Merge branch 'grpo_optimization' of https://github.com/hpcaitech/ColossalAI into grpo_optimization

Author: YeAnbang

applications/ColossalChat/coati/distributed/consumer.py

@@ -6,6 +6,7 @@
 import ray.util.collective as cc
 import torch
 import torch.distributed as dist
+from coati.distributed.profiling_utils import CustomProfiler
 from tqdm import tqdm
 from transformers import AutoModelForCausalLM
 
@@ -36,6 +37,8 @@ def __init__(
         minibatch_size: int = 1,
         save_interval: int = 100,
         save_dir: str = "./model",
+        enable_profiling: bool = False,
+        n_behind: int = 0,
     ):
         self.num_producers = num_producers
         self.num_episodes = num_episodes
@@ -49,6 +52,7 @@ def __init__(
         self.minibatch_size = minibatch_size
         self.save_interval = save_interval
         self.save_dir = save_dir
+        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
 
@@ -57,6 +61,7 @@ def __init__(
 
         self.device = get_current_device()
         self.lr_scheduler = None
+        self.n_behind = n_behind
 
     def setup(self) -> None:
         launch(self.rank, self.world_size, self.master_addr, self.master_port, local_rank=0)
@@ -94,13 +99,45 @@ def setup(self) -> None:
 
         self.buffer = []
         self.recv_cnt = 0
+        self.profiler = CustomProfiler(f"C{self.rank}", disabled=not self.enable_profiling)
 
     def state_dict(self) -> Dict[str, torch.Tensor]:
         raise NotImplementedError
 
     def step(self, step_idx: int, **kwargs) -> Optional[float]:
         raise NotImplementedError
 
+    def prepare_mini_batch(self, effective_group_to_raw_group_mapping: Dict[int, int]) -> Dict[str, torch.Tensor]:
+        """
+        Prepare a mini-batch from the effective group to raw group mapping.
+        This method is used to create a mini-batch for training.
+        """
+        batches = [
+            self.buffer[effective_group_to_raw_group_mapping[i]]
+            for i in range(self.dp_rank * self.minibatch_size, (self.dp_rank + 1) * self.minibatch_size)
+        ]
+        # every dp_rank will receive a complete mini-batch, no need to sync within step() later
+        # each mini-batch use the first self.dp_size * minibatch_size effective samples
+        raw_mini_batches = self.buffer[
+            : effective_group_to_raw_group_mapping[self.dp_size * self.minibatch_size - 1] + 1
+        ]  # include the last effective sample
+        raw_mini_batches_metric_dict = {
+            "raw_train_mini_batch_reward": [t[1] for t in raw_mini_batches],
+            "raw_train_mini_batch_format_acc": [t[2] for t in raw_mini_batches],
+            "raw_train_mini_batch_ans_acc": [t[3] for t in raw_mini_batches],
+            "raw_train_mini_batch_response_len": [t[4] for t in raw_mini_batches],
+        }
+        batch = bind_batch([t[0] for t in batches])
+        batch = post_recv(batch)
+        return batch, raw_mini_batches_metric_dict
+
+    def calculate_effective_group_to_raw_group_mapping(self):
+        effective_group_to_raw_group_mapping = {}
+        for buffer_idx in range(len(self.buffer)):
+            if self.buffer[buffer_idx][0] is not None:
+                effective_group_to_raw_group_mapping[len(effective_group_to_raw_group_mapping)] = buffer_idx
+        return effective_group_to_raw_group_mapping
+
     def loop(self) -> None:
         print(
             f"Consumer{self.rank} num_update: {self.num_update_per_episode}, num_recv: {self.num_recv_per_update}, nmb: {self.num_microbatches}"
@@ -112,14 +149,49 @@ def loop(self) -> None:
                 disable=self.rank != 0,
             ) as pbar:
                 for step in pbar:
+                    torch.cuda.reset_peak_memory_stats()
                     i = 0
                     for _ in range(self.num_recv_per_update):
+                        # after sync model, do not wait for more data to arrive as rollout takes time, use buffered data
+                        effective_group_to_raw_group_mapping = self.calculate_effective_group_to_raw_group_mapping()
+                        while len(effective_group_to_raw_group_mapping) > max(
+                            self.dp_size * self.batch_size
+                            - self.dp_size
+                            * self.minibatch_size
+                            * self.grpo_config.get("num_minibatch_during_rollout", 1),
+                            self.dp_size * self.minibatch_size,
+                        ):
+                            self.profiler.log(
+                                f"Still have {len(effective_group_to_raw_group_mapping)} effective groups, greater than {self.dp_size * self.minibatch_size}, start training"
+                            )
+                            batch, raw_mini_batches_metric_dict = self.prepare_mini_batch(
+                                effective_group_to_raw_group_mapping
+                            )
+                            self.profiler.enter("step")
+                            loss = self.step(i, pbar, **batch, **raw_mini_batches_metric_dict)
+                            self.profiler.exit("step")
+                            self.buffer = self.buffer[
+                                effective_group_to_raw_group_mapping[self.dp_size * self.minibatch_size - 1] + 1 :
+                            ]
+                            # recalculate the effective group to raw group mapping
+                            effective_group_to_raw_group_mapping_size_before = len(effective_group_to_raw_group_mapping)
+                            effective_group_to_raw_group_mapping = self.calculate_effective_group_to_raw_group_mapping()
+                            assert (
+                                len(effective_group_to_raw_group_mapping)
+                                == effective_group_to_raw_group_mapping_size_before - self.dp_size * self.minibatch_size
+                            )
+                            if loss is not None:
+                                pbar.set_postfix({"loss": loss})
+                            i += 1
+
                         # receive data from producers
                         for r in range(self.num_producers):
                             print(f"[T{dist.get_rank()}] Recv data episode {episode} step {step} from {r}")
+                            self.profiler.enter(f"recv_broadcast_data_P{r}")
                             raw_batch = ray_broadcast_tensor_dict(
                                 None, src=0, device=self.device, group_name=f"sync_data_{r}"
                             )
+                            self.profiler.exit(f"recv_broadcast_data_P{r}")
                             # calculate group reward et al. filtering. As only the filtered group will be used for training (which is incomplete),
                             # we need to calculate the metrics before filtering here for logging
                             # [batch_size, num_generations, ...] -> [batch_size * num_generations, ...]
@@ -161,49 +233,29 @@ def loop(self) -> None:
                                     f"[T{dist.get_rank()}] Filter recv data: {len(raw_batch)} -> {torch.sum(effective_group_mask).cpu().item()} effective groups"
                                 )
                         # mapping the effective group to the raw group for indexing
-                        effective_group_to_raw_group_mapping = {}
-                        for buffer_idx in range(len(self.buffer)):
-                            if self.buffer[buffer_idx][0] is not None:
-                                effective_group_to_raw_group_mapping[len(effective_group_to_raw_group_mapping)] = (
-                                    buffer_idx
-                                )
+                        effective_group_to_raw_group_mapping = self.calculate_effective_group_to_raw_group_mapping()
                         print(
                             f"[T{dist.get_rank()}] Collect Effective Prompt: {len(effective_group_to_raw_group_mapping)}/{self.dp_size * self.minibatch_size}"
                         )
 
-                        while len(effective_group_to_raw_group_mapping) >= self.dp_size * self.minibatch_size:
+                        while len(effective_group_to_raw_group_mapping) > self.dp_size * self.batch_size:
+                            self.profiler.log(
+                                f"Received {len(effective_group_to_raw_group_mapping)} effective groups, greater than {self.dp_size * self.batch_size}, start training after recv"
+                            )
+                            # always keep at least dp_size * batch_size effective samples in the buffer for training during the rollout times after each sync model
                             # on each dp_rank, we use minibatch_size effective samples to form a batch
-                            batches = [
-                                self.buffer[effective_group_to_raw_group_mapping[i]]
-                                for i in range(
-                                    self.dp_rank * self.minibatch_size, (self.dp_rank + 1) * self.minibatch_size
-                                )
-                            ]
-                            # every dp_rank will receive a complete mini-batch, no need to sync within step() later
-                            # each mini-batch use the first self.dp_size * minibatch_size effective samples
-                            raw_mini_batches = self.buffer[
-                                : effective_group_to_raw_group_mapping[self.dp_size * self.minibatch_size - 1] + 1
-                            ]  # include the last effective sample
-                            raw_mini_batches_metric_dict = {
-                                "raw_train_mini_batch_reward": [t[1] for t in raw_mini_batches],
-                                "raw_train_mini_batch_format_acc": [t[2] for t in raw_mini_batches],
-                                "raw_train_mini_batch_ans_acc": [t[3] for t in raw_mini_batches],
-                                "raw_train_mini_batch_response_len": [t[4] for t in raw_mini_batches],
-                            }
-                            batch = bind_batch([t[0] for t in batches])
-                            batch = post_recv(batch)
+                            batch, raw_mini_batches_metric_dict = self.prepare_mini_batch(
+                                effective_group_to_raw_group_mapping
+                            )
+                            self.profiler.enter("step")
                             loss = self.step(i, pbar, **batch, **raw_mini_batches_metric_dict)
+                            self.profiler.exit("step")
                             self.buffer = self.buffer[
                                 effective_group_to_raw_group_mapping[self.dp_size * self.minibatch_size - 1] + 1 :
                             ]
                             # recalculate the effective group to raw group mapping
                             effective_group_to_raw_group_mapping_size_before = len(effective_group_to_raw_group_mapping)
-                            effective_group_to_raw_group_mapping = {}
-                            for buffer_idx in range(len(self.buffer)):
-                                if self.buffer[buffer_idx][0] is not None:
-                                    effective_group_to_raw_group_mapping[len(effective_group_to_raw_group_mapping)] = (
-                                        buffer_idx
-                                    )
+                            effective_group_to_raw_group_mapping = self.calculate_effective_group_to_raw_group_mapping()
                             assert (
                                 len(effective_group_to_raw_group_mapping)
                                 == effective_group_to_raw_group_mapping_size_before - self.dp_size * self.minibatch_size
@@ -221,13 +273,16 @@ def loop(self) -> None:
                         if self.rank == 0:
                             print(f"Saved model checkpoint at step {step + 1} in folder {save_path}")
 
-                    if episode != self.num_episodes - 1 or step != self.num_update_per_episode - 1:
+                    if (episode != self.num_episodes - 1 or step != self.num_update_per_episode - 1) and (
+                        episode != 0 or step >= self.n_behind
+                    ):
                         if self.pp_size > 1:
                             print(
                                 f"[T{dist.get_rank()}] Sync model PP stage {self.pp_rank} episode {episode} step {step}"
                             )
                         else:
                             print(f"[T{dist.get_rank()}] Sync model episode {episode} step {step}")
+                        self.profiler.enter("sync_model")
                         torch.cuda.empty_cache()
                         state_dict = self.state_dict()
                         if self.pp_size > 1:
@@ -245,6 +300,12 @@ def loop(self) -> None:
                                 )
                         del state_dict
                         torch.cuda.empty_cache()
+                        self.profiler.exit("sync_model")
+                    self.profiler.log(f"Peak memory usage: {torch.cuda.max_memory_allocated() / 1024 ** 2:.2f} MB")
+
+    def __del__(self):
+        if hasattr(self, "profiler"):
+            self.profiler.close()
 
 
 @ray.remote

applications/ColossalChat/coati/distributed/grpo_consumer.py

@@ -38,6 +38,8 @@ def __init__(
         project_name: str = None,
         run_name: str = None,
         wandb_group_name: str = None,
+        enable_profiling: bool = False,
+        n_behind: int = 0,
     ):
         print(f"Using GRPO config: {grpo_config}")
         if (
@@ -63,6 +65,8 @@ def __init__(
             minibatch_size,
             save_interval=save_interval,
             save_dir=save_dir,
+            enable_profiling=enable_profiling,
+            n_behind=n_behind,
         )
         path = model_config.pop("path")
         self.policy_model = AutoModelForCausalLM.from_pretrained(path, **model_config)

applications/ColossalChat/coati/distributed/launch.py

@@ -57,6 +57,8 @@ def launch_distributed(
     eval_generation_config: Optional[Dict[str, Any]] = None,
     log_rollout_interval: int = 20,
     rollout_save_dir: str = "./rollout",
+    enable_profiling: bool = False,
+    n_behind: int = 0,
 ):
     if core_algo not in ALGO_MAP:
         raise NotImplementedError(f"{core_algo} is not supported yet.")
@@ -132,6 +134,8 @@ def launch_distributed(
             wandb_group_name=wandb_group_name,
             log_rollout_interval=log_rollout_interval,
             rollout_log_file=rollout_log_file,
+            enable_profiling=enable_profiling,
+            n_behind=n_behind,
         )
         producer_procs.append(producer)
     ray.get([p.setup.remote() for p in producer_procs])
@@ -171,6 +175,8 @@ def launch_distributed(
             project_name=project_name,
             run_name=run_name,
             wandb_group_name=wandb_group_name,
+            enable_profiling=enable_profiling,
+            n_behind=n_behind,
         )
         consumer_procs.append(consumer)
     ray.get([p.setup.remote() for p in consumer_procs])

applications/ColossalChat/coati/distributed/producer.py

@@ -9,6 +9,7 @@
 import tqdm
 import wandb
 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 ray.util.collective import allreduce
@@ -52,6 +53,8 @@ def __init__(
         wandb_group_name: str = None,
         log_rollout_interval: int = 20,
         rollout_log_file: str = "./rollout_log.jsonl",
+        enable_profiling: bool = False,
+        n_behind: int = 0,
     ):
         self.producer_idx = producer_idx
         self.num_producers = num_producers
@@ -62,6 +65,7 @@ def __init__(
         assert batch_size % microbatch_size == 0
         self.num_microbatches = batch_size // microbatch_size
         self.latest_eval_step = -1
+        self.profiler = CustomProfiler(f"P{self.producer_idx}", disabled=not enable_profiling)
 
         self.train_dataset_config = train_dataset_config
         self.model_config = model_config
@@ -75,6 +79,7 @@ def __init__(
         self.log_rollout_interval = log_rollout_interval
         self.latest_rollout_log_step = -1
         self.grpo_config = grpo_config
+        self.n_behind = n_behind
         reward_model_kwargs = {
             k: v
             for k, v in grpo_config.items()
@@ -268,11 +273,14 @@ def loop(self) -> None:
                             self.wandb_run.log(to_log_msg, step=self.consumer_global_step)
                         self.eval_mode = False
                         self.latest_eval_step = self.consumer_global_step
+                self.profiler.enter("rollout")
                 outputs = self.rollout(**batch)
+                self.profiler.exit("rollout")
                 outputs["temperature"] = torch.tensor(
                     [self.model.generate_config["temperature"]] * outputs["input_ids"].size(0)
                 ).to(outputs["input_ids"].device)
                 bs, num_gen = outputs["input_ids"].size(0), outputs["input_ids"].size(1)
+                self.profiler.enter("calculate_reward")
                 if self.grpo_config["reward_fn_type"] == "code":
                     test_cases = []
                     for prompt_id in range(bs):
@@ -310,20 +318,26 @@ def loop(self) -> None:
                     outputs.pop("gt_answer")
                 if "test_cases" in outputs:
                     outputs.pop("test_cases")
+                self.profiler.exit("calculate_reward")
 
                 print(f"[P{self.producer_idx}] Send data {[(k, v.shape) for k, v in outputs.items()]}")
                 outputs = pre_send(outputs)
+                self.profiler.enter("send_broadcast_data")
                 ray_broadcast_tensor_dict(
                     outputs, src=0, device=self.device, group_name=f"sync_data_{self.producer_idx}"
                 )
-                if (i + 1) % self.num_microbatches == 0 and (
-                    episode != self.num_episodes - 1 or i != num_valid_microbatches - 1
+                self.profiler.exit("send_broadcast_data")
+                if (
+                    (i + 1) % self.num_microbatches == 0
+                    and (episode != self.num_episodes - 1 or i != num_valid_microbatches - 1)
+                    and (episode != 0 or (i + 1) > self.n_behind * self.num_microbatches)
                 ):
                     if isinstance(self.model, BACKEND_MAP["vllm"]) and self.model.model_config.get(
                         "enable_sleep_mode", False
                     ):
                         self.model.llm.sleep()  # revict KV_cache to avoid OOM
                     # don't sync model for last iteration
+                    self.profiler.enter("sync_model")
                     torch.cuda.empty_cache()
 
                     if self.consumer_pp_size > 1:
@@ -349,6 +363,7 @@ def loop(self) -> None:
                         self.load_state_dict(state_dict)
                     del state_dict
                     torch.cuda.empty_cache()
+                    self.profiler.exit("sync_model")
                     if isinstance(self.model, BACKEND_MAP["vllm"]) and self.model.model_config.get(
                         "enable_sleep_mode", False
                     ):
@@ -364,6 +379,9 @@ def loop(self) -> None:
                             "temperature"
                         ] + ratio * 0.9
 
+    def __del__(self):
+        self.profiler.close()
+
 
 @ray.remote
 class SimpleProducer(BaseProducer):
@@ -392,6 +410,8 @@ def __init__(
         wandb_group_name: str = None,
         log_rollout_interval: int = 20,
         rollout_log_file: str = "./rollout_log.jsonl",
+        enable_profiling: bool = False,
+        n_behind: int = 0,
     ):
         super().__init__(
             producer_idx,
@@ -415,6 +435,8 @@ def __init__(
             wandb_group_name=wandb_group_name,
             log_rollout_interval=log_rollout_interval,
             rollout_log_file=rollout_log_file,
+            enable_profiling=enable_profiling,
+            n_behind=n_behind,
         )
         self.model = self.backend_cls(model_config, generate_config, self.tokenizer, num_generations)
         self.eval_generation_config = copy.deepcopy(self.model.generate_config)