Commit graph

  • ae33438729 examples : do not print meaningless defaults Cebtenzzre 2023-09-17 20:55:17 -04:00
  • 76988cdb9a llama : allow gguf rope keys to be overridden with defaults Cebtenzzre 2023-09-13 22:37:24 -04:00
  • 78c45b7975 llama : remove unused llama_hparams defaults Cebtenzzre 2023-09-13 22:36:42 -04:00
  • f80820f047
    Merge e283c20285 into 7ddf185537 staviq 2023-09-18 00:43:47 +00:00
  • 7ddf185537
    ci : switch cudatoolkit install on windows to networked (#3236) b1255 Erik Scholz 2023-09-18 02:21:47 +02:00
  • ee66942d7e
    CUDA: fix peer access logic (#3231) b1254 Johannes Gäßler 2023-09-17 23:35:20 +02:00
  • bfab8ee055
    ci : switch cudatoolkit install on windows to networked Green Sky 2023-09-17 23:16:48 +02:00
  • 1470286e4f
    Merge 784d14ed31 into fad56936d4 Georgi Gerganov 2023-09-17 20:50:14 +00:00
  • 784d14ed31
    llama : store non-RoPEd K cache (WIP) custom-attention-mask-no-roped-cache Georgi Gerganov 2023-09-17 23:12:28 +03:00
  • fad56936d4
    metal : add rope_f16 kernel + optimize cpy kernels Georgi Gerganov 2023-09-17 23:09:48 +03:00
  • 1fb033fd85
    ggml : ggml_rope now takes a vector with positions instead of n_past Georgi Gerganov 2023-09-17 21:12:51 +03:00
  • 1e53622fa7 CUDA: fix peer access logic JohannesGaessler 2023-09-17 20:16:35 +02:00
  • 5ce74ee461
    initialize opt->loss_after with zero xaedes 2023-09-17 19:42:19 +02:00
  • 3b4bab6a38
    llama : replace ggml_diag_mask_inf with ggml_add (custom -inf mask) Georgi Gerganov 2023-09-17 19:42:39 +03:00
  • e0eba91bea ggml : ggml_log_callback typedef Rickard Hallerbäck 2023-09-17 18:19:37 +02:00
  • 3b9d97484c
    remove unused options and equalize train-text-from-scratch with finetune xaedes 2023-09-17 18:01:16 +02:00
  • c993246bfd
    train-text-from-scratch: automatically allocate compute memory xaedes 2023-09-17 17:52:22 +02:00
  • f9b5d9b760
    train-text-from-scratch: automatically allocate input tensors xaedes 2023-09-17 17:41:53 +02:00
  • db38d2bce4
    train-text-from-scratch: automatically allocate opt context xaedes 2023-09-17 17:33:11 +02:00
  • 9e10fa977e
    train-text-from-scratch: automatically allocate model tensors, remove option '--mem-model N' xaedes 2023-09-17 17:08:36 +02:00
  • e283c20285 mv handlers back to .h for inlining staviq 2023-09-17 16:59:37 +02:00
  • c5df72e848
    tests : verify that RoPE is "additive" Georgi Gerganov 2023-09-17 17:54:14 +03:00
  • dd94ce4ec0
    remove now unnecessary llama API functions to get model params that where added by this PR xaedes 2023-09-17 16:49:52 +02:00
  • 934ad8d35d
    move some params from lora hparams into model hparams and load model params from gguf xaedes 2023-09-17 16:51:03 +02:00
  • b0ee563748
    assert correct base model tensor shapes xaedes 2023-09-17 16:43:12 +02:00
  • 78de0dff08 Makefile : cleanup ggml-metal recipe Rickard Hallerbäck 2023-09-17 16:42:24 +02:00
  • 5ed309810e
    align code xaedes 2023-09-17 16:41:25 +02:00
  • 1dbd6bc3d5
    remove n_rot hparam, as it must always be hparam.n_embd_head() xaedes 2023-09-17 16:40:40 +02:00
  • 92b39b47dd Merge branch 'master' of github.com:Ricardicus/llama.cpp Rickard Hallerbäck 2023-09-17 16:39:22 +02:00
  • d266e15c81 ggml : log level enum used by llama Rickard Hallerbäck 2023-09-17 16:38:46 +02:00
  • 111163e246
    CUDA: enable peer access between devices (#2470) b1253 Johannes Gäßler 2023-09-17 16:37:53 +02:00
  • 56a03faf5f
    deduplicate code into function xaedes 2023-09-17 16:37:21 +02:00
  • 10ad994447 fix cmake staviq 2023-09-17 16:32:11 +02:00
  • 5b4db36130 mv defs to log.cpp staviq 2023-09-17 16:27:53 +02:00
  • 02540fff64 CUDA: enable peer access between devices JohannesGaessler 2023-07-31 16:07:02 +02:00
  • 06e5c1902a formatting: break overly long LOG lines staviq 2023-09-17 14:45:25 +02:00
  • 34930bfdc2 updated lite Concedo 2023-09-17 20:43:04 +08:00
  • d1bb6fb349
    add train option "--sample-random-offsets" xaedes 2023-09-17 14:37:41 +02:00
  • 8b428c9bc8
    llama.cpp : show model size and BPW on load (#3223) b1252 slaren 2023-09-17 14:33:28 +02:00
  • bf2ad65836
    fix frand to return value in interval [0,1) xaedes 2023-09-17 14:28:58 +02:00
  • 578d8c8f5c
    CUDA: fix scratch malloced on non-main device (#3220) b1251 Johannes Gäßler 2023-09-17 14:16:22 +02:00
  • 151bfe9ee1
    assert that sample_count > 0, avoiding division by zero xaedes 2023-09-17 13:07:17 +02:00
  • ddf5ac257a
    use new/delete for train_state instead of malloc/free xaedes 2023-09-17 12:48:17 +02:00
  • 391dab70d7 CUDA: fix scratch malloced on non-main device JohannesGaessler 2023-09-16 21:48:29 +02:00
  • e0fcc9a725 fixed all issues with class.py Concedo 2023-09-17 15:23:35 +08:00
  • 1b6884336a Update examples/embedding/README.md yuiseki 2023-09-17 09:41:21 +09:00
  • a08e1a922d llama.cpp : show model size and BPW on load slaren 2023-09-17 01:07:23 +02:00
  • 28cdd8c3b0 formatting staviq 2023-09-16 22:36:04 +02:00
  • 8721785c52
    fix compile warnings xaedes 2023-09-16 22:28:23 +02:00
  • 83061fbdbe
    fix compile warnings xaedes 2023-09-16 22:19:46 +02:00
  • dd3e7634f0
    remove terminating '\0' from tokenization xaedes 2023-09-16 21:30:49 +02:00
  • 9db2664dd1
    fix saving and loading of training type xaedes 2023-09-16 21:21:04 +02:00
  • 1d09965179
    use die("msg") instead of replace GGML_ASSERT(!"msg") or throw std::runtime_error("msg") xaedes 2023-09-16 21:12:16 +02:00
  • 1d33ec5b1c
    fix condition in load_train_state_gguf xaedes 2023-09-16 21:10:16 +02:00
  • 9139fec7ff
    fix code formating of long function declarations xaedes 2023-09-16 20:38:23 +02:00
  • 8d82d4c8e6
    remove static from process_escape since we need it exposed in header xaedes 2023-09-16 20:37:56 +02:00
  • 9f630c6681 formatting staviq 2023-09-16 20:37:38 +02:00
  • 7930caf24c
    fix usage of llama_tokenize xaedes 2023-09-16 20:36:43 +02:00
  • d3e06d3e73
    Merge branch 'master' into finetune-lora xaedes 2023-09-16 20:31:58 +02:00
  • 571dc94da9
    increase train_samples by used_samples instead of number of batches xaedes 2023-09-16 20:23:05 +02:00
  • 48d3509190
    save and load head_count_kv in lora checkpoints xaedes 2023-09-16 20:20:23 +02:00
  • b541b4f0b1
    Enable BUILD_SHARED_LIBS=ON on all Windows builds (#3215) b1250 IsaacDynamo 2023-09-16 19:35:25 +02:00
  • 130f5fb1ed mv LogTargetWrapper private, simplify set_target staviq 2023-09-16 19:12:20 +02:00
  • 7aa9ea7f20
    fix consume_common_train_arg xaedes 2023-09-16 19:08:51 +02:00
  • bef1e97875
    move common opt_callback into common/train xaedes 2023-09-16 18:51:16 +02:00
  • e9758ae1d2
    move common train params into common/train xaedes 2023-09-16 18:45:59 +02:00
  • 71c49a692a simplify tee funcs, LogTargetWrapper yield lock staviq 2023-09-16 18:17:53 +02:00
  • 0631ea363c Don't crash on available devices if we can't even create an instance. master-0631ea3 Adam Treat 2023-09-16 12:17:29 -04:00
  • 65116c6015
    Merge branch 'ggerganov:master' into betterlogs2 staviq 2023-09-16 18:09:24 +02:00
  • ee27333b16
    move train data saving code into callback to unify code of opt_callback xaedes 2023-09-16 17:50:16 +02:00
  • a8c8907c62
    move train state into struct train_state xaedes 2023-09-16 17:08:18 +02:00
  • e107bce105 add another missing field Concedo 2023-09-16 23:17:22 +08:00
  • f01a75b563 added missing field Concedo 2023-09-16 23:15:44 +08:00
  • 733127b160 class.py off by 1 Concedo 2023-09-16 23:01:12 +08:00
  • 5dbc2b3213
    Enable build with CUDA 11.0 (make) (#3132) b1249 Vlad 2023-09-16 17:55:43 +03:00
  • 9f4b1bf88d
    move common train functions into common/train.[h|cpp] xaedes 2023-09-16 14:58:34 +02:00
  • 00b656f6db
    remove lbfgs related train parameters xaedes 2023-09-16 15:59:38 +02:00
  • b08e75baea
    Fixing the last deviations from sentencepiece indicated by test-tokenizer-1 (#3170) b1248 goerch 2023-09-16 13:41:33 +02:00
  • 11eb9fbd8b Enable BUILD_SHARED_LIBS=ON on all Windows builds IsaacDynamo 2023-09-16 13:02:23 +02:00
  • f6ba36dff6
    Reduce warnings. (#439) Ycros 2023-09-16 20:52:09 +10:00
  • 8d90072a2a updated class.py Concedo 2023-09-16 18:22:28 +08:00
  • c96fb3984d Merge branch 'master' into concedo_experimental Concedo 2023-09-16 12:14:19 +08:00
  • 32bc3f4fcf llama : enable mmap in quantize on Linux -> 31% faster Cebtenzzre 2023-09-09 23:04:53 -04:00
  • 53885de6db added multiuser mode Concedo 2023-09-16 11:23:39 +08:00
  • e7c6774ca7 make : restore build-info.h dependency for several targets Cebtenzzre 2023-09-15 23:16:54 -04:00
  • 4218641d97
    Separate CuBLAS/hipBLAS (#438) YellowRoseCx 2023-09-15 21:13:44 -05:00
  • afc0d0d160 Adding huichen's test case goerch 2023-09-16 02:16:06 +02:00
  • ab56b63b27
    update train-text-from-scratch with tokenization, sample selection and shuffling from finetune xaedes 2023-09-15 23:45:54 +02:00
  • cc60b3f639
    remove outcommented old code xaedes 2023-09-15 23:45:05 +02:00
  • 4f2ce91b9e
    add static keywords xaedes 2023-09-15 23:25:13 +02:00
  • 401badac40 Merge branch 'master' of github.com:ggerganov/llama.cpp Laura 2023-09-15 23:05:33 +02:00
  • e6616cf0db
    examples : add compiler version and target to build info (#2998) b1247 Cebtenzzre 2023-09-15 16:59:49 -04:00
  • e75ba80552 move comment Cebtenzzre 2023-09-15 16:36:41 -04:00
  • 6adc3b6ea6 do not use pipefail bashism Cebtenzzre 2023-09-15 16:35:56 -04:00
  • 7aa146ad90 fix build when git is unavailable Cebtenzzre 2023-09-15 16:13:02 -04:00
  • 3aefaab9e5
    check C++ code with -Wmissing-declarations (#3184) b1246 Cebtenzzre 2023-09-15 15:38:27 -04:00
  • bbb9511705 Merge branch 'master' of https://github.com/ggerganov/llama.cpp into build-info-compiler Cebtenzzre 2023-09-15 15:35:45 -04:00
  • 69eb67e282
    fix build numbers by setting fetch-depth=0 (#3197) b1245 Cebtenzzre 2023-09-15 15:18:15 -04:00
  • a9cc37806c
    Merge cec628e7fd into 4fe09dfe66 Erik Scholz 2023-09-15 22:10:19 +03:00
  • 06ebae8f86 server : revert some unnecessary style changes Cebtenzzre 2023-09-15 15:09:34 -04:00