vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Merge branch 'ggerganov:master' into master

Browse source
This commit is contained in:
hsnmkls 2024-02-29 16:21:59 +08:00 • committed by GitHub
commit e063bce2b6
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
60 changed files with 7591 additions and 1989 deletions
Download patch file Download diff file Expand all files Collapse all files

2
.github/ISSUE_TEMPLATE/bug.md vendored
View file

@ -7,3 +7,5 @@ assignees: ''
---
Please include information about your system, the steps to reproduce the bug, and the version of llama.cpp that you are using. If possible, please provide a minimal code example that reproduces the bug.
If the bug concerns the server, please try to reproduce it first using the [server test scenario framework](https://github.com/ggerganov/llama.cpp/tree/master/examples/server/tests).

3
.github/workflows/build.yml vendored
View file

@ -669,8 +669,7 @@ jobs:
run: |
cd examples/llama.android
# Skip armeabi-v7a for now (https://github.com/llvm/llvm-project/issues/65820).
./gradlew build --no-daemon -Pskip-armeabi-v7a
./gradlew build --no-daemon
# freeBSD-latest:
# runs-on: macos-12

83
.github/workflows/server.yml vendored Normal file
View file

@ -0,0 +1,83 @@
# Server build and tests
name: Server
on:
workflow_dispatch: # allows manual triggering
push:
branches:
- master
paths: ['.github/workflows/server.yml', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.swift', '**/*.m', 'examples/server/tests/**.*']
pull_request:
types: [opened, synchronize, reopened]
paths: ['.github/workflows/server.yml', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.swift', '**/*.m', 'examples/server/tests/**.*']
jobs:
server:
runs-on: ubuntu-latest
strategy:
matrix:
sanitizer: [ADDRESS, THREAD, UNDEFINED]
build_type: [Debug, Release]
include:
- build_type: Release
sanitizer: ""
exclude:
- build_type: Release
sanitizer: ADDRESS
- build_type: Release
sanitizer: THREAD
- build_type: Release
sanitizer: UNDEFINED
container:
image: ubuntu:latest
ports:
- 8888
options: --cpus 4
steps:
- name: Clone
id: checkout
uses: actions/checkout@v3
- name: Dependencies
id: depends
run: |
apt-get update
apt-get -y install \
build-essential \
git \
cmake \
python3-pip \
wget \
psmisc
- name: Build
id: cmake_build
run: |
mkdir build
cd build
cmake .. \
-DLLAMA_NATIVE=OFF \
-DLLAMA_BUILD_SERVER=ON \
-DCMAKE_BUILD_TYPE=${{ matrix.build_type }} \
-DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON ;
cmake --build . --config ${{ matrix.build_type }} -j $(nproc) --target server
- name: Tests dependencies
id: test_dependencies
run: |
pip install -r examples/server/tests/requirements.txt
- name: Download models
id: download_models
run: |
cd examples/server/tests
../../../scripts/hf.sh --repo ggml-org/models --file tinyllamas/stories260K.gguf
- name: Tests
id: server_integration_test
run: |
cd examples/server/tests
PORT=8888 ./tests.sh

10
CMakeLists.txt
View file

@ -936,10 +936,16 @@ if (CMAKE_OSX_ARCHITECTURES STREQUAL "arm64" OR CMAKE_GENERATOR_PLATFORM_LWR STR
list(APPEND ARCH_FLAGS -mfpu=neon-fp-armv8 -mno-unaligned-access)
endif()
if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "armv7")
# Raspberry Pi 2
list(APPEND ARCH_FLAGS -mfpu=neon-fp-armv8 -mno-unaligned-access -funsafe-math-optimizations)
if ("${CMAKE_SYSTEM_NAME}" STREQUAL "Android")
# Android armeabi-v7a
list(APPEND ARCH_FLAGS -mfpu=neon-vfpv4 -mno-unaligned-access -funsafe-math-optimizations)
else()
# Raspberry Pi 2
list(APPEND ARCH_FLAGS -mfpu=neon-fp-armv8 -mno-unaligned-access -funsafe-math-optimizations)
endif()
endif()
if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "armv8")
# Android arm64-v8a
# Raspberry Pi 3, 4, Zero 2 (32-bit)
list(APPEND ARCH_FLAGS -mno-unaligned-access)
endif()

11
Makefile
View file

@ -381,8 +381,13 @@ ifdef LLAMA_BLIS
endif # LLAMA_BLIS
ifdef LLAMA_CUBLAS
MK_CPPFLAGS += -DGGML_USE_CUBLAS -I/usr/local/cuda/include -I/opt/cuda/include -I$(CUDA_PATH)/targets/x86_64-linux/include -I/usr/local/cuda/targets/aarch64-linux/include
MK_LDFLAGS += -lcuda -lcublas -lculibos -lcudart -lcublasLt -lpthread -ldl -lrt -L/usr/local/cuda/lib64 -L/opt/cuda/lib64 -L$(CUDA_PATH)/targets/x86_64-linux/lib -L/usr/local/cuda/targets/aarch64-linux/lib -L/usr/lib/wsl/lib
ifneq ('', '$(wildcard /opt/cuda)')
CUDA_PATH ?= /opt/cuda
else
CUDA_PATH ?= /usr/local/cuda
endif
MK_CPPFLAGS += -DGGML_USE_CUBLAS -I$(CUDA_PATH)/include -I$(CUDA_PATH)/targets/$(UNAME_M)-linux/include
MK_LDFLAGS += -lcuda -lcublas -lculibos -lcudart -lcublasLt -lpthread -ldl -lrt -L$(CUDA_PATH)/lib64 -L/usr/lib64 -L$(CUDA_PATH)/targets/$(UNAME_M)-linux/lib -L/usr/lib/wsl/lib
OBJS += ggml-cuda.o
MK_NVCCFLAGS += -use_fast_math
ifdef LLAMA_FATAL_WARNINGS
@ -597,7 +602,7 @@ $(info I CC: $(shell $(CC) --version | head -n 1))
$(info I CXX: $(shell $(CXX) --version | head -n 1))
ifdef LLAMA_CUBLAS
$(info I NVCC: $(shell $(NVCC) --version | tail -n 1))
CUDA_VERSION := $(shell nvcc --version | grep -oP 'release (\K[0-9]+\.[0-9])')
CUDA_VERSION := $(shell $(NVCC) --version | grep -oP 'release (\K[0-9]+\.[0-9])')
ifeq ($(shell awk -v "v=$(CUDA_VERSION)" 'BEGIN { print (v < 11.7) }'),1)
ifndef CUDA_DOCKER_ARCH
ifndef CUDA_POWER_ARCH

7
README.md
View file

@ -107,13 +107,16 @@ Typically finetunes of the base models below are supported as well.
**Multimodal models:**
- [x] [LLaVA 1.5 models](https://huggingface.co/collections/liuhaotian/llava-15-653aac15d994e992e2677a7e)
- [x] [LLaVA 1.5 models](https://huggingface.co/collections/liuhaotian/llava-15-653aac15d994e992e2677a7e), [LLaVA 1.6 models](https://huggingface.co/collections/liuhaotian/llava-16-65b9e40155f60fd046a5ccf2)
- [x] [BakLLaVA](https://huggingface.co/models?search=SkunkworksAI/Bakllava)
- [x] [Obsidian](https://huggingface.co/NousResearch/Obsidian-3B-V0.5)
- [x] [ShareGPT4V](https://huggingface.co/models?search=Lin-Chen/ShareGPT4V)
- [x] [MobileVLM 1.7B/3B models](https://huggingface.co/models?search=mobileVLM)
- [x] [Yi-VL](https://huggingface.co/models?search=Yi-VL)
**HTTP server**
[llama.cpp web server](./examples/server) is a lightweight [OpenAI API](https://github.com/openai/openai-openapi) compatible HTTP server that can be used to serve local models and easily connect them to existing clients.
**Bindings:**
@ -155,6 +158,8 @@ Unless otherwise noted these projects are open-source with permissive licensing:
- [semperai/amica](https://github.com/semperai/amica)
- [withcatai/catai](https://github.com/withcatai/catai)
- [Mobile-Artificial-Intelligence/maid](https://github.com/Mobile-Artificial-Intelligence/maid) (MIT)
- [Msty](https://msty.app) (proprietary)
- [LLMFarm](https://github.com/guinmoon/LLMFarm?tab=readme-ov-file) (MIT)
---

116
awq-py/README.md
View file

@ -1,116 +0,0 @@
# AWQ: Activation-aware Weight Quantization for LLM - version apply to llamacpp
[[Paper](https://arxiv.org/abs/2306.00978)][[Original Repo](https://github.com/mit-han-lab/llm-awq)][[Easy-to-use Repo](https://github.com/casper-hansen/AutoAWQ)]
**Supported models:**
- [X] LLaMA
- [x] LLaMA 2
- [X] MPT
- [X] Mistral AI v0.1
- [ ] Bloom
- [ ] Mixtral MoE
**TODO:**
- [x] Update version work with both MPT and MPT-AWQ model
- [ ] Add OPT model
- [ ] Add Bloom model
- [ ] Add Mixtral MoE
- [ ] Support w3, w2
## Contents
- [Install](##Install)
- [Convert](##Convert)
- [Quantize](##Quantize)
- [Test](##Test)
- [Benchmark](##Benchmark)
- [Results](##Results)
## Install
Install requirements
```bash
pip install -r requirements.txt
```
Get the pre-computed AWQ search results for multiple model families, including LLaMA, LLaMA2, MPT, OPT
```bash
git clone https://huggingface.co/datasets/mit-han-lab/awq-model-zoo awq_cache
```
## Convert
Example for llama model
```bash
# For llama7b and llama2 models
python convert.py models/llama-7b/ --awq-path awq_cache/llama-7b-w4-g128.pt --outfile models/llama_7b_fp16.gguf
# For mistral and mpt models
python convert-hf-to-gguf.py models/mpt-7b/ --awq-path awq_cache/mpt-7b-w4-g128.pt --outfile models/mpt_7b_fp16.gguf
```
## Quantize
```bash
# We only benchmark and confirm the results on q4_0, q4_1, and q2_k types.
./quantize models/llama_7b_fp16.gguf models/llama_7b_q4_0.gguf q4_0
```
## Test
```bash
# For all models.
./build/bin/main -m models/llama_7b_q4_0.gguf -n 128 --prompt "Once upon a time"
```
## Benchmark
The perplexity measurements in table above are done against the `wikitext2` test dataset (https://paperswithcode.com/dataset/wikitext-2), with context length of 512.
```bash
# For llama and llama2, and mistral models.
./perplexity -m models/llama_7b_q4_0.gguf -f datasets/wikitext-2-raw/wiki.test.raw
```
## Results
Results are run on OpenBLAS (CPU) and CuBLAS (GPU) for fair comparison
We use three types of llamacpp quantization methods to work with our version, including q4_0, q4_1, and q2_k
### Llama 7B (Build with OpenBLAS)
| Model | Measure | F16 | Q4_0 | Q4_1 | Q2_K |
|-----------:|--------------|-------:|-------:|-------:|-------:|
|Llama 7B | perplexity | 5.9066 | 6.1214 | 6.0643 | 6.5808 |
|Llama 7B | file size | 12.9G | 3.5G | 3.9G | 2.7G |
|Llama 7B | bits/weight | 16.0 | 4.5 | 5.0 | 2.6 |
|AWQ-LLama 7B| perplexity | 5.9175 | 6.0252 | 5.9987 | 6.3692 |
|AWQ-LLama 7B| file size | 12.9G | 3.5G | 3.9G | 2.7G |
|AWQ-LLama 7B| bits/weight | 16.0 | 4.5 | 5.0 | 2.6 |
### Llama2 7B (Build with CuBLAS)
| Model | Measure | F16 | Q4_0 | Q4_1 | Q2_K |
|------------:|--------------|-------:|-------:|-------:|-------:|
|Llama2 7B | perplexity | 5.8664 | 6.0260 | 6.0656 | 6.4496 |
|Llama2 7B | file size | 12.9G | 3.5G | 3.9G | 2.7G |
|Llama2 7B | bits/weight | 16.0 | 4.5 | 5.0 | 2.6 |
|AWQ-LLama2 7B| perplexity | 5.8801 | 6.0054 | 5.9849 | 6.3650 |
|AWQ-LLama2 7B| file size | 12.9G | 3.5G | 3.9G | 2.7G |
|AWQ-LLama2 7B| bits/weight | 16.0 | 4.5 | 5.0 | 2.6 |
### Mistral 7B v0.1 (Build with CuBLAS)
| Model | Measure | F16 | Q4_0 | Q4_1 | Q2_K |
|-------------:|--------------|-------:|-------:|-------:|-------:|
|Mistral 7B | perplexity | 5.6931 | 5.8202 | 5.8268 | 6.1645 |
|Mistral 7B | file size | 14.5G | 4.1G | 4.5G | 3.1G |
|Mistral 7B | bits/weight | 16.0 | 4.5 | 5.0 | 2.6 |
|AWQ-Mistral 7B| perplexity | 5.6934 | 5.8020 | 5.7691 | 6.0426 |
|AWQ-Mistral 7B| file size | 14.5G | 4.1G | 4.5G | 3.1G |
|AWQ-Mistral 7B| bits/weight | 16.0 | 4.5 | 5.0 | 2.6 |
### MPT 7B (Build with OpenBLAS)
| Model | Measure | F16 | Q4_0 | Q4_1 | Q2_K |
|---------:|--------------|-------:|-------:|-------:|--------:|
|MPT 7B | perplexity | 8.4369 | 8.7956 | 8.6265 | 11.4913 |
|MPT 7B | file size | 13.7G | 3.9G | 4.3G | 2.8G |
|MPT 7B | bits/weight | 16.0 | 4.5 | 5.0 | 2.6 |
|AWQ-MPT 7B| perplexity | 8.4944 | 8.7053 | 8.6750 | 10.2873|
|AWQ-MPT 7B| file size | 13.7G | 3.9G | 4.3G | 2.8G |
|AWQ-MPT 7B| bits/weight | 16.0 | 4.5 | 5.0 | 2.6 |

254
awq-py/awq/apply_awq.py
View file

@ -1,254 +0,0 @@
"""
Implements the AWQ for llama.cpp use cases.
Original paper: https://arxiv.org/abs/2306.00978
This code is based on versions of the AWQ implementation found in the following repositories:
* https://github.com/mit-han-lab/llm-awq
* https://github.com/casper-hansen/AutoAWQ
"""
import os
import torch
import torch.nn as nn
from transformers import AutoModelForCausalLM, AutoConfig
from transformers.models.bloom.modeling_bloom import BloomGelu
from transformers.models.llama.modeling_llama import LlamaRMSNorm
from transformers.activations import GELUActivation
class ScaledActivation(nn.Module):
"""
ScaledActivation module wraps an existing activation function and applies a
scale factor to its output.
Args:
module (nn.Module): The activation function to be scaled.
scales (torch.Tensor): A tensor of size (num_features,) containing the initial
scale factors for each feature.
Returns:
torch.Tensor: The scaled output of the activation function.
"""
def __init__(self, module, scales):
super().__init__()
self.act = module
self.scales = nn.Parameter(scales.data)
def forward(self, x):
return self.act(x) / self.scales.view(1, 1, -1).to(x.device)
def set_op_by_name(layer, name, new_module):
"""
Set the new module for given module's name.
Args:
layer (nn.Module): The layer in which to replace the submodule.
name (str): The path to the submodule to be replaced, using dot notation
to access nested modules.
new_module (nn.Module): The new module to replace the existing one.
"""
levels = name.split(".")
if len(levels) > 1:
mod_ = layer
for l_idx in range(len(levels) - 1):
if levels[l_idx].isdigit():
mod_ = mod_[int(levels[l_idx])]
else:
mod_ = getattr(mod_, levels[l_idx])
setattr(mod_, levels[-1], new_module)
else:
setattr(layer, name, new_module)
def get_op_by_name(module, op_name):
"""
Retrieves a submodule within a given layer based on its name.
Args:
module (nn.Module): The layer containing the submodule to find.
op_name (str): The name of the submodule.
Returns:
nn.Module: The requested submodule found within the given layer.
Raises:
ValueError: If the specified submodule cannot be found within the layer.
"""
for name, m in module.named_modules():
if name == op_name:
return m
raise ValueError(f"Cannot find op {op_name} in module {module}")
@torch.no_grad()
def scale_ln_fcs(ln, fcs, scales):
"""
Scales the weights of a LayerNorm and a list of fully-connected layers proportionally.
Args:
ln (nn.LayerNorm): The LayerNorm module to be scaled.
fcs (List[nn.Linear]): A list of fully-connected layers to be scaled.
scales (torch.Tensor): A 1D tensor of size (num_features,).
"""
if not isinstance(fcs, list):
fcs = [fcs]
scales = scales.to(ln.weight.device)
ln.weight.div_(scales)
if hasattr(ln, "bias") and ln.bias is not None:
ln.bias.div_(scales)
for fc in fcs:
fc.weight.mul_(scales.view(1, -1))
for p in ln.parameters():
assert torch.isnan(p).sum() == 0
for fc in fcs:
for p in fc.parameters():
assert torch.isnan(p).sum() == 0
@torch.no_grad()
def scale_fc_fc(fc1, fc2, scales):
"""
Scales the weights of two fully-connected layers in a specific pattern.
Args:
fc1 (nn.Linear): The first fully-connected layer to be scaled.
fc2 (nn.Linear): The second fully-connected layer to be scaled.
scales (torch.Tensor): A 1D tensor of size (num_features,).
"""
assert isinstance(fc1, nn.Linear)
assert isinstance(fc2, nn.Linear)
scales = scales.to(fc1.weight.device)
fc1.weight[-scales.size(0):].div_(scales.view(-1, 1))
if fc1.bias is not None:
fc1.bias.div_(scales.view(-1))
fc2.weight.mul_(scales.view(1, -1))
for p in fc1.parameters():
assert torch.isnan(p).sum() == 0
for p in fc2.parameters():
assert torch.isnan(p).sum() == 0
@torch.no_grad()
def scale_gelu_fc(gelu, fc, scales):
"""
Scales the weight of a GELU activation and a fully-connected layer proportionally.
Args:
gelu (Union[nn.GELU, BloomGelu, GELUActivation]): The GELU activation module to be scaled.
fc (nn.Linear): The fully-connected layer to be scaled.
scales (torch.Tensor): A 1D tensor of size (num_features,).
Raises:
TypeError: If the `gelu` module is not of type `nn.GELU`, `BloomGelu`, or `GELUActivation`.
TypeError: If the `fc` module is not of type `nn.Linear`.
"""
assert isinstance(gelu, (nn.GELU, BloomGelu, GELUActivation))
assert isinstance(fc, nn.Linear)
fc.weight.mul_(scales.view(1, -1).to(fc.weight.device))
for p in fc.parameters():
assert torch.isnan(p).sum() == 0
def apply_scale(module, scales_list, input_feat_dict=None):
"""
Applies different scaling strategies to layers based on their type and hierarchy within a given module.
Args:
module (nn.Module): The module containing the layers to be scaled.
scales_list (List[Tuple[str, List[str], torch.Tensor]]): A list of tuples containing:
* prev_op_name (str): The name of the preceding operation or module,
relative to which the layers to be scaled are located.
* layer_names (List[str]): A list of names of the layers to be scaled, relative to the preceding operation.
* scales (torch.Tensor): A 1D tensor of size (num_features,) containing the scaling factors for each feature.
input_feat_dict (Optional[Dict[str, torch.Tensor]]): A dictionary mapping layer names to their corresponding
input features (optional).
"""
for prev_op_name, layer_names, scales in scales_list:
prev_op = get_op_by_name(module, prev_op_name)
layers = [get_op_by_name(module, name) for name in layer_names]
prev_op.cuda()
for layer in layers:
layer.cuda()
scales.cuda()
if isinstance(prev_op, nn.Linear):
assert len(layers) == 1
scale_fc_fc(prev_op, layers[0], scales)
elif isinstance(prev_op, (nn.LayerNorm, LlamaRMSNorm)) or "rmsnorm" in str(prev_op.__class__).lower():
scale_ln_fcs(prev_op, layers, scales)
elif isinstance(prev_op, (nn.GELU, BloomGelu, GELUActivation)):
new_module = ScaledActivation(prev_op, scales)
set_op_by_name(module, prev_op_name, new_module)
scale_gelu_fc(prev_op, layers[0], scales)
else:
raise NotImplementedError(f"prev_op {type(prev_op)} not supported yet!")
# apply the scaling to input feat if given; prepare it for clipping
if input_feat_dict is not None:
for layer_name in layer_names:
inp = input_feat_dict[layer_name]
inp.div_(scales.view(1, -1).to(inp.device))
prev_op.cpu()
for layer in layers:
layer.cpu()
scales.cpu()
@torch.no_grad()
def apply_clip(module, clip_list):
"""
Applies element-wise clipping to the weight of a specific layer within a given module.
Args:
module (nn.Module): The module containing the layer to be clipped.
clip_list (List[Tuple[str, torch.Tensor]]): A list of tuples containing:
* name (str): The name of the layer to be clipped, relative to the root of the module.
* max_val (torch.Tensor): A 1D or 2D tensor defining the upper bound for each element of the layer's weight.
"""
for name, max_val in clip_list:
layer = get_op_by_name(module, name)
layer.cuda()
max_val = max_val.to(layer.weight.device)
org_shape = layer.weight.shape
layer.weight.data = layer.weight.data.reshape(*max_val.shape[:2], -1)
layer.weight.data = torch.clamp(layer.weight.data, -max_val, max_val)
layer.weight.data = layer.weight.data.reshape(org_shape)
layer.cpu()
def add_scale_weights(model_path, scale_path, tmp_path):
"""
Adds pre-computed Activation Weight Quantization (AWQ) results to a model,
including scaling factors and clipping bounds.
Args:
model_path (str): Path to the pre-trained model to be equipped with AWQ.
scale_path (str): Path to the AWQ scale factors (.pt file).
tmp_path (str): Path to the temporary directory where the equipped model will be saved.
"""
config = AutoConfig.from_pretrained(model_path, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
model_path, config=config, trust_remote_code=True
)
model.eval()
awq_results = torch.load(str(scale_path), map_location="cpu")
apply_scale(model, awq_results["scale"])
apply_clip(model, awq_results["clip"])
model.save_pretrained(str(tmp_path))
os.system(f"cp {str(model_path)}/tokenizer* {str(tmp_path)}")

2
awq-py/requirements.txt
View file

@ -1,2 +0,0 @@
torch>=2.1.1
transformers>=4.32.0

34
ci/run.sh
View file

@ -272,19 +272,19 @@ function gg_run_open_llama_3b_v2 {
(time ./bin/main --model ${model_q5_k} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
(time ./bin/main --model ${model_q6_k} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
(time ./bin/perplexity --model ${model_f16} -f ${wiki_test_60} -c 128 -b 128 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
(time ./bin/perplexity --model ${model_q8_0} -f ${wiki_test_60} -c 128 -b 128 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
(time ./bin/perplexity --model ${model_q4_0} -f ${wiki_test_60} -c 128 -b 128 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
(time ./bin/perplexity --model ${model_q4_1} -f ${wiki_test_60} -c 128 -b 128 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
(time ./bin/perplexity --model ${model_q5_0} -f ${wiki_test_60} -c 128 -b 128 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
(time ./bin/perplexity --model ${model_q5_1} -f ${wiki_test_60} -c 128 -b 128 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
(time ./bin/perplexity --model ${model_q2_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
(time ./bin/perplexity --model ${model_q3_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
(time ./bin/perplexity --model ${model_q4_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
(time ./bin/perplexity --model ${model_q5_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
(time ./bin/perplexity --model ${model_q6_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
(time ./bin/perplexity --model ${model_f16} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
(time ./bin/perplexity --model ${model_q8_0} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
(time ./bin/perplexity --model ${model_q4_0} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
(time ./bin/perplexity --model ${model_q4_1} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
(time ./bin/perplexity --model ${model_q5_0} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
(time ./bin/perplexity --model ${model_q5_1} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
(time ./bin/perplexity --model ${model_q2_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
(time ./bin/perplexity --model ${model_q3_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
(time ./bin/perplexity --model ${model_q4_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
(time ./bin/perplexity --model ${model_q5_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
(time ./bin/perplexity --model ${model_q6_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
(time ./bin/imatrix --model ${model_f16} -f ${wiki_test_60} -c 128 -b 128 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-imatrix.log
(time ./bin/imatrix --model ${model_f16} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-imatrix.log
(time ./bin/save-load-state --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
@ -343,17 +343,17 @@ function gg_run_open_llama_3b_v2 {
python3 ../convert-lora-to-ggml.py ${path_lora}
# f16
(time ./bin/perplexity --model ${model_f16} -f ${shakespeare} -c 128 -b 128 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-ppl-shakespeare-f16.log
(time ./bin/perplexity --model ${model_f16} -f ${shakespeare} --lora ${lora_shakespeare} -c 128 -b 128 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-ppl-shakespeare-lora-f16.log
(time ./bin/perplexity --model ${model_f16} -f ${shakespeare} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-ppl-shakespeare-f16.log
(time ./bin/perplexity --model ${model_f16} -f ${shakespeare} --lora ${lora_shakespeare} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-ppl-shakespeare-lora-f16.log
compare_ppl "f16 shakespeare" "$(cat $OUT/${ci}-ppl-shakespeare-f16.log | grep "^\[1\]")" "$(cat $OUT/${ci}-ppl-shakespeare-lora-f16.log | grep "^\[1\]")" | tee -a $OUT/${ci}-lora-ppl.log
# q8_0
(time ./bin/perplexity --model ${model_q8_0} -f ${shakespeare} -c 128 -b 128 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-ppl-shakespeare-q8_0.log
(time ./bin/perplexity --model ${model_q8_0} -f ${shakespeare} --lora ${lora_shakespeare} -c 128 -b 128 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-ppl-shakespeare-lora-q8_0.log
(time ./bin/perplexity --model ${model_q8_0} -f ${shakespeare} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-ppl-shakespeare-q8_0.log
(time ./bin/perplexity --model ${model_q8_0} -f ${shakespeare} --lora ${lora_shakespeare} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-ppl-shakespeare-lora-q8_0.log
compare_ppl "q8_0 shakespeare" "$(cat $OUT/${ci}-ppl-shakespeare-q8_0.log | grep "^\[1\]")" "$(cat $OUT/${ci}-ppl-shakespeare-lora-q8_0.log | grep "^\[1\]")" | tee -a $OUT/${ci}-lora-ppl.log
# q8_0 + f16 lora-base
(time ./bin/perplexity --model ${model_q8_0} -f ${shakespeare} --lora ${lora_shakespeare} --lora-base ${model_f16} -c 128 -b 128 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-ppl-shakespeare-lora-q8_0-f16.log
(time ./bin/perplexity --model ${model_q8_0} -f ${shakespeare} --lora ${lora_shakespeare} --lora-base ${model_f16} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-ppl-shakespeare-lora-q8_0-f16.log
compare_ppl "q8_0 / f16 base shakespeare" "$(cat $OUT/${ci}-ppl-shakespeare-q8_0.log | grep "^\[1\]")" "$(cat $OUT/${ci}-ppl-shakespeare-lora-q8_0-f16.log | grep "^\[1\]")" | tee -a $OUT/${ci}-lora-ppl.log
set +e

27
common/common.cpp
View file

@ -295,9 +295,9 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) {
break;
}
std::string value(argv[i]);
/**/ if (value == "none") { params.rope_scaling_type = LLAMA_ROPE_SCALING_NONE; }
else if (value == "linear") { params.rope_scaling_type = LLAMA_ROPE_SCALING_LINEAR; }
else if (value == "yarn") { params.rope_scaling_type = LLAMA_ROPE_SCALING_YARN; }
/**/ if (value == "none") { params.rope_scaling_type = LLAMA_ROPE_SCALING_TYPE_NONE; }
else if (value == "linear") { params.rope_scaling_type = LLAMA_ROPE_SCALING_TYPE_LINEAR; }
else if (value == "yarn") { params.rope_scaling_type = LLAMA_ROPE_SCALING_TYPE_YARN; }
else { invalid_param = true; break; }
} else if (arg == "--rope-scale") {
if (++i >= argc) {
@ -335,6 +335,12 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) {
break;
}
params.yarn_beta_slow = std::stof(argv[i]);
} else if (arg == "--defrag-thold" || arg == "-dt") {
if (++i >= argc) {
invalid_param = true;
break;
}
params.defrag_thold = std::stof(argv[i]);
} else if (arg == "--samplers") {
if (++i >= argc) {
invalid_param = true;
@ -630,11 +636,11 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) {
}
std::string arg_next = argv[i];
if (arg_next == "none") {
params.split_mode = LLAMA_SPLIT_NONE;
params.split_mode = LLAMA_SPLIT_MODE_NONE;
} else if (arg_next == "layer") {
params.split_mode = LLAMA_SPLIT_LAYER;
params.split_mode = LLAMA_SPLIT_MODE_LAYER;
} else if (arg_next == "row") {
params.split_mode = LLAMA_SPLIT_ROW;
params.split_mode = LLAMA_SPLIT_MODE_ROW;
} else {
invalid_param = true;
break;
@ -837,15 +843,15 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) {
sep++;
if (strncmp(sep, "int:", 4) == 0) {
sep += 4;
kvo.tag = LLAMA_KV_OVERRIDE_INT;
kvo.tag = LLAMA_KV_OVERRIDE_TYPE_INT;
kvo.int_value = std::atol(sep);
} else if (strncmp(sep, "float:", 6) == 0) {
sep += 6;
kvo.tag = LLAMA_KV_OVERRIDE_FLOAT;
kvo.tag = LLAMA_KV_OVERRIDE_TYPE_FLOAT;
kvo.float_value = std::atof(sep);
} else if (strncmp(sep, "bool:", 5) == 0) {
sep += 5;
kvo.tag = LLAMA_KV_OVERRIDE_BOOL;
kvo.tag = LLAMA_KV_OVERRIDE_TYPE_BOOL;
if (std::strcmp(sep, "true") == 0) {
kvo.bool_value = true;
} else if (std::strcmp(sep, "false") == 0) {
@ -1004,6 +1010,8 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
printf(" --yarn-attn-factor N YaRN: scale sqrt(t) or attention magnitude (default: 1.0)\n");
printf(" --yarn-beta-slow N YaRN: high correction dim or alpha (default: %.1f)\n", params.yarn_beta_slow);
printf(" --yarn-beta-fast N YaRN: low correction dim or beta (default: %.1f)\n", params.yarn_beta_fast);
printf(" -dt N, --defrag-thold N\n");
printf(" KV cache defragmentation threshold (default: %.1f, < 0 - disabled)\n", params.defrag_thold);
printf(" --ignore-eos ignore end of stream token and continue generating (implies --logit-bias 2-inf)\n");
printf(" --no-penalize-nl do not penalize newline token\n");
printf(" --temp N temperature (default: %.1f)\n", (double)sparams.temp);
@ -1285,6 +1293,7 @@ struct llama_context_params llama_context_params_from_gpt_params(const gpt_param
cparams.yarn_beta_fast = params.yarn_beta_fast;
cparams.yarn_beta_slow = params.yarn_beta_slow;
cparams.yarn_orig_ctx = params.yarn_orig_ctx;
cparams.defrag_thold = params.defrag_thold;
cparams.offload_kqv = !params.no_kv_offload;
cparams.type_k = kv_cache_type_from_str(params.cache_type_k);

5
common/common.h
View file

@ -61,7 +61,7 @@ struct gpt_params {
float p_split = 0.1f; // speculative decoding split probability
int32_t n_gpu_layers = -1; // number of layers to store in VRAM (-1 - use default)
int32_t n_gpu_layers_draft = -1; // number of layers to store in VRAM for the draft model (-1 - use default)
llama_split_mode split_mode = LLAMA_SPLIT_LAYER; // how to split the model across GPUs
llama_split_mode split_mode = LLAMA_SPLIT_MODE_LAYER; // how to split the model across GPUs
int32_t main_gpu = 0; // the GPU that is used for scratch and small tensors
float tensor_split[128] = {0}; // how split tensors should be distributed across GPUs
int32_t n_beams = 0; // if non-zero then use beam search of given width.
@ -75,7 +75,8 @@ struct gpt_params {
float yarn_beta_fast = 32.0f; // YaRN low correction dim
float yarn_beta_slow = 1.0f; // YaRN high correction dim
int32_t yarn_orig_ctx = 0; // YaRN original context length
int32_t rope_scaling_type = LLAMA_ROPE_SCALING_UNSPECIFIED;
float defrag_thold = -1.0f; // KV cache defragmentation threshold
int32_t rope_scaling_type = LLAMA_ROPE_SCALING_TYPE_UNSPECIFIED;
ggml_numa_strategy numa = GGML_NUMA_STRATEGY_DISABLED;
// // sampling parameters

2
common/sampling.cpp
View file

@ -266,7 +266,7 @@ static llama_token llama_sampling_sample_impl(
// }
//}
LOG("sampled token: %5d: '%s'\n", id, llama_token_to_piece(ctx_main, id).c_str());
//LOG("sampled token: %5d: '%s'\n", id, llama_token_to_piece(ctx_main, id).c_str());
}
}

10
common/train.cpp
View file

@ -31,7 +31,7 @@ struct train_state * init_train_state() {
state->opt = new struct ggml_opt_context;
state->opt->ctx = NULL;
state->opt->params = ggml_opt_default_params(GGML_OPT_ADAM);
state->opt->params = ggml_opt_default_params(GGML_OPT_TYPE_ADAM);
state->opt->params.graph_size = LLAMA_TRAIN_MAX_NODES;
state->opt->loss_after = 0.0f;
@ -556,7 +556,7 @@ void load_opt_context_gguf(struct gguf_context * fctx, struct ggml_context * f_g
std::string opt_type;
GGUF_GET_KEY(fctx, opt_type, gguf_get_val_str, GGUF_TYPE_STRING, true, LLM_KV_OPTIMIZER_TYPE);
if (opt_type == LLM_KV_OPTIMIZER_TYPE_ADAM) {
opt->params.type = GGML_OPT_ADAM;
opt->params.type = GGML_OPT_TYPE_ADAM;
GGUF_GET_KEY(fctx, opt->adam.fx_best, gguf_get_val_f32, GGUF_TYPE_FLOAT32, true, LLM_KV_OPTIMIZER_ADAM_BEST_LOSS);
GGUF_GET_KEY(fctx, opt->adam.fx_prev, gguf_get_val_f32, GGUF_TYPE_FLOAT32, true, LLM_KV_OPTIMIZER_ADAM_PREVIOUS_LOSS);
@ -568,7 +568,7 @@ void load_opt_context_gguf(struct gguf_context * fctx, struct ggml_context * f_g
copy_tensor_by_name(opt->adam.v, f_ggml_ctx, LLM_TENSOR_OPTIMIZER_ADAM_SECOND_MOMENTS);
copy_tensor_by_name(opt->adam.pf, f_ggml_ctx, LLM_TENSOR_OPTIMIZER_ADAM_PAST_LOSS_VALUES);
} else if (opt_type == LLM_KV_OPTIMIZER_TYPE_LBFGS) {
opt->params.type = GGML_OPT_LBFGS;
opt->params.type = GGML_OPT_TYPE_LBFGS;
GGUF_GET_KEY(fctx, opt->params.lbfgs.m, gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_OPTIMIZER_LBFGS_APPROX_HESSIAN_COUNT);
GGUF_GET_KEY(fctx, opt->lbfgs.fx_best, gguf_get_val_f32, GGUF_TYPE_FLOAT32, true, LLM_KV_OPTIMIZER_LBFGS_BEST_LOSS);
@ -603,7 +603,7 @@ void save_opt_context_gguf(struct gguf_context * fctx, struct ggml_opt_context *
gguf_set_val_bool(fctx, LLM_KV_OPTIMIZER_JUST_INITIALIZED, opt->just_initialized);
switch (opt->params.type) {
case GGML_OPT_ADAM:
case GGML_OPT_TYPE_ADAM:
{
gguf_set_val_str(fctx, LLM_KV_OPTIMIZER_TYPE, LLM_KV_OPTIMIZER_TYPE_ADAM);
gguf_set_val_f32(fctx, LLM_KV_OPTIMIZER_ADAM_BEST_LOSS, opt->adam.fx_best);
@ -622,7 +622,7 @@ void save_opt_context_gguf(struct gguf_context * fctx, struct ggml_opt_context *
gguf_add_tensor(fctx, opt->adam.pf);
}
} break;
case GGML_OPT_LBFGS:
case GGML_OPT_TYPE_LBFGS:
{
gguf_set_val_str(fctx, LLM_KV_OPTIMIZER_TYPE, LLM_KV_OPTIMIZER_TYPE_LBFGS);
gguf_set_val_u32(fctx, LLM_KV_OPTIMIZER_LBFGS_APPROX_HESSIAN_COUNT, opt->params.lbfgs.m);

10
convert-hf-to-gguf.py
View file

@ -192,7 +192,7 @@ class Model:
return RefactModel
if model_architecture == "PersimmonForCausalLM":
return PersimmonModel
if model_architecture in ("StableLMEpochForCausalLM", "LlavaStableLMEpochForCausalLM"):
if model_architecture in ("StableLmForCausalLM", "StableLMEpochForCausalLM", "LlavaStableLMEpochForCausalLM"):
return StableLMModel
if model_architecture == "QWenLMHeadModel":
return QwenModel
@ -253,7 +253,7 @@ class Model:
return gguf.MODEL_ARCH.REFACT
if arch == "PersimmonForCausalLM":
return gguf.MODEL_ARCH.PERSIMMON
if arch in ("StableLMEpochForCausalLM", "LlavaStableLMEpochForCausalLM"):
if arch in ("StableLmForCausalLM", "StableLMEpochForCausalLM", "LlavaStableLMEpochForCausalLM"):
return gguf.MODEL_ARCH.STABLELM
if arch == "QWenLMHeadModel":
return gguf.MODEL_ARCH.QWEN
@ -1074,10 +1074,11 @@ class StableLMModel(Model):
self.gguf_writer.add_embedding_length(hparams["hidden_size"])
self.gguf_writer.add_block_count(block_count)
self.gguf_writer.add_feed_forward_length(hparams["intermediate_size"])
self.gguf_writer.add_rope_dimension_count(int(hparams["rope_pct"] * (hparams["hidden_size"] // hparams["num_attention_heads"])))
rotary_factor = self.find_hparam(["partial_rotary_factor", "rope_pct"])
self.gguf_writer.add_rope_dimension_count(int(rotary_factor * (hparams["hidden_size"] // hparams["num_attention_heads"])))
self.gguf_writer.add_head_count(hparams["num_attention_heads"])
self.gguf_writer.add_parallel_residual(hparams["use_parallel_residual"] if "use_parallel_residual" in hparams else True)
self.gguf_writer.add_layer_norm_eps(1e-5)
self.gguf_writer.add_layer_norm_eps(self.find_hparam(["layer_norm_eps", "norm_eps"]))
class MixtralModel(Model):
@ -1803,6 +1804,7 @@ class GemmaModel(Model):
self.gguf_writer.add_layer_norm_rms_eps(self.hparams["rms_norm_eps"])
self.gguf_writer.add_key_length(hparams["head_dim"])
self.gguf_writer.add_value_length(hparams["head_dim"])
self.gguf_writer.add_file_type(self.ftype)
def write_tensors(self):
block_count = self.hparams.get("n_layers", self.hparams.get("num_hidden_layers", self.hparams.get("n_layer")))

2
examples/baby-llama/baby-llama.cpp
View file

@ -1547,7 +1547,7 @@ int main(int argc, char ** argv) {
float error_before_opt = ggml_get_f32_1d(e, 0);
struct ggml_opt_params opt_params_lbfgs = ggml_opt_default_params(GGML_OPT_LBFGS);
struct ggml_opt_params opt_params_lbfgs = ggml_opt_default_params(GGML_OPT_TYPE_LBFGS);
opt_params_lbfgs.print_forward_graph = false;
opt_params_lbfgs.print_backward_graph = false;
opt_params_lbfgs.lbfgs.n_iter = 16;

2
examples/finetune/finetune.cpp
View file

@ -1531,7 +1531,7 @@ int main(int argc, char ** argv) {
lora.hparams.n_rank_output = n_rank_output;
// set opt params from command line
opt->params = ggml_opt_default_params(GGML_OPT_ADAM);
opt->params = ggml_opt_default_params(GGML_OPT_TYPE_ADAM);
opt->params.print_forward_graph = false;
opt->params.print_backward_graph = false;
opt->params.graph_size = LLAMA_TRAIN_MAX_NODES;

4
examples/infill/infill.cpp
View file

@ -447,8 +447,8 @@ int main(int argc, char ** argv) {
LOG("context full, swapping: n_past = %d, n_left = %d, n_ctx = %d, n_keep = %d, n_discard = %d\n",
n_past, n_left, n_ctx, params.n_keep, n_discard);
llama_kv_cache_seq_rm (ctx, 0, params.n_keep + 1 , params.n_keep + n_discard + 1);
llama_kv_cache_seq_shift(ctx, 0, params.n_keep + 1 + n_discard, n_past, -n_discard);
llama_kv_cache_seq_rm (ctx, 0, params.n_keep + 1 , params.n_keep + n_discard + 1);
llama_kv_cache_seq_add(ctx, 0, params.n_keep + 1 + n_discard, n_past, -n_discard);
n_past -= n_discard;

14
examples/llama-bench/llama-bench.cpp
View file

@ -157,9 +157,9 @@ static const char * output_format_str(output_formats format) {
static const char * split_mode_str(llama_split_mode mode) {
switch (mode) {
case LLAMA_SPLIT_NONE: return "none";
case LLAMA_SPLIT_LAYER: return "layer";
case LLAMA_SPLIT_ROW: return "row";
case LLAMA_SPLIT_MODE_NONE: return "none";
case LLAMA_SPLIT_MODE_LAYER: return "layer";
case LLAMA_SPLIT_MODE_ROW: return "row";
default: GGML_ASSERT(!"invalid split mode");
}
}
@ -193,7 +193,7 @@ static const cmd_params cmd_params_defaults = {
/* type_v */ {GGML_TYPE_F16},
/* n_threads */ {get_num_physical_cores()},
/* n_gpu_layers */ {99},
/* split_mode */ {LLAMA_SPLIT_LAYER},
/* split_mode */ {LLAMA_SPLIT_MODE_LAYER},
/* main_gpu */ {0},
/* no_kv_offload */ {false},
/* mul_mat_q */ {true},
@ -358,11 +358,11 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
for (const auto & m : p) {
llama_split_mode mode;
if (m == "none") {
mode = LLAMA_SPLIT_NONE;
mode = LLAMA_SPLIT_MODE_NONE;
} else if (m == "layer") {
mode = LLAMA_SPLIT_LAYER;
mode = LLAMA_SPLIT_MODE_LAYER;
} else if (m == "row") {
mode = LLAMA_SPLIT_ROW;
mode = LLAMA_SPLIT_MODE_ROW;
} else {
invalid_param = true;
break;

8
examples/llama.android/app/build.gradle.kts
View file

@ -21,12 +21,8 @@ android {
useSupportLibrary = true
}
ndk {
// Workaround for https://github.com/llvm/llvm-project/issues/65820
// affecting armeabi-v7a. Skip armeabi-v7a when invoked with
// -Pskip-armeabi-v7a (e.g., ./gradlew build -Pskip-armeabi-v7a).
if (project.hasProperty("skip-armeabi-v7a")) {
abiFilters += listOf("arm64-v8a", "x86_64", "x86")
}
// Add NDK properties if wanted, e.g.
// abiFilters += listOf("arm64-v8a")
}
externalNativeBuild {
cmake {

2
examples/llava/llava.cpp
View file

@ -152,7 +152,7 @@ static bool clip_llava_handle_patches(clip_ctx * ctx_clip, std::vector<float *>
ggml_tensor * newline_tmp = clip_get_newline_tensor(ctx_clip);
model.newline = ggml_new_tensor_1d(model.ctx, GGML_TYPE_F32, newline_tmp->ne[0]);
if (newline_tmp->backend != GGML_BACKEND_CPU) {
if (newline_tmp->backend != GGML_BACKEND_TYPE_CPU) {
if (newline_tmp->buffer == NULL) {
printf("newline_tmp tensor buffer is NULL\n");
}

10
examples/main/main.cpp
View file

@ -548,8 +548,8 @@ int main(int argc, char ** argv) {
LOG("context full, swapping: n_past = %d, n_left = %d, n_ctx = %d, n_keep = %d, n_discard = %d\n",
n_past, n_left, n_ctx, params.n_keep, n_discard);
llama_kv_cache_seq_rm (ctx, 0, params.n_keep , params.n_keep + n_discard);
llama_kv_cache_seq_shift(ctx, 0, params.n_keep + n_discard, n_past, -n_discard);
llama_kv_cache_seq_rm (ctx, 0, params.n_keep , params.n_keep + n_discard);
llama_kv_cache_seq_add(ctx, 0, params.n_keep + n_discard, n_past, -n_discard);
n_past -= n_discard;
@ -576,9 +576,9 @@ int main(int argc, char ** argv) {
LOG("div: [%6d, %6d] / %6d -> [%6d, %6d]\n", ga_i + ib*bd, ga_i + ib*bd + ga_w, ga_n, (ga_i + ib*bd)/ga_n, (ga_i + ib*bd + ga_w)/ga_n);
LOG("shift: [%6d, %6d] + %6d -> [%6d, %6d]\n", ga_i + ib*bd + ga_w, n_past + ib*bd, dd, ga_i + ib*bd + ga_w + dd, n_past + ib*bd + dd);
llama_kv_cache_seq_shift(ctx, 0, ga_i, n_past, ib*bd);
llama_kv_cache_seq_div (ctx, 0, ga_i + ib*bd, ga_i + ib*bd + ga_w, ga_n);
llama_kv_cache_seq_shift(ctx, 0, ga_i + ib*bd + ga_w, n_past + ib*bd, dd);
llama_kv_cache_seq_add(ctx, 0, ga_i, n_past, ib*bd);
llama_kv_cache_seq_div(ctx, 0, ga_i + ib*bd, ga_i + ib*bd + ga_w, ga_n);
llama_kv_cache_seq_add(ctx, 0, ga_i + ib*bd + ga_w, n_past + ib*bd, dd);
n_past -= bd;

25
examples/passkey/passkey.cpp
View file

@ -126,7 +126,7 @@ int main(int argc, char ** argv) {
const int n_batch = ctx_params.n_batch;
const int n_batch_grp = ctx_params.n_batch/n_grp;
LOG_TEE("\n%s: n_len = %d, n_ctx = %d, n_kv_req = %d, n_grp = %d, n_batch = %d\n", __func__, n_len, n_ctx, n_kv_req, n_grp, n_batch);
LOG_TEE("\n%s: n_len = %d, n_ctx = %d, n_kv_req = %d, n_grp = %d, n_batch = %d, n_junk = %d, i_pos = %d\n", __func__, n_len, n_ctx, n_kv_req, n_grp, n_batch, n_junk, i_pos);
// print the prompt token-by-token
@ -146,10 +146,11 @@ int main(int argc, char ** argv) {
const int ib = i/n_batch - 1;
const int bd = n_batch_grp*(n_grp - 1);
llama_kv_cache_seq_shift(ctx, 0, n_past - n_batch, n_past, ib*bd);
llama_kv_cache_seq_div (ctx, 0, n_past - n_batch + ib*bd, n_past + ib*bd, n_grp);
llama_kv_cache_seq_add (ctx, 0, n_past - n_batch, n_past, ib*bd);
llama_kv_cache_seq_div (ctx, 0, n_past - n_batch + ib*bd, n_past + ib*bd, n_grp);
llama_kv_cache_update (ctx);
n_past -= bd;
n_past = llama_kv_cache_seq_pos_max(ctx, 0) + 1;
}
llama_batch_clear(batch);
@ -179,10 +180,12 @@ int main(int argc, char ** argv) {
LOG_TEE("%s: shifting KV cache with %d\n", __func__, n_discard);
llama_kv_cache_seq_rm (ctx, 0, n_keep , n_keep + n_discard);
llama_kv_cache_seq_shift(ctx, 0, n_keep + n_discard, n_ctx, -n_discard);
llama_kv_cache_seq_rm (ctx, 0, n_keep , n_keep + n_discard);
llama_kv_cache_seq_add(ctx, 0, n_keep + n_discard, n_ctx, -n_discard);
//llama_kv_cache_defrag (ctx);
llama_kv_cache_update (ctx);
n_past -= n_discard;
n_past = llama_kv_cache_seq_pos_max(ctx, 0) + 1;
llama_batch_clear(batch);
@ -208,10 +211,12 @@ int main(int argc, char ** argv) {
if (n_discard > 0) {
LOG_TEE("%s: shifting KV cache with %d to free space for the answer\n", __func__, n_discard);
llama_kv_cache_seq_rm (ctx, 0, n_keep , n_keep + n_discard);
llama_kv_cache_seq_shift(ctx, 0, n_keep + n_discard, n_ctx, -n_discard);
llama_kv_cache_seq_rm (ctx, 0, n_keep , n_keep + n_discard);
llama_kv_cache_seq_add(ctx, 0, n_keep + n_discard, n_ctx, -n_discard);
//llama_kv_cache_defrag (ctx);
llama_kv_cache_update (ctx);
n_past -= n_discard;
n_past = llama_kv_cache_seq_pos_max(ctx, 0) + 1;
}
}

10
examples/quantize/quantize.cpp
View file

@ -23,16 +23,21 @@ static const std::vector<struct quant_option> QUANT_OPTIONS = {
{ "Q5_1", LLAMA_FTYPE_MOSTLY_Q5_1, " 4.70G, +0.0349 ppl @ LLaMA-v1-7B", },
{ "IQ2_XXS",LLAMA_FTYPE_MOSTLY_IQ2_XXS," 2.06 bpw quantization", },
{ "IQ2_XS", LLAMA_FTYPE_MOSTLY_IQ2_XS, " 2.31 bpw quantization", },
{ "IQ2_S", LLAMA_FTYPE_MOSTLY_IQ2_S, " 2.5 bpw quantization", },
{ "IQ2_M", LLAMA_FTYPE_MOSTLY_IQ2_M, " 2.7 bpw quantization", },
{ "IQ1_S", LLAMA_FTYPE_MOSTLY_IQ1_S, " 1.56 bpw quantization", },
{ "Q2_K", LLAMA_FTYPE_MOSTLY_Q2_K, " 2.63G, +0.6717 ppl @ LLaMA-v1-7B", },
{ "Q2_K_S", LLAMA_FTYPE_MOSTLY_Q2_K_S, " 2.16G, +9.0634 ppl @ LLaMA-v1-7B", },
{ "IQ3_XXS",LLAMA_FTYPE_MOSTLY_IQ3_XXS," 3.06 bpw quantization", },
{ "IQ3_S", LLAMA_FTYPE_MOSTLY_IQ3_S, " 3.44 bpw quantization", },
{ "IQ3_M", LLAMA_FTYPE_MOSTLY_IQ3_M, " 3.66 bpw quantization mix", },
{ "Q3_K", LLAMA_FTYPE_MOSTLY_Q3_K_M, "alias for Q3_K_M" },
{ "Q3_K_XS",LLAMA_FTYPE_MOSTLY_Q3_K_XS,"3-bit extra small quantization" , },
{ "IQ3_XS", LLAMA_FTYPE_MOSTLY_IQ3_XS, " 3.3 bpw quantization" , },
{ "Q3_K_S", LLAMA_FTYPE_MOSTLY_Q3_K_S, " 2.75G, +0.5551 ppl @ LLaMA-v1-7B", },
{ "Q3_K_M", LLAMA_FTYPE_MOSTLY_Q3_K_M, " 3.07G, +0.2496 ppl @ LLaMA-v1-7B", },
{ "Q3_K_L", LLAMA_FTYPE_MOSTLY_Q3_K_L, " 3.35G, +0.1764 ppl @ LLaMA-v1-7B", },
{ "IQ4_NL", LLAMA_FTYPE_MOSTLY_IQ4_NL, " 4.25 bpw non-linear quantization", },
{ "IQ4_NL", LLAMA_FTYPE_MOSTLY_IQ4_NL, " 4.50 bpw non-linear quantization", },
{ "IQ4_XS", LLAMA_FTYPE_MOSTLY_IQ4_XS, " 4.25 bpw non-linear quantization", },
{ "Q4_K", LLAMA_FTYPE_MOSTLY_Q4_K_M, "alias for Q4_K_M", },
{ "Q4_K_S", LLAMA_FTYPE_MOSTLY_Q4_K_S, " 3.59G, +0.0992 ppl @ LLaMA-v1-7B", },
{ "Q4_K_M", LLAMA_FTYPE_MOSTLY_Q4_K_M, " 3.80G, +0.0532 ppl @ LLaMA-v1-7B", },
@ -290,6 +295,7 @@ int main(int argc, char ** argv) {
}
if ((params.ftype == LLAMA_FTYPE_MOSTLY_IQ2_XS || params.ftype == LLAMA_FTYPE_MOSTLY_IQ2_XXS ||
params.ftype == LLAMA_FTYPE_MOSTLY_IQ2_S ||
params.ftype == LLAMA_FTYPE_MOSTLY_Q2_K_S || params.ftype == LLAMA_FTYPE_MOSTLY_IQ1_S) && imatrix_data.empty()) {
fprintf(stderr, "\n===============================================================================================\n");
fprintf(stderr, "Please do not use IQ1_S, IQ2_XXS, IQ2_XS or Q2_K_S quantization without an importance matrix\n");

41
examples/server/README.md
View file

@ -1,8 +1,20 @@
# llama.cpp/example/server
# LLaMA.cpp HTTP Server
This example demonstrates a simple HTTP API server and a simple web front end to interact with llama.cpp.
Fast, lightweight, pure C/C++ HTTP server based on [httplib](https://github.com/yhirose/cpp-httplib), [nlohmann::json](https://github.com/nlohmann/json) and **llama.cpp**.
Command line options:
Set of LLM REST APIs and a simple web front end to interact with llama.cpp.
**Features:**
* LLM inference of F16 and quantum models on GPU and CPU
* [OpenAI API](https://github.com/openai/openai-openapi) compatible chat completions and embeddings routes
* Parallel decoding with multi-user support
* Continuous batching
* Multimodal (wip)
* Monitoring endpoints
The project is under active development, and we are [looking for feedback and contributors](https://github.com/ggerganov/llama.cpp/issues/4216).
**Command line options:**
- `--threads N`, `-t N`: Set the number of threads to use during generation.
- `-tb N, --threads-batch N`: Set the number of threads to use during batch and prompt processing. If not specified, the number of threads will be set to the number of threads used for generation.
@ -39,9 +51,12 @@ see https://github.com/ggerganov/llama.cpp/issues/1437
- `--mmproj MMPROJ_FILE`: Path to a multimodal projector file for LLaVA.
- `--grp-attn-n`: Set the group attention factor to extend context size through self-extend(default: 1=disabled), used together with group attention width `--grp-attn-w`
- `--grp-attn-w`: Set the group attention width to extend context size through self-extend(default: 512), used together with group attention factor `--grp-attn-n`
- `-n, --n-predict`: Set the maximum tokens to predict (default: -1)
- `-n N, --n-predict N`: Set the maximum tokens to predict (default: -1)
- `--slots-endpoint-disable`: To disable slots state monitoring endpoint. Slots state may contain user data, prompts included.
- `--metrics`: enable prometheus `/metrics` compatible endpoint (default: disabled)
- `--chat-template JINJA_TEMPLATE`: Set custom jinja chat template. This parameter accepts a string, not a file name (default: template taken from model's metadata). We only support [some pre-defined templates](https://github.com/ggerganov/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template)
- `--log-disable`: Output logs to stdout only, default: enabled.
- `--log-format FORMAT`: Define the log output to FORMAT: json or text (default: json)
## Build
@ -98,6 +113,12 @@ curl --request POST \
--data '{"prompt": "Building a website can be done in 10 simple steps:","n_predict": 128}'
```
## Advanced testing
We implemented a [server test framework](./tests/README.md) using human-readable scenario.
*Before submitting an issue, please try to reproduce it with this format.*
## Node JS Test
You need to have [Node.js](https://nodejs.org/en) installed.
@ -451,6 +472,18 @@ Notice that each `probs` is an array of length `n_probs`.
]
```
- **GET** `/metrics`: [Prometheus](https://prometheus.io/) compatible metrics exporter endpoint if `--metrics` is enabled:
Available metrics:
- `llamacpp:prompt_tokens_total`: Number of prompt tokens processed.
- `llamacpp:tokens_predicted_total`: Number of generation tokens processed.
- `llamacpp:prompt_tokens_seconds`: Average prompt throughput in tokens/s.
- `llamacpp:predicted_tokens_seconds`: Average generation throughput in tokens/s.
- `llamacpp:kv_cache_usage_ratio`: KV-cache usage. 1 means 100 percent usage.
- `llamacpp:kv_cache_tokens`: KV-cache tokens.
- `llamacpp:requests_processing`: Number of request processing.
- `llamacpp:requests_deferred`: Number of request deferred.
## More examples
### Change system prompt on runtime

618
examples/server/server.cpp
View file

@ -43,9 +43,11 @@ struct server_params
int32_t read_timeout = 600;
int32_t write_timeout = 600;
bool slots_endpoint = true;
bool metrics_endpoint = false;
};
bool server_verbose = false;
bool server_log_json = true;
static size_t common_part(const std::vector<llama_token> &a, const std::vector<llama_token> &b)
{
@ -301,12 +303,76 @@ struct llama_client_slot
}
void print_timings() const {
LOG_TEE("\n");
LOG_TEE("%s: prompt eval time = %10.2f ms / %5d tokens (%8.2f ms per token, %8.2f tokens per second)\n",
__func__, t_prompt_processing, num_prompt_tokens_processed, t_prompt_processing / num_prompt_tokens_processed, 1e3 / t_prompt_processing * num_prompt_tokens_processed);
LOG_TEE("%s: eval time = %10.2f ms / %5d runs (%8.2f ms per token, %8.2f tokens per second)\n",
__func__, t_token_generation, n_decoded,t_token_generation / n_decoded, 1e3 / t_token_generation * n_decoded);
LOG_TEE("%s: total time = %10.2f ms\n", __func__, t_prompt_processing + t_token_generation);
char buffer[512];
double t_token = t_prompt_processing / num_prompt_tokens_processed;
double n_tokens_second = 1e3 / t_prompt_processing * num_prompt_tokens_processed;
sprintf(buffer, "prompt eval time = %10.2f ms / %5d tokens (%8.2f ms per token, %8.2f tokens per second)",
t_prompt_processing, num_prompt_tokens_processed,
t_token, n_tokens_second);
LOG_INFO(buffer, {
{"slot_id", id},
{"task_id", task_id},
{"t_prompt_processing", t_prompt_processing},
{"num_prompt_tokens_processed", num_prompt_tokens_processed},
{"t_token", t_token},
{"n_tokens_second", n_tokens_second},
});
t_token = t_token_generation / n_decoded;
n_tokens_second = 1e3 / t_token_generation * n_decoded;
sprintf(buffer, "generation eval time = %10.2f ms / %5d runs (%8.2f ms per token, %8.2f tokens per second)",
t_token_generation, n_decoded,
t_token, n_tokens_second);
LOG_INFO(buffer, {
{"slot_id", id},
{"task_id", task_id},
{"t_token_generation", t_token_generation},
{"n_decoded", n_decoded},
{"t_token", t_token},
{"n_tokens_second", n_tokens_second},
});
sprintf(buffer, " total time = %10.2f ms", t_prompt_processing + t_token_generation);
LOG_INFO(buffer, {
{"slot_id", id},
{"task_id", task_id},
{"t_prompt_processing", t_prompt_processing},
{"t_token_generation", t_token_generation},
{"t_total", t_prompt_processing + t_token_generation},
});
}
};
struct llama_metrics {
uint64_t n_prompt_tokens_processed_total = 0;
uint64_t n_tokens_predicted_total = 0;
uint64_t n_prompt_tokens_processed = 0;
uint64_t t_prompt_processing = 0;
uint64_t n_tokens_predicted = 0;
uint64_t t_tokens_generation = 0;
void on_prompt_eval(const llama_client_slot &slot) {
n_prompt_tokens_processed_total += slot.num_prompt_tokens_processed;
n_prompt_tokens_processed += slot.num_prompt_tokens_processed;
t_prompt_processing += slot.t_prompt_processing;
}
void on_prediction(const llama_client_slot &slot) {
n_tokens_predicted_total += slot.n_decoded;
n_tokens_predicted += slot.n_decoded;
t_tokens_generation += slot.t_token_generation;
}
void reset_bucket() {
n_prompt_tokens_processed = 0;
t_prompt_processing = 0;
n_tokens_predicted = 0;
t_tokens_generation = 0;
}
};
@ -344,6 +410,8 @@ struct llama_server_context
llama_server_queue queue_tasks;
llama_server_response queue_results;
llama_metrics metrics;
~llama_server_context()
{
if (ctx)
@ -363,7 +431,7 @@ struct llama_server_context
params = params_;
if (!params.mmproj.empty()) {
multimodal = true;
LOG_TEE("Multi Modal Mode Enabled");
LOG_INFO("Multi Modal Mode Enabled", {});
clp_ctx = clip_model_load(params.mmproj.c_str(), /*verbosity=*/ 1);
if(clp_ctx == nullptr) {
LOG_ERROR("unable to load clip model", {{"model", params.mmproj}});
@ -416,7 +484,7 @@ struct llama_server_context
const int32_t n_ctx_slot = n_ctx / params.n_parallel;
LOG_TEE("Available slots:\n");
LOG_INFO("initializing slots", {{"n_slots", params.n_parallel}});
for (int i = 0; i < params.n_parallel; i++)
{
llama_client_slot slot;
@ -425,7 +493,10 @@ struct llama_server_context
slot.n_ctx = n_ctx_slot;
slot.n_predict = params.n_predict;
LOG_TEE(" -> Slot %i - max context: %i\n", slot.id, n_ctx_slot);
LOG_INFO("new slot", {
{"slot_id", slot.id},
{"n_ctx_slot", slot.n_ctx}
});
const int ga_n = params.grp_attn_n;
const int ga_w = params.grp_attn_w;
@ -435,7 +506,12 @@ struct llama_server_context
GGML_ASSERT(ga_w % ga_n == 0 && "ga_w must be a multiple of ga_n"); // NOLINT
//GGML_ASSERT(n_ctx_train % ga_w == 0 && "n_ctx_train must be a multiple of ga_w"); // NOLINT
//GGML_ASSERT(n_ctx >= n_ctx_train * ga_n && "n_ctx must be at least n_ctx_train * ga_n"); // NOLINT
LOG_TEE(" -> Slot %i - self-extend: ga_n = %d, ga_w = %d\n", slot.id, ga_n, ga_w);
LOG_INFO("slot self-extend", {
{"slot_id", slot.id},
{"ga_n", ga_n},
{"ga_w", ga_w}
});
}
slot.ga_i = 0;
@ -729,10 +805,16 @@ struct llama_server_context
img_sl.img_data = clip_image_u8_init();
if (!clip_image_load_from_bytes(image_buffer.data(), image_buffer.size(), img_sl.img_data))
{
LOG_TEE("slot %i - failed to load image [id: %i]\n", slot->id, img_sl.id);
LOG_ERROR("failed to load image", {
{"slot_id", slot->id},
{"img_sl_id", img_sl.id}
});
return false;
}
LOG_TEE("slot %i - loaded image\n", slot->id);
LOG_VERBOSE("image loaded", {
{"slot_id", slot->id},
{"img_sl_id", img_sl.id}
});
img_sl.request_encode_image = true;
slot->images.push_back(img_sl);
}
@ -792,7 +874,10 @@ struct llama_server_context
all_slots_are_idle = false;
LOG_TEE("slot %i is processing [task id: %i]\n", slot->id, slot->task_id);
LOG_INFO("slot is processing task", {
{"slot_id", slot->id},
{"task_id", slot->task_id},
});
return true;
}
@ -817,10 +902,24 @@ struct llama_server_context
llama_batch_add(batch, system_tokens[i], i, { 0 }, false);
}
if (llama_decode(ctx, batch) != 0)
for (int32_t i = 0; i < (int32_t) batch.n_tokens; i += params.n_batch)
{
LOG_TEE("%s: llama_decode() failed\n", __func__);
return;
const int32_t n_tokens = std::min(params.n_batch, (int32_t) (batch.n_tokens - i));
llama_batch batch_view = {
n_tokens,
batch.token + i,
nullptr,
batch.pos + i,
batch.n_seq_id + i,
batch.seq_id + i,
batch.logits + i,
0, 0, 0, // unused
};
if (llama_decode(ctx, batch_view) != 0)
{
LOG_TEE("%s: llama_decode() failed\n", __func__);
return;
}
}
// assign the system KV cache to all parallel sequences
@ -1237,6 +1336,10 @@ struct llama_server_context
split_multiprompt_task(task_id, task);
}
} else {
// an empty prompt can make slot become buggy
if (task.data.contains("prompt") && task.data["prompt"].is_string() && task.data["prompt"].get<std::string>().empty()) {
task.data["prompt"] = " "; // add a space so that we have one token
}
queue_tasks.post(task);
}
}
@ -1355,7 +1458,7 @@ struct llama_server_context
if (slot == nullptr)
{
// if no slot is available, we defer this task for processing later
LOG_VERBOSE("no slot is available", {});
LOG_VERBOSE("no slot is available", {{"task_id", task.id}});
queue_tasks.defer(task);
break;
}
@ -1404,17 +1507,12 @@ struct llama_server_context
case TASK_TYPE_NEXT_RESPONSE: {
// do nothing
} break;
case TASK_TYPE_SLOTS_DATA: {
case TASK_TYPE_METRICS: {
json slots_data = json::array();
int n_idle_slots = 0;
int n_processing_slots = 0;
for (llama_client_slot &slot: slots) {
if (slot.available()) {
n_idle_slots++;
} else {
n_processing_slots++;
}
json slot_data = get_formated_generation(slot);
slot_data["id"] = slot.id;
slot_data["task_id"] = slot.task_id;
@ -1429,19 +1527,48 @@ struct llama_server_context
{"stopped_limit", slot.stopped_limit},
{"stopping_word", slot.stopping_word},
};
if (slot_data["state"] == IDLE) {
n_idle_slots++;
} else {
n_processing_slots++;
}
slots_data.push_back(slot_data);
}
LOG_TEE("task %i - slots data: idle=%i processing=%i\n", task.id, n_idle_slots, n_processing_slots);
LOG_INFO("slot data", {
{"task_id", task.id},
{"n_idle_slots", n_idle_slots},
{"n_processing_slots", n_processing_slots}
});
LOG_VERBOSE("slot data", {
{"task_id", task.id},
{"n_idle_slots", n_idle_slots},
{"n_processing_slots", n_processing_slots},
{"slots", slots_data}
});
task_result res;
res.id = task.id;
res.multitask_id = task.multitask_id;
res.stop = true;
res.error = false;
res.result_json = {
{ "idle", n_idle_slots },
{ "processing", n_processing_slots },
{ "slots", slots_data }
{ "idle", n_idle_slots },
{ "processing", n_processing_slots },
{ "deferred", queue_tasks.queue_tasks_deferred.size() },
{ "n_prompt_tokens_processed_total", metrics.n_prompt_tokens_processed_total},
{ "n_tokens_predicted_total", metrics.n_tokens_predicted_total},
{ "n_prompt_tokens_processed", metrics.n_prompt_tokens_processed},
{ "t_prompt_processing", metrics.t_prompt_processing},
{ "n_tokens_predicted", metrics.n_tokens_predicted},
{ "t_tokens_generation", metrics.t_tokens_generation},
{ "kv_cache_tokens_count", llama_get_kv_cache_token_count(ctx)},
{ "kv_cache_used_cells", llama_get_kv_cache_used_cells(ctx)},
{ "slots", slots_data },
};
metrics.reset_bucket();
queue_results.send(res);
} break;
}
@ -1469,7 +1596,7 @@ struct llama_server_context
bool update_slots() {
if (system_need_update)
{
LOG_TEE("updating system prompt\n");
LOG_INFO("updating system prompt", {});
update_system_prompt();
}
@ -1479,12 +1606,13 @@ struct llama_server_context
{
if (system_prompt.empty() && clean_kv_cache)
{
LOG_TEE("all slots are idle and system prompt is empty, clear the KV cache\n");
LOG_INFO("all slots are idle and system prompt is empty, clear the KV cache", {});
kv_cache_clear();
}
return true;
}
LOG_VERBOSE("posting NEXT_RESPONSE", {});
task_server task;
task.type = TASK_TYPE_NEXT_RESPONSE;
task.target_id = -1;
@ -1498,12 +1626,22 @@ struct llama_server_context
{
// Shift context
const int n_keep = slot.params.n_keep + add_bos_token;
const int n_left = system_tokens.size() + slot.n_past - n_keep;
const int n_left = (int) system_tokens.size() + slot.n_past - n_keep;
const int n_discard = n_left / 2;
LOG_TEE("slot %d: context shift - n_keep = %d, n_left = %d, n_discard = %d\n", slot.id, n_keep, n_left, n_discard);
llama_kv_cache_seq_rm (ctx, slot.id, n_keep , n_keep + n_discard);
llama_kv_cache_seq_shift(ctx, slot.id, n_keep + n_discard, system_tokens.size() + slot.n_past, -n_discard);
LOG_INFO("slot context shift", {
{"slot_id", slot.id},
{"task_id", slot.task_id},
{"n_keep", n_keep},
{"n_left", n_left},
{"n_discard", n_discard},
{"n_ctx", n_ctx},
{"n_past", slot.n_past},
{"n_system_tokens", system_tokens.size()},
{"n_cache_tokens", slot.cache_tokens.size()}
});
llama_kv_cache_seq_rm (ctx, slot.id, n_keep , n_keep + n_discard);
llama_kv_cache_seq_add(ctx, slot.id, n_keep + n_discard, system_tokens.size() + slot.n_past, -n_discard);
for (size_t i = n_keep + n_discard; i < slot.cache_tokens.size(); i++)
{
@ -1515,17 +1653,12 @@ struct llama_server_context
slot.n_past -= n_discard;
slot.truncated = true;
LOG_VERBOSE("context shift", {
{ "n_ctx", n_ctx },
{ "n_keep", n_keep },
{ "n_left", n_left },
});
}
}
}
// decode any currently ongoing sequences
LOG_VERBOSE("decoding ongoing sequences", {});
for (auto & slot : slots)
{
// release the slot
@ -1535,7 +1668,15 @@ struct llama_server_context
slot.command = NONE;
slot.t_last_used = ggml_time_us();
LOG_TEE("slot %d released (%d tokens in cache)\n", slot.id, (int) slot.cache_tokens.size());
LOG_INFO("slot released", {
{"slot_id", slot.id},
{"task_id", slot.task_id},
{"n_ctx", n_ctx},
{"n_past", slot.n_past},
{"n_system_tokens", system_tokens.size()},
{"n_cache_tokens", slot.cache_tokens.size()},
{"truncated", slot.truncated}
});
queue_tasks.notify_slot_changed();
continue;
@ -1662,6 +1803,14 @@ struct llama_server_context
}
slot.n_past = common_part(slot.cache_tokens, prompt_tokens);
// the last token of the cache is not in the KV cache until the next call to llama_decode
// (it was sampled, pushed into the "cache_tokens", but not yet put in the context)
if (slot.n_past > 0 && slot.n_past == (int32_t) slot.cache_tokens.size())
{
slot.n_past -= 1;
}
slot.num_prompt_tokens_processed = slot.num_prompt_tokens - slot.n_past;
if (slot.ga_n != 1)
@ -1683,7 +1832,12 @@ struct llama_server_context
slot.ga_i = ga_i;
}
LOG_TEE("slot %d : in cache: %i tokens | to process: %i tokens\n", slot.id, slot.n_past, slot.num_prompt_tokens_processed);
LOG_INFO("slot progression", {
{ "slot_id", slot.id },
{ "task_id", slot.task_id },
{ "n_past", slot.n_past },
{ "num_prompt_tokens_processed", slot.num_prompt_tokens_processed }
});
}
slot.cache_tokens = prompt_tokens;
@ -1691,7 +1845,10 @@ struct llama_server_context
if (slot.n_past == slot.num_prompt_tokens && slot.n_past > 0)
{
// we have to evaluate at least 1 token to generate logits.
LOG_TEE("slot %d : we have to evaluate at least 1 token to generate logits\n", slot.id);
LOG_INFO("we have to evaluate at least 1 token to generate logits", {
{ "slot_id", slot.id },
{ "task_id", slot.task_id }
});
slot.n_past--;
if (slot.ga_i > 0)
{
@ -1699,9 +1856,13 @@ struct llama_server_context
}
}
LOG_TEE("slot %d : kv cache rm - [%d, end)\n", slot.id, (int) system_tokens.size() + slot.n_past);
llama_kv_cache_seq_rm(ctx, slot.id, system_tokens.size() + slot.n_past, -1);
int p0 = (int) system_tokens.size() + slot.n_past;
LOG_INFO("kv cache rm [p0, end)", {
{ "slot_id", slot.id },
{ "task_id", slot.task_id },
{ "p0", p0 }
});
llama_kv_cache_seq_rm(ctx, slot.id, p0, -1);
LOG_VERBOSE("prompt ingested", {
{"n_past", slot.n_past},
@ -1736,7 +1897,13 @@ struct llama_server_context
if (has_images && !ingest_images(slot, n_batch))
{
LOG_TEE("failed processing images\n");
LOG_ERROR("failed processing images", {
"slot_id", slot.id,
"task_id", slot.task_id,
});
// FIXME @phymbert: to be properly tested
// early returning without changing the slot state will block the slot for ever
// no one at the moment is checking the return value
return false;
}
@ -1778,9 +1945,9 @@ struct llama_server_context
LOG_TEE("div: [%6d, %6d] / %6d -> [%6d, %6d]\n", slot.ga_i + ib * bd, slot.ga_i + ib * bd + slot.ga_w, slot.ga_n, (slot.ga_i + ib * bd) / slot.ga_n, (slot.ga_i + ib * bd + slot.ga_w) / slot.ga_n);
LOG_TEE("shift: [%6d, %6d] + %6d -> [%6d, %6d]\n", slot.ga_i + ib * bd + slot.ga_w, slot.n_past_se + ib * bd, dd, slot.ga_i + ib * bd + slot.ga_w + dd, slot.n_past_se + ib * bd + dd);
llama_kv_cache_seq_shift(ctx, slot.id, slot.ga_i, slot.n_past_se, ib * bd);
llama_kv_cache_seq_add(ctx, slot.id, slot.ga_i, slot.n_past_se, ib * bd);
llama_kv_cache_seq_div(ctx, slot.id, slot.ga_i + ib * bd, slot.ga_i + ib * bd + slot.ga_w,slot.ga_n);
llama_kv_cache_seq_shift(ctx, slot.id, slot.ga_i + ib * bd + slot.ga_w,slot.n_past_se + ib * bd, dd);
llama_kv_cache_seq_add(ctx, slot.id, slot.ga_i + ib * bd + slot.ga_w,slot.n_past_se + ib * bd, dd);
slot.n_past_se -= bd;
@ -1836,7 +2003,7 @@ struct llama_server_context
send_embedding(slot);
slot.release();
slot.i_batch = -1;
return true;
continue;
}
completion_token_output result;
@ -1849,6 +2016,7 @@ struct llama_server_context
{
slot.t_start_genereration = ggml_time_us();
slot.t_prompt_processing = (slot.t_start_genereration - slot.t_start_process_prompt) / 1e3;
metrics.on_prompt_eval(slot);
}
llama_token_data_array cur_p = { slot.ctx_sampling->cur.data(), slot.ctx_sampling->cur.size(), false };
@ -1871,11 +2039,14 @@ struct llama_server_context
slot.release();
slot.print_timings();
send_final_response(slot);
metrics.on_prediction(slot);
}
slot.i_batch = -1;
}
}
LOG_VERBOSE("slots updated", {});
return true;
}
@ -1948,9 +2119,15 @@ static void server_print_usage(const char *argv0, const gpt_params &params,
printf(" -cb, --cont-batching enable continuous batching (a.k.a dynamic batching) (default: disabled)\n");
printf(" -spf FNAME, --system-prompt-file FNAME\n");
printf(" set a file to load a system prompt (initial prompt of all slots), this is useful for chat applications.\n");
printf(" -ctk TYPE, --cache-type-k TYPE\n");
printf(" KV cache data type for K (default: f16)\n");
printf(" -ctv TYPE, --cache-type-v TYPE\n");
printf(" KV cache data type for V (default: f16)\n");
printf(" --mmproj MMPROJ_FILE path to a multimodal projector file for LLaVA.\n");
printf(" --log-format log output format: json or text (default: json)\n");
printf(" --log-disable disables logging to a file.\n");
printf(" --slots-endpoint-disable disables slots monitoring endpoint.\n");
printf(" --metrics enable prometheus compatible metrics endpoint (default: %s).\n", sparams.metrics_endpoint ? "enabled" : "disabled");
printf("\n");
printf(" -n, --n-predict maximum tokens to predict (default: %d)\n", params.n_predict);
printf(" --override-kv KEY=TYPE:VALUE\n");
@ -2082,9 +2259,9 @@ static void server_params_parse(int argc, char **argv, server_params &sparams,
break;
}
std::string value(argv[i]);
/**/ if (value == "none") { params.rope_scaling_type = LLAMA_ROPE_SCALING_NONE; }
else if (value == "linear") { params.rope_scaling_type = LLAMA_ROPE_SCALING_LINEAR; }
else if (value == "yarn") { params.rope_scaling_type = LLAMA_ROPE_SCALING_YARN; }
/**/ if (value == "none") { params.rope_scaling_type = LLAMA_ROPE_SCALING_TYPE_NONE; }
else if (value == "linear") { params.rope_scaling_type = LLAMA_ROPE_SCALING_TYPE_LINEAR; }
else if (value == "yarn") { params.rope_scaling_type = LLAMA_ROPE_SCALING_TYPE_YARN; }
else { invalid_param = true; break; }
}
else if (arg == "--rope-freq-base")
@ -2208,15 +2385,15 @@ static void server_params_parse(int argc, char **argv, server_params &sparams,
std::string arg_next = argv[i];
if (arg_next == "none")
{
params.split_mode = LLAMA_SPLIT_NONE;
params.split_mode = LLAMA_SPLIT_MODE_NONE;
}
else if (arg_next == "layer")
{
params.split_mode = LLAMA_SPLIT_LAYER;
params.split_mode = LLAMA_SPLIT_MODE_LAYER;
}
else if (arg_next == "row")
{
params.split_mode = LLAMA_SPLIT_ROW;
params.split_mode = LLAMA_SPLIT_MODE_ROW;
}
else {
invalid_param = true;
@ -2386,6 +2563,12 @@ static void server_params_parse(int argc, char **argv, server_params &sparams,
);
llama.process_system_prompt_data(json::parse(systm_content));
}
else if (arg == "-ctk" || arg == "--cache-type-k") {
params.cache_type_k = argv[++i];
}
else if (arg == "-ctv" || arg == "--cache-type-v") {
params.cache_type_v = argv[++i];
}
else if(arg == "--mmproj")
{
if (++i >= argc)
@ -2395,6 +2578,27 @@ static void server_params_parse(int argc, char **argv, server_params &sparams,
}
params.mmproj = argv[i];
}
else if (arg == "--log-format")
{
if (++i >= argc)
{
invalid_param = true;
break;
}
if (std::strcmp(argv[i], "json") == 0)
{
server_log_json = true;
}
else if (std::strcmp(argv[i], "text") == 0)
{
server_log_json = false;
}
else
{
invalid_param = true;
break;
}
}
else if (arg == "--log-disable")
{
log_set_target(stdout);
@ -2404,6 +2608,10 @@ static void server_params_parse(int argc, char **argv, server_params &sparams,
{
sparams.slots_endpoint = false;
}
else if (arg == "--metrics")
{
sparams.metrics_endpoint = true;
}
else if (arg == "--chat-template")
{
if (++i >= argc)
@ -2437,15 +2645,15 @@ static void server_params_parse(int argc, char **argv, server_params &sparams,
sep++;
if (strncmp(sep, "int:", 4) == 0) {
sep += 4;
kvo.tag = LLAMA_KV_OVERRIDE_INT;
kvo.tag = LLAMA_KV_OVERRIDE_TYPE_INT;
kvo.int_value = std::atol(sep);
} else if (strncmp(sep, "float:", 6) == 0) {
sep += 6;
kvo.tag = LLAMA_KV_OVERRIDE_FLOAT;
kvo.tag = LLAMA_KV_OVERRIDE_TYPE_FLOAT;
kvo.float_value = std::atof(sep);
} else if (strncmp(sep, "bool:", 5) == 0) {
sep += 5;
kvo.tag = LLAMA_KV_OVERRIDE_BOOL;
kvo.tag = LLAMA_KV_OVERRIDE_TYPE_BOOL;
if (std::strcmp(sep, "true") == 0) {
kvo.bool_value = true;
} else if (std::strcmp(sep, "false") == 0) {
@ -2504,32 +2712,40 @@ static json format_partial_response(
static json format_tokenizer_response(const std::vector<llama_token> &tokens)
{
return json{
{"tokens", tokens}};
return json {
{"tokens", tokens}
};
}
static json format_detokenized_response(std::string content)
{
return json{
{"content", content}};
return json {
{"content", content}
};
}
static void log_server_request(const httplib::Request &req, const httplib::Response &res)
{
// skip GH copilot requests when using default port
if (req.path == "/v1/health" || req.path == "/v1/completions")
{
return;
}
LOG_INFO("request", {
{"remote_addr", req.remote_addr},
{"remote_port", req.remote_port},
{"status", res.status},
{"method", req.method},
{"path", req.path},
{"params", req.params},
});
{"remote_addr", req.remote_addr},
{"remote_port", req.remote_port},
{"status", res.status},
{"method", req.method},
{"path", req.path},
{"params", req.params},
});
LOG_VERBOSE("request", {
{"request", req.body},
{"response", res.body},
});
{"request", req.body},
{"response", res.body},
});
}
struct token_translator
@ -2556,7 +2772,16 @@ static void append_to_generated_text_from_generated_token_probs(llama_server_con
}
std::function<void(int)> shutdown_handler;
inline void signal_handler(int signal) { shutdown_handler(signal); }
std::atomic_flag is_terminating = ATOMIC_FLAG_INIT;
inline void signal_handler(int signal) {
if (is_terminating.test_and_set()) {
// in case it hangs, we can force terminate the server by hitting Ctrl+C twice
// this is for better developer experience, we can remove when the server is stable enough
fprintf(stderr, "Received second interrupt, terminating immediately.\n");
exit(1);
}
shutdown_handler(signal);
}
int main(int argc, char **argv)
{
@ -2611,7 +2836,7 @@ int main(int argc, char **argv)
// request slots data using task queue
task_server task;
task.id = llama.queue_tasks.get_new_id();
task.type = TASK_TYPE_SLOTS_DATA;
task.type = TASK_TYPE_METRICS;
task.target_id = -1;
llama.queue_results.add_waiting_task_id(task.id);
@ -2658,7 +2883,7 @@ int main(int argc, char **argv)
// request slots data using task queue
task_server task;
task.id = llama.queue_tasks.get_new_id();
task.type = TASK_TYPE_SLOTS_DATA;
task.type = TASK_TYPE_METRICS;
task.target_id = -1;
llama.queue_results.add_waiting_task_id(task.id);
@ -2673,6 +2898,87 @@ int main(int argc, char **argv)
});
}
if (sparams.metrics_endpoint) {
svr.Get("/metrics", [&](const httplib::Request&, httplib::Response& res) {
// request slots data using task queue
task_server task;
task.id = llama.queue_tasks.get_new_id();
task.type = TASK_TYPE_METRICS;
task.target_id = -1;
llama.queue_results.add_waiting_task_id(task.id);
llama.queue_tasks.post(task);
// get the result
task_result result = llama.queue_results.recv(task.id);
llama.queue_results.remove_waiting_task_id(task.id);
json data = result.result_json;
uint64_t n_prompt_tokens_processed = data["n_prompt_tokens_processed"];
uint64_t t_prompt_processing = data["t_prompt_processing"];
uint64_t n_tokens_predicted = data["n_tokens_predicted"];
uint64_t t_tokens_generation = data["t_tokens_generation"];
int32_t kv_cache_used_cells = data["kv_cache_used_cells"];
// metrics definition: https://prometheus.io/docs/practices/naming/#metric-names
json all_metrics_def = json {
{"counter", {{
{"name", "prompt_tokens_total"},
{"help", "Number of prompt tokens processed."},
{"value", data["n_prompt_tokens_processed_total"]}
}, {
{"name", "tokens_predicted_total"},
{"help", "Number of generation tokens processed."},
{"value", data["n_tokens_predicted_total"]}
}}},
{"gauge", {{
{"name", "prompt_tokens_seconds"},
{"help", "Average prompt throughput in tokens/s."},
{"value", n_prompt_tokens_processed ? 1e3 / t_prompt_processing * n_prompt_tokens_processed : 0}
},{
{"name", "predicted_tokens_seconds"},
{"help", "Average generation throughput in tokens/s."},
{"value", n_tokens_predicted ? 1e3 / t_tokens_generation * n_tokens_predicted : 0}
},{
{"name", "kv_cache_usage_ratio"},
{"help", "KV-cache usage. 1 means 100 percent usage."},
{"value", 1. * kv_cache_used_cells / params.n_ctx}
},{
{"name", "kv_cache_tokens"},
{"help", "KV-cache tokens."},
{"value", data["kv_cache_tokens_count"]}
},{
{"name", "requests_processing"},
{"help", "Number of request processing."},
{"value", data["processing"]}
},{
{"name", "requests_deferred"},
{"help", "Number of request deferred."},
{"value", data["deferred"]}
}}}
};
std::stringstream prometheus;
for (const auto& el : all_metrics_def.items()) {
const auto& type = el.key();
const auto& metrics_def = el.value();
for (const auto& metric_def : metrics_def) {
std::string name = metric_def["name"];
std::string help = metric_def["help"];
prometheus << "# HELP llamacpp:" << name << " " << help << "\n"
<< "# TYPE llamacpp:" << name << " " << type << "\n"
<< "llamacpp:" << name << " " << metric_def["value"] << "\n";
}
}
res.set_content(prometheus.str(), "text/plain; version=0.0.4");
res.status = 200; // HTTP OK
});
}
svr.set_logger(log_server_request);
svr.set_exception_handler([](const httplib::Request &, httplib::Response &res, std::exception_ptr ep)
@ -2725,9 +3031,6 @@ int main(int argc, char **argv)
// Set the base directory for serving static files
svr.set_base_dir(sparams.public_path);
// to make it ctrl+clickable:
LOG_TEE("\nllama server listening at http://%s:%d\n\n", sparams.hostname.c_str(), sparams.port);
std::unordered_map<std::string, std::string> log_data;
log_data["hostname"] = sparams.hostname;
log_data["port"] = std::to_string(sparams.port);
@ -2738,19 +3041,6 @@ int main(int argc, char **argv)
log_data["api_key"] = "api_key: " + std::to_string(sparams.api_keys.size()) + " keys loaded";
}
LOG_INFO("HTTP server listening", log_data);
// run the HTTP server in a thread - see comment below
std::thread t([&]()
{
if (!svr.listen_after_bind())
{
state.store(SERVER_STATE_ERROR);
return 1;
}
return 0;
});
// load the model
if (!llama.load_model(params))
{
@ -2930,87 +3220,88 @@ int main(int argc, char **argv)
res.set_content(models.dump(), "application/json; charset=utf-8");
});
const auto chat_completions = [&llama, &validate_api_key, &sparams](const httplib::Request &req, httplib::Response &res)
{
res.set_header("Access-Control-Allow-Origin", req.get_header_value("Origin"));
if (!validate_api_key(req, res)) {
return;
}
json data = oaicompat_completion_params_parse(llama.model, json::parse(req.body), sparams.chat_template);
// TODO: add mount point without "/v1" prefix -- how?
svr.Post("/v1/chat/completions", [&llama, &validate_api_key, &sparams](const httplib::Request &req, httplib::Response &res)
{
res.set_header("Access-Control-Allow-Origin", req.get_header_value("Origin"));
if (!validate_api_key(req, res)) {
return;
}
json data = oaicompat_completion_params_parse(llama.model, json::parse(req.body), sparams.chat_template);
const int task_id = llama.queue_tasks.get_new_id();
llama.queue_results.add_waiting_task_id(task_id);
llama.request_completion(task_id, data, false, false, -1);
const int task_id = llama.queue_tasks.get_new_id();
llama.queue_results.add_waiting_task_id(task_id);
llama.request_completion(task_id, data, false, false, -1);
if (!json_value(data, "stream", false)) {
std::string completion_text;
task_result result = llama.queue_results.recv(task_id);
if (!json_value(data, "stream", false)) {
std::string completion_text;
task_result result = llama.queue_results.recv(task_id);
if (!result.error && result.stop) {
json oaicompat_result = format_final_response_oaicompat(data, result);
if (!result.error && result.stop) {
json oaicompat_result = format_final_response_oaicompat(data, result);
res.set_content(oaicompat_result.dump(-1, ' ', false,
json::error_handler_t::replace),
"application/json; charset=utf-8");
} else {
res.status = 500;
res.set_content(result.result_json["content"], "text/plain; charset=utf-8");
}
llama.queue_results.remove_waiting_task_id(task_id);
} else {
const auto chunked_content_provider = [task_id, &llama](size_t, httplib::DataSink &sink) {
while (true) {
task_result llama_result = llama.queue_results.recv(task_id);
if (!llama_result.error) {
std::vector<json> result_array = format_partial_response_oaicompat( llama_result);
res.set_content(oaicompat_result.dump(-1, ' ', false,
json::error_handler_t::replace),
"application/json; charset=utf-8");
} else {
res.status = 500;
res.set_content(result.result_json["content"], "text/plain; charset=utf-8");
}
llama.queue_results.remove_waiting_task_id(task_id);
} else {
const auto chunked_content_provider = [task_id, &llama](size_t, httplib::DataSink &sink) {
while (true) {
task_result llama_result = llama.queue_results.recv(task_id);
if (!llama_result.error) {
std::vector<json> result_array = format_partial_response_oaicompat( llama_result);
for (auto it = result_array.begin(); it != result_array.end(); ++it)
{
if (!it->empty()) {
const std::string str =
"data: " +
it->dump(-1, ' ', false, json::error_handler_t::replace) +
"\n\n";
LOG_VERBOSE("data stream", {{"to_send", str}});
if (!sink.write(str.c_str(), str.size())) {
llama.queue_results.remove_waiting_task_id(task_id);
return false;
}
}
}
if (llama_result.stop) {
break;
}
} else {
for (auto it = result_array.begin(); it != result_array.end(); ++it)
{
if (!it->empty()) {
const std::string str =
"error: " +
llama_result.result_json.dump(-1, ' ', false,
json::error_handler_t::replace) +
"data: " +
it->dump(-1, ' ', false, json::error_handler_t::replace) +
"\n\n";
LOG_VERBOSE("data stream", {{"to_send", str}});
if (!sink.write(str.c_str(), str.size())) {
llama.queue_results.remove_waiting_task_id(task_id);
return false;
}
break;
}
}
sink.done();
llama.queue_results.remove_waiting_task_id(task_id);
return true;
};
auto on_complete = [task_id, &llama](bool) {
// cancel request
llama.request_cancel(task_id);
llama.queue_results.remove_waiting_task_id(task_id);
};
res.set_chunked_content_provider("text/event-stream", chunked_content_provider, on_complete);
if (llama_result.stop) {
break;
}
} else {
const std::string str =
"error: " +
llama_result.result_json.dump(-1, ' ', false,
json::error_handler_t::replace) +
"\n\n";
LOG_VERBOSE("data stream", {{"to_send", str}});
if (!sink.write(str.c_str(), str.size())) {
llama.queue_results.remove_waiting_task_id(task_id);
return false;
}
break;
}
}
});
sink.done();
llama.queue_results.remove_waiting_task_id(task_id);
return true;
};
auto on_complete = [task_id, &llama](bool) {
// cancel request
llama.request_cancel(task_id);
llama.queue_results.remove_waiting_task_id(task_id);
};
res.set_chunked_content_provider("text/event-stream", chunked_content_provider, on_complete);
}
};
svr.Post("/chat/completions", chat_completions);
svr.Post("/v1/chat/completions", chat_completions);
svr.Post("/infill", [&llama, &validate_api_key](const httplib::Request &req, httplib::Response &res)
{
@ -3218,6 +3509,19 @@ int main(int argc, char **argv)
}*/
//);
LOG_INFO("HTTP server listening", log_data);
// run the HTTP server in a thread - see comment below
std::thread t([&]()
{
if (!svr.listen_after_bind())
{
state.store(SERVER_STATE_ERROR);
return 1;
}
return 0;
});
llama.queue_tasks.on_new_task(std::bind(
&llama_server_context::process_single_task, &llama, std::placeholders::_1));
llama.queue_tasks.on_finish_multitask(std::bind(

47
examples/server/tests/README.md Normal file
View file

@ -0,0 +1,47 @@
# Server tests
Python based server tests scenario using [BDD](https://en.wikipedia.org/wiki/Behavior-driven_development) and [behave](https://behave.readthedocs.io/en/latest/):
* [issues.feature](./features/issues.feature) Pending issues scenario
* [parallel.feature](./features/parallel.feature) Scenario involving multi slots and concurrent requests
* [security.feature](./features/security.feature) Security, CORS and API Key
* [server.feature](./features/server.feature) Server base scenario: completion, embedding, tokenization, etc...
Tests target GitHub workflows job runners with 4 vCPU.
Requests are using [aiohttp](https://docs.aiohttp.org/en/stable/client_reference.html), [asyncio](https://docs.python.org/fr/3/library/asyncio.html) based http client.
Note: If the host architecture inference speed is faster than GitHub runners one, parallel scenario may randomly fail. To mitigate it, you can increase values in `n_predict`, `kv_size`.
### Install dependencies
`pip install -r requirements.txt`
### Run tests
1. Build the server
```shell
cd ../../..
mkdir build
cd build
cmake ../
cmake --build . --target server
```
2. download required models:
1. `../../../scripts/hf.sh --repo ggml-org/models --file tinyllamas/stories260K.gguf`
3. Start the test: `./tests.sh`
It's possible to override some scenario steps values with environment variables:
- `PORT` -> `context.server_port` to set the listening port of the server during scenario, default: `8080`
- `LLAMA_SERVER_BIN_PATH` -> to change the server binary path, default: `../../../build/bin/server`
- `DEBUG` -> "ON" to enable steps and server verbose mode `--verbose`
- `SERVER_LOG_FORMAT_JSON` -> if set switch server logs to json format
### Run @bug, @wip or @wrong_usage annotated scenario
Feature or Scenario must be annotated with `@llama.cpp` to be included in the default scope.
- `@bug` annotation aims to link a scenario with a GitHub issue.
- `@wrong_usage` are meant to show user issue that are actually an expected behavior
- `@wip` to focus on a scenario working in progress
To run a scenario annotated with `@bug`, start:
`DEBUG=ON ./tests.sh --no-skipped --tags bug`
After changing logic in `steps.py`, ensure that `@bug` and `@wrong_usage` scenario are updated.

69
examples/server/tests/features/environment.py Normal file
View file

@ -0,0 +1,69 @@
import os
import socket
import subprocess
import time
from contextlib import closing
from signal import SIGKILL
def before_scenario(context, scenario):
print(f"\x1b[33;42mStarting new scenario: {scenario.name}!\x1b[0m")
port = 8080
if 'PORT' in os.environ:
port = int(os.environ['PORT'])
if is_server_listening("localhost", port):
assert False, "Server already started"
def after_scenario(context, scenario):
if context.server_process is None:
return
if scenario.status == "failed":
if 'GITHUB_ACTIONS' in os.environ:
print(f"\x1b[33;101mSCENARIO FAILED: {scenario.name} server logs:\x1b[0m\n\n")
if os.path.isfile('llama.log'):
with closing(open('llama.log', 'r')) as f:
for line in f:
print(line)
if not is_server_listening(context.server_fqdn, context.server_port):
print("\x1b[33;101mERROR: Server stopped listening\x1b[0m")
if not pid_exists(context.server_process.pid):
assert False, f"Server not running pid={context.server_process.pid} ..."
print(f"stopping server pid={context.server_process.pid} ...")
context.server_process.kill()
# Wait few for socket to free up
time.sleep(0.05)
attempts = 0
while is_server_listening(context.server_fqdn, context.server_port):
print(f"stopping server pid={context.server_process.pid} ...")
os.kill(context.server_process.pid, SIGKILL)
time.sleep(0.1)
attempts += 1
if attempts > 5:
print(f"Server dangling exits, killing all {context.server_path} ...")
process = subprocess.run(['killall', '-9', context.server_path],
stderr=subprocess.PIPE,
universal_newlines=True)
print(process)
def is_server_listening(server_fqdn, server_port):
with closing(socket.socket(socket.AF_INET, socket.SOCK_STREAM)) as sock:
result = sock.connect_ex((server_fqdn, server_port))
return result == 0
def pid_exists(pid):
"""Check whether pid exists in the current process table."""
import errno
if pid < 0:
return False
try:
os.kill(pid, 0)
except OSError as e:
return e.errno == errno.EPERM
else:
return True

4
examples/server/tests/features/issues.feature Normal file
View file

@ -0,0 +1,4 @@
# List of ongoing issues
@bug
Feature: Issues
# No confirmed issue at the moment

145
examples/server/tests/features/parallel.feature Normal file
View file

@ -0,0 +1,145 @@
@llama.cpp
Feature: Parallel
Background: Server startup
Given a server listening on localhost:8080
And a model file stories260K.gguf
And a model alias tinyllama-2
And 42 as server seed
And 64 KV cache size
And 2 slots
And embeddings extraction
And continuous batching
Then the server is starting
Then the server is healthy
Scenario Outline: Multi users completion
Given a prompt:
"""
Write a very long story about AI.
"""
And a prompt:
"""
Write another very long music lyrics.
"""
And <n_predict> max tokens to predict
Given concurrent completion requests
Then the server is busy
Then the server is idle
And all slots are idle
Then all prompts are predicted with <n_predict> tokens
Examples:
| n_predict |
| 128 |
Scenario Outline: Multi users OAI completions compatibility
Given a system prompt You are a writer.
And a model tinyllama-2
Given a prompt:
"""
Write a very long book.
"""
And a prompt:
"""
Write another a poem.
"""
And <n_predict> max tokens to predict
And streaming is <streaming>
Given concurrent OAI completions requests
Then the server is busy
Then the server is idle
Then all prompts are predicted with <n_predict> tokens
Examples:
| streaming | n_predict |
| disabled | 128 |
| enabled | 64 |
Scenario Outline: Multi users OAI completions compatibility no v1
Given a system prompt You are a writer.
And a model tinyllama-2
Given a prompt:
"""
Write a very long book.
"""
And a prompt:
"""
Write another a poem.
"""
And <n_predict> max tokens to predict
And streaming is <streaming>
Given concurrent OAI completions requests no v1
Then the server is busy
Then the server is idle
Then all prompts are predicted with <n_predict> tokens
Examples:
| streaming | n_predict |
| disabled | 128 |
| enabled | 64 |
Scenario: Multi users with total number of tokens to predict exceeds the KV Cache size #3969
Given a prompt:
"""
Write a very long story about AI.
"""
And a prompt:
"""
Write another very long music lyrics.
"""
And a prompt:
"""
Write a very long poem.
"""
And a prompt:
"""
Write a very long joke.
"""
And 128 max tokens to predict
Given concurrent completion requests
Then the server is busy
Then the server is idle
Then all prompts are predicted
Scenario: Multi users embeddings
Given a prompt:
"""
Write a very long story about AI.
"""
And a prompt:
"""
Write another very long music lyrics.
"""
And a prompt:
"""
Write a very long poem.
"""
And a prompt:
"""
Write a very long joke.
"""
Given concurrent embedding requests
Then the server is busy
Then the server is idle
Then all embeddings are generated
Scenario: Multi users OAI compatibility embeddings
Given a prompt:
"""
In which country Paris is located ?
"""
And a prompt:
"""
Is Madrid the capital of Spain ?
"""
And a prompt:
"""
What is the biggest US city ?
"""
And a prompt:
"""
What is the capital of Bulgaria ?
"""
And a model tinyllama-2
Given concurrent OAI embedding requests
Then the server is busy
Then the server is idle
Then all embeddings are generated

50
examples/server/tests/features/security.feature Normal file
View file

@ -0,0 +1,50 @@
@llama.cpp
Feature: Security
Background: Server startup with an api key defined
Given a server listening on localhost:8080
And a model file stories260K.gguf
And a server api key llama.cpp
Then the server is starting
Then the server is healthy
Scenario Outline: Completion with some user api key
Given a prompt test
And a user api key <api_key>
And 4 max tokens to predict
And a completion request with <api_error> api error
Examples: Prompts
| api_key | api_error |
| llama.cpp | no |
| llama.cpp | no |
| hackeme | raised |
| | raised |
Scenario Outline: OAI Compatibility
Given a system prompt test
And a user prompt test
And a model test
And 2 max tokens to predict
And streaming is disabled
And a user api key <api_key>
Given an OAI compatible chat completions request with <api_error> api error
Examples: Prompts
| api_key | api_error |
| llama.cpp | no |
| llama.cpp | no |
| hackme | raised |
Scenario Outline: CORS Options
When an OPTIONS request is sent from <origin>
Then CORS header <cors_header> is set to <cors_header_value>
Examples: Headers
| origin | cors_header | cors_header_value |
| localhost | Access-Control-Allow-Origin | localhost |
| web.mydomain.fr | Access-Control-Allow-Origin | web.mydomain.fr |
| origin | Access-Control-Allow-Credentials | true |
| web.mydomain.fr | Access-Control-Allow-Methods | POST |
| web.mydomain.fr | Access-Control-Allow-Headers | * |

84
examples/server/tests/features/server.feature Normal file
View file

@ -0,0 +1,84 @@
@llama.cpp
Feature: llama.cpp server
Background: Server startup
Given a server listening on localhost:8080
And a model file stories260K.gguf
And a model alias tinyllama-2
And 42 as server seed
# KV Cache corresponds to the total amount of tokens
# that can be stored across all independent sequences: #4130
# see --ctx-size and #5568
And 32 KV cache size
And 1 slots
And embeddings extraction
And 32 server max tokens to predict
And prometheus compatible metrics exposed
Then the server is starting
Then the server is healthy
Scenario: Health
Then the server is ready
And all slots are idle
Scenario Outline: Completion
Given a prompt <prompt>
And <n_predict> max tokens to predict
And a completion request with no api error
Then <n_predicted> tokens are predicted matching <re_content>
And prometheus metrics are exposed
Examples: Prompts
| prompt | n_predict | re_content | n_predicted |
| I believe the meaning of life is | 8 | (read<or>going)+ | 8 |
| Write a joke about AI | 64 | (park<or>friends<or>scared<or>always)+ | 32 |
Scenario Outline: OAI Compatibility
Given a model <model>
And a system prompt <system_prompt>
And a user prompt <user_prompt>
And <max_tokens> max tokens to predict
And streaming is <enable_streaming>
Given an OAI compatible chat completions request with no api error
Then <n_predicted> tokens are predicted matching <re_content>
Examples: Prompts
| model | system_prompt | user_prompt | max_tokens | re_content | n_predicted | enable_streaming |
| llama-2 | Book | What is the best book | 8 | (Mom<or>what)+ | 8 | disabled |
| codellama70b | You are a coding assistant. | Write the fibonacci function in c++. | 64 | (thanks<or>happy<or>bird)+ | 32 | enabled |
Scenario: Embedding
When embeddings are computed for:
"""
What is the capital of Bulgaria ?
"""
Then embeddings are generated
Scenario: OAI Embeddings compatibility
Given a model tinyllama-2
When an OAI compatible embeddings computation request for:
"""
What is the capital of Spain ?
"""
Then embeddings are generated
Scenario: OAI Embeddings compatibility with multiple inputs
Given a model tinyllama-2
Given a prompt:
"""
In which country Paris is located ?
"""
And a prompt:
"""
Is Madrid the capital of Spain ?
"""
When an OAI compatible embeddings computation request for multiple inputs
Then embeddings are generated
Scenario: Tokenize / Detokenize
When tokenizing:
"""
What is the capital of France ?
"""
Then tokens can be detokenize

827
examples/server/tests/features/steps/steps.py Normal file
View file

@ -0,0 +1,827 @@
import asyncio
import collections
import json
import os
import re
import socket
import subprocess
import time
from contextlib import closing
from re import RegexFlag
import aiohttp
import openai
from behave import step
from behave.api.async_step import async_run_until_complete
from prometheus_client import parser
@step(u"a server listening on {server_fqdn}:{server_port}")
def step_server_config(context, server_fqdn, server_port):
context.server_fqdn = server_fqdn
context.server_port = int(server_port)
if 'PORT' in os.environ:
context.server_port = int(os.environ['PORT'])
print(f"$PORT set, overriding server port with to {context.server_port}")
context.base_url = f'http://{context.server_fqdn}:{context.server_port}'
context.debug = 'DEBUG' in os.environ and os.environ['DEBUG'] == 'ON'
context.model_alias = None
context.n_ctx = None
context.n_predict = None
context.n_server_predict = None
context.n_slots = None
context.server_api_key = None
context.server_continuous_batching = False
context.server_embeddings = False
context.server_metrics = False
context.server_process = None
context.server_seed = None
context.user_api_key = None
context.tasks_result = []
context.concurrent_tasks = []
context.prompts = []
@step(u'a model file {model_file}')
def step_model_file(context, model_file):
context.model_file = model_file
@step(u'a model alias {model_alias}')
def step_model_alias(context, model_alias):
context.model_alias = model_alias
@step(u'{seed} as server seed')
def step_seed(context, seed):
context.server_seed = int(seed)
@step(u'{n_ctx} KV cache size')
def step_n_ctx(context, n_ctx):
context.n_ctx = int(n_ctx)
@step(u'{n_slots} slots')
def step_n_slots(context, n_slots):
context.n_slots = int(n_slots)
@step(u'{n_predict} server max tokens to predict')
def step_server_n_predict(context, n_predict):
context.n_server_predict = int(n_predict)
@step(u'continuous batching')
def step_server_continuous_batching(context):
context.server_continuous_batching = True
@step(u'embeddings extraction')
def step_server_embeddings(context):
context.server_embeddings = True
@step(u'prometheus compatible metrics exposed')
def step_server_metrics(context):
context.server_metrics = True
@step(u"the server is starting")
def step_start_server(context):
start_server_background(context)
attempts = 0
while True:
with closing(socket.socket(socket.AF_INET, socket.SOCK_STREAM)) as sock:
result = sock.connect_ex((context.server_fqdn, context.server_port))
if result == 0:
print("\x1b[33;46mserver started!\x1b[0m")
return
attempts += 1
if attempts > 20:
assert False, "server not started"
print(f"waiting for server to start, connect error code = {result}...")
time.sleep(0.1)
@step(u"the server is {expecting_status}")
@async_run_until_complete
async def step_wait_for_the_server_to_be_started(context, expecting_status):
match expecting_status:
case 'healthy':
await wait_for_health_status(context, context.base_url, 200, 'ok')
case 'ready' | 'idle':
await wait_for_health_status(context, context.base_url, 200, 'ok',
params={'fail_on_no_slot': 0, 'include_slots': 0},
slots_idle=context.n_slots,
slots_processing=0,
expected_slots=[{'id': slot_id, 'state': 0}
for slot_id in range(context.n_slots)])
case 'busy':
await wait_for_health_status(context, context.base_url, 503,
'no slot available',
params={'fail_on_no_slot': 0, 'include_slots': 0},
slots_idle=0,
slots_processing=context.n_slots,
expected_slots=[{'id': slot_id, 'state': 1}
for slot_id in range(context.n_slots)])
case _:
assert False, "unknown status"
@step(u'all slots are {expected_slot_status_string}')
@async_run_until_complete
async def step_all_slots_status(context, expected_slot_status_string):
match expected_slot_status_string:
case 'idle':
expected_slot_status = 0
case 'busy':
expected_slot_status = 1
case _:
assert False, "unknown status"
expected_slots = [{'id': slot_id, 'state': expected_slot_status}
for slot_id in range(context.n_slots)]
await request_slots_status(context, expected_slots)
@step(u'a completion request with {api_error} api error')
@async_run_until_complete
async def step_request_completion(context, api_error):
expect_api_error = api_error == 'raised'
completion = await request_completion(context.prompts.pop(),
context.base_url,
debug=context.debug,
n_predict=context.n_predict,
server_seed=context.server_seed,
expect_api_error=expect_api_error,
user_api_key=context.user_api_key)
context.tasks_result.append(completion)
if context.debug:
print(f"Completion response: {completion}")
if expect_api_error:
assert completion == 401, f"completion must be an 401 status code: {completion}"
@step(u'{predicted_n} tokens are predicted matching {re_content}')
def step_n_tokens_predicted_with_content(context, predicted_n, re_content):
assert_n_tokens_predicted(context.tasks_result.pop(), int(predicted_n), re_content)
@step(u'{predicted_n} tokens are predicted')
def step_n_tokens_predicted(context, predicted_n):
assert_n_tokens_predicted(context.tasks_result.pop(), int(predicted_n))
@step(u'a user prompt {user_prompt}')
def step_user_prompt(context, user_prompt):
context.prompts.append(user_prompt)
@step(u'a system prompt {system_prompt}')
def step_system_prompt(context, system_prompt):
context.system_prompt = system_prompt
@step(u'a model {model}')
def step_model(context, model):
context.model = model
@step(u'{max_tokens} max tokens to predict')
def step_max_tokens(context, max_tokens):
context.n_predict = int(max_tokens)
@step(u'streaming is {enable_streaming}')
def step_streaming(context, enable_streaming):
context.enable_streaming = enable_streaming == 'enabled'
@step(u'a user api key {user_api_key}')
def step_user_api_key(context, user_api_key):
context.user_api_key = user_api_key
@step(u'no user api key')
def step_no_user_api_key(context):
context.user_api_key = None
@step(u'a user api key ')
def step_no_user_api_key_space(context):
context.user_api_key = None
@step(u'a server api key {server_api_key}')
def step_server_api_key(context, server_api_key):
context.server_api_key = server_api_key
@step(u'an OAI compatible chat completions request with {api_error} api error')
@async_run_until_complete
async def step_oai_chat_completions(context, api_error):
if context.debug:
print(f"Submitting OAI compatible completions request...")
expect_api_error = api_error == 'raised'
completion = await oai_chat_completions(context.prompts.pop(),
context.system_prompt,
context.base_url,
'/v1/chat',
False,
model=context.model if hasattr(context, 'model') else None,
n_predict=context.n_predict
if hasattr(context, 'n_predict') else None,
enable_streaming=context.enable_streaming
if hasattr(context, 'enable_streaming') else None,
server_seed=context.server_seed
if hasattr(context, 'server_seed') else None,
user_api_key=context.user_api_key
if hasattr(context, 'user_api_key') else None,
expect_api_error=expect_api_error)
context.tasks_result.append(completion)
if context.debug:
print(f"Completion response: {completion}")
if expect_api_error:
assert completion == 401, f"completion must be an 401 status code: {completion}"
if context.debug:
print(f"Completion response: {completion}")
@step(u'a prompt')
def step_a_prompt(context):
context.prompts.append(context.text)
@step(u'a prompt {prompt}')
def step_a_prompt_prompt(context, prompt):
context.prompts.append(prompt)
@step(u'concurrent completion requests')
@async_run_until_complete()
async def step_concurrent_completion_requests(context):
await concurrent_requests(context,
request_completion,
# prompt is inserted automatically
context.base_url,
debug=context.debug,
n_predict=context.n_predict if hasattr(context, 'n_predict') else None,
server_seed=context.server_seed if hasattr(context, 'server_seed') else None,
user_api_key=context.user_api_key if hasattr(context,
'user_api_key') else None)
@step(u'concurrent OAI completions requests')
@async_run_until_complete
async def step_oai_chat_completions(context):
await concurrent_requests(context, oai_chat_completions,
# user_prompt is inserted automatically
context.system_prompt,
context.base_url,
'/v1/chat/completions',
True, # async_client
model=context.model
if hasattr(context, 'model') else None,
n_predict=context.n_predict
if hasattr(context, 'n_predict') else None,
enable_streaming=context.enable_streaming
if hasattr(context, 'enable_streaming') else None,
server_seed=context.server_seed
if hasattr(context, 'server_seed') else None,
user_api_key=context.user_api_key
if hasattr(context, 'user_api_key') else None)
@step(u'concurrent OAI completions requests no v1')
@async_run_until_complete
async def step_oai_chat_completions(context):
await concurrent_requests(context, oai_chat_completions,
# user_prompt is inserted automatically
context.system_prompt,
context.base_url,
'/chat/completions',
True, # async_client
model=context.model
if hasattr(context, 'model') else None,
n_predict=context.n_predict
if hasattr(context, 'n_predict') else None,
enable_streaming=context.enable_streaming
if hasattr(context, 'enable_streaming') else None,
server_seed=context.server_seed
if hasattr(context, 'server_seed') else None,
user_api_key=context.user_api_key
if hasattr(context, 'user_api_key') else None)
@step(u'all prompts are predicted')
@async_run_until_complete
async def step_all_prompts_are_predicted(context):
await all_prompts_are_predicted(context)
@step(u'all prompts are predicted with {n_predict} tokens')
@async_run_until_complete
async def step_all_prompts_are_predicted_with_n_tokens(context, n_predict):
expected_predicted_n = int(n_predict)
await all_prompts_are_predicted(context, expected_predicted_n)
async def all_prompts_are_predicted(context, expected_predicted_n=None):
n_completions = await gather_tasks_results(context)
assert n_completions > 0
for i in range(n_completions):
assert_n_tokens_predicted(context.tasks_result.pop(), expected_predicted_n=expected_predicted_n)
assert len(context.concurrent_tasks) == 0, f"{len(context.concurrent_tasks)} pending requests"
@step(u'embeddings are computed for')
@async_run_until_complete
async def step_compute_embedding(context):
context.embeddings = await request_embedding(context.text, base_url=context.base_url)
@step(u'embeddings are generated')
def step_assert_embeddings(context):
if len(context.prompts) == 0:
assert_embeddings(context.embeddings)
else:
assert len(context.embeddings) == len(context.prompts), (f"unexpected response:\n"
f"context.prompts={context.prompts}\n"
f"context.embeddings={context.embeddings}")
for embedding in context.embeddings:
context.prompts.pop()
assert_embeddings(embedding)
@step(u'an OAI compatible embeddings computation request for')
@async_run_until_complete
async def step_oai_compute_embeddings(context):
context.embeddings = await request_oai_embeddings(context.text,
base_url=context.base_url,
user_api_key=context.user_api_key,
model=context.model)
@step(u'an OAI compatible embeddings computation request for multiple inputs')
@async_run_until_complete
async def step_oai_compute_embeddings_multiple_inputs(context):
context.embeddings = await request_oai_embeddings(context.prompts,
base_url=context.base_url,
user_api_key=context.user_api_key,
model=context.model)
@step(u'concurrent embedding requests')
@async_run_until_complete()
async def step_concurrent_embedding_requests(context):
await concurrent_requests(context,
request_embedding,
# prompt is inserted automatically
base_url=context.base_url)
@step(u'concurrent OAI embedding requests')
@async_run_until_complete()
async def step_concurrent_oai_embedding_requests(context):
await concurrent_requests(context,
request_oai_embeddings,
# prompt is inserted automatically
base_url=context.base_url,
async_client=True,
model=context.model)
@step(u'all embeddings are generated')
@async_run_until_complete()
async def all_embeddings_are_generated(context):
n_embedding_requests = await gather_tasks_results(context)
assert n_embedding_requests > 0
for i in range(n_embedding_requests):
assert_embeddings(context.tasks_result.pop())
@step(u'tokenizing')
@async_run_until_complete
async def step_tokenize(context):
context.tokenized_text = context.text
async with aiohttp.ClientSession() as session:
async with session.post(f'{context.base_url}/tokenize',
json={
"content": context.tokenized_text,
}) as response:
assert response.status == 200
tokenize_json = await response.json()
context.tokens = tokenize_json['tokens']
@step(u'tokens can be detokenize')
@async_run_until_complete
async def step_detokenize(context):
assert len(context.tokens) > 0
async with aiohttp.ClientSession() as session:
async with session.post(f'{context.base_url}/detokenize',
json={
"tokens": context.tokens,
}) as response:
assert response.status == 200
detokenize_json = await response.json()
# SPM tokenizer adds a whitespace prefix: https://github.com/google/sentencepiece/issues/15
assert context.tokenized_text == detokenize_json['content'].strip()
@step(u'an OPTIONS request is sent from {origin}')
@async_run_until_complete
async def step_options_request(context, origin):
async with aiohttp.ClientSession() as session:
async with session.options(f'{context.base_url}/v1/chat/completions',
headers={"Origin": origin}) as response:
assert response.status == 200
context.options_response = response
@step(u'CORS header {cors_header} is set to {cors_header_value}')
def step_check_options_header_value(context, cors_header, cors_header_value):
assert context.options_response.headers[cors_header] == cors_header_value
@step(u'prometheus metrics are exposed')
@async_run_until_complete
async def step_prometheus_metrics_exported(context):
async with aiohttp.ClientSession() as session:
async with await session.get(f'{context.base_url}/metrics') as metrics_response:
assert metrics_response.status == 200
assert metrics_response.headers['Content-Type'] == "text/plain; version=0.0.4"
metrics_raw = await metrics_response.text()
metric_exported = False
for metric in parser.text_string_to_metric_families(metrics_raw):
match metric.name:
case "llamacpp:kv_cache_usage_ratio":
assert len(metric.samples) > 0
metric_exported = True
assert metric_exported, "No metrics exported"
async def concurrent_requests(context, f_completion, *args, **kwargs):
n_prompts = len(context.prompts)
if context.debug:
print(f"starting {n_prompts} concurrent completion requests...")
assert n_prompts > 0
for prompt_no in range(n_prompts):
shifted_args = [context.prompts.pop(), *args]
context.concurrent_tasks.append(asyncio.create_task(f_completion(*shifted_args, **kwargs)))
await asyncio.sleep(0.1)
async def request_completion(prompt,
base_url,
debug=False,
n_predict=None,
server_seed=None,
expect_api_error=None,
user_api_key=None):
if debug:
print(f"Sending completion request: {prompt}")
origin = "my.super.domain"
headers = {
'Origin': origin
}
if user_api_key is not None:
if debug:
print(f"Set user_api_key: {user_api_key}")
headers['Authorization'] = f'Bearer {user_api_key}'
async with aiohttp.ClientSession() as session:
async with session.post(f'{base_url}/completion',
json={
"prompt": prompt,
"n_predict": int(n_predict) if n_predict is not None else -1,
"seed": server_seed if server_seed is not None else 42
},
headers=headers) as response:
if expect_api_error is None or not expect_api_error:
assert response.status == 200
assert response.headers['Access-Control-Allow-Origin'] == origin
return await response.json()
else:
return response.status
async def oai_chat_completions(user_prompt,
system_prompt,
base_url,
base_path,
async_client,
debug=False,
model=None,
n_predict=None,
enable_streaming=None,
server_seed=None,
user_api_key=None,
expect_api_error=None):
if debug:
print(f"Sending OAI Chat completions request: {user_prompt}")
# openai client always expects an api key
user_api_key = user_api_key if user_api_key is not None else 'nope'
seed = server_seed if server_seed is not None else 42
enable_streaming = enable_streaming if enable_streaming is not None else False
payload = {
"messages": [
{
"role": "system",
"content": system_prompt,
},
{
"role": "user",
"content": user_prompt,
}
],
"model": model,
"max_tokens": n_predict,
"stream": enable_streaming,
"seed": seed
}
completion_response = {
'content': '',
'timings': {
'predicted_n': 0
}
}
if async_client:
origin = 'llama.cpp'
headers = {'Authorization': f'Bearer {user_api_key}', 'Origin': origin}
async with aiohttp.ClientSession() as session:
async with session.post(f'{base_url}{base_path}',
json=payload,
headers=headers) as response:
if enable_streaming:
assert response.status == 200
assert response.headers['Access-Control-Allow-Origin'] == origin
assert response.headers['Content-Type'] == "text/event-stream"
event_received = True
while event_received:
event_received = False
async for line_in_bytes in response.content:
line = line_in_bytes.decode('utf8')
line = line.rstrip('\n').rstrip('\r')
if line == '':
continue
event_data = line.split(': ', 1)
assert event_data[0] == 'data', f'Bad event code received: ```{event_data}```'
chunk_raw = event_data[1]
chunk = json.loads(chunk_raw)
assert len(chunk['choices']) == 1, f"no choices provided, line ```{line}```"
delta = chunk['choices'][0]['delta']
if 'content' in delta:
completion_response['content'] += delta['content']
completion_response['timings']['predicted_n'] += 1
else:
if expect_api_error is None or not expect_api_error:
assert response.status == 200
assert response.headers['Access-Control-Allow-Origin'] == origin
assert response.headers['Content-Type'] == "application/json; charset=utf-8"
chat_completion_raw = await response.json()
completion_response = {
'content': chat_completion_raw['choices'][0]['message'],
'timings': {
'predicted_n': chat_completion_raw['usage']['completion_tokens']
}
}
else:
return response.status
else:
try:
openai.api_key = user_api_key
openai.api_base = f'{base_url}{base_path}'
chat_completion = openai.Completion.create(
messages=payload['messages'],
model=model,
max_tokens=n_predict,
stream=enable_streaming,
seed=seed
)
except openai.error.APIError as e:
if expect_api_error is not None and expect_api_error:
return 401
else:
assert False, f'error raised: {e}'
if enable_streaming:
for chunk in chat_completion:
assert len(chunk.choices) == 1
delta = chunk.choices[0].delta
if 'content' in delta:
completion_response['content'] += delta['content']
completion_response['timings']['predicted_n'] += 1
else:
assert len(chat_completion.choices) == 1
completion_response = {
'content': chat_completion.choices[0].message.content,
'timings': {
'predicted_n': chat_completion.usage.completion_tokens
}
}
if debug:
print("OAI response formatted to llama.cpp:", completion_response)
return completion_response
async def request_embedding(content, base_url=None):
async with aiohttp.ClientSession() as session:
async with session.post(f'{base_url}/embedding',
json={
"content": content,
}) as response:
assert response.status == 200
response_json = await response.json()
return response_json['embedding']
async def request_oai_embeddings(input,
base_url=None, user_api_key=None,
model=None, async_client=False):
# openai client always expects an api_key
user_api_key = user_api_key if user_api_key is not None else 'nope'
if async_client:
origin = 'llama.cpp'
if user_api_key is not None:
headers = {'Authorization': f'Bearer {user_api_key}', 'Origin': origin}
async with aiohttp.ClientSession() as session:
async with session.post(f'{base_url}/v1/embeddings',
json={
"input": input,
"model": model,
},
headers=headers) as response:
assert response.status == 200, f"received status code not expected: {response.status}"
assert response.headers['Access-Control-Allow-Origin'] == origin
assert response.headers['Content-Type'] == "application/json; charset=utf-8"
response_json = await response.json()
assert response_json['model'] == model, f"invalid model received: {response_json['model']}"
assert response_json['object'] == 'list'
return response_json['data']
else:
openai.api_key = user_api_key
openai.api_base = f'{base_url}/v1'
oai_embeddings = openai.Embedding.create(
model=model,
input=input,
)
if isinstance(input, collections.abc.Sequence):
embeddings = []
for an_oai_embeddings in oai_embeddings.data:
embeddings.append(an_oai_embeddings.embedding)
else:
embeddings = oai_embeddings.data.embedding
return embeddings
def assert_n_tokens_predicted(completion_response, expected_predicted_n=None, re_content=None):
content = completion_response['content']
n_predicted = completion_response['timings']['predicted_n']
assert len(content) > 0, "no token predicted"
if expected_predicted_n is not None:
assert n_predicted == expected_predicted_n, (f'invalid number of tokens predicted:'
f' {n_predicted} <> {expected_predicted_n}')
if re_content is not None:
re_content = '^.*' + re_content.replace('<or>', '|') + '.*$'
assert re.match(re_content, content, flags=RegexFlag.IGNORECASE | RegexFlag.MULTILINE | RegexFlag.DOTALL), (
f'invalid tokens predicted:'
f' ```\n{content}\n``` do not match /{re_content}/')
async def gather_tasks_results(context):
n_tasks = len(context.concurrent_tasks)
if context.debug:
print(f"Waiting for all {n_tasks} tasks results...")
for task_no in range(n_tasks):
context.tasks_result.append(await context.concurrent_tasks.pop())
n_completions = len(context.tasks_result)
return n_completions
async def wait_for_health_status(context,
base_url,
expected_http_status_code,
expected_health_status,
params=None,
slots_idle=None,
slots_processing=None,
expected_slots=None):
if context.debug:
print(f"Starting checking for health for expected_health_status={expected_health_status}")
timeout = 3 # seconds
if expected_health_status == 'ok':
timeout = 10 # CI slow inference
interval = 0.5
counter = 0
async with aiohttp.ClientSession() as session:
while True:
async with await session.get(f'{base_url}/health', params=params) as health_response:
status_code = health_response.status
health = await health_response.json()
if context.debug:
print(f"HEALTH - response for expected health status='{expected_health_status}' on "
f"'{base_url}/health'?{params} is {health}")
if (status_code == expected_http_status_code
and health['status'] == expected_health_status
and (slots_idle is None or health['slots_idle'] == slots_idle)
and (slots_processing is None or health['slots_processing'] == slots_processing)):
if expected_slots is not None:
assert_slots_status(health['slots'], expected_slots)
return
if (status_code == expected_http_status_code
and health['status'] == expected_health_status
and (slots_idle is None or health['slots_idle'] == slots_idle)
and (slots_processing is None or health['slots_processing'] == slots_processing)):
if expected_slots is not None:
assert_slots_status(health['slots'], expected_slots)
return
await asyncio.sleep(interval)
counter += interval
if counter >= timeout:
# Sometimes health requests are triggered after completions are predicted
if expected_http_status_code == 503:
if len(context.tasks_result) == 0:
print("\x1b[5;37;43mWARNING: forcing concurrent tasks,"
" busy health check missed, probably too fast inference\x1b[0m")
n_completions = await gather_tasks_results(context)
if n_completions > 0:
return
assert False, f'{expected_health_status} timeout exceeded {counter}s>={timeout}'
def assert_embeddings(embeddings):
assert len(embeddings) > 0
embeddings_computed = False
for emb in embeddings:
if emb != 0:
embeddings_computed = True
assert embeddings_computed, f"Embeddings: {embeddings}"
async def request_slots_status(context, expected_slots):
async with aiohttp.ClientSession() as session:
async with await session.get(f'{context.base_url}/slots') as slots_response:
assert slots_response.status == 200
slots = await slots_response.json()
assert_slots_status(slots, expected_slots)
def assert_slots_status(slots, expected_slots):
assert len(slots) == len(expected_slots)
for slot_id, (expected, slot) in enumerate(zip(expected_slots, slots)):
for key in expected:
assert expected[key] == slot[key], (f"invalid slot {slot_id}"
f" expected[{key}] != slot[{key}]"
f" = {expected[key]} != {slot[key]}")
def start_server_background(context):
context.server_path = '../../../build/bin/server'
if 'LLAMA_SERVER_BIN_PATH' in os.environ:
context.server_path = os.environ['LLAMA_SERVER_BIN_PATH']
server_args = [
'--host', context.server_fqdn,
'--port', context.server_port,
'--model', context.model_file
]
if context.server_continuous_batching:
server_args.append('--cont-batching')
if context.server_embeddings:
server_args.append('--embedding')
if context.server_metrics:
server_args.append('--metrics')
if context.model_alias is not None:
server_args.extend(['--alias', context.model_alias])
if context.n_ctx is not None:
server_args.extend(['--ctx-size', context.n_ctx])
if context.n_slots is not None:
server_args.extend(['--parallel', context.n_slots])
if context.n_server_predict is not None:
server_args.extend(['--n-predict', context.n_server_predict])
if context.server_api_key is not None:
server_args.extend(['--api-key', context.server_api_key])
if context.debug:
server_args.append('--verbose')
if 'SERVER_LOG_FORMAT_JSON' not in os.environ:
server_args.extend(['--log-format', "text"])
print(f"starting server with: {context.server_path}", *server_args)
context.server_process = subprocess.Popen(
[str(arg) for arg in [context.server_path, *server_args]],
close_fds=True)
print(f"server pid={context.server_process.pid}")

21
examples/server/tests/features/wrong_usages.feature Normal file
View file

@ -0,0 +1,21 @@
# run with ./test.sh --tags wrong_usage
@wrong_usage
Feature: Wrong usage of llama.cpp server
#3969 The user must always set --n-predict option
# to cap the number of tokens any completion request can generate
# or pass n_predict/max_tokens in the request.
Scenario: Infinite loop
Given a server listening on localhost:8080
And a model file stories260K.gguf
# Uncomment below to fix the issue
#And 64 server max tokens to predict
Then the server is starting
Given a prompt:
"""
Go to: infinite loop
"""
# Uncomment below to fix the issue
#And 128 max tokens to predict
Given concurrent completion requests
Then all prompts are predicted

4
examples/server/tests/requirements.txt Normal file
View file

@ -0,0 +1,4 @@
aiohttp~=3.9.3
behave~=1.2.6
openai~=0.25.0
prometheus-client~=0.20.0

12
examples/server/tests/tests.sh Executable file
View file

@ -0,0 +1,12 @@
#!/bin/bash
set -eu
if [ $# -lt 1 ]
then
# Start @llama.cpp scenario
behave --summary --stop --no-capture --exclude 'issues|wrong_usages' --tags llama.cpp
else
behave "$@"
fi

84
examples/server/utils.hpp
View file

@ -14,6 +14,7 @@
using json = nlohmann::json;
extern bool server_verbose;
extern bool server_log_json;
#ifndef SERVER_VERBOSE
#define SERVER_VERBOSE 1
@ -27,14 +28,14 @@ extern bool server_verbose;
{ \
if (server_verbose) \
{ \
server_log("VERBOSE", __func__, __LINE__, MSG, __VA_ARGS__); \
server_log("VERB", __func__, __LINE__, MSG, __VA_ARGS__); \
} \
} while (0)
#endif
#define LOG_ERROR( MSG, ...) server_log("ERROR", __func__, __LINE__, MSG, __VA_ARGS__)
#define LOG_WARNING(MSG, ...) server_log("WARNING", __func__, __LINE__, MSG, __VA_ARGS__)
#define LOG_INFO( MSG, ...) server_log("INFO", __func__, __LINE__, MSG, __VA_ARGS__)
#define LOG_ERROR( MSG, ...) server_log("ERR", __func__, __LINE__, MSG, __VA_ARGS__)
#define LOG_WARNING(MSG, ...) server_log("WARN", __func__, __LINE__, MSG, __VA_ARGS__)
#define LOG_INFO( MSG, ...) server_log("INFO", __func__, __LINE__, MSG, __VA_ARGS__)
//
// parallel
@ -50,7 +51,7 @@ enum task_type {
TASK_TYPE_COMPLETION,
TASK_TYPE_CANCEL,
TASK_TYPE_NEXT_RESPONSE,
TASK_TYPE_SLOTS_DATA
TASK_TYPE_METRICS
};
struct task_server {
@ -133,26 +134,48 @@ struct completion_token_output
std::string text_to_send;
};
static inline void server_log(const char *level, const char *function, int line,
const char *message, const nlohmann::ordered_json &extra)
static inline void server_log(const char *level, const char *function, int line, const char *message, const nlohmann::ordered_json &extra)
{
nlohmann::ordered_json log
{
std::stringstream ss_tid;
ss_tid << std::this_thread::get_id();
json log = nlohmann::ordered_json{
{"tid", ss_tid.str()},
{"timestamp", time(nullptr)},
{"level", level},
{"function", function},
{"line", line},
{"message", message},
};
if (!extra.empty())
{
log.merge_patch(extra);
}
if (server_log_json) {
log.merge_patch(
{
{"level", level},
{"function", function},
{"line", line},
{"msg", message},
});
if (!extra.empty()) {
log.merge_patch(extra);
}
const std::string str = log.dump(-1, ' ', false, json::error_handler_t::replace);
printf("%.*s\n", (int)str.size(), str.data());
fflush(stdout);
std::cout << log.dump(-1, ' ', false, json::error_handler_t::replace) << "\n" << std::flush;
} else {
char buf[1024];
snprintf(buf, 1024, "%4s [%24s] %s", level, function, message);
if (!extra.empty()) {
log.merge_patch(extra);
}
std::stringstream ss;
ss << buf << " |";
for (const auto& el : log.items())
{
const std::string value = el.value().dump(-1, ' ', false, json::error_handler_t::replace);
snprintf(buf, 1024, " %s=%s", el.key().c_str(), value.c_str());
ss << buf;
}
const std::string str = ss.str();
printf("%.*s\n", (int)str.size(), str.data());
fflush(stdout);
}
}
//
@ -234,6 +257,7 @@ struct llama_server_queue {
std::unique_lock<std::mutex> lock(mutex_tasks);
if (task.id == -1) {
task.id = id++;
LOG_VERBOSE("new task id", {{"new_id", task.id}});
}
queue_tasks.push_back(std::move(task));
condition_tasks.notify_one();
@ -249,7 +273,9 @@ struct llama_server_queue {
// Get the next id for creating anew task
int get_new_id() {
std::unique_lock<std::mutex> lock(mutex_tasks);
return id++;
int new_id = id++;
LOG_VERBOSE("new task id", {{"new_id", new_id}});
return new_id;
}
// Register function to process a new task
@ -290,8 +316,7 @@ struct llama_server_queue {
void start_loop() {
running = true;
while (true) {
// new task arrived
LOG_VERBOSE("have new task", {});
LOG_VERBOSE("new task may arrive", {});
{
while (true)
{
@ -303,7 +328,7 @@ struct llama_server_queue {
task_server task = queue_tasks.front();
queue_tasks.erase(queue_tasks.begin());
lock.unlock();
LOG_VERBOSE("callback_new_task", {});
LOG_VERBOSE("callback_new_task", {{"task_id", task.id}});
callback_new_task(task);
}
LOG_VERBOSE("callback_all_task_finished", {});
@ -384,11 +409,13 @@ struct llama_server_response {
std::condition_variable condition_results;
void add_waiting_task_id(int task_id) {
LOG_VERBOSE("waiting for task id", {{"task_id", task_id}});
std::unique_lock<std::mutex> lock(mutex_results);
waiting_task_ids.insert(task_id);
}
void remove_waiting_task_id(int task_id) {
LOG_VERBOSE("remove waiting for task id", {{"task_id", task_id}});
std::unique_lock<std::mutex> lock(mutex_results);
waiting_task_ids.erase(task_id);
}
@ -401,7 +428,6 @@ struct llama_server_response {
condition_results.wait(lock, [&]{
return !queue_results.empty();
});
LOG_VERBOSE("condition_results unblock", {});
for (int i = 0; i < (int) queue_results.size(); i++)
{
@ -426,22 +452,22 @@ struct llama_server_response {
// Send a new result to a waiting task_id
void send(task_result result) {
std::unique_lock<std::mutex> lock(mutex_results);
LOG_VERBOSE("send new result", {});
LOG_VERBOSE("send new result", {{"task_id", result.id}});
for (auto& task_id : waiting_task_ids) {
// LOG_TEE("waiting task id %i \n", task_id);
// for now, tasks that have associated parent multitasks just get erased once multitask picks up the result
if (result.multitask_id == task_id)
{
LOG_VERBOSE("callback_update_multitask", {});
LOG_VERBOSE("callback_update_multitask", {{"task_id", task_id}});
callback_update_multitask(task_id, result.id, result);
continue;
}
if (result.id == task_id)
{
LOG_VERBOSE("queue_results.push_back", {});
LOG_VERBOSE("queue_results.push_back", {{"task_id", task_id}});
queue_results.push_back(result);
condition_results.notify_one();
condition_results.notify_all();
return;
}
}

2
examples/train-text-from-scratch/train-text-from-scratch.cpp
View file

@ -960,7 +960,7 @@ int main(int argc, char ** argv) {
struct ggml_opt_context * opt = train->opt;
// set opt params from command line
opt->params = ggml_opt_default_params(GGML_OPT_ADAM);
opt->params = ggml_opt_default_params(GGML_OPT_TYPE_ADAM);
opt->params.print_forward_graph = false;
opt->params.print_backward_graph = false;
opt->params.graph_size = LLAMA_TRAIN_MAX_NODES;

6
flake.lock generated
View file

@ -20,11 +20,11 @@
},
"nixpkgs": {
"locked": {
"lastModified": 1708118438,
"narHash": "sha256-kk9/0nuVgA220FcqH/D2xaN6uGyHp/zoxPNUmPCMmEE=",
"lastModified": 1708655239,
"narHash": "sha256-ZrP/yACUvDB+zbqYJsln4iwotbH6CTZiTkANJ0AgDv4=",
"owner": "NixOS",
"repo": "nixpkgs",
"rev": "5863c27340ba4de8f83e7e3c023b9599c3cb3c80",
"rev": "cbc4211f0afffe6dfd2478a62615dd5175a13f9a",
"type": "github"
},
"original": {

2
ggml-backend-impl.h
View file

@ -104,6 +104,8 @@ extern "C" {
};
struct ggml_backend {
ggml_guid_t guid;
struct ggml_backend_i iface;
ggml_backend_context_t context;

16
ggml-backend.c
View file

@ -12,7 +12,6 @@
#define MAX(a, b) ((a) > (b) ? (a) : (b))
// backend buffer type
const char * ggml_backend_buft_name(ggml_backend_buffer_type_t buft) {
@ -159,6 +158,13 @@ bool ggml_backend_buffer_copy_tensor(const struct ggml_tensor * src, struct ggml
// backend
ggml_guid_t ggml_backend_guid(ggml_backend_t backend) {
if (backend == NULL) {
return NULL;
}
return backend->guid;
}
const char * ggml_backend_name(ggml_backend_t backend) {
if (backend == NULL) {
return "NULL";
@ -781,6 +787,11 @@ static struct ggml_backend_i cpu_backend_i = {
/* .supports_op = */ ggml_backend_cpu_supports_op,
};
static ggml_guid_t ggml_backend_cpu_guid(void) {
static ggml_guid guid = { 0xaa, 0x67, 0xc7, 0x43, 0x96, 0xe6, 0xa3, 0x8a, 0xe3, 0xaf, 0xea, 0x92, 0x36, 0xbc, 0xfc, 0x89 };
return &guid;
}
ggml_backend_t ggml_backend_cpu_init(void) {
struct ggml_backend_cpu_context * ctx = malloc(sizeof(struct ggml_backend_cpu_context));
if (ctx == NULL) {
@ -800,6 +811,7 @@ ggml_backend_t ggml_backend_cpu_init(void) {
}
*cpu_backend = (struct ggml_backend) {
/* .guid = */ ggml_backend_cpu_guid(),
/* .interface = */ cpu_backend_i,
/* .context = */ ctx
};
@ -807,7 +819,7 @@ ggml_backend_t ggml_backend_cpu_init(void) {
}
GGML_CALL bool ggml_backend_is_cpu(ggml_backend_t backend) {
return backend && backend->iface.get_name == ggml_backend_cpu_name;
return backend != NULL && ggml_guid_matches(backend->guid, ggml_backend_cpu_guid());
}
void ggml_backend_cpu_set_n_threads(ggml_backend_t backend_cpu, int n_threads) {

2
ggml-backend.h
View file

@ -49,7 +49,7 @@ extern "C" {
// Backend
//
GGML_API ggml_guid_t ggml_backend_guid(ggml_backend_t backend);
GGML_API const char * ggml_backend_name(ggml_backend_t backend);
GGML_API void ggml_backend_free(ggml_backend_t backend);

884
ggml-cuda.cu
View file

File diff suppressed because it is too large Load diff

8
ggml-kompute.cpp
View file

@ -1953,11 +1953,17 @@ static struct ggml_backend_i kompute_backend_i = {
/* .supports_op = */ ggml_backend_kompute_supports_op,
};
static ggml_guid_t ggml_backend_kompute_guid() {
static ggml_guid guid = { 0x7b, 0x57, 0xdc, 0xaf, 0xde, 0x12, 0x1d, 0x49, 0xfb, 0x35, 0xfa, 0x9b, 0x18, 0x31, 0x1d, 0xca };
return &guid;
}
ggml_backend_t ggml_backend_kompute_init(int device) {
GGML_ASSERT(s_kompute_context == nullptr);
s_kompute_context = new ggml_kompute_context(device);
ggml_backend_t kompute_backend = new ggml_backend {
/* .guid = */ ggml_backend_kompute_guid(),
/* .interface = */ kompute_backend_i,
/* .context = */ s_kompute_context,
};
@ -1966,7 +1972,7 @@ ggml_backend_t ggml_backend_kompute_init(int device) {
}
bool ggml_backend_is_kompute(ggml_backend_t backend) {
return backend && backend->iface.get_name == ggml_backend_kompute_name;
return backend != NULL && ggml_guid_matches(backend->guid, ggml_backend_kompute_guid());
}
static ggml_backend_t ggml_backend_reg_kompute_init(const char * params, void * user_data) {

111
ggml-metal.m
View file

@ -61,8 +61,11 @@ enum ggml_metal_kernel_type {
GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ2_XXS,
GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ2_XS,
GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ3_XXS,
GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ3_S,
GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ2_S,
GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ1_S,
GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ4_NL,
GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ4_XS,
GGML_METAL_KERNEL_TYPE_GET_ROWS_I32,
GGML_METAL_KERNEL_TYPE_RMS_NORM,
GGML_METAL_KERNEL_TYPE_GROUP_NORM,
@ -85,8 +88,11 @@ enum ggml_metal_kernel_type {
GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_XXS_F32,
GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_XS_F32,
GGML_METAL_KERNEL_TYPE_MUL_MV_IQ3_XXS_F32,
GGML_METAL_KERNEL_TYPE_MUL_MV_IQ3_S_F32,
GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_S_F32,
GGML_METAL_KERNEL_TYPE_MUL_MV_IQ1_S_F32,
GGML_METAL_KERNEL_TYPE_MUL_MV_IQ4_NL_F32,
GGML_METAL_KERNEL_TYPE_MUL_MV_IQ4_XS_F32,
GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F32_F32,
//GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F16_F16,
GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F16_F32,
@ -105,8 +111,11 @@ enum ggml_metal_kernel_type {
GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ2_XXS_F32,
GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ2_XS_F32,
GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ3_XXS_F32,
GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ3_S_F32,
GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ2_S_F32,
GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ1_S_F32,
GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ4_NL_F32,
GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ4_XS_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_F32_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_F16_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_Q4_0_F32,
@ -122,8 +131,11 @@ enum ggml_metal_kernel_type {
GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_XXS_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_XS_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_IQ3_XXS_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_IQ3_S_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_S_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_IQ1_S_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_IQ4_NL_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_IQ4_XS_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_ID_F32_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_ID_F16_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_0_F32,
@ -139,8 +151,11 @@ enum ggml_metal_kernel_type {
GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_XXS_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_XS_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ3_XXS_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ3_S_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_S_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ1_S_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ4_NL_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ4_XS_F32,
GGML_METAL_KERNEL_TYPE_ROPE_F32,
GGML_METAL_KERNEL_TYPE_ROPE_F16,
GGML_METAL_KERNEL_TYPE_ALIBI_F32,
@ -452,8 +467,11 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) {
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ2_XXS, get_rows_iq2_xxs, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ2_XS, get_rows_iq2_xs, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ3_XXS, get_rows_iq3_xxs, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ3_S, get_rows_iq3_s, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ2_S, get_rows_iq2_s, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ1_S, get_rows_iq1_s, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ4_NL, get_rows_iq4_nl, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ4_XS, get_rows_iq4_xs, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_I32, get_rows_i32, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_RMS_NORM, rms_norm, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GROUP_NORM, group_norm, ctx->support_simdgroup_reduction);
@ -476,8 +494,11 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) {
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_XXS_F32, mul_mv_iq2_xxs_f32, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_XS_F32, mul_mv_iq2_xs_f32, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ3_XXS_F32, mul_mv_iq3_xxs_f32, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ3_S_F32, mul_mv_iq3_s_f32, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_S_F32, mul_mv_iq2_s_f32, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ1_S_F32, mul_mv_iq1_s_f32, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ4_NL_F32, mul_mv_iq4_nl_f32, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ4_XS_F32, mul_mv_iq4_xs_f32, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F32_F32, mul_mv_id_f32_f32, ctx->support_simdgroup_reduction);
//GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F16_F16, mul_mv_id_f16_f16, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F16_F32, mul_mv_id_f16_f32, ctx->support_simdgroup_reduction);
@ -496,8 +517,11 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) {
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ2_XXS_F32, mul_mv_id_iq2_xxs_f32, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ2_XS_F32, mul_mv_id_iq2_xs_f32, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ3_XXS_F32, mul_mv_id_iq3_xxs_f32, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ3_S_F32, mul_mv_id_iq3_s_f32, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ2_S_F32, mul_mv_id_iq2_s_f32, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ1_S_F32, mul_mv_id_iq1_s_f32, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ4_NL_F32, mul_mv_id_iq4_nl_f32, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ4_XS_F32, mul_mv_id_iq4_xs_f32, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_F32_F32, mul_mm_f32_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_F16_F32, mul_mm_f16_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q4_0_F32, mul_mm_q4_0_f32, ctx->support_simdgroup_mm);
@ -513,8 +537,11 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) {
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_XXS_F32, mul_mm_iq2_xxs_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_XS_F32, mul_mm_iq2_xs_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ3_XXS_F32, mul_mm_iq3_xxs_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ3_S_F32, mul_mm_iq3_s_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_S_F32, mul_mm_iq2_s_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ1_S_F32, mul_mm_iq1_s_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ4_NL_F32, mul_mm_iq4_nl_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ4_XS_F32, mul_mm_iq4_xs_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_F32_F32, mul_mm_id_f32_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_F16_F32, mul_mm_id_f16_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_0_F32, mul_mm_id_q4_0_f32, ctx->support_simdgroup_mm);
@ -530,8 +557,11 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) {
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_XXS_F32, mul_mm_id_iq2_xxs_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_XS_F32, mul_mm_id_iq2_xs_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ3_XXS_F32, mul_mm_id_iq3_xxs_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ3_S_F32, mul_mm_id_iq3_s_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_S_F32, mul_mm_id_iq2_s_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ1_S_F32, mul_mm_id_iq1_s_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ4_NL_F32, mul_mm_id_iq4_nl_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ4_XS_F32, mul_mm_id_iq4_xs_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ROPE_F32, rope_f32, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ROPE_F16, rope_f16, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ALIBI_F32, alibi_f32, true);
@ -1347,8 +1377,11 @@ static bool ggml_metal_graph_compute(
case GGML_TYPE_IQ2_XXS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_XXS_F32].pipeline; break;
case GGML_TYPE_IQ2_XS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_XS_F32 ].pipeline; break;
case GGML_TYPE_IQ3_XXS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ3_XXS_F32].pipeline; break;
case GGML_TYPE_IQ3_S: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ3_S_F32 ].pipeline; break;
case GGML_TYPE_IQ2_S: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_S_F32 ].pipeline; break;
case GGML_TYPE_IQ1_S: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ1_S_F32 ].pipeline; break;
case GGML_TYPE_IQ4_NL: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ4_NL_F32 ].pipeline; break;
case GGML_TYPE_IQ4_XS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ4_XS_F32 ].pipeline; break;
default: GGML_ASSERT(false && "MUL MAT-MAT not implemented");
}
@ -1483,6 +1516,18 @@ static bool ggml_metal_graph_compute(
nth1 = 16;
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ3_XXS_F32].pipeline;
} break;
case GGML_TYPE_IQ3_S:
{
nth0 = 4;
nth1 = 16;
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ3_S_F32].pipeline;
} break;
case GGML_TYPE_IQ2_S:
{
nth0 = 4;
nth1 = 16;
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_S_F32].pipeline;
} break;
case GGML_TYPE_IQ1_S:
{
nth0 = 4;
@ -1495,6 +1540,12 @@ static bool ggml_metal_graph_compute(
nth1 = 16;
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ4_NL_F32].pipeline;
} break;
case GGML_TYPE_IQ4_XS:
{
nth0 = 4;
nth1 = 16;
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ4_XS_F32].pipeline;
} break;
default:
{
GGML_METAL_LOG_ERROR("Asserting on type %d\n", (int)src0t);
@ -1527,9 +1578,9 @@ static bool ggml_metal_graph_compute(
[encoder setBytes:&r2 length:sizeof(r2) atIndex:17];
[encoder setBytes:&r3 length:sizeof(r3) atIndex:18];
if (src0t == GGML_TYPE_Q4_0 || src0t == GGML_TYPE_Q4_1 ||
src0t == GGML_TYPE_Q5_0 || src0t == GGML_TYPE_Q5_1 || src0t == GGML_TYPE_Q8_0 ||
src0t == GGML_TYPE_Q2_K || src0t == GGML_TYPE_IQ1_S) { // || src0t == GGML_TYPE_Q4_K) {
if (src0t == GGML_TYPE_Q4_0 || src0t == GGML_TYPE_Q4_1 ||
src0t == GGML_TYPE_Q5_0 || src0t == GGML_TYPE_Q5_1 || src0t == GGML_TYPE_Q8_0 ||
src0t == GGML_TYPE_Q2_K || src0t == GGML_TYPE_IQ1_S || src0t == GGML_TYPE_IQ2_S) {
[encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7)/8, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
}
else if (src0t == GGML_TYPE_IQ2_XXS || src0t == GGML_TYPE_IQ2_XS) {
@ -1537,12 +1588,12 @@ static bool ggml_metal_graph_compute(
[encoder setThreadgroupMemoryLength:mem_size atIndex:0];
[encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7)/8, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
}
else if (src0t == GGML_TYPE_IQ3_XXS) {
const int mem_size = 256*4+128;
else if (src0t == GGML_TYPE_IQ3_XXS || src0t == GGML_TYPE_IQ3_S) {
const int mem_size = src0t == GGML_TYPE_IQ3_XXS ? 256*4+128 : 512*4;
[encoder setThreadgroupMemoryLength:mem_size atIndex:0];
[encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7)/8, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
}
else if (src0t == GGML_TYPE_IQ4_NL) {
else if (src0t == GGML_TYPE_IQ4_NL || src0t == GGML_TYPE_IQ4_XS) {
const int mem_size = 32*sizeof(float);
[encoder setThreadgroupMemoryLength:mem_size atIndex:0];
[encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
@ -1640,8 +1691,11 @@ static bool ggml_metal_graph_compute(
case GGML_TYPE_IQ2_XXS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_XXS_F32].pipeline; break;
case GGML_TYPE_IQ2_XS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_XS_F32 ].pipeline; break;
case GGML_TYPE_IQ3_XXS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ3_XXS_F32].pipeline; break;
case GGML_TYPE_IQ3_S: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ3_S_F32 ].pipeline; break;
case GGML_TYPE_IQ2_S: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_S_F32 ].pipeline; break;
case GGML_TYPE_IQ1_S: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ1_S_F32 ].pipeline; break;
case GGML_TYPE_IQ4_NL: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ4_NL_F32 ].pipeline; break;
case GGML_TYPE_IQ4_XS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ4_XS_F32 ].pipeline; break;
default: GGML_ASSERT(false && "MUL_MAT_ID not implemented");
}
@ -1779,6 +1833,18 @@ static bool ggml_metal_graph_compute(
nth1 = 16;
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ3_XXS_F32].pipeline;
} break;
case GGML_TYPE_IQ3_S:
{
nth0 = 4;
nth1 = 16;
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ3_S_F32].pipeline;
} break;
case GGML_TYPE_IQ2_S:
{
nth0 = 4;
nth1 = 16;
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ2_S_F32].pipeline;
} break;
case GGML_TYPE_IQ1_S:
{
nth0 = 4;
@ -1791,6 +1857,12 @@ static bool ggml_metal_graph_compute(
nth1 = 16;
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ4_NL_F32].pipeline;
} break;
case GGML_TYPE_IQ4_XS:
{
nth0 = 4;
nth1 = 16;
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ4_XS_F32].pipeline;
} break;
default:
{
GGML_METAL_LOG_ERROR("Asserting on type %d\n", (int)src2t);
@ -1839,9 +1911,9 @@ static bool ggml_metal_graph_compute(
[encoder setBuffer:id_src_cur offset:offs_src_cur atIndex:23 + j];
}
if (src2t == GGML_TYPE_Q4_0 || src2t == GGML_TYPE_Q4_1 ||
src2t == GGML_TYPE_Q5_0 || src2t == GGML_TYPE_Q5_1 || src2t == GGML_TYPE_Q8_0 ||
src2t == GGML_TYPE_Q2_K || src2t == GGML_TYPE_IQ1_S) { // || src2t == GGML_TYPE_Q4_K) {
if (src2t == GGML_TYPE_Q4_0 || src2t == GGML_TYPE_Q4_1 ||
src2t == GGML_TYPE_Q5_0 || src2t == GGML_TYPE_Q5_1 || src2t == GGML_TYPE_Q8_0 ||
src2t == GGML_TYPE_Q2_K || src2t == GGML_TYPE_IQ1_S || src2t == GGML_TYPE_IQ2_S) {
[encoder dispatchThreadgroups:MTLSizeMake((ne21 + 7)/8, _ne1, ne01*ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
}
else if (src2t == GGML_TYPE_IQ2_XXS || src2t == GGML_TYPE_IQ2_XS) {
@ -1849,12 +1921,12 @@ static bool ggml_metal_graph_compute(
[encoder setThreadgroupMemoryLength:mem_size atIndex:0];
[encoder dispatchThreadgroups:MTLSizeMake((ne21 + 7)/8, _ne1, ne01*ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
}
else if (src2t == GGML_TYPE_IQ3_XXS) {
const int mem_size = 256*4+128;
else if (src2t == GGML_TYPE_IQ3_XXS || src2t == GGML_TYPE_IQ3_S) {
const int mem_size = src2t == GGML_TYPE_IQ3_XXS ? 256*4+128 : 512*4;
[encoder setThreadgroupMemoryLength:mem_size atIndex:0];
[encoder dispatchThreadgroups:MTLSizeMake((ne21 + 7)/8, _ne1, ne01*ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
}
else if (src2t == GGML_TYPE_IQ4_NL) {
else if (src2t == GGML_TYPE_IQ4_NL || src2t == GGML_TYPE_IQ4_XS) {
const int mem_size = 32*sizeof(float);
[encoder setThreadgroupMemoryLength:mem_size atIndex:0];
[encoder dispatchThreadgroups:MTLSizeMake((ne21 + 3)/4, _ne1, ne01*ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
@ -1900,8 +1972,11 @@ static bool ggml_metal_graph_compute(
case GGML_TYPE_IQ2_XXS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ2_XXS].pipeline; break;
case GGML_TYPE_IQ2_XS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ2_XS ].pipeline; break;
case GGML_TYPE_IQ3_XXS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ3_XXS].pipeline; break;
case GGML_TYPE_IQ3_S: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ3_S ].pipeline; break;
case GGML_TYPE_IQ2_S: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ2_S ].pipeline; break;
case GGML_TYPE_IQ1_S: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ1_S ].pipeline; break;
case GGML_TYPE_IQ4_NL: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ4_NL ].pipeline; break;
case GGML_TYPE_IQ4_XS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ4_XS ].pipeline; break;
case GGML_TYPE_I32: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_I32 ].pipeline; break;
default: GGML_ASSERT(false && "not implemented");
}
@ -2237,8 +2312,8 @@ static bool ggml_metal_graph_compute(
id<MTLComputePipelineState> pipeline = nil;
switch (order) {
case GGML_SORT_ASC: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ARGSORT_F32_I32_ASC].pipeline; break;
case GGML_SORT_DESC: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ARGSORT_F32_I32_DESC].pipeline; break;
case GGML_SORT_ORDER_ASC: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ARGSORT_F32_I32_ASC].pipeline; break;
case GGML_SORT_ORDER_DESC: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ARGSORT_F32_I32_DESC].pipeline; break;
default: GGML_ASSERT(false);
};
@ -2696,6 +2771,11 @@ void ggml_backend_metal_log_set_callback(ggml_log_callback log_callback, void *
ggml_metal_log_user_data = user_data;
}
static ggml_guid_t ggml_backend_metal_guid(void) {
static ggml_guid guid = { 0x81, 0xa1, 0x8b, 0x1e, 0x71, 0xec, 0x79, 0xed, 0x2b, 0x85, 0xdc, 0x8a, 0x61, 0x98, 0x30, 0xe6 };
return &guid;
}
ggml_backend_t ggml_backend_metal_init(void) {
struct ggml_metal_context * ctx = ggml_metal_init(GGML_DEFAULT_N_THREADS);
@ -2706,6 +2786,7 @@ ggml_backend_t ggml_backend_metal_init(void) {
ggml_backend_t metal_backend = malloc(sizeof(struct ggml_backend));
*metal_backend = (struct ggml_backend) {
/* .guid = */ ggml_backend_metal_guid(),
/* .interface = */ ggml_backend_metal_i,
/* .context = */ ctx,
};
@ -2714,7 +2795,7 @@ ggml_backend_t ggml_backend_metal_init(void) {
}
bool ggml_backend_is_metal(ggml_backend_t backend) {
return backend && backend->iface.get_name == ggml_backend_metal_name;
return backend != NULL && ggml_guid_matches(backend->guid, ggml_backend_metal_guid());
}
void ggml_backend_metal_set_n_cb(ggml_backend_t backend, int n_cb) {

1073
ggml-metal.metal
View file

File diff suppressed because it is too large Load diff

50
ggml-opencl.cpp
View file

@ -1354,7 +1354,7 @@ static void ggml_cl_pool_free(cl_mem mem, size_t size) {
}
void ggml_cl_free_data(const struct ggml_tensor* tensor) {
if (tensor->backend != GGML_BACKEND_GPU) {
if (tensor->backend != GGML_BACKEND_TYPE_GPU) {
return;
}
@ -1412,7 +1412,7 @@ static cl_int ggml_cl_h2d_tensor_2d(cl_command_queue queue, cl_mem dst, size_t o
}
static void ggml_cl_mul_f32(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
GGML_ASSERT(src1->backend == GGML_BACKEND_GPU);
GGML_ASSERT(src1->backend == GGML_BACKEND_TYPE_GPU);
const int64_t ne00 = src0->ne[0];
const int64_t ne01 = src0->ne[1];
const int64_t ne02 = src0->ne[2];
@ -1476,7 +1476,7 @@ void ggml_cl_mul(const struct ggml_tensor * src0, const struct ggml_tensor * src
}
static void ggml_cl_add_f32(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
GGML_ASSERT(src1->backend == GGML_BACKEND_GPU);
GGML_ASSERT(src1->backend == GGML_BACKEND_TYPE_GPU);
const int64_t ne00 = src0->ne[0];
const int64_t ne01 = src0->ne[1];
const int64_t ne02 = src0->ne[2];
@ -1566,13 +1566,13 @@ static void ggml_cl_mul_mat_f32(const ggml_tensor * src0, const ggml_tensor * sr
size_t y_size;
size_t d_size;
cl_mem d_X;
if (src0->backend == GGML_BACKEND_GPU) { // NOLINT
if (src0->backend == GGML_BACKEND_TYPE_GPU) { // NOLINT
d_X = (cl_mem) src0->extra;
} else {
d_X = ggml_cl_pool_malloc(sizeof(float) * x_ne, &x_size);
}
cl_mem d_Y = src1->backend == GGML_BACKEND_GPU ? (cl_mem) src1->extra : ggml_cl_pool_malloc(sizeof(float) * y_ne, &y_size);
cl_mem d_D = dst->backend == GGML_BACKEND_GPU ? (cl_mem) dst->extra : ggml_cl_pool_malloc(sizeof(float) * d_ne, &d_size);
cl_mem d_Y = src1->backend == GGML_BACKEND_TYPE_GPU ? (cl_mem) src1->extra : ggml_cl_pool_malloc(sizeof(float) * y_ne, &y_size);
cl_mem d_D = dst->backend == GGML_BACKEND_TYPE_GPU ? (cl_mem) dst->extra : ggml_cl_pool_malloc(sizeof(float) * d_ne, &d_size);
size_t x_offset = 0;
@ -1580,7 +1580,7 @@ static void ggml_cl_mul_mat_f32(const ggml_tensor * src0, const ggml_tensor * sr
// TODO: copy src0 here when r3>1
for (int64_t i13 = i03 * r3, e13 = i13 + r3; i13 < e13; i13++) {
for (int64_t i02 = 0; i02 < ne02; i02++) {
if (src0->backend == GGML_BACKEND_GPU) {
if (src0->backend == GGML_BACKEND_TYPE_GPU) {
x_offset = (i03 * ne02 + i02) * x_ne;
} else {
// copy src0 to device
@ -1589,7 +1589,7 @@ static void ggml_cl_mul_mat_f32(const ggml_tensor * src0, const ggml_tensor * sr
for (int64_t i12 = i02 * r2, e12 = i12 + r2; i12 < e12; i12++) {
// copy src1 to device
if (src1->backend == GGML_BACKEND_CPU) {
if (src1->backend == GGML_BACKEND_TYPE_CPU) {
CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_Y, 0, src1, i13, i12, NULL));
}
@ -1612,7 +1612,7 @@ static void ggml_cl_mul_mat_f32(const ggml_tensor * src0, const ggml_tensor * sr
}
// copy dst to host
if (dst->backend == GGML_BACKEND_CPU) {
if (dst->backend == GGML_BACKEND_TYPE_CPU) {
float * d = (float *) ((char *) dst->data + i12*nb2 + i13*nb3);
CL_CHECK(clEnqueueReadBuffer(queue, d_D, true, 0, sizeof(float) * d_ne, d, 1, &ev_sgemm, NULL));
}
@ -1621,13 +1621,13 @@ static void ggml_cl_mul_mat_f32(const ggml_tensor * src0, const ggml_tensor * sr
}
}
if (src0->backend != GGML_BACKEND_GPU) {
if (src0->backend != GGML_BACKEND_TYPE_GPU) {
ggml_cl_pool_free(d_X, x_size);
}
if (src1->backend != GGML_BACKEND_GPU) {
if (src1->backend != GGML_BACKEND_TYPE_GPU) {
ggml_cl_pool_free(d_Y, y_size);
}
if (dst->backend != GGML_BACKEND_GPU) {
if (dst->backend != GGML_BACKEND_TYPE_GPU) {
ggml_cl_pool_free(d_D, d_size);
}
}
@ -1670,7 +1670,7 @@ static void ggml_cl_mul_mat_f16(const ggml_tensor * src0, const ggml_tensor * sr
size_t y_size;
size_t d_size;
cl_mem d_X;
if (src0->backend == GGML_BACKEND_GPU) { // NOLINT
if (src0->backend == GGML_BACKEND_TYPE_GPU) { // NOLINT
d_X = (cl_mem) src0->extra;
} else {
d_X = ggml_cl_pool_malloc(sizeof(ggml_fp16_t) * x_ne, &x_size);
@ -1687,7 +1687,7 @@ static void ggml_cl_mul_mat_f16(const ggml_tensor * src0, const ggml_tensor * sr
// TODO: copy src0 here when r3>1
for (int64_t i13 = i03 * r3, e13 = i13 + r3; i13 < e13; i13++) {
for (int64_t i02 = 0; i02 < ne02; i02++) {
if (src0->backend == GGML_BACKEND_GPU) {
if (src0->backend == GGML_BACKEND_TYPE_GPU) {
x_offset = (i03 * ne02 + i02) * x_ne;
} else {
// copy src0 to device
@ -1741,7 +1741,7 @@ static void ggml_cl_mul_mat_f16(const ggml_tensor * src0, const ggml_tensor * sr
}
// copy dst to host, then convert to float
if (dst->backend == GGML_BACKEND_CPU) {
if (dst->backend == GGML_BACKEND_TYPE_CPU) {
CL_CHECK(clEnqueueReadBuffer(queue, d_D, true, 0, sizeof(ggml_fp16_t) * d_ne, tmp, 1, &ev_sgemm, NULL));
float * d = (float *) ((char *) dst->data + i12*nb2 + i13*nb3);
ggml_fp16_to_fp32_row(tmp, d, d_ne);
@ -1753,7 +1753,7 @@ static void ggml_cl_mul_mat_f16(const ggml_tensor * src0, const ggml_tensor * sr
}
}
if (src0->backend != GGML_BACKEND_GPU) {
if (src0->backend != GGML_BACKEND_TYPE_GPU) {
ggml_cl_pool_free(d_X, x_size);
}
ggml_cl_pool_free(d_Y, y_size);
@ -1798,7 +1798,7 @@ static void ggml_cl_mul_mat_q_f32(const ggml_tensor * src0, const ggml_tensor *
cl_mem d_Y = ggml_cl_pool_malloc(sizeof(float) * y_ne, &y_size);
cl_mem d_D = ggml_cl_pool_malloc(sizeof(float) * d_ne, &d_size);
cl_mem d_Q;
if (src0->backend == GGML_BACKEND_CPU) {
if (src0->backend == GGML_BACKEND_TYPE_CPU) {
d_Q = ggml_cl_pool_malloc(q_sz, &q_size);
}
@ -1817,10 +1817,10 @@ static void ggml_cl_mul_mat_q_f32(const ggml_tensor * src0, const ggml_tensor *
for (int64_t i13 = i03 * r3, e13 = i13 + r3; i13 < e13; i13++) {
for (int64_t i02 = 0; i02 < ne02; i02++) {
// copy src0 to device if necessary
if (src0->backend == GGML_BACKEND_CPU) {
if (src0->backend == GGML_BACKEND_TYPE_CPU) {
events.emplace_back();
CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_Q, 0, src0, i03, i02, events.data() + ev_idx++));
} else if (src0->backend == GGML_BACKEND_GPU) {
} else if (src0->backend == GGML_BACKEND_TYPE_GPU) {
d_Q = (cl_mem) src0->extra;
} else {
GGML_ASSERT(false);
@ -1829,7 +1829,7 @@ static void ggml_cl_mul_mat_q_f32(const ggml_tensor * src0, const ggml_tensor *
if (!mul_mat_vec) {
// convert src0 to fp32 on device
const size_t global = x_ne / global_denom;
const size_t offset = src0->backend == GGML_BACKEND_GPU ? (i03 * ne02 + i02) * x_bps : 0;
const size_t offset = src0->backend == GGML_BACKEND_TYPE_GPU ? (i03 * ne02 + i02) * x_bps : 0;
CL_CHECK(clSetKernelArg(*to_fp32_cl, 0, sizeof(cl_mem), &d_Q));
CL_CHECK(clSetKernelArg(*to_fp32_cl, 1, sizeof(cl_mem), &d_X));
CL_CHECK(clEnqueueNDRangeKernel(queue, *to_fp32_cl, 1, &offset, &global, local > 0 ? &local : NULL, events.size(), !events.empty() ? events.data() : NULL, NULL));
@ -1843,7 +1843,7 @@ static void ggml_cl_mul_mat_q_f32(const ggml_tensor * src0, const ggml_tensor *
// compute
const size_t global = ne01 * local;
const size_t offset = src0->backend == GGML_BACKEND_GPU ? (i03 * ne02 + i02) * x_bps : 0;
const size_t offset = src0->backend == GGML_BACKEND_TYPE_GPU ? (i03 * ne02 + i02) * x_bps : 0;
const cl_int ncols = ne00;
events.emplace_back();
CL_CHECK(clSetKernelArg(*dmmv, 0, sizeof(cl_mem), &d_Q));
@ -1895,7 +1895,7 @@ static void ggml_cl_mul_mat_q_f32(const ggml_tensor * src0, const ggml_tensor *
}
ggml_cl_pool_free(d_Y, y_size);
ggml_cl_pool_free(d_D, d_size);
if (src0->backend == GGML_BACKEND_CPU) {
if (src0->backend == GGML_BACKEND_TYPE_CPU) {
ggml_cl_pool_free(d_Q, q_size);
}
}
@ -1911,7 +1911,7 @@ bool ggml_cl_can_mul_mat(const struct ggml_tensor * src0, const struct ggml_tens
if ((src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16 || ggml_is_quantized(src0->type)) &&
src1->type == GGML_TYPE_F32 &&
dst->type == GGML_TYPE_F32 &&
((ne0 >= 32 && ne1 >= 32 && ne10 >= 32) || src0->backend == GGML_BACKEND_GPU)) {
((ne0 >= 32 && ne1 >= 32 && ne10 >= 32) || src0->backend == GGML_BACKEND_TYPE_GPU)) {
return true;
}
@ -1993,7 +1993,7 @@ void ggml_cl_transform_tensor(void * data, ggml_tensor * tensor) {
CL_CHECK(clFinish(queue));
tensor->extra = dst;
GGML_ASSERT(tensor->backend == GGML_BACKEND_GPU);
GGML_ASSERT(tensor->backend == GGML_BACKEND_TYPE_GPU);
}
// ggml-backend
@ -2045,7 +2045,7 @@ static void ggml_backend_opencl_buffer_init_tensor(ggml_backend_buffer_t buffer,
ctx->sub_buffers.push_back(sub_buffer);
tensor->extra = sub_buffer;
}
tensor->backend = GGML_BACKEND_GPU;
tensor->backend = GGML_BACKEND_TYPE_GPU;
}
static void ggml_backend_opencl_buffer_set_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor, const void * data, size_t offset, size_t size) {

1887
ggml-quants.c
View file

File diff suppressed because it is too large Load diff

52
ggml-quants.h
View file

@ -182,6 +182,15 @@ typedef struct {
} block_iq2_xs;
static_assert(sizeof(block_iq2_xs) == sizeof(ggml_fp16_t) + QK_K/8*sizeof(uint16_t) + QK_K/32, "wrong iq2_xs block size/padding");
// 2.5625 bpw quants
typedef struct {
ggml_fp16_t d;
uint8_t qs[QK_K/4];
uint8_t qh[QK_K/32];
uint8_t scales[QK_K/32];
} block_iq2_s;
static_assert(sizeof(block_iq2_s) == sizeof(ggml_fp16_t) + QK_K/4 + QK_K/16, "wrong iq2_s block size/padding");
// (Almost) "true" 3-bit quantization.
// Due to the need to use blocks as per ggml design, it ends up using
// 3.0625 bpw because of the 16-bit scale for each block of 256.
@ -191,6 +200,21 @@ typedef struct {
} block_iq3_xxs;
static_assert(sizeof(block_iq3_xxs) == sizeof(ggml_fp16_t) + 3*(QK_K/8), "wrong iq3_xxs block size/padding");
// 3.4375 bpw
#if QK_K == 64
#define IQ3S_N_SCALE 2
#else
#define IQ3S_N_SCALE QK_K/64
#endif
typedef struct {
ggml_fp16_t d;
uint8_t qs[QK_K/4];
uint8_t qh[QK_K/32];
uint8_t signs[QK_K/8];
uint8_t scales[IQ3S_N_SCALE];
} block_iq3_s;
static_assert(sizeof(block_iq3_s) == sizeof(ggml_fp16_t) + 13*(QK_K/32) + IQ3S_N_SCALE, "wrong iq3_s block size/padding");
typedef struct {
ggml_fp16_t d;
uint8_t qs[QK_K/8];
@ -206,6 +230,19 @@ typedef struct {
} block_iq4_nl;
static_assert(sizeof(block_iq4_nl) == sizeof(ggml_fp16_t) + QK4_NL/2, "wrong iq4_nl block size/padding");
#if QK_K == 64
#define block_iq4_xs block_iq4_nl
//typedef struct block_iq4_nl block_iq4_xs;
#else
typedef struct {
ggml_fp16_t d;
uint16_t scales_h;
uint8_t scales_l[QK_K/64];
uint8_t qs[QK_K/2];
} block_iq4_xs;
static_assert(sizeof(block_iq4_xs) == sizeof(ggml_fp16_t) + sizeof(uint16_t) + QK_K/64 + QK_K/2, "wrong iq4_xs block size/padding");
#endif
#ifdef __cplusplus
extern "C" {
#endif
@ -226,6 +263,9 @@ void quantize_row_q6_K_reference(const float * GGML_RESTRICT x, block_q6_K * GGM
void quantize_row_q8_K_reference(const float * GGML_RESTRICT x, block_q8_K * GGML_RESTRICT y, int k);
void quantize_row_iq3_xxs_reference(const float * GGML_RESTRICT x, block_iq3_xxs * GGML_RESTRICT y, int k);
void quantize_row_iq4_nl_reference (const float * GGML_RESTRICT x, block_iq4_nl * GGML_RESTRICT y, int k);
void quantize_row_iq4_xs_reference (const float * GGML_RESTRICT x, block_iq4_xs * GGML_RESTRICT y, int k);
void quantize_row_iq3_s_reference (const float * GGML_RESTRICT x, block_iq3_s * GGML_RESTRICT y, int k);
void quantize_row_iq2_s_reference (const float * GGML_RESTRICT x, block_iq2_s * GGML_RESTRICT y, int k);
void quantize_row_q4_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int k);
void quantize_row_q4_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int k);
@ -242,6 +282,9 @@ void quantize_row_q6_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, in
void quantize_row_q8_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int k);
void quantize_row_iq3_xxs(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int k);
void quantize_row_iq4_nl (const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int k);
void quantize_row_iq4_xs (const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int k);
void quantize_row_iq3_s (const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int k);
void quantize_row_iq2_s (const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int k);
// Dequantization
void dequantize_row_q4_0(const block_q4_0 * GGML_RESTRICT x, float * GGML_RESTRICT y, int k);
@ -259,9 +302,12 @@ void dequantize_row_q6_K(const block_q6_K * GGML_RESTRICT x, float * GGML_RESTRI
void dequantize_row_q8_K(const block_q8_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int k);
void dequantize_row_iq2_xxs(const block_iq2_xxs * GGML_RESTRICT x, float * GGML_RESTRICT y, int k);
void dequantize_row_iq2_xs (const block_iq2_xs * GGML_RESTRICT x, float * GGML_RESTRICT y, int k);
void dequantize_row_iq2_s (const block_iq2_s * GGML_RESTRICT x, float * GGML_RESTRICT y, int k);
void dequantize_row_iq3_xxs(const block_iq3_xxs * GGML_RESTRICT x, float * GGML_RESTRICT y, int k);
void dequantize_row_iq1_s (const block_iq1_s * GGML_RESTRICT x, float * GGML_RESTRICT y, int k);
void dequantize_row_iq4_nl (const block_iq4_nl * GGML_RESTRICT x, float * GGML_RESTRICT y, int k);
void dequantize_row_iq4_xs (const block_iq4_xs * GGML_RESTRICT x, float * GGML_RESTRICT y, int k);
void dequantize_row_iq3_s (const block_iq3_s * GGML_RESTRICT x, float * GGML_RESTRICT y, int k);
// Dot product
void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
@ -277,18 +323,24 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_iq2_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_iq2_xs_q8_K (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_iq2_s_q8_K (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_iq3_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_iq1_s_q8_K (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_iq4_nl_q8_0 (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_iq4_xs_q8_K (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_iq3_s_q8_K (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
//
// Quantization utilizing an importance matrix (a.k.a. "Activation aWare Quantization")
//
size_t quantize_iq2_xxs(const float * src, void * dst, int nrows, int n_per_row, int64_t * hist, const float * imatrix);
size_t quantize_iq2_xs (const float * src, void * dst, int nrows, int n_per_row, int64_t * hist, const float * imatrix);
size_t quantize_iq2_s (const float * src, void * dst, int nrows, int n_per_row, int64_t * hist, const float * imatrix);
size_t quantize_iq3_xxs(const float * src, void * dst, int nrows, int n_per_row, int64_t * hist, const float * imatrix);
size_t quantize_iq1_s (const float * src, void * dst, int nrows, int n_per_row, int64_t * hist, const float * imatrix);
size_t quantize_iq4_nl (const float * src, void * dst, int nrows, int n_per_row, int64_t * hist, const float * imatrix);
size_t quantize_iq4_xs (const float * src, void * dst, int nrows, int n_per_row, int64_t * hist, const float * imatrix);
size_t quantize_iq3_s (const float * src, void * dst, int nrows, int n_per_row, int64_t * hist, const float * imatrix);
size_t quantize_q2_K (const float * src, void * dst, int nrows, int n_per_row, int64_t * hist, const float * imatrix);
size_t quantize_q3_K (const float * src, void * dst, int nrows, int n_per_row, int64_t * hist, const float * imatrix);
size_t quantize_q4_K (const float * src, void * dst, int nrows, int n_per_row, int64_t * hist, const float * imatrix);

408
ggml-sycl.cpp
View file

@ -3338,7 +3338,7 @@ void print_ggml_tensor(const char*name, struct ggml_tensor *src){
size_t total_elements = ggml_nelements(src);
const bool src_on_device = src->backend == GGML_BACKEND_GPU || src->backend == GGML_BACKEND_GPU_SPLIT;
const bool src_on_device = src->backend == GGML_BACKEND_TYPE_GPU || src->backend == GGML_BACKEND_TYPE_GPU_SPLIT;
float *src_data =NULL;
if(src_on_device) {
ggml_tensor_extra_gpu * src_extra = (ggml_tensor_extra_gpu *) src->extra;
@ -8086,11 +8086,11 @@ static void k_argsort_f32_i32(const float * x, int * dst, const int ncols,
int ixj = col ^ j;
if (ixj > col) {
if ((col & k) == 0) {
if (order == GGML_SORT_ASC ? x_row[dst_row[col]] > x_row[dst_row[ixj]] : x_row[dst_row[col]] < x_row[dst_row[ixj]]) {
if (order == GGML_SORT_ORDER_ASC ? x_row[dst_row[col]] > x_row[dst_row[ixj]] : x_row[dst_row[col]] < x_row[dst_row[ixj]]) {
swap(dst_row[col], dst_row[ixj]);
}
} else {
if (order == GGML_SORT_ASC ? x_row[dst_row[col]] < x_row[dst_row[ixj]] : x_row[dst_row[col]] > x_row[dst_row[ixj]]) {
if (order == GGML_SORT_ORDER_ASC ? x_row[dst_row[col]] < x_row[dst_row[ixj]] : x_row[dst_row[col]] > x_row[dst_row[ixj]]) {
swap(dst_row[col], dst_row[ixj]);
}
}
@ -8126,23 +8126,51 @@ static void diag_mask_inf_f32(const float * x, float * dst, const int ncols, con
dst[i] = x[i] - (col > n_past + row % rows_per_channel) * FLT_MAX;
}
static void soft_max_f32(const float * x, const float * y, float * dst, const int ncols, const int nrows_y, const float scale,
const sycl::nd_item<3> &item_ct1, float *buf) {
template <bool vals_smem, int ncols_template, int block_size_template>
static void soft_max_f32(const float * x, const float * mask, const float *pos, float * dst, const int ncols_par,
const int nrows_y, const float scale, const float max_bias, const float m0,
const float m1, uint32_t n_head_log2, const sycl::nd_item<3> &item_ct1, float *buf) {
const int ncols = ncols_template == 0 ? ncols_par : ncols_template;
const int tid = item_ct1.get_local_id(2);
const int rowx = item_ct1.get_group(2);
const int rowy = rowx % nrows_y; // broadcast the mask (y) in the row dimension
const int block_size = item_ct1.get_local_range(2);
const int block_size = block_size_template == 0 ? item_ct1.get_local_range(2) : block_size_template;
const int warp_id = item_ct1.get_local_id(2) / WARP_SIZE;
const int lane_id = item_ct1.get_local_id(2) % WARP_SIZE;
float slope = 0.0f;
// ALiBi
if (max_bias > 0.0f) {
const uint32_t h = rowx/nrows_y; // head index
const float base = h < n_head_log2 ? m0 : m1;
const int exp = h < n_head_log2 ? h + 1 : 2*(h - n_head_log2) + 1;
slope = sycl::pow(base, float(exp));
}
float * vals = vals_smem ? buf + WARP_SIZE : dst + rowx*ncols;
float max_val = -INFINITY;
for (int col = tid; col < ncols; col += block_size) {
for (int col0 = 0; col0 < ncols; col0 += block_size) {
const int col = col0 + tid;
if (ncols_template == 0 && col >= ncols) {
break;
}
const int ix = rowx*ncols + col;
const int iy = rowy*ncols + col;
max_val = sycl::max(max_val, x[ix] * scale + (y ? y[iy] : 0.0f));
const float val = x[ix]*scale + (mask ? mask[iy] : 0.0f) + (pos ? slope*pos[col] : 0.0f);
vals[col] = val;
max_val = sycl::max(max_val, val);
}
// find the max value in the block
@ -8151,30 +8179,12 @@ static void soft_max_f32(const float * x, const float * y, float * dst, const in
if (warp_id == 0) {
buf[lane_id] = -INFINITY;
}
/*
DPCT1118:12: SYCL group functions and algorithms must be encountered in
converged control flow. You may need to adjust the code.
*/
/*
DPCT1065:60: Consider replacing sycl::nd_item::barrier() with
sycl::nd_item::barrier(sycl::access::fence_space::local_space) for
better performance if there is no access to global memory.
*/
item_ct1.barrier();
item_ct1.barrier(sycl::access::fence_space::local_space);
if (lane_id == 0) {
buf[warp_id] = max_val;
}
/*
DPCT1118:13: SYCL group functions and algorithms must be encountered in
converged control flow. You may need to adjust the code.
*/
/*
DPCT1065:61: Consider replacing sycl::nd_item::barrier() with
sycl::nd_item::barrier(sycl::access::fence_space::local_space) for
better performance if there is no access to global memory.
*/
item_ct1.barrier();
item_ct1.barrier(sycl::access::fence_space::local_space);
max_val = buf[lane_id];
max_val = warp_reduce_max(max_val, item_ct1);
@ -8182,13 +8192,16 @@ static void soft_max_f32(const float * x, const float * y, float * dst, const in
float tmp = 0.f;
for (int col = tid; col < ncols; col += block_size) {
const int ix = rowx*ncols + col;
const int iy = rowy*ncols + col;
const float val =
sycl::native::exp((x[ix] * scale + (y ? y[iy] : 0.0f)) - max_val);
#pragma unroll
for (int col0 = 0; col0 < ncols; col0 += block_size) {
const int col = col0 + tid;
if (ncols_template == 0 && col >= ncols) {
break;
}
const float val = sycl::native::exp(vals[col] - max_val);
tmp += val;
dst[ix] = val;
vals[col] = val;
}
// find the sum of exps in the block
@ -8197,40 +8210,29 @@ static void soft_max_f32(const float * x, const float * y, float * dst, const in
if (warp_id == 0) {
buf[lane_id] = 0.f;
}
/*
DPCT1118:14: SYCL group functions and algorithms must be encountered in
converged control flow. You may need to adjust the code.
*/
/*
DPCT1065:62: Consider replacing sycl::nd_item::barrier() with
sycl::nd_item::barrier(sycl::access::fence_space::local_space) for
better performance if there is no access to global memory.
*/
item_ct1.barrier();
item_ct1.barrier(sycl::access::fence_space::local_space);
if (lane_id == 0) {
buf[warp_id] = tmp;
}
/*
DPCT1118:15: SYCL group functions and algorithms must be encountered in
converged control flow. You may need to adjust the code.
*/
/*
DPCT1065:63: Consider replacing sycl::nd_item::barrier() with
sycl::nd_item::barrier(sycl::access::fence_space::local_space) for
better performance if there is no access to global memory.
*/
item_ct1.barrier();
item_ct1.barrier(sycl::access::fence_space::local_space);
tmp = buf[lane_id];
tmp = warp_reduce_sum(tmp, item_ct1);
}
const float inv_tmp = 1.f / tmp;
const float inv_sum = 1.f / tmp;
for (int col = tid; col < ncols; col += block_size) {
const int i = rowx*ncols + col;
dst[i] *= inv_tmp;
#pragma unroll
for (int col0 = 0; col0 < ncols; col0 += block_size) {
const int col = col0 + tid;
if (ncols_template == 0 && col >= ncols) {
return;
}
const int idst = rowx*ncols + col;
dst[idst] = vals[col] * inv_sum;
}
}
@ -10825,7 +10827,7 @@ static void argsort_f32_i32_sycl(const float *x, int *dst, const int ncols,
const sycl::range<3> block_dims(1, 1, ncols);
const sycl::range<3> block_nums(1, nrows, 1);
if (order == GGML_SORT_ASC) {
if (order == GGML_SORT_ORDER_ASC) {
/*
DPCT1049:44: The work-group size passed to the SYCL kernel may exceed
the limit. To get the device limit, query
@ -10834,9 +10836,9 @@ static void argsort_f32_i32_sycl(const float *x, int *dst, const int ncols,
stream->parallel_for(
sycl::nd_range<3>(block_nums * block_dims, block_dims),
[=](sycl::nd_item<3> item_ct1) {
k_argsort_f32_i32<GGML_SORT_ASC>(x, dst, ncols, item_ct1);
k_argsort_f32_i32<GGML_SORT_ORDER_ASC>(x, dst, ncols, item_ct1);
});
} else if (order == GGML_SORT_DESC) {
} else if (order == GGML_SORT_ORDER_DESC) {
/*
DPCT1049:45: The work-group size passed to the SYCL kernel may exceed
the limit. To get the device limit, query
@ -10845,7 +10847,7 @@ static void argsort_f32_i32_sycl(const float *x, int *dst, const int ncols,
stream->parallel_for(
sycl::nd_range<3>(block_nums * block_dims, block_dims),
[=](sycl::nd_item<3> item_ct1) {
k_argsort_f32_i32<GGML_SORT_DESC>(x, dst, ncols, item_ct1);
k_argsort_f32_i32<GGML_SORT_ORDER_DESC>(x, dst, ncols, item_ct1);
});
} else {
GGML_ASSERT(false);
@ -10867,35 +10869,96 @@ static void diag_mask_inf_f32_sycl(const float *x, float *dst,
});
}
static void soft_max_f32_sycl(const float *x, const float *y, float *dst,
const int ncols_x, const int nrows_x,
const int nrows_y, const float scale,
template <bool vals_smem, int ncols_template, int block_size_template>
static void soft_max_f32_submitter(const float * x, const float * mask, const float *pos, float * dst, const int ncols_par,
const int nrows_y, const float scale, const float max_bias, const float m0,
const float m1, uint32_t n_head_log2, sycl::range<3> block_nums, sycl::range<3> block_dims,
const size_t n_local_scratch, dpct::queue_ptr stream) {
stream->submit([&](sycl::handler &cgh) {
sycl::local_accessor<float, 1> local_buf_acc(n_local_scratch, cgh);
cgh.parallel_for(
sycl::nd_range<3>(block_nums * block_dims, block_dims),
[=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(32)]] {
soft_max_f32<vals_smem, ncols_template, block_size_template>(x, mask, pos, dst, ncols_par,
nrows_y, scale, max_bias, m0,
m1, n_head_log2, item_ct1,
local_buf_acc.get_pointer());
});
});
}
static void soft_max_f32_sycl(const float * x, const float * mask, const float * pos,
float * dst, const int ncols_x, const int nrows_x,
const int nrows_y, const float scale, const float max_bias,
dpct::queue_ptr stream) {
int nth = WARP_SIZE;
while (nth < ncols_x && nth < SYCL_SOFT_MAX_BLOCK_SIZE) nth *= 2;
const sycl::range<3> block_dims(1, 1, nth);
const sycl::range<3> block_nums(1, 1, nrows_x);
/*
DPCT1049:46: The work-group size passed to the SYCL kernel may exceed the
limit. To get the device limit, query info::device::max_work_group_size.
Adjust the work-group size if needed.
*/
stream->submit([&](sycl::handler &cgh) {
/*
DPCT1101:96: 'SYCL_SOFT_MAX_BLOCK_SIZE/WARP_SIZE' expression was
replaced with a value. Modify the code to use the original expression,
provided in comments, if it is correct.
*/
sycl::local_accessor<float, 1> buf_acc_ct1(
sycl::range<1>(32 /*SYCL_SOFT_MAX_BLOCK_SIZE/WARP_SIZE*/), cgh);
const size_t n_local_scratch = (GGML_PAD(ncols_x, WARP_SIZE) + WARP_SIZE);
static_assert(SYCL_SOFT_MAX_BLOCK_SIZE == 1024, "These values need to be adjusted.");
cgh.parallel_for(
sycl::nd_range<3>(block_nums * block_dims, block_dims),
[=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(32)]] {
soft_max_f32(x, y, dst, ncols_x, nrows_y, scale, item_ct1,
buf_acc_ct1.get_pointer());
});
});
const uint32_t n_head_kv = nrows_x/nrows_y;
const uint32_t n_head_log2 = 1u << (uint32_t) floorf(log2f((float) n_head_kv));
const float m0 = powf(2.0f, -(max_bias ) / n_head_log2);
const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2);
const size_t local_mem_size = stream->get_device().get_info<sycl::info::device::local_mem_size>();
if (n_local_scratch*sizeof(float) < local_mem_size) {
switch (ncols_x) {
case 32:
soft_max_f32_submitter<true, 32, 32>(x, mask, pos, dst, ncols_x, nrows_y, scale,
max_bias, m0, m1, n_head_log2, block_nums,
block_dims, n_local_scratch, stream);
break;
case 64:
soft_max_f32_submitter<true, 64, 64>(x, mask, pos, dst, ncols_x, nrows_y, scale,
max_bias, m0, m1, n_head_log2, block_nums,
block_dims, n_local_scratch, stream);
break;
case 128:
soft_max_f32_submitter<true, 128, 128>(x, mask, pos, dst, ncols_x, nrows_y, scale,
max_bias, m0, m1, n_head_log2, block_nums,
block_dims, n_local_scratch, stream);
break;
case 256:
soft_max_f32_submitter<true, 256, 256>(x, mask, pos, dst, ncols_x, nrows_y, scale,
max_bias, m0, m1, n_head_log2, block_nums,
block_dims, n_local_scratch, stream);
break;
case 512:
soft_max_f32_submitter<true, 512, 512>(x, mask, pos, dst, ncols_x, nrows_y, scale,
max_bias, m0, m1, n_head_log2, block_nums,
block_dims, n_local_scratch, stream);
break;
case 1024:
soft_max_f32_submitter<true, 1024, 1024>(x, mask, pos, dst, ncols_x, nrows_y, scale,
max_bias, m0, m1, n_head_log2, block_nums,
block_dims, n_local_scratch, stream);
break;
case 2048:
soft_max_f32_submitter<true, 2048, 1024>(x, mask, pos, dst, ncols_x, nrows_y, scale,
max_bias, m0, m1, n_head_log2, block_nums,
block_dims, n_local_scratch, stream);
break;
case 4096:
soft_max_f32_submitter<true, 4096, 1024>(x, mask, pos, dst, ncols_x, nrows_y, scale,
max_bias, m0, m1, n_head_log2, block_nums,
block_dims, n_local_scratch, stream);
break;
default:
soft_max_f32_submitter<true, 0, 0>(x, mask, pos, dst, ncols_x, nrows_y, scale,
max_bias, m0, m1, n_head_log2, block_nums,
block_dims, n_local_scratch, stream);
break;
}
} else {
soft_max_f32_submitter<false, 0, 0>(x, mask, pos, dst, ncols_x, nrows_y, scale,
max_bias, m0, m1, n_head_log2, block_nums,
block_dims, WARP_SIZE, stream);
}
}
template <typename T>
@ -11407,12 +11470,12 @@ static dpct::err0 ggml_sycl_cpy_tensor_2d(void *dst,
dpct::memcpy_direction kind;
char * src_ptr;
if (src->backend == GGML_BACKEND_CPU) {
if (src->backend == GGML_BACKEND_TYPE_CPU) {
kind = dpct::host_to_device;
src_ptr = (char *) src->data;
// GGML_SYCL_DEBUG("ggml_sycl_cpy_tensor_2d GGML_BACKEND_CPU src_ptr %p\n", src_ptr);
} else if (src->backend == GGML_BACKEND_GPU || src->backend == GGML_BACKEND_GPU_SPLIT) {
GGML_ASSERT(src->backend != GGML_BACKEND_GPU_SPLIT || (i1_low == 0 && i1_high == src->ne[1]));
// GGML_SYCL_DEBUG("ggml_sycl_cpy_tensor_2d GGML_BACKEND_TYPE_CPU src_ptr %p\n", src_ptr);
} else if (src->backend == GGML_BACKEND_TYPE_GPU || src->backend == GGML_BACKEND_TYPE_GPU_SPLIT) {
GGML_ASSERT(src->backend != GGML_BACKEND_TYPE_GPU_SPLIT || (i1_low == 0 && i1_high == src->ne[1]));
kind = dpct::device_to_device;
ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) src->extra;
int id;
@ -11846,7 +11909,7 @@ inline void ggml_sycl_op_mul_mat_q(
// the main device has a larger memory buffer to hold the results from all GPUs
// nrows_dst == nrows of the matrix that the dequantize_mul_mat kernel writes into
const int64_t nrows_dst = dst->backend == GGML_BACKEND_GPU && device_id == g_main_device ? ne0 : row_diff;
const int64_t nrows_dst = dst->backend == GGML_BACKEND_TYPE_GPU && device_id == g_main_device ? ne0 : row_diff;
switch (src0->type) {
case GGML_TYPE_Q4_0:
@ -12119,7 +12182,7 @@ inline void ggml_sycl_op_mul_mat_sycl(
// the main device has a larger memory buffer to hold the results from all GPUs
// ldc == nrows of the matrix that cuBLAS writes into
int ldc = dst->backend == GGML_BACKEND_GPU && device_id == g_main_device ? ne0 : row_diff;
int ldc = dst->backend == GGML_BACKEND_TYPE_GPU && device_id == g_main_device ? ne0 : row_diff;
#ifdef GGML_SYCL_F16
bool use_fp16 = true; // TODO(Yu) SYCL capability check
@ -12435,14 +12498,35 @@ inline void ggml_sycl_op_soft_max(const ggml_tensor *src0,
const int64_t ne00 = src0->ne[0];
const int64_t nrows_x = ggml_nrows(src0);
const int64_t nrows_y = src1 ? ggml_nrows(src1) : 1;
const int64_t nrows_y = src0->ne[1];
float scale = 1.0f;
memcpy(&scale, dst->op_params, sizeof(float));
float max_bias = 0.0f;
soft_max_f32_sycl(src0_dd, src1 ? src1_dd : nullptr, dst_dd, ne00, nrows_x, nrows_y, scale, main_stream);
memcpy(&scale, dst->op_params + 0, sizeof(float));
memcpy(&max_bias, dst->op_params + 1, sizeof(float));
(void) dst;
// positions tensor
float * src2_dd = nullptr;
sycl_pool_alloc<float> src2_f;
ggml_tensor * src2 = dst->src[2];
const bool use_src2 = src2 != nullptr;
if (use_src2) {
const bool src2_on_device = src2->backend == GGML_BACKEND_TYPE_GPU;
if (src2_on_device) {
ggml_tensor_extra_gpu * src2_extra = (ggml_tensor_extra_gpu *) src2->extra;
src2_dd = (float *) src2_extra->data_device[g_main_device];
} else {
src2_dd = src2_f.alloc(ggml_nelements(src2));
SYCL_CHECK(ggml_sycl_cpy_tensor_2d(src2_dd, src2, 0, 0, 0, 1, main_stream));
}
}
soft_max_f32_sycl(src0_dd, src1 ? src1_dd : nullptr, src2_dd, dst_dd, ne00,
nrows_x, nrows_y, scale, max_bias, main_stream);
}
inline void ggml_sycl_op_scale(const ggml_tensor *src0, const ggml_tensor *src1,
@ -12501,16 +12585,16 @@ static void ggml_sycl_op_flatten(const ggml_tensor *src0,
const bool use_src1 = src1 != nullptr;
const int64_t nrows1 = use_src1 ? ggml_nrows(src1) : 1;
GGML_ASSERT(!use_src1 || src1->backend != GGML_BACKEND_GPU_SPLIT);
GGML_ASSERT( dst->backend != GGML_BACKEND_GPU_SPLIT);
GGML_ASSERT(!use_src1 || src1->backend != GGML_BACKEND_TYPE_GPU_SPLIT);
GGML_ASSERT( dst->backend != GGML_BACKEND_TYPE_GPU_SPLIT);
ggml_tensor_extra_gpu * src0_extra = (ggml_tensor_extra_gpu *) src0->extra;
ggml_tensor_extra_gpu * src1_extra = use_src1 ? (ggml_tensor_extra_gpu *) src1->extra : nullptr;
ggml_tensor_extra_gpu * dst_extra = (ggml_tensor_extra_gpu *) dst->extra;
const bool src0_on_device = src0->backend == GGML_BACKEND_GPU || src0->backend == GGML_BACKEND_GPU_SPLIT;
const bool src1_on_device = use_src1 && src1->backend == GGML_BACKEND_GPU;
const bool dst_on_device = dst->backend == GGML_BACKEND_GPU;
const bool src0_on_device = src0->backend == GGML_BACKEND_TYPE_GPU || src0->backend == GGML_BACKEND_TYPE_GPU_SPLIT;
const bool src1_on_device = use_src1 && src1->backend == GGML_BACKEND_TYPE_GPU;
const bool dst_on_device = dst->backend == GGML_BACKEND_TYPE_GPU;
// dd = data device
float * src0_ddf = nullptr;
@ -12565,7 +12649,7 @@ static void ggml_sycl_op_flatten(const ggml_tensor *src0,
main_stream->memcpy(dst->data, dst_ddf, ggml_nbytes(dst))));
}
if (dst->backend == GGML_BACKEND_CPU) {
if (dst->backend == GGML_BACKEND_TYPE_CPU) {
SYCL_CHECK(CHECK_TRY_ERROR(
dpct::get_current_device().queues_wait_and_throw()));
}
@ -12640,8 +12724,8 @@ static void ggml_sycl_op_mul_mat(const ggml_tensor *src0,
const int nb2 = dst->nb[2];
const int nb3 = dst->nb[3];
GGML_ASSERT(dst->backend != GGML_BACKEND_GPU_SPLIT);
GGML_ASSERT(src1->backend != GGML_BACKEND_GPU_SPLIT);
GGML_ASSERT(dst->backend != GGML_BACKEND_TYPE_GPU_SPLIT);
GGML_ASSERT(src1->backend != GGML_BACKEND_TYPE_GPU_SPLIT);
GGML_ASSERT(ne12 >= ne02 && ne12 % ne02 == 0);
@ -12656,13 +12740,13 @@ static void ggml_sycl_op_mul_mat(const ggml_tensor *src0,
ggml_tensor_extra_gpu * src1_extra = (ggml_tensor_extra_gpu *) src1->extra;
ggml_tensor_extra_gpu * dst_extra = (ggml_tensor_extra_gpu *) dst->extra;
const bool src0_on_device = src0->backend == GGML_BACKEND_GPU || src0->backend == GGML_BACKEND_GPU_SPLIT;
const bool src0_on_device = src0->backend == GGML_BACKEND_TYPE_GPU || src0->backend == GGML_BACKEND_TYPE_GPU_SPLIT;
const bool src0_is_contiguous = ggml_is_contiguous(src0);
const bool src1_is_contiguous = ggml_is_contiguous(src1);
int64_t src1_padded_col_size = GGML_PAD(ne10, MATRIX_ROW_PADDING);
const bool split = src0->backend == GGML_BACKEND_GPU_SPLIT;
const bool split = src0->backend == GGML_BACKEND_TYPE_GPU_SPLIT;
GGML_ASSERT(!(split && ne02 > 1));
GGML_ASSERT(!(split && ne03 > 1));
GGML_ASSERT(!(split && ne02 < ne12));
@ -12717,8 +12801,8 @@ static void ggml_sycl_op_mul_mat(const ggml_tensor *src0,
used_devices++;
const bool src1_on_device = src1->backend == GGML_BACKEND_GPU && id == g_main_device_index;
const bool dst_on_device = dst->backend == GGML_BACKEND_GPU && id == g_main_device_index;
const bool src1_on_device = src1->backend == GGML_BACKEND_TYPE_GPU && id == g_main_device_index;
const bool dst_on_device = dst->backend == GGML_BACKEND_TYPE_GPU && id == g_main_device_index;
ggml_sycl_set_device(get_device_id_by_index(id));
const dpct::queue_ptr stream = g_syclStreams[id][0];
@ -12782,8 +12866,8 @@ static void ggml_sycl_op_mul_mat(const ggml_tensor *src0,
continue;
}
const bool src1_on_device = src1->backend == GGML_BACKEND_GPU && id == g_main_device_index;
const bool dst_on_device = dst->backend == GGML_BACKEND_GPU && id == g_main_device_index;
const bool src1_on_device = src1->backend == GGML_BACKEND_TYPE_GPU && id == g_main_device_index;
const bool dst_on_device = dst->backend == GGML_BACKEND_TYPE_GPU && id == g_main_device_index;
const int64_t row_diff = row_high[id] - row_low[id];
ggml_sycl_set_device(get_device_id_by_index(id));
@ -12809,12 +12893,12 @@ static void ggml_sycl_op_mul_mat(const ggml_tensor *src0,
// the main device memory buffer can be on VRAM scratch, with space for all partial results
// in that case an offset on dst_ddf_i is needed
if (dst->backend == GGML_BACKEND_GPU && id == g_main_device_index) {
if (dst->backend == GGML_BACKEND_TYPE_GPU && id == g_main_device_index) {
dst_dd_i += row_low[id]; // offset is 0 if no tensor split
}
// copy src0, src1 to device if necessary
if (src1->backend == GGML_BACKEND_GPU && src1_is_contiguous) {
if (src1->backend == GGML_BACKEND_TYPE_GPU && src1_is_contiguous) {
if (id != g_main_device_index) {
if (convert_src1_to_q8_1) {
char * src1_ddq_i_source = src1_ddq[g_main_device_index] + src1_ddq_i_offset;
@ -12830,14 +12914,14 @@ static void ggml_sycl_op_mul_mat(const ggml_tensor *src0,
src1_ncols * ne10 * sizeof(float))));
}
}
} else if (src1->backend == GGML_BACKEND_CPU || (src1_on_device && !src1_is_contiguous)) {
} else if (src1->backend == GGML_BACKEND_TYPE_CPU || (src1_on_device && !src1_is_contiguous)) {
SYCL_CHECK(ggml_sycl_cpy_tensor_2d(
src1_ddf_i, src1, i03, i02, src1_col_0, src1_col_0+src1_ncols, stream));
} else {
GGML_ASSERT(false);
}
if (convert_src1_to_q8_1 && (src1->backend == GGML_BACKEND_CPU || !src1_is_contiguous)) {
if (convert_src1_to_q8_1 && (src1->backend == GGML_BACKEND_TYPE_CPU || !src1_is_contiguous)) {
quantize_row_q8_1_sycl(src1_ddf_i, src1_ddq_i, ne10, src1_ncols, src1_padded_col_size, stream);
/*
DPCT1010:92: SYCL uses exceptions to report errors and does
@ -12867,10 +12951,10 @@ static void ggml_sycl_op_mul_mat(const ggml_tensor *src0,
if (!dst_on_device) {
void * dst_off_device;
dpct::memcpy_direction kind;
if (dst->backend == GGML_BACKEND_CPU) {
if (dst->backend == GGML_BACKEND_TYPE_CPU) {
dst_off_device = dst->data;
kind = dpct::device_to_host;
} else if (dst->backend == GGML_BACKEND_GPU) {
} else if (dst->backend == GGML_BACKEND_TYPE_GPU) {
dst_off_device = dst_extra->data_device[g_main_device_index];
kind = dpct::device_to_device;
} else {
@ -12954,7 +13038,7 @@ static void ggml_sycl_op_mul_mat(const ggml_tensor *src0,
}
}
if (dst->backend == GGML_BACKEND_CPU) {
if (dst->backend == GGML_BACKEND_TYPE_CPU) {
SYCL_CHECK(ggml_sycl_set_device(g_main_device));
SYCL_CHECK(CHECK_TRY_ERROR(
dpct::get_current_device().queues_wait_and_throw()));
@ -13091,7 +13175,7 @@ static void ggml_sycl_mul_mat_vec_p021(const ggml_tensor *src0,
const ggml_tensor *src1,
ggml_tensor *dst) try {
GGML_ASSERT(ggml_is_permuted(src0) && ggml_is_permuted(src1));
GGML_ASSERT(src0->backend != GGML_BACKEND_GPU_SPLIT);
GGML_ASSERT(src0->backend != GGML_BACKEND_TYPE_GPU_SPLIT);
GGML_ASSERT(src0->nb[0] <= src0->nb[1] && src0->nb[2] <= src0->nb[3]); // 0213 permutation
GGML_ASSERT(src1->nb[0] <= src1->nb[1] && src1->nb[2] <= src1->nb[3]); // 0213 permutation
GGML_ASSERT(src0->type == GGML_TYPE_F16);
@ -13129,7 +13213,7 @@ static void ggml_sycl_mul_mat_vec_nc(const ggml_tensor *src0,
GGML_ASSERT(!ggml_is_transposed(src0));
GGML_ASSERT(!ggml_is_transposed(src1));
GGML_ASSERT(!ggml_is_permuted(src0));
GGML_ASSERT(src0->backend != GGML_BACKEND_GPU_SPLIT);
GGML_ASSERT(src0->backend != GGML_BACKEND_TYPE_GPU_SPLIT);
GGML_ASSERT(src0->type == GGML_TYPE_F16);
GGML_ASSERT(src1->type == GGML_TYPE_F32);
@ -13196,7 +13280,7 @@ static void ggml_sycl_mul_mat_mat_batched_sycl(const ggml_tensor *src0,
GGML_ASSERT(!ggml_is_transposed(src0));
GGML_ASSERT(!ggml_is_transposed(src1));
GGML_ASSERT(src0->backend != GGML_BACKEND_GPU_SPLIT);
GGML_ASSERT(src0->backend != GGML_BACKEND_TYPE_GPU_SPLIT);
GGML_ASSERT(src0->type == GGML_TYPE_F16);
GGML_ASSERT(src1->type == GGML_TYPE_F32);
@ -13372,11 +13456,11 @@ catch (sycl::exception const &exc) {
static void ggml_sycl_mul_mat(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
const bool all_on_device =
(src0->backend == GGML_BACKEND_GPU || src0->backend == GGML_BACKEND_GPU_SPLIT) &&
(src1->backend == GGML_BACKEND_GPU) &&
( dst->backend == GGML_BACKEND_GPU);
(src0->backend == GGML_BACKEND_TYPE_GPU || src0->backend == GGML_BACKEND_TYPE_GPU_SPLIT) &&
(src1->backend == GGML_BACKEND_TYPE_GPU) &&
( dst->backend == GGML_BACKEND_TYPE_GPU);
const bool split = src0->backend == GGML_BACKEND_GPU_SPLIT;
const bool split = src0->backend == GGML_BACKEND_TYPE_GPU_SPLIT;
int64_t min_compute_capability = INT_MAX;
for (int64_t id = 0; id < g_device_count; ++id) {
@ -13505,7 +13589,7 @@ static void ggml_sycl_mul_mat_id_sycl(ggml_tensor * dst) {
GGML_ASSERT(!ggml_is_transposed(src00));
GGML_ASSERT(!ggml_is_transposed(src1));
GGML_ASSERT(src00->backend != GGML_BACKEND_GPU_SPLIT);
GGML_ASSERT(src00->backend != GGML_BACKEND_TYPE_GPU_SPLIT);
GGML_ASSERT(src1->type == GGML_TYPE_F32);
GGML_TENSOR_LOCALS(int64_t, ne0, src00, ne);
@ -13643,7 +13727,7 @@ static void ggml_sycl_mul_mat_id(const ggml_tensor *src0,
const dpct::queue_ptr stream = g_syclStreams[g_main_device_index][0];
if (ids->backend == GGML_BACKEND_GPU) {
if (ids->backend == GGML_BACKEND_TYPE_GPU) {
const char * ids_dev = (const char *)((const ggml_tensor_extra_gpu *)ids->extra)->data_device[g_main_device_index];
SYCL_CHECK(CHECK_TRY_ERROR(
stream->memcpy(ids_host.data(), ids_dev, ggml_nbytes(ids))));
@ -13661,20 +13745,20 @@ static void ggml_sycl_mul_mat_id(const ggml_tensor *src0,
ggml_tensor src1_row = *src1;
ggml_tensor dst_row = *dst;
src1_row.backend = GGML_BACKEND_GPU;
dst_row.backend = GGML_BACKEND_GPU;
src1_row.backend = GGML_BACKEND_TYPE_GPU;
dst_row.backend = GGML_BACKEND_TYPE_GPU;
src1_row.extra = &src1_row_extra;
dst_row.extra = &dst_row_extra;
char * src1_original = src1->backend == GGML_BACKEND_CPU ?
char * src1_original = src1->backend == GGML_BACKEND_TYPE_CPU ?
(char *) src1->data : (char *) src1_extra->data_device[g_main_device_index];
char * dst_original = dst->backend == GGML_BACKEND_CPU ?
char * dst_original = dst->backend == GGML_BACKEND_TYPE_CPU ?
(char *) dst->data : (char *) dst_extra->data_device[g_main_device_index];
if (src1->ne[1] == 1) {
GGML_ASSERT(src1->backend == GGML_BACKEND_GPU);
GGML_ASSERT(dst->backend == GGML_BACKEND_GPU);
GGML_ASSERT(src1->backend == GGML_BACKEND_TYPE_GPU);
GGML_ASSERT(dst->backend == GGML_BACKEND_TYPE_GPU);
for (int64_t i01 = 0; i01 < ids->ne[1]; i01++) {
//int32_t row_id;
@ -13756,7 +13840,7 @@ static void ggml_sycl_mul_mat_id(const ggml_tensor *src0,
}
}
if (dst->backend == GGML_BACKEND_CPU) {
if (dst->backend == GGML_BACKEND_TYPE_CPU) {
SYCL_CHECK(CHECK_TRY_ERROR(stream->wait()));
}
}
@ -13779,8 +13863,8 @@ static void ggml_sycl_cpy(const ggml_tensor *src0, const ggml_tensor *src1,
const int64_t ne = ggml_nelements(src0);
GGML_ASSERT(ne == ggml_nelements(src1));
GGML_ASSERT(src0->backend == GGML_BACKEND_GPU);
GGML_ASSERT(src1->backend == GGML_BACKEND_GPU);
GGML_ASSERT(src0->backend == GGML_BACKEND_TYPE_GPU);
GGML_ASSERT(src1->backend == GGML_BACKEND_TYPE_GPU);
GGML_ASSERT(ggml_nbytes(src0) <= INT_MAX);
GGML_ASSERT(ggml_nbytes(src1) <= INT_MAX);
@ -13887,17 +13971,17 @@ void ggml_sycl_transform_tensor(void *data, struct ggml_tensor *tensor) try {
memset(extra, 0, sizeof(*extra));
for (int64_t id = 0; id < g_device_count; ++id) {
if (backend == GGML_BACKEND_GPU && id != g_main_device_index) {
if (backend == GGML_BACKEND_TYPE_GPU && id != g_main_device_index) {
continue;
}
ggml_sycl_set_device(get_device_id_by_index(id));
const dpct::queue_ptr stream = g_syclStreams[id][0];
int64_t row_low, row_high;
if (backend == GGML_BACKEND_GPU) {
if (backend == GGML_BACKEND_TYPE_GPU) {
row_low = 0;
row_high = nrows;
} else if (backend == GGML_BACKEND_GPU_SPLIT) {
} else if (backend == GGML_BACKEND_TYPE_GPU_SPLIT) {
const int64_t rounding = get_row_rounding(tensor->type);
row_low = id == 0 ? 0 : nrows*g_tensor_split[id];
@ -13946,7 +14030,7 @@ void ggml_sycl_transform_tensor(void *data, struct ggml_tensor *tensor) try {
extra->data_device[id] = buf;
if (backend == GGML_BACKEND_GPU_SPLIT) {
if (backend == GGML_BACKEND_TYPE_GPU_SPLIT) {
for (int64_t is = 0; is < MAX_STREAMS; ++is) {
SYCL_CHECK(CHECK_TRY_ERROR(extra->events[id][is] =
new sycl::event()));
@ -13963,7 +14047,7 @@ catch (sycl::exception const &exc) {
}
void ggml_sycl_free_data(struct ggml_tensor *tensor) try {
if (!tensor || !tensor->extra || (tensor->backend != GGML_BACKEND_GPU && tensor->backend != GGML_BACKEND_GPU_SPLIT) ) {
if (!tensor || !tensor->extra || (tensor->backend != GGML_BACKEND_TYPE_GPU && tensor->backend != GGML_BACKEND_TYPE_GPU_SPLIT) ) {
return;
}
@ -14016,15 +14100,15 @@ static void ggml_sycl_assign_buffers_impl(struct ggml_tensor *tensor,
return;
}
tensor->backend = GGML_BACKEND_GPU;
tensor->backend = GGML_BACKEND_TYPE_GPU;
if (tensor->src[0] != nullptr && tensor->src[0]->backend == GGML_BACKEND_CPU) {
if (tensor->src[0] != nullptr && tensor->src[0]->backend == GGML_BACKEND_TYPE_CPU) {
const ggml_op src0_op = tensor->src[0]->op;
if (src0_op == GGML_OP_RESHAPE || src0_op == GGML_OP_TRANSPOSE || src0_op == GGML_OP_VIEW || src0_op == GGML_OP_PERMUTE) {
ggml_sycl_assign_buffers_impl(tensor->src[0], scratch, force_inplace, no_alloc);
}
}
if (tensor->op == GGML_OP_CPY && tensor->src[1]->backend == GGML_BACKEND_CPU) {
if (tensor->op == GGML_OP_CPY && tensor->src[1]->backend == GGML_BACKEND_TYPE_CPU) {
ggml_sycl_assign_buffers_impl(tensor->src[1], scratch, force_inplace, no_alloc);
}
@ -14042,7 +14126,7 @@ static void ggml_sycl_assign_buffers_impl(struct ggml_tensor *tensor,
SYCL_CHECK(ggml_sycl_set_device(g_main_device));
const dpct::queue_ptr stream = g_syclStreams[g_main_device_index][0];
if (inplace && (tensor->src[0]->backend == GGML_BACKEND_GPU || tensor->src[0]->backend == GGML_BACKEND_GPU_SPLIT)) {
if (inplace && (tensor->src[0]->backend == GGML_BACKEND_TYPE_GPU || tensor->src[0]->backend == GGML_BACKEND_TYPE_GPU_SPLIT)) {
ggml_tensor_extra_gpu * src0_extra = (ggml_tensor_extra_gpu * ) tensor->src[0]->extra;
char * src0_ddc = (char *) src0_extra->data_device[g_main_device_index];
size_t offset = 0;
@ -14111,7 +14195,7 @@ void ggml_sycl_assign_scratch_offset(struct ggml_tensor *tensor,
const bool inplace = tensor->view_src != nullptr;
if (inplace && (tensor->view_src->backend == GGML_BACKEND_GPU || tensor->view_src->backend == GGML_BACKEND_GPU_SPLIT)) {
if (inplace && (tensor->view_src->backend == GGML_BACKEND_TYPE_GPU || tensor->view_src->backend == GGML_BACKEND_TYPE_GPU_SPLIT)) {
ggml_tensor_extra_gpu * src0_extra = (ggml_tensor_extra_gpu * ) tensor->view_src->extra;
char * src0_ddc = (char *) src0_extra->data_device[g_main_device_index];
size_t view_offset = 0;
@ -14132,7 +14216,7 @@ catch (sycl::exception const &exc) {
}
void ggml_sycl_copy_to_device(struct ggml_tensor *tensor) try {
GGML_ASSERT(tensor->backend == GGML_BACKEND_GPU);
GGML_ASSERT(tensor->backend == GGML_BACKEND_TYPE_GPU);
GGML_ASSERT(ggml_is_contiguous(tensor));
ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) tensor->extra;
@ -14219,9 +14303,9 @@ bool ggml_sycl_compute_forward(struct ggml_compute_params * params, struct ggml_
if (!g_sycl_loaded) return false;
ggml_sycl_func_t func;
const bool any_on_device = tensor->backend == GGML_BACKEND_GPU
|| (tensor->src[0] != nullptr && (tensor->src[0]->backend == GGML_BACKEND_GPU || tensor->src[0]->backend == GGML_BACKEND_GPU_SPLIT))
|| (tensor->src[1] != nullptr && tensor->src[1]->backend == GGML_BACKEND_GPU);
const bool any_on_device = tensor->backend == GGML_BACKEND_TYPE_GPU
|| (tensor->src[0] != nullptr && (tensor->src[0]->backend == GGML_BACKEND_TYPE_GPU || tensor->src[0]->backend == GGML_BACKEND_TYPE_GPU_SPLIT))
|| (tensor->src[1] != nullptr && tensor->src[1]->backend == GGML_BACKEND_TYPE_GPU);
if (!any_on_device && tensor->op != GGML_OP_MUL_MAT && tensor->op != GGML_OP_MUL_MAT_ID) {
return false;
@ -14359,14 +14443,14 @@ bool ggml_sycl_compute_forward(struct ggml_compute_params * params, struct ggml_
return false;
}
if (tensor->src[0] != nullptr && tensor->src[0]->backend == GGML_BACKEND_GPU_SPLIT) {
if (tensor->src[0] != nullptr && tensor->src[0]->backend == GGML_BACKEND_TYPE_GPU_SPLIT) {
ggml_sycl_set_peer_access(tensor->src[1]->ne[1]);
}
if (params->ith != 0) {
return true;
}
if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
if (params->type == GGML_TASK_TYPE_INIT || params->type == GGML_TASK_TYPE_FINALIZE) {
return true;
}
func(tensor->src[0], tensor->src[1], tensor);
@ -14517,7 +14601,7 @@ static void ggml_backend_sycl_buffer_init_tensor(ggml_backend_buffer_t buffer,
extra->data_device[ctx->device] = tensor->data;
tensor->backend = GGML_BACKEND_GPU;
tensor->backend = GGML_BACKEND_TYPE_GPU;
tensor->extra = extra;
if (ggml_is_quantized(tensor->type)) {
@ -14548,7 +14632,7 @@ static void ggml_backend_sycl_buffer_set_tensor(ggml_backend_buffer_t buffer,
ggml_tensor *tensor,
const void *data, size_t offset,
size_t size) try {
GGML_ASSERT(tensor->backend == GGML_BACKEND_GPU);
GGML_ASSERT(tensor->backend == GGML_BACKEND_TYPE_GPU);
ggml_backend_sycl_buffer_context * ctx = ( ggml_backend_sycl_buffer_context *)buffer->context;
@ -14573,7 +14657,7 @@ static void ggml_backend_sycl_buffer_get_tensor(ggml_backend_buffer_t buffer,
const ggml_tensor *tensor,
void *data, size_t offset,
size_t size) try {
GGML_ASSERT(tensor->backend == GGML_BACKEND_GPU);
GGML_ASSERT(tensor->backend == GGML_BACKEND_TYPE_GPU);
ggml_backend_sycl_buffer_context * ctx = ( ggml_backend_sycl_buffer_context *)buffer->context;
@ -14809,7 +14893,7 @@ static void ggml_backend_sycl_set_tensor_async(ggml_backend_t backend,
ggml_backend_sycl_context * sycl_ctx = (ggml_backend_sycl_context *)backend->context;
GGML_ASSERT(tensor->buffer->buft == ggml_backend_sycl_buffer_type(sycl_ctx->device) && "unsupported buffer type");
GGML_ASSERT(tensor->backend == GGML_BACKEND_GPU);
GGML_ASSERT(tensor->backend == GGML_BACKEND_TYPE_GPU);
SYCL_CHECK(CHECK_TRY_ERROR(g_syclStreams[sycl_ctx->device][0]->memcpy(
(char *)tensor->data + offset, data, size)));
@ -14827,7 +14911,7 @@ static void ggml_backend_sycl_get_tensor_async(ggml_backend_t backend,
ggml_backend_sycl_context * sycl_ctx = (ggml_backend_sycl_context *)backend->context;
GGML_ASSERT(tensor->buffer->buft == ggml_backend_sycl_buffer_type(sycl_ctx->device) && "unsupported buffer type");
GGML_ASSERT(tensor->backend == GGML_BACKEND_GPU);
GGML_ASSERT(tensor->backend == GGML_BACKEND_TYPE_GPU);
SYCL_CHECK(CHECK_TRY_ERROR(g_syclStreams[sycl_ctx->device][0]->memcpy(
data, (const char *)tensor->data + offset, size)));
@ -14880,7 +14964,7 @@ static bool ggml_backend_sycl_graph_compute(ggml_backend_t backend, ggml_cgraph
ggml_sycl_set_main_device(sycl_ctx->device);
ggml_compute_params params = {};
params.type = GGML_TASK_COMPUTE;
params.type = GGML_TASK_TYPE_COMPUTE;
params.ith = 0;
for (int i = 0; i < cgraph->n_nodes; i++) {
ggml_tensor * node = cgraph->nodes[i];
@ -14888,13 +14972,13 @@ static bool ggml_backend_sycl_graph_compute(ggml_backend_t backend, ggml_cgraph
if (node->op == GGML_OP_RESHAPE || node->op == GGML_OP_TRANSPOSE || node->op == GGML_OP_VIEW || node->op == GGML_OP_PERMUTE)
continue;
assert(node->backend == GGML_BACKEND_GPU);
assert(node->backend == GGML_BACKEND_TYPE_GPU);
assert(node->buffer->buft == ggml_backend_sycl_buffer_type(sycl_ctx->device));
assert(node->extra != nullptr);
for (int j = 0; j < GGML_MAX_SRC; j++) {
if (node->src[j] != nullptr) {
assert(node->src[j]->backend == GGML_BACKEND_GPU);
assert(node->src[j]->backend == GGML_BACKEND_TYPE_GPU);
assert(node->src[j]->buffer->buft == ggml_backend_sycl_buffer_type(sycl_ctx->device));
assert(node->src[j]->extra != nullptr);
}
@ -15078,6 +15162,11 @@ static ggml_backend_i ggml_backend_sycl_interface = {
/* .supports_op = */ ggml_backend_sycl_supports_op,
};
static ggml_guid_t ggml_backend_sycl_guid() {
static ggml_guid guid = { 0x58, 0x05, 0x13, 0x8f, 0xcd, 0x3a, 0x61, 0x9d, 0xe7, 0xcd, 0x98, 0xa9, 0x03, 0xfd, 0x7c, 0x53 };
return &guid;
}
ggml_backend_t ggml_backend_sycl_init(int device) {
ggml_init_sycl(); // TODO: remove from ggml.c
@ -15095,6 +15184,7 @@ ggml_backend_t ggml_backend_sycl_init(int device) {
};
ggml_backend_t sycl_backend = new ggml_backend {
/* .guid = */ ggml_backend_sycl_guid(),
/* .interface = */ ggml_backend_sycl_interface,
/* .context = */ ctx
};
@ -15103,7 +15193,7 @@ ggml_backend_t ggml_backend_sycl_init(int device) {
}
bool ggml_backend_is_sycl(ggml_backend_t backend) {
return backend->iface.get_name == ggml_backend_sycl_name;
return backend != NULL && ggml_guid_matches(backend->guid, ggml_backend_sycl_guid());
}
static ggml_backend_t ggml_backend_reg_sycl_init(const char * params, void * user_data) {

116
ggml-vulkan.cpp
View file

@ -1106,7 +1106,9 @@ void ggml_vk_instance_init() {
const std::vector<vk::ExtensionProperties> instance_extensions = vk::enumerateInstanceExtensionProperties();
const bool validation_ext = ggml_vk_instance_validation_ext_available(instance_extensions);
#ifdef __APPLE__
const bool portability_enumeration_ext = ggml_vk_instance_portability_enumeration_ext_available(instance_extensions);
#endif
std::vector<const char*> layers;
@ -1117,13 +1119,17 @@ void ggml_vk_instance_init() {
if (validation_ext) {
extensions.push_back("VK_EXT_validation_features");
}
#ifdef __APPLE__
if (portability_enumeration_ext) {
extensions.push_back("VK_KHR_portability_enumeration");
}
#endif
vk::InstanceCreateInfo instance_create_info(vk::InstanceCreateFlags{}, &app_info, layers, extensions);
#ifdef __APPLE__
if (portability_enumeration_ext) {
instance_create_info.flags |= vk::InstanceCreateFlagBits::eEnumeratePortabilityKHR;
}
#endif
std::vector<vk::ValidationFeatureEnableEXT> features_enable;
vk::ValidationFeaturesEXT validation_features;
@ -2320,8 +2326,8 @@ static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context * su
src1_uma = d_Qy != nullptr;
}
const bool load_x = src0->backend != GGML_BACKEND_GPU && !src0_uma;
const bool load_y = src1->backend != GGML_BACKEND_GPU && !src1_uma;
const bool load_x = src0->backend != GGML_BACKEND_TYPE_GPU && !src0_uma;
const bool load_y = src1->backend != GGML_BACKEND_TYPE_GPU && !src1_uma;
const bool x_non_contig = !load_x && !ggml_vk_dim01_contiguous(src0);
const bool y_non_contig = !load_y && !ggml_vk_dim01_contiguous(src1);
@ -2453,7 +2459,7 @@ static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context * su
// compute
ggml_vk_matmul(ctx, subctx, *pipeline, { d_X, x_buf_offset, x_sz * ne02 * ne03 }, { d_Y, y_buf_offset, y_sz * ne12 * ne13 }, { d_D, d_buf_offset, d_sz * ne12 * ne13 }, { ctx->prealloc_split_k, 0, d_sz * ne12 * ne13 * split_k }, ne01, ne11, ne10, ne10, ne10, ne01, split_k, ne12*ne13, ne02, ne12, r2, r3, stride_batch_x, stride_batch_y, ne20*ne21); // NOLINT
if (dst->backend == GGML_BACKEND_CPU) {
if (dst->backend == GGML_BACKEND_TYPE_CPU) {
// copy dst to host
float * d = (float *) ((char *) dst->data);
ggml_vk_buffer_read_async(ctx, subctx, d_D, 0, d, sizeof(float) * d_ne * ne12 * ne13);
@ -2506,8 +2512,8 @@ static void ggml_vk_mul_mat_vec_q_f16(ggml_backend_vk_context * ctx, vk_context
src1_uma = d_Qy != nullptr;
}
const bool load_x = src0->backend != GGML_BACKEND_GPU && !src0_uma;
const bool load_y = src1->backend != GGML_BACKEND_GPU && !src1_uma;
const bool load_x = src0->backend != GGML_BACKEND_TYPE_GPU && !src0_uma;
const bool load_y = src1->backend != GGML_BACKEND_TYPE_GPU && !src1_uma;
const bool x_non_contig = !load_x && !ggml_vk_dim01_contiguous(src0);
const bool y_non_contig = !load_y && !ggml_vk_dim01_contiguous(src1);
@ -2630,7 +2636,7 @@ static void ggml_vk_mul_mat_vec_q_f16(ggml_backend_vk_context * ctx, vk_context
ggml_vk_sync_buffers(subctx);
ggml_vk_dispatch_pipeline(ctx, subctx, *dmmv, { { d_X, x_offset, x_sz }, { d_Y, y_buffer_offset, y_sz + y_shader_offset }, { d_D, d_buffer_offset, d_sz + d_shader_offset } }, 3 * sizeof(int), &pc, { (uint32_t)ne01, 1, 1});
if (dst->backend == GGML_BACKEND_CPU) {
if (dst->backend == GGML_BACKEND_TYPE_CPU) {
// copy dst to host
float * d = (float *) ((char *) dst->data + i12*nb2 + i13*nb3);
ggml_vk_sync_buffers(subctx);
@ -2647,7 +2653,7 @@ static void ggml_vk_mul_mat_vec_p021_f16_f32(ggml_backend_vk_context * ctx, vk_c
std::cerr << "), (" << dst << ", name=" << dst->name << ", type=" << dst->type << ", backend=" << dst->backend << ", ne0=" << dst->ne[0] << ", ne1=" << dst->ne[1] << ", ne2=" << dst->ne[2] << ", ne3=" << dst->ne[3] << ", nb0=" << dst->nb[0] << ", nb1=" << dst->nb[1] << ", nb2=" << dst->nb[2] << ", nb3=" << dst->nb[3] << "),)" << std::endl;
#endif
GGML_ASSERT(ggml_is_permuted(src0) && ggml_is_permuted(src1));
GGML_ASSERT(src0->backend == GGML_BACKEND_GPU);
GGML_ASSERT(src0->backend == GGML_BACKEND_TYPE_GPU);
GGML_ASSERT(src0->nb[0] <= src0->nb[1] && src0->nb[2] <= src0->nb[3]); // NOLINT
GGML_ASSERT(src1->nb[0] <= src1->nb[1] && src1->nb[2] <= src1->nb[3]); // NOLINT
GGML_ASSERT(src0->type == GGML_TYPE_F16);
@ -2679,7 +2685,7 @@ static void ggml_vk_mul_mat_vec_p021_f16_f32(ggml_backend_vk_context * ctx, vk_c
src1_uma = d_Qy != nullptr;
}
const bool load_y = src1->backend != GGML_BACKEND_GPU && !src1_uma;
const bool load_y = src1->backend != GGML_BACKEND_TYPE_GPU && !src1_uma;
const uint64_t x_ne = ne00 * ne01 * ne02;
const uint64_t y_ne = ne10 * ne11 * ne12;
@ -2721,7 +2727,7 @@ static void ggml_vk_mul_mat_vec_p021_f16_f32(ggml_backend_vk_context * ctx, vk_c
ggml_vk_sync_buffers(subctx);
ggml_vk_dispatch_pipeline(ctx, subctx, ctx->pipeline_mul_mat_vec_p021_f16_f32, { { d_Qx, qx_buf_offset, qx_sz }, { d_Qy, qy_buffer_offset, qy_sz + qy_shader_offset }, { d_D, d_buffer_offset, d_sz + d_shader_offset } }, 6 * sizeof(uint32_t), &pc, { 1, (uint32_t)ne01, (uint32_t)ne12 });
if (dst->backend == GGML_BACKEND_CPU) {
if (dst->backend == GGML_BACKEND_TYPE_CPU) {
// copy dst to host
float * d = (float *) dst->data;
ggml_vk_sync_buffers(subctx);
@ -2738,7 +2744,7 @@ static void ggml_vk_mul_mat_vec_nc_f16_f32(ggml_backend_vk_context * ctx, vk_con
GGML_ASSERT(!ggml_is_transposed(src0));
GGML_ASSERT(!ggml_is_transposed(src1));
GGML_ASSERT(!ggml_is_permuted(src0));
GGML_ASSERT(src0->backend == GGML_BACKEND_GPU);
GGML_ASSERT(src0->backend == GGML_BACKEND_TYPE_GPU);
GGML_ASSERT(src0->type == GGML_TYPE_F16);
GGML_ASSERT(src1->type == GGML_TYPE_F32);
@ -2771,7 +2777,7 @@ static void ggml_vk_mul_mat_vec_nc_f16_f32(ggml_backend_vk_context * ctx, vk_con
src1_uma = d_Qy != nullptr;
}
const bool load_y = src1->backend != GGML_BACKEND_GPU && !src1_uma;
const bool load_y = src1->backend != GGML_BACKEND_TYPE_GPU && !src1_uma;
const uint64_t d_ne = ne01 * ne11 * ne12;
@ -2814,7 +2820,7 @@ static void ggml_vk_mul_mat_vec_nc_f16_f32(ggml_backend_vk_context * ctx, vk_con
ggml_vk_sync_buffers(subctx);
ggml_vk_dispatch_pipeline(ctx, subctx, ctx->pipeline_mul_mat_vec_nc_f16_f32, { { d_Qx, qx_buf_offset, qx_sz }, { d_Qy, qy_buffer_offset, qy_sz + qy_shader_offset }, { d_D, d_buffer_offset, d_sz + d_shader_offset } }, 7 * sizeof(uint32_t), &pc, { 1, (uint32_t)ne01, (uint32_t)ne12 });
if (dst->backend == GGML_BACKEND_CPU) {
if (dst->backend == GGML_BACKEND_TYPE_CPU) {
// copy dst to host
float * d = (float *) dst->data;
ggml_vk_sync_buffers(subctx);
@ -2832,7 +2838,7 @@ static bool ggml_vk_can_mul_mat(const ggml_tensor * src0, const ggml_tensor * sr
return (src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16 || ggml_is_quantized(src0->type)) &&
(src1->type == GGML_TYPE_F32 || src1->type == GGML_TYPE_F16 || ggml_is_quantized(src1->type)) &&
dst->type == GGML_TYPE_F32 &&
((ne0 >= 32 && ne1 >= 32 && ne10 >= 32) || src0->backend == GGML_BACKEND_GPU);
((ne0 >= 32 && ne1 >= 32 && ne10 >= 32) || src0->backend == GGML_BACKEND_TYPE_GPU);
}
static void ggml_vk_mul_mat(ggml_backend_vk_context * ctx, vk_context * subctx, const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst) {
@ -2880,8 +2886,8 @@ static void ggml_vk_op_repeat(ggml_backend_vk_context * ctx, vk_context * subctx
// TODO: support for transposed / permuted tensors
GGML_ASSERT(nb0 == sizeof(float));
GGML_ASSERT(nb00 == sizeof(float));
GGML_ASSERT(src0->backend == GGML_BACKEND_GPU);
GGML_ASSERT(dst->backend == GGML_BACKEND_GPU);
GGML_ASSERT(src0->backend == GGML_BACKEND_TYPE_GPU);
GGML_ASSERT(dst->backend == GGML_BACKEND_TYPE_GPU);
ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) dst->extra;
ggml_tensor_extra_gpu * extra_src0 = (ggml_tensor_extra_gpu *) src0->extra;
@ -3110,8 +3116,8 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context * subctx, c
}
}
const bool transfer_src0 = src0->backend != GGML_BACKEND_GPU && !src0_uma;
const bool transfer_src1 = use_src1 && src1->backend != GGML_BACKEND_GPU && !src1_uma;
const bool transfer_src0 = src0->backend != GGML_BACKEND_TYPE_GPU && !src0_uma;
const bool transfer_src1 = use_src1 && src1->backend != GGML_BACKEND_TYPE_GPU && !src1_uma;
uint64_t x_sz = ggml_vk_align_size(ggml_type_size(src0->type) * ne0, ctx->device.lock()->properties.limits.minStorageBufferOffsetAlignment);
uint64_t y_sz = use_src1 ? ggml_vk_align_size(ggml_type_size(src1->type) * ne1, ctx->device.lock()->properties.limits.minStorageBufferOffsetAlignment) : 0;
@ -3120,7 +3126,7 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context * subctx, c
vk_buffer d_D = extra->buffer_gpu.lock();
// Workaround for tiny tensor inputs on ROPE
if (use_src1 && src1->backend == GGML_BACKEND_GPU && y_sz > d_D->size) {
if (use_src1 && src1->backend == GGML_BACKEND_TYPE_GPU && y_sz > d_D->size) {
y_sz = VK_WHOLE_SIZE;
}
@ -3209,9 +3215,9 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context * subctx, c
ggml_vk_sync_buffers(subctx);
ggml_vk_dispatch_pipeline(ctx, subctx, *pipeline, { { d_X, x_buf_offset, x_sz }, { d_D, d_buf_offset, d_sz } }, sizeof(PC), &pc, elements);
}
if (dst->backend == GGML_BACKEND_CPU && op == GGML_OP_CPY) {
if (dst->backend == GGML_BACKEND_TYPE_CPU && op == GGML_OP_CPY) {
ggml_vk_d2h_tensor_2d(ctx, subctx, d_D, 0, dst);
} else if(dst->backend == GGML_BACKEND_CPU) {
} else if(dst->backend == GGML_BACKEND_TYPE_CPU) {
// copy dst to host
float * d = (float *) dst->data;
ggml_vk_buffer_read_async(ctx, subctx, d_D, 0, d, d_sz);
@ -3253,7 +3259,7 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context * subctx, c
ggml_vk_sync_buffers(subctx);
ggml_vk_dispatch_pipeline(ctx, subctx, *pipeline, { { d_X, x_buf_offset + x_offset, x_sz }, { d_D, d_buf_offset + d_offset, d_sz } }, sizeof(PC), &pc, elements);
}
if (dst->backend == GGML_BACKEND_CPU) {
if (dst->backend == GGML_BACKEND_TYPE_CPU) {
// copy dst to host
ggml_vk_buffer_read_async(ctx, subctx, d_D, d_buf_offset + d_offset, (char *) dst->data + i02*nb2 + i03*nb3, d_sz);
}
@ -3359,7 +3365,7 @@ static void ggml_vk_rope(ggml_backend_vk_context * ctx, vk_context * subctx, con
static void ggml_vk_nop(ggml_backend_vk_context * ctx, vk_context * subctx, const ggml_tensor * src0, ggml_tensor * dst) {
// If backend is CPU, data from src0 has to be copied off the device
if (dst->backend == GGML_BACKEND_CPU) {
if (dst->backend == GGML_BACKEND_TYPE_CPU) {
ggml_tensor_extra_gpu * extra_src0 = (ggml_tensor_extra_gpu *) src0->extra;
vk_buffer d_D = extra_src0->buffer_gpu.lock();
ggml_vk_sync_buffers(subctx);
@ -3994,9 +4000,9 @@ static void ggml_vk_preallocate_buffers_graph(ggml_backend_vk_context * ctx, ggm
#ifdef GGML_VULKAN_DEBUG
std::cerr << "ggml_vk_preallocate_buffers_graph(" << node << ")" << std::endl;
#endif
const bool any_on_device = node->backend == GGML_BACKEND_GPU
|| (node->src[0] != nullptr && (node->src[0]->backend == GGML_BACKEND_GPU || node->src[0]->backend == GGML_BACKEND_GPU_SPLIT))
|| (node->src[1] != nullptr && (node->src[1]->backend == GGML_BACKEND_GPU));
const bool any_on_device = node->backend == GGML_BACKEND_TYPE_GPU
|| (node->src[0] != nullptr && (node->src[0]->backend == GGML_BACKEND_TYPE_GPU || node->src[0]->backend == GGML_BACKEND_TYPE_GPU_SPLIT))
|| (node->src[1] != nullptr && (node->src[1]->backend == GGML_BACKEND_TYPE_GPU));
if (ctx->disable || (!any_on_device && node->op != GGML_OP_MUL_MAT)) {
return;
@ -4215,9 +4221,9 @@ static void ggml_vk_preallocate_buffers(ggml_backend_vk_context * ctx) {
}
static void ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * node, bool last_node){
const bool any_on_device = node->backend == GGML_BACKEND_GPU
|| (node->src[0] != nullptr && (node->src[0]->backend == GGML_BACKEND_GPU || node->src[0]->backend == GGML_BACKEND_GPU_SPLIT))
|| (node->src[1] != nullptr && node->src[1]->backend == GGML_BACKEND_GPU);
const bool any_on_device = node->backend == GGML_BACKEND_TYPE_GPU
|| (node->src[0] != nullptr && (node->src[0]->backend == GGML_BACKEND_TYPE_GPU || node->src[0]->backend == GGML_BACKEND_TYPE_GPU_SPLIT))
|| (node->src[1] != nullptr && node->src[1]->backend == GGML_BACKEND_TYPE_GPU);
if (ctx->disable || (!any_on_device && node->op != GGML_OP_MUL_MAT) || (node->op == GGML_OP_MUL_MAT && !any_on_device && !ggml_vk_can_mul_mat(node->src[0], node->src[1], node))) {
return;
@ -4371,7 +4377,7 @@ static void ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod
last_node = true;
#endif
if (node->backend == GGML_BACKEND_CPU || last_node) {
if (node->backend == GGML_BACKEND_TYPE_CPU || last_node) {
ggml_vk_ctx_end(ctx->compute_ctx);
ctx->compute_ctx->exit_tensor = node;
ctx->compute_ctx = nullptr;
@ -4379,9 +4385,9 @@ static void ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod
}
static bool ggml_vk_compute_forward(ggml_backend_vk_context * ctx, ggml_compute_params * params, ggml_tensor * tensor){
const bool any_on_device = tensor->backend == GGML_BACKEND_GPU
|| (tensor->src[0] != nullptr && (tensor->src[0]->backend == GGML_BACKEND_GPU || tensor->src[0]->backend == GGML_BACKEND_GPU_SPLIT))
|| (tensor->src[1] != nullptr && tensor->src[1]->backend == GGML_BACKEND_GPU);
const bool any_on_device = tensor->backend == GGML_BACKEND_TYPE_GPU
|| (tensor->src[0] != nullptr && (tensor->src[0]->backend == GGML_BACKEND_TYPE_GPU || tensor->src[0]->backend == GGML_BACKEND_TYPE_GPU_SPLIT))
|| (tensor->src[1] != nullptr && tensor->src[1]->backend == GGML_BACKEND_TYPE_GPU);
if (ctx->disable || (!any_on_device && tensor->op != GGML_OP_MUL_MAT)) {
return false;
@ -4442,7 +4448,7 @@ static bool ggml_vk_compute_forward(ggml_backend_vk_context * ctx, ggml_compute_
if (params->ith != 0) {
return true;
}
if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
if (params->type == GGML_TASK_TYPE_INIT || params->type == GGML_TASK_TYPE_FINALIZE) {
return true;
}
@ -4745,7 +4751,7 @@ GGML_CALL static void ggml_backend_vk_buffer_init_tensor(ggml_backend_buffer_t b
extra->offset = (uint8_t *) tensor->data - (uint8_t *) vk_ptr_base;
}
tensor->backend = GGML_BACKEND_GPU;
tensor->backend = GGML_BACKEND_TYPE_GPU;
tensor->extra = extra;
}
@ -4753,7 +4759,7 @@ GGML_CALL static void ggml_backend_vk_buffer_set_tensor(ggml_backend_buffer_t bu
#ifdef GGML_VULKAN_DEBUG
std::cerr << "ggml_backend_vk_buffer_set_tensor(" << buffer << ", " << tensor << ", " << data << ", " << offset << ", " << size << ")" << std::endl;
#endif
GGML_ASSERT(tensor->backend == GGML_BACKEND_GPU);
GGML_ASSERT(tensor->backend == GGML_BACKEND_TYPE_GPU);
ggml_backend_vk_buffer_context * ctx = (ggml_backend_vk_buffer_context *)buffer->context;
@ -4768,7 +4774,7 @@ GGML_CALL static void ggml_backend_vk_buffer_get_tensor(ggml_backend_buffer_t bu
#ifdef GGML_VULKAN_DEBUG
std::cerr << "ggml_backend_vk_buffer_get_tensor(" << buffer << ", " << tensor << ", " << data << ", " << offset << ", " << size << ")" << std::endl;
#endif
GGML_ASSERT(tensor->backend == GGML_BACKEND_GPU);
GGML_ASSERT(tensor->backend == GGML_BACKEND_TYPE_GPU);
ggml_backend_vk_buffer_context * ctx = (ggml_backend_vk_buffer_context *)buffer->context;
@ -4999,7 +5005,7 @@ GGML_CALL static void ggml_backend_vk_set_tensor_async(ggml_backend_t backend, g
#endif
ggml_backend_vk_context * ctx = (ggml_backend_vk_context *)backend->context;
GGML_ASSERT((tensor->buffer->buft == ggml_backend_vk_buffer_type(ctx->idx) || tensor->buffer->buft == ggml_backend_vk_host_buffer_type()) && "unsupported buffer type");
GGML_ASSERT(tensor->backend == GGML_BACKEND_GPU);
GGML_ASSERT(tensor->backend == GGML_BACKEND_TYPE_GPU);
ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) tensor->extra;
@ -5020,7 +5026,7 @@ GGML_CALL static void ggml_backend_vk_get_tensor_async(ggml_backend_t backend, c
#endif
ggml_backend_vk_context * ctx = (ggml_backend_vk_context *)backend->context;
GGML_ASSERT((tensor->buffer->buft == ggml_backend_vk_buffer_type(ctx->idx) || tensor->buffer->buft == ggml_backend_vk_host_buffer_type()) && "unsupported buffer type");
GGML_ASSERT(tensor->backend == GGML_BACKEND_GPU);
GGML_ASSERT(tensor->backend == GGML_BACKEND_TYPE_GPU);
ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) tensor->extra;
@ -5097,7 +5103,7 @@ GGML_CALL static bool ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml
int last_node = cgraph->n_nodes - 1;
// If the last op in the cgraph isn't backend GPU, the command buffer doesn't get closed properly
while (last_node > 0 && cgraph->nodes[last_node]->backend != GGML_BACKEND_GPU) {
while (last_node > 0 && cgraph->nodes[last_node]->backend != GGML_BACKEND_TYPE_GPU) {
last_node -= 1;
}
@ -5106,7 +5112,7 @@ GGML_CALL static bool ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml
}
ggml_compute_params params = {};
params.type = GGML_TASK_COMPUTE;
params.type = GGML_TASK_TYPE_COMPUTE;
params.ith = 0;
for (int i = 0; i < cgraph->n_nodes; i++) {
ggml_tensor * node = cgraph->nodes[i];
@ -5244,6 +5250,11 @@ static ggml_backend_i ggml_backend_vk_interface = {
/* .supports_op = */ ggml_backend_vk_supports_op,
};
static ggml_guid_t ggml_backend_vk_guid() {
static ggml_guid guid = { 0xb8, 0xf7, 0x4f, 0x86, 0x40, 0x3c, 0xe1, 0x02, 0x91, 0xc8, 0xdd, 0xe9, 0x02, 0x3f, 0xc0, 0x2b };
return &guid;
}
GGML_CALL ggml_backend_t ggml_backend_vk_init(size_t idx) {
if (vk_instance.initialized[idx]) {
return vk_instance.backends[idx];
@ -5262,6 +5273,7 @@ GGML_CALL ggml_backend_t ggml_backend_vk_init(size_t idx) {
vk_instance.initialized[idx] = true;
ggml_backend_t vk_backend = new ggml_backend {
/* .guid = */ ggml_backend_vk_guid(),
/* .interface = */ ggml_backend_vk_interface,
/* .context = */ &vk_instance.contexts[ctx->idx],
};
@ -5272,7 +5284,7 @@ GGML_CALL ggml_backend_t ggml_backend_vk_init(size_t idx) {
}
GGML_CALL bool ggml_backend_is_vk(ggml_backend_t backend) {
return backend && backend->iface.get_name == ggml_backend_vk_name;
return backend != NULL && ggml_guid_matches(backend->guid, ggml_backend_vk_guid());
}
GGML_CALL int ggml_backend_vk_get_device_count() {
@ -5410,7 +5422,7 @@ static void ggml_vk_print_tensor_area(const ggml_tensor * tensor, const void * d
static void ggml_vk_print_tensor(ggml_backend_vk_context * ctx, const ggml_tensor * tensor, const char * name) {
void * tensor_data = tensor->data;
if (tensor->backend == GGML_BACKEND_GPU) {
if (tensor->backend == GGML_BACKEND_TYPE_GPU) {
const size_t tensor_size = ggml_nbytes(tensor);
tensor_data = malloc(tensor_size);
@ -5436,14 +5448,14 @@ static void ggml_vk_print_tensor(ggml_backend_vk_context * ctx, const ggml_tenso
std::vector<const ggml_tensor *> done;
ggml_vk_print_graph_origin(tensor, done);
if (tensor->backend == GGML_BACKEND_GPU) {
if (tensor->backend == GGML_BACKEND_TYPE_GPU) {
free(tensor_data);
}
}
static void ggml_vk_check_tensor(const std::string& name, const ggml_tensor * tensor) {
return;
GGML_ASSERT(tensor->backend == GGML_BACKEND_CPU);
GGML_ASSERT(tensor->backend == GGML_BACKEND_TYPE_CPU);
if (tensor->type != GGML_TYPE_F32 && tensor->type != GGML_TYPE_F16) {
return;
}
@ -5481,7 +5493,7 @@ static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_compute_
if (params->ith != 0) {
return;
}
if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE || tensor->op == GGML_OP_TRANSPOSE) {
if (params->type == GGML_TASK_TYPE_INIT || params->type == GGML_TASK_TYPE_FINALIZE || tensor->op == GGML_OP_TRANSPOSE) {
return;
}
@ -5518,10 +5530,10 @@ static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_compute_
src0_buffer = malloc(src0_size);
src0_clone->data = src0_buffer;
if (src0->backend == GGML_BACKEND_CPU) {
if (src0->backend == GGML_BACKEND_TYPE_CPU) {
memcpy(src0_clone->data, src0->data, src0_size);
memcpy(src0_clone->nb, src0->nb, sizeof(size_t) * GGML_MAX_DIMS);
} else if (src0->backend == GGML_BACKEND_GPU) {
} else if (src0->backend == GGML_BACKEND_TYPE_GPU) {
ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) src0->extra;
uint64_t offset = extra->offset;
if (!ggml_is_contiguous(src0) && ggml_vk_dim01_contiguous(src0)) {
@ -5561,10 +5573,10 @@ static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_compute_
src1_buffer = malloc(src1_size);
src1_clone->data = src1_buffer;
if (src1->backend == GGML_BACKEND_CPU) {
if (src1->backend == GGML_BACKEND_TYPE_CPU) {
memcpy(src1_clone->data, src1->data, src1_size);
memcpy(src1_clone->nb, src1->nb, sizeof(size_t) * GGML_MAX_DIMS);
} else if (src1->backend == GGML_BACKEND_GPU) {
} else if (src1->backend == GGML_BACKEND_TYPE_GPU) {
ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) src1->extra;
uint64_t offset = extra->offset;
if (!ggml_is_contiguous(src1) && ggml_vk_dim01_contiguous(src1)) {
@ -5723,7 +5735,7 @@ static void ggml_vk_check_results_1(ggml_backend_vk_context * ctx, ggml_compute_
if (params->ith != 0) {
return;
}
if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE || tensor->op == GGML_OP_TRANSPOSE) {
if (params->type == GGML_TASK_TYPE_INIT || params->type == GGML_TASK_TYPE_FINALIZE || tensor->op == GGML_OP_TRANSPOSE) {
return;
}
if (!(vk_output_tensor > 0 && vk_output_tensor == check_counter) && check_counter <= vk_skip_checks) {
@ -5735,7 +5747,7 @@ static void ggml_vk_check_results_1(ggml_backend_vk_context * ctx, ggml_compute_
void * tensor_data = tensor->data;
if (tensor->backend == GGML_BACKEND_GPU) {
if (tensor->backend == GGML_BACKEND_TYPE_GPU) {
size_t tensor_size = ggml_nbytes(tensor);
tensor_data = malloc(tensor_size);
@ -5868,7 +5880,7 @@ static void ggml_vk_check_results_1(ggml_backend_vk_context * ctx, ggml_compute_
comp_result = nullptr;
comp_size = 0;
if (tensor->backend == GGML_BACKEND_GPU) {
if (tensor->backend == GGML_BACKEND_TYPE_GPU) {
free(tensor_data);
}
}

513
ggml.c
View file

File diff suppressed because it is too large Load diff

54
ggml.h
View file

@ -350,6 +350,9 @@ extern "C" {
GGML_TYPE_IQ3_XXS = 18,
GGML_TYPE_IQ1_S = 19,
GGML_TYPE_IQ4_NL = 20,
GGML_TYPE_IQ3_S = 21,
GGML_TYPE_IQ2_S = 22,
GGML_TYPE_IQ4_XS = 23,
GGML_TYPE_I8,
GGML_TYPE_I16,
GGML_TYPE_I32,
@ -363,9 +366,9 @@ extern "C" {
};
enum ggml_backend_type {
GGML_BACKEND_CPU = 0,
GGML_BACKEND_GPU = 10,
GGML_BACKEND_GPU_SPLIT = 20,
GGML_BACKEND_TYPE_CPU = 0,
GGML_BACKEND_TYPE_GPU = 10,
GGML_BACKEND_TYPE_GPU_SPLIT = 20,
};
// model file types
@ -389,6 +392,9 @@ extern "C" {
GGML_FTYPE_MOSTLY_IQ3_XXS = 17, // except 1d tensors
GGML_FTYPE_MOSTLY_IQ1_S = 18, // except 1d tensors
GGML_FTYPE_MOSTLY_IQ4_NL = 19, // except 1d tensors
GGML_FTYPE_MOSTLY_IQ3_S = 20, // except 1d tensors
GGML_FTYPE_MOSTLY_IQ2_S = 21, // except 1d tensors
GGML_FTYPE_MOSTLY_IQ4_XS = 22, // except 1d tensors
};
// available tensor operations:
@ -496,9 +502,9 @@ extern "C" {
};
enum ggml_object_type {
GGML_OBJECT_TENSOR,
GGML_OBJECT_GRAPH,
GGML_OBJECT_WORK_BUFFER
GGML_OBJECT_TYPE_TENSOR,
GGML_OBJECT_TYPE_GRAPH,
GGML_OBJECT_TYPE_WORK_BUFFER
};
enum ggml_log_level {
@ -640,9 +646,9 @@ extern "C" {
// NOTE: the INIT or FINALIZE pass is not scheduled unless explicitly enabled.
// This behavior was changed since https://github.com/ggerganov/llama.cpp/pull/1995.
enum ggml_task_type {
GGML_TASK_INIT = 0,
GGML_TASK_COMPUTE,
GGML_TASK_FINALIZE,
GGML_TASK_TYPE_INIT = 0,
GGML_TASK_TYPE_COMPUTE,
GGML_TASK_TYPE_FINALIZE,
};
struct ggml_compute_params {
@ -666,6 +672,16 @@ extern "C" {
GGML_NUMA_STRATEGY_COUNT
};
//
// GUID
//
// GUID types
typedef uint8_t ggml_guid[16];
typedef ggml_guid * ggml_guid_t;
GGML_API bool ggml_guid_matches(ggml_guid_t guid_a, ggml_guid_t guid_b);
// misc
GGML_API void ggml_time_init(void); // call this once at the beginning of the program
@ -1647,8 +1663,8 @@ extern "C" {
// sort rows
enum ggml_sort_order {
GGML_SORT_ASC,
GGML_SORT_DESC,
GGML_SORT_ORDER_ASC,
GGML_SORT_ORDER_DESC,
};
GGML_API struct ggml_tensor * ggml_argsort(
@ -1941,8 +1957,8 @@ extern "C" {
// optimization methods
enum ggml_opt_type {
GGML_OPT_ADAM,
GGML_OPT_LBFGS,
GGML_OPT_TYPE_ADAM,
GGML_OPT_TYPE_LBFGS,
};
// linesearch methods
@ -1956,12 +1972,12 @@ extern "C" {
// optimization return values
enum ggml_opt_result {
GGML_OPT_OK = 0,
GGML_OPT_DID_NOT_CONVERGE,
GGML_OPT_NO_CONTEXT,
GGML_OPT_INVALID_WOLFE,
GGML_OPT_FAIL,
GGML_OPT_CANCEL,
GGML_OPT_RESULT_OK = 0,
GGML_OPT_RESULT_DID_NOT_CONVERGE,
GGML_OPT_RESULT_NO_CONTEXT,
GGML_OPT_RESULT_INVALID_WOLFE,
GGML_OPT_RESULT_FAIL,
GGML_OPT_RESULT_CANCEL,
GGML_LINESEARCH_FAIL = -128,
GGML_LINESEARCH_MINIMUM_STEP,

1337
llama.cpp
View file

File diff suppressed because it is too large Load diff

117
llama.h
View file

@ -64,6 +64,15 @@ extern "C" {
LLAMA_VOCAB_TYPE_WPM = 2, // WordPiece
};
// note: these values should be synchronized with ggml_rope
// TODO: maybe move this enum to ggml.h (ggml_rope_type)
enum llama_rope_type {
LLAMA_ROPE_TYPE_NONE = -1,
LLAMA_ROPE_TYPE_NORM = 0,
LLAMA_ROPE_TYPE_NEOX = 2,
LLAMA_ROPE_TYPE_GLM = 4,
};
enum llama_token_type {
LLAMA_TOKEN_TYPE_UNDEFINED = 0,
LLAMA_TOKEN_TYPE_NORMAL = 1,
@ -98,32 +107,37 @@ extern "C" {
LLAMA_FTYPE_MOSTLY_IQ2_XXS = 19, // except 1d tensors
LLAMA_FTYPE_MOSTLY_IQ2_XS = 20, // except 1d tensors
LLAMA_FTYPE_MOSTLY_Q2_K_S = 21, // except 1d tensors
LLAMA_FTYPE_MOSTLY_Q3_K_XS = 22, // except 1d tensors
LLAMA_FTYPE_MOSTLY_IQ3_XS = 22, // except 1d tensors
LLAMA_FTYPE_MOSTLY_IQ3_XXS = 23, // except 1d tensors
LLAMA_FTYPE_MOSTLY_IQ1_S = 24, // except 1d tensors
LLAMA_FTYPE_MOSTLY_IQ4_NL = 25, // except 1d tensors
LLAMA_FTYPE_MOSTLY_IQ3_S = 26, // except 1d tensors
LLAMA_FTYPE_MOSTLY_IQ3_M = 27, // except 1d tensors
LLAMA_FTYPE_MOSTLY_IQ2_S = 28, // except 1d tensors
LLAMA_FTYPE_MOSTLY_IQ2_M = 29, // except 1d tensors
LLAMA_FTYPE_MOSTLY_IQ4_XS = 30, // except 1d tensors
LLAMA_FTYPE_GUESSED = 1024, // not specified in the model file
};
enum llama_rope_scaling_type {
LLAMA_ROPE_SCALING_UNSPECIFIED = -1,
LLAMA_ROPE_SCALING_NONE = 0,
LLAMA_ROPE_SCALING_LINEAR = 1,
LLAMA_ROPE_SCALING_YARN = 2,
LLAMA_ROPE_SCALING_MAX_VALUE = LLAMA_ROPE_SCALING_YARN,
LLAMA_ROPE_SCALING_TYPE_UNSPECIFIED = -1,
LLAMA_ROPE_SCALING_TYPE_NONE = 0,
LLAMA_ROPE_SCALING_TYPE_LINEAR = 1,
LLAMA_ROPE_SCALING_TYPE_YARN = 2,
LLAMA_ROPE_SCALING_TYPE_MAX_VALUE = LLAMA_ROPE_SCALING_TYPE_YARN,
};
enum llama_pooling_type {
LLAMA_POOLING_NONE = 0,
LLAMA_POOLING_MEAN = 1,
LLAMA_POOLING_CLS = 2,
LLAMA_POOLING_TYPE_NONE = 0,
LLAMA_POOLING_TYPE_MEAN = 1,
LLAMA_POOLING_TYPE_CLS = 2,
};
enum llama_split_mode {
LLAMA_SPLIT_NONE = 0, // single GPU
LLAMA_SPLIT_LAYER = 1, // split layers and KV across GPUs
LLAMA_SPLIT_ROW = 2, // split rows across GPUs
LLAMA_SPLIT_MODE_NONE = 0, // single GPU
LLAMA_SPLIT_MODE_LAYER = 1, // split layers and KV across GPUs
LLAMA_SPLIT_MODE_ROW = 2, // split rows across GPUs
};
typedef struct llama_token_data {
@ -171,9 +185,9 @@ extern "C" {
} llama_batch;
enum llama_model_kv_override_type {
LLAMA_KV_OVERRIDE_INT,
LLAMA_KV_OVERRIDE_FLOAT,
LLAMA_KV_OVERRIDE_BOOL,
LLAMA_KV_OVERRIDE_TYPE_INT,
LLAMA_KV_OVERRIDE_TYPE_FLOAT,
LLAMA_KV_OVERRIDE_TYPE_BOOL,
};
struct llama_model_kv_override {
@ -232,6 +246,7 @@ extern "C" {
float yarn_beta_fast; // YaRN low correction dim
float yarn_beta_slow; // YaRN high correction dim
uint32_t yarn_orig_ctx; // YaRN original context size
float defrag_thold; // defragment the KV cache if holes/size > thold, < 0 disabled (default)
ggml_backend_sched_eval_callback cb_eval;
void * cb_eval_user_data;
@ -349,15 +364,13 @@ extern "C" {
LLAMA_API bool llama_supports_mlock (void);
LLAMA_API bool llama_supports_gpu_offload(void);
LLAMA_API DEPRECATED(bool llama_mmap_supported (void), "use llama_supports_mmap() instead");
LLAMA_API DEPRECATED(bool llama_mlock_supported(void), "use llama_supports_mlock() instead");
LLAMA_API const struct llama_model * llama_get_model(const struct llama_context * ctx);
LLAMA_API uint32_t llama_n_ctx (const struct llama_context * ctx);
LLAMA_API uint32_t llama_n_batch (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 int32_t llama_n_vocab (const struct llama_model * model);
LLAMA_API int32_t llama_n_ctx_train(const struct llama_model * model);
@ -407,14 +420,6 @@ extern "C" {
// The model needs to be reloaded before applying a new adapter, otherwise the adapter
// will be applied on top of the previous one
// Returns 0 on success
LLAMA_API DEPRECATED(int32_t llama_apply_lora_from_file(
struct llama_context * ctx,
const char * path_lora,
float scale,
const char * path_base_model,
int32_t n_threads),
"use llama_model_apply_lora_from_file instead");
LLAMA_API int32_t llama_model_apply_lora_from_file(
const struct llama_model * model,
const char * path_lora,
@ -512,10 +517,12 @@ extern "C" {
llama_seq_id seq_id);
// Adds relative position "delta" to all tokens that belong to the specified sequence and have positions in [p0, p1)
// If the KV cache is RoPEd, the KV data is updated accordingly
// If the KV cache is RoPEd, the KV data is updated accordingly:
// - lazily on next llama_decode()
// - explicitly with llama_kv_cache_update()
// p0 < 0 : [0, p1]
// p1 < 0 : [p0, inf)
LLAMA_API void llama_kv_cache_seq_shift(
LLAMA_API void llama_kv_cache_seq_add(
struct llama_context * ctx,
llama_seq_id seq_id,
llama_pos p0,
@ -523,7 +530,9 @@ extern "C" {
llama_pos delta);
// Integer division of the positions by factor of `d > 1`
// If the KV cache is RoPEd, the KV data is updated accordingly
// If the KV cache is RoPEd, the KV data is updated accordingly:
// - lazily on next llama_decode()
// - explicitly with llama_kv_cache_update()
// p0 < 0 : [0, p1]
// p1 < 0 : [p0, inf)
LLAMA_API void llama_kv_cache_seq_div(
@ -533,6 +542,20 @@ extern "C" {
llama_pos p1,
int d);
// Returns the largest position present in the KV cache for the specified sequence
LLAMA_API llama_pos llama_kv_cache_seq_pos_max(
struct llama_context * ctx,
llama_seq_id seq_id);
// Defragment the KV cache
// This will be applied:
// - lazily on next llama_decode()
// - explicitly with llama_kv_cache_update()
LLAMA_API void llama_kv_cache_defrag(struct llama_context * ctx);
// Apply the KV cache updates (such as K-shifts, defragmentation, etc.)
LLAMA_API void llama_kv_cache_update(struct llama_context * ctx);
//
// State / sessions
//
@ -552,7 +575,7 @@ extern "C" {
// Returns the number of bytes read
LLAMA_API size_t llama_set_state_data(
struct llama_context * ctx,
uint8_t * src);
const uint8_t * src);
// Save/load session file
LLAMA_API bool llama_load_session_file(
@ -572,27 +595,6 @@ extern "C" {
// Decoding
//
// Run the llama inference to obtain the logits and probabilities for the next token(s).
// 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
// Returns 0 on success
// DEPRECATED: use llama_decode() instead
LLAMA_API DEPRECATED(int llama_eval(
struct llama_context * ctx,
llama_token * tokens,
int32_t n_tokens,
int32_t n_past),
"use llama_decode() instead");
// Same as llama_eval, but use float matrix input directly.
// DEPRECATED: use llama_decode() instead
LLAMA_API DEPRECATED(int llama_eval_embd(
struct llama_context * ctx,
float * embd,
int32_t n_tokens,
int32_t n_past),
"use llama_decode() instead");
// Return batch for single sequence of tokens starting at pos_0
//
// NOTE: this is a helper function to facilitate transition to the new batch API - avoid using it
@ -766,13 +768,6 @@ extern "C" {
float * logits_guidance,
float scale);
LLAMA_API DEPRECATED(void llama_sample_classifier_free_guidance(
struct llama_context * ctx,
llama_token_data_array * candidates,
struct llama_context * guidance_ctx,
float scale),
"use llama_sample_apply_guidance() instead");
/// @details Sorts candidate tokens by their logits in descending order and calculate probabilities based on logits.
LLAMA_API void llama_sample_softmax(
struct llama_context * ctx,
@ -826,12 +821,6 @@ extern "C" {
llama_token_data_array * candidates,
float temp);
LLAMA_API DEPRECATED(void llama_sample_temperature(
struct llama_context * ctx,
llama_token_data_array * candidates,
float temp),
"use llama_sample_temp instead");
/// @details Apply constraints from grammar
LLAMA_API void llama_sample_grammar(
struct llama_context * ctx,

2
scripts/sync-ggml.last
View file

@ -1 +1 @@
8cdf783f288a98eddf521b0ab1b4d405be9e18ba
b458250b736a7473f7ff3560d47c93f1644f3290

8
tests/test-backend-ops.cpp
View file

@ -1264,7 +1264,7 @@ struct test_argsort : public test_case {
test_argsort(ggml_type type = GGML_TYPE_F32,
std::array<int64_t, 4> ne = {16, 10, 10, 10},
ggml_sort_order order = GGML_SORT_ASC)
ggml_sort_order order = GGML_SORT_ORDER_ASC)
: type(type), ne(ne), order(order) {}
ggml_tensor * build_graph(ggml_context * ctx) override {
@ -1916,9 +1916,9 @@ static bool test_backend(ggml_backend_t backend, test_mode mode, const char * op
GGML_TYPE_Q2_K, GGML_TYPE_Q3_K,
GGML_TYPE_Q4_K, GGML_TYPE_Q5_K,
GGML_TYPE_Q6_K,
GGML_TYPE_IQ2_XXS, GGML_TYPE_IQ2_XS,
GGML_TYPE_IQ2_XXS, GGML_TYPE_IQ2_XS, GGML_TYPE_IQ2_S,
GGML_TYPE_IQ3_XXS, GGML_TYPE_IQ1_S,
GGML_TYPE_IQ4_NL,
GGML_TYPE_IQ4_NL, GGML_TYPE_IQ3_S, GGML_TYPE_IQ4_XS,
};
// unary ops
@ -2116,7 +2116,7 @@ static bool test_backend(ggml_backend_t backend, test_mode mode, const char * op
test_cases.emplace_back(new test_concat(GGML_TYPE_F32));
test_cases.emplace_back(new test_concat(GGML_TYPE_I32));
for (ggml_sort_order order : {GGML_SORT_ASC, GGML_SORT_DESC}) {
for (ggml_sort_order order : {GGML_SORT_ORDER_ASC, GGML_SORT_ORDER_DESC}) {
test_cases.emplace_back(new test_argsort(GGML_TYPE_F32, {8, 1, 1, 1}, order));
test_cases.emplace_back(new test_argsort(GGML_TYPE_F32, {16, 10, 10, 10}, order));
}

2
tests/test-opt.cpp
View file

@ -118,7 +118,7 @@ int main(void) {
const float fe = ggml_get_f32_1d(e, 0);
printf("%s: e = %.4f\n", __func__, fe);
struct ggml_opt_params opt_params = ggml_opt_default_params(GGML_OPT_ADAM);
struct ggml_opt_params opt_params = ggml_opt_default_params(GGML_OPT_TYPE_ADAM);
ggml_opt(ctx, opt_params, e);

6
tests/test-quantize-fns.cpp
View file

@ -150,7 +150,9 @@ int main(int argc, char * argv[]) {
const float total_error = total_quantization_error(qfns, test_size, test_data.data());
const float max_quantization_error =
type == GGML_TYPE_Q2_K ? MAX_QUANTIZATION_TOTAL_ERROR_2BITS :
type == GGML_TYPE_IQ2_S ? MAX_QUANTIZATION_TOTAL_ERROR_2BITS :
type == GGML_TYPE_Q3_K ? MAX_QUANTIZATION_TOTAL_ERROR_3BITS :
type == GGML_TYPE_IQ3_S ? MAX_QUANTIZATION_TOTAL_ERROR_3BITS :
type == GGML_TYPE_IQ3_XXS ? MAX_QUANTIZATION_TOTAL_ERROR_3BITS_XXS : MAX_QUANTIZATION_TOTAL_ERROR;
failed = !(total_error < max_quantization_error);
num_failed += failed;
@ -167,7 +169,9 @@ int main(int argc, char * argv[]) {
const float vec_dot_error = dot_product_error(qfns, test_size, test_data.data(), test_data2.data());
const float max_allowed_error = type == GGML_TYPE_Q2_K || type == GGML_TYPE_IQ2_XS || type == GGML_TYPE_IQ2_XXS ||
type == GGML_TYPE_IQ3_XXS ? MAX_DOT_PRODUCT_ERROR_LOWBIT : MAX_DOT_PRODUCT_ERROR;
type == GGML_TYPE_IQ3_XXS || type == GGML_TYPE_IQ3_S || type == GGML_TYPE_IQ2_S
? MAX_DOT_PRODUCT_ERROR_LOWBIT
: MAX_DOT_PRODUCT_ERROR;
failed = !(vec_dot_error < max_allowed_error);
num_failed += failed;
if (failed || verbose) {

265
unicode.h
View file

@ -223,6 +223,268 @@ static const std::vector<std::pair<uint32_t, uint32_t>> control_ranges = {
{0x2B81E, 0x2B81F}, {0x2CEA2, 0x2CEAF}, {0x2EBE1, 0x2F7FF}, {0x2FA1E, 0x2FFFF}, {0x3134B, 0xE00FF}, {0xE01F0, 0x10FFFF},
};
static const std::unordered_map<uint32_t, std::vector<uint32_t>> nfd_map = {
{0xC0, {0x41, 0x300}}, {0xC1, {0x41, 0x301}}, {0xC2, {0x41, 0x302}}, {0xC3, {0x41, 0x303}}, {0xC4, {0x41, 0x308}}, {0xC5, {0x41, 0x30A}}, {0xC7, {0x43, 0x327}}, {0xC8, {0x45, 0x300}},
{0xC9, {0x45, 0x301}}, {0xCA, {0x45, 0x302}}, {0xCB, {0x45, 0x308}}, {0xCC, {0x49, 0x300}}, {0xCD, {0x49, 0x301}}, {0xCE, {0x49, 0x302}}, {0xCF, {0x49, 0x308}}, {0xD1, {0x4E, 0x303}},
{0xD2, {0x4F, 0x300}}, {0xD3, {0x4F, 0x301}}, {0xD4, {0x4F, 0x302}}, {0xD5, {0x4F, 0x303}}, {0xD6, {0x4F, 0x308}}, {0xD9, {0x55, 0x300}}, {0xDA, {0x55, 0x301}}, {0xDB, {0x55, 0x302}},
{0xDC, {0x55, 0x308}}, {0xDD, {0x59, 0x301}}, {0xE0, {0x61, 0x300}}, {0xE1, {0x61, 0x301}}, {0xE2, {0x61, 0x302}}, {0xE3, {0x61, 0x303}}, {0xE4, {0x61, 0x308}}, {0xE5, {0x61, 0x30A}},
{0xE7, {0x63, 0x327}}, {0xE8, {0x65, 0x300}}, {0xE9, {0x65, 0x301}}, {0xEA, {0x65, 0x302}}, {0xEB, {0x65, 0x308}}, {0xEC, {0x69, 0x300}}, {0xED, {0x69, 0x301}}, {0xEE, {0x69, 0x302}},
{0xEF, {0x69, 0x308}}, {0xF1, {0x6E, 0x303}}, {0xF2, {0x6F, 0x300}}, {0xF3, {0x6F, 0x301}}, {0xF4, {0x6F, 0x302}}, {0xF5, {0x6F, 0x303}}, {0xF6, {0x6F, 0x308}}, {0xF9, {0x75, 0x300}},
{0xFA, {0x75, 0x301}}, {0xFB, {0x75, 0x302}}, {0xFC, {0x75, 0x308}}, {0xFD, {0x79, 0x301}}, {0xFF, {0x79, 0x308}}, {0x100, {0x41, 0x304}}, {0x101, {0x61, 0x304}}, {0x102, {0x41, 0x306}},
{0x103, {0x61, 0x306}}, {0x104, {0x41, 0x328}}, {0x105, {0x61, 0x328}}, {0x106, {0x43, 0x301}}, {0x107, {0x63, 0x301}}, {0x108, {0x43, 0x302}}, {0x109, {0x63, 0x302}}, {0x10A, {0x43, 0x307}},
{0x10B, {0x63, 0x307}}, {0x10C, {0x43, 0x30C}}, {0x10D, {0x63, 0x30C}}, {0x10E, {0x44, 0x30C}}, {0x10F, {0x64, 0x30C}}, {0x112, {0x45, 0x304}}, {0x113, {0x65, 0x304}}, {0x114, {0x45, 0x306}},
{0x115, {0x65, 0x306}}, {0x116, {0x45, 0x307}}, {0x117, {0x65, 0x307}}, {0x118, {0x45, 0x328}}, {0x119, {0x65, 0x328}}, {0x11A, {0x45, 0x30C}}, {0x11B, {0x65, 0x30C}}, {0x11C, {0x47, 0x302}},
{0x11D, {0x67, 0x302}}, {0x11E, {0x47, 0x306}}, {0x11F, {0x67, 0x306}}, {0x120, {0x47, 0x307}}, {0x121, {0x67, 0x307}}, {0x122, {0x47, 0x327}}, {0x123, {0x67, 0x327}}, {0x124, {0x48, 0x302}},
{0x125, {0x68, 0x302}}, {0x128, {0x49, 0x303}}, {0x129, {0x69, 0x303}}, {0x12A, {0x49, 0x304}}, {0x12B, {0x69, 0x304}}, {0x12C, {0x49, 0x306}}, {0x12D, {0x69, 0x306}}, {0x12E, {0x49, 0x328}},
{0x12F, {0x69, 0x328}}, {0x130, {0x49, 0x307}}, {0x134, {0x4A, 0x302}}, {0x135, {0x6A, 0x302}}, {0x136, {0x4B, 0x327}}, {0x137, {0x6B, 0x327}}, {0x139, {0x4C, 0x301}}, {0x13A, {0x6C, 0x301}},
{0x13B, {0x4C, 0x327}}, {0x13C, {0x6C, 0x327}}, {0x13D, {0x4C, 0x30C}}, {0x13E, {0x6C, 0x30C}}, {0x143, {0x4E, 0x301}}, {0x144, {0x6E, 0x301}}, {0x145, {0x4E, 0x327}}, {0x146, {0x6E, 0x327}},
{0x147, {0x4E, 0x30C}}, {0x148, {0x6E, 0x30C}}, {0x14C, {0x4F, 0x304}}, {0x14D, {0x6F, 0x304}}, {0x14E, {0x4F, 0x306}}, {0x14F, {0x6F, 0x306}}, {0x150, {0x4F, 0x30B}}, {0x151, {0x6F, 0x30B}},
{0x154, {0x52, 0x301}}, {0x155, {0x72, 0x301}}, {0x156, {0x52, 0x327}}, {0x157, {0x72, 0x327}}, {0x158, {0x52, 0x30C}}, {0x159, {0x72, 0x30C}}, {0x15A, {0x53, 0x301}}, {0x15B, {0x73, 0x301}},
{0x15C, {0x53, 0x302}}, {0x15D, {0x73, 0x302}}, {0x15E, {0x53, 0x327}}, {0x15F, {0x73, 0x327}}, {0x160, {0x53, 0x30C}}, {0x161, {0x73, 0x30C}}, {0x162, {0x54, 0x327}}, {0x163, {0x74, 0x327}},
{0x164, {0x54, 0x30C}}, {0x165, {0x74, 0x30C}}, {0x168, {0x55, 0x303}}, {0x169, {0x75, 0x303}}, {0x16A, {0x55, 0x304}}, {0x16B, {0x75, 0x304}}, {0x16C, {0x55, 0x306}}, {0x16D, {0x75, 0x306}},
{0x16E, {0x55, 0x30A}}, {0x16F, {0x75, 0x30A}}, {0x170, {0x55, 0x30B}}, {0x171, {0x75, 0x30B}}, {0x172, {0x55, 0x328}}, {0x173, {0x75, 0x328}}, {0x174, {0x57, 0x302}}, {0x175, {0x77, 0x302}},
{0x176, {0x59, 0x302}}, {0x177, {0x79, 0x302}}, {0x178, {0x59, 0x308}}, {0x179, {0x5A, 0x301}}, {0x17A, {0x7A, 0x301}}, {0x17B, {0x5A, 0x307}}, {0x17C, {0x7A, 0x307}}, {0x17D, {0x5A, 0x30C}},
{0x17E, {0x7A, 0x30C}}, {0x1A0, {0x4F, 0x31B}}, {0x1A1, {0x6F, 0x31B}}, {0x1AF, {0x55, 0x31B}}, {0x1B0, {0x75, 0x31B}}, {0x1CD, {0x41, 0x30C}}, {0x1CE, {0x61, 0x30C}}, {0x1CF, {0x49, 0x30C}},
{0x1D0, {0x69, 0x30C}}, {0x1D1, {0x4F, 0x30C}}, {0x1D2, {0x6F, 0x30C}}, {0x1D3, {0x55, 0x30C}}, {0x1D4, {0x75, 0x30C}}, {0x1D5, {0x55, 0x308, 0x304}}, {0x1D6, {0x75, 0x308, 0x304}},
{0x1D7, {0x55, 0x308, 0x301}}, {0x1D8, {0x75, 0x308, 0x301}}, {0x1D9, {0x55, 0x308, 0x30C}}, {0x1DA, {0x75, 0x308, 0x30C}}, {0x1DB, {0x55, 0x308, 0x300}}, {0x1DC, {0x75, 0x308, 0x300}},
{0x1DE, {0x41, 0x308, 0x304}}, {0x1DF, {0x61, 0x308, 0x304}}, {0x1E0, {0x41, 0x307, 0x304}}, {0x1E1, {0x61, 0x307, 0x304}}, {0x1E2, {0xC6, 0x304}}, {0x1E3, {0xE6, 0x304}}, {0x1E6, {0x47, 0x30C}},
{0x1E7, {0x67, 0x30C}}, {0x1E8, {0x4B, 0x30C}}, {0x1E9, {0x6B, 0x30C}}, {0x1EA, {0x4F, 0x328}}, {0x1EB, {0x6F, 0x328}}, {0x1EC, {0x4F, 0x328, 0x304}}, {0x1ED, {0x6F, 0x328, 0x304}},
{0x1EE, {0x1B7, 0x30C}}, {0x1EF, {0x292, 0x30C}}, {0x1F0, {0x6A, 0x30C}}, {0x1F4, {0x47, 0x301}}, {0x1F5, {0x67, 0x301}}, {0x1F8, {0x4E, 0x300}}, {0x1F9, {0x6E, 0x300}}, {0x1FA, {0x41, 0x30A, 0x301}},
{0x1FB, {0x61, 0x30A, 0x301}}, {0x1FC, {0xC6, 0x301}}, {0x1FD, {0xE6, 0x301}}, {0x1FE, {0xD8, 0x301}}, {0x1FF, {0xF8, 0x301}}, {0x200, {0x41, 0x30F}}, {0x201, {0x61, 0x30F}}, {0x202, {0x41, 0x311}},
{0x203, {0x61, 0x311}}, {0x204, {0x45, 0x30F}}, {0x205, {0x65, 0x30F}}, {0x206, {0x45, 0x311}}, {0x207, {0x65, 0x311}}, {0x208, {0x49, 0x30F}}, {0x209, {0x69, 0x30F}}, {0x20A, {0x49, 0x311}},
{0x20B, {0x69, 0x311}}, {0x20C, {0x4F, 0x30F}}, {0x20D, {0x6F, 0x30F}}, {0x20E, {0x4F, 0x311}}, {0x20F, {0x6F, 0x311}}, {0x210, {0x52, 0x30F}}, {0x211, {0x72, 0x30F}}, {0x212, {0x52, 0x311}},
{0x213, {0x72, 0x311}}, {0x214, {0x55, 0x30F}}, {0x215, {0x75, 0x30F}}, {0x216, {0x55, 0x311}}, {0x217, {0x75, 0x311}}, {0x218, {0x53, 0x326}}, {0x219, {0x73, 0x326}}, {0x21A, {0x54, 0x326}},
{0x21B, {0x74, 0x326}}, {0x21E, {0x48, 0x30C}}, {0x21F, {0x68, 0x30C}}, {0x226, {0x41, 0x307}}, {0x227, {0x61, 0x307}}, {0x228, {0x45, 0x327}}, {0x229, {0x65, 0x327}}, {0x22A, {0x4F, 0x308, 0x304}},
{0x22B, {0x6F, 0x308, 0x304}}, {0x22C, {0x4F, 0x303, 0x304}}, {0x22D, {0x6F, 0x303, 0x304}}, {0x22E, {0x4F, 0x307}}, {0x22F, {0x6F, 0x307}}, {0x230, {0x4F, 0x307, 0x304}},
{0x231, {0x6F, 0x307, 0x304}}, {0x232, {0x59, 0x304}}, {0x233, {0x79, 0x304}}, {0x340, {0x300}}, {0x341, {0x301}}, {0x343, {0x313}}, {0x344, {0x308, 0x301}}, {0x374, {0x2B9}}, {0x37E, {0x3B}},
{0x385, {0xA8, 0x301}}, {0x386, {0x391, 0x301}}, {0x387, {0xB7}}, {0x388, {0x395, 0x301}}, {0x389, {0x397, 0x301}}, {0x38A, {0x399, 0x301}}, {0x38C, {0x39F, 0x301}}, {0x38E, {0x3A5, 0x301}},
{0x38F, {0x3A9, 0x301}}, {0x390, {0x3B9, 0x308, 0x301}}, {0x3AA, {0x399, 0x308}}, {0x3AB, {0x3A5, 0x308}}, {0x3AC, {0x3B1, 0x301}}, {0x3AD, {0x3B5, 0x301}}, {0x3AE, {0x3B7, 0x301}},
{0x3AF, {0x3B9, 0x301}}, {0x3B0, {0x3C5, 0x308, 0x301}}, {0x3CA, {0x3B9, 0x308}}, {0x3CB, {0x3C5, 0x308}}, {0x3CC, {0x3BF, 0x301}}, {0x3CD, {0x3C5, 0x301}}, {0x3CE, {0x3C9, 0x301}},
{0x3D3, {0x3D2, 0x301}}, {0x3D4, {0x3D2, 0x308}}, {0x400, {0x415, 0x300}}, {0x401, {0x415, 0x308}}, {0x403, {0x413, 0x301}}, {0x407, {0x406, 0x308}}, {0x40C, {0x41A, 0x301}}, {0x40D, {0x418, 0x300}},
{0x40E, {0x423, 0x306}}, {0x419, {0x418, 0x306}}, {0x439, {0x438, 0x306}}, {0x450, {0x435, 0x300}}, {0x451, {0x435, 0x308}}, {0x453, {0x433, 0x301}}, {0x457, {0x456, 0x308}}, {0x45C, {0x43A, 0x301}},
{0x45D, {0x438, 0x300}}, {0x45E, {0x443, 0x306}}, {0x476, {0x474, 0x30F}}, {0x477, {0x475, 0x30F}}, {0x4C1, {0x416, 0x306}}, {0x4C2, {0x436, 0x306}}, {0x4D0, {0x410, 0x306}}, {0x4D1, {0x430, 0x306}},
{0x4D2, {0x410, 0x308}}, {0x4D3, {0x430, 0x308}}, {0x4D6, {0x415, 0x306}}, {0x4D7, {0x435, 0x306}}, {0x4DA, {0x4D8, 0x308}}, {0x4DB, {0x4D9, 0x308}}, {0x4DC, {0x416, 0x308}}, {0x4DD, {0x436, 0x308}},
{0x4DE, {0x417, 0x308}}, {0x4DF, {0x437, 0x308}}, {0x4E2, {0x418, 0x304}}, {0x4E3, {0x438, 0x304}}, {0x4E4, {0x418, 0x308}}, {0x4E5, {0x438, 0x308}}, {0x4E6, {0x41E, 0x308}}, {0x4E7, {0x43E, 0x308}},
{0x4EA, {0x4E8, 0x308}}, {0x4EB, {0x4E9, 0x308}}, {0x4EC, {0x42D, 0x308}}, {0x4ED, {0x44D, 0x308}}, {0x4EE, {0x423, 0x304}}, {0x4EF, {0x443, 0x304}}, {0x4F0, {0x423, 0x308}}, {0x4F1, {0x443, 0x308}},
{0x4F2, {0x423, 0x30B}}, {0x4F3, {0x443, 0x30B}}, {0x4F4, {0x427, 0x308}}, {0x4F5, {0x447, 0x308}}, {0x4F8, {0x42B, 0x308}}, {0x4F9, {0x44B, 0x308}}, {0x622, {0x627, 0x653}}, {0x623, {0x627, 0x654}},
{0x624, {0x648, 0x654}}, {0x625, {0x627, 0x655}}, {0x626, {0x64A, 0x654}}, {0x6C0, {0x6D5, 0x654}}, {0x6C2, {0x6C1, 0x654}}, {0x6D3, {0x6D2, 0x654}}, {0x929, {0x928, 0x93C}}, {0x931, {0x930, 0x93C}},
{0x934, {0x933, 0x93C}}, {0x958, {0x915, 0x93C}}, {0x959, {0x916, 0x93C}}, {0x95A, {0x917, 0x93C}}, {0x95B, {0x91C, 0x93C}}, {0x95C, {0x921, 0x93C}}, {0x95D, {0x922, 0x93C}}, {0x95E, {0x92B, 0x93C}},
{0x95F, {0x92F, 0x93C}}, {0x9CB, {0x9C7, 0x9BE}}, {0x9CC, {0x9C7, 0x9D7}}, {0x9DC, {0x9A1, 0x9BC}}, {0x9DD, {0x9A2, 0x9BC}}, {0x9DF, {0x9AF, 0x9BC}}, {0xA33, {0xA32, 0xA3C}}, {0xA36, {0xA38, 0xA3C}},
{0xA59, {0xA16, 0xA3C}}, {0xA5A, {0xA17, 0xA3C}}, {0xA5B, {0xA1C, 0xA3C}}, {0xA5E, {0xA2B, 0xA3C}}, {0xB48, {0xB47, 0xB56}}, {0xB4B, {0xB47, 0xB3E}}, {0xB4C, {0xB47, 0xB57}}, {0xB5C, {0xB21, 0xB3C}},
{0xB5D, {0xB22, 0xB3C}}, {0xB94, {0xB92, 0xBD7}}, {0xBCA, {0xBC6, 0xBBE}}, {0xBCB, {0xBC7, 0xBBE}}, {0xBCC, {0xBC6, 0xBD7}}, {0xC48, {0xC46, 0xC56}}, {0xCC0, {0xCBF, 0xCD5}}, {0xCC7, {0xCC6, 0xCD5}},
{0xCC8, {0xCC6, 0xCD6}}, {0xCCA, {0xCC6, 0xCC2}}, {0xCCB, {0xCC6, 0xCC2, 0xCD5}}, {0xD4A, {0xD46, 0xD3E}}, {0xD4B, {0xD47, 0xD3E}}, {0xD4C, {0xD46, 0xD57}}, {0xDDA, {0xDD9, 0xDCA}},
{0xDDC, {0xDD9, 0xDCF}}, {0xDDD, {0xDD9, 0xDCF, 0xDCA}}, {0xDDE, {0xDD9, 0xDDF}}, {0xF43, {0xF42, 0xFB7}}, {0xF4D, {0xF4C, 0xFB7}}, {0xF52, {0xF51, 0xFB7}}, {0xF57, {0xF56, 0xFB7}},
{0xF5C, {0xF5B, 0xFB7}}, {0xF69, {0xF40, 0xFB5}}, {0xF73, {0xF71, 0xF72}}, {0xF75, {0xF71, 0xF74}}, {0xF76, {0xFB2, 0xF80}}, {0xF78, {0xFB3, 0xF80}}, {0xF81, {0xF71, 0xF80}}, {0xF93, {0xF92, 0xFB7}},
{0xF9D, {0xF9C, 0xFB7}}, {0xFA2, {0xFA1, 0xFB7}}, {0xFA7, {0xFA6, 0xFB7}}, {0xFAC, {0xFAB, 0xFB7}}, {0xFB9, {0xF90, 0xFB5}}, {0x1026, {0x1025, 0x102E}}, {0x1B06, {0x1B05, 0x1B35}},
{0x1B08, {0x1B07, 0x1B35}}, {0x1B0A, {0x1B09, 0x1B35}}, {0x1B0C, {0x1B0B, 0x1B35}}, {0x1B0E, {0x1B0D, 0x1B35}}, {0x1B12, {0x1B11, 0x1B35}}, {0x1B3B, {0x1B3A, 0x1B35}}, {0x1B3D, {0x1B3C, 0x1B35}},
{0x1B40, {0x1B3E, 0x1B35}}, {0x1B41, {0x1B3F, 0x1B35}}, {0x1B43, {0x1B42, 0x1B35}}, {0x1E00, {0x41, 0x325}}, {0x1E01, {0x61, 0x325}}, {0x1E02, {0x42, 0x307}}, {0x1E03, {0x62, 0x307}},
{0x1E04, {0x42, 0x323}}, {0x1E05, {0x62, 0x323}}, {0x1E06, {0x42, 0x331}}, {0x1E07, {0x62, 0x331}}, {0x1E08, {0x43, 0x327, 0x301}}, {0x1E09, {0x63, 0x327, 0x301}}, {0x1E0A, {0x44, 0x307}},
{0x1E0B, {0x64, 0x307}}, {0x1E0C, {0x44, 0x323}}, {0x1E0D, {0x64, 0x323}}, {0x1E0E, {0x44, 0x331}}, {0x1E0F, {0x64, 0x331}}, {0x1E10, {0x44, 0x327}}, {0x1E11, {0x64, 0x327}}, {0x1E12, {0x44, 0x32D}},
{0x1E13, {0x64, 0x32D}}, {0x1E14, {0x45, 0x304, 0x300}}, {0x1E15, {0x65, 0x304, 0x300}}, {0x1E16, {0x45, 0x304, 0x301}}, {0x1E17, {0x65, 0x304, 0x301}}, {0x1E18, {0x45, 0x32D}},
{0x1E19, {0x65, 0x32D}}, {0x1E1A, {0x45, 0x330}}, {0x1E1B, {0x65, 0x330}}, {0x1E1C, {0x45, 0x327, 0x306}}, {0x1E1D, {0x65, 0x327, 0x306}}, {0x1E1E, {0x46, 0x307}}, {0x1E1F, {0x66, 0x307}},
{0x1E20, {0x47, 0x304}}, {0x1E21, {0x67, 0x304}}, {0x1E22, {0x48, 0x307}}, {0x1E23, {0x68, 0x307}}, {0x1E24, {0x48, 0x323}}, {0x1E25, {0x68, 0x323}}, {0x1E26, {0x48, 0x308}}, {0x1E27, {0x68, 0x308}},
{0x1E28, {0x48, 0x327}}, {0x1E29, {0x68, 0x327}}, {0x1E2A, {0x48, 0x32E}}, {0x1E2B, {0x68, 0x32E}}, {0x1E2C, {0x49, 0x330}}, {0x1E2D, {0x69, 0x330}}, {0x1E2E, {0x49, 0x308, 0x301}},
{0x1E2F, {0x69, 0x308, 0x301}}, {0x1E30, {0x4B, 0x301}}, {0x1E31, {0x6B, 0x301}}, {0x1E32, {0x4B, 0x323}}, {0x1E33, {0x6B, 0x323}}, {0x1E34, {0x4B, 0x331}}, {0x1E35, {0x6B, 0x331}},
{0x1E36, {0x4C, 0x323}}, {0x1E37, {0x6C, 0x323}}, {0x1E38, {0x4C, 0x323, 0x304}}, {0x1E39, {0x6C, 0x323, 0x304}}, {0x1E3A, {0x4C, 0x331}}, {0x1E3B, {0x6C, 0x331}}, {0x1E3C, {0x4C, 0x32D}},
{0x1E3D, {0x6C, 0x32D}}, {0x1E3E, {0x4D, 0x301}}, {0x1E3F, {0x6D, 0x301}}, {0x1E40, {0x4D, 0x307}}, {0x1E41, {0x6D, 0x307}}, {0x1E42, {0x4D, 0x323}}, {0x1E43, {0x6D, 0x323}}, {0x1E44, {0x4E, 0x307}},
{0x1E45, {0x6E, 0x307}}, {0x1E46, {0x4E, 0x323}}, {0x1E47, {0x6E, 0x323}}, {0x1E48, {0x4E, 0x331}}, {0x1E49, {0x6E, 0x331}}, {0x1E4A, {0x4E, 0x32D}}, {0x1E4B, {0x6E, 0x32D}},
{0x1E4C, {0x4F, 0x303, 0x301}}, {0x1E4D, {0x6F, 0x303, 0x301}}, {0x1E4E, {0x4F, 0x303, 0x308}}, {0x1E4F, {0x6F, 0x303, 0x308}}, {0x1E50, {0x4F, 0x304, 0x300}}, {0x1E51, {0x6F, 0x304, 0x300}},
{0x1E52, {0x4F, 0x304, 0x301}}, {0x1E53, {0x6F, 0x304, 0x301}}, {0x1E54, {0x50, 0x301}}, {0x1E55, {0x70, 0x301}}, {0x1E56, {0x50, 0x307}}, {0x1E57, {0x70, 0x307}}, {0x1E58, {0x52, 0x307}},
{0x1E59, {0x72, 0x307}}, {0x1E5A, {0x52, 0x323}}, {0x1E5B, {0x72, 0x323}}, {0x1E5C, {0x52, 0x323, 0x304}}, {0x1E5D, {0x72, 0x323, 0x304}}, {0x1E5E, {0x52, 0x331}}, {0x1E5F, {0x72, 0x331}},
{0x1E60, {0x53, 0x307}}, {0x1E61, {0x73, 0x307}}, {0x1E62, {0x53, 0x323}}, {0x1E63, {0x73, 0x323}}, {0x1E64, {0x53, 0x301, 0x307}}, {0x1E65, {0x73, 0x301, 0x307}}, {0x1E66, {0x53, 0x30C, 0x307}},
{0x1E67, {0x73, 0x30C, 0x307}}, {0x1E68, {0x53, 0x323, 0x307}}, {0x1E69, {0x73, 0x323, 0x307}}, {0x1E6A, {0x54, 0x307}}, {0x1E6B, {0x74, 0x307}}, {0x1E6C, {0x54, 0x323}}, {0x1E6D, {0x74, 0x323}},
{0x1E6E, {0x54, 0x331}}, {0x1E6F, {0x74, 0x331}}, {0x1E70, {0x54, 0x32D}}, {0x1E71, {0x74, 0x32D}}, {0x1E72, {0x55, 0x324}}, {0x1E73, {0x75, 0x324}}, {0x1E74, {0x55, 0x330}}, {0x1E75, {0x75, 0x330}},
{0x1E76, {0x55, 0x32D}}, {0x1E77, {0x75, 0x32D}}, {0x1E78, {0x55, 0x303, 0x301}}, {0x1E79, {0x75, 0x303, 0x301}}, {0x1E7A, {0x55, 0x304, 0x308}}, {0x1E7B, {0x75, 0x304, 0x308}},
{0x1E7C, {0x56, 0x303}}, {0x1E7D, {0x76, 0x303}}, {0x1E7E, {0x56, 0x323}}, {0x1E7F, {0x76, 0x323}}, {0x1E80, {0x57, 0x300}}, {0x1E81, {0x77, 0x300}}, {0x1E82, {0x57, 0x301}}, {0x1E83, {0x77, 0x301}},
{0x1E84, {0x57, 0x308}}, {0x1E85, {0x77, 0x308}}, {0x1E86, {0x57, 0x307}}, {0x1E87, {0x77, 0x307}}, {0x1E88, {0x57, 0x323}}, {0x1E89, {0x77, 0x323}}, {0x1E8A, {0x58, 0x307}}, {0x1E8B, {0x78, 0x307}},
{0x1E8C, {0x58, 0x308}}, {0x1E8D, {0x78, 0x308}}, {0x1E8E, {0x59, 0x307}}, {0x1E8F, {0x79, 0x307}}, {0x1E90, {0x5A, 0x302}}, {0x1E91, {0x7A, 0x302}}, {0x1E92, {0x5A, 0x323}}, {0x1E93, {0x7A, 0x323}},
{0x1E94, {0x5A, 0x331}}, {0x1E95, {0x7A, 0x331}}, {0x1E96, {0x68, 0x331}}, {0x1E97, {0x74, 0x308}}, {0x1E98, {0x77, 0x30A}}, {0x1E99, {0x79, 0x30A}}, {0x1E9B, {0x17F, 0x307}}, {0x1EA0, {0x41, 0x323}},
{0x1EA1, {0x61, 0x323}}, {0x1EA2, {0x41, 0x309}}, {0x1EA3, {0x61, 0x309}}, {0x1EA4, {0x41, 0x302, 0x301}}, {0x1EA5, {0x61, 0x302, 0x301}}, {0x1EA6, {0x41, 0x302, 0x300}},
{0x1EA7, {0x61, 0x302, 0x300}}, {0x1EA8, {0x41, 0x302, 0x309}}, {0x1EA9, {0x61, 0x302, 0x309}}, {0x1EAA, {0x41, 0x302, 0x303}}, {0x1EAB, {0x61, 0x302, 0x303}}, {0x1EAC, {0x41, 0x323, 0x302}},
{0x1EAD, {0x61, 0x323, 0x302}}, {0x1EAE, {0x41, 0x306, 0x301}}, {0x1EAF, {0x61, 0x306, 0x301}}, {0x1EB0, {0x41, 0x306, 0x300}}, {0x1EB1, {0x61, 0x306, 0x300}}, {0x1EB2, {0x41, 0x306, 0x309}},
{0x1EB3, {0x61, 0x306, 0x309}}, {0x1EB4, {0x41, 0x306, 0x303}}, {0x1EB5, {0x61, 0x306, 0x303}}, {0x1EB6, {0x41, 0x323, 0x306}}, {0x1EB7, {0x61, 0x323, 0x306}}, {0x1EB8, {0x45, 0x323}},
{0x1EB9, {0x65, 0x323}}, {0x1EBA, {0x45, 0x309}}, {0x1EBB, {0x65, 0x309}}, {0x1EBC, {0x45, 0x303}}, {0x1EBD, {0x65, 0x303}}, {0x1EBE, {0x45, 0x302, 0x301}}, {0x1EBF, {0x65, 0x302, 0x301}},
{0x1EC0, {0x45, 0x302, 0x300}}, {0x1EC1, {0x65, 0x302, 0x300}}, {0x1EC2, {0x45, 0x302, 0x309}}, {0x1EC3, {0x65, 0x302, 0x309}}, {0x1EC4, {0x45, 0x302, 0x303}}, {0x1EC5, {0x65, 0x302, 0x303}},
{0x1EC6, {0x45, 0x323, 0x302}}, {0x1EC7, {0x65, 0x323, 0x302}}, {0x1EC8, {0x49, 0x309}}, {0x1EC9, {0x69, 0x309}}, {0x1ECA, {0x49, 0x323}}, {0x1ECB, {0x69, 0x323}}, {0x1ECC, {0x4F, 0x323}},
{0x1ECD, {0x6F, 0x323}}, {0x1ECE, {0x4F, 0x309}}, {0x1ECF, {0x6F, 0x309}}, {0x1ED0, {0x4F, 0x302, 0x301}}, {0x1ED1, {0x6F, 0x302, 0x301}}, {0x1ED2, {0x4F, 0x302, 0x300}},
{0x1ED3, {0x6F, 0x302, 0x300}}, {0x1ED4, {0x4F, 0x302, 0x309}}, {0x1ED5, {0x6F, 0x302, 0x309}}, {0x1ED6, {0x4F, 0x302, 0x303}}, {0x1ED7, {0x6F, 0x302, 0x303}}, {0x1ED8, {0x4F, 0x323, 0x302}},
{0x1ED9, {0x6F, 0x323, 0x302}}, {0x1EDA, {0x4F, 0x31B, 0x301}}, {0x1EDB, {0x6F, 0x31B, 0x301}}, {0x1EDC, {0x4F, 0x31B, 0x300}}, {0x1EDD, {0x6F, 0x31B, 0x300}}, {0x1EDE, {0x4F, 0x31B, 0x309}},
{0x1EDF, {0x6F, 0x31B, 0x309}}, {0x1EE0, {0x4F, 0x31B, 0x303}}, {0x1EE1, {0x6F, 0x31B, 0x303}}, {0x1EE2, {0x4F, 0x31B, 0x323}}, {0x1EE3, {0x6F, 0x31B, 0x323}}, {0x1EE4, {0x55, 0x323}},
{0x1EE5, {0x75, 0x323}}, {0x1EE6, {0x55, 0x309}}, {0x1EE7, {0x75, 0x309}}, {0x1EE8, {0x55, 0x31B, 0x301}}, {0x1EE9, {0x75, 0x31B, 0x301}}, {0x1EEA, {0x55, 0x31B, 0x300}},
{0x1EEB, {0x75, 0x31B, 0x300}}, {0x1EEC, {0x55, 0x31B, 0x309}}, {0x1EED, {0x75, 0x31B, 0x309}}, {0x1EEE, {0x55, 0x31B, 0x303}}, {0x1EEF, {0x75, 0x31B, 0x303}}, {0x1EF0, {0x55, 0x31B, 0x323}},
{0x1EF1, {0x75, 0x31B, 0x323}}, {0x1EF2, {0x59, 0x300}}, {0x1EF3, {0x79, 0x300}}, {0x1EF4, {0x59, 0x323}}, {0x1EF5, {0x79, 0x323}}, {0x1EF6, {0x59, 0x309}}, {0x1EF7, {0x79, 0x309}},
{0x1EF8, {0x59, 0x303}}, {0x1EF9, {0x79, 0x303}}, {0x1F00, {0x3B1, 0x313}}, {0x1F01, {0x3B1, 0x314}}, {0x1F02, {0x3B1, 0x313, 0x300}}, {0x1F03, {0x3B1, 0x314, 0x300}}, {0x1F04, {0x3B1, 0x313, 0x301}},
{0x1F05, {0x3B1, 0x314, 0x301}}, {0x1F06, {0x3B1, 0x313, 0x342}}, {0x1F07, {0x3B1, 0x314, 0x342}}, {0x1F08, {0x391, 0x313}}, {0x1F09, {0x391, 0x314}}, {0x1F0A, {0x391, 0x313, 0x300}},
{0x1F0B, {0x391, 0x314, 0x300}}, {0x1F0C, {0x391, 0x313, 0x301}}, {0x1F0D, {0x391, 0x314, 0x301}}, {0x1F0E, {0x391, 0x313, 0x342}}, {0x1F0F, {0x391, 0x314, 0x342}}, {0x1F10, {0x3B5, 0x313}},
{0x1F11, {0x3B5, 0x314}}, {0x1F12, {0x3B5, 0x313, 0x300}}, {0x1F13, {0x3B5, 0x314, 0x300}}, {0x1F14, {0x3B5, 0x313, 0x301}}, {0x1F15, {0x3B5, 0x314, 0x301}}, {0x1F18, {0x395, 0x313}},
{0x1F19, {0x395, 0x314}}, {0x1F1A, {0x395, 0x313, 0x300}}, {0x1F1B, {0x395, 0x314, 0x300}}, {0x1F1C, {0x395, 0x313, 0x301}}, {0x1F1D, {0x395, 0x314, 0x301}}, {0x1F20, {0x3B7, 0x313}},
{0x1F21, {0x3B7, 0x314}}, {0x1F22, {0x3B7, 0x313, 0x300}}, {0x1F23, {0x3B7, 0x314, 0x300}}, {0x1F24, {0x3B7, 0x313, 0x301}}, {0x1F25, {0x3B7, 0x314, 0x301}}, {0x1F26, {0x3B7, 0x313, 0x342}},
{0x1F27, {0x3B7, 0x314, 0x342}}, {0x1F28, {0x397, 0x313}}, {0x1F29, {0x397, 0x314}}, {0x1F2A, {0x397, 0x313, 0x300}}, {0x1F2B, {0x397, 0x314, 0x300}}, {0x1F2C, {0x397, 0x313, 0x301}},
{0x1F2D, {0x397, 0x314, 0x301}}, {0x1F2E, {0x397, 0x313, 0x342}}, {0x1F2F, {0x397, 0x314, 0x342}}, {0x1F30, {0x3B9, 0x313}}, {0x1F31, {0x3B9, 0x314}}, {0x1F32, {0x3B9, 0x313, 0x300}},
{0x1F33, {0x3B9, 0x314, 0x300}}, {0x1F34, {0x3B9, 0x313, 0x301}}, {0x1F35, {0x3B9, 0x314, 0x301}}, {0x1F36, {0x3B9, 0x313, 0x342}}, {0x1F37, {0x3B9, 0x314, 0x342}}, {0x1F38, {0x399, 0x313}},
{0x1F39, {0x399, 0x314}}, {0x1F3A, {0x399, 0x313, 0x300}}, {0x1F3B, {0x399, 0x314, 0x300}}, {0x1F3C, {0x399, 0x313, 0x301}}, {0x1F3D, {0x399, 0x314, 0x301}}, {0x1F3E, {0x399, 0x313, 0x342}},
{0x1F3F, {0x399, 0x314, 0x342}}, {0x1F40, {0x3BF, 0x313}}, {0x1F41, {0x3BF, 0x314}}, {0x1F42, {0x3BF, 0x313, 0x300}}, {0x1F43, {0x3BF, 0x314, 0x300}}, {0x1F44, {0x3BF, 0x313, 0x301}},
{0x1F45, {0x3BF, 0x314, 0x301}}, {0x1F48, {0x39F, 0x313}}, {0x1F49, {0x39F, 0x314}}, {0x1F4A, {0x39F, 0x313, 0x300}}, {0x1F4B, {0x39F, 0x314, 0x300}}, {0x1F4C, {0x39F, 0x313, 0x301}},
{0x1F4D, {0x39F, 0x314, 0x301}}, {0x1F50, {0x3C5, 0x313}}, {0x1F51, {0x3C5, 0x314}}, {0x1F52, {0x3C5, 0x313, 0x300}}, {0x1F53, {0x3C5, 0x314, 0x300}}, {0x1F54, {0x3C5, 0x313, 0x301}},
{0x1F55, {0x3C5, 0x314, 0x301}}, {0x1F56, {0x3C5, 0x313, 0x342}}, {0x1F57, {0x3C5, 0x314, 0x342}}, {0x1F59, {0x3A5, 0x314}}, {0x1F5B, {0x3A5, 0x314, 0x300}}, {0x1F5D, {0x3A5, 0x314, 0x301}},
{0x1F5F, {0x3A5, 0x314, 0x342}}, {0x1F60, {0x3C9, 0x313}}, {0x1F61, {0x3C9, 0x314}}, {0x1F62, {0x3C9, 0x313, 0x300}}, {0x1F63, {0x3C9, 0x314, 0x300}}, {0x1F64, {0x3C9, 0x313, 0x301}},
{0x1F65, {0x3C9, 0x314, 0x301}}, {0x1F66, {0x3C9, 0x313, 0x342}}, {0x1F67, {0x3C9, 0x314, 0x342}}, {0x1F68, {0x3A9, 0x313}}, {0x1F69, {0x3A9, 0x314}}, {0x1F6A, {0x3A9, 0x313, 0x300}},
{0x1F6B, {0x3A9, 0x314, 0x300}}, {0x1F6C, {0x3A9, 0x313, 0x301}}, {0x1F6D, {0x3A9, 0x314, 0x301}}, {0x1F6E, {0x3A9, 0x313, 0x342}}, {0x1F6F, {0x3A9, 0x314, 0x342}}, {0x1F70, {0x3B1, 0x300}},
{0x1F71, {0x3B1, 0x301}}, {0x1F72, {0x3B5, 0x300}}, {0x1F73, {0x3B5, 0x301}}, {0x1F74, {0x3B7, 0x300}}, {0x1F75, {0x3B7, 0x301}}, {0x1F76, {0x3B9, 0x300}}, {0x1F77, {0x3B9, 0x301}},
{0x1F78, {0x3BF, 0x300}}, {0x1F79, {0x3BF, 0x301}}, {0x1F7A, {0x3C5, 0x300}}, {0x1F7B, {0x3C5, 0x301}}, {0x1F7C, {0x3C9, 0x300}}, {0x1F7D, {0x3C9, 0x301}}, {0x1F80, {0x3B1, 0x313, 0x345}},
{0x1F81, {0x3B1, 0x314, 0x345}}, {0x1F82, {0x3B1, 0x313, 0x300, 0x345}}, {0x1F83, {0x3B1, 0x314, 0x300, 0x345}}, {0x1F84, {0x3B1, 0x313, 0x301, 0x345}}, {0x1F85, {0x3B1, 0x314, 0x301, 0x345}},
{0x1F86, {0x3B1, 0x313, 0x342, 0x345}}, {0x1F87, {0x3B1, 0x314, 0x342, 0x345}}, {0x1F88, {0x391, 0x313, 0x345}}, {0x1F89, {0x391, 0x314, 0x345}}, {0x1F8A, {0x391, 0x313, 0x300, 0x345}},
{0x1F8B, {0x391, 0x314, 0x300, 0x345}}, {0x1F8C, {0x391, 0x313, 0x301, 0x345}}, {0x1F8D, {0x391, 0x314, 0x301, 0x345}}, {0x1F8E, {0x391, 0x313, 0x342, 0x345}}, {0x1F8F, {0x391, 0x314, 0x342, 0x345}},
{0x1F90, {0x3B7, 0x313, 0x345}}, {0x1F91, {0x3B7, 0x314, 0x345}}, {0x1F92, {0x3B7, 0x313, 0x300, 0x345}}, {0x1F93, {0x3B7, 0x314, 0x300, 0x345}}, {0x1F94, {0x3B7, 0x313, 0x301, 0x345}},
{0x1F95, {0x3B7, 0x314, 0x301, 0x345}}, {0x1F96, {0x3B7, 0x313, 0x342, 0x345}}, {0x1F97, {0x3B7, 0x314, 0x342, 0x345}}, {0x1F98, {0x397, 0x313, 0x345}}, {0x1F99, {0x397, 0x314, 0x345}},
{0x1F9A, {0x397, 0x313, 0x300, 0x345}}, {0x1F9B, {0x397, 0x314, 0x300, 0x345}}, {0x1F9C, {0x397, 0x313, 0x301, 0x345}}, {0x1F9D, {0x397, 0x314, 0x301, 0x345}}, {0x1F9E, {0x397, 0x313, 0x342, 0x345}},
{0x1F9F, {0x397, 0x314, 0x342, 0x345}}, {0x1FA0, {0x3C9, 0x313, 0x345}}, {0x1FA1, {0x3C9, 0x314, 0x345}}, {0x1FA2, {0x3C9, 0x313, 0x300, 0x345}}, {0x1FA3, {0x3C9, 0x314, 0x300, 0x345}},
{0x1FA4, {0x3C9, 0x313, 0x301, 0x345}}, {0x1FA5, {0x3C9, 0x314, 0x301, 0x345}}, {0x1FA6, {0x3C9, 0x313, 0x342, 0x345}}, {0x1FA7, {0x3C9, 0x314, 0x342, 0x345}}, {0x1FA8, {0x3A9, 0x313, 0x345}},
{0x1FA9, {0x3A9, 0x314, 0x345}}, {0x1FAA, {0x3A9, 0x313, 0x300, 0x345}}, {0x1FAB, {0x3A9, 0x314, 0x300, 0x345}}, {0x1FAC, {0x3A9, 0x313, 0x301, 0x345}}, {0x1FAD, {0x3A9, 0x314, 0x301, 0x345}},
{0x1FAE, {0x3A9, 0x313, 0x342, 0x345}}, {0x1FAF, {0x3A9, 0x314, 0x342, 0x345}}, {0x1FB0, {0x3B1, 0x306}}, {0x1FB1, {0x3B1, 0x304}}, {0x1FB2, {0x3B1, 0x300, 0x345}}, {0x1FB3, {0x3B1, 0x345}},
{0x1FB4, {0x3B1, 0x301, 0x345}}, {0x1FB6, {0x3B1, 0x342}}, {0x1FB7, {0x3B1, 0x342, 0x345}}, {0x1FB8, {0x391, 0x306}}, {0x1FB9, {0x391, 0x304}}, {0x1FBA, {0x391, 0x300}}, {0x1FBB, {0x391, 0x301}},
{0x1FBC, {0x391, 0x345}}, {0x1FBE, {0x3B9}}, {0x1FC1, {0xA8, 0x342}}, {0x1FC2, {0x3B7, 0x300, 0x345}}, {0x1FC3, {0x3B7, 0x345}}, {0x1FC4, {0x3B7, 0x301, 0x345}}, {0x1FC6, {0x3B7, 0x342}},
{0x1FC7, {0x3B7, 0x342, 0x345}}, {0x1FC8, {0x395, 0x300}}, {0x1FC9, {0x395, 0x301}}, {0x1FCA, {0x397, 0x300}}, {0x1FCB, {0x397, 0x301}}, {0x1FCC, {0x397, 0x345}}, {0x1FCD, {0x1FBF, 0x300}},
{0x1FCE, {0x1FBF, 0x301}}, {0x1FCF, {0x1FBF, 0x342}}, {0x1FD0, {0x3B9, 0x306}}, {0x1FD1, {0x3B9, 0x304}}, {0x1FD2, {0x3B9, 0x308, 0x300}}, {0x1FD3, {0x3B9, 0x308, 0x301}}, {0x1FD6, {0x3B9, 0x342}},
{0x1FD7, {0x3B9, 0x308, 0x342}}, {0x1FD8, {0x399, 0x306}}, {0x1FD9, {0x399, 0x304}}, {0x1FDA, {0x399, 0x300}}, {0x1FDB, {0x399, 0x301}}, {0x1FDD, {0x1FFE, 0x300}}, {0x1FDE, {0x1FFE, 0x301}},
{0x1FDF, {0x1FFE, 0x342}}, {0x1FE0, {0x3C5, 0x306}}, {0x1FE1, {0x3C5, 0x304}}, {0x1FE2, {0x3C5, 0x308, 0x300}}, {0x1FE3, {0x3C5, 0x308, 0x301}}, {0x1FE4, {0x3C1, 0x313}}, {0x1FE5, {0x3C1, 0x314}},
{0x1FE6, {0x3C5, 0x342}}, {0x1FE7, {0x3C5, 0x308, 0x342}}, {0x1FE8, {0x3A5, 0x306}}, {0x1FE9, {0x3A5, 0x304}}, {0x1FEA, {0x3A5, 0x300}}, {0x1FEB, {0x3A5, 0x301}}, {0x1FEC, {0x3A1, 0x314}},
{0x1FED, {0xA8, 0x300}}, {0x1FEE, {0xA8, 0x301}}, {0x1FEF, {0x60}}, {0x1FF2, {0x3C9, 0x300, 0x345}}, {0x1FF3, {0x3C9, 0x345}}, {0x1FF4, {0x3C9, 0x301, 0x345}}, {0x1FF6, {0x3C9, 0x342}},
{0x1FF7, {0x3C9, 0x342, 0x345}}, {0x1FF8, {0x39F, 0x300}}, {0x1FF9, {0x39F, 0x301}}, {0x1FFA, {0x3A9, 0x300}}, {0x1FFB, {0x3A9, 0x301}}, {0x1FFC, {0x3A9, 0x345}}, {0x1FFD, {0xB4}}, {0x2000, {0x2002}},
{0x2001, {0x2003}}, {0x2126, {0x3A9}}, {0x212A, {0x4B}}, {0x212B, {0x41, 0x30A}}, {0x219A, {0x2190, 0x338}}, {0x219B, {0x2192, 0x338}}, {0x21AE, {0x2194, 0x338}}, {0x21CD, {0x21D0, 0x338}},
{0x21CE, {0x21D4, 0x338}}, {0x21CF, {0x21D2, 0x338}}, {0x2204, {0x2203, 0x338}}, {0x2209, {0x2208, 0x338}}, {0x220C, {0x220B, 0x338}}, {0x2224, {0x2223, 0x338}}, {0x2226, {0x2225, 0x338}},
{0x2241, {0x223C, 0x338}}, {0x2244, {0x2243, 0x338}}, {0x2247, {0x2245, 0x338}}, {0x2249, {0x2248, 0x338}}, {0x2260, {0x3D, 0x338}}, {0x2262, {0x2261, 0x338}}, {0x226D, {0x224D, 0x338}},
{0x226E, {0x3C, 0x338}}, {0x226F, {0x3E, 0x338}}, {0x2270, {0x2264, 0x338}}, {0x2271, {0x2265, 0x338}}, {0x2274, {0x2272, 0x338}}, {0x2275, {0x2273, 0x338}}, {0x2278, {0x2276, 0x338}},
{0x2279, {0x2277, 0x338}}, {0x2280, {0x227A, 0x338}}, {0x2281, {0x227B, 0x338}}, {0x2284, {0x2282, 0x338}}, {0x2285, {0x2283, 0x338}}, {0x2288, {0x2286, 0x338}}, {0x2289, {0x2287, 0x338}},
{0x22AC, {0x22A2, 0x338}}, {0x22AD, {0x22A8, 0x338}}, {0x22AE, {0x22A9, 0x338}}, {0x22AF, {0x22AB, 0x338}}, {0x22E0, {0x227C, 0x338}}, {0x22E1, {0x227D, 0x338}}, {0x22E2, {0x2291, 0x338}},
{0x22E3, {0x2292, 0x338}}, {0x22EA, {0x22B2, 0x338}}, {0x22EB, {0x22B3, 0x338}}, {0x22EC, {0x22B4, 0x338}}, {0x22ED, {0x22B5, 0x338}}, {0x2329, {0x3008}}, {0x232A, {0x3009}},
{0x2ADC, {0x2ADD, 0x338}}, {0x304C, {0x304B, 0x3099}}, {0x304E, {0x304D, 0x3099}}, {0x3050, {0x304F, 0x3099}}, {0x3052, {0x3051, 0x3099}}, {0x3054, {0x3053, 0x3099}}, {0x3056, {0x3055, 0x3099}},
{0x3058, {0x3057, 0x3099}}, {0x305A, {0x3059, 0x3099}}, {0x305C, {0x305B, 0x3099}}, {0x305E, {0x305D, 0x3099}}, {0x3060, {0x305F, 0x3099}}, {0x3062, {0x3061, 0x3099}}, {0x3065, {0x3064, 0x3099}},
{0x3067, {0x3066, 0x3099}}, {0x3069, {0x3068, 0x3099}}, {0x3070, {0x306F, 0x3099}}, {0x3071, {0x306F, 0x309A}}, {0x3073, {0x3072, 0x3099}}, {0x3074, {0x3072, 0x309A}}, {0x3076, {0x3075, 0x3099}},
{0x3077, {0x3075, 0x309A}}, {0x3079, {0x3078, 0x3099}}, {0x307A, {0x3078, 0x309A}}, {0x307C, {0x307B, 0x3099}}, {0x307D, {0x307B, 0x309A}}, {0x3094, {0x3046, 0x3099}}, {0x309E, {0x309D, 0x3099}},
{0x30AC, {0x30AB, 0x3099}}, {0x30AE, {0x30AD, 0x3099}}, {0x30B0, {0x30AF, 0x3099}}, {0x30B2, {0x30B1, 0x3099}}, {0x30B4, {0x30B3, 0x3099}}, {0x30B6, {0x30B5, 0x3099}}, {0x30B8, {0x30B7, 0x3099}},
{0x30BA, {0x30B9, 0x3099}}, {0x30BC, {0x30BB, 0x3099}}, {0x30BE, {0x30BD, 0x3099}}, {0x30C0, {0x30BF, 0x3099}}, {0x30C2, {0x30C1, 0x3099}}, {0x30C5, {0x30C4, 0x3099}}, {0x30C7, {0x30C6, 0x3099}},
{0x30C9, {0x30C8, 0x3099}}, {0x30D0, {0x30CF, 0x3099}}, {0x30D1, {0x30CF, 0x309A}}, {0x30D3, {0x30D2, 0x3099}}, {0x30D4, {0x30D2, 0x309A}}, {0x30D6, {0x30D5, 0x3099}}, {0x30D7, {0x30D5, 0x309A}},
{0x30D9, {0x30D8, 0x3099}}, {0x30DA, {0x30D8, 0x309A}}, {0x30DC, {0x30DB, 0x3099}}, {0x30DD, {0x30DB, 0x309A}}, {0x30F4, {0x30A6, 0x3099}}, {0x30F7, {0x30EF, 0x3099}}, {0x30F8, {0x30F0, 0x3099}},
{0x30F9, {0x30F1, 0x3099}}, {0x30FA, {0x30F2, 0x3099}}, {0x30FE, {0x30FD, 0x3099}}, {0xF900, {0x8C48}}, {0xF901, {0x66F4}}, {0xF902, {0x8ECA}}, {0xF903, {0x8CC8}}, {0xF904, {0x6ED1}},
{0xF905, {0x4E32}}, {0xF906, {0x53E5}}, {0xF907, {0x9F9C}}, {0xF908, {0x9F9C}}, {0xF909, {0x5951}}, {0xF90A, {0x91D1}}, {0xF90B, {0x5587}}, {0xF90C, {0x5948}}, {0xF90D, {0x61F6}}, {0xF90E, {0x7669}},
{0xF90F, {0x7F85}}, {0xF910, {0x863F}}, {0xF911, {0x87BA}}, {0xF912, {0x88F8}}, {0xF913, {0x908F}}, {0xF914, {0x6A02}}, {0xF915, {0x6D1B}}, {0xF916, {0x70D9}}, {0xF917, {0x73DE}}, {0xF918, {0x843D}},
{0xF919, {0x916A}}, {0xF91A, {0x99F1}}, {0xF91B, {0x4E82}}, {0xF91C, {0x5375}}, {0xF91D, {0x6B04}}, {0xF91E, {0x721B}}, {0xF91F, {0x862D}}, {0xF920, {0x9E1E}}, {0xF921, {0x5D50}}, {0xF922, {0x6FEB}},
{0xF923, {0x85CD}}, {0xF924, {0x8964}}, {0xF925, {0x62C9}}, {0xF926, {0x81D8}}, {0xF927, {0x881F}}, {0xF928, {0x5ECA}}, {0xF929, {0x6717}}, {0xF92A, {0x6D6A}}, {0xF92B, {0x72FC}}, {0xF92C, {0x90CE}},
{0xF92D, {0x4F86}}, {0xF92E, {0x51B7}}, {0xF92F, {0x52DE}}, {0xF930, {0x64C4}}, {0xF931, {0x6AD3}}, {0xF932, {0x7210}}, {0xF933, {0x76E7}}, {0xF934, {0x8001}}, {0xF935, {0x8606}}, {0xF936, {0x865C}},
{0xF937, {0x8DEF}}, {0xF938, {0x9732}}, {0xF939, {0x9B6F}}, {0xF93A, {0x9DFA}}, {0xF93B, {0x788C}}, {0xF93C, {0x797F}}, {0xF93D, {0x7DA0}}, {0xF93E, {0x83C9}}, {0xF93F, {0x9304}}, {0xF940, {0x9E7F}},
{0xF941, {0x8AD6}}, {0xF942, {0x58DF}}, {0xF943, {0x5F04}}, {0xF944, {0x7C60}}, {0xF945, {0x807E}}, {0xF946, {0x7262}}, {0xF947, {0x78CA}}, {0xF948, {0x8CC2}}, {0xF949, {0x96F7}}, {0xF94A, {0x58D8}},
{0xF94B, {0x5C62}}, {0xF94C, {0x6A13}}, {0xF94D, {0x6DDA}}, {0xF94E, {0x6F0F}}, {0xF94F, {0x7D2F}}, {0xF950, {0x7E37}}, {0xF951, {0x964B}}, {0xF952, {0x52D2}}, {0xF953, {0x808B}}, {0xF954, {0x51DC}},
{0xF955, {0x51CC}}, {0xF956, {0x7A1C}}, {0xF957, {0x7DBE}}, {0xF958, {0x83F1}}, {0xF959, {0x9675}}, {0xF95A, {0x8B80}}, {0xF95B, {0x62CF}}, {0xF95C, {0x6A02}}, {0xF95D, {0x8AFE}}, {0xF95E, {0x4E39}},
{0xF95F, {0x5BE7}}, {0xF960, {0x6012}}, {0xF961, {0x7387}}, {0xF962, {0x7570}}, {0xF963, {0x5317}}, {0xF964, {0x78FB}}, {0xF965, {0x4FBF}}, {0xF966, {0x5FA9}}, {0xF967, {0x4E0D}}, {0xF968, {0x6CCC}},
{0xF969, {0x6578}}, {0xF96A, {0x7D22}}, {0xF96B, {0x53C3}}, {0xF96C, {0x585E}}, {0xF96D, {0x7701}}, {0xF96E, {0x8449}}, {0xF96F, {0x8AAA}}, {0xF970, {0x6BBA}}, {0xF971, {0x8FB0}}, {0xF972, {0x6C88}},
{0xF973, {0x62FE}}, {0xF974, {0x82E5}}, {0xF975, {0x63A0}}, {0xF976, {0x7565}}, {0xF977, {0x4EAE}}, {0xF978, {0x5169}}, {0xF979, {0x51C9}}, {0xF97A, {0x6881}}, {0xF97B, {0x7CE7}}, {0xF97C, {0x826F}},
{0xF97D, {0x8AD2}}, {0xF97E, {0x91CF}}, {0xF97F, {0x52F5}}, {0xF980, {0x5442}}, {0xF981, {0x5973}}, {0xF982, {0x5EEC}}, {0xF983, {0x65C5}}, {0xF984, {0x6FFE}}, {0xF985, {0x792A}}, {0xF986, {0x95AD}},
{0xF987, {0x9A6A}}, {0xF988, {0x9E97}}, {0xF989, {0x9ECE}}, {0xF98A, {0x529B}}, {0xF98B, {0x66C6}}, {0xF98C, {0x6B77}}, {0xF98D, {0x8F62}}, {0xF98E, {0x5E74}}, {0xF98F, {0x6190}}, {0xF990, {0x6200}},
{0xF991, {0x649A}}, {0xF992, {0x6F23}}, {0xF993, {0x7149}}, {0xF994, {0x7489}}, {0xF995, {0x79CA}}, {0xF996, {0x7DF4}}, {0xF997, {0x806F}}, {0xF998, {0x8F26}}, {0xF999, {0x84EE}}, {0xF99A, {0x9023}},
{0xF99B, {0x934A}}, {0xF99C, {0x5217}}, {0xF99D, {0x52A3}}, {0xF99E, {0x54BD}}, {0xF99F, {0x70C8}}, {0xF9A0, {0x88C2}}, {0xF9A1, {0x8AAA}}, {0xF9A2, {0x5EC9}}, {0xF9A3, {0x5FF5}}, {0xF9A4, {0x637B}},
{0xF9A5, {0x6BAE}}, {0xF9A6, {0x7C3E}}, {0xF9A7, {0x7375}}, {0xF9A8, {0x4EE4}}, {0xF9A9, {0x56F9}}, {0xF9AA, {0x5BE7}}, {0xF9AB, {0x5DBA}}, {0xF9AC, {0x601C}}, {0xF9AD, {0x73B2}}, {0xF9AE, {0x7469}},
{0xF9AF, {0x7F9A}}, {0xF9B0, {0x8046}}, {0xF9B1, {0x9234}}, {0xF9B2, {0x96F6}}, {0xF9B3, {0x9748}}, {0xF9B4, {0x9818}}, {0xF9B5, {0x4F8B}}, {0xF9B6, {0x79AE}}, {0xF9B7, {0x91B4}}, {0xF9B8, {0x96B8}},
{0xF9B9, {0x60E1}}, {0xF9BA, {0x4E86}}, {0xF9BB, {0x50DA}}, {0xF9BC, {0x5BEE}}, {0xF9BD, {0x5C3F}}, {0xF9BE, {0x6599}}, {0xF9BF, {0x6A02}}, {0xF9C0, {0x71CE}}, {0xF9C1, {0x7642}}, {0xF9C2, {0x84FC}},
{0xF9C3, {0x907C}}, {0xF9C4, {0x9F8D}}, {0xF9C5, {0x6688}}, {0xF9C6, {0x962E}}, {0xF9C7, {0x5289}}, {0xF9C8, {0x677B}}, {0xF9C9, {0x67F3}}, {0xF9CA, {0x6D41}}, {0xF9CB, {0x6E9C}}, {0xF9CC, {0x7409}},
{0xF9CD, {0x7559}}, {0xF9CE, {0x786B}}, {0xF9CF, {0x7D10}}, {0xF9D0, {0x985E}}, {0xF9D1, {0x516D}}, {0xF9D2, {0x622E}}, {0xF9D3, {0x9678}}, {0xF9D4, {0x502B}}, {0xF9D5, {0x5D19}}, {0xF9D6, {0x6DEA}},
{0xF9D7, {0x8F2A}}, {0xF9D8, {0x5F8B}}, {0xF9D9, {0x6144}}, {0xF9DA, {0x6817}}, {0xF9DB, {0x7387}}, {0xF9DC, {0x9686}}, {0xF9DD, {0x5229}}, {0xF9DE, {0x540F}}, {0xF9DF, {0x5C65}}, {0xF9E0, {0x6613}},
{0xF9E1, {0x674E}}, {0xF9E2, {0x68A8}}, {0xF9E3, {0x6CE5}}, {0xF9E4, {0x7406}}, {0xF9E5, {0x75E2}}, {0xF9E6, {0x7F79}}, {0xF9E7, {0x88CF}}, {0xF9E8, {0x88E1}}, {0xF9E9, {0x91CC}}, {0xF9EA, {0x96E2}},
{0xF9EB, {0x533F}}, {0xF9EC, {0x6EBA}}, {0xF9ED, {0x541D}}, {0xF9EE, {0x71D0}}, {0xF9EF, {0x7498}}, {0xF9F0, {0x85FA}}, {0xF9F1, {0x96A3}}, {0xF9F2, {0x9C57}}, {0xF9F3, {0x9E9F}}, {0xF9F4, {0x6797}},
{0xF9F5, {0x6DCB}}, {0xF9F6, {0x81E8}}, {0xF9F7, {0x7ACB}}, {0xF9F8, {0x7B20}}, {0xF9F9, {0x7C92}}, {0xF9FA, {0x72C0}}, {0xF9FB, {0x7099}}, {0xF9FC, {0x8B58}}, {0xF9FD, {0x4EC0}}, {0xF9FE, {0x8336}},
{0xF9FF, {0x523A}}, {0xFA00, {0x5207}}, {0xFA01, {0x5EA6}}, {0xFA02, {0x62D3}}, {0xFA03, {0x7CD6}}, {0xFA04, {0x5B85}}, {0xFA05, {0x6D1E}}, {0xFA06, {0x66B4}}, {0xFA07, {0x8F3B}}, {0xFA08, {0x884C}},
{0xFA09, {0x964D}}, {0xFA0A, {0x898B}}, {0xFA0B, {0x5ED3}}, {0xFA0C, {0x5140}}, {0xFA0D, {0x55C0}}, {0xFA10, {0x585A}}, {0xFA12, {0x6674}}, {0xFA15, {0x51DE}}, {0xFA16, {0x732A}}, {0xFA17, {0x76CA}},
{0xFA18, {0x793C}}, {0xFA19, {0x795E}}, {0xFA1A, {0x7965}}, {0xFA1B, {0x798F}}, {0xFA1C, {0x9756}}, {0xFA1D, {0x7CBE}}, {0xFA1E, {0x7FBD}}, {0xFA20, {0x8612}}, {0xFA22, {0x8AF8}}, {0xFA25, {0x9038}},
{0xFA26, {0x90FD}}, {0xFA2A, {0x98EF}}, {0xFA2B, {0x98FC}}, {0xFA2C, {0x9928}}, {0xFA2D, {0x9DB4}}, {0xFA2E, {0x90DE}}, {0xFA2F, {0x96B7}}, {0xFA30, {0x4FAE}}, {0xFA31, {0x50E7}}, {0xFA32, {0x514D}},
{0xFA33, {0x52C9}}, {0xFA34, {0x52E4}}, {0xFA35, {0x5351}}, {0xFA36, {0x559D}}, {0xFA37, {0x5606}}, {0xFA38, {0x5668}}, {0xFA39, {0x5840}}, {0xFA3A, {0x58A8}}, {0xFA3B, {0x5C64}}, {0xFA3C, {0x5C6E}},
{0xFA3D, {0x6094}}, {0xFA3E, {0x6168}}, {0xFA3F, {0x618E}}, {0xFA40, {0x61F2}}, {0xFA41, {0x654F}}, {0xFA42, {0x65E2}}, {0xFA43, {0x6691}}, {0xFA44, {0x6885}}, {0xFA45, {0x6D77}}, {0xFA46, {0x6E1A}},
{0xFA47, {0x6F22}}, {0xFA48, {0x716E}}, {0xFA49, {0x722B}}, {0xFA4A, {0x7422}}, {0xFA4B, {0x7891}}, {0xFA4C, {0x793E}}, {0xFA4D, {0x7949}}, {0xFA4E, {0x7948}}, {0xFA4F, {0x7950}}, {0xFA50, {0x7956}},
{0xFA51, {0x795D}}, {0xFA52, {0x798D}}, {0xFA53, {0x798E}}, {0xFA54, {0x7A40}}, {0xFA55, {0x7A81}}, {0xFA56, {0x7BC0}}, {0xFA57, {0x7DF4}}, {0xFA58, {0x7E09}}, {0xFA59, {0x7E41}}, {0xFA5A, {0x7F72}},
{0xFA5B, {0x8005}}, {0xFA5C, {0x81ED}}, {0xFA5D, {0x8279}}, {0xFA5E, {0x8279}}, {0xFA5F, {0x8457}}, {0xFA60, {0x8910}}, {0xFA61, {0x8996}}, {0xFA62, {0x8B01}}, {0xFA63, {0x8B39}}, {0xFA64, {0x8CD3}},
{0xFA65, {0x8D08}}, {0xFA66, {0x8FB6}}, {0xFA67, {0x9038}}, {0xFA68, {0x96E3}}, {0xFA69, {0x97FF}}, {0xFA6A, {0x983B}}, {0xFA6B, {0x6075}}, {0xFA6C, {0x242EE}}, {0xFA6D, {0x8218}}, {0xFA70, {0x4E26}},
{0xFA71, {0x51B5}}, {0xFA72, {0x5168}}, {0xFA73, {0x4F80}}, {0xFA74, {0x5145}}, {0xFA75, {0x5180}}, {0xFA76, {0x52C7}}, {0xFA77, {0x52FA}}, {0xFA78, {0x559D}}, {0xFA79, {0x5555}}, {0xFA7A, {0x5599}},
{0xFA7B, {0x55E2}}, {0xFA7C, {0x585A}}, {0xFA7D, {0x58B3}}, {0xFA7E, {0x5944}}, {0xFA7F, {0x5954}}, {0xFA80, {0x5A62}}, {0xFA81, {0x5B28}}, {0xFA82, {0x5ED2}}, {0xFA83, {0x5ED9}}, {0xFA84, {0x5F69}},
{0xFA85, {0x5FAD}}, {0xFA86, {0x60D8}}, {0xFA87, {0x614E}}, {0xFA88, {0x6108}}, {0xFA89, {0x618E}}, {0xFA8A, {0x6160}}, {0xFA8B, {0x61F2}}, {0xFA8C, {0x6234}}, {0xFA8D, {0x63C4}}, {0xFA8E, {0x641C}},
{0xFA8F, {0x6452}}, {0xFA90, {0x6556}}, {0xFA91, {0x6674}}, {0xFA92, {0x6717}}, {0xFA93, {0x671B}}, {0xFA94, {0x6756}}, {0xFA95, {0x6B79}}, {0xFA96, {0x6BBA}}, {0xFA97, {0x6D41}}, {0xFA98, {0x6EDB}},
{0xFA99, {0x6ECB}}, {0xFA9A, {0x6F22}}, {0xFA9B, {0x701E}}, {0xFA9C, {0x716E}}, {0xFA9D, {0x77A7}}, {0xFA9E, {0x7235}}, {0xFA9F, {0x72AF}}, {0xFAA0, {0x732A}}, {0xFAA1, {0x7471}}, {0xFAA2, {0x7506}},
{0xFAA3, {0x753B}}, {0xFAA4, {0x761D}}, {0xFAA5, {0x761F}}, {0xFAA6, {0x76CA}}, {0xFAA7, {0x76DB}}, {0xFAA8, {0x76F4}}, {0xFAA9, {0x774A}}, {0xFAAA, {0x7740}}, {0xFAAB, {0x78CC}}, {0xFAAC, {0x7AB1}},
{0xFAAD, {0x7BC0}}, {0xFAAE, {0x7C7B}}, {0xFAAF, {0x7D5B}}, {0xFAB0, {0x7DF4}}, {0xFAB1, {0x7F3E}}, {0xFAB2, {0x8005}}, {0xFAB3, {0x8352}}, {0xFAB4, {0x83EF}}, {0xFAB5, {0x8779}}, {0xFAB6, {0x8941}},
{0xFAB7, {0x8986}}, {0xFAB8, {0x8996}}, {0xFAB9, {0x8ABF}}, {0xFABA, {0x8AF8}}, {0xFABB, {0x8ACB}}, {0xFABC, {0x8B01}}, {0xFABD, {0x8AFE}}, {0xFABE, {0x8AED}}, {0xFABF, {0x8B39}}, {0xFAC0, {0x8B8A}},
{0xFAC1, {0x8D08}}, {0xFAC2, {0x8F38}}, {0xFAC3, {0x9072}}, {0xFAC4, {0x9199}}, {0xFAC5, {0x9276}}, {0xFAC6, {0x967C}}, {0xFAC7, {0x96E3}}, {0xFAC8, {0x9756}}, {0xFAC9, {0x97DB}}, {0xFACA, {0x97FF}},
{0xFACB, {0x980B}}, {0xFACC, {0x983B}}, {0xFACD, {0x9B12}}, {0xFACE, {0x9F9C}}, {0xFACF, {0x2284A}}, {0xFAD0, {0x22844}}, {0xFAD1, {0x233D5}}, {0xFAD2, {0x3B9D}}, {0xFAD3, {0x4018}},
{0xFAD4, {0x4039}}, {0xFAD5, {0x25249}}, {0xFAD6, {0x25CD0}}, {0xFAD7, {0x27ED3}}, {0xFAD8, {0x9F43}}, {0xFAD9, {0x9F8E}}, {0xFB1D, {0x5D9, 0x5B4}}, {0xFB1F, {0x5F2, 0x5B7}}, {0xFB2A, {0x5E9, 0x5C1}},
{0xFB2B, {0x5E9, 0x5C2}}, {0xFB2C, {0x5E9, 0x5BC, 0x5C1}}, {0xFB2D, {0x5E9, 0x5BC, 0x5C2}}, {0xFB2E, {0x5D0, 0x5B7}}, {0xFB2F, {0x5D0, 0x5B8}}, {0xFB30, {0x5D0, 0x5BC}}, {0xFB31, {0x5D1, 0x5BC}},
{0xFB32, {0x5D2, 0x5BC}}, {0xFB33, {0x5D3, 0x5BC}}, {0xFB34, {0x5D4, 0x5BC}}, {0xFB35, {0x5D5, 0x5BC}}, {0xFB36, {0x5D6, 0x5BC}}, {0xFB38, {0x5D8, 0x5BC}}, {0xFB39, {0x5D9, 0x5BC}},
{0xFB3A, {0x5DA, 0x5BC}}, {0xFB3B, {0x5DB, 0x5BC}}, {0xFB3C, {0x5DC, 0x5BC}}, {0xFB3E, {0x5DE, 0x5BC}}, {0xFB40, {0x5E0, 0x5BC}}, {0xFB41, {0x5E1, 0x5BC}}, {0xFB43, {0x5E3, 0x5BC}},
{0xFB44, {0x5E4, 0x5BC}}, {0xFB46, {0x5E6, 0x5BC}}, {0xFB47, {0x5E7, 0x5BC}}, {0xFB48, {0x5E8, 0x5BC}}, {0xFB49, {0x5E9, 0x5BC}}, {0xFB4A, {0x5EA, 0x5BC}}, {0xFB4B, {0x5D5, 0x5B9}},
{0xFB4C, {0x5D1, 0x5BF}}, {0xFB4D, {0x5DB, 0x5BF}}, {0xFB4E, {0x5E4, 0x5BF}}, {0x1109A, {0x11099, 0x110BA}}, {0x1109C, {0x1109B, 0x110BA}}, {0x110AB, {0x110A5, 0x110BA}},
{0x1112E, {0x11131, 0x11127}}, {0x1112F, {0x11132, 0x11127}}, {0x1134B, {0x11347, 0x1133E}}, {0x1134C, {0x11347, 0x11357}}, {0x114BB, {0x114B9, 0x114BA}}, {0x114BC, {0x114B9, 0x114B0}},
{0x114BE, {0x114B9, 0x114BD}}, {0x115BA, {0x115B8, 0x115AF}}, {0x115BB, {0x115B9, 0x115AF}}, {0x1D15E, {0x1D157, 0x1D165}}, {0x1D15F, {0x1D158, 0x1D165}}, {0x1D160, {0x1D158, 0x1D165, 0x1D16E}},
{0x1D161, {0x1D158, 0x1D165, 0x1D16F}}, {0x1D162, {0x1D158, 0x1D165, 0x1D170}}, {0x1D163, {0x1D158, 0x1D165, 0x1D171}}, {0x1D164, {0x1D158, 0x1D165, 0x1D172}}, {0x1D1BB, {0x1D1B9, 0x1D165}},
{0x1D1BC, {0x1D1BA, 0x1D165}}, {0x1D1BD, {0x1D1B9, 0x1D165, 0x1D16E}}, {0x1D1BE, {0x1D1BA, 0x1D165, 0x1D16E}}, {0x1D1BF, {0x1D1B9, 0x1D165, 0x1D16F}}, {0x1D1C0, {0x1D1BA, 0x1D165, 0x1D16F}},
{0x2F800, {0x4E3D}}, {0x2F801, {0x4E38}}, {0x2F802, {0x4E41}}, {0x2F803, {0x20122}}, {0x2F804, {0x4F60}}, {0x2F805, {0x4FAE}}, {0x2F806, {0x4FBB}}, {0x2F807, {0x5002}}, {0x2F808, {0x507A}},
{0x2F809, {0x5099}}, {0x2F80A, {0x50E7}}, {0x2F80B, {0x50CF}}, {0x2F80C, {0x349E}}, {0x2F80D, {0x2063A}}, {0x2F80E, {0x514D}}, {0x2F80F, {0x5154}}, {0x2F810, {0x5164}}, {0x2F811, {0x5177}},
{0x2F812, {0x2051C}}, {0x2F813, {0x34B9}}, {0x2F814, {0x5167}}, {0x2F815, {0x518D}}, {0x2F816, {0x2054B}}, {0x2F817, {0x5197}}, {0x2F818, {0x51A4}}, {0x2F819, {0x4ECC}}, {0x2F81A, {0x51AC}},
{0x2F81B, {0x51B5}}, {0x2F81C, {0x291DF}}, {0x2F81D, {0x51F5}}, {0x2F81E, {0x5203}}, {0x2F81F, {0x34DF}}, {0x2F820, {0x523B}}, {0x2F821, {0x5246}}, {0x2F822, {0x5272}}, {0x2F823, {0x5277}},
{0x2F824, {0x3515}}, {0x2F825, {0x52C7}}, {0x2F826, {0x52C9}}, {0x2F827, {0x52E4}}, {0x2F828, {0x52FA}}, {0x2F829, {0x5305}}, {0x2F82A, {0x5306}}, {0x2F82B, {0x5317}}, {0x2F82C, {0x5349}},
{0x2F82D, {0x5351}}, {0x2F82E, {0x535A}}, {0x2F82F, {0x5373}}, {0x2F830, {0x537D}}, {0x2F831, {0x537F}}, {0x2F832, {0x537F}}, {0x2F833, {0x537F}}, {0x2F834, {0x20A2C}}, {0x2F835, {0x7070}},
{0x2F836, {0x53CA}}, {0x2F837, {0x53DF}}, {0x2F838, {0x20B63}}, {0x2F839, {0x53EB}}, {0x2F83A, {0x53F1}}, {0x2F83B, {0x5406}}, {0x2F83C, {0x549E}}, {0x2F83D, {0x5438}}, {0x2F83E, {0x5448}},
{0x2F83F, {0x5468}}, {0x2F840, {0x54A2}}, {0x2F841, {0x54F6}}, {0x2F842, {0x5510}}, {0x2F843, {0x5553}}, {0x2F844, {0x5563}}, {0x2F845, {0x5584}}, {0x2F846, {0x5584}}, {0x2F847, {0x5599}},
{0x2F848, {0x55AB}}, {0x2F849, {0x55B3}}, {0x2F84A, {0x55C2}}, {0x2F84B, {0x5716}}, {0x2F84C, {0x5606}}, {0x2F84D, {0x5717}}, {0x2F84E, {0x5651}}, {0x2F84F, {0x5674}}, {0x2F850, {0x5207}},
{0x2F851, {0x58EE}}, {0x2F852, {0x57CE}}, {0x2F853, {0x57F4}}, {0x2F854, {0x580D}}, {0x2F855, {0x578B}}, {0x2F856, {0x5832}}, {0x2F857, {0x5831}}, {0x2F858, {0x58AC}}, {0x2F859, {0x214E4}},
{0x2F85A, {0x58F2}}, {0x2F85B, {0x58F7}}, {0x2F85C, {0x5906}}, {0x2F85D, {0x591A}}, {0x2F85E, {0x5922}}, {0x2F85F, {0x5962}}, {0x2F860, {0x216A8}}, {0x2F861, {0x216EA}}, {0x2F862, {0x59EC}},
{0x2F863, {0x5A1B}}, {0x2F864, {0x5A27}}, {0x2F865, {0x59D8}}, {0x2F866, {0x5A66}}, {0x2F867, {0x36EE}}, {0x2F868, {0x36FC}}, {0x2F869, {0x5B08}}, {0x2F86A, {0x5B3E}}, {0x2F86B, {0x5B3E}},
{0x2F86C, {0x219C8}}, {0x2F86D, {0x5BC3}}, {0x2F86E, {0x5BD8}}, {0x2F86F, {0x5BE7}}, {0x2F870, {0x5BF3}}, {0x2F871, {0x21B18}}, {0x2F872, {0x5BFF}}, {0x2F873, {0x5C06}}, {0x2F874, {0x5F53}},
{0x2F875, {0x5C22}}, {0x2F876, {0x3781}}, {0x2F877, {0x5C60}}, {0x2F878, {0x5C6E}}, {0x2F879, {0x5CC0}}, {0x2F87A, {0x5C8D}}, {0x2F87B, {0x21DE4}}, {0x2F87C, {0x5D43}}, {0x2F87D, {0x21DE6}},
{0x2F87E, {0x5D6E}}, {0x2F87F, {0x5D6B}}, {0x2F880, {0x5D7C}}, {0x2F881, {0x5DE1}}, {0x2F882, {0x5DE2}}, {0x2F883, {0x382F}}, {0x2F884, {0x5DFD}}, {0x2F885, {0x5E28}}, {0x2F886, {0x5E3D}},
{0x2F887, {0x5E69}}, {0x2F888, {0x3862}}, {0x2F889, {0x22183}}, {0x2F88A, {0x387C}}, {0x2F88B, {0x5EB0}}, {0x2F88C, {0x5EB3}}, {0x2F88D, {0x5EB6}}, {0x2F88E, {0x5ECA}}, {0x2F88F, {0x2A392}},
{0x2F890, {0x5EFE}}, {0x2F891, {0x22331}}, {0x2F892, {0x22331}}, {0x2F893, {0x8201}}, {0x2F894, {0x5F22}}, {0x2F895, {0x5F22}}, {0x2F896, {0x38C7}}, {0x2F897, {0x232B8}}, {0x2F898, {0x261DA}},
{0x2F899, {0x5F62}}, {0x2F89A, {0x5F6B}}, {0x2F89B, {0x38E3}}, {0x2F89C, {0x5F9A}}, {0x2F89D, {0x5FCD}}, {0x2F89E, {0x5FD7}}, {0x2F89F, {0x5FF9}}, {0x2F8A0, {0x6081}}, {0x2F8A1, {0x393A}},
{0x2F8A2, {0x391C}}, {0x2F8A3, {0x6094}}, {0x2F8A4, {0x226D4}}, {0x2F8A5, {0x60C7}}, {0x2F8A6, {0x6148}}, {0x2F8A7, {0x614C}}, {0x2F8A8, {0x614E}}, {0x2F8A9, {0x614C}}, {0x2F8AA, {0x617A}},
{0x2F8AB, {0x618E}}, {0x2F8AC, {0x61B2}}, {0x2F8AD, {0x61A4}}, {0x2F8AE, {0x61AF}}, {0x2F8AF, {0x61DE}}, {0x2F8B0, {0x61F2}}, {0x2F8B1, {0x61F6}}, {0x2F8B2, {0x6210}}, {0x2F8B3, {0x621B}},
{0x2F8B4, {0x625D}}, {0x2F8B5, {0x62B1}}, {0x2F8B6, {0x62D4}}, {0x2F8B7, {0x6350}}, {0x2F8B8, {0x22B0C}}, {0x2F8B9, {0x633D}}, {0x2F8BA, {0x62FC}}, {0x2F8BB, {0x6368}}, {0x2F8BC, {0x6383}},
{0x2F8BD, {0x63E4}}, {0x2F8BE, {0x22BF1}}, {0x2F8BF, {0x6422}}, {0x2F8C0, {0x63C5}}, {0x2F8C1, {0x63A9}}, {0x2F8C2, {0x3A2E}}, {0x2F8C3, {0x6469}}, {0x2F8C4, {0x647E}}, {0x2F8C5, {0x649D}},
{0x2F8C6, {0x6477}}, {0x2F8C7, {0x3A6C}}, {0x2F8C8, {0x654F}}, {0x2F8C9, {0x656C}}, {0x2F8CA, {0x2300A}}, {0x2F8CB, {0x65E3}}, {0x2F8CC, {0x66F8}}, {0x2F8CD, {0x6649}}, {0x2F8CE, {0x3B19}},
{0x2F8CF, {0x6691}}, {0x2F8D0, {0x3B08}}, {0x2F8D1, {0x3AE4}}, {0x2F8D2, {0x5192}}, {0x2F8D3, {0x5195}}, {0x2F8D4, {0x6700}}, {0x2F8D5, {0x669C}}, {0x2F8D6, {0x80AD}}, {0x2F8D7, {0x43D9}},
{0x2F8D8, {0x6717}}, {0x2F8D9, {0x671B}}, {0x2F8DA, {0x6721}}, {0x2F8DB, {0x675E}}, {0x2F8DC, {0x6753}}, {0x2F8DD, {0x233C3}}, {0x2F8DE, {0x3B49}}, {0x2F8DF, {0x67FA}}, {0x2F8E0, {0x6785}},
{0x2F8E1, {0x6852}}, {0x2F8E2, {0x6885}}, {0x2F8E3, {0x2346D}}, {0x2F8E4, {0x688E}}, {0x2F8E5, {0x681F}}, {0x2F8E6, {0x6914}}, {0x2F8E7, {0x3B9D}}, {0x2F8E8, {0x6942}}, {0x2F8E9, {0x69A3}},
{0x2F8EA, {0x69EA}}, {0x2F8EB, {0x6AA8}}, {0x2F8EC, {0x236A3}}, {0x2F8ED, {0x6ADB}}, {0x2F8EE, {0x3C18}}, {0x2F8EF, {0x6B21}}, {0x2F8F0, {0x238A7}}, {0x2F8F1, {0x6B54}}, {0x2F8F2, {0x3C4E}},
{0x2F8F3, {0x6B72}}, {0x2F8F4, {0x6B9F}}, {0x2F8F5, {0x6BBA}}, {0x2F8F6, {0x6BBB}}, {0x2F8F7, {0x23A8D}}, {0x2F8F8, {0x21D0B}}, {0x2F8F9, {0x23AFA}}, {0x2F8FA, {0x6C4E}}, {0x2F8FB, {0x23CBC}},
{0x2F8FC, {0x6CBF}}, {0x2F8FD, {0x6CCD}}, {0x2F8FE, {0x6C67}}, {0x2F8FF, {0x6D16}}, {0x2F900, {0x6D3E}}, {0x2F901, {0x6D77}}, {0x2F902, {0x6D41}}, {0x2F903, {0x6D69}}, {0x2F904, {0x6D78}},
{0x2F905, {0x6D85}}, {0x2F906, {0x23D1E}}, {0x2F907, {0x6D34}}, {0x2F908, {0x6E2F}}, {0x2F909, {0x6E6E}}, {0x2F90A, {0x3D33}}, {0x2F90B, {0x6ECB}}, {0x2F90C, {0x6EC7}}, {0x2F90D, {0x23ED1}},
{0x2F90E, {0x6DF9}}, {0x2F90F, {0x6F6E}}, {0x2F910, {0x23F5E}}, {0x2F911, {0x23F8E}}, {0x2F912, {0x6FC6}}, {0x2F913, {0x7039}}, {0x2F914, {0x701E}}, {0x2F915, {0x701B}}, {0x2F916, {0x3D96}},
{0x2F917, {0x704A}}, {0x2F918, {0x707D}}, {0x2F919, {0x7077}}, {0x2F91A, {0x70AD}}, {0x2F91B, {0x20525}}, {0x2F91C, {0x7145}}, {0x2F91D, {0x24263}}, {0x2F91E, {0x719C}}, {0x2F91F, {0x243AB}},
{0x2F920, {0x7228}}, {0x2F921, {0x7235}}, {0x2F922, {0x7250}}, {0x2F923, {0x24608}}, {0x2F924, {0x7280}}, {0x2F925, {0x7295}}, {0x2F926, {0x24735}}, {0x2F927, {0x24814}}, {0x2F928, {0x737A}},
{0x2F929, {0x738B}}, {0x2F92A, {0x3EAC}}, {0x2F92B, {0x73A5}}, {0x2F92C, {0x3EB8}}, {0x2F92D, {0x3EB8}}, {0x2F92E, {0x7447}}, {0x2F92F, {0x745C}}, {0x2F930, {0x7471}}, {0x2F931, {0x7485}},
{0x2F932, {0x74CA}}, {0x2F933, {0x3F1B}}, {0x2F934, {0x7524}}, {0x2F935, {0x24C36}}, {0x2F936, {0x753E}}, {0x2F937, {0x24C92}}, {0x2F938, {0x7570}}, {0x2F939, {0x2219F}}, {0x2F93A, {0x7610}},
{0x2F93B, {0x24FA1}}, {0x2F93C, {0x24FB8}}, {0x2F93D, {0x25044}}, {0x2F93E, {0x3FFC}}, {0x2F93F, {0x4008}}, {0x2F940, {0x76F4}}, {0x2F941, {0x250F3}}, {0x2F942, {0x250F2}}, {0x2F943, {0x25119}},
{0x2F944, {0x25133}}, {0x2F945, {0x771E}}, {0x2F946, {0x771F}}, {0x2F947, {0x771F}}, {0x2F948, {0x774A}}, {0x2F949, {0x4039}}, {0x2F94A, {0x778B}}, {0x2F94B, {0x4046}}, {0x2F94C, {0x4096}},
{0x2F94D, {0x2541D}}, {0x2F94E, {0x784E}}, {0x2F94F, {0x788C}}, {0x2F950, {0x78CC}}, {0x2F951, {0x40E3}}, {0x2F952, {0x25626}}, {0x2F953, {0x7956}}, {0x2F954, {0x2569A}}, {0x2F955, {0x256C5}},
{0x2F956, {0x798F}}, {0x2F957, {0x79EB}}, {0x2F958, {0x412F}}, {0x2F959, {0x7A40}}, {0x2F95A, {0x7A4A}}, {0x2F95B, {0x7A4F}}, {0x2F95C, {0x2597C}}, {0x2F95D, {0x25AA7}}, {0x2F95E, {0x25AA7}},
{0x2F95F, {0x7AEE}}, {0x2F960, {0x4202}}, {0x2F961, {0x25BAB}}, {0x2F962, {0x7BC6}}, {0x2F963, {0x7BC9}}, {0x2F964, {0x4227}}, {0x2F965, {0x25C80}}, {0x2F966, {0x7CD2}}, {0x2F967, {0x42A0}},
{0x2F968, {0x7CE8}}, {0x2F969, {0x7CE3}}, {0x2F96A, {0x7D00}}, {0x2F96B, {0x25F86}}, {0x2F96C, {0x7D63}}, {0x2F96D, {0x4301}}, {0x2F96E, {0x7DC7}}, {0x2F96F, {0x7E02}}, {0x2F970, {0x7E45}},
{0x2F971, {0x4334}}, {0x2F972, {0x26228}}, {0x2F973, {0x26247}}, {0x2F974, {0x4359}}, {0x2F975, {0x262D9}}, {0x2F976, {0x7F7A}}, {0x2F977, {0x2633E}}, {0x2F978, {0x7F95}}, {0x2F979, {0x7FFA}},
{0x2F97A, {0x8005}}, {0x2F97B, {0x264DA}}, {0x2F97C, {0x26523}}, {0x2F97D, {0x8060}}, {0x2F97E, {0x265A8}}, {0x2F97F, {0x8070}}, {0x2F980, {0x2335F}}, {0x2F981, {0x43D5}}, {0x2F982, {0x80B2}},
{0x2F983, {0x8103}}, {0x2F984, {0x440B}}, {0x2F985, {0x813E}}, {0x2F986, {0x5AB5}}, {0x2F987, {0x267A7}}, {0x2F988, {0x267B5}}, {0x2F989, {0x23393}}, {0x2F98A, {0x2339C}}, {0x2F98B, {0x8201}},
{0x2F98C, {0x8204}}, {0x2F98D, {0x8F9E}}, {0x2F98E, {0x446B}}, {0x2F98F, {0x8291}}, {0x2F990, {0x828B}}, {0x2F991, {0x829D}}, {0x2F992, {0x52B3}}, {0x2F993, {0x82B1}}, {0x2F994, {0x82B3}},
{0x2F995, {0x82BD}}, {0x2F996, {0x82E6}}, {0x2F997, {0x26B3C}}, {0x2F998, {0x82E5}}, {0x2F999, {0x831D}}, {0x2F99A, {0x8363}}, {0x2F99B, {0x83AD}}, {0x2F99C, {0x8323}}, {0x2F99D, {0x83BD}},
{0x2F99E, {0x83E7}}, {0x2F99F, {0x8457}}, {0x2F9A0, {0x8353}}, {0x2F9A1, {0x83CA}}, {0x2F9A2, {0x83CC}}, {0x2F9A3, {0x83DC}}, {0x2F9A4, {0x26C36}}, {0x2F9A5, {0x26D6B}}, {0x2F9A6, {0x26CD5}},
{0x2F9A7, {0x452B}}, {0x2F9A8, {0x84F1}}, {0x2F9A9, {0x84F3}}, {0x2F9AA, {0x8516}}, {0x2F9AB, {0x273CA}}, {0x2F9AC, {0x8564}}, {0x2F9AD, {0x26F2C}}, {0x2F9AE, {0x455D}}, {0x2F9AF, {0x4561}},
{0x2F9B0, {0x26FB1}}, {0x2F9B1, {0x270D2}}, {0x2F9B2, {0x456B}}, {0x2F9B3, {0x8650}}, {0x2F9B4, {0x865C}}, {0x2F9B5, {0x8667}}, {0x2F9B6, {0x8669}}, {0x2F9B7, {0x86A9}}, {0x2F9B8, {0x8688}},
{0x2F9B9, {0x870E}}, {0x2F9BA, {0x86E2}}, {0x2F9BB, {0x8779}}, {0x2F9BC, {0x8728}}, {0x2F9BD, {0x876B}}, {0x2F9BE, {0x8786}}, {0x2F9BF, {0x45D7}}, {0x2F9C0, {0x87E1}}, {0x2F9C1, {0x8801}},
{0x2F9C2, {0x45F9}}, {0x2F9C3, {0x8860}}, {0x2F9C4, {0x8863}}, {0x2F9C5, {0x27667}}, {0x2F9C6, {0x88D7}}, {0x2F9C7, {0x88DE}}, {0x2F9C8, {0x4635}}, {0x2F9C9, {0x88FA}}, {0x2F9CA, {0x34BB}},
{0x2F9CB, {0x278AE}}, {0x2F9CC, {0x27966}}, {0x2F9CD, {0x46BE}}, {0x2F9CE, {0x46C7}}, {0x2F9CF, {0x8AA0}}, {0x2F9D0, {0x8AED}}, {0x2F9D1, {0x8B8A}}, {0x2F9D2, {0x8C55}}, {0x2F9D3, {0x27CA8}},
{0x2F9D4, {0x8CAB}}, {0x2F9D5, {0x8CC1}}, {0x2F9D6, {0x8D1B}}, {0x2F9D7, {0x8D77}}, {0x2F9D8, {0x27F2F}}, {0x2F9D9, {0x20804}}, {0x2F9DA, {0x8DCB}}, {0x2F9DB, {0x8DBC}}, {0x2F9DC, {0x8DF0}},
{0x2F9DD, {0x208DE}}, {0x2F9DE, {0x8ED4}}, {0x2F9DF, {0x8F38}}, {0x2F9E0, {0x285D2}}, {0x2F9E1, {0x285ED}}, {0x2F9E2, {0x9094}}, {0x2F9E3, {0x90F1}}, {0x2F9E4, {0x9111}}, {0x2F9E5, {0x2872E}},
{0x2F9E6, {0x911B}}, {0x2F9E7, {0x9238}}, {0x2F9E8, {0x92D7}}, {0x2F9E9, {0x92D8}}, {0x2F9EA, {0x927C}}, {0x2F9EB, {0x93F9}}, {0x2F9EC, {0x9415}}, {0x2F9ED, {0x28BFA}}, {0x2F9EE, {0x958B}},
{0x2F9EF, {0x4995}}, {0x2F9F0, {0x95B7}}, {0x2F9F1, {0x28D77}}, {0x2F9F2, {0x49E6}}, {0x2F9F3, {0x96C3}}, {0x2F9F4, {0x5DB2}}, {0x2F9F5, {0x9723}}, {0x2F9F6, {0x29145}}, {0x2F9F7, {0x2921A}},
{0x2F9F8, {0x4A6E}}, {0x2F9F9, {0x4A76}}, {0x2F9FA, {0x97E0}}, {0x2F9FB, {0x2940A}}, {0x2F9FC, {0x4AB2}}, {0x2F9FD, {0x29496}}, {0x2F9FE, {0x980B}}, {0x2F9FF, {0x980B}}, {0x2FA00, {0x9829}},
{0x2FA01, {0x295B6}}, {0x2FA02, {0x98E2}}, {0x2FA03, {0x4B33}}, {0x2FA04, {0x9929}}, {0x2FA05, {0x99A7}}, {0x2FA06, {0x99C2}}, {0x2FA07, {0x99FE}}, {0x2FA08, {0x4BCE}}, {0x2FA09, {0x29B30}},
{0x2FA0A, {0x9B12}}, {0x2FA0B, {0x9C40}}, {0x2FA0C, {0x9CFD}}, {0x2FA0D, {0x4CCE}}, {0x2FA0E, {0x4CED}}, {0x2FA0F, {0x9D67}}, {0x2FA10, {0x2A0CE}}, {0x2FA11, {0x4CF8}}, {0x2FA12, {0x2A105}},
{0x2FA13, {0x2A20E}}, {0x2FA14, {0x2A291}}, {0x2FA15, {0x9EBB}}, {0x2FA16, {0x4D56}}, {0x2FA17, {0x9EF9}}, {0x2FA18, {0x9EFE}}, {0x2FA19, {0x9F05}}, {0x2FA1A, {0x9F0F}}, {0x2FA1B, {0x9F16}},
{0x2FA1D, {0x2A600}},
};
static std::string codepoint_to_utf8(uint32_t cp) {
std::string result;
if (/* 0x00 <= cp && */ cp <= 0x7f) {
@ -404,7 +666,8 @@ static std::unordered_map<uint32_t, int> codepoint_type_map() {
static int codepoint_type(uint32_t cp) {
static std::unordered_map<uint32_t, int> codepoint_types = codepoint_type_map();
return codepoint_types.find(cp) == codepoint_types.end() ? CODEPOINT_TYPE_UNIDENTIFIED : codepoint_types.at(cp);
const auto it = codepoint_types.find(cp);
return it == codepoint_types.end() ? CODEPOINT_TYPE_UNIDENTIFIED : it->second;
}
static int codepoint_type(const std::string & utf8) {