| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | import time |
| 4 | import uuid |
| 5 | |
| 6 | import pytest |
| 7 | import torch |
| 8 | |
| 9 | import vllm.distributed.ec_transfer.ec_connector.cpu.scheduler as sched_mod |
| 10 | from tests.v1.ec_connector.unit.utils import create_ec_vllm_config |
| 11 | from vllm.distributed.ec_transfer.ec_connector.cpu.ec_shared_region import ( |
| 12 | ECSharedRegion, |
| 13 | ) |
| 14 | from vllm.distributed.ec_transfer.ec_connector.cpu.scheduler import ECCPUScheduler |
| 15 | from vllm.multimodal.inputs import ( |
| 16 | MultiModalFieldElem, |
| 17 | MultiModalKwargsItem, |
| 18 | MultiModalSharedField, |
| 19 | ) |
| 20 | |
| 21 | _N, _BS, _HID, _ES = 16, 64, 32, 2 |
| 22 | |
| 23 | |
| 24 | class _Pos: |
| 25 | def __init__(self, offset, length): |
| 26 | self.offset, self.length = offset, length |
| 27 | |
| 28 | |
| 29 | class _Feature: |
| 30 | def __init__(self, mm_hash, length=1, data=None): |
| 31 | self.mm_hash = mm_hash |
| 32 | self.identifier = mm_hash |
| 33 | self.mm_position = _Pos(0, length) |
| 34 | self.data = data |
| 35 | self.modality = "image" |
| 36 | |
| 37 | |
| 38 | class _Request: |
| 39 | def __init__(self, features, params=None, req_id="r1"): |
| 40 | self.mm_features = features |
| 41 | self.request_id = req_id |
| 42 | self.ec_transfer_params = params |
| 43 | |
| 44 | |
| 45 | class _FakeResults: |
| 46 | def __init__(self): |
| 47 | self.completed = set() |
| 48 | self.tombstoned = set() |
| 49 | self.quarantined = set() |
| 50 | self.retryable = set() |
| 51 | self.settled = [] |
| 52 | |
| 53 | |
| 54 | class _FakeSession: |
| 55 | """Records started transfers; lets the test flip them to completed.""" |
| 56 | |
| 57 | def __init__(self): |
| 58 | self.started = [] |
| 59 | self._results = _FakeResults() |
| 60 | |
| 61 | def start_xfer(self, mm_hash, block_indices, deadline): |
| 62 | self.started.append(mm_hash) |
| 63 | |
| 64 | def poll(self, messages, now): |
| 65 | pass |
| 66 | |
| 67 | def take_results(self): |
| 68 | r, self._results = self._results, _FakeResults() |
| 69 | return r |
| 70 | |
| 71 | def close(self): |
| 72 | pass |
| 73 | |
| 74 | |
| 75 | class _FakeTransport: |
| 76 | """Stand-in for ZmqClientTransport so _poll_step runs without ZMQ.""" |
| 77 | |
| 78 | def poll(self): |
| 79 | return {} |
| 80 | |
| 81 | def poll_dead(self): |
| 82 | return [] |
| 83 | |
| 84 | def close(self): |
| 85 | pass |
| 86 | |
| 87 | |
| 88 | def _consumer_sched(monkeypatch): |
| 89 | def _region(cfg): |
| 90 | return ECSharedRegion( |
| 91 | engine_id="eng-" + str(uuid.uuid4()), num_blocks=_N, block_size_bytes=_BS |
| 92 | ) |
| 93 | |
| 94 | monkeypatch.setattr(sched_mod, "create_ec_shared_region", _region) |
| 95 | |
| 96 | # ec_enable_nixl defaults to False, so this builds gate-off; the fields |
| 97 | # below flip on the NIXL consumer state directly. |
| 98 | s = ECCPUScheduler(create_ec_vllm_config(ec_role="ec_consumer")) |
| 99 | # Turn on NIXL consumer state without constructing real transports. |
| 100 | s._nixl_enabled = True |
| 101 | s._transport = _FakeTransport() |
| 102 | s._data = None |
| 103 | s._compat_hash = "c" |
| 104 | # _setup_nixl normally computes these from model_config; set them |
| 105 | # directly since this helper builds gate-off then flips fields on. |
| 106 | s._hidden_dim = _HID |
| 107 | s._element_size = _ES |
| 108 | s._metadata_resolver._cache["image"] = {"image_grid_thw"} |
| 109 | return s |
| 110 | |
| 111 | |
| 112 | def _params(mm_hash, length): |
| 113 | return { |
| 114 | mm_hash: { |
| 115 | "peer_host": "h", |
| 116 | "peer_port": 1, |
| 117 | "size_bytes": length * _HID * _ES, |
| 118 | } |
| 119 | } |
| 120 | |
| 121 | |
| 122 | def test_new_remote_read_defers_then_completes(monkeypatch): |
| 123 | s = _consumer_sched(monkeypatch) |
| 124 | fake = _FakeSession() |
| 125 | |
| 126 | # Route _start_xfer to our fake session instead of real ZMQ/NIXL, but |
| 127 | # still reserve a real not-ready cache entry so mark_ready works. |
| 128 | def _fake_start(mm_hash, info, size): |
| 129 | entry = s._cache.alloc(mm_hash, 1) |
| 130 | assert entry is not None |
| 131 | fake.started.append(mm_hash) |
| 132 | return True |
| 133 | |
| 134 | monkeypatch.setattr(s, "_start_xfer", _fake_start) |
| 135 | req = _Request([_Feature("h1", 1)], params=_params("h1", 1)) |
| 136 | |
| 137 | # Step 1: unseen remote item -> read started, request deferred. |
| 138 | assert s.ensure_cache_available(req, 0) is False |
| 139 | assert "h1" in s._in_flight |
| 140 | assert "h1" in fake.started |
| 141 | s.build_connector_meta(scheduler_output=None) # re-arms _first_in_batch |
| 142 | |
| 143 | # Session reports completion this step. |
| 144 | fake._results.completed.add("h1") |
| 145 | s._sessions[("h", 1)] = fake |
| 146 | |
| 147 | # Step 2: poll drains completion -> entry marked ready, request admitted. |
| 148 | assert s.ensure_cache_available(req, 0) is True |
| 149 | assert "h1" not in s._in_flight |
| 150 | assert "h1" in s._step_completed |
| 151 | entry = s._cache.get("h1") |
| 152 | assert entry is not None and entry.ready |
| 153 | # The scheduler now allocates encoder input; the ready entry loads through |
| 154 | # the same local path as a natively cached encoding. |
| 155 | s.update_state_after_alloc(req, 0) |
| 156 | meta = s.build_connector_meta(scheduler_output=None) |
| 157 | assert "h1" in meta.loads |
| 158 | assert "h1" not in s._step_completed # cleared by promote |
| 159 | |
| 160 | # Step 3: still cached -> admitted directly, no new transfer. |
| 161 | assert s.ensure_cache_available(req, 0) is True |
| 162 | s.shutdown() |
| 163 | |
| 164 | |
| 165 | @pytest.mark.parametrize("params", [None, {"x": 123}]) |
| 166 | def test_no_remote_announcement_is_ready(monkeypatch, params): |
| 167 | s = _consumer_sched(monkeypatch) |
| 168 | assert s.ensure_cache_available(_Request([_Feature("x")], params=params), 0) |
| 169 | assert s.take_unavailable_requests() == set() |
| 170 | s.shutdown() |
| 171 | |
| 172 | |
| 173 | @pytest.mark.parametrize("payload", [None, "metadata", "pixels", "embeds", "shm"]) |
| 174 | def test_failed_read_falls_back_only_with_local_input(monkeypatch, payload): |
| 175 | """A remote miss is fatal only when local model input is unavailable.""" |
| 176 | s = _consumer_sched(monkeypatch) |
| 177 | fake = _FakeSession() |
| 178 | |
| 179 | def _fake_start(mm_hash, info, size): |
| 180 | entry = s._cache.alloc(mm_hash, 1) |
| 181 | assert entry is not None |
| 182 | fake.started.append(mm_hash) |
| 183 | return True |
| 184 | |
| 185 | monkeypatch.setattr(s, "_start_xfer", _fake_start) |
| 186 | fields = { |
| 187 | "metadata": {"image_grid_thw": torch.tensor([1, 2, 2])}, |
| 188 | "pixels": {"pixel_values": torch.ones(1, 3)}, |
| 189 | "embeds": {"image_embeds": torch.ones(1, _HID)}, |
| 190 | "shm": {"address": 123, "monotonic_id": 1}, |
| 191 | } |
| 192 | data = ( |
| 193 | MultiModalKwargsItem( |
| 194 | { |
| 195 | key: MultiModalFieldElem(value, MultiModalSharedField(batch_size=1)) |
| 196 | for key, value in fields[payload].items() |
| 197 | } |
| 198 | ) |
| 199 | if payload is not None |
| 200 | else None |
| 201 | ) |
| 202 | req = _Request([_Feature("h1", 1, data)], params=_params("h1", 1)) |
| 203 | |
| 204 | # Step 1: start the read. |
| 205 | assert s.ensure_cache_available(req, 0) is False |
| 206 | s.build_connector_meta(scheduler_output=None) |
| 207 | |
| 208 | fake._results.tombstoned.add("h1") |
| 209 | s._sessions[("h", 1)] = fake |
| 210 | |
| 211 | can_fallback = payload in ("pixels", "embeds", "shm") |
| 212 | assert s.ensure_cache_available(req, 0) is can_fallback |
| 213 | assert s._cache.get("h1") is None |
| 214 | assert "h1" not in s._in_flight |
| 215 | assert s.take_unavailable_requests() == (set() if can_fallback else {"r1"}) |
| 216 | # Drained by the read, so the scheduler cannot abort it twice. |
| 217 | assert s.take_unavailable_requests() == set() |
| 218 | if can_fallback: |
| 219 | # A later scheduling step must not retry the failed remote source. |
| 220 | s.build_connector_meta(scheduler_output=None) |
| 221 | assert s.ensure_cache_available(req, 0) |
| 222 | assert fake.started == ["h1"] |
| 223 | s.shutdown() |
| 224 | |
| 225 | |
| 226 | @pytest.mark.parametrize("retry", [False, True], ids=["in-flight", "retryable"]) |
| 227 | def test_remote_wait_budget_survives_retries_and_long_steps(monkeypatch, retry): |
| 228 | """Neither retries nor delayed scheduling may restart the request budget.""" |
| 229 | s = _consumer_sched(monkeypatch) |
| 230 | fake = _FakeSession() |
| 231 | s._sessions[("h", 1)] = fake |
| 232 | |
| 233 | def start(mm_hash, info, size): |
| 234 | assert s._cache.alloc(mm_hash, 1) is not None |
| 235 | fake.started.append(mm_hash) |
| 236 | return True |
| 237 | |
| 238 | monkeypatch.setattr(s, "_start_xfer", start) |
| 239 | now = [1000.0] |
| 240 | monkeypatch.setattr(time, "monotonic", lambda: now[0]) |
| 241 | req = _Request([_Feature("h1")], params=_params("h1", 1)) |
| 242 | assert not s.ensure_cache_available(req, 0) |
| 243 | for elapsed in (20, 40, 121): |
| 244 | now[0] = 1000.0 + elapsed |
| 245 | if retry: |
| 246 | fake._results.retryable.add("h1") |
| 247 | s.build_connector_meta(scheduler_output=None) |
| 248 | assert not s.ensure_cache_available(req, 0) |
| 249 | assert s.take_unavailable_requests() == ({"r1"} if elapsed > 60 else set()) |
| 250 | assert len(fake.started) == (3 if retry else 1) |
| 251 | s.shutdown() |
| 252 | |
| 253 | |
| 254 | def test_retryable_read_re_requests_without_admitting(monkeypatch): |
| 255 | """A not-ready NACK defers the request and re-requests the read. |
| 256 | |
| 257 | Unlike a tombstone it never admits the request, because an admitted request |
| 258 | whose media was rewritten away upstream has no embedding and no way to |
| 259 | recompute one. |
| 260 | """ |
| 261 | s = _consumer_sched(monkeypatch) |
| 262 | fake = _FakeSession() |
| 263 | |
| 264 | def _fake_start(mm_hash, info, size): |
| 265 | entry = s._cache.alloc(mm_hash, 1) |
| 266 | assert entry is not None |
| 267 | fake.started.append(mm_hash) |
| 268 | return True |
| 269 | |
| 270 | monkeypatch.setattr(s, "_start_xfer", _fake_start) |
| 271 | req = _Request([_Feature("h1", 1)], params=_params("h1", 1)) |
| 272 | |
| 273 | assert s.ensure_cache_available(req, 0) is False |
| 274 | s.build_connector_meta(scheduler_output=None) |
| 275 | |
| 276 | # Producer's save has not landed: retryable, not tombstoned. |
| 277 | fake._results.retryable.add("h1") |
| 278 | s._sessions[("h", 1)] = fake |
| 279 | |
| 280 | # The entry is released and the read re-issued in the same step, so the |
| 281 | # request stays deferred rather than being admitted unfetched. |
| 282 | assert s.ensure_cache_available(req, 0) is False |
| 283 | assert "h1" in s._in_flight |
| 284 | assert "h1" not in s._tombstones |
| 285 | assert fake.started.count("h1") == 2 |
| 286 | s.shutdown() |
| 287 | |
| 288 | |
| 289 | @pytest.mark.parametrize( |
| 290 | "size_fields", |
| 291 | [{"size_bytes": 999}, {"size_bytes": None}, {"size_bytes": "big"}, {}], |
| 292 | ids=["mismatch", "null-size", "text-size", "no-size"], |
| 293 | ) |
| 294 | def test_invalid_remote_size_fails_the_request(monkeypatch, size_fields): |
| 295 | """Invalid remote sizes fail the request without raising in the scheduler.""" |
| 296 | s = _consumer_sched(monkeypatch) |
| 297 | announced = {"peer_host": "h", "peer_port": 1, **size_fields} |
| 298 | req = _Request([_Feature("h1", 1)], params={"h1": announced}) |
| 299 | assert s.ensure_cache_available(req, 0) is False |
| 300 | assert s.take_unavailable_requests() == {"r1"} |
| 301 | assert "h1" not in s._in_flight |
| 302 | assert s._cache.get("h1") is None |
| 303 | s.shutdown() |
| 304 | |
| 305 | |
| 306 | def test_deferral_budget_is_per_request(monkeypatch): |
| 307 | """Shared hashes have independent wait budgets and cancellation cleanup.""" |
| 308 | s = _consumer_sched(monkeypatch) |
| 309 | monkeypatch.setattr(s._cache, "alloc", lambda key, n: None) |
| 310 | params = _params("h1", 1) |
| 311 | old = _Request([_Feature("h1", 1)], params=params, req_id="old") |
| 312 | new = _Request([_Feature("h1", 1)], params=params, req_id="new") |
| 313 | |
| 314 | now = [1000.0] |
| 315 | monkeypatch.setattr(time, "monotonic", lambda: now[0]) |
| 316 | assert s.ensure_cache_available(old, 0) is False |
| 317 | now[0] += sched_mod._ADMIT_DEFER_TIMEOUT_S - 1 |
| 318 | assert s.ensure_cache_available(new, 0) is False |
| 319 | assert s.take_unavailable_requests() == set() |
| 320 | |
| 321 | now[0] += 2 |
| 322 | # `old` is now past its budget; `new` is not. |
| 323 | assert s.ensure_cache_available(new, 0) is False |
| 324 | assert s.take_unavailable_requests() == set() |
| 325 | assert s.request_finished(new) == (False, None) |
| 326 | assert set(s._deferred_since) == {("old", "h1")} |
| 327 | assert s.ensure_cache_available(old, 0) is False |
| 328 | assert s.take_unavailable_requests() == {"old"} |
| 329 | assert "h1" not in s._in_flight |
| 330 | s.shutdown() |
| 331 | |
| 332 | |
| 333 | def test_metadata_only_entry_admits_without_transfer(monkeypatch): |
| 334 | # A producer that reported placeholder metadata but had no cache entry |
| 335 | # to transfer omits peer_host/peer_port/size_bytes entirely; the consumer |
| 336 | # must fall back to local compute rather than treating that as a size |
| 337 | # mismatch. |
| 338 | s = _consumer_sched(monkeypatch) |
| 339 | metadata_only = {"h1": {"image_grid_thw": [1, 2, 3]}} |
| 340 | req = _Request([_Feature("h1", 1)], params=metadata_only) |
| 341 | assert s.ensure_cache_available(req, 0) is True |
| 342 | assert "h1" not in s._in_flight |
| 343 | assert s._cache.get("h1") is None |
| 344 | s.shutdown() |
| 345 | |
| 346 | |
| 347 | def test_orphan_not_ready_entry_defers_no_realloc(monkeypatch): |
| 348 | # Post-quarantine, post-tombstone-consumed orphan: a not-ready cache |
| 349 | # entry exists for the mm_hash, but the hash is tracked in none of |
| 350 | # _in_flight / _step_completed / _tombstones. Its blocks are held by a |
| 351 | # quarantined/settling DMA and must not be re-allocated. |
| 352 | s = _consumer_sched(monkeypatch) |
| 353 | entry = s._cache.alloc("h1", 1) |
| 354 | assert entry is not None and not entry.ready |
| 355 | assert "h1" not in s._in_flight |
| 356 | assert "h1" not in s._step_completed |
| 357 | assert "h1" not in s._tombstones |
| 358 | |
| 359 | # Spy on alloc: record calls, delegate to the real allocator (which |
| 360 | # asserts on a duplicate key). A re-alloc for "h1" is the bug. |
| 361 | real_alloc = s._cache.alloc |
| 362 | calls: list[str] = [] |
| 363 | |
| 364 | def _spy_alloc(key, n): |
| 365 | calls.append(key) |
| 366 | return real_alloc(key, n) |
| 367 | |
| 368 | monkeypatch.setattr(s._cache, "alloc", _spy_alloc) |
| 369 | req = _Request([_Feature("h1", 1)], params=_params("h1", 1)) |
| 370 | |
| 371 | # Must not raise; the request is deferred while the DMA settles, the |
| 372 | # orphan entry is untouched, and alloc is never called again. |
| 373 | assert s.ensure_cache_available(req, 0) is False |
| 374 | assert calls == [] |
| 375 | assert s._cache.get("h1") is entry |
| 376 | assert not entry.ready |
| 377 | assert s.take_unavailable_requests() == set() |
| 378 | s.shutdown() |
| 379 | |
| 380 | |
| 381 | def test_already_computed_feature_admits_without_transfer(monkeypatch): |
| 382 | # A feature entirely within num_computed_tokens is skipped by admit; no |
| 383 | # cache lookup/alloc or transfer should be triggered for it. |
| 384 | s = _consumer_sched(monkeypatch) |
| 385 | calls: list[str] = [] |
| 386 | real_alloc = s._cache.alloc |
| 387 | |
| 388 | def _spy_alloc(key, n): |
| 389 | calls.append(key) |
| 390 | return real_alloc(key, n) |
| 391 | |
| 392 | monkeypatch.setattr(s._cache, "alloc", _spy_alloc) |
| 393 | # Feature spans [0, 4); with num_computed_tokens=4 it is fully computed. |
| 394 | req = _Request([_Feature("h1", 4)], params=_params("h1", 4)) |
| 395 | |
| 396 | assert s.ensure_cache_available(req, 4) is True |
| 397 | assert "h1" not in s._in_flight |
| 398 | assert calls == [] |
| 399 | s.shutdown() |
| 400 | |
| 401 | |
| 402 | def test_local_ready_entry_admits_without_transfer(monkeypatch): |
| 403 | # A READY entry already present in the cache is a local hit: admit |
| 404 | # returns True and no new transfer is started for that mm_hash. |
| 405 | s = _consumer_sched(monkeypatch) |
| 406 | entry = s._cache.alloc("h1", 1) |
| 407 | assert entry is not None |
| 408 | s._cache.mark_ready("h1") |
| 409 | req = _Request([_Feature("h1", 1)], params=_params("h1", 1)) |
| 410 | |
| 411 | assert s.ensure_cache_available(req, 0) is True |
| 412 | assert "h1" not in s._in_flight |
| 413 | s.shutdown() |
| 414 | |
| 415 | |
| 416 | def test_in_flight_hash_defers_without_second_transfer(monkeypatch): |
| 417 | # A mm_hash already tracked in _in_flight must defer the request rather |
| 418 | # than starting a duplicate transfer. |
| 419 | s = _consumer_sched(monkeypatch) |
| 420 | started: list[str] = [] |
| 421 | |
| 422 | def _spy_start_xfer(mm_hash, info, size): |
| 423 | started.append(mm_hash) |
| 424 | return True |
| 425 | |
| 426 | monkeypatch.setattr(s, "_start_xfer", _spy_start_xfer) |
| 427 | s._in_flight.add("h1") |
| 428 | req = _Request([_Feature("h1", 1)], params=_params("h1", 1)) |
| 429 | |
| 430 | assert s.ensure_cache_available(req, 0) is False |
| 431 | assert started == [] |
| 432 | s.shutdown() |