cuda : fix tensor size calculation for non-split buffer (llama/5145)

ggml-alloc : add 10% margin to the buffer sizes (llama/5149)

ggml : update softmax n_task calculation (llama/5126)

metal : remove unused `n_buffers` and `buffers` (llama/5129)

metal : show compile log messages

cuda : fix 2-bit quants on amd hip (llama/5105)

llama : pre-allocate input tensors in a separate buffer (llama/5100)

metal : disable support for MUL_MAT F32 x F16

CUDA: more info when no device code (llama/5088)

minor : clean-up some warnings and style (llama/5094)

ggml : parallelize FP32 conversion when using BLAS (llama/5045)

llava : MobileVLM support (llama/4954)

llama : run all KQV ops on the CPU with no KV offload (llama/5049)

cuda : fix compile error in jetson platform (llama/4975)

ggml : check ggml_add src1 type (ggml/708)

docs : make model options / model install methods clearer (#1806)

cmake : make libwhisper.so position independent (#1792)

cmake : temporary remove VLA check (#1795)

whisper.android : return output from benchmarks (#1785)

server : implement "verbose_json" format with token details (#1781)

ggml : sync ggml-metal.m

sync : llama.cpp

sync : ggml

ggml : add IQ2 to test-backend-ops + refactoring (llama/4990)

imatrix : offload to GPU support (llama/4957)

backend : add eval callback (llama/4935)

metal : create autorelease pool during library build (llama/4970)

ggml : importance matrix support for legacy quants (llama/4969)

metal : log `recommendedMaxWorkingSetSize` on iOS 16+ (llama/4936)

ggml : introduce GGML_CALL function annotation (llama/4850)

cuda : fix dequantize kernel names (llama/4938)

CUDA: faster dequantize kernels for Q4_0 and Q4_1 (llama/4938)

Add ability to use importance matrix for all k-quants (llama/4930)

talk-llama : optional wake-up command and audio confirmation (#1765)

server : fix building and simplify lib deps on Windows (#1772)

talk-llama : sync llama.cpp

talk-llama : llama.cpp

sync : ggml

metal : correctly set SIMD support flags on iOS (llama/4923)

2-bit quantizations (llama/4897)

scripts : sync-ggml-am.sh add option to skip commits

talk-llama : sync llama.cpp

sync : ggml

examples : adapt to metal API

ggml: cache sin/cos for RoPE (llama/4908)

metal : remove old API (llama/4919)

metal : disable log for loaded kernels (llama/4794)

gguf : fix potential infinite for-loop (llama/4600)

metal : refactor kernel loading code (llama/4794)

CUDA: faster q8_0 -> f16 dequantization (llama/4895)