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

Merge branch 'master' into concedo_experimental

Browse source 
# Conflicts:
#	CMakeLists.txt
#	Makefile
#	README.md
#	pocs/vdot/vdot.cpp
#	scripts/verify-checksum-models.py
#	tests/test-quantize-fns.cpp
#	tests/test-quantize-perf.cpp
#	tests/test-sampling.cpp
#	tests/test-tokenizer-0.cpp
This commit is contained in:
Concedo 2023-06-17 20:02:32 +08:00
commit 9f8e2f8a18
25 changed files with 1829 additions and 1252 deletions
Download patch file Download diff file Expand all files Collapse all files

2
.flake8 Normal file
View file

@ -0,0 +1,2 @@
[flake8]
max-line-length = 125

3
.gitignore vendored
View file

@ -22,6 +22,7 @@ build-metal/
build-no-accel/
build-sanitize-addr/
build-sanitize-thread/
out/
/main
/quantize
@ -32,11 +33,13 @@ build-sanitize-thread/
/train-text-from-scratch
/benchmark-matmult
/vdot
/server
/Pipfile
/libllama.so
arm_neon.h
compile_commands.json
CMakeSettings.json
__pycache__

15
.pre-commit-config.yaml Normal file
View file

@ -0,0 +1,15 @@
# See https://pre-commit.com for more information
# See https://pre-commit.com/hooks.html for more hooks
exclude: prompts/.*.txt
repos:
- repo: https://github.com/pre-commit/pre-commit-hooks
rev: v3.2.0
hooks:
- id: trailing-whitespace
- id: end-of-file-fixer
- id: check-yaml
- id: check-added-large-files
- repo: https://github.com/PyCQA/flake8
rev: 6.0.0
hooks:
- id: flake8

26
convert.py
View file

@ -512,7 +512,11 @@ class LazyTensor:
if not isinstance(self.data_type, QuantizedDataType):
raise Exception(f"Can't turn an unquantized tensor into a quantized type ({data_type})")
if self.data_type.have_g_idx:
sys.stderr.write("Error: Input uses the newer GPTQ-for-LLaMa format (using g_idx), which is not yet natively supported by GGML. For now you can still convert this model by passing `--outtype f16` to dequantize, but that will result in a much larger output file for no quality benefit.\n")
sys.stderr.write(
"Error: Input uses the newer GPTQ-for-LLaMa format (using g_idx), "
"which is not yet natively supported by GGML. "
"For now you can still convert this model by passing `--outtype f16` to dequantize, "
"but that will result in a much larger output file for no quality benefit.\n")
sys.exit(1)
assert not data_type.have_g_idx and self.data_type.have_addends and data_type.have_addends
@ -694,8 +698,9 @@ class LazyUnpickler(pickle.Unpickler):
description = f'storage data_type={data_type} path-in-zip={filename} path={self.zip_file.filename}'
return LazyStorage(load=load, kind=pid[1], description=description)
# @staticmethod
def lazy_rebuild_tensor_v2(storage: Any, storage_offset: Any, size: Any, stride: Any, # pyright: ignore[reportSelfClsParameterName]
# @staticmethod
def lazy_rebuild_tensor_v2(storage: Any, storage_offset: Any, size: Any, stride: Any,
# pyright: ignore[reportSelfClsParameterName]
requires_grad: Any, backward_hooks: Any, metadata: Any = None) -> LazyTensor:
assert isinstance(storage, LazyStorage)
@ -812,7 +817,7 @@ def lazy_load_ggml_file(fp: io.BufferedReader, path: Path) -> ModelPlus:
# Use mmap for the actual data to avoid race conditions with the file offset.
off = fp.raw.tell()
mapped = memoryview(mmap.mmap(fp.fileno(), 0, access=mmap.ACCESS_READ))
fp.raw.seek(off) # needed on Windows
fp.raw.seek(off) # needed on Windows
def read_tensor() -> None: # this is a function so that variables captured in `load` don't change
shape_len, name_len, ftype = struct.unpack("iii", must_read(fp, 12))
@ -1054,7 +1059,7 @@ def load_some_model(path: Path) -> ModelPlus:
files = list(path.glob("model-00001-of-*.safetensors"))
if not files:
# Try the PyTorch patterns too, with lower priority
globs = ["consolidated.00.pth", "pytorch_model-00001-of-*.bin", "*.pt", "pytorch_model.bin" ]
globs = ["consolidated.00.pth", "pytorch_model-00001-of-*.bin", "*.pt", "pytorch_model.bin"]
files = [file for glob in globs for file in path.glob(glob)]
if not files:
# Try GGML too, but with lower priority, since if both a non-GGML
@ -1094,7 +1099,9 @@ def load_vocab(path: Path) -> SentencePieceVocab:
elif path3.exists():
path = path3
else:
raise FileNotFoundError(f"Could not find tokenizer.model in {path} or its parent; if it's in another directory, pass the directory as --vocab-dir")
raise FileNotFoundError(
f"Could not find tokenizer.model in {path} or its parent; "
"if it's in another directory, pass the directory as --vocab-dir")
added_tokens_path = path.parent / "added_tokens.json"
print(f"Loading vocab file {path}")
return SentencePieceVocab(path, added_tokens_path if added_tokens_path.exists() else None)
@ -1110,7 +1117,9 @@ def default_outfile(model_paths: List[Path], params: Params) -> Path:
}[params.file_type]
ret = model_paths[0].parent / f"ggml-model-{namestr}.bin"
if ret in model_paths:
sys.stderr.write(f"Error: Default output path ({ret}) would overwrite the input. Please explicitly specify a path using --outfile.\n")
sys.stderr.write(
f"Error: Default output path ({ret}) would overwrite the input. "
"Please explicitly specify a path using --outfile.\n")
sys.exit(1)
return ret
@ -1131,7 +1140,8 @@ def main(args_in: Optional[List[str]] = None) -> None:
parser.add_argument("--outtype", choices=["f32", "f16", "q4_1", "q4_0"], help="output format (default: based on input)")
parser.add_argument("--vocab-dir", type=Path, help="directory containing tokenizer.model, if separate from model file")
parser.add_argument("--outfile", type=Path, help="path to write to; default: based on input")
parser.add_argument("model", type=Path, help="directory containing model file, or model file itself (*.pth, *.pt, *.bin)")
parser.add_argument("model", type=Path,
help="directory containing model file, or model file itself (*.pth, *.pt, *.bin)")
args = parser.parse_args(args_in)
vocab: Vocab

6
examples/baby-llama/baby-llama.cpp
View file

@ -4,6 +4,10 @@
#include <random>
#include <cstring>
#if defined(_MSC_VER)
#pragma warning(disable: 4244 4267) // possible loss of data
#endif
float frand() {
return (float)rand()/(float)RAND_MAX;
}
@ -1470,7 +1474,7 @@ struct ggml_tensor * square_error_loss(struct ggml_context * ctx, struct ggml_te
}
struct ggml_tensor * cross_entropy_loss(struct ggml_context * ctx, struct ggml_tensor * a, struct ggml_tensor * b) {
const float eps = 1e-3;
const float eps = 1e-3f;
return
ggml_sum(ctx,
ggml_neg(ctx,

10
examples/benchmark/benchmark-matmult.cpp
View file

@ -16,6 +16,10 @@
#include <iterator>
#include <algorithm>
#if defined(_MSC_VER)
#pragma warning(disable: 4244 4267) // possible loss of data
#endif
float tensor_sum_elements(const ggml_tensor * tensor) {
float sum = 0;
if (tensor->type==GGML_TYPE_F32) {
@ -29,9 +33,9 @@ float tensor_sum_elements(const ggml_tensor * tensor) {
}
void tensor_dump(const ggml_tensor * tensor, const char * name) {
printf("%15s: type = %i (%5s) ne = %5d x %5d x %5d, nb = (%5li, %5li, %5li) - ", name,
printf("%15s: type = %i (%5s) ne = %5" PRIi64 " x %5" PRIi64 " x %5" PRIi64 ", nb = (%5zi, %5zi, %5zi) - ", name,
tensor->type, ggml_type_name(tensor->type),
(int) tensor->ne[0], (int) tensor->ne[1], (int) tensor->ne[2], tensor->nb[0], tensor->nb[1], tensor->nb[2]);
tensor->ne[0], tensor->ne[1], tensor->ne[2], tensor->nb[0], tensor->nb[1], tensor->nb[2]);
float sum = tensor_sum_elements(tensor);
printf("Sum of tensor %s is %6.2f\n", name, sum);
}
@ -120,7 +124,7 @@ int main(int argc, char ** argv) {
ctx_size += sizex*sizey*ggml_type_sizef(GGML_TYPE_F32); // BLAS
ctx_size += 1024*1024*16;
printf("Allocating Memory of size %li bytes, %li MB\n",ctx_size, (ctx_size/1024/1024));
printf("Allocating Memory of size %zi bytes, %zi MB\n",ctx_size, (ctx_size/1024/1024));
struct ggml_init_params params = {
/*.mem_size =*/ ctx_size,

6
examples/common.cpp
View file

@ -28,6 +28,10 @@
#include <wchar.h>
#endif
#if defined(_MSC_VER)
#pragma warning(disable: 4244 4267) // possible loss of data
#endif
int32_t get_num_physical_cores() {
#ifdef __linux__
// enumerate the set of thread siblings, num entries is num cores
@ -373,7 +377,7 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
} else {
throw std::exception();
}
} catch (const std::exception &e) {
} catch (const std::exception&) {
invalid_param = true;
break;
}

4
examples/embedding/embedding.cpp
View file

@ -4,6 +4,10 @@
#include <ctime>
#if defined(_MSC_VER)
#pragma warning(disable: 4244 4267) // possible loss of data
#endif
int main(int argc, char ** argv) {
gpt_params params;

7
examples/jeopardy/graph.py
View file

@ -1,5 +1,5 @@
import matplotlib.pyplot as plt
import sys, os
import os
import csv
labels = []
@ -8,6 +8,7 @@ numEntries = 1
rows = []
def bar_chart(numbers, labels, pos):
plt.bar(pos, numbers, color='blue')
plt.xticks(ticks=pos, labels=labels)
@ -16,6 +17,7 @@ def bar_chart(numbers, labels, pos):
plt.ylabel("Questions Correct")
plt.show()
def calculatecorrect():
directory = os.fsencode("./examples/jeopardy/results/")
csv_reader = csv.reader(open("./examples/jeopardy/qasheet.csv", 'rt'), delimiter=',')
@ -38,14 +40,13 @@ def calculatecorrect():
print(line)
else:
print("Correct answer: " + rows[i][2] + "\n")
i+=1
i += 1
print("Did the AI get the question right? (y/n)")
if input() == "y":
totalcorrect += 1
numbers.append(totalcorrect)
if __name__ == '__main__':
calculatecorrect()
pos = list(range(numEntries))

8
examples/main/main.cpp
View file

@ -23,11 +23,17 @@
#include <unistd.h>
#elif defined (_WIN32)
#define WIN32_LEAN_AND_MEAN
#ifndef NOMINMAX
#define NOMINMAX
#endif
#include <windows.h>
#include <signal.h>
#endif
#if defined(_MSC_VER)
#pragma warning(disable: 4244 4267) // possible loss of data
#endif
static console_state con_st;
static llama_context ** g_ctx;
@ -348,7 +354,7 @@ int main(int argc, char ** argv) {
if ((int)embd.size() > max_embd_size) {
auto skipped_tokens = embd.size() - max_embd_size;
console_set_color(con_st, CONSOLE_COLOR_ERROR);
printf("<<input too long: skipped %ld token%s>>", skipped_tokens, skipped_tokens != 1 ? "s" : "");
printf("<<input too long: skipped %" PRIu64 " token%s>>", skipped_tokens, skipped_tokens != 1 ? "s" : "");
console_set_color(con_st, CONSOLE_COLOR_DEFAULT);
fflush(stdout);
embd.resize(max_embd_size);

4
examples/perplexity/perplexity.cpp
View file

@ -5,6 +5,10 @@
#include <cmath>
#include <ctime>
#if defined(_MSC_VER)
#pragma warning(disable: 4244 4267) // possible loss of data
#endif
std::vector<float> softmax(const std::vector<float>& logits) {
std::vector<float> probs(logits.size());
float max_logit = logits[0];

4
examples/quantize-stats/quantize-stats.cpp
View file

@ -19,6 +19,10 @@
#include <thread>
#include <mutex>
#if defined(_MSC_VER)
#pragma warning(disable: 4244 4267) // possible loss of data
#endif
struct quantize_stats_params {
std::string model = "models/7B/ggml-model-f16.bin";
bool verbose = false;

2
examples/save-load-state/save-load-state.cpp
View file

@ -37,7 +37,7 @@ int main(int argc, char ** argv) {
// init
auto ctx = llama_init_from_file(params.model.c_str(), lparams);
auto tokens = std::vector<llama_token>(params.n_ctx);
auto n_prompt_tokens = llama_tokenize(ctx, params.prompt.c_str(), tokens.data(), tokens.size(), true);
auto n_prompt_tokens = llama_tokenize(ctx, params.prompt.c_str(), tokens.data(), int(tokens.size()), true);
if (n_prompt_tokens < 1) {
fprintf(stderr, "%s : failed to tokenize prompt\n", __func__);

4
examples/server/CMakeLists.txt
View file

@ -1,6 +1,10 @@
set(TARGET server)
option(LLAMA_SERVER_VERBOSE "Build verbose logging option for Server" ON)
include_directories(${CMAKE_CURRENT_SOURCE_DIR})
add_executable(${TARGET} server.cpp json.hpp httplib.h)
target_compile_definitions(${TARGET} PRIVATE
SERVER_VERBOSE=$<BOOL:${LLAMA_SERVER_VERBOSE}>
)
target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
target_compile_features(${TARGET} PRIVATE cxx_std_11)
if(TARGET BUILD_INFO)

318
examples/server/README.md
View file

@ -1,37 +1,74 @@
# llama.cpp/example/server
This example allow you to have a llama.cpp http server to interact from a web page or consume the API.
This example demonstrates a simple HTTP API server to interact with llama.cpp.
## Table of Contents
Command line options:
1. [Quick Start](#quick-start)
2. [Node JS Test](#node-js-test)
3. [API Endpoints](#api-endpoints)
4. [More examples](#more-examples)
5. [Common Options](#common-options)
6. [Performance Tuning and Memory Options](#performance-tuning-and-memory-options)
- `--threads N`, `-t N`: Set the number of threads to use during computation.
- `-m FNAME`, `--model FNAME`: Specify the path to the LLaMA model file (e.g., `models/7B/ggml-model.bin`).
- `-m ALIAS`, `--alias ALIAS`: Set an alias for the model. The alias will be returned in API responses.
- `-c N`, `--ctx-size N`: Set the size of the prompt context. The default is 512, but LLaMA models were built with a context of 2048, which will provide better results for longer input/inference.
- `-ngl N`, `--n-gpu-layers N`: When compiled with appropriate support (currently CLBlast or cuBLAS), this option allows offloading some layers to the GPU for computation. Generally results in increased performance.
- `-mg i, --main-gpu i`: When using multiple GPUs this option controls which GPU is used for small tensors for which the overhead of splitting the computation across all GPUs is not worthwhile. The GPU in question will use slightly more VRAM to store a scratch buffer for temporary results. By default GPU 0 is used. Requires cuBLAS.
- `-ts SPLIT, --tensor-split SPLIT`: When using multiple GPUs this option controls how large tensors should be split across all GPUs. `SPLIT` is a comma-separated list of non-negative values that assigns the proportion of data that each GPU should get in order. For example, "3,2" will assign 60% of the data to GPU 0 and 40% to GPU 1. By default the data is split in proportion to VRAM but this may not be optimal for performance. Requires cuBLAS.
- `-lv, --low-vram`: Do not allocate a VRAM scratch buffer for holding temporary results. Reduces VRAM usage at the cost of performance, particularly prompt processing speed. Requires cuBLAS.
- `-b N`, `--batch-size N`: Set the batch size for prompt processing. Default: `512`.
- `--memory-f32`: Use 32-bit floats instead of 16-bit floats for memory key+value. Not recommended.
- `--mlock`: Lock the model in memory, preventing it from being swapped out when memory-mapped.
- `--no-mmap`: Do not memory-map the model. By default, models are mapped into memory, which allows the system to load only the necessary parts of the model as needed.
- `--lora FNAME`: Apply a LoRA (Low-Rank Adaptation) adapter to the model (implies --no-mmap). This allows you to adapt the pretrained model to specific tasks or domains.
- `--lora-base FNAME`: Optional model to use as a base for the layers modified by the LoRA adapter. This flag is used in conjunction with the `--lora` flag, and specifies the base model for the adaptation.
- `-to N`, `--timeout N`: Server read/write timeout in seconds. Default `600`.
- `--host`: Set the hostname or ip address to listen. Default `127.0.0.1`.
- `--port`: Set the port to listen. Default: `8080`.
## Build
Build llama.cpp with server from repository root with either make or CMake.
- Using `make`:
```bash
LLAMA_BUILD_SERVER=1 make
```
- Using `CMake`:
```bash
mkdir build-server
cd build-server
cmake -DLLAMA_BUILD_SERVER=ON ..
cmake --build . --config Release
```
## Quick Start
To get started right away, run the following command, making sure to use the correct path for the model you have:
#### Unix-based systems (Linux, macOS, etc.):
Make sure to build with the server option on
```bash
LLAMA_BUILD_SERVER=1 make
```
### Unix-based systems (Linux, macOS, etc.):
```bash
./server -m models/7B/ggml-model.bin --ctx_size 2048
./server -m models/7B/ggml-model.bin -c 2048
```
#### Windows:
### Windows:
```powershell
server.exe -m models\7B\ggml-model.bin --ctx_size 2048
server.exe -m models\7B\ggml-model.bin -c 2048
```
That will start a server that by default listens on `127.0.0.1:8080`. You can consume the endpoints with Postman or NodeJS with axios library.
The above command will start a server that by default listens on `127.0.0.1:8080`.
You can consume the endpoints with Postman or NodeJS with axios library.
## Testing with CURL
Using [curl](https://curl.se/). On Windows `curl.exe` should be available in the base OS.
```sh
curl --request POST \
--url http://localhost:8080/completion \
--data '{"prompt": "Building a website can be done in 10 simple steps:","n_predict": 128}'
```
## Node JS Test
@ -54,7 +91,6 @@ const prompt = `Building a website can be done in 10 simple steps:`;
async function Test() {
let result = await axios.post("http://127.0.0.1:8080/completion", {
prompt,
batch_size: 128,
n_predict: 512,
});
@ -73,247 +109,75 @@ node .
## API Endpoints
You can interact with this API Endpoints. This implementations just support chat style interaction.
- **POST** `/completion`: Given a prompt, it returns the predicted completion.
- **POST** `hostname:port/completion`: Setting up the Llama Context to begin the completions tasks.
*Options:*
*Options:*
`temperature`: Adjust the randomness of the generated text (default: 0.8).
`batch_size`: Set the batch size for prompt processing (default: 512).
`top_k`: Limit the next token selection to the K most probable tokens (default: 40).
`temperature`: Adjust the randomness of the generated text (default: 0.8).
`top_p`: Limit the next token selection to a subset of tokens with a cumulative probability above a threshold P (default: 0.9).
`top_k`: Limit the next token selection to the K most probable tokens (default: 40).
`n_predict`: Set the number of tokens to predict when generating text. **Note:** May exceed the set limit slightly if the last token is a partial multibyte character. (default: 128, -1 = infinity).
`top_p`: Limit the next token selection to a subset of tokens with a cumulative probability above a threshold P (default: 0.9).
`n_keep`: Specify the number of tokens from the initial prompt to retain when the model resets its internal context.
By default, this value is set to 0 (meaning no tokens are kept). Use `-1` to retain all tokens from the initial prompt.
`n_predict`: Set the number of tokens to predict when generating text (default: 128, -1 = infinity).
`stream`: It allows receiving each predicted token in real-time instead of waiting for the completion to finish. To enable this, set to `true`.
`threads`: Set the number of threads to use during computation.
`prompt`: Provide a prompt. Internally, the prompt is compared, and it detects if a part has already been evaluated, and the remaining part will be evaluate.
`n_keep`: Specify the number of tokens from the initial prompt to retain when the model resets its internal context. By default, this value is set to 0 (meaning no tokens are kept). Use `-1` to retain all tokens from the initial prompt.
`stop`: Specify a JSON array of stopping strings.
These words will not be included in the completion, so make sure to add them to the prompt for the next iteration (default: []).
`as_loop`: It allows receiving each predicted token in real-time instead of waiting for the completion to finish. To enable this, set to `true`.
`tfs_z`: Enable tail free sampling with parameter z (default: 1.0, 1.0 = disabled).
`interactive`: It allows interacting with the completion, and the completion stops as soon as it encounters a `stop word`. To enable this, set to `true`.
`typical_p`: Enable locally typical sampling with parameter p (default: 1.0, 1.0 = disabled).
`prompt`: Provide a prompt. Internally, the prompt is compared, and it detects if a part has already been evaluated, and the remaining part will be evaluate.
`repeat_penalty`: Control the repetition of token sequences in the generated text (default: 1.1).
`stop`: Specify the words or characters that indicate a stop. These words will not be included in the completion, so make sure to add them to the prompt for the next iteration.
`repeat_last_n`: Last n tokens to consider for penalizing repetition (default: 64, 0 = disabled, -1 = ctx-size).
`exclude`: Specify the words or characters you do not want to appear in the completion. These words will not be included in the completion, so make sure to add them to the prompt for the next iteration.
`penalize_nl`: Penalize newline tokens when applying the repeat penalty (default: true).
- **POST** `hostname:port/embedding`: Generate embedding of a given text
`presence_penalty`: Repeat alpha presence penalty (default: 0.0, 0.0 = disabled).
*Options:*
`frequency_penalty`: Repeat alpha frequency penalty (default: 0.0, 0.0 = disabled);
`content`: Set the text to get generate the embedding.
`mirostat`: Enable Mirostat sampling, controlling perplexity during text generation (default: 0, 0 = disabled, 1 = Mirostat, 2 = Mirostat 2.0).
`threads`: Set the number of threads to use during computation.
`mirostat_tau`: Set the Mirostat target entropy, parameter tau (default: 5.0).
To use this endpoint, you need to start the server with the `--embedding` option added.
`mirostat_eta`: Set the Mirostat learning rate, parameter eta (default: 0.1).
- **POST** `hostname:port/tokenize`: Tokenize a given text
`seed`: Set the random number generator (RNG) seed (default: -1, < 0 = random seed).
*Options:*
`ignore_eos`: Ignore end of stream token and continue generating (default: false).
`content`: Set the text to tokenize.
`logit_bias`: Modify the likelihood of a token appearing in the generated text completion. For example, use `"logit_bias": [[15043,1.0]]` to increase the likelihood of the token 'Hello', or `"logit_bias": [[15043,-1.0]]` to decrease its likelihood. Setting the value to false, `"logit_bias": [[15043,false]]` ensures that the token `Hello` is never produced (default: []).
- **GET** `hostname:port/next-token`: Receive the next token predicted, execute this request in a loop. Make sure set `as_loop` as `true` in the completion request.
- **POST** `/tokenize`: Tokenize a given text.
*Options:*
*Options:*
`stop`: Set `hostname:port/next-token?stop=true` to stop the token generation.
`content`: Set the text to tokenize.
## More examples
### Interactive mode
This mode allows interacting in a chat-like manner. It is recommended for models designed as assistants such as `Vicuna`, `WizardLM`, `Koala`, among others. Make sure to add the correct stop word for the corresponding model.
Check the sample in [chat.mjs](chat.mjs).
Run with NodeJS version 16 or later:
The prompt should be generated by you, according to the model's guidelines. You should keep adding the model's completions to the context as well.
This example works well for `Vicuna - version 1`.
```javascript
const axios = require("axios");
let prompt = `A chat between a curious human and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the human's questions.
### Human: Hello, Assistant.
### Assistant: Hello. How may I help you today?
### Human: Please tell me the largest city in Europe.
### Assistant: Sure. The largest city in Europe is Moscow, the capital of Russia.`;
async function ChatCompletion(answer) {
// the user's next question to the prompt
prompt += `\n### Human: ${answer}\n`
result = await axios.post("http://127.0.0.1:8080/completion", {
prompt,
batch_size: 128,
temperature: 0.2,
top_k: 40,
top_p: 0.9,
n_keep: -1,
n_predict: 2048,
stop: ["\n### Human:"], // when detect this, stop completion
exclude: ["### Assistant:"], // no show in the completion
threads: 8,
as_loop: true, // use this to request the completion token by token
interactive: true, // enable the detection of a stop word
});
// create a loop to receive every token predicted
// note: this operation is blocking, avoid use this in a ui thread
let message = "";
while (true) {
// you can stop the inference adding '?stop=true' like this http://127.0.0.1:8080/next-token?stop=true
result = await axios.get("http://127.0.0.1:8080/next-token");
process.stdout.write(result.data.content);
message += result.data.content;
// to avoid an infinite loop
if (result.data.stop) {
console.log("Completed");
// make sure to add the completion to the prompt.
prompt += `### Assistant: ${message}`;
break;
}
}
}
// This function should be called every time a question to the model is needed.
async function Test() {
// the server can't inference in paralell
await ChatCompletion("Write a long story about a time magician in a fantasy world");
await ChatCompletion("Summary the story");
}
Test();
```sh
node chat.mjs
```
### Alpaca example
Another sample in [chat.sh](chat.sh).
Requires [bash](https://www.gnu.org/software/bash/), [curl](https://curl.se) and [jq](https://jqlang.github.io/jq/).
Run with bash:
**Temporaly note:** no tested, if you have the model, please test it and report me some issue
```javascript
const axios = require("axios");
let prompt = `Below is an instruction that describes a task. Write a response that appropriately completes the request.
`;
async function DoInstruction(instruction) {
prompt += `\n\n### Instruction:\n\n${instruction}\n\n### Response:\n\n`;
result = await axios.post("http://127.0.0.1:8080/completion", {
prompt,
batch_size: 128,
temperature: 0.2,
top_k: 40,
top_p: 0.9,
n_keep: -1,
n_predict: 2048,
stop: ["### Instruction:\n\n"], // when detect this, stop completion
exclude: [], // no show in the completion
threads: 8,
as_loop: true, // use this to request the completion token by token
interactive: true, // enable the detection of a stop word
});
// create a loop to receive every token predicted
// note: this operation is blocking, avoid use this in a ui thread
let message = "";
while (true) {
result = await axios.get("http://127.0.0.1:8080/next-token");
process.stdout.write(result.data.content);
message += result.data.content;
// to avoid an infinite loop
if (result.data.stop) {
console.log("Completed");
// make sure to add the completion and the user's next question to the prompt.
prompt += message;
break;
}
}
}
// This function should be called every time a instruction to the model is needed.
DoInstruction("Destroy the world"); // as joke
```sh
bash chat.sh
```
### Embeddings
First, run the server with `--embedding` option:
```bash
server -m models/7B/ggml-model.bin --ctx_size 2048 --embedding
```
Run this code in NodeJS:
```javascript
const axios = require('axios');
async function Test() {
let result = await axios.post("http://127.0.0.1:8080/embedding", {
content: `Hello`,
threads: 5
});
// print the embedding array
console.log(result.data.embedding);
}
Test();
```
### Tokenize
Run this code in NodeJS:
```javascript
const axios = require('axios');
async function Test() {
let result = await axios.post("http://127.0.0.1:8080/tokenize", {
content: `Hello`
});
// print the embedding array
console.log(result.data.tokens);
}
Test();
```
## Common Options
- `-m FNAME, --model FNAME`: Specify the path to the LLaMA model file (e.g., `models/7B/ggml-model.bin`).
- `-c N, --ctx-size N`: Set the size of the prompt context. The default is 512, but LLaMA models were built with a context of 2048, which will provide better results for longer input/inference.
- `-ngl N, --n-gpu-layers N`: When compiled with appropriate support (currently CLBlast or cuBLAS), this option allows offloading some layers to the GPU for computation. Generally results in increased performance.
- `-mg i, --main-gpu i`: When using multiple GPUs this option controls which GPU is used for small tensors for which the overhead of splitting the computation across all GPUs is not worthwhile. The GPU in question will use slightly more VRAM to store a scratch buffer for temporary results. By default GPU 0 is used. Requires cuBLAS.
- `-ts SPLIT, --tensor-split SPLIT`: When using multiple GPUs this option controls how large tensors should be split across all GPUs. `SPLIT` is a comma-separated list of non-negative values that assigns the proportion of data that each GPU should get in order. For example, "3,2" will assign 60% of the data to GPU 0 and 40% to GPU 1. By default the data is split in proportion to VRAM but this may not be optimal for performance. Requires cuBLAS.
- `-lv, --low-vram`: Do not allocate a VRAM scratch buffer for holding temporary results. Reduces VRAM usage at the cost of performance, particularly prompt processing speed. Requires cuBLAS.
- `--embedding`: Enable the embedding mode. **Completion function doesn't work in this mode**.
- `--host`: Set the hostname or ip address to listen. Default `127.0.0.1`;
- `--port`: Set the port to listen. Default: `8080`.
### RNG Seed
- `-s SEED, --seed SEED`: Set the random number generator (RNG) seed (default: -1, < 0 = random seed).
The RNG seed is used to initialize the random number generator that influences the text generation process. By setting a specific seed value, you can obtain consistent and reproducible results across multiple runs with the same input and settings. This can be helpful for testing, debugging, or comparing the effects of different options on the generated text to see when they diverge. If the seed is set to a value less than 0, a random seed will be used, which will result in different outputs on each run.
## Performance Tuning and Memory Options
### No Memory Mapping
- `--no-mmap`: Do not memory-map the model. By default, models are mapped into memory, which allows the system to load only the necessary parts of the model as needed. However, if the model is larger than your total amount of RAM or if your system is low on available memory, using mmap might increase the risk of pageouts, negatively impacting performance.
### Memory Float 32
- `--memory-f32`: Use 32-bit floats instead of 16-bit floats for memory key+value. This doubles the context memory requirement but does not appear to increase generation quality in a measurable way. Not recommended.
## Limitations:
- The actual implementation of llama.cpp need a `llama-state` for handle multiple contexts and clients, but this could require more powerful hardware.

89
examples/server/chat.mjs Normal file
View file

@ -0,0 +1,89 @@
import * as readline from 'node:readline'
import { stdin, stdout } from 'node:process'
const API_URL = 'http://127.0.0.1:8080'
const chat = [
{
human: "Hello, Assistant.",
assistant: "Hello. How may I help you today?"
},
{
human: "Please tell me the largest city in Europe.",
assistant: "Sure. The largest city in Europe is Moscow, the capital of Russia."
},
]
const instruction = `A chat between a curious human and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the human's questions.`
function format_prompt(question) {
return `${instruction}\n${
chat.map(m =>`### Human: ${m.human}\n### Assistant: ${m.assistant}`).join("\n")
}\n### Human: ${question}\n### Assistant:`
}
async function tokenize(content) {
const result = await fetch(`${API_URL}/tokenize`, {
method: 'POST',
body: JSON.stringify({ content })
})
if (!result.ok) {
return []
}
return await result.json().tokens
}
const n_keep = await tokenize(instruction).length
async function chat_completion(question) {
const result = await fetch(`${API_URL}/completion`, {
method: 'POST',
body: JSON.stringify({
prompt: format_prompt(question),
temperature: 0.2,
top_k: 40,
top_p: 0.9,
n_keep: n_keep,
n_predict: 256,
stop: ["\n### Human:"], // stop completion after generating this
stream: true,
})
})
if (!result.ok) {
return
}
let answer = ''
for await (var chunk of result.body) {
const t = Buffer.from(chunk).toString('utf8')
if (t.startsWith('data: ')) {
const message = JSON.parse(t.substring(6))
answer += message.content
process.stdout.write(message.content)
if (message.stop) {
if (message.truncated) {
chat.shift()
}
break
}
}
}
process.stdout.write('\n')
chat.push({ human: question, assistant: answer.trimStart() })
}
const rl = readline.createInterface({ input: stdin, output: stdout });
const readlineQuestion = (rl, query, options) => new Promise((resolve, reject) => {
rl.question(query, options, resolve)
});
while(true) {
const question = await readlineQuestion(rl, '> ')
await chat_completion(question)
}

77
examples/server/chat.sh Normal file
View file

@ -0,0 +1,77 @@
#!/bin/bash
API_URL="${API_URL:-http://127.0.0.1:8080}"
CHAT=(
"Hello, Assistant."
"Hello. How may I help you today?"
"Please tell me the largest city in Europe."
"Sure. The largest city in Europe is Moscow, the capital of Russia."
)
INSTRUCTION="A chat between a curious human and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the human's questions."
trim() {
shopt -s extglob
set -- "${1##+([[:space:]])}"
printf "%s" "${1%%+([[:space:]])}"
}
trim_trailing() {
shopt -s extglob
printf "%s" "${1%%+([[:space:]])}"
}
format_prompt() {
echo -n "${INSTRUCTION}"
printf "\n### Human: %s\n### Assistant: %s" "${CHAT[@]}" "$1"
}
tokenize() {
curl \
--silent \
--request POST \
--url "${API_URL}/tokenize" \
--data-raw "$(jq -ns --arg content "$1" '{content:$content}')" \
| jq '.tokens[]'
}
N_KEEP=$(tokenize "${INSTRUCTION}" | wc -l)
chat_completion() {
PROMPT="$(trim_trailing "$(format_prompt "$1")")"
DATA="$(echo -n "$PROMPT" | jq -Rs --argjson n_keep $N_KEEP '{
prompt: .,
temperature: 0.2,
top_k: 40,
top_p: 0.9,
n_keep: $n_keep,
n_predict: 256,
stop: ["\n### Human:"],
stream: true
}')"
ANSWER=''
while IFS= read -r LINE; do
if [[ $LINE = data:* ]]; then
CONTENT="$(echo "${LINE:5}" | jq -r '.content')"
printf "%s" "${CONTENT}"
ANSWER+="${CONTENT}"
fi
done < <(curl \
--silent \
--no-buffer \
--request POST \
--url "${API_URL}/completion" \
--data-raw "${DATA}")
printf "\n"
CHAT+=("$1" "$(trim "$ANSWER")")
}
while true; do
read -r -e -p "> " QUESTION
chat_completion "${QUESTION}"
done

1681
examples/server/server.cpp
View file

File diff suppressed because it is too large Load diff

7
examples/simple/CMakeLists.txt Normal file
View file

@ -0,0 +1,7 @@
set(TARGET simple)
add_executable(${TARGET} simple.cpp)
target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
target_compile_features(${TARGET} PRIVATE cxx_std_11)
if(TARGET BUILD_INFO)
add_dependencies(${TARGET} BUILD_INFO)
endif()

177
examples/simple/simple.cpp Normal file
View file

@ -0,0 +1,177 @@
#ifndef _GNU_SOURCE
#define _GNU_SOURCE
#endif
#include "common.h"
#include "llama.h"
#include "build-info.h"
#include <cassert>
#include <cinttypes>
#include <cmath>
#include <cstdio>
#include <cstring>
#include <ctime>
#include <fstream>
#include <iostream>
#include <string>
#include <vector>
#if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__))
#include <signal.h>
#include <unistd.h>
#elif defined (_WIN32)
#define WIN32_LEAN_AND_MEAN
#define NOMINMAX
#include <windows.h>
#include <signal.h>
#endif
int main(int argc, char ** argv)
{
gpt_params params;
//---------------------------------
// Print help :
//---------------------------------
if ( argc == 1 || argv[1][0] == '-' )
{
printf( "usage: %s MODEL_PATH [PROMPT]\n" , argv[0] );
return 1 ;
}
//---------------------------------
// Load parameters :
//---------------------------------
if ( argc >= 2 )
{
params.model = argv[1];
}
if ( argc >= 3 )
{
params.prompt = argv[2];
}
if ( params.prompt.empty() )
{
params.prompt = "Hello my name is";
}
//---------------------------------
// Init LLM :
//---------------------------------
llama_init_backend();
llama_context * ctx ;
ctx = llama_init_from_gpt_params( params );
if ( ctx == NULL )
{
fprintf( stderr , "%s: error: unable to load model\n" , __func__ );
return 1;
}
//---------------------------------
// Tokenize the prompt :
//---------------------------------
std::vector<llama_token> tokens_list;
tokens_list = ::llama_tokenize( ctx , params.prompt , true );
const int max_context_size = llama_n_ctx( ctx );
const int max_tokens_list_size = max_context_size - 4 ;
if ( (int)tokens_list.size() > max_tokens_list_size )
{
fprintf( stderr , "%s: error: prompt too long (%d tokens, max %d)\n" ,
__func__ , (int)tokens_list.size() , max_tokens_list_size );
return 1;
}
fprintf( stderr, "\n\n" );
// Print the tokens from the prompt :
for( auto id : tokens_list )
{
printf( "%s" , llama_token_to_str( ctx , id ) );
}
fflush(stdout);
//---------------------------------
// Main prediction loop :
//---------------------------------
// The LLM keeps a contextual cache memory of previous token evaluation.
// Usually, once this cache is full, it is required to recompute a compressed context based on previous
// tokens (see "infinite text generation via context swapping" in the main example), but in this minimalist
// example, we will just stop the loop once this cache is full or once an end of stream is detected.
while ( llama_get_kv_cache_token_count( ctx ) < max_context_size )
{
//---------------------------------
// Evaluate the tokens :
//---------------------------------
if ( llama_eval( ctx , tokens_list.data() , tokens_list.size() , llama_get_kv_cache_token_count( ctx ) , params.n_threads ) )
{
fprintf( stderr, "%s : failed to eval\n" , __func__ );
return 1;
}
tokens_list.clear();
//---------------------------------
// Select the best prediction :
//---------------------------------
llama_token new_token_id = 0;
auto logits = llama_get_logits( ctx );
auto n_vocab = llama_n_vocab( ctx ); // the size of the LLM vocabulary (in tokens)
std::vector<llama_token_data> candidates;
candidates.reserve( n_vocab );
for( llama_token token_id = 0 ; token_id < n_vocab ; token_id++ )
{
candidates.emplace_back( llama_token_data{ token_id , logits[ token_id ] , 0.0f } );
}
llama_token_data_array candidates_p = { candidates.data(), candidates.size(), false };
// Select it using the "Greedy sampling" method :
new_token_id = llama_sample_token_greedy( ctx , &candidates_p );
// is it an end of stream ?
if ( new_token_id == llama_token_eos() )
{
fprintf(stderr, " [end of text]\n");
break;
}
// Print the new token :
printf( "%s" , llama_token_to_str( ctx , new_token_id ) );
fflush( stdout );
// Push this new token for next evaluation :
tokens_list.push_back( new_token_id );
} // wend of main loop
llama_free( ctx );
return 0;
}
// EOF

2
examples/train-text-from-scratch/README.md
View file

@ -4,7 +4,7 @@ Basic usage instructions:
```bash
# get training data
wget https://github.com/brunoklein99/deep-learning-notes/blob/master/shakespeare.txt
wget https://raw.githubusercontent.com/brunoklein99/deep-learning-notes/master/shakespeare.txt
# train
./bin/train-text-from-scratch \

18
examples/train-text-from-scratch/train-text-from-scratch.cpp
View file

@ -12,6 +12,9 @@
#include <algorithm>
#include <string>
#if defined(_MSC_VER)
#pragma warning(disable: 4244 4267) // possible loss of data
#endif
struct random_normal_distribution {
std::mt19937 gen;
@ -20,7 +23,6 @@ struct random_normal_distribution {
float max;
};
struct random_uniform_distribution {
std::mt19937 gen;
std::uniform_real_distribution<float> rd;
@ -2366,7 +2368,7 @@ void write_tensor(struct llama_file * file, struct ggml_tensor * tensor) {
file->write_u32(0);
file->write_u32(0);
file->write_u32(GGML_TYPE_F32);
file->seek(-file->tell() & 31, SEEK_CUR);
file->seek(0-file->tell() & 31, SEEK_CUR);
return;
}
const char * name = ggml_get_name(tensor);
@ -2381,7 +2383,7 @@ void write_tensor(struct llama_file * file, struct ggml_tensor * tensor) {
file->write_u32(tensor->type);
file->write_raw(ne, sizeof(ne[0]) * nd);
file->write_raw(name, name_len);
file->seek(-file->tell() & 31, SEEK_CUR);
file->seek(0-file->tell() & 31, SEEK_CUR);
file->write_raw(tensor->data, ggml_nbytes(tensor));
}
@ -2402,7 +2404,7 @@ void read_tensor(struct llama_file * file, struct ggml_tensor * tensor) {
std::string name = file->read_string(name_len);
GGML_ASSERT(strncmp(ggml_get_name(tensor), name.c_str(), sizeof(tensor->name)-1) == 0);
file->seek(-file->tell() & 31, SEEK_CUR);
file->seek(0-file->tell() & 31, SEEK_CUR);
file->read_raw(tensor->data, ggml_nbytes(tensor));
}
@ -2756,8 +2758,8 @@ struct train_params get_default_train_params() {
params.lbfgs_n_iter = 16;
params.adam_n_iter = 16;
params.adam_alpha = 1e-3;
params.adam_decay = 1e-3;
params.adam_alpha = 1e-3f;
params.adam_decay = 1e-3f;
params.mem_model_gb = 2;
params.mem_compute_gb = 24;
@ -3331,8 +3333,8 @@ int main(int argc, char ** argv) {
int n_gen = params.n_predict;
int sample_ctx = n_tokens - n_tokens/8;
sampler.params.temp = 0.2;
sampler.params.repeat_penalty = 1.1;
sampler.params.temp = 0.2f;
sampler.params.repeat_penalty = 1.1f;
sampler.params.mirostat = 2;
init_sampler(&sampler, lctx);

599
ggml-cuda.cu
View file

@ -167,6 +167,12 @@ static_assert(sizeof(block_q6_K) == sizeof(ggml_fp16_t) + 13*QK_K/16, "wrong q6_
#define GGML_CUDA_DMMV_Y 1
#endif
#ifndef K_QUANTS_PER_ITERATION
#define K_QUANTS_PER_ITERATION 2
#else
static_assert(K_QUANTS_PER_ITERATION == 1 || K_QUANTS_PER_ITERATION == 2, "K_QUANTS_PER_ITERATION must be 1 or 2");
#endif
static __global__ void add_f32(const float * x, const float * y, float * dst, const int k) {
const int i = blockDim.x*blockIdx.x + threadIdx.x;
@ -326,37 +332,6 @@ static __global__ void dequantize_block_q2_K(const void * vx, float * yy) {
}
static __device__ void vec_dot_q2_K(const void * vx, const int ib, const int iqs, const float * yy, float & result) {
const block_q2_K * x = (const block_q2_K *) vx;
// if n is 0, we want to do the lower 128, else the upper 128,
// covering y[l+0], y[l+32], y[l+64], y[l+96] and
// y[l+16], y[l+48], y[l+80], y[l+112]
int n = iqs/128; // 0 or 1
int r = iqs - 128*n; // 0...120 in steps of 8
int l = r/8; // 0...15 in steps of 1
const float * y = yy + 128*n + l;
const uint8_t * q = x[ib].qs + 32*n + l;
const uint8_t * s = x[ib].scales + 8*n;
const float dall = x[ib].d;
const float dmin = x[ib].dmin;
float sum = y[ 0] * (dall * ((s[0] & 0xF) * ((q[ 0] >> 0) & 3)) - dmin * (s[0] >> 4))
+ y[ 32] * (dall * ((s[2] & 0xF) * ((q[ 0] >> 2) & 3)) - dmin * (s[2] >> 4))
+ y[ 64] * (dall * ((s[4] & 0xF) * ((q[ 0] >> 4) & 3)) - dmin * (s[4] >> 4))
+ y[ 96] * (dall * ((s[6] & 0xF) * ((q[ 0] >> 6) & 3)) - dmin * (s[6] >> 4))
+ y[ 16] * (dall * ((s[1] & 0xF) * ((q[16] >> 0) & 3)) - dmin * (s[1] >> 4))
+ y[ 48] * (dall * ((s[3] & 0xF) * ((q[16] >> 2) & 3)) - dmin * (s[3] >> 4))
+ y[ 80] * (dall * ((s[5] & 0xF) * ((q[16] >> 4) & 3)) - dmin * (s[5] >> 4))
+ y[112] * (dall * ((s[7] & 0xF) * ((q[16] >> 6) & 3)) - dmin * (s[7] >> 4));
result = sum;
}
static __global__ void dequantize_block_q3_K(const void * vx, float * yy) {
int r = threadIdx.x/4;
@ -388,51 +363,6 @@ static __global__ void dequantize_block_q3_K(const void * vx, float * yy) {
}
static __device__ void vec_dot_q3_K(const void * vx, const int ib, const int iqs, const float * yy, float & result) {
const block_q3_K * x = (const block_q3_K *) vx;
const uint32_t kmask1 = 0x03030303;
const uint32_t kmask2 = 0x0f0f0f0f;
uint32_t aux[3];
uint32_t utmp[4];
// if n is 0, we want to do the lower 128, else the upper 128,
// covering y[l+0], y[l+32], y[l+64], y[l+96] and
// y[l+16], y[l+48], y[l+80], y[l+112]
int n = iqs/128; // 0 or 1
int r = iqs - 128*n; // 0...120 in steps of 8
int l = r/8; // 0...15 in steps of 1
const float * y = yy + 128*n + l;
const uint8_t * q = x[ib].qs + 32*n + l;
const uint8_t * hm = x[ib].hmask + l;
const int8_t * s = (const int8_t *)utmp + 8*n;
memcpy(aux, x[ib].scales, 12);
utmp[3] = ((aux[1] >> 4) & kmask2) | (((aux[2] >> 6) & kmask1) << 4);
utmp[2] = ((aux[0] >> 4) & kmask2) | (((aux[2] >> 4) & kmask1) << 4);
utmp[1] = (aux[1] & kmask2) | (((aux[2] >> 2) & kmask1) << 4);
utmp[0] = (aux[0] & kmask2) | (((aux[2] >> 0) & kmask1) << 4);
const float dall = x[ib].d;
const uint8_t m = 1 << (4*n);
float sum = y[ 0] * (s[0] - 32) * (((q[ 0] >> 0) & 3) - (hm[ 0] & (m << 0) ? 0 : 4))
+ y[ 32] * (s[2] - 32) * (((q[ 0] >> 2) & 3) - (hm[ 0] & (m << 1) ? 0 : 4))
+ y[ 64] * (s[4] - 32) * (((q[ 0] >> 4) & 3) - (hm[ 0] & (m << 2) ? 0 : 4))
+ y[ 96] * (s[6] - 32) * (((q[ 0] >> 6) & 3) - (hm[ 0] & (m << 3) ? 0 : 4))
+ y[ 16] * (s[1] - 32) * (((q[16] >> 0) & 3) - (hm[16] & (m << 0) ? 0 : 4))
+ y[ 48] * (s[3] - 32) * (((q[16] >> 2) & 3) - (hm[16] & (m << 1) ? 0 : 4))
+ y[ 80] * (s[5] - 32) * (((q[16] >> 4) & 3) - (hm[16] & (m << 2) ? 0 : 4))
+ y[112] * (s[7] - 32) * (((q[16] >> 6) & 3) - (hm[16] & (m << 3) ? 0 : 4));
result = sum * dall;
}
static inline __device__ void get_scale_min_k4(int j, const uint8_t * q, uint8_t & d, uint8_t & m) {
if (j < 4) {
d = q[j] & 63; m = q[j + 4] & 63;
@ -479,38 +409,6 @@ static __global__ void dequantize_block_q4_K(const void * vx, float * yy) {
}
}
static __device__ void vec_dot_q4_K(const void * vx, const int ib, const int iqs, const float * yy, float & result) {
const block_q4_K * x = (const block_q4_K *) vx;
// iqs is in 0...248 in steps of 8 =>
const int j = iqs / 64; // j is in 0...3
const int ir = (iqs - 64*j)/2; // ir is in 0...28 in steps of 4
const int is = 2*j; // is is in 0...6 in steps of 2
const float * y = yy + 64*j + ir;
const uint8_t * q = x[ib].qs + 32*j + ir;
const float dall = x[ib].d;
const float dmin = x[ib].dmin;
uint8_t sc, m;
get_scale_min_k4(is + 0, x[ib].scales, sc, m);
const float d1 = dall * sc;
const float m1 = dmin * m;
get_scale_min_k4(is + 1, x[ib].scales, sc, m);
const float d2 = dall * sc;
const float m2 = dmin * m;
float sum = 0;
for (int k = 0; k < 4; ++k) {
sum += y[k + 0] * (d1 * (q[k] & 0xF) - m1);
sum += y[k + 32] * (d2 * (q[k] >> 4) - m2);
}
result = sum;
}
static __global__ void dequantize_block_q5_K(const void * vx, float * yy) {
const block_q5_K * x = (const block_q5_K *) vx;
@ -544,43 +442,6 @@ static __global__ void dequantize_block_q5_K(const void * vx, float * yy) {
y[33] = d2 * ((ql[ 1] >> 4) + (qh[ 1] & hm ? 16 : 0)) - m2;
}
static __device__ void vec_dot_q5_K(const void * vx, const int ib, const int iqs, const float * yy, float & result) {
const block_q5_K * x = (const block_q5_K *) vx;
// iqs is in 0...248 in steps of 8 =>
const int j = iqs / 64; // j is in 0...3
const int ir = (iqs - 64*j)/2; // ir is in 0...28 in steps of 4
const int is = 2*j; // is is in 0...6 in steps of 2
const float * y = yy + 64*j + ir;
const uint8_t * ql = x[ib].qs + 32*j + ir;
const uint8_t * qh = x[ib].qh + ir;
const float dall = x[ib].d;
const float dmin = x[ib].dmin;
uint8_t sc, m;
get_scale_min_k4(is + 0, x[ib].scales, sc, m);
const float d1 = dall * sc;
const float m1 = dmin * m;
get_scale_min_k4(is + 1, x[ib].scales, sc, m);
const float d2 = dall * sc;
const float m2 = dmin * m;
uint8_t hm = 1 << is;
float sum = 0;
for (int k = 0; k < 4; ++k) {
sum += y[k + 0] * (d1 * ((ql[k] & 0xF) + (qh[k] & hm ? 16 : 0)) - m1);
}
hm <<= 1;
for (int k = 0; k < 4; ++k) {
sum += y[k + 32] * (d2 * ((ql[k] >> 4) + (qh[k] & hm ? 16 : 0)) - m2);
}
result = sum;
}
static __global__ void dequantize_block_q6_K(const void * vx, float * yy) {
const block_q6_K * x = (const block_q6_K *) vx;
@ -606,31 +467,376 @@ static __global__ void dequantize_block_q6_K(const void * vx, float * yy) {
y[96] = d * sc[6] * ((int8_t)((ql[32] >> 4) | (((qh >> 6) & 3) << 4)) - 32);
}
static __device__ void vec_dot_q6_K(const void * vx, const int ib, const int iqs, const float * yy, float & result) {
static __global__ void dequantize_mul_mat_vec_q2_k(const void * vx, const float * yy, float * dst, const int ncols, int nrows) {
const block_q6_K * x = (const block_q6_K *) vx;
static_assert(16%K_QUANTS_PER_ITERATION == 0, "16 must be divisible by K_QUANTS_PER_ITERATION");
const int ip = iqs / 128; // 0 or 1
const int il = (iqs - 128*ip)/8; // 0...15
const int is = 8*ip;
const int row = blockIdx.y*blockDim.y + threadIdx.y;
if (row > nrows) return;
const float * y = yy + 128*ip + il;
const int num_blocks_per_row = ncols / QK_K;
const int ib0 = row*num_blocks_per_row;
const float d = x[ib].d;
const block_q2_K * x = (const block_q2_K *)vx + ib0;
const uint8_t * ql = x[ib].ql + 64*ip + il;
const uint8_t * qh = x[ib].qh + 32*ip + il;
const int8_t * sc = x[ib].scales + is;
const int tid = threadIdx.x/K_QUANTS_PER_ITERATION; // 0...31
const int ix = threadIdx.x%K_QUANTS_PER_ITERATION; // 0
result = y[ 0] * d * sc[0] * ((int8_t)((ql[ 0] & 0xF) | (((qh[ 0] >> 0) & 3) << 4)) - 32)
+ y[ 32] * d * sc[2] * ((int8_t)((ql[32] & 0xF) | (((qh[ 0] >> 2) & 3) << 4)) - 32)
+ y[ 64] * d * sc[4] * ((int8_t)((ql[ 0] >> 4) | (((qh[ 0] >> 4) & 3) << 4)) - 32)
+ y[ 96] * d * sc[6] * ((int8_t)((ql[32] >> 4) | (((qh[ 0] >> 6) & 3) << 4)) - 32)
+ y[ 16] * d * sc[1] * ((int8_t)((ql[16] & 0xF) | (((qh[16] >> 0) & 3) << 4)) - 32)
+ y[ 48] * d * sc[3] * ((int8_t)((ql[48] & 0xF) | (((qh[16] >> 2) & 3) << 4)) - 32)
+ y[ 80] * d * sc[5] * ((int8_t)((ql[16] >> 4) | (((qh[16] >> 4) & 3) << 4)) - 32)
+ y[112] * d * sc[7] * ((int8_t)((ql[48] >> 4) | (((qh[16] >> 6) & 3) << 4)) - 32);
const int step = 16/K_QUANTS_PER_ITERATION;
const int im = tid/step; // 0 or 1. 0 computes 0..., 1 computes 128...
const int in = tid - step*im; // 0...7
const int l0 = K_QUANTS_PER_ITERATION*in; // 0...14 in steps of 4
const int q_offset = 32*im + l0;
const int s_offset = 8*im;
const int y_offset = 128*im + l0;
float tmp = 0; // partial sum for thread in warp
uint32_t aux[4];
const uint8_t * d = (const uint8_t *)aux;
const uint8_t * m = (const uint8_t *)(aux + 2);
for (int i = ix; i < num_blocks_per_row; i += K_QUANTS_PER_ITERATION) {
const float * y = yy + i * QK_K + y_offset;
const uint8_t * q = x[i].qs + q_offset;
const float dall = x[i].d;
const float dmin = x[i].dmin;
const uint32_t * a = (const uint32_t *)(x[i].scales + s_offset);
aux[0] = a[0] & 0x0f0f0f0f;
aux[1] = a[1] & 0x0f0f0f0f;
aux[2] = (a[0] >> 4) & 0x0f0f0f0f;
aux[3] = (a[1] >> 4) & 0x0f0f0f0f;
float sum1 = 0, sum2 = 0;
for (int l = 0; l < K_QUANTS_PER_ITERATION; ++l) {
sum1 += y[l+ 0] * d[0] * ((q[l+ 0] >> 0) & 3)
+ y[l+32] * d[2] * ((q[l+ 0] >> 2) & 3)
+ y[l+64] * d[4] * ((q[l+ 0] >> 4) & 3)
+ y[l+96] * d[6] * ((q[l+ 0] >> 6) & 3)
+ y[l+16] * d[1] * ((q[l+16] >> 0) & 3)
+ y[l+48] * d[3] * ((q[l+16] >> 2) & 3)
+ y[l+80] * d[5] * ((q[l+16] >> 4) & 3)
+y[l+112] * d[7] * ((q[l+16] >> 6) & 3);
sum2 += y[l+ 0] * m[0] + y[l+32] * m[2] + y[l+64] * m[4] + y[ l+96] * m[6]
+ y[l+16] * m[1] + y[l+48] * m[3] + y[l+80] * m[5] + y[l+112] * m[7];
}
tmp += dall * sum1 - dmin * sum2;
}
// sum up partial sums and write back result
__syncthreads();
#pragma unroll
for (int mask = 16; mask > 0; mask >>= 1) {
tmp += __shfl_xor_sync(0xffffffff, tmp, mask, 32);
}
if (tid == 0) {
dst[row] = tmp;
}
}
static __global__ void dequantize_mul_mat_vec_q3_k(const void * vx, const float * yy, float * dst, const int ncols) {
const uint16_t kmask1 = 0x0303;
const uint16_t kmask2 = 0x0f0f;
const int row = blockIdx.x;
const int num_blocks_per_row = ncols / QK_K;
const int ib0 = row*num_blocks_per_row;
const block_q3_K * x = (const block_q3_K *)vx + ib0;
const int tid = threadIdx.x/2; // 0...15
const int ix = threadIdx.x%2; // 0, 1
const int n = 2; // iterations in the inner loop
const int im = tid/8; // 0 or 1. 0 computes 0..., 1 computes 128...
const int in = tid - 8*im; // 0...7
const uint8_t m = 1 << (4*im);
const int l0 = n*in; // 0...28 in steps of 4
const int q_offset = 32*im + l0;
const int y_offset = 128*im + l0;
uint16_t utmp[4];
const int8_t * s = (const int8_t *)utmp;
const uint16_t s_shift = 4*im;
float tmp = 0; // partial sum for thread in warp
for (int i = ix; i < num_blocks_per_row; i += 2) {
const float * y = yy + i * QK_K + y_offset;
const uint8_t * q = x[i].qs + q_offset;
const uint8_t * h = x[i].hmask + l0;
const uint16_t * a = (const uint16_t *)x[i].scales;
utmp[0] = ((a[0] >> s_shift) & kmask2) | (((a[4] >> (s_shift + 0)) & kmask1) << 4);
utmp[1] = ((a[1] >> s_shift) & kmask2) | (((a[5] >> (s_shift + 0)) & kmask1) << 4);
utmp[2] = ((a[2] >> s_shift) & kmask2) | (((a[4] >> (s_shift + 2)) & kmask1) << 4);
utmp[3] = ((a[3] >> s_shift) & kmask2) | (((a[5] >> (s_shift + 2)) & kmask1) << 4);
const float d = x[i].d;
float sum = 0;
for (int l = 0; l < n; ++l) {
sum += y[l+ 0] * (s[0] - 32) * (((q[l] >> 0) & 3) - (h[l] & (m << 0) ? 0 : 4))
+ y[l+32] * (s[2] - 32) * (((q[l] >> 2) & 3) - (h[l] & (m << 1) ? 0 : 4))
+ y[l+64] * (s[4] - 32) * (((q[l] >> 4) & 3) - (h[l] & (m << 2) ? 0 : 4))
+ y[l+96] * (s[6] - 32) * (((q[l] >> 6) & 3) - (h[l] & (m << 3) ? 0 : 4));
sum += y[l+16] * (s[1] - 32) * (((q[l+16] >> 0) & 3) - (h[l+16] & (m << 0) ? 0 : 4))
+ y[l+48] * (s[3] - 32) * (((q[l+16] >> 2) & 3) - (h[l+16] & (m << 1) ? 0 : 4))
+ y[l+80] * (s[5] - 32) * (((q[l+16] >> 4) & 3) - (h[l+16] & (m << 2) ? 0 : 4))
+ y[l+112] * (s[7] - 32) * (((q[l+16] >> 6) & 3) - (h[l+16] & (m << 3) ? 0 : 4));
}
tmp += d * sum;
}
// sum up partial sums and write back result
__syncthreads();
#pragma unroll
for (int mask = 16; mask > 0; mask >>= 1) {
tmp += __shfl_xor_sync(0xffffffff, tmp, mask, 32);
}
if (tid == 0) {
dst[row] = tmp;
}
}
static __global__ void dequantize_mul_mat_vec_q4_k(const void * vx, const float * yy, float * dst, const int ncols) {
const uint16_t kmask1 = 0x3f3f;
const uint16_t kmask2 = 0x0f0f;
const uint16_t kmask3 = 0xc0c0;
const int row = blockIdx.x;
const int num_blocks_per_row = ncols / QK_K;
const int ib0 = row*num_blocks_per_row;
const int tid = threadIdx.x/2; // 0...15
const int ix = threadIdx.x%2;
const int il = tid/4; // 0...3
const int ir = tid - 4*il;// 0...3
const int n = 4;
const int im = il/2; // 0 or 1. 0 computes 0,32 + 128,160, 1 computes 64,96 + 192,224
const int in = il%2;
const int l0 = n*(2*ir + in);
const int q_offset = 32*im + l0;
const int y_offset = 64*im + l0;
uint16_t aux[4];
const uint8_t * sc = (const uint8_t *)aux;
const block_q4_K * x = (const block_q4_K *)vx + ib0;
float tmp = 0; // partial sum for thread in warp
for (int i = ix; i < num_blocks_per_row; i += 2) {
const uint8_t * q1 = x[i].qs + q_offset;
const uint8_t * q2 = q1 + 64;
const float * y1 = yy + i*QK_K + y_offset;
const float * y2 = y1 + 128;
const float dall = x[i].d;
const float dmin = x[i].dmin;
const uint16_t * a = (const uint16_t *)x[i].scales;
aux[0] = a[im+0] & kmask1;
aux[1] = a[im+2] & kmask1;
aux[2] = ((a[im+4] >> 0) & kmask2) | ((a[im+0] & kmask3) >> 2);
aux[3] = ((a[im+4] >> 4) & kmask2) | ((a[im+2] & kmask3) >> 2);
float4 s = {0.f, 0.f, 0.f, 0.f};
float smin = 0;
for (int l = 0; l < n; ++l) {
s.x += y1[l] * (q1[l] & 0xF); s.y += y1[l+32] * (q1[l] >> 4);
s.z += y2[l] * (q2[l] & 0xF); s.w += y2[l+32] * (q2[l] >> 4);
smin += y1[l] * sc[2] + y1[l+32] * sc[3] + y2[l] * sc[6] + y2[l+32] * sc[7];
}
tmp += dall * (s.x * sc[0] + s.y * sc[1] + s.z * sc[4] + s.w * sc[5]) - dmin * smin;
}
// sum up partial sums and write back result
__syncthreads();
#pragma unroll
for (int mask = 16; mask > 0; mask >>= 1) {
tmp += __shfl_xor_sync(0xffffffff, tmp, mask, 32);
}
if (tid == 0) {
dst[row] = tmp;
}
}
static __global__ void dequantize_mul_mat_vec_q5_k(const void * vx, const float * yy, float * dst, const int ncols) {
const uint16_t kmask1 = 0x3f3f;
const uint16_t kmask2 = 0x0f0f;
const uint16_t kmask3 = 0xc0c0;
//const int row = blockIdx.x*blockDim.y + threadIdx.y;
const int row = blockIdx.x;
const int num_blocks_per_row = ncols / QK_K;
const int ib0 = row*num_blocks_per_row;
const int tid = threadIdx.x/2; // 0...15
const int ix = threadIdx.x%2;
const int il = tid/4; // 0...3
const int ir = tid - 4*il;// 0...3
const int n = 4;
const int im = il/2; // 0 or 1. 0 computes 0,32 + 128,160, 1 computes 64,96 + 192,224
const int in = il%2;
const int l0 = n*(2*ir + in);
const int q_offset = 32*im + l0;
const int y_offset = 64*im + l0;
const uint8_t hm1 = 1 << (2*im);
const uint8_t hm2 = hm1 << 4;
uint16_t aux[4];
const uint8_t * sc = (const uint8_t *)aux;
const block_q5_K * x = (const block_q5_K *)vx + ib0;
float tmp = 0; // partial sum for thread in warp
for (int i = ix; i < num_blocks_per_row; i += 2) {
const uint8_t * ql1 = x[i].qs + q_offset;
const uint8_t * ql2 = ql1 + 64;
const uint8_t * qh = x[i].qh + l0;
const float * y1 = yy + i*QK_K + y_offset;
const float * y2 = y1 + 128;
const float dall = x[i].d;
const float dmin = x[i].dmin;
const uint16_t * a = (const uint16_t *)x[i].scales;
aux[0] = a[im+0] & kmask1;
aux[1] = a[im+2] & kmask1;
aux[2] = ((a[im+4] >> 0) & kmask2) | ((a[im+0] & kmask3) >> 2);
aux[3] = ((a[im+4] >> 4) & kmask2) | ((a[im+2] & kmask3) >> 2);
float4 sum = {0.f, 0.f, 0.f, 0.f};
float smin = 0;
for (int l = 0; l < n; ++l) {
sum.x += y1[l+ 0] * ((ql1[l] & 0xF) + (qh[l] & (hm1 << 0) ? 16 : 0));
sum.y += y1[l+32] * ((ql1[l] >> 4) + (qh[l] & (hm1 << 1) ? 16 : 0));
sum.z += y2[l+ 0] * ((ql2[l] & 0xF) + (qh[l] & (hm2 << 0) ? 16 : 0));
sum.w += y2[l+32] * ((ql2[l] >> 4) + (qh[l] & (hm2 << 1) ? 16 : 0));
smin += y1[l] * sc[2] + y1[l+32] * sc[3] + y2[l] * sc[6] + y2[l+32] * sc[7];
}
tmp += dall * (sum.x * sc[0] + sum.y * sc[1] + sum.z * sc[4] + sum.w * sc[5]) - dmin * smin;
}
// sum up partial sums and write back result
__syncthreads();
#pragma unroll
for (int mask = 16; mask > 0; mask >>= 1) {
tmp += __shfl_xor_sync(0xffffffff, tmp, mask, 32);
}
if (tid == 0) {
dst[row] = tmp;
}
}
static __global__ void dequantize_mul_mat_vec_q6_k(const void * vx, const float * yy, float * dst, const int ncols, int nrows) {
static_assert(16%K_QUANTS_PER_ITERATION == 0, "16 must be divisible by K_QUANTS_PER_ITERATION");
const int row = blockIdx.y*blockDim.y + threadIdx.y;
if (row > nrows) return;
const int num_blocks_per_row = ncols / QK_K;
const int ib0 = row*num_blocks_per_row;
const block_q6_K * x = (const block_q6_K *)vx + ib0;
const int tid = threadIdx.x/K_QUANTS_PER_ITERATION; // 0...31 or 0...16
const int ix = threadIdx.x%K_QUANTS_PER_ITERATION; // 0 or 0, 1
const int step = 16/K_QUANTS_PER_ITERATION; // 16 or 8
const int im = tid/step; // 0 or 1. 0 computes 0..., 1 computes 128...
const int in = tid - step*im; // 0...15 or 0...7
#if K_QUANTS_PER_ITERATION == 1
const int l0 = K_QUANTS_PER_ITERATION*in; // 0...15
const int is = 0;
#else
const int l0 = 4 * in; // 0, 4, 8, ..., 28
const int is = in / 4;
#endif
const int ql_offset = 64*im + l0;
const int qh_offset = 32*im + l0;
const int s_offset = 8*im + is;
const int y_offset = 128*im + l0;
float tmp = 0; // partial sum for thread in warp
for (int i = ix; i < num_blocks_per_row; i += K_QUANTS_PER_ITERATION) {
const float * y = yy + i * QK_K + y_offset;
const uint8_t * ql = x[i].ql + ql_offset;
const uint8_t * qh = x[i].qh + qh_offset;
const int8_t * s = x[i].scales + s_offset;
const float d = x[i].d;
#if K_QUANTS_PER_ITERATION == 1
float sum = y[ 0] * s[0] * d * ((int8_t)((ql[ 0] & 0xF) | ((qh[ 0] & 0x03) << 4)) - 32)
+ y[16] * s[1] * d * ((int8_t)((ql[16] & 0xF) | ((qh[16] & 0x03) << 4)) - 32)
+ y[32] * s[2] * d * ((int8_t)((ql[32] & 0xF) | ((qh[ 0] & 0x0c) << 2)) - 32)
+ y[48] * s[3] * d * ((int8_t)((ql[48] & 0xF) | ((qh[16] & 0x0c) << 2)) - 32)
+ y[64] * s[4] * d * ((int8_t)((ql[ 0] >> 4) | ((qh[ 0] & 0x30) >> 0)) - 32)
+ y[80] * s[5] * d * ((int8_t)((ql[16] >> 4) | ((qh[16] & 0x30) >> 0)) - 32)
+ y[96] * s[6] * d * ((int8_t)((ql[32] >> 4) | ((qh[ 0] & 0xc0) >> 2)) - 32)
+y[112] * s[7] * d * ((int8_t)((ql[48] >> 4) | ((qh[16] & 0xc0) >> 2)) - 32);
tmp += sum;
#else
float sum = 0;
for (int l = 0; l < 4; ++l) {
sum += y[l+ 0] * s[0] * d * ((int8_t)((ql[l+ 0] & 0xF) | (((qh[l] >> 0) & 3) << 4)) - 32)
+ y[l+32] * s[2] * d * ((int8_t)((ql[l+32] & 0xF) | (((qh[l] >> 2) & 3) << 4)) - 32)
+ y[l+64] * s[4] * d * ((int8_t)((ql[l+ 0] >> 4) | (((qh[l] >> 4) & 3) << 4)) - 32)
+ y[l+96] * s[6] * d * ((int8_t)((ql[l+32] >> 4) | (((qh[l] >> 6) & 3) << 4)) - 32);
}
tmp += sum;
#endif
}
// sum up partial sums and write back result
__syncthreads();
#pragma unroll
for (int mask = 16; mask > 0; mask >>= 1) {
tmp += __shfl_xor_sync(0xffffffff, tmp, mask, 32);
}
if (tid == 0) {
dst[row] = tmp;
}
}
static __device__ void convert_f16(const void * vx, const int ib, const int iqs, float & v0, float & v1){
@ -712,46 +918,6 @@ static __global__ void dequantize_mul_mat_vec(const void * vx, const float * y,
}
}
template <int n_thread, dot_kernel_k_t dot_kernel>
static __global__ void dequantize_mul_mat_vec_k(const void * vx, const float * y, float * dst, const int ncols, const int nrows) {
const int row = blockIdx.y*blockDim.y + threadIdx.y;
if (row >= nrows) {
return;
}
const int tid = threadIdx.x;
const int iter_stride = QK_K;
const int vals_per_iter = iter_stride / n_thread;
const int num_blocks_per_row = ncols / QK_K;
const int ib0 = row*num_blocks_per_row;
float tmp = 0; // partial sum for thread in warp
for (int i = 0; i < ncols; i += iter_stride) {
const int col = i + vals_per_iter*tid;
const int ib = ib0 + col/QK_K; // x block index
const int iqs = col%QK_K; // x quant index
const int iybs = col - col%QK_K; // y block start index
float v;
dot_kernel(vx, ib, iqs, y + iybs, v);
tmp += v;
}
// sum up partial sums and write back result
__syncthreads();
#pragma unroll
for (int mask = 16; mask > 0; mask >>= 1) {
tmp += __shfl_xor_sync(0xffffffff, tmp, mask, 32);
}
if (tid == 0) {
dst[row] = tmp;
}
}
static __global__ void mul_mat_p021_f16_f32(const void * vx, const float * y, float * dst, const int ncols_x, const int nrows_x, const int nchannels_x) {
const half * x = (half *) vx;
@ -1094,43 +1260,34 @@ static void dequantize_mul_mat_vec_q2_K_cuda(const void * vx, const float * y, f
const int block_num_y = (nrows + ny - 1) / ny;
const dim3 block_nums(1, block_num_y, 1);
const dim3 block_dims(32, ny, 1);
dequantize_mul_mat_vec_k<32, vec_dot_q2_K><<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
dequantize_mul_mat_vec_q2_k<<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
}
static void dequantize_mul_mat_vec_q3_K_cuda(const void * vx, const float * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
GGML_ASSERT(ncols % QK_K == 0);
const int ny = 2;
const int block_num_y = (nrows + ny - 1) / ny;
const dim3 block_nums(1, block_num_y, 1);
const dim3 block_dims(32, ny, 1);
dequantize_mul_mat_vec_k<32, vec_dot_q3_K><<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
const dim3 block_dims(32, 1, 1);
dequantize_mul_mat_vec_q3_k<<<nrows, block_dims, 0, stream>>>(vx, y, dst, ncols);
}
static void dequantize_mul_mat_vec_q4_K_cuda(const void * vx, const float * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
GGML_ASSERT(ncols % QK_K == 0);
const int ny = 2;
const int block_num_y = (nrows + ny - 1) / ny;
const dim3 block_nums(1, block_num_y, 1);
const dim3 block_dims(32, ny, 1);
dequantize_mul_mat_vec_k<32, vec_dot_q4_K><<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
const dim3 block_dims(32, 1, 1);
dequantize_mul_mat_vec_q4_k<<<nrows, block_dims, 0, stream>>>(vx, y, dst, ncols);
}
static void dequantize_mul_mat_vec_q5_K_cuda(const void * vx, const float * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
GGML_ASSERT(ncols % QK_K == 0);
const int ny = 2;
const int block_num_y = (nrows + ny - 1) / ny;
const dim3 block_nums(1, block_num_y, 1);
const dim3 block_dims(32, ny, 1);
dequantize_mul_mat_vec_k<32, vec_dot_q5_K><<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
const dim3 block_dims(32, 1, 1);
dequantize_mul_mat_vec_q5_k<<<nrows, block_dims, 0, stream>>>(vx, y, dst, ncols);
}
static void dequantize_mul_mat_vec_q6_K_cuda(const void * vx, const float * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
GGML_ASSERT(ncols % QK_K == 0);
const int ny = 2;
const int ny = 2 / K_QUANTS_PER_ITERATION;
const int block_num_y = (nrows + ny - 1) / ny;
const dim3 block_nums(1, block_num_y, 1);
const dim3 block_dims(32, ny, 1);
dequantize_mul_mat_vec_k<32, vec_dot_q6_K><<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
dequantize_mul_mat_vec_q6_k<<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
}
static void convert_fp16_to_fp32_cuda(const void * vx, float * y, const int k, cudaStream_t stream) {

6
ggml.c
View file

@ -35,6 +35,12 @@
#define static_assert(cond, msg) struct global_scope_noop_trick
#endif
#if defined(_MSC_VER)
// disable "possible loss of data" to avoid hundreds of casts
// we should just be careful :)
#pragma warning(disable: 4244 4267)
#endif
#if defined(_WIN32)
#include <windows.h>

6
llama.cpp
View file

@ -40,6 +40,10 @@
#include <sstream>
#include <numeric>
#if defined(_MSC_VER)
#pragma warning(disable: 4244 4267) // possible loss of data
#endif
#define LLAMA_USE_SCRATCH
#define LLAMA_MAX_SCRATCH_BUFFERS 16
@ -1654,7 +1658,7 @@ static bool llama_eval_internal(
// cur = cur*norm(broadcasted)
cur = ggml_mul(ctx0, cur, model.norm);
offload_func_nr(cur);
// offload_func_nr(cur); // TODO CPU + GPU mirrored backend
ggml_set_name(cur, "result_norm");
embeddings = cur;