app.py

import spaces
import os
import json
import subprocess
from functools import lru_cache

from cachetools import LRUCache
import threading

MODEL_CACHE = LRUCache(maxsize=3)   # max 10 models
CACHE_LOCK = threading.Lock()


### monkey patch 
from llama_cpp.llama_chat_format import Qwen35ChatHandler
import llama_cpp._internals as internals


# 2️⃣ Monkey patch BEFORE creating Llama()

_original_close = internals.LlamaModel.close

def safe_close(self):
    try:
        if hasattr(self, "sampler") and self.sampler is not None:
            return _original_close(self)
    except Exception:
        pass

internals.LlamaModel.close = safe_close


def safe_del(self):
    try:
        self.close()
    except Exception:
        pass

internals.LlamaModel.__del__ = safe_del




##### final verdict

# qwen 3 next 80b infrence 
#qwen 3 VL
#mini max 2.5
#gpt oss 120B     #seems unstable 










import llama_cpp._internals as internals
from llama_cpp.llama_chat_format import Qwen3VLChatHandler


from llama_cpp import Llama
from llama_cpp_agent import LlamaCppAgent, MessagesFormatterType
from llama_cpp_agent.providers import LlamaCppPythonProvider
from llama_cpp_agent.chat_history import BasicChatHistory
from llama_cpp_agent.chat_history.messages import Roles
import gradio as gr
from huggingface_hub import hf_hub_download

import llama_cpp
print(llama_cpp.__file__)
print(llama_cpp.__version__)

huggingface_token = os.getenv("HUGGINGFACE_TOKEN")

# hf_hub_download(
#     repo_id="bartowski/gemma-2-9b-it-GGUF",
#     filename="gemma-2-9b-it-Q5_K_M.gguf",
#     local_dir="./models"
# )

# hf_hub_download(
#     repo_id="bartowski/gemma-2-27b-it-GGUF",
#     filename="gemma-2-27b-it-Q5_K_M.gguf",
#     local_dir="./models"
# )

# hf_hub_download(
#     repo_id="google/gemma-2-2b-it-GGUF",
#     filename="2b_it_v2.gguf",
#     local_dir="./models",
#     token=huggingface_token
# )

# hf_hub_donload(
#     repo_id="unsloth/GLM-4.7-Flash-GGUF",
#     filename="GLM-4.7-Flash-UD-Q8_K_XL.gguf",
#     local_dir="./models",
#     token=huggingface_token
# )

# hf_hub_download(
#     repo_id="unsloth/gpt-oss-20b-GGUF",
#     filename="gpt-oss-20b-Q4_K_M.gguf",
#     local_dir="./models",
#     token=huggingface_token
# )

# hf_hub_download(
#     repo_id="unsloth/gpt-oss-20b-GGUF",
#     filename="gpt-oss-20b-Q4_K_M.gguf",
#     local_dir="./models",
#     token=huggingface_token
# )




# hf_hub_download(
#     repo_id="unsloth/Qwen3-Next-80B-A3B-Instruct-GGUF",
#     filename="Qwen3-Next-80B-A3B-Instruct-Q4_K_M.gguf",
#     local_dir="./models",
#     token=huggingface_token
# )


# hf_hub_download(
#     repo_id="unsloth/Qwen3-VL-32B-Thinking-GGUF",
#     filename="Qwen3-VL-32B-Thinking-Q8_0.gguf",
#     local_dir="./models"
# )


# hf_hub_download(
#     repo_id="unsloth/Qwen3-Coder-Next-GGUF",
#     filename="Qwen3-Coder-Next-Q4_K_M.gguf",
#     local_dir="./models"
# )
from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="unsloth/Qwen3.5-122B-A10B-GGUF",
    repo_type="model",
    local_dir="./models/",
    allow_patterns=["UD-Q3_K_XL/*"],   # 👈 folder inside repo
    token=huggingface_token      # only if gated/private
)

hf_hub_download(
    repo_id="unsloth/Qwen3.5-122B-A10B-GGUF",
    filename="mmproj-BF16.gguf",
    local_dir="./models"
)


_IMAGE_MIME_TYPES = {
    # Most common formats
    '.png':  'image/png',
    '.jpg':  'image/jpeg',
    '.jpeg': 'image/jpeg',
    '.gif':  'image/gif',
    '.webp': 'image/webp',

    # Next-generation formats
    '.avif': 'image/avif',
    '.jp2':  'image/jp2',
    '.j2k':  'image/jp2',
    '.jpx':  'image/jp2',

    # Legacy / Windows formats
    '.bmp':  'image/bmp',
    '.ico':  'image/x-icon',
    '.pcx':  'image/x-pcx',
    '.tga':  'image/x-tga',
    '.icns': 'image/icns',

    # Professional / Scientific imaging
    '.tif':  'image/tiff',
    '.tiff': 'image/tiff',
    '.eps':  'application/postscript',
    '.dds':  'image/vnd-ms.dds',
    '.dib':  'image/dib',
    '.sgi':  'image/sgi',

    # Portable Map formats (PPM/PGM/PBM)
    '.pbm':  'image/x-portable-bitmap',
    '.pgm':  'image/x-portable-graymap',
    '.ppm':  'image/x-portable-pixmap',

    # Miscellaneous / Older formats
    '.xbm':  'image/x-xbitmap',
    '.mpo':  'image/mpo',
    '.msp':  'image/msp',
    '.im':   'image/x-pillow-im',
    '.qoi':  'image/qoi',
}
import os
import base64

def image_to_base64_data_uri(
    file_path: str,
    *,
    fallback_mime: str = "application/octet-stream"
) -> str:
    """
    Convert a local image file to a base64-encoded data URI with the correct MIME type.
    Supports 20+ image formats (PNG, JPEG, WebP, AVIF, BMP, ICO, TIFF, etc.).
    Args:
        file_path: Path to the image file on disk.
        fallback_mime: MIME type used when the file extension is unknown.
    Returns:
        A valid data URI string (e.g., data:image/webp;base64,...).
    Raises:
        FileNotFoundError: If the file does not exist.
        OSError: If reading the file fails.
    """
    if not os.path.isfile(file_path):
        raise FileNotFoundError(f"Image file not found: {file_path}")

    extension = os.path.splitext(file_path)[1].lower()
    mime_type = _IMAGE_MIME_TYPES.get(extension, fallback_mime)

    if mime_type == fallback_mime:
        print(f"Warning: Unknown extension '{extension}' for '{file_path}'. "
              f"Using fallback MIME type: {fallback_mime}")

    try:
        with open(file_path, "rb") as img_file:
            encoded_data = base64.b64encode(img_file.read()).decode("utf-8")
    except OSError as e:
        raise OSError(f"Failed to read image file '{file_path}': {e}") from e

    return f"data:{mime_type};base64,{encoded_data}"



import os
import tempfile
import requests
from urllib.parse import urlparse





def handle_image_input(image_input):
    """
    image_input can be:
    - URL (http/https)
    - Data URI (data:image/png;base64,...)
    """

    # Case 1: If it's a Data URI → do nothing
    if image_input.startswith("data:"):
        print("Data URI detected. No download needed.")
        return process_image(image_input)

    # Case 2: If it's a URL → download temporarily
    parsed = urlparse(image_input)

    if parsed.scheme in ("http", "https"):
        print("URL detected. Downloading temporarily...")

        response = requests.get(image_input)
        response.raise_for_status()

        # Create temporary file
        with tempfile.NamedTemporaryFile(delete=False, suffix=".png") as tmp_file:
            tmp_file.write(response.content)
            temp_path = tmp_file.name

        try:
            tmp_tmp=image_to_base64_data_uri(temp_path)
        finally:
            # Ensure deletion
            os.remove(temp_path)
            return tmp_tmp

    else:
        raise ValueError("Unsupported image format.")




import os

def print_tree(start_path="models"):
    for root, dirs, files in os.walk(start_path):
        level = root.replace(start_path, "").count(os.sep)
        indent = "│   " * level
        print(f"{indent}├── {os.path.basename(root)}/")
        
        subindent = "│   " * (level + 1)
        for f in files:
            print(f"{subindent}├── {f}")

print_tree("models")


llm = None
llm_model = None

import json

def str_to_json(str_obj):
    json_obj = json.loads(str_obj)
    return json_obj

MODES = {

    # ===============================
    # THINKING MODE - GENERAL
    # ===============================
    "thinking_general": {
        "temperature": 1.0,
        "top_p": 0.95,
        "top_k": 20,
        "min_p": 0.0,
        "presence_penalty": 1.5,
        "repeat_penalty": 1.0,
        "enable_thinking": True,
    },

    # ===============================
    # THINKING MODE - CODING
    # ===============================
    "thinking_coding": {
        "temperature": 0.6,
        "top_p": 0.95,
        "top_k": 20,
        "min_p": 0.0,
        "presence_penalty": 0.0,
        "repeat_penalty": 1.0,
        "enable_thinking": True,
    },

    # ===============================
    # INSTRUCT MODE - GENERAL
    # ===============================
    "instruct_general": {
        "temperature": 0.7,
        "top_p": 0.8,
        "top_k": 20,
        "min_p": 0.0,
        "presence_penalty": 1.5,
        "repeat_penalty": 1.0,
        "enable_thinking": False,
    },

    # ===============================
    # INSTRUCT MODE - REASONING
    # ===============================
    "instruct_reasoning": {
        "temperature": 1.0,
        "top_p": 0.95,
        "top_k": 20,
        "min_p": 0.0,
        "presence_penalty": 1.5,
        "repeat_penalty": 1.0,
        "enable_thinking": False,
    },
}

import os




def load_llama_model(model_path, enable_thinking):

    key = (model_path, enable_thinking)

    with CACHE_LOCK:

        if key in MODEL_CACHE:
            return MODEL_CACHE[key]

        print("Loading model:", key)

        llm = Llama(
            model_path=model_path,
            flash_attn=True,
            n_gpu_layers=-1,
            n_batch=2048,
            n_ctx=8196,
            n_threads=os.cpu_count(),
            n_threads_batch=os.cpu_count(),
            use_mmap=True,
            use_mlock=False,
            mul_mat_q=True,
            chat_handler=Qwen35ChatHandler(
                clip_model_path="models/mmproj-BF16.gguf",
                enable_thinking=enable_thinking,
                image_min_tokens=1024,
            ),
        )

        MODEL_CACHE[key] = llm

        return llm
# @lru_cache(maxsize=1)  # adjust if you want to keep multiple models
# def load_llama_model(
#     model_path: str,
#     enable_thinking: bool,
# ):
#     print("🔹 Loading model ONCE:", model_path, "thinking=", enable_thinking)

#     return Llama(
#         model_path=model_path,
#         flash_attn=True,
#         n_gpu_layers=-1,
#         n_batch=2048,
#         n_ctx=8196,
#         n_threads=os.cpu_count(),
#         n_threads_batch=os.cpu_count(),
#         use_mlock=False,
#         use_mmap=True,
#         mul_mat_q=True,
#         chat_handler=Qwen35ChatHandler(
#             clip_model_path="models/mmproj-BF16.gguf",
#             enable_thinking=enable_thinking,
#             image_min_tokens=1024,
#         ),
#     )
@spaces.GPU(duration=70)
def respond(
    message,
    history: list[tuple[str, str]],
    model,
    system_message,
    max_tokens,
    temperature,
    top_p,
    top_k,
    repeat_penalty,
):
    # chat_template = MessagesFormatterType.GEMMA_2
    chat_template = MessagesFormatterType.CHATML
    # chat_template = 

    # global llm
    # global llm_model
    global MODES
    try:
        j_r= str_to_json(message)
        messages= j_r['messages']
        max_token= j_r['max_token']
        try:
            mode = j_r['mode']
        except:
            mode="thinking_general"
    except Exception as e:
        print(e)
        print("not valid json")
        max_token=8196
        mode="instruct_general"
        messages=[
            {
                "role":"user",
                "content":str(message)
            }
        ]
        
    final_messages=[]

    for message in messages:
        content= message['content']
        if isinstance(content, str):
            final_messages.append(message)
            continue
        ####   its a list so it contains images 
        tmp_list= []
        for item in content:
            if item['type']=="text":
                tmp_list.append(item)
                continue
            #### its a definitely a image
            if item['type']=="image":
                try:
                    if item['url']:
                        data_uri= handle_image_input(item['url'])
                        tmp_list.append({"type": "image_url", "image_url": {"url": data_uri}})
                except Exception as e:
                    print(e)
                try:
                    if item['uri']:
                        data_uri= handle_image_input(item['url'])
                        tmp_list.append({"type": "image_url", "image_url": {"url": data_uri}})
                except Exception as e:
                    print(e)
        
                    
        final_messages.append({"role": message['role'], "content": tmp_list})      
                    
    
    
    # if llm is None or llm_model != model:
    config = MODES[mode]
    # llm = Llama(
    #        model_path=f"models/{model}",
    #          flash_attn=True,  
    #         n_gpu_layers=-1,
    #             n_batch=2048,        # increase
    #             n_ctx= 8196,          # reduce if you don’t need 8k
    #             # n_threads= os.cpu_count(),        # set to your CPU cores
    #              n_threads=os.cpu_count(),        # use all cores
    #             n_threads_batch=os.cpu_count(),
    #             use_mlock=False,
    #             # verbose=True,
    #                 mul_mat_q=True,
    #              use_mmap=True,
    #        chat_handler=Qwen35ChatHandler(
    #   clip_model_path=f"models/mmproj-BF16.gguf",
    #      enable_thinking= config["enable_thinking"],
    #            image_min_tokens=1024, # Note: Qwen-VL models require at minimum 1024 image tokens to function correctly on bbox grounding tasks
    # ),
    #     )
    # llm_model = model
    llm = load_llama_model(
            model_path=f"models/{model}",
            enable_thinking=config["enable_thinking"],
            )

    x=llm.create_chat_completion(
      messages = final_messages,
        max_tokens= max_token,
        temperature=config["temperature"],
                top_p=config["top_p"],
                top_k=config["top_k"],
                min_p=config["min_p"],
                # presence_penalty=config["presence_penalty"],
                repeat_penalty=config["repeat_penalty"]
        )
    print(x)
    print(x)
    x= x['choices'][0]['message']['content']
    yield str(x)

    # provider = LlamaCppPythonProvider(llm)

    # agent = LlamaCppAgent(
    #     provider,
    #     system_prompt=f"{system_message}",
    #     predefined_messages_formatter_type=chat_template,
    #     debug_output=False
    # )
    
    # settings = provider.get_provider_default_settings()
    # settings.temperature = temperature
    # settings.top_k = top_k
    # settings.top_p = top_p
    # settings.max_tokens = max_tokens
    # settings.repeat_penalty = repeat_penalty
    # # settings.stream = True
    # # settings.reasoning_effort ="low"

    # messages = BasicChatHistory()

    # for msn in history:
    #     user = {
    #         'role': Roles.user,
    #         'content': msn[0]
    #     }
    #     assistant = {
    #         'role': Roles.assistant,
    #         'content': msn[1]
    #     }
    #     messages.add_message(user)
    #     messages.add_message(assistant)
    
    # stream = agent.get_chat_response(
    #     message,
    #     # llm_sampling_settings=settings,
    #     chat_history=messages,
    #     # returns_streaming_generator=True,
    #     print_output=False
    # )
    
    # outputs = ""
    # for output in stream:
    #     outputs += output
    #     yield outputs

description = """<p align="center">Defaults to 2B (you can switch to 9B or 27B from additional inputs)</p>
<p><center>
<a href="https://huggingface.co/google/gemma-2-27b-it" target="_blank">[27B it Model]</a>
<a href="https://huggingface.co/google/gemma-2-9b-it" target="_blank">[9B it Model]</a>
<a href="https://huggingface.co/google/gemma-2-2b-it" target="_blank">[2B it Model]</a>
<a href="https://huggingface.co/bartowski/gemma-2-27b-it-GGUF" target="_blank">[27B it Model GGUF]</a>
<a href="https://huggingface.co/bartowski/gemma-2-9b-it-GGUF" target="_blank">[9B it Model GGUF]</a>
<a href="https://huggingface.co/google/gemma-2-2b-it-GGUF" target="_blank">[2B it Model GGUF]</a>
</center></p>
"""
import gradio as gr

demo = gr.ChatInterface(
    fn=respond,
    additional_inputs=[
        gr.Dropdown(
            [
                # "gemma-2-9b-it-Q5_K_M.gguf",
                # "gemma-2-27b-it-Q5_K_M.gguf",
                # # "2b_it_v2.gguf",
                # "GLM-4.7-Flash-UD-Q8_K_XL.gguf",
                # "Qwen3-Coder-Next-Q4_K_M.gguf",
                # "gpt-oss-20b-Q4_K_M.gguf",
                # "Qwen3-Next-80B-A3B-Instruct-Q4_K_M.gguf",
                "UD-Q3_K_XL/Qwen3.5-122B-A10B-UD-Q3_K_XL-00001-of-00003.gguf",
                # "Qwen3-VL-32B-Thinking-Q8_0.gguf",
                # "Q8_0/gpt-oss-120b-Q8_0-00001-of-00002.gguf"
            ],
            value= "UD-Q3_K_XL/Qwen3.5-122B-A10B-UD-Q3_K_XL-00001-of-00003.gguf",
            label="Model",
        ),
        gr.Textbox(
            value="You are a helpful assistant.",
            label="System message",
        ),
        gr.Slider(1, 4096, value=2048, step=1, label="Max tokens"),
        gr.Slider(0.1, 4.0, value=0.7, step=0.1, label="Temperature"),
        gr.Slider(0.1, 1.0, value=0.95, step=0.05, label="Top-p"),
        gr.Slider(0, 100, value=40, step=1, label="Top-k"),
        gr.Slider(0.0, 2.0, value=1.1, step=0.1, label="Repetition penalty"),
    ],
    title="Chat with Gemma 2 using llama.cpp",
    description=description,
)


# demo.launch()

if __name__ == "__main__":
    demo.launch()