HIP: bump requirement to rocm 6.1 (llama/15296)

CUDA: Optimize `reduce_rows_f32` kernel, leading up to 25x perf improvement on kernel-level and 10% perf increase for Gemma3n (llama/15132)

musa: fix failures in test-backend-ops for mul_mat_id op (llama/15236)

CUDA: GEMM for FP32/FP16/BF16 and ne11 <= 16 (llama/15131)

llama : add gpt-oss (llama/15091)

HIP: enable mfma mmq on gfx908 and gfx90a for select datatypes and shapes (llama/14949)

CUDA: skip masked KV slices for all FA kernels (llama/14924)

HIP: remove the use of __HIP_PLATFORM_AMD__, explicitly support only AMD targets (llama/14945)

HIP: add GGML_HIP_MMQ_MFMA option to allow disableing the MFMA path. (llama/14930)

HIP: Enable Matrix cores for MMQ Kernels, Enable stream-K for CDNA 3 (llama/14624)

musa: upgrade musa sdk to rc4.2.0 (llama/14498)

musa: fix build warnings (unused variable) (llama/14561)

CUDA: add dynamic shared mem to softmax, refactor general usage (llama/14497)

musa: enable fp16 mma (all) and cublas on qy2 (llama/13842)

CUDA/HIP: optimize mmv paths taken for HIP devices (llama/14324)

CUDA: mul_mat_v support for batch sizes > 1 (llama/14262)

HIP: enable vec fattn on RDNA4 (llama/14323)

CUDA: add mean operation (llama/14313)

cuda : synchronize graph capture and cublas handle destruction (llama/14288)

HIP: disable rocwmma on gfx12 by default until rocm 7.0 (llama/14202)

HIP: Replace usage of depricated preprocessor macro __AMDGCN_WAVEFRONT_SIZE__ (llama/14183)

ggml-cpu : split arch-specific implementations (llama/13892)

CUDA: add a prop in ggml_cuda_device_infor for distinguish iGPU or dGPU in cuda (#13856) (llama/13895)

cuda : avoid cuGetErrorString (llama/13791)

CUDA: FA support for Deepseek (Ampere or newer) (llama/13306)

whisper: remove MSVC warnings pragmas (#3090)

musa: fix typo in cc control (llama/13144)

Simplify and improve CUDA graphs through use of indirect copy pointers (llama/9017)

musa: fix all warnings, re-enable `-DLLAMA_FATAL_WARNINGS=ON` in ci and update doc (llama/12611)

HIP: Add support for RDNA4 targets (llama/12372)

CUDA: Fix clang warnings (llama/12540)

musa: refine compute capability (llama/12493)

cuda : enable CUDA Graph on CUDA Toolkit < 12.x (llama/12394)

CUDA/HIP: refractor mmqv to unify the calculation of nwarps and rows per block between host and device code. (llama/12177)

HIP: implement FlashAttention via rocWMMA for CDNA and RDNA3+ (llama/12032)

CUDA: app option to compile without FlashAttention (llama/12025)

MUSA: support ARM64 and enable dp4a .etc (llama/11843)

CUDA: use async data loading for FlashAttention (llama/11894)

CUDA: fix CUDART_VERSION checks (llama/11821)

CUDA: use arch list for compatibility check (llama/11775)

CUDA/HIP: add support for selectable warp size to mmv (llama/11519)

HIP: add GGML_CUDA_CC_IS_* for amd familys as increasing cc archtectures for amd gpus are not supersets of eatch other (llama/11601)

CUDA: use mma PTX instructions for FlashAttention (llama/11583)

HIP: Prepare reduction operators for wave 64

CUDA/HIP: add warp_size to cuda_device_info

AMD: parse the architecture as supplied by gcnArchName (llama/11244)

Hip: disable VMM on hip as it seams that it dosent work in some configurations (llama/11420)

hip : Add hipGraph and VMM support to ROCM (llama/11362)

CUDA: rename macros to avoid conflicts with WinAPI (llama/10736)

Add some minimal optimizations for CDNA (llama/10498)