This commit is contained in:
@@ -10,7 +10,7 @@ from vllm.multimodal import MultiModalRegistry
|
||||
from vllm.v1.request import Request
|
||||
|
||||
if TYPE_CHECKING:
|
||||
from vllm.config import RendererConfig, SchedulerConfig
|
||||
from vllm.config import ModelConfig, SchedulerConfig
|
||||
|
||||
logger = init_logger(__name__)
|
||||
|
||||
@@ -250,7 +250,7 @@ class EncoderCacheManager:
|
||||
|
||||
|
||||
def compute_encoder_budget(
|
||||
renderer_config: "RendererConfig",
|
||||
model_config: "ModelConfig",
|
||||
scheduler_config: "SchedulerConfig",
|
||||
mm_registry: MultiModalRegistry,
|
||||
) -> tuple[int, int]:
|
||||
@@ -263,9 +263,9 @@ def compute_encoder_budget(
|
||||
- Space budget for encoder cache size, measured in number of tokens
|
||||
from the input sequence.
|
||||
"""
|
||||
if mm_registry.supports_multimodal_inputs(renderer_config):
|
||||
if mm_registry.supports_multimodal_inputs(model_config):
|
||||
max_tokens_by_modality = mm_registry.get_max_tokens_per_item_by_modality(
|
||||
renderer_config
|
||||
model_config
|
||||
)
|
||||
|
||||
return compute_mm_encoder_budget(
|
||||
|
||||
@@ -164,7 +164,7 @@ class Scheduler(SchedulerInterface):
|
||||
# This can be changed when we make encoder cache for embedding caching
|
||||
# across requests.
|
||||
encoder_compute_budget, encoder_cache_size = compute_encoder_budget(
|
||||
renderer_config=vllm_config.renderer_config,
|
||||
model_config=vllm_config.model_config,
|
||||
scheduler_config=vllm_config.scheduler_config,
|
||||
mm_registry=mm_registry,
|
||||
)
|
||||
|
||||
@@ -91,7 +91,6 @@ class AsyncLLM(EngineClient):
|
||||
# Ensure we can serialize custom transformer configs
|
||||
maybe_register_config_serialize_by_value()
|
||||
|
||||
self.renderer_config = vllm_config.renderer_config
|
||||
self.model_config = vllm_config.model_config
|
||||
self.vllm_config = vllm_config
|
||||
self.observability_config = vllm_config.observability_config
|
||||
@@ -109,15 +108,15 @@ class AsyncLLM(EngineClient):
|
||||
"enabling logging without default stat loggers."
|
||||
)
|
||||
|
||||
if self.renderer_config.skip_tokenizer_init:
|
||||
if self.model_config.skip_tokenizer_init:
|
||||
tokenizer = None
|
||||
else:
|
||||
tokenizer = init_tokenizer_from_config(self.renderer_config)
|
||||
tokenizer = init_tokenizer_from_config(self.model_config)
|
||||
|
||||
self.input_processor = InputProcessor(self.vllm_config, tokenizer)
|
||||
self.io_processor = get_io_processor(
|
||||
self.vllm_config,
|
||||
self.renderer_config.io_processor_plugin,
|
||||
self.model_config.io_processor_plugin,
|
||||
)
|
||||
|
||||
# OutputProcessor (converts EngineCoreOutputs --> RequestOutput).
|
||||
|
||||
@@ -43,7 +43,6 @@ class InputProcessor:
|
||||
mm_registry: MultiModalRegistry = MULTIMODAL_REGISTRY,
|
||||
) -> None:
|
||||
self.vllm_config = vllm_config
|
||||
self.renderer_config = vllm_config.renderer_config
|
||||
self.model_config = vllm_config.model_config
|
||||
self.cache_config = vllm_config.cache_config
|
||||
self.lora_config = vllm_config.lora_config
|
||||
@@ -55,7 +54,7 @@ class InputProcessor:
|
||||
self.mm_processor_cache = processor_cache_from_config(vllm_config, mm_registry)
|
||||
|
||||
self.input_preprocessor = InputPreprocessor(
|
||||
self.renderer_config,
|
||||
self.model_config,
|
||||
tokenizer,
|
||||
mm_registry,
|
||||
mm_processor_cache=self.mm_processor_cache,
|
||||
@@ -253,7 +252,7 @@ class InputProcessor:
|
||||
if not params.structured_outputs or not self.structured_outputs_config:
|
||||
return
|
||||
|
||||
if self.renderer_config.skip_tokenizer_init and params.structured_outputs:
|
||||
if self.model_config.skip_tokenizer_init and params.structured_outputs:
|
||||
raise ValueError(
|
||||
"Structured outputs requires a tokenizer so it can't be used with 'skip_tokenizer_init'" # noqa: E501
|
||||
)
|
||||
@@ -583,7 +582,7 @@ class InputProcessor:
|
||||
if prompt_type == "encoder" and model_config.is_multimodal_model:
|
||||
mm_registry = self.input_preprocessor.mm_registry
|
||||
mm_processor = mm_registry.create_processor(
|
||||
self.renderer_config,
|
||||
model_config,
|
||||
tokenizer=tokenizer,
|
||||
)
|
||||
assert isinstance(mm_processor, EncDecMultiModalProcessor)
|
||||
|
||||
@@ -60,7 +60,6 @@ class LLMEngine:
|
||||
) -> None:
|
||||
self.vllm_config = vllm_config
|
||||
self.observability_config = vllm_config.observability_config
|
||||
self.renderer_config = vllm_config.renderer_config
|
||||
self.model_config = vllm_config.model_config
|
||||
self.cache_config = vllm_config.cache_config
|
||||
|
||||
@@ -84,15 +83,15 @@ class LLMEngine:
|
||||
self.dp_group = None
|
||||
self.should_execute_dummy_batch = False
|
||||
|
||||
if self.renderer_config.skip_tokenizer_init:
|
||||
if self.model_config.skip_tokenizer_init:
|
||||
tokenizer = None
|
||||
else:
|
||||
tokenizer = init_tokenizer_from_config(self.renderer_config)
|
||||
tokenizer = init_tokenizer_from_config(self.model_config)
|
||||
|
||||
self.input_processor = InputProcessor(self.vllm_config, tokenizer)
|
||||
self.io_processor = get_io_processor(
|
||||
self.vllm_config,
|
||||
self.renderer_config.io_processor_plugin,
|
||||
self.model_config.io_processor_plugin,
|
||||
)
|
||||
|
||||
# OutputProcessor (convert EngineCoreOutputs --> RequestOutput).
|
||||
|
||||
@@ -85,7 +85,7 @@ class EagleProposer:
|
||||
# Multi-modal data support
|
||||
self.mm_registry = MULTIMODAL_REGISTRY
|
||||
self.supports_mm_inputs = self.mm_registry.supports_multimodal_inputs(
|
||||
vllm_config.renderer_config
|
||||
vllm_config.model_config
|
||||
)
|
||||
|
||||
self.attn_metadata_builder: AttentionMetadataBuilder | None = None
|
||||
|
||||
@@ -63,7 +63,7 @@ class StructuredOutputManager:
|
||||
max_workers = max(1, min(multiprocessing.cpu_count() // 2, 8))
|
||||
self.executor_for_fillmask = ThreadPoolExecutor(max_workers=max_workers)
|
||||
|
||||
if not vllm_config.renderer_config.skip_tokenizer_init:
|
||||
if not self.vllm_config.model_config.skip_tokenizer_init:
|
||||
# The default max_workers if not specified is the number of
|
||||
# CPUs * 5, which is way too high since these tasks are CPU-bound,
|
||||
# not I/O bound. We also know we would never dominate CPU usage
|
||||
@@ -71,15 +71,21 @@ class StructuredOutputManager:
|
||||
# of CPUs.
|
||||
max_workers = max(1, (multiprocessing.cpu_count() + 1) // 2)
|
||||
self.executor = ThreadPoolExecutor(max_workers=max_workers)
|
||||
self.tokenizer = init_tokenizer_from_config(vllm_config.renderer_config)
|
||||
reasoning_parser = vllm_config.structured_outputs_config.reasoning_parser
|
||||
self.tokenizer = init_tokenizer_from_config(
|
||||
model_config=self.vllm_config.model_config
|
||||
)
|
||||
reasoning_parser = (
|
||||
self.vllm_config.structured_outputs_config.reasoning_parser
|
||||
)
|
||||
reasoning_parser_plugin = (
|
||||
vllm_config.structured_outputs_config.reasoning_parser_plugin
|
||||
self.vllm_config.structured_outputs_config.reasoning_parser_plugin
|
||||
)
|
||||
if reasoning_parser_plugin and len(reasoning_parser_plugin) > 3:
|
||||
ReasoningParserManager.import_reasoning_parser(reasoning_parser_plugin)
|
||||
|
||||
reasoning_parser = vllm_config.structured_outputs_config.reasoning_parser
|
||||
reasoning_parser = (
|
||||
self.vllm_config.structured_outputs_config.reasoning_parser
|
||||
)
|
||||
if reasoning_parser:
|
||||
reasoner_cls = ReasoningParserManager.get_reasoning_parser(
|
||||
reasoning_parser
|
||||
@@ -87,7 +93,7 @@ class StructuredOutputManager:
|
||||
self.reasoner = reasoner_cls(tokenizer=self.tokenizer)
|
||||
|
||||
self.enable_in_reasoning = (
|
||||
vllm_config.structured_outputs_config.enable_in_reasoning
|
||||
self.vllm_config.structured_outputs_config.enable_in_reasoning
|
||||
)
|
||||
|
||||
def grammar_init(self, request: Request) -> None:
|
||||
|
||||
@@ -271,7 +271,6 @@ class GPUModelRunner(
|
||||
device: torch.device,
|
||||
):
|
||||
self.vllm_config = vllm_config
|
||||
self.renderer_config = vllm_config.renderer_config
|
||||
self.model_config = vllm_config.model_config
|
||||
self.cache_config = vllm_config.cache_config
|
||||
self.compilation_config = vllm_config.compilation_config
|
||||
@@ -336,7 +335,7 @@ class GPUModelRunner(
|
||||
self.uses_mrope = model_config.uses_mrope
|
||||
self.uses_xdrope_dim = model_config.uses_xdrope_dim
|
||||
self.supports_mm_inputs = self.mm_registry.supports_multimodal_inputs(
|
||||
self.renderer_config
|
||||
model_config
|
||||
)
|
||||
|
||||
if self.model_config.is_encoder_decoder:
|
||||
@@ -559,7 +558,7 @@ class GPUModelRunner(
|
||||
|
||||
self.mm_budget = (
|
||||
MultiModalBudget(
|
||||
self.renderer_config,
|
||||
self.model_config,
|
||||
self.scheduler_config,
|
||||
self.mm_registry,
|
||||
)
|
||||
@@ -3874,7 +3873,7 @@ class GPUModelRunner(
|
||||
assert self.mm_budget is not None
|
||||
|
||||
dummy_decoder_data = self.mm_registry.get_decoder_dummy_data(
|
||||
renderer_config=self.renderer_config,
|
||||
model_config=self.model_config,
|
||||
seq_len=self.max_model_len,
|
||||
mm_counts={modality: 1},
|
||||
cache=self.mm_budget.cache,
|
||||
|
||||
@@ -143,7 +143,6 @@ class TPUModelRunner(LoRAModelRunnerMixin, KVConnectorModelRunnerMixin):
|
||||
original_parallel_config: ParallelConfig | None = None,
|
||||
):
|
||||
self.vllm_config = vllm_config
|
||||
self.renderer_config = vllm_config.renderer_config
|
||||
self.model_config = vllm_config.model_config
|
||||
self.cache_config = vllm_config.cache_config
|
||||
self.lora_config = vllm_config.lora_config
|
||||
@@ -223,7 +222,7 @@ class TPUModelRunner(LoRAModelRunnerMixin, KVConnectorModelRunnerMixin):
|
||||
self.mm_registry = MULTIMODAL_REGISTRY
|
||||
self.uses_mrope = model_config.uses_mrope
|
||||
self.supports_mm_inputs = self.mm_registry.supports_multimodal_inputs(
|
||||
self.renderer_config
|
||||
model_config
|
||||
)
|
||||
# TODO: Support M-RoPE (e.g, Qwen2-VL)
|
||||
assert not self.uses_mrope, "TPU does not support M-RoPE yet."
|
||||
@@ -354,7 +353,7 @@ class TPUModelRunner(LoRAModelRunnerMixin, KVConnectorModelRunnerMixin):
|
||||
|
||||
self.mm_budget = (
|
||||
MultiModalBudget(
|
||||
self.renderer_config,
|
||||
self.model_config,
|
||||
self.scheduler_config,
|
||||
self.mm_registry,
|
||||
)
|
||||
@@ -2039,7 +2038,7 @@ class TPUModelRunner(LoRAModelRunnerMixin, KVConnectorModelRunnerMixin):
|
||||
assert self.mm_budget is not None
|
||||
|
||||
dummy_decoder_data = self.mm_registry.get_decoder_dummy_data(
|
||||
renderer_config=self.renderer_config,
|
||||
model_config=self.model_config,
|
||||
seq_len=self.max_model_len,
|
||||
mm_counts={modality: 1},
|
||||
cache=self.mm_budget.cache,
|
||||
|
||||
+7
-12
@@ -7,7 +7,7 @@ import torch
|
||||
|
||||
from vllm.attention.backends.abstract import AttentionBackend
|
||||
from vllm.attention.layer import Attention
|
||||
from vllm.config import RendererConfig, SchedulerConfig, VllmConfig
|
||||
from vllm.config import ModelConfig, SchedulerConfig, VllmConfig
|
||||
from vllm.model_executor.models.interfaces import MultiModalEmbeddings
|
||||
from vllm.model_executor.models.utils import extract_layer_index
|
||||
from vllm.multimodal.cache import processor_only_cache_from_config
|
||||
@@ -23,29 +23,24 @@ class MultiModalBudget:
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
renderer_config: RendererConfig,
|
||||
model_config: ModelConfig,
|
||||
scheduler_config: SchedulerConfig,
|
||||
mm_registry: MultiModalRegistry,
|
||||
) -> None:
|
||||
super().__init__()
|
||||
|
||||
self.renderer_config = renderer_config
|
||||
self.model_config = renderer_config.model_config
|
||||
self.model_config = model_config
|
||||
self.scheduler_config = scheduler_config
|
||||
self.mm_registry = mm_registry
|
||||
self.cache = cache = processor_only_cache_from_config(
|
||||
renderer_config, mm_registry
|
||||
)
|
||||
self.cache = cache = processor_only_cache_from_config(model_config, mm_registry)
|
||||
|
||||
self.max_model_len = self.model_config.max_model_len
|
||||
self.max_model_len = model_config.max_model_len
|
||||
self.max_num_reqs = scheduler_config.max_num_seqs
|
||||
|
||||
self.mm_limits = mm_registry.get_mm_limits_per_prompt(
|
||||
renderer_config, cache=cache
|
||||
)
|
||||
self.mm_limits = mm_registry.get_mm_limits_per_prompt(model_config, cache=cache)
|
||||
|
||||
max_tokens_by_modality = mm_registry.get_max_tokens_per_item_by_modality(
|
||||
renderer_config,
|
||||
model_config,
|
||||
cache=cache,
|
||||
profiler_limits=self.mm_limits,
|
||||
)
|
||||
|
||||
Reference in New Issue
Block a user