The llama.cpp project has released version b11122, which includes significant updates to the SYCL backend. The release extends the MMVQ GLU fusion to support mixed quantization types and introduces new fusions for rms_norm+scale and ssm_conv+silu operations.
The llama.cpp project released version b11120, which addresses a critical bug in the Vulkan backend by hiding internal symbols. This change prevents duplicate-dlopen state destruction issues that occurred when internal symbols were exported.
The llama.cpp project has released build b11118, which includes a new hex-dma feature introducing a direct-mapped DMA cache better suited for Hexagon Vector eXtended (HVX) FA mask handling.
The llama.cpp project released build b11111, which introduces Vulkan-based Intel Xe flash attention optimization kernels for split k paths on Xe-LPG Plus, Xe2, and Xe3 architectures.
The llama.cpp project has added support for DFlash speculative decoding with the HunyuanOCR model by exposing layer input tensors in the target graph. This change allows the draft model to read residual streams, enabling image requests to run successfully with a draft acceptance rate of approximately 0.5 and byte-identical OCR output compared to non-speculative runs.
The llama.cpp project released version b11104, introducing a new feature that allows the llama-server component to bind to multiple network addresses simultaneously.
The llama.cpp project released build b11102, which introduces support for converting MiMo-V2.6 models.
The llama.cpp project has released version b11101, which includes a build system fix allowing repeated find_package calls for the llama library.
The llama.cpp project released build b11100, which addresses a specific parsing issue with the Muse Glimmer model. The primary change involves fixing the first parser error encountered during tool-call execution for this model.
The llama.cpp project has released version b11097, which introduces an OpenCL binary kernel for the A8 Q4_0 non-MoE dp4a format.
The llama.cpp b11095 release introduces a new HMX-optimized GATED_DELTA_NET (GDN) implementation, alongside various optimizations for Hexagon and Flash Attention kernels.
The llama.cpp project released version b11090, which includes a fix for a sm_70 tile compilation error. The update generalizes the tile shape of the 5-argument load_ldmatrix function to resolve mismatches with Volta architectures.
The llama.cpp project has released build b11081, which includes several updates to the `test-llama-archs` testing infrastructure. Key changes involve making tensor data standard deviation configurable, expanding usage examples, and adding support for architecture regex patterns.
The llama.cpp project released build b11078, which introduces environment variable support for configuring key generation parameters in the `llama-server` component. This update allows users to set temperature, top-p, min-p, and various penalties via `LLAMA_ARG_*` variables.
The llama.cpp b11080 release updates the test backend operations to support regular expressions in the `-o` filter, allowing pattern matching for op names while preserving exact matches for plain names. It also modifies the `cont` tool to stop printing Fast Attention vector slice logs when they are not needed.
The llama.cpp b11077 release addresses a bug in router mode where the `--api-key-file` argument was incorrectly forwarded to router-spawned child instances. This caused clients using `--api-key` to receive 401 errors on chat completions and prevented internal router stream calls from working.
The llama.cpp project released build b11074, which includes a fix for JSON enum handling. The update adds common_json_value support for enum values and simplifies the code by delegating the enum constructor to the underlying type's constructor.
The llama.cpp project released version b11073, which includes a change to the SYCL backend. This update coalesces MKL-FA softmax loads instead of assigning one work-item per row.
The llama.cpp project released version b11070, featuring a significant overhaul of the Hexagon backend to support 64-bit memory mappings and extended buffer handling. This update enables proper DMA (Direct Memory Access) support for larger models and complex operations on Qualcomm Hexagon DSPs.
The llama.cpp project has released build b11067, which introduces a fused GDN + copy operation for the WebGPU backend.