class GPUModelRunner(LoRAModelRunnerMixin, KVConnectorModelRunnerMixin):
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.lora_config = vllm_config.lora_config
self.load_config = vllm_config.load_config
self.parallel_config = vllm_config.parallel_config
self.scheduler_config = vllm_config.scheduler_config
self.speculative_config = vllm_config.speculative_config
self.observability_config = vllm_config.observability_config
from vllm.model_executor.models.utils import set_cpu_offload_max_bytes
set_cpu_offload_max_bytes(int(self.cache_config.cpu_offload_gb * 1024**3))
model_config = self.model_config
cache_config = self.cache_config
scheduler_config = self.scheduler_config
parallel_config = self.parallel_config
self.device = device
self.pin_memory = is_pin_memory_available()
self.dtype = self.model_config.dtype
self.kv_cache_dtype = kv_cache_dtype_str_to_dtype(
cache_config.cache_dtype, self.model_config
)
self.is_pooling_model = model_config.runner_type == "pooling"
self.enable_prompt_embeds = model_config.enable_prompt_embeds
self.is_multimodal_raw_input_only_model = (
model_config.is_multimodal_raw_input_only_model
)
# This will be overridden in load_model()
self.is_multimodal_pruning_enabled = False
self.max_model_len = model_config.max_model_len
self.dcp_world_size = self.parallel_config.decode_context_parallel_size
self.max_num_tokens = scheduler_config.max_num_batched_tokens
self.max_num_reqs = scheduler_config.max_num_seqs
# Broadcast PP output for external_launcher (torchrun)
# to make sure we are synced across pp ranks
# TODO: Support overlapping mirco-batches
# https://github.com/vllm-project/vllm/issues/18019
self.broadcast_pp_output = (
self.parallel_config.distributed_executor_backend == "external_launcher"
and len(get_pp_group().ranks) > 0
)
# Model-related.
self.num_query_heads = model_config.get_num_attention_heads(parallel_config)
self.hidden_size = model_config.get_hidden_size()
self.attention_chunk_size = model_config.attention_chunk_size
# Only relevant for models using ALiBi (e.g, MPT)
self.use_alibi = check_use_alibi(model_config)
self.cascade_attn_enabled = not self.model_config.disable_cascade_attn
# Multi-modal data support
self.mm_registry = MULTIMODAL_REGISTRY
self.uses_mrope = model_config.uses_mrope
self.supports_mm_inputs = self.mm_registry.supports_multimodal_inputs(
model_config
)
if self.model_config.is_encoder_decoder:
# Maximum length of the encoder input, only for encoder-decoder
# models.
self.max_encoder_len = scheduler_config.max_num_encoder_input_tokens
else:
self.max_encoder_len = 0
# Sampler
self.sampler = Sampler(logprobs_mode=self.model_config.logprobs_mode)
self.eplb_state: EplbState | None = None
"""
State of the expert parallelism load balancer.
Will be lazily initialized when the model is loaded.
"""
# Lazy initializations
# self.model: nn.Module # Set after load_model
# Initialize in initialize_kv_cache
self.kv_caches: list[torch.Tensor] = []
# indexes: [kv_cache_group_id][attn_group]
self.attn_groups: list[list[AttentionGroup]] = []
# self.kv_cache_config: KVCacheConfig
# mm_hash -> encoder_output
self.encoder_cache: dict[str, torch.Tensor] = {}
self.use_aux_hidden_state_outputs = False
# Set up speculative decoding.
# NOTE(Jiayi): currently we put the entire draft model on
# the last PP rank. This is not ideal if there are many
# layers in the draft model.
if self.speculative_config and get_pp_group().is_last_rank:
if self.speculative_config.method == "ngram":
self.drafter = NgramProposer(self.vllm_config)
elif self.speculative_config.use_eagle():
self.drafter = EagleProposer(self.vllm_config, self.device, self) # type: ignore
if self.speculative_config.method == "eagle3":
self.use_aux_hidden_state_outputs = True
elif self.speculative_config.method == "medusa":
self.drafter = MedusaProposer(
vllm_config=self.vllm_config, device=self.device
) # type: ignore
else:
raise ValueError(
"Unknown speculative decoding method: "
f"{self.speculative_config.method}"
)
self.rejection_sampler = RejectionSampler(self.sampler)
# Request states.
self.requests: dict[str, CachedRequestState] = {}
self.comm_stream = torch.cuda.Stream()
# 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.
custom_logitsprocs = model_config.logits_processors
self.input_batch = InputBatch(
max_num_reqs=self.max_num_reqs,
# We need to use the encoder length for encoder-decoer
# because of KV cache for cross-attention.
max_model_len=max(self.max_model_len, self.max_encoder_len),
max_num_batched_tokens=self.max_num_tokens,
device=self.device,
pin_memory=self.pin_memory,
vocab_size=self.model_config.get_vocab_size(),
block_sizes=[self.cache_config.block_size],
kernel_block_sizes=[self.cache_config.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,
custom_logitsprocs,
),
# We currently don't know whether a particular custom logits processor
# uses output token ids so we set this conservatively.
logitsprocs_need_output_token_ids=bool(custom_logitsprocs),
is_pooling_model=self.is_pooling_model,
)
self.use_async_scheduling = self.scheduler_config.async_scheduling
# Separate cuda stream for overlapping transfer of sampled token ids from
# GPU to CPU when async scheduling is enabled.
self.async_output_copy_stream: torch.cuda.Stream | None = None
# cuda event to synchronize use of reused CPU tensors between steps
# when async scheduling is enabled.
self.prepare_inputs_event: torch.cuda.Event | None = None
if self.use_async_scheduling:
self.async_output_copy_stream = torch.cuda.Stream()
self.prepare_inputs_event = torch.cuda.Event()
# TODO(woosuk): Provide an option to tune the max cudagraph batch size.
# The convention is different.
# self.cudagraph_batch_sizes sorts in ascending order.
# The batch sizes in the config are in descending order.
if (
self.compilation_config.cudagraph_capture_sizes
and self.compilation_config.cudagraph_mode != CUDAGraphMode.NONE
):
self.cudagraph_batch_sizes = list(
reversed(self.compilation_config.cudagraph_capture_sizes)
)
# Cache the device properties.
self._init_device_properties()
# Persistent buffers for CUDA graphs.
self.input_ids = self._make_buffer(self.max_num_tokens, dtype=torch.int32)
self.positions = self._make_buffer(self.max_num_tokens, dtype=torch.int64)
self.query_start_loc = self._make_buffer(
self.max_num_reqs + 1, dtype=torch.int32
)
self.seq_lens = self._make_buffer(self.max_num_reqs, dtype=torch.int32)
if self.dcp_world_size > 1:
self.dcp_local_seq_lens = self._make_buffer(
self.max_num_reqs, dtype=torch.int32
)
# Because inputs_embeds may be bfloat16 and we don't need a numpy
# version of this tensor, avoid a RuntimeError by not creating a
# numpy buffer.
self.inputs_embeds = self._make_buffer(
self.max_num_tokens, self.hidden_size, dtype=self.dtype, numpy=False
)
self.is_token_ids = self._make_buffer(self.max_num_tokens, dtype=torch.bool)
self.discard_request_indices = self._make_buffer(
self.max_num_reqs, dtype=torch.int64
)
self.num_discarded_requests = 0
self.num_decode_draft_tokens = self._make_buffer(
self.max_num_reqs, dtype=torch.int32
)
self.num_accepted_tokens = self._make_buffer(
self.max_num_reqs, dtype=torch.int64
)
# Only relevant for multimodal models
if self.supports_mm_inputs:
self.is_mm_embed = self._make_buffer(self.max_num_tokens, dtype=torch.bool)
# Only relevant for models using M-RoPE (e.g, Qwen2-VL)
if self.uses_mrope:
# NOTE: `mrope_positions` is implemented with one additional dummy
# position on purpose to make it non-contiguous so that it can work
# with torch compile.
# See detailed explanation in https://github.com/vllm-project/vllm/pull/12128#discussion_r1926431923
# NOTE: When M-RoPE is enabled, position ids are 3D regardless of
# the modality of inputs. For text-only inputs, each dimension has
# identical position IDs, making M-RoPE functionally equivalent to
# 1D-RoPE.
# See page 5 of https://arxiv.org/abs/2409.12191
self.mrope_positions = self._make_buffer(
(3, self.max_num_tokens + 1), dtype=torch.int64
)
# None in the first PP rank. The rest are set after load_model.
self.intermediate_tensors: IntermediateTensors | None = None
# OPTIMIZATION: Cache the tensors rather than creating them every step.
# Keep in int64 to avoid overflow with long context
self.arange_np = np.arange(
max(self.max_num_reqs + 1, self.max_model_len, self.max_num_tokens),
dtype=np.int64,
)
# Layer pairings for cross-layer KV sharing.
# If an Attention layer `layer_name` is in the keys of this dict, it
# means this layer will perform attention using the keys and values
# from the KV cache of `shared_kv_cache_layers[layer_name]`.
self.shared_kv_cache_layers: dict[str, str] = {}
self.kv_sharing_fast_prefill_eligible_layers: set[str] = set()
self.kv_sharing_fast_prefill_logits_indices = None
if self.cache_config.kv_sharing_fast_prefill:
self.kv_sharing_fast_prefill_logits_indices = torch.zeros(
self.max_num_tokens, dtype=torch.int32, device=self.device
)
self.uniform_decode_query_len = (
1
if not self.speculative_config
else 1 + self.speculative_config.num_speculative_tokens
)
# Cudagraph dispatcher for runtime cudagraph dispatching.
self.cudagraph_dispatcher = CudagraphDispatcher(self.vllm_config)
self.mm_budget = (
MultiModalBudget(
self.model_config,
self.scheduler_config,
self.mm_registry,
)
if self.supports_mm_inputs
else None
)
self.reorder_batch_threshold: int | None = None
# Attention layers that are only in the KVCacheConfig of the runner
# (e.g., KV sharing, encoder-only attention), but not in the
# KVCacheConfig of the scheduler.
self.runner_only_attn_layers: set[str] = set()
# Cached outputs.
self._draft_token_ids: list[list[int]] | torch.Tensor | None = None
self.transfer_event = torch.cuda.Event()
self.sampled_token_ids_pinned_cpu = torch.empty(
(self.max_model_len, 1),
dtype=torch.int64,
device="cpu",
pin_memory=self.pin_memory,
)
def reset_mm_cache(self) -> None:
if self.mm_budget:
self.mm_budget.reset_cache()
def _get_positions(self, num_tokens: Any):
if isinstance(num_tokens, int):
if self.uses_mrope:
return self.mrope_positions.gpu[:, :num_tokens]
return self.positions.gpu[:num_tokens]
else:
if self.uses_mrope:
return self.mrope_positions.gpu[:, num_tokens]
return self.positions.gpu[num_tokens]
def _make_buffer(
self, *size: int | torch.SymInt, dtype: torch.dtype, numpy: bool = True
) -> CpuGpuBuffer:
return CpuGpuBuffer(
*size,
dtype=dtype,
device=self.device,
pin_memory=self.pin_memory,
with_numpy=numpy,
)
def _init_model_kwargs(self, num_tokens: int):
model_kwargs = dict[str, Any]()
if not self.is_pooling_model:
return model_kwargs
num_reqs = self.input_batch.num_reqs
pooling_params = self.input_batch.get_pooling_params()
token_type_id_requests = dict[int, Any]()
for i, param in enumerate(pooling_params):
if (
param.extra_kwargs is not None
and (token_types := param.extra_kwargs.get("compressed_token_type_ids"))
is not None
):
token_type_id_requests[i] = token_types
if len(token_type_id_requests) == 0:
return model_kwargs
seq_lens = self.seq_lens.gpu[:num_reqs]
token_type_ids = []
for i in range(num_reqs):
pos = token_type_id_requests.get(i, seq_lens[i])
ids = (torch.arange(seq_lens[i]) >= pos).int()
token_type_ids.append(ids)
model_kwargs["token_type_ids"] = torch.concat(token_type_ids).to(
device=self.device
)
return model_kwargs
def _may_reorder_batch(self, scheduler_output: "SchedulerOutput") -> None:
"""
Update the order of requests in the batch based on the attention
backend's needs. For example, some attention backends (namely MLA) may
want to separate requests based on if the attention computation will be
compute-bound or memory-bound.
Args:
scheduler_output: The scheduler output.
"""
# Attention free models have zero kv_cache_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:
# NOTE(lucas): currently no backend supports the custom masking
# required for DCP with q_len > 1, so we assert here. Remove this
# assert once the custom mask is support is added to FA3.
if (
self.dcp_world_size > 1
and envs.VLLM_ATTENTION_BACKEND != "FLASH_ATTN_MLA"
):
assert self.reorder_batch_threshold == 1, (
"DCP not support reorder_batch_threshold > 1 now."
)
reorder_batch_to_split_decodes_and_prefills(
self.input_batch,
scheduler_output,
decode_threshold=self.reorder_batch_threshold,
)
# Note: used for model runner override.
def _init_device_properties(self) -> None:
"""Initialize attributes from torch.cuda.get_device_properties"""
self.device_properties = torch.cuda.get_device_properties(self.device)
self.num_sms = self.device_properties.multi_processor_count
# Note: used for model runner override.
def _sync_device(self) -> None:
torch.cuda.synchronize()
def _update_states(self, scheduler_output: "SchedulerOutput") -> None:
"""Update the cached states and the persistent batch with the scheduler
output.
The updated states are used by the `_prepare_inputs` function to create
the input GPU tensors for the model.
The SamplingMetadata is updated and copied to the GPU if there is a
new/resumed/paused/finished request in the batch.
"""
# Remove finished requests from the cached states.
for req_id in scheduler_output.finished_req_ids:
self.requests.pop(req_id, None)
# 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)
# Free the cached encoder outputs.
for mm_hash in scheduler_output.free_encoder_mm_hashes:
self.encoder_cache.pop(mm_hash, None)
# Remove the unscheduled requests from the persistent batch.
# NOTE(woosuk): The unscheduled requests are either preempted requests
# or running requests that are not scheduled in this step. We remove
# them from the persistent batch but keep their cached states since
# they will be scheduled again sometime in the future.
scheduled_req_ids = scheduler_output.num_scheduled_tokens.keys()
cached_req_ids = self.input_batch.req_id_to_index.keys()
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)
reqs_to_add: list[CachedRequestState] = []
# 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 self.is_pooling_model:
assert pooling_params is not None
task = pooling_params.task
assert task is not None, "You did not set `task` in the API"
model = cast(VllmModelForPooling, self.get_model())
to_update = model.pooler.get_pooling_updates(task)
to_update.apply(pooling_params)
req_state = CachedRequestState(
req_id=req_id,
prompt_token_ids=new_req_data.prompt_token_ids,
prompt_embeds=new_req_data.prompt_embeds,
mm_features=new_req_data.mm_features,
sampling_params=sampling_params,
pooling_params=pooling_params,
generator=generator,
block_ids=new_req_data.block_ids,
num_computed_tokens=new_req_data.num_computed_tokens,
output_token_ids=[],
lora_request=new_req_data.lora_request,
)
self.requests[req_id] = req_state
# Only relevant for models using M-RoPE (e.g, Qwen2-VL)
if self.uses_mrope:
self._init_mrope_positions(req_state)
reqs_to_add.append(req_state)
# Update the states of the running/resumed requests.
is_last_rank = get_pp_group().is_last_rank
req_data = scheduler_output.scheduled_cached_reqs
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]
num_output_tokens = req_data.num_output_tokens[i]
# Update the cached states.
req_state.num_computed_tokens = num_computed_tokens
req_index = self.input_batch.req_id_to_index.get(req_id)
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:])
elif num_output_tokens < len(req_state.output_token_ids):
# Some output tokens were discarded due to a sync-KV-load
# failure. Align the cached state.
del req_state.output_token_ids[num_output_tokens:]
if req_index is not None:
end_idx = (
self.input_batch.num_prompt_tokens[req_index]
+ num_output_tokens
)
self.input_batch.num_tokens[req_index] = end_idx
self.input_batch.num_tokens_no_spec[req_index] = end_idx
# Update the block IDs.
if not resumed_from_preemption:
if new_block_ids is not None:
# Append the new blocks to the existing block IDs.
for block_ids, new_ids in zip(req_state.block_ids, new_block_ids):
block_ids.extend(new_ids)
else:
assert req_index is None
assert new_block_ids is not None
# The request is resumed from preemption.
# Replace the existing block IDs with the new ones.
req_state.block_ids = new_block_ids
if self.use_async_scheduling and num_output_tokens > 0:
# We must recover the output token ids for resumed requests in the
# async scheduling case, so that correct input_ids are obtained.
resumed_token_ids = req_data.resumed_req_token_ids[i]
assert resumed_token_ids is not None
req_state.output_token_ids = resumed_token_ids[-num_output_tokens:]
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.
reqs_to_add.append(req_state)
continue
# Update the persistent batch.
self.input_batch.num_computed_tokens_cpu[req_index] = num_computed_tokens
if new_block_ids is not None:
self.input_batch.block_table.append_row(new_block_ids, req_index)
# For the last rank, we don't need to update the token_ids_cpu
# because the sampled tokens are already cached.
if not is_last_rank:
# 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
# When speculative decoding is used with structured output,
# the scheduler can drop draft tokens that do not
# conform to the schema. This can result in
# scheduler_output.scheduled_spec_decode_tokens being empty,
# even when speculative decoding is enabled.
self.input_batch.spec_token_ids[req_index] = spec_token_ids
# Add the new or resumed requests to the persistent batch.
# The smaller empty indices are filled first.
for request in reqs_to_add:
self.input_batch.add_request(request)
# 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 _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 _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
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
assert supports_mrope(self.get_model()), "M-RoPE support is not implemented."
req_state.mrope_positions, req_state.mrope_position_delta = (
self.model.get_mrope_input_positions(
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 _extract_mm_kwargs(
self,
scheduler_output: "SchedulerOutput",
) -> BatchedTensorInputs:
if not scheduler_output or not self.is_multimodal_raw_input_only_model:
return {}
mm_kwargs = list[MultiModalKwargsItem]()
for req in scheduler_output.scheduled_new_reqs:
for feature in req.mm_features:
if feature.data is not None:
mm_kwargs.append(feature.data)
# Input all modalities at once
model = cast(SupportsMultiModal, self.model)
mm_kwargs_combined: BatchedTensorInputs = {}
for _, _, mm_kwargs_group in group_mm_kwargs_by_modality(
mm_kwargs,
device=self.device,
pin_memory=self.pin_memory,
merge_by_field_config=model.merge_by_field_config,
):
mm_kwargs_combined.update(mm_kwargs_group)
return mm_kwargs_combined
def _dummy_mm_kwargs(self, num_seqs: int) -> BatchedTensorInputs:
if not self.is_multimodal_raw_input_only_model:
return {}
mm_budget = self.mm_budget
assert mm_budget is not None
dummy_modality = mm_budget.get_modality_with_max_tokens()
return self._get_mm_dummy_batch(dummy_modality, num_seqs)
def _get_cumsum_and_arange(
self,
num_tokens: np.ndarray,
cumsum_dtype: np.dtype | None = 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
GPU 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.copy_to_gpu(total_num_scheduled_tokens)
if self.enable_prompt_embeds:
self.inputs_embeds.copy_to_gpu(total_num_scheduled_tokens)
self.is_token_ids.copy_to_gpu(total_num_scheduled_tokens)
return
# Async scheduling case, where some decode requests from the previous
# iteration won't have entries in input_ids_cpu and need to be copied
# on the GPU from prev_sampled_token_ids.
prev_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 GPU first.
self.input_ids.copy_to_gpu(total_num_scheduled_tokens)
if self.enable_prompt_embeds:
self.inputs_embeds.copy_to_gpu(total_num_scheduled_tokens)
self.is_token_ids.copy_to_gpu(total_num_scheduled_tokens)
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.gpu[:num_commmon_tokens].copy_(
self.input_batch.prev_sampled_token_ids[:num_commmon_tokens, 0],
non_blocking=True,
)
if self.enable_prompt_embeds:
self.is_token_ids.gpu[:num_commmon_tokens] = 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.gpu.scatter_(
dim=0,
index=input_ids_index_tensor,
src=self.input_batch.prev_sampled_token_ids[
prev_common_req_indices_tensor, 0
],
)
def _get_encoder_seq_lens(
self,
scheduler_output: "SchedulerOutput",
kv_cache_spec: KVCacheSpec,
num_reqs: int,
) -> np.ndarray | None:
if not isinstance(kv_cache_spec, CrossAttentionSpec):
return None
# Build encoder_seq_lens array mapping request indices to
# encoder lengths for inputs scheduled in this batch
encoder_seq_lens = np.zeros(num_reqs, dtype=np.int32)
for req_id in scheduler_output.scheduled_encoder_inputs:
req_index = self.input_batch.req_id_to_index[req_id]
encoder_seq_lens[req_index] = self.max_encoder_len
return encoder_seq_lens
def _prepare_inputs(
self, scheduler_output: "SchedulerOutput"
) -> tuple[
PerLayerAttnMetadata,
torch.Tensor,
SpecDecodeMetadata | None,
np.ndarray,
CommonAttentionMetadata | None,
int,
UBatchSlices | None,
torch.Tensor | None,
bool,
]:
"""
:return: tuple[
attn_metadata: layer-to-attention_metadata mapping,
logits_indices, spec_decode_metadata,
num_scheduled_tokens, spec_decode_common_attn_metadata,
max_num_scheduled_tokens, use_cascade_attn
]
"""
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 = max(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)
# Get positions.
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)
# Get token indices.
# E.g., [0, 1, 0, 1, 2, 3, 4, 0, 1, 2]
# -> [0, 1, M, M + 1, M + 2, M + 3, M + 4, 2 * M, 2 * M + 1, 2 * M + 2]
# where M is the max_model_len.
token_indices = (
positions_np + req_indices * self.input_batch.token_ids_cpu.shape[1]
)
token_indices_tensor = torch.from_numpy(token_indices)
# NOTE(woosuk): We use torch.index_select instead of np.take here
# because torch.index_select is much faster than np.take for large
# tensors.
torch.index_select(
self.input_batch.token_ids_cpu_tensor.flatten(),
0,
token_indices_tensor,
out=self.input_ids.cpu[:total_num_scheduled_tokens],
)
if self.enable_prompt_embeds:
is_token_ids = self.input_batch.is_token_ids.flatten()
torch.index_select(
is_token_ids,
0,
token_indices_tensor,
out=self.is_token_ids.cpu[:total_num_scheduled_tokens],
)
# Because we did not pre-allocate a massive prompt_embeds CPU tensor on
# the InputBatch, we need to fill in the prompt embeds into the expected
# spots in the GpuModelRunner's pre-allocated prompt_embeds tensor.
if self.input_batch.req_prompt_embeds:
output_idx = 0
for req_idx in range(num_reqs):
num_sched = num_scheduled_tokens[req_idx]
# Skip if this request doesn't have embeddings
if req_idx not in self.input_batch.req_prompt_embeds:
output_idx += num_sched
continue
# Skip if no tokens scheduled
if num_sched <= 0:
output_idx += num_sched
continue
req_embeds = self.input_batch.req_prompt_embeds[req_idx]
start_pos = self.input_batch.num_computed_tokens_cpu[req_idx]
# Skip if trying to read beyond available embeddings
if start_pos >= req_embeds.shape[0]:
output_idx += num_sched
continue
# Copy available embeddings
end_pos = start_pos + num_sched
actual_end = min(end_pos, req_embeds.shape[0])
actual_num_sched = actual_end - start_pos
if actual_num_sched > 0:
self.inputs_embeds.cpu[
output_idx : output_idx + actual_num_sched
].copy_(req_embeds[start_pos:actual_end])
output_idx += num_sched
self.input_batch.block_table.compute_slot_mapping(req_indices, positions_np)
self.input_batch.block_table.commit_slot_mapping(total_num_scheduled_tokens)
# Prepare the attention metadata.
self.query_start_loc.np[0] = 0
self.query_start_loc.np[1 : num_reqs + 1] = cu_num_tokens
# Note: pad query_start_loc to be non-decreasing, as kernels
# like FlashAttention requires that
self.query_start_loc.np[num_reqs + 1 :].fill(cu_num_tokens[-1])
self.query_start_loc.copy_to_gpu()
query_start_loc = self.query_start_loc.gpu[: num_reqs + 1]
num_tokens_unpadded = scheduler_output.total_num_scheduled_tokens
num_tokens_padded = self._get_num_input_tokens(num_tokens_unpadded)
uniform_decode = (
max_num_scheduled_tokens == self.uniform_decode_query_len
) and (total_num_scheduled_tokens == num_reqs * max_num_scheduled_tokens)
# Disable DP padding when running eager to avoid excessive padding when
# running prefills. This lets us set enforce_eager on the prefiller in
# a P/D setup and still use CUDA graphs (enabled by this padding) on the
# decoder.
allow_dp_padding = self.compilation_config.cudagraph_mode != CUDAGraphMode.NONE
ubatch_slices, num_tokens_across_dp = coordinate_batch_across_dp(
num_tokens_unpadded=num_tokens_unpadded,
parallel_config=self.parallel_config,
allow_microbatching=True,
allow_dp_padding=allow_dp_padding,
num_tokens_padded=num_tokens_padded,
uniform_decode=uniform_decode,
num_scheduled_tokens_per_request=num_scheduled_tokens,
)
self.seq_lens.np[:num_reqs] = (
self.input_batch.num_computed_tokens_cpu[:num_reqs] + num_scheduled_tokens
)
# Fill unused with 0 for full cuda graph mode.
self.seq_lens.np[num_reqs:].fill(0)
self.seq_lens.copy_to_gpu()
seq_lens = self.seq_lens.gpu[:num_reqs]
max_seq_len = self.seq_lens.np[:num_reqs].max().item()
num_tokens = [self.requests[r].num_tokens for r in self.input_batch.req_ids]
num_tokens_np = np.array(num_tokens, dtype=np.int32)
# Record the index of requests that should not be sampled,
# so that we could clear the sampled tokens before returning
discard_requests_mask = self.seq_lens.np[:num_reqs] < num_tokens_np
discard_request_indices = np.nonzero(discard_requests_mask)[0]
self.num_discarded_requests = len(discard_request_indices)
self.discard_request_indices.np[: self.num_discarded_requests] = (
discard_request_indices
)
self.discard_request_indices.copy_to_gpu(self.num_discarded_requests)
# Copy the tensors to the GPU.
self._prepare_input_ids(total_num_scheduled_tokens, cu_num_tokens)
if self.uses_mrope:
# Only relevant for models using M-RoPE (e.g, Qwen2-VL)
self.mrope_positions.gpu[:, :total_num_scheduled_tokens].copy_(
self.mrope_positions.cpu[:, :total_num_scheduled_tokens],
non_blocking=True,
)
else:
# Common case (1D positions)
self.positions.copy_to_gpu(total_num_scheduled_tokens)
use_spec_decode = len(scheduler_output.scheduled_spec_decode_tokens) > 0
if not use_spec_decode:
# NOTE(woosuk): Due to chunked prefills, the batch may contain
# partial requests. While we should not sample any token
# from these partial requests, we do so for simplicity.
# We will ignore the sampled tokens from the partial requests.
# TODO: Support prompt logprobs.
logits_indices = query_start_loc[1:] - 1
num_draft_tokens = None
spec_decode_metadata = None
else:
# Get the number of draft tokens for each request.
# Iterate over the dictionary rather than all requests since not all
# requests have draft tokens.
num_draft_tokens = np.zeros(num_reqs, dtype=np.int32)
# For chunked prefills, use -1 as mask rather than 0, as guided
# decoding may rollback speculative tokens.
num_decode_draft_tokens = np.full(num_reqs, -1, dtype=np.int32)
for (
req_id,
draft_token_ids,
) in scheduler_output.scheduled_spec_decode_tokens.items():
req_idx = self.input_batch.req_id_to_index[req_id]
num_draft_tokens[req_idx] = len(draft_token_ids)
num_decode_draft_tokens[req_idx] = (
len(draft_token_ids)
if (
self.input_batch.num_computed_tokens_cpu[req_idx]
>= self.input_batch.num_prompt_tokens[req_idx]
)
else -1
)
spec_decode_metadata = self._calc_spec_decode_metadata(
num_draft_tokens, cu_num_tokens
)
logits_indices = spec_decode_metadata.logits_indices
# For DECODE only cuda graph of some attention backends (e.g., GDN).
self.num_decode_draft_tokens.np[:num_reqs] = num_decode_draft_tokens
self.num_decode_draft_tokens.np[num_reqs:].fill(-1)
self.num_decode_draft_tokens.copy_to_gpu()
logits_indices_padded = None
if self.cache_config.kv_sharing_fast_prefill:
logits_indices_padded = self._prepare_kv_sharing_fast_prefill(
logits_indices
)
attn_metadata: PerLayerAttnMetadata = {}
if ubatch_slices is not None:
attn_metadata = [dict() for _ in range(len(ubatch_slices))]
use_cascade_attn = False
# Used in the below loop.
query_start_loc_cpu = self.query_start_loc.cpu[: num_reqs + 1]
seq_lens_cpu = self.seq_lens.cpu[:num_reqs]
num_computed_tokens_cpu = self.input_batch.num_computed_tokens_cpu_tensor[
:num_reqs
]
spec_decode_common_attn_metadata = None
if use_spec_decode:
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
):
encoder_seq_lens = self._get_encoder_seq_lens(
scheduler_output, kv_cache_group_spec.kv_cache_spec, num_reqs
)
if isinstance(kv_cache_group_spec.kv_cache_spec, EncoderOnlyAttentionSpec):
# Encoder-only layers do not have KV cache, so we need to
# create a dummy block table and slot mapping for them.
blk_table_tensor = torch.zeros(
(num_reqs, 1),
dtype=torch.int32,
device=self.device,
)
slot_mapping = torch.zeros(
(total_num_scheduled_tokens,),
dtype=torch.int64,
device=self.device,
)
num_common_prefix_blocks = 0
else:
blk_table = self.input_batch.block_table[kv_cache_group_id]
blk_table_tensor = blk_table.get_device_tensor(num_reqs)
slot_mapping = blk_table.slot_mapping.gpu[:total_num_scheduled_tokens]
# Fill unused with -1. Needed for reshape_and_cache in full cuda
# graph mode.
blk_table.slot_mapping.gpu[total_num_scheduled_tokens:].fill_(-1)
num_common_prefix_blocks = scheduler_output.num_common_prefix_blocks[
kv_cache_group_id
]
common_attn_metadata = CommonAttentionMetadata(
query_start_loc=query_start_loc,
query_start_loc_cpu=query_start_loc_cpu,
seq_lens=seq_lens,
seq_lens_cpu=seq_lens_cpu,
num_computed_tokens_cpu=num_computed_tokens_cpu,
num_reqs=num_reqs,
num_actual_tokens=total_num_scheduled_tokens,
max_query_len=max_num_scheduled_tokens,
max_seq_len=max_seq_len,
block_table_tensor=blk_table_tensor,
slot_mapping=slot_mapping,
logits_indices_padded=logits_indices_padded,
num_logits_indices=logits_indices.size(0),
causal=True,
encoder_seq_lens=encoder_seq_lens,
dcp_local_seq_lens=self.dcp_local_seq_lens.gpu[:num_reqs]
if self.dcp_world_size > 1
else None,
)
if self.speculative_config and spec_decode_common_attn_metadata is None:
if isinstance(self.drafter, EagleProposer):
if (
self.drafter.attn_layer_names[0]
in kv_cache_group_spec.layer_names
):
spec_decode_common_attn_metadata = common_attn_metadata
else:
spec_decode_common_attn_metadata = common_attn_metadata
for attn_group in self.attn_groups[kv_cache_group_id]:
# Prepare for cascade attention if enabled & beneficial.
common_prefix_len = 0
builder = attn_group.get_metadata_builder()
if self.cascade_attn_enabled:
common_prefix_len = self._compute_cascade_attn_prefix_len(
num_scheduled_tokens,
num_common_prefix_blocks,
attn_group.kv_cache_spec,
builder,
)
extra_attn_metadata_args = {}
if use_spec_decode and isinstance(builder, GDNAttentionMetadataBuilder):
extra_attn_metadata_args = dict(
num_accepted_tokens=self.num_accepted_tokens.gpu[:num_reqs],
num_decode_draft_tokens_cpu=self.num_decode_draft_tokens.cpu[
:num_reqs
],
)
if ubatch_slices is not None:
common_attn_metadata_list = split_attn_metadata(
ubatch_slices, common_attn_metadata
)
for ubid, common_attn_metadata in enumerate(
common_attn_metadata_list
):
attn_metadata_i = attn_group.get_metadata_builder(
ubatch_id=ubid
).build(
common_prefix_len=common_prefix_len,
common_attn_metadata=common_attn_metadata,
)
for layer_name in kv_cache_group_spec.layer_names:
assert type(attn_metadata) is list
attn_metadata[ubid][layer_name] = attn_metadata_i
else:
assert isinstance(attn_metadata, dict)
attn_metadata_i = builder.build(
common_prefix_len=common_prefix_len,
common_attn_metadata=common_attn_metadata,
**extra_attn_metadata_args,
)
use_cascade_attn |= getattr(attn_metadata_i, "use_cascade", False)
for layer_name in attn_group.layer_names:
attn_metadata[layer_name] = attn_metadata_i
# disable cascade attention when DBO
if ubatch_slices is not None:
use_cascade_attn = False
# Hot-Swap lora model
if self.lora_config:
self.set_active_loras(self.input_batch, num_scheduled_tokens)
return (
attn_metadata,
logits_indices,
spec_decode_metadata,
num_scheduled_tokens,
spec_decode_common_attn_metadata,
max_num_scheduled_tokens,
ubatch_slices,
num_tokens_across_dp,
use_cascade_attn,
)
def _compute_cascade_attn_prefix_len(
self,
num_scheduled_tokens: np.ndarray,
num_common_prefix_blocks: int,
kv_cache_spec: KVCacheSpec,
attn_metadata_builder: AttentionMetadataBuilder,
) -> int:
"""Compute the length of the common prefix for cascade attention.
NOTE(woosuk): The common prefix length returned by this function
represents the length used specifically for cascade attention, not the
actual number of tokens shared between requests. When cascade attention
is disabled (use_cascade=False), this function returns 0 even if
requests share common tokens. Additionally, the common prefix length is
truncated to a multiple of the block size and may be further truncated
due to implementation details explained below.
Args:
num_scheduled_tokens: Number of tokens scheduled per request.
num_common_prefix_blocks: Number of shared KV cache blocks.
Returns:
int: Length of common prefix in tokens.
"""
common_prefix_len = num_common_prefix_blocks * kv_cache_spec.block_size
if common_prefix_len == 0:
# Common case.
return 0
# NOTE(woosuk): Cascade attention uses two attention kernels: one
# for the common prefix and the other for the rest. For the first
# kernel, we concatenate all the query tokens (possibly from
# different requests) and treat them as if they are from the same
# request. Then, we use bi-directional attention to process the
# common prefix in the KV cache. Importantly, this means that the
# first kernel does not do any masking.
# Consider the following example:
# Request 1's input query: [D, E, X]
# Request 1's kv cache: [A, B, C, D, E, X]
# Request 1's num_computed_tokens: 3 (i.e., [A, B, C])
# Request 2's input query: [E, Y]
# Request 2's kv cache: [A, B, C, D, E, Y]
# Request 2's num_computed_tokens: 4 (i.e., [A, B, C, D])
# If we use [A, B, C, D, E] as the common prefix, then the
# first kernel will compute the bi-directional attention between
# input query [D, E, X, E, Y] and common prefix [A, B, C, D, E].
# However, this is wrong because D in Request 1 should not attend to
# E in the common prefix (i.e., we need masking).
# To avoid this, [A, B, C, D] should be the common prefix.
# That is, the common prefix should be capped by the minimum
# num_computed_tokens among the requests, and plus one to include
# the first token of the query.
# In practice, we use [A, B, C] as the common prefix, instead of
# [A, B, C, D] (i.e., the common prefix is capped by the minimum
# num_computed_tokens, without plus one).
# This is because of an implementation detail: We want to always
# use two kernels for cascade attention. Let's imagine:
# Request 3's input query: [D]
# Request 3's kv cache: [A, B, C, D]
# Request 3's num_computed_tokens: 3 (i.e., [A, B, C])
# If we use [A, B, C, D] as the common prefix for Request 1-3,
# then Request 3 will be processed only by the first kernel,
# and the second kernel will get an empty input. While this is not
# a fundamental problem, our current implementation does not support
# this case.
num_reqs = len(num_scheduled_tokens)
common_prefix_len = min(
common_prefix_len, self.input_batch.num_computed_tokens_cpu[:num_reqs].min()
)
# common_prefix_len should be a multiple of the block size.
common_prefix_len = (
common_prefix_len // kv_cache_spec.block_size * kv_cache_spec.block_size
)
use_sliding_window = isinstance(kv_cache_spec, SlidingWindowSpec) or (
isinstance(kv_cache_spec, FullAttentionSpec)
and kv_cache_spec.sliding_window is not None
)
use_local_attention = isinstance(kv_cache_spec, ChunkedLocalAttentionSpec) or (
isinstance(kv_cache_spec, FullAttentionSpec)
and kv_cache_spec.attention_chunk_size is not None
)
assert isinstance(kv_cache_spec, AttentionSpec)
use_cascade = attn_metadata_builder.use_cascade_attention(
common_prefix_len=common_prefix_len,
query_lens=num_scheduled_tokens,
num_query_heads=self.num_query_heads,
num_kv_heads=kv_cache_spec.num_kv_heads,
use_alibi=self.use_alibi,
use_sliding_window=use_sliding_window,
use_local_attention=use_local_attention,
num_sms=self.num_sms,
dcp_world_size=self.dcp_world_size,
)
return common_prefix_len if use_cascade else 0
def _calc_mrope_positions(self, scheduler_output: "SchedulerOutput"):
mrope_pos_ptr = 0
for index, req_id in enumerate(self.input_batch.req_ids):
req = self.requests[req_id]
assert req.mrope_positions is not None
num_computed_tokens = self.input_batch.num_computed_tokens_cpu[index]
num_scheduled_tokens = scheduler_output.num_scheduled_tokens[req_id]
num_prompt_tokens = length_from_prompt_token_ids_or_embeds(
req.prompt_token_ids, req.prompt_embeds
)
if num_computed_tokens + num_scheduled_tokens > num_prompt_tokens:
prompt_part_len = max(0, num_prompt_tokens - num_computed_tokens)
completion_part_len = max(0, num_scheduled_tokens - prompt_part_len)
else:
prompt_part_len = num_scheduled_tokens
completion_part_len = 0
assert num_scheduled_tokens == prompt_part_len + completion_part_len
if prompt_part_len > 0:
# prompt's mrope_positions are pre-computed
dst_start = mrope_pos_ptr
dst_end = mrope_pos_ptr + prompt_part_len
src_start = num_computed_tokens
src_end = num_computed_tokens + prompt_part_len
self.mrope_positions.cpu[:, dst_start:dst_end] = req.mrope_positions[
:, src_start:src_end
]
mrope_pos_ptr += prompt_part_len
if completion_part_len > 0:
# compute completion's mrope_positions on-the-fly
dst_start = mrope_pos_ptr
dst_end = mrope_pos_ptr + completion_part_len
MRotaryEmbedding.get_next_input_positions_tensor(
out=self.mrope_positions.np,
out_offset=dst_start,
mrope_position_delta=req.mrope_position_delta,
context_len=num_computed_tokens + prompt_part_len,
num_new_tokens=completion_part_len,
)
mrope_pos_ptr += completion_part_len
def _calc_spec_decode_metadata(
self,
num_draft_tokens: np.ndarray,
cu_num_scheduled_tokens: np.ndarray,
) -> SpecDecodeMetadata:
# Inputs:
# cu_num_scheduled_tokens: [ 4, 104, 107, 207, 209]
# num_draft_tokens: [ 3, 0, 2, 0, 1]
# Outputs:
# cu_num_draft_tokens: [ 3, 3, 5, 5, 6]
# logits_indices: [ 0, 1, 2, 3, 103, 104, 105, 106,
# 206, 207, 208]
# target_logits_indices: [ 0, 1, 2, 5, 6, 9]
# bonus_logits_indices: [ 3, 4, 7, 8, 10]
# Compute the logits indices.
# [4, 1, 3, 1, 2]
num_sampled_tokens = num_draft_tokens + 1
# Step 1. cu_num_sampled_tokens: [4, 5, 8, 9, 11]
# arange: [0, 1, 2, 3, 0, 0, 1, 2, 0, 0, 1]
cu_num_sampled_tokens, arange = self._get_cumsum_and_arange(
num_sampled_tokens, cumsum_dtype=np.int32
)
# Step 2. [0, 0, 0, 0, 103, 104, 104, 104, 206, 207, 207]
logits_indices = np.repeat(
cu_num_scheduled_tokens - num_sampled_tokens, num_sampled_tokens
)
# Step 3. [0, 1, 2, 3, 103, 104, 105, 106, 206, 207, 208]
logits_indices += arange
# Compute the bonus logits indices.
bonus_logits_indices = cu_num_sampled_tokens - 1
# Compute the draft logits indices.
# cu_num_draft_tokens: [3, 3, 5, 5, 6]
# arange: [0, 1, 2, 0, 1, 0]
cu_num_draft_tokens, arange = self._get_cumsum_and_arange(
num_draft_tokens, cumsum_dtype=np.int32
)
# [0, 0, 0, 5, 5, 9]
target_logits_indices = np.repeat(
cu_num_sampled_tokens - num_sampled_tokens, num_draft_tokens
)
# [0, 1, 2, 5, 6, 9]
target_logits_indices += arange
# TODO: Optimize the CPU -> GPU copy.
cu_num_draft_tokens = torch.from_numpy(cu_num_draft_tokens).to(
self.device, non_blocking=True
)
cu_num_sampled_tokens = torch.from_numpy(cu_num_sampled_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.gpu[logits_indices]
draft_token_ids = draft_token_ids[target_logits_indices + 1]
return SpecDecodeMetadata(
draft_token_ids=draft_token_ids,
num_draft_tokens=num_draft_tokens.tolist(),
cu_num_draft_tokens=cu_num_draft_tokens,
cu_num_sampled_tokens=cu_num_sampled_tokens,
target_logits_indices=target_logits_indices,
bonus_logits_indices=bonus_logits_indices,
logits_indices=logits_indices,
)
def _prepare_kv_sharing_fast_prefill(
self,
logits_indices: torch.Tensor,
) -> torch.Tensor:
assert self.kv_sharing_fast_prefill_logits_indices is not None
num_logits = logits_indices.shape[0]
assert num_logits > 0
self.kv_sharing_fast_prefill_logits_indices[:num_logits].copy_(logits_indices)
# There might have leftover indices in logits_indices[num_logits:]
# from previous iterations, whose values may be greater than the
# batch size in the current iteration. To ensure indices are always
# valid, we fill the padded indices with the last index.
self.kv_sharing_fast_prefill_logits_indices[num_logits:].fill_(
logits_indices[-1].item()
)
if (
self.compilation_config.cudagraph_mode != CUDAGraphMode.NONE
and num_logits <= self.cudagraph_batch_sizes[-1]
):
# Use piecewise CUDA graphs.
# Add padding to the batch size.
num_logits_padded = self.vllm_config.pad_for_cudagraph(num_logits)
else:
num_logits_padded = num_logits
logits_indices_padded = self.kv_sharing_fast_prefill_logits_indices[
:num_logits_padded
]
return logits_indices_padded
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]
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 _execute_mm_encoder(self, scheduler_output: "SchedulerOutput"):
# Batch the multi-modal inputs using the helper method.
mm_kwargs, mm_hashes_pos = self._batch_mm_kwargs_from_scheduler(
scheduler_output
)
if not mm_kwargs:
return
# Batch mm inputs as much as we can: if a request in the batch has
# multiple modalities or a different modality than the previous one,
# we process it separately to preserve item order.
# FIXME(ywang96): This is a hacky way to deal with multiple modalities
# in the same batch while still being able to benefit from batching
# multimodal inputs. The proper solution should be reordering the
# encoder outputs.
model = cast(SupportsMultiModal, self.model)
encoder_outputs = []
for modality, num_items, mm_kwargs_group in group_mm_kwargs_by_modality(
mm_kwargs,
device=self.device,
pin_memory=self.pin_memory,
merge_by_field_config=model.merge_by_field_config,
):
curr_group_outputs = []
# EVS-related change.
# (ekhvedchenia): Temporary hack to limit peak memory usage when
# processing multimodal data. This solves the issue with scheduler
# putting too many video samples into a single batch. Scheduler
# uses pruned vision tokens count to compare it versus compute
# budget which is incorrect (Either input media size or non-pruned
# output vision tokens count should be considered)
# TODO(ywang96): Fix memory profiling to take EVS into account and
# remove this hack.
if (
self.is_multimodal_pruning_enabled
and modality == "video"
and num_items > 1
):
for video_mm_kwargs_item in filter(
lambda item: item.modality == "video", mm_kwargs
):
_, _, micro_batch_mm_inputs = next(
group_mm_kwargs_by_modality(
[video_mm_kwargs_item],
device=self.device,
pin_memory=self.pin_memory,
merge_by_field_config=model.merge_by_field_config,
)
)
micro_batch_outputs = model.get_multimodal_embeddings(
**micro_batch_mm_inputs
)
curr_group_outputs.extend(micro_batch_outputs)
else:
# 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 = model.get_multimodal_embeddings(**mm_kwargs_group)
sanity_check_mm_encoder_outputs(
curr_group_outputs,
expected_num_items=num_items,
)
encoder_outputs.extend(curr_group_outputs)
# Cache the encoder outputs by mm_hash
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 _gather_mm_embeddings(
self,
scheduler_output: "SchedulerOutput",
shift_computed_tokens: int = 0,
) -> tuple[list[torch.Tensor], torch.Tensor]:
total_num_scheduled_tokens = scheduler_output.total_num_scheduled_tokens
mm_embeds = list[torch.Tensor]()
is_mm_embed = self.is_mm_embed.cpu
is_mm_embed[:total_num_scheduled_tokens] = False
req_start_idx = 0
should_sync_mrope_positions = False
for req_id in self.input_batch.req_ids:
mm_embeds_req: list[torch.Tensor] = []
num_scheduled_tokens = scheduler_output.num_scheduled_tokens[req_id]
req_state = self.requests[req_id]
num_computed_tokens = req_state.num_computed_tokens + shift_computed_tokens
for mm_feature in req_state.mm_features:
pos_info = mm_feature.mm_position
start_pos = pos_info.offset
num_encoder_tokens = pos_info.length
# The encoder output is needed if the two ranges overlap:
# [num_computed_tokens,
# num_computed_tokens + num_scheduled_tokens) and
# [start_pos, start_pos + num_encoder_tokens)
if start_pos >= num_computed_tokens + num_scheduled_tokens:
# The encoder output is not needed in this step.
break
if start_pos + num_encoder_tokens <= num_computed_tokens:
# The encoder output is already processed and stored
# in the decoder's KV cache.
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
mm_hash = mm_feature.identifier
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 := pos_info.is_embed) is not None:
is_embed = is_embed[start_idx:end_idx]
req_start_pos = req_start_idx + start_pos - num_computed_tokens
is_mm_embed[req_start_pos + start_idx : req_start_pos + end_idx] = (
True if is_embed is None else is_embed
)
mm_embeds_item = gather_mm_placeholders(
encoder_output[start_idx:end_idx],
is_embed=is_embed,
)
mm_embeds_req.append(mm_embeds_item)
if self.is_multimodal_pruning_enabled and self.uses_mrope:
assert req_state.mrope_positions is not None
should_sync_mrope_positions = True
mm_embeds_req, new_mrope_positions, new_delta = (
self.model.recompute_mrope_positions(
input_ids=req_state.prompt_token_ids,
multimodal_embeddings=mm_embeds_req,
mrope_positions=req_state.mrope_positions,
num_computed_tokens=req_state.num_computed_tokens,
)
)
req_state.mrope_positions.copy_(new_mrope_positions)
req_state.mrope_position_delta = new_delta
mm_embeds.extend(mm_embeds_req)
req_start_idx += num_scheduled_tokens
is_mm_embed = self.is_mm_embed.copy_to_gpu(total_num_scheduled_tokens)
if should_sync_mrope_positions:
self._calc_mrope_positions(scheduler_output)
self.mrope_positions.copy_to_gpu(total_num_scheduled_tokens)
return mm_embeds, is_mm_embed
def _extract_encoder_inputs(
self,
scheduler_output: "SchedulerOutput",
) -> dict[str, torch.Tensor]:
"""Extract encoder inputs for encoder-decoder models.
This method extracts multimodal input features from scheduled encoder
inputs and formats them for the encoder-decoder model forward pass.
"""
# Batch the multi-modal inputs using the helper method.
mm_kwargs, _ = self._batch_mm_kwargs_from_scheduler(scheduler_output)
if not mm_kwargs:
return {}
# Group MM kwargs by modality and extract features
model = cast(SupportsMultiModal, self.model)
encoder_features = {}
for _, _, mm_kwargs_group in group_mm_kwargs_by_modality(
mm_kwargs,
device=self.device,
pin_memory=self.pin_memory,
merge_by_field_config=model.merge_by_field_config,
):
# Add the grouped features to encoder_features dict
# This allows the model to receive them as kwargs (e.g.,
# input_features=...)
encoder_features.update(mm_kwargs_group)
return encoder_features
def get_model(self) -> nn.Module:
# get raw model out of the cudagraph wrapper.
if isinstance(self.model, (CUDAGraphWrapper, UBatchWrapper)):
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_pooling_tasks(self) -> list[PoolingTask]:
model = self.get_model()
if not is_pooling_model(model):
return []
supported_tasks = list(model.pooler.get_supported_tasks())
if self.scheduler_config.chunked_prefill_enabled:
if "token_embed" in supported_tasks:
supported_tasks.remove("token_embed")
if "token_classify" in supported_tasks:
supported_tasks.remove("token_classify")
logger.debug_once(
"Chunked prefill is not supported with "
"token_embed and token_classify tasks "
"which using ALL pooling. "
"Please turn off chunked prefill by "
"`--no-enable-chunked-prefill` before using it."
)
if "score" in supported_tasks:
num_labels = getattr(self.model_config.hf_config, "num_labels", 0)
if num_labels != 1:
supported_tasks.remove("score")
logger.debug_once("Score API is only enabled for num_labels == 1.")
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 sync_and_slice_intermediate_tensors(
self,
num_tokens: int,
intermediate_tensors: IntermediateTensors,
sync_self: bool,
) -> IntermediateTensors:
assert self.intermediate_tensors is not None
tp = self.vllm_config.parallel_config.tensor_parallel_size
is_rs = is_residual_scattered_for_sp(self.vllm_config, num_tokens)
# When sequence parallelism is enabled, the "residual" tensor is sharded
# across tensor parallel ranks, so each rank only needs its own slice.
if sync_self:
assert intermediate_tensors is not None
for k, v in intermediate_tensors.items():
is_scattered = k == "residual" and is_rs
copy_len = num_tokens // tp if is_scattered else num_tokens
self.intermediate_tensors[k][:copy_len].copy_(
v[:copy_len], non_blocking=True
)
return IntermediateTensors(
{
k: v[: num_tokens // tp]
if k == "residual" and is_rs
else v[:num_tokens]
for k, v in self.intermediate_tensors.items()
}
)
def eplb_step(self, is_dummy: bool = False, is_profile: bool = False) -> None:
"""
Step for the EPLB (Expert Parallelism Load Balancing) state.
"""
if not self.parallel_config.enable_eplb:
return
assert self.eplb_state is not None
model = self.get_model()
assert is_mixture_of_experts(model)
self.eplb_state.step(
model,
is_dummy,
is_profile,
log_stats=self.parallel_config.eplb_config.log_balancedness,
)
# This is where the second ubatch is adjusted to account for the padding.
# Should be called after attention metadata creation. This just pads
# the second ubatch slice out to the total number of tokens
# (num_tokens + padding)
@staticmethod
def pad_out_ubatch_slice(ubatch_slices: UBatchSlices, num_total_tokens: int):
padded_second_ubatch_slice = slice(
ubatch_slices[1].token_slice.start, num_total_tokens
)
ubatch_slices[1] = UBatchSlice(
padded_second_ubatch_slice, padded_second_ubatch_slice
)
def _pool(
self,
hidden_states: torch.Tensor,
num_scheduled_tokens: int,
num_scheduled_tokens_np: np.ndarray,
) -> 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.get_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: PoolerOutput = 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,
)
self._sync_device()
pooler_output: list[torch.Tensor | None] = []
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,
)
def _get_num_input_tokens(self, num_scheduled_tokens: int) -> int:
if (
self.compilation_config.cudagraph_mode != CUDAGraphMode.NONE
and hasattr(self, "cudagraph_batch_sizes")
and self.cudagraph_batch_sizes
and num_scheduled_tokens <= self.cudagraph_batch_sizes[-1]
):
# Use CUDA graphs.
# Add padding to the batch size.
return self.vllm_config.pad_for_cudagraph(num_scheduled_tokens)
# Eager mode.
# Pad tokens to multiple of tensor_parallel_size when
# enabled collective fusion for SP
tp_size = self.vllm_config.parallel_config.tensor_parallel_size
if (
self.compilation_config.pass_config.enable_sequence_parallelism
and tp_size > 1
):
return round_up(num_scheduled_tokens, tp_size)
return num_scheduled_tokens
def _preprocess(
self,
scheduler_output: "SchedulerOutput",
num_input_tokens: int, # Padded
intermediate_tensors: IntermediateTensors | None = None,
) -> tuple[
int,
torch.Tensor | None,
torch.Tensor | None,
torch.Tensor,
IntermediateTensors | None,
dict[str, Any],
]:
num_scheduled_tokens = scheduler_output.total_num_scheduled_tokens
is_first_rank = get_pp_group().is_first_rank
# _prepare_inputs may reorder the batch, so we must gather multi
# modal outputs after that to ensure the correct order
if (
self.supports_mm_inputs
and is_first_rank
and not self.model_config.is_encoder_decoder
):
# Run the multimodal encoder if any.
self._execute_mm_encoder(scheduler_output)
mm_embeds, is_mm_embed = self._gather_mm_embeddings(scheduler_output)
# NOTE(woosuk): To unify token ids and soft tokens (vision
# embeddings), we always use embeddings (rather than token ids)
# as input to the multimodal model, even when the input is text.
inputs_embeds_scheduled = self.model.get_input_embeddings(
self.input_ids.gpu[:num_scheduled_tokens],
multimodal_embeddings=mm_embeds,
is_multimodal=is_mm_embed,
)
# TODO(woosuk): Avoid the copy. Optimize.
self.inputs_embeds.gpu[:num_scheduled_tokens].copy_(inputs_embeds_scheduled)
input_ids = None
inputs_embeds = self.inputs_embeds.gpu[:num_input_tokens]
model_kwargs = {
**self._init_model_kwargs(num_scheduled_tokens),
**self._extract_mm_kwargs(scheduler_output),
}
elif self.enable_prompt_embeds and is_first_rank:
# Get the input embeddings for the tokens that are not input embeds,
# then put them into the appropriate positions.
# TODO(qthequartermasterman): Since even when prompt embeds are
# enabled, (a) not all requests will use prompt embeds, and (b)
# after the initial prompt is processed, the rest of the generated
# tokens will be token ids, it is not desirable to have the
# embedding layer outside of the CUDA graph all the time. The v0
# engine avoids this by "double compiling" the CUDA graph, once
# with input_ids and again with inputs_embeds, for all num_tokens.
# If a batch only has token ids, then including the embedding layer
# in the CUDA graph will be more performant (like in the else case
# below).
token_ids_idx = (
self.is_token_ids.gpu[:num_scheduled_tokens]
.nonzero(as_tuple=False)
.squeeze(1)
)
# Some tokens ids may need to become embeds
if token_ids_idx.numel() > 0:
token_ids = self.input_ids.gpu[token_ids_idx]
tokens_to_embeds = self.model.get_input_embeddings(input_ids=token_ids)
self.inputs_embeds.gpu[token_ids_idx] = tokens_to_embeds
inputs_embeds = self.inputs_embeds.gpu[:num_input_tokens]
model_kwargs = self._init_model_kwargs(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 CUDA graph.
input_ids = self.input_ids.gpu[:num_input_tokens]
inputs_embeds = None
model_kwargs = self._init_model_kwargs(num_input_tokens)
if self.uses_mrope:
positions = self.mrope_positions.gpu[:, :num_input_tokens]
else:
positions = self.positions.gpu[:num_input_tokens]
if is_first_rank:
intermediate_tensors = None
else:
intermediate_tensors = self.sync_and_slice_intermediate_tensors(
num_input_tokens, intermediate_tensors, True
)
if (
self.model_config.is_encoder_decoder
and scheduler_output.scheduled_encoder_inputs
):
encoder_inputs = self._extract_encoder_inputs(scheduler_output)
model_kwargs.update(encoder_inputs)
return (
num_scheduled_tokens,
input_ids,
inputs_embeds,
positions,
intermediate_tensors,
model_kwargs,
)
def _sample(
self,
logits: torch.Tensor | None,
spec_decode_metadata: SpecDecodeMetadata | None,
) -> SamplerOutput:
# Sample the next token and get logprobs if needed.
sampling_metadata = self.input_batch.sampling_metadata
if spec_decode_metadata is None:
# Update output token ids with tokens sampled in last step
# if async scheduling and required by current sampling params.
self.input_batch.update_async_output_token_ids()
return self.sampler(
logits=logits,
sampling_metadata=sampling_metadata,
)
sampler_output = self.rejection_sampler(
spec_decode_metadata,
None, # draft_probs
logits,
sampling_metadata,
)
self._update_states_after_model_execute(sampler_output.sampled_token_ids)
return sampler_output
def _bookkeeping_sync(
self,
scheduler_output: "SchedulerOutput",
sampler_output: SamplerOutput,
logits: torch.Tensor | None,
hidden_states: torch.Tensor,
num_scheduled_tokens: int,
spec_decode_metadata: SpecDecodeMetadata | None,
) -> tuple[
dict[str, int],
LogprobsLists | None,
list[list[int]],
dict[str, LogprobsTensors | None],
list[str],
dict[str, int],
list[int],
]:
num_nans_in_logits = {}
if envs.VLLM_COMPUTE_NANS_IN_LOGITS:
num_nans_in_logits = self._get_nans_in_logits(logits)
discard_sampled_tokens_req_indices = self.discard_request_indices.np[
: self.num_discarded_requests
]
for i in discard_sampled_tokens_req_indices:
gen = self.input_batch.generators.get(int(i))
if gen is not None:
gen.set_offset(gen.get_offset() - 4)
# Copy some objects so they don't get modified after returning.
# This is important when using async scheduling.
req_ids_output_copy = self.input_batch.req_ids.copy()
req_id_to_index_output_copy = self.input_batch.req_id_to_index.copy()
num_sampled_tokens = sampler_output.sampled_token_ids.shape[0]
sampled_token_ids = sampler_output.sampled_token_ids
invalid_req_indices = []
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 = self._to_list(sampled_token_ids)
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[int(i)].clear()
else:
valid_sampled_token_ids = []
invalid_req_indices = discard_sampled_tokens_req_indices.tolist()
invalid_req_indices_set = set(invalid_req_indices)
assert sampled_token_ids.shape[-1] == 1
# Cache the sampled tokens on the GPU and avoid CPU sync.
# These will be copied into input_ids in the next step
# when preparing inputs.
self.input_batch.prev_sampled_token_ids = sampled_token_ids
self.input_batch.prev_req_id_to_index = {
req_id: i
for i, req_id in enumerate(self.input_batch.req_ids)
if i not in invalid_req_indices_set
}
# Cache the sampled tokens in the model runner, so that the scheduler
# doesn't need to send them back.
# NOTE(woosuk): As an exception, when using PP, the scheduler sends
# the sampled tokens back, because there's no direct communication
# between the first-stage worker and the last-stage worker.
req_ids = self.input_batch.req_ids
logprobs_tensors = sampler_output.logprobs_tensors
cu_num_accepted_tokens = (
[0] if spec_decode_metadata and logprobs_tensors else None
)
for req_idx in range(num_sampled_tokens):
if self.use_async_scheduling:
sampled_ids = [-1] if req_idx not in invalid_req_indices_set else None
else:
sampled_ids = valid_sampled_token_ids[req_idx]
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.max_model_len, (
"Sampled token IDs exceed the max model length. "
f"Total number of tokens: {end_idx} > max_model_len: "
f"{self.max_model_len}"
)
self.input_batch.token_ids_cpu[req_idx, start_idx:end_idx] = sampled_ids
self.input_batch.is_token_ids[req_idx, start_idx:end_idx] = True
self.input_batch.num_tokens_no_spec[req_idx] = end_idx
self.input_batch.num_tokens[req_idx] = end_idx
req_id = req_ids[req_idx]
req_state = self.requests[req_id]
req_state.output_token_ids.extend(sampled_ids)
if cu_num_accepted_tokens is not None:
cu_num_accepted_tokens.append(
cu_num_accepted_tokens[-1] + len(sampled_ids)
)
# NOTE: GPU -> CPU Sync happens here.
# Move as many CPU operations as possible before this sync point.
logprobs_lists = (
logprobs_tensors.tolists(cu_num_accepted_tokens)
if logprobs_tensors is not None
else None
)
# Compute prompt logprobs if needed.
prompt_logprobs_dict = self._get_prompt_logprobs_dict(
hidden_states[:num_scheduled_tokens],
scheduler_output.num_scheduled_tokens,
)
return (
num_nans_in_logits,
logprobs_lists,
valid_sampled_token_ids,
prompt_logprobs_dict,
req_ids_output_copy,
req_id_to_index_output_copy,
invalid_req_indices,
)
@contextmanager
def synchronize_input_prep(self):
if self.prepare_inputs_event is None:
yield
return
# Ensure prior step has finished with reused CPU tensors.
# This is required in the async scheduling case because
# the CPU->GPU transfer happens async.
self.prepare_inputs_event.synchronize()
try:
yield
finally:
self.prepare_inputs_event.record()
def _model_forward(
self,
input_ids: torch.Tensor | None = None,
positions: torch.Tensor | None = None,
intermediate_tensors: IntermediateTensors | None = None,
inputs_embeds: torch.Tensor | None = None,
**model_kwargs: dict[str, Any],
) -> Any:
"""Helper method to call the model forward pass.
This method can be overridden by subclasses for model execution.
Motivation: We can inspect only this method versus
the whole execute_model, which has additional logic.
Args:
input_ids: Input token IDs
positions: Token positions
intermediate_tensors: Tensors from previous pipeline stages
inputs_embeds: Input embeddings (alternative to input_ids)
**model_kwargs: Additional model arguments
Returns:
Model output tensor
"""
return self.model(
input_ids=input_ids,
positions=positions,
intermediate_tensors=intermediate_tensors,
inputs_embeds=inputs_embeds,
**model_kwargs,
)
@torch.inference_mode()
def execute_model(
self,
scheduler_output: "SchedulerOutput",
intermediate_tensors: IntermediateTensors | None = None,
) -> ModelRunnerOutput | AsyncModelRunnerOutput | IntermediateTensors:
with record_function_or_nullcontext("Preprocess"):
with self.synchronize_input_prep():
# Update persistent batch states.
self._update_states(scheduler_output)
if not scheduler_output.total_num_scheduled_tokens:
if not has_kv_transfer_group():
# Return empty ModelRunnerOutput if no work to do.
return EMPTY_MODEL_RUNNER_OUTPUT
return self.kv_connector_no_forward(
scheduler_output, self.vllm_config
)
if self.cache_config.kv_sharing_fast_prefill:
assert not self.input_batch.num_prompt_logprobs, (
"--kv-sharing-fast-prefill produces incorrect "
"logprobs for prompt tokens, tokens, please disable "
"it when the requests need prompt logprobs"
)
# Prepare the decoder inputs.
(
attn_metadata,
logits_indices,
spec_decode_metadata,
num_scheduled_tokens_np,
spec_decode_common_attn_metadata,
max_query_len,
ubatch_slices,
num_tokens_across_dp,
use_cascade_attn,
) = self._prepare_inputs(scheduler_output)
dp_rank = self.parallel_config.data_parallel_rank
if ubatch_slices:
assert num_tokens_across_dp is not None
num_input_tokens = int(num_tokens_across_dp[dp_rank].item())
self.pad_out_ubatch_slice(ubatch_slices, num_input_tokens)
elif num_tokens_across_dp is not None:
num_input_tokens = int(num_tokens_across_dp[dp_rank].item())
else:
num_input_tokens = self._get_num_input_tokens(
scheduler_output.total_num_scheduled_tokens
)
(
num_scheduled_tokens,
input_ids,
inputs_embeds,
positions,
intermediate_tensors,
model_kwargs,
) = self._preprocess(
scheduler_output, num_input_tokens, intermediate_tensors
)
uniform_decode = (max_query_len == self.uniform_decode_query_len) and (
num_scheduled_tokens == self.input_batch.num_reqs * max_query_len
)
batch_descriptor = BatchDescriptor(
num_tokens=num_input_tokens,
uniform_decode=uniform_decode,
has_lora=len(self.input_batch.lora_id_to_lora_request) > 0,
)
cudagraph_runtime_mode, batch_descriptor = (
self.cudagraph_dispatcher.dispatch(batch_descriptor, use_cascade_attn)
)
# Set cudagraph mode to none if calc_kv_scales is true.
if attn_metadata is not None:
metadata_list = (
attn_metadata.values()
if isinstance(attn_metadata, dict)
else [attn_metadata]
)
if any(
getattr(m, "enable_kv_scales_calculation", False) for m in metadata_list
):
cudagraph_runtime_mode = CUDAGraphMode.NONE
# Run the model.
# Use persistent buffers for CUDA graphs.
with (
set_forward_context(
attn_metadata,
self.vllm_config,
num_tokens=num_input_tokens,
num_tokens_across_dp=num_tokens_across_dp,
cudagraph_runtime_mode=cudagraph_runtime_mode,
batch_descriptor=batch_descriptor,
ubatch_slices=ubatch_slices,
),
record_function_or_nullcontext("Forward"),
self.maybe_get_kv_connector_output(scheduler_output) as kv_connector_output,
):
model_output = self._model_forward(
input_ids=input_ids,
positions=positions,
intermediate_tensors=intermediate_tensors,
inputs_embeds=inputs_embeds,
**model_kwargs,
)
with record_function_or_nullcontext("Postprocess"):
if self.use_aux_hidden_state_outputs:
# True when EAGLE 3 is used.
hidden_states, aux_hidden_states = model_output
else:
# Common case.
hidden_states = model_output
aux_hidden_states = None
if not self.broadcast_pp_output:
# Common case.
if not get_pp_group().is_last_rank:
# Return the intermediate tensors.
assert isinstance(hidden_states, IntermediateTensors)
hidden_states.kv_connector_output = kv_connector_output
return hidden_states
if self.is_pooling_model:
# Return the pooling output.
output = self._pool(
hidden_states, num_scheduled_tokens, num_scheduled_tokens_np
)
output.kv_connector_output = kv_connector_output
return output
sample_hidden_states = hidden_states[logits_indices]
logits = self.model.compute_logits(sample_hidden_states)
else:
# Rare case.
assert not self.is_pooling_model
if not get_pp_group().is_last_rank:
all_gather_tensors = {
"residual": not is_residual_scattered_for_sp(
self.vllm_config, num_input_tokens
)
}
get_pp_group().send_tensor_dict(
hidden_states.tensors,
all_gather_group=get_tp_group(),
all_gather_tensors=all_gather_tensors,
)
logits = None
else:
sample_hidden_states = hidden_states[logits_indices]
logits = self.model.compute_logits(sample_hidden_states)
model_output_broadcast_data = {}
if logits is not None:
model_output_broadcast_data["logits"] = logits.contiguous()
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.structured_output_request_ids:
apply_grammar_bitmask(scheduler_output, self.input_batch, logits)
with record_function_or_nullcontext("Sample"):
sampler_output = self._sample(logits, spec_decode_metadata)
def propose_draft_token_ids(sampled_token_ids):
assert spec_decode_common_attn_metadata is not None
with record_function_or_nullcontext("Draft"):
self._draft_token_ids = self.propose_draft_token_ids(
scheduler_output,
sampled_token_ids,
self.input_batch.sampling_metadata,
hidden_states,
sample_hidden_states,
aux_hidden_states,
spec_decode_metadata,
spec_decode_common_attn_metadata,
)
use_padded_batch_for_eagle = (
self.speculative_config
and self.speculative_config.use_eagle()
and not self.speculative_config.disable_padded_drafter_batch
)
effective_drafter_max_model_len = self.max_model_len
if effective_drafter_max_model_len is None:
effective_drafter_max_model_len = self.model_config.max_model_len
if (
self.speculative_config
and self.speculative_config.draft_model_config is not None
and self.speculative_config.draft_model_config.max_model_len is not None
):
effective_drafter_max_model_len = (
self.speculative_config.draft_model_config.max_model_len
)
input_fits_in_drafter = spec_decode_common_attn_metadata and (
spec_decode_common_attn_metadata.max_seq_len
+ self.speculative_config.num_speculative_tokens
<= effective_drafter_max_model_len
)
if use_padded_batch_for_eagle and input_fits_in_drafter:
# EAGLE speculative decoding can use the GPU sampled tokens
# as inputs, and does not need to wait for bookkeeping to finish.
propose_draft_token_ids(sampler_output.sampled_token_ids)
with record_function_or_nullcontext("Bookkeep"):
(
num_nans_in_logits,
logprobs_lists,
valid_sampled_token_ids,
prompt_logprobs_dict,
req_ids_output_copy,
req_id_to_index_output_copy,
invalid_req_indices,
) = self._bookkeeping_sync(
scheduler_output,
sampler_output,
logits,
hidden_states,
num_scheduled_tokens,
spec_decode_metadata,
)
if (
self.speculative_config
and not use_padded_batch_for_eagle
and input_fits_in_drafter
):
# ngram and other speculative decoding methods use the sampled
# tokens on the CPU, so they are run after bookkeeping.
propose_draft_token_ids(valid_sampled_token_ids)
with record_function_or_nullcontext("EPLB"):
self.eplb_step()
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=[],
kv_connector_output=kv_connector_output,
num_nans_in_logits=num_nans_in_logits,
)
if not self.use_async_scheduling:
return output
async_output = AsyncGPUModelRunnerOutput(
model_runner_output=output,
sampled_token_ids=sampler_output.sampled_token_ids,
invalid_req_indices=invalid_req_indices,
async_output_copy_stream=self.async_output_copy_stream,
)
# Save ref of sampled_token_ids CPU tensor if the batch contains
# any requests with sampling params that that require output ids.
self.input_batch.set_async_sampled_token_ids(
async_output.sampled_token_ids_cpu,
async_output.async_copy_ready_event,
)
return async_output
def take_draft_token_ids(self) -> DraftTokenIds | None:
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 propose_draft_token_ids(
self,
scheduler_output: "SchedulerOutput",
sampled_token_ids: torch.Tensor | list[list[int]],
sampling_metadata: SamplingMetadata,
hidden_states: torch.Tensor,
sample_hidden_states: torch.Tensor,
aux_hidden_states: list[torch.Tensor] | None,
spec_decode_metadata: SpecDecodeMetadata | None,
common_attn_metadata: CommonAttentionMetadata,
) -> list[list[int]] | torch.Tensor:
num_scheduled_tokens = scheduler_output.total_num_scheduled_tokens
if self.speculative_config.method == "ngram":
assert isinstance(sampled_token_ids, list)
assert isinstance(self.drafter, NgramProposer)
draft_token_ids = self.drafter.propose(
sampled_token_ids,
self.input_batch.req_ids,
self.input_batch.num_tokens_no_spec,
self.input_batch.token_ids_cpu,
self.input_batch.spec_decode_unsupported_reqs,
)
elif self.speculative_config.method == "medusa":
assert isinstance(sampled_token_ids, list)
assert isinstance(self.drafter, MedusaProposer)
if sample_hidden_states.shape[0] == len(sampled_token_ids):
# The input to the target model does not include draft tokens.
hidden_states = sample_hidden_states
else:
indices = []
offset = 0
assert spec_decode_metadata is not None
for num_draft, tokens in zip(
spec_decode_metadata.num_draft_tokens, sampled_token_ids
):
indices.append(offset + len(tokens) - 1)
offset += num_draft + 1
indices = torch.tensor(indices, device=self.device)
hidden_states = sample_hidden_states[indices]
draft_token_ids = self.drafter.propose(
target_hidden_states=hidden_states,
sampling_metadata=sampling_metadata,
)
elif self.speculative_config.use_eagle():
assert isinstance(self.drafter, EagleProposer)
if self.speculative_config.disable_padded_drafter_batch:
# When padded-batch is disabled, the sampled_token_ids should be
# the cpu-side list[list[int]] of valid sampled tokens for each
# request, with invalid requests having empty lists.
assert isinstance(sampled_token_ids, list), (
"sampled_token_ids should be a python list when"
"padded-batch is disabled."
)
next_token_ids = self.drafter.prepare_next_token_ids_cpu(
sampled_token_ids,
self.requests,
self.input_batch,
scheduler_output.num_scheduled_tokens,
)
else:
# When using padded-batch, the sampled_token_ids should be
# the gpu tensor of sampled tokens for each request, of shape
# (num_reqs, num_spec_tokens + 1) with rejected tokens having
# value -1.
assert isinstance(sampled_token_ids, torch.Tensor), (
"sampled_token_ids should be a torch.Tensor when"
"padded-batch is enabled."
)
next_token_ids, valid_sampled_tokens_count = (
self.drafter.prepare_next_token_ids_padded(
common_attn_metadata,
sampled_token_ids,
self.requests,
self.input_batch,
self.discard_request_indices.gpu,
self.num_discarded_requests,
)
)
if spec_decode_metadata is None:
token_indices_to_sample = None
# input_ids can be None for multimodal models.
target_token_ids = self.input_ids.gpu[:num_scheduled_tokens]
target_positions = self._get_positions(num_scheduled_tokens)
if self.use_aux_hidden_state_outputs:
assert aux_hidden_states is not None
target_hidden_states = torch.cat(
[h[:num_scheduled_tokens] for h in aux_hidden_states], dim=-1
)
else:
target_hidden_states = hidden_states[:num_scheduled_tokens]
else:
if self.speculative_config.disable_padded_drafter_batch:
token_indices_to_sample = None
common_attn_metadata, token_indices = self.drafter.prepare_inputs(
common_attn_metadata,
sampled_token_ids,
spec_decode_metadata.num_draft_tokens,
)
else:
common_attn_metadata, token_indices, token_indices_to_sample = (
self.drafter.prepare_inputs_padded(
common_attn_metadata,
spec_decode_metadata,
valid_sampled_tokens_count,
)
)
target_token_ids = self.input_ids.gpu[token_indices]
target_positions = self._get_positions(token_indices)
if self.use_aux_hidden_state_outputs:
assert aux_hidden_states is not None
target_hidden_states = torch.cat(
[h[token_indices] for h in aux_hidden_states], dim=-1
)
else:
target_hidden_states = hidden_states[token_indices]
if self.supports_mm_inputs:
mm_embed_inputs = self._gather_mm_embeddings(
scheduler_output,
shift_computed_tokens=1,
)
else:
mm_embed_inputs = None
draft_token_ids = self.drafter.propose(
target_token_ids=target_token_ids,
target_positions=target_positions,
target_hidden_states=target_hidden_states,
next_token_ids=next_token_ids,
last_token_indices=token_indices_to_sample,
sampling_metadata=sampling_metadata,
common_attn_metadata=common_attn_metadata,
mm_embed_inputs=mm_embed_inputs,
)
return draft_token_ids
def update_config(self, overrides: dict[str, Any]) -> None:
allowed_config_names = {"load_config", "model_config"}
for config_name, config_overrides in overrides.items():
assert config_name in allowed_config_names, (
f"Config `{config_name}` not supported. "
f"Allowed configs: {allowed_config_names}"
)
config = getattr(self, config_name)
new_config = update_config(config, config_overrides)
setattr(self, config_name, new_config)
def load_model(self, eep_scale_up: bool = False) -> None:
"""
Args:
eep_scale_up: the model loading is for elastic EP scale up.
"""
logger.info("Starting to load model %s...", self.model_config.model)
if eep_scale_up:
from vllm.distributed.parallel_state import get_ep_group
num_local_physical_experts = torch.empty(1, dtype=torch.int32, device="cpu")
torch.distributed.broadcast(
num_local_physical_experts, group=get_ep_group().cpu_group, group_src=0
)
num_local_physical_experts = int(num_local_physical_experts.item())
new_ep_size = get_ep_group().world_size
global_expert_load, old_global_expert_indices = EplbState.recv_state()
num_logical_experts = global_expert_load.shape[1]
self.parallel_config.eplb_config.num_redundant_experts = (
num_local_physical_experts * new_ep_size - num_logical_experts
)
assert old_global_expert_indices.shape[1] % num_local_physical_experts == 0
old_ep_size = (
old_global_expert_indices.shape[1] // num_local_physical_experts
)
rank_mapping = {
old_ep_rank: old_ep_rank for old_ep_rank in range(old_ep_size)
}
else:
global_expert_load = None
old_global_expert_indices = None
rank_mapping = None
with DeviceMemoryProfiler() as m:
time_before_load = time.perf_counter()
model_loader = get_model_loader(self.load_config)
self.model = model_loader.load_model(
vllm_config=self.vllm_config, model_config=self.model_config
)
if self.lora_config:
self.model = self.load_lora_model(
self.model, self.vllm_config, self.device
)
if hasattr(self, "drafter"):
logger.info("Loading drafter model...")
self.drafter.load_model(self.model)
if self.use_aux_hidden_state_outputs:
if not supports_eagle3(self.get_model()):
raise RuntimeError(
"Model does not support EAGLE3 interface but "
"aux_hidden_state_outputs was requested"
)
# Try to get auxiliary layers from speculative config,
# otherwise use model's default layers
aux_layers = self._get_eagle3_aux_layers_from_config()
if aux_layers:
logger.info(
"Using auxiliary layers from speculative config: %s",
aux_layers,
)
else:
aux_layers = self.model.get_eagle3_aux_hidden_state_layers()
self.model.set_aux_hidden_state_layers(aux_layers)
time_after_load = time.perf_counter()
self.model_memory_usage = m.consumed_memory
logger.info(
"Model loading took %.4f GiB and %.6f seconds",
self.model_memory_usage / GiB_bytes,
time_after_load - time_before_load,
)
prepare_communication_buffer_for_model(self.model)
self.is_multimodal_pruning_enabled = (
supports_multimodal_pruning(self.get_model())
and self.model_config.multimodal_config.is_multimodal_pruning_enabled()
)
if is_mixture_of_experts(self.model) and self.parallel_config.enable_eplb:
logger.info("EPLB is enabled for model %s.", self.model_config.model)
self.eplb_state = EplbState.build(
self.model,
self.device,
self.parallel_config,
global_expert_load,
old_global_expert_indices,
rank_mapping,
)
if (
self.vllm_config.compilation_config.mode
== CompilationMode.STOCK_TORCH_COMPILE
and supports_dynamo()
):
backend = self.vllm_config.compilation_config.init_backend(self.vllm_config)
compilation_counter.stock_torch_compile_count += 1
self.model.compile(fullgraph=True, backend=backend)
return
# for other compilation modes, cudagraph behavior is controlled by
# CudagraphWraper and CudagraphDispatcher of vllm.
# wrap the model with full cudagraph wrapper if needed.
if (
self.compilation_config.cudagraph_mode.has_full_cudagraphs()
and not self.parallel_config.enable_dbo
):
self.model = CUDAGraphWrapper(
self.model, self.vllm_config, runtime_mode=CUDAGraphMode.FULL
)
elif self.parallel_config.enable_dbo:
if self.compilation_config.cudagraph_mode.has_full_cudagraphs():
self.model = UBatchWrapper(
self.model, self.vllm_config, CUDAGraphMode.FULL, self.device
)
else:
self.model = UBatchWrapper(
self.model, self.vllm_config, CUDAGraphMode.NONE, self.device
)
def _get_eagle3_aux_layers_from_config(self) -> tuple[int, ...] | None:
"""Extract Eagle3 auxiliary layer indices from speculative config.
These indices specify which hidden states from the base model should
be used as auxiliary inputs for the Eagle3 drafter model during
speculative decoding.
Returns:
Tuple of layer indices if found in draft model config,
None otherwise.
"""
if not (self.speculative_config and self.speculative_config.draft_model_config):
return None
hf_config = self.speculative_config.draft_model_config.hf_config
if not hasattr(hf_config, "eagle_aux_hidden_state_layer_ids"):
return None
layer_ids = hf_config.eagle_aux_hidden_state_layer_ids
if layer_ids and isinstance(layer_ids, (list, tuple)):
return tuple(layer_ids)
return None
def reload_weights(self) -> None:
assert getattr(self, "model", None) is not None, (
"Cannot reload weights before model is loaded."
)
model_loader = get_model_loader(self.load_config)
logger.info("Reloading weights inplace...")
model_loader.load_weights(self.get_model(), model_config=self.model_config)
def save_tensorized_model(
self,
tensorizer_config: "TensorizerConfig",
) -> None:
TensorizerLoader.save_model(
self.get_model(),
tensorizer_config=tensorizer_config,
model_config=self.model_config,
)
def _get_prompt_logprobs_dict(
self,
hidden_states: torch.Tensor,
num_scheduled_tokens: dict[str, int],
) -> dict[str, LogprobsTensors | None]:
num_prompt_logprobs_dict = self.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, LogprobsTensors | None] = {}
# Since prompt logprobs are a rare feature, prioritize simple,
# maintainable loop over optimal performance.
completed_prefill_reqs = []
for req_id, num_prompt_logprobs in num_prompt_logprobs_dict.items():
num_tokens = num_scheduled_tokens[req_id]
# Get metadata for this request.
request = self.requests[req_id]
if request.prompt_token_ids is None:
# Prompt logprobs is incompatible with prompt embeddings
continue
num_prompt_tokens = len(request.prompt_token_ids)
prompt_token_ids = 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 GPU->CPU async.
chunk_slice = slice(start_idx, start_idx + num_logits)
logprobs_tensors.logprob_token_ids[chunk_slice].copy_(
token_ids, non_blocking=True
)
logprobs_tensors.logprobs[chunk_slice].copy_(logprobs, non_blocking=True)
logprobs_tensors.selected_token_ranks[chunk_slice].copy_(
ranks, non_blocking=True
)
# Remove requests that have completed prefill from the batch
# num_prompt_logprobs_dict.
for req_id in completed_prefill_reqs:
del num_prompt_logprobs_dict[req_id]
del in_progress_dict[req_id]
# Must synchronize the non-blocking GPU->CPU transfers.
if prompt_logprobs_dict:
self._sync_device()
return prompt_logprobs_dict
def _get_nans_in_logits(
self,
logits: torch.Tensor | None,
) -> dict[str, int]:
try:
if logits is None:
return {req_id: 0 for req_id in self.input_batch.req_ids}
num_nans_in_logits = {}
num_nans_for_index = logits.isnan().sum(dim=-1).cpu().numpy()
for req_id in self.input_batch.req_ids:
req_index = self.input_batch.req_id_to_index[req_id]
num_nans_in_logits[req_id] = (
int(num_nans_for_index[req_index])
if num_nans_for_index is not None and req_index < logits.shape[0]
else 0
)
return num_nans_in_logits
except IndexError:
return {}
@contextmanager
def maybe_randomize_inputs(self, input_ids: torch.Tensor):
"""
Randomize input_ids if VLLM_RANDOMIZE_DP_DUMMY_INPUTS is set.
This is to help balance expert-selection
- during profile_run
- during DP rank dummy run
"""
dp_size = self.vllm_config.parallel_config.data_parallel_size
randomize_inputs = envs.VLLM_RANDOMIZE_DP_DUMMY_INPUTS and dp_size > 1
if not randomize_inputs:
yield
else:
import functools
@functools.cache
def rand_input_ids() -> torch.Tensor:
return torch.randint_like(
self.input_ids.gpu,
low=0,
high=self.model_config.get_vocab_size(),
dtype=input_ids.dtype,
)
logger.debug_once("Randomizing dummy data for DP Rank")
input_ids.copy_(rand_input_ids()[: input_ids.size(0)], non_blocking=True)
yield
input_ids.fill_(0)
def _get_mm_dummy_batch(
self,
modality: str,
max_items_per_batch: int,
) -> BatchedTensorInputs:
"""Dummy data for profiling and precompiling multimodal models."""
assert self.mm_budget is not None
dummy_decoder_data = self.mm_registry.get_decoder_dummy_data(
model_config=self.model_config,
seq_len=self.max_model_len,
mm_counts={modality: 1},
cache=self.mm_budget.cache,
)
dummy_mm_data = dummy_decoder_data.multi_modal_data
# Result in the maximum GPU consumption of the model
dummy_mm_item = dummy_mm_data[modality][0]
dummy_mm_items = [dummy_mm_item] * max_items_per_batch
model = cast(SupportsMultiModal, self.model)
return next(
mm_kwargs_group
for _, _, mm_kwargs_group in group_mm_kwargs_by_modality(
dummy_mm_items,
device=self.device,
pin_memory=self.pin_memory,
merge_by_field_config=model.merge_by_field_config,
)
)
@torch.inference_mode()
def _dummy_run(
self,
num_tokens: int,
cudagraph_runtime_mode: CUDAGraphMode | None = None,
force_attention: bool = False,
uniform_decode: bool = False,
allow_microbatching: bool = True,
skip_eplb: bool = False,
is_profile: bool = False,
create_mixed_batch: bool = False,
remove_lora: bool = True,
activate_lora: bool = False,
) -> tuple[torch.Tensor, torch.Tensor]:
"""
Run a dummy forward pass to warm up/profile run or capture the
CUDA graph for the model.
Args:
num_tokens: Number of tokens to run the dummy forward pass.
cudagraph_runtime_mode: used to control the behavior.
- if not set will determine the cudagraph mode based on using
the self.cudagraph_dispatcher.
- CUDAGraphMode.NONE: No cudagraph, for warm up and profile run
- CUDAGraphMode.PIECEWISE: Piecewise cudagraph.
- CUDAGraphMode.FULL: Full cudagraph, attention metadata is
needed.
force_attention: If True, always create attention metadata. Used to
warm up attention backend when mode is NONE.
uniform_decode: If True, the batch is a uniform decode batch.
skip_eplb: If True, skip EPLB state update.
is_profile: If True, this is a profile run.
create_mixed_batch: If True, create a mixed batch with both decode
(1 token) and prefill (multiple tokens) requests.
remove_lora: If False, dummy LoRAs are not destroyed after the run
activate_lora: If False, dummy_run is performed without LoRAs.
"""
assert (
cudagraph_runtime_mode is None
or cudagraph_runtime_mode.valid_runtime_modes()
)
# If cudagraph_mode.decode_mode() == FULL and
# cudagraph_mode.separate_routine(). This means that we are using
# different graphs and/or modes for mixed prefill-decode batches vs.
# uniform decode batches. A uniform decode batch means that all
# requests have identical query length, except a potential virtual
# request (shorter) in the batch account for padding.
# Uniform decode batch could either be common pure decode, where
# max_query_len == 1, or speculative decode, where
# max_query_len == 1 + num_spec_decode_tokens.
# When setting max_query_len = 1, we switch to and capture the optimized
# routine of FA2 for pure decode, i.e., Flashdecode + an optimization
# for GQA/MQA.
max_query_len = self.uniform_decode_query_len if uniform_decode else num_tokens
# Set num_scheduled_tokens based on num_tokens and max_num_seqs
# for dummy run with LoRA so that the num_reqs collectively
# has num_tokens in total.
assert num_tokens <= self.scheduler_config.max_num_batched_tokens
max_num_reqs = self.scheduler_config.max_num_seqs
if create_mixed_batch:
assert not uniform_decode
# Create mixed batch:
# first half decode tokens, second half one prefill
num_decode_tokens = min(max_num_reqs - 1, num_tokens // 2)
num_prefill_tokens = num_tokens - num_decode_tokens
num_reqs = num_decode_tokens + 1
# Create decode requests (1 token each) followed by prefill request
num_scheduled_tokens_list = [1] * num_decode_tokens + [num_prefill_tokens]
# Note: Overriding max_query_len to be the prefill tokens
max_query_len = num_prefill_tokens
elif uniform_decode:
assert not create_mixed_batch
num_reqs = min(max_num_reqs, cdiv(num_tokens, max_query_len))
num_scheduled_tokens_list = [max_query_len] * num_reqs
if num_tokens % max_query_len != 0:
num_scheduled_tokens_list[-1] = num_tokens % max_query_len
else:
num_reqs = min(num_tokens, max_num_reqs)
min_tokens_per_req = num_tokens // num_reqs
num_scheduled_tokens_list = [min_tokens_per_req] * num_reqs
num_scheduled_tokens_list[-1] += num_tokens % num_reqs
assert sum(num_scheduled_tokens_list) == num_tokens
assert len(num_scheduled_tokens_list) == num_reqs
num_scheduled_tokens = np.array(num_scheduled_tokens_list, dtype=np.int32)
total_num_scheduled_tokens = int(num_scheduled_tokens.sum())
# Disable DP padding when running eager
allow_dp_padding = self.compilation_config.cudagraph_mode != CUDAGraphMode.NONE
# We currently only microbatch if the number of tokens is
# over a certain threshold.
ubatch_slices, num_tokens_across_dp = coordinate_batch_across_dp(
num_tokens_unpadded=total_num_scheduled_tokens,
parallel_config=self.vllm_config.parallel_config,
allow_microbatching=allow_microbatching,
allow_dp_padding=allow_dp_padding,
num_tokens_padded=total_num_scheduled_tokens,
uniform_decode=uniform_decode,
num_scheduled_tokens_per_request=num_scheduled_tokens,
)
num_tokens_after_padding = num_tokens
if num_tokens_across_dp is not None:
dp_rank = self.parallel_config.data_parallel_rank
num_tokens_after_padding = int(num_tokens_across_dp[dp_rank])
attn_metadata: PerLayerAttnMetadata | None = None
# If force_attention is True, we always capture attention. Otherwise,
# it only happens for cudagraph_runtime_mode=FULL.
if force_attention or cudagraph_runtime_mode == CUDAGraphMode.FULL:
attn_metadata = {}
if ubatch_slices is not None:
attn_metadata = [dict() for _ in range(len(ubatch_slices))]
if create_mixed_batch:
# In the mixed batch mode (used for FI warmup), we use
# shorter sequence lengths to run faster.
# TODO(luka) better system for describing dummy batches
seq_lens = [1] * num_decode_tokens + [num_prefill_tokens + 1]
else:
seq_lens = max_query_len
self.seq_lens.np[:num_reqs] = seq_lens
self.seq_lens.np[num_reqs:] = 0
self.seq_lens.copy_to_gpu()
cum_num_tokens, _ = self._get_cumsum_and_arange(num_scheduled_tokens)
self.query_start_loc.np[1 : num_reqs + 1] = cum_num_tokens
self.query_start_loc.copy_to_gpu()
for kv_cache_group_id, kv_cache_group_spec in enumerate(
self.kv_cache_config.kv_cache_groups
):
common_attn_metadata = CommonAttentionMetadata(
query_start_loc=self.query_start_loc.gpu[: num_reqs + 1],
query_start_loc_cpu=self.query_start_loc.cpu[: num_reqs + 1],
seq_lens=self.seq_lens.gpu[:num_reqs],
seq_lens_cpu=self.seq_lens.cpu[:num_reqs],
num_computed_tokens_cpu=self.input_batch.num_computed_tokens_cpu_tensor[
:num_reqs
],
num_reqs=num_reqs,
num_actual_tokens=num_tokens,
max_query_len=max_query_len,
max_seq_len=self.max_model_len,
block_table_tensor=self.input_batch.block_table[
kv_cache_group_id
].get_device_tensor(num_reqs),
slot_mapping=self.input_batch.block_table[
kv_cache_group_id
].slot_mapping.gpu[:num_tokens],
causal=True,
dcp_local_seq_lens=self.dcp_local_seq_lens.gpu[:num_reqs]
if self.dcp_world_size > 1
else None,
)
for attn_group in self.attn_groups[kv_cache_group_id]:
if ubatch_slices is not None:
common_attn_metadata_list = split_attn_metadata(
ubatch_slices, common_attn_metadata
)
for ubid, common_attn_metadata in enumerate(
common_attn_metadata_list
):
assert common_attn_metadata.max_query_len == 1
attn_metadata_i = attn_group.get_metadata_builder(
ubatch_id=ubid
).build_for_cudagraph_capture(common_attn_metadata)
for layer_name in attn_group.layer_names:
assert type(attn_metadata) is list
attn_metadata[ubid][layer_name] = attn_metadata_i
else:
assert type(attn_metadata) is dict
metadata_builder = attn_group.get_metadata_builder()
attn_metadata_i = metadata_builder.build_for_cudagraph_capture(
common_attn_metadata
)
for layer_name in attn_group.layer_names:
attn_metadata[layer_name] = attn_metadata_i
with self.maybe_dummy_run_with_lora(
self.lora_config, num_scheduled_tokens, activate_lora, remove_lora
):
# Make sure padding doesn't exceed max_num_tokens
assert num_tokens_after_padding <= self.max_num_tokens
model_kwargs = self._init_model_kwargs(num_tokens_after_padding)
if self.supports_mm_inputs and not self.model_config.is_encoder_decoder:
input_ids = None
inputs_embeds = self.inputs_embeds.gpu[:num_tokens_after_padding]
model_kwargs = {
**model_kwargs,
**self._dummy_mm_kwargs(num_reqs),
}
elif self.enable_prompt_embeds:
input_ids = None
inputs_embeds = self.inputs_embeds.gpu[:num_tokens_after_padding]
model_kwargs = self._init_model_kwargs(num_tokens_after_padding)
else:
input_ids = self.input_ids.gpu[:num_tokens_after_padding]
inputs_embeds = None
if self.uses_mrope:
positions = self.mrope_positions.gpu[:, :num_tokens_after_padding]
else:
positions = self.positions.gpu[:num_tokens_after_padding]
if get_pp_group().is_first_rank:
intermediate_tensors = None
else:
if self.intermediate_tensors is None:
self.intermediate_tensors = (
self.model.make_empty_intermediate_tensors(
batch_size=self.max_num_tokens,
dtype=self.model_config.dtype,
device=self.device,
)
)
intermediate_tensors = self.sync_and_slice_intermediate_tensors(
num_tokens_after_padding, None, False
)
# filter out the valid batch descriptor
_cg_mode, batch_descriptor = (
self.cudagraph_dispatcher.dispatch(
BatchDescriptor(
num_tokens=num_tokens_after_padding,
uniform_decode=uniform_decode,
has_lora=activate_lora and self.lora_config is not None,
)
)
if not is_profile
else (CUDAGraphMode.NONE, None)
)
if cudagraph_runtime_mode is not None:
# we allow forcing NONE when the dispatcher disagrees to support
# warm ups for cudagraph capture
assert (
cudagraph_runtime_mode == CUDAGraphMode.NONE
or cudagraph_runtime_mode == _cg_mode
), (
f"Cudagraph runtime mode mismatch at dummy_run. "
f"Expected {_cg_mode}, but got {cudagraph_runtime_mode}."
)
else:
cudagraph_runtime_mode = _cg_mode
if ubatch_slices is not None:
# Adjust values to reflect a single ubatch.
# TODO(sage,lucas): this is cruft that should be addressed in
# the padding refactor.
num_tokens_after_padding = ubatch_slices[0].num_tokens
if num_tokens_across_dp is not None:
num_tokens_across_dp[:] = num_tokens_after_padding
with (
self.maybe_randomize_inputs(input_ids),
set_forward_context(
attn_metadata,
self.vllm_config,
num_tokens=num_tokens_after_padding,
num_tokens_across_dp=num_tokens_across_dp,
cudagraph_runtime_mode=cudagraph_runtime_mode,
batch_descriptor=batch_descriptor,
ubatch_slices=ubatch_slices,
),
):
outputs = self.model(
input_ids=input_ids,
positions=positions,
intermediate_tensors=intermediate_tensors,
inputs_embeds=inputs_embeds,
**model_kwargs,
)
if self.use_aux_hidden_state_outputs:
hidden_states, _ = outputs
else:
hidden_states = outputs
if self.speculative_config and self.speculative_config.use_eagle():
assert isinstance(self.drafter, EagleProposer)
use_cudagraphs = (
cudagraph_runtime_mode == CUDAGraphMode.PIECEWISE
and not self.speculative_config.enforce_eager
)
self.drafter.dummy_run(num_tokens, use_cudagraphs=use_cudagraphs)
# This is necessary to avoid blocking DP.
# For dummy runs, we typically skip EPLB since we don't have any real
# requests to process.
# However, in DP settings, there may be cases when some DP ranks do
# not have any requests to process, so they're executing dummy batches.
# In such cases, we still have to trigger EPLB to make sure
# ranks execute the rearrangement in synchronization.
if not skip_eplb:
self.eplb_step(is_dummy=True, is_profile=is_profile)
logit_indices = np.cumsum(num_scheduled_tokens) - 1
return hidden_states, hidden_states[logit_indices]
@torch.inference_mode()
def _dummy_sampler_run(
self,
hidden_states: torch.Tensor,
) -> torch.Tensor:
# The dummy hidden states may contain special values,
# like `inf` or `nan`.
# To avoid breaking the sampler, we use a random tensor here instead.
hidden_states = torch.rand_like(hidden_states)
logits = self.model.compute_logits(hidden_states)
num_reqs = logits.size(0)
dummy_tensors = lambda v: torch.full((num_reqs,), v, device=self.device)
dummy_metadata = SamplingMetadata(
temperature=dummy_tensors(0.5),
all_greedy=False,
all_random=False,
top_p=dummy_tensors(0.9),
top_k=dummy_tensors(logits.size(1) - 1),
generators={},
max_num_logprobs=None,
no_penalties=True,
prompt_token_ids=None,
frequency_penalties=dummy_tensors(0.1),
presence_penalties=dummy_tensors(0.1),
repetition_penalties=dummy_tensors(0.1),
output_token_ids=[[] for _ in range(num_reqs)],
spec_token_ids=[[] for _ in range(num_reqs)],
allowed_token_ids_mask=None,
bad_words_token_ids={},
logitsprocs=LogitsProcessors(),
)
try:
sampler_output = self.sampler(
logits=logits, sampling_metadata=dummy_metadata
)
except RuntimeError as e:
if "out of memory" in str(e):
raise RuntimeError(
"CUDA out of memory occurred when warming up sampler with "
f"{num_reqs} dummy requests. Please try lowering "
"`max_num_seqs` or `gpu_memory_utilization` when "
"initializing the engine."
) from e
else:
raise e
if self.speculative_config:
draft_token_ids = [[0] for _ in range(num_reqs)]
dummy_spec_decode_metadata = SpecDecodeMetadata.make_dummy(
draft_token_ids, self.device
)
num_tokens = sum(len(ids) for ids in draft_token_ids)
# draft_probs = torch.randn(
# num_tokens, logits.shape[-1], device=self.device,
# dtype=logits.dtype)
draft_probs = None
logits = torch.randn(
num_tokens + num_reqs,
logits.shape[-1],
device=self.device,
dtype=logits.dtype,
)
self.rejection_sampler(
dummy_spec_decode_metadata,
draft_probs,
logits,
dummy_metadata,
)
return sampler_output
def _dummy_pooler_run_task(
self,
hidden_states: torch.Tensor,
task: PoolingTask,
) -> PoolerOutput:
num_tokens = hidden_states.shape[0]
max_num_reqs = self.scheduler_config.max_num_seqs
num_reqs = min(num_tokens, max_num_reqs)
min_tokens_per_req = num_tokens // num_reqs
num_scheduled_tokens_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_prompt_lens = torch.tensor(
num_scheduled_tokens_list,
device="cpu",
)
dummy_token_ids = torch.zeros(
(num_reqs, req_num_tokens), dtype=torch.int32, device=self.device
)
model = cast(VllmModelForPooling, self.get_model())
dummy_pooling_params = PoolingParams(task=task)
dummy_pooling_params.verify(task=task, model_config=self.model_config)
to_update = model.pooler.get_pooling_updates(task)
to_update.apply(dummy_pooling_params)
dummy_metadata = PoolingMetadata(
prompt_lens=dummy_prompt_lens,
prompt_token_ids=dummy_token_ids,
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
supported_pooling_tasks = self.get_supported_pooling_tasks()
if not supported_pooling_tasks:
if self.scheduler_config.chunked_prefill_enabled:
raise RuntimeError(
f"Model {self.model_config.model} does not support "
"any pooling tasks with chunked prefill enabled. "
"Please add --no-enable-chunked-prefill to your "
"config or CLI args. See "
"https://docs.vllm.ai/en/latest/models/pooling_models.html "
"to learn more."
)
else:
raise RuntimeError(
f"Model {self.model_config.model} does not support "
"any pooling tasks. See "
"https://docs.vllm.ai/en/latest/models/pooling_models.html "
"to learn more."
)
output_size = dict[PoolingTask, float]()
for task in supported_pooling_tasks:
# Run a full batch with each task to ensure none of them OOMs
output = self._dummy_pooler_run_task(hidden_states, task)
output_size[task] = sum(o.nbytes for o in output)
del output # Allow GC
max_task = max(output_size.items(), key=lambda x: x[1])[0]
return self._dummy_pooler_run_task(hidden_states, max_task)
def profile_run(self) -> None:
# Profile with multimodal encoder & encoder cache.
if self.supports_mm_inputs:
if self.model_config.multimodal_config.skip_mm_profiling:
logger.info(
"Skipping memory profiling for multimodal encoder and "
"encoder cache."
)
else:
mm_budget = self.mm_budget
assert mm_budget is not None
if (encoder_budget := mm_budget.get_encoder_budget()) > 0:
# NOTE: Currently model is profiled with a single non-text
# modality with the max possible input tokens even when
# it supports multiple.
dummy_modality = mm_budget.get_modality_with_max_tokens()
max_mm_items_per_batch = mm_budget.max_items_per_batch_by_modality[
dummy_modality
]
logger.info(
"Encoder cache will be initialized with a budget of "
"%s tokens, and profiled with %s %s items of the "
"maximum feature size.",
encoder_budget,
max_mm_items_per_batch,
dummy_modality,
)
# Create dummy batch of multimodal inputs.
batched_dummy_mm_inputs = self._get_mm_dummy_batch(
dummy_modality,
max_mm_items_per_batch,
)
# Run multimodal encoder.
dummy_encoder_outputs = self.model.get_multimodal_embeddings(
**batched_dummy_mm_inputs
)
sanity_check_mm_encoder_outputs(
dummy_encoder_outputs,
expected_num_items=max_mm_items_per_batch,
)
# NOTE: This happens when encoder cache needs to store
# the embeddings that encoder outputs are scattered onto.
# In this case we create dummy embeddings of size
# (encode_budget, hidden_size) and scatter encoder
# output into it.
encoder_output_shape = dummy_encoder_outputs[0].shape
if encoder_output_shape[0] < encoder_budget:
expanded_outputs = []
for output in dummy_encoder_outputs:
expanded = output.new_zeros(
(encoder_budget, encoder_output_shape[-1])
)
num_tokens = output.shape[0]
expanded[:num_tokens].copy_(output)
expanded_outputs.append(expanded)
dummy_encoder_outputs = expanded_outputs
# Cache the dummy encoder outputs.
self.encoder_cache["tmp"] = dict(enumerate(dummy_encoder_outputs))
# Add `is_profile` here to pre-allocate communication buffers
hidden_states, last_hidden_states = self._dummy_run(
self.max_num_tokens, is_profile=True
)
if get_pp_group().is_last_rank:
if self.is_pooling_model:
output = self._dummy_pooler_run(hidden_states)
else:
output = self._dummy_sampler_run(last_hidden_states)
else:
output = None
self._sync_device()
del hidden_states, output
self.encoder_cache.clear()
gc.collect()
def capture_model(self) -> int:
if self.compilation_config.cudagraph_mode == CUDAGraphMode.NONE:
logger.warning(
"Skipping CUDA graph capture. To turn on CUDA graph capture, "
"ensure `cudagraph_mode` was not manually set to `NONE`"
)
return 0
else:
self.initialize_cudagraph_capture()
compilation_counter.num_gpu_runner_capture_triggers += 1
start_time = time.perf_counter()
@contextmanager
def freeze_gc():
# Optimize garbage collection during CUDA graph capture.
# Clean up, then freeze all remaining objects from being included
# in future collections.
gc.collect()
should_freeze = not envs.VLLM_ENABLE_CUDAGRAPH_GC
if should_freeze:
gc.freeze()
try:
yield
finally:
if should_freeze:
gc.unfreeze()
gc.collect()
# Trigger CUDA graph capture for specific shapes.
# Capture the large shapes first so that the smaller shapes
# can reuse the memory pool allocated for the large shapes.
set_cudagraph_capturing_enabled(True)
with freeze_gc(), graph_capture(device=self.device):
start_free_gpu_memory = torch.cuda.mem_get_info()[0]
cudagraph_mode = self.compilation_config.cudagraph_mode
assert cudagraph_mode is not None
if self.lora_config:
if self.compilation_config.cudagraph_specialize_lora:
lora_cases = [True, False]
else:
lora_cases = [True]
else:
lora_cases = [False]
if cudagraph_mode.mixed_mode() != CUDAGraphMode.NONE:
cudagraph_runtime_mode = cudagraph_mode.mixed_mode()
compilation_cases = list(
product(reversed(self.cudagraph_batch_sizes), lora_cases)
)
self._capture_cudagraphs(
compilation_cases,
cudagraph_runtime_mode=cudagraph_runtime_mode,
uniform_decode=False,
)
# Capture full cudagraph for uniform decode batches if we
# don't already have full mixed prefill-decode cudagraphs.
if (
cudagraph_mode.decode_mode() == CUDAGraphMode.FULL
and cudagraph_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.cudagraph_batch_sizes
if max_num_tokens >= x >= self.uniform_decode_query_len
]
compilation_cases_decode = list(
product(reversed(decode_cudagraph_batch_sizes), lora_cases)
)
self._capture_cudagraphs(
compilation_cases=compilation_cases_decode,
cudagraph_runtime_mode=CUDAGraphMode.FULL,
uniform_decode=True,
)
torch.cuda.synchronize()
end_free_gpu_memory = torch.cuda.mem_get_info()[0]
# Disable cudagraph capturing globally, so any unexpected cudagraph
# capturing will be detected and raise an error after here.
# Note: We don't put it into graph_capture context manager because
# we may do lazy capturing in future that still allows capturing
# after here.
set_cudagraph_capturing_enabled(False)
end_time = time.perf_counter()
elapsed_time = end_time - start_time
cuda_graph_size = start_free_gpu_memory - end_free_gpu_memory
# This usually takes 5~20 seconds.
logger.info(
"Graph capturing finished in %.0f secs, took %.2f GiB",
elapsed_time,
cuda_graph_size / (1 << 30),
)
return cuda_graph_size
def _capture_cudagraphs(
self,
compilation_cases: list[tuple[int, bool]],
cudagraph_runtime_mode: CUDAGraphMode,
uniform_decode: bool,
):
assert (
cudagraph_runtime_mode != CUDAGraphMode.NONE
and cudagraph_runtime_mode.valid_runtime_modes()
), f"Invalid cudagraph runtime mode: {cudagraph_runtime_mode}"
# Only rank 0 should print progress bar during capture
if is_global_first_rank():
compilation_cases = tqdm(
compilation_cases,
disable=not self.load_config.use_tqdm_on_load,
desc="Capturing CUDA graphs ({}, {})".format(
"decode" if uniform_decode else "mixed prefill-decode",
cudagraph_runtime_mode.name,
),
)
# We skip EPLB here since we don't want to record dummy metrics
for num_tokens, activate_lora in compilation_cases:
# We currently only capture ubatched graphs when its a FULL
# cudagraph, a uniform decode batch, and the number of tokens
# is above the threshold. Otherwise we just capture a non-ubatched
# version of the graph
allow_microbatching = (
self.parallel_config.enable_dbo
and cudagraph_runtime_mode == CUDAGraphMode.FULL
and uniform_decode
and check_ubatch_thresholds(
config=self.vllm_config.parallel_config,
num_tokens=num_tokens,
uniform_decode=uniform_decode,
)
)
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 = cudagraph_runtime_mode == CUDAGraphMode.FULL
self._dummy_run(
num_tokens,
cudagraph_runtime_mode=CUDAGraphMode.NONE,
force_attention=force_attention,
uniform_decode=uniform_decode,
allow_microbatching=allow_microbatching,
skip_eplb=True,
remove_lora=False,
activate_lora=activate_lora,
)
self._dummy_run(
num_tokens,
cudagraph_runtime_mode=cudagraph_runtime_mode,
uniform_decode=uniform_decode,
allow_microbatching=allow_microbatching,
skip_eplb=True,
remove_lora=False,
activate_lora=activate_lora,
)
self.maybe_remove_all_loras(self.lora_config)
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()
if layer_name in self.kv_sharing_fast_prefill_eligible_layers:
attn_backend = create_fast_prefill_custom_backend(
"FastPrefill",
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[AttentionGroupKey, list[str]],
) -> list[AttentionGroup]:
attn_groups: list[AttentionGroup] = []
for (attn_backend, kv_cache_spec), layer_names in attn_backends_map.items():
attn_group = AttentionGroup.create_with_metadata_builders(
attn_backend,
layer_names,
kv_cache_spec,
self.vllm_config,
self.device,
num_metadata_builders=1
if not self.parallel_config.enable_dbo
else 2,
)
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(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 initialize_cudagraph_capture(self) -> None:
"""
Resolve the cudagraph_mode when there are multiple attention
backends with potential conflicting CUDA graph support.
Then initialize the cudagraph_dispatcher based on the resolved
cudagraph_mode.
"""
min_cg_support = AttentionCGSupport.ALWAYS
min_cg_builder_name = None
for attn_group in self._attn_group_iterator():
builder = attn_group.get_metadata_builder()
if builder.cudagraph_support.value < min_cg_support.value:
min_cg_support = builder.cudagraph_support
min_cg_builder_name = builder.__class__.__name__
# Flexible resolve the cudagraph mode
cudagraph_mode = self.compilation_config.cudagraph_mode
# check cudagraph for mixed batch is supported
if (
cudagraph_mode.mixed_mode() == CUDAGraphMode.FULL
and min_cg_support != AttentionCGSupport.ALWAYS
):
msg = (
f"CUDAGraphMode.{cudagraph_mode.name} is not supported "
f"with {min_cg_builder_name} backend (support: "
f"{min_cg_support})"
)
if min_cg_support == AttentionCGSupport.NEVER:
# if not supported any full cudagraphs, just raise it.
msg += (
"; please try cudagraph_mode=PIECEWISE, and "
"make sure compilation mode is VLLM_COMPILE"
)
raise ValueError(msg)
# attempt to resolve the full cudagraph related mode
if self.compilation_config.splitting_ops_contain_attention():
msg += "; setting cudagraph_mode=FULL_AND_PIECEWISE"
cudagraph_mode = self.compilation_config.cudagraph_mode = (
CUDAGraphMode.FULL_AND_PIECEWISE
)
else:
msg += "; setting cudagraph_mode=FULL_DECODE_ONLY"
cudagraph_mode = self.compilation_config.cudagraph_mode = (
CUDAGraphMode.FULL_DECODE_ONLY
)
logger.warning(msg)
# check that if we are doing decode full-cudagraphs it is supported
if (
cudagraph_mode.decode_mode() == CUDAGraphMode.FULL
and min_cg_support == AttentionCGSupport.NEVER
):
msg = (
f"CUDAGraphMode.{cudagraph_mode.name} is not supported "
f"with {min_cg_builder_name} backend (support: "
f"{min_cg_support})"
)
if self.compilation_config.mode == CompilationMode.VLLM_COMPILE and (
self.compilation_config.splitting_ops_contain_attention()
or self.compilation_config.use_inductor_graph_partition
):
msg += (
"; setting cudagraph_mode=PIECEWISE because "
"attention is compiled piecewise"
)
cudagraph_mode = self.compilation_config.cudagraph_mode = (
CUDAGraphMode.PIECEWISE
)
else:
msg += (
"; setting cudagraph_mode=NONE because "
"attention is not compiled piecewise"
)
cudagraph_mode = self.compilation_config.cudagraph_mode = (
CUDAGraphMode.NONE
)
logger.warning(msg)
# check that if we are doing spec-decode + decode full-cudagraphs it is
# supported
if (
cudagraph_mode.decode_mode() == CUDAGraphMode.FULL
and self.uniform_decode_query_len > 1
and min_cg_support.value < AttentionCGSupport.UNIFORM_BATCH.value
):
msg = (
f"CUDAGraphMode.{cudagraph_mode.name} is not supported"
f" with spec-decode for attention backend "
f"{min_cg_builder_name} (support: {min_cg_support})"
)
if self.compilation_config.splitting_ops_contain_attention():
msg += "; setting cudagraph_mode=PIECEWISE"
cudagraph_mode = self.compilation_config.cudagraph_mode = (
CUDAGraphMode.PIECEWISE
)
else:
msg += "; setting cudagraph_mode=NONE"
cudagraph_mode = self.compilation_config.cudagraph_mode = (
CUDAGraphMode.NONE
)
logger.warning(msg)
# double check that we can support full cudagraph if they are requested
# even after automatic downgrades
if (
cudagraph_mode.has_full_cudagraphs()
and min_cg_support == AttentionCGSupport.NEVER
):
raise ValueError(
f"CUDAGraphMode.{cudagraph_mode.name} is not "
f"supported with {min_cg_builder_name} backend ("
f"support:{min_cg_support}) "
"; please try cudagraph_mode=PIECEWISE, "
"and make sure compilation mode is VLLM_COMPILE"
)
# Trigger cudagraph dispatching keys initialization here (after
# initializing attn backends).
self.cudagraph_dispatcher.initialize_cudagraph_keys(
self.compilation_config.cudagraph_mode, self.uniform_decode_query_len
)
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()
# 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 _find_compatible_block_sizes(
self,
kv_manager_block_size: int,
backend_cls: type[AttentionBackend],
return_all: bool = False,
) -> list[int]:
"""
Find compatible block sizes for a backend.
Args:
kv_manager_block_size: Physical block size of KV cache
backend_cls: Attention backend class
return_all: Return all compatible sizes if True, max size if False
Returns:
Compatible block size(s) based on return_all parameter
Raises:
ValueError: If no compatible block size found
"""
supported_block_size = backend_cls.get_supported_kernel_block_size()
compatible_sizes = []
for block_size in supported_block_size:
if isinstance(block_size, int):
if kv_manager_block_size % block_size == 0:
compatible_sizes.append(block_size)
elif (
isinstance(block_size, MultipleOf)
and kv_manager_block_size % block_size.base == 0
):
compatible_sizes.append(kv_manager_block_size)
if not compatible_sizes:
raise ValueError(f"No compatible block size for {kv_manager_block_size}")
return compatible_sizes if return_all else [max(compatible_sizes)]
def _select_common_block_size(
self, kv_manager_block_size: int, attn_groups: list[AttentionGroup]
) -> int:
"""
Select common block size for all backends.
Args:
kv_manager_block_size: Block size of KV cache
attn_groups: List of attention groups
Returns:
Block size supported by all backends,
prioritizing cache_config.block_size
Raises:
ValueError: If no common block size found
"""
all_backend_supports = []
for attn_group in attn_groups:
compatible_sizes = self._find_compatible_block_sizes(
kv_manager_block_size, attn_group.backend, return_all=True
)
supported_sizes = sorted(list(set(compatible_sizes)), reverse=True)
all_backend_supports.append(set(supported_sizes))
common_supported_sizes = set.intersection(*all_backend_supports)
if not common_supported_sizes:
error_msg = f"No common block size for {kv_manager_block_size}. "
for i, attn_group in enumerate(attn_groups):
supported = all_backend_supports[i]
error_msg += (
f"Backend {attn_group.backend} supports: {sorted(supported)}. "
)
raise ValueError(error_msg)
if self.cache_config.block_size in common_supported_sizes:
return self.cache_config.block_size
return max(common_supported_sizes)
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
if not isinstance(kv_cache_group.kv_cache_spec, EncoderOnlyAttentionSpec)
]
# Generate kernel_block_sizes that matches each block_size
kernel_block_sizes = self._prepare_kernel_block_sizes(kv_cache_config)
if block_sizes != [self.cache_config.block_size] or 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=max(self.max_model_len, self.max_encoder_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,
kernel_block_sizes=kernel_block_sizes,
is_spec_decode=bool(self.vllm_config.speculative_config),
logitsprocs=self.input_batch.logitsprocs,
logitsprocs_need_output_token_ids=self.input_batch.logitsprocs_need_output_token_ids,
is_pooling_model=self.is_pooling_model,
num_speculative_tokens=(
self.vllm_config.speculative_config.num_speculative_tokens
if self.vllm_config.speculative_config
else 0
),
)
def _allocate_kv_cache_tensors(
self, kv_cache_config: KVCacheConfig
) -> dict[str, torch.Tensor]:
"""
Initializes the KV cache buffer with the correct size. The buffer needs
to be reshaped to the desired shape before being used by the models.
Args:
kv_cache_config: The KV cache config
Returns:
dict[str, torch.Tensor]: A map between layer names to their
corresponding memory buffer for KV cache.
"""
kv_cache_raw_tensors: dict[str, torch.Tensor] = {}
for kv_cache_tensor in kv_cache_config.kv_cache_tensors:
tensor = torch.zeros(
kv_cache_tensor.size, dtype=torch.int8, device=self.device
)
for layer_name in kv_cache_tensor.shared_by:
kv_cache_raw_tensors[layer_name] = tensor
layer_names = set()
for group in kv_cache_config.kv_cache_groups:
for layer_name in group.layer_names:
if layer_name in self.runner_only_attn_layers:
continue
layer_names.add(layer_name)
assert layer_names == set(kv_cache_raw_tensors.keys()), (
"Some layers are not correctly initialized"
)
return kv_cache_raw_tensors
def _attn_group_iterator(self) -> Iterator[AttentionGroup]:
return itertools.chain.from_iterable(self.attn_groups)
def _kv_cache_spec_attn_group_iterator(self) -> Iterator[AttentionGroup]:
if not self.kv_cache_config.kv_cache_groups:
return
for attn_groups in self.attn_groups:
yield from attn_groups
def _prepare_kernel_block_sizes(self, kv_cache_config: KVCacheConfig) -> list[int]:
"""
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 the same block size (no splitting).
Args:
kv_cache_config: The KV cache configuration.
Returns:
list[int]: List of kernel block sizes for each cache group.
"""
kernel_block_sizes = []
for kv_cache_group_id, kv_cache_group in enumerate(
kv_cache_config.kv_cache_groups
):
kv_cache_spec = kv_cache_group.kv_cache_spec
if isinstance(kv_cache_spec, UniformTypeKVCacheSpecs):
# All layers in the UniformTypeKVCacheSpecs have the same type,
# Pick an arbitrary one to dispatch.
kv_cache_spec = next(iter(kv_cache_spec.kv_cache_specs.values()))
if isinstance(kv_cache_spec, EncoderOnlyAttentionSpec):
continue
elif isinstance(kv_cache_spec, AttentionSpec):
# This is an attention backend that supports virtual
# block splitting. Get the supported block sizes from
# all backends in the group.
attn_groups = self.attn_groups[kv_cache_group_id]
kv_manager_block_size = kv_cache_group.kv_cache_spec.block_size
selected_kernel_size = self._select_common_block_size(
kv_manager_block_size, attn_groups
)
kernel_block_sizes.append(selected_kernel_size)
elif isinstance(kv_cache_spec, MambaSpec):
# This is likely Mamba or other non-attention cache,
# no splitting.
kernel_block_sizes.append(kv_cache_spec.block_size)
else:
raise NotImplementedError(
f"unknown kv cache spec {kv_cache_group.kv_cache_spec}"
)
return kernel_block_sizes
def _reshape_kv_cache_tensors(
self,
kv_cache_config: KVCacheConfig,
kv_cache_raw_tensors: dict[str, torch.Tensor],
) -> dict[str, torch.Tensor]:
"""
Reshape the KV cache tensors to the desired shape and dtype.
Args:
kv_cache_config: The KV cache config
kv_cache_raw_tensors: The KV cache buffer of each layer, with
correct size but uninitialized shape.
Returns:
Dict[str, torch.Tensor]: A map between layer names to their
corresponding memory buffer for KV cache.
"""
kv_caches: dict[str, torch.Tensor] = {}
has_attn, has_mamba = False, False
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
raw_tensor = kv_cache_raw_tensors[layer_name]
assert raw_tensor.numel() % kv_cache_spec.page_size_bytes == 0
num_blocks = raw_tensor.numel() // kv_cache_spec.page_size_bytes
if isinstance(kv_cache_spec, AttentionSpec):
has_attn = True
kv_manager_block_size = kv_cache_spec.block_size
kernel_size_list = self._find_compatible_block_sizes(
kv_manager_block_size, attn_backend, return_all=False
)
kernel_size = kernel_size_list[0]
num_blocks_per_kv_block = kv_manager_block_size // kernel_size
kernel_num_blocks = num_blocks * num_blocks_per_kv_block
kv_cache_shape = attn_backend.get_kv_cache_shape(
kernel_num_blocks,
kernel_size,
kv_cache_spec.num_kv_heads,
kv_cache_spec.head_size,
cache_dtype_str=self.cache_config.cache_dtype,
)
dtype = kv_cache_spec.dtype
try:
kv_cache_stride_order = attn_backend.get_kv_cache_stride_order() # noqa: E501
assert len(kv_cache_stride_order) == len(kv_cache_shape)
except (AttributeError, NotImplementedError):
kv_cache_stride_order = tuple(range(len(kv_cache_shape)))
# The allocation respects the backend-defined stride order
# to ensure the semantic remains consistent for each
# backend. We first obtain the generic kv cache shape and
# then permute it according to the stride order which could
# result in a non-contiguous tensor.
kv_cache_shape = tuple(
kv_cache_shape[i] for i in kv_cache_stride_order
)
# Maintain original KV shape view.
inv_order = [
kv_cache_stride_order.index(i)
for i in range(len(kv_cache_stride_order))
]
kv_caches[layer_name] = (
kv_cache_raw_tensors[layer_name]
.view(dtype)
.view(kv_cache_shape)
.permute(*inv_order)
)
elif isinstance(kv_cache_spec, MambaSpec):
has_mamba = True
raw_tensor = kv_cache_raw_tensors[layer_name]
state_tensors = []
storage_offset_bytes = 0
for shape, dtype in zip(kv_cache_spec.shapes, kv_cache_spec.dtypes):
dtype_size = get_dtype_size(dtype)
num_element_per_page = (
kv_cache_spec.page_size_bytes // dtype_size
)
target_shape = (num_blocks, *shape)
stride = torch.empty(target_shape).stride()
target_stride = (num_element_per_page, *stride[1:])
assert storage_offset_bytes % dtype_size == 0
tensor = torch.as_strided(
raw_tensor.view(dtype),
size=target_shape,
stride=target_stride,
storage_offset=storage_offset_bytes // dtype_size,
)
state_tensors.append(tensor)
storage_offset_bytes += stride[0] * dtype_size
kv_caches[layer_name] = state_tensors
else:
raise NotImplementedError
if has_attn and has_mamba:
self._update_hybrid_attention_mamba_layout(kv_caches)
return kv_caches
def _update_hybrid_attention_mamba_layout(
self, kv_caches: dict[str, torch.Tensor]
) -> None:
"""
Update the layout of attention layers from (2, num_blocks, ...) to
(num_blocks, 2, ...).
Args:
kv_caches: The KV cache buffer of each layer.
"""
for group in self._kv_cache_spec_attn_group_iterator():
kv_cache_spec = group.kv_cache_spec
for layer_name in group.layer_names:
kv_cache = kv_caches[layer_name]
if isinstance(kv_cache_spec, AttentionSpec) and kv_cache.shape[0] == 2:
assert kv_cache.shape[1] != 2, (
"Fail to determine whether the layout is "
"(2, num_blocks, ...) or (num_blocks, 2, ...) for "
f"a tensor of shape {kv_cache.shape}"
)
hidden_size = kv_cache.shape[2:].numel()
kv_cache.as_strided_(
size=kv_cache.shape,
stride=(hidden_size, 2 * hidden_size, *kv_cache.stride()[2:]),
)
def initialize_kv_cache_tensors(
self, kv_cache_config: KVCacheConfig
) -> 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.
"""
# Initialize the memory buffer for KV cache
kv_cache_raw_tensors = self._allocate_kv_cache_tensors(kv_cache_config)
# Change the memory buffer to the desired shape
kv_caches = self._reshape_kv_cache_tensors(
kv_cache_config, kv_cache_raw_tensors
)
# Set up cross-layer KV cache sharing
for layer_name, target_layer_name in self.shared_kv_cache_layers.items():
logger.debug("%s reuses KV cache of %s", layer_name, target_layer_name)
kv_caches[layer_name] = kv_caches[target_layer_name]
num_attn_module = (
2 if self.model_config.hf_config.model_type == "longcat_flash" else 1
)
bind_kv_cache(
kv_caches,
self.compilation_config.static_forward_context,
self.kv_caches,
num_attn_module,
)
return kv_caches
def maybe_add_kv_sharing_layers_to_kv_cache_groups(
self, kv_cache_config: KVCacheConfig
) -> None:
"""
Add layers that re-use KV cache to KV cache group of its target layer.
Mapping of KV cache tensors happens in `initialize_kv_cache_tensors()`
"""
if not self.shared_kv_cache_layers:
# No cross-layer KV sharing, return
return
add_kv_sharing_layers_to_kv_cache_groups(
self.shared_kv_cache_layers,
kv_cache_config.kv_cache_groups,
self.runner_only_attn_layers,
)
if self.cache_config.kv_sharing_fast_prefill:
# In You Only Cache Once (https://arxiv.org/abs/2405.05254) or other
# similar KV sharing setups, only the layers that generate KV caches
# are involved in the prefill phase, enabling prefill to early exit.
attn_layers = get_layers_from_vllm_config(self.vllm_config, Attention)
for layer_name in reversed(attn_layers):
if layer_name in self.shared_kv_cache_layers:
self.kv_sharing_fast_prefill_eligible_layers.add(layer_name)
else:
break
def initialize_kv_cache(self, kv_cache_config: KVCacheConfig) -> 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.may_add_encoder_only_layers_to_kv_cache_config()
self.maybe_add_kv_sharing_layers_to_kv_cache_groups(kv_cache_config)
self.initialize_attn_backend(kv_cache_config)
# Reinitialize need to after initialize_attn_backend
self.may_reinitialize_input_batch(kv_cache_config)
kv_caches = self.initialize_kv_cache_tensors(kv_cache_config)
if self.speculative_config and self.speculative_config.use_eagle():
assert isinstance(self.drafter, EagleProposer)
# validate all draft model layers belong to the same kv cache
# group
self.drafter.validate_same_kv_cache_group(kv_cache_config)
if has_kv_transfer_group():
kv_transfer_group = get_kv_transfer_group()
kv_transfer_group.register_kv_caches(kv_caches)
kv_transfer_group.set_host_xfer_buffer_ops(copy_kv_blocks)
if self.dcp_world_size > 1:
layer_names = self.attn_groups[0][0].layer_names
layers = get_layers_from_vllm_config(
self.vllm_config, AttentionLayerBase, layer_names
)
for layer in layers.values():
assert layer.impl.need_to_return_lse_for_decode, (
"DCP requires attention impls to return"
" the softmax lse for decode, but the impl "
f"{layer.impl.__class__.__name__} "
"does not return the softmax lse for decode."
)
def may_add_encoder_only_layers_to_kv_cache_config(self) -> None:
"""
Add encoder-only layers to the KV cache config.
"""
block_size = self.vllm_config.cache_config.block_size
encoder_only_attn_specs: dict[AttentionSpec, list[str]] = defaultdict(list)
attn_layers = get_layers_from_vllm_config(self.vllm_config, Attention)
for layer_name, attn_module in attn_layers.items():
if attn_module.attn_type == AttentionType.ENCODER_ONLY:
attn_spec: AttentionSpec = EncoderOnlyAttentionSpec(
block_size=block_size,
num_kv_heads=attn_module.num_kv_heads,
head_size=attn_module.head_size,
dtype=self.kv_cache_dtype,
)
encoder_only_attn_specs[attn_spec].append(layer_name)
self.runner_only_attn_layers.add(layer_name)
if len(encoder_only_attn_specs) > 0:
assert len(encoder_only_attn_specs) == 1, (
"Only support one encoder-only attention spec now"
)
spec, layer_names = encoder_only_attn_specs.popitem()
self.kv_cache_config.kv_cache_groups.append(
KVCacheGroupSpec(layer_names=layer_names, kv_cache_spec=spec)
)
def get_kv_cache_spec(self) -> dict[str, KVCacheSpec]:
"""
Generates the KVCacheSpec by parsing the kv cache format from each
Attention module in the static forward context.
Returns:
KVCacheSpec: A dictionary mapping layer names to their KV cache
format. Layers that do not need KV cache are not included.
"""
kv_cache_spec: dict[str, KVCacheSpec] = {}
attn_layers = get_layers_from_vllm_config(self.vllm_config, AttentionLayerBase)
for layer_name, attn_module in attn_layers.items():
if isinstance(attn_module, Attention) and (
kv_tgt_layer := attn_module.kv_sharing_target_layer_name
):
# The layer doesn't need its own KV cache and will use that of
# the target layer. We skip creating a KVCacheSpec for it, so
# that KV cache management logic will act as this layer does
# not exist, and doesn't allocate KV cache for the layer. This
# enables the memory saving of cross-layer kv sharing, allowing
# a given amount of memory to accommodate longer context lengths
# or enable more requests to be processed simultaneously.
self.shared_kv_cache_layers[layer_name] = kv_tgt_layer
continue
# Skip modules that don't need KV cache (eg encoder-only attention)
if spec := attn_module.get_kv_cache_spec(self.vllm_config):
kv_cache_spec[layer_name] = spec
return kv_cache_spec
def _to_list(self, sampled_token_ids: torch.Tensor) -> list[list[int]]:
# This is a short term mitigation for issue mentioned in
# https://github.com/vllm-project/vllm/issues/22754.
# `tolist` would trigger a cuda wise stream sync, which
# would block other copy ops from other cuda streams.
# A cuda event sync would avoid such a situation. Since
# this is in the critical path of every single model
# forward loop, this has caused perf issue for a disagg
# setup.
pinned = self.sampled_token_ids_pinned_cpu[: sampled_token_ids.shape[0]]
pinned.copy_(sampled_token_ids, non_blocking=True)
self.transfer_event.record()
self.transfer_event.synchronize()
return pinned.tolist()