Merge pull request #6378 from hpcaitech/grpo-latest-rebase-fix-resume

[feat] fix resume training

Author: YeAnbang

applications/ColossalChat/coati/distributed/consumer.py

@@ -1,4 +1,3 @@
-import os
 from contextlib import nullcontext
 from typing import Any, Dict, Optional
 
@@ -7,11 +6,13 @@
 import torch
 import torch.distributed as dist
 from coati.distributed.profiling_utils import CustomProfiler
+from coati.utils import save_checkpoint
 from tqdm import tqdm
 from transformers import AutoModelForCausalLM
 
 from colossalai.booster import Booster
 from colossalai.booster.plugin import HybridParallelPlugin
+from colossalai.cluster import DistCoordinator
 from colossalai.initialize import launch
 from colossalai.nn.optimizer import HybridAdam
 from colossalai.utils import get_current_device
@@ -55,16 +56,19 @@ def __init__(
         self.enable_profiling = enable_profiling
         assert batch_size % minibatch_size == 0, "batch_size should be divisible by microbatch_size"
         self.num_microbatches = batch_size // minibatch_size
+        self.checkpoint_path = model_config.pop("checkpoint_path", None)
 
         self.model_config = model_config
         self.plugin_config = plugin_config
 
         self.device = get_current_device()
         self.lr_scheduler = None
         self.n_behind = n_behind
+        self.total_prompt_trained = 0  # for setting start index when resume training
 
     def setup(self) -> None:
         launch(self.rank, self.world_size, self.master_addr, self.master_port, local_rank=0)
+        self.coordinator = DistCoordinator()
 
         plugin_config = dict(tp_size=1, pp_size=1, precision="bf16", zero_stage=2)
         if (
@@ -143,6 +147,26 @@ def calculate_effective_group_to_raw_group_mapping(self, step):
         return effective_group_to_raw_group_mapping
 
     def loop(self) -> None:
+        self.profiler.enter("sync_model")
+        torch.cuda.empty_cache()
+        state_dict = self.state_dict()
+        if self.pp_size > 1:
+            if self.tp_rank == 0 and self.dp_rank == 0:
+                ray_broadcast_tensor_dict(
+                    state_dict,
+                    src=self.num_producers,
+                    device=self.device,
+                    group_name=f"sync_model_{self.pp_rank}",
+                )
+        else:
+            if self.rank == 0:
+                ray_broadcast_tensor_dict(
+                    state_dict, src=self.num_producers, device=self.device, group_name="sync_model"
+                )
+        del state_dict
+        torch.cuda.empty_cache()
+        self.profiler.exit("sync_model")
+
         print(
             f"Consumer{self.rank} num_update: {self.num_update_per_episode}, num_recv: {self.num_recv_per_update}, nmb: {self.num_microbatches}"
         )
@@ -208,6 +232,7 @@ def loop(self) -> None:
                                 for k, v in raw_batch.items()
                             }
                             # [batch_size, num_generations] -> [batch_size]
+                            self.total_prompt_trained += raw_batch["reward"].size(0)
                             reward = raw_batch["reward"][:, :, 0]
                             format_acc = raw_batch["format_acc"][:, :, 0]
                             ans_acc = raw_batch["ans_acc"][:, :, 0]
@@ -285,10 +310,19 @@ def loop(self) -> None:
                     if (step + 1) % self.save_interval == 0 or (step + 1) == self.num_update_per_episode:
                         if self.rank == 0:
                             print(f"Start saving policy model at step {step + 1}.")
-                        save_path = os.path.join(self.save_dir, f"modeling-episode-{episode}-step-{step + 1}")
-                        self.booster.save_model(self.policy_model, save_path, shard=True)
+                        save_checkpoint(
+                            save_dir=self.save_dir,
+                            booster=self.booster,
+                            model=self.policy_model,
+                            optimizer=self.optimizer,
+                            lr_scheduler=self.lr_scheduler,
+                            epoch=episode,
+                            step=step,
+                            batch_size=int(self.total_prompt_trained / step),
+                            coordinator=self.coordinator,
+                        )  # for setting start index when resuming training
                         if self.rank == 0:
-                            print(f"Saved model checkpoint at step {step + 1} in folder {save_path}")
+                            print(f"Saved model checkpoint at step {step + 1} in folder {self.save_dir}")
 
                     if (episode != self.num_episodes - 1 or step != self.num_update_per_episode - 1) and (
                         episode != 0 or step >= self.n_behind

applications/ColossalChat/coati/distributed/grpo_consumer.py

@@ -8,6 +8,7 @@
 from coati.distributed.loss import PolicyLoss
 from coati.distributed.utils import entropy_from_logits, memory_efficient_logprob
 from coati.trainer.utils import all_reduce_mean, all_reduce_sum
+from coati.utils import load_checkpoint
 from transformers import AutoModelForCausalLM, AutoTokenizer
 
 from colossalai.nn.lr_scheduler import CosineAnnealingWarmupLR
@@ -157,6 +158,14 @@ def setup(self):
         )
         if self.policy_loss_fn.beta > 0:
             self.reference_model, *_ = self.booster.boost(self.reference_model)
+        if self.checkpoint_path is not None:
+            load_checkpoint(
+                self.checkpoint_path,
+                self.booster,
+                self.policy_model,
+                self.optimizer,
+                self.lr_scheduler,
+            )
         self.plugin.logger.set_level("ERROR")
 
     def step(self, step_idx: int, pbar: Any, **kwargs) -> Optional[float]:

applications/ColossalChat/coati/distributed/producer.py

@@ -8,10 +8,12 @@
 import torch
 import tqdm
 import wandb
+from coati.dataset import StatefulDistributedSampler
 from coati.dataset.loader import RawConversationDataset, collate_fn_grpo
 from coati.distributed.profiling_utils import CustomProfiler
 from coati.distributed.reward.reward_fn import boxed_math_reward_fn, code_reward_fn, math_reward_fn
 from coati.distributed.reward.verifiable_reward import VerifiableReward
+from coati.utils import load_checkpoint
 from ray.util.collective import allreduce
 from ray.util.collective.types import Backend, ReduceOp
 from torch.utils.data import DataLoader, DistributedSampler
@@ -68,6 +70,7 @@ def __init__(
         self.profiler = CustomProfiler(f"P{self.producer_idx}", disabled=not enable_profiling)
 
         self.train_dataset_config = train_dataset_config
+        self.checkpoint_path = model_config.pop("checkpoint_path", None)
         self.model_config = model_config
         self.generate_config = generate_config
         self.tokenizer_config = tokenizer_config
@@ -121,7 +124,7 @@ def __init__(
         self.train_dataloader = DataLoader(
             self.train_dataset,
             batch_size=microbatch_size,
-            sampler=DistributedSampler(
+            sampler=StatefulDistributedSampler(
                 self.train_dataset,
                 num_replicas=num_producers,
                 rank=producer_idx,
@@ -133,6 +136,13 @@ def __init__(
             drop_last=True,
             collate_fn=collate_fn_grpo,
         )
+        if self.checkpoint_path is not None:
+            # resume training from checkpoint
+            start_epoch, start_step, sampler_start_idx = load_checkpoint(self.checkpoint_path, None, None, None, None)
+            self.train_dataloader.sampler.set_start_index(sampler_start_idx)
+            print(
+                f"[P{self.producer_idx}] Resume training from checkpoint {self.checkpoint_path}, start epoch {start_epoch}, start step {start_step}, sampler start index {sampler_start_idx}"
+            )
         if grpo_config["reward_fn_type"] == "think_answer_tags":
             self.evaluation_function = math_reward_fn
         elif grpo_config["reward_fn_type"] == "boxed":
@@ -203,6 +213,29 @@ def load_state_dict(self, state_dict: Dict[str, torch.Tensor]) -> None:
         raise NotImplementedError
 
     def loop(self) -> None:
+
+        torch.cuda.empty_cache()
+        self.profiler.enter("sync_model")
+        if self.consumer_pp_size > 1:
+            for pp_idx in range(self.consumer_pp_size):
+                state_dict = ray_broadcast_tensor_dict(
+                    None, self.num_producers, device=self.device, group_name=f"sync_model_{pp_idx}"
+                )
+                if "consumer_global_step" in state_dict:
+                    self.consumer_global_step = state_dict.pop("consumer_global_step").item()
+                self.load_state_dict(state_dict)
+        else:
+            state_dict = ray_broadcast_tensor_dict(
+                None, self.num_producers, device=self.device, group_name="sync_model"
+            )
+            if "consumer_global_step" in state_dict:
+                self.consumer_global_step = state_dict.pop("consumer_global_step").item()
+            self.load_state_dict(state_dict)
+        self.profiler.exit("sync_model")
+        print(f"[P{self.producer_idx}] Sync initial model done.")
+        del state_dict
+        torch.cuda.empty_cache()
+
         num_update_per_episode = len(self.train_dataloader) // self.num_microbatches
         num_valid_microbatches = num_update_per_episode * self.num_microbatches
 

applications/ColossalChat/coati/utils/ckpt_io.py

@@ -81,9 +81,12 @@ def load_checkpoint(
     """
 
     # Update booster params states.
-    booster.load_model(model=model, checkpoint=os.path.join(load_dir, "modeling"))
-    booster.load_optimizer(optimizer=optimizer, checkpoint=os.path.join(load_dir, "optimizer"))
-    booster.load_lr_scheduler(lr_scheduler=lr_scheduler, checkpoint=os.path.join(load_dir, "lr_scheduler"))
+    if model is not None:
+        booster.load_model(model=model, checkpoint=os.path.join(load_dir, "modeling"))
+    if optimizer is not None:
+        booster.load_optimizer(optimizer=optimizer, checkpoint=os.path.join(load_dir, "optimizer"))
+    if lr_scheduler is not None:
+        booster.load_lr_scheduler(lr_scheduler=lr_scheduler, checkpoint=os.path.join(load_dir, "lr_scheduler"))
 
     running_states = load_json(file_path=os.path.join(load_dir, "running_states.json"))
     return (

applications/ColossalChat/rl_example.py

@@ -18,6 +18,13 @@
 if __name__ == "__main__":
     parser = argparse.ArgumentParser()
     parser.add_argument("-m", "--model", type=str, default="Qwen/Qwen2.5-7B")
+    parser.add_argument(
+        "-cp",
+        "--checkpoint-path",
+        type=str,
+        default=None,
+        help="Path to the checkpoint to load the model from. If not provided, the model will be loaded from the model path.",
+    )
     parser.add_argument("-d", "--dataset", type=str, default="data.jsonl")
     parser.add_argument(
         "-ed",
@@ -226,8 +233,10 @@
 
     os.environ["TOKENIZERS_PARALLELISM"] = "false"  # Disable tokenizers parallelism to avoid deadlock
 
-    inference_model_config = dict(path=args.model)
-    train_model_config = dict(path=args.model, use_flash_attention_2=True, use_cache=False)
+    inference_model_config = dict(path=args.model, checkpoint_path=args.checkpoint_path)
+    train_model_config = dict(
+        path=args.model, use_flash_attention_2=True, use_cache=False, checkpoint_path=args.checkpoint_path
+    )
     generate_config = dict(top_k=args.top_k, top_p=args.top_p, temperature=args.temperature)
 
     if args.backend == "transformers":