Code

[pull] master from ggerganov:master#121

Code

Closed

pull[bot] wants to merge 55 commits intoteleprint-me:masterteleprint-me/llama.cpp:masterfrom ggml-org:masterggml-org/llama.cpp:masterCopy head branch name to clipboard

+8,434-4,312

Commits

Commits on Jun 24, 2024

• gguf-py, convert-hf : model conversion support for T5 and FLAN-T5 model variants (#5763)

  Show description for de0d6a6
  

  fairydreaming

  and

  sszymczy

  authored
  de0d6a6

  Copy full SHA for de0d6a6

• embedding : more cli arguments (#7458)

  Show description for 646ef4a
  

  YannFollet

  and

  mofosyne

  authored
  646ef4a

  Copy full SHA for 646ef4a

• disable publishing the full-rocm docker image (#8083)

  slaren

  authored
  8cb508d

  Copy full SHA for 8cb508d

• Option to split during conversion (#6942)

  Show description for 52fc870
  

  3 peopleauthored
  52fc870

  Copy full SHA for 52fc870

• CUDA: optimize MMQ int8 tensor core performance (#8062)

  Show description for 9a590c8
  JohannesGaessler

  authored
  9a590c8

  Copy full SHA for 9a590c8

• gguf-py : fix tensor groups for encoder-decoder models in gguf-dump.py (#8090)

  Show description for d62e4aa
  

  3 peopleauthored
  d62e4aa

  Copy full SHA for d62e4aa

• CUDA: use MMQ instead of cuBLAS by default (#8075)

  JohannesGaessler

  authored
  a818f30

  Copy full SHA for a818f30

• CUDA: fix MMQ writeback for int8 tensor cores (#8100)

  JohannesGaessler

  authored
  3b099bc

  Copy full SHA for 3b099bc

• CUDA: fix matrix multiplication algorithm choice (#8102)

  JohannesGaessler

  authored
  2df373a

  Copy full SHA for 2df373a

Commits on Jun 25, 2024

• [SYCL] Re-enabled mul_mat_batched_sycl (#8095)

  airMeng

  authored
  083bacc

  Copy full SHA for 083bacc

• Update control vector help (#8104)

  HatsuneMikuUwU33

  authored
  f702a90

  Copy full SHA for f702a90

• SimpleChat v3.1: Boolean chat request options in Settings UI, cache_prompt (#7950)

  Show description for 3791ad2
  hanishkvc

  authored
  3791ad2

  Copy full SHA for 3791ad2

• Add chat template support for llama-cli (#8068)

  Show description for 48e6b92
  

  ngxson

  and

  ggerganov

  authored
  48e6b92

  Copy full SHA for 48e6b92

• cvector: better prompt handling, add "mean vector" method (#8069)

  Show description for 49c03c7
  ngxson

  authored
  49c03c7

  Copy full SHA for 49c03c7

• Gguf dump start data offset via --data-offset and some extra refactor (#8054)

  Show description for c8ad359
  mofosyne

  authored
  c8ad359

  Copy full SHA for c8ad359

• Add healthchecks to llama-server containers (#8081)

  Show description for 925c309
  codearranger

  authored
  925c309

  Copy full SHA for 925c309

• disable docker CI on pull requests (#8110)

  slaren

  authored
  dd047b4

  Copy full SHA for dd047b4

• json: support integer minimum, maximum, exclusiveMinimum, exclusiveMaximum (#7797)

  Show description for 84631fe
  ochafik

  authored
  84631fe

  Copy full SHA for 84631fe

• llama : return nullptr from llama_grammar_init (#8093)

  Show description for e6bf007
  

  danbev

  and

  HanClinto

  authored
  e6bf007

  Copy full SHA for e6bf007

• llama : implement Unigram tokenizer needed by T5 and FLAN-T5 model families (#5763)

  Show description for 6fcbf68
  

  fairydreaming

  and

  sszymczy

  authored
  6fcbf68

  Copy full SHA for 6fcbf68

• fixes #7999 (adds control vectors to all build_XXX() functions in llama.cpp [needs testing] (#8060)

  Show description for 163d50a
  jukofyork

  authored
  163d50a

  Copy full SHA for 163d50a

Commits on Jun 26, 2024

• json: fix additionalProperties, allow space after enum/const (#7840)

  Show description for 6777c54
  ochafik

  authored
  6777c54

  Copy full SHA for 6777c54

• json: better support for "type" unions (e.g. nullable arrays w/ typed items) (#7863)

  Show description for 9b2f16f
  ochafik

  authored
  9b2f16f

  Copy full SHA for 9b2f16f

• llama : extend llm_build_ffn() to support _scale tensors (#8103)

  Eddie-Wang1120

  authored
  494165f

  Copy full SHA for 494165f

• CUDA: fix misaligned shared memory read (#8123)

  JohannesGaessler

  authored
  c8771ab

  Copy full SHA for c8771ab

• Clarify default MMQ for CUDA and LLAMA_CUDA_FORCE_MMQ flag (#8115)

  Show description for 8854044
  

  isaac-mcfadyen

  and

  JohannesGaessler

  authored
  8854044

  Copy full SHA for 8854044

• llama : reorganize source code + improve CMake (#8006)

  Show description for f3f6542
  

  ggerganov

  and

  slaren

  authored
  f3f6542

  Copy full SHA for f3f6542

• readme : update API notes

  ggerganov

  authored
  a95631e

  Copy full SHA for a95631e

• devops : remove clblast + LLAMA_CUDA -> GGML_CUDA (#8139)

  Show description for 0e814df
  ggerganov

  authored
  0e814df

  Copy full SHA for 0e814df

• authors : regen

  ggerganov

  committed
  4713bf3

  Copy full SHA for 4713bf3

• sync : ggml

  ggerganov

  committed
  f2d48ff

  Copy full SHA for f2d48ff

• make : fix missing -O3 (#8143)

  slaren

  authored
  c7ab7b6

  Copy full SHA for c7ab7b6

• ggml : add GGML_CUDA_USE_GRAPHS option, restore GGML_CUDA_FORCE_CUBLAS (cmake) (#8140)

  slaren

  authored
  31ec399

  Copy full SHA for 31ec399

• ci : publish new docker images only when the files change (#8142)

  slaren

  authored
  ae5d0f4

  Copy full SHA for ae5d0f4

• scripts : fix filename sync

  ggerganov

  committed
  c70d117

  Copy full SHA for c70d117

• clip : suppress unused variable warnings (#8105)

  Show description for 9b31a40
  danbev

  authored
  9b31a40

  Copy full SHA for 9b31a40

Commits on Jun 27, 2024

• Fix llama-android.cpp for error - "common/common.h not found" (#8145)

  Show description for ac14662
  criminact

  authored
  ac14662

  Copy full SHA for ac14662

• llama : fix CodeLlama FIM token checks (#8144)

  Show description for 911e35b
  CISC

  authored
  911e35b

  Copy full SHA for 911e35b

• Added support for Viking pre-tokenizer (#8135)

  Show description for f675b20
  

  kustaaya

  and

  kustaaya

  authored
  f675b20

  Copy full SHA for f675b20

• CUDA: fix MMQ stream-k for --split-mode row (#8167)

  JohannesGaessler

  authored
  85a267d

  Copy full SHA for 85a267d

• Add Qwen2MoE 57B-A14B model identifier (#8158)

  Show description for 6030c61
  CISC

  authored
  6030c61

  Copy full SHA for 6030c61

• Delete examples/llama.android/llama/CMakeLists.txt (#8165)

  Show description for 3879526
  criminact

  authored
  3879526

  Copy full SHA for 3879526

• Control vector loading fixes (#8137)

  Show description for 97877eb
  jukofyork

  authored
  97877eb

  Copy full SHA for 97877eb

• flake.lock: Update (#8071)

  Show description for ab36791
  

  3 peopleauthored
  ab36791

  Copy full SHA for ab36791

• Add chatml fallback for cpp llama_chat_apply_template (#8160)

  Show description for 16791b8
  ngxson

  authored
  16791b8

  Copy full SHA for 16791b8

• cmake : fix deprecated option names not working (#8171)

  Show description for 8172ee9
  slaren

  authored
  8172ee9

  Copy full SHA for 8172ee9

• CI: fix release build (Ubuntu+Mac) (#8170)

  Show description for 558f44b
  loonerin

  authored
  558f44b

  Copy full SHA for 558f44b

• json: update grammars/README w/ examples & note about additionalProperties (#8132)

  Show description for cb0b06a
  ochafik

  authored
  cb0b06a

  Copy full SHA for cb0b06a

Commits on Jun 28, 2024

• Add missing items in makefile (#8177)

  ngxson

  authored
  a27aa50

  Copy full SHA for a27aa50

• llama: Add support for Gemma2ForCausalLM (#8156)

  Show description for e57dc62
  

  3 peopleauthored
  e57dc62

  Copy full SHA for e57dc62

• json: restore default additionalProperties to false, fix some pattern escapes (#8180)

  Show description for 139cc62
  ochafik

  authored
  139cc62

  Copy full SHA for 139cc62

• cmake : allow user to override default options (#8178)

  slaren

  authored
  b851b3f

  Copy full SHA for b851b3f

• Add SPM infill support (#8016)

  Show description for 38373cf
  CISC

  authored
  38373cf

  Copy full SHA for 38373cf

• Add MiniCPM, Deepseek V2 chat template + clean up llama_chat_apply_template_internal (#8172)

  Show description for 26a39bb
  ngxson

  authored
  26a39bb

  Copy full SHA for 26a39bb

• json: attempt to skip slow tests when running under emulator (#8189)

  ochafik

  authored
  8748d8a

  Copy full SHA for 8748d8a