diff --git a/README.md b/README.md index dbc8c59c..c9834558 100644 --- a/README.md +++ b/README.md @@ -60,7 +60,9 @@ pip3 install torch torchvision torchaudio --extra-index-url https://download.pyt conda install pytorch torchvision torchaudio git -c pytorch ``` -See also: [Installation instructions for human beings](https://github.com/oobabooga/text-generation-webui/wiki/Installation-instructions-for-human-beings). +> **Note** +> 1. If you are on Windows, it may be easier to run the commands above in a WSL environment. The performance may also be better. +> 2. For a more detailed, user-contributed guide, see: [Installation instructions for human beings](https://github.com/oobabooga/text-generation-webui/wiki/Installation-instructions-for-human-beings). ## Installation option 2: one-click installers @@ -140,8 +142,9 @@ Optionally, you can use the following command-line flags: | `--cai-chat` | Launch the web UI in chat mode with a style similar to Character.AI's. If the file `img_bot.png` or `img_bot.jpg` exists in the same folder as server.py, this image will be used as the bot's profile picture. Similarly, `img_me.png` or `img_me.jpg` will be used as your profile picture. | | `--cpu` | Use the CPU to generate text.| | `--load-in-8bit` | Load the model with 8-bit precision.| -| `--load-in-4bit` | Load the model with 4-bit precision. Currently only works with LLaMA.| -| `--gptq-bits GPTQ_BITS` | Load a pre-quantized model with specified precision. 2, 3, 4 and 8 (bit) are supported. Currently only works with LLaMA. | +| `--load-in-4bit` | DEPRECATED: use `--gptq-bits 4` instead. | +| `--gptq-bits GPTQ_BITS` | Load a pre-quantized model with specified precision. 2, 3, 4 and 8 (bit) are supported. Currently only works with LLaMA and OPT. | +| `--gptq-model-type MODEL_TYPE` | Model type of pre-quantized model. Currently only LLaMa and OPT are supported. | | `--bf16` | Load the model with bfloat16 precision. Requires NVIDIA Ampere GPU. | | `--auto-devices` | Automatically split the model across the available GPU(s) and CPU.| | `--disk` | If the model is too large for your GPU(s) and CPU combined, send the remaining layers to the disk. | diff --git a/modules/quantized_LLaMA.py b/modules/GPTQ_loader.py similarity index 54% rename from modules/quantized_LLaMA.py rename to modules/GPTQ_loader.py index a5757c68..c2723490 100644 --- a/modules/quantized_LLaMA.py +++ b/modules/GPTQ_loader.py @@ -7,28 +7,40 @@ import torch import modules.shared as shared sys.path.insert(0, str(Path("repositories/GPTQ-for-LLaMa"))) -from llama import load_quant +import llama +import opt -# 4-bit LLaMA -def load_quantized_LLaMA(model_name): - if shared.args.load_in_4bit: - bits = 4 +def load_quantized(model_name): + if not shared.args.gptq_model_type: + # Try to determine model type from model name + model_type = model_name.split('-')[0].lower() + if model_type not in ('llama', 'opt'): + print("Can't determine model type from model name. Please specify it manually using --gptq-model-type " + "argument") + exit() else: - bits = shared.args.gptq_bits + model_type = shared.args.gptq_model_type.lower() + + if model_type == 'llama': + load_quant = llama.load_quant + elif model_type == 'opt': + load_quant = opt.load_quant + else: + print("Unknown pre-quantized model type specified. Only 'llama' and 'opt' are supported") + exit() path_to_model = Path(f'models/{model_name}') - pt_model = '' if path_to_model.name.lower().startswith('llama-7b'): - pt_model = f'llama-7b-{bits}bit.pt' + pt_model = f'llama-7b-{shared.args.gptq_bits}bit.pt' elif path_to_model.name.lower().startswith('llama-13b'): - pt_model = f'llama-13b-{bits}bit.pt' + pt_model = f'llama-13b-{shared.args.gptq_bits}bit.pt' elif path_to_model.name.lower().startswith('llama-30b'): - pt_model = f'llama-30b-{bits}bit.pt' + pt_model = f'llama-30b-{shared.args.gptq_bits}bit.pt' elif path_to_model.name.lower().startswith('llama-65b'): - pt_model = f'llama-65b-{bits}bit.pt' + pt_model = f'llama-65b-{shared.args.gptq_bits}bit.pt' else: - pt_model = f'{model_name}-{bits}bit.pt' + pt_model = f'{model_name}-{shared.args.gptq_bits}bit.pt' # Try to find the .pt both in models/ and in the subfolder pt_path = None @@ -40,7 +52,7 @@ def load_quantized_LLaMA(model_name): print(f"Could not find {pt_model}, exiting...") exit() - model = load_quant(str(path_to_model), str(pt_path), bits) + model = load_quant(str(path_to_model), str(pt_path), shared.args.gptq_bits) # Multiple GPUs or GPU+CPU if shared.args.gpu_memory: diff --git a/modules/models.py b/modules/models.py index 7d094ed5..f4bb11fd 100644 --- a/modules/models.py +++ b/modules/models.py @@ -1,6 +1,5 @@ import json import os -import sys import time import zipfile from pathlib import Path @@ -35,6 +34,7 @@ if shared.args.deepspeed: ds_config = generate_ds_config(shared.args.bf16, 1 * world_size, shared.args.nvme_offload_dir) dschf = HfDeepSpeedConfig(ds_config) # Keep this object alive for the Transformers integration + def load_model(model_name): print(f"Loading {model_name}...") t0 = time.time() @@ -42,7 +42,7 @@ def load_model(model_name): shared.is_RWKV = model_name.lower().startswith('rwkv-') # Default settings - if not any([shared.args.cpu, shared.args.load_in_8bit, shared.args.load_in_4bit, shared.args.gptq_bits > 0, shared.args.auto_devices, shared.args.disk, shared.args.gpu_memory is not None, shared.args.cpu_memory is not None, shared.args.deepspeed, shared.args.flexgen, shared.is_RWKV]): + if not any([shared.args.cpu, shared.args.load_in_8bit, shared.args.gptq_bits, shared.args.auto_devices, shared.args.disk, shared.args.gpu_memory is not None, shared.args.cpu_memory is not None, shared.args.deepspeed, shared.args.flexgen, shared.is_RWKV]): if any(size in shared.model_name.lower() for size in ('13b', '20b', '30b')): model = AutoModelForCausalLM.from_pretrained(Path(f"models/{shared.model_name}"), device_map='auto', load_in_8bit=True) else: @@ -87,11 +87,11 @@ def load_model(model_name): return model, tokenizer - # 4-bit LLaMA - elif shared.args.gptq_bits > 0 or shared.args.load_in_4bit: - from modules.quantized_LLaMA import load_quantized_LLaMA + # Quantized model + elif shared.args.gptq_bits > 0: + from modules.GPTQ_loader import load_quantized - model = load_quantized_LLaMA(model_name) + model = load_quantized(model_name) # Custom else: diff --git a/modules/shared.py b/modules/shared.py index 5411009a..ea2eb50b 100644 --- a/modules/shared.py +++ b/modules/shared.py @@ -69,8 +69,9 @@ parser.add_argument('--chat', action='store_true', help='Launch the web UI in ch parser.add_argument('--cai-chat', action='store_true', help='Launch the web UI in chat mode with a style similar to Character.AI\'s. If the file img_bot.png or img_bot.jpg exists in the same folder as server.py, this image will be used as the bot\'s profile picture. Similarly, img_me.png or img_me.jpg will be used as your profile picture.') parser.add_argument('--cpu', action='store_true', help='Use the CPU to generate text.') parser.add_argument('--load-in-8bit', action='store_true', help='Load the model with 8-bit precision.') -parser.add_argument('--load-in-4bit', action='store_true', help='Load the model with 4-bit precision. Currently only works with LLaMA.') -parser.add_argument('--gptq-bits', type=int, default=0, help='Load a pre-quantized model with specified precision. 2, 3, 4 and 8bit are supported. Currently only works with LLaMA.') +parser.add_argument('--load-in-4bit', action='store_true', help='DEPRECATED: use --gptq-bits 4 instead.') +parser.add_argument('--gptq-bits', type=int, default=0, help='Load a pre-quantized model with specified precision. 2, 3, 4 and 8bit are supported. Currently only works with LLaMA and OPT.') +parser.add_argument('--gptq-model-type', type=str, help='Model type of pre-quantized model. Currently only LLaMa and OPT are supported.') parser.add_argument('--bf16', action='store_true', help='Load the model with bfloat16 precision. Requires NVIDIA Ampere GPU.') parser.add_argument('--auto-devices', action='store_true', help='Automatically split the model across the available GPU(s) and CPU.') parser.add_argument('--disk', action='store_true', help='If the model is too large for your GPU(s) and CPU combined, send the remaining layers to the disk.') @@ -95,3 +96,8 @@ parser.add_argument('--share', action='store_true', help='Create a public URL. T parser.add_argument('--auto-launch', action='store_true', default=False, help='Open the web UI in the default browser upon launch.') parser.add_argument('--verbose', action='store_true', help='Print the prompts to the terminal.') args = parser.parse_args() + +# Provisional, this will be deleted later +if args.load_in_4bit: + print("Warning: --load-in-4bit is deprecated and will be removed. Use --gptq-bits 4 instead.\n") + args.gptq_bits = 4 diff --git a/modules/text_generation.py b/modules/text_generation.py index d64481b2..70a51d91 100644 --- a/modules/text_generation.py +++ b/modules/text_generation.py @@ -122,7 +122,7 @@ def generate_reply(question, max_new_tokens, do_sample, temperature, top_p, typi input_ids = encode(question, max_new_tokens) original_input_ids = input_ids output = input_ids[0] - cuda = "" if (shared.args.cpu or shared.args.deepspeed or shared.args.flexgen) else ".cuda()" + cuda = not any((shared.args.cpu, shared.args.deepspeed, shared.args.flexgen)) eos_token_ids = [shared.tokenizer.eos_token_id] if shared.tokenizer.eos_token_id is not None else [] if eos_token is not None: eos_token_ids.append(int(encode(eos_token)[0][-1])) @@ -132,45 +132,48 @@ def generate_reply(question, max_new_tokens, do_sample, temperature, top_p, typi t = encode(stopping_string, 0, add_special_tokens=False) stopping_criteria_list.append(_SentinelTokenStoppingCriteria(sentinel_token_ids=t, starting_idx=len(input_ids[0]))) + generate_params = {} if not shared.args.flexgen: - generate_params = [ - f"max_new_tokens=max_new_tokens", - f"eos_token_id={eos_token_ids}", - f"stopping_criteria=stopping_criteria_list", - f"do_sample={do_sample}", - f"temperature={temperature}", - f"top_p={top_p}", - f"typical_p={typical_p}", - f"repetition_penalty={repetition_penalty}", - f"top_k={top_k}", - f"min_length={min_length if shared.args.no_stream else 0}", - f"no_repeat_ngram_size={no_repeat_ngram_size}", - f"num_beams={num_beams}", - f"penalty_alpha={penalty_alpha}", - f"length_penalty={length_penalty}", - f"early_stopping={early_stopping}", - ] + generate_params.update({ + "max_new_tokens": max_new_tokens, + "eos_token_id": eos_token_ids, + "stopping_criteria": stopping_criteria_list, + "do_sample": do_sample, + "temperature": temperature, + "top_p": top_p, + "typical_p": typical_p, + "repetition_penalty": repetition_penalty, + "top_k": top_k, + "min_length": min_length if shared.args.no_stream else 0, + "no_repeat_ngram_size": no_repeat_ngram_size, + "num_beams": num_beams, + "penalty_alpha": penalty_alpha, + "length_penalty": length_penalty, + "early_stopping": early_stopping, + }) else: - generate_params = [ - f"max_new_tokens={max_new_tokens if shared.args.no_stream else 8}", - f"do_sample={do_sample}", - f"temperature={temperature}", - f"stop={eos_token_ids[-1]}", - ] + generate_params.update({ + "max_new_tokens": max_new_tokens if shared.args.no_stream else 8, + "do_sample": do_sample, + "temperature": temperature, + "stop": eos_token_ids[-1], + }) if shared.args.deepspeed: - generate_params.append("synced_gpus=True") + generate_params.update({"synced_gpus": True}) if shared.soft_prompt: inputs_embeds, filler_input_ids = generate_softprompt_input_tensors(input_ids) - generate_params.insert(0, "inputs_embeds=inputs_embeds") - generate_params.insert(0, "inputs=filler_input_ids") + generate_params.update({"inputs_embeds": inputs_embeds}) + generate_params.update({"inputs": filler_input_ids}) else: - generate_params.insert(0, "inputs=input_ids") + generate_params.update({"inputs": input_ids}) try: # Generate the entire reply at once. if shared.args.no_stream: with torch.no_grad(): - output = eval(f"shared.model.generate({', '.join(generate_params)}){cuda}")[0] + output = shared.model.generate(**generate_params)[0] + if cuda: + output = output.cuda() if shared.soft_prompt: output = torch.cat((input_ids[0], output[filler_input_ids.shape[1]:])) @@ -194,7 +197,7 @@ def generate_reply(question, max_new_tokens, do_sample, temperature, top_p, typi return Iteratorize(generate_with_callback, kwargs, callback=None) yield formatted_outputs(original_question, shared.model_name) - with eval(f"generate_with_streaming({', '.join(generate_params)})") as generator: + with generate_with_streaming(**generate_params) as generator: for output in generator: if shared.soft_prompt: output = torch.cat((input_ids[0], output[filler_input_ids.shape[1]:])) @@ -214,7 +217,7 @@ def generate_reply(question, max_new_tokens, do_sample, temperature, top_p, typi for i in range(max_new_tokens//8+1): clear_torch_cache() with torch.no_grad(): - output = eval(f"shared.model.generate({', '.join(generate_params)})")[0] + output = shared.model.generate(**generate_params)[0] if shared.soft_prompt: output = torch.cat((input_ids[0], output[filler_input_ids.shape[1]:])) reply = decode(output) diff --git a/server.py b/server.py index 08b1a478..a54e3b62 100644 --- a/server.py +++ b/server.py @@ -269,7 +269,7 @@ if shared.args.chat or shared.args.cai_chat: function_call = 'chat.cai_chatbot_wrapper' if shared.args.cai_chat else 'chat.chatbot_wrapper' - gen_events.append(shared.gradio['Generate'].click(eval(function_call), shared.input_params, shared.gradio['display'], show_progress=shared.args.no_stream, api_name='textgen')) + gen_events.append(shared.gradio['Generate'].click(eval(function_call), shared.input_params, shared.gradio['display'], show_progress=shared.args.no_stream)) gen_events.append(shared.gradio['textbox'].submit(eval(function_call), shared.input_params, shared.gradio['display'], show_progress=shared.args.no_stream)) gen_events.append(shared.gradio['Regenerate'].click(chat.regenerate_wrapper, shared.input_params, shared.gradio['display'], show_progress=shared.args.no_stream)) gen_events.append(shared.gradio['Impersonate'].click(chat.impersonate_wrapper, shared.input_params, shared.gradio['textbox'], show_progress=shared.args.no_stream))