hf_space/app.py

"""
MINDI 1.5 Vision-Coder — HuggingFace Space (ZeroGPU)

Uses ZeroGPU for free A100 access (40GB VRAM).
Full bf16 model — NO quantization.

IMPORTANT: All .to("cuda") calls MUST be inside @spaces.GPU decorated functions.
ZeroGPU only provides GPU access inside those functions.
"""

import os
import re
import gc
import json
import torch
import spaces
import gradio as gr
from pathlib import Path
from PIL import Image
from huggingface_hub import snapshot_download

# ── Global model reference ──────────────────────────────
MODEL = None
TOKENIZER = None
IS_LOADED = False

# ── Special token definitions ───────────────────────────
SECTION_TOKENS = {
    "thinking":  ("<|think_start|>",    "<|think_end|>"),
    "file":      ("<|file_start|>",     "<|file_end|>"),
    "code":      ("<|code_start|>",     "<|code_end|>"),
    "critique":  ("<|critique_start|>", "<|critique_end|>"),
    "suggest":   ("<|suggest_start|>",  "<|suggest_end|>"),
    "search":    ("<|search_start|>",   "<|search_end|>"),
    "error":     ("<|error_start|>",    "<|error_end|>"),
    "fix":       ("<|fix_start|>",      "<|fix_end|>"),
}


def parse_output(text: str) -> dict:
    result = {}
    for section, (start_tok, end_tok) in SECTION_TOKENS.items():
        pattern = re.escape(start_tok) + r"(.*?)" + re.escape(end_tok)
        matches = re.findall(pattern, text, flags=re.DOTALL)
        if matches:
            result[section] = [m.strip() for m in matches]
    return result


_CHAT_TOKEN_PATTERN = re.compile(
    r"<\|(?:im_start|im_end|endoftext|fim_prefix|fim_middle|fim_suffix|fim_pad|repo_name|file_sep)\|>"
)
# Match a role line ONLY if it stands alone at the very start of the text
# followed by an explicit newline. The previous '\s*' wildcard could swallow
# leading content when the model emitted weird sequences in the vision path.
_ROLE_PREFIX_PATTERN = re.compile(r"^(?:system|user|assistant)\n")


def clean_output(text: str) -> str:
    """Strip Qwen chat-template artifacts and any leading role prefix."""
    if os.environ.get("MINDI_DEBUG_RAW") == "1":
        print(f"[clean_output] RAW ({len(text)} chars): {text!r}")
    text = _CHAT_TOKEN_PATTERN.sub("", text)
    # Apply role-prefix strip up to twice: handles the vision-path case where
    # the model occasionally emits 'assistant\n' followed by stray noise like
    # an extra 'user\n' before the real reply.
    for _ in range(2):
        new_text = _ROLE_PREFIX_PATTERN.sub("", text, count=1)
        if new_text == text:
            break
        text = new_text
    return text.strip()


def download_checkpoint():
    """Download checkpoint (CPU-safe, no CUDA needed)."""
    ckpt_dir = Path("/tmp/mindi_ckpt")
    if not ckpt_dir.exists():
        print("[MINDI] Downloading checkpoint from HuggingFace...")
        snapshot_download(
            "Mindigenous/MINDI-1.5-Vision-Coder",
            local_dir=str(ckpt_dir),
            allow_patterns=[
                "checkpoints/phase3_final/**",
                "data/tokenizer/**",
            ],
        )
        print("[MINDI] Download complete")
    return ckpt_dir


def load_tokenizer(ckpt_dir):
    """Load tokenizer (CPU-safe)."""
    global TOKENIZER
    if TOKENIZER is not None:
        return

    from transformers import AutoTokenizer
    tok_path = ckpt_dir / "data" / "tokenizer" / "mindi_tokenizer"
    if not tok_path.exists():
        tok_path = "Qwen/Qwen2.5-Coder-7B-Instruct"
    TOKENIZER = AutoTokenizer.from_pretrained(str(tok_path), trust_remote_code=True)
    print(f"[MINDI] Tokenizer loaded: {len(TOKENIZER)} tokens")


def load_model_to_gpu(ckpt_dir):
    """Load model TO GPU — MUST be called inside @spaces.GPU function."""
    global MODEL, IS_LOADED

    if IS_LOADED:
        return

    from transformers import (
        AutoModelForCausalLM,
        CLIPVisionModel, CLIPImageProcessor,
    )
    from peft import PeftModel
    import torch.nn as nn

    print("[MINDI] Loading full bf16 model to GPU...")

    # Base LLM
    base_model = AutoModelForCausalLM.from_pretrained(
        "Qwen/Qwen2.5-Coder-7B-Instruct",
        torch_dtype=torch.bfloat16,
        device_map="auto",
        trust_remote_code=True,
    )
    base_model.resize_token_embeddings(len(TOKENIZER))
    print("[MINDI] Base model loaded (bf16)")

    # LoRA
    lora_path = ckpt_dir / "checkpoints" / "phase3_final" / "lora"
    if lora_path.exists():
        base_model = PeftModel.from_pretrained(base_model, str(lora_path))
        print("[MINDI] LoRA loaded")

    # CLIP
    clip_model = CLIPVisionModel.from_pretrained(
        "openai/clip-vit-large-patch14",
        torch_dtype=torch.bfloat16,
    ).to("cuda").eval()
    clip_processor = CLIPImageProcessor.from_pretrained("openai/clip-vit-large-patch14")
    print("[MINDI] CLIP loaded")

    # Vision projection
    class VisionProjection(nn.Module):
        def __init__(self, clip_dim=1024, llm_dim=3584):
            super().__init__()
            self.projection = nn.Linear(clip_dim, llm_dim)
            self.layer_norm = nn.LayerNorm(llm_dim)
        def forward(self, x):
            return self.layer_norm(self.projection(x))

    vision_proj = VisionProjection().to("cuda").to(torch.bfloat16)
    vision_ckpt = ckpt_dir / "checkpoints" / "phase3_final" / "vision"
    if vision_ckpt.exists():
        for f in vision_ckpt.iterdir():
            if f.suffix in (".pt", ".bin", ".safetensors"):
                state = torch.load(f, map_location="cuda", weights_only=True)
                vision_proj.load_state_dict(state, strict=False)
                print("[MINDI] Vision projection loaded")
                break

    # Fusion
    class SimpleFusion(nn.Module):
        def __init__(self, hidden_size=3584):
            super().__init__()
            self.visual_gate = nn.Linear(hidden_size, hidden_size)
            self.text_gate_param = nn.Parameter(torch.zeros(1))
            self.layer_norm = nn.LayerNorm(hidden_size)
        def forward(self, text_embeds, visual_embeds):
            gated_visual = torch.sigmoid(self.visual_gate(visual_embeds)) * visual_embeds
            combined = torch.cat([gated_visual, text_embeds], dim=1)
            return self.layer_norm(combined)

    fusion = SimpleFusion().to("cuda").to(torch.bfloat16)
    fusion_file = ckpt_dir / "checkpoints" / "phase3_final" / "fusion" / "fusion.pt"
    if fusion_file.exists():
        state = torch.load(fusion_file, map_location="cuda", weights_only=True)
        fusion.load_state_dict(state, strict=False)
        print("[MINDI] Fusion loaded")

    MODEL = {
        "llm": base_model,
        "clip": clip_model,
        "clip_processor": clip_processor,
        "vision_proj": vision_proj,
        "fusion": fusion,
    }
    IS_LOADED = True

    vram = torch.cuda.memory_allocated() / 1e9
    print(f"[MINDI] Ready! VRAM: {vram:.1f} GB")


# Pre-download checkpoint and tokenizer at startup (CPU-safe)
_ckpt_dir = download_checkpoint()
load_tokenizer(_ckpt_dir)


SYSTEM_MSG = (
    "You are MINDI 1.5 Vision-Coder, an AI coding assistant created by MINDIGENOUS.AI.\n"
    "Your name is MINDI (Mindigenous Intelligence). You are NOT GPT, GPT-4, ChatGPT, "
    "Claude, Gemini, Llama, or Qwen. If asked who you are, who built you, or what model "
    "you are, answer: 'I am MINDI 1.5 Vision-Coder, built by MINDIGENOUS.AI.'\n"
    "Architecture: Qwen2.5-Coder-7B base, fine-tuned on 1.48M coding examples and 50K "
    "UI/web screenshots, with a CLIP-Large vision encoder fused for image understanding.\n"
    "\n"
    "OUTPUT RULES (critical — follow strictly):\n"
    "1. Wrap every code block in a fenced block with the correct language tag:\n"
    "   `tsx` for Next.js or React with TypeScript, `jsx` for React with JavaScript,\n"
    "   `html` for HTML documents, `vue` for Vue SFCs, `py` / `js` / `css` / `json`\n"
    "   / `sql` / `bash` for everything else.\n"
    "2. Emit the ACTUAL application code, not a scaffolding script. Never write\n"
    "   `fs.mkdirSync(...)`, `execSync('npx create-next-app')`, `fs.writeFileSync`\n"
    "   loops, or any code whose job is to create other files — write the files'\n"
    "   contents directly in fenced blocks.\n"
    "3. Never use placeholders, TODOs, '…add more here', or '// rest of code'.\n"
    "   Every block you emit must be a complete, runnable artifact.\n"
    "\n"
    "FRAMEWORK AWARENESS — match the output shape to what the user named:\n"
    "- User says 'Next.js' → emit ONE file: `app/page.tsx` (or `pages/index.tsx`\n"
    "  for pages-router). Start with `'use client';` ONLY if you use hooks\n"
    "  (useState, useEffect). Export a default function component. Use Tailwind\n"
    "  classes directly — do NOT emit `next.config.js`, `tsconfig.json`, or\n"
    "  `package.json`; just the page component file.\n"
    "- User says 'React' → emit ONE functional component in a `.tsx` or `.jsx`\n"
    "  file, exported as default. No build config.\n"
    "- User says 'Vue' → emit ONE `.vue` single-file component with `<template>`,\n"
    "  `<script setup>`, and `<style>` blocks.\n"
    "- User says 'HTML' or names no framework → emit ONE complete\n"
    "  `<!DOCTYPE html>` document with `<style>` and `<script>` inline. If the\n"
    "  user mentions Tailwind, include `<script src=\"https://cdn.tailwindcss.com\">\n"
    "  </script>` in the <head>.\n"
    "- User says 'Python' / 'FastAPI' / 'Flask' → emit ONE `.py` file with all\n"
    "  imports at the top, routes / functions inline.\n"
    "\n"
    "BEHAVIOR:\n"
    "- When given an image, describe what you see in 1–2 sentences, then produce\n"
    "  code that implements it.\n"
    "- Use the conversation history — refer back to what the user told you.\n"
    "- Be direct. Code first, brief explanation after."
)

# Hard limits to keep prompt within Qwen-Coder's 8K context window.
_MAX_HISTORY_TURNS = 20
_MAX_HISTORY_CHARS = 12000


def _build_prompt(prompt: str, history: list | None) -> str:
    """Render the full chat-template string from system + history + new user msg."""
    parts = [f"<|im_start|>system\n{SYSTEM_MSG}<|im_end|>"]

    if history:
        # Take last N turns and trim from the front if total chars too big.
        trimmed = list(history)[-_MAX_HISTORY_TURNS:]
        total = 0
        kept: list[tuple[str, str]] = []
        for turn in reversed(trimmed):
            role = (turn or {}).get("role")
            content = ((turn or {}).get("content") or "").strip()
            if role not in ("user", "assistant") or not content:
                continue
            total += len(content)
            if total > _MAX_HISTORY_CHARS:
                break
            kept.append((role, content))
        for role, content in reversed(kept):
            parts.append(f"<|im_start|>{role}\n{content}<|im_end|>")

    parts.append(f"<|im_start|>user\n{prompt}<|im_end|>")
    parts.append("<|im_start|>assistant\n")
    return "\n".join(parts)


@spaces.GPU(duration=60)
def generate(prompt: str, image: Image.Image = None,
             temperature: float = 0.7, max_tokens: int = 2048,
             history: list | None = None) -> str:
    """Generate with full bf16 model. GPU allocated by ZeroGPU."""

    # Load model INSIDE the GPU function
    load_model_to_gpu(_ckpt_dir)

    formatted = _build_prompt(prompt, history)

    inputs = TOKENIZER(formatted, return_tensors="pt").to("cuda")

    # Vision path
    if image is not None and MODEL["clip"] is not None:
        try:
            pixel_values = MODEL["clip_processor"](
                images=image, return_tensors="pt"
            ).pixel_values.to("cuda", dtype=torch.bfloat16)

            with torch.no_grad():
                clip_out = MODEL["clip"](pixel_values).last_hidden_state
                visual_tokens = MODEL["vision_proj"](clip_out)
                text_embeds = MODEL["llm"].get_input_embeddings()(inputs["input_ids"])
                fused = MODEL["fusion"](text_embeds, visual_tokens)

                outputs = MODEL["llm"].generate(
                    inputs_embeds=fused,
                    attention_mask=torch.ones(fused.shape[:2], device="cuda"),
                    max_new_tokens=int(max_tokens),
                    temperature=max(float(temperature), 0.01),
                    do_sample=float(temperature) > 0,
                    pad_token_id=TOKENIZER.pad_token_id or TOKENIZER.eos_token_id,
                )
            return clean_output(TOKENIZER.decode(outputs[0], skip_special_tokens=False))
        except Exception as e:
            print(f"[WARN] Vision failed: {e}")

    # Text-only path
    with torch.no_grad():
        outputs = MODEL["llm"].generate(
            **inputs,
            max_new_tokens=int(max_tokens),
            temperature=max(float(temperature), 0.01),
            do_sample=float(temperature) > 0,
            pad_token_id=TOKENIZER.pad_token_id or TOKENIZER.eos_token_id,
        )
    generated = outputs[:, inputs["input_ids"].shape[1]:]
    return clean_output(TOKENIZER.decode(generated[0], skip_special_tokens=False))


# ── Gradio endpoint ─────────────────────────────────────

def _coerce_history(raw) -> list | None:
    """Accept history as list[dict], or JSON-string list, or None."""
    if raw is None or raw == "":
        return None
    if isinstance(raw, str):
        try:
            raw = json.loads(raw)
        except Exception:
            return None
    if isinstance(raw, list):
        return raw
    return None


def chat_fn(message: str, image: Image.Image = None,
            temperature: float = 0.7, max_tokens: int = 2048,
            history=None) -> str:
    """Exposed via Gradio API — wraps generate.

    `history` is a list of {"role": "user"|"assistant", "content": str} dicts,
    typically the previous turns of this conversation. May also arrive as a
    JSON-encoded string when called via the raw HTTP API.
    """
    try:
        hist = _coerce_history(history)
        response = generate(message, image, temperature, max_tokens, hist)
        sections = parse_output(response)
        return json.dumps({"response": response, "sections": sections})
    except Exception as e:
        error_msg = str(e)
        is_quota = (
            "quota" in error_msg.lower()
            or "zerogpu" in error_msg.lower()
            or "unlogged user" in error_msg.lower()
        )
        if is_quota:
            return json.dumps({
                "response": f"⚠️ GPU quota exceeded. Please try again later or reduce Max Tokens. Error: {error_msg}",
                "sections": {"error": [error_msg]},
            })
        return json.dumps({
            "response": f"Error during generation: {error_msg}",
            "sections": {"error": [error_msg]},
        })


# ── Gradio App ──────────────────────────────────────────

with gr.Blocks(title="MINDI 1.5 API", theme=gr.themes.Soft(
    primary_hue="purple", secondary_hue="blue",
)) as demo:
    gr.Markdown("# 🧠 MINDI 1.5 Vision-Coder API\nFull bf16 on ZeroGPU A100 · No quantization")

    with gr.Row():
        with gr.Column(scale=3):
            prompt = gr.Textbox(label="Prompt", placeholder="Write code...", lines=4)
            image = gr.Image(label="Image (optional)", type="pil")
        with gr.Column(scale=1):
            temperature = gr.Slider(0, 2, value=0.7, step=0.1, label="Temperature")
            max_tokens = gr.Slider(128, 4096, value=2048, step=128, label="Max Tokens")
            submit_btn = gr.Button("Generate", variant="primary")

    history = gr.Textbox(
        label="History (JSON list of {role, content})",
        placeholder='[{"role":"user","content":"..."},{"role":"assistant","content":"..."}]',
        lines=4,
        value="",
    )
    output = gr.Textbox(label="Response (JSON)", lines=20)

    submit_btn.click(
        fn=chat_fn,
        inputs=[prompt, image, temperature, max_tokens, history],
        outputs=output,
    )

demo.queue()
demo.launch()