sync : ggml#3319

Merged

ggerganov merged 50 commits intomasterfrom

sync-ggml-25-07-12

Jul 12, 2025

Commits on Jul 12, 2025

ggml : add version function to get lib version (ggml/1286)

danbev
and
ggerganov
committed
vulkan : implement ggml_roll (ggml/1290)

Acly
authored and
ggerganov
committed
vulkan : implement bilinear interpolation for ggml_upscale/ggml_interpolate (ggml/1291)

Acly
authored and
ggerganov
committed
add GELU_ERF (llama/14455)

CISC
authored and
ggerganov
committed
vulkan: Split large mul_mat_id to fit in shared memory (llama/14451)

jeffbolznv
authored and
ggerganov
committed
CANN: update aclnnGroupedMatmulV2 to aclnnGroupedMatmulV3 (llama/14411)

noemotiovon
authored and
ggerganov
committed
ci : disable fast-math for Metal GHA CI (llama/14478)
ggerganov
committed
ggml : Callback before abort (llama/14481)

authored and
ggerganov
committed
opencl : update upscale to support align corners (llama/14488)

lhez
authored and
ggerganov
committed
opencl : skip empty nodes on cgraph compute (llama/14491)

EZForever
authored and
ggerganov
committed
opencl : fix possible buffer overflow in dump_tensor (llama/14490)

zhouwg
authored and
ggerganov
committed
ggml : support bcast ggml_soft_max_ext, ggml_flash_attn_ext (llama/14435)
ggerganov
committed
vulkan: support softmax/FA batch and broadcast (llama/14449)

jeffbolznv
authored and
ggerganov
committed
CUDA: broadcasting for FlashAttention mask (llama/14500)

JohannesGaessler
authored and
ggerganov
committed
CUDA: add softmax broadcast (llama/14475)

am17an
authored and
ggerganov
committed
llama : initial Mamba-2 support (llama/9126)

compilade
authored and
ggerganov
committed
CUDA: add dynamic shared mem to softmax, refactor general usage (llama/14497)

am17an
authored and
ggerganov
committed
ggml : fix FA mask dim 2 and 3 (llama/14505)
ggerganov
committed
kv-cache : use ggml_set_rows (llama/14285)
ggerganov
committed
Fix conditional enabling following arch checks for ggml-sycl (llama/14504)

s-Nick
authored and
ggerganov
committed
ggml: backward pass for split swiglu (llama/14483)

JohannesGaessler
authored and
ggerganov
committed
vulkan: support mixed/deepseekR1 FA head sizes (llama/14509)

jeffbolznv
authored and
ggerganov
committed
opencl : broadcast for soft_max (llama/14510)

lhez
authored and
ggerganov
committed
ggml : implement GEGLU_ERF and GEGLU_QUICK ops (llama/14445)

CISC
authored and
ggerganov
committed
CANN: Replace aclrtMemsetSync with aclnnInplaceZero operator (llama/14002)

authored and
ggerganov
committed
metal : disable fast math in all quantize kernels (llama/14528)
ggerganov
committed
opencl: add GELU_ERF (llama/14476)

CISC
authored and
ggerganov
committed
vulkan: Handle updated FA dim2/3 definition (llama/14518)

jeffbolznv
authored and
ggerganov
committed
vulkan: fix rms_norm+mul fusion (llama/14545)

jeffbolznv
authored and
ggerganov
committed
vulkan: increase LOAD_VEC_A to 8 (IQ1/IQ2) or 4 (IQ3) (llama/14485)

authored and
ggerganov
committed
CUDA: add bf16 and i32 to getrows (llama/14529)

am17an
authored and
ggerganov
committed
musa: fix build warnings (unused variable) (llama/14561)

yeahdongcn
authored and
ggerganov
committed
CUDA: add bilinear interpolation for upscale (llama/14563)

am17an
authored and
ggerganov
committed
cuda : fix rope with partial rotation and non-cont src (llama/14580)
ggerganov
committed
vulkan : fix rope with partial rotation and non-cont src (llama/14582)

jeffbolznv
authored and
ggerganov
committed
vulkan: optimize flash attention split_k_reduce (llama/14554)

jeffbolznv
authored and
ggerganov
committed
ggml : prevent integer overflow in gguf tensor size calculation (llama/14595)

Yuuoniy
authored and
ggerganov
committed
ggml : add ggml_scale_bias (llama/14417)

ngxson
authored and
ggerganov
committed
cuda : support Falcon-H1 state size for SSM_SCAN (llama/14602)

compilade
authored and
ggerganov
committed
SYCL: Initial set_rows kernel implementation (llama/14562)

qnixsynapse
authored and
ggerganov
committed
opencl: add `set_rows` for `f16` and `f32` (llama/14547)

lhez
authored and
ggerganov
committed
opencl: add tiled mul_mat_f16_f32 (llama/14535)

rmatif
authored and
ggerganov
committed
HIP : Add HIP 7.0+ compatibility for hipBLAS compute types (llama/14634)

slojosic-amd
authored and
ggerganov
committed
model : support LiquidAI LFM2 hybrid family (llama/14620)

tdakhran
authored and
ggerganov
committed
vulkan: optimizations for deepseek prompt processing (llama/14555)

jeffbolznv
authored and
ggerganov
committed
vulkan: support SET_ROWS (llama/14587)

jeffbolznv
authored and
ggerganov
committed
sync : resolve conflicts (ggml/0)
ggerganov
committed
sync : ggml
ggerganov
committed
talk-llama : sync llama.cpp
ggerganov
committed
sync : resolve conflicts (#0)
ggerganov
committed