| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | from dataclasses import dataclass |
| 5 | from functools import partial |
| 6 | from itertools import accumulate |
| 7 | from types import SimpleNamespace |
| 8 | |
| 9 | import pytest |
| 10 | import torch |
| 11 | from torch import nn |
| 12 | from torch.nn import functional as F |
| 13 | from transformers import Qwen4ExpTextConfig |
| 14 | |
| 15 | import vllm.model_executor.layers.vocab_parallel_embedding as embedding_module |
| 16 | import vllm.model_executor.parameter as parameter_module |
| 17 | import vllm.models.qwen4_exp.common.ngram_embedding as ngram_embedding_module |
| 18 | from tests.v1.attention.utils import BatchSpec, create_common_attn_metadata |
| 19 | from vllm.config import SpeculativeConfig, VllmConfig |
| 20 | from vllm.config.compilation import CUDAGraphMode |
| 21 | from vllm.config.quantization import QuantizationConfigArgs |
| 22 | from vllm.model_executor.layers.quantization.fp8 import Fp8Config |
| 23 | from vllm.model_executor.layers.quantization.modelopt import ( |
| 24 | ModelOptMixedPrecisionConfig, |
| 25 | ModelOptNvFp4Config, |
| 26 | ) |
| 27 | from vllm.model_executor.layers.quantization.online.base import OnlineQuantizationConfig |
| 28 | from vllm.models.qwen4_exp.amd import ple_layer as amd_ple_layer |
| 29 | from vllm.models.qwen4_exp.amd.ple_layer import ( |
| 30 | Qwen4ExpPLELayer as Qwen4ExpPLELayerAMD, |
| 31 | ) |
| 32 | from vllm.models.qwen4_exp.common.ple import ( |
| 33 | PLEShardOverlap, |
| 34 | compute_ple_shard_overlap, |
| 35 | copy_ple_embedding_shard_, |
| 36 | ) |
| 37 | from vllm.models.qwen4_exp.nvidia.ngram_embedding import ( |
| 38 | Qwen4ExpNGramEmbedding, |
| 39 | Qwen4ExpPLEDeviceEmbedding, |
| 40 | Qwen4ExpPLEEmbeddingMethod, |
| 41 | Qwen4ExpPLEFp8EmbeddingMethod, |
| 42 | Qwen4ExpPLEPinnedHostEmbedding, |
| 43 | Qwen4ExpPLEUnquantizedEmbeddingMethod, |
| 44 | ) |
| 45 | from vllm.models.qwen4_exp.nvidia.ple_layer import Qwen4ExpPLELayer |
| 46 | from vllm.utils.torch_utils import weak_ref_tensor |
| 47 | from vllm.v1.attention.backends.short_conv_attn import ( |
| 48 | PleShortConvAttentionMetadata, |
| 49 | PleShortConvAttentionMetadataBuilder, |
| 50 | ) |
| 51 | from vllm.v1.attention.backends.utils import NULL_BLOCK_ID |
| 52 | from vllm.v1.kv_cache_interface import MambaSpec |
| 53 | |
| 54 | |
| 55 | @pytest.mark.parametrize("device", ["cpu", "cuda"]) |
| 56 | @pytest.mark.parametrize("full_graph", [False, True]) |
| 57 | @pytest.mark.parametrize( |
| 58 | "query_lens,drafts,spec_rows,non_spec_rows,token_order", |
| 59 | [ |
| 60 | pytest.param([4, 2, 0], [3, 1, -1], [0, 1], [], None, id="spec-padding"), |
| 61 | pytest.param( |
| 62 | [3, 1, 2, 5, 0], |
| 63 | [2, -1, 1, -1, -1], |
| 64 | [0, 2], |
| 65 | [1, 3], |
| 66 | [0, 1, 2, 4, 5, 3, 6, 7, 8, 9, 10], |
| 67 | id="mixed-padding", |
| 68 | ), |
| 69 | pytest.param([1, 5], [-1, -1], [], [0, 1], None, id="non-spec"), |
| 70 | ], |
| 71 | ) |
| 72 | def test_ple_metadata_preserves_request_and_token_order( |
| 73 | device, full_graph, query_lens, drafts, spec_rows, non_spec_rows, token_order |
| 74 | ): |
| 75 | """Preserve state routing, acceptance and prefill offsets without GPU sync.""" |
| 76 | if device == "cuda" and not torch.cuda.is_available(): |
| 77 | pytest.skip("CUDA required") |
| 78 | device = torch.device(device) |
| 79 | config = VllmConfig() |
| 80 | config.speculative_config = SpeculativeConfig( |
| 81 | method="ngram", num_speculative_tokens=3 |
| 82 | ) |
| 83 | config.compilation_config.cudagraph_mode = ( |
| 84 | CUDAGraphMode.FULL if full_graph else CUDAGraphMode.NONE |
| 85 | ) |
| 86 | config.compilation_config.max_cudagraph_capture_size = 32 |
| 87 | builder = PleShortConvAttentionMetadataBuilder( |
| 88 | MambaSpec(block_size=16, shapes=((16, 64),), dtypes=(torch.float32,)), |
| 89 | ["ple"], |
| 90 | config, |
| 91 | device, |
| 92 | ) |
| 93 | common = create_common_attn_metadata( |
| 94 | BatchSpec(seq_lens=[q + 16 for q in query_lens], query_lens=query_lens), |
| 95 | 16, |
| 96 | device, |
| 97 | arange_block_indices=True, |
| 98 | ) |
| 99 | common.is_prefilling = torch.tensor( |
| 100 | [q > 1 and d < 0 for q, d in zip(query_lens, drafts)] |
| 101 | ) |
| 102 | # Token padding is independent of the sum of per-request query lengths. |
| 103 | common.num_actual_tokens += 3 |
| 104 | accepted = torch.arange(1, len(query_lens) + 1, dtype=torch.int32, device=device) |
| 105 | drafts_cpu = torch.tensor(drafts, dtype=torch.int32) |
| 106 | sync_mode = torch.cuda.get_sync_debug_mode() if device.type == "cuda" else None |
| 107 | if sync_mode is not None: |
| 108 | torch.cuda.set_sync_debug_mode("error") |
| 109 | try: |
| 110 | metadata = builder.build( |
| 111 | 0, |
| 112 | common, |
| 113 | num_accepted_tokens=accepted, |
| 114 | num_decode_draft_tokens_cpu=drafts_cpu, |
| 115 | ) |
| 116 | finally: |
| 117 | if sync_mode is not None: |
| 118 | torch.cuda.set_sync_debug_mode(sync_mode) |
| 119 | |
| 120 | def check(actual, expected): |
| 121 | torch.testing.assert_close( |
| 122 | actual, torch.tensor(expected, dtype=actual.dtype, device=device) |
| 123 | ) |
| 124 | |
| 125 | assert metadata.num_spec_decodes == len(spec_rows) |
| 126 | assert metadata.num_decodes == sum(query_lens[i] == 1 for i in non_spec_rows) |
| 127 | assert metadata.num_prefills == sum(query_lens[i] > 1 for i in non_spec_rows) |
| 128 | if spec_rows: |
| 129 | torch.testing.assert_close( |
| 130 | metadata.spec_state_indices_tensor[: len(spec_rows)], |
| 131 | common.block_table_tensor[spec_rows, 0], |
| 132 | ) |
| 133 | torch.testing.assert_close( |
| 134 | metadata.num_accepted_tokens[: len(spec_rows)], accepted[spec_rows] |
| 135 | ) |
| 136 | check( |
| 137 | metadata.spec_query_start_loc[: len(spec_rows) + 1], |
| 138 | [0, *accumulate(query_lens[i] for i in spec_rows)], |
| 139 | ) |
| 140 | if full_graph and not non_spec_rows: |
| 141 | check(metadata.spec_state_indices_tensor[len(spec_rows) :], [NULL_BLOCK_ID]) |
| 142 | check(metadata.num_accepted_tokens[len(spec_rows) :], [1]) |
| 143 | check( |
| 144 | metadata.spec_query_start_loc[len(spec_rows) + 1 :], [sum(query_lens)] |
| 145 | ) |
| 146 | assert ( |
| 147 | metadata.spec_state_indices_tensor.data_ptr() |
| 148 | == builder.spec_state_indices_tensor.data_ptr() |
| 149 | ) |
| 150 | if non_spec_rows: |
| 151 | torch.testing.assert_close( |
| 152 | metadata.state_indices_tensor, common.block_table_tensor[non_spec_rows, 0] |
| 153 | ) |
| 154 | prefill_lens = [query_lens[i] for i in non_spec_rows if query_lens[i] > 1] |
| 155 | check(metadata.query_start_loc_p, [0, *accumulate(prefill_lens)]) |
| 156 | check(metadata.has_initial_states_p, [True] * len(prefill_lens)) |
| 157 | if token_order is not None: |
| 158 | check( |
| 159 | torch.cat((metadata.spec_token_indx, metadata.non_spec_token_indx)), |
| 160 | token_order, |
| 161 | ) |
| 162 | else: |
| 163 | assert metadata.spec_token_indx is None |
| 164 | assert metadata.non_spec_token_indx is None |
| 165 | assert metadata.nums_dict is None |
| 166 | assert metadata.batch_ptr is None |
| 167 | assert metadata.token_chunk_offset_ptr is None |
| 168 | |
| 169 | |
| 170 | def _mock_etp_group( |
| 171 | monkeypatch: pytest.MonkeyPatch, |
| 172 | world_size: int = 1, |
| 173 | all_reduce=lambda tensor: tensor, |
| 174 | ) -> None: |
| 175 | group = SimpleNamespace( |
| 176 | rank_in_group=0, |
| 177 | world_size=world_size, |
| 178 | all_reduce=all_reduce, |
| 179 | ) |
| 180 | monkeypatch.setattr(ngram_embedding_module, "get_etp_group", lambda: group) |
| 181 | monkeypatch.setattr( |
| 182 | ngram_embedding_module, |
| 183 | "get_tp_group", |
| 184 | lambda: SimpleNamespace(world_size=world_size), |
| 185 | ) |
| 186 | |
| 187 | |
| 188 | def _make_ngram_embedding_for_load_test() -> Qwen4ExpNGramEmbedding: |
| 189 | module = Qwen4ExpNGramEmbedding.__new__(Qwen4ExpNGramEmbedding) |
| 190 | nn.Module.__init__(module) |
| 191 | module.split_ngram_parts = 2 |
| 192 | module.register_buffer("layer_multipliers", torch.zeros(1, dtype=torch.long)) |
| 193 | module.register_buffer("ngram_heads_offsets", torch.zeros(1, dtype=torch.long)) |
| 194 | module.register_buffer("ngram_heads_vocab_sizes", torch.zeros(1, dtype=torch.long)) |
| 195 | module.ngram_embedding = SimpleNamespace( |
| 196 | org_vocab_size=8, |
| 197 | embedding_dim=2, |
| 198 | weight=nn.Parameter(torch.full((4, 2), -1.0)), |
| 199 | shard_indices=SimpleNamespace( |
| 200 | org_vocab_start_index=2, |
| 201 | org_vocab_end_index=6, |
| 202 | ), |
| 203 | ) |
| 204 | _set_test_embedding_weight_loader(module.ngram_embedding) |
| 205 | return module |
| 206 | |
| 207 | |
| 208 | def _set_test_embedding_weight_loader(embedding) -> None: |
| 209 | embedding.weight.weight_loader = partial( |
| 210 | copy_ple_embedding_shard_, |
| 211 | tp_start=embedding.shard_indices.org_vocab_start_index, |
| 212 | tp_end=embedding.shard_indices.org_vocab_end_index, |
| 213 | ) |
| 214 | |
| 215 | |
| 216 | def _make_fp8_ngram_embedding_for_load_test() -> Qwen4ExpNGramEmbedding: |
| 217 | module = _make_ngram_embedding_for_load_test() |
| 218 | embedding = nn.Module() |
| 219 | embedding.org_vocab_size = 8 |
| 220 | embedding.embedding_dim = 2 |
| 221 | embedding.shard_indices = SimpleNamespace( |
| 222 | org_vocab_start_index=2, |
| 223 | org_vocab_end_index=6, |
| 224 | ) |
| 225 | embedding.register_parameter( |
| 226 | "weight", |
| 227 | nn.Parameter( |
| 228 | torch.full((4, 2), -1.0).to(torch.float8_e4m3fn), |
| 229 | requires_grad=False, |
| 230 | ), |
| 231 | ) |
| 232 | embedding.register_parameter( |
| 233 | "weight_scale", |
| 234 | nn.Parameter(torch.zeros(1, dtype=torch.float32), requires_grad=False), |
| 235 | ) |
| 236 | _set_test_embedding_weight_loader(embedding) |
| 237 | module.ngram_embedding = embedding |
| 238 | return module |
| 239 | |
| 240 | |
| 241 | def test_ple_shard_overlap_and_copy() -> None: |
| 242 | overlap = compute_ple_shard_overlap( |
| 243 | checkpoint_start=2, checkpoint_rows=5, tp_start=4, tp_end=8 |
| 244 | ) |
| 245 | assert overlap == PLEShardOverlap(source_start=2, destination_start=0, row_count=3) |
| 246 | |
| 247 | destination = torch.full((4, 2), -1.0) |
| 248 | loaded = torch.arange(10, dtype=torch.float64).reshape(5, 2) |
| 249 | copied = copy_ple_embedding_shard_( |
| 250 | destination, |
| 251 | loaded, |
| 252 | checkpoint_start=2, |
| 253 | tp_start=4, |
| 254 | tp_end=8, |
| 255 | ) |
| 256 | |
| 257 | assert copied == 3 |
| 258 | torch.testing.assert_close(destination[:3], loaded[2:5].float()) |
| 259 | torch.testing.assert_close(destination[3], torch.tensor([-1.0, -1.0])) |
| 260 | |
| 261 | |
| 262 | def test_ple_shard_copy_is_a_noop_without_overlap() -> None: |
| 263 | destination = torch.ones(4, 2) |
| 264 | copied = copy_ple_embedding_shard_( |
| 265 | destination, |
| 266 | torch.zeros(2, 2), |
| 267 | checkpoint_start=10, |
| 268 | tp_start=4, |
| 269 | tp_end=8, |
| 270 | ) |
| 271 | |
| 272 | assert copied == 0 |
| 273 | assert torch.equal(destination, torch.ones_like(destination)) |
| 274 | |
| 275 | |
| 276 | def test_ngram_embedding_loads_shards_and_ignores_legacy_token_lookup() -> None: |
| 277 | module = _make_ngram_embedding_for_load_test() |
| 278 | shard_0 = torch.arange(8, dtype=torch.float32).reshape(4, 2) |
| 279 | shard_1 = torch.arange(8, 16, dtype=torch.float32).reshape(4, 2) |
| 280 | |
| 281 | loaded = module.load_weights( |
| 282 | [ |
| 283 | ("ngram_embedding.shard_0.weight", shard_0), |
| 284 | ("ngram_embedding.shard_1.weight", shard_1), |
| 285 | ("token_lookup", torch.tensor([2, 1, 0])), |
| 286 | ] |
| 287 | ) |
| 288 | |
| 289 | assert loaded == {"ngram_embedding.weight"} |
| 290 | torch.testing.assert_close( |
| 291 | module.ngram_embedding.weight, |
| 292 | torch.cat((shard_0[2:4], shard_1[0:2])), |
| 293 | ) |
| 294 | |
| 295 | |
| 296 | def test_ngram_embedding_loads_fp8_shards_and_global_scale() -> None: |
| 297 | module = _make_fp8_ngram_embedding_for_load_test() |
| 298 | shard_0 = torch.arange(8, dtype=torch.float32).reshape(4, 2).to(torch.float8_e4m3fn) |
| 299 | shard_1 = ( |
| 300 | torch.arange(8, 16, dtype=torch.float32).reshape(4, 2).to(torch.float8_e4m3fn) |
| 301 | ) |
| 302 | weight_scale = torch.tensor([0.25], dtype=torch.bfloat16) |
| 303 | |
| 304 | loaded = module.load_weights( |
| 305 | [ |
| 306 | ("ngram_embedding.shard_0.weight", shard_0), |
| 307 | ("ngram_embedding.shard_1.weight", shard_1), |
| 308 | ("ngram_embedding.weight_scale", weight_scale), |
| 309 | ] |
| 310 | ) |
| 311 | |
| 312 | assert loaded == {"ngram_embedding.weight", "ngram_embedding.weight_scale"} |
| 313 | assert module.ngram_embedding.weight.dtype == torch.float8_e4m3fn |
| 314 | assert torch.equal( |
| 315 | module.ngram_embedding.weight.float(), |
| 316 | torch.cat((shard_0[2:4], shard_1[0:2])).float(), |
| 317 | ) |
| 318 | assert module.ngram_embedding.weight_scale.dtype == torch.float32 |
| 319 | torch.testing.assert_close( |
| 320 | module.ngram_embedding.weight_scale, |
| 321 | weight_scale.float(), |
| 322 | ) |
| 323 | |
| 324 | |
| 325 | @pytest.mark.parametrize(("dp_rank", "local_tokens"), [(0, 2), (1, 3)]) |
| 326 | def test_etp_lookup_gathers_and_returns_dp_local_rows( |
| 327 | monkeypatch: pytest.MonkeyPatch, |
| 328 | dp_rank: int, |
| 329 | local_tokens: int, |
| 330 | ) -> None: |
| 331 | embedding = Qwen4ExpPLEDeviceEmbedding.__new__(Qwen4ExpPLEDeviceEmbedding) |
| 332 | nn.Module.__init__(embedding) |
| 333 | embedding.tp_size = 2 |
| 334 | embedding.etp_data_parallel_size = 2 |
| 335 | embedding.data_parallel_rank = dp_rank |
| 336 | gathered_ids = torch.tensor([[10], [11], [0], [20], [21], [22]]) |
| 337 | group = SimpleNamespace( |
| 338 | rank_in_group=dp_rank, |
| 339 | all_gather=lambda input_ids, dim: gathered_ids, |
| 340 | all_reduce=lambda embeddings: embeddings, |
| 341 | ) |
| 342 | embedding.parallel_group = group |
| 343 | forward_context = SimpleNamespace( |
| 344 | dp_metadata=SimpleNamespace( |
| 345 | num_tokens_across_dp_cpu=torch.tensor([2, 3]), |
| 346 | ) |
| 347 | ) |
| 348 | monkeypatch.setattr(ngram_embedding_module, "get_dp_group", lambda: group) |
| 349 | monkeypatch.setattr( |
| 350 | ngram_embedding_module, "get_forward_context", lambda: forward_context |
| 351 | ) |
| 352 | local_ids = gathered_ids[dp_rank * 3 : dp_rank * 3 + local_tokens] |
| 353 | embeddings = torch.arange(12).reshape(6, 2) |
| 354 | |
| 355 | def forward(self, input_ids: torch.Tensor) -> torch.Tensor: |
| 356 | assert torch.equal(input_ids, gathered_ids) |
| 357 | return embeddings |
| 358 | |
| 359 | monkeypatch.setattr( |
| 360 | ngram_embedding_module.PLEVocabParallelEmbedding, |
| 361 | "forward", |
| 362 | forward, |
| 363 | ) |
| 364 | output = embedding(local_ids) |
| 365 | |
| 366 | expected = embeddings[dp_rank * 3 : dp_rank * 3 + local_tokens] |
| 367 | assert torch.equal(output, expected) |
| 368 | |
| 369 | |
| 370 | def _make_fp8_embedding_layer( |
| 371 | monkeypatch: pytest.MonkeyPatch, |
| 372 | *, |
| 373 | load_scale: bool = True, |
| 374 | ) -> Qwen4ExpPLEDeviceEmbedding: |
| 375 | _mock_etp_group(monkeypatch) |
| 376 | monkeypatch.setattr(embedding_module, "get_tensor_model_parallel_rank", lambda: 0) |
| 377 | monkeypatch.setattr( |
| 378 | embedding_module, "get_tensor_model_parallel_world_size", lambda: 1 |
| 379 | ) |
| 380 | monkeypatch.setattr(parameter_module, "get_tensor_model_parallel_rank", lambda: 0) |
| 381 | monkeypatch.setattr( |
| 382 | parameter_module, "get_tensor_model_parallel_world_size", lambda: 1 |
| 383 | ) |
| 384 | method = Qwen4ExpPLEFp8EmbeddingMethod() |
| 385 | layer = Qwen4ExpPLEDeviceEmbedding( |
| 386 | 3, |
| 387 | 2, |
| 388 | params_dtype=torch.bfloat16, |
| 389 | padding_size=1, |
| 390 | prefix="test.ple_embedding", |
| 391 | embedding_method=method, |
| 392 | ) |
| 393 | weight = torch.tensor([[1.0, 2.0], [4.0, 8.0], [16.0, 32.0]]) |
| 394 | layer.weight.data.copy_(weight.to(torch.float8_e4m3fn)) |
| 395 | if load_scale: |
| 396 | layer.weight_scale.data.copy_(torch.tensor([0.25], dtype=torch.bfloat16)) |
| 397 | return layer |
| 398 | |
| 399 | |
| 400 | def test_ple_fp8_embedding_dequantizes_in_ple_layer(monkeypatch) -> None: |
| 401 | layer = _make_fp8_embedding_layer(monkeypatch) |
| 402 | layer.quant_method.process_weights_after_loading(layer) |
| 403 | quantized_output = layer(torch.tensor([2, 0])) |
| 404 | ple_layer = Qwen4ExpPLELayer.__new__(Qwen4ExpPLELayer) |
| 405 | nn.Module.__init__(ple_layer) |
| 406 | ple_layer.ple_embedding = nn.Module() |
| 407 | ple_layer.ple_embedding.ngram_embedding = layer |
| 408 | |
| 409 | output = ple_layer._dequantize_embeddings( |
| 410 | quantized_output, |
| 411 | torch.bfloat16, |
| 412 | ) |
| 413 | |
| 414 | assert layer.weight.dtype == torch.float8_e4m3fn |
| 415 | assert layer.weight_scale.dtype == torch.float32 |
| 416 | assert quantized_output.dtype == torch.float8_e4m3fn |
| 417 | assert output.dtype == torch.bfloat16 |
| 418 | weight = torch.tensor([[1.0, 2.0], [4.0, 8.0], [16.0, 32.0]]) |
| 419 | torch.testing.assert_close(output, (weight[[2, 0]] * 0.25).bfloat16()) |
| 420 | |
| 421 | |
| 422 | def test_ple_fp8_embedding_rejects_missing_global_scale(monkeypatch) -> None: |
| 423 | layer = _make_fp8_embedding_layer(monkeypatch, load_scale=False) |
| 424 | |
| 425 | with pytest.raises(ValueError, match="missing its global scale"): |
| 426 | layer.quant_method.process_weights_after_loading(layer) |
| 427 | |
| 428 | |
| 429 | @pytest.mark.parametrize("requires_device_loading", [False, True]) |
| 430 | @pytest.mark.parametrize("load_scale", [False, True]) |
| 431 | def test_pinned_ple_post_load_validates_scale_without_staging( |
| 432 | monkeypatch: pytest.MonkeyPatch, requires_device_loading: bool, load_scale: bool |
| 433 | ) -> None: |
| 434 | """Validate scales through the loader while honoring the staging policy.""" |
| 435 | from contextlib import nullcontext |
| 436 | |
| 437 | from vllm.model_executor.layers.quantization.base_config import QuantizeMethodBase |
| 438 | from vllm.model_executor.model_loader import utils as loader_utils |
| 439 | |
| 440 | source = _make_fp8_embedding_layer(monkeypatch, load_scale=load_scale) |
| 441 | embedding = Qwen4ExpPLEPinnedHostEmbedding.__new__(Qwen4ExpPLEPinnedHostEmbedding) |
| 442 | nn.Module.__init__(embedding) |
| 443 | embedding.weight = source.weight |
| 444 | embedding.weight_scale = source.weight_scale |
| 445 | embedding.embedding_method = source.embedding_method |
| 446 | embedding.quant_method = source.embedding_method |
| 447 | embedding.tp_rank = source.tp_rank |
| 448 | embedding.tp_size = source.tp_size |
| 449 | assert QuantizeMethodBase.requires_device_loading |
| 450 | assert not embedding.quant_method.requires_device_loading |
| 451 | if requires_device_loading: |
| 452 | embedding.quant_method.requires_device_loading = True |
| 453 | model = nn.Module() |
| 454 | model.embedding = embedding |
| 455 | weight_ptr = embedding.weight.data_ptr() |
| 456 | staged_modules = [] |
| 457 | |
| 458 | def track_staging(module, target_device): |
| 459 | staged_modules.append(module) |
| 460 | return nullcontext() |
| 461 | |
| 462 | monkeypatch.setattr(loader_utils, "device_loading_context", track_staging) |
| 463 | monkeypatch.setattr(loader_utils, "maybe_retie_word_embeddings", lambda *args: None) |
| 464 | monkeypatch.setattr( |
| 465 | loader_utils, "release_device_memory_under_pressure", lambda *args: None |
| 466 | ) |
| 467 | config = SimpleNamespace(quantization="fp8") |
| 468 | if load_scale: |
| 469 | loader_utils.process_weights_after_loading(model, config, torch.device("cuda")) |
| 470 | else: |
| 471 | with pytest.raises(ValueError, match="missing its global scale"): |
| 472 | loader_utils.process_weights_after_loading( |
| 473 | model, config, torch.device("cuda") |
| 474 | ) |
| 475 | assert embedding.weight.device.type == "cpu" |
| 476 | assert embedding.weight.data_ptr() == weight_ptr |
| 477 | assert staged_modules == ([embedding] if requires_device_loading else []) |
| 478 | |
| 479 | |
| 480 | def test_ple_fp8_embedding_uses_int8_for_parallel_reduce(monkeypatch) -> None: |
| 481 | monkeypatch.setattr(embedding_module, "get_tensor_model_parallel_rank", lambda: 0) |
| 482 | monkeypatch.setattr( |
| 483 | embedding_module, "get_tensor_model_parallel_world_size", lambda: 2 |
| 484 | ) |
| 485 | monkeypatch.setattr(parameter_module, "get_tensor_model_parallel_rank", lambda: 0) |
| 486 | monkeypatch.setattr( |
| 487 | parameter_module, "get_tensor_model_parallel_world_size", lambda: 2 |
| 488 | ) |
| 489 | monkeypatch.setattr( |
| 490 | embedding_module, |
| 491 | "get_masked_input_and_mask", |
| 492 | lambda *args: ( |
| 493 | torch.tensor([0, 0]), |
| 494 | torch.tensor([False, True]), |
| 495 | ), |
| 496 | ) |
| 497 | reduced_dtypes = [] |
| 498 | |
| 499 | def all_reduce(tensor: torch.Tensor) -> torch.Tensor: |
| 500 | reduced_dtypes.append(tensor.dtype) |
| 501 | return tensor.clone() |
| 502 | |
| 503 | _mock_etp_group(monkeypatch, world_size=2, all_reduce=all_reduce) |
| 504 | layer = Qwen4ExpPLEDeviceEmbedding( |
| 505 | 4, |
| 506 | 2, |
| 507 | params_dtype=torch.bfloat16, |
| 508 | padding_size=1, |
| 509 | prefix="test.ple_embedding", |
| 510 | embedding_method=Qwen4ExpPLEFp8EmbeddingMethod(), |
| 511 | ) |
| 512 | layer.weight.data.copy_( |
| 513 | torch.tensor([[1.0, 2.0], [4.0, 8.0]]).to(torch.float8_e4m3fn) |
| 514 | ) |
| 515 | |
| 516 | output = layer(torch.tensor([0, 2])) |
| 517 | |
| 518 | assert reduced_dtypes == [torch.int8] |
| 519 | assert output.dtype == torch.float8_e4m3fn |
| 520 | torch.testing.assert_close(output[0].float(), layer.weight[0].float()) |
| 521 | assert torch.count_nonzero(output[1].float()) == 0 |
| 522 | |
| 523 | |
| 524 | def test_ple_fp8_embedding_respects_checkpoint_shard_exclusions() -> None: |
| 525 | prefix = "model.layers.1.ple.ple_embedding.ngram_embedding" |
| 526 | quant_config = Fp8Config( |
| 527 | ignored_layers=[], |
| 528 | weight_block_size=[128, 128], |
| 529 | ) |
| 530 | assert isinstance( |
| 531 | Qwen4ExpPLEEmbeddingMethod.from_quant_config(quant_config, prefix), |
| 532 | Qwen4ExpPLEFp8EmbeddingMethod, |
| 533 | ) |
| 534 | |
| 535 | quant_config.ignored_layers = [f"{prefix}.shard_0"] |
| 536 | assert isinstance( |
| 537 | Qwen4ExpPLEEmbeddingMethod.from_quant_config(quant_config, prefix), |
| 538 | Qwen4ExpPLEUnquantizedEmbeddingMethod, |
| 539 | ) |
| 540 | assert isinstance( |
| 541 | Qwen4ExpPLEEmbeddingMethod.from_quant_config(None, prefix), |
| 542 | Qwen4ExpPLEUnquantizedEmbeddingMethod, |
| 543 | ) |
| 544 | |
| 545 | |
| 546 | def test_ple_embedding_rejects_unsupported_quantization_configs() -> None: |
| 547 | prefix = "model.layers.1.ple.ple_embedding.ngram_embedding" |
| 548 | nvfp4_config = ModelOptNvFp4Config(exclude_modules=[]) |
| 549 | with pytest.raises(NotImplementedError, match="ModelOptNvFp4Config"): |
| 550 | Qwen4ExpPLEEmbeddingMethod.from_quant_config(nvfp4_config, prefix) |
| 551 | |
| 552 | online_fp8_config = OnlineQuantizationConfig( |
| 553 | QuantizationConfigArgs(linear="fp8_per_tensor") |
| 554 | ) |
| 555 | with pytest.raises(NotImplementedError, match="OnlineQuantizationConfig"): |
| 556 | Qwen4ExpPLEEmbeddingMethod.from_quant_config(online_fp8_config, prefix) |
| 557 | |
| 558 | |
| 559 | def test_ple_embedding_respects_modelopt_exclusion() -> None: |
| 560 | prefix = "model.layers.1.ple.ple_embedding.ngram_embedding" |
| 561 | quant_config = ModelOptNvFp4Config(exclude_modules=[prefix]) |
| 562 | |
| 563 | assert isinstance( |
| 564 | Qwen4ExpPLEEmbeddingMethod.from_quant_config(quant_config, prefix), |
| 565 | Qwen4ExpPLEUnquantizedEmbeddingMethod, |
| 566 | ) |
| 567 | |
| 568 | |
| 569 | def test_ple_embedding_dtype_overrides_modelopt_exclusion() -> None: |
| 570 | prefix = "model.layers.1.ple.ple_embedding.ngram_embedding" |
| 571 | quant_config = ModelOptNvFp4Config(exclude_modules=[prefix]) |
| 572 | |
| 573 | assert isinstance( |
| 574 | Qwen4ExpPLEEmbeddingMethod.from_quant_config( |
| 575 | quant_config, |
| 576 | prefix, |
| 577 | "float8_e4m3fn", |
| 578 | ), |
| 579 | Qwen4ExpPLEFp8EmbeddingMethod, |
| 580 | ) |
| 581 | |
| 582 | |
| 583 | def test_pinned_embedding_forward_finalizes_prefetched_output( |
| 584 | monkeypatch: pytest.MonkeyPatch, |
| 585 | ) -> None: |
| 586 | embedding = Qwen4ExpPLEPinnedHostEmbedding.__new__(Qwen4ExpPLEPinnedHostEmbedding) |
| 587 | nn.Module.__init__(embedding) |
| 588 | embedding._prefetch_buffer = torch.empty(4, 2, 3, dtype=torch.float8_e4m3fn) |
| 589 | embedding._output_dim = 6 |
| 590 | hidden_states = torch.zeros(2, 4, dtype=torch.bfloat16) |
| 591 | expected = torch.arange(12).reshape(2, 6).to(torch.float8_e4m3fn) |
| 592 | |
| 593 | def finalize_prefetched( |
| 594 | prefetch_output: torch.Tensor, |
| 595 | output: torch.Tensor, |
| 596 | ) -> None: |
| 597 | assert prefetch_output is embedding._prefetch_buffer |
| 598 | output.copy_(expected) |
| 599 | |
| 600 | monkeypatch.setattr( |
| 601 | embedding, |
| 602 | "_finalize_prefetch", |
| 603 | finalize_prefetched, |
| 604 | ) |
| 605 | |
| 606 | output = embedding(hidden_states) |
| 607 | |
| 608 | assert output.dtype == embedding._prefetch_buffer.dtype |
| 609 | assert torch.equal(output, expected) |
| 610 | |
| 611 | |
| 612 | def test_pinned_embedding_forward_requires_prior_start_prefetch() -> None: |
| 613 | """Forward before any start_prefetch must fail loudly, not read garbage.""" |
| 614 | embedding = Qwen4ExpPLEPinnedHostEmbedding.__new__(Qwen4ExpPLEPinnedHostEmbedding) |
| 615 | nn.Module.__init__(embedding) |
| 616 | embedding._prefetch_buffer = None |
| 617 | embedding._prefetch_stream = None |
| 618 | embedding._output_dim = 6 |
| 619 | hidden_states = torch.zeros(2, 4, dtype=torch.bfloat16) |
| 620 | |
| 621 | with pytest.raises(RuntimeError, match="prior start_prefetch"): |
| 622 | embedding(hidden_states) |
| 623 | |
| 624 | |
| 625 | def test_pinned_embedding_finalize_requires_prior_start_prefetch() -> None: |
| 626 | """_finalize_prefetch before any start_prefetch must fail loudly.""" |
| 627 | embedding = Qwen4ExpPLEPinnedHostEmbedding.__new__(Qwen4ExpPLEPinnedHostEmbedding) |
| 628 | nn.Module.__init__(embedding) |
| 629 | embedding._prefetch_buffer = None |
| 630 | embedding._prefetch_stream = None |
| 631 | embedding.tp_size = 1 |
| 632 | output = torch.zeros(2, 6, dtype=torch.bfloat16) |
| 633 | |
| 634 | with pytest.raises(RuntimeError, match="prior start_prefetch"): |
| 635 | embedding._finalize_prefetch(torch.empty(4, 2, 3), output) |
| 636 | |
| 637 | |
| 638 | def test_pinned_fp8_embedding_uses_int8_for_parallel_reduce() -> None: |
| 639 | embedding = Qwen4ExpPLEPinnedHostEmbedding.__new__(Qwen4ExpPLEPinnedHostEmbedding) |
| 640 | nn.Module.__init__(embedding) |
| 641 | embedding.tp_size = 2 |
| 642 | reduced_dtypes = [] |
| 643 | |
| 644 | def all_reduce(tensor: torch.Tensor) -> torch.Tensor: |
| 645 | reduced_dtypes.append(tensor.dtype) |
| 646 | return tensor.clone() |
| 647 | |
| 648 | embedding.parallel_group = SimpleNamespace(all_reduce=all_reduce) |
| 649 | embeddings = torch.arange(8).reshape(2, 4).to(torch.float8_e4m3fn) |
| 650 | |
| 651 | output = embedding._reduce_etp_embeddings(embeddings) |
| 652 | |
| 653 | assert reduced_dtypes == [torch.int8] |
| 654 | assert output.dtype == embeddings.dtype |
| 655 | torch.testing.assert_close(output.float(), embeddings.float()) |
| 656 | |
| 657 | |
| 658 | @pytest.mark.skipif(not torch.cuda.is_available(), reason="requires CUDA") |
| 659 | def test_ple_device_embedding_allocates_on_active_device( |
| 660 | monkeypatch: pytest.MonkeyPatch, |
| 661 | ) -> None: |
| 662 | _mock_etp_group(monkeypatch) |
| 663 | monkeypatch.setattr(embedding_module, "get_tensor_model_parallel_rank", lambda: 0) |
| 664 | monkeypatch.setattr( |
| 665 | embedding_module, "get_tensor_model_parallel_world_size", lambda: 1 |
| 666 | ) |
| 667 | monkeypatch.setattr(parameter_module, "get_tensor_model_parallel_rank", lambda: 0) |
| 668 | monkeypatch.setattr( |
| 669 | parameter_module, "get_tensor_model_parallel_world_size", lambda: 1 |
| 670 | ) |
| 671 | embedding_method = Qwen4ExpPLEUnquantizedEmbeddingMethod() |
| 672 | |
| 673 | with torch.device("cuda:0"): |
| 674 | embedding = Qwen4ExpPLEDeviceEmbedding( |
| 675 | 4, |
| 676 | 3, |
| 677 | params_dtype=torch.bfloat16, |
| 678 | padding_size=1, |
| 679 | prefix="test.ple_embedding", |
| 680 | embedding_method=embedding_method, |
| 681 | ) |
| 682 | |
| 683 | assert embedding.weight.device == torch.device("cuda:0") |
| 684 | |
| 685 | |
| 686 | @pytest.mark.skipif(not torch.cuda.is_available(), reason="requires CUDA") |
| 687 | @pytest.mark.parametrize("fp8_checkpoint", [False, True]) |
| 688 | def test_ple_pinned_embedding_loads_on_cpu_and_looks_up_through_uva( |
| 689 | monkeypatch: pytest.MonkeyPatch, |
| 690 | fp8_checkpoint: bool, |
| 691 | ) -> None: |
| 692 | _mock_etp_group(monkeypatch) |
| 693 | monkeypatch.setattr(embedding_module, "get_tensor_model_parallel_rank", lambda: 0) |
| 694 | monkeypatch.setattr( |
| 695 | embedding_module, "get_tensor_model_parallel_world_size", lambda: 1 |
| 696 | ) |
| 697 | monkeypatch.setattr(parameter_module, "get_tensor_model_parallel_rank", lambda: 0) |
| 698 | monkeypatch.setattr( |
| 699 | parameter_module, "get_tensor_model_parallel_world_size", lambda: 1 |
| 700 | ) |
| 701 | |
| 702 | with torch.device("cuda:0"): |
| 703 | embedding_method = ( |
| 704 | Qwen4ExpPLEFp8EmbeddingMethod() |
| 705 | if fp8_checkpoint |
| 706 | else Qwen4ExpPLEUnquantizedEmbeddingMethod() |
| 707 | ) |
| 708 | embedding = Qwen4ExpPLEPinnedHostEmbedding( |
| 709 | 4, |
| 710 | 3, |
| 711 | params_dtype=torch.bfloat16, |
| 712 | padding_size=1, |
| 713 | prefix="test.ple_embedding", |
| 714 | embedding_method=embedding_method, |
| 715 | ) |
| 716 | |
| 717 | storage_dtype = torch.float8_e4m3fn if fp8_checkpoint else torch.bfloat16 |
| 718 | loaded_weight = ( |
| 719 | torch.arange(12, dtype=torch.float32).reshape(4, 3).to(storage_dtype) |
| 720 | ) |
| 721 | copied = copy_ple_embedding_shard_( |
| 722 | embedding.weight, |
| 723 | loaded_weight, |
| 724 | checkpoint_start=0, |
| 725 | tp_start=0, |
| 726 | tp_end=4, |
| 727 | ) |
| 728 | input_ids = torch.tensor([[3, 0], [1, 2]], device="cuda:0") |
| 729 | output = embedding._lookup(input_ids) |
| 730 | if fp8_checkpoint: |
| 731 | embedding.weight_scale.data.fill_(0.25) |
| 732 | weight_ptr = embedding.weight.data_ptr() |
| 733 | embedding.quant_method.process_weights_after_loading(embedding) |
| 734 | assert embedding.weight.data_ptr() == weight_ptr |
| 735 | dequantized = embedding.dequantize(output, torch.bfloat16) |
| 736 | |
| 737 | assert copied == 4 |
| 738 | assert embedding.weight.device.type == "cpu" |
| 739 | assert embedding.weight.is_pinned() |
| 740 | assert output.dtype == storage_dtype |
| 741 | expected = loaded_weight[input_ids.cpu()].to(device="cuda:0") |
| 742 | torch.testing.assert_close(output.float(), expected.float(), rtol=0, atol=0) |
| 743 | expected_scale = 0.25 if fp8_checkpoint else 1.0 |
| 744 | torch.testing.assert_close( |
| 745 | dequantized, |
| 746 | expected.to(torch.bfloat16) * expected_scale, |
| 747 | rtol=0, |
| 748 | atol=0, |
| 749 | ) |
| 750 | |
| 751 | |
| 752 | @pytest.mark.skipif(not torch.cuda.is_available(), reason="requires CUDA") |
| 753 | def test_pinned_embedding_start_prefetch_allocates_lazily_and_feeds_forward( |
| 754 | monkeypatch: pytest.MonkeyPatch, |
| 755 | ) -> None: |
| 756 | """start_prefetch lazily allocates the stream/buffer, then forward consumes it. |
| 757 | |
| 758 | Exercises the full pinned-prefetch contract end to end: first call |
| 759 | allocates the side stream and buffer with the constructor geometry, |
| 760 | and forward finalizes the lookup into the returned output. |
| 761 | """ |
| 762 | _mock_etp_group(monkeypatch) |
| 763 | monkeypatch.setattr(embedding_module, "get_tensor_model_parallel_rank", lambda: 0) |
| 764 | monkeypatch.setattr( |
| 765 | embedding_module, "get_tensor_model_parallel_world_size", lambda: 1 |
| 766 | ) |
| 767 | monkeypatch.setattr(parameter_module, "get_tensor_model_parallel_rank", lambda: 0) |
| 768 | monkeypatch.setattr( |
| 769 | parameter_module, "get_tensor_model_parallel_world_size", lambda: 1 |
| 770 | ) |
| 771 | |
| 772 | with torch.device("cuda:0"): |
| 773 | embedding = Qwen4ExpPLEPinnedHostEmbedding( |
| 774 | 4, |
| 775 | 3, |
| 776 | params_dtype=torch.bfloat16, |
| 777 | padding_size=1, |
| 778 | prefix="test.ple_embedding", |
| 779 | embedding_method=Qwen4ExpPLEUnquantizedEmbeddingMethod(), |
| 780 | num_ngram_heads=2, |
| 781 | max_total_tokens=4, |
| 782 | ) |
| 783 | assert embedding._prefetch_buffer is None |
| 784 | assert embedding._prefetch_stream is None |
| 785 | |
| 786 | loaded_weight = torch.arange(12, dtype=torch.float32).reshape(4, 3) |
| 787 | loaded_weight = loaded_weight.to(torch.bfloat16) |
| 788 | copy_ple_embedding_shard_( |
| 789 | embedding.weight, |
| 790 | loaded_weight, |
| 791 | checkpoint_start=0, |
| 792 | tp_start=0, |
| 793 | tp_end=4, |
| 794 | ) |
| 795 | |
| 796 | ngram_ids = torch.tensor([[3, 0], [1, 2]], device="cuda:0") |
| 797 | hidden_states = torch.zeros(2, 4, dtype=torch.bfloat16, device="cuda:0") |
| 798 | embedding.start_prefetch(hidden_states, ngram_ids) |
| 799 | |
| 800 | buffer = embedding._prefetch_buffer |
| 801 | assert buffer is not None |
| 802 | assert buffer.shape == (4, 2, 3) |
| 803 | assert buffer.dtype == torch.bfloat16 |
| 804 | assert buffer.device.type == "cuda" |
| 805 | assert embedding._prefetch_stream is not None |
| 806 | |
| 807 | output = embedding(hidden_states) |
| 808 | |
| 809 | expected = loaded_weight[ngram_ids.cpu()].to(device="cuda:0").flatten(-2) |
| 810 | torch.testing.assert_close(output, expected, rtol=0, atol=0) |
| 811 | |
| 812 | |
| 813 | def test_ple_fp8_embedding_supports_mixed_precision_config() -> None: |
| 814 | prefix = "model.language_model.layers.1.ple.ple_embedding.ngram_embedding" |
| 815 | quant_config = ModelOptMixedPrecisionConfig.from_config( |
| 816 | { |
| 817 | "quantization": { |
| 818 | "quant_algo": "MIXED_PRECISION", |
| 819 | "exclude_modules": [], |
| 820 | "group_size": 16, |
| 821 | "quantized_layers": { |
| 822 | prefix: {"quant_algo": "FP8"}, |
| 823 | "model.language_model.layers.2.moe.gate_proj": { |
| 824 | "quant_algo": "NVFP4" |
| 825 | }, |
| 826 | }, |
| 827 | } |
| 828 | } |
| 829 | ) |
| 830 | |
| 831 | assert isinstance( |
| 832 | Qwen4ExpPLEEmbeddingMethod.from_quant_config(quant_config, prefix), |
| 833 | Qwen4ExpPLEFp8EmbeddingMethod, |
| 834 | ) |
| 835 | assert isinstance( |
| 836 | Qwen4ExpPLEEmbeddingMethod.from_quant_config( |
| 837 | quant_config, |
| 838 | "model.language_model.layers.2.moe.gate_proj", |
| 839 | ), |
| 840 | Qwen4ExpPLEUnquantizedEmbeddingMethod, |
| 841 | ) |
| 842 | |
| 843 | |
| 844 | def test_dilated_ple_spec_state_rolls_back_before_next_forward() -> None: |
| 845 | conv_state_len = 6 |
| 846 | dilation = 2 |
| 847 | |
| 848 | conv_weights = torch.tensor([[0.25, -0.5, 0.75, 1.0]]) |
| 849 | conv_state = torch.zeros(2, 1, 9) |
| 850 | conv_state[1] = torch.arange(1, 10, dtype=torch.float32).reshape(1, 9) |
| 851 | first_inputs = torch.tensor([[10.0], [20.0], [30.0], [40.0]]) |
| 852 | initial_state = conv_state[1:].clone() |
| 853 | first_history = torch.cat( |
| 854 | (initial_state[..., :conv_state_len], first_inputs.T.unsqueeze(0)), |
| 855 | dim=-1, |
| 856 | ) |
| 857 | |
| 858 | graph_padded_inputs = F.pad(first_inputs, (0, 0, 0, 4)) |
| 859 | first_output = _short_conv_dilated_spec_pytorch( |
| 860 | graph_padded_inputs, |
| 861 | conv_state, |
| 862 | conv_weights, |
| 863 | torch.tensor([1, 0]), |
| 864 | torch.tensor([0, 4, 4]), |
| 865 | torch.tensor([1, 0]), |
| 866 | spec_query_len=4, |
| 867 | conv_state_len=conv_state_len, |
| 868 | dilation=dilation, |
| 869 | ) |
| 870 | |
| 871 | expected_first_output = F.silu( |
| 872 | F.conv1d( |
| 873 | first_history, |
| 874 | conv_weights.unsqueeze(1), |
| 875 | groups=1, |
| 876 | dilation=dilation, |
| 877 | ) |
| 878 | ).transpose(1, 2)[0] |
| 879 | expected_first_state = first_history[..., 1:10] |
| 880 | torch.testing.assert_close(first_output[:4], expected_first_output) |
| 881 | assert torch.count_nonzero(first_output[4:]) == 0 |
| 882 | assert torch.count_nonzero(conv_state[0]) == 0 |
| 883 | torch.testing.assert_close(conv_state[1:], expected_first_state) |
| 884 | |
| 885 | second_inputs = torch.tensor([[50.0], [60.0]]) |
| 886 | rollback_state = expected_first_state[..., 1:7] |
| 887 | padded_second_inputs = F.pad(second_inputs.T.unsqueeze(0), (0, 2)) |
| 888 | second_history = torch.cat((rollback_state, padded_second_inputs), dim=-1) |
| 889 | expected_second_state = expected_first_state.clone() |
| 890 | expected_second_state[..., :7] = second_history[..., 1:8] |
| 891 | |
| 892 | second_output = _short_conv_dilated_spec_pytorch( |
| 893 | second_inputs, |
| 894 | conv_state, |
| 895 | conv_weights, |
| 896 | torch.tensor([1]), |
| 897 | torch.tensor([0, 2]), |
| 898 | torch.tensor([2]), |
| 899 | spec_query_len=4, |
| 900 | conv_state_len=conv_state_len, |
| 901 | dilation=dilation, |
| 902 | ) |
| 903 | |
| 904 | expected_second_output = F.silu( |
| 905 | F.conv1d( |
| 906 | second_history, |
| 907 | conv_weights.unsqueeze(1), |
| 908 | groups=1, |
| 909 | dilation=dilation, |
| 910 | ) |
| 911 | ).transpose(1, 2)[0, :2] |
| 912 | torch.testing.assert_close(second_output, expected_second_output) |
| 913 | torch.testing.assert_close(conv_state[1:], expected_second_state) |
| 914 | |
| 915 | |
| 916 | def _reference_ngram_ids( |
| 917 | input_ids: torch.Tensor, |
| 918 | query_start_loc: torch.Tensor, |
| 919 | ngram_context: torch.Tensor, |
| 920 | layer_multipliers: torch.Tensor, |
| 921 | ngram_heads_vocab_sizes: torch.Tensor, |
| 922 | ngram_heads_offsets: torch.Tensor, |
| 923 | eos_token_id: int, |
| 924 | heads_per_ngram: int, |
| 925 | ) -> torch.Tensor: |
| 926 | tokens = input_ids.tolist() |
| 927 | starts = query_start_loc.tolist() |
| 928 | contexts = ngram_context.tolist() |
| 929 | multipliers = layer_multipliers.cpu() |
| 930 | sizes = ngram_heads_vocab_sizes.cpu() |
| 931 | offsets = ngram_heads_offsets.cpu() |
| 932 | rows = [] |
| 933 | context_len = len(contexts[0]) |
| 934 | for req, (start, end) in enumerate(zip(starts, starts[1:])): |
| 935 | history = contexts[req] + tokens[start:end] |
| 936 | for pos in range(context_len, len(history)): |
| 937 | shifted = [history[pos]] |
| 938 | crossed_eos = False |
| 939 | for shift in range(1, context_len + 1): |
| 940 | token = eos_token_id if crossed_eos else history[pos - shift] |
| 941 | shifted.append(token) |
| 942 | crossed_eos |= token == eos_token_id |
| 943 | |
| 944 | row = [] |
| 945 | mixed = torch.tensor(shifted[0], dtype=torch.int64) * multipliers[0] |
| 946 | for ngram_order in range(2, context_len + 2): |
| 947 | shift = ngram_order - 1 |
| 948 | mixed ^= shifted[shift] * multipliers[shift] |
| 949 | head_start = (ngram_order - 2) * heads_per_ngram |
| 950 | for head in range(head_start, head_start + heads_per_ngram): |
| 951 | row.append(torch.remainder(mixed, sizes[head]) + offsets[head]) |
| 952 | rows.append(torch.stack(row)) |
| 953 | return torch.stack(rows).to(input_ids.device) |
| 954 | |
| 955 | |
| 956 | _NGRAM_MULTIPLIERS = ( |
| 957 | 18_014_398_509_481_983, |
| 958 | 17_114_398_509_481_981, |
| 959 | 16_214_398_509_481_979, |
| 960 | 15_314_398_509_481_977, |
| 961 | ) |
| 962 | _NGRAM_HEADS_VOCAB_SIZES = ( |
| 963 | 101, |
| 964 | 103, |
| 965 | 107, |
| 966 | 109, |
| 967 | 113, |
| 968 | 127, |
| 969 | 131, |
| 970 | 137, |
| 971 | 139, |
| 972 | 149, |
| 973 | 151, |
| 974 | 157, |
| 975 | 163, |
| 976 | 167, |
| 977 | 173, |
| 978 | 179, |
| 979 | 181, |
| 980 | 191, |
| 981 | 193, |
| 982 | 197, |
| 983 | 199, |
| 984 | 211, |
| 985 | 223, |
| 986 | 227, |
| 987 | ) |
| 988 | _NGRAM_EOS_TOKEN_ID = 251 |
| 989 | _NGRAM_HEADS_PER_NGRAM = 8 |
| 990 | |
| 991 | |
| 992 | def _ngram_hash_params(device: torch.device, context_len: int) -> dict: |
| 993 | num_heads = context_len * _NGRAM_HEADS_PER_NGRAM |
| 994 | sizes = torch.tensor( |
| 995 | _NGRAM_HEADS_VOCAB_SIZES[:num_heads], dtype=torch.int64, device=device |
| 996 | ) |
| 997 | offsets = torch.zeros_like(sizes) |
| 998 | offsets[1:] = torch.cumsum(sizes, dim=0)[:-1] |
| 999 | return { |
| 1000 | "layer_multipliers": torch.tensor( |
| 1001 | _NGRAM_MULTIPLIERS[: context_len + 1], |
| 1002 | dtype=torch.int64, |
| 1003 | device=device, |
| 1004 | ), |
| 1005 | "ngram_heads_vocab_sizes": sizes, |
| 1006 | "ngram_heads_offsets": offsets, |
| 1007 | "eos_token_id": _NGRAM_EOS_TOKEN_ID, |
| 1008 | "heads_per_ngram": _NGRAM_HEADS_PER_NGRAM, |
| 1009 | } |
| 1010 | |
| 1011 | |
| 1012 | @pytest.mark.skipif(not torch.cuda.is_available(), reason="fused PLE needs CUDA") |
| 1013 | @pytest.mark.parametrize( |
| 1014 | ("query_lens", "eos_offsets", "contexts", "first_token_id"), |
| 1015 | [ |
| 1016 | ([1], [], [[11, 12]], 20), |
| 1017 | ([3], [1], [[11]], 20), |
| 1018 | ([4, 4], [0, 7], [[11, 12], [13, 14]], 20), |
| 1019 | ([4, 0, 3], [], [[11, 12], [13, 14], [15, 16]], 20), |
| 1020 | ( |
| 1021 | [3, 2, 0, 0], |
| 1022 | [], |
| 1023 | [ |
| 1024 | [11, 12], |
| 1025 | [13, 14], |
| 1026 | [_NGRAM_EOS_TOKEN_ID, _NGRAM_EOS_TOKEN_ID], |
| 1027 | [_NGRAM_EOS_TOKEN_ID, _NGRAM_EOS_TOKEN_ID], |
| 1028 | ], |
| 1029 | 20, |
| 1030 | ), |
| 1031 | ( |
| 1032 | [1, 33, 2], |
| 1033 | [5, 32], |
| 1034 | [[_NGRAM_EOS_TOKEN_ID, 11], [12, 13], [14, _NGRAM_EOS_TOKEN_ID]], |
| 1035 | 20, |
| 1036 | ), |
| 1037 | ( |
| 1038 | [5, 12, 16, 1, 16, 17], |
| 1039 | [10, 40], |
| 1040 | [[11, 12], [13, 14], [15, 16], [17, 18], [19, 20], [21, 22]], |
| 1041 | 20, |
| 1042 | ), |
| 1043 | ([5, 3], [3], [[11, 12, 13], [14, 15, 16]], 20), |
| 1044 | ( |
| 1045 | [4, 0, 3], |
| 1046 | [], |
| 1047 | [ |
| 1048 | [200_000, 200_001], |
| 1049 | [250_000, 250_001], |
| 1050 | [300_000, 300_001], |
| 1051 | ], |
| 1052 | 350_000, |
| 1053 | ), |
| 1054 | ([3], [], [[1_000_000_000, 1_000_000_001]], 1_000_000_002), |
| 1055 | ], |
| 1056 | ids=[ |
| 1057 | "single-token", |
| 1058 | "bigram-only", |
| 1059 | "power-of-two", |
| 1060 | "empty-request", |
| 1061 | "trailing-padded-requests", |
| 1062 | "three-requests", |
| 1063 | "six-requests", |
| 1064 | "four-gram", |
| 1065 | "int64-overflow", |
| 1066 | "large-int32-ids", |
| 1067 | ], |
| 1068 | ) |
| 1069 | def test_fused_ngram_ids_correctness( |
| 1070 | query_lens: list[int], |
| 1071 | eos_offsets: list[int], |
| 1072 | contexts: list[list[int]], |
| 1073 | first_token_id: int, |
| 1074 | ) -> None: |
| 1075 | from vllm.models.qwen4_exp.nvidia.ops.ple import ple_ngram_ids |
| 1076 | |
| 1077 | device = torch.device("cuda") |
| 1078 | query_start_loc = torch.tensor( |
| 1079 | [0, *accumulate(query_lens)], dtype=torch.int32, device=device |
| 1080 | ) |
| 1081 | input_ids = torch.arange( |
| 1082 | first_token_id, |
| 1083 | first_token_id + sum(query_lens), |
| 1084 | dtype=torch.int32, |
| 1085 | device=device, |
| 1086 | ) |
| 1087 | if eos_offsets: |
| 1088 | input_ids[eos_offsets] = _NGRAM_EOS_TOKEN_ID |
| 1089 | ngram_context = torch.tensor(contexts, dtype=torch.int32, device=device) |
| 1090 | params = _ngram_hash_params(device, ngram_context.shape[1]) |
| 1091 | |
| 1092 | expected = _reference_ngram_ids(input_ids, query_start_loc, ngram_context, **params) |
| 1093 | actual = ple_ngram_ids(input_ids, query_start_loc, ngram_context, **params) |
| 1094 | |
| 1095 | assert torch.equal(actual, expected) |
| 1096 | offsets = params["ngram_heads_offsets"] |
| 1097 | sizes = params["ngram_heads_vocab_sizes"] |
| 1098 | assert torch.all((actual >= offsets) & (actual < offsets + sizes)) |
| 1099 | |
| 1100 | |
| 1101 | @pytest.mark.skipif(not torch.cuda.is_available(), reason="fused PLE needs CUDA") |
| 1102 | def test_ngram_prefetch_ids_outlive_start_prefetch() -> None: |
| 1103 | """Eager breaks hand the side-stream lookup weak refs, so the ids must stay |
| 1104 | valid after start_prefetch returns rather than be reused from the pool.""" |
| 1105 | device = torch.device("cuda") |
| 1106 | query_start_loc = torch.tensor([0, 3, 5], dtype=torch.int32, device=device) |
| 1107 | input_ids = torch.arange(20, 25, dtype=torch.int32, device=device) |
| 1108 | ngram_context = torch.tensor([[11, 12], [13, 14]], dtype=torch.int32, device=device) |
| 1109 | params = _ngram_hash_params(device, ngram_context.shape[1]) |
| 1110 | module = Qwen4ExpNGramEmbedding.__new__(Qwen4ExpNGramEmbedding) |
| 1111 | nn.Module.__init__(module) |
| 1112 | for name in ("layer_multipliers", "ngram_heads_vocab_sizes", "ngram_heads_offsets"): |
| 1113 | module.register_buffer(name, params[name]) |
| 1114 | module.eos_token_id = params["eos_token_id"] |
| 1115 | module.heads_per_ngram = params["heads_per_ngram"] |
| 1116 | num_heads = params["ngram_heads_vocab_sizes"].numel() |
| 1117 | module._prefetch_ids = torch.empty(8, num_heads, dtype=torch.long, device=device) |
| 1118 | prefetched: list[torch.Tensor] = [] |
| 1119 | module.ngram_embedding = SimpleNamespace( |
| 1120 | supports_prefetch=True, |
| 1121 | start_prefetch=lambda _, ids: prefetched.append(weak_ref_tensor(ids)), |
| 1122 | ) |
| 1123 | expected = _reference_ngram_ids(input_ids, query_start_loc, ngram_context, **params) |
| 1124 | |
| 1125 | # A fresh pool makes the next same-size allocation reuse any freed ids, |
| 1126 | # like a later CUDA graph segment would. |
| 1127 | pool = torch.cuda.MemPool() |
| 1128 | with torch.cuda.use_mem_pool(pool): |
| 1129 | module.start_prefetch(None, input_ids, query_start_loc, ngram_context) |
| 1130 | torch.full_like(expected, -1) |
| 1131 | |
| 1132 | assert torch.equal(prefetched[0], expected) |
| 1133 | |
| 1134 | |
| 1135 | def _short_conv_dilated_decode_pytorch( |
| 1136 | x_d: torch.Tensor, |
| 1137 | conv_state: torch.Tensor, |
| 1138 | conv_weights: torch.Tensor, |
| 1139 | state_indices_tensor_d: torch.Tensor, |
| 1140 | has_initial_states_d: torch.Tensor | None, |
| 1141 | *, |
| 1142 | conv_state_len: int, |
| 1143 | dilation: int, |
| 1144 | ) -> torch.Tensor: |
| 1145 | state_indices = state_indices_tensor_d.to( |
| 1146 | device=conv_state.device, dtype=torch.int64 |
| 1147 | ) |
| 1148 | # FULL cudagraph padded decode rows use NULL_BLOCK_ID. Remap them to |
| 1149 | # slot 0 for a safe gather, then zero output and skip write-back. |
| 1150 | valid_state = state_indices != NULL_BLOCK_ID |
| 1151 | state_indices = torch.where( |
| 1152 | valid_state, state_indices, torch.zeros_like(state_indices) |
| 1153 | ) |
| 1154 | if has_initial_states_d is None: |
| 1155 | has_initial_state = valid_state |
| 1156 | else: |
| 1157 | if has_initial_states_d.numel() < state_indices_tensor_d.numel(): |
| 1158 | raise ValueError( |
| 1159 | "has_initial_states_d size mismatch: " |
| 1160 | f"got {has_initial_states_d.numel()}, " |
| 1161 | f"need >= {state_indices_tensor_d.numel()}." |
| 1162 | ) |
| 1163 | has_initial_state = has_initial_states_d[: state_indices_tensor_d.numel()].to( |
| 1164 | device=conv_state.device, dtype=torch.bool |
| 1165 | ) |
| 1166 | has_initial_state = has_initial_state & valid_state |
| 1167 | |
| 1168 | cached_state = conv_state.index_select(0, state_indices) |
| 1169 | state = cached_state[..., :conv_state_len].to(x_d.dtype) |
| 1170 | if conv_state_len > 0: |
| 1171 | initial_state = torch.where( |
| 1172 | has_initial_state.view(-1, 1, 1), |
| 1173 | state, |
| 1174 | torch.zeros_like(state), |
| 1175 | ) |
| 1176 | history = torch.cat((initial_state, x_d.unsqueeze(-1)), dim=-1) |
| 1177 | else: |
| 1178 | history = x_d.unsqueeze(-1) |
| 1179 | |
| 1180 | conv_output = F.conv1d( |
| 1181 | history, |
| 1182 | conv_weights.unsqueeze(1).contiguous(), |
| 1183 | groups=history.size(1), |
| 1184 | dilation=dilation, |
| 1185 | ).squeeze(-1) |
| 1186 | output = F.silu(conv_output) |
| 1187 | output = output * valid_state.view(-1, 1).to(output.dtype) |
| 1188 | |
| 1189 | if conv_state_len > 0: |
| 1190 | next_state = history[..., -conv_state_len:] |
| 1191 | # Padded rows are remapped to the reserved null slot. Preserve its |
| 1192 | # existing value while writing the new states for valid rows. |
| 1193 | existing_base_state = cached_state[..., :conv_state_len] |
| 1194 | safe_next_state = torch.where( |
| 1195 | valid_state.view(-1, 1, 1), |
| 1196 | next_state.to(conv_state.dtype), |
| 1197 | existing_base_state, |
| 1198 | ) |
| 1199 | cached_state[..., :conv_state_len] = safe_next_state |
| 1200 | conv_state.index_copy_(0, state_indices, cached_state) |
| 1201 | |
| 1202 | return output |
| 1203 | |
| 1204 | |
| 1205 | def _short_conv_dilated_prefill_pytorch( |
| 1206 | x_p: torch.Tensor, |
| 1207 | metadata: PleShortConvAttentionMetadata, |
| 1208 | conv_state: torch.Tensor, |
| 1209 | conv_weights: torch.Tensor, |
| 1210 | state_indices_tensor_p: torch.Tensor, |
| 1211 | num_prefills: int, |
| 1212 | num_decode_tokens: int, |
| 1213 | num_prefill_tokens: int, |
| 1214 | *, |
| 1215 | conv_state_len: int, |
| 1216 | dilation: int, |
| 1217 | ) -> torch.Tensor: |
| 1218 | # ``non_spec_query_start_loc`` covers the non-spec (decode + prefill) |
| 1219 | # requests and equals ``query_start_loc`` when spec-decode is inactive. |
| 1220 | non_spec_query_start_loc = metadata.non_spec_query_start_loc |
| 1221 | if non_spec_query_start_loc is None: |
| 1222 | raise ValueError("query_start_loc is required for prefill short-conv") |
| 1223 | query_start_loc_p = ( |
| 1224 | non_spec_query_start_loc[-num_prefills - 1 :] - num_decode_tokens |
| 1225 | ) |
| 1226 | # The metadata builder guarantees that the prefill query offsets start |
| 1227 | # at 0 and end at num_prefill_tokens. Avoid reading those values here, |
| 1228 | # since doing so would force a device-to-host synchronization. |
| 1229 | has_initial_states_p = metadata.has_initial_states_p |
| 1230 | if has_initial_states_p is None: |
| 1231 | raise ValueError("has_initial_states_p is required for prefill short-conv") |
| 1232 | |
| 1233 | output = torch.empty_like(x_p) |
| 1234 | q_starts = query_start_loc_p.to(torch.int64) |
| 1235 | if state_indices_tensor_p.numel() < num_prefills: |
| 1236 | raise ValueError( |
| 1237 | "state_indices_tensor_p size mismatch: " |
| 1238 | f"got {state_indices_tensor_p.numel()}, " |
| 1239 | f"need >= {num_prefills}." |
| 1240 | ) |
| 1241 | if has_initial_states_p.numel() < num_prefills: |
| 1242 | raise ValueError( |
| 1243 | "has_initial_states_p size mismatch: " |
| 1244 | f"got {has_initial_states_p.numel()}, " |
| 1245 | f"need >= {num_prefills}." |
| 1246 | ) |
| 1247 | if num_prefills == 0 or x_p.numel() == 0: |
| 1248 | return output |
| 1249 | lengths = q_starts[1:] - q_starts[:-1] |
| 1250 | # Use the CPU-computed packing width from the metadata builder instead |
| 1251 | # of synchronizing on lengths.max(). |
| 1252 | max_len = metadata.max_prefill_query_len |
| 1253 | if max_len <= 0: |
| 1254 | return output |
| 1255 | |
| 1256 | hidden_size = x_p.shape[1] |
| 1257 | positions = torch.arange(num_prefill_tokens, device=x_p.device, dtype=torch.int64) |
| 1258 | req_indices = torch.searchsorted(q_starts[1:], positions, right=True) |
| 1259 | col_indices = positions - q_starts[req_indices] |
| 1260 | |
| 1261 | packed_tokens = x_p.new_zeros((num_prefills, max_len, hidden_size)) |
| 1262 | packed_tokens[req_indices, col_indices] = x_p |
| 1263 | packed_tokens = packed_tokens.transpose(1, 2).contiguous() |
| 1264 | |
| 1265 | state_indices = state_indices_tensor_p[:num_prefills].to( |
| 1266 | device=conv_state.device, dtype=torch.int64 |
| 1267 | ) |
| 1268 | valid_state = state_indices != NULL_BLOCK_ID |
| 1269 | state_indices = torch.where( |
| 1270 | valid_state, state_indices, torch.zeros_like(state_indices) |
| 1271 | ) |
| 1272 | has_initial = has_initial_states_p[:num_prefills].to( |
| 1273 | device=conv_state.device, dtype=torch.bool |
| 1274 | ) |
| 1275 | if conv_state_len > 0: |
| 1276 | if conv_state.shape[0] == 0: |
| 1277 | state = conv_state.new_zeros( |
| 1278 | (num_prefills, hidden_size, conv_state_len), |
| 1279 | dtype=x_p.dtype, |
| 1280 | ) |
| 1281 | else: |
| 1282 | state = conv_state.index_select(0, state_indices)[..., :conv_state_len].to( |
| 1283 | x_p.dtype |
| 1284 | ) |
| 1285 | use_initial_mask = (valid_state & has_initial).view(num_prefills, 1, 1) |
| 1286 | initial_state = torch.where( |
| 1287 | use_initial_mask, |
| 1288 | state, |
| 1289 | torch.zeros_like(state), |
| 1290 | ) |
| 1291 | history = torch.cat((initial_state, packed_tokens), dim=-1) |
| 1292 | else: |
| 1293 | history = packed_tokens |
| 1294 | |
| 1295 | conv_output = F.conv1d( |
| 1296 | history, |
| 1297 | conv_weights.unsqueeze(1).contiguous(), |
| 1298 | groups=history.size(1), |
| 1299 | dilation=dilation, |
| 1300 | ) |
| 1301 | conv_output = F.silu(conv_output).transpose(1, 2).contiguous() |
| 1302 | |
| 1303 | token_positions = torch.arange(max_len, device=x_p.device, dtype=torch.int64) |
| 1304 | valid_tokens = token_positions.view(1, max_len) < lengths.view(num_prefills, 1) |
| 1305 | valid_output_mask = valid_tokens & valid_state.to(device=x_p.device).view( |
| 1306 | num_prefills, 1 |
| 1307 | ) |
| 1308 | conv_output.masked_fill_(~valid_output_mask.unsqueeze(-1), 0) |
| 1309 | output.copy_(conv_output[req_indices, col_indices]) |
| 1310 | |
| 1311 | if conv_state_len > 0 and conv_state.shape[0] > 0: |
| 1312 | state_starts = lengths.to(device=history.device, dtype=torch.int64).view( |
| 1313 | num_prefills, 1, 1 |
| 1314 | ) |
| 1315 | state_offsets = torch.arange( |
| 1316 | conv_state_len, device=history.device, dtype=torch.int64 |
| 1317 | ).view(1, 1, conv_state_len) |
| 1318 | next_state = history.gather( |
| 1319 | dim=2, |
| 1320 | index=(state_starts + state_offsets).expand(-1, history.size(1), -1), |
| 1321 | ) |
| 1322 | # Write back without a host synchronization. Valid, non-empty rows |
| 1323 | # receive their new state; padding and zero-length rows keep the |
| 1324 | # current cache value. |
| 1325 | existing_state = conv_state.index_select(0, state_indices) |
| 1326 | existing_base_state = existing_state[..., :conv_state_len] |
| 1327 | update_mask = valid_state & (lengths.to(device=conv_state.device) > 0) |
| 1328 | safe_next_state = torch.where( |
| 1329 | update_mask.view(num_prefills, 1, 1), |
| 1330 | next_state.to(conv_state.dtype), |
| 1331 | existing_base_state, |
| 1332 | ) |
| 1333 | existing_state[..., :conv_state_len] = safe_next_state |
| 1334 | conv_state.index_copy_(0, state_indices, existing_state) |
| 1335 | return output |
| 1336 | |
| 1337 | |
| 1338 | def _short_conv_dilated_spec_pytorch( |
| 1339 | x_spec: torch.Tensor, |
| 1340 | conv_state: torch.Tensor, |
| 1341 | conv_weights: torch.Tensor, |
| 1342 | spec_state_indices_tensor: torch.Tensor, |
| 1343 | spec_query_start_loc: torch.Tensor, |
| 1344 | num_accepted_tokens: torch.Tensor, |
| 1345 | spec_query_len: int, |
| 1346 | *, |
| 1347 | conv_state_len: int, |
| 1348 | dilation: int, |
| 1349 | ) -> torch.Tensor: |
| 1350 | """Dilated short-conv for speculative-decode (MTP) requests. |
| 1351 | |
| 1352 | Each spec request feeds multiple (draft + 1) query tokens. The conv |
| 1353 | outputs are computed causally after rolling back the previous draft |
| 1354 | state by ``num_accepted_tokens - 1``. The current candidate inputs stay |
| 1355 | in the extended cache for the next forward, matching |
| 1356 | ``causal_conv1d_update``. |
| 1357 | |
| 1358 | ``spec_query_len`` (== num_speculative_tokens + 1) is the maximum query |
| 1359 | length and is a Python int, so no host synchronization is needed; this |
| 1360 | keeps the path safe for full CUDA-graph capture/replay where the buffers |
| 1361 | are padded at the request level. |
| 1362 | """ |
| 1363 | num_reqs = spec_state_indices_tensor.numel() |
| 1364 | hidden_size = x_spec.size(-1) |
| 1365 | # Use a fixed packing width instead of synchronizing on lengths.max(). |
| 1366 | max_len = spec_query_len |
| 1367 | # Full CUDA graphs can pad these buffers. Only the first num_reqs |
| 1368 | # accepted-token counts belong to actual speculative requests. |
| 1369 | num_accepted_tokens = num_accepted_tokens[:num_reqs] |
| 1370 | q_starts = spec_query_start_loc[: num_reqs + 1].to(torch.int64) |
| 1371 | # Keep the number of real speculative tokens on the device. |
| 1372 | total_real_tokens = q_starts[num_reqs] |
| 1373 | |
| 1374 | state_indices = spec_state_indices_tensor.to( |
| 1375 | device=conv_state.device, dtype=torch.int64 |
| 1376 | ) |
| 1377 | valid_state = state_indices != NULL_BLOCK_ID |
| 1378 | state_indices = torch.where( |
| 1379 | valid_state, state_indices, torch.zeros_like(state_indices) |
| 1380 | ) |
| 1381 | positions = torch.arange(x_spec.size(0), device=x_spec.device, dtype=torch.int64) |
| 1382 | # Route graph-padded token rows to the discarded dummy request so that |
| 1383 | # they cannot overwrite real packed data. |
| 1384 | req_indices = torch.searchsorted(q_starts[1:], positions, right=True) |
| 1385 | valid_tokens = (positions < total_real_tokens) & (req_indices < num_reqs) |
| 1386 | clamped_req_indices = req_indices.clamp_max(max(num_reqs - 1, 0)) |
| 1387 | col_indices = (positions - q_starts[clamped_req_indices]).clamp_(0, max_len - 1) |
| 1388 | pack_req_indices = torch.where( |
| 1389 | valid_tokens, |
| 1390 | clamped_req_indices, |
| 1391 | torch.full_like(req_indices, num_reqs), |
| 1392 | ) |
| 1393 | pack_col_indices = torch.where( |
| 1394 | valid_tokens, col_indices, torch.zeros_like(col_indices) |
| 1395 | ) |
| 1396 | |
| 1397 | # The last request row is the dummy sink for graph padding. |
| 1398 | packed = x_spec.new_zeros((num_reqs + 1, max_len, hidden_size)) |
| 1399 | packed[pack_req_indices, pack_col_indices] = x_spec |
| 1400 | packed = packed.transpose(1, 2).contiguous() |
| 1401 | |
| 1402 | if conv_state_len > 0: |
| 1403 | cached_state = conv_state.index_select(0, state_indices) |
| 1404 | rollback_offsets = num_accepted_tokens.to( |
| 1405 | device=conv_state.device, dtype=torch.int64 |
| 1406 | ).sub(1) |
| 1407 | rollback_offsets = torch.where( |
| 1408 | valid_state, |
| 1409 | rollback_offsets.clamp_(0, max_len - 1), |
| 1410 | torch.zeros_like(rollback_offsets), |
| 1411 | ) |
| 1412 | state_offsets = torch.arange( |
| 1413 | conv_state_len, device=conv_state.device, dtype=torch.int64 |
| 1414 | ).view(1, 1, conv_state_len) |
| 1415 | rollback_indices = rollback_offsets.view(-1, 1, 1) + state_offsets |
| 1416 | state = cached_state.gather(2, rollback_indices.expand(-1, hidden_size, -1)).to( |
| 1417 | x_spec.dtype |
| 1418 | ) |
| 1419 | state = torch.where( |
| 1420 | valid_state.view(num_reqs, 1, 1), |
| 1421 | state, |
| 1422 | torch.zeros_like(state), |
| 1423 | ) |
| 1424 | # Append a zeroed dummy-row state to match the [num_reqs + 1] pack. |
| 1425 | dummy_state = state.new_zeros((1, hidden_size, conv_state_len)) |
| 1426 | state_full = torch.cat((state, dummy_state), dim=0) |
| 1427 | history = torch.cat((state_full, packed), dim=-1) |
| 1428 | else: |
| 1429 | history = packed |
| 1430 | |
| 1431 | conv_output = F.conv1d( |
| 1432 | history, |
| 1433 | conv_weights.unsqueeze(1).contiguous(), |
| 1434 | groups=history.size(1), |
| 1435 | dilation=dilation, |
| 1436 | ) |
| 1437 | conv_output = F.silu(conv_output).transpose(1, 2).contiguous() |
| 1438 | |
| 1439 | output = conv_output[pack_req_indices, pack_col_indices] |
| 1440 | output = output * valid_tokens.view(-1, 1).to(output.dtype) |
| 1441 | |
| 1442 | # Keep all current candidate inputs in the extended state. On the next |
| 1443 | # target forward, ``num_accepted_tokens - 1`` selects the rollback |
| 1444 | # window before processing the newly scheduled tokens. |
| 1445 | if conv_state_len > 0: |
| 1446 | state_capacity = conv_state_len + max_len - 1 |
| 1447 | if conv_state.size(-1) < state_capacity: |
| 1448 | raise RuntimeError( |
| 1449 | "PLE short-conv cache cannot retain speculative tokens: " |
| 1450 | f"got {conv_state.size(-1)}, need {state_capacity}." |
| 1451 | ) |
| 1452 | candidate_state = history[:num_reqs, :, 1 : state_capacity + 1] |
| 1453 | query_lengths = q_starts[1:] - q_starts[:-1] |
| 1454 | state_positions = torch.arange( |
| 1455 | state_capacity, device=history.device, dtype=torch.int64 |
| 1456 | ).view(1, 1, state_capacity) |
| 1457 | update_lengths = (conv_state_len + query_lengths - 1).view(num_reqs, 1, 1) |
| 1458 | update_mask = valid_state.view(num_reqs, 1, 1) & ( |
| 1459 | state_positions < update_lengths |
| 1460 | ) |
| 1461 | existing_state = cached_state[..., :state_capacity] |
| 1462 | next_state = torch.where( |
| 1463 | update_mask, |
| 1464 | candidate_state.to(conv_state.dtype), |
| 1465 | existing_state, |
| 1466 | ) |
| 1467 | cached_state[..., :state_capacity] = next_state |
| 1468 | conv_state.index_copy_(0, state_indices, cached_state) |
| 1469 | |
| 1470 | return output |
| 1471 | |
| 1472 | |
| 1473 | def _short_conv_dilated_dispatch_pytorch( |
| 1474 | inputs: torch.Tensor, |
| 1475 | residual: torch.Tensor, |
| 1476 | metadata: PleShortConvAttentionMetadata, |
| 1477 | conv_state: torch.Tensor, |
| 1478 | conv_weights: torch.Tensor, |
| 1479 | *, |
| 1480 | conv_state_len: int, |
| 1481 | dilation: int, |
| 1482 | ) -> None: |
| 1483 | """PyTorch reference for fused short-convolution dispatch.""" |
| 1484 | num_prefills = metadata.num_prefills |
| 1485 | num_decodes = metadata.num_decodes |
| 1486 | num_decode_tokens = metadata.num_decode_tokens |
| 1487 | num_prefill_tokens = metadata.num_prefill_tokens |
| 1488 | has_prefill = num_prefills > 0 |
| 1489 | has_decode = num_decodes > 0 |
| 1490 | has_spec = metadata.spec_sequence_masks is not None |
| 1491 | x = inputs[: metadata.num_actual_tokens] |
| 1492 | residual = residual[: metadata.num_actual_tokens] |
| 1493 | |
| 1494 | if has_spec: |
| 1495 | if has_prefill or has_decode: |
| 1496 | assert metadata.spec_token_indx is not None |
| 1497 | assert metadata.non_spec_token_indx is not None |
| 1498 | spec_token_indices = metadata.spec_token_indx.long() |
| 1499 | non_spec_token_indices = metadata.non_spec_token_indx.long() |
| 1500 | x_spec = x.index_select(0, spec_token_indices) |
| 1501 | x_non_spec = x.index_select(0, non_spec_token_indices) |
| 1502 | residual_spec = residual.index_select(0, spec_token_indices) |
| 1503 | residual_non_spec = residual.index_select(0, non_spec_token_indices) |
| 1504 | else: |
| 1505 | spec_token_indices = None |
| 1506 | non_spec_token_indices = None |
| 1507 | x_spec = x |
| 1508 | x_non_spec = None |
| 1509 | residual_spec = residual |
| 1510 | residual_non_spec = None |
| 1511 | else: |
| 1512 | spec_token_indices = None |
| 1513 | non_spec_token_indices = None |
| 1514 | x_spec = None |
| 1515 | x_non_spec = x |
| 1516 | residual_spec = None |
| 1517 | residual_non_spec = residual |
| 1518 | |
| 1519 | if has_spec: |
| 1520 | assert metadata.spec_state_indices_tensor is not None |
| 1521 | assert metadata.spec_query_start_loc is not None |
| 1522 | assert metadata.num_accepted_tokens is not None |
| 1523 | assert x_spec is not None |
| 1524 | assert residual_spec is not None |
| 1525 | spec_state_indices = metadata.spec_state_indices_tensor[ |
| 1526 | : metadata.num_spec_decodes |
| 1527 | ] |
| 1528 | conv_output = _short_conv_dilated_spec_pytorch( |
| 1529 | x_spec=x_spec, |
| 1530 | conv_state=conv_state, |
| 1531 | conv_weights=conv_weights, |
| 1532 | spec_state_indices_tensor=spec_state_indices, |
| 1533 | spec_query_start_loc=metadata.spec_query_start_loc, |
| 1534 | num_accepted_tokens=metadata.num_accepted_tokens, |
| 1535 | spec_query_len=metadata.spec_query_len, |
| 1536 | conv_state_len=conv_state_len, |
| 1537 | dilation=dilation, |
| 1538 | ) |
| 1539 | residual_spec.add_(conv_output) |
| 1540 | |
| 1541 | state_indices = metadata.state_indices_tensor |
| 1542 | if x_non_spec is not None: |
| 1543 | assert state_indices is not None |
| 1544 | assert residual_non_spec is not None |
| 1545 | if has_prefill: |
| 1546 | state_indices_d, state_indices_p = torch.split( |
| 1547 | state_indices, [num_decodes, num_prefills], dim=0 |
| 1548 | ) |
| 1549 | x_d, x_p = torch.split( |
| 1550 | x_non_spec, [num_decode_tokens, num_prefill_tokens], dim=0 |
| 1551 | ) |
| 1552 | residual_d, residual_p = torch.split( |
| 1553 | residual_non_spec, |
| 1554 | [num_decode_tokens, num_prefill_tokens], |
| 1555 | dim=0, |
| 1556 | ) |
| 1557 | if has_decode: |
| 1558 | conv_output = _short_conv_dilated_decode_pytorch( |
| 1559 | x_d=x_d, |
| 1560 | conv_state=conv_state, |
| 1561 | conv_weights=conv_weights, |
| 1562 | state_indices_tensor_d=state_indices_d, |
| 1563 | has_initial_states_d=metadata.has_initial_states_d, |
| 1564 | conv_state_len=conv_state_len, |
| 1565 | dilation=dilation, |
| 1566 | ) |
| 1567 | residual_d.add_(conv_output) |
| 1568 | conv_output = _short_conv_dilated_prefill_pytorch( |
| 1569 | x_p=x_p, |
| 1570 | metadata=metadata, |
| 1571 | conv_state=conv_state, |
| 1572 | conv_weights=conv_weights, |
| 1573 | state_indices_tensor_p=state_indices_p, |
| 1574 | num_prefills=num_prefills, |
| 1575 | num_decode_tokens=num_decode_tokens, |
| 1576 | num_prefill_tokens=num_prefill_tokens, |
| 1577 | conv_state_len=conv_state_len, |
| 1578 | dilation=dilation, |
| 1579 | ) |
| 1580 | residual_p.add_(conv_output) |
| 1581 | else: |
| 1582 | conv_output = _short_conv_dilated_decode_pytorch( |
| 1583 | x_d=x_non_spec, |
| 1584 | conv_state=conv_state, |
| 1585 | conv_weights=conv_weights, |
| 1586 | state_indices_tensor_d=state_indices[: x_non_spec.size(0)], |
| 1587 | has_initial_states_d=metadata.has_initial_states_d, |
| 1588 | conv_state_len=conv_state_len, |
| 1589 | dilation=dilation, |
| 1590 | ) |
| 1591 | residual_non_spec.add_(conv_output) |
| 1592 | |
| 1593 | if has_spec and residual_non_spec is not None: |
| 1594 | assert spec_token_indices is not None |
| 1595 | assert non_spec_token_indices is not None |
| 1596 | assert residual_spec is not None |
| 1597 | residual.index_copy_(0, spec_token_indices, residual_spec) |
| 1598 | residual.index_copy_(0, non_spec_token_indices, residual_non_spec) |
| 1599 | |
| 1600 | |
| 1601 | _KERNEL_STATE_BLOCKS = 64 |
| 1602 | |
| 1603 | |
| 1604 | def _make_conv_case( |
| 1605 | device: torch.device, |
| 1606 | seed: int, |
| 1607 | channels: int, |
| 1608 | kernel_size: int, |
| 1609 | dilation: int, |
| 1610 | state_layout: str, |
| 1611 | spec_query_len: int = 1, |
| 1612 | ) -> tuple[torch.Generator, torch.Tensor, torch.Tensor, torch.Tensor]: |
| 1613 | rng = torch.Generator(device=device).manual_seed(seed) |
| 1614 | state_len = (kernel_size - 1) * dilation |
| 1615 | state_width = state_len + spec_query_len - 1 |
| 1616 | state_shape = ( |
| 1617 | (_KERNEL_STATE_BLOCKS, state_width, channels) |
| 1618 | if state_layout == "SD" |
| 1619 | else (_KERNEL_STATE_BLOCKS, channels, state_width) |
| 1620 | ) |
| 1621 | state_kernel = torch.randn( |
| 1622 | state_shape, device=device, dtype=torch.bfloat16, generator=rng |
| 1623 | ) |
| 1624 | conv_state = ( |
| 1625 | state_kernel.transpose(-1, -2) if state_layout == "SD" else state_kernel |
| 1626 | ) |
| 1627 | state_reference = conv_state.clone() |
| 1628 | weights = torch.randn( |
| 1629 | channels, kernel_size, device=device, dtype=torch.bfloat16, generator=rng |
| 1630 | ) |
| 1631 | return rng, state_reference, conv_state, weights |
| 1632 | |
| 1633 | |
| 1634 | @dataclass(frozen=True) |
| 1635 | class _ConvBatchCase: |
| 1636 | spec_query_lens: tuple[int, ...] = () |
| 1637 | num_accepted: tuple[int, ...] = () |
| 1638 | num_decodes: int = 0 |
| 1639 | prefill_query_lens: tuple[int, ...] = () |
| 1640 | channels: int = 2048 |
| 1641 | kernel_size: int = 4 |
| 1642 | dilation: int = 3 |
| 1643 | spec_query_len: int = 1 |
| 1644 | graph_padding: int = 0 |
| 1645 | state_index_stride: int = 1 |
| 1646 | include_null_state: bool = True |
| 1647 | |
| 1648 | |
| 1649 | def _make_conv_metadata( |
| 1650 | case: _ConvBatchCase, |
| 1651 | device: torch.device, |
| 1652 | ) -> tuple[SimpleNamespace, int]: |
| 1653 | request_groups: dict[str, list[int]] = { |
| 1654 | "spec": [], |
| 1655 | "decode": [], |
| 1656 | "prefill": [], |
| 1657 | } |
| 1658 | token_offset = 0 |
| 1659 | max_num_reqs = max( |
| 1660 | len(case.spec_query_lens), |
| 1661 | case.num_decodes, |
| 1662 | len(case.prefill_query_lens), |
| 1663 | ) |
| 1664 | request_kinds = [] |
| 1665 | for req_idx in range(max_num_reqs): |
| 1666 | if req_idx < len(case.spec_query_lens): |
| 1667 | request_kinds.append(("spec", case.spec_query_lens[req_idx])) |
| 1668 | if req_idx < case.num_decodes: |
| 1669 | request_kinds.append(("decode", 1)) |
| 1670 | if req_idx < len(case.prefill_query_lens): |
| 1671 | request_kinds.append(("prefill", case.prefill_query_lens[req_idx])) |
| 1672 | for kind, query_len in request_kinds: |
| 1673 | request_groups[kind].extend(range(token_offset, token_offset + query_len)) |
| 1674 | token_offset += query_len |
| 1675 | |
| 1676 | num_spec_reqs = len(case.spec_query_lens) |
| 1677 | num_prefills = len(case.prefill_query_lens) |
| 1678 | has_spec = num_spec_reqs > 0 |
| 1679 | has_non_spec = case.num_decodes > 0 or num_prefills > 0 |
| 1680 | mixed_batch = has_spec and has_non_spec |
| 1681 | |
| 1682 | spec_state_indices = torch.arange( |
| 1683 | 1, num_spec_reqs + 1, dtype=torch.int32, device=device |
| 1684 | ) |
| 1685 | if num_spec_reqs >= 3: |
| 1686 | spec_state_indices[1] = NULL_BLOCK_ID |
| 1687 | non_spec_state_indices = torch.arange( |
| 1688 | num_spec_reqs + 1, |
| 1689 | num_spec_reqs + case.num_decodes + num_prefills + 1, |
| 1690 | dtype=torch.int32, |
| 1691 | device=device, |
| 1692 | ) |
| 1693 | if case.state_index_stride > 1: |
| 1694 | strided_indices = torch.full( |
| 1695 | (non_spec_state_indices.numel(), case.state_index_stride), |
| 1696 | NULL_BLOCK_ID, |
| 1697 | dtype=torch.int32, |
| 1698 | device=device, |
| 1699 | ) |
| 1700 | strided_indices[:, 0] = non_spec_state_indices |
| 1701 | non_spec_state_indices = strided_indices[:, 0] |
| 1702 | if case.num_decodes > 1: |
| 1703 | non_spec_state_indices[case.num_decodes - 1] = NULL_BLOCK_ID |
| 1704 | if num_prefills > 1 and case.include_null_state: |
| 1705 | non_spec_state_indices[case.num_decodes + 1] = NULL_BLOCK_ID |
| 1706 | |
| 1707 | spec_query_start_loc = torch.tensor( |
| 1708 | [0, *accumulate(case.spec_query_lens)], dtype=torch.int32, device=device |
| 1709 | ) |
| 1710 | non_spec_query_lens = (1,) * case.num_decodes + case.prefill_query_lens |
| 1711 | non_spec_query_start_loc = torch.tensor( |
| 1712 | [0, *accumulate(non_spec_query_lens)], dtype=torch.int32, device=device |
| 1713 | ) |
| 1714 | spec_token_indices = torch.tensor( |
| 1715 | request_groups["spec"], dtype=torch.int64, device=device |
| 1716 | ) |
| 1717 | non_spec_token_indices = torch.tensor( |
| 1718 | request_groups["decode"] + request_groups["prefill"], |
| 1719 | dtype=torch.int64, |
| 1720 | device=device, |
| 1721 | ) |
| 1722 | |
| 1723 | metadata = SimpleNamespace( |
| 1724 | num_prefills=num_prefills, |
| 1725 | num_decodes=case.num_decodes, |
| 1726 | num_decode_tokens=case.num_decodes, |
| 1727 | num_prefill_tokens=sum(case.prefill_query_lens), |
| 1728 | spec_sequence_masks=( |
| 1729 | torch.tensor([kind == "spec" for kind, _ in request_kinds], device=device) |
| 1730 | if has_spec |
| 1731 | else None |
| 1732 | ), |
| 1733 | spec_token_indx=spec_token_indices if mixed_batch else None, |
| 1734 | non_spec_token_indx=non_spec_token_indices if mixed_batch else None, |
| 1735 | num_actual_tokens=token_offset + case.graph_padding, |
| 1736 | spec_state_indices_tensor=spec_state_indices if has_spec else None, |
| 1737 | spec_query_start_loc=spec_query_start_loc if has_spec else None, |
| 1738 | num_accepted_tokens=( |
| 1739 | torch.tensor(case.num_accepted, dtype=torch.int32, device=device) |
| 1740 | if has_spec |
| 1741 | else None |
| 1742 | ), |
| 1743 | spec_query_len=case.spec_query_len, |
| 1744 | state_indices_tensor=non_spec_state_indices if has_non_spec else None, |
| 1745 | has_initial_states_d=( |
| 1746 | torch.arange(case.num_decodes, device=device) % 2 == 0 |
| 1747 | if case.num_decodes |
| 1748 | else None |
| 1749 | ), |
| 1750 | non_spec_query_start_loc=(non_spec_query_start_loc if has_non_spec else None), |
| 1751 | query_start_loc_p=( |
| 1752 | non_spec_query_start_loc[case.num_decodes :] - case.num_decodes |
| 1753 | if num_prefills |
| 1754 | else None |
| 1755 | ), |
| 1756 | has_initial_states_p=( |
| 1757 | torch.arange(num_prefills, device=device) % 2 == 0 if num_prefills else None |
| 1758 | ), |
| 1759 | max_prefill_query_len=max(case.prefill_query_lens, default=0), |
| 1760 | num_spec_decodes=num_spec_reqs, |
| 1761 | ) |
| 1762 | return metadata, token_offset |
| 1763 | |
| 1764 | |
| 1765 | @pytest.mark.skipif(not torch.cuda.is_available(), reason="fused conv needs CUDA") |
| 1766 | @pytest.mark.parametrize("state_layout", ["SD", "DS"]) |
| 1767 | @pytest.mark.parametrize( |
| 1768 | "case", |
| 1769 | [ |
| 1770 | pytest.param( |
| 1771 | _ConvBatchCase(num_decodes=1, channels=512), |
| 1772 | id="decode-single", |
| 1773 | ), |
| 1774 | pytest.param( |
| 1775 | _ConvBatchCase( |
| 1776 | num_decodes=33, |
| 1777 | channels=2048, |
| 1778 | kernel_size=5, |
| 1779 | dilation=1, |
| 1780 | ), |
| 1781 | id="decode-padded", |
| 1782 | ), |
| 1783 | pytest.param( |
| 1784 | _ConvBatchCase(prefill_query_lens=(37, 0, 5, 128)), |
| 1785 | id="prefill", |
| 1786 | ), |
| 1787 | pytest.param( |
| 1788 | _ConvBatchCase( |
| 1789 | prefill_query_lens=(16, 676), |
| 1790 | channels=10240, |
| 1791 | state_index_stride=4, |
| 1792 | include_null_state=False, |
| 1793 | ), |
| 1794 | id="prefill-strided-state-indices", |
| 1795 | ), |
| 1796 | pytest.param( |
| 1797 | _ConvBatchCase( |
| 1798 | spec_query_lens=(1, 4, 4, 0), |
| 1799 | num_accepted=(1, 2, 4, 1), |
| 1800 | spec_query_len=4, |
| 1801 | graph_padding=4, |
| 1802 | ), |
| 1803 | id="spec-graph-padded", |
| 1804 | ), |
| 1805 | pytest.param( |
| 1806 | _ConvBatchCase( |
| 1807 | spec_query_lens=(3, 4), |
| 1808 | num_accepted=(2, 4), |
| 1809 | num_decodes=1, |
| 1810 | prefill_query_lens=(5,), |
| 1811 | spec_query_len=4, |
| 1812 | ), |
| 1813 | id="mixed", |
| 1814 | ), |
| 1815 | pytest.param( |
| 1816 | _ConvBatchCase( |
| 1817 | spec_query_lens=(1, 2), |
| 1818 | num_accepted=(0, 1), |
| 1819 | num_decodes=2, |
| 1820 | prefill_query_lens=(0, 8, 3), |
| 1821 | channels=640, |
| 1822 | kernel_size=3, |
| 1823 | dilation=2, |
| 1824 | spec_query_len=3, |
| 1825 | ), |
| 1826 | id="mixed-varied", |
| 1827 | ), |
| 1828 | ], |
| 1829 | ) |
| 1830 | def test_fused_conv_correctness( |
| 1831 | case: _ConvBatchCase, |
| 1832 | state_layout: str, |
| 1833 | ) -> None: |
| 1834 | device = torch.device("cuda") |
| 1835 | metadata, num_real_tokens = _make_conv_metadata(case, device) |
| 1836 | if case.state_index_stride > 1: |
| 1837 | assert metadata.state_indices_tensor.stride(0) == case.state_index_stride |
| 1838 | module = Qwen4ExpPLELayer.__new__(Qwen4ExpPLELayer) |
| 1839 | nn.Module.__init__(module) |
| 1840 | module.conv_state_len = (case.kernel_size - 1) * case.dilation |
| 1841 | module.short_conv_dilation = case.dilation |
| 1842 | |
| 1843 | rng, state_reference, conv_state, weights = _make_conv_case( |
| 1844 | device, |
| 1845 | seed=num_real_tokens + case.channels, |
| 1846 | channels=case.channels, |
| 1847 | kernel_size=case.kernel_size, |
| 1848 | dilation=case.dilation, |
| 1849 | state_layout=state_layout, |
| 1850 | spec_query_len=case.spec_query_len, |
| 1851 | ) |
| 1852 | inputs = torch.randn( |
| 1853 | metadata.num_actual_tokens, |
| 1854 | case.channels, |
| 1855 | device=device, |
| 1856 | dtype=torch.bfloat16, |
| 1857 | generator=rng, |
| 1858 | ) |
| 1859 | residual = torch.randn( |
| 1860 | inputs.shape, |
| 1861 | device=device, |
| 1862 | dtype=torch.bfloat16, |
| 1863 | generator=rng, |
| 1864 | ) |
| 1865 | outer_residual = torch.randn( |
| 1866 | inputs.shape, |
| 1867 | device=device, |
| 1868 | dtype=torch.bfloat16, |
| 1869 | generator=rng, |
| 1870 | ) |
| 1871 | null_state = conv_state[NULL_BLOCK_ID].clone() |
| 1872 | residual_kernel = residual.clone() |
| 1873 | residual_reference = residual.clone() |
| 1874 | |
| 1875 | module._short_conv_dilated_dispatch( |
| 1876 | inputs=inputs, |
| 1877 | residual=residual_kernel, |
| 1878 | outer_residual=outer_residual, |
| 1879 | metadata=metadata, |
| 1880 | conv_state=conv_state, |
| 1881 | conv_weights=weights, |
| 1882 | ) |
| 1883 | _short_conv_dilated_dispatch_pytorch( |
| 1884 | inputs=inputs, |
| 1885 | residual=residual_reference, |
| 1886 | metadata=metadata, |
| 1887 | conv_state=state_reference, |
| 1888 | conv_weights=weights, |
| 1889 | conv_state_len=module.conv_state_len, |
| 1890 | dilation=module.short_conv_dilation, |
| 1891 | ) |
| 1892 | residual_reference = outer_residual + residual_reference |
| 1893 | |
| 1894 | assert torch.equal( |
| 1895 | residual_kernel[:num_real_tokens], residual_reference[:num_real_tokens] |
| 1896 | ) |
| 1897 | assert torch.equal(conv_state, state_reference) |
| 1898 | assert torch.equal(conv_state[NULL_BLOCK_ID], null_state) |
| 1899 | if case.graph_padding: |
| 1900 | assert torch.equal( |
| 1901 | residual_kernel[num_real_tokens:], |
| 1902 | (outer_residual + residual)[num_real_tokens:], |
| 1903 | ) |
| 1904 | |
| 1905 | |
| 1906 | @pytest.mark.skipif(not torch.cuda.is_available(), reason="fused gate needs CUDA") |
| 1907 | @pytest.mark.parametrize(("num_tokens", "strided_kv"), [(1, False), (64, True)]) |
| 1908 | def test_fused_gate_correctness(num_tokens: int, strided_kv: bool) -> None: |
| 1909 | import math |
| 1910 | |
| 1911 | from vllm.models.qwen4_exp.nvidia.ops.ple import ple_gate |
| 1912 | |
| 1913 | device = torch.device("cuda") |
| 1914 | hc, h = 4, 2560 |
| 1915 | generator = torch.Generator(device=device).manual_seed(num_tokens) |
| 1916 | kv = torch.randn( |
| 1917 | num_tokens, |
| 1918 | hc * h + h, |
| 1919 | device=device, |
| 1920 | dtype=torch.bfloat16, |
| 1921 | generator=generator, |
| 1922 | ) |
| 1923 | key = kv[:, : hc * h] |
| 1924 | value = kv[:, hc * h :] |
| 1925 | if not strided_kv: |
| 1926 | key = key.contiguous() |
| 1927 | value = value.contiguous() |
| 1928 | hidden = torch.randn( |
| 1929 | num_tokens, |
| 1930 | hc * h, |
| 1931 | device=device, |
| 1932 | dtype=torch.bfloat16, |
| 1933 | generator=generator, |
| 1934 | ) |
| 1935 | |
| 1936 | def make_norm_weight() -> torch.Tensor: |
| 1937 | weight = torch.empty(hc * h, device=device, dtype=torch.bfloat16) |
| 1938 | return weight.normal_(mean=-0.1, std=0.1, generator=generator) |
| 1939 | |
| 1940 | norm_key = make_norm_weight() |
| 1941 | norm_query = make_norm_weight() |
| 1942 | norm_conv = make_norm_weight() |
| 1943 | |
| 1944 | def grouped_norm(inputs: torch.Tensor, weight: torch.Tensor) -> torch.Tensor: |
| 1945 | grouped = inputs.float().unflatten(-1, (hc, h)) |
| 1946 | var = grouped.square().mean(dim=-1, keepdim=True) |
| 1947 | normalized = grouped * torch.rsqrt(var + 1e-6) |
| 1948 | return (normalized.flatten(-2) * (1.0 + weight.float())).to(inputs.dtype) |
| 1949 | |
| 1950 | gated, normed = ple_gate( |
| 1951 | key, |
| 1952 | value, |
| 1953 | hidden, |
| 1954 | norm_key, |
| 1955 | norm_query, |
| 1956 | norm_conv, |
| 1957 | 1e-6, |
| 1958 | ) |
| 1959 | key_normalized = grouped_norm(key, norm_key).reshape(num_tokens, hc, h) |
| 1960 | query_normalized = grouped_norm(hidden, norm_query).reshape(num_tokens, hc, h) |
| 1961 | dot = (key_normalized * query_normalized).sum(dim=-1, keepdim=True) |
| 1962 | dot = (dot / math.sqrt(h)).to(torch.bfloat16) |
| 1963 | gate = torch.sigmoid(dot.sign() * dot.abs().clamp_min(1e-6).sqrt()).to( |
| 1964 | torch.bfloat16 |
| 1965 | ) |
| 1966 | expected_gated = (gate * value.unsqueeze(-2)).flatten(-2) |
| 1967 | expected_normed = grouped_norm(expected_gated, norm_conv) |
| 1968 | assert torch.equal(gated, expected_gated) |
| 1969 | torch.testing.assert_close(normed, expected_normed, atol=1e-2, rtol=1e-2) |
| 1970 | |
| 1971 | |
| 1972 | def _build_amd_ngram_embedding( |
| 1973 | monkeypatch: pytest.MonkeyPatch, |
| 1974 | *, |
| 1975 | device: str, |
| 1976 | cpu_offload: bool, |
| 1977 | fp8_checkpoint: bool, |
| 1978 | ) -> tuple[amd_ple_layer.Qwen4ExpNGramEmbedding, torch.Tensor]: |
| 1979 | """Construct and load a small AMD embedding through its checkpoint loader.""" |
| 1980 | _mock_etp_group(monkeypatch) |
| 1981 | monkeypatch.setattr(embedding_module, "get_tensor_model_parallel_rank", lambda: 0) |
| 1982 | monkeypatch.setattr( |
| 1983 | embedding_module, "get_tensor_model_parallel_world_size", lambda: 1 |
| 1984 | ) |
| 1985 | monkeypatch.setattr(parameter_module, "get_tensor_model_parallel_rank", lambda: 0) |
| 1986 | monkeypatch.setattr( |
| 1987 | parameter_module, "get_tensor_model_parallel_world_size", lambda: 1 |
| 1988 | ) |
| 1989 | |
| 1990 | monkeypatch.setattr( |
| 1991 | amd_ple_layer, |
| 1992 | "get_current_vllm_config", |
| 1993 | lambda: SimpleNamespace(engram_config=SimpleNamespace(cpu_offload=cpu_offload)), |
| 1994 | ) |
| 1995 | config = Qwen4ExpTextConfig( |
| 1996 | ngram_size=3, |
| 1997 | heads_per_ngram=1, |
| 1998 | ngram_vocab_size_base=5, |
| 1999 | make_ngram_vocab_size_divisible_by=8, |
| 2000 | split_ngram_parts=2, |
| 2001 | ple_embed_dim=6, |
| 2002 | eos_token_id=0, |
| 2003 | vocab_size=64, |
| 2004 | ) |
| 2005 | with torch.device(device): |
| 2006 | module = amd_ple_layer.Qwen4ExpNGramEmbedding( |
| 2007 | config, |
| 2008 | embedding_dim=config.ple_embed_dim, |
| 2009 | ple_dense_layer_id=0, |
| 2010 | max_total_tokens=4, |
| 2011 | max_num_reqs=2, |
| 2012 | prefix="test.ple_embedding", |
| 2013 | layer_name="test.ple", |
| 2014 | quant_config=( |
| 2015 | Fp8Config(is_checkpoint_fp8_serialized=True) if fp8_checkpoint else None |
| 2016 | ), |
| 2017 | params_dtype=torch.bfloat16, |
| 2018 | ) |
| 2019 | # Only the outer owner is a stub; the embedding constructor and loader run. |
| 2020 | layer = Qwen4ExpPLELayerAMD.__new__(Qwen4ExpPLELayerAMD) |
| 2021 | nn.Module.__init__(layer) |
| 2022 | layer.ple_embedding = module |
| 2023 | |
| 2024 | monkeypatch.setattr( |
| 2025 | amd_ple_layer, |
| 2026 | "get_forward_context", |
| 2027 | lambda: SimpleNamespace(no_compile_layers={module.layer_name: layer}), |
| 2028 | ) |
| 2029 | |
| 2030 | embedding = module.ngram_embedding |
| 2031 | loaded_weight = ( |
| 2032 | torch.arange(embedding.org_vocab_size * embedding.embedding_dim) |
| 2033 | .reshape(embedding.org_vocab_size, embedding.embedding_dim) |
| 2034 | .to(embedding.weight.dtype) |
| 2035 | ) |
| 2036 | shard_size = (embedding.org_vocab_size + 1) // 2 |
| 2037 | weights = [ |
| 2038 | (f"ngram_embedding.shard_{i}.weight", shard) |
| 2039 | for i, shard in enumerate(loaded_weight.split(shard_size)) |
| 2040 | ] |
| 2041 | expected_loaded = {"ngram_embedding.weight"} |
| 2042 | if fp8_checkpoint: |
| 2043 | weights.append(("ngram_embedding.weight_scale", torch.tensor([0.25]))) |
| 2044 | expected_loaded.add("ngram_embedding.weight_scale") |
| 2045 | assert module.load_weights(weights) == expected_loaded |
| 2046 | embedding.quant_method.process_weights_after_loading(embedding) |
| 2047 | return module, loaded_weight |
| 2048 | |
| 2049 | |
| 2050 | @pytest.mark.parametrize( |
| 2051 | "device,cpu_offload", |
| 2052 | [ |
| 2053 | pytest.param("cpu", False, id="cpu-resident"), |
| 2054 | pytest.param( |
| 2055 | "cuda:0", |
| 2056 | False, |
| 2057 | id="gpu-resident", |
| 2058 | marks=pytest.mark.skipif( |
| 2059 | not torch.cuda.is_available(), reason="requires CUDA or ROCm" |
| 2060 | ), |
| 2061 | ), |
| 2062 | pytest.param( |
| 2063 | "cuda:0", |
| 2064 | True, |
| 2065 | id="gpu-offloaded", |
| 2066 | marks=pytest.mark.skipif( |
| 2067 | not torch.cuda.is_available(), reason="requires CUDA or ROCm" |
| 2068 | ), |
| 2069 | ), |
| 2070 | ], |
| 2071 | ) |
| 2072 | def test_amd_fp8_embedding_loads_checkpoint_shards_and_global_scale( |
| 2073 | monkeypatch: pytest.MonkeyPatch, |
| 2074 | device: str, |
| 2075 | cpu_offload: bool, |
| 2076 | ) -> None: |
| 2077 | """AMD checkpoint loading must accept and apply the global FP8 scale.""" |
| 2078 | module, loaded_weight = _build_amd_ngram_embedding( |
| 2079 | monkeypatch, device=device, cpu_offload=cpu_offload, fp8_checkpoint=True |
| 2080 | ) |
| 2081 | embedding = module.ngram_embedding |
| 2082 | assert embedding.weight.dtype == torch.float8_e4m3fn |
| 2083 | assert embedding.weight.device == torch.device("cpu" if cpu_offload else device) |
| 2084 | if cpu_offload: |
| 2085 | assert embedding.weight.is_pinned() |
| 2086 | assert embedding.weight_scale.device == torch.device(device) |
| 2087 | torch.testing.assert_close(embedding.weight.float().cpu(), loaded_weight.float()) |
| 2088 | dequantized = embedding.dequantize(embedding.weight.to(device), torch.bfloat16) |
| 2089 | torch.testing.assert_close( |
| 2090 | dequantized.cpu(), loaded_weight.to(torch.bfloat16) * 0.25, rtol=0, atol=0 |
| 2091 | ) |
| 2092 | |
| 2093 | |
| 2094 | @pytest.mark.skipif(not torch.cuda.is_available(), reason="requires CUDA or ROCm") |
| 2095 | def test_amd_pinned_embedding_prefetch_machinery_is_lazy( |
| 2096 | monkeypatch: pytest.MonkeyPatch, |
| 2097 | ) -> None: |
| 2098 | """The pinned embedding must not allocate the prefetch stream/buffer in init. |
| 2099 | |
| 2100 | Paths that never prefetch (AMD uses the synchronous pinned-lookup op) |
| 2101 | keep the stream/buffer unallocated for the process lifetime. NVIDIA's |
| 2102 | first ``start_prefetch`` — the eager profile run, always before cudagraph |
| 2103 | capture — allocates them lazily. |
| 2104 | """ |
| 2105 | module, _ = _build_amd_ngram_embedding( |
| 2106 | monkeypatch, device="cuda:0", cpu_offload=True, fp8_checkpoint=False |
| 2107 | ) |
| 2108 | embedding = module.ngram_embedding |
| 2109 | assert embedding._prefetch_buffer is None |
| 2110 | assert embedding._prefetch_stream is None |
| 2111 | |
| 2112 | |
| 2113 | @pytest.mark.skipif(not torch.cuda.is_available(), reason="requires CUDA or ROCm") |
| 2114 | @pytest.mark.parametrize("fp8_checkpoint", [False, True], ids=["bf16", "fp8"]) |
| 2115 | def test_amd_pinned_embedding_output_written_under_compile( |
| 2116 | monkeypatch: pytest.MonkeyPatch, |
| 2117 | fp8_checkpoint: bool, |
| 2118 | ) -> None: |
| 2119 | """The AMD pinned lookup op must survive aot_autograd/Inductor. |
| 2120 | |
| 2121 | The op writes via the declared-mutated ``output``. Without |
| 2122 | ``mutates_args=["output"]`` the call is dead-code-eliminated, the UVA lookup |
| 2123 | never runs, and the returned ``output`` stays untouched. |
| 2124 | """ |
| 2125 | module, loaded_weight = _build_amd_ngram_embedding( |
| 2126 | monkeypatch, device="cuda:0", cpu_offload=True, fp8_checkpoint=fp8_checkpoint |
| 2127 | ) |
| 2128 | |
| 2129 | ngram_ids = torch.tensor([[3, 0], [1, 2]], device="cuda:0") |
| 2130 | output = torch.zeros( |
| 2131 | ngram_ids.shape[0], |
| 2132 | module.embedding_dim, |
| 2133 | dtype=module.ngram_embedding.weight.dtype, |
| 2134 | device="cuda:0", |
| 2135 | ) |
| 2136 | |
| 2137 | def lookup(ids: torch.Tensor, out: torch.Tensor) -> torch.Tensor: |
| 2138 | torch.ops.vllm.qwen4_exp_amd_ple_ngram_embedding_pinned( |
| 2139 | ids, out, module.layer_name |
| 2140 | ) |
| 2141 | return out |
| 2142 | |
| 2143 | torch.compile(lookup, fullgraph=True)(ngram_ids, output) |
| 2144 | torch.cuda.current_stream().synchronize() |
| 2145 | |
| 2146 | expected = loaded_weight[ngram_ids.cpu()].to(device="cuda:0").flatten(-2) |
| 2147 | torch.testing.assert_close(output.float(), expected.float(), rtol=0, atol=0) |
| 2148 | |
| 2149 | |
| 2150 | @pytest.mark.skipif(not torch.cuda.is_available(), reason="requires CUDA or ROCm") |
| 2151 | @pytest.mark.parametrize("fp8_checkpoint", [False, True], ids=["bf16", "fp8"]) |
| 2152 | def test_amd_pinned_embedding_output_written_under_cudagraph_capture( |
| 2153 | monkeypatch: pytest.MonkeyPatch, |
| 2154 | fp8_checkpoint: bool, |
| 2155 | ) -> None: |
| 2156 | """The AMD pinned lookup op must be capture-safe. |
| 2157 | |
| 2158 | It launches a single kernel on the current stream, so a cudagraph capture |
| 2159 | reproduces the lookup into ``output`` without cross-stream work. |
| 2160 | """ |
| 2161 | module, loaded_weight = _build_amd_ngram_embedding( |
| 2162 | monkeypatch, device="cuda:0", cpu_offload=True, fp8_checkpoint=fp8_checkpoint |
| 2163 | ) |
| 2164 | |
| 2165 | ngram_ids = torch.tensor([[3, 0], [1, 2]], device="cuda:0") |
| 2166 | output = torch.zeros( |
| 2167 | ngram_ids.shape[0], |
| 2168 | module.embedding_dim, |
| 2169 | dtype=module.ngram_embedding.weight.dtype, |
| 2170 | device="cuda:0", |
| 2171 | ) |
| 2172 | |
| 2173 | def lookup(ids: torch.Tensor, out: torch.Tensor) -> torch.Tensor: |
| 2174 | torch.ops.vllm.qwen4_exp_amd_ple_ngram_embedding_pinned( |
| 2175 | ids, out, module.layer_name |
| 2176 | ) |
| 2177 | return out |
| 2178 | |
| 2179 | graph = torch.cuda.CUDAGraph() |
| 2180 | warmup_stream = torch.cuda.Stream(device="cuda:0") |
| 2181 | warmup_stream.wait_stream(torch.cuda.current_stream()) |
| 2182 | with torch.cuda.stream(warmup_stream): |
| 2183 | for _ in range(3): |
| 2184 | lookup(ngram_ids, output) |
| 2185 | torch.cuda.current_stream().wait_stream(warmup_stream) |
| 2186 | |
| 2187 | with torch.cuda.graph(graph): |
| 2188 | lookup(ngram_ids, output) |
| 2189 | for _ in range(2): |
| 2190 | ngram_ids.copy_(ngram_ids.flip(-1) + 3) |
| 2191 | output.fill_(float("nan")) |
| 2192 | graph.replay() |
| 2193 | torch.cuda.current_stream().synchronize() |
| 2194 | |
| 2195 | expected = loaded_weight[ngram_ids.cpu()].to(device="cuda:0").flatten(-2) |
| 2196 | torch.testing.assert_close(output.float(), expected.float(), rtol=0, atol=0) |