[Disagg][StagingBuffer][1/2] Robustness and failure handling (#31217)

test/registered/disaggregation/test_disaggregation_different_tp.py

@@ -336,9 +336,10 @@ class TestDisaggregationMooncakeMHADecodeLargerTP(PDDisaggregationServerBase):
         self.assertGreater(metrics["score"], 0.60)
 
 
+# The prefill staging buffer is auto-sized to one chunk (chunked_prefill_size);
+# the decode ring is sized manually.
 STAGING_ENV = {
     "SGLANG_DISAGG_STAGING_BUFFER": "1",
-    "SGLANG_DISAGG_STAGING_BUFFER_SIZE_MB": "64",
     "SGLANG_DISAGG_STAGING_POOL_SIZE_MB": "1024",
 }
 

test/registered/unit/disaggregation/test_nixl_backend_basic.py

@@ -791,6 +791,7 @@ class TestNixlStaging(CustomTestCase):
 
     def test_do_staging_transfer_requeues_when_allocation_not_ready(self):
         mgr = self._make_manager()
+        mgr._staging_ctx = PrefillStagingContext()
         strategy = MagicMock()
         strategy.check_ready.return_value = (False, 0, -1, 0, -1)
         kv_chunk = TransferKVChunk(