from h2o_wave import main, app, Q, ui, data from gradio_client import Client import ast async def init_ui(q: Q) -> None: q.page['meta'] = ui.meta_card( box='', layouts=[ ui.layout(breakpoint='xs', min_height='100vh', zones=[ ui.zone('main', size='1', direction=ui.ZoneDirection.ROW, zones=[ ui.zone('sidebar', size='250px'), ui.zone('body', direction=ui.ZoneDirection.COLUMN, zones=[ ui.zone('title', size='55px'), ui.zone('content', size='1'), ui.zone('footer'), ]), ]) ]) ], title='AchyuthGPT', ) q.page['sidebar'] = ui.nav_card( box='sidebar', color='primary', title='AchyuthGPT', subtitle='Programmed by Achyuth', value=f"#{q.args['#']}' if q.args['#'] else '#page1", image='https://huggingface.co/spaces/AchyuthGamer/AchyuthGPT-v1/resolve/main/achyuthailogo.png', items=[ ui.nav_group('', items=[ ui.nav_item(name='dwave-docs', label='Wave docs', path='https://AchyuthGPT.blogspot.com/'), ui.nav_item(name='Achyuth-GPT', label='Achyuth GPT', path='https://github.com/achyuth4/AchyuthGPT-llmstudio'), ui.nav_item(name='fine-tune', label='LLM Studio', path='https://github.com/achyuth4/AchyuthGPT-llmstudio'), ui.nav_item(name='more-models', label='More models', path='https://huggingface.co/achyuthgamer'), ]), ], secondary_items=[ ui.toggle(name='dark_mode', label='Dark mode', trigger=True), ui.text('
Developer - N.Achyuth Reddy.
') ] ) q.page['chatbot'] = ui.chatbot_card( box=ui.box('content'), data=data('content from_user', t='list'), name='chatbot' ) q.page['title'] = ui.section_card( box='title', title='', subtitle='', items=[ ui.dropdown(name='model', trigger=True, label='', value='gpt', choices=[ ui.choice(name='gpt', label='GPT Model-1'), ui.choice(name='falcon', label='GPT Model-2'), ui.choice(name='mpt', label='GPT Model-3'), ]), ui.button(name='clear', label='Clear', icon='Delete'), ], ) """ :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 :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 langchain_action: Mode langchain operations in on documents. Query: Make query of document(s) Summarize or Summarize_map_reduce: Summarize document(s) via map_reduce Summarize_all: Summarize document(s) using entire document at once Summarize_refine: Summarize document(s) using entire document, and try to refine before returning summary :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 visible_langchain_actions: Which actions to allow :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: """ @app('/') async def serve(q: Q): if not q.client.initialized: await init_ui(q) q.client.model_client = Client('https://gpt.h2o.ai/') q.client.initialized = True # A new message arrived. if q.args.chatbot: # Append user message. q.page['chatbot'].data += [q.args.chatbot, True] # Append bot response. kwargs = dict(instruction_nochat=q.args.chatbot) try: res = q.client.model_client.predict(str(dict(kwargs)), api_name='/submit_nochat_api') bot_res = ast.literal_eval(res)['response'] q.page['chatbot'].data += [bot_res, False] except: q.page['meta'] = ui.meta_card(box='', notification_bar=ui.notification_bar( text='An error occurred during prediction. Please try later or a different model.', type='error', )) elif q.args.clear: # Recreate the card. q.page['chatbot'] = ui.chatbot_card( box=ui.box('content'), data=data('content from_user', t='list'), name='chatbot' ) elif q.args.dark_mode is not None: q.page['meta'].theme = 'achyuthgpt-dark' if q.args.dark_mode else 'light' q.page['sidebar'].color = 'card' if q.args.dark_mode else 'primary' elif q.args.model: try: q.client.model_client = Client(f'https://{q.args.model}.h2o.ai/') q.page['meta'] = ui.meta_card(box='', notification_bar=ui.notification_bar( text='Model changed successfully.', type='success', )) except: q.page['meta'] = ui.meta_card(box='', notification_bar=ui.notification_bar( text='An error occurred while changing the model. Please try a different one.', type='error', )) await q.page.save()