support resume training, still buggy

applications/ColossalChat/coati/distributed/consumer.py

@@ -55,6 +55,7 @@ class BaseConsumer:
         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
@@ -143,6 +144,26 @@ class BaseConsumer:
         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}"
         )
@@ -286,7 +307,7 @@ class BaseConsumer:
                         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)
+                        self.booster.save_model(self.policy_model, save_path, shard=True, use_safetensors=True)
                         if self.rank == 0:
                             print(f"Saved model checkpoint at step {step + 1} in folder {save_path}")
 
@@ -365,7 +386,7 @@ class SimpleConsumer(BaseConsumer):
         self.model = AutoModelForCausalLM.from_pretrained(path, **model_config)
         self.model.train()
         self.model.gradient_checkpointing_enable()
-        self.optimizer = HybridAdam(self.model.parameters(), lr=1e-3)
+        self.optimizer = HybridAdam(self.model.parameters(), lr=1e-3, weight_decay=0.01)
         self.accum_loss = torch.zeros(1, device=self.device)
 
     def setup(self):

applications/ColossalChat/coati/distributed/grpo_consumer.py

@@ -72,7 +72,11 @@ 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=grpo_config.get("lr", 1e-6))
+        self.optimizer = HybridAdam(
+            self.policy_model.parameters(),
+            lr=grpo_config.get("lr", 1e-6),
+            weight_decay=grpo_config.get("weight_decay", 0.01),
+        )
         self.accum_loss = torch.zeros(1, device=self.device)
         self.accum_kl = torch.zeros(1, device=self.device)
         self.accum_entropy = torch.zeros(1, device=self.device)
@@ -153,6 +157,8 @@ class GRPOConsumer(BaseConsumer):
         )
         if self.policy_loss_fn.beta > 0:
             self.reference_model, *_ = self.booster.boost(self.reference_model)
+        if self.checkpoint_path is not None:
+            self.booster.load_model(self.policy_model, self.checkpoint_path)
         self.plugin.logger.set_level("ERROR")
 
     def step(self, step_idx: int, pbar: Any, **kwargs) -> Optional[float]:

applications/ColossalChat/coati/distributed/launch_zero_bubble.py

@@ -55,7 +55,6 @@ def launch_distributed(
     eval_dataset_config: Optional[Dict[str, Any]] = None,
     eval_interval: int = 100,
     eval_save_dir: Optional[str] = None,
-    eval_generation_config: Optional[Dict[str, Any]] = None,
     log_rollout_interval: int = 20,
     rollout_save_dir: str = "./rollout",
     enable_profiling: bool = False,
@@ -139,7 +138,6 @@ def launch_distributed(
             eval_interval=eval_interval,
             grpo_config=grpo_config,
             eval_save_dir=eval_save_dir,
-            eval_generation_config=eval_generation_config,
             project_name=project_name,
             run_name=run_name,
             wandb_group_name=wandb_group_name,

applications/ColossalChat/coati/distributed/producer.py

@@ -203,6 +203,28 @@ class BaseProducer:
         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")
+        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/distributed/reward/reward_fn.py

@@ -25,7 +25,12 @@ from latex2sympy2_extended import NormalizationConfig
 from math_verify import ExprExtractionConfig, LatexExtractionConfig, parse, verify
 
 from .code_reward.utils import check_correctness_code_api as check_correctness_code
-from .reward_utils import extract_boxed_solution, extract_solution, validate_response_structure
+from .reward_utils import (
+    extract_boxed_solution,
+    extract_solution,
+    find_infinite_loop_start,
+    validate_response_structure,
+)
 
 CANNOT_PARSE_GT_ANSWER = -1
 CANNOT_PARSE_PREDICTION = -2
@@ -122,6 +127,8 @@ def math_reward_fn(input_ids, gt_answer, response_idx, **kwargs):
 
     decoded_final_answer = tokenizer.decode(input_ids[s : e + 1], skip_special_tokens=True)
 
+    repetition_reward = 1.0 if detect_repetition(decoded_final_answer) == [] else 0.0
+
     final_answer, processed_str = extract_solution(decoded_final_answer)
 
     format_valid = validate_response_structure(processed_str, kwargs["tags"])
@@ -137,6 +144,10 @@ def math_reward_fn(input_ids, gt_answer, response_idx, **kwargs):
     if format_valid:
         format_acc += 1
 
+    # Add repetition reward
+    if not eval_mode:
+        reward += repetition_reward
+
     # Check if the sequence is over length
     if not eval_mode and res_length >= max_new_tokens:
         reward *= 0.0
@@ -182,6 +193,8 @@ def boxed_math_reward_fn(input_ids, gt_answer, response_idx, **kwargs):
         raise ValueError("no gt_answer is provided, please check your training dataset.")
 
     decoded_final_answer = tokenizer.decode(input_ids[s : e + 1], skip_special_tokens=True)
+    print(f"Decoded final answer: {decoded_final_answer[-500:]}")
+    repetition_score = find_infinite_loop_start(input_ids[s : e + 1], min_repeats=2, distance=False)
 
     final_answer = extract_boxed_solution(decoded_final_answer)
     format_valid = final_answer is not None
@@ -202,6 +215,10 @@ def boxed_math_reward_fn(input_ids, gt_answer, response_idx, **kwargs):
     if format_valid:
         format_acc += 1
 
+    if not repetition_score > 0 and not eval_mode:
+        # award for non-repetition
+        reward += 2
+
     # Check if the sequence is over length
     if not eval_mode and res_length >= max_new_tokens:
         reward *= 0.0

applications/ColossalChat/coati/distributed/reward/reward_utils.py

@@ -14,7 +14,9 @@
 # limitations under the License.
 
 import re
-from typing import Dict, Optional, Tuple
+from typing import Dict, List, Optional, Tuple
+
+import torch
 
 
 def validate_response_structure(processed_str: str, tags: Dict = None) -> bool:
@@ -122,3 +124,51 @@ def extract_boxed_solution(text: str) -> Optional[str]:
     except Exception:
         # Any other unexpected error
         return None
+
+
+import Levenshtein
+
+
+def is_similar(seq1: List[int], seq2: List[int], threshold: float = 0.9) -> bool:
+    ratio = Levenshtein.ratio(seq1, seq2)
+    return ratio >= threshold
+
+
+def find_infinite_loop_start(token_ids: List[int], min_repeats: int = 2, distance: bool = False) -> float:
+    n = len(token_ids)
+
+    # Step 1: Detect the repeating segment at the end using two pointers
+    longest_valid_length = 0
+    start_of_loop = n
+
+    for length in range(1, n // min_repeats + 1):  # Try different phrase lengths
+        count = 1  # Reset repetition counter
+        right = n - length  # Start comparing from the second last occurrence
+
+        while right - length >= 0:
+            # Check if the current phrase matches the previous phrase
+            if distance:
+                if is_similar(token_ids[right - length : right], token_ids[right : right + length]):
+                    count += 1
+                else:
+                    break  # Stop if repetition is broken
+            else:
+                # Use torch.equal() for tensor comparison
+                if torch.equal(token_ids[right - length : right], token_ids[right : right + length]):
+                    count += 1
+                else:
+                    break  # Stop if repetition is broken
+
+            right -= length  # Move left to check further
+
+        if count >= min_repeats:  # Found a valid repeating phrase
+            longest_valid_length = length
+            start_of_loop = right  # This is where the first cycle of the repetition begins
+
+    if longest_valid_length == 0:
+        return 0.0  # No infinite loop found, return repetition ratio as 0
+
+    # Step 2: Compute the repetition ratio
+    repetition_ratio = (n - start_of_loop) / n
+
+    return repetition_ratio

applications/ColossalChat/coati/distributed/zero_bubble/consumer.py

@@ -55,7 +55,7 @@ class BaseConsumer:
         self.num_microbatches = batch_size // minibatch_size
         self.data_uid = 0
         self.sync_model_thread_started = False
-
+        self.checkpoint_path = model_config.pop("checkpoint_path", None)
         self.model_config = model_config
         self.plugin_config = plugin_config
 
@@ -79,6 +79,7 @@ class BaseConsumer:
         plugin_config.update(self.plugin_config)
         self.plugin = HybridParallelPlugin(**plugin_config)
         self.booster = Booster(plugin=self.plugin)
+
         self.dp_rank = dist.get_rank(self.plugin.dp_group)
         self.tp_rank = dist.get_rank(self.plugin.tp_group)
         self.pp_rank = dist.get_rank(self.plugin.pp_group)
@@ -165,8 +166,93 @@ class BaseConsumer:
                 desc=f"Episode {episode} with rollout step(s)",
                 disable=self.rank != 0,
             ) as pbar:
+                need_sync_model = True
                 while self.received_prompts < self.train_dataset_size:
                     torch.cuda.reset_peak_memory_stats()
+                    if need_sync_model and (
+                        (self.global_step + 1) % self.save_interval == 0
+                        or self.received_prompts >= self.train_dataset_size
+                    ):
+                        if self.rank == 0:
+                            print(f"Start saving policy model at step {self.global_step + 1}.")
+                        save_path = os.path.join(
+                            self.save_dir, f"modeling-episode-{episode}-step-{self.global_step + 1}"
+                        )
+                        self.booster.save_model(self.policy_model, save_path, shard=True, use_safetensors=True)
+                        if self.rank == 0:
+                            print(f"Saved model checkpoint at step {self.global_step + 1} in folder {save_path}")
+
+                    if need_sync_model and (
+                        episode != self.num_episodes - 1 or self.received_prompts != self.train_dataset_size
+                    ):
+
+                        def sync_model_thread():
+                            # sync model weights to all producers, if no model update or it is the last training step, skip syncing
+                            if self.pp_size > 1:
+                                print(
+                                    f"[T{dist.get_rank()}] Sync model PP stage {self.pp_rank} episode {episode} step {self.global_step}"
+                                )
+                            else:
+                                print(f"[T{dist.get_rank()}] Sync model episode {episode} step {self.global_step}")
+                            torch.cuda.empty_cache()
+                            if self.pp_size > 1:
+                                if self.tp_rank == 0 and self.dp_rank == 0:
+                                    self.profiler.enter("sync_model")
+                                    ray.get(
+                                        self.shared_signal_actor.set_signal.remote(
+                                            f"consumer_pp_{self.pp_rank}", "ready_sync_model"
+                                        )
+                                    )
+                                    print(
+                                        f"[T{dist.get_rank()}] Sync model PP stage {self.pp_rank} episode {episode} step {self.global_step}"
+                                    )
+                                    ray_broadcast_tensor_dict(
+                                        self.state_dict_cpu,
+                                        src=0,
+                                        device=torch.device("cpu"),
+                                        group_name=f"sync_model_consumer_pp_{self.pp_rank}",
+                                        backend="gloo",
+                                    )
+                                    self.profiler.exit("sync_model")
+                            else:
+                                if self.rank == 0:
+                                    self.profiler.enter("sync_model")
+                                    ray.get(self.shared_signal_actor.set_signal.remote("consumer", "ready_sync_model"))
+                                    print(f"[T{dist.get_rank()}] Sync model episode {episode} step {self.global_step}")
+                                    ray_broadcast_tensor_dict(
+                                        self.state_dict_cpu,
+                                        src=0,
+                                        device=torch.device("cpu"),
+                                        group_name="sync_model_consumer",
+                                        backend="gloo",
+                                    )
+                                    self.profiler.exit("sync_model")
+
+                        if not self.sync_model_thread_started:
+                            # only sync model when the thread is not started and no other thread is broadcasting
+                            self.sync_model_thread_started = True
+                            state_dict_ = self.state_dict()
+                            if (self.pp_size > 1 and self.tp_rank == 0 and self.dp_rank == 0) or (
+                                self.pp_size == 1 and self.rank == 0
+                            ):
+                                if len(self.state_dict_cpu) == 0:
+                                    # use pinned memory to speed up the transfer
+                                    self.state_dict_cpu = {k: v.cpu().pin_memory() for k, v in state_dict_.items()}
+                                    torch.cuda.synchronize()
+                                for k, v in state_dict_.items():
+                                    self.state_dict_cpu[k].copy_(v, non_blocking=True)
+                                torch.cuda.synchronize()
+                            cc.barrier(
+                                group_name="consumer_pg"
+                            )  # to make sure all ranks have state dict offloaded to CPU before starting the thread
+                            time_before_starting_thread = time.time()
+                            threading.Thread(target=sync_model_thread).start()
+                            # sync_model_thread()
+                            self.profiler.log(
+                                f"Sync model, took {time.time() - time_before_starting_thread:.2f} seconds"
+                            )
+                            self.sync_model_thread_started = False
+                            # ray.get(self.shared_signal_actor.release_process_lock.remote("broadcasting_lock"))
                     effective_group_to_raw_group_mapping = {}
                     self.profiler.enter(f"recv_data")
                     while len(effective_group_to_raw_group_mapping) < self.dp_size * self.minibatch_size:
@@ -254,90 +340,6 @@ class BaseConsumer:
                             pbar.set_postfix({"loss": loss})
                             need_sync_model = True
                             ray.get(self.shared_signal_actor.set_signal.remote("global_step", self.global_step + 1))
-                    if need_sync_model and (
-                        (self.global_step + 1) % self.save_interval == 0
-                        or self.received_prompts >= self.train_dataset_size
-                    ):
-                        if self.rank == 0:
-                            print(f"Start saving policy model at step {self.global_step + 1}.")
-                        save_path = os.path.join(
-                            self.save_dir, f"modeling-episode-{episode}-step-{self.global_step + 1}"
-                        )
-                        self.booster.save_model(self.policy_model, save_path, shard=True)
-                        if self.rank == 0:
-                            print(f"Saved model checkpoint at step {self.global_step + 1} in folder {save_path}")
-
-                    if need_sync_model and (
-                        episode != self.num_episodes - 1 or self.received_prompts != self.train_dataset_size
-                    ):
-
-                        def sync_model_thread():
-                            # sync model weights to all producers, if no model update or it is the last training step, skip syncing
-                            if self.pp_size > 1:
-                                print(
-                                    f"[T{dist.get_rank()}] Sync model PP stage {self.pp_rank} episode {episode} step {self.global_step}"
-                                )
-                            else:
-                                print(f"[T{dist.get_rank()}] Sync model episode {episode} step {self.global_step}")
-                            torch.cuda.empty_cache()
-                            if self.pp_size > 1:
-                                if self.tp_rank == 0 and self.dp_rank == 0:
-                                    self.profiler.enter("sync_model")
-                                    ray.get(
-                                        self.shared_signal_actor.set_signal.remote(
-                                            f"consumer_pp_{self.pp_rank}", "ready_sync_model"
-                                        )
-                                    )
-                                    print(
-                                        f"[T{dist.get_rank()}] Sync model PP stage {self.pp_rank} episode {episode} step {self.global_step}"
-                                    )
-                                    ray_broadcast_tensor_dict(
-                                        self.state_dict_cpu,
-                                        src=0,
-                                        device=torch.device("cpu"),
-                                        group_name=f"sync_model_consumer_pp_{self.pp_rank}",
-                                        backend="gloo",
-                                    )
-                                    self.profiler.exit("sync_model")
-                            else:
-                                if self.rank == 0:
-                                    self.profiler.enter("sync_model")
-                                    ray.get(self.shared_signal_actor.set_signal.remote("consumer", "ready_sync_model"))
-                                    print(f"[T{dist.get_rank()}] Sync model episode {episode} step {self.global_step}")
-                                    ray_broadcast_tensor_dict(
-                                        self.state_dict_cpu,
-                                        src=0,
-                                        device=torch.device("cpu"),
-                                        group_name="sync_model_consumer",
-                                        backend="gloo",
-                                    )
-                                    self.profiler.exit("sync_model")
-
-                        if not self.sync_model_thread_started:
-                            # only sync model when the thread is not started and no other thread is broadcasting
-                            self.sync_model_thread_started = True
-                            state_dict_ = self.state_dict()
-                            if (self.pp_size > 1 and self.tp_rank == 0 and self.dp_rank == 0) or (
-                                self.pp_size == 1 and self.rank == 0
-                            ):
-                                if len(self.state_dict_cpu) == 0:
-                                    # use pinned memory to speed up the transfer
-                                    self.state_dict_cpu = {k: v.cpu().pin_memory() for k, v in state_dict_.items()}
-                                    torch.cuda.synchronize()
-                                for k, v in state_dict_.items():
-                                    self.state_dict_cpu[k].copy_(v, non_blocking=True)
-                                torch.cuda.synchronize()
-                            cc.barrier(
-                                group_name="consumer_pg"
-                            )  # to make sure all ranks have state dict offloaded to CPU before starting the thread
-                            time_before_starting_thread = time.time()
-                            threading.Thread(target=sync_model_thread).start()
-                            # sync_model_thread()
-                            self.profiler.log(
-                                f"Sync model, took {time.time() - time_before_starting_thread:.2f} seconds"
-                            )
-                            self.sync_model_thread_started = False
-                            # ray.get(self.shared_signal_actor.release_process_lock.remote("broadcasting_lock"))
                     self.profiler.log(f"Peak memory usage: {torch.cuda.max_memory_allocated() / 1024 ** 2:.2f} MB")
                 self.received_prompts = 0
         ray.get(self.shared_signal_actor.set_signal.remote("consumer", "terminate"))

applications/ColossalChat/coati/distributed/zero_bubble/distributor.py

@@ -20,7 +20,7 @@ class Distributor:
         enable_profiling: bool = True,
     ):
         self.distributor_id = distributor_id
-        self.weight_version = [0] * consumer_pp_size
+        self.weight_version = [-1] * consumer_pp_size
         self.consumer_pp_size = consumer_pp_size
         self.state_dict_cpu = {}
         self.num_producers = num_producers

applications/ColossalChat/coati/distributed/zero_bubble/grpo_consumer.py

@@ -75,7 +75,11 @@ class GRPOConsumer(BaseConsumer):
         self.policy_model.train()
         self.policy_model.gradient_checkpointing_enable()
         self.vocab_size = self.policy_model.config.vocab_size
-        self.optimizer = HybridAdam(self.policy_model.parameters(), lr=grpo_config.get("lr", 1e-6))
+        self.optimizer = HybridAdam(
+            self.policy_model.parameters(),
+            lr=grpo_config.get("lr", 1e-6),
+            weight_decay=grpo_config.get("weight_decay", 0.01),
+        )
         self.accum_loss = torch.zeros(1, device=self.device)
         self.accum_kl = torch.zeros(1, device=self.device)
         self.accum_entropy = torch.zeros(1, device=self.device)
@@ -157,6 +161,9 @@ class GRPOConsumer(BaseConsumer):
         self.policy_model, self.optimizer, _, _, self.lr_scheduler = self.booster.boost(
             self.policy_model, self.optimizer, lr_scheduler=self.lr_scheduler
         )
+        if self.checkpoint_path is not None:
+            print(f"Resume training from checkpoint: {self.checkpoint_path}")
+            self.booster.load_model(self.policy_model, self.checkpoint_path)
         if self.policy_loss_fn.beta > 0:
             self.reference_model, *_ = self.booster.boost(self.reference_model)
         self.plugin.logger.set_level("ERROR")

applications/ColossalChat/coati/distributed/zero_bubble/producer.py

@@ -82,7 +82,7 @@ class BaseProducer:
         self.eval_interval = eval_interval
         self.eval_save_dir = eval_save_dir
         self.consumer_global_step = 0
-        self.producer_weight_version = 0
+        self.producer_weight_version = -1
         self.eval_mode = False
         self.log_rollout_interval = log_rollout_interval
         self.latest_rollout_log_step = -1
@@ -148,19 +148,38 @@ class BaseProducer:
             self.evaluation_function = code_reward_fn
         else:
             raise ValueError(f"Unknown evaluation function type {grpo_config['reward_fn_type']}")
-
         self.eval_dataset_config = eval_dataset_config
         if self.eval_dataset_config is not None:
             self.eval_dataloaders = {}
+            self.eval_sample_params = {}
             for eval_task_name in self.eval_dataset_config:
+                eval_sampling_params = copy.deepcopy(generate_config)
+                eval_sampling_params["n"] = eval_dataset_config[eval_task_name].pop(
+                    "n", 1
+                )  # use 1 generation for evaluation
+                if "max_tokens" in eval_sampling_params:
+                    eval_sampling_params["max_tokens"] = eval_dataset_config[eval_task_name].pop(
+                        "max_tokens", eval_sampling_params["max_tokens"]
+                    )
+                elif "max_new_tokens" in eval_sampling_params:
+                    eval_sampling_params["max_new_tokens"] = eval_dataset_config[eval_task_name].pop(
+                        "max_new_tokens", eval_sampling_params["max_new_tokens"]
+                    )
+                eval_sampling_params["temperature"] = eval_dataset_config[eval_task_name].pop("temperature", 0.6)
+                eval_sampling_params["logprobs"] = 0  # force parameter
+                self.eval_sample_params[eval_task_name] = SamplingParams(**eval_sampling_params)
+
                 eval_dataset_path = eval_dataset_config[eval_task_name].pop("path")
+                eval_inference_batch_size = eval_dataset_config[eval_task_name].pop("batch_size", microbatch_size)
                 eval_dataset = RawConversationDataset(
                     self.tokenizer, eval_dataset_path, **eval_dataset_config[eval_task_name]
                 )
-                print(f"[P{self.producer_idx}] eval dataset {eval_task_name} size: {len(eval_dataset)}")
+                print(
+                    f"[P{self.producer_idx}] eval dataset {eval_task_name} size: {len(eval_dataset)}, eval_batch_size: {eval_inference_batch_size} num_producers:{num_producers}"
+                )
                 self.eval_dataloaders[eval_task_name] = DataLoader(
                     eval_dataset,
-                    batch_size=microbatch_size,
+                    batch_size=eval_inference_batch_size,
                     sampler=DistributedSampler(
                         eval_dataset,
                         num_replicas=num_producers,
@@ -280,7 +299,17 @@ class BaseProducer:
                         self.sync_model_thread_started = False
 
                     distributor_weight_version = ray.get(self.shared_signal_actor.get_signal.remote()).get(
-                        f"distributor_weight_version", 0
+                        f"distributor_weight_version", -1
+                    )
+                    print(f"[P{self.producer_idx}] Distributor weight version: {distributor_weight_version}")
+                    is_initial_sync = False
+                    while distributor_weight_version < 0:
+                        # wait for distributor to set weight version
+                        is_initial_sync = True
+                        print(f"[P{self.producer_idx}] Waiting for intial model sync before training")
+                        time.sleep(1)
+                        distributor_weight_version = ray.get(self.shared_signal_actor.get_signal.remote()).get(
+                            f"distributor_weight_version", -1
                         )
                     if (
                         not self.sync_model_thread_started
@@ -291,6 +320,16 @@ class BaseProducer:
                         self.sync_model_thread = threading.Thread(target=sync_model_thread)
                         self.producer_weight_version = distributor_weight_version
                         self.sync_model_thread.start()
+                    if is_initial_sync:
+                        # wait till initial model sync is done and load the model state dict
+                        while self.sync_model_thread_started:
+                            print(f"[P{self.producer_idx}] Waiting for model sync to finish before evaluation")
+                            time.sleep(1)
+                        for pp_idx in range(self.consumer_pp_size):
+                            if self.state_dict_cpu[pp_idx] is not None and self.state_dict_cpu[pp_idx] != {}:
+                                self.load_state_dict(self.state_dict_cpu[pp_idx])
+                        print(f"[P{self.producer_idx}] loaded initial model state dict")
+
                     torch.cuda.empty_cache()
                     if isinstance(self.model, BACKEND_MAP["vllm"]) and self.model.model_config.get(
                         "enable_sleep_mode", False
@@ -314,7 +353,9 @@ class BaseProducer:
                             for eval_batch in tqdm.tqdm(
                                 self.eval_dataloaders[eval_task_name], disable=self.producer_idx != 0
                             ):
-                                eval_outputs = self.rollout(**eval_batch, sample_params=self.eval_sample_params)
+                                eval_outputs = self.rollout(
+                                    **eval_batch, sample_params=copy.deepcopy(self.eval_sample_params[eval_task_name])
+                                )
                                 eval_results = eval_results + [
                                     self.evaluation_function(
                                         eval_outputs["input_ids"][m][n],
@@ -343,6 +384,7 @@ class BaseProducer:
                                 print(
                                     f"[P{self.producer_idx}]: Accuracy on {eval_task_name}: {to_log_msg[f'eval/{eval_task_name}']}"
                                 )
+
                             # save eval results
                             safe_append_to_jsonl_file(
                                 os.path.join(
@@ -351,7 +393,7 @@ class BaseProducer:
                                 ),
                                 eval_results,
                             )
-
+                            raise ValueError()
                         if self.producer_idx == 0:
                             self.wandb_run.log(to_log_msg, step=self.consumer_global_step)
                         self.eval_mode = False
@@ -427,15 +469,15 @@ class BaseProducer:
                                 self.load_state_dict(self.state_dict_cpu[pp_idx])
 
                 # linear annealing for 1 episode, temperature from initial to 0.9
-                if episode <= 0:
+                # if episode <= 0:
-                    ratio = 1 - (len(self.train_dataloader) - i) / len(self.train_dataloader)
+                #     ratio = 1 - (len(self.train_dataloader) - i) / len(self.train_dataloader)
-                    self.model.generate_config["temperature"] = (1 - ratio) * self.generate_config[
+                #     self.model.generate_config["temperature"] = (1 - ratio) * self.generate_config[
-                        "temperature"
+                #         "temperature"
-                    ] + ratio * 0.9
+                #     ] + ratio * 0.9
-                    if isinstance(self.model, BACKEND_MAP["vllm"]):
+                #     if isinstance(self.model, BACKEND_MAP["vllm"]):
-                        self.model.sample_params.temperature = (1 - ratio) * self.generate_config[
+                #         self.model.sample_params.temperature = (1 - ratio) * self.generate_config[
-                            "temperature"
+                #             "temperature"
-                        ] + ratio * 0.9
+                #         ] + ratio * 0.9
 
     def __del__(self):
         self.profiler.close()
@@ -464,7 +506,6 @@ class SimpleProducer(BaseProducer):
         eval_interval=-1,  # disable evaluation
         grpo_config: Dict[str, Any] = None,
         eval_save_dir: str = "./eval",
-        eval_generation_config={},
         project_name: str = None,
         run_name: str = None,
         wandb_group_name: str = None,
@@ -500,10 +541,6 @@ class SimpleProducer(BaseProducer):
         )
         print("tokenizer_config", tokenizer_config)
         self.model = self.backend_cls(model_config, generate_config, self.tokenizer, num_generations, tokenizer_config)
-        self.eval_generation_config = copy.deepcopy(self.model.generate_config)
-        self.eval_generation_config["n"] = 1  # use 1 generation for evaluation
-        self.eval_generation_config.update(eval_generation_config)
-        self.eval_sample_params = SamplingParams(**self.eval_generation_config)
 
     @torch.no_grad()
     def rollout(self, input_ids, attention_mask, **kwargs):

applications/ColossalChat/profiling.sh

@@ -11,3 +11,9 @@ rm -rf *.prof
 MAX_NEW_TOKENS=$((4096-512))
 python rl_example.py --dataset /mnt/nfs/yeanbang/experiments/RLHF/grpo/train-alignment-samll.jsonl --model /home/share/data/model/Qwen2.5-Math-7B/ -t 4 -i 4 -b vllm -a GRPO -ibs 8 -tbs 8 -e 1 -rt boxed -si 100 -s "Please reason step by step, and put your final answer within \\boxed{}." -tMbs 4 -tmbs 4 -p GRPO-Math-Profile -ei -5 -zero 2 -mnt $MAX_NEW_TOKENS -nb 1 --enable_profiling 2>&1| tee ibs_32_tbs_32_tmbs_4_zero_2_4096_GRPO_profile.txt
 python visualization.py --visualization actor_timelines_ibs_32_tbs_32_tmbs_4_zero_2_4096_GRPO_profile.png
+
+
+python rl_example_zero_bubble.py --dataset /home/litong/workspace/data/rl-math-train-final.jsonl --model /home/litong/workspace/model/DeepSeek-R1-Distill-Qwen-7B -t 4 -i 4 -b vllm -a DAPO -imbs 4 -ibs 8 -tbs 32 -e 2 -rt boxed -si 25 -s "Please reason step by step, and put your final answer within \\boxed{}." -tMbs 2 -tmbs 2 -p DAPO-Experiment-Baseline -ei 15 -ed '{"Math_500_level_1": "/mnt/nfs/yeanbang/data/RLHF_data/Math-500/Math_500_level_1.jsonl", "Math_500_level_2": "/mnt/nfs/yeanbang/data/RLHF_data/Math-500/Math_500_level_2.jsonl", "Math_500_level_3": "/mnt/nfs/yeanbang/data/RLHF_data/Math-500/Math_500_level_3.jsonl", "Math_500_level_4": "/mnt/nfs/yeanbang/data/RLHF_data/Math-500/Math_500_level_4.jsonl", "Math_500_level_5": "/mnt/nfs/yeanbang/data/RLHF_data/Math-500/Math_500_level_5.jsonl"}' -zero 1 -tp 2 -pp 2 -mpt 512 -mnt 15872 --data_actor_buffer_size_limit 48
+
+
+python rl_example_zero_bubble.py --dataset /mnt/nfs/yeanbang/data/RLHF_data/miromina/miromina.jsonl --model /home/litong/workspace/model/DeepSeek-R1-Distill-Qwen-7B -t 4 -i 4 -b vllm -a DAPO -imbs 4 -ibs 8 -tbs 32 -e 2 -g 16 -rt boxed -si 25 -s "Please reason step by step, and put your final answer within \\boxed{}." -tMbs 2 -tmbs 2 -p DAPO-Experiment-Baseline-V3 -ei 15 -ed '{"AIME2024": "/mnt/nfs/yeanbang/data/RLHF_data/AIME/2024/AIME2024.jsonl"}' -zero 1 -tp 2 -pp 2 -mpt 1024 -mnt 16384 --data_actor_buffer_size_limit 64 --enable_profiling 2>&1| tee ibs_64_tbs_32_tmbs_2_pp_2_pp_2_16K_DAPO_profile_v3.txt

applications/ColossalChat/reproduce.sh

@@ -0,0 +1,9 @@
+export https_proxy=http://vpn.luchentech.com:32171
+export http_proxy=http://vpn.luchentech.com:32171
+export all_proxy=socks5://vpn.luchentech.com:32170
+
+# 输出正常
+# python rl_example.py --dataset /mnt/nfs/yeanbang/data/RLHF_data/miromina/miromina.jsonl --model /mnt/nfs/share/data/models/Miromind-M1-SFT-7B/ --checkpoint-path /mnt/nfs/share/data/models/Miromind-M1-SFT-7B/ -t 4 -i 4 -b vllm -a DAPO -imbs 8 -ibs 8 -tbs 32 -e 2 -g 16 -rt boxed -si 25 -s "Please reason step by step, and put your final answer within \\boxed{}." -tMbs 8 -tmbs 8 -p DAPO-Experiment-Baseline-V3_1 -zero 1 -pp 2 -tp 2 -mpt 1024 -mnt 4096 -nb 1 --enable_profiling
+
+# 输出乱码
+python rl_example.py --dataset /mnt/nfs/yeanbang/data/RLHF_data/miromina/miromina.jsonl --model /mnt/nfs/share/data/models/Miromind-M1-SFT-7B/ --checkpoint-path /mnt/nfs/yeanbang/experiments/rlhf/dapo/v3/model/DAPO-Experiment-Baseline-V3/modeling-episode-0-step-200 -t 4 -i 4 -b vllm -a DAPO -imbs 8 -ibs 8 -tbs 32 -e 2 -g 16 -rt boxed -si 25 -s "Please reason step by step, and put your final answer within \\boxed{}." -tMbs 8 -tmbs 8 -p DAPO-Experiment-Baseline-V3_1 -zero 1 -pp 2 -tp 2 -mpt 1024 -mnt 4096 -nb 1 --enable_profiling

applications/ColossalChat/rl_example.py

@@ -18,6 +18,13 @@ os.environ["no_proxy"] = "127.0.0.1,localhost"
 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",
@@ -227,7 +234,9 @@ if __name__ == "__main__":
     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)
+    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":

applications/ColossalChat/rl_example_zero_bubble.py

@@ -261,7 +261,6 @@ if __name__ == "__main__":
                 stop=["</answer>"] if args.reward_type == "think_answer_tags" else None,
             )
         )
-        eval_generation_config = {"temperature": 0.6}  # used to update generation config for evaluation
     else:
         raise ValueError(f"Unsupported backend: {args.backend}")
 
@@ -290,7 +289,7 @@ if __name__ == "__main__":
     elif args.algo == "DAPO":
         # DAPO variant settings
         grpo_config = {
-            "filter_range": [0.01, 0.7],  # only filter out all zero batch and all one batch
+            "filter_range": [0.01, 0.8],  # only filter out all zero batch and all one batch
             "lr": args.learning_rate,
             "train_microbatch_size": args.train_microbatch_size,
             "dynamic_batching": True,
@@ -302,8 +301,8 @@ if __name__ == "__main__":
             "soft_over_length_punishment": True,
             "max_length": args.max_new_tokens + args.max_prompt_tokens,
             "max_new_tokens": args.max_new_tokens,
-            "cache_length": min(1024, int(args.max_new_tokens / 4)),
+            "cache_length": max(1024, int(args.max_new_tokens / 4)),
-            "filter_truncated_response": True,
+            "filter_truncated_response": False,
             "reward_fn_type": args.reward_type,
             "response_format_tags": (
                 {
@@ -362,15 +361,22 @@ if __name__ == "__main__":
         save_dir=os.path.join(args.save_dir, args.project.replace(" ", "_")),
         eval_dataset_config=(
             {
-                k: {"path": v, "max_length": args.max_prompt_tokens, "system_prompt": args.system_prompt}
+                k: {
+                    "path": v,
+                    "max_length": args.max_prompt_tokens,
+                    "system_prompt": args.system_prompt,
+                    "batch_size": 1,
+                    "n": 16,
+                    "max_tokens": 32 * 1024 - 512,
+                    "temperature": 0.6,
+                }
                 for k, v in json.loads(args.eval_dataset).items()
             }
             if args.eval_dataset
             else None
-        ),
+        ),  # support using different evaluation prompts/ parameters for each task
         eval_interval=args.eval_interval,
         eval_save_dir=os.path.join(args.eval_save_dir, args.project.replace(" ", "_")),
-        eval_generation_config=eval_generation_config,
         log_rollout_interval=20,
         rollout_save_dir=args.rollout_save_dir,
         enable_profiling=args.enable_profiling,

applications/ColossalChat/run_eval.sh

@@ -0,0 +1,54 @@
+# run eval for every saved checkpoint
+
+#!/bin/bash
+
+# Set the base directory where models are saved
+BASE_DIR="/mnt/nfs/yeanbang/ColossalAI/applications/ColossalChat/model/DAPO-Experiment-Baseline"
+BASE_DIR="/mnt/nfs/yeanbang/experiments/rlhf/dapo/v3/model/DAPO-Experiment-Baseline-V3"
+DATASET="/home/litong/workspace/data/rl-math-train-final.jsonl"
+EVAL_SCRIPT="rl_example_zero_bubble.py"
+# /mnt/nfs/yeanbang/ColossalAI/applications/ColossalChat/model/DAPO-Experiment-Baseline/modeling-episode-0-step-105
+# /mnt/nfs/yeanbang/ColossalAI/applications/ColossalChat/model/DAPO-Experiment-Baseline/modeling-episode-0-step-120
+# /mnt/nfs/yeanbang/ColossalAI/applications/ColossalChat/model/DAPO-Experiment-Baseline/modeling-episode-0-step-15
+# /mnt/nfs/yeanbang/ColossalAI/applications/ColossalChat/model/DAPO-Experiment-Baseline/modeling-episode-0-step-30
+# /mnt/nfs/yeanbang/ColossalAI/applications/ColossalChat/model/DAPO-Experiment-Baseline/modeling-episode-0-step-45
+# /mnt/nfs/yeanbang/ColossalAI/applications/ColossalChat/model/DAPO-Experiment-Baseline/modeling-episode-0-step-60
+# /mnt/nfs/yeanbang/ColossalAI/applications/ColossalChat/model/DAPO-Experiment-Baseline/modeling-episode-0-step-75
+# /mnt/nfs/yeanbang/ColossalAI/applications/ColossalChat/model/DAPO-Experiment-Baseline/modeling-episode-0-step-90
+# /mnt/nfs/yeanbang/ColossalAI/applications/ColossalChat/model/DAPO-Experiment-Baseline/modeling-episode-1-step-150
+# Loop through each model directory matching the pattern
+# 64K 65024
+# 32K 32256
+for model_path in "$BASE_DIR"/modeling-episode-0-step-150; do
+    if [ -d "$model_path" ]; then
+        echo "Evaluating model: $model_path"
+
+        python "$EVAL_SCRIPT" \
+            --dataset "$DATASET" \
+            --tokenizer-path "/home/litong/workspace/model/DeepSeek-R1-Distill-Qwen-7B" \
+            --model "$model_path" \
+            -t 1 \
+            -i 6 \
+            -b vllm \
+            -a GRPO \
+            -imbs 1 \
+            -ibs 8 \
+            -tbs 8 \
+            -e 2 \
+            -g 16 \
+            -rt boxed \
+            -si 25 \
+            -s "Please reason step by step, and put your final answer within \\boxed{}." \
+            -tMbs 2 \
+            -tmbs 1 \
+            -p DAPO-Experiment-Baseline-Eval \
+            -ei 15 \
+            -ed '{"AIME2024": "/mnt/nfs/yeanbang/data/RLHF_data/AIME/2024/AIME2024.jsonl"}' \
+            -zero 2 \
+            -mpt 512 \
+            -mnt 65024 \
+            --data_actor_buffer_size_limit 4 | grep "Accuracy on"
+
+        echo "Finished evaluating: $model_path"
+    fi
+done

colossalai/shardformer/modeling/qwen2.py

@@ -824,7 +824,7 @@ def get_lm_forward_with_dist_cross_entropy(shard_config: ShardConfig):
             output_attentions=output_attentions,
             output_hidden_states=output_hidden_states,
             return_dict=return_dict,
-            force_sp_output_gather=False,
+            # force_sp_output_gather=False,
         )
 
         hidden_states = outputs[0]