sync : ggml#3148

Merged

ggerganov merged 21 commits intomasterfrom

sync-ggml-25-05-13

May 13, 2025

Commits on May 13, 2025

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

Alcpz
authored and
ggerganov
committed
vulkan: Allow up to 4096 elements for mul_mat_id row_ids (llama/13326)

jeffbolznv
authored and
ggerganov
committed
rpc : add rpc_msg_set_tensor_hash_req (llama/13353)

rgerganov
authored and
ggerganov
committed
CUDA: fix crash on large batch size for MoE models (llama/13384)

JohannesGaessler
authored and
ggerganov
committed
CUDA: FA support for Deepseek (Ampere or newer) (llama/13306)

JohannesGaessler
authored and
ggerganov
committed
sycl : implementation of reordered Q4_0 MMVQ for Intel GPUs (llama/12858)

authored and
ggerganov
committed
vulkan: scalar flash attention implementation (llama/13324)

jeffbolznv
authored and
ggerganov
committed
CUDA: fix FlashAttention on Turing (llama/13415)

JohannesGaessler
authored and
ggerganov
committed
CUDA: fix race conditions FlashAttention kernels (llama/13438)

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

hjc4869
authored and
ggerganov
committed
CUDA: fix crash with partial offloading of MoE (llama/13439)

JohannesGaessler
authored and
ggerganov
committed
enable dpcpp nightly builds with libraries (llama/13406)

AD2605
authored and
ggerganov
committed
CUDA: fix misaligned synchronization in FA (llama/13469)

JohannesGaessler
authored and
ggerganov
committed
ggml-cpu: Integrate fp32=bf16xbf16 SME KleidiAI kernel (llama/13053)

authored and
ggerganov
committed
llama/ggml: add LLM training support (llama/10544)

JohannesGaessler
authored and
ggerganov
committed
opencl: remove unnecessary assert for `add` (llama/13257)

lhez
authored and
ggerganov
committed
metal : optimize MoE for large batches (llama/13388)
ggerganov
committed
ggml : add mrope kernel for metal (llama/13457)

ngxson
authored and
ggerganov
committed
sync : ggml
ggerganov
committed
whisper : update to ggml-backend changes (#0)
ggerganov
committed
talk-llama : sync llama.cpp
ggerganov
committed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

sync : ggml#3148

sync : ggml#3148
ggerganov merged 21 commits intomasterfrom
sync-ggml-25-05-13

Commits on May 13, 2025

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

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

rpc : add rpc_msg_set_tensor_hash_req (llama/13353)

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

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

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

vulkan: scalar flash attention implementation (llama/13324)

CUDA: fix FlashAttention on Turing (llama/13415)

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

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

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

enable dpcpp nightly builds with libraries (llama/13406)

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

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

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

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

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

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

sync : ggml

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

talk-llama : sync llama.cpp

sync : ggml#3148ggerganov merged 21 commits intomasterggml-org/whisper.cpp:masterfrom sync-ggml-25-05-13ggml-org/whisper.cpp:sync-ggml-25-05-13Copy head branch name to clipboard

Commits

Commits on May 13, 2025

sync : ggml#3148
ggerganov merged 21 commits intomasterfrom
sync-ggml-25-05-13