228 lines
9.5 KiB
Python
228 lines
9.5 KiB
Python
"""
|
||
训练工具函数集合
|
||
"""
|
||
import os
|
||
import sys
|
||
__package__ = "trainer"
|
||
sys.path.append(os.path.abspath(os.path.join(os.path.dirname(__file__), '..')))
|
||
import random
|
||
import math
|
||
import numpy as np
|
||
import torch
|
||
import torch.distributed as dist
|
||
from torch.nn.parallel import DistributedDataParallel
|
||
from torch.utils.data import Sampler
|
||
from transformers import AutoTokenizer, AutoModel, AutoModelForSequenceClassification
|
||
from model.model_minimind import MiniMindForCausalLM
|
||
|
||
def get_model_params(model, config):
|
||
total = sum(p.numel() for p in model.parameters()) / 1e6
|
||
n_routed = getattr(config, 'n_routed_experts', getattr(config, 'num_experts', 0))
|
||
n_active = getattr(config, 'num_experts_per_tok', 0)
|
||
n_shared = getattr(config, 'n_shared_experts', 0)
|
||
expert = sum(p.numel() for n, p in model.named_parameters() if 'mlp.experts.0.' in n) / 1e6
|
||
shared_expert = sum(p.numel() for n, p in model.named_parameters() if 'mlp.shared_experts.0.' in n) / 1e6
|
||
base = total - (expert * n_routed) - (shared_expert * n_shared)
|
||
active = base + (expert * n_active) + (shared_expert * n_shared)
|
||
if active < total: Logger(f'Model Params: {total:.2f}M-A{active:.2f}M')
|
||
else: Logger(f'Model Params: {total:.2f}M')
|
||
|
||
|
||
def is_main_process():
|
||
return not dist.is_initialized() or dist.get_rank() == 0
|
||
|
||
|
||
def Logger(content):
|
||
if is_main_process():
|
||
print(content)
|
||
|
||
|
||
def get_lr(current_step, total_steps, lr):
|
||
return lr*(0.1 + 0.45*(1 + math.cos(math.pi * current_step / total_steps)))
|
||
|
||
|
||
# A100/H100/3090 等常见卡的 bf16 理论峰值 TFLOPS。MoE 等价 FLOPs 是 active params 不是 total，
|
||
# active params 算出来的 MFU 才有意义；MoE 的 active 比例由 num_experts_per_tok / num_experts 决定。
|
||
_GPU_PEAK_BF16_TFLOPS = {
|
||
'A100': 312, 'A100-SXM4-80GB': 312, 'A100-SXM4-40GB': 312, 'A100-PCIE-80GB': 312, 'A100-PCIE-40GB': 312,
|
||
'H100': 989, 'H100-SXM5-80GB': 989, 'H100-PCIE-80GB': 756, 'H100 80GB HBM3': 989,
|
||
'H200': 989,
|
||
'RTX 3090': 71, 'GeForce RTX 3090': 71, 'RTX 3090 Ti': 80,
|
||
'RTX 4090': 165, 'GeForce RTX 4090': 165,
|
||
'V100': 125, 'V100-SXM2-32GB': 125,
|
||
'L40': 181, 'L40S': 362,
|
||
}
|
||
|
||
|
||
def detect_gpu_peak_tflops_bf16(device_index: int = 0, override: float = None) -> float:
|
||
"""Look up the bf16 TensorCore peak for the current GPU; fall back to A100 (312)."""
|
||
if override is not None and override > 0:
|
||
return float(override)
|
||
if not torch.cuda.is_available():
|
||
return 312.0
|
||
name = torch.cuda.get_device_name(device_index)
|
||
for key, peak in _GPU_PEAK_BF16_TFLOPS.items():
|
||
if key in name:
|
||
return float(peak)
|
||
Logger(f'[MFU] Unknown GPU "{name}", falling back to A100 (312 TFLOPS). Pass --gpu_peak_tflops to override.')
|
||
return 312.0
|
||
|
||
|
||
def compute_model_flops_per_token(model, config=None) -> float:
|
||
"""Compute the "model FLOPs" per training token (forward + backward).
|
||
|
||
Standard approximation for dense transformers: 6 × P, where P is the total
|
||
trainable params count. For MoE we use active params = base + experts_per_tok × per_expert.
|
||
"""
|
||
raw = model.module if isinstance(model, DistributedDataParallel) else model
|
||
raw = getattr(raw, '_orig_mod', raw)
|
||
total = sum(p.numel() for p in raw.parameters())
|
||
if config is not None and getattr(config, 'use_moe', False):
|
||
n_routed = getattr(config, 'num_experts', 0)
|
||
n_active = getattr(config, 'num_experts_per_tok', 1)
|
||
if n_routed > 0:
|
||
expert = sum(p.numel() for n, p in raw.named_parameters() if 'mlp.experts.0.' in n)
|
||
base = total - expert * n_routed
|
||
active = base + expert * n_active
|
||
return 6.0 * active
|
||
return 6.0 * total
|
||
|
||
|
||
def init_distributed_mode():
|
||
if int(os.environ.get("RANK", -1)) == -1:
|
||
return 0 # 非DDP模式
|
||
|
||
dist.init_process_group(backend="nccl")
|
||
local_rank = int(os.environ["LOCAL_RANK"])
|
||
torch.cuda.set_device(local_rank)
|
||
return local_rank
|
||
|
||
|
||
def setup_seed(seed: int, deterministic: bool = False):
|
||
# cudnn 决定性 + 禁 benchmark 是预训练吞吐杀手（A100 上 ~3-5x 慢）。
|
||
# 默认走 benchmark=True 让 cudnn 选最快 kernel；要严格复现传 deterministic=True。
|
||
random.seed(seed)
|
||
np.random.seed(seed)
|
||
torch.manual_seed(seed)
|
||
torch.cuda.manual_seed(seed)
|
||
torch.cuda.manual_seed_all(seed)
|
||
torch.backends.cudnn.deterministic = deterministic
|
||
torch.backends.cudnn.benchmark = not deterministic
|
||
# A100 TF32 矩阵乘加速（autocast bf16 路径之外的 fp32 matmul 才生效，但无副作用）
|
||
torch.set_float32_matmul_precision('high')
|
||
|
||
def lm_checkpoint(lm_config, weight='full_sft', model=None, optimizer=None, epoch=0, step=0, wandb=None, save_dir='../checkpoints', **kwargs):
|
||
os.makedirs(save_dir, exist_ok=True)
|
||
moe_path = '_moe' if lm_config.use_moe else ''
|
||
ckp_path = f'{save_dir}/{weight}_{lm_config.hidden_size}{moe_path}.pth'
|
||
resume_path = f'{save_dir}/{weight}_{lm_config.hidden_size}{moe_path}_resume.pth'
|
||
|
||
if model is not None:
|
||
raw_model = model.module if isinstance(model, DistributedDataParallel) else model
|
||
raw_model = getattr(raw_model, '_orig_mod', raw_model)
|
||
state_dict = raw_model.state_dict()
|
||
state_dict = {k: v.half().cpu() for k, v in state_dict.items()}
|
||
ckp_tmp = ckp_path + '.tmp'
|
||
torch.save(state_dict, ckp_tmp)
|
||
os.replace(ckp_tmp, ckp_path)
|
||
wandb_id = None
|
||
if wandb:
|
||
if hasattr(wandb, 'get_run'):
|
||
run = wandb.get_run()
|
||
wandb_id = getattr(run, 'id', None) if run else None
|
||
else:
|
||
wandb_id = getattr(wandb, 'id', None)
|
||
|
||
resume_data = {
|
||
'model': state_dict,
|
||
'optimizer': optimizer.state_dict(),
|
||
'epoch': epoch,
|
||
'step': step,
|
||
'world_size': dist.get_world_size() if dist.is_initialized() else 1,
|
||
'wandb_id': wandb_id
|
||
}
|
||
for key, value in kwargs.items():
|
||
if value is not None:
|
||
if hasattr(value, 'state_dict'):
|
||
raw_value = value.module if isinstance(value, DistributedDataParallel) else value
|
||
raw_value = getattr(raw_value, '_orig_mod', raw_value)
|
||
resume_data[key] = raw_value.state_dict()
|
||
else:
|
||
resume_data[key] = value
|
||
|
||
resume_tmp = resume_path + '.tmp'
|
||
torch.save(resume_data, resume_tmp)
|
||
os.replace(resume_tmp, resume_path)
|
||
del state_dict, resume_data
|
||
torch.cuda.empty_cache()
|
||
else: # 加载模式
|
||
if os.path.exists(resume_path):
|
||
ckp_data = torch.load(resume_path, map_location='cpu')
|
||
saved_ws = ckp_data.get('world_size', 1)
|
||
current_ws = dist.get_world_size() if dist.is_initialized() else 1
|
||
if saved_ws != current_ws:
|
||
ckp_data['step'] = ckp_data['step'] * saved_ws // current_ws
|
||
Logger(f'GPU数量变化({saved_ws}→{current_ws})，step已自动转换为{ckp_data["step"]}')
|
||
return ckp_data
|
||
return None
|
||
|
||
|
||
def init_model(lm_config, from_weight='pretrain', tokenizer_path='../model', save_dir='../out', device='cuda'):
|
||
tokenizer = AutoTokenizer.from_pretrained(tokenizer_path)
|
||
model = MiniMindForCausalLM(lm_config)
|
||
|
||
if from_weight!= 'none':
|
||
moe_suffix = '_moe' if lm_config.use_moe else ''
|
||
weight_path = f'{save_dir}/{from_weight}_{lm_config.hidden_size}{moe_suffix}.pth'
|
||
weights = torch.load(weight_path, map_location=device)
|
||
model.load_state_dict(weights, strict=False)
|
||
|
||
get_model_params(model, lm_config)
|
||
Logger(f'Trainable Params: {sum(p.numel() for p in model.parameters() if p.requires_grad) / 1e6:.3f}M')
|
||
return model.to(device), tokenizer
|
||
|
||
|
||
class SkipBatchSampler(Sampler):
|
||
def __init__(self, sampler, batch_size, skip_batches=0):
|
||
self.sampler = sampler
|
||
self.batch_size = batch_size
|
||
self.skip_batches = skip_batches
|
||
|
||
def __iter__(self):
|
||
batch = []
|
||
skipped = 0
|
||
for idx in self.sampler:
|
||
batch.append(idx)
|
||
if len(batch) == self.batch_size:
|
||
if skipped < self.skip_batches:
|
||
skipped += 1
|
||
batch = []
|
||
continue
|
||
yield batch
|
||
batch = []
|
||
if len(batch) > 0 and skipped >= self.skip_batches:
|
||
yield batch
|
||
|
||
def __len__(self):
|
||
total_batches = (len(self.sampler) + self.batch_size - 1) // self.batch_size
|
||
return max(0, total_batches - self.skip_batches)
|
||
|
||
|
||
class LMForRewardModel:
|
||
def __init__(self, model_path, device="cuda", dtype=torch.float16):
|
||
self.tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
|
||
self.model = AutoModel.from_pretrained(model_path, torch_dtype=dtype, trust_remote_code=True)
|
||
self.model = self.model.to(device).eval()
|
||
self.device = device
|
||
|
||
@torch.no_grad()
|
||
def get_score(self, messages, response):
|
||
history_text = "\n".join([f"{m['role']}: {m['content']}" for m in messages[:-1]])
|
||
last_query = messages[-1]['content'] if messages else ""
|
||
message_context = f"{history_text}\n以上是对话历史。我的新问题是：\n{last_query}" if history_text else last_query
|
||
eval_messages = [
|
||
{"role": "user", "content": message_context},
|
||
{"role": "assistant", "content": response}
|
||
]
|
||
score = self.model.get_score(self.tokenizer, eval_messages)
|
||
return max(min(score, 3.0), -3.0) |