sync : ggml#2608

Merged

ggerganov merged 65 commits intomasterfrom

Dec 8, 2024

Commits on Dec 5, 2024

ggml-opt: fix data corruption (ggml/1022)

JohannesGaessler
authored and
ggerganov
committed
Do not include arm_neon.h when compiling CUDA code (ggml/1028)

frankier
authored and
ggerganov
committed
metal : add `GGML_OP_CONV_TRANSPOSE_1D` kernels (ggml/1026)

PABannier
authored and
ggerganov
committed
feat: add `GGML_UNARY_OP_ARGMAX` Metal kernel (ggml/1019)

authored and
ggerganov
committed
CUDA: remove unnecessary warp reduce in FA (ggml/1032)

authored and
ggerganov
committed
add cmake rvv support (llama/10411)

lhpqaq
authored and
ggerganov
committed
vulkan: further optimize mul_mat_vec using larger loads (llama/10387)

jeffbolznv
authored and
ggerganov
committed
vulkan: copy iq4_nl LUT into shared memory (llama/10409)

jeffbolznv
authored and
ggerganov
committed
vulkan: predicate max operation in soft_max shaders/soft_max (llama/10437)

jeffbolznv
authored and
ggerganov
committed
cuda : optimize argmax (llama/10441)

authored and
ggerganov
committed
CANN: Support Ascend310P to accelerate F32 and F16 Model (llama/10216)

leo-pony
authored and
ggerganov
committed
ggml : do not use ARM features not included in the build (llama/10457)

slaren
authored and
ggerganov
committed
metal : minor code formatting
ggerganov
committed
ggml : add support for dynamic loading of backends (llama/10469)

slaren
and
ggerganov
committed
llama : accept a list of devices to use to offload a model (llama/10497)

slaren
authored and
ggerganov
committed
metal : enable mat-vec kernels for bs <= 4 (llama/10491)
ggerganov
committed
vulkan: Fix a vulkan-shaders-gen arugment parsing error (llama/10484)

sparkleholic
authored and
ggerganov
committed
CANN: RoPE and CANCAT operator optimization (llama/10488)

authored and
ggerganov
committed
CANN: Improve the Inferencing Performance for Ascend NPU Device (llama/10454)

authored and
ggerganov
committed
ggml-cpu: cmake add arm64 cpu feature check for macos (llama/10487)

chaxu01
authored and
ggerganov
committed
cmake : enable warnings in llama (llama/10474)
ggerganov
committed
vulkan: fix group_norm (llama/10496)

jeffbolznv
authored and
ggerganov
committed
mtgpu: Add MUSA_DOCKER_ARCH in Dockerfiles && update cmake and make (llama/10516)

yeahdongcn
authored and
ggerganov
committed
vulkan: optimize Q2_K and Q3_K mul_mat_vec (llama/10459)

jeffbolznv
authored and
ggerganov
committed
vulkan: skip integer div/mod in get_offsets for batch_idx==0 (llama/10506)

jeffbolznv
authored and
ggerganov
committed
vulkan: further optimize q5_k mul_mat_vec (llama/10479)

jeffbolznv
authored and
ggerganov
committed
vulkan: Handle GPUs with less shared memory (llama/10468)

jeffbolznv
authored and
ggerganov
committed
vulkan: define all quant data structures in types.comp (llama/10440)

jeffbolznv
authored and
ggerganov
committed
metal : fix group_norm support condition (llama/0)
ggerganov
committed
Add some minimal optimizations for CDNA (llama/10498)

IMbackK
authored and
ggerganov
committed
CANN: ROPE operator optimization (llama/10540)

authored and
ggerganov
committed
CANN: Fix SOC_TYPE compile bug (llama/10519)

leo-pony
authored and
ggerganov
committed
kompute : improve backend to pass test_backend_ops (llama/10542)

slp
authored and
ggerganov
committed
ggml-cpu: support IQ4_NL_4_4 by runtime repack (llama/10541)

FanShupei
authored and
ggerganov
committed
cmake : fix ARM feature detection (llama/10543)
ggerganov
committed
ggml : fix row condition for i8mm kernels (llama/10561)
ggerganov
committed
ggml : remove redundant copyright notice + update authors
ggerganov
committed
vulkan: get the first command buffer submitted sooner (llama/10499)

jeffbolznv
authored and
ggerganov
committed
CANN: RoPE operator optimization (llama/10563)

authored and
ggerganov
committed
sycl : Reroute permuted mul_mats through oneMKL (llama/10408)

Alcpz
authored and
ggerganov
committed
sycl : offload of get_rows set to 0 (llama/10432)

Alcpz
authored and
ggerganov
committed
ggml-cpu: fix typo in gemv/gemm iq4_nl_4_4 (llama/10580)

FanShupei
authored and
ggerganov
committed
ggml : fix I8MM Q4_1 scaling factor conversion (llama/10562)
ggerganov
committed
vulkan: Dynamic subgroup size support for Q6_K mat_vec (llama/10536)

netrunnereve
authored and
ggerganov
committed
ggml-cpu: replace AArch64 NEON assembly with intrinsics in ggml_gemv_q4_0_4x4_q8_0() (llama/10567)

angt
authored and
ggerganov
committed
SYCL: Fix and switch to GGML_LOG system instead of fprintf (llama/10579)

qnixsynapse
authored and
ggerganov
committed
metal : small-batch mat-mul kernels (llama/10581)
ggerganov
committed
ggml : move AMX to the CPU backend (llama/10570)

slaren
authored and
ggerganov
committed
common : fix compile warning
ggerganov
committed
files : remove make artifacts
ggerganov
committed
ggml : add `GGML_PAD_REFLECT_1D` operation (ggml/1034)

PABannier
authored and
ggerganov
committed
ggml: add `GGML_SET` Metal kernel + i32 CPU kernel (ggml/1037)

PABannier
authored and
ggerganov
committed
vulkan: optimize and reenable split_k (llama/10637)

jeffbolznv
authored and
ggerganov
committed
Avoid using __fp16 on ARM with old nvcc (llama/10616)

frankier
authored and
ggerganov
committed
SYCL : Move to compile time oneMKL interface backend selection for NVIDIA backend (llama/10584)

s-Nick
authored and
ggerganov
committed
vulkan: Implement "fast divide" (mul+shift) for unary ops like copy (llama/10642)

jeffbolznv
authored and
ggerganov
committed
ggml-cpu : fix HWCAP2_I8MM value (llama/10646)

slaren
authored and
ggerganov
committed
ggml : add predefined list of CPU backend variants to build (llama/10626)

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

Commits on Dec 6, 2024

make : shim cmake
ggerganov
committed