[V0 Deprecation] Remove Prompt Adapters (#20588)

Signed-off-by: mgoin <mgoin64@gmail.com>
This commit is contained in:
Michael Goin
2025-07-23 16:36:48 -07:00
committed by GitHub
parent 78c13e30e1
commit 82ec66f514
60 changed files with 126 additions and 1727 deletions
+3 -4
View File
@@ -91,10 +91,9 @@ class EncoderDecoderModelRunner(GPUModelRunnerBase[EncoderDecoderModelInput]):
'''
EncoderDecoderModelRunner constructor.
`lora_config` and `prompt_adapter_config` are
unused (since these features are not yet supported for encoder/decoder
models) but these arguments are present here for compatibility with
the base-class constructor.
`lora_config` is unused (since these features are not yet supported
for encoder/decoder models) but these arguments are present here for
compatibility with the base-class constructor.
'''
self._maybe_force_supported_attention_backend()
+2 -149
View File
@@ -45,10 +45,6 @@ from vllm.model_executor.models.utils import set_cpu_offload_max_bytes
from vllm.multimodal import (MULTIMODAL_REGISTRY, BatchedTensorInputs,
MultiModalKwargs, MultiModalPlaceholderMap,
MultiModalRegistry)
from vllm.prompt_adapter.layers import PromptAdapterMapping
from vllm.prompt_adapter.request import PromptAdapterRequest
from vllm.prompt_adapter.worker_manager import (
LRUCacheWorkerPromptAdapterManager)
from vllm.sampling_params import SamplingParams
from vllm.sequence import IntermediateTensors, SequenceGroupMetadata
from vllm.utils import (DeviceMemoryProfiler, GiB_bytes, PyObjectCache,
@@ -95,8 +91,6 @@ class ModelInputForGPU(ModelRunnerInputBase):
lora_mapping: Optional["LoRAMapping"] = None
lora_requests: Optional[Set[LoRARequest]] = None
attn_metadata: Optional["AttentionMetadata"] = None
prompt_adapter_mapping: Optional[PromptAdapterMapping] = None
prompt_adapter_requests: Optional[Set[PromptAdapterRequest]] = None
multi_modal_kwargs: Optional[BatchedTensorInputs] = None
request_ids_to_seq_ids: Optional[Dict[str, List[int]]] = None
finished_requests_ids: Optional[List[str]] = None
@@ -113,8 +107,6 @@ class ModelInputForGPU(ModelRunnerInputBase):
"lora_requests": self.lora_requests,
"lora_mapping": self.lora_mapping,
"multi_modal_kwargs": self.multi_modal_kwargs,
"prompt_adapter_mapping": self.prompt_adapter_mapping,
"prompt_adapter_requests": self.prompt_adapter_requests,
"virtual_engine": self.virtual_engine,
"request_ids_to_seq_ids": self.request_ids_to_seq_ids,
"finished_requests_ids": self.finished_requests_ids,
@@ -164,8 +156,6 @@ class ModelInputForGPUWithSamplingMetadata(ModelInputForGPU):
"lora_requests": self.lora_requests,
"lora_mapping": self.lora_mapping,
"multi_modal_kwargs": self.multi_modal_kwargs,
"prompt_adapter_mapping": self.prompt_adapter_mapping,
"prompt_adapter_requests": self.prompt_adapter_requests,
"virtual_engine": self.virtual_engine,
"request_ids_to_seq_ids": self.request_ids_to_seq_ids,
"finished_requests_ids": self.finished_requests_ids,
@@ -212,8 +202,6 @@ class ModelInputForGPUBuilder(ModelRunnerInputBuilderBase[ModelInputForGPU]):
self.lora_index_mapping.clear() # type: ignore
self.lora_prompt_mapping.clear() # type: ignore
self.lora_requests.clear() # type: ignore
self.prompt_adapter_index_mapping.clear() # type: ignore
self.prompt_adapter_prompt_mapping.clear() # type: ignore
def __init__(
self,
@@ -252,11 +240,6 @@ class ModelInputForGPUBuilder(ModelRunnerInputBuilderBase[ModelInputForGPU]):
lora_prompt_mapping: Optional[List[List[int]]] = None,
lora_requests: Optional[Set[LoRARequest]] = None,
# Prompt adapter inputs.
prompt_adapter_index_mapping: Optional[List[int]] = None,
prompt_adapter_prompt_mapping: Optional[List[int]] = None,
prompt_adapter_request: Optional[PromptAdapterRequest] = None,
# Multi-modal inputs.
multi_modal_kwargs: Optional[MultiModalKwargs] = None,
multi_modal_placeholder_maps: Optional[Dict[
@@ -360,18 +343,6 @@ class ModelInputForGPUBuilder(ModelRunnerInputBuilderBase[ModelInputForGPU]):
else:
self.lora_requests.clear()
if prompt_adapter_index_mapping:
self.prompt_adapter_index_mapping = \
prompt_adapter_index_mapping
else:
self.prompt_adapter_index_mapping.clear()
if prompt_adapter_prompt_mapping:
self.prompt_adapter_prompt_mapping = \
prompt_adapter_prompt_mapping
else:
self.prompt_adapter_prompt_mapping.clear()
else:
self.input_tokens = input_tokens or []
self.inputs_embeds = inputs_embeds
@@ -390,12 +361,6 @@ class ModelInputForGPUBuilder(ModelRunnerInputBuilderBase[ModelInputForGPU]):
self.lora_prompt_mapping = lora_prompt_mapping or []
self.lora_requests = lora_requests or set()
self.prompt_adapter_index_mapping = (
prompt_adapter_index_mapping or [])
self.prompt_adapter_prompt_mapping = (
prompt_adapter_prompt_mapping or [])
self.prompt_adapter_request = prompt_adapter_request
self.multi_modal_kwargs = multi_modal_kwargs
self.multi_modal_placeholder_maps = multi_modal_placeholder_maps
self.prefix_cache_hit = prefix_cache_hit
@@ -485,7 +450,6 @@ class ModelInputForGPUBuilder(ModelRunnerInputBuilderBase[ModelInputForGPU]):
# Compute functions for each sequence group.
# WARNING: The order of the functions matters!
self.per_seq_group_compute_fns = [
self._compute_prompt_adapter_input,
self._compute_multi_modal_input,
]
@@ -496,8 +460,6 @@ class ModelInputForGPUBuilder(ModelRunnerInputBuilderBase[ModelInputForGPU]):
self.sliding_window = self.runner.sliding_window
self.block_size = self.runner.block_size
self.enable_lora = self.runner.lora_config is not None
self.enable_prompt_adapter = (self.runner.prompt_adapter_config
is not None)
# Attention metadata inputs.
if self.attn_backend is not None:
@@ -693,34 +655,6 @@ class ModelInputForGPUBuilder(ModelRunnerInputBuilderBase[ModelInputForGPU]):
else:
inter_data.lora_prompt_mapping.append([])
def _compute_prompt_adapter_input(
self, inter_data: InterDataForSeqGroup,
seq_group_metadata: SequenceGroupMetadata):
"""If prompt adapter is enabled, compute index and prompt mapping.
"""
# Note that when is_prompt=True, we expect only one sequence
# in the group.
if not self.enable_prompt_adapter:
return
prompt_adapter_id = seq_group_metadata.prompt_adapter_id
if prompt_adapter_id <= 0 or not inter_data.is_prompt:
return
# We expect only one sequence in the group when is_prompt=True.
assert inter_data.n_seqs == 1
query_len = inter_data.query_lens[0]
inter_data.prompt_adapter_request = (
seq_group_metadata.prompt_adapter_request)
num_tokens = seq_group_metadata.prompt_adapter_num_virtual_tokens
inter_data.prompt_adapter_index_mapping = [
prompt_adapter_id
] * num_tokens + [0] * (query_len - num_tokens)
inter_data.prompt_adapter_prompt_mapping = [prompt_adapter_id] * (
query_len if seq_group_metadata.sampling_params
and seq_group_metadata.sampling_params.prompt_logprobs else 1)
def _compute_multi_modal_input(self, inter_data: InterDataForSeqGroup,
seq_group_metadata: SequenceGroupMetadata):
"""If multi-modal data is given, add it to the input."""
@@ -1009,29 +943,6 @@ class ModelInputForGPUBuilder(ModelRunnerInputBuilderBase[ModelInputForGPU]):
prompt_mapping=lora_prompt_mapping,
is_prefill=not self.decode_only))
# Prompt adapter data.
prompt_adapter_requests: Set[PromptAdapterRequest] = set()
prompt_adapter_mapping = None
if self.enable_prompt_adapter:
prompt_adapter_requests = set(
data.prompt_adapter_request for data in self.inter_data_list
if data.prompt_adapter_request is not None)
prompt_adapter_index_mapping = flatten_2d_lists([
inter_data.prompt_adapter_index_mapping
for inter_data in self.inter_data_list
])
if cuda_graph_pad_size:
prompt_adapter_index_mapping.extend(
itertools.repeat(0, cuda_graph_pad_size))
prompt_adapter_prompt_mapping = flatten_2d_lists([
inter_data.prompt_adapter_prompt_mapping
for inter_data in self.inter_data_list
])
prompt_adapter_mapping = PromptAdapterMapping(
prompt_adapter_index_mapping,
prompt_adapter_prompt_mapping,
)
# Multi-modal data.
multi_modal_kwargs_list = [
data.multi_modal_kwargs for data in self.inter_data_list
@@ -1051,9 +962,7 @@ class ModelInputForGPUBuilder(ModelRunnerInputBuilderBase[ModelInputForGPU]):
lora_requests=lora_requests,
multi_modal_kwargs=multi_modal_kwargs,
request_ids_to_seq_ids=request_ids_to_seq_ids,
finished_requests_ids=self.finished_requests_ids,
prompt_adapter_mapping=prompt_adapter_mapping,
prompt_adapter_requests=prompt_adapter_requests)
finished_requests_ids=self.finished_requests_ids)
class GPUModelRunnerBase(ModelRunnerBase[TModelInputForGPU]):
@@ -1148,7 +1057,6 @@ class GPUModelRunnerBase(ModelRunnerBase[TModelInputForGPU]):
self.model: nn.Module # Set after load_model
# Set after load_model.
self.lora_manager: Optional[LRUCacheWorkerLoRAManager] = None
self.prompt_adapter_manager: LRUCacheWorkerPromptAdapterManager = None
self.sampler = get_sampler()
set_cpu_offload_max_bytes(
@@ -1207,14 +1115,7 @@ class GPUModelRunnerBase(ModelRunnerBase[TModelInputForGPU]):
logger.info("Model loading took %.4f GiB and %.6f seconds",
self.model_memory_usage / GiB_bytes,
time_after_load - time_before_load)
if self.prompt_adapter_config:
self.prompt_adapter_manager = LRUCacheWorkerPromptAdapterManager(
self.scheduler_config.max_num_seqs,
self.scheduler_config.max_num_batched_tokens, self.device,
self.prompt_adapter_config)
self.model = (
self.prompt_adapter_manager.create_prompt_adapter_manager(
self.model))
if self.vllm_config.compilation_config.level ==\
CompilationLevel.DYNAMO_AS_IS and supports_dynamo():
@@ -1466,40 +1367,6 @@ class GPUModelRunnerBase(ModelRunnerBase[TModelInputForGPU]):
raise RuntimeError("LoRA is not enabled.")
return self.lora_manager.list_adapters()
def remove_all_prompt_adapters(self):
if not self.prompt_adapter_manager:
raise RuntimeError("PromptAdapter is not enabled.")
self.prompt_adapter_manager.remove_all_adapters()
def set_active_prompt_adapters(
self, prompt_adapter_requests: Set[PromptAdapterRequest],
prompt_adapter_mapping: PromptAdapterMapping) -> None:
if not self.prompt_adapter_manager:
raise RuntimeError("PromptAdapter is not enabled.")
self.prompt_adapter_manager.set_active_adapters(
prompt_adapter_requests, prompt_adapter_mapping)
def add_prompt_adapter(
self, prompt_adapter_request: PromptAdapterRequest) -> bool:
if not self.prompt_adapter_manager:
raise RuntimeError("PromptAdapter is not enabled.")
return self.prompt_adapter_manager.add_adapter(prompt_adapter_request)
def remove_prompt_adapter(self, prompt_adapter_id: int) -> bool:
if not self.prompt_adapter_manager:
raise RuntimeError("PromptAdapter is not enabled.")
return self.prompt_adapter_manager.remove_adapter(prompt_adapter_id)
def pin_prompt_adapter(self, prompt_adapter_id: int) -> bool:
if not self.prompt_adapter_manager:
raise RuntimeError("PromptAdapter is not enabled.")
return self.prompt_adapter_manager.pin_adapter(prompt_adapter_id)
def list_prompt_adapters(self) -> Set[int]:
if not self.prompt_adapter_manager:
raise RuntimeError("PromptAdapter is not enabled.")
return self.prompt_adapter_manager.list_adapters()
@torch.inference_mode()
def capture_model(self, kv_caches: List[List[torch.Tensor]]) -> None:
"""Cuda graph capture a model.
@@ -1609,13 +1476,6 @@ class GPUModelRunnerBase(ModelRunnerBase[TModelInputForGPU]):
self.set_active_loras(set([dummy_lora_request]),
lora_mapping)
if self.prompt_adapter_config:
prompt_adapter_mapping = PromptAdapterMapping(
[-1] * batch_size,
[-1] * batch_size,
)
self.set_active_prompt_adapters(
set(), prompt_adapter_mapping)
graph_runner = CUDAGraphRunner(
self.model, self.attn_backend.get_name(),
self.attn_state.graph_clone(batch_size),
@@ -1776,13 +1636,6 @@ class ModelRunner(GPUModelRunnerBase[ModelInputForGPUWithSamplingMetadata]):
self.set_active_loras(model_input.lora_requests,
model_input.lora_mapping)
if self.prompt_adapter_config:
assert model_input.prompt_adapter_requests is not None
assert model_input.prompt_adapter_mapping is not None
self.set_active_prompt_adapters(
model_input.prompt_adapter_requests,
model_input.prompt_adapter_mapping)
self.attn_state.begin_forward(model_input)
# Currently cuda graph is only supported by the decode phase.
-1
View File
@@ -190,7 +190,6 @@ class ModelRunnerBase(ABC, Generic[T]):
self.scheduler_config = vllm_config.scheduler_config
self.device_config = vllm_config.device_config
self.speculative_config = vllm_config.speculative_config
self.prompt_adapter_config = vllm_config.prompt_adapter_config
self.observability_config = vllm_config.observability_config
# Map of request_id -> generator used for seeded random sampling
-3
View File
@@ -288,9 +288,6 @@ class StatefulModelInput(BroadcastableModelInput):
assert fmi.lora_requests is not None
assert len(fmi.lora_requests) == 0
assert fmi.attn_metadata is not None
assert fmi.prompt_adapter_mapping is None
assert fmi.prompt_adapter_requests is not None
assert len(fmi.prompt_adapter_requests) == 0
assert fmi.multi_modal_kwargs is not None
assert len(fmi.multi_modal_kwargs) == 0
-7
View File
@@ -64,13 +64,6 @@ class PoolingModelRunner(
self.set_active_loras(model_input.lora_requests,
model_input.lora_mapping)
if self.prompt_adapter_config:
assert model_input.prompt_adapter_requests is not None
assert model_input.prompt_adapter_mapping is not None
self.set_active_prompt_adapters(
model_input.prompt_adapter_requests,
model_input.prompt_adapter_mapping)
# Currently cuda graph is only supported by the decode phase.
assert model_input.attn_metadata is not None
prefill_meta = model_input.attn_metadata.prefill_metadata
-4
View File
@@ -47,7 +47,3 @@ def assert_enc_dec_mr_supported_scenario(
if enc_dec_mr.scheduler_config.num_lookahead_slots > 0:
raise NotImplementedError(
STR_NOT_IMPL_ENC_DEC_ERR_STRS['STR_NOT_IMPL_ENC_DEC_SPEC_DEC'])
if enc_dec_mr.prompt_adapter_config is not None:
raise NotImplementedError(STR_NOT_IMPL_ENC_DEC_ERR_STRS[
'STR_NOT_IMPL_ENC_DEC_PROMPT_ADAPTER'])
-14
View File
@@ -22,7 +22,6 @@ from vllm.model_executor import set_random_seed
from vllm.model_executor.layers.sampler import SamplerOutput
from vllm.model_executor.model_loader.tensorizer import TensorizerConfig
from vllm.platforms import current_platform
from vllm.prompt_adapter.request import PromptAdapterRequest
from vllm.sequence import (ExecuteModelRequest, IntermediateTensors,
SequenceGroupMetadata, SequenceGroupMetadataDelta)
from vllm.utils import (GiB_bytes, MemorySnapshot, bind_kv_cache,
@@ -513,19 +512,6 @@ class Worker(LocalOrDistributedWorkerBase):
def list_loras(self) -> Set[int]:
return self.model_runner.list_loras()
def add_prompt_adapter(
self, prompt_adapter_request: PromptAdapterRequest) -> bool:
return self.model_runner.add_prompt_adapter(prompt_adapter_request)
def remove_prompt_adapter(self, prompt_adapter_id: int) -> bool:
return self.model_runner.remove_lora(prompt_adapter_id)
def pin_prompt_adapter(self, prompt_adapter_id: int) -> bool:
return self.model_runner.pin_prompt_adapter(prompt_adapter_id)
def list_prompt_adapters(self) -> Set[int]:
return self.model_runner.list_prompt_adapters()
@property
def max_model_len(self) -> int:
return self.model_config.max_model_len
-1
View File
@@ -49,7 +49,6 @@ class WorkerBase:
self.scheduler_config = vllm_config.scheduler_config
self.device_config = vllm_config.device_config
self.speculative_config = vllm_config.speculative_config
self.prompt_adapter_config = vllm_config.prompt_adapter_config
self.observability_config = vllm_config.observability_config
self.kv_transfer_config = vllm_config.kv_transfer_config
self.compilation_config = vllm_config.compilation_config