import os import json import torch from tqdm import tqdm from transformers import AutoTokenizer, AutoModel from torch import nn from safetensors.torch import load_file # ========================= # 配置 #加速方法: #真正的加速来自: #✅ 批量并行(一次性喂多个样本) #✅ AMP 半精度加速 #✅ 显存优化(禁用 cache) ''' 1.批量并行 (Batching) 传统做法:一条数据跑一次前向,GPU 每次只处理一条输入。 现在的做法:把 cur_bs 条数据拼成一个 batch,一次送进 GPU。 GPU 的矩阵乘法 (GEMM) 对大 batch 并行效率高,吞吐量提升明显。 换句话说,“加速”来自 利用 GPU 的并行算力,不是 CPU 开线程。 2.CUDA 内核并行 PyTorch 在 GPU 上的算子(比如 Attention、Linear)本身就是用 CUDA 内核实现的。 CUDA 内核里有成千上万的线程在跑矩阵乘法和 softmax,这部分完全由 GPU 管理。 你虽然没写 threading 或 multiprocessing,但 GPU 在底层自动开了成百上千个线程。 3.自动混合精度 (AMP) with torch.cuda.amp.autocast(dtype=torch.float16): 这让很多矩阵运算走 FP16 / TensorCore,比 FP32 快 2~8 倍,还能节省显存。 所以 速度 + 批量大小 叠加,整体吞吐量大幅提高。 4.KV Cache 关闭 训练 Reward Model 不需要自回归生成,只做一次全序列前向。 关掉 KV Cache → 避免显存里存一堆 key/value,减少显存占用。 这样就能跑更大的 batch,间接提升速度。 ''' # ========================= #MODEL_PATH="/root/test/weitiao/output/qwen3_4B_rm" #v1 #MODEL_PATH="/root/test/weitiao/data_processing_hsichen/data_process_bq/model/qwen3_4B_rm_merged_v2" #v2 MODEL_PATH="/root/test/weitiao/data_processing_hsichen/data_process_bq/model/qwen3_4b_rm_v2.2_merged" #v2.2复现v2看稳定性 INPUT_FILE="/root/test/weitiao/data_processing_hsichen/data_process_bq/data/rm_dpo_scored_quota_filtered.json" OUTPUT_FILE = "/root/test/weitiao/data_processing_hsichen/data_process_bq/data/rm_dpo_scored_infered_v1.json" # 输出结果文件 INIT_BATCH_SIZE = 32 # 初始 batch,后面会根据显存自动回退 MAX_SEQ_LEN = 1024 # 先把长度砍到 1024(从 2048 降一半,显存压力四分之一) USE_FP16 = True # 建议保持 True USE_COMPILE = False # 先关闭 compile,Qwen3 + inductor 在长序列上显存放大明显 # 可选:让 float32 matmul 走 TF32(速度更快;仅 Ampere+ 生效) torch.set_float32_matmul_precision("high") print(f"🚀 Loading reward model from: {MODEL_PATH}") device = "cuda" if torch.cuda.is_available() else "cpu" tokenizer = AutoTokenizer.from_pretrained(MODEL_PATH, trust_remote_code=True) torch_dtype = torch.float16 if (USE_FP16 and torch.cuda.is_available()) else torch.float32 model = AutoModel.from_pretrained( MODEL_PATH, trust_remote_code=True, torch_dtype=torch_dtype ) # === 加载奖励头 === value_head_path = f"{MODEL_PATH}/value_head.safetensors" value_head_state = load_file(value_head_path) print("✅ Loaded value head keys:", value_head_state.keys()) hidden_size = model.config.hidden_size value_head = nn.Linear(hidden_size, 1, bias=True) value_head.load_state_dict({ "weight": value_head_state["v_head.summary.weight"].to(dtype=torch_dtype), "bias": value_head_state["v_head.summary.bias"].to(dtype=torch_dtype), }) value_head.to(device=device, dtype=torch_dtype) model.v_head = value_head model.eval() model.to(device) # 关键:禁用 KV cache;不做自回归生成,只做全序列前向,开 cache 反而占显存 if hasattr(model.config, "use_cache"): model.config.use_cache = False # 先不要 compile,稳定为主 if USE_COMPILE and hasattr(torch, "compile"): # 对于 Qwen3 + 长序列,compile 可能放大峰值显存;确认显存足够再打开 model = torch.compile(model, mode="max-autotune") # ========================= # 构造 prompt(保持简单,长度靠 tokenizer 截断) # ========================= def build_prompt(conversations, response): sys = [c["value"] for c in conversations if c["from"] == "system"] sys_text = "\n".join([f"System: {s}" for s in sys]) + "\n" if sys else "" hist = [ f"{c['from'].capitalize()}: {c['value']}\n" for c in conversations if c["from"] != "system" ] return sys_text + "".join(hist) + f"Assistant: {response}" # ========================= # 批量打分(强制 memory-efficient attention + 自动半精度) # ========================= def get_reward_score(prompts): """ 输入: list[str] 输出: list[float] """ # 注意:decoder-only 模型下 padding='longest' 足够;让 tokenizer 自己截断到 MAX_SEQ_LEN inputs = tokenizer( prompts, return_tensors="pt", padding=True, truncation=True, max_length=MAX_SEQ_LEN ).to(device) # 强制 SDPA 使用 memory-efficient kernel,避免 flash 不可用时退到 math 内核吃更多显存 with torch.backends.cuda.sdp_kernel(enable_flash=True, enable_math=False, enable_mem_efficient=True): with torch.no_grad(): # autocast 能保证偶发的 float32 op 也尽量走半精度/TF32 autocast_dtype = torch.float16 if torch_dtype == torch.float16 else torch.bfloat16 with torch.cuda.amp.autocast(dtype=autocast_dtype): outputs = model(**inputs, use_cache=False) # 再次显式禁用 cache last_hidden = outputs.last_hidden_state # [B, L, H] last_token_hidden = last_hidden[:, -1, :] # [B, H] scores = model.v_head(last_token_hidden).squeeze(-1) # [B] return scores.detach().cpu().tolist() # ========================= # 加载数据 # ========================= print(f"📄 Loading dataset: {INPUT_FILE}") with open(INPUT_FILE, "r", encoding="utf-8") as f: data = json.load(f) # ========================= # 自适应 batch 的主循环(OOM 自动回退) # ========================= def score_all(data): scored = [] i = 0 cur_bs = INIT_BATCH_SIZE pbar = tqdm(total=len(data), desc="Scoring", dynamic_ncols=True) try: while i < len(data): batch = data[i : i + cur_bs] # 1) 打平 prompts(chosen + rejected 一起跑) prompts = [] for sample in batch: pc = build_prompt(sample["conversations"], sample["chosen"]["value"]) pr = build_prompt(sample["conversations"], sample["rejected"]["value"]) prompts.extend([pc, pr]) # 2) 推理(若 OOM,缩小 batch 重试) try: scores = get_reward_score(prompts) # [2 * cur_bs] except torch.cuda.OutOfMemoryError: torch.cuda.empty_cache() # 动态降批次 if cur_bs == 1: # 仍然 OOM:再把序列长降一档再试 raise RuntimeError( f"OOM at batch_size=1, MAX_SEQ_LEN={MAX_SEQ_LEN}. " f"请进一步降低 MAX_SEQ_LEN 或确保输入更短。" ) cur_bs = max(1, cur_bs // 2) print(f"\n⚠️ OOM detected. Reducing batch_size to {cur_bs} and retrying...") continue # 3) 回填 for j, sample in enumerate(batch): rc = float(scores[2*j]) rr = float(scores[2*j + 1]) sample["reward_chosen"] = rc sample["reward_rejected"] = rr sample["delta_score"] = rc - rr scored.append(sample) i += cur_bs pbar.update(cur_bs) finally: pbar.close() return scored print("⚡ Scoring samples...") scored_data = score_all(data) # ========================= # 保存结果 # ========================= with open(OUTPUT_FILE, "w", encoding="utf-8") as f: json.dump(scored_data, f, ensure_ascii=False, indent=2) print(f"✅ Scoring finished! Results saved to: {OUTPUT_FILE}")