app macOS/iOS: macOS App
Knowledge catalogue
server: (router) add LRU scheduler (#26572) add lru_sched handle coalescing (req leaves waiting queue) add tests fix stream case address review comments Website: https://llama.app macOS/iOS: macOS App
metal : fix NORM/RMS_NORM for row lengths that leave a partial simdgroup (#26708) ggml_metal_op_norm sized the threadgroup with nth = std::min(nth, args.ne00_t), which can leave nth not a multiple of
sycl: coalesce the ssm_conv window loads (#26612) test-backend-ops perf -o SSM_CONV on an Arc Pro B70, interleaved A/B against master, 6 reps, us/run: ne_a=[515,3328,1,1] ne_b=[4,3328,1,1] n_t=512 97.
tts: account for the vocoder pass in the timings line (#26733) get_output runs the waveform work the pipeline defers to it, from a single trailing window to a full pass depending on the model. Measuri
mtmd/ggml: add ggml_build_forward_order (#26649) ggml: add ggml_build_forward_order ggml_build_forward_expand marks the tensor and all its ancestors for compute, so using it as a pure ordering hint (k
vulkan: fix submission batching size, add debug tools for diagnosing causes of DeviceLost drivers errors (#26371) vulkan: add debug tooling to get more information about a DeviceLost error fix submiss
ci : onboard AMD ROCm CI with gfx1151 fixes (#26544) ci: prepare for amd rocm ci Signed-off-by: Aaron Teo aaron.teo1@ibm.com ci: fix editorconfig-checker Signed-off-by: Aaron Teo aaron.teo1@ibm.com ci
arXiv:2608.04407v1 Announce Type: cross Abstract: Memory-efficient matrix optimizers such as Sinkhorn gradient descent remove most AdamW optimizer state for dense Transformer matrices, but direct appl
server: Adding spec-decode counters to /metrics endpoint (#26389) server: add spec-decode counters to /metrics endpoint server: fixed review comments and now aligned param names exactly with vLLM. Web
mtmd: support multi-row batching for deepseek-ocr (#26154) mtmd: support multi-row batching for deepseek-ocr mtmd: weave deepseek-ocr rows in one shot instead of per row (#26615) Co-authored-by: Saba
server: harden the file_glob_search directory walk (#26626) server: don't walk Windows junctions in file_glob_search std::filesystem reports a junction as a plain directory, so the symlink guard misse
ggml: use dynamic allocation for split graph inputs (#22789) ggml: use dynamic allocation for split graph inputs Replace fixed-size GGML_SCHED_MAX_SPLIT_INPUTS arrays with dynamically allocated buffer
server: add get_info tool (#26522) server: add get_info tool fix --rpc in docs server: harden get_info probe result handling Report the OS as unknown when the probe process fails to spawn or times out
tests: add model resolution test on synthetic repo listings (#26172) tests: add model resolution test on synthetic repo listings Include download.cpp and arg.cpp inside a namespace with hf_cache monke
chat : add new template for DeepSeek V4 Flash 0731 (#26398) common/chat: update DeepSeek V4 templates Align the DeepSeek V4 templates with the official encoders while keeping parser behavior out of th
Extended SYCL oneDNN SDPA to non-FP16 KV caches (Q4_0–Q8_0 and FP32) (#25874) sycl: extend oneDNN SDPA to Q4_0-Q8_0 and F32 KV caches Extends the oneDNN SDPA path (PR #25222) to handle non-F16 KV cach
sycl: parallelize the non-contiguous concat kernel (#25852) sycl: parallelize the non-contiguous concat kernel Launch geometry only: the non-contiguous concat kernel launched a single-lane work-group
llama : move n_vocab from llama_sampler_data to penalty_sampler (#26520) This matches how it is done for logit_bias and mirostat samplers, see #25262 (comment) Website: https://llama.app macOS/iOS: ma
vulkan backend ops: implemented GATED_LINEAR_ATTN (#25601) vulkan : add GATED_LINEAR_ATTN op docs : update Vulkan ops vulkan : remove unused GLA spec constant Updated ops.md ops.md update Website: htt
speculative : refactor enabled configs common_speculative_init (#26510) This commit contains a suggestion to reduce some code duplication in common_speculative_init when adding the enabled speculative
ci: fix pre-built binaries no longer working on macOS 15 and below (#26375) ci: fix pre-built binaries no longer working on macOS 15 and below ci: add macOS deployment target to disabled KleidiAI buil
mtmd: support Qwen3-TTS (note: breaking change to llama-tts binary) (#26254) convert text model main model load ok convert encoder ok speaker encoder loading ok speaker enc graph adapt vocab for backb
ui: CWD for agent (#26518) server : extend file_glob_search for UI pickers ui : add per-conversation working directory with picker ui : add path navigation and search scope to cwd picker Treat path-li
sampler : remove 'full-context windows' from history-based samplers (#26524) Resolve -1 to 1024 instead of ctx-len for samplers Because of backend-sampling we initialize samplers before the complete l
server: decode Windows OEM output to UTF-8 in built-in tools (#26597) a child process writes in the OEM code page, which is not UTF-8 on a western Windows install, so accented output reaches the JSON
Disclosure: I’m the author and maintainer of QuarkStar. I built QuarkStar, a small native inference engine inspired by Antirez’s DwarfStar. QuarkStar currently supports: Qwen3.6-35B-A3B, using the sam
metal: implement DSv4 Lightning Indexer (#25893) metal: implement F16 Lightning Indexer Implement GGML_OP_LIGHTNING_INDEXER for 128-dimensional, 64-head inputs with F32 queries and weights plus F16 ke
llama : MTP support for DeepSeek V3.2 (#26457) llama : MTP support for DeepSeek V3.2 model : no need to include MTP layers during DeepSeek V3.2 model type discovery Co-authored-by: Stanisław Szymczyk
model: MTP support for Qwen3-Next (#25589) mtp for qwen3nex fix for python type-check Fix to compute num_mtp from directly mtp layer define opt_num_mtp_layers in _QwenMtpMixin and fix some comments Fi
server: add notice for upcoming default port change 8080 --> 9931 (#26508) server: add notice for upcoming default port change 8080 --> 9931 add link to PR correct to 9931 Website: https://llama.app m
CUDA: Fix data-races when reusing SMEM in block_reduce (#26385) CUDA: Fix data-races when reusing block_reduce block_reduce currently doesn't resync after reading from SMEM, causing potential data-rac
CUDA: Add backend sampler for penalties sampler (#25262) sampling: enhance penalty handling in common_sampler_init Set default value for penalty_last_n based on model context if not specified. Ensure
model: M3: Move MSA into a new memory implementation (#26338) Move MSA logic from llama-kv-cache into llama-kv-cache-msa cont : minor cont : ws fix Co-authored-by: Georgi Gerganov ggerganov@gmail.com
opencl: route large q6_K lm_head to the flat GEMV (#26427) add a direct size condition for large weights; the original dimension condition is insufficient -- q6_K lm_head for gemma-4 E2B has [1536, 26
chat : add qwen3 specialized parser (#26252) Add tagged thinking tool parser chat : refactor and add permute helper cont : add support for <tool_call> omission cont : update tool delimiters cont : add
opencl: bugfix increment ref_count in ggml_backend_opencl_init() (#26162) Incrementing ref_count at the beginning is important later in the free() method of the ggml_backend_opencl_context at program
common: support the DSpark sidecar resolution (#26458) The dspark- files resolve like the other speculative sidecars: the -hfd tag applies to them, a requested sidecar resolves without a full model at
metal: implement DeepSeek V4 hyper-connections (#26459) Implement GGML_OP_DSV4_HC_COMB, GGML_OP_DSV4_HC_PRE, and GGML_OP_DSV4_HC_POST with SIMDgroup register and shuffle optimized kernels. Add Metal d
metal : add SILU_BACK (#25982) feat(silu_back): implemented silu_back op for f32 fix(silu_back): removed redundant asserts in ggml-metal-ops.cpp function ggml_metal_op_silu_back. Website: https://llam
New York Times: A look at the deluge of AI computing power set to come online in the coming years; Epoch AI expects the number of AI chips in use to double every nine months — The milestones for artif
mtmd: add minicpmv46 downsample (#25993) add minicpmv46 downsample Signed-off-by: tc-mb tianchi_cai@icloud.com put downsample mode inside gguf. Signed-off-by: tc-mb tianchi_cai@icloud.com build mtmd_i
cli : persist reasoning_content in chat history (#26362) cli : persist reasoning_content in chat history llama-cli collected reasoning from the stream for display but only stored assistant content in
ggml-webgpu: improve flash_attn_vec for quantized KV at long contexts (#25956) improve fa of quantized kv cache Fix some bugs and some comments. fix v type check and some comments Fix build error caus
ggml-zendnn : group matmul direct API for mul_mat_id (#25918) ggml-zendnn : group matmul API for mul_mat_id ggml-zendnn : scale MUL_MAT_ID fallback threshold by expert count Website: https://llama.app
llama : enforce the same K and V cache types for DeepSeek V4; enable FA if V cache is quantized (#25871) llama : enforce the same K and V cache types for DeepSeek V4; enable FA if V cache is quantized
SYCL: add oneMKL GEMM flash attention for XMX-accelerated prompt proc… (#25025) SYCL: add oneMKL GEMM flash attention for XMX-accelerated prompt processing fattn-mkl: fix interleaved dst layout in nor
server: correct accepted tokens when need draft token replay (#26320) spec: correct accepted tokens when need draft token replay cont : naming Co-authored-by: Georgi Gerganov ggerganov@gmail.com Websi
llama : load MTP tensors only if they are really used (#26296) llama : load MTP tensors only if they are really used llama : skip loading MTP (if not used) in remaining models that support MTP Co-auth
vulkan: Introduce driver version check for Windows Intel GPU to mitigate crashing (#25192) Removed crash guard for Intel Crash fixed from driver 32.0.101.8860 Added driver version check for windows Ch
vulkan: add POOL_1D op (#25431) vulkan : add pool1d push constants and pipeline field Declared data structures needed for POOL1D OP, which are the vk_op_pool1d_push_constants struct and pipeline_pool1
DeepSeek V4 Flash: Preview → 2026-07-31 Benchmark Preview 0731 Δ Terminal Bench* 56.9 82.7 +25.8 Toolathlon 51.8 70.3 +18.5 NL2Repo — 54.2 new Cybergym — 76.7 new DeepSWE — 54.4 new Agent Last Exam —
metal: fix memory unwire if model is freed without any GPU operations (#26082) metal: fix memory leak if model is freed without any GPU operations metal: run dummy work only if residency sets are used
ggml-cuda: Allow transpose-free gemmv computation (#26171) When matrix's weights are shaped 1xK is leverage a transpose-free computation to use mat_mul_vec_f. Website: https://llama.app macOS/iOS: mac
Test support for alternative conv layout (#25617) add bool cwhn = true to conv_2d test cases add layout check at graph building time extend layout checks for conv2d.cu kernel in CPU back-end kernel ne
I've always found it a little strange that AI is everywhere, but the codecs we use in practice are the traditional hand-engineered systems like h.264, h.265, av1. Alexnet started the wave of neural ne
model: add NextN/MTP speculative decoding support for GLM_DSA (GLM-5.2) (#25980) model: add NextN/MTP speculative decoding support for GLM_DSA (GLM-5.2) Adds GLM-5.2 NextN/MTP as a --spec-type draft-m
server : add trace logging for slot similarity checking (#26271) Adds trace logging in server-context.cpp for slot similarity checking during prompt cache slot selection, including skip reasons and si
sycl: contiguous fast path + 32-bit index math for unary elementwise ops (#25946) sycl: contiguous fast path + 32-bit index math for unary elementwise ops sycl: use fastdiv for elementwise index math
ggml-cuda : disable MMQ on devices with less than 48 KiB shared memory (#26141) ggml_cuda_should_use_mmq() selects MMQ purely from the quantization type. The current MMQ configurations are designed an
llama: move suppress_tokens handling to common/sampling (#26276) llama: move suppress_tokens handling to common/sampling address security issues rm has_logit_bias Website: https://llama.app macOS/iOS: