import ast
import copy
import functools
import glob
import inspect
import queue
import sys
import os
import time
import traceback
import types
import typing
import warnings
from datetime import datetime
import filelock
import requests
import psutil
from requests import ConnectTimeout, JSONDecodeError
from urllib3.exceptions import ConnectTimeoutError, MaxRetryError, ConnectionError
from requests.exceptions import ConnectionError as ConnectionError2
from requests.exceptions import ReadTimeout as ReadTimeout2
if os.path.dirname(os.path.abspath(__file__)) not in sys.path:
sys.path.append(os.path.dirname(os.path.abspath(__file__)))
os.environ['HF_HUB_DISABLE_TELEMETRY'] = '1'
os.environ['BITSANDBYTES_NOWELCOME'] = '1'
warnings.filterwarnings('ignore', category=UserWarning, message='TypedStorage is deprecated')
from enums import DocumentChoices, LangChainMode, no_lora_str, model_token_mapping, no_model_str, source_prefix, \
source_postfix
from loaders import get_loaders
from utils import set_seed, clear_torch_cache, save_generate_output, NullContext, wrapped_partial, EThread, get_githash, \
import_matplotlib, get_device, makedirs, get_kwargs, start_faulthandler, get_hf_server, FakeTokenizer, remove
start_faulthandler()
import_matplotlib()
SEED = 1236
set_seed(SEED)
from typing import Union
import fire
import torch
from transformers import GenerationConfig, AutoModel, TextIteratorStreamer
from prompter import Prompter, inv_prompt_type_to_model_lower, non_hf_types, PromptType, get_prompt, generate_prompt
from stopping import get_stopping
eval_extra_columns = ['prompt', 'response', 'score']
langchain_modes = [x.value for x in list(LangChainMode)]
scratch_base_dir = '/tmp/'
def main(
load_8bit: bool = False,
load_4bit: bool = False,
load_half: bool = True,
infer_devices: bool = True,
base_model: str = '',
tokenizer_base_model: str = '',
lora_weights: str = "",
gpu_id: int = 0,
compile_model: bool = True,
use_cache: bool = None,
inference_server: str = "",
prompt_type: Union[int, str] = None,
prompt_dict: typing.Dict = None,
model_lock: typing.List[typing.Dict[str, str]] = None,
model_lock_columns: int = None,
fail_if_cannot_connect: bool = False,
# input to generation
temperature: float = None,
top_p: float = None,
top_k: int = None,
num_beams: int = None,
repetition_penalty: float = None,
num_return_sequences: int = None,
do_sample: bool = None,
max_new_tokens: int = None,
min_new_tokens: int = None,
early_stopping: Union[bool, str] = None,
max_time: float = None,
memory_restriction_level: int = None,
debug: bool = False,
save_dir: str = None,
share: bool = True,
local_files_only: bool = False,
resume_download: bool = True,
use_auth_token: Union[str, bool] = False,
trust_remote_code: Union[str, bool] = True,
offload_folder: str = "offline_folder",
src_lang: str = "English",
tgt_lang: str = "Russian",
cli: bool = False,
cli_loop: bool = True,
gradio: bool = True,
gradio_offline_level: int = 0,
chat: bool = True,
chat_context: bool = False,
stream_output: bool = True,
show_examples: bool = None,
verbose: bool = False,
h2ocolors: bool = True,
height: int = 600,
show_lora: bool = True,
login_mode_if_model0: bool = False,
block_gradio_exit: bool = True,
concurrency_count: int = 1,
api_open: bool = False,
allow_api: bool = True,
input_lines: int = 1,
gradio_size: str = None,
auth: typing.List[typing.Tuple[str, str]] = None,
max_max_time=None,
max_max_new_tokens=None,
sanitize_user_prompt: bool = False,
sanitize_bot_response: bool = False,
extra_model_options: typing.List[str] = [],
extra_lora_options: typing.List[str] = [],
extra_server_options: typing.List[str] = [],
score_model: str = 'OpenAssistant/reward-model-deberta-v3-large-v2',
eval_filename: str = None,
eval_prompts_only_num: int = 0,
eval_prompts_only_seed: int = 1234,
eval_as_output: bool = False,
langchain_mode: str = 'Disabled',
force_langchain_evaluate: bool = False,
visible_langchain_modes: list = ['UserData', 'MyData'],
document_choice: list = [DocumentChoices.All_Relevant.name],
user_path: str = None,
detect_user_path_changes_every_query: bool = False,
load_db_if_exists: bool = True,
keep_sources_in_context: bool = False,
db_type: str = 'chroma',
use_openai_embedding: bool = False,
use_openai_model: bool = False,
hf_embedding_model: str = None,
allow_upload_to_user_data: bool = True,
allow_upload_to_my_data: bool = True,
enable_url_upload: bool = True,
enable_text_upload: bool = True,
enable_sources_list: bool = True,
chunk: bool = True,
chunk_size: int = 512,
top_k_docs: int = None,
reverse_docs: bool = True,
auto_reduce_chunks: bool = True,
max_chunks: int = 100,
n_jobs: int = -1,
enable_captions: bool = True,
captions_model: str = "Salesforce/blip-image-captioning-base",
pre_load_caption_model: bool = False,
caption_gpu: bool = True,
enable_ocr: bool = False,
):
"""
:param load_8bit: load model in 8-bit using bitsandbytes
:param load_4bit: load model in 4-bit using bitsandbytes
:param load_half: load model in float16
:param infer_devices: whether to control devices with gpu_id. If False, then spread across GPUs
:param base_model: model HF-type name. If use --base_model to preload model, cannot unload in gradio in models tab
:param tokenizer_base_model: tokenizer HF-type name. Usually not required, inferred from base_model.
:param lora_weights: LORA weights path/HF link
:param gpu_id: if infer_devices, then use gpu_id for cuda device ID, or auto mode if gpu_id != -1
:param compile_model Whether to compile the model
:param use_cache: Whether to use caching in model (some models fail when multiple threads use)
:param inference_server: Consume base_model as type of model at this address
Address can be text-generation-server hosting that base_model
e.g. python generate.py --inference_server="http://192.168.1.46:6112" --base_model=h2oai/h2ogpt-oasst1-512-12b
Or Address can be "openai_chat" or "openai" for OpenAI API
e.g. python generate.py --inference_server="openai_chat" --base_model=gpt-3.5-turbo
e.g. python generate.py --inference_server="openai" --base_model=text-davinci-003
:param prompt_type: type of prompt, usually matched to fine-tuned model or plain for foundational model
:param prompt_dict: If prompt_type=custom, then expects (some) items returned by get_prompt(..., return_dict=True)
:param model_lock: Lock models to specific combinations, for ease of use and extending to many models
Only used if gradio = True
List of dicts, each dict has base_model, tokenizer_base_model, lora_weights, inference_server, prompt_type, and prompt_dict
If all models have same prompt_type, and prompt_dict, can still specify that once in CLI outside model_lock as default for dict
Can specify model_lock instead of those items on CLI
As with CLI itself, base_model can infer prompt_type and prompt_dict if in prompter.py.
Also, tokenizer_base_model and lora_weights are optional.
Also, inference_server is optional if loading model from local system.
All models provided will automatically appear in compare model mode
Model loading-unloading and related choices will be disabled. Model/lora/server adding will be disabled
:param model_lock_columns: How many columns to show if locking models (and so showing all at once)
If None, then defaults to up to 3
if -1, then all goes into 1 row
Maximum value is 4 due to non-dynamic gradio rendering elements
:param fail_if_cannot_connect: if doing model locking (e.g. with many models), fail if True. Otherwise ignore.
Useful when many endpoints and want to just see what works, but still have to wait for timeout.
:param temperature: generation temperature
:param top_p: generation top_p
:param top_k: generation top_k
:param num_beams: generation number of beams
:param repetition_penalty: generation repetition penalty
:param num_return_sequences: generation number of sequences (1 forced for chat)
:param do_sample: generation sample
:param max_new_tokens: generation max new tokens
:param min_new_tokens: generation min tokens
:param early_stopping: generation early stopping
:param max_time: maximum time to allow for generation
:param memory_restriction_level: 0 = no restriction to tokens or model, 1 = some restrictions on token 2 = HF like restriction 3 = very low memory case
:param debug: enable debug mode
:param save_dir: directory chat data is saved to
:param share: whether to share the gradio app with sharable URL
:param local_files_only: whether to only use local files instead of doing to HF for models
:param resume_download: whether to resume downloads from HF for models
:param use_auth_token: whether to use HF auth token (requires CLI did huggingface-cli login before)
:param trust_remote_code: whether to use trust any code needed for HF model
:param offload_folder: path for spilling model onto disk
:param src_lang: source languages to include if doing translation (None = all)
:param tgt_lang: target languages to include if doing translation (None = all)
:param cli: whether to use CLI (non-gradio) interface.
:param cli_loop: whether to loop for CLI (False usually only for testing)
:param gradio: whether to enable gradio, or to enable benchmark mode
:param gradio_offline_level: > 0, then change fonts so full offline
== 1 means backend won't need internet for fonts, but front-end UI might if font not cached
== 2 means backend and frontend don't need internet to download any fonts.
Note: Some things always disabled include HF telemetry, gradio telemetry, chromadb posthog that involve uploading.
This option further disables google fonts for downloading, which is less intrusive than uploading,
but still required in air-gapped case. The fonts don't look as nice as google fonts, but ensure full offline behavior.
Also set --share=False to avoid sharing a gradio live link.
:param chat: whether to enable chat mode with chat history
:param chat_context: whether to use extra helpful context if human_bot
:param stream_output: whether to stream output from generate
:param show_examples: whether to show clickable examples in gradio
:param verbose: whether to show verbose prints
:param h2ocolors: whether to use H2O.ai theme
:param height: height of chat window
:param show_lora: whether to show LORA options in UI (expert so can be hard to understand)
:param login_mode_if_model0: set to True to load --base_model after client logs in, to be able to free GPU memory when model is swapped
:param block_gradio_exit: whether to block gradio exit (used for testing)
:param concurrency_count: gradio concurrency count (1 is optimal for LLMs)
:param api_open: If False, don't let API calls skip gradio queue
:param allow_api: whether to allow API calls at all to gradio server
:param input_lines: how many input lines to show for chat box (>1 forces shift-enter for submit, else enter is submit)
:param gradio_size: Overall size of text and spaces: "xsmall", "small", "medium", "large".
Small useful for many chatbots in model_lock mode
:param auth: gradio auth for launcher in form [(user1, pass1), (user2, pass2), ...]
e.g. --auth=[('jon','password')] with no spaces
:param max_max_time: Maximum max_time for gradio slider
:param max_max_new_tokens: Maximum max_new_tokens for gradio slider
:param sanitize_user_prompt: whether to remove profanity from user input (slows down input processing)
:param sanitize_bot_response: whether to remove profanity and repeat lines from bot output (about 2x slower generation for long streaming cases due to better_profanity being slow)
:param extra_model_options: extra models to show in list in gradio
:param extra_lora_options: extra LORA to show in list in gradio
:param extra_server_options: extra servers to show in list in gradio
:param score_model: which model to score responses (None means no scoring)
:param eval_filename: json file to use for evaluation, if None is sharegpt
:param eval_prompts_only_num: for no gradio benchmark, if using eval_filename prompts for eval instead of examples
:param eval_prompts_only_seed: for no gradio benchmark, seed for eval_filename sampling
:param eval_as_output: for no gradio benchmark, whether to test eval_filename output itself
:param langchain_mode: Data source to include. Choose "UserData" to only consume files from make_db.py.
WARNING: wiki_full requires extra data processing via read_wiki_full.py and requires really good workstation to generate db, unless already present.
:param force_langchain_evaluate: Whether to force langchain LLM use even if not doing langchain, mostly for testing.
:param user_path: user path to glob from to generate db for vector search, for 'UserData' langchain mode.
If already have db, any new/changed files are added automatically if path set, does not have to be same path used for prior db sources
:param detect_user_path_changes_every_query: whether to detect if any files changed or added every similarity search (by file hashes).
Expensive for large number of files, so not done by default. By default only detect changes during db loading.
:param visible_langchain_modes: dbs to generate at launch to be ready for LLM
Can be up to ['wiki', 'wiki_full', 'UserData', 'MyData', 'github h2oGPT', 'DriverlessAI docs']
But wiki_full is expensive and requires preparation
To allow scratch space only live in session, add 'MyData' to list
Default: If only want to consume local files, e.g. prepared by make_db.py, only include ['UserData']
FIXME: Avoid 'All' for now, not implemented
:param document_choice: Default document choice when taking subset of collection
:param load_db_if_exists: Whether to load chroma db if exists or re-generate db
:param keep_sources_in_context: Whether to keep url sources in context, not helpful usually
:param db_type: 'faiss' for in-memory or 'chroma' or 'weaviate' for persisted on disk
:param use_openai_embedding: Whether to use OpenAI embeddings for vector db
:param use_openai_model: Whether to use OpenAI model for use with vector db
:param hf_embedding_model: Which HF embedding model to use for vector db
Default is instructor-large with 768 parameters per embedding if have GPUs, else all-MiniLM-L6-v1 if no GPUs
Can also choose simpler model with 384 parameters per embedding: "sentence-transformers/all-MiniLM-L6-v2"
Can also choose even better embedding with 1024 parameters: 'hkunlp/instructor-xl'
We support automatically changing of embeddings for chroma, with a backup of db made if this is done
:param allow_upload_to_user_data: Whether to allow file uploads to update shared vector db
:param allow_upload_to_my_data: Whether to allow file uploads to update scratch vector db
:param enable_url_upload: Whether to allow upload from URL
:param enable_text_upload: Whether to allow upload of text
:param enable_sources_list: Whether to allow list (or download for non-shared db) of list of sources for chosen db
:param chunk: Whether to chunk data (True unless know data is already optimally chunked)
:param chunk_size: Size of chunks, with typically top-4 passed to LLM, so neesd to be in context length
:param top_k_docs: number of chunks to give LLM
:param reverse_docs: whether to reverse docs order so most relevant is closest to question.
Best choice for sufficiently smart model, and truncation occurs for oldest context, so best then too.
But smaller 6_9 models fail to use newest context and can get stuck on old information.
:param auto_reduce_chunks: Whether to automatically reduce top_k_docs to fit context given prompt
:param max_chunks: If top_k_docs=-1, maximum number of chunks to allow
:param n_jobs: Number of processors to use when consuming documents (-1 = all, is default)
:param enable_captions: Whether to support captions using BLIP for image files as documents, then preloads that model
:param captions_model: Which model to use for captions.
captions_model: str = "Salesforce/blip-image-captioning-base", # continue capable
captions_model: str = "Salesforce/blip2-flan-t5-xl", # question/answer capable, 16GB state
captions_model: str = "Salesforce/blip2-flan-t5-xxl", # question/answer capable, 60GB state
Note: opt-based blip2 are not permissive license due to opt and Meta license restrictions
:param pre_load_caption_model: Whether to preload caption model, or load after forking parallel doc loader
parallel loading disabled if preload and have images, to prevent deadlocking on cuda context
Recommended if using larger caption model
:param caption_gpu: If support caption, then use GPU if exists
:param enable_ocr: Whether to support OCR on images
:return:
"""
if base_model is None:
base_model = ''
if tokenizer_base_model is None:
tokenizer_base_model = ''
if lora_weights is None:
lora_weights = ''
if inference_server is None:
inference_server = ''
# listen to env if set
model_lock = os.getenv('model_lock', str(model_lock))
model_lock = ast.literal_eval(model_lock)
if model_lock:
assert gradio, "model_lock only supported for gradio=True"
if len(model_lock) > 1:
assert chat, "model_lock only works for multiple models for chat=True"
assert not cli, "model_lock only supported for cli=False"
assert not (not cli and not gradio), "model_lock only supported for eval (cli=gradio=False)"
assert not base_model, "Don't specify model_lock and base_model"
assert not tokenizer_base_model, "Don't specify model_lock and tokenizer_base_model"
assert not lora_weights, "Don't specify model_lock and lora_weights"
assert not inference_server, "Don't specify model_lock and inference_server"
# assert not prompt_type, "Don't specify model_lock and prompt_type"
# assert not prompt_dict, "Don't specify model_lock and prompt_dict"
is_hf = bool(int(os.getenv("HUGGINGFACE_SPACES", '0')))
is_gpth2oai = bool(int(os.getenv("GPT_H2O_AI", '0')))
is_public = is_hf or is_gpth2oai # multi-user case with fixed model and disclaimer
if memory_restriction_level is None:
memory_restriction_level = 2 if is_hf else 0 # 2 assumes run on 24GB consumer GPU
else:
assert 0 <= memory_restriction_level <= 3, "Bad memory_restriction_level=%s" % memory_restriction_level
admin_pass = os.getenv("ADMIN_PASS")
# will sometimes appear in UI or sometimes actual generation, but maybe better than empty result
# but becomes unrecoverable sometimes if raise, so just be silent for now
raise_generate_gpu_exceptions = True
# allow set token directly
use_auth_token = os.environ.get("HUGGINGFACE_API_TOKEN", use_auth_token)
allow_upload_to_user_data = bool(
int(os.environ.get("allow_upload_to_user_data", str(int(allow_upload_to_user_data)))))
allow_upload_to_my_data = bool(int(os.environ.get("allow_upload_to_my_data", str(int(allow_upload_to_my_data)))))
height = int(os.environ.get("HEIGHT", height))
h2ocolors = bool(int(os.getenv('h2ocolors', h2ocolors)))
# allow enabling langchain via ENV
# FIRST PLACE where LangChain referenced, but no imports related to it
langchain_mode = os.environ.get("LANGCHAIN_MODE", langchain_mode)
assert langchain_mode in langchain_modes, "Invalid langchain_mode %s" % langchain_mode
visible_langchain_modes = ast.literal_eval(os.environ.get("visible_langchain_modes", str(visible_langchain_modes)))
if langchain_mode not in visible_langchain_modes and langchain_mode in langchain_modes:
visible_langchain_modes += [langchain_mode]
# if specifically chose not to show My or User Data, disable upload, so gradio elements are simpler
if LangChainMode.MY_DATA.value not in visible_langchain_modes:
allow_upload_to_my_data = False
if LangChainMode.USER_DATA.value not in visible_langchain_modes:
allow_upload_to_user_data = False
if is_public:
allow_upload_to_user_data = False
input_lines = 1 # ensure set, for ease of use
temperature = 0.2 if temperature is None else temperature
top_p = 0.85 if top_p is None else top_p
top_k = 70 if top_k is None else top_k
if is_hf:
do_sample = True if do_sample is None else do_sample
top_k_docs = 3 if top_k_docs is None else top_k_docs
else:
# by default don't sample, too chatty
do_sample = False if do_sample is None else do_sample
top_k_docs = 4 if top_k_docs is None else top_k_docs
if memory_restriction_level == 2:
if not base_model and not inference_server:
base_model = 'h2oai/h2ogpt-oasst1-512-12b'
# don't set load_8bit if passed base_model, doesn't always work so can't just override
load_8bit = True
load_4bit = False # FIXME - consider using 4-bit instead of 8-bit
elif not inference_server:
top_k_docs = 10 if top_k_docs is None else top_k_docs
if memory_restriction_level >= 2:
load_8bit = True
load_4bit = False # FIXME - consider using 4-bit instead of 8-bit
if hf_embedding_model is None:
hf_embedding_model = "sentence-transformers/all-MiniLM-L6-v2"
top_k_docs = 3 if top_k_docs is None else top_k_docs
if top_k_docs is None:
top_k_docs = 3
if is_public:
if not max_time:
max_time = 60 * 2
if not max_max_time:
max_max_time = max_time
if not max_new_tokens:
max_new_tokens = 256
if not max_max_new_tokens:
max_max_new_tokens = 256
else:
if not max_max_time:
max_max_time = 60 * 20
if not max_max_new_tokens:
max_max_new_tokens = 512
if is_hf:
# must override share if in spaces
share = False
if not max_time:
max_time = 60 * 1
if not max_max_time:
max_max_time = max_time
# HF accounted for later in get_max_max_new_tokens()
save_dir = os.getenv('SAVE_DIR', save_dir)
score_model = os.getenv('SCORE_MODEL', score_model)
if score_model == 'None' or score_model is None:
score_model = ''
concurrency_count = int(os.getenv('CONCURRENCY_COUNT', concurrency_count))
api_open = bool(int(os.getenv('API_OPEN', str(int(api_open)))))
allow_api = bool(int(os.getenv('ALLOW_API', str(int(allow_api)))))
n_gpus = torch.cuda.device_count() if torch.cuda.is_available else 0
if n_gpus == 0:
gpu_id = None
load_8bit = False
load_4bit = False
load_half = False
infer_devices = False
torch.backends.cudnn.benchmark = True
torch.backends.cudnn.enabled = False
torch.set_default_dtype(torch.float32)
if psutil.virtual_memory().available < 94 * 1024 ** 3 and not inference_server:
# 12B uses ~94GB
# 6.9B uses ~47GB
base_model = 'h2oai/h2ogpt-oig-oasst1-512-6_9b' if not base_model else base_model
if hf_embedding_model is None:
# if no GPUs, use simpler embedding model to avoid cost in time
hf_embedding_model = "sentence-transformers/all-MiniLM-L6-v2"
else:
if hf_embedding_model is None:
# if still None, then set default
hf_embedding_model = 'hkunlp/instructor-large'
# get defaults
model_lower = base_model.lower()
if not gradio:
# force, else not single response like want to look at
stream_output = False
# else prompt removal can mess up output
chat = False
# hard-coded defaults
first_para = False
text_limit = None
if offload_folder:
makedirs(offload_folder)
if user_path:
makedirs(user_path)
placeholder_instruction, placeholder_input, \
stream_output, show_examples, \
prompt_type, prompt_dict, \
temperature, top_p, top_k, num_beams, \
max_new_tokens, min_new_tokens, early_stopping, max_time, \
repetition_penalty, num_return_sequences, \
do_sample, \
src_lang, tgt_lang, \
examples, \
task_info = \
get_generate_params(model_lower, chat,
stream_output, show_examples,
prompt_type, prompt_dict,
temperature, top_p, top_k, num_beams,
max_new_tokens, min_new_tokens, early_stopping, max_time,
repetition_penalty, num_return_sequences,
do_sample,
top_k_docs,
chunk,
chunk_size,
verbose,
)
git_hash = get_githash()
locals_dict = locals()
locals_print = '\n'.join(['%s: %s' % (k, v) for k, v in locals_dict.items()])
if verbose:
print(f"Generating model with params:\n{locals_print}", flush=True)
print("Command: %s\nHash: %s" % (str(' '.join(sys.argv)), git_hash), flush=True)
if langchain_mode != "Disabled":
# SECOND PLACE where LangChain referenced, but all imports are kept local so not required
from gpt_langchain import prep_langchain, get_some_dbs_from_hf
if is_hf:
get_some_dbs_from_hf()
dbs = {}
for langchain_mode1 in visible_langchain_modes:
if langchain_mode1 in ['MyData']:
# don't use what is on disk, remove it instead
for gpath1 in glob.glob(os.path.join(scratch_base_dir, 'db_dir_%s*' % langchain_mode1)):
if os.path.isdir(gpath1):
print("Removing old MyData: %s" % gpath1, flush=True)
remove(gpath1)
continue
if langchain_mode1 in ['All']:
# FIXME: All should be avoided until scans over each db, shouldn't be separate db
continue
persist_directory1 = 'db_dir_%s' % langchain_mode1 # single place, no special names for each case
try:
db = prep_langchain(persist_directory1,
load_db_if_exists,
db_type, use_openai_embedding,
langchain_mode1, user_path,
hf_embedding_model,
kwargs_make_db=locals())
finally:
# in case updated embeddings or created new embeddings
clear_torch_cache()
dbs[langchain_mode1] = db
# remove None db's so can just rely upon k in dbs for if hav db
dbs = {k: v for k, v in dbs.items() if v is not None}
else:
dbs = {}
# import control
if os.environ.get("TEST_LANGCHAIN_IMPORT"):
assert 'gpt_langchain' not in sys.modules, "Dev bug, import of langchain when should not have"
assert 'langchain' not in sys.modules, "Dev bug, import of langchain when should not have"
model_state_none = dict(model=None, tokenizer=None, device=None,
base_model=None, tokenizer_base_model=None, lora_weights=None,
inference_server=None, prompt_type=None, prompt_dict=None)
if cli:
from cli import run_cli
return run_cli(**get_kwargs(run_cli, exclude_names=['model_state0'], **locals()))
elif not gradio:
from eval import run_eval
return run_eval(**get_kwargs(run_eval, exclude_names=['model_state0'], **locals()))
elif gradio:
# imported here so don't require gradio to run generate
from gradio_runner import go_gradio
# get default model
model_states = []
model_list = [dict(base_model=base_model, tokenizer_base_model=tokenizer_base_model, lora_weights=lora_weights,
inference_server=inference_server, prompt_type=prompt_type, prompt_dict=prompt_dict)]
model_list0 = copy.deepcopy(model_list) # just strings, safe to deepcopy
model_state0 = model_state_none.copy()
assert len(model_state_none) == len(model_state0)
if model_lock:
model_list = model_lock
for model_dict in reversed(model_list):
# do reverse, so first is default base_model etc., so some logic works in go_gradio() more easily
# handles defaults user didn't have to pass
model_dict['base_model'] = base_model = model_dict.get('base_model', '')
model_dict['tokenizer_base_model'] = tokenizer_base_model = model_dict.get('tokenizer_base_model', '')
model_dict['lora_weights'] = lora_weights = model_dict.get('lora_weights', '')
model_dict['inference_server'] = inference_server = model_dict.get('inference_server', '')
prompt_type = model_dict.get('prompt_type', model_list0[0]['prompt_type']) # don't use mutated value
# try to infer, ignore empty initial state leading to get_generate_params -> 'plain'
if model_dict.get('prompt_type') is None:
model_lower = base_model.lower()
if model_lower in inv_prompt_type_to_model_lower:
prompt_type = inv_prompt_type_to_model_lower[model_lower]
prompt_dict, error0 = get_prompt(prompt_type, '',
chat=False, context='', reduced=False, making_context=False,
return_dict=True)
model_dict['prompt_type'] = prompt_type
model_dict['prompt_dict'] = prompt_dict = model_dict.get('prompt_dict', prompt_dict)
all_kwargs = locals().copy()
if base_model and not login_mode_if_model0:
model0, tokenizer0, device = get_model(reward_type=False,
**get_kwargs(get_model, exclude_names=['reward_type'],
**all_kwargs))
else:
# if empty model, then don't load anything, just get gradio up
model0, tokenizer0, device = None, None, None
if model0 is None:
if fail_if_cannot_connect:
raise RuntimeError("Could not connect, see logs")
# skip
if isinstance(model_lock, list):
model_lock.remove(model_dict)
continue
model_state_trial = dict(model=model0, tokenizer=tokenizer0, device=device)
model_state_trial.update(model_dict)
assert len(model_state_none) == len(model_state_trial)
print("Model %s" % model_dict, flush=True)
if model_lock:
# last in iteration will be first
model_states.insert(0, model_state_trial)
# fill model_state0 so go_gradio() easier, manage model_states separately
model_state0 = model_state_trial.copy()
else:
model_state0 = model_state_trial.copy()
assert len(model_state_none) == len(model_state0)
# get score model
all_kwargs = locals().copy()
smodel, stokenizer, sdevice = get_score_model(reward_type=True,
**get_kwargs(get_score_model, exclude_names=['reward_type'],
**all_kwargs))
score_model_state0 = dict(model=smodel, tokenizer=stokenizer, device=sdevice,
base_model=score_model, tokenizer_base_model='', lora_weights='',
inference_server='', prompt_type='', prompt_dict='')
if enable_captions:
if pre_load_caption_model:
from image_captions import H2OImageCaptionLoader
caption_loader = H2OImageCaptionLoader(caption_gpu=caption_gpu).load_model()
else:
caption_loader = 'gpu' if caption_gpu else 'cpu'
else:
caption_loader = False
# assume gradio needs everything
go_gradio(**locals())
def get_config(base_model,
use_auth_token=False,
trust_remote_code=True,
offload_folder=None,
triton_attn=False,
long_sequence=True,
return_model=False,
raise_exception=False,
):
from accelerate import init_empty_weights
with init_empty_weights():
from transformers import AutoConfig
try:
config = AutoConfig.from_pretrained(base_model, use_auth_token=use_auth_token,
trust_remote_code=trust_remote_code,
offload_folder=offload_folder)
except OSError as e:
if raise_exception:
raise
if 'not a local folder and is not a valid model identifier listed on' in str(
e) or '404 Client Error' in str(e):
# e.g. llama, gpjt, etc.
# e.g. HF TGI but not model on HF or private etc.
# HF TGI server only should really require prompt_type, not HF model state
return None, None
else:
raise
if triton_attn and 'mpt-' in base_model.lower():
config.attn_config['attn_impl'] = 'triton'
if long_sequence:
if 'mpt-7b-storywriter' in base_model.lower():
config.update({"max_seq_len": 83968})
if 'mosaicml/mpt-7b-chat' in base_model.lower():
config.update({"max_seq_len": 4096})
if 'mpt-30b' in base_model.lower():
config.update({"max_seq_len": 2 * 8192})
if return_model and \
issubclass(config.__class__, tuple(AutoModel._model_mapping.keys())):
model = AutoModel.from_config(
config,
trust_remote_code=trust_remote_code,
)
else:
# can't infer
model = None
if 'falcon' in base_model.lower():
config.use_cache = False
return config, model
def get_non_lora_model(base_model, model_loader, load_half, model_kwargs, reward_type,
config, model,
gpu_id=0,
):
"""
Ensure model gets on correct device
"""
if model is not None:
# NOTE: Can specify max_memory={0: max_mem, 1: max_mem}, to shard model
# NOTE: Some models require avoiding sharding some layers,
# then would pass no_split_module_classes and give list of those layers.
from accelerate import infer_auto_device_map
device_map = infer_auto_device_map(
model,
dtype=torch.float16 if load_half else torch.float32,
)
if hasattr(model, 'model'):
device_map_model = infer_auto_device_map(
model.model,
dtype=torch.float16 if load_half else torch.float32,
)
device_map.update(device_map_model)
else:
device_map = "auto"
n_gpus = torch.cuda.device_count() if torch.cuda.is_available else 0
if n_gpus > 0:
if gpu_id >= 0:
# FIXME: If really distributes model, tend to get things like: ValueError: gpt_neox.embed_in.weight doesn't have any device set.
# So avoid for now, just put on first GPU, unless score_model, put on last
if reward_type:
device_map = {'': n_gpus - 1}
else:
device_map = {'': min(n_gpus - 1, gpu_id)}
if gpu_id == -1:
device_map = {'': 'cuda'}
else:
device_map = {'': 'cpu'}
model_kwargs['load_in_8bit'] = False
model_kwargs['load_in_4bit'] = False
print('device_map: %s' % device_map, flush=True)
load_in_8bit = model_kwargs.get('load_in_8bit', False)
load_in_4bit = model_kwargs.get('load_in_4bit', False)
model_kwargs['device_map'] = device_map
pop_unused_model_kwargs(model_kwargs)
if load_in_8bit or load_in_4bit or not load_half:
model = model_loader.from_pretrained(
base_model,
config=config,
**model_kwargs,
)
else:
model = model_loader.from_pretrained(
base_model,
config=config,
**model_kwargs,
).half()
return model
def get_client_from_inference_server(inference_server, raise_connection_exception=False):
inference_server, headers = get_hf_server(inference_server)
# preload client since slow for gradio case especially
from gradio_utils.grclient import GradioClient
gr_client = None
hf_client = None
if headers is None:
try:
print("GR Client Begin: %s" % inference_server, flush=True)
# first do sanity check if alive, else gradio client takes too long by default
requests.get(inference_server, timeout=int(os.getenv('REQUEST_TIMEOUT', '30')))
gr_client = GradioClient(inference_server)
print("GR Client End: %s" % inference_server, flush=True)
except (OSError, ValueError) as e:
# Occurs when wrong endpoint and should have been HF client, so don't hard raise, just move to HF
gr_client = None
print("GR Client Failed %s: %s" % (inference_server, str(e)), flush=True)
except (ConnectTimeoutError, ConnectTimeout, MaxRetryError, ConnectionError, ConnectionError2,
JSONDecodeError, ReadTimeout2, KeyError) as e:
t, v, tb = sys.exc_info()
ex = ''.join(traceback.format_exception(t, v, tb))
print("GR Client Failed %s: %s" % (inference_server, str(ex)), flush=True)
if raise_connection_exception:
raise
if gr_client is None:
res = None
from text_generation import Client as HFClient
print("HF Client Begin: %s" % inference_server)
try:
hf_client = HFClient(inference_server, headers=headers, timeout=int(os.getenv('REQUEST_TIMEOUT', '30')))
# quick check valid TGI endpoint
res = hf_client.generate('What?', max_new_tokens=1)
hf_client = HFClient(inference_server, headers=headers, timeout=300)
except (ConnectTimeoutError, ConnectTimeout, MaxRetryError, ConnectionError, ConnectionError2,
JSONDecodeError, ReadTimeout2, KeyError) as e:
hf_client = None
t, v, tb = sys.exc_info()
ex = ''.join(traceback.format_exception(t, v, tb))
print("HF Client Failed %s: %s" % (inference_server, str(ex)))
if raise_connection_exception:
raise
print("HF Client End: %s %s" % (inference_server, res))
return inference_server, gr_client, hf_client
def get_model(
load_8bit: bool = False,
load_4bit: bool = False,
load_half: bool = True,
infer_devices: bool = True,
base_model: str = '',
inference_server: str = "",
tokenizer_base_model: str = '',
lora_weights: str = "",
gpu_id: int = 0,
reward_type: bool = None,
local_files_only: bool = False,
resume_download: bool = True,
use_auth_token: Union[str, bool] = False,
trust_remote_code: bool = True,
offload_folder: str = None,
compile_model: bool = True,
verbose: bool = False,
):
"""
:param load_8bit: load model in 8-bit, not supported by all models
:param load_4bit: load model in 4-bit, not supported by all models
:param load_half: load model in 16-bit
:param infer_devices: Use torch infer of optimal placement of layers on devices (for non-lora case)
For non-LORA case, False will spread shards across multiple GPUs, but this can lead to cuda:x cuda:y mismatches
So it is not the default
:param base_model: name/path of base model
:param inference_server: whether base_model is hosted locally ('') or via http (url)
:param tokenizer_base_model: name/path of tokenizer
:param lora_weights: name/path
:param gpu_id: which GPU (0..n_gpus-1) or allow all GPUs if relevant (-1)
:param reward_type: reward type model for sequence classification
:param local_files_only: use local files instead of from HF
:param resume_download: resume downloads from HF
:param use_auth_token: assumes user did on CLI `huggingface-cli login` to access private repo
:param trust_remote_code: trust code needed by model
:param offload_folder: offload folder
:param compile_model: whether to compile torch model
:param verbose:
:return:
"""
if verbose:
print("Get %s model" % base_model, flush=True)
triton_attn = False
long_sequence = True
config_kwargs = dict(use_auth_token=use_auth_token,
trust_remote_code=trust_remote_code,
offload_folder=offload_folder,
triton_attn=triton_attn,
long_sequence=long_sequence)
config, _ = get_config(base_model, **config_kwargs, raise_exception=False)
if base_model in non_hf_types:
assert config is None, "Expected config None for %s" % base_model
llama_type_from_config = 'llama' in str(config).lower()
llama_type_from_name = "llama" in base_model.lower()
llama_type = llama_type_from_config or llama_type_from_name
if "xgen" in base_model.lower():
llama_type = False
if llama_type:
if verbose:
print("Detected as llama type from"
" config (%s) or name (%s)" % (llama_type_from_config, llama_type_from_name), flush=True)
model_loader, tokenizer_loader = get_loaders(model_name=base_model, reward_type=reward_type, llama_type=llama_type)
tokenizer_kwargs = dict(local_files_only=local_files_only,
resume_download=resume_download,
use_auth_token=use_auth_token,
trust_remote_code=trust_remote_code,
offload_folder=offload_folder,
padding_side='left',
config=config,
)
if not tokenizer_base_model:
tokenizer_base_model = base_model
if config is not None and tokenizer_loader is not None and not isinstance(tokenizer_loader, str):
tokenizer = tokenizer_loader.from_pretrained(tokenizer_base_model, **tokenizer_kwargs)
# sets raw (no cushion) limit
set_model_max_len(config, tokenizer, verbose=False)
# if using fake tokenizer, not really accurate when lots of numbers, give a bit of buffer, else get:
# Generation Failed: Input validation error: `inputs` must have less than 2048 tokens. Given: 2233
tokenizer.model_max_length = tokenizer.model_max_length - 50
else:
tokenizer = FakeTokenizer()
if isinstance(inference_server, str) and inference_server.startswith("http"):
inference_server, gr_client, hf_client = get_client_from_inference_server(inference_server)
client = gr_client or hf_client
# Don't return None, None for model, tokenizer so triggers
return client, tokenizer, 'http'
if isinstance(inference_server, str) and inference_server.startswith('openai'):
assert os.getenv('OPENAI_API_KEY'), "Set environment for OPENAI_API_KEY"
# Don't return None, None for model, tokenizer so triggers
# include small token cushion
tokenizer = FakeTokenizer(model_max_length=model_token_mapping[base_model] - 50)
return inference_server, tokenizer, inference_server
assert not inference_server, "Malformed inference_server=%s" % inference_server
if base_model in non_hf_types:
from gpt4all_llm import get_model_tokenizer_gpt4all
model, tokenizer, device = get_model_tokenizer_gpt4all(base_model)
return model, tokenizer, device
# get local torch-HF model
return get_hf_model(load_8bit=load_8bit,
load_4bit=load_4bit,
load_half=load_half,
infer_devices=infer_devices,
base_model=base_model,
tokenizer_base_model=tokenizer_base_model,
lora_weights=lora_weights,
gpu_id=gpu_id,
reward_type=reward_type,
local_files_only=local_files_only,
resume_download=resume_download,
use_auth_token=use_auth_token,
trust_remote_code=trust_remote_code,
offload_folder=offload_folder,
compile_model=compile_model,
llama_type=llama_type,
config_kwargs=config_kwargs,
tokenizer_kwargs=tokenizer_kwargs,
verbose=verbose)
def get_hf_model(load_8bit: bool = False,
load_4bit: bool = False,
load_half: bool = True,
infer_devices: bool = True,
base_model: str = '',
tokenizer_base_model: str = '',
lora_weights: str = "",
gpu_id: int = 0,
reward_type: bool = None,
local_files_only: bool = False,
resume_download: bool = True,
use_auth_token: Union[str, bool] = False,
trust_remote_code: bool = True,
offload_folder: str = None,
compile_model: bool = True,
llama_type: bool = False,
config_kwargs=None,
tokenizer_kwargs=None,
verbose: bool = False,
):
assert config_kwargs is not None
assert tokenizer_kwargs is not None
if lora_weights is not None and lora_weights.strip():
if verbose:
print("Get %s lora weights" % lora_weights, flush=True)
device = get_device()
if 'gpt2' in base_model.lower():
# RuntimeError: where expected condition to be a boolean tensor, but got a tensor with dtype Half
load_8bit = False
load_4bit = False
assert base_model.strip(), (
"Please choose a base model with --base_model (CLI) or load one from Models Tab (gradio)"
)
model_loader, tokenizer_loader = get_loaders(model_name=base_model, reward_type=reward_type, llama_type=llama_type)
config, _ = get_config(base_model, return_model=False, raise_exception=True, **config_kwargs)
if tokenizer_loader is not None and not isinstance(tokenizer_loader, str):
tokenizer = tokenizer_loader.from_pretrained(tokenizer_base_model,
**tokenizer_kwargs)
else:
tokenizer = tokenizer_loader
if isinstance(tokenizer, str):
# already a pipeline, tokenizer_loader is string for task
model = model_loader(tokenizer,
model=base_model,
device=0 if device == "cuda" else -1,
torch_dtype=torch.float16 if device == 'cuda' else torch.float32)
else:
assert device in ["cuda", "cpu"], "Unsupported device %s" % device
model_kwargs = dict(local_files_only=local_files_only,
torch_dtype=torch.float16 if device == 'cuda' else torch.float32,
resume_download=resume_download,
use_auth_token=use_auth_token,
trust_remote_code=trust_remote_code,
offload_folder=offload_folder,
)
if 'mbart-' not in base_model.lower() and 'mpt-' not in base_model.lower():
model_kwargs.update(dict(load_in_8bit=load_8bit,
load_in_4bit=load_4bit,
device_map={"": 0} if (load_8bit or load_4bit) and device == 'cuda' else "auto",
))
if 'mpt-' in base_model.lower() and gpu_id is not None and gpu_id >= 0:
model_kwargs.update(dict(device_map={"": gpu_id} if device == 'cuda' else "cpu"))
if 'OpenAssistant/reward-model'.lower() in base_model.lower():
# FIXME: could put on other GPUs
model_kwargs['device_map'] = {"": 0} if device == 'cuda' else {"": 'cpu'}
model_kwargs.pop('torch_dtype', None)
pop_unused_model_kwargs(model_kwargs)
if not lora_weights:
with torch.device(device):
if infer_devices:
config, model = get_config(base_model, return_model=True, raise_exception=True, **config_kwargs)
model = get_non_lora_model(base_model, model_loader, load_half, model_kwargs, reward_type,
config, model,
gpu_id=gpu_id,
)
else:
config, _ = get_config(base_model, **config_kwargs)
if load_half and not (load_8bit or load_4bit):
model = model_loader.from_pretrained(
base_model,
config=config,
**model_kwargs).half()
else:
model = model_loader.from_pretrained(
base_model,
config=config,
**model_kwargs)
elif load_8bit or load_4bit:
config, _ = get_config(base_model, **config_kwargs)
model = model_loader.from_pretrained(
base_model,
config=config,
**model_kwargs
)
from peft import PeftModel # loads cuda, so avoid in global scope
model = PeftModel.from_pretrained(
model,
lora_weights,
torch_dtype=torch.float16 if device == 'cuda' else torch.float32,
local_files_only=local_files_only,
resume_download=resume_download,
use_auth_token=use_auth_token,
trust_remote_code=trust_remote_code,
offload_folder=offload_folder,
device_map={"": 0} if device == 'cuda' else {"": 'cpu'}, # seems to be required
)
else:
with torch.device(device):
config, _ = get_config(base_model, raise_exception=True, **config_kwargs)
model = model_loader.from_pretrained(
base_model,
config=config,
**model_kwargs
)
from peft import PeftModel # loads cuda, so avoid in global scope
model = PeftModel.from_pretrained(
model,
lora_weights,
torch_dtype=torch.float16 if device == 'cuda' else torch.float32,
local_files_only=local_files_only,
resume_download=resume_download,
use_auth_token=use_auth_token,
trust_remote_code=trust_remote_code,
offload_folder=offload_folder,
device_map="auto",
)
if load_half:
model.half()
# unwind broken decapoda-research config
if llama_type:
model.config.pad_token_id = tokenizer.pad_token_id = 0 # unk
model.config.bos_token_id = 1
model.config.eos_token_id = 2
if 'gpt2' in base_model.lower():
# add special tokens that otherwise all share the same id
tokenizer.add_special_tokens({'bos_token': '<bos>',
'eos_token': '<eos>',
'pad_token': '<pad>'})
if not isinstance(tokenizer, str):
model.eval()
if torch.__version__ >= "2" and sys.platform != "win32" and compile_model:
model = torch.compile(model)
set_model_max_len(config, tokenizer, verbose=False, reward_type=reward_type)
return model, tokenizer, device
def set_model_max_len(config, tokenizer, verbose=False, reward_type=False):
if reward_type:
# limit deberta, else uses too much memory and not worth response score
tokenizer.model_max_length = 512
if hasattr(config, 'max_seq_len') and isinstance(config.max_seq_len, int):
tokenizer.model_max_length = config.max_seq_len
elif hasattr(config, 'max_position_embeddings') and isinstance(config.max_position_embeddings, int):
# help automatically limit inputs to generate
tokenizer.model_max_length = config.max_position_embeddings
else:
if verbose:
print("Could not determine model_max_length, setting to 2048", flush=True)
tokenizer.model_max_length = 2048
# for bug in HF transformers
if tokenizer.model_max_length > 100000000:
tokenizer.model_max_length = 2048
def pop_unused_model_kwargs(model_kwargs):
"""
in-place pop unused kwargs that are not dependency-upgrade friendly
no point passing in False, is default, and helps avoid needing to update requirements for new deps
:param model_kwargs:
:return:
"""
check_list = ['load_in_8bit', 'load_in_4bit']
for k in check_list:
if k in model_kwargs and not model_kwargs[k]:
model_kwargs.pop(k)
def get_score_model(score_model: str = None,
load_8bit: bool = False,
load_4bit: bool = False,
load_half: bool = True,
infer_devices: bool = True,
base_model: str = '',
inference_server: str = '',
tokenizer_base_model: str = '',
lora_weights: str = "",
gpu_id: int = 0,
reward_type: bool = None,
local_files_only: bool = False,
resume_download: bool = True,
use_auth_token: Union[str, bool] = False,
trust_remote_code: bool = True,
offload_folder: str = None,
compile_model: bool = True,
verbose: bool = False,
):
if score_model is not None and score_model.strip():
load_8bit = False
load_4bit = False
load_half = False
base_model = score_model.strip()
tokenizer_base_model = ''
lora_weights = ''
inference_server = ''
llama_type = False
compile_model = False
smodel, stokenizer, sdevice = get_model(reward_type=True,
**get_kwargs(get_model, exclude_names=['reward_type'], **locals()))
else:
smodel, stokenizer, sdevice = None, None, None
return smodel, stokenizer, sdevice
no_default_param_names = [
'instruction',
'iinput',
'context',
'instruction_nochat',
'iinput_nochat',
]
gen_hyper = ['temperature',
'top_p',
'top_k',
'num_beams',
'max_new_tokens',
'min_new_tokens',
'early_stopping',
'max_time',
'repetition_penalty',
'num_return_sequences',
'do_sample',
]
eval_func_param_names = ['instruction',
'iinput',
'context',
'stream_output',
'prompt_type',
'prompt_dict'] + \
gen_hyper + \
['chat',
'instruction_nochat',
'iinput_nochat',
'langchain_mode',
'top_k_docs',
'chunk',
'chunk_size',
'document_choice',
]
# form evaluate defaults for submit_nochat_api
eval_func_param_names_defaults = eval_func_param_names.copy()
for k in no_default_param_names:
if k in eval_func_param_names_defaults:
eval_func_param_names_defaults.remove(k)
def evaluate_from_str(
model_state,
my_db_state,
# START NOTE: Examples must have same order of parameters
user_kwargs,
# END NOTE: Examples must have same order of parameters
default_kwargs=None,
src_lang=None,
tgt_lang=None,
debug=False,
concurrency_count=None,
save_dir=None,
sanitize_bot_response=False,
model_state0=None,
memory_restriction_level=None,
max_max_new_tokens=None,
is_public=None,
max_max_time=None,
raise_generate_gpu_exceptions=None,
chat_context=None,
lora_weights=None,
load_db_if_exists=True,
dbs=None,
user_path=None,
detect_user_path_changes_every_query=None,
use_openai_embedding=None,
use_openai_model=None,
hf_embedding_model=None,
db_type=None,
n_jobs=None,
first_para=None,
text_limit=None,
verbose=False,
cli=False,
reverse_docs=True,
use_cache=None,
auto_reduce_chunks=None,
max_chunks=None,
model_lock=None,
force_langchain_evaluate=None,
model_state_none=None,
):
if isinstance(user_kwargs, str):
user_kwargs = ast.literal_eval(user_kwargs)
# only used for submit_nochat_api
user_kwargs['chat'] = False
if 'stream_output' not in user_kwargs:
user_kwargs['stream_output'] = False
if 'langchain_mode' not in user_kwargs:
# if user doesn't specify, then assume disabled, not use default
user_kwargs['langchain_mode'] = 'Disabled'
assert set(list(default_kwargs.keys())) == set(eval_func_param_names)
# correct ordering. Note some things may not be in default_kwargs, so can't be default of user_kwargs.get()
args_list = [user_kwargs[k] if k in user_kwargs else default_kwargs[k] for k in eval_func_param_names]
ret = evaluate(
model_state,
my_db_state,
# START NOTE: Examples must have same order of parameters
*tuple(args_list),
# END NOTE: Examples must have same order of parameters
src_lang=src_lang,
tgt_lang=tgt_lang,
debug=debug,
concurrency_count=concurrency_count,
save_dir=save_dir,
sanitize_bot_response=sanitize_bot_response,
model_state0=model_state0,
memory_restriction_level=memory_restriction_level,
max_max_new_tokens=max_max_new_tokens,
is_public=is_public,
max_max_time=max_max_time,
raise_generate_gpu_exceptions=raise_generate_gpu_exceptions,
chat_context=chat_context,
lora_weights=lora_weights,
load_db_if_exists=load_db_if_exists,
dbs=dbs,
user_path=user_path,
detect_user_path_changes_every_query=detect_user_path_changes_every_query,
use_openai_embedding=use_openai_embedding,
use_openai_model=use_openai_model,
hf_embedding_model=hf_embedding_model,
db_type=db_type,
n_jobs=n_jobs,
first_para=first_para,
text_limit=text_limit,
verbose=verbose,
cli=cli,
reverse_docs=reverse_docs,
use_cache=use_cache,
auto_reduce_chunks=auto_reduce_chunks,
max_chunks=max_chunks,
model_lock=model_lock,
force_langchain_evaluate=force_langchain_evaluate,
model_state_none=model_state_none,
)
try:
for ret1 in ret:
yield ret1
finally:
# clear before return, in finally in case GPU OOM exception
clear_torch_cache()
def evaluate(
model_state,
my_db_state,
# START NOTE: Examples must have same order of parameters
instruction,
iinput,
context,
stream_output,
prompt_type,
prompt_dict,
temperature,
top_p,
top_k,
num_beams,
max_new_tokens,
min_new_tokens,
early_stopping,
max_time,
repetition_penalty,
num_return_sequences,
do_sample,
chat,
instruction_nochat,
iinput_nochat,
langchain_mode,
top_k_docs,
chunk,
chunk_size,
document_choice,
# END NOTE: Examples must have same order of parameters
src_lang=None,
tgt_lang=None,
debug=False,
concurrency_count=None,
save_dir=None,
sanitize_bot_response=False,
model_state0=None,
memory_restriction_level=None,
max_max_new_tokens=None,
is_public=None,
max_max_time=None,
raise_generate_gpu_exceptions=None,
chat_context=None,
lora_weights=None,
load_db_if_exists=True,
dbs=None,
user_path=None,
detect_user_path_changes_every_query=None,
use_openai_embedding=None,
use_openai_model=None,
hf_embedding_model=None,
db_type=None,
n_jobs=None,
first_para=None,
text_limit=None,
verbose=False,
cli=False,
reverse_docs=True,
use_cache=None,
auto_reduce_chunks=None,
max_chunks=None,
model_lock=None,
force_langchain_evaluate=None,
model_state_none=None,
):
# ensure passed these
assert concurrency_count is not None
assert memory_restriction_level is not None
assert raise_generate_gpu_exceptions is not None
assert chat_context is not None
assert use_openai_embedding is not None
assert use_openai_model is not None
assert hf_embedding_model is not None
assert db_type is not None
assert top_k_docs is not None and isinstance(top_k_docs, int)
assert chunk is not None and isinstance(chunk, bool)
assert chunk_size is not None and isinstance(chunk_size, int)
assert n_jobs is not None
assert first_para is not None
if debug:
locals_dict = locals().copy()
locals_dict.pop('model_state', None)
locals_dict.pop('model_state0', None)
locals_dict.pop('model_states', None)
print(locals_dict)
no_model_msg = "Please choose a base model with --base_model (CLI) or load in Models Tab (gradio).\n" \
"Then start New Conversation"
if model_state is None:
model_state = model_state_none.copy()
if model_state0 is None:
# e.g. for no gradio case, set dummy value, else should be set
model_state0 = model_state_none.copy()
# model_state['model] is only 'model' if should use model_state0
# model could also be None
have_model_lock = model_lock is not None
have_fresh_model = model_state['model'] not in [None, 'model', no_model_str]
# for gradio UI control, expect model_state and model_state0 to match, so if have_model_lock=True, then should have_fresh_model=True
# but gradio API control will only use nochat api etc. and won't use fresh model, so can't assert in general
# if have_model_lock:
# assert have_fresh_model, "Expected model_state and model_state0 to match if have_model_lock"
have_cli_model = model_state0['model'] not in [None, 'model', no_model_str]
if have_fresh_model:
# USE FRESH MODEL
if not have_model_lock:
# model_state0 is just one of model_state if model_lock, so don't nuke
# try to free-up original model (i.e. list was passed as reference)
if model_state0['model'] and hasattr(model_state0['model'], 'cpu'):
model_state0['model'].cpu()
model_state0['model'] = None
# try to free-up original tokenizer (i.e. list was passed as reference)
if model_state0['tokenizer']:
model_state0['tokenizer'] = None
clear_torch_cache()
chosen_model_state = model_state
elif have_cli_model:
# USE MODEL SETUP AT CLI
assert isinstance(model_state['model'], str) # expect no fresh model
chosen_model_state = model_state0
else:
raise AssertionError(no_model_msg)
# get variables
model = chosen_model_state['model']
tokenizer = chosen_model_state['tokenizer']
device = chosen_model_state['device']
base_model = chosen_model_state['base_model']
tokenizer_base_model = chosen_model_state['tokenizer_base_model']
lora_weights = chosen_model_state['lora_weights']
inference_server = chosen_model_state['inference_server']
# prefer use input from API over model state
prompt_type = prompt_type or chosen_model_state['prompt_type']
prompt_dict = prompt_dict or chosen_model_state['prompt_dict']
if base_model is None:
raise AssertionError(no_model_msg)
assert base_model.strip(), no_model_msg
assert model, "Model is missing"
assert tokenizer, "Tokenizer is missing"
# choose chat or non-chat mode
if not chat:
instruction = instruction_nochat
iinput = iinput_nochat
# in some cases, like lean nochat API, don't want to force sending prompt_type, allow default choice
model_lower = base_model.lower()
if not prompt_type and model_lower in inv_prompt_type_to_model_lower:
prompt_type = inv_prompt_type_to_model_lower[model_lower]
if verbose:
print("Auto-selecting prompt_type=%s for %s" % (prompt_type, model_lower), flush=True)
assert prompt_type is not None, "prompt_type was None"
# Control generation hyperparameters
# adjust for bad inputs, e.g. in case also come from API that doesn't get constrained by gradio sliders
# below is for TGI server, not required for HF transformers
# limits are chosen similar to gradio_runner.py sliders/numbers
top_p = min(max(1e-3, top_p), 1.0 - 1e-3)
top_k = min(max(1, int(top_k)), 100)
temperature = min(max(0.01, temperature), 2.0)
# FIXME: https://github.com/h2oai/h2ogpt/issues/106
num_beams = 1 if stream_output else num_beams # See max_beams in gradio_runner
max_max_new_tokens = get_max_max_new_tokens(chosen_model_state,
memory_restriction_level=memory_restriction_level,
max_new_tokens=max_new_tokens,
max_max_new_tokens=max_max_new_tokens)
model_max_length = get_model_max_length(chosen_model_state)
max_new_tokens = min(max(1, int(max_new_tokens)), max_max_new_tokens)
min_new_tokens = min(max(0, int(min_new_tokens)), max_new_tokens)
max_time = min(max(0, max_time), max_max_time)
repetition_penalty = min(max(0.01, repetition_penalty), 3.0)
num_return_sequences = 1 if chat else min(max(1, int(num_return_sequences)), 10)
min_top_k_docs, max_top_k_docs, label_top_k_docs = get_minmax_top_k_docs(is_public)
top_k_docs = min(max(min_top_k_docs, int(top_k_docs)), max_top_k_docs)
chunk_size = min(max(128, int(chunk_size)), 2048)
if not context:
# get hidden context if have one
context = get_context(chat_context, prompt_type)
# restrict instruction, typically what has large input
from h2oai_pipeline import H2OTextGenerationPipeline
instruction, num_prompt_tokens1 = H2OTextGenerationPipeline.limit_prompt(instruction, tokenizer)
context, num_prompt_tokens2 = H2OTextGenerationPipeline.limit_prompt(context, tokenizer)
iinput, num_prompt_tokens3 = H2OTextGenerationPipeline.limit_prompt(iinput, tokenizer)
num_prompt_tokens = (num_prompt_tokens1 or 0) + (num_prompt_tokens2 or 0) + (num_prompt_tokens3 or 0)
# get prompt
prompter = Prompter(prompt_type, prompt_dict, debug=debug, chat=chat, stream_output=stream_output)
data_point = dict(context=context, instruction=instruction, input=iinput)
prompt = prompter.generate_prompt(data_point)
# THIRD PLACE where LangChain referenced, but imports only occur if enabled and have db to use
assert langchain_mode in langchain_modes, "Invalid langchain_mode %s" % langchain_mode
if langchain_mode in ['MyData'] and my_db_state is not None and len(my_db_state) > 0 and my_db_state[0] is not None:
db1 = my_db_state[0]
elif dbs is not None and langchain_mode in dbs:
db1 = dbs[langchain_mode]
else:
db1 = None
do_langchain_path = langchain_mode not in [False, 'Disabled', 'ChatLLM', 'LLM'] and \
db1 is not None or \
base_model in non_hf_types or \
force_langchain_evaluate
if do_langchain_path:
query = instruction if not iinput else "%s\n%s" % (instruction, iinput)
outr = ""
# use smaller cut_distanct for wiki_full since so many matches could be obtained, and often irrelevant unless close
from gpt_langchain import run_qa_db
gen_hyper_langchain = dict(do_sample=do_sample,
temperature=temperature,
repetition_penalty=repetition_penalty,
top_k=top_k,
top_p=top_p,
num_beams=num_beams,
min_new_tokens=min_new_tokens,
max_new_tokens=max_new_tokens,
early_stopping=early_stopping,
max_time=max_time,
num_return_sequences=num_return_sequences,
)
for r in run_qa_db(query=query,
model_name=base_model, model=model, tokenizer=tokenizer,
inference_server=inference_server,
stream_output=stream_output,
prompter=prompter,
load_db_if_exists=load_db_if_exists,
db=db1,
user_path=user_path,
detect_user_path_changes_every_query=detect_user_path_changes_every_query,
cut_distanct=1.1 if langchain_mode in ['wiki_full'] else 1.64, # FIXME, too arbitrary
use_openai_embedding=use_openai_embedding,
use_openai_model=use_openai_model,
hf_embedding_model=hf_embedding_model,
first_para=first_para,
text_limit=text_limit,
chunk=chunk,
chunk_size=chunk_size,
langchain_mode=langchain_mode,
document_choice=document_choice,
db_type=db_type,
top_k_docs=top_k_docs,
**gen_hyper_langchain,
prompt_type=prompt_type,
prompt_dict=prompt_dict,
n_jobs=n_jobs,
verbose=verbose,
cli=cli,
sanitize_bot_response=sanitize_bot_response,
reverse_docs=reverse_docs,
lora_weights=lora_weights,
auto_reduce_chunks=auto_reduce_chunks,
max_chunks=max_chunks,
):
outr, extra = r # doesn't accumulate, new answer every yield, so only save that full answer
yield dict(response=outr, sources=extra)
if save_dir:
extra_dict = gen_hyper_langchain.copy()
extra_dict.update(prompt_type=prompt_type, inference_server=inference_server,
langchain_mode=langchain_mode, document_choice=document_choice,
num_prompt_tokens=num_prompt_tokens)
save_generate_output(prompt=query, output=outr, base_model=base_model, save_dir=save_dir,
where_from='run_qa_db',
extra_dict=extra_dict)
if verbose:
print(
'Post-Generate Langchain: %s decoded_output: %s' % (str(datetime.now()), len(outr) if outr else -1),
flush=True)
if outr or base_model in non_hf_types:
# if got no response (e.g. not showing sources and got no sources,
# so nothing to give to LLM), then slip through and ask LLM
# Or if llama/gptj, then just return since they had no response and can't go down below code path
# clear before return, since .then() never done if from API
clear_torch_cache()
return
if inference_server.startswith('openai') or inference_server.startswith('http'):
if inference_server.startswith('openai'):
import openai
where_from = "openai_client"
openai.api_key = os.getenv("OPENAI_API_KEY")
stop_sequences = list(set(prompter.terminate_response + [prompter.PreResponse]))
# OpenAI will complain if ask for too many new tokens, takes it as min in some sense, wrongly so.
max_new_tokens_openai = min(max_new_tokens, model_max_length - num_prompt_tokens)
gen_server_kwargs = dict(temperature=temperature if do_sample else 0,
max_tokens=max_new_tokens_openai,
top_p=top_p if do_sample else 1,
frequency_penalty=0,
n=num_return_sequences,
presence_penalty=1.07 - repetition_penalty + 0.6, # so good default
)
if inference_server == 'openai':
response = openai.Completion.create(
model=base_model,
prompt=prompt,
**gen_server_kwargs,
stop=stop_sequences,
stream=stream_output,
)
if not stream_output:
text = response['choices'][0]['text']
yield dict(response=prompter.get_response(prompt + text, prompt=prompt,
sanitize_bot_response=sanitize_bot_response),
sources='')
else:
collected_events = []
text = ''
for event in response:
collected_events.append(event) # save the event response
event_text = event['choices'][0]['text'] # extract the text
text += event_text # append the text
yield dict(response=prompter.get_response(prompt + text, prompt=prompt,
sanitize_bot_response=sanitize_bot_response),
sources='')
elif inference_server == 'openai_chat':
response = openai.ChatCompletion.create(
model=base_model,
messages=[
{"role": "system", "content": "You are a helpful assistant."},
{'role': 'user',
'content': prompt,
}
],
stream=stream_output,
**gen_server_kwargs,
)
if not stream_output:
text = response["choices"][0]["message"]["content"]
yield dict(response=prompter.get_response(prompt + text, prompt=prompt,
sanitize_bot_response=sanitize_bot_response),
sources='')
else:
text = ""
for chunk in response:
delta = chunk["choices"][0]["delta"]
if 'content' in delta:
text += delta['content']
yield dict(response=prompter.get_response(prompt + text, prompt=prompt,
sanitize_bot_response=sanitize_bot_response),
sources='')
else:
raise RuntimeError("No such OpenAI mode: %s" % inference_server)
elif inference_server.startswith('http'):
inference_server, headers = get_hf_server(inference_server)
from gradio_utils.grclient import GradioClient
from text_generation import Client as HFClient
if isinstance(model, GradioClient):
gr_client = model
hf_client = None
elif isinstance(model, HFClient):
gr_client = None
hf_client = model
else:
inference_server, gr_client, hf_client = get_client_from_inference_server(inference_server)
# quick sanity check to avoid long timeouts, just see if can reach server
requests.get(inference_server, timeout=int(os.getenv('REQUEST_TIMEOUT_FAST', '10')))
if gr_client is not None:
# Note: h2oGPT gradio server could handle input token size issues for prompt,
# but best to handle here so send less data to server
chat_client = False
where_from = "gr_client"
client_langchain_mode = 'Disabled'
gen_server_kwargs = dict(temperature=temperature,
top_p=top_p,
top_k=top_k,
num_beams=num_beams,
max_new_tokens=max_new_tokens,
min_new_tokens=min_new_tokens,
early_stopping=early_stopping,
max_time=max_time,
repetition_penalty=repetition_penalty,
num_return_sequences=num_return_sequences,
do_sample=do_sample,
chat=chat_client,
)
# account for gradio into gradio that handles prompting, avoid duplicating prompter prompt injection
if prompt_type in [None, '', PromptType.plain.name, PromptType.plain.value,
str(PromptType.plain.value)]:
# if our prompt is plain, assume either correct or gradio server knows different prompt type,
# so pass empty prompt_Type
gr_prompt_type = ''
gr_prompt_dict = ''
gr_prompt = prompt # already prepared prompt
gr_context = ''
gr_iinput = ''
else:
# if already have prompt_type that is not plain, None, or '', then already applied some prompting
# But assume server can handle prompting, and need to avoid double-up.
# Also assume server can do better job of using stopping.py to stop early, so avoid local prompting, let server handle
# So avoid "prompt" and let gradio server reconstruct from prompt_type we passed
# Note it's ok that prompter.get_response() has prompt+text, prompt=prompt passed,
# because just means extra processing and removal of prompt, but that has no human-bot prompting doesn't matter
# since those won't appear
gr_context = context
gr_prompt = instruction
gr_iinput = iinput
gr_prompt_type = prompt_type
gr_prompt_dict = prompt_dict
client_kwargs = dict(instruction=gr_prompt if chat_client else '', # only for chat=True
iinput=gr_iinput, # only for chat=True
context=gr_context,
# streaming output is supported, loops over and outputs each generation in streaming mode
# but leave stream_output=False for simple input/output mode
stream_output=stream_output,
**gen_server_kwargs,
prompt_type=gr_prompt_type,
prompt_dict=gr_prompt_dict,
instruction_nochat=gr_prompt if not chat_client else '',
iinput_nochat=gr_iinput, # only for chat=False
langchain_mode=client_langchain_mode,
top_k_docs=top_k_docs,
chunk=chunk,
chunk_size=chunk_size,
document_choice=[DocumentChoices.All_Relevant.name],
)
api_name = '/submit_nochat_api' # NOTE: like submit_nochat but stable API for string dict passing
if not stream_output:
res = gr_client.predict(str(dict(client_kwargs)), api_name=api_name)
res_dict = ast.literal_eval(res)
text = res_dict['response']
sources = res_dict['sources']
yield dict(response=prompter.get_response(prompt + text, prompt=prompt,
sanitize_bot_response=sanitize_bot_response),
sources=sources)
else:
job = gr_client.submit(str(dict(client_kwargs)), api_name=api_name)
text = ''
sources = ''
res_dict = dict(response=text, sources=sources)
while not job.done():
outputs_list = job.communicator.job.outputs
if outputs_list:
res = job.communicator.job.outputs[-1]
res_dict = ast.literal_eval(res)
text = res_dict['response']
sources = res_dict['sources']
if gr_prompt_type == 'plain':
# then gradio server passes back full prompt + text
prompt_and_text = text
else:
prompt_and_text = prompt + text
yield dict(response=prompter.get_response(prompt_and_text, prompt=prompt,
sanitize_bot_response=sanitize_bot_response),
sources=sources)
time.sleep(0.01)
# ensure get last output to avoid race
res_all = job.outputs()
if len(res_all) > 0:
res = res_all[-1]
res_dict = ast.literal_eval(res)
text = res_dict['response']
sources = res_dict['sources']
else:
# go with old text if last call didn't work
e = job.future._exception
if e is not None:
stre = str(e)
strex = ''.join(traceback.format_tb(e.__traceback__))
else:
stre = ''
strex = ''
print("Bad final response: %s %s %s %s %s: %s %s" % (base_model, inference_server,
res_all, prompt, text, stre, strex),
flush=True)
if gr_prompt_type == 'plain':
# then gradio server passes back full prompt + text
prompt_and_text = text
else:
prompt_and_text = prompt + text
yield dict(response=prompter.get_response(prompt_and_text, prompt=prompt,
sanitize_bot_response=sanitize_bot_response),
sources=sources)
elif hf_client:
# HF inference server needs control over input tokens
where_from = "hf_client"
# prompt must include all human-bot like tokens, already added by prompt
# https://github.com/huggingface/text-generation-inference/tree/main/clients/python#types
stop_sequences = list(set(prompter.terminate_response + [prompter.PreResponse]))
gen_server_kwargs = dict(do_sample=do_sample,
max_new_tokens=max_new_tokens,
# best_of=None,
repetition_penalty=repetition_penalty,
return_full_text=True,
seed=SEED,
stop_sequences=stop_sequences,
temperature=temperature,
top_k=top_k,
top_p=top_p,
# truncate=False, # behaves oddly
# typical_p=top_p,
# watermark=False,
# decoder_input_details=False,
)
# work-around for timeout at constructor time, will be issue if multi-threading,
# so just do something reasonable or max_time if larger
# lower bound because client is re-used if multi-threading
hf_client.timeout = max(300, max_time)
if not stream_output:
text = hf_client.generate(prompt, **gen_server_kwargs).generated_text
yield dict(response=prompter.get_response(text, prompt=prompt,
sanitize_bot_response=sanitize_bot_response),
sources='')
else:
text = ""
for response in hf_client.generate_stream(prompt, **gen_server_kwargs):
if not response.token.special:
# stop_sequences
text_chunk = response.token.text
text += text_chunk
yield dict(response=prompter.get_response(prompt + text, prompt=prompt,
sanitize_bot_response=sanitize_bot_response),
sources='')
else:
raise RuntimeError("Failed to get client: %s" % inference_server)
else:
raise RuntimeError("No such inference_server %s" % inference_server)
if save_dir and text:
# save prompt + new text
extra_dict = gen_server_kwargs.copy()
extra_dict.update(dict(inference_server=inference_server, num_prompt_tokens=num_prompt_tokens))
save_generate_output(prompt=prompt, output=text, base_model=base_model, save_dir=save_dir,
where_from=where_from, extra_dict=extra_dict)
return
else:
assert not inference_server, "inferene_server=%s not supported" % inference_server
if isinstance(tokenizer, str):
# pipeline
if tokenizer == "summarization":
key = 'summary_text'
else:
raise RuntimeError("No such task type %s" % tokenizer)
# NOTE: uses max_length only
yield dict(response=model(prompt, max_length=max_new_tokens)[0][key], sources='')
if 'mbart-' in base_model.lower():
assert src_lang is not None
tokenizer.src_lang = languages_covered()[src_lang]
stopping_criteria = get_stopping(prompt_type, prompt_dict, tokenizer, device,
model_max_length=tokenizer.model_max_length)
inputs = tokenizer(prompt, return_tensors="pt")
if debug and len(inputs["input_ids"]) > 0:
print('input_ids length', len(inputs["input_ids"][0]), flush=True)
input_ids = inputs["input_ids"].to(device)
# CRITICAL LIMIT else will fail
max_max_tokens = tokenizer.model_max_length
max_input_tokens = max_max_tokens - min_new_tokens
# NOTE: Don't limit up front due to max_new_tokens, let go up to max or reach max_max_tokens in stopping.py
input_ids = input_ids[:, -max_input_tokens:]
# required for falcon if multiple threads or asyncio accesses to model during generation
if use_cache is None:
use_cache = False if 'falcon' in base_model else True
gen_config_kwargs = dict(temperature=float(temperature),
top_p=float(top_p),
top_k=top_k,
num_beams=num_beams,
do_sample=do_sample,
repetition_penalty=float(repetition_penalty),
num_return_sequences=num_return_sequences,
renormalize_logits=True,
remove_invalid_values=True,
use_cache=use_cache,
)
token_ids = ['eos_token_id', 'pad_token_id', 'bos_token_id', 'cls_token_id', 'sep_token_id']
for token_id in token_ids:
if hasattr(tokenizer, token_id) and getattr(tokenizer, token_id) is not None:
gen_config_kwargs.update({token_id: getattr(tokenizer, token_id)})
generation_config = GenerationConfig(**gen_config_kwargs)
gen_kwargs = dict(input_ids=input_ids,
generation_config=generation_config,
return_dict_in_generate=True,
output_scores=True,
max_new_tokens=max_new_tokens, # prompt + new
min_new_tokens=min_new_tokens, # prompt + new
early_stopping=early_stopping, # False, True, "never"
max_time=max_time,
stopping_criteria=stopping_criteria,
)
if 'gpt2' in base_model.lower():
gen_kwargs.update(dict(bos_token_id=tokenizer.bos_token_id, pad_token_id=tokenizer.eos_token_id))
elif 'mbart-' in base_model.lower():
assert tgt_lang is not None
tgt_lang = languages_covered()[tgt_lang]
gen_kwargs.update(dict(forced_bos_token_id=tokenizer.lang_code_to_id[tgt_lang]))
else:
token_ids = ['eos_token_id', 'bos_token_id', 'pad_token_id']
for token_id in token_ids:
if hasattr(tokenizer, token_id) and getattr(tokenizer, token_id) is not None:
gen_kwargs.update({token_id: getattr(tokenizer, token_id)})
decoder_kwargs = dict(skip_special_tokens=True,
clean_up_tokenization_spaces=True)
decoder = functools.partial(tokenizer.decode,
**decoder_kwargs
)
decoder_raw_kwargs = dict(skip_special_tokens=False,
clean_up_tokenization_spaces=True)
decoder_raw = functools.partial(tokenizer.decode,
**decoder_raw_kwargs
)
with torch.no_grad():
have_lora_weights = lora_weights not in [no_lora_str, '', None]
context_class_cast = NullContext if device == 'cpu' or have_lora_weights else torch.autocast
with context_class_cast(device):
# protection for gradio not keeping track of closed users,
# else hit bitsandbytes lack of thread safety:
# https://github.com/h2oai/h2ogpt/issues/104
# but only makes sense if concurrency_count == 1
context_class = NullContext # if concurrency_count > 1 else filelock.FileLock
if verbose:
print('Pre-Generate: %s' % str(datetime.now()), flush=True)
decoded_output = None
with context_class("generate.lock"):
if verbose:
print('Generate: %s' % str(datetime.now()), flush=True)
# decoded tokenized prompt can deviate from prompt due to special characters
inputs_decoded = decoder(input_ids[0])
inputs_decoded_raw = decoder_raw(input_ids[0])
if inputs_decoded == prompt:
# normal
pass
elif inputs_decoded.lstrip() == prompt.lstrip():
# sometimes extra space in front, make prompt same for prompt removal
prompt = inputs_decoded
elif inputs_decoded_raw == prompt:
# some models specify special tokens that are part of normal prompt, so can't skip them
inputs_decoded = prompt = inputs_decoded_raw
decoder = decoder_raw
decoder_kwargs = decoder_raw_kwargs
elif inputs_decoded_raw.replace("<unk> ", "").replace("<unk>", "").replace('\n', ' ').replace(' ',
'') == prompt.replace(
'\n', ' ').replace(' ', ''):
inputs_decoded = prompt = inputs_decoded_raw
decoder = decoder_raw
decoder_kwargs = decoder_raw_kwargs
else:
if verbose:
print("WARNING: Special characters in prompt", flush=True)
if stream_output:
skip_prompt = False
streamer = H2OTextIteratorStreamer(tokenizer, skip_prompt=skip_prompt, block=False,
**decoder_kwargs)
gen_kwargs.update(dict(streamer=streamer))
target = wrapped_partial(generate_with_exceptions, model.generate,
prompt=prompt, inputs_decoded=inputs_decoded,
raise_generate_gpu_exceptions=raise_generate_gpu_exceptions,
**gen_kwargs)
bucket = queue.Queue()
thread = EThread(target=target, streamer=streamer, bucket=bucket)
thread.start()
outputs = ""
try:
for new_text in streamer:
if bucket.qsize() > 0 or thread.exc:
thread.join()
outputs += new_text
yield dict(response=prompter.get_response(outputs, prompt=inputs_decoded,
sanitize_bot_response=sanitize_bot_response),
sources='')
except BaseException:
# if any exception, raise that exception if was from thread, first
if thread.exc:
raise thread.exc
raise
finally:
# clear before return, since .then() never done if from API
clear_torch_cache()
# in case no exception and didn't join with thread yet, then join
if not thread.exc:
thread.join()
# in case raise StopIteration or broke queue loop in streamer, but still have exception
if thread.exc:
raise thread.exc
decoded_output = outputs
else:
try:
outputs = model.generate(**gen_kwargs)
finally:
clear_torch_cache() # has to be here for API submit_nochat_api since.then() not called
outputs = [decoder(s) for s in outputs.sequences]
yield dict(response=prompter.get_response(outputs, prompt=inputs_decoded,
sanitize_bot_response=sanitize_bot_response), sources='')
if outputs and len(outputs) >= 1:
decoded_output = prompt + outputs[0]
if save_dir and decoded_output:
extra_dict = gen_config_kwargs.copy()
extra_dict.update(dict(num_prompt_tokens=num_prompt_tokens))
save_generate_output(prompt=prompt, output=decoded_output, base_model=base_model, save_dir=save_dir,
where_from="evaluate_%s" % str(stream_output),
extra_dict=gen_config_kwargs)
if verbose:
print('Post-Generate: %s decoded_output: %s' % (
str(datetime.now()), len(decoded_output) if decoded_output else -1), flush=True)
inputs_list_names = list(inspect.signature(evaluate).parameters)
state_names = ['model_state', 'my_db_state']
inputs_kwargs_list = [x for x in inputs_list_names if x not in eval_func_param_names + state_names]
def get_cutoffs(memory_restriction_level, for_context=False, model_max_length=2048):
# help to avoid errors like:
# RuntimeError: The size of tensor a (2048) must match the size of tensor b (2049) at non-singleton dimension 3
# RuntimeError: expected scalar type Half but found Float
# with - 256
if memory_restriction_level > 0:
max_length_tokenize = 768 - 256 if memory_restriction_level <= 2 else 512 - 256
else:
# at least give room for 1 paragraph output
max_length_tokenize = model_max_length - 256
cutoff_len = max_length_tokenize * 4 # if reaches limit, then can't generate new tokens
output_smallest = 30 * 4
max_prompt_length = cutoff_len - output_smallest
if for_context:
# then lower even more to avoid later chop, since just estimate tokens in context bot
max_prompt_length = max(64, int(max_prompt_length * 0.8))
return cutoff_len, output_smallest, max_length_tokenize, max_prompt_length
class H2OTextIteratorStreamer(TextIteratorStreamer):
"""
normally, timeout required for now to handle exceptions, else get()
but with H2O version of TextIteratorStreamer, loop over block to handle
"""
def __init__(self, tokenizer, skip_prompt: bool = False, timeout: typing.Optional[float] = None,
block=True, **decode_kwargs):
super().__init__(tokenizer, skip_prompt, **decode_kwargs)
self.text_queue = queue.Queue()
self.stop_signal = None
self.do_stop = False
self.timeout = timeout
self.block = block
def on_finalized_text(self, text: str, stream_end: bool = False):
"""Put the new text in the queue. If the stream is ending, also put a stop signal in the queue."""
self.text_queue.put(text, timeout=self.timeout)
if stream_end:
self.text_queue.put(self.stop_signal, timeout=self.timeout)
def __iter__(self):
return self
def __next__(self):
while True:
try:
value = self.stop_signal # value looks unused in pycharm, not true
if self.do_stop:
print("hit stop", flush=True)
# could raise or break, maybe best to raise and make parent see if any exception in thread
raise StopIteration()
# break
value = self.text_queue.get(block=self.block, timeout=self.timeout)
break
except queue.Empty:
time.sleep(0.01)
if value == self.stop_signal:
raise StopIteration()
else:
return value
def generate_with_exceptions(func, *args, prompt='', inputs_decoded='', raise_generate_gpu_exceptions=True, **kwargs):
try:
func(*args, **kwargs)
except torch.cuda.OutOfMemoryError as e:
print("GPU OOM 2: prompt: %s inputs_decoded: %s exception: %s" % (prompt, inputs_decoded, str(e)),
flush=True)
if 'input_ids' in kwargs:
if kwargs['input_ids'] is not None:
kwargs['input_ids'].cpu()
kwargs['input_ids'] = None
traceback.print_exc()
clear_torch_cache()
return
except (Exception, RuntimeError) as e:
if 'Expected all tensors to be on the same device' in str(e) or \
'expected scalar type Half but found Float' in str(e) or \
'probability tensor contains either' in str(e) or \
'cublasLt ran into an error!' in str(e) or \
'mat1 and mat2 shapes cannot be multiplied' in str(e):
print(
"GPU Error: prompt: %s inputs_decoded: %s exception: %s" % (prompt, inputs_decoded, str(e)),
flush=True)
traceback.print_exc()
clear_torch_cache()
if raise_generate_gpu_exceptions:
raise
return
else:
clear_torch_cache()
if raise_generate_gpu_exceptions:
raise
def get_generate_params(model_lower, chat,
stream_output, show_examples,
prompt_type, prompt_dict,
temperature, top_p, top_k, num_beams,
max_new_tokens, min_new_tokens, early_stopping, max_time,
repetition_penalty, num_return_sequences,
do_sample,
top_k_docs, chunk, chunk_size,
verbose):
use_defaults = False
use_default_examples = True
examples = []
task_info = 'LLM'
if model_lower:
print(f"Using Model {model_lower}", flush=True)
else:
if verbose:
print("No model defined yet", flush=True)
min_new_tokens = min_new_tokens if min_new_tokens is not None else 0
early_stopping = early_stopping if early_stopping is not None else False
max_time_defaults = 60 * 3
max_time = max_time if max_time is not None else max_time_defaults
if not prompt_type and model_lower in inv_prompt_type_to_model_lower:
prompt_type = inv_prompt_type_to_model_lower[model_lower]
if verbose:
print("Auto-selecting prompt_type=%s for %s" % (prompt_type, model_lower), flush=True)
# examples at first don't include chat, instruction_nochat, iinput_nochat, added at end
if show_examples is None:
if chat:
show_examples = False
else:
show_examples = True
summarize_example1 = """Jeff: Can I train a ? Transformers model on Amazon SageMaker?
Philipp: Sure you can use the new Hugging Face Deep Learning Container.
Jeff: ok.
Jeff: and how can I get started?
Jeff: where can I find documentation?
Philipp: ok, ok you can find everything here. https://huggingface.co/blog/the-partnership-amazon-sagemaker-and-hugging-face"""
use_placeholder_instruction_as_example = False
if 'bart-large-cnn-samsum' in model_lower or 'flan-t5-base-samsum' in model_lower:
placeholder_instruction = summarize_example1
placeholder_input = ""
use_defaults = True
use_default_examples = False
use_placeholder_instruction_as_example = True
task_info = "Summarization"
elif 't5-' in model_lower or 't5' == model_lower or 'flan-' in model_lower:
placeholder_instruction = "The square root of x is the cube root of y. What is y to the power of 2, if x = 4?"
placeholder_input = ""
use_defaults = True
use_default_examples = True
task_info = "Multi-Task: Q/A, translation, Chain-of-Thought, Logical Reasoning, Summarization, etc. Best to use task prefix as trained on, e.g. `translate English to German: ` (space after colon)"
elif 'mbart-' in model_lower:
placeholder_instruction = "The girl has long hair."
placeholder_input = ""
use_defaults = True
use_default_examples = False
use_placeholder_instruction_as_example = True
elif 'gpt2' in model_lower:
placeholder_instruction = "The sky is"
placeholder_input = ""
prompt_type = prompt_type or 'plain'
use_default_examples = True # some will be odd "continuations" but can be ok
use_placeholder_instruction_as_example = True
task_info = "Auto-complete phrase, code, etc."
use_defaults = True
else:
if chat:
placeholder_instruction = ""
else:
placeholder_instruction = "Give detailed answer for whether Einstein or Newton is smarter."
placeholder_input = ""
if model_lower in inv_prompt_type_to_model_lower:
prompt_type = inv_prompt_type_to_model_lower[model_lower]
elif model_lower:
# default is plain, because might rely upon trust_remote_code to handle prompting
prompt_type = prompt_type or 'plain'
else:
prompt_type = ''
task_info = "No task"
if prompt_type == 'instruct':
task_info = "Answer question or follow imperative as instruction with optionally input."
elif prompt_type == 'plain':
task_info = "Auto-complete phrase, code, etc."
elif prompt_type == 'human_bot':
if chat:
task_info = "Chat (Shift-Enter to give question/imperative, input concatenated with instruction)"
else:
task_info = "Ask question/imperative (input concatenated with instruction)"
# revert to plain if still nothing
prompt_type = prompt_type or 'plain'
if use_defaults:
temperature = 1.0 if temperature is None else temperature
top_p = 1.0 if top_p is None else top_p
top_k = 40 if top_k is None else top_k
num_beams = num_beams or 1
max_new_tokens = max_new_tokens or 128
repetition_penalty = repetition_penalty or 1.07
num_return_sequences = min(num_beams, num_return_sequences or 1)
do_sample = False if do_sample is None else do_sample
else:
temperature = 0.1 if temperature is None else temperature
top_p = 0.75 if top_p is None else top_p
top_k = 40 if top_k is None else top_k
num_beams = num_beams or 1
max_new_tokens = max_new_tokens or 256
repetition_penalty = repetition_penalty or 1.07
num_return_sequences = min(num_beams, num_return_sequences or 1)
do_sample = False if do_sample is None else do_sample
# doesn't include chat, instruction_nochat, iinput_nochat, added later
params_list = ["",
stream_output,
prompt_type, prompt_dict,
temperature, top_p, top_k, num_beams,
max_new_tokens, min_new_tokens,
early_stopping, max_time, repetition_penalty, num_return_sequences, do_sample]
if use_placeholder_instruction_as_example:
examples += [[placeholder_instruction, ''] + params_list]
if use_default_examples:
examples += [
["Translate English to French", "Good morning"] + params_list,
["Give detailed answer for whether Einstein or Newton is smarter.", ''] + params_list,
["Explain in detailed list, all the best practices for coding in python.", ''] + params_list,
[
"Create a markdown table with 3 rows for the primary colors, and 2 columns, with color name and hex codes.",
''] + params_list,
['Translate to German: My name is Arthur', ''] + params_list,
["Please answer to the following question. Who is going to be the next Ballon d'or?", ''] + params_list,
['Can Geoffrey Hinton have a conversation with George Washington? Give the rationale before answering.',
''] + params_list,
['Please answer the following question. What is the boiling point of Nitrogen?', ''] + params_list,
['Answer the following yes/no question. Can you write a whole Haiku in a single tweet?', ''] + params_list,
["Simplify the following expression: (False or False and True). Explain your answer.", ''] + params_list,
[
"Premise: At my age you will probably have learnt one lesson. Hypothesis: It's not certain how many lessons you'll learn by your thirties. Does the premise entail the hypothesis?",
''] + params_list,
['The square root of x is the cube root of y. What is y to the power of 2, if x = 4?', ''] + params_list,
[
'Answer the following question by reasoning step by step. The cafeteria had 23 apples. If they used 20 for lunch, and bought 6 more, how many apple do they have?',
''] + params_list,
["""def area_of_rectangle(a: float, b: float):
\"\"\"Return the area of the rectangle.\"\"\"""", ''] + params_list,
["""# a function in native python:
def mean(a):
return sum(a)/len(a)
# the same function using numpy:
import numpy as np
def mean(a):""", ''] + params_list,
["""X = np.random.randn(100, 100)
y = np.random.randint(0, 1, 100)
# fit random forest classifier with 20 estimators""", ''] + params_list,
]
# add summary example
examples += [
[summarize_example1, 'Summarize' if prompt_type not in ['plain', 'instruct_simple'] else ''] + params_list]
src_lang = "English"
tgt_lang = "Russian"
# move to correct position
for example in examples:
example += [chat, '', '', 'Disabled', top_k_docs, chunk, chunk_size, [DocumentChoices.All_Relevant.name]]
# adjust examples if non-chat mode
if not chat:
example[eval_func_param_names.index('instruction_nochat')] = example[
eval_func_param_names.index('instruction')]
example[eval_func_param_names.index('instruction')] = ''
example[eval_func_param_names.index('iinput_nochat')] = example[eval_func_param_names.index('iinput')]
example[eval_func_param_names.index('iinput')] = ''
assert len(example) == len(eval_func_param_names), "Wrong example: %s %s" % (
len(example), len(eval_func_param_names))
if prompt_type == PromptType.custom.name and not prompt_dict:
raise ValueError("Unexpected to get non-empty prompt_dict=%s for prompt_type=%s" % (prompt_dict, prompt_type))
# get prompt_dict from prompt_type, so user can see in UI etc., or for custom do nothing except check format
prompt_dict, error0 = get_prompt(prompt_type, prompt_dict,
chat=False, context='', reduced=False, making_context=False, return_dict=True)
if error0:
raise RuntimeError("Prompt wrong: %s" % error0)
return placeholder_instruction, placeholder_input, \
stream_output, show_examples, \
prompt_type, prompt_dict, \
temperature, top_p, top_k, num_beams, \
max_new_tokens, min_new_tokens, early_stopping, max_time, \
repetition_penalty, num_return_sequences, \
do_sample, \
src_lang, tgt_lang, \
examples, \
task_info
def languages_covered():
# https://huggingface.co/facebook/mbart-large-50-many-to-many-mmt#languages-covered
covered = """Arabic (ar_AR), Czech (cs_CZ), German (de_DE), English (en_XX), Spanish (es_XX), Estonian (et_EE), Finnish (fi_FI), French (fr_XX), Gujarati (gu_IN), Hindi (hi_IN), Italian (it_IT), Japanese (ja_XX), Kazakh (kk_KZ), Korean (ko_KR), Lithuanian (lt_LT), Latvian (lv_LV), Burmese (my_MM), Nepali (ne_NP), Dutch (nl_XX), Romanian (ro_RO), Russian (ru_RU), Sinhala (si_LK), Turkish (tr_TR), Vietnamese (vi_VN), Chinese (zh_CN), Afrikaans (af_ZA), Azerbaijani (az_AZ), Bengali (bn_IN), Persian (fa_IR), Hebrew (he_IL), Croatian (hr_HR), Indonesian (id_ID), Georgian (ka_GE), Khmer (km_KH), Macedonian (mk_MK), Malayalam (ml_IN), Mongolian (mn_MN), Marathi (mr_IN), Polish (pl_PL), Pashto (ps_AF), Portuguese (pt_XX), Swedish (sv_SE), Swahili (sw_KE), Tamil (ta_IN), Telugu (te_IN), Thai (th_TH), Tagalog (tl_XX), Ukrainian (uk_UA), Urdu (ur_PK), Xhosa (xh_ZA), Galician (gl_ES), Slovene (sl_SI)"""
covered = covered.split(', ')
covered = {x.split(' ')[0]: x.split(' ')[1].replace(')', '').replace('(', '') for x in covered}
return covered
def get_context(chat_context, prompt_type):
if chat_context and prompt_type == 'human_bot':
context0 = """<bot>: I am an intelligent, helpful, truthful, and fair assistant named h2oGPT, who will give accurate, balanced, and reliable responses. I will not respond with I don't know or I don't understand.
<human>: I am a human person seeking useful assistance and request all questions be answered completely, and typically expect detailed responses. Give answers in numbered list format if several distinct but related items are being listed."""
else:
context0 = ''
return context0
def score_qa(smodel, stokenizer, max_length_tokenize, question, answer, cutoff_len):
question = question[-cutoff_len:]
answer = answer[-cutoff_len:]
inputs = stokenizer(question, answer,
return_tensors="pt",
truncation=True,
max_length=max_length_tokenize).to(smodel.device)
try:
score = torch.sigmoid(smodel(**inputs).logits[0]).cpu().detach().numpy()[0]
except torch.cuda.OutOfMemoryError as e:
print("GPU OOM 3: question: %s answer: %s exception: %s" % (question, answer, str(e)), flush=True)
del inputs
traceback.print_exc()
clear_torch_cache()
return 'Response Score: GPU OOM'
except (Exception, RuntimeError) as e:
if 'Expected all tensors to be on the same device' in str(e) or \
'expected scalar type Half but found Float' in str(e) or \
'probability tensor contains either' in str(e) or \
'cublasLt ran into an error!' in str(e) or \
'device-side assert triggered' in str(e):
print("GPU Error: question: %s answer: %s exception: %s" % (question, answer, str(e)),
flush=True)
traceback.print_exc()
clear_torch_cache()
return 'Response Score: GPU Error'
else:
raise
os.environ['TOKENIZERS_PARALLELISM'] = 'true'
return score
def check_locals(**kwargs):
# ensure everything in evaluate is here
can_skip_because_locally_generated = no_default_param_names + [
# get_model:
'reward_type'
]
for k in eval_func_param_names:
if k in can_skip_because_locally_generated:
continue
assert k in kwargs, "Missing %s" % k
for k in inputs_kwargs_list:
if k in can_skip_because_locally_generated:
continue
assert k in kwargs, "Missing %s" % k
for k in list(inspect.signature(get_model).parameters):
if k in can_skip_because_locally_generated:
continue
assert k in kwargs, "Missing %s" % k
def get_model_max_length(model_state):
if not isinstance(model_state['tokenizer'], (str, types.NoneType)):
return model_state['tokenizer'].model_max_length
else:
return 2048
def get_max_max_new_tokens(model_state, **kwargs):
if not isinstance(model_state['tokenizer'], (str, types.NoneType)):
max_max_new_tokens = model_state['tokenizer'].model_max_length
else:
max_max_new_tokens = None
if kwargs['max_max_new_tokens'] is not None and max_max_new_tokens is not None:
return min(max_max_new_tokens, kwargs['max_max_new_tokens'])
elif kwargs['max_max_new_tokens'] is not None:
return kwargs['max_max_new_tokens']
elif kwargs['memory_restriction_level'] == 1:
return 768
elif kwargs['memory_restriction_level'] == 2:
return 512
elif kwargs['memory_restriction_level'] >= 3:
return 256
else:
# FIXME: Need to update after new model loaded, so user can control with slider
return 2048
def get_minmax_top_k_docs(is_public):
if is_public:
min_top_k_docs = 1
max_top_k_docs = 3
label_top_k_docs = "Number of document chunks"
else:
min_top_k_docs = -1
max_top_k_docs = 100
label_top_k_docs = "Number of document chunks (-1 = auto fill model context)"
return min_top_k_docs, max_top_k_docs, label_top_k_docs
def history_to_context(history, langchain_mode1, prompt_type1, prompt_dict1, chat1, model_max_length1,
memory_restriction_level1, keep_sources_in_context1):
"""
consumes all history up to (but not including) latest history item that is presumed to be an [instruction, None] pair
:param history:
:param langchain_mode1:
:param prompt_type1:
:param prompt_dict1:
:param chat1:
:param model_max_length1:
:param memory_restriction_level1:
:param keep_sources_in_context1:
:return:
"""
# ensure output will be unique to models
_, _, _, max_prompt_length = get_cutoffs(memory_restriction_level1,
for_context=True, model_max_length=model_max_length1)
context1 = ''
if max_prompt_length is not None and langchain_mode1 not in ['LLM']:
context1 = ''
# - 1 below because current instruction already in history from user()
for histi in range(0, len(history) - 1):
data_point = dict(instruction=history[histi][0], input='', output=history[histi][1])
prompt, pre_response, terminate_response, chat_sep, chat_turn_sep = generate_prompt(data_point,
prompt_type1,
prompt_dict1,
chat1,
reduced=True,
making_context=True)
# md -> back to text, maybe not super important if model trained enough
if not keep_sources_in_context1 and langchain_mode1 != 'Disabled' and prompt.find(source_prefix) >= 0:
# FIXME: This is relatively slow even for small amount of text, like 0.3s each history item
import re
prompt = re.sub(f'{re.escape(source_prefix)}.*?{re.escape(source_postfix)}', '', prompt,
flags=re.DOTALL)
if prompt.endswith('\n<p>'):
prompt = prompt[:-4]
prompt = prompt.replace('<br>', chat_turn_sep)
if not prompt.endswith(chat_turn_sep):
prompt += chat_turn_sep
# most recent first, add older if can
# only include desired chat history
if len(prompt + context1) > max_prompt_length:
break
context1 += prompt
_, pre_response, terminate_response, chat_sep, chat_turn_sep = generate_prompt({}, prompt_type1, prompt_dict1,
chat1, reduced=True,
making_context=True)
if context1 and not context1.endswith(chat_turn_sep):
context1 += chat_turn_sep # ensure if terminates abruptly, then human continues on next line
return context1
def entrypoint_main():
"""
Examples:
WORLD_SIZE=4 CUDA_VISIBLE_DEVICES="0,1,2,3" torchrun --nproc_per_node=4 --master_port=1234 generate.py --base_model='EleutherAI/gpt-j-6B' --lora_weights=lora-alpaca_6B
python generate.py --base_model='EleutherAI/gpt-j-6B' --lora_weights='lora-alpaca_6B'
python generate.py --base_model='EleutherAI/gpt-neox-20b' --lora_weights='lora-alpaca_20B'
# generate without lora weights, no prompt
python generate.py --base_model='EleutherAI/gpt-neox-20b' --prompt_type='plain'
python generate.py --base_model='togethercomputer/GPT-NeoXT-Chat-Base-20B' --prompt_type='dai_faq'
python generate.py --base_model='togethercomputer/GPT-NeoXT-Chat-Base-20B' --prompt_type='dai_faq' --lora_weights='lora_20B_daifaq'
# OpenChatKit settings:
python generate.py --base_model='togethercomputer/GPT-NeoXT-Chat-Base-20B' --prompt_type='human_bot --debug=True --num_beams=1 --temperature=0.6 --top_k=40 --top_p=1.0
python generate.py --base_model='distilgpt2' --prompt_type='plain' --debug=True --num_beams=1 --temperature=0.6 --top_k=40 --top_p=1.0 --share=False
python generate.py --base_model='t5-large' --prompt_type='simple_instruct'
python generate.py --base_model='philschmid/bart-large-cnn-samsum'
python generate.py --base_model='philschmid/flan-t5-base-samsum'
python generate.py --base_model='facebook/mbart-large-50-many-to-many-mmt'
python generate.py --base_model='togethercomputer/GPT-NeoXT-Chat-Base-20B' --prompt_type='human_bot' --lora_weights='GPT-NeoXT-Chat-Base-20B.merged.json.8_epochs.57b2892c53df5b8cefac45f84d019cace803ef26.28'
must have 4*48GB GPU and run without 8bit in order for sharding to work with infer_devices=False
can also pass --prompt_type='human_bot' and model can somewhat handle instructions without being instruct tuned
python generate.py --base_model=decapoda-research/llama-65b-hf --load_8bit=False --infer_devices=False --prompt_type='human_bot'
python generate.py --base_model=h2oai/h2ogpt-oig-oasst1-512-6_9b
"""
fire.Fire(main)
if __name__ == "__main__":
entrypoint_main()