# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
import functools
import gc
import itertools
import threading
import time
from collections import defaultdict
from collections.abc import Callable, Iterable, Iterator, Sequence
from contextlib import contextmanager
from copy import copy, deepcopy
from dataclasses import dataclass, replace
from functools import reduce
from typing import TYPE_CHECKING, Any, NamedTuple, TypeAlias, cast
import numpy as np
import torch
import torch.distributed
import torch.nn as nn
from tqdm import tqdm
import vllm.envs as envs
from vllm.compilation.breakable_cudagraph import (
BreakableCUDAGraphWrapper,
is_breakable_cudagraph_enabled,
)
from vllm.compilation.counter import compilation_counter
from vllm.compilation.cuda_graph import CUDAGraphStat, CUDAGraphWrapper
from vllm.compilation.monitor import set_cudagraph_capturing_enabled
from vllm.config import (
CompilationMode,
CUDAGraphMode,
VllmConfig,
get_layers_from_vllm_config,
set_current_vllm_config,
update_config,
)
from vllm.config.cache import CacheConfig
from vllm.distributed.ec_transfer import get_ec_transfer, has_ec_transfer
from vllm.distributed.eplb.eplb_state import EplbState
from vllm.distributed.kv_transfer import get_kv_transfer_group, has_kv_transfer_group
from vllm.distributed.kv_transfer.kv_connector.utils import copy_kv_blocks
from vllm.distributed.parallel_state import (
GraphCaptureContext,
get_dcp_group,
get_pp_group,
get_tp_group,
graph_capture,
is_global_first_rank,
prepare_communication_buffer_for_model,
)
from vllm.forward_context import (
BatchDescriptor,
set_forward_context,
)
from vllm.logger import init_logger
from vllm.lora.layers import LoRAMapping, LoRAMappingType
from vllm.model_executor.layers.attention import Attention, MLAAttention
from vllm.model_executor.layers.attention_layer_base import AttentionLayerBase
from vllm.model_executor.layers.fused_moe.all2all_utils import get_ep_all2all_manager
from vllm.model_executor.layers.fused_moe.routed_experts_capturer import (
RoutedExpertsCapturer,
)
from vllm.model_executor.layers.mamba.ops.ssu_dispatch import (
initialize_mamba_ssu_backend,
)
from vllm.model_executor.layers.rotary_embedding import (
MRotaryEmbedding,
XDRotaryEmbedding,
)
from vllm.model_executor.model_loader import get_model_loader
from vllm.model_executor.model_loader.reload import (
finalize_layerwise_reload,
initialize_layerwise_reload,
)
from vllm.model_executor.models.interfaces import (
MixtureOfExperts,
MultiModalEmbeddings,
SupportsMRoPE,
SupportsMultiModal,
SupportsXDRoPE,
is_mixture_of_experts,
supports_eagle3,
supports_mrope,
supports_multimodal_pruning,
supports_realtime,
supports_transcription,
supports_xdrope,
)
from vllm.model_executor.models.interfaces_base import (
VllmModelForPooling,
is_pooling_model,
is_text_generation_model,
)
from vllm.model_executor.offloader import (
create_offloader,
get_offloader,
set_offloader,
)
from vllm.multimodal import MULTIMODAL_REGISTRY
from vllm.multimodal.encoder_budget import MultiModalBudget
from vllm.multimodal.inputs import (
BatchedTensorInputs,
MultiModalKwargsItem,
PlaceholderRange,
)
from vllm.multimodal.utils import get_mm_features_in_window, group_and_batch_mm_kwargs
from vllm.platforms import current_platform
from vllm.pooling_params import PoolingParams
from vllm.sampling_params import SamplingType
from vllm.sequence import IntermediateTensors
from vllm.tasks import GenerationTask, PoolingTask, SupportedTask
from vllm.tracing import instrument
from vllm.utils import length_from_prompt_token_ids_or_embeds
from vllm.utils.math_utils import cdiv, round_up
from vllm.utils.mem_utils import DeviceMemoryProfiler, format_gib
from vllm.utils.nvtx_pytorch_hooks import PytHooks
from vllm.utils.platform_utils import num_compute_units
from vllm.utils.torch_utils import (
PIN_MEMORY,
async_tensor_h2d,
get_dtype_size,
is_quantized_kv_cache,
kv_cache_dtype_str_to_dtype,
)
from vllm.v1.attention.backend import (
AttentionBackend,
AttentionCGSupport,
AttentionMetadata,
AttentionMetadataBuilder,
AttentionType,
CommonAttentionMetadata,
)
from vllm.v1.attention.backends.gdn_attn import GDNAttentionMetadataBuilder
from vllm.v1.attention.backends.linear_attn import (
BailingLinearAttentionMetadataBuilder,
)
from vllm.v1.attention.backends.mamba2_attn import Mamba2AttentionMetadataBuilder
from vllm.v1.attention.backends.utils import (
NULL_BLOCK_ID,
create_fast_prefill_custom_backend,
get_dcp_local_seq_lens,
reorder_batch_to_split_decodes_and_prefills,
)
from vllm.v1.core.sched.output import NewRequestData
from vllm.v1.cudagraph_dispatcher import CudagraphDispatcher
from vllm.v1.kv_cache_interface import (
AttentionSpec,
ChunkedLocalAttentionSpec,
CrossAttentionSpec,
EncoderOnlyAttentionSpec,
FullAttentionSpec,
KVCacheConfig,
KVCacheGroupSpec,
KVCacheSpec,
KVCacheSpecKind,
KVQuantMode,
MambaSpec,
SlidingWindowSpec,
UniformTypeKVCacheSpecs,
get_kv_cache_spec_kind,
)
from vllm.v1.kv_cache_spec_registry import KVCacheSpecRegistry
from vllm.v1.outputs import (
EMPTY_MODEL_RUNNER_OUTPUT,
AsyncModelRunnerOutput,
DraftTokenIds,
ECConnectorOutput,
KVConnectorOutput,
LogprobsLists,
LogprobsTensors,
ModelRunnerOutput,
PoolerOutput,
RoutedExpertsLists,
RoutedExpertsTensors,
SamplerOutput,
make_empty_encoder_model_runner_output,
)
from vllm.v1.pool.late_interaction_runner import LateInteractionRunner
from vllm.v1.pool.metadata import PoolingMetadata, PoolingStates
from vllm.v1.sample.logits_processor import LogitsProcessors, build_logitsprocs
from vllm.v1.sample.logits_processor.interface import LogitsProcessor
from vllm.v1.sample.metadata import SamplingMetadata
from vllm.v1.sample.rejection_sampler import RejectionSampler
from vllm.v1.sample.sampler import Sampler
from vllm.v1.spec_decode.custom_class_proposer import create_custom_proposer
from vllm.v1.spec_decode.dflash import DFlashProposer
from vllm.v1.spec_decode.draft_model import DraftModelProposer
from vllm.v1.spec_decode.eagle import EagleProposer
from vllm.v1.spec_decode.extract_hidden_states import ExtractHiddenStatesProposer
from vllm.v1.spec_decode.gemma4 import Gemma4Proposer
from vllm.v1.spec_decode.medusa import MedusaProposer
from vllm.v1.spec_decode.metadata import SpecDecodeMetadata
from vllm.v1.spec_decode.ngram_proposer_gpu import (
NgramProposerGPU,
copy_num_valid_draft_tokens,
update_ngram_gpu_tensors_incremental,
update_scheduler_for_invalid_drafts,
)
from vllm.v1.spec_decode.step3p5 import Step3p5MTPProposer
from vllm.v1.spec_decode.suffix_decoding import SuffixDecodingProposer
from vllm.v1.spec_decode.utils import update_num_computed_tokens_for_batch_change
from vllm.v1.structured_output.utils import apply_grammar_bitmask
from vllm.v1.utils import CpuGpuBuffer, record_function_or_nullcontext
from vllm.v1.worker import mamba_utils
from vllm.v1.worker.block_table import SlotMappingMode
from vllm.v1.worker.cp_utils import (
check_attention_cp_compatibility,
get_dcp_dummy_context_len,
prepare_dcp_dummy_context_metadata,
)
from vllm.v1.worker.dp_utils import coordinate_batch_across_dp
from vllm.v1.worker.ec_connector_model_runner_mixin import ECConnectorModelRunnerMixin
from vllm.v1.worker.gpu.attn_utils import _reshape_attention_kv_cache
from vllm.v1.worker.gpu_input_batch import CachedRequestState, InputBatch
from vllm.v1.worker.gpu_ubatch_wrapper import UBatchWrapper
from vllm.v1.worker.kv_connector_model_runner_mixin import KVConnectorModelRunnerMixin
from vllm.v1.worker.lora_model_runner_mixin import LoRAModelRunnerMixin
from vllm.v1.worker.ubatch_utils import (
UBatchSlices,
check_ubatch_thresholds,
maybe_create_ubatch_slices,
split_attn_metadata,
)
from vllm.v1.worker.utils import is_residual_scattered_for_sp
from vllm.v1.worker.workspace import lock_workspace
from .utils import (
AttentionGroup,
KVBlockZeroer,
add_kv_sharing_layers_to_kv_cache_groups,
bind_kv_cache,
copy_kv_cache_blocks_inplace,
prepare_kernel_block_sizes,
sanity_check_mm_encoder_outputs,
)
if TYPE_CHECKING:
from vllm.v1.core.sched.output import GrammarOutput, SchedulerOutput
from vllm.v1.spec_decode.ngram_proposer import NgramProposer
from vllm.v1.worker.encoder_cudagraph import EncoderCudaGraphManager
logger = init_logger(__name__)
AttnMetadataDict: TypeAlias = dict[str, AttentionMetadata]
# list when ubatching is enabled
PerLayerAttnMetadata: TypeAlias = list[AttnMetadataDict] | AttnMetadataDict
# Wrapper for ModelRunnerOutput to support overlapped execution.
class AsyncGPUModelRunnerOutput(AsyncModelRunnerOutput):
def __init__(
self,
model_runner_output: ModelRunnerOutput,
sampled_token_ids: torch.Tensor,
logprobs_tensors: LogprobsTensors | None,
invalid_req_indices: list[int],
async_output_copy_stream: torch.cuda.Stream,
vocab_size: int,
routed_experts: RoutedExpertsTensors | None = None,
check_ep_fault: bool = False,
):
self._model_runner_output = model_runner_output
self._invalid_req_indices = invalid_req_indices
# Event on the copy stream so we can synchronize the non-blocking copy.
# Blocking (sleep) event to avoid busy-polling the CUDA driver lock.
self.async_copy_ready_event = torch.cuda.Event(blocking=True)
# Keep a reference to the device tensor to avoid it being
# deallocated until we finish copying it to the host.
self._sampled_token_ids = sampled_token_ids
self.vocab_size = vocab_size
self._logprobs_tensors = logprobs_tensors
self._routed_experts = routed_experts
self._has_fault: torch.Tensor | None = None
# Initiate the copy on a separate stream, but do not synchronize it.
default_stream = torch.cuda.current_stream()
with torch.cuda.stream(async_output_copy_stream):
async_output_copy_stream.wait_stream(default_stream)
self.sampled_token_ids_cpu = self._sampled_token_ids.to(
"cpu", non_blocking=True
)
self._logprobs_tensors_cpu = (
self._logprobs_tensors.to_cpu_nonblocking()
if self._logprobs_tensors
else None
)
self._routed_experts_cpu = (
self._routed_experts.to_cpu_nonblocking()
if self._routed_experts is not None
else None
)
if check_ep_fault:
has_fault = get_ep_all2all_manager().query_fault()
self._has_fault = has_fault.to("cpu", non_blocking=True)
self.async_copy_ready_event.record()
def get_output(self) -> ModelRunnerOutput:
"""Copy the device tensors to the host and return a ModelRunnerOutput.
This function blocks until the copy is finished.
"""
max_gen_len = self.sampled_token_ids_cpu.shape[-1]
self.async_copy_ready_event.synchronize()
# Release the device tensors once the copy has completed.
del self._logprobs_tensors
del self._sampled_token_ids
if max_gen_len == 1:
valid_sampled_token_ids = self.sampled_token_ids_cpu.tolist()
for i in self._invalid_req_indices:
valid_sampled_token_ids[i].clear()
logprobs_lists = None
if self._logprobs_tensors_cpu is not None:
logprobs_lists = self._logprobs_tensors_cpu.tolists()
else:
valid_sampled_token_ids, logprobs_lists = RejectionSampler.parse_output(
self.sampled_token_ids_cpu,
self.vocab_size,
self._invalid_req_indices,
logprobs_tensors=self._logprobs_tensors_cpu,
)
output = self._model_runner_output
output.sampled_token_ids = valid_sampled_token_ids
output.logprobs = logprobs_lists
if self._routed_experts_cpu is not None:
output.routed_experts = self._routed_experts_cpu.tolists()
del self._routed_experts
if self._has_fault is not None and self._has_fault.item():
mask = get_ep_all2all_manager().query_active_mask()
raise RuntimeError(
"Fault detected in EP all2all communication: "
"one or more ranks timed out during dispatch/combine. "
f"Mask: {mask.cpu().tolist()}"
)
return output
def _copy_pooler_output_to_cpu(
raw_pooler_output: PoolerOutput, finished_mask: list[bool]
) -> list[torch.Tensor | None]:
num_reqs = len(finished_mask)
if isinstance(raw_pooler_output, torch.Tensor):
if raw_pooler_output.shape[0] != num_reqs:
raise ValueError(
"Pooler output batch size does not match finished mask size: "
f"{raw_pooler_output.shape[0]} != {num_reqs}."
)
num_finished = sum(finished_mask)
if num_finished == 0:
return [None] * num_reqs
if num_finished == num_reqs:
return list(raw_pooler_output.to("cpu", non_blocking=True))
# partial finished
finished_indices = [i for i, include in enumerate(finished_mask) if include]
index_tensor = torch.tensor(
finished_indices, device=raw_pooler_output.device, dtype=torch.long
)
finished_outputs = raw_pooler_output.index_select(0, index_tensor).to(
"cpu", non_blocking=True
)
partial_pooler_output: list[torch.Tensor | None] = [None] * num_reqs
for i, out in zip(finished_indices, finished_outputs):
partial_pooler_output[i] = out
return partial_pooler_output
assert isinstance(raw_pooler_output, list)
if len(raw_pooler_output) != num_reqs:
raise ValueError(
"Pooler output batch size does not match finished mask size: "
f"{len(raw_pooler_output)} != {num_reqs}."
)
pooler_output: list[torch.Tensor | None] = [None] * num_reqs
for i, (out, include) in enumerate(zip(raw_pooler_output, finished_mask)):
if include and out is not None:
pooler_output[i] = out.to("cpu", non_blocking=True)
return pooler_output
class AsyncGPUPoolingModelRunnerOutput(AsyncModelRunnerOutput):
def __init__(
self,
model_runner_output: ModelRunnerOutput,
raw_pooler_output: PoolerOutput,
finished_mask: list[bool],
async_output_copy_stream: torch.cuda.Stream,
):
self._model_runner_output = model_runner_output
# Event on the copy stream so we can synchronize the non-blocking copy.
# Blocking (sleep) event to avoid busy-polling the CUDA driver lock.
self.async_copy_ready_event = torch.cuda.Event(blocking=True)
# Keep a reference to the device tensors to avoid them being
# deallocated until we finish copying it to the host.
self._raw_pooler_output = raw_pooler_output
# Initiate the copy on a separate stream, but do not synchronize it.
default_stream = torch.cuda.current_stream()
with torch.cuda.stream(async_output_copy_stream):
async_output_copy_stream.wait_stream(default_stream)
self._model_runner_output.pooler_output = _copy_pooler_output_to_cpu(
raw_pooler_output=self._raw_pooler_output,
finished_mask=finished_mask,
)
self.async_copy_ready_event.record()
def get_output(self) -> ModelRunnerOutput:
"""Copy the device tensors to the host and return a ModelRunnerOutput.
This function blocks until the copy is finished.
"""
self.async_copy_ready_event.synchronize()
# Release the device tensors once the copy has completed.
del self._raw_pooler_output
return self._model_runner_output
class ExecuteModelState(NamedTuple):
"""Ephemeral cached state transferred between execute_model() and
sample_tokens(), after execute_model() returns None."""
scheduler_output: "SchedulerOutput"
logits: torch.Tensor
spec_decode_metadata: SpecDecodeMetadata | None
spec_decode_common_attn_metadata: CommonAttentionMetadata | None
hidden_states: torch.Tensor
sample_hidden_states: torch.Tensor
aux_hidden_states: list[torch.Tensor] | None
ec_connector_output: ECConnectorOutput | None
cudagraph_stats: CUDAGraphStat | None
slot_mappings: dict[str, torch.Tensor] | list[dict[str, torch.Tensor]] | None
class GPUModelRunner(
LoRAModelRunnerMixin, KVConnectorModelRunnerMixin, ECConnectorModelRunnerMixin
):
def __init__(
self,
vllm_config: VllmConfig,
device: torch.device,
):
self.vllm_config = vllm_config
self.model_config = vllm_config.model_config
self.cache_config = vllm_config.cache_config
self.offload_config = vllm_config.offload_config
self.compilation_config = vllm_config.compilation_config
self.lora_config = vllm_config.lora_config
self.load_config = vllm_config.load_config
self.parallel_config = vllm_config.parallel_config
self.scheduler_config = vllm_config.scheduler_config
self.speculative_config = vllm_config.speculative_config
self.observability_config = vllm_config.observability_config
model_config = self.model_config
cache_config = self.cache_config
scheduler_config = self.scheduler_config
parallel_config = self.parallel_config
self.device = device
self.dtype = self.model_config.dtype
self.check_ep_fault = False
if parallel_config.data_parallel_size > 1 and self.model_config.is_moe:
self.check_ep_fault = get_ep_all2all_manager().support_fault_tolerance
self.kv_cache_dtype = kv_cache_dtype_str_to_dtype(
cache_config.cache_dtype, self.model_config
)
self.is_pooling_model = model_config.runner_type == "pooling"
self.enable_prompt_embeds = model_config.enable_prompt_embeds
self.is_multimodal_raw_input_only_model = (
model_config.is_multimodal_raw_input_only_model
)
# These will be overridden in load_model()
self.is_multimodal_pruning_enabled = False
self.requires_sequential_video_encoding = False
# Set to True after init_routed_experts_capturer() completes.
# Prevents routed experts code from running during profiling/dummy run.
self.routed_experts_initialized = False
self.max_model_len = model_config.max_model_len
# Always set to false after the first forward pass
self.calculate_kv_scales = self.cache_config.calculate_kv_scales
self.dcp_world_size = self.parallel_config.decode_context_parallel_size
self.dcp_rank = 0 if self.dcp_world_size <= 1 else get_dcp_group().rank_in_group
self.max_num_tokens = scheduler_config.max_num_batched_tokens
self.max_num_reqs = scheduler_config.max_num_seqs
# Broadcast PP output for external_launcher (torchrun)
# to make sure we are synced across pp ranks
# TODO: Support overlapping micro-batches
# https://github.com/vllm-project/vllm/issues/18019
self.broadcast_pp_output = (
self.parallel_config.distributed_executor_backend == "external_launcher"
and len(get_pp_group().ranks) > 1
)
# Model-related.
self.num_query_heads = model_config.get_num_attention_heads(parallel_config)
self.inputs_embeds_size = model_config.get_inputs_embeds_size()
# Only relevant for models using ALiBi (e.g, MPT)
self.use_alibi = model_config.uses_alibi
self.cascade_attn_enabled = not self.model_config.disable_cascade_attn
self.is_mm_prefix_lm = self.model_config.is_mm_prefix_lm
# Multi-modal data support
self.mm_registry = MULTIMODAL_REGISTRY
self.uses_mrope = model_config.uses_mrope
self.uses_xdrope_dim = model_config.uses_xdrope_dim
self.supports_mm_inputs = self.mm_registry.supports_multimodal_inputs(
model_config
)
if self.model_config.is_encoder_decoder:
# Maximum length of the encoder input, only for encoder-decoder
# models.
self.max_encoder_len = scheduler_config.max_num_encoder_input_tokens
else:
self.max_encoder_len = 0
# Async scheduling
self.use_async_scheduling = self.scheduler_config.async_scheduling
# Sampler
self.sampler = Sampler(
logprobs_mode=self.model_config.logprobs_mode,
use_fp64_gumbel=self.model_config.use_fp64_gumbel,
)
self.eplb_state: EplbState | None = None
self._moe_model: MixtureOfExperts | None = None
# NOTE(yongji): flag to temporarily disable EPLB during scaling up/down
self.eep_eplb_suppressed = False
"""
State of the expert parallelism load balancer.
Will be lazily initialized when the model is loaded.
"""
# Lazy initializations
# self.model: nn.Module # Set after load_model
# Initialize in initialize_kv_cache
self.kv_caches: list[torch.Tensor] = []
# Initialize in initialize_kv_cache_tensors
self.cross_layers_kv_cache: torch.Tensor | None = None
self.cross_layers_attn_backend: type[AttentionBackend] | None = None
# indexes: [kv_cache_group_id][attn_group]
self.attn_groups: list[list[AttentionGroup]] = []
# self.kv_cache_config: KVCacheConfig
# mm_hash -> encoder_output
self.encoder_cache: dict[str, torch.Tensor] = {}
self.late_interaction_runner = LateInteractionRunner()
# Encoder CUDA graph manager (initialized after model load if enabled)
self.encoder_cudagraph_manager: EncoderCudaGraphManager | None = None
self.use_aux_hidden_state_outputs = False
# Set up speculative decoding.
# NOTE(Jiayi): currently we put the entire draft model on
# the last PP rank. This is not ideal if there are many
# layers in the draft model.
if self.speculative_config and get_pp_group().is_last_rank:
self.drafter: (
NgramProposer # noqa: F823
| NgramProposerGPU
| SuffixDecodingProposer
| EagleProposer
| DFlashProposer
| DraftModelProposer
| MedusaProposer
| ExtractHiddenStatesProposer
| Gemma4Proposer
| Step3p5MTPProposer
)
if self.speculative_config.method == "custom_class":
self.drafter = create_custom_proposer( # type: ignore[assignment]
self.vllm_config
)
elif self.speculative_config.method == "ngram":
from vllm.v1.spec_decode.ngram_proposer import NgramProposer
self.drafter = NgramProposer(self.vllm_config)
elif self.speculative_config.uses_draft_model():
self.drafter = DraftModelProposer(
vllm_config=self.vllm_config,
device=self.device,
runner=self,
)
elif self.speculative_config.use_ngram_gpu():
self.drafter = NgramProposerGPU(self.vllm_config, self.device, self)
self.num_tokens_no_spec_gpu = torch.zeros(
self.max_num_reqs, dtype=torch.int32, device=device
)
self.token_ids_gpu_tensor = torch.zeros(
self.max_num_reqs,
self.max_model_len,
dtype=torch.int32,
device=device,
)
self._ngram_pinned_idx_buf = torch.zeros(
self.max_num_reqs, dtype=torch.long, pin_memory=True
)
self._ngram_pinned_val_buf = torch.zeros(
self.max_num_reqs, dtype=torch.int32, pin_memory=True
)
elif self.speculative_config.use_gemma4_mtp():
self.drafter = Gemma4Proposer(self.vllm_config, self.device, self)
elif self.speculative_config.use_step3p5_mtp():
self.drafter = Step3p5MTPProposer(self.vllm_config, self.device, self)
elif self.speculative_config.use_dflash():
self.drafter = DFlashProposer(self.vllm_config, self.device, self)
self.use_aux_hidden_state_outputs = True
elif self.speculative_config.method == "suffix":
self.drafter = SuffixDecodingProposer(self.vllm_config)
elif self.speculative_config.use_eagle():
self.drafter = EagleProposer(self.vllm_config, self.device, self)
if self.speculative_config.method == "eagle3":
self.use_aux_hidden_state_outputs = (
self.drafter.eagle3_use_aux_hidden_state
)
elif self.speculative_config.method == "medusa":
self.drafter = MedusaProposer(
vllm_config=self.vllm_config, device=self.device
)
elif self.speculative_config.method == "extract_hidden_states":
self.drafter = ExtractHiddenStatesProposer(
vllm_config=self.vllm_config, device=self.device
)
self.use_aux_hidden_state_outputs = True
else:
raise ValueError(
"Unknown speculative decoding method: "
f"{self.speculative_config.method}"
)
self.rejection_sampler = RejectionSampler(
self.sampler, self.speculative_config, self.device
)
self.num_spec_tokens = 0
self.prev_num_spec_tokens = 0
self.valid_sampled_token_count_gpu: torch.Tensor | None = None
if self.speculative_config:
self.num_spec_tokens = self.speculative_config.num_speculative_tokens
self.prev_num_spec_tokens = self.num_spec_tokens
draft_config = self.speculative_config.draft_model_config
if draft_config is not None and draft_config.max_model_len is not None:
self.effective_drafter_max_model_len = draft_config.max_model_len
else:
self.effective_drafter_max_model_len = self.max_model_len
self.use_async_spec_decode = (
self.use_async_scheduling and self.num_spec_tokens > 0
)
# Request states.
self.requests: dict[str, CachedRequestState] = {}
# NOTE(rob): num_prompt_logprobs only includes reqs
# that are currently in the prefill phase.
self.num_prompt_logprobs: dict[str, int] = {}
# Input Batch
# NOTE(Chen): Ideally, we should initialize the input batch inside
# `initialize_kv_cache` based on the kv cache config. However, as in
# https://github.com/vllm-project/vllm/pull/18298, due to some unknown
# reasons, we have to initialize the input batch before `load_model`,
# quantization + weight offloading will fail otherwise. As a temporary
# solution, we initialize the input batch here, and re-initialize it
# in `initialize_kv_cache` if the block_sizes here is different from
# the block_sizes in the kv cache config.
logits_processors = model_config.logits_processors
custom_logitsprocs: Sequence[str | type[LogitsProcessor]] = (
tuple(logits_processors) if logits_processors is not None else ()
)
placeholder_block_size = (
self.cache_config.block_size or CacheConfig.DEFAULT_BLOCK_SIZE
)
placeholder_max_num_blocks = cdiv(
max(self.max_model_len, self.max_encoder_len), placeholder_block_size
)
self._init_block_sizes = [placeholder_block_size]
self._init_kernel_block_sizes = [placeholder_block_size]
self._init_max_num_blocks = [placeholder_max_num_blocks]
self._init_slot_mapping_modes = [SlotMappingMode.TOKEN_TO_KV_SLOT]
self.input_batch = InputBatch(
max_num_reqs=self.max_num_reqs,
# We need to use the encoder length for encoder-decoder
# because of KV cache for cross-attention.
max_model_len=max(self.max_model_len, self.max_encoder_len),
max_num_batched_tokens=self.max_num_tokens,
device=self.device,
vocab_size=self.model_config.get_vocab_size(),
block_sizes=[placeholder_block_size],
kernel_block_sizes=[placeholder_block_size],
max_num_blocks_per_req=[placeholder_max_num_blocks],
num_spec_tokens=self.num_spec_tokens,
logitsprocs=build_logitsprocs(
self.vllm_config,
self.device,
PIN_MEMORY,
self.is_pooling_model,
custom_logitsprocs,
),
# We currently don't know whether a particular custom logits processor
# uses output token ids so we set this conservatively. Thinking-budget
# tracking is requested dynamically when a budgeted request is in the batch.
logitsprocs_need_output_token_ids=bool(custom_logitsprocs),
is_pooling_model=self.is_pooling_model,
cp_kv_cache_interleave_size=self.parallel_config.cp_kv_cache_interleave_size,
reasoning_config=self.vllm_config.reasoning_config,
)
# Separate cuda stream for overlapping transfer of sampled token ids from
# GPU to CPU when async scheduling is enabled.
self.async_output_copy_stream: torch.cuda.Stream | None = None
# cuda event to synchronize use of reused CPU tensors between steps
# when async scheduling is enabled.
self.prepare_inputs_event: torch.Event | None = None
if self.use_async_scheduling:
self.async_output_copy_stream = torch.cuda.Stream()
# Blocking (sleep) event to avoid busy-polling the CUDA driver lock;
# under TP contention that spin can balloon and make the rank a straggler.
self.prepare_inputs_event = torch.cuda.Event(blocking=True)
# self.cudagraph_batch_sizes sorts in ascending order.
if (
self.compilation_config.cudagraph_capture_sizes
and self.compilation_config.cudagraph_mode != CUDAGraphMode.NONE
):
self.cudagraph_batch_sizes = sorted(
self.compilation_config.cudagraph_capture_sizes
)
else:
self.cudagraph_batch_sizes = []
# Cache the device properties.
self._init_device_properties()
# Encoder timing registry for observability
self.encoder_timing_registry: dict[str, EncoderTimingStats] = {}
self._encoder_timing_lock = threading.Lock()
# Persistent buffers for CUDA graphs.
self.input_ids = self._make_buffer(self.max_num_tokens, dtype=torch.int32)
self.positions = torch.zeros(
self.max_num_tokens, dtype=torch.int64, device=self.device
)
self.query_start_loc = self._make_buffer(
self.max_num_reqs + 1, dtype=torch.int32
)
self.seq_lens = torch.zeros(
self.max_num_reqs, dtype=torch.int32, device=self.device
)
self.optimistic_seq_lens_cpu = torch.zeros(
self.max_num_reqs, dtype=torch.int32, pin_memory=PIN_MEMORY
)
self.num_computed_tokens = torch.zeros(
self.max_num_reqs, dtype=torch.int32, device=self.device
)
self.prev_num_draft_tokens = self._make_buffer(
self.max_num_reqs, dtype=torch.int32
)
self.req_indices = self._make_buffer(self.max_num_tokens, dtype=torch.int64)
# Maps current batch position -> previous batch position (-1 for new reqs)
self.prev_positions = self._make_buffer(self.max_num_reqs, dtype=torch.int64)
self.num_scheduled_tokens = self._make_buffer(
self.max_num_reqs, dtype=torch.int32
)
self.encoder_seq_lens = self._make_buffer(self.max_num_reqs, dtype=torch.int32)
if self.dcp_world_size > 1:
self.dcp_local_seq_lens = self._make_buffer(
self.max_num_reqs, dtype=torch.int32
)
# Because inputs_embeds may be bfloat16 and we don't need a numpy
# version of this tensor, avoid a RuntimeError by not creating a
# numpy buffer.
self.inputs_embeds = self._make_buffer(
self.max_num_tokens, self.inputs_embeds_size, dtype=self.dtype, numpy=False
)
self.is_token_ids = self._make_buffer(self.max_num_tokens, dtype=torch.bool)
self.discard_request_mask = self._make_buffer(
self.max_num_reqs, dtype=torch.bool
)
self.num_decode_draft_tokens = self._make_buffer(
self.max_num_reqs, dtype=torch.int32
)
self.num_accepted_tokens = self._make_buffer(
self.max_num_reqs, dtype=torch.int32
)
# Only relevant for models using M-RoPE (e.g, Qwen2-VL)
if self.uses_mrope:
# 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
self.mrope_positions = self._make_buffer(
(3, self.max_num_tokens + 1), dtype=torch.int64
)
# Only relevant for models using XD-RoPE (e.g, HunYuan-VL)
if self.uses_xdrope_dim > 0:
# Similar to mrope but use assigned dimension number for RoPE, 4 as default.
self.xdrope_positions = self._make_buffer(
(self.uses_xdrope_dim, self.max_num_tokens + 1), dtype=torch.int64
)
# None in the first PP rank. The rest are set after load_model.
self.intermediate_tensors: IntermediateTensors | None = None
# OPTIMIZATION: Cache the arange tensors rather than creating them
# every step. Keep in int64 to avoid overflow with long context.
# - arange_np: immutable [0, 1, 2, ...] used as source for batched computation
# - query_pos: CpuGpuBuffer for the computed batched arange result
arange_size = max(self.max_num_reqs + 1, self.max_num_tokens)
self.arange_np = np.arange(arange_size, dtype=np.int64)
self.query_pos = self._make_buffer(arange_size, dtype=torch.int64)
self._arange_scratch = np.empty(arange_size, dtype=np.int64)
# Layer pairings for cross-layer KV sharing.
# If an Attention layer `layer_name` is in the keys of this dict, it
# means this layer will perform attention using the keys and values
# from the KV cache of `shared_kv_cache_layers[layer_name]`.
self.shared_kv_cache_layers: dict[str, str] = {}
self.kv_sharing_fast_prefill_eligible_layers: set[str] = set()
self.kv_sharing_fast_prefill_logits_indices = None
if self.cache_config.kv_sharing_fast_prefill:
self.kv_sharing_fast_prefill_logits_indices = torch.zeros(
self.max_num_tokens, dtype=torch.int32, device=self.device
)
self.uniform_decode_query_len = 1 + self.num_spec_tokens
# Cudagraph dispatcher for runtime cudagraph dispatching.
self.cudagraph_dispatcher = CudagraphDispatcher(self.vllm_config)
self.mm_budget = (
MultiModalBudget(self.vllm_config, self.mm_registry)
if self.supports_mm_inputs
else None
)
self.reorder_batch_threshold: int | None = None
# Attention layers that are only in the KVCacheConfig of the runner
# (e.g., KV sharing, encoder-only attention), but not in the
# KVCacheConfig of the scheduler.
self.runner_only_attn_layers: set[str] = set()
# Cached outputs.
self._draft_token_ids: list[list[int]] | torch.Tensor | None = None
self._draft_probs: torch.Tensor | None = None
self._draft_prob_req_ids: list[str] | None = None
# N-gram GPU path: async D2H buffer/event for per-request valid draft counts.
self._num_valid_draft_tokens: torch.Tensor | None = None
self._num_valid_draft_tokens_cpu: torch.Tensor | None = None
self._num_valid_draft_tokens_event: torch.cuda.Event | None = None
self._num_valid_draft_tokens_copy_stream: torch.cuda.Stream | None = None
if (
self.speculative_config is not None
and self.speculative_config.use_ngram_gpu()
):
self._num_valid_draft_tokens_cpu = torch.empty(
self.max_num_reqs, dtype=torch.int32, pin_memory=PIN_MEMORY
)
self._num_valid_draft_tokens_event = torch.cuda.Event()
self._num_valid_draft_tokens_copy_stream = torch.cuda.Stream()
self._draft_token_req_ids: list[str] | None = None
self.transfer_event = torch.Event()
self.sampled_token_ids_pinned_cpu = torch.empty(
(self.max_num_reqs, 1),
dtype=torch.int64,
device="cpu",
pin_memory=PIN_MEMORY,
)
# Pre-allocated tensor for copying valid sampled token counts to CPU,
# with dedicated stream for overlapping and event for coordination.
self.valid_sampled_token_count_event: torch.Event | None = None
self.valid_sampled_token_count_copy_stream: torch.cuda.Stream | None = None
# We also copy the drafted tokens to the CPU asynchronously,
# in case we need them for structured outputs.
self.draft_token_ids_event: torch.Event | None = None
self.draft_token_ids_copy_stream: torch.cuda.Stream | None = None
self.valid_sampled_token_count_cpu: torch.Tensor | None = None
self.draft_token_ids_cpu: torch.Tensor | None = None
self.num_accepted_tokens_event: torch.Event | None = None
if self.num_spec_tokens:
self.draft_token_ids_event = torch.Event()
self.num_accepted_tokens_event = torch.Event()
self.draft_token_ids_copy_stream = torch.cuda.Stream()
self.draft_token_ids_cpu = torch.empty(
(self.max_num_reqs, self.num_spec_tokens),
dtype=torch.int64,
device="cpu",
pin_memory=PIN_MEMORY,
)
if self.use_async_scheduling:
self.valid_sampled_token_count_event = torch.Event()
self.valid_sampled_token_count_copy_stream = torch.cuda.Stream()
self.valid_sampled_token_count_cpu = torch.empty(
self.max_num_reqs,
dtype=torch.int32,
device="cpu",
pin_memory=PIN_MEMORY,
)
# Model weight offloader
# Make sure this is called before any get_offloader call
set_offloader(create_offloader(self.offload_config))
# Ephemeral state transferred between execute_model() and sample_tokens().
self.execute_model_state: ExecuteModelState | None = None
self.kv_connector_output: KVConnectorOutput | None = None
self.mamba_state_idx: dict[str, int] = {}
self._mamba_bufs: mamba_utils.MambaBuffers | None = None
self.mamba_prev_last_scheduled_idx: CpuGpuBuffer | None = None
if self.cache_config.mamba_cache_mode == "all" and self.num_spec_tokens > 0:
self.mamba_prev_last_scheduled_idx = self._make_buffer(
self.max_num_reqs, dtype=torch.int32
)
self.layerwise_nvtx_hooks_registered = False
def update_max_model_len(self, max_model_len: int) -> None:
self.max_model_len = max_model_len
if self.speculative_config:
draft_config = self.speculative_config.draft_model_config
if draft_config is None or draft_config.max_model_len is None:
self.effective_drafter_max_model_len = self.max_model_len
def reset_mm_cache(self) -> None:
"""
Clear the multi-modal cache that was used during profiling,
but no longer needed during inference.
"""
if self.mm_budget:
self.mm_budget.reset_cache()
self.late_interaction_runner.clear()
def reset_encoder_cache(self) -> None:
"""Clear the GPU-side encoder cache storing vision embeddings.
This should be called when model weights are updated to ensure
stale embeddings computed with old weights are not reused.
"""
self.encoder_cache.clear()
self.late_interaction_runner.clear()
def post_kv_cache_wake_up(self) -> None:
self.init_fp8_kv_scales()
@torch.inference_mode()
def init_fp8_kv_scales(self) -> None:
"""
Re-initialize the KV cache and FP8 scales after waking from sleep.
1. Zero out the KV cache tensors to remove garbage data from re-allocation.
2. Reset Attention layer scaling factors (_k_scale, _v_scale) to 1.0.
If these are left at 0.0 (default after wake_up), all KV cache values
become effectively zero, causing gibberish output.
"""
if not is_quantized_kv_cache(self.cache_config.cache_dtype):
return
kv_caches = getattr(self, "kv_caches", [])
for cache_tensor in kv_caches:
if cache_tensor is not None:
cache_tensor.zero_()
k_attr_names = ("_k_scale", "k_scale")
v_attr_names = ("_v_scale", "v_scale")
attn_layers = self.compilation_config.static_forward_context
for name, module in attn_layers.items():
if isinstance(module, (Attention, MLAAttention)):
# TODO: Generally, scale is 1.0 if user uses on-the-fly fp8
# kvcache quant. However, to get better accuracy, compression
# frameworks like llm-compressors allow users to tune the
# scale. We may need to restore the specific calibrated scales
# here in the future.
k_scale_val, v_scale_val = 1.0, 1.0
# Processing K Scale
for attr in k_attr_names:
if hasattr(module, attr):
param = getattr(module, attr)
if isinstance(param, torch.Tensor):
param.fill_(k_scale_val)
# Processing V Scale
for attr in v_attr_names:
if hasattr(module, attr):
param = getattr(module, attr)
if isinstance(param, torch.Tensor):
param.fill_(v_scale_val)
def _get_positions(self, num_tokens: Any):
if isinstance(num_tokens, int):
if self.uses_mrope:
return self.mrope_positions.gpu[:, :num_tokens]
if self.uses_xdrope_dim > 0:
return self.xdrope_positions.gpu[:, :num_tokens]
return self.positions[:num_tokens]
else:
if self.uses_mrope:
return self.mrope_positions.gpu[:, num_tokens]
if self.uses_xdrope_dim > 0:
return self.xdrope_positions.gpu[:, num_tokens]
return self.positions[num_tokens]
def _make_buffer(
self, *size: int | torch.SymInt, dtype: torch.dtype, numpy: bool = True
) -> CpuGpuBuffer:
return CpuGpuBuffer(
*size,
dtype=dtype,
device=self.device,
with_numpy=numpy,
)
def _get_mamba_bufs(self) -> mamba_utils.MambaBuffers:
# Only reachable on the ``mamba_cache_mode == "align"`` path.
# The postprocess sub-object is additionally gated on spec
# decode + hybrid model.
assert self.cache_config.mamba_cache_mode == "align"
if self._mamba_bufs is None:
self._mamba_bufs = mamba_utils.MambaBuffers.create(
max_num_reqs=self.max_num_reqs,
kv_cache_config=self.kv_cache_config,
copy_funcs=self.model.get_mamba_state_copy_func(),
make_buffer=self._make_buffer,
device=self.device,
with_postprocess_align=(
self.speculative_config is not None and self.model_config.is_hybrid
),
)
return self._mamba_bufs
def _init_model_kwargs(self):
model_kwargs = dict[str, Any]()
if not self.is_pooling_model:
return model_kwargs
num_reqs = self.input_batch.num_reqs
pooling_params = self.input_batch.get_pooling_params()
token_type_id_requests = dict[int, Any]()
for i, param in enumerate(pooling_params):
if (
param.extra_kwargs is not None
and (token_types := param.extra_kwargs.get("compressed_token_type_ids"))
is not None
):
token_type_id_requests[i] = token_types
if len(token_type_id_requests) == 0:
return model_kwargs
# Build ids on CPU using the CPU-resident upper bound for seq_lens;
# `torch.arange(seq_lens[i])` with a GPU scalar would force a sync.
seq_lens_cpu = self.optimistic_seq_lens_cpu[:num_reqs].tolist()
token_type_ids = []
for i in range(num_reqs):
seq_len_i = seq_lens_cpu[i]
pos = token_type_id_requests.get(i, seq_len_i)
ids = (torch.arange(seq_len_i) >= pos).int()
token_type_ids.append(ids)
token_type_ids_cpu = torch.empty(
sum(seq_lens_cpu), dtype=torch.int32, pin_memory=PIN_MEMORY
)
torch.cat(token_type_ids, out=token_type_ids_cpu)
model_kwargs["token_type_ids"] = token_type_ids_cpu.to(
device=self.device, non_blocking=True
)
return model_kwargs
def _may_reorder_batch(self, scheduler_output: "SchedulerOutput") -> None:
"""
Update the order of requests in the batch based on the attention
backend's needs. For example, some attention backends (namely MLA) may
want to separate requests based on if the attention computation will be
compute-bound or memory-bound.
Args:
scheduler_output: The scheduler output.
"""
# Attention free models have zero kv_cache_groups, however models
# like Mamba are also attention free but use the kv_cache for
# keeping its internal state. This is why we check the number
# of kv_cache groups instead of solely checking
# for self.model_config.is_attention_free.
if len(self.kv_cache_config.kv_cache_groups) == 0:
return
if self.reorder_batch_threshold is not None:
reorder_batch_to_split_decodes_and_prefills(
self.input_batch,
scheduler_output,
decode_threshold=self.reorder_batch_threshold,
)
def _init_kv_zero_meta(self) -> None:
"""One-time precomputation for _zero_block_ids.
Called from gpu_worker.py outside the CuMem pool context.
"""
self._kv_block_zeroer = KVBlockZeroer(
self.device,
pin_memory=PIN_MEMORY,
attn_groups_iter=self._kv_cache_spec_attn_group_iterator(),
kernel_block_sizes=self._kernel_block_sizes,
cache_dtype=self.cache_config.cache_dtype,
runner_only_attn_layers=self.runner_only_attn_layers,
static_forward_context=self.compilation_config.static_forward_context,
max_concurrency=self.vllm_config.max_concurrent_batches,
)
def _zero_block_ids(self, block_ids: list[int]) -> None:
"""Zero the KV cache memory for the given block IDs."""
if hasattr(self, "_kv_block_zeroer"):
self._kv_block_zeroer.zero_block_ids(block_ids)
# Note: used for model runner override.
def _init_device_properties(self) -> None:
"""Initialize attributes from torch.cuda.get_device_properties"""
self.num_sms = num_compute_units(self.device.index)
# Note: used for model runner override.
def _sync_device(self) -> None:
torch.accelerator.synchronize()
def _get_or_create_async_output_copy_stream(self) -> torch.cuda.Stream:
stream = self.async_output_copy_stream
if stream is None:
stream = torch.cuda.Stream()
self.async_output_copy_stream = stream
return stream
def _update_states(self, scheduler_output: "SchedulerOutput") -> Callable | None:
"""Update the cached states and the persistent batch with the scheduler
output.
The updated states are used by the `_prepare_inputs` function to create
the input GPU tensors for the model.
The SamplingMetadata is updated and copied to the GPU if there is a
new/resumed/paused/finished request in the batch.
"""
# Remove finished requests from the cached states.
for req_id in scheduler_output.finished_req_ids:
self.requests.pop(req_id, None)
self.num_prompt_logprobs.pop(req_id, None)
self.late_interaction_runner.on_requests_finished(
scheduler_output.finished_req_ids
)
# Remove the finished requests from the persistent batch.
# NOTE(woosuk): There could be an edge case where finished_req_ids and
# scheduled_req_ids overlap. This happens when a request is aborted and
# then resubmitted with the same ID. In this case, we treat them as two
# distinct requests - clearing the cached states for the first request
# and handling the second as a new request.
for req_id in scheduler_output.finished_req_ids:
self.input_batch.remove_request(req_id)
# Zero GPU memory for freshly allocated cache blocks to prevent
# stale NaN/data from corrupting attention or SSM computation.
if scheduler_output.new_block_ids_to_zero:
self._zero_block_ids(scheduler_output.new_block_ids_to_zero)
if scheduler_output.kv_cache_block_copies:
copy_kv_cache_blocks_inplace(
self.kv_caches,
self.kv_cache_config.num_blocks,
scheduler_output.kv_cache_block_copies,
)
# Free the cached encoder outputs.
for mm_hash in scheduler_output.free_encoder_mm_hashes:
self.encoder_cache.pop(mm_hash, None)
# Remove the unscheduled requests from the persistent batch.
# NOTE(woosuk): The unscheduled requests are either preempted requests
# or running requests that are not scheduled in this step. We remove
# them from the persistent batch but keep their cached states since
# they will be scheduled again sometime in the future.
scheduled_req_ids = scheduler_output.num_scheduled_tokens.keys()
cached_req_ids = self.input_batch.req_id_to_index.keys()
resumed_req_ids = scheduler_output.scheduled_cached_reqs.resumed_req_ids
# NOTE(zhuohan): cached_req_ids and resumed_req_ids are usually disjoint,
# so `(scheduled_req_ids - resumed_req_ids) == scheduled_req_ids` holds
# apart from the forced-preemption case in reset_prefix_cache. And in
# that case we include the resumed_req_ids in the unscheduled set so
# that they get cleared from the persistent batch before being re-scheduled
# in the normal resumed request path.
unscheduled_req_ids = cached_req_ids - (scheduled_req_ids - resumed_req_ids)
# NOTE(woosuk): The persistent batch optimization assumes that
# consecutive batches contain mostly the same requests. If batches
# have low request overlap (e.g., alternating between two distinct
# sets of requests), this optimization becomes very inefficient.
for req_id in unscheduled_req_ids:
self.input_batch.remove_request(req_id)
is_ngram_gpu = (
self.speculative_config is not None
and self.speculative_config.use_ngram_gpu()
)
if is_ngram_gpu:
ngram_gpu_new_reqs: list[CachedRequestState] = []
reqs_to_add: list[CachedRequestState] = []
deferred_spec_decode_corrections = []
# Add new requests to the cached states.
for new_req_data in scheduler_output.scheduled_new_reqs:
req_id = new_req_data.req_id
if req_id in self.requests:
# For streaming case only.
req_state = self._update_streaming_request(req_id, new_req_data)
reqs_to_add.append(req_state)
continue
sampling_params = new_req_data.sampling_params
pooling_params = new_req_data.pooling_params
if (
sampling_params
and sampling_params.sampling_type == SamplingType.RANDOM_SEED
):
generator = torch.Generator(device=self.device)
generator.manual_seed(sampling_params.seed)
else:
generator = None
if self.is_pooling_model:
assert pooling_params is not None
task = pooling_params.task
assert task is not None, "You did not set `task` in the API"
model = cast(VllmModelForPooling, self.get_model())
to_update = model.pooler.get_pooling_updates(task)
to_update.apply(pooling_params)
req_state = CachedRequestState(
req_id=req_id,
prompt_token_ids=new_req_data.prompt_token_ids,
prompt_embeds=new_req_data.prompt_embeds,
prompt_is_token_ids=new_req_data.prompt_is_token_ids,
mm_features=new_req_data.mm_features,
sampling_params=sampling_params,
pooling_params=pooling_params,
generator=generator,
block_ids=new_req_data.block_ids,
num_computed_tokens=new_req_data.num_computed_tokens,
output_token_ids=[],
lora_request=new_req_data.lora_request,
)
self.requests[req_id] = req_state
self.late_interaction_runner.register_request(req_id, pooling_params)
if sampling_params and sampling_params.prompt_logprobs is not None:
self.num_prompt_logprobs[req_id] = (
self.input_batch.vocab_size
if sampling_params.prompt_logprobs == -1
else sampling_params.prompt_logprobs
)
# Only relevant for models using M-RoPE (e.g, Qwen2-VL)
if self.uses_mrope:
self._init_mrope_positions(req_state)
# Only relevant for models using XD-RoPE (e.g, HunYuan-VL)
if self.uses_xdrope_dim > 0:
self._init_xdrope_positions(req_state)
reqs_to_add.append(req_state)
# Track new requests for ngram_gpu full tensor copy
if is_ngram_gpu:
ngram_gpu_new_reqs.append(req_state)
# Update the states of the running/resumed requests.
is_last_rank = get_pp_group().is_last_rank
req_data = scheduler_output.scheduled_cached_reqs
scheduled_spec_tokens = scheduler_output.scheduled_spec_decode_tokens
# Save scheduler-allocated spec lengths before trimming so
# prev_num_draft_len keeps the optimistic count for rejection correction.
original_num_spec_per_req: dict[str, int] = {}
if (
self.speculative_config is not None
and self.speculative_config.use_ngram_gpu()
):
for req_id, toks in scheduled_spec_tokens.items():
original_num_spec_per_req[req_id] = len(toks)
update_scheduler_for_invalid_drafts(
self._num_valid_draft_tokens_event,
self._num_valid_draft_tokens_cpu,
scheduler_output,
self.input_batch.req_id_to_index,
)
if self.use_async_spec_decode:
self.prev_num_draft_tokens.np.fill(0)
for i, req_id in enumerate(req_data.req_ids):
req_state = self.requests[req_id]
num_computed_tokens = req_data.num_computed_tokens[i]
new_block_ids = req_data.new_block_ids[i]
resumed_from_preemption = req_id in req_data.resumed_req_ids
num_output_tokens = req_data.num_output_tokens[i]
req_index = self.input_batch.req_id_to_index.get(req_id)
if req_state.prev_num_draft_len and self.use_async_scheduling:
# prev_num_draft_len is used in async scheduling mode with
# spec decode. it indicates if need to update num_computed_tokens
# of the request. for example:
# first step: num_computed_tokens = 0, spec_tokens = [],
# prev_num_draft_len = 0.
# second step: num_computed_tokens = 100(prompt length),
# spec_tokens = [a,b], prev_num_draft_len = 0.
# third step: num_computed_tokens = 100 + 2, spec_tokens = [c,d],
# prev_num_draft_len = 2.
# num_computed_tokens in first step and second step doesn't contain
# the spec tokens length, but in third step it contains the
# spec tokens length. we only need to update num_computed_tokens
# when prev_num_draft_len > 0.
if req_index is None:
req_state.prev_num_draft_len = 0
else:
# Optimistically assume all accepted; queue up a correction
# to be called after the model forward to preserve async
# scheduling. Corrected on GPU in _prepare_inputs.
optimistic_num_accepted = req_state.prev_num_draft_len
req_state.output_token_ids.extend([-1] * optimistic_num_accepted)
deferred_spec_decode_corrections.append(
(req_id, optimistic_num_accepted, req_state)
)
prev_req_index = (
self.input_batch.prev_req_id_to_index.get(req_id)
if self.input_batch.prev_req_id_to_index
else None
)
if prev_req_index is not None:
self.prev_num_draft_tokens.np[prev_req_index] = (
optimistic_num_accepted
)
if is_ngram_gpu and optimistic_num_accepted > 0:
self.input_batch.num_tokens_no_spec[req_index] += (
optimistic_num_accepted
)
# Update the cached states.
req_state.num_computed_tokens = num_computed_tokens
if not is_last_rank:
if not req_data.new_token_ids:
# Async scheduled PP: Sampled tokens propagated via GPU broadcast.
new_token_ids: list[int] = []
else:
# Non-async scheduling with PP: The scheduler sends
# sampled token ids back because there's no direct communication
# between the first-stage worker and the last-stage worker.
new_token_ids = req_data.new_token_ids[i]
# Add the sampled token(s) from the previous step (if any).
# This doesn't include "unverified" tokens like spec tokens.
num_new_tokens = (
num_computed_tokens + len(new_token_ids) - req_state.num_tokens
)
if num_new_tokens == 1:
# Avoid slicing list in most common case.
req_state.output_token_ids.append(new_token_ids[-1])
elif num_new_tokens > 0:
req_state.output_token_ids.extend(
new_token_ids[-num_new_tokens:]
)
elif num_output_tokens < len(req_state.output_token_ids):
# Some output tokens were discarded due to a sync-KV-load
# failure, or output_token_ids was inflated by the optimistic
# extend above (async spec decode). Align the cached state.
del req_state.output_token_ids[num_output_tokens:]
if req_index is not None:
end_idx = (
self.input_batch.num_prompt_tokens[req_index]
+ num_output_tokens
)
self.input_batch.num_tokens_no_spec[req_index] = end_idx
# Update the block IDs.
if not resumed_from_preemption:
if new_block_ids is not None:
# Append the new blocks to the existing block IDs.
for block_ids, new_ids in zip(req_state.block_ids, new_block_ids):
block_ids.extend(new_ids)
else:
assert req_index is None
assert new_block_ids is not None
# The request is resumed from preemption.
# Replace the existing block IDs with the new ones.
req_state.block_ids = new_block_ids
if req_index is None:
# The request is not in the persistent batch.
# The request was either preempted and resumed later, or was not
# scheduled in the previous step and needs to be added again.
if self.use_async_scheduling and num_output_tokens > 0:
# We must recover the output token ids for resumed requests in the
# async scheduling case, so that correct input_ids are obtained.
resumed_token_ids = req_data.all_token_ids[req_id]
req_state.output_token_ids = resumed_token_ids[-num_output_tokens:]
reqs_to_add.append(req_state)
# Track resumed requests for ngram_gpu full tensor copy
if is_ngram_gpu:
ngram_gpu_new_reqs.append(req_state)
continue
# Update the persistent batch.
self.input_batch.num_computed_tokens_cpu[req_index] = num_computed_tokens
if new_block_ids is not None:
self.input_batch.block_table.append_row(new_block_ids, req_index)
# For the last rank, we don't need to update the token_ids_cpu
# because the sampled tokens are already cached.
if not is_last_rank:
start_token_index = self.input_batch.num_tokens_no_spec[req_index]
# For chunked prefill, num_computed_tokens may less
# than num_tokens_no_spec.
# Async scheduled PP: no new_token_ids, advance num_tokens_no_spec
# according to num_computed_tokens.
end_token_index = max(
start_token_index,
num_computed_tokens + len(new_token_ids),
)
if end_token_index > start_token_index:
if new_token_ids:
# Add new_token_ids to token_ids_cpu.
num_new_tokens = end_token_index - start_token_index
tokens_to_append = new_token_ids[-num_new_tokens:]
self.input_batch.token_ids_cpu[
req_index, start_token_index:end_token_index
] = tokens_to_append
self.input_batch.is_token_ids[
req_index, start_token_index:end_token_index
] = True
self.input_batch.num_tokens_no_spec[req_index] = end_token_index
# Add spec_token_ids to token_ids_cpu.
self.input_batch.update_req_spec_token_ids(req_state, scheduled_spec_tokens)
# Restore scheduler-side draft count after ngram trimming.
if original_num_spec_per_req:
orig = original_num_spec_per_req.get(req_id, 0)
if orig != req_state.prev_num_draft_len:
req_state.prev_num_draft_len = orig
# Add the new or resumed requests to the persistent batch.
# The smaller empty indices are filled first.
for request in reqs_to_add:
self.input_batch.add_request(request)
self.input_batch.update_req_spec_token_ids(request, scheduled_spec_tokens)
# Condense the batched states if there are gaps left by removed requests
self.input_batch.condense()
# Allow attention backend to reorder the batch, potentially
self._may_reorder_batch(scheduler_output)
# Refresh batch metadata with any pending updates.
self.input_batch.refresh_metadata()
# Incrementally update ngram_gpu tensors after batch is stable
if is_ngram_gpu:
update_ngram_gpu_tensors_incremental(
self.input_batch,
self.token_ids_gpu_tensor,
self.num_tokens_no_spec_gpu,
ngram_gpu_new_reqs,
self.device,
_pinned_idx_buf=self._ngram_pinned_idx_buf,
_pinned_val_buf=self._ngram_pinned_val_buf,
)
if deferred_spec_decode_corrections:
def correct_spec_decode_token_counts():
valid_sampled_token_count = self._get_valid_sampled_token_count()
if not valid_sampled_token_count:
return
prev_req_id_to_index = self.input_batch.prev_req_id_to_index
if not prev_req_id_to_index:
return
for (
req_id,
optimistic_num_accepted,
req_state,
) in deferred_spec_decode_corrections:
prev_req_index = prev_req_id_to_index.get(req_id)
if prev_req_index is None:
continue
num_accepted = valid_sampled_token_count[prev_req_index] - 1
correction = optimistic_num_accepted - num_accepted
req_state.num_computed_tokens -= correction
cur_req_index = self.input_batch.req_id_to_index.get(req_id)
if cur_req_index is None:
continue
self.input_batch.num_computed_tokens_cpu[cur_req_index] -= (
correction
)
if is_ngram_gpu and correction > 0:
self.input_batch.num_tokens_no_spec[cur_req_index] -= correction
self.num_tokens_no_spec_gpu[cur_req_index] -= correction
return correct_spec_decode_token_counts
else:
return None
def _update_states_after_model_execute(
self, output_token_ids: torch.Tensor, scheduler_output: "SchedulerOutput"
) -> None:
"""Update the cached states after model execution.
This is used for MTP/EAGLE for hybrid models, as in linear attention,
only the last token's state is kept. In MTP/EAGLE, for draft tokens
the state are kept util we decide how many tokens are accepted for
each sequence, and a shifting is done during the next iteration
based on the number of accepted tokens.
"""
if not self.speculative_config or not self.model_config.is_hybrid:
return
# Count the number of accepted tokens for each sequence.
# Valid tokens are contiguous from position 0, so counting non-(-1)
# tokens gives us the first -1 position (i.e., number of accepted).
num_reqs = output_token_ids.size(0)
self.num_accepted_tokens.gpu[:num_reqs] = (output_token_ids != -1).sum(dim=1)
if self.cache_config.mamba_cache_mode == "align":
# Fused GPU postprocess: state copies + per-request accepted-token
# update without CPU-GPU sync. The metadata
# (num_scheduled_tokens, num_draft_tokens, num_computed_tokens) is
# pre-staged to GPU buffers in _prepare_inputs.
mamba_utils.postprocess_mamba_align_gpu(
bufs=self._get_mamba_bufs(),
num_reqs=num_reqs,
num_accepted_tokens_gpu=self.num_accepted_tokens.gpu,
num_accepted_tokens_cpu_tensor=(
self.input_batch.num_accepted_tokens_cpu_tensor
),
input_batch=self.input_batch,
kv_cache_config=self.kv_cache_config,
forward_context=self.compilation_config.static_forward_context,
mamba_state_copy_funcs=self.model.get_mamba_state_copy_func(),
)
assert self.num_accepted_tokens_event is not None
self.num_accepted_tokens_event.record()
else:
self.input_batch.num_accepted_tokens_cpu_tensor[:num_reqs].copy_(
self.num_accepted_tokens.gpu[:num_reqs], non_blocking=True
)
assert self.num_accepted_tokens_event is not None
self.num_accepted_tokens_event.record()
if self.cache_config.mamba_cache_mode == "all":
mamba_utils.postprocess_mamba_all(
scheduler_output,
self.kv_cache_config,
self.input_batch,
self.requests,
self.mamba_state_idx,
self.num_spec_tokens,
num_reqs,
)
def _update_streaming_request(
self, req_id: str, new_req_data: NewRequestData
) -> CachedRequestState:
"""Updates streaming session request from `scheduled_new_reqs`.
Removes the request from InputBatch (if present), updates the cached
state, and prepares it for re-addition to the batch.
NOTE: prompt_token_ids includes intermediate output tokens - tokens
previously generated but now are input context (part of the prompt).
"""
self.input_batch.remove_request(req_id)
req_state = self.requests[req_id]
req_state.prompt_token_ids = new_req_data.prompt_token_ids
req_state.mm_features = new_req_data.mm_features
req_state.prompt_embeds = new_req_data.prompt_embeds
req_state.sampling_params = new_req_data.sampling_params
req_state.pooling_params = new_req_data.pooling_params
self.late_interaction_runner.register_request(req_id, req_state.pooling_params)
req_state.block_ids = new_req_data.block_ids
req_state.num_computed_tokens = new_req_data.num_computed_tokens
req_state.num_prompt_tokens = length_from_prompt_token_ids_or_embeds(
req_state.prompt_token_ids, req_state.prompt_embeds
)
# Clear `output_token_ids` as previous output tokens are now part of
# `prompt_token_ids`.
req_state.output_token_ids.clear()
if self.uses_mrope:
self._init_mrope_positions(req_state)
return req_state
def _init_mrope_positions(self, req_state: CachedRequestState):
model = self.get_model()
assert supports_mrope(model), "M-RoPE support is not implemented."
mrope_model = cast(SupportsMRoPE, model)
# `prompt_embeds` is a passthrough modality (no grid_thw), models'
# M-RoPE code assumes per-feature grid info, so filter it out. The
# prompt_embeds positions are treated as text positions for M-RoPE.
mrope_features = [
f for f in req_state.mm_features if f.modality != "prompt_embeds"
]
if req_state.prompt_token_ids is not None:
input_tokens = req_state.prompt_token_ids
elif req_state.prompt_embeds is not None:
# For embeddings-only inputs, get_mrope_input_positions only
# needs the sequence length when mm_features is empty (which is
# the case here since prompt_embeds are filtered out above).
seq_len = req_state.prompt_embeds.shape[0]
input_tokens = list(range(seq_len))
else:
raise ValueError(
"M-RoPE requires either prompt_token_ids or prompt_embeds."
)
req_state.mrope_positions, req_state.mrope_position_delta = (
mrope_model.get_mrope_input_positions(
input_tokens,
mrope_features,
)
)
def _init_xdrope_positions(self, req_state: CachedRequestState):
model = self.get_model()
xdrope_model = cast(SupportsXDRoPE, model)
assert req_state.prompt_token_ids is not None, (
"XD-RoPE requires prompt_token_ids to be available."
)
assert supports_xdrope(model), "XD-RoPE support is not implemented."
req_state.xdrope_positions = xdrope_model.get_xdrope_input_positions(
req_state.prompt_token_ids,
req_state.mm_features,
)
def _extract_mm_kwargs(
self,
scheduler_output: "SchedulerOutput",
) -> BatchedTensorInputs:
if not scheduler_output or not self.is_multimodal_raw_input_only_model:
return {}
mm_kwargs = list[tuple[str, MultiModalKwargsItem]]()
for req in scheduler_output.scheduled_new_reqs:
for feature in req.mm_features:
if feature.data is not None:
mm_kwargs.append((feature.modality, feature.data))
# Input all modalities at once
mm_kwargs_combined: BatchedTensorInputs = {}
for _, _, mm_kwargs_batch in group_and_batch_mm_kwargs(
mm_kwargs,
device=self.device,
pin_memory=PIN_MEMORY,
):
mm_kwargs_combined.update(mm_kwargs_batch)
return mm_kwargs_combined
def _dummy_mm_kwargs(self, num_seqs: int) -> BatchedTensorInputs:
if not self.is_multimodal_raw_input_only_model:
return {}
mm_budget = self.mm_budget
assert mm_budget is not None
if not mm_budget.mm_max_toks_per_item:
return {} # No tower modalities (embed-only mode)
dummy_modality = mm_budget.get_modality_with_max_tokens()
return self._get_mm_dummy_batch(dummy_modality, num_seqs)
def _get_cumsum_and_arange(
self,
num_tokens: np.ndarray,
arange_out: np.ndarray,
cumsum_dtype: np.dtype | None = None,
) -> np.ndarray:
"""Get the cumulative sum and batched arange of the given array.
E.g., [2, 5, 3] -> [2, 7, 10], arange written to
arange_out[:10] as [0, 1, 0, 1, 2, 3, 4, 0, 1, 2].
Equivalent to but faster than:
np.concatenate([np.arange(n) for n in num_tokens])
"""
# Step 1. [2, 5, 3] -> [2, 7, 10]
cu_num_tokens = np.cumsum(num_tokens, dtype=cumsum_dtype)
total_num_tokens = cu_num_tokens[-1]
# Step 2. [2, 7, 10] -> [0, 0, 2, 2, 2, 2, 2, 7, 7, 7]
cumsums_offsets = np.repeat(cu_num_tokens - num_tokens, num_tokens)
# Step 3. [0, 1, 0, 1, 2, 3, 4, 0, 1, 2]
np.subtract(
self.arange_np[:total_num_tokens],
cumsums_offsets,
out=arange_out[:total_num_tokens],
)
return cu_num_tokens
def _compute_prev_positions(self, num_reqs: int) -> None:
"""Build prev_positions mapping: current pos -> previous pos (-1 if new).
Populates self.prev_positions.np[:num_reqs] with the mapping.
"""
prev_req_id_to_index = self.input_batch.prev_req_id_to_index
prev_positions = self.prev_positions.np[:num_reqs]
if not prev_req_id_to_index:
prev_positions.fill(-1)
return
for i, req_id in enumerate(self.input_batch.req_ids[:num_reqs]):
prev_positions[i] = prev_req_id_to_index.get(req_id, -1)
def _prepare_input_ids(
self,
scheduler_output: "SchedulerOutput",
num_reqs: int,
total_num_scheduled_tokens: int,
cu_num_tokens: np.ndarray,
) -> None:
"""Prepare the input IDs for the current batch.
Carefully handles the `prev_sampled_token_ids` which can be cached
from the previous engine iteration, in which case those tokens on the
GPU need to be copied into the corresponding slots into input_ids.
Uses self.prev_positions[:num_reqs] which maps current pos -> prev pos
(-1 for new requests).
"""
if self.input_batch.prev_sampled_token_ids is None:
# Normal scheduling case
self.input_ids.copy_to_gpu(total_num_scheduled_tokens)
if self.enable_prompt_embeds:
self.inputs_embeds.copy_to_gpu(total_num_scheduled_tokens)
self.is_token_ids.copy_to_gpu(total_num_scheduled_tokens)
return
# Async scheduling case, where some decode requests from the previous
# iteration won't have entries in input_ids_cpu and need to be copied
# on the GPU from prev_sampled_token_ids.
prev_positions = self.prev_positions.np[:num_reqs]
scheduled_spec_tokens = scheduler_output.scheduled_spec_decode_tokens
sample_flattened_indices: list[int] = []
spec_flattened_indices: list[int] = []
prev_draft_token_indices: list[int] = []
prev_indices: list[int] = []
common_indices_match = True
max_flattened_index = -1
total_num_spec_tokens = 0
for cur_index in range(num_reqs):
prev_index = prev_positions[cur_index]
if prev_index < 0:
continue
prev_indices.append(prev_index)
req_id = self.input_batch.req_ids[cur_index]
# We need to compute the flattened input_ids index of the
# last token in each common request.
draft_len = len(scheduled_spec_tokens.get(req_id, ()))
total_num_spec_tokens += draft_len
flattened_index = cu_num_tokens[cur_index].item() - 1
# example: cu_num_tokens = [2, 5, 8], draft_tokens = [1, 2, 2]
# sample_flattened_indices = [0, 2, 5]
# spec_flattened_indices = [1, 3, 4, 6, 7]
sample_flattened_indices.append(flattened_index - draft_len)
spec_flattened_indices.extend(
range(flattened_index - draft_len + 1, flattened_index + 1)
)
start = prev_index * self.prev_num_spec_tokens
# prev_draft_token_indices is used to find which draft_tokens_id
# should be copied to input_ids
# example: prev draft_tokens_id [[1,2], [3,4], [5, 6]]
# flatten draft_tokens_id [1,2,3,4,5,6]
# draft_len of each request [1, 2, 1]
# then prev_draft_token_indices is [0, 2, 3, 4]
prev_draft_token_indices.extend(range(start, start + draft_len))
common_indices_match &= prev_index == flattened_index
max_flattened_index = max(max_flattened_index, flattened_index)
num_common_tokens = len(sample_flattened_indices)
total_without_spec = total_num_scheduled_tokens - total_num_spec_tokens
if self.enable_prompt_embeds:
# The multimodal embed path reads is_token_ids.gpu; its .cpu copy is
# refreshed every step but the async fast paths below only scatter
# input_ids.gpu, so refresh is_token_ids.gpu here too.
self.is_token_ids.copy_to_gpu(total_num_scheduled_tokens)
if num_common_tokens < total_without_spec:
# If not all requests are decodes from the last iteration,
# we need to copy the input_ids_cpu to the GPU first.
self.input_ids.copy_to_gpu(total_num_scheduled_tokens)
if self.enable_prompt_embeds:
self.inputs_embeds.copy_to_gpu(total_num_scheduled_tokens)
if num_common_tokens == 0:
# No requests in common with the previous iteration
# So input_ids.cpu will have all the input ids.
return
if common_indices_match and max_flattened_index == (num_common_tokens - 1):
# Common-case optimization: the batch is unchanged
# and no reordering happened.
# The indices are both the same permutation of 0..N-1 so
# we can copy directly using a single slice.
self.input_ids.gpu[:num_common_tokens].copy_(
self.input_batch.prev_sampled_token_ids[:num_common_tokens, 0],
non_blocking=True,
)
return
# Upload the index tensors asynchronously so the scatter can be non-blocking.
sampled_tokens_index_tensor = torch.tensor(
sample_flattened_indices, dtype=torch.int64, pin_memory=PIN_MEMORY
).to(self.device, non_blocking=True)
prev_common_req_indices_tensor = torch.tensor(
prev_indices, dtype=torch.int64, pin_memory=PIN_MEMORY
).to(self.device, non_blocking=True)
self.input_ids.gpu.scatter_(
dim=0,
index=sampled_tokens_index_tensor,
src=self.input_batch.prev_sampled_token_ids[
prev_common_req_indices_tensor, 0
],
)
# Scatter the draft tokens after the sampled tokens are scattered.
if self._draft_token_ids is None or not spec_flattened_indices:
return
assert isinstance(self._draft_token_ids, torch.Tensor)
draft_tokens_index_tensor = torch.tensor(
spec_flattened_indices, dtype=torch.int64, pin_memory=PIN_MEMORY
).to(self.device, non_blocking=True)
prev_draft_token_indices_tensor = torch.tensor(
prev_draft_token_indices, dtype=torch.int64, pin_memory=PIN_MEMORY
).to(self.device, non_blocking=True)
# because input_ids dtype is torch.int32,
# so convert draft_token_ids to torch.int32 here.
draft_token_ids = self._draft_token_ids.to(dtype=torch.int32)
self.input_ids.gpu.scatter_(
dim=0,
index=draft_tokens_index_tensor,
src=draft_token_ids.flatten()[prev_draft_token_indices_tensor],
)
def _get_encoder_seq_lens(
self,
num_scheduled_tokens: dict[str, int],
kv_cache_spec: KVCacheSpec,
num_reqs: int,
for_cudagraph_capture: bool = False,
) -> tuple[torch.Tensor | None, np.ndarray | None]:
if not isinstance(kv_cache_spec, CrossAttentionSpec):
return None, None
# Zero out buffer for padding requests that are not actually scheduled (CGs)
self.encoder_seq_lens.np[:num_reqs] = 0
# Build encoder_seq_lens array mapping request indices to
# encoder lengths for inputs scheduled in this batch
for req_id in num_scheduled_tokens:
req_index = self.input_batch.req_id_to_index[req_id]
req_state = self.requests[req_id]
if req_state.mm_features is None:
self.encoder_seq_lens.np[req_index] = 0
continue
# Get the total number of encoder input tokens for running encoder requests
# whether encoding is finished or not so that cross-attention knows how
# many encoder tokens to attend to.
encoder_input_tokens = sum(
feature.mm_position.length for feature in req_state.mm_features
)
self.encoder_seq_lens.np[req_index] = encoder_input_tokens
if for_cudagraph_capture:
# During CUDA graph capture, we need to use realistic encoder lengths
# so that max_seqlen_k is captured with the correct value.
max_encoder_len = getattr(
self.model_config.hf_config,
"max_source_positions",
self.max_encoder_len,
)
self.encoder_seq_lens.np[:num_reqs] = max_encoder_len
self.encoder_seq_lens.copy_to_gpu(num_reqs)
encoder_seq_lens = self.encoder_seq_lens.gpu[:num_reqs]
encoder_seq_lens_cpu = self.encoder_seq_lens.np[:num_reqs]
return encoder_seq_lens, encoder_seq_lens_cpu
def _prepare_inputs(
self,
scheduler_output: "SchedulerOutput",
num_scheduled_tokens: np.ndarray,
) -> tuple[
torch.Tensor,
SpecDecodeMetadata | None,
]:
"""
Returns:
tuple[logits_indices, spec_decode_metadata]
"""
total_num_scheduled_tokens = scheduler_output.total_num_scheduled_tokens
assert total_num_scheduled_tokens > 0
num_reqs = self.input_batch.num_reqs
assert num_reqs > 0
# OPTIMIZATION: Start copying the block table first.
# This way, we can overlap the copy with the following CPU operations.
self.input_batch.block_table.commit_block_table(num_reqs)
# Get request indices.
# E.g., [2, 5, 3] -> [0, 0, 1, 1, 1, 1, 1, 2, 2, 2]
req_indices = np.repeat(self.arange_np[:num_reqs], num_scheduled_tokens)
# cu_num_tokens: [2, 5, 3] -> [2, 7, 10]
# self.query_pos.np[:10]: [0, 1, 0, 1, 2, 3, 4, 0, 1, 2]
cu_num_tokens = self._get_cumsum_and_arange(
num_scheduled_tokens, self.query_pos.np
)
# Get positions.
positions_np = (
self.input_batch.num_computed_tokens_cpu[req_indices]
+ self.query_pos.np[: cu_num_tokens[-1]]
)
# Calculate M-RoPE positions.
# Only relevant for models using M-RoPE (e.g, Qwen2-VL)
if self.uses_mrope:
self._calc_mrope_positions(scheduler_output)
# Calculate XD-RoPE positions.
# Only relevant for models using XD-RoPE (e.g, HunYuan-VL)
if self.uses_xdrope_dim > 0:
self._calc_xdrope_positions(scheduler_output)
# Get token indices.
# E.g., [0, 1, 0, 1, 2, 3, 4, 0, 1, 2]
# -> [0, 1, M, M + 1, M + 2, M + 3, M + 4, 2 * M, 2 * M + 1, 2 * M + 2]
# where M is the max_model_len.
token_indices = (
positions_np + req_indices * self.input_batch.token_ids_cpu.shape[1]
)
token_indices_tensor = torch.from_numpy(token_indices)
# NOTE(woosuk): We use torch.index_select instead of np.take here
# because torch.index_select is much faster than np.take for large
# tensors.
torch.index_select(
self.input_batch.token_ids_cpu_tensor.flatten(),
0,
token_indices_tensor,
out=self.input_ids.cpu[:total_num_scheduled_tokens],
)
if self.enable_prompt_embeds:
is_token_ids = self.input_batch.is_token_ids_tensor.flatten()
torch.index_select(
is_token_ids,
0,
token_indices_tensor,
out=self.is_token_ids.cpu[:total_num_scheduled_tokens],
)
# Because we did not pre-allocate a massive prompt_embeds CPU tensor on
# the InputBatch, we need to fill in the prompt embeds into the expected
# spots in the GpuModelRunner's pre-allocated prompt_embeds tensor.
if self.input_batch.req_prompt_embeds:
output_idx = 0
for req_idx in range(num_reqs):
num_sched = num_scheduled_tokens[req_idx]
# Skip if this request doesn't have embeddings
if req_idx not in self.input_batch.req_prompt_embeds:
output_idx += num_sched
continue
# Skip if no tokens scheduled
if num_sched <= 0:
output_idx += num_sched
continue
req_embeds = self.input_batch.req_prompt_embeds[req_idx]
start_pos = self.input_batch.num_computed_tokens_cpu[req_idx]
# Skip if trying to read beyond available embeddings
if start_pos >= req_embeds.shape[0]:
output_idx += num_sched
continue
# Copy available embeddings
end_pos = start_pos + num_sched
actual_end = min(end_pos, req_embeds.shape[0])
actual_num_sched = actual_end - start_pos
if actual_num_sched > 0:
self.inputs_embeds.cpu[
output_idx : output_idx + actual_num_sched
].copy_(req_embeds[start_pos:actual_end])
output_idx += num_sched
# Prepare the attention metadata.
self.query_start_loc.np[0] = 0
self.query_start_loc.np[1 : num_reqs + 1] = cu_num_tokens
# Note: pad query_start_loc to be non-decreasing, as kernels
# like FlashAttention requires that
self.query_start_loc.np[num_reqs + 1 :].fill(cu_num_tokens[-1])
self.query_start_loc.copy_to_gpu()
query_start_loc = self.query_start_loc.gpu[: num_reqs + 1]
# Compute optimistic seq_lens (assumes all draft tokens from previous
# iteration accepted). Store in optimistic_seq_lens_cpu for use by
# _build_attention_metadata (max_seq_len) and discard_request_mask.
# seq_lens (GPU) will be computed later using the same optimistic values.
torch.add(
self.input_batch.num_computed_tokens_cpu_tensor[:num_reqs],
torch.from_numpy(num_scheduled_tokens),
out=self.optimistic_seq_lens_cpu[:num_reqs],
)
self.optimistic_seq_lens_cpu[num_reqs:].fill_(0)
# Build prev_positions mapping: current pos -> prev pos (-1 if new).
# Used for gathering from previous iteration's GPU tensors.
prev_req_id_to_index = self.input_batch.prev_req_id_to_index
self._compute_prev_positions(num_reqs)
num_tokens = [self.requests[r].num_tokens for r in self.input_batch.req_ids]
num_tokens_np = np.array(num_tokens, dtype=np.int32)
# Record which requests should not be sampled,
# so that we could clear the sampled tokens before returning
self.discard_request_mask.np[:num_reqs] = (
self.optimistic_seq_lens_cpu[:num_reqs].numpy() < num_tokens_np
)
self.discard_request_mask.copy_to_gpu(num_reqs)
# Sync num_accepted_tokens from CPU (set by
# _update_states_after_model_execute for hybrid models).
# Skipped under async scheduling (non-align): the CPU copy races with
# the in-flight D2H copy and with input-batch row moves.
needs_cpu_accepted_counts = self.num_accepted_tokens_event is not None and not (
self.use_async_scheduling and self.cache_config.mamba_cache_mode != "align"
)
if needs_cpu_accepted_counts:
assert self.num_accepted_tokens_event is not None
self.num_accepted_tokens_event.synchronize()
# Async mode: condense() reordered indices, use prev_positions mapping
if self.use_async_scheduling and prev_req_id_to_index:
prev_idx = self.prev_positions.np[:num_reqs]
new_mask = prev_idx < 0
self.num_accepted_tokens.np[:num_reqs] = (
self.input_batch.num_accepted_tokens_cpu[
np.where(new_mask, 0, prev_idx)
]
)
self.num_accepted_tokens.np[:num_reqs][new_mask] = 1
self.input_batch.num_accepted_tokens_cpu[:num_reqs] = (
self.num_accepted_tokens.np[:num_reqs]
)
else:
# Non-async mode: use values directly
self.num_accepted_tokens.np[:num_reqs] = (
self.input_batch.num_accepted_tokens_cpu[:num_reqs]
)
self.num_accepted_tokens.np[num_reqs:].fill(1)
self.num_accepted_tokens.copy_to_gpu()
else:
# Default to 1; update_num_computed_tokens_for_batch_change below
# corrects rows that had drafts from valid_sampled_token_count.
self.num_accepted_tokens.np.fill(1)
self.num_accepted_tokens.gpu.fill_(1)
if self.mamba_prev_last_scheduled_idx is not None:
mamba_utils.preprocess_mamba_all_specdec(
scheduler_output,
self.input_batch,
self.mamba_state_idx,
num_reqs,
self.mamba_prev_last_scheduled_idx,
)
# Update num_computed_tokens on GPU. In async spec decode,
# CPU values are optimistic (all drafts accepted). The kernel
# corrects on GPU using the previous step's
# valid_sampled_token_count_gpu. Otherwise, just copy from CPU.
if (
self.use_async_spec_decode
and self.valid_sampled_token_count_gpu is not None
and prev_req_id_to_index
):
self.prev_positions.copy_to_gpu(num_reqs)
self.prev_num_draft_tokens.copy_to_gpu()
cpu_values = self.input_batch.num_computed_tokens_cpu_tensor[:num_reqs].to(
device=self.device, non_blocking=True
)
update_num_computed_tokens_for_batch_change(
self.num_computed_tokens,
self.num_accepted_tokens.gpu[:num_reqs],
self.prev_positions.gpu[:num_reqs],
self.valid_sampled_token_count_gpu,
self.prev_num_draft_tokens.gpu,
cpu_values,
)
else:
self.num_computed_tokens[:num_reqs].copy_(
self.input_batch.num_computed_tokens_cpu_tensor[:num_reqs],
non_blocking=True,
)
self.req_indices.np[:total_num_scheduled_tokens] = req_indices
self.req_indices.copy_to_gpu(total_num_scheduled_tokens)
req_indices_gpu = self.req_indices.gpu[:total_num_scheduled_tokens]
self.query_pos.copy_to_gpu(total_num_scheduled_tokens)
self.num_scheduled_tokens.np[:num_reqs] = num_scheduled_tokens
self.num_scheduled_tokens.copy_to_gpu(num_reqs)
num_scheduled_tokens_gpu = self.num_scheduled_tokens.gpu[:num_reqs]
self.positions[:total_num_scheduled_tokens] = (
self.num_computed_tokens[req_indices_gpu].to(torch.int64)
+ self.query_pos.gpu[:total_num_scheduled_tokens]
)
self.seq_lens[:num_reqs] = (
self.num_computed_tokens[:num_reqs] + num_scheduled_tokens_gpu
)
self.seq_lens[num_reqs:].fill_(0)
self.input_batch.block_table.compute_slot_mapping(
num_reqs,
self.query_start_loc.gpu[: num_reqs + 1],
self.positions[:total_num_scheduled_tokens],
)
# Copy the tensors to the GPU.
self._prepare_input_ids(
scheduler_output,
num_reqs,
total_num_scheduled_tokens,
cu_num_tokens,
)
if self.uses_mrope:
# Only relevant for models using M-RoPE (e.g, Qwen2-VL)
self.mrope_positions.gpu[:, :total_num_scheduled_tokens].copy_(
self.mrope_positions.cpu[:, :total_num_scheduled_tokens],
non_blocking=True,
)
elif self.uses_xdrope_dim > 0:
# Only relevant for models using XD-RoPE (e.g, HunYuan-VL)
self.xdrope_positions.gpu[:, :total_num_scheduled_tokens].copy_(
self.xdrope_positions.cpu[:, :total_num_scheduled_tokens],
non_blocking=True,
)
if self.use_async_spec_decode and (self.uses_mrope or self.uses_xdrope_dim > 0):
drift = self.num_computed_tokens[req_indices_gpu].to(
torch.int64
) - self.input_batch.num_computed_tokens_cpu_tensor[req_indices].to(
device=self.device, dtype=torch.int64, non_blocking=True
)
target = self.mrope_positions if self.uses_mrope else self.xdrope_positions
target.gpu[:, :total_num_scheduled_tokens] += drift
use_spec_decode = len(scheduler_output.scheduled_spec_decode_tokens) > 0
if not use_spec_decode:
# NOTE(woosuk): Due to chunked prefills, the batch may contain
# partial requests. While we should not sample any token
# from these partial requests, we do so for simplicity.
# We will ignore the sampled tokens from the partial requests.
# TODO: Support prompt logprobs.
logits_indices = query_start_loc[1:] - 1
spec_decode_metadata = None
num_sampled_tokens = np.ones(num_reqs, dtype=np.int32)
else:
# Get the number of draft tokens for each request.
# Iterate over the dictionary rather than all requests since not all
# requests have draft tokens.
num_draft_tokens = np.zeros(num_reqs, dtype=np.int32)
# For chunked prefills, use -1 as mask rather than 0, as guided
# decoding may rollback speculative tokens.
num_decode_draft_tokens = np.full(num_reqs, -1, dtype=np.int32)
for (
req_id,
draft_token_ids,
) in scheduler_output.scheduled_spec_decode_tokens.items():
req_idx = self.input_batch.req_id_to_index[req_id]
draft_len = len(draft_token_ids)
num_draft_tokens[req_idx] = draft_len
if num_scheduled_tokens[req_idx] == draft_len + 1:
num_decode_draft_tokens[req_idx] = draft_len
spec_decode_metadata = self._calc_spec_decode_metadata(
num_draft_tokens, cu_num_tokens
)
logits_indices = spec_decode_metadata.logits_indices
num_sampled_tokens = num_draft_tokens + 1
# For DECODE only cuda graph of some attention backends (e.g., GDN).
self.num_decode_draft_tokens.np[:num_reqs] = num_decode_draft_tokens
self.num_decode_draft_tokens.np[num_reqs:].fill(-1)
self.num_decode_draft_tokens.copy_to_gpu()
# Hot-Swap lora model
if self.lora_config:
assert (
np.sum(num_sampled_tokens)
<= self.vllm_config.scheduler_config.max_num_batched_tokens
)
self.set_active_loras(
self.input_batch, num_scheduled_tokens, num_sampled_tokens
)
return (
logits_indices,
spec_decode_metadata,
)
def _build_attention_metadata(
self,
num_tokens: int,
num_reqs: int,
max_query_len: int,
num_tokens_padded: int | None = None,
num_reqs_padded: int | None = None,
ubatch_slices: UBatchSlices | None = None,
logits_indices: torch.Tensor | None = None,
use_spec_decode: bool = False,
for_cudagraph_capture: bool = False,
num_scheduled_tokens: dict[str, int] | None = None,
cascade_attn_prefix_lens: list[list[int]] | None = None,
slot_mappings: dict[int, torch.Tensor] | None = None,
) -> tuple[PerLayerAttnMetadata, CommonAttentionMetadata | None]:
"""
Returns:
tuple[attn_metadata, spec_decode_common_attn_metadata]
"""
# Attention metadata is not needed for attention free models
if len(self.kv_cache_config.kv_cache_groups) == 0:
return {}, None
num_tokens_padded = num_tokens_padded or num_tokens
num_reqs_padded = num_reqs_padded or num_reqs
assert num_reqs_padded is not None and num_tokens_padded is not None
attn_metadata: PerLayerAttnMetadata = {}
if ubatch_slices is not None:
attn_metadata = [dict() for _ in range(len(ubatch_slices))]
if for_cudagraph_capture:
# For some attention backends (e.g. FA) with sliding window models we need
# to make sure the backend see a max_seq_len that is larger to the sliding
# window size when capturing to make sure the correct kernel is selected.
max_seq_len = self.max_model_len
else:
max_seq_len = self.optimistic_seq_lens_cpu.numpy()[:num_reqs].max().item()
kv_cache_groups = self.kv_cache_config.kv_cache_groups
def _get_block_table(kv_cache_gid: int):
assert num_reqs_padded is not None and num_tokens_padded is not None
kv_cache_spec = kv_cache_groups[kv_cache_gid].kv_cache_spec
if isinstance(kv_cache_spec, EncoderOnlyAttentionSpec):
blk_table_tensor = torch.zeros(
(num_reqs_padded, 1),
dtype=torch.int32,
device=self.device,
)
else:
blk_table = self.input_batch.block_table[kv_cache_gid]
blk_table_tensor = blk_table.get_device_tensor(num_reqs_padded)
# Fill unused block table entries with NULL_BLOCK_ID (null block)
# for CUDAGraph padding. Block 0 is reserved for padding.
blk_table_tensor[num_reqs:num_reqs_padded].fill_(NULL_BLOCK_ID)
return blk_table_tensor
assert slot_mappings is not None
block_table_gid_0 = _get_block_table(0)
slot_mapping_gid_0 = slot_mappings[0]
if self.routed_experts_initialized:
# Copy this step's attention slot_mapping into our private
# device buffer. The shared ``slot_mappings[attn_gid]`` is
# owned by the attention block table and will be overwritten
# by the next ``_prepare_inputs``; we need a stable snapshot
# because the async D2H may still be in flight on the copy
# stream when the next step runs.
attn_gid = self.routed_experts_attn_gid
slot_mapping_attn = slot_mappings[attn_gid]
self.routed_experts_slot_mapping_device[:num_tokens].copy_(
slot_mapping_attn[:num_tokens]
)
num_computed_tokens_cpu = self.input_batch.num_computed_tokens_cpu_tensor[
:num_reqs_padded
]
num_prompt_tokens_cpu = self.input_batch.num_prompt_tokens_cpu_tensor[
:num_reqs_padded
]
seq_lens_cpu = self.optimistic_seq_lens_cpu[:num_reqs_padded]
seq_lens_cpu_upper_bound = seq_lens_cpu
# is_prefilling: True if request is still in prefill phase.
# Used by mamba backends to distinguish actual decodes from
# short extends.
is_prefilling = num_computed_tokens_cpu < num_prompt_tokens_cpu
# Zero out padded rows so stale data from condense() doesn't
# misclassify padding as prefill in CUDA graph mode.
is_prefilling[num_reqs:] = False
if self.use_async_spec_decode:
# GPU tensors are authoritative in async mode.
seq_lens_cpu = None
num_computed_tokens_cpu = None
# Compute mm_prefix bidirectional ranges before building
# attention metadata so builders handle them during build().
# By default, ranges exceeding sliding_window are skipped to prevent
# early tokens from attending across the entire image span. Models that
# clamp mm_prefix to the sliding window *in-kernel* (e.g. Gemma4, which
# needs HF's (causal OR blockwise) AND sliding_window on sliding layers)
# opt out of the skip so the bidirectional range survives for images
# larger than the window; the kernel then bounds it per-query.
req_doc_ranges: dict[int, list[tuple[int, int]]] | None = None
if self.is_mm_prefix_lm:
req_doc_ranges = {}
hf_text_config = self.model_config.hf_text_config
_bidi_sw = getattr(hf_text_config, "sliding_window", None)
_clamps_in_kernel = getattr(
self.model, "mm_prefix_clamp_sliding_window", False
)
for req_id in self.input_batch.req_ids:
image_doc_ranges = []
req_state = self.requests[req_id]
for mm_feature in req_state.mm_features:
if mm_feature.modality == "audio":
continue
pos_info = mm_feature.mm_position
img_doc_range = pos_info.extract_embeds_range()
for r in img_doc_range:
if (
not _clamps_in_kernel
and _bidi_sw is not None
and (r[1] - r[0] + 1) > _bidi_sw
):
continue
image_doc_ranges.append(r)
req_idx = self.input_batch.req_id_to_index[req_id]
req_doc_ranges[req_idx] = image_doc_ranges
# Reference Sliding Window Attention (R-SWA): pass per-request prompt
# lengths so the attention backend can keep the prefix globally visible.
# The backend owns the persistent CUDA-graph-safe GPU buffer.
rswa_prefix_lens = None
if self.model_config.rswa_window is not None:
rswa_prefix_lens = num_prompt_tokens_cpu
cm_base = CommonAttentionMetadata(
query_start_loc=self.query_start_loc.gpu[: num_reqs_padded + 1],
query_start_loc_cpu=self.query_start_loc.cpu[: num_reqs_padded + 1],
seq_lens=self.seq_lens[:num_reqs_padded],
_seq_lens_cpu=seq_lens_cpu,
_num_computed_tokens_cpu=num_computed_tokens_cpu,
seq_lens_cpu_upper_bound=seq_lens_cpu_upper_bound,
num_reqs=num_reqs_padded,
num_actual_tokens=num_tokens_padded,
max_query_len=max_query_len,
max_seq_len=max_seq_len,
block_table_tensor=block_table_gid_0,
slot_mapping=slot_mapping_gid_0,
causal=True,
is_prefilling=is_prefilling,
positions=self.positions[:num_tokens_padded],
mm_req_doc_ranges=req_doc_ranges,
rswa_prefix_lens=rswa_prefix_lens,
)
if self.dcp_world_size > 1:
self.dcp_local_seq_lens.cpu[:num_reqs] = get_dcp_local_seq_lens(
self.optimistic_seq_lens_cpu[:num_reqs],
self.dcp_world_size,
self.dcp_rank,
self.parallel_config.cp_kv_cache_interleave_size,
)
self.dcp_local_seq_lens.cpu[num_reqs:].fill_(0)
self.dcp_local_seq_lens.copy_to_gpu(num_reqs_padded)
cm_base.dcp_local_seq_lens = self.dcp_local_seq_lens.gpu[:num_reqs_padded]
cm_base.dcp_local_seq_lens_cpu = self.dcp_local_seq_lens.cpu[
:num_reqs_padded
]
if logits_indices is not None and self.cache_config.kv_sharing_fast_prefill:
cm_base.num_logits_indices = logits_indices.size(0)
cm_base.logits_indices_padded = self._prepare_kv_sharing_fast_prefill(
logits_indices
)
# Cache attention metadata builds across hybrid KV-cache groups
# The only thing that changes between different hybrid KV-cache groups when the
# same metadata builder and KVCacheSpec is the same is the block table, so we
# can cache the attention metadata builds and just update the block table using
# `builder.update_block_table` if the builder supports it.
cached_attn_metadata: dict[
tuple[KVCacheSpec, type[AttentionMetadataBuilder]], AttentionMetadata
] = {}
def _build_attn_group_metadata(
kv_cache_gid: int,
attn_gid: int,
common_attn_metadata: CommonAttentionMetadata,
ubid: int | None = None,
) -> None:
attn_group = self.attn_groups[kv_cache_gid][attn_gid]
builder = attn_group.get_metadata_builder(ubid or 0)
kv_cache_spec = kv_cache_groups[kv_cache_gid].kv_cache_spec
if isinstance(kv_cache_spec, UniformTypeKVCacheSpecs):
kv_cache_spec = kv_cache_spec.kv_cache_specs[attn_group.layer_names[0]]
cache_key = (kv_cache_spec, type(builder))
cascade_attn_prefix_len = (
cascade_attn_prefix_lens[kv_cache_gid][attn_gid]
if cascade_attn_prefix_lens
else 0
)
extra_attn_metadata_args = {}
if use_spec_decode and isinstance(
builder,
(
Mamba2AttentionMetadataBuilder,
GDNAttentionMetadataBuilder,
BailingLinearAttentionMetadataBuilder,
),
):
assert ubid is None, (
"UBatching not supported with GDN or linear attn yet"
)
extra_attn_metadata_args = dict(
num_accepted_tokens=self.num_accepted_tokens.gpu[:num_reqs_padded],
num_decode_draft_tokens_cpu=self.num_decode_draft_tokens.cpu[
:num_reqs_padded
],
)
if (
isinstance(builder, Mamba2AttentionMetadataBuilder)
and self.mamba_prev_last_scheduled_idx is not None
):
extra_attn_metadata_args["prev_last_scheduled_idx"] = (
self.mamba_prev_last_scheduled_idx.gpu[:num_reqs_padded]
)
if for_cudagraph_capture:
attn_metadata_i = builder.build_for_cudagraph_capture(
common_attn_metadata
)
elif (
cache_key in cached_attn_metadata
and builder.supports_update_block_table
):
attn_metadata_i = builder.update_block_table(
cached_attn_metadata[cache_key],
common_attn_metadata.block_table_tensor,
common_attn_metadata.slot_mapping,
)
else:
attn_metadata_i = builder.build(
common_prefix_len=cascade_attn_prefix_len,
common_attn_metadata=common_attn_metadata,
**extra_attn_metadata_args,
)
if builder.supports_update_block_table:
cached_attn_metadata[cache_key] = attn_metadata_i
if ubid is None:
assert isinstance(attn_metadata, dict)
attn_metadata_dict = attn_metadata
else:
assert isinstance(attn_metadata, list)
attn_metadata_dict = attn_metadata[ubid]
for layer_name in attn_group.layer_names:
attn_metadata_dict[layer_name] = attn_metadata_i
# Prepare the attention metadata for each KV cache group and make layers
# in the same group share the same metadata.
spec_decode_common_attn_metadata = None
for kv_cache_gid, kv_cache_group in enumerate(kv_cache_groups):
cm = copy(cm_base) # shallow copy
# Basically only the encoder seq_lens, block_table and slot_mapping change
# for each kv_cache_group.
cm.encoder_seq_lens, cm.encoder_seq_lens_cpu = self._get_encoder_seq_lens(
num_scheduled_tokens or {},
kv_cache_group.kv_cache_spec,
num_reqs_padded,
for_cudagraph_capture=for_cudagraph_capture,
)
if kv_cache_gid > 0:
cm.block_table_tensor = _get_block_table(kv_cache_gid)
cm.slot_mapping = slot_mappings[kv_cache_gid]
if self.speculative_config and spec_decode_common_attn_metadata is None:
if isinstance(
self.drafter,
(
EagleProposer,
DFlashProposer,
Gemma4Proposer,
ExtractHiddenStatesProposer,
),
):
if self.drafter.kv_cache_gid == kv_cache_gid:
spec_decode_common_attn_metadata = cm
else:
spec_decode_common_attn_metadata = cm
# Capture per-group block tables for multi-group proposers.
if self.speculative_config and isinstance(self.drafter, Step3p5MTPProposer):
self.drafter.set_per_group_attn_metadata(
kv_cache_gid, cm.block_table_tensor, cm.slot_mapping
)
elif self.speculative_config and isinstance(self.drafter, Gemma4Proposer):
self.drafter.set_per_group_block_table(
kv_cache_gid, cm.block_table_tensor
)
for attn_gid in range(len(self.attn_groups[kv_cache_gid])):
if ubatch_slices is not None:
for ubid, _cm in enumerate(split_attn_metadata(ubatch_slices, cm)):
_build_attn_group_metadata(kv_cache_gid, attn_gid, _cm, ubid)
else:
_build_attn_group_metadata(kv_cache_gid, attn_gid, cm)
if spec_decode_common_attn_metadata is not None and (
num_reqs != num_reqs_padded or num_tokens != num_tokens_padded
):
# Currently the drafter still only uses piecewise cudagraphs (and modifies
# the attention metadata in directly), and therefore does not want to use
# padded attention metadata.
spec_decode_common_attn_metadata = (
spec_decode_common_attn_metadata.unpadded(num_tokens, num_reqs)
)
return attn_metadata, spec_decode_common_attn_metadata
def _compute_cascade_attn_prefix_lens(
self,
num_scheduled_tokens: np.ndarray,
num_computed_tokens: np.ndarray,
num_common_prefix_blocks: list[int],
) -> list[list[int]] | None:
"""
Returns:
Optional[cascade_attn_prefix_lens]
cascade_attn_prefix_lens is 2D:
``[kv_cache_group_id][attn_group_idx]``,
None if we should not use cascade attention
"""
use_cascade_attn = False
num_kv_cache_groups = len(self.kv_cache_config.kv_cache_groups)
cascade_attn_prefix_lens: list[list[int]] = [
[] for _ in range(num_kv_cache_groups)
]
for kv_cache_gid in range(num_kv_cache_groups):
for attn_group in self.attn_groups[kv_cache_gid]:
if isinstance(attn_group.kv_cache_spec, EncoderOnlyAttentionSpec):
cascade_attn_prefix_len = 0
else:
# 0 if cascade attention should not be used
cascade_attn_prefix_len = self._compute_cascade_attn_prefix_len(
num_scheduled_tokens,
num_computed_tokens,
num_common_prefix_blocks[kv_cache_gid],
attn_group.kv_cache_spec,
attn_group.get_metadata_builder(),
)
cascade_attn_prefix_lens[kv_cache_gid].append(cascade_attn_prefix_len)
use_cascade_attn |= cascade_attn_prefix_len > 0
return cascade_attn_prefix_lens if use_cascade_attn else None
def _compute_cascade_attn_prefix_len(
self,
num_scheduled_tokens: np.ndarray,
num_computed_tokens: np.ndarray,
num_common_prefix_blocks: int,
kv_cache_spec: KVCacheSpec,
attn_metadata_builder: AttentionMetadataBuilder,
) -> int:
"""Compute the length of the common prefix for cascade attention.
NOTE(woosuk): The common prefix length returned by this function
represents the length used specifically for cascade attention, not the
actual number of tokens shared between requests. When cascade attention
is disabled (use_cascade=False), this function returns 0 even if
requests share common tokens. Additionally, the common prefix length is
truncated to a multiple of the block size and may be further truncated
due to implementation details explained below.
Args:
num_scheduled_tokens: Number of tokens scheduled per request.
num_common_prefix_blocks: Number of shared KV cache blocks.
Returns:
int: Length of common prefix in tokens.
"""
common_prefix_len = num_common_prefix_blocks * kv_cache_spec.block_size
if common_prefix_len == 0:
# Common case.
return 0
# NOTE(woosuk): Cascade attention uses two attention kernels: one
# for the common prefix and the other for the rest. For the first
# kernel, we concatenate all the query tokens (possibly from
# different requests) and treat them as if they are from the same
# request. Then, we use bi-directional attention to process the
# common prefix in the KV cache. Importantly, this means that the
# first kernel does not do any masking.
# Consider the following example:
# Request 1's input query: [D, E, X]
# Request 1's kv cache: [A, B, C, D, E, X]
# Request 1's num_computed_tokens: 3 (i.e., [A, B, C])
# Request 2's input query: [E, Y]
# Request 2's kv cache: [A, B, C, D, E, Y]
# Request 2's num_computed_tokens: 4 (i.e., [A, B, C, D])
# If we use [A, B, C, D, E] as the common prefix, then the
# first kernel will compute the bi-directional attention between
# input query [D, E, X, E, Y] and common prefix [A, B, C, D, E].
# However, this is wrong because D in Request 1 should not attend to
# E in the common prefix (i.e., we need masking).
# To avoid this, [A, B, C, D] should be the common prefix.
# That is, the common prefix should be capped by the minimum
# num_computed_tokens among the requests, and plus one to include
# the first token of the query.
# In practice, we use [A, B, C] as the common prefix, instead of
# [A, B, C, D] (i.e., the common prefix is capped by the minimum
# num_computed_tokens, without plus one).
# This is because of an implementation detail: We want to always
# use two kernels for cascade attention. Let's imagine:
# Request 3's input query: [D]
# Request 3's kv cache: [A, B, C, D]
# Request 3's num_computed_tokens: 3 (i.e., [A, B, C])
# If we use [A, B, C, D] as the common prefix for Request 1-3,
# then Request 3 will be processed only by the first kernel,
# and the second kernel will get an empty input. While this is not
# a fundamental problem, our current implementation does not support
# this case.
common_prefix_len = min(common_prefix_len, num_computed_tokens.min())
# common_prefix_len should be a multiple of the block size.
common_prefix_len = (
common_prefix_len // kv_cache_spec.block_size * kv_cache_spec.block_size
)
use_sliding_window = isinstance(kv_cache_spec, SlidingWindowSpec) or (
isinstance(kv_cache_spec, FullAttentionSpec)
and kv_cache_spec.sliding_window is not None
)
use_local_attention = isinstance(kv_cache_spec, ChunkedLocalAttentionSpec) or (
isinstance(kv_cache_spec, FullAttentionSpec)
and kv_cache_spec.attention_chunk_size is not None
)
assert isinstance(kv_cache_spec, AttentionSpec)
use_cascade = attn_metadata_builder.use_cascade_attention(
common_prefix_len=common_prefix_len,
query_lens=num_scheduled_tokens,
num_query_heads=self.num_query_heads,
num_kv_heads=kv_cache_spec.num_kv_heads,
use_alibi=self.use_alibi,
use_sliding_window=use_sliding_window,
use_local_attention=use_local_attention,
num_sms=self.num_sms,
dcp_world_size=self.dcp_world_size,
)
return common_prefix_len if use_cascade else 0
def _calc_mrope_positions(self, scheduler_output: "SchedulerOutput"):
mrope_pos_ptr = 0
for index, req_id in enumerate(self.input_batch.req_ids):
req = self.requests[req_id]
assert req.mrope_positions is not None
num_computed_tokens = self.input_batch.num_computed_tokens_cpu[index]
num_scheduled_tokens = scheduler_output.num_scheduled_tokens[req_id]
num_prompt_tokens = length_from_prompt_token_ids_or_embeds(
req.prompt_token_ids, req.prompt_embeds
)
if num_computed_tokens + num_scheduled_tokens > num_prompt_tokens:
prompt_part_len = max(0, num_prompt_tokens - num_computed_tokens)
completion_part_len = max(0, num_scheduled_tokens - prompt_part_len)
else:
prompt_part_len = num_scheduled_tokens
completion_part_len = 0
assert num_scheduled_tokens == prompt_part_len + completion_part_len
if prompt_part_len > 0:
# prompt's mrope_positions are pre-computed
dst_start = mrope_pos_ptr
dst_end = mrope_pos_ptr + prompt_part_len
src_start = num_computed_tokens
src_end = num_computed_tokens + prompt_part_len
self.mrope_positions.cpu[:, dst_start:dst_end] = req.mrope_positions[
:, src_start:src_end
]
mrope_pos_ptr += prompt_part_len
if completion_part_len > 0:
# compute completion's mrope_positions on-the-fly
dst_start = mrope_pos_ptr
dst_end = mrope_pos_ptr + completion_part_len
assert req.mrope_position_delta is not None
MRotaryEmbedding.get_next_input_positions_tensor(
out=self.mrope_positions.np,
out_offset=dst_start,
mrope_position_delta=req.mrope_position_delta,
context_len=num_computed_tokens + prompt_part_len,
num_new_tokens=completion_part_len,
)
mrope_pos_ptr += completion_part_len
def _calc_xdrope_positions(self, scheduler_output: "SchedulerOutput"):
xdrope_pos_ptr = 0
for index, req_id in enumerate(self.input_batch.req_ids):
req = self.requests[req_id]
assert req.xdrope_positions is not None
num_computed_tokens = self.input_batch.num_computed_tokens_cpu[index]
num_scheduled_tokens = scheduler_output.num_scheduled_tokens[req_id]
num_prompt_tokens = length_from_prompt_token_ids_or_embeds(
req.prompt_token_ids, req.prompt_embeds
)
if num_computed_tokens + num_scheduled_tokens > num_prompt_tokens:
prompt_part_len = max(0, num_prompt_tokens - num_computed_tokens)
completion_part_len = max(0, num_scheduled_tokens - prompt_part_len)
else:
prompt_part_len = num_scheduled_tokens
completion_part_len = 0
assert num_scheduled_tokens == prompt_part_len + completion_part_len
if prompt_part_len > 0:
# prompt's xdrope_positions are pre-computed
dst_start = xdrope_pos_ptr
dst_end = xdrope_pos_ptr + prompt_part_len
src_start = num_computed_tokens
src_end = num_computed_tokens + prompt_part_len
self.xdrope_positions.cpu[:, dst_start:dst_end] = req.xdrope_positions[
:, src_start:src_end
]
xdrope_pos_ptr += prompt_part_len
if completion_part_len > 0:
# compute completion's xdrope_positions on-the-fly
dst_start = xdrope_pos_ptr
dst_end = xdrope_pos_ptr + completion_part_len
XDRotaryEmbedding.get_next_input_positions_tensor(
out=self.xdrope_positions.np,
out_offset=dst_start,
context_len=num_computed_tokens + prompt_part_len,
num_new_tokens=completion_part_len,
)
xdrope_pos_ptr += completion_part_len
def _calc_spec_decode_metadata(
self,
num_draft_tokens: np.ndarray,
cu_num_scheduled_tokens: np.ndarray,
) -> SpecDecodeMetadata:
# Inputs:
# cu_num_scheduled_tokens: [ 4, 104, 107, 207, 209]
# num_draft_tokens: [ 3, 0, 2, 0, 1]
# Outputs:
# cu_num_draft_tokens: [ 3, 3, 5, 5, 6]
# logits_indices: [ 0, 1, 2, 3, 103, 104, 105, 106,
# 206, 207, 208]
# target_logits_indices: [ 0, 1, 2, 5, 6, 9]
# bonus_logits_indices: [ 3, 4, 7, 8, 10]
# Compute the logits indices.
# [4, 1, 3, 1, 2]
num_sampled_tokens = num_draft_tokens + 1
# Step 1.
# cu_num_sampled_tokens: [4, 5, 8, 9, 11]
# _arange_scratch[:11]: [0, 1, 2, 3, 0, 0, 1, 2, 0, 0, 1]
cu_num_sampled_tokens = self._get_cumsum_and_arange(
num_sampled_tokens, self._arange_scratch, cumsum_dtype=np.int32
)
# Step 2. [0, 0, 0, 0, 103, 104, 104, 104, 206, 207, 207]
logits_indices = np.repeat(
cu_num_scheduled_tokens - num_sampled_tokens, num_sampled_tokens
)
# Step 3. [0, 1, 2, 3, 103, 104, 105, 106, 206, 207, 208]
logits_indices += self._arange_scratch[: cu_num_sampled_tokens[-1]]
# Compute the bonus logits indices.
bonus_logits_indices = cu_num_sampled_tokens - 1
# Compute the draft logits indices.
# cu_num_draft_tokens: [3, 3, 5, 5, 6]
# _arange_scratch[:6]: [0, 1, 2, 0, 1, 0]
cu_num_draft_tokens = self._get_cumsum_and_arange(
num_draft_tokens, self._arange_scratch, cumsum_dtype=np.int32
)
# [0, 0, 0, 5, 5, 9]
target_logits_indices = np.repeat(
cu_num_sampled_tokens - num_sampled_tokens, num_draft_tokens
)
# [0, 1, 2, 5, 6, 9]
target_logits_indices += self._arange_scratch[: cu_num_draft_tokens[-1]]
cu_num_draft_tokens = async_tensor_h2d(cu_num_draft_tokens, device=self.device)
cu_num_sampled_tokens = async_tensor_h2d(
cu_num_sampled_tokens, device=self.device
)
logits_indices = async_tensor_h2d(logits_indices, device=self.device)
target_logits_indices = async_tensor_h2d(
target_logits_indices, device=self.device
)
bonus_logits_indices = async_tensor_h2d(
bonus_logits_indices, device=self.device
)
# Compute the draft token ids.
# draft_token_indices: [ 1, 2, 3, 105, 106, 208]
draft_token_ids = self.input_ids.gpu[logits_indices]
draft_token_ids = draft_token_ids[target_logits_indices + 1]
return SpecDecodeMetadata(
draft_token_ids=draft_token_ids,
num_draft_tokens=num_draft_tokens.tolist(),
cu_num_draft_tokens=cu_num_draft_tokens,
cu_num_sampled_tokens=cu_num_sampled_tokens,
target_logits_indices=target_logits_indices,
bonus_logits_indices=bonus_logits_indices,
logits_indices=logits_indices,
)
def _prepare_kv_sharing_fast_prefill(
self,
logits_indices: torch.Tensor,
) -> torch.Tensor:
assert self.kv_sharing_fast_prefill_logits_indices is not None
num_logits = logits_indices.shape[0]
assert num_logits > 0
self.kv_sharing_fast_prefill_logits_indices[:num_logits].copy_(logits_indices)
# There might have leftover indices in logits_indices[num_logits:]
# from previous iterations, whose values may be greater than the
# batch size in the current iteration. To ensure indices are always
# valid, fill the padded indices with the last index. Broadcast the
# scalar GPU-side to avoid a D2H sync on `.item()`.
self.kv_sharing_fast_prefill_logits_indices[num_logits:] = logits_indices[-1]
# Dispatch for the decoder portion of the model.
_, batch_desc = self.cudagraph_dispatcher.dispatch(
num_logits, invalid_modes={CUDAGraphMode.FULL}
)
num_logits_padded = batch_desc.num_tokens
logits_indices_padded = self.kv_sharing_fast_prefill_logits_indices[
:num_logits_padded
]
return logits_indices_padded
def _batch_mm_inputs_from_scheduler(
self,
scheduler_output: "SchedulerOutput",
) -> tuple[
list[str],
list[tuple[str, MultiModalKwargsItem]],
list[tuple[str, PlaceholderRange]],
]:
"""Batch multimodal inputs from scheduled encoder inputs.
Args:
scheduler_output: The scheduler output containing scheduled encoder
inputs.
Returns:
A tuple of (mm_hashes, mm_kwargs, mm_lora_refs) where:
- mm_hashes: List of multimodal hashes for each item
- mm_kwargs: List of multimodal kwargs for each item
- mm_lora_refs: List of (req_id, placeholder_range) for each item
"""
scheduled_encoder_inputs = scheduler_output.scheduled_encoder_inputs
if not scheduled_encoder_inputs:
return [], [], []
mm_hashes = list[str]()
mm_kwargs = list[tuple[str, MultiModalKwargsItem]]()
# Multimodal LoRA reference info to map each multimodal item
# back to its request & position
mm_lora_refs = list[tuple[str, PlaceholderRange]]()
for req_id, encoder_input_ids in scheduled_encoder_inputs.items():
req_state = self.requests[req_id]
for mm_input_id in encoder_input_ids:
mm_feature = req_state.mm_features[mm_input_id]
if mm_feature.data is None:
continue
mm_hashes.append(mm_feature.identifier)
mm_kwargs.append((mm_feature.modality, mm_feature.data))
mm_lora_refs.append((req_id, mm_feature.mm_position))
return mm_hashes, mm_kwargs, mm_lora_refs
def _execute_mm_encoder(
self, scheduler_output: "SchedulerOutput"
) -> list[torch.Tensor]:
mm_hashes, mm_kwargs, mm_lora_refs = self._batch_mm_inputs_from_scheduler(
scheduler_output
)
if not mm_kwargs:
return []
# `prompt_embeds` is a passthrough modality, the tensor is already in
# the model embedding space, so no encoder runs. Inject each
# `prompt_embeds` tensor directly into the encoder cache here so that
# `_gather_mm_embeddings` can splice it via the standard `is_mm_embed`
# path.
pe_indices = [
i
for i, (modality, _) in enumerate(mm_kwargs)
if modality == "prompt_embeds"
]
if pe_indices:
for i in pe_indices:
pe_tensor = mm_kwargs[i][1]["embedding"].data
assert isinstance(pe_tensor, torch.Tensor)
self.encoder_cache[mm_hashes[i]] = pe_tensor.to(self.device)
self.maybe_save_ec_to_connector(self.encoder_cache, mm_hashes[i])
# Filter out `prompt_embeds` items from mm_kwargs/mm_hashes/mm_lora_refs
# since they don't require further encoder processing.
mm_hashes = [h for i, h in enumerate(mm_hashes) if i not in pe_indices]
mm_kwargs = [k for i, k in enumerate(mm_kwargs) if i not in pe_indices]
mm_lora_refs = [
r for i, r in enumerate(mm_lora_refs) if i not in pe_indices
]
if not mm_kwargs:
return [] # nothing left to encode after filtering out `prompt_embeds`
should_time = bool(
self.observability_config
and self.observability_config.enable_mm_processor_stats
and scheduler_output.scheduled_encoder_inputs
)
# Batch mm inputs as much as we can: if a request in the batch has
# multiple modalities or a different modality than the previous one,
# we process it separately to preserve item order.
# FIXME(ywang96): This is a hacky way to deal with multiple modalities
# in the same batch while still being able to benefit from batching
# multimodal inputs. The proper solution should be reordering the
# encoder outputs.
model = cast(SupportsMultiModal, self.model)
if self.lora_config and self.lora_manager.supports_tower_connector_lora():
# Build LoRA mappings independently for encoder inputs
# (encoder batch structure is different from main batch)
prompt_lora_mapping = []
token_lora_mapping = []
lora_requests = set()
encoder_token_counts = []
for req_id, pos_info in mm_lora_refs:
req_idx = self.input_batch.req_id_to_index[req_id]
lora_id = int(self.input_batch.request_lora_mapping[req_idx])
# Prefer pos_info.get_num_embeds to count precise MM embedding tokens.
num_tokens = self.model.get_num_mm_encoder_tokens( # type: ignore[attr-defined]
pos_info.get_num_embeds()
)
prompt_lora_mapping.append(lora_id)
token_lora_mapping.extend([lora_id] * num_tokens)
encoder_token_counts.append(num_tokens)
if lora_id > 0:
lora_request = self.input_batch.lora_id_to_lora_request.get(lora_id)
if lora_request is not None:
lora_requests.add(lora_request)
# Set tower adapter mapping
tower_mapping = LoRAMapping(
tuple(token_lora_mapping),
tuple(prompt_lora_mapping),
is_prefill=True,
type=LoRAMappingType.TOWER,
)
self.lora_manager.set_active_adapters(lora_requests, tower_mapping)
# Only set connector mapping if the model actually has a connector.
# Some multimodal models inherit a stub `get_num_mm_connector_tokens`
# from `SupportsMultiModal`, which returns None and should not be
# treated as a signal that connector LoRA is supported.
mm_mapping = (
self.model.get_mm_mapping() # type: ignore[attr-defined]
if hasattr(self.model, "get_mm_mapping")
else None
)
if (
mm_mapping is not None
and mm_mapping.connector
and hasattr(self.model, "get_num_mm_connector_tokens")
):
post_op_counts = [
self.model.get_num_mm_connector_tokens(num_tokens) # type: ignore[attr-defined]
for num_tokens in encoder_token_counts
]
connector_token_mapping = np.repeat(
np.array(prompt_lora_mapping, dtype=np.int32),
np.array(post_op_counts, dtype=np.int32),
)
connector_mapping = LoRAMapping(
index_mapping=tuple(connector_token_mapping.tolist()),
prompt_mapping=tuple(prompt_lora_mapping),
is_prefill=True,
type=LoRAMappingType.CONNECTOR,
)
self.lora_manager.set_active_adapters(
lora_requests,
connector_mapping,
)
encoder_outputs: list[torch.Tensor] = []
# Track the current index in mm_kwargs/mm_lora_refs to map groups to request IDs
current_item_idx = 0
for modality, num_items, mm_kwargs_batch in group_and_batch_mm_kwargs(
mm_kwargs, device=self.device, pin_memory=PIN_MEMORY
):
batch_outputs: MultiModalEmbeddings
# EVS and dynamic res video related change.
# (ekhvedchenia): Temporary hack to limit peak memory usage when
# processing multimodal data. This solves the issue with scheduler
# putting too many video samples into a single batch. Scheduler
# uses pruned vision tokens count to compare it versus compute
# budget which is incorrect (Either input media size or non-pruned
# output vision tokens count should be considered)
# dynamic res video for nemotron temporarily uses this hack via
# requires_sequential_video_encoding
# because it doesn't yet support video batching.
# TODO(ywang96): Fix memory profiling to take EVS into account and
# remove this hack.
if (
(
self.is_multimodal_pruning_enabled
or self.requires_sequential_video_encoding
)
and modality == "video"
and num_items > 1
):
batch_outputs_lst = list[torch.Tensor]()
for video_idx in range(num_items):
video_mm_kwargs_item = mm_kwargs[current_item_idx + video_idx]
with self.timed_encoder_operation(
should_time, mm_lora_refs, current_item_idx + video_idx, 1
):
_, _, micro_batch_mm_inputs = next(
group_and_batch_mm_kwargs(
[video_mm_kwargs_item],
device=self.device,
pin_memory=PIN_MEMORY,
)
)
micro_batch_outputs = model.embed_multimodal(
**micro_batch_mm_inputs
)
batch_outputs_lst.extend(micro_batch_outputs)
batch_outputs = batch_outputs_lst
else:
# Run the encoder.
# `batch_outputs` is either of the following:
# 1. A tensor of shape (num_items, feature_size, hidden_size)
# in case feature_size is fixed across all multimodal items.
# 2. A list or tuple (length: num_items) of tensors,
# each of shape (feature_size, hidden_size) in case the feature
# size is dynamic depending on the input multimodal items.
with self.timed_encoder_operation(
should_time, mm_lora_refs, current_item_idx, num_items
):
cudagraph_output = None
if (
self.encoder_cudagraph_manager is not None
and self.encoder_cudagraph_manager.supports_modality(modality)
):
cudagraph_output = self.encoder_cudagraph_manager.execute(
mm_kwargs_batch,
)
if cudagraph_output is not None:
batch_outputs = cudagraph_output
else:
batch_outputs = model.embed_multimodal(**mm_kwargs_batch)
sanity_check_mm_encoder_outputs(batch_outputs, expected_num_items=num_items)
encoder_outputs.extend(batch_outputs)
current_item_idx += num_items
# Cache the encoder outputs by mm_hash
for mm_hash, output in zip(mm_hashes, encoder_outputs):
self.encoder_cache[mm_hash] = output
logger.debug("Finish execute for mm hash %s", mm_hash)
self.maybe_save_ec_to_connector(self.encoder_cache, mm_hash)
return encoder_outputs
def _gather_mm_embeddings(
self,
scheduler_output: "SchedulerOutput",
shift_computed_tokens: int = 0,
) -> tuple[list[torch.Tensor], torch.Tensor]:
total_num_scheduled_tokens = scheduler_output.total_num_scheduled_tokens
mm_embeds = list[torch.Tensor]()
is_mm_embed = torch.zeros(
total_num_scheduled_tokens,
dtype=torch.bool,
device="cpu",
pin_memory=PIN_MEMORY,
)
req_start_idx = 0
should_sync_mrope_positions = False
should_sync_xdrope_positions = False
for req_id in self.input_batch.req_ids:
mm_embeds_req: list[torch.Tensor] = []
num_scheduled_tokens = scheduler_output.num_scheduled_tokens[req_id]
req_state = self.requests[req_id]
num_computed_tokens = req_state.num_computed_tokens + shift_computed_tokens
mm_features = req_state.mm_features
lo, hi = get_mm_features_in_window(
mm_features,
start=num_computed_tokens,
end=num_computed_tokens + num_scheduled_tokens,
)
for i in range(lo, hi):
mm_feature = mm_features[i]
pos_info = mm_feature.mm_position
start_pos = pos_info.offset
num_encoder_tokens = pos_info.length
start_idx = max(num_computed_tokens - start_pos, 0)
end_idx = min(
num_computed_tokens - start_pos + num_scheduled_tokens,
num_encoder_tokens,
)
assert start_idx < end_idx
curr_embeds_start, curr_embeds_end = (
pos_info.get_embeds_indices_in_range(start_idx, end_idx)
)
# If there are no embeddings in the current range, we skip
# gathering the embeddings.
if curr_embeds_start == curr_embeds_end:
continue
mm_hash = mm_feature.identifier
encoder_output = self.encoder_cache.get(mm_hash, None)
if encoder_output is None:
# A feature starting at/after the processed boundary is only
# reached via the drafter's +1 look-ahead and might not be
# encoded yet; fall back to the token embedding for drafting.
if (
start_pos
>= req_state.num_computed_tokens + num_scheduled_tokens
):
continue
raise RuntimeError(f"Encoder cache miss for {mm_hash}.")
if (is_embed := pos_info.is_embed) is not None:
is_embed = is_embed[start_idx:end_idx]
mm_embeds_item = encoder_output[curr_embeds_start:curr_embeds_end]
else:
mm_embeds_item = encoder_output[start_idx:end_idx]
req_start_pos = req_start_idx + start_pos - num_computed_tokens
# OR mask for overlapping mm_features (use_audio_in_video)
if is_embed is None:
is_mm_embed[req_start_pos + start_idx : req_start_pos + end_idx] = (
True
)
else:
is_mm_embed[
req_start_pos + start_idx : req_start_pos + end_idx
] |= is_embed
mm_embeds_req.append(mm_embeds_item)
if self.is_multimodal_pruning_enabled and self.uses_mrope:
assert req_state.mrope_positions is not None
should_sync_mrope_positions = True
mm_embeds_req, new_mrope_positions, new_delta = (
self.model.recompute_mrope_positions(
input_ids=req_state.prompt_token_ids,
multimodal_embeddings=mm_embeds_req,
mrope_positions=req_state.mrope_positions,
num_computed_tokens=req_state.num_computed_tokens,
)
)
req_state.mrope_positions.copy_(new_mrope_positions)
req_state.mrope_position_delta = new_delta
mm_embeds.extend(mm_embeds_req)
req_start_idx += num_scheduled_tokens
if should_sync_mrope_positions:
self._calc_mrope_positions(scheduler_output)
self.mrope_positions.copy_to_gpu(total_num_scheduled_tokens)
if should_sync_xdrope_positions:
self._calc_xdrope_positions(scheduler_output)
self.xdrope_positions.copy_to_gpu(total_num_scheduled_tokens)
return mm_embeds, is_mm_embed
def get_model(self) -> nn.Module:
if not hasattr(self, "model"):
raise ValueError("Cannot get model before model has been initialized")
if isinstance(
self.model, (CUDAGraphWrapper, UBatchWrapper, BreakableCUDAGraphWrapper)
):
# get raw model out of the cudagraph wrapper.
return self.model.unwrap()
return self.model
def get_draft_model(self) -> nn.Module | None:
drafter = getattr(self, "drafter", None)
if drafter is None:
return None
model = getattr(drafter, "model", None)
if isinstance(
model, (CUDAGraphWrapper, UBatchWrapper, BreakableCUDAGraphWrapper)
):
return cast(nn.Module, model.unwrap())
return cast(nn.Module | None, model)
def get_supported_generation_tasks(self) -> list[GenerationTask]:
model = self.get_model()
supported_tasks = list[GenerationTask]()
if is_text_generation_model(model):
supported_tasks.append("generate")
if supports_transcription(model):
if model.supports_transcription_only:
return ["transcription"]
supported_tasks.append("transcription")
if supports_realtime(model):
supported_tasks.append("realtime")
return supported_tasks
def get_supported_pooling_tasks(self) -> list[PoolingTask]:
model = self.get_model()
if not is_pooling_model(model):
return []
return list(model.pooler.get_supported_tasks())
def get_supported_tasks(self) -> tuple[SupportedTask, ...]:
tasks = list[SupportedTask]()
if self.model_config.runner_type == "generate":
tasks.extend(self.get_supported_generation_tasks())
if self.model_config.runner_type == "pooling":
tasks.extend(self.get_supported_pooling_tasks())
return tuple(tasks)
def sync_and_gather_intermediate_tensors(
self,
num_tokens: int,
intermediate_tensors: IntermediateTensors | None,
sync_self: bool,
) -> IntermediateTensors:
assert self.intermediate_tensors is not None
tp = self.vllm_config.parallel_config.tensor_parallel_size
is_rs = is_residual_scattered_for_sp(self.vllm_config, num_tokens)
# When sequence parallelism is enabled, the "residual" tensor is
# sharded across TP ranks. All-gather it here because downstream
# QKV + Attention needs the full residual before the SP split point.
if sync_self:
assert intermediate_tensors is not None
for k, v in intermediate_tensors.items():
is_scattered = k == "residual" and is_rs
if is_scattered:
local_len = num_tokens // tp
v = get_tp_group().all_gather(v[:local_len], dim=0)
self.intermediate_tensors[k][:num_tokens].copy_(
v[:num_tokens], non_blocking=True
)
return IntermediateTensors(
{k: v[:num_tokens] for k, v in self.intermediate_tensors.items()}
)
def eplb_step(self, is_dummy: bool = False, is_profile: bool = False) -> None:
"""
Step for the EPLB (Expert Parallelism Load Balancing) state.
"""
if not self.parallel_config.enable_eplb or self.eep_eplb_suppressed:
return
assert self.eplb_state is not None
assert self._moe_model is not None
self.eplb_state.step(
is_dummy,
is_profile,
log_stats=self.parallel_config.eplb_config.log_balancedness,
)
def setup_eplb_from_mapping(
self,
expanded_physical_to_logical: torch.Tensor,
old_num_physical_experts: int,
) -> None:
assert self._moe_model is not None
self.eplb_state = EplbState.from_mapping(
model=self._moe_model,
model_config=self.model_config,
device=self.device,
parallel_config=self.parallel_config,
expanded_physical_to_logical=expanded_physical_to_logical,
num_valid_physical_experts=old_num_physical_experts,
)
def _pool(
self,
hidden_states: torch.Tensor,
num_scheduled_tokens: int,
num_scheduled_tokens_np: np.ndarray,
kv_connector_output: KVConnectorOutput | None,
) -> ModelRunnerOutput | AsyncModelRunnerOutput:
num_reqs = self.input_batch.num_reqs
assert num_reqs == len(self.input_batch.pooling_params), (
"Either all or none of the requests in a batch must be pooling request"
)
hidden_states = hidden_states[:num_scheduled_tokens]
seq_lens_cpu = self.optimistic_seq_lens_cpu[:num_reqs]
pooling_metadata = self.input_batch.get_pooling_metadata()
pooling_metadata.build_pooling_cursor(
num_scheduled_tokens_np,
seq_lens_cpu,
device=hidden_states.device,
query_start_loc_gpu=self.query_start_loc.gpu[: num_reqs + 1],
)
model = cast(VllmModelForPooling, self.model)
raw_pooler_output: PoolerOutput = model.pooler(
hidden_states=hidden_states, pooling_metadata=pooling_metadata
)
finished_mask = [
seq_len == prompt_len
for seq_len, prompt_len in zip(seq_lens_cpu, pooling_metadata.prompt_lens)
]
raw_pooler_output = self.late_interaction_runner.postprocess_pooler_output(
raw_pooler_output=raw_pooler_output,
pooling_params=pooling_metadata.pooling_params,
req_ids=self.input_batch.req_ids,
finished_mask=finished_mask,
)
model_runner_output = ModelRunnerOutput(
req_ids=self.input_batch.req_ids.copy(),
req_id_to_index=self.input_batch.req_id_to_index.copy(),
kv_connector_output=kv_connector_output,
)
if raw_pooler_output is None or not any(finished_mask):
model_runner_output.pooler_output = [None] * num_reqs
return model_runner_output
if not current_platform.is_cuda_alike():
# cpu/xpu runners cannot use the CUDA stream/event-based wrapper.
model_runner_output.pooler_output = _copy_pooler_output_to_cpu(
raw_pooler_output=raw_pooler_output,
finished_mask=finished_mask,
)
self._sync_device()
return model_runner_output
return AsyncGPUPoolingModelRunnerOutput(
model_runner_output=model_runner_output,
raw_pooler_output=raw_pooler_output,
finished_mask=finished_mask,
async_output_copy_stream=self._get_or_create_async_output_copy_stream(),
)
def _pad_for_sequence_parallelism(self, num_scheduled_tokens: int) -> int:
# Pad tokens to multiple of tensor_parallel_size when
# enabled collective fusion for SP
tp_size = self.vllm_config.parallel_config.tensor_parallel_size
if self.compilation_config.pass_config.enable_sp and tp_size > 1:
return round_up(num_scheduled_tokens, tp_size)
return num_scheduled_tokens
def _prepare_mm_inputs(
self, num_tokens: int
) -> tuple[torch.Tensor | None, torch.Tensor]:
if self.model.requires_raw_input_tokens:
input_ids = self.input_ids.gpu[:num_tokens]
else:
input_ids = None
inputs_embeds = self.inputs_embeds.gpu[:num_tokens]
return input_ids, inputs_embeds
def _preprocess(
self,
scheduler_output: "SchedulerOutput",
num_input_tokens: int, # Padded
intermediate_tensors: IntermediateTensors | None = None,
) -> tuple[
torch.Tensor | None,
torch.Tensor | None,
torch.Tensor,
IntermediateTensors | None,
dict[str, Any],
ECConnectorOutput | None,
]:
num_scheduled_tokens = scheduler_output.total_num_scheduled_tokens
is_first_rank = get_pp_group().is_first_rank
is_encoder_decoder = self.model_config.is_encoder_decoder
# Clamp speculative scheduler placeholders (-1) before embedding lookup.
if self.speculative_config is not None:
self.input_ids.gpu[:num_input_tokens].clamp_(min=0)
# _prepare_inputs may reorder the batch, so we must gather multi
# modal outputs after that to ensure the correct order
ec_connector_output = None
if self.supports_mm_inputs and is_first_rank and not is_encoder_decoder:
# Run the multimodal encoder if any.
with self.maybe_get_ec_connector_output(
scheduler_output,
encoder_cache=self.encoder_cache,
) as ec_connector_output:
self._execute_mm_encoder(scheduler_output)
mm_embeds, is_mm_embed = self._gather_mm_embeddings(scheduler_output)
# NOTE(woosuk): To unify token ids and soft tokens (vision
# embeddings), we always use embeddings (rather than token ids)
# as input to the multimodal model, even when the input is text.
if self.enable_prompt_embeds and self.input_batch.req_prompt_embeds:
# Some positions carry precomputed prompt_embeds: they are
# already in self.inputs_embeds and marked is_token_ids=False.
# Embed only the token-id positions (zeroing the placeholder ids
# at prompt_embeds positions so the embedding gather cannot read
# out-of-range ids), and write them back without clobbering the
# prompt_embeds positions.
is_token_ids = self.is_token_ids.gpu[:num_scheduled_tokens]
safe_input_ids = torch.where(
is_token_ids,
self.input_ids.gpu[:num_scheduled_tokens],
0,
)
inputs_embeds_scheduled = self.model.embed_input_ids(
safe_input_ids,
multimodal_embeddings=mm_embeds,
is_multimodal=is_mm_embed,
)
target = self.inputs_embeds.gpu[:num_scheduled_tokens]
self.inputs_embeds.gpu[:num_scheduled_tokens] = torch.where(
is_token_ids.unsqueeze(-1),
inputs_embeds_scheduled,
target,
)
else:
inputs_embeds_scheduled = self.model.embed_input_ids(
self.input_ids.gpu[:num_scheduled_tokens],
multimodal_embeddings=mm_embeds,
is_multimodal=is_mm_embed,
)
# TODO(woosuk): Avoid the copy. Optimize.
self.inputs_embeds.gpu[:num_scheduled_tokens].copy_(
inputs_embeds_scheduled
)
input_ids, inputs_embeds = self._prepare_mm_inputs(num_input_tokens)
model_kwargs = {
**self._init_model_kwargs(),
**self._extract_mm_kwargs(scheduler_output),
}
elif self.enable_prompt_embeds and is_first_rank:
# Get the input embeddings for the tokens that are not input embeds,
# then put them into the appropriate positions.
# TODO(qthequartermasterman): Since even when prompt embeds are
# enabled, (a) not all requests will use prompt embeds, and (b)
# after the initial prompt is processed, the rest of the generated
# tokens will be token ids, it is not desirable to have the
# embedding layer outside of the CUDA graph all the time. The v0
# engine avoids this by "double compiling" the CUDA graph, once
# with input_ids and again with inputs_embeds, for all num_tokens.
# If a batch only has token ids, then including the embedding layer
# in the CUDA graph will be more performant (like in the else case
# below).
is_token_ids = self.is_token_ids.np[:num_scheduled_tokens]
token_ids_idx_np = np.nonzero(is_token_ids)[0]
# Some tokens ids may need to become embeds
if token_ids_idx_np.size > 0:
token_ids_idx = async_tensor_h2d(token_ids_idx_np, device=self.device)
token_ids = self.input_ids.gpu[token_ids_idx]
tokens_to_embeds = self.model.embed_input_ids(input_ids=token_ids)
self.inputs_embeds.gpu[token_ids_idx] = tokens_to_embeds
inputs_embeds = self.inputs_embeds.gpu[:num_input_tokens]
model_kwargs = self._init_model_kwargs()
input_ids = None
else:
# For text-only models, we use token ids as input.
# While it is possible to use embeddings as input just like the
# multimodal models, it is not desirable for performance since
# then the embedding layer is not included in the CUDA graph.
input_ids = self.input_ids.gpu[:num_input_tokens]
inputs_embeds = None
model_kwargs = self._init_model_kwargs()
if self.uses_mrope:
positions = self.mrope_positions.gpu[:, :num_input_tokens]
elif self.uses_xdrope_dim > 0:
positions = self.xdrope_positions.gpu[:, :num_input_tokens]
else:
positions = self.positions[:num_input_tokens]
if num_input_tokens > num_scheduled_tokens:
self.positions[num_scheduled_tokens:num_input_tokens].zero_()
if is_first_rank:
intermediate_tensors = None
else:
assert intermediate_tensors is not None
intermediate_tensors = self.sync_and_gather_intermediate_tensors(
num_input_tokens, intermediate_tensors, True
)
if is_encoder_decoder and scheduler_output.scheduled_encoder_inputs:
# Run the encoder, just like we do with other multimodal inputs.
# For an encoder-decoder model, our processing here is a bit
# simpler, because the outputs are just passed to the decoder.
# We are not doing any prompt replacement. We also will only
# ever have a single encoder input.
encoder_outputs = self._execute_mm_encoder(scheduler_output)
model_kwargs.update({"encoder_outputs": encoder_outputs})
return (
input_ids,
inputs_embeds,
positions,
intermediate_tensors,
model_kwargs,
ec_connector_output,
)
def _sample(
self,
logits: torch.Tensor | None,
spec_decode_metadata: SpecDecodeMetadata | None,
) -> SamplerOutput:
# Sample the next token and get logprobs if needed.
sampling_metadata = self.input_batch.sampling_metadata
# Update output token ids with tokens sampled in last step
# if async scheduling and required by current sampling params.
self.input_batch.update_async_output_token_ids()
if spec_decode_metadata is None:
return self.sampler(
logits=logits,
sampling_metadata=sampling_metadata,
)
# Update spec_token_ids with real draft tokens from pre step only when
# output_token_ids is needed (penalties or bad_words are in use).
if self.use_async_scheduling and self._draft_token_req_ids is not None:
draft_token_ids_cpu, _ = self._get_draft_token_ids_cpu()
self.input_batch.update_async_spec_token_ids(draft_token_ids_cpu)
draft_probs = self._get_spec_decode_draft_probs(spec_decode_metadata)
sampler_output = self.rejection_sampler(
spec_decode_metadata,
draft_probs,
logits,
sampling_metadata,
)
return sampler_output
def _bookkeeping_sync(
self,
scheduler_output: "SchedulerOutput",
sampler_output: SamplerOutput,
logits: torch.Tensor | None,
hidden_states: torch.Tensor,
num_scheduled_tokens: int,
) -> tuple[
dict[str, int],
LogprobsLists | None,
list[list[int]],
dict[str, LogprobsTensors | None],
list[str],
dict[str, int],
list[int],
]:
num_nans_in_logits = {}
if envs.VLLM_COMPUTE_NANS_IN_LOGITS:
num_nans_in_logits = self._get_nans_in_logits(logits)
num_reqs = self.input_batch.num_reqs
discard_sampled_tokens_req_indices = np.nonzero(
self.discard_request_mask.np[:num_reqs]
)[0]
for i in discard_sampled_tokens_req_indices:
gen = self.input_batch.generators.get(int(i))
if gen is not None:
gen.set_offset(gen.get_offset() - 4)
# Copy some objects so they don't get modified after returning.
# This is important when using async scheduling.
req_ids_output_copy = self.input_batch.req_ids.copy()
req_id_to_index_output_copy = self.input_batch.req_id_to_index.copy()
num_sampled_tokens = sampler_output.sampled_token_ids.shape[0]
sampled_token_ids = sampler_output.sampled_token_ids
logprobs_tensors = sampler_output.logprobs_tensors
invalid_req_indices = []
logprobs_lists = None
if not self.use_async_scheduling:
# Sync scheduling: issue routed experts D2H into the pinned
# CPU buffer BEFORE ``_to_list`` below. ``_to_list`` does
# ``event.synchronize()`` on the async copy stream which
# waits for every D2H queued on the default stream since
# the last sync, so this enqueue is naturally covered
# without requiring its own synchronize.
if self.routed_experts_initialized:
buf = self.routed_experts_capturer.get_device_buffer()
total = scheduler_output.total_num_scheduled_tokens
self.routed_experts_cpu[:total].copy_(buf[:total], non_blocking=True)
self.routed_experts_slot_mapping_cpu[:total].copy_(
self.routed_experts_slot_mapping_device[:total],
non_blocking=True,
)
# Get the valid generated tokens.
max_gen_len = sampled_token_ids.shape[-1]
if max_gen_len == 1:
# No spec decode tokens.
valid_sampled_token_ids = self._to_list(sampled_token_ids)
# Mask out the sampled tokens that should not be sampled.
for i in discard_sampled_tokens_req_indices:
valid_sampled_token_ids[int(i)].clear()
if logprobs_tensors is not None:
logprobs_lists = logprobs_tensors.tolists()
else:
# Includes spec decode tokens.
valid_sampled_token_ids, logprobs_lists = RejectionSampler.parse_output(
sampled_token_ids,
self.input_batch.vocab_size,
discard_sampled_tokens_req_indices,
logprobs_tensors=logprobs_tensors,
)
else:
valid_sampled_token_ids = []
invalid_req_indices = discard_sampled_tokens_req_indices.tolist()
invalid_req_indices_set = set(invalid_req_indices)
# Cache the sampled tokens on the GPU and avoid CPU sync.
# These will be copied into input_ids in the next step
# when preparing inputs.
# With spec decoding, this is done in propose_draft_token_ids().
if self.input_batch.prev_sampled_token_ids is None:
assert sampled_token_ids.shape[-1] == 1
self.input_batch.prev_sampled_token_ids = sampled_token_ids
self.input_batch.prev_req_id_to_index = {
req_id: i
for i, req_id in enumerate(self.input_batch.req_ids)
if i not in invalid_req_indices_set
}
# Cache the sampled tokens in the model runner, so that the scheduler
# doesn't need to send them back.
# NOTE(woosuk): As an exception, when using PP, the scheduler sends
# the sampled tokens back, because there's no direct communication
# between the first-stage worker and the last-stage worker.
req_ids = self.input_batch.req_ids
for req_idx in range(num_sampled_tokens):
if self.use_async_scheduling:
sampled_ids = [-1] if req_idx not in invalid_req_indices_set else None
else:
sampled_ids = valid_sampled_token_ids[req_idx]
num_sampled_ids: int = len(sampled_ids) if sampled_ids else 0
if not sampled_ids:
continue
start_idx = self.input_batch.num_tokens_no_spec[req_idx]
end_idx = start_idx + num_sampled_ids
assert end_idx <= self.max_model_len, (
"Sampled token IDs exceed the max model length. "
f"Total number of tokens: {end_idx} > max_model_len: "
f"{self.max_model_len}"
)
self.input_batch.token_ids_cpu[req_idx, start_idx:end_idx] = sampled_ids
self.input_batch.is_token_ids[req_idx, start_idx:end_idx] = True
self.input_batch.num_tokens_no_spec[req_idx] = end_idx
req_id = req_ids[req_idx]
req_state = self.requests[req_id]
req_state.output_token_ids.extend(sampled_ids)
# Compute prompt logprobs if needed.
prompt_logprobs_dict = self._get_prompt_logprobs_dict(
hidden_states[:num_scheduled_tokens],
scheduler_output.num_scheduled_tokens,
)
return (
num_nans_in_logits,
logprobs_lists,
valid_sampled_token_ids,
prompt_logprobs_dict,
req_ids_output_copy,
req_id_to_index_output_copy,
invalid_req_indices,
)
@contextmanager
def synchronize_input_prep(self):
if self.prepare_inputs_event is None:
yield
return
# Ensure prior step has finished with reused CPU tensors.
# This is required in the async scheduling case because
# the CPU->GPU transfer happens async.
self.prepare_inputs_event.synchronize()
try:
yield
finally:
self.prepare_inputs_event.record()
def _model_forward(
self,
input_ids: torch.Tensor | None = None,
positions: torch.Tensor | None = None,
intermediate_tensors: IntermediateTensors | None = None,
inputs_embeds: torch.Tensor | None = None,
**model_kwargs: dict[str, Any],
) -> Any:
"""Helper method to call the model forward pass.
This method can be overridden by subclasses for model execution.
Motivation: We can inspect only this method versus
the whole execute_model, which has additional logic.
Args:
input_ids: Input token IDs
positions: Token positions
intermediate_tensors: Tensors from previous pipeline stages
inputs_embeds: Input embeddings (alternative to input_ids)
**model_kwargs: Additional model arguments
Returns:
Model output tensor
"""
return self.model(
input_ids=input_ids,
positions=positions,
intermediate_tensors=intermediate_tensors,
inputs_embeds=inputs_embeds,
**model_kwargs,
)
@staticmethod
def _is_uniform_decode(
max_num_scheduled_tokens: int,
uniform_decode_query_len: int,
num_tokens: int,
num_reqs: int,
force_uniform_decode: bool | None = None,
) -> bool:
"""
Checks if it's a decode batch with same amount scheduled tokens
across all requests.
"""
return (
(
(max_num_scheduled_tokens == uniform_decode_query_len)
and (num_tokens == max_num_scheduled_tokens * num_reqs)
)
if force_uniform_decode is None
else force_uniform_decode
)
def _determine_batch_execution_and_padding(
self,
num_tokens: int,
num_reqs: int,
num_scheduled_tokens_np: np.ndarray,
max_num_scheduled_tokens: int,
use_cascade_attn: bool,
allow_microbatching: bool = True,
force_eager: bool = False,
# For cudagraph capture TODO(lucas): Refactor how we capture cudagraphs (will
# be improved in model runner v2)
force_uniform_decode: bool | None = None,
force_has_lora: bool | None = None,
force_num_active_loras: int | None = None,
num_encoder_reqs: int = 0,
) -> tuple[
CUDAGraphMode,
BatchDescriptor,
bool,
torch.Tensor | None,
CUDAGraphStat | None,
]:
uniform_decode = self._is_uniform_decode(
max_num_scheduled_tokens=max_num_scheduled_tokens,
uniform_decode_query_len=self.uniform_decode_query_len,
num_tokens=num_tokens,
num_reqs=num_reqs,
force_uniform_decode=force_uniform_decode,
)
# Encoder-decoder models only support CG for decoder_step > 0 (no enc_output
# is present). Also, chunked-prefill is disabled, so batch are uniform.
has_encoder_output = (
self.model_config.is_encoder_decoder and num_encoder_reqs > 0
)
# Compute LoRA state for cudagraph dispatch
num_active_loras = (
force_num_active_loras
if force_num_active_loras is not None
else len(self.input_batch.lora_id_to_lora_request)
)
has_lora = num_active_loras > 0 if force_has_lora is None else force_has_lora
num_tokens_padded = self._pad_for_sequence_parallelism(num_tokens)
def dispatch_cudagraph(num_tokens, disable_full=False, valid_modes=None):
return self.cudagraph_dispatcher.dispatch(
num_tokens=num_tokens,
has_lora=has_lora,
uniform_decode=uniform_decode,
num_active_loras=num_active_loras,
valid_modes={CUDAGraphMode.NONE} if force_eager else valid_modes,
invalid_modes={CUDAGraphMode.FULL} if disable_full else None,
)
cudagraph_mode, batch_descriptor = dispatch_cudagraph(
num_tokens_padded, disable_full=use_cascade_attn or has_encoder_output
)
num_tokens_padded = batch_descriptor.num_tokens
if self.compilation_config.pass_config.enable_sp:
assert (
batch_descriptor.num_tokens
% self.vllm_config.parallel_config.tensor_parallel_size
== 0
), (
"Sequence parallelism requires num_tokens to be "
"a multiple of tensor parallel size"
)
# Extra coordination when running data-parallel since we need to coordinate
# across ranks
should_ubatch, num_tokens_across_dp = False, None
if self.vllm_config.parallel_config.data_parallel_size > 1:
should_ubatch, num_tokens_across_dp, synced_cudagraph_mode = (
coordinate_batch_across_dp(
num_tokens_unpadded=num_tokens,
parallel_config=self.parallel_config,
allow_microbatching=allow_microbatching,
num_tokens_padded=num_tokens_padded,
uniform_decode=uniform_decode,
cudagraph_mode=cudagraph_mode.value,
)
)
# Extract DP-synced values
if num_tokens_across_dp is not None:
dp_rank = self.parallel_config.data_parallel_rank
num_tokens_padded = int(num_tokens_across_dp[dp_rank].item())
# Re-dispatch with DP padding so we have the correct batch_descriptor
cudagraph_mode, batch_descriptor = dispatch_cudagraph(
num_tokens_padded,
valid_modes={CUDAGraphMode(synced_cudagraph_mode)},
)
# Assert to make sure the agreed upon token count is correct otherwise
# num_tokens_across_dp will no-longer be valid
assert batch_descriptor.num_tokens == num_tokens_padded
cudagraph_stats = None
if self.vllm_config.observability_config.cudagraph_metrics:
cudagraph_stats = CUDAGraphStat(
num_unpadded_tokens=num_tokens,
num_padded_tokens=batch_descriptor.num_tokens,
num_paddings=batch_descriptor.num_tokens - num_tokens,
runtime_mode=str(cudagraph_mode),
)
return (
cudagraph_mode,
batch_descriptor,
should_ubatch,
num_tokens_across_dp,
cudagraph_stats,
)
def _register_layerwise_nvtx_hooks(self) -> None:
"""
Register layerwise NVTX hooks if --enable-layerwise-nvtx-tracing is enabled
to trace detailed information of each layer or module in the model.
"""
if (
self.vllm_config.observability_config.enable_layerwise_nvtx_tracing
and not self.layerwise_nvtx_hooks_registered
):
if self.compilation_config.cudagraph_mode != CUDAGraphMode.NONE:
logger.debug_once(
"layerwise NVTX tracing is not supported when CUDA graph is "
"turned off; you may observe part or all of the model "
"missing NVTX markers"
)
# In STOCK_TORCH_COMPILE mode, after registering hooks here,
# the __call__ function of nn.module will be recompiled with
# fullgraph=True. Since nvtx.range_push/pop are not traceable
# by torch dynamo, we can't register hook functions here
# because hook functions will also be traced by torch dynamo.
if (
self.vllm_config.compilation_config.mode
== CompilationMode.STOCK_TORCH_COMPILE
):
logger.debug_once(
"layerwise NVTX tracing is not supported when "
"CompilationMode is STOCK_TORCH_COMPILE, skipping "
"function hooks registration"
)
else:
pyt_hooks = PytHooks()
pyt_hooks.register_hooks(self.model, self.model.__class__.__name__)
self.layerwise_nvtx_hooks_registered = True
def _get_slot_mappings(
self,
num_tokens_padded: int,
num_reqs_padded: int,
num_tokens_unpadded: int,
ubatch_slices: "UBatchSlices | None" = None,
) -> tuple[
dict[int, torch.Tensor] | None,
dict[str, torch.Tensor] | list[dict[str, torch.Tensor]] | None,
]:
"""
Build slot mappings in both formats needed by the system.
Args:
num_tokens_padded: Total number of tokens (padded)
num_reqs_padded: Total number of requests (padded)
num_tokens_unpadded: Actual number of tokens (unpadded)
ubatch_slices: Optional ubatch slicing info for DBO
Returns:
A tuple of:
- slot_mappings_by_gid: dict[int, torch.Tensor] for attention metadata
- slot_mappings_by_layer: dict[str, torch.Tensor] or list for ForwardContext
"""
if not (
hasattr(self, "kv_cache_config")
and self.kv_cache_config is not None
and len(self.kv_cache_config.kv_cache_groups) > 0
):
return None, None
def _get_slot_mapping(kv_cache_gid: int):
assert num_reqs_padded is not None and num_tokens_padded is not None
kv_cache_spec = self.kv_cache_config.kv_cache_groups[
kv_cache_gid
].kv_cache_spec
if isinstance(kv_cache_spec, EncoderOnlyAttentionSpec):
slot_mapping = torch.zeros(
(num_tokens_padded,),
dtype=torch.int64,
device=self.device,
)
else:
blk_table = self.input_batch.block_table[kv_cache_gid]
slot_mapping = blk_table.slot_mapping.gpu[:num_tokens_padded]
# Fill unused with -1. Needed for reshape_and_cache in full cuda
# graph mode. `blk_table_tensor` -1 to match mamba PAD_SLOT_ID
slot_mapping[num_tokens_unpadded:num_tokens_padded].fill_(-1)
return slot_mapping
slot_mappings_by_gid = {
gid: _get_slot_mapping(gid)
for gid, _ in enumerate(self.kv_cache_config.kv_cache_groups)
}
slot_mappings_by_layer: dict[str, torch.Tensor] = {}
for gid, kv_cache_group in enumerate(self.kv_cache_config.kv_cache_groups):
slot_mapping = slot_mappings_by_gid[gid]
for layer_name in kv_cache_group.layer_names:
slot_mappings_by_layer[layer_name] = slot_mapping
if ubatch_slices is not None:
result: list[dict[str, torch.Tensor]] = []
for ubatch in ubatch_slices:
sliced_mappings: dict[str, torch.Tensor] = {}
for layer_name, slot_mapping in slot_mappings_by_layer.items():
sliced_mappings[layer_name] = slot_mapping[ubatch.token_slice]
result.append(sliced_mappings)
return slot_mappings_by_gid, result
return slot_mappings_by_gid, slot_mappings_by_layer
def _is_all_reqs_chunked_prefill(self) -> bool:
"""Check if all scheduled requests are marked to discard sampled tokens.
This is true when `discard_request_mask` is set for every scheduled
request (e.g., for chunked prefill requests that are not the last
prefill chunk)."""
num_reqs = self.input_batch.num_reqs
return bool(self.discard_request_mask.np[:num_reqs].all())
@torch.inference_mode()
def execute_model(
self,
scheduler_output: "SchedulerOutput",
intermediate_tensors: IntermediateTensors | None = None,
) -> ModelRunnerOutput | AsyncModelRunnerOutput | IntermediateTensors | None:
if self.execute_model_state is not None:
raise RuntimeError(
"State error: sample_tokens() must be called "
"after execute_model() returns None."
)
if self.routed_experts_initialized:
self.routed_experts_capturer.clear_buffer()
# If ngram_gpu is used, we need to copy the scheduler_output to avoid
# the modification has influence on the scheduler_output in engine core process.
# The replace is much faster than deepcopy.
if (
self.speculative_config is not None
and self.speculative_config.use_ngram_gpu()
):
num_scheduled_tokens_copy = scheduler_output.num_scheduled_tokens.copy()
spec_decode_tokens_copy = (
scheduler_output.scheduled_spec_decode_tokens.copy()
)
scheduler_output = replace(
scheduler_output,
num_scheduled_tokens=num_scheduled_tokens_copy,
scheduled_spec_decode_tokens=spec_decode_tokens_copy,
)
if has_kv_transfer_group():
kv_connector_metadata = scheduler_output.kv_connector_metadata
assert kv_connector_metadata is not None
get_kv_transfer_group().handle_preemptions(kv_connector_metadata)
num_scheduled_tokens = scheduler_output.total_num_scheduled_tokens
with (
record_function_or_nullcontext("gpu_model_runner: preprocess"),
self.synchronize_input_prep(),
):
# Update persistent batch states.
deferred_state_corrections_fn = self._update_states(scheduler_output)
if has_ec_transfer() and not get_ec_transfer().is_consumer:
with self.maybe_get_ec_connector_output(
scheduler_output,
encoder_cache=self.encoder_cache,
) as ec_connector_output:
self._execute_mm_encoder(scheduler_output)
return make_empty_encoder_model_runner_output(scheduler_output)
if not num_scheduled_tokens:
if (
self.parallel_config.distributed_executor_backend
== "external_launcher"
and self.parallel_config.data_parallel_size > 1
):
# this is a corner case when both external launcher
# and DP are enabled, num_scheduled_tokens could be
# 0, and has_unfinished_requests in the outer loop
# returns True. before returning early here we call
# dummy run to ensure coordinate_batch_across_dp
# is called into to avoid out of sync issues.
self._dummy_run(1)
if not has_kv_transfer_group():
# Return empty ModelRunnerOutput if no work to do.
return EMPTY_MODEL_RUNNER_OUTPUT
return self.kv_connector_no_forward(scheduler_output, self.vllm_config)
if self.cache_config.kv_sharing_fast_prefill:
assert not self.num_prompt_logprobs, (
"--kv-sharing-fast-prefill produces incorrect "
"logprobs for prompt tokens, tokens, please disable "
"it when the requests need prompt logprobs"
)
num_reqs = self.input_batch.num_reqs
req_ids = self.input_batch.req_ids
tokens = [scheduler_output.num_scheduled_tokens[i] for i in req_ids]
num_scheduled_tokens_np = np.array(tokens, dtype=np.int32)
max_num_scheduled_tokens = int(num_scheduled_tokens_np.max())
num_tokens_unpadded = scheduler_output.total_num_scheduled_tokens
logits_indices, spec_decode_metadata = self._prepare_inputs(
scheduler_output,
num_scheduled_tokens_np,
)
cascade_attn_prefix_lens = None
# Disable cascade attention when using microbatching (DBO)
if self.cascade_attn_enabled and not self.parallel_config.use_ubatching:
# Pre-compute cascade attention prefix lengths
cascade_attn_prefix_lens = self._compute_cascade_attn_prefix_lens(
num_scheduled_tokens_np,
self.input_batch.num_computed_tokens_cpu[:num_reqs],
scheduler_output.num_common_prefix_blocks,
)
(
cudagraph_mode,
batch_desc,
should_ubatch,
num_tokens_across_dp,
cudagraph_stats,
) = self._determine_batch_execution_and_padding(
num_tokens=num_tokens_unpadded,
num_reqs=num_reqs,
num_scheduled_tokens_np=num_scheduled_tokens_np,
max_num_scheduled_tokens=max_num_scheduled_tokens,
use_cascade_attn=cascade_attn_prefix_lens is not None,
num_encoder_reqs=len(scheduler_output.scheduled_encoder_inputs),
)
logger.debug(
"Running batch with cudagraph_mode: %s, batch_descriptor: %s, "
"should_ubatch: %s, num_tokens_across_dp: %s",
cudagraph_mode,
batch_desc,
should_ubatch,
num_tokens_across_dp,
)
num_tokens_padded = batch_desc.num_tokens
num_reqs_padded = (
batch_desc.num_reqs if batch_desc.num_reqs is not None else num_reqs
)
ubatch_slices, ubatch_slices_padded = maybe_create_ubatch_slices(
should_ubatch,
num_scheduled_tokens_np,
num_tokens_padded,
num_reqs_padded,
self.parallel_config.num_ubatches,
)
logger.debug(
"ubatch_slices: %s, ubatch_slices_padded: %s",
ubatch_slices,
ubatch_slices_padded,
)
# True if any attention backend handles KV cache update separately
# from forward() (i.e., forward_includes_kv_cache_update=False). When true,
# slot_mappings must use padded dimensions to match the key/value tensors.
has_separate_kv_update = not all(
all(
g.backend.forward_includes_kv_cache_update
for g in self.attn_groups[id]
)
for id, spec in enumerate(self.kv_cache_config.kv_cache_groups)
if not isinstance(spec.kv_cache_spec, EncoderOnlyAttentionSpec)
)
pad_attn = cudagraph_mode == CUDAGraphMode.FULL
if self.cache_config.mamba_cache_mode == "align":
# preprocess_mamba reads req_state.num_computed_tokens (CPU)
# to decide copy operations, so we must apply deferred
# corrections before it runs.
if deferred_state_corrections_fn:
deferred_state_corrections_fn()
deferred_state_corrections_fn = None
mamba_bufs = self._get_mamba_bufs()
mamba_utils.preprocess_mamba(
scheduler_output,
self.kv_cache_config,
self.cache_config,
self.mamba_state_idx,
self.input_batch,
self.requests,
self.compilation_config.static_forward_context,
self.model.get_mamba_state_copy_func(),
mamba_bufs.preprocess,
)
# preprocess_mamba resets num_accepted_tokens_cpu to 1
# for requests whose state was copied to a new block.
# Re-sync to GPU so the mamba kernel reads from the
# correct initial state slot (init_token_idx = 0).
self.num_accepted_tokens.np[:num_reqs] = (
self.input_batch.num_accepted_tokens_cpu[:num_reqs]
)
self.num_accepted_tokens.copy_to_gpu(num_reqs)
# Stage per-request inputs for the fused postprocess kernel
# only when that kernel will actually run. The kernel is
# gated on spec-decode + hybrid (see MambaBuffers.create);
# without it, ``mamba_bufs.postprocess_align`` is None and
# the staging buffers don't exist.
if mamba_bufs.postprocess_align is not None:
mamba_utils.stage_postprocess_inputs_to_gpu(
mamba_bufs.postprocess_align,
scheduler_output,
self.input_batch.req_ids,
num_reqs,
self.requests,
self.mamba_state_idx,
)
use_spec_decode = len(scheduler_output.scheduled_spec_decode_tokens) > 0
ubatch_slices_attn = ubatch_slices_padded if pad_attn else ubatch_slices
slot_mappings_by_group, slot_mappings = self._get_slot_mappings(
num_tokens_padded=num_tokens_padded
if pad_attn or has_separate_kv_update
else num_tokens_unpadded,
num_reqs_padded=(
num_reqs_padded if pad_attn or has_separate_kv_update else num_reqs
),
num_tokens_unpadded=num_tokens_unpadded,
ubatch_slices=ubatch_slices_padded,
)
attn_metadata, spec_decode_common_attn_metadata = (
self._build_attention_metadata(
num_tokens=num_tokens_unpadded,
num_tokens_padded=num_tokens_padded if pad_attn else None,
num_reqs=num_reqs,
num_reqs_padded=num_reqs_padded if pad_attn else None,
max_query_len=max_num_scheduled_tokens,
ubatch_slices=ubatch_slices_attn,
logits_indices=logits_indices,
use_spec_decode=use_spec_decode,
num_scheduled_tokens=scheduler_output.num_scheduled_tokens,
cascade_attn_prefix_lens=cascade_attn_prefix_lens,
slot_mappings=slot_mappings_by_group,
)
)
(
input_ids,
inputs_embeds,
positions,
intermediate_tensors,
model_kwargs,
ec_connector_output,
) = self._preprocess(
scheduler_output, num_tokens_padded, intermediate_tensors
)
# Set cudagraph mode to none if calc_kv_scales is true.
# KV scales calculation involves dynamic operations that are incompatible
# with CUDA graph capture.
if self.calculate_kv_scales:
cudagraph_mode = CUDAGraphMode.NONE
# Mark KV scales as calculated after the first forward pass
self.calculate_kv_scales = False
# Encoder-decoder models can only compile the pure decode steps where no
# encoder inputs are present. Use eager for the first pass.
num_encoder_reqs = len(scheduler_output.scheduled_encoder_inputs)
has_encoder_input = (
self.model_config.is_encoder_decoder and num_encoder_reqs > 0
)
# Run the model.
# Use persistent buffers for CUDA graphs.
# When spec decode is enabled, defer connector finalization
# (wait_for_save + clear metadata) until after draft model runs.
defer_kv_connector_finalize = self.speculative_config is not None
# Update the EPLB meta.
if self.eplb_state is not None:
self.eplb_state.prepare_forward(
self.model_config,
num_tokens_unpadded,
ubatch_slices_padded,
)
with (
set_forward_context(
attn_metadata,
self.vllm_config,
num_tokens=num_tokens_padded,
num_tokens_across_dp=num_tokens_across_dp,
cudagraph_runtime_mode=cudagraph_mode,
batch_descriptor=batch_desc,
ubatch_slices=ubatch_slices_padded,
slot_mapping=slot_mappings,
skip_compiled=has_encoder_input,
),
record_function_or_nullcontext("gpu_model_runner: forward"),
self.maybe_get_kv_connector_output(
scheduler_output,
defer_finalize=defer_kv_connector_finalize,
) as kv_connector_output,
):
model_output = self._model_forward(
input_ids=input_ids,
positions=positions,
intermediate_tensors=intermediate_tensors,
inputs_embeds=inputs_embeds,
**model_kwargs,
)
with record_function_or_nullcontext("gpu_model_runner: postprocess"):
if self.use_aux_hidden_state_outputs:
# True when EAGLE 3 is used.
hidden_states, aux_hidden_states = model_output
else:
# Common case.
hidden_states = model_output
aux_hidden_states = None
if not self.broadcast_pp_output:
# Common case.
if not get_pp_group().is_last_rank:
# Return the intermediate tensors.
assert isinstance(hidden_states, IntermediateTensors)
self.kv_connector_output = kv_connector_output
return hidden_states
if self.is_pooling_model:
# Return the pooling output.
return self._pool(
hidden_states,
num_scheduled_tokens,
num_scheduled_tokens_np,
kv_connector_output,
)
sample_hidden_states = hidden_states[logits_indices]
logits = self.model.compute_logits(sample_hidden_states)
else:
# Rare case.
assert not self.is_pooling_model
sample_hidden_states = hidden_states[logits_indices]
if not get_pp_group().is_last_rank:
all_gather_tensors = {
"residual": not is_residual_scattered_for_sp(
self.vllm_config, num_tokens_padded
)
}
get_pp_group().send_tensor_dict(
hidden_states.tensors,
all_gather_group=get_tp_group(),
all_gather_tensors=all_gather_tensors,
)
logits = None
else:
logits = self.model.compute_logits(sample_hidden_states)
model_output_broadcast_data: dict[str, Any] = {}
if logits is not None:
model_output_broadcast_data["logits"] = logits.contiguous()
broadcasted = get_pp_group().broadcast_tensor_dict(
model_output_broadcast_data, src=len(get_pp_group().ranks) - 1
)
assert broadcasted is not None
logits = broadcasted["logits"]
self.execute_model_state = ExecuteModelState(
scheduler_output,
logits,
spec_decode_metadata,
spec_decode_common_attn_metadata,
hidden_states,
sample_hidden_states,
aux_hidden_states,
ec_connector_output,
cudagraph_stats,
slot_mappings,
)
self.kv_connector_output = kv_connector_output
# Now the batch has been launched we can wait for corrections from the
# previous model forward without breaking async scheduling.
if deferred_state_corrections_fn:
deferred_state_corrections_fn()
return None
def _input_fits_in_drafter(
self, common_attn_metadata: CommonAttentionMetadata | None
) -> bool:
if common_attn_metadata is None:
return False
assert self.speculative_config is not None
# DFlash queries one extra token (the bonus token) beyond num_spec_tokens
num_drafter_query_tokens = self.num_spec_tokens + (
1 if self.speculative_config.use_dflash() else 0
)
return (
common_attn_metadata.max_seq_len + num_drafter_query_tokens
<= self.effective_drafter_max_model_len
)
@torch.inference_mode
def sample_tokens(
self, grammar_output: "GrammarOutput | None"
) -> ModelRunnerOutput | AsyncModelRunnerOutput | IntermediateTensors:
if self.execute_model_state is None:
kv_connector_output = self.kv_connector_output
self.kv_connector_output = None
# receive sampled token ids from the last PP rank.
if self.use_async_scheduling and not get_pp_group().is_last_rank:
self._pp_receive_prev_sampled_token_ids_to_input_batch()
# In case of PP with kv transfer, we need to pass through the
# kv_connector_output
return ModelRunnerOutput.with_kv_conn_output_only(kv_connector_output)
# Unpack ephemeral state.
(
scheduler_output,
logits,
spec_decode_metadata,
spec_decode_common_attn_metadata,
hidden_states,
sample_hidden_states,
aux_hidden_states,
ec_connector_output,
cudagraph_stats,
slot_mappings,
) = self.execute_model_state
# Clear ephemeral state.
self.execute_model_state = None
# Apply structured output bitmasks if present.
if grammar_output is not None:
apply_grammar_bitmask(
scheduler_output, grammar_output, self.input_batch, logits
)
with record_function_or_nullcontext("gpu_model_runner: sample"):
sampler_output = self._sample(logits, spec_decode_metadata)
self._update_states_after_model_execute(
sampler_output.sampled_token_ids, scheduler_output
)
if self.use_async_scheduling:
pp = get_pp_group()
# For torchrun external_launcher PP mode with broadcast_pp_output=True,
# PP outputs have been broadcasted to all ranks at logits computation.
# Therefore, here is no need to send sampled token ids again in this case.
if not self.broadcast_pp_output and pp.world_size > 1 and pp.is_last_rank:
self._pp_broadcast_prev_sampled_token_ids(
sampler_output.sampled_token_ids
)
self._draft_token_ids = None
self._draft_probs = None
self._draft_prob_req_ids = None
self._draft_token_req_ids = None
self.valid_sampled_token_count_gpu = None
self.input_batch.prev_sampled_token_ids = None
def propose_draft_token_ids(sampled_token_ids):
assert spec_decode_common_attn_metadata is not None
with record_function_or_nullcontext("gpu_model_runner: draft"):
self._draft_token_ids = self.propose_draft_token_ids(
scheduler_output,
sampled_token_ids,
self.input_batch.sampling_metadata,
hidden_states,
sample_hidden_states,
aux_hidden_states,
spec_decode_metadata,
spec_decode_common_attn_metadata,
slot_mappings,
)
self._copy_draft_token_ids_to_cpu(scheduler_output)
spec_config = self.speculative_config
draft_after_bookkeeping = False
if spec_config is not None:
# Decide whether to run the drafter or zero out draft tokens.
input_fits_in_drafter = self._input_fits_in_drafter(
spec_decode_common_attn_metadata
)
# Whether the drafter runs a GPU model forward (and thus carries
# TP/EP/DP collectives), independent of padded-batch timing.
drafter_runs_model_forward = (
spec_config.use_eagle()
or spec_config.uses_draft_model()
or spec_config.uses_extract_hidden_states()
)
use_gpu_toks = (
drafter_runs_model_forward
and not spec_config.disable_padded_drafter_batch
)
if use_gpu_toks:
# EAGLE/DraftModel speculative decoding can use the GPU sampled tokens
# as inputs, and does not need to wait for bookkeeping to finish.
assert isinstance(
self.drafter,
EagleProposer
| DFlashProposer
| DraftModelProposer
| ExtractHiddenStatesProposer
| Gemma4Proposer,
)
sampled_token_ids = sampler_output.sampled_token_ids
if input_fits_in_drafter:
propose_draft_token_ids(sampled_token_ids)
else:
if self.valid_sampled_token_count_event is not None:
assert spec_decode_common_attn_metadata is not None
next_token_ids, valid_sampled_tokens_count = (
self.drafter.prepare_next_token_ids_padded(
sampled_token_ids,
self.requests,
self.input_batch,
self.discard_request_mask.gpu,
)
)
self._copy_valid_sampled_token_count(
next_token_ids, valid_sampled_tokens_count
)
if self.parallel_config.data_parallel_size > 1:
# Prevent hang when DP ranks disagree on input_fits_in_drafter
self.drafter.dummy_run(num_tokens=1)
elif (
spec_config.use_ngram_gpu()
and not spec_config.disable_padded_drafter_batch
):
assert isinstance(self.drafter, NgramProposerGPU)
sampled_token_ids = sampler_output.sampled_token_ids
if input_fits_in_drafter:
propose_draft_token_ids(sampled_token_ids)
elif self.valid_sampled_token_count_event is not None:
assert spec_decode_common_attn_metadata is not None
next_token_ids, valid_sampled_tokens_count, _ = (
self.drafter.update_token_ids_ngram(
sampled_token_ids,
self.input_batch,
self.token_ids_gpu_tensor,
self.num_tokens_no_spec_gpu,
self.discard_request_mask.gpu,
)
)
self._copy_valid_sampled_token_count(
next_token_ids, valid_sampled_tokens_count
)
else:
# These drafters consume CPU sampled tokens, so they run
# after bookkeeping.
draft_after_bookkeeping = True
if not input_fits_in_drafter:
# Zero out draft tokens so the scheduler doesn't schedule
# stale drafts from the previous step.
# For Nemotron-H: it is necessary to zero out the draft tokens,
# otherwise the stale tokens will corrupt Mamba recurrent
# state and logprobs for sequences near max_model_len.
self._draft_token_ids = torch.zeros(
1, device=self.device, dtype=torch.int32
).expand(len(self.input_batch.req_ids), self.num_spec_tokens)
self._draft_probs = None
self._draft_prob_req_ids = None
self._copy_draft_token_ids_to_cpu(scheduler_output, zeros_only=True)
with record_function_or_nullcontext("gpu_model_runner: bookkeep"):
(
num_nans_in_logits,
logprobs_lists,
valid_sampled_token_ids,
prompt_logprobs_dict,
req_ids_output_copy,
req_id_to_index_output_copy,
invalid_req_indices,
) = self._bookkeeping_sync(
scheduler_output,
sampler_output,
logits,
hidden_states,
scheduler_output.total_num_scheduled_tokens,
)
if draft_after_bookkeeping:
# ngram and other speculative decoding methods use the sampled
# tokens on the CPU, so they are run after bookkeeping.
if input_fits_in_drafter:
propose_draft_token_ids(valid_sampled_token_ids)
elif (
drafter_runs_model_forward
and self.parallel_config.data_parallel_size > 1
):
# Prevent hang when DP ranks disagree on input_fits_in_drafter
assert isinstance(
self.drafter,
EagleProposer
| DFlashProposer
| DraftModelProposer
| ExtractHiddenStatesProposer
| Gemma4Proposer,
)
self.drafter.dummy_run(num_tokens=1)
# Finalize KV connector (wait_for_save + clear metadata) after
# draft model runs. Deferred from target model forward to allow
# draft model to also save its KV cache.
if spec_config is not None:
self.finalize_kv_connector()
with record_function_or_nullcontext("gpu_model_runner: eplb"):
self.eplb_step()
# self.kv_connector_output may be modified during drafting
kv_connector_output = self.kv_connector_output
self.kv_connector_output = None
with record_function_or_nullcontext("gpu_model_runner: ModelRunnerOutput"):
output = ModelRunnerOutput(
req_ids=req_ids_output_copy,
req_id_to_index=req_id_to_index_output_copy,
sampled_token_ids=valid_sampled_token_ids,
logprobs=logprobs_lists,
prompt_logprobs_dict=prompt_logprobs_dict,
kv_connector_output=kv_connector_output,
ec_connector_output=ec_connector_output
if self.supports_mm_inputs
else None,
num_nans_in_logits=num_nans_in_logits,
cudagraph_stats=cudagraph_stats,
routed_experts=None,
)
if not self.use_async_scheduling:
if self.routed_experts_initialized:
# Sync path: D2H was issued in ``_bookkeeping_sync`` and
# synchronized by ``_to_list``'s event.synchronize(), so
# the pinned buffers are ready to be wrapped as numpy.
total = scheduler_output.total_num_scheduled_tokens
output.routed_experts = RoutedExpertsLists(
routing_data=self.routed_experts_cpu[:total].numpy(),
slot_mapping=self.routed_experts_slot_mapping_cpu[:total].numpy(),
)
return output
with record_function_or_nullcontext(
"gpu_model_runner: AsyncGPUModelRunnerOutput"
):
# Async path: produce a device-side snapshot that the async
# copy stream can D2H later. Both tensors must be private
# clones because:
# - ``routing_data`` source is the shared capturer buffer,
# which is ``clear_buffer()``-ed at the start of the
# next step on the default stream.
# - ``slot_mapping`` source is our own
# ``routed_experts_slot_mapping_device``, which the
# next ``_prepare_inputs`` overwrites on the default
# stream while the D2H is still pending on the copy
# stream.
# Without clones, the copy stream would read torn data.
routed_experts_snapshot = None
if self.routed_experts_initialized:
buf = self.routed_experts_capturer.get_device_buffer()
total = scheduler_output.total_num_scheduled_tokens
routed_experts_snapshot = RoutedExpertsTensors(
routing_data=buf[:total].clone(),
slot_mapping=self.routed_experts_slot_mapping_device[
:total
].clone(),
)
async_output = AsyncGPUModelRunnerOutput(
model_runner_output=output,
sampled_token_ids=sampler_output.sampled_token_ids,
logprobs_tensors=sampler_output.logprobs_tensors,
invalid_req_indices=invalid_req_indices,
async_output_copy_stream=self._get_or_create_async_output_copy_stream(),
vocab_size=self.input_batch.vocab_size,
routed_experts=routed_experts_snapshot,
check_ep_fault=self.check_ep_fault,
)
with record_function_or_nullcontext(
"gpu_model_runner: set_async_sampled_token_ids"
):
# Save ref of sampled_token_ids CPU tensor if the batch contains
# any requests with sampling params that require output ids.
self.input_batch.set_async_sampled_token_ids(
async_output.sampled_token_ids_cpu,
async_output.async_copy_ready_event,
)
return async_output
def _pp_broadcast_prev_sampled_token_ids(
self, sampled_token_ids: torch.Tensor
) -> None:
"""Broadcast sampled token ids (GPU) from last PP stage"""
pp = get_pp_group()
assert pp.is_last_rank
# `prev_sampled_token_ids` is expected to have shape [num_reqs, 1].
assert sampled_token_ids.dim() == 2 and sampled_token_ids.shape[-1] == 1, (
"PP+async expects sampled_token_ids to have shape [num_reqs, 1]"
)
# Skip for chunked prefill: sampled tokens are dummy
# and will be discarded, no need to broadcast.
if not self._is_all_reqs_chunked_prefill():
torch.distributed.broadcast(
sampled_token_ids, src=pp.rank, group=pp.device_group
)
def _pp_receive_prev_sampled_token_ids_to_input_batch(self) -> None:
"""Receive sampled token ids broadcast from last PP stage"""
pp = get_pp_group()
assert not pp.is_last_rank
num_reqs = self.input_batch.num_reqs
# `prev_sampled_token_ids` is expected to have shape [num_reqs, 1].
recv = torch.empty((num_reqs, 1), dtype=torch.int32, device=self.device)
# skip for chunked prefill.
if not self._is_all_reqs_chunked_prefill():
torch.distributed.broadcast(recv, src=pp.last_rank, group=pp.device_group)
self.input_batch.prev_sampled_token_ids = recv
# construct `prev_req_id_to_index` here so `_prepare_input_ids`
# can map req_id -> previous batch row
discard_req_indices = np.nonzero(self.discard_request_mask.np[:num_reqs])[0]
discard_req_indices_set = set(discard_req_indices)
prev_req_id_to_index: dict[str, int] = {}
for i, req_id in enumerate(self.input_batch.req_ids):
if i in discard_req_indices_set:
continue
prev_req_id_to_index[req_id] = i
# PP+async scheduling: advance per-request local cached output length by
# appending a placeholder (-1) token id.
if (req_state := self.requests.get(req_id)) is not None:
req_state.output_token_ids.append(-1)
pos = self.input_batch.num_tokens_no_spec[i]
self.input_batch.is_token_ids[i, pos] = True
self.input_batch.num_tokens_no_spec[i] = pos + 1
self.input_batch.prev_req_id_to_index = prev_req_id_to_index
def take_draft_token_ids(self) -> DraftTokenIds | None:
if not self.num_spec_tokens or not self._draft_token_req_ids:
return None
draft_token_ids, req_ids = self._get_draft_token_ids_cpu()
return DraftTokenIds(req_ids, draft_token_ids)
def _copy_draft_token_ids_to_cpu(
self, scheduler_output: "SchedulerOutput", zeros_only: bool = False
) -> None:
if torch.is_tensor(self._draft_token_ids):
assert isinstance(self._draft_token_ids, torch.Tensor)
self.prev_num_spec_tokens = self._draft_token_ids.shape[1]
# Check if we need to copy draft tokens to CPU. In async scheduling,
# we only copy when needed for structured output, penalties or bad_words.
if self.use_async_scheduling and not (
scheduler_output.has_structured_output_requests
or self.input_batch.sampling_metadata.output_token_ids
):
return
# We must also set the corresponding request ids.
self._draft_token_req_ids = self.input_batch.req_ids.copy()
draft_token_ids: torch.Tensor = self._draft_token_ids
if not torch.is_tensor(draft_token_ids):
return
assert self.draft_token_ids_event is not None
assert self.draft_token_ids_copy_stream is not None
assert self.draft_token_ids_cpu is not None
default_stream = torch.cuda.current_stream()
num_reqs = draft_token_ids.shape[0]
num_spec_tokens = draft_token_ids.shape[1]
with torch.cuda.stream(self.draft_token_ids_copy_stream):
if not zeros_only:
# Trigger async copy of draft token ids to cpu.
self.draft_token_ids_copy_stream.wait_stream(default_stream)
self.draft_token_ids_cpu[:num_reqs, :num_spec_tokens].copy_(
draft_token_ids, non_blocking=True
)
else:
# No copy needed, just zero-out cpu tensor.
self.draft_token_ids_cpu[:num_reqs, :num_spec_tokens] = 0
self.draft_token_ids_event.record()
def _get_draft_token_ids_cpu(self) -> tuple[list[list[int]], list[str]]:
if isinstance(self._draft_token_ids, list):
return self._draft_token_ids, self.input_batch.req_ids
req_ids = self._draft_token_req_ids
if req_ids is None:
return [], []
assert self.draft_token_ids_event is not None
assert self.draft_token_ids_cpu is not None
self.draft_token_ids_event.synchronize()
assert isinstance(self._draft_token_ids, torch.Tensor)
num_spec_tokens = self._draft_token_ids.shape[1]
return self.draft_token_ids_cpu[
: len(req_ids), :num_spec_tokens
].tolist(), req_ids
def _copy_valid_sampled_token_count(
self, next_token_ids: torch.Tensor, valid_sampled_tokens_count: torch.Tensor
) -> None:
if self.valid_sampled_token_count_event is None:
return
default_stream = torch.cuda.current_stream()
# Initialize a new stream to overlap the copy operation with
# prepare_input of draft model.
with torch.cuda.stream(self.valid_sampled_token_count_copy_stream):
self.valid_sampled_token_count_copy_stream.wait_stream(default_stream) # type: ignore
counts = valid_sampled_tokens_count
counts_cpu = self.valid_sampled_token_count_cpu
assert counts_cpu is not None
counts_cpu[: counts.shape[0]].copy_(counts, non_blocking=True)
self.valid_sampled_token_count_event.record()
if self.use_async_spec_decode:
# Stash for GPU-side correction in _prepare_inputs.
self.valid_sampled_token_count_gpu = valid_sampled_tokens_count
self.input_batch.prev_sampled_token_ids = next_token_ids.unsqueeze(1)
def _get_valid_sampled_token_count(self) -> list[int]:
# Wait until valid_sampled_tokens_count is copied to cpu,
prev_sampled_token_ids = self.input_batch.prev_sampled_token_ids
sampled_count_event = self.valid_sampled_token_count_event
if sampled_count_event is None or prev_sampled_token_ids is None:
return []
counts_cpu = self.valid_sampled_token_count_cpu
assert counts_cpu is not None
sampled_count_event.synchronize()
return counts_cpu[: prev_sampled_token_ids.shape[0]].tolist()
def _get_spec_decode_draft_probs(
self, spec_decode_metadata: SpecDecodeMetadata
) -> torch.Tensor | None:
if self._draft_probs is None or self._draft_prob_req_ids is None:
return None
row_by_req_id = {
req_id: idx for idx, req_id in enumerate(self._draft_prob_req_ids)
}
draft_probs_rows: list[torch.Tensor] = []
for req_id, num_draft in zip(
self.input_batch.req_ids, spec_decode_metadata.num_draft_tokens
):
if num_draft == 0:
continue
row_idx = row_by_req_id.get(req_id)
if row_idx is None:
logger.warning(
"Missing cached draft probabilities for request %s; "
"falling back to legacy speculative rejection behavior.",
req_id,
)
return None
draft_probs_rows.append(self._draft_probs[row_idx, :num_draft])
if not draft_probs_rows:
return None
return torch.cat(draft_probs_rows, dim=0).contiguous()
def propose_draft_token_ids(
self,
scheduler_output: "SchedulerOutput",
sampled_token_ids: torch.Tensor | list[list[int]],
sampling_metadata: SamplingMetadata,
hidden_states: torch.Tensor,
sample_hidden_states: torch.Tensor,
aux_hidden_states: list[torch.Tensor] | None,
spec_decode_metadata: SpecDecodeMetadata | None,
common_attn_metadata: CommonAttentionMetadata,
slot_mappings: dict[str, torch.Tensor] | list[dict[str, torch.Tensor]] | None,
) -> list[list[int]] | torch.Tensor:
num_scheduled_tokens = scheduler_output.total_num_scheduled_tokens
spec_config = self.speculative_config
assert spec_config is not None
num_spec_tokens_to_schedule = scheduler_output.num_spec_tokens_to_schedule
self._draft_probs = None
self._draft_prob_req_ids = None
if spec_config.method == "ngram":
from vllm.v1.spec_decode.ngram_proposer import NgramProposer
assert isinstance(sampled_token_ids, list)
assert isinstance(self.drafter, NgramProposer)
draft_token_ids = self.drafter.propose(
num_spec_tokens_to_schedule,
sampled_token_ids,
self.input_batch.num_tokens_no_spec,
self.input_batch.token_ids_cpu,
slot_mappings=slot_mappings,
)
elif spec_config.method == "custom_class":
assert isinstance(sampled_token_ids, list)
draft_token_ids = cast(Any, self.drafter).propose(
sampled_token_ids,
self.input_batch.num_tokens_no_spec,
self.input_batch.token_ids_cpu,
slot_mappings=slot_mappings,
)
elif spec_config.use_ngram_gpu():
assert isinstance(self.drafter, NgramProposerGPU)
(
next_token_ids,
valid_sampled_tokens_count,
valid_sampled_token_ids_gpu,
) = self.drafter.update_token_ids_ngram(
sampled_token_ids,
self.input_batch,
self.token_ids_gpu_tensor,
self.num_tokens_no_spec_gpu,
self.discard_request_mask.gpu,
)
self._copy_valid_sampled_token_count(
next_token_ids, valid_sampled_tokens_count
)
batch_size = next_token_ids.shape[0]
draft_token_ids, num_valid_draft_tokens = self.drafter.propose(
num_spec_tokens_to_schedule,
self.num_tokens_no_spec_gpu[:batch_size],
self.token_ids_gpu_tensor[:batch_size],
valid_sampled_token_ids_gpu,
valid_sampled_tokens_count,
)
# Cache valid draft counts for scheduler-side trimming.
self._num_valid_draft_tokens = num_valid_draft_tokens
# Async D2H copy on a dedicated stream.
copy_num_valid_draft_tokens(
self._num_valid_draft_tokens_cpu,
self._num_valid_draft_tokens_copy_stream,
self._num_valid_draft_tokens_event,
self._num_valid_draft_tokens,
self.input_batch.num_reqs,
)
elif spec_config.method == "suffix":
assert isinstance(sampled_token_ids, list)
assert isinstance(self.drafter, SuffixDecodingProposer)
draft_token_ids = self.drafter.propose(
num_spec_tokens_to_schedule,
self.input_batch,
sampled_token_ids,
slot_mappings=slot_mappings,
)
elif spec_config.method == "medusa":
assert isinstance(sampled_token_ids, list)
assert isinstance(self.drafter, MedusaProposer)
if sample_hidden_states.shape[0] == len(sampled_token_ids):
# The input to the target model does not include draft tokens.
hidden_states = sample_hidden_states
else:
indices = []
offset = 0
assert spec_decode_metadata is not None, (
"No spec decode metadata for medusa"
)
for num_draft, tokens in zip(
spec_decode_metadata.num_draft_tokens, sampled_token_ids
):
indices.append(offset + len(tokens) - 1)
offset += num_draft + 1
indices = async_tensor_h2d(indices, device=self.device)
hidden_states = sample_hidden_states[indices]
draft_token_ids = self.drafter.propose(
num_speculative_tokens=num_spec_tokens_to_schedule,
target_hidden_states=hidden_states,
sampling_metadata=sampling_metadata,
slot_mappings=slot_mappings,
)
elif spec_config.uses_extract_hidden_states():
assert isinstance(self.drafter, ExtractHiddenStatesProposer)
assert isinstance(sampled_token_ids, torch.Tensor), (
"sampled_token_ids should be a torch.Tensor for "
"extract_hidden_states method."
)
if not self.use_aux_hidden_state_outputs or aux_hidden_states is None:
raise ValueError(
"aux_hidden_states are required when using `extract_hidden_states`"
)
target_hidden_states = [h[:num_scheduled_tokens] for h in aux_hidden_states]
draft_token_ids = self.drafter.propose(
num_speculative_tokens=num_spec_tokens_to_schedule,
sampled_token_ids=sampled_token_ids,
target_hidden_states=target_hidden_states,
common_attn_metadata=common_attn_metadata,
slot_mappings=slot_mappings,
)
next_token_ids, valid_sampled_tokens_count = (
self.drafter.prepare_next_token_ids_padded(
sampled_token_ids,
self.requests,
self.input_batch,
self.discard_request_mask.gpu,
)
)
self._copy_valid_sampled_token_count(
next_token_ids, valid_sampled_tokens_count
)
elif (
spec_config.use_eagle()
or spec_config.use_dflash()
or spec_config.uses_draft_model()
):
assert isinstance(
self.drafter,
EagleProposer | DFlashProposer | DraftModelProposer | Gemma4Proposer,
)
if spec_config.disable_padded_drafter_batch:
# When padded-batch is disabled, the sampled_token_ids should be
# the cpu-side list[list[int]] of valid sampled tokens for each
# request, with invalid requests having empty lists.
assert isinstance(sampled_token_ids, list), (
"sampled_token_ids should be a python list when"
"padded-batch is disabled."
)
next_token_ids = self.drafter.prepare_next_token_ids_cpu(
sampled_token_ids,
self.requests,
self.input_batch,
scheduler_output.num_scheduled_tokens,
)
else:
# When using padded-batch, the sampled_token_ids should be
# the gpu tensor of sampled tokens for each request, of shape
# (num_reqs, num_spec_tokens + 1) with rejected tokens having
# value -1.
assert isinstance(sampled_token_ids, torch.Tensor), (
"sampled_token_ids should be a torch.Tensor when"
"padded-batch is enabled."
)
next_token_ids, valid_sampled_tokens_count = (
self.drafter.prepare_next_token_ids_padded(
sampled_token_ids,
self.requests,
self.input_batch,
self.discard_request_mask.gpu,
)
)
self._copy_valid_sampled_token_count(
next_token_ids, valid_sampled_tokens_count
)
# Let the target override the hidden state fed to the drafter
# (e.g. DeepSeek V4 MTP needs the pre-hc_head residual). Safe to
# rebind here: hidden_states was already consumed for sampling
# above and is not used again in this branch.
alt = getattr(
self.get_model(), "get_mtp_target_hidden_states", lambda: None
)()
if alt is not None:
hidden_states = alt
num_rejected_tokens_gpu = None
if spec_decode_metadata is None:
token_indices_to_sample = None
# input_ids can be None for multimodal models.
target_token_ids = self.input_ids.gpu[:num_scheduled_tokens]
target_positions = self._get_positions(num_scheduled_tokens)
if self.use_aux_hidden_state_outputs:
assert aux_hidden_states is not None
target_hidden_states = torch.cat(
[h[:num_scheduled_tokens] for h in aux_hidden_states], dim=-1
)
else:
target_hidden_states = hidden_states[:num_scheduled_tokens]
else:
if spec_config.disable_padded_drafter_batch:
token_indices_to_sample = None
common_attn_metadata, token_indices = self.drafter.prepare_inputs(
common_attn_metadata,
sampled_token_ids,
spec_decode_metadata.num_draft_tokens,
)
target_token_ids = self.input_ids.gpu[token_indices]
target_positions = self._get_positions(token_indices)
if self.use_aux_hidden_state_outputs:
assert aux_hidden_states is not None
target_hidden_states = torch.cat(
[h[token_indices] for h in aux_hidden_states], dim=-1
)
else:
target_hidden_states = hidden_states[token_indices]
else:
(
common_attn_metadata,
token_indices_to_sample,
num_rejected_tokens_gpu,
) = self.drafter.prepare_inputs_padded(
common_attn_metadata,
spec_decode_metadata,
valid_sampled_tokens_count,
)
total_num_tokens = common_attn_metadata.num_actual_tokens
# When padding the batch, token_indices is just a range
target_token_ids = self.input_ids.gpu[:total_num_tokens]
target_positions = self._get_positions(total_num_tokens)
if self.use_aux_hidden_state_outputs:
assert aux_hidden_states is not None
target_hidden_states = torch.cat(
[h[:total_num_tokens] for h in aux_hidden_states], dim=-1
)
else:
target_hidden_states = hidden_states[:total_num_tokens]
if self.supports_mm_inputs and self.drafter.supports_mm_inputs:
mm_embed_inputs = self._gather_mm_embeddings(
scheduler_output,
shift_computed_tokens=1,
)
else:
mm_embed_inputs = None
draft_token_ids = self.drafter.propose(
num_speculative_tokens=num_spec_tokens_to_schedule,
target_token_ids=target_token_ids,
target_positions=target_positions,
target_hidden_states=target_hidden_states,
next_token_ids=next_token_ids,
token_indices_to_sample=token_indices_to_sample,
sampling_metadata=sampling_metadata,
common_attn_metadata=common_attn_metadata,
mm_embed_inputs=mm_embed_inputs,
num_rejected_tokens_gpu=num_rejected_tokens_gpu,
slot_mappings=slot_mappings,
)
if hasattr(self.drafter, "take_last_draft_probs"):
draft_probs = self.drafter.take_last_draft_probs()
if draft_probs is not None:
self._draft_probs = draft_probs
self._draft_prob_req_ids = self.input_batch.req_ids.copy()
return draft_token_ids
def update_config(self, overrides: dict[str, Any]) -> None:
allowed_config_names = {"load_config", "model_config"}
for config_name, config_overrides in overrides.items():
assert config_name in allowed_config_names, (
f"Config `{config_name}` not supported. "
f"Allowed configs: {allowed_config_names}"
)
config = getattr(self, config_name)
new_config = update_config(config, config_overrides)
setattr(self, config_name, new_config)
@instrument(span_name="Loading (GPU)")
def load_model(self, load_dummy_weights: bool = False) -> None:
"""
Args:
load_dummy_weights: load dummy weights instead of real weights.
"""
logger.info_once(
"Starting to load model %s...",
self.model_config.model,
scope="global",
)
if self.parallel_config.enable_eplb:
self.eplb_state = EplbState(self.parallel_config, self.device)
eplb_models = 0
try:
with DeviceMemoryProfiler() as m:
time_before_load = time.perf_counter()
if load_dummy_weights:
self.load_config.load_format = "dummy"
model_loader = get_model_loader(self.load_config)
self.model = model_loader.load_model(
vllm_config=self.vllm_config, model_config=self.model_config
)
if self.lora_config:
self.model = self.load_lora_model(
self.model, self.vllm_config, self.device
)
if hasattr(self, "drafter"):
logger.info_once("Loading drafter model...")
if hasattr(self.drafter, "load_model"):
self.drafter.load_model(self.model)
if (
hasattr(self.drafter, "model")
and is_mixture_of_experts(self.drafter.model)
and self.parallel_config.enable_eplb
):
assert not self.parallel_config.enable_elastic_ep, (
"Elastic EP is not supported with drafter model."
)
spec_config = self.vllm_config.speculative_config
assert spec_config is not None
assert spec_config.draft_model_config is not None
logger.info_once(
"EPLB is enabled for drafter model %s.",
spec_config.draft_model_config.model,
)
if self.eplb_state is None:
self.eplb_state = EplbState(
self.parallel_config, self.device
)
self.eplb_state.add_model(
self.drafter.model,
spec_config.draft_model_config,
)
assert hasattr(self.drafter, "set_eplb_state")
self.drafter.set_eplb_state(self.eplb_state)
eplb_models += 1
self._setup_eagle3_aux_hidden_state_outputs()
# Resolve the MoE model, unwrapping VLM wrappers if needed.
# VLM models (e.g. KimiK25ForConditionalGeneration) wrap the
# actual MoE language model but don't implement
# MixtureOfExperts themselves.
moe_candidate = self.model
if not is_mixture_of_experts(moe_candidate) and isinstance(
moe_candidate, SupportsMultiModal
):
moe_candidate = moe_candidate.get_language_model()
if is_mixture_of_experts(moe_candidate):
self._moe_model = moe_candidate
if (
self._moe_model is not None
and self.parallel_config.enable_eplb
and not load_dummy_weights
):
logger.info_once(
"EPLB is enabled for model %s.",
self.model_config.model,
)
assert self.eplb_state is not None
self.eplb_state.add_model(
self._moe_model,
self.model_config,
)
eplb_models += 1
time_after_load = time.perf_counter()
self.model_memory_usage = m.consumed_memory
except torch.cuda.OutOfMemoryError as e:
msg = (
"Failed to load model - not enough GPU memory. "
"Try lowering --gpu-memory-utilization to free memory for weights, "
"increasing --tensor-parallel-size, or using --quantization. "
"See https://docs.vllm.ai/en/latest/configuration/conserving_memory/ "
"for more tips."
)
combined_msg = f"{msg} (original error: {e})"
logger.error(combined_msg)
raise e
logger.info_once(
"Model loading took %s GiB memory and %.6f seconds",
format_gib(self.model_memory_usage),
time_after_load - time_before_load,
)
if not load_dummy_weights:
prepare_communication_buffer_for_model(self.model)
if (drafter := getattr(self, "drafter", None)) and (
drafter_model := getattr(drafter, "model", None)
):
prepare_communication_buffer_for_model(drafter_model)
mm_config = self.model_config.multimodal_config
self.is_multimodal_pruning_enabled = (
supports_multimodal_pruning(self.get_model())
and mm_config is not None
and mm_config.is_multimodal_pruning_enabled()
)
self.requires_sequential_video_encoding = hasattr(
self.get_model(), "requires_sequential_video_encoding"
) # Temporary hack for dynamic res video w/o support for bs>1 yet
if (
self._moe_model is not None
and self.parallel_config.enable_eplb
and not load_dummy_weights
and self.eplb_state is not None
and self.eplb_state.is_async
):
self.eplb_state.start_async_loop()
if (
self.vllm_config.compilation_config.mode
== CompilationMode.STOCK_TORCH_COMPILE
):
from vllm.env_override import _apply_constrain_to_fx_strides_patch
_apply_constrain_to_fx_strides_patch()
backend = self.vllm_config.compilation_config.init_backend(self.vllm_config)
compilation_counter.stock_torch_compile_count += 1
self.model.compile(fullgraph=True, backend=backend)
return
# for other compilation modes, cudagraph behavior is controlled by
# CudagraphWrapper and CudagraphDispatcher of vllm.
# wrap the model with full cudagraph wrapper if needed.
cudagraph_mode = self.compilation_config.cudagraph_mode
assert cudagraph_mode is not None
if (
is_breakable_cudagraph_enabled()
and cudagraph_mode != CUDAGraphMode.NONE
and not self.parallel_config.use_ubatching
):
self.model = BreakableCUDAGraphWrapper(self.model, self.vllm_config)
drafter = getattr(self, "drafter", None)
if drafter is not None and hasattr(drafter, "model"):
drafter.model = BreakableCUDAGraphWrapper(
drafter.model, self.vllm_config
)
elif (
cudagraph_mode.has_full_cudagraphs()
and not self.parallel_config.use_ubatching
):
self.model = CUDAGraphWrapper(
self.model, self.vllm_config, runtime_mode=CUDAGraphMode.FULL
)
elif self.parallel_config.use_ubatching:
if cudagraph_mode.has_full_cudagraphs():
self.model = UBatchWrapper(
self.model, self.vllm_config, CUDAGraphMode.FULL, self.device
)
else:
self.model = UBatchWrapper(
self.model, self.vllm_config, CUDAGraphMode.NONE, self.device
)
get_offloader().post_init()
def _setup_eagle3_aux_hidden_state_outputs(self) -> None:
if not self.use_aux_hidden_state_outputs:
return
if not supports_eagle3(self.get_model()):
raise RuntimeError(
"Model does not support EAGLE3 interface but "
"aux_hidden_state_outputs was requested"
)
# Try to get auxiliary layers from speculative config,
# otherwise use model's default layers
aux_layers = self._get_eagle3_aux_layers_from_config()
if aux_layers:
logger.info(
"Using auxiliary layers from speculative config: %s", aux_layers
)
else:
aux_layers = self.model.get_eagle3_default_aux_hidden_state_layers()
self.model.set_aux_hidden_state_layers(aux_layers)
def _get_eagle3_aux_layers_from_config(self) -> tuple[int, ...] | None:
"""Extract Eagle3 auxiliary layer indices from speculative config.
These indices specify which hidden states from the base model should
be used as auxiliary inputs for the Eagle3 drafter model during
speculative decoding.
Returns:
Tuple of layer indices if found in draft model config,
None otherwise.
"""
if not (self.speculative_config and self.speculative_config.draft_model_config):
return None
hf_config = self.speculative_config.draft_model_config.hf_config
layer_ids = getattr(hf_config, "eagle_aux_hidden_state_layer_ids", None)
if not layer_ids:
dflash_config = getattr(hf_config, "dflash_config", None)
eagle_config = getattr(hf_config, "eagle_config", None)
if dflash_config and isinstance(dflash_config, dict):
# Add 1 to convert DFlash's aux layer id semantics
layer_ids = [
i + 1 for i in (dflash_config.get("target_layer_ids") or [])
]
if eagle_config and isinstance(eagle_config, dict):
layer_ids = eagle_config.get("eagle_aux_hidden_state_layer_ids")
if layer_ids and isinstance(layer_ids, (list, tuple)):
return tuple(layer_ids)
return None
def reload_weights(
self,
weights_iterator: Iterable[tuple[str, torch.Tensor]] | None = None,
weights_path: str | None = None,
is_checkpoint_format: bool = True,
) -> None:
"""
Reload weights from a weights iterator or from disk
Args:
weights_iterator: weights to load into model
weights_path: path to load weights from if weights_iterator is not
provided. Use path of original model if neither is provided.
is_checkpoint_format: set to False if weights have already been
processed into kernel format (repacking, renaming, etc.)
"""
# TODO(@kylesayrs): generalize to all runners and loaders
# argument validation
if weights_iterator is None and not is_checkpoint_format:
logger.warning(
"Reloading from disk means that weights will be in checkpoint format. "
"Please use `is_checkpoint_format=True` "
"to avoid weight reloading errors"
)
model = self.get_model()
weights_to_load = {name for name, _ in model.named_parameters()}
counter_before_reloading = time.perf_counter()
# load weights from disk if none are provided
if weights_iterator is None:
model_loader = get_model_loader(self.load_config)
if not hasattr(model_loader, "get_all_weights"):
raise NotImplementedError(
f"Model reloading with `{self.load_config.load_format}` format"
)
if weights_path is not None:
self.model_config.model = weights_path
weights_iterator = model_loader.get_all_weights(self.model_config, model)
weights_iterator = cast(
Iterable[tuple[str, torch.Tensor]], weights_iterator
)
# begin loading weights
logger.info_once("Reloading weights inplace...")
if is_checkpoint_format:
# load weights from checkpoint/ original model format
initialize_layerwise_reload(model)
loaded_weights = model.load_weights(weights_iterator)
finalize_layerwise_reload(model, self.model_config)
else:
# load weights from kernel format
logger.warning_once(
"Reloading with `is_checkpoint_format=True` requires that "
"weights be in kernel format and already sharded",
)
loaded_weights = set()
for name, loaded_weight in weights_iterator:
param = model.get_parameter(name) # TODO: buffers?
param.copy_(loaded_weight)
loaded_weights.add(name)
# logging and validation
counter_after_reloading = time.perf_counter()
diff_seconds = counter_after_reloading - counter_before_reloading
logger.info_once(
"Reloading and processing weights took %.2f seconds",
diff_seconds,
)
if self.model_config.quantization is None and loaded_weights is not None:
weights_not_loaded = weights_to_load - loaded_weights
if weights_not_loaded:
logger.warning(
"Following weights were not loaded from checkpoint: %s",
weights_not_loaded,
)
self.reset_encoder_cache()
self.reset_mm_cache()
def _get_prompt_logprobs_dict(
self,
hidden_states: torch.Tensor,
num_scheduled_tokens: dict[str, int],
) -> dict[str, LogprobsTensors | None]:
num_prompt_logprobs_dict = self.num_prompt_logprobs
if not num_prompt_logprobs_dict:
return {}
prompt_logprobs_dict: dict[str, LogprobsTensors | None] = {}
# Since prompt logprobs are a rare feature, prioritize simple,
# maintainable loop over optimal performance.
completed_prefill_reqs = []
for req_id, num_prompt_logprobs in num_prompt_logprobs_dict.items():
num_tokens = num_scheduled_tokens.get(req_id)
if num_tokens is None:
# This can happen if the request was preempted in prefill stage.
continue
# Get metadata for this request.
request = self.requests[req_id]
if request.prompt_token_ids is None:
# Prompt logprobs is incompatible with prompt embeddings
continue
num_prompt_tokens = len(request.prompt_token_ids)
prompt_token_ids = async_tensor_h2d(
request.prompt_token_ids, device=self.device
)
# Set up target LogprobsTensors object.
logprobs_tensors = request.in_progress_prompt_logprobs_cpu
if logprobs_tensors is None:
# Create empty logprobs CPU tensors for the entire prompt.
# If chunked, we'll copy in slice by slice.
logprobs_tensors = LogprobsTensors.empty_cpu(
num_prompt_tokens - 1, num_prompt_logprobs + 1
)
request.in_progress_prompt_logprobs_cpu = logprobs_tensors
# Determine number of logits to retrieve.
start_idx = request.num_computed_tokens
start_tok = start_idx + 1
num_remaining_tokens = num_prompt_tokens - start_tok
if num_tokens <= num_remaining_tokens:
# This is a chunk, more tokens remain.
# In the == case, there are no more prompt logprobs to produce
# but we want to defer returning them to the next step where we
# have new generated tokens to return.
num_logits = num_tokens
else:
# This is the last chunk of prompt tokens to return.
num_logits = num_remaining_tokens
completed_prefill_reqs.append(req_id)
prompt_logprobs_dict[req_id] = logprobs_tensors
if num_logits <= 0:
# This can happen for the final chunk if we prefilled exactly
# (num_prompt_tokens - 1) tokens for this request in the prior
# step. There are no more prompt logprobs to produce.
continue
# Get the logits corresponding to this req's prompt tokens.
# If this is a partial request (i.e. chunked prefill),
# then there is prompt logprob generated for each index.
req_idx = self.input_batch.req_id_to_index[req_id]
offset = self.query_start_loc.np[req_idx].item()
prompt_hidden_states = hidden_states[offset : offset + num_logits]
logits = self.model.compute_logits(prompt_hidden_states)
# Get the "target" tokens for each index. For prompt at index i,
# the token at prompt index i+1 is the "sampled" token we want
# to gather the logprob for.
tgt_token_ids = prompt_token_ids[start_tok : start_tok + num_logits]
# Compute prompt logprobs.
logprobs = self.sampler.compute_logprobs(logits)
token_ids, logprobs, ranks, _ = self.sampler.gather_logprobs(
logprobs, num_prompt_logprobs, tgt_token_ids
)
# Transfer GPU->CPU async.
chunk_slice = slice(start_idx, start_idx + num_logits)
logprobs_tensors.logprob_token_ids[chunk_slice].copy_(
token_ids, non_blocking=True
)
logprobs_tensors.logprobs[chunk_slice].copy_(logprobs, non_blocking=True)
logprobs_tensors.selected_token_ranks[chunk_slice].copy_(
ranks, non_blocking=True
)
# Remove requests that have completed prefill from the batch
# num_prompt_logprobs_dict.
for req_id in completed_prefill_reqs:
del num_prompt_logprobs_dict[req_id]
self.requests[req_id].in_progress_prompt_logprobs_cpu = None
# Must synchronize the non-blocking GPU->CPU transfers.
if prompt_logprobs_dict:
self._sync_device()
return prompt_logprobs_dict
def _get_nans_in_logits(
self,
logits: torch.Tensor | None,
) -> dict[str, int]:
try:
if logits is None:
return {req_id: 0 for req_id in self.input_batch.req_ids}
num_nans_in_logits = {}
num_nans_for_index = logits.isnan().sum(dim=-1).cpu().numpy()
for req_id in self.input_batch.req_ids:
req_index = self.input_batch.req_id_to_index[req_id]
num_nans_in_logits[req_id] = (
int(num_nans_for_index[req_index])
if num_nans_for_index is not None and req_index < logits.shape[0]
else 0
)
return num_nans_in_logits
except IndexError:
return {}
@contextmanager
def maybe_randomize_inputs(
self, input_ids: torch.Tensor | None, inputs_embeds: torch.Tensor | None
):
"""
Randomize input_ids if VLLM_RANDOMIZE_DP_DUMMY_INPUTS is set.
This is to help balance expert-selection
- during profile_run
- during DP rank dummy run
"""
dp_size = self.vllm_config.parallel_config.data_parallel_size
randomize_inputs = envs.VLLM_RANDOMIZE_DP_DUMMY_INPUTS and dp_size > 1
if not randomize_inputs:
yield
elif input_ids is not None:
@functools.cache
def rand_input_ids() -> torch.Tensor:
return torch.randint_like(
self.input_ids.gpu,
low=0,
high=self.model_config.get_vocab_size(),
)
logger.debug_once("Randomizing dummy input_ids for DP Rank")
input_ids.copy_(rand_input_ids()[: input_ids.size(0)], non_blocking=True)
yield
input_ids.fill_(0)
else:
@functools.cache
def rand_inputs_embeds() -> torch.Tensor:
return torch.randn_like(
self.inputs_embeds.gpu,
)
assert inputs_embeds is not None
logger.debug_once("Randomizing dummy inputs_embeds for DP Rank")
inputs_embeds.copy_(
rand_inputs_embeds()[: inputs_embeds.size(0)], non_blocking=True
)
yield
inputs_embeds.fill_(0)
def _get_mm_dummy_batch(
self,
modality: str,
max_items_per_batch: int,
) -> BatchedTensorInputs:
"""Dummy data for profiling and precompiling multimodal models."""
assert self.mm_budget is not None
# Don't use `max_items_per_batch` here to avoid redundant computation
dummy_mm_inputs = self.mm_registry.get_dummy_mm_inputs(
self.model_config,
mm_counts={modality: 1},
cache=self.mm_budget.cache,
)
dummy_mm_item = dummy_mm_inputs["mm_kwargs"][modality][0]
# We use the cache so that the item is saved to the cache,
# but not read from the cache
assert dummy_mm_item is not None, "Item should not already be cached"
return next(
mm_kwargs_batch
for _, _, mm_kwargs_batch in group_and_batch_mm_kwargs(
[(modality, dummy_mm_item)] * max_items_per_batch,
device=self.device,
pin_memory=PIN_MEMORY,
)
)
@torch.inference_mode()
def _dummy_run(
self,
num_tokens: int,
cudagraph_runtime_mode: CUDAGraphMode | None = None,
force_attention: bool = False,
uniform_decode: bool = False,
allow_microbatching: bool = True,
skip_eplb: bool = False,
is_profile: bool = False,
create_mixed_batch: bool = False,
remove_lora: bool = True,
is_graph_capturing: bool = False,
num_active_loras: int = 0,
profile_seq_lens: int | None = None,
) -> tuple[torch.Tensor, torch.Tensor]:
"""
Run a dummy forward pass to warm up/profile run or capture the
CUDA graph for the model.
Args:
num_tokens: Number of tokens to run the dummy forward pass.
cudagraph_runtime_mode: used to control the behavior.
- if not set will determine the cudagraph mode based on using
the self.cudagraph_dispatcher.
- CUDAGraphMode.NONE: No cudagraph, for warm up and profile run
- CUDAGraphMode.PIECEWISE: Piecewise cudagraph.
- CUDAGraphMode.FULL: Full cudagraph, attention metadata is
needed.
force_attention: If True, always create attention metadata. Used to
warm up attention backend when mode is NONE.
uniform_decode: If True, the batch is a uniform decode batch.
skip_eplb: If True, skip EPLB state update.
is_profile: If True, this is a profile run.
create_mixed_batch: If True, create a mixed batch with both decode
(1 token) and prefill (multiple tokens) requests.
remove_lora: If False, dummy LoRAs are not destroyed after the run
num_active_loras: Number of distinct active LoRAs to capture for.
LoRA is activated when num_active_loras > 0.
profile_seq_lens: If provided, use this value for seq_lens instead
of max_query_len. Used to profile attention workspace that
scales with context length.
"""
mm_config = self.vllm_config.model_config.multimodal_config
if mm_config and mm_config.mm_encoder_only:
# The current dummy run only covers LM execution, so we can skip it.
# mm encoder dummy run may need to add in the future.
return torch.tensor([]), torch.tensor([])
assert (
cudagraph_runtime_mode is None
or cudagraph_runtime_mode.is_valid_runtime_mode()
)
# If cudagraph_mode.decode_mode() == FULL and
# cudagraph_mode.separate_routine(). This means that we are using
# different graphs and/or modes for mixed prefill-decode batches vs.
# uniform decode batches. A uniform decode batch means that all
# requests have identical query length, except a potential virtual
# request (shorter) in the batch account for padding.
# Uniform decode batch could either be common pure decode, where
# max_query_len == 1, or speculative decode, where
# max_query_len == 1 + num_spec_decode_tokens.
# When setting max_query_len = 1, we switch to and capture the optimized
# routine of FA2 for pure decode, i.e., Flashdecode + an optimization
# for GQA/MQA.
max_query_len = self.uniform_decode_query_len if uniform_decode else num_tokens
# Set num_scheduled_tokens based on num_tokens and max_num_seqs
# for dummy run with LoRA so that the num_reqs collectively
# has num_tokens in total.
assert num_tokens <= self.max_num_tokens
max_num_reqs = self.scheduler_config.max_num_seqs
if create_mixed_batch:
assert not uniform_decode
# Create mixed batch:
# first half decode tokens, second half one prefill
num_decode_tokens = min(max_num_reqs - 1, num_tokens // 2)
num_prefill_tokens = num_tokens - num_decode_tokens
num_reqs = num_decode_tokens + 1
# Create decode requests (1 token each) followed by prefill request
num_scheduled_tokens_list = [1] * num_decode_tokens + [num_prefill_tokens]
# Note: Overriding max_query_len to be the prefill tokens
max_query_len = num_prefill_tokens
elif uniform_decode:
assert not create_mixed_batch
num_reqs = min(max_num_reqs, cdiv(num_tokens, max_query_len))
num_scheduled_tokens_list = [max_query_len] * num_reqs
if num_tokens % max_query_len != 0:
num_scheduled_tokens_list[-1] = num_tokens % max_query_len
else:
num_reqs = min(num_tokens, max_num_reqs)
min_tokens_per_req = num_tokens // num_reqs
num_scheduled_tokens_list = [min_tokens_per_req] * num_reqs
num_scheduled_tokens_list[-1] += num_tokens % num_reqs
assert sum(num_scheduled_tokens_list) == num_tokens
assert len(num_scheduled_tokens_list) == num_reqs
num_scheduled_tokens = np.array(num_scheduled_tokens_list, dtype=np.int32)
num_tokens_unpadded = int(num_scheduled_tokens.sum())
num_sampled_tokens = np.ones(num_reqs, dtype=np.int32)
_cudagraph_mode, batch_desc, should_ubatch, num_tokens_across_dp, _ = (
self._determine_batch_execution_and_padding(
num_tokens=num_tokens_unpadded,
num_reqs=num_reqs,
num_scheduled_tokens_np=num_scheduled_tokens,
max_num_scheduled_tokens=max_query_len,
use_cascade_attn=False,
allow_microbatching=allow_microbatching,
force_eager=is_profile
or (cudagraph_runtime_mode == CUDAGraphMode.NONE),
# `force_uniform_decode` is used for cudagraph capture; because for
# capturing mixed prefill-decode batches, we sometimes use
# num_tokens == num_reqs which looks like a uniform decode batch to the
# dispatcher; but we actually want to capture a piecewise cudagraph
force_uniform_decode=uniform_decode,
# `force_has_lora` is used for cudagraph capture; because LoRA is
# activated later in the context manager, but we need to know the
# LoRA state when determining the batch descriptor for capture
force_has_lora=num_active_loras > 0,
# `force_num_active_loras` is used for cudagraph capture; because we
# need to capture graphs for specific num_active_loras counts
force_num_active_loras=num_active_loras,
)
)
if cudagraph_runtime_mode is None:
cudagraph_runtime_mode = _cudagraph_mode
else:
assert cudagraph_runtime_mode == _cudagraph_mode, (
f"Cudagraph runtime mode mismatch in dummy_run. "
f"Expected {_cudagraph_mode}, but got {cudagraph_runtime_mode}."
)
num_tokens_padded = batch_desc.num_tokens
num_reqs_padded = (
batch_desc.num_reqs if batch_desc.num_reqs is not None else num_reqs
)
dcp_dummy_context_len = get_dcp_dummy_context_len(
self.dcp_world_size,
self.parallel_config.cp_kv_cache_interleave_size,
hasattr(self, "kv_cache_config"),
create_mixed_batch,
is_graph_capturing,
uniform_decode,
)
ubatch_slices, ubatch_slices_padded = maybe_create_ubatch_slices(
should_ubatch,
num_scheduled_tokens,
num_tokens_padded,
num_reqs_padded,
self.vllm_config.parallel_config.num_ubatches,
)
logger.debug(
"ubatch_slices: %s, ubatch_slices_padded: %s",
ubatch_slices,
ubatch_slices_padded,
)
attn_metadata: PerLayerAttnMetadata | None = None
slot_mappings_by_group, slot_mappings = self._get_slot_mappings(
num_tokens_padded=num_tokens_padded,
num_reqs_padded=num_reqs_padded,
num_tokens_unpadded=num_tokens_unpadded,
ubatch_slices=ubatch_slices_padded,
)
# Dummy runs have no real slot assignments — fill with -1 so
# concat_and_cache kernels skip the KV write.
if slot_mappings_by_group is not None:
for sm in slot_mappings_by_group.values():
sm.fill_(-1)
# _dummy_run shares pinned CPU buffers (seq_lens, query_start_loc,
# etc.) with execute_model. It must participate in the same event
# protocol so that back-to-back dummy/real steps don't overwrite
# pinned memory while a prior non_blocking H2D DMA is still reading.
with self.synchronize_input_prep():
# If force_attention is True, we always capture attention.
# Otherwise, it only happens for cudagraph_runtime_mode=FULL.
if force_attention or cudagraph_runtime_mode == CUDAGraphMode.FULL:
if profile_seq_lens is not None:
seq_lens = profile_seq_lens # type: ignore[assignment]
elif create_mixed_batch:
# In the mixed batch mode (used for FI warmup), we use
# shorter sequence lengths to run faster.
# TODO(luka) better system for describing dummy batches
if dcp_dummy_context_len > 0:
seq_lens = torch.tensor( # type: ignore[assignment]
[1 + dcp_dummy_context_len] * num_decode_tokens
+ [num_prefill_tokens + dcp_dummy_context_len],
dtype=torch.int,
)
else:
seq_lens = torch.tensor( # type: ignore[assignment]
[1] * num_decode_tokens + [num_prefill_tokens + 1],
dtype=torch.int,
)
elif dcp_dummy_context_len > 0:
seq_lens = max_query_len + dcp_dummy_context_len # type: ignore[assignment]
else:
seq_lens = max_query_len # type: ignore[assignment]
self.optimistic_seq_lens_cpu[:num_reqs] = seq_lens
self.optimistic_seq_lens_cpu[num_reqs:].fill_(0)
self.seq_lens.copy_(self.optimistic_seq_lens_cpu, non_blocking=True)
cum_num_tokens = self._get_cumsum_and_arange(
num_scheduled_tokens, self.query_pos.np
)
self.query_start_loc.np[1 : num_reqs + 1] = cum_num_tokens
self.query_start_loc.np[num_reqs + 1 : num_reqs_padded + 1].fill(
cum_num_tokens[-1]
)
self.query_start_loc.copy_to_gpu()
prepare_dcp_dummy_context_metadata(
input_batch=self.input_batch,
kv_cache_config=getattr(self, "kv_cache_config", None),
query_pos=self.query_pos,
positions=self.positions,
query_start_loc=self.query_start_loc,
num_reqs=num_reqs,
num_tokens_unpadded=num_tokens_unpadded,
dcp_dummy_context_len=dcp_dummy_context_len,
)
# Sync block table CPU->GPU so cleared rows from
# remove_request() are visible to the attention metadata
# builder. Without this, stale block IDs from finished
# requests can corrupt Mamba state.
self.input_batch.block_table.commit_block_table(num_reqs_padded)
pad_attn = cudagraph_runtime_mode == CUDAGraphMode.FULL
attn_metadata, _ = self._build_attention_metadata(
num_tokens=num_tokens_unpadded,
num_tokens_padded=num_tokens_padded if pad_attn else None,
num_reqs=num_reqs_padded,
max_query_len=max_query_len,
ubatch_slices=(ubatch_slices_padded if pad_attn else ubatch_slices),
for_cudagraph_capture=is_graph_capturing,
slot_mappings=slot_mappings_by_group,
use_spec_decode=self.speculative_config is not None,
)
with self.maybe_dummy_run_with_lora(
self.lora_config,
num_scheduled_tokens,
num_sampled_tokens,
remove_lora,
num_active_loras,
):
# Make sure padding doesn't exceed max_num_tokens
assert num_tokens_padded <= self.max_num_tokens
model_kwargs = self._init_model_kwargs()
if self.supports_mm_inputs and not self.model_config.is_encoder_decoder:
input_ids, inputs_embeds = self._prepare_mm_inputs(num_tokens_padded)
model_kwargs = {
**model_kwargs,
**self._dummy_mm_kwargs(num_reqs),
}
elif self.enable_prompt_embeds:
input_ids = None
inputs_embeds = self.inputs_embeds.gpu[:num_tokens_padded]
model_kwargs = self._init_model_kwargs()
else:
input_ids = self.input_ids.gpu[:num_tokens_padded]
inputs_embeds = None
if self.uses_mrope:
positions = self.mrope_positions.gpu[:, :num_tokens_padded]
elif self.uses_xdrope_dim > 0:
positions = self.xdrope_positions.gpu[:, :num_tokens_padded]
else:
positions = self.positions[:num_tokens_padded]
if get_pp_group().is_first_rank:
intermediate_tensors = None
else:
if self.intermediate_tensors is None:
self.intermediate_tensors = (
self.model.make_empty_intermediate_tensors(
batch_size=self.max_num_tokens,
dtype=self.model_config.dtype,
device=self.device,
)
)
intermediate_tensors = self.sync_and_gather_intermediate_tensors(
num_tokens_padded, None, False
)
if ubatch_slices_padded is not None:
# Adjust values to reflect a single ubatch.
# TODO(sage,lucas): this is cruft that should be addressed in
# the padding refactor.
num_tokens_padded = ubatch_slices_padded[0].num_tokens
if num_tokens_across_dp is not None:
num_tokens_across_dp[:] = num_tokens_padded
with (
self.maybe_randomize_inputs(input_ids, inputs_embeds),
set_forward_context(
attn_metadata,
self.vllm_config,
num_tokens=num_tokens_padded,
num_tokens_across_dp=num_tokens_across_dp,
cudagraph_runtime_mode=cudagraph_runtime_mode,
batch_descriptor=batch_desc,
ubatch_slices=ubatch_slices_padded,
slot_mapping=slot_mappings,
),
):
outputs = self.model(
input_ids=input_ids,
positions=positions,
intermediate_tensors=intermediate_tensors,
inputs_embeds=inputs_embeds,
**model_kwargs,
)
if self.use_aux_hidden_state_outputs:
hidden_states, _ = outputs
else:
hidden_states = outputs
if self.speculative_config and (
self.speculative_config.use_eagle()
or self.speculative_config.uses_draft_model()
or self.speculative_config.uses_extract_hidden_states()
):
assert isinstance(
self.drafter,
EagleProposer
| DFlashProposer
| DraftModelProposer
| ExtractHiddenStatesProposer
| Gemma4Proposer,
)
assert self.speculative_config is not None
# Eagle currently only supports PIECEWISE cudagraphs.
# Therefore only use cudagraphs if the main model uses PIECEWISE
# NOTE(lucas): this is a hack, need to clean up.
use_cudagraphs = (
(
is_graph_capturing
and cudagraph_runtime_mode == CUDAGraphMode.PIECEWISE
)
or (
not is_graph_capturing
and cudagraph_runtime_mode != CUDAGraphMode.NONE
)
) and not self.speculative_config.enforce_eager
# Note(gnovack) - We need to disable cudagraphs for one of the two
# lora cases when cudagraph_specialize_lora is enabled. This is a
# short term mitigation for issue mentioned in
# https://github.com/vllm-project/vllm/issues/28334
if (
self.compilation_config.cudagraph_specialize_lora
and num_active_loras > 0
):
use_cudagraphs = False
self.drafter.dummy_run(
num_tokens,
use_cudagraphs=use_cudagraphs,
is_graph_capturing=is_graph_capturing,
slot_mappings=slot_mappings,
)
# We register layerwise NVTX hooks here after the first dynamo tracing is
# done to avoid nvtx operations in hook functions being traced by
# torch dynamo and causing graph breaks.
# Note that for DYNAMO_ONCE and VLLM_COMPILE mode,
# compiled model's dynamo tracing is only done once and the compiled model's
# __call__ function is replaced by calling the compiled function.
# So it's safe to register hooks here. Hooks will be registered to
# both compiled and uncompiled models but they will never
# be called on the compiled model execution path.
self._register_layerwise_nvtx_hooks()
# This is necessary to avoid blocking DP.
# For dummy runs, we typically skip EPLB since we don't have any real
# requests to process.
# However, in DP settings, there may be cases when some DP ranks do
# not have any requests to process, so they're executing dummy batches.
# In such cases, we still have to trigger EPLB to make sure
# ranks execute the rearrangement in synchronization.
if not skip_eplb:
self.eplb_step(is_dummy=True, is_profile=is_profile)
logit_indices = np.cumsum(num_scheduled_tokens) - 1
logit_indices_device = torch.from_numpy(logit_indices).to(
self.device, non_blocking=True
)
return hidden_states, hidden_states[logit_indices_device]
@torch.inference_mode()
def _dummy_sampler_run(
self,
hidden_states: torch.Tensor,
) -> torch.Tensor:
# The dummy hidden states may contain special values,
# like `inf` or `nan`.
# To avoid breaking the sampler, we use a random tensor here instead.
mm_config = self.vllm_config.model_config.multimodal_config
if mm_config and mm_config.mm_encoder_only:
# MM Encoder only model no need to run sampler.
return torch.tensor([])
hidden_states = torch.rand_like(hidden_states)
logits = self.model.compute_logits(hidden_states)
num_reqs = logits.size(0)
dummy_tensors = lambda v: torch.full((num_reqs,), v, device=self.device)
dummy_metadata = SamplingMetadata(
temperature=dummy_tensors(0.5),
all_greedy=False,
all_random=False,
top_p=dummy_tensors(0.9),
top_k=dummy_tensors(logits.size(1) - 1),
generators={},
max_num_logprobs=None,
logprob_token_ids=None,
no_penalties=True,
prompt_token_ids=None,
frequency_penalties=dummy_tensors(0.1),
presence_penalties=dummy_tensors(0.1),
repetition_penalties=dummy_tensors(0.1),
output_token_ids=[[] for _ in range(num_reqs)],
spec_token_ids=[[] for _ in range(num_reqs)],
allowed_token_ids_mask=None,
bad_words_token_ids={},
logitsprocs=LogitsProcessors(),
)
try:
sampler_output = self.sampler(
logits=logits, sampling_metadata=dummy_metadata
)
# Also warm forward_native (taken when generators dict is non-empty),
# but skip the extra call in 'processed_logits' / 'processed_logprobs'
# modes — there TopKTopPSampler binds forward = forward_native at
# init time, so the warmup call is redundant and only inflates peak
# memory during profile_run.
# No .clone() of logits: warmup output is discarded, so any in-place
# mutation by forward_native does not affect correctness.
if self.sampler.logprobs_mode not in (
"processed_logits",
"processed_logprobs",
):
self.sampler(
logits=logits,
sampling_metadata=replace(
dummy_metadata,
generators={
0: torch.Generator(device=self.device).manual_seed(0)
},
),
)
except RuntimeError as e:
if "out of memory" in str(e):
raise RuntimeError(
"CUDA out of memory occurred when warming up sampler with "
f"{num_reqs} dummy requests. Please try lowering "
"`max_num_seqs` or `gpu_memory_utilization` when "
"initializing the engine."
) from e
else:
raise e
if self.speculative_config:
draft_token_ids = [[0] for _ in range(num_reqs)]
dummy_spec_decode_metadata = SpecDecodeMetadata.make_dummy(
draft_token_ids, self.device
)
num_tokens = sum(len(ids) for ids in draft_token_ids)
draft_probs = None
if (
self.speculative_config.rejection_sample_method == "standard"
and self.speculative_config.draft_sample_method == "probabilistic"
):
draft_probs = torch.rand(
num_tokens,
logits.shape[-1],
device=self.device,
dtype=torch.float32,
)
draft_probs = torch.softmax(draft_probs, dim=-1)
logits = torch.randn(
num_tokens + num_reqs,
logits.shape[-1],
device=self.device,
dtype=logits.dtype,
)
self.rejection_sampler(
dummy_spec_decode_metadata,
draft_probs,
logits,
dummy_metadata,
)
return sampler_output
def _dummy_pooler_run_task(
self,
hidden_states: torch.Tensor,
task: PoolingTask,
) -> PoolerOutput:
num_tokens = hidden_states.shape[0]
max_num_reqs = self.scheduler_config.max_num_seqs
num_reqs = min(num_tokens, max_num_reqs)
min_tokens_per_req = num_tokens // num_reqs
num_scheduled_tokens_np = np.full(num_reqs, min_tokens_per_req)
num_scheduled_tokens_np[-1] += num_tokens % num_reqs
assert np.sum(num_scheduled_tokens_np) == num_tokens
assert len(num_scheduled_tokens_np) == num_reqs
req_num_tokens = num_tokens // num_reqs
dummy_prompt_lens = torch.from_numpy(num_scheduled_tokens_np)
dummy_token_ids = torch.zeros(
(num_reqs, req_num_tokens), dtype=torch.int32, device=self.device
)
model = cast(VllmModelForPooling, self.get_model())
dummy_pooling_params = PoolingParams(task=task)
dummy_pooling_params.verify(self.model_config)
to_update = model.pooler.get_pooling_updates(task)
to_update.apply(dummy_pooling_params)
dummy_metadata = PoolingMetadata(
prompt_lens=dummy_prompt_lens,
prompt_token_ids=dummy_token_ids,
prompt_token_ids_cpu=dummy_token_ids.cpu(),
pooling_params=[dummy_pooling_params] * num_reqs,
pooling_states=[PoolingStates() for i in range(num_reqs)],
)
dummy_metadata.build_pooling_cursor(
num_scheduled_tokens_np,
seq_lens_cpu=dummy_prompt_lens,
device=hidden_states.device,
)
try:
return model.pooler(
hidden_states=hidden_states, pooling_metadata=dummy_metadata
)
except RuntimeError as e:
if "out of memory" in str(e):
raise RuntimeError(
"CUDA out of memory occurred when warming up pooler "
f"({task=}) with {num_reqs} dummy requests. Please try "
"lowering `max_num_seqs` or `gpu_memory_utilization` when "
"initializing the engine."
) from e
else:
raise e
@torch.inference_mode()
def _dummy_pooler_run(
self,
hidden_states: torch.Tensor,
) -> PoolerOutput:
mm_config = self.vllm_config.model_config.multimodal_config
if mm_config and mm_config.mm_encoder_only:
# MM Encoder only model not need to run pooler.
return torch.tensor([])
# Find the task that has the largest output for subsequent steps
supported_pooling_tasks = self.get_supported_pooling_tasks()
if not supported_pooling_tasks:
raise RuntimeError(
f"Model {self.model_config.model} does not support "
"any pooling tasks. See "
"https://docs.vllm.ai/en/latest/models/pooling_models.html "
"to learn more."
)
output_size = dict[PoolingTask, float]()
for task in supported_pooling_tasks:
# Run a full batch with each task to ensure none of them OOMs
output = self._dummy_pooler_run_task(hidden_states, task)
output_size[task] = sum(o.nbytes for o in output if o is not None)
del output # Allow GC
max_task = max(output_size.items(), key=lambda x: x[1])[0]
return self._dummy_pooler_run_task(hidden_states, max_task)
def profile_run(self) -> None:
# Profile with multimodal encoder & encoder cache.
if self.supports_mm_inputs:
mm_config = self.model_config.multimodal_config
if mm_config is not None and mm_config.skip_mm_profiling:
logger.info(
"Skipping memory profiling for multimodal encoder and "
"encoder cache."
)
else:
mm_budget = self.mm_budget
assert mm_budget is not None
if (encoder_budget := mm_budget.get_encoder_budget()) > 0:
if not mm_budget.mm_max_toks_per_item:
# All modality limits are 0 — embedding-only mode.
# Budget is non-zero for embedding storage, but
# there's no encoder to profile.
logger.info(
"Skipping encoder profiling for embedding-only "
"mode (all modality limits=0 with "
"enable_mm_embeds=True).",
)
else:
# NOTE: Currently model is profiled with a single
# non-text modality with the max possible input
# tokens even when it supports multiple.
dummy_modality = mm_budget.get_modality_with_max_tokens()
max_mm_items_per_batch = mm_budget.mm_max_items_per_batch[
dummy_modality
]
logger.info_once(
"Encoder cache will be initialized with a "
"budget of %s tokens, and profiled with "
"%s %s items of the maximum feature size.",
encoder_budget,
max_mm_items_per_batch,
dummy_modality,
)
# Create dummy batch of multimodal inputs.
batched_dummy_mm_inputs = self._get_mm_dummy_batch(
dummy_modality,
max_mm_items_per_batch,
)
# Run multimodal encoder.
dummy_encoder_outputs = self.model.embed_multimodal(
**batched_dummy_mm_inputs
)
sanity_check_mm_encoder_outputs(
dummy_encoder_outputs,
expected_num_items=max_mm_items_per_batch,
)
for i, output in enumerate(dummy_encoder_outputs):
self.encoder_cache[f"tmp_{i}"] = output
# Add `is_profile` here to pre-allocate communication buffers
hidden_states, last_hidden_states = self._dummy_run(
self.max_num_tokens, is_profile=True
)
if get_pp_group().is_last_rank:
if self.is_pooling_model:
output = self._dummy_pooler_run(hidden_states)
else:
output = self._dummy_sampler_run(last_hidden_states)
else:
output = None
self._sync_device()
del hidden_states, output
self.encoder_cache.clear()
gc.collect()
def _init_minimal_kv_cache_for_profiling(self) -> None:
from vllm.v1.core.kv_cache_utils import (
get_kv_cache_config_from_groups,
get_kv_cache_groups,
)
kv_cache_spec = self.get_kv_cache_spec()
KVCacheSpecRegistry.check_kv_cache_spec_registry(kv_cache_spec)
kv_cache_groups = get_kv_cache_groups(self.vllm_config, kv_cache_spec)
# the minimum number of blocks required is 1 block *per sequence*
min_blocks = (
min(self.max_num_reqs, self.compilation_config.max_cudagraph_capture_size)
or 1
)
# Temporarily change num_gpu_blocks_override to allocate a minimal KV cache
saved_override = self.cache_config.num_gpu_blocks_override
self.cache_config.num_gpu_blocks_override = min_blocks
minimal_config = get_kv_cache_config_from_groups(
self.vllm_config, kv_cache_groups, available_memory=0
)
self.cache_config.num_gpu_blocks_override = saved_override
self.initialize_kv_cache(minimal_config, is_profiling=True)
self.cache_config.num_gpu_blocks = minimal_config.num_blocks
logger.debug("Initialized minimal KV cache for CUDA graph profiling")
@staticmethod
@contextmanager
def _freeze_gc():
gc.collect()
should_freeze = not envs.VLLM_ENABLE_CUDAGRAPH_GC
if should_freeze:
gc.freeze()
try:
yield
finally:
if should_freeze:
gc.unfreeze()
gc.collect()
def shutdown(self) -> None:
"""Release GPU tensors (model weights, KV caches, workspace) so that
memory is reclaimable when running in the same process."""
from vllm.model_executor.layers.rotary_embedding import _ROPE_DICT
from vllm.v1.worker.workspace import reset_workspace_manager
# Calls torch.accelerator.synchronize()
self._cleanup_profiling_kv_cache()
if current_platform.is_rocm():
# Drop captured graphs before distributed teardown. On ROCm, delayed
# graph destruction can surface HSA faults in the next engine startup.
CUDAGraphWrapper.clear_all_graphs()
BreakableCUDAGraphWrapper.clear_all_graphs()
self.encoder_cudagraph_manager = None
self.compilation_config.static_forward_context.clear()
self.model = None # type: ignore[assignment]
_ROPE_DICT.clear()
reset_workspace_manager()
if current_platform.is_rocm() or current_platform.is_xpu():
gc.collect()
torch.accelerator.empty_cache()
torch.accelerator.synchronize()
def _cleanup_profiling_kv_cache(self) -> None:
torch.accelerator.synchronize()
if hasattr(self, "kv_caches") and self.kv_caches:
for i in range(len(self.kv_caches)):
self.kv_caches[i] = None # type: ignore
self.kv_caches.clear()
if hasattr(self, "cross_layers_kv_cache"):
self.cross_layers_kv_cache = None
self.cross_layers_attn_backend = None
if hasattr(self, "attn_groups"):
self.attn_groups.clear()
if hasattr(self, "kv_cache_config"):
delattr(self, "kv_cache_config")
self.cache_config.num_gpu_blocks = None
for layer in self.compilation_config.static_forward_context.values():
if hasattr(layer, "kv_cache"):
kv_cache = layer.kv_cache
layer.kv_cache = (
torch.tensor([]) if isinstance(kv_cache, torch.Tensor) else []
)
# Clean up quantized KV cache scale views
# (int8_per_token_head, fp8_per_token_head)
if hasattr(layer, "impl"):
if hasattr(layer.impl, "_k_scale_cache"):
layer.impl._k_scale_cache = None
if hasattr(layer.impl, "_v_scale_cache"):
layer.impl._v_scale_cache = None
gc.collect()
torch.accelerator.empty_cache()
logger.debug("Cleaned up profiling KV cache and CUDA graphs")
@torch.inference_mode()
def _create_encoder_cudagraph_manager(self) -> "EncoderCudaGraphManager | None":
if not (
self.compilation_config.cudagraph_mm_encoder and self.supports_mm_inputs
):
return None
# Use get_model() to unwrap CUDAGraphWrapper/UBatchWrapper, because
# @runtime_checkable Protocol isinstance() checks do not work through
# __getattr__ forwarding.
from vllm.model_executor.models.interfaces import (
SupportsEncoderCudaGraph,
supports_encoder_cudagraph,
)
from vllm.v1.worker.encoder_cudagraph import (
EncoderCudaGraphManager,
)
raw_model = self.get_model()
if not supports_encoder_cudagraph(raw_model):
return None
return EncoderCudaGraphManager(
vllm_config=self.vllm_config,
device=self.device,
dtype=self.dtype,
model=cast(SupportsEncoderCudaGraph, raw_model),
)
@torch.inference_mode()
def _maybe_init_encoder_cudagraph_manager(self) -> None:
if self.encoder_cudagraph_manager is None:
self.encoder_cudagraph_manager = self._create_encoder_cudagraph_manager()
if self.encoder_cudagraph_manager is not None:
logger.info("Initialized EncoderCudaGraphManager for vision encoder")
@torch.inference_mode()
def profile_cudagraph_memory(self) -> int:
with set_current_vllm_config(self.vllm_config):
self._init_minimal_kv_cache_for_profiling()
saved_num_cudagraph_captured = compilation_counter.num_cudagraph_captured
capture_descs = self.cudagraph_dispatcher.get_capture_descs()
# Use a temporary manager for memory profiling. The persistent manager
# is initialized later so it does not keep profiling-only graph state.
encoder_cudagraph_manager = self._create_encoder_cudagraph_manager()
decoder_graphs = sum(len(descs) for _, descs in capture_descs)
encoder_graphs = (
encoder_cudagraph_manager.get_num_graphs_to_capture()
if encoder_cudagraph_manager is not None
else 0
)
total_graphs = decoder_graphs + encoder_graphs
if total_graphs == 0:
logger.debug("No CUDA graphs will be captured, skipping profiling")
self._cleanup_profiling_kv_cache()
return 0
graph_groups = [
*(
f"{mode.name}={len(descs)} (largest={descs[0].num_tokens})"
for mode, descs in capture_descs
if descs
),
]
if encoder_graphs > 0:
graph_groups.append(
f"ENCODER={encoder_graphs} "
f"(largest={encoder_cudagraph_manager.token_budgets[-1]})"
)
logger.info("Profiling CUDA graph memory: %s", ", ".join(graph_groups))
# Use a temporary pool for profiling to avoid fragmentation in the main pool.
profiling_pool = current_platform.graph_pool_handle()
encoder_profiling_pool = current_platform.graph_pool_handle()
original_pools: dict[int, Any] = {}
all_wrappers = list(CUDAGraphWrapper._all_instances) + list(
BreakableCUDAGraphWrapper._all_instances
)
for instance in all_wrappers:
original_pools[id(instance)] = instance.graph_pool
instance.graph_pool = profiling_pool
shared_memory_estimate = {}
per_graph_estimate = {}
encoder_memory_estimate = 0
# On ROCm, capture these throwaway profiling graphs on the current stream
# instead of the fresh side stream graph_capture() allocates by default.
# torch's allocator pools free blocks per stream, so a side-stream forward
# strands a persistent aiter scratch buffer in a separate pool, shifting
# the physical placement of the real KV cache allocated afterward and
# slowing bandwidth-bound decode ~20%. The graphs are discarded, so a
# side stream is unnecessary here.
# cap_ctx=None keeps the side-stream path on CUDA, where the current
# stream is the legacy default stream, on which capture cannot begin.
cap_ctx = (
GraphCaptureContext(torch.cuda.current_stream(self.device))
if current_platform.is_rocm()
else None
)
# Cleanup-only guard: CUDA graph capture errors should still propagate
# because encoder graph capture is opt-in.
try:
set_cudagraph_capturing_enabled(True)
with (
self._freeze_gc(),
graph_capture(device=self.device, graph_capture_context=cap_ctx),
):
torch.accelerator.synchronize()
torch.accelerator.empty_cache()
for mode, descs in capture_descs:
profile_descs = descs[:2]
mem_samples: list[int] = []
for i, desc in enumerate(profile_descs):
mem_before = torch.accelerator.get_memory_info()[0]
self._warmup_and_capture(
desc,
cudagraph_runtime_mode=mode,
profile_seq_lens=(
min(
self.max_model_len,
self.max_num_tokens // desc.num_tokens,
)
if mode == CUDAGraphMode.FULL and i == 0
else None
),
)
torch.accelerator.synchronize()
free_after = torch.accelerator.get_memory_info()[0]
mem_samples.append(mem_before - free_after)
first_capture = mem_samples[0]
# Use at least 1 MiB per graph for driver overhead
per_graph = max(
mem_samples[1] if len(mem_samples) > 1 else 0, 1 << 20
)
shared_memory_estimate[mode] = first_capture
per_graph_estimate[mode] = per_graph * (len(descs) - 1)
logger.debug(
"Estimated %s CUDA graph memory: "
"%.2f MiB first-capture + (%d-1) × %.2f MiB per-graph",
mode.name,
first_capture / (1 << 20),
len(descs),
per_graph / (1 << 20),
)
if encoder_cudagraph_manager is not None:
mem_before = torch.accelerator.get_memory_info()[0]
encoder_cudagraph_manager.capture(graph_pool=encoder_profiling_pool)
torch.accelerator.synchronize()
free_after = torch.accelerator.get_memory_info()[0]
encoder_memory_estimate = max(mem_before - free_after, 0)
logger.debug(
"Estimated encoder CUDA graph memory: %.2f MiB for %d graphs",
encoder_memory_estimate / (1 << 20),
encoder_graphs,
)
finally:
set_cudagraph_capturing_enabled(False)
CUDAGraphWrapper.clear_all_graphs()
BreakableCUDAGraphWrapper.clear_all_graphs()
if encoder_cudagraph_manager is not None:
encoder_cudagraph_manager.clear()
all_wrappers = list(CUDAGraphWrapper._all_instances) + list(
BreakableCUDAGraphWrapper._all_instances
)
for instance in all_wrappers:
if id(instance) in original_pools:
instance.graph_pool = original_pools[id(instance)]
for key_set in self.cudagraph_dispatcher.cudagraph_keys.values():
key_set.clear()
self.cudagraph_dispatcher.keys_initialized = False
self.maybe_remove_all_loras(self.lora_config)
self._cleanup_profiling_kv_cache()
compilation_counter.num_cudagraph_captured = saved_num_cudagraph_captured
# FULL and PIECEWISE graphs share the global pool at runtime and are
# never replayed concurrently, so the pool overlays their memory.
# Take the max to avoid double-counting the overlap.
decoder_estimate = max(shared_memory_estimate.values(), default=0) + sum(
per_graph_estimate.values()
)
# Encoder graphs use a manager-local pool at runtime, separate from the
# decoder pool, so add their estimate instead of overlaying it.
total_estimate = decoder_estimate + encoder_memory_estimate
logger.info(
"Estimated CUDA graph memory: %.2f GiB total",
total_estimate / (1 << 30),
)
return int(total_estimate)
@instrument(span_name="Capture model")
def capture_model(self) -> int:
if self.compilation_config.cudagraph_mode == CUDAGraphMode.NONE:
logger.warning(
"Skipping CUDA graph capture. To turn on CUDA graph capture, "
"ensure `cudagraph_mode` was not manually set to `NONE`"
)
return 0
# Initialize encoder CUDA graph manager if enabled.
self._maybe_init_encoder_cudagraph_manager()
compilation_counter.num_gpu_runner_capture_triggers += 1
start_time = time.perf_counter()
# Trigger CUDA graph capture for specific shapes.
# Capture the large shapes first so that the smaller shapes
# can reuse the memory pool allocated for the large shapes.
set_cudagraph_capturing_enabled(True)
with self._freeze_gc(), graph_capture(device=self.device):
torch.accelerator.synchronize()
torch.accelerator.empty_cache()
start_free_gpu_memory = torch.accelerator.get_memory_info()[0]
for (
runtime_mode,
batch_descs,
) in self.cudagraph_dispatcher.get_capture_descs():
self._capture_cudagraphs(
batch_descriptors=batch_descs,
cudagraph_runtime_mode=runtime_mode,
)
torch.accelerator.synchronize()
# Capture encoder CUDA graphs if enabled
if self.encoder_cudagraph_manager is not None:
encoder_graph_pool = current_platform.graph_pool_handle()
self.encoder_cudagraph_manager.capture(graph_pool=encoder_graph_pool)
torch.accelerator.synchronize()
end_free_gpu_memory = torch.accelerator.get_memory_info()[0]
# Disable cudagraph capturing globally, so any unexpected cudagraph
# capturing will be detected and raise an error after here.
# Note: We don't put it into graph_capture context manager because
# we may do lazy capturing in future that still allows capturing
# after here.
set_cudagraph_capturing_enabled(False)
torch.accelerator.synchronize()
torch.accelerator.empty_cache()
# Lock workspace to prevent resizing during execution.
# Max workspace sizes should have been captured during warmup/profiling.
lock_workspace()
end_time = time.perf_counter()
elapsed_time = end_time - start_time
cuda_graph_size = start_free_gpu_memory - end_free_gpu_memory
# This usually takes 5~20 seconds.
logger.info_once(
"Graph capturing finished in %.0f secs, took %.2f GiB",
elapsed_time,
cuda_graph_size / (1 << 30),
)
return cuda_graph_size
def _warmup_and_capture(
self,
desc: BatchDescriptor,
cudagraph_runtime_mode: CUDAGraphMode,
profile_seq_lens: int | None = None,
allow_microbatching: bool = False,
num_warmups: int | None = None,
):
if num_warmups is None:
num_warmups = self.compilation_config.cudagraph_num_of_warmups
force_attention = cudagraph_runtime_mode == CUDAGraphMode.FULL
for _ in range(num_warmups):
self._dummy_run(
desc.num_tokens,
cudagraph_runtime_mode=CUDAGraphMode.NONE,
force_attention=force_attention,
uniform_decode=desc.uniform,
allow_microbatching=allow_microbatching,
skip_eplb=True,
remove_lora=False,
num_active_loras=desc.num_active_loras,
profile_seq_lens=profile_seq_lens,
)
self._dummy_run(
desc.num_tokens,
cudagraph_runtime_mode=cudagraph_runtime_mode,
uniform_decode=desc.uniform,
allow_microbatching=allow_microbatching,
skip_eplb=True,
remove_lora=False,
num_active_loras=desc.num_active_loras,
is_graph_capturing=True,
profile_seq_lens=profile_seq_lens,
)
def _capture_cudagraphs(
self,
batch_descriptors: list[BatchDescriptor],
cudagraph_runtime_mode: CUDAGraphMode,
):
assert (
cudagraph_runtime_mode != CUDAGraphMode.NONE
and cudagraph_runtime_mode.is_valid_runtime_mode()
), f"Invalid cudagraph runtime mode: {cudagraph_runtime_mode}"
if not batch_descriptors:
return
uniform_decode = batch_descriptors[0].uniform
# Only rank 0 should print progress bar during capture
if is_global_first_rank():
batch_descriptors = tqdm(
batch_descriptors,
disable=not self.load_config.use_tqdm_on_load,
desc="Capturing CUDA graphs ({}, {})".format(
"decode" if uniform_decode else "mixed prefill-decode",
cudagraph_runtime_mode.name,
),
)
# We skip EPLB here since we don't want to record dummy metrics
for batch_desc in batch_descriptors:
# We currently only capture ubatched graphs when its a FULL
# cudagraph, a uniform decode batch, and the number of tokens
# is above the threshold. Otherwise we just capture a non-ubatched
# version of the graph
allow_microbatching = (
self.parallel_config.use_ubatching
and cudagraph_runtime_mode == CUDAGraphMode.FULL
and uniform_decode
and check_ubatch_thresholds(
config=self.vllm_config.parallel_config,
num_tokens=batch_desc.num_tokens,
uniform_decode=uniform_decode,
)
)
self._warmup_and_capture(
batch_desc,
cudagraph_runtime_mode=cudagraph_runtime_mode,
allow_microbatching=allow_microbatching,
)
torch.accelerator.synchronize()
self.maybe_remove_all_loras(self.lora_config)
def initialize_attn_backend(
self,
kv_cache_config: KVCacheConfig,
is_profiling: bool = False,
) -> None:
"""
Initialize the attention backends and attention metadata builders.
"""
assert len(self.attn_groups) == 0, "Attention backends are already initialized"
class AttentionGroupKey(NamedTuple):
"""Deduplication key for attention groups within a KV cache group.
Splits on per-rank ``num_heads_q`` in addition to backend + spec
so layers with different Q-head counts (e.g. a spec-decode draft
with fewer attention heads than its target) get separate metadata
builders. The builders' scratch (e.g. ``softmax_segm_*`` in
``triton_attn``, ``num_qo_heads`` in FlashInfer) is sized by
``num_heads_q`` and assumes uniformity within the group; see
``get_num_attention_heads_from_layers`` in
``vllm/v1/attention/backends/utils.py``.
"""
attn_backend: type[AttentionBackend]
kv_cache_spec: KVCacheSpec
num_heads_q: int
def get_attn_backends_for_group(
kv_cache_group_spec: KVCacheGroupSpec,
) -> tuple[dict[AttentionGroupKey, list[str]], set[type[AttentionBackend]]]:
layer_type = cast(type[Any], AttentionLayerBase)
layers = get_layers_from_vllm_config(
self.vllm_config, layer_type, kv_cache_group_spec.layer_names
)
attn_backends = {}
attn_backend_layers = defaultdict(list)
# Dedupe based on full class name; this is a bit safer than
# using the class itself as the key because when we create dynamic
# attention backend subclasses (e.g. ChunkedLocalAttention) unless
# they are cached correctly, there will be different objects per
# layer.
for layer_name in kv_cache_group_spec.layer_names:
attn_backend = layers[layer_name].get_attn_backend()
if layer_name in self.kv_sharing_fast_prefill_eligible_layers:
attn_backend = create_fast_prefill_custom_backend(
"FastPrefill",
attn_backend, # type: ignore[arg-type]
)
full_cls_name = attn_backend.full_cls_name()
layer_kv_cache_spec = kv_cache_group_spec.kv_cache_spec
if isinstance(layer_kv_cache_spec, UniformTypeKVCacheSpecs):
layer_kv_cache_spec = layer_kv_cache_spec.kv_cache_specs[layer_name]
# Non-Attention layer types (e.g. Mamba1, ShortConv) do not
# expose ``num_heads``; fall back to 0 so they cluster as
# before. Such layers never coexist with Attention in a
# single KV cache group (different KVCacheSpec), so the
# fallback can never spuriously merge them with attention
# layers.
num_heads_q = getattr(layers[layer_name], "num_heads", 0)
key = (full_cls_name, layer_kv_cache_spec, num_heads_q)
attn_backends[key] = AttentionGroupKey(
attn_backend, layer_kv_cache_spec, num_heads_q
)
attn_backend_layers[key].append(layer_name)
return (
{attn_backends[k]: v for k, v in attn_backend_layers.items()},
set(group_key.attn_backend for group_key in attn_backends.values()),
)
def create_attn_groups(
attn_backends_map: dict[AttentionGroupKey, list[str]],
kv_cache_group_id: int,
) -> list[AttentionGroup]:
attn_groups: list[AttentionGroup] = []
for key, layer_names in attn_backends_map.items():
attn_group = AttentionGroup(
key.attn_backend,
layer_names,
key.kv_cache_spec,
kv_cache_group_id,
)
attn_groups.append(attn_group)
return attn_groups
attention_backend_maps = []
attention_backend_list = []
for kv_cache_group_spec in kv_cache_config.kv_cache_groups:
attn_backends = get_attn_backends_for_group(kv_cache_group_spec)
attention_backend_maps.append(attn_backends[0])
attention_backend_list.append(attn_backends[1])
# Resolve cudagraph_mode before actually initialize metadata_builders
self._check_and_update_cudagraph_mode(
attention_backend_list,
kv_cache_config.kv_cache_groups,
is_profiling=is_profiling,
)
# Check if attention backend supports PCP&DCP and related features.
check_attention_cp_compatibility(self.vllm_config)
for i, attn_backend_map in enumerate(attention_backend_maps):
self.attn_groups.append(create_attn_groups(attn_backend_map, i))
def initialize_metadata_builders(
self, kv_cache_config: KVCacheConfig, kernel_block_sizes: list[int]
) -> None:
"""
Create the metadata builders for all KV cache groups and attn groups.
"""
for kv_cache_group_id in range(len(kv_cache_config.kv_cache_groups)):
for attn_group in self.attn_groups[kv_cache_group_id]:
attn_group.create_metadata_builders(
self.vllm_config,
self.device,
kernel_block_sizes[kv_cache_group_id]
if kv_cache_group_id < len(kernel_block_sizes)
else None,
num_metadata_builders=1
if not self.parallel_config.use_ubatching
else self.parallel_config.num_ubatches,
)
# Calculate reorder batch threshold (if needed)
# Note (tdoublep): do this *after* constructing builders,
# because some of them change the threshold at init time.
self.calculate_reorder_batch_threshold()
# Initialize drafter attention backend
if self.speculative_config and (
self.speculative_config.use_eagle()
or self.speculative_config.uses_draft_model()
):
assert isinstance(
self.drafter,
EagleProposer | DFlashProposer | DraftModelProposer | Gemma4Proposer,
)
self.drafter.initialize_attn_backend(kv_cache_config, kernel_block_sizes)
def _check_and_update_cudagraph_mode(
self,
attention_backends: list[set[type[AttentionBackend]]],
kv_cache_groups: list[KVCacheGroupSpec],
is_profiling: bool = False,
) -> None:
"""
Resolve the cudagraph_mode when there are multiple attention
groups with potential conflicting CUDA graph support.
Then initialize the cudagraph_dispatcher based on the resolved
cudagraph_mode.
"""
min_cg_support = AttentionCGSupport.ALWAYS
min_cg_attn_backend = None
for attn_backend_set, kv_cache_group in zip(
attention_backends, kv_cache_groups
):
for attn_backend in attn_backend_set:
builder_cls = attn_backend.get_builder_cls()
cg_support = builder_cls.get_cudagraph_support(
self.vllm_config, kv_cache_group.kv_cache_spec
)
if cg_support.value < min_cg_support.value:
min_cg_support = cg_support
min_cg_attn_backend = attn_backend.__name__
cudagraph_mode = self.compilation_config.resolve_cudagraph_mode_and_sizes(
min_cg_support,
min_cg_attn_backend,
self.uniform_decode_query_len,
use_v2_model_runner=False,
tensor_parallel_size=self.parallel_config.tensor_parallel_size,
kv_cache_config=self.kv_cache_config,
max_num_reqs=self.max_num_reqs,
is_profiling=is_profiling,
)
# Trigger cudagraph dispatching keys initialization after
# resolved cudagraph mode.
self.cudagraph_dispatcher.initialize_cudagraph_keys(
cudagraph_mode, self.uniform_decode_query_len
)
# Initialize drafter's cudagraph dispatcher if using spec decode.
if self.speculative_config and (
self.speculative_config.use_eagle()
or self.speculative_config.uses_extract_hidden_states()
):
assert isinstance(
self.drafter,
EagleProposer
| DFlashProposer
| ExtractHiddenStatesProposer
| Gemma4Proposer,
)
self.drafter.initialize_cudagraph_keys(cudagraph_mode)
def calculate_reorder_batch_threshold(self) -> None:
"""
Choose the minimum reorder batch threshold from all attention groups.
Backends should be able to support lower threshold then what they request
just may have a performance penalty due to that backend treating decodes
as prefills.
"""
min_none_high = lambda a, b: a if b is None else b if a is None else min(a, b)
reorder_batch_thresholds: list[int | None] = [
group.get_metadata_builder().reorder_batch_threshold
for group in self._attn_group_iterator()
]
# If there are no attention groups (attention-free model) or no backend
# reports a threshold, leave reordering disabled.
if len(reorder_batch_thresholds) == 0:
self.reorder_batch_threshold = None
return
self.reorder_batch_threshold = reduce(min_none_high, reorder_batch_thresholds) # type: ignore[assignment]
def may_reinitialize_input_batch(
self, kv_cache_config: KVCacheConfig, kernel_block_sizes: list[int]
) -> None:
"""
Re-initialize the input batch if the block sizes are different from
what it was originally created with. This happens when the final
block size (determined after model loading) differs from the
placeholder used during __init__, or when there are multiple
KV cache groups.
Args:
kv_cache_config: The KV cache configuration.
kernel_block_sizes: The kernel block sizes for each KV cache group.
"""
block_sizes = []
max_num_blocks = []
slot_mapping_modes = []
max_model_len = max(self.max_model_len, self.max_encoder_len)
for kv_cache_group in kv_cache_config.kv_cache_groups:
kv_cache_spec = kv_cache_group.kv_cache_spec
kv_cache_spec_kind = get_kv_cache_spec_kind(kv_cache_spec)
if kv_cache_spec_kind == KVCacheSpecKind.ENCODER_ONLY_ATTENTION:
continue
block_size = kv_cache_spec.block_size
block_sizes.append(block_size)
if kv_cache_spec_kind == KVCacheSpecKind.MAMBA:
slot_mapping_modes.append(SlotMappingMode.NONE)
else:
slot_mapping_modes.append(SlotMappingMode.TOKEN_TO_KV_SLOT)
max_num_blocks_per_req = kv_cache_spec.max_num_blocks_per_req(
self.vllm_config, max_model_len
)
max_num_blocks.append(max_num_blocks_per_req)
if (
block_sizes != self._init_block_sizes
or kernel_block_sizes != self._init_kernel_block_sizes
or max_num_blocks != self._init_max_num_blocks
or slot_mapping_modes != self._init_slot_mapping_modes
):
self._init_block_sizes = block_sizes
self._init_kernel_block_sizes = kernel_block_sizes
self._init_max_num_blocks = max_num_blocks
self._init_slot_mapping_modes = slot_mapping_modes
self.input_batch = InputBatch(
max_num_reqs=self.max_num_reqs,
max_model_len=max_model_len,
max_num_batched_tokens=self.max_num_tokens,
device=self.device,
vocab_size=self.model_config.get_vocab_size(),
block_sizes=block_sizes,
kernel_block_sizes=kernel_block_sizes,
max_num_blocks_per_req=max_num_blocks,
num_spec_tokens=self.num_spec_tokens,
logitsprocs=self.input_batch.logitsprocs,
logitsprocs_need_output_token_ids=self.input_batch.logitsprocs_need_output_token_ids,
is_pooling_model=self.is_pooling_model,
cp_kv_cache_interleave_size=self.parallel_config.cp_kv_cache_interleave_size,
reasoning_config=self.vllm_config.reasoning_config,
slot_mapping_modes=slot_mapping_modes,
)
assert self._init_block_sizes == block_sizes, (
f"InputBatch block_sizes {self._init_block_sizes} != "
f"kv_cache block_sizes {block_sizes}"
)
assert self._init_kernel_block_sizes == kernel_block_sizes, (
f"InputBatch kernel_block_sizes {self._init_kernel_block_sizes} "
f"!= kv_cache kernel_block_sizes {kernel_block_sizes}"
)
def _allocate_kv_cache_tensors(
self, kv_cache_config: KVCacheConfig
) -> dict[str, torch.Tensor]:
"""
Initializes the KV cache buffer with the correct size. The buffer needs
to be reshaped to the desired shape before being used by the models.
Args:
kv_cache_config: The KV cache config
Returns:
dict[str, torch.Tensor]: A map between layer names to their
corresponding memory buffer for KV cache.
"""
kv_cache_raw_tensors: dict[str, torch.Tensor] = {}
packed_backing: torch.Tensor | None = None
for kv_cache_tensor in kv_cache_config.kv_cache_tensors:
if kv_cache_tensor.block_stride > 0:
# Allocate once; all packed tensors alias the same backing.
if packed_backing is None:
packed_backing = torch.zeros(
kv_cache_tensor.size,
dtype=torch.int8,
device=self.device,
)
tensor = packed_backing
else:
tensor = torch.zeros(
kv_cache_tensor.size, dtype=torch.int8, device=self.device
)
for layer_name in kv_cache_tensor.shared_by:
kv_cache_raw_tensors[layer_name] = tensor
layer_names = set()
for group in kv_cache_config.kv_cache_groups:
for layer_name in group.layer_names:
if layer_name in self.runner_only_attn_layers:
continue
layer_names.add(layer_name)
assert layer_names == set(kv_cache_raw_tensors.keys()), (
"Some layers are not correctly initialized"
)
return kv_cache_raw_tensors
def _attn_group_iterator(self) -> Iterator[AttentionGroup]:
return itertools.chain.from_iterable(self.attn_groups)
def _kv_cache_spec_attn_group_iterator(self) -> Iterator[AttentionGroup]:
if not self.kv_cache_config.kv_cache_groups:
return
for attn_groups in self.attn_groups:
yield from attn_groups
def _reshape_kv_cache_tensors(
self,
kv_cache_raw_tensors: dict[str, torch.Tensor],
kernel_block_sizes: list[int],
) -> dict[str, torch.Tensor]:
"""
Reshape the KV cache tensors to the desired shape and dtype.
Args:
kv_cache_raw_tensors: The KV cache buffer of each layer, with
correct size but uninitialized shape.
kernel_block_sizes: The kernel block sizes for each KV cache group.
Returns:
Dict[str, torch.Tensor]: A map between layer names to their
corresponding memory buffer for KV cache.
"""
kv_caches: dict[str, torch.Tensor] = {}
has_attn, has_mamba = False, False
# Map layer names to (offset, block_stride) within the packed
# backing tensor so we can create strided views per layer.
layer_packing: dict[str, tuple[int, int]] = {}
for kv_tensor in self.kv_cache_config.kv_cache_tensors:
if kv_tensor.block_stride > 0:
for ln in kv_tensor.shared_by:
layer_packing[ln] = (kv_tensor.offset, kv_tensor.block_stride)
for group in self._kv_cache_spec_attn_group_iterator():
kv_cache_spec = group.kv_cache_spec
attn_backend = group.backend
if group.kv_cache_group_id == len(kernel_block_sizes):
# There may be a last group for layers without kv cache.
continue
kernel_block_size = kernel_block_sizes[group.kv_cache_group_id]
for layer_name in group.layer_names:
if layer_name in self.runner_only_attn_layers:
continue
raw_tensor = kv_cache_raw_tensors[layer_name]
packing = layer_packing.get(layer_name)
if packing is not None:
_, blk_stride = packing
num_blocks = raw_tensor.numel() // blk_stride
else:
assert raw_tensor.numel() % kv_cache_spec.page_size_bytes == 0
num_blocks = raw_tensor.numel() // kv_cache_spec.page_size_bytes
if isinstance(kv_cache_spec, AttentionSpec):
has_attn = True
num_blocks_per_kv_block = (
kv_cache_spec.block_size // kernel_block_size
)
kernel_num_blocks = num_blocks * num_blocks_per_kv_block
# For MLA with compression, storage_block_size != block_size
if kv_cache_spec.storage_block_size != kv_cache_spec.block_size:
shape_block_size = kv_cache_spec.storage_block_size
else:
shape_block_size = kernel_block_size
# Skipped layers (--kv-cache-dtype-skip-layers) need
# the unquantized shape.
layer_cache_dtype_str = (
"auto"
if kv_cache_spec.kv_quant_mode == KVQuantMode.NONE
else getattr(
kv_cache_spec,
"cache_dtype_str",
None,
)
or self.cache_config.cache_dtype
)
kv_cache_shape = attn_backend.get_kv_cache_shape(
kernel_num_blocks,
shape_block_size,
kv_cache_spec.num_kv_heads,
kv_cache_spec.head_size,
cache_dtype_str=layer_cache_dtype_str,
)
try:
kv_cache_stride_order = attn_backend.get_kv_cache_stride_order()
assert len(kv_cache_stride_order) == len(kv_cache_shape)
except (AttributeError, NotImplementedError):
kv_cache_stride_order = tuple(range(len(kv_cache_shape)))
raw_tensor = kv_cache_raw_tensors[layer_name]
kv_caches[layer_name] = _reshape_attention_kv_cache(
raw_tensor,
kv_cache_spec,
kv_cache_shape,
kv_cache_stride_order,
kernel_num_blocks,
packing,
)
elif isinstance(kv_cache_spec, MambaSpec):
has_mamba = True
raw_tensor = kv_cache_raw_tensors[layer_name]
state_tensors = []
storage_offset_bytes = 0
for shape, dtype in zip(kv_cache_spec.shapes, kv_cache_spec.dtypes):
dtype_size = get_dtype_size(dtype)
num_element_per_page = (
kv_cache_spec.page_size_bytes // dtype_size
)
target_shape = (num_blocks, *shape)
stride = torch.empty(target_shape).stride()
target_stride = (num_element_per_page, *stride[1:])
assert storage_offset_bytes % dtype_size == 0
tensor = torch.as_strided(
raw_tensor.view(dtype),
size=target_shape,
stride=target_stride,
storage_offset=storage_offset_bytes // dtype_size,
)
state_tensors.append(tensor)
storage_offset_bytes += stride[0] * dtype_size
kv_caches[layer_name] = state_tensors
else:
raise NotImplementedError
# Reconcile divergent KV layouts to blocks-first. Triggered by hybrid
# attention/mamba models, and by encoder-decoder models whose shared
# decoder/cross-attention allocation mixes K/V-first and blocks-first
# backends (see _has_mixed_attention_kv_layout).
if has_attn and (
has_mamba or self._has_mixed_attention_kv_layout(kernel_block_sizes)
):
self._update_hybrid_attention_mamba_layout(kv_caches, kernel_block_sizes)
return kv_caches
def _has_mixed_attention_kv_layout(self, kernel_block_sizes: list[int]) -> bool:
"""Whether attention groups disagree on the physical KV cache layout.
Encoder-decoder models (e.g. Whisper) share one raw KV allocation
between a decoder self-attention layer (K/V-first ROCM_ATTN, block dim
1) and a cross-attention layer (blocks-first, block dim 0). Mixed block
dims mean a block ID maps to different bytes per layer, so the shared
buffer must be normalized to a single (blocks-first) layout.
"""
block_dims: set[int] = set()
for group in self._kv_cache_spec_attn_group_iterator():
kv_cache_spec = group.kv_cache_spec
if not isinstance(kv_cache_spec, AttentionSpec):
continue
if group.kv_cache_group_id == len(kernel_block_sizes):
continue
block_dims.add(
group.backend.get_kv_cache_block_dim(
kernel_block_sizes[group.kv_cache_group_id],
kv_cache_spec.num_kv_heads,
kv_cache_spec.head_size,
cache_dtype_str=self.cache_config.cache_dtype,
)
)
return len(block_dims) > 1
def _update_hybrid_attention_mamba_layout(
self, kv_caches: dict[str, torch.Tensor], kernel_block_sizes: list[int]
) -> None:
"""
Update the layout of attention layers from (2, num_blocks, ...) to
(num_blocks, 2, ...).
Args:
kv_caches: The KV cache buffer of each layer.
kernel_block_sizes: The kernel block sizes for each KV cache group.
"""
for group in self._kv_cache_spec_attn_group_iterator():
kv_cache_spec = group.kv_cache_spec
if not isinstance(kv_cache_spec, AttentionSpec):
continue
block_dim = group.backend.get_kv_cache_block_dim(
kernel_block_sizes[group.kv_cache_group_id],
kv_cache_spec.num_kv_heads,
kv_cache_spec.head_size,
cache_dtype_str=self.cache_config.cache_dtype,
)
# block_dim: 0 means (num_blocks, 2, ...); 1 means (2, num_blocks, ...).
if block_dim == 0:
continue
assert block_dim == 1
for layer_name in group.layer_names:
kv_cache = kv_caches[layer_name]
hidden_size = kv_cache.shape[2:].numel()
kv_cache.as_strided_(
size=kv_cache.shape,
stride=(hidden_size, 2 * hidden_size, *kv_cache.stride()[2:]),
)
def initialize_kv_cache_tensors(
self, kv_cache_config: KVCacheConfig, kernel_block_sizes: list[int]
) -> dict[str, torch.Tensor]:
"""
Initialize the memory buffer for KV cache.
Args:
kv_cache_config: The KV cache config
kernel_block_sizes: The kernel block sizes for each KV cache group.
Returns:
Dict[str, torch.Tensor]: A map between layer names to their
corresponding memory buffer for KV cache.
"""
# Try creating KV caches optimized for kv-connector transfers
cache_dtype = self.cache_config.cache_dtype
if self.use_uniform_kv_cache(self.attn_groups):
kv_caches, cross_layers_kv_cache, attn_backend = (
self.allocate_uniform_kv_caches(
kv_cache_config,
self.attn_groups,
cache_dtype,
self.device,
kernel_block_sizes,
)
)
self.cross_layers_kv_cache = cross_layers_kv_cache
self.cross_layers_attn_backend = attn_backend
else:
# Fallback to the general case
# Initialize the memory buffer for KV cache
kv_cache_raw_tensors = self._allocate_kv_cache_tensors(kv_cache_config)
# Change the memory buffer to the desired shape
kv_caches = self._reshape_kv_cache_tensors(
kv_cache_raw_tensors, kernel_block_sizes
)
# Set up cross-layer KV cache sharing
for layer_name, target_layer_name in self.shared_kv_cache_layers.items():
logger.debug("%s reuses KV cache of %s", layer_name, target_layer_name)
kv_caches[layer_name] = kv_caches[target_layer_name]
num_attn_module = (
2 if self.model_config.hf_config.model_type == "longcat_flash" else 1
)
bind_kv_cache(
kv_caches,
self.compilation_config.static_forward_context,
self.kv_caches,
num_attn_module,
)
return kv_caches
def maybe_add_kv_sharing_layers_to_kv_cache_groups(
self, kv_cache_config: KVCacheConfig
) -> None:
"""
Add layers that re-use KV cache to KV cache group of its target layer.
Mapping of KV cache tensors happens in `initialize_kv_cache_tensors()`
"""
if not self.shared_kv_cache_layers:
# No cross-layer KV sharing, return
return
add_kv_sharing_layers_to_kv_cache_groups(
self.shared_kv_cache_layers,
kv_cache_config.kv_cache_groups,
self.runner_only_attn_layers,
)
if self.cache_config.kv_sharing_fast_prefill:
# In You Only Cache Once (https://arxiv.org/abs/2405.05254) or other
# similar KV sharing setups, only the layers that generate KV caches
# are involved in the prefill phase, enabling prefill to early exit.
attn_layers = get_layers_from_vllm_config(self.vllm_config, Attention)
for layer_name in reversed(attn_layers):
if layer_name in self.shared_kv_cache_layers:
self.kv_sharing_fast_prefill_eligible_layers.add(layer_name)
else:
break
def initialize_kv_cache(
self,
kv_cache_config: KVCacheConfig,
is_profiling: bool = False,
) -> None:
"""
Initialize KV cache based on `kv_cache_config`.
Args:
kv_cache_config: Configuration for the KV cache, including the KV
cache size of each layer
"""
kv_cache_config = deepcopy(kv_cache_config)
self.kv_cache_config = kv_cache_config
self._mamba_bufs = None
self.may_add_encoder_only_layers_to_kv_cache_config()
self.maybe_add_kv_sharing_layers_to_kv_cache_groups(kv_cache_config)
self.initialize_attn_backend(kv_cache_config, is_profiling=is_profiling)
initialize_mamba_ssu_backend(
self.vllm_config.mamba_config, self.kv_cache_config
)
# The kernel block size for all KV cache groups. For example, if
# kv_cache_manager uses block_size 256 for a given group, but the attention
# backends for that group only supports block_size 64, we will return
# kernel_block_size 64 and split the 256-token-block to 4 blocks with 64
# tokens each.
kernel_block_sizes = prepare_kernel_block_sizes(
kv_cache_config, self.attn_groups
)
self._kernel_block_sizes = kernel_block_sizes
# create metadata builders
self.initialize_metadata_builders(kv_cache_config, kernel_block_sizes)
# Reinitialize need to after initialize_attn_backend
self.may_reinitialize_input_batch(kv_cache_config, kernel_block_sizes)
kv_caches = self.initialize_kv_cache_tensors(
kv_cache_config, kernel_block_sizes
)
if (
self.speculative_config
and self.speculative_config.uses_extract_hidden_states()
):
assert isinstance(self.drafter, ExtractHiddenStatesProposer)
# validate all draft model layers belong to the same kv cache
# group
self.drafter.validate_same_kv_cache_group(kv_cache_config)
if has_kv_transfer_group() and not is_profiling:
kv_transfer_group = get_kv_transfer_group()
if self.cross_layers_kv_cache is not None:
assert self.cross_layers_attn_backend is not None
kv_transfer_group.register_cross_layers_kv_cache(
self.cross_layers_kv_cache, self.cross_layers_attn_backend
)
else:
kv_transfer_group.register_kv_caches(kv_caches)
kv_transfer_group.set_host_xfer_buffer_ops(copy_kv_blocks)
def _get_attention_kv_cache_gid(self) -> int:
"""Find the KV cache group index for attention layers.
Must match :attr:`RoutedExpertsManager.attn_gid` in the scheduler:
both pick the first ``FullAttentionSpec`` group so hybrid models
(Mamba / linear-attention layers that use other AttentionSpec
subclasses) end up indexing the same slot layout on both sides.
Falls back to 0 only for legacy single-group configs.
"""
for gid, group in enumerate(self.kv_cache_config.kv_cache_groups):
if isinstance(group.kv_cache_spec, FullAttentionSpec):
return gid
return 0
def init_routed_experts_capturer(self):
logger.info(
"Initializing routed experts capturer, enable_return_routed_experts: %s",
self.model_config.enable_return_routed_experts,
)
self.routed_experts_capturer = RoutedExpertsCapturer(
max_num_batched_tokens=self.scheduler_config.max_num_batched_tokens,
vllm_config=self.vllm_config,
)
self.routed_experts_attn_gid = self._get_attention_kv_cache_gid()
self._bind_routed_experts_capturer(self.routed_experts_capturer)
# Pinned CPU buffer for non-blocking D2H of ``routing_data`` on
# the sync scheduling path. Shape / dtype mirror the device
# capturer exactly so ``copy_`` is a straight memcpy.
self.routed_experts_cpu = torch.empty(
self.routed_experts_capturer.device_buffer.shape,
dtype=self.routed_experts_capturer.device_buffer.dtype,
device="cpu",
pin_memory=PIN_MEMORY,
)
# ``slot_mapping`` dtype is fixed to int64 by
# ``block_table.slot_mapping``; we mirror that here.
max_tokens = self.scheduler_config.max_num_batched_tokens
self.routed_experts_slot_mapping_cpu = torch.empty(
(max_tokens,),
dtype=torch.int64,
device="cpu",
pin_memory=PIN_MEMORY,
)
# Private device buffer so the shared ``block_table.slot_mapping``
# can be overwritten by the next ``_prepare_inputs`` while the
# D2H is still pending on the copy stream. Written in
# ``_prepare_inputs``, read in ``_bookkeeping_sync`` (sync path)
# or cloned into a snapshot (async path).
self.routed_experts_slot_mapping_device = torch.empty(
(max_tokens,),
dtype=torch.int64,
device=self.device,
)
self.routed_experts_initialized = True
def _bind_routed_experts_capturer(self, capturer: RoutedExpertsCapturer) -> None:
from vllm.model_executor.layers.fused_moe.layer import MoERunner
from vllm.model_executor.layers.fused_moe.router.base_router import (
BaseRouter,
)
for module in self.model.modules():
if isinstance(module, MoERunner) and isinstance(module.router, BaseRouter):
layer_id = module.layer_id
def _capture_fn(topk_ids, _layer_id=layer_id, _capturer=capturer):
_capturer.capture(_layer_id, topk_ids)
module.router.set_capture_fn(_capture_fn)
def may_add_encoder_only_layers_to_kv_cache_config(self) -> None:
"""
Add encoder-only layers to the KV cache config.
"""
block_size = self.vllm_config.cache_config.block_size
encoder_only_attn_specs: dict[AttentionSpec, list[str]] = defaultdict(list)
attn_layers = get_layers_from_vllm_config(self.vllm_config, Attention)
for layer_name, attn_module in attn_layers.items():
if attn_module.attn_type == AttentionType.ENCODER_ONLY:
attn_spec: AttentionSpec = EncoderOnlyAttentionSpec(
block_size=block_size,
num_kv_heads=attn_module.num_kv_heads,
head_size=attn_module.head_size,
dtype=self.kv_cache_dtype,
)
encoder_only_attn_specs[attn_spec].append(layer_name)
self.runner_only_attn_layers.add(layer_name)
if len(encoder_only_attn_specs) > 0:
assert len(encoder_only_attn_specs) == 1, (
"Only support one encoder-only attention spec now"
)
spec, layer_names = encoder_only_attn_specs.popitem()
self.kv_cache_config.kv_cache_groups.append(
KVCacheGroupSpec(layer_names=layer_names, kv_cache_spec=spec)
)
def get_kv_cache_spec(self) -> dict[str, KVCacheSpec]:
"""
Generates the KVCacheSpec by parsing the kv cache format from each
Attention module in the static forward context.
Returns:
KVCacheSpec: A dictionary mapping layer names to their KV cache
format. Layers that do not need KV cache are not included.
"""
if has_ec_transfer() and not get_ec_transfer().is_consumer:
return {}
kv_cache_spec: dict[str, KVCacheSpec] = {}
layer_type = cast(type[Any], AttentionLayerBase)
attn_layers = get_layers_from_vllm_config(self.vllm_config, layer_type)
for layer_name, attn_module in attn_layers.items():
if isinstance(attn_module, Attention) and (
kv_tgt_layer := attn_module.kv_sharing_target_layer_name
):
# The layer doesn't need its own KV cache and will use that of
# the target layer. We skip creating a KVCacheSpec for it, so
# that KV cache management logic will act as this layer does
# not exist, and doesn't allocate KV cache for the layer. This
# enables the memory saving of cross-layer kv sharing, allowing
# a given amount of memory to accommodate longer context lengths
# or enable more requests to be processed simultaneously.
self.shared_kv_cache_layers[layer_name] = kv_tgt_layer
continue
# Skip modules that don't need KV cache (eg encoder-only attention)
if spec := attn_module.get_kv_cache_spec(self.vllm_config):
if isinstance(spec, AttentionSpec):
backend = attn_module.get_attn_backend()
# indexes_kv_by_block_stride() -> get_kv_cache_stride_order()
# -> get_kv_cache_layout() needs the current vLLM config.
with set_current_vllm_config(self.vllm_config):
indexes = backend.indexes_kv_by_block_stride()
spec = replace(spec, indexes_kv_by_block_stride=indexes)
kv_cache_spec[layer_name] = spec
return kv_cache_spec
def _to_list(self, sampled_token_ids: torch.Tensor) -> list[list[int]]:
# This is a short term mitigation for issue mentioned in
# https://github.com/vllm-project/vllm/issues/22754.
# `tolist` would trigger a cuda wise stream sync, which
# would block other copy ops from other cuda streams.
# A cuda event sync would avoid such a situation. Since
# this is in the critical path of every single model
# forward loop, this has caused perf issue for a disagg
# setup.
pinned = self.sampled_token_ids_pinned_cpu[: sampled_token_ids.shape[0]]
pinned.copy_(sampled_token_ids, non_blocking=True)
self.transfer_event.record()
self.transfer_event.synchronize()
return pinned.tolist()
def get_encoder_timing_stats(self) -> dict[str, dict[str, float | int]]:
"""
Get encoder timing stats for all requests and clear the registry.
Returns:
Dictionary mapping request_id to stats dict.
"""
with self._encoder_timing_lock:
stats = {
req_id: stats_obj.to_dict()
for req_id, stats_obj in self.encoder_timing_registry.items()
}
self.encoder_timing_registry.clear()
return stats
@contextmanager
def timed_encoder_operation(
self,
should_time: bool,
group_lora_refs: list[tuple[str, Any]],
current_item_idx: int,
num_items: int,
):
"""
Context manager to time encoder forward operations.
Args:
should_time: Whether timing is enabled
group_lora_refs: Full list of (request_id, pos_info) tuples
current_item_idx: Starting index for this group
num_items: Number of items in this group
"""
if not should_time:
yield
return
group_refs = group_lora_refs[current_item_idx : current_item_idx + num_items]
group_request_ids = {req_id for req_id, _ in group_refs}
torch.accelerator.synchronize()
start_time = time.perf_counter()
try:
yield
finally:
torch.accelerator.synchronize()
elapsed = time.perf_counter() - start_time
per_request_time = elapsed / max(len(group_request_ids), 1)
with self._encoder_timing_lock:
for req_id in group_request_ids:
if req_id not in self.encoder_timing_registry:
self.encoder_timing_registry[req_id] = EncoderTimingStats()
stats = self.encoder_timing_registry[req_id]
stats.encoder_forward_secs += per_request_time
stats.num_encoder_calls += 1
@dataclass
class EncoderTimingStats:
"""Per-request timing statistics for encoder forward pass."""
encoder_forward_secs: float = 0.0
"""Time spent in vision encoder forward pass (seconds)."""
num_encoder_calls: int = 0
"""Number of times encoder was called for this request."""
def to_dict(self) -> dict[str, float | int]:
return {
"encoder_forward_secs": self.encoder_forward_secs,
"num_encoder_calls": self.num_encoder_calls,
}