diff --git a/verl/workers/config/rollout.py b/verl/workers/config/rollout.py
index 19cf7f5e322..549f357d2e1 100644
--- a/verl/workers/config/rollout.py
+++ b/verl/workers/config/rollout.py
@@ -13,9 +13,12 @@
 # limitations under the License.
 import warnings
 from dataclasses import dataclass, field
+from importlib.metadata import PackageNotFoundError
+from importlib.metadata import version as package_version
 from typing import Optional
 
 from omegaconf import MISSING, DictConfig, OmegaConf
+from packaging import version as vs
 
 from verl.base_config import BaseConfig
 from verl.utils.profiler import ProfilerConfig
@@ -35,6 +38,18 @@
 ]
 
 
+_VLLM_MM_PROCESSOR_CACHE_FIX_VERSION = "0.22.0"
+
+
+def _should_disable_vllm_mm_processor_cache() -> bool:
+    try:
+        vllm_version = package_version("vllm")
+    except PackageNotFoundError:
+        return True
+
+    return vs.parse(vllm_version) < vs.parse(_VLLM_MM_PROCESSOR_CACHE_FIX_VERSION)
+
+
 @dataclass
 class SamplingConfig(BaseConfig):
     temperature: float = 1.0
@@ -274,6 +289,8 @@ class RolloutConfig(BaseConfig):
 
     def __post_init__(self):
         """Validate the rollout config"""
+        self._maybe_disable_vllm_mm_processor_cache()
+
         # Deprecation warning for mode field - only async mode is supported
         if self.mode == "sync":
             raise ValueError(
@@ -339,3 +356,18 @@ def __post_init__(self):
             raise ValueError(
                 f"rollout.disaggregation.enabled=True requires rollout.name in ('sglang', 'vllm'); got {self.name!r}."
             )
+
+    def _maybe_disable_vllm_mm_processor_cache(self) -> None:
+        if self.name != "vllm" or not _should_disable_vllm_mm_processor_cache():
+            return
+
+        engine_kwargs = dict(self.engine_kwargs or {})
+        vllm_engine_kwargs = dict(engine_kwargs.get("vllm", {}) or {})
+        if "mm_processor_cache_gb" in vllm_engine_kwargs:
+            return
+
+        # Disable caching for vLLM < 0.22.0 to work around vLLM multimodal processor cache desync across pause/resume.
+        # See: https://github.com/vllm-project/vllm/pull/43001/
+        vllm_engine_kwargs["mm_processor_cache_gb"] = 0
+        engine_kwargs["vllm"] = vllm_engine_kwargs
+        self.engine_kwargs = engine_kwargs