gguf-py : add Numpy MXFP4 de/quantization support (llama/15111)

ggml : fix field name when new ggml_backend (llama/14944)

CUDA: attention sinks for mma FlashAttention (llama/15157)

opencl: support sink in `soft_max` (attn sinks) (llama/15152)

vulkan: support fattn sinks (llama/15126)

vulkan: Add env var to disable host visible vidmem (llama/15109)

HIP: add cmake option to enable compiler output of kernel resource usage metrics (llama/15103)

ggml: Skip backend library linking code when GGML_BACKEND_DL=ON (llama/15094)

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

fix profiling crash (llama/15072)

opencl: add `swiglu_oai` and `add_id` (llama/15121)

ggml : fix fallback to CPU for ununsupported ops (llama/15118)

CANN: add support for ACL Graph (llama/15065)

llama : add gpt-oss (llama/15091)

sycl: fix mul_mat selection (llama/15092)

cmake: Add GGML_BACKEND_DIR option (llama/15074)

vulkan: fix build when using glslang that does not support coopmat2 (llama/15062)

vulkan: Use coopmat2 for conv2d (llama/14982)

opencl: fix adreno compiler detection logic (llama/15029)

CUDA: use mma FA kernel for gqa > 4 on RTX 4000 (llama/15035)

cuda: make im2col a little faster (llama/15025)

cuda, sycl : fix batched gemm when ne02 == 1 && ne03 > 1 (llama/15038)

vulkan: coopmat2 mul_mat optimizations (llama/14934)

vulkan: Support ne[3]>1 in noncontig matrix-vector multiply (llama/15015)

vulkan: optimizations for direct convolution (llama/14933)

CUDA: fix MMQ nwarps for AMD with warp_size==32 (llama/15014)

opencl: add f16 for `add`, `sub`, `mul`, `div` (llama/14984)

ggml : Q2k interleaving implementation - x86/x64 SIMD (llama/14373)

docker : add cann build pipline (llama/14591)

Vulkan: Fix minor debug mode issues (llama/14899)

CANN: Improve loading efficiency after converting weights to NZ format. (llama/14985)

opencl: add `mul_mat_f32_f32_l4_lm` and `mul_mat_f16_f32_l4_lm` (llama/14809)

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: Ignore unsupported unroll transformation in fattn-vec (llama/14931)

CANN: Add ggml_set_rows (llama/14943)

cuda : add softcap fusion (llama/14907)

CUDA: add roll (llama/14919)

ggml-cpu : deduplicate scalar implementations (llama/14897)

SYCL: Add set_rows support for quantized types (llama/14883)

CUDA: fix pointer incrementation in FA (llama/14916)

sycl: refactor quantization to q8_1 (llama/14815)

cmake : Fix BLAS link interface (ggml/1316)

vulkan : fix 32-bit builds (ggml/1313)

scripts : update sync scripts

node : add win platform check for require path (#3363)

ci : update main-cuda.Dockerfile (#3371)

whisper : fixed crash in GPU device selection on multi-GPU systems (#3372)