Commit Graph

999 Commits

Author SHA1 Message Date
slaren
41c674161f
make rms_norm_eps a parameter (#2374)
* make rms_norm_eps a parameter

* add rms_norm_eps to command line

* fix baby llama, test-grad0

* use scientific notation for eps param in the help

ggml-ci
2023-07-24 17:57:12 +02:00
Aarni Koskela
b3f138d058
Chat UI extras (#2366)
* makefile: correct deps for server

* server: tighten settings layout a little

* server: expose all currently configured generation params in UI

* server: expose remaining generation params, for the adventurous

* server: embetter mirostat fields
2023-07-24 17:54:22 +03:00
Georgi Gerganov
5b2b2dc6ae
ggml : sync (unary ops refactor, static-correctness) (#2370)
* ggml : sync (unary ops, tests)

ggml-ci

* tests : remove unnecessary funcs
2023-07-24 14:46:21 +03:00
Kawrakow
42f70cb2f6
Fix scalar version of Q5_K when QK_K = 64 (#2362)
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-07-24 12:55:02 +03:00
Evan Jones
84e09a7d8b
llama : add grammar-based sampling (#1773)
* llama, main : constrain sampling to grammar

* allow loading grammar from file

* fix whitespace errors

* handle & print parser errors

* add comments to grammar syntax and allow newlines where unambiguous

* add missing include

* support alternates in root rule

* fix bugs with empty token and EOS

* adjust JSON grammar

* remove swp file

* rewrite ternary expressions

Co-authored-by: Henri Vasserman <henv@hot.ee>

* use struct for grammar elements and add Unicode support

* add unicode escapes

* add inverse char ranges

* only sample full tokens (no peeking or truncation)

* llama : minor style changes

blindly applied in online editor - hopefully I didn't break something

* update help text

* add warning message if EOS is disabled

---------

Co-authored-by: Henri Vasserman <henv@hot.ee>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-07-23 23:58:10 -04:00
Kawrakow
2f9cf974a0
Some more Q4_K and Q5_K speedup on CUDA (#2346)
* Faster Q5_K on CUDA

* Small Q5_K improvement on older GPUs

* Spped up Q4_K on CUDA

GTX1660: 29.5 ms/t -> 25.6 ms/t
RTX4080: 8.40 ms/t -> 8.25 ms/t

* Spped up Q4_K on CUDA

GTX1660: 36.7 ms/t -> 35.6 ms/t
RTX4080:  9.8 ms/t ->  9.5 ms/t

* Address PR comments

* Add some comments to satisfy PR reviewer

---------

Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-07-24 00:19:47 +03:00
IgnacioFDM
4f06592cc6
Add gqa parameter support to the server (#2351)
* Add gqa parameter support to the server
* Change help from stderr to stdout
2023-07-23 23:31:17 +03:00
Johannes Gäßler
70d26ac388
Fix __dp4a documentation (#2348) 2023-07-23 17:49:06 +02:00
wzy
57921ca6db
common : n_threads == -1 uses std:🧵:hardware_concurrency() (#2347)
* Fix #2345, fix incorrect n_threads

* Update examples/common.cpp

---------

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-07-23 16:33:02 +03:00
slaren
3602ac4255
fix n_tasks (#2342)
ggml-ci
2023-07-23 15:19:39 +02:00
slaren
95a6c595e7
ggml: move op parameters from tensors to ggml_tensor::op_params (#2333)
* ggml: move op parameters from tensors to ggml_tensor::op_params

* alibi: use memcpy for float params

* remove `src[1] = NULL` in ops
2023-07-23 14:36:02 +02:00
Georgi Gerganov
e76d630df1
llama : grouped-query attention + LLaMAv2 70B support (#2276)
* CUDA: GQA implementation

* llama : support for GQA and LLaMAv2 70B

ggml-ci

* py : fix hparams parsing (if-else blocks)

ggml-ci

* py : oh boy ..

ggml-ci

* help : fix gqa value for 70B

ggml-ci

---------

Co-authored-by: JohannesGaessler <johannesg@5d6.de>
2023-07-23 15:09:47 +03:00
maddes8cht
1d0824b247
llama : print help to stdout (#2338) 2023-07-23 14:59:48 +03:00
wzy
bc3ec2cdc9
flake : support nix build '.#opencl' (#2337) 2023-07-23 14:57:02 +03:00
Christian Demsar
a940458e48
llama : print max tensor size to stderr (#2336) 2023-07-23 14:56:34 +03:00
Jose Maldonado
91171b8072
make : fix CLBLAST compile support in FreeBSD (#2331)
* Fix Makefile for CLBLAST compile support and instructions for compile llama.cpp FreeBSD

* More general use-case for CLBLAST support (Linux and FreeBSD)
2023-07-23 14:52:08 +03:00
AustinMroz
355c80f49e
examples : simplify vim plugin (#2327)
Uses builtin json_encode and json_decode functions to simplify escaping
Removes the need for temp files
2023-07-23 14:16:48 +03:00
Jiahao Li
83a00ce69b
metal : support bcast add & dup & cont op (#2323) 2023-07-23 14:00:37 +03:00
Kawrakow
d2a43664f9
Speed up Q4_K (#2322)
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-07-23 08:49:20 +03:00
Johannes Gäßler
b9b7d94fc1
CUDA: Fixed 7b q3_K_S with mul_mat_vec_q (#2313) 2023-07-22 21:27:34 +02:00
Georgi Gerganov
b47b8a9cfe
llama : optimize memory buffers (#2325) 2023-07-22 21:17:57 +03:00
klosax
b5fe67f8c6
Perplexity: Compute scores correlated to HellaSwag (#2312)
* Add parameter --perplexity-lines to perplexity.cpp
2023-07-22 14:21:24 +02:00
whoreson
24baa54ac1
examples : basic VIM plugin
VIM plugin for server exe
2023-07-22 13:34:51 +03:00
Georgi Gerganov
dd6c67d3cb
ci : fix args 2023-07-22 12:00:56 +03:00
Georgi Gerganov
5d500e8ccf
ci : add 7B CUDA tests (#2319)
* ci : add 7B CUDA tests

ggml-ci

* ci : add Q2_K to the tests

* ci : bump CUDA ppl chunks

ggml-ci

* ci : increase CUDA TG len + add --ignore-eos

* ci : reduce CUDA ppl cunks down to 4 to save time
2023-07-22 11:48:22 +03:00
Richard Roberson
7d5f18468c
examples : add easy python script to create quantized (k-bit support) GGML models from local HF Transformer models (#2311)
* Resync my fork with new llama.cpp commits

* examples : rename to use dash instead of underscore

---------

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-07-21 22:01:10 +03:00
Kawrakow
d924522a46
Custom RoPE + bettter memory management for CUDA (#2295)
* Custom RoPE + bettter memory management for CUDA

* Adjusted look ahead in ggml_cuda_pool_malloc to 5%

This is sufficient it seems.
We end up using about 200 MB less VRAM that way when running
the 13B model with context 8192.

---------

Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-07-21 17:27:51 +03:00
Kawrakow
4d76a5f49b
Faster Q3_K implementation on Metal (#2307)
* Faster Q3_K on Metal

* Additional Q3_K speedup on Metal

* Q3_K for QK_K = 64

* Better Q3_K for QK_K = 64

21.6 ms/t -> 21.1 ms/t

---------

Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-07-21 17:05:30 +03:00
Georgi Gerganov
0db14fef06
ggml : fix the rope fix (513f861953) 2023-07-21 15:16:55 +03:00
Ikko Eltociear Ashimine
03e566977b
examples : fix typo in minigpt4.py (#2298)
promt -> prompt
2023-07-21 14:53:07 +03:00
Georgi Gerganov
513f861953
ggml : fix rope args order + assert (#2054) 2023-07-21 14:51:34 +03:00
Georgi Gerganov
3973b25a64
gitignore : fix final newline 2023-07-21 14:42:41 +03:00
Guillaume "Vermeille" Sanchez
ab0e26bdfb
llama : remove cfg smooth factor as it is only a reparameterization of the guidance scale (#2280) 2023-07-21 13:58:36 +03:00
Jose Maldonado
73643f5fb1
gitignore : changes for Poetry users + chat examples (#2284)
A fix in Makefile for FreeBSD users. In the platfrom x86_64 is amd64. This fix resolve compilation using CFLAGS and CXXFLAGS with -march=native and -mtune=native
Add two examples for interactive mode using Llama2 models (thx TheBloke for models)

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-07-21 13:53:27 +03:00
Georgi Gerganov
a814d04f81
make : fix indentation 2023-07-21 13:50:55 +03:00
Georgi Gerganov
4c013bb738
ci : fix MNT realpath usage (#2250) 2023-07-21 13:49:18 +03:00
Sky Yan
42c7c2e2e9
make : support customized LLAMA_CUDA_NVCC and LLAMA_CUDA_CCBIN (#2275)
Under certain environment, nvcc and gcc is installed under customized path but not standard path

Co-authored-by: Yan Lin <yanlin@baidu.com>
2023-07-21 13:38:57 +03:00
wzy
78a3d13424
flake : remove intel mkl from flake.nix due to missing files (#2277)
NixOS's mkl misses some libraries like mkl-sdl.pc. See #2261
Currently NixOS doesn't have intel C compiler (icx, icpx). See https://discourse.nixos.org/t/packaging-intel-math-kernel-libraries-mkl/975
So remove it from flake.nix

Some minor changes:

- Change pkgs.python310 to pkgs.python3 to keep latest
- Add pkgconfig to devShells.default
- Remove installPhase because we have `cmake --install` from #2256
2023-07-21 13:26:34 +03:00
Georgi Gerganov
ae178ab46b
llama : make tensor_split ptr instead of array (#2272) 2023-07-21 13:10:51 +03:00
Jiří Podivín
54e3bc76fe
make : add new target for test binaries (#2244)
Programs in the tests directory are now build with target tests
and placed in the same location.

* clean target was expanded to remove new binaries

* test target binaries are listed in a variable

* Locations of binaries were added to the .gitignore

Signed-off-by: Jiri Podivin <jpodivin@gmail.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
2023-07-21 13:09:16 +03:00
Hatsune Miku
019fe257bb
MIKU MAYHEM: Upgrading the Default Model for Maximum Fun 🎉 (#2287)
* Miku.sh: Set default model to llama-2-7b-chat

* Miku.sh: Set ctx_size to 4096

* Miku.sh: Add in-prefix/in-suffix opts

* Miku.sh: Switch sampler to mirostat_v2 and tiny prompt improvements
2023-07-21 11:13:18 +03:00
Kawrakow
e68c96f7fe
Faster Q2_K on Metal (#2297)
* Faster Q2_K on Metal

* Deleting unnoticed and dangereous trailing white space

* Fixed bug in new metal Q2_K implementation

---------

Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-07-21 10:44:40 +03:00
Przemysław Pawełczyk
9cf022a188
make : fix embdinput library and server examples building on MSYS2 (#2235)
* make : fix embdinput library and server examples building on MSYS2

* cmake : fix server example building on MSYS2
2023-07-21 10:42:21 +03:00
Kawrakow
e782c9e735
Faster Q5_K and Q6_K on Metal (#2294)
* Faster Q6_K on Metal

* Faster Q5_K on Metal

* Another Q5_K speedup

---------

Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-07-20 18:19:45 +03:00
Kawrakow
785829dfe8
Faster Q4_K on Metal (#2290)
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
2023-07-20 15:18:43 +03:00
Georgi Gerganov
fff0e0eafe llama : fix regression from #2000 - could not load no-mmap models 2023-07-20 13:47:26 +03:00
Shouzheng Liu
417a85a001
metal: minor q4 optimization and reduce code size (#2248)
* metal: use uint16_t instead of uint8_t.

Apple GPU doesn't like uint8_t. For every operation on uint8_t
the gpu need to copy the uint8_t to an empty 16 bit register, then
it can issue other instructions.

For the matrix-vector multiplication kernel only, we observed a
340~350 GB/s memory read speed on M1 Max after this commit, which is
very close to the reported hardware limit.

* metal: update rms_norm kernel

This commit double the speed of rms_norm operations by using 512 threads
per threadgroup, combining with SIMD primitives to minimize the need for
thread group barriers.

* metal: use template to reduce size

Revert modifications on block_q4_0 and block_q4_1.
2023-07-20 13:32:22 +03:00
Rinne
294f424554
llama : extend API to get max devices at runtime (#2253) 2023-07-19 10:06:40 +03:00
wzy
45a1b07e9b
flake : update flake.nix (#2270)
When `isx86_32 || isx86_64`, it will use mkl, else openblas

According to
https://discourse.nixos.org/t/rpath-of-binary-contains-a-forbidden-reference-to-build/12200/3,
add -DCMAKE_SKIP_BUILD_RPATH=ON

Fix #2261, Nix doesn't provide mkl-sdl.pc.
When we build with -DBUILD_SHARED_LIBS=ON, -DLLAMA_BLAS_VENDOR=Intel10_lp64
replace mkl-sdl.pc by mkl-dynamic-lp64-iomp.pc
2023-07-19 10:01:55 +03:00
wzy
b1f4290953
cmake : install targets (#2256)
fix #2252
2023-07-19 10:01:11 +03:00