diff --git a/tests/plugins/vllm_add_dummy_model/vllm_add_dummy_model/my_llava.py b/tests/plugins/vllm_add_dummy_model/vllm_add_dummy_model/my_llava.py index da97cf7e2b40b..b431ad1ed0928 100644 --- a/tests/plugins/vllm_add_dummy_model/vllm_add_dummy_model/my_llava.py +++ b/tests/plugins/vllm_add_dummy_model/vllm_add_dummy_model/my_llava.py @@ -9,7 +9,6 @@ from vllm.model_executor.models.llava import (LlavaDummyInputsBuilder, LlavaForConditionalGeneration, LlavaMultiModalProcessor, LlavaProcessingInfo) -from vllm.model_executor.sampling_metadata import SamplingMetadata from vllm.multimodal import MULTIMODAL_REGISTRY @@ -18,11 +17,10 @@ from vllm.multimodal import MULTIMODAL_REGISTRY dummy_inputs=LlavaDummyInputsBuilder) class MyLlava(LlavaForConditionalGeneration): - def compute_logits( - self, hidden_states: torch.Tensor, - sampling_metadata: SamplingMetadata) -> Optional[torch.Tensor]: + def compute_logits(self, + hidden_states: torch.Tensor) -> Optional[torch.Tensor]: # this dummy model always predicts the first token - logits = super().compute_logits(hidden_states, sampling_metadata) + logits = super().compute_logits(hidden_states) if logits is not None: logits.zero_() logits[:, 0] += 1.0 diff --git a/tests/plugins/vllm_add_dummy_model/vllm_add_dummy_model/my_opt.py b/tests/plugins/vllm_add_dummy_model/vllm_add_dummy_model/my_opt.py index 8c34407e3e071..a6fafff98e9c5 100644 --- a/tests/plugins/vllm_add_dummy_model/vllm_add_dummy_model/my_opt.py +++ b/tests/plugins/vllm_add_dummy_model/vllm_add_dummy_model/my_opt.py @@ -6,16 +6,14 @@ from typing import Optional import torch from vllm.model_executor.models.opt import OPTForCausalLM -from vllm.model_executor.sampling_metadata import SamplingMetadata class MyOPTForCausalLM(OPTForCausalLM): - def compute_logits( - self, hidden_states: torch.Tensor, - sampling_metadata: SamplingMetadata) -> Optional[torch.Tensor]: + def compute_logits(self, + hidden_states: torch.Tensor) -> Optional[torch.Tensor]: # this dummy model always predicts the first token - logits = super().compute_logits(hidden_states, sampling_metadata) + logits = super().compute_logits(hidden_states) if logits is not None: logits.zero_() logits[:, 0] += 1.0 diff --git a/vllm/model_executor/__init__.py b/vllm/model_executor/__init__.py index a59aebfac4ff9..3c094cfdb553f 100644 --- a/vllm/model_executor/__init__.py +++ b/vllm/model_executor/__init__.py @@ -3,11 +3,9 @@ from vllm.model_executor.parameter import (BasevLLMParameter, PackedvLLMParameter) -from vllm.model_executor.sampling_metadata import SamplingMetadata from vllm.model_executor.utils import set_random_seed __all__ = [ - "SamplingMetadata", "set_random_seed", "BasevLLMParameter", "PackedvLLMParameter", diff --git a/vllm/model_executor/layers/logits_processor.py b/vllm/model_executor/layers/logits_processor.py index 8226437cb1898..2110aa2769b93 100644 --- a/vllm/model_executor/layers/logits_processor.py +++ b/vllm/model_executor/layers/logits_processor.py @@ -10,7 +10,6 @@ from vllm.distributed import (tensor_model_parallel_all_gather, from vllm.model_executor.custom_op import CustomOp from vllm.model_executor.layers.vocab_parallel_embedding import ( VocabParallelEmbedding) -from vllm.model_executor.sampling_metadata import SamplingMetadata from vllm.platforms import current_platform @@ -50,7 +49,6 @@ class LogitsProcessor(CustomOp): self, lm_head: VocabParallelEmbedding, hidden_states: torch.Tensor, - sampling_metadata: Optional[SamplingMetadata] = None, embedding_bias: Optional[torch.Tensor] = None, ) -> Optional[torch.Tensor]: if self.logits_as_input: diff --git a/vllm/model_executor/models/apertus.py b/vllm/model_executor/models/apertus.py index f6400b05e110a..6dab4ed14345f 100644 --- a/vllm/model_executor/models/apertus.py +++ b/vllm/model_executor/models/apertus.py @@ -48,7 +48,6 @@ from vllm.model_executor.layers.vocab_parallel_embedding import ( DEFAULT_VOCAB_PADDING_SIZE, ParallelLMHead, VocabParallelEmbedding) from vllm.model_executor.model_loader.weight_utils import ( default_weight_loader, maybe_remap_kv_scale_name) -from vllm.model_executor.sampling_metadata import SamplingMetadata from vllm.sequence import IntermediateTensors from .interfaces import SupportsLoRA, SupportsPP @@ -566,10 +565,8 @@ class ApertusForCausalLM(nn.Module, SupportsLoRA, SupportsPP): def compute_logits( self, hidden_states: torch.Tensor, - sampling_metadata: SamplingMetadata, ) -> Optional[torch.Tensor]: - logits = self.logits_processor(self.lm_head, hidden_states, - sampling_metadata) + logits = self.logits_processor(self.lm_head, hidden_states) return logits def load_weights(self, weights: Iterable[tuple[str, diff --git a/vllm/model_executor/models/arcee.py b/vllm/model_executor/models/arcee.py index be82c2fd59644..1ee378af76c9f 100644 --- a/vllm/model_executor/models/arcee.py +++ b/vllm/model_executor/models/arcee.py @@ -399,11 +399,10 @@ class ArceeForCausalLM(nn.Module, SupportsLoRA, SupportsPP): inputs_embeds=inputs_embeds) return model_output - def compute_logits(self, hidden_states: torch.Tensor, - sampling_metadata) -> Optional[torch.Tensor]: + def compute_logits(self, + hidden_states: torch.Tensor) -> Optional[torch.Tensor]: # Compute final logits from hidden states (last pipeline rank only) - logits = self.logits_processor(self.lm_head, hidden_states, - sampling_metadata) + logits = self.logits_processor(self.lm_head, hidden_states) return logits def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor: diff --git a/vllm/model_executor/models/arctic.py b/vllm/model_executor/models/arctic.py index b6dd559968415..55d16fd75cebb 100644 --- a/vllm/model_executor/models/arctic.py +++ b/vllm/model_executor/models/arctic.py @@ -30,7 +30,6 @@ from vllm.model_executor.layers.rotary_embedding import get_rope from vllm.model_executor.layers.vocab_parallel_embedding import ( ParallelLMHead, VocabParallelEmbedding) from vllm.model_executor.model_loader.weight_utils import default_weight_loader -from vllm.model_executor.sampling_metadata import SamplingMetadata from vllm.model_executor.utils import set_weight_attrs from vllm.platforms import current_platform from vllm.sequence import IntermediateTensors @@ -456,10 +455,8 @@ class ArcticForCausalLM(nn.Module, SupportsPP, SupportsQuant): def compute_logits( self, hidden_states: torch.Tensor, - sampling_metadata: SamplingMetadata, ) -> Optional[torch.Tensor]: - logits = self.logits_processor(self.lm_head, hidden_states, - sampling_metadata) + logits = self.logits_processor(self.lm_head, hidden_states) return logits def load_weights(self, weights: Iterable[tuple[str, diff --git a/vllm/model_executor/models/aria.py b/vllm/model_executor/models/aria.py index a7cb6b35a4ab4..35c1adbdd00b6 100644 --- a/vllm/model_executor/models/aria.py +++ b/vllm/model_executor/models/aria.py @@ -19,7 +19,6 @@ from vllm.model_executor.layers.logits_processor import LogitsProcessor from vllm.model_executor.layers.vocab_parallel_embedding import ParallelLMHead from vllm.model_executor.model_loader.weight_utils import ( default_weight_loader, maybe_remap_kv_scale_name) -from vllm.model_executor.sampling_metadata import SamplingMetadata from vllm.multimodal import MULTIMODAL_REGISTRY from vllm.multimodal.inputs import (MultiModalDataDict, MultiModalFieldConfig, MultiModalKwargsItems) @@ -644,10 +643,8 @@ class AriaForConditionalGeneration(nn.Module, SupportsMultiModal): return hidden_states - def compute_logits(self, hidden_states: torch.Tensor, - sampling_metadata: SamplingMetadata) -> torch.Tensor: - logits = self.logits_processor(self.lm_head, hidden_states, - sampling_metadata) + def compute_logits(self, hidden_states: torch.Tensor) -> torch.Tensor: + logits = self.logits_processor(self.lm_head, hidden_states) return logits def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]): diff --git a/vllm/model_executor/models/aya_vision.py b/vllm/model_executor/models/aya_vision.py index 687c82ded9d0a..0f05f9b4efcd6 100644 --- a/vllm/model_executor/models/aya_vision.py +++ b/vllm/model_executor/models/aya_vision.py @@ -16,7 +16,6 @@ from transformers.models.got_ocr2.image_processing_got_ocr2 import ( get_optimal_tiled_canvas) from vllm.config import VllmConfig -from vllm.model_executor.sampling_metadata import SamplingMetadata from vllm.multimodal import MULTIMODAL_REGISTRY from vllm.multimodal.inputs import MultiModalDataDict, MultiModalKwargsItems from vllm.multimodal.parse import (ImageProcessorItems, ImageSize, @@ -464,7 +463,5 @@ class AyaVisionForConditionalGeneration(nn.Module, SupportsMultiModal, def compute_logits( self, hidden_states: torch.Tensor, - sampling_metadata: SamplingMetadata, ) -> Optional[torch.Tensor]: - return self.language_model.compute_logits(hidden_states, - sampling_metadata) + return self.language_model.compute_logits(hidden_states) diff --git a/vllm/model_executor/models/baichuan.py b/vllm/model_executor/models/baichuan.py index ae25033410407..db8d0a8710471 100644 --- a/vllm/model_executor/models/baichuan.py +++ b/vllm/model_executor/models/baichuan.py @@ -46,7 +46,6 @@ from vllm.model_executor.layers.vocab_parallel_embedding import ( ParallelLMHead, VocabParallelEmbedding) from vllm.model_executor.model_loader.weight_utils import ( default_weight_loader, row_parallel_weight_loader) -from vllm.model_executor.sampling_metadata import SamplingMetadata from vllm.sequence import IntermediateTensors from .interfaces import SupportsLoRA, SupportsPP, SupportsQuant @@ -421,10 +420,8 @@ class BaiChuanBaseForCausalLM(nn.Module, SupportsLoRA, SupportsPP, def compute_logits( self, hidden_states: torch.Tensor, - sampling_metadata: SamplingMetadata, ) -> Optional[torch.Tensor]: - logits = self.logits_processor(self.lm_head, hidden_states, - sampling_metadata) + logits = self.logits_processor(self.lm_head, hidden_states) return logits def load_weights(self, weights: Iterable[tuple[str, diff --git a/vllm/model_executor/models/bailing_moe.py b/vllm/model_executor/models/bailing_moe.py index 5f6025abf315c..82cd4a26a1baa 100644 --- a/vllm/model_executor/models/bailing_moe.py +++ b/vllm/model_executor/models/bailing_moe.py @@ -51,7 +51,6 @@ from vllm.model_executor.layers.rotary_embedding import get_rope from vllm.model_executor.layers.vocab_parallel_embedding import ( ParallelLMHead, VocabParallelEmbedding) from vllm.model_executor.model_loader.weight_utils import default_weight_loader -from vllm.model_executor.sampling_metadata import SamplingMetadata from vllm.sequence import IntermediateTensors from .interfaces import SupportsLoRA, SupportsPP @@ -623,10 +622,8 @@ class BailingMoeForCausalLM(nn.Module, SupportsPP, SupportsLoRA): def compute_logits( self, hidden_states: torch.Tensor, - sampling_metadata: SamplingMetadata, ) -> Optional[torch.Tensor]: - logits = self.logits_processor(self.lm_head, hidden_states, - sampling_metadata) + logits = self.logits_processor(self.lm_head, hidden_states) return logits def load_weights(self, weights: Iterable[tuple[str, diff --git a/vllm/model_executor/models/bamba.py b/vllm/model_executor/models/bamba.py index 397089f31cdf6..584981ef3ebfd 100644 --- a/vllm/model_executor/models/bamba.py +++ b/vllm/model_executor/models/bamba.py @@ -34,7 +34,6 @@ from vllm.model_executor.layers.vocab_parallel_embedding import ( from vllm.model_executor.model_loader.weight_utils import default_weight_loader from vllm.model_executor.models.mamba_cache import (MambaCacheManager, MambaCacheParams) -from vllm.model_executor.sampling_metadata import SamplingMetadata from vllm.sequence import IntermediateTensors from vllm.utils import LayerBlockType @@ -571,10 +570,8 @@ class BambaForCausalLM(nn.Module, HasInnerState, SupportsLoRA, SupportsPP, def compute_logits( self, hidden_states: torch.Tensor, - sampling_metadata: SamplingMetadata, ) -> Optional[torch.Tensor]: - logits = self.logits_processor(self.lm_head, hidden_states, - sampling_metadata) + logits = self.logits_processor(self.lm_head, hidden_states) return logits def load_weights(self, weights: Iterable[tuple[str, diff --git a/vllm/model_executor/models/blip2.py b/vllm/model_executor/models/blip2.py index a3131aa3812ef..b7455fba62c02 100644 --- a/vllm/model_executor/models/blip2.py +++ b/vllm/model_executor/models/blip2.py @@ -12,7 +12,6 @@ from transformers import (BatchFeature, Blip2Config, Blip2QFormerConfig, from vllm.config import CacheConfig, VllmConfig from vllm.model_executor.layers.activation import get_act_fn from vllm.model_executor.layers.quantization import QuantizationConfig -from vllm.model_executor.sampling_metadata import SamplingMetadata from vllm.multimodal import MULTIMODAL_REGISTRY from vllm.multimodal.inputs import (MultiModalDataDict, MultiModalFieldConfig, MultiModalKwargsItems) @@ -704,10 +703,8 @@ class Blip2ForConditionalGeneration(nn.Module, SupportsMultiModal, SupportsPP, def compute_logits( self, hidden_states: torch.Tensor, - sampling_metadata: SamplingMetadata, ) -> Optional[torch.Tensor]: - return self.language_model.compute_logits(hidden_states, - sampling_metadata) + return self.language_model.compute_logits(hidden_states) def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]: diff --git a/vllm/model_executor/models/bloom.py b/vllm/model_executor/models/bloom.py index 4c37622b049c8..30816f72a2678 100644 --- a/vllm/model_executor/models/bloom.py +++ b/vllm/model_executor/models/bloom.py @@ -41,7 +41,6 @@ from vllm.model_executor.layers.quantization import QuantizationConfig from vllm.model_executor.layers.vocab_parallel_embedding import ( ParallelLMHead, VocabParallelEmbedding) from vllm.model_executor.model_loader.weight_utils import default_weight_loader -from vllm.model_executor.sampling_metadata import SamplingMetadata from vllm.sequence import IntermediateTensors from .interfaces import SupportsPP, SupportsQuant @@ -355,10 +354,8 @@ class BloomForCausalLM(nn.Module, SupportsPP, SupportsQuant): def compute_logits( self, hidden_states: torch.Tensor, - sampling_metadata: SamplingMetadata, ) -> Optional[torch.Tensor]: - logits = self.logits_processor(self.lm_head, hidden_states, - sampling_metadata) + logits = self.logits_processor(self.lm_head, hidden_states) return logits def load_weights(self, weights: Iterable[tuple[str, diff --git a/vllm/model_executor/models/chameleon.py b/vllm/model_executor/models/chameleon.py index 7a56236483749..79d648d749c6a 100644 --- a/vllm/model_executor/models/chameleon.py +++ b/vllm/model_executor/models/chameleon.py @@ -28,7 +28,6 @@ from vllm.model_executor.layers.vocab_parallel_embedding import ( ParallelLMHead, VocabParallelEmbedding) from vllm.model_executor.model_loader.weight_utils import ( default_weight_loader, row_parallel_weight_loader) -from vllm.model_executor.sampling_metadata import SamplingMetadata from vllm.model_executor.utils import set_weight_attrs from vllm.multimodal import MULTIMODAL_REGISTRY from vllm.multimodal.inputs import (MultiModalDataDict, MultiModalFieldConfig, @@ -1046,10 +1045,8 @@ class ChameleonForConditionalGeneration(nn.Module, SupportsMultiModal, def compute_logits( self, hidden_states: torch.Tensor, - sampling_metadata: SamplingMetadata, ) -> Optional[torch.Tensor]: - logits = self.logits_processor(self.lm_head, hidden_states, - sampling_metadata) + logits = self.logits_processor(self.lm_head, hidden_states) # Disallow image tokens which does not include special # begin-image and end-image tokens diff --git a/vllm/model_executor/models/chatglm.py b/vllm/model_executor/models/chatglm.py index 1fc2da3e4d7ca..879508400222f 100644 --- a/vllm/model_executor/models/chatglm.py +++ b/vllm/model_executor/models/chatglm.py @@ -27,7 +27,6 @@ from vllm.model_executor.layers.rotary_embedding import get_rope from vllm.model_executor.layers.vocab_parallel_embedding import ( ParallelLMHead, VocabParallelEmbedding) from vllm.model_executor.model_loader.weight_utils import default_weight_loader -from vllm.model_executor.sampling_metadata import SamplingMetadata from vllm.sequence import IntermediateTensors from vllm.transformers_utils.configs import ChatGLMConfig @@ -437,10 +436,8 @@ class ChatGLMBaseModel(nn.Module): def compute_logits( self, hidden_states: torch.Tensor, - sampling_metadata: SamplingMetadata, ) -> Optional[torch.Tensor]: - logits = self.logits_processor(self.lm_head, hidden_states, - sampling_metadata) + logits = self.logits_processor(self.lm_head, hidden_states) return logits def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]): diff --git a/vllm/model_executor/models/cohere2_vision.py b/vllm/model_executor/models/cohere2_vision.py index 179cc2af8eb3f..6d67eb68d51a8 100644 --- a/vllm/model_executor/models/cohere2_vision.py +++ b/vllm/model_executor/models/cohere2_vision.py @@ -21,7 +21,6 @@ from vllm.model_executor.layers.linear import (MergedColumnParallelLinear, RowParallelLinear) from vllm.model_executor.layers.quantization import QuantizationConfig from vllm.model_executor.layers.quantization.awq import AWQConfig -from vllm.model_executor.sampling_metadata import SamplingMetadata from vllm.multimodal import MULTIMODAL_REGISTRY from vllm.multimodal.inputs import MultiModalDataDict, MultiModalKwargsItems from vllm.multimodal.parse import (ImageProcessorItems, ImageSize, @@ -478,7 +477,5 @@ class Cohere2VisionForConditionalGeneration(nn.Module, SupportsMultiModal, def compute_logits( self, hidden_states: torch.Tensor, - sampling_metadata: SamplingMetadata, ) -> Optional[torch.Tensor]: - return self.language_model.compute_logits(hidden_states, - sampling_metadata) + return self.language_model.compute_logits(hidden_states) diff --git a/vllm/model_executor/models/commandr.py b/vllm/model_executor/models/commandr.py index 7f87e31abdcd3..f3929ef3b5938 100644 --- a/vllm/model_executor/models/commandr.py +++ b/vllm/model_executor/models/commandr.py @@ -46,7 +46,6 @@ from vllm.model_executor.layers.vocab_parallel_embedding import ( from vllm.model_executor.model_loader.weight_utils import ( default_weight_loader, maybe_remap_kv_scale_name, row_parallel_weight_loader) -from vllm.model_executor.sampling_metadata import SamplingMetadata from vllm.model_executor.utils import set_weight_attrs from vllm.platforms import current_platform from vllm.sequence import IntermediateTensors @@ -448,15 +447,14 @@ class CohereForCausalLM(nn.Module, SupportsLoRA, SupportsPP, SupportsQuant): def compute_logits( self, hidden_states: torch.Tensor, - sampling_metadata: SamplingMetadata, ) -> Optional[torch.Tensor]: is_not_lora = hasattr(self.model.embed_tokens, 'weight') if is_not_lora: logits = self.logits_processor(self.model.embed_tokens, - hidden_states, sampling_metadata) + hidden_states) else: logits = self.logits_processor(self.model.embed_tokens.base_layer, - hidden_states, sampling_metadata) + hidden_states) return logits diff --git a/vllm/model_executor/models/dbrx.py b/vllm/model_executor/models/dbrx.py index 003cf4563a22f..f863b1da5505c 100644 --- a/vllm/model_executor/models/dbrx.py +++ b/vllm/model_executor/models/dbrx.py @@ -24,7 +24,6 @@ from vllm.model_executor.layers.vocab_parallel_embedding import ( DEFAULT_VOCAB_PADDING_SIZE, ParallelLMHead, VocabParallelEmbedding) from vllm.model_executor.model_loader.weight_utils import ( default_weight_loader, maybe_remap_kv_scale_name) -from vllm.model_executor.sampling_metadata import SamplingMetadata from vllm.sequence import IntermediateTensors from .interfaces import SupportsPP @@ -462,10 +461,8 @@ class DbrxForCausalLM(nn.Module, SupportsPP): def compute_logits( self, hidden_states: torch.Tensor, - sampling_metadata: SamplingMetadata, ) -> Optional[torch.Tensor]: - logits = self.logits_processor(self.lm_head, hidden_states, - sampling_metadata) + logits = self.logits_processor(self.lm_head, hidden_states) return logits def load_weights(self, weights: Iterable[tuple[str, diff --git a/vllm/model_executor/models/deepseek.py b/vllm/model_executor/models/deepseek.py index 59c9921881497..ffc843fe033cc 100644 --- a/vllm/model_executor/models/deepseek.py +++ b/vllm/model_executor/models/deepseek.py @@ -49,7 +49,6 @@ from vllm.model_executor.layers.rotary_embedding import get_rope from vllm.model_executor.layers.vocab_parallel_embedding import ( ParallelLMHead, VocabParallelEmbedding) from vllm.model_executor.model_loader.weight_utils import default_weight_loader -from vllm.model_executor.sampling_metadata import SamplingMetadata from vllm.sequence import IntermediateTensors from .interfaces import SupportsLoRA, SupportsPP @@ -488,10 +487,8 @@ class DeepseekForCausalLM(nn.Module, SupportsLoRA, SupportsPP): def compute_logits( self, hidden_states: torch.Tensor, - sampling_metadata: SamplingMetadata, ) -> Optional[torch.Tensor]: - logits = self.logits_processor(self.lm_head, hidden_states, - sampling_metadata) + logits = self.logits_processor(self.lm_head, hidden_states) return logits def load_weights(self, weights: Iterable[tuple[str, diff --git a/vllm/model_executor/models/deepseek_eagle.py b/vllm/model_executor/models/deepseek_eagle.py index 2770ddebc48ab..ed7e7614800fc 100644 --- a/vllm/model_executor/models/deepseek_eagle.py +++ b/vllm/model_executor/models/deepseek_eagle.py @@ -19,7 +19,6 @@ from vllm.model_executor.model_loader.weight_utils import ( default_weight_loader, maybe_remap_kv_scale_name) from vllm.model_executor.models.deepseek_v2 import (DeepseekV2DecoderLayer, DeepseekV3ForCausalLM) -from vllm.model_executor.sampling_metadata import SamplingMetadata from .utils import AutoWeightsLoader, maybe_prefix @@ -222,10 +221,8 @@ class EagleDeepseekV3ForCausalLM(DeepseekV3ForCausalLM): def compute_logits( self, hidden_states: torch.Tensor, - sampling_metadata: SamplingMetadata, ) -> Optional[torch.Tensor]: - logits = self.logits_processor(self.lm_head, hidden_states, - sampling_metadata) + logits = self.logits_processor(self.lm_head, hidden_states) return logits def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]): diff --git a/vllm/model_executor/models/deepseek_mtp.py b/vllm/model_executor/models/deepseek_mtp.py index 8fbf16d206a86..92f311ab465b5 100644 --- a/vllm/model_executor/models/deepseek_mtp.py +++ b/vllm/model_executor/models/deepseek_mtp.py @@ -15,7 +15,6 @@ from vllm.model_executor.layers.quantization import QuantizationConfig from vllm.model_executor.layers.vocab_parallel_embedding import ( ParallelLMHead, VocabParallelEmbedding) from vllm.model_executor.model_loader.weight_utils import default_weight_loader -from vllm.model_executor.sampling_metadata import SamplingMetadata from vllm.sequence import IntermediateTensors from .deepseek_v2 import (DeepseekV2DecoderLayer, @@ -124,15 +123,13 @@ class DeepSeekMultiTokenPredictor(nn.Module): def compute_logits( self, hidden_states: torch.Tensor, - sampling_metadata: SamplingMetadata, spec_step_idx: int = 0, ) -> torch.Tensor: current_step_idx = (spec_step_idx % self.num_mtp_layers) mtp_layer = self.layers[str(self.mtp_start_layer_idx + current_step_idx)] logits = self.logits_processor(mtp_layer.shared_head.head, - mtp_layer.shared_head(hidden_states), - sampling_metadata) + mtp_layer.shared_head(hidden_states)) return logits @@ -161,11 +158,9 @@ class DeepSeekMTP(nn.Module, SupportsPP): def compute_logits( self, hidden_states: torch.Tensor, - sampling_metadata: SamplingMetadata, spec_step_idx: int = 0, ) -> Optional[torch.Tensor]: - return self.model.compute_logits(hidden_states, sampling_metadata, - spec_step_idx) + return self.model.compute_logits(hidden_states, spec_step_idx) def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]: diff --git a/vllm/model_executor/models/deepseek_v2.py b/vllm/model_executor/models/deepseek_v2.py index 636554bd648f2..a99a6679a5696 100644 --- a/vllm/model_executor/models/deepseek_v2.py +++ b/vllm/model_executor/models/deepseek_v2.py @@ -56,7 +56,6 @@ from vllm.model_executor.layers.vocab_parallel_embedding import ( ParallelLMHead, VocabParallelEmbedding) from vllm.model_executor.model_loader.weight_utils import ( default_weight_loader, maybe_remap_kv_scale_name) -from vllm.model_executor.sampling_metadata import SamplingMetadata from vllm.platforms import current_platform from vllm.sequence import IntermediateTensors from vllm.utils import cdiv, direct_register_custom_op @@ -914,10 +913,8 @@ class DeepseekV2ForCausalLM(nn.Module, SupportsPP, MixtureOfExperts, def compute_logits( self, hidden_states: torch.Tensor, - sampling_metadata: SamplingMetadata, ) -> Optional[torch.Tensor]: - logits = self.logits_processor(self.lm_head, hidden_states, - sampling_metadata) + logits = self.logits_processor(self.lm_head, hidden_states) return logits def load_weights(self, weights: Iterable[tuple[str, diff --git a/vllm/model_executor/models/deepseek_vl2.py b/vllm/model_executor/models/deepseek_vl2.py index d7ae8206baca5..c8ed759d2e972 100644 --- a/vllm/model_executor/models/deepseek_vl2.py +++ b/vllm/model_executor/models/deepseek_vl2.py @@ -15,7 +15,6 @@ from transformers import BatchFeature from vllm.config import VllmConfig from vllm.distributed import get_tensor_model_parallel_world_size -from vllm.model_executor import SamplingMetadata from vllm.model_executor.layers.quantization import QuantizationConfig from vllm.model_executor.model_loader.utils import set_default_torch_dtype from vllm.model_executor.models.transformers import replace_linear_class @@ -647,10 +646,8 @@ class DeepseekVLV2ForCausalLM(nn.Module, SupportsMultiModal, SupportsPP): def compute_logits( self, hidden_states: torch.Tensor, - sampling_metadata: SamplingMetadata, ) -> Optional[torch.Tensor]: - return self.language_model.compute_logits(hidden_states, - sampling_metadata) + return self.language_model.compute_logits(hidden_states) def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]: diff --git a/vllm/model_executor/models/dots1.py b/vllm/model_executor/models/dots1.py index 20555e48b73d4..2a09234b59ed1 100644 --- a/vllm/model_executor/models/dots1.py +++ b/vllm/model_executor/models/dots1.py @@ -52,7 +52,6 @@ from vllm.model_executor.layers.vocab_parallel_embedding import ( ParallelLMHead, VocabParallelEmbedding) from vllm.model_executor.model_loader.weight_utils import ( default_weight_loader, maybe_remap_kv_scale_name) -from vllm.model_executor.sampling_metadata import SamplingMetadata from vllm.sequence import IntermediateTensors from .interfaces import SupportsLoRA, SupportsPP @@ -534,10 +533,8 @@ class Dots1ForCausalLM(nn.Module, SupportsPP, SupportsLoRA): def compute_logits( self, hidden_states: torch.Tensor, - sampling_metadata: SamplingMetadata, ) -> Optional[torch.Tensor]: - logits = self.logits_processor(self.lm_head, hidden_states, - sampling_metadata) + logits = self.logits_processor(self.lm_head, hidden_states) return logits def load_weights(self, weights: Iterable[tuple[str, diff --git a/vllm/model_executor/models/ernie45_moe.py b/vllm/model_executor/models/ernie45_moe.py index ebab018ed67e7..d262e9e9da50e 100644 --- a/vllm/model_executor/models/ernie45_moe.py +++ b/vllm/model_executor/models/ernie45_moe.py @@ -49,7 +49,6 @@ from vllm.model_executor.layers.vocab_parallel_embedding import ( ParallelLMHead, VocabParallelEmbedding) from vllm.model_executor.model_loader.weight_utils import ( default_weight_loader, maybe_remap_kv_scale_name) -from vllm.model_executor.sampling_metadata import SamplingMetadata from vllm.sequence import IntermediateTensors from .interfaces import SupportsLoRA, SupportsPP @@ -591,10 +590,8 @@ class Ernie4_5_MoeForCausalLM(nn.Module, SupportsPP, SupportsLoRA): def compute_logits( self, hidden_states: torch.Tensor, - sampling_metadata: SamplingMetadata, ) -> Optional[torch.Tensor]: - logits = self.logits_processor(self.lm_head, hidden_states, - sampling_metadata) + logits = self.logits_processor(self.lm_head, hidden_states) return logits def load_weights(self, weights: Iterable[tuple[str, diff --git a/vllm/model_executor/models/ernie45_vl.py b/vllm/model_executor/models/ernie45_vl.py index 0d4aced93ca1c..74b358034ef3d 100644 --- a/vllm/model_executor/models/ernie45_vl.py +++ b/vllm/model_executor/models/ernie45_vl.py @@ -39,7 +39,6 @@ from vllm.config import VllmConfig from vllm.distributed import parallel_state from vllm.distributed import utils as dist_utils from vllm.logger import init_logger -from vllm.model_executor import SamplingMetadata from vllm.model_executor.layers.activation import QuickGELU from vllm.model_executor.layers.layernorm import RMSNorm from vllm.model_executor.layers.linear import (ColumnParallelLinear, @@ -1292,11 +1291,9 @@ class Ernie4_5_VLMoeForConditionalGeneration(nn.Module, SupportsMultiModal, def compute_logits( self, hidden_states: torch.Tensor, - sampling_metadata: SamplingMetadata, ) -> Optional[torch.Tensor]: """compute logits""" - return self.language_model.compute_logits(hidden_states, - sampling_metadata) + return self.language_model.compute_logits(hidden_states) def _vision_forward( self, diff --git a/vllm/model_executor/models/ernie45_vl_moe.py b/vllm/model_executor/models/ernie45_vl_moe.py index 7f791852ceb91..f55016f7ccb36 100644 --- a/vllm/model_executor/models/ernie45_vl_moe.py +++ b/vllm/model_executor/models/ernie45_vl_moe.py @@ -48,7 +48,6 @@ from vllm.model_executor.layers.vocab_parallel_embedding import ( ParallelLMHead, VocabParallelEmbedding) from vllm.model_executor.model_loader.weight_utils import ( default_weight_loader, maybe_remap_kv_scale_name) -from vllm.model_executor.sampling_metadata import SamplingMetadata from vllm.sequence import IntermediateTensors from .ernie45_moe import Ernie4_5_MoeMLP @@ -587,10 +586,8 @@ class Ernie4_5_VLMoeForCausalLM(nn.Module, SupportsPP): def compute_logits( self, hidden_states: torch.Tensor, - sampling_metadata: SamplingMetadata, ) -> Optional[torch.Tensor]: - logits = self.logits_processor(self.lm_head, hidden_states, - sampling_metadata) + logits = self.logits_processor(self.lm_head, hidden_states) return logits def load_weights(self, weights: Iterable[tuple[str, diff --git a/vllm/model_executor/models/ernie_mtp.py b/vllm/model_executor/models/ernie_mtp.py index c446265230313..288fbe736c32f 100644 --- a/vllm/model_executor/models/ernie_mtp.py +++ b/vllm/model_executor/models/ernie_mtp.py @@ -36,7 +36,6 @@ from vllm.model_executor.layers.quantization import QuantizationConfig from vllm.model_executor.layers.vocab_parallel_embedding import ( ParallelLMHead, VocabParallelEmbedding) from vllm.model_executor.model_loader.weight_utils import default_weight_loader -from vllm.model_executor.sampling_metadata import SamplingMetadata from vllm.sequence import IntermediateTensors from .interfaces import SupportsPP @@ -138,12 +137,10 @@ class ErnieMultiTokenPredictor(nn.Module): self, hidden_states: torch.Tensor, lm_head: ParallelLMHead, - sampling_metadata: SamplingMetadata, spec_step_idx: int = 0, ) -> torch.Tensor: self.layers[str(self.mtp_start_layer_idx + spec_step_idx)] - logits = self.logits_processor(lm_head, hidden_states, - sampling_metadata) + logits = self.logits_processor(lm_head, hidden_states) return logits @@ -180,11 +177,10 @@ class ErnieMTP(nn.Module, SupportsPP): def compute_logits( self, hidden_states: torch.Tensor, - sampling_metadata: SamplingMetadata, spec_step_idx: int = 0, ) -> Optional[torch.Tensor]: return self.model.compute_logits(hidden_states, self.lm_head, - sampling_metadata, spec_step_idx) + spec_step_idx) def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]: diff --git a/vllm/model_executor/models/exaone.py b/vllm/model_executor/models/exaone.py index f503fb0f9364a..5dafcd595e4a5 100644 --- a/vllm/model_executor/models/exaone.py +++ b/vllm/model_executor/models/exaone.py @@ -49,7 +49,6 @@ from vllm.model_executor.layers.vocab_parallel_embedding import ( DEFAULT_VOCAB_PADDING_SIZE, ParallelLMHead, VocabParallelEmbedding) from vllm.model_executor.model_loader.weight_utils import ( default_weight_loader, maybe_remap_kv_scale_name) -from vllm.model_executor.sampling_metadata import SamplingMetadata from vllm.sequence import IntermediateTensors from .interfaces import SupportsLoRA, SupportsPP @@ -534,10 +533,8 @@ class ExaoneForCausalLM(nn.Module, SupportsLoRA, SupportsPP): def compute_logits( self, hidden_states: torch.Tensor, - sampling_metadata: SamplingMetadata, ) -> Optional[torch.Tensor]: - logits = self.logits_processor(self.lm_head, hidden_states, - sampling_metadata) + logits = self.logits_processor(self.lm_head, hidden_states) return logits def load_weights(self, weights: Iterable[tuple[str, diff --git a/vllm/model_executor/models/exaone4.py b/vllm/model_executor/models/exaone4.py index 9f7d57d938140..c78eedff66700 100644 --- a/vllm/model_executor/models/exaone4.py +++ b/vllm/model_executor/models/exaone4.py @@ -45,7 +45,6 @@ from vllm.model_executor.layers.vocab_parallel_embedding import ( DEFAULT_VOCAB_PADDING_SIZE, ParallelLMHead, VocabParallelEmbedding) from vllm.model_executor.model_loader.weight_utils import ( default_weight_loader, maybe_remap_kv_scale_name) -from vllm.model_executor.sampling_metadata import SamplingMetadata from vllm.sequence import IntermediateTensors from .interfaces import SupportsLoRA, SupportsPP @@ -517,10 +516,8 @@ class Exaone4ForCausalLM(nn.Module, SupportsLoRA, SupportsPP): def compute_logits( self, hidden_states: torch.Tensor, - sampling_metadata: SamplingMetadata, ) -> Optional[torch.Tensor]: - logits = self.logits_processor(self.lm_head, hidden_states, - sampling_metadata) + logits = self.logits_processor(self.lm_head, hidden_states) return logits def load_weights(self, weights: Iterable[tuple[str, diff --git a/vllm/model_executor/models/falcon.py b/vllm/model_executor/models/falcon.py index 42c378e5c389a..0c50056d1c527 100644 --- a/vllm/model_executor/models/falcon.py +++ b/vllm/model_executor/models/falcon.py @@ -46,7 +46,6 @@ from vllm.model_executor.layers.rotary_embedding import get_rope from vllm.model_executor.layers.vocab_parallel_embedding import ( ParallelLMHead, VocabParallelEmbedding) from vllm.model_executor.model_loader.weight_utils import default_weight_loader -from vllm.model_executor.sampling_metadata import SamplingMetadata from vllm.sequence import IntermediateTensors from vllm.transformers_utils.configs import RWConfig @@ -496,10 +495,8 @@ class FalconForCausalLM(nn.Module, SupportsPP): def compute_logits( self, hidden_states: torch.Tensor, - sampling_metadata: SamplingMetadata, ) -> Optional[torch.Tensor]: - logits = self.logits_processor(self.lm_head, hidden_states, - sampling_metadata) + logits = self.logits_processor(self.lm_head, hidden_states) return logits def load_weights(self, weights: Iterable[tuple[str, diff --git a/vllm/model_executor/models/falcon_h1.py b/vllm/model_executor/models/falcon_h1.py index 757051b3b1447..83efdd2e433fc 100644 --- a/vllm/model_executor/models/falcon_h1.py +++ b/vllm/model_executor/models/falcon_h1.py @@ -33,7 +33,6 @@ from vllm.model_executor.layers.vocab_parallel_embedding import ( from vllm.model_executor.model_loader.weight_utils import default_weight_loader from vllm.model_executor.models.mamba_cache import (MambaCacheManager, MambaCacheParams) -from vllm.model_executor.sampling_metadata import SamplingMetadata from vllm.sequence import IntermediateTensors from .interfaces import HasInnerState, IsHybrid, SupportsLoRA, SupportsPP @@ -675,10 +674,8 @@ class FalconH1ForCausalLM(nn.Module, HasInnerState, SupportsLoRA, SupportsPP, def compute_logits( self, hidden_states: torch.Tensor, - sampling_metadata: SamplingMetadata, ) -> Optional[torch.Tensor]: - logits = self.logits_processor(self.lm_head, hidden_states, - sampling_metadata) + logits = self.logits_processor(self.lm_head, hidden_states) return logits diff --git a/vllm/model_executor/models/fuyu.py b/vllm/model_executor/models/fuyu.py index 90af859ab92ec..53e9e6fe6e460 100644 --- a/vllm/model_executor/models/fuyu.py +++ b/vllm/model_executor/models/fuyu.py @@ -29,7 +29,6 @@ from transformers import (BatchFeature, FuyuConfig, FuyuImageProcessor, from vllm.config import VllmConfig from vllm.model_executor.layers.linear import ColumnParallelLinear from vllm.model_executor.models.persimmon import PersimmonForCausalLM -from vllm.model_executor.sampling_metadata import SamplingMetadata from vllm.multimodal import MULTIMODAL_REGISTRY from vllm.multimodal.inputs import (MultiModalDataDict, MultiModalFieldConfig, MultiModalKwargsItems) @@ -389,10 +388,9 @@ class FuyuForCausalLM(nn.Module, SupportsMultiModal, SupportsPP): def compute_logits( self, hidden_states: torch.Tensor, - sampling_metadata: SamplingMetadata, ) -> Optional[torch.Tensor]: logits = self.language_model.logits_processor( - self.language_model.lm_head, hidden_states, sampling_metadata) + self.language_model.lm_head, hidden_states) return logits def load_weights(self, weights: Iterable[tuple[str, diff --git a/vllm/model_executor/models/gemma.py b/vllm/model_executor/models/gemma.py index 12eb27503870c..c19425b6cb6d6 100644 --- a/vllm/model_executor/models/gemma.py +++ b/vllm/model_executor/models/gemma.py @@ -41,7 +41,6 @@ from vllm.model_executor.layers.rotary_embedding import get_rope from vllm.model_executor.layers.vocab_parallel_embedding import ( VocabParallelEmbedding) from vllm.model_executor.model_loader.weight_utils import default_weight_loader -from vllm.model_executor.sampling_metadata import SamplingMetadata from vllm.sequence import IntermediateTensors from .interfaces import SupportsLoRA, SupportsPP @@ -412,10 +411,8 @@ class GemmaForCausalLM(nn.Module, SupportsLoRA, SupportsPP): def compute_logits( self, hidden_states: torch.Tensor, - sampling_metadata: SamplingMetadata, ) -> Optional[torch.Tensor]: - logits = self.logits_processor(self.model.embed_tokens, hidden_states, - sampling_metadata) + logits = self.logits_processor(self.model.embed_tokens, hidden_states) return logits def load_weights(self, weights: Iterable[tuple[str, diff --git a/vllm/model_executor/models/gemma2.py b/vllm/model_executor/models/gemma2.py index 0bdb6c6bf7ae9..3f76e1e7d42a2 100644 --- a/vllm/model_executor/models/gemma2.py +++ b/vllm/model_executor/models/gemma2.py @@ -41,7 +41,6 @@ from vllm.model_executor.layers.vocab_parallel_embedding import ( VocabParallelEmbedding) from vllm.model_executor.model_loader.weight_utils import ( default_weight_loader, maybe_remap_kv_scale_name) -from vllm.model_executor.sampling_metadata import SamplingMetadata from vllm.sequence import IntermediateTensors from .interfaces import SupportsLoRA, SupportsPP @@ -409,10 +408,8 @@ class Gemma2ForCausalLM(nn.Module, SupportsLoRA, SupportsPP): def compute_logits( self, hidden_states: torch.Tensor, - sampling_metadata: SamplingMetadata, ) -> Optional[torch.Tensor]: - logits = self.logits_processor(self.model.embed_tokens, hidden_states, - sampling_metadata) + logits = self.logits_processor(self.model.embed_tokens, hidden_states) return logits def load_weights(self, weights: Iterable[tuple[str, diff --git a/vllm/model_executor/models/gemma3.py b/vllm/model_executor/models/gemma3.py index 7246308d59028..77c0ef8cb91d2 100644 --- a/vllm/model_executor/models/gemma3.py +++ b/vllm/model_executor/models/gemma3.py @@ -41,7 +41,6 @@ from vllm.model_executor.layers.vocab_parallel_embedding import ( VocabParallelEmbedding) from vllm.model_executor.model_loader.weight_utils import ( default_weight_loader, maybe_remap_kv_scale_name) -from vllm.model_executor.sampling_metadata import SamplingMetadata from vllm.sequence import IntermediateTensors from ...attention.layers.encoder_only_attention import EncoderOnlyAttention @@ -542,10 +541,8 @@ class Gemma3ForCausalLM(nn.Module, SupportsLoRA, SupportsPP): def compute_logits( self, hidden_states: torch.Tensor, - sampling_metadata: SamplingMetadata, ) -> Optional[torch.Tensor]: - logits = self.logits_processor(self.model.embed_tokens, hidden_states, - sampling_metadata) + logits = self.logits_processor(self.model.embed_tokens, hidden_states) return logits def load_weights(self, weights: Iterable[tuple[str, diff --git a/vllm/model_executor/models/gemma3_mm.py b/vllm/model_executor/models/gemma3_mm.py index bee9fbd2c084a..0630ee07c347e 100644 --- a/vllm/model_executor/models/gemma3_mm.py +++ b/vllm/model_executor/models/gemma3_mm.py @@ -14,7 +14,6 @@ from vllm.config import VllmConfig from vllm.logger import init_logger from vllm.model_executor.layers.layernorm import GemmaRMSNorm from vllm.model_executor.models.module_mapping import MultiModelKeys -from vllm.model_executor.sampling_metadata import SamplingMetadata from vllm.multimodal import MULTIMODAL_REGISTRY from vllm.multimodal.inputs import (MultiModalDataDict, MultiModalFieldConfig, MultiModalKwargsItems) @@ -704,10 +703,8 @@ class Gemma3ForConditionalGeneration(nn.Module, SupportsMultiModal, SupportsPP, def compute_logits( self, hidden_states: torch.Tensor, - sampling_metadata: SamplingMetadata, ) -> Optional[torch.Tensor]: - return self.language_model.compute_logits(hidden_states, - sampling_metadata) + return self.language_model.compute_logits(hidden_states) def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]: diff --git a/vllm/model_executor/models/gemma3n.py b/vllm/model_executor/models/gemma3n.py index ffec3408702c9..f4d288fd887e9 100644 --- a/vllm/model_executor/models/gemma3n.py +++ b/vllm/model_executor/models/gemma3n.py @@ -43,7 +43,6 @@ from vllm.model_executor.layers.vocab_parallel_embedding import ( VocabParallelEmbedding) from vllm.model_executor.model_loader.weight_utils import ( default_weight_loader, maybe_remap_kv_scale_name) -from vllm.model_executor.sampling_metadata import SamplingMetadata from vllm.sequence import IntermediateTensors from .interfaces import SupportsQuant @@ -814,10 +813,8 @@ class Gemma3nForCausalLM(nn.Module): def compute_logits( self, hidden_states: torch.Tensor, - sampling_metadata: Optional[SamplingMetadata], ) -> Optional[torch.Tensor]: - logits = self.logits_processor(self.model.embed_tokens, hidden_states, - sampling_metadata) + logits = self.logits_processor(self.model.embed_tokens, hidden_states) return logits def load_weights(self, weights: Iterable[tuple[str, diff --git a/vllm/model_executor/models/gemma3n_mm.py b/vllm/model_executor/models/gemma3n_mm.py index 8d3079aee0dfb..2acdba54a257d 100644 --- a/vllm/model_executor/models/gemma3n_mm.py +++ b/vllm/model_executor/models/gemma3n_mm.py @@ -25,7 +25,6 @@ from vllm.model_executor.layers.vocab_parallel_embedding import ( from vllm.model_executor.models.gemma3n import Gemma3nForCausalLM from vllm.model_executor.models.module_mapping import MultiModelKeys from vllm.model_executor.models.whisper import ISO639_1_SUPPORTED_LANGS -from vllm.model_executor.sampling_metadata import SamplingMetadata from vllm.multimodal import MULTIMODAL_REGISTRY from vllm.multimodal.inputs import (MultiModalDataDict, MultiModalFieldConfig, MultiModalKwargsItems) @@ -685,10 +684,8 @@ class Gemma3nForConditionalGeneration(nn.Module, SupportsMultiModal, def compute_logits( self, hidden_states: torch.Tensor, - sampling_metadata: SamplingMetadata, ) -> Optional[torch.Tensor]: - return self.language_model.compute_logits(hidden_states, - sampling_metadata) + return self.language_model.compute_logits(hidden_states) def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]: diff --git a/vllm/model_executor/models/glm4.py b/vllm/model_executor/models/glm4.py index 5e2908a82c418..b9d5e24e9f6fa 100644 --- a/vllm/model_executor/models/glm4.py +++ b/vllm/model_executor/models/glm4.py @@ -40,7 +40,6 @@ from vllm.model_executor.layers.logits_processor import LogitsProcessor from vllm.model_executor.layers.quantization import QuantizationConfig from vllm.model_executor.layers.rotary_embedding import get_rope from vllm.model_executor.layers.vocab_parallel_embedding import ParallelLMHead -from vllm.model_executor.sampling_metadata import SamplingMetadata from vllm.sequence import IntermediateTensors from .interfaces import SupportsLoRA, SupportsPP @@ -289,10 +288,8 @@ class Glm4ForCausalLM(nn.Module, SupportsLoRA, SupportsPP): def compute_logits( self, hidden_states: torch.Tensor, - sampling_metadata: SamplingMetadata, ) -> Optional[torch.Tensor]: - logits = self.logits_processor(self.lm_head, hidden_states, - sampling_metadata) + logits = self.logits_processor(self.lm_head, hidden_states) return logits def load_weights(self, weights: Iterable[tuple[str, diff --git a/vllm/model_executor/models/glm4_1v.py b/vllm/model_executor/models/glm4_1v.py index 308b0cb602bc9..56ec634386909 100644 --- a/vllm/model_executor/models/glm4_1v.py +++ b/vllm/model_executor/models/glm4_1v.py @@ -52,7 +52,6 @@ from vllm.distributed import (get_tensor_model_parallel_world_size, parallel_state) from vllm.distributed import utils as dist_utils from vllm.logger import init_logger -from vllm.model_executor import SamplingMetadata from vllm.model_executor.layers.layernorm import RMSNorm from vllm.model_executor.layers.linear import (ColumnParallelLinear, MergedColumnParallelLinear, @@ -1654,10 +1653,8 @@ class Glm4vForConditionalGeneration(nn.Module, SupportsMultiModal, def compute_logits( self, hidden_states: torch.Tensor, - sampling_metadata: SamplingMetadata, ) -> Optional[torch.Tensor]: - return self.language_model.compute_logits(hidden_states, - sampling_metadata) + return self.language_model.compute_logits(hidden_states) def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]: diff --git a/vllm/model_executor/models/glm4_moe.py b/vllm/model_executor/models/glm4_moe.py index 1acbd18091fb3..947c6ce62f551 100644 --- a/vllm/model_executor/models/glm4_moe.py +++ b/vllm/model_executor/models/glm4_moe.py @@ -51,7 +51,6 @@ from vllm.model_executor.layers.vocab_parallel_embedding import ( ParallelLMHead, VocabParallelEmbedding) from vllm.model_executor.model_loader.weight_utils import ( default_weight_loader, maybe_remap_kv_scale_name) -from vllm.model_executor.sampling_metadata import SamplingMetadata from vllm.sequence import IntermediateTensors from .interfaces import SupportsLoRA, SupportsPP @@ -703,10 +702,8 @@ class Glm4MoeForCausalLM(nn.Module, SupportsPP, SupportsLoRA): def compute_logits( self, hidden_states: torch.Tensor, - sampling_metadata: SamplingMetadata, ) -> Optional[torch.Tensor]: - logits = self.logits_processor(self.lm_head, hidden_states, - sampling_metadata) + logits = self.logits_processor(self.lm_head, hidden_states) return logits def load_weights(self, weights: Iterable[tuple[str, diff --git a/vllm/model_executor/models/glm4_moe_mtp.py b/vllm/model_executor/models/glm4_moe_mtp.py index 322c5619c1783..c572978e62206 100644 --- a/vllm/model_executor/models/glm4_moe_mtp.py +++ b/vllm/model_executor/models/glm4_moe_mtp.py @@ -38,7 +38,6 @@ from vllm.model_executor.layers.quantization import QuantizationConfig from vllm.model_executor.layers.vocab_parallel_embedding import ( ParallelLMHead, VocabParallelEmbedding) from vllm.model_executor.model_loader.weight_utils import default_weight_loader -from vllm.model_executor.sampling_metadata import SamplingMetadata from vllm.sequence import IntermediateTensors from .glm4_moe import Glm4MoeDecoderLayer, get_spec_layer_idx_from_weight_name @@ -155,15 +154,13 @@ class Glm4MoeMultiTokenPredictor(nn.Module): def compute_logits( self, hidden_states: torch.Tensor, - sampling_metadata: SamplingMetadata, spec_step_idx: int = 0, ) -> torch.Tensor: current_step_idx = (spec_step_idx % self.num_mtp_layers) mtp_layer = self.layers[str(self.mtp_start_layer_idx + current_step_idx)] logits = self.logits_processor(mtp_layer.shared_head.head, - mtp_layer.shared_head(hidden_states), - sampling_metadata) + mtp_layer.shared_head(hidden_states)) return logits @@ -192,11 +189,9 @@ class Glm4MoeMTP(nn.Module, SupportsPP): def compute_logits( self, hidden_states: torch.Tensor, - sampling_metadata: SamplingMetadata, spec_step_idx: int = 0, ) -> Optional[torch.Tensor]: - return self.model.compute_logits(hidden_states, sampling_metadata, - spec_step_idx) + return self.model.compute_logits(hidden_states, spec_step_idx) def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]: diff --git a/vllm/model_executor/models/gpt2.py b/vllm/model_executor/models/gpt2.py index 0f6521e44e6be..24274db148bd7 100644 --- a/vllm/model_executor/models/gpt2.py +++ b/vllm/model_executor/models/gpt2.py @@ -41,7 +41,6 @@ from vllm.model_executor.layers.quantization import QuantizationConfig from vllm.model_executor.layers.vocab_parallel_embedding import ( ParallelLMHead, VocabParallelEmbedding) from vllm.model_executor.model_loader.weight_utils import default_weight_loader -from vllm.model_executor.sampling_metadata import SamplingMetadata from vllm.sequence import IntermediateTensors from ..layers.pooler import DispatchPooler, Pooler @@ -307,10 +306,8 @@ class GPT2LMHeadModel(nn.Module, SupportsPP): def compute_logits( self, hidden_states: torch.Tensor, - sampling_metadata: SamplingMetadata, ) -> Optional[torch.Tensor]: - logits = self.logits_processor(self.lm_head, hidden_states, - sampling_metadata) + logits = self.logits_processor(self.lm_head, hidden_states) return logits def load_weights(self, weights: Iterable[tuple[str, diff --git a/vllm/model_executor/models/gpt_bigcode.py b/vllm/model_executor/models/gpt_bigcode.py index 745d0b7759991..162018450e7c0 100644 --- a/vllm/model_executor/models/gpt_bigcode.py +++ b/vllm/model_executor/models/gpt_bigcode.py @@ -41,7 +41,6 @@ from vllm.model_executor.layers.quantization import QuantizationConfig from vllm.model_executor.layers.vocab_parallel_embedding import ( ParallelLMHead, VocabParallelEmbedding) from vllm.model_executor.model_loader.weight_utils import default_weight_loader -from vllm.model_executor.sampling_metadata import SamplingMetadata from vllm.sequence import IntermediateTensors from .interfaces import SupportsLoRA, SupportsPP @@ -329,10 +328,8 @@ class GPTBigCodeForCausalLM(nn.Module, SupportsLoRA, SupportsPP): def compute_logits( self, hidden_states: torch.Tensor, - sampling_metadata: SamplingMetadata, ) -> Optional[torch.Tensor]: - logits = self.logits_processor(self.lm_head, hidden_states, - sampling_metadata) + logits = self.logits_processor(self.lm_head, hidden_states) return logits def load_weights(self, weights: Iterable[tuple[str, diff --git a/vllm/model_executor/models/gpt_j.py b/vllm/model_executor/models/gpt_j.py index 77df6ae6f30c8..698387fab946c 100644 --- a/vllm/model_executor/models/gpt_j.py +++ b/vllm/model_executor/models/gpt_j.py @@ -41,7 +41,6 @@ from vllm.model_executor.layers.vocab_parallel_embedding import ( ParallelLMHead, VocabParallelEmbedding) from vllm.model_executor.model_loader.weight_utils import ( default_weight_loader, maybe_remap_kv_scale_name) -from vllm.model_executor.sampling_metadata import SamplingMetadata from vllm.sequence import IntermediateTensors from .interfaces import SupportsPP @@ -329,10 +328,9 @@ class GPTJForCausalLM(nn.Module, SupportsPP): def compute_logits( self, hidden_states: torch.Tensor, - sampling_metadata: SamplingMetadata, ) -> Optional[torch.Tensor]: logits = self.logits_processor(self.lm_head, hidden_states, - sampling_metadata, self.lm_head.bias) + self.lm_head.bias) return logits def load_weights(self, weights: Iterable[tuple[str, diff --git a/vllm/model_executor/models/gpt_neox.py b/vllm/model_executor/models/gpt_neox.py index e97db188e27eb..7570aefb6e96e 100644 --- a/vllm/model_executor/models/gpt_neox.py +++ b/vllm/model_executor/models/gpt_neox.py @@ -40,7 +40,6 @@ from vllm.model_executor.layers.rotary_embedding import get_rope from vllm.model_executor.layers.vocab_parallel_embedding import ( ParallelLMHead, VocabParallelEmbedding) from vllm.model_executor.model_loader.weight_utils import default_weight_loader -from vllm.model_executor.sampling_metadata import SamplingMetadata from vllm.sequence import IntermediateTensors from .interfaces import SupportsPP @@ -321,10 +320,8 @@ class GPTNeoXForCausalLM(nn.Module, SupportsPP): def compute_logits( self, hidden_states: torch.Tensor, - sampling_metadata: SamplingMetadata, ) -> Optional[torch.Tensor]: - logits = self.logits_processor(self.embed_out, hidden_states, - sampling_metadata) + logits = self.logits_processor(self.embed_out, hidden_states) return logits def load_weights(self, weights: Iterable[tuple[str, diff --git a/vllm/model_executor/models/gpt_oss.py b/vllm/model_executor/models/gpt_oss.py index b49fd0d8f88af..4fe59f91124dd 100644 --- a/vllm/model_executor/models/gpt_oss.py +++ b/vllm/model_executor/models/gpt_oss.py @@ -24,7 +24,6 @@ from vllm.model_executor.layers.rotary_embedding import get_rope from vllm.model_executor.layers.vocab_parallel_embedding import ( ParallelLMHead, VocabParallelEmbedding) from vllm.model_executor.model_loader.weight_utils import default_weight_loader -from vllm.model_executor.sampling_metadata import SamplingMetadata from vllm.sequence import IntermediateTensors from vllm.utils import cdiv @@ -670,10 +669,8 @@ class GptOssForCausalLM(nn.Module, SupportsPP): return self.model(input_ids, positions, intermediate_tensors, inputs_embeds) - def compute_logits(self, hidden_states: torch.Tensor, - sampling_metadata: SamplingMetadata) -> torch.Tensor: - logits = self.logits_processor(self.lm_head, hidden_states, - sampling_metadata) + def compute_logits(self, hidden_states: torch.Tensor) -> torch.Tensor: + logits = self.logits_processor(self.lm_head, hidden_states) return logits def load_weights(self, weights: Iterable[tuple[str, diff --git a/vllm/model_executor/models/granite.py b/vllm/model_executor/models/granite.py index 4f9cc2532bd8c..795b38e724eab 100644 --- a/vllm/model_executor/models/granite.py +++ b/vllm/model_executor/models/granite.py @@ -48,7 +48,6 @@ from vllm.model_executor.layers.vocab_parallel_embedding import ( DEFAULT_VOCAB_PADDING_SIZE, ParallelLMHead, VocabParallelEmbedding) from vllm.model_executor.model_loader.weight_utils import ( default_weight_loader, maybe_remap_kv_scale_name) -from vllm.model_executor.sampling_metadata import SamplingMetadata from vllm.sequence import IntermediateTensors from .interfaces import SupportsLoRA, SupportsPP @@ -463,11 +462,9 @@ class GraniteForCausalLM(nn.Module, SupportsLoRA, SupportsPP): inputs_embeds) return model_output - def compute_logits( - self, hidden_states: torch.Tensor, - sampling_metadata: SamplingMetadata) -> Optional[torch.Tensor]: - logits = self.logits_processor(self.lm_head, hidden_states, - sampling_metadata) + def compute_logits(self, + hidden_states: torch.Tensor) -> Optional[torch.Tensor]: + logits = self.logits_processor(self.lm_head, hidden_states) return logits def make_empty_intermediate_tensors( diff --git a/vllm/model_executor/models/granite_speech.py b/vllm/model_executor/models/granite_speech.py index 221023f1fb657..a5849184339b1 100644 --- a/vllm/model_executor/models/granite_speech.py +++ b/vllm/model_executor/models/granite_speech.py @@ -37,7 +37,6 @@ from vllm.model_executor.layers.linear import (ColumnParallelLinear, RowParallelLinear) from vllm.model_executor.layers.quantization import QuantizationConfig from vllm.model_executor.models.module_mapping import MultiModelKeys -from vllm.model_executor.sampling_metadata import SamplingMetadata from vllm.multimodal import MULTIMODAL_REGISTRY from vllm.multimodal.inputs import (MultiModalDataDict, MultiModalFieldConfig, MultiModalKwargsItems) @@ -776,12 +775,8 @@ class GraniteSpeechForConditionalGeneration( def compute_logits( self, hidden_states: torch.Tensor, - sampling_metadata: SamplingMetadata, ) -> Optional[torch.Tensor]: - return self.language_model.compute_logits( - hidden_states, - sampling_metadata, - ) + return self.language_model.compute_logits(hidden_states) def load_weights( self, diff --git a/vllm/model_executor/models/granitemoe.py b/vllm/model_executor/models/granitemoe.py index da16c72000c0e..07200fef4799d 100644 --- a/vllm/model_executor/models/granitemoe.py +++ b/vllm/model_executor/models/granitemoe.py @@ -48,7 +48,6 @@ from vllm.model_executor.layers.vocab_parallel_embedding import ( DEFAULT_VOCAB_PADDING_SIZE, ParallelLMHead, VocabParallelEmbedding) from vllm.model_executor.model_loader.weight_utils import ( default_weight_loader, maybe_remap_kv_scale_name) -from vllm.model_executor.sampling_metadata import SamplingMetadata from vllm.sequence import IntermediateTensors from .interfaces import SupportsLoRA, SupportsPP @@ -511,11 +510,9 @@ class GraniteMoeForCausalLM(nn.Module, SupportsLoRA, SupportsPP): inputs_embeds) return hidden_states - def compute_logits( - self, hidden_states: torch.Tensor, - sampling_metadata: SamplingMetadata) -> Optional[torch.Tensor]: - logits = self.logits_processor(self.lm_head, hidden_states, - sampling_metadata) + def compute_logits(self, + hidden_states: torch.Tensor) -> Optional[torch.Tensor]: + logits = self.logits_processor(self.lm_head, hidden_states) return logits def make_empty_intermediate_tensors( diff --git a/vllm/model_executor/models/granitemoehybrid.py b/vllm/model_executor/models/granitemoehybrid.py index 79c6d8146ba9c..e89a1a4a0f7d3 100644 --- a/vllm/model_executor/models/granitemoehybrid.py +++ b/vllm/model_executor/models/granitemoehybrid.py @@ -32,7 +32,6 @@ from vllm.model_executor.layers.vocab_parallel_embedding import ( from vllm.model_executor.model_loader.weight_utils import default_weight_loader from vllm.model_executor.models.mamba_cache import (MambaCacheManager, MambaCacheParams) -from vllm.model_executor.sampling_metadata import SamplingMetadata from vllm.sequence import IntermediateTensors from vllm.utils import LayerBlockType @@ -672,10 +671,8 @@ class GraniteMoeHybridForCausalLM(nn.Module, HasInnerState, SupportsLoRA, def compute_logits( self, hidden_states: torch.Tensor, - sampling_metadata: SamplingMetadata, ) -> Optional[torch.Tensor]: - logits = self.logits_processor(self.lm_head, hidden_states, - sampling_metadata) + logits = self.logits_processor(self.lm_head, hidden_states) return logits def load_weights(self, weights: Iterable[tuple[str, diff --git a/vllm/model_executor/models/granitemoeshared.py b/vllm/model_executor/models/granitemoeshared.py index 0b568a4b22685..a5d118f084e6c 100644 --- a/vllm/model_executor/models/granitemoeshared.py +++ b/vllm/model_executor/models/granitemoeshared.py @@ -25,7 +25,6 @@ from vllm.model_executor.layers.quantization.base_config import ( QuantizationConfig) from vllm.model_executor.layers.vocab_parallel_embedding import ( DEFAULT_VOCAB_PADDING_SIZE, ParallelLMHead, VocabParallelEmbedding) -from vllm.model_executor.sampling_metadata import SamplingMetadata from vllm.sequence import IntermediateTensors from .granitemoe import GraniteMoeAttention, GraniteMoeModel, GraniteMoeMoE @@ -311,11 +310,9 @@ class GraniteMoeSharedForCausalLM(nn.Module, SupportsLoRA, SupportsPP): inputs_embeds) return hidden_states - def compute_logits( - self, hidden_states: torch.Tensor, - sampling_metadata: SamplingMetadata) -> Optional[torch.Tensor]: - logits = self.logits_processor(self.lm_head, hidden_states, - sampling_metadata) + def compute_logits(self, + hidden_states: torch.Tensor) -> Optional[torch.Tensor]: + logits = self.logits_processor(self.lm_head, hidden_states) return logits def make_empty_intermediate_tensors( diff --git a/vllm/model_executor/models/grok1.py b/vllm/model_executor/models/grok1.py index a591134383371..996e41fe84ff3 100644 --- a/vllm/model_executor/models/grok1.py +++ b/vllm/model_executor/models/grok1.py @@ -46,7 +46,6 @@ from vllm.model_executor.layers.vocab_parallel_embedding import ( DEFAULT_VOCAB_PADDING_SIZE, ParallelLMHead, VocabParallelEmbedding) from vllm.model_executor.model_loader.weight_utils import ( default_weight_loader, maybe_remap_kv_scale_name) -from vllm.model_executor.sampling_metadata import SamplingMetadata from vllm.sequence import IntermediateTensors from .interfaces import SupportsLoRA, SupportsPP @@ -528,10 +527,8 @@ class Grok1ForCausalLM(nn.Module, SupportsLoRA, SupportsPP): def compute_logits( self, hidden_states: torch.Tensor, - sampling_metadata: SamplingMetadata, ) -> Optional[torch.Tensor]: - logits = self.logits_processor(self.lm_head, hidden_states, - sampling_metadata) + logits = self.logits_processor(self.lm_head, hidden_states) return logits def load_weights(self, weights: Iterable[tuple[str, diff --git a/vllm/model_executor/models/hunyuan_v1.py b/vllm/model_executor/models/hunyuan_v1.py index 4110c8a1fd08d..8a23a6b45bc70 100644 --- a/vllm/model_executor/models/hunyuan_v1.py +++ b/vllm/model_executor/models/hunyuan_v1.py @@ -54,7 +54,6 @@ from vllm.model_executor.layers.vocab_parallel_embedding import ( DEFAULT_VOCAB_PADDING_SIZE, ParallelLMHead, VocabParallelEmbedding) from vllm.model_executor.model_loader.weight_utils import ( default_weight_loader, maybe_remap_kv_scale_name) -from vllm.model_executor.sampling_metadata import SamplingMetadata from vllm.sequence import IntermediateTensors from .interfaces import MixtureOfExperts, SupportsLoRA, SupportsPP @@ -1004,10 +1003,8 @@ class HunYuanV1Base(nn.Module, SupportsLoRA, SupportsPP, MixtureOfExperts): def compute_logits( self, hidden_states: torch.Tensor, - sampling_metadata: SamplingMetadata, ) -> Optional[torch.Tensor]: - logits = self.logits_processor(self.lm_head, hidden_states, - sampling_metadata) + logits = self.logits_processor(self.lm_head, hidden_states) return logits def make_empty_intermediate_tensors( diff --git a/vllm/model_executor/models/hyperclovax_vision.py b/vllm/model_executor/models/hyperclovax_vision.py index 870addd0dcbca..54167f9f10995 100644 --- a/vllm/model_executor/models/hyperclovax_vision.py +++ b/vllm/model_executor/models/hyperclovax_vision.py @@ -31,7 +31,6 @@ from transformers.modeling_utils import no_init_weights from vllm.config import VllmConfig from vllm.inputs import InputProcessingContext from vllm.model_executor.layers.quantization import QuantizationConfig -from vllm.model_executor.sampling_metadata import SamplingMetadata from vllm.multimodal import MULTIMODAL_REGISTRY from vllm.multimodal.cache import BaseMultiModalProcessorCache from vllm.multimodal.inputs import (MultiModalDataDict, MultiModalFieldConfig, @@ -962,10 +961,8 @@ class HCXVisionForCausalLM(nn.Module, SupportsMultiModal, SupportsPP): def compute_logits( self, hidden_states: torch.Tensor, - sampling_metadata: SamplingMetadata, ) -> Optional[torch.Tensor]: - return self.language_model.compute_logits(hidden_states, - sampling_metadata) + return self.language_model.compute_logits(hidden_states) def load_weights( self, diff --git a/vllm/model_executor/models/idefics3.py b/vllm/model_executor/models/idefics3.py index 9153a0e2c1e5a..18446d126b51a 100644 --- a/vllm/model_executor/models/idefics3.py +++ b/vllm/model_executor/models/idefics3.py @@ -31,7 +31,6 @@ from vllm.model_executor.layers.logits_processor import LogitsProcessor from vllm.model_executor.layers.quantization import QuantizationConfig from vllm.model_executor.layers.vocab_parallel_embedding import ParallelLMHead from vllm.model_executor.models.module_mapping import MultiModelKeys -from vllm.model_executor.sampling_metadata import SamplingMetadata from vllm.multimodal import MULTIMODAL_REGISTRY from vllm.multimodal.inputs import (MultiModalDataDict, MultiModalFieldConfig, MultiModalKwargsItems) @@ -738,10 +737,8 @@ class Idefics3ForConditionalGeneration(nn.Module, SupportsMultiModal, return hidden_states - def compute_logits(self, hidden_states: torch.Tensor, - sampling_metadata: SamplingMetadata) -> torch.Tensor: - logits = self.logits_processor(self.lm_head, hidden_states, - sampling_metadata) + def compute_logits(self, hidden_states: torch.Tensor) -> torch.Tensor: + logits = self.logits_processor(self.lm_head, hidden_states) return logits def load_weights(self, weights: Iterable[tuple[str, diff --git a/vllm/model_executor/models/interfaces_base.py b/vllm/model_executor/models/interfaces_base.py index 19a3ef1a3b800..8fdf70e35a2b8 100644 --- a/vllm/model_executor/models/interfaces_base.py +++ b/vllm/model_executor/models/interfaces_base.py @@ -13,11 +13,9 @@ from vllm.utils import supports_kw if TYPE_CHECKING: from vllm.config import VllmConfig from vllm.model_executor.layers.pooler import Pooler - from vllm.model_executor.sampling_metadata import SamplingMetadata else: VllmConfig = Any Pooler = Any - SamplingMetadata = Any logger = init_logger(__name__) @@ -100,7 +98,6 @@ class VllmModelForTextGeneration(VllmModel[T], Protocol[T]): def compute_logits( self, hidden_states: T, - sampling_metadata: SamplingMetadata, ) -> Optional[T]: """Return `None` if TP rank > 0.""" ... diff --git a/vllm/model_executor/models/internlm2.py b/vllm/model_executor/models/internlm2.py index ce94328797ed6..221ff08b43843 100644 --- a/vllm/model_executor/models/internlm2.py +++ b/vllm/model_executor/models/internlm2.py @@ -29,7 +29,6 @@ from vllm.model_executor.layers.rotary_embedding import get_rope from vllm.model_executor.layers.vocab_parallel_embedding import ( ParallelLMHead, VocabParallelEmbedding) from vllm.model_executor.model_loader.weight_utils import default_weight_loader -from vllm.model_executor.sampling_metadata import SamplingMetadata from vllm.sequence import IntermediateTensors from .interfaces import SupportsLoRA, SupportsPP @@ -358,10 +357,8 @@ class InternLM2ForCausalLM(nn.Module, SupportsPP, SupportsLoRA): def compute_logits( self, hidden_states: torch.Tensor, - sampling_metadata: SamplingMetadata, ) -> Optional[torch.Tensor]: - logits = self.logits_processor(self.output, hidden_states, - sampling_metadata) + logits = self.logits_processor(self.output, hidden_states) return logits def load_weights(self, weights: Iterable[tuple[str, diff --git a/vllm/model_executor/models/interns1.py b/vllm/model_executor/models/interns1.py index b59d1b88cf5ce..ba72c288b2b12 100644 --- a/vllm/model_executor/models/interns1.py +++ b/vllm/model_executor/models/interns1.py @@ -21,7 +21,6 @@ from vllm.config import VllmConfig from vllm.model_executor.layers.quantization import QuantizationConfig from vllm.model_executor.models.interns1_vit import InternS1VisionModel from vllm.model_executor.models.module_mapping import MultiModelKeys -from vllm.model_executor.sampling_metadata import SamplingMetadata from vllm.multimodal import MULTIMODAL_REGISTRY from vllm.multimodal.inputs import (MultiModalDataDict, MultiModalFieldConfig, MultiModalKwargsItems, NestedTensors) @@ -812,10 +811,8 @@ class InternS1ForConditionalGeneration(nn.Module, SupportsMultiModal, def compute_logits( self, hidden_states: torch.Tensor, - sampling_metadata: SamplingMetadata, ) -> Optional[torch.Tensor]: - return self.language_model.compute_logits(hidden_states, - sampling_metadata) + return self.language_model.compute_logits(hidden_states) def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]: diff --git a/vllm/model_executor/models/internvl.py b/vllm/model_executor/models/internvl.py index 6a5c565b52e85..f4004e518e3ba 100644 --- a/vllm/model_executor/models/internvl.py +++ b/vllm/model_executor/models/internvl.py @@ -25,7 +25,6 @@ from vllm.model_executor.layers.quantization.awq import AWQConfig from vllm.model_executor.models.intern_vit import (InternVisionModel, InternVisionPatchModel) from vllm.model_executor.models.module_mapping import MultiModelKeys -from vllm.model_executor.sampling_metadata import SamplingMetadata from vllm.multimodal import MULTIMODAL_REGISTRY from vllm.multimodal.image import convert_image_mode from vllm.multimodal.inputs import (MultiModalDataDict, MultiModalFieldConfig, @@ -1399,10 +1398,8 @@ class InternVLChatModel(nn.Module, SupportsMultiModal, SupportsPP, def compute_logits( self, hidden_states: torch.Tensor, - sampling_metadata: SamplingMetadata, ) -> Optional[torch.Tensor]: - return self.language_model.compute_logits(hidden_states, - sampling_metadata) + return self.language_model.compute_logits(hidden_states) def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]: diff --git a/vllm/model_executor/models/jais.py b/vllm/model_executor/models/jais.py index 4fee8c32fd581..0eb1578b43610 100644 --- a/vllm/model_executor/models/jais.py +++ b/vllm/model_executor/models/jais.py @@ -42,7 +42,6 @@ from vllm.model_executor.layers.quantization import QuantizationConfig from vllm.model_executor.layers.vocab_parallel_embedding import ( ParallelLMHead, VocabParallelEmbedding) from vllm.model_executor.model_loader.weight_utils import default_weight_loader -from vllm.model_executor.sampling_metadata import SamplingMetadata from vllm.sequence import IntermediateTensors from vllm.transformers_utils.configs import JAISConfig @@ -332,10 +331,8 @@ class JAISLMHeadModel(nn.Module, SupportsPP): def compute_logits( self, hidden_states: torch.Tensor, - sampling_metadata: SamplingMetadata, ) -> Optional[torch.Tensor]: - logits = self.logits_processor(self.lm_head, hidden_states, - sampling_metadata) + logits = self.logits_processor(self.lm_head, hidden_states) return logits def load_weights(self, weights: Iterable[tuple[str, diff --git a/vllm/model_executor/models/jamba.py b/vllm/model_executor/models/jamba.py index 5b8fbc7226866..12a49029195ff 100644 --- a/vllm/model_executor/models/jamba.py +++ b/vllm/model_executor/models/jamba.py @@ -32,7 +32,6 @@ from vllm.model_executor.model_loader.weight_utils import default_weight_loader from vllm.model_executor.models.llama import LlamaMLP as JambaMLP from vllm.model_executor.models.mamba_cache import (MambaCacheManager, MambaCacheParams) -from vllm.model_executor.sampling_metadata import SamplingMetadata from vllm.sequence import IntermediateTensors from vllm.utils import LayerBlockType @@ -581,10 +580,8 @@ class JambaForCausalLM(nn.Module, HasInnerState, SupportsLoRA, SupportsPP, def compute_logits( self, hidden_states: torch.Tensor, - sampling_metadata: SamplingMetadata, ) -> Optional[torch.Tensor]: - logits = self.logits_processor(self.lm_head, hidden_states, - sampling_metadata) + logits = self.logits_processor(self.lm_head, hidden_states) return logits def load_weights(self, weights: Iterable[tuple[str, diff --git a/vllm/model_executor/models/keye.py b/vllm/model_executor/models/keye.py index afe33b4d4ad26..2e5e276cc1c7d 100644 --- a/vllm/model_executor/models/keye.py +++ b/vllm/model_executor/models/keye.py @@ -21,7 +21,6 @@ from vllm.attention.layer import check_upstream_fa_availability from vllm.config import VllmConfig from vllm.distributed import get_tensor_model_parallel_world_size from vllm.logger import init_logger -from vllm.model_executor import SamplingMetadata from vllm.model_executor.layers.linear import (ColumnParallelLinear, QKVParallelLinear, RowParallelLinear) @@ -1556,10 +1555,8 @@ class BaseKeyeModule(nn.Module): def compute_logits( self, hidden_states: torch.Tensor, - sampling_metadata: SamplingMetadata, ) -> Optional[torch.Tensor]: - return self.language_model.compute_logits(hidden_states, - sampling_metadata) + return self.language_model.compute_logits(hidden_states) def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]: diff --git a/vllm/model_executor/models/kimi_vl.py b/vllm/model_executor/models/kimi_vl.py index 94a5933a61416..f554077935bf3 100644 --- a/vllm/model_executor/models/kimi_vl.py +++ b/vllm/model_executor/models/kimi_vl.py @@ -67,7 +67,6 @@ from vllm.model_executor.models.interfaces import (SupportsMultiModal, SupportsPP) from vllm.model_executor.models.moonvit import MoonVitPretrainedModel from vllm.model_executor.models.utils import merge_multimodal_embeddings -from vllm.model_executor.sampling_metadata import SamplingMetadata from vllm.multimodal import MULTIMODAL_REGISTRY from vllm.multimodal.inputs import (MultiModalDataDict, MultiModalFieldConfig, MultiModalKwargsItems, NestedTensors) @@ -484,10 +483,8 @@ class KimiVLForConditionalGeneration(nn.Module, SupportsMultiModal, return hidden_states def compute_logits(self, hidden_states: torch.Tensor, - sampling_metadata: SamplingMetadata, **kwargs) -> torch.Tensor: - logits = self.logits_processor(self.lm_head, hidden_states, - sampling_metadata, **kwargs) + logits = self.logits_processor(self.lm_head, hidden_states, **kwargs) return logits def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]): diff --git a/vllm/model_executor/models/lfm2.py b/vllm/model_executor/models/lfm2.py index 927f78c4e4b45..dd97afbeb668a 100644 --- a/vllm/model_executor/models/lfm2.py +++ b/vllm/model_executor/models/lfm2.py @@ -27,7 +27,6 @@ from vllm.model_executor.layers.rotary_embedding import get_rope from vllm.model_executor.layers.vocab_parallel_embedding import ( DEFAULT_VOCAB_PADDING_SIZE, ParallelLMHead, VocabParallelEmbedding) from vllm.model_executor.model_loader.weight_utils import default_weight_loader -from vllm.model_executor.sampling_metadata import SamplingMetadata from vllm.sequence import IntermediateTensors from .interfaces import (HasInnerState, IsHybrid, SupportsLoRA, SupportsPP, @@ -542,10 +541,8 @@ class Lfm2ForCausalLM(nn.Module, HasInnerState, SupportsLoRA, SupportsPP, inputs_embeds) return hidden_states - def compute_logits(self, hidden_states: torch.Tensor, - sampling_metadata: SamplingMetadata) -> torch.Tensor: - logits = self.logits_processor(self.lm_head, hidden_states, - sampling_metadata) + def compute_logits(self, hidden_states: torch.Tensor) -> torch.Tensor: + logits = self.logits_processor(self.lm_head, hidden_states) return logits def load_weights(self, weights: Iterable[tuple[str, diff --git a/vllm/model_executor/models/llama.py b/vllm/model_executor/models/llama.py index f8ea2111fed57..1b03cbef501b3 100644 --- a/vllm/model_executor/models/llama.py +++ b/vllm/model_executor/models/llama.py @@ -48,7 +48,6 @@ from vllm.model_executor.layers.vocab_parallel_embedding import ( DEFAULT_VOCAB_PADDING_SIZE, ParallelLMHead, VocabParallelEmbedding) from vllm.model_executor.model_loader.weight_utils import ( default_weight_loader, maybe_remap_kv_scale_name) -from vllm.model_executor.sampling_metadata import SamplingMetadata from vllm.sequence import IntermediateTensors from .interfaces import SupportsEagle3, SupportsLoRA, SupportsPP @@ -601,10 +600,8 @@ class LlamaForCausalLM(nn.Module, SupportsLoRA, SupportsPP, SupportsEagle3): def compute_logits( self, hidden_states: torch.Tensor, - sampling_metadata: SamplingMetadata, ) -> Optional[torch.Tensor]: - logits = self.logits_processor(self.lm_head, hidden_states, - sampling_metadata) + logits = self.logits_processor(self.lm_head, hidden_states) return logits def load_weights(self, weights: Iterable[tuple[str, diff --git a/vllm/model_executor/models/llama_eagle3.py b/vllm/model_executor/models/llama_eagle3.py index 7027138dfcb17..fb10af6c53c90 100644 --- a/vllm/model_executor/models/llama_eagle3.py +++ b/vllm/model_executor/models/llama_eagle3.py @@ -21,7 +21,6 @@ from vllm.model_executor.layers.vocab_parallel_embedding import ( from vllm.model_executor.model_loader.weight_utils import default_weight_loader from vllm.model_executor.models.llama import (LlamaDecoderLayer, LlamaForCausalLM) -from vllm.v1.sample.metadata import SamplingMetadata from .utils import AutoWeightsLoader, maybe_prefix @@ -244,10 +243,8 @@ class Eagle3LlamaForCausalLM(LlamaForCausalLM): def compute_logits( self, hidden_states: torch.Tensor, - sampling_metadata: SamplingMetadata, ) -> Optional[torch.Tensor]: - logits = self.logits_processor(self.lm_head, hidden_states, - sampling_metadata) + logits = self.logits_processor(self.lm_head, hidden_states) if self.draft_id_to_target_id is None: assert logits.shape[1] == self.config.vocab_size, \ "Expected logits to have shape " \ diff --git a/vllm/model_executor/models/llava.py b/vllm/model_executor/models/llava.py index 4f15e1b5762ea..e2d7b9f23b28a 100644 --- a/vllm/model_executor/models/llava.py +++ b/vllm/model_executor/models/llava.py @@ -20,7 +20,6 @@ from vllm.model_executor.layers.activation import get_act_fn from vllm.model_executor.layers.linear import (ColumnParallelLinear, RowParallelLinear) from vllm.model_executor.layers.quantization import QuantizationConfig -from vllm.model_executor.sampling_metadata import SamplingMetadata from vllm.multimodal import MULTIMODAL_REGISTRY from vllm.multimodal.cache import BaseMultiModalProcessorCache from vllm.multimodal.inputs import (MultiModalDataDict, MultiModalFieldConfig, @@ -760,10 +759,8 @@ class LlavaForConditionalGeneration(nn.Module, SupportsMultiModal, SupportsPP): def compute_logits( self, hidden_states: torch.Tensor, - sampling_metadata: SamplingMetadata, ) -> Optional[torch.Tensor]: - return self.language_model.compute_logits(hidden_states, - sampling_metadata) + return self.language_model.compute_logits(hidden_states) def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]: diff --git a/vllm/model_executor/models/llava_next.py b/vllm/model_executor/models/llava_next.py index beb3c33100595..c9133fde14552 100644 --- a/vllm/model_executor/models/llava_next.py +++ b/vllm/model_executor/models/llava_next.py @@ -13,7 +13,6 @@ from transformers.models.llava_next.modeling_llava_next import ( get_anyres_image_grid_shape, unpad_image) from vllm.config import VllmConfig -from vllm.model_executor.sampling_metadata import SamplingMetadata from vllm.multimodal import MULTIMODAL_REGISTRY from vllm.multimodal.inputs import MultiModalFieldConfig from vllm.multimodal.parse import ImageSize @@ -563,10 +562,8 @@ model_executor.models.llava_next.LlavaNextProcessingInfo.get_num_image_tokens]. def compute_logits( self, hidden_states: torch.Tensor, - sampling_metadata: SamplingMetadata, ) -> Optional[torch.Tensor]: - return self.language_model.compute_logits(hidden_states, - sampling_metadata) + return self.language_model.compute_logits(hidden_states) def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]: diff --git a/vllm/model_executor/models/llava_next_video.py b/vllm/model_executor/models/llava_next_video.py index cf9852de633f3..610fb188d57d2 100644 --- a/vllm/model_executor/models/llava_next_video.py +++ b/vllm/model_executor/models/llava_next_video.py @@ -13,7 +13,6 @@ from transformers import (BatchFeature, LlavaNextVideoConfig, from vllm.config import VllmConfig from vllm.model_executor.layers.activation import get_act_fn from vllm.model_executor.models.clip import CLIPVisionModel -from vllm.model_executor.sampling_metadata import SamplingMetadata from vllm.multimodal import MULTIMODAL_REGISTRY from vllm.multimodal.inputs import (MultiModalDataDict, MultiModalFieldConfig, MultiModalKwargsItems) @@ -464,10 +463,8 @@ class LlavaNextVideoForConditionalGeneration(nn.Module, SupportsMultiModal, def compute_logits( self, hidden_states: torch.Tensor, - sampling_metadata: SamplingMetadata, ) -> Optional[torch.Tensor]: - return self.language_model.compute_logits(hidden_states, - sampling_metadata) + return self.language_model.compute_logits(hidden_states) def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]: diff --git a/vllm/model_executor/models/llava_onevision.py b/vllm/model_executor/models/llava_onevision.py index 46d54452a52d8..cee9ddaf94cc4 100644 --- a/vllm/model_executor/models/llava_onevision.py +++ b/vllm/model_executor/models/llava_onevision.py @@ -14,7 +14,6 @@ from transformers.models.llava_onevision.modeling_llava_onevision import ( from vllm.config import VllmConfig from vllm.model_executor.layers.activation import get_act_fn -from vllm.model_executor.sampling_metadata import SamplingMetadata from vllm.multimodal import MULTIMODAL_REGISTRY from vllm.multimodal.inputs import (MultiModalDataDict, MultiModalFieldConfig, MultiModalKwargsItems) @@ -934,10 +933,8 @@ class LlavaOnevisionForConditionalGeneration(nn.Module, SupportsMultiModal, def compute_logits( self, hidden_states: torch.Tensor, - sampling_metadata: SamplingMetadata, ) -> Optional[torch.Tensor]: - return self.language_model.compute_logits(hidden_states, - sampling_metadata) + return self.language_model.compute_logits(hidden_states) def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]: diff --git a/vllm/model_executor/models/mamba.py b/vllm/model_executor/models/mamba.py index 9d1017dac8aa1..36141a5d50641 100644 --- a/vllm/model_executor/models/mamba.py +++ b/vllm/model_executor/models/mamba.py @@ -26,7 +26,6 @@ from vllm.model_executor.models.interfaces import (HasInnerState, IsAttentionFree, SupportsPP) from vllm.model_executor.models.mamba_cache import (MambaCacheManager, MambaCacheParams) -from vllm.model_executor.sampling_metadata import SamplingMetadata from vllm.sequence import IntermediateTensors from vllm.utils import LayerBlockType @@ -299,10 +298,8 @@ class MambaForCausalLM(nn.Module, HasInnerState, IsAttentionFree, SupportsPP): def get_seqlen_agnostic_capture_inputs(self, batch_size: int): return self.mamba_cache.get_seqlen_agnostic_capture_inputs(batch_size) - def compute_logits(self, hidden_states: torch.Tensor, - sampling_metadata: SamplingMetadata) -> torch.Tensor: - logits = self.logits_processor(self.lm_head, hidden_states, - sampling_metadata) + def compute_logits(self, hidden_states: torch.Tensor) -> torch.Tensor: + logits = self.logits_processor(self.lm_head, hidden_states) return logits def load_weights(self, weights: Iterable[tuple[str, diff --git a/vllm/model_executor/models/mamba2.py b/vllm/model_executor/models/mamba2.py index b1a4138cb8f6c..9c3108146d2e5 100644 --- a/vllm/model_executor/models/mamba2.py +++ b/vllm/model_executor/models/mamba2.py @@ -30,7 +30,6 @@ from vllm.model_executor.models.interfaces import (HasInnerState, IsAttentionFree) from vllm.model_executor.models.mamba_cache import (MambaCacheManager, MambaCacheParams) -from vllm.model_executor.sampling_metadata import SamplingMetadata from vllm.sequence import IntermediateTensors from vllm.utils import LayerBlockType @@ -335,10 +334,8 @@ class Mamba2ForCausalLM(nn.Module, HasInnerState, IsAttentionFree): def get_seqlen_agnostic_capture_inputs(self, batch_size: int): return self.mamba_cache.get_seqlen_agnostic_capture_inputs(batch_size) - def compute_logits(self, hidden_states: torch.Tensor, - sampling_metadata: SamplingMetadata) -> torch.Tensor: - logits = self.logits_processor(self.lm_head, hidden_states, - sampling_metadata) + def compute_logits(self, hidden_states: torch.Tensor) -> torch.Tensor: + logits = self.logits_processor(self.lm_head, hidden_states) return logits def load_weights(self, weights: Iterable[tuple[str, diff --git a/vllm/model_executor/models/medusa.py b/vllm/model_executor/models/medusa.py index b0a96fca2ff8a..0ae59dc8dfc23 100644 --- a/vllm/model_executor/models/medusa.py +++ b/vllm/model_executor/models/medusa.py @@ -104,12 +104,11 @@ class Medusa(nn.Module): def compute_logits( self, hidden_states: list[torch.Tensor], - sampling_metadata, ) -> list[torch.Tensor]: logits_lst: list[torch.Tensor] = [] for hs, lm_head in zip(hidden_states, self.lm_heads): - _logits = self.logits_processor(lm_head, hs, sampling_metadata) + _logits = self.logits_processor(lm_head, hs) if _logits is None: # _logits should only be None on rank > 0, in which case diff --git a/vllm/model_executor/models/midashenglm.py b/vllm/model_executor/models/midashenglm.py index 140800dd41c76..82648ba668ca5 100644 --- a/vllm/model_executor/models/midashenglm.py +++ b/vllm/model_executor/models/midashenglm.py @@ -42,7 +42,6 @@ from vllm.model_executor.layers.linear import (ColumnParallelLinear, RowParallelLinear) from vllm.model_executor.layers.quantization import QuantizationConfig from vllm.model_executor.model_loader.utils import set_default_torch_dtype -from vllm.model_executor.sampling_metadata import SamplingMetadata from vllm.multimodal import MULTIMODAL_REGISTRY from vllm.multimodal.inputs import (MultiModalDataDict, MultiModalFieldConfig, MultiModalKwargsItems) @@ -784,9 +783,8 @@ class MiDashengLMModel(nn.Module, SupportsMultiModal, SupportsPP): def compute_logits( self, hidden_states: torch.Tensor, - sampling_metadata: SamplingMetadata, ) -> Optional[torch.Tensor]: - return self.decoder.compute_logits(hidden_states, sampling_metadata) + return self.decoder.compute_logits(hidden_states) def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]: diff --git a/vllm/model_executor/models/mimo.py b/vllm/model_executor/models/mimo.py index ea5292d0df202..d256c1f3eed7b 100644 --- a/vllm/model_executor/models/mimo.py +++ b/vllm/model_executor/models/mimo.py @@ -41,7 +41,6 @@ from vllm.model_executor.layers.vocab_parallel_embedding import ParallelLMHead from vllm.model_executor.model_loader.weight_utils import ( default_weight_loader, maybe_remap_kv_scale_name) from vllm.model_executor.models.qwen2 import Qwen2ForCausalLM, Qwen2Model -from vllm.model_executor.sampling_metadata import SamplingMetadata from vllm.sequence import IntermediateTensors from .utils import PPMissingLayer, is_pp_missing_parameter, maybe_prefix @@ -183,9 +182,7 @@ class MiMoForCausalLM(Qwen2ForCausalLM, nn.Module): def compute_logits( self, hidden_states: torch.Tensor, - sampling_metadata: SamplingMetadata, ) -> Optional[torch.Tensor]: hidden_states = self.model.norm(hidden_states) - logits = self.logits_processor(self.lm_head, hidden_states, - sampling_metadata) + logits = self.logits_processor(self.lm_head, hidden_states) return logits diff --git a/vllm/model_executor/models/mimo_mtp.py b/vllm/model_executor/models/mimo_mtp.py index 09194e9f95d0e..b4abe458e4771 100644 --- a/vllm/model_executor/models/mimo_mtp.py +++ b/vllm/model_executor/models/mimo_mtp.py @@ -34,7 +34,6 @@ from vllm.model_executor.layers.vocab_parallel_embedding import ( ParallelLMHead, VocabParallelEmbedding) from vllm.model_executor.model_loader.weight_utils import default_weight_loader from vllm.model_executor.models.qwen2 import Qwen2DecoderLayer -from vllm.model_executor.sampling_metadata import SamplingMetadata from vllm.sequence import IntermediateTensors from .utils import maybe_prefix @@ -140,12 +139,10 @@ class MiMoMultiTokenPredictor(nn.Module): self, hidden_states: torch.Tensor, lm_head: ParallelLMHead, - sampling_metadata: SamplingMetadata, spec_step_idx: int = 0, ) -> torch.Tensor: self.mtp_layers[str(self.mtp_start_layer_idx + spec_step_idx)] - logits = self.logits_processor(lm_head, hidden_states, - sampling_metadata) + logits = self.logits_processor(lm_head, hidden_states) return logits @@ -178,11 +175,10 @@ class MiMoMTP(nn.Module): def compute_logits( self, hidden_states: torch.Tensor, - sampling_metadata: SamplingMetadata, spec_step_idx: int = 0, ) -> Optional[torch.Tensor]: return self.model.compute_logits(hidden_states, self.lm_head, - sampling_metadata, spec_step_idx) + spec_step_idx) def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]: diff --git a/vllm/model_executor/models/minicpm.py b/vllm/model_executor/models/minicpm.py index 240c23ea2b25d..0986ea07406a9 100644 --- a/vllm/model_executor/models/minicpm.py +++ b/vllm/model_executor/models/minicpm.py @@ -51,7 +51,6 @@ from vllm.model_executor.layers.rotary_embedding import get_rope from vllm.model_executor.layers.vocab_parallel_embedding import ( DEFAULT_VOCAB_PADDING_SIZE, ParallelLMHead, VocabParallelEmbedding) from vllm.model_executor.model_loader.weight_utils import default_weight_loader -from vllm.model_executor.sampling_metadata import SamplingMetadata from vllm.model_executor.utils import set_weight_attrs from vllm.platforms import current_platform from vllm.sequence import IntermediateTensors @@ -583,10 +582,8 @@ class MiniCPMForCausalLM(nn.Module, SupportsLoRA, SupportsPP): def compute_logits( self, hidden_states: torch.Tensor, - sampling_metadata: SamplingMetadata, ) -> Optional[torch.Tensor]: - logits = self.logits_processor(self.lm_head, hidden_states, - sampling_metadata) + logits = self.logits_processor(self.lm_head, hidden_states) return logits def load_weights(self, weights: Iterable[tuple[str, diff --git a/vllm/model_executor/models/minicpm_eagle.py b/vllm/model_executor/models/minicpm_eagle.py index 848a97b8bb2a0..2af0d546ce63d 100644 --- a/vllm/model_executor/models/minicpm_eagle.py +++ b/vllm/model_executor/models/minicpm_eagle.py @@ -39,7 +39,6 @@ from vllm.model_executor.layers.quantization import QuantizationConfig from vllm.model_executor.layers.vocab_parallel_embedding import ( DEFAULT_VOCAB_PADDING_SIZE, ParallelLMHead, VocabParallelEmbedding) from vllm.model_executor.model_loader.weight_utils import default_weight_loader -from vllm.model_executor.sampling_metadata import SamplingMetadata from vllm.sequence import IntermediateTensors from .interfaces import SupportsLoRA, SupportsPP @@ -376,10 +375,8 @@ class EagleMiniCPMForCausalLM(nn.Module, SupportsLoRA, SupportsPP): def compute_logits( self, hidden_states: torch.Tensor, - sampling_metadata: SamplingMetadata, ) -> Optional[torch.Tensor]: - logits = self.logits_processor(self.lm_head, hidden_states, - sampling_metadata) + logits = self.logits_processor(self.lm_head, hidden_states) return logits def load_weights(self, weights: Iterable[tuple[str, diff --git a/vllm/model_executor/models/minicpmv.py b/vllm/model_executor/models/minicpmv.py index 9b2d84e32151a..a17c4f004d75c 100644 --- a/vllm/model_executor/models/minicpmv.py +++ b/vllm/model_executor/models/minicpmv.py @@ -50,7 +50,6 @@ from vllm.model_executor.models.minicpm import MiniCPMForCausalLM from vllm.model_executor.models.module_mapping import MultiModelKeys from vllm.model_executor.models.qwen2 import Qwen2ForCausalLM from vllm.model_executor.models.qwen3 import Qwen3ForCausalLM -from vllm.model_executor.sampling_metadata import SamplingMetadata from vllm.multimodal import MULTIMODAL_REGISTRY from vllm.multimodal.inputs import (MultiModalDataDict, MultiModalFieldConfig, MultiModalKwargsItems, NestedTensors) @@ -1194,9 +1193,8 @@ class MiniCPMVBaseModel(nn.Module, SupportsMultiModal, SupportsPP): def compute_logits( self, hidden_states: torch.Tensor, - sampling_metadata: SamplingMetadata, ) -> Optional[torch.Tensor]: - return self.llm.compute_logits(hidden_states, sampling_metadata) + return self.llm.compute_logits(hidden_states) def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]: diff --git a/vllm/model_executor/models/minimax_text_01.py b/vllm/model_executor/models/minimax_text_01.py index 6ce883be0a83c..1d2c7dea811e0 100644 --- a/vllm/model_executor/models/minimax_text_01.py +++ b/vllm/model_executor/models/minimax_text_01.py @@ -41,7 +41,6 @@ from vllm.model_executor.layers.vocab_parallel_embedding import ( DEFAULT_VOCAB_PADDING_SIZE, ParallelLMHead, VocabParallelEmbedding) from vllm.model_executor.model_loader.weight_utils import default_weight_loader from vllm.model_executor.models.utils import maybe_prefix -from vllm.model_executor.sampling_metadata import SamplingMetadata from vllm.sequence import IntermediateTensors from .interfaces import HasInnerState, IsHybrid @@ -742,10 +741,8 @@ class MiniMaxText01ForCausalLM(nn.Module, HasInnerState, IsHybrid): return hidden_states - def compute_logits(self, hidden_states: torch.Tensor, - sampling_metadata: SamplingMetadata) -> torch.Tensor: - logits = self.logits_processor(self.lm_head, hidden_states.float(), - sampling_metadata) + def compute_logits(self, hidden_states: torch.Tensor) -> torch.Tensor: + logits = self.logits_processor(self.lm_head, hidden_states.float()) return logits diff --git a/vllm/model_executor/models/minimax_vl_01.py b/vllm/model_executor/models/minimax_vl_01.py index cc7db849a28bf..b2f020f3323e8 100644 --- a/vllm/model_executor/models/minimax_vl_01.py +++ b/vllm/model_executor/models/minimax_vl_01.py @@ -14,7 +14,6 @@ from vllm.model_executor.layers.activation import get_act_fn from vllm.model_executor.layers.linear import (ColumnParallelLinear, RowParallelLinear) from vllm.model_executor.layers.quantization import QuantizationConfig -from vllm.model_executor.sampling_metadata import SamplingMetadata from vllm.multimodal import MULTIMODAL_REGISTRY from vllm.multimodal.inputs import MultiModalFieldConfig from vllm.sequence import IntermediateTensors @@ -420,10 +419,8 @@ class MiniMaxVL01ForConditionalGeneration(nn.Module, SupportsMultiModal, def compute_logits( self, hidden_states: torch.Tensor, - sampling_metadata: SamplingMetadata, ) -> Optional[torch.Tensor]: - return self.language_model.compute_logits(hidden_states, - sampling_metadata) + return self.language_model.compute_logits(hidden_states) def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]: diff --git a/vllm/model_executor/models/mistral3.py b/vllm/model_executor/models/mistral3.py index d15776a39362d..94e3d7234b6f4 100644 --- a/vllm/model_executor/models/mistral3.py +++ b/vllm/model_executor/models/mistral3.py @@ -20,7 +20,6 @@ from vllm.model_executor.layers.linear import (ColumnParallelLinear, RowParallelLinear) from vllm.model_executor.layers.quantization import QuantizationConfig from vllm.model_executor.models.module_mapping import MultiModelKeys -from vllm.model_executor.sampling_metadata import SamplingMetadata from vllm.multimodal import MULTIMODAL_REGISTRY from vllm.multimodal.cache import BaseMultiModalProcessorCache from vllm.multimodal.inputs import (MultiModalDataDict, MultiModalFieldConfig, @@ -606,10 +605,8 @@ class Mistral3ForConditionalGeneration(nn.Module, SupportsLoRA, def compute_logits( self, hidden_states: torch.Tensor, - sampling_metadata: SamplingMetadata, ) -> Optional[torch.Tensor]: - return self.language_model.compute_logits(hidden_states, - sampling_metadata) + return self.language_model.compute_logits(hidden_states) def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]: diff --git a/vllm/model_executor/models/mixtral.py b/vllm/model_executor/models/mixtral.py index 8b3474d809532..bebf0b5adac52 100644 --- a/vllm/model_executor/models/mixtral.py +++ b/vllm/model_executor/models/mixtral.py @@ -49,7 +49,6 @@ from vllm.model_executor.layers.vocab_parallel_embedding import ( DEFAULT_VOCAB_PADDING_SIZE, ParallelLMHead, VocabParallelEmbedding) from vllm.model_executor.model_loader.weight_utils import ( default_weight_loader, maybe_remap_kv_scale_name) -from vllm.model_executor.sampling_metadata import SamplingMetadata from vllm.sequence import IntermediateTensors from .interfaces import MixtureOfExperts, SupportsLoRA, SupportsPP @@ -594,10 +593,8 @@ class MixtralForCausalLM(nn.Module, SupportsLoRA, SupportsPP, def compute_logits( self, hidden_states: torch.Tensor, - sampling_metadata: SamplingMetadata, ) -> Optional[torch.Tensor]: - logits = self.logits_processor(self.lm_head, hidden_states, - sampling_metadata) + logits = self.logits_processor(self.lm_head, hidden_states) return logits def load_weights(self, weights: Iterable[tuple[str, diff --git a/vllm/model_executor/models/mllama4.py b/vllm/model_executor/models/mllama4.py index 2f0e8a2a5e575..131a66b713235 100644 --- a/vllm/model_executor/models/mllama4.py +++ b/vllm/model_executor/models/mllama4.py @@ -41,7 +41,6 @@ from vllm.model_executor.layers.quantization import QuantizationConfig from vllm.model_executor.layers.rotary_embedding import get_rope from vllm.model_executor.model_loader.utils import initialize_model from vllm.model_executor.model_loader.weight_utils import default_weight_loader -from vllm.model_executor.sampling_metadata import SamplingMetadata from vllm.multimodal import MULTIMODAL_REGISTRY from vllm.multimodal.inputs import (MultiModalDataDict, MultiModalFieldConfig, MultiModalKwargsItems, NestedTensors) @@ -856,10 +855,8 @@ class Llama4ForConditionalGeneration(nn.Module, SupportsMultiModal, def compute_logits( self, hidden_states: torch.Tensor, - sampling_metadata: SamplingMetadata, ) -> Optional[torch.Tensor]: - return self.language_model.compute_logits(hidden_states, - sampling_metadata) + return self.language_model.compute_logits(hidden_states) def separate_weights( self, diff --git a/vllm/model_executor/models/molmo.py b/vllm/model_executor/models/molmo.py index 2475fe1316097..201bf83cac581 100644 --- a/vllm/model_executor/models/molmo.py +++ b/vllm/model_executor/models/molmo.py @@ -26,7 +26,6 @@ from vllm.distributed import (get_pp_group, get_tensor_model_parallel_rank, get_tensor_model_parallel_world_size, split_tensor_along_last_dim, tensor_model_parallel_all_gather) -from vllm.model_executor import SamplingMetadata from vllm.model_executor.layers.activation import (MulAndSilu, QuickGELU, SiluAndMul) from vllm.model_executor.layers.layernorm import RMSNorm @@ -1527,10 +1526,8 @@ class MolmoForCausalLM(nn.Module, SupportsMultiModal, SupportsPP, SupportsLoRA, return hidden_states - def compute_logits(self, hidden_states: torch.Tensor, - sampling_metadata: SamplingMetadata) -> torch.Tensor: - logits = self.logits_processor(self.lm_head, hidden_states, - sampling_metadata) + def compute_logits(self, hidden_states: torch.Tensor) -> torch.Tensor: + logits = self.logits_processor(self.lm_head, hidden_states) return logits def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]): diff --git a/vllm/model_executor/models/mpt.py b/vllm/model_executor/models/mpt.py index 48ac91fa6dde0..64d669e8ac3e1 100644 --- a/vllm/model_executor/models/mpt.py +++ b/vllm/model_executor/models/mpt.py @@ -25,7 +25,6 @@ from vllm.model_executor.layers.quantization import QuantizationConfig from vllm.model_executor.layers.vocab_parallel_embedding import ( VocabParallelEmbedding) from vllm.model_executor.model_loader.weight_utils import default_weight_loader -from vllm.model_executor.sampling_metadata import SamplingMetadata from vllm.sequence import IntermediateTensors from .interfaces import SupportsPP @@ -320,10 +319,8 @@ class MPTForCausalLM(nn.Module, SupportsPP): def compute_logits( self, hidden_states: torch.Tensor, - sampling_metadata: SamplingMetadata, ) -> Optional[torch.Tensor]: - logits = self.logits_processor(self.lm_head, hidden_states, - sampling_metadata) + logits = self.logits_processor(self.lm_head, hidden_states) return logits def load_weights(self, weights: Iterable[tuple[str, diff --git a/vllm/model_executor/models/nano_nemotron_vl.py b/vllm/model_executor/models/nano_nemotron_vl.py index 4f8652c006941..ae50f1aefc6f7 100644 --- a/vllm/model_executor/models/nano_nemotron_vl.py +++ b/vllm/model_executor/models/nano_nemotron_vl.py @@ -37,7 +37,6 @@ from vllm.model_executor.models.utils import (flatten_bn, init_vllm_registered_model, maybe_prefix, merge_multimodal_embeddings) -from vllm.model_executor.sampling_metadata import SamplingMetadata from vllm.multimodal import MULTIMODAL_REGISTRY from vllm.multimodal.inputs import (MultiModalDataDict, MultiModalFieldConfig, MultiModalKwargs, MultiModalKwargsItems, @@ -1192,10 +1191,8 @@ class NemotronH_Nano_VL(nn.Module, HasInnerState, IsHybrid, def compute_logits( self, hidden_states: torch.Tensor, - sampling_metadata: SamplingMetadata, ) -> Optional[torch.Tensor]: - return self.language_model.compute_logits(hidden_states, - sampling_metadata) + return self.language_model.compute_logits(hidden_states) def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]): adapter_dict = dict(self.mlp1.named_parameters()) diff --git a/vllm/model_executor/models/nemotron.py b/vllm/model_executor/models/nemotron.py index 21f785e4b91af..6bb2f7392cb49 100644 --- a/vllm/model_executor/models/nemotron.py +++ b/vllm/model_executor/models/nemotron.py @@ -45,7 +45,6 @@ from vllm.model_executor.layers.vocab_parallel_embedding import ( DEFAULT_VOCAB_PADDING_SIZE, ParallelLMHead, VocabParallelEmbedding) from vllm.model_executor.model_loader.weight_utils import ( default_weight_loader, maybe_remap_kv_scale_name) -from vllm.model_executor.sampling_metadata import SamplingMetadata from vllm.sequence import IntermediateTensors from vllm.transformers_utils.configs import NemotronConfig @@ -498,10 +497,8 @@ class NemotronForCausalLM(nn.Module, SupportsLoRA, SupportsPP): def compute_logits( self, hidden_states: torch.Tensor, - sampling_metadata: SamplingMetadata, ) -> Optional[torch.Tensor]: - logits = self.logits_processor(self.lm_head, hidden_states, - sampling_metadata) + logits = self.logits_processor(self.lm_head, hidden_states) return logits def load_weights(self, weights: Iterable[tuple[str, diff --git a/vllm/model_executor/models/nemotron_h.py b/vllm/model_executor/models/nemotron_h.py index 1e1f0524bd063..ff571541a60a5 100644 --- a/vllm/model_executor/models/nemotron_h.py +++ b/vllm/model_executor/models/nemotron_h.py @@ -54,7 +54,6 @@ from vllm.model_executor.models.mamba_cache import (MambaCacheManager, from vllm.model_executor.models.utils import ( AutoWeightsLoader, WeightsMapper, make_empty_intermediate_tensors_factory, make_layers, maybe_prefix) -from vllm.model_executor.sampling_metadata import SamplingMetadata from vllm.sequence import IntermediateTensors from vllm.transformers_utils.configs import NemotronHConfig from vllm.utils import LayerBlockType @@ -622,10 +621,8 @@ class NemotronHForCausalLM(nn.Module, HasInnerState, SupportsLoRA, SupportsPP, def compute_logits( self, hidden_states: torch.Tensor, - sampling_metadata: SamplingMetadata, ) -> Optional[torch.Tensor]: - logits = self.logits_processor(self.lm_head, hidden_states, - sampling_metadata) + logits = self.logits_processor(self.lm_head, hidden_states) return logits def load_weights(self, weights: Iterable[tuple[str, diff --git a/vllm/model_executor/models/nemotron_nas.py b/vllm/model_executor/models/nemotron_nas.py index f8e38dcd80b5a..d474c8db41b2f 100644 --- a/vllm/model_executor/models/nemotron_nas.py +++ b/vllm/model_executor/models/nemotron_nas.py @@ -44,7 +44,6 @@ from vllm.model_executor.layers.vocab_parallel_embedding import ( from vllm.model_executor.model_loader.weight_utils import ( default_weight_loader, maybe_remap_kv_scale_name) from vllm.model_executor.models.llama import LlamaAttention, LlamaMLP -from vllm.model_executor.sampling_metadata import SamplingMetadata from vllm.sequence import IntermediateTensors from .interfaces import HasNoOps, SupportsLoRA, SupportsPP @@ -468,10 +467,8 @@ class DeciLMForCausalLM(nn.Module, SupportsLoRA, SupportsPP, HasNoOps): def compute_logits( self, hidden_states: torch.Tensor, - sampling_metadata: SamplingMetadata, ) -> Optional[torch.Tensor]: - logits = self.logits_processor(self.lm_head, hidden_states, - sampling_metadata) + logits = self.logits_processor(self.lm_head, hidden_states) return logits def load_weights(self, weights: Iterable[tuple[str, diff --git a/vllm/model_executor/models/nemotron_vl.py b/vllm/model_executor/models/nemotron_vl.py index acda2027401d9..3abbff8c717d4 100644 --- a/vllm/model_executor/models/nemotron_vl.py +++ b/vllm/model_executor/models/nemotron_vl.py @@ -26,7 +26,6 @@ from vllm.model_executor.models.internvl import ( BaseInternVLProcessingInfo, InternVLImageEmbeddingInputs, InternVLImageInputs, InternVLImagePixelInputs, InternVLProcessor) from vllm.model_executor.models.module_mapping import MultiModelKeys -from vllm.model_executor.sampling_metadata import SamplingMetadata from vllm.multimodal import MULTIMODAL_REGISTRY from vllm.multimodal.image import convert_image_mode from vllm.multimodal.inputs import NestedTensors @@ -632,10 +631,8 @@ class LlamaNemotronVLChatModel(nn.Module, SupportsMultiModal, SupportsPP, def compute_logits( self, hidden_states: torch.Tensor, - sampling_metadata: SamplingMetadata, ) -> Optional[torch.Tensor]: - return self.language_model.compute_logits(hidden_states, - sampling_metadata) + return self.language_model.compute_logits(hidden_states) def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]: diff --git a/vllm/model_executor/models/olmo.py b/vllm/model_executor/models/olmo.py index 7be3c16528b52..9fa8760073c15 100644 --- a/vllm/model_executor/models/olmo.py +++ b/vllm/model_executor/models/olmo.py @@ -45,7 +45,6 @@ from vllm.model_executor.layers.rotary_embedding import get_rope from vllm.model_executor.layers.vocab_parallel_embedding import ( ParallelLMHead, VocabParallelEmbedding) from vllm.model_executor.model_loader.weight_utils import default_weight_loader -from vllm.model_executor.sampling_metadata import SamplingMetadata from vllm.sequence import IntermediateTensors from .interfaces import SupportsLoRA, SupportsPP @@ -391,10 +390,8 @@ class OlmoForCausalLM(nn.Module, SupportsPP, SupportsLoRA): def compute_logits( self, hidden_states: torch.Tensor, - sampling_metadata: SamplingMetadata, ) -> Optional[torch.Tensor]: - logits = self.logits_processor(self.lm_head, hidden_states, - sampling_metadata) + logits = self.logits_processor(self.lm_head, hidden_states) return logits def load_weights(self, weights: Iterable[tuple[str, diff --git a/vllm/model_executor/models/olmo2.py b/vllm/model_executor/models/olmo2.py index 3e4c580a11211..2e0b1fb2a13f7 100644 --- a/vllm/model_executor/models/olmo2.py +++ b/vllm/model_executor/models/olmo2.py @@ -54,7 +54,6 @@ from vllm.model_executor.models.interfaces import SupportsLoRA, SupportsPP from vllm.model_executor.models.utils import ( AutoWeightsLoader, extract_layer_index, is_pp_missing_parameter, make_empty_intermediate_tensors_factory, make_layers, maybe_prefix) -from vllm.model_executor.sampling_metadata import SamplingMetadata from vllm.sequence import IntermediateTensors from vllm.transformers_utils.configs import Olmo3Config @@ -427,10 +426,8 @@ class Olmo2ForCausalLM(nn.Module, SupportsPP, SupportsLoRA): def compute_logits( self, hidden_states: torch.Tensor, - sampling_metadata: SamplingMetadata, ) -> Optional[torch.Tensor]: - logits = self.logits_processor(self.lm_head, hidden_states, - sampling_metadata) + logits = self.logits_processor(self.lm_head, hidden_states) return logits def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]): diff --git a/vllm/model_executor/models/olmoe.py b/vllm/model_executor/models/olmoe.py index 892e967e4a21f..77ece544d4900 100644 --- a/vllm/model_executor/models/olmoe.py +++ b/vllm/model_executor/models/olmoe.py @@ -41,7 +41,6 @@ from vllm.model_executor.layers.rotary_embedding import get_rope from vllm.model_executor.layers.vocab_parallel_embedding import ( ParallelLMHead, VocabParallelEmbedding) from vllm.model_executor.model_loader.weight_utils import default_weight_loader -from vllm.model_executor.sampling_metadata import SamplingMetadata from vllm.sequence import IntermediateTensors from .interfaces import SupportsPP @@ -471,10 +470,8 @@ class OlmoeForCausalLM(nn.Module, SupportsPP): inputs_embeds) return hidden_states - def compute_logits(self, hidden_states: torch.Tensor, - sampling_metadata: SamplingMetadata) -> torch.Tensor: - logits = self.logits_processor(self.lm_head, hidden_states, - sampling_metadata) + def compute_logits(self, hidden_states: torch.Tensor) -> torch.Tensor: + logits = self.logits_processor(self.lm_head, hidden_states) return logits def load_weights(self, weights: Iterable[tuple[str, diff --git a/vllm/model_executor/models/opt.py b/vllm/model_executor/models/opt.py index 365aab205b211..4c3ce9f61efb3 100644 --- a/vllm/model_executor/models/opt.py +++ b/vllm/model_executor/models/opt.py @@ -41,7 +41,6 @@ from vllm.model_executor.layers.quantization import QuantizationConfig from vllm.model_executor.layers.vocab_parallel_embedding import ( ParallelLMHead, VocabParallelEmbedding) from vllm.model_executor.model_loader.weight_utils import default_weight_loader -from vllm.model_executor.sampling_metadata import SamplingMetadata from vllm.sequence import IntermediateTensors from .interfaces import SupportsPP @@ -399,10 +398,8 @@ class OPTForCausalLM(nn.Module, SupportsPP): def compute_logits( self, hidden_states: torch.Tensor, - sampling_metadata: SamplingMetadata, ) -> Optional[torch.Tensor]: - logits = self.logits_processor(self.lm_head, hidden_states, - sampling_metadata) + logits = self.logits_processor(self.lm_head, hidden_states) return logits def load_weights(self, weights: Iterable[tuple[str, diff --git a/vllm/model_executor/models/orion.py b/vllm/model_executor/models/orion.py index 944a9151d75d3..586fea343d6f9 100644 --- a/vllm/model_executor/models/orion.py +++ b/vllm/model_executor/models/orion.py @@ -28,7 +28,6 @@ from vllm.model_executor.layers.rotary_embedding import get_rope from vllm.model_executor.layers.vocab_parallel_embedding import ( ParallelLMHead, VocabParallelEmbedding) from vllm.model_executor.model_loader.weight_utils import default_weight_loader -from vllm.model_executor.sampling_metadata import SamplingMetadata from vllm.sequence import IntermediateTensors from .interfaces import SupportsPP @@ -339,10 +338,8 @@ class OrionForCausalLM(nn.Module, SupportsPP): def compute_logits( self, hidden_states: torch.Tensor, - sampling_metadata: SamplingMetadata, ) -> Optional[torch.Tensor]: - logits = self.logits_processor(self.lm_head, hidden_states, - sampling_metadata) + logits = self.logits_processor(self.lm_head, hidden_states) return logits def load_weights(self, weights: Iterable[tuple[str, diff --git a/vllm/model_executor/models/ovis.py b/vllm/model_executor/models/ovis.py index f1bb18716b40d..052e143b27f6e 100644 --- a/vllm/model_executor/models/ovis.py +++ b/vllm/model_executor/models/ovis.py @@ -39,7 +39,6 @@ from vllm.model_executor.models.siglip import SiglipVisionModel from vllm.model_executor.models.utils import (AutoWeightsLoader, flatten_bn, init_vllm_registered_model, maybe_prefix) -from vllm.model_executor.sampling_metadata import SamplingMetadata from vllm.multimodal import MULTIMODAL_REGISTRY from vllm.multimodal.inputs import (MultiModalDataDict, MultiModalFieldConfig, MultiModalKwargsItems) @@ -558,9 +557,8 @@ class Ovis(nn.Module, SupportsMultiModal, SupportsPP): def compute_logits( self, hidden_states: torch.Tensor, - sampling_metadata: SamplingMetadata, ) -> Optional[torch.Tensor]: - logits = self.llm.compute_logits(hidden_states, sampling_metadata) + logits = self.llm.compute_logits(hidden_states) return logits def load_weights(self, weights: Iterable[tuple[str, diff --git a/vllm/model_executor/models/ovis2_5.py b/vllm/model_executor/models/ovis2_5.py index 5e4758ef8ea5d..f18e38ce154d2 100644 --- a/vllm/model_executor/models/ovis2_5.py +++ b/vllm/model_executor/models/ovis2_5.py @@ -19,7 +19,6 @@ from vllm.model_executor.models.siglip2navit import Siglip2NavitModel from vllm.model_executor.models.utils import (AutoWeightsLoader, flatten_bn, init_vllm_registered_model, maybe_prefix) -from vllm.model_executor.sampling_metadata import SamplingMetadata from vllm.multimodal import MULTIMODAL_REGISTRY from vllm.multimodal.inputs import (MultiModalDataDict, MultiModalFieldConfig, MultiModalKwargsItems) @@ -630,9 +629,8 @@ class Ovis2_5(nn.Module, SupportsMultiModal, SupportsPP): def compute_logits( self, hidden_states: torch.Tensor, - sampling_metadata: SamplingMetadata, ) -> Optional[torch.Tensor]: - logits = self.llm.compute_logits(hidden_states, sampling_metadata) + logits = self.llm.compute_logits(hidden_states) return logits def load_weights(self, weights: Iterable[tuple[str, diff --git a/vllm/model_executor/models/paligemma.py b/vllm/model_executor/models/paligemma.py index d6eec77ebcee5..aef5102304614 100644 --- a/vllm/model_executor/models/paligemma.py +++ b/vllm/model_executor/models/paligemma.py @@ -9,7 +9,6 @@ from transformers import BatchFeature, PaliGemmaConfig from vllm.config import VllmConfig from vllm.logger import init_logger -from vllm.model_executor.sampling_metadata import SamplingMetadata from vllm.multimodal import MULTIMODAL_REGISTRY from vllm.multimodal.inputs import (MultiModalDataDict, MultiModalFieldConfig, MultiModalInputs, MultiModalKwargsItems, @@ -403,10 +402,8 @@ class PaliGemmaForConditionalGeneration(nn.Module, SupportsMultiModal, def compute_logits( self, hidden_states: torch.Tensor, - sampling_metadata: SamplingMetadata, ) -> Optional[torch.Tensor]: - return self.language_model.compute_logits(hidden_states, - sampling_metadata) + return self.language_model.compute_logits(hidden_states) def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]: diff --git a/vllm/model_executor/models/persimmon.py b/vllm/model_executor/models/persimmon.py index 3e854e4d561ff..23fb7bb85215c 100644 --- a/vllm/model_executor/models/persimmon.py +++ b/vllm/model_executor/models/persimmon.py @@ -44,7 +44,6 @@ from vllm.model_executor.layers.rotary_embedding import get_rope from vllm.model_executor.layers.vocab_parallel_embedding import ( ParallelLMHead, VocabParallelEmbedding) from vllm.model_executor.model_loader.weight_utils import default_weight_loader -from vllm.model_executor.sampling_metadata import SamplingMetadata from vllm.sequence import IntermediateTensors from .interfaces import SupportsPP @@ -334,10 +333,8 @@ class PersimmonForCausalLM(nn.Module, SupportsPP): def compute_logits( self, hidden_states: torch.Tensor, - sampling_metadata: SamplingMetadata, ) -> Optional[torch.Tensor]: - logits = self.logits_processor(self.lm_head, hidden_states, - sampling_metadata) + logits = self.logits_processor(self.lm_head, hidden_states) return logits def load_weights(self, weights: Iterable[tuple[str, diff --git a/vllm/model_executor/models/phi.py b/vllm/model_executor/models/phi.py index 6f39afbecf35b..9cf288e850057 100644 --- a/vllm/model_executor/models/phi.py +++ b/vllm/model_executor/models/phi.py @@ -59,7 +59,6 @@ from vllm.model_executor.layers.rotary_embedding import get_rope from vllm.model_executor.layers.vocab_parallel_embedding import ( ParallelLMHead, VocabParallelEmbedding) from vllm.model_executor.model_loader.weight_utils import default_weight_loader -from vllm.model_executor.sampling_metadata import SamplingMetadata from vllm.sequence import IntermediateTensors from .interfaces import SupportsLoRA, SupportsPP @@ -346,10 +345,9 @@ class PhiForCausalLM(nn.Module, SupportsLoRA, SupportsPP): def compute_logits( self, hidden_states: torch.Tensor, - sampling_metadata: SamplingMetadata, ) -> Optional[torch.Tensor]: logits = self.logits_processor(self.lm_head, hidden_states, - sampling_metadata, self.lm_head.bias) + self.lm_head.bias) return logits def load_weights(self, weights: Iterable[tuple[str, diff --git a/vllm/model_executor/models/phi3v.py b/vllm/model_executor/models/phi3v.py index 4522c7043d01a..a2b201fe4228d 100644 --- a/vllm/model_executor/models/phi3v.py +++ b/vllm/model_executor/models/phi3v.py @@ -29,7 +29,6 @@ from vllm.logger import init_logger from vllm.model_executor.layers.quantization import QuantizationConfig from vllm.model_executor.layers.vocab_parallel_embedding import ( VocabParallelEmbedding) -from vllm.model_executor.sampling_metadata import SamplingMetadata from vllm.multimodal import MULTIMODAL_REGISTRY from vllm.multimodal.inputs import (MultiModalDataDict, MultiModalFieldConfig, MultiModalKwargsItems) @@ -681,10 +680,8 @@ class Phi3VForCausalLM(nn.Module, SupportsMultiModal, SupportsPP, def compute_logits( self, hidden_states: torch.Tensor, - sampling_metadata: SamplingMetadata, ) -> Optional[torch.Tensor]: - return self.language_model.compute_logits(hidden_states, - sampling_metadata) + return self.language_model.compute_logits(hidden_states) def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]: diff --git a/vllm/model_executor/models/phi4_multimodal.py b/vllm/model_executor/models/phi4_multimodal.py index 25df9e9261d91..d2a3a8cc04969 100644 --- a/vllm/model_executor/models/phi4_multimodal.py +++ b/vllm/model_executor/models/phi4_multimodal.py @@ -27,7 +27,6 @@ from vllm.model_executor.layers.linear import (ColumnParallelLinear, from vllm.model_executor.layers.quantization import QuantizationConfig from vllm.model_executor.model_loader.weight_utils import default_weight_loader from vllm.model_executor.models.module_mapping import MultiModelKeys -from vllm.model_executor.sampling_metadata import SamplingMetadata from vllm.multimodal import MULTIMODAL_REGISTRY from vllm.multimodal.inputs import (MultiModalDataDict, MultiModalFieldConfig, MultiModalKwargsItems, NestedTensors) @@ -1451,10 +1450,8 @@ class Phi4MultimodalForCausalLM(nn.Module, SupportsLoRA, SupportsMultiModal): def compute_logits( self, hidden_states: torch.Tensor, - sampling_metadata: SamplingMetadata, ) -> Optional[torch.Tensor]: - return self.language_model.compute_logits(hidden_states, - sampling_metadata) + return self.language_model.compute_logits(hidden_states) def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]: diff --git a/vllm/model_executor/models/phi4flash.py b/vllm/model_executor/models/phi4flash.py index aa7c434a44aeb..ae153558e37aa 100644 --- a/vllm/model_executor/models/phi4flash.py +++ b/vllm/model_executor/models/phi4flash.py @@ -29,7 +29,6 @@ from vllm.model_executor.models.interfaces import (HasInnerState, IsHybrid, SupportsV0Only) from vllm.model_executor.models.mamba_cache import (MambaCacheManager, MambaCacheParams) -from vllm.model_executor.sampling_metadata import SamplingMetadata from vllm.sequence import IntermediateTensors from .utils import make_layers, maybe_prefix @@ -695,12 +694,10 @@ class Phi4FlashForCausalLM(nn.Module, HasInnerState, IsHybrid, SupportsV0Only): def compute_logits( self, hidden_states: torch.Tensor, - sampling_metadata: SamplingMetadata, ) -> Optional[torch.Tensor]: processed_logits = self.logits_processor( self.lm_head, hidden_states, - sampling_metadata, self.embedding_bias, ) return processed_logits diff --git a/vllm/model_executor/models/phi4mm.py b/vllm/model_executor/models/phi4mm.py index b3fc55dab6eca..47b5ad55ab2d0 100644 --- a/vllm/model_executor/models/phi4mm.py +++ b/vllm/model_executor/models/phi4mm.py @@ -18,7 +18,6 @@ from vllm.model_executor.layers.vocab_parallel_embedding import ( DEFAULT_VOCAB_PADDING_SIZE, ParallelLMHead) from vllm.model_executor.models.llama import LlamaModel from vllm.model_executor.models.module_mapping import MultiModelKeys -from vllm.model_executor.sampling_metadata import SamplingMetadata from vllm.multimodal import MULTIMODAL_REGISTRY from vllm.multimodal.inputs import (MultiModalDataDict, MultiModalFieldConfig, MultiModalKwargsItems, NestedTensors) @@ -1257,10 +1256,8 @@ class Phi4MMForCausalLM(nn.Module, SupportsLoRA, SupportsMultiModal): def compute_logits( self, hidden_states: torch.Tensor, - sampling_metadata: SamplingMetadata, ) -> Optional[torch.Tensor]: - logits = self.logits_processor(self.lm_head, hidden_states, - sampling_metadata) + logits = self.logits_processor(self.lm_head, hidden_states) return logits def load_weights(self, weights: Iterable[tuple[str, diff --git a/vllm/model_executor/models/phimoe.py b/vllm/model_executor/models/phimoe.py index 01d16f1f2c387..3ce67ce37a7ab 100644 --- a/vllm/model_executor/models/phimoe.py +++ b/vllm/model_executor/models/phimoe.py @@ -47,7 +47,6 @@ from vllm.model_executor.layers.vocab_parallel_embedding import ( DEFAULT_VOCAB_PADDING_SIZE, ParallelLMHead, VocabParallelEmbedding) from vllm.model_executor.model_loader.weight_utils import ( default_weight_loader, maybe_remap_kv_scale_name) -from vllm.model_executor.sampling_metadata import SamplingMetadata from vllm.sequence import IntermediateTensors from .interfaces import SupportsLoRA, SupportsPP @@ -667,10 +666,8 @@ class PhiMoEForCausalLM(nn.Module, SupportsLoRA, SupportsPP): inputs_embeds) return hidden_states - def compute_logits(self, hidden_states: torch.Tensor, - sampling_metadata: SamplingMetadata) -> torch.Tensor: - logits = self.logits_processor(self.lm_head, hidden_states, - sampling_metadata) + def compute_logits(self, hidden_states: torch.Tensor) -> torch.Tensor: + logits = self.logits_processor(self.lm_head, hidden_states) return logits def load_weights(self, weights: Iterable[tuple[str, diff --git a/vllm/model_executor/models/pixtral.py b/vllm/model_executor/models/pixtral.py index 142d3251bc67a..7b197844c8b63 100644 --- a/vllm/model_executor/models/pixtral.py +++ b/vllm/model_executor/models/pixtral.py @@ -32,7 +32,6 @@ from vllm.model_executor.layers.linear import (MergedColumnParallelLinear, RowParallelLinear) from vllm.model_executor.layers.quantization import QuantizationConfig from vllm.model_executor.model_loader.weight_utils import default_weight_loader -from vllm.model_executor.sampling_metadata import SamplingMetadata from vllm.multimodal import MULTIMODAL_REGISTRY, MultiModalKwargsItems from vllm.multimodal.inputs import (MultiModalDataDict, MultiModalFieldConfig, MultiModalUUIDDict, NestedTensors) @@ -480,10 +479,8 @@ class PixtralForConditionalGeneration(nn.Module, SupportsMultiModal, def compute_logits( self, hidden_states: torch.Tensor, - sampling_metadata: SamplingMetadata, ) -> Optional[torch.Tensor]: - return self.language_model.compute_logits(hidden_states, - sampling_metadata) + return self.language_model.compute_logits(hidden_states) def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]): diff --git a/vllm/model_executor/models/plamo2.py b/vllm/model_executor/models/plamo2.py index 9f1ee36366fdd..33ee1cf44afd1 100644 --- a/vllm/model_executor/models/plamo2.py +++ b/vllm/model_executor/models/plamo2.py @@ -52,7 +52,6 @@ from vllm.model_executor.models.mamba_cache import (MambaCacheManager, from vllm.model_executor.models.utils import ( is_pp_missing_parameter, make_empty_intermediate_tensors_factory, make_layers, maybe_prefix) -from vllm.model_executor.sampling_metadata import SamplingMetadata from vllm.model_executor.utils import set_weight_attrs from vllm.platforms import current_platform from vllm.sequence import IntermediateTensors @@ -1022,10 +1021,8 @@ class Plamo2ForCausalLM(torch.nn.Module, HasInnerState, SupportsPP, IsHybrid): def compute_logits( self, hidden_states: torch.Tensor, - sampling_metadata: SamplingMetadata, ) -> Optional[torch.Tensor]: - logits = self.logits_processor(self.lm_head, hidden_states, - sampling_metadata) + logits = self.logits_processor(self.lm_head, hidden_states) return logits def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]): diff --git a/vllm/model_executor/models/qwen.py b/vllm/model_executor/models/qwen.py index 7470948499005..e0c08a6a88271 100644 --- a/vllm/model_executor/models/qwen.py +++ b/vllm/model_executor/models/qwen.py @@ -30,7 +30,6 @@ from vllm.model_executor.layers.rotary_embedding import get_rope from vllm.model_executor.layers.vocab_parallel_embedding import ( ParallelLMHead, VocabParallelEmbedding) from vllm.model_executor.model_loader.weight_utils import default_weight_loader -from vllm.model_executor.sampling_metadata import SamplingMetadata from vllm.sequence import IntermediateTensors from .interfaces import SupportsLoRA, SupportsPP @@ -282,10 +281,8 @@ class QWenBaseModel(nn.Module): def compute_logits( self, hidden_states: torch.Tensor, - sampling_metadata: SamplingMetadata, ) -> Optional[torch.Tensor]: - logits = self.logits_processor(self.lm_head, hidden_states, - sampling_metadata) + logits = self.logits_processor(self.lm_head, hidden_states) return logits def load_weights(self, weights: Iterable[tuple[str, diff --git a/vllm/model_executor/models/qwen2.py b/vllm/model_executor/models/qwen2.py index e13e87b93429d..c536b0f60c30d 100644 --- a/vllm/model_executor/models/qwen2.py +++ b/vllm/model_executor/models/qwen2.py @@ -49,7 +49,6 @@ from vllm.model_executor.layers.vocab_parallel_embedding import ( ParallelLMHead, VocabParallelEmbedding) from vllm.model_executor.model_loader.weight_utils import ( default_weight_loader, maybe_remap_kv_scale_name) -from vllm.model_executor.sampling_metadata import SamplingMetadata from vllm.sequence import IntermediateTensors from vllm.transformers_utils.config import is_interleaved @@ -510,10 +509,8 @@ class Qwen2ForCausalLM(nn.Module, SupportsLoRA, SupportsPP, SupportsEagle3): def compute_logits( self, hidden_states: torch.Tensor, - sampling_metadata: SamplingMetadata, ) -> Optional[torch.Tensor]: - logits = self.logits_processor(self.lm_head, hidden_states, - sampling_metadata) + logits = self.logits_processor(self.lm_head, hidden_states) return logits def load_weights(self, weights: Iterable[tuple[str, diff --git a/vllm/model_executor/models/qwen2_5_omni_thinker.py b/vllm/model_executor/models/qwen2_5_omni_thinker.py index a7e71309b6074..5f27230c913b4 100644 --- a/vllm/model_executor/models/qwen2_5_omni_thinker.py +++ b/vllm/model_executor/models/qwen2_5_omni_thinker.py @@ -50,7 +50,6 @@ from vllm.model_executor.models.qwen2_5_vl import ( from vllm.model_executor.models.qwen2_audio import ( Qwen2AudioProcessingInfo, _get_feat_extract_output_lengths) from vllm.model_executor.models.qwen2_vl import Qwen2VLMultiModalDataParser -from vllm.model_executor.sampling_metadata import SamplingMetadata from vllm.multimodal import MULTIMODAL_REGISTRY from vllm.multimodal.inputs import (ImageItem, ModalityData, MultiModalDataDict, MultiModalFieldConfig, @@ -955,10 +954,8 @@ class Qwen2_5OmniThinkerForConditionalGeneration( def compute_logits( self, hidden_states: torch.Tensor, - sampling_metadata: SamplingMetadata, ) -> Optional[torch.Tensor]: - return self.language_model.compute_logits(hidden_states, - sampling_metadata) + return self.language_model.compute_logits(hidden_states) def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]: diff --git a/vllm/model_executor/models/qwen2_5_vl.py b/vllm/model_executor/models/qwen2_5_vl.py index dbf486374bcf3..73b27572a8ebd 100644 --- a/vllm/model_executor/models/qwen2_5_vl.py +++ b/vllm/model_executor/models/qwen2_5_vl.py @@ -43,7 +43,6 @@ from vllm.config import VllmConfig from vllm.distributed import parallel_state from vllm.distributed import utils as dist_utils from vllm.logger import init_logger -from vllm.model_executor import SamplingMetadata from vllm.model_executor.layers.activation import get_act_and_mul_fn from vllm.model_executor.layers.layernorm import RMSNorm # yapf: disable @@ -1256,10 +1255,8 @@ class Qwen2_5_VLForConditionalGeneration(nn.Module, SupportsMultiModal, def compute_logits( self, hidden_states: torch.Tensor, - sampling_metadata: SamplingMetadata, ) -> Optional[torch.Tensor]: - return self.language_model.compute_logits(hidden_states, - sampling_metadata) + return self.language_model.compute_logits(hidden_states) def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]: diff --git a/vllm/model_executor/models/qwen2_audio.py b/vllm/model_executor/models/qwen2_audio.py index c797b71b5d2e1..762ab42e5929e 100644 --- a/vllm/model_executor/models/qwen2_audio.py +++ b/vllm/model_executor/models/qwen2_audio.py @@ -34,7 +34,6 @@ from transformers.models.qwen2_audio import (Qwen2AudioConfig, from transformers.models.whisper import WhisperFeatureExtractor from vllm.config import VllmConfig -from vllm.model_executor.sampling_metadata import SamplingMetadata from vllm.multimodal import MULTIMODAL_REGISTRY from vllm.multimodal.inputs import (AudioItem, ModalityData, MultiModalDataDict, MultiModalFieldConfig, @@ -481,10 +480,8 @@ class Qwen2AudioForConditionalGeneration(nn.Module, SupportsMultiModal, def compute_logits( self, hidden_states: torch.Tensor, - sampling_metadata: SamplingMetadata, ) -> Optional[torch.Tensor]: - return self.language_model.compute_logits(hidden_states, - sampling_metadata) + return self.language_model.compute_logits(hidden_states) def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]: diff --git a/vllm/model_executor/models/qwen2_moe.py b/vllm/model_executor/models/qwen2_moe.py index 6c6276a930453..6a9acaf2c3fe0 100644 --- a/vllm/model_executor/models/qwen2_moe.py +++ b/vllm/model_executor/models/qwen2_moe.py @@ -51,7 +51,6 @@ from vllm.model_executor.layers.rotary_embedding import get_rope from vllm.model_executor.layers.vocab_parallel_embedding import ( ParallelLMHead, VocabParallelEmbedding) from vllm.model_executor.model_loader.weight_utils import default_weight_loader -from vllm.model_executor.sampling_metadata import SamplingMetadata from vllm.sequence import IntermediateTensors from .interfaces import SupportsLoRA, SupportsPP @@ -546,10 +545,8 @@ class Qwen2MoeForCausalLM(nn.Module, SupportsPP, SupportsLoRA): def compute_logits( self, hidden_states: torch.Tensor, - sampling_metadata: SamplingMetadata, ) -> Optional[torch.Tensor]: - logits = self.logits_processor(self.lm_head, hidden_states, - sampling_metadata) + logits = self.logits_processor(self.lm_head, hidden_states) return logits def load_weights(self, weights: Iterable[tuple[str, diff --git a/vllm/model_executor/models/qwen2_vl.py b/vllm/model_executor/models/qwen2_vl.py index dd4e7731e0b08..b3c42c2572566 100644 --- a/vllm/model_executor/models/qwen2_vl.py +++ b/vllm/model_executor/models/qwen2_vl.py @@ -46,7 +46,6 @@ from vllm.config import VllmConfig from vllm.distributed import parallel_state, tensor_model_parallel_all_gather from vllm.distributed import utils as dist_utils from vllm.logger import init_logger -from vllm.model_executor import SamplingMetadata from vllm.model_executor.layers.activation import QuickGELU from vllm.model_executor.layers.linear import (ColumnParallelLinear, RowParallelLinear) @@ -1527,10 +1526,8 @@ class Qwen2VLForConditionalGeneration(nn.Module, SupportsMultiModal, def compute_logits( self, hidden_states: torch.Tensor, - sampling_metadata: SamplingMetadata, ) -> Optional[torch.Tensor]: - return self.language_model.compute_logits(hidden_states, - sampling_metadata) + return self.language_model.compute_logits(hidden_states) def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]: diff --git a/vllm/model_executor/models/qwen3.py b/vllm/model_executor/models/qwen3.py index dddb47048a1fc..ae72fd30c3993 100644 --- a/vllm/model_executor/models/qwen3.py +++ b/vllm/model_executor/models/qwen3.py @@ -41,7 +41,6 @@ from vllm.model_executor.layers.logits_processor import LogitsProcessor from vllm.model_executor.layers.quantization import QuantizationConfig from vllm.model_executor.layers.rotary_embedding import get_rope from vllm.model_executor.layers.vocab_parallel_embedding import ParallelLMHead -from vllm.model_executor.sampling_metadata import SamplingMetadata from vllm.sequence import IntermediateTensors from .interfaces import SupportsEagle3, SupportsLoRA, SupportsPP @@ -328,10 +327,8 @@ class Qwen3ForCausalLM(nn.Module, SupportsLoRA, SupportsPP, SupportsEagle3): def compute_logits( self, hidden_states: torch.Tensor, - sampling_metadata: SamplingMetadata, ) -> Optional[torch.Tensor]: - logits = self.logits_processor(self.lm_head, hidden_states, - sampling_metadata) + logits = self.logits_processor(self.lm_head, hidden_states) return logits def load_weights(self, weights: Iterable[tuple[str, diff --git a/vllm/model_executor/models/qwen3_moe.py b/vllm/model_executor/models/qwen3_moe.py index 029309c49efd4..0661b3707ff44 100644 --- a/vllm/model_executor/models/qwen3_moe.py +++ b/vllm/model_executor/models/qwen3_moe.py @@ -54,7 +54,6 @@ from vllm.model_executor.layers.vocab_parallel_embedding import ( ParallelLMHead, VocabParallelEmbedding) from vllm.model_executor.model_loader.weight_utils import ( default_weight_loader, maybe_remap_kv_scale_name) -from vllm.model_executor.sampling_metadata import SamplingMetadata from vllm.sequence import IntermediateTensors from .interfaces import MixtureOfExperts, SupportsLoRA, SupportsPP @@ -690,10 +689,8 @@ class Qwen3MoeForCausalLM(nn.Module, SupportsPP, SupportsLoRA, def compute_logits( self, hidden_states: torch.Tensor, - sampling_metadata: SamplingMetadata, ) -> Optional[torch.Tensor]: - logits = self.logits_processor(self.lm_head, hidden_states, - sampling_metadata) + logits = self.logits_processor(self.lm_head, hidden_states) return logits def load_weights(self, weights: Iterable[tuple[str, diff --git a/vllm/model_executor/models/qwen3_next.py b/vllm/model_executor/models/qwen3_next.py index ce917f92bd2e5..24cebc5bfdd82 100644 --- a/vllm/model_executor/models/qwen3_next.py +++ b/vllm/model_executor/models/qwen3_next.py @@ -53,7 +53,6 @@ from vllm.model_executor.model_loader.weight_utils import ( default_weight_loader, sharded_weight_loader) from vllm.model_executor.models.mamba_cache import MambaCacheParams from vllm.model_executor.models.qwen2_moe import Qwen2MoeMLP as Qwen3NextMLP -from vllm.model_executor.sampling_metadata import SamplingMetadata from vllm.model_executor.utils import set_weight_attrs from vllm.platforms import current_platform from vllm.sequence import IntermediateTensors @@ -1208,10 +1207,8 @@ class Qwen3NextForCausalLM(nn.Module, HasInnerState, SupportsLoRA, SupportsPP, def compute_logits( self, hidden_states: torch.Tensor, - sampling_metadata: SamplingMetadata, ) -> Optional[torch.Tensor]: - return self.logits_processor(self.lm_head, hidden_states, - sampling_metadata) + return self.logits_processor(self.lm_head, hidden_states) def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]: diff --git a/vllm/model_executor/models/qwen3_next_mtp.py b/vllm/model_executor/models/qwen3_next_mtp.py index c755eeb9b4eaa..c054339842e64 100644 --- a/vllm/model_executor/models/qwen3_next_mtp.py +++ b/vllm/model_executor/models/qwen3_next_mtp.py @@ -19,7 +19,6 @@ from vllm.model_executor.layers.vocab_parallel_embedding import ( from vllm.model_executor.model_loader.weight_utils import default_weight_loader from vllm.model_executor.models.qwen3_next import (Qwen3NextDecoderLayer, Qwen3NextRMSNorm) -from vllm.model_executor.sampling_metadata import SamplingMetadata from vllm.sequence import IntermediateTensors from vllm.transformers_utils.configs import Qwen3NextConfig @@ -266,11 +265,9 @@ class Qwen3NextMTP(nn.Module, SupportsPP): def compute_logits( self, hidden_states: torch.Tensor, - sampling_metadata: SamplingMetadata, spec_step_idx: int = 0, ) -> Optional[torch.Tensor]: - return self.logits_processor(self.lm_head, hidden_states, - sampling_metadata) + return self.logits_processor(self.lm_head, hidden_states) def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]: diff --git a/vllm/model_executor/models/qwen3_vl.py b/vllm/model_executor/models/qwen3_vl.py index ca232e03767b1..aa28c07ddcebe 100644 --- a/vllm/model_executor/models/qwen3_vl.py +++ b/vllm/model_executor/models/qwen3_vl.py @@ -45,7 +45,6 @@ from vllm.compilation.decorators import support_torch_compile from vllm.config import VllmConfig from vllm.distributed import get_pp_group from vllm.logger import init_logger -from vllm.model_executor import SamplingMetadata from vllm.model_executor.layers.activation import _ACTIVATION_REGISTRY from vllm.model_executor.layers.linear import (ColumnParallelLinear, RowParallelLinear) @@ -1493,10 +1492,8 @@ class Qwen3VLForConditionalGeneration(nn.Module, SupportsMultiModal, def compute_logits( self, hidden_states: torch.Tensor, - sampling_metadata: SamplingMetadata, ) -> Optional[torch.Tensor]: - return self.language_model.compute_logits(hidden_states, - sampling_metadata) + return self.language_model.compute_logits(hidden_states) def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]: diff --git a/vllm/model_executor/models/seed_oss.py b/vllm/model_executor/models/seed_oss.py index e3c7c700f8fa1..a217c820fedf0 100644 --- a/vllm/model_executor/models/seed_oss.py +++ b/vllm/model_executor/models/seed_oss.py @@ -47,7 +47,6 @@ from vllm.model_executor.layers.vocab_parallel_embedding import ( ParallelLMHead, VocabParallelEmbedding) from vllm.model_executor.model_loader.weight_utils import ( default_weight_loader, maybe_remap_kv_scale_name) -from vllm.model_executor.sampling_metadata import SamplingMetadata from vllm.sequence import IntermediateTensors from .interfaces import SupportsLoRA, SupportsPP @@ -472,10 +471,8 @@ class SeedOssForCausalLM(nn.Module, SupportsLoRA, SupportsPP): def compute_logits( self, hidden_states: torch.Tensor, - sampling_metadata: SamplingMetadata, ) -> Optional[torch.Tensor]: - logits = self.logits_processor(self.lm_head, hidden_states, - sampling_metadata) + logits = self.logits_processor(self.lm_head, hidden_states) return logits def load_weights(self, weights: Iterable[tuple[str, diff --git a/vllm/model_executor/models/skyworkr1v.py b/vllm/model_executor/models/skyworkr1v.py index 9857ccdcbe2d4..893ce4497c319 100644 --- a/vllm/model_executor/models/skyworkr1v.py +++ b/vllm/model_executor/models/skyworkr1v.py @@ -22,7 +22,6 @@ from vllm.model_executor.layers.quantization import QuantizationConfig from vllm.model_executor.layers.quantization.awq import AWQConfig from vllm.model_executor.models.intern_vit import (InternVisionModel, InternVisionPatchModel) -from vllm.model_executor.sampling_metadata import SamplingMetadata from vllm.multimodal import MULTIMODAL_REGISTRY from vllm.multimodal.image import convert_image_mode from vllm.multimodal.inputs import (MultiModalDataDict, MultiModalFieldConfig, @@ -897,10 +896,8 @@ class SkyworkR1VChatModel(nn.Module, SupportsMultiModal, SupportsPP): def compute_logits( self, hidden_states: torch.Tensor, - sampling_metadata: SamplingMetadata, ) -> Optional[torch.Tensor]: - return self.language_model.compute_logits(hidden_states, - sampling_metadata) + return self.language_model.compute_logits(hidden_states) def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]: diff --git a/vllm/model_executor/models/solar.py b/vllm/model_executor/models/solar.py index 94c862258b7ad..c774171b9dcd2 100644 --- a/vllm/model_executor/models/solar.py +++ b/vllm/model_executor/models/solar.py @@ -47,7 +47,6 @@ from vllm.model_executor.layers.vocab_parallel_embedding import ( DEFAULT_VOCAB_PADDING_SIZE, ParallelLMHead, VocabParallelEmbedding) from vllm.model_executor.model_loader.weight_utils import ( default_weight_loader, maybe_remap_kv_scale_name) -from vllm.model_executor.sampling_metadata import SamplingMetadata from vllm.sequence import IntermediateTensors from .interfaces import SupportsLoRA, SupportsPP @@ -495,10 +494,8 @@ class SolarForCausalLM(nn.Module, SupportsLoRA, SupportsPP): inputs_embeds) return model_output - def compute_logits(self, hidden_states: torch.Tensor, - sampling_metadata: SamplingMetadata) -> torch.Tensor: - logits = self.logits_processor(self.lm_head, hidden_states, - sampling_metadata) + def compute_logits(self, hidden_states: torch.Tensor) -> torch.Tensor: + logits = self.logits_processor(self.lm_head, hidden_states) return logits def load_weights(self, weights: Iterable[tuple[str, diff --git a/vllm/model_executor/models/stablelm.py b/vllm/model_executor/models/stablelm.py index 9e880ebd50813..e4dfe8d5a9a3b 100644 --- a/vllm/model_executor/models/stablelm.py +++ b/vllm/model_executor/models/stablelm.py @@ -42,7 +42,6 @@ from vllm.model_executor.layers.rotary_embedding import get_rope from vllm.model_executor.layers.vocab_parallel_embedding import ( ParallelLMHead, VocabParallelEmbedding) from vllm.model_executor.model_loader.weight_utils import default_weight_loader -from vllm.model_executor.sampling_metadata import SamplingMetadata from vllm.sequence import IntermediateTensors from .interfaces import SupportsPP @@ -332,10 +331,8 @@ class StablelmForCausalLM(nn.Module, SupportsPP): def compute_logits( self, hidden_states: torch.Tensor, - sampling_metadata: SamplingMetadata, ) -> Optional[torch.Tensor]: - logits = self.logits_processor(self.lm_head, hidden_states, - sampling_metadata) + logits = self.logits_processor(self.lm_head, hidden_states) return logits def load_weights(self, weights: Iterable[tuple[str, diff --git a/vllm/model_executor/models/starcoder2.py b/vllm/model_executor/models/starcoder2.py index 62ff9b6182755..7f379ab95a03d 100644 --- a/vllm/model_executor/models/starcoder2.py +++ b/vllm/model_executor/models/starcoder2.py @@ -43,7 +43,6 @@ from vllm.model_executor.layers.vocab_parallel_embedding import ( DEFAULT_VOCAB_PADDING_SIZE, ParallelLMHead, VocabParallelEmbedding) from vllm.model_executor.model_loader.weight_utils import ( default_weight_loader, maybe_remap_kv_scale_name) -from vllm.model_executor.sampling_metadata import SamplingMetadata from vllm.sequence import IntermediateTensors from .interfaces import SupportsPP @@ -339,10 +338,8 @@ class Starcoder2ForCausalLM(nn.Module, SupportsPP): def compute_logits( self, hidden_states: torch.Tensor, - sampling_metadata: SamplingMetadata, ) -> Optional[torch.Tensor]: - logits = self.logits_processor(self.lm_head, hidden_states, - sampling_metadata) + logits = self.logits_processor(self.lm_head, hidden_states) return logits def load_weights(self, weights: Iterable[tuple[str, diff --git a/vllm/model_executor/models/step3_text.py b/vllm/model_executor/models/step3_text.py index 6a5b540fc8174..0cce0c78f8dc6 100644 --- a/vllm/model_executor/models/step3_text.py +++ b/vllm/model_executor/models/step3_text.py @@ -29,7 +29,6 @@ from vllm.model_executor.layers.rotary_embedding import get_rope from vllm.model_executor.layers.vocab_parallel_embedding import ( DEFAULT_VOCAB_PADDING_SIZE, ParallelLMHead, VocabParallelEmbedding) from vllm.model_executor.model_loader.weight_utils import default_weight_loader -from vllm.model_executor.sampling_metadata import SamplingMetadata from vllm.sequence import IntermediateTensors from .interfaces import SupportsPP @@ -405,10 +404,8 @@ class Step3TextForCausalLM(nn.Module, SupportsPP): inputs_embeds) return hidden_states - def compute_logits(self, hidden_states: torch.Tensor, - sampling_metadata: SamplingMetadata) -> torch.Tensor: - logits = self.logits_processor(self.lm_head, hidden_states, - sampling_metadata) + def compute_logits(self, hidden_states: torch.Tensor) -> torch.Tensor: + logits = self.logits_processor(self.lm_head, hidden_states) return logits def load_weights(self, weights: Iterable[tuple[str, diff --git a/vllm/model_executor/models/step3_vl.py b/vllm/model_executor/models/step3_vl.py index c2940f8e44450..f667266b77bfa 100644 --- a/vllm/model_executor/models/step3_vl.py +++ b/vllm/model_executor/models/step3_vl.py @@ -23,7 +23,6 @@ from vllm.model_executor.layers.linear import (ColumnParallelLinear, QKVParallelLinear, RowParallelLinear) from vllm.model_executor.layers.quantization import QuantizationConfig -from vllm.model_executor.sampling_metadata import SamplingMetadata from vllm.multimodal import MULTIMODAL_REGISTRY from vllm.multimodal.inputs import (MultiModalDataDict, MultiModalFieldConfig, MultiModalKwargsItems, NestedTensors) @@ -1055,10 +1054,8 @@ class Step3VLForConditionalGeneration(nn.Module, SupportsMultiModal, def compute_logits( self, hidden_states: torch.Tensor, - sampling_metadata: SamplingMetadata, ) -> Optional[torch.Tensor]: - return self.language_model.compute_logits(hidden_states, - sampling_metadata) + return self.language_model.compute_logits(hidden_states) def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]): diff --git a/vllm/model_executor/models/tarsier.py b/vllm/model_executor/models/tarsier.py index c66867315e553..67cf3ccf315d1 100644 --- a/vllm/model_executor/models/tarsier.py +++ b/vllm/model_executor/models/tarsier.py @@ -23,7 +23,6 @@ from vllm.model_executor.layers.linear import (ColumnParallelLinear, RowParallelLinear) from vllm.model_executor.layers.quantization import QuantizationConfig from vllm.model_executor.models.llava import LlavaDummyInputsBuilder -from vllm.model_executor.sampling_metadata import SamplingMetadata from vllm.multimodal import MULTIMODAL_REGISTRY from vllm.multimodal.cache import BaseMultiModalProcessorCache from vllm.multimodal.inputs import MultiModalFieldConfig, MultiModalKwargsItems @@ -638,10 +637,8 @@ class TarsierForConditionalGeneration(nn.Module, SupportsMultiModal, def compute_logits( self, hidden_states: torch.Tensor, - sampling_metadata: SamplingMetadata, ) -> Optional[torch.Tensor]: - return self.language_model.compute_logits(hidden_states, - sampling_metadata) + return self.language_model.compute_logits(hidden_states) def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]: diff --git a/vllm/model_executor/models/transformers.py b/vllm/model_executor/models/transformers.py index 3bd4d10316ec6..475a68bc642b9 100644 --- a/vllm/model_executor/models/transformers.py +++ b/vllm/model_executor/models/transformers.py @@ -41,7 +41,6 @@ from vllm.model_executor.layers.logits_processor import LogitsProcessor from vllm.model_executor.layers.quantization import QuantizationConfig from vllm.model_executor.layers.vocab_parallel_embedding import ( ParallelLMHead, VocabParallelEmbedding) -from vllm.model_executor.sampling_metadata import SamplingMetadata from vllm.multimodal import MULTIMODAL_REGISTRY, MultiModalKwargsItems from vllm.multimodal.inputs import (MultiModalDataDict, MultiModalFieldConfig, MultiModalInputs, MultiModalUUIDDict, @@ -798,10 +797,8 @@ class TransformersForCausalLM(TransformersBase): def compute_logits( self, hidden_states: torch.Tensor, - sampling_metadata: SamplingMetadata, ) -> Optional[torch.Tensor]: - logits = self.logits_processor(self.lm_head, hidden_states, - sampling_metadata) + logits = self.logits_processor(self.lm_head, hidden_states) return logits diff --git a/vllm/model_executor/models/ultravox.py b/vllm/model_executor/models/ultravox.py index f1f11c5fe8f00..12ae9487ad9dc 100644 --- a/vllm/model_executor/models/ultravox.py +++ b/vllm/model_executor/models/ultravox.py @@ -18,7 +18,6 @@ from vllm.model_executor.layers.activation import MulAndSilu, get_act_fn from vllm.model_executor.layers.layernorm import RMSNorm from vllm.model_executor.model_loader import DefaultModelLoader from vllm.model_executor.models.module_mapping import MultiModelKeys -from vllm.model_executor.sampling_metadata import SamplingMetadata from vllm.multimodal import MULTIMODAL_REGISTRY from vllm.multimodal.inputs import (MultiModalDataDict, MultiModalFieldConfig, MultiModalKwargsItems, NestedTensors) @@ -616,10 +615,8 @@ class UltravoxModel(nn.Module, SupportsMultiModal, SupportsPP, SupportsLoRA): inputs_embeds=inputs_embeds) return hidden_states - def compute_logits(self, hidden_states: torch.Tensor, - sampling_metadata: SamplingMetadata) -> torch.Tensor: - return self.language_model.compute_logits(hidden_states, - sampling_metadata) + def compute_logits(self, hidden_states: torch.Tensor) -> torch.Tensor: + return self.language_model.compute_logits(hidden_states) def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]: diff --git a/vllm/model_executor/models/voxtral.py b/vllm/model_executor/models/voxtral.py index 16a97389cd21b..b33e8d09c4be1 100644 --- a/vllm/model_executor/models/voxtral.py +++ b/vllm/model_executor/models/voxtral.py @@ -30,7 +30,6 @@ from vllm.model_executor.models.module_mapping import MultiModelKeys # yapf: disable from vllm.model_executor.models.whisper import WhisperEncoder # yapf: enable -from vllm.model_executor.sampling_metadata import SamplingMetadata from vllm.multimodal import MULTIMODAL_REGISTRY from vllm.multimodal.inputs import (MultiModalDataDict, MultiModalFieldConfig, MultiModalKwargsItems, MultiModalUUIDDict, @@ -454,10 +453,8 @@ class VoxtralForConditionalGeneration(nn.Module, SupportsMultiModal, def compute_logits( self, hidden_states: torch.Tensor, - sampling_metadata: SamplingMetadata, ) -> Optional[torch.Tensor]: - return self.language_model.compute_logits(hidden_states, - sampling_metadata) + return self.language_model.compute_logits(hidden_states) @classmethod def get_speech_to_text_config(cls, model_config: ModelConfig, diff --git a/vllm/model_executor/models/whisper.py b/vllm/model_executor/models/whisper.py index 41ae7b129782d..de3e4f0592a62 100644 --- a/vllm/model_executor/models/whisper.py +++ b/vllm/model_executor/models/whisper.py @@ -31,7 +31,6 @@ from vllm.model_executor.layers.quantization.base_config import ( from vllm.model_executor.layers.vocab_parallel_embedding import ParallelLMHead from vllm.model_executor.model_loader.utils import set_default_torch_dtype from vllm.model_executor.model_loader.weight_utils import default_weight_loader -from vllm.model_executor.sampling_metadata import SamplingMetadata from vllm.multimodal import MULTIMODAL_REGISTRY, NestedTensors from vllm.multimodal.inputs import (MultiModalDataDict, MultiModalFieldConfig, MultiModalKwargsItems) @@ -936,10 +935,8 @@ class WhisperForConditionalGeneration(nn.Module, SupportsTranscription, return WhisperAudioInputs(input_features=input_features) - def compute_logits(self, hidden_states: torch.Tensor, - sampling_metadata: SamplingMetadata) -> torch.Tensor: - logits = self.logits_processor(self.proj_out, hidden_states, - sampling_metadata) + def compute_logits(self, hidden_states: torch.Tensor) -> torch.Tensor: + logits = self.logits_processor(self.proj_out, hidden_states) return logits def load_weights(self, weights: Iterable[tuple[str, diff --git a/vllm/model_executor/models/zamba2.py b/vllm/model_executor/models/zamba2.py index e601bc3adb6e9..4350e38e02f96 100644 --- a/vllm/model_executor/models/zamba2.py +++ b/vllm/model_executor/models/zamba2.py @@ -41,7 +41,6 @@ from vllm.model_executor.layers.vocab_parallel_embedding import ( from vllm.model_executor.model_loader.weight_utils import default_weight_loader from vllm.model_executor.models.mamba_cache import (MambaCacheManager, MambaCacheParams) -from vllm.model_executor.sampling_metadata import SamplingMetadata from vllm.sequence import IntermediateTensors from .interfaces import HasInnerState, IsHybrid @@ -1036,7 +1035,6 @@ class Zamba2ForCausalLM(nn.Module, HasInnerState, IsHybrid): def compute_logits( self, hidden_states: torch.Tensor, - sampling_metadata: SamplingMetadata, ) -> Optional[torch.Tensor]: """Compute logits for next token prediction. @@ -1047,8 +1045,7 @@ class Zamba2ForCausalLM(nn.Module, HasInnerState, IsHybrid): Returns: Logits for next token prediction """ - logits = self.logits_processor(self.lm_head, hidden_states, - sampling_metadata) + logits = self.logits_processor(self.lm_head, hidden_states) return logits def load_weights(self, weights: Iterable[tuple[str, diff --git a/vllm/model_executor/sampling_metadata.py b/vllm/model_executor/sampling_metadata.py deleted file mode 100644 index 8c4548ff7f7dc..0000000000000 --- a/vllm/model_executor/sampling_metadata.py +++ /dev/null @@ -1,7 +0,0 @@ -# SPDX-License-Identifier: Apache-2.0 -# SPDX-FileCopyrightText: Copyright contributors to the vLLM project - - -class SamplingMetadata: - # Placeholder until it can be safely removed. - pass diff --git a/vllm/v1/spec_decode/eagle.py b/vllm/v1/spec_decode/eagle.py index 2a178ddf48777..5dacf60886966 100644 --- a/vllm/v1/spec_decode/eagle.py +++ b/vllm/v1/spec_decode/eagle.py @@ -239,7 +239,7 @@ class EagleProposer: else: last_hidden_states, hidden_states = ret_hidden_states sample_hidden_states = last_hidden_states[last_token_indices] - logits = self.model.compute_logits(sample_hidden_states, None) + logits = self.model.compute_logits(sample_hidden_states) # Early exit if there is only one draft token to be generated. if self.num_speculative_tokens == 1: @@ -367,8 +367,7 @@ class EagleProposer: else: last_hidden_states, hidden_states = ret_hidden_states hidden_states = hidden_states[:batch_size] - logits = self.model.compute_logits(last_hidden_states[:batch_size], - None) + logits = self.model.compute_logits(last_hidden_states[:batch_size]) draft_token_ids = logits.argmax(dim=-1) draft_token_ids_list.append(draft_token_ids) @@ -678,9 +677,7 @@ class EagleProposer: # Get the output logits for the draft tokens. logits = self.model.compute_logits( draft_last_hidden_states.reshape(batch_size * level_num_drafts, - -1), - None, - ) + -1)) # Sample a draft token for each child at the next tree level. num_children = self.child_drafts_per_level[level + 1] diff --git a/vllm/v1/spec_decode/medusa.py b/vllm/v1/spec_decode/medusa.py index 3e90179e78d99..70b29c05c2a50 100644 --- a/vllm/v1/spec_decode/medusa.py +++ b/vllm/v1/spec_decode/medusa.py @@ -41,7 +41,7 @@ class MedusaProposer: ) -> list[list[int]]: # Generate blocks and compute logits blocks = self.model(target_hidden_states) - logits = self.model.compute_logits(blocks, None) + logits = self.model.compute_logits(blocks) # Get draft tokens and transpose the result # TODO(woosuk): OPTIMIZATION: Return GPU tensor without GPU-CPU diff --git a/vllm/v1/worker/gpu_model_runner.py b/vllm/v1/worker/gpu_model_runner.py index d0946e8c5d7d8..b0cd0f4133079 100644 --- a/vllm/v1/worker/gpu_model_runner.py +++ b/vllm/v1/worker/gpu_model_runner.py @@ -2240,7 +2240,7 @@ class GPUModelRunner(LoRAModelRunnerMixin, KVConnectorModelRunnerMixin): return output sample_hidden_states = hidden_states[logits_indices] - logits = self.model.compute_logits(sample_hidden_states, None) + logits = self.model.compute_logits(sample_hidden_states) else: # Rare case. assert not self.is_pooling_model @@ -2258,8 +2258,7 @@ class GPUModelRunner(LoRAModelRunnerMixin, KVConnectorModelRunnerMixin): logits = None else: sample_hidden_states = hidden_states[logits_indices] - logits = self.model.compute_logits(sample_hidden_states, - None) + logits = self.model.compute_logits(sample_hidden_states) model_output_broadcast_data = {} if logits is not None: @@ -2706,7 +2705,7 @@ class GPUModelRunner(LoRAModelRunnerMixin, KVConnectorModelRunnerMixin): 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, None) + 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 @@ -3105,7 +3104,7 @@ class GPUModelRunner(LoRAModelRunnerMixin, KVConnectorModelRunnerMixin): # 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, None) + logits = self.model.compute_logits(hidden_states) num_reqs = logits.size(0) dummy_tensors = lambda v: torch.full( diff --git a/vllm/v1/worker/tpu_model_runner.py b/vllm/v1/worker/tpu_model_runner.py index 48070c1e3e7cb..dd11b1dcbe94c 100644 --- a/vllm/v1/worker/tpu_model_runner.py +++ b/vllm/v1/worker/tpu_model_runner.py @@ -1692,7 +1692,7 @@ class TPUModelRunner(LoRAModelRunnerMixin, KVConnectorModelRunnerMixin): @torch.compile(backend="openxla", fullgraph=True, dynamic=False) def compute_logits(self, sample_hidden_states: torch.Tensor) -> torch.Tensor: - return self.model.compute_logits(sample_hidden_states, None) + return self.model.compute_logits(sample_hidden_states) # TODO: Under SPMD mode, sample_from_logits has correctness issue. # Re-enable the torch.compile once the issue is fixed in torchxla.