]>
git.djapps.eu Git - pkg/ggml/sources/ggml/log
Georgi Gerganov [Sun, 25 Jun 2023 10:07:18 +0000 (13:07 +0300)]
readme : add encodec.cpp link
Georgi Gerganov [Sat, 24 Jun 2023 17:58:42 +0000 (20:58 +0300)]
readme : add BioGPT example link
LoganDark [Sat, 24 Jun 2023 17:47:53 +0000 (10:47 -0700)]
ggml : add custom mapping functions (#264)
* Add custom mapping functions
The current mapping functions are basically jokes, add some real
ones. These ones get access to the actual tensor structs so they
can do things like
- Know the dimensions they are operating on
- Work with tensors with more than 2 dimensions, or transposed
- Operate on two differently sized tensors (like matmul)
- Use their own thread pool that does a better job than ggml does.
Among other things ...
* fix ordering mistake
* ggml : custom operators support scratch buffers
---------
Co-authored-by: Georgi Gerganov <redacted>
sjinzh [Sat, 24 Jun 2023 17:03:13 +0000 (01:03 +0800)]
zig : add zig build system support (#279)
* add zig build system support
* add zig build system support
Georgi Gerganov [Sat, 24 Jun 2023 16:39:32 +0000 (19:39 +0300)]
tests : allow to set threads to test-grad0
Borislav Stanimirov [Sat, 24 Jun 2023 16:11:35 +0000 (19:11 +0300)]
build : fix compilation errors and warnigns when building with MSVC (#275)
Borislav Stanimirov [Sat, 24 Jun 2023 16:06:13 +0000 (19:06 +0300)]
tests : increase stack size for test1 when building with MSVC (#277)
Georgi Gerganov [Sat, 24 Jun 2023 16:03:09 +0000 (19:03 +0300)]
tests : use LBFGS optimizer instead of ADAM (close #276)
ADAM seems to behave differently since the recent training changes.
Need to see how to make it work again for test2 - probably some
parameters need to be adjusted
AmbientL [Sat, 24 Jun 2023 15:31:38 +0000 (15:31 +0000)]
ggml : more verbose memory allocation failure (#270)
AmbientL [Sat, 24 Jun 2023 15:30:23 +0000 (15:30 +0000)]
starcoder : add special tokens for fill-in-the-middle task (#269)
Georgi Gerganov [Sat, 24 Jun 2023 15:27:46 +0000 (18:27 +0300)]
ggml : sync llama.cpp (tensor names)
Georgi Gerganov [Mon, 19 Jun 2023 18:28:16 +0000 (21:28 +0300)]
ci : reduce GGML_NLOOP to 3
Georgi Gerganov [Mon, 19 Jun 2023 17:43:19 +0000 (20:43 +0300)]
tests : sync test-grad0 from llama.cpp
Georgi Gerganov [Mon, 19 Jun 2023 17:43:12 +0000 (20:43 +0300)]
ggml : fix bug in LBFGS optimizer
Georgi Gerganov [Mon, 19 Jun 2023 17:35:08 +0000 (20:35 +0300)]
ggml : sync latest llama.cpp
Ebey Abraham [Sun, 18 Jun 2023 10:33:38 +0000 (11:33 +0100)]
gpt-2 : fix typo (#261)
Co-authored-by: Ebey Abraham <redacted>
Avi Lumelsky [Sun, 18 Jun 2023 10:32:09 +0000 (13:32 +0300)]
whisper : removed duplicate lines in convert-pt-to-ggml.py (#256)
Deleted 2 lines of .astype(float32) conversion to the model weights (No real impact, just for cleaner code)
Lukas Möller [Sun, 18 Jun 2023 08:34:21 +0000 (10:34 +0200)]
replit : update inference code to match reference (#218)
* Update replit inference code to match reference
* Add qntvr printf
Adam Tazi [Sun, 18 Jun 2023 08:15:58 +0000 (01:15 -0700)]
ci : introduce Github Actions CI workflow (#247)
* Introduce Github Actions CI workflow for the ggml repo
This commit integrates a Github Actions CI workflow that compiles and tests the codebase on both Ubuntu 22.04 and macOS 12 Monterey. The workflow is triggered on pull requests against the main branch and on every push to the main branch.
To accommodate the resource constraints of the Github-hosted runners, a `GGML_NITER` environment variable is introduced, allowing tests to run within a reasonable time frame. `test-grad0.c` is modified to use this variable instead of `GGML_NLOOP`.
The workflow file includes:
- A build strategy for both Ubuntu and MacOS.
- An environment setup with variables `GGML_NLOOP` and `GGML_NITER`.
- A step to limit the number of threads used by `test2.c` for efficient execution.
- A typical build process with steps for environment creation, CMake configuration, building, and verbose testing with a timeout.
* main to master
Tanmay [Sun, 18 Jun 2023 08:09:48 +0000 (13:39 +0530)]
ggml : convert interleaved addressing to sequential addressing for reduce functions (#117)
* Convert interleaved addressing to sequential addressing for REDUCE
* update addressing on new archs
Ravindra Marella [Sun, 18 Jun 2023 07:54:59 +0000 (13:24 +0530)]
examples : fix c++ standard errors and pedantic warnings (#239)
Cristiano Calcagno [Sun, 18 Jun 2023 07:45:11 +0000 (09:45 +0200)]
ggml : fix minor resource leak reported by static analysis (#237)
Ravindra Marella [Sun, 18 Jun 2023 07:37:09 +0000 (13:07 +0530)]
starcoder : add support for starchat special tokens (#246)
* starcoder : add support for starchat special tokens
* examples : fix `gpt_tokenize()` for special tokens
LoganDark [Fri, 16 Jun 2023 19:39:09 +0000 (12:39 -0700)]
ggml : return input tensor in ggml_set_name (#262)
this is SO USEFUL for debugging. in order to find any cgraph node,
I can wrap it in ggml_set_name and set a conditional breakpoint.
but I can only wrap existing code if this returns its input.
otherwise the barrier becomes annoyingly high (have to move a
bunch of code around to add name to a tensor)
LoganDark [Fri, 16 Jun 2023 19:17:30 +0000 (12:17 -0700)]
ggml : fix ggml_clamp (#263)
This unconditionally failed before
M. Yusuf Sarıgöz [Fri, 16 Jun 2023 17:36:46 +0000 (20:36 +0300)]
ggml : add quick GELU (#254)
* Implement Quick GELU
* Revert "Implement Quick GELU"
This reverts commit
ff220cc1f91a184f195d19b17ed4c352cc72a6f0 .
* Tidy up ggml.h
* Respect to the style of ggml
* Fix: Fix minor typo
* Rename `quick_gelu` -> `gelu_quick`
Andrei [Thu, 8 Jun 2023 18:51:39 +0000 (14:51 -0400)]
cmake : export all symbols on windows when building shared library (#234)
Currently building ggml on windows as a shared library does not export all symbols by default.
LoganDark [Wed, 7 Jun 2023 16:16:19 +0000 (09:16 -0700)]
ggml : correct off-by-one bounds check in ggml_compute_forward_set_f32 (#229)
without this fix you will be unable to set a zero-length tensor to the end of another tensor
this sounds stupid, but is used in my testing
klosax [Wed, 7 Jun 2023 16:15:50 +0000 (18:15 +0200)]
gpt-neox : fix ctx size calculation (#228)
Georgi Gerganov [Wed, 7 Jun 2023 16:14:50 +0000 (19:14 +0300)]
ggml : fix ggml_clamp thresholds being read as ints instead of floats (#221)
Jiahao Li [Wed, 7 Jun 2023 16:14:27 +0000 (00:14 +0800)]
ggml : add inplace ops api in header file (#219)
Georgi Gerganov [Fri, 2 Jun 2023 12:46:59 +0000 (15:46 +0300)]
ggml : add ggml_conv_2d_sk_p0(), ggml_win_part(), ggml_win_unpart()
Georgi Gerganov [Tue, 30 May 2023 10:49:08 +0000 (13:49 +0300)]
ggml : fix ggml op conv_1d enum names
Georgi Gerganov [Tue, 30 May 2023 10:19:55 +0000 (13:19 +0300)]
ggml : better conv_1d naming
Georgi Gerganov [Tue, 30 May 2023 07:18:31 +0000 (10:18 +0300)]
ggml : rename conv_1d ops to reflect half-padding used
Georgi Gerganov [Tue, 30 May 2023 07:03:30 +0000 (10:03 +0300)]
ggml : fix compiler warnings for printf
Georgi Gerganov [Mon, 29 May 2023 18:14:52 +0000 (21:14 +0300)]
mnist : remove redundant stuff + rename ctx0
Eldar Yusupov [Mon, 29 May 2023 16:55:13 +0000 (19:55 +0300)]
mnist : add missing header (#213)
Eldar Yusupov [Mon, 29 May 2023 16:47:57 +0000 (19:47 +0300)]
common : fix compilation on Linux (#212)
Georgi Gerganov [Mon, 29 May 2023 16:28:07 +0000 (19:28 +0300)]
ggml : cgraph export/import/eval example + GPU support (#108)
* ggml : cgraph export brainstorming
* mnist : code style
* mnist : minor
* ggml : initial cgraph export
* ggml : initial graph import (wip)
* ggml : import op args correctly
* ggml : add ggml_get_tensor_by_name()
* mnist : add compute graph evaluation on CPU example
* ggml : add ggml_tensor_overhead()
* ggml : rename new functions to ggml_cgraph_...
* mnist : add Metal inference skeleton (WIP)
* mnist : working on the Metal pipeline (WIP)
* mnist : prepare the Metal encoder (WIP)
* mnist : first Metal kernel for F32 ADD
* mnist : looks like MTLHeap does not work
* mnist : initial full pass of MNIST on the GPU (not verified)
* mnist : minor cleanup
* mnist : full GPU inference works
* mnist : use custom soft_max kernel since MPSMatrixSoftMax is bugged
* mnist : use constant for soft_max instead of hardcoded 10
* mnist : check multiple predictions (Metal)
* mnist : minor
* ggml : move cgraph import / export to ggml
* mnist : remove common dependencies
* mnist : fix soft_max threadgroup size
* mnist : init no_alloc member
* ggml : improve "get tensor" API
Tyé singwa [Sun, 28 May 2023 17:41:11 +0000 (20:41 +0300)]
fix : fix ggml_alibi (#204)
Skyler Celestinian-Sterling [Sun, 28 May 2023 10:45:30 +0000 (03:45 -0700)]
readme : add "development" (#203)
You are welcome lol
apcameron [Sat, 27 May 2023 13:48:33 +0000 (14:48 +0100)]
ggml : add CLBLAST support (#197)
Enable support for the RISCV architecture
This addresses https://github.com/ggerganov/ggml/issues/129
Georgi Gerganov [Sat, 27 May 2023 13:20:24 +0000 (16:20 +0300)]
cuda : sync latest llama.cpp (control DMMV X/Y sizes)
Georgi Gerganov [Sat, 27 May 2023 13:18:28 +0000 (16:18 +0300)]
ggml : add ggml_tensor_overhead() + ggml_get_tensort_by_name()
Georgi Gerganov [Sat, 27 May 2023 08:55:25 +0000 (11:55 +0300)]
ggml : sync llama.cpp (OpenCL support for GPU offload)
Georgi Gerganov [Sat, 27 May 2023 08:51:29 +0000 (11:51 +0300)]
mnist : gitignore stuff
jaeminSon [Sat, 27 May 2023 08:47:34 +0000 (17:47 +0900)]
examples : add tokenization tests and refactor codes (#186)
* examples : [refactor] remove unnecessary lines and segments
* examples : [feature] add tokenization test for gpt-neox
* examples : [feature] handle multibyte character set
* examples : [refactor] find the longest token for word
* examples : [refactor] move test_tokenizer to common.cpp as the function affects other models
* add 'test_tokenizer' function after loading the model
* examples : [feature] add test cases for checking tokenization
* examples : [feature] tokenize with huggingface tokenizers for currently supported models
* examples : add tokenization test cases for each model
* revert conversion from string to utf-8 encoded byte strings
* [refactor] make util functions for testing tokenizers available
* [bug fix] test replit using functions and variables (e.g. tokenizer struct, tokenization method) defined in its main.cpp
* [refactor] modify function name test_tokenizer -> test_gpt_tokenizer
* [refactor] put parenthesis on single line for-loops and if-statements
* [refactor] withdraw <filesystem> and use <iostream> and <dirent.h>
* [refactor] remove 'find_test_file' function and directly set test file path from 'test_gpt_tokenizer' function
* call a function for testing tokenizer with filename specified
* revert test tokenizer in replit (replit uses seperate methods for tokenzation and decoding)
* compare vector of id to check if two tokenizations are identical.
* write token ids instead of strings.
* [refactor] use --token_test rather than --test for token-test argument
* add english test cases
* update test cases with more english prompts
* examples : tokenizer testing fixes
---------
Co-authored-by: Georgi Gerganov <redacted>
Georgi Gerganov [Sat, 27 May 2023 06:11:10 +0000 (09:11 +0300)]
Update README.md
Georgi Gerganov [Fri, 26 May 2023 09:34:29 +0000 (12:34 +0300)]
minor : fix printf warnings
Radoslav Gerganov [Fri, 26 May 2023 08:53:18 +0000 (11:53 +0300)]
mnist : smooth user input (#199)
Drawing on the canvas is now smooth. The final image which is used for
prediction is obtained by down-scaling the canvas to 28x28 pixels.
Download button is aslo added for downloading raw image values.
Radoslav Gerganov [Fri, 26 May 2023 06:36:40 +0000 (09:36 +0300)]
examples : add missing header file (#198)
Some of the examples are missing the cstring header which is needed for
memcpy().
Radoslav Gerganov [Thu, 25 May 2023 08:27:15 +0000 (11:27 +0300)]
mnist : add progress indicator on the web page (#194)
Prevent user actions before the model and the data set is loaded
Georgi Gerganov [Wed, 24 May 2023 08:49:53 +0000 (11:49 +0300)]
mnist : add WASM instructions + web-page link
Radoslav Gerganov [Wed, 24 May 2023 08:40:47 +0000 (11:40 +0300)]
mnist : add web page for the MNIST example (#190)
The web page is using WASM for model inference.
Users can draw digits on an HTML canvas and load random digits from the
MNIST dataset.
Georgi Gerganov [Wed, 24 May 2023 08:38:41 +0000 (11:38 +0300)]
mnist : cleanup main.cpp
Ettore Di Giacinto [Wed, 24 May 2023 08:01:31 +0000 (10:01 +0200)]
docs : add golang transformer bindings (#191)
This PR adds golang bindings to transformers in ggml
Georgi Gerganov [Wed, 24 May 2023 07:54:45 +0000 (10:54 +0300)]
mpt : fix n_ctx (close #165)
Georgi Gerganov [Wed, 24 May 2023 07:41:06 +0000 (10:41 +0300)]
examples : remove prompt pipe-in support
Need cross-platform solution, factored out in common
Georgi Gerganov [Wed, 24 May 2023 07:40:27 +0000 (10:40 +0300)]
common : add missing declarations
klosax [Wed, 24 May 2023 07:27:36 +0000 (09:27 +0200)]
mpt : utf-8 support, perplexity testing, repeat penalty sampling (#184)
* common: utf-8 decoder, reverted gpt_toeknize utf-8 convert
* Update common.h
* main: decode utf-8 tokens on load
* mpt import: bug fix
* common: style fixes
* common: style fix
* Update common.h
* common: revert gpt_tokenize utf-8 convert
* Update common.cpp
* Update common.cpp
* Update common.cpp
* Add perplexity to mpt
* Update CMakeLists: perplexity
* mpt-perplexity: fixes
* Update perplexity.cpp
* common: add sampling with repeat penalty
* mpt-main: add repeat penalty sampling, add commandline parameters
* Update common.h
* mpt-main: style fixes
* Update perplexity.cpp
* Delete perplexity.cpp
* mpt: move perplexity to main
* mpt: move perplexity to main
* common.cpp: Use codecvt utf-8 converter
* main.cpp: Use codecvt utf-8 converter
* mpt : code style changes
---------
Co-authored-by: Georgi Gerganov <redacted>
Georgi Gerganov [Mon, 22 May 2023 14:57:21 +0000 (17:57 +0300)]
readme : update Features
Ravindra Marella [Sun, 21 May 2023 12:32:05 +0000 (18:02 +0530)]
readme : add link to python bindings (#181)
klosax [Sun, 21 May 2023 08:21:51 +0000 (10:21 +0200)]
common : support utf-8 + fix gpt_tokenize + fix MPT model import (#179)
* Update convert-h5-to-ggml.py
* Import tokens correctly
* gpt_tokenize: Convert input to utf-8 + bug fix
* common : minor style fixes
---------
Co-authored-by: Georgi Gerganov <redacted>
Dan Forbes [Sat, 20 May 2023 18:25:25 +0000 (11:25 -0700)]
readme : add link to GGML format docs (#177)
Georgi Gerganov [Sat, 20 May 2023 17:56:35 +0000 (20:56 +0300)]
examples : use scratch buffers to reduce memory usage (#176)
* starcoder : example for using scratch buffers to reduce memory usage
* starcoder : bump scratch buffers to 256 MB
* examples : add scratch buffers to MPT and GPT-NeoX
Georgi Gerganov [Sat, 20 May 2023 17:00:27 +0000 (20:00 +0300)]
ggml : update WASM SIMD
Georgi Gerganov [Sat, 20 May 2023 15:59:04 +0000 (18:59 +0300)]
whisper : fix Hebrew lang id
Georgi Gerganov [Sat, 20 May 2023 15:01:40 +0000 (18:01 +0300)]
examples : add quantize version to MPT and Replit examples (ref #168)
Georgi Gerganov [Sat, 20 May 2023 14:45:49 +0000 (17:45 +0300)]
common : force --top_k to be at least 1
Georgi Gerganov [Sat, 20 May 2023 14:33:07 +0000 (17:33 +0300)]
examples : fix vocab loading (close #163)
Georgi Gerganov [Sat, 20 May 2023 14:22:58 +0000 (17:22 +0300)]
common : fix gpt_tokenize (ref #170)
Michael Verrilli [Sat, 20 May 2023 14:12:24 +0000 (10:12 -0400)]
dolly-v2 : par_res and neox changes (#167)
* dolly-v2 example: par_res and neox changes
* Update examples/dolly-v2/quantize.cpp
---------
Co-authored-by: Georgi Gerganov <redacted>
Georgi Gerganov [Sat, 20 May 2023 14:09:41 +0000 (17:09 +0300)]
examples : call ggml_time_init() (close #166)
Georgi Gerganov [Sat, 20 May 2023 13:48:03 +0000 (16:48 +0300)]
Update README.md
Georgi Gerganov [Sat, 20 May 2023 12:59:34 +0000 (15:59 +0300)]
ggml : sync llama.cpp - CUDA improvements + ggml minor fixes
Georgi Gerganov [Sat, 20 May 2023 11:56:14 +0000 (14:56 +0300)]
ggml : sync llama.cpp - new quantization formats Q4 + Q8
pikalover6 [Thu, 18 May 2023 06:52:22 +0000 (23:52 -0700)]
readme : update roadmap (#164)
+ MPT & Replit
Lukas Möller [Wed, 17 May 2023 19:58:21 +0000 (21:58 +0200)]
examples : sample replit + MPT inference (#145)
* Add replit model
* Add unigram tokenization support
* Remove debug log
* Port alibi attn bias fix
* Remove torch input
* Fix hardcoded path
* Remove unsupported hyperparams
* Add mpt
* Add replit quantization script
* Remove debug print
* Add quantization support to mpt
* Reformat
* Remove trailing return type
* Implement stylistic changes
* use f16 in k/v memory calculations for replit/mpt
* Update context size calculation
* Add clip_qkv and alibi_bias_max support
* fix clamping implementation, remove implicit conversions
* Fix qkv if condition
* Fix replit context size calculation
* Potentially fix gcc compilation error
* Fix warning
* Adjust object overhead
* Remove dead code
jaeminSon [Wed, 17 May 2023 15:49:37 +0000 (00:49 +0900)]
examples : fix a hyperparameter value in gpt-neox (#161) (#162)
Andrei [Wed, 17 May 2023 06:27:11 +0000 (02:27 -0400)]
ggml : fix typo in ggml_diag_mask_zero_inplace() (#159)
Georgi Gerganov [Mon, 15 May 2023 04:50:54 +0000 (07:50 +0300)]
readme : add link to training example
Georgi Gerganov [Sun, 14 May 2023 15:56:18 +0000 (18:56 +0300)]
ggml : add AVX dot products
Georgi Gerganov [Sun, 14 May 2023 15:55:29 +0000 (18:55 +0300)]
whisper : sync whisper.cpp
IGUILIZ Salah-Eddine [Sun, 14 May 2023 15:31:08 +0000 (17:31 +0200)]
starcoder : detect santacoder fix end of text token (#155)
Co-authored-by: IGUILIZ Salah-Eddine <redacted>
Georgi Gerganov [Sun, 14 May 2023 14:26:41 +0000 (17:26 +0300)]
readme : add re-quantization warning
Georgi Gerganov [Sun, 14 May 2023 12:10:32 +0000 (15:10 +0300)]
examples : use inplace calls explicitly
Georgi Gerganov [Sun, 14 May 2023 11:55:28 +0000 (14:55 +0300)]
tests : add tests from llama.cpp
Georgi Gerganov [Sun, 14 May 2023 11:45:13 +0000 (14:45 +0300)]
ggml : fix multi-threaded ggml_compute_forward_diag_mask_f32()
Georgi Gerganov [Sun, 14 May 2023 11:16:47 +0000 (14:16 +0300)]
ggml : fix rope calculation (!inplace + GPT-NeoX mode)
Georgi Gerganov [Sun, 14 May 2023 08:23:02 +0000 (11:23 +0300)]
ggml : new Q4 and Q5 quantization formats + backward ops
sync llama.cpp
- bump GGML_QNT_VERSION -> 1
- increase cwggml object overhead size from 256 to 512 in examples
- drop Q4_2 support
- tensor backend support CUDA
Georgi Gerganov [Sun, 14 May 2023 07:07:27 +0000 (10:07 +0300)]
ggml : add GGML_QNT_VERSION for tracking changes to the quantization format
ref #150
Georgi Gerganov [Sun, 14 May 2023 07:06:19 +0000 (10:06 +0300)]
whisper : sync whisper.cpp minor changes
Ravindra Marella [Sat, 13 May 2023 13:47:02 +0000 (19:17 +0530)]
starcoder : update example to follow the naming convention of other examples (#153)
Georgi Gerganov [Sat, 13 May 2023 13:02:49 +0000 (16:02 +0300)]
readme : fix gpt-neox example link
Ravindra Marella [Sat, 13 May 2023 12:24:47 +0000 (17:54 +0530)]
examples : fix warnings (#152)
Nouamane Tazi [Sat, 13 May 2023 10:46:10 +0000 (12:46 +0200)]
readme : add BLOOM example (#151)
Georgi Gerganov [Sat, 13 May 2023 10:08:56 +0000 (13:08 +0300)]
examples : update readme with new quantization usage + remove bug alert
Georgi Gerganov [Sat, 13 May 2023 10:04:57 +0000 (13:04 +0300)]
readme : update example list (#146)
Nouamane Tazi [Sat, 13 May 2023 09:54:03 +0000 (11:54 +0200)]
examples : add StarCoder/SantaCoder sample inference (#146)
* init commit
* fix building starcoder
* gen work
* fix vocab
* santacoder mha
* .
* fix quantize
* offload_state_dict
* endoftext
* rename scripts
* fix main
* scripts
* update README
* quickfixes