[perf] fix: set use_reentrant=False when enable gradient checkpointing (#114)

- Set use_reentrant=False to avoid duplicate allgather in backward when
gradient checkpointing is enabled.
- Optimize temperature computation by using inplace op
- Fix testing logics

verl/trainer/ppo/ray_trainer.py

@@ -611,7 +611,8 @@ class RayPPOTrainer(object):
                         metrics.update(actor_output_metrics)
 
                     # validate
-                    if self.val_reward_fn is not None and self.global_steps % self.config.trainer.test_freq == 0:
+                    if self.val_reward_fn is not None and self.config.trainer.test_freq > 0 and \
+                        self.global_steps % self.config.trainer.test_freq == 0:
                         with _timer('testing', timing_raw):
                             val_metrics: dict = self._validate()
                         metrics.update(val_metrics)

verl/workers/actor/dp_actor.py

@@ -92,7 +92,7 @@ class DataParallelPPOActor(BasePPOActor):
                                            use_cache=False)  # prevent model thinks we are generating
                 logits_rmpad = output.logits.squeeze(0)  # (total_nnz, vocab_size)
 
-                logits_rmpad /= temperature
+                logits_rmpad.div_(temperature)
 
                 # compute entropy
                 entropy_rmpad = verl_F.entropy_from_logits(logits_rmpad)  # ((total_nnz / sp) + pad)
@@ -127,7 +127,8 @@ class DataParallelPPOActor(BasePPOActor):
                                            attention_mask=attention_mask,
                                            position_ids=position_ids,
                                            use_cache=False)  # prevent model thinks we are generating
-                logits = output.logits / temperature
+                logits = output.logits
+                logits.div_(temperature)
                 logits = logits[:, -response_length - 1:-1]  # (bsz, response_length)
                 log_probs = logprobs_from_logits(logits, micro_batch['responses'])
                 entropy = verl_F.entropy_from_logits(logits)  # (bsz, response_length)

verl/workers/fsdp_workers.py

@@ -164,7 +164,7 @@ class ActorRolloutRefWorker(Worker):
             actor_module.to(torch_dtype)
 
             if enable_gradient_checkpointing:
-                actor_module.gradient_checkpointing_enable()
+                actor_module.gradient_checkpointing_enable(gradient_checkpointing_kwargs={'use_reentrant': False})
         torch.distributed.barrier()
 
         if self.rank == 0:
@@ -212,7 +212,8 @@ class ActorRolloutRefWorker(Worker):
             sharding_strategy=sharding_strategy,  # zero3
             mixed_precision=mixed_precision,
             sync_module_states=True,
-            device_mesh=self.device_mesh)
+            device_mesh=self.device_mesh,
+            forward_prefetch=False)
 
         log_gpu_memory_usage('After Actor FSDP init', logger=logger)
 
@@ -575,7 +576,7 @@ class CriticWorker(Worker):
             critic_module.to(torch_dtype)
 
             if config.model.get('enable_gradient_checkpointing', False):
-                critic_module.gradient_checkpointing_enable()
+                critic_module.gradient_checkpointing_enable(gradient_checkpointing_kwargs={'use_reentrant': False})
         if self.rank == 0:
             print_model_size(critic_module)
 
@@ -603,7 +604,8 @@ class CriticWorker(Worker):
                              device_id=torch.cuda.current_device(),
                              sharding_strategy=ShardingStrategy.FULL_SHARD,
                              mixed_precision=mixed_precision,
-                             sync_module_states=True)
+                             sync_module_states=True,
+                             forward_prefetch=False)
 
         log_gpu_memory_usage('After critic FSDP', logger=None)
 
@@ -806,7 +808,8 @@ class RewardModelWorker(Worker):
             device_id=torch.cuda.current_device(),
             sharding_strategy=ShardingStrategy.FULL_SHARD,  # zero3
             sync_module_states=True,
-            cpu_offload=CPUOffload(offload_params=self.config.model.fsdp_config.param_offload))
+            cpu_offload=CPUOffload(offload_params=self.config.model.fsdp_config.param_offload),
+            forward_prefetch=False)
 
         return reward_module