PR #3148 sync : ggml - SemanticDiff

sycl: addressing non-contiguous src1 mul_mats (nc and batched) (llama/13343)

ggerganov committed 1 year ago

vulkan: Allow up to 4096 elements for mul_mat_id row_ids (llama/13326)

ggerganov committed 1 year ago

rpc : add rpc_msg_set_tensor_hash_req (llama/13353)

ggerganov committed 1 year ago

CUDA: fix crash on large batch size for MoE models (llama/13384)

ggerganov committed 1 year ago

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

ggerganov committed 1 year ago

sycl : implementation of reordered Q4_0 MMVQ for Intel GPUs (llama/12858)

ggerganov committed 1 year ago

vulkan: scalar flash attention implementation (llama/13324)

ggerganov committed 1 year ago

CUDA: fix FlashAttention on Turing (llama/13415)

ggerganov committed 1 year ago

CUDA: fix race conditions FlashAttention kernels (llama/13438)

ggerganov committed 1 year ago

Add `--no-op-offload` to improve `-ot` pp perf in MoE models like llama4 400B (llama/13386)

ggerganov committed 1 year ago

CUDA: fix crash with partial offloading of MoE (llama/13439)

ggerganov committed 1 year ago

enable dpcpp nightly builds with libraries (llama/13406)

ggerganov committed 1 year ago

CUDA: fix misaligned synchronization in FA (llama/13469)

ggerganov committed 1 year ago

ggml-cpu: Integrate fp32=bf16xbf16 SME KleidiAI kernel (llama/13053)

ggerganov committed 1 year ago

llama/ggml: add LLM training support (llama/10544)

ggerganov committed 1 year ago

opencl: remove unnecessary assert for `add` (llama/13257)

ggerganov committed 1 year ago

metal : optimize MoE for large batches (llama/13388)

ggerganov committed 1 year ago

ggml : add mrope kernel for metal (llama/13457)

ggerganov committed 1 year ago

sync : ggml

ggerganov committed 1 year ago

whisper : update to ggml-backend changes (#0)

ggerganov committed 1 year ago

talk-llama : sync llama.cpp

ggerganov committed 1 year ago

whisper.cpp sync : ggml #3148 Merged