[BREAKING][vllm, fsdp] feat: add Rollout-Training Mismatch Fix -- Truncated importance sampling

verl/trainer/config/actor/actor.yaml

@@ -71,6 +71,9 @@ loss_agg_mode: token-mean
 # Entropy regularization coefficient in PPO loss
 entropy_coeff: 0
 
+# whether to apply the truncated Importance Sampling (-1 for no importance sampling)
+imp_ratio_cap: 5
+
 # Whether to use KL loss instead of KL reward penalty. True for GRPO
 use_kl_loss: false
 

verl/trainer/config/rollout/rollout.yaml

@@ -159,7 +159,8 @@ multi_turn:
   format: hermes
 
 # support logging rollout prob for debugging purpose
-calculate_log_probs: False
+# "Truncated importance sampling" requires rollout log probs, set to True when turning on Truncated importance sampling
+calculate_log_probs: True 
 
 # [Experimental] agent loop based rollout configs
 agent:

verl/trainer/ppo/core_algos.py

@@ -747,6 +747,8 @@ def compute_policy_loss(
     cliprange_high=None,
     clip_ratio_c=3.0,
     loss_agg_mode: str = "token-mean",
+    rollout_log_probs=None,
+    imp_ratio_cap=-1,
 ):
     """
     Compute the clipped policy objective and related metrics for PPO.
@@ -807,6 +809,13 @@ def compute_policy_loss(
     )
 
     pg_losses = torch.where(advantages < 0, clip_pg_losses2, clip_pg_losses1)
+
+    if imp_ratio_cap > 0 and rollout_log_probs is not None:
+        # Apply truncated importance sampling
+        imp_ratio = torch.exp(old_log_prob - rollout_log_probs)
+        imp_ratio = torch.clamp(imp_ratio, max=imp_ratio_cap)
+        pg_losses = pg_losses * imp_ratio
+
     pg_loss = agg_loss(loss_mat=pg_losses, loss_mask=response_mask, loss_agg_mode=loss_agg_mode)
 
     return pg_loss, pg_clipfrac, ppo_kl, pg_clipfrac_lower
@@ -820,6 +829,8 @@ def compute_policy_loss_vanilla(
     response_mask: torch.Tensor,
     loss_agg_mode: str = "token-mean",
     config: Optional[DictConfig | AlgoConfig] = None,
+    rollout_log_probs=None,
+    imp_ratio_cap=-1,
 ) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
     """
     Compute the clipped policy objective and related metrics for PPO.
@@ -884,6 +895,13 @@ def compute_policy_loss_vanilla(
     )
 
     pg_losses = torch.where(advantages < 0, clip_pg_losses2, clip_pg_losses1)
+
+    if imp_ratio_cap > 0 and rollout_log_probs is not None:
+        # Apply truncated importance sampling
+        imp_ratio = torch.exp(old_log_prob - rollout_log_probs)
+        imp_ratio = torch.clamp(imp_ratio, max=imp_ratio_cap)
+        pg_losses = pg_losses * imp_ratio
+
     pg_loss = agg_loss(loss_mat=pg_losses, loss_mask=response_mask, loss_agg_mode=loss_agg_mode)
 
     return pg_loss, pg_clipfrac, ppo_kl, pg_clipfrac_lower
@@ -1270,6 +1288,19 @@ def compute_value_loss(
 
 
 def kl_penalty(logprob: torch.FloatTensor, ref_logprob: torch.FloatTensor, kl_penalty) -> torch.FloatTensor:
+    """
+    The expectation of k1 and k3 estimator is the expectaed value of KL, 
+    but the expected gradient of k1 and k3 estimator is not the expectaed gradient of KL!
+    On the other hand k2 estimator gives right gradient estimator, 
+    so we use a straight through trick here
+    """
+    forward_score = kl_penalty_forward(logprob, ref_logprob, kl_penalty)
+    backward_score = 0.5 * (logprob - ref_logprob).square()
+
+    return backward_score - backward_score.detach() + forward_score.detach()
+
+
+def kl_penalty_forward(logprob: torch.FloatTensor, ref_logprob: torch.FloatTensor, kl_penalty) -> torch.FloatTensor:
     """Compute KL divergence given logprob and ref_logprob.
     Copied from https://github.com/huggingface/trl/blob/main/trl/trainer/ppo_trainer.py#L1104
     See more description in http://joschu.net/blog/kl-approx.html

verl/workers/actor/dp_actor.py

@@ -376,6 +376,8 @@ def update_policy(self, data: DataProto):
         ]
         if self.config.use_kl_loss:
             select_keys.append("ref_log_prob")
+        if self.config.imp_ratio_cap > 0:
+            select_keys.append("rollout_log_probs")
 
         has_multi_modal_inputs = "multi_modal_inputs" in data.non_tensor_batch.keys()
         non_tensor_select_keys = ["multi_modal_inputs"] if has_multi_modal_inputs else []
@@ -405,6 +407,7 @@ def update_policy(self, data: DataProto):
                     model_inputs = {**micro_batch.batch, **micro_batch.non_tensor_batch}
                     response_mask = model_inputs["response_mask"]
                     old_log_prob = model_inputs["old_log_probs"]
+                    rollout_log_probs = model_inputs["rollout_log_probs"] if self.config.imp_ratio_cap > 0 else None
                     advantages = model_inputs["advantages"]
 
                     entropy_coeff = self.config.entropy_coeff
@@ -435,6 +438,8 @@ def update_policy(self, data: DataProto):
                         response_mask=response_mask,
                         loss_agg_mode=loss_agg_mode,
                         config=self.config,
+                        rollout_log_probs=rollout_log_probs,
+                        imp_ratio_cap=self.config.imp_ratio_cap,
                     )
 
                     if entropy_coeff != 0: