wrap if else into the function

Author: mxc19912008

verl/utils/transformers_compat.py

@@ -22,6 +22,10 @@
 
 from packaging import version
 
+from verl.utils.logging import get_logger
+
+logger = get_logger(__name__)
+
 # Handle version compatibility for flash_attn_supports_top_left_mask
 # This function was added in newer versions of transformers
 try:
@@ -57,22 +61,25 @@ def is_transformers_version_in_range(min_version: Optional[str] = None, max_vers
     return lower_bound_check and upper_bound_check
 
 
-def get_max_position_embeddings(hf_config: Any) -> Optional[int]:
-    """Best-effort resolution of model context length from HF configs.
-
-    Works for:
-      - text-only configs where max_position_embeddings is top-level
-      - multimodal wrapper configs (e.g., Qwen3-VL) where it lives in text_config
-    """
+def resolve_max_model_len_from_hf_config(hf_config: Any) -> int | None:
     mpe = getattr(hf_config, "max_position_embeddings", None)
     if isinstance(mpe, int):
         return mpe
-
-    # Common wrappers for VLMs / composite configs
     for subname in ("text_config", "language_config", "llm_config"):
-        subcfg = getattr(hf_config, subname, None)
-        mpe = getattr(subcfg, "max_position_embeddings", None) if subcfg is not None else None
+        sub = getattr(hf_config, subname, None)
+        mpe = getattr(sub, "max_position_embeddings", None) if sub is not None else None
         if isinstance(mpe, int):
             return mpe
-
     return None
+
+
+def maybe_set_max_model_len_from_hf_config(config: Any, hf_config: Any) -> None:
+    mpe = resolve_max_model_len_from_hf_config(hf_config)
+    if mpe is not None:
+        config.max_model_len = mpe
+    else:
+        logger.warning(
+            "Cannot infer max_model_len from hf_config=%s; keeping max_model_len=%s",
+            type(hf_config),
+            getattr(config, "max_model_len", None),
+        )

verl/workers/rollout/sglang_rollout/async_sglang_server.py

@@ -41,7 +41,7 @@
 
 from verl.single_controller.ray import RayClassWithInitArgs
 from verl.utils.config import omega_conf_to_dataclass
-from verl.utils.transformers_compat import get_max_position_embeddings
+from verl.utils.transformers_compat import maybe_set_max_model_len_from_hf_config
 from verl.workers.config import HFModelConfig, RolloutConfig
 from verl.workers.rollout.replica import RolloutMode, RolloutReplica, TokenOutput
 from verl.workers.rollout.sglang_rollout.sglang_rollout import ServerAdapter, _set_envs_and_config
@@ -84,19 +84,8 @@ def __init__(
 
         self.config: RolloutConfig = omega_conf_to_dataclass(config)
         self.model_config: HFModelConfig = omega_conf_to_dataclass(model_config, dataclass_type=HFModelConfig)
-        mpe = get_max_position_embeddings(self.model_config.hf_config)
-        if mpe is not None:
-            # Don't accidentally exceed model limit; clamp if user set something smaller.
-            if getattr(self.config, "max_model_len", None) is not None:
-                self.config.max_model_len = min(self.config.max_model_len, mpe)
-            else:
-                self.config.max_model_len = mpe
-        else:
-            logger.warning(
-                "Cannot infer max_position_embeddings from hf_config=%s; keeping max_model_len=%s",
-                type(self.model_config.hf_config),
-                getattr(self.config, "max_model_len", None),
-            )
+        # safely make sure config.max_model_len doesn't exceed hf_config's max_position_embeddings + prompt_length
+        maybe_set_max_model_len_from_hf_config(self.config, self.model_config.hf_config)
         self.rollout_mode = rollout_mode
         self.workers = workers
 

verl/workers/rollout/vllm_rollout/vllm_async_server.py

@@ -45,7 +45,7 @@
 
 from verl.single_controller.ray import RayClassWithInitArgs
 from verl.utils.config import omega_conf_to_dataclass
-from verl.utils.transformers_compat import get_max_position_embeddings
+from verl.utils.transformers_compat import maybe_set_max_model_len_from_hf_config
 from verl.utils.vllm.vllm_fp8_utils import apply_vllm_fp8_patches
 from verl.workers.config import HFModelConfig, RolloutConfig
 from verl.workers.rollout.replica import RolloutMode, RolloutReplica, TokenOutput
@@ -196,19 +196,8 @@ def __init__(
 
         self.config: RolloutConfig = omega_conf_to_dataclass(config)
         self.model_config: HFModelConfig = omega_conf_to_dataclass(model_config, dataclass_type=HFModelConfig)
-        mpe = get_max_position_embeddings(self.model_config.hf_config)
-        if mpe is not None:
-            # Don't accidentally exceed model limit; clamp if user set something smaller.
-            if getattr(self.config, "max_model_len", None) is not None:
-                self.config.max_model_len = min(self.config.max_model_len, mpe)
-            else:
-                self.config.max_model_len = mpe
-        else:
-            logger.warning(
-                "Cannot infer max_position_embeddings from hf_config=%s; keeping max_model_len=%s",
-                type(self.model_config.hf_config),
-                getattr(self.config, "max_model_len", None),
-            )
+        # safely make sure config.max_model_len doesn't exceed hf_config's max_position_embeddings + prompt_length
+        maybe_set_max_model_len_from_hf_config(self.config, self.model_config.hf_config)
         self.rollout_mode = rollout_mode
         self.workers = workers