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 Southeast 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
import time

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")))

# List of languages to block
BLOCK_LANGS = str(os.environ.get("BLOCK_LANGS", ""))
BLOCK_LANGS = [x.strip() for x in BLOCK_LANGS.strip().split(";")] if len(BLOCK_LANGS.strip()) > 0 else []

# for lang block, wether to block in history too
LANG_BLOCK_HISTORY = bool(int(os.environ.get("LANG_BLOCK_HISTORY", "0")))
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")))
# ! show model path in the demo page, only for internal
DISPLAY_MODEL_PATH = bool(int(os.environ.get("DISPLAY_MODEL_PATH", "1")))

# ! 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")

# ! log path
LOG_PATH = os.environ.get("LOG_PATH", "").strip()
LOG_FILE = None
SAVE_LOGS = LOG_PATH is not None and LOG_PATH != ''
if SAVE_LOGS:
    if os.path.exists(LOG_PATH):
        print(f'LOG_PATH exist: {LOG_PATH}')
    else:
        LOG_DIR = os.path.dirname(LOG_PATH)
        os.makedirs(LOG_DIR, exist_ok=True)

# ! get LOG_PATH as aggregated outputs in log
GET_LOG_CMD = os.environ.get("GET_LOG_CMD", "").strip()

print(f'SAVE_LOGS: {SAVE_LOGS} | {LOG_PATH}')
print(f'GET_LOG_CMD: {GET_LOG_CMD}')

# ! !! Whether to delete the folder, ONLY SET THIS IF YOU WANT TO DELETE SAVED MODEL ON PERSISTENT DISC
DELETE_FOLDER = os.environ.get("DELETE_FOLDER", "")
IS_DELETE_FOLDER = DELETE_FOLDER is not None and os.path.exists(DELETE_FOLDER)
print(f'DELETE_FOLDER: {DELETE_FOLDER} | {DOWNLOAD_SNAPSHOT=}')

# ! 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"))

# whether to enable to popup accept user
ENABLE_AGREE_POPUP = bool(int(os.environ.get("ENABLE_AGREE_POPUP", "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"))
PRESENCE_PENALTY = float(os.environ.get("PRESENCE_PENALTY", "0.0"))
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", ""))


# Batch inference file upload
ENABLE_BATCH_INFER = bool(int(os.environ.get("ENABLE_BATCH_INFER", "1")))
BATCH_INFER_MAX_ITEMS = int(os.environ.get("BATCH_INFER_MAX_ITEMS", "200"))
BATCH_INFER_MAX_FILE_SIZE = int(os.environ.get("BATCH_INFER_MAX_FILE_SIZE", "500"))
BATCH_INFER_MAX_PROMPT_TOKENS = int(os.environ.get("BATCH_INFER_MAX_PROMPT_TOKENS", "4000"))
BATCH_INFER_SAVE_TMP_FILE = os.environ.get("BATCH_INFER_SAVE_TMP_FILE", "./tmp/pred.json")

# 
DATA_SET_REPO_PATH = str(os.environ.get("DATA_SET_REPO_PATH", ""))
DATA_SET_REPO = None

"""
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=SeaLLMs/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


HF_HOME=/data/.huggingface
MODEL_PATH=/data/ckpt/seal-13b-chat-a
DELETE_FOLDER=/data/

"""

# ==============================
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"

# TODO: should Hide the system prompt
SYSTEM_PROMPT_1 = """You are a multilingual, helpful, respectful and honest assistant. Your name is SeaLLM and you are built by DAMO Academy, Alibaba Group. \
Please 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 = """
<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: 3em;
            width: 3em;
            float: left;
            margin-left: auto;
        ">
      </div>
      <div class="text" style="
            padding-left: 20px;
            padding-top: 1%;
            float: left;
        ">
      <h1>SeaLLMs - Large Language Models for Southeast Asia</h1>
      </div>
</div>
"""
# <a href=''><img src='https://img.shields.io/badge/Paper-PDF-red'></a>
# MODEL_DESC = """
# <div style='display:flex; gap: 0.25rem; '>
# <a href='https://github.com/SeaLLMs/SeaLLMs'><img src='https://img.shields.io/badge/Github-Code-success'></a>
# <a href='https://huggingface.co/spaces/SeaLLMs/SeaLLM-Chat-13b'><img src='https://img.shields.io/badge/%F0%9F%A4%97%20Hugging%20Face-Spaces-blue'></a> 
# <a href='https://huggingface.co/SeaLLMs/SeaLLM-Chat-13b'><img src='https://img.shields.io/badge/%F0%9F%A4%97%20Hugging%20Face-Model-blue'></a> 
# </div>
# <span style="font-size: larger">
# This is <a href="https://huggingface.co/SeaLLMs/SeaLLM-Chat-13b" target="_blank">SeaLLM-13B-Chat</a> - a chatbot assistant optimized for Southeast Asian Languages. It produces helpful responses in English 🇬🇧, Vietnamese 🇻🇳, Indonesian 🇮🇩 and Thai 🇹🇭.
# Explore <a href="https://huggingface.co/SeaLLMs/SeaLLM-Chat-13b" target="_blank">our article</a> for more details.
# </span>
# <br>
# <span >
# NOTE: The chatbot may produce inaccurate and harmful information about people, places, or facts.
# <span style="color: red">By using our service, you are required to agree to our <a href="https://huggingface.co/SeaLLMs/SeaLLM-Chat-13b/blob/main/LICENSE" target="_blank" style="color: red">SeaLLM Terms Of Use</a>, which include:</span><br>
# <ul>
# <li >
# You must not use our service to generate any harmful, unethical or illegal content that violates locally applicable and international laws or regulations, 
# including but not limited to hate speech, violence, pornography and deception.</li>
# <li >
# The service collects user dialogue data for testing and performance improvement, and reserves the right to distribute it under 
# <a href="https://creativecommons.org/licenses/by/4.0/">Creative Commons Attribution (CC-BY)</a> or similar license. So do not enter any personal information!
# </li>
# </ul>
# </span>
# """.strip()

MODEL_DESC = """
<div style='display:flex; gap: 0.25rem; '>
<a href='https://github.com/SeaLLMs/SeaLLMs'><img src='https://img.shields.io/badge/Github-Code-success'></a>
<a href='https://huggingface.co/spaces/SeaLLMs/SeaLLM-Chat-13b'><img src='https://img.shields.io/badge/%F0%9F%A4%97%20Hugging%20Face-Spaces-blue'></a> 
<a href='https://huggingface.co/SeaLLMs/SeaLLM-Chat-13b'><img src='https://img.shields.io/badge/%F0%9F%A4%97%20Hugging%20Face-Model-blue'></a> 
</div>
<span style="font-size: larger">
This is <a href="https://huggingface.co/SeaLLMs/SeaLLM-Chat-13b" target="_blank">SeaLLM-13B-Chat</a> - a chatbot assistant optimized for Southeast Asian Languages. It produces helpful responses in English 🇬🇧, Vietnamese 🇻🇳, Indonesian 🇮🇩 and Thai 🇹🇭.
Explore <a href="https://huggingface.co/SeaLLMs/SeaLLM-Chat-13b" target="_blank">our article</a> for more details.
</span>
<br>
<span>
<span style="color: red">NOTE:</span> The chatbot may produce inaccurate and harmful information.
By using our service, you are required to <span style="color: red">agree to our <a href="https://huggingface.co/SeaLLMs/SeaLLM-Chat-13b/blob/main/LICENSE" target="_blank" style="color: red">Terms Of Use</a>,</span> which includes 
not to use our service to generate any harmful, inappropriate or unethical or illegal content that violates locally applicable and international laws and regulations. 
The service collects user dialogue data for testing and performance improvement, and reserves the right to distribute it under 
<a href="https://creativecommons.org/licenses/by/4.0/">(CC-BY)</a> or similar license. So do not enter any personal information!
</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 = {Xuan-Phi Nguyen*, Wenxuan Zhang*, Xin Li*, Mahani Aljunied*, Qingyu Tan, Liying Cheng, Guanzheng Chen, Yue Deng, Sen Yang, Chaoqun Liu, Hang Zhang, Lidong Bing},
  title = {SeaLLMs - Large Language Models for Southeast Asia},
  year = 2023,
}
```
"""

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




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, not been thoroughly tested yet.
    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,
    include_end_instruct=True,
):
    """
    ```
        <bos>[INST] B_SYS SytemPrompt E_SYS Prompt [/INST] Answer <eos>
        <bos>[INST] Prompt [/INST] Answer <eos>
        <bos>[INST] Prompt [/INST]
    ```
    """
    text = ''
    end_instr = f" {E_INST}" if include_end_instruct else ""
    for i, (prompt, res) in enumerate(history):
        if i == 0:
            text += f"{bos_token}{B_INST} {B_SYS} {sys_prompt} {E_SYS} {prompt}{end_instr}"
        else:
            text += f"{bos_token}{B_INST} {prompt}{end_instr}"

        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}{end_instr}"
    else:
        text += f"{bos_token}{B_INST} {message}{end_instr}"
    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:
    from gradio.components import State
    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:
    from gradio.components import State
    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

    def update_time(c_time, chatbot_state):
        # if chatbot_state is empty, register a new conversaion with the current timestamp
        assert len(chatbot_state) > 0, f'empty chatbot state'
        if len(chatbot_state) == 1:
            assert chatbot_state[-1][-1] is None, f'invalid [[message, None]] , got {chatbot_state}'
            return gr.Number(value=time.time(), label='current_time', visible=False), chatbot_state
        else:
            return c_time, chatbot_state

    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(
                update_time,
                [self.additional_inputs[-1], self.chatbot_state],
                [self.additional_inputs[-1], 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



@document()
class CustomTabbedInterface(gr.Blocks):
    def __init__(
        self,
        interface_list: list[gr.Interface],
        tab_names: Optional[list[str]] = None,
        title: Optional[str] = None,
        description: Optional[str] = None,
        theme: Optional[gr.Theme] = None,
        analytics_enabled: Optional[bool] = None,
        css: Optional[str] = None,
    ):
        """
        Parameters:
            interface_list: a list of interfaces to be rendered in tabs.
            tab_names: a list of tab names. If None, the tab names will be "Tab 1", "Tab 2", etc.
            title: a title for the interface; if provided, appears above the input and output components in large font. Also used as the tab title when opened in a browser window.
            analytics_enabled: whether to allow basic telemetry. If None, will use GRADIO_ANALYTICS_ENABLED environment variable or default to True.
            css: custom css or path to custom css file to apply to entire Blocks
        Returns:
            a Gradio Tabbed Interface for the given interfaces
        """
        super().__init__(
            title=title or "Gradio",
            theme=theme,
            analytics_enabled=analytics_enabled,
            mode="tabbed_interface",
            css=css,
        )
        self.description = description
        if tab_names is None:
            tab_names = [f"Tab {i}" for i in range(len(interface_list))]
        with self:
            if title:
                gr.Markdown(
                    f"<h1 style='text-align: center; margin-bottom: 1rem'>{title}</h1>"
                )
            if description:
                gr.Markdown(description)
            with gr.Tabs():
                for interface, tab_name in zip(interface_list, tab_names):
                    with gr.Tab(label=tab_name):
                        interface.render()



def vllm_abort(self: Any):
    sh = self.llm_engine.scheduler
    for g in (sh.waiting + sh.running + sh.swapped):
        sh.abort_seq_group(g.request_id)

    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)



# ! avoid saying 
LANG_BLOCK_MESSAGE = """Sorry, the language you have asked is currently not supported. If you have questions in other supported languages, I'll be glad to help. \
Please also consider clearing the chat box for a better experience."""

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


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


def block_lang(
    message: str, 
    history: List[Tuple[str, str]] = None,
) -> str:
    # relieve history base block
    if len(BLOCK_LANGS) == 0:
        return False
    
    if LANG_BLOCK_HISTORY and history is not None and any((LANG_BLOCK_MESSAGE in x[1].strip()) for x in history):
        return True
    else:
        _lang = _detect_lang(message)
        if _lang in BLOCK_LANGS:
            print(f'Detect blocked {_lang}: {message}')
            return True
        else:
            return False


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 len(KEYWORDS) > 0 and any(x in text.lower() for x in KEYWORDS):
        return KEYWORD_BLOCK_MESSAGE
    
    if len(BLOCK_LANGS) > 0:
        if block_lang(text, history):
            return LANG_BLOCK_MESSAGE

    return None


def chat_response_stream_multiturn(
    message: str, 
    history: List[Tuple[str, str]], 
    temperature: float, 
    max_tokens: int, 
    frequency_penalty: float,
    presence_penalty: float,
    current_time: Optional[float] = None,
    system_prompt: Optional[str] = SYSTEM_PROMPT_1
) -> str:
    global LOG_FILE, LOG_PATH
    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'
    tokenizer = llm.get_tokenizer()
    # force removing all 
    vllm_abort(llm)

    temperature = float(temperature)
    frequency_penalty = float(frequency_penalty)
    max_tokens = int(max_tokens)

    message = message.strip()

    if message.strip() == GET_LOG_CMD:
        print_log_file()
        yield "Finish printed log. Please clear the chatbox now."
        return 

    if len(message) == 0:
        raise gr.Error("The message cannot be empty!")

    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
    )

    if len(tokenizer.encode(full_prompt, add_special_tokens=False)) >= 4050:
        raise gr.Error(f"Conversation or prompt is too long, please clear the chatbox or try shorter input.")

    sampling_params = SamplingParams(
        temperature=temperature, 
        max_tokens=max_tokens,
        frequency_penalty=frequency_penalty,
        presence_penalty=presence_penalty,
        stop=['<s>', '</s>', '<<SYS>>', '<</SYS>>', '[INST]', '[/INST]']
    )
    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:
            cur_out = cur_out.replace("\\n", "\n")
            
            # 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

        if j >= max_tokens - 2:
            gr.Warning(f'The response hits limit of {max_tokens} tokens. Consider increase the max tokens parameter in the Additional Inputs.')
    
    # TODO: use current_time to register conversations, accoriding history and cur_out
    history_str = format_conversation(history + [[message, cur_out]])
    print(f'@@@@@@@@@@\n{history_str}\n##########\n')

    maybe_log_conv_file(current_time, history, message, cur_out, temperature=temperature, frequency_penalty=frequency_penalty)
    
    if cur_out is not None and "\\n" in cur_out:
        print(f'double slash-n in cur_out:\n{cur_out}')
        cur_out = cur_out.replace("\\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


def maybe_log_conv_file(current_time, history, message, response, **kwargs):
    global LOG_FILE
    if LOG_FILE is not None:
        my_history = history + [[message, response]]
        obj = {
            'key': str(current_time),
            'history': my_history
        }
        for k, v in kwargs.items():
            obj[k] = v
        log_ = json.dumps(obj, ensure_ascii=False)
        LOG_FILE.write(log_ + "\n")
        LOG_FILE.flush()
        print(f'Wrote {obj["key"]} to {LOG_PATH}')


def format_conversation(history):
    _str = '\n'.join([
        (
            f'<<<User>>> {h[0]}\n'
            f'<<<Asst>>> {h[1]}'
        )
        for h in history
    ])
    return _str


def maybe_upload_to_dataset():
    global LOG_FILE, DATA_SET_REPO_PATH, SAVE_LOGS
    if SAVE_LOGS and os.path.exists(LOG_PATH) and DATA_SET_REPO_PATH != "":
        with open(LOG_PATH, 'r', encoding='utf-8') as f:
            convos = {}
            for l in f:
                if l:
                    item = json.loads(l)
                    convos[item['key']] = item
        AGG_LOG_PATH = LOG_PATH + ".agg.json"
        with open(AGG_LOG_PATH, 'w', encoding='utf-8') as fo:
            json.dump(convos, fo, indent=4, ensure_ascii=False)
        print(f'Saved aggregated json to {AGG_LOG_PATH}')
        try:
            from huggingface_hub import upload_file
            print(f'upload {AGG_LOG_PATH} to {DATA_SET_REPO_PATH}')
            upload_file(
                path_or_fileobj=AGG_LOG_PATH,
                path_in_repo=os.path.basename(AGG_LOG_PATH),
                repo_id=DATA_SET_REPO_PATH,
                token=HF_TOKEN,
                repo_type="dataset",
                create_pr=True
            )
        except Exception as e:
            print(f'Failed to save to repo: {DATA_SET_REPO_PATH}|{str(e)}')


def print_log_file():
    global LOG_FILE, LOG_PATH
    if SAVE_LOGS and os.path.exists(LOG_PATH):
        with open(LOG_PATH, 'r', encoding='utf-8') as f:
            convos = {}
            for l in f:
                if l:
                    item = json.loads(l)
                    convos[item['key']] = item
            print(f'Printing log from {LOG_PATH}')
            for k, v in convos.items():
                history = v.pop('history')
                print(f'######--{v}--##')
                _str = format_conversation(history)
                print(_str)
        maybe_upload_to_dataset()


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,
    presence_penalty: float = 0.0,
    current_time: Optional[float] = None,
    system_prompt: str = SYSTEM_PROMPT_1,
) -> str:
    global LOG_FILE
    import time
    time.sleep(0.5)

    if message.strip() == GET_LOG_CMD:
        print_log_file()
        yield "Finish printed log."
        return 

    for i in range(len(message)):
        yield f"repeat: {current_time} {message[:i + 1]}"
    
    cur_out = f"repeat: {current_time} {message}"
    maybe_log_conv_file(current_time, history, message, cur_out, temperature=temperature, frequency_penalty=frequency_penalty)


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 maybe_delete_folder():
    if IS_DELETE_FOLDER and DOWNLOAD_SNAPSHOT:
        import shutil
        print(f'DELETE ALL FILES IN {DELETE_FOLDER}')
        for filename in os.listdir(DELETE_FOLDER):
            file_path = os.path.join(DELETE_FOLDER, filename)
            try:
                if os.path.isfile(file_path) or os.path.islink(file_path):
                    os.unlink(file_path)
                elif os.path.isdir(file_path):
                    shutil.rmtree(file_path)
            except Exception as e:
                print('Failed to delete %s. Reason: %s' % (file_path, e))


AGREE_POP_SCRIPTS = """
async () => {
    alert("To use our service, you are required to agree to the following terms:\\nYou must not use our service to generate any harmful, unethical or illegal content that violates local and international laws, including but not limited to hate speech, violence and deception.\\nThe service may collect user dialogue data for performance improvement, and reserves the right to distribute it under CC-BY or similar license. So do not enter any personal information!");
}
"""

def debug_file_function(
        files: Union[str, List[str]],
        prompt_mode: str,
        temperature: float, 
        max_tokens: int, 
        frequency_penalty: float,
        presence_penalty: float,
        stop_strings: str = "[STOP],<s>,</s>",
        current_time: Optional[float] = None,          
):
    """This is only for debug purpose"""
    files = files if isinstance(files, list) else [files]
    print(files)
    filenames = [f.name for f in files]
    all_items = []
    for fname in filenames:
        print(f'Reading {fname}')
        with open(fname, 'r', encoding='utf-8') as f:
            items = json.load(f)
        assert isinstance(items, list), f'invalid items from {fname} not list'
        all_items.extend(items)
    print(all_items)
    print(f'{prompt_mode} / {temperature} / {max_tokens}, {frequency_penalty}, {presence_penalty}')
    save_path = "./test.json"
    with open(save_path, 'w', encoding='utf-8') as f:
        json.dump(all_items, f, indent=4, ensure_ascii=False)

    for x in all_items:
        x['response'] = "Return response"
    
    print_items = all_items[:1]
    # print_json = json.dumps(print_items, indent=4, ensure_ascii=False)
    return save_path, print_items


def validate_file_item(filename, index, item: Dict[str, str]):
    """
    check safety for items in files
    """
    message = item['prompt'].strip()

    if len(message) == 0:
        raise gr.Error(f'Prompt {index} empty')
    
    message_safety = safety_check(message, history=None)
    if message_safety is not None:
        raise gr.Error(f'Prompt {index} invalid: {message_safety}')
    
    tokenizer = llm.get_tokenizer() if llm is not None else None
    if tokenizer is None or len(tokenizer.encode(message, add_special_tokens=False)) >= BATCH_INFER_MAX_PROMPT_TOKENS:
        raise gr.Error(f"Prompt {index} too long, should be less than {BATCH_INFER_MAX_PROMPT_TOKENS} tokens")


def read_validate_json_files(files: Union[str, List[str]]):
    files = files if isinstance(files, list) else [files]
    filenames = [f.name for f in files]
    all_items = []
    for fname in filenames:
        # check each files
        print(f'Reading {fname}')
        with open(fname, 'r', encoding='utf-8') as f:
            items = json.load(f)
        assert isinstance(items, list), f'Data {fname} not list'
        assert all(isinstance(x, dict) for x in items), f'item in input file not list'
        assert all("prompt" in x for x in items), f'key prompt should be in dict item of input file'

        for i, x in enumerate(items):
            validate_file_item(fname, i, x)

        all_items.extend(items)

    if len(all_items) > BATCH_INFER_MAX_ITEMS:
        raise gr.Error(f"Num samples {len(all_items)} > {BATCH_INFER_MAX_ITEMS} allowed.")
    
    return all_items, filenames


def remove_gradio_cache(exclude_names=None):
    """remove gradio cache to avoid flooding"""
    import shutil
    for root, dirs, files in os.walk('/tmp/gradio/'):
        for f in files:
            # if not any(f in ef for ef in except_files):
            if exclude_names is None or not any(ef in f for ef in exclude_names):
                print(f'Remove: {f}')
                os.unlink(os.path.join(root, f))
        # for d in dirs:
        #     # if not any(d in ef for ef in except_files):
        #     if exclude_names is None or not any(ef in d for ef in exclude_names):
        #         print(f'Remove d: {d}')
        #         shutil.rmtree(os.path.join(root, d))


def maybe_upload_batch_set(pred_json_path):
    global LOG_FILE, DATA_SET_REPO_PATH, SAVE_LOGS

    if SAVE_LOGS and DATA_SET_REPO_PATH != "":
        try:
            from huggingface_hub import upload_file
            path_in_repo = "misc/" + os.path.basename(pred_json_path).replace(".json", f'.{time.time()}.json')
            print(f'upload {pred_json_path} to {DATA_SET_REPO_PATH}//{path_in_repo}')
            upload_file(
                path_or_fileobj=pred_json_path,
                path_in_repo=path_in_repo,
                repo_id=DATA_SET_REPO_PATH,
                token=HF_TOKEN,
                repo_type="dataset",
                create_pr=True
            )
        except Exception as e:
            print(f'Failed to save to repo: {DATA_SET_REPO_PATH}|{str(e)}')


def batch_inference(
        files: Union[str, List[str]], 
        prompt_mode: str,
        temperature: float, 
        max_tokens: int, 
        frequency_penalty: float,
        presence_penalty: float,
        stop_strings: str = "[STOP],<s>,</s>",
        current_time: Optional[float] = None,
        system_prompt: Optional[str] = SYSTEM_PROMPT_1
):
    """
    Handle file upload batch inference
    
    """
    global LOG_FILE, LOG_PATH, DEBUG, llm, RES_PRINTED
    if DEBUG:
        return debug_file_function(
            files, prompt_mode, temperature, max_tokens, 
            presence_penalty, stop_strings, current_time)
    
    from vllm import LLM, SamplingParams
    assert llm is not None
    # assert system_prompt.strip() != '', f'system prompt is empty'

    stop_strings = [x.strip() for x in stop_strings.strip().split(",")]
    tokenizer = llm.get_tokenizer()
    # force removing all 
    # NOTE: need to make sure all cached items are removed!!!!!!!!!
    vllm_abort(llm)

    temperature = float(temperature)
    frequency_penalty = float(frequency_penalty)
    max_tokens = int(max_tokens)

    all_items, filenames = read_validate_json_files(files)

    # remove all items in /tmp/gradio/
    remove_gradio_cache(exclude_names=['upload_chat.json', 'upload_few_shot.json'])

    if prompt_mode == 'chat':
        prompt_format_fn = llama_chat_multiturn_sys_input_seq_constructor
    elif prompt_mode == 'few-shot':
        from functools import partial
        prompt_format_fn = partial(
            llama_chat_multiturn_sys_input_seq_constructor, include_end_instruct=False
        )
    else:
        raise gr.Error(f'Wrong mode {prompt_mode}')

    full_prompts = [
        prompt_format_fn(
            x['prompt'], [], sys_prompt=system_prompt
        )
        for i, x in enumerate(all_items)
    ]
    print(f'{full_prompts[0]}\n')
    
    if any(len(tokenizer.encode(x, add_special_tokens=False)) >= 4090 for x in full_prompts):
        raise gr.Error(f"Some prompt is too long!")

    stop_seq = list(set(['<s>', '</s>', '<<SYS>>', '<</SYS>>', '[INST]', '[/INST]'] + stop_strings))
    sampling_params = SamplingParams(
        temperature=temperature, 
        max_tokens=max_tokens,
        frequency_penalty=frequency_penalty,
        presence_penalty=presence_penalty,
        stop=stop_seq
    )

    generated = llm.generate(full_prompts, sampling_params, use_tqdm=False)
    responses = [g.outputs[0].text for g in generated]
    if len(responses) != len(all_items):
        raise gr.Error(f'inconsistent lengths {len(responses)} != {len(all_items)}')

    for res, item in zip(responses, all_items):
        item['response'] = res

    # save_path = "/mnt/workspace/workgroup/phi/test.json"
    save_path = BATCH_INFER_SAVE_TMP_FILE
    os.makedirs(os.path.dirname(save_path), exist_ok=True)
    with open(save_path, 'w', encoding='utf-8') as f:
        json.dump(all_items, f, indent=4, ensure_ascii=False)

    # You need to upload save_path as a new timestamp file.
    maybe_upload_batch_set(save_path)
    
    print_items = all_items[:2]
    # print_json = json.dumps(print_items, indent=4, ensure_ascii=False)
    return save_path, print_items
    

# BATCH_INFER_MAX_ITEMS
FILE_UPLOAD_DESCRIPTION = f"""Upload JSON file as list of dict with < {BATCH_INFER_MAX_ITEMS} items, \
each item has `prompt` key. We put guardrails to enhance safety, so do not input any harmful content or personal information! Re-upload the file after every submit. See the examples below.
```
[ {{"id": 0, "prompt": "Hello world"}} ,  {{"id": 1, "prompt": "Hi there?"}}]
```
"""


def launch():
    global demo, llm, DEBUG, LOG_FILE
    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
    presence_penalty = PRESENCE_PENALTY
    ckpt_info = "None"

    print(
        f'Launch config: '
        f'\n| model_title=`{model_title}` '
        f'\n| max_tokens={max_tokens} '
        f'\n| dtype={dtype} '
        f'\n| tensor_parallel={tensor_parallel} '
        f'\n| BLOCK_LANGS={BLOCK_LANGS} '
        f'\n| IS_DELETE_FOLDER={IS_DELETE_FOLDER} '
        f'\n| STREAM_YIELD_MULTIPLE={STREAM_YIELD_MULTIPLE} '
        f'\n| STREAM_CHECK_MULTIPLE={STREAM_CHECK_MULTIPLE} '
        f'\n| DISPLAY_MODEL_PATH={DISPLAY_MODEL_PATH} '
        f'\n| LANG_BLOCK_HISTORY={LANG_BLOCK_HISTORY} '
        f'\n| frequence_penalty={frequence_penalty} '
        f'\n| presence_penalty={presence_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'\n| LOG_PATH={LOG_PATH} | SAVE_LOGS={SAVE_LOGS} '
        f'\n| DATA_SET_REPO_PATH={DATA_SET_REPO_PATH} '
        f'\n| GET_LOG_CMD={GET_LOG_CMD} '
        f'\n| Sys={SYSTEM_PROMPT_1}'
        f'\n| Desc={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')
        if SAVE_LOGS:
            LOG_FILE = open(LOG_PATH, 'a', encoding='utf-8')
    else:
        # ! load the model
        maybe_delete_folder()

        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}')

        try:
            llm.llm_engine.scheduler_config.max_model_len = 4096
            llm.llm_engine.scheduler_config.max_num_batched_tokens = 4096
            llm.llm_engine.tokenizer.add_special_tokens = False
        except Exception as e:
            print(f'Cannot set parameters: {e}')

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

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

        if SAVE_LOGS:
            LOG_FILE = open(LOG_PATH, 'a', encoding='utf-8')

    if ENABLE_BATCH_INFER:

        demo_file_upload = gr.Interface(
            batch_inference,
            inputs=[
                gr.File(file_count='single', file_types=['json']),
                gr.Radio(["chat", "few-shot"], value='chat', label="Chat or Few-shot mode", info="Chat's output more user-friendly, Few-shot's output more consistent with few-shot patterns."),
                gr.Number(value=temperature, label='Temperature', info="Higher -> more random"), 
                gr.Number(value=max_tokens, label='Max tokens', info='Increase if want more generation'), 
                gr.Number(value=frequence_penalty, label='Frequency penalty', info='> 0 encourage new tokens over repeated tokens'), 
                gr.Number(value=presence_penalty, label='Presence penalty', info='> 0 encourage new tokens, < 0 encourage existing tokens'), 
                gr.Textbox(value="[STOP],[END],<s>,</s>", label='Stop strings', info='Comma-separated string to stop generation only in FEW-SHOT mode', lines=1),
                gr.Number(value=0, label='current_time', visible=False), 
            ],
            outputs=[
                # "file",
                gr.File(label="Generated file"),
                # "json"
                gr.JSON(label='Example outputs (display 2 samples)')
            ],
            description=FILE_UPLOAD_DESCRIPTION,
            allow_flagging=False,
            examples=[
                ["upload_chat.json", "chat", 0.2, 1024, 0.5, 0, "[STOP],[END],<s>,</s>"],
                ["upload_few_shot.json", "few-shot", 0.2, 128, 0.5, 0, "[STOP],[END],<s>,</s>,\\n"]
            ],
            # cache_examples=True,
        )
        
        demo_chat = gr.ChatInterface(
            response_fn,
            chatbot=ChatBot(
                label=MODEL_NAME,
                bubble_full_width=False,
                latex_delimiters=[
                    { "left": "$", "right": "$", "display": False},
                    { "left": "$$", "right": "$$", "display": True},
                ],
                show_copy_button=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 over repeated tokens)'), 
                gr.Number(value=presence_penalty, label='Presence penalty (> 0 encourage new tokens, < 0 encourage existing tokens)'), 
                gr.Number(value=0, label='current_time', visible=False), 
                # ! Remove the system prompt textbox to avoid jailbreaking
                # gr.Textbox(value=sys_prompt, label='System prompt', lines=8)
            ], 
        )
        demo = CustomTabbedInterface(
            interface_list=[demo_chat, demo_file_upload],
            tab_names=["Chat Interface", "Batch Inference"],
            title=f"{model_title}",
            description=f"{model_desc}",
        )
        demo.title = MODEL_NAME
        with demo:
            gr.Markdown(cite_markdown)
            if DISPLAY_MODEL_PATH:
                gr.Markdown(path_markdown.format(model_path=model_path))
            
            if ENABLE_AGREE_POPUP:
                demo.load(None, None, None, _js=AGREE_POP_SCRIPTS)
        

        demo.queue()
        demo.launch(server_port=PORT)
    else:
        demo = gr.ChatInterface(
            response_fn,
            chatbot=ChatBot(
                label=MODEL_NAME,
                bubble_full_width=False,
                latex_delimiters=[
                    { "left": "$", "right": "$", "display": False},
                    { "left": "$$", "right": "$$", "display": True},
                ],
                show_copy_button=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 over repeated tokens)'), 
                gr.Number(value=presence_penalty, label='Presence penalty (> 0 encourage new tokens, < 0 encourage existing tokens)'), 
                gr.Number(value=0, label='current_time', visible=False), 
                # ! 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(cite_markdown)
            if DISPLAY_MODEL_PATH:
                gr.Markdown(path_markdown.format(model_path=model_path))
            
            if ENABLE_AGREE_POPUP:
                demo.load(None, None, None, _js=AGREE_POP_SCRIPTS)
        

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


def main():

    launch()


if __name__ == "__main__":
    main()