ggerganov/llama.cpp

cann: upgrade flash attention to V5 and add SWIGLU_OAI/ADD_ID ops devops ggml Ascend NPU

vulkan: Use native e2m1 and e4m3 conversions for mxfp4/nvfp4 Vulkan ggml

perplexity : Printing PPL when stride is used (#25335) [no release] examples

ggml : add E4M3 (fp8) CPU quantization type testing examples ggml Apple Metal conversion

fix: load qwen35 Ollama GGUF exports model

chat : allow gpt-oss non-streaming final without channel markup (#25321) testing

llama_dsv4: write only used rows in state

ggml : use kernel type for conv_1d im2col to support f32 ggml

ggml: asynchronous scheduler memory copies ggml CUDA

SYCL: oneDNN SDPA Flash Attention for Quantized KV Caches documentation ggml SYCL

loader: map sparse mmap tensor ranges

json-schema-to-grammar: harden _visit_pattern against malformed regex patterns

kv-cache : fix SWA state save/load round-trip past n_swa

llama : stream MoE routed experts from disk

common : resolve non-positive --threads to the number of math cores documentation testing server

CANN: Refactor `#ifdef` blocks to avoid unreachable code after `return` ggml Ascend NPU

ggml, server: add ggml_backend_dev_reset() for sleep mode Vulkan server ggml SYCL Apple Metal Ascend NPU OpenCL IBM zDNN Hexagon CUDA AMD ZenDNN OpenVINO WebGPU

[SYCL] support OP OPT_STEP_ADAMW, OPT_STEP_SGD documentation ggml merge ready SYCL

[SYC:] support op get_rows_back, only support fp32/fp16 documentation ggml SYCL

CUDA: Add backend sampler for penalties sampler testing

feat: add --threads-all option to llama-bench examples

common,server : fix custom preset dedup against cached models server

[UT] enhance UT to show all real unsupported backends testing

llama : clear error when MTP draft shares KV cache across backends

vulkan: when using transfer queue for async copies, sync on event_wait to avoid race Vulkan ggml

CUDA: Support CUDA Virtual Devices ggml CUDA

server : don't list cached models when a preset is used server

[SYCL] Flash Attention with XMX engine via oneDNN graph API (SDPA) on KV f16 for Xe2 ; Qwen3.6-27b-Q8_0 prefill speed up x1.21 at p=512 and x4.26 at p=80k ggml merge ready SYCL

sycl: add fused top-k MoE documentation ggml merge ready SYCL

server: add --no-sleep flag for GPU heartbeat on headless GPUs Vulkan server ggml SYCL CUDA