merge
This commit is contained in:
mgroeber9110
commit
e0814f1224
24 changed files with 1483 additions and 457 deletions
16
.github/workflows/build.yml
vendored
16
.github/workflows/build.yml
vendored
|
|
@ -12,7 +12,7 @@ on:
|
|||
- master
|
||||
paths: ['.github/workflows/**', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.c', '**/*.cpp']
|
||||
pull_request:
|
||||
types: [opened, synchronize, edited, reopened, review_requested, ready_for_review]
|
||||
types: [opened, synchronize, reopened]
|
||||
paths: ['**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.c', '**/*.cpp']
|
||||
|
||||
env:
|
||||
|
|
@ -20,8 +20,6 @@ env:
|
|||
|
||||
jobs:
|
||||
ubuntu-latest-make:
|
||||
if: github.event.pull_request.draft == false
|
||||
|
||||
runs-on: ubuntu-latest
|
||||
|
||||
steps:
|
||||
|
|
@ -41,8 +39,6 @@ jobs:
|
|||
make
|
||||
|
||||
ubuntu-latest-cmake:
|
||||
if: github.event.pull_request.draft == false
|
||||
|
||||
runs-on: ubuntu-latest
|
||||
|
||||
steps:
|
||||
|
|
@ -71,8 +67,6 @@ jobs:
|
|||
ctest --verbose
|
||||
|
||||
ubuntu-latest-cmake-sanitizer:
|
||||
if: github.event.pull_request.draft == false
|
||||
|
||||
runs-on: ubuntu-latest
|
||||
|
||||
continue-on-error: true
|
||||
|
|
@ -108,8 +102,6 @@ jobs:
|
|||
ctest --verbose
|
||||
|
||||
macOS-latest-make:
|
||||
if: github.event.pull_request.draft == false
|
||||
|
||||
runs-on: macos-latest
|
||||
|
||||
steps:
|
||||
|
|
@ -128,8 +120,6 @@ jobs:
|
|||
make
|
||||
|
||||
macOS-latest-cmake:
|
||||
if: github.event.pull_request.draft == false
|
||||
|
||||
runs-on: macOS-latest
|
||||
|
||||
steps:
|
||||
|
|
@ -157,8 +147,6 @@ jobs:
|
|||
ctest --verbose
|
||||
|
||||
windows-latest-cmake:
|
||||
if: github.event.pull_request.draft == false
|
||||
|
||||
runs-on: windows-latest
|
||||
|
||||
strategy:
|
||||
|
|
@ -169,7 +157,7 @@ jobs:
|
|||
- build: 'avx'
|
||||
defines: '-DLLAMA_AVX2=OFF'
|
||||
- build: 'avx512'
|
||||
defines: '-DLLAMA_AVX512=ON'
|
||||
defines: '-DLLAMA_AVX512=ON -DBUILD_SHARED_LIBS=ON'
|
||||
|
||||
steps:
|
||||
- name: Clone
|
||||
|
|
|
|||
|
|
@ -201,6 +201,10 @@ endif()
|
|||
|
||||
if (MSVC)
|
||||
add_compile_definitions(_CRT_SECURE_NO_WARNINGS)
|
||||
|
||||
if (BUILD_SHARED_LIBS)
|
||||
set(CMAKE_WINDOWS_EXPORT_ALL_SYMBOLS ON)
|
||||
endif()
|
||||
endif()
|
||||
|
||||
if (LLAMA_LTO)
|
||||
|
|
@ -307,7 +311,8 @@ add_library(ggml OBJECT
|
|||
|
||||
target_include_directories(ggml PUBLIC .)
|
||||
target_compile_features(ggml PUBLIC c_std_11) # don't bump
|
||||
target_link_libraries(ggml PRIVATE Threads::Threads ${LLAMA_EXTRA_LIBS})
|
||||
target_link_libraries(ggml PUBLIC Threads::Threads ${LLAMA_EXTRA_LIBS})
|
||||
|
||||
if (BUILD_SHARED_LIBS)
|
||||
set_target_properties(ggml PROPERTIES POSITION_INDEPENDENT_CODE ON)
|
||||
endif()
|
||||
|
|
@ -320,6 +325,7 @@ add_library(llama
|
|||
target_include_directories(llama PUBLIC .)
|
||||
target_compile_features(llama PUBLIC cxx_std_11) # don't bump
|
||||
target_link_libraries(llama PRIVATE ggml ${LLAMA_EXTRA_LIBS})
|
||||
|
||||
if (BUILD_SHARED_LIBS)
|
||||
set_target_properties(llama PROPERTIES POSITION_INDEPENDENT_CODE ON)
|
||||
target_compile_definitions(llama PRIVATE LLAMA_SHARED LLAMA_BUILD)
|
||||
|
|
|
|||
24
Makefile
24
Makefile
|
|
@ -74,13 +74,17 @@ endif
|
|||
# feel free to update the Makefile for your architecture and send a pull request or issue
|
||||
ifeq ($(UNAME_M),$(filter $(UNAME_M),x86_64 i686))
|
||||
# Use all CPU extensions that are available:
|
||||
CFLAGS += -march=native -mtune=native
|
||||
CFLAGS += -march=native -mtune=native
|
||||
CXXFLAGS += -march=native -mtune=native
|
||||
|
||||
# Usage AVX-only
|
||||
#CFLAGS += -mfma -mf16c -mavx
|
||||
#CXXFLAGS += -mfma -mf16c -mavx
|
||||
endif
|
||||
ifneq ($(filter ppc64%,$(UNAME_M)),)
|
||||
POWER9_M := $(shell grep "POWER9" /proc/cpuinfo)
|
||||
ifneq (,$(findstring POWER9,$(POWER9_M)))
|
||||
CFLAGS += -mcpu=power9
|
||||
CFLAGS += -mcpu=power9
|
||||
CXXFLAGS += -mcpu=power9
|
||||
endif
|
||||
# Require c++23's std::byteswap for big-endian support.
|
||||
|
|
@ -101,18 +105,24 @@ ifdef LLAMA_OPENBLAS
|
|||
LDFLAGS += -lopenblas
|
||||
endif
|
||||
ifdef LLAMA_CUBLAS
|
||||
CFLAGS += -DGGML_USE_CUBLAS -I/usr/local/cuda/include
|
||||
LDFLAGS += -lcublas -lculibos -lcudart -lcublasLt -lpthread -ldl -lrt -L/usr/local/cuda/lib64
|
||||
OBJS += ggml-cuda.o
|
||||
CFLAGS += -DGGML_USE_CUBLAS -I/usr/local/cuda/include
|
||||
LDFLAGS += -lcublas -lculibos -lcudart -lcublasLt -lpthread -ldl -lrt -L/usr/local/cuda/lib64
|
||||
OBJS += ggml-cuda.o
|
||||
NVCC = nvcc
|
||||
NVCCFLAGS = --forward-unknown-to-host-linker -arch=native
|
||||
ggml-cuda.o: ggml-cuda.cu ggml-cuda.h
|
||||
nvcc -arch=native -c -o $@ $<
|
||||
$(NVCC) $(NVCCFLAGS) $(CXXFLAGS) -c $< -o $@
|
||||
endif
|
||||
ifdef LLAMA_GPROF
|
||||
CFLAGS += -pg
|
||||
CXXFLAGS += -pg
|
||||
endif
|
||||
ifdef LLAMA_PERF
|
||||
CFLAGS += -DGGML_PERF
|
||||
CXXFLAGS += -DGGML_PERF
|
||||
endif
|
||||
ifneq ($(filter aarch64%,$(UNAME_M)),)
|
||||
CFLAGS += -mcpu=native
|
||||
CFLAGS += -mcpu=native
|
||||
CXXFLAGS += -mcpu=native
|
||||
endif
|
||||
ifneq ($(filter armv6%,$(UNAME_M)),)
|
||||
|
|
|
|||
21
README.md
21
README.md
|
|
@ -275,18 +275,19 @@ cadaver, cauliflower, cabbage (vegetable), catalpa (tree) and Cailleach.
|
|||
|
||||
### Using [GPT4All](https://github.com/nomic-ai/gpt4all)
|
||||
|
||||
- Obtain the `gpt4all-lora-quantized.bin` model
|
||||
- It is distributed in the old `ggml` format, which is now obsoleted
|
||||
- You have to convert it to the new format using [./convert-gpt4all-to-ggml.py](./convert-gpt4all-to-ggml.py). You may also need to
|
||||
convert the model from the old format to the new format with [./migrate-ggml-2023-03-30-pr613.py](./migrate-ggml-2023-03-30-pr613.py):
|
||||
- Obtain the `tokenizer.model` file from LLaMA model and put it to `models`
|
||||
- Obtain the `added_tokens.json` file from Alpaca model and put it to `models`
|
||||
- Obtain the `gpt4all-lora-quantized.bin` file from GPT4All model and put it to `models/gpt4all-7B`
|
||||
- It is distributed in the old `ggml` format which is now obsoleted
|
||||
- You have to convert it to the new format using `convert.py`:
|
||||
|
||||
```bash
|
||||
python3 convert-gpt4all-to-ggml.py models/gpt4all-7B/gpt4all-lora-quantized.bin ./models/tokenizer.model
|
||||
python3 migrate-ggml-2023-03-30-pr613.py models/gpt4all-7B/gpt4all-lora-quantized.bin models/gpt4all-7B/gpt4all-lora-quantized-new.bin
|
||||
```
|
||||
```bash
|
||||
python3 convert.py models/gpt4all-7B/gpt4all-lora-quantized.bin
|
||||
```
|
||||
|
||||
- You can now use the newly generated `gpt4all-lora-quantized-new.bin` model in exactly the same way as all other models
|
||||
- The original model is saved in the same folder with a suffix `.orig`
|
||||
- You can now use the newly generated `models/gpt4all-7B/ggml-model-q4_0.bin` model in exactly the same way as all other models
|
||||
|
||||
- The newer GPT4All-J model is not yet supported!
|
||||
|
||||
### Obtaining and verifying the Facebook LLaMA original model and Stanford Alpaca model data
|
||||
|
||||
|
|
|
|||
20
SHA256SUMS
20
SHA256SUMS
|
|
@ -1,12 +1,27 @@
|
|||
700df0d3013b703a806d2ae7f1bfb8e59814e3d06ae78be0c66368a50059f33d models/7B/consolidated.00.pth
|
||||
666a4bb533b303bdaf89e1b6a3b6f93535d868de31d903afdc20983dc526c847 models/7B/ggml-model-f16.bin
|
||||
fcb7664c2e69776920b526362a243e912f73c36b1ec892eb354bab940f5edb5a models/7B/ggml-model-q4_0.bin
|
||||
cc061458339a3eb8bcecbf0a825e9924fb7d1a8150f63cd5d091caa99215aafe models/7B/ggml-model-q4_1.bin
|
||||
1bc7484c24a87612726d756f1761890e7acf5f412e23378577ce50fbe789b5b8 models/7B/ggml-model-q4_2.bin
|
||||
3429bf198ec771886cf81a574df45245f3ebf04f0ce0956b73ef5d0ab01ff48b models/7B/ggml-model-q4_3.bin
|
||||
7e89e242ddc0dd6f060b43ca219ce8b3e8f08959a72cb3c0855df8bb04d46265 models/7B/params.json
|
||||
745bf4e29a4dd6f411e72976d92b452da1b49168a4f41c951cfcc8051823cf08 models/13B/consolidated.00.pth
|
||||
d5ccbcc465c71c0de439a5aeffebe8344c68a519bce70bc7f9f92654ee567085 models/13B/consolidated.01.pth
|
||||
2b206e9b21fb1076f11cafc624e2af97c9e48ea09312a0962153acc20d45f808 models/13B/ggml-model-f16.bin
|
||||
4b69e4d6b6e3275230955997b90407fceca7e5ab3daf2e63a2c9e7270a8e1e3e models/13B/ggml-model-q4_0.bin
|
||||
d9581b5b88e5622532fe897c9f9b0e67a317d22dd27a6f90fa4ab8c6d23ccdbb models/13B/ggml-model-q4_1.bin
|
||||
8d55a2077317ec9a928c7851d6a43e08e51f7e9e08360f2a7a7e1deefea3134f models/13B/ggml-model-q4_2.bin
|
||||
4208cdec9788ffa48dc1a17af2c36a0299f5bf3eb0e2b87889dda7fad591fca3 models/13B/ggml-model-q4_3.bin
|
||||
4ab77bec4d4405ccb66a97b282574c89a94417e3c32e5f68f37e2876fc21322f models/13B/params.json
|
||||
e23294a58552d8cdec5b7e8abb87993b97ea6eced4178ff2697c02472539d067 models/30B/consolidated.00.pth
|
||||
4e077b7136c7ae2302e954860cf64930458d3076fcde9443f4d0e939e95903ff models/30B/consolidated.01.pth
|
||||
24a87f01028cbd3a12de551dcedb712346c0b5cbdeff1454e0ddf2df9b675378 models/30B/consolidated.02.pth
|
||||
1adfcef71420886119544949767f6a56cb6339b4d5fcde755d80fe68b49de93b models/30B/consolidated.03.pth
|
||||
7e1b524061a9f4b27c22a12d6d2a5bf13b8ebbea73e99f218809351ed9cf7d37 models/30B/ggml-model-f16.bin
|
||||
7a679908ce31c9d6ae2e38d6059bcd4d0ad3a870cd58cc1c8f7b36f2b2f51c73 models/30B/ggml-model-q4_0.bin
|
||||
7b75ac615fa369ee593493a7e6ef87542bf0350255db928b22c5a24f6d598bcd models/30B/ggml-model-q4_1.bin
|
||||
2c82b4954a94a6a284f452f6011c1e4f0d20362c194a0b1eb5737f5fd8a20fb3 models/30B/ggml-model-q4_2.bin
|
||||
a6188660199dbcb8d5658abe7d89169869e50423494385830d9e6b330ea7fc33 models/30B/ggml-model-q4_3.bin
|
||||
2c07118ea98d69dbe7810d88520e30288fa994751b337f8fca02b171955f44cb models/30B/params.json
|
||||
135c563f6b3938114458183afb01adc9a63bef3d8ff7cccc3977e5d3664ecafe models/65B/consolidated.00.pth
|
||||
9a600b37b19d38c7e43809485f70d17d1dc12206c07efa83bc72bb498a568bde models/65B/consolidated.01.pth
|
||||
|
|
@ -16,5 +31,10 @@ e7babf7c5606f165a3756f527cb0fedc4f83e67ef1290391e52fb1cce5f26770 models/65B/con
|
|||
a287c0dfe49081626567c7fe87f74cce5831f58e459b427b5e05567641f47b78 models/65B/consolidated.05.pth
|
||||
72b4eba67a1a3b18cb67a85b70f8f1640caae9b40033ea943fb166bd80a7b36b models/65B/consolidated.06.pth
|
||||
d27f5b0677d7ff129ceacd73fd461c4d06910ad7787cf217b249948c3f3bc638 models/65B/consolidated.07.pth
|
||||
60758f2384d74e423dffddfd020ffed9d3bb186ebc54506f9c4a787d0f5367b0 models/65B/ggml-model-f16.bin
|
||||
c671fe1bce71499ac732ec999770ebe53ac486623a7891e42c9dfdb6962d2c64 models/65B/ggml-model-q4_0.bin
|
||||
4743a28aac3e5f32a6e838a815f51d3779de44fbbe251d745251e66c23c5950f models/65B/ggml-model-q4_1.bin
|
||||
4a145a210c56982389b1ed34387e0590c3e0d7325fa9be4f2284fe4d244a3633 models/65B/ggml-model-q4_2.bin
|
||||
305e91a4608b4f627b9b8ad5b4af75187d2684254bfd76dcb9db571618ef293c models/65B/ggml-model-q4_3.bin
|
||||
999ed1659b469ccc2a941714c0a9656fa571d17c9f7c8c7589817ca90edef51b models/65B/params.json
|
||||
9e556afd44213b6bd1be2b850ebbbd98f5481437a8021afaf58ee7fb1818d347 models/tokenizer.model
|
||||
|
|
|
|||
|
|
@ -7,4 +7,13 @@
|
|||
cd `dirname $0`
|
||||
cd ..
|
||||
|
||||
./main -m ./models/ggml-alpaca-7b-q4.bin --color -f ./prompts/alpaca.txt --ctx_size 2048 -n -1 -ins -b 256 --top_k 10000 --temp 0.2 --repeat_penalty 1 -t 7
|
||||
./main -m ./models/ggml-alpaca-7b-q4.bin \
|
||||
--color \
|
||||
-f ./prompts/alpaca.txt \
|
||||
--ctx_size 2048 \
|
||||
-n -1 \
|
||||
-ins -b 256 \
|
||||
--top_k 10000 \
|
||||
--temp 0.2 \
|
||||
--repeat_penalty 1.1 \
|
||||
-t 7
|
||||
|
|
|
|||
|
|
@ -20,7 +20,7 @@ struct gpt_params {
|
|||
int32_t repeat_last_n = 64; // last n tokens to penalize
|
||||
int32_t n_parts = -1; // amount of model parts (-1 = determine from model dimensions)
|
||||
int32_t n_ctx = 512; // context size
|
||||
int32_t n_batch = 8; // batch size for prompt processing
|
||||
int32_t n_batch = 512; // batch size for prompt processing (must be >=32 to use BLAS)
|
||||
int32_t n_keep = 0; // number of tokens to keep from initial prompt
|
||||
|
||||
// sampling parameters
|
||||
|
|
|
|||
|
|
@ -1,65 +1,181 @@
|
|||
# main
|
||||
# llama.cpp/example/main
|
||||
|
||||
This example shows how to run a LLaMA-like model for chatting or generating text. There are two basic modes of operation:
|
||||
This example program allows you to use various LLaMA language models in an easy and efficient way. It is specifically designed to work with the [llama.cpp](https://github.com/ggerganov/llama.cpp) project, which provides a plain C/C++ implementation with optional 4-bit quantization support for faster, lower memory inference, and is optimized for desktop CPUs. This program can be used to perform various inference tasks with LLaMA models, including generating text based on user-provided prompts and chat-like interactions with reverse prompts.
|
||||
|
||||
- Text generation: starting from an initial prompt, produce additional text as predicted by the model
|
||||
- Chat: alternate between user input to complete a prompt and generating some text based on the input
|
||||
## Table of Contents
|
||||
|
||||
## Essential parameters
|
||||
1. [Quick Start](#quick-start)
|
||||
2. [Common Options](#common-options)
|
||||
3. [Input Prompts](#input-prompts)
|
||||
4. [Interaction](#interaction)
|
||||
5. [Context Management](#context-management)
|
||||
6. [Generation Flags](#generation-flags)
|
||||
7. [Performance Tuning and Memory Options](#performance-tuning-and-memory-options)
|
||||
8. [Additional Options](#additional-options)
|
||||
|
||||
- ``--model``: the pathname of the model file to load (if not specified: ``models/llama-7B/ggml-model.bin``)
|
||||
- ``--prompt``: the first prompt used to initialize the model.
|
||||
- Alternatively, you can use ``--file`` to load the prompt from a file (UTF-8 encoded, may include line breaks).
|
||||
## Quick Start
|
||||
|
||||
**Note**: Most parameters also have short forms (such as ``-m`` for the model filepath). For clarity, all examples use the long form.
|
||||
To get started right away, run the following command, making sure to use the correct path for the model you have:
|
||||
|
||||
## Text generation
|
||||
```bash
|
||||
./main -m models/7B/ggml-model.bin --prompt "Once upon a time"
|
||||
```
|
||||
|
||||
The most basic application for an LLM is producing more text from a given prompt.
|
||||
For an interactive experience, try this command:
|
||||
|
||||
./main --model models/alpaca-7B/ggml-alpaca-7b-q4.bin --prompt "Hello World"
|
||||
```bash
|
||||
./main -m models/7B/ggml-model.bin -n -1 --color -r "User:" --in-prefix " " --prompt $'User: Hi\nAI: Hello. I am an AI chatbot. Would you like to talk?\nUser: Sure!\nAI: What would you like to talk about?\nUser:'
|
||||
```
|
||||
|
||||
This will run until the model predicts an end-of-text token, or until ``--n_predict`` tokens have been generated (a value of ``-1`` means unlimited). If you want the model to keep going without ever predicting end-of-text on its own, use the ``--ignore-eos`` parameter.
|
||||
## Common Options
|
||||
|
||||
When generating "infinite" text, the model will at some point exhaust its context size, that is, the number of past tokens it can remember. When this happens, the oldest half of the context is forgotten, and the most recent half is used to seamlessly continue the text. Text generation will temporarily
|
||||
slow down until the context is recomputed. If you want the model to remember the initial prompt (rather than just continuing from the most recent text), you can pass ``--keep -1`` on the command line (to remember the full prompt), or give a specific number of tokens to remember. In theory, this should lead to more consistency across longer texts.
|
||||
In this section, we cover the most commonly used options for running the `main` program with the LLaMA models:
|
||||
|
||||
So, a useful starting point for long text generation (using the default llama-7B model) would be:
|
||||
- `-m FNAME, --model FNAME`: Specify the path to the LLaMA model file (e.g., `models/7B/ggml-model.bin`).
|
||||
- `-i, --interactive`: Run the program in interactive mode, allowing you to provide input directly and receive real-time responses.
|
||||
- `-ins, --instruct`: Run the program in instruction mode, which is particularly useful when working with Alpaca models.
|
||||
- `-t N, --threads N`: Set the number of threads to use during computation. It is recommended to set this to the number of physical cores your CPU has.
|
||||
- `-n N, --n_predict N`: Set the number of tokens to predict when generating text. Adjusting this value can influence the length of the generated text.
|
||||
- `-c N, --ctx_size N`: Set the size of the prompt context. The default is 512, but LLaMA models were built with a context of 2048, which will provide better results for longer input/inference.
|
||||
|
||||
./main --ignore-eos --n_predict -1 --keep -1 --prompt "Hello World"
|
||||
## Input Prompts
|
||||
|
||||
## Chat mode
|
||||
The `main` program provides several ways to interact with the LLaMA models using input prompts:
|
||||
|
||||
Chat can be viewed as "co-creating" a stream of text between the user and the model: first, the context is initialized with a starting prompt (``--prompt`` or ``--file``, for example describing the character of the assistant, and giving a brief example of a typical conversation). Control is then passed to the user, who completes this prompt with their request. From this point, the model is used to generate the tokens of the reply, until the model predicts either end-of-text or emits a phrase that is defined with the ``--reverse-prompt`` parameter. This is usually a phrase such as "User:" that indicates a dialog turn:
|
||||
- `--prompt PROMPT`: Provide a prompt directly as a command-line option.
|
||||
- `--file FNAME`: Provide a file containing a prompt or multiple prompts.
|
||||
- `--interactive-first`: Run the program in interactive mode and wait for input right away. (More on this below.)
|
||||
- `--random-prompt`: Start with a randomized prompt.
|
||||
|
||||
User: Please tell me the largest city in Europe.
|
||||
Bob: Sure. The largest city in Europe is Moscow, the capital of Russia.
|
||||
User:
|
||||
## Interaction
|
||||
|
||||
There are three different chat modes, mostly with small differences in how phrases are handled that indicate conversation turns:
|
||||
The `main` program offers a seamless way to interact with LLaMA models, allowing users to engage in real-time conversations or provide instructions for specific tasks. The interactive mode can be triggered using various options, including `--interactive`, `--interactive-first`, and `--instruct`.
|
||||
|
||||
1. ``--interactive-first``: initializes the context with the starting prompt, then passes control to the user. From the user's input, keeps generating a response until a reverse prompt or end-of-text has been predicted, and then passes control back.
|
||||
- If the response ends with an end-of-text token, the user's next input is prefixed with the reverse prompt, if any, to preserve the turn structure of the dialog.
|
||||
2. ``--interactive``: same as ``--interactive-first``, but immediately generates text after the initial prompt, until a reverse prompt or end-of-text is encountered.
|
||||
3. ``--instruct``: same as ``--interactive-first``, but with the following differences:
|
||||
- When the context is full (see above), the initial prompt is remembered when switching to the new context (same as ``--keep -1``)
|
||||
- The user input is quietly prefixed with the reverse prompt (or ``### Instruction:`` as the default), and followed by ``### Response:`` (except if you just press Return without any input, to keep generating a longer response).
|
||||
In interactive mode, users can participate in text generation by injecting their input during the process. Users can press `Ctrl+C` at any time to interject and type their input, followed by pressing `Return` to submit it to the LLaMA model. To submit additional lines without finalizing input, users can end the current line with a backslash (`\`) and continue typing.
|
||||
|
||||
The ``--n_predict`` parameter limits the length of the model's responses before giving control back to the user. In *interactive* mode, this does not insert the reverse prompt before the user's input. The same happens if you use Ctrl-C to interrupt a response.
|
||||
### Interaction Options
|
||||
|
||||
If your terminal supports it, you can use ``--color`` to visually distinguish user input from model output.
|
||||
- `-i, --interactive`: Run the program in interactive mode, allowing users to engage in real-time conversations or provide specific instructions to the model.
|
||||
- `--interactive-first`: Run the program in interactive mode and immediately wait for user input before starting the text generation.
|
||||
- `-ins, --instruct`: Run the program in instruction mode, which is specifically designed to work with Alpaca models that excel in completing tasks based on user instructions.
|
||||
- `--color`: Enable colorized output to differentiate visually distinguishing between prompts, user input, and generated text.
|
||||
|
||||
Example:
|
||||
By understanding and utilizing these interaction options, you can create engaging and dynamic experiences with the LLaMA models, tailoring the text generation process to your specific needs.
|
||||
|
||||
./main --file .\prompts\chat-with-bob.txt --interactive -r "User:" --keep -1
|
||||
### Reverse Prompts
|
||||
|
||||
runs a basic chat where the prompt primes the model to expect one of the dialog participants to be named "User:", and the initial prompt is kept on context switches. In many cases, model authors suggest reasonable values for these parameters. However, the defaults are usually fine to get started.
|
||||
Reverse prompts are a powerful way to create a chat-like experience with a LLaMA model by pausing the text generation when specific text strings are encountered:
|
||||
|
||||
## Sampling parameters
|
||||
- `-r PROMPT, --reverse-prompt PROMPT`: Specify one or multiple reverse prompts to pause text generation and switch to interactive mode. For example, `-r "User:"` can be used to jump back into the conversation whenever it's the user's turn to speak. This helps create a more interactive and conversational experience. However, the reverse prompt doesn't work when it ends with a space.
|
||||
|
||||
The following parameters control sampling, that is, how a token is randomly selected from the most probable candidates that are predicted by the model. For more background, see [https://huggingface.co/blog/how-to-generate](https://huggingface.co/blog/how-to-generate).
|
||||
To overcome this limitation, you can use the `--in-prefix` flag to add a space or any other characters after the reverse prompt.
|
||||
|
||||
- ``--seed``: the starting value of the random number generator. If you use a positive value, a given prompt will produce the same output in each run. A negative value (the default) will usually produce somewhat different output in each run.
|
||||
- ``--temp``: the "temperature" parameter of the softmax function. A higher temperature means less likely words are being picked more often. Conversely, a temperature of 0 will always pick the most likely next token, leading to identical outputs in each run.
|
||||
- ``--top_k``, ``--top_p``: restrict the selection of the final token candidates to the *k* most likely, or the tokens that combine to a probability mass of at least *p*.
|
||||
- ``--repeat_last_n``, ``--repeat_penalty``: reduces repetitions within the last *n* tokes by applying a penalty to repeated tokens.
|
||||
### In-Prefix
|
||||
|
||||
The `--in-prefix` flag is used to add a prefix to your input, primarily, this is used to insert a space after the reverse prompt. Here's an example of how to use the `--in-prefix` flag in conjunction with the `--reverse-prompt` flag:
|
||||
|
||||
```sh
|
||||
./main -r "User:" --in-prefix " "
|
||||
```
|
||||
|
||||
### Instruction Mode
|
||||
|
||||
Instruction mode is particularly useful when working with Alpaca models, which are designed to follow user instructions for specific tasks:
|
||||
|
||||
- `-ins, --instruct`: Enable instruction mode to leverage the capabilities of Alpaca models in completing tasks based on user-provided instructions.
|
||||
|
||||
By understanding and utilizing these interaction options, you can create engaging and dynamic experiences with the LLaMA models, tailoring the text generation process to your specific needs.
|
||||
|
||||
## Context Management
|
||||
|
||||
During text generation, LLaMA models have a limited context size, which means they can only consider a certain number of tokens from the input and generated text. When the context fills up, the model resets internally, potentially losing some information from the beginning of the conversation or instructions. Context management options help maintain continuity and coherence in these situations.
|
||||
|
||||
### Context Size
|
||||
|
||||
The `--ctx_size` option allows you to set the size of the prompt context used by the LLaMA models during text generation. A larger context size helps the model to better comprehend and generate responses for longer input or conversations.
|
||||
|
||||
- `-c N, --ctx_size N`: Set the size of the prompt context (default: 512). The LLaMA models were built with a context of 2048, which will yield the best results on longer input/inference. However, increasing the context size beyond 2048 may lead to unpredictable results.
|
||||
|
||||
### Keep Prompt
|
||||
|
||||
The `--keep` option allows users to retain the original prompt when the model runs out of context, ensuring a connection to the initial instruction or conversation topic is maintained.
|
||||
|
||||
- `--keep N`: Specify the number of tokens from the initial prompt to retain when the model resets its internal context. By default, this value is set to 0 (meaning no tokens are kept). Use `-1` to retain all tokens from the initial prompt.
|
||||
|
||||
By utilizing context management options like `--ctx_size` and `--keep`, you can maintain a more coherent and consistent interaction with the LLaMA models, ensuring that the generated text remains relevant to the original prompt or conversation.
|
||||
|
||||
## Generation Flags
|
||||
|
||||
The following options are related to controlling the text generation process, influencing the diversity, creativity, and quality of the generated text. Understanding these options will help you fine-tune the output according to your needs:
|
||||
|
||||
### Number of Tokens to Predict
|
||||
|
||||
- `-n N, --n_predict N`: Set the number of tokens to predict when generating text (default: 128, -1 = infinity).
|
||||
|
||||
The `--n_predict` option controls the number of tokens the model generates in response to the input prompt. By adjusting this value, you can influence the length of the generated text. A higher value will result in longer text, while a lower value will produce shorter text. A value of -1 will cause text to be generated without limit.
|
||||
|
||||
It is important to note that the generated text may be shorter than the specified number of tokens if an End-of-Sequence (EOS) token or a reverse prompt is encountered. In interactive mode text generation will pause and control will be returned to the user. In non-interactive mode, the program will end. In both cases, the text generation may stop before reaching the specified `n_predict` value.
|
||||
|
||||
### RNG Seed
|
||||
|
||||
- `-s SEED, --seed SEED`: Set the random number generator (RNG) seed (default: -1).
|
||||
|
||||
The RNG seed is used to initialize the random number generator that influences the text generation process. By setting a specific seed value, you can obtain consistent and reproducible results across multiple runs with the same input and settings. This can be helpful for testing, debugging, or comparing the effects of different options on the generated text to see when they diverge. If the seed is set to a value less than or equal to 0, a random seed will be used, which will result in different outputs on each run.
|
||||
|
||||
### Temperature
|
||||
|
||||
- `--temp N`: Adjust the randomness of the generated text (default: 0.8).
|
||||
|
||||
Temperature is a hyperparameter that controls the randomness of the generated text. It affects the probability distribution of the model's output tokens. A higher temperature (e.g., 1.5) makes the output more random and creative, while a lower temperature (e.g., 0.5) makes the output more focused, deterministic, and conservative. The default value is 0.8, which provides a balance between randomness and determinism.
|
||||
|
||||
Example usage: `--temp 0.8`
|
||||
|
||||
### Repeat Penalty
|
||||
|
||||
- `--repeat_penalty N`: Control the repetition of token sequences in the generated text (default: 1.1).
|
||||
|
||||
Repeat penalty is a hyperparameter used to penalize the repetition of token sequences during text generation. It helps prevent the model from generating repetitive or monotonous text. A higher value (e.g., 1.5) will penalize repetitions more strongly, while a lower value (e.g., 0.9) will be more lenient. The default value is 1.1.
|
||||
|
||||
Example usage: `--repeat_penalty 1.1`
|
||||
|
||||
### Top-K Sampling
|
||||
|
||||
- `--top_k N`: Limit the next token selection to the K most probable tokens (default: 40).
|
||||
|
||||
Top-k sampling is a text generation method that selects the next token only from the top k most likely tokens predicted by the model. It helps reduce the risk of generating low-probability or nonsensical tokens, but it may also limit the diversity of the output. A higher value for top_k (e.g., 100) will consider more tokens and lead to more diverse text, while a lower value (e.g., 10) will focus on the most probable tokens and generate more conservative text. The default value is 40.
|
||||
|
||||
Example usage: `--top_k 40`
|
||||
|
||||
### Top-P Sampling
|
||||
|
||||
- `--top_p N`: Limit the next token selection to a subset of tokens with a cumulative probability above a threshold P (default: 0.9).
|
||||
|
||||
Top-p sampling, also known as nucleus sampling, is another text generation method that selects the next token from a subset of tokens that together have a cumulative probability of at least p. This method provides a balance between diversity and quality by considering both the probabilities of tokens and the number of tokens to sample from. A higher value for top_p (e.g., 0.95) will lead to more diverse text, while a lower value (e.g., 0.5) will generate more focused and conservative text. The default value is 0.9.
|
||||
|
||||
Example usage: `--top_p 0.9`
|
||||
|
||||
By adjusting these options, you can control the diversity, quality, and creativity of the generated text to better suit your needs. You can experiment with different combinations of values to find the best settings for your specific use case.
|
||||
|
||||
## Performance Tuning and Memory Options
|
||||
|
||||
These options help improve the performance and memory usage of the LLaMA models:
|
||||
|
||||
- `-t N, --threads N`: Set the number of threads to use during computation. Using the correct number of threads can greatly improve performance. It is recommended to set this value to the number of CPU cores.
|
||||
- `--mlock`: Lock the model in memory, preventing it from being swapped out when mmaped. This can improve performance.
|
||||
- `--no-mmap`: Do not memory-map the model. This results in a slower load time but may reduce pageouts if you're not using `mlock`.
|
||||
- `--memory_f32`: Use 32 bit floats instead of 16 bit floats for memory key+value, allowing higher quality inference at the cost of memory.
|
||||
- `-b N, --batch_size N`: Set the batch size for prompt processing (default: 512). This large batch size benefits users who have BLAS installed and enabled it during the build. If you don't have BLAS enabled ("BLAS=0"), you can use a smaller number, such as 8, to see the prompt progress as it's evaluated in some situations.
|
||||
|
||||
For information about 4-bit quantization, which can significantly improve performance and reduce memory usage, please refer to llama.cpp's primary [README](../../README.md#prepare-data--run).
|
||||
|
||||
By understanding and using these performance tuning settings, you can optimize the LLaMA model's behavior to achieve the best performance for your specific needs.
|
||||
|
||||
## Additional Options
|
||||
|
||||
These options provide extra functionality and customization when running the LLaMA models:
|
||||
|
||||
- `-h, --help`: Display a help message showing all available options and their default values. This is particularly useful for checking the latest options and default values, as they can change frequently, and the information in this document may become outdated.
|
||||
- `--verbose-prompt`: Print the prompt before generating text.
|
||||
- `--mtest`: Test the model's functionality by running a series of tests to ensure it's working properly.
|
||||
- `--lora FNAME`: Apply a LoRA (Layer-wise Relevance Approximation) adapter to the model (implies --no-mmap). This allows you to adapt the pretrained model to specific tasks or domains.
|
||||
- `--lora-base FNAME`: Optional model to use as a base for the layers modified by the LoRA adapter. This flag is used in conjunction with the `--lora` flag, and specifies the base model for the adaptation.
|
||||
|
|
|
|||
|
|
@ -25,6 +25,7 @@
|
|||
#endif
|
||||
|
||||
static console_state con_st;
|
||||
static llama_context ** g_ctx;
|
||||
|
||||
static bool is_interacting = false;
|
||||
|
||||
|
|
@ -36,6 +37,7 @@ void sigint_handler(int signo) {
|
|||
if (!is_interacting) {
|
||||
is_interacting=true;
|
||||
} else {
|
||||
llama_print_timings(*g_ctx);
|
||||
_exit(130);
|
||||
}
|
||||
}
|
||||
|
|
@ -94,6 +96,7 @@ int main(int argc, char ** argv) {
|
|||
//bool is_prime(int n) {)";
|
||||
|
||||
llama_context * ctx;
|
||||
g_ctx = &ctx;
|
||||
|
||||
// load the model
|
||||
{
|
||||
|
|
@ -264,7 +267,7 @@ int main(int argc, char ** argv) {
|
|||
// infinite text generation via context swapping
|
||||
// if we run out of context:
|
||||
// - take the n_keep first tokens from the original prompt (via n_past)
|
||||
// - take half of the last (n_ctx - n_keep) tokens and recompute the logits in a batch
|
||||
// - take half of the last (n_ctx - n_keep) tokens and recompute the logits in batches
|
||||
if (n_past + (int) embd.size() > n_ctx) {
|
||||
const int n_left = n_past - params.n_keep;
|
||||
|
||||
|
|
@ -282,13 +285,21 @@ int main(int argc, char ** argv) {
|
|||
//printf("\n---\n");
|
||||
}
|
||||
|
||||
if (llama_eval(ctx, embd.data(), embd.size(), n_past, params.n_threads)) {
|
||||
fprintf(stderr, "%s : failed to eval\n", __func__);
|
||||
return 1;
|
||||
// evaluate tokens in batches
|
||||
// embd is typically prepared beforehand to fit within a batch, but not always
|
||||
for (int i = 0; i < (int) embd.size(); i += params.n_batch) {
|
||||
int n_eval = (int) embd.size() - i;
|
||||
if (n_eval > params.n_batch) {
|
||||
n_eval = params.n_batch;
|
||||
}
|
||||
if (llama_eval(ctx, &embd[i], n_eval, n_past, params.n_threads)) {
|
||||
fprintf(stderr, "%s : failed to eval\n", __func__);
|
||||
return 1;
|
||||
}
|
||||
n_past += n_eval;
|
||||
}
|
||||
}
|
||||
|
||||
n_past += embd.size();
|
||||
embd.clear();
|
||||
|
||||
if ((int) embd_inp.size() <= n_consumed && !is_interacting) {
|
||||
|
|
|
|||
|
|
@ -53,7 +53,13 @@ void perplexity(llama_context * ctx, const gpt_params & params) {
|
|||
auto end_t = std::chrono::high_resolution_clock::now();
|
||||
if (i == 0) {
|
||||
const float seconds = std::chrono::duration<float>(end_t - start_t).count();
|
||||
printf("%.2f seconds per pass - ETA %.2f hours\n", seconds, (seconds * seq_count) / (60.0*60.0));
|
||||
printf("%.2f seconds per pass - ETA ", seconds);
|
||||
int total_seconds = (int)(seconds * seq_count);
|
||||
if (total_seconds >= 60*60) {
|
||||
printf("%d hours ", total_seconds / (60*60));
|
||||
total_seconds = total_seconds % (60*60);
|
||||
}
|
||||
printf("%d minutes\n", total_seconds / 60);
|
||||
}
|
||||
// We get the logits for all the tokens in the context window (params.n_ctx)
|
||||
// from llama_eval above. Now, based on https://huggingface.co/docs/transformers/perplexity,
|
||||
|
|
|
|||
116
ggml-cuda.cu
116
ggml-cuda.cu
|
|
@ -1,5 +1,7 @@
|
|||
#include <stdint.h>
|
||||
#include <stdio.h>
|
||||
#include <cuda_fp16.h>
|
||||
#include <atomic>
|
||||
#include "ggml-cuda.h"
|
||||
|
||||
typedef uint16_t ggml_fp16_t;
|
||||
|
|
@ -29,14 +31,12 @@ static_assert(sizeof(block_q4_2) == sizeof(ggml_fp16_t) + QK4_2 / 2, "wrong q4_2
|
|||
|
||||
#define QK4_3 16
|
||||
typedef struct {
|
||||
__half d; // delta
|
||||
__half m; // min
|
||||
uint8_t qs[QK4_3 / 2]; // nibbles / quants
|
||||
__half d; // delta
|
||||
__half m; // min
|
||||
uint8_t qs[QK4_3 / 2]; // nibbles / quants
|
||||
} block_q4_3;
|
||||
static_assert(sizeof(block_q4_3) == 2 * sizeof(ggml_fp16_t) + QK4_3 / 2, "wrong q4_3 block size/padding");
|
||||
|
||||
|
||||
|
||||
static __global__ void dequantize_block_q4_0(const void * vx, float * y) {
|
||||
const block_q4_0 * x = (const block_q4_0 *) vx;
|
||||
|
||||
|
|
@ -131,24 +131,98 @@ static __global__ void dequantize_block_q4_3(const void * vx, float * y) {
|
|||
}
|
||||
}
|
||||
|
||||
extern "C" {
|
||||
__host__ void dequantize_row_q4_0_cuda(const void * vx, float * y, int k, cudaStream_t stream) {
|
||||
const int nb = k / QK4_0;
|
||||
dequantize_block_q4_0<<<nb, 1, 0, stream>>>(vx, y);
|
||||
}
|
||||
void dequantize_row_q4_0_cuda(const void * vx, float * y, int k, cudaStream_t stream) {
|
||||
const int nb = k / QK4_0;
|
||||
dequantize_block_q4_0<<<nb, 1, 0, stream>>>(vx, y);
|
||||
}
|
||||
|
||||
__host__ void dequantize_row_q4_1_cuda(const void * vx, float * y, int k, cudaStream_t stream) {
|
||||
const int nb = k / QK4_1;
|
||||
dequantize_block_q4_1<<<nb, 1, 0, stream>>>(vx, y);
|
||||
}
|
||||
void dequantize_row_q4_1_cuda(const void * vx, float * y, int k, cudaStream_t stream) {
|
||||
const int nb = k / QK4_1;
|
||||
dequantize_block_q4_1<<<nb, 1, 0, stream>>>(vx, y);
|
||||
}
|
||||
|
||||
__host__ void dequantize_row_q4_2_cuda(const void * vx, float * y, int k, cudaStream_t stream) {
|
||||
const int nb = k / QK4_2;
|
||||
dequantize_block_q4_2<<<nb, 1, 0, stream>>>(vx, y);
|
||||
}
|
||||
void dequantize_row_q4_2_cuda(const void * vx, float * y, int k, cudaStream_t stream) {
|
||||
const int nb = k / QK4_2;
|
||||
dequantize_block_q4_2<<<nb, 1, 0, stream>>>(vx, y);
|
||||
}
|
||||
|
||||
__host__ void dequantize_row_q4_3_cuda(const void * vx, float * y, int k, cudaStream_t stream) {
|
||||
const int nb = k / QK4_3;
|
||||
dequantize_block_q4_3<<<nb, 1, 0, stream>>>(vx, y);
|
||||
void dequantize_row_q4_3_cuda(const void * vx, float * y, int k, cudaStream_t stream) {
|
||||
const int nb = k / QK4_3;
|
||||
dequantize_block_q4_3<<<nb, 1, 0, stream>>>(vx, y);
|
||||
}
|
||||
|
||||
// buffer pool for cuda
|
||||
#define MAX_CUDA_BUFFERS 16
|
||||
|
||||
struct scoped_spin_lock {
|
||||
std::atomic_flag& lock;
|
||||
scoped_spin_lock(std::atomic_flag& lock) : lock(lock) {
|
||||
while (lock.test_and_set(std::memory_order_acquire)) {
|
||||
; // spin
|
||||
}
|
||||
}
|
||||
~scoped_spin_lock() {
|
||||
lock.clear(std::memory_order_release);
|
||||
}
|
||||
scoped_spin_lock(const scoped_spin_lock&) = delete;
|
||||
scoped_spin_lock& operator=(const scoped_spin_lock&) = delete;
|
||||
};
|
||||
|
||||
struct cuda_buffer {
|
||||
void * ptr = nullptr;
|
||||
size_t size = 0;
|
||||
};
|
||||
|
||||
static cuda_buffer g_cuda_buffer_pool[MAX_CUDA_BUFFERS];
|
||||
static std::atomic_flag g_cuda_pool_lock = ATOMIC_FLAG_INIT;
|
||||
|
||||
void * ggml_cuda_pool_malloc(size_t size, size_t * actual_size) {
|
||||
scoped_spin_lock lock(g_cuda_pool_lock);
|
||||
|
||||
for (int i = 0; i < MAX_CUDA_BUFFERS; ++i) {
|
||||
cuda_buffer& b = g_cuda_buffer_pool[i];
|
||||
if (b.size >= size && b.ptr != nullptr) {
|
||||
void * ptr = b.ptr;
|
||||
*actual_size = b.size;
|
||||
b.ptr = nullptr;
|
||||
b.size = 0;
|
||||
return ptr;
|
||||
}
|
||||
}
|
||||
void * ptr;
|
||||
CUDA_CHECK(cudaMalloc((void **) &ptr, size));
|
||||
*actual_size = size;
|
||||
return ptr;
|
||||
}
|
||||
|
||||
void ggml_cuda_pool_free(void * ptr, size_t size) {
|
||||
scoped_spin_lock lock(g_cuda_pool_lock);
|
||||
|
||||
for (int i = 0; i < MAX_CUDA_BUFFERS; ++i) {
|
||||
cuda_buffer& b = g_cuda_buffer_pool[i];
|
||||
if (b.ptr == nullptr) {
|
||||
b.ptr = ptr;
|
||||
b.size = size;
|
||||
return;
|
||||
}
|
||||
}
|
||||
fprintf(stderr, "WARNING: cuda buffer pool full, increase MAX_CUDA_BUFFERS\n");
|
||||
CUDA_CHECK(cudaFree(ptr));
|
||||
}
|
||||
|
||||
cublasHandle_t g_cublasH = NULL;
|
||||
cudaStream_t g_cudaStream = NULL;
|
||||
|
||||
void ggml_init_cublas(void) {
|
||||
if (g_cublasH == NULL) {
|
||||
// create cublas handle, bind a stream
|
||||
CUBLAS_CHECK(cublasCreate(&g_cublasH));
|
||||
|
||||
CUDA_CHECK(cudaStreamCreateWithFlags(&g_cudaStream, cudaStreamNonBlocking));
|
||||
|
||||
CUBLAS_CHECK(cublasSetStream(g_cublasH, g_cudaStream));
|
||||
|
||||
// configure logging to stdout
|
||||
// CUBLAS_CHECK(cublasLoggerConfigure(1, 1, 0, NULL));
|
||||
}
|
||||
}
|
||||
|
|
|
|||
29
ggml-cuda.h
29
ggml-cuda.h
|
|
@ -1,7 +1,36 @@
|
|||
#include <cublas_v2.h>
|
||||
#include <cuda_runtime.h>
|
||||
|
||||
#ifdef __cplusplus
|
||||
extern "C" {
|
||||
#endif
|
||||
|
||||
#define CUDA_CHECK(err) \
|
||||
do { \
|
||||
cudaError_t err_ = (err); \
|
||||
if (err_ != cudaSuccess) { \
|
||||
fprintf(stderr, "CUDA error %d at %s:%d: %s\n", err_, __FILE__, __LINE__, \
|
||||
cudaGetErrorString(err_)); \
|
||||
exit(1); \
|
||||
} \
|
||||
} while (0)
|
||||
|
||||
#define CUBLAS_CHECK(err) \
|
||||
do { \
|
||||
cublasStatus_t err_ = (err); \
|
||||
if (err_ != CUBLAS_STATUS_SUCCESS) { \
|
||||
fprintf(stderr, "cuBLAS error %d at %s:%d\n", err_, __FILE__, __LINE__); \
|
||||
exit(1); \
|
||||
} \
|
||||
} while (0)
|
||||
|
||||
extern cublasHandle_t g_cublasH;
|
||||
extern cudaStream_t g_cudaStream;
|
||||
|
||||
void ggml_init_cublas(void);
|
||||
void * ggml_cuda_pool_malloc(size_t size, size_t * actual_size);
|
||||
void ggml_cuda_pool_free(void * ptr, size_t size);
|
||||
|
||||
void dequantize_row_q4_0_cuda(const void * vx, float * y, int k, cudaStream_t stream);
|
||||
void dequantize_row_q4_1_cuda(const void * vx, float * y, int k, cudaStream_t stream);
|
||||
void dequantize_row_q4_2_cuda(const void * vx, float * y, int k, cudaStream_t stream);
|
||||
|
|
|
|||
3
ggml.h
3
ggml.h
|
|
@ -630,7 +630,8 @@ struct ggml_tensor * ggml_soft_max(
|
|||
|
||||
// rotary position embedding
|
||||
// in-place, returns view(a)
|
||||
// if mode == 1, skip n_past elements
|
||||
// if mode & 1 == 1, skip n_past elements
|
||||
// if mode & 2 == 1, GPT-NeoX style
|
||||
// TODO: avoid creating a new tensor every time
|
||||
struct ggml_tensor * ggml_rope(
|
||||
struct ggml_context * ctx,
|
||||
|
|
|
|||
140
llama.cpp
140
llama.cpp
|
|
@ -27,6 +27,7 @@
|
|||
#include <thread>
|
||||
#include <atomic>
|
||||
#include <mutex>
|
||||
#include <sstream>
|
||||
|
||||
#define LLAMA_USE_SCRATCH
|
||||
#define LLAMA_MAX_SCRATCH_BUFFERS 16
|
||||
|
|
@ -67,7 +68,7 @@ static const std::map<e_model, size_t> & MEM_REQ_SCRATCH1()
|
|||
{ MODEL_65B, 512ull * MB },
|
||||
};
|
||||
return _MEM_REQ_SCRATCH1;
|
||||
};
|
||||
}
|
||||
|
||||
// 2*n_embd*n_ctx*n_layer*sizeof(float16)
|
||||
static const std::map<e_model, size_t> & MEM_REQ_KV_SELF()
|
||||
|
|
@ -79,7 +80,7 @@ static const std::map<e_model, size_t> & MEM_REQ_KV_SELF()
|
|||
{ MODEL_65B, 5120ull * MB },
|
||||
};
|
||||
return _MEM_REQ_KV_SELF;
|
||||
};
|
||||
}
|
||||
|
||||
// this is mostly needed for temporary mul_mat buffers to dequantize the data
|
||||
// not actually needed if BLAS is disabled
|
||||
|
|
@ -92,7 +93,7 @@ static const std::map<e_model, size_t> & MEM_REQ_EVAL()
|
|||
{ MODEL_65B, 1536ull * MB },
|
||||
};
|
||||
return _MEM_REQ_EVAL;
|
||||
};
|
||||
}
|
||||
|
||||
// default hparams (LLaMA 7B)
|
||||
struct llama_hparams {
|
||||
|
|
@ -1249,9 +1250,11 @@ static bool llama_eval_internal(
|
|||
ggml_build_forward_expand(&gf, inpL);
|
||||
ggml_graph_compute (ctx0, &gf);
|
||||
|
||||
#ifdef GGML_PERF
|
||||
// print timing information per ggml operation (for debugging purposes)
|
||||
// requires GGML_PERF to be defined
|
||||
//ggml_graph_print(&gf);
|
||||
ggml_graph_print(&gf);
|
||||
#endif
|
||||
|
||||
// plot the computation graph in dot format (for debugging purposes)
|
||||
//if (n_past%100 == 0) {
|
||||
|
|
@ -1618,6 +1621,11 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
|
|||
// quantize only 2D tensors
|
||||
quantize &= (tensor.ne.size() == 2);
|
||||
|
||||
// uncomment this to keep the output layer in FP16
|
||||
//if (tensor.name == "output.weight") {
|
||||
// quantize = false;
|
||||
//}
|
||||
|
||||
enum ggml_type new_type;
|
||||
void * new_data;
|
||||
size_t new_size;
|
||||
|
|
@ -1782,7 +1790,7 @@ struct llama_context * llama_init_from_file(
|
|||
if (params.logits_all) {
|
||||
ctx->logits.reserve(hparams.n_ctx*hparams.n_vocab);
|
||||
} else {
|
||||
ctx->logits.reserve(hparams.n_ctx);
|
||||
ctx->logits.reserve(hparams.n_vocab);
|
||||
}
|
||||
|
||||
if (params.embedding){
|
||||
|
|
@ -2087,7 +2095,11 @@ void llama_set_kv_cache(
|
|||
int n_token_count) {
|
||||
// Make sure we have the same kv cache setup
|
||||
LLAMA_ASSERT(ctx->model.kv_self.buf.size == n_size);
|
||||
void * k_data = ctx->model.kv_self.k->data; // remember data pointers
|
||||
void * v_data = ctx->model.kv_self.v->data; // because their value is stored in buf and overwritten by memcpy
|
||||
memcpy(ctx->model.kv_self.buf.addr, kv_cache, n_size);
|
||||
ctx->model.kv_self.k->data = k_data; // restore correct data pointers
|
||||
ctx->model.kv_self.v->data = v_data;
|
||||
ctx->model.kv_self.n = n_token_count;
|
||||
}
|
||||
|
||||
|
|
@ -2243,3 +2255,121 @@ const char * llama_print_system_info(void) {
|
|||
std::vector<std::pair<std::string, struct ggml_tensor *>>& llama_internal_get_tensor_map(struct llama_context * ctx) {
|
||||
return ctx->model.tensors_by_name;
|
||||
}
|
||||
|
||||
// Returns the size of the state
|
||||
size_t llama_get_state_size(struct llama_context * ctx) {
|
||||
// we don't know size of rng until we actually serialize it. so reserve more than enough memory for its serialized state.
|
||||
// for reference, std::mt19937(1337) serializes to 6701 bytes.
|
||||
const size_t s_rng_size = sizeof(size_t);
|
||||
const size_t s_rng = 64*1024;
|
||||
const size_t s_logits_capacity = sizeof(size_t);
|
||||
const size_t s_logits_size = sizeof(size_t);
|
||||
const size_t s_logits = ctx->logits.capacity() * sizeof(float);
|
||||
const size_t s_embedding_size = sizeof(size_t);
|
||||
const size_t s_embedding = ctx->embedding.size() * sizeof(float);
|
||||
const size_t s_kv_size = sizeof(size_t);
|
||||
const size_t s_kv_ntok = sizeof(int);
|
||||
const size_t s_kv = llama_get_kv_cache_size(ctx);
|
||||
const size_t s_total = (
|
||||
+ s_rng_size
|
||||
+ s_rng
|
||||
+ s_logits_capacity
|
||||
+ s_logits_size
|
||||
+ s_logits
|
||||
+ s_embedding_size
|
||||
+ s_embedding
|
||||
+ s_kv_size
|
||||
+ s_kv_ntok
|
||||
+ s_kv
|
||||
);
|
||||
return s_total;
|
||||
}
|
||||
|
||||
// Copies the state to the specified destination address
|
||||
size_t llama_copy_state_data(struct llama_context * ctx, uint8_t * dest) {
|
||||
std::stringstream rng_ss;
|
||||
rng_ss << ctx->rng;
|
||||
const size_t rng_size = rng_ss.str().size();
|
||||
char rng_buf[64*1024];
|
||||
memset(&rng_buf[0], 0, 64*1024);
|
||||
memcpy(&rng_buf[0], rng_ss.str().data(), rng_ss.str().size());
|
||||
const size_t logits_capacity = ctx->logits.capacity();
|
||||
const size_t logits_size = ctx->logits.size();
|
||||
const size_t embedding_size = ctx->embedding.size();
|
||||
const size_t kv_size = llama_get_kv_cache_size(ctx);
|
||||
const int kv_ntok = llama_get_kv_cache_token_count(ctx);
|
||||
|
||||
uint8_t * out = dest;
|
||||
memcpy(out, &rng_size, sizeof(size_t)); out += sizeof(size_t);
|
||||
memcpy(out, &rng_buf[0], 64*1024); out += 64*1024;
|
||||
memcpy(out, &logits_capacity, sizeof(size_t)); out += sizeof(size_t);
|
||||
memcpy(out, &logits_size, sizeof(size_t)); out += sizeof(size_t);
|
||||
if (logits_size) {
|
||||
memcpy(out, ctx->logits.data(), logits_size * sizeof(float));
|
||||
}
|
||||
out += logits_capacity * sizeof(float);
|
||||
memcpy(out, &embedding_size, sizeof(size_t)); out += sizeof(size_t);
|
||||
if (embedding_size) {
|
||||
memcpy(out, ctx->embedding.data(), embedding_size * sizeof(float)); out += embedding_size * sizeof(float);
|
||||
}
|
||||
memcpy(out, &kv_size, sizeof(size_t)); out += sizeof(size_t);
|
||||
memcpy(out, &kv_ntok, sizeof(int)); out += sizeof(int);
|
||||
if (kv_size) {
|
||||
memcpy(out, llama_get_kv_cache(ctx), kv_size); out += kv_size;
|
||||
}
|
||||
const size_t written = out - dest;
|
||||
const size_t expected = llama_get_state_size(ctx);
|
||||
LLAMA_ASSERT(written == expected);
|
||||
return written;
|
||||
}
|
||||
|
||||
// Sets the state reading from the specified source address
|
||||
size_t llama_set_state_data(struct llama_context * ctx, const uint8_t * src) {
|
||||
size_t rng_size;
|
||||
char rng_buf[64*1024];
|
||||
std::stringstream rng_ss;
|
||||
|
||||
const uint8_t * in = src;
|
||||
memcpy(&rng_size, in, sizeof(size_t)); in += sizeof(size_t);
|
||||
memcpy(&rng_buf[0], in, 64*1024); in += 64*1024;
|
||||
rng_ss.str(std::string(&rng_buf[0], rng_size));
|
||||
rng_ss >> ctx->rng;
|
||||
LLAMA_ASSERT(rng_ss.fail() == false);
|
||||
|
||||
size_t logits_capacity;
|
||||
size_t logits_size;
|
||||
size_t embedding_size;
|
||||
size_t kv_size;
|
||||
int kv_ntok;
|
||||
|
||||
memcpy(&logits_capacity, in, sizeof(size_t)); in += sizeof(size_t);
|
||||
memcpy(&logits_size, in, sizeof(size_t)); in += sizeof(size_t);
|
||||
LLAMA_ASSERT(ctx->logits.capacity() == logits_capacity);
|
||||
if (logits_size) {
|
||||
ctx->logits.resize(logits_size);
|
||||
memcpy(ctx->logits.data(), in, logits_size * sizeof(float));
|
||||
}
|
||||
in += logits_capacity * sizeof(float);
|
||||
memcpy(&embedding_size, in, sizeof(size_t)); in += sizeof(size_t);
|
||||
LLAMA_ASSERT(ctx->embedding.capacity() == embedding_size);
|
||||
if (embedding_size) {
|
||||
memcpy(ctx->embedding.data(), in, embedding_size * sizeof(float));
|
||||
in += embedding_size * sizeof(float);
|
||||
}
|
||||
memcpy(&kv_size, in, sizeof(size_t)); in += sizeof(size_t);
|
||||
memcpy(&kv_ntok, in, sizeof(int)); in += sizeof(int);
|
||||
if (kv_size) {
|
||||
LLAMA_ASSERT(ctx->model.kv_self.buf.size == kv_size);
|
||||
void * k_data = ctx->model.kv_self.k->data; // remember data pointers
|
||||
void * v_data = ctx->model.kv_self.v->data; // because their value is stored in buf and overwritten by memcpy
|
||||
memcpy(ctx->model.kv_self.buf.addr, in, kv_size);
|
||||
ctx->model.kv_self.k->data = k_data; // restore correct data pointers
|
||||
ctx->model.kv_self.v->data = v_data;
|
||||
in += kv_size;
|
||||
}
|
||||
ctx->model.kv_self.n = kv_ntok;
|
||||
const size_t nread = in - src;
|
||||
const size_t expected = llama_get_state_size(ctx);
|
||||
LLAMA_ASSERT(nread == expected);
|
||||
return nread;
|
||||
}
|
||||
|
|
|
|||
12
llama.h
12
llama.h
|
|
@ -129,6 +129,18 @@ extern "C" {
|
|||
size_t n_size,
|
||||
int n_token_count);
|
||||
|
||||
// Returns the size in bytes of the state (rng, logits, embedding and kv_cache)
|
||||
LLAMA_API size_t llama_get_state_size(struct llama_context * ctx);
|
||||
|
||||
// Copies the state to the specified destination address.
|
||||
// Destination needs to have allocated enough memory.
|
||||
// Returns the number of bytes copied
|
||||
LLAMA_API size_t llama_copy_state_data(struct llama_context * ctx, uint8_t * dest);
|
||||
|
||||
// Set the state reading from the specified address
|
||||
// Returns the number of bytes read
|
||||
LLAMA_API size_t llama_set_state_data(struct llama_context * ctx, const uint8_t * src);
|
||||
|
||||
// Run the llama inference to obtain the logits and probabilities for the next token.
|
||||
// tokens + n_tokens is the provided batch of new tokens to process
|
||||
// n_past is the number of tokens to use from previous eval calls
|
||||
|
|
|
|||
17
llama_util.h
17
llama_util.h
|
|
@ -21,6 +21,9 @@
|
|||
#if defined(_POSIX_MAPPED_FILES)
|
||||
#include <sys/mman.h>
|
||||
#endif
|
||||
#if defined(_POSIX_MEMLOCK_RANGE)
|
||||
#include <sys/resource.h>
|
||||
#endif
|
||||
#endif
|
||||
#endif
|
||||
|
||||
|
|
@ -303,8 +306,18 @@ struct llama_mlock {
|
|||
if (!mlock(addr, size)) {
|
||||
return true;
|
||||
} else {
|
||||
fprintf(stderr, "warning: failed to mlock %zu-byte buffer (after previously locking %zu bytes): %s\n" MLOCK_SUGGESTION,
|
||||
size, this->size, std::strerror(errno));
|
||||
char* errmsg = std::strerror(errno);
|
||||
bool suggest = (errno == ENOMEM);
|
||||
|
||||
// Check if the resource limit is fine after all
|
||||
struct rlimit lock_limit;
|
||||
if (suggest && getrlimit(RLIMIT_MEMLOCK, &lock_limit))
|
||||
suggest = false;
|
||||
if (suggest && (lock_limit.rlim_max > lock_limit.rlim_cur + size))
|
||||
suggest = false;
|
||||
|
||||
fprintf(stderr, "warning: failed to mlock %zu-byte buffer (after previously locking %zu bytes): %s\n%s",
|
||||
size, this->size, errmsg, suggest ? MLOCK_SUGGESTION : "");
|
||||
return false;
|
||||
}
|
||||
}
|
||||
|
|
|
|||
|
|
@ -2,3 +2,8 @@ set(TARGET vdot)
|
|||
add_executable(${TARGET} vdot.cpp)
|
||||
target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
|
||||
target_compile_features(${TARGET} PRIVATE cxx_std_11)
|
||||
|
||||
set(TARGET q8dot)
|
||||
add_executable(${TARGET} q8dot.cpp)
|
||||
target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
|
||||
target_compile_features(${TARGET} PRIVATE cxx_std_11)
|
||||
|
|
|
|||
172
pocs/vdot/q8dot.cpp
Normal file
172
pocs/vdot/q8dot.cpp
Normal file
|
|
@ -0,0 +1,172 @@
|
|||
#include <cstdio>
|
||||
#include <type_traits>
|
||||
#include <vector>
|
||||
#include <random>
|
||||
#include <chrono>
|
||||
#include <cstdlib>
|
||||
#include <cmath>
|
||||
#include <cassert>
|
||||
#include <cstring>
|
||||
#include <array>
|
||||
#include <type_traits>
|
||||
|
||||
#include <ggml.h>
|
||||
|
||||
constexpr int kVecSize = 1 << 16;
|
||||
|
||||
// Copy-pasted from ggml.c
|
||||
#define QK4_0 32
|
||||
typedef struct {
|
||||
float d; // delta
|
||||
uint8_t qs[QK4_0 / 2]; // nibbles / quants
|
||||
} block_q4_0;
|
||||
static_assert(sizeof(block_q4_0) == sizeof(float) + QK4_0 / 2, "wrong q4_0 block size/padding");
|
||||
|
||||
#define QK4_1 32
|
||||
typedef struct {
|
||||
float d; // delta
|
||||
float m; // min
|
||||
uint8_t qs[QK4_1 / 2]; // nibbles / quants
|
||||
} block_q4_1;
|
||||
static_assert(sizeof(block_q4_1) == sizeof(float) * 2 + QK4_1 / 2, "wrong q4_1 block size/padding");
|
||||
|
||||
// Copy-pasted from ggml.c
|
||||
#define QK8_0 32
|
||||
typedef struct {
|
||||
float d; // delta
|
||||
float s; // d * sum(qs[i])
|
||||
int8_t qs[QK8_0]; // quants
|
||||
} block_q8_0;
|
||||
static_assert(sizeof(block_q8_0) == 2*sizeof(float) + QK8_0, "wrong q8_0 block size/padding");
|
||||
|
||||
static_assert(QK4_1 == QK8_0, "QK4_1 and QK8_0 must be the same");
|
||||
static_assert(QK4_0 == QK8_0, "QK4_0 and QK8_0 must be the same");
|
||||
|
||||
template <typename T>
|
||||
void fillQ4blocks(std::vector<T>& blocks, std::mt19937& rndm) {
|
||||
for (auto& b : blocks) {
|
||||
b.d = 1;
|
||||
for (int i=0; i<QK4_1/2; ++i) {
|
||||
uint8_t v1 = rndm() >> 28;
|
||||
uint8_t v2 = rndm() >> 28;
|
||||
b.qs[i] = v1 | (v2 << 4);
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
void fillQ80blocks(std::vector<block_q8_0>& blocks, std::mt19937& rndm) {
|
||||
for (auto& b : blocks) {
|
||||
b.d = 1;
|
||||
int sum = 0;
|
||||
for (int i=0; i<QK8_0; ++i) {
|
||||
b.qs[i] = (rndm() >> 24) - 128;
|
||||
sum += b.qs[i];
|
||||
}
|
||||
b.s = b.d * sum;
|
||||
}
|
||||
}
|
||||
|
||||
float simpleDot(const block_q4_0& x, const block_q8_0& y) {
|
||||
int s1 = 0; //, s2 = 0;
|
||||
for (int i=0; i<QK4_1/2; i+=2) {
|
||||
int v1 = x.qs[i+0] & 0xf;
|
||||
int v2 = x.qs[i+0] >> 4;
|
||||
int v3 = x.qs[i+1] & 0xf;
|
||||
int v4 = x.qs[i+1] >> 4;
|
||||
int j = 2*i;
|
||||
s1 += v1*y.qs[j] + v2*y.qs[j+1] + v3*y.qs[j+2] + v4*y.qs[j+3];
|
||||
//s2 += y.qs[j] + y.qs[j+1] + y.qs[j+2] + y.qs[j+3];
|
||||
}
|
||||
return y.d * x.d * s1 - 8 * x.d * y.s;
|
||||
//return y.d * x.d * (s1 - 8 * s2);
|
||||
}
|
||||
|
||||
float simpleDot(const block_q4_1& x, const block_q8_0& y) {
|
||||
int s1 = 0; //, s2 = 0;
|
||||
for (int i=0; i<QK4_1/2; i+=2) {
|
||||
int v1 = x.qs[i+0] & 0xf;
|
||||
int v2 = x.qs[i+0] >> 4;
|
||||
int v3 = x.qs[i+1] & 0xf;
|
||||
int v4 = x.qs[i+1] >> 4;
|
||||
int j = 2*i;
|
||||
s1 += v1*y.qs[j] + v2*y.qs[j+1] + v3*y.qs[j+2] + v4*y.qs[j+3];
|
||||
//s2 += y.qs[j] + y.qs[j+1] + y.qs[j+2] + y.qs[j+3];
|
||||
}
|
||||
return y.d * x.d * s1 + y.s * x.m;
|
||||
//return y.d * (x.d * s1 + x.m * s2);
|
||||
}
|
||||
|
||||
struct Stat {
|
||||
double sum = 0, sumt = 0, sumt2 = 0, maxt = 0;
|
||||
int nloop = 0;
|
||||
void addResult(double s, double t) {
|
||||
sum += s;
|
||||
sumt += t; sumt2 += t*t; maxt = std::max(maxt, t);
|
||||
++nloop;
|
||||
}
|
||||
void reportResult(const char* title) const {
|
||||
if (nloop < 1) {
|
||||
printf("%s(%s): no result\n",__func__,title);
|
||||
return;
|
||||
}
|
||||
printf("============ %s\n",title);
|
||||
printf("<dot> = %g\n",sum/nloop);
|
||||
auto t = sumt/nloop, dt = sumt2/nloop - t*t;
|
||||
if (dt > 0) dt = sqrt(dt);
|
||||
printf("<time> = %g +/- %g us. Max. time = %g us.\n",t,dt,maxt);
|
||||
}
|
||||
};
|
||||
|
||||
|
||||
int main(int argc, char** argv) {
|
||||
|
||||
int nloop = argc > 1 ? atoi(argv[1]) : 10;
|
||||
int type = argc > 2 ? atoi(argv[2]) : 1;
|
||||
|
||||
std::mt19937 rndm(1234);
|
||||
|
||||
std::vector<block_q4_1> x41;
|
||||
std::vector<block_q4_0> x40;
|
||||
std::vector<block_q8_0> y(kVecSize);
|
||||
if (type == 0) x40.resize(kVecSize);
|
||||
else {
|
||||
x41.resize(kVecSize);
|
||||
for (auto& b : x41) b.m = 1;
|
||||
}
|
||||
|
||||
auto ggml_type = type == 0 ? GGML_TYPE_Q4_0 : GGML_TYPE_Q4_1;
|
||||
|
||||
auto funcs = ggml_internal_get_quantize_fn(ggml_type);
|
||||
|
||||
Stat simple, ggml;
|
||||
|
||||
for (int iloop=0; iloop<nloop; ++iloop) {
|
||||
|
||||
if (type == 0) fillQ4blocks(x40, rndm);
|
||||
else fillQ4blocks(x41, rndm);
|
||||
fillQ80blocks(y, rndm);
|
||||
|
||||
auto t1 = std::chrono::high_resolution_clock::now();
|
||||
double s = 0;
|
||||
if (type == 0) for (int i=0; i<kVecSize; ++i) s += simpleDot(x40[i], y[i]);
|
||||
else for (int i=0; i<kVecSize; ++i) s += simpleDot(x41[i], y[i]);
|
||||
auto t2 = std::chrono::high_resolution_clock::now();
|
||||
auto t = 1e-3*std::chrono::duration_cast<std::chrono::nanoseconds>(t2-t1).count();
|
||||
if (iloop > 3) simple.addResult(s, t);
|
||||
|
||||
t1 = std::chrono::high_resolution_clock::now();
|
||||
float fs;
|
||||
if (type == 0) funcs.vec_dot_q(kVecSize * QK4_1, &fs, x40.data(), y.data());
|
||||
else funcs.vec_dot_q(kVecSize * QK4_1, &fs, x41.data(), y.data());
|
||||
t2 = std::chrono::high_resolution_clock::now();
|
||||
t = 1e-3*std::chrono::duration_cast<std::chrono::nanoseconds>(t2-t1).count();
|
||||
if (iloop > 3) ggml.addResult(fs, t);
|
||||
|
||||
}
|
||||
|
||||
// Report the time (and the average of the dot products so the compiler does not come up with the idea
|
||||
// of optimizing away the function calls after figuring that the result is not used).
|
||||
simple.reportResult("Simple");
|
||||
ggml.reportResult("ggml");
|
||||
return 0;
|
||||
}
|
||||
6
scripts/sync-ggml.sh
Executable file
6
scripts/sync-ggml.sh
Executable file
|
|
@ -0,0 +1,6 @@
|
|||
#!/bin/bash
|
||||
|
||||
cp -rpv ../ggml/src/ggml.c ./ggml.c
|
||||
cp -rpv ../ggml/src/ggml-cuda.cu ./ggml-cuda.cu
|
||||
cp -rpv ../ggml/src/ggml-cuda.h ./ggml-cuda.h
|
||||
cp -rpv ../ggml/include/ggml/ggml.h ./ggml.h
|
||||
|
|
@ -6,5 +6,6 @@ function(llama_add_test source)
|
|||
endfunction()
|
||||
|
||||
# llama_add_test(test-double-float.c) # SLOW
|
||||
llama_add_test(test-quantize.c)
|
||||
llama_add_test(test-quantize-fns.cpp)
|
||||
llama_add_test(test-quantize-perf.cpp)
|
||||
llama_add_test(test-tokenizer-0.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab.bin)
|
||||
|
|
|
|||
154
tests/test-quantize-fns.cpp
Normal file
154
tests/test-quantize-fns.cpp
Normal file
|
|
@ -0,0 +1,154 @@
|
|||
// Unit tests for quantization specific functions - quantize, dequantize and dot product
|
||||
|
||||
#include "ggml.h"
|
||||
|
||||
#undef NDEBUG
|
||||
#include <assert.h>
|
||||
#include <math.h>
|
||||
#include <stdio.h>
|
||||
#include <string>
|
||||
#include <vector>
|
||||
|
||||
|
||||
const float MAX_QUANTIZATION_REFERENCE_ERROR = 0.0001;
|
||||
const float MAX_QUANTIZATION_TOTAL_ERROR = 0.002;
|
||||
const float MAX_DOT_PRODUCT_ERROR = 0.02;
|
||||
|
||||
const char* RESULT_STR[] = {"ok", "FAILED"};
|
||||
|
||||
|
||||
// Generate synthetic data
|
||||
void generate_data(float offset, size_t n, float * dst) {
|
||||
for (size_t i = 0; i < n; i++) {
|
||||
dst[i] = 0.1 + 2*cosf(i + offset);
|
||||
}
|
||||
}
|
||||
|
||||
// Calculate RMSE between two float arrays
|
||||
float array_rmse(const float * a1, const float * a2, size_t n) {
|
||||
double sum = 0;
|
||||
for (size_t i = 0; i < n; i++) {
|
||||
double diff = a1[i] - a2[i];
|
||||
sum += diff * diff;
|
||||
}
|
||||
return sqrtf(sum) / n;
|
||||
}
|
||||
|
||||
// Total quantization error on test data
|
||||
float total_quantization_error(quantize_fns_t & qfns, size_t test_size, const float * test_data) {
|
||||
std::vector<uint8_t> tmp_q(test_size);
|
||||
std::vector<float> tmp_out(test_size);
|
||||
|
||||
qfns.quantize_row_q(test_data, tmp_q.data(), test_size);
|
||||
qfns.dequantize_row_q(tmp_q.data(), tmp_out.data(), test_size);
|
||||
return array_rmse(test_data, tmp_out.data(), test_size);
|
||||
}
|
||||
|
||||
// Total quantization error on test data
|
||||
float reference_quantization_error(quantize_fns_t & qfns, size_t test_size, const float * test_data) {
|
||||
std::vector<uint8_t> tmp_q(test_size);
|
||||
std::vector<float> tmp_out(test_size);
|
||||
std::vector<float> tmp_out_ref(test_size);
|
||||
|
||||
qfns.quantize_row_q(test_data, tmp_q.data(), test_size);
|
||||
qfns.dequantize_row_q(tmp_q.data(), tmp_out.data(), test_size);
|
||||
|
||||
qfns.quantize_row_q_reference(test_data, tmp_q.data(), test_size);
|
||||
qfns.dequantize_row_q(tmp_q.data(), tmp_out_ref.data(), test_size);
|
||||
|
||||
return array_rmse(tmp_out.data(), tmp_out_ref.data(), test_size);
|
||||
}
|
||||
|
||||
float dot_product(const float * a1, const float * a2, size_t test_size) {
|
||||
double sum = 0;
|
||||
for (size_t i = 0; i < test_size; i++) {
|
||||
sum += a1[i] * a2[i];
|
||||
}
|
||||
return sum;
|
||||
}
|
||||
|
||||
// Total dot product error
|
||||
float dot_product_error(quantize_fns_t & qfns, size_t test_size, const float * test_data1, const float *test_data2) {
|
||||
std::vector<uint8_t> tmp_q1(test_size);
|
||||
std::vector<uint8_t> tmp_q2(test_size*2);
|
||||
|
||||
qfns.quantize_row_q(test_data1, tmp_q1.data(), test_size);
|
||||
qfns.quantize_row_q_dot(test_data2, tmp_q2.data(), test_size);
|
||||
|
||||
float result = INFINITY;
|
||||
qfns.vec_dot_q(test_size, &result, tmp_q1.data(), tmp_q2.data());
|
||||
|
||||
const float dot_ref = dot_product(test_data1, test_data2, test_size);
|
||||
|
||||
return fabsf(result - dot_ref) / test_size;
|
||||
}
|
||||
|
||||
int main(int argc, char * argv[]) {
|
||||
bool verbose = false;
|
||||
const size_t test_size = 32 * 128;
|
||||
|
||||
std::string arg;
|
||||
for (int i = 1; i < argc; i++) {
|
||||
arg = argv[i];
|
||||
|
||||
if (arg == "-v") {
|
||||
verbose = true;
|
||||
} else {
|
||||
fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
|
||||
return 1;
|
||||
}
|
||||
}
|
||||
|
||||
std::vector<float> test_data(test_size);
|
||||
std::vector<float> test_data2(test_size);
|
||||
|
||||
generate_data(0.0, test_data.size(), test_data.data());
|
||||
generate_data(1.0, test_data2.size(), test_data2.data());
|
||||
|
||||
// Initialize GGML, ensures float conversion tables are initialized
|
||||
struct ggml_init_params ggml_params = {
|
||||
/* .mem_size = */ 1*1024,
|
||||
/* .mem_buffer = */ NULL,
|
||||
/* .no_alloc = */ true,
|
||||
};
|
||||
struct ggml_context * ctx = ggml_init(ggml_params);
|
||||
|
||||
int num_failed = 0;
|
||||
bool failed = false;
|
||||
|
||||
for (int i = 0; i < GGML_TYPE_COUNT; i++) {
|
||||
ggml_type type = (ggml_type) i;
|
||||
quantize_fns_t qfns = ggml_internal_get_quantize_fn(i);
|
||||
|
||||
if (qfns.quantize_row_q && qfns.dequantize_row_q) {
|
||||
const float total_error = total_quantization_error(qfns, test_size, test_data.data());
|
||||
failed = !(total_error < MAX_QUANTIZATION_TOTAL_ERROR);
|
||||
num_failed += failed;
|
||||
if (failed || verbose) {
|
||||
printf("%5s absolute quantization error: %s (%f)\n", ggml_type_name(type), RESULT_STR[failed], total_error);
|
||||
}
|
||||
|
||||
const float reference_error = reference_quantization_error(qfns, test_size, test_data.data());
|
||||
failed = !(reference_error < MAX_QUANTIZATION_REFERENCE_ERROR);
|
||||
num_failed += failed;
|
||||
if (failed || verbose) {
|
||||
printf("%5s reference implementation error: %s (%f)\n", ggml_type_name(type), RESULT_STR[failed], reference_error);
|
||||
}
|
||||
|
||||
const float vec_dot_error = dot_product_error(qfns, test_size, test_data.data(), test_data2.data());
|
||||
failed = !(vec_dot_error < MAX_DOT_PRODUCT_ERROR);
|
||||
num_failed += failed;
|
||||
if (failed || verbose) {
|
||||
printf("%5s dot product error: %s (%f)\n", ggml_type_name(type), RESULT_STR[failed], vec_dot_error);
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
if (num_failed || verbose) {
|
||||
printf("%d tests failed\n", num_failed);
|
||||
}
|
||||
|
||||
ggml_free(ctx);
|
||||
|
||||
return num_failed > 0;
|
||||
}
|
||||
310
tests/test-quantize-perf.cpp
Normal file
310
tests/test-quantize-perf.cpp
Normal file
|
|
@ -0,0 +1,310 @@
|
|||
// Benchmark quantization specific functions on synthetic data
|
||||
|
||||
#include "ggml.h"
|
||||
|
||||
#undef NDEBUG
|
||||
#include <algorithm>
|
||||
#include <assert.h>
|
||||
#include <functional>
|
||||
#include <inttypes.h>
|
||||
#include <math.h>
|
||||
#include <memory>
|
||||
#include <stdio.h>
|
||||
#include <string>
|
||||
#include <vector>
|
||||
|
||||
#define MAX_ALIGNMENT 64
|
||||
#define QK 32
|
||||
#define WARMUP 5
|
||||
#define ITERATIONS 10
|
||||
|
||||
#define L1_SIZE 32*128
|
||||
#define L2_SIZE 32*2048
|
||||
#define L3_SIZE 32*20480
|
||||
#define MEM_SIZE 32*2048000
|
||||
|
||||
struct quantize_perf_params {
|
||||
std::vector<std::string> include_types;
|
||||
std::vector<size_t> test_sizes;
|
||||
size_t alignment_offset = 0;
|
||||
bool op_quantize_row_q_reference = false;
|
||||
bool op_quantize_row_q = false;
|
||||
bool op_dequantize_row_q = false;
|
||||
bool op_quantize_row_q_dot = false;
|
||||
bool op_vec_dot_q = false;
|
||||
};
|
||||
|
||||
|
||||
#if defined(__x86_64__) || defined(__i386__)
|
||||
|
||||
#include <x86intrin.h>
|
||||
inline int64_t cpu_cycles() {
|
||||
// Rough way to detect new-ish CPUs
|
||||
#ifdef __POPCNT__
|
||||
unsigned int dummy;
|
||||
return __rdtscp(&dummy);
|
||||
#else
|
||||
return __rdtsc();
|
||||
#endif
|
||||
}
|
||||
|
||||
#else
|
||||
|
||||
#define cpu_cycles() 0
|
||||
|
||||
#endif
|
||||
|
||||
|
||||
// Generate synthetic data
|
||||
void generate_data(float offset, size_t n, float * dst) {
|
||||
for (size_t i = 0; i < n; i++) {
|
||||
dst[i] = 0.1 + 2*cosf(i + offset);
|
||||
}
|
||||
}
|
||||
|
||||
float gigabytes_per_second(size_t bytes, int64_t usecs) {
|
||||
return bytes / (float) usecs * 1000000 / (1024*1024*1024);
|
||||
}
|
||||
|
||||
void * align_with_offset(void * ptr, int offset) {
|
||||
size_t dummy_size = MAX_ALIGNMENT * 4;
|
||||
return (char *) std::align(MAX_ALIGNMENT, MAX_ALIGNMENT, ptr, dummy_size) + offset;
|
||||
}
|
||||
|
||||
void benchmark_function(size_t size, size_t q_size, std::function<size_t(void)> function) {
|
||||
int64_t min_time_us = INT64_MAX;
|
||||
int64_t total_time_us = 0;
|
||||
int64_t min_time_cycles = INT64_MAX;
|
||||
int64_t total_time_cycles = 0;
|
||||
|
||||
for (int i = 0; i < WARMUP; i++) {
|
||||
function();
|
||||
}
|
||||
|
||||
|
||||
for (int i = 0; i < ITERATIONS; i++) {
|
||||
const int64_t start_time = ggml_time_us();
|
||||
const int64_t start_cycles = cpu_cycles();
|
||||
|
||||
function();
|
||||
|
||||
const int64_t end_cycles = cpu_cycles();
|
||||
const int64_t end_time = ggml_time_us();
|
||||
|
||||
total_time_cycles += end_cycles - start_cycles;
|
||||
min_time_cycles = std::min(min_time_cycles, end_cycles - start_cycles);
|
||||
total_time_us += end_time - start_time;
|
||||
min_time_us = std::min(min_time_us, end_time - start_time);
|
||||
}
|
||||
|
||||
printf(" min cycles/%d vals : %9.2f\n", QK, QK * min_time_cycles / (float) size);
|
||||
printf(" avg cycles/%d vals : %9.2f\n", QK, QK * total_time_cycles / (float) (size * ITERATIONS));
|
||||
printf(" float32 throughput : %9.2f GB/s\n", gigabytes_per_second(4 * size * ITERATIONS, total_time_us));
|
||||
printf(" quantized throughput : %9.2f GB/s\n", gigabytes_per_second(q_size * ITERATIONS, total_time_us));
|
||||
}
|
||||
|
||||
int main(int argc, char * argv[]) {
|
||||
quantize_perf_params params {};
|
||||
|
||||
// read command line
|
||||
|
||||
bool invalid_param = false;
|
||||
std::string arg;
|
||||
for (int i = 1; i < argc; i++) {
|
||||
arg = argv[i];
|
||||
|
||||
if (arg == "--size") {
|
||||
if (++i >= argc) {
|
||||
invalid_param = true;
|
||||
break;
|
||||
}
|
||||
size_t size = std::stoi(argv[i]);
|
||||
if (size % 32 != 0) {
|
||||
fprintf(stderr, "error: size %zu not divisible by 32\n", size);
|
||||
invalid_param = true;
|
||||
break;
|
||||
}
|
||||
params.test_sizes.push_back(size);
|
||||
} else if (arg == "-3") {
|
||||
// quick select sizes that probably fit in CPU caches
|
||||
params.test_sizes.push_back(L1_SIZE);
|
||||
params.test_sizes.push_back(L2_SIZE);
|
||||
params.test_sizes.push_back(L3_SIZE);
|
||||
} else if (arg == "-4") {
|
||||
// quick select cache sizes + memory
|
||||
params.test_sizes.push_back(L1_SIZE);
|
||||
params.test_sizes.push_back(L2_SIZE);
|
||||
params.test_sizes.push_back(L3_SIZE);
|
||||
params.test_sizes.push_back(MEM_SIZE);
|
||||
} else if (arg == "--op") {
|
||||
if (++i >= argc) {
|
||||
invalid_param = true;
|
||||
break;
|
||||
}
|
||||
std::string op {argv[i]};
|
||||
if (op == "quantize_row_q_reference") {
|
||||
params.op_quantize_row_q_reference = true;
|
||||
} else if (op == "quantize_row_q") {
|
||||
params.op_quantize_row_q = true;
|
||||
} else if (op == "dequantize_row_q") {
|
||||
params.op_dequantize_row_q = true;
|
||||
} else if (op == "quantize_row_q_dot") {
|
||||
params.op_quantize_row_q_dot = true;
|
||||
} else if (op == "vec_dot_q") {
|
||||
params.op_vec_dot_q = true;
|
||||
} else {
|
||||
invalid_param = true;
|
||||
break;
|
||||
}
|
||||
} else if (arg == "--type") {
|
||||
if (++i >= argc) {
|
||||
invalid_param = true;
|
||||
break;
|
||||
}
|
||||
params.include_types.push_back(argv[i]);
|
||||
} else if (arg == "--alignment-offset") {
|
||||
if (++i >= argc) {
|
||||
invalid_param = true;
|
||||
break;
|
||||
}
|
||||
int alignment = std::stoi(argv[i]);
|
||||
if (alignment < 0 || alignment > MAX_ALIGNMENT) {
|
||||
fprintf(stderr, "error: aligment-offset must be less than %d\n", MAX_ALIGNMENT);
|
||||
invalid_param = true;
|
||||
break;
|
||||
}
|
||||
params.alignment_offset = alignment;
|
||||
} else {
|
||||
fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
|
||||
return 1;
|
||||
}
|
||||
}
|
||||
if (invalid_param) {
|
||||
fprintf(stderr, "error: invalid parameter for argument: %s\n", arg.c_str());
|
||||
return 1;
|
||||
}
|
||||
|
||||
if (params.test_sizes.empty()) {
|
||||
params.test_sizes.push_back(L1_SIZE);
|
||||
}
|
||||
if (!(params.op_quantize_row_q_reference || params.op_quantize_row_q || params.op_dequantize_row_q || params.op_quantize_row_q_dot || params.op_vec_dot_q)) {
|
||||
params.op_quantize_row_q_reference = params.op_quantize_row_q = params.op_dequantize_row_q = params.op_quantize_row_q_dot = params.op_vec_dot_q = true;
|
||||
}
|
||||
|
||||
std::sort(params.test_sizes.begin(), params.test_sizes.end());
|
||||
size_t largest = params.test_sizes.back();
|
||||
|
||||
std::vector<uint8_t> test_data1_v(largest*4 + MAX_ALIGNMENT*2);
|
||||
std::vector<uint8_t> test_data2_v(largest*4 + MAX_ALIGNMENT*2);
|
||||
std::vector<uint8_t> test_q1_v(largest*4 + MAX_ALIGNMENT*2);
|
||||
std::vector<uint8_t> test_q2_v(largest*4 + MAX_ALIGNMENT*2);
|
||||
std::vector<uint8_t> test_out_v(largest*4 + MAX_ALIGNMENT*2);
|
||||
|
||||
float * test_data1 = (float *) align_with_offset(test_data1_v.data(), params.alignment_offset);
|
||||
float * test_data2 = (float *) align_with_offset(test_data2_v.data(), params.alignment_offset);
|
||||
float * test_q1 = (float *) align_with_offset(test_q1_v.data(), params.alignment_offset);
|
||||
float * test_q2 = (float *) align_with_offset(test_q2_v.data(), params.alignment_offset);
|
||||
float * test_out = (float *) align_with_offset(test_out_v.data(), params.alignment_offset);
|
||||
|
||||
generate_data(0, largest, test_data1);
|
||||
generate_data(1, largest, test_data2);
|
||||
|
||||
|
||||
// Initialize GGML, ensures float conversion tables are initialized
|
||||
struct ggml_init_params ggml_params = {
|
||||
/* .mem_size = */ 1*1024,
|
||||
/* .mem_buffer = */ NULL,
|
||||
/* .no_alloc = */ true,
|
||||
};
|
||||
struct ggml_context * ctx = ggml_init(ggml_params);
|
||||
|
||||
for (int i = 0; i < GGML_TYPE_COUNT; i++) {
|
||||
ggml_type type = (ggml_type) i;
|
||||
quantize_fns_t qfns = ggml_internal_get_quantize_fn(i);
|
||||
if (!params.include_types.empty() && std::find(params.include_types.begin(), params.include_types.end(), ggml_type_name(type)) == params.include_types.end()) {
|
||||
continue;
|
||||
}
|
||||
|
||||
if (qfns.quantize_row_q && qfns.dequantize_row_q) {
|
||||
printf("%s\n", ggml_type_name(type));
|
||||
|
||||
if (params.op_quantize_row_q_reference) {
|
||||
printf(" quantize_row_q_reference\n");
|
||||
for (size_t size : params.test_sizes) {
|
||||
printf(" %zu values (%.2f MB)\n", size, 4*size/(float)(1024*1024));
|
||||
auto quantize_fn = [&](void ) {
|
||||
qfns.quantize_row_q_reference(test_data1, test_q1, size);
|
||||
return test_q1[0];
|
||||
};
|
||||
size_t quantized_size = size / ggml_blck_size(type) * ggml_type_size(type);
|
||||
benchmark_function(size, quantized_size, quantize_fn);
|
||||
}
|
||||
printf("\n");
|
||||
}
|
||||
|
||||
if (params.op_quantize_row_q) {
|
||||
printf(" quantize_row_q\n");
|
||||
for (size_t size : params.test_sizes) {
|
||||
printf(" %zu values (%.2f MB)\n", size, 4*size/(float)(1024*1024));
|
||||
auto quantize_fn = [&](void ) {
|
||||
qfns.quantize_row_q(test_data1, test_q1, size);
|
||||
return test_q1[0];
|
||||
};
|
||||
size_t quantized_size = size / ggml_blck_size(type) * ggml_type_size(type);
|
||||
benchmark_function(size, quantized_size, quantize_fn);
|
||||
}
|
||||
printf("\n");
|
||||
}
|
||||
|
||||
if (params.op_dequantize_row_q) {
|
||||
printf(" dequantize_row_q\n");
|
||||
qfns.quantize_row_q(test_data1, test_q1, largest);
|
||||
for (size_t size : params.test_sizes) {
|
||||
printf(" %zu values (%.2f MB)\n", size, 4*size/(float)(1024*1024));
|
||||
auto quantize_fn = [&](void ) {
|
||||
qfns.dequantize_row_q(test_q1, test_out, size);
|
||||
return test_out[0];
|
||||
};
|
||||
size_t quantized_size = size / ggml_blck_size(type) * ggml_type_size(type);
|
||||
benchmark_function(size, quantized_size, quantize_fn);
|
||||
}
|
||||
printf("\n");
|
||||
}
|
||||
|
||||
if (params.op_quantize_row_q_dot) {
|
||||
printf(" quantize_row_q_dot\n");
|
||||
for (size_t size : params.test_sizes) {
|
||||
printf(" %zu values (%.2f MB)\n", size, 4*size/(float)(1024*1024));
|
||||
auto quantize_fn = [&](void ) {
|
||||
qfns.quantize_row_q_dot(test_data1, test_q1, size);
|
||||
return test_q1[0];
|
||||
};
|
||||
size_t quantized_size = size / ggml_blck_size(type) * ggml_type_size(type);
|
||||
benchmark_function(size, quantized_size, quantize_fn);
|
||||
}
|
||||
printf("\n");
|
||||
}
|
||||
|
||||
if (params.op_vec_dot_q) {
|
||||
printf(" vec_dot_q\n");
|
||||
qfns.quantize_row_q(test_data1, test_q1, largest);
|
||||
qfns.quantize_row_q(test_data2, test_q2, largest);
|
||||
for (size_t size : params.test_sizes) {
|
||||
printf(" %zu values (%.2f MB)\n", size, 4*size/(float)(1024*1024));
|
||||
auto quantize_fn = [&](void ) {
|
||||
float result;
|
||||
qfns.vec_dot_q(size, &result, test_q1, test_q2);
|
||||
return result;
|
||||
};
|
||||
size_t quantized_size = size / ggml_blck_size(type) * ggml_type_size(type);
|
||||
benchmark_function(size, quantized_size, quantize_fn);
|
||||
}
|
||||
printf("\n");
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
ggml_free(ctx);
|
||||
|
||||
return 0;
|
||||
}
|
||||
|
|
@ -1,42 +0,0 @@
|
|||
#include "ggml.h"
|
||||
#undef NDEBUG
|
||||
#include <assert.h>
|
||||
#include <math.h>
|
||||
|
||||
int main(void) {
|
||||
#define QK 32
|
||||
float src[QK];
|
||||
uint8_t dst[24];
|
||||
int64_t hist[16];
|
||||
|
||||
for (int i = 0; i < QK; i++) {
|
||||
src[i] = (float)(i + 1);
|
||||
}
|
||||
|
||||
size_t size = ggml_quantize_q4_0(src, dst, QK, QK, hist);
|
||||
assert(size == 20);
|
||||
float max_result = ((float *)dst)[0];
|
||||
float max_expected = src[31] / ((1 << 3) - 1);
|
||||
assert(max_result == max_expected);
|
||||
for (int i = 0; i < QK; i++) {
|
||||
uint8_t q4_result = (i % 2) ? (dst[sizeof(float) + i/2] >> 4) : (dst[sizeof(float) + i/2] & 0xF);
|
||||
uint8_t q4_expected = roundf(src[i] / max_expected) + 8;
|
||||
assert(q4_result == q4_expected);
|
||||
}
|
||||
|
||||
size = ggml_quantize_q4_1(src, dst, QK, QK, hist);
|
||||
assert(size == 24);
|
||||
float delta_result = ((float *)dst)[0];
|
||||
float delta_expected = (src[31] - src[0]) / ((1 << 4) - 1);
|
||||
assert(delta_result == delta_expected);
|
||||
float min_result = ((float *)dst)[1];
|
||||
float min_expected = src[0];
|
||||
assert(min_result == min_expected);
|
||||
for (int i = 0; i < QK; i++) {
|
||||
uint8_t q4_result = (i % 2) ? (dst[sizeof(float)*2 + i/2] >> 4) : (dst[sizeof(float)*2 + i/2] & 0xF);
|
||||
uint8_t q4_expected = roundf((src[i] - min_expected) / delta_expected);
|
||||
assert(q4_result == q4_expected);
|
||||
}
|
||||
|
||||
return 0;
|
||||
}
|
||||
Loading…
Add table
Add a link
Reference in a new issue