sync : ggml#3940

Merged

ggerganov merged 72 commits into

sync-ggml-26-07-10

Jul 10, 2026

Commits on Jul 10, 2026

vulkan: add INTEL_XE1 arch enum and enable coopmat1 on Intel Xe-LPG Plus (llama/24404)

authored and
ggerganov
committed
vulkan: opt mul_mat_vecq for mi50 (llama/22933)

chraac
authored and
ggerganov
committed
openvino: Update to OV 2026.2.1, self-contained release packages, operator improvements (llama/24974)

authored and
ggerganov
committed
sched : reintroduce less synchronizations during split compute (llama/20793)

committed
vulkan: fix step operator for 0 input (llama/25036)

0cc4m
authored and
ggerganov
committed
sycl : fix failed ut cases of norm (llama/25044)

arthw
authored and
ggerganov
committed
Added a cudaMemcpy2DAsync fast path to ggml_cuda_cpy (llama/25057)

gaugarg-nv
authored and
ggerganov
committed
opencl: flash attention improvement (llama/25069)

authored and
ggerganov
committed
vulkan: use flops instead of weight tensor size for submission heuristic (llama/25005)

0cc4m
authored and
ggerganov
committed
Revert "sched : reintroduce less synchronizations during split compute (#20793)" (llama/25138)

ORippler
authored and
ggerganov
committed
ggml-webgpu: add support for NVFP4 (llama/25143)

yomaytk
authored and
ggerganov
committed
HIP: use hipBLAS for dense prefill on gfx900, keep MMQ for MoE (llama/24588)

DEV-DUFORD
authored and
ggerganov
committed
vulkan: roll bk loop in matmul for asahi linux (llama/24663)

xingjianll
authored and
ggerganov
committed
CUDA: fix Gemma E4B MTP FlashAttention (llama/25148)

JohannesGaessler
authored and
ggerganov
committed
CUDA: fix get_rows_back for tables with more than 65535 rows (grid-y clamp + stride) (llama/25103)

mattjallo
authored and
ggerganov
committed
cuda : prevent integer truncation and overflow errors when using KQ mask strides in flash_attn_mask_to_KV_max kernel (llama/24945)

authored and
ggerganov
committed
opencl: initial q1_0 support (llama/25160)

lhez
authored and
ggerganov
committed
ggml-cpu: add AVX2 optimization for nvfp4 dot product and use UE4M3 LUT (llama/23961)

ragz4125
authored and
ggerganov
committed
CUDA: consistent use of __restrict__ + PDL for FA (llama/25185)

JohannesGaessler
authored and
ggerganov
committed
hexagon: flash attention rework (optimizations, accuracy improvements, etc) (llama/25085)

max-krasnyansky
authored and
ggerganov
committed
opencl: allow loading precompiled binary kernels from library (llama/23042)

lhez
authored and
ggerganov
committed
Remove redundant CUDA copies after gated_delta_net. (llama/23940)

gaugarg-nv
authored and
ggerganov
committed
cuda: enable topk-moe fusion for 288 experts (llama/25267)

authored and
ggerganov
committed
ggml : fix broken CPU concat implementation for quantized types (llama/25247)

authored and
ggerganov
committed
cuda : concat implementation for quantized types (llama/25303)

authored and
ggerganov
committed
ggml: Update VMM Pool allocation ggml-cuda.cu - Turing P2P access fix (fixes #24489) (llama/24491)

authored and
ggerganov
committed
ggml : fix tensor-parallel + -ncmoe crash on MoE models (llama/25028)

liminfei-amd
authored and
ggerganov
committed
abort if we see a multi buffer (llama/25276)

netrunnereve
authored and
ggerganov
committed
ggml-cuda: optimize conv_transpose_1d indexing (llama/25310)

adavyas
authored and
ggerganov
committed
ggml-hip: enable -ffast-math for HIP builds (llama/23862)

a-huk
authored and
ggerganov
committed
vulkan: fix 32-bit integer overflow in CEIL_DIV (llama/25245)

hokanosekai
authored and
ggerganov
committed
ggml-cpu: Enable tiled matmul on AIX (llama/25199)

shalinib-ibm
authored and
ggerganov
committed
ggml-cpu: use UE4M3 LUT in ARM NVFP4 dot product (llama/25331)

ragz4125
authored and
ggerganov
committed
CUDA: extend K-type validation to V-types for flash attention (llama/24403)

sanmai
authored and
ggerganov
committed
CUDA: remove -sm row, refactor cuBLAS (llama/24216)

JohannesGaessler
authored and
ggerganov
committed
metal: add col2im_1d op (f32/f16/bf16) (llama/25176)

authored and
ggerganov
committed
opencl: general flash attention decode performance optimizations (llama/25366)

authored and
ggerganov
committed
vulkan : check src0 type in GGML_OP_SET_ROWS to avoid failures due to unimplemented f16 support (llama/25351)

authored and
ggerganov
committed
ggml : make ggml_time_init idempotent (llama/24422)

aisk
authored and
ggerganov
committed
sycl : rename the env vars from "disable" to "enable" (llama/25042)

arthw
authored and
ggerganov
committed
sycl : use sycl func to fix AOT double type issue (llama/25081)

arthw
authored and
ggerganov
committed
sycl : enhance argsort to support all UT cases (llama/25125)

arthw
authored and
ggerganov
committed
fix unsupport ACC UT cases for noncontiguous (llama/25124)

arthw
authored and
ggerganov
committed
sycl : set K_QUANTS_PER_ITERATION to 1 on DMMV path (llama/25063)

malsbat
authored and
ggerganov
committed
support OP cross_entropy_loss, cross_entropy_loss_back (llama/25236)

arthw
authored and
ggerganov
committed
support op col2im_1d (llama/25264)

arthw
authored and
ggerganov
committed
fix unsupported UT cases of CONT & CPY (llama/25231)

arthw
authored and
ggerganov
committed
ggml-hip : add -fno-finite-math-only alongside -ffast-math (llama/25373)

asf0
authored and
ggerganov
committed
CUDA: Fuse MMVQ post-scale for NVFP4 (llama/24481)

committed
Add Q2_0 quantization: type definition and CPU backend (llama/24448)

khosravipasha
authored and
ggerganov
committed
opencl: fix potential crash in aos reconstruct (llama/25383)

lhez
authored and
ggerganov
committed
ggml : add support for CPU f16->f16 GGML_OP_SET_ROWS (llama/25344)

authored and
ggerganov
committed
ggml : fix A indexing in simd_gemm scalar tail-column path (llama/25390)

tyronecai
authored and
ggerganov
committed
metal : add set_rows with src0 f16 (llama/25434)

fairydreaming
and
ggerganov
committed
cuda : add support for f16->f16 GGML_OP_SET_ROWS (llama/25367)

fairydreaming
authored and
ggerganov
committed
hexagon: new vtcm layouts and improved pipelines for MUL_MAT, MUL_MAT_ID and FLASH_ATTN_EXT (llama/25425)

max-krasnyansky
authored and
ggerganov
committed
vulkan: for small AMD GPUs, reduce submission threshold based on CU count (llama/25240)

0cc4m
authored and
ggerganov
committed
opencl: ragged-tile MoE prefill FP16 GEMM optimization (skip padded expert tiles) (llama/25433)

authored and
ggerganov
committed
vulkan: disable FA mask_opt on GCN to improve performance (llama/24362)

0cc4m
authored and
ggerganov
committed
opencl: Q6_K GEMM/GEMV fix for ne01 of weights that are not multiples of 128. (llama/25464)

authored and
ggerganov
committed
ggml-webgpu: tune subgroup split (d_split) in flash_attn_vec (llama/25418)

yomaytk
authored and
ggerganov
committed
hexagon: add VISION RoPE support (llama/25216)

aparmp-quic
authored and
ggerganov
committed
cuda: align snake fusion matcher with the other backends (llama/25460)

ServeurpersoCom
authored and
ggerganov
committed
ggml-hip: enable -funsafe-math-optimizations (llama/24668)

authored and
ggerganov
committed
metal : add CONV_2D_DW (depthwise convolution) support (llama/21565)

Sou-ly
and
ggerganov
committed
Only index by compile times + always multiply/add (llama/25445)

ORippler
authored and
ggerganov
committed
common : adapt to q2_0 (ggml/0)
ggerganov
committed
ggml : fix conv 2d dw (llama/25490)
ggerganov
committed
ggml : bump version to 0.16.0 (ggml/1559)
ggerganov
committed
sync : ggml
ggerganov
committed
talk-llama : sync llama.cpp
ggerganov
committed
ggml : use ggml_vqtbl1q_u8 for 32-bit compat (#0)
ggerganov
committed