Update app.py

app.py

@@ -1,61 +1,272 @@
 import gradio as gr
-from transformers import AutoModelForCausalLM, AutoTokenizer
 import torch
+import os
+from huggingface_hub import hf_hub_download
+from transformers import AutoTokenizer
+import torch.nn as nn
+import torch.nn.functional as F
 
-# --- 直接加载模型和分词器 ---
+# ================ 第一步：重新定义模型结构 (必须与训练时完全一致) ================
+# 注意：这些类定义必须与你原始训练脚本中的完全相同
+
+class GELU(nn.Module):
+    def __init__(self):
+        super().__init__()
+    def forward(self, x):
+        return 0.5 * x * (1 + torch.tanh(
+            torch.sqrt(torch.tensor(2.0 / torch.pi)) *
+            (x + 0.044715 * torch.pow(x, 3))
+        ))
+
+class FeedForward(nn.Module):
+    def __init__(self, cfg):
+        super().__init__()
+        self.layers = nn.Sequential(
+            nn.Linear(cfg["emb_dim"], 4 * cfg["emb_dim"]),
+            GELU(),
+            nn.Linear(4 * cfg["emb_dim"], cfg["emb_dim"]),
+        )
+    def forward(self, x):
+        return self.layers(x)
+
+class MultiHeadAttention(nn.Module):
+    def __init__(self, d_in, d_out, 
+                context_length, dropout, num_heads, qkv_bias=False):
+        super().__init__()
+        assert (d_out % num_heads == 0), \
+            "d_out must be divisible by num_heads"
+        
+        self.d_out = d_out
+        self.num_heads = num_heads
+        self.head_dim = d_out // num_heads 
+        self.W_query = nn.Linear(d_in, d_out, bias=qkv_bias)
+        self.W_key = nn.Linear(d_in, d_out, bias=qkv_bias)
+        self.W_value = nn.Linear(d_in, d_out, bias=qkv_bias)
+        self.out_proj = nn.Linear(d_out, d_out) 
+        self.dropout_p = dropout
+
+    def forward(self, x):
+        b, num_tokens, d_in = x.shape
+        keys = self.W_key(x) 
+        queries = self.W_query(x) 
+        values = self.W_value(x)
+        
+        # Transpose into [B, num_heads, num_tokens, head_dim] for SDPA
+        keys = keys.view(b, num_tokens, self.num_heads, self.head_dim).transpose(1, 2)
+        values = values.view(b, num_tokens, self.num_heads, self.head_dim).transpose(1, 2)
+        queries = queries.view(b, num_tokens, self.num_heads, self.head_dim).transpose(1, 2)
+
+        # Use F.scaled_dot_product_attention
+        context_vec = F.scaled_dot_product_attention(
+            queries, keys, values,
+            attn_mask=None,
+            dropout_p=self.dropout_p if self.training else 0.0,
+            is_causal=True
+        )
+
+        # Transpose back to [B, num_tokens, num_heads * head_dim] = [B, T, d_out]
+        context_vec = context_vec.transpose(1, 2).contiguous().view(b, num_tokens, self.d_out)
+        # Apply output projection
+        context_vec = self.out_proj(context_vec)
+
+        return context_vec
+
+class LayerNorm(nn.Module):
+    def __init__(self, emb_dim):
+        super().__init__()
+        self.eps = 1e-5
+        self.scale = nn.Parameter(torch.ones(emb_dim)) 
+        self.shift = nn.Parameter(torch.zeros(emb_dim))
+
+    def forward(self, x):
+        mean = x.mean(dim=-1, keepdim=True)
+        var = x.var(dim=-1, keepdim=True, unbiased=False)
+        norm_x = (x - mean) / torch.sqrt(var + self.eps)
+        return self.scale * norm_x + self.shift
+
+class TransformerBlock(nn.Module):
+    def __init__(self, cfg):
+        super().__init__()
+        self.att = MultiHeadAttention(
+            d_in=cfg["emb_dim"],
+            d_out=cfg["emb_dim"],
+            context_length=cfg["context_length"],
+            num_heads=cfg["n_heads"],
+            dropout=cfg["drop_rate"],
+            qkv_bias=cfg["qkv_bias"])
+        self.ff = FeedForward(cfg)
+        self.norm1 = LayerNorm(cfg["emb_dim"])
+        self.norm2 = LayerNorm(cfg["emb_dim"])
+        self.drop_shortcut = nn.Dropout(cfg["drop_rate"])
+
+    def forward(self, x):
+        shortcut = x
+        x = self.norm1(x)
+        x = self.att(x)
+        x = self.drop_shortcut(x)
+        x = x + shortcut
+        shortcut = x
+        x = self.norm2(x)
+        x = self.ff(x)
+        x = self.drop_shortcut(x)
+        x = x + shortcut
+        return x
+
+class GPTModel(nn.Module):
+    def __init__(self, cfg):
+        super().__init__()
+        self.tok_emb = nn.Embedding(cfg["vocab_size"], cfg["emb_dim"]) 
+        self.pos_emb = nn.Embedding(cfg["context_length"], cfg["emb_dim"]) 
+        self.drop_emb = nn.Dropout(cfg["drop_rate"])
+        
+        self.trf_blocks = nn.Sequential(
+            *[TransformerBlock(cfg) for _ in range(cfg["n_layers"])])
+        
+        self.final_norm = LayerNorm(cfg["emb_dim"])
+        self.out_head = nn.Linear(
+            cfg["emb_dim"], cfg["vocab_size"], bias=False
+        )
+    def forward(self, in_idx):
+        batch_size, seq_len = in_idx.shape
+        tok_embeds = self.tok_emb(in_idx)
+        pos_embeds = self.pos_emb( 
+            torch.arange(seq_len, device=in_idx.device)
+        )
+        x = tok_embeds + pos_embeds
+        x = self.drop_emb(x)
+        x = self.trf_blocks(x)
+        x = self.final_norm(x)
+        logits = self.out_head(x)
+        return logits
+
+# 用于生成的函数
+def generate_text_simple(model, idx, max_new_tokens, context_size):
+    device = idx.device
+    current_device_type = str(device).split(':')[0]
+    
+    for _ in range(max_new_tokens):
+        idx_cond = idx[:, -context_size:]
+        with torch.no_grad():
+            # 推理时不需要混合精度
+            logits = model(idx_cond)
+        logits = logits[:, -1, :]
+        probas = torch.softmax(logits, dim=-1)
+        idx_next = torch.argmax(probas, dim=-1, keepdim=True)
+        idx = torch.cat((idx, idx_next), dim=1)
+    return idx
+
+def text_to_token_ids(text, tokenizer):
+    encoded = tokenizer.encode(text)
+    encoded_tensor = torch.tensor(encoded).unsqueeze(0)
+    return encoded_tensor
+
+def token_ids_to_text(token_ids, tokenizer):
+    flat = token_ids.squeeze(0)
+    return tokenizer.decode(flat.tolist(), skip_special_tokens=True)
+
+# ================ 第二步：设置模型加载和推理 ================
+
+# 模型 ID
 model_id = "xingyu1996/tiger-gpt2"
-tokenizer = AutoTokenizer.from_pretrained("gpt2") # 使用原始的 GPT-2 分词器
-model = AutoModelForCausalLM.from_pretrained(model_id)
-
-def respond(
-    message,
-    history: list[tuple[str, str]],
-    max_tokens,
-    temperature,
-    top_p,
-):
-    # 将输入文本转换为 token ID
-    input_ids = tokenizer.encode(message, return_tensors="pt")
-    
-    # 准备生成参数
-    gen_kwargs = {
-        "max_length": input_ids.shape[1] + max_tokens,
-        "do_sample": True if temperature > 0 else False,
+
+# 从 Hugging Face Hub 下载模型权重文件
+def load_model_from_hub():
+    print("开始从 Hugging Face Hub 下载模型权重...")
+    
+    # 下载 pytorch_model.bin 文件
+    model_file = hf_hub_download(model_id, "pytorch_model.bin")
+    print(f"模型权重文件下载完成：{model_file}")
+    
+    # 下载 config.json 文件
+    config_file = hf_hub_download(model_id, "config.json")
+    print(f"配置文件下载完成：{config_file}")
+    
+    # 加载权重
+    state_dict = torch.load(model_file, map_location="cpu")
+    
+    # 加载配置
+    import json
+    with open(config_file, 'r') as f:
+        config = json.load(f)
+    
+    # 将 Hugging Face 格式的配置转换为我们的格式
+    # 注意：这里的映射需要根据实际情况调整
+    my_config = {
+        "vocab_size": config.get("vocab_size", 50257),
+        "context_length": config.get("n_positions", 512),
+        "emb_dim": config.get("n_embd", 768),
+        "n_heads": config.get("n_head", 12),
+        "n_layers": config.get("n_layer", 12),
+        "drop_rate": config.get("resid_pdrop", 0.1),
+        "qkv_bias": config.get("qkv_bias", False),
     }
     
-    if temperature > 0:
-        gen_kwargs["temperature"] = temperature
-    if top_p < 1.0:
-        gen_kwargs["top_p"] = top_p
-        
-    # 生成文本
-    with torch.no_grad():
-        output_ids = model.generate(input_ids, **gen_kwargs)
-        
-    # 只保留新生成的部分
-    new_tokens = output_ids[0, input_ids.shape[1]:]
+    # 创建模型
+    model = GPTModel(my_config)
+    
+    # 加载权重到模型
+    # 检查状态字典中是否有 _orig_mod. 前缀
+    if any(k.startswith('_orig_mod.') for k in state_dict.keys()):
+        state_dict = {k.replace('_orig_mod.', ''): v for k, v in state_dict.items()}
+        print("已去除权重中的 _orig_mod. 前缀")
     
-    # 解码生成的 token ID
-    response = tokenizer.decode(new_tokens, skip_special_tokens=True)
-    return response
+    # 加载权重
+    try:
+        model.load_state_dict(state_dict)
+        print("模型权重加载成功！")
+    except Exception as e:
+        print(f"模型权重加载失败: {e}")
+        # 尝试加载部分权重
+        model.load_state_dict(state_dict, strict=False)
+        print("模型已使用非严格模式加载权重，可能有部分参数没有加载。")
+    
+    model.eval()  # 设置为评估模式
+    return model, my_config
+
+# 加载模型和分词器
+print("正在初始化...")
+model, config = load_model_from_hub()
+tokenizer = AutoTokenizer.from_pretrained("gpt2")
+print("模型和分词器加载完成！")
 
+# ================ 第三步：设置 Gradio 接口 ================
+
+def respond(message, history, max_tokens, temperature):
+    input_ids = text_to_token_ids(message, tokenizer).to("cpu")  # Hugging Face Space 可能没有 GPU
+    context_size = config["context_length"]
+    
+    try:
+        # 生成文本
+        output_ids = generate_text_simple(
+            model=model,
+            idx=input_ids,
+            max_new_tokens=max_tokens,
+            context_size=context_size
+        )
+        
+        # 解码生成的文本
+        full_text = token_ids_to_text(output_ids, tokenizer)
+        
+        # 分离提示和生成部分
+        if message in full_text:
+            generated = full_text[len(message):]
+        else:
+            generated = full_text
+            
+        return generated
+    except Exception as e:
+        print(f"生成过程中出错: {type(e).__name__} - {e}")
+        return f"抱歉，生成文本时出错: {type(e).__name__}"
 
-# 其他 Gradio 界面代码不变
+# 创建 Gradio 界面
 demo = gr.ChatInterface(
     respond,
     additional_inputs=[
-        gr.Slider(minimum=1, maximum=512, value=325, step=1, label="Max new tokens"),
-        gr.Slider(minimum=0.1, maximum=2.0, value=0.7, step=0.1, label="Temperature"),
-        gr.Slider(
-            minimum=0.1,
-            maximum=1.0,
-            value=0.95,
-            step=0.05,
-            label="Top-p (nucleus sampling)",
-        ),
+        gr.Slider(minimum=1, maximum=100, value=30, step=1, label="生成长度"),
+        gr.Slider(minimum=0.1, maximum=1.0, value=0.7, step=0.1, label="温度"),
     ],
-    title=f"推理测试: {model_id}",
-    description="输入中文文本，模型将进行补全。"
+    title=f"Tiger-GPT2 推理测试",
+    description="输入中文文本，模型将生成后续内容。此演示直接加载了原始模型权重，与本地推理行为一致。",
 )
 
 if __name__ == "__main__":