diff --git a/tests/v1/attention/test_flashinfer_metadata_builder.py b/tests/v1/attention/test_flashinfer_metadata_builder.py
new file mode 100644
index 0000000..5e85c81
--- /dev/null
+++ b/tests/v1/attention/test_flashinfer_metadata_builder.py
@@ -0,0 +1,120 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+"""Regression tests for FlashInfer's persistent host metadata buffers."""
+
+import unittest
+from itertools import product
+from unittest.mock import MagicMock, patch
+
+import torch
+
+from vllm.config import (
+    AttentionConfig,
+    CUDAGraphMode,
+    ModelConfig,
+    SpeculativeConfig,
+    VllmConfig,
+    set_current_vllm_config,
+)
+from vllm.platforms import current_platform
+from vllm.v1.attention.backends.utils import PerLayerParameters
+from vllm.v1.kv_cache_interface import FullAttentionSpec
+
+
+def check_persistent_host_buffer_pinning(
+    cuda_device: torch.device,
+    use_v2_model_runner: bool,
+    has_speculation: bool,
+    pin_memory: bool,
+):
+    from vllm.v1.attention.backends import flashinfer
+
+    # Mock configuration only: no model config resolution or weight downloads.
+    config = MagicMock(spec=VllmConfig)
+    config.model_config = MagicMock(spec=ModelConfig)
+    config.model_config.dtype = torch.float16
+    config.model_config.max_model_len = 32
+    config.model_config.get_num_attention_heads.return_value = 4
+    config.cache_config = MagicMock()
+    config.scheduler_config = MagicMock(max_num_seqs=2)
+    config.compilation_config = MagicMock(cudagraph_mode=CUDAGraphMode.NONE)
+    config.parallel_config = MagicMock(
+        decode_context_parallel_size=1,
+        dcp_kv_cache_interleave_size=1,
+    )
+    config.attention_config = AttentionConfig(use_trtllm_attention=False)
+    config.use_v2_model_runner = use_v2_model_runner
+    config.speculative_config = (
+        MagicMock(
+            spec=SpeculativeConfig,
+            num_speculative_tokens=1,
+            parallel_drafting=False,
+        )
+        if has_speculation
+        else None
+    )
+    kv_cache_spec = FullAttentionSpec(
+        block_size=16,
+        num_kv_heads=2,
+        head_size=64,
+        dtype=torch.float16,
+    )
+    per_layer_parameters = {
+        "layer.0": PerLayerParameters(
+            window_left=-1, logits_soft_cap=0.0, sm_scale=64**-0.5
+        )
+    }
+
+    # Exercise the real constructor and allocators; only layer registration
+    # and the platform pinning flag are patched. GPU buffers total 36 bytes.
+    with (
+        set_current_vllm_config(config),
+        patch.object(flashinfer, "PIN_MEMORY", pin_memory),
+        patch.object(
+            flashinfer, "get_per_layer_parameters", return_value=per_layer_parameters
+        ),
+    ):
+        builder = flashinfer.FlashInferMetadataBuilder(
+            kv_cache_spec, ["layer.0"], config, cuda_device
+        )
+
+    # V2 and speculative proposals can reuse host metadata while an earlier
+    # asynchronous H2D copy is in flight. Legacy non-spec decoding stays pinned.
+    expected_pinned = pin_memory and not use_v2_model_runner and not has_speculation
+    assert builder.pin_memory is expected_pinned
+    for name in ("paged_kv_indptr", "paged_kv_indices", "paged_kv_last_page_len"):
+        buffer = getattr(builder, name)
+        assert buffer.gpu.device == cuda_device, name
+        assert buffer.cpu.device.type == "cpu", name
+        assert buffer.cpu.is_pinned() is expected_pinned, name
+    assert builder.paged_kv_indptr_cpu_buffer.device.type == "cpu"
+    assert builder.paged_kv_indptr_cpu_buffer.is_pinned() is expected_pinned
+
+
+class TestFlashInferMetadataBuilder(unittest.TestCase):
+    @classmethod
+    def setUpClass(cls):
+        if not current_platform.is_cuda() or not torch.cuda.is_available():
+            raise unittest.SkipTest("Requires CUDA for real pinned host allocations")
+        try:
+            import flashinfer  # noqa: F401
+        except ImportError as exc:
+            raise unittest.SkipTest("Requires FlashInfer") from exc
+
+    def test_persistent_host_buffer_pinning(self):
+        device = torch.device("cuda:0")
+        for use_v2_model_runner, has_speculation, pin_memory in product(
+            (False, True), repeat=3
+        ):
+            with self.subTest(
+                use_v2_model_runner=use_v2_model_runner,
+                has_speculation=has_speculation,
+                pin_memory=pin_memory,
+            ):
+                check_persistent_host_buffer_pinning(
+                    device, use_v2_model_runner, has_speculation, pin_memory
+                )
+
+
+if __name__ == "__main__":
+    unittest.main()
diff --git a/vllm/v1/attention/backends/flashinfer.py b/vllm/v1/attention/backends/flashinfer.py
index 8031900..f11a1bd 100755
--- a/vllm/v1/attention/backends/flashinfer.py
+++ b/vllm/v1/attention/backends/flashinfer.py
@@ -718,10 +718,14 @@ class FlashInferMetadataBuilder(AttentionMetadataBuilder[FlashInferMetadata]):
                 "earlier GPUs."
             )
         # Preparing persistent buffers
-        # Since we do not have explicit synchronization in ModelRunnerV2, we do not pin
-        # reused CPU buffers to avoid a race condition between step N async copies to
-        # GPU and step N+1 buffer updates.
-        self.pin_memory = not vllm_config.use_v2_model_runner and PIN_MEMORY
+        # ModelRunnerV2 and speculative proposals can reuse host metadata before
+        # an earlier asynchronous H2D copy completes. Pageable staging prevents
+        # the next proposal's metadata from overwriting that copy's source.
+        self.pin_memory = (
+            PIN_MEMORY
+            and not vllm_config.use_v2_model_runner
+            and speculative_config is None
+        )
         self.paged_kv_indptr = self._make_buffer(max_num_reqs + 1)
         self.paged_kv_indptr_cpu_buffer = torch.zeros_like(
             self.paged_kv_indptr.cpu, pin_memory=self.pin_memory