Public/koboldcpp
1
0
Fork 0
mirror of https://github.com/LostRuins/koboldcpp.git synced 2026-09-19 17:25:07 +02:00
Code Issues Packages Projects Releases Wiki Activity

Merge branch 'upstream' into concedo_experimental

Browse Source 
# Conflicts:
#	.devops/full-cuda.Dockerfile
#	.devops/full-rocm.Dockerfile
#	.devops/full.Dockerfile
#	.devops/llama-cli-cuda.Dockerfile
#	.devops/llama-cli-intel.Dockerfile
#	.devops/llama-cli-rocm.Dockerfile
#	.devops/llama-cli-vulkan.Dockerfile
#	.devops/llama-cli.Dockerfile
#	.devops/llama-server-cuda.Dockerfile
#	.devops/llama-server-intel.Dockerfile
#	.devops/llama-server-rocm.Dockerfile
#	.devops/llama-server-vulkan.Dockerfile
#	.devops/llama-server.Dockerfile
#	CMakeLists.txt
#	CONTRIBUTING.md
#	Makefile
#	ggml/CMakeLists.txt
#	ggml/src/CMakeLists.txt
#	requirements.txt
#	src/llama.cpp
#	tests/test-backend-ops.cpp
This commit is contained in:
Concedo
2024-07-19 14:23:33 +08:00
parent a998588f3a a15ef8f8a0
commit 24b9616344
61 changed files with 12994 additions and 936 deletions
Download Patch File Download Diff File Expand all files Collapse all files
examples/batched/batched.cpp
+1 -1
View File
@@ -31,7 +31,7 @@ int main(int argc, char ** argv) {
int n_parallel = params.n_parallel;
// total length of the sequences including the prompt
int n_predict = 32;
int n_predict = params.n_predict;
// init LLM
examples/convert_legacy_llama.py
+126 -105
View File
@@ -24,7 +24,7 @@ from abc import ABC, abstractmethod
from concurrent.futures import ProcessPoolExecutor, ThreadPoolExecutor
from dataclasses import dataclass
from pathlib import Path
from typing import TYPE_CHECKING, Any, Callable, IO, Iterable, Literal, TypeVar, Optional
from typing import TYPE_CHECKING, Any, Callable, IO, Iterable, Literal, TypeVar
import numpy as np
@@ -346,42 +346,6 @@ class Params:
return params
@dataclass
class Metadata:
name: Optional[str] = None
author: Optional[str] = None
version: Optional[str] = None
url: Optional[str] = None
description: Optional[str] = None
license: Optional[str] = None
source_url: Optional[str] = None
source_hf_repo: Optional[str] = None
@staticmethod
def load(metadata_path: Path) -> Metadata:
if metadata_path is None or not metadata_path.exists():
return Metadata()
with open(metadata_path, 'r') as file:
data = json.load(file)
# Create a new Metadata instance
metadata = Metadata()
# Assigning values to Metadata attributes if they exist in the JSON file
# This is based on LLM_KV_NAMES mapping in llama.cpp
metadata.name = data.get("general.name")
metadata.author = data.get("general.author")
metadata.version = data.get("general.version")
metadata.url = data.get("general.url")
metadata.description = data.get("general.description")
metadata.license = data.get("general.license")
metadata.source_url = data.get("general.source.url")
metadata.source_hf_repo = data.get("general.source.huggingface.repository")
return metadata
#
# data loading
# TODO: reuse (probably move to gguf.py?)
@@ -806,7 +770,7 @@ class OutputFile:
def __init__(self, fname_out: Path, endianess:gguf.GGUFEndian = gguf.GGUFEndian.LITTLE):
self.gguf = gguf.GGUFWriter(fname_out, gguf.MODEL_ARCH_NAMES[ARCH], endianess=endianess)
def add_meta_model(self, params: Params, metadata: Metadata | None) -> None:
def add_meta_model(self, params: Params, metadata: gguf.Metadata | None) -> None:
# Metadata About The Model And Its Provenence
name = "LLaMA"
if metadata is not None and metadata.name is not None:
@@ -824,16 +788,73 @@ class OutputFile:
self.gguf.add_author(metadata.author)
if metadata.version is not None:
self.gguf.add_version(metadata.version)
if metadata.url is not None:
self.gguf.add_url(metadata.url)
if metadata.organization is not None:
self.gguf.add_organization(metadata.organization)
if metadata.finetune is not None:
self.gguf.add_finetune(metadata.finetune)
if metadata.basename is not None:
self.gguf.add_basename(metadata.basename)
if metadata.description is not None:
self.gguf.add_description(metadata.description)
if metadata.quantized_by is not None:
self.gguf.add_quantized_by(metadata.quantized_by)
if metadata.size_label is not None:
self.gguf.add_size_label(metadata.size_label)
if metadata.license is not None:
self.gguf.add_licence(metadata.license)
self.gguf.add_license(metadata.license)
if metadata.license_name is not None:
self.gguf.add_license_name(metadata.license_name)
if metadata.license_link is not None:
self.gguf.add_license_link(metadata.license_link)
if metadata.url is not None:
self.gguf.add_url(metadata.url)
if metadata.doi is not None:
self.gguf.add_doi(metadata.doi)
if metadata.uuid is not None:
self.gguf.add_uuid(metadata.uuid)
if metadata.repo_url is not None:
self.gguf.add_repo_url(metadata.repo_url)
if metadata.source_url is not None:
self.gguf.add_source_url(metadata.source_url)
if metadata.source_hf_repo is not None:
self.gguf.add_source_hf_repo(metadata.source_hf_repo)
if metadata.source_doi is not None:
self.gguf.add_source_doi(metadata.source_doi)
if metadata.source_uuid is not None:
self.gguf.add_source_uuid(metadata.source_uuid)
if metadata.source_repo_url is not None:
self.gguf.add_source_repo_url(metadata.source_repo_url)
if metadata.base_models is not None:
self.gguf.add_base_model_count(len(metadata.base_models))
for key, base_model_entry in enumerate(metadata.base_models):
if "name" in base_model_entry:
self.gguf.add_base_model_name(key, base_model_entry["name"])
if "author" in base_model_entry:
self.gguf.add_base_model_author(key, base_model_entry["author"])
if "version" in base_model_entry:
self.gguf.add_base_model_version(key, base_model_entry["version"])
if "organization" in base_model_entry:
self.gguf.add_base_model_organization(key, base_model_entry["organization"])
if "url" in base_model_entry:
self.gguf.add_base_model_url(key, base_model_entry["url"])
if "doi" in base_model_entry:
self.gguf.add_base_model_doi(key, base_model_entry["doi"])
if "uuid" in base_model_entry:
self.gguf.add_base_model_uuid(key, base_model_entry["uuid"])
if "repo_url" in base_model_entry:
self.gguf.add_base_model_repo_url(key, base_model_entry["repo_url"])
if metadata.tags is not None:
self.gguf.add_tags(metadata.tags)
if metadata.languages is not None:
self.gguf.add_languages(metadata.languages)
if metadata.datasets is not None:
self.gguf.add_datasets(metadata.datasets)
def add_meta_arch(self, params: Params) -> None:
# Metadata About The Neural Architecture Itself
@@ -944,7 +965,7 @@ class OutputFile:
@staticmethod
def write_vocab_only(
fname_out: Path, params: Params, vocab: Vocab, svocab: gguf.SpecialVocab,
endianess: gguf.GGUFEndian = gguf.GGUFEndian.LITTLE, pad_vocab: bool = False, metadata: Metadata | None = None,
endianess: gguf.GGUFEndian = gguf.GGUFEndian.LITTLE, pad_vocab: bool = False, metadata: gguf.Metadata | None = None,
) -> None:
check_vocab_size(params, vocab, pad_vocab=pad_vocab)
@@ -978,7 +999,7 @@ class OutputFile:
fname_out: Path, ftype: GGMLFileType, params: Params, model: LazyModel, vocab: BaseVocab, svocab: gguf.SpecialVocab,
concurrency: int = DEFAULT_CONCURRENCY, endianess: gguf.GGUFEndian = gguf.GGUFEndian.LITTLE,
pad_vocab: bool = False,
metadata: Metadata | None = None,
metadata: gguf.Metadata | None = None,
) -> None:
check_vocab_size(params, vocab, pad_vocab=pad_vocab)
@@ -1021,35 +1042,32 @@ def pick_output_type(model: LazyModel, output_type_str: str | None) -> GGMLFileT
raise ValueError(f"Unexpected combination of types: {name_to_type}")
def model_parameter_count(model: LazyModel) -> int:
total_model_parameters = 0
for i, (name, lazy_tensor) in enumerate(model.items()):
sum_weights_in_tensor = 1
def per_model_weight_count_estimation(tensors: Iterable[tuple[str, LazyTensor]]) -> tuple[int, int, int]:
total_params = 0
shared_params = 0
expert_params = 0
for name, lazy_tensor in tensors:
# We don't need these
if name.endswith((".attention.masked_bias", ".attention.bias", ".rotary_emb.inv_freq")):
continue
# Got A Tensor
sum_weights_in_tensor: int = 1
# Tensor Volume
for dim in lazy_tensor.shape:
sum_weights_in_tensor *= dim
total_model_parameters += sum_weights_in_tensor
return total_model_parameters
if ".experts." in name:
if ".experts.0." in name:
expert_params += sum_weights_in_tensor
else:
shared_params += sum_weights_in_tensor
def model_parameter_count_rounded_notation(model_params_count: int) -> str:
if model_params_count > 1e12 :
# Trillions Of Parameters
scaled_model_params = model_params_count * 1e-12
scale_suffix = "T"
elif model_params_count > 1e9 :
# Billions Of Parameters
scaled_model_params = model_params_count * 1e-9
scale_suffix = "B"
elif model_params_count > 1e6 :
# Millions Of Parameters
scaled_model_params = model_params_count * 1e-6
scale_suffix = "M"
else:
# Thousands Of Parameters
scaled_model_params = model_params_count * 1e-3
scale_suffix = "K"
total_params += sum_weights_in_tensor
return f"{round(scaled_model_params)}{scale_suffix}"
return total_params, shared_params, expert_params
def convert_to_output_type(model: LazyModel, output_type: GGMLFileType) -> LazyModel:
@@ -1231,34 +1249,24 @@ class VocabFactory:
return vocab, special_vocab
def default_convention_outfile(file_type: GGMLFileType, params: Params, model_params_count: int, metadata: Metadata) -> str:
quantization = {
def default_convention_outfile(file_type: GGMLFileType, expert_count: int | None, model_params_count: tuple[int, int, int], metadata: gguf.Metadata) -> str:
name = metadata.name if metadata.name is not None else None
basename = metadata.basename if metadata.basename is not None else None
finetune = metadata.finetune if metadata.finetune is not None else None
version = metadata.version if metadata.version is not None else None
size_label = metadata.size_label if metadata.size_label is not None else gguf.size_label(*model_params_count, expert_count=expert_count or 0)
output_type = {
GGMLFileType.AllF32: "F32",
GGMLFileType.MostlyF16: "F16",
GGMLFileType.MostlyQ8_0: "Q8_0",
}[file_type]
parameters = model_parameter_count_rounded_notation(model_params_count)
expert_count = ""
if params.n_experts is not None:
expert_count = f"{params.n_experts}x"
version = ""
if metadata is not None and metadata.version is not None:
version = f"-{metadata.version}"
name = "ggml-model"
if metadata is not None and metadata.name is not None:
name = metadata.name
elif params.path_model is not None:
name = params.path_model.name
return f"{name}{version}-{expert_count}{parameters}-{quantization}"
return gguf.naming_convention(name, basename, finetune, version, size_label, output_type)
def default_outfile(model_paths: list[Path], file_type: GGMLFileType, params: Params, model_params_count: int, metadata: Metadata) -> Path:
default_filename = default_convention_outfile(file_type, params, model_params_count, metadata)
def default_outfile(model_paths: list[Path], file_type: GGMLFileType, expert_count: int | None, model_params_count: tuple[int, int, int], metadata: gguf.Metadata) -> Path:
default_filename = default_convention_outfile(file_type, expert_count, model_params_count, metadata)
ret = model_paths[0].parent / f"{default_filename}.gguf"
if ret in model_paths:
logger.error(
@@ -1297,8 +1305,9 @@ def main(args_in: list[str] | None = None) -> None:
parser.add_argument("--pad-vocab", action="store_true", help="add pad tokens when model vocab expects more than tokenizer metadata provides")
parser.add_argument("--skip-unknown", action="store_true", help="skip unknown tensor names instead of failing")
parser.add_argument("--verbose", action="store_true", help="increase output verbosity")
parser.add_argument("--metadata", type=Path, help="Specify the path for a metadata file")
parser.add_argument("--metadata", type=Path, help="Specify the path for an authorship metadata override file")
parser.add_argument("--get-outfile", action="store_true", help="get calculated default outfile name")
parser.add_argument("--model-name", type=str, default=None, help="name of the model")
args = parser.parse_args(args_in)
@@ -1310,32 +1319,36 @@ def main(args_in: list[str] | None = None) -> None:
else:
logging.basicConfig(level=logging.INFO)
metadata = Metadata.load(args.metadata)
model_name = args.model_name
dir_model = args.model
metadata = gguf.Metadata.load(args.metadata, dir_model, model_name)
if args.get_outfile:
model_plus = load_some_model(args.model)
model_plus = load_some_model(dir_model)
params = Params.load(model_plus)
model = convert_model_names(model_plus.model, params, args.skip_unknown)
model_params_count = model_parameter_count(model_plus.model)
ftype = pick_output_type(model, args.outtype)
print(f"{default_convention_outfile(ftype, params, model_params_count, metadata)}") # noqa: NP100
model = convert_model_names(model_plus.model, params, args.skip_unknown)
model_params_count = per_model_weight_count_estimation(model_plus.model.items())
ftype = pick_output_type(model, args.outtype)
if (metadata is None or metadata.name is None) and params.path_model is not None:
metadata.name = params.path_model.name
print(f"{default_convention_outfile(ftype, params.n_experts, model_params_count, metadata)}") # noqa: NP100
return
if args.no_vocab and args.vocab_only:
raise ValueError("--vocab-only does not make sense with --no-vocab")
if args.dump_single:
model_plus = lazy_load_file(args.model)
model_plus = lazy_load_file(dir_model)
do_dump_model(model_plus)
return
if not args.vocab_only:
model_plus = load_some_model(args.model)
model_plus = load_some_model(dir_model)
else:
model_plus = ModelPlus(model = {}, paths = [args.model / 'dummy'], format = 'none', vocab = None)
model_params_count = model_parameter_count(model_plus.model)
logger.info(f"model parameters count : {model_params_count} ({model_parameter_count_rounded_notation(model_params_count)})")
model_plus = ModelPlus(model = {}, paths = [dir_model / 'dummy'], format = 'none', vocab = None)
if args.dump:
do_dump_model(model_plus)
@@ -1368,7 +1381,7 @@ def main(args_in: list[str] | None = None) -> None:
logger.info(f"params = {params}")
model_parent_path = model_plus.paths[0].parent
vocab_path = Path(args.vocab_dir or args.model or model_parent_path)
vocab_path = Path(args.vocab_dir or dir_model or model_parent_path)
vocab_factory = VocabFactory(vocab_path)
vocab_types = None if args.no_vocab else args.vocab_type.split(",")
vocab, special_vocab = vocab_factory.load_vocab(vocab_types, model_parent_path)
@@ -1399,13 +1412,21 @@ def main(args_in: list[str] | None = None) -> None:
assert params is not None
if metadata.name is None and params.path_model is not None:
metadata.name = params.path_model.name
model_params_count = per_model_weight_count_estimation(model_plus.model.items())
logger.info(f"model parameters count : {model_params_count} ({gguf.model_weight_count_rounded_notation(model_params_count[0])})")
logger.info(f"Vocab info: {vocab}")
logger.info(f"Special vocab info: {special_vocab}")
model = model_plus.model
model = convert_model_names(model, params, args.skip_unknown)
ftype = pick_output_type(model, args.outtype)
model = convert_to_output_type(model, ftype)
outfile = args.outfile or default_outfile(model_plus.paths, ftype, params, model_params_count, metadata)
outfile = args.outfile or default_outfile(model_plus.paths, ftype, params.n_experts, model_params_count, metadata=metadata)
metadata.size_label = gguf.size_label(*model_params_count, expert_count=params.n_experts or 0)
params.ftype = ftype
logger.info(f"Writing {outfile}, format {ftype}")
examples/export-lora/export-lora.cpp
+3
View File
@@ -190,6 +190,9 @@ static bool export_lora_params_parse(int argc, char ** argv, struct export_lora_
if (params->n_threads <= 0) {
params->n_threads = std::thread::hardware_concurrency();
}
} else if (arg == "-h" || arg == "--help") {
export_lora_print_usage(argc, argv, &default_params);
exit(0);
} else {
fprintf(stderr, "error: unknown argument: '%s'\n", arg.c_str());
export_lora_print_usage(argc, argv, &default_params);
examples/gguf-hash/README.md
+1 -1
View File
@@ -201,6 +201,6 @@ Verification results for test.gguf.manifest - Success
These micro c libraries dependencies was installed via the [clib c package manager](https://github.com/clibs)
- https://github.com/mofosyne/xxHash (From: https://github.com/Cyan4973/xxHash)
- https://github.com/Cyan4973/xxHash
- https://github.com/clibs/sha1/
- https://github.com/jb55/sha256.c
examples/gguf-hash/deps/xxhash/clib.json
+1 -1
View File
@@ -1,7 +1,7 @@
{
"name": "xxhash",
"version": "0.8.2",
"repo": "mofosyne/xxhash",
"repo": "Cyan4973/xxhash",
"description": "Extremely fast non-cryptographic hash algorithm",
"keywords": ["xxhash", "hashing"],
"license": "BSD-2-Clause",
examples/llama-bench/llama-bench.cpp
+15
View File
@@ -24,6 +24,10 @@
#include "ggml-cuda.h"
#include "ggml-sycl.h"
#ifdef GGML_USE_CANN
#include "ggml-cann.h"
#endif
// utils
static uint64_t get_time_ns() {
using clock = std::chrono::high_resolution_clock;
@@ -121,6 +125,17 @@ static std::string get_gpu_info() {
id += "/";
}
}
#endif
#ifdef GGML_USE_CANN
uint32_t count = ggml_backend_cann_get_device_count();
for (uint32_t i = 0; i < count; i++) {
char buf[128];
ggml_backend_cann_get_device_description(i, buf, sizeof(buf));
id += buf;
if (i < count - 1) {
id += "/";
}
}
#endif
// TODO: other backends
return id;
examples/llava/clip.cpp
+9
View File
@@ -16,6 +16,10 @@
#include "ggml-metal.h"
#endif
#ifdef GGML_USE_CANN
#include "ggml-cann.h"
#endif
#define STB_IMAGE_IMPLEMENTATION
#include "stb_image.h"
@@ -1001,6 +1005,11 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
LOG_TEE("%s: CLIP using Metal backend\n", __func__);
#endif
#ifdef GGML_USE_CANN
new_clip->backend = ggml_backend_cann_init(0);
LOG_TEE("%s: CLIP using CANN backend\n", __func__);
#endif
if (!new_clip->backend) {
new_clip->backend = ggml_backend_cpu_init();
examples/lookup/lookup-stats.cpp
+1 -2
View File
@@ -31,7 +31,6 @@ int main(int argc, char ** argv){
// load the model
std::tie(model, ctx) = llama_init_from_gpt_params(params);
GGML_ASSERT(llama_n_vocab(model) < (1 << 16));
// tokenize the prompt
std::vector<llama_token> inp;
@@ -65,7 +64,7 @@ int main(int argc, char ** argv){
}
const int n_input = inp.size();
const int n_ctx = params.n_ctx;
const int n_ctx = llama_n_ctx(ctx);
int n_drafted = 0;
int n_accept = 0;
examples/lookup/lookup.cpp
-1
View File
@@ -39,7 +39,6 @@ int main(int argc, char ** argv){
// load the model
std::tie(model, ctx) = llama_init_from_gpt_params(params);
GGML_ASSERT(llama_n_vocab(model) < (1 << 16));
// tokenize the prompt
std::vector<llama_token> inp;
examples/quantize/quantize.cpp
+37 -37
View File
@@ -17,44 +17,44 @@ struct quant_option {
};
static const std::vector<struct quant_option> QUANT_OPTIONS = {
{ "Q4_0", LLAMA_FTYPE_MOSTLY_Q4_0, " 4.34G, +0.4685 ppl @ Llama-3-8B", },
{ "Q4_1", LLAMA_FTYPE_MOSTLY_Q4_1, " 4.78G, +0.4511 ppl @ Llama-3-8B", },
{ "Q5_0", LLAMA_FTYPE_MOSTLY_Q5_0, " 5.21G, +0.1316 ppl @ Llama-3-8B", },
{ "Q5_1", LLAMA_FTYPE_MOSTLY_Q5_1, " 5.65G, +0.1062 ppl @ Llama-3-8B", },
{ "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", },
{ "IQ1_M", LLAMA_FTYPE_MOSTLY_IQ1_M, " 1.75 bpw quantization", },
{ "Q2_K", LLAMA_FTYPE_MOSTLY_Q2_K, " 2.96G, +3.5199 ppl @ Llama-3-8B", },
{ "Q2_K_S", LLAMA_FTYPE_MOSTLY_Q2_K_S, " 2.96G, +3.1836 ppl @ Llama-3-8B", },
{ "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" },
{ "IQ3_XS", LLAMA_FTYPE_MOSTLY_IQ3_XS, " 3.3 bpw quantization", },
{ "Q3_K_S", LLAMA_FTYPE_MOSTLY_Q3_K_S, " 3.41G, +1.6321 ppl @ Llama-3-8B", },
{ "Q3_K_M", LLAMA_FTYPE_MOSTLY_Q3_K_M, " 3.74G, +0.6569 ppl @ Llama-3-8B", },
{ "Q3_K_L", LLAMA_FTYPE_MOSTLY_Q3_K_L, " 4.03G, +0.5562 ppl @ Llama-3-8B", },
{ "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, " 4.37G, +0.2689 ppl @ Llama-3-8B", },
{ "Q4_K_M", LLAMA_FTYPE_MOSTLY_Q4_K_M, " 4.58G, +0.1754 ppl @ Llama-3-8B", },
{ "Q5_K", LLAMA_FTYPE_MOSTLY_Q5_K_M, "alias for Q5_K_M", },
{ "Q5_K_S", LLAMA_FTYPE_MOSTLY_Q5_K_S, " 5.21G, +0.1049 ppl @ Llama-3-8B", },
{ "Q5_K_M", LLAMA_FTYPE_MOSTLY_Q5_K_M, " 5.33G, +0.0569 ppl @ Llama-3-8B", },
{ "Q6_K", LLAMA_FTYPE_MOSTLY_Q6_K, " 6.14G, +0.0217 ppl @ Llama-3-8B", },
{ "Q8_0", LLAMA_FTYPE_MOSTLY_Q8_0, " 7.96G, +0.0026 ppl @ Llama-3-8B", },
{ "Q4_0_4_4", LLAMA_FTYPE_MOSTLY_Q4_0_4_4, " 4.34G, +0.4685 ppl @ Llama-3-8B", },
{ "Q4_0_4_8", LLAMA_FTYPE_MOSTLY_Q4_0_4_8, " 4.34G, +0.4685 ppl @ Llama-3-8B", },
{ "Q4_0_8_8", LLAMA_FTYPE_MOSTLY_Q4_0_8_8, " 4.34G, +0.4685 ppl @ Llama-3-8B", },
{ "F16", LLAMA_FTYPE_MOSTLY_F16, "14.00G, +0.0020 ppl @ Mistral-7B", },
{ "BF16", LLAMA_FTYPE_MOSTLY_BF16, "14.00G, -0.0050 ppl @ Mistral-7B", },
{ "F32", LLAMA_FTYPE_ALL_F32, "26.00G @ 7B", },
{ "Q4_0", LLAMA_FTYPE_MOSTLY_Q4_0, " 4.34G, +0.4685 ppl @ Llama-3-8B", },
{ "Q4_1", LLAMA_FTYPE_MOSTLY_Q4_1, " 4.78G, +0.4511 ppl @ Llama-3-8B", },
{ "Q5_0", LLAMA_FTYPE_MOSTLY_Q5_0, " 5.21G, +0.1316 ppl @ Llama-3-8B", },
{ "Q5_1", LLAMA_FTYPE_MOSTLY_Q5_1, " 5.65G, +0.1062 ppl @ Llama-3-8B", },
{ "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", },
{ "IQ1_M", LLAMA_FTYPE_MOSTLY_IQ1_M, " 1.75 bpw quantization", },
{ "Q2_K", LLAMA_FTYPE_MOSTLY_Q2_K, " 2.96G, +3.5199 ppl @ Llama-3-8B", },
{ "Q2_K_S", LLAMA_FTYPE_MOSTLY_Q2_K_S, " 2.96G, +3.1836 ppl @ Llama-3-8B", },
{ "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" },
{ "IQ3_XS", LLAMA_FTYPE_MOSTLY_IQ3_XS, " 3.3 bpw quantization", },
{ "Q3_K_S", LLAMA_FTYPE_MOSTLY_Q3_K_S, " 3.41G, +1.6321 ppl @ Llama-3-8B", },
{ "Q3_K_M", LLAMA_FTYPE_MOSTLY_Q3_K_M, " 3.74G, +0.6569 ppl @ Llama-3-8B", },
{ "Q3_K_L", LLAMA_FTYPE_MOSTLY_Q3_K_L, " 4.03G, +0.5562 ppl @ Llama-3-8B", },
{ "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, " 4.37G, +0.2689 ppl @ Llama-3-8B", },
{ "Q4_K_M", LLAMA_FTYPE_MOSTLY_Q4_K_M, " 4.58G, +0.1754 ppl @ Llama-3-8B", },
{ "Q5_K", LLAMA_FTYPE_MOSTLY_Q5_K_M, "alias for Q5_K_M", },
{ "Q5_K_S", LLAMA_FTYPE_MOSTLY_Q5_K_S, " 5.21G, +0.1049 ppl @ Llama-3-8B", },
{ "Q5_K_M", LLAMA_FTYPE_MOSTLY_Q5_K_M, " 5.33G, +0.0569 ppl @ Llama-3-8B", },
{ "Q6_K", LLAMA_FTYPE_MOSTLY_Q6_K, " 6.14G, +0.0217 ppl @ Llama-3-8B", },
{ "Q8_0", LLAMA_FTYPE_MOSTLY_Q8_0, " 7.96G, +0.0026 ppl @ Llama-3-8B", },
{ "Q4_0_4_4", LLAMA_FTYPE_MOSTLY_Q4_0_4_4, " 4.34G, +0.4685 ppl @ Llama-3-8B", },
{ "Q4_0_4_8", LLAMA_FTYPE_MOSTLY_Q4_0_4_8, " 4.34G, +0.4685 ppl @ Llama-3-8B", },
{ "Q4_0_8_8", LLAMA_FTYPE_MOSTLY_Q4_0_8_8, " 4.34G, +0.4685 ppl @ Llama-3-8B", },
{ "F16", LLAMA_FTYPE_MOSTLY_F16, "14.00G, +0.0020 ppl @ Mistral-7B", },
{ "BF16", LLAMA_FTYPE_MOSTLY_BF16, "14.00G, -0.0050 ppl @ Mistral-7B", },
{ "F32", LLAMA_FTYPE_ALL_F32, "26.00G @ 7B", },
// Note: Ensure COPY comes after F32 to avoid ftype 0 from matching.
{ "COPY", LLAMA_FTYPE_ALL_F32, "only copy tensors, no quantizing", },
{ "COPY", LLAMA_FTYPE_ALL_F32, "only copy tensors, no quantizing", },
};
static const char * const LLM_KV_QUANTIZE_IMATRIX_FILE = "quantize.imatrix.file";
examples/server/README.md
+2 -2
View File
@@ -211,7 +211,7 @@ retrieval:
--context-file FNAME file to load context from (repeat to specify multiple files)
--chunk-size N minimum length of embedded text chunks (default: 64)
--chunk-separator STRING
--chunk-separator STRING
separator between chunks (default: '
')
@@ -256,7 +256,7 @@ server:
--threads-http N number of threads used to process HTTP requests (default: -1)
--system-prompt-file FNAME
set a file to load a system prompt (initial prompt of all slots), this is useful for chat applications
--log-format {text,json}
--log-format {text,json}
log output format: json or text (default: json)
--metrics enable prometheus compatible metrics endpoint (default: disabled)
--no-slots disables slots monitoring endpoint (default: enabled)
examples/server/public/completion.js
+2 -2
View File
@@ -21,7 +21,7 @@ let generation_settings = null;
//
export async function* llama(prompt, params = {}, config = {}) {
let controller = config.controller;
const api_url = config.api_url || "";
const api_url = config.api_url?.replace(/\/+$/, '') || "";
if (!controller) {
controller = new AbortController();
@@ -196,7 +196,7 @@ export const llamaComplete = async (params, controller, callback) => {
// Get the model info from the server. This is useful for getting the context window and so on.
export const llamaModelInfo = async (config = {}) => {
if (!generation_settings) {
const api_url = config.api_url || "";
const api_url = config.api_url?.replace(/\/+$/, '') || "";
const props = await fetch(`${api_url}/props`).then(r => r.json());
generation_settings = props.default_generation_settings;
}
examples/server/public/index-new.html
+4 -4
View File
@@ -14,10 +14,10 @@
<script type="module">
import {
html, h, signal, effect, computed, render, useSignal, useEffect, useRef, Component
} from '/index.js';
} from './index.js';
import { llama } from '/completion.js';
import { SchemaConverter } from '/json-schema-to-grammar.mjs';
import { llama } from './completion.js';
import { SchemaConverter } from './json-schema-to-grammar.mjs';
import { promptFormats } from './prompt-formats.js';
import { systemPrompts } from './system-prompts.js'; // multilingual is wip
let selected_image = false;
@@ -225,7 +225,7 @@
throw new Error("already running");
}
controller.value = new AbortController();
for await (const chunk of llama(prompt, llamaParams, { controller: controller.value })) {
for await (const chunk of llama(prompt, llamaParams, { controller: controller.value, api_url: URL.parse('.', document.baseURI).href })) {
const data = chunk.data;
if (data.stop) {
while (
examples/server/public/index.html
+1 -1
View File
@@ -479,7 +479,7 @@
throw new Error("already running");
}
controller.value = new AbortController();
for await (const chunk of llama(prompt, llamaParams, { controller: controller.value, api_url: location.pathname.replace(/\/+$/, '') })) {
for await (const chunk of llama(prompt, llamaParams, { controller: controller.value, api_url: URL.parse('.', document.baseURI).href })) {
const data = chunk.data;
if (data.stop) {
examples/server/server.cpp
+1 -1
View File
@@ -1183,7 +1183,7 @@ struct server_context {
bool process_token(completion_token_output & result, server_slot & slot) {
// remember which tokens were sampled - used for repetition penalties during sampling
const std::string token_str = llama_token_to_piece(ctx, result.tok, false);
const std::string token_str = llama_token_to_piece(ctx, result.tok, params.special);
slot.sampled = result.tok;
// search stop word and delete it