Upload 4 files

app.py

@@ -1,37 +1,135 @@
-import subprocess
 import gradio as gr
+import subprocess
+import threading
+import time
 
-def train():
-    # 安装依赖
-    process = subprocess.Popen(
-        ['pip', 'install', '-r', 'requirements.txt'],
-        stdout=subprocess.PIPE,
-        stderr=subprocess.STDOUT,
-        text=True
-    )
-    for line in iter(process.stdout.readline, ''):
-        yield line
-    process.wait()
-    
-    yield "---依赖安装完成，开始训练---"
+def get_md_content(file_path):
+    try:
+        with open(file_path, 'r', encoding='utf-8') as f:
+            return f.read()
+    except FileNotFoundError:
+        return f"Error: {file_path} not found."
+    except Exception as e:
+        return f"An error occurred: {e}"
 
-    # 运行训练脚本
+def run_script():
+    """Function to run the fine-tuning script and stream output."""
     process = subprocess.Popen(
         ['python3', 'fine_tune.py'],
         stdout=subprocess.PIPE,
         stderr=subprocess.STDOUT,
-        text=True
+        text=True,
+        bufsize=1,
+        universal_newlines=True
     )
-    for line in iter(process.stdout.readline, ''):
-        yield line
+    output = ""
+    for line in process.stdout:
+        output += line
+        yield output
     process.wait()
+
+# JavaScript to find and render Mermaid diagrams
+js_script = """
+() => {
+    function initMermaidAndConvert() {
+        // Wait for mermaid to be available
+        if (typeof mermaid === 'undefined') {
+            console.log('Mermaid not loaded yet, retrying...');
+            setTimeout(initMermaidAndConvert, 100);
+            return;
+        }
+        
+        console.log('Mermaid loaded successfully');
+        // Initialize mermaid
+        mermaid.initialize({ 
+            startOnLoad: false, 
+            theme: 'default',
+            securityLevel: 'loose'
+        });
+        
+        function convertMermaidCodeBlocks() {
+            console.log('Converting Mermaid code blocks...');
+            let processedCount = 0;
+            
+            // Look for pre blocks that contain mermaid syntax
+            document.querySelectorAll('pre').forEach((pre, index) => {
+                const code = pre.querySelector('code');
+                if (code && !pre.classList.contains('mermaid-processed')) {
+                    const text = code.textContent.trim();
+                    // Check if it contains mermaid syntax
+                    const isMermaid = text.includes('graph ') || 
+                                    text.includes('flowchart ') || 
+                                    text.includes('subgraph ') ||
+                                    text.startsWith('graph') ||
+                                    text.startsWith('flowchart') ||
+                                    text.includes('classDiagram') ||
+                                    text.includes('sequenceDiagram');
+                    
+                    if (isMermaid) {
+                        console.log(`Found Mermaid diagram ${processedCount + 1}:`, text.substring(0, 50) + '...');
+                        pre.classList.add('mermaid');
+                        pre.classList.add('mermaid-processed');
+                        pre.textContent = text;
+                        processedCount++;
+                    }
+                }
+            });
+            
+            console.log(`Processed ${processedCount} Mermaid diagrams`);
+            // Run Mermaid
+            try {
+                mermaid.run();
+            } catch (e) {
+                console.log('Mermaid rendering error:', e);
+            }
+        }
+
+        // Use a MutationObserver to re-run the conversion when Gradio updates the page
+        const observer = new MutationObserver((mutations) => {
+            // A simple debounce to avoid excessive re-renders
+            clearTimeout(window.mermaidTimeout);
+            window.mermaidTimeout = setTimeout(convertMermaidCodeBlocks, 100);
+        });
+        observer.observe(document.body, { childList: true, subtree: true });
+
+        // Initial run
+        convertMermaidCodeBlocks();
+    }
     
-    yield "---训练完成！---"
+    // Start the initialization
+    initMermaidAndConvert();
+}
+"""
+
+# HTML to include the Mermaid.js library
+head_script = '<script src="https://cdn.jsdelivr.net/npm/mermaid/dist/mermaid.min.js"></script>'
+
+with gr.Blocks(theme=gr.themes.Soft(), head=head_script, js=js_script) as demo:
+    gr.Markdown("# 微调技术分享")
+
+    with gr.Tabs():
+        with gr.TabItem("分享大纲"):
+            gr.Markdown(get_md_content("outline.md"))
+        
+        with gr.TabItem("核心技术概览"):
+            gr.Markdown(get_md_content("presentation.md"))
+
+        with gr.TabItem("LoRA & QLoRA 深度解析"):
+            gr.Markdown(get_md_content("lora_qlora_deep_dive.md"))
+
+
 
-with gr.Blocks() as demo:
-    gr.Markdown("点击按钮开始微调")
-    output = gr.Textbox(label="训练日志", lines=20)
-    train_button = gr.Button("开始微调")
-    train_button.click(fn=train, inputs=[], outputs=output)
+        with gr.TabItem("动手实战：模型微调"):
+            with gr.Row():
+                start_button = gr.Button("开始微调", variant="primary")
+            
+            log_output = gr.Textbox(
+                label="训练日志",
+                interactive=False,
+                lines=20,
+                show_copy_button=True
+            )
+            start_button.click(fn=run_script, outputs=log_output)
 
-demo.launch()
+if __name__ == "__main__":
+    demo.launch()

fine_tune.py

@@ -4,20 +4,15 @@ from peft import LoraConfig, get_peft_model, prepare_model_for_kbit_training
 from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig, TrainingArguments
 from trl import SFTTrainer
 
-# 1. 加载模型和分词器
-model_name = "NousResearch/Llama-2-7b-chat-hf"
-
-# BitsAndBytesConfig for QLoRA
-bnb_config = BitsAndBytesConfig(
-    load_in_4bit=True,
-    bnb_4bit_quant_type="nf4",
-    bnb_4bit_compute_dtype=torch.float16,
-)
+# 1. 加载模型和分词器 (CPU优化版本)
+# 使用更小的模型以适配CPU环境
+model_name = "microsoft/DialoGPT-small"  # 更小的模型，适合CPU训练
 
+# CPU环境下不需要量化配置
 model = AutoModelForCausalLM.from_pretrained(
     model_name,
-    # quantization_config=bnb_config, # Deactivated for CPU
-    # device_map="auto" # Deactivated for CPU
+    torch_dtype=torch.float32,  # CPU使用float32
+    low_cpu_mem_usage=True,     # 优化CPU内存使用
 )
 model.config.use_cache = False
 
@@ -34,31 +29,34 @@ def formatting_prompts_func(example):
 
 dataset = load_dataset("json", data_files="data.json", split="train")
 
-# 3. 配置LoRA参数
+# 3. 配置LoRA参数 (适配DialoGPT)
 lora_config = LoraConfig(
     r=8, # Rank
     lora_alpha=32,
     lora_dropout=0.1,
     bias="none",
     task_type="CAUSAL_LM",
-    target_modules=["q_proj", "k_proj", "v_proj", "o_proj"], # Llama-2 specific modules
+    target_modules=["c_attn", "c_proj"],  # DialoGPT/GPT-2 架构的注意力模块
 )
 
-# 4. 创建PEFT模型
-model = prepare_model_for_kbit_training(model)
+# 4. 创建PEFT模型 (CPU版本)
+# CPU环境下不需要量化准备
 model = get_peft_model(model, lora_config)
 
-# 5. 配置训练参数
-output_dir = "./llama-2-7b-chat-json"
+# 5. 配置训练参数 (CPU优化)
+output_dir = "./dialogpt-small-lora"
 training_args = TrainingArguments(
     output_dir=output_dir,
-    per_device_train_batch_size=4,
-    gradient_accumulation_steps=4,
-    learning_rate=2e-4,
-    logging_steps=10,
-    max_steps=100, # for demo
-    save_strategy="epoch",
-    # num_train_epochs=1, # use max_steps for demo
+    per_device_train_batch_size=1,  # CPU环境使用更小的批次
+    gradient_accumulation_steps=8,   # 增加梯度累积以补偿小批次
+    learning_rate=5e-4,             # 稍微提高学习率
+    logging_steps=5,
+    max_steps=50,                   # 减少训练步数用于演示
+    save_strategy="steps",
+    save_steps=25,
+    dataloader_num_workers=0,       # CPU环境下设为0
+    fp16=False,                     # CPU不支持fp16
+    report_to=None,                 # 禁用wandb等报告
 )
 
 # 6. 创建Trainer并开始训练
@@ -74,6 +72,6 @@ trainer = SFTTrainer(
 trainer.train()
 
 # 7. 保存模型
-print("Saving LoRA adapter...")
+print("Saving DialoGPT LoRA adapter...")
 trainer.save_model(output_dir)
-print(f"LoRA adapter saved to {output_dir}")
+print(f"DialoGPT LoRA adapter saved to {output_dir}")