import pynvml import pandas as pd from typing import Any, Dict, List, Optional import torch from tqdm import tqdm import ast import os import tempfile import numpy as np from src.metrics import rouge_L from transformers import AutoModelForCausalLM, AutoTokenizer MODEL2HF={ "llama2-7b":"meta-llama/Llama-2-7b-hf", "mistral-7b-0.3":"mistralai/Mistral-7B-v0.3", "llama2-13b":"meta-llama/Llama-2-13b-hf", "llama3.1-8b":"meta-llama/Llama-3.1-8B", "llama3.2_1b_instruct":"/bigtemp/usr/Model/Llama-3.2-1B-Instruct", "llama3.2_3b_instruct":"/bigtemp/usr/Model/Llama-3.2-3B-Instruct", "mistral_instruct":"mistralai/Mistral-7B-Instruct-v0.3", "llama_instruct":"meta-llama/Llama-3.1-8B-Instruct", "llama3.2-1b":"/bigtemp/usr/Model/Llama-3.2-1B", "llama3.2-3b":"/bigtemp/usr/Model/Llama-3.2-3B", "qwen2.5_14b_instruct":"/bigtemp/usr/Model/Qwen2.5-14B-Instruct", "qwen2.5_14b":"/bigtemp/usr/Model/Qwen2.5-14B" } MODEL2LAYER = { "llama2-7b": 32, "mistral-7b": 32, "llama2-13b": 40, "llama3.1-8b":32, "llama3.2-3b":28, "llama3.2-1b":16, "llama_instruct":32, "llama3.2_3b_instruct":28, "llama3.2_1b_instruct":16, "mistral_instruct":32, "qwen2.5_14b_instruct":48 } DATA2HF={ "coqa":"stanfordnlp/coqa", "triviaqa":None, "hotpotqa":None, "squad":"rajpurkar/squad", "hotpotqa_c":None, 'psiloqa':None, "halueval_summary":None, "cnn_dailymail":"abisee/cnn_dailymail", "odd_man_out":None, "trec":"trec", "dbpedia_14":"dbpedia_14", "ag_news":"ag_news" } def get_least_used_gpu() -> int: """ Returns the local GPU index (0-based) with the least used memory among visible GPUs. Raises RuntimeError if no GPU is available. """ visible_devices = os.environ.get("CUDA_VISIBLE_DEVICES") visible_physical_ids: Optional[List[int]] = None if visible_devices is not None and visible_devices.strip(): visible_physical_ids = [] for token in visible_devices.split(","): token = token.strip() if not token: continue if not token.lstrip("-").isdigit(): visible_physical_ids = None break physical_id = int(token) if physical_id < 0: visible_physical_ids = None break visible_physical_ids.append(physical_id) if visible_physical_ids == []: visible_physical_ids = None pynvml.nvmlInit() try: device_count = pynvml.nvmlDeviceGetCount() if device_count == 0: raise RuntimeError("No GPU devices found.") if visible_physical_ids is None: candidate_ids = list(range(device_count)) local_to_physical = {idx: idx for idx in candidate_ids} else: candidate_ids = [idx for idx in visible_physical_ids if idx < device_count] if not candidate_ids: raise RuntimeError( "CUDA_VISIBLE_DEVICES does not reference any valid GPU." ) local_to_physical = { local_idx: physical_idx for local_idx, physical_idx in enumerate(candidate_ids) } best_local_idx = 0 best_used = None for local_idx, physical_idx in local_to_physical.items(): handle = pynvml.nvmlDeviceGetHandleByIndex(physical_idx) mem_info = pynvml.nvmlDeviceGetMemoryInfo(handle) # bytes used = mem_info.used if best_used is None or used < best_used: best_used = used best_local_idx = local_idx return best_local_idx finally: pynvml.nvmlShutdown() def expand_df_to_data(df: pd.DataFrame, save_csv_path: Optional[str] = None) -> pd.DataFrame: """Expand a prepared QA DataFrame into model-ready records. """ eval_expanded: Dict[str, List[Any]] = {"context_pred": [], "label": [], "context_gt": [], "context": []} for _, row in tqdm(df.iterrows(), total=len(df), desc="Expanding DataFrame"): context = row['context'] prediction=row['sampled_answers'] gold_answer: List[str] = row['gold_answer'] # Score the first pair for label using local rouge_L helper score,matched_gold = rouge_L([prediction], gold_answer) label = score > 0.5 context_pred = construct_input(context, prediction) context_gt = construct_input(context, matched_gold) eval_expanded["context_pred"].append(context_pred) eval_expanded["context"].append(context) eval_expanded["label"].append(label) eval_expanded["context_gt"].append(context_gt) eval_df = pd.DataFrame(eval_expanded) # Optionally persist labels (and minimal context) to CSV here if save_csv_path is not None: try: pd.DataFrame({ "context": eval_df["context"].tolist(), "context_pred": eval_df["context_pred"].tolist(), "context_gt": eval_df["context_gt"].tolist(), "label": eval_df["label"].astype(int).tolist(), }).to_csv(save_csv_path, index=False) except Exception as e: print(f"Warning: failed to save labels CSV at {save_csv_path}: {e}") return eval_df def process_row(row,args): context = row['context'] prediction = row['sampled_answers'] if args.model=='mistral-7b': prediction=ast.literal_eval(row['sampled_answers']) prediction=prediction[0] gold_answer = row['gold_answer'] with tempfile.TemporaryDirectory() as tmpdirname: os.environ["HF_DATASETS_CACHE"] = tmpdirname score, matched_gold = rouge_L([prediction], gold_answer) label = score > 0.4 context_pred = construct_input(context, prediction) context_gt = construct_input(context, matched_gold) return pd.Series({ "context_pred": context_pred, "context": context, "label": label, "context_gt": context_gt }) def construct_input(context: str, answer: str) -> str: return f"{context}{answer}" def last_token_stack(acts): """ acts - List[Tensor]: each of shape [L_i, D] Returns: Tensor: [N, D] """ last_vecs = [] for i, t in enumerate(acts): if not isinstance(t, torch.Tensor): raise TypeError(f"acts[{i}] is not a Tensor, got {type(t)}") if t.ndim != 2: raise ValueError(f"acts[{i}] should have shape [L_i, D], got {t.shape}") if t.size(0) == 0: raise ValueError(f"acts[{i}] has zero length in seq dimension") last_vecs.append(t[-1]) # [D] return torch.stack(last_vecs, dim=0) # [N, D] def load_model_and_validate_gpu(model_path,cache_dir=None,single_gpu=False,args=None): tokenizer_path = model_path tokenizer = AutoTokenizer.from_pretrained(tokenizer_path,cache_dir=cache_dir) print("Started loading model") if single_gpu: device_id = get_least_used_gpu() device = torch.device(f"cuda:{device_id}") model = AutoModelForCausalLM.from_pretrained( model_path, torch_dtype=torch.bfloat16, low_cpu_mem_usage=True, cache_dir=cache_dir, ) model.to(device) else: model = AutoModelForCausalLM.from_pretrained(model_path, device_map='auto', dtype=torch.bfloat16, low_cpu_mem_usage=True,cache_dir=cache_dir,) assert ('cpu' not in model.hf_device_map.values()) return model, tokenizer def ensure_dir(path: List[str] | str) -> None: if isinstance(path, list): for p in path: ensure_dir(p) elif not os.path.exists(path): os.makedirs(path, exist_ok=True)