Merge tag 'v0.8.5' into v0.8.5-dev

vllm/model_executor/models/nemotron_nas.py

@@ -34,7 +34,6 @@ from vllm.distributed import get_pp_group
 from vllm.model_executor.layers.layernorm import RMSNorm
 from vllm.model_executor.layers.logits_processor import LogitsProcessor
 from vllm.model_executor.layers.quantization import QuantizationConfig
-from vllm.model_executor.layers.sampler import SamplerOutput, get_sampler
 from vllm.model_executor.layers.vocab_parallel_embedding import (
     DEFAULT_VOCAB_PADDING_SIZE, ParallelLMHead, VocabParallelEmbedding)
 from vllm.model_executor.model_loader.weight_utils import (
@@ -408,8 +407,6 @@ class DeciLMForCausalLM(nn.Module, SupportsLoRA, SupportsPP, HasNoOps):
         else:
             self.lm_head = PPMissingLayer()
 
-        self.sampler = get_sampler()
-
         self.make_empty_intermediate_tensors = (
             self.model.make_empty_intermediate_tensors)
 
@@ -439,11 +436,6 @@ class DeciLMForCausalLM(nn.Module, SupportsLoRA, SupportsPP, HasNoOps):
                                        sampling_metadata)
         return logits
 
-    def sample(self, logits: torch.Tensor,
-               sampling_metadata: SamplingMetadata) -> Optional[SamplerOutput]:
-        next_tokens = self.sampler(logits, sampling_metadata)
-        return next_tokens
-
     def load_weights(self, weights: Iterable[Tuple[str,
                                                    torch.Tensor]]) -> Set[str]:
         loader = AutoWeightsLoader(

vllm/model_executor/models/olmo.py

@@ -39,7 +39,6 @@ from vllm.model_executor.layers.linear import (MergedColumnParallelLinear,
 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.sampler import SamplerOutput, get_sampler
 from vllm.model_executor.layers.vocab_parallel_embedding import (
     ParallelLMHead, VocabParallelEmbedding)
 from vllm.model_executor.model_loader.weight_utils import default_weight_loader
@@ -309,7 +308,6 @@ class OlmoForCausalLM(nn.Module, SupportsPP):
                 quant_config=quant_config,
             )
         self.logits_processor = LogitsProcessor(config.vocab_size)
-        self.sampler = get_sampler()
         self.make_empty_intermediate_tensors = (
             self.model.make_empty_intermediate_tensors)
 
@@ -340,14 +338,6 @@ class OlmoForCausalLM(nn.Module, SupportsPP):
                                        sampling_metadata)
         return logits
 
-    def sample(
-        self,
-        logits: torch.Tensor,
-        sampling_metadata: SamplingMetadata,
-    ) -> Optional[SamplerOutput]:
-        next_tokens = self.sampler(logits, sampling_metadata)
-        return next_tokens
-
     def load_weights(self, weights: Iterable[Tuple[str,
                                                    torch.Tensor]]) -> Set[str]:
         stacked_params_mapping = [

vllm/model_executor/models/olmo2.py

@@ -28,6 +28,7 @@ from typing import Iterable, Optional, Tuple, Union
 
 import torch
 from torch import nn
+from transformers import Olmo2Config
 
 from vllm.attention import Attention
 from vllm.config import VllmConfig
@@ -42,7 +43,6 @@ from vllm.model_executor.layers.linear import (MergedColumnParallelLinear,
                                                RowParallelLinear)
 from vllm.model_executor.layers.logits_processor import LogitsProcessor
 from vllm.model_executor.layers.rotary_embedding import get_rope
-from vllm.model_executor.layers.sampler import SamplerOutput, get_sampler
 from vllm.model_executor.layers.vocab_parallel_embedding import (
     ParallelLMHead, VocabParallelEmbedding)
 from vllm.model_executor.model_loader.weight_utils import default_weight_loader
@@ -52,7 +52,6 @@ from vllm.model_executor.models.utils import (
     make_layers, maybe_prefix)
 from vllm.model_executor.sampling_metadata import SamplingMetadata
 from vllm.sequence import IntermediateTensors
-from vllm.transformers_utils.configs.olmo2 import Olmo2Config
 
 
 class Olmo2Attention(nn.Module):
@@ -339,7 +338,6 @@ class Olmo2ForCausalLM(nn.Module, SupportsPP):
                 prefix=maybe_prefix(prefix, "lm_head"),
             )
         self.logits_processor = LogitsProcessor(config.vocab_size)
-        self.sampler = get_sampler()
         self.make_empty_intermediate_tensors = (
             self.model.make_empty_intermediate_tensors)
 
@@ -367,14 +365,6 @@ class Olmo2ForCausalLM(nn.Module, SupportsPP):
                                        sampling_metadata)
         return logits
 
-    def sample(
-        self,
-        logits: torch.Tensor,
-        sampling_metadata: SamplingMetadata,
-    ) -> Optional[SamplerOutput]:
-        next_tokens = self.sampler(logits, sampling_metadata)
-        return next_tokens
-
     def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
         stacked_params_mapping = [
             # (param_name, shard_name, shard_id)

vllm/model_executor/models/olmoe.py

@@ -31,7 +31,6 @@ from vllm.model_executor.layers.linear import (QKVParallelLinear,
 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.sampler import SamplerOutput, get_sampler
 from vllm.model_executor.layers.vocab_parallel_embedding import (
     ParallelLMHead, VocabParallelEmbedding)
 from vllm.model_executor.model_loader.weight_utils import default_weight_loader
@@ -39,7 +38,7 @@ from vllm.model_executor.sampling_metadata import SamplingMetadata
 from vllm.sequence import IntermediateTensors
 
 from .interfaces import SupportsPP
-from .utils import (is_pp_missing_parameter,
+from .utils import (AutoWeightsLoader, is_pp_missing_parameter,
                     make_empty_intermediate_tensors_factory, make_layers,
                     maybe_prefix)
 
@@ -255,7 +254,7 @@ class OlmoeModel(nn.Module):
         quant_config = vllm_config.quant_config
 
         self.vocab_size = config.vocab_size
-
+        self.config = config
         self.embed_tokens = VocabParallelEmbedding(
             config.vocab_size,
             config.hidden_size,
@@ -308,56 +307,6 @@ class OlmoeModel(nn.Module):
         hidden_states, _ = self.norm(hidden_states, residual)
         return hidden_states
 
-
-class OlmoeForCausalLM(nn.Module, SupportsPP):
-
-    fall_back_to_pt_during_load = False
-
-    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
-        super().__init__()
-        config = vllm_config.model_config.hf_config
-        quant_config = vllm_config.quant_config
-        self.config = config
-        self.quant_config = quant_config
-        self.model = OlmoeModel(vllm_config=vllm_config,
-                                prefix=maybe_prefix(prefix, "model"))
-        self.lm_head = ParallelLMHead(config.vocab_size,
-                                      config.hidden_size,
-                                      quant_config=quant_config)
-        self.logits_processor = LogitsProcessor(config.vocab_size)
-        self.sampler = get_sampler()
-
-        self.make_empty_intermediate_tensors = (
-            self.model.make_empty_intermediate_tensors)
-
-    def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
-        return self.model.get_input_embeddings(input_ids)
-
-    def forward(
-        self,
-        input_ids: torch.Tensor,
-        positions: torch.Tensor,
-        intermediate_tensors: Optional[IntermediateTensors] = None,
-        inputs_embeds: Optional[torch.Tensor] = None,
-    ) -> Union[torch.Tensor, IntermediateTensors]:
-        hidden_states = self.model(input_ids, positions, intermediate_tensors,
-                                   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)
-        return logits
-
-    def sample(
-        self,
-        logits: Optional[torch.Tensor],
-        sampling_metadata: SamplingMetadata,
-    ) -> Optional[SamplerOutput]:
-        next_tokens = self.sampler(logits, sampling_metadata)
-        return next_tokens
-
     def load_weights(self, weights: Iterable[Tuple[str,
                                                    torch.Tensor]]) -> Set[str]:
         stacked_params_mapping = [
@@ -380,8 +329,6 @@ class OlmoeForCausalLM(nn.Module, SupportsPP):
         params_dict = dict(self.named_parameters())
         loaded_params: Set[str] = set()
         for name, loaded_weight in weights:
-            if "rotary_emb.inv_freq" in name:
-                continue
             for (param_name, weight_name, shard_id) in stacked_params_mapping:
                 # Skip non-stacked layers and experts (experts handled below).
                 if weight_name not in name:
@@ -453,3 +400,50 @@ class OlmoeForCausalLM(nn.Module, SupportsPP):
                     weight_loader(param, loaded_weight)
             loaded_params.add(name)
         return loaded_params
+
+
+class OlmoeForCausalLM(nn.Module, SupportsPP):
+
+    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
+        super().__init__()
+        config = vllm_config.model_config.hf_config
+        quant_config = vllm_config.quant_config
+        self.config = config
+        self.quant_config = quant_config
+        self.model = OlmoeModel(vllm_config=vllm_config,
+                                prefix=maybe_prefix(prefix, "model"))
+        self.lm_head = ParallelLMHead(config.vocab_size,
+                                      config.hidden_size,
+                                      quant_config=quant_config)
+        self.logits_processor = LogitsProcessor(config.vocab_size)
+
+        self.make_empty_intermediate_tensors = (
+            self.model.make_empty_intermediate_tensors)
+
+    def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
+        return self.model.get_input_embeddings(input_ids)
+
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        positions: torch.Tensor,
+        intermediate_tensors: Optional[IntermediateTensors] = None,
+        inputs_embeds: Optional[torch.Tensor] = None,
+    ) -> Union[torch.Tensor, IntermediateTensors]:
+        hidden_states = self.model(input_ids, positions, intermediate_tensors,
+                                   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)
+        return logits
+
+    def load_weights(self, weights: Iterable[Tuple[str,
+                                                   torch.Tensor]]) -> Set[str]:
+        loader = AutoWeightsLoader(
+            self,
+            skip_prefixes=["rotary_emb.inv_freq"],
+        )
+        return loader.load_weights(weights)

vllm/model_executor/models/opt.py

@@ -35,7 +35,6 @@ from vllm.model_executor.layers.linear import (ColumnParallelLinear,
                                                RowParallelLinear)
 from vllm.model_executor.layers.logits_processor import LogitsProcessor
 from vllm.model_executor.layers.quantization import QuantizationConfig
-from vllm.model_executor.layers.sampler import SamplerOutput, get_sampler
 from vllm.model_executor.layers.vocab_parallel_embedding import (
     ParallelLMHead, VocabParallelEmbedding)
 from vllm.model_executor.model_loader.weight_utils import default_weight_loader
@@ -43,7 +42,7 @@ from vllm.model_executor.sampling_metadata import SamplingMetadata
 from vllm.sequence import IntermediateTensors
 
 from .interfaces import SupportsPP
-from .utils import (is_pp_missing_parameter,
+from .utils import (AutoWeightsLoader, WeightsMapper, is_pp_missing_parameter,
                     make_empty_intermediate_tensors_factory, make_layers,
                     maybe_prefix)
 
@@ -313,6 +312,43 @@ class OPTModel(nn.Module):
                             intermediate_tensors,
                             inputs_embeds=inputs_embeds)
 
+    def load_weights(self, weights: Iterable[Tuple[str,
+                                                   torch.Tensor]]) -> Set[str]:
+        stacked_params_mapping = [
+            # (param_name, shard_name, shard_id)
+            ("qkv_proj", "q_proj", "q"),
+            ("qkv_proj", "k_proj", "k"),
+            ("qkv_proj", "v_proj", "v"),
+        ]
+        params_dict = dict(self.named_parameters(remove_duplicate=False))
+        loaded_params: Set[str] = set()
+        for name, loaded_weight in weights:
+            for (param_name, weight_name, shard_id) in stacked_params_mapping:
+                if weight_name not in name:
+                    continue
+                name = name.replace(weight_name, param_name)
+                # Skip loading extra bias for GPTQ models.
+                if name.endswith(".bias") and name not in params_dict:
+                    continue
+                if is_pp_missing_parameter(name, self):
+                    continue
+                param = params_dict[name]
+                weight_loader = param.weight_loader
+                weight_loader(param, loaded_weight, shard_id)
+                break
+            else:
+                # Skip loading extra bias for GPTQ models.
+                if name.endswith(".bias") and name not in params_dict:
+                    continue
+                if is_pp_missing_parameter(name, self):
+                    continue
+                param = params_dict[name]
+                weight_loader = getattr(param, "weight_loader",
+                                        default_weight_loader)
+                weight_loader(param, loaded_weight)
+            loaded_params.add(name)
+        return loaded_params
+
 
 class OPTForCausalLM(nn.Module, SupportsPP):
     packed_modules_mapping = {
@@ -320,6 +356,10 @@ class OPTForCausalLM(nn.Module, SupportsPP):
         "gate_up_proj": ["gate_proj", "up_proj"]
     }
 
+    hf_to_vllm_mapper = WeightsMapper(orig_to_new_prefix={
+        "decoder.": "model.decoder.",
+    })
+
     def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
         super().__init__()
         config = vllm_config.model_config.hf_config
@@ -334,7 +374,6 @@ class OPTForCausalLM(nn.Module, SupportsPP):
             self.lm_head = ParallelLMHead(config.vocab_size,
                                           config.word_embed_proj_dim)
         self.logits_processor = LogitsProcessor(config.vocab_size)
-        self.sampler = get_sampler()
         self.make_empty_intermediate_tensors = (
             self.model.make_empty_intermediate_tensors)
 
@@ -361,52 +400,11 @@ class OPTForCausalLM(nn.Module, SupportsPP):
                                        sampling_metadata)
         return logits
 
-    def sample(
-        self,
-        logits: torch.Tensor,
-        sampling_metadata: SamplingMetadata,
-    ) -> Optional[SamplerOutput]:
-        next_tokens = self.sampler(logits, sampling_metadata)
-        return next_tokens
-
     def load_weights(self, weights: Iterable[Tuple[str,
                                                    torch.Tensor]]) -> Set[str]:
-        stacked_params_mapping = [
-            # (param_name, shard_name, shard_id)
-            ("qkv_proj", "q_proj", "q"),
-            ("qkv_proj", "k_proj", "k"),
-            ("qkv_proj", "v_proj", "v"),
-        ]
-        params_dict = dict(self.named_parameters(remove_duplicate=False))
-        loaded_params: Set[str] = set()
-        for name, loaded_weight in weights:
-            if "lm_head.weight" in name and self.config.tie_word_embeddings:
-                continue
-            if name.startswith("decoder."):
-                name = "model." + name
-
-            for (param_name, weight_name, shard_id) in stacked_params_mapping:
-                if weight_name not in name:
-                    continue
-                name = name.replace(weight_name, param_name)
-                # Skip loading extra bias for GPTQ models.
-                if name.endswith(".bias") and name not in params_dict:
-                    continue
-                if is_pp_missing_parameter(name, self):
-                    continue
-                param = params_dict[name]
-                weight_loader = param.weight_loader
-                weight_loader(param, loaded_weight, shard_id)
-                break
-            else:
-                # Skip loading extra bias for GPTQ models.
-                if name.endswith(".bias") and name not in params_dict:
-                    continue
-                if is_pp_missing_parameter(name, self):
-                    continue
-                param = params_dict[name]
-                weight_loader = getattr(param, "weight_loader",
-                                        default_weight_loader)
-                weight_loader(param, loaded_weight)
-            loaded_params.add(name)
-        return loaded_params
+        loader = AutoWeightsLoader(
+            self,
+            skip_prefixes=(["lm_head.weight"]
+                           if self.config.tie_word_embeddings else None),
+        )
+        return loader.load_weights(weights, mapper=self.hf_to_vllm_mapper)

vllm/model_executor/models/orion.py

@@ -22,7 +22,6 @@ from vllm.model_executor.layers.linear import (MergedColumnParallelLinear,
 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.sampler import SamplerOutput, get_sampler
 from vllm.model_executor.layers.vocab_parallel_embedding import (
     ParallelLMHead, VocabParallelEmbedding)
 from vllm.model_executor.model_loader.weight_utils import default_weight_loader
@@ -30,7 +29,7 @@ from vllm.model_executor.sampling_metadata import SamplingMetadata
 from vllm.sequence import IntermediateTensors
 
 from .interfaces import SupportsPP
-from .utils import (is_pp_missing_parameter,
+from .utils import (AutoWeightsLoader, is_pp_missing_parameter,
                     make_empty_intermediate_tensors_factory, make_layers,
                     maybe_prefix)
 
@@ -260,6 +259,45 @@ class OrionModel(nn.Module):
         hidden_states = self.norm(hidden_states)
         return hidden_states
 
+    def load_weights(self, weights: Iterable[Tuple[str,
+                                                   torch.Tensor]]) -> Set[str]:
+        stacked_params_mapping = [
+            # (param_name, shard_name, shard_id)
+            ("qkv_proj", "q_proj", "q"),
+            ("qkv_proj", "k_proj", "k"),
+            ("qkv_proj", "v_proj", "v"),
+            ("gate_up_proj", "gate_proj", 0),
+            ("gate_up_proj", "up_proj", 1),
+        ]
+        params_dict = dict(self.named_parameters())
+        loaded_params: Set[str] = set()
+        for name, loaded_weight in weights:
+            for (param_name, weight_name, shard_id) in stacked_params_mapping:
+                if weight_name not in name:
+                    continue
+                name = name.replace(weight_name, param_name)
+                # Skip loading extra bias for GPTQ models.
+                if name.endswith(".bias") and name not in params_dict:
+                    continue
+                if is_pp_missing_parameter(name, self):
+                    continue
+                param = params_dict[name]
+                weight_loader = param.weight_loader
+                weight_loader(param, loaded_weight, shard_id)
+                break
+            else:
+                # Skip loading extra bias for GPTQ models.
+                if name.endswith(".bias") and name not in params_dict:
+                    continue
+                if is_pp_missing_parameter(name, self):
+                    continue
+                param = params_dict[name]
+                weight_loader = getattr(param, "weight_loader",
+                                        default_weight_loader)
+                weight_loader(param, loaded_weight)
+            loaded_params.add(name)
+        return loaded_params
+
 
 class OrionForCausalLM(nn.Module, SupportsPP):
 
@@ -277,7 +315,6 @@ class OrionForCausalLM(nn.Module, SupportsPP):
         if self.config.tie_word_embeddings:
             self.lm_head.weight = self.model.embed_tokens.weight
         self.logits_processor = LogitsProcessor(config.vocab_size)
-        self.sampler = get_sampler()
         self.make_empty_intermediate_tensors = (
             self.model.make_empty_intermediate_tensors)
 
@@ -304,56 +341,16 @@ class OrionForCausalLM(nn.Module, SupportsPP):
                                        sampling_metadata)
         return logits
 
-    def sample(
-        self,
-        logits: torch.Tensor,
-        sampling_metadata: SamplingMetadata,
-    ) -> Optional[SamplerOutput]:
-        next_tokens = self.sampler(logits, sampling_metadata)
-        return next_tokens
-
     def load_weights(self, weights: Iterable[Tuple[str,
                                                    torch.Tensor]]) -> Set[str]:
-        stacked_params_mapping = [
-            # (param_name, shard_name, shard_id)
-            ("qkv_proj", "q_proj", "q"),
-            ("qkv_proj", "k_proj", "k"),
-            ("qkv_proj", "v_proj", "v"),
-            ("gate_up_proj", "gate_proj", 0),
-            ("gate_up_proj", "up_proj", 1),
-        ]
-        params_dict = dict(self.named_parameters())
-        loaded_params: Set[str] = set()
-        for name, loaded_weight in weights:
-            if "rotary_emb.inv_freq" in name:
-                continue
-            if ("rotary_emb.cos_cached" in name
-                    or "rotary_emb.sin_cached" in name):
+        loader = AutoWeightsLoader(
+            self,
+            skip_prefixes=([
+                "rotary_emb.inv_freq",
                 # Models trained using ColossalAI may include these tensors in
                 # the checkpoint. Skip them.
-                continue
-            for (param_name, weight_name, shard_id) in stacked_params_mapping:
-                if weight_name not in name:
-                    continue
-                name = name.replace(weight_name, param_name)
-                # Skip loading extra bias for GPTQ models.
-                if name.endswith(".bias") and name not in params_dict:
-                    continue
-                if is_pp_missing_parameter(name, self):
-                    continue
-                param = params_dict[name]
-                weight_loader = param.weight_loader
-                weight_loader(param, loaded_weight, shard_id)
-                break
-            else:
-                # Skip loading extra bias for GPTQ models.
-                if name.endswith(".bias") and name not in params_dict:
-                    continue
-                if is_pp_missing_parameter(name, self):
-                    continue
-                param = params_dict[name]
-                weight_loader = getattr(param, "weight_loader",
-                                        default_weight_loader)
-                weight_loader(param, loaded_weight)
-            loaded_params.add(name)
-        return loaded_params
+                "rotary_emb.cos_cached",
+                "rotary_emb.sin_cached"
+            ]),
+        )
+        return loader.load_weights(weights)

vllm/model_executor/models/paligemma.py

@@ -8,7 +8,6 @@ from transformers import BatchFeature, PaliGemmaConfig
 
 from vllm.config import VllmConfig
 from vllm.logger import init_logger
-from vllm.model_executor.layers.sampler import SamplerOutput
 from vllm.model_executor.sampling_metadata import SamplingMetadata
 from vllm.multimodal import MULTIMODAL_REGISTRY
 from vllm.multimodal.inputs import (MultiModalDataDict, MultiModalFieldConfig,
@@ -260,10 +259,6 @@ class PaliGemmaForConditionalGeneration(nn.Module, SupportsMultiModal,
         self.make_empty_intermediate_tensors = (
             self.language_model.make_empty_intermediate_tensors)
 
-    @property
-    def sampler(self):
-        return self.language_model.sampler
-
     def _validate_pixel_values(self, data: torch.Tensor) -> torch.Tensor:
         h = w = self.config.vision_config.image_size
         expected_dims = (3, h, w)
@@ -369,7 +364,7 @@ class PaliGemmaForConditionalGeneration(nn.Module, SupportsMultiModal,
                 positions: torch.Tensor,
                 intermediate_tensors: Optional[IntermediateTensors] = None,
                 inputs_embeds: Optional[torch.Tensor] = None,
-                **kwargs: object) -> Union[SamplerOutput, IntermediateTensors]:
+                **kwargs: object) -> IntermediateTensors:
         if intermediate_tensors is not None:
             inputs_embeds = None
 
@@ -396,13 +391,6 @@ class PaliGemmaForConditionalGeneration(nn.Module, SupportsMultiModal,
         return self.language_model.compute_logits(hidden_states,
                                                   sampling_metadata)
 
-    def sample(
-        self,
-        logits: torch.Tensor,
-        sampling_metadata: SamplingMetadata,
-    ) -> Optional[SamplerOutput]:
-        return self.language_model.sample(logits, sampling_metadata)
-
     def load_weights(self, weights: Iterable[Tuple[str,
                                                    torch.Tensor]]) -> Set[str]:
         loader = AutoWeightsLoader(self)

vllm/model_executor/models/persimmon.py

@@ -38,7 +38,6 @@ from vllm.model_executor.layers.linear import (ColumnParallelLinear,
 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.sampler import SamplerOutput, get_sampler
 from vllm.model_executor.layers.vocab_parallel_embedding import (
     ParallelLMHead, VocabParallelEmbedding)
 from vllm.model_executor.model_loader.weight_utils import default_weight_loader
@@ -46,7 +45,7 @@ from vllm.model_executor.sampling_metadata import SamplingMetadata
 from vllm.sequence import IntermediateTensors
 
 from .interfaces import SupportsPP
-from .utils import (is_pp_missing_parameter,
+from .utils import (AutoWeightsLoader, is_pp_missing_parameter,
                     make_empty_intermediate_tensors_factory, make_layers,
                     maybe_prefix)
 
@@ -116,9 +115,10 @@ class PersimmonAttention(nn.Module):
 
         self.rotary_emb = get_rope(
             self.head_dim,
-            rotary_dim=int(self.partial_rotary_factor * self.head_dim),
+            rotary_dim=self.head_dim,
             max_position=self.max_position_embeddings,
             base=self.rope_theta,
+            partial_rotary_factor=self.partial_rotary_factor,
         )
         self.scaling = self.head_dim**-0.5
         self.attn = Attention(self.num_heads,
@@ -221,7 +221,7 @@ class PersimmonModel(nn.Module):
         quant_config = vllm_config.quant_config
 
         self.vocab_size = config.vocab_size
-
+        self.config = config
         self.embed_tokens = VocabParallelEmbedding(config.vocab_size,
                                                    config.hidden_size)
         self.start_layer, self.end_layer, self.layers = make_layers(
@@ -260,6 +260,38 @@ class PersimmonModel(nn.Module):
         hidden_states = self.final_layernorm(hidden_states)
         return hidden_states
 
+    def load_weights(self, weights: Iterable[Tuple[str,
+                                                   torch.Tensor]]) -> Set[str]:
+        params_dict = dict(self.named_parameters(remove_duplicate=False))
+        loaded_params: Set[str] = set()
+        for name, loaded_weight in weights:
+            if is_pp_missing_parameter(name, self):
+                continue
+            param = params_dict[name]
+
+            if "query_key_value" in name:
+                # copy from vllm/model_executor/models/bloom.py
+                # NOTE: Persimmon's fused QKV's output_dim has the shape of
+                # (num_heads * 3 * head_size), while the
+                # required shape is (3 * num_heads * head_size).
+                # Thus, we need weight conversion.
+                output_dim = getattr(param, "output_dim", None)
+                num_heads = self.config.num_attention_heads
+                if output_dim is not None:
+                    loaded_weight_shape = loaded_weight.shape
+                    loaded_weight = loaded_weight.view(
+                        loaded_weight_shape[:output_dim] + (num_heads, 3, -1) +
+                        loaded_weight_shape[output_dim + 1:])
+                    loaded_weight = loaded_weight.transpose(
+                        output_dim, output_dim + 1)
+                    loaded_weight = loaded_weight.reshape(loaded_weight_shape)
+
+            weight_loader = getattr(param, "weight_loader",
+                                    default_weight_loader)
+            weight_loader(param, loaded_weight)
+            loaded_params.add(name)
+        return loaded_params
+
 
 class PersimmonForCausalLM(nn.Module, SupportsPP):
 
@@ -274,7 +306,6 @@ class PersimmonForCausalLM(nn.Module, SupportsPP):
                                       config.hidden_size,
                                       bias=False)
         self.logits_processor = LogitsProcessor(config.vocab_size)
-        self.sampler = get_sampler()
         self.make_empty_intermediate_tensors = (
             self.model.make_empty_intermediate_tensors)
 
@@ -305,49 +336,7 @@ class PersimmonForCausalLM(nn.Module, SupportsPP):
                                        sampling_metadata)
         return logits
 
-    def sample(
-        self,
-        logits: torch.Tensor,
-        sampling_metadata: SamplingMetadata,
-    ) -> Optional[SamplerOutput]:
-        next_tokens = self.sampler(logits, sampling_metadata)
-        return next_tokens
-
     def load_weights(self, weights: Iterable[Tuple[str,
                                                    torch.Tensor]]) -> Set[str]:
-        params_dict = dict(self.named_parameters(remove_duplicate=False))
-        loaded_params: Set[str] = set()
-        for name, loaded_weight in weights:
-            if "rotary_emb.inv_freq" in name:
-                continue
-            if ("rotary_emb.cos_cached" in name
-                    or "rotary_emb.sin_cached" in name):
-                # Models trained using ColossalAI may include these tensors in
-                # the checkpoint. Skip them.
-                continue
-            if is_pp_missing_parameter(name, self):
-                continue
-            param = params_dict[name]
-
-            if "query_key_value" in name:
-                # copy from vllm/model_executor/models/bloom.py
-                # NOTE: Persimmon's fused QKV's output_dim has the shape of
-                # (num_heads * 3 * head_size), while the
-                # required shape is (3 * num_heads * head_size).
-                # Thus, we need weight conversion.
-                output_dim = getattr(param, "output_dim", None)
-                num_heads = self.config.num_attention_heads
-                if output_dim is not None:
-                    loaded_weight_shape = loaded_weight.shape
-                    loaded_weight = loaded_weight.view(
-                        loaded_weight_shape[:output_dim] + (num_heads, 3, -1) +
-                        loaded_weight_shape[output_dim + 1:])
-                    loaded_weight = loaded_weight.transpose(
-                        output_dim, output_dim + 1)
-                    loaded_weight = loaded_weight.reshape(loaded_weight_shape)
-
-            weight_loader = getattr(param, "weight_loader",
-                                    default_weight_loader)
-            weight_loader(param, loaded_weight)
-            loaded_params.add(name)
-        return loaded_params
+        loader = AutoWeightsLoader(self)
+        return loader.load_weights(weights)

vllm/model_executor/models/phi.py

@@ -53,7 +53,6 @@ from vllm.model_executor.layers.linear import (ColumnParallelLinear,
 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.sampler import SamplerOutput, get_sampler
 from vllm.model_executor.layers.vocab_parallel_embedding import (
     ParallelLMHead, VocabParallelEmbedding)
 from vllm.model_executor.model_loader.weight_utils import default_weight_loader
@@ -322,7 +321,6 @@ class PhiForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
                                       bias=True,
                                       quant_config=quant_config)
         self.logits_processor = LogitsProcessor(config.vocab_size)
-        self.sampler = get_sampler()
         self.make_empty_intermediate_tensors = (
             self.model.make_empty_intermediate_tensors)
 
@@ -350,14 +348,6 @@ class PhiForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
                                        sampling_metadata, self.lm_head.bias)
         return logits
 
-    def sample(
-        self,
-        logits: torch.Tensor,
-        sampling_metadata: SamplingMetadata,
-    ) -> Optional[SamplerOutput]:
-        next_tokens = self.sampler(logits, sampling_metadata)
-        return next_tokens
-
     def load_weights(self, weights: Iterable[Tuple[str,
                                                    torch.Tensor]]) -> Set[str]:
         loader = AutoWeightsLoader(self)

vllm/model_executor/models/phi3_small.py

@@ -17,7 +17,6 @@ from vllm.model_executor.layers.linear import (MergedColumnParallelLinear,
 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.sampler import SamplerOutput, get_sampler
 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
@@ -26,7 +25,7 @@ from vllm.platforms import current_platform
 from vllm.sequence import IntermediateTensors
 
 from .interfaces import SupportsPP
-from .utils import (is_pp_missing_parameter,
+from .utils import (AutoWeightsLoader, WeightsMapper, is_pp_missing_parameter,
                     make_empty_intermediate_tensors_factory, make_layers,
                     maybe_prefix)
 
@@ -353,10 +352,29 @@ class Phi3SmallModel(nn.Module):
         hidden_states = self.final_layernorm(hidden_states)
         return hidden_states
 
+    def load_weights(self, weights: Iterable[Tuple[str,
+                                                   torch.Tensor]]) -> Set[str]:
+        params_dict = dict(self.named_parameters())
+        loaded_params: Set[str] = set()
+        for name, loaded_weight in weights:
+            if name.endswith(".bias") and name not in params_dict:
+                continue
+            if is_pp_missing_parameter(name, self):
+                continue
+            param = params_dict[name]
+            weight_loader = getattr(param, "weight_loader",
+                                    default_weight_loader)
+            weight_loader(param, loaded_weight)
+            loaded_params.add(name)
+        return loaded_params
+
 
 class Phi3SmallForCausalLM(nn.Module, SupportsPP):
     _tied_weights_keys = ["lm_head.weight"]
 
+    hf_to_vllm_mapper = WeightsMapper(
+        orig_to_new_suffix={"rotary_emb.inv_freq": None})
+
     def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
         super().__init__()
         config = vllm_config.model_config.hf_config
@@ -377,7 +395,6 @@ class Phi3SmallForCausalLM(nn.Module, SupportsPP):
         if self.config.tie_word_embeddings:
             self.lm_head.weight = self.model.embed_tokens.weight
         self.logits_processor = LogitsProcessor(config.vocab_size)
-        self.sampler = get_sampler()
         self.make_empty_intermediate_tensors = (
             self.model.make_empty_intermediate_tensors)
 
@@ -418,6 +435,7 @@ class Phi3SmallForCausalLM(nn.Module, SupportsPP):
                                        sampling_metadata)
         if self.dummy_token_indices is not None and logits is not None:
             logits.index_fill_(-1, self.dummy_token_indices, -torch.inf)
+        logits = logits / self.mup_width_multiplier
         return logits
 
     def forward(
@@ -436,33 +454,10 @@ class Phi3SmallForCausalLM(nn.Module, SupportsPP):
         output_hidden_states = output_hidden_states
         return output_hidden_states
 
-    def sample(
-        self,
-        logits: torch.Tensor,
-        sampling_metadata: SamplingMetadata,
-    ) -> Optional[SamplerOutput]:
-
-        next_tokens = self.sampler(logits / self.mup_width_multiplier,
-                                   sampling_metadata)
-        return next_tokens
-
     def load_weights(self, weights: Iterable[Tuple[str,
                                                    torch.Tensor]]) -> Set[str]:
-
-        params_dict = dict(self.named_parameters())
-        loaded_params: Set[str] = set()
-        for name, loaded_weight in weights:
-            if "rotary_emb.inv_freq" in name:
-                continue
-            if name.endswith(".bias") and name not in params_dict:
-                continue
-            if is_pp_missing_parameter(name, self):
-                continue
-            if "lm_head.weight" in name and self.config.tie_word_embeddings:
-                continue
-            param = params_dict[name]
-            weight_loader = getattr(param, "weight_loader",
-                                    default_weight_loader)
-            weight_loader(param, loaded_weight)
-            loaded_params.add(name)
-        return loaded_params
+        loader = AutoWeightsLoader(
+            self,
+            skip_prefixes=(["lm_head.weight"]
+                           if self.config.tie_word_embeddings else None))
+        return loader.load_weights(weights, mapper=self.hf_to_vllm_mapper)

vllm/model_executor/models/phi3v.py

@@ -16,7 +16,6 @@
 # limitations under the License.
 import re
 from collections.abc import Iterable, Mapping, Sequence
-from functools import cached_property
 from typing import Any, List, Literal, Optional, Set, Tuple, TypedDict, Union
 
 import torch
@@ -27,7 +26,6 @@ from transformers import (BatchFeature, CLIPVisionConfig, PretrainedConfig,
 from vllm.config import VllmConfig
 from vllm.logger import init_logger
 from vllm.model_executor.layers.quantization import QuantizationConfig
-from vllm.model_executor.layers.sampler import SamplerOutput, get_sampler
 from vllm.model_executor.layers.vocab_parallel_embedding import (
     VocabParallelEmbedding)
 from vllm.model_executor.sampling_metadata import SamplingMetadata
@@ -327,7 +325,7 @@ class Phi3VProcessingInfo(BaseProcessingInfo):
         *,
         image_width: int,
         image_height: int,
-        processor: Optional[ProcessorMixin],
+        processor: Optional[ProcessorMixin] = None,
     ) -> int:
         if processor is None:
             processor = self.get_hf_processor()
@@ -555,13 +553,6 @@ class Phi3VForCausalLM(nn.Module, SupportsMultiModal, SupportsPP,
         self.make_empty_intermediate_tensors = (
             self.language_model.make_empty_intermediate_tensors)
 
-    @cached_property
-    def sampler(self):
-        if hasattr(self.language_model, "sampler"):
-            return self.language_model.sampler
-
-        return get_sampler()
-
     def _validate_image_sizes(self, data: torch.Tensor) -> torch.Tensor:
         expected_dims = (2, )
 
@@ -716,13 +707,6 @@ class Phi3VForCausalLM(nn.Module, SupportsMultiModal, SupportsPP,
         return self.language_model.compute_logits(hidden_states,
                                                   sampling_metadata)
 
-    def sample(
-        self,
-        logits: torch.Tensor,
-        sampling_metadata: SamplingMetadata,
-    ) -> Optional[SamplerOutput]:
-        return self.language_model.sample(logits, sampling_metadata)
-
     def load_weights(self, weights: Iterable[Tuple[str,
                                                    torch.Tensor]]) -> Set[str]: