app.py

# Copyright: DAMO Academy, Alibaba Group
# By Xuan Phi Nguyen at DAMO Academy, Alibaba Group

# Description:
"""
VLLM-based demo script to launch Language chat model for South East Asian Languages
"""


import os
import numpy as np
import argparse
import torch
import gradio as gr
from typing import Any, Iterator
from typing import Iterator, List, Optional, Tuple
import filelock
import glob
import json

from gradio_client.documentation import document, set_documentation_group

from typing import List, Optional, Union, Dict, Tuple
from tqdm.auto import tqdm
from huggingface_hub import snapshot_download


# @@ environments ================

DEBUG = bool(int(os.environ.get("DEBUG", "1")))
BLOCK_ZH = bool(int(os.environ.get("BLOCK_ZH", "1")))
TENSOR_PARALLEL = int(os.environ.get("TENSOR_PARALLEL", "1"))
DTYPE = os.environ.get("DTYPE", "bfloat16")

# ! (no debug) whether to download HF_MODEL_NAME and save to MODEL_PATH
DOWNLOAD_SNAPSHOT = bool(int(os.environ.get("DOWNLOAD_SNAPSHOT", "0")))
LOG_RESPONSE = bool(int(os.environ.get("LOG_RESPONSE", "0")))

# ! uploaded model path, will be downloaded to MODEL_PATH
HF_MODEL_NAME = os.environ.get("HF_MODEL_NAME", "DAMO-NLP-SG/seal-13b-chat-a")
# ! if model is private, need HF_TOKEN to access the model
HF_TOKEN = os.environ.get("HF_TOKEN", None)
# ! path where the model is downloaded, either on ./ or persistent disc
MODEL_PATH = os.environ.get("MODEL_PATH", "./seal-13b-chat-a")

# ! list of keywords to disabled as security measures to comply with local regulation
KEYWORDS = os.environ.get("KEYWORDS", "").strip()
KEYWORDS = KEYWORDS.split(";") if len(KEYWORDS) > 0 else []
KEYWORDS = [x.lower() for x in KEYWORDS]

# gradio config
PORT = int(os.environ.get("PORT", "7860"))
# how many iterations to yield response
STREAM_YIELD_MULTIPLE = int(os.environ.get("STREAM_YIELD_MULTIPLE", "1"))
# how many iterations to perform safety check on response
STREAM_CHECK_MULTIPLE = int(os.environ.get("STREAM_CHECK_MULTIPLE", "0"))

# self explanatory
MAX_TOKENS = int(os.environ.get("MAX_TOKENS", "2048"))
TEMPERATURE = float(os.environ.get("TEMPERATURE", "0.1"))
FREQUENCE_PENALTY = float(os.environ.get("FREQUENCE_PENALTY", "0.4"))
gpu_memory_utilization = float(os.environ.get("gpu_memory_utilization", "0.9"))

# whether to enable quantization, currently not in use
QUANTIZATION = str(os.environ.get("QUANTIZATION", ""))


"""
Internal instructions of how to configure the DEMO

1. Upload SFT model as a model to huggingface: hugginface/models/seal_13b_a
2. If the model weights is private, set HF_TOKEN=<your private hf token> in https://huggingface.co/spaces/????/?????/settings
3. space config env: `HF_MODEL_NAME=DAMO-NLP-SG/seal-13b-chat-a` or the underlining model
4. If enable persistent storage: set
HF_HOME=/data/.huggingface
MODEL_PATH=/data/.huggingface/seal-13b-chat-a
if not:
MODEL_PATH=./seal-13b-chat-a

"""



# ==============================
print(f'DEBUG mode: {DEBUG}')
print(f'Torch version: {torch.__version__}')
try:
    print(f'Torch CUDA version: {torch.version.cuda}')
except Exception as e:
    print(f'Failed to print cuda version: {e}')

try:
    compute_capability = torch.cuda.get_device_capability()
    print(f'Torch CUDA compute_capability: {compute_capability}')
except Exception as e:
    print(f'Failed to print compute_capability version: {e}')


# @@ constants ================

DTYPES = {
    'float16': torch.float16,
    'bfloat16': torch.bfloat16
}

llm = None
demo = None


BOS_TOKEN = '<s>'
EOS_TOKEN = '</s>'

B_INST, E_INST = "[INST]", "[/INST]"
B_SYS, E_SYS = "<<SYS>>\n", "\n<</SYS>>\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.
"""

# ============ CONSTANT ============
# https://github.com/gradio-app/gradio/issues/884
MODEL_NAME = "SeaLLM-13B"
MODEL_TITLE = "SeaLLM-13B - An Assistant for South East Asian Languages"
# ! add icon: "<img  src='file/lion.jpg' alt='image One'>"
MODEL_TITLE = """
<div class="container" style="
    align-items: center;
    justify-content: center;
    display: flex;
">
    <div class="image" >
        <img src="file/seal_logo.png" style="
            max-width: 10em;
            max-height: 5%;
            height: 5em;
            width: 5em;
            float: left;
            margin-left: auto;
        ">
      </div>
      <div class="text" style="
            padding-left: 20px;
            padding-top: 2%;
            float: left;
        ">
      <h1>SeaLLM-13B - An Assistant for South East Asian Languages</h1>
      </div>
</div>
"""
MODEL_DESC = """
<span style="font-size: larger">
This is SeaLLM-13B - a chatbot assistant optimized for South East Asian Languages. It can produce helpful responses in English 🇬🇧, Vietnamese 🇻🇳, Indonesian 🇮🇩 and Thai 🇹🇭.
</span>
<br>
<span style="color: red">NOTICE: The chatbot may produce inaccurate and harmful information about people, places, or facts. \
We strongly advise against misuse of the chatbot to knowingly generate harmful or unethical content, \
or content that violates locally applicable and international laws or regulations, including hate speech, violence, pornography, deception, etc!</span>
""".strip()


cite_markdown = """
## Citation
If you find our project useful, hope you can star our repo and cite our paper as follows:
```
@article{damonlpsg2023seallm,
  author = {???},
  title = {SeaLLM: A language model for South East Asian Languages},
  year = 2023,
}
```
"""

# warning_markdown = """
# ## Warning:
# <span style="color: red">The chatbot may produce inaccurate and harmful information about people, places, or facts.</span>
# <span style="color: red">We strongly advise against misuse of the chatbot to knowingly generate harmful or unethical content, \
# or content that violates locally applicable and international laws or regulations, including hate speech, violence, pornography, deception, etc!</span>
# """

path_markdown = """
#### Model path:
{model_path}
"""


def _detect_lang(text):
    from langdetect import detect as detect_lang
    dlang = None
    try:
        dlang = detect_lang(text)
    except Exception as e:
        # No features in text.
        print(f'Error: {e}')
        if "No features in text." in str(e):
            return "en"
        else:
            return "zh"
    return dlang


def custom_hf_model_weights_iterator(
    model_name_or_path: str,
    cache_dir: Optional[str] = None,
    use_np_cache: bool = False,
) -> Iterator[Tuple[str, torch.Tensor]]:
    # ! if use vllm==0.1.4, use this to augment hf_model_weights_iterator loader
    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, "*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")
    ]

    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",
        revision: Optional[str] = None
):
    # if use vllm==0.1.4
    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
    iterator = custom_hf_model_weights_iterator(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=}')

        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
            )
            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)
            )
            _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}')


def new_llama_load_weights(
    self,
    model_name_or_path: str,
    cache_dir: Optional[str] = None,
    load_format: str = "auto",
    revision: Optional[str] = None
):
    # If use newest vllm
    from vllm.model_executor.weight_utils import (
        load_tensor_parallel_weights, hf_model_weights_iterator
    )
    from vllm.model_executor.parallel_utils.parallel_state import (
        get_tensor_model_parallel_rank, get_tensor_model_parallel_world_size)
    
    if self.quant_config is None:
        weight_suffixes = ["weight"]
    else:
        weight_suffixes = self.quant_config.get_tp_tensor_names()

    column_parallel_weights: List[str] = []
    for layer in self._column_parallel_layers:
        for suffix in weight_suffixes:
            column_parallel_weights.append(f"{layer}.{suffix}")
    row_parallel_weights: List[str] = []
    for layer in self._row_parallel_layers:
        for suffix in weight_suffixes:
            row_parallel_weights.append(f"{layer}.{suffix}")

    tp_size = get_tensor_model_parallel_world_size()
    tp_rank = get_tensor_model_parallel_rank()
    assert tp_size == 1, f'tensorparallel >=2 not allowed. {tp_size}'
    q_proj_shard_size = (self.config.hidden_size // tp_size)
    num_kv_heads_replicas = max(1,
                                tp_size // self.config.num_key_value_heads)
    num_kv_heads_per_gpu = max(1,
                                self.config.num_key_value_heads // tp_size)
    kv_proj_shard_size = (self.config.hidden_size //
                            self.config.num_attention_heads *
                            num_kv_heads_per_gpu)
    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

    for name, loaded_weight in hf_model_weights_iterator(
            model_name_or_path, cache_dir, load_format, revision):
        if "rotary_emb.inv_freq" in name:
            continue

        is_packed = False
        is_transposed = False
        if self.quant_config is not None:
            is_packed = self.quant_config.is_packed(name)
            is_transposed = self.quant_config.is_transposed(name)
        if is_transposed:
            loaded_weight = convert_pyslice_to_tensor(loaded_weight)
            loaded_weight = loaded_weight.T

        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")]
            if is_transposed:
                param = param.T

            if is_packed:
                shard_size //= self.quant_config.pack_factor
                offset //= self.quant_config.pack_factor

            if weight_name in ["k_proj", "v_proj"]:
                shard_id = tp_rank // num_kv_heads_replicas
            else:
                shard_id = tp_rank
            loaded_weight = loaded_weight[shard_size *
                                            shard_id:shard_size *
                                            (shard_id + 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
        
        # TODO: need to figure out to do sharding with qkv_proj fused

        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")]
            if is_transposed:
                param = param.T

            shard_size = param.shape[0] // 2
            loaded_weight = loaded_weight[shard_size * tp_rank:shard_size *
                                            (tp_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

        # TODO: need to figure out to do sharding with gate_up_proj fused

        param = state_dict[name]
        if is_transposed:
            param = param.T

        if "embed_tokens" in name or "lm_head" in name:
            load_padded_tensor_parallel_vocab(param, loaded_weight,
                                                tp_rank)
            loaded += 1
            continue

        load_tensor_parallel_weights(param, loaded_weight, name,
                                        column_parallel_weights,
                                        row_parallel_weights, tp_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
if not DEBUG:
    
    try:
        import vllm
        from vllm.model_executor.model_loader import _MODEL_REGISTRY
        from vllm.model_executor.models import LlamaForCausalLM

        _MODEL_REGISTRY['FasterLlamaForCausalLM'] = LlamaForCausalLM
        if vllm.__version__ == "0.1.4":
            LlamaForCausalLM.load_weights = llama_load_weights
        else:
            LlamaForCausalLM.load_weights = new_llama_load_weights

        if DTYPE == "bfloat16":
            try:
                compute_capability = torch.cuda.get_device_capability()
                if compute_capability[0] < 8:
                    gpu_name = torch.cuda.get_device_name()
                    print(
                        "Bfloat16 is only supported on GPUs with compute capability "
                        f"of at least 8.0. Your {gpu_name} GPU has compute capability "
                        f"{compute_capability[0]}.{compute_capability[1]}. --> Move to FLOAT16")
                    DTYPE = "float16"
            except Exception as e:
                print(f'Unable to obtain compute_capability: {e}')
    except Exception as e:
        print(f'Failing import and reconfigure VLLM: {str(e)}')
    

# ! ==================================================================

set_documentation_group("component")


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,
):
    """
    ```
        <bos>[INST] B_SYS SytemPrompt E_SYS Prompt [/INST] Answer <eos>
        <bos>[INST] Prompt [/INST] Answer <eos>
        <bos>[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.strip().replace("\n", "<br>")
        return x


from gradio.components import Button
from gradio.events import Dependency, EventListenerMethod

# replace events so that submit button is disabled during generation, if stop_btn not found
# this prevent weird behavior
def _setup_stop_events(
    self, event_triggers: list[EventListenerMethod], event_to_cancel: Dependency
) -> None:
    event_triggers = event_triggers if isinstance(event_triggers, (list, tuple)) else [event_triggers]
    if self.stop_btn and self.is_generator:
        if self.submit_btn:
            for event_trigger in event_triggers:
                event_trigger(
                    lambda: (
                        Button.update(visible=False),
                        Button.update(visible=True),
                    ),
                    None,
                    [self.submit_btn, self.stop_btn],
                    api_name=False,
                    queue=False,
                )
            event_to_cancel.then(
                lambda: (Button.update(visible=True), Button.update(visible=False)),
                None,
                [self.submit_btn, self.stop_btn],
                api_name=False,
                queue=False,
            )
        else:
            for event_trigger in event_triggers:
                event_trigger(
                    lambda: Button.update(visible=True),
                    None,
                    [self.stop_btn],
                    api_name=False,
                    queue=False,
                )
            event_to_cancel.then(
                lambda: Button.update(visible=False),
                None,
                [self.stop_btn],
                api_name=False,
                queue=False,
            )
        self.stop_btn.click(
            None,
            None,
            None,
            cancels=event_to_cancel,
            api_name=False,
        )
    else:
        if self.submit_btn:
            for event_trigger in event_triggers:
                event_trigger(
                    lambda: Button.update(interactive=False),
                    None,
                    [self.submit_btn],
                    api_name=False,
                    queue=False,
                )
            event_to_cancel.then(
                lambda: Button.update(interactive=True),
                None,
                [self.submit_btn],
                api_name=False,
                queue=False,
            )
    # upon clear, cancel the submit event as well
    if self.clear_btn:
        self.clear_btn.click(
            lambda: ([], [], None, Button.update(interactive=True)),
            None,
            [self.chatbot, self.chatbot_state, self.saved_input, self.submit_btn],
            queue=False,
            api_name=False,
            cancels=event_to_cancel,
        )

# TODO: reconfigure clear button as stop and clear button
def _setup_events(self) -> None:
    has_on = False
    try:
        from gradio.events import Dependency, EventListenerMethod, on
        has_on = True
    except ImportError as ie:
        has_on = False
    submit_fn = self._stream_fn if self.is_generator else self._submit_fn


    if has_on:
        # new version
        submit_triggers = (
            [self.textbox.submit, self.submit_btn.click]
            if self.submit_btn
            else [self.textbox.submit]
        )
        submit_event = (
            on(
                submit_triggers,
                self._clear_and_save_textbox,
                [self.textbox],
                [self.textbox, self.saved_input],
                api_name=False,
                queue=False,
            )
            .then(
                self._display_input,
                [self.saved_input, self.chatbot_state],
                [self.chatbot, self.chatbot_state],
                api_name=False,
                queue=False,
            )
            .then(
                submit_fn,
                [self.saved_input, self.chatbot_state] + self.additional_inputs,
                [self.chatbot, self.chatbot_state],
                api_name=False,
            )
        )
        self._setup_stop_events(submit_triggers, submit_event)
    else:
        raise ValueError(f'Better install new gradio version than 3.44.0')

    if self.retry_btn:
        retry_event = (
            self.retry_btn.click(
                self._delete_prev_fn,
                [self.chatbot_state],
                [self.chatbot, self.saved_input, self.chatbot_state],
                api_name=False,
                queue=False,
            )
            .then(
                self._display_input,
                [self.saved_input, self.chatbot_state],
                [self.chatbot, self.chatbot_state],
                api_name=False,
                queue=False,
            )
            .then(
                submit_fn,
                [self.saved_input, self.chatbot_state] + self.additional_inputs,
                [self.chatbot, self.chatbot_state],
                api_name=False,
            )
        )
        self._setup_stop_events([self.retry_btn.click], retry_event)

    if self.undo_btn:
        self.undo_btn.click(
            self._delete_prev_fn,
            [self.chatbot_state],
            [self.chatbot, self.saved_input, self.chatbot_state],
            api_name=False,
            queue=False,
        ).then(
            lambda x: x,
            [self.saved_input],
            [self.textbox],
            api_name=False,
            queue=False,
        )

    # Reconfigure clear_btn to stop and clear text box
    # if self.clear_btn:
    #     self.clear_btn.click(
    #         lambda: ([], [], None),
    #         None,
    #         [self.chatbot, self.chatbot_state, self.saved_input],
    #         queue=False,
    #         api_name=False,
    #         cancels=submit_event,
    #     )


# replace
gr.ChatInterface._setup_stop_events = _setup_stop_events
gr.ChatInterface._setup_events = _setup_events

def chat_response(message, history, temperature: float, max_tokens: int, system_prompt: str = '') -> str:
    global llm
    assert llm is not None
    from vllm import LLM, SamplingParams
    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
    return f'{out}'


def vllm_abort(self: Any):
    from vllm.sequence import SequenceStatus
    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: Any, use_tqdm: bool = False) -> Dict[str, Any]:
    from vllm.outputs import 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:
            outputs[output.request_id] = output
        if len(outputs) > 0:
            yield outputs



def vllm_generate_stream(
    self: Any,
    prompts: Optional[Union[str, List[str]]] = None,
    sampling_params: Optional[Any] = None,
    prompt_token_ids: Optional[List[List[int]]] = None,
    use_tqdm: bool = False,
) -> Dict[str, Any]:
    """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.
    """
    from vllm import LLM, SamplingParams
    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)


BLOCK_MESSAGE = """Sorry, Chinese is not currently supported. Please clear the chat box for a new conversation.
抱歉，目前不支持中文。 请清除聊天框以进行新对话。"""

KEYWORD_BLOCK_MESSAGE = "Sorry, I cannot fulfill your request. If you have any unrelated questions, I'll be glad to help."

def block_zh(
    message: str, 
    history: List[Tuple[str, str]]
) -> str:
    if history is not None and any((BLOCK_MESSAGE in x[1].strip()) for x in history):
        return True
    elif 'zh' in _detect_lang(message):
        print(f'Detect zh: {message}')
        return True
    else:
        return False


def log_responses(history, message, response):
    pass


def safety_check(text, history=None, ) -> Optional[str]:
    """
    Despite our effort in safety tuning and red teaming, our models may still generate harmful or illegal content.
    This provides an additional security measure to enhance safety and compliance with local regulations.
    """
    if BLOCK_ZH:
        if history is not None:
            if block_zh(text, history):
                return BLOCK_MESSAGE
        else:
            if "zh" in _detect_lang(text):
                return BLOCK_MESSAGE

    if len(KEYWORDS) > 0 and any(x in text.lower() for x in KEYWORDS):
        return KEYWORD_BLOCK_MESSAGE

    return None


def chat_response_stream_multiturn(
    message: str, 
    history: List[Tuple[str, str]], 
    temperature: float, 
    max_tokens: int, 
    frequency_penalty: float,
    system_prompt: Optional[str] = SYSTEM_PROMPT_1
) -> str:
    from vllm import LLM, SamplingParams
    """Build multi turn
    <bos>[INST] B_SYS SytemPrompt E_SYS Prompt [/INST] Answer <eos>
    <bos>[INST] Prompt [/INST] Answer <eos>
    <bos>[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)

    message = message.strip()

    message_safety = safety_check(message, history=history)
    if message_safety is not None:
        yield message_safety
        return

    # 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 j, gen in enumerate(vllm_generate_stream(llm, full_prompt, sampling_params)):
        if cur_out is not None and (STREAM_YIELD_MULTIPLE < 1 or j % STREAM_YIELD_MULTIPLE == 0) and j > 0:
            # optionally check safety, and respond
            if STREAM_CHECK_MULTIPLE > 0 and j % STREAM_CHECK_MULTIPLE == 0:
                message_safety = safety_check(cur_out, history=None)
                if message_safety is not None:
                    yield message_safety
                    return

            yield cur_out
        assert len(gen) == 1, f'{gen}'
        item = next(iter(gen.values()))
        cur_out = item.outputs[0].text
        
    print(f'{full_prompt}<<<{cur_out}>>>\n\n')
    if cur_out is not None:
        yield cur_out
    
    message_safety = safety_check(cur_out, history=None)
    if message_safety is not None:
        yield message_safety
        return
    
    if LOG_RESPONSE:
        log_responses(history, message, 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:
    import time
    time.sleep(0.5)
    yield f"repeat: {message}"


def check_model_path(model_path) -> str:
    assert os.path.exists(model_path), f'{model_path} not found'
    ckpt_info = "None"
    if os.path.isdir(model_path):
        if os.path.exists(f'{model_path}/info.txt'):
            with open(f'{model_path}/info.txt', 'r') as f:
                ckpt_info = f.read()
                print(f'Checkpoint info:\n{ckpt_info}\n-----')
        else:
            print(f'info.txt not found in {model_path}')
        print(f'model path dir: {list(os.listdir(model_path))}')
    
    return ckpt_info


def launch():
    global demo, llm, DEBUG
    model_desc = MODEL_DESC
    model_path = MODEL_PATH
    model_title = MODEL_TITLE
    hf_model_name = HF_MODEL_NAME
    tensor_parallel = TENSOR_PARALLEL
    assert tensor_parallel > 0 , f'{tensor_parallel} invalid'
    dtype = DTYPE
    sys_prompt = SYSTEM_PROMPT_1
    max_tokens = MAX_TOKENS
    temperature = TEMPERATURE
    frequence_penalty = FREQUENCE_PENALTY
    ckpt_info = "None"

    print(
        f'Launch config: {model_title=} / {tensor_parallel=} / {dtype=} / {max_tokens} | {BLOCK_ZH=} '
        f'\n| STREAM_YIELD_MULTIPLE={STREAM_YIELD_MULTIPLE} '
        f'\n| STREAM_CHECK_MULTIPLE={STREAM_CHECK_MULTIPLE} '
        f'\n| frequence_penalty={frequence_penalty} '
        f'\n| temperature={temperature} '
        f'\n| hf_model_name={hf_model_name} '
        f'\n| model_path={model_path} '
        f'\n| DOWNLOAD_SNAPSHOT={DOWNLOAD_SNAPSHOT} '
        f'\n| gpu_memory_utilization={gpu_memory_utilization} '
        f'\n| KEYWORDS={KEYWORDS} '
        f'\nsys={SYSTEM_PROMPT_1}'
        f'\ndesc={model_desc}'
    )

    if DEBUG:
        model_desc += "\n<br>!!!!! This is in debug mode, responses will copy original"
        response_fn = debug_chat_response_echo
        print(f'Creating in DEBUG MODE')
    else:
        # ! load the model

        if DOWNLOAD_SNAPSHOT:
            print(f'Downloading from HF_MODEL_NAME={hf_model_name} -> {model_path}')
            if HF_TOKEN is not None:
                print(f'Load with HF_TOKEN: {HF_TOKEN}')
                snapshot_download(hf_model_name, local_dir=model_path, use_auth_token=True, token=HF_TOKEN)
            else:
                snapshot_download(hf_model_name, local_dir=model_path)

        import vllm
        from vllm import LLM

        print(F'VLLM: {vllm.__version__}')
        ckpt_info = check_model_path(model_path)

        print(f'Load path: {model_path} | {ckpt_info}')

        if QUANTIZATION == 'awq':
            print(F'Load model in int4 quantization')
            llm = LLM(model=model_path, dtype=dtype, tensor_parallel_size=tensor_parallel, gpu_memory_utilization=gpu_memory_utilization, quantization="awq")
        else:
            llm = LLM(model=model_path, dtype=dtype, tensor_parallel_size=tensor_parallel, gpu_memory_utilization=gpu_memory_utilization)

        try:
            print(llm.llm_engine.workers[0].model)
        except Exception as e:
            print(f'Cannot print model worker: {e}')

        print(f'Use system prompt:\n{sys_prompt}')

        response_fn = chat_response_stream_multiturn
        print(F'respond: {response_fn}')

    demo = gr.ChatInterface(
        response_fn,
        chatbot=ChatBot(
            label=MODEL_NAME,
            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),
        # ! consider preventing the stop button
        stop_btn=None,
        title=f"{model_title}",
        description=f"{model_desc}",
        additional_inputs=[
            gr.Number(value=temperature, label='Temperature (higher -> more random)'), 
            gr.Number(value=max_tokens, label='Max generated tokens (increase if want more generation)'), 
            gr.Number(value=frequence_penalty, label='Frequency penalty (> 0 encourage new tokens)'), 
            # ! Remove the system prompt textbox to avoid jailbreaking
            # gr.Textbox(value=sys_prompt, label='System prompt', lines=8)
        ], 
    )
    demo.title = MODEL_NAME
    with demo:
        # gr.Markdown(warning_markdown)
        gr.Markdown(cite_markdown)
        gr.Markdown(path_markdown.format(model_path=model_path))

    demo.queue()
    demo.launch(server_port=PORT)


def main():

    launch()


if __name__ == "__main__":
    main()