app.py

import gradio as gr
import os, gc, copy, torch
from huggingface_hub import hf_hub_download
from pynvml import *

# Flag to check if GPU is present
HAS_GPU = False

# Model title and context size limit
ctx_limit = 2000
title = "RWKV-5-World-1B5-v2-Translator"
model_file = "RWKV-5-World-1B5-v2-20231025-ctx4096"

# Get the GPU count
try:
    nvmlInit()
    GPU_COUNT = nvmlDeviceGetCount()
    if GPU_COUNT > 0:
        HAS_GPU = True
        gpu_h = nvmlDeviceGetHandleByIndex(0)
except NVMLError as error:
    print(error)

os.environ["RWKV_JIT_ON"] = '1'

# Model strategy to use
MODEL_STRAT = "cpu bf16"
os.environ["RWKV_CUDA_ON"] = '0'  # if '1' then use CUDA kernel for seq mode (much faster)

# Switch to GPU mode
if HAS_GPU:
    os.environ["RWKV_CUDA_ON"] = '1'
    MODEL_STRAT = "cuda bf16"

# Load the model
from rwkv.model import RWKV
model_path = hf_hub_download(repo_id="BlinkDL/rwkv-5-world", filename=f"{model_file}.pth")
model = RWKV(model=model_path, strategy=MODEL_STRAT)
from rwkv.utils import PIPELINE
pipeline = PIPELINE(model, "rwkv_vocab_v20230424")

# State copy
def universal_deepcopy(obj):
    if hasattr(obj, 'clone'):  # Assuming it's a tensor if it has a clone method
        return obj.clone()
    elif isinstance(obj, list):
        return [universal_deepcopy(item) for item in obj]
    else:
        return copy.deepcopy(obj)

# For debgging mostly
def inspect_structure(obj, depth=0):
    indent = "  " * depth
    obj_type = type(obj).__name__
    
    if isinstance(obj, list):
        print(f"{indent}List (length {len(obj)}):")
        for item in obj:
            inspect_structure(item, depth + 1)
    elif isinstance(obj, dict):
        print(f"{indent}Dict (length {len(obj)}):")
        for key, value in obj.items():
            print(f"{indent}  Key: {key}")
            inspect_structure(value, depth + 1)
    else:
        print(f"{indent}{obj_type}")

# Precomputation of the state
def precompute_state(text):
    state = None
    text_encoded = pipeline.encode(text)
    _, state = model.forward(text_encoded, state)
    return state

# Precomputing the base instruction set
INSTRUCT_PREFIX = f'''
You are a translator bot that can translate text to any language.
And will respond only with the translated text, without additional comments.

## From English:
It is not enough to know, we must also apply; it is not enough to will, we must also do.
## To Polish:
Nie wystarczy wiedzieć, trzeba także zastosować; nie wystarczy chcieć, trzeba też działać.

## From Spanish:
La muerte no nos concierne, porque mientras existamos, la muerte no está aquí. Y cuando llega, ya no existimos.
## To English:
Death does not concern us, because as long as we exist, death is not here. And when it does come, we no longer exist.


'''

# Get the prefix state
PREFIX_STATE = precompute_state(INSTRUCT_PREFIX)

# Translation logic
def translate(
        text, source_language, target_language, 
        inState=PREFIX_STATE,
        temperature=0.2,
        top_p=0.5,
        presencePenalty = 0.1,
        countPenalty = 0.1,    
    ):
    prompt = f"## From {source_language}:\n{text}\n\n## To {target_language}:\n"
    ctx = prompt.strip()
    all_tokens = []
    out_last = 0
    out_str = ''
    occurrence = {}

    alpha_frequency = countPenalty
    alpha_presence = presencePenalty

    state = None
    if inState != None:
        state = universal_deepcopy(inState)
    
    # Clear GC
    gc.collect()
    if HAS_GPU == True :
        torch.cuda.empty_cache()

    # Generate things token by token
    for i in range(ctx_limit):
        out, state = model.forward(pipeline.encode(ctx)[-ctx_limit:] if i == 0 else [token], state)
        for n in occurrence:
            out[n] -= (alpha_presence + occurrence[n] * alpha_frequency)
        token = pipeline.sample_logits(out, temperature=temperature, top_p=top_p)
        
        if token in [0]:  # EOS token
            break

        all_tokens += [token]
        for xxx in occurrence:
            occurrence[xxx] *= 0.996        
        if token not in occurrence:
            occurrence[token] = 1
        else:
            occurrence[token] += 1

        tmp = pipeline.decode(all_tokens[out_last:])
        if '\ufffd' not in tmp:
            out_str += tmp
            out_last = i + 1
        else:
            return out_str.strip()

        if "\n:" in out_str :
            out_str = out_str.split("\n\nHuman:")[0].split("\nHuman:")[0]
            return out_str.strip()

        if "{source_language}:" in out_str :
            out_str = out_str.split("{source_language}:")[0]
            return out_str.strip()

        if "{target_language}:" in out_str :
            out_str = out_str.split("{target_language}:")[0]
            return out_str.strip()

        if "\nHuman:" in out_str :
            out_str = out_str.split("\n\nHuman:")[0].split("\nHuman:")[0]
            return out_str.strip()

        if "\nAssistant:" in out_str :
            out_str = out_str.split("\n\nAssistant:")[0].split("\nAssistant:")[0]
            return out_str.strip()

        if "\n#" in out_str :
            out_str = out_str.split("\n\n#")[0].split("\n#")[0]
            return out_str.strip()

        # Yield for streaming
        yield out_str.strip()

    del out
    del state

    # # Clear GC
    # gc.collect()
    # if HAS_GPU == True :
    #     torch.cuda.empty_cache()
    
    # yield out_str.strip()
    return out_str.strip()

# Languages
LANGUAGES = [
  "English",
  "Chinese",
  "Spanish",
  "Bengali",
  "Hindi", 
  "Portuguese",
  "Russian",
  "Japanese",
  "German",
  "Chinese (Wu)",
  "Javanese",
  "Korean",
  "French",
  "Vietnamese",
  "Telugu",
  "Chinese (Yue)",
  "Marathi", 
  "Tamil",
  "Turkish",
  "Urdu",
  "Chinese (Min Nan)",
  "Chinese (Jin Yu)",
  "Gujarati",
  "Polish",
  "Arabic (Egyptian Spoken)",
  "Ukrainian",
  "Italian",
  "Chinese (Xiang)",
  "Malayalam",
  "Chinese (Hakka)",
  "Kannada",
  "Oriya",
  "Panjabi (Western)",
  "Panjabi (Eastern)",
  "Sunda",
  "Romanian",
  "Bhojpuri",
  "Azerbaijani (South)",
  "Farsi (Western)",
  "Maithili",
  "Hausa",
  "Arabic (Algerian Spoken)",
  "Burmese",
  "Serbo-Croatian",
  "Chinese (Gan)",
  "Awadhi",
  "Thai",
  "Dutch", 
  "Yoruba",
  "Sindhi",
  "Arabic (Moroccan Spoken)",
  "Arabic (Saidi Spoken)",
  "Uzbek, Northern",
  "Malay",
  "Amharic",
  "Indonesian",
  "Igbo",
  "Tagalog",
  "Nepali",
  "Arabic (Sudanese Spoken)",
  "Saraiki",
  "Cebuano",
  "Arabic (North Levantine Spoken)",
  "Thai (Northeastern)",
  "Assamese",
  "Hungarian",
  "Chittagonian",
  "Arabic (Mesopotamian Spoken)",
  "Madura",
  "Sinhala",
  "Haryanvi",
  "Marwari",
  "Czech",
  "Greek",
  "Magahi",
  "Chhattisgarhi",
  "Deccan",
  "Chinese (Min Bei)",
  "Belarusan",
  "Zhuang (Northern)",
  "Arabic (Najdi Spoken)",
  "Pashto (Northern)",
  "Somali",
  "Malagasy",
  "Arabic (Tunisian Spoken)",
  "Rwanda",
  "Zulu",
  "Latin",
  "Bulgarian",
  "Swedish",
  "Lombard",
  "Oromo (West-central)",
  "Pashto (Southern)",
  "Kazakh",
  "Ilocano",
  "Tatar",
  "Fulfulde (Nigerian)",
  "Arabic (Sanaani Spoken)",
  "Uyghur",
  "Haitian Creole French",
  "Azerbaijani, North",
  "Napoletano-calabrese",
  "Khmer (Central)",
  "Farsi (Eastern)",
  "Akan",
  "Hiligaynon",
  "Kurmanji",
  "Shona"
]

# Example data
EXAMPLES = [
    # More people would learn from their mistakes if they weren't so busy denying them.
    ["Többen tanulnának a hibáikból, ha nem lennének annyira elfoglalva, hogy tagadják azokat.", "Hungarian", "English"],
    ["La mejor venganza es el éxito masivo.", "Spanish", "English"],
    ["Tout est bien qui finit bien.", "French", "English"],
    ["Lasciate ogne speranza, voi ch'intrate.", "Italian", "English"],
    ["Errare humanum est.", "Latin", "English"],

    # ["Brargh-ains argh-uh foo-duh", "English"],
    # ["I Want to eat your brains", "Zombie Speak"],
    # ["Bonjour, comment ça va?", "English"],
    # ["Hola, ¿cómo estás?", "English"],
    # ["你好吗?", "English"],
    # ["Guten Tag, wie geht es Ihnen?", "English"],
    # ["Привет, как ты?", "English"],
    # ["مرحبًا ، كيف حالك؟", "English"],
]
# Gradio interface
with gr.Blocks(title=title) as demo:
    gr.HTML(f"<div style=\"text-align: center;\"><h1>RWKV-5 World v2 - {title}</h1></div>")
    gr.Markdown("This is the RWKV-5 World v2 1B5 model tailored for translation tasks. All on 8 vCPUs")
    
    # Input and output components
    text = gr.Textbox(lines=5, label="Source Text", placeholder="Enter the text you want to translate...", value=EXAMPLES[0][0])
    source_language = gr.Dropdown(choices=LANGUAGES, label="Source Language", value=EXAMPLES[0][1])
    target_language = gr.Dropdown(choices=LANGUAGES, label="Target Language", value=EXAMPLES[0][2])
    output = gr.Textbox(lines=5, label="Translated Text")

    # Submission
    submit = gr.Button("Translate", variant="primary")
    
    # Example data
    data = gr.Dataset(components=[text, source_language, target_language], samples=EXAMPLES, label="Example Translations", headers=["Source Text", "Source Language", "Target Language"])
    
    # Button action
    submit.click(translate, [text, source_language, target_language], [output])
    data.click(lambda x: x, [data], [text, source_language, target_language])

# Gradio launch
demo.queue(concurrency_count=1, max_size=10)
demo.launch(share=False, debug=True)