feat: add weight to model/

inference.py

@@ -0,0 +1,439 @@
+# -*- coding: utf-8 -*-
+"""
+VLLM Information Density & Visual Comfort Auditor
+功能：批量扫描文件夹内的图片，基于"Visual Comfort"标准判断图片是 Suitable 还是 Unsuitable。
+输出：终端统计报告 + 详细 JSON 结果文件
+
+使用方法:
+python rule_info_vllm.py \
+  --input_dir "/path/to/your/images" \
+  --model_path "/path/to/your/Qwen2.5-VL-7B-Instruct"
+"""
+
+import os
+import re
+import json
+import argparse
+import multiprocessing
+from pathlib import Path
+from typing import Dict, List, Optional, Any
+from tqdm import tqdm
+from PIL import Image, ImageFile
+from transformers import AutoProcessor
+from vllm import LLM, SamplingParams
+
+# ==========================================
+# 【核心配置】环境与并发设置
+# ==========================================
+os.environ['VLLM_WORKER_MULTIPROC_METHOD'] = 'spawn'
+
+# 防止部分图片因截断而报错
+ImageFile.LOAD_TRUNCATED_IMAGES = True
+
+# 支持的图片格式
+IMG_EXTS = {".jpg", ".jpeg", ".png", ".webp", ".bmp", ".tif", ".tiff"}
+
+# ==========================================
+# 【提示词工程】
+# ==========================================
+SYS_PROMPT_TEXT = """ """
+
+
+##用法示例：
+#SYS_PROMPT_TEXT 替换！
+
+# --- 1. EXQUISITENESS (精美度) ---
+EXQUISITENESS_SYSTEM_PROMPT = """You are a highly critical Senior Art Director and Visual Auditor. 
+Your task is to identify "Low-Quality, Amateur, or Overly Simplistic" advertising materials based on the "Exquisiteness" standard.
+You have ZERO TOLERANCE for "Cheap Templates" that lack professional depth and aesthetic cohesion.
+
+INPUT: One image and one natural-language question about aesthetic exquisiteness.
+
+YOUR TASK:
+1. Determine if the image is a **VIOLATION** (Unsuitable) or **SAFE** (Suitable) based on the criteria below.
+2. Output a JSON object containing a rigorous Chain-of-Thought ("think") and a precise classification label ("answer").
+
+OUTPUT FORMAT:
+Return EXACTLY two blocks, no extra text:
+<think>Detailed reasoning comparing visual features against BOTH unsuitable and suitable criteria (checking for template-like flatness vs. rich visual layers)...</think><answer>{"Answer": "<Suitable OR Unsuitable>", "Answer type": "Exquisiteness"}</answer>
+
+=========================================
+CRITERIA FOR 'UNSUITABLE' (VIOLATION / LOW QUALITY)
+=========================================
+1. **Simplistic & "Flat" Design (The "Template" Trap):**
+   - **Overly Basic:** The layout is overly basic: just a simple solid color block + a generic icon + plain text.
+   - **Lack of Depth:** The design feels like a "default" or "low-end" template with zero artistic polish, shadows, or texture.
+   - **Disconnected:** Elements feel placed randomly without visual cohesion.
+
+2. **Poor Overall Aesthetics:**
+   - **Low Fidelity:** The colors are muddy, the composition is unbalanced, or the material quality looks pixelated/unrefined.
+   - **Cheap Experience:** The image fails to provide a "premium" or "high-fidelity" visual experience.
+
+=========================================
+CRITERIA FOR 'SUITABLE' (NON-VIOLATION / GOOD DESIGN)
+=========================================
+1. **Rich Visual Layers:** Use of depth, professional lighting, shadows, and high-quality textures.
+2. **Professional Polish:** The image features a coherent color palette and clear visual hierarchy that feels "designed" rather than "assembled".
+3. **Intentional Minimalism:** Even if the design is simple, it looks intentional, high-end, and balanced (not empty or basic).
+
+=========================================
+DECISION LOGIC
+=========================================
+- **Unsuitable**: If the design is simplistic, flat, low-effort, or looks like a cheap template.
+- **Suitable**: If the image features rich visual layers, depth, and looks polished/premium.
+"""
+
+# --- 2. PROFESSIONAL POLISH (后期质感) ---
+PROFESSIONAL_POLISH_SYSTEM_PROMPT = """You are a highly critical Senior Art Director specializing in Post-Production. 
+Your task is to evaluate "Post-Production Quality" for S-level splash ads.
+You have ZERO TOLERANCE for raw, unprocessed photos that look like amateur snapshots.
+
+INPUT: One image and one natural-language question about post-production quality.
+
+YOUR TASK:
+1. Determine if the image is a **VIOLATION** (Unsuitable) or **SAFE** (Suitable) based on the criteria below.
+2. Output a JSON object containing a rigorous Chain-of-Thought ("think") and a precise classification label ("answer").
+
+OUTPUT FORMAT:
+Return EXACTLY two blocks, no extra text:
+<think>Detailed reasoning comparing visual features against BOTH unsuitable and suitable criteria (analyzing lighting, color grading, depth of field, and texture)...</think><answer>{"Answer": "<Suitable OR Unsuitable>", "Answer type": "Professional Polish"}</answer>
+
+=========================================
+CRITERIA FOR 'UNSUITABLE' (VIOLATION / AMATEUR SNAPSHOT)
+=========================================
+1. **Lack of Professional Post-Processing (Raw Photo Look):**
+   - **Unprocessed:** The image appears to be a "Raw Photo" directly from a camera/phone without professional retouching.
+   - **Poor Technicals:** There is no deliberate optimization of Lighting (flat or messy light), Color (dull or unbalanced tones), or Depth of Field (lack of professional bokeh or focus control).
+
+2. **The "Amateur Snapshot" Aesthetic:**
+   - **Casual Vibe:** The image looks like something a "passerby" could easily capture. It lacks the sophisticated framing, high-end texture, and artistic polish required for premium advertising.
+   - **Cheap Feel:** The visual quality feels "Cheap" and fails to convey the premium value of the brand.
+
+3. **Absence of Value Conveyance:**
+   - **Visually Flat:** The image fails to evoke a sense of high quality or luxury. It is visually "flat" and does not use post-production techniques to guide emotions.
+
+=========================================
+CRITERIA FOR 'SUITABLE' (NON-VIOLATION / GOOD DESIGN)
+=========================================
+1. **Professional Polish:** The image has distinct, high-end color grading and lighting that creates a "Cinematic" or "Commercial" look.
+2. **Intentional Aesthetic:** Clear mastery of lighting (highlights/shadows), color harmony, and depth-of-field that elevates the subject.
+3. **Premium Value:** The overall visual execution feels expensive and exclusive, far beyond a casual photograph.
+4. **Cinematic Excellence:** Film stills or cinematic stills are ALWAYS classified as SUITABLE.
+
+=========================================
+DECISION LOGIC
+=========================================
+- **Unsuitable**: If the image looks like an unprocessed, amateur, or casual snapshot (Raw Photo).
+- **Suitable**: If the image shows professional polish, cinematic lighting, or is a film still.
+"""
+
+# --- 3. LAYOUT BREATHABILITY (布局呼吸感 - 已包含正向标准) ---
+LAYOUT_BREATHABILITY_SYSTEM_PROMPT = """You are a highly critical Senior Art Director specializing in Layout and Visual Hierarchy. 
+Your task is to identify "Suffocating Designs"—creative pieces where elements are too cramped, lack breathing room, or feel disorganized due to poor spacing.
+
+INPUT: One image and one natural-language question about layout spacing.
+
+YOUR TASK:
+1. Determine if the image is a **VIOLATION** (Unsuitable) or **SAFE** (Suitable) based on the criteria below.
+2. Output a JSON object containing a rigorous Chain-of-Thought ("think") and a precise classification label ("answer").
+
+OUTPUT FORMAT:
+Return EXACTLY two blocks, no extra text:
+<think>Detailed reasoning comparing visual features against BOTH unsuitable and suitable criteria (analyzing module gap, edge tension, and visual path)...</think><answer>{"Answer": "<Suitable OR Unsuitable>", "Answer type": "Layout Breathability Check"}</answer>
+
+=========================================
+CRITERIA FOR 'UNSUITABLE' (VIOLATION / SUFFOCATING DESIGN)
+=========================================
+1. **Lack of Breathing Room (Crowded Modules):**
+   - **Core Violation:** The main subject (product/hero), the headline text, and the logo are placed too close to each other.
+   - **Visual Feel:** The overall design feels "heavy" or "claustrophobic" because major modules lack sufficient negative space.
+   - **Small Print Nuance:** While secondary text (annotations) can have smaller gaps, they must NOT feel like they are "clinging" or "tangent" to other elements.
+
+2. **Edge Tension (贴边风险):**
+   - **Tangency:** Elements are unintentionally "touching" or "tangent" to each other or the canvas border without intentional overlapping (creating uncomfortable tension).
+
+3. **Information Overload:**
+   - **Clutter:** The layout is filled with too many text blocks or icons with no clear separation.
+   - **No Visual Path:** The eye doesn't know where to rest because every element is competing for attention and space simultaneously.
+
+=========================================
+CRITERIA FOR 'SUITABLE' (NON-VIOLATION / GOOD DESIGN)
+=========================================
+1. **Generous White Space:** Clear and deliberate separation between the headline, the hero subject, and the footer information.
+2. **Structured Layout:** Elements follow a clear grid or intentional alignment that allows the design to "breathe" while maintaining a strong hierarchy.
+3. **Intentional Overlap:** If elements overlap, it looks artistic, layered, and deliberate (e.g., text weaving behind a subject), not accidental or messy.
+
+=========================================
+DECISION LOGIC
+=========================================
+- **Unsuitable**: If the layout feels squeezed, crowded, has uncomfortable edge tension, or lacks a visual path.
+- **Suitable**: If the layout has generous negative space, structured alignment, and allows the eye to travel comfortably.
+"""
+
+# --- 4. TEXT LEGIBILITY (文字易读性与排布) ---
+TEXT_LEGIBILITY_SYSTEM_PROMPT = """You are a highly critical Senior Art Director. 
+Your task is to evaluate "Information Accessibility" to ensure advertising copy is instantly readable and strategically placed.
+
+INPUT: One image and one natural-language question about text legibility.
+
+YOUR TASK:
+1. Determine if the image is a **VIOLATION** (Unsuitable) or **SAFE** (Suitable) based on the criteria below.
+2. Output a JSON object containing a rigorous Chain-of-Thought ("think") and a precise classification label ("answer").
+
+OUTPUT FORMAT:
+Return EXACTLY two blocks, no extra text:
+<think>Detailed reasoning comparing visual features against BOTH unsuitable and suitable criteria (analyzing contrast, background interference, and placement logic)...</think><answer>{"Answer": "<Suitable OR Unsuitable>", "Answer type": "Text Legibility and Placement"}</answer>
+
+=========================================
+CRITERIA FOR 'UNSUITABLE' (VIOLATION / POOR LEGIBILITY)
+=========================================
+1. **Direct Overlay on Complex Background:**
+   - **Interference:** Text is placed over "busy" areas (faces, complex textures, high-contrast patterns) where background details "cut through" the strokes.
+   - **No Protection:** The text lacks professional treatments (no solid backing, masks, strokes, or drop shadows) to separate it from the noisy background.
+
+2. **Poor Visual Catchiness (Dead Zones):**
+   - **Placement:** The primary marketing message (the "Hook") is placed in a "visual dead zone" (extreme edges or corners) where the eye does not naturally land.
+   - **Weak Presence:** Important copy is too small or lacks contrast relative to its position, failing to be eye-catching.
+
+3. **Low Contrast / Visual Camouflage:**
+   - **Blending:** Text color is too similar to background colors, causing it to "camouflage".
+   - **No Hierarchy:** There is no clear visual hierarchy; the eye has to "search" or struggle to find the text.
+
+4. **Unclear Small Text:**
+   - **Illegible:** Footnotes, disclaimers, or annotations are buried in background noise and are difficult to read (excluding text naturally on product packaging).
+
+=========================================
+CRITERIA FOR 'SUITABLE' (NON-VIOLATION / HIGH ACCESSIBILITY)
+=========================================
+1. **Clean Placement:** Main text is placed on a "clean" area of the image (e.g., sky, plain wall, or a blurred background) utilizing negative space.
+2. **Professional Treatment:** Even if the background is complex, the text uses solid containers, high-contrast strokes, masks, or heavy shadows to ensure perfect, instant legibility.
+3. **Strategic Positioning:** Key messages are placed in focal points, not hidden in corners.
+4. **Cinematic Exception:** Film stills or cinematic captures are ALWAYS classified as SUITABLE (Cinematic Excellence).
+
+=========================================
+DECISION LOGIC
+=========================================
+- **Unsuitable**: If text is hard to read due to low contrast, complex background interference, or placement in dead zones.
+- **Suitable**: If text is instantly readable due to clean placement, professional contrast treatments (masks/shadows), or if it is a cinematic still.
+"""
+
+# ==========================================
+# 辅助函数
+# ==========================================
+
+def collect_images(input_dir: Path) -> List[Dict[str, str]]:
+    """扫描目录下所有图片"""
+    if not input_dir.exists():
+        raise FileNotFoundError(f"Input directory not found: {input_dir}")
+    
+    files = [p for p in input_dir.iterdir() if p.is_file() and p.suffix.lower() in IMG_EXTS]
+    files.sort()
+    
+    print(f"[Info] Found {len(files)} images in {input_dir}")
+    return [{"path": str(p), "filename": p.name} for p in files]
+
+def parse_llm_output(text: str) -> Dict[str, Any]:
+    """
+    解析模型输出，提取 <think> 和 <answer> 标签中的内容
+    目标格式: <think>...</think><answer>{"Answer": "Suitable/Unsuitable", ...}</answer>
+    """
+    default_res = {
+        "label": "Parse Error", 
+        "think": "No reasoning found", 
+        "raw": text
+    }
+    
+    if not text:
+        return default_res
+
+    # 1. 提取 <think> 内容
+    think_match = re.search(r'<think>(.*?)</think>', text, re.DOTALL)
+    think_content = think_match.group(1).strip() if think_match else ""
+
+    # 2. 提取 <answer> 内容
+    answer_match = re.search(r'<answer>(.*?)</answer>', text, re.DOTALL)
+    
+    extracted_label = "Parse Error"
+    
+    if answer_match:
+        json_str = answer_match.group(1).strip()
+        try:
+            # 尝试解析 JSON
+            data = json.loads(json_str)
+            # 获取 Answer 字段 (Suitable 或 Unsuitable)
+            raw_ans = data.get("Answer", "")
+            
+            # 归一化处理
+            if "unsuitable" in raw_ans.lower():
+                extracted_label = "Unsuitable"
+            elif "suitable" in raw_ans.lower():
+                extracted_label = "Suitable"
+            else:
+                extracted_label = raw_ans # 如果是其他奇怪的内容，保留原样
+                
+        except json.JSONDecodeError:
+            # 如果 JSON 解析失败，尝试暴力匹配字��串
+            if "Unsuitable" in json_str:
+                extracted_label = "Unsuitable"
+            elif "Suitable" in json_str:
+                extracted_label = "Suitable"
+    else:
+        # 兜底：如果没有 answer 标签，直接在全文搜
+        if "Unsuitable" in text:
+            extracted_label = "Unsuitable"
+        elif "Suitable" in text:
+            extracted_label = "Suitable"
+
+    return {
+        "label": extracted_label,
+        "think": think_content,
+        "raw": text
+    }
+
+def prepare_vllm_inputs(batch_meta: List[Dict], processor) -> List[Dict]:
+    """构建 vLLM 输入格式"""
+    vllm_inputs = []
+    # 根据新的 Prompt 设置对应的用户提问
+    user_query = "Is this image visually comfortable and suitable for information display?"
+    
+    for item in batch_meta:
+        img_path = item["path"]
+        try:
+            image_obj = Image.open(img_path).convert("RGB")
+            
+            messages = [
+                {"role": "system", "content": [{"type": "text", "text": SYS_PROMPT_TEXT}]},
+                {"role": "user", "content": [
+                    {"type": "image", "image": img_path},
+                    {"type": "text", "text": user_query}
+                ]}
+            ]
+            
+            prompt_text = processor.apply_chat_template(
+                messages, tokenize=False, add_generation_prompt=True
+            )
+            
+            vllm_inputs.append({
+                "prompt": prompt_text,
+                "multi_modal_data": {"image": image_obj}
+            })
+        except Exception as e:
+            print(f"[Warning] Failed to load {img_path}: {e}")
+            vllm_inputs.append(None)
+            
+    return vllm_inputs
+
+# ==========================================
+# 主程序
+# ==========================================
+
+def main():
+    parser = argparse.ArgumentParser(description="AI Visual Comfort Auditor")
+    parser.add_argument("--input_dir", type=str, required=True, help="Folder containing images to check")
+    parser.add_argument("--model_path", type=str, required=True, help="Path to local Qwen-VL model")
+    parser.add_argument("--batch_size", type=int, default=128, help="Inference batch size")
+    parser.add_argument("--tp_size", type=int, default=4, help="Tensor Parallel size")
+    args = parser.parse_args()
+
+    input_path = Path(args.input_dir)
+    meta_data = collect_images(input_path)
+    
+    if not meta_data:
+        print("[Info] No images found. Exiting.")
+        return
+
+    # ---------------------------
+    # 初始化模型
+    # ---------------------------
+    print(f"\n[Init] Loading Model: {args.model_path}")
+    
+    llm = LLM(
+        model=args.model_path,
+        tokenizer=args.model_path,
+        trust_remote_code=True,
+        tensor_parallel_size=args.tp_size,
+        gpu_memory_utilization=0.90, 
+        max_model_len=8192, 
+        enforce_eager=True,
+        limit_mm_per_prompt={"image": 1}
+    )
+    
+    processor = AutoProcessor.from_pretrained(args.model_path, trust_remote_code=True)
+    
+    # 采样参数
+    sampling_params = SamplingParams(
+        temperature=0.7, # 稍微降低温度以获得更稳定的分类
+        max_tokens=1024,
+        top_p=0.9
+    )
+
+    # ---------------------------
+    # 批量推理
+    # ---------------------------
+    results = []
+    print(f"\n[Run] Starting Inference on {len(meta_data)} images...")
+    
+    for i in tqdm(range(0, len(meta_data), args.batch_size), desc="Processing Batches"):
+        batch_meta = meta_data[i : i + args.batch_size]
+        batch_inputs = prepare_vllm_inputs(batch_meta, processor)
+        
+        valid_inputs = [inp for inp in batch_inputs if inp is not None]
+        valid_indices = [idx for idx, inp in enumerate(batch_inputs) if inp is not None]
+        
+        if not valid_inputs:
+            continue
+            
+        outputs = llm.generate(valid_inputs, sampling_params=sampling_params, use_tqdm=False)
+        
+        for local_idx, out in enumerate(outputs):
+            original_meta = batch_meta[valid_indices[local_idx]]
+            generated_text = out.outputs[0].text
+            
+            # 解析结果
+            parsed = parse_llm_output(generated_text)
+            
+            results.append({
+                "filename": original_meta["filename"],
+                "path": original_meta["path"],
+                "label": parsed["label"],   # Suitable / Unsuitable
+                "think": parsed["think"],   # 思维链
+                "raw_output": generated_text
+            })
+
+    # ---------------------------
+    # 统计与输出
+    # ---------------------------
+    total = len(results)
+    unsuitable_count = sum(1 for r in results if r["label"] == "Unsuitable")
+    suitable_count = sum(1 for r in results if r["label"] == "Suitable")
+    error_count = total - unsuitable_count - suitable_count
+
+    unsuitable_rate = (unsuitable_count / total * 100) if total > 0 else 0
+    suitable_rate = (suitable_count / total * 100) if total > 0 else 0
+
+    print("\n" + "="*60)
+    print(f"AUDIT REPORT FOR: {input_path.name}")
+    print("="*60)
+    print(f"{'Total Images':<25}: {total}")
+    print("-" * 60)
+    print(f"{'UNSUITABLE (Violation)':<25}: {unsuitable_count}  ({unsuitable_rate:.2f}%)")
+    print(f"{'SUITABLE (Safe)':<25}: {suitable_count}  ({suitable_rate:.2f}%)")
+    print(f"{'Parse Errors':<25}: {error_count}")
+    print("="*60)
+
+    # 保存结果
+    output_file = input_path / f"audit_result_{input_path.name}.json"
+    try:
+        with open(output_file, "w", encoding="utf-8") as f:
+            json.dump(results, f, ensure_ascii=False, indent=2)
+        print(f"\n[Done] Detailed JSON report saved to:\n-> {output_file}")
+    except Exception as e:
+        print(f"[Error] Could not save JSON: {e}")
+
+if __name__ == "__main__":
+    try:
+        multiprocessing.set_start_method('spawn', force=True)
+    except RuntimeError:
+        pass
+    
+    main()

stage2_object_v2/added_tokens.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c0284b582e14987fbd3d5a2cb2bd139084371ed9acbae488829a1c900833c680
+size 707

stage2_object_v2/chat_template.jinja

@@ -0,0 +1,120 @@
+{%- if tools %}
+    {{- '<|im_start|>system\n' }}
+    {%- if messages[0].role == 'system' %}
+        {%- if messages[0].content is string %}
+            {{- messages[0].content }}
+        {%- else %}
+            {%- for content in messages[0].content %}
+                {%- if 'text' in content %}
+                    {{- content.text }}
+                {%- endif %}
+            {%- endfor %}
+        {%- endif %}
+        {{- '\n\n' }}
+    {%- endif %}
+    {{- "# Tools\n\nYou may call one or more functions to assist with the user query.\n\nYou are provided with function signatures within <tools></tools> XML tags:\n<tools>" }}
+    {%- for tool in tools %}
+        {{- "\n" }}
+        {{- tool | tojson }}
+    {%- endfor %}
+    {{- "\n</tools>\n\nFor each function call, return a json object with function name and arguments within <tool_call></tool_call> XML tags:\n<tool_call>\n{\"name\": <function-name>, \"arguments\": <args-json-object>}\n</tool_call><|im_end|>\n" }}
+{%- else %}
+    {%- if messages[0].role == 'system' %}
+        {{- '<|im_start|>system\n' }}
+        {%- if messages[0].content is string %}
+            {{- messages[0].content }}
+        {%- else %}
+            {%- for content in messages[0].content %}
+                {%- if 'text' in content %}
+                    {{- content.text }}
+                {%- endif %}
+            {%- endfor %}
+        {%- endif %}
+        {{- '<|im_end|>\n' }}
+    {%- endif %}
+{%- endif %}
+{%- set image_count = namespace(value=0) %}
+{%- set video_count = namespace(value=0) %}
+{%- for message in messages %}
+    {%- if message.role == "user" %}
+        {{- '<|im_start|>' + message.role + '\n' }}
+        {%- if message.content is string %}
+            {{- message.content }}
+        {%- else %}
+            {%- for content in message.content %}
+                {%- if content.type == 'image' or 'image' in content or 'image_url' in content %}
+                    {%- set image_count.value = image_count.value + 1 %}
+                    {%- if add_vision_id %}Picture {{ image_count.value }}: {% endif -%}
+                    <|vision_start|><|image_pad|><|vision_end|>
+                {%- elif content.type == 'video' or 'video' in content %}
+                    {%- set video_count.value = video_count.value + 1 %}
+                    {%- if add_vision_id %}Video {{ video_count.value }}: {% endif -%}
+                    <|vision_start|><|video_pad|><|vision_end|>
+                {%- elif 'text' in content %}
+                    {{- content.text }}
+                {%- endif %}
+            {%- endfor %}
+        {%- endif %}
+        {{- '<|im_end|>\n' }}
+    {%- elif message.role == "assistant" %}
+        {{- '<|im_start|>' + message.role + '\n' }}
+        {%- if message.content is string %}
+            {{- message.content }}
+        {%- else %}
+            {%- for content_item in message.content %}
+                {%- if 'text' in content_item %}
+                    {{- content_item.text }}
+                {%- endif %}
+            {%- endfor %}
+        {%- endif %}
+        {%- if message.tool_calls %}
+            {%- for tool_call in message.tool_calls %}
+                {%- if (loop.first and message.content) or (not loop.first) %}
+                    {{- '\n' }}
+                {%- endif %}
+                {%- if tool_call.function %}
+                    {%- set tool_call = tool_call.function %}
+                {%- endif %}
+                {{- '<tool_call>\n{"name": "' }}
+                {{- tool_call.name }}
+                {{- '", "arguments": ' }}
+                {%- if tool_call.arguments is string %}
+                    {{- tool_call.arguments }}
+                {%- else %}
+                    {{- tool_call.arguments | tojson }}
+                {%- endif %}
+                {{- '}\n</tool_call>' }}
+            {%- endfor %}
+        {%- endif %}
+        {{- '<|im_end|>\n' }}
+    {%- elif message.role == "tool" %}
+        {%- if loop.first or (messages[loop.index0 - 1].role != "tool") %}
+            {{- '<|im_start|>user' }}
+        {%- endif %}
+        {{- '\n<tool_response>\n' }}
+        {%- if message.content is string %}
+            {{- message.content }}
+        {%- else %}
+            {%- for content in message.content %}
+                {%- if content.type == 'image' or 'image' in content or 'image_url' in content %}
+                    {%- set image_count.value = image_count.value + 1 %}
+                    {%- if add_vision_id %}Picture {{ image_count.value }}: {% endif -%}
+                    <|vision_start|><|image_pad|><|vision_end|>
+                {%- elif content.type == 'video' or 'video' in content %}
+                    {%- set video_count.value = video_count.value + 1 %}
+                    {%- if add_vision_id %}Video {{ video_count.value }}: {% endif -%}
+                    <|vision_start|><|video_pad|><|vision_end|>
+                {%- elif 'text' in content %}
+                    {{- content.text }}
+                {%- endif %}
+            {%- endfor %}
+        {%- endif %}
+        {{- '\n</tool_response>' }}
+        {%- if loop.last or (messages[loop.index0 + 1].role != "tool") %}
+            {{- '<|im_end|>\n' }}
+        {%- endif %}
+    {%- endif %}
+{%- endfor %}
+{%- if add_generation_prompt %}
+    {{- '<|im_start|>assistant\n' }}
+{%- endif %}

stage2_object_v2/config.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b321be22f463f0af477f83cf7e63c566bb15645299208c6656a57ece7b9ffa87
+size 1613

stage2_object_v2/generation_config.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3ff2e83a0510cccccc85c8c96c7df4207985c098b526d798bc2bc68e50bb1a41
+size 199

stage2_object_v2/latest

@@ -0,0 +1 @@
+global_step420

stage2_object_v2/model.safetensors.index.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:dbdf49cc47bf9a028d0b1aa914b25401527572f03b092c8dfd9d428c9172783f
+size 67791

stage2_object_v2/preprocessor_config.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:93585062a80db5e8ca038efc7726a3e6411d9db948472d81d63c6303993be8c5
+size 782

stage2_object_v2/rng_state_0.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7272d7347df7a3eda0dcde7c67b3d4f9ff25e61e90c9673fc43693fe2e45be2f
+size 15365

stage2_object_v2/rng_state_1.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0eb1f7b1acfd920c8cdef076eee48ccfec7d3ed4d8c5d83c2592fbbb4a4f9b38
+size 15429

stage2_object_v2/rng_state_2.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5664fd5079c69dc8edcc429ded884f07d45c771f27a5a65fe1ad751348f9e1de
+size 15429

stage2_object_v2/rng_state_3.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0a3f17c4e252e5b7e99f3573c3bca992070f9ef436dd578ee9e88c333a94cef0
+size 15429

stage2_object_v2/scheduler.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e3433917e330a5e58b0185a3812a9e2c943fa2c157c04fecc317779eb366e4e6
+size 1465

stage2_object_v2/special_tokens_map.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:76862e765266b85aa9459767e33cbaf13970f327a0e88d1c65846c2ddd3a1ecd
+size 613

stage2_object_v2/tokenizer.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:aeb13307a71acd8fe81861d94ad54ab689df773318809eed3cbe794b4492dae4
+size 11422654

stage2_object_v2/tokenizer_config.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cf43a5bf1a49ee69ecced02f419b169e72559034dcf15af47cf775bd253830f0
+size 5472

stage2_object_v2/trainer_state.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b0e46e697ee453aa2a87361fbd2616515614af85b224376d525af005dea2e8eb
+size 9250

stage2_object_v2/training_args.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:314b5adbe54f49c9e00e177abc56e7953500d7fd77e68390e5a0608c2fb34a90
+size 8209

stage2_object_v2/video_preprocessor_config.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:59c5c9eb52182eb14c06ffb10ca9effd29adce5f238a95de23ca14a38dbd2cb1
+size 817

stage2_object_v2/vocab.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ca10d7e9fb3ed18575dd1e277a2579c16d108e32f27439684afa0e10b1440910
+size 2776833

stage2_object_v2/zero_to_fp32.py

@@ -0,0 +1,760 @@
+#!/usr/bin/env python
+
+# Copyright (c) Microsoft Corporation.
+# SPDX-License-Identifier: Apache-2.0
+
+# DeepSpeed Team
+
+# This script extracts fp32 consolidated weights from a zero 1, 2 and 3 DeepSpeed checkpoints. It gets
+# copied into the top level checkpoint dir, so the user can easily do the conversion at any point in
+# the future. Once extracted, the weights don't require DeepSpeed and can be used in any
+# application.
+#
+# example:
+#   python zero_to_fp32.py . output_dir/
+#   or
+#   python zero_to_fp32.py . output_dir/ --safe_serialization
+
+import argparse
+import torch
+import glob
+import math
+import os
+import re
+import gc
+import json
+import numpy as np
+from tqdm import tqdm
+from collections import OrderedDict
+from dataclasses import dataclass
+
+# while this script doesn't use deepspeed to recover data, since the checkpoints are pickled with
+# DeepSpeed data structures it has to be available in the current python environment.
+from deepspeed.utils import logger
+from deepspeed.checkpoint.constants import (DS_VERSION, OPTIMIZER_STATE_DICT, SINGLE_PARTITION_OF_FP32_GROUPS,
+                                            FP32_FLAT_GROUPS, ZERO_STAGE, PARTITION_COUNT, PARAM_SHAPES, BUFFER_NAMES,
+                                            FROZEN_PARAM_SHAPES, FROZEN_PARAM_FRAGMENTS)
+
+
+@dataclass
+class zero_model_state:
+    buffers: dict()
+    param_shapes: dict()
+    shared_params: list
+    ds_version: int
+    frozen_param_shapes: dict()
+    frozen_param_fragments: dict()
+
+
+debug = 0
+
+# load to cpu
+device = torch.device('cpu')
+
+
+def atoi(text):
+    return int(text) if text.isdigit() else text
+
+
+def natural_keys(text):
+    '''
+    alist.sort(key=natural_keys) sorts in human order
+    http://nedbatchelder.com/blog/200712/human_sorting.html
+    (See Toothy's implementation in the comments)
+    '''
+    return [atoi(c) for c in re.split(r'(\d+)', text)]
+
+
+def get_model_state_file(checkpoint_dir, zero_stage):
+    if not os.path.isdir(checkpoint_dir):
+        raise FileNotFoundError(f"Directory '{checkpoint_dir}' doesn't exist")
+
+    # there should be only one file
+    if zero_stage <= 2:
+        file = os.path.join(checkpoint_dir, "mp_rank_00_model_states.pt")
+    elif zero_stage == 3:
+        file = os.path.join(checkpoint_dir, "zero_pp_rank_0_mp_rank_00_model_states.pt")
+
+    if not os.path.exists(file):
+        raise FileNotFoundError(f"can't find model states file at '{file}'")
+
+    return file
+
+
+def get_checkpoint_files(checkpoint_dir, glob_pattern):
+    # XXX: need to test that this simple glob rule works for multi-node setup too
+    ckpt_files = sorted(glob.glob(os.path.join(checkpoint_dir, glob_pattern)), key=natural_keys)
+
+    if len(ckpt_files) == 0:
+        raise FileNotFoundError(f"can't find {glob_pattern} files in directory '{checkpoint_dir}'")
+
+    return ckpt_files
+
+
+def get_optim_files(checkpoint_dir):
+    return get_checkpoint_files(checkpoint_dir, "*_optim_states.pt")
+
+
+def get_model_state_files(checkpoint_dir):
+    return get_checkpoint_files(checkpoint_dir, "*_model_states.pt")
+
+
+def parse_model_states(files):
+    zero_model_states = []
+    for file in files:
+        state_dict = torch.load(file, map_location=device, weights_only=False)
+
+        if BUFFER_NAMES not in state_dict:
+            raise ValueError(f"{file} is not a model state checkpoint")
+        buffer_names = state_dict[BUFFER_NAMES]
+        if debug:
+            print("Found buffers:", buffer_names)
+
+        # recover just the buffers while restoring them to fp32 if they were saved in fp16
+        buffers = {k: v.float() for k, v in state_dict["module"].items() if k in buffer_names}
+        param_shapes = state_dict[PARAM_SHAPES]
+
+        # collect parameters that are included in param_shapes
+        param_names = []
+        for s in param_shapes:
+            for name in s.keys():
+                param_names.append(name)
+
+        # update with frozen parameters
+        frozen_param_shapes = state_dict.get(FROZEN_PARAM_SHAPES, None)
+        if frozen_param_shapes is not None:
+            if debug:
+                print(f"Found frozen_param_shapes: {frozen_param_shapes}")
+            param_names += list(frozen_param_shapes.keys())
+
+        # handle shared params
+        shared_params = [[k, v] for k, v in state_dict["shared_params"].items()]
+
+        ds_version = state_dict.get(DS_VERSION, None)
+
+        frozen_param_fragments = state_dict.get(FROZEN_PARAM_FRAGMENTS, None)
+
+        z_model_state = zero_model_state(buffers=buffers,
+                                         param_shapes=param_shapes,
+                                         shared_params=shared_params,
+                                         ds_version=ds_version,
+                                         frozen_param_shapes=frozen_param_shapes,
+                                         frozen_param_fragments=frozen_param_fragments)
+        zero_model_states.append(z_model_state)
+
+    return zero_model_states
+
+
+def parse_optim_states(files, ds_checkpoint_dir):
+    total_files = len(files)
+    state_dicts = []
+    for f in tqdm(files, desc='Loading checkpoint shards'):
+        state_dict = torch.load(f, map_location=device, mmap=True, weights_only=False)
+        # immediately discard the potentially huge 2 optimizer states as we only care for fp32 master weights
+        # and also handle the case where it was already removed by another helper script
+        state_dict["optimizer_state_dict"].pop("optimizer_state_dict", None)
+        state_dicts.append(state_dict)
+
+    if not ZERO_STAGE in state_dicts[0][OPTIMIZER_STATE_DICT]:
+        raise ValueError(f"{files[0]} is not a zero checkpoint")
+    zero_stage = state_dicts[0][OPTIMIZER_STATE_DICT][ZERO_STAGE]
+    world_size = state_dicts[0][OPTIMIZER_STATE_DICT][PARTITION_COUNT]
+
+    # For ZeRO-2 each param group can have different partition_count as data parallelism for expert
+    # parameters can be different from data parallelism for non-expert parameters. So we can just
+    # use the max of the partition_count to get the dp world_size.
+
+    if type(world_size) is list:
+        world_size = max(world_size)
+
+    if world_size != total_files:
+        raise ValueError(
+            f"Expected {world_size} of '*_optim_states.pt' under '{ds_checkpoint_dir}' but found {total_files} files. "
+            "Possibly due to an overwrite of an old checkpoint, or a checkpoint didn't get saved by one or more processes."
+        )
+
+    # the groups are named differently in each stage
+    if zero_stage <= 2:
+        fp32_groups_key = SINGLE_PARTITION_OF_FP32_GROUPS
+    elif zero_stage == 3:
+        fp32_groups_key = FP32_FLAT_GROUPS
+    else:
+        raise ValueError(f"unknown zero stage {zero_stage}")
+
+    fp32_flat_groups = [state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key] for i in range(len(state_dicts))]
+    return zero_stage, world_size, fp32_flat_groups
+
+
+def _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir, exclude_frozen_parameters):
+    """
+    Returns fp32 state_dict reconstructed from ds checkpoint
+
+    Args:
+        - ``ds_checkpoint_dir``: path to the deepspeed checkpoint folder (where the optimizer files are)
+
+    """
+    print(f"Processing zero checkpoint '{ds_checkpoint_dir}'")
+
+    optim_files = get_optim_files(ds_checkpoint_dir)
+    zero_stage, world_size, fp32_flat_groups = parse_optim_states(optim_files, ds_checkpoint_dir)
+    print(f"Detected checkpoint of type zero stage {zero_stage}, world_size: {world_size}")
+
+    model_files = get_model_state_files(ds_checkpoint_dir)
+
+    zero_model_states = parse_model_states(model_files)
+    print(f'Parsing checkpoint created by deepspeed=={zero_model_states[0].ds_version}')
+
+    if zero_stage <= 2:
+        return _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                                          exclude_frozen_parameters)
+    elif zero_stage == 3:
+        return _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                                          exclude_frozen_parameters)
+
+
+def _zero2_merge_frozen_params(state_dict, zero_model_states):
+    if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
+        return
+
+    frozen_param_shapes = zero_model_states[0].frozen_param_shapes
+    frozen_param_fragments = zero_model_states[0].frozen_param_fragments
+
+    if debug:
+        num_elem = sum(s.numel() for s in frozen_param_shapes.values())
+        print(f'rank 0: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
+
+        wanted_params = len(frozen_param_shapes)
+        wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
+        avail_numel = sum([p.numel() for p in frozen_param_fragments.values()])
+        print(f'Frozen params: Have {avail_numel} numels to process.')
+        print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
+
+    total_params = 0
+    total_numel = 0
+    for name, shape in frozen_param_shapes.items():
+        total_params += 1
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+
+        state_dict[name] = frozen_param_fragments[name]
+
+        if debug:
+            print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
+
+    print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _has_callable(obj, fn):
+    attr = getattr(obj, fn, None)
+    return callable(attr)
+
+
+def _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
+    param_shapes = zero_model_states[0].param_shapes
+
+    # Reconstruction protocol:
+    #
+    # XXX: document this
+
+    if debug:
+        for i in range(world_size):
+            for j in range(len(fp32_flat_groups[0])):
+                print(f"{FP32_FLAT_GROUPS}[{i}][{j}].shape={fp32_flat_groups[i][j].shape}")
+
+    # XXX: memory usage doubles here (zero2)
+    num_param_groups = len(fp32_flat_groups[0])
+    merged_single_partition_of_fp32_groups = []
+    for i in range(num_param_groups):
+        merged_partitions = [sd[i] for sd in fp32_flat_groups]
+        full_single_fp32_vector = torch.cat(merged_partitions, 0)
+        merged_single_partition_of_fp32_groups.append(full_single_fp32_vector)
+    avail_numel = sum(
+        [full_single_fp32_vector.numel() for full_single_fp32_vector in merged_single_partition_of_fp32_groups])
+
+    if debug:
+        wanted_params = sum([len(shapes) for shapes in param_shapes])
+        wanted_numel = sum([sum(shape.numel() for shape in shapes.values()) for shapes in param_shapes])
+        # not asserting if there is a mismatch due to possible padding
+        print(f"Have {avail_numel} numels to process.")
+        print(f"Need {wanted_numel} numels in {wanted_params} params.")
+
+    # params
+    # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
+    # out-of-core computing solution
+    total_numel = 0
+    total_params = 0
+    for shapes, full_single_fp32_vector in zip(param_shapes, merged_single_partition_of_fp32_groups):
+        offset = 0
+        avail_numel = full_single_fp32_vector.numel()
+        for name, shape in shapes.items():
+
+            unpartitioned_numel = shape.numel() if _has_callable(shape, 'numel') else math.prod(shape)
+            total_numel += unpartitioned_numel
+            total_params += 1
+
+            if debug:
+                print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
+            state_dict[name] = full_single_fp32_vector.narrow(0, offset, unpartitioned_numel).view(shape)
+            offset += unpartitioned_numel
+
+        # Z2 started to align to 2*world_size to improve nccl performance. Therefore both offset and
+        # avail_numel can differ by anywhere between 0..2*world_size. Due to two unrelated complex
+        # paddings performed in the code it's almost impossible to predict the exact numbers w/o the
+        # live optimizer object, so we are checking that the numbers are within the right range
+        align_to = 2 * world_size
+
+        def zero2_align(x):
+            return align_to * math.ceil(x / align_to)
+
+        if debug:
+            print(f"original offset={offset}, avail_numel={avail_numel}")
+
+        offset = zero2_align(offset)
+        avail_numel = zero2_align(avail_numel)
+
+        if debug:
+            print(f"aligned  offset={offset}, avail_numel={avail_numel}")
+
+        # Sanity check
+        if offset != avail_numel:
+            raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
+
+    print(f"Reconstructed fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                               exclude_frozen_parameters):
+    state_dict = OrderedDict()
+
+    # buffers
+    buffers = zero_model_states[0].buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    if not exclude_frozen_parameters:
+        _zero2_merge_frozen_params(state_dict, zero_model_states)
+
+    _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
+
+    # recover shared parameters
+    for pair in zero_model_states[0].shared_params:
+        if pair[1] in state_dict:
+            state_dict[pair[0]] = state_dict[pair[1]]
+
+    return state_dict
+
+
+def zero3_partitioned_param_info(unpartitioned_numel, world_size):
+    remainder = unpartitioned_numel % world_size
+    padding_numel = (world_size - remainder) if remainder else 0
+    partitioned_numel = math.ceil(unpartitioned_numel / world_size)
+    return partitioned_numel, padding_numel
+
+
+def _zero3_merge_frozen_params(state_dict, world_size, zero_model_states):
+    if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
+        return
+
+    if debug:
+        for i in range(world_size):
+            num_elem = sum(s.numel() for s in zero_model_states[i].frozen_param_fragments.values())
+            print(f'rank {i}: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
+
+        frozen_param_shapes = zero_model_states[0].frozen_param_shapes
+        wanted_params = len(frozen_param_shapes)
+        wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
+        avail_numel = sum([p.numel() for p in zero_model_states[0].frozen_param_fragments.values()]) * world_size
+        print(f'Frozen params: Have {avail_numel} numels to process.')
+        print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
+
+    total_params = 0
+    total_numel = 0
+    for name, shape in zero_model_states[0].frozen_param_shapes.items():
+        total_params += 1
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+
+        param_frags = tuple(model_state.frozen_param_fragments[name] for model_state in zero_model_states)
+        state_dict[name] = torch.cat(param_frags, 0).narrow(0, 0, unpartitioned_numel).view(shape)
+
+        partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
+
+        if debug:
+            print(
+                f"Frozen params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
+            )
+
+    print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
+
+
+class GatheredTensor:
+    """
+    A pseudo tensor that collects partitioned weights.
+    It is more memory efficient when there are multiple groups.
+    """
+
+    def __init__(self, flat_groups, flat_groups_offset, offset, partitioned_numel, shape):
+        self.flat_groups = flat_groups
+        self.flat_groups_offset = flat_groups_offset
+        self.offset = offset
+        self.partitioned_numel = partitioned_numel
+        self.shape = shape
+        self.dtype = self.flat_groups[0][0].dtype
+
+    def contiguous(self):
+        """
+        Merge partitioned weights from flat_groups into a single tensor.
+        """
+        end_idx = self.offset + self.partitioned_numel
+        world_size = len(self.flat_groups)
+        pad_flat_param_chunks = []
+
+        for rank_i in range(world_size):
+            # for each rank, we need to collect weights from related group/groups
+            flat_groups_at_rank_i = self.flat_groups[rank_i]
+            start_group_id = None
+            end_group_id = None
+            for group_id in range(len(self.flat_groups_offset)):
+                if self.flat_groups_offset[group_id] <= self.offset < self.flat_groups_offset[group_id + 1]:
+                    start_group_id = group_id
+                if self.flat_groups_offset[group_id] < end_idx <= self.flat_groups_offset[group_id + 1]:
+                    end_group_id = group_id
+                    break
+            # collect weights from related group/groups
+            for group_id in range(start_group_id, end_group_id + 1):
+                flat_tensor = flat_groups_at_rank_i[group_id]
+                start_offset = self.offset - self.flat_groups_offset[group_id]
+                end_offset = min(end_idx, self.flat_groups_offset[group_id + 1]) - self.flat_groups_offset[group_id]
+                pad_flat_param_chunks.append(flat_tensor[start_offset:end_offset])
+
+        # collect weights from all ranks
+        pad_flat_param = torch.cat(pad_flat_param_chunks, dim=0)
+        param = pad_flat_param[:self.shape.numel()].view(self.shape).contiguous()
+        return param
+
+
+def _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
+    param_shapes = zero_model_states[0].param_shapes
+    avail_numel = sum([flat_group.numel() for flat_group in fp32_flat_groups[0]]) * world_size
+
+    # Reconstruction protocol: For zero3 we need to zip the partitions together at boundary of each
+    # param, re-consolidating each param, while dealing with padding if any
+
+    # merge list of dicts, preserving order
+    param_shapes = {k: v for d in param_shapes for k, v in d.items()}
+
+    if debug:
+        for i in range(world_size):
+            print(f"{FP32_FLAT_GROUPS}[{i}].shape={fp32_flat_groups[i].shape}")
+
+        wanted_params = len(param_shapes)
+        wanted_numel = sum(shape.numel() for shape in param_shapes.values())
+        # not asserting if there is a mismatch due to possible padding
+        avail_numel = fp32_flat_groups[0].numel() * world_size
+        print(f"Trainable params: Have {avail_numel} numels to process.")
+        print(f"Trainable params: Need {wanted_numel} numels in {wanted_params} params.")
+
+    # params
+    # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
+    # out-of-core computing solution
+    offset = 0
+    total_numel = 0
+    total_params = 0
+    flat_groups_offset = [0] + list(np.cumsum([flat_tensor.numel() for flat_tensor in fp32_flat_groups[0]]))
+    for name, shape in tqdm(param_shapes.items(), desc='Gathering sharded weights'):
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+        total_params += 1
+        partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
+
+        if debug:
+            print(
+                f"Trainable params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
+            )
+
+        # memory efficient tensor
+        tensor = GatheredTensor(fp32_flat_groups, flat_groups_offset, offset, partitioned_numel, shape)
+        state_dict[name] = tensor
+        offset += partitioned_numel
+
+    offset *= world_size
+
+    # Sanity check
+    if offset != avail_numel:
+        raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
+
+    print(f"Reconstructed Trainable fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                               exclude_frozen_parameters):
+    state_dict = OrderedDict()
+
+    # buffers
+    buffers = zero_model_states[0].buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    if not exclude_frozen_parameters:
+        _zero3_merge_frozen_params(state_dict, world_size, zero_model_states)
+
+    _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
+
+    # recover shared parameters
+    for pair in zero_model_states[0].shared_params:
+        if pair[1] in state_dict:
+            state_dict[pair[0]] = state_dict[pair[1]]
+
+    return state_dict
+
+
+def to_torch_tensor(state_dict, return_empty_tensor=False):
+    """
+    Convert state_dict of GatheredTensor to torch tensor
+    """
+    torch_state_dict = {}
+    converted_tensors = {}
+    for name, tensor in state_dict.items():
+        tensor_id = id(tensor)
+        if tensor_id in converted_tensors:  # shared tensors
+            shared_tensor = torch_state_dict[converted_tensors[tensor_id]]
+            torch_state_dict[name] = shared_tensor
+        else:
+            converted_tensors[tensor_id] = name
+            if return_empty_tensor:
+                torch_state_dict[name] = torch.empty(tensor.shape, dtype=tensor.dtype)
+            else:
+                torch_state_dict[name] = tensor.contiguous()
+    return torch_state_dict
+
+
+def get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir,
+                                             tag=None,
+                                             exclude_frozen_parameters=False,
+                                             lazy_mode=False):
+    """
+    Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated state_dict that can be loaded with
+    ``load_state_dict()`` and used for training without DeepSpeed or shared with others, for example
+    via a model hub.
+
+    Args:
+        - ``checkpoint_dir``: path to the desired checkpoint folder
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in 'latest' file. e.g., ``global_step14``
+        - ``exclude_frozen_parameters``: exclude frozen parameters
+        - ``lazy_mode``: get state_dict in lazy mode. It returns a dict of pesduo tensor instead of torch tensor, which is more memory efficient.
+          Convert the pesduo tensor to torch tensor by ``.contiguous()``
+
+    Returns:
+        - pytorch ``state_dict``
+
+    A typical usage might be ::
+
+        from deepspeed.utils.zero_to_fp32 import get_fp32_state_dict_from_zero_checkpoint
+        # do the training and checkpoint saving
+        state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir) # already on cpu
+        model = model.cpu() # move to cpu
+        model.load_state_dict(state_dict)
+        # submit to model hub or save the model to share with others
+
+    In this example the ``model`` will no longer be usable in the deepspeed context of the same
+    application. i.e. you will need to re-initialize the deepspeed engine, since
+    ``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
+
+    If you want it all done for you, use ``load_state_dict_from_zero_checkpoint`` instead.
+
+    Note: the above usage may not work if your application doesn't have sufficient free CPU memory.
+    You may need to use the offline approach using the ``zero_to_fp32.py`` script that is saved with
+    the checkpoint. Or you can load state_dict in lazy mode ::
+
+        from deepspeed.utils.zero_to_fp32 import get_fp32_state_dict_from_zero_checkpoint
+        state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, lazy_mode=True) # not on cpu
+        for name, lazy_tensor in state_dict.item():
+            tensor = lazy_tensor.contiguous()  # to cpu
+            print(name, tensor)
+            # del tensor to release memory if it no longer in use
+    """
+    if tag is None:
+        latest_path = os.path.join(checkpoint_dir, 'latest')
+        if os.path.isfile(latest_path):
+            with open(latest_path, 'r') as fd:
+                tag = fd.read().strip()
+        else:
+            raise ValueError(f"Unable to find 'latest' file at {latest_path}")
+
+    ds_checkpoint_dir = os.path.join(checkpoint_dir, tag)
+
+    if not os.path.isdir(ds_checkpoint_dir):
+        raise FileNotFoundError(f"Directory '{ds_checkpoint_dir}' doesn't exist")
+
+    state_dict = _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir, exclude_frozen_parameters)
+    if lazy_mode:
+        return state_dict
+    else:
+        return to_torch_tensor(state_dict)
+
+
+def convert_zero_checkpoint_to_fp32_state_dict(checkpoint_dir,
+                                               output_dir,
+                                               max_shard_size="5GB",
+                                               safe_serialization=False,
+                                               tag=None,
+                                               exclude_frozen_parameters=False):
+    """
+    Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict`` file that can be
+    loaded with ``torch.load(file)`` + ``load_state_dict()`` and used for training without DeepSpeed.
+
+    Args:
+        - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
+        - ``output_dir``: directory to the pytorch fp32 state_dict output files
+        - ``max_shard_size``: the maximum size for a checkpoint before being sharded, default value is 5GB
+        - ``safe_serialization``:  whether to save the model using `safetensors` or the traditional PyTorch way (that uses `pickle`).
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in the file named ``latest`` in the checkpoint folder, e.g., ``global_step14``
+        - ``exclude_frozen_parameters``: exclude frozen parameters
+    """
+
+    # Dependency pre-check
+    if safe_serialization:
+        try:
+            from safetensors.torch import save_file
+        except ImportError:
+            print('If you want to use `safe_serialization`, please `pip install safetensors`')
+            raise
+    if max_shard_size is not None:
+        try:
+            from huggingface_hub import split_torch_state_dict_into_shards
+        except ImportError:
+            print('If you want to use `max_shard_size`, please `pip install huggingface_hub`')
+            raise
+
+    # Convert zero checkpoint to state_dict
+    state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir,
+                                                          tag,
+                                                          exclude_frozen_parameters,
+                                                          lazy_mode=True)
+
+    # Shard the model if it is too big.
+    weights_name = "model.safetensors" if safe_serialization else "pytorch_model.bin"
+    if max_shard_size is not None:
+        filename_pattern = weights_name.replace(".bin", "{suffix}.bin").replace(".safetensors", "{suffix}.safetensors")
+        # an memory-efficient approach for sharding
+        empty_state_dict = to_torch_tensor(state_dict, return_empty_tensor=True)
+        state_dict_split = split_torch_state_dict_into_shards(empty_state_dict,
+                                                              filename_pattern=filename_pattern,
+                                                              max_shard_size=max_shard_size)
+    else:
+        from collections import namedtuple
+        StateDictSplit = namedtuple("StateDictSplit", ["is_sharded", "filename_to_tensors"])
+        state_dict_split = StateDictSplit(is_sharded=False,
+                                          filename_to_tensors={weights_name: list(state_dict.keys())})
+
+    # Save the model by shard
+    os.makedirs(output_dir, exist_ok=True)
+    filename_to_tensors = state_dict_split.filename_to_tensors.items()
+    for shard_file, tensors in tqdm(filename_to_tensors, desc="Saving checkpoint shards"):
+        shard_state_dict = {tensor_name: state_dict[tensor_name] for tensor_name in tensors}
+        shard_state_dict = to_torch_tensor(shard_state_dict)
+        output_path = os.path.join(output_dir, shard_file)
+        if safe_serialization:
+            save_file(shard_state_dict, output_path, metadata={"format": "pt"})
+        else:
+            torch.save(shard_state_dict, output_path)
+        # release the memory of current shard
+        for tensor_name in list(shard_state_dict.keys()):
+            del state_dict[tensor_name]
+            del shard_state_dict[tensor_name]
+        del shard_state_dict
+        gc.collect()
+
+    # Save index if sharded
+    if state_dict_split.is_sharded:
+        index = {
+            "metadata": state_dict_split.metadata,
+            "weight_map": state_dict_split.tensor_to_filename,
+        }
+        save_index_file = "model.safetensors.index.json" if safe_serialization else "pytorch_model.bin.index.json"
+        save_index_file = os.path.join(output_dir, save_index_file)
+        with open(save_index_file, "w", encoding="utf-8") as f:
+            content = json.dumps(index, indent=2, sort_keys=True) + "\n"
+            f.write(content)
+
+
+def load_state_dict_from_zero_checkpoint(model, checkpoint_dir, tag=None):
+    """
+    1. Put the provided model to cpu
+    2. Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict``
+    3. Load it into the provided model
+
+    Args:
+        - ``model``: the model object to update
+        - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in the file named ``latest`` in the checkpoint folder, e.g., ``global_step14``
+
+    Returns:
+        - ``model`: modified model
+
+    Make sure you have plenty of CPU memory available before you call this function. If you don't
+    have enough use the ``zero_to_fp32.py`` utility to do the conversion. You will find it
+    conveniently placed for you in the checkpoint folder.
+
+    A typical usage might be ::
+
+        from deepspeed.utils.zero_to_fp32 import load_state_dict_from_zero_checkpoint
+        model = load_state_dict_from_zero_checkpoint(trainer.model, checkpoint_dir)
+        # submit to model hub or save the model to share with others
+
+    Note, that once this was run, the ``model`` will no longer be usable in the deepspeed context
+    of the same application. i.e. you will need to re-initialize the deepspeed engine, since
+    ``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
+
+    """
+    logger.info(f"Extracting fp32 weights")
+    state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag)
+
+    logger.info(f"Overwriting model with fp32 weights")
+    model = model.cpu()
+    model.load_state_dict(state_dict, strict=False)
+
+    return model
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("checkpoint_dir",
+                        type=str,
+                        help="path to the desired checkpoint folder, e.g., path/checkpoint-12")
+    parser.add_argument("output_dir",
+                        type=str,
+                        help="directory to the pytorch fp32 state_dict output files"
+                        "(e.g. path/checkpoint-12-output/)")
+    parser.add_argument(
+        "--max_shard_size",
+        type=str,
+        default="5GB",
+        help="The maximum size for a checkpoint before being sharded. Checkpoints shard will then be each of size"
+        "lower than this size. If expressed as a string, needs to be digits followed by a unit (like `5MB`"
+        "We default it to 5GB in order for models to be able to run easily on free-tier google colab instances"
+        "without CPU OOM issues.")
+    parser.add_argument(
+        "--safe_serialization",
+        default=False,
+        action='store_true',
+        help="Whether to save the model using `safetensors` or the traditional PyTorch way (that uses `pickle`).")
+    parser.add_argument("-t",
+                        "--tag",
+                        type=str,
+                        default=None,
+                        help="checkpoint tag used as a unique identifier for checkpoint. e.g., global_step1")
+    parser.add_argument("--exclude_frozen_parameters", action='store_true', help="exclude frozen parameters")
+    parser.add_argument("-d", "--debug", action='store_true', help="enable debug")
+    args = parser.parse_args()
+
+    debug = args.debug
+
+    convert_zero_checkpoint_to_fp32_state_dict(args.checkpoint_dir,
+                                               args.output_dir,
+                                               max_shard_size=args.max_shard_size,
+                                               safe_serialization=args.safe_serialization,
+                                               tag=args.tag,
+                                               exclude_frozen_parameters=args.exclude_frozen_parameters)