From 7e85d3a42cc180d8b8fa85ca815c3e90bf2cb970 Mon Sep 17 00:00:00 2001
From: Flora Cui <flora.cui@amd.com>
Date: Fri, 7 Aug 2026 23:43:42 +0800
Subject: [PATCH] [ROCm] Enable pinned memory on supported WSL2 kernels
 (#50126)

Signed-off-by: Flora Cui <flora.cui@amd.com>
---
 .../scripts/xpu/create-xpu-ecr-manifest.sh    |  4 +--
 vllm/platforms/rocm.py                        | 28 ++++++++++++++++++-
 2 files changed, 29 insertions(+), 3 deletions(-)

diff --git a/.buildkite/scripts/xpu/create-xpu-ecr-manifest.sh b/.buildkite/scripts/xpu/create-xpu-ecr-manifest.sh
index 78f2c33b36..9f21118d50 100644
--- a/.buildkite/scripts/xpu/create-xpu-ecr-manifest.sh
+++ b/.buildkite/scripts/xpu/create-xpu-ecr-manifest.sh
@@ -8,6 +8,6 @@ ARCH_TAG="${BUILDKITE_COMMIT}-$(uname -m)-xpu"
 PLATFORM_TAG="${BUILDKITE_COMMIT}-xpu"
 
 aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin ${REGISTRY}
-docker manifest rm ${REGISTRY}/${REPO}:${PLATFORM_TAG} || true
-docker manifest create ${REGISTRY}/${REPO}:${PLATFORM_TAG} ${REGISTRY}/${REPO}:${ARCH_TAG} --amend
+docker manifest rm "${REGISTRY}/${REPO}:${PLATFORM_TAG}" || true
+docker manifest create "${REGISTRY}/${REPO}:${PLATFORM_TAG}" "${REGISTRY}/${REPO}:${ARCH_TAG}" --amend
 docker manifest push ${REGISTRY}/${REPO}:${PLATFORM_TAG}
\ No newline at end of file
diff --git a/vllm/platforms/rocm.py b/vllm/platforms/rocm.py
index 35bc244261..dd577e1cd2 100644
--- a/vllm/platforms/rocm.py
+++ b/vllm/platforms/rocm.py
@@ -2,6 +2,7 @@
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
 import os
+import platform
 from datetime import timedelta
 from functools import cache, lru_cache, wraps
 from typing import TYPE_CHECKING
@@ -15,7 +16,7 @@ import vllm.envs as envs
 from vllm.logger import init_logger
 from vllm.v1.attention.backends.registry import AttentionBackendEnum
 
-from .interface import DeviceCapability, Platform, PlatformEnum
+from .interface import DeviceCapability, Platform, PlatformEnum, in_wsl
 
 if TYPE_CHECKING:
     from vllm.config import VllmConfig
@@ -113,6 +114,16 @@ def _rocm_device_count_stateless(cuda_visible_devices: str | None = None) -> int
     return r
 
 
+@cache
+def _get_wsl_kernel_version() -> tuple[int, ...] | None:
+    try:
+        release = platform.uname().release
+        parts = release.split("-")[0].split(".")
+        return tuple(int(part) for part in parts[:3])
+    except (TypeError, ValueError):
+        return None
+
+
 def _sync_hip_cuda_env_vars():
     """Ensure HIP_VISIBLE_DEVICES and CUDA_VISIBLE_DEVICES are consistent.
     Treats empty string as unset. Raises on genuine conflicts."""
@@ -536,6 +547,21 @@ class RocmPlatform(Platform):
         with contextlib.suppress(ImportError):
             import vllm._rocm_C  # noqa: F401
 
+    @classmethod
+    def is_pin_memory_available(cls) -> bool:
+        if in_wsl():
+            version = _get_wsl_kernel_version()
+            if version is None or version < (4, 19, 121):
+                # warning_once() causes a circular import on WSL, see #48397.
+                logger.warning(
+                    "Using 'pin_memory=False' as WSL is detected and the "
+                    "WSL2 kernel version is below 4.19.121. This may slow "
+                    "down performance. Please run `wsl --update`."
+                )
+                return False
+
+        return True
+
     @classmethod
     def get_valid_backends(
         cls,