[XPU] Support register KV offload mmap region as pinned host memory on XPU (#51956)

Signed-off-by: Chaojun Zhang <chaojun.zhang@intel.com>
This commit is contained in:
Chaojun Zhang
2026-10-09 18:58:07 +08:00
committed by GitHub
parent 5cadabfafd
commit b027ac8312
6 changed files with 151 additions and 100 deletions
@@ -772,17 +772,14 @@ def test_cleanup_idempotent(iid):
def test_cleanup_unregisters_every_pinned_chunk(iid, monkeypatch):
"""cleanup() must release every chunk from a split host registration."""
r = _make_region(iid)
cudart = MagicMock()
cudart.cudaHostUnregister.return_value = MagicMock(value=0)
monkeypatch.setattr(region_module, "current_platform", MagicMock())
region_module.current_platform.is_cuda_alike.return_value = True
monkeypatch.setattr(region_module.torch.cuda, "cudart", lambda: cudart)
unregister = MagicMock()
monkeypatch.setattr(region_module, "host_unregister", unregister)
r.pinned_addresses = [0x100000, 0x200000, 0x300000]
r.is_pinned = True
r.cleanup()
assert cudart.cudaHostUnregister.call_args_list == [
assert unregister.call_args_list == [
call(0x300000),
call(0x200000),
call(0x100000),
@@ -791,40 +788,35 @@ def test_cleanup_unregisters_every_pinned_chunk(iid, monkeypatch):
@pytest.fixture
def host_register(monkeypatch):
"""Pretend to run on CUDA with host registrations capped at seven pages.
Registration goes to a mock CudaRTLibrary and the region's cleanup to a
separate mock torch runtime. Both stay installed through the cleanup that
``_region`` runs on exit.
"""Mock the cross-platform host_register/host_unregister helpers with
registrations capped at seven pages. The same unregister mock backs both
pin_mmap_region's rollback path and the region's cleanup() on exit from
``_region``.
"""
cudart = MagicMock()
cudart.cudaHostRegister.return_value = 0
cudart.cudaHostUnregister.return_value = 0
cudart.cudaGetLastError.return_value = 2
torch_cudart = MagicMock()
torch_cudart.cudaHostUnregister.return_value = MagicMock(value=0)
monkeypatch.setattr(gpu_worker.current_platform, "is_cuda_alike", lambda: True)
register = MagicMock(return_value=True)
unregister = MagicMock()
monkeypatch.setattr(gpu_worker, "MAX_HOST_REGISTER_CHUNK_BYTES", 7 * PAGE_SIZE)
monkeypatch.setattr(gpu_worker, "CudaRTLibrary", lambda: cudart)
monkeypatch.setattr(region_module.torch.cuda, "cudart", lambda: torch_cudart)
return cudart, torch_cudart
monkeypatch.setattr(gpu_worker, "host_register", register)
monkeypatch.setattr(gpu_worker, "host_unregister", unregister)
monkeypatch.setattr(region_module, "host_unregister", unregister)
return register, unregister
def test_pin_mmap_region_registers_row_aligned_chunks(iid, host_register):
"""Chunks end on row boundaries: 3-page rows under a 7-page cap register
as 6 + 6 + 3 pages, where a raw byte cap would give 7 + 7 + 1."""
cudart, torch_cudart = host_register
register, unregister = host_register
with _region(iid, num_chunks=5, cpu_page_size=3 * PAGE_SIZE) as region:
gpu_worker.pin_mmap_region(region)
base = region._base.data_ptr()
assert cudart.mock_calls == [
call.cudaHostRegister(base, 6 * PAGE_SIZE),
call.cudaHostRegister(base + 6 * PAGE_SIZE, 6 * PAGE_SIZE),
call.cudaHostRegister(base + 12 * PAGE_SIZE, 3 * PAGE_SIZE),
assert register.call_args_list == [
call(base, 6 * PAGE_SIZE),
call(base + 6 * PAGE_SIZE, 6 * PAGE_SIZE),
call(base + 12 * PAGE_SIZE, 3 * PAGE_SIZE),
]
assert region.is_pinned
assert torch_cudart.cudaHostUnregister.call_args_list == [
assert unregister.call_args_list == [
call(base + 12 * PAGE_SIZE),
call(base + 6 * PAGE_SIZE),
call(base),
@@ -833,10 +825,10 @@ def test_pin_mmap_region_registers_row_aligned_chunks(iid, host_register):
@pytest.mark.parametrize("fail_at", [0, 1, 2])
def test_pin_mmap_region_failure_leaves_region_pageable(iid, host_register, fail_at):
"""A failed chunk drains its pending error and unregisters the chunks
before it on the same handle, so the region is pinned whole or not at all."""
cudart, torch_cudart = host_register
cudart.cudaHostRegister.side_effect = [0] * fail_at + [2]
"""A failed chunk unregisters the chunks before it on the same range, so
the region is pinned whole or not at all."""
register, unregister = host_register
register.side_effect = [True] * fail_at + [False]
with _region(iid, num_chunks=5, cpu_page_size=3 * PAGE_SIZE) as region:
gpu_worker.pin_mmap_region(region)
base = region._base.data_ptr()
@@ -845,25 +837,22 @@ def test_pin_mmap_region_failure_leaves_region_pageable(iid, host_register, fail
registered = [base + i * 6 * PAGE_SIZE for i in range(fail_at)]
failed_size = min(6, 15 - 6 * fail_at) * PAGE_SIZE
assert cudart.mock_calls == [
*(call.cudaHostRegister(address, 6 * PAGE_SIZE) for address in registered),
call.cudaHostRegister(base + fail_at * 6 * PAGE_SIZE, failed_size),
call.cudaGetLastError(),
*(call.cudaHostUnregister(address) for address in reversed(registered)),
assert register.call_args_list == [
*(call(address, 6 * PAGE_SIZE) for address in registered),
call(base + fail_at * 6 * PAGE_SIZE, failed_size),
]
assert unregister.call_args_list == [
call(address) for address in reversed(registered)
]
torch_cudart.cudaHostUnregister.assert_not_called()
def test_pin_mmap_region_without_cuda_runtime_stays_pageable(
iid, host_register, monkeypatch
def test_pin_mmap_region_when_host_register_unsupported_stays_pageable(
iid, monkeypatch
):
"""A CUDA runtime that cannot be loaded leaves the region pageable instead
of failing worker startup."""
def missing_runtime():
raise AssertionError("libcudart is not loaded in the current process")
monkeypatch.setattr(gpu_worker, "CudaRTLibrary", missing_runtime)
"""A platform without host registration support (host_register()
returning False) leaves the region pageable instead of failing worker
startup."""
monkeypatch.setattr(gpu_worker, "host_register", lambda ptr, num_bytes: False)
with _region(iid) as region:
gpu_worker.pin_mmap_region(region)
assert not region.is_pinned
+3
View File
@@ -307,6 +307,7 @@ def test_cpu_spec_replicated_disabled_without_shared_region(
# those private buffers empty and corrupting subsequent loads.
import vllm.v1.kv_offload.cpu.spec as cpu_spec_module
monkeypatch.setattr(cpu_spec_module.current_platform, "is_xpu", lambda: False)
monkeypatch.setattr(cpu_spec_module.current_platform, "is_cuda_alike", lambda: rocm)
monkeypatch.setattr(cpu_spec_module.current_platform, "is_rocm", lambda: rocm)
worker_kv_bytes_per_block = SharedOffloadRegion.BLOCK_SIZE_ALIGNMENT
@@ -341,6 +342,7 @@ def test_cpu_spec_replicated_layout_truth_matrix(
# actually allocates on the shared region.
import vllm.v1.kv_offload.cpu.spec as cpu_spec_module
monkeypatch.setattr(cpu_spec_module.current_platform, "is_xpu", lambda: False)
monkeypatch.setattr(
cpu_spec_module.current_platform, "is_cuda_alike", lambda: cuda_alike
)
@@ -417,6 +419,7 @@ def test_cpu_spec_create_worker_uses_tensor_path_without_shared_region(
worker_calls.append(kwargs)
return MagicMock()
monkeypatch.setattr(cpu_spec_module.current_platform, "is_xpu", lambda: False)
monkeypatch.setattr(cpu_spec_module.current_platform, "is_cuda_alike", lambda: rocm)
monkeypatch.setattr(cpu_spec_module.current_platform, "is_rocm", lambda: rocm)
monkeypatch.setattr(cpu_spec_module, "SharedOffloadRegion", fake_region_ctor)