File size: 3,487 Bytes
c72da75
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
"""Pinned teacher-distillation implementation, shared by job and CPU smoke."""
import json,sys
from pathlib import Path
sys.modules['vllm']=None

def trajectories(rows):
 from agenttune.agentic.events import Event,EventKind,EventLog
 logs=[]
 for row in rows:
  log=EventLog(tier='full');log.append(Event(EventKind.OBSERVATION,{'text':row['initial_prompt']}))
  for i,c in enumerate(row['calls']):
   log.append(Event(EventKind.TEXT,{'text':json.dumps(c['action'],sort_keys=True)}))
   if i<len(row['calls'])-1:log.append(Event(EventKind.OBSERVATION,{'text':json.dumps({'tool_result':c['result']},sort_keys=True)}))
  logs.append(log)
 return logs

def assistant_labels(tokenizer,messages):
 ids=tokenizer.apply_chat_template(messages,tokenize=True,return_dict=True)['input_ids'];labels=[-100]*len(ids)
 for i,message in enumerate(messages):
  if message['role']!='assistant':continue
  prefix=tokenizer.apply_chat_template(messages[:i],tokenize=True,add_generation_prompt=True,return_dict=True)['input_ids']
  full=tokenizer.apply_chat_template(messages[:i+1],tokenize=True,return_dict=True)['input_ids']
  assert ids[:len(prefix)]==prefix and ids[:len(full)]==full
  labels[len(prefix):len(full)]=ids[len(prefix):len(full)]
 return tuple(labels)

def train(student,rows,output,config):
 import torch,trl
 from transformers import AutoTokenizer
 from agenttune.agentic.distill_trainer import create_distill_trainer
 tok=AutoTokenizer.from_pretrained(student);logs=trajectories(rows)
 trainer=create_distill_trainer(student=str(student),trajectories=logs,output_dir=str(output),lora_r=16,lora_alpha=32,lora_dropout=.05,lora_targets=['q_proj','k_proj','v_proj','o_proj'],max_steps=config['max_steps'],per_device_train_batch_size=1,gradient_accumulation_steps=config['gradient_accumulation_steps'],learning_rate=2e-4,seed=20261104,bf16=config['bf16'],gradient_checkpointing=config['gradient_checkpointing'],report_to='none',assistant_only_loss=True,save_strategy='no',max_length=config['max_length'],device_map={'':'cuda'} if config['bf16'] else None)
 expected={assistant_labels(tok,row['messages']) for row in trainer.sft_dataset()};assert all(any(v!=-100 for v in e) for e in expected)
 assert max(map(len,expected))<=config['max_length'],'No truncation permitted'
 seen=[];original=trl.SFTTrainer.compute_loss
 def check(self,model,inputs,*a,**k):
  cfg=self.args
  assert cfg.assistant_only_loss and cfg.learning_rate==2e-4 and cfg.max_steps==config['max_steps']
  assert cfg.gradient_accumulation_steps==config['gradient_accumulation_steps'] and cfg.per_device_train_batch_size==1 and cfg.seed==20261104
  assert cfg.bf16==config['bf16'] and cfg.gradient_checkpointing==config['gradient_checkpointing']
  lora=model.peft_config['default'];assert lora.r==16 and lora.lora_alpha==32 and set(lora.target_modules)=={'q_proj','k_proj','v_proj','o_proj'}
  for labels in inputs['labels']:
   values=tuple(labels.tolist());assert any(values[:len(e)]==e and all(v==-100 for v in values[len(e):]) for e in expected),'Loss mask mismatch'
  seen.append(1);return original(self,model,inputs,*a,**k)
 trl.SFTTrainer.compute_loss=check
 try:trainer.train()
 finally:trl.SFTTrainer.compute_loss=original
 assert seen
 path=Path(output)/'adapter';trainer.trainer_factory.sft_trainer.model.save_pretrained(path)
 assert (path/'adapter_model.safetensors').exists() and (path/'adapter_config.json').exists()
 return trainer,{'verified_batches':len(seen),'training_rows':len(rows),'saved_adapter':True}