[vllm, sglang, rollout] fix: Fix a mistake when running run_qwen3_vl-30b-megatron.sh with latest verl and vllm0.12 (#4810)

Author: cboss6

verl/workers/rollout/sglang_rollout/async_sglang_server.py

@@ -44,7 +44,12 @@
 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
-from verl.workers.rollout.utils import get_free_port, is_valid_ipv6_address, run_unvicorn
+from verl.workers.rollout.utils import (
+    get_free_port,
+    get_max_position_embeddings,
+    is_valid_ipv6_address,
+    run_unvicorn,
+)
 
 logger = logging.getLogger(__file__)
 logger.setLevel(logging.INFO)
@@ -83,7 +88,7 @@ def __init__(
 
         self.config: RolloutConfig = omega_conf_to_dataclass(config)
         self.model_config: HFModelConfig = omega_conf_to_dataclass(model_config, dataclass_type=HFModelConfig)
-        self.config.max_model_len = self.model_config.hf_config.max_position_embeddings
+        self.config.max_model_len = get_max_position_embeddings(self.model_config.hf_config)
         self.rollout_mode = rollout_mode
         self.workers = workers
 

verl/workers/rollout/utils.py

@@ -23,6 +23,18 @@
 logger = logging.getLogger(__file__)
 
 
+def get_max_position_embeddings(hf_config) -> int:
+    max_len = getattr(hf_config, "max_position_embeddings", None)
+    if max_len is None:
+        text_config = getattr(hf_config, "text_config", None)
+        if text_config is not None:
+            max_len = getattr(text_config, "max_position_embeddings", None)
+
+    if max_len is None:
+        raise ValueError("max_position_embeddings not found in HFModelConfig!")
+    return int(max_len)
+
+
 def is_valid_ipv6_address(address: str) -> bool:
     try:
         ipaddress.IPv6Address(address)

verl/workers/rollout/vllm_rollout/vllm_async_server.py

@@ -48,7 +48,12 @@
 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
-from verl.workers.rollout.utils import get_free_port, is_valid_ipv6_address, run_unvicorn
+from verl.workers.rollout.utils import (
+    get_free_port,
+    get_max_position_embeddings,
+    is_valid_ipv6_address,
+    run_unvicorn,
+)
 from verl.workers.rollout.vllm_rollout import vLLMAsyncRollout
 from verl.workers.rollout.vllm_rollout.utils import (
     VLLM_LORA_INT_ID,
@@ -195,7 +200,7 @@ def __init__(
 
         self.config: RolloutConfig = omega_conf_to_dataclass(config)
         self.model_config: HFModelConfig = omega_conf_to_dataclass(model_config, dataclass_type=HFModelConfig)
-        self.config.max_model_len = self.model_config.hf_config.max_position_embeddings
+        self.config.max_model_len = get_max_position_embeddings(self.model_config.hf_config)
         self.rollout_mode = rollout_mode
         self.workers = workers