Code

sync : llama.cpp#1039

Code

Merged

ggerganov merged 52 commits intomasterggml-org/ggml:masterfrom syncggml-org/ggml:syncCopy head branch name to clipboardDec 3, 2024

+5,253-2,375

Commits

Commits on Dec 3, 2024

• add cmake rvv support (llama/10411)

  

  lhpqaq

  authored and

  ggerganov

  committed
  2dde9cc

  Copy full SHA for 2dde9cc

• vulkan: further optimize mul_mat_vec using larger loads (llama/10387)

  Show description for 01d961f
  

  jeffbolznv

  authored and

  ggerganov

  committed
  01d961f

  Copy full SHA for 01d961f

• vulkan: copy iq4_nl LUT into shared memory (llama/10409)

  

  jeffbolznv

  authored and

  ggerganov

  committed
  a37e9dc

  Copy full SHA for a37e9dc

• vulkan: predicate max operation in soft_max shaders/soft_max (llama/10437)

  Show description for 826679f
  

  jeffbolznv

  authored and

  ggerganov

  committed
  826679f

  Copy full SHA for 826679f

• cuda : optimize argmax (llama/10441)

  Show description for 1dada6d
  

  2 peopleauthored and

  ggerganov

  committed
  1dada6d

  Copy full SHA for 1dada6d

• CANN: Support Ascend310P to accelerate F32 and F16 Model (llama/10216)

  Show description for 1a9b04c
  

  leo-pony

  authored and

  ggerganov

  committed
  1a9b04c

  Copy full SHA for 1a9b04c

• ggml : do not use ARM features not included in the build (llama/10457)

  

  slaren

  authored and

  ggerganov

  committed
  0e4f353

  Copy full SHA for 0e4f353

• metal : minor code formatting

  ggerganov

  committed
  27b3394

  Copy full SHA for 27b3394

• tests : fix compile warning

  ggerganov

  committed
  f21071b

  Copy full SHA for f21071b

• ggml : add support for dynamic loading of backends (llama/10469)

  Show description for 3f44610
  

  slaren

  and

  ggerganov

  committed
  3f44610

  Copy full SHA for 3f44610

• llama : accept a list of devices to use to offload a model (llama/10497)

  Show description for 0df7ac3
  

  slaren

  authored and

  ggerganov

  committed
  0df7ac3

  Copy full SHA for 0df7ac3

• metal : enable mat-vec kernels for bs <= 4 (llama/10491)

  ggerganov

  committed
  41b84d7

  Copy full SHA for 41b84d7

• vulkan: Fix a vulkan-shaders-gen arugment parsing error (llama/10484)

  Show description for 93a0782
  

  sparkleholic

  authored and

  ggerganov

  committed
  93a0782

  Copy full SHA for 93a0782

• CANN: RoPE and CANCAT operator optimization (llama/10488)

  Show description for d3ccd11
  

  2 peopleauthored and

  ggerganov

  committed
  d3ccd11

  Copy full SHA for d3ccd11

• CANN: Improve the Inferencing Performance for Ascend NPU Device (llama/10454)

  Show description for 8c59e6b
  

  3 peopleauthored and

  ggerganov

  committed
  8c59e6b

  Copy full SHA for 8c59e6b

• ggml-cpu: cmake add arm64 cpu feature check for macos (llama/10487)

  Show description for d1976ec
  

  chaxu01

  authored and

  ggerganov

  committed
  d1976ec

  Copy full SHA for d1976ec

• cmake : enable warnings in llama (llama/10474)

  Show description for a1d5c10
  ggerganov

  committed
  a1d5c10

  Copy full SHA for a1d5c10

• vulkan: fix group_norm (llama/10496)

  Show description for 965432c
  

  jeffbolznv

  authored and

  ggerganov

  committed
  965432c

  Copy full SHA for 965432c

• mtgpu: Add MUSA_DOCKER_ARCH in Dockerfiles && update cmake and make (llama/10516)

  Show description for 76301a7
  

  yeahdongcn

  authored and

  ggerganov

  committed
  76301a7

  Copy full SHA for 76301a7

• vulkan: optimize Q2_K and Q3_K mul_mat_vec (llama/10459)

  

  jeffbolznv

  authored and

  ggerganov

  committed
  2480d18

  Copy full SHA for 2480d18

• vulkan: skip integer div/mod in get_offsets for batch_idx==0 (llama/10506)

  

  jeffbolznv

  authored and

  ggerganov

  committed
  5ff537f

  Copy full SHA for 5ff537f

• vulkan: further optimize q5_k mul_mat_vec (llama/10479)

  

  jeffbolznv

  authored and

  ggerganov

  committed
  02dfbe1

  Copy full SHA for 02dfbe1

• vulkan: Handle GPUs with less shared memory (llama/10468)

  Show description for 40699b9
  

  jeffbolznv

  authored and

  ggerganov

  committed
  40699b9

  Copy full SHA for 40699b9

• vulkan: define all quant data structures in types.comp (llama/10440)

  

  jeffbolznv

  authored and

  ggerganov

  committed
  b314f1f

  Copy full SHA for b314f1f

• metal : fix group_norm support condition (llama/0)

  ggerganov

  committed
  24bc6e0

  Copy full SHA for 24bc6e0

• Add some minimal optimizations for CDNA (llama/10498)

  Show description for 3602400
  

  IMbackK

  authored and

  ggerganov

  committed
  3602400

  Copy full SHA for 3602400

• CANN: ROPE operator optimization (llama/10540)

  Show description for ac2fc33
  

  2 peopleauthored and

  ggerganov

  committed
  ac2fc33

  Copy full SHA for ac2fc33

• CANN: Fix SOC_TYPE compile bug (llama/10519)

  Show description for abcb176
  

  leo-pony

  authored and

  ggerganov

  committed
  abcb176

  Copy full SHA for abcb176

• kompute : improve backend to pass test_backend_ops (llama/10542)

  Show description for a6b17f2
  

  slp

  authored and

  ggerganov

  committed
  a6b17f2

  Copy full SHA for a6b17f2

• ggml-cpu: support IQ4_NL_4_4 by runtime repack (llama/10541)

  Show description for f7ffcd2
  

  FanShupei

  authored and

  ggerganov

  committed
  f7ffcd2

  Copy full SHA for f7ffcd2

• cmake : fix ARM feature detection (llama/10543)

  Show description for 4def5cb
  ggerganov

  committed
  4def5cb

  Copy full SHA for 4def5cb

• ggml : fix row condition for i8mm kernels (llama/10561)

  Show description for 5848c4c
  ggerganov

  committed
  5848c4c

  Copy full SHA for 5848c4c

• ggml : remove redundant copyright notice + update authors

  ggerganov

  committed
  816485a

  Copy full SHA for 816485a

• vulkan: get the first command buffer submitted sooner (llama/10499)

  Show description for 6eff9fb
  

  jeffbolznv

  authored and

  ggerganov

  committed
  6eff9fb

  Copy full SHA for 6eff9fb

• CANN: RoPE operator optimization (llama/10563)

  Show description for 5a8c20b
  

  2 peopleauthored and

  ggerganov

  committed
  5a8c20b

  Copy full SHA for 5a8c20b

• sycl : Reroute permuted mul_mats through oneMKL (llama/10408)

  Show description for c17ffb7
  

  Alcpz

  authored and

  ggerganov

  committed
  c17ffb7

  Copy full SHA for c17ffb7

• sycl : offload of get_rows set to 0 (llama/10432)

  

  Alcpz

  authored and

  ggerganov

  committed
  f6e4e07

  Copy full SHA for f6e4e07

• ggml-cpu: fix typo in gemv/gemm iq4_nl_4_4 (llama/10580)

  

  FanShupei

  authored and

  ggerganov

  committed
  1602102

  Copy full SHA for 1602102

• ggml : fix I8MM Q4_1 scaling factor conversion (llama/10562)

  Show description for d2896d6
  ggerganov

  committed
  d2896d6

  Copy full SHA for d2896d6

• vulkan: Dynamic subgroup size support for Q6_K mat_vec (llama/10536)

  Show description for cf66ee0
  

  netrunnereve

  authored and

  ggerganov

  committed
  cf66ee0

  Copy full SHA for cf66ee0

• ggml-cpu: replace AArch64 NEON assembly with intrinsics in ggml_gemv_q4_0_4x4_q8_0() (llama/10567)

  Show description for a5c2af2
  

  angt

  authored and

  ggerganov

  committed
  a5c2af2

  Copy full SHA for a5c2af2

• SYCL: Fix and switch to GGML_LOG system instead of fprintf (llama/10579)

  Show description for 1bc5829
  

  qnixsynapse

  authored and

  ggerganov

  committed
  1bc5829

  Copy full SHA for 1bc5829

• metal : small-batch mat-mul kernels (llama/10581)

  Show description for c908240
  ggerganov

  committed
  c908240

  Copy full SHA for c908240

• scripts : remove amx from sync

  ggerganov

  committed
  8ce7ebf

  Copy full SHA for 8ce7ebf

• ggml : move AMX to the CPU backend (llama/10570)

  Show description for a842931
  

  slaren

  authored and

  ggerganov

  committed
  a842931

  Copy full SHA for a842931

• sync : llama.cpp

  Show description for 6aa7be5
  ggerganov

  committed
  6aa7be5

  Copy full SHA for 6aa7be5

• authors : update

  ggerganov

  committed
  23caf78

  Copy full SHA for 23caf78

• common : fix compile warning

  Show description for 7756deb
  ggerganov

  committed
  7756deb

  Copy full SHA for 7756deb

• files : remove make artifacts

  ggerganov

  committed
  9666d68

  Copy full SHA for 9666d68

• ci : fix pip env

  Show description for 0c55ec2
  ggerganov

  committed
  0c55ec2

  Copy full SHA for 0c55ec2

• ci : remove opencl workflow

  ggerganov

  committed
  65f6241

  Copy full SHA for 65f6241

• ci : update requirements.txt

  Show description for d54f335
  ggerganov

  committed
  d54f335

  Copy full SHA for d54f335