update#2

Open

drzo wants to merge 575 commits intocogpy:masterfrom

ggml-org:master

This pull request is big! We're only showing the most recent 250 commits

Commits on Dec 14, 2025

Commits on Dec 17, 2025

vulkan: fix mul_mat_vec_iq1_s formatting (llama/18026)

0cc4m
authored and
ggerganov
committed
Support gpt-oss by OPs add-id, mul_mat for mxfp4, swiglu_oai (llama/17826)

NeoZhangJianyu
authored and
ggerganov
committed
llama: automatically set parameters not set by the user in such a way that maximizes GPU utilization (llama/16653)

JohannesGaessler
authored and
ggerganov
committed
metal: use shared buffers on eGPU (llama/17866)

jdemeule
authored and
ggerganov
committed
ggml-hexagon: mm for mtmd (llama/17894)

joeldushouyu
authored and
ggerganov
committed
ggml : use WARP_SIZE/2 for argmax reduction offset (llama/18092)

Aadeshveer
authored and
ggerganov
committed
llama.android : Rewrite Android binding (w/o cpu_features dep) (llama/17413)

authored and
ggerganov
committed
sync : llama.cpp
ggerganov
committed

Commits on Dec 31, 2025

HIP: Refactor mma for RDNA and CDNA (llama/17990)

authored and
ggerganov
committed
ggml-cpu: ARM64: repack version of q8_0 (dotprod and i8mm) (llama/18096)

Alcpz
authored and
ggerganov
committed
ggml-hexagon: gelu operation (llama/17921)

joeldushouyu
authored and
ggerganov
committed
ggml-hexagon: swiglu_oai operation (llama/18114)

joeldushouyu
authored and
ggerganov
committed
remove i_major_dual (llama/18157)

authored and
ggerganov
committed
ggml-cpu: extend support for RVV floating-point kernels (llama/17318)

authored and
ggerganov
committed
model : add ASR support for LFM2-Audio-1.5B (conformer) (llama/18106)

authored and
ggerganov
committed
vulkan: Add perf logger mode with concurrency (llama/17944)

jeffbolznv
authored and
ggerganov
committed
ggml-hexagon: Implement true Q8_0 quantization on Hexagon NPU for more accurate mixed-precision matmul operations (llama/17977)

authored and
ggerganov
committed
Added comments explaining thread block size selection logic based on row count and column size, derived from historical commit context (llama/18212)

Aadeshveer
authored and
ggerganov
committed
test-backend-ops: improve msvc build time (llama/18209)

jeffbolznv
authored and
ggerganov
committed
tests: Avoid floating point precision false positives in SUM (llama/17471)

jeffbolznv
authored and
ggerganov
committed
Vulkan: some improvement on mul_mat_iq2_xs (llama/18031)

lovedheart
authored and
ggerganov
committed
vulkan: in graph_optimize, try to group ADD operations (llama/18060)

jeffbolznv
authored and
ggerganov
committed
vulkan: support GGML_UNARY_OP_XIELU (llama/18062)

jeffbolznv
authored and
ggerganov
committed
vulkan/cuda: fix topk_moe with exp_probs_b (llama/18071)

jeffbolznv
authored and
ggerganov
committed
vulkan: fix im2col overflowing maxworkgroupcount (llama/18180)

jeffbolznv
authored and
ggerganov
committed
llama: fix RPC for -fit on (llama/18233)

JohannesGaessler
authored and
ggerganov
committed
vulkan: Implement set_tensor_async and the event interfaces (llama/18047)

jeffbolznv
authored and
ggerganov
committed
vulkan: Extend rope fusions to allow mrope (llama/18264)

jeffbolznv
authored and
ggerganov
committed
opencl: unpack q4_0 for adreno in get_tensor (llama/18278)

lhez
authored and
ggerganov
committed
llamafile: add rvv support for sgemm kernels (llama/18199)

authored and
ggerganov
committed
ggml-hexagon: gelu optimization (llama/18151)

authored and
ggerganov
committed
ggml-hexagon: create generalized functions for cpu side op (llama/17500)

authored and
ggerganov
committed
rpc : add check for rpc buffer type (llama/18242)

struct
authored and
ggerganov
committed
CANN: Uses yarn_ramp cache in ROPE (llama/17725)

TianHao324
authored and
ggerganov
committed
vulkan: use fewer FA rows for small cache runs (llama/18280)

0cc4m
authored and
ggerganov
committed
CANN : refactor ACL graph cache (llama/17752)

wangweixuan
authored and
ggerganov
committed
vulkan: fix command buffer corruption in ggml_backend_vk_event_wait (llama/18302)

jeffbolznv
authored and
ggerganov
committed
CUDA: experimental native mxfp4 support for blackwell (llama/17906)

am17an
authored and
ggerganov
committed
ggml : optimize cuda cumsum fallback kernel (llama/18343)

Aadeshveer
authored and
ggerganov
committed
CANN: Add support for CONV_TRANSPOSE_1D when kernel size > 255 (llama/17934)

Intellouis
authored and
ggerganov
committed
ggml-cuda: fix blackwell native builds (llama/18361)

am17an
authored and
ggerganov
committed
cuda: optimize cumsum cub path (llama/18362)

am17an
authored and
ggerganov
committed
ggml-cuda: fix regex for arch list (llama/18371)

am17an
authored and
ggerganov
committed
CANN: implement the SSM_CONV operator (llama/17737)

authored and
ggerganov
committed
vulkan: handle rope with large number of rows (llama/18306)

jeffbolznv
authored and
ggerganov
committed
vulkan: Support UPSCALE w/antialias (llama/18327)

jeffbolznv
authored and
ggerganov
committed
vulkan: small dequantization improvements (llama/18380)

netrunnereve
authored and
ggerganov
committed
vulkan: Use BK=32 for coopmat2 mul_mat_id (llama/18332)

jeffbolznv
authored and
ggerganov
committed
vulkan: optimize decodeFuncB in coopmat2 mul_mat_id shader (llama/18349)

jeffbolznv
authored and
ggerganov
committed
vulkan: preprocess mul_mat_id experts and discard workgroups more quickly (llama/18352)

jeffbolznv
authored and
ggerganov
committed
ggml-cuda: Use same regex for GGML_NATIVE=OFF (llama/18407)

am17an
authored and
ggerganov
committed
opencl: allow resizing transpose buffers (llama/18384)

lhez
authored and
ggerganov
committed
ggml-cuda: use CMAKE_CUDA_ARCHITECTURES if set when GGML_NATIVE=ON (llama/18413)

QDelta
authored and
ggerganov
committed
cmake: Added more x86_64 CPU backends when building with `GGML_CPU_ALL_VARIANTS=On` (llama/18186)

bberberov
authored and
ggerganov
committed
rpc: fix segfault on invalid endpoint format (llama/18387)

o7si
authored and
ggerganov
committed
Revert "ggml-cuda: use CMAKE_CUDA_ARCHITECTURES if set when GGML_NATIVE=ON (#18413)" (llama/18426)

am17an
authored and
ggerganov
committed
HIP: Use mmq on MFMA devices for MUL_MAT_ID in cases where a lot of splits would be generated (llama/18202)

IMbackK
authored and
ggerganov
committed
cuda: fix race condition in cumsum (llama/18448)

am17an
authored and
ggerganov
committed
CUDA: Blackwell features for non-native builds (llama/18436)

JohannesGaessler
authored and
ggerganov
committed
CUDA: fix replacment of bad archs in CMake (llama/18457)

JohannesGaessler
authored and
ggerganov
committed
CUDA: add log line when mxfp4 acceleration is used (llama/18483)

am17an
authored and
ggerganov
committed
kleidiai: add and integrate SVE 256-bit vector-length kernel (llama/18458)

chaxu01
authored and
ggerganov
committed
Work around broken IntelSYCLConfig.cmake in Intel oneAPI 2025.x (llama/18345)

rrsathe
authored and
ggerganov
committed
sycl: add newline at the end of CMakeLists.txt (llama/18503)

am17an
authored and
ggerganov
committed
sync : llama.cpp
ggerganov
committed
metal : remove BF16 x F16 kernels (llama/18456)
ggerganov
committed
CUDA: fix KQ max calculation (llama/18487)

JohannesGaessler
authored and
ggerganov
committed
metal : add count_equal op (llama/18314)

gatbontonpc
and
ggerganov
committed
sync : llama.cpp
ggerganov
committed
sync : whisper.cpp
ggerganov
committed
ggml : bump version to 0.9.5 (#1410)
ggerganov
authored

Commits on Jan 11, 2026

vulkan: extend topk_moe to handle sigmoid w/exp_probs_b for nemotron (llama/18295)

jeffbolznv
authored and
ggerganov
committed
ggml-cuda: remove unneccesary prints on ggml_cuda_init (llama/18502)

am17an
authored and
ggerganov
committed
cuda : fix copy of large tensors (ggml_nbytes <= INT_MAX assertion) (llama/18433)

Meet91721
authored and
ggerganov
committed
rpc : use unordered_map::reserve and emplace (llama/18513)

struct
authored and
ggerganov
committed
metal : adjust extra size for FA buffer to avoid reallocations (llama/18545)
ggerganov
committed
vulkan: Implement mmvq for iq1_s/iq1_m (llama/18450)

jeffbolznv
authored and
ggerganov
committed
vulkan: Optimize GGML_OP_CUMSUM (llama/18417)

jeffbolznv
authored and
ggerganov
committed
ggml-hexagon: optimize activation function (llama/18393)

joeldushouyu
authored and
ggerganov
committed
(Bugfix, ggml-cuda) Pool alloc count fix + small size computation type adjustment (llama/18559)

authored and
ggerganov
committed
CUDA: only allocate FA tmp buffer if needed (llama/18564)

JohannesGaessler
authored and
ggerganov
committed
ggml-cuda: fixes for concurrent streams (llama/18496)

am17an
authored and
ggerganov
committed
ggml-cuda: remove unused params in ggml_cuda_graph (llama/18579)

am17an
authored and
ggerganov
committed
CUDA: disable cuda graph when using n-cpu-moe (llama/18593)

am17an
authored and
ggerganov
committed
sampling : add support for backend sampling (llama/17004)

danbev
authored and
ggerganov
committed
CANN: add operator fusion support for ADD + RMS_NORM (llama/17512)

noemotiovon
authored and
ggerganov
committed
vulkan: handle quantize_q8_1 overflowing the max workgroup count (llama/18515)

jeffbolznv
authored and
ggerganov
committed
vulkan: fix topk_moe_sigmoid_norm_bias failures in GLM-4.6 (llama/18582)

jeffbolznv
authored and
ggerganov
committed
ggml-cuda: check for srcs outside the cgraph (llama/18583)

am17an
authored and
ggerganov
committed
CUDA: fix FA FP16 accumulator overflow for Granite (llama/18614)

JohannesGaessler
authored and
ggerganov
committed
ggml webgpu: add CEIL operation support (llama/18605)

tnguyen21
authored and
ggerganov
committed
CANN: Make `valid_values` variable `static const` (llama/18627)

rauletorresc
authored and
ggerganov
committed
ggml : fix avx512bf16 build (llama/18623)

angt
authored and
ggerganov
committed
mmq.cu: tune mmq/rocblas switching for RDNA (llama/18537)

authored and
ggerganov
committed
ggml-cuda: refactor cuda graph usage (llama/18637)

am17an
authored and
ggerganov
committed
vulkan: support buffer_from_host_ptr (llama/18467)

jeffbolznv
authored and
ggerganov
committed
ggml : optimize cuda ssm_scan using warp-level reduction (llama/18505)

Aadeshveer
authored and
ggerganov
committed
Hexagon add support for f16/f32 flash attention, scale, set-rows and improve f16/32 matmul (llama/18611)

max-krasnyansky
authored and
ggerganov
committed
CANN: Rename `get_env` to `get_env_as_lowercase` (llama/18624)

rauletorresc
authored and
ggerganov
committed
CANN: Fix rename for get_env (llama/18652)

hipudding
authored and
ggerganov
committed
vulkan: more mul mat optimizations (llama/18533)

netrunnereve
authored and
ggerganov
committed
vulkan: Warptile tuning for Intel Xe2/Xe3 (llama/18178)

virajwad
authored and
ggerganov
committed
vulkan: reject ops when a tensor is too large to allocate (llama/18646)

jeffbolznv
authored and
ggerganov
committed
cuda : fix build on cuda 12.8 (llama/18672)

olliewalsh
authored and
ggerganov
committed
opencl: add FILL op support (llama/18682)

shaofeiqi
authored and
ggerganov
committed
ggml: add env var GGML_OP_OFFLOAD_MIN_BATCH (llama/18535)

authored and
ggerganov
committed
metal : add MoE kernel specialization for ne20=5 (llama/18667)

dororodoroddo
authored and
ggerganov
committed
vulkan: optimize ssm_scan (llama/18630)

jeffbolznv
authored and
ggerganov
committed
vulkan: fix push constant size for quantize_q8_1 (llama/18687)

jeffbolznv
authored and
ggerganov
committed
ggml webgpu: initial flashattention implementation (llama/18610)

reeselevine
authored and
ggerganov
committed
ggml-webgpu: Fix GGML_MEM_ALIGN to 8 for emscripten. (llama/18628)

yomaytk
authored and
ggerganov
committed
llama: use host memory if device reports 0 memory (llama/18587)

taronaeo
authored and
ggerganov
committed
sync : llama.cpp
ggerganov
committed
Updates to webgpu get_memory (llama/18707)

reeselevine
authored and
ggerganov
committed
opencl: add EXPM1 op (llama/18704)

shaofeiqi
authored and
ggerganov
committed
Corrected: changed s13 = src1->nb[3] instead of nb[2] (llama/18724)

michaelw9999
authored and
ggerganov
committed
cmake : update blas logic (llama/18205)

DaAwesomeP
authored and
ggerganov
committed
HIP: adjust RDNA3.5 MMQ kernel selction logic (llama/18666)

JohannesGaessler
authored and
ggerganov
committed
test-backend-ops: fix mxfp4 tests on blackwell (llama/18736)

am17an
authored and
ggerganov
committed
opencl: add SOFTPLUS op support (llama/18726)

shaofeiqi
authored and
ggerganov
committed
sync : llma.cpp
ggerganov
committed

Commits on Jan 13, 2026

Commits on Jan 30, 2026

HIP: add fattn-mma-f16 for RDNA4 (llama/18481)

authored and
ggerganov
committed
ggml-metal: do not copy headers for embedded, use current binary dir for embedded (llama/18705)

DaAwesomeP
authored and
ggerganov
committed
vulkan: work around Intel fp16 bug in mmq (llama/18814)

0cc4m
authored and
ggerganov
committed
CUDA : fix typo in clang pragma comment [no ci] (llama/18830)

danbev
authored and
ggerganov
committed
vulkan: Check maxStorageBufferRange in supports_op (llama/18709)

jeffbolznv
authored and
ggerganov
committed
CUDA: Factor out and re-use `block_reduce` function (llama/18785)

authored and
ggerganov
committed
sync : llama.cpp
ggerganov
committed
hexagon: support for OP_CPY, host buffers now optional (llama/18822)

max-krasnyansky
authored and
ggerganov
committed
sync : llama.cpp
ggerganov
committed
ggml-cpu: optimize ggml_vec_dot_bf16 for Power9 (llama/18837)

shalinib-ibm
authored and
ggerganov
committed
CUDA: fix allignment on register spill for FA (llama/18815)

JohannesGaessler
authored and
ggerganov
committed
cuda : print less debug logs when disabling cuda graphs (llama/18868)
ggerganov
committed
OpenCL: add SOLVE_TRI op support (llama/18846)

shaofeiqi
authored and
ggerganov
committed
CANN: support gated linear attn (llama/18653)

authored and
ggerganov
committed
CANN: fix an issue where get_env was not fully renamed (llama/18796)

noemotiovon
authored and
ggerganov
committed
CANN: Remove unused `ggml_cann_get_device` function (llama/18625)

rauletorresc
authored and
ggerganov
committed
ggml-blas: hide warnings from included BLAS headers (llama/18818)

DaAwesomeP
authored and
ggerganov
committed
ggml : extend ggml_pool_1d + metal (llama/16429)

ThoreKoritzius
and
ggerganov
committed
sync : llama.cpp
ggerganov
committed
ggml webgpu: support for backend sampling (llama/18880)

reeselevine
authored and
ggerganov
committed
sync : llama.cpp
ggerganov
committed
opencl: fix q6_K mv for m=1 (llama/18893)

lhez
authored and
ggerganov
committed
ggml : add ggml_build_forward_select (llama/18550)
ggerganov
committed
metal : enable FA for MLA heads (llama/18950)
ggerganov
committed
ggml : cleanup path_str() (llama/18928)

angt
authored and
ggerganov
committed
CUDA: Replace init_offsets kernel with iterators in cub-based argsort (llama/18930)

ORippler
authored and
ggerganov
committed
CUDA: Fix builds for older CCCL versions by ifdefing strided_iterator (llama/18964)

ORippler
authored and
ggerganov
committed
vulkan: Use mul_mat_vec_id for small values of n (llama/18918)

jeffbolznv
authored and
ggerganov
committed
Revert "vulkan: force full subgroups for flash attention to fix intel subgroup crash (#17356)" (llama/18831)

rillomas
authored and
ggerganov
committed
vulkan: support flash attention GQA/split_k with small batches (llama/18938)

jeffbolznv
authored and
ggerganov
committed
vulkan: Remove transfer_ctx, do everything in compute_ctx. (llama/18945)

jeffbolznv
authored and
ggerganov
committed
ggml-zdnn : mark zDNN buffers as non-host (llama/18967)

AlekseiNikiforovIBM
authored and
ggerganov
committed
opencl: add TRI op support (llama/18979)

shaofeiqi
authored and
ggerganov
committed
CUDA: add gqa_ratio 4 for GLM 4.7 flash (llama/18953)

am17an
authored and
ggerganov
committed
opencl: enable the general fp mm for non-cont input and as a fallback for specialized kqv kernel for adreno (llama/18970)

lhez
authored and
ggerganov
committed
CUDA: fix alignment check for FA (llama/19023)

JohannesGaessler
authored and
ggerganov
committed
mla : make the V tensor a view of K (llama/18986)

ggerganov
and
JohannesGaessler
committed
ggml-cpu: aarm64: q5_K repack gemm and gemv (and generic) implementations (i8mm) (llama/18860)

Alcpz
authored and
ggerganov
committed
use malloc to support both iGPU and dGPU in same time (llama/18992)

authored and
ggerganov
committed
ggml-hexagon: flash-attn opt (llama/19025)

chraac
authored and
ggerganov
committed
ggml-cuda: enable cuda-graphs for `n-cpu-moe` (llama/18934)

am17an
authored and
ggerganov
committed
CUDA: re-use MLA K data for V in MMA FA (llama/19057)

JohannesGaessler
authored and
ggerganov
committed
kv-cache : support V-less cache (llama/19067)
ggerganov
committed
ggml-cpu: Use tiled FA for prompt-processing (llama/19012)

am17an
authored and
ggerganov
committed
metal : fix recommendedMaxWorkingSetSize availability on legacy iOS/macOS (llama/19088)

authored and
ggerganov
committed
CUDA: faster FA for GQA > 1 but not power of 2 (llama/19092)

JohannesGaessler
authored and
ggerganov
committed
CUDA: fix padding of GQA to power of 2 in FA (llama/19115)

JohannesGaessler
authored and
ggerganov
committed
sync : llama.cpp
ggerganov
committed
opencl: add flattened q6_K mv (llama/19054)

lhez
authored and
ggerganov
committed
ggml-cpu: Enable FP16 MMA kernels on PPC (llama/19060)

shalinib-ibm
authored and
ggerganov
committed
Reduce CPU-side stalls due to the CUDA command buffer being full (llama/19042)

gaugarg-nv
authored and
ggerganov
committed
ggml-cpu: aarm64: q6_K repack gemm and gemv (and generic) implementations (i8mm) #18860 (llama/18888)

Alcpz
and
ggerganov
committed
CUDA: tune GLM 4.7 Flash FA kernel selection logic (llama/19097)

JohannesGaessler
authored and
ggerganov
committed
ggml-zendnn : update ZenDNN git tag to main branch (llama/19133)

z-vishal
authored and
ggerganov
committed
ggml webgpu: Split shared state (webgpu_context) into global state and per-thread state (llama/18976)

authored and
ggerganov
committed
CUDA: tune GLM 4.7 Flash FA kernel selection logic (DGX Spark) (llama/19142)
ggerganov
committed
cuda : fix "V is K view" check for non-unified KV cache (llama/19145)
ggerganov
committed
ggml-cpu: arm64: Q4_K scale unroll and vectorization (llama/19108)

Alcpz
authored and
ggerganov
committed
ggml: new backend for Virglrenderer API Remoting acceleration (v2) (llama/18718)

kpouget
authored and
ggerganov
committed
vulkan: handle device dedup on MacOS + Vega II Duo cards (llama/19058)

okuvshynov
authored and
ggerganov
committed
ggml-sycl: remove unused syclcompat header (llama/19140)

PatKamin
authored and
ggerganov
committed
Vulkan Flash Attention Coopmat1 Refactor (llama/19075)

0cc4m
authored and
ggerganov
committed
sycl: fix norm kernels: l2_norm, group_norm, rms_norm by remove assert to support more cases (llama/19154)

authored and
ggerganov
committed
CUDA: refactor topk-moe to enable more models (GLM 4.7, Nemotron etc.) (llama/19126)

am17an
authored and
ggerganov
committed
ggml-zendnn : resolve ZenDNN backend cross-module symbol dependency (llama/19159)

z-vishal
authored and
ggerganov
committed
HIP: add mmf for CDNA (llama/18896)

authored and
ggerganov
committed
cuda : fix nkvo, offload and cuda graph node properties matching (llama/19165)
ggerganov
committed
hexagon: enable offloading to Hexagon on Windows on Snapdragon (llama/19150)

authored and
ggerganov
committed
ggml-webgpu: improve flastAttention performance by software pipelining (llama/19151)

authored and
ggerganov
committed
sycl: implement GGML_OP_TRI (llama/19089)

RachelMantel
authored and
ggerganov
committed
sycl: implement GGML_UNARY_OP_SOFTPLUS (llama/19114)

s8322
authored and
ggerganov
committed
add tensor type checking as part of cuda graph properties (llama/19186)

bssrdf
authored and
ggerganov
committed
sync : llama.cpp
ggerganov
committed
cuda : fix compile warnings (whisper/0)
ggerganov
committed
sync : whisper.cpp
ggerganov
committed
cmake : remove unused file (#1419)
ggerganov
authored

Commits on Feb 7, 2026

ggml : bump version to 0.9.6 (#1423)
ggerganov
authored
ci : remove "Release" word from the title of the release
ggerganov
committed
tests : add GQA=20 FA test (llama/19095)
ggerganov
committed
Correctly fetch q8_1 quantize pipeline in test as needed by 8a3519b (llama/19194)

sredman
authored and
ggerganov
committed
opencl: add optimized q8_0 mm kernel for adreno (llama/18871)

authored and
ggerganov
committed
ggml-hexagon: flash-attention and reduce-sum optimizations (llama/19141)

authored and
ggerganov
committed
Bump cmake max version (needed for Windows on Snapdragon builds) (llama/19188)

max-krasnyansky
authored and
ggerganov
committed
Remove pipeline cache mutexes (llama/19195)

nikhilJain17
authored and
ggerganov
committed
docs : Minor cleanups (llama/19252)

ckastner
authored and
ggerganov
committed
ggml-backend: fix async set/get fallback sync (llama/19179)

JohannesGaessler
authored and
ggerganov
committed
metal : support virtual devices (llama/18919)
ggerganov
committed
sycl: implement GGML_OP_TOP_K (llama/19242)

tdevelope
authored and
ggerganov
committed
Remove support for Nvidia & AMD GPU, because the oneAPI plugin for Nvidia & AMD GPU is unavailable: download/installation channels are out of work. (llama/19246)

arthw
authored and
ggerganov
committed
ggml-cpu: FA split across kv for faster TG (llama/19209)

am17an
authored and
ggerganov
committed
opencl: refactor some ops, concat, repeat, tanh and scale (llama/19226)

lhez
authored and
ggerganov
committed
cuda : revert CUDA_SCALE_LAUNCH_QUEUES override until investigated (llama/19227)

gaugarg-nv
authored and
ggerganov
committed
ggml: added cleanups in ggml_quantize_free (llama/19278)

noctrex
authored and
ggerganov
committed
CUDA: Fix loop unrolling for BW in mul_mat_q_stream_k_fixup (llama/19053)

ORippler
authored and
ggerganov
committed
metal : minor cleanup (llama/19251)
ggerganov
committed
CUDA: use mmvq for mul-mat-id for small batch sizes (llama/18958)

am17an
authored and
ggerganov
committed
vulkan: disable coopmat1 fa on Nvidia Turing (llama/19290)

0cc4m
authored and
ggerganov
committed
metal : add solve_tri (llama/19302)
ggerganov
committed
ggml-cpu: use LUT for converting e8->f32 scales on x86 (llama/19288)

am17an
authored and
ggerganov
committed
ggml-virtgpu: make the code thread safe (llama/19204)

kpouget
authored and
ggerganov
committed
tests : add non-cont, inplace rope tests (llama/19296)

ggerganov
and
jeffbolznv
committed
metal : add missing includes (llama/19348)

will-lms
authored and
ggerganov
committed
vulkan: fix non-contig rope (llama/19299)

jeffbolznv
authored and
ggerganov
committed
vulkan: Set k_load_shmem to false when K is too large (llama/19301)

jeffbolznv
authored and
ggerganov
committed
vulkan: fix GPU deduplication logic. (llama/19222)

okuvshynov
authored and
ggerganov
committed
metal : add diag (llama/19330)
ggerganov
committed
vulkan: Preprocess FA mask to detect all-neg-inf and all-zero. (llama/19281)

jeffbolznv
authored and
ggerganov
committed
metal : adaptive CPU/GPU interleave based on number of nodes (llama/19369)
ggerganov
committed
cuda : cuda graphs now compare all node params (llama/19383)
ggerganov
committed
metal : skip loading all-zero mask (llama/19337)
ggerganov
committed
vulkan: make FA mask/softcap enables spec constants (llama/19309)

jeffbolznv
authored and
ggerganov
committed
vulkan: For coopmat2 FA, use fp16 accumulators for the final result (llama/19376)

jeffbolznv
authored and
ggerganov
committed
tests: reduce number of FA test permutations (llama/19381)

jeffbolznv
authored and
ggerganov
committed
sycl: add F16 support for GGML_OP_CEIL (llama/19306)

NechamaKrashinski
authored and
ggerganov
committed
sync : llama.cpp
ggerganov
committed
ggml-webgpu: JIT compile binary operators and handle binding overlaps (llama/19310)

abhijitramesh
authored and
ggerganov
committed
metal : fix event synchronization in cpy_tensor_async (llama/19402)
ggerganov
committed
metal : consolidate bin kernels (llama/19390)
ggerganov
committed
sync : llama.cpp
ggerganov
committed