diff --git a/.devops/full-cuda.Dockerfile b/.devops/full-cuda.Dockerfile index 059fd2695..c01006efe 100644 --- a/.devops/full-cuda.Dockerfile +++ b/.devops/full-cuda.Dockerfile @@ -31,6 +31,6 @@ ENV LLAMA_CUDA=1 # Enable cURL ENV LLAMA_CURL=1 -RUN make +RUN make -j$(nproc) ENTRYPOINT ["/app/.devops/tools.sh"] diff --git a/.devops/full-rocm.Dockerfile b/.devops/full-rocm.Dockerfile index 6ecf3bcc7..0314d469b 100644 --- a/.devops/full-rocm.Dockerfile +++ b/.devops/full-rocm.Dockerfile @@ -45,6 +45,6 @@ ENV LLAMA_CURL=1 RUN apt-get update && \ apt-get install -y libcurl4-openssl-dev -RUN make +RUN make -j$(nproc) ENTRYPOINT ["/app/.devops/tools.sh"] diff --git a/.devops/full.Dockerfile b/.devops/full.Dockerfile index 432fb5dad..6d5943a2f 100644 --- a/.devops/full.Dockerfile +++ b/.devops/full.Dockerfile @@ -18,7 +18,7 @@ COPY . . ENV LLAMA_CURL=1 -RUN make +RUN make -j$(nproc) ENV LC_ALL=C.utf8 diff --git a/.devops/main-cuda.Dockerfile b/.devops/main-cuda.Dockerfile index b937a4829..23f428944 100644 --- a/.devops/main-cuda.Dockerfile +++ b/.devops/main-cuda.Dockerfile @@ -23,7 +23,7 @@ ENV CUDA_DOCKER_ARCH=${CUDA_DOCKER_ARCH} # Enable CUDA ENV LLAMA_CUDA=1 -RUN make +RUN make -j$(nproc) FROM ${BASE_CUDA_RUN_CONTAINER} as runtime diff --git a/.devops/main-rocm.Dockerfile b/.devops/main-rocm.Dockerfile index 0a706dc73..37576d68e 100644 --- a/.devops/main-rocm.Dockerfile +++ b/.devops/main-rocm.Dockerfile @@ -40,6 +40,6 @@ ENV LLAMA_HIPBLAS=1 ENV CC=/opt/rocm/llvm/bin/clang ENV CXX=/opt/rocm/llvm/bin/clang++ -RUN make +RUN make -j$(nproc) ENTRYPOINT [ "/app/main" ] diff --git a/.devops/main.Dockerfile b/.devops/main.Dockerfile index 3ab1decd6..763d75fce 100644 --- a/.devops/main.Dockerfile +++ b/.devops/main.Dockerfile @@ -9,7 +9,7 @@ WORKDIR /app COPY . . -RUN make +RUN make -j$(nproc) FROM ubuntu:$UBUNTU_VERSION as runtime diff --git a/.devops/server-cuda.Dockerfile b/.devops/server-cuda.Dockerfile index 59a52ba21..7f5228185 100644 --- a/.devops/server-cuda.Dockerfile +++ b/.devops/server-cuda.Dockerfile @@ -25,7 +25,7 @@ ENV LLAMA_CUDA=1 # Enable cURL ENV LLAMA_CURL=1 -RUN make +RUN make -j$(nproc) FROM ${BASE_CUDA_RUN_CONTAINER} as runtime diff --git a/.devops/server-rocm.Dockerfile b/.devops/server-rocm.Dockerfile index c02a31dd8..a6b76dee8 100644 --- a/.devops/server-rocm.Dockerfile +++ b/.devops/server-rocm.Dockerfile @@ -45,6 +45,6 @@ ENV LLAMA_CURL=1 RUN apt-get update && \ apt-get install -y libcurl4-openssl-dev -RUN make +RUN make -j$(nproc) ENTRYPOINT [ "/app/server" ] diff --git a/.devops/server.Dockerfile b/.devops/server.Dockerfile index be964e0e8..0d09d3627 100644 --- a/.devops/server.Dockerfile +++ b/.devops/server.Dockerfile @@ -11,7 +11,7 @@ COPY . . ENV LLAMA_CURL=1 -RUN make +RUN make -j$(nproc) FROM ubuntu:$UBUNTU_VERSION as runtime diff --git a/.devops/tools.sh b/.devops/tools.sh index 3a7d274e4..97424c3aa 100755 --- a/.devops/tools.sh +++ b/.devops/tools.sh @@ -8,7 +8,7 @@ arg1="$1" shift if [[ "$arg1" == '--convert' || "$arg1" == '-c' ]]; then - python3 ./convert.py "$@" + python3 ./convert-hf-to-gguf.py "$@" elif [[ "$arg1" == '--quantize' || "$arg1" == '-q' ]]; then ./quantize "$@" elif [[ "$arg1" == '--run' || "$arg1" == '-r' ]]; then diff --git a/.github/ISSUE_TEMPLATE/06-question.yml b/.github/ISSUE_TEMPLATE/06-question.yml deleted file mode 100644 index 9d3ff4972..000000000 --- a/.github/ISSUE_TEMPLATE/06-question.yml +++ /dev/null @@ -1,38 +0,0 @@ -name: Question -description: Used to ask questions about llama.cpp -title: "Question: " -labels: ["question"] -body: - - type: markdown - attributes: - value: | - [Please search your question first in Discussion if you got a common general question.](https://github.com/ggerganov/llama.cpp/discussions/categories/q-a) - - - type: checkboxes - id: prerequisites - attributes: - label: Prerequisites - description: Please confirm the following before submitting your question. - options: - - label: I searched using keywords relevant to my issue to make sure that I am creating a new issue that is not already open (or closed). - required: true - - label: I reviewed the [Discussions](https://github.com/ggerganov/llama.cpp/discussions), and have a new useful question to share that cannot be answered within Discussions. - required: true - - - type: textarea - id: background-description - attributes: - label: Background Description - description: Please provide a detailed written description of what you were trying to do, and what you expected `llama.cpp` to do as an question. - placeholder: Detailed description of your question - validations: - required: true - - - type: textarea - id: possible-answer - attributes: - label: Possible Answer - description: If you have some idea of possible answers you want to confirm, that would also be appreciated. - placeholder: Your idea of possible answers - validations: - required: false diff --git a/.github/ISSUE_TEMPLATE/06-research.yml b/.github/ISSUE_TEMPLATE/06-research.yml new file mode 100644 index 000000000..3ae4e9f8c --- /dev/null +++ b/.github/ISSUE_TEMPLATE/06-research.yml @@ -0,0 +1,52 @@ +name: Research +description: Track new technical research area +title: "Research: " +labels: ["research 🔬"] +body: + - type: markdown + attributes: + value: | + Don't forget to check for any [duplicate research issue tickets](https://github.com/ggerganov/llama.cpp/issues?q=is%3Aopen+is%3Aissue+label%3A%22research+%F0%9F%94%AC%22) + + - type: checkboxes + id: research-stage + attributes: + label: Research Stage + description: Track general state of this research ticket + options: + - label: Background Research (Let's try to avoid reinventing the wheel) + - label: Hypothesis Formed (How do you think this will work and it's effect?) + - label: Strategy / Implementation Forming + - label: Analysis of results + - label: Debrief / Documentation (So people in the future can learn from us) + + - type: textarea + id: background + attributes: + label: Previous existing literature and research + description: Whats the current state of the art and whats the motivation for this research? + + - type: textarea + id: hypothesis + attributes: + label: Hypothesis + description: How do you think this will work and it's effect? + + - type: textarea + id: implementation + attributes: + label: Implementation + description: Got an approach? e.g. a PR ready to go? + + - type: textarea + id: analysis + attributes: + label: Analysis + description: How does the proposed implementation behave? + + - type: textarea + id: logs + attributes: + label: Relevant log output + description: Please copy and paste any relevant log output. This will be automatically formatted into code, so no need for backticks. + render: shell diff --git a/.github/ISSUE_TEMPLATE/config.yml b/.github/ISSUE_TEMPLATE/config.yml new file mode 100644 index 000000000..c88134dbb --- /dev/null +++ b/.github/ISSUE_TEMPLATE/config.yml @@ -0,0 +1,13 @@ +blank_issues_enabled: true +contact_links: + - name: Got an idea? + url: https://github.com/ggerganov/llama.cpp/discussions/categories/ideas + about: Pop it there. It may then become an enhancement ticket. + - name: Got a question? + url: https://github.com/ggerganov/llama.cpp/discussions/categories/q-a + about: Ask a question there! + - name: Want to contribute? + url: https://github.com/ggerganov/llama.cpp/wiki/contribute + about: Head to the contribution guide page of the wiki for areas you can help with + + diff --git a/CMakeLists.txt b/CMakeLists.txt index fbbc38644..60cf7bdc4 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -1314,7 +1314,7 @@ set_target_properties(llama PROPERTIES PUBLIC_HEADER ${CMAKE_CURRENT_SOURCE_DIR} install(TARGETS llama LIBRARY PUBLIC_HEADER) install( - FILES convert.py + FILES convert-hf-to-gguf.py PERMISSIONS OWNER_READ OWNER_WRITE diff --git a/README.md b/README.md index 1cab7f19d..e8bf1e246 100644 --- a/README.md +++ b/README.md @@ -2,7 +2,9 @@ ![llama](https://user-images.githubusercontent.com/1991296/230134379-7181e485-c521-4d23-a0d6-f7b3b61ba524.png) -[![License: MIT](https://img.shields.io/badge/license-MIT-blue.svg)](https://opensource.org/licenses/MIT) [![Server](https://github.com/ggerganov/llama.cpp/actions/workflows/server.yml/badge.svg?branch=master&event=schedule)](https://github.com/ggerganov/llama.cpp/actions/workflows/server.yml) +[![License: MIT](https://img.shields.io/badge/license-MIT-blue.svg)](https://opensource.org/licenses/MIT) +[![Server](https://github.com/ggerganov/llama.cpp/actions/workflows/server.yml/badge.svg?branch=master&event=schedule)](https://github.com/ggerganov/llama.cpp/actions/workflows/server.yml) +[![Conan Center](https://shields.io/conan/v/llama-cpp)](https://conan.io/center/llama-cpp) [Roadmap](https://github.com/users/ggerganov/projects/7) / [Project status](https://github.com/ggerganov/llama.cpp/discussions/3471) / [Manifesto](https://github.com/ggerganov/llama.cpp/discussions/205) / [ggml](https://github.com/ggerganov/ggml) @@ -20,7 +22,8 @@ Inference of Meta's [LLaMA](https://arxiv.org/abs/2302.13971) model (and others) ### Hot topics -- **Initial Flash-Attention support: https://github.com/ggerganov/llama.cpp/pull/5021** +- **`convert.py` has been deprecated and moved to `examples/convert-legacy-llama.py`, please use `convert-hf-to-gguf.py`** https://github.com/ggerganov/llama.cpp/pull/7430 +- Initial Flash-Attention support: https://github.com/ggerganov/llama.cpp/pull/5021 - BPE pre-tokenization support has been added: https://github.com/ggerganov/llama.cpp/pull/6920 - MoE memory layout has been updated - reconvert models for `mmap` support and regenerate `imatrix` https://github.com/ggerganov/llama.cpp/pull/6387 - Model sharding instructions using `gguf-split` https://github.com/ggerganov/llama.cpp/discussions/6404 @@ -200,6 +203,7 @@ Unless otherwise noted these projects are open-source with permissive licensing: - [KodiBot](https://github.com/firatkiral/kodibot) (GPL) - [eva](https://github.com/ylsdamxssjxxdd/eva) (MIT) - [AI Sublime Text plugin](https://github.com/yaroslavyaroslav/OpenAI-sublime-text) (MIT) +- [AIKit](https://github.com/sozercan/aikit) (MIT) *(to have a project listed here, it should clearly state that it depends on `llama.cpp`)* @@ -315,8 +319,6 @@ In order to build llama.cpp you have four different options. make ``` - **Note**: for `Debug` builds, run `make LLAMA_DEBUG=1` - - On Windows: 1. Download the latest fortran version of [w64devkit](https://github.com/skeeto/w64devkit/releases). @@ -328,23 +330,32 @@ In order to build llama.cpp you have four different options. make ``` + - Notes: + - For faster compilation, add the `-j` argument to run multiple jobs in parallel. For example, `make -j 8` will run 8 jobs in parallel. + - For faster repeated compilation, install [ccache](https://ccache.dev/). + - For debug builds, run `make LLAMA_DEBUG=1` + - Using `CMake`: - ```bash - cmake -B build - cmake --build build --config Release - ``` + ```bash + cmake -B build + cmake --build build --config Release + ``` - **Note**: for `Debug` builds, there are two cases: + **Notes**: - - Single-config generators (e.g. default = `Unix Makefiles`; note that they just ignore the `--config` flag): + - For faster compilation, add the `-j` argument to run multiple jobs in parallel. For example, `cmake --build build --config Release -j 8` will run 8 jobs in parallel. + - For faster repeated compilation, install [ccache](https://ccache.dev/). + - For debug builds, there are two cases: + + 1. Single-config generators (e.g. default = `Unix Makefiles`; note that they just ignore the `--config` flag): ```bash cmake -B build -DCMAKE_BUILD_TYPE=Debug cmake --build build ``` - - Multi-config generators (`-G` param set to Visual Studio, XCode...): + 2. Multi-config generators (`-G` param set to Visual Studio, XCode...): ```bash cmake -B build -G "Xcode" @@ -379,6 +390,14 @@ In order to build llama.cpp you have four different options. CLBLAST support for use OpenCL GPU acceleration in FreeBSD. Please read the instructions for use and activate this options in this document below. +### Homebrew + +On Mac and Linux, the homebrew package manager can be used via +``` +brew install llama.cpp +``` +The formula is automatically updated with new `llama.cpp` releases. More info: https://github.com/ggerganov/llama.cpp/discussions/7668 + ### Metal Build On MacOS, Metal is enabled by default. Using Metal makes the computation run on the GPU. @@ -697,7 +716,8 @@ Building the program with BLAS support may lead to some performance improvements To obtain the official LLaMA 2 weights please see the Obtaining and using the Facebook LLaMA 2 model section. There is also a large selection of pre-quantized `gguf` models available on Hugging Face. -Note: `convert.py` does not support LLaMA 3, you can use `convert-hf-to-gguf.py` with LLaMA 3 downloaded from Hugging Face. +Note: `convert.py` has been moved to `examples/convert-legacy-llama.py` and shouldn't be used for anything other than `Llama/Llama2/Mistral` models and their derievatives. +It does not support LLaMA 3, you can use `convert-hf-to-gguf.py` with LLaMA 3 downloaded from Hugging Face. ```bash # obtain the official LLaMA model weights and place them in ./models @@ -714,10 +734,10 @@ ls ./models python3 -m pip install -r requirements.txt # convert the model to ggml FP16 format -python3 convert.py models/mymodel/ +python3 convert-hf-to-gguf.py models/mymodel/ # [Optional] for models using BPE tokenizers -python convert.py models/mymodel/ --vocab-type bpe +python convert-hf-to-gguf.py models/mymodel/ --vocab-type bpe # quantize the model to 4-bits (using Q4_K_M method) ./quantize ./models/mymodel/ggml-model-f16.gguf ./models/mymodel/ggml-model-Q4_K_M.gguf Q4_K_M diff --git a/ci/run.sh b/ci/run.sh index 940299025..3fc5f48b2 100755 --- a/ci/run.sh +++ b/ci/run.sh @@ -287,7 +287,7 @@ function gg_run_open_llama_7b_v2 { (time cmake -DCMAKE_BUILD_TYPE=Release ${CMAKE_EXTRA} -DLLAMA_CUDA=1 .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log (time make -j ) 2>&1 | tee -a $OUT/${ci}-make.log - python3 ../convert.py ${path_models} --outfile ${path_models}/ggml-model-f16.gguf + python3 ../examples/convert-legacy-llama.py ${path_models} --outfile ${path_models}/ggml-model-f16.gguf model_f16="${path_models}/ggml-model-f16.gguf" model_q8_0="${path_models}/ggml-model-q8_0.gguf" diff --git a/convert-hf-to-gguf.py b/convert-hf-to-gguf.py index 98b50d150..9f29cda23 100755 --- a/convert-hf-to-gguf.py +++ b/convert-hf-to-gguf.py @@ -25,8 +25,6 @@ if 'NO_LOCAL_GGUF' not in os.environ: sys.path.insert(1, str(Path(__file__).parent / 'gguf-py')) import gguf -from convert import LlamaHfVocab - logger = logging.getLogger("hf-to-gguf") @@ -634,7 +632,7 @@ class Model: special_vocab.add_to_gguf(self.gguf_writer) def _set_vocab_llama_hf(self): - vocab = LlamaHfVocab(self.dir_model) + vocab = gguf.LlamaHfVocab(self.dir_model) tokens = [] scores = [] toktypes = [] diff --git a/docs/HOWTO-add-model.md b/docs/HOWTO-add-model.md index 48769cdf6..138124248 100644 --- a/docs/HOWTO-add-model.md +++ b/docs/HOWTO-add-model.md @@ -17,7 +17,7 @@ Also, it is important to check that the examples and main ggml backends (CUDA, M ### 1. Convert the model to GGUF This step is done in python with a `convert` script using the [gguf](https://pypi.org/project/gguf/) library. -Depending on the model architecture, you can use either [convert.py](../convert.py) or [convert-hf-to-gguf.py](../convert-hf-to-gguf.py). +Depending on the model architecture, you can use either [convert-hf-to-gguf.py](../convert-hf-to-gguf.py) or [examples/convert-legacy-llama.py](../examples/convert-legacy-llama.py) (for `llama/llama2` models in `.pth` format). The convert script reads the model configuration, tokenizer, tensor names+data and converts them to GGUF metadata and tensors. diff --git a/convert.py b/examples/convert-legacy-llama.py similarity index 82% rename from convert.py rename to examples/convert-legacy-llama.py index da1247957..fd8401015 100755 --- a/convert.py +++ b/examples/convert-legacy-llama.py @@ -24,14 +24,16 @@ from abc import ABC, abstractmethod from concurrent.futures import ProcessPoolExecutor, ThreadPoolExecutor from dataclasses import dataclass from pathlib import Path -from typing import TYPE_CHECKING, Any, Callable, ClassVar, IO, Iterable, Literal, Protocol, TypeVar, runtime_checkable, Optional +from typing import TYPE_CHECKING, Any, Callable, IO, Iterable, Literal, TypeVar, Optional import numpy as np -from sentencepiece import SentencePieceProcessor if 'NO_LOCAL_GGUF' not in os.environ: - sys.path.insert(1, str(Path(__file__).parent / 'gguf-py')) + # use .parent.parent since we are in "examples" directory + sys.path.insert(1, str(Path(__file__).parent.parent / 'gguf-py')) + import gguf +from gguf import BaseVocab, Vocab, NoVocab, BpeVocab, SentencePieceVocab, LlamaHfVocab if TYPE_CHECKING: from typing_extensions import Self, TypeAlias @@ -380,306 +382,6 @@ class Metadata: return metadata -# -# vocab -# - - -@runtime_checkable -class BaseVocab(Protocol): - tokenizer_model: ClassVar[str] - name: ClassVar[str] - - -class NoVocab(BaseVocab): - tokenizer_model = "no_vocab" - name = "no_vocab" - - def __repr__(self) -> str: - return "" - - -@runtime_checkable -class Vocab(BaseVocab, Protocol): - vocab_size: int - added_tokens_dict: dict[str, int] - added_tokens_list: list[str] - fname_tokenizer: Path - - def __init__(self, base_path: Path): ... - def all_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]: ... - - -class BpeVocab(Vocab): - tokenizer_model = "gpt2" - name = "bpe" - - def __init__(self, base_path: Path): - added_tokens: dict[str, int] = {} - - if (fname_tokenizer := base_path / 'vocab.json').exists(): - # "slow" tokenizer - with open(fname_tokenizer, encoding="utf-8") as f: - self.vocab = json.load(f) - - try: - # FIXME: Verify that added tokens here _cannot_ overlap with the main vocab. - with open(base_path / ADDED_TOKENS_FILE, encoding="utf-8") as f: - added_tokens = json.load(f) - except FileNotFoundError: - pass - else: - # "fast" tokenizer - fname_tokenizer = base_path / FAST_TOKENIZER_FILE - - # if this fails, FileNotFoundError propagates to caller - with open(fname_tokenizer, encoding="utf-8") as f: - tokenizer_json = json.load(f) - - tokenizer_model: dict[str, Any] = tokenizer_json['model'] - if ( - tokenizer_model['type'] != 'BPE' or tokenizer_model.get('byte_fallback', False) - or tokenizer_json['decoder']['type'] != 'ByteLevel' - ): - raise FileNotFoundError('Cannot find GPT-2 BPE tokenizer') - - self.vocab = tokenizer_model["vocab"] - - if (added := tokenizer_json.get('added_tokens')) is not None: - # Added tokens here can be duplicates of the main vocabulary. - added_tokens = {item['content']: item['id'] - for item in added - if item['content'] not in self.vocab} - - vocab_size = len(self.vocab) - expected_ids = list(range(vocab_size, vocab_size + len(added_tokens))) - actual_ids = sorted(added_tokens.values()) - if expected_ids != actual_ids: - expected_end_id = vocab_size + len(actual_ids) - 1 - raise ValueError(f"Expected the {len(actual_ids)} added token ID(s) to be sequential in the range " - f"{vocab_size} - {expected_end_id}; got {actual_ids}") - - items = sorted(added_tokens.items(), key=lambda text_idx: text_idx[1]) - self.added_tokens_dict = added_tokens - self.added_tokens_list = [text for (text, idx) in items] - self.vocab_size_base = vocab_size - self.vocab_size = self.vocab_size_base + len(self.added_tokens_list) - self.fname_tokenizer = fname_tokenizer - - def bpe_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]: - reverse_vocab = {id: encoded_tok for encoded_tok, id in self.vocab.items()} - - for i, _ in enumerate(self.vocab): - yield reverse_vocab[i], 0.0, gguf.TokenType.NORMAL - - def added_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]: - for text in self.added_tokens_list: - score = -1000.0 - yield text.encode("utf-8"), score, gguf.TokenType.CONTROL - - def all_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]: - yield from self.bpe_tokens() - yield from self.added_tokens() - - def __repr__(self) -> str: - return f"" - - -class SentencePieceVocab(Vocab): - tokenizer_model = "llama" - name = "spm" - - def __init__(self, base_path: Path): - added_tokens: dict[str, int] = {} - if (fname_tokenizer := base_path / 'tokenizer.model').exists(): - # normal location - try: - with open(base_path / ADDED_TOKENS_FILE, encoding="utf-8") as f: - added_tokens = json.load(f) - except FileNotFoundError: - pass - elif not (fname_tokenizer := base_path.parent / 'tokenizer.model').exists(): - # not found in alternate location either - raise FileNotFoundError('Cannot find tokenizer.model') - - self.sentencepiece_tokenizer = SentencePieceProcessor() - self.sentencepiece_tokenizer.LoadFromFile(str(fname_tokenizer)) - vocab_size = self.sentencepiece_tokenizer.vocab_size() - - new_tokens = {id: piece for piece, id in added_tokens.items() if id >= vocab_size} - expected_new_ids = list(range(vocab_size, vocab_size + len(new_tokens))) - actual_new_ids = sorted(new_tokens.keys()) - - if expected_new_ids != actual_new_ids: - raise ValueError(f"Expected new token IDs {expected_new_ids} to be sequential; got {actual_new_ids}") - - # Token pieces that were added to the base vocabulary. - self.added_tokens_dict = added_tokens - self.added_tokens_list = [new_tokens[id] for id in actual_new_ids] - self.vocab_size_base = vocab_size - self.vocab_size = self.vocab_size_base + len(self.added_tokens_list) - self.fname_tokenizer = fname_tokenizer - - def sentencepiece_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]: - tokenizer = self.sentencepiece_tokenizer - for i in range(tokenizer.vocab_size()): - piece = tokenizer.IdToPiece(i) - text = piece.encode("utf-8") - score: float = tokenizer.GetScore(i) - - toktype = gguf.TokenType.NORMAL - if tokenizer.IsUnknown(i): - toktype = gguf.TokenType.UNKNOWN - if tokenizer.IsControl(i): - toktype = gguf.TokenType.CONTROL - - # NOTE: I think added_tokens are user defined. - # ref: https://github.com/google/sentencepiece/blob/master/src/sentencepiece_model.proto - # if tokenizer.is_user_defined(i): toktype = gguf.TokenType.USER_DEFINED - - if tokenizer.IsUnused(i): - toktype = gguf.TokenType.UNUSED - if tokenizer.IsByte(i): - toktype = gguf.TokenType.BYTE - - yield text, score, toktype - - def added_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]: - for text in self.added_tokens_list: - score = -1000.0 - yield text.encode("utf-8"), score, gguf.TokenType.USER_DEFINED - - def all_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]: - yield from self.sentencepiece_tokens() - yield from self.added_tokens() - - def __repr__(self) -> str: - return f"" - - -class LlamaHfVocab(Vocab): - tokenizer_model = "llama" - name = "hfft" - - def __init__(self, base_path: Path): - fname_tokenizer = base_path / FAST_TOKENIZER_FILE - # if this fails, FileNotFoundError propagates to caller - with open(fname_tokenizer, encoding='utf-8') as f: - tokenizer_json = json.load(f) - - # pre-check so we know if we need transformers - tokenizer_model: dict[str, Any] = tokenizer_json['model'] - is_llama3 = ( - tokenizer_model['type'] == 'BPE' and tokenizer_model.get('ignore_merges', False) - and not tokenizer_model.get('byte_fallback', True) - ) - if is_llama3: - raise TypeError('Llama 3 must be converted with BpeVocab') - - if not is_llama3 and ( - tokenizer_model['type'] != 'BPE' or not tokenizer_model.get('byte_fallback', False) - or tokenizer_json['decoder']['type'] != 'Sequence' - ): - raise FileNotFoundError('Cannot find Llama BPE tokenizer') - - try: - from transformers import AutoTokenizer - except ImportError as e: - raise ImportError( - "To use LlamaHfVocab, please install the `transformers` package. " - "You can install it with `pip install transformers`." - ) from e - - # Allow the tokenizer to default to slow or fast versions. - # Explicitly set tokenizer to use local paths. - self.tokenizer = AutoTokenizer.from_pretrained( - base_path, - cache_dir=base_path, - local_files_only=True, - ) - assert self.tokenizer.is_fast # assume tokenizer.json is used - - # Initialize lists and dictionaries for added tokens - self.added_tokens_list = [] - self.added_tokens_dict = dict() - self.added_tokens_ids = set() - - # Process added tokens - for tok, tokidx in sorted( - self.tokenizer.get_added_vocab().items(), key=lambda x: x[1] - ): - # Only consider added tokens that are not in the base vocabulary - if tokidx >= self.tokenizer.vocab_size: - self.added_tokens_list.append(tok) - self.added_tokens_dict[tok] = tokidx - self.added_tokens_ids.add(tokidx) - - # Store special tokens and their IDs - self.specials = { - tok: self.tokenizer.get_vocab()[tok] - for tok in self.tokenizer.all_special_tokens - } - self.special_ids = set(self.tokenizer.all_special_ids) - - # Set vocabulary sizes - self.vocab_size_base = self.tokenizer.vocab_size - self.vocab_size = self.vocab_size_base + len(self.added_tokens_list) - - self.fname_tokenizer = fname_tokenizer - - def hf_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]: - reverse_vocab = { - id: encoded_tok for encoded_tok, id in self.tokenizer.get_vocab().items() - } - - for token_id in range(self.vocab_size_base): - # Skip processing added tokens here - if token_id in self.added_tokens_ids: - continue - - # Convert token text to bytes - token_text = reverse_vocab[token_id].encode("utf-8") - - # Yield token text, score, and type - yield token_text, self.get_token_score(token_id), self.get_token_type( - token_id, token_text, self.special_ids # Reuse already stored special IDs - ) - - def get_token_type(self, token_id: int, token_text: bytes, special_ids: set[int]) -> gguf.TokenType: - # Special case for byte tokens - if re.fullmatch(br"<0x[0-9A-Fa-f]{2}>", token_text): - return gguf.TokenType.BYTE - - # Determine token type based on whether it's a special token - return gguf.TokenType.CONTROL if token_id in special_ids else gguf.TokenType.NORMAL - - def get_token_score(self, token_id: int) -> float: - # Placeholder for actual logic to determine the token's score - # This needs to be implemented based on specific requirements - return -1000.0 # Default score - - def added_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]: - for text in self.added_tokens_list: - if text in self.specials: - toktype = self.get_token_type(self.specials[text], b'', self.special_ids) - score = self.get_token_score(self.specials[text]) - else: - toktype = gguf.TokenType.USER_DEFINED - score = -1000.0 - - yield text.encode("utf-8"), score, toktype - - def has_newline_token(self): - return "<0x0A>" in self.tokenizer.vocab or "\n" in self.tokenizer.vocab - - def all_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]: - yield from self.hf_tokens() - yield from self.added_tokens() - - def __repr__(self) -> str: - return f"" - - # # data loading # TODO: reuse (probably move to gguf.py?) diff --git a/examples/llava/MobileVLM-README.md b/examples/llava/MobileVLM-README.md index 413e433dd..74f021dec 100644 --- a/examples/llava/MobileVLM-README.md +++ b/examples/llava/MobileVLM-README.md @@ -54,10 +54,10 @@ python ./examples/llava/convert-image-encoder-to-gguf \ --projector-type ldpv2 ``` -4. Use `convert.py` to convert the LLaMA part of LLaVA to GGUF: +4. Use `examples/convert-legacy-llama.py` to convert the LLaMA part of LLaVA to GGUF: ```sh -python ./convert.py path/to/MobileVLM-1.7B +python ./examples/convert-legacy-llama.py path/to/MobileVLM-1.7B ``` 5. Use `quantize` to convert LLaMA part's DataType from `fp16` to `q4_k` diff --git a/examples/llava/README.md b/examples/llava/README.md index 4fb0cf381..8d1ae5270 100644 --- a/examples/llava/README.md +++ b/examples/llava/README.md @@ -50,10 +50,10 @@ python ./examples/llava/llava-surgery.py -m ../llava-v1.5-7b python ./examples/llava/convert-image-encoder-to-gguf.py -m ../clip-vit-large-patch14-336 --llava-projector ../llava-v1.5-7b/llava.projector --output-dir ../llava-v1.5-7b ``` -5. Use `convert.py` to convert the LLaMA part of LLaVA to GGUF: +5. Use `examples/convert-legacy-llama.py` to convert the LLaMA part of LLaVA to GGUF: ```sh -python ./convert.py ../llava-v1.5-7b --skip-unknown +python ./examples/convert-legacy-llama.py ../llava-v1.5-7b --skip-unknown ``` Now both the LLaMA part and the image encoder are in the `llava-v1.5-7b` directory. @@ -92,7 +92,7 @@ python ./examples/llava/convert-image-encoder-to-gguf.py -m vit --llava-projecto 6) Then convert the model to gguf format: ```console -python ./convert.py ../llava-v1.6-vicuna-7b/ --skip-unknown +python ./examples/convert-legacy-llama.py ../llava-v1.6-vicuna-7b/ --skip-unknown ``` 7) And finally we can run the llava-cli using the 1.6 model version: diff --git a/examples/llava/requirements.txt b/examples/llava/requirements.txt index f80f727a7..17cb4d5e5 100644 --- a/examples/llava/requirements.txt +++ b/examples/llava/requirements.txt @@ -1,3 +1,3 @@ --r ../../requirements/requirements-convert.txt +-r ../../requirements/requirements-convert-legacy-llama.txt pillow~=10.2.0 torch~=2.1.1 diff --git a/examples/make-ggml.py b/examples/make-ggml.py deleted file mode 100755 index c73485ebf..000000000 --- a/examples/make-ggml.py +++ /dev/null @@ -1,98 +0,0 @@ -#!/usr/bin/env python3 -""" -This script converts Hugging Face Llama, StarCoder, Falcon, Baichuan, and GPT-NeoX models to GGUF and quantizes them. - -Usage: -python make-ggml.py {model_dir_or_hf_repo_name} --model_type {model_type} [--outname {output_name} (Optional)] [--outdir {output_directory} (Optional)] [--quants {quant_types} (Optional)] [--keep_fp16 (Optional)] - -Arguments: -- model: (Required) The directory of the downloaded Hugging Face model or the name of the Hugging Face model repository. If the model directory does not exist, it will be downloaded from the Hugging Face model hub. -- --model_type: (Required) The type of the model to be converted. Choose from llama, starcoder, falcon, baichuan, or gptneox. -- --outname: (Optional) The name of the output model. If not specified, the last part of the model directory path or the Hugging Face model repo name will be used. -- --outdir: (Optional) The directory where the output model(s) will be stored. If not specified, '../models/{outname}' will be used. -- --quants: (Optional) The types of quantization to apply. This should be a space-separated list. The default is 'Q4_K_M Q5_K_S'. -- --keep_fp16: (Optional) If specified, the FP16 model will not be deleted after the quantized models are created. - -Old quant types (some base model types require these): -- Q4_0: small, very high quality loss - legacy, prefer using Q3_K_M -- Q4_1: small, substantial quality loss - legacy, prefer using Q3_K_L -- Q5_0: medium, balanced quality - legacy, prefer using Q4_K_M -- Q5_1: medium, low quality loss - legacy, prefer using Q5_K_M - -New quant types (recommended): -- Q2_K: smallest, extreme quality loss - not recommended -- Q3_K: alias for Q3_K_M -- Q3_K_S: very small, very high quality loss -- Q3_K_M: very small, very high quality loss -- Q3_K_L: small, substantial quality loss -- Q4_K: alias for Q4_K_M -- Q4_K_S: small, significant quality loss -- Q4_K_M: medium, balanced quality - recommended -- Q5_K: alias for Q5_K_M -- Q5_K_S: large, low quality loss - recommended -- Q5_K_M: large, very low quality loss - recommended -- Q6_K: very large, extremely low quality loss -- Q8_0: very large, extremely low quality loss - not recommended -- F16: extremely large, virtually no quality loss - not recommended -- F32: absolutely huge, lossless - not recommended -""" -import subprocess -subprocess.run(f"pip install huggingface-hub==0.16.4", shell=True, check=True) - -import argparse -import os -from huggingface_hub import snapshot_download - -def main(model, model_type, outname, outdir, quants, keep_fp16): - if not os.path.isdir(model): - print(f"Model not found at {model}. Downloading...") - try: - if outname is None: - outname = model.split('/')[-1] - model = snapshot_download(repo_id=model, cache_dir='../models/hf_cache') - except Exception as e: - raise Exception(f"Could not download the model: {e}") - - if outdir is None: - outdir = f'../models/{outname}' - - if not os.path.isfile(f"{model}/config.json"): - raise Exception(f"Could not find config.json in {model}") - - os.makedirs(outdir, exist_ok=True) - - print("Building llama.cpp") - subprocess.run(f"cd .. && make quantize", shell=True, check=True) - - fp16 = f"{outdir}/{outname}.gguf.fp16.bin" - - print(f"Making unquantised GGUF at {fp16}") - if not os.path.isfile(fp16): - if model_type != "llama": - subprocess.run(f"python3 ../convert-{model_type}-hf-to-gguf.py {model} 1 --outfile {fp16}", shell=True, check=True) - else: - subprocess.run(f"python3 ../convert.py {model} --outtype f16 --outfile {fp16}", shell=True, check=True) - else: - print(f"Unquantised GGML already exists at: {fp16}") - - print("Making quants") - for type in quants: - outfile = f"{outdir}/{outname}.gguf.{type}.bin" - print(f"Making {type} : {outfile}") - subprocess.run(f"../quantize {fp16} {outfile} {type}", shell=True, check=True) - - if not keep_fp16: - os.remove(fp16) - -if __name__ == "__main__": - parser = argparse.ArgumentParser(description='Convert/Quantize HF models to GGUF. If you have the HF model downloaded already, pass the path to the model dir. Otherwise, pass the Hugging Face model repo name. You need to be in the /examples folder for it to work.') - parser.add_argument('model', help='Downloaded model dir or Hugging Face model repo name') - parser.add_argument('--model_type', required=True, choices=['llama', 'starcoder', 'falcon', 'baichuan', 'gptneox'], help='Type of the model to be converted. Choose from llama, starcoder, falcon, baichuan, or gptneox.') - parser.add_argument('--outname', default=None, help='Output model(s) name') - parser.add_argument('--outdir', default=None, help='Output directory') - parser.add_argument('--quants', nargs='*', default=["Q4_K_M", "Q5_K_S"], help='Quant types') - parser.add_argument('--keep_fp16', action='store_true', help='Keep fp16 model', default=False) - - args = parser.parse_args() - - main(args.model, args.model_type, args.outname, args.outdir, args.quants, args.keep_fp16) diff --git a/ggml-quants.c b/ggml-quants.c index 4f2c7224c..9f864e5c4 100644 --- a/ggml-quants.c +++ b/ggml-quants.c @@ -6088,6 +6088,7 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * restrict s, size_t bs, const void * r const uint8_t * restrict q2 = x[i].qs; const int8_t * restrict q8 = y[i].qs; + const __m128i mins_and_scales = __lsx_vld((const __m128i*)x[i].scales, 0); const __m128i scales8 = __lsx_vand_v(mins_and_scales, m4); const __m128i mins8 = __lsx_vand_v(__lsx_vsrli_h(mins_and_scales, 4), m4); @@ -6807,6 +6808,8 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * restrict s, size_t bs, const void * r for (int i = 0; i < nb; ++i) { const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); + const uint8_t * restrict q3 = x[i].qs; + const int8_t * restrict q8 = y[i].qs; // Set up scales memcpy(aux, x[i].scales, 12); __m128i scales128 = lsx_set_w( @@ -6828,29 +6831,32 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * restrict s, size_t bs, const void * r int bit = 0; int is = 0; + __m256i xvbit; - const uint8_t * restrict q3 = x[i].qs; - const int8_t * restrict q8 = y[i].qs; for (int j = 0; j < QK_K/128; ++j) { // load low 2 bits const __m256i q3bits = __lasx_xvld((const __m256i*)q3, 0); q3 += 32; + xvbit = __lasx_xvreplgr2vr_h(bit); // prepare low and high bits const __m256i q3l_0 = __lasx_xvand_v(q3bits, m3); - const __m256i q3h_0 = __lasx_xvslli_h(__lasx_xvsrli_h(__lasx_xvandn_v(hbits, __lasx_xvslli_h(mone, bit)), bit), 2); + const __m256i q3h_0 = __lasx_xvslli_h(__lasx_xvsrl_h(__lasx_xvandn_v(hbits, __lasx_xvsll_h(mone, xvbit)), xvbit), 2); ++bit; + xvbit = __lasx_xvreplgr2vr_h(bit); const __m256i q3l_1 = __lasx_xvand_v(__lasx_xvsrli_h(q3bits, 2), m3); - const __m256i q3h_1 = __lasx_xvslli_h(__lasx_xvsrli_h(__lasx_xvandn_v(hbits, __lasx_xvslli_h(mone, bit)), bit), 2); + const __m256i q3h_1 = __lasx_xvslli_h(__lasx_xvsrl_h(__lasx_xvandn_v(hbits, __lasx_xvsll_h(mone, xvbit)), xvbit), 2); ++bit; + xvbit = __lasx_xvreplgr2vr_h(bit); const __m256i q3l_2 = __lasx_xvand_v(__lasx_xvsrli_h(q3bits, 4), m3); - const __m256i q3h_2 = __lasx_xvslli_h(__lasx_xvsrli_h(__lasx_xvandn_v(hbits, __lasx_xvslli_h(mone, bit)), bit), 2); + const __m256i q3h_2 = __lasx_xvslli_h(__lasx_xvsrl_h(__lasx_xvandn_v(hbits, __lasx_xvsll_h(mone, xvbit)), xvbit), 2); ++bit; + xvbit = __lasx_xvreplgr2vr_h(bit); const __m256i q3l_3 = __lasx_xvand_v(__lasx_xvsrli_h(q3bits, 6), m3); - const __m256i q3h_3 = __lasx_xvslli_h(__lasx_xvsrli_h(__lasx_xvandn_v(hbits, __lasx_xvslli_h(mone, bit)), bit), 2); + const __m256i q3h_3 = __lasx_xvslli_h(__lasx_xvsrl_h(__lasx_xvandn_v(hbits, __lasx_xvsll_h(mone, xvbit)), xvbit), 2); ++bit; // load Q8 quants @@ -7399,6 +7405,9 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * restrict s, size_t bs, const void * r *s = vec_extract(vsumf0, 0); #elif defined __loongarch_asx + GGML_UNUSED(kmask1); + GGML_UNUSED(kmask2); + GGML_UNUSED(kmask3); const __m256i m4 = __lasx_xvreplgr2vr_b(0xF); @@ -7411,6 +7420,11 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * restrict s, size_t bs, const void * r const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin); memcpy(utmp, x[i].scales, 12); + utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4); + const uint32_t uaux = utmp[1] & kmask1; + utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4); + utmp[2] = uaux; + utmp[0] &= kmask1; const uint8_t * restrict q4 = x[i].qs; const int8_t * restrict q8 = y[i].qs; @@ -7450,16 +7464,17 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * restrict s, size_t bs, const void * r __m256 vd = __lasx_xvreplfr2vr_s(d); acc = __lasx_xvfmadd_s(vd, __lasx_xvffint_s_w(sumi), acc); + } acc_m = __lsx_vfadd_s(acc_m, (__m128)__lsx_vpermi_w((__m128i)acc_m, (__m128i)acc_m, 0xee)); __m128i tmp1 = __lsx_vinsgr2vr_w(__lsx_vldi(0), __lsx_vpickve2gr_w((__m128i)acc_m, 1), 0); acc_m = __lsx_vfadd_s(acc_m, (__m128)tmp1); + ft_union fi; fi.i = __lsx_vpickve2gr_w(acc_m, 0); *s = hsum_float_8(acc) + fi.f ; - #else const uint8_t * scales = (const uint8_t*)&utmp[0]; @@ -7997,6 +8012,9 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * restrict s, size_t bs, const void * r *s = vec_extract(vsumf0, 0); #elif defined __loongarch_asx + GGML_UNUSED(kmask1); + GGML_UNUSED(kmask2); + GGML_UNUSED(kmask3); const __m256i m4 = __lasx_xvreplgr2vr_b(0xF); const __m128i mzero = __lsx_vldi(0); @@ -8015,6 +8033,11 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * restrict s, size_t bs, const void * r const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin); memcpy(utmp, x[i].scales, 12); + utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4); + const uint32_t uaux = utmp[1] & kmask1; + utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4); + utmp[2] = uaux; + utmp[0] &= kmask1; const __m256i mins_and_scales = lasx_extu8_16(lsx_set_w(utmp[3], utmp[2], utmp[1], utmp[0])); @@ -8033,6 +8056,7 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * restrict s, size_t bs, const void * r __m256i sumi = __lasx_xvldi(0); int bit = 0; + __m256i xvbit; for (int j = 0; j < QK_K/64; ++j) { @@ -8041,13 +8065,15 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * restrict s, size_t bs, const void * r const __m256i q5bits = __lasx_xvld((const __m256i*)q5, 0); q5 += 32; + xvbit = __lasx_xvreplgr2vr_h(bit++); const __m256i q5l_0 = __lasx_xvand_v(q5bits, m4); - const __m256i q5h_0 = __lasx_xvslli_h(__lasx_xvsrli_h(__lasx_xvand_v(hbits, hmask), bit++), 4); + const __m256i q5h_0 = __lasx_xvslli_h(__lasx_xvsrl_h(__lasx_xvand_v(hbits, hmask), xvbit), 4); const __m256i q5_0 = __lasx_xvadd_b(q5l_0, q5h_0); hmask = __lasx_xvslli_h(hmask, 1); + xvbit = __lasx_xvreplgr2vr_h(bit++); const __m256i q5l_1 = __lasx_xvand_v(__lasx_xvsrli_h(q5bits, 4), m4); - const __m256i q5h_1 = __lasx_xvslli_h(__lasx_xvsrli_h(__lasx_xvand_v(hbits, hmask), bit++), 4); + const __m256i q5h_1 = __lasx_xvslli_h(__lasx_xvsrl_h(__lasx_xvand_v(hbits, hmask), xvbit), 4); const __m256i q5_1 = __lasx_xvadd_b(q5l_1, q5h_1); hmask = __lasx_xvslli_h(hmask, 1); @@ -8061,10 +8087,12 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * restrict s, size_t bs, const void * r p16_1 = lasx_madd_h(scale_1, p16_1); sumi = __lasx_xvadd_w(sumi, __lasx_xvadd_w(p16_0, p16_1)); + } __m256 vd = __lasx_xvreplfr2vr_s(d); acc = __lasx_xvfmadd_s(vd, __lasx_xvffint_s_w(sumi), acc); + } *s = hsum_float_8(acc) + summs; diff --git a/ggml.c b/ggml.c index b2b725f65..f479dc3e1 100644 --- a/ggml.c +++ b/ggml.c @@ -1576,11 +1576,11 @@ do { \ // F16 arithmetic is not supported by AVX, so we use F32 instead -#define GGML_F32Cx8 __m256 +#define GGML_F32Cx8 __m256 #define GGML_F32Cx8_ZERO (__m256)__lasx_xvldi(0) #define GGML_F32Cx8_SET1(x) (__m256)__lasx_xvreplgr2vr_w((x)) -static inline __m256 __lasx_f32cx8_load(ggml_fp16_t *x) { +static inline __m256 __lasx_f32cx8_load(const ggml_fp16_t * x) { float tmp[8]; for (int i = 0; i < 8; i++) { @@ -1589,13 +1589,14 @@ static inline __m256 __lasx_f32cx8_load(ggml_fp16_t *x) { return (__m256)__lasx_xvld(tmp, 0); } -static inline void __lasx_f32cx8_store(ggml_fp16_t *x, __m256 y) { +static inline void __lasx_f32cx8_store(ggml_fp16_t * x, __m256 y) { float arr[8]; __lasx_xvst(y, arr, 0); - for (int i = 0; i < 8; i++) + for (int i = 0; i < 8; i++) { x[i] = GGML_FP32_TO_FP16(arr[i]); + } } #define GGML_F32Cx8_LOAD(x) __lasx_f32cx8_load(x) #define GGML_F32Cx8_STORE(x, y) __lasx_f32cx8_store(x, y) @@ -1671,7 +1672,7 @@ static inline void __lasx_f32cx8_store(ggml_fp16_t *x, __m256 y) { #define GGML_F16_STEP 32 #define GGML_F16_EPR 4 -static inline __m128 __lsx_f16x4_load(ggml_fp16_t *x) { +static inline __m128 __lsx_f16x4_load(const ggml_fp16_t * x) { float tmp[4]; tmp[0] = GGML_FP16_TO_FP32(x[0]); @@ -1682,7 +1683,7 @@ static inline __m128 __lsx_f16x4_load(ggml_fp16_t *x) { return __lsx_vld(tmp, 0); } -static inline void __lsx_f16x4_store(ggml_fp16_t *x, __m128 y) { +static inline void __lsx_f16x4_store(ggml_fp16_t * x, __m128 y) { float arr[4]; __lsx_vst(y, arr, 0); @@ -2315,32 +2316,27 @@ inline static __m512 ggml_v_expf(__m512 x) { const __m512 r = _mm512_set1_ps(0x1.8p23f); const __m512 z = _mm512_fmadd_ps(x, _mm512_set1_ps(0x1.715476p+0f), r); const __m512 n = _mm512_sub_ps(z, r); - const __m512 b = _mm512_fnmadd_ps(n, _mm512_set1_ps(0x1.7f7d1cp-20f), - _mm512_fnmadd_ps(n, _mm512_set1_ps(0x1.62e4p-1f), x)); - const __m512i e = _mm512_slli_epi32(_mm512_castps_si512(z), 23); - const __m512 k = _mm512_castsi512_ps(_mm512_add_epi32(e, _mm512_castps_si512(_mm512_set1_ps(1)))); - const __mmask16 c = _mm512_cmp_ps_mask(_mm512_abs_ps(n), _mm512_set1_ps(126), _CMP_GT_OQ); - const __m512 u = _mm512_mul_ps(b, b); - const __m512 j = _mm512_fmadd_ps(_mm512_fmadd_ps(_mm512_fmadd_ps(_mm512_set1_ps(0x1.0e4020p-7f), b, - _mm512_set1_ps(0x1.573e2ep-5f)), u, - _mm512_fmadd_ps(_mm512_set1_ps(0x1.555e66p-3f), b, - _mm512_set1_ps(0x1.fffdb6p-2f))), - u, _mm512_mul_ps(_mm512_set1_ps(0x1.ffffecp-1f), b)); - if (_mm512_kortestz(c, c)) - return _mm512_fmadd_ps(j, k, k); - const __m512i g = _mm512_and_si512( - _mm512_movm_epi32(_mm512_cmp_ps_mask(n, _mm512_setzero_ps(), _CMP_LE_OQ)), - _mm512_set1_epi32(0x82000000u)); - const __m512 s1 = - _mm512_castsi512_ps(_mm512_add_epi32(g, _mm512_set1_epi32(0x7f000000u))); - const __m512 s2 = _mm512_castsi512_ps(_mm512_sub_epi32(e, g)); + const __m512 b = + _mm512_fnmadd_ps(n, _mm512_set1_ps(0x1.7f7d1cp-20f), + _mm512_fnmadd_ps(n, _mm512_set1_ps(0x1.62e4p-1f), x)); const __mmask16 d = _mm512_cmp_ps_mask(_mm512_abs_ps(n), _mm512_set1_ps(192), _CMP_GT_OQ); - return _mm512_mask_blend_ps( - d, _mm512_mask_blend_ps( - c, _mm512_fmadd_ps(k, j, k), - _mm512_mul_ps(_mm512_fmadd_ps(s2, j, s2), s1)), - _mm512_mul_ps(s1, s1)); + const __m512 u = _mm512_mul_ps(b, b); + const __m512 j = _mm512_fmadd_ps( + _mm512_fmadd_ps(_mm512_fmadd_ps(_mm512_set1_ps(0x1.0e4020p-7f), b, + _mm512_set1_ps(0x1.573e2ep-5f)), + u, + _mm512_fmadd_ps(_mm512_set1_ps(0x1.555e66p-3f), b, + _mm512_set1_ps(0x1.fffdb6p-2f))), + u, + _mm512_fmadd_ps(_mm512_set1_ps(0x1.ffffecp-1f), b, _mm512_set1_ps(1.0F))); + const __m512 res = _mm512_scalef_ps(j, n); + if (_mm512_kortestz(d, d)) + return res; + const __m512 zero = _mm512_setzero_ps(); + const __m512 alt = _mm512_mask_blend_ps( + _mm512_cmp_ps_mask(n, zero, _CMP_LE_OQ), _mm512_set1_ps(INFINITY), zero); + return _mm512_mask_blend_ps(d, res, alt); } // computes silu x/(1+exp(-x)) in single precision vector diff --git a/gguf-py/gguf/vocab.py b/gguf-py/gguf/vocab.py index 3ba99be4f..dc5749913 100644 --- a/gguf-py/gguf/vocab.py +++ b/gguf-py/gguf/vocab.py @@ -1,10 +1,15 @@ from __future__ import annotations +import re import logging import json import os from pathlib import Path -from typing import Any, Callable, Sequence, Mapping, Iterable +from typing import Any, Callable, Sequence, Mapping, Iterable, Protocol, ClassVar, runtime_checkable + +from sentencepiece import SentencePieceProcessor + +import gguf from .gguf_writer import GGUFWriter @@ -163,3 +168,298 @@ class SpecialVocab: for typ in self.special_token_types: self._set_special_token(typ, config.get(f'{typ}_token_id')) return True + + +@runtime_checkable +class BaseVocab(Protocol): + tokenizer_model: ClassVar[str] + name: ClassVar[str] + + +@runtime_checkable +class Vocab(BaseVocab, Protocol): + vocab_size: int + added_tokens_dict: dict[str, int] + added_tokens_list: list[str] + fname_tokenizer: Path + + def __init__(self, base_path: Path): ... + def all_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]: ... + + +class NoVocab(BaseVocab): + tokenizer_model = "no_vocab" + name = "no_vocab" + + def __repr__(self) -> str: + return "" + + +class BpeVocab(Vocab): + tokenizer_model = "gpt2" + name = "bpe" + + def __init__(self, base_path: Path): + added_tokens: dict[str, int] = {} + + if (fname_tokenizer := base_path / 'vocab.json').exists(): + # "slow" tokenizer + with open(fname_tokenizer, encoding="utf-8") as f: + self.vocab = json.load(f) + + try: + # FIXME: Verify that added tokens here _cannot_ overlap with the main vocab. + with open(base_path / 'added_tokens.json', encoding="utf-8") as f: + added_tokens = json.load(f) + except FileNotFoundError: + pass + else: + # "fast" tokenizer + fname_tokenizer = base_path / 'tokenizer.json' + + # if this fails, FileNotFoundError propagates to caller + with open(fname_tokenizer, encoding="utf-8") as f: + tokenizer_json = json.load(f) + + tokenizer_model: dict[str, Any] = tokenizer_json['model'] + if ( + tokenizer_model['type'] != 'BPE' or tokenizer_model.get('byte_fallback', False) + or tokenizer_json['decoder']['type'] != 'ByteLevel' + ): + raise FileNotFoundError('Cannot find GPT-2 BPE tokenizer') + + self.vocab = tokenizer_model["vocab"] + + if (added := tokenizer_json.get('added_tokens')) is not None: + # Added tokens here can be duplicates of the main vocabulary. + added_tokens = {item['content']: item['id'] + for item in added + if item['content'] not in self.vocab} + + vocab_size = len(self.vocab) + expected_ids = list(range(vocab_size, vocab_size + len(added_tokens))) + actual_ids = sorted(added_tokens.values()) + if expected_ids != actual_ids: + expected_end_id = vocab_size + len(actual_ids) - 1 + raise ValueError(f"Expected the {len(actual_ids)} added token ID(s) to be sequential in the range " + f"{vocab_size} - {expected_end_id}; got {actual_ids}") + + items = sorted(added_tokens.items(), key=lambda text_idx: text_idx[1]) + self.added_tokens_dict = added_tokens + self.added_tokens_list = [text for (text, idx) in items] + self.vocab_size_base = vocab_size + self.vocab_size = self.vocab_size_base + len(self.added_tokens_list) + self.fname_tokenizer = fname_tokenizer + + def bpe_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]: + reverse_vocab = {id: encoded_tok for encoded_tok, id in self.vocab.items()} + + for i, _ in enumerate(self.vocab): + yield reverse_vocab[i], 0.0, gguf.TokenType.NORMAL + + def added_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]: + for text in self.added_tokens_list: + score = -1000.0 + yield text.encode("utf-8"), score, gguf.TokenType.CONTROL + + def all_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]: + yield from self.bpe_tokens() + yield from self.added_tokens() + + def __repr__(self) -> str: + return f"" + + +class SentencePieceVocab(Vocab): + tokenizer_model = "llama" + name = "spm" + + def __init__(self, base_path: Path): + added_tokens: dict[str, int] = {} + if (fname_tokenizer := base_path / 'tokenizer.model').exists(): + # normal location + try: + with open(base_path / 'added_tokens.json', encoding="utf-8") as f: + added_tokens = json.load(f) + except FileNotFoundError: + pass + elif not (fname_tokenizer := base_path.parent / 'tokenizer.model').exists(): + # not found in alternate location either + raise FileNotFoundError('Cannot find tokenizer.model') + + self.sentencepiece_tokenizer = SentencePieceProcessor() + self.sentencepiece_tokenizer.LoadFromFile(str(fname_tokenizer)) + vocab_size = self.sentencepiece_tokenizer.vocab_size() + + new_tokens = {id: piece for piece, id in added_tokens.items() if id >= vocab_size} + expected_new_ids = list(range(vocab_size, vocab_size + len(new_tokens))) + actual_new_ids = sorted(new_tokens.keys()) + + if expected_new_ids != actual_new_ids: + raise ValueError(f"Expected new token IDs {expected_new_ids} to be sequential; got {actual_new_ids}") + + # Token pieces that were added to the base vocabulary. + self.added_tokens_dict = added_tokens + self.added_tokens_list = [new_tokens[id] for id in actual_new_ids] + self.vocab_size_base = vocab_size + self.vocab_size = self.vocab_size_base + len(self.added_tokens_list) + self.fname_tokenizer = fname_tokenizer + + def sentencepiece_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]: + tokenizer = self.sentencepiece_tokenizer + for i in range(tokenizer.vocab_size()): + piece = tokenizer.IdToPiece(i) + text = piece.encode("utf-8") + score: float = tokenizer.GetScore(i) + + toktype = gguf.TokenType.NORMAL + if tokenizer.IsUnknown(i): + toktype = gguf.TokenType.UNKNOWN + if tokenizer.IsControl(i): + toktype = gguf.TokenType.CONTROL + + # NOTE: I think added_tokens are user defined. + # ref: https://github.com/google/sentencepiece/blob/master/src/sentencepiece_model.proto + # if tokenizer.is_user_defined(i): toktype = gguf.TokenType.USER_DEFINED + + if tokenizer.IsUnused(i): + toktype = gguf.TokenType.UNUSED + if tokenizer.IsByte(i): + toktype = gguf.TokenType.BYTE + + yield text, score, toktype + + def added_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]: + for text in self.added_tokens_list: + score = -1000.0 + yield text.encode("utf-8"), score, gguf.TokenType.USER_DEFINED + + def all_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]: + yield from self.sentencepiece_tokens() + yield from self.added_tokens() + + def __repr__(self) -> str: + return f"" + + +class LlamaHfVocab(Vocab): + tokenizer_model = "llama" + name = "hfft" + + def __init__(self, base_path: Path): + fname_tokenizer = base_path / 'tokenizer.json' + # if this fails, FileNotFoundError propagates to caller + with open(fname_tokenizer, encoding='utf-8') as f: + tokenizer_json = json.load(f) + + # pre-check so we know if we need transformers + tokenizer_model: dict[str, Any] = tokenizer_json['model'] + is_llama3 = ( + tokenizer_model['type'] == 'BPE' and tokenizer_model.get('ignore_merges', False) + and not tokenizer_model.get('byte_fallback', True) + ) + if is_llama3: + raise TypeError('Llama 3 must be converted with BpeVocab') + + if not is_llama3 and ( + tokenizer_model['type'] != 'BPE' or not tokenizer_model.get('byte_fallback', False) + or tokenizer_json['decoder']['type'] != 'Sequence' + ): + raise FileNotFoundError('Cannot find Llama BPE tokenizer') + + try: + from transformers import AutoTokenizer + except ImportError as e: + raise ImportError( + "To use LlamaHfVocab, please install the `transformers` package. " + "You can install it with `pip install transformers`." + ) from e + + # Allow the tokenizer to default to slow or fast versions. + # Explicitly set tokenizer to use local paths. + self.tokenizer = AutoTokenizer.from_pretrained( + base_path, + cache_dir=base_path, + local_files_only=True, + ) + assert self.tokenizer.is_fast # assume tokenizer.json is used + + # Initialize lists and dictionaries for added tokens + self.added_tokens_list = [] + self.added_tokens_dict = dict() + self.added_tokens_ids = set() + + # Process added tokens + for tok, tokidx in sorted( + self.tokenizer.get_added_vocab().items(), key=lambda x: x[1] + ): + # Only consider added tokens that are not in the base vocabulary + if tokidx >= self.tokenizer.vocab_size: + self.added_tokens_list.append(tok) + self.added_tokens_dict[tok] = tokidx + self.added_tokens_ids.add(tokidx) + + # Store special tokens and their IDs + self.specials = { + tok: self.tokenizer.get_vocab()[tok] + for tok in self.tokenizer.all_special_tokens + } + self.special_ids = set(self.tokenizer.all_special_ids) + + # Set vocabulary sizes + self.vocab_size_base = self.tokenizer.vocab_size + self.vocab_size = self.vocab_size_base + len(self.added_tokens_list) + + self.fname_tokenizer = fname_tokenizer + + def hf_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]: + reverse_vocab = { + id: encoded_tok for encoded_tok, id in self.tokenizer.get_vocab().items() + } + + for token_id in range(self.vocab_size_base): + # Skip processing added tokens here + if token_id in self.added_tokens_ids: + continue + + # Convert token text to bytes + token_text = reverse_vocab[token_id].encode("utf-8") + + # Yield token text, score, and type + yield token_text, self.get_token_score(token_id), self.get_token_type( + token_id, token_text, self.special_ids # Reuse already stored special IDs + ) + + def get_token_type(self, token_id: int, token_text: bytes, special_ids: set[int]) -> gguf.TokenType: + # Special case for byte tokens + if re.fullmatch(br"<0x[0-9A-Fa-f]{2}>", token_text): + return gguf.TokenType.BYTE + + # Determine token type based on whether it's a special token + return gguf.TokenType.CONTROL if token_id in special_ids else gguf.TokenType.NORMAL + + def get_token_score(self, token_id: int) -> float: + # Placeholder for actual logic to determine the token's score + # This needs to be implemented based on specific requirements + return -1000.0 # Default score + + def added_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]: + for text in self.added_tokens_list: + if text in self.specials: + toktype = self.get_token_type(self.specials[text], b'', self.special_ids) + score = self.get_token_score(self.specials[text]) + else: + toktype = gguf.TokenType.USER_DEFINED + score = -1000.0 + + yield text.encode("utf-8"), score, toktype + + def has_newline_token(self): + return "<0x0A>" in self.tokenizer.vocab or "\n" in self.tokenizer.vocab + + def all_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]: + yield from self.hf_tokens() + yield from self.added_tokens() + + def __repr__(self) -> str: + return f"" diff --git a/llama.cpp b/llama.cpp index e7412de4b..40d2ec2c9 100644 --- a/llama.cpp +++ b/llama.cpp @@ -1702,12 +1702,13 @@ struct llama_mlock { }; using llama_mlocks = std::vector>; -static std::string llama_token_to_piece(const struct llama_context * ctx, llama_token token, bool special) { +// NOTE: avoid ever using this except for building the token_to_piece caches +static std::string llama_token_to_piece(const struct llama_model * model, llama_token token, bool special) { std::vector result(8, 0); - const int n_tokens = llama_token_to_piece(llama_get_model(ctx), token, result.data(), result.size(), special); + const int n_tokens = llama_token_to_piece(model, token, result.data(), result.size(), special); if (n_tokens < 0) { result.resize(-n_tokens); - int check = llama_token_to_piece(llama_get_model(ctx), token, result.data(), result.size(), special); + int check = llama_token_to_piece(model, token, result.data(), result.size(), special); GGML_ASSERT(check == -n_tokens); } else { @@ -2162,7 +2163,9 @@ struct llama_vocab { std::unordered_map token_to_id; std::vector id_to_token; - std::vector special_tokens_cache; + std::vector cache_special_tokens; + std::vector cache_token_to_piece; // llama_token_to_piece(special = false); + std::vector cache_token_to_piece_special; // llama_token_to_piece(special = true); std::map, int> bpe_ranks; @@ -4592,20 +4595,14 @@ static void llm_load_vocab( vocab.special_cls_id = 101; vocab.special_mask_id = 103; vocab.add_space_prefix = false; - } else { - if (tokenizer_model == "gpt2") { - vocab.type = LLAMA_VOCAB_TYPE_BPE; + } else if (tokenizer_model == "gpt2") { + vocab.type = LLAMA_VOCAB_TYPE_BPE; - const int add_space_prefix_keyidx = gguf_find_key(ctx, kv(LLM_KV_TOKENIZER_ADD_PREFIX).c_str()); - if (add_space_prefix_keyidx != -1) { - vocab.add_space_prefix = gguf_get_val_bool(ctx, add_space_prefix_keyidx); - } - } else { - LLAMA_LOG_WARN("%s: unknown tokenizer: '%s'", __func__, tokenizer_model.c_str()); - LLAMA_LOG_WARN("%s: using default tokenizer: 'llama'", __func__); - vocab.type = LLAMA_VOCAB_TYPE_SPM; - return; + const int add_space_prefix_keyidx = gguf_find_key(ctx, kv(LLM_KV_TOKENIZER_ADD_PREFIX).c_str()); + if (add_space_prefix_keyidx != -1) { + vocab.add_space_prefix = gguf_get_val_bool(ctx, add_space_prefix_keyidx); } + // read bpe merges and populate bpe ranks const int merges_keyidx = gguf_find_key(ctx, kv(LLM_KV_TOKENIZER_MERGES).c_str()); if (merges_keyidx == -1) { @@ -4639,6 +4636,8 @@ static void llm_load_vocab( vocab.special_pad_id = -1; vocab.special_cls_id = -1; vocab.special_mask_id = -1; + } else { + throw std::runtime_error(format("unknown tokenizer: '%s'", tokenizer_model.c_str())); } // for now, only BPE models have pre-tokenizers @@ -4833,17 +4832,38 @@ static void llm_load_vocab( { for (llama_vocab::id id = 0; id < (llama_vocab::id)n_vocab; ++id) { if (vocab.id_to_token[id].type != LLAMA_TOKEN_TYPE_NORMAL) { - vocab.special_tokens_cache.push_back(id); + vocab.cache_special_tokens.push_back(id); } } - std::sort( vocab.special_tokens_cache.begin(), vocab.special_tokens_cache.end(), + std::sort( vocab.cache_special_tokens.begin(), vocab.cache_special_tokens.end(), [&] (const llama_vocab::id a, const llama_vocab::id b) { return vocab.id_to_token[a].text.size() > vocab.id_to_token[b].text.size(); } ); - LLAMA_LOG_INFO("%s: special tokens cache size = %u.\n", __func__, (uint32_t)vocab.special_tokens_cache.size()); + LLAMA_LOG_INFO("%s: special tokens cache size = %u\n", __func__, (uint32_t)vocab.cache_special_tokens.size()); + } + + // build token to piece caches + { + size_t size_cache = 0; + + std::vector cache_token_to_piece (n_vocab); + std::vector cache_token_to_piece_special(n_vocab); + + for (uint32_t id = 0; id < n_vocab; ++id) { + cache_token_to_piece[id] = llama_token_to_piece(&model, id, false); + cache_token_to_piece_special[id] = llama_token_to_piece(&model, id, true); + + size_cache += cache_token_to_piece[id].size(); + size_cache += cache_token_to_piece_special[id].size(); + } + + std::swap(vocab.cache_token_to_piece, cache_token_to_piece); + std::swap(vocab.cache_token_to_piece_special, cache_token_to_piece_special); + + LLAMA_LOG_INFO("%s: token to piece cache size = %.4f MB\n", __func__, size_cache / 1024.0 / 1024.0); } } @@ -13233,7 +13253,7 @@ struct fragment_buffer_variant { static void tokenizer_st_partition(const llama_vocab & vocab, std::forward_list & buffer) { // for each special token - for (const llama_vocab::id special_id : vocab.special_tokens_cache) { + for (const llama_vocab::id special_id : vocab.cache_special_tokens) { const auto & special_token = vocab.id_to_token[special_id].text; // for each text fragment @@ -14392,7 +14412,7 @@ void llama_sample_repetition_penalties( void llama_sample_grammar(struct llama_context * ctx, llama_token_data_array * candidates, const struct llama_grammar * grammar) { GGML_ASSERT(ctx); - const int64_t t_start_sample_us = ggml_time_us(); + int64_t t_start_sample_us = ggml_time_us(); bool allow_eog = false; for (const auto & stack : grammar->stacks) { @@ -14404,12 +14424,13 @@ void llama_sample_grammar(struct llama_context * ctx, llama_token_data_array * c std::vector, llama_partial_utf8>> candidates_decoded; candidates_decoded.reserve(candidates->size); - std::vector candidates_grammar; + + std::vector candidates_grammar; candidates_grammar.reserve(candidates->size); for (size_t i = 0; i < candidates->size; ++i) { - const llama_token id = candidates->data[i].id; - const std::string piece = llama_token_to_piece(ctx, id, false); + const llama_token id = candidates->data[i].id; + const std::string & piece = ctx->model.vocab.cache_token_to_piece.at(id); if (llama_token_is_eog(&ctx->model, id)) { if (!allow_eog) { @@ -14609,7 +14630,7 @@ void llama_grammar_accept_token(struct llama_context * ctx, struct llama_grammar GGML_ASSERT(false); } - const std::string piece = llama_token_to_piece(ctx, token, false); + const std::string & piece = ctx->model.vocab.cache_token_to_piece.at(token); // Note terminating 0 in decoded string const auto decoded = decode_utf8(piece, grammar->partial_utf8); @@ -18292,69 +18313,83 @@ static std::string llama_decode_text(const std::string & text) { // does not write null-terminator to buf int32_t llama_token_to_piece(const struct llama_model * model, llama_token token, char * buf, int32_t length, bool special) { + // if we have a cache - use it + { + const auto & cache = special ? model->vocab.cache_token_to_piece_special : model->vocab.cache_token_to_piece; + + if (!cache.empty()) { + const auto & res = cache.at(token); + if (length < (int) res.size()) { + return -(int) res.size(); + } + memcpy(buf, res.c_str(), res.size()); + return res.size(); + } + } + if (0 <= token && token < llama_n_vocab(model)) { switch (llama_vocab_get_type(model->vocab)) { - case LLAMA_VOCAB_TYPE_WPM: - case LLAMA_VOCAB_TYPE_SPM: { - // NOTE: we accept all unsupported token types, - // suppressing them like CONTROL tokens. - if (llama_is_normal_token(model->vocab, token)) { - std::string result = model->vocab.id_to_token[token].text; - llama_unescape_whitespace(result); - if (length < (int) result.length()) { - return -(int) result.length(); + case LLAMA_VOCAB_TYPE_WPM: + case LLAMA_VOCAB_TYPE_SPM: { + // NOTE: we accept all unsupported token types, + // suppressing them like CONTROL tokens. + if (llama_is_normal_token(model->vocab, token)) { + std::string result = model->vocab.id_to_token[token].text; + llama_unescape_whitespace(result); + if (length < (int) result.length()) { + return -(int) result.length(); + } + memcpy(buf, result.c_str(), result.length()); + return result.length(); + } else if ( + (llama_is_user_defined_token(model->vocab, token)) || + (llama_is_control_token (model->vocab, token) && special)) { + std::string result = model->vocab.id_to_token[token].text; + if (length < (int) result.length()) { + return -(int) result.length(); + } + memcpy(buf, result.c_str(), result.length()); + return result.length(); + } else if (llama_is_unknown_token(model->vocab, token)) { // NOLINT + if (length < 3) { + return -3; + } + memcpy(buf, "\xe2\x96\x85", 3); + return 3; + } else if (llama_is_byte_token(model->vocab, token)) { + if (length < 1) { + return -1; + } + buf[0] = llama_token_to_byte(model->vocab, token); + return 1; } - memcpy(buf, result.c_str(), result.length()); - return result.length(); - } else if ( - (llama_is_user_defined_token(model->vocab, token)) || - (llama_is_control_token (model->vocab, token) && special)) { - std::string result = model->vocab.id_to_token[token].text; - if (length < (int) result.length()) { - return -(int) result.length(); - } - memcpy(buf, result.c_str(), result.length()); - return result.length(); - } else if (llama_is_unknown_token(model->vocab, token)) { // NOLINT - if (length < 3) { - return -3; - } - memcpy(buf, "\xe2\x96\x85", 3); - return 3; - } else if (llama_is_byte_token(model->vocab, token)) { - if (length < 1) { - return -1; - } - buf[0] = llama_token_to_byte(model->vocab, token); - return 1; + break; } - break; - } - case LLAMA_VOCAB_TYPE_BPE: { - // NOTE: we accept all unsupported token types, - // suppressing them like CONTROL tokens. - if (llama_is_normal_token(model->vocab, token)) { - std::string result = model->vocab.id_to_token[token].text; - result = llama_decode_text(result); - if (length < (int) result.length()) { - return -(int) result.length(); + case LLAMA_VOCAB_TYPE_BPE: { + // NOTE: we accept all unsupported token types, + // suppressing them like CONTROL tokens. + if (llama_is_normal_token(model->vocab, token)) { + std::string result = model->vocab.id_to_token[token].text; + result = llama_decode_text(result); + if (length < (int) result.length()) { + return -(int) result.length(); + } + memcpy(buf, result.c_str(), result.length()); + return result.length(); + } else if ( + (llama_is_user_defined_token(model->vocab, token)) || + (llama_is_control_token (model->vocab, token) && special)) { + std::string result = model->vocab.id_to_token[token].text; + if (length < (int) result.length()) { + return -(int) result.length(); + } + memcpy(buf, result.c_str(), result.length()); + return result.length(); } - memcpy(buf, result.c_str(), result.length()); - return result.length(); - } else if ( - (llama_is_user_defined_token(model->vocab, token)) || - (llama_is_control_token (model->vocab, token) && special)) { - std::string result = model->vocab.id_to_token[token].text; - if (length < (int) result.length()) { - return -(int) result.length(); - } - memcpy(buf, result.c_str(), result.length()); - return result.length(); + break; } - break; - } - default: - GGML_ASSERT(false); + default: + GGML_ASSERT(false); } } return 0; diff --git a/llama.h b/llama.h index 3e4474bb9..95105c28e 100644 --- a/llama.h +++ b/llama.h @@ -424,8 +424,8 @@ extern "C" { LLAMA_API enum llama_pooling_type llama_pooling_type(const struct llama_context * ctx); - LLAMA_API enum llama_vocab_type llama_vocab_type (const struct llama_model * model); - LLAMA_API enum llama_rope_type llama_rope_type (const struct llama_model * model); + LLAMA_API enum llama_vocab_type llama_vocab_type (const struct llama_model * model); + LLAMA_API enum llama_rope_type llama_rope_type (const struct llama_model * model); LLAMA_API int32_t llama_n_vocab (const struct llama_model * model); LLAMA_API int32_t llama_n_ctx_train(const struct llama_model * model); diff --git a/requirements.txt b/requirements.txt index 43f82dc2e..e5cfbf10b 100644 --- a/requirements.txt +++ b/requirements.txt @@ -4,7 +4,7 @@ # Package versions must stay compatible across all top-level python scripts. # --r ./requirements/requirements-convert.txt +-r ./requirements/requirements-convert-legacy-llama.txt -r ./requirements/requirements-convert-hf-to-gguf.txt -r ./requirements/requirements-convert-hf-to-gguf-update.txt diff --git a/requirements/requirements-convert-hf-to-gguf-update.txt b/requirements/requirements-convert-hf-to-gguf-update.txt index 6ac402610..6eacaf429 100644 --- a/requirements/requirements-convert-hf-to-gguf-update.txt +++ b/requirements/requirements-convert-hf-to-gguf-update.txt @@ -1,2 +1,2 @@ --r ./requirements-convert.txt +-r ./requirements-convert-legacy-llama.txt torch~=2.1.1 diff --git a/requirements/requirements-convert-hf-to-gguf.txt b/requirements/requirements-convert-hf-to-gguf.txt index 6ac402610..6eacaf429 100644 --- a/requirements/requirements-convert-hf-to-gguf.txt +++ b/requirements/requirements-convert-hf-to-gguf.txt @@ -1,2 +1,2 @@ --r ./requirements-convert.txt +-r ./requirements-convert-legacy-llama.txt torch~=2.1.1 diff --git a/requirements/requirements-convert.txt b/requirements/requirements-convert-legacy-llama.txt similarity index 100% rename from requirements/requirements-convert.txt rename to requirements/requirements-convert-legacy-llama.txt diff --git a/requirements/requirements-convert-llama-ggml-to-gguf.txt b/requirements/requirements-convert-llama-ggml-to-gguf.txt index a0f37cd1c..e80c29012 100644 --- a/requirements/requirements-convert-llama-ggml-to-gguf.txt +++ b/requirements/requirements-convert-llama-ggml-to-gguf.txt @@ -1 +1 @@ --r ./requirements-convert.txt +-r ./requirements-convert-legacy-llama.txt diff --git a/scripts/check-requirements.sh b/scripts/check-requirements.sh index 6a7400d3c..0c6afdd59 100755 --- a/scripts/check-requirements.sh +++ b/scripts/check-requirements.sh @@ -166,7 +166,7 @@ if (( do_cleanup )); then rm -rf -- "$all_venv" fi -check_convert_script convert.py +check_convert_script examples/convert-legacy-llama.py for py in convert-*.py; do # skip convert-hf-to-gguf-update.py # TODO: the check is failing for some reason: diff --git a/scripts/convert-gg.sh b/scripts/convert-gg.sh index 01fda16fd..8a0168432 100755 --- a/scripts/convert-gg.sh +++ b/scripts/convert-gg.sh @@ -3,20 +3,20 @@ set -e # LLaMA v1 -python3 convert.py ../llama1/7B --outfile models/llama-7b/ggml-model-f16.gguf --outtype f16 -python3 convert.py ../llama1/13B --outfile models/llama-13b/ggml-model-f16.gguf --outtype f16 -python3 convert.py ../llama1/30B --outfile models/llama-30b/ggml-model-f16.gguf --outtype f16 -python3 convert.py ../llama1/65B --outfile models/llama-65b/ggml-model-f16.gguf --outtype f16 +python3 examples/convert-legacy-llama.py ../llama1/7B --outfile models/llama-7b/ggml-model-f16.gguf --outtype f16 +python3 examples/convert-legacy-llama.py ../llama1/13B --outfile models/llama-13b/ggml-model-f16.gguf --outtype f16 +python3 examples/convert-legacy-llama.py ../llama1/30B --outfile models/llama-30b/ggml-model-f16.gguf --outtype f16 +python3 examples/convert-legacy-llama.py ../llama1/65B --outfile models/llama-65b/ggml-model-f16.gguf --outtype f16 # LLaMA v2 -python3 convert.py ../llama2/llama-2-7b --outfile models/llama-7b-v2/ggml-model-f16.gguf --outtype f16 -python3 convert.py ../llama2/llama-2-13b --outfile models/llama-13b-v2/ggml-model-f16.gguf --outtype f16 -python3 convert.py ../llama2/llama-2-70b --outfile models/llama-70b-v2/ggml-model-f16.gguf --outtype f16 +python3 examples/convert-legacy-llama.py ../llama2/llama-2-7b --outfile models/llama-7b-v2/ggml-model-f16.gguf --outtype f16 +python3 examples/convert-legacy-llama.py ../llama2/llama-2-13b --outfile models/llama-13b-v2/ggml-model-f16.gguf --outtype f16 +python3 examples/convert-legacy-llama.py ../llama2/llama-2-70b --outfile models/llama-70b-v2/ggml-model-f16.gguf --outtype f16 # Code Llama -python3 convert.py ../codellama/CodeLlama-7b/ --outfile models/codellama-7b/ggml-model-f16.gguf --outtype f16 -python3 convert.py ../codellama/CodeLlama-13b/ --outfile models/codellama-13b/ggml-model-f16.gguf --outtype f16 -python3 convert.py ../codellama/CodeLlama-34b/ --outfile models/codellama-34b/ggml-model-f16.gguf --outtype f16 +python3 examples/convert-legacy-llama.py ../codellama/CodeLlama-7b/ --outfile models/codellama-7b/ggml-model-f16.gguf --outtype f16 +python3 examples/convert-legacy-llama.py ../codellama/CodeLlama-13b/ --outfile models/codellama-13b/ggml-model-f16.gguf --outtype f16 +python3 examples/convert-legacy-llama.py ../codellama/CodeLlama-34b/ --outfile models/codellama-34b/ggml-model-f16.gguf --outtype f16 # Falcon python3 convert-falcon-hf-to-gguf.py ../falcon/falcon-7b 1 diff --git a/scripts/pod-llama.sh b/scripts/pod-llama.sh index 2058ceabf..5dabbf60e 100644 --- a/scripts/pod-llama.sh +++ b/scripts/pod-llama.sh @@ -75,7 +75,7 @@ if [ "$1" -eq "1" ]; then cd /workspace/llama.cpp - python3 convert.py ./models/tinyllama-1b --outfile ./models/tinyllama-1b/ggml-model-f16.gguf --outtype f16 + python3 examples/convert-legacy-llama.py ./models/tinyllama-1b --outfile ./models/tinyllama-1b/ggml-model-f16.gguf --outtype f16 ./quantize ./models/tinyllama-1b/ggml-model-f16.gguf ./models/tinyllama-1b/ggml-model-q4_0.gguf q4_0 ./quantize ./models/tinyllama-1b/ggml-model-f16.gguf ./models/tinyllama-1b/ggml-model-q4_k.gguf q4_k @@ -90,7 +90,7 @@ if [ "$1" -eq "2" ]; then cd /workspace/llama.cpp - python3 convert.py ./models/codellama-7b --outfile ./models/codellama-7b/ggml-model-f16.gguf --outtype f16 + python3 examples/convert-legacy-llama.py ./models/codellama-7b --outfile ./models/codellama-7b/ggml-model-f16.gguf --outtype f16 ./quantize ./models/codellama-7b/ggml-model-f16.gguf ./models/codellama-7b/ggml-model-q4_0.gguf q4_0 ./quantize ./models/codellama-7b/ggml-model-f16.gguf ./models/codellama-7b/ggml-model-q4_k.gguf q4_k @@ -105,7 +105,7 @@ if [ "$1" -eq "3" ]; then cd /workspace/llama.cpp - python3 convert.py ./models/codellama-13b --outfile ./models/codellama-13b/ggml-model-f16.gguf --outtype f16 + python3 examples/convert-legacy-llama.py ./models/codellama-13b --outfile ./models/codellama-13b/ggml-model-f16.gguf --outtype f16 ./quantize ./models/codellama-13b/ggml-model-f16.gguf ./models/codellama-13b/ggml-model-q4_0.gguf q4_0 ./quantize ./models/codellama-13b/ggml-model-f16.gguf ./models/codellama-13b/ggml-model-q4_k.gguf q4_k @@ -120,7 +120,7 @@ if [ "$1" -eq "4" ]; then cd /workspace/llama.cpp - python3 convert.py ./models/codellama-34b --outfile ./models/codellama-34b/ggml-model-f16.gguf --outtype f16 + python3 examples/convert-legacy-llama.py ./models/codellama-34b --outfile ./models/codellama-34b/ggml-model-f16.gguf --outtype f16 ./quantize ./models/codellama-34b/ggml-model-f16.gguf ./models/codellama-34b/ggml-model-q4_0.gguf q4_0 ./quantize ./models/codellama-34b/ggml-model-f16.gguf ./models/codellama-34b/ggml-model-q4_k.gguf q4_k @@ -135,7 +135,7 @@ if [ "$1" -eq "5" ]; then cd /workspace/llama.cpp - python3 convert.py ./models/codellama-7b-instruct --outfile ./models/codellama-7b-instruct/ggml-model-f16.gguf --outtype f16 + python3 examples/convert-legacy-llama.py ./models/codellama-7b-instruct --outfile ./models/codellama-7b-instruct/ggml-model-f16.gguf --outtype f16 ./quantize ./models/codellama-7b-instruct/ggml-model-f16.gguf ./models/codellama-7b-instruct/ggml-model-q4_0.gguf q4_0 ./quantize ./models/codellama-7b-instruct/ggml-model-f16.gguf ./models/codellama-7b-instruct/ggml-model-q4_k.gguf q4_k @@ -150,7 +150,7 @@ if [ "$1" -eq "6" ]; then cd /workspace/llama.cpp - python3 convert.py ./models/codellama-13b-instruct --outfile ./models/codellama-13b-instruct/ggml-model-f16.gguf --outtype f16 + python3 examples/convert-legacy-llama.py ./models/codellama-13b-instruct --outfile ./models/codellama-13b-instruct/ggml-model-f16.gguf --outtype f16 ./quantize ./models/codellama-13b-instruct/ggml-model-f16.gguf ./models/codellama-13b-instruct/ggml-model-q4_0.gguf q4_0 ./quantize ./models/codellama-13b-instruct/ggml-model-f16.gguf ./models/codellama-13b-instruct/ggml-model-q4_k.gguf q4_k @@ -165,7 +165,7 @@ if [ "$1" -eq "7" ]; then cd /workspace/llama.cpp - python3 convert.py ./models/codellama-34b-instruct --outfile ./models/codellama-34b-instruct/ggml-model-f16.gguf --outtype f16 + python3 examples/convert-legacy-llama.py ./models/codellama-34b-instruct --outfile ./models/codellama-34b-instruct/ggml-model-f16.gguf --outtype f16 ./quantize ./models/codellama-34b-instruct/ggml-model-f16.gguf ./models/codellama-34b-instruct/ggml-model-q4_0.gguf q4_0 ./quantize ./models/codellama-34b-instruct/ggml-model-f16.gguf ./models/codellama-34b-instruct/ggml-model-q4_k.gguf q4_k diff --git a/tests/CMakeLists.txt b/tests/CMakeLists.txt index 766a01752..cfa707315 100644 --- a/tests/CMakeLists.txt +++ b/tests/CMakeLists.txt @@ -129,8 +129,11 @@ llama_target_and_test(test-rope.cpp) llama_target_and_test(test-model-load-cancel.cpp LABEL "model") llama_target_and_test(test-autorelease.cpp LABEL "model") -llama_target_and_test(test-json-schema-to-grammar.cpp WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}/..) -target_include_directories(test-json-schema-to-grammar PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/../examples/server) +# TODO: disabled on loongarch64 because the ggml-ci node lacks Python 3.8 +if (NOT ${CMAKE_SYSTEM_PROCESSOR} MATCHES "loongarch64") + llama_target_and_test(test-json-schema-to-grammar.cpp WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}/..) + target_include_directories(test-json-schema-to-grammar PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/../examples/server) +endif() # dummy executable - not installed get_filename_component(TEST_TARGET test-c.c NAME_WE)