[BUG] fix sglang veRL GptOSS rollout problem

python/sglang/srt/managers/tp_worker.py

@@ -15,6 +15,7 @@
 from __future__ import annotations
 
 import logging
+import os
 from abc import ABC, abstractmethod
 from typing import TYPE_CHECKING, Optional
 
@@ -56,6 +57,38 @@
 logger = logging.getLogger(__name__)
 
 
+def safe_load():
+    import torch.multiprocessing as mp
+
+    # 1) setting process sharing policy
+    try:
+        mp.set_sharing_strategy("file_system")
+        mp.set_start_method("fork", force=True)
+    except Exception:
+        pass
+
+    # 2) using forks method
+    try:
+        mp.set_start_method("fork", force=True)
+    except RuntimeError:
+        pass
+
+    import ctypes
+
+    try:
+        libc = ctypes.CDLL("libc.so.6", use_errno=True)
+        PR_SET_PTRACER = 0x59616D61
+        PR_SET_PTRACER_ANY = -1
+        libc.prctl(PR_SET_PTRACER, PR_SET_PTRACER_ANY, 0, 0, 0)
+    except Exception:
+        pass
+    pass
+
+
+if os.environ.get("SAFE_LOAD", "0") == "1":
+    safe_load()
+
+
 class BaseTpWorker(ABC):
     @abstractmethod
     def forward_batch_generation(self, forward_batch: ForwardBatch):

python/sglang/srt/utils/patch_torch.py

@@ -16,6 +16,7 @@
 import torch
 from packaging import version
 from torch.multiprocessing import reductions
+from torch.storage import UntypedStorage
 
 from sglang.srt.utils.common import is_npu
 
@@ -32,11 +33,15 @@ def monkey_patch_torch_reductions():
     if hasattr(reductions, "_reduce_tensor_original"):
         return
 
+    UntypedStorage._new_shared_cuda_original = UntypedStorage._new_shared_cuda
+
     reductions._reduce_tensor_original = reductions.reduce_tensor
     reductions._rebuild_cuda_tensor_original = reductions.rebuild_cuda_tensor
 
+    UntypedStorage._new_shared_cuda = _new_shared_cuda_safe
+
     reductions.reduce_tensor = _reduce_tensor_modified
-    reductions.rebuild_cuda_tensor = _rebuild_cuda_tensor_modified
+    reductions._rebuild_cuda_tensor = _rebuild_cuda_tensor_safe
 
     reductions.init_reductions()
 
@@ -59,6 +64,44 @@ def _rebuild_cuda_tensor_modified(*args):
     return reductions._rebuild_cuda_tensor_original(*args)
 
 
+# NOTE (yiakwy)
+def _new_shared_cuda_safe(*args, **kwargs):
+    """
+    Monkey-patch for torch.UntypedStorage._new_shared_cuda to avoid pidfd_getfd
+    failures in Docker / restricted environments.
+
+    Instead of using pidfd, fallback to CPU tensor and move to GPU manually later.
+    """
+    # Allocate CPU storage of the same size
+    try:
+        return UntypedStorage._new_shared_cuda_original(*args, **kwargs)
+    except:
+        # original impl of https://github.com/pytorch/pytorch/blob/f47dd0ddef1359e5b43e4b962412f67b30ecde56/torch/csrc/StorageSharing.cpp#L423
+        size_bytes = args[2]
+        if size_bytes is None or not isinstance(size_bytes, int):
+            raise RuntimeError("Cannot determine size for safe CUDA storage fallback")
+
+        return torch.UntypedStorage._new_shared(size_bytes)
+
+
+# NOTE (yiakwy)
+def _rebuild_cuda_tensor_safe(*args):
+    try:
+        return _rebuild_cuda_tensor_modified(*args)
+    except RuntimeError:
+        args = _modify_tuple(
+            args, _REDUCE_TENSOR_ARG_DEVICE_INDEX, _device_from_maybe_uuid
+        )
+
+        storage, metadata = args[1], args[2]
+
+        if storage.device.type != "cuda":
+            storage = storage.cuda()
+
+        tensor = torch._utils._rebuild_tensor_v2(storage, *metadata)
+        return tensor
+
+
 def _device_to_uuid(device: int) -> str:
     return str(torch.cuda.get_device_properties(device).uuid)