eagle.py 63.3 KB
Newer Older
1
# SPDX-License-Identifier: Apache-2.0
2
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3
4
import ast
from dataclasses import replace
5
from importlib.util import find_spec
6
from typing import cast
7

8
import numpy as np
9
10
11
import torch
import torch.nn as nn

12
13
14
15
16
from vllm.config import (
    CUDAGraphMode,
    VllmConfig,
    get_layers_from_vllm_config,
)
17
from vllm.distributed.parallel_state import get_pp_group
18
from vllm.forward_context import set_forward_context
19
from vllm.logger import init_logger
20
from vllm.model_executor.layers.attention_layer_base import AttentionLayerBase
21
from vllm.model_executor.model_loader import get_model
22
from vllm.model_executor.models import supports_multimodal
23
from vllm.model_executor.models.deepseek_v2 import DeepseekV32IndexerCache
24
from vllm.model_executor.models.interfaces import SupportsMultiModal
25
from vllm.model_executor.models.llama_eagle3 import Eagle3LlamaForCausalLM
26
from vllm.multimodal import MULTIMODAL_REGISTRY
27
from vllm.platforms import current_platform
28
from vllm.triton_utils import triton
29
from vllm.utils.platform_utils import is_pin_memory_available
30
31
32
33
from vllm.v1.attention.backend import (
    AttentionMetadataBuilder,
    CommonAttentionMetadata,
)
34
from vllm.v1.attention.backends.registry import AttentionBackendEnum
35
36
37
38
from vllm.v1.attention.backends.tree_attn import (
    TreeAttentionMetadata,
    TreeAttentionMetadataBuilder,
)
39
from vllm.v1.attention.backends.triton_attn import TritonAttentionMetadata
40
from vllm.v1.cudagraph_dispatcher import CudagraphDispatcher
41
from vllm.v1.kv_cache_interface import KVCacheConfig
42
from vllm.v1.sample.metadata import SamplingMetadata
43
from vllm.v1.sample.sampler import _SAMPLING_EPS
44
from vllm.v1.spec_decode.metadata import SpecDecodeMetadata
45
46
47
48
from vllm.v1.spec_decode.utils import (
    eagle_prepare_inputs_padded_kernel,
    eagle_prepare_next_token_padded_kernel,
)
49
from vllm.v1.utils import CpuGpuBuffer
Rémi Delacourt's avatar
Rémi Delacourt committed
50
from vllm.v1.worker.dp_utils import coordinate_batch_across_dp
51
from vllm.v1.worker.gpu_input_batch import CachedRequestState, InputBatch
52

53
54
logger = init_logger(__name__)

55
56
PADDING_SLOT_ID = -1

57

58
class SpecDecodeBaseProposer:
59
60
61
62
    def __init__(
        self,
        vllm_config: VllmConfig,
        device: torch.device,
63
        pass_hidden_states_to_model: bool,
Jiayi Yao's avatar
Jiayi Yao committed
64
        runner=None,
65
66
    ):
        self.vllm_config = vllm_config
67
        assert vllm_config.speculative_config is not None
68
69
70
        self.speculative_config = vllm_config.speculative_config
        self.draft_model_config = self.speculative_config.draft_model_config
        self.method = self.speculative_config.method
71
        self.pass_hidden_states_to_model = pass_hidden_states_to_model
72

Jiayi Yao's avatar
Jiayi Yao committed
73
        self.runner = runner
74
        self.device = device
75
        self.dtype = vllm_config.model_config.dtype
76
        self.max_model_len = vllm_config.model_config.max_model_len
Rémi Delacourt's avatar
Rémi Delacourt committed
77
        self.dp_rank = vllm_config.parallel_config.data_parallel_rank
78
        self.num_speculative_tokens = self.speculative_config.num_speculative_tokens
79
80
81
82
83
        # The drafter can get longer sequences than the target model.
        max_batch_size = vllm_config.scheduler_config.max_num_seqs
        self.max_num_tokens = (
            vllm_config.scheduler_config.max_num_batched_tokens + max_batch_size
        )
84
        self.token_arange_np = np.arange(self.max_num_tokens)
85
86
87
88
        # We need to get the hidden size from the draft model config because
        # the draft model's hidden size can be different from the target model's
        # hidden size (e.g., Llama 3.3 70B).
        self.hidden_size = self.draft_model_config.get_hidden_size()
89
        self.inputs_embeds_size = self.draft_model_config.get_inputs_embeds_size()
90

91
92
93
        # Multi-modal data support
        self.mm_registry = MULTIMODAL_REGISTRY
        self.supports_mm_inputs = self.mm_registry.supports_multimodal_inputs(
94
            vllm_config.model_config
95
        )
96

97
98
        self.attn_metadata_builder: AttentionMetadataBuilder | None = None
        self.draft_indexer_metadata_builder: AttentionMetadataBuilder | None = None
99
100
        self.attn_layer_names: list[str] = []
        self.indexer_layer_names: list[str] = []
101
102
103
        self.eagle3_use_aux_hidden_state: bool = (
            self._get_eagle3_use_aux_hidden_state_from_config()
        )
104

105
        self.compilation_config = self.vllm_config.compilation_config
106
107
108
109
110
111

        # Cudagraph dispatcher for PIECEWISE-only dispatching in eagle.
        # Keys are initialized later via initialize_cudagraph_keys() called from
        # gpu_model_runner._check_and_update_cudagraph_mode after
        # adjust_cudagraph_sizes_for_spec_decode is called.
        self.cudagraph_dispatcher = CudagraphDispatcher(self.vllm_config)
112

113
        # persistent buffers for cuda graph
114
115
116
        self.input_ids = torch.zeros(
            self.max_num_tokens, dtype=torch.int32, device=device
        )
117
118
119
        # Use draft model's M-RoPE setting, not target model's
        # Draft models may be text-only even if target is multimodal
        self.uses_mrope = self.draft_model_config.uses_mrope
120
121
        self.uses_xdrope_dim = self.vllm_config.model_config.uses_xdrope_dim
        self.draft_uses_xdrope_dim = self.draft_model_config.uses_xdrope_dim
122
        if self.uses_mrope:
123
124
125
126
127
128
129
130
131
132
            # NOTE: `mrope_positions` is implemented with one additional dummy
            # position on purpose to make it non-contiguous so that it can work
            # with torch compile.
            # See detailed explanation in https://github.com/vllm-project/vllm/pull/12128#discussion_r1926431923

            # NOTE: When M-RoPE is enabled, position ids are 3D regardless of
            # the modality of inputs. For text-only inputs, each dimension has
            # identical position IDs, making M-RoPE functionally equivalent to
            # 1D-RoPE.
            # See page 5 of https://arxiv.org/abs/2409.12191
133
            self.mrope_positions = torch.zeros(
134
                (3, self.max_num_tokens + 1), dtype=torch.int64, device=device
135
            )
136
137
138
139
140
141
        elif self.uses_xdrope_dim > 0 and self.draft_uses_xdrope_dim > 0:
            self.xdrope_positions = torch.zeros(
                (self.uses_xdrope_dim, self.max_num_tokens + 1),
                dtype=torch.int64,
                device=device,
            )
142
143
        else:
            # RoPE need (max_num_tokens,)
144
145
146
            self.positions = torch.zeros(
                self.max_num_tokens, dtype=torch.int64, device=device
            )
147
        self.hidden_states = torch.zeros(
148
149
            (self.max_num_tokens, self.hidden_size), dtype=self.dtype, device=device
        )
150

151
152
153
        # We need +1 here because the arange is used to set query_start_loc,
        # which has one more element than batch_size.
        max_num_slots_for_arange = max(max_batch_size + 1, self.max_num_tokens)
154
155
156
        self.arange = torch.arange(
            max_num_slots_for_arange, device=device, dtype=torch.int32
        )
157

158
        self.inputs_embeds = torch.zeros(
159
160
161
            (self.max_num_tokens, self.inputs_embeds_size),
            dtype=self.dtype,
            device=device,
162
        )
163

164
165
166
167
168
        self.backup_next_token_ids = CpuGpuBuffer(
            max_batch_size,
            dtype=torch.int32,
            pin_memory=is_pin_memory_available(),
            device=device,
169
170
            with_numpy=True,
        )
171

172
173
174
175
        self._slot_mapping_buffer = torch.zeros(
            self.max_num_tokens, dtype=torch.int64, device=device
        )

176
        # Determine allowed attention backends once during initialization.
177
        self.allowed_attn_types: tuple | None = None
178
        if current_platform.is_rocm():
179
180
181
182
183
184
            from vllm.v1.attention.backends.rocm_attn import RocmAttentionMetadata

            rocm_types = [
                TritonAttentionMetadata,
                RocmAttentionMetadata,
            ]
185
186
187
188
            # ROCM_AITER_FA is an optional backend
            if find_spec(
                AttentionBackendEnum.ROCM_AITER_FA.get_path(include_classname=False)
            ):
189
                from vllm.v1.attention.backends.rocm_aiter_fa import (
190
191
192
                    AiterFlashAttentionMetadata,
                )

193
                rocm_types.append(AiterFlashAttentionMetadata)
194
195

            # TRITON_MLA backend support for MLA models (e.g., DeepSeek)
196
197
198
            from vllm.model_executor.layers.attention.mla_attention import (
                MLACommonMetadata,
            )
199
200
201

            rocm_types.append(MLACommonMetadata)

202
203
204
205
206
            # FlexAttention backend support
            from vllm.v1.attention.backends.flex_attention import FlexAttentionMetadata

            rocm_types.append(FlexAttentionMetadata)

207
208
            self.allowed_attn_types = tuple(rocm_types)

209
210
        # Parse the speculative token tree.
        spec_token_tree = self.speculative_config.speculative_token_tree
211
        assert spec_token_tree is not None
212
        self.tree_choices: list[tuple[int, ...]] = ast.literal_eval(spec_token_tree)
213
214
215
216
217
218
219
220
        tree_depth = len(self.tree_choices[-1])
        # Precompute per-level properties of the tree.
        num_drafts_per_level = [0] * tree_depth
        for node in self.tree_choices:
            num_drafts_per_level[len(node) - 1] += 1
        self.cu_drafts_per_level = [num_drafts_per_level[0]]
        self.child_drafts_per_level = [num_drafts_per_level[0]]
        for level in range(1, tree_depth):
221
222
223
224
225
226
            self.cu_drafts_per_level.append(
                self.cu_drafts_per_level[-1] + num_drafts_per_level[level]
            )
            self.child_drafts_per_level.append(
                num_drafts_per_level[level] // num_drafts_per_level[level - 1]
            )
227
228
        # Precompute draft position offsets in flattened tree.
        self.tree_draft_pos_offsets = torch.arange(
229
            1, len(self.tree_choices) + 1, device=device, dtype=torch.int32
230
231
        ).repeat(max_batch_size, 1)

232
233
234
    def _get_positions(self, num_tokens: int):
        if self.uses_mrope:
            return self.mrope_positions[:, :num_tokens]
235
236
        if self.uses_xdrope_dim > 0 and self.draft_uses_xdrope_dim > 0:
            return self.xdrope_positions[:, :num_tokens]
237
238
239
240
241
        return self.positions[:num_tokens]

    def _set_positions(self, num_tokens: int, positions: torch.Tensor):
        if self.uses_mrope:
            self.mrope_positions[:, :num_tokens] = positions
242
243
        elif self.uses_xdrope_dim > 0 and self.draft_uses_xdrope_dim > 0:
            self.xdrope_positions[:, :num_tokens] = positions
244
        else:
245
246
247
248
249
            # Convert M-RoPE positions if target model uses M-RoPE
            # but draft doesn't, For text inputs, all M-RoPE
            # dimensions are identical
            if self.vllm_config.model_config.uses_mrope:
                positions = positions[0]
250
251
            self.positions[:num_tokens] = positions

252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
    def _get_slot_mapping(
        self,
        num_tokens: int,
        slot_mapping: torch.Tensor | None = None,
    ) -> dict[str, torch.Tensor]:
        """Return slot_mapping dict for EAGLE layers.

        If slot_mapping is provided, copies it into the buffer first.
        """
        if slot_mapping is not None:
            num_actual = slot_mapping.shape[0]
            self._slot_mapping_buffer[:num_actual].copy_(slot_mapping)
            if num_tokens > num_actual:
                self._slot_mapping_buffer[num_actual:num_tokens].fill_(PADDING_SLOT_ID)

        view = self._slot_mapping_buffer[:num_tokens]
        return {name: view for name in self.attn_layer_names + self.indexer_layer_names}

270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
    def initialize_cudagraph_keys(self, cudagraph_mode: CUDAGraphMode) -> None:
        """Initialize cudagraph dispatcher keys for eagle.

        Eagle only supports PIECEWISE cudagraphs (via mixed_mode).
        This should be called after adjust_cudagraph_sizes_for_spec_decode.
        """
        if (
            not self.speculative_config.enforce_eager
            and cudagraph_mode.mixed_mode()
            in [CUDAGraphMode.PIECEWISE, CUDAGraphMode.FULL]
        ):
            eagle_cudagraph_mode = CUDAGraphMode.PIECEWISE
        else:
            eagle_cudagraph_mode = CUDAGraphMode.NONE

        self.cudagraph_dispatcher.initialize_cudagraph_keys(eagle_cudagraph_mode)

287
288
289
290
    def propose(
        self,
        # [num_tokens]
        target_token_ids: torch.Tensor,
291
        # [num_tokens] or [3, num_tokens] when M-RoPE is enabled
292
293
294
295
296
        target_positions: torch.Tensor,
        # [num_tokens, hidden_size]
        target_hidden_states: torch.Tensor,
        # [batch_size]
        next_token_ids: torch.Tensor,
297
        last_token_indices: torch.Tensor | None,
298
        common_attn_metadata: CommonAttentionMetadata,
299
        sampling_metadata: SamplingMetadata,
300
        mm_embed_inputs: tuple[list[torch.Tensor], torch.Tensor] | None = None,
301
        num_rejected_tokens_gpu: torch.Tensor | None = None,
302
303
304
        slot_mappings: dict[str, torch.Tensor]
        | list[dict[str, torch.Tensor]]
        | None = None,
305
    ) -> torch.Tensor:
306
        batch_size = common_attn_metadata.batch_size()
307

308
309
310
        if self.method == "eagle3":
            assert isinstance(self.model, Eagle3LlamaForCausalLM)
            target_hidden_states = self.model.combine_hidden_states(
311
312
                target_hidden_states
            )
313
            assert target_hidden_states.shape[-1] == self.hidden_size
314
315
316
317
318
319
320
321
322
323
324

        num_tokens, last_token_indices, common_attn_metadata = (
            self.set_inputs_first_pass(
                target_token_ids=target_token_ids,
                next_token_ids=next_token_ids,
                target_positions=target_positions,
                last_token_indices=last_token_indices,
                cad=common_attn_metadata,
                num_rejected_tokens_gpu=num_rejected_tokens_gpu,
            )
        )
325

326
        assert self.runner is not None
Jiayi Yao's avatar
Jiayi Yao committed
327

328
329
330
331
332
        if self.attn_metadata_builder is None:
            attn_metadata_builder = self._get_attention_metadata_builder()
        else:
            attn_metadata_builder = self.attn_metadata_builder

333
        attn_metadata = attn_metadata_builder.build_for_drafting(
334
335
            common_attn_metadata=common_attn_metadata, draft_index=0
        )
336
337
338
339
340
341
        # FIXME: support hybrid kv for draft model (remove separate indexer)
        if self.draft_indexer_metadata_builder:
            draft_indexer_metadata = (
                self.draft_indexer_metadata_builder.build_for_drafting(
                    common_attn_metadata=common_attn_metadata,
                    draft_index=0,
342
343
                )
            )
344
345
        else:
            draft_indexer_metadata = None
346
347
348
349
350
        # At this moment, we assume all eagle layers belong to the same KV
        # cache group, thus using the same attention metadata.
        per_layer_attn_metadata = {}
        for layer_name in self.attn_layer_names:
            per_layer_attn_metadata[layer_name] = attn_metadata
351

352
353
354
355
        for layer_name in self.indexer_layer_names:
            assert draft_indexer_metadata is not None
            per_layer_attn_metadata[layer_name] = draft_indexer_metadata

Rémi Delacourt's avatar
Rémi Delacourt committed
356
        num_tokens_dp_padded, num_tokens_across_dp = self._pad_batch_across_dp(
357
            num_tokens_unpadded=num_tokens, num_tokens_padded=num_tokens
Rémi Delacourt's avatar
Rémi Delacourt committed
358
359
        )

360
361
362
363
        cudagraph_runtime_mode, batch_desc = self.cudagraph_dispatcher.dispatch(
            num_tokens_dp_padded
        )
        num_input_tokens = batch_desc.num_tokens
Rémi Delacourt's avatar
Rémi Delacourt committed
364
365
366
        if num_tokens_across_dp is not None:
            num_tokens_across_dp[self.dp_rank] = num_input_tokens

367
368
369
370
        if self.pass_hidden_states_to_model:
            # target_hidden_states and self.hidden_states can have different
            # hidden dims. E.g. large target model and small draft model.
            self.hidden_states[:num_tokens] = target_hidden_states
371
372
373
374

        if self.supports_mm_inputs:
            mm_embeds, is_mm_embed = mm_embed_inputs or (None, None)

375
            self.inputs_embeds[:num_tokens] = self.model.embed_input_ids(
376
377
378
                self.input_ids[:num_tokens],
                multimodal_embeddings=mm_embeds,
                is_multimodal=is_mm_embed,
379
            )
380

381
            input_ids = None
382
            inputs_embeds = self.inputs_embeds[:num_input_tokens]
383
384
        else:
            input_ids = self.input_ids[:num_input_tokens]
385
            inputs_embeds = None
386

387
388
389
390
391
392
393
394
        model_kwargs = {
            "input_ids": input_ids,
            "positions": self._get_positions(num_input_tokens),
            "inputs_embeds": inputs_embeds,
        }
        if self.pass_hidden_states_to_model:
            model_kwargs["hidden_states"] = self.hidden_states[:num_input_tokens]

395
        with set_forward_context(
396
397
398
            per_layer_attn_metadata,
            self.vllm_config,
            num_tokens=num_input_tokens,
Rémi Delacourt's avatar
Rémi Delacourt committed
399
            num_tokens_across_dp=num_tokens_across_dp,
400
            cudagraph_runtime_mode=cudagraph_runtime_mode,
401
402
403
            slot_mapping=self._get_slot_mapping(
                num_input_tokens, common_attn_metadata.slot_mapping
            ),
404
        ):
405
406
            ret_hidden_states = self.model(**model_kwargs)
            if not self.model_returns_tuple():
Jiayi Yao's avatar
Jiayi Yao committed
407
                last_hidden_states = ret_hidden_states
408
                hidden_states = last_hidden_states
Jiayi Yao's avatar
Jiayi Yao committed
409
410
            else:
                last_hidden_states, hidden_states = ret_hidden_states
411

412
        sample_hidden_states = last_hidden_states[last_token_indices]
413
        logits = self.model.compute_logits(sample_hidden_states)
414
415
416
417
418
419

        # Early exit if there is only one draft token to be generated.
        if self.num_speculative_tokens == 1:
            draft_token_ids = logits.argmax(dim=-1)
            return draft_token_ids.view(-1, 1)

420
        if self.uses_mrope:
421
            positions = self.mrope_positions[:, last_token_indices]
422
        else:
423
            positions = self.positions[last_token_indices]
424
425
426
427
428
429
        if self.method in (
            "deepseek_mtp",
            "ernie_mtp",
            "longcat_flash_mtp",
            "pangu_ultra_moe_mtp",
        ):
XuruiYang's avatar
XuruiYang committed
430
431
432
            hidden_states = self.hidden_states[last_token_indices]
        else:
            hidden_states = hidden_states[last_token_indices]
433
434
435

        if isinstance(attn_metadata, TreeAttentionMetadata):
            # Draft using tree attention.
436
437
438
439
440
441
            draft_token_ids_list = self.propose_tree(
                batch_size=batch_size,
                logits=logits,
                positions=positions,
                hidden_states=hidden_states,
                common_attn_metadata=common_attn_metadata,
442
                slot_mappings=slot_mappings,
443
444
445
446
            )
            # [batch_size, num_tree_tokens]
            return torch.cat(draft_token_ids_list, dim=1)

447
        draft_token_ids = logits.argmax(dim=-1)
448

449
450
451
        if self.allowed_attn_types is not None and not isinstance(
            attn_metadata, self.allowed_attn_types
        ):
452
453
454
455
            raise ValueError(
                f"Unsupported attention metadata type for speculative "
                "decoding with num_speculative_tokens > 1: "
                f"{type(attn_metadata)}. Supported types are: "
456
457
                f"{self.allowed_attn_types}"
            )
458

459
460
461
        # Generate the remaining draft tokens.
        draft_token_ids_list = [draft_token_ids]

Rémi Delacourt's avatar
Rémi Delacourt committed
462
        batch_size_dp_padded, batch_size_across_dp = self._pad_batch_across_dp(
463
            num_tokens_unpadded=batch_size, num_tokens_padded=batch_size
Rémi Delacourt's avatar
Rémi Delacourt committed
464
465
        )

466
467
468
469
        cudagraph_runtime_mode, batch_desc = self.cudagraph_dispatcher.dispatch(
            batch_size_dp_padded
        )
        input_batch_size = batch_desc.num_tokens
Rémi Delacourt's avatar
Rémi Delacourt committed
470
471
        if batch_size_across_dp is not None:
            batch_size_across_dp[self.dp_rank] = input_batch_size
472
473
474

        common_attn_metadata.num_actual_tokens = batch_size
        common_attn_metadata.max_query_len = 1
475
        common_attn_metadata.query_start_loc = self.arange[: batch_size + 1]
476
        common_attn_metadata.query_start_loc_cpu = torch.from_numpy(
477
478
            self.token_arange_np[: batch_size + 1]
        ).clone()
479
480
481
482
483
484
485
486
487
488
489

        # In padded drafter batch, we need to adjust the sequence lengths
        # to remove the "padding" (i.e. rejected tokens).
        # Only apply this adjustment when we have rejected tokens
        # (i.e., not the first proposal).
        if self.num_speculative_tokens > 1 and num_rejected_tokens_gpu is not None:
            common_attn_metadata.seq_lens -= num_rejected_tokens_gpu
            # Invalidate the CPU-side shadows to avoid H<>D sync.
            common_attn_metadata._seq_lens_cpu = None
            common_attn_metadata._num_computed_tokens_cpu = None

490
        for token_index in range(self.num_speculative_tokens - 1):
491
            # Update the inputs.
492
493
494
            # cast to int32 is crucial when eagle model is compiled.
            # tensor.argmax() returns int64 by default.
            input_ids = draft_token_ids_list[-1].int()
495
496
497
498
499
500
501
502
503
504
505
506
507
            if self.uses_mrope:
                positions += 1
                # NOTE(woosuk): We should handle the case where the draft model
                # generates tokens beyond the max model length.
                # Since it is complex to remove such requests from the batch,
                # we keep them in the batch but adjust the position ids
                # and slot mappings to avoid the
                # out-of-range access during the model execution.
                # The draft tokens generated with this adjustment
                # should be ignored.
                exceeds_max_model_len = positions[0] >= self.max_model_len
                # Mask out the position ids that exceed the max model length.
                # Otherwise, we may get out-of-range error in RoPE.
508
509
510
511
512
                clamped_positions = torch.where(
                    exceeds_max_model_len.unsqueeze(0),
                    torch.zeros_like(positions),
                    positions,
                )
513
514
515
            else:
                positions += 1
                exceeds_max_model_len = positions >= self.max_model_len
516
                clamped_positions = torch.where(exceeds_max_model_len, 0, positions)
517
518
519
            # For data integrity when async scheduling, we shouldn't use in place
            # operations in case they are modified in next step's `prepare_input`
            # of main model.
520
            # Increment the sequence lengths.
521
            common_attn_metadata.seq_lens += 1
522
523
            # For the requests that exceed the max model length, we set the
            # sequence length to 1 to minimize their overheads in attention.
524
            common_attn_metadata.seq_lens.masked_fill_(exceeds_max_model_len, 1)
525
526
527
528
529
530
            # Increment the maximum sequence length. We increment max_seq_len
            # unconditionally even though some seq_lens may have been capped above,
            # as max_seq_len serves as an upper bound for sequence lengths.
            common_attn_metadata.max_seq_len = min(
                common_attn_metadata.max_seq_len + 1, self.max_model_len
            )
531

532
533
534
535
536
537
            # Also update the CPU-side shadow; NOTE: this is hacky and should be
            # removed in when common_attn_metadata.seq_lens_cpu is deprecated.
            if common_attn_metadata._seq_lens_cpu is not None:
                common_attn_metadata._seq_lens_cpu += 1
            if common_attn_metadata._num_computed_tokens_cpu is not None:
                common_attn_metadata._num_computed_tokens_cpu += 1
538

539
            # Compute the slot mapping.
540
            block_size = attn_metadata_builder.kv_cache_spec.block_size
541
542
            if self.uses_mrope:
                # all dimensions of positions are the same
543
                block_numbers = clamped_positions[0] // block_size
544
            else:
545
                block_numbers = clamped_positions // block_size
546
            block_ids = common_attn_metadata.block_table_tensor.gather(
547
548
                dim=1, index=block_numbers.view(-1, 1)
            )
549
            block_ids = block_ids.view(-1)
550
551
            if self.uses_mrope:
                common_attn_metadata.slot_mapping = (
552
                    block_ids * block_size + clamped_positions[0] % block_size
553
                )
554
555
            else:
                common_attn_metadata.slot_mapping = (
556
                    block_ids * block_size + clamped_positions % block_size
557
                )
558
559
560
            # Mask out the slot mappings that exceed the max model length.
            # Otherwise, the KV cache will be inadvertently updated with the
            # padding tokens.
561
            common_attn_metadata.slot_mapping.masked_fill_(
562
563
                exceeds_max_model_len, PADDING_SLOT_ID
            )
564
565

            # Rebuild attention metadata
566
            attn_metadata = attn_metadata_builder.build_for_drafting(  # type: ignore
567
568
                common_attn_metadata=common_attn_metadata, draft_index=token_index + 1
            )
569
570
            for layer_name in self.attn_layer_names:
                per_layer_attn_metadata[layer_name] = attn_metadata
571

572
573
            # copy inputs to buffer for cudagraph
            self.input_ids[:batch_size] = input_ids
574
            self._set_positions(batch_size, clamped_positions)
575
            self.hidden_states[:batch_size] = hidden_states
576
            if self.supports_mm_inputs:
577
                self.inputs_embeds[:batch_size] = self.model.embed_input_ids(input_ids)
578

579
                input_ids = None
580
                inputs_embeds = self.inputs_embeds[:input_batch_size]
581
582
            else:
                input_ids = self.input_ids[:input_batch_size]
583
                inputs_embeds = None
584

585
            # Run the model.
586
587
588
589
590
591
592
593
            model_kwargs = {
                "input_ids": input_ids,
                "positions": self._get_positions(input_batch_size),
                "inputs_embeds": inputs_embeds,
            }
            if self.pass_hidden_states_to_model:
                model_kwargs["hidden_states"] = self.hidden_states[:input_batch_size]

594
            with set_forward_context(
595
596
597
                per_layer_attn_metadata,
                self.vllm_config,
                num_tokens=input_batch_size,
Rémi Delacourt's avatar
Rémi Delacourt committed
598
                num_tokens_across_dp=batch_size_across_dp,
599
                cudagraph_runtime_mode=cudagraph_runtime_mode,
600
601
602
                slot_mapping=self._get_slot_mapping(
                    input_batch_size, common_attn_metadata.slot_mapping
                ),
603
            ):
604
605
                ret_hidden_states = self.model(**model_kwargs)
                if not self.model_returns_tuple():
606
607
608
609
                    last_hidden_states = ret_hidden_states
                    hidden_states = ret_hidden_states
                else:
                    last_hidden_states, hidden_states = ret_hidden_states
610

611
            hidden_states = hidden_states[:batch_size]
612
            logits = self.model.compute_logits(last_hidden_states[:batch_size])
613
            draft_token_ids = logits.argmax(dim=-1)
614
615
616
617
            draft_token_ids_list.append(draft_token_ids)

        # [batch_size, num_speculative_tokens]
        draft_token_ids = torch.stack(draft_token_ids_list, dim=1)
618
        return draft_token_ids
619

620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
    def set_inputs_first_pass(
        self,
        target_token_ids: torch.Tensor,
        next_token_ids: torch.Tensor,
        target_positions: torch.Tensor,
        last_token_indices: torch.Tensor | None,
        cad: CommonAttentionMetadata,
        num_rejected_tokens_gpu: torch.Tensor | None,
    ) -> tuple[int, torch.Tensor, CommonAttentionMetadata]:
        if last_token_indices is None:
            last_token_indices = cad.query_start_loc[1:] - 1

        num_tokens = target_token_ids.shape[0]
        # Shift the input ids by one token.
        # E.g., [a1, b1, b2, c1, c2, c3] -> [b1, b2, c1, c2, c3, c3]
        self.input_ids[: num_tokens - 1] = target_token_ids[1:]
        # Replace the last token with the next token.
        # E.g., [b1, b2, c1, c2, c3, c3] -> [a2, b2, b3, c2, c3, c4]
        self.input_ids[last_token_indices] = next_token_ids

        # copy inputs to buffer for cudagraph
641
642
        if self.uses_xdrope_dim > 0 and self.draft_uses_xdrope_dim == 0:
            target_positions = target_positions[0]
643
644
645
646
647
648
649
        self._set_positions(num_tokens, target_positions)

        return num_tokens, last_token_indices, cad

    def model_returns_tuple(self) -> bool:
        return self.method not in ("mtp", "draft_model")

650
    def prepare_next_token_ids_cpu(
651
        self,
652
        sampled_token_ids: list[list[int]],
653
654
655
656
        requests: dict[str, CachedRequestState],
        gpu_input_batch: InputBatch,
        num_scheduled_tokens: dict[str, int],
    ) -> torch.Tensor:
657
658
659
660
661
662
663
664
665
666
        """
        This function is used to prepare the inputs for speculative decoding.
        It calculates the next token ids for each request based on the sampled
        token ids from the CPU. If a request has no sampled token ids (e.g.,
        during the initial decoding steps), it falls back to using the request
        state to get the next token id.
        """
        req_ids = gpu_input_batch.req_ids
        next_token_ids: list[int] = []
        for i, token_ids in enumerate(sampled_token_ids):
667
            if token_ids:
668
669
670
671
672
673
674
                # Common case.
                next_token_id = token_ids[-1]
            else:
                # Partial prefill (rare case).
                # Get the next token id from the request state.
                req_id = req_ids[i]
                req_state = requests[req_id]
675
                seq_len = req_state.num_computed_tokens + num_scheduled_tokens[req_id]
676
677
                next_token_id = req_state.get_token_id(seq_len)
            next_token_ids.append(next_token_id)
678
        next_token_ids = torch.tensor(
679
680
            next_token_ids, dtype=torch.int32, device=self.input_ids.device
        )
681
        return next_token_ids
682

683
684
685
686
687
688
    def prepare_next_token_ids_padded(
        self,
        common_attn_metadata: CommonAttentionMetadata,
        sampled_token_ids: torch.Tensor,
        requests: dict[str, CachedRequestState],
        gpu_input_batch: InputBatch,
689
        discard_request_mask: torch.Tensor,
690
    ) -> tuple[torch.Tensor, torch.Tensor]:
691
692
693
694
        """
        This function is used to prepare the inputs for speculative decoding.
        It calculates the next token ids and the number of valid sampled tokens
        for each request, considering the "discarded" requests whose next token
695
696
        is not sampled and comes from `request.get_token_id()` instead. This is denoted
        the "backup" token id. It also counts rejected tokens via `sampled_token_ids`.
697
698
699
        """
        # Precompute get_token_id for when there is no valid next token
        num_reqs = gpu_input_batch.num_reqs
700
701
702
703
704
705
        self.backup_next_token_ids.np[:num_reqs] = np.array(
            [
                requests[gpu_input_batch.req_ids[i]].get_token_id(
                    common_attn_metadata.seq_lens_cpu[i].item()
                )
                for i in range(num_reqs)
706
707
            ],
            dtype=np.int32,
708
        )
709
        self.backup_next_token_ids.copy_to_gpu(num_reqs)
710
        backup_tokens_gpu = self.backup_next_token_ids.gpu
711

712
713
        batch_size, num_tokens = sampled_token_ids.shape
        device = sampled_token_ids.device
714

715
716
        assert discard_request_mask.dtype == torch.bool
        assert backup_tokens_gpu.dtype == torch.int32
717

718
719
        next_token_ids = torch.empty(batch_size, dtype=torch.int32, device=device)
        valid_sampled_tokens_count = next_token_ids.new_empty(batch_size)
720

721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
        # Kernel grid: one program per request (row)
        grid = (batch_size,)

        # Find the next power of 2 for block sizes
        BLOCK_SIZE_TOKENS = triton.next_power_of_2(num_tokens)
        eagle_prepare_next_token_padded_kernel[grid](
            sampled_token_ids,
            discard_request_mask,
            backup_tokens_gpu,
            next_token_ids,
            valid_sampled_tokens_count,
            gpu_input_batch.vocab_size,
            num_tokens,
            batch_size,
            sampled_token_ids.stride(0),
            BLOCK_SIZE_TOKENS=BLOCK_SIZE_TOKENS,
737
        )
738
739
740

        return next_token_ids, valid_sampled_tokens_count

741
742
743
744
745
    def prepare_inputs_padded(
        self,
        common_attn_metadata: CommonAttentionMetadata,
        spec_decode_metadata: SpecDecodeMetadata,
        valid_sampled_tokens_count: torch.Tensor,
746
    ) -> tuple[CommonAttentionMetadata, torch.Tensor, torch.Tensor]:
747
748
749
750
751
752
        """
        This function is used to prepare the inputs for speculative decoding
        It updates the common_attn_metadata for speculative decoding,
        but does not consider the rejected tokens. Instead, all tokens
        are included as inputs to the speculator, with the rejected tokens
        used as padding and filtered out later by `token_indices_to_sample`.
753
        No blocking CPU operations should be introduced in this function.
754
        """
755
756
757
758
759
        num_reqs = common_attn_metadata.num_reqs
        device = valid_sampled_tokens_count.device

        token_indices_to_sample = torch.empty(
            (num_reqs,), dtype=torch.int32, device=device
760
        )
761
762
763
        num_rejected_tokens_gpu = torch.empty(
            (num_reqs,), dtype=torch.int32, device=device
        )
764

765
766
767
768
769
770
        grid = (num_reqs,)
        eagle_prepare_inputs_padded_kernel[grid](
            spec_decode_metadata.cu_num_draft_tokens,
            valid_sampled_tokens_count,
            common_attn_metadata.query_start_loc,
            token_indices_to_sample,
771
            num_rejected_tokens_gpu,
772
            num_reqs,
773
        )
774
775

        query_start_loc_cpu = common_attn_metadata.query_start_loc_cpu
776
        new_query_len_per_req = query_start_loc_cpu[1:] - query_start_loc_cpu[:-1]
777
778
779
780
781
782
783

        total_num_tokens = query_start_loc_cpu[-1].item()

        spec_common_attn_metadata = CommonAttentionMetadata(
            query_start_loc=common_attn_metadata.query_start_loc,
            seq_lens=common_attn_metadata.seq_lens,
            query_start_loc_cpu=query_start_loc_cpu,
784
785
            _seq_lens_cpu=common_attn_metadata._seq_lens_cpu,
            _num_computed_tokens_cpu=common_attn_metadata._num_computed_tokens_cpu,
786
787
788
789
790
            num_reqs=common_attn_metadata.num_reqs,
            num_actual_tokens=total_num_tokens,
            max_query_len=new_query_len_per_req.max().item(),
            max_seq_len=common_attn_metadata.seq_lens_cpu.max().item(),
            block_table_tensor=common_attn_metadata.block_table_tensor,
791
            slot_mapping=common_attn_metadata.slot_mapping[:total_num_tokens],
792
            causal=True,
793
            dcp_local_seq_lens=common_attn_metadata.dcp_local_seq_lens,
794
795
        )

796
797
798
799
800
        return (
            spec_common_attn_metadata,
            token_indices_to_sample,
            num_rejected_tokens_gpu,
        )
801

802
803
804
805
806
807
808
809
810
811
    def propose_tree(
        self,
        batch_size: int,
        # [num_tokens, vocab_size]
        logits: torch.Tensor,
        # [num_tokens]
        positions: torch.Tensor,
        # [num_tokens, hidden_size]
        hidden_states: torch.Tensor,
        common_attn_metadata: CommonAttentionMetadata,
812
813
814
        slot_mappings: dict[str, torch.Tensor]
        | list[dict[str, torch.Tensor]]
        | None = None,
815
    ) -> list[torch.Tensor]:
816
817
818
819
        tree_attn_metadata_builder = self.runner.attn_groups[0][
            0
        ].get_metadata_builder()
        assert isinstance(tree_attn_metadata_builder, TreeAttentionMetadataBuilder)
820

821
        total_num_drafts = self.cu_drafts_per_level[0]
822
823
        level_num_drafts = total_num_drafts
        # Sample a draft token for each child at the tree root level.
824
        num_children = self.child_drafts_per_level[0]
825
826
827
        if num_children == 1:
            draft_token_ids = logits.argmax(dim=-1).view(batch_size, -1)
        else:
828
829
830
            draft_token_ids = torch.topk(logits, num_children, dim=-1).indices.view(
                batch_size, -1
            )
831
832
833
834
        draft_token_ids_list = [draft_token_ids]
        draft_hidden_states = hidden_states.view(batch_size, 1, -1)

        # Initialize empty tensors for concatenation with the level outputs.
835
836
837
838
839
840
841
842
843
        tree_input_ids = torch.empty(
            0, device=self.input_ids.device, dtype=self.input_ids.dtype
        )
        tree_positions = torch.empty(
            0, device=self.positions.device, dtype=self.positions.dtype
        )
        tree_hidden_states = torch.empty(
            0, device=self.hidden_states.device, dtype=self.hidden_states.dtype
        )
844
845
        # Precompute the draft token positions.
        flattened_draft_positions = (
846
847
            positions.view(batch_size, -1) + self.tree_draft_pos_offsets[:batch_size, :]
        )
848
        tree_depth = len(self.cu_drafts_per_level)
849
        for level in range(tree_depth - 1):
850
851
            # Get draft positions for RoPE.
            draft_positions = positions + (level + 1)
852
            exceeds_max_model_len = (positions + total_num_drafts) >= self.max_model_len
853
854
            # Mask out the position ids that exceed the max model length.
            # Otherwise, we may get out-of-range error in RoPE.
855
            draft_positions = torch.where(
856
857
858
                exceeds_max_model_len,
                0,
                draft_positions,
859
860
            ).view(batch_size, -1)

861
862
            if level_num_drafts > 1:
                # Repeat the positions for each draft at this level.
863
                draft_positions = draft_positions.repeat_interleave(
864
865
                    level_num_drafts, dim=1
                )
866
867
868
869

            if num_children > 1:
                # Repeat draft hidden states for each child.
                draft_hidden_states = draft_hidden_states.repeat_interleave(
870
871
                    num_children, dim=1
                )
872
873

            # Concatenate the draft tokens, positions, and hidden states.
874
875
            tree_input_ids = torch.cat([tree_input_ids, draft_token_ids], dim=1)
            tree_positions = torch.cat([tree_positions, draft_positions], dim=1)
876
            tree_hidden_states = torch.cat(
877
878
                [tree_hidden_states, draft_hidden_states], dim=1
            )
879
880
881

            # Build new attention metadata for the next level of drafts.
            # This is necessary to support tree attention.
882
            query_len = total_num_drafts
883
884
            common_attn_metadata = replace(
                common_attn_metadata,
885
                query_start_loc=query_len * self.arange[: batch_size + 1],
886
887
888
889
890
                seq_lens=common_attn_metadata.seq_lens + level_num_drafts,
                num_actual_tokens=batch_size * query_len,
                max_query_len=query_len,
            )
            attn_metadata = tree_attn_metadata_builder.build_for_drafting(
891
                common_attn_metadata=common_attn_metadata, draft_index=level + 1
892
893
894
895
896
897
898
899
            )

            # Apply new attention metadata to all layers.
            per_layer_attn_metadata = {}
            for layer_name in self.attn_layer_names:
                per_layer_attn_metadata[layer_name] = attn_metadata

            # Consider max model length.
900
901
902
            attn_metadata.max_seq_len = min(
                attn_metadata.max_seq_len, self.max_model_len
            )
903
904
905
906
907
            # For the requests that exceed the max model length, we set the
            # sequence length to 1 to minimize their overheads in attention.
            attn_metadata.seq_lens.masked_fill_(exceeds_max_model_len, 1)

            # Compute the slot mapping.
908
            block_size = tree_attn_metadata_builder.kv_cache_spec.block_size
909
            query_positions = flattened_draft_positions[:, level : level + query_len]
910
            block_numbers = query_positions // block_size
911
            block_ids = attn_metadata.block_table.gather(dim=1, index=block_numbers)
912
            slot_mapping = block_ids * block_size + query_positions % block_size
913
914
915
916
917
918
919
920
921
922
923
            # Mask out the slot mappings that exceed the max model length.
            # Otherwise, the KV cache will be inadvertently updated with the
            # padding tokens.
            slot_mapping[exceeds_max_model_len] = PADDING_SLOT_ID
            attn_metadata.slot_mapping = slot_mapping.view(-1)

            # Copy inputs to buffer for cudagraph.
            num_tokens = attn_metadata.num_actual_tokens
            input_ids = tree_input_ids.view(-1)
            self.input_ids[:num_tokens] = input_ids
            self.positions[:num_tokens] = tree_positions.view(-1)
924
            self.hidden_states[:num_tokens] = tree_hidden_states.view(num_tokens, -1)
925

926
927
928
929
            cudagraph_runtime_mode, batch_desc = self.cudagraph_dispatcher.dispatch(
                num_tokens
            )
            num_input_tokens = batch_desc.num_tokens
930
            # Run the model.
931
            with set_forward_context(
932
933
934
935
                per_layer_attn_metadata,
                self.vllm_config,
                num_tokens=num_input_tokens,
                cudagraph_runtime_mode=cudagraph_runtime_mode,
936
937
938
                slot_mapping=self._get_slot_mapping(
                    num_input_tokens, attn_metadata.slot_mapping
                ),
939
            ):
940
941
942
943
944
945
946
947
948
                last_hidden_states, hidden_states = self.model(
                    input_ids=self.input_ids[:num_input_tokens],
                    positions=self.positions[:num_input_tokens],
                    hidden_states=self.hidden_states[:num_input_tokens],
                    inputs_embeds=None,
                )

            # Get the output hidden states for the draft tokens.
            draft_hidden_states = hidden_states[:num_tokens].view(
949
950
                batch_size, query_len, -1
            )[:, -level_num_drafts:]
951
            draft_last_hidden_states = last_hidden_states[:num_tokens].view(
952
953
                batch_size, query_len, -1
            )[:, -level_num_drafts:]
954
955
956

            # Get the output logits for the draft tokens.
            logits = self.model.compute_logits(
957
958
                draft_last_hidden_states.reshape(batch_size * level_num_drafts, -1)
            )
959
960
961
962
963
964

            # Sample a draft token for each child at the next tree level.
            num_children = self.child_drafts_per_level[level + 1]
            if num_children == 1:
                draft_token_ids = logits.argmax(dim=-1).view(batch_size, -1)
            else:
965
966
967
                draft_token_ids = torch.topk(logits, num_children, dim=-1).indices.view(
                    batch_size, -1
                )
968
969
970
            draft_token_ids_list.append(draft_token_ids)

            # Update the # drafts counters for the next tree level.
971
            level_num_drafts = self.cu_drafts_per_level[level + 1] - total_num_drafts
972
973
974
            total_num_drafts = self.cu_drafts_per_level[level + 1]
        return draft_token_ids_list

975
    def prepare_inputs(
976
977
        self,
        common_attn_metadata: CommonAttentionMetadata,
978
979
        sampled_token_ids: list[list[int]],
        num_draft_tokens: list[int],
980
981
    ) -> tuple[CommonAttentionMetadata, torch.Tensor]:
        """
982
        This function is used to prepare the inputs for speculative decoding.
983
984
985
986
987
988
        It updates to the common_attn_metadata to account for the rejected
        tokens (and newly sampled tokens). It also returns the token indices
        of the tokens that should be fed to the speculator.
        """
        # E.g.
        #  common_attn_metadata.query_start_loc{_cpu}:
989
        #       [0, q1, q1 + q2, q1 + q2 + q3]
990
991
992
993
994
995
        #  common_attn_metadata.seq_lens{_cpu}: [s1, s2, s3]
        #  num_rejected_tokens: [n1, n2, n3]
        # This function computes the intermediate values:
        #  num_tokens_per_req: [q1 - n1, q2 - n2, q3 - n3]
        # And returns:
        #  common_attn_metadata.query_start_loc{_cpu}:
996
        #       [0, q1 - n1, q1 + q2 - n1 - n2, q1 + q2 + q3 - n1 - n2 - n3]
997
        #  common_attn_metadata.seq_lens{_cpu}:
998
        #       [s1 - n1 + 1, s2 - n2 + 1, s3 - n3 + 1]
999
        #  token_indices: [0, 1, ..., q1 - n1 - 1,
1000
1001
        #                 q1, q1 + 1, ..., q1 + q2 - n2 - 1,
        #                 q1 + q2, q1 + q2 + 1, ..., q1 + q2 + q3 - n3 - 1]
1002

1003
1004
1005
1006
        num_rejected_tokens = [
            n + 1 - len(sampled_token_ids[i]) if n > 0 else 0
            for i, n in enumerate(num_draft_tokens)
        ]
1007
        num_rejected_tokens = torch.tensor(num_rejected_tokens, dtype=torch.int32)
1008

1009
1010
        device = common_attn_metadata.query_start_loc.device
        query_start_loc_cpu = common_attn_metadata.query_start_loc_cpu
1011
        new_seq_lens_cpu = common_attn_metadata.seq_lens_cpu - num_rejected_tokens
1012
1013

        # [0, q1, q1 + q2, q1 + q2 + q3] -> [q1, q2, q3]
1014
        new_query_len_per_req = query_start_loc_cpu[1:] - query_start_loc_cpu[:-1]
1015
1016
1017
1018
1019
1020
1021
1022
        # [q1, q2, q3] -> [q1 - n1, q2 - n2, q3 - n3]
        new_num_tokens_per_req = new_query_len_per_req - num_rejected_tokens
        new_num_tokens_per_req_np = new_num_tokens_per_req.numpy()

        # [q1 - n1, q2 - n2, q3 - n3] ->
        # [0, q1 - n1, q1 + q2 - n1 - n2, q1 + q2 + q3 - n1 - n2 - n3]
        new_query_start_loc_cpu = torch.zeros(
            query_start_loc_cpu.shape,
1023
            dtype=torch.int32,
1024
1025
            pin_memory=is_pin_memory_available(),
        )
1026
1027
1028
1029
1030
1031
1032
1033
1034
        new_query_start_loc_np = new_query_start_loc_cpu.numpy()
        np.cumsum(new_num_tokens_per_req_np, out=new_query_start_loc_np[1:])

        total_num_tokens = new_query_start_loc_np[-1]
        # Example assuming num_tokens_per_req_np = [2, 4, 3]
        # this implies that `new_query_start_locs` is:
        # [0, 2, 6, 9] ->
        # [0, 0, 2, 2, 2, 2, 6, 6, 6]
        #  _r1_  ____r2____  ___r3__
1035
1036
1037
        new_query_start_locs_expanded = np.repeat(
            new_query_start_loc_np[:-1], new_num_tokens_per_req_np
        )
1038
1039
1040
        # [0, 1, 2, 3, 4, 5, 6, 7, 8] ->
        # [0, 1, 0, 1, 2, 3, 0, 1, 2]
        #  _r1_  ____r2____  ___r3__
1041
        token_offsets = (
1042
1043
            self.token_arange_np[:total_num_tokens] - new_query_start_locs_expanded
        )
1044
1045
1046
1047
1048
1049

        # Expand starting positions to match token pattern
        # [0, q1, q1 + q2] ->
        # [0, 0, q1, q1, q1, q1, q1 + q2, q1 + q2, q1 + q2]
        #  _r1_  _____r2_______  ___________r3____________
        old_query_start_locs_expanded = np.repeat(
1050
1051
            query_start_loc_cpu[:-1].numpy(), new_num_tokens_per_req_np
        )
1052
        # Final token indices are:
1053
1054
1055
        # [0, 1,                                // req 1
        #  q1 + 0, q1 + 1, q1 + 2, q1 + 3,       // req 2
        #  q1 + q2 + 0, q1 + q2 + 1, q1 + q2 + 2] // req 3
1056
        token_indices_np = token_offsets + old_query_start_locs_expanded
1057
        token_indices = torch.from_numpy(token_indices_np).to(device, non_blocking=True)
1058
1059

        spec_common_attn_metadata = CommonAttentionMetadata(
1060
            query_start_loc=new_query_start_loc_cpu.to(device, non_blocking=True),
1061
1062
            seq_lens=new_seq_lens_cpu.to(device, non_blocking=True),
            query_start_loc_cpu=new_query_start_loc_cpu,
1063
1064
            _seq_lens_cpu=new_seq_lens_cpu,
            _num_computed_tokens_cpu=common_attn_metadata._num_computed_tokens_cpu,
1065
1066
1067
            num_reqs=common_attn_metadata.num_reqs,
            num_actual_tokens=total_num_tokens,
            max_query_len=new_query_len_per_req.max().item(),
1068
            max_seq_len=new_seq_lens_cpu.max().item(),
1069
1070
            block_table_tensor=common_attn_metadata.block_table_tensor,
            slot_mapping=common_attn_metadata.slot_mapping[token_indices],
1071
            causal=True,
1072
            dcp_local_seq_lens=common_attn_metadata.dcp_local_seq_lens,
1073
        )
1074
1075

        return spec_common_attn_metadata, token_indices
1076

1077
    def get_model_name(self, model: nn.Module) -> str:
1078
        if hasattr(model, "module"):  # multi-GPU
1079
1080
1081
            model = model.module
        return model.__class__.__name__

1082
    def load_model(self, target_model: nn.Module) -> None:
1083
        draft_model_config = self.speculative_config.draft_model_config
1084
        target_attn_layer_names = set(
1085
1086
1087
1088
            get_layers_from_vllm_config(
                self.vllm_config,
                AttentionLayerBase,  # type: ignore[type-abstract]
            ).keys()
1089
        )
1090
1091
        # FIXME: support hybrid kv for draft model
        target_indexer_layer_names = set(
1092
1093
1094
1095
            get_layers_from_vllm_config(
                self.vllm_config, DeepseekV32IndexerCache
            ).keys()
        )
1096

1097
        from vllm.compilation.backends import set_model_tag
1098

1099
        with set_model_tag("eagle_head"):
1100
1101
1102
            self.model = get_model(
                vllm_config=self.vllm_config, model_config=draft_model_config
            )
1103

1104
        draft_attn_layer_names = (
1105
1106
1107
1108
            get_layers_from_vllm_config(
                self.vllm_config,
                AttentionLayerBase,  # type: ignore[type-abstract]
            ).keys()
1109
1110
1111
1112
1113
1114
            - target_attn_layer_names
        )
        indexer_layers = get_layers_from_vllm_config(
            self.vllm_config, DeepseekV32IndexerCache
        )
        draft_indexer_layer_names = indexer_layers.keys() - target_indexer_layer_names
1115
        self.attn_layer_names = list(draft_attn_layer_names - draft_indexer_layer_names)
1116
1117
1118
1119
1120
        self.indexer_layer_names = list(draft_indexer_layer_names)

        if self.indexer_layer_names:
            first_layer = self.indexer_layer_names[0]
            self.draft_indexer_metadata_builder = (
1121
1122
1123
                indexer_layers[first_layer]
                .get_attn_backend()
                .get_builder_cls()(
1124
                    indexer_layers[first_layer].get_kv_cache_spec(self.vllm_config),
1125
1126
1127
                    self.indexer_layer_names,
                    self.vllm_config,
                    self.device,
1128
1129
                )
            )
1130
1131
        else:
            self.draft_indexer_metadata_builder = None
1132

1133
        if self.supports_mm_inputs:
1134
1135
1136
            # Even if the target model is multimodal, we can also use
            # text-only draft models
            try:
1137
                dummy_input_ids = torch.tensor([[1]], device=self.input_ids.device)
1138
                self.model.embed_input_ids(dummy_input_ids, multimodal_embeddings=None)
1139
1140
1141
            except (NotImplementedError, AttributeError, TypeError):
                logger.warning(
                    "Draft model does not support multimodal inputs, "
1142
1143
                    "falling back to text-only mode"
                )
1144
                self.supports_mm_inputs = False
1145

1146
1147
        if supports_multimodal(target_model):
            # handle multimodality
1148
            assert hasattr(target_model, "config")
1149
1150
1151
            if self.get_model_name(target_model) in [
                "Qwen2_5_VLForConditionalGeneration",
                "Qwen3VLForConditionalGeneration",
1152
                "Qwen3VLMoeForConditionalGeneration",
1153
                "HunYuanVLForConditionalGeneration",
1154
                "GlmOcrForConditionalGeneration",
1155
            ]:
1156
                self.model.config.image_token_index = target_model.config.image_token_id
1157
1158
1159
1160
            elif self.get_model_name(target_model) == "PixtralForConditionalGeneration":
                self.model.config.image_token_index = (
                    target_model.config.vision_config.image_token_id
                )
1161
1162
            else:
                self.model.config.image_token_index = (
1163
1164
                    target_model.config.image_token_index
                )
1165
1166
1167
            target_language_model = cast(
                SupportsMultiModal, target_model
            ).get_language_model()
1168
1169
        else:
            target_language_model = target_model
1170

1171
        # share embed_tokens with the target model if needed
1172
        if get_pp_group().world_size == 1:
1173
1174
1175
1176
1177
1178
1179
            inner_model = getattr(target_language_model, "model", None)
            if inner_model is None:
                raise AttributeError("Target model does not have 'model' attribute")
            if hasattr(inner_model, "embed_tokens"):
                target_embed_tokens = inner_model.embed_tokens
            elif hasattr(inner_model, "embedding"):
                target_embed_tokens = inner_model.embedding
1180
1181
            else:
                raise AttributeError(
1182
1183
                    "Target model does not have 'embed_tokens' or 'embedding' attribute"
                )
1184

1185
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196
1197
            share_embeddings = False
            if hasattr(self.model, "has_own_embed_tokens"):
                # EAGLE model
                if not self.model.has_own_embed_tokens:
                    share_embeddings = True
                    logger.info(
                        "Detected EAGLE model without its own embed_tokens in the"
                        " checkpoint. Sharing target model embedding weights with the"
                        " draft model."
                    )
                elif (
                    isinstance(target_embed_tokens.weight, torch.Tensor)
                    and isinstance(self.model.model.embed_tokens.weight, torch.Tensor)
1198
1199
1200
                    # TODO: Offload to CPU for comparison to avoid extra GPU memory
                    # usage in CI testing environments with limited GPU memory
                    and torch.equal(
1201
1202
                        target_embed_tokens.weight.cpu(),
                        self.model.model.embed_tokens.weight.cpu(),
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
                    )
                ):
                    share_embeddings = True
                    logger.info(
                        "Detected EAGLE model with embed_tokens identical to the target"
                        " model. Sharing target model embedding weights with the draft"
                        " model."
                    )
                else:
                    logger.info(
                        "Detected EAGLE model with distinct embed_tokens weights. "
                        "Keeping separate embedding weights from the target model."
                    )
1216
            else:
1217
1218
                # MTP model
                share_embeddings = True
1219
                logger.info(
1220
1221
                    "Detected MTP model. "
                    "Sharing target model embedding weights with the draft model."
1222
                )
1223
1224
1225
1226
1227

            if share_embeddings:
                if hasattr(self.model.model, "embed_tokens"):
                    del self.model.model.embed_tokens
                self.model.model.embed_tokens = target_embed_tokens
1228
        else:
1229
            logger.info(
1230
                "The draft model's vocab embedding will be loaded separately"
1231
1232
                " from the target model."
            )
1233
1234

        # share lm_head with the target model if needed
1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
1245
1246
1247
        share_lm_head = False
        if hasattr(self.model, "has_own_lm_head"):
            # EAGLE model
            if not self.model.has_own_lm_head:
                share_lm_head = True
                logger.info(
                    "Detected EAGLE model without its own lm_head in the checkpoint. "
                    "Sharing target model lm_head weights with the draft model."
                )
            elif (
                hasattr(target_language_model, "lm_head")
                and isinstance(target_language_model.lm_head.weight, torch.Tensor)
                and isinstance(self.model.lm_head.weight, torch.Tensor)
1248
1249
                # TODO: Offload to CPU for comparison to avoid extra GPU memory
                # usage in CI testing environments with limited GPU memory
1250
                and torch.equal(
1251
1252
                    target_language_model.lm_head.weight.cpu(),
                    self.model.lm_head.weight.cpu(),
1253
                )
1254
            ):
1255
                share_lm_head = True
1256
                logger.info(
1257
1258
                    "Detected EAGLE model with lm_head identical to the target model. "
                    "Sharing target model lm_head weights with the draft model."
1259
                )
1260
1261
            else:
                logger.info(
1262
1263
                    "Detected EAGLE model with distinct lm_head weights. "
                    "Keeping separate lm_head weights from the target model."
1264
                )
1265
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275
1276
        else:
            # MTP model
            share_lm_head = True
            logger.info(
                "Detected MTP model. "
                "Sharing target model lm_head weights with the draft model."
            )

        if share_lm_head and hasattr(target_language_model, "lm_head"):
            if hasattr(self.model, "lm_head"):
                del self.model.lm_head
            self.model.lm_head = target_language_model.lm_head
1277

1278
1279
1280
1281
    @torch.inference_mode()
    def dummy_run(
        self,
        num_tokens: int,
1282
1283
        use_cudagraphs: bool = True,
        is_graph_capturing: bool = False,
1284
        slot_mappings: dict[str, torch.Tensor] | None = None,
1285
    ) -> None:
Rémi Delacourt's avatar
Rémi Delacourt committed
1286
1287
1288
1289
        # FIXME: when using tree-based specdec, adjust number of forward-passes
        # according to the depth of the tree.
        for fwd_idx in range(
            self.num_speculative_tokens if not is_graph_capturing else 1
1290
        ):
Rémi Delacourt's avatar
Rémi Delacourt committed
1291
1292
            if fwd_idx <= 1:
                num_tokens_dp_padded, num_tokens_across_dp = self._pad_batch_across_dp(
1293
                    num_tokens_unpadded=num_tokens, num_tokens_padded=num_tokens
Rémi Delacourt's avatar
Rémi Delacourt committed
1294
                )
1295
1296
1297
1298
1299
1300
1301
1302
                if use_cudagraphs:
                    cudagraph_runtime_mode, batch_desc = (
                        self.cudagraph_dispatcher.dispatch(num_tokens_dp_padded)
                    )
                    num_input_tokens = batch_desc.num_tokens
                else:
                    cudagraph_runtime_mode = CUDAGraphMode.NONE
                    num_input_tokens = num_tokens_dp_padded
Rémi Delacourt's avatar
Rémi Delacourt committed
1303
1304
                if num_tokens_across_dp is not None:
                    num_tokens_across_dp[self.dp_rank] = num_input_tokens
1305

1306
1307
1308
1309
1310
1311
1312
1313
1314
1315
            # Make sure to use EAGLE's own buffer during cudagraph capture.
            if (
                self.attn_layer_names
                and slot_mappings is not None
                and self.attn_layer_names[0] in slot_mappings
            ):
                slot_mapping_dict = self._get_slot_mapping(num_input_tokens)
            else:
                slot_mapping_dict = slot_mappings or {}

Rémi Delacourt's avatar
Rémi Delacourt committed
1316
1317
1318
1319
1320
            with set_forward_context(
                None,
                self.vllm_config,
                num_tokens=num_input_tokens,
                num_tokens_across_dp=num_tokens_across_dp,
1321
                cudagraph_runtime_mode=cudagraph_runtime_mode,
1322
                slot_mapping=slot_mapping_dict,
Rémi Delacourt's avatar
Rémi Delacourt committed
1323
1324
1325
1326
1327
1328
1329
1330
            ):
                if self.supports_mm_inputs:
                    input_ids = None
                    inputs_embeds = self.inputs_embeds[:num_input_tokens]
                else:
                    input_ids = self.input_ids[:num_input_tokens]
                    inputs_embeds = None

1331
                kwargs = dict(
Rémi Delacourt's avatar
Rémi Delacourt committed
1332
1333
1334
1335
                    input_ids=input_ids,
                    positions=self._get_positions(num_input_tokens),
                    inputs_embeds=inputs_embeds,
                )
1336
1337
1338
                if self.pass_hidden_states_to_model:
                    kwargs["hidden_states"] = self.hidden_states[:num_input_tokens]
                self.model(**kwargs)
1339

1340
    def _get_attention_metadata_builder(self) -> AttentionMetadataBuilder:
1341
        """Find and return the attention metadata builders for EAGLE layers.
1342

1343
1344
        Returns:
            The metadata builders for EAGLE layers.
1345

1346
1347
1348
1349
1350
1351
1352
1353
1354
1355
1356
1357
1358
1359
1360
        Raises:
            AssertionError: If no metadata builders are found for EAGLE layers.
        """
        builder = None
        chosen_layer = self.attn_layer_names[0]

        for kv_cache_group in self.runner.attn_groups:
            for attn_group in kv_cache_group:
                if chosen_layer in attn_group.layer_names:
                    builder = attn_group.get_metadata_builder()
                    break
            if builder is not None:
                break

        assert builder is not None, (
1361
1362
            "Failed to find attention metadata builder for EAGLE layers."
        )
1363
1364
        return builder

1365
1366
1367
1368
1369
1370
1371
1372
1373
1374
1375
1376
1377
1378
1379
1380
    def _get_eagle3_use_aux_hidden_state_from_config(self) -> bool:
        """
        Some eagle3 heads (e.g., nvidia/gpt-oss-120b-Eagle3-v2) do not use auxiliary
        hidden states and directly uses the last layer output just like eagle1.
        They might indicate this by setting "use_aux_hidden_state" to False
        inside the "eagle_config" dict of their hf_config.
        """
        if self.method != "eagle3":
            return False
        # Assume that eagle3 heads use aux hidden states by default
        use_aux_hidden_state = True
        eagle_config = getattr(self.draft_model_config.hf_config, "eagle_config", None)
        if eagle_config is not None:
            use_aux_hidden_state = eagle_config.get("use_aux_hidden_state", True)
        return use_aux_hidden_state

1381
    def validate_same_kv_cache_group(self, kv_cache_config: KVCacheConfig) -> None:
1382
        """
1383
1384
        Validate that all drafting layers belong to the same KVCacheGroup.
        Need this assumption to ensure all drafting layers can use the
1385
1386
1387
1388
1389
1390
1391
        same AttentionMetadata.
        May extend to multiple AttentionMetadata in the future.
        """
        kv_cache_groups: dict[str, int] = {}
        for id, kv_cache_group in enumerate(kv_cache_config.kv_cache_groups):
            for layer_name in kv_cache_group.layer_names:
                kv_cache_groups[layer_name] = id
1392
1393
1394
1395
1396
1397
1398
1399
1400
1401
        assert (
            len(
                set(
                    [
                        kv_cache_groups[layer_name]
                        for layer_name in self.attn_layer_names
                    ]
                )
            )
            == 1
1402
        ), "All drafting layers should belong to the same kv cache group"
1403

Rémi Delacourt's avatar
Rémi Delacourt committed
1404
1405
1406
1407
1408
1409
    def _pad_batch_across_dp(
        self,
        num_tokens_unpadded: int,
        num_tokens_padded: int,
    ) -> tuple[int, torch.Tensor]:
        # TODO(Flechman): support DBO ubatching
1410
        should_ubatch, num_toks_across_dp, _ = coordinate_batch_across_dp(
Rémi Delacourt's avatar
Rémi Delacourt committed
1411
1412
1413
            num_tokens_unpadded=num_tokens_unpadded,
            parallel_config=self.vllm_config.parallel_config,
            allow_microbatching=False,
1414
1415
            allow_dp_padding=self.cudagraph_dispatcher.cudagraph_mode
            != CUDAGraphMode.NONE,
Rémi Delacourt's avatar
Rémi Delacourt committed
1416
1417
1418
1419
            num_tokens_padded=num_tokens_padded,
            uniform_decode=None,
            num_scheduled_tokens_per_request=None,
        )
1420
        assert not should_ubatch, "DBO ubatching not implemented for EAGLE"
Rémi Delacourt's avatar
Rémi Delacourt committed
1421
1422
1423
1424
1425
1426

        num_tokens_dp_padded = num_tokens_padded
        if num_toks_across_dp is not None:
            num_tokens_dp_padded = int(num_toks_across_dp[self.dp_rank].item())
        return num_tokens_dp_padded, num_toks_across_dp

1427

1428
1429
1430
1431
1432
1433
1434
1435
1436
1437
1438
1439
1440
1441
1442
class EagleProposer(SpecDecodeBaseProposer):
    def __init__(
        self,
        vllm_config: VllmConfig,
        device: torch.device,
        runner=None,
    ):
        super().__init__(
            vllm_config,
            device,
            pass_hidden_states_to_model=True,
            runner=runner,
        )


1443
1444
1445
1446
# NOTE(woosuk): Currently, the below code is not used and we always use argmax
# to sample the draft tokens. We will use this after we find a way to manage
# the draft prob tensor.
# Refer to https://github.com/vllm-project/vllm/pull/16899 for the details.
1447
1448
1449
1450
1451
1452
1453
1454
1455
1456
1457
1458
1459
# FIXME(woosuk): The logic here is duplicated with the main sampling code.
# We should refactor this to reuse the same sampling implementation.
def compute_probs_and_sample_next_token(
    logits: torch.Tensor,
    sampling_metadata: SamplingMetadata,
) -> tuple[torch.Tensor, torch.Tensor]:
    if sampling_metadata.all_greedy:
        # For greedy requests, draft_probs is not used in rejection sampling.
        # Therefore, we can just return the logits.
        probs = logits
        next_token_ids = logits.argmax(dim=-1)
        return next_token_ids, probs

1460
1461
1462
1463
1464
1465
1466
1467
1468
    assert sampling_metadata.temperature is not None

    # Use epsilon comparison to detect greedy sampling (temperature ~ 0.0)
    # consistent with sampler.py's _SAMPLING_EPS threshold
    temperature = sampling_metadata.temperature
    # Avoid division by zero if there are greedy requests.
    if not sampling_metadata.all_random:
        is_greedy = temperature < _SAMPLING_EPS
        temperature = torch.where(is_greedy, 1.0, temperature)
1469
1470
1471
1472
1473
1474
1475
1476
1477
1478
1479
    logits.div_(temperature.view(-1, 1))
    probs = logits.softmax(dim=-1, dtype=torch.float32)

    # NOTE(woosuk): Currently, we ignore most of the sampling parameters in
    # generating the draft tokens. We only use the temperature. While this
    # could degrade the acceptance rate, it does not affect the distribution
    # of the generated tokens after rejection sampling.

    # TODO(woosuk): Consider seeds.
    q = torch.empty_like(probs)
    q.exponential_()
1480
1481
1482
    # NOTE(woosuk): We shouldn't use `probs.div_(q)` because the draft_probs
    # will be used later for rejection sampling.
    next_token_ids = probs.div(q).argmax(dim=-1).view(-1)
1483
1484
    if not sampling_metadata.all_random:
        greedy_token_ids = probs.argmax(dim=-1)
1485
        next_token_ids = torch.where(is_greedy, greedy_token_ids, next_token_ids)
1486
    return next_token_ids, probs