Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

llama : add functions to get the model's metadata #4013

Merged
merged 3 commits into from
Nov 17, 2023
Merged

Conversation

slaren
Copy link
Collaborator

@slaren slaren commented Nov 9, 2023

These functions are meant to provide an easy way to read the model's metadata without needing to use one of the GGUF libraries.

  • Arrays are excluded
  • Also prints metadata values on load and adds bounds checks to some gguf functions

@KerfuffleV2
Copy link
Collaborator

I kind of have mixed feelings about this. I definitely agree that being able to access those values is sorely needed, but most of the time I'd really rather access it through a function like llama_n_vocab instead of having to know details about the GGUF file format.

@slaren
Copy link
Collaborator Author

slaren commented Nov 10, 2023

In part, the motivation to add this is to be able to add the jinja prompt template metadata in the future, which will also require having a reasonable way to access it. I don't really see an issue with being able to enumerate metadata directly, and I am not sure that we should add specific functions to obtain metadata that is not directly relevant to the functions of llama.cpp.

@KerfuffleV2
Copy link
Collaborator

I don't really see an issue with being able to enumerate metadata directly

Sorry I wasn't clear. I wasn't saying there was an issue with adding it, just that I'd like to access metadata (that is relavant to the functions of llama.cpp) via functions like llama_n_vocab rather than having to know details about the GGUF format and what specific metadata keys are called.

Just for example, there's no way (as far as I know) to determine how many layers are in the currently loaded model.

@KerfuffleV2
Copy link
Collaborator

KerfuffleV2 commented Nov 11, 2023

Hmm, ironically I actually need this now but it doesn't seem to add enough metadata access methods.

I can use llama_model_meta_count and llama_model_meta_key_by_index to get the keys but I don't have a way to find the value's type. Being able to fetch the value as a string is also pretty awkward and apparently I'll get "???" in my string result if the value happened to be an array. It would also be ambiguous whether the value is the string "1" or an integer value 1.

I think this kind of needs to work like the GGUF metadata functions in ggml except that it takes a const struct llama_model *. An alternative would be to nuke most of this and just add a function to access the struct gguf_context in the model. It's kind of unfortunate the functions in ggml take struct gguf_context* instead of struct gguf_kv * otherwise you wouldn't have to export the whole gguf_context.

I hate to say it but accessing non-string values as a string with the weird stuff like having to do escaping seems like it's not going to be useful most of the time.

@slaren
Copy link
Collaborator Author

slaren commented Nov 11, 2023

Arrays are explicitly excluded from the metadata returned by these functions. You won't get a "???" if you try to read an array, you will get an error. I think you are misunderstanding the code, the string "???" is only returned with arrays of arrays, which is not supported by GGUF (or at least not by the C library). And even then, that is only relevant during in the info prints during loading.

The reason for returning only strings is for simplicity from the user perspective. If you need something more advanced, you are free to use ggml or gguf.py directly to read all the metadata. There is no point replicating the GGUF library in llama.cpp. The goal here is to provide a very simply API that will allow applications to access metadata such as the jinja templates.

@KerfuffleV2
Copy link
Collaborator

I think you are misunderstanding the code

You're correct, sorry, I did misread it.

There is no point replicating the GGUF library in llama.cpp.

I agree, but can't we at least make it so if someone already has a llama_context * or a llama_model * they can use the existing metadata functions from GGML? That would be as simple as adding some way to get the gguf_context *... Or from looking at the code, it looks like that doesn't get saved after the loading process? So I guess it would need to also be saved in the model or context structure.

@slaren
Copy link
Collaborator Author

slaren commented Nov 11, 2023

I don't really think there is a good enough reason to do that. As you say, currently the gguf_context is released after loading, and this would force us to keep it in memory. Additionally, this would mean exposing internal details that may restrict our ability to make changes in the future. In contrast, this exposes an abstraction instead of the internals, and the implementation can be easily changed without changing the interface. The applications can easily do the same with just an additional call to gguf_init_from_file.

@KerfuffleV2
Copy link
Collaborator

KerfuffleV2 commented Nov 11, 2023

The applications can easily do the same with just an additional call to gguf_init_from_file.

It seems incredibly strange to me that after we've loaded the file and all the metadata and returned a handle to the model that the only way for the user to access the metadata is to load the file a second time.

I don't really like the return the whole gguf_context idea either. The problem is the gguf_get_val_ functions expect a gguf_context instead of what they really need, just the KV stuff. If we are allowed to change that API, then I think there should be a struct containing the number of KVs and struct gguf_kv array and then accessor functions for it. Then llama.cpp could just wrap those with it's own get_val type functions, or save and allow accessing the KV structure so that the functions in GGML could be used on it.

But literally having to reload the file that was already loaded just makes absolutely no sense to me.

@slaren
Copy link
Collaborator Author

slaren commented Nov 11, 2023

The cost of loading the metadata again from disk is not really significant, and it doesn't happen in a performance sensitive path, it happens during load time. Nobody really cares if loading takes 10ms more because the metadata needs to be loaded twice. So I don't think that this is something that needs to be optimized at this point. Never forget the first rule of optimization!

@@ -1297,6 +1351,9 @@ struct llama_model {

int n_gpu_layers;

// gguf metadata
std::unordered_map<std::string, std::string> gguf_kv;
Copy link
Collaborator

@KerfuffleV2 KerfuffleV2 Nov 11, 2023

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

What's this doing here then? Why aren't you taking your own suggestion to just gguf_init_from_file again?

All I'm saying is save this information in a format that's not only useful for Jinja templates or whatever. Where there's a way to reliably access it and something like 1 can be differentiated from "1".


Never forget the first rule of optimization!

And yes, yes, I know. "Premature optimization is the root of all evil." To be honest, I never forget it but pretty much always disregard it. My objection here isn't really about optimization, but more for a logical relatively intuitive API where you don't have to do weird stuff reload the same file from disk multiple times.

When you do that, there's also no way to even know for sure you reloaded the same file as the model that's in memory.

Copy link
Collaborator Author

@slaren slaren Nov 11, 2023

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I am not trying to be snarky, but it would take a lot more code to re-load the gguf when using the metadata reading functions, and the only purpose would be to save a few bytes of memory (the reason arrays are excluded is precisely because it would waste too much memory for something of questionable utility). So doing that would be another case of premature optimization.

To be clear, I am not opposed to extending the API to support more types of metadata, that's why the function are called llama_model_meta_val_str and not just llama_model_meta_val, I expected that somebody may want to extend the API in the future. I am just not convinced that it is really worth the effort and complexity to do that because the expected use of these functions is to get very specific metadata like the jinja templates, which is trivial to convert to a number if needed, or just to dump a bunch of useless info to the user, in which case it doesn't matter what the type is. Making everything accesible as a string is simpler than having to check the correct function to call depending on the type. If this was a C++17 library I would be happy to make it return a std::variant, but that's not what it is.

@ggerganov ggerganov merged commit e85bb1a into master Nov 17, 2023
33 checks passed
olexiyb pushed a commit to Sanctum-AI/llama.cpp that referenced this pull request Nov 23, 2023
* llama : add functions to get the model's metadata

* format -> std::to_string

* better documentation
hodlen added a commit to hodlen/llama.cpp that referenced this pull request Apr 1, 2024
llama : restore prefix space in llama tokenizer (ggerganov#4081)

gguf : fix potential infinite loops while parsing (ggerganov#4100)

Co-authored-by: Bernhard Gstrein <[email protected]>

Respect tokenizer.ggml.add_bos_token value when tokenizing (ggerganov#4040)

* gguf-py: gguf-dump: Respect --no-tensor flag in JSON mode.

* Respect add_bos_token GGUF metadata value

* gguf-py: Try to fix SpecialVocab giving up too easily for the Nth time

llama : fix data units (ggerganov#4101)

* llama : fix data units

ggml-ci

* Revert "llama : fix data units"

This reverts commit f5feac8.

* llama : disambiguate data units

ggml-ci

cuda : get_row_rounding F32 (ggerganov#4095)

* Fix ggerganov#4017

* Update ggml-cuda.cu

Co-authored-by: Jared Van Bortel <[email protected]>

* Update ggml-cuda.cu

Co-authored-by: Jared Van Bortel <[email protected]>

---------

Co-authored-by: Jared Van Bortel <[email protected]>

finetune : zero the loraB initial vectors (ggerganov#4082)

* finetune : zero the loraB initial vectors

Without this, the first iteration is starting out far from the base model, instead of exactly on it.
Zeroing loraB is what the paper recommends. loralib also zeroes at least one of the init vector pairs
(though it departs from the paper in using a different distribution for the other vector, in some cases).

* tabs to spaces

* Use ggml_set_zero instead of adding a new function

finetune : speed-up ggml_compute_forward_out_prod_f32 via BLAS (ggerganov#4079)

* Remove logically superfluous assertions and order by dimension

* Use cblas_sgemm() to implement ggml_compute_forward_out_prod()

* Remove ggml_compute_forward_out_prod_use_blas(), fix compiling errors on cmake/zig, remove trailing whitespace

* Add openBLAS support for sgemm() in compute_forward_out_prod()

llama : add functions to get the model's metadata (ggerganov#4013)

* llama : add functions to get the model's metadata

* format -> std::to_string

* better documentation

train : move number of gpu layers argument parsing to common/train.cpp (ggerganov#4074)

- introduces help entry for the argument
 - cuts '--gpu-layers' form in order to simplify usage and documentation.

Signed-off-by: Jiri Podivin <[email protected]>
Co-authored-by: Jiri Podivin <[email protected]>

py : remove superfluous import statements (ggerganov#4076)

Signed-off-by: Jiri Podivin <[email protected]>
Co-authored-by: Jiri Podivin <[email protected]>

llava : fix compilation warning that fread return value is not used (ggerganov#4069)

common : improve yaml log escaping (ggerganov#4080)

* logging: improve escaping in yaml output

* logging: include review feedback

py : Falcon HF compatibility (ggerganov#4104)

Falcon HF compatibility

convert : use 'model' value if it exists. This allows karpathy/tinyllamas to load (ggerganov#4089)

Co-authored-by: Don Mahurin <@>

examples : add tokenize (ggerganov#4039)

tokenize : fix trailing whitespace

build : support ppc64le build for make and CMake (ggerganov#3963)

* build: support ppc64le build for make and CMake

* build: keep __POWER9_VECTOR__ ifdef and extend with __powerpc64__

Co-authored-by: Georgi Gerganov <[email protected]>

---------

Co-authored-by: Georgi Gerganov <[email protected]>

llama : increase max nodes (ggerganov#4115)

Clean up ggml-cuda.cu warnings when compiling with clang (for ROCM) (ggerganov#4124)

* ggml-cuda.cu: Clean up warnings when compiling with clang

* ggml-cuda.cu: Move static items into anonymous namespace

* ggml-cuda.cu: Fix use of namespace start macro

* Revert "ggml-cuda.cu: Fix use of namespace start macro"

This reverts commit 26c1149.

* Revert "ggml-cuda.cu: Move static items into anonymous namespace"

This reverts commit e29757e.

scripts : Remove missed baichuan convert script (ggerganov#4127)

tokenize example: Respect normal add BOS token behavior (ggerganov#4126)

Allow building with Makefile

gguf-py : export chat templates (ggerganov#4125)

* gguf-py : export chat templates

* llama.cpp : escape new lines in gguf kv info prints

* gguf-py : bump version

* gguf-py : check chat_template type

* gguf-py : initialize chat_template

gitignore : tokenize

common : comma should be semicolon (ggerganov#4137)

server : relay error messages (ggerganov#4131)

finetune : add --n-gpu-layers flag info to --help (ggerganov#4128)

Revert "finetune : add --n-gpu-layers flag info to --help (ggerganov#4128)"

This reverts commit 05e8301.

speculative : fix prompt tokenization in speculative example (ggerganov#4025)

* Support special tokens and not adding BOS to prompt in speculative

* Adapt to new should_add_bos function

* Ensure tgt and dft have same add_bos setting

ci : add flake8 to github actions (python linting) (ggerganov#4129)

Disabled rules:

* E203 Whitespace before ':' - disabled because we often use 'C' Style where values are aligned

* E211 Whitespace before '(' (E211) - disabled because we often use 'C' Style where values are aligned

* E221 Multiple spaces before operator - disabled because we often use 'C' Style where values are aligned

* E225 Missing whitespace around operator - disabled because it's broken so often it seems like a standard

* E231 Missing whitespace after ',', ';', or ':' - disabled because we often use 'C' Style where values are aligned

* E241 Multiple spaces after ',' - disabled because we often use 'C' Style where values are aligned

* E251 Unexpected spaces around keyword / parameter equals - disabled because it's broken so often it seems like a standard

* E261 At least two spaces before inline comment - disabled because it's broken so often it seems like a standard

* E266 Too many leading '#' for block comment - sometimes used as "section" separator

* E501 Line too long - disabled because it's broken so often it seems like a standard

* E701 Multiple statements on one line (colon) - broken only in convert.py when defining abstract methods (we can use# noqa instead)

* E704 Multiple statements on one line - broken only in convert.py when defining abstract methods (we can use# noqa instead)

main : Add ChatML functionality to main example (ggerganov#4046)

Co-authored-by: Sebastian Cramond <[email protected]>

readme : update ROCm Windows instructions (ggerganov#4122)

* Update README.md

* Update README.md

Co-authored-by: Jared Van Bortel <[email protected]>

---------

Co-authored-by: Jared Van Bortel <[email protected]>

finetune - update readme to mention llama support only (ggerganov#4148)

stablelm : simplify + speedup generation (ggerganov#4153)

docs : add llama-star arch idea

examples : fix typo in parallel example doc comment (ggerganov#4181)

Signed-off-by: Daniel Bevenius <[email protected]>

readme : update hot topics

llama : KV cache view API + better KV cache management (ggerganov#4170)

* llama : keep track of used KV cells + better KV cache management

* llama : zero KV cache used upon clear

ggml-ci

* llama : allow exporting a view of the KV cache (ggerganov#4180)

* Allow exporting a view of the KV cache

* Allow dumping the sequences per cell in common

* Track max contiguous cells value and position as well

* Fix max contiguous empty cells index calculation

Make dump functions deal with lengths or sequences counts > 10 better

* Fix off by one error in dump_kv_cache_view

* Add doc comments for KV cache view functions

Eliminate cell sequence struct; use llama_seq_id directly

Minor cleanups

* common : add -dkvc arg for enabling kv cache dumps

---------

Co-authored-by: Kerfuffle <[email protected]>

Fix incorrect format strings and uninitialized variables. (ggerganov#4133)

* Fix incorrect format strings and uninitialized variables.

* Address comments

* Add the missing include statement

readme : use PATH for Windows ROCm (ggerganov#4195)

* Update README.md to use PATH for Windows ROCm

* Update README.md

* Update README.md

main.swift : fix eos checking (ggerganov#4197)

llama_token_eos(const struct llama_model *) is currently getting struct llama_context type variable context as a parameter.

convert : fix tensors using grad in some models (ggerganov#4173)

ggml-cuda : support stablelm rope (ggerganov#4156)

* ggml-cuda : support stablelm rope

* remove unused freq_base kernel parameter

* add n_dims parameter to llm_build_k_shift, default to n_rot via overload

* llama : fix llm_build_k_shift args

---------

Co-authored-by: Georgi Gerganov <[email protected]>

llama : set metal log callback correctly (ggerganov#4204)

server : OAI API compatibility (ggerganov#4198)

* Add openai-compatible POST /v1/chat/completions API endpoint to server example

* fix code style

* Update server README.md

* Improve server README.md

* Fix server.cpp code style according to review

* server : some style changes

* server : indentation

* server : enable special tokens during tokenization by default

* server : minor code style

* server : change random string generator

* straightforward /v1/models endpoint

---------

Co-authored-by: kir-gadjello <[email protected]>
Co-authored-by: Tobi Lütke <[email protected]>

readme : update hot topics

Update docs for yarn_ext_factor <0.0 as unspecified instead of NaN (ggerganov#4189)

llama : grammar `reserve` space in `decode_utf8` (ggerganov#4210)

* reserve space for codepoints

* improvement for the appended 0

scripts : Use mmap in torch load (ggerganov#4202)

* Use mmap in torch load, prefer .bin files when loading

* Revert .bin > .safetensors preference

metal : fix yarn (ggerganov#4220)

get the correct n_orig_ctx in metal

lookahead : add example for lookahead decoding (ggerganov#4207)

* lookahead : init

* lookahead : generate and store n-grams

* lookahead : use loop instead recursion to generate n-grams

* lookahead : initial working implementation

* lookahead : filter repeating n-grams

* lookahead : use deterministic init

* lookahead : add to Makefile

* lookahead : fix a bug in the seq_id of the lookahead tokens

* lookahead : add comments

---------

Co-authored-by: slaren <[email protected]>

readme : update hot topics

lookahead : support `-n -1` infinite generation

ggml : fix -Warray-bounds warning with gcc (ggerganov#4231)

examples : iOS example with swift ui (ggerganov#4159)

* copy to llama.cpp as subdir

* attempt enabling metal, fails

* ggml metal compiles!

* Update README.md

* initial conversion to new format, utf8 errors?

* bug fixes, but now has an invalid memory access :(

* added O3, now has insufficient memory access

* begin sync with master

* update to match latest code, new errors

* fixed it!

* fix for loop conditionals, increase result size

* fix current workflow errors

* attempt a llama.swiftui workflow

* Update .github/workflows/build.yml

Co-authored-by: Georgi Gerganov <[email protected]>

---------

Co-authored-by: Georgi Gerganov <[email protected]>

readme : add Amica to UI list (ggerganov#4230)

cmake : fix issue with version info not getting baked into LlamaConfig.cmake (ggerganov#3970)

* Split CPP generation from build-info query

* Remove blank lines

* Add BUILD_SHARED_LIBS option

ggml : re-enable BLAS for CPU when src0 != F32 + remove redundant full offload checks in llama.cpp (ggerganov#4240)

* ggml : use blas even if src0 is not F32

* llama : use n_threads_batch only when n_tokens >= 32

ggml-ci

* llama : revert n_threads_batch logic

ggml-ci

ggml : restore abort() in GGML_ASSERT (ggerganov#4242)

readme : add FreeChat (ggerganov#4248)

examples : add readme files

py : fix oai proxy (ggerganov#3972)

* fix oai proxy

fix generation not stoped while bot stop talking in chat mode

fix possible `slot_id` not exist

response for cors (and pre flight)

* oai proxy: workaround for some client (such as Chatbox)

* use stop as separator to replace hardcoded `\n`

llama : fix typical sampling (ggerganov#4261)

Typical sampling was broken because after copying new_candidates into canditates, the "sorted" bool is left at "true", but the new data is no longer sorted according to probability. Patch to set "sorted" to false.

Test: Generating with temp=0.0001 (approx. argmax)  should generate the same sequence at typical>=1.0 and typical=0.9999 (approx. disabled, but enters the typical sampling codepath).

convert.py : fix llama/llama2 conversion due to vocab_size=-1 (ggerganov#4258)

llama : fix alignment of general.name in print meta (ggerganov#4254)

* llama: fix alignment of general.name in print meta

This commit fixes the alignment of the general.name field in the
llm_load_print_meta function.

Currently the output looks like this:
```console
llm_load_print_meta: model ftype      = mostly Q4_0
llm_load_print_meta: model params     = 13.02 B
llm_load_print_meta: model size       = 6.86 GiB (4.53 BPW)
llm_load_print_meta: general.name   = LLaMA v2
```
And with this commit it looks like this:
```console
llm_load_print_meta: model ftype      = mostly Q4_0
llm_load_print_meta: model params     = 13.02 B
llm_load_print_meta: model size       = 6.86 GiB (4.53 BPW)
llm_load_print_meta: general.name     = LLaMA v2
```

Signed-off-by: Daniel Bevenius <[email protected]>

* llama: fix alignment of special tokens

Signed-off-by: Daniel Bevenius <[email protected]>

---------

Signed-off-by: Daniel Bevenius <[email protected]>

readme : fix typo (ggerganov#4253)

llama.cpp uses GitHub Actions, not Gitlab Actions.

cmake : fix the metal file foder path (ggerganov#4217)

batched.swift : update README.md (ggerganov#4214)

docs: update how to run

docker : add finetune option (ggerganov#4211)

readme : fix (ggerganov#4135)

* fix: readme

* chore: resolve comments

* chore: resolve comments

main : pass LOG_TEE callback to llama.cpp log (ggerganov#4033)

* main : Call llama_log_set to use LOG_TEE

* tabs to spaces

llava : ShareGPT4V compatibility (vision encoder only loading) (ggerganov#4172)

* ShareGPT4 compatibility (vision encoder only loading)

Load only a CLIP vision encoder (as supplied by ShareGPT finetunes)
Corrects the argument parsing for --img_mean and --img_std (which were previously not parsed but attempted to access)
Defines defaults for img_mean and img_std which are equal to the llava 1.5 CLIP encoder, so you do not have to provide them

* Update convert-image-encoder-to-gguf.py

build : fix build info generation and cleanup Makefile (ggerganov#3920)

* cmake : fix joining of REAL_GIT_DIR

* fix includes with help from include-what-you-use

* make : remove unneeded deps and add test-rope target

* fix C includes in C++ source files

* Revert "fix includes with help from include-what-you-use"

This reverts commit 635e9fa.

make : fix Apple clang determination bug (ggerganov#4272)

Co-authored-by: Will Findley <[email protected]>

server : add single-client multi-prompt support (ggerganov#4232)

* * add multiprompt support

* * cleanup

* * more cleanup

* * remove atomicity of id_gen, and change lock_guard to unique_lock on completion requests

* * remove all references to mutex_multitasks

* Update examples/server/server.cpp

Co-authored-by: Jared Van Bortel <[email protected]>

* Update examples/server/server.cpp

Co-authored-by: Jared Van Bortel <[email protected]>

* Update examples/server/server.cpp

Co-authored-by: Jared Van Bortel <[email protected]>

* Update examples/server/server.cpp

Co-authored-by: Jared Van Bortel <[email protected]>

* * change to set

---------

Co-authored-by: Jared Van Bortel <[email protected]>

server : add --log-disable to disable logging to file (ggerganov#4260)

* * add --log-disable to disable logging to file in the server example

* * typo fix

ggml : add ggml_soft_max_ext (ggerganov#4256)

* metal : implement soft_max_ext

* cuda : implement soft_max_ext

* ggml : implement soft_max_ext (CPU)

* batched-bench : print threads

ggml-ci

* metal : simplify soft_max encoding

ggml-ci

* cuda : use 512 threads for soft_max instead of 32

* ggml : update soft max cpu

* cuda : do warp-based block reduce

* cuda : increase max block size to 1024

* cuda : fix warp reduction initialization of shared mem

* metal : warp-based reduction for soft max kernel

* metal : warp-based reduce for rms_norm

* metal : simplify soft max kernel

ggml-ci

* alloc : fix build with debug

py : add requirements file for convert-hf-to-gguf.py (ggerganov#4277)

This commit adds a requirements file for the convert-hf-to-gguf.py
script, and also add the torch and transformers packages to it.

The motivation for this is that currently running convert-hf-to-gguf.py
will produce the following error:
```console
$ python3 -m venv venv
$ source venv/bin/activate
(venv) $ pip install -r requirements.txt
Collecting numpy==1.24.4
Collecting sentencepiece==0.1.98
Collecting gguf>=0.1.0
Installing collected packages: sentencepiece, numpy, gguf
Successfully installed gguf-0.5.1 numpy-1.24.4 sentencepiece-0.1.98

(venv) $ python convert-hf-to-gguf.py --help
Traceback (most recent call last):
  File "llama.cpp/convert-hf-to-gguf.py", line 16, in <module>
    import torch
ModuleNotFoundError: No module named 'torch'
```
With this commit, and using requirements-hf-to-gguf.txt instead of
requirements.txt, the script can be run and shows the help output.

Signed-off-by: Daniel Bevenius <[email protected]>

llama : fix integer overflow during quantization (ggerganov#4284)

happens with multi-threaded quantization of Qwen-72B

ggml-ci

llama : add Qwen support (ggerganov#4281)

* enable qwen to llama.cpp

* llama : do not GPU split bias tensors

---------

Co-authored-by: Georgi Gerganov <[email protected]>

llama : support attention bias on LLaMA architecture (ggerganov#4283)

* Support attention_bias on LLaMA architecture

QKVO bias, should fix InternLM (ggerganov#3133) and works for LLaMAfied Qwen models (ggerganov#3743 (comment)).

* check existence of qkvo bias while loading llama models

Tested on LLaMA2, CUDA and CPU.

* Update llama.cpp

build : enable libstdc++ assertions for debug builds (ggerganov#4275)

swift : fix token_to_piece implementation (ggerganov#4278)

* Fix token_to_piece implementation in Swift

* Fix errors

llama : support optional tensors (ggerganov#4283)

llama : avoid using "optional" keyword (ggerganov#4283)

llama : pad KV cache size (ggerganov#4280)

* llama : pad KV cache size to 32

* metal : try to improve batched decoding

py : add grammar to oai like api (ggerganov#4294)

server : fix OpenAI API `stop` field to be optional (ggerganov#4299)

(cherry picked from commit Mozilla-Ocho/llamafile@e8c92bc)

ggml : fix soft max out-of-bounds access (ggerganov#4307)

ggml-ci

ggml : reuse ggml_get_n_tasks() in ggml_graph_plan() (ggerganov#4308)

* ggml : fix soft max out-of-bounds access

ggml-ci

* ggml : reuse ggml_get_n_tasks() in ggml_graph_plan()

ggml-ci

grammar-parser : fix typo (ggerganov#4318)

preceeding -> preceding

swift : fix prompt tokenization logic (ggerganov#4321)

swift : fix concatenation method to avoid invalid UTF8 stringfication (ggerganov#4325)

simple : update error message for KV cache check (ggerganov#4324)

This commit updates the error message that is printed when the
KV cache is not big enough to hold all the prompt and generated
tokens. Specifically it removes the reference to n_parallel and
replaces it with n_len.

Signed-off-by: Daniel Bevenius <[email protected]>

swift : revert compiler checks for swift package (ggerganov#4332)

sampling : custom samplers order (ggerganov#4285)

* Samplers sequence order w parameter

* Cleaned commented code

* Fixed formatting

* Rewrote with unordered_map

* Revert and rewrite, too many problems and safeguards would be needed

* Fixed code style

* Code style fixes according to review

* More readable samplers input string, fixed help

* Style fix in sampler_queue

* Formatting fixes

* Fixing whitespaces

llama : allow overriding GGUF metadata when loading model (ggerganov#4092)

* feat: Allow overriding GGUF metadata when loading model

* Fix the one time GCC is stricter than clang about something

* Step1

* Refactor... basically everything!

* Nuke obsolete GetArrayLen struct

* simplify std::string specialization

* Various cleanups

Add informational output when overrides are applied

Warn user when an override with the wrong type is specified

* Fix broken logic for parsing bool KV overrides
Fix issue where overrides didn't apply when key missing in GGUF metadata
Resolve merge changes

* llama : rearrange model params

* Update new GET_KEY call

Add note that metadata KV overrides aren't reflected in initial metadata KV info dump

---------

Co-authored-by: cebtenzzre <[email protected]>
Co-authored-by: Georgi Gerganov <[email protected]>

grammar : pre-computed pieces + reserve mem + less string copies (ggerganov#4330)

* reserve space for codepoints

* improvement for the appended 0

* used precomputed token text for grammar sample

* reserve canidates_decoded

* reserve canidates_grammar

* remove candidates_decoded

* Revert "remove candidates_decoded"

This reverts commit 3773328.

* changed decode_utf8 to take src by ref

speculative : support `--color` (ggerganov#4343)

* speculative: add some colors

* minor : add braces

---------

Co-authored-by: Georgi Gerganov <[email protected]>

common : fix compile warning

server : recognize cache_prompt parameter in OAI API (ggerganov#4347)

train : fix ggerganov#4227 (double free in examples/train-text-from-scratch/train-text-from-scratch.cpp) (ggerganov#4351)

On commit b1108 (44c117f) xaedes added

    ggml_allocr * alloc = NULL;

    ... (many lines in between)

    if (alloc) {
        ggml_allocr_free(alloc);
    }

Which is correct, but it's easy to lose context after many lines in between.

On commit b1287 (0e76a899) xaedes made a big change. From here on, alloc is freed eagerly.

    alloc = ggml_allocr_new(...)
    ... (short lines of code)
    ggml_allocr_free(alloc)

This happens a few times, but alloc is never set to NULL, and many lines below,
we still have

    if (alloc) {
        ggml_allocr_free(alloc);
    }

which causes a double-free.

llama : per-layer KV cache + quantum K cache (ggerganov#4309)

* per-layer KV

* remove unnecessary copies

* less code duplication, offload k and v separately

* llama : offload KV cache per-layer

* llama : offload K shift tensors

* llama : offload for rest of the model arches

* llama : enable offload debug temporarily

* llama : keep the KV related layers on the device

* llama : remove mirrors, perform Device -> Host when partial offload

* common : add command-line arg to disable KV cache offloading

* llama : update session save/load

* llama : support quantum K cache (ggerganov#4312)

* llama : support quantum K cache (wip)

* metal : add F32 -> Q8_0 copy kernel

* cuda : add F32 -> Q8_0 copy kernel

ggml-ci

* cuda : use mmv kernel for quantum cache ops

* llama : pass KV cache type through API

* llama : fix build

ggml-ci

* metal : add F32 -> Q4_0 copy kernel

* metal : add F32 -> Q4_1 copy kernel

* cuda : wip

* cuda : add F32 -> Q4_0 and F32 -> Q4_1 copy kernels

* llama-bench : support type_k/type_v

* metal : use mm kernel only for quantum KV cache

* cuda : add comment

* llama : remove memory_f16 and kv_f16 flags

---------

Co-authored-by: slaren <[email protected]>

* readme : add API change notice

---------

Co-authored-by: slaren <[email protected]>

sync : ggml (new ops, tests, backend, etc.) (ggerganov#4359)

* sync : ggml (part 1)

* sync : ggml (part 2, CUDA)

* sync : ggml (part 3, Metal)

* ggml : build fixes

ggml-ci

* cuda : restore lost changes

* cuda : restore lost changes (StableLM rope)

* cmake : enable separable compilation for CUDA

ggml-ci

* ggml-cuda : remove device side dequantize

* Revert "cmake : enable separable compilation for CUDA"

This reverts commit 09e35d0.

* cuda : remove assert for rope

* tests : add test-backend-ops

* ggml : fix bug in ggml_concat

* ggml : restore `ggml_get_n_tasks()` logic in `ggml_graph_plan()`

* ci : try to fix macOS

* ggml-backend : remove backend self-registration

* ci : disable Metal for macOS cmake build

ggml-ci

* metal : fix "supports family" call

* metal : fix assert

* metal : print resource path

ggml-ci

---------

Co-authored-by: slaren <[email protected]>

grammar : revert the replacement of llama_token_to_piece with id_to_token (ggerganov#4396)

Update README.md (ggerganov#4388)

Fix small typo.

ggml : increased GGML_MAX_PARAMS to allow finetuning of 70b models (ggerganov#4424)

server : fix local model name in server (ggerganov#4420)

llama : document logits_all deprecation (ggerganov#4418)

llama_context_params.logits_all is a parameter for controlling
llama_eval. This documents that logits_all should not be used with
llama_decode and llama_batch.

build : target Windows 8 for standard mingw-w64 (ggerganov#4405)

* build : target Windows 8 for standard mingw-w64

* make : fix missing console.o deps

This was causing a link error with `make all` on Windows.

english : use `typos` to fix comments and logs (ggerganov#4354)

server : tweak default sampling parameters (ggerganov#4367)

* Set a more typical Top P setting as the default

* Update temp max

llama : add Mixtral support (ggerganov#4406)

* convert : support Mixtral as LLAMA arch

* convert : fix n_ff typo

* llama : model loading

* ggml : sync latest ggml_mul_mat_id

* llama : update graph to support MoE

* llama : fix cur -> cur_expert

* llama : first working version

* llama : fix expert weighting in the FFN

* ggml : ggml_get_rows support 2D indexing [n_tokens, n_experts] (cpu only)

* ggml : add n_as argument to ggml_mul_mat_id

* ggml : fix ggml_get_rows to take into account ne02 / ne11

* metal : add more general support for ggml_get_rows + tests

* llama : add basic support for offloading moe with CUDA

* metal : add/mul/div use general kernel when src1 not cont

* metal : reduce the kernel launches for ggml_mul_mat_id

* ggml : get_rows : support non-contiguos tensors with gaps, generalize up to 3D

* ggml : update get_rows f16 and q

* cuda : support non-contiguous src1 in get_rows

* llama : offload missing ffn_moe_silu

* metal : fix ggml_get_rows to work with non-cont src1

* metal : add indirect mat-vec kernels for all quantization types

* llama : do not quantize expert gating tensors

* llama : add n_expert and n_expert_used to hparams + change quants

* test-backend-ops : add moe test

* cuda : fix get_rows when ncols is odd

* convert : determine n_ctx correctly

* metal : fix ggml_mul_mat_id for F32

* test-backend-ops : make experts more evenly probable (test_moe)

* test-backend-ops : cleanup, add moe test for batches

* test-backend-ops : add cpy from f32 -> all types test

* test-backend-ops : fix dequantize block offset

* llama : fix hard-coded number of experts

* test-backend-ops : simplify and disable slow tests to avoid CI timeout

* test-backend-ops : disable MOE test with thread sanitizer

* cuda : fix mul_mat_id with multi gpu

* convert : use 1e6 rope_freq_base for mixtral

* convert : fix style

* convert : support safetensors format

* gguf-py : bump version

* metal : add cpy f16 -> f32 kernel

* metal : fix binary ops for ne10 % 4 != 0

* test-backend-ops : add one more sum_rows test

* ggml : do not use BLAS with ggml_mul_mat_id

* convert-hf : support for mixtral-instruct (ggerganov#4428)

* convert : typo fix, add additional hyperparameters, use LLaMA arch for Mixtral-instruct

* convert : use sentencepiece tokenizer for Mixtral-instruct

* convert : make flake8 happy

* metal : fix soft_max kernels

ref: ggerganov/ggml@1914017

* metal : limit kernels to not use more than the allowed threads

---------

Co-authored-by: Georgi Gerganov <[email protected]>
Co-authored-by: Radek Pilar <[email protected]>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

Successfully merging this pull request may close these issues.

3 participants