From 2c05ed4e7776c876478f4b2db61acb12b9a27d01 Mon Sep 17 00:00:00 2001
From: JohnQinAMD <yanyuan.qin@amd.com>
Date: Sun, 6 Sep 2026 12:28:58 -0700
Subject: [PATCH] [ROCm] Stage large pageable H2D copies instead of pinning
 them in place (#37720)

---
 python/sglang/srt/arg_groups/platform_hook.py |  5 ++
 .../unit/test_rocm_pageable_h2d_staging.py    | 60 +++++++++++++++++++
 2 files changed, 65 insertions(+)
 create mode 100644 test/registered/unit/test_rocm_pageable_h2d_staging.py

diff --git a/python/sglang/srt/arg_groups/platform_hook.py b/python/sglang/srt/arg_groups/platform_hook.py
index 32d7028a0..f32876aaa 100644
--- a/python/sglang/srt/arg_groups/platform_hook.py
+++ b/python/sglang/srt/arg_groups/platform_hook.py
@@ -4,6 +4,7 @@
 from __future__ import annotations
 
 import logging
+import os
 from typing import Any
 
 from sglang.srt.arg_groups.overrides import (
@@ -64,6 +65,10 @@ def handle_amd_specifics(server_args: Any):
         declare_resolution(
             server_args, "_handle_amd_specifics", triton_attention_num_kv_splits=16
         )
+        # Above this the HIP runtime registers a pageable H2D source with the
+        # GPU rather than staging it, and the MMU notifier on that registration
+        # evicts our KFD queues once per tensor while weights load. In KB.
+        os.environ.setdefault("GPU_PINNED_MIN_XFER_SIZE", str(4 * 1024 * 1024))
 
 
 def handle_nccl_pre_warm(server_args: Any):
diff --git a/test/registered/unit/test_rocm_pageable_h2d_staging.py b/test/registered/unit/test_rocm_pageable_h2d_staging.py
new file mode 100644
index 000000000..d8693df01
--- /dev/null
+++ b/test/registered/unit/test_rocm_pageable_h2d_staging.py
@@ -0,0 +1,60 @@
+"""The HIP weight loader must stage large pageable copies, not pin them in place.
+
+The default raised here is what keeps the loader out of the MMU-notifier eviction
+loop, so what is worth pinning is that it is a default and not an override: an
+operator who has tuned the threshold must keep their value.
+"""
+
+import os
+import unittest
+from unittest.mock import patch
+
+from sglang.srt.arg_groups.platform_hook import handle_amd_specifics
+from sglang.test.ci.ci_register import register_cpu_ci
+from sglang.test.test_utils import CustomTestCase
+
+register_cpu_ci(est_time=5, suite="base-a-test-cpu")
+
+VAR = "GPU_PINNED_MIN_XFER_SIZE"
+
+
+class _Platform:
+    def __init__(self, is_hip: bool):
+        self.is_hip = is_hip
+
+
+class TestRocmPageableH2DStaging(CustomTestCase):
+    def setUp(self):
+        self._saved = os.environ.pop(VAR, None)
+
+    def tearDown(self):
+        os.environ.pop(VAR, None)
+        if self._saved is not None:
+            os.environ[VAR] = self._saved
+
+    def _run(self, is_hip: bool):
+        with (
+            patch(
+                "sglang.srt.arg_groups.platform_hook.get_platform",
+                return_value=_Platform(is_hip),
+            ),
+            patch("sglang.srt.arg_groups.platform_hook.declare_resolution"),
+        ):
+            handle_amd_specifics(object())
+
+    def test_set_on_hip(self):
+        self._run(is_hip=True)
+        self.assertEqual(os.environ.get(VAR), str(4 * 1024 * 1024))
+
+    def test_absent_off_hip(self):
+        self._run(is_hip=False)
+        self.assertIsNone(os.environ.get(VAR))
+
+    def test_an_operators_value_survives(self):
+        os.environ[VAR] = "12345"
+        self._run(is_hip=True)
+        self.assertEqual(os.environ[VAR], "12345")
+
+
+if __name__ == "__main__":
+    unittest.main()