Merge pull request #6250 from hpcaitech/grpo-latest-dev

[feat] Fix Vllm, logprob, add filtering, temperature annealing, lr descent

applications/ColossalChat/coati/distributed/consumer.py

@@ -57,6 +57,7 @@ class BaseConsumer:
         assert self.plugin_config.get("pp_size", 1) == 1, "pp_size > 1 is not supported now"
 
         self.device = get_current_device()
+        self.lr_scheduler = None
 
     def setup(self) -> None:
         for i in range(self.num_producers):
@@ -121,6 +122,8 @@ class BaseConsumer:
                                 pbar.set_postfix({"loss": loss})
                             i += 1
                     assert len(self.buffer) == 0
+                    if self.lr_scheduler is not None:
+                        self.lr_scheduler.step()
                     if (step + 1) % self.save_interval == 0:
                         if self.rank == 0:
                             print(f"Start saving policy model at step {step + 1}.")

applications/ColossalChat/coati/distributed/grpo_consumer.py

@@ -1,8 +1,11 @@
+import json
+import os
 from contextlib import nullcontext
 from typing import Optional
 
 import ray
 import torch
+import torch.distributed as dist
 import wandb
 from coati.distributed.consumer import BaseConsumer
 from coati.distributed.loss import PolicyLoss
@@ -12,6 +15,7 @@ from coati.distributed.utils import calc_action_log_probs
 from coati.trainer.utils import all_reduce_mean
 from transformers import AutoModelForCausalLM, AutoTokenizer
 
+from colossalai.nn.lr_scheduler import CosineAnnealingWarmupLR
 from colossalai.nn.optimizer import HybridAdam
 
 
@@ -31,8 +35,10 @@ class GRPOConsumer(BaseConsumer):
         model_config,
         plugin_config,
         microbatch_size=1,
-        num_generations=4,
+        num_generations=8,
         use_wandb=True,
+        generate_config=None,
+        training_config={},
     ):
         super().__init__(
             num_producers,
@@ -52,7 +58,7 @@ class GRPOConsumer(BaseConsumer):
         self.policy_model = AutoModelForCausalLM.from_pretrained(path, **model_config)
         self.policy_model.train()
         self.policy_model.gradient_checkpointing_enable()
-        self.optimizer = HybridAdam(self.policy_model.parameters(), lr=1e-6)
+        self.optimizer = HybridAdam(self.policy_model.parameters(), lr=training_config.get("lr", 1e-6))
         self.accum_loss = torch.zeros(1, device=self.device)
         self.accum_reward = torch.zeros(1, device=self.device)
         self.accum_kl = torch.zeros(1, device=self.device)
@@ -61,6 +67,7 @@ class GRPOConsumer(BaseConsumer):
         self.accum_advantages = torch.zeros(1, device=self.device)
         self.accum_response_length = torch.zeros(1, device=self.device)
         self.accum_count = 0
+        self.generate_config = generate_config
 
         # Reference model is initialized from policy model.
         self.reference_model = AutoModelForCausalLM.from_pretrained(path, **model_config)
@@ -69,6 +76,9 @@ class GRPOConsumer(BaseConsumer):
         self.tokenizer = AutoTokenizer.from_pretrained(path)
         self.pad_token_id = self.tokenizer.pad_token_id
         self.num_generations = num_generations
+        self.filter_range = training_config.get("filter_range", None)
+        if self.filter_range is not None:
+            assert len(self.filter_range) == 2, "Filter range should have 2 values."
 
         # Initialize verifiable reward.
         response_format_tags = {
@@ -84,11 +94,21 @@ class GRPOConsumer(BaseConsumer):
         self.policy_loss_fn = PolicyLoss()
         self.global_step = 0
         if use_wandb and self.rank == 0:
-            self.wandb_run = wandb.init(project="GRPO-V1", sync_tensorboard=True)
+            name = f"{generate_config['backend']}_bs_{self.batch_size*self.world_size}_temp_{generate_config['temperature']:.01f}_top_p_{generate_config['top_p']:.02f}"
+            self.wandb_run = wandb.init(project="GRPO-V1", sync_tensorboard=True, dir="./wandb", name=name)
+
+        self.lr_scheduler = CosineAnnealingWarmupLR(
+            optimizer=self.optimizer,
+            total_steps=min(self.num_episodes, 4) * self.num_update_per_episode,
+            warmup_steps=0,
+            eta_min=0.1 * training_config.get("lr", 1e-6),
+        )
 
     def setup(self):
         super().setup()
-        self.policy_model, self.optimizer, *_ = self.booster.boost(self.policy_model, self.optimizer)
+        self.policy_model, self.optimizer, _, _, self.lr_scheduler = self.booster.boost(
+            self.policy_model, self.optimizer, lr_scheduler=self.lr_scheduler
+        )
         self.reference_model, *_ = self.booster.boost(self.reference_model)
 
     def step(self, step_idx: int, **kwargs) -> Optional[float]:
@@ -113,7 +133,6 @@ class GRPOConsumer(BaseConsumer):
         response_length = torch.sum(action_mask, dim=1).to(torch.float32)
 
         need_update = (step_idx + 1) % self.num_microbatches == 0
-
         ctx = nullcontext() if need_update else self.booster.no_sync(self.policy_model, self.optimizer)
         with ctx:
             policy_model_logits = self.policy_model(
@@ -121,7 +140,10 @@ class GRPOConsumer(BaseConsumer):
                 attention_mask=data["attention_mask"],
             )["logits"]
             action_log_probs = calc_action_log_probs(
-                policy_model_logits, data["input_ids"], num_action, self.plugin.shard_config
+                policy_model_logits / self.generate_config["temperature"],
+                data["input_ids"],
+                num_action,
+                self.plugin.shard_config,
             )
 
             with torch.no_grad():
@@ -130,7 +152,10 @@ class GRPOConsumer(BaseConsumer):
                     attention_mask=data["attention_mask"],
                 )["logits"]
             reference_action_log_probs = calc_action_log_probs(
-                reference_model_logits, data["input_ids"], num_action, self.plugin.shard_config
+                reference_model_logits / self.generate_config["temperature"],
+                data["input_ids"],
+                num_action,
+                self.plugin.shard_config,
             )
 
             per_token_kl = (
@@ -149,21 +174,31 @@ class GRPOConsumer(BaseConsumer):
             acc_reward = torch.tensor([value[2] for value in reward_group]).to(data["input_ids"].device)
 
             # [batch_size, num_generations]
+
             group_reward = reward.view(-1, self.num_generations)
+            reward_mean = group_reward.mean(dim=1)
+            # filter out the reward that is too high (all sample gets full score) or too low (all sample gets 0 score),
+            loss_mask = (
+                None
+                if self.filter_range is None
+                else torch.logical_and(
+                    reward_mean > self.filter_range[0], reward_mean < self.filter_range[1]
+                ).repeat_interleave(self.num_generations, dim=0)
+            )
 
             # [batch_size x num_generations]
-            reward_mean = group_reward.mean(dim=1).repeat_interleave(self.num_generations, dim=0)
+            reward_mean = reward_mean.repeat_interleave(self.num_generations, dim=0)
             reward_std = group_reward.std(dim=1).repeat_interleave(self.num_generations, dim=0)
             # [batch_size x num_generations]
             advantages = (reward - reward_mean) / (reward_std + 1e-4)
 
-            # Calculate Loss
             loss, skip_update, _ = self.policy_loss_fn(
                 action_log_probs,
                 old_action_log_probs,
                 advantages.unsqueeze(dim=-1).repeat_interleave(action_log_probs.size(-1), dim=-1),
                 per_token_kl,
                 action_mask,
+                loss_mask=loss_mask,
             )
 
             if not skip_update:
@@ -207,13 +242,15 @@ class GRPOConsumer(BaseConsumer):
                 )
                 self.wandb_run.log(
                     {
+                        "metrics/reward": self.accum_reward.item() / self.accum_count,
+                        "metrics/format_reward": self.accum_format_reward.item() / self.accum_count,
+                        "metrics/acc_reward": self.accum_acc_reward.item() / self.accum_count,
+                        "metrics/response_length": self.accum_response_length.item() / self.accum_count,
                         "train/loss": self.accum_loss.item() / self.accum_count,
-                        "train/reward": self.accum_reward.item() / self.accum_count,
-                        "train/format_reward": self.accum_format_reward.item() / self.accum_count,
-                        "train/acc_reward": self.accum_acc_reward.item() / self.accum_count,
                         "train/kl": self.accum_kl.item() / self.accum_count,
                         "train/advantages": self.accum_advantages.item() / self.accum_count,
-                        "train/response_length": self.accum_response_length.item() / self.accum_count,
+                        "train/learning_rate": self.lr_scheduler.get_last_lr()[0],
+                        "rollout/temperature": data["temperature"].cpu().numpy()[0][0],
                     }
                 )
             self.accum_loss.zero_()
@@ -232,3 +269,125 @@ class GRPOConsumer(BaseConsumer):
         model = self.policy_model.unwrap()
         state_dict = model.state_dict()
         return state_dict
+
+
+@ray.remote
+class GRPOEvalConsumer(BaseConsumer):
+    def __init__(
+        self,
+        num_producers,
+        num_episodes,
+        rank,
+        world_size,
+        master_addr,
+        master_port,
+        num_update_per_episode,
+        num_recv_per_update,
+        batch_size,
+        model_config,
+        plugin_config,
+        microbatch_size=1,
+        num_generations=4,
+        use_wandb=True,
+        log_dir="./results",
+    ):
+        super().__init__(
+            num_producers,
+            num_episodes,
+            rank,
+            world_size,
+            master_addr,
+            master_port,
+            num_update_per_episode,
+            num_recv_per_update,
+            batch_size,
+            model_config,
+            plugin_config,
+            microbatch_size,
+        )
+        path = model_config.pop("path")
+        self.policy_model = AutoModelForCausalLM.from_pretrained(path, **model_config)
+        self.policy_model.train()
+        self.accum_reward = torch.zeros(1, device=self.device)
+        self.accum_format_reward = torch.zeros(1, device=self.device)
+        self.accum_acc_reward = torch.zeros(1, device=self.device)
+        self.accum_response_length = torch.zeros(1, device=self.device)
+        self.accum_count = torch.zeros(1, device=self.device)
+
+        self.tokenizer = AutoTokenizer.from_pretrained(path)
+        self.pad_token_id = self.tokenizer.pad_token_id
+        self.num_generations = num_generations
+
+        # Initialize verifiable reward.
+        response_format_tags = {
+            "think_start": {"text": "<think>", "num_occur": 1},
+            "think_end": {"text": "</think>", "num_occur": 1},
+            "answer_start": {"text": "<answer>", "num_occur": 1},
+            "answer_end": {"text": "</answer>", "num_occur": 1},
+        }
+        self.reward_model = VerifiableReward(
+            reward_fns=[math_reward_fn], tokenizer=self.tokenizer, tags=response_format_tags
+        )
+
+        self.log_dir = log_dir
+        if not os.path.exists(self.log_dir):
+            os.makedirs(self.log_dir)
+        else:
+            os.system(f"rm -rf {self.log_dir}/*")
+
+    def setup(self):
+        super().setup()
+        self.policy_model, _, *_ = self.booster.boost(self.policy_model)
+
+    def step(self, step_idx: int, **kwargs) -> Optional[float]:
+        rank = dist.get_rank()
+        data = {k: v.view(-1, v.size(-1)).cpu() for k, v in kwargs.items()}
+        kwargs["input_ids"].size(0)
+        reward_group = self.reward_model(
+            data["input_ids"], gt_answer=data["gt_answer"], response_idx=data["response_idx"]
+        )
+        reward = [value[0].item() for value in reward_group]
+        format_reward = [value[1].item() for value in reward_group]
+        acc_reward = [value[2].item() for value in reward_group]
+        response_length = [(data["response_idx"][i][1] - data["response_idx"][i][0]).item() for i in range(len(reward))]
+
+        response = self.tokenizer.batch_decode(data["input_ids"], skip_special_tokens=True)
+        with open(f"{self.log_dir}/eval_results_rank_{rank}.jsonl", "a", encoding="utf8") as f:
+            for i in range(len(response)):
+                f.write(
+                    json.dumps(
+                        {
+                            "response": response[i],
+                            "reward": reward[i],
+                            "format_reward": format_reward[i],
+                            "acc_reward": acc_reward[i],
+                            "response_length": response_length[i],
+                        },
+                        ensure_ascii=False,
+                    )
+                    + "\n"
+                )
+
+        self.accum_reward += sum(reward)
+        self.accum_format_reward += sum(format_reward)
+        self.accum_acc_reward += sum(acc_reward)
+        self.accum_response_length += sum(response_length)
+        self.accum_count += len(reward)
+
+        # print results
+        total_count = all_reduce_mean(self.accum_count, self.plugin)
+        mean_reward = all_reduce_mean(self.accum_reward, self.plugin) / total_count
+        mean_format_reward = all_reduce_mean(self.accum_format_reward, self.plugin) / total_count
+        mean_acc_reward = all_reduce_mean(self.accum_acc_reward, self.plugin) / total_count
+        mean_response_length = all_reduce_mean(self.accum_response_length, self.plugin) / total_count
+        if rank == 0:
+            print(
+                f"Step {step_idx}: Mean Reward: {mean_reward}, Mean Format Reward: {mean_format_reward}, Mean Acc Reward: {mean_acc_reward}, Mean Response Length: {mean_response_length}"
+            )
+        return None
+
+    def state_dict(self):
+        self.policy_model._force_wait_all_gather()
+        model = self.policy_model.unwrap()
+        state_dict = model.state_dict()
+        return state_dict

applications/ColossalChat/coati/distributed/inference_backend.py

@@ -53,7 +53,13 @@ class TransformersInferenceBackend(BaseInferenceBackend):
     )
     FORCE_GENERATE_CONFIG = dict(output_logits=True, return_dict_in_generate=True)
 
-    def __init__(self, model_config: Dict[str, Any], generate_config: Dict[str, Any], tokenizer: PreTrainedTokenizer):
+    def __init__(
+        self,
+        model_config: Dict[str, Any],
+        generate_config: Dict[str, Any],
+        tokenizer: PreTrainedTokenizer,
+        num_generations: int = 8,
+    ):
         model_config = update_by_default(model_config, self.DEFAULT_MODEL_CONFIG)
         model_config.update(self.FORCE_MODEL_CONFIG)
         path = model_config.pop("path")
@@ -61,7 +67,7 @@ class TransformersInferenceBackend(BaseInferenceBackend):
         self.generate_config = generate_config.copy()
         self.generate_config.update(self.FORCE_GENERATE_CONFIG)
         self.tokenizer = tokenizer
-        self.num_generations = 8
+        self.num_generations = num_generations
 
     @torch.no_grad()
     def generate(self, input_ids: torch.Tensor, attention_mask: torch.Tensor, **kwargs) -> Dict[str, torch.Tensor]:
@@ -120,7 +126,13 @@ class TransformersInferenceBackend(BaseInferenceBackend):
 
 
 class SGLangInferenceBackend(BaseInferenceBackend):
-    def __init__(self, model_config: Dict[str, Any], generate_config: Dict[str, Any], tokenizer: PreTrainedTokenizer):
+    def __init__(
+        self,
+        model_config: Dict[str, Any],
+        generate_config: Dict[str, Any],
+        tokenizer: PreTrainedTokenizer,
+        num_generations: int = 8,
+    ):
         if sgl is None:
             raise ImportError("sglang is not installed")
         path = model_config.pop("path")
@@ -175,27 +187,38 @@ class VLLMInferenceBackend(BaseInferenceBackend):
     )
     FORCE_GENERATE_CONFIG = dict(
         logprobs=0,
-        n=8,
     )
 
-    def __init__(self, model_config: Dict[str, Any], generate_config: Dict[str, Any], tokenizer: PreTrainedTokenizer):
+    def __init__(
+        self,
+        model_config: Dict[str, Any],
+        generate_config: Dict[str, Any],
+        tokenizer: PreTrainedTokenizer,
+        num_generations: int = 8,
+    ):
         if LLM is None:
             raise ImportError("vllm is not installed")
         model_config = update_by_default(model_config, self.DEFAULT_MODEL_CONFIG)
         path = model_config.pop("path")
-        self.llm = LLM(path, **model_config)
+        self.llm = LLM(model=path, **model_config)
         generate_config = generate_config.copy()
         generate_config.update(self.FORCE_GENERATE_CONFIG)
+        generate_config.update({"n": num_generations})
         self.generate_config = SamplingParams(**generate_config)
         self.tokenizer = tokenizer
-        self.num_generations = self.FORCE_GENERATE_CONFIG["n"]
+        self.num_generations = num_generations
 
     @torch.no_grad()
     def generate(self, input_ids: torch.Tensor, attention_mask: torch.Tensor, **kwargs) -> Dict[str, torch.Tensor]:
         micro_batch_size = input_ids.size(0)
         response_start_idx = input_ids.size(1)
+        first_non_padding_token_idx = (input_ids != self.tokenizer.pad_token_id).int().argmax(dim=1)
+        micro_batch_input_ids = input_ids.tolist()
+        micro_batch_input_ids_no_padding = [
+            micro_batch_input_ids[i][first_non_padding_token_idx[i] :] for i in range(micro_batch_size)
+        ]
         outputs = self.llm.generate(
-            prompt_token_ids=input_ids.tolist(), sampling_params=self.generate_config, use_tqdm=False
+            prompt_token_ids=micro_batch_input_ids_no_padding, sampling_params=self.generate_config, use_tqdm=False
         )
         out_tokens = []
         out_len = []

applications/ColossalChat/coati/distributed/launch.py

@@ -1,15 +1,13 @@
+import copy
 from typing import Any, Dict, Optional
 
 import ray
 
 from .consumer import SimpleConsumer
-from .grpo_consumer import GRPOConsumer
+from .grpo_consumer import GRPOConsumer, GRPOEvalConsumer
 from .producer import SimpleProducer
 
-ALGO_MAP = {
+ALGO_MAP = {"Simple": SimpleConsumer, "GRPO": GRPOConsumer, "EvalGRPO": GRPOEvalConsumer}
-    "Simple": SimpleConsumer,
-    "GRPO": GRPOConsumer,
-}
 
 
 def get_jsonl_size_fast(path: str) -> int:
@@ -44,6 +42,7 @@ def launch_distributed(
     plugin_config: Dict[str, Any],
     tokenizer_config: Optional[Dict[str, Any]] = None,
     inference_backend: str = "transformers",
+    num_generations: int = 8,
     master_addr: str = "localhost",
     master_port: int = 29500,
     core_algo: str = "GRPO",
@@ -78,8 +77,15 @@ def launch_distributed(
             tokenizer_config=tokenizer_config,
             microbatch_size=inference_microbatch_size,
             backend=inference_backend,
+            num_generations=num_generations,
         )
         procs.append(producer)
+    generate_config_consumer = copy.deepcopy(generate_config)
+    generate_config_consumer.update(
+        dict(
+            backend=inference_backend,
+        )
+    )
     for i in range(num_consumer_procs):
         consumer = core_consumer.options(num_gpus=1).remote(
             num_producers=num_producers,
@@ -94,6 +100,9 @@ def launch_distributed(
             model_config=train_model_config,
             plugin_config=plugin_config,
             microbatch_size=train_microbatch_size,
+            generate_config=generate_config_consumer,
+            training_config={"filter_range": [0.05, 9.0], "lr": 1e-6},
+            num_generations=num_generations,
         )
         procs.append(consumer)
     ray.get([p.setup.remote() for p in procs])

applications/ColossalChat/coati/distributed/loss.py

@@ -23,6 +23,7 @@ class PolicyLoss(nn.Module):
         advantages: torch.Tensor,
         per_token_kl: torch.Tensor,
         action_mask: Optional[torch.Tensor] = None,
+        loss_mask: Optional[torch.Tensor] = None,
     ) -> torch.Tensor:
         skip = False
         if action_mask is None:
@@ -38,5 +39,7 @@ class PolicyLoss(nn.Module):
             loss = masked_mean(loss, action_mask)
         else:
             loss = loss.mean(dim=1)
+        if loss_mask is not None:
+            loss = loss * loss_mask
         loss = loss.mean()
         return loss, skip, ratio.max()

applications/ColossalChat/coati/distributed/producer.py

@@ -101,6 +101,9 @@ class BaseProducer:
                     break
                 outputs = self.rollout(**batch)
                 print(f"[P{self.producer_idx}] Send data {[(k, v.shape) for k, v in outputs.items()]}")
+                outputs["temperature"] = torch.tensor(
+                    [self.model.generate_config.temperature] * outputs["input_ids"].size(0)
+                ).to(outputs["input_ids"].device)
                 outputs = pre_send(outputs)
                 ray_broadcast_tensor_dict(
                     outputs, src=0, device=self.device, group_name=f"sync_data_{self.producer_idx}"
@@ -117,6 +120,12 @@ class BaseProducer:
                         None, self.num_producers, device=self.device, group_name="sync_model"
                     )
                     self.load_state_dict(state_dict)
+                # linear annealing for 1 episode, temperature from initial to 0.7
+                if episode <= 0:
+                    ratio = 1 - (len(self.dataloader) - i) / len(self.dataloader)
+                    self.model.generate_config.temperature = (
+                        ratio * self.generate_config["temperature"] + (1 - ratio) * 0.7
+                    )
 
 
 @ray.remote
@@ -135,6 +144,7 @@ class SimpleProducer(BaseProducer):
         tokenizer_config=None,
         microbatch_size=1,
         backend="transformers",
+        num_generations: int = 8,
     ):
         super().__init__(
             producer_idx,
@@ -150,7 +160,7 @@ class SimpleProducer(BaseProducer):
             microbatch_size,
             backend,
         )
-        self.model = self.backend_cls(model_config, generate_config, self.tokenizer)
+        self.model = self.backend_cls(model_config, generate_config, self.tokenizer, num_generations)
 
     @torch.no_grad()
     def rollout(self, input_ids, attention_mask, **kwargs):

applications/ColossalChat/rl_example.py

@@ -15,18 +15,14 @@ if __name__ == "__main__":
     parser.add_argument("-tbs", "--train-batch-size", type=int, default=32)
     parser.add_argument("-tmbs", "--train-microbatch-size", type=int, default=1)
     parser.add_argument("-b", "--backend", type=str, default="transformers")
-    parser.add_argument("-a", "--algo", type=str, default="GRPO", choices=["Simple, GRPO"])
+    parser.add_argument("-a", "--algo", type=str, default="GRPO", choices=["Simple", "GRPO", "EvalGRPO"])
     args = parser.parse_args()
 
     ray.init(address="local", namespace="ray-example")
 
     inference_model_config = dict(path=args.model)
     train_model_config = dict(path=args.model)
-    generate_config = dict(
+    generate_config = dict(top_k=50, top_p=0.75, temperature=0.9)
-        top_k=50,
-        top_p=0.9,
-        temperature=1.0,
-    )
 
     if args.backend == "transformers":
         inference_model_config.update(
@@ -52,19 +48,13 @@ if __name__ == "__main__":
             )
         )
     elif args.backend == "vllm":
-        inference_model_config.update(
+        inference_model_config.update(dict(gpu_memory_utilization=0.7, enforce_eager=True, enable_chunked_prefill=True))
-            dict(
-                gpu_memory_utilization=0.7,
-            )
-        )
         generate_config.update(
             dict(
                 max_tokens=2048,
                 ignore_eos=True,
                 include_stop_str_in_output=True,
                 stop=["</answer>"],
-                temperature=0.7,
-                top_p=0.95,
             )
         )
     else:
@@ -97,6 +87,6 @@ if __name__ == "__main__":
         plugin_config={},
         inference_backend=args.backend,
         master_addr="localhost",
-        master_port=29504,
+        master_port=29503,
         core_algo=args.algo,
     )

colossalai/shardformer/layer/loss.py

@@ -387,7 +387,7 @@ def dist_log_prob(
             dtype=dtype,
         )
     else:
-        log_prob = log_softmax(logits)
+        log_prob = log_softmax(logits, dim=-1)
         log_prob = log_prob.gather(dim=-1, index=labels.unsqueeze(-1))
 
     return log_prob