| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | """Consumer-side unit tests for the sharded-RDT backend. |
| 4 | |
| 5 | The consumer engine's planning core is pure — it turns the baked plan plus the |
| 6 | driver's gather-group partition into a static `_CallPlan` with no pulls, no |
| 7 | model and no Ray — so most of it is exercised here on CPU/meta tensors only. |
| 8 | These tests pin the values the engine produces today (chunk boundaries, packed |
| 9 | byte offsets, recorded op chains) so a refactor that changes them fails loudly |
| 10 | instead of silently shipping different bytes. |
| 11 | |
| 12 | `TestBakeOnARealModel` at the end is the exception: it needs a GPU, because the |
| 13 | bake only exists to be driven through a real `model.load_weights`, and the parts |
| 14 | that matter there — that the monkeypatched loaders are put back, and that the |
| 15 | recorded op chains reproduce what the loaders did — cannot be faked. |
| 16 | |
| 17 | `tests/distributed/test_sharded_rdt_trainer.py` covers the trainer (producer) |
| 18 | side; `test_sharded_rdt_producer.py` covers its serve-actor protocol. |
| 19 | """ |
| 20 | |
| 21 | from types import SimpleNamespace |
| 22 | |
| 23 | import pytest |
| 24 | import torch |
| 25 | |
| 26 | from vllm.distributed.weight_transfer.base import layerwise_groups |
| 27 | from vllm.distributed.weight_transfer.sharded_rdt_common import ( |
| 28 | ALLOWED_OPS, |
| 29 | SUPPORTED_OPS, |
| 30 | RdtRouter, |
| 31 | assign_producer_indices, |
| 32 | buffer_alloc_bytes, |
| 33 | ) |
| 34 | from vllm.distributed.weight_transfer.sharded_rdt_engine import ( |
| 35 | ShardedRDTWeightTransferEngine, |
| 36 | ShardedRDTWeightTransferInitInfo, |
| 37 | _dtype_from_name, |
| 38 | ) |
| 39 | from vllm.distributed.weight_transfer.sharded_rdt_fake import ( |
| 40 | BakeSink, |
| 41 | FakeRDTTensor, |
| 42 | _Scatter, |
| 43 | _UnsupportedFakeOp, |
| 44 | ) |
| 45 | |
| 46 | META = torch.device("meta") |
| 47 | |
| 48 | |
| 49 | # --------------------------------------------------------------------------- |
| 50 | # Harness |
| 51 | # --------------------------------------------------------------------------- |
| 52 | |
| 53 | |
| 54 | class _FakeLayer: |
| 55 | """Stand-in for a leaf module. The planner only stores it and keys on |
| 56 | ``id()``; nothing in the plan path touches its attributes.""" |
| 57 | |
| 58 | def __init__(self, tag: str) -> None: |
| 59 | self.tag = tag |
| 60 | |
| 61 | def __repr__(self) -> str: |
| 62 | return f"_FakeLayer({self.tag!r})" |
| 63 | |
| 64 | |
| 65 | def _module(layer, copies): |
| 66 | """What the bake records for one leaf module: a scatter list whose entries |
| 67 | all carry that module as their ``layer``.""" |
| 68 | for c in copies: |
| 69 | c.layer = layer |
| 70 | return copies |
| 71 | |
| 72 | |
| 73 | def _copy( |
| 74 | name, |
| 75 | layer_param="weight", |
| 76 | *, |
| 77 | offset=0, |
| 78 | shape=(4,), |
| 79 | ops=(), |
| 80 | layer=None, |
| 81 | dtype=torch.bfloat16, |
| 82 | ): |
| 83 | """A `_Scatter` as the bake would record it: source key, owning layer, and |
| 84 | the meta destination region.""" |
| 85 | stride = ( |
| 86 | (1,) |
| 87 | if len(shape) == 1 |
| 88 | else tuple( |
| 89 | int(torch.empty(shape, device=META).stride()[i]) for i in range(len(shape)) |
| 90 | ) |
| 91 | ) |
| 92 | return _Scatter( |
| 93 | layer=layer, |
| 94 | param_name=layer_param, |
| 95 | src=(name, tuple(ops)), |
| 96 | offset=offset, |
| 97 | shape=tuple(shape), |
| 98 | stride=stride, |
| 99 | dtype=dtype, |
| 100 | ) |
| 101 | |
| 102 | |
| 103 | def _planner(baked, *, name_meta, live=None, held_by=None, group_lens=None): |
| 104 | """A planning-only engine. |
| 105 | |
| 106 | `__init__` needs a config/model/device the planner never reads, so build the |
| 107 | instance without it and set exactly the state `_build_call_plan` consumes: |
| 108 | the baked plan, the name metadata, the live-name set, and a bound router. |
| 109 | |
| 110 | ``held_by``: name -> the ranks holding it. Names left out are held by every |
| 111 | producer (the replicated class 0). Default: one producer holding everything, |
| 112 | the degenerate router. Stand-in ranks 0-3 let plan tests assert chunk |
| 113 | structure; owner VALUES are pinned by TestCallPlanRouting. |
| 114 | """ |
| 115 | names = list(name_meta) |
| 116 | if held_by: |
| 117 | n_prod = max(r for owners in held_by.values() for r in owners) + 1 |
| 118 | owner_sets = [list(range(n_prod))] |
| 119 | seen = {tuple(range(n_prod)): 0} |
| 120 | classes = [] |
| 121 | for n in names: |
| 122 | key = tuple(sorted(held_by.get(n, range(n_prod)))) |
| 123 | if key not in seen: |
| 124 | seen[key] = len(owner_sets) |
| 125 | owner_sets.append(list(key)) |
| 126 | classes.append(seen[key]) |
| 127 | else: |
| 128 | n_prod, owner_sets, classes = 1, [[0]], [0] * len(names) |
| 129 | eng = object.__new__(ShardedRDTWeightTransferEngine) |
| 130 | eng._name_to_plan = dict(baked) |
| 131 | eng._name_meta = dict(name_meta) |
| 132 | eng._live_names = set(live or name_meta) |
| 133 | router = RdtRouter( |
| 134 | n_prod, 1, owner_sets, classes, names, group_lens or [1] * len(names) |
| 135 | ) |
| 136 | router.bind([object()] * n_prod, [object()] * n_prod, 0) |
| 137 | eng._router = router |
| 138 | return eng |
| 139 | |
| 140 | |
| 141 | def _one_module_per_layer(n_layers, *, dtype="bfloat16", numel=4): |
| 142 | """`(baked, name_meta, names, group_lens)` for a pre / N-layer / post model |
| 143 | where every gather group is one fully-baked leaf module.""" |
| 144 | names = ["embed.weight"] |
| 145 | names += [f"model.layers.{i}.w" for i in range(n_layers)] |
| 146 | names += ["norm.weight"] |
| 147 | baked = {} |
| 148 | for n in names: |
| 149 | layer = _FakeLayer(n) |
| 150 | baked[n] = _module(layer, [_copy(n, shape=(numel,))]) |
| 151 | name_meta = {n: (dtype, [numel]) for n in names} |
| 152 | group_lens = [1] * len(names) |
| 153 | return baked, name_meta, names, group_lens |
| 154 | |
| 155 | |
| 156 | # --------------------------------------------------------------------------- |
| 157 | # FakeRDTTensor: op-chain recording |
| 158 | # --------------------------------------------------------------------------- |
| 159 | |
| 160 | |
| 161 | class TestFakeOpChains: |
| 162 | """Every allowlisted op must append itself to the chain and hand back a |
| 163 | child whose shape/dtype PyTorch itself computed. The chain is the wire |
| 164 | format the producer replays, so its exact contents are load-bearing.""" |
| 165 | |
| 166 | def _fake(self, shape=(4, 6), dtype=torch.bfloat16, sink=None): |
| 167 | return FakeRDTTensor( |
| 168 | name="w", shape=torch.Size(shape), dtype=dtype, device=META, sink=sink |
| 169 | ) |
| 170 | |
| 171 | def test_bare_fake_has_metadata_but_no_chain(self): |
| 172 | t = self._fake() |
| 173 | assert t.shape == (4, 6) |
| 174 | assert t.dtype is torch.bfloat16 |
| 175 | assert t._key() == ("w", ()) |
| 176 | |
| 177 | @pytest.mark.parametrize( |
| 178 | "call,expect_op,expect_shape", |
| 179 | [ |
| 180 | (lambda t: t.narrow(0, 1, 2), ("narrow", (0, 1, 2), ()), (2, 6)), |
| 181 | (lambda t: t.view(6, 4), ("view", (6, 4), ()), (6, 4)), |
| 182 | (lambda t: t.reshape(24), ("reshape", (24,), ()), (24,)), |
| 183 | (lambda t: t[1], ("__getitem__", (1,), ()), (6,)), |
| 184 | (lambda t: t.unsqueeze(0), ("unsqueeze", (0,), ()), (1, 4, 6)), |
| 185 | (lambda t: t.transpose(0, 1), ("transpose", (0, 1), ()), (6, 4)), |
| 186 | (lambda t: t.t(), ("t", (), ()), (6, 4)), |
| 187 | (lambda t: t.permute(1, 0), ("permute", (1, 0), ()), (6, 4)), |
| 188 | (lambda t: t.flatten(), ("flatten", (), ()), (24,)), |
| 189 | (lambda t: t.contiguous(), ("contiguous", (), ()), (4, 6)), |
| 190 | ], |
| 191 | ) |
| 192 | def test_single_return_op_appends_one_spec(self, call, expect_op, expect_shape): |
| 193 | child = call(self._fake()) |
| 194 | assert isinstance(child, FakeRDTTensor) |
| 195 | assert child._key() == ("w", (expect_op,)) |
| 196 | assert tuple(child.shape) == expect_shape |
| 197 | |
| 198 | def test_squeeze_records_its_argument(self): |
| 199 | child = self._fake(shape=(1, 4)).squeeze(0) |
| 200 | assert child._key() == ("w", (("squeeze", (0,), ()),)) |
| 201 | assert tuple(child.shape) == (4,) |
| 202 | |
| 203 | def test_chains_compose_in_call_order(self): |
| 204 | child = self._fake().t().narrow(0, 0, 3).flatten() |
| 205 | assert child._key() == ( |
| 206 | "w", |
| 207 | ( |
| 208 | ("t", (), ()), |
| 209 | ("narrow", (0, 0, 3), ()), |
| 210 | ("flatten", (), ()), |
| 211 | ), |
| 212 | ) |
| 213 | assert tuple(child.shape) == (12,) |
| 214 | |
| 215 | def test_kwargs_are_frozen_sorted_for_hashability(self): |
| 216 | child = self._fake().narrow(dim=0, start=1, length=2) |
| 217 | (op, args, kwargs) = child._key()[1][0] |
| 218 | assert (op, args) == ("narrow", ()) |
| 219 | assert kwargs == (("dim", 0), ("length", 2), ("start", 1)) |
| 220 | hash(child._key()) # the chain is used as a dict key on both sides |
| 221 | |
| 222 | @pytest.mark.parametrize("op,n", [("chunk", 2), ("unbind", 4)]) |
| 223 | def test_multi_return_op_emits_one_child_per_output(self, op, n): |
| 224 | """chunk/unbind hand back a tuple; each child carries the base op plus a |
| 225 | trailing __getitem__(i) so the producer can index the replayed result.""" |
| 226 | parts = getattr(self._fake(), op)(*((n,) if op == "chunk" else ()), 0) |
| 227 | assert isinstance(parts, tuple) |
| 228 | assert len(parts) == n |
| 229 | for i, part in enumerate(parts): |
| 230 | base, index = part._key()[1] |
| 231 | assert base[0] == op |
| 232 | assert index == ("__getitem__", (i,), ()) |
| 233 | |
| 234 | def test_op_chain_is_hashable_as_a_fetch_key(self): |
| 235 | keys = {self._fake().t()._key(), self._fake().t()._key()} |
| 236 | assert len(keys) == 1, "equal chains must collapse — they dedup pull keys" |
| 237 | |
| 238 | |
| 239 | class TestFakeUnsupportedOps: |
| 240 | """Anything that needs real data must fail loudly at bake time rather than |
| 241 | silently transferring the wrong bytes.""" |
| 242 | |
| 243 | def _fake(self): |
| 244 | return FakeRDTTensor( |
| 245 | name="w", |
| 246 | shape=torch.Size((4,)), |
| 247 | dtype=torch.bfloat16, |
| 248 | device=META, |
| 249 | sink=None, |
| 250 | ) |
| 251 | |
| 252 | @pytest.mark.parametrize( |
| 253 | "call", |
| 254 | [ |
| 255 | lambda t: t.to(torch.float32), |
| 256 | lambda t: t.float(), |
| 257 | lambda t: t + 1, |
| 258 | lambda t: t * 2, |
| 259 | lambda t: t.sum(), |
| 260 | ], |
| 261 | ids=["to", "float", "add", "mul", "sum"], |
| 262 | ) |
| 263 | def test_data_dependent_ops_raise(self, call): |
| 264 | with pytest.raises(_UnsupportedFakeOp): |
| 265 | call(self._fake()) |
| 266 | |
| 267 | def test_error_names_the_weight_and_the_chain(self): |
| 268 | with pytest.raises(_UnsupportedFakeOp) as exc: |
| 269 | self._fake().narrow(0, 0, 2).float() |
| 270 | msg = str(exc.value) |
| 271 | assert "'w'" in msg |
| 272 | assert "narrow" in msg |
| 273 | |
| 274 | def test_unsupported_is_a_notimplementederror(self): |
| 275 | """Callers distinguish "this backend can't handle the loader" from bugs.""" |
| 276 | assert issubclass(_UnsupportedFakeOp, NotImplementedError) |
| 277 | |
| 278 | |
| 279 | class TestBakeRecording: |
| 280 | """During the dry run the fake's ``copy_`` is the data sink: it records the |
| 281 | source chain plus the meta destination's strided region and moves nothing.""" |
| 282 | |
| 283 | def _recorder_and_fake(self, shape=(4, 6)): |
| 284 | rec = BakeSink() |
| 285 | fake = FakeRDTTensor( |
| 286 | name="w", |
| 287 | shape=torch.Size(shape), |
| 288 | dtype=torch.bfloat16, |
| 289 | device=META, |
| 290 | sink=rec, |
| 291 | ) |
| 292 | return rec, fake |
| 293 | |
| 294 | def test_copy_records_the_destination_region(self): |
| 295 | rec, fake = self._recorder_and_fake() |
| 296 | layer = _FakeLayer("q_proj") |
| 297 | param = torch.empty((8, 6), dtype=torch.bfloat16, device=META) |
| 298 | dest = param.narrow(0, 4, 4) # the second half of a fused param |
| 299 | rec.current = (layer, "weight") |
| 300 | dest.copy_(fake) |
| 301 | |
| 302 | (recorded,) = rec.copies_by_layer[layer] |
| 303 | assert recorded.src == ("w", ()) |
| 304 | assert recorded.param_name == "weight" |
| 305 | assert recorded.offset == dest.storage_offset() == 24 |
| 306 | assert recorded.shape == (4, 6) |
| 307 | assert recorded.stride == (6, 1) |
| 308 | |
| 309 | def test_copy_marks_the_source_name_live(self): |
| 310 | rec, fake = self._recorder_and_fake() |
| 311 | rec.current = (_FakeLayer("l"), "weight") |
| 312 | torch.empty((4, 6), dtype=torch.bfloat16, device=META).copy_(fake) |
| 313 | assert rec.copied_names == {"w"} |
| 314 | |
| 315 | def test_unattributed_copy_is_live_but_unrecorded(self): |
| 316 | """A copy_ with no loader stamp cannot be attributed to a param, so its |
| 317 | module must fall back to the plain load — but the name still moved data.""" |
| 318 | rec, fake = self._recorder_and_fake() |
| 319 | rec.current = None |
| 320 | torch.empty((4, 6), dtype=torch.bfloat16, device=META).copy_(fake) |
| 321 | assert rec.copied_names == {"w"} |
| 322 | assert dict(rec.copies_by_layer) == {} |
| 323 | |
| 324 | def test_copies_are_grouped_by_module_in_call_order(self): |
| 325 | rec, _ = self._recorder_and_fake() |
| 326 | layer = _FakeLayer("gate_up") |
| 327 | for i, name in enumerate(("gate", "up")): |
| 328 | fake = FakeRDTTensor( |
| 329 | name=name, |
| 330 | shape=torch.Size((4,)), |
| 331 | dtype=torch.bfloat16, |
| 332 | device=META, |
| 333 | sink=rec, |
| 334 | ) |
| 335 | param = torch.empty((8,), dtype=torch.bfloat16, device=META) |
| 336 | rec.current = (layer, "weight") |
| 337 | param.narrow(0, 4 * i, 4).copy_(fake) |
| 338 | assert [c.src[0] for c in rec.copies_by_layer[layer]] == ["gate", "up"] |
| 339 | assert [c.offset for c in rec.copies_by_layer[layer]] == [0, 4] |
| 340 | |
| 341 | def test_recording_a_sliced_source(self): |
| 342 | """The chain on the source and the region on the dest are independent.""" |
| 343 | rec, fake = self._recorder_and_fake(shape=(8, 6)) |
| 344 | layer = _FakeLayer("k_proj") |
| 345 | rec.current = (layer, "weight") |
| 346 | param = torch.empty((4, 6), dtype=torch.bfloat16, device=META) |
| 347 | param.copy_(fake.narrow(0, 2, 4)) |
| 348 | (recorded,) = rec.copies_by_layer[layer] |
| 349 | assert recorded.src == ("w", (("narrow", (0, 2, 4), ()),)) |
| 350 | assert recorded.shape == (4, 6) |
| 351 | |
| 352 | def test_the_recorded_dtype_is_the_chains_output_not_the_sources(self): |
| 353 | """``view(dtype)`` is allowlisted, so a chain can reinterpret dtype. The |
| 354 | producer packs what the replay yields, so the record must carry the |
| 355 | POST-chain dtype: taking it from the source name's metadata instead sizes |
| 356 | the slice with the wrong itemsize and shifts every later slice in the |
| 357 | chunk, carving the packed blob differently on the two sides.""" |
| 358 | rec, fake = self._recorder_and_fake(shape=(4,)) |
| 359 | layer = _FakeLayer("reinterpreted") |
| 360 | rec.current = (layer, "weight") |
| 361 | viewed = fake.view(torch.float32) # 4 x bf16 -> 2 x f32 |
| 362 | param = torch.empty((2,), dtype=torch.float32, device=META) |
| 363 | param.copy_(viewed) |
| 364 | |
| 365 | (recorded,) = rec.copies_by_layer[layer] |
| 366 | assert recorded.dtype is torch.float32, "recorded the source dtype" |
| 367 | assert recorded.shape == (2,) |
| 368 | # What the source name alone would have said, and what the producer |
| 369 | # would NOT have sent. |
| 370 | assert 2 * torch.float32.itemsize == 8 != 2 * torch.bfloat16.itemsize |
| 371 | |
| 372 | def test_a_broadcasting_copy_is_refused_at_bake(self): |
| 373 | """The packed layout sizes each slice from the DESTINATION shape while |
| 374 | the producer packs what the chain yields, and nothing downstream can |
| 375 | catch a mismatch — the consumer's buffer view is exactly prod(dest.shape) |
| 376 | elements, so it reshapes cleanly over bytes laid out at other offsets. |
| 377 | So the bake refuses rather than recording a slice it cannot carve.""" |
| 378 | rec, fake = self._recorder_and_fake(shape=(1, 6)) |
| 379 | rec.current = (_FakeLayer("broadcast"), "weight") |
| 380 | param = torch.empty((4, 6), dtype=torch.bfloat16, device=META) |
| 381 | with pytest.raises(_UnsupportedFakeOp, match="broadcasting"): |
| 382 | param.copy_(fake) |
| 383 | |
| 384 | |
| 385 | # --------------------------------------------------------------------------- |
| 386 | # Consumer identity |
| 387 | # --------------------------------------------------------------------------- |
| 388 | |
| 389 | |
| 390 | class TestConsumerIdentity: |
| 391 | """Every worker in the fleet needs a DISTINCT id in 0..C-1: it selects the |
| 392 | worker's producer block and keys the producer's per-consumer serve ring, so |
| 393 | a collision silently serves two workers out of one ring.""" |
| 394 | |
| 395 | def _engine(self, *, dp_index, rank, world_size): |
| 396 | eng = object.__new__(ShardedRDTWeightTransferEngine) |
| 397 | eng.parallel_config = SimpleNamespace( |
| 398 | data_parallel_index=dp_index, rank=rank, world_size=world_size |
| 399 | ) |
| 400 | return eng |
| 401 | |
| 402 | def _ids(self, *, num_consumers, workers, replica_rank=0, num_replicas=1): |
| 403 | """Consumer ids for a DP-only engine whose workers are ``dp_index``es.""" |
| 404 | info = ShardedRDTWeightTransferInitInfo( |
| 405 | num_consumers=num_consumers, |
| 406 | replica_rank=replica_rank, |
| 407 | num_replicas=num_replicas, |
| 408 | ) |
| 409 | out = [] |
| 410 | for dp_index in workers: |
| 411 | eng = self._engine(dp_index=dp_index, rank=0, world_size=1) |
| 412 | eng._num_consumers_override = num_consumers |
| 413 | out.append(eng._resolve_consumer_id(info)) |
| 414 | return out |
| 415 | |
| 416 | def test_one_engine_indexes_by_dp_and_tp(self): |
| 417 | """dense-via-TP and MoE-via-DP+EP both flatten to 0..C-1.""" |
| 418 | eng = self._engine(dp_index=0, rank=3, world_size=8) |
| 419 | eng._num_consumers_override = 8 |
| 420 | info = ShardedRDTWeightTransferInitInfo(num_consumers=8) |
| 421 | assert eng._resolve_consumer_id(info) == 3 |
| 422 | |
| 423 | eng = self._engine(dp_index=5, rank=0, world_size=1) |
| 424 | eng._num_consumers_override = 8 |
| 425 | assert eng._resolve_consumer_id(info) == 5 |
| 426 | |
| 427 | def test_independent_engines_offset_into_distinct_ranges(self): |
| 428 | """A fleet of separate engines restarts _global_worker_index at 0 in |
| 429 | each, so without the replica offset every engine would claim 0..w-1. |
| 430 | The driver sets replica_rank on the payload; these are the ids it buys.""" |
| 431 | ids = [ |
| 432 | self._ids(num_consumers=8, workers=range(4), replica_rank=r, num_replicas=2) |
| 433 | for r in (0, 1) |
| 434 | ] |
| 435 | assert ids == [[0, 1, 2, 3], [4, 5, 6, 7]] |
| 436 | assert len(set(ids[0] + ids[1])) == 8, "every worker distinct fleet-wide" |
| 437 | |
| 438 | def test_the_default_is_a_single_engine_with_no_offset(self): |
| 439 | assert self._ids(num_consumers=4, workers=range(4)) == [0, 1, 2, 3] |
| 440 | |
| 441 | |
| 442 | # --------------------------------------------------------------------------- |
| 443 | # RdtRouter: per-name ownership -> one producer per pull |
| 444 | # --------------------------------------------------------------------------- |
| 445 | |
| 446 | |
| 447 | def _router( |
| 448 | num_producers, |
| 449 | num_consumers, |
| 450 | owner_sets=None, |
| 451 | name_owner_class=None, |
| 452 | names=None, |
| 453 | group_lens=None, |
| 454 | ): |
| 455 | """A router over ``n`` single-name groups unless told otherwise, so a test |
| 456 | can state expectations in group indices.""" |
| 457 | if names is None: |
| 458 | n = len(name_owner_class or []) or (len(group_lens or []) or 1) |
| 459 | names = [f"g{i}" for i in range(n)] |
| 460 | if group_lens is None: |
| 461 | group_lens = [1] * len(names) |
| 462 | return RdtRouter( |
| 463 | num_producers, num_consumers, owner_sets, name_owner_class, names, group_lens |
| 464 | ) |
| 465 | |
| 466 | |
| 467 | class TestRdtRouter: |
| 468 | """Who serves each name. |
| 469 | |
| 470 | A wrong answer is not a wrong number but a hang or a loud misroute: a |
| 471 | consumer pulling from a producer that never gathered the name trips its |
| 472 | served-names guard. Freeing does NOT route this way — every consumer signals |
| 473 | every owner of a group and the producers count to the live total — so these |
| 474 | pin PULL routing, plus ``group_owners`` for the free fan-out. |
| 475 | """ |
| 476 | |
| 477 | def test_identity_when_fleets_match(self): |
| 478 | r = _router(8, 8, group_lens=[1] * 6) |
| 479 | assert all(r.producer_for(3, f"g{g}") == 3 for g in range(6)) |
| 480 | assert all(r.producer_for(c, "g0") == c for c in range(8)) |
| 481 | |
| 482 | def test_every_producer_nic_carries_traffic(self): |
| 483 | """16 producers / 8 consumers: the block rule spreads each consumer's |
| 484 | pulls over its block, alternating by group, so no producer sits idle. |
| 485 | Plain `consumer_id % len(owners)` would idle half of them.""" |
| 486 | r = _router(16, 8, group_lens=[1] * 95) |
| 487 | names = [f"g{i}" for i in range(95)] |
| 488 | for c in range(8): |
| 489 | block = assign_producer_indices(16, 8, c) |
| 490 | assert {r.producer_for(c, n) for n in names} == set(block) |
| 491 | assert [r.producer_for(0, f"g{g}") for g in range(4)] == [0, 1, 0, 1] |
| 492 | assert {r.producer_for(c, n) for c in range(8) for n in names} == set(range(16)) |
| 493 | |
| 494 | def test_fan_in_shares_one_producer(self): |
| 495 | r = _router(2, 8, group_lens=[1] * 5) |
| 496 | names = [f"g{i}" for i in range(5)] |
| 497 | assert [{r.producer_for(c, n) for n in names} for c in range(8)] == [ |
| 498 | {c // 4} for c in range(8) |
| 499 | ] |
| 500 | |
| 501 | def test_pipeline_stages_route_to_the_owning_stage(self): |
| 502 | """2 stages x 8 ranks: groups 0-2 on stage 0, 3-5 on stage 1. Two owner |
| 503 | sets, one per stage; each consumer reaches both.""" |
| 504 | r = _router( |
| 505 | 16, |
| 506 | 8, |
| 507 | owner_sets=[list(range(8)), list(range(8, 16))], |
| 508 | name_owner_class=[0, 0, 0, 1, 1, 1], |
| 509 | ) |
| 510 | for c in range(8): |
| 511 | picks = [r.producer_for(c, f"g{g}") for g in range(6)] |
| 512 | assert picks == [c] * 3 + [c + 8] * 3 |
| 513 | assert r.group_owners(0) == list(range(8)) |
| 514 | assert r.group_owners(3) == list(range(8, 16)) |
| 515 | r.validate() |
| 516 | |
| 517 | def test_expert_names_route_to_the_holding_ranks(self): |
| 518 | """A name held by only some ranks routes to those; a replicated name |
| 519 | keeps the full set.""" |
| 520 | r = RdtRouter( |
| 521 | 8, |
| 522 | 4, |
| 523 | [[0, 1, 2, 3], [2, 3], [4, 5, 6, 7], [6, 7]], |
| 524 | [0, 1, 2, 3], |
| 525 | ["repl0", "expert0", "repl1", "expert1"], |
| 526 | [2, 2], |
| 527 | ) |
| 528 | r.validate() |
| 529 | assert r.owners("expert0") == [2, 3] |
| 530 | assert r.owners("expert1") == [6, 7] |
| 531 | assert r.owners("repl1") == [4, 5, 6, 7] |
| 532 | for c in range(4): |
| 533 | assert r.producer_for(c, "expert0") in (2, 3) |
| 534 | assert r.producer_for(c, "repl1") in (4, 5, 6, 7) |
| 535 | |
| 536 | def test_consumers_spread_over_an_owner_set(self): |
| 537 | """Several ranks hold a name (its TP peers); the block rule must spread |
| 538 | consumers across them, not funnel through one.""" |
| 539 | r = _router( |
| 540 | 4, 4, owner_sets=[[0, 1]], name_owner_class=[0] * 6, group_lens=[1] * 6 |
| 541 | ) |
| 542 | served = {r.producer_for(c, f"g{g}") for c in range(4) for g in range(6)} |
| 543 | assert served == {0, 1} |
| 544 | |
| 545 | def test_a_group_frees_at_every_owner_of_any_of_its_names(self): |
| 546 | """The free barrier is per group, so its fan-out is the UNION over the |
| 547 | group's names: a rank holding none of them is not signalled.""" |
| 548 | r = RdtRouter(4, 2, [[0, 1], [2, 3]], [0, 1, 0], ["a", "b", "c"], [2, 1]) |
| 549 | assert r.group_owners(0) == [0, 1, 2, 3] |
| 550 | assert r.group_owners(1) == [0, 1] |
| 551 | |
| 552 | def test_an_unowned_name_raises(self): |
| 553 | r = RdtRouter(2, 2, [[]], [0], ["w"], [1]) |
| 554 | with pytest.raises(ValueError, match="has no owner"): |
| 555 | r.producer_for(0, "w") |
| 556 | |
| 557 | def test_validate_rejects_an_empty_owner_set(self): |
| 558 | with pytest.raises(ValueError, match="empty"): |
| 559 | RdtRouter( |
| 560 | 4, 2, [[0, 1], [], [2, 3]], [0, 1, 2], ["a", "b", "c"], [3] |
| 561 | ).validate() |
| 562 | |
| 563 | def test_validate_rejects_an_out_of_range_owner(self): |
| 564 | with pytest.raises(ValueError, match="out of range"): |
| 565 | RdtRouter(2, 2, [[0, 5]], [0], ["a"], [1]).validate() |
| 566 | |
| 567 | def test_validate_rejects_an_unresolvable_class(self): |
| 568 | with pytest.raises(ValueError, match="owner class"): |
| 569 | RdtRouter(4, 2, [[0, 1]], [0, 7], ["a", "b"], [2]).validate() |
| 570 | |
| 571 | |
| 572 | class TestReplicaOverlay: |
| 573 | """[RDT-SHARE-SLOTS] Several inference deployments carve the same blocks. |
| 574 | |
| 575 | The producer can serve R deployments out of ONE slot only if the R copies of |
| 576 | a worker meet on the same producer asking for the same bytes. Identical |
| 577 | plans give the second half for free; this is the first half. Carving over |
| 578 | the whole fleet instead spreads the multi-owner (replicated) names of |
| 579 | different deployments onto different producers, so a producer ends up |
| 580 | holding a ring per distinct worker index and nothing can merge. |
| 581 | """ |
| 582 | |
| 583 | def _fleet( |
| 584 | self, |
| 585 | num_producers, |
| 586 | workers, |
| 587 | replicas, |
| 588 | owner_sets=None, |
| 589 | name_owner_class=None, |
| 590 | n_names=6, |
| 591 | ): |
| 592 | names = [f"g{i}" for i in range(n_names)] |
| 593 | return ( |
| 594 | RdtRouter( |
| 595 | num_producers, |
| 596 | workers * replicas, |
| 597 | owner_sets, |
| 598 | name_owner_class, |
| 599 | names, |
| 600 | [1] * n_names, |
| 601 | workers_per_replica=workers, |
| 602 | ), |
| 603 | names, |
| 604 | ) |
| 605 | |
| 606 | def test_every_replica_of_a_worker_resolves_the_same_producer(self): |
| 607 | r, names = self._fleet(16, 8, 4) |
| 608 | for worker in range(8): |
| 609 | for name in names: |
| 610 | picks = {r.producer_for(worker + rep * 8, name) for rep in range(4)} |
| 611 | assert len(picks) == 1, ( |
| 612 | f"worker {worker} split across producers {picks}" |
| 613 | ) |
| 614 | |
| 615 | def test_adding_a_deployment_does_not_move_the_first_one(self): |
| 616 | """The strongest form: an existing consumer's routes are byte-identical |
| 617 | before and after another deployment is provisioned.""" |
| 618 | one, names = self._fleet(16, 8, 1) |
| 619 | two, _ = self._fleet(16, 8, 2) |
| 620 | for consumer_id in range(8): |
| 621 | assert [one.producer_for(consumer_id, n) for n in names] == [ |
| 622 | two.producer_for(consumer_id, n) for n in names |
| 623 | ] |
| 624 | |
| 625 | def test_one_deployment_is_the_historical_carve(self): |
| 626 | """Default (no width) and width == the whole fleet must agree with the |
| 627 | plain block rule, so a single-deployment fleet is untouched.""" |
| 628 | plain = _router(16, 8, group_lens=[1] * 6) |
| 629 | explicit, names = self._fleet(16, 8, 1) |
| 630 | for consumer_id in range(8): |
| 631 | block = assign_producer_indices(16, 8, consumer_id) |
| 632 | for name in names: |
| 633 | assert plain.producer_for(consumer_id, name) == explicit.producer_for( |
| 634 | consumer_id, name |
| 635 | ) |
| 636 | assert explicit.producer_for(consumer_id, name) in block |
| 637 | |
| 638 | def test_the_overlay_holds_for_partially_owned_names(self): |
| 639 | """Pipeline stages and experts too: two owner sets, one narrow. The |
| 640 | narrow one already forced agreement (one owner, no block to carve); the |
| 641 | wide one is what the carve decides.""" |
| 642 | r, names = self._fleet( |
| 643 | 8, |
| 644 | 4, |
| 645 | 2, |
| 646 | owner_sets=[[0, 1, 2, 3], [5]], |
| 647 | name_owner_class=[0, 1, 0, 1, 0, 1], |
| 648 | ) |
| 649 | for worker in range(4): |
| 650 | for name in names: |
| 651 | assert r.producer_for(worker, name) == r.producer_for(worker + 4, name) |
| 652 | assert {r.producer_for(c, "g1") for c in range(8)} == {5} |
| 653 | |
| 654 | def test_a_width_that_does_not_divide_the_fleet_raises(self): |
| 655 | """A uniform fleet is asserted upstream in ``get_world_size``; if it ever |
| 656 | is not, the overlay would map two workers of one deployment onto the same |
| 657 | block index, so refuse rather than serve the wrong bytes.""" |
| 658 | with pytest.raises(ValueError, match="does not divide"): |
| 659 | RdtRouter(4, 6, None, None, ["a"], [1], workers_per_replica=4) |
| 660 | |
| 661 | def test_pulls_still_carry_the_global_consumer_id(self): |
| 662 | """The carve uses the index WITHIN a deployment, but the wire must keep |
| 663 | the fleet-global id: it is what lets the producer tell the sharers of a |
| 664 | slot apart, and count their arrivals separately.""" |
| 665 | |
| 666 | class _Method: |
| 667 | def __init__(self): |
| 668 | self.calls = [] |
| 669 | |
| 670 | def remote(self, keys, consumer_id, seq): |
| 671 | self.calls.append((keys, consumer_id, seq)) |
| 672 | |
| 673 | r, _ = self._fleet(4, 2, 2) |
| 674 | methods = [_Method() for _ in range(4)] |
| 675 | r.bind([object()] * 4, methods, consumer_id=3) # replica 1, worker 1 |
| 676 | r.pull(2, ["k"], 7) |
| 677 | |
| 678 | assert methods[2].calls == [(["k"], 3, 7)] |
| 679 | |
| 680 | |
| 681 | # --------------------------------------------------------------------------- |
| 682 | # Packed layout: the cross-process invariant |
| 683 | # --------------------------------------------------------------------------- |
| 684 | |
| 685 | |
| 686 | def _producer_pack_offsets(slices): |
| 687 | """The producer's rule, transcribed from ``rdt_produce_weights_batched`` |
| 688 | (sharded_rdt_trainer.py): 16B-aligned offsets in specs order. |
| 689 | |
| 690 | Kept as an independent implementation on purpose: the consumer computing the |
| 691 | same offsets is the invariant that makes the packed blob readable, and this |
| 692 | is the only thing that catches the two rules drifting apart. A divergence is |
| 693 | silent at runtime — the bytes still arrive, they are just carved up wrong. |
| 694 | """ |
| 695 | pack_cur = 0 |
| 696 | offsets = [] |
| 697 | for numel, dtype in slices: |
| 698 | off = (pack_cur + 15) & ~15 |
| 699 | pack_cur = off + numel * dtype.itemsize |
| 700 | offsets.append(off) |
| 701 | return offsets, pack_cur |
| 702 | |
| 703 | |
| 704 | class TestPackedLayout: |
| 705 | def test_consumer_layout_matches_the_producer_rule(self): |
| 706 | """Mixed dtypes, sizes that do not land on 16B — i.e. the Kimi-style |
| 707 | group (fp8 weights + fp32 scales + bf16 norms) that exposes alignment.""" |
| 708 | specs = [ |
| 709 | ("w.fp8", "float8_e4m3fn", [17]), |
| 710 | ("w.scale", "float32", [3]), |
| 711 | ("w.norm", "bfloat16", [5]), |
| 712 | ("w.big", "bfloat16", [1024]), |
| 713 | ] |
| 714 | baked, name_meta = {}, {} |
| 715 | layer = _FakeLayer("mixed") |
| 716 | copies = [] |
| 717 | for name, dtype_name, shape in specs: |
| 718 | name_meta[name] = (dtype_name, shape) |
| 719 | copies.append( |
| 720 | _copy(name, shape=tuple(shape), dtype=_dtype_from_name(dtype_name)) |
| 721 | ) |
| 722 | mod_plan = _module(layer, copies) |
| 723 | for name, _d, _s in specs: |
| 724 | baked[name] = mod_plan |
| 725 | |
| 726 | eng = _planner(baked, name_meta=name_meta) |
| 727 | plan = eng._build_call_plan([n for n, _, _ in specs], [len(specs)]) |
| 728 | (chunk,) = plan.chunks |
| 729 | |
| 730 | want_offsets, want_bytes = _producer_pack_offsets( |
| 731 | [(s[0], _dtype_from_name(d)) for _n, d, s in specs] |
| 732 | ) |
| 733 | assert [off for off, _dt, _n, _shape in chunk.pack_layout] == want_offsets |
| 734 | assert chunk.pack_bytes == want_bytes |
| 735 | |
| 736 | def test_every_offset_is_16b_aligned(self): |
| 737 | specs = [(f"w{i}", "bfloat16", [i * 3 + 1]) for i in range(6)] |
| 738 | layer = _FakeLayer("l") |
| 739 | copies = [_copy(n, shape=tuple(s)) for n, _d, s in specs] |
| 740 | mod_plan = _module(layer, copies) |
| 741 | eng = _planner( |
| 742 | {n: mod_plan for n, _d, _s in specs}, |
| 743 | name_meta={n: (d, s) for n, d, s in specs}, |
| 744 | ) |
| 745 | (chunk,) = eng._build_call_plan([n for n, _, _ in specs], [len(specs)]).chunks |
| 746 | assert all(off % 16 == 0 for off, _dt, _n, _s in chunk.pack_layout) |
| 747 | |
| 748 | def test_layout_carries_dtype_numel_and_shape_for_the_view_carve(self): |
| 749 | baked, name_meta, names, group_lens = _one_module_per_layer(1, numel=6) |
| 750 | eng = _planner(baked, name_meta=name_meta) |
| 751 | chunk = eng._build_call_plan(names, group_lens).chunks[0] |
| 752 | (off, dt, numel, shape) = chunk.pack_layout[0] |
| 753 | assert (off, dt, numel, shape) == (0, torch.bfloat16, 6, (6,)) |
| 754 | |
| 755 | def test_duplicate_source_keys_are_packed_once(self): |
| 756 | """Two destinations fed by the SAME (name, chain) pull one slice.""" |
| 757 | layer = _FakeLayer("l") |
| 758 | shared = ("w", ()) |
| 759 | copies = [ |
| 760 | _copy("w", "a", shape=(4,)), |
| 761 | _copy("w", "b", shape=(4,)), |
| 762 | ] |
| 763 | assert copies[0].src == shared and copies[1].src == shared |
| 764 | eng = _planner( |
| 765 | {"w": _module(layer, copies)}, |
| 766 | name_meta={"w": ("bfloat16", [4])}, |
| 767 | ) |
| 768 | (chunk,) = eng._build_call_plan(["w"], [1]).chunks |
| 769 | assert len(chunk.scatters) == 2 |
| 770 | assert chunk.keys == [shared] |
| 771 | assert len(chunk.pack_layout) == 1 |
| 772 | |
| 773 | |
| 774 | # --------------------------------------------------------------------------- |
| 775 | # Chunk splitting |
| 776 | # --------------------------------------------------------------------------- |
| 777 | |
| 778 | |
| 779 | class TestChunkModuleScatters: |
| 780 | """The chunk cut: one chunk per distinct owner class present in the copies, |
| 781 | ascending by class index. Derived purely from the bake plus the ownership |
| 782 | table, so vLLM's expert placement needs no cases.""" |
| 783 | |
| 784 | def test_unstamped_copies_form_one_chunk(self): |
| 785 | layer = _FakeLayer("l") |
| 786 | copies = [_copy(f"w{i}", shape=(4,)) for i in range(5)] |
| 787 | eng = _planner({}, name_meta={f"w{i}": ("bfloat16", [4]) for i in range(5)}) |
| 788 | chunks = eng._chunk_module_scatters([_module(layer, copies)]) |
| 789 | assert [ci for ci, _s in chunks] == [0] |
| 790 | assert len(chunks[0][1]) == 5 |
| 791 | |
| 792 | def test_one_chunk_per_owner_class_ascending(self): |
| 793 | """Classes are numbered by first appearance in metadata order, so the |
| 794 | replicated class (0) leads and the rest follow — the ordering the old |
| 795 | "-1 first, then ascending stamp" rule produced, now falling out of the |
| 796 | table instead of a special case.""" |
| 797 | layer = _FakeLayer("moe") |
| 798 | names = ["norm", "e5", "e0", "e2"] |
| 799 | copies = [_copy(n, shape=(4,)) for n in names] |
| 800 | eng = _planner( |
| 801 | {}, |
| 802 | name_meta={n: ("bfloat16", [4]) for n in names}, |
| 803 | held_by={"e0": [0], "e2": [1], "e5": [2]}, |
| 804 | ) |
| 805 | chunks = eng._chunk_module_scatters([_module(layer, copies)]) |
| 806 | assert [ci for ci, _s in chunks] == [0, 1, 2, 3] |
| 807 | assert [[s.src[0] for s in sc] for _ci, sc in chunks] == [ |
| 808 | ["norm"], |
| 809 | ["e5"], |
| 810 | ["e0"], |
| 811 | ["e2"], |
| 812 | ] |
| 813 | |
| 814 | def test_copy_order_within_a_chunk_is_bake_order(self): |
| 815 | """Scatter order within a chunk must follow the bake, not the stamp |
| 816 | sort: the packed layout and the producer's replay agree on keys order.""" |
| 817 | layer = _FakeLayer("moe") |
| 818 | names = ["e3", "e1", "e2"] # one stamp, deliberately unsorted names |
| 819 | copies = [_copy(n, shape=(4,)) for n in names] |
| 820 | eng = _planner( |
| 821 | {}, |
| 822 | name_meta={n: ("bfloat16", [4]) for n in names}, |
| 823 | held_by={n: [0] for n in names}, |
| 824 | ) |
| 825 | ((_ci, scatters),) = eng._chunk_module_scatters([_module(layer, copies)]) |
| 826 | assert [s.src[0] for s in scatters] == names |
| 827 | |
| 828 | def test_stamps_interleave_across_modules_without_reordering_within(self): |
| 829 | """Two modules' copies bucket by stamp; within each bucket the order is |
| 830 | module-then-bake order.""" |
| 831 | a, b = _FakeLayer("a"), _FakeLayer("b") |
| 832 | mods = [ |
| 833 | _module(a, [_copy("a0", shape=(4,)), _copy("a1", shape=(4,))]), |
| 834 | _module(b, [_copy("b0", shape=(4,)), _copy("b1", shape=(4,))]), |
| 835 | ] |
| 836 | eng = _planner( |
| 837 | {}, |
| 838 | name_meta={n: ("bfloat16", [4]) for n in ("a0", "a1", "b0", "b1")}, |
| 839 | held_by={"a0": [0], "a1": [1], "b0": [0], "b1": [1]}, |
| 840 | ) |
| 841 | chunks = eng._chunk_module_scatters(mods) |
| 842 | assert [(ci, [s.src[0] for s in sc]) for ci, sc in chunks] == [ |
| 843 | (1, ["a0", "b0"]), |
| 844 | (2, ["a1", "b1"]), |
| 845 | ] |
| 846 | |
| 847 | def test_scatters_carry_their_own_dtype(self): |
| 848 | """Dtype rides the record from the bake, where it is the fake's dtype |
| 849 | AFTER its op chain — not a plan-time lookup of the source name, which |
| 850 | would be wrong for any chain that reinterprets dtype.""" |
| 851 | layer = _FakeLayer("l") |
| 852 | eng = _planner({}, name_meta={"w": ("float32", [10])}) |
| 853 | ((_er, scatters),) = eng._chunk_module_scatters( |
| 854 | [_module(layer, [_copy("w", shape=(10,), dtype=torch.float32)])] |
| 855 | ) |
| 856 | (sc,) = scatters |
| 857 | assert sc.dtype is torch.float32 |
| 858 | assert sc.layer is layer |
| 859 | |
| 860 | |
| 861 | # --------------------------------------------------------------------------- |
| 862 | # _build_call_plan |
| 863 | # --------------------------------------------------------------------------- |
| 864 | |
| 865 | |
| 866 | class TestBuildCallPlan: |
| 867 | def test_one_chunk_per_group_when_unstamped(self): |
| 868 | baked, name_meta, names, group_lens = _one_module_per_layer(3) |
| 869 | eng = _planner(baked, name_meta=name_meta) |
| 870 | plan = eng._build_call_plan(names, group_lens) |
| 871 | assert len(plan.chunks) == len(group_lens) == 5 |
| 872 | assert plan.pre_free == [] |
| 873 | |
| 874 | def test_stamps_multiply_chunks_within_a_group(self): |
| 875 | layer = _FakeLayer("l") |
| 876 | names = [f"w{i}" for i in range(4)] |
| 877 | copies = [_copy(n, shape=(4,)) for n in names] |
| 878 | mod_plan = _module(layer, copies) |
| 879 | eng = _planner( |
| 880 | {n: mod_plan for n in names}, |
| 881 | name_meta={n: ("bfloat16", [4]) for n in names}, |
| 882 | held_by={"w2": [0], "w3": [1]}, |
| 883 | ) |
| 884 | plan = eng._build_call_plan(names, [4]) |
| 885 | assert len(plan.chunks) == 3 # replicated, w2's class, w3's class |
| 886 | |
| 887 | def test_materialize_fires_on_a_modules_first_chunk_only(self): |
| 888 | """Empty HF params are allocated once per module, by construction — |
| 889 | including a FusedMoE-like module whose copies span owner-class chunks.""" |
| 890 | layer = _FakeLayer("spanning") |
| 891 | names = [f"w{i}" for i in range(4)] |
| 892 | copies = [_copy(n, shape=(4,)) for n in names] |
| 893 | mod_plan = _module(layer, copies) |
| 894 | eng = _planner( |
| 895 | {n: mod_plan for n in names}, |
| 896 | name_meta={n: ("bfloat16", [4]) for n in names}, |
| 897 | held_by={"w0": [0], "w1": [0], "w2": [1], "w3": [1]}, |
| 898 | ) |
| 899 | plan = eng._build_call_plan(names, [4]) |
| 900 | assert [c.materialize for c in plan.chunks] == [[layer], []] |
| 901 | |
| 902 | def test_quant_defers_to_a_modules_last_chunk(self): |
| 903 | layer = _FakeLayer("spanning") |
| 904 | names = [f"w{i}" for i in range(4)] |
| 905 | copies = [_copy(n, shape=(4,)) for n in names] |
| 906 | mod_plan = _module(layer, copies) |
| 907 | eng = _planner( |
| 908 | {n: mod_plan for n in names}, |
| 909 | name_meta={n: ("bfloat16", [4]) for n in names}, |
| 910 | held_by={"w0": [0], "w1": [0], "w2": [1], "w3": [1]}, |
| 911 | ) |
| 912 | plan = eng._build_call_plan(names, [4]) |
| 913 | assert [c.quant for c in plan.chunks] == [[], [layer]] |
| 914 | |
| 915 | def test_free_signal_fires_on_each_groups_last_chunk(self): |
| 916 | baked, name_meta, names, group_lens = _one_module_per_layer(2) |
| 917 | eng = _planner(baked, name_meta=name_meta) |
| 918 | plan = eng._build_call_plan(names, group_lens) |
| 919 | assert [c.free for c in plan.chunks] == [[0], [1], [2], [3]] |
| 920 | |
| 921 | def test_free_signal_waits_for_a_groups_last_chunk_when_it_is_split(self): |
| 922 | """The load-bearing case: with the group cut across owner-class chunks the |
| 923 | signal must hang off the LAST one. Signaling earlier lets the producers |
| 924 | drop the gather buffers while a later chunk's RDMA is still reading.""" |
| 925 | layer = _FakeLayer("l") |
| 926 | names = [f"w{i}" for i in range(4)] |
| 927 | copies = [_copy(n, shape=(4,)) for n in names] |
| 928 | mod_plan = _module(layer, copies) |
| 929 | eng = _planner( |
| 930 | {n: mod_plan for n in names}, |
| 931 | name_meta={n: ("bfloat16", [4]) for n in names}, |
| 932 | held_by={"w2": [0], "w3": [1]}, |
| 933 | ) |
| 934 | plan = eng._build_call_plan(names, [4]) |
| 935 | assert len(plan.chunks) == 3 |
| 936 | assert [c.free for c in plan.chunks] == [[], [], [0]] |
| 937 | |
| 938 | def test_free_signal_timing_across_two_split_groups(self): |
| 939 | """Two groups x 2 owner-class chunks each: signals land on each group's |
| 940 | last chunk only.""" |
| 941 | names_a = [f"model.layers.0.w{i}" for i in range(4)] |
| 942 | names_b = [f"model.layers.1.w{i}" for i in range(4)] |
| 943 | baked, name_meta, stamps = {}, {}, {} |
| 944 | for names in (names_a, names_b): |
| 945 | layer = _FakeLayer(names[0]) |
| 946 | mod_plan = _module(layer, [_copy(n, shape=(4,)) for n in names]) |
| 947 | for i, n in enumerate(names): |
| 948 | baked[n] = mod_plan |
| 949 | name_meta[n] = ("bfloat16", [4]) |
| 950 | stamps[n] = [i // 2] # two owner classes per group |
| 951 | eng = _planner(baked, name_meta=name_meta, held_by=stamps) |
| 952 | plan = eng._build_call_plan(names_a + names_b, [4, 4]) |
| 953 | assert [c.free for c in plan.chunks] == [[], [0], [], [1]] |
| 954 | |
| 955 | def test_a_group_with_nothing_local_is_signaled_at_sync_start(self): |
| 956 | """The owners still published it, so this consumer must signal it — |
| 957 | immediately, before the pipeline (signal-before-publish is tolerated), |
| 958 | never hung off another group's chunk.""" |
| 959 | baked, name_meta, names, group_lens = _one_module_per_layer(2) |
| 960 | del baked["model.layers.1.w"] # nothing baked, and not live => no pull |
| 961 | eng = _planner( |
| 962 | baked, name_meta=name_meta, live=set(names) - {"model.layers.1.w"} |
| 963 | ) |
| 964 | plan = eng._build_call_plan(names, group_lens) |
| 965 | assert len(plan.chunks) == 3 |
| 966 | assert plan.pre_free == [2] |
| 967 | assert [c.free for c in plan.chunks] == [[0], [1], [3]] |
| 968 | |
| 969 | def test_leading_groups_with_nothing_local_go_to_pre_free(self): |
| 970 | baked, name_meta, names, group_lens = _one_module_per_layer(2) |
| 971 | del baked["embed.weight"] |
| 972 | eng = _planner(baked, name_meta=name_meta, live=set(names) - {"embed.weight"}) |
| 973 | plan = eng._build_call_plan(names, group_lens) |
| 974 | assert plan.pre_free == [0] |
| 975 | assert len(plan.chunks) == 3 |
| 976 | |
| 977 | def test_modules_are_deduped_within_a_group(self): |
| 978 | """Several names of one fused module map to the same scatter list; the |
| 979 | plan must not chunk it twice.""" |
| 980 | layer = _FakeLayer("fused") |
| 981 | names = ["qkv.q", "qkv.k", "qkv.v"] |
| 982 | copies = [_copy(n, shape=(4,)) for n in names] |
| 983 | mod_plan = _module(layer, copies) |
| 984 | eng = _planner( |
| 985 | {n: mod_plan for n in names}, |
| 986 | name_meta={n: ("bfloat16", [4]) for n in names}, |
| 987 | ) |
| 988 | plan = eng._build_call_plan(names, [3]) |
| 989 | (chunk,) = plan.chunks |
| 990 | assert chunk.materialize == [layer] |
| 991 | assert len(chunk.scatters) == 3 |
| 992 | |
| 993 | def test_plan_is_a_pure_function_of_its_inputs(self): |
| 994 | """Called twice with the same arguments it must produce the same plan — |
| 995 | that is what lets the engine build it once and reuse it every sync.""" |
| 996 | baked, name_meta, names, group_lens = _one_module_per_layer(2) |
| 997 | eng = _planner(baked, name_meta=name_meta) |
| 998 | a = eng._build_call_plan(names, group_lens) |
| 999 | b = eng._build_call_plan(names, group_lens) |
| 1000 | assert [c.pack_bytes for c in a.chunks] == [c.pack_bytes for c in b.chunks] |
| 1001 | assert [c.keys for c in a.chunks] == [c.keys for c in b.chunks] |
| 1002 | assert [c.free for c in a.chunks] == [c.free for c in b.chunks] |
| 1003 | assert a.pre_free == b.pre_free |
| 1004 | |
| 1005 | def test_single_producer_routes_every_chunk_to_local_owner_zero(self): |
| 1006 | baked, name_meta, names, group_lens = _one_module_per_layer(2) |
| 1007 | eng = _planner(baked, name_meta=name_meta) |
| 1008 | plan = eng._build_call_plan(names, group_lens) |
| 1009 | assert [c.owner for c in plan.chunks] == [0] * len(plan.chunks) |
| 1010 | |
| 1011 | |
| 1012 | class TestUnbakedNamesGuard: |
| 1013 | """A live name with no baked plan has no way to load — there is no fallback, |
| 1014 | and its pull would target groups the pipeline already freed — so the plan |
| 1015 | build fails at init, naming the names.""" |
| 1016 | |
| 1017 | def test_a_live_unbaked_name_fails_the_plan_build(self): |
| 1018 | baked, name_meta, names, group_lens = _one_module_per_layer(2) |
| 1019 | del baked["norm.weight"] # unbaked but live |
| 1020 | eng = _planner(baked, name_meta=name_meta, live=set(names)) |
| 1021 | with pytest.raises(RuntimeError, match="norm.weight"): |
| 1022 | eng._build_call_plan(names, group_lens) |
| 1023 | |
| 1024 | def test_never_copied_names_are_dropped_entirely(self): |
| 1025 | """Experts owned by another EP rank no-op in their loader: not an error, |
| 1026 | and not a chunk — their group is pre-freed if nothing else fills it.""" |
| 1027 | baked, name_meta, names, group_lens = _one_module_per_layer(2) |
| 1028 | del baked["norm.weight"] |
| 1029 | eng = _planner(baked, name_meta=name_meta, live=set(names) - {"norm.weight"}) |
| 1030 | plan = eng._build_call_plan(names, group_lens) |
| 1031 | assert len(names) - 1 == len(plan.chunks) |
| 1032 | assert plan.pre_free == [len(group_lens) - 1] |
| 1033 | |
| 1034 | |
| 1035 | class TestCallPlanRouting: |
| 1036 | """Under partial ownership each chunk must be pulled from a producer that |
| 1037 | actually holds every name in it. Consumers bind EVERY producer (signals fan |
| 1038 | out to all owners), so local index == trainer rank.""" |
| 1039 | |
| 1040 | def _routed_planner( |
| 1041 | self, held_by_group, num_consumers, consumer_id, *, held_by=None |
| 1042 | ): |
| 1043 | """``held_by_group[g]`` = the ranks holding group g's names, so a test |
| 1044 | can still speak in groups; ``held_by`` overrides individual names.""" |
| 1045 | baked, name_meta, names, group_lens = _one_module_per_layer( |
| 1046 | len(held_by_group) - 2 |
| 1047 | ) |
| 1048 | per_name = {n: list(held_by_group[gi]) for gi, n in enumerate(names)} |
| 1049 | per_name.update(held_by or {}) |
| 1050 | n_prod = max(max(o) for o in per_name.values()) + 1 |
| 1051 | owner_sets: list[list[int]] = [] |
| 1052 | seen: dict[tuple, int] = {} |
| 1053 | classes = [] |
| 1054 | for n in names: |
| 1055 | key = tuple(sorted(per_name[n])) |
| 1056 | if key not in seen: |
| 1057 | seen[key] = len(owner_sets) |
| 1058 | owner_sets.append(list(key)) |
| 1059 | classes.append(seen[key]) |
| 1060 | eng = _planner(baked, name_meta=name_meta) |
| 1061 | router = RdtRouter( |
| 1062 | n_prod, num_consumers, owner_sets, classes, names, group_lens |
| 1063 | ) |
| 1064 | router.validate() |
| 1065 | router.bind([object()] * n_prod, [object()] * n_prod, consumer_id) |
| 1066 | eng._router = router |
| 1067 | return eng, names, group_lens |
| 1068 | |
| 1069 | def test_two_stage_ownership_splits_chunks_between_owners(self): |
| 1070 | """Two PP stages, 4 groups: stage 0 holds the first half, stage 1 the |
| 1071 | second. Every chunk goes to a holder of its names.""" |
| 1072 | eng, names, group_lens = self._routed_planner([[0], [0], [1], [1]], 1, 0) |
| 1073 | plan = eng._build_call_plan(names, group_lens) |
| 1074 | assert [c.owner for c in plan.chunks] == [0, 0, 1, 1] |
| 1075 | |
| 1076 | def test_a_multi_owner_group_still_resolves_to_one_producer(self): |
| 1077 | """Every chunk is served by exactly ONE producer: splitting a pull only |
| 1078 | multiplies produce calls, since the consumer's NIC bounds it.""" |
| 1079 | eng, names, group_lens = self._routed_planner([[0, 1]] * 4, 1, 0) |
| 1080 | plan = eng._build_call_plan(names, group_lens) |
| 1081 | assert all(0 <= c.owner < 2 for c in plan.chunks) |
| 1082 | |
| 1083 | def test_an_expert_name_routes_to_the_rank_holding_it(self): |
| 1084 | """PP ∩ EP: stage 0 = ranks 0-1, stage 1 = ranks 2-3, and one expert |
| 1085 | name inside a stage-1 group is held by rank 3 alone. The name's own owner |
| 1086 | set is what routes it.""" |
| 1087 | eng, names, group_lens = self._routed_planner( |
| 1088 | [[0, 1], [0, 1], [2, 3], [2, 3]], |
| 1089 | 1, |
| 1090 | 0, |
| 1091 | held_by={"model.layers.1.w": [3]}, |
| 1092 | ) |
| 1093 | plan = eng._build_call_plan(names, group_lens) |
| 1094 | assert plan.chunks[2].owner == 3 |
| 1095 | assert plan.chunks[0].owner in (0, 1) |
| 1096 | assert plan.chunks[3].owner in (2, 3) |
| 1097 | |
| 1098 | def test_an_unheld_name_is_rejected_at_init(self): |
| 1099 | """A name no rank holds can never be served. The trainer raises when it |
| 1100 | transposes the holdings; the router's validate() is the consumer-side |
| 1101 | backstop, and either way it fails at init rather than at first pull.""" |
| 1102 | with pytest.raises(ValueError, match="empty"): |
| 1103 | RdtRouter(2, 1, [[]], [0], ["w"], [1]).validate() |
| 1104 | |
| 1105 | |
| 1106 | class TestSignalCompleteness: |
| 1107 | """Every gather group is signaled exactly once — after its last chunk when |
| 1108 | the worker pulls from it, at sync start otherwise — for ANY expert |
| 1109 | placement. A group signaled twice over-credits the barrier; a group never |
| 1110 | signaled parks a producer credit and hangs end_sync.""" |
| 1111 | |
| 1112 | def _moe_planner(self, worker_experts, *, ep_size=4, n_experts=8, n_layers=2): |
| 1113 | """Pre / n_layers MoE layers / post. Each layer: one norm (its own |
| 1114 | module, stamp -1) + n_experts expert names stamped ``e // n_local``. |
| 1115 | The worker's bake covers only ``worker_experts`` (its placement); |
| 1116 | foreign expert names never copied => dropped entirely, as in the real |
| 1117 | bake.""" |
| 1118 | n_local = n_experts // ep_size |
| 1119 | names = ["embed.weight"] |
| 1120 | group_lens = [1] |
| 1121 | baked = { |
| 1122 | "embed.weight": _module( |
| 1123 | _FakeLayer("embed"), [_copy("embed.weight", shape=(4,))] |
| 1124 | ) |
| 1125 | } |
| 1126 | name_meta = {"embed.weight": ("bfloat16", [4])} |
| 1127 | stamps: dict = {} |
| 1128 | live = {"embed.weight"} |
| 1129 | for li in range(n_layers): |
| 1130 | norm = f"model.layers.{li}.norm" |
| 1131 | enames = [f"model.layers.{li}.experts.{e}.w" for e in range(n_experts)] |
| 1132 | names += [norm] + enames |
| 1133 | group_lens.append(1 + n_experts) |
| 1134 | name_meta[norm] = ("bfloat16", [4]) |
| 1135 | live.add(norm) |
| 1136 | baked[norm] = _module(_FakeLayer(f"norm{li}"), [_copy(norm, shape=(4,))]) |
| 1137 | fused = _module( |
| 1138 | _FakeLayer(f"moe{li}"), |
| 1139 | [ |
| 1140 | _copy(f"model.layers.{li}.experts.{e}.w", shape=(4,)) |
| 1141 | for e in worker_experts |
| 1142 | ], |
| 1143 | ) |
| 1144 | for e in range(n_experts): |
| 1145 | n = f"model.layers.{li}.experts.{e}.w" |
| 1146 | name_meta[n] = ("bfloat16", [4]) |
| 1147 | stamps[n] = [e // n_local] # held by that EP rank alone |
| 1148 | if e in worker_experts: |
| 1149 | baked[n] = fused |
| 1150 | live.add(n) |
| 1151 | names.append("lm_head.weight") |
| 1152 | group_lens.append(1) |
| 1153 | baked["lm_head.weight"] = _module( |
| 1154 | _FakeLayer("head"), [_copy("lm_head.weight", shape=(4,))] |
| 1155 | ) |
| 1156 | name_meta["lm_head.weight"] = ("bfloat16", [4]) |
| 1157 | live.add("lm_head.weight") |
| 1158 | eng = _planner(baked, name_meta=name_meta, live=live, held_by=stamps) |
| 1159 | return eng, names, group_lens |
| 1160 | |
| 1161 | @staticmethod |
| 1162 | def _signals(plan): |
| 1163 | return sorted(list(plan.pre_free) + [gi for c in plan.chunks for gi in c.free]) |
| 1164 | |
| 1165 | def test_linear_placement_signals_every_group_exactly_once(self): |
| 1166 | eng, names, group_lens = self._moe_planner(worker_experts=[0, 1, 2, 3]) |
| 1167 | plan = eng._build_call_plan(names, group_lens) |
| 1168 | assert self._signals(plan) == list(range(len(group_lens))) |
| 1169 | |
| 1170 | def test_round_robin_placement_signals_every_group_exactly_once(self): |
| 1171 | eng, names, group_lens = self._moe_planner(worker_experts=[0, 2, 4, 6]) |
| 1172 | plan = eng._build_call_plan(names, group_lens) |
| 1173 | assert self._signals(plan) == list(range(len(group_lens))) |
| 1174 | |
| 1175 | def test_placement_changes_only_the_chunk_count_never_the_signals(self): |
| 1176 | """Linear experts 0-3 hit 2 owner classes; round_robin 0,2,4,6 |
| 1177 | hits all 4. More chunks per group, identical signal set.""" |
| 1178 | plans = {} |
| 1179 | for label, experts in (("linear", [0, 1, 2, 3]), ("round_robin", [0, 2, 4, 6])): |
| 1180 | eng, names, group_lens = self._moe_planner(worker_experts=experts) |
| 1181 | plans[label] = eng._build_call_plan(names, group_lens) |
| 1182 | # per MoE group: the replicated chunk + one per owner class present |
| 1183 | assert ( |
| 1184 | len(plans["linear"].chunks) == 2 + 2 * 3 |
| 1185 | ) # embed+head + 2 layers x (1 + 2) |
| 1186 | assert len(plans["round_robin"].chunks) == 2 + 2 * 5 # 2 layers x (1 + 4) |
| 1187 | assert self._signals(plans["linear"]) == self._signals(plans["round_robin"]) |
| 1188 | |
| 1189 | def test_the_fused_module_materializes_first_and_quants_last_across_chunks(self): |
| 1190 | """The FusedMoE shape: one module's copies span every owner-class chunk of |
| 1191 | its group; materialize on its first, quant on its last.""" |
| 1192 | eng, names, group_lens = self._moe_planner( |
| 1193 | worker_experts=[0, 2, 4, 6], n_layers=1 |
| 1194 | ) |
| 1195 | plan = eng._build_call_plan(names, group_lens) |
| 1196 | # group 1 = [norm chunk, coord 0..3 chunks] at chunk indices 1..5 |
| 1197 | fused_chunks = [ |
| 1198 | c |
| 1199 | for c in plan.chunks[1:6] |
| 1200 | if any("experts" in sc.src[0] for sc in c.scatters) |
| 1201 | ] |
| 1202 | assert len(fused_chunks) == 4 |
| 1203 | (fused_layer,) = fused_chunks[0].materialize |
| 1204 | assert fused_layer.tag == "moe0" |
| 1205 | assert fused_chunks[0].quant == [] |
| 1206 | assert fused_chunks[-1].quant == [fused_layer] |
| 1207 | |
| 1208 | |
| 1209 | # --------------------------------------------------------------------------- |
| 1210 | # Shared helpers |
| 1211 | # --------------------------------------------------------------------------- |
| 1212 | |
| 1213 | |
| 1214 | class TestBufferAllocBytes: |
| 1215 | def test_rounds_up_to_a_coarse_256mb_boundary(self): |
| 1216 | assert buffer_alloc_bytes(1) == 256 << 20 |
| 1217 | assert buffer_alloc_bytes(256 << 20) == 256 << 20 |
| 1218 | assert buffer_alloc_bytes((256 << 20) + 1) == 512 << 20 |
| 1219 | |
| 1220 | def test_presize_is_a_floor_not_a_cap(self): |
| 1221 | assert buffer_alloc_bytes(1, presize=3 << 30) == 3 << 30 |
| 1222 | big = 4 << 30 |
| 1223 | assert buffer_alloc_bytes(big, presize=1 << 30) == big |
| 1224 | |
| 1225 | def test_never_returns_less_than_requested(self): |
| 1226 | for nbytes in (1, 1 << 20, (700 << 20) + 3): |
| 1227 | assert buffer_alloc_bytes(nbytes) >= nbytes |
| 1228 | |
| 1229 | |
| 1230 | class TestLayerwiseGroups: |
| 1231 | def test_partition_is_one_group_per_layer_between_unindexed_blocks(self): |
| 1232 | names = [ |
| 1233 | "embed.weight", |
| 1234 | "model.layers.0.a", |
| 1235 | "model.layers.0.b", |
| 1236 | "model.layers.1.a", |
| 1237 | "norm.weight", |
| 1238 | ] |
| 1239 | assert layerwise_groups(names) == [ |
| 1240 | ["embed.weight"], |
| 1241 | ["model.layers.0.a", "model.layers.0.b"], |
| 1242 | ["model.layers.1.a"], |
| 1243 | ["norm.weight"], |
| 1244 | ] |
| 1245 | |
| 1246 | def test_flattening_a_partition_returns_the_input_order(self): |
| 1247 | """Un-indexed names group by POSITION, not by name class — so one that |
| 1248 | looks like a pre-block name still lands after the layers it follows. |
| 1249 | Group index therefore means the same thing on every rank.""" |
| 1250 | names = [ |
| 1251 | "embed.weight", |
| 1252 | "model.layers.0.a", |
| 1253 | "embed.weight.tied", # after the layers => its own trailing block |
| 1254 | ] |
| 1255 | groups = layerwise_groups(names) |
| 1256 | assert [n for g in groups for n in g] == names |
| 1257 | |
| 1258 | def test_layers_are_ordered_by_index_not_appearance(self): |
| 1259 | names = ["model.layers.10.a", "model.layers.2.a"] |
| 1260 | assert layerwise_groups(names) == [ |
| 1261 | ["model.layers.2.a"], |
| 1262 | ["model.layers.10.a"], |
| 1263 | ] |
| 1264 | |
| 1265 | def test_stacks_keep_prefix_appearance_order_and_sort_within(self): |
| 1266 | """Two stacks stay in the order they appear — the vision tower before the |
| 1267 | text stack — while each sorts by index internally.""" |
| 1268 | names = [ |
| 1269 | "visual.blocks.1.w", |
| 1270 | "visual.blocks.0.w", |
| 1271 | "model.language_model.layers.1.w", |
| 1272 | "model.language_model.layers.0.w", |
| 1273 | ] |
| 1274 | assert layerwise_groups(names) == [ |
| 1275 | ["visual.blocks.0.w"], |
| 1276 | ["visual.blocks.1.w"], |
| 1277 | ["model.language_model.layers.0.w"], |
| 1278 | ["model.language_model.layers.1.w"], |
| 1279 | ] |
| 1280 | |
| 1281 | def test_the_post_block_lands_last_however_early_it_arrives(self): |
| 1282 | """Megatron-Bridge streams the last pipeline stage's output block before |
| 1283 | its layers. Sweeping un-indexed names after the first layer into a |
| 1284 | trailing group is what lets the gather loop walk groups ascending.""" |
| 1285 | names = ["model.layers.0.w", "model.norm.weight", "model.layers.1.w"] |
| 1286 | assert layerwise_groups(names) == [ |
| 1287 | ["model.layers.0.w"], |
| 1288 | ["model.layers.1.w"], |
| 1289 | ["model.norm.weight"], |
| 1290 | ] |
| 1291 | |
| 1292 | def test_names_sharing_a_layer_coalesce_when_the_source_interleaves(self): |
| 1293 | """A raw-checkpoint source may yield in shard-packing order. Coalescing on |
| 1294 | the key keeps that from shattering one layer into many groups.""" |
| 1295 | names = [ |
| 1296 | "model.layers.0.a", |
| 1297 | "model.layers.1.a", |
| 1298 | "model.layers.0.b", |
| 1299 | ] |
| 1300 | assert layerwise_groups(names) == [ |
| 1301 | ["model.layers.0.a", "model.layers.0.b"], |
| 1302 | ["model.layers.1.a"], |
| 1303 | ] |
| 1304 | |
| 1305 | def test_no_empty_groups(self): |
| 1306 | for names in ([], ["only.pre"], ["model.layers.0.a"]): |
| 1307 | assert all(g for g in layerwise_groups(names)) |
| 1308 | |
| 1309 | def test_a_model_with_no_layers_is_one_pre_group(self): |
| 1310 | assert layerwise_groups(["a", "b"]) == [["a", "b"]] |
| 1311 | |
| 1312 | @pytest.mark.parametrize( |
| 1313 | "prefix", |
| 1314 | [ |
| 1315 | "model.layers.", # Llama, Qwen3, GLM-4.5, DeepSeek, Mixtral |
| 1316 | "model.language_model.layers.", # Qwen3.5/3.6, Qwen2.5-VL, Gemma3 |
| 1317 | "language_model.layers.", # Kimi-K2.5, Nemotron-VL |
| 1318 | "thinker.model.layers.", # Qwen2.5/3-Omni |
| 1319 | "transformer.h.", # GPT-2, GPT-J, Bloom, Falcon |
| 1320 | "transformer.blocks.", # MPT, DBRX |
| 1321 | "transformer.encoder.layers.", # ChatGLM |
| 1322 | "backbone.layers.", # Mamba, Mamba2 |
| 1323 | "gpt_neox.layers.", # GPT-NeoX |
| 1324 | "model.decoder.layers.", # OPT |
| 1325 | "model.h.", # Exaone |
| 1326 | "model.layers.layers.", # Plamo3 |
| 1327 | "bert.encoder.layer.", # BERT, RoBERTa |
| 1328 | "visual.blocks.", # Qwen-VL vision tower |
| 1329 | "vision_tower.vision_model.encoder.layers.", # SigLIP/CLIP towers |
| 1330 | "audio_tower.layers.", # Whisper-style audio towers |
| 1331 | ], |
| 1332 | ) |
| 1333 | def test_every_supported_naming_convention_partitions_per_layer(self, prefix): |
| 1334 | """One decoder layer per group under any prefix, with no per-architecture |
| 1335 | table. A literal `model.layers.` match silently yields one whole-model |
| 1336 | group for most of these.""" |
| 1337 | names = [f"{prefix}{i}.self_attn.q_proj.weight" for i in range(3)] |
| 1338 | assert layerwise_groups(names) == [[n] for n in names] |
| 1339 | |
| 1340 | def test_a_moe_layer_stays_whole(self): |
| 1341 | """The OUTERMOST index wins, so per-expert names group by layer. Splitting |
| 1342 | on the expert would cut FusedMoE — one leaf module — across groups.""" |
| 1343 | names = [ |
| 1344 | "model.layers.0.mlp.experts.0.w1.weight", |
| 1345 | "model.layers.0.mlp.experts.1.w1.weight", |
| 1346 | "model.layers.1.mlp.experts.0.w1.weight", |
| 1347 | ] |
| 1348 | assert layerwise_groups(names) == [ |
| 1349 | [ |
| 1350 | "model.layers.0.mlp.experts.0.w1.weight", |
| 1351 | "model.layers.0.mlp.experts.1.w1.weight", |
| 1352 | ], |
| 1353 | ["model.layers.1.mlp.experts.0.w1.weight"], |
| 1354 | ] |
| 1355 | |
| 1356 | def test_a_vlm_partitions_the_vision_tower_and_the_text_stack(self): |
| 1357 | """Qwen-VL layout. Today's literal `model.layers.` match puts every one of |
| 1358 | these in ONE group — the whole model — because the text stack ships as |
| 1359 | `model.language_model.layers.`.""" |
| 1360 | names = [ |
| 1361 | "visual.patch_embed.proj.weight", |
| 1362 | "visual.blocks.0.attn.qkv.weight", |
| 1363 | "visual.blocks.1.attn.qkv.weight", |
| 1364 | "visual.merger.mlp.weight", |
| 1365 | "model.language_model.layers.0.self_attn.q_proj.weight", |
| 1366 | "lm_head.weight", |
| 1367 | ] |
| 1368 | assert layerwise_groups(names) == [ |
| 1369 | ["visual.patch_embed.proj.weight"], |
| 1370 | ["visual.blocks.0.attn.qkv.weight"], |
| 1371 | ["visual.blocks.1.attn.qkv.weight"], |
| 1372 | ["model.language_model.layers.0.self_attn.q_proj.weight"], |
| 1373 | ["visual.merger.mlp.weight", "lm_head.weight"], |
| 1374 | ] |
| 1375 | |
| 1376 | def test_two_stacks_sharing_an_index_do_not_merge(self): |
| 1377 | """The key carries the prefix, so vision block 0 and text layer 0 are |
| 1378 | different groups even though both are index 0.""" |
| 1379 | names = [ |
| 1380 | "visual.blocks.0.attn.qkv.weight", |
| 1381 | "model.language_model.layers.0.self_attn.q_proj.weight", |
| 1382 | ] |
| 1383 | assert layerwise_groups(names) == [[n] for n in names] |
| 1384 | |
| 1385 | |
| 1386 | class TestDtypeFromName: |
| 1387 | def test_resolves_wire_names(self): |
| 1388 | assert _dtype_from_name("bfloat16") is torch.bfloat16 |
| 1389 | assert _dtype_from_name("float8_e4m3fn") is torch.float8_e4m3fn |
| 1390 | |
| 1391 | def test_rejects_a_non_dtype_attribute(self): |
| 1392 | with pytest.raises(ValueError): |
| 1393 | _dtype_from_name("nn") # torch.nn exists but is not a dtype |
| 1394 | |
| 1395 | def test_rejects_an_unknown_name(self): |
| 1396 | with pytest.raises(ValueError): |
| 1397 | _dtype_from_name("float9") |
| 1398 | |
| 1399 | |
| 1400 | class TestOpAllowlistAgreement: |
| 1401 | """The consumer records op chains from ``SUPPORTED_OPS``; the producer |
| 1402 | replays a chain only if every op is in ``ALLOWED_OPS``. The two must describe |
| 1403 | the same contract, so they are derived from one table. |
| 1404 | |
| 1405 | A divergence is silent until first pull: an op the consumer can emit but the |
| 1406 | producer rejects bakes fine at init, then fails a whole sync later with |
| 1407 | "disallowed op". |
| 1408 | """ |
| 1409 | |
| 1410 | def test_the_two_allowlists_describe_the_same_contract(self): |
| 1411 | assert set(SUPPORTED_OPS.values()) == set(ALLOWED_OPS) |
| 1412 | |
| 1413 | def test_every_bake_recordable_op_is_replayable(self): |
| 1414 | """Anything the bake can record must be serveable, or the failure lands a |
| 1415 | whole sync later than the mistake.""" |
| 1416 | assert not set(SUPPORTED_OPS.values()) - set(ALLOWED_OPS) |
| 1417 | |
| 1418 | def test_the_producer_allows_nothing_the_consumer_cannot_emit(self): |
| 1419 | """The allowlist is also a guard against a misbehaving or spoofed |
| 1420 | consumer invoking arbitrary methods on trainer tensors, so it must not be |
| 1421 | wider than the ops the bake can actually produce. ``to`` in particular |
| 1422 | would let a replay change dtype or device.""" |
| 1423 | assert not set(ALLOWED_OPS) - set(SUPPORTED_OPS.values()) |
| 1424 | assert "to" not in ALLOWED_OPS |
| 1425 | |
| 1426 | def test_transpose_via_t_is_serveable(self): |
| 1427 | """The specific regression: a chain recorded from ``.t()`` must pass the |
| 1428 | producer's guard.""" |
| 1429 | fake = FakeRDTTensor( |
| 1430 | name="w", |
| 1431 | shape=torch.Size((4, 6)), |
| 1432 | dtype=torch.bfloat16, |
| 1433 | device=META, |
| 1434 | sink=None, |
| 1435 | ) |
| 1436 | (op, _args, _kwargs) = fake.t()._key()[1][0] |
| 1437 | assert op in ALLOWED_OPS |
| 1438 | |
| 1439 | |
| 1440 | # --------------------------------------------------------------------------- |
| 1441 | # Bake against a real model (GPU) |
| 1442 | # --------------------------------------------------------------------------- |
| 1443 | |
| 1444 | TINY_QWEN = "Qwen/Qwen3-0.6B" |
| 1445 | |
| 1446 | # One small model per attention/FFN shape the bake has to survive. The point is |
| 1447 | # the checkpoint->module mapping, which differs structurally between them: GQA |
| 1448 | # fuses Q/K/V of unequal head counts into one param, MLA replaces them with the |
| 1449 | # compressed kv_a/kv_b pair, and MoE turns per-expert checkpoint entries into |
| 1450 | # stacked expert params. |
| 1451 | # |
| 1452 | # Keep them small: `load_format="dummy"` skips the download but not the |
| 1453 | # allocation, so every parameter is still built on the device at full size. |
| 1454 | BAKE_MODELS = [ |
| 1455 | pytest.param(TINY_QWEN, id="gqa-dense"), |
| 1456 | pytest.param("TitanML/tiny-mixtral", id="moe"), |
| 1457 | # MLA (kv_lora_rank set, so `ModelConfig.use_mla`) plus routed experts. Only |
| 1458 | # the STRUCTURE is under test -- `_bake_probe` synthesizes the checkpoint |
| 1459 | # from the safetensors headers and never reads a weight value -- so a random |
| 1460 | # 0.06B model does the job, as long as it has the MLA head shape the prefill |
| 1461 | # backends accept: (qk_nope=128, qk_rope=64, v=128). |
| 1462 | # |
| 1463 | # Not an fp8 checkpoint: `process_weights_after_loading` swaps the vLLM |
| 1464 | # parameter subclasses out on those, and the plain second `load_weights` |
| 1465 | # this test compares against then cannot run. |
| 1466 | pytest.param("hmellor/tiny-random-DeepseekV2ForCausalLM", id="mla-moe"), |
| 1467 | ] |
| 1468 | |
| 1469 | |
| 1470 | class TestRequiresTheRayExecutor: |
| 1471 | """The engine's data plane is Ray's, so a non-Ray executor cannot work. The |
| 1472 | check is at construction because the alternative is an opaque failure during |
| 1473 | the first handshake, long after the misconfiguration.""" |
| 1474 | |
| 1475 | def _construct(self, backend): |
| 1476 | cfg = SimpleNamespace( |
| 1477 | parallel_config=SimpleNamespace(distributed_executor_backend=backend) |
| 1478 | ) |
| 1479 | eng = object.__new__(ShardedRDTWeightTransferEngine) |
| 1480 | # Only the guard is under test, so stand in for the base __init__. |
| 1481 | with pytest.MonkeyPatch.context() as m: |
| 1482 | m.setattr( |
| 1483 | ShardedRDTWeightTransferEngine.__bases__[0], |
| 1484 | "__init__", |
| 1485 | lambda self, *a, **k: None, |
| 1486 | ) |
| 1487 | ShardedRDTWeightTransferEngine.__init__(eng, None, cfg, META, None) |
| 1488 | return eng |
| 1489 | |
| 1490 | @pytest.mark.parametrize("backend", ["uni", "mp", "external_launcher"]) |
| 1491 | def test_a_non_ray_executor_is_refused(self, backend): |
| 1492 | with pytest.raises(ValueError, match="requires distributed_executor"): |
| 1493 | self._construct(backend) |
| 1494 | |
| 1495 | def test_ray_is_accepted(self): |
| 1496 | self._construct("ray") |
| 1497 | |
| 1498 | def test_a_custom_executor_class_is_left_alone(self): |
| 1499 | """A `type[Executor]` override is deliberate and unjudgeable here, so it |
| 1500 | must not be rejected for merely not being the string "ray".""" |
| 1501 | |
| 1502 | class _CustomExecutor: |
| 1503 | pass |
| 1504 | |
| 1505 | self._construct(_CustomExecutor) |
| 1506 | |
| 1507 | |
| 1508 | def _bake_probe(self, model_name): # noqa: PLR0915 (runs in the vLLM worker) |
| 1509 | """Drive `_bake` against the worker's real model and check the three |
| 1510 | properties. Returns a small summary; every assertion fires here, in the |
| 1511 | worker, so a failure surfaces as the RPC's exception. |
| 1512 | |
| 1513 | Self-contained on purpose: this is pickled to the worker, so it imports what |
| 1514 | it needs rather than closing over this module's globals. |
| 1515 | """ |
| 1516 | import torch |
| 1517 | from huggingface_hub import get_safetensors_metadata |
| 1518 | |
| 1519 | from vllm.distributed.weight_transfer.sharded_rdt_common import ALLOWED_OPS |
| 1520 | from vllm.distributed.weight_transfer.sharded_rdt_engine import ( |
| 1521 | ShardedRDTWeightTransferEngine, |
| 1522 | ShardedRDTWeightTransferInitInfo, |
| 1523 | ) |
| 1524 | from vllm.model_executor.model_loader.reload.utils import get_layer_tensors |
| 1525 | |
| 1526 | model = self.model_runner.model |
| 1527 | device = next(model.parameters()).device |
| 1528 | |
| 1529 | # ---- the checkpoint, exactly as a trainer would describe and serve it ---- |
| 1530 | # Only the safetensors HEADERS are fetched (a few KB); the values are |
| 1531 | # synthesized. Nothing here depends on real weights -- both the reference |
| 1532 | # load and the replay read this same dict, so the comparison is unaffected, |
| 1533 | # and it keeps a multi-GB download out of CI. |
| 1534 | st_dtype = { |
| 1535 | "F64": torch.float64, |
| 1536 | "F32": torch.float32, |
| 1537 | "F16": torch.float16, |
| 1538 | "BF16": torch.bfloat16, |
| 1539 | "F8_E4M3": torch.float8_e4m3fn, |
| 1540 | "F8_E5M2": torch.float8_e5m2, |
| 1541 | "I64": torch.int64, |
| 1542 | "I32": torch.int32, |
| 1543 | "I8": torch.int8, |
| 1544 | "U8": torch.uint8, |
| 1545 | "BOOL": torch.bool, |
| 1546 | } |
| 1547 | meta = get_safetensors_metadata(model_name) |
| 1548 | specs = {} |
| 1549 | for file_meta in meta.files_metadata.values(): |
| 1550 | for name, t in file_meta.tensors.items(): |
| 1551 | specs[name] = (st_dtype[t.dtype], tuple(t.shape)) |
| 1552 | |
| 1553 | gen = torch.Generator(device="cpu").manual_seed(0) |
| 1554 | ckpt: dict[str, torch.Tensor] = {} |
| 1555 | for name, (dt, shape) in specs.items(): |
| 1556 | if dt.is_floating_point: |
| 1557 | # Small magnitudes so any narrowing cast stays finite. |
| 1558 | v = torch.randn(shape, generator=gen, dtype=torch.float32) * 0.1 |
| 1559 | ckpt[name] = v.to(dtype=dt) |
| 1560 | else: |
| 1561 | ckpt[name] = torch.zeros(shape, dtype=dt) |
| 1562 | names = sorted(ckpt) |
| 1563 | dtype_names = [str(ckpt[n].dtype).split(".")[-1] for n in names] |
| 1564 | shapes = [list(ckpt[n].shape) for n in names] |
| 1565 | |
| 1566 | def _feed(): |
| 1567 | return iter([(n, ckpt[n]) for n in names]) |
| 1568 | |
| 1569 | # ---- reference: what the model's OWN loader produces from that checkpoint -- |
| 1570 | # Poison first, so "which params did load_weights actually write?" is |
| 1571 | # answerable. Some params have no checkpoint source at all -- derived ones, |
| 1572 | # or ones process_weights_after_loading builds -- and those keep their NaNs |
| 1573 | # and are legitimately outside the plan. That set is what the plan must |
| 1574 | # cover, not every parameter of the model. |
| 1575 | def _poison(): |
| 1576 | with torch.no_grad(): |
| 1577 | for prm in model.parameters(): |
| 1578 | if prm.is_floating_point(): |
| 1579 | prm.fill_(float("nan")) |
| 1580 | else: |
| 1581 | prm.zero_() |
| 1582 | |
| 1583 | _poison() |
| 1584 | model.load_weights(_feed()) |
| 1585 | loaded = { |
| 1586 | n |
| 1587 | for n, prm in model.named_parameters() |
| 1588 | if not (prm.is_floating_point() and torch.isnan(prm).any()) |
| 1589 | } |
| 1590 | # On the host: the weights themselves already fill the GPU. |
| 1591 | reference = { |
| 1592 | n: prm.detach().to("cpu", copy=True) |
| 1593 | for n, prm in model.named_parameters() |
| 1594 | if n in loaded |
| 1595 | } |
| 1596 | |
| 1597 | # ---- pre-bake state, for the restoration check -------------------------- |
| 1598 | param_name_of = {} |
| 1599 | for mod_name, mod in model.named_modules(): |
| 1600 | for pname in get_layer_tensors(mod): |
| 1601 | param_name_of[(id(mod), pname)] = f"{mod_name}.{pname}".lstrip(".") |
| 1602 | loaders_before = {} |
| 1603 | for mod in model.modules(): |
| 1604 | for pname, tensor in get_layer_tensors(mod).items(): |
| 1605 | loaders_before[(id(mod), pname)] = getattr(tensor, "weight_loader", None) |
| 1606 | |
| 1607 | # ---- bake --------------------------------------------------------------- |
| 1608 | eng = object.__new__(ShardedRDTWeightTransferEngine) |
| 1609 | eng.model = model |
| 1610 | eng.device = device |
| 1611 | eng._name_to_plan = {} |
| 1612 | eng._name_meta = {} |
| 1613 | eng._live_names = set() |
| 1614 | eng._bake( |
| 1615 | ShardedRDTWeightTransferInitInfo( |
| 1616 | names=names, dtype_names=dtype_names, shapes=shapes |
| 1617 | ) |
| 1618 | ) |
| 1619 | |
| 1620 | # (1) the bake ran and produced a plan |
| 1621 | assert eng._name_to_plan, "bake produced no plan" |
| 1622 | assert eng._live_names, "bake recorded no live names" |
| 1623 | n_modules = len({id(v) for v in eng._name_to_plan.values()}) |
| 1624 | |
| 1625 | # (2) every intermediate change to model state was undone |
| 1626 | for mod in model.modules(): |
| 1627 | for pname, tensor in get_layer_tensors(mod).items(): |
| 1628 | key = (id(mod), pname) |
| 1629 | loader = getattr(tensor, "weight_loader", None) |
| 1630 | assert not hasattr(loader, "_rdt_stamp_inner"), ( |
| 1631 | f"recording stamp left on {param_name_of.get(key, key)}" |
| 1632 | ) |
| 1633 | assert loader is loaders_before[key], ( |
| 1634 | f"weight_loader not restored on {param_name_of.get(key, key)}" |
| 1635 | ) |
| 1636 | assert not tensor.is_meta, ( |
| 1637 | f"{param_name_of.get(key, key)} left on meta after the dry run" |
| 1638 | ) |
| 1639 | for pname, p in model.named_parameters(): |
| 1640 | assert torch.equal(p.detach().cpu(), reference[pname]), f"bake mutated {pname}" |
| 1641 | |
| 1642 | # (3) replaying the recorded op chains reproduces that same load |
| 1643 | scatters = {id(v): v for v in eng._name_to_plan.values()} |
| 1644 | |
| 1645 | # No deferred-attention layer may be a plan target. The engine's quant |
| 1646 | # thread calls `info.reset()` on every module it processes, which makes |
| 1647 | # finalize_layerwise_reload SKIP that module -- and finalize is the only |
| 1648 | # thing that calls the LAYER's process_weights_after_loading, which is where |
| 1649 | # MLA derives W_UK_T/W_UV from kv_b_proj. An attention layer in the plan |
| 1650 | # would therefore leave those derived weights stale after a sync, silently. |
| 1651 | from vllm.model_executor.layers.attention import is_deferred_attention_layer |
| 1652 | |
| 1653 | attn_in_plan = sorted( |
| 1654 | { |
| 1655 | type(c.layer).__name__ |
| 1656 | for plan in scatters.values() |
| 1657 | for c in plan |
| 1658 | if is_deferred_attention_layer(c.layer) |
| 1659 | } |
| 1660 | ) |
| 1661 | assert not attn_in_plan, ( |
| 1662 | f"plan targets deferred-attention layer(s) {attn_in_plan}; the quant " |
| 1663 | f"thread would reset them and finalize would skip their " |
| 1664 | f"process_weights_after_loading, staling any derived weights" |
| 1665 | ) |
| 1666 | |
| 1667 | touched = set() |
| 1668 | _poison() |
| 1669 | with torch.no_grad(): |
| 1670 | for plan in scatters.values(): |
| 1671 | for c in plan: |
| 1672 | src_name, chain = c.src |
| 1673 | t = ckpt[src_name] |
| 1674 | for op, args, kw in chain: |
| 1675 | assert op in ALLOWED_OPS, f"disallowed op {op!r}" |
| 1676 | t = getattr(t, op)(*args, **dict(kw)) |
| 1677 | dest = getattr(c.layer, c.param_name) |
| 1678 | dest.as_strided(c.shape, c.stride, c.offset).copy_(t) |
| 1679 | touched.add(param_name_of[(id(c.layer), c.param_name)]) |
| 1680 | |
| 1681 | assert loaded, "the reference load wrote nothing" |
| 1682 | # The plan owns everything with a checkpoint source. Anything else the model |
| 1683 | # ends up holding is DERIVED -- under MLA, `process_weights_after_loading` |
| 1684 | # splits kv_b_proj into W_UK_T/W_UV via `replace_parameter` -- and belongs to |
| 1685 | # finalize_layerwise_reload, which `finish_weight_update` runs after the |
| 1686 | # replay. So the uncovered set must contain only params owned by a module |
| 1687 | # that post-processes; a plain loadable weight going missing still fails. |
| 1688 | post_processed = { |
| 1689 | id(mod) |
| 1690 | for mod in model.modules() |
| 1691 | if hasattr(mod, "process_weights_after_loading") |
| 1692 | } |
| 1693 | modules_by_name = dict(model.named_modules()) |
| 1694 | uncovered = sorted(loaded - touched) |
| 1695 | derived: list[str] = [] |
| 1696 | unexplained: list[str] = [] |
| 1697 | for n in uncovered: |
| 1698 | mod_name = n.rsplit(".", 1)[0] |
| 1699 | mod = modules_by_name.get(mod_name) |
| 1700 | is_derived = mod is not None and id(mod) in post_processed |
| 1701 | (derived if is_derived else unexplained).append(n) |
| 1702 | assert not unexplained, ( |
| 1703 | f"the plan misses {len(unexplained)} param(s) load_weights wrote that " |
| 1704 | f"nothing derives: {unexplained[:5]}" |
| 1705 | ) |
| 1706 | mismatched = [ |
| 1707 | n |
| 1708 | for n, prm in model.named_parameters() |
| 1709 | if n in loaded and not torch.equal(prm.detach().cpu(), reference[n]) |
| 1710 | ] |
| 1711 | assert not mismatched, f"replay differs from load_weights for: {mismatched[:5]}" |
| 1712 | |
| 1713 | return { |
| 1714 | "names": len(names), |
| 1715 | "baked": len(eng._name_to_plan), |
| 1716 | "live": len(eng._live_names), |
| 1717 | "modules": n_modules, |
| 1718 | "loaded_params": len(loaded), |
| 1719 | "derived_not_in_plan": len(derived), |
| 1720 | "replayed_params": len(touched), |
| 1721 | "total_params": sum(1 for _ in model.named_parameters()), |
| 1722 | } |
| 1723 | |
| 1724 | |
| 1725 | @pytest.mark.skipif( |
| 1726 | not torch.cuda.is_available(), reason="the bake drives a real model on GPU" |
| 1727 | ) |
| 1728 | class TestBakeOnARealModel: |
| 1729 | """`_bake` against a real vLLM model, which the CPU tests above cannot reach: |
| 1730 | they drive the planner with hand-built scatters, so nothing there exercises |
| 1731 | the dry run through a real `load_weights` — the part that monkeypatches the |
| 1732 | model and has to put it back. |
| 1733 | """ |
| 1734 | |
| 1735 | @pytest.mark.parametrize("model", BAKE_MODELS) |
| 1736 | def test_bake_records_a_replayable_plan_and_restores_the_model( |
| 1737 | self, vllm_runner, monkeypatch, model |
| 1738 | ): |
| 1739 | monkeypatch.setenv("VLLM_ALLOW_INSECURE_SERIALIZATION", "1") |
| 1740 | with vllm_runner( |
| 1741 | model, load_format="dummy", enforce_eager=True, max_model_len=1024 |
| 1742 | ) as llm: |
| 1743 | (summary,) = llm.llm.collective_rpc(_bake_probe, args=(model,)) |
| 1744 | print(f"bake summary [{model}]: {summary}") |
| 1745 | # Non-vacuity: the checks above are only worth something if the plan |
| 1746 | # covers the model. Fusion means far fewer leaf modules than names -- |
| 1747 | # QKV and gate/up for GQA, the stacked experts for MoE. |
| 1748 | assert summary["baked"] >= 0.95 * summary["names"], summary |
| 1749 | assert summary["live"] >= 0.95 * summary["names"], summary |
| 1750 | assert 0 < summary["modules"] < summary["baked"], summary |
| 1751 | # The probe already asserts the plan covers every param load_weights |
| 1752 | # wrote; this just keeps that from being vacuously few. |
| 1753 | assert summary["loaded_params"] > 0.5 * summary["total_params"], summary |