# A mirror to gradio launch stream # By Xuan Phi Nguyen at DAMO Academy, Alibaba Group """ Load FasterLlama with original VLLM codebase, require changing config names to LlamaForCausalLM tensor_parallel must == 1 """ import torch import os import numpy as np import argparse from vllm import LLM, SamplingParams import gradio as gr from gradio_client.documentation import document, set_documentation_group from typing import List, Optional, Union, Dict, Tuple from tqdm import tqdm from transformers import PreTrainedTokenizer, PreTrainedTokenizerFast from vllm.engine.arg_utils import EngineArgs from vllm.engine.llm_engine import LLMEngine from vllm.outputs import RequestOutput from vllm.sampling_params import SamplingParams from vllm.utils import Counter from vllm.sequence import (Sequence, SequenceData, SequenceGroup, SequenceGroupMetadata, SequenceOutputs, SequenceStatus) # ! reconfigure vllm to faster llama from typing import Any, Iterator from typing import Iterator, List, Optional, Tuple import filelock import glob import json import os from huggingface_hub import snapshot_download from tqdm.auto import tqdm from vllm.model_executor.model_loader import _MODEL_REGISTRY from vllm.model_executor.models import LlamaForCausalLM _MODEL_REGISTRY['FasterLlamaForCausalLM'] = LlamaForCausalLM def hf_model_weights_iterator( model_name_or_path: str, cache_dir: Optional[str] = None, use_np_cache: bool = False, ) -> Iterator[Tuple[str, torch.Tensor]]: from vllm.model_executor.weight_utils import Disabledtqdm # Prepare file lock directory to prevent multiple processes from # downloading the same model weights at the same time. lock_dir = cache_dir if cache_dir is not None else "/tmp" lock_file_name = model_name_or_path.replace("/", "-") + ".lock" lock = filelock.FileLock(os.path.join(lock_dir, lock_file_name)) # Download model weights from huggingface. is_local = os.path.isdir(model_name_or_path) if not is_local: with lock: hf_folder = snapshot_download(model_name_or_path, allow_patterns="*.bin", cache_dir=cache_dir, local_files_only=True, tqdm_class=Disabledtqdm) else: hf_folder = model_name_or_path hf_bin_files = [ # x for x in glob.glob(os.path.join(hf_folder, "*.bin")) x for x in glob.glob(os.path.join(hf_folder, "*model*.bin")) if not x.endswith("training_args.bin") ] hf_safetensors_files = [ x for x in glob.glob(os.path.join(hf_folder, "*model*.safetensors")) if not x.endswith("training_args.bin") ] # print(F'Load bin files: {hf_bin_files} // safetensors: {hf_safetensors_files}') if use_np_cache: # Convert the model weights from torch tensors to numpy arrays for # faster loading. np_folder = os.path.join(hf_folder, "np") os.makedirs(np_folder, exist_ok=True) weight_names_file = os.path.join(np_folder, "weight_names.json") with lock: if not os.path.exists(weight_names_file): weight_names = [] for bin_file in hf_bin_files: state = torch.load(bin_file, map_location="cpu") for name, param in state.items(): param_path = os.path.join(np_folder, name) with open(param_path, "wb") as f: np.save(f, param.cpu().detach().numpy()) weight_names.append(name) with open(weight_names_file, "w") as f: json.dump(weight_names, f) with open(weight_names_file, "r") as f: weight_names = json.load(f) for name in weight_names: param_path = os.path.join(np_folder, name) with open(param_path, "rb") as f: param = np.load(f) yield name, torch.from_numpy(param) else: if len(hf_bin_files) > 0: print(F'Load bin files: {hf_bin_files}') for bin_file in hf_bin_files: state = torch.load(bin_file, map_location="cpu") for name, param in state.items(): yield name, param del state torch.cuda.empty_cache() elif len(hf_safetensors_files) > 0: print(F'Load safetensor files: {hf_safetensors_files}') from safetensors.torch import load_file for safe_file in hf_safetensors_files: # state = torch.load(bin_file, map_location="cpu") state = load_file(safe_file) for name, param in state.items(): yield name, param del state torch.cuda.empty_cache() else: raise ValueError(f'no files available either bin or safe') def convert_pyslice_to_tensor(x: Any) -> torch.Tensor: """convert PySafeSlice object from safetensors to torch.Tensor PySafeSlice object supports indexing, which is done before loading the actual tensor and can reduce the amount of memory being read into the memory. However, it does not support more advanced functionalities like `.view()` or `.t()`. Therefore, if we need to modify the loaded tensor with these more complicated operators, we need to convert to tensor first. """ if not isinstance(x, torch.Tensor): x = x[:] return x def load_padded_tensor_parallel_vocab( param: torch.Tensor, loaded_weight: Any, # `torch.Tensor` or `PySafeSlice` tensor_model_parallel_rank: int, ) -> None: shard_size = param.shape[0] start_idx = tensor_model_parallel_rank * shard_size end_idx = (tensor_model_parallel_rank + 1) * shard_size loaded_weight = loaded_weight[start_idx:end_idx] loaded_weight = convert_pyslice_to_tensor(loaded_weight) param[:loaded_weight.shape[0]].copy_(loaded_weight) def llama_load_weights( self, model_name_or_path: str, cache_dir: Optional[str] = None, use_np_cache: bool = False, load_format: str = "auto", # load_format: str = "pt", revision: Optional[str] = None ): from vllm.model_executor.weight_utils import ( load_tensor_parallel_weights ) from vllm.model_executor.parallel_utils.parallel_state import ( get_tensor_model_parallel_rank, get_tensor_model_parallel_world_size) tp_size = get_tensor_model_parallel_world_size() tensor_model_parallel_rank = get_tensor_model_parallel_rank() q_proj_shard_size = (self.config.hidden_size // tp_size) kv_proj_shard_size = (self.config.hidden_size // self.config.num_attention_heads * getattr(self.config, "num_key_value_heads", self.config.num_attention_heads) // tp_size) attention_weight_specs = [ # (weight_name, shard_size, offset) ("q_proj", q_proj_shard_size, 0), ("k_proj", kv_proj_shard_size, q_proj_shard_size), ("v_proj", kv_proj_shard_size, q_proj_shard_size + kv_proj_shard_size), ] state_dict = self.state_dict() need_to_load = len(state_dict) loaded = 0 # try: # iterator = hf_model_weights_iterator(model_name_or_path, cache_dir, use_np_cache) # except Exception as e: # iterator = hf_model_weights_iterator(model_name_or_path, cache_dir, load_format, revision) iterator = hf_model_weights_iterator(model_name_or_path, cache_dir, use_np_cache) # for name, loaded_weight in hf_model_weights_iterator( # model_name_or_path, cache_dir, load_format, revision): # model_name_or_path, cache_dir, use_np_cache): for name, loaded_weight in iterator: if "rotary_emb.inv_freq" in name: continue # if "embed_tokens" in name or "lm_head" in name: # param = state_dict[name] # # Consider padding in the vocab size. # padded_vocab_size = (param.shape[0] * tp_size) # # num_extra_rows = padded_vocab_size - self.config.vocab_size # num_extra_rows = padded_vocab_size - loaded_weight.size(0) # load_size = loaded_weight.size() # extra_rows = torch.empty(num_extra_rows, # loaded_weight.shape[1]) # extra_rows = extra_rows.to(loaded_weight) # loaded_weight = torch.cat([loaded_weight, extra_rows], dim=0) # if num_extra_rows > 0: # print(f'Add empty to {num_extra_rows} extra row for {name}') # print(f'Load: {name} | {padded_vocab_size=} | {self.config.vocab_size=} | {num_extra_rows=} | {param.size()=} | {loaded_weight.size()=} | {load_size=}') if "embed_tokens" in name or "lm_head" in name: param = state_dict[name] load_padded_tensor_parallel_vocab(param, loaded_weight, tensor_model_parallel_rank) loaded += 1 continue is_attention_weight = False for weight_name, shard_size, offset in attention_weight_specs: if weight_name not in name or "qkv_proj" in name: continue param = state_dict[name.replace(weight_name, "qkv_proj")] loaded_weight = loaded_weight[ shard_size * tensor_model_parallel_rank:shard_size * (tensor_model_parallel_rank + 1)] param_slice = param.data[offset:offset + shard_size] assert param_slice.shape == loaded_weight.shape param_slice.copy_(loaded_weight) loaded += 1.0 / 3 is_attention_weight = True break if is_attention_weight: continue # ! qkv_proj is sharded differently if concatenated into qkv # qkv: qqqq kkkk vvvv # lweight: qq0qq1 kk0kk1 vv0vv1 # q_shard_size: hidden_size // tp_size = qq # qkv_s0: qq0_kk0_vv0 # qkv_s1: qq1_kk1_vv1 if "qkv_proj" in name: param = state_dict[name] # loaded_weight qsize = self.config.hidden_size kvsize = self.config.hidden_size // self.config.num_attention_heads * getattr(self.config, "num_key_value_heads", self.config.num_attention_heads) q_offsets = ( q_proj_shard_size * tensor_model_parallel_rank, q_proj_shard_size * (tensor_model_parallel_rank + 1) ) k_offsets = ( qsize + kv_proj_shard_size * tensor_model_parallel_rank, qsize + kv_proj_shard_size * (tensor_model_parallel_rank + 1) ) v_offsets = ( qsize + kvsize + kv_proj_shard_size * tensor_model_parallel_rank, qsize + kvsize + kv_proj_shard_size * (tensor_model_parallel_rank + 1) ) _loaded_weight = torch.cat( [ loaded_weight[q_offsets[0]:q_offsets[1]], loaded_weight[k_offsets[0]:k_offsets[1]], loaded_weight[v_offsets[0]:v_offsets[1]], ], 0 ) # print(f'{name} | {q_offsets} | {k_offsets} | {v_offsets}') assert param.shape == _loaded_weight.shape, f'{param.shape=} != {_loaded_weight.shape=}' param.data.copy_(_loaded_weight) loaded += 1.0 is_attention_weight = True if is_attention_weight: continue is_gate_up_weight = False for stride_id, weight_name in enumerate(["gate_proj", "up_proj"]): if weight_name not in name or "gate_up_proj" in name: continue param = state_dict[name.replace(weight_name, "gate_up_proj")] shard_size = param.shape[0] // 2 loaded_weight = loaded_weight[ shard_size * tensor_model_parallel_rank:shard_size * (tensor_model_parallel_rank + 1)] param_slice = param.data[shard_size * stride_id:shard_size * (stride_id + 1)] assert param_slice.shape == loaded_weight.shape param_slice.copy_(loaded_weight) loaded += 1.0 / 2 is_gate_up_weight = True break if is_gate_up_weight: continue if "gate_up_proj" in name: param = state_dict[name] shard_size = param.shape[0] // 2 intermediate_size = self.config.intermediate_size g_offsets = ( shard_size * tensor_model_parallel_rank, shard_size * (tensor_model_parallel_rank + 1) ) u_offsets = ( intermediate_size + shard_size * tensor_model_parallel_rank, intermediate_size + shard_size * (tensor_model_parallel_rank + 1) ) # print(f'{name} {param.size()} | {g_offsets} | {u_offsets}') _loaded_weight = torch.cat( [ loaded_weight[g_offsets[0]:g_offsets[1]], loaded_weight[u_offsets[0]:u_offsets[1]], ], 0 ) assert param.shape == _loaded_weight.shape param.data.copy_(_loaded_weight) loaded += 1.0 is_gate_up_weight = True if is_gate_up_weight: continue param = state_dict[name] load_tensor_parallel_weights(param, loaded_weight, name, self._column_parallel_weights, self._row_parallel_weights, tensor_model_parallel_rank) loaded += 1 if np.abs(loaded - need_to_load) < 0.01: print(f'WARNING: only {loaded} params loaded out of {need_to_load}') else: print(f'Loaded all {loaded} params loaded out of {need_to_load}') # Reassign LlamaForCausalLM.load_weights with llama_load_weights LlamaForCausalLM.load_weights = llama_load_weights # ! ================================================================== set_documentation_group("component") DATA_ROOT = os.environ.get("dataroot", "/mnt/workspace/workgroup/phi") MODEL_CACHE_DIR = os.path.join(DATA_ROOT, "pret_models") DTYPES = { 'float16': torch.float16, 'bfloat16': torch.bfloat16 } llm = None demo = None RELOAD_SIGNAL = '<<>\n\n" SYSTEM_PROMPT_1 = """You are a multilingual, helpful, respectful and honest assistant. Your name is SeaL and you are built by DAMO Academy, Alibaba Group. Always answer as helpfully as possible, while being safe. Your \ answers should not include any harmful, unethical, racist, sexist, toxic, dangerous, or illegal content. Please ensure\ that your responses are socially unbiased and positive in nature. If a question does not make any sense, or is not factually coherent, explain why instead of answering something not \ correct. If you don't know the answer to a question, please don't share false information. As a multilingual assistant, you must respond and follow instructions in the native language of the user by default, unless told otherwise. \ Your response should adapt to the norms and customs of the respective language and culture. """ RES_PRINTED = False def llama_chat_sys_input_seq_constructor(text, sys_prompt=SYSTEM_PROMPT_1, bos_token=BOS_TOKEN, eos_token=EOS_TOKEN): return f"{bos_token}{B_INST} {B_SYS} {sys_prompt} {E_SYS} {text} {E_INST}" def llama_chat_multiturn_sys_input_seq_constructor( message: str, history: List[Tuple[str, str]], sys_prompt=SYSTEM_PROMPT_1, bos_token=BOS_TOKEN, eos_token=EOS_TOKEN, ): """ ``` [INST] B_SYS SytemPrompt E_SYS Prompt [/INST] Answer [INST] Prompt [/INST] Answer [INST] Prompt [/INST] ``` """ text = '' for i, (prompt, res) in enumerate(history): if i == 0: text += f"{bos_token}{B_INST} {B_SYS} {sys_prompt} {E_SYS} {prompt} {E_INST}" else: text += f"{bos_token}{B_INST} {prompt} {E_INST}" if res is not None: text += f" {res} {eos_token} " if len(history) == 0 or text.strip() == '': text = f"{bos_token}{B_INST} {B_SYS} {sys_prompt} {E_SYS} {message} {E_INST}" else: text += f"{bos_token}{B_INST} {message} {E_INST}" return text @document() class ChatBot(gr.Chatbot): def _postprocess_chat_messages( self, chat_message ): x = super()._postprocess_chat_messages(chat_message) if isinstance(x, str): x = x.replace("\n", "
") return x def load_ckpt(ckpt_file: str) -> str: global llm status = "Failed" if not os.path.exists(ckpt_file): status = f"Failed - file not found: {ckpt_file}" elif not ckpt_file.endswith(".bin"): status = f"Failed - file not .bin: {ckpt_file}" else: try: state_dict = torch.load(ckpt_file, map_location='cpu') print(f'loaded state_dict: {ckpt_file}') llm.llm_engine.workers[0].model.load_state_dict(state_dict) status = f'Success. Loaded {ckpt_file}' except Exception as e: status = f'Failed - {str(e)}' return status def chat_response(message, history, temperature: float, max_tokens: int, system_prompt: str = '') -> str: global llm assert llm is not None temperature = float(temperature) max_tokens = int(max_tokens) if system_prompt.strip() != '': # chat version, add system prompt message = llama_chat_sys_input_seq_constructor( message.strip(), sys_prompt=system_prompt ) sampling_params = SamplingParams(temperature=temperature, max_tokens=max_tokens) gen = llm.generate(message, sampling_params) out = gen[0].outputs[0].text # print(f'{message}<<<{out}>>>') return f'{out}' def vllm_abort(self: LLM): scheduler = self.llm_engine.scheduler for state_queue in [scheduler.waiting, scheduler.running, scheduler.swapped]: for seq_group in state_queue: # if seq_group.request_id == request_id: # Remove the sequence group from the state queue. state_queue.remove(seq_group) for seq in seq_group.seqs: if seq.is_finished(): continue scheduler.free_seq(seq, SequenceStatus.FINISHED_ABORTED) def _vllm_run_engine(self: LLM, use_tqdm: bool = False) -> Dict[str, RequestOutput]: # Initialize tqdm. if use_tqdm: num_requests = self.llm_engine.get_num_unfinished_requests() pbar = tqdm(total=num_requests, desc="Processed prompts") # Run the engine. outputs: Dict[str, RequestOutput] = {} while self.llm_engine.has_unfinished_requests(): step_outputs = self.llm_engine.step() for output in step_outputs: # if output.finished: # outputs.append(output) # if use_tqdm: # pbar.update(1) outputs[output.request_id] = output # outputs = sorted(outputs, key=lambda x: int(x.request_id)) if len(outputs) > 0: yield outputs # if use_tqdm: # pbar.close() # Sort the outputs by request ID. # This is necessary because some requests may be finished earlier than # its previous requests. # outputs = sorted(outputs, key=lambda x: int(x.request_id)) # return outputs def vllm_generate_stream( self: LLM, prompts: Optional[Union[str, List[str]]] = None, sampling_params: Optional[SamplingParams] = None, prompt_token_ids: Optional[List[List[int]]] = None, use_tqdm: bool = False, ) -> Dict[str, RequestOutput]: """Generates the completions for the input prompts. NOTE: This class automatically batches the given prompts, considering the memory constraint. For the best performance, put all of your prompts into a single list and pass it to this method. Args: prompts: A list of prompts to generate completions for. sampling_params: The sampling parameters for text generation. If None, we use the default sampling parameters. prompt_token_ids: A list of token IDs for the prompts. If None, we use the tokenizer to convert the prompts to token IDs. use_tqdm: Whether to use tqdm to display the progress bar. Returns: A list of `RequestOutput` objects containing the generated completions in the same order as the input prompts. """ if prompts is None and prompt_token_ids is None: raise ValueError("Either prompts or prompt_token_ids must be " "provided.") if isinstance(prompts, str): # Convert a single prompt to a list. prompts = [prompts] if prompts is not None and prompt_token_ids is not None: if len(prompts) != len(prompt_token_ids): raise ValueError("The lengths of prompts and prompt_token_ids " "must be the same.") if sampling_params is None: # Use default sampling params. sampling_params = SamplingParams() # Add requests to the engine. if prompts is not None: num_requests = len(prompts) else: num_requests = len(prompt_token_ids) for i in range(num_requests): prompt = prompts[i] if prompts is not None else None if prompt_token_ids is None: token_ids = None else: token_ids = prompt_token_ids[i] self._add_request(prompt, sampling_params, token_ids) # return self._run_engine(use_tqdm) yield from _vllm_run_engine(self, use_tqdm) def chat_response_stream( message: str, history: List[Tuple[str, str]], temperature: float, max_tokens: int, frequency_penalty: float, system_prompt: str ) -> str: global llm, RES_PRINTED assert llm is not None # force removing all vllm_abort(llm) temperature = float(temperature) frequency_penalty = float(frequency_penalty) max_tokens = int(max_tokens) if system_prompt.strip() != '': # chat version, add system prompt message = llama_chat_sys_input_seq_constructor( message.strip(), sys_prompt=system_prompt ) sampling_params = SamplingParams( temperature=temperature, max_tokens=max_tokens, frequency_penalty=frequency_penalty, ) cur_out = None for gen in vllm_generate_stream(llm, message, sampling_params): if cur_out is not None: yield cur_out assert len(gen) == 1, f'{gen}' item = next(iter(gen.values())) cur_out = item.outputs[0].text if not RES_PRINTED: print(f'{message}<<<{cur_out}>>>') RES_PRINTED = True if cur_out is not None: yield cur_out def chat_response_stream_multiturn( message: str, history: List[Tuple[str, str]], temperature: float, max_tokens: int, frequency_penalty: float, system_prompt: str ) -> str: """Build multi turn [INST] B_SYS SytemPrompt E_SYS Prompt [/INST] Answer [INST] Prompt [/INST] Answer [INST] Prompt [/INST] message is incoming prompt history don't have the current messauge """ global llm, RES_PRINTED assert llm is not None assert system_prompt.strip() != '', f'system prompt is empty' # force removing all vllm_abort(llm) temperature = float(temperature) frequency_penalty = float(frequency_penalty) max_tokens = int(max_tokens) # history.append([message, None]) # history will be appended with message later on full_prompt = llama_chat_multiturn_sys_input_seq_constructor( message, history, sys_prompt=system_prompt ) sampling_params = SamplingParams( temperature=temperature, max_tokens=max_tokens, frequency_penalty=frequency_penalty, ) cur_out = None for gen in vllm_generate_stream(llm, full_prompt, sampling_params): if cur_out is not None: yield cur_out assert len(gen) == 1, f'{gen}' item = next(iter(gen.values())) cur_out = item.outputs[0].text if not RES_PRINTED: print(f'{full_prompt}<<<{cur_out}>>>') RES_PRINTED = True if cur_out is not None: yield cur_out def debug_chat_response_echo( message: str, history: List[Tuple[str, str]], temperature: float = 0.0, max_tokens: int = 4096, frequency_penalty: float = 0.4, system_prompt: str = SYSTEM_PROMPT_1, ) -> str: yield message MODEL_TITLE = "DAMO-SeaL-13B - An Assistant for South East Asian Languages" MODEL_DESC = """ This is a 13B DAMO-SeaL-Chat assistant model built by DAMO Academy, Alibaba Group. It can produce helpful responses in English, Vietnamese, Indonesian and Thai. """.strip() TENSOR_PARALLEL = int(os.environ.get("TENSOR_PARALLEL", "1")) DTYPE = 'bfloat16' DTYPE = 'float16' MODEL_PATH = os.environ.get("MODEL_PATH", "notfound, please set `export MODEL_PATH=`") DEBUG = 1 def launch(): global demo, llm, DEBUG if DEBUG: model_desc + "
This is in debug mode, responses will be copy original" response_fn = debug_chat_response_echo else: model_desc = MODEL_DESC model_path = MODEL_PATH assert os.path.exists(model_path), f'{model_path} not found' model_title = MODEL_TITLE tensor_parallel = TENSOR_PARALLEL assert tensor_parallel > 0 , f'{tensor_parallel} invalid' dtype = DTYPE # ! load the model llm = LLM(model=model_path, dtype=dtype, tensor_parallel_size=tensor_parallel) sys_prompt = SYSTEM_PROMPT_1 max_tokens = 4096 print(f'Use system prompt:\n{sys_prompt}') # response_fn = chat_response_stream_multiturn if args.multiturn else chat_response_stream response_fn = chat_response_stream_multiturn print(F'respond: {response_fn}') demo = gr.ChatInterface( response_fn, chatbot=ChatBot( bubble_full_width=False, latex_delimiters=[ { "left": "$", "right": "$", "display": False}, { "left": "$$", "right": "$$", "display": True}, ] ), textbox=gr.Textbox(placeholder='Type message', lines=8, max_lines=128, min_width=200), submit_btn=gr.Button(value='Submit', variant="primary", scale=0), # stop_btn=None, title=f"{model_title}", description=f"{model_desc}", # ! decide if can change the system prompt. additional_inputs=[ gr.Number(value=0, label='Temperature (higher -> more random)'), gr.Number(value=max_tokens, label='Max generated tokens (increase if want more generation)'), gr.Number(value=0.4, label='Frequency penalty (> 0 encourage new tokens)'), gr.Textbox(value=sys_prompt, label='System prompt', lines=8)], ) demo.queue() # demo.launch(server_port=args.port) demo.launch() def main(): # launch(parser.parse_args()) launch() if __name__ == "__main__": main() """ export CUDA_VISIBLE_DEVICES=0 export MODEL_PATH=${dataroot}/hf_train/pretrain_lm/swpn/merlion13s108Hi8kPretFlCW8k.LMFromHf.a.gc.t5k0.vizhthid.mean_std.TrainTask.NLNL.Multi.Vi.FSePlCq13M.FSePlCq13M.m4k.b8.lr1e5.linear.wa0k.ms858k.grac1.se1.8g.v4c.zfsdp/step_4000 export MODEL_PATH=${dataroot}/llama-2-7b-lxxp-faster export MODEL_PATH=${dataroot}/llama-2-7b-chat-xp python app.py """