Files
vllm-ascend/vllm_ascend/worker/model_runner_v1.py
fan2956 434059e417 [BugFix] Fix multimodal model support fullgraph error (#3425)
### What this PR does / why we need it?
Because the update_attn_params function requires passing the num_tokens
parameter, and num_tokens is obtained via postions.shape[0]. However,
the multimodal model uses mrope (Multidimensional Rotary Position
Embedding), which results in the postions having a shape of 2.
Consequently, postions.shape[0] retrieves an incorrect value.We resolve
this issue by replacing positions.shape[0] with maybe_padded_num_tokens.

- vLLM version: v0.11.0rc3
- vLLM main: https://github.com/vllm-project/vllm/commit/v0.11.0

Signed-off-by: fan2956 <zhoufan53@huawei.com>
2025-10-14 21:51:09 +08:00

3580 lines
170 KiB
Python

#
# Copyright (c) 2025 Huawei Technologies Co., Ltd. All Rights Reserved.
# Copyright 2023 The vLLM team.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# This file is a part of the vllm-ascend project.
# Adapted from vllm-project/vllm/vllm/worker/gpu_model_runner.py
#
import copy
import gc
import itertools
import re
import time
from collections import defaultdict
from collections.abc import Iterator
from contextlib import contextmanager, nullcontext
from copy import deepcopy
from dataclasses import dataclass
from multiprocessing import Manager
from typing import (TYPE_CHECKING, Any, Dict, List, NamedTuple, Optional,
Union, cast)
import numpy as np
import numpy.typing as npt
import torch
import torch._dynamo.cache_size
import torch.distributed as dist
import torch.nn as nn
from tqdm import tqdm # type: ignore
from vllm.attention import AttentionType, get_attn_backend
from vllm.attention.backends.abstract import AttentionBackend
from vllm.attention.layer import Attention
from vllm.compilation.counter import compilation_counter
from vllm.compilation.monitor import set_cudagraph_capturing_enabled
from vllm.config import (CompilationLevel, CUDAGraphMode, VllmConfig,
get_layers_from_vllm_config)
from vllm.distributed import tensor_model_parallel_all_gather
from vllm.distributed.kv_transfer import (get_kv_transfer_group,
has_kv_transfer_group)
from vllm.distributed.kv_transfer.kv_connector.v1 import KVConnectorBase_V1
from vllm.distributed.parallel_state import (get_dp_group, get_pp_group,
get_tp_group,
is_global_first_rank)
from vllm.forward_context import BatchDescriptor, get_forward_context
from vllm.logger import logger
from vllm.model_executor.layers.attention_layer_base import AttentionLayerBase
from vllm.model_executor.layers.mamba.abstract import MambaBase
from vllm.model_executor.layers.rotary_embedding import MRotaryEmbedding
from vllm.model_executor.model_loader import get_model
from vllm.model_executor.models.interfaces import supports_transcription
from vllm.model_executor.models.interfaces_base import (
VllmModelForPooling, is_pooling_model, is_text_generation_model)
from vllm.multimodal.inputs import MultiModalKwargsItem, PlaceholderRange
from vllm.multimodal.utils import group_mm_kwargs_by_modality
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.utils import (STR_DTYPE_TO_TORCH_DTYPE, DeviceMemoryProfiler,
LazyLoader, cdiv, get_dtype_size,
is_pin_memory_available)
from vllm.utils.jsontree import json_map_leaves
from vllm.v1.attention.backends.gdn_attn import GDNAttentionMetadataBuilder
from vllm.v1.attention.backends.utils import (
AttentionCGSupport, reorder_batch_to_split_decodes_and_prefills)
from vllm.v1.cudagraph_dispatcher import CudagraphDispatcher
# yapf conflicts with isort for this block
# yapf: disable
from vllm.v1.kv_cache_interface import (AttentionSpec, FullAttentionSpec,
KVCacheConfig, KVCacheGroupSpec,
KVCacheSpec, MambaSpec,
MLAAttentionSpec,
UniformTypeKVCacheSpecs)
# yapf: enable
from vllm.v1.outputs import (EMPTY_MODEL_RUNNER_OUTPUT, AsyncModelRunnerOutput,
DraftTokenIds, LogprobsTensors, ModelRunnerOutput,
PoolerOutput)
from vllm.v1.pool.metadata import PoolingMetadata
from vllm.v1.sample.metadata import SamplingMetadata
from vllm.v1.spec_decode.metadata import SpecDecodeMetadata
from vllm.v1.spec_decode.ngram_proposer import NgramProposer
from vllm.v1.utils import CpuGpuBuffer
from vllm.v1.worker.kv_connector_model_runner_mixin import KVConnectorOutput
from vllm.v1.worker.lora_model_runner_mixin import LoRAModelRunnerMixin
from vllm.v1.worker.utils import (AttentionGroup, bind_kv_cache,
gather_mm_placeholders,
sanity_check_mm_encoder_outputs,
scatter_mm_placeholders)
import vllm_ascend.envs as envs_ascend
from vllm_ascend.ascend_config import get_ascend_config
from vllm_ascend.ascend_forward_context import (MoECommType,
set_ascend_forward_context)
from vllm_ascend.attention.attention_mask import AttentionMaskBuilder
from vllm_ascend.attention.attention_v1 import AscendAttentionState
from vllm_ascend.attention.utils import AscendCommonAttentionMetadata
from vllm_ascend.compilation.acl_graph import (ACLGraphWrapper,
set_graph_params,
update_attn_params,
update_mla_attn_params)
from vllm_ascend.eplb.adaptor.vllm_adaptor import VllmEplbAdaptor
from vllm_ascend.eplb.core.eplb_device_transfer_loader import \
D2DExpertWeightLoader
from vllm_ascend.eplb.core.eplb_worker import EplbProcess
from vllm_ascend.eplb.eplb_updator import EplbUpdator
from vllm_ascend.eplb.utils import model_register
from vllm_ascend.models.layers.mla import AscendMultiHeadLatentAttention
from vllm_ascend.multistream.ms_split import compute_split_seq_index
from vllm_ascend.ops.weight_prefetch import WeightPrefetchMethod
from vllm_ascend.platform import NPUPlatform
from vllm_ascend.sample.logits_processor import build_logitsprocs
from vllm_ascend.sample.rejection_sampler import AscendRejectionSampler
from vllm_ascend.spec_decode import get_spec_decode_method
from vllm_ascend.spec_decode.eagle_proposer import EagleProposer
from vllm_ascend.spec_decode.interface import SpecDcodeType
from vllm_ascend.spec_decode.mtp_proposer import MtpProposer
from vllm_ascend.utils import (ACL_FORMAT_FRACTAL_ND, ACL_FORMAT_FRACTAL_NZ,
AscendSocVersion, ProfileExecuteDuration,
get_ascend_soc_version, is_310p, is_enable_nz,
lmhead_tp_enable)
from vllm_ascend.worker.npu_input_batch import CachedRequestState, InputBatch
if TYPE_CHECKING:
import xgrammar as xgr # type: ignore[import-untyped]
from vllm.v1.core.sched.output import SchedulerOutput
else:
xgr = LazyLoader("xgr", globals(), "xgrammar")
import torch_npu
# if true, allow tensor initialization and casting with internal format (e.g., NZ)
torch.npu.config.allow_internal_format = True
if is_310p():
torch_npu.npu.set_compile_mode(jit_compile=False)
ACL_FORMAT = ACL_FORMAT_FRACTAL_NZ
else:
ACL_FORMAT = ACL_FORMAT_FRACTAL_ND
@dataclass
class GraphCaptureContext:
stream: torch.npu.Stream
@contextmanager
def graph_capture(device: torch.device):
"""
`graph_capture` is a context manager which should surround the code that
is capturing the NPU graph. Its main purpose is to ensure that the
some operations will be run after the graph is captured, before the graph
is replayed. It returns a `GraphCaptureContext` object which contains the
necessary data for the graph capture. Currently, it only contains the
stream that the graph capture is running on. This stream is set to the
current NPU stream when the context manager is entered and reset to the
default stream when the context manager is exited. This is to ensure that
the graph capture is running on a separate stream from the default stream,
in order to explicitly distinguish the kernels to capture
from other kernels possibly launched on background in the default stream.
"""
graph_capture_context = GraphCaptureContext(
torch.npu.Stream(device=device))
stream = graph_capture_context.stream
# we use nullcontext now
maybe_ca_context = nullcontext()
# ensure all initialization operations complete before attempting to
# capture the graph on another stream
curr_stream = torch.npu.current_stream()
if curr_stream != stream:
stream.wait_stream(curr_stream)
with torch.npu.stream(stream), maybe_ca_context:
yield graph_capture_context
# Wrapper for ModelRunnerOutput to support overlapped execution.
class AsyncNPUModelRunnerOutput(AsyncModelRunnerOutput):
def __init__(
self,
model_runner_output: ModelRunnerOutput,
sampled_token_ids: torch.Tensor,
invalid_req_indices: list[int],
async_output_copy_stream: torch.npu.Stream,
):
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.
self._async_copy_ready_event = torch.npu.Event()
# 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
# Initiate the copy on a separate stream, but do not synchronize it.
default_stream = torch.npu.current_stream()
with torch.npu.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._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 tensor once the copy has completed
del self._sampled_token_ids
valid_sampled_token_ids = self._sampled_token_ids_cpu.tolist()
for i in self._invalid_req_indices:
valid_sampled_token_ids[i].clear()
output = self._model_runner_output
output.sampled_token_ids = valid_sampled_token_ids
return output
class NPUModelRunner(LoRAModelRunnerMixin):
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.compilation_config = vllm_config.compilation_config
self.load_config = vllm_config.load_config
self.lora_config = vllm_config.lora_config
self.parallel_config = vllm_config.parallel_config
self.pin_memory = is_pin_memory_available()
self.scheduler_config = vllm_config.scheduler_config
self.speculative_config = vllm_config.speculative_config
self.block_size = vllm_config.cache_config.block_size
self.max_num_blocks_per_req = cdiv(self.model_config.max_model_len,
self.block_size)
self.max_num_tokens = self.scheduler_config.max_num_batched_tokens
decode_max_num_seqs = getattr(self.scheduler_config,
'decode_max_num_seqs', 0)
self.max_num_reqs = max(self.scheduler_config.max_num_seqs,
decode_max_num_seqs)
self.dp_size = vllm_config.parallel_config.data_parallel_size
self.dp_rank = vllm_config.parallel_config.data_parallel_rank
self.device = device
if envs_ascend.VLLM_ASCEND_ENABLE_PREFETCH_MLP:
self.prefetch_stream = torch.npu.Stream(device=device)
else:
self.prefetch_stream = None
self.dtype = self.model_config.dtype
if envs_ascend.VLLM_ASCEND_ENABLE_TOPK_TOPP_OPTIMIZATION:
# TODO: drop the env config to use ascend sampler by default
from vllm_ascend.sample.sampler import AscendSampler
self.sampler = AscendSampler()
else:
from vllm.v1.sample.sampler import Sampler
self.sampler = Sampler()
self.reorder_batch_threshold: Optional[int] = None
# Lazy initialization, these will be set after __init__
self.kv_caches: List[torch.Tensor] = []
self.attn_groups: list[list[AttentionGroup]] = []
self.encoder_cache: Dict[str, torch.Tensor] = {}
self.attn_mask = None
self.attn_state = None
self.requests: Dict[str, CachedRequestState] = {}
self.intermediate_tensors: Optional[IntermediateTensors] = None
self.runner_only_attn_layers: set[str] = set()
self.ascend_config = get_ascend_config()
if self.ascend_config.ascend_scheduler_config.enabled:
self.chunked_prefill_enabled = self.scheduler_config.chunked_prefill_enabled
else:
self.chunked_prefill_enabled = True
self.weight_prefetch_method = WeightPrefetchMethod(
self.ascend_config.weight_prefetch_config)
if self.cache_config.cache_dtype == "auto":
self.kv_cache_dtype = self.dtype
else:
self.kv_cache_dtype = STR_DTYPE_TO_TORCH_DTYPE[
self.cache_config.cache_dtype]
# use_hybrid_blocks: if hybrid blocks is used.
self.use_hybrid_blocks: bool = False
self.need_accepted_tokens: bool = False
self.is_multimodal_model = self.model_config.is_multimodal_model
self.is_pooling_model = self.model_config.pooler_config is not None
if self.is_multimodal_model:
self.inputs_embeds = torch.zeros(
(self.max_num_tokens, self.model_config.get_hidden_size()),
dtype=self.dtype,
device=self.device)
# Set up Attention
self.attn_backend = get_attn_backend(
0,
self.dtype,
None,
self.block_size,
use_mla=self.model_config.use_mla,
use_sfa=self.ascend_config.use_sfa)
if torch.version.cann.startswith("8.3"):
self.attn_mask_builder = AttentionMaskBuilder(
self.scheduler_config.max_num_batched_tokens, self.dtype,
self.device)
else:
self.attn_mask_builder = AttentionMaskBuilder(
self.model_config.max_model_len, self.dtype)
# Set up speculative decoding.
self.spec_attn_mask = None
self.drafter: Optional[Union[NgramProposer, EagleProposer,
MtpProposer]] = None
self.actual_seq_lengths_q: list[int] = []
self.decode_token_per_req = 1
if self.speculative_config:
spec_token_num = self.speculative_config.num_speculative_tokens
assert spec_token_num > 0
self.decode_token_per_req = 1 + spec_token_num
self.spec_attn_mask = torch.triu(torch.ones(2048,
2048,
dtype=torch.bool),
diagonal=1).to(self.device)
if get_pp_group().is_last_rank:
self.drafter = get_spec_decode_method(
self.speculative_config.method, self.vllm_config,
self.device, self)
self.rejection_sampler = AscendRejectionSampler()
# Persistent batch.
self.input_ids = torch.zeros(self.max_num_tokens,
dtype=torch.int32,
device=self.device)
self.positions = torch.zeros(self.max_num_tokens,
dtype=torch.int64,
device=self.device)
self.query_start_loc = torch.zeros(self.max_num_reqs + 1,
dtype=torch.int32,
device=self.device)
self.seq_lens = torch.zeros(self.max_num_reqs,
dtype=torch.int32,
device=self.device)
self.slot_mapping = torch.zeros(self.max_num_tokens,
dtype=torch.int32,
device=self.device)
if self.vllm_config.model_config.use_mla and \
self.compilation_config.cudagraph_mode == CUDAGraphMode.FULL_DECODE_ONLY:
rope_dim = self.model_config.hf_text_config.qk_rope_head_dim
self.cos = torch.ones(self.max_num_reqs,
1,
1,
rope_dim,
dtype=self.dtype,
device=self.device)
self.sin = torch.zeros(self.max_num_reqs,
1,
1,
rope_dim,
dtype=self.dtype,
device=self.device)
else:
self.cos = None
self.sin = None
self.uses_mrope = self.model_config.uses_mrope
# 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 = torch.zeros((3, self.max_num_tokens + 1),
dtype=torch.int64,
device=self.device)
self.mrope_positions_cpu = torch.zeros(
(3, self.max_num_tokens + 1),
dtype=torch.int64,
device="cpu",
pin_memory=True)
self.mrope_positions_np = self.mrope_positions_cpu.numpy()
# OPTIMIZATION: Cache the tensors rather than creating them every step.
self.arange_np: npt.NDArray[np.int32] = np.arange(max(
self.max_num_reqs + 1, self.model_config.max_model_len,
self.max_num_tokens),
dtype=np.int32)
# NOTE(woosuk): These tensors are "stateless", i.e., they are literally
# a faster version of creating a new tensor every time. Thus, we should
# not make any assumptions about the values in these tensors.
self.input_ids_cpu = torch.zeros(self.max_num_tokens,
dtype=torch.int32,
device="cpu",
pin_memory=True)
self.positions_cpu = torch.zeros(self.max_num_tokens,
dtype=torch.int64,
device="cpu",
pin_memory=True)
self.positions_np = self.positions_cpu.numpy()
self.slot_mapping_cpu = torch.zeros(self.max_num_tokens,
dtype=torch.int32,
device="cpu",
pin_memory=True)
self.slot_mapping_np = self.slot_mapping_cpu.numpy()
self.query_start_loc_cpu = torch.zeros(self.max_num_reqs + 1,
dtype=torch.int32,
device="cpu",
pin_memory=True)
self.query_start_loc_np = self.query_start_loc_cpu.numpy()
self.seq_lens_cpu = torch.zeros(self.max_num_reqs,
dtype=torch.int32,
device="cpu",
pin_memory=True)
self.seq_lens_np = self.seq_lens_cpu.numpy()
self.use_aclgraph = self._use_aclgraph()
self.aclgraph_batch_sizes = list(
reversed(self.compilation_config.cudagraph_capture_sizes))
self.uniform_decode_query_len = 1 if not self.speculative_config else \
1 + self.speculative_config.num_speculative_tokens
# aclgraph dispatcher for runtime aclgraph dispatching.
self.aclgraph_dispatcher = CudagraphDispatcher(self.vllm_config)
# Cached outputs.
self._draft_token_ids: Optional[Union[list[list[int]],
torch.Tensor]] = None
# NOTE: we need to use `in_profile_run` to determine whether `enable_force_load_balance` is True
self.in_profile_run = False
# kv role
self.is_kv_producer = False
self.is_kv_consumer = False
if vllm_config.kv_transfer_config is not None:
self.is_kv_producer = vllm_config.kv_transfer_config.is_kv_producer
self.is_kv_consumer = vllm_config.kv_transfer_config.is_kv_consumer
# NOTE: To be clear, we need to make sure that during graph capture, the number of
# tokens is less than or equal to mc2_tokens_capacity. According to _set_cudagraph_sizes,
# the max number of tokens in graph is min(max_num_seqs * 2, 512).
if self.compilation_config.cudagraph_capture_sizes:
max_num_tokens = self.compilation_config.cudagraph_capture_sizes[0]
else:
max_num_tokens = self.max_num_reqs * self.uniform_decode_query_len
tp_size = self.parallel_config.tensor_parallel_size
# Use integer arithmetic for ceiling division.
num_tokens_per_tp_rank = (max_num_tokens + tp_size - 1) // tp_size
self.mc2_tokens_capacity = num_tokens_per_tp_rank * tp_size
self.reserved_mc2_mask = torch.zeros(
self.mc2_tokens_capacity,
dtype=torch.bool,
device=self.device,
)
self.dynamic_eplb = self.ascend_config.dynamic_eplb
if self.dynamic_eplb:
self.is_eplb_warmuped = False
self.policy_type = self.ascend_config.eplb_policy_type
self.eplb_loader = D2DExpertWeightLoader()
self.manager = Manager()
self.shared_dict = self.manager.dict({
"expert_map": None,
"moe_load": None,
"expert_maps": None
})
self.eplb_process = EplbProcess(shared_dict=self.shared_dict,
policy_type=self.policy_type,
enable_d2d=True)
self.process = self.eplb_process._launch_process()
ascend_config = get_ascend_config()
self.eplb_updator = EplbUpdator(ascend_config, self.eplb_loader,
self.eplb_process, self.process)
self.use_async_scheduling = self.scheduler_config.async_scheduling
self.async_output_copy_stream = torch.npu.Stream() if \
self.use_async_scheduling else None
# 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.
self.input_batch = InputBatch(
max_num_reqs=self.max_num_reqs,
max_model_len=self.model_config.max_model_len,
max_num_batched_tokens=self.max_num_tokens,
device=self.device,
pin_memory=self.pin_memory,
vocab_size=self.model_config.get_vocab_size(),
block_sizes=[self.block_size],
is_spec_decode=bool(self.vllm_config.speculative_config),
logitsprocs=build_logitsprocs(
self.vllm_config, self.device, self.pin_memory,
self.is_pooling_model,
self.vllm_config.model_config.logits_processors),
is_pooling_model=self.is_pooling_model,
kernel_block_sizes=[[self.vllm_config.cache_config.block_size]],
)
self.num_accepted_tokens = self._make_buffer(self.max_num_reqs,
dtype=torch.int64)
self.num_draft_tokens = self._make_buffer(self.max_num_reqs,
dtype=torch.int32)
def _make_buffer(self,
*size: Union[int, torch.SymInt],
dtype: torch.dtype,
numpy: bool = True) -> CpuGpuBuffer:
# Bfloat16 torch tensors cannot be directly cast to a numpy array, so
# if a bfloat16 buffer is needed without a corresponding numpy array,
# don't bother instantiating the numpy array.
return CpuGpuBuffer(*size,
dtype=dtype,
device=self.device,
pin_memory=self.pin_memory,
with_numpy=numpy)
def _update_states_after_model_execute(
self, output_token_ids: torch.Tensor) -> 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.model_config.is_hybrid or not self.speculative_config:
return
# Find the number of accepted tokens for each sequence.
num_accepted_tokens = (torch.cat(
[
output_token_ids,
torch.full((output_token_ids.size(0), 1),
-1,
device=output_token_ids.device),
],
dim=1) == -1).int().argmax(-1).cpu().numpy()
for i, num_tokens in enumerate(num_accepted_tokens):
self.input_batch.num_accepted_tokens_cpu[i] = num_tokens
def _use_aclgraph(self) -> bool:
return self.compilation_config.cudagraph_mode != CUDAGraphMode.NONE and self.compilation_config.level == CompilationLevel.PIECEWISE and not self.model_config.enforce_eager
def _update_states(self, scheduler_output: "SchedulerOutput") -> None:
# Remove finished requests from the cached states.
for req_id in scheduler_output.finished_req_ids:
self.requests.pop(req_id, None)
# 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)
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()
unscheduled_req_ids = cached_req_ids - scheduled_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)
req_ids_to_add: list[str] = []
# Add new requests to the cached states.
for new_req_data in scheduler_output.scheduled_new_reqs:
req_id = new_req_data.req_id
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 pooling_params:
assert (task := pooling_params.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)
backward_kwargs = {}
backward_kwargs["mm_features"] = new_req_data.mm_features
self.requests[req_id] = CachedRequestState(
req_id=req_id,
prompt_token_ids=new_req_data.prompt_token_ids,
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,
**backward_kwargs,
)
# Only relevant for models using M-RoPE (e.g, Qwen2-VL)
if self.uses_mrope:
self._init_mrope_positions(self.requests[req_id])
req_ids_to_add.append(req_id)
# Update the states of the running/resumed requests.
is_last_rank = get_pp_group().is_last_rank
req_data = scheduler_output.scheduled_cached_reqs
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_data.resumed_from_preemption[i]
# Update the cached states.
req_state.num_computed_tokens = num_computed_tokens
if not is_last_rank:
# 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.
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:])
# 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 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
req_index = self.input_batch.req_id_to_index.get(req_id)
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.
req_ids_to_add.append(req_id)
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:
# Add new_token_ids to token_ids_cpu.
start_token_index = num_computed_tokens
end_token_index = num_computed_tokens + len(new_token_ids)
self.input_batch.token_ids_cpu[
req_index,
start_token_index:end_token_index] = new_token_ids
self.input_batch.num_tokens_no_spec[
req_index] = end_token_index
self.input_batch.num_tokens[req_index] = end_token_index
# Add spec_token_ids to token_ids_cpu.
spec_token_ids = (
scheduler_output.scheduled_spec_decode_tokens.get(req_id, ()))
if spec_token_ids:
num_spec_tokens = len(spec_token_ids)
start_index = self.input_batch.num_tokens_no_spec[req_index]
end_token_index = start_index + num_spec_tokens
self.input_batch.token_ids_cpu[
req_index, start_index:end_token_index] = spec_token_ids
# NOTE(woosuk): `num_tokens` here may include spec tokens.
self.input_batch.num_tokens[req_index] += num_spec_tokens
# Add the new or resumed requests to the persistent batch.
# The smaller empty indices are filled first.
for req_id in req_ids_to_add:
req_state = self.requests[req_id]
self.input_batch.add_request(req_state)
# 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()
def _init_mrope_positions(self, req_state: CachedRequestState):
image_grid_thw = []
video_grid_thw = []
second_per_grid_ts = []
audio_feature_lengths = []
use_audio_in_video = False
assert req_state.mm_features is not None
for mm_feature in req_state.mm_features:
mm_item = mm_feature.data
if mm_item is None:
continue
mm_input = mm_item.get_data()
if (t := mm_input.get("image_grid_thw")) is not None:
image_grid_thw.append(t.tolist())
if (t := mm_input.get("video_grid_thw")) is not None:
video_grid_thw.append(t.tolist())
if (t := mm_input.get("second_per_grid_ts")) is not None:
second_per_grid_ts.append(t)
if (t := mm_input.get("audio_feature_lengths")) is not None:
audio_feature_lengths.append(t)
if mm_input.get("use_audio_in_video") is True:
use_audio_in_video = True
req_state.mrope_positions, req_state.mrope_position_delta = \
MRotaryEmbedding.get_input_positions_tensor(
req_state.prompt_token_ids,
hf_config=self.model_config.hf_config,
image_grid_thw=image_grid_thw,
video_grid_thw=video_grid_thw,
second_per_grid_ts=second_per_grid_ts,
audio_feature_lengths=audio_feature_lengths,
use_audio_in_video=use_audio_in_video,
)
def _sync_metadata_across_dp(
self, num_tokens: int, with_prefill: bool, enable_dbo: bool
) -> tuple[int, Optional[torch.Tensor], bool, bool]:
# TODO: In vLLM, the only thing that needs to be synced is num_tokens, but in
# our case, we still need to sync the other two flags as well. So we need to
# include them in the all_reduce operation, and more over, we CANNOT skip it
# even if we are running in eager mode, which harms performance.
# FIXME: Restore the `or self.vllm_config.model_config.enforce_eager` here
# immediately once the other two flags are no longer needed.
if self.dp_size == 1:
return num_tokens, None, with_prefill, enable_dbo
# Sync num_tokens, with_prefill, enable_dbo across dp ranks
num_tokens_tensor = torch.tensor([
num_tokens if i == self.dp_rank else 0 for i in range(self.dp_size)
],
dtype=torch.int32,
device="npu")
flags_tensor = torch.tensor(
[int(with_prefill), int(not enable_dbo)],
dtype=torch.int32,
device="npu")
packed_tensor = torch.cat([num_tokens_tensor, flags_tensor])
dist.all_reduce(packed_tensor, group=get_dp_group().device_group)
# Unpack the results
num_tokens_across_dp = packed_tensor[:-2]
synced_flags = packed_tensor[-2:]
max_tokens_across_dp = torch.max(num_tokens_across_dp).item()
global_with_prefill = bool(synced_flags[0])
global_enable_dbo = not bool(synced_flags[1])
# Create a tensor for num_tokens_after_padding
num_tokens_after_padding = torch.tensor([max_tokens_across_dp] *
self.dp_size,
device="npu",
dtype=torch.int32)
return max_tokens_across_dp, num_tokens_after_padding, global_with_prefill, global_enable_dbo
def _check_dbo_is_valid(self, query_lens: torch.Tensor,
attn_state: AscendAttentionState,
num_tokens: int) -> bool:
# do the checks for dp + dbo
if attn_state in [
AscendAttentionState.DecodeOnly,
AscendAttentionState.SpecDecoding
]:
return False
# considering the case that one dp rank may enable dbo while others may not
if not self.vllm_config.model_config.use_mla or not envs_ascend.VLLM_ASCEND_ENABLE_DBO:
return False
# TODO: remove it if token-level microbatch is enabled
[token_index,
seq_index] = compute_split_seq_index(query_lens, attn_state,
num_tokens)
if token_index == 0 or seq_index == 0 or seq_index == len(
query_lens) or num_tokens < 256:
return False
return True
def get_model(self) -> nn.Module:
# get raw model out of the aclgraph wrapper.
if isinstance(self.model, ACLGraphWrapper):
return self.model.unwrap()
return self.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")
return 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 _make_attention_mask(self, seq_lens, position,
attn_state) -> torch.Tensor:
# Chunk Prefill situation.
if attn_state == AscendAttentionState.ChunkedPrefill and not self.vllm_config.model_config.use_mla and not self.ascend_config.use_sfa:
if torch.version.cann.startswith("8.3"):
return self.attn_mask_builder.get_splitfuse_attn_mask()
else:
return self.attn_mask_builder.get_splitfuse_attn_mask(
seq_lens, position, self.dtype, self.device)
# Prefill without cache situation.
elif attn_state == AscendAttentionState.PrefillNoCache:
max_seq_len = max(seq_lens.max().item(), 0)
return self.attn_mask_builder.get_attn_mask(
max_seq_len, self.dtype, self.device)
# Prefill with cache hit.
elif attn_state == AscendAttentionState.PrefillCacheHit:
return self.attn_mask_builder.get_attn_mask(
128, self.dtype, self.device)
# Decode-only situation.
else:
return None
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 = len(req.prompt_token_ids)
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
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 _execute_mm_encoder(self, scheduler_output: "SchedulerOutput"):
scheduled_encoder_inputs = scheduler_output.scheduled_encoder_inputs
if not scheduled_encoder_inputs:
return
# Batch the multi-modal inputs.
mm_kwargs, mm_hashes_pos = self._batch_mm_kwargs_from_scheduler(
scheduler_output)
encoder_outputs = []
for _, num_items, mm_kwargs_group in group_mm_kwargs_by_modality(
mm_kwargs,
device=self.device,
pin_memory=True,
):
# Run the encoder.
# `curr_group_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.
curr_group_outputs = self.model.get_multimodal_embeddings(
**mm_kwargs_group)
sanity_check_mm_encoder_outputs(
curr_group_outputs,
expected_num_items=num_items,
)
for output in curr_group_outputs:
encoder_outputs.append(output)
for (mm_hash, pos_info), output in zip(mm_hashes_pos, encoder_outputs):
self.encoder_cache[mm_hash] = scatter_mm_placeholders(
output,
is_embed=pos_info.is_embed,
)
def _batch_mm_kwargs_from_scheduler(
self,
scheduler_output: "SchedulerOutput",
) -> tuple[list[MultiModalKwargsItem], list[tuple[str, PlaceholderRange]]]:
"""Batch multimodal kwargs from scheduled encoder inputs.
Args:
scheduler_output: The scheduler output containing scheduled encoder
inputs.
Returns:
A tuple of (mm_kwargs, req_ids_pos) where:
- mm_kwargs: List of multimodal kwargs items to be batched
- mm_hashes_pos: List of (mm_hash, position_info) tuples
"""
scheduled_encoder_inputs = scheduler_output.scheduled_encoder_inputs
if not scheduled_encoder_inputs:
return [], []
# Batch the multi-modal inputs.
mm_kwargs = list[MultiModalKwargsItem]()
# list of tuple (mm_hash, position_info)
mm_hashes_pos = list[tuple[str, PlaceholderRange]]()
for req_id, encoder_input_ids in scheduled_encoder_inputs.items():
req_state = self.requests[req_id]
assert req_state.mm_features is not None
for mm_input_id in encoder_input_ids:
mm_feature = req_state.mm_features[mm_input_id]
mm_hash = mm_feature.identifier
mm_kwargs.append(mm_feature.data)
mm_hashes_pos.append((mm_hash, mm_feature.mm_position))
return mm_kwargs, mm_hashes_pos
def _gather_mm_embeddings(
self,
scheduler_output: "SchedulerOutput",
) -> list[torch.Tensor]:
def _iter_mm_features(req_state: CachedRequestState):
assert req_state.mm_features is not None
for mm_feature in req_state.mm_features:
pos_info = mm_feature.mm_position
yield mm_feature.identifier, pos_info, getattr(
pos_info, "is_embed", None)
mm_embeds: list[torch.Tensor] = []
for req_id in self.input_batch.req_ids:
num_scheduled_tokens = scheduler_output.num_scheduled_tokens[
req_id]
req_state = self.requests[req_id]
num_computed_tokens = req_state.num_computed_tokens
for mm_hash, pos_info, is_embed in _iter_mm_features(req_state):
start_pos = pos_info.offset
num_encoder_tokens = pos_info.length
if start_pos >= num_computed_tokens + num_scheduled_tokens:
break
if start_pos + num_encoder_tokens <= num_computed_tokens:
continue
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
encoder_output = self.encoder_cache.get(mm_hash, None)
assert encoder_output is not None, \
f"Encoder cache miss for {mm_hash}."
if is_embed is not None:
is_embed = is_embed[start_idx:end_idx]
mm_embeds_item = gather_mm_placeholders(
encoder_output[start_idx:end_idx],
is_embed=is_embed,
)
mm_embeds.append(mm_embeds_item)
return mm_embeds
def _get_cumsum_and_arange(
self,
num_tokens: np.ndarray,
cumsum_dtype: Optional[np.dtype] = None,
) -> tuple[np.ndarray, np.ndarray]:
"""Get the cumulative sum and batched arange of the given array.
# E.g., [2, 5, 3] -> ([2, 7, 10], [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]
arange = self.arange_np[:total_num_tokens] - cumsums_offsets
return cu_num_tokens, arange
def _prepare_input_ids(self, 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
NPU need to be copied into the corresponding slots into input_ids."""
if self.input_batch.prev_sampled_token_ids is None:
# Normal scheduling case
self.input_ids[:total_num_scheduled_tokens].copy_(
self.input_ids_cpu[:total_num_scheduled_tokens],
non_blocking=True)
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 NPU from prev_sampled_token_ids.
prev_req_id_to_index = self.input_batch.prev_req_id_to_index
assert prev_req_id_to_index is not None
flattened_indices = []
prev_common_req_indices = []
indices_match = True
max_flattened_index = -1
for req_id, cur_index in self.input_batch.req_id_to_index.items():
if (prev_index := prev_req_id_to_index.get(req_id)) is not None:
prev_common_req_indices.append(prev_index)
# We need to compute the flattened input_ids index of the
# last token in each common request.
flattened_index = cu_num_tokens[cur_index].item() - 1
flattened_indices.append(flattened_index)
indices_match &= (prev_index == flattened_index)
max_flattened_index = max(max_flattened_index, flattened_index)
num_commmon_tokens = len(flattened_indices)
if num_commmon_tokens < total_num_scheduled_tokens:
# If not all requests are decodes from the last iteration,
# We need to copy the input_ids_cpu to the NPU first.
self.input_ids[:total_num_scheduled_tokens].copy_(
self.input_ids_cpu[:total_num_scheduled_tokens],
non_blocking=True)
if num_commmon_tokens == 0:
# No requests in common with the previous iteration
# So input_ids_cpu will have all the input ids.
return
if indices_match and max_flattened_index == (num_commmon_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[:num_commmon_tokens].copy_(
self.input_batch.prev_sampled_token_ids[:num_commmon_tokens,
0],
non_blocking=True)
return
# Upload the index tensors asynchronously
# so the scatter can be non-blocking.
input_ids_index_tensor = torch.tensor(flattened_indices,
dtype=torch.int64,
pin_memory=self.pin_memory).to(
self.device,
non_blocking=True)
prev_common_req_indices_tensor = torch.tensor(
prev_common_req_indices,
dtype=torch.int64,
pin_memory=self.pin_memory).to(self.device, non_blocking=True)
self.input_ids.scatter_(dim=0,
index=input_ids_index_tensor,
src=self.input_batch.prev_sampled_token_ids[
prev_common_req_indices_tensor, 0])
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_goups, 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 _prepare_inputs(
self,
scheduler_output: "SchedulerOutput",
intermediate_tensors: Optional[IntermediateTensors] = None,
) -> tuple[dict[str, Any], torch.Tensor, np.ndarray, int, torch.Tensor,
int, torch.Tensor, SpecDecodeMetadata, Optional[torch.Tensor],
Optional[torch.Tensor], Optional[torch.Tensor], int]:
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 the number of scheduled tokens for each request.
req_ids = self.input_batch.req_ids
tokens = [scheduler_output.num_scheduled_tokens[i] for i in req_ids]
num_scheduled_tokens = np.array(tokens, dtype=np.int32)
max_num_scheduled_tokens = num_scheduled_tokens.max()
num_valid_tokens = np.array([
num_tokens -
len(scheduler_output.scheduled_spec_decode_tokens.get(i, []))
for num_tokens, i in zip(tokens, req_ids)
],
dtype=np.int32)
if (self.use_aclgraph and total_num_scheduled_tokens
<= self.aclgraph_batch_sizes[-1]):
# Add padding to the batch size.
num_input_tokens = self.vllm_config.pad_for_cudagraph(
total_num_scheduled_tokens)
else:
# Eager mode.
num_input_tokens = total_num_scheduled_tokens
# Get the attention state.
attn_state = self._build_attn_state(num_reqs, num_scheduled_tokens,
num_valid_tokens)
self.attn_state = attn_state # type: ignore
# Determine if it's a splitfuse batch
with_prefill = attn_state not in [
AscendAttentionState.DecodeOnly, AscendAttentionState.SpecDecoding
]
self.query_lens = torch.from_numpy(num_scheduled_tokens)
enable_dbo = self._check_dbo_is_valid(self.query_lens.tolist(),
attn_state,
total_num_scheduled_tokens)
# Get info across DP ranks.
# NOTE: maybe_padded_num_tokens is only used when using TorchAir with DP,
# Otherwise, it's just max_tokens_across_dp_cpu
(maybe_padded_num_tokens, num_tokens_across_dp, with_prefill,
enable_dbo) = self._sync_metadata_across_dp(num_input_tokens,
with_prefill, enable_dbo)
# TODO: Now that num_input_tokens is basically identical with maybe_padded_num_tokens
# We should consider removing maybe_padded_num_tokens later
num_input_tokens = maybe_padded_num_tokens
# Hot-Swap lora model
if self.lora_config:
self.set_active_loras(self.input_batch, num_scheduled_tokens)
# 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]
# arange: [0, 1, 0, 1, 2, 3, 4, 0, 1, 2]
cu_num_tokens, arange = self._get_cumsum_and_arange(
num_scheduled_tokens)
positions_np = self.positions_np[:total_num_scheduled_tokens]
np.add(self.input_batch.num_computed_tokens_cpu[req_indices],
arange,
out=positions_np)
# 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)
# Only relevant for models using M-RoPE (e.g, Qwen2-VL)
self.mrope_positions[:, :total_num_scheduled_tokens].copy_(
self.mrope_positions_cpu[:, :total_num_scheduled_tokens],
non_blocking=True)
# 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])
# Prepare input_ids.
# 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,
torch.from_numpy(token_indices),
out=self.input_ids_cpu[:total_num_scheduled_tokens])
# Prepare some information for building Attention-Metadata
# Compute and commit slot mapping
self.input_batch.block_table.compute_slot_mapping(
req_indices, positions_np)
self.input_batch.block_table.commit_slot_mapping(
total_num_scheduled_tokens)
self.query_start_loc_np[0] = 0
self.query_start_loc_np[1:num_reqs + 1] = cu_num_tokens
self.query_start_loc[:num_reqs + 1].copy_(
self.query_start_loc_cpu[:num_reqs + 1], non_blocking=True)
self.seq_lens_np[:num_reqs] = (
self.input_batch.num_computed_tokens_cpu[:num_reqs] +
num_scheduled_tokens)
self.seq_lens[:num_reqs].copy_(self.seq_lens_cpu[:num_reqs],
non_blocking=True)
# Fill unused with -1. Needed for reshape_and_cache
self.query_start_loc[num_reqs + 1:].fill_(-1)
self.seq_lens[num_reqs:].fill_(0)
self.query_lens = torch.from_numpy(num_scheduled_tokens)
# Copy the tensors to the NPU.
self.input_ids[:total_num_scheduled_tokens].copy_(
self.input_ids_cpu[:total_num_scheduled_tokens], non_blocking=True)
self.positions_cpu[total_num_scheduled_tokens:num_input_tokens].zero_()
self.positions[:num_input_tokens].copy_(
self.positions_cpu[:num_input_tokens], non_blocking=True)
# Make Attention metadata
positions_cpu = self.positions_cpu[:num_input_tokens]
positions = self.positions[:num_input_tokens]
seq_lens_cpu = self.seq_lens_cpu[:num_reqs]
self.attn_mask = self._make_attention_mask(seq_lens=seq_lens_cpu,
position=positions_cpu,
attn_state=attn_state)
self.attn_state = attn_state # type: ignore
self.with_prefill = with_prefill
self.num_tokens_across_dp = num_tokens_across_dp
self._update_graph_pad_size(with_prefill, maybe_padded_num_tokens)
attn_metadata: dict[str, Any] = {}
# Prepare input_ids
token_indices = (positions_np +
req_indices * self.input_batch.token_ids_cpu.shape[1])
torch.index_select(self.input_batch.token_ids_cpu_tensor.flatten(),
0,
torch.from_numpy(token_indices),
out=self.input_ids_cpu[:total_num_scheduled_tokens])
# Copy the tensors to the NPU.
self._prepare_input_ids(total_num_scheduled_tokens, cu_num_tokens)
# _prepare_inputs may reorder the batch, so we must gather
# multi-modal outputs after that to ensure the correct order
if self.is_multimodal_model:
# Run the multimodal encoder if any.
self._execute_mm_encoder(scheduler_output)
mm_embeds = 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.
input_ids = self.input_ids[:total_num_scheduled_tokens]
if mm_embeds:
inputs_embeds = self.model.get_input_embeddings(
input_ids, mm_embeds)
else:
inputs_embeds = self.model.get_input_embeddings(input_ids)
# TODO(woosuk): Avoid the copy. Optimize.
self.inputs_embeds[:total_num_scheduled_tokens].copy_(
inputs_embeds)
inputs_embeds = self.inputs_embeds[:num_input_tokens]
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 ACL graph.
input_ids = self.input_ids[:num_input_tokens]
inputs_embeds = None
positions = self.positions[:num_input_tokens]
input_ids, positions = self._update_input_ids_and_positions(
input_ids, positions, num_input_tokens, with_prefill,
maybe_padded_num_tokens)
if get_pp_group().is_first_rank:
intermediate_tensors = None
else:
assert intermediate_tensors is not None
assert self.intermediate_tensors is not None
for k, v in intermediate_tensors.items():
self.intermediate_tensors[k][:num_input_tokens].copy_(
v[:num_input_tokens], non_blocking=True)
intermediate_tensors = IntermediateTensors({
k: v[:num_input_tokens]
for k, v in self.intermediate_tensors.items()
})
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.
spec_decode_metadata = None
logits_indices = torch.from_numpy(cu_num_tokens - 1).to(
self.device, non_blocking=True)
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 req_id, draft_token_ids in (
scheduler_output.scheduled_spec_decode_tokens.items()):
req_idx = self.input_batch.req_id_to_index[req_id]
num_draft_tokens[req_idx] = len(draft_token_ids)
spec_decode_metadata = self._calc_spec_decode_metadata(
num_draft_tokens, cu_num_tokens)
logits_indices = spec_decode_metadata.logits_indices
self.num_draft_tokens.np[:num_reqs] = num_draft_tokens
self.num_draft_tokens.np[num_reqs:].fill(0)
self.num_draft_tokens.copy_to_gpu()
# Used in the below loop.
# query_start_loc_cpu = self.query_start_loc.cpu[:num_reqs + 1]
num_computed_tokens_cpu = (
self.input_batch.num_computed_tokens_cpu_tensor[:num_reqs])
spec_decode_common_attn_metadata = None
if use_spec_decode and self.need_accepted_tokens:
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()
# Prepare the attention metadata for each KV cache group and make layers
# in the same group share the same metadata.
for kv_cache_group_id, kv_cache_group_spec in enumerate(
self.kv_cache_config.kv_cache_groups):
blk_table = self.input_batch.block_table[kv_cache_group_id]
blk_table_tensor = blk_table.get_device_tensor()
slot_mapping = blk_table.slot_mapping_cpu[:
total_num_scheduled_tokens]
self.slot_mapping[:total_num_scheduled_tokens].copy_(
slot_mapping[:total_num_scheduled_tokens],
non_blocking=True,
)
# Make AscendCommonAttentionMetadata
common_attn_metadata = AscendCommonAttentionMetadata(
query_start_loc=self.query_start_loc[:num_reqs + 1],
query_start_loc_cpu=self.query_start_loc_cpu[:num_reqs + 1],
seq_lens_cpu=self.seq_lens_cpu,
seq_lens=self.seq_lens_cpu[:num_reqs],
num_reqs=num_reqs,
num_actual_tokens=total_num_scheduled_tokens,
actual_seq_lengths_q=self.actual_seq_lengths_q,
# TODO: change this to the right block table for linear attn
block_table_tensor=blk_table_tensor[:num_reqs],
slot_mapping=self.slot_mapping,
num_computed_tokens_cpu=num_computed_tokens_cpu,
positions=self.positions,
attn_mask=self.attn_mask,
spec_attn_mask=self.spec_attn_mask,
attn_state=self.attn_state,
enable_dbo_across_dp=enable_dbo,
is_only_prefill=bool(np.all(num_valid_tokens != 1)),
max_query_len=max_num_scheduled_tokens,
graph_pad_size=self.graph_pad_size,
decode_token_per_req=self.decode_token_per_req,
cos=self.cos,
sin=self.sin,
)
if self.speculative_config and \
spec_decode_common_attn_metadata is None:
spec_decode_common_attn_metadata = common_attn_metadata
for attn_group in self.attn_groups[kv_cache_group_id]:
common_prefix_len = 0
extra_attn_metadata_args = {}
builder = attn_group.get_metadata_builder()
if isinstance(builder, GDNAttentionMetadataBuilder):
if use_spec_decode:
extra_attn_metadata_args = dict(
num_accepted_tokens=self.num_accepted_tokens.
gpu[:num_reqs],
num_draft_tokens=self.num_draft_tokens.
gpu[:num_reqs],
)
attn_metadata_i = builder.build(
common_prefix_len=common_prefix_len,
common_attn_metadata=common_attn_metadata,
**extra_attn_metadata_args)
else:
attn_metadata_i = builder.build(
common_prefix_len=common_prefix_len,
common_attn_metadata=common_attn_metadata,
model=self.get_model(),
**extra_attn_metadata_args)
if self.vllm_config.model_config.use_mla or self.ascend_config.use_sfa:
attn_metadata_i.num_input_tokens = num_input_tokens
for layer_name in attn_group.layer_names:
attn_metadata[layer_name] = attn_metadata_i
if lmhead_tp_enable():
max_num_reqs_across_dp = maybe_padded_num_tokens if not with_prefill else self.max_num_reqs
logits_indices = nn.functional.pad(
logits_indices,
(0, max_num_reqs_across_dp - logits_indices.shape[0]))
return (attn_metadata, positions, num_scheduled_tokens,
num_input_tokens, num_tokens_across_dp,
maybe_padded_num_tokens, logits_indices, spec_decode_metadata,
input_ids, inputs_embeds, intermediate_tensors,
max_num_scheduled_tokens)
def _generate_process_reqs_hidden_states(self, attn_metadata, with_prefill,
maybe_padded_num_tokens,
input_ids, positions,
intermediate_tensors,
inputs_embeds):
assert self.model is not None
hidden_states = self.model(
input_ids=input_ids,
positions=positions,
intermediate_tensors=intermediate_tensors,
inputs_embeds=inputs_embeds,
)
forward_context = get_forward_context()
if forward_context.cudagraph_runtime_mode == CUDAGraphMode.FULL:
# TODO: maybe_padded_num_tokens will be removed, use num_input_tokens instead
if self.vllm_config.model_config.use_mla:
# FIXME: Try using `auto_dispatch_capture=True`
update_mla_attn_params(self.update_stream, forward_context,
maybe_padded_num_tokens)
else:
update_attn_params(self.update_stream, forward_context,
maybe_padded_num_tokens)
if get_forward_context().sp_enabled:
hidden_states = tensor_model_parallel_all_gather(hidden_states, 0)
pad_size = get_forward_context().pad_size
if pad_size > 0:
hidden_states = hidden_states[:-pad_size, :]
return hidden_states
def _build_attn_state(self, num_reqs, num_scheduled_tokens,
num_valid_tokens):
ascend_config = get_ascend_config()
if np.array_equal(self.seq_lens_np[:num_reqs], num_scheduled_tokens):
attn_state = AscendAttentionState.PrefillNoCache
# We assume it is the decode stage, where prefill occurs but only one token is not hit in cache.
elif np.all(num_scheduled_tokens == 1):
attn_state = AscendAttentionState.DecodeOnly
if self.speculative_config and self.speculative_config.method == 'deepseek_mtp':
# SpecDecoding now supports seq_len=1 and seq_len=2
# In Prefilling Decoding Disaggregation scenario, SpecDecoding need to supports seq_len=1
attn_state = AscendAttentionState.SpecDecoding
# Speculative decoding.
elif np.all(num_valid_tokens == 1):
if self.drafter and (self.drafter.name == SpecDcodeType.EAGLE
or self.drafter.name == SpecDcodeType.EAGLE3):
attn_state = AscendAttentionState.ChunkedPrefill
else:
attn_state = AscendAttentionState.SpecDecoding
# splitfuse
elif not ascend_config.ascend_scheduler_config.enabled or self.chunked_prefill_enabled:
attn_state = AscendAttentionState.ChunkedPrefill
else:
attn_state = AscendAttentionState.PrefillCacheHit
return attn_state
def _update_graph_pad_size(self, with_prefill, graph_pad_size):
self.graph_pad_size = -1
def _update_input_ids_and_positions(self, input_ids, positions,
num_input_tokens, with_prefill,
maybe_padded_num_tokens):
if self.uses_mrope:
positions = self.mrope_positions[:, :num_input_tokens]
return input_ids, positions
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. [4, 5, 8, 9, 11]
cu_num_sampled_tokens = np.cumsum(num_sampled_tokens, dtype=np.int32)
total_num_sampled_tokens = cu_num_sampled_tokens[-1]
# Step 2. [0, 0, 0, 0, 4, 5, 5, 5, 8, 9, 9]
cumsums_offsets = np.repeat(cu_num_sampled_tokens - num_sampled_tokens,
num_sampled_tokens)
# Step 3. [0, 1, 2, 3, 0, 0, 1, 2, 0, 0, 1]
arange = self.arange_np[:total_num_sampled_tokens] - cumsums_offsets
# Step 4. [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 5. [0, 1, 2, 3, 103, 104, 105, 106, 206, 207, 208]
logits_indices += arange
# Compute the bonus logits indices.
bonus_logits_indices = cu_num_sampled_tokens - 1
# Compute the draft logits indices.
# [3, 3, 5, 5, 6]
cu_num_draft_tokens = np.cumsum(num_draft_tokens, dtype=np.int32)
total_num_draft_tokens = cu_num_draft_tokens[-1]
# [0, 0, 0, 3, 3, 5]
cumsums_offsets = np.repeat(cu_num_draft_tokens - num_draft_tokens,
num_draft_tokens)
# [0, 1, 2, 0, 1, 0]
arange = self.arange_np[:total_num_draft_tokens] - cumsums_offsets
# [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 += arange
# TODO: Optimize the CPU -> NPU copy.
cu_num_draft_tokens = torch.from_numpy(cu_num_draft_tokens).to(
self.device, non_blocking=True)
logits_indices = torch.from_numpy(logits_indices).to(self.device,
non_blocking=True)
target_logits_indices = torch.from_numpy(target_logits_indices).to(
self.device, non_blocking=True)
bonus_logits_indices = torch.from_numpy(bonus_logits_indices).to(
self.device, non_blocking=True)
# Compute the draft token ids.
# draft_token_indices: [ 1, 2, 3, 105, 106, 208]
draft_token_ids = self.input_ids[logits_indices]
draft_token_ids = draft_token_ids[target_logits_indices + 1]
metadata = SpecDecodeMetadata(
draft_token_ids=draft_token_ids,
num_draft_tokens=num_draft_tokens.tolist(),
cu_num_draft_tokens=cu_num_draft_tokens,
target_logits_indices=target_logits_indices,
bonus_logits_indices=bonus_logits_indices,
logits_indices=logits_indices,
)
return metadata
def apply_grammar_bitmask(
self,
scheduler_output: "SchedulerOutput",
logits: torch.Tensor,
) -> torch.Tensor:
grammar_bitmask = scheduler_output.grammar_bitmask
# We receive the structured output bitmask from the scheduler,
# compacted to contain bitmasks only for structured output requests.
# The order of the requests in the bitmask is not guaranteed to be the
# same as the order of the requests in the gpu runner's batch. We need
# to sort the bitmask to match the order of the requests used here.
# Get the batch indices of the structured output requests.
# Keep track of the number of speculative tokens scheduled for every
# request in the batch, as the logit indices are offset by this amount.
struct_out_req_batch_indices: dict[str, int] = {}
cumulative_offset = 0
seq = sorted(self.input_batch.req_id_to_index.items(),
key=lambda x: x[1])
for req_id, batch_index in seq:
logit_index = batch_index + cumulative_offset
cumulative_offset += len(
scheduler_output.scheduled_spec_decode_tokens.get(req_id, []))
if req_id in scheduler_output.structured_output_request_ids:
struct_out_req_batch_indices[req_id] = logit_index
out_indices = []
# Reorder the bitmask to match the order of the requests in the batch.
sorted_bitmask = np.zeros_like(grammar_bitmask,
shape=(logits.shape[0],
grammar_bitmask.shape[1]))
cumulative_index = 0
seq = sorted(scheduler_output.structured_output_request_ids.items(),
key=lambda x: x[1])
for req_id, _ in seq:
logit_index = struct_out_req_batch_indices[req_id]
num_spec_tokens = len(
scheduler_output.scheduled_spec_decode_tokens.get(req_id, []))
for i in range(1 + num_spec_tokens):
sorted_bitmask[logit_index + i] = \
grammar_bitmask[cumulative_index + i]
out_indices.append(logit_index + i)
cumulative_index += 1 + num_spec_tokens
grammar_bitmask = sorted_bitmask
# Serialization of np.ndarray is much more efficient than a tensor,
# so we receive it in that format.
grammar_bitmask = torch.from_numpy(grammar_bitmask)
# NOTE:
# 1. XGrammar bitmask applying only supports CPU and GPU.
# 2. The logits and bitmask should be on the same device.
# 3. XGrammar logits on CPU only supports float32 dtype.
logits_dtype = logits.dtype
logits = logits.to("cpu").float()
xgr.apply_token_bitmask_inplace(
logits,
grammar_bitmask,
indices=out_indices,
)
return logits.to(self.device).to(logits_dtype)
def propose_draft_token_ids(
self,
valid_sampled_token_ids: list[list[int]],
sampling_metadata: SamplingMetadata,
scheduler_output: "SchedulerOutput",
spec_decode_metadata: SpecDecodeMetadata,
positions: torch.Tensor,
num_scheduled_tokens: int,
hidden_states: torch.Tensor,
attn_metadata: dict[str, Any],
aux_hidden_states: torch.Tensor = None,
) -> Optional[list[list[int]]]:
if not self.drafter:
# Speculative decoding is not enabled.
draft_token_ids = None
else:
draft_token_ids = self.drafter.generate_token_ids(
valid_sampled_token_ids, sampling_metadata, scheduler_output,
spec_decode_metadata, positions, num_scheduled_tokens,
hidden_states, attn_metadata, aux_hidden_states)
return draft_token_ids
def _pool(
self,
hidden_states: torch.Tensor,
num_scheduled_tokens: int,
num_scheduled_tokens_np: np.ndarray,
finished_sending: Optional[set[str]] = None,
finished_recving: Optional[set[str]] = None,
kv_connector_output: Optional["KVConnectorOutput"] = None,
) -> ModelRunnerOutput:
assert self.input_batch.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]
pooling_metadata = self.input_batch.pooling_metadata
pooling_metadata.build_pooling_cursor(num_scheduled_tokens_np.tolist(),
device=hidden_states.device)
seq_lens_cpu = self.seq_lens_cpu[:self.input_batch.num_reqs]
model = cast(VllmModelForPooling, self.model)
raw_pooler_output = model.pooler(
hidden_states=hidden_states,
pooling_metadata=pooling_metadata,
)
raw_pooler_output = json_map_leaves(
lambda x: x.to("cpu", non_blocking=True),
raw_pooler_output,
)
torch.npu.synchronize()
pooler_output: list[Optional[torch.Tensor]] = []
for raw_output, seq_len, prompt_len in zip(
raw_pooler_output, seq_lens_cpu, pooling_metadata.prompt_lens):
output = raw_output if seq_len == prompt_len else None
pooler_output.append(output)
return ModelRunnerOutput(
req_ids=self.input_batch.req_ids,
req_id_to_index=self.input_batch.req_id_to_index,
sampled_token_ids=[],
logprobs=None,
prompt_logprobs_dict={},
pooler_output=pooler_output,
kv_connector_output=kv_connector_output,
)
def _select_moe_comm_method(self, num_tokens: int,
with_prefill: bool) -> MoECommType:
"""1. If expert parallel is not enabled, we use all-gather since MC2 and all-to-all
are designed for expert parallelism.
2. If expert parallel is enabled, we need to consider the soc version and the
number of tokens. This is based on the observation that all-gather is more
efficient than all-to-all when running on A2.
a. For A2, we choose from MC2 and all-gather.
b. For A3, we choose from MC2 and all-to-all.
In both cases, we use MC2 when the number of tokens is smaller than
a its capacity threshold.
Args:
num_tokens (int): The number of tokens in the current batch.
Raises:
ValueError: If the soc version is unsupported.
Returns:
MoECommType: The selected MoE communication method.
"""
soc_version = get_ascend_soc_version()
quant_type = getattr(self.vllm_config.model_config.hf_config,
'moe_quantize', None)
model_type = self.vllm_config.model_config.hf_config.model_type
if not self.parallel_config.enable_expert_parallel:
moe_comm_type = MoECommType.ALLGATHER
elif soc_version in {AscendSocVersion.A2}:
if (num_tokens <= self.mc2_tokens_capacity
and self.parallel_config.world_size_across_dp >= 16):
moe_comm_type = MoECommType.MC2
else:
# Currently, w4a8_dynamic does not support allgatherep
if quant_type == "w4a8_dynamic":
moe_comm_type = MoECommType.ALLTOALL
else:
moe_comm_type = MoECommType.ALLGATHER
elif soc_version in {AscendSocVersion.A3}:
moe_comm_type = (MoECommType.MC2
if num_tokens <= self.mc2_tokens_capacity else
MoECommType.ALLTOALL)
else:
raise ValueError(f"Unsupported soc_version: {soc_version}")
if moe_comm_type == MoECommType.ALLGATHER and with_prefill:
moe_comm_type = MoECommType.NAIVE_MULTICAST
# PanguProMoE only supports allgather
if model_type == "PanguProMoE":
moe_comm_type = MoECommType.ALLGATHER
if is_global_first_rank():
logger.debug(f"num_tokens: {num_tokens}, "
f"moe_comm_type: {moe_comm_type}")
return moe_comm_type
@torch.inference_mode()
def execute_model(
self,
scheduler_output: "SchedulerOutput",
intermediate_tensors: Optional[IntermediateTensors] = None,
) -> Union[ModelRunnerOutput, AsyncModelRunnerOutput, IntermediateTensors]:
with ProfileExecuteDuration().capture_async("prepare input"):
self._update_states(scheduler_output)
if not scheduler_output.total_num_scheduled_tokens:
if not has_kv_transfer_group():
logger.debug(
"skip this step for we receive the data from remote disaggregate prefill node"
)
# Return empty ModelRunnerOuptut if there's no work to do.
return EMPTY_MODEL_RUNNER_OUTPUT
return self.kv_connector_no_forward(scheduler_output)
if self.dynamic_eplb:
self.eplb_updator.forward_before()
(attn_metadata, positions, num_scheduled_tokens_np,
num_input_tokens, num_tokens_across_dp, maybe_padded_num_tokens,
logits_indices, spec_decode_metadata, input_ids, inputs_embeds,
intermediate_tensors,
max_query_len) = (self._prepare_inputs(scheduler_output,
intermediate_tensors))
if self.dynamic_eplb:
self.eplb_updator.take_update_info_from_eplb_process()
moe_comm_type = self._select_moe_comm_method(num_input_tokens,
self.with_prefill)
uniform_decode = (max_query_len == self.uniform_decode_query_len) and (
scheduler_output.total_num_scheduled_tokens
== self.input_batch.num_reqs * max_query_len)
batch_descriptor = BatchDescriptor(num_tokens=num_input_tokens,
uniform_decode=uniform_decode)
aclgraph_runtime_mode, batch_descriptor = \
self.aclgraph_dispatcher.dispatch(batch_descriptor)
# Run forward pass
with ProfileExecuteDuration().capture_async("forward"):
with set_ascend_forward_context(
attn_metadata,
self.vllm_config,
num_tokens=num_input_tokens,
num_tokens_across_dp=num_tokens_across_dp,
with_prefill=self.with_prefill,
reserved_mc2_mask=self.reserved_mc2_mask,
moe_comm_type=moe_comm_type,
aclgraph_runtime_mode=aclgraph_runtime_mode,
batch_descriptor=batch_descriptor,
num_actual_tokens=scheduler_output.
total_num_scheduled_tokens,
prefetch_stream=self.prefetch_stream,
model_instance=self.model,
weight_prefetch_method=self.weight_prefetch_method):
self.maybe_setup_kv_connector(scheduler_output)
hidden_states = self._generate_process_reqs_hidden_states(
attn_metadata, self.with_prefill, maybe_padded_num_tokens,
input_ids, positions, intermediate_tensors, inputs_embeds)
self.maybe_wait_for_kv_save()
finished_sending, finished_recving = self.get_finished_kv_transfer(
scheduler_output)
aux_hidden_states = None
if self.drafter and self.drafter.name == SpecDcodeType.EAGLE3:
hidden_states, aux_hidden_states = hidden_states
kv_connector_output = KVConnectorOutput(
finished_sending=finished_sending,
finished_recving=finished_recving)
finished_sending = None
finished_recving = None
with ProfileExecuteDuration().capture_async("post process"):
# Broadcast PP output for external_launcher (torchrun)
# to make sure we are synced across pp ranks
# TODO: Support overlapping mirco-batches
# https://github.com/vllm-project/vllm/issues/18019
broadcast_pp_output = \
self.parallel_config.distributed_executor_backend \
== "external_launcher" and len(get_pp_group().ranks) > 0
if not get_pp_group().is_last_rank:
# For mid-pipeline stages, return the hidden states.
if not broadcast_pp_output:
hidden_states.kv_connector_output = kv_connector_output
return hidden_states
assert isinstance(hidden_states, IntermediateTensors)
get_pp_group().send_tensor_dict(
hidden_states.tensors, all_gather_group=get_tp_group())
logits = None
else:
if self.input_batch.pooling_params:
return self._pool(
hidden_states,
scheduler_output.total_num_scheduled_tokens,
num_scheduled_tokens_np, finished_sending,
finished_recving, kv_connector_output)
sample_hidden_states = hidden_states[logits_indices]
logits = self.model.compute_logits(sample_hidden_states)
if broadcast_pp_output:
model_output_broadcast_data = {
"logits": logits.contiguous(),
} if logits is not None else {}
model_output_broadcast_data = get_pp_group(
).broadcast_tensor_dict(model_output_broadcast_data,
src=len(get_pp_group().ranks) - 1)
assert model_output_broadcast_data is not None
logits = model_output_broadcast_data["logits"]
# Apply structured output bitmasks if present
if scheduler_output.grammar_bitmask is not None:
logits = self.apply_grammar_bitmask(scheduler_output, logits)
# Sample the next token and get logprobs if needed.
sampling_metadata = self.input_batch.sampling_metadata
if spec_decode_metadata is None:
if lmhead_tp_enable() and logits is not None:
logits = logits[:self.input_batch.num_reqs]
sampler_output = self.sampler(
logits=logits,
sampling_metadata=sampling_metadata,
)
else:
if lmhead_tp_enable() and logits is not None:
logits = logits[:len(spec_decode_metadata.logits_indices)]
# When indexing with a tensor (bonus_logits_indices), PyTorch
# creates a new tensor with separate storage from the original
# logits tensor. This means any in-place operations on bonus_logits
# won't affect the original logits tensor.
assert logits is not None
bonus_logits = logits[
spec_decode_metadata.bonus_logits_indices]
sampler_output = self.sampler(
logits=bonus_logits,
sampling_metadata=sampling_metadata,
)
bonus_token_ids = sampler_output.sampled_token_ids
# Just like `bonus_logits`, `target_logits` is a new tensor with
# separate storage from the original `logits` tensor. Therefore,
# it is safe to update `target_logits` in place.
target_logits = logits[
spec_decode_metadata.target_logits_indices]
output_token_ids = self.rejection_sampler(
spec_decode_metadata,
None, # draft_probs
target_logits,
bonus_token_ids,
sampling_metadata,
)
sampler_output.sampled_token_ids = output_token_ids
if self.need_accepted_tokens:
self._update_states_after_model_execute(output_token_ids)
discard_sampled_tokens_req_indices: list[int] = []
# TODO(woosuk): The following loop can be slow since it iterates over
# the requests one by one. Optimize.
discard_sampled_tokens_req_indices = []
for i, req_id in enumerate(self.input_batch.req_ids):
req_state = self.requests[req_id]
seq_len = (req_state.num_computed_tokens +
scheduler_output.num_scheduled_tokens[req_id])
if seq_len < req_state.num_tokens:
# Ignore the sampled token.
# Rewind the generator state as if the token was not sampled.
generator = self.input_batch.generators.get(i)
if generator is not None:
generator.set_offset(generator.get_offset() - 4)
discard_sampled_tokens_req_indices.append(i)
# 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()
# NOTE: NPU -> CPU Sync happens here.
# Move as many CPU operations as possible before this sync point.
logprobs_tensors = sampler_output.logprobs_tensors
logprobs_lists = logprobs_tensors.tolists() \
if logprobs_tensors is not None else None
# Compute prompt logprobs if needed.
prompt_logprobs_dict = self._get_prompt_logprobs_dict(
hidden_states[:scheduler_output.total_num_scheduled_tokens],
scheduler_output,
)
num_sampled_tokens = sampler_output.sampled_token_ids.shape[0]
sampled_token_ids = sampler_output.sampled_token_ids
if not self.use_async_scheduling:
# 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 = sampled_token_ids.tolist()
else:
# Includes spec decode tokens.
valid_sampled_token_ids = self.rejection_sampler.parse_output(
sampled_token_ids,
self.input_batch.vocab_size,
)
# Mask out the sampled tokens that should not be sampled.
for i in discard_sampled_tokens_req_indices:
valid_sampled_token_ids[i].clear()
else:
valid_sampled_token_ids = []
invalid_req_indices = list(discard_sampled_tokens_req_indices)
invalid_req_indices_set = set(invalid_req_indices)
assert sampled_token_ids.shape[-1] == 1
# Cache the sampled tokens on the NPU and avoid CPU sync.
# These will be copied into input_ids in the next step
# when preparing inputs.
self.input_batch.prev_sampled_token_ids = \
sampled_token_ids
self.input_batch.prev_sampled_token_ids_invalid_indices = \
invalid_req_indices_set
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.
for req_idx in range(num_sampled_tokens):
if self.use_async_scheduling:
sampled_ids = [-1] * 1 if \
req_idx not in invalid_req_indices_set else None
else:
sampled_ids = valid_sampled_token_ids[req_idx]
if not sampled_ids:
continue
start_idx = self.input_batch.num_tokens_no_spec[req_idx]
end_idx = start_idx + len(sampled_ids)
assert end_idx <= self.model_config.max_model_len, (
"Sampled token IDs exceed the max model length. "
f"Total number of tokens: {end_idx} > max_model_len: "
f"{self.model_config.max_model_len}")
self.input_batch.token_ids_cpu[req_idx,
start_idx:end_idx] = sampled_ids
self.input_batch.num_tokens_no_spec[req_idx] = end_idx
self.input_batch.num_tokens[req_idx] = end_idx
req_id = self.input_batch.req_ids[req_idx]
req_state = self.requests[req_id]
req_state.output_token_ids.extend(sampled_ids)
if self.speculative_config:
self._draft_token_ids = self.propose_draft_token_ids(
valid_sampled_token_ids,
sampling_metadata,
scheduler_output,
spec_decode_metadata,
positions,
scheduler_output.total_num_scheduled_tokens,
hidden_states,
attn_metadata,
aux_hidden_states,
)
if has_kv_transfer_group():
get_kv_transfer_group().clear_connector_metadata()
extra_args = ({"kv_connector_output": kv_connector_output})
model_runner_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,
pooler_output=[],
**extra_args,
)
durations = ProfileExecuteDuration().pop_captured_sync()
if durations:
dr_str = [
f"[{tag}]:{duration:.2f}ms"
for tag, duration in durations.items()
]
captured_name = "Decode" if self.attn_state == AscendAttentionState.DecodeOnly else "Prefill"
logger.info("Profile execute duration [%s]:%s", captured_name,
" ".join(dr_str))
if self.dynamic_eplb:
self.eplb_updator.forward_end()
if not self.use_async_scheduling:
return model_runner_output
return AsyncNPUModelRunnerOutput(
model_runner_output=model_runner_output,
sampled_token_ids=sampled_token_ids,
invalid_req_indices=invalid_req_indices,
async_output_copy_stream=self.async_output_copy_stream,
)
def take_draft_token_ids(self) -> Optional[DraftTokenIds]:
if self._draft_token_ids is None:
return None
req_ids = self.input_batch.req_ids
if isinstance(self._draft_token_ids, torch.Tensor):
draft_token_ids = self._draft_token_ids.tolist()
else:
draft_token_ids = self._draft_token_ids
self._draft_token_ids = None
return DraftTokenIds(req_ids, draft_token_ids)
def kv_connector_no_forward(
self, scheduler_output: "SchedulerOutput") -> ModelRunnerOutput:
with set_ascend_forward_context(None, self.vllm_config):
self.maybe_setup_kv_connector(scheduler_output)
finished_sending, finished_recving = (
self.get_finished_kv_transfer(scheduler_output))
# For the case of no forward caused by receiving remote kv,
# one round of dummy inference is necessary
# to prevent hang over the collective calls.
output = copy.copy(EMPTY_MODEL_RUNNER_OUTPUT)
output.kv_connector_output = KVConnectorOutput(
finished_sending=finished_sending,
finished_recving=finished_recving)
return output
@staticmethod
def maybe_setup_kv_connector(scheduler_output: "SchedulerOutput"):
# Update KVConnector with the KVConnector metadata forward().
if has_kv_transfer_group():
kv_connector = get_kv_transfer_group()
assert isinstance(kv_connector, KVConnectorBase_V1)
assert scheduler_output.kv_connector_metadata is not None
kv_connector.bind_connector_metadata(
scheduler_output.kv_connector_metadata)
kv_connector.start_load_kv(get_forward_context())
@staticmethod
def maybe_wait_for_kv_save() -> None:
if has_kv_transfer_group():
get_kv_transfer_group().wait_for_save()
@staticmethod
def get_finished_kv_transfer(
scheduler_output: "SchedulerOutput",
) -> tuple[Optional[set[str]], Optional[set[str]]]:
if has_kv_transfer_group():
return get_kv_transfer_group().get_finished(
scheduler_output.finished_req_ids)
return None, None
def _build_attention_metadata(self, create_mixed_batch, num_reqs,
num_tokens, max_query_len, force_attention):
attn_metadata: Optional[dict[str, Any]] = None
if force_attention:
attn_metadata = {}
if create_mixed_batch:
raise NotImplementedError(
"force_attention=True is not supported for mixed batches.")
else:
seq_lens = self.model_config.max_model_len
self.seq_lens_np[:num_reqs] = seq_lens
self.seq_lens_np[num_reqs:] = 0
num_computed_tokens_cpu = (
self.input_batch.num_computed_tokens_cpu_tensor[:num_reqs])
for kv_cache_group_id, kv_cache_group_spec in enumerate(
self.kv_cache_config.kv_cache_groups):
block_table_tensor = self.input_batch.block_table[
kv_cache_group_id].get_device_tensor()
common_attn_metadata = AscendCommonAttentionMetadata(
query_start_loc=self.query_start_loc[:num_reqs + 1],
query_start_loc_cpu=self.query_start_loc_cpu[:num_reqs +
1],
seq_lens_cpu=self.seq_lens_cpu,
seq_lens=self.seq_lens_cpu[:num_reqs],
num_reqs=num_reqs,
num_actual_tokens=num_tokens,
actual_seq_lengths_q=self.actual_seq_lengths_q,
block_table_tensor=block_table_tensor[:num_reqs],
slot_mapping=self.slot_mapping,
num_computed_tokens_cpu=num_computed_tokens_cpu,
positions=self.positions,
attn_mask=self.attn_mask,
spec_attn_mask=self.spec_attn_mask,
attn_state=self.attn_state,
max_query_len=max_query_len,
decode_token_per_req=self.decode_token_per_req,
cos=self.cos,
sin=self.sin,
)
for attn_group in self.attn_groups[kv_cache_group_id]:
builder = attn_group.get_metadata_builder()
attn_metadata_i = builder.build_for_graph_capture(
common_attn_metadata, AscendAttentionState.DecodeOnly,
self.get_model())
for layer_name in kv_cache_group_spec.layer_names:
attn_metadata[layer_name] = attn_metadata_i
return attn_metadata
def _generate_dummy_run_hidden_states(self, with_prefill,
is_torchair_compile, input_ids,
positions, attn_metadata, num_tokens,
intermediate_tensors, inputs_embeds):
hidden_states = self.model(input_ids=input_ids,
positions=positions,
intermediate_tensors=intermediate_tensors,
inputs_embeds=inputs_embeds)
forward_context = get_forward_context()
assert forward_context is not None
if forward_context.cudagraph_runtime_mode == CUDAGraphMode.FULL and \
not forward_context.capturing:
if self.vllm_config.model_config.use_mla:
# FIXME: Try using `auto_dispatch_capture=True`
update_mla_attn_params(self.update_stream, forward_context,
positions.shape[0])
else:
update_attn_params(self.update_stream, forward_context,
positions.shape[0])
if self.drafter and self.drafter.name == SpecDcodeType.EAGLE3:
hidden_states, _ = hidden_states
else:
hidden_states = hidden_states
return hidden_states
@torch.inference_mode()
def _dummy_run(
self,
num_tokens: int,
with_prefill: bool = False,
is_torchair_compile: bool = False,
aclgraph_runtime_mode: Optional[CUDAGraphMode] = None,
force_attention: bool = False,
uniform_decode: bool = False,
) -> torch.Tensor:
# only support eager mode and piecewise graph now
assert aclgraph_runtime_mode is None or aclgraph_runtime_mode in {
CUDAGraphMode.NONE, CUDAGraphMode.PIECEWISE, CUDAGraphMode.FULL
}
# Padding for DP
(num_tokens, num_tokens_across_dp, with_prefill,
_) = self._sync_metadata_across_dp(num_tokens, with_prefill, False)
moe_comm_type = self._select_moe_comm_method(num_tokens, with_prefill)
# If cudagraph_mode.decode_mode() == FULL and
# cudagraph_mode.seperate_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.scheduler_config.max_num_batched_tokens
max_num_reqs = self.scheduler_config.max_num_seqs
if uniform_decode:
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:
if with_prefill:
num_reqs = num_tokens
else:
num_reqs = (num_tokens + self.decode_token_per_req -
1) // self.decode_token_per_req
num_reqs = min(num_reqs, 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)
# Force dummy run on prefill stage when this node is deemed as kv producer.
if self.is_kv_producer and not self.is_kv_consumer:
with_prefill = True
# TODO(cmq): check if with_prefill is reasonable
attn_metadata = self._build_attention_metadata(
False,
num_reqs=num_reqs,
num_tokens=num_tokens,
max_query_len=max_query_len,
force_attention=force_attention,
)
if not self.in_profile_run and self.dynamic_eplb:
self.eplb_updator.forward_before()
with self.maybe_dummy_run_with_lora(self.lora_config,
num_scheduled_tokens):
if self.is_multimodal_model:
input_ids = None
inputs_embeds = self.inputs_embeds[:num_tokens]
else:
input_ids = self.input_ids[:num_tokens]
inputs_embeds = None
if self.uses_mrope:
positions = self.mrope_positions[:, :num_tokens]
else:
positions = self.positions[:num_tokens]
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=num_tokens,
dtype=self.dtype,
device=self.device))
intermediate_tensors = IntermediateTensors({
k: v[:num_tokens]
for k, v in self.intermediate_tensors.items()
})
# filter out the valid batch descriptor
_ag_mode, batch_descriptor = \
self.aclgraph_dispatcher.dispatch(
BatchDescriptor(num_tokens=num_tokens,
uniform_decode=uniform_decode))
if aclgraph_runtime_mode is not None:
# we allow forcing NONE when the dispatcher disagrees to support
# warm ups for aclgraph capture
assert aclgraph_runtime_mode == CUDAGraphMode.NONE or \
aclgraph_runtime_mode == _ag_mode, (
f"Aclgraph runtime mode mismatch at dummy_run. "
f"Expected {_ag_mode}, but got {aclgraph_runtime_mode}.")
else:
aclgraph_runtime_mode = _ag_mode
need_dummy_logits = (not self.in_profile_run
and lmhead_tp_enable())
if need_dummy_logits:
max_num_reqs_across_dp = num_tokens if not with_prefill else max_num_reqs
dummy_indices = torch.zeros(max_num_reqs_across_dp,
dtype=torch.int32)
def dummy_compute_logits(hidden_states):
return self.model.compute_logits(
hidden_states[dummy_indices])
with set_ascend_forward_context(
attn_metadata,
self.vllm_config,
num_tokens=num_tokens,
num_tokens_across_dp=num_tokens_across_dp,
with_prefill=with_prefill,
in_profile_run=self.in_profile_run,
reserved_mc2_mask=self.reserved_mc2_mask,
moe_comm_type=moe_comm_type,
num_actual_tokens=0,
aclgraph_runtime_mode=aclgraph_runtime_mode,
batch_descriptor=batch_descriptor,
prefetch_stream=self.prefetch_stream,
model_instance=self.model,
weight_prefetch_method=self.weight_prefetch_method):
hidden_states = self._generate_dummy_run_hidden_states(
with_prefill, is_torchair_compile, input_ids, positions,
attn_metadata, num_tokens, intermediate_tensors,
inputs_embeds)
if need_dummy_logits:
dummy_compute_logits(hidden_states)
if self.drafter:
self.drafter.dummy_run(
num_tokens=num_tokens,
with_prefill=with_prefill,
skip_attn=True,
num_reqs=num_reqs,
num_tokens_across_dp=num_tokens_across_dp)
if need_dummy_logits:
dummy_compute_logits(hidden_states)
if self.in_profile_run and self.dynamic_eplb:
self.model.clear_all_moe_loads()
if not self.in_profile_run and self.dynamic_eplb:
self.eplb_updator.take_update_info_from_eplb_process()
self.eplb_updator.forward_end()
return hidden_states
@contextmanager
def set_in_profile_run(self):
self.in_profile_run = True
try:
yield
finally:
self.in_profile_run = False
def profile_run(self) -> None:
# Trigger compilation for general shape.
with self.set_in_profile_run():
hidden_states = self._dummy_run(self.max_num_tokens,
with_prefill=True)
# MC2 will consume additional NPU memory.
# Therefore, we need to run the MC2 path once here to complete its initialization,
# allowing vLLM to correctly estimate the maximum memory required.
if self.max_num_tokens > self.mc2_tokens_capacity and \
self._select_moe_comm_method(
self.mc2_tokens_capacity,
with_prefill=True) == MoECommType.MC2:
self._dummy_run(self.mc2_tokens_capacity, with_prefill=True)
output = None
if get_pp_group().is_last_rank:
if self.is_pooling_model:
output = self._dummy_pooler_run(hidden_states)
else:
# For profile, have maximum num_reqs and that collectively have
# maximum num_tokens.
min_tokens_per_req = self.max_num_tokens // self.max_num_reqs
num_scheduled_tokens_list = [min_tokens_per_req
] * self.max_num_reqs
num_scheduled_tokens_list[
-1] += self.max_num_tokens % self.max_num_reqs
num_scheduled_tokens = np.array(num_scheduled_tokens_list,
dtype=np.int32)
logit_indices = np.cumsum(num_scheduled_tokens) - 1
# TODO: need to rum a dummy sampler for generate task
hidden_states = hidden_states[logit_indices]
output = self.model.compute_logits(hidden_states)
NPUPlatform.synchronize()
del hidden_states, output
self.encoder_cache.clear()
gc.collect()
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_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
req_num_tokens = num_tokens // num_reqs
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)
to_update = model.pooler.get_pooling_updates(task)
to_update.apply(dummy_pooling_params)
dummy_prompt_lens = torch.tensor(
num_scheduled_tokens_list,
device="cpu",
)
dummy_metadata = PoolingMetadata(
prompt_lens=dummy_prompt_lens,
prompt_token_ids=dummy_token_ids,
pooling_params=[dummy_pooling_params] * num_reqs,
)
dummy_metadata.build_pooling_cursor(num_scheduled_tokens_list,
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:
# Find the task that has the largest output for subsequent steps
output_size = dict[PoolingTask, float]()
for task in self.get_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)
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 eplb_warmup(self):
if self.dynamic_eplb and not self.is_eplb_warmuped:
self.is_eplb_warmuped = True
self.eplb_adaptor = VllmEplbAdaptor(model=self.model)
self.eplb_loader.set_adator(self.eplb_adaptor)
self.eplb_updator.set_adaptor(self.eplb_adaptor)
self.eplb_updator.warm_up_eplb()
def load_model(self) -> None:
logger.info("Starting to load model %s...", self.model_config.model)
with DeviceMemoryProfiler() as m: # noqa: SIM117
self.model = get_model(vllm_config=self.vllm_config)
if self.dynamic_eplb:
model_register(self.model, self.model_config)
if is_310p():
from vllm.model_executor.layers.linear import (
MergedColumnParallelLinear, QKVParallelLinear,
RowParallelLinear)
for module in self.model.modules():
if isinstance(module,
(MergedColumnParallelLinear,
QKVParallelLinear, RowParallelLinear)):
module.weight.data = self._convert_torch_format(
module.weight.data)
if self.drafter:
logger.info("Loading drafter model...")
self.drafter.load_model(self.model)
if self.drafter.name == SpecDcodeType.EAGLE3:
self.model.set_aux_hidden_state_layers(
self.model.get_eagle3_aux_hidden_state_layers())
if self.lora_config:
self.model = self.load_lora_model(self.model, self.vllm_config,
self.device)
logger.info("Loading model weights took %.4f GB",
m.consumed_memory / float(2**30))
# wrap the model with full graph wrapper if needed.
if self.compilation_config.cudagraph_mode.has_full_cudagraphs():
self.update_stream = torch.npu.Stream()
set_graph_params(self.compilation_config.cudagraph_capture_sizes)
self.model = ACLGraphWrapper(self.model,
self.vllm_config,
runtime_mode=CUDAGraphMode.FULL)
def _convert_torch_format(self, tensor):
if ACL_FORMAT == ACL_FORMAT_FRACTAL_NZ \
and not is_enable_nz():
return tensor
tensor = torch_npu.npu_format_cast(tensor, ACL_FORMAT)
return tensor
def initialize_kv_cache(self, kv_cache_config: KVCacheConfig) -> 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.initialize_attn_backend(kv_cache_config)
self.use_hybrid_blocks = (len(self.attn_groups) > 1)
# NOTE: Currently, we determine whether we need `num_accepted_tokens` through `MambaSpec`.
self.need_accepted_tokens = any([
isinstance(attn_group[0].kv_cache_spec, MambaSpec)
for attn_group in self.attn_groups
])
self.may_reinitialize_input_batch(kv_cache_config)
if self.ascend_config.is_deepseek_sfa:
kv_caches = self.initialize_kv_cache_tensors_deepseek_sfa(
kv_cache_config)
elif self.model_config.is_deepseek_mla:
kv_caches = self.initialize_kv_cache_tensors_deepseek_mla(
kv_cache_config)
else:
kv_caches = self.initialize_kv_cache_tensors(kv_cache_config)
if has_kv_transfer_group():
get_kv_transfer_group().register_kv_caches(kv_caches)
def _align_memory(self, tensor: torch.Tensor,
alignment: int) -> torch.Tensor:
data_ptr = tensor.data_ptr()
aligned_addr = (data_ptr + alignment - 1) // alignment * alignment
offset = (aligned_addr - data_ptr) // tensor.element_size()
return tensor[int(offset):]
def initialize_kv_cache_tensors_deepseek_sfa(
self, kv_cache_config: KVCacheConfig) -> dict[str, torch.Tensor]:
kv_cache_sizes = {}
for kv_cache_tensor in kv_cache_config.kv_cache_tensors:
assert len(kv_cache_tensor.shared_by) == 1, (
"KV cache tensor shared by multiple layers is not supported in "
"NPU.")
kv_cache_sizes[kv_cache_tensor.shared_by[0]] = kv_cache_tensor.size
kv_caches: Dict[str, torch.Tensor] = {}
for group in self._kv_cache_spec_attn_group_iterator():
kv_cache_spec = group.kv_cache_spec
attn_backend = group.backend
for layer_name in group.layer_names:
if layer_name in self.runner_only_attn_layers:
continue
tensor_size = kv_cache_sizes[layer_name]
num_blocks = tensor_size // kv_cache_spec.page_size_bytes
if self.vllm_config.additional_config.get(
"kv_cache_dtype", None) == 'int8':
kv_cache_shape = attn_backend.get_bsh_kv_cache_shape(
num_blocks, kv_cache_spec.block_size,
kv_cache_spec.num_kv_heads, kv_cache_spec.head_size)
elif hasattr(
attn_backend, "get_supported_block_size"
) and not self.model_config.is_deepseek_mla and not self.ascend_config.is_deepseek_sfa:
block_size = attn_backend.get_supported_block_size()[0]
block_size_chunk = kv_cache_spec.block_size // block_size
kv_cache_shape = attn_backend.get_kv_cache_shape(
num_blocks * block_size_chunk, block_size,
kv_cache_spec.num_kv_heads, kv_cache_spec.head_size)
else:
kv_cache_shape = self.attn_backend.get_kv_cache_shape(
num_blocks, kv_cache_spec.block_size,
kv_cache_spec.num_kv_heads, kv_cache_spec.head_size)
dtype = kv_cache_spec.dtype
alignment = 2 * 1024 * 1024
num_blocks, block_size, num_kv_heads, head_size = kv_cache_shape
rope_dim = self.model_config.hf_text_config.qk_rope_head_dim
nope_dim = head_size - rope_dim
nope_cache_shape = (num_blocks, block_size, num_kv_heads,
nope_dim)
rope_cache_shape = (num_blocks, block_size, num_kv_heads,
rope_dim)
#### k cache
# TODO(zzzzwwjj): wait transformers add these params
k_cache_shape = (num_blocks, block_size, 1, 128)
if self.vllm_config.kv_transfer_config is None:
# For no disaggregate pd scenario, allocate kv cache in normal way
rope_cache = torch.zeros(rope_cache_shape,
dtype=dtype,
device=self.device)
nope_cache = torch.zeros(nope_cache_shape,
dtype=dtype,
device=self.device)
rope_cache = self._convert_torch_format(rope_cache)
nope_cache = self._convert_torch_format(nope_cache)
#### k cache
k_cache = torch.zeros(k_cache_shape,
dtype=dtype,
device=self.device)
k_cache = self._convert_torch_format(k_cache)
else:
# In order to transfer kv cache through the reigster_memory api from llmdatadist, the memory
# address should be aligned by 2M. In most case, torch_npu can allocate 2M aligned memory, but
# we found there are also some exceptions during test, so we manual align those memory here, this part
# of code may consume 2M * 2 * elem_size memory every layer.
nope_allocate_shape = num_blocks * block_size * num_kv_heads * nope_dim
nope_allocate_shape_alignment = nope_allocate_shape + alignment
rope_allocate_shape = num_blocks * block_size * num_kv_heads * rope_dim
rope_allocate_shape_alignment = rope_allocate_shape + alignment
nope_cache = torch.zeros(nope_allocate_shape_alignment,
dtype=dtype,
device=self.device)
rope_cache = torch.zeros(rope_allocate_shape_alignment,
dtype=dtype,
device=self.device)
#### k cache
# TODO(zzzzwwjj): wait transformers add these params
k_allocate_shape = num_blocks * block_size * 1 * 128
k_allocate_shape_alignment = k_allocate_shape + alignment
k_cache = torch.zeros(k_allocate_shape_alignment,
dtype=dtype,
device=self.device)
nope_cache = self._align_memory(
nope_cache,
alignment)[:nope_allocate_shape].view(nope_cache_shape)
rope_cache = self._align_memory(
rope_cache,
alignment)[:rope_allocate_shape].view(rope_cache_shape)
k_cache = self._align_memory(
k_cache,
alignment)[:k_allocate_shape].view(k_cache_shape)
kv_caches[layer_name] = (nope_cache, rope_cache, k_cache)
bind_kv_cache(kv_caches,
self.compilation_config.static_forward_context,
self.kv_caches)
return kv_caches
def initialize_kv_cache_tensors_deepseek_mla(
self, kv_cache_config: KVCacheConfig) -> dict[str, torch.Tensor]:
kv_cache_sizes = {}
for kv_cache_tensor in kv_cache_config.kv_cache_tensors:
assert len(kv_cache_tensor.shared_by) == 1, (
"KV cache tensor shared by multiple layers is not supported in "
"NPU.")
kv_cache_sizes[kv_cache_tensor.shared_by[0]] = kv_cache_tensor.size
kv_caches: Dict[str, torch.Tensor] = {}
for group in self._kv_cache_spec_attn_group_iterator():
kv_cache_spec = group.kv_cache_spec
attn_backend = group.backend
for layer_name in group.layer_names:
if layer_name in self.runner_only_attn_layers:
continue
tensor_size = kv_cache_sizes[layer_name]
num_blocks = tensor_size // kv_cache_spec.page_size_bytes
if self.vllm_config.additional_config.get(
"kv_cache_dtype", None) == 'int8':
kv_cache_shape = attn_backend.get_bsh_kv_cache_shape(
num_blocks, kv_cache_spec.block_size,
kv_cache_spec.num_kv_heads, kv_cache_spec.head_size)
elif hasattr(attn_backend, "get_supported_block_size"
) and not self.model_config.is_deepseek_mla:
block_size = attn_backend.get_supported_block_size()[0]
block_size_chunk = kv_cache_spec.block_size // block_size
kv_cache_shape = attn_backend.get_kv_cache_shape(
num_blocks * block_size_chunk, block_size,
kv_cache_spec.num_kv_heads, kv_cache_spec.head_size)
else:
kv_cache_shape = self.attn_backend.get_kv_cache_shape(
num_blocks, kv_cache_spec.block_size,
kv_cache_spec.num_kv_heads, kv_cache_spec.head_size)
dtype = kv_cache_spec.dtype
alignment = 2 * 1024 * 1024
num_blocks, block_size, num_kv_heads, head_size = kv_cache_shape
rope_dim = self.model_config.hf_text_config.qk_rope_head_dim
nope_dim = head_size - rope_dim
nope_cache_shape = (num_blocks, block_size, num_kv_heads,
nope_dim)
rope_cache_shape = (num_blocks, block_size, num_kv_heads,
rope_dim)
if self.vllm_config.kv_transfer_config is None:
# For no disaggregate pd scenario, allocate kv cache in normal way
rope_cache = torch.zeros(rope_cache_shape,
dtype=dtype,
device=self.device)
nope_cache = torch.zeros(nope_cache_shape,
dtype=dtype,
device=self.device)
rope_cache = self._convert_torch_format(rope_cache)
nope_cache = self._convert_torch_format(nope_cache)
else:
# In order to transfer kv cache through the reigster_memory api from llmdatadist, the memory
# address should be aligned by 2M. In most case, torch_npu can allocate 2M aligned memory, but
# we found there are also some exceptions during test, so we manual align those memory here, this part
# of code may consume 2M * 2 * elem_size memory every layer.
nope_allocate_shape = num_blocks * block_size * num_kv_heads * nope_dim
nope_allocate_shape_alignment = nope_allocate_shape + alignment
rope_allocate_shape = num_blocks * block_size * num_kv_heads * rope_dim
rope_allocate_shape_alignment = rope_allocate_shape + alignment
nope_cache = torch.zeros(nope_allocate_shape_alignment,
dtype=dtype,
device=self.device)
rope_cache = torch.zeros(rope_allocate_shape_alignment,
dtype=dtype,
device=self.device)
nope_cache = self._align_memory(
nope_cache,
alignment)[:nope_allocate_shape].view(nope_cache_shape)
rope_cache = self._align_memory(
rope_cache,
alignment)[:rope_allocate_shape].view(rope_cache_shape)
kv_caches[layer_name] = (nope_cache, rope_cache)
bind_kv_cache(kv_caches,
self.compilation_config.static_forward_context,
self.kv_caches)
return kv_caches
def initialize_kv_cache_tensors(
self, kv_cache_config: KVCacheConfig) -> dict[str, torch.Tensor]:
"""
Initialize the memory buffer for KV cache.
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.
"""
# init kv cache tensors
kv_cache_raw_tensors: dict[str, Union[torch.Tensor,
Optional[torch.Tensor]]] = {}
# llmdatadist need the addr of cache tensor be aligned with 2M
alignment = 2 * 1024 * 1024
for kv_cache_tensor in kv_cache_config.kv_cache_tensors:
# TODO: REFACTOR ME to sharing hybrid cache
for idx in range(len(kv_cache_tensor.shared_by)):
layer_name = kv_cache_tensor.shared_by[idx]
if "linear_attn" in layer_name:
# for mamba linear attention
for layer_name_inner in kv_cache_tensor.shared_by:
if ("attn" in layer_name_inner and "linear_attn" not in layer_name_inner) or \
layer_name_inner in kv_cache_raw_tensors.keys():
continue
if self.vllm_config.kv_transfer_config is None:
tensor = torch.zeros(kv_cache_tensor.size,
dtype=torch.int8,
device=self.device)
else:
cache_size_aligned = kv_cache_tensor.size + alignment
tensor = torch.zeros(cache_size_aligned,
dtype=torch.int8,
device=self.device)
tensor = self._align_memory(
tensor, alignment)[:kv_cache_tensor.size]
kv_cache_raw_tensors[layer_name_inner] = tensor
elif "attn" in layer_name:
# for other attentions, e.g., self_attn, sliding window attn
if self.vllm_config.kv_transfer_config is None:
k_tensor = torch.zeros(kv_cache_tensor.size // 2,
dtype=torch.int8,
device=self.device)
v_tensor = torch.zeros(kv_cache_tensor.size // 2,
dtype=torch.int8,
device=self.device)
else:
cache_size = kv_cache_tensor.size // 2
cache_size_aligned = kv_cache_tensor.size // 2 + alignment
k_tensor = torch.zeros(cache_size_aligned,
dtype=torch.int8,
device=self.device)
v_tensor = torch.zeros(cache_size_aligned,
dtype=torch.int8,
device=self.device)
k_tensor = self._align_memory(k_tensor,
alignment)[:cache_size]
v_tensor = self._align_memory(v_tensor,
alignment)[:cache_size]
kv_cache_raw_tensors[layer_name] = (k_tensor, v_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"
kv_caches: Dict[str, torch.Tensor] = {}
for group in self._kv_cache_spec_attn_group_iterator():
kv_cache_spec = group.kv_cache_spec
attn_backend = group.backend
for layer_name in group.layer_names:
if layer_name in self.runner_only_attn_layers:
continue
# TODO: remove this after the OOM issue is located and fixed, otherwise, some model may
# encounter OOM issue
if isinstance(kv_cache_spec, FullAttentionSpec):
raw_k_tensor, raw_v_tensor = kv_cache_raw_tensors[ # type: ignore
layer_name]
assert raw_k_tensor is not None
assert raw_v_tensor is not None
assert (raw_k_tensor.numel() + raw_v_tensor.numel()
) % kv_cache_spec.page_size_bytes == 0
num_blocks = (raw_k_tensor.numel() + raw_v_tensor.numel()
) // kv_cache_spec.page_size_bytes
# `num_blocks` is the number of blocks the model runner can use.
# `kv_cache_config.num_blocks` is the number of blocks that
# KVCacheManager may allocate.
# Since different GPUs may have different number of layers and
# different memory capacities, `num_blocks` can be different on
# different GPUs, and `kv_cache_config.num_blocks` is set to
# the min of all `num_blocks`. Verify it here.
assert num_blocks >= kv_cache_config.num_blocks
if self.vllm_config.additional_config.get(
"kv_cache_dtype", None) == 'int8':
kv_cache_shape = attn_backend.get_bsh_kv_cache_shape(
num_blocks, kv_cache_spec.block_size,
kv_cache_spec.num_kv_heads,
kv_cache_spec.head_size)
elif hasattr(attn_backend, "get_supported_block_size"
) and self.use_hybrid_blocks:
block_size = attn_backend.get_supported_block_size()[0]
block_size_chunk = kv_cache_spec.block_size // block_size
kv_cache_shape = attn_backend.get_kv_cache_shape(
num_blocks * block_size_chunk, block_size,
kv_cache_spec.num_kv_heads,
kv_cache_spec.head_size)
else:
kv_cache_shape = self.attn_backend.get_kv_cache_shape(
num_blocks, kv_cache_spec.block_size,
kv_cache_spec.num_kv_heads,
kv_cache_spec.head_size)
dtype = kv_cache_spec.dtype
k_cache = raw_k_tensor.view(dtype).view(kv_cache_shape[1:])
k_cache = self._convert_torch_format(k_cache)
v_cache = raw_v_tensor.view(dtype).view(kv_cache_shape[1:])
v_cache = self._convert_torch_format(v_cache)
kv_caches[layer_name] = (k_cache, v_cache)
elif isinstance(kv_cache_spec, MambaSpec):
raw_tensor = kv_cache_raw_tensors[layer_name]
assert raw_tensor is not None
assert raw_tensor.numel(
) % kv_cache_spec.page_size_bytes == 0
num_blocks = raw_tensor.numel(
) // kv_cache_spec.page_size_bytes
# `num_blocks` is the number of blocks the model runner can use.
# `kv_cache_config.num_blocks` is the number of blocks that
# KVCacheManager may allocate.
# Since different GPUs may have different number of layers and
# different memory capacities, `num_blocks` can be different on
# different GPUs, and `kv_cache_config.num_blocks` is set to
# the min of all `num_blocks`. Verify it here.
assert num_blocks >= kv_cache_config.num_blocks
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 ValueError("Unknown KV cache spec type.")
bind_kv_cache(kv_caches,
self.compilation_config.static_forward_context,
self.kv_caches)
return kv_caches
def may_reinitialize_input_batch(self,
kv_cache_config: KVCacheConfig) -> None:
"""
Re-initialize the input batch if the block sizes are different from
`[self.cache_config.block_size]`. This usually happens when there
are multiple KV cache groups.
Args:
kv_cache_config: The KV cache configuration.
"""
block_sizes = [
kv_cache_group.kv_cache_spec.block_size
for kv_cache_group in kv_cache_config.kv_cache_groups
]
# Generate kernel_block_sizes that matches each block_size
# For attention backends that support virtual block splitting,
# use the supported block sizes from the backend
# For other backends (like Mamba), use [0] (no splitting)
kernel_block_sizes = []
for kv_cache_group_id, kv_cache_group in enumerate(
kv_cache_config.kv_cache_groups):
if isinstance(kv_cache_group.kv_cache_spec, AttentionSpec):
# This is an attention backend that supports virtual
# block splitting. Get the supported block sizes from
# the backend.
try:
attn_groups = self.attn_groups[kv_cache_group_id]
except IndexError:
attn_groups = None
if attn_groups and self.use_hybrid_blocks:
# Use the backend's supported block size list
backend = attn_groups[0].backend
supported_sizes = backend.get_supported_block_size()
# If no specific sizes supported, use cache config
# block_size
kernel_block_size_list = (supported_sizes
if supported_sizes else
[self.cache_config.block_size])
else:
# Fallback to cache config block_size if no backend found
kernel_block_size_list = [self.cache_config.block_size]
kernel_block_sizes.append(kernel_block_size_list)
else:
# This is likely Mamba or other non-attention cache,
# no splitting.
# NOTE: set kernel_block_sizes to 0 to disable slotmapping computation
# of mamba block. In this case, BlockTable.block_size will never equal
# to kernel_block_sizes[0]
kernel_block_sizes.append([0])
if kernel_block_sizes != [[self.cache_config.block_size]]:
assert self.cache_config.cpu_offload_gb == 0, (
"Cannot re-initialize the input batch when CPU weight "
"offloading is enabled. See https://github.com/vllm-project/vllm/pull/18298 " # noqa: E501
"for more details.")
self.input_batch = InputBatch(
max_num_reqs=self.max_num_reqs,
max_model_len=self.model_config.max_model_len,
max_num_batched_tokens=self.max_num_tokens,
device=self.device,
pin_memory=self.pin_memory,
vocab_size=self.model_config.get_vocab_size(),
block_sizes=block_sizes,
is_spec_decode=bool(self.vllm_config.speculative_config),
logitsprocs=self.input_batch.logitsprocs,
is_pooling_model=self.is_pooling_model,
num_speculative_tokens=(
self.vllm_config.speculative_config.num_speculative_tokens
if self.vllm_config.speculative_config else 0),
kernel_block_sizes=kernel_block_sizes,
)
def initialize_attn_backend(self, kv_cache_config: KVCacheConfig) -> None:
"""
Initialize the attention backends and attention metadata builders.
"""
assert len(self.attn_groups) == 0, \
"Attention backends are already initialized"
class AttentionGroupKey(NamedTuple):
attn_backend: type[AttentionBackend]
kv_cache_spec: KVCacheSpec
def get_attn_backends_for_group(
kv_cache_group_spec: KVCacheGroupSpec,
) -> dict[AttentionGroupKey, list[str]]:
layers = get_layers_from_vllm_config(
self.vllm_config, AttentionLayerBase,
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()
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]
key = (full_cls_name, layer_kv_cache_spec)
attn_backends[key] = AttentionGroupKey(attn_backend,
layer_kv_cache_spec)
attn_backend_layers[key].append(layer_name)
return {
attn_backends[k]: v
for k, v in attn_backend_layers.items()
}
def create_attn_groups(
attn_backends_map: dict[AttentionBackend, list[str]],
) -> list[AttentionGroup]:
attn_groups: list[AttentionGroup] = []
for (attn_backend,
kv_cache_spec), layer_names in attn_backends_map.items():
attn_metadata_builders = []
attn_metadata_builders.append(attn_backend.get_builder_cls()(
kv_cache_spec,
layer_names,
self.vllm_config,
self.device,
))
attn_group = AttentionGroup(attn_backend,
attn_metadata_builders,
layer_names, kv_cache_spec)
attn_groups.append(attn_group)
return attn_groups
for kv_cache_group_spec in kv_cache_config.kv_cache_groups:
attn_backends = get_attn_backends_for_group( # type: ignore
kv_cache_group_spec)
self.attn_groups.append(create_attn_groups(attn_backends))
# Calculate reorder batch threshold (if needed)
self.calculate_reorder_batch_threshold()
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 calculate_reorder_batch_threshold(self) -> None:
"""
Check that if any backends reorder batches; that the reordering
is compatible (e.g., decode threshold is the same)
"""
for group in self._attn_group_iterator():
attn_metadata_builder_i = group.get_metadata_builder()
if hasattr(attn_metadata_builder_i, "reorder_batch_threshold"):
# check that if any backends reorder batches; that the reordering
# is compatible (e.g., decode threshold is the same)
reorder_batch_threshold_i = (
attn_metadata_builder_i.reorder_batch_threshold)
if reorder_batch_threshold_i is not None:
if self.reorder_batch_threshold is not None:
if reorder_batch_threshold_i != \
self.reorder_batch_threshold:
raise ValueError(
f"Attention backend reorders decodes with "
f"threshold {reorder_batch_threshold_i} but other "
f"backend uses threshold "
f"{self.reorder_batch_threshold}")
else:
self.reorder_batch_threshold = reorder_batch_threshold_i
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.
"""
block_size = self.vllm_config.cache_config.block_size
use_mla = self.vllm_config.model_config.use_mla
use_sfa = self.ascend_config.use_sfa
kv_cache_spec: dict[str, KVCacheSpec] = {}
attn_layers = get_layers_from_vllm_config(self.vllm_config, Attention)
for layer_name, attn_module in attn_layers.items():
if (kv_tgt_layer :=
attn_module.kv_sharing_target_layer_name) is not None:
# 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
if isinstance(attn_module, AscendMultiHeadLatentAttention):
continue
# TODO: Support other attention modules, e.g., cross-attention
# TODO(lucas): move the attention specs into the model layers like
# the attention backends
if attn_module.attn_type == AttentionType.DECODER:
if use_mla and not use_sfa:
kv_cache_spec[layer_name] = MLAAttentionSpec(
block_size=block_size,
num_kv_heads=attn_module.num_kv_heads,
head_size=attn_module.head_size,
dtype=self.kv_cache_dtype,
cache_dtype_str=self.cache_config.cache_dtype)
else:
# TODO(cmq): This is a hack way to fix deepseek kvcache when
# using DSA. Fix the spec in vLLM is a finnal way.
kv_cache_spec[layer_name] = FullAttentionSpec(
block_size=block_size,
num_kv_heads=attn_module.num_kv_heads,
head_size=attn_module.head_size,
dtype=self.kv_cache_dtype)
elif attn_module.attn_type in (AttentionType.ENCODER,
AttentionType.ENCODER_ONLY):
# encoder-only attention does not need KV cache.
continue
elif attn_module.attn_type == AttentionType.ENCODER_DECODER:
raise NotImplementedError
else:
raise ValueError(
f"Unknown attention type: {attn_module.attn_type}")
mamba_layers = get_layers_from_vllm_config(self.vllm_config, MambaBase)
if len(mamba_layers) > 0:
if (self.vllm_config.speculative_config is not None
and self.vllm_config.model_config.hf_config.model_type
not in ["qwen3_next"]):
raise NotImplementedError(
"Mamba with speculative decoding is not supported yet.")
if self.vllm_config.cache_config.enable_prefix_caching:
raise NotImplementedError(
"Prefix caching is not supported for Mamba yet.")
max_model_len = self.vllm_config.model_config.max_model_len
page_size_padded = (
self.vllm_config.cache_config.mamba_page_size_padded)
# Set block_size to max_model_len, so that mamba model will always
# have only one block in the KV cache.
for layer_name, mamba_module in mamba_layers.items():
kv_cache_spec[layer_name] = MambaSpec(
shapes=mamba_module.get_state_shape(),
dtypes=mamba_module.get_state_dtype(),
block_size=max_model_len,
page_size_padded=page_size_padded,
mamba_type=mamba_module.mamba_type,
num_speculative_blocks=(
self.speculative_config.num_speculative_tokens
if self.speculative_config else 0),
)
return kv_cache_spec
def initialize_aclgraph_capture(self) -> None:
min_ag_support = AttentionCGSupport.ALWAYS
min_ag_builder_name = None
for attn_group in self._attn_group_iterator():
builder = attn_group.get_metadata_builder()
if builder.aclgraph_support.value < min_ag_support.value:
min_ag_support = builder.aclgraph_support
min_ag_builder_name = builder.__class__.__name__
# This is an imitation of compilation_config.splitting_ops_contain_attention()
splitting_ops_contain_attention = (
self.compilation_config.splitting_ops is not None
and all(op in self.compilation_config.splitting_ops for op in [
"vllm.unified_ascend_attention_with_output",
"vllm.mla_forward",
]))
# Flexible resolve the aclgraph mode
aclgraph_mode = self.compilation_config.cudagraph_mode
# check graph for mixed batch is supported
if aclgraph_mode.mixed_mode() == CUDAGraphMode.FULL \
and min_ag_support != AttentionCGSupport.ALWAYS:
msg = (f"ACLGraphMode.{aclgraph_mode.name} is not supported "
f"with {min_ag_builder_name} backend (support: "
f"{min_ag_support})")
if min_ag_support == AttentionCGSupport.NEVER:
# if not supported any full graphs, just raise it.
msg += "; please try cudagraph_mode=PIECEWISE, and "\
"make sure compilation level is piecewise"
raise ValueError(msg)
# attempt to resolve the full graph related mode
if splitting_ops_contain_attention:
msg += "; setting cudagraph_mode=FULL_AND_PIECEWISE"
aclgraph_mode = self.compilation_config.cudagraph_mode = (
CUDAGraphMode.FULL_AND_PIECEWISE)
else:
msg += "; setting cudagraph_mode=FULL_DECODE_ONLY"
aclgraph_mode = self.compilation_config.cudagraph_mode = (
CUDAGraphMode.FULL_DECODE_ONLY)
logger.warning(msg)
# check that if spec-decode + decode full-graphs is supported
if (aclgraph_mode.decode_mode() == CUDAGraphMode.FULL
and self.uniform_decode_query_len > 1 and min_ag_support.value
< AttentionCGSupport.UNIFORM_BATCH.value):
msg = (f"CUDAGraphMode.{aclgraph_mode.name} is not supported"
f" with spec-decode for attention backend "
f"{min_ag_builder_name} (support: {min_ag_support})")
if splitting_ops_contain_attention:
msg += "; setting cudagraph_mode=PIECEWISE"
aclgraph_mode = self.compilation_config.cudagraph_mode = \
CUDAGraphMode.PIECEWISE
else:
msg += "; setting cudagraph_mode=NONE"
aclgraph_mode = self.compilation_config.cudagraph_mode = \
CUDAGraphMode.NONE
logger.warning(msg)
# double check that we can support full graph if they are requested
# even after automatic downgrades
if aclgraph_mode.has_full_cudagraphs() \
and min_ag_support == AttentionCGSupport.NEVER:
raise ValueError(f"CUDAGraphMode.{aclgraph_mode.name} is not "
f"supported with {min_ag_builder_name} backend ("
f"support:{min_ag_support}) "
"; please try cudagraph_mode=PIECEWISE, "
"and make sure compilation level is piecewise")
self.aclgraph_dispatcher.initialize_cudagraph_keys(
self.compilation_config.cudagraph_mode,
self.uniform_decode_query_len)
def _capture_aclgraphs(self, compilation_cases: list[int],
aclgraph_runtime_mode: CUDAGraphMode,
uniform_decode: bool):
assert aclgraph_runtime_mode != CUDAGraphMode.NONE and \
aclgraph_runtime_mode in [CUDAGraphMode.FULL,
CUDAGraphMode.PIECEWISE]
# Only rank 0 should print progress bar during capture
if is_global_first_rank():
logger.info(
"Starting to capture ACL graphs for cases: %s, "
"mode: %s, uniform_decode: %s", compilation_cases,
aclgraph_runtime_mode.name, uniform_decode)
compilation_cases = tqdm(
compilation_cases,
disable=not self.load_config.use_tqdm_on_load,
desc="Capturing ACL graphs ({}, {})".format(
"decode" if uniform_decode else "mixed prefill-decode",
aclgraph_runtime_mode.name))
# We skip EPLB here since we don't want to record dummy metrics
for num_tokens in compilation_cases:
for _ in range(self.compilation_config.cudagraph_num_of_warmups):
# Use CUDAGraphRuntimeStyle.NONE (default) for warmup.
# But be careful, warm up with `NONE`is orthogonal to
# if we want to warm up attention or not. This is
# different from the case where `FULL` implies capture
# attention while `PIECEWISE` implies no attention.
force_attention = (aclgraph_runtime_mode == CUDAGraphMode.FULL)
self._dummy_run(num_tokens,
aclgraph_runtime_mode=CUDAGraphMode.NONE,
force_attention=force_attention,
uniform_decode=uniform_decode)
self._dummy_run(num_tokens,
aclgraph_runtime_mode=aclgraph_runtime_mode,
force_attention=force_attention,
uniform_decode=uniform_decode)
def _capture_model(self):
if not self.use_aclgraph:
logger.warning(
"Skipping ACL graph capture. To turn on ACL graph capture, "
"ensure `aclraph_mode` was not manually set to `NONE`")
return
else:
self.initialize_aclgraph_capture()
set_cudagraph_capturing_enabled(True)
# Trigger ACL 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.
with graph_capture(device=self.device):
aclgraph_mode = self.compilation_config.cudagraph_mode
if aclgraph_mode.mixed_mode() != CUDAGraphMode.NONE:
aclgraph_runtime_mode = aclgraph_mode.mixed_mode()
compilation_cases = list(reversed(self.aclgraph_batch_sizes))
try:
self._capture_aclgraphs(
compilation_cases,
aclgraph_runtime_mode=aclgraph_runtime_mode,
uniform_decode=False)
except Exception as e:
error_msg = str(e)
error_code = '0x7020023'
pattern = r'retCode=([^,\s\.]+)'
match = re.search(pattern, error_msg)
if match:
retCode = match.group(1)
# Determine whether the error message is caused by stream capture failure.
if match and retCode == error_code:
logger.error(
f"ACLgraph sizes capture fail: {type(e).__name__}:\n"
"ACLgraph has insufficient available streams to capture the configured number of sizes. "
"Please verify both the availability of adequate streams and the appropriateness of the configured size count.\n\n"
"Recommended solutions:\n"
"1. Manually configure the compilation_config parameter "
"with a reduced set of sizes: '{\"cudagraph_capture_sizes\":[size1, size2, size3, ...]}'.\n"
"2. Utilize ACLgraph's full graph mode as an alternative to the piece-wise approach.\n\n"
f"{str(e)}")
raise
if aclgraph_mode.decode_mode() == CUDAGraphMode.FULL and \
aclgraph_mode.separate_routine():
max_num_tokens = self.scheduler_config.max_num_seqs * \
self.uniform_decode_query_len
decode_cudagraph_batch_sizes = [
x for x in self.aclgraph_batch_sizes if x <= max_num_tokens
and x >= self.uniform_decode_query_len
]
compilation_cases_decode = list(
reversed(decode_cudagraph_batch_sizes))
self._capture_aclgraphs(
compilation_cases=compilation_cases_decode,
aclgraph_runtime_mode=CUDAGraphMode.FULL,
uniform_decode=True)
# Disable aclgraph capturing globally, so any unexpected aclgraph
# capturing will be detected and raise an error after here.
# Note: We don't put it into graph_capture context manager because
# we may doing lazy capturing in future that still allows capturing
# after here.
set_cudagraph_capturing_enabled(False)
def capture_model(self) -> None:
compilation_counter.num_gpu_runner_capture_triggers += 1
start_time = time.perf_counter()
start_free_npu_memory = torch.npu.mem_get_info()[0]
self._capture_model()
end_time = time.perf_counter()
end_free_npu_memory = torch.npu.mem_get_info()[0]
elapsed_time = end_time - start_time
npu_graph_size = start_free_npu_memory - end_free_npu_memory
# This usually takes 5~20 seconds.
logger.info("Graph capturing finished in %.0f secs, took %.2f GiB",
elapsed_time, npu_graph_size / (1 << 30))
def _get_prompt_logprobs_dict(
self,
hidden_states: torch.Tensor,
scheduler_output: "SchedulerOutput",
) -> dict[str, Optional[LogprobsTensors]]:
num_prompt_logprobs_dict = self.input_batch.num_prompt_logprobs
if not num_prompt_logprobs_dict:
return {}
in_progress_dict = self.input_batch.in_progress_prompt_logprobs_cpu
prompt_logprobs_dict: dict[str, Optional[LogprobsTensors]] = {}
# 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 = scheduler_output.num_scheduled_tokens[req_id]
# Get metadata for this request.
request = self.requests[req_id]
num_prompt_tokens = len(request.prompt_token_ids)
prompt_token_ids = torch.tensor(request.prompt_token_ids).to(
self.device, non_blocking=True)
# Set up target LogprobsTensors object.
logprobs_tensors = in_progress_dict.get(req_id)
if not logprobs_tensors:
# 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)
in_progress_dict[req_id] = 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 NPU->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]
del in_progress_dict[req_id]
# Must synchronize the non-blocking NPU->CPU transfers.
if prompt_logprobs_dict:
torch.npu.synchronize()
return prompt_logprobs_dict
def get_supported_pooling_tasks(self):
model = self.get_model()
if not is_pooling_model(model):
return []
return list(model.pooler.get_supported_tasks())
def _build_drafter_prepare_inputs_torchair_param(self):
return False