[ModelRunner V2] Deduplicate ModelState init logic (#46776)
Signed-off-by: Nick Hill <nickhill123@gmail.com>
This commit is contained in:
@@ -776,33 +776,7 @@ class DiffusionGemmaModelState(ModelState):
|
||||
encoder_cache: Any,
|
||||
device: torch.device,
|
||||
) -> None:
|
||||
self.vllm_config = vllm_config
|
||||
self.model_config = vllm_config.model_config
|
||||
self.scheduler_config = vllm_config.scheduler_config
|
||||
self.model = model
|
||||
self.device = device
|
||||
|
||||
self.supports_mm_inputs = encoder_cache is not None
|
||||
self.max_num_reqs = self.scheduler_config.max_num_seqs
|
||||
self.max_num_tokens = self.scheduler_config.max_num_batched_tokens
|
||||
self.max_model_len = self.model_config.max_model_len
|
||||
self.inputs_embeds_size = self.model_config.get_inputs_embeds_size()
|
||||
self.dtype = self.model_config.dtype
|
||||
|
||||
if self.supports_mm_inputs:
|
||||
from vllm.v1.worker.gpu.mm.encoder_cache import EncoderCache
|
||||
from vllm.v1.worker.gpu.mm.encoder_runner import EncoderRunner
|
||||
|
||||
assert isinstance(encoder_cache, EncoderCache)
|
||||
self.encoder_cache = encoder_cache
|
||||
self.encoder_runner = EncoderRunner(
|
||||
model=self.model,
|
||||
max_num_tokens=self.max_num_tokens,
|
||||
hidden_size=self.inputs_embeds_size,
|
||||
encoder_cache=encoder_cache,
|
||||
dtype=self.dtype,
|
||||
device=self.device,
|
||||
)
|
||||
super().__init__(vllm_config, model, encoder_cache, device)
|
||||
|
||||
# Per-step MM data produced by get_mm_embeddings and consumed by
|
||||
# prepare_inputs. Stored as raw (mm_embeds, is_mm_embed) so that
|
||||
|
||||
@@ -12,7 +12,6 @@ from vllm.v1.kv_cache_interface import KVCacheConfig
|
||||
from vllm.v1.worker.gpu.attn_utils import build_attn_metadata
|
||||
from vllm.v1.worker.gpu.input_batch import InputBatch
|
||||
from vllm.v1.worker.gpu.mm.encoder_cache import EncoderCache
|
||||
from vllm.v1.worker.gpu.mm.encoder_runner import EncoderRunner
|
||||
from vllm.v1.worker.gpu.mm.rope import get_rope_state
|
||||
from vllm.v1.worker.gpu.model_states.interface import ModelState
|
||||
from vllm.v1.worker.gpu.model_states.mm_pruning import maybe_create_mm_pruner
|
||||
@@ -28,30 +27,7 @@ class DefaultModelState(ModelState):
|
||||
encoder_cache: EncoderCache | None,
|
||||
device: torch.device,
|
||||
):
|
||||
self.vllm_config = vllm_config
|
||||
self.model_config = vllm_config.model_config
|
||||
self.scheduler_config = vllm_config.scheduler_config
|
||||
self.model = model
|
||||
self.device = device
|
||||
|
||||
self.supports_mm_inputs = encoder_cache is not None
|
||||
self.max_model_len = self.model_config.max_model_len
|
||||
self.max_num_reqs = self.scheduler_config.max_num_seqs
|
||||
self.max_num_tokens = self.scheduler_config.max_num_batched_tokens
|
||||
self.inputs_embeds_size = self.model_config.get_inputs_embeds_size()
|
||||
self.dtype = self.model_config.dtype
|
||||
|
||||
if self.supports_mm_inputs:
|
||||
assert encoder_cache is not None
|
||||
self.encoder_cache = encoder_cache
|
||||
self.encoder_runner = EncoderRunner(
|
||||
model=self.model,
|
||||
max_num_tokens=self.max_num_tokens,
|
||||
hidden_size=self.inputs_embeds_size,
|
||||
encoder_cache=encoder_cache,
|
||||
dtype=self.dtype,
|
||||
device=self.device,
|
||||
)
|
||||
super().__init__(vllm_config, model, encoder_cache, device)
|
||||
|
||||
self.rope_state = get_rope_state(
|
||||
self.model_config,
|
||||
|
||||
@@ -13,7 +13,6 @@ from vllm.v1.kv_cache_interface import CrossAttentionSpec, KVCacheConfig
|
||||
from vllm.v1.worker.gpu.attn_utils import build_attn_metadata
|
||||
from vllm.v1.worker.gpu.input_batch import InputBatch
|
||||
from vllm.v1.worker.gpu.mm.encoder_cache import EncoderCache
|
||||
from vllm.v1.worker.gpu.mm.encoder_runner import EncoderRunner
|
||||
from vllm.v1.worker.gpu.model_states.interface import (
|
||||
ModelSpecificAttnMetadata,
|
||||
ModelState,
|
||||
@@ -53,25 +52,8 @@ class EncoderDecoderModelState(ModelState):
|
||||
encoder_cache: EncoderCache | None,
|
||||
device: torch.device,
|
||||
) -> None:
|
||||
self.vllm_config = vllm_config
|
||||
self.model_config = vllm_config.model_config
|
||||
self.scheduler_config = vllm_config.scheduler_config
|
||||
self.model = model
|
||||
self.max_num_reqs = vllm_config.scheduler_config.max_num_seqs
|
||||
self.max_num_tokens = self.scheduler_config.max_num_batched_tokens
|
||||
self.max_model_len = self.model_config.max_model_len
|
||||
self.device = device
|
||||
|
||||
assert encoder_cache is not None
|
||||
self.encoder_cache = encoder_cache
|
||||
self.encoder_runner = EncoderRunner(
|
||||
model=self.model,
|
||||
max_num_tokens=self.max_num_tokens,
|
||||
hidden_size=self.model_config.get_inputs_embeds_size(),
|
||||
encoder_cache=self.encoder_cache,
|
||||
dtype=self.model_config.dtype,
|
||||
device=self.device,
|
||||
)
|
||||
super().__init__(vllm_config, model, encoder_cache, device)
|
||||
|
||||
self.max_encoder_len = getattr(
|
||||
self.model_config.hf_config,
|
||||
|
||||
@@ -37,7 +37,6 @@ class ModelSpecificAttnMetadata:
|
||||
|
||||
|
||||
class ModelState(ABC):
|
||||
@abstractmethod
|
||||
def __init__(
|
||||
self,
|
||||
vllm_config: VllmConfig,
|
||||
@@ -45,11 +44,29 @@ class ModelState(ABC):
|
||||
encoder_cache: EncoderCache | None,
|
||||
device: torch.device,
|
||||
) -> None:
|
||||
raise NotImplementedError
|
||||
self.vllm_config = vllm_config
|
||||
self.model_config = vllm_config.model_config
|
||||
self.scheduler_config = vllm_config.scheduler_config
|
||||
self.model = model
|
||||
self.device = device
|
||||
|
||||
model: nn.Module
|
||||
# Set by mm-capable states; used by the default gather_mm_embeddings().
|
||||
encoder_runner: EncoderRunner
|
||||
self.max_model_len = self.model_config.max_model_len
|
||||
self.max_num_reqs = self.scheduler_config.max_num_seqs
|
||||
self.max_num_tokens = self.scheduler_config.max_num_batched_tokens
|
||||
self.inputs_embeds_size = self.model_config.get_inputs_embeds_size()
|
||||
self.dtype = self.model_config.dtype
|
||||
|
||||
self.supports_mm_inputs = encoder_cache is not None
|
||||
if encoder_cache is not None:
|
||||
self.encoder_cache = encoder_cache
|
||||
self.encoder_runner = EncoderRunner(
|
||||
model=self.model,
|
||||
max_num_tokens=self.max_num_tokens,
|
||||
hidden_size=self.inputs_embeds_size,
|
||||
encoder_cache=encoder_cache,
|
||||
dtype=self.dtype,
|
||||
device=self.device,
|
||||
)
|
||||
|
||||
def get_supported_generation_tasks(self) -> tuple[GenerationTask, ...]:
|
||||
from vllm.model_executor.models.interfaces import (
|
||||
|
||||
Reference in New Issue
Block a user