eagle.py 61.9 KB
Newer Older
1
# SPDX-License-Identifier: Apache-2.0
2
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3

4
5
import ast
from dataclasses import replace
6
from importlib.util import find_spec
7

8
import numpy as np
9

10
11
12
import torch
import torch.nn as nn

13
14
15
16
17
from vllm.config import (
    CUDAGraphMode,
    VllmConfig,
    get_layers_from_vllm_config,
)
18
from vllm.distributed.parallel_state import get_pp_group
19
from vllm.forward_context import set_forward_context
20
from vllm.logger import init_logger
21
from vllm.model_executor.layers.attention_layer_base import AttentionLayerBase
22
from vllm.model_executor.model_loader import get_model
23
from vllm.model_executor.models import supports_multimodal
24
from vllm.model_executor.models.deepseek_v2 import DeepseekV32IndexerCache
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
        self.speculative_config = vllm_config.speculative_config
68
        assert self.speculative_config is not None
69
70
        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
95
            vllm_config.model_config
        )
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
        if self.uses_mrope:
121
122
123
124
125
126
127
128
129
130
            # 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
131
            self.mrope_positions = torch.zeros(
132
                (3, self.max_num_tokens + 1), dtype=torch.int64, device=device
133
            )
134
135
        else:
            # RoPE need (max_num_tokens,)
136
137
138
            self.positions = torch.zeros(
                self.max_num_tokens, dtype=torch.int64, device=device
            )
139
        self.hidden_states = torch.zeros(
140
141
            (self.max_num_tokens, self.hidden_size), dtype=self.dtype, device=device
        )
142

143
144
145
        # 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)
146
147
148
        self.arange = torch.arange(
            max_num_slots_for_arange, device=device, dtype=torch.int32
        )
149

150
        self.inputs_embeds = torch.zeros(
151
            (self.max_num_tokens, self.inputs_embeds_size),
152
            dtype=self.dtype,
153
            device=device,
154
        )
155

156
157
158
159
160
        self.backup_next_token_ids = CpuGpuBuffer(
            max_batch_size,
            dtype=torch.int32,
            pin_memory=is_pin_memory_available(),
            device=device,
161
162
            with_numpy=True,
        )
163

164
165
166
167
        self._slot_mapping_buffer = torch.zeros(
            self.max_num_tokens, dtype=torch.int64, device=device
        )

168
        # Determine allowed attention backends once during initialization.
169
        self.allowed_attn_types: tuple | None = None
170
        if current_platform.is_rocm():
171
172
173
174
175
176
            from vllm.v1.attention.backends.rocm_attn import RocmAttentionMetadata

            rocm_types = [
                TritonAttentionMetadata,
                RocmAttentionMetadata,
            ]
177
178
179
180
            # ROCM_AITER_FA is an optional backend
            if find_spec(
                AttentionBackendEnum.ROCM_AITER_FA.get_path(include_classname=False)
            ):
181
                from vllm.v1.attention.backends.rocm_aiter_fa import (
182
183
184
                    AiterFlashAttentionMetadata,
                )

185
                rocm_types.append(AiterFlashAttentionMetadata)
186
187

            # TRITON_MLA backend support for MLA models (e.g., DeepSeek)
188
189
190
            from vllm.model_executor.layers.attention.mla_attention import (
                MLACommonMetadata,
            )
191
192
193

            rocm_types.append(MLACommonMetadata)

194
195
196
197
198
            # FlexAttention backend support
            from vllm.v1.attention.backends.flex_attention import FlexAttentionMetadata

            rocm_types.append(FlexAttentionMetadata)

199
200
            self.allowed_attn_types = tuple(rocm_types)

201
202
        # Parse the speculative token tree.
        spec_token_tree = self.speculative_config.speculative_token_tree
203
        self.tree_choices: list[tuple[int, ...]] = ast.literal_eval(spec_token_tree)
204
205
206
207
208
209
210
211
        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):
212
213
214
215
216
217
            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]
            )
218
219
        # Precompute draft position offsets in flattened tree.
        self.tree_draft_pos_offsets = torch.arange(
220
            1, len(self.tree_choices) + 1, device=device, dtype=torch.int32
221
        ).repeat(max_batch_size, 1)
222

223
224
225
226
227
228
229
230
231
    def _get_positions(self, num_tokens: int):
        if self.uses_mrope:
            return self.mrope_positions[:, :num_tokens]
        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
        else:
232
233
234
235
236
            # 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]
237
238
            self.positions[:num_tokens] = positions

239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
    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}

257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
    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)

274
275
276
277
    def propose(
        self,
        # [num_tokens]
        target_token_ids: torch.Tensor,
278
        # [num_tokens] or [3, num_tokens] when M-RoPE is enabled
279
280
281
282
283
        target_positions: torch.Tensor,
        # [num_tokens, hidden_size]
        target_hidden_states: torch.Tensor,
        # [batch_size]
        next_token_ids: torch.Tensor,
284
        last_token_indices: torch.Tensor | None,
285
        common_attn_metadata: CommonAttentionMetadata,
286
        sampling_metadata: SamplingMetadata,
287
        mm_embed_inputs: tuple[list[torch.Tensor], torch.Tensor] | None = None,
288
        num_rejected_tokens_gpu: torch.Tensor | None = None,
289
290
291
        slot_mappings: dict[str, torch.Tensor]
        | list[dict[str, torch.Tensor]]
        | None = None,
292
    ) -> torch.Tensor:
293
        batch_size = common_attn_metadata.batch_size()
294

295
296
297
        if self.method == "eagle3":
            assert isinstance(self.model, Eagle3LlamaForCausalLM)
            target_hidden_states = self.model.combine_hidden_states(
298
299
                target_hidden_states
            )
300
            assert target_hidden_states.shape[-1] == self.hidden_size
301
302
303
304
305
306
307
308
309
310
311

        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,
            )
        )
312

313
        assert self.runner is not None
Jiayi Yao's avatar
Jiayi Yao committed
314

315
316
317
318
319
        if self.attn_metadata_builder is None:
            attn_metadata_builder = self._get_attention_metadata_builder()
        else:
            attn_metadata_builder = self.attn_metadata_builder

320
        attn_metadata = attn_metadata_builder.build_for_drafting(
321
322
            common_attn_metadata=common_attn_metadata, draft_index=0
        )
323
324
325
326
327
328
        # 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,
329
330
                )
            )
331
332
        else:
            draft_indexer_metadata = None
333
334
335
336
337
        # 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
338

339
340
341
342
        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
343
        num_tokens_dp_padded, num_tokens_across_dp = self._pad_batch_across_dp(
344
            num_tokens_unpadded=num_tokens, num_tokens_padded=num_tokens
Rémi Delacourt's avatar
Rémi Delacourt committed
345
346
        )

347
348
349
350
        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
351
352
353
        if num_tokens_across_dp is not None:
            num_tokens_across_dp[self.dp_rank] = num_input_tokens

354
355
356
357
        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
358
359
360
361

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

362
            self.inputs_embeds[:num_tokens] = self.model.embed_input_ids(
363
364
365
                self.input_ids[:num_tokens],
                multimodal_embeddings=mm_embeds,
                is_multimodal=is_mm_embed,
366
            )
367

368
            input_ids = None
369
            inputs_embeds = self.inputs_embeds[:num_input_tokens]
370
371
        else:
            input_ids = self.input_ids[:num_input_tokens]
372
            inputs_embeds = None
373

374
375
376
377
378
379
380
381
        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]

382
        with set_forward_context(
383
384
385
            per_layer_attn_metadata,
            self.vllm_config,
            num_tokens=num_input_tokens,
Rémi Delacourt's avatar
Rémi Delacourt committed
386
            num_tokens_across_dp=num_tokens_across_dp,
387
            cudagraph_runtime_mode=cudagraph_runtime_mode,
388
389
390
            slot_mapping=self._get_slot_mapping(
                num_input_tokens, common_attn_metadata.slot_mapping
            ),
391
        ):
392
393
            ret_hidden_states = self.model(**model_kwargs)
            if not self.model_returns_tuple():
Jiayi Yao's avatar
Jiayi Yao committed
394
                last_hidden_states = ret_hidden_states
395
                hidden_states = last_hidden_states
Jiayi Yao's avatar
Jiayi Yao committed
396
397
            else:
                last_hidden_states, hidden_states = ret_hidden_states
398

399
        sample_hidden_states = last_hidden_states[last_token_indices]
400
        logits = self.model.compute_logits(sample_hidden_states)
401
402
403
404
405

        # 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)
406

407
        if self.uses_mrope:
408
            positions = self.mrope_positions[:, last_token_indices]
409
        else:
410
            positions = self.positions[last_token_indices]
411
412
413
414
415
416
        if self.method in (
            "deepseek_mtp",
            "ernie_mtp",
            "longcat_flash_mtp",
            "pangu_ultra_moe_mtp",
        ):
zhuwenwen's avatar
zhuwenwen committed
417
418
419
            hidden_states = self.hidden_states[last_token_indices]
        else:
            hidden_states = hidden_states[last_token_indices]
420
421
422

        if isinstance(attn_metadata, TreeAttentionMetadata):
            # Draft using tree attention.
423
424
425
426
427
428
            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,
429
                slot_mappings=slot_mappings,
430
431
432
433
            )
            # [batch_size, num_tree_tokens]
            return torch.cat(draft_token_ids_list, dim=1)

434
        draft_token_ids = logits.argmax(dim=-1)
435

436
437
438
        if self.allowed_attn_types is not None and not isinstance(
            attn_metadata, self.allowed_attn_types
        ):
439
440
441
442
            raise ValueError(
                f"Unsupported attention metadata type for speculative "
                "decoding with num_speculative_tokens > 1: "
                f"{type(attn_metadata)}. Supported types are: "
443
444
                f"{self.allowed_attn_types}"
            )
445

446
447
        # Generate the remaining draft tokens.
        draft_token_ids_list = [draft_token_ids]
448

Rémi Delacourt's avatar
Rémi Delacourt committed
449
        batch_size_dp_padded, batch_size_across_dp = self._pad_batch_across_dp(
450
            num_tokens_unpadded=batch_size, num_tokens_padded=batch_size
Rémi Delacourt's avatar
Rémi Delacourt committed
451
452
        )

453
454
455
456
        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
457
458
        if batch_size_across_dp is not None:
            batch_size_across_dp[self.dp_rank] = input_batch_size
459

460
461
        common_attn_metadata.num_actual_tokens = batch_size
        common_attn_metadata.max_query_len = 1
462
        common_attn_metadata.query_start_loc = self.arange[: batch_size + 1]
463
        common_attn_metadata.query_start_loc_cpu = torch.from_numpy(
464
465
            self.token_arange_np[: batch_size + 1]
        ).clone()
466
467
468
469
470
471
472
473
474
475
476

        # 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

477
        for token_index in range(self.num_speculative_tokens - 1):
478
            # Update the inputs.
479
480
481
            # cast to int32 is crucial when eagle model is compiled.
            # tensor.argmax() returns int64 by default.
            input_ids = draft_token_ids_list[-1].int()
482
483
484
485
486
487
488
489
490
491
492
493
494
            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.
495
496
497
498
499
                clamped_positions = torch.where(
                    exceeds_max_model_len.unsqueeze(0),
                    torch.zeros_like(positions),
                    positions,
                )
500
            else:
501
502
                positions += 1
                exceeds_max_model_len = positions >= self.max_model_len
503
                clamped_positions = torch.where(exceeds_max_model_len, 0, positions)
504
505
506
            # 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.
507
            # Increment the sequence lengths.
508
            common_attn_metadata.seq_lens += 1
509
510
            # For the requests that exceed the max model length, we set the
            # sequence length to 1 to minimize their overheads in attention.
511
            common_attn_metadata.seq_lens.masked_fill_(exceeds_max_model_len, 1)
512
513
514
515
516
517
            # 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
            )
518

519
520
521
522
523
524
            # 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
525

526
            # Compute the slot mapping.
527
            block_size = attn_metadata_builder.kv_cache_spec.block_size
528
529
            if self.uses_mrope:
                # all dimensions of positions are the same
530
                block_numbers = clamped_positions[0] // block_size
531
            else:
532
                block_numbers = clamped_positions // block_size
533
            block_ids = common_attn_metadata.block_table_tensor.gather(
534
535
                dim=1, index=block_numbers.view(-1, 1)
            )
536
            block_ids = block_ids.view(-1)
537
538
            if self.uses_mrope:
                common_attn_metadata.slot_mapping = (
539
                    block_ids * block_size + clamped_positions[0] % block_size
540
                )
541
542
            else:
                common_attn_metadata.slot_mapping = (
543
                    block_ids * block_size + clamped_positions % block_size
544
                )
545
546
547
            # Mask out the slot mappings that exceed the max model length.
            # Otherwise, the KV cache will be inadvertently updated with the
            # padding tokens.
548
            common_attn_metadata.slot_mapping.masked_fill_(
549
550
                exceeds_max_model_len, PADDING_SLOT_ID
            )
551
552

            # Rebuild attention metadata
553
            attn_metadata = attn_metadata_builder.build_for_drafting(  # type: ignore
554
555
                common_attn_metadata=common_attn_metadata, draft_index=token_index + 1
            )
556
557
            for layer_name in self.attn_layer_names:
                per_layer_attn_metadata[layer_name] = attn_metadata
558

559
560
            # copy inputs to buffer for cudagraph
            self.input_ids[:batch_size] = input_ids
561
            self._set_positions(batch_size, clamped_positions)
562
            self.hidden_states[:batch_size] = hidden_states
563
            if self.supports_mm_inputs:
564
                self.inputs_embeds[:batch_size] = self.model.embed_input_ids(input_ids)
565

566
                input_ids = None
567
                inputs_embeds = self.inputs_embeds[:input_batch_size]
568
569
            else:
                input_ids = self.input_ids[:input_batch_size]
570
                inputs_embeds = None
571

572
            # Run the model.
573
574
575
576
577
578
579
580
            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]

581
            with set_forward_context(
582
583
584
                per_layer_attn_metadata,
                self.vllm_config,
                num_tokens=input_batch_size,
Rémi Delacourt's avatar
Rémi Delacourt committed
585
                num_tokens_across_dp=batch_size_across_dp,
586
                cudagraph_runtime_mode=cudagraph_runtime_mode,
587
588
589
                slot_mapping=self._get_slot_mapping(
                    input_batch_size, common_attn_metadata.slot_mapping
                ),
590
            ):
591
592
                ret_hidden_states = self.model(**model_kwargs)
                if not self.model_returns_tuple():
593
                    last_hidden_states = ret_hidden_states
594
                    hidden_states = ret_hidden_states
595
596
                else:
                    last_hidden_states, hidden_states = ret_hidden_states
597

598
            hidden_states = hidden_states[:batch_size]
599
            logits = self.model.compute_logits(last_hidden_states[:batch_size])
600
            draft_token_ids = logits.argmax(dim=-1)
601
602
603
604
            draft_token_ids_list.append(draft_token_ids)

        # [batch_size, num_speculative_tokens]
        draft_token_ids = torch.stack(draft_token_ids_list, dim=1)
605
        return draft_token_ids
606

607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
    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
        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")

635
    def prepare_next_token_ids_cpu(
636
        self,
637
        sampled_token_ids: list[list[int]],
638
639
640
641
        requests: dict[str, CachedRequestState],
        gpu_input_batch: InputBatch,
        num_scheduled_tokens: dict[str, int],
    ) -> torch.Tensor:
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
        """
        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):
            if token_ids:
                # 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]
660
                seq_len = req_state.num_computed_tokens + num_scheduled_tokens[req_id]
661
662
                next_token_id = req_state.get_token_id(seq_len)
            next_token_ids.append(next_token_id)
663
        next_token_ids = torch.tensor(
664
665
            next_token_ids, dtype=torch.int32, device=self.input_ids.device
        )
666
667
        return next_token_ids

668
669
670
671
672
673
    def prepare_next_token_ids_padded(
        self,
        common_attn_metadata: CommonAttentionMetadata,
        sampled_token_ids: torch.Tensor,
        requests: dict[str, CachedRequestState],
        gpu_input_batch: InputBatch,
674
        discard_request_mask: torch.Tensor,
675
    ) -> tuple[torch.Tensor, torch.Tensor]:
676
677
678
679
        """
        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
680
681
        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`.
682
683
684
        """
        # Precompute get_token_id for when there is no valid next token
        num_reqs = gpu_input_batch.num_reqs
685
686
687
688
689
690
        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)
691
692
            ],
            dtype=np.int32,
693
        )
694
        self.backup_next_token_ids.copy_to_gpu(num_reqs)
695
        backup_tokens_gpu = self.backup_next_token_ids.gpu
696

697
698
        batch_size, num_tokens = sampled_token_ids.shape
        device = sampled_token_ids.device
699

700
701
        assert discard_request_mask.dtype == torch.bool
        assert backup_tokens_gpu.dtype == torch.int32
702

703
704
        next_token_ids = torch.empty(batch_size, dtype=torch.int32, device=device)
        valid_sampled_tokens_count = next_token_ids.new_empty(batch_size)
705

706
707
        # Kernel grid: one program per request (row)
        grid = (batch_size,)
708

709
710
711
712
713
714
715
716
717
718
719
720
721
        # 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,
722
        )
723
724
725

        return next_token_ids, valid_sampled_tokens_count

726
727
728
729
730
    def prepare_inputs_padded(
        self,
        common_attn_metadata: CommonAttentionMetadata,
        spec_decode_metadata: SpecDecodeMetadata,
        valid_sampled_tokens_count: torch.Tensor,
731
    ) -> tuple[CommonAttentionMetadata, torch.Tensor, torch.Tensor]:
732
733
734
735
736
737
        """
        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`.
738
        No blocking CPU operations should be introduced in this function.
739
        """
740
741
        num_reqs = common_attn_metadata.num_reqs
        device = valid_sampled_tokens_count.device
742

743
744
        token_indices_to_sample = torch.empty(
            (num_reqs,), dtype=torch.int32, device=device
745
        )
746
747
748
        num_rejected_tokens_gpu = torch.empty(
            (num_reqs,), dtype=torch.int32, device=device
        )
749

750
751
752
753
754
755
        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,
756
            num_rejected_tokens_gpu,
757
            num_reqs,
758
        )
759
760

        query_start_loc_cpu = common_attn_metadata.query_start_loc_cpu
761
        new_query_len_per_req = query_start_loc_cpu[1:] - query_start_loc_cpu[:-1]
762
763
764
765
766
767
768

        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,
769
770
            _seq_lens_cpu=common_attn_metadata._seq_lens_cpu,
            _num_computed_tokens_cpu=common_attn_metadata._num_computed_tokens_cpu,
771
772
773
774
775
            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,
776
            slot_mapping=common_attn_metadata.slot_mapping[:total_num_tokens],
777
            causal=True,
778
            dcp_local_seq_lens=common_attn_metadata.dcp_local_seq_lens,
779
780
        )

781
782
783
784
785
        return (
            spec_common_attn_metadata,
            token_indices_to_sample,
            num_rejected_tokens_gpu,
        )
786

787
788
789
790
791
792
793
794
795
796
    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,
797
798
799
        slot_mappings: dict[str, torch.Tensor]
        | list[dict[str, torch.Tensor]]
        | None = None,
800
    ) -> list[torch.Tensor]:
801
802
803
804
        tree_attn_metadata_builder = self.runner.attn_groups[0][
            0
        ].get_metadata_builder()
        assert isinstance(tree_attn_metadata_builder, TreeAttentionMetadataBuilder)
805

806
        total_num_drafts = self.cu_drafts_per_level[0]
807
808
        level_num_drafts = total_num_drafts
        # Sample a draft token for each child at the tree root level.
809
        num_children = self.child_drafts_per_level[0]
810
811
812
        if num_children == 1:
            draft_token_ids = logits.argmax(dim=-1).view(batch_size, -1)
        else:
813
814
815
            draft_token_ids = torch.topk(logits, num_children, dim=-1).indices.view(
                batch_size, -1
            )
816
817
818
819
        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.
820
821
822
823
824
825
826
827
828
        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
        )
829
830
        # Precompute the draft token positions.
        flattened_draft_positions = (
831
832
            positions.view(batch_size, -1) + self.tree_draft_pos_offsets[:batch_size, :]
        )
833
        tree_depth = len(self.cu_drafts_per_level)
834
        for level in range(tree_depth - 1):
835
836
            # Get draft positions for RoPE.
            draft_positions = positions + (level + 1)
837
            exceeds_max_model_len = (positions + total_num_drafts) >= self.max_model_len
838
839
            # Mask out the position ids that exceed the max model length.
            # Otherwise, we may get out-of-range error in RoPE.
840
            draft_positions = torch.where(
841
842
843
                exceeds_max_model_len,
                0,
                draft_positions,
844
845
            ).view(batch_size, -1)

846
847
            if level_num_drafts > 1:
                # Repeat the positions for each draft at this level.
848
                draft_positions = draft_positions.repeat_interleave(
849
850
                    level_num_drafts, dim=1
                )
851
852
853
854

            if num_children > 1:
                # Repeat draft hidden states for each child.
                draft_hidden_states = draft_hidden_states.repeat_interleave(
855
856
                    num_children, dim=1
                )
857
858

            # Concatenate the draft tokens, positions, and hidden states.
859
860
            tree_input_ids = torch.cat([tree_input_ids, draft_token_ids], dim=1)
            tree_positions = torch.cat([tree_positions, draft_positions], dim=1)
861
            tree_hidden_states = torch.cat(
862
863
                [tree_hidden_states, draft_hidden_states], dim=1
            )
864
865
866

            # Build new attention metadata for the next level of drafts.
            # This is necessary to support tree attention.
867
            query_len = total_num_drafts
868
869
            common_attn_metadata = replace(
                common_attn_metadata,
870
                query_start_loc=query_len * self.arange[: batch_size + 1],
871
872
873
874
875
                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(
876
                common_attn_metadata=common_attn_metadata, draft_index=level + 1
877
878
879
880
881
882
883
884
            )

            # 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.
885
886
887
            attn_metadata.max_seq_len = min(
                attn_metadata.max_seq_len, self.max_model_len
            )
888
889
890
891
892
            # 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.
893
            block_size = tree_attn_metadata_builder.kv_cache_spec.block_size
894
            query_positions = flattened_draft_positions[:, level : level + query_len]
895
            block_numbers = query_positions // block_size
896
            block_ids = attn_metadata.block_table.gather(dim=1, index=block_numbers)
897
            slot_mapping = block_ids * block_size + query_positions % block_size
898
899
900
901
902
903
904
905
906
907
908
            # 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)
909
            self.hidden_states[:num_tokens] = tree_hidden_states.view(num_tokens, -1)
910

911
912
913
914
            cudagraph_runtime_mode, batch_desc = self.cudagraph_dispatcher.dispatch(
                num_tokens
            )
            num_input_tokens = batch_desc.num_tokens
915
            # Run the model.
916
            with set_forward_context(
917
918
919
920
                per_layer_attn_metadata,
                self.vllm_config,
                num_tokens=num_input_tokens,
                cudagraph_runtime_mode=cudagraph_runtime_mode,
921
922
923
                slot_mapping=self._get_slot_mapping(
                    num_input_tokens, attn_metadata.slot_mapping
                ),
924
            ):
925
926
927
928
929
930
931
932
933
                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(
934
935
                batch_size, query_len, -1
            )[:, -level_num_drafts:]
936
            draft_last_hidden_states = last_hidden_states[:num_tokens].view(
937
938
                batch_size, query_len, -1
            )[:, -level_num_drafts:]
939
940
941

            # Get the output logits for the draft tokens.
            logits = self.model.compute_logits(
942
943
                draft_last_hidden_states.reshape(batch_size * level_num_drafts, -1)
            )
944
945
946
947
948
949

            # 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:
950
951
952
                draft_token_ids = torch.topk(logits, num_children, dim=-1).indices.view(
                    batch_size, -1
                )
953
954
955
            draft_token_ids_list.append(draft_token_ids)

            # Update the # drafts counters for the next tree level.
956
            level_num_drafts = self.cu_drafts_per_level[level + 1] - total_num_drafts
957
958
959
            total_num_drafts = self.cu_drafts_per_level[level + 1]
        return draft_token_ids_list

960
    def prepare_inputs(
961
962
        self,
        common_attn_metadata: CommonAttentionMetadata,
963
964
        sampled_token_ids: list[list[int]],
        num_draft_tokens: list[int],
965
966
    ) -> tuple[CommonAttentionMetadata, torch.Tensor]:
        """
967
        This function is used to prepare the inputs for speculative decoding.
968
969
970
971
972
973
        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}:
974
        #       [0, q1, q1 + q2, q1 + q2 + q3]
975
976
977
978
979
980
        #  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}:
981
        #       [0, q1 - n1, q1 + q2 - n1 - n2, q1 + q2 + q3 - n1 - n2 - n3]
982
        #  common_attn_metadata.seq_lens{_cpu}:
983
        #       [s1 - n1 + 1, s2 - n2 + 1, s3 - n3 + 1]
984
        #  token_indices: [0, 1, ..., q1 - n1 - 1,
985
986
        #                 q1, q1 + 1, ..., q1 + q2 - n2 - 1,
        #                 q1 + q2, q1 + q2 + 1, ..., q1 + q2 + q3 - n3 - 1]
987

988
989
990
991
        num_rejected_tokens = [
            n + 1 - len(sampled_token_ids[i]) if n > 0 else 0
            for i, n in enumerate(num_draft_tokens)
        ]
992
        num_rejected_tokens = torch.tensor(num_rejected_tokens, dtype=torch.int32)
993

994
995
        device = common_attn_metadata.query_start_loc.device
        query_start_loc_cpu = common_attn_metadata.query_start_loc_cpu
996
        new_seq_lens_cpu = common_attn_metadata.seq_lens_cpu - num_rejected_tokens
997
998

        # [0, q1, q1 + q2, q1 + q2 + q3] -> [q1, q2, q3]
999
        new_query_len_per_req = query_start_loc_cpu[1:] - query_start_loc_cpu[:-1]
1000
1001
1002
1003
1004
1005
1006
1007
        # [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,
1008
            dtype=torch.int32,
1009
1010
            pin_memory=is_pin_memory_available(),
        )
1011
1012
1013
1014
1015
1016
1017
1018
1019
        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__
1020
1021
1022
        new_query_start_locs_expanded = np.repeat(
            new_query_start_loc_np[:-1], new_num_tokens_per_req_np
        )
1023
1024
1025
        # [0, 1, 2, 3, 4, 5, 6, 7, 8] ->
        # [0, 1, 0, 1, 2, 3, 0, 1, 2]
        #  _r1_  ____r2____  ___r3__
1026
1027
1028
        token_offests = (
            self.token_arange_np[:total_num_tokens] - new_query_start_locs_expanded
        )
1029
1030
1031
1032
1033
1034

        # 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(
1035
1036
            query_start_loc_cpu[:-1].numpy(), new_num_tokens_per_req_np
        )
1037
        # Final token indices are:
1038
1039
1040
        # [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
1041
        token_indices_np = token_offests + old_query_start_locs_expanded
1042
        token_indices = torch.from_numpy(token_indices_np).to(device, non_blocking=True)
1043
1044

        spec_common_attn_metadata = CommonAttentionMetadata(
1045
            query_start_loc=new_query_start_loc_cpu.to(device, non_blocking=True),
1046
1047
            seq_lens=new_seq_lens_cpu.to(device, non_blocking=True),
            query_start_loc_cpu=new_query_start_loc_cpu,
1048
1049
            _seq_lens_cpu=new_seq_lens_cpu,
            _num_computed_tokens_cpu=common_attn_metadata._num_computed_tokens_cpu,
1050
1051
1052
            num_reqs=common_attn_metadata.num_reqs,
            num_actual_tokens=total_num_tokens,
            max_query_len=new_query_len_per_req.max().item(),
1053
            max_seq_len=new_seq_lens_cpu.max().item(),
1054
1055
            block_table_tensor=common_attn_metadata.block_table_tensor,
            slot_mapping=common_attn_metadata.slot_mapping[token_indices],
1056
            causal=True,
1057
            dcp_local_seq_lens=common_attn_metadata.dcp_local_seq_lens,
1058
        )
1059
1060

        return spec_common_attn_metadata, token_indices
1061

1062
    def get_model_name(self, model: nn.Module) -> str:
1063
        if hasattr(model, "module"):  # multi-GPU
1064
1065
1066
            model = model.module
        return model.__class__.__name__

1067
    def load_model(self, target_model: nn.Module) -> None:
1068
        draft_model_config = self.vllm_config.speculative_config.draft_model_config
1069
        target_attn_layer_names = set(
1070
            get_layers_from_vllm_config(self.vllm_config, AttentionLayerBase).keys()
1071
        )
1072
1073
        # FIXME: support hybrid kv for draft model
        target_indexer_layer_names = set(
1074
1075
1076
1077
            get_layers_from_vllm_config(
                self.vllm_config, DeepseekV32IndexerCache
            ).keys()
        )
1078

1079
        from vllm.compilation.backends import set_model_tag
1080

1081
        with set_model_tag("eagle_head"):
1082
1083
1084
            self.model = get_model(
                vllm_config=self.vllm_config, model_config=draft_model_config
            )
1085

1086
        draft_attn_layer_names = (
1087
            get_layers_from_vllm_config(self.vllm_config, AttentionLayerBase).keys()
1088
1089
1090
1091
1092
1093
            - 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
1094
        self.attn_layer_names = list(draft_attn_layer_names - draft_indexer_layer_names)
1095
1096
1097
1098
1099
        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 = (
1100
1101
1102
                indexer_layers[first_layer]
                .get_attn_backend()
                .get_builder_cls()(
1103
                    indexer_layers[first_layer].get_kv_cache_spec(self.vllm_config),
1104
1105
1106
                    self.indexer_layer_names,
                    self.vllm_config,
                    self.device,
1107
1108
                )
            )
1109
1110
        else:
            self.draft_indexer_metadata_builder = None
1111

1112
        if self.supports_mm_inputs:
1113
1114
1115
            # Even if the target model is multimodal, we can also use
            # text-only draft models
            try:
1116
                dummy_input_ids = torch.tensor([[1]], device=self.input_ids.device)
1117
                self.model.embed_input_ids(dummy_input_ids, multimodal_embeddings=None)
1118
1119
1120
            except (NotImplementedError, AttributeError, TypeError):
                logger.warning(
                    "Draft model does not support multimodal inputs, "
1121
1122
                    "falling back to text-only mode"
                )
1123
                self.supports_mm_inputs = False
1124

1125
1126
        if supports_multimodal(target_model):
            # handle multimodality
1127
1128
1129
            if self.get_model_name(target_model) in [
                "Qwen2_5_VLForConditionalGeneration",
                "Qwen3VLForConditionalGeneration",
1130
                "Qwen3VLMoeForConditionalGeneration",
1131
                "GlmOcrForConditionalGeneration",
1132
            ]:
1133
                self.model.config.image_token_index = target_model.config.image_token_id
1134
1135
1136
1137
            elif self.get_model_name(target_model) == "PixtralForConditionalGeneration":
                self.model.config.image_token_index = (
                    target_model.config.vision_config.image_token_id
                )
1138
1139
            else:
                self.model.config.image_token_index = (
1140
1141
                    target_model.config.image_token_index
                )
1142
1143
1144
            target_language_model = target_model.get_language_model()
        else:
            target_language_model = target_model
1145

1146
        # share embed_tokens with the target model if needed
1147
        if get_pp_group().world_size == 1:
1148
            if hasattr(target_language_model.model, "embed_tokens"):
1149
                target_embed_tokens = target_language_model.model.embed_tokens
1150
            elif hasattr(target_language_model.model, "embedding"):
1151
1152
1153
                target_embed_tokens = target_language_model.model.embedding
            else:
                raise AttributeError(
1154
1155
                    "Target model does not have 'embed_tokens' or 'embedding' attribute"
                )
1156

1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
            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)
1170
1171
1172
                    # TODO: Offload to CPU for comparison to avoid extra GPU memory
                    # usage in CI testing environments with limited GPU memory
                    and torch.equal(
1173
1174
                        target_embed_tokens.weight.cpu(),
                        self.model.model.embed_tokens.weight.cpu(),
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
                    )
                ):
                    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."
                    )
1188
            else:
1189
1190
                # MTP model
                share_embeddings = True
1191
                logger.info(
1192
1193
                    "Detected MTP model. "
                    "Sharing target model embedding weights with the draft model."
1194
                )
1195
1196
1197
1198
1199

            if share_embeddings:
                if hasattr(self.model.model, "embed_tokens"):
                    del self.model.model.embed_tokens
                self.model.model.embed_tokens = target_embed_tokens
1200
        else:
1201
            logger.info(
1202
                "The draft model's vocab embedding will be loaded separately"
1203
1204
                " from the target model."
            )
1205
1206

        # share lm_head with the target model if needed
1207
1208
1209
1210
1211
        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
1212
                logger.info(
1213
1214
1215
1216
1217
1218
1219
                    "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)
1220
1221
                # TODO: Offload to CPU for comparison to avoid extra GPU memory
                # usage in CI testing environments with limited GPU memory
1222
                and torch.equal(
1223
1224
                    target_language_model.lm_head.weight.cpu(),
                    self.model.lm_head.weight.cpu(),
1225
                )
1226
            ):
1227
                share_lm_head = True
1228
                logger.info(
1229
1230
                    "Detected EAGLE model with lm_head identical to the target model. "
                    "Sharing target model lm_head weights with the draft model."
1231
                )
1232
1233
            else:
                logger.info(
1234
1235
                    "Detected EAGLE model with distinct lm_head weights. "
                    "Keeping separate lm_head weights from the target model."
1236
                )
1237
1238
1239
1240
1241
1242
1243
1244
1245
1246
1247
1248
        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
1249

1250
1251
1252
1253
    @torch.inference_mode()
    def dummy_run(
        self,
        num_tokens: int,
1254
1255
        use_cudagraphs: bool = True,
        is_graph_capturing: bool = False,
1256
        slot_mappings: dict[str, torch.Tensor] | None = None,
1257
    ) -> None:
Rémi Delacourt's avatar
Rémi Delacourt committed
1258
1259
1260
1261
        # 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
1262
        ):
Rémi Delacourt's avatar
Rémi Delacourt committed
1263
1264
            if fwd_idx <= 1:
                num_tokens_dp_padded, num_tokens_across_dp = self._pad_batch_across_dp(
1265
                    num_tokens_unpadded=num_tokens, num_tokens_padded=num_tokens
Rémi Delacourt's avatar
Rémi Delacourt committed
1266
                )
1267
1268
1269
                if use_cudagraphs:
                    cudagraph_runtime_mode, batch_desc = (
                        self.cudagraph_dispatcher.dispatch(num_tokens_dp_padded)
Rémi Delacourt's avatar
Rémi Delacourt committed
1270
                    )
1271
                    num_input_tokens = batch_desc.num_tokens
Rémi Delacourt's avatar
Rémi Delacourt committed
1272
                else:
1273
                    cudagraph_runtime_mode = CUDAGraphMode.NONE
Rémi Delacourt's avatar
Rémi Delacourt committed
1274
1275
1276
                    num_input_tokens = num_tokens_dp_padded
                if num_tokens_across_dp is not None:
                    num_tokens_across_dp[self.dp_rank] = num_input_tokens
1277

1278
1279
1280
1281
1282
1283
1284
1285
1286
1287
            # 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
1288
1289
1290
1291
1292
            with set_forward_context(
                None,
                self.vllm_config,
                num_tokens=num_input_tokens,
                num_tokens_across_dp=num_tokens_across_dp,
1293
                cudagraph_runtime_mode=cudagraph_runtime_mode,
1294
                slot_mapping=slot_mapping_dict,
Rémi Delacourt's avatar
Rémi Delacourt committed
1295
1296
1297
1298
1299
1300
1301
            ):
                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
1302

1303
                kwargs = dict(
Rémi Delacourt's avatar
Rémi Delacourt committed
1304
1305
1306
1307
                    input_ids=input_ids,
                    positions=self._get_positions(num_input_tokens),
                    inputs_embeds=inputs_embeds,
                )
1308
1309
1310
                if self.pass_hidden_states_to_model:
                    kwargs["hidden_states"] = self.hidden_states[:num_input_tokens]
                self.model(**kwargs)
1311

1312
    def _get_attention_metadata_builder(self) -> AttentionMetadataBuilder:
1313
        """Find and return the attention metadata builders for EAGLE layers.
1314

1315
1316
        Returns:
            The metadata builders for EAGLE layers.
1317

1318
1319
1320
1321
1322
1323
1324
1325
1326
1327
1328
1329
1330
1331
1332
        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, (
1333
1334
            "Failed to find attention metadata builder for EAGLE layers."
        )
1335
1336
        return builder

1337
1338
1339
1340
1341
1342
1343
1344
1345
1346
1347
1348
1349
1350
1351
1352
    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

1353
    def validate_same_kv_cache_group(self, kv_cache_config: KVCacheConfig) -> None:
1354
        """
1355
1356
        Validate that all drafting layers belong to the same KVCacheGroup.
        Need this assumption to ensure all drafting layers can use the
1357
1358
1359
1360
1361
1362
1363
        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
1364
1365
1366
1367
1368
1369
1370
1371
1372
1373
        assert (
            len(
                set(
                    [
                        kv_cache_groups[layer_name]
                        for layer_name in self.attn_layer_names
                    ]
                )
            )
            == 1
1374
        ), "All drafting layers should belong to the same kv cache group"
1375

Rémi Delacourt's avatar
Rémi Delacourt committed
1376
1377
1378
1379
1380
1381
    def _pad_batch_across_dp(
        self,
        num_tokens_unpadded: int,
        num_tokens_padded: int,
    ) -> tuple[int, torch.Tensor]:
        # TODO(Flechman): support DBO ubatching
1382
        should_ubatch, num_toks_across_dp, _ = coordinate_batch_across_dp(
Rémi Delacourt's avatar
Rémi Delacourt committed
1383
1384
1385
            num_tokens_unpadded=num_tokens_unpadded,
            parallel_config=self.vllm_config.parallel_config,
            allow_microbatching=False,
1386
1387
            allow_dp_padding=self.cudagraph_dispatcher.cudagraph_mode
            != CUDAGraphMode.NONE,
Rémi Delacourt's avatar
Rémi Delacourt committed
1388
1389
1390
1391
            num_tokens_padded=num_tokens_padded,
            uniform_decode=None,
            num_scheduled_tokens_per_request=None,
        )
1392
        assert not should_ubatch, "DBO ubatching not implemented for EAGLE"
Rémi Delacourt's avatar
Rémi Delacourt committed
1393
1394
1395
1396
1397

        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
1398
1399


1400
1401
1402
1403
1404
1405
1406
1407
1408
1409
1410
1411
1412
1413
1414
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,
        )


1415
1416
1417
1418
# 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.
1419
1420
1421
1422
1423
1424
1425
1426
1427
1428
1429
1430
1431
# 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

1432
1433
1434
1435
1436
1437
1438
1439
1440
    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)
1441
1442
1443
1444
1445
1446
1447
1448
1449
1450
1451
    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_()
1452
1453
1454
    # 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)
1455
1456
    if not sampling_metadata.all_random:
        greedy_token_ids = probs.argmax(dim=-1)
1457
        next_token_ids = torch.where(is_greedy, greedy_token_ids, next_token_ids)
1458
    return next_token_ids, probs