sync : ggml#3803

Merged

ggerganov merged 73 commits into

sync-ggml-26-05-10

May 14, 2026

Commits on May 10, 2026

opencl: Adreno optimization for MoE - MxFP4 (llama/22301)

authored and
ggerganov
committed
ggml-virtgpu: fix circular dependency in headers (llama/22557)

Juste-Leo2
authored and
ggerganov
committed
fix: CUDA device PCI bus ID de-dupe OOMing (ignoring other 3 gpus entirely) (llama/22533)

authored and
ggerganov
committed
ggml-webgpu: add layer norm ops (llama/22406)

Constannnnnt
authored and
ggerganov
committed
vulkan: delete dead GGML_VK_MAX_NODES def (llama/22621)

Atomic-Germ
authored and
ggerganov
committed
CUDA: use fastdiv for batch index split in get_rows (llama/22650)

leonardHONG
authored and
ggerganov
committed
kleidiai : update to v1.24.0 and use release archive (llama/22549)

chaxu01
authored and
ggerganov
committed
ggml : implement fast walsh-hadamard transform for kv rotation (#21352) (llama/22631)

AlrIsmail
authored and
ggerganov
committed
llama : add option to save memory in device buffers (llama/22679)
ggerganov
committed
ggml : bump version to 0.11.0 (ggml/1478)
ggerganov
committed
rpc : use graph uid instead of graph cache (llama/22701)

rgerganov
authored and
ggerganov
committed
opencl: refactor Adreno q4_0 (llama/22335)

lhez
authored and
ggerganov
committed
Hexagon: Process M-tail rows on HMX instead of HVX (llama/22724)

authored and
ggerganov
committed
ggml : use `CL_DEVICE_GLOBAL_MEM_SIZE` as memory estimate for OpenCL --fit (llama/22688)

fl0rianr
authored and
ggerganov
committed
ggml-cpu: fuse RMS_NORM + MUL on CPU backend (llama/22423)

zzzzwc
authored and
ggerganov
committed
ggml-cpu: Optimized risc-v cpu q1_0 dot

pl752
authored and
ggerganov
committed
sycl: add FILL, CUMSUM, DIAG, SOLVE_TRI, SSM_SCAN, GATED_DELTA_NET (llama/22149)

authored and
ggerganov
committed
opencl: add opfilter regex for debugging (llama/22782)

shaofeiqi
authored and
ggerganov
committed
llama : fix device state save/load (llama/22805)
ggerganov
committed
CUDA: batch out_prod inner loop with cublasSgemmStridedBatched (llama/22651)

leonardHONG
authored and
ggerganov
committed
opencl: add q4_0 MoE GEMM for Adreno (llama/22731)

authored and
ggerganov
committed
ggml: update SCHED_DEBUG output to use ggml_op_desc() (llama/22825)

max-krasnyansky
authored and
ggerganov
committed
vulkan: fix spv shadowing (llama/22760)

miyanyan
authored and
ggerganov
committed
CUDA: lower-case PCI bus id, standardize for ggml (llama/22820)

JohannesGaessler
authored and
ggerganov
committed
cuda: fuse snake activation (mul, sin, sqr, mul, add) (llama/22667)

authored and
ggerganov
committed
Feature hexagon l2 norm (llama/22816)

authored and
ggerganov
committed
sycl: support non-contiguous input in PAD op (llama/22148)

authored and
ggerganov
committed
hexagon: add HTP kernel for GGML_OP_GATED_DELTA_NET (llama/22837)

wyanzhao
authored and
ggerganov
committed
Add flash attention MMA / Tiles to support MiMo-V2.5 (llama/22812)

AesSedai
authored and
ggerganov
committed
sycl: Battlemage AOT build via spir64_gen + MMQ subgroup annotations (llama/22147)

authored and
ggerganov
committed
sycl: Q5_K reorder MMVQ/dequant + Q8_0 reorder MMVQ path (llama/22152)

authored and
ggerganov
committed
Add BF16 support to GET_ROWS operation (llama/21391)

devedse
authored and
ggerganov
committed
SYCL: reduce allocation overhead during flash attention (llama/22732)

sanmai
authored and
ggerganov
committed
internal AllReduce kernel for CUDA provider (llama/22299)

authored and
ggerganov
committed
ggml : bump version to 0.11.1 (ggml/1484)
ggerganov
committed
sync : ggml
ggerganov
committed
talk-llama : sync llama.cpp
ggerganov
committed

Commits on May 14, 2026

Revert "ggml-cuda : add ar_add() to avoid ambiguous operator+ for half/bfloat16 in CUDA 11.8"
danbev
committed
Revert "ggml-cuda : add explicit casts to -INFINITY for float and half2 types"
danbev
committed
ggml: install ggml.pc in <libdir>/pkgconfig (ggml/1480)

robUx4
authored and
ggerganov
committed
metal : tighten input-position loop in kernel_conv_transpose_1d (ggml/1477)

CrispStrobe
authored and
ggerganov
committed
ggml-virtgpu : include missing mutex header (llama/22810)

olliewalsh
authored and
ggerganov
committed
Add OP im2col_3d (llama/22903)

arthw
authored and
ggerganov
committed
CUDA: directly include cuda/iterator (llama/22936)

ORippler
authored and
ggerganov
committed
vulkan: Support asymmetric FA in scalar/mmq/coopmat1 paths (llama/22589)

jeffbolznv
authored and
ggerganov
committed
Ggml/cuda snake fusion hardening (llama/22912)

ServeurpersoCom
authored and
ggerganov
committed
CUDA: handle OW > 65535 in im2col (2D and 3D) (llama/22944)

CrispStrobe
authored and
ggerganov
committed
opencl: add q4_1 MoE for Adreno (llama/22856)

authored and
ggerganov
committed
metal : promote mul_mv/mul_mm batch divisors to function constants (llama/22711)

guyfischman
authored and
ggerganov
committed
vulkan: Check shared memory size for mmq shaders (llama/22693)

jeffbolznv
authored and
ggerganov
committed
vulkan: Fix Windows performance regression on Intel GPU BF16 workloads for Xe2 and newer (llama/22461)

rillomas
authored and
ggerganov
committed
ggml-webgpu: address precision issues for multimodal (llama/22808)

Constannnnnt
authored and
ggerganov
committed
ggml-webgpu: Enables running gpt-oss-20b (llama/22906)

yomaytk
authored and
ggerganov
committed
opencl: add opt-in Adreno xmem F16xF32 GEMM for prefill (llama/22755)

authored and
ggerganov
committed
hexagon: eliminate scalar VTCM loads via HVX splat helpers (llama/22993)

authored and
ggerganov
committed
ggml-zendnn : adaptive fallback to CPU backend for small batch sizes (llama/22681)

z-sachin
authored and
ggerganov
committed
hexagon: add unary tanh op (llama/22999)

max-krasnyansky
authored and
ggerganov
committed
flush the gpu profile timestamp before the queryset is overflowed (llama/22995)

yomaytk
authored and
ggerganov
committed
opencl: fix crash when warming up MoE on Adreno (llama/22876)

lhez
authored and
ggerganov
committed
opencl: add q5_0 and q5_1 MoE for Adreno (llama/22985)

authored and
ggerganov
committed
Fix for issue #22974. Cast intermediate results to float before adding and casting the result to the destination type. Avoids half+half operator ambiguity. (llama/22994)

scutler-nv
authored and
ggerganov
committed
ggml-webgpu: only use subgroup-matrix path when head dims are divisible by sg_mat_k / sg_mat_n (llama/23020)

ArberSephirotheca
authored and
ggerganov
committed
sync : ggml
ggerganov
committed
talk-llama : sync llama.cpp
ggerganov
committed