Add files using upload-large-folder tool

.venv/lib/python3.11/site-packages/vllm/model_executor/models/__init__.py

@@ -0,0 +1,24 @@
+# SPDX-License-Identifier: Apache-2.0
+
+from .interfaces import (HasInnerState, SupportsLoRA, SupportsMultiModal,
+                         SupportsPP, has_inner_state, supports_lora,
+                         supports_multimodal, supports_pp)
+from .interfaces_base import (VllmModelForPooling, VllmModelForTextGeneration,
+                              is_pooling_model, is_text_generation_model)
+from .registry import ModelRegistry
+
+__all__ = [
+    "ModelRegistry",
+    "VllmModelForPooling",
+    "is_pooling_model",
+    "VllmModelForTextGeneration",
+    "is_text_generation_model",
+    "HasInnerState",
+    "has_inner_state",
+    "SupportsLoRA",
+    "supports_lora",
+    "SupportsMultiModal",
+    "supports_multimodal",
+    "SupportsPP",
+    "supports_pp",
+]

.venv/lib/python3.11/site-packages/vllm/model_executor/models/adapters.py

@@ -0,0 +1,250 @@
+# SPDX-License-Identifier: Apache-2.0
+
+from collections.abc import Iterable
+from typing import TYPE_CHECKING, Any, Optional, TypeVar
+
+import torch
+import torch.nn as nn
+
+from .interfaces_base import VllmModelForPooling, is_pooling_model
+
+if TYPE_CHECKING:
+    from vllm.model_executor.layers.pooler import PoolingType
+
+_T = TypeVar("_T", bound=type[nn.Module])
+
+_GENERATE_SUFFIXES = [
+    "ForCausalLM",
+    "ForConditionalGeneration",
+    "ChatModel",
+    "LMHeadModel",
+]
+
+
+def _get_pooling_model_name(orig_model_name: str, pooling_suffix: str) -> str:
+    model_name = orig_model_name
+
+    for generate_suffix in _GENERATE_SUFFIXES:
+        model_name = model_name.removesuffix(generate_suffix)
+
+    return model_name + pooling_suffix
+
+
+def _create_pooling_model_cls(
+    orig_cls: _T,
+    *,
+    default_pooling_type: "PoolingType",
+    default_normalize: bool,
+    default_softmax: bool,
+) -> _T:
+    # Lazy import
+    from vllm.config import VllmConfig
+    from vllm.model_executor.layers.pooler import Pooler, PoolerOutput
+    from vllm.model_executor.pooling_metadata import PoolingMetadata
+
+    from .utils import AutoWeightsLoader, WeightsMapper
+
+    class ModelForPooling(orig_cls, VllmModelForPooling):
+
+        def __init__(
+            self,
+            *,
+            vllm_config: "VllmConfig",
+            prefix: str = "",
+            **kwargs: Any,
+        ) -> None:
+            super().__init__(vllm_config=vllm_config, prefix=prefix, **kwargs)
+
+            # These are not used in pooling models
+            for attr in ("lm_head", "logits_processor"):
+                if hasattr(self, attr):
+                    delattr(self, attr)
+
+            pooler_config = vllm_config.model_config.pooler_config
+            assert pooler_config is not None
+
+            # If the model already defines a pooler instance, don't overwrite it
+            if not getattr(self, "_pooler", None):
+                self._pooler = Pooler.from_config_with_defaults(
+                    pooler_config,
+                    pooling_type=default_pooling_type,
+                    normalize=default_normalize,
+                    softmax=default_softmax,
+                )
+
+        def pooler(
+            self,
+            hidden_states: torch.Tensor,
+            pooling_metadata: PoolingMetadata,
+        ) -> PoolerOutput:
+            return self._pooler(hidden_states, pooling_metadata)
+
+        def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]):
+            # TODO: Support uninitialized params tracking
+
+            # We have deleted this attribute, so don't load it
+            weights = ((name, data) for name, data in weights
+                       if not name.startswith("lm_head."))
+
+            # If `*ForCausalLM` defines `load_weights` on the inner model
+            # and there are no other inner modules with parameters,
+            # we support loading from both `*Model` and `*ForCausalLM`
+            if hasattr(self, "model") and hasattr(self.model, "load_weights"):
+                # Whether only `self.model` contains parameters
+                model_is_only_param = all(
+                    name == "model" or next(child.parameters(), None) is None
+                    for name, child in self.named_children())
+
+                if model_is_only_param:
+                    mapper = WeightsMapper(orig_to_new_prefix={"model.": ""})
+                    weights = mapper.apply(weights)
+
+                    self.model.load_weights(weights)
+                    return
+
+            # For most other models
+            if hasattr(orig_cls, "load_weights"):
+                orig_cls.load_weights(self, weights)  # type: ignore
+            # Fallback
+            else:
+                loader = AutoWeightsLoader(self)
+                loader.load_weights(weights)
+
+    return ModelForPooling  # type: ignore
+
+
+def as_embedding_model(cls: _T) -> _T:
+    """
+    Subclass an existing vLLM model to support embeddings.
+
+    By default, the embeddings of the whole prompt are extracted from the
+    normalized hidden state corresponding to the last token.
+
+    Note:
+        We assume that no extra layers are added to the original model;
+        please implement your own model if this is not the case.
+    """
+    # Avoid modifying existing embedding models
+    if is_pooling_model(cls):
+        return cls
+
+    # Lazy import
+    from vllm.model_executor.layers.pooler import PoolingType
+
+    ModelForEmbedding = _create_pooling_model_cls(
+        cls,
+        default_pooling_type=PoolingType.LAST,
+        default_normalize=True,
+        default_softmax=False,
+    )
+    ModelForEmbedding.__name__ = \
+        _get_pooling_model_name(cls.__name__, "ForEmbedding")
+
+    return ModelForEmbedding  # type: ignore
+
+
+def as_classification_model(cls: _T) -> _T:
+    """
+    Subclass an existing vLLM model to support classification.
+
+    By default, the class probabilities are extracted from the softmaxed
+    hidden state corresponding to the last token.
+
+    Note:
+        We assume that the classification head is a single linear layer
+        stored as the attribute `score` of the top-level model;
+        please implement your own model if this is not the case.
+    """
+    # Avoid modifying existing classification models
+    if is_pooling_model(cls):
+        return cls
+
+    # Lazy import
+    from vllm.attention import AttentionMetadata
+    from vllm.config import VllmConfig
+    from vllm.model_executor.layers.linear import RowParallelLinear
+    from vllm.model_executor.layers.pooler import PoolingType
+    from vllm.sequence import IntermediateTensors
+
+    from .utils import maybe_prefix
+
+    ModelForPooling = _create_pooling_model_cls(
+        cls,
+        default_pooling_type=PoolingType.LAST,
+        default_normalize=False,
+        default_softmax=True,
+    )
+
+    class ModelForClassification(ModelForPooling):
+
+        def __init__(
+            self,
+            *,
+            vllm_config: "VllmConfig",
+            prefix: str = "",
+            **kwargs: Any,
+        ) -> None:
+            super().__init__(vllm_config=vllm_config, prefix=prefix, **kwargs)
+
+            config = vllm_config.model_config.hf_config
+            quant_config = vllm_config.quant_config
+
+            self.score = RowParallelLinear(config.hidden_size,
+                                           config.num_labels,
+                                           quant_config=quant_config,
+                                           input_is_parallel=False,
+                                           bias=False,
+                                           prefix=maybe_prefix(
+                                               prefix, "score"))
+
+        def forward(
+            self,
+            input_ids: torch.Tensor,
+            positions: torch.Tensor,
+            kv_caches: list[torch.Tensor],
+            attn_metadata: AttentionMetadata,
+            intermediate_tensors: Optional[IntermediateTensors] = None,
+            inputs_embeds: Optional[torch.Tensor] = None,
+        ) -> torch.Tensor:
+            hidden_states = super().forward(input_ids, positions, kv_caches,
+                                            attn_metadata,
+                                            intermediate_tensors,
+                                            inputs_embeds)
+            logits, _ = self.score(hidden_states)
+            return logits
+
+
+    ModelForClassification.__name__ = \
+        _get_pooling_model_name(cls.__name__, "ForClassification")
+
+    return ModelForClassification  # type: ignore
+
+
+def as_reward_model(cls: _T) -> _T:
+    """
+    Subclass an existing vLLM model to support reward modeling.
+
+    By default, we return the hidden states of each token directly.
+
+    Note:
+        We assume that no extra layers are added to the original model;
+        please implement your own model if this is not the case.
+    """
+    # Avoid modifying existing reward models
+    if is_pooling_model(cls):
+        return cls
+
+    # Lazy import
+    from vllm.model_executor.layers.pooler import PoolingType
+
+    ModelForReward = _create_pooling_model_cls(
+        cls,
+        default_pooling_type=PoolingType.ALL,
+        default_normalize=False,
+        default_softmax=False,
+    )
+
+    ModelForReward.__name__ = \
+        _get_pooling_model_name(cls.__name__, "ForReward")
+
+    return ModelForReward  # type: ignore

.venv/lib/python3.11/site-packages/vllm/model_executor/models/aria.py

@@ -0,0 +1,663 @@
+# SPDX-License-Identifier: Apache-2.0
+
+from typing import (Iterable, List, Mapping, Optional, Set, Tuple, TypedDict,
+                    Union)
+
+import torch
+import torch.nn as nn
+from transformers import AriaConfig, AriaTextConfig, BatchFeature
+from transformers.models.aria.modeling_aria import AriaCrossAttention
+from transformers.models.aria.processing_aria import AriaProcessor
+
+from vllm.attention import AttentionMetadata
+from vllm.config import CacheConfig, QuantizationConfig, VllmConfig
+from vllm.distributed import get_tensor_model_parallel_rank
+from vllm.model_executor.layers.activation import get_act_fn
+from vllm.model_executor.layers.fused_moe import FusedMoE
+from vllm.model_executor.layers.linear import (ColumnParallelLinear,
+                                               RowParallelLinear)
+from vllm.model_executor.layers.logits_processor import LogitsProcessor
+from vllm.model_executor.layers.sampler import (SamplerOutput,
+                                                SamplingMetadata, get_sampler)
+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.multimodal import MULTIMODAL_REGISTRY
+from vllm.multimodal.inputs import (MultiModalFieldConfig, MultiModalKwargs,
+                                    NestedTensors)
+from vllm.multimodal.parse import MultiModalDataItems
+from vllm.multimodal.processing import (BaseMultiModalProcessor,
+                                        BaseProcessingInfo, PromptReplacement)
+from vllm.multimodal.profiling import BaseDummyInputsBuilder, ProcessorInputs
+from vllm.sequence import IntermediateTensors
+
+# yapf: disable
+from .idefics2_vision_model import Idefics2VisionConfig
+from .idefics2_vision_model import (
+    Idefics2VisionTransformer as Idefics3VisionTransformer)
+# yapf: enable
+from .interfaces import SupportsMultiModal
+from .llama import LlamaDecoderLayer, LlamaMLP, LlamaModel
+from .utils import (AutoWeightsLoader, WeightsMapper, flatten_bn,
+                    is_pp_missing_parameter, maybe_prefix,
+                    merge_multimodal_embeddings)
+
+
+class AriaImagePixelInputs(TypedDict):
+    pixel_values: torch.Tensor
+    pixel_mask: Optional[torch.Tensor]
+    """
+    Shape: 
+        pixel_values: `(batch_size * num_images, num_channels, height, width)`
+        pixel_mask: `(batch_size * num_images, height, width)`
+    """
+
+
+class AriaVisionTransformer(Idefics3VisionTransformer):
+
+    def __init__(
+        self,
+        config: Idefics2VisionConfig,
+        quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
+    ) -> None:
+        super().__init__(config, quant_config, prefix)
+        # Unlike Idefics3VisionTransformer which uses LayerNorm after the
+        # final layer, Aria omits this normalization, so we replace it with an
+        # Identity layer
+        self.post_layernorm = nn.Identity()
+
+    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())
+        loaded_params: Set[str] = set()
+        for name, loaded_weight in weights:
+
+            # NOTE: post_layernorm is not used in Aria
+            if "post_layernorm" in name:
+                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)
+                param = params_dict[name]
+                weight_loader = param.weight_loader
+                weight_loader(param, loaded_weight, shard_id)
+                break
+            else:
+                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 AriaProjectorMLP(nn.Module):
+
+    def __init__(
+        self,
+        in_features: int,
+        hidden_features: int,
+        output_dim: int,
+    ) -> None:
+        super().__init__()
+
+        self.linear_in = ColumnParallelLinear(in_features,
+                                              hidden_features,
+                                              bias=False)
+        self.linear_out = RowParallelLinear(hidden_features,
+                                            output_dim,
+                                            bias=False)
+        self.act = get_act_fn("gelu_new")
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        hidden_states, _ = self.linear_in(hidden_states)
+        hidden_states = self.act(hidden_states)
+        hidden_states, _ = self.linear_out(hidden_states)
+        return hidden_states
+
+
+class AriaProjector(nn.Module):
+    """
+    A projection module with one cross attention layer and one FFN layer, which
+    projects ViT's outputs into MoE's inputs.
+
+    Args:
+        patch_to_query_dict (dict): Maps patch numbers to their corresponding
+        query numbers,
+            e.g., {1225: 128, 4900: 256}. This allows for different query sizes
+            based on image resolution.
+        embed_dim (int): Embedding dimension. 
+        num_heads (int): Number of attention heads. 
+        kv_dim (int): Dimension of key and value. 
+        ff_dim (int): Hidden dimension of the feed-forward network. 
+        output_dim (int): Output dimension. 
+        norm_layer (nn.Module): Normalization layer. Default is nn.LayerNorm.
+
+    Outputs:
+        A tensor with the shape of (batch_size, query_number, output_dim)
+    """
+
+    def __init__(self, config: AriaConfig) -> None:
+        super().__init__()
+
+        self.patch_to_query_dict = config.projector_patch_to_query_dict
+        self.in_features = config.vision_config.hidden_size
+        self.num_heads = config.vision_config.num_attention_heads
+        self.kv_dim = config.vision_config.hidden_size
+        self.hidden_features = config.text_config.hidden_size
+        self.output_dim = config.text_config.hidden_size
+
+        self.query = nn.Parameter(
+            torch.empty(config.max_value_projector_patch_to_query_dict,
+                        self.in_features))
+
+        self.cross_attn = AriaCrossAttention(config)
+
+        self.layer_norm = nn.LayerNorm(self.in_features)
+        self.feed_forward = AriaProjectorMLP(self.in_features,
+                                             self.hidden_features,
+                                             self.output_dim)
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        attn_mask: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        batch_size, num_patches = x.shape[0], x.shape[1]
+
+        if num_patches not in self.patch_to_query_dict:
+            raise KeyError(f"Number of patches {num_patches} not found in "
+                           "patch_to_query_dict amongst possible values "
+                           f"{self.patch_to_query_dict.keys()}.")
+
+        query_num = self.patch_to_query_dict[num_patches]
+
+        queries = self.query[:query_num].unsqueeze(0).repeat(batch_size, 1, 1)
+
+        if attn_mask is not None:
+            attn_mask = attn_mask.repeat_interleave(self.num_heads, 0)
+            attn_mask = attn_mask.unsqueeze(1).expand(-1, queries.size(1), -1)
+
+        attention_out = self.cross_attn(x, queries, attn_mask=attn_mask)
+
+        out = self.feed_forward(self.layer_norm(attention_out))
+
+        return out
+
+
+class AriaFusedMoE(FusedMoE):
+
+    def weight_loader(self, param: nn.Parameter, loaded_weight: torch.Tensor,
+                      shard_id: str) -> None:
+        # Override the weight_loader to handle the expert weights in the Aria
+        # model, which are already packed with experts, and merge the gate and
+        # up weights for each expert.
+        # Note: Loading expert weights with quantization is not supported
+        tp_rank = get_tensor_model_parallel_rank()
+        if shard_id == 'w13':
+            # the shape of loaded_weight is
+            # (num_experts, hidden_size, 2 * moe_intermediate_size)
+            if self.tp_size > 1:
+                up, gate = loaded_weight.chunk(2, dim=-1)
+                up_current_rank = up.chunk(self.tp_size, dim=-1)[tp_rank]
+                gate_current_rank = gate.chunk(self.tp_size, dim=-1)[tp_rank]
+                up_and_gate = torch.cat([up_current_rank, gate_current_rank],
+                                        dim=-1).transpose(1, 2)
+                param.data.copy_(up_and_gate)
+            else:
+                param.data.copy_(loaded_weight.transpose(1, 2))
+        elif shard_id == 'w2':
+            # the shape of loaded_weight is
+            # (num_experts, moe_intermediate_size, hidden_size)
+            if self.tp_size > 1:
+                down_current_rank = loaded_weight.chunk(self.tp_size,
+                                                        dim=1)[tp_rank]
+                param.data.copy_(down_current_rank.transpose(1, 2))
+            else:
+                param.data.copy_(loaded_weight.transpose(1, 2))
+
+
+class AriaTextMoELayer(nn.Module):
+    """
+    Mixture of Experts (MoE) Layer for the AriaMoE model.
+
+    This layer implements the MoE mechanism, which routes input tokens to
+    different experts based on a routing algorithm, processes them through the
+    experts, and then combines the outputs.
+    """
+
+    def __init__(
+        self,
+        config: AriaTextConfig,
+        quant_config: Optional[QuantizationConfig],
+    ) -> None:
+        super().__init__()
+        self.config = config
+
+        self.router_weight = nn.Parameter(
+            torch.empty(
+                (self.config.moe_num_experts, self.config.hidden_size)))
+
+        self.experts = AriaFusedMoE(
+            num_experts=config.moe_num_experts,
+            top_k=config.moe_topk,
+            hidden_size=config.hidden_size,
+            intermediate_size=config.intermediate_size,
+            quant_config=quant_config,
+            reduce_results=True,
+        )
+        self.shared_experts = LlamaMLP(
+            config.hidden_size,
+            config.intermediate_size * config.moe_num_shared_experts,
+            "silu",
+            quant_config=quant_config,
+            bias=config.mlp_bias,
+        )
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        """
+        Forward pass of the MoE Layer.
+
+        Args:
+            hidden_states (torch.Tensor): Input tensor of shape (batch_size,
+            sequence_length, hidden_size).
+
+        Returns:
+            torch.Tensor: Output tensor after passing through the MoE layer.
+        """
+
+        router_output = torch.nn.functional.linear(hidden_states,
+                                                   self.router_weight)
+
+        hidden_states_copy = hidden_states.clone()
+        # NOTE: hidden_states will be modified inplace by `FusedMoE`
+        sparse_expert_output = self.experts(hidden_states, router_output)
+        shared_expert_output = self.shared_experts(hidden_states_copy)
+
+        return sparse_expert_output + shared_expert_output
+
+
+class AriaTextDecoderLayer(LlamaDecoderLayer):
+    """
+    Custom Decoder Layer for the AriaMoE model which modifies the standard
+    `LlamaDecoderLayer` by replacing the traditional MLP with a Mixture of
+    Experts (MoE) Layer.
+    """
+
+    def __init__(
+        self,
+        config: AriaTextConfig,
+        cache_config: Optional[CacheConfig] = None,
+        quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
+    ) -> None:
+        super().__init__(config, cache_config, quant_config, prefix)
+        self.mlp = AriaTextMoELayer(config, quant_config=quant_config)
+
+
+class AriaTextModel(LlamaModel):
+    """
+    Custom LlamaModel for the AriaMoE model which modifies the standard
+    LlamaModel by replacing the `LlamaDecoderLayer` with `MoEDecoderLayer`.
+    """
+
+    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
+        super().__init__(vllm_config=vllm_config,
+                         prefix=prefix,
+                         layer_type=AriaTextDecoderLayer)
+
+    # Adapted from LlamaModel.load_weights with the modification of adding
+    # the expert weights mapping to `stacked_params_mapping`
+    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),
+            ("experts.w13_weight", "experts.fc1.weight", 'w13'),
+            ("experts.w2_weight", "experts.fc2.weight", 'w2'),
+        ]
+        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):
+                # Models trained using ColossalAI may include these tensors in
+                # the checkpoint. Skip them.
+                continue
+            if (self.quant_config is not None and
+                (scale_name := self.quant_config.get_cache_scale(name))):
+                # Loading kv cache quantization scales
+                param = params_dict[scale_name]
+                weight_loader = getattr(param, "weight_loader",
+                                        default_weight_loader)
+                loaded_weight = (loaded_weight if loaded_weight.dim() == 0 else
+                                 loaded_weight[0])
+                weight_loader(param, loaded_weight)
+                loaded_params.add(scale_name)
+                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
+                # Remapping the name of FP8 kv-scale.
+                name = maybe_remap_kv_scale_name(name, params_dict)
+                if name is None:
+                    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 AriaProcessingInfo(BaseProcessingInfo):
+
+    def get_hf_config(self):
+        return self.ctx.get_hf_config(AriaConfig)
+
+    def get_vision_config(self):
+        return self.get_hf_config().vision_config
+
+    def get_hf_processor(self):
+        return self.ctx.get_hf_processor(AriaProcessor)
+
+    def get_supported_mm_limits(self) -> Mapping[str, Optional[int]]:
+        return {"image": None}
+
+    def get_mm_max_tokens_per_item(
+        self,
+        seq_len: int,
+        mm_counts: Mapping[str, int],
+    ) -> Mapping[str, int]:
+        return {"image": self.get_num_image_tokens()}
+
+    def get_num_image_tokens(self) -> int:
+        hf_config = self.get_hf_config()
+        return max(hf_config.projector_patch_to_query_dict.values())
+
+
+class AriaDummyInputsBuilder(BaseDummyInputsBuilder[AriaProcessingInfo]):
+
+    def get_dummy_processor_inputs(
+        self,
+        seq_len: int,
+        mm_counts: Mapping[str, int],
+    ) -> ProcessorInputs:
+        vision_config = self.info.get_vision_config()
+
+        max_image_size = vision_config.image_size
+        num_images = mm_counts.get("image", 0)
+
+        mm_data = {
+            "image":
+            self._get_dummy_images(width=max_image_size,
+                                   height=max_image_size,
+                                   num_images=num_images)
+        }
+
+        hf_processor = self.info.get_hf_processor()
+        image_token: str = hf_processor.tokenizer.image_token  # type: ignore
+
+        return ProcessorInputs(
+            prompt_text=image_token * num_images,
+            mm_data=mm_data,
+        )
+
+
+class AriaMultiModalProcessor(BaseMultiModalProcessor[AriaProcessingInfo]):
+
+    def _get_mm_fields_config(
+        self,
+        hf_inputs: BatchFeature,
+        hf_processor_mm_kwargs: Mapping[str, object],
+    ) -> Mapping[str, MultiModalFieldConfig]:
+        return dict(
+            pixel_values=MultiModalFieldConfig.batched("image"),
+            pixel_mask=MultiModalFieldConfig.batched("image"),
+        )
+
+    def _get_prompt_replacements(
+        self,
+        mm_items: MultiModalDataItems,
+        hf_processor_mm_kwargs: Mapping[str, object],
+        out_mm_kwargs: MultiModalKwargs,
+    ) -> list[PromptReplacement]:
+        hf_config = self.info.get_hf_config()
+        image_token_id = hf_config.image_token_index
+
+        num_image_tokens = self.info.get_num_image_tokens()
+
+        return [
+            PromptReplacement(
+                modality="image",
+                target=[image_token_id],
+                replacement=[image_token_id] * num_image_tokens,
+            )
+        ]
+
+
+@MULTIMODAL_REGISTRY.register_processor(AriaMultiModalProcessor,
+                                        info=AriaProcessingInfo,
+                                        dummy_inputs=AriaDummyInputsBuilder)
+class AriaForConditionalGeneration(nn.Module, SupportsMultiModal):
+    """
+    Aria model for conditional generation tasks.
+
+    This model combines a vision tower, a multi-modal projector, and a language
+    model to perform tasks that involve both image and text inputs.
+    """
+    hf_to_vllm_mapper = WeightsMapper(
+        orig_to_new_prefix={
+            "language_model.model": "language_model",
+            "language_model.lm_head": "lm_head",
+        },
+        orig_to_new_suffix={
+            "router.weight": "router_weight",
+        },
+    )
+
+    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.vision_tower = AriaVisionTransformer(
+            config.vision_config,
+            quant_config,
+            prefix=f"{prefix}.vision_tower",
+        )
+        self.multi_modal_projector = AriaProjector(config)
+        self.vocab_size = config.text_config.vocab_size
+        self.language_model = AriaTextModel(
+            vllm_config=vllm_config.with_hf_config(config.text_config),
+            prefix=maybe_prefix(prefix, "language_model.model"),
+        )
+        self.pad_token_id = (self.config.pad_token_id
+                             if self.config.pad_token_id is not None else -1)
+        self.unpadded_vocab_size = config.text_config.vocab_size
+        self.lm_head = ParallelLMHead(
+            self.unpadded_vocab_size,
+            config.text_config.hidden_size,
+            org_num_embeddings=self.language_model.org_vocab_size,
+            quant_config=quant_config,
+        )
+        logit_scale = getattr(config, "logit_scale", 1.0)
+        self.logits_processor = LogitsProcessor(self.unpadded_vocab_size,
+                                                self.vocab_size, logit_scale)
+        self.sampler = get_sampler()
+
+    def _validate_image_sizes(
+            self, images: List[torch.Tensor]) -> List[torch.Tensor]:
+        if not all(img.shape == images[0].shape for img in images):
+            raise ValueError("All images must be the same size")
+        return images
+
+    def _parse_and_validate_image_input(
+            self, **kwargs: object) -> Optional[AriaImagePixelInputs]:
+        pixel_values = kwargs.pop("pixel_values", None)
+        pixel_mask = kwargs.pop("pixel_mask", None)
+
+        if pixel_values is None:
+            return None
+
+        if not isinstance(pixel_values, (torch.Tensor, list)):
+            raise ValueError("Incorrect type of pixel values. "
+                             f"Got type: {type(pixel_values)}")
+
+        pixel_values = self._validate_image_sizes(pixel_values)
+        pixel_values = flatten_bn(pixel_values, concat=True)
+
+        if pixel_mask is not None:
+            if not isinstance(pixel_mask, (torch.Tensor, list)):
+                raise ValueError("Incorrect type of pixel mask. "
+                                 f"Got type: {type(pixel_mask)}")
+
+            pixel_mask = flatten_bn(pixel_mask, concat=True)
+
+        return AriaImagePixelInputs(
+            pixel_values=pixel_values,
+            pixel_mask=pixel_mask,
+        )
+
+    def _create_patch_attention_mask(
+            self, pixel_mask: Optional[torch.Tensor]) -> torch.Tensor:
+        if pixel_mask is None:
+            return None
+
+        patches_subgrid = pixel_mask.unfold(
+            dimension=1,
+            size=self.vision_tower.config.patch_size,
+            step=self.vision_tower.config.patch_size,
+        ).unfold(
+            dimension=2,
+            size=self.vision_tower.config.patch_size,
+            step=self.vision_tower.config.patch_size,
+        )
+        return (patches_subgrid.sum(dim=(-1, -2)) > 0).bool()
+
+    def _process_image_input(
+        self, image_input: AriaImagePixelInputs
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        assert self.vision_tower is not None
+
+        pixel_values = image_input['pixel_values']
+        pixel_mask = image_input['pixel_mask']
+
+        patch_attention_mask = self._create_patch_attention_mask(pixel_mask)
+
+        image_outputs = self.vision_tower(
+            pixel_values=pixel_values,
+            patch_attention_mask=patch_attention_mask,
+        )
+        image_attn_mask = None
+        if patch_attention_mask is not None:
+            flattened_mask = patch_attention_mask.flatten(1)
+            image_attn_mask = torch.logical_not(flattened_mask)
+
+        return self.multi_modal_projector(image_outputs, image_attn_mask)
+
+    def get_multimodal_embeddings(self, **kwargs) -> Optional[NestedTensors]:
+        image_input = self._parse_and_validate_image_input(**kwargs)
+        if image_input is None:
+            return None
+        multimodal_embeddings = self._process_image_input(image_input)
+        return multimodal_embeddings
+
+    def get_input_embeddings(
+        self,
+        input_ids: torch.Tensor,
+        multimodal_embeddings: Optional[NestedTensors] = None,
+    ) -> torch.Tensor:
+        inputs_embeds = self.language_model.get_input_embeddings(input_ids)
+        if multimodal_embeddings is not None:
+            inputs_embeds = merge_multimodal_embeddings(
+                input_ids, inputs_embeds, multimodal_embeddings,
+                self.config.image_token_index)
+        return inputs_embeds
+
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        positions: torch.Tensor,
+        kv_caches: List[torch.Tensor],
+        attn_metadata: AttentionMetadata,
+        intermediate_tensors: Optional[IntermediateTensors] = None,
+        inputs_embeds: Optional[torch.Tensor] = None,
+        **kwargs: object,
+    ) -> Union[torch.Tensor, IntermediateTensors]:
+        if inputs_embeds is None:
+            multimodal_embeddings = self.get_multimodal_embeddings(**kwargs)
+            # always pass the input via `inputs_embeds`
+            # to make sure the computation graph is consistent
+            inputs_embeds = self.get_input_embeddings(input_ids,
+                                                      multimodal_embeddings)
+            input_ids = None
+
+        hidden_states = self.language_model(
+            input_ids,
+            positions,
+            kv_caches,
+            attn_metadata,
+            intermediate_tensors,
+            inputs_embeds=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: 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]]):
+        loader = AutoWeightsLoader(self)
+        loader.load_weights(weights, mapper=self.hf_to_vllm_mapper)

.venv/lib/python3.11/site-packages/vllm/model_executor/models/baichuan.py

@@ -0,0 +1,493 @@
+# SPDX-License-Identifier: Apache-2.0
+
+# Copyright 2022 EleutherAI and the HuggingFace Inc. team. All rights reserved.
+#
+# This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX
+# and OPT implementations in this library. It has been modified from its
+# original forms to accommodate minor architectural differences compared
+# to GPT-NeoX and OPT used by the Meta AI team that trained the model.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Inference-only BaiChuan model compatible with HuggingFace weights."""
+import math
+from typing import Iterable, List, Optional, Set, Tuple, Union
+
+import torch
+from torch import nn
+from transformers import PretrainedConfig
+
+from vllm.attention import Attention, AttentionMetadata
+from vllm.compilation.decorators import support_torch_compile
+from vllm.config import CacheConfig, VllmConfig
+from vllm.distributed import (get_pp_group, get_tensor_model_parallel_rank,
+                              get_tensor_model_parallel_world_size)
+from vllm.model_executor.layers.activation import SiluAndMul
+from vllm.model_executor.layers.layernorm import RMSNorm
+from vllm.model_executor.layers.linear import (MergedColumnParallelLinear,
+                                               QKVParallelLinear,
+                                               RowParallelLinear)
+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
+from vllm.model_executor.sampling_metadata import SamplingMetadata
+from vllm.sequence import IntermediateTensors
+
+from .interfaces import SupportsLoRA, SupportsPP
+from .utils import (is_pp_missing_parameter,
+                    make_empty_intermediate_tensors_factory, make_layers)
+
+
+def _get_alibi_slopes(total_num_heads: int) -> torch.Tensor:
+    closest_power_of_2 = 2**math.floor(math.log2(total_num_heads))
+    base = torch.tensor(
+        2**(-(2**-(math.log2(closest_power_of_2) - 3))),
+        dtype=torch.float32,
+    )
+    powers = torch.arange(1, 1 + closest_power_of_2, dtype=torch.int32)
+    slopes = torch.pow(base, powers)
+
+    if closest_power_of_2 != total_num_heads:
+        extra_base = torch.tensor(
+            2**(-(2**-(math.log2(2 * closest_power_of_2) - 3))),
+            dtype=torch.float32,
+        )
+        num_remaining_heads = min(closest_power_of_2,
+                                  total_num_heads - closest_power_of_2)
+        extra_powers = torch.arange(start=1,
+                                    end=1 + 2 * num_remaining_heads,
+                                    step=2,
+                                    dtype=torch.int32)
+        slopes = torch.cat(
+            [slopes, torch.pow(extra_base, extra_powers)], dim=0)
+    return slopes
+
+
+class BaiChuanMLP(nn.Module):
+
+    def __init__(
+        self,
+        hidden_size: int,
+        intermediate_size: int,
+        hidden_act: str,
+        quant_config: Optional[QuantizationConfig] = None,
+    ):
+        super().__init__()
+        self.gate_up_proj = MergedColumnParallelLinear(
+            hidden_size, [intermediate_size] * 2,
+            bias=False,
+            quant_config=quant_config)
+        self.down_proj = RowParallelLinear(intermediate_size,
+                                           hidden_size,
+                                           bias=False,
+                                           quant_config=quant_config)
+        if hidden_act != "silu":
+            raise ValueError(f"Unsupported activation: {hidden_act}. "
+                             "Only silu is supported for now.")
+        self.act_fn = SiluAndMul()
+
+    def forward(self, x):
+        gate_up, _ = self.gate_up_proj(x)
+        x = self.act_fn(gate_up)
+        x, _ = self.down_proj(x)
+        return x
+
+
+class BaiChuanAttention(nn.Module):
+    """Multi-headed attention from 'Attention Is All You Need' paper"""
+
+    def __init__(
+        self,
+        hidden_size: int,
+        num_heads: int,
+        position_embedding: str,
+        rope_theta: float = 10000,
+        max_position_embeddings: int = 8192,
+        cache_config: Optional[CacheConfig] = None,
+        quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
+    ):
+        super().__init__()
+        self.hidden_size = hidden_size
+        tensor_model_parallel_world_size = get_tensor_model_parallel_world_size(
+        )
+        self.total_num_heads = num_heads
+        assert self.total_num_heads % tensor_model_parallel_world_size == 0
+        self.num_heads = (self.total_num_heads //
+                          tensor_model_parallel_world_size)
+        self.head_dim = hidden_size // self.total_num_heads
+        self.postion_embedding = position_embedding
+        self.rope_theta = rope_theta
+        self.max_position_embeddings = max_position_embeddings
+
+        # pylint: disable=invalid-name
+        self.W_pack = QKVParallelLinear(
+            hidden_size,
+            self.head_dim,
+            self.total_num_heads,
+            self.total_num_heads,
+            bias=False,
+            quant_config=quant_config,
+        )
+        self.o_proj = RowParallelLinear(
+            self.total_num_heads * self.head_dim,
+            hidden_size,
+            bias=False,
+            quant_config=quant_config,
+        )
+        # Create the alibi slopes and slice them.
+        if self.postion_embedding == "ALIBI":
+            tp_rank = get_tensor_model_parallel_rank()
+            head_start = tp_rank * self.num_heads
+            head_end = (tp_rank + 1) * self.num_heads
+            alibi_slopes = _get_alibi_slopes(self.total_num_heads)
+            alibi_slopes = alibi_slopes[head_start:head_end].tolist()
+
+            scaling = self.head_dim**-0.5
+            self.attn = Attention(self.num_heads,
+                                  self.head_dim,
+                                  scaling,
+                                  alibi_slopes=alibi_slopes,
+                                  quant_config=quant_config,
+                                  prefix=f"{prefix}.attn")
+        else:
+            self.rotary_emb = get_rope(
+                self.head_dim,
+                rotary_dim=self.head_dim,
+                max_position=self.max_position_embeddings,
+                base=self.rope_theta,
+            )
+            self.scaling = self.head_dim**-0.5
+            self.attn = Attention(self.num_heads,
+                                  self.head_dim,
+                                  self.scaling,
+                                  cache_config=cache_config,
+                                  quant_config=quant_config,
+                                  prefix=f"{prefix}.attn")
+
+    def forward(
+        self,
+        positions: torch.Tensor,
+        hidden_states: torch.Tensor,
+        kv_cache: torch.Tensor,
+        attn_metadata: AttentionMetadata,
+    ) -> torch.Tensor:
+        qkv, _ = self.W_pack(hidden_states)
+        q, k, v = qkv.chunk(chunks=3, dim=-1)
+        if self.postion_embedding != "ALIBI":
+            q, k = self.rotary_emb(positions, q, k)
+        attn_output = self.attn(q, k, v, kv_cache, attn_metadata)
+        output, _ = self.o_proj(attn_output)
+        return output
+
+
+class BaiChuanDecoderLayer(nn.Module):
+
+    def __init__(self,
+                 config: PretrainedConfig,
+                 position_embedding: str,
+                 cache_config: Optional[CacheConfig] = None,
+                 quant_config: Optional[QuantizationConfig] = None,
+                 prefix: str = ""):
+        super().__init__()
+        self.hidden_size = config.hidden_size
+        rope_theta = getattr(config, "rope_theta", 10000)
+        max_position_embeddings = getattr(config, "max_position_embeddings",
+                                          8192)
+        self.self_attn = BaiChuanAttention(
+            hidden_size=self.hidden_size,
+            num_heads=config.num_attention_heads,
+            position_embedding=position_embedding,
+            rope_theta=rope_theta,
+            max_position_embeddings=max_position_embeddings,
+            cache_config=cache_config,
+            quant_config=quant_config,
+            prefix=f"{prefix}.self_attn",
+        )
+        self.mlp = BaiChuanMLP(
+            hidden_size=self.hidden_size,
+            intermediate_size=config.intermediate_size,
+            hidden_act=config.hidden_act,
+            quant_config=quant_config,
+        )
+        self.input_layernorm = RMSNorm(config.hidden_size,
+                                       eps=config.rms_norm_eps)
+        self.post_attention_layernorm = RMSNorm(config.hidden_size,
+                                                eps=config.rms_norm_eps)
+
+    def forward(
+        self,
+        positions: torch.Tensor,
+        hidden_states: torch.Tensor,
+        kv_cache: torch.Tensor,
+        attn_metadata: AttentionMetadata,
+        residual: Optional[torch.Tensor],
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        # Self Attention
+        if residual is None:
+            residual = hidden_states
+            hidden_states = self.input_layernorm(hidden_states)
+        else:
+            hidden_states, residual = self.input_layernorm(
+                hidden_states, residual)
+        hidden_states = self.self_attn(
+            positions=positions,
+            hidden_states=hidden_states,
+            kv_cache=kv_cache,
+            attn_metadata=attn_metadata,
+        )
+
+        # Fully Connected
+        hidden_states, residual = self.post_attention_layernorm(
+            hidden_states, residual)
+        hidden_states = self.mlp(hidden_states)
+        return hidden_states, residual
+
+
+@support_torch_compile
+class BaiChuanModel(nn.Module):
+
+    def __init__(
+        self,
+        vllm_config: VllmConfig,
+        prefix: str = "",
+        position_embedding: str = "ROPE",
+    ) -> None:
+        super().__init__()
+
+        config = vllm_config.model_config.hf_config
+        cache_config = vllm_config.cache_config
+        quant_config = vllm_config.quant_config
+
+        self.config = config
+        self.padding_idx = config.pad_token_id
+        self.vocab_size = config.vocab_size
+
+        self.embed_tokens = VocabParallelEmbedding(
+            config.vocab_size,
+            config.hidden_size,
+        )
+        self.start_layer, self.end_layer, self.layers = make_layers(
+            config.num_hidden_layers,
+            lambda prefix: BaiChuanDecoderLayer(config,
+                                                position_embedding,
+                                                cache_config,
+                                                quant_config,
+                                                prefix=prefix),
+            prefix=f"{prefix}.layers",
+        )
+        self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.make_empty_intermediate_tensors = (
+            make_empty_intermediate_tensors_factory(
+                ["hidden_states", "residual"], config.hidden_size))
+
+    def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
+        return self.embed_tokens(input_ids)
+
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        positions: torch.Tensor,
+        kv_caches: List[torch.Tensor],
+        attn_metadata: AttentionMetadata,
+        intermediate_tensors: Optional[IntermediateTensors],
+        inputs_embeds: Optional[torch.Tensor] = None,
+    ) -> Union[torch.Tensor, IntermediateTensors]:
+        if get_pp_group().is_first_rank:
+            if inputs_embeds is not None:
+                hidden_states = inputs_embeds
+            else:
+                hidden_states = self.get_input_embeddings(input_ids)
+            residual = None
+        else:
+            assert intermediate_tensors is not None
+            hidden_states = intermediate_tensors["hidden_states"]
+            residual = intermediate_tensors["residual"]
+        for i in range(self.start_layer, self.end_layer):
+            layer = self.layers[i]
+            hidden_states, residual = layer(
+                positions,
+                hidden_states,
+                kv_caches[i - self.start_layer],
+                attn_metadata,
+                residual,
+            )
+        if not get_pp_group().is_last_rank:
+            return IntermediateTensors({
+                "hidden_states": hidden_states,
+                "residual": residual,
+            })
+        hidden_states, _ = self.norm(hidden_states, residual)
+        return hidden_states
+
+
+class BaiChuanBaseForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
+    packed_modules_mapping = {
+        "W_pack": ["W_pack"],
+        "gate_up_proj": [
+            "gate_proj",
+            "up_proj",
+        ],
+    }
+    # LoRA specific attributes
+    supported_lora_modules = [
+        "W_pack",
+        "o_proj",
+        "gate_up_proj",
+        "down_proj",
+    ]
+    embedding_modules = {}
+    embedding_padding_modules = []
+
+    def __init__(
+        self,
+        *,
+        vllm_config: VllmConfig,
+        prefix: str = "",
+        position_embedding: str = "ROPE",
+    ):
+        super().__init__()
+        config = vllm_config.model_config.hf_config
+        quant_config = vllm_config.quant_config
+        lora_config = vllm_config.lora_config
+        self.config = config
+        self.lora_config = lora_config
+
+        self.quant_config = quant_config
+        self.model = BaiChuanModel(vllm_config=vllm_config,
+                                   prefix=prefix,
+                                   position_embedding=position_embedding)
+        self.lm_head = ParallelLMHead(config.vocab_size,
+                                      config.hidden_size,
+                                      quant_config=quant_config)
+        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)
+
+    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,
+        kv_caches: List[torch.Tensor],
+        attn_metadata: AttentionMetadata,
+        intermediate_tensors: Optional[IntermediateTensors] = None,
+        inputs_embeds: Optional[torch.Tensor] = None,
+    ) -> Union[torch.Tensor, IntermediateTensors]:
+        hidden_states = self.model(input_ids, positions, kv_caches,
+                                   attn_metadata, intermediate_tensors,
+                                   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)
+        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)
+            ("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 name == "lm_head.weight":
+                # Unlike Baichuan, Baichuan2 normalizes the head weights.
+                # Refer to:
+                # https://huggingface.co/baichuan-inc/Baichuan2-7B-Chat/blob/84603cde5ebffb6084e476cfaeceaf0b8b91fe54/modeling_baichuan.py#L508
+                # Distinguish between Baichuan and Baichuan2 by checking the
+                # vocab size. This is suggested by
+                # https://github.com/vllm-project/vllm/pull/1022#discussion_r1325652704
+                is_baichuan2 = self.config.vocab_size == 125696
+                if is_baichuan2:
+                    loaded_weight = torch.nn.functional.normalize(
+                        loaded_weight)
+
+            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 BaichuanForCausalLM(BaiChuanBaseForCausalLM):
+    """Baichuan 13B and Baichuan2 7B/13B.
+    NOTE: the class name has a lower case 'c'.
+    """
+
+    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
+        config = vllm_config.model_config.hf_config
+        if config.hidden_size == 4096:  # baichuan2 7b
+            super().__init__(vllm_config=vllm_config,
+                             prefix=prefix,
+                             position_embedding="ROPE")
+        else:  # baichuan 13b, baichuan2 13b
+            super().__init__(vllm_config=vllm_config,
+                             prefix=prefix,
+                             position_embedding="ALIBI")
+
+
+class BaiChuanForCausalLM(BaiChuanBaseForCausalLM):
+    """Baichuan 7B.
+    NOTE: the class name has an upper case 'C'.
+    """
+
+    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
+        super().__init__(vllm_config=vllm_config,
+                         prefix=prefix,
+                         position_embedding="ROPE")

.venv/lib/python3.11/site-packages/vllm/model_executor/models/blip.py

@@ -0,0 +1,334 @@
+# SPDX-License-Identifier: Apache-2.0
+"""Minimal implementation of BlipVisionModel intended to be only used 
+within a vision language model."""
+from typing import Iterable, Optional, Set, Tuple, Union
+
+import torch
+import torch.nn as nn
+from transformers import Blip2VisionConfig, BlipVisionConfig
+
+from vllm.attention.layer import MultiHeadAttention
+from vllm.distributed import divide, get_tensor_model_parallel_world_size
+from vllm.model_executor.layers.activation import get_act_fn
+from vllm.model_executor.layers.linear import (ColumnParallelLinear,
+                                               QKVParallelLinear,
+                                               RowParallelLinear)
+from vllm.model_executor.layers.quantization import QuantizationConfig
+from vllm.model_executor.model_loader.weight_utils import default_weight_loader
+
+
+def get_blip_patch_grid_length(*, image_size: int, patch_size: int) -> int:
+    assert image_size % patch_size == 0
+    return image_size // patch_size
+
+
+def get_blip_num_patches(*, image_size: int, patch_size: int) -> int:
+    grid_length = get_blip_patch_grid_length(image_size=image_size,
+                                             patch_size=patch_size)
+    return grid_length * grid_length
+
+
+# Adapted from https://github.com/huggingface/transformers/blob/v4.39.0/src/transformers/models/blip/modeling_blip.py#L164 # noqa
+class BlipVisionEmbeddings(nn.Module):
+
+    def __init__(self, config: Union[BlipVisionConfig, Blip2VisionConfig]):
+        super().__init__()
+
+        self.config = config
+        self.embed_dim = config.hidden_size
+        self.image_size = config.image_size
+        self.patch_size = config.patch_size
+
+        self.class_embedding = nn.Parameter(torch.randn(1, 1, self.embed_dim))
+
+        self.patch_embedding = nn.Conv2d(
+            in_channels=3,
+            out_channels=self.embed_dim,
+            kernel_size=self.patch_size,
+            stride=self.patch_size,
+        )
+
+        self.num_patches = get_blip_num_patches(image_size=self.image_size,
+                                                patch_size=self.patch_size)
+        self.num_positions = self.num_patches + 1
+
+        self.position_embedding = nn.Parameter(
+            torch.randn(1, self.num_positions, self.embed_dim))
+
+    def forward(self, pixel_values: torch.Tensor) -> torch.Tensor:
+        batch_size = pixel_values.shape[0]
+        target_dtype = self.patch_embedding.weight.dtype
+        patch_embeds = self.patch_embedding(pixel_values.to(
+            dtype=target_dtype))  # shape = [*, width, grid, grid]
+        patch_embeds = patch_embeds.flatten(2).transpose(1, 2)
+
+        class_embeds = self.class_embedding.expand(batch_size, 1, -1)
+        embeddings = torch.cat([class_embeds, patch_embeds], dim=1)
+
+        position_embeds = self.position_embedding.to(target_dtype)
+        embeddings = embeddings + position_embeds[:, :embeddings.size(1), :]
+
+        return embeddings
+
+
+class BlipAttention(nn.Module):
+    """Multi-headed attention from 'Attention Is All You Need' paper"""
+
+    def __init__(
+        self,
+        config: Union[BlipVisionConfig, Blip2VisionConfig],
+        quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
+    ) -> None:
+        super().__init__()
+        self.config = config
+        self.embed_dim = config.hidden_size
+        self.num_heads = config.num_attention_heads
+        self.head_dim = self.embed_dim // self.num_heads
+        if self.head_dim * self.num_heads != self.embed_dim:
+            raise ValueError(
+                "embed_dim must be divisible by num_heads "
+                f"(got `embed_dim`: {self.embed_dim} and `num_heads`:"
+                f" {self.num_heads}).")
+        self.scale = self.head_dim**-0.5
+        self.dropout = config.attention_dropout
+
+        self.qkv = QKVParallelLinear(
+            self.embed_dim,
+            self.head_dim,
+            self.num_heads,
+            bias=config.qkv_bias,
+            quant_config=quant_config,
+            prefix=f"{prefix}.qkv",
+        )
+        self.projection = RowParallelLinear(
+            self.embed_dim,
+            self.embed_dim,
+            quant_config=quant_config,
+            prefix=f"{prefix}.projection",
+        )
+
+        self.tp_size = get_tensor_model_parallel_world_size()
+        self.num_heads_per_partition = divide(self.num_heads, self.tp_size)
+
+        self.attn = MultiHeadAttention(self.num_heads_per_partition,
+                                       self.head_dim, self.scale)
+
+    def _shape(self, tensor: torch.Tensor, seq_len: int, bsz: int):
+        return tensor.view(bsz, seq_len, self.num_heads,
+                           self.head_dim).transpose(1, 2).contiguous()
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+    ):
+        """Input shape: Batch x Time x Channel"""
+
+        qkv_states, _ = self.qkv(hidden_states)
+        query_states, key_states, value_states = qkv_states.chunk(3, dim=-1)
+        out = self.attn(query_states, key_states, value_states)
+        attn_output, _ = self.projection(out)
+
+        return attn_output, None
+
+
+class BlipMLP(nn.Module):
+
+    def __init__(
+        self,
+        config: BlipVisionConfig,
+        quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
+    ) -> None:
+        super().__init__()
+
+        self.config = config
+
+        self.activation_fn = get_act_fn(config.hidden_act)
+        self.fc1 = ColumnParallelLinear(config.hidden_size,
+                                        config.intermediate_size,
+                                        bias=True,
+                                        quant_config=quant_config,
+                                        prefix=f"{prefix}.fc1")
+        self.fc2 = RowParallelLinear(config.intermediate_size,
+                                     config.hidden_size,
+                                     bias=True,
+                                     quant_config=quant_config,
+                                     prefix=f"{prefix}.fc2")
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        hidden_states, _ = self.fc1(hidden_states)
+        hidden_states = self.activation_fn(hidden_states)
+        hidden_states, _ = self.fc2(hidden_states)
+
+        return hidden_states
+
+
+class BlipEncoderLayer(nn.Module):
+
+    def __init__(
+        self,
+        config: BlipVisionConfig,
+        quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
+    ) -> None:
+        super().__init__()
+
+        # fallback to sdpa attention if tp unavailable
+        self.self_attn = BlipAttention(
+            config,
+            quant_config=quant_config,
+            prefix=f"{prefix}.self_attn",
+        )
+        self.layer_norm1 = nn.LayerNorm(config.hidden_size,
+                                        eps=config.layer_norm_eps)
+        self.mlp = BlipMLP(config,
+                           quant_config=quant_config,
+                           prefix=f"{prefix}.mlp")
+        self.layer_norm2 = nn.LayerNorm(config.hidden_size,
+                                        eps=config.layer_norm_eps)
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        residual = hidden_states
+
+        hidden_states = self.layer_norm1(hidden_states)
+        hidden_states, _ = self.self_attn(hidden_states=hidden_states)
+        hidden_states = residual + hidden_states
+
+        residual = hidden_states
+        hidden_states = self.layer_norm2(hidden_states)
+        hidden_states = self.mlp(hidden_states)
+        hidden_states = residual + hidden_states
+
+        return hidden_states
+
+
+class BlipEncoder(nn.Module):
+    """
+    Transformer encoder consisting of `config.num_hidden_layers` self 
+    attention layers. Each layer is a [`BlipEncoderLayer`].
+
+    Args:
+        config: BlipConfig
+    """
+
+    def __init__(
+        self,
+        config: BlipVisionConfig,
+        quant_config: Optional[QuantizationConfig] = None,
+        num_hidden_layers_override: Optional[int] = None,
+        prefix: str = "",
+    ) -> None:
+        super().__init__()
+
+        self.config = config
+
+        if num_hidden_layers_override is None:
+            num_hidden_layers = config.num_hidden_layers
+        else:
+            num_hidden_layers = num_hidden_layers_override
+
+        self.layers = nn.ModuleList([
+            BlipEncoderLayer(config=config,
+                             quant_config=quant_config,
+                             prefix=f"{prefix}.layers.{layer_idx}")
+            for layer_idx in range(num_hidden_layers)
+        ])
+
+    def forward(self, inputs_embeds: torch.Tensor):
+        hidden_states = inputs_embeds
+        for encoder_layer in self.layers:
+            hidden_states = encoder_layer(hidden_states)
+
+        return hidden_states
+
+
+class BlipVisionModel(nn.Module):
+    config_class = BlipVisionConfig
+    main_input_name = "pixel_values"
+
+    def __init__(
+        self,
+        config: BlipVisionConfig,
+        quant_config: Optional[QuantizationConfig] = None,
+        *,
+        num_hidden_layers_override: Optional[int] = None,
+        require_post_norm: Optional[bool] = None,
+        prefix: str = "",
+    ) -> None:
+        super().__init__()
+        self.config = config
+
+        self.embeddings = BlipVisionEmbeddings(config)
+        self.encoder = BlipEncoder(
+            config=config,
+            quant_config=quant_config,
+            num_hidden_layers_override=num_hidden_layers_override,
+            prefix=f"{prefix}.encoder",
+        )
+
+        num_hidden_layers = config.num_hidden_layers
+        if len(self.encoder.layers) > config.num_hidden_layers:
+            raise ValueError(
+                f"The original encoder only has {num_hidden_layers} "
+                f"layers, but you requested {len(self.encoder.layers)} layers."
+            )
+
+        # If possible, skip post_layernorm to conserve memory
+        if require_post_norm is None:
+            require_post_norm = len(self.encoder.layers) == num_hidden_layers
+
+        if require_post_norm:
+            self.post_layernorm = nn.LayerNorm(config.hidden_size,
+                                               eps=config.layer_norm_eps)
+        else:
+            self.post_layernorm = None
+
+    def forward(self, pixel_values: torch.Tensor) -> torch.Tensor:
+        hidden_states = self.embeddings(pixel_values)
+        hidden_states = self.encoder(inputs_embeds=hidden_states)
+
+        if self.post_layernorm is None:
+            return hidden_states
+
+        return self.post_layernorm(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"),
+        ]
+        params_dict = dict(self.named_parameters())
+        loaded_params: Set[str] = set()
+        layer_count = len(self.encoder.layers)
+
+        for name, loaded_weight in weights:
+            # post_layernorm is not needed in BlipVisionModel
+            if (name.startswith("post_layernorm")
+                    and self.post_layernorm is None):
+                continue
+
+            # omit layers when num_hidden_layers_override is set
+            if name.startswith("encoder.layers"):
+                layer_idx = int(name.split(".")[2])
+                if layer_idx >= layer_count:
+                    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)
+                param = params_dict[name]
+                weight_loader = param.weight_loader
+                weight_loader(param, loaded_weight, shard_id)
+                break
+            else:
+                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

.venv/lib/python3.11/site-packages/vllm/model_executor/models/clip.py

@@ -0,0 +1,545 @@
+# SPDX-License-Identifier: Apache-2.0
+"""Minimal implementation of CLIPVisionModel intended to be only used
+within a vision language model."""
+from typing import Iterable, List, Optional, Set, Tuple, Union
+
+import numpy as np
+import torch
+import torch.nn as nn
+from PIL import Image
+from transformers import CLIPVisionConfig
+
+from vllm.attention.layer import MultiHeadAttention
+from vllm.config import ModelConfig
+from vllm.distributed import divide, get_tensor_model_parallel_world_size
+from vllm.inputs import DecoderOnlyInputs, token_inputs
+from vllm.model_executor.layers.activation import get_act_fn
+from vllm.model_executor.layers.linear import (ColumnParallelLinear,
+                                               QKVParallelLinear,
+                                               RowParallelLinear)
+from vllm.model_executor.layers.quantization import QuantizationConfig
+from vllm.model_executor.model_loader.weight_utils import default_weight_loader
+from vllm.multimodal.utils import (cached_get_tokenizer,
+                                   consecutive_placeholder_ranges,
+                                   repeat_and_pad_placeholder_tokens)
+from vllm.sequence import SequenceData
+
+from .vision import VisionEncoderInfo, resolve_visual_encoder_outputs
+
+
+def get_clip_patch_grid_length(*, image_size: int, patch_size: int) -> int:
+    assert image_size % patch_size == 0
+    return image_size // patch_size
+
+
+def get_clip_num_patches(*, image_size: int, patch_size: int) -> int:
+    grid_length = get_clip_patch_grid_length(image_size=image_size,
+                                             patch_size=patch_size)
+    return grid_length * grid_length
+
+
+def get_clip_image_feature_size(hf_config: CLIPVisionConfig) -> int:
+    return get_clip_num_patches(image_size=hf_config.image_size,
+                                patch_size=hf_config.patch_size) + 1
+
+
+def get_max_clip_image_tokens(hf_config: CLIPVisionConfig) -> int:
+    return get_clip_image_feature_size(hf_config)
+
+
+def dummy_seq_data_for_clip(hf_config: CLIPVisionConfig,
+                            seq_len: int,
+                            num_images: int,
+                            *,
+                            image_token_id: int,
+                            image_feature_size_override: Optional[int] = None,
+                            mm_key: str = "image"):
+    if image_feature_size_override is None:
+        image_feature_size = get_clip_image_feature_size(hf_config)
+    else:
+        image_feature_size = image_feature_size_override
+
+    return SequenceData.from_prompt_token_counts(
+        (image_token_id, image_feature_size * num_images),
+        (0, seq_len - image_feature_size * num_images),
+    ), {
+        mm_key:
+        consecutive_placeholder_ranges(num_items=num_images,
+                                       item_size=image_feature_size)
+    }
+
+
+def dummy_image_for_clip(
+    hf_config: CLIPVisionConfig,
+    num_images: int,
+    *,
+    image_width_override: Optional[int] = None,
+    image_height_override: Optional[int] = None,
+):
+    width = height = hf_config.image_size
+    if image_width_override is not None:
+        width = image_width_override
+    if image_height_override is not None:
+        height = image_height_override
+
+    image = Image.new("RGB", (width, height), color=0)
+    return {"image": image if num_images == 1 else [image] * num_images}
+
+
+def dummy_video_for_clip(
+    hf_config: CLIPVisionConfig,
+    num_frames: int,
+    num_videos: int = 1,
+    *,
+    image_width_override: Optional[int] = None,
+    image_height_override: Optional[int] = None,
+):
+    pil_frame = dummy_image_for_clip(
+        hf_config,
+        num_images=1,
+        image_width_override=image_width_override,
+        image_height_override=image_height_override)
+    np_frame = np.array(pil_frame["image"])
+    mm_data_per_video = np.repeat([np_frame], num_frames, axis=0)
+    video_data = [mm_data_per_video] * num_videos
+    mm_data = {"video": video_data}
+    return mm_data
+
+
+def input_processor_for_clip(
+    model_config: ModelConfig,
+    hf_config: CLIPVisionConfig,
+    inputs: DecoderOnlyInputs,
+    *,
+    image_token_id: int,
+    image_feature_size_override: Optional[Union[int, List[int]]] = None,
+):
+    multi_modal_data = inputs.get("multi_modal_data")
+    if multi_modal_data is None or "image" not in multi_modal_data:
+        return inputs
+
+    if "multi_modal_placeholders" in inputs and "image" in inputs[
+            "multi_modal_placeholders"]:
+        # The inputs already have placeholders.
+        return inputs
+
+    tokenizer = cached_get_tokenizer(model_config.tokenizer)
+
+    if image_feature_size_override is None:
+        image_data = multi_modal_data["image"]
+        if isinstance(image_data, Image.Image):
+            image_feature_size = get_clip_image_feature_size(hf_config)
+        elif isinstance(image_data, torch.Tensor):
+            num_images, image_feature_size, hidden_size = image_data.shape
+        else:
+            raise TypeError(f"Invalid image type: {type(image_data)}")
+    else:
+        image_feature_size = image_feature_size_override
+
+    new_prompt, new_token_ids, ranges = repeat_and_pad_placeholder_tokens(
+        tokenizer,
+        inputs.get("prompt"),
+        inputs["prompt_token_ids"],
+        placeholder_token_id=image_token_id,
+        repeat_count=image_feature_size,
+    )
+
+    # NOTE: Create a defensive copy of the original inputs
+    return token_inputs(prompt_token_ids=new_token_ids,
+                        prompt=new_prompt,
+                        multi_modal_data=multi_modal_data,
+                        multi_modal_placeholders={"image": ranges})
+
+
+class CLIPEncoderInfo(VisionEncoderInfo[CLIPVisionConfig]):
+
+    def get_num_image_tokens(
+        self,
+        *,
+        image_width: int,
+        image_height: int,
+    ) -> int:
+        return get_clip_image_feature_size(self.vision_config)
+
+    def get_max_image_tokens(self) -> int:
+        return get_max_clip_image_tokens(self.vision_config)
+
+    def get_image_size(self) -> int:
+        return self.vision_config.image_size
+
+    def get_patch_size(self) -> int:
+        return self.vision_config.patch_size
+
+    def get_patch_grid_length(self) -> int:
+        return get_clip_patch_grid_length(
+            image_size=self.vision_config.image_size,
+            patch_size=self.vision_config.patch_size,
+        )
+
+
+# Adapted from https://github.com/huggingface/transformers/blob/v4.39.0/src/transformers/models/clip/modeling_clip.py#L164 # noqa
+class CLIPVisionEmbeddings(nn.Module):
+
+    def __init__(self, config: CLIPVisionConfig):
+        super().__init__()
+        self.config = config
+        self.embed_dim = config.hidden_size
+        self.image_size = config.image_size
+        self.patch_size = config.patch_size
+
+        self.class_embedding = nn.Parameter(torch.randn(self.embed_dim))
+
+        self.patch_embedding = nn.Conv2d(
+            in_channels=config.num_channels,
+            out_channels=self.embed_dim,
+            kernel_size=self.patch_size,
+            stride=self.patch_size,
+            bias=False,
+        )
+
+        self.num_patches = get_clip_num_patches(image_size=self.image_size,
+                                                patch_size=self.patch_size)
+        self.num_positions = self.num_patches + 1
+        self.position_embedding = nn.Embedding(self.num_positions,
+                                               self.embed_dim)
+        self.register_buffer("position_ids",
+                             torch.arange(self.num_positions).expand((1, -1)),
+                             persistent=False)
+
+    def forward(self, pixel_values: torch.Tensor) -> torch.Tensor:
+        batch_size = pixel_values.shape[0]
+        target_dtype = self.patch_embedding.weight.dtype
+        patch_embeds = self.patch_embedding(pixel_values.to(
+            dtype=target_dtype))  # shape = [*, width, grid, grid]
+        patch_embeds = patch_embeds.flatten(2).transpose(1, 2)
+
+        class_embeds = self.class_embedding.expand(batch_size, 1, -1)
+        embeddings = torch.cat([class_embeds, patch_embeds], dim=1)
+        embeddings = embeddings + self.position_embedding(self.position_ids)
+
+        return embeddings
+
+
+class CLIPAttention(nn.Module):
+    """Multi-headed attention from 'Attention Is All You Need' paper"""
+
+    def __init__(
+        self,
+        config: CLIPVisionConfig,
+        quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
+    ):
+        super().__init__()
+        self.config = config
+        self.embed_dim = config.hidden_size
+        self.num_heads = config.num_attention_heads
+        self.head_dim = self.embed_dim // self.num_heads
+        if self.head_dim * self.num_heads != self.embed_dim:
+            raise ValueError(
+                "embed_dim must be divisible by num_heads "
+                f"(got `embed_dim`: {self.embed_dim} and `num_heads`:"
+                f" {self.num_heads}).")
+        self.scale = self.head_dim**-0.5
+        self.dropout = config.attention_dropout
+
+        self.qkv_proj = QKVParallelLinear(
+            hidden_size=self.embed_dim,
+            head_size=self.head_dim,
+            total_num_heads=self.num_heads,
+            quant_config=quant_config,
+            prefix=f"{prefix}.qkv_proj",
+        )
+
+        self.out_proj = RowParallelLinear(
+            input_size=self.embed_dim,
+            output_size=self.embed_dim,
+            quant_config=quant_config,
+            prefix=f"{prefix}.out_proj",
+        )
+
+        self.tp_size = get_tensor_model_parallel_world_size()
+        self.num_heads_per_partition = divide(self.num_heads, self.tp_size)
+
+        self.attn = MultiHeadAttention(self.num_heads_per_partition,
+                                       self.head_dim, self.scale)
+
+    def _shape(self, tensor: torch.Tensor, seq_len: int, bsz: int):
+        return tensor.view(bsz, seq_len, self.num_heads,
+                           self.head_dim).transpose(1, 2).contiguous()
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+    ):
+        """Input shape: Batch x Time x Channel"""
+
+        qkv_states, _ = self.qkv_proj(hidden_states)
+        query_states, key_states, value_states = qkv_states.chunk(3, dim=-1)
+        out = self.attn(query_states, key_states, value_states)
+        attn_output, _ = self.out_proj(out)
+
+        return attn_output, None
+
+
+class CLIPMLP(nn.Module):
+
+    def __init__(
+        self,
+        config: CLIPVisionConfig,
+        quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
+    ) -> None:
+        super().__init__()
+        self.config = config
+        self.activation_fn = get_act_fn(config.hidden_act)
+        self.fc1 = ColumnParallelLinear(config.hidden_size,
+                                        config.intermediate_size,
+                                        bias=True,
+                                        quant_config=quant_config,
+                                        prefix=f"{prefix}.fc1")
+        self.fc2 = RowParallelLinear(config.intermediate_size,
+                                     config.hidden_size,
+                                     bias=True,
+                                     quant_config=quant_config,
+                                     prefix=f"{prefix}.fc2")
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        hidden_states, _ = self.fc1(hidden_states)
+        hidden_states = self.activation_fn(hidden_states)
+        hidden_states, _ = self.fc2(hidden_states)
+
+        return hidden_states
+
+
+class CLIPEncoderLayer(nn.Module):
+
+    def __init__(
+        self,
+        config: CLIPVisionConfig,
+        quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
+    ) -> None:
+        super().__init__()
+        self.self_attn = CLIPAttention(
+            config,
+            quant_config=quant_config,
+            prefix=f"{prefix}.self_attn",
+        )
+        self.layer_norm1 = nn.LayerNorm(config.hidden_size,
+                                        eps=config.layer_norm_eps)
+        self.mlp = CLIPMLP(config,
+                           quant_config=quant_config,
+                           prefix=f"{prefix}.mlp")
+        self.layer_norm2 = nn.LayerNorm(config.hidden_size,
+                                        eps=config.layer_norm_eps)
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+
+        residual = hidden_states
+
+        hidden_states = self.layer_norm1(hidden_states)
+        hidden_states, _ = self.self_attn(hidden_states=hidden_states)
+        hidden_states = residual + hidden_states
+
+        residual = hidden_states
+        hidden_states = self.layer_norm2(hidden_states)
+        hidden_states = self.mlp(hidden_states)
+        hidden_states = residual + hidden_states
+
+        return hidden_states
+
+
+class CLIPEncoder(nn.Module):
+    """
+    Transformer encoder consisting of `config.num_hidden_layers` self
+    attention layers. Each layer is a [`CLIPEncoderLayer`].
+
+    Args:
+        config: CLIPConfig
+    """
+
+    def __init__(
+        self,
+        config: CLIPVisionConfig,
+        quant_config: Optional[QuantizationConfig] = None,
+        num_hidden_layers_override: Optional[int] = None,
+        prefix: str = "",
+    ) -> None:
+        super().__init__()
+
+        self.config = config
+
+        if num_hidden_layers_override is None:
+            num_hidden_layers = config.num_hidden_layers
+        else:
+            num_hidden_layers = num_hidden_layers_override
+        self.layers = nn.ModuleList([
+            CLIPEncoderLayer(config=config,
+                             quant_config=quant_config,
+                             prefix=f"{prefix}.layers.{layer_idx}")
+            for layer_idx in range(num_hidden_layers)
+        ])
+
+    def forward(
+        self, inputs_embeds: torch.Tensor, return_all_hidden_states: bool
+    ) -> Union[torch.Tensor, list[torch.Tensor]]:
+        hidden_states_pool = []
+        hidden_states = inputs_embeds
+
+        for encoder_layer in self.layers:
+            hidden_states = encoder_layer(hidden_states)
+            if return_all_hidden_states:
+                hidden_states_pool.append(hidden_states)
+        # If we have multiple feature sample layers, we return all hidden
+        # states in order and grab the ones we need by index.
+        if return_all_hidden_states:
+            return hidden_states_pool
+        return hidden_states
+
+
+class CLIPVisionTransformer(nn.Module):
+
+    def __init__(
+        self,
+        config: CLIPVisionConfig,
+        quant_config: Optional[QuantizationConfig] = None,
+        *,
+        num_hidden_layers_override: Optional[int] = None,
+        require_post_norm: Optional[bool] = None,
+        prefix: str = "",
+    ) -> None:
+        super().__init__()
+
+        self.config = config
+        embed_dim = config.hidden_size
+
+        self.embeddings = CLIPVisionEmbeddings(config)
+
+        # NOTE: This typo of "layrnorm" is not fixed on purpose to match
+        # the original transformers code and name of the model weights.
+        self.pre_layrnorm = nn.LayerNorm(embed_dim, eps=config.layer_norm_eps)
+
+        self.encoder = CLIPEncoder(
+            config=config,
+            quant_config=quant_config,
+            num_hidden_layers_override=num_hidden_layers_override,
+            prefix=f"{prefix}.encoder",
+        )
+
+        num_hidden_layers = config.num_hidden_layers
+        if len(self.encoder.layers) > config.num_hidden_layers:
+            raise ValueError(
+                f"The original encoder only has {num_hidden_layers} "
+                f"layers, but you requested {len(self.encoder.layers)} layers."
+            )
+
+        # If possible, skip post_layernorm to conserve memory
+        if require_post_norm is None:
+            require_post_norm = len(self.encoder.layers) == num_hidden_layers
+
+        if require_post_norm:
+            self.post_layernorm = nn.LayerNorm(embed_dim,
+                                               eps=config.layer_norm_eps)
+        else:
+            self.post_layernorm = None
+
+    def forward(
+        self,
+        pixel_values: torch.Tensor,
+        feature_sample_layers: Optional[list[int]] = None,
+    ) -> torch.Tensor:
+
+        hidden_states = self.embeddings(pixel_values)
+        hidden_states = self.pre_layrnorm(hidden_states)
+
+        return_all_hidden_states = feature_sample_layers is not None
+
+        # Produces either the last layer output or all of the hidden states,
+        # depending on if we have feature_sample_layers or not
+        encoder_outputs = self.encoder(
+            inputs_embeds=hidden_states,
+            return_all_hidden_states=return_all_hidden_states)
+
+        # Handle post-norm (if applicable) and stacks feature layers if needed
+        encoder_outputs = resolve_visual_encoder_outputs(
+            encoder_outputs, feature_sample_layers, self.post_layernorm,
+            self.config.num_hidden_layers)
+
+        return encoder_outputs
+
+
+class CLIPVisionModel(nn.Module):
+
+    config_class = CLIPVisionConfig
+    main_input_name = "pixel_values"
+
+    def __init__(
+        self,
+        config: CLIPVisionConfig,
+        quant_config: Optional[QuantizationConfig] = None,
+        *,
+        num_hidden_layers_override: Optional[int] = None,
+        require_post_norm: Optional[bool] = None,
+        prefix: str = "",
+    ) -> None:
+        super().__init__()
+        self.vision_model = CLIPVisionTransformer(
+            config=config,
+            quant_config=quant_config,
+            num_hidden_layers_override=num_hidden_layers_override,
+            require_post_norm=require_post_norm,
+            prefix=f"{prefix}.vision_model")
+
+    def forward(
+        self,
+        pixel_values: torch.Tensor,
+        feature_sample_layers: Optional[list[int]] = None,
+    ) -> torch.Tensor:
+        return self.vision_model(pixel_values, feature_sample_layers)
+
+    @property
+    def device(self):
+        return next(self.parameters()).device
+
+    # (TODO) Add prefix argument for filtering out weights to be loaded
+    #        ref: https://github.com/vllm-project/vllm/pull/7186#discussion_r1734163986
+    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())
+        loaded_params: Set[str] = set()
+        layer_count = len(self.vision_model.encoder.layers)
+
+        for name, loaded_weight in weights:
+            # post_layernorm is not needed in CLIPVisionModel
+            if (name.startswith("vision_model.post_layernorm")
+                    and self.vision_model.post_layernorm is None):
+                continue
+
+            # omit layers when num_hidden_layers_override is set
+            if name.startswith("vision_model.encoder.layers"):
+                layer_idx = int(name.split(".")[3])
+                if layer_idx >= layer_count:
+                    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)
+
+                param = params_dict[name]
+                weight_loader = param.weight_loader
+                weight_loader(param, loaded_weight, shard_id)
+                break
+            else:
+                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

.venv/lib/python3.11/site-packages/vllm/model_executor/models/commandr.py

@@ -0,0 +1,488 @@
+# SPDX-License-Identifier: Apache-2.0
+
+# Copyright 2024 Cohere and the HuggingFace Inc. team. All rights reserved.
+#
+# This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX
+# and OPT implementations in this library. It has been modified from its
+# original forms to accommodate minor architectural differences compared
+# to GPT-NeoX and OPT used by the Meta AI team that trained the model.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# This file is based on the LLama model definition file in transformers
+"""PyTorch Cohere model."""
+from typing import Iterable, List, Optional, Set, Tuple, Union
+
+import torch
+import torch.utils.checkpoint
+from torch import nn
+from transformers import CohereConfig
+
+from vllm.attention import Attention, AttentionMetadata
+from vllm.compilation.decorators import support_torch_compile
+from vllm.config import CacheConfig, VllmConfig
+from vllm.distributed import get_pp_group, get_tensor_model_parallel_world_size
+from vllm.model_executor.layers.activation import SiluAndMul
+from vllm.model_executor.layers.linear import (MergedColumnParallelLinear,
+                                               QKVParallelLinear,
+                                               RowParallelLinear)
+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 (
+    VocabParallelEmbedding)
+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
+
+from .interfaces import SupportsLoRA, SupportsPP
+from .utils import (extract_layer_index, is_pp_missing_parameter,
+                    make_empty_intermediate_tensors_factory, make_layers,
+                    maybe_prefix)
+
+
+@torch.compile(backend=current_platform.simple_compile_backend)
+def layer_norm_func(hidden_states, weight, variance_epsilon):
+    input_dtype = hidden_states.dtype
+    hidden_states = hidden_states.to(torch.float32)
+    mean = hidden_states.mean(-1, keepdim=True)
+    variance = (hidden_states - mean).pow(2).mean(-1, keepdim=True)
+    hidden_states = (hidden_states - mean) * torch.rsqrt(variance +
+                                                         variance_epsilon)
+    hidden_states = weight.to(torch.float32) * hidden_states
+    return hidden_states.to(input_dtype)
+
+
+class LayerNorm(nn.Module):
+
+    def __init__(self, param_shape=None, eps=1e-5):
+        super().__init__()
+        self.weight = nn.Parameter(torch.ones(param_shape))
+        self.variance_epsilon = eps
+        set_weight_attrs(self.weight,
+                         {"weight_loader": row_parallel_weight_loader})
+
+    def forward(self, hidden_states, residuals=None):
+        hidden_states = layer_norm_func(hidden_states, self.weight,
+                                        self.variance_epsilon)
+        return hidden_states, residuals
+
+
+# Copied from transformers.models.llama.modeling_llama.LlamaMLP Llama->Cohere
+class CohereMLP(nn.Module):
+
+    def __init__(
+        self,
+        config: CohereConfig,
+        quant_config: Optional[QuantizationConfig] = None,
+    ):
+        super().__init__()
+        self.config = config
+        self.hidden_size = config.hidden_size
+        self.intermediate_size = config.intermediate_size
+        self.gate_up_proj = MergedColumnParallelLinear(
+            self.hidden_size,
+            [self.intermediate_size] * 2,
+            bias=False,
+            quant_config=quant_config,
+        )
+        self.down_proj = RowParallelLinear(
+            self.intermediate_size,
+            self.hidden_size,
+            bias=False,
+            quant_config=quant_config,
+        )
+        self.act_fn = SiluAndMul()
+
+    def forward(self, x):
+        gate_up, _ = self.gate_up_proj(x)
+        x = self.act_fn(gate_up)
+        x, _ = self.down_proj(x)
+        return x
+
+
+class CohereAttention(nn.Module):
+
+    def __init__(
+        self,
+        config: CohereConfig,
+        cache_config: Optional[CacheConfig] = None,
+        quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
+    ):
+        super().__init__()
+        tp_size = get_tensor_model_parallel_world_size()
+        self.config = config
+        self.attention_dropout = config.attention_dropout
+        self.hidden_size = config.hidden_size
+        self.total_num_heads = config.num_attention_heads
+        self.num_heads = self.total_num_heads // tp_size
+        self.head_dim = self.hidden_size // self.total_num_heads
+        self.total_num_kv_heads = config.num_key_value_heads
+        if self.total_num_kv_heads >= tp_size:
+            # Number of KV heads is greater than TP size, so we partition
+            # the KV heads across multiple tensor parallel GPUs.
+            assert self.total_num_kv_heads % tp_size == 0
+        else:
+            # Number of KV heads is less than TP size, so we replicate
+            # the KV heads across multiple tensor parallel GPUs.
+            assert tp_size % self.total_num_kv_heads == 0
+        self.num_kv_heads = max(1, self.total_num_kv_heads // tp_size)
+        self.q_size = self.num_heads * self.head_dim
+        self.kv_size = self.num_kv_heads * self.head_dim
+        self.scaling = self.head_dim**-0.5
+        self.max_position_embeddings = getattr(
+            config, "model_max_length", None) or getattr(
+                config, "max_position_embeddings", 8192)
+        self.rope_theta = config.rope_theta
+        self.rope_scaling = getattr(config, "rope_scaling", None)
+        self.use_qk_norm = getattr(config, "use_qk_norm", False)
+        self.qkv_proj = QKVParallelLinear(
+            self.hidden_size,
+            self.head_dim,
+            self.total_num_heads,
+            self.total_num_kv_heads,
+            bias=False,
+            quant_config=quant_config,
+        )
+        self.o_proj = RowParallelLinear(
+            self.total_num_heads * self.head_dim,
+            self.hidden_size,
+            bias=False,
+            quant_config=quant_config,
+        )
+        self.rotary_emb = get_rope(
+            self.head_dim,
+            rotary_dim=self.head_dim,
+            max_position=self.max_position_embeddings,
+            base=self.rope_theta,
+            rope_scaling=self.rope_scaling,
+            is_neox_style=False,
+        )
+
+        # Model v2 has interleaved sliding windows, v1 does not
+        interleaved_sliding_window = getattr(config,
+                                             "interleaved_sliding_window",
+                                             None)
+        self.v1 = interleaved_sliding_window is None
+
+        layer_idx = extract_layer_index(prefix)
+        layer_has_sliding_window = (
+            getattr(config, "sliding_window_pattern", False)
+            and (layer_idx + 1) % self.config.sliding_window_pattern != 0)
+
+        self.sliding_window = (interleaved_sliding_window
+                               if layer_has_sliding_window else None)
+
+        self.attn = Attention(self.num_heads,
+                              self.head_dim,
+                              self.scaling,
+                              num_kv_heads=self.num_kv_heads,
+                              cache_config=cache_config,
+                              quant_config=quant_config,
+                              per_layer_sliding_window=self.sliding_window,
+                              prefix=f"{prefix}.attn")
+        if self.use_qk_norm:
+            self.q_norm = LayerNorm(param_shape=(self.num_heads,
+                                                 self.head_dim),
+                                    eps=config.layer_norm_eps)
+            self.k_norm = LayerNorm(param_shape=(self.num_kv_heads,
+                                                 self.head_dim),
+                                    eps=config.layer_norm_eps)
+
+    def _apply_qk_norm(self, q, k):
+        q = q.view(*q.shape[:-1], -1, self.head_dim)
+        k = k.view(*k.shape[:-1], -1, self.head_dim)
+        q, _ = self.q_norm(q)
+        k, _ = self.k_norm(k)
+        q = q.view(*q.shape[:-2], -1)
+        k = k.view(*k.shape[:-2], -1)
+        return q, k
+
+    def forward(
+        self,
+        positions: torch.Tensor,
+        hidden_states: torch.Tensor,
+        kv_cache: torch.Tensor,
+        attn_metadata: AttentionMetadata,
+    ) -> torch.Tensor:
+        qkv, _ = self.qkv_proj(hidden_states)
+        q, k, v = qkv.split([self.q_size, self.kv_size, self.kv_size], dim=-1)
+        if self.use_qk_norm:
+            q, k = self._apply_qk_norm(q, k)
+        if self.v1 or self.sliding_window:
+            q, k = self.rotary_emb(positions, q, k)
+        attn_output = self.attn(q, k, v, kv_cache, attn_metadata)
+        output, _ = self.o_proj(attn_output)
+        return output
+
+
+class CohereDecoderLayer(nn.Module):
+
+    def __init__(self,
+                 config: CohereConfig,
+                 cache_config: Optional[CacheConfig] = None,
+                 quant_config: Optional[QuantizationConfig] = None,
+                 prefix: str = ""):
+        super().__init__()
+        self.hidden_size = config.hidden_size
+
+        self.self_attn = CohereAttention(config,
+                                         cache_config,
+                                         quant_config=quant_config,
+                                         prefix=f"{prefix}.self_attn")
+
+        self.mlp = CohereMLP(config, quant_config=quant_config)
+        self.input_layernorm = LayerNorm(param_shape=(config.hidden_size),
+                                         eps=config.layer_norm_eps)
+
+    def forward(
+        self,
+        positions: torch.Tensor,
+        hidden_states: torch.Tensor,
+        kv_cache: torch.Tensor,
+        attn_metadata: AttentionMetadata,
+        residual: Optional[torch.Tensor],
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        # Self Attention
+        residual = hidden_states
+        hidden_states, residual = self.input_layernorm(hidden_states, residual)
+        hidden_states_attention = self.self_attn(
+            positions=positions,
+            hidden_states=hidden_states,
+            kv_cache=kv_cache,
+            attn_metadata=attn_metadata,
+        )
+        hidden_states_mlp = self.mlp(hidden_states)
+        # Add everything together
+        hidden_states = residual + hidden_states_attention + hidden_states_mlp
+
+        return hidden_states, residual
+
+
+@support_torch_compile
+class CohereModel(nn.Module):
+
+    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
+        super().__init__()
+
+        config = vllm_config.model_config.hf_config
+        cache_config = vllm_config.cache_config
+        quant_config = vllm_config.quant_config
+        lora_config = vllm_config.lora_config
+
+        self.config = config
+        lora_vocab = (lora_config.lora_extra_vocab_size *
+                      (lora_config.max_loras or 1)) if lora_config else 0
+        self.vocab_size = config.vocab_size + lora_vocab
+        self.org_vocab_size = config.vocab_size
+        self.embed_tokens = VocabParallelEmbedding(config.vocab_size,
+                                                   config.hidden_size)
+        self.start_layer, self.end_layer, self.layers = make_layers(
+            config.num_hidden_layers,
+            lambda prefix: CohereDecoderLayer(
+                config, cache_config, quant_config, prefix=prefix),
+            prefix=f"{prefix}.layers")
+        self.norm = LayerNorm(param_shape=(config.hidden_size),
+                              eps=config.layer_norm_eps)
+        self.make_empty_intermediate_tensors = (
+            make_empty_intermediate_tensors_factory(
+                ["hidden_states", "residual"], config.hidden_size))
+
+    def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
+        return self.embed_tokens(input_ids)
+
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        positions: torch.Tensor,
+        kv_caches: List[torch.Tensor],
+        attn_metadata: AttentionMetadata,
+        intermediate_tensors: Optional[IntermediateTensors],
+        inputs_embeds: Optional[torch.Tensor] = None,
+    ) -> Union[torch.Tensor, IntermediateTensors]:
+        if get_pp_group().is_first_rank:
+            if inputs_embeds is not None:
+                hidden_states = inputs_embeds
+            else:
+                hidden_states = self.get_input_embeddings(input_ids)
+            residual = None
+        else:
+            assert intermediate_tensors is not None
+            hidden_states = intermediate_tensors["hidden_states"]
+            residual = intermediate_tensors["residual"]
+        for i in range(self.start_layer, self.end_layer):
+            layer = self.layers[i]
+            hidden_states, residual = layer(
+                positions,
+                hidden_states,
+                kv_caches[i - self.start_layer],
+                attn_metadata,
+                residual,
+            )
+        if not get_pp_group().is_last_rank:
+            return IntermediateTensors({
+                "hidden_states": hidden_states,
+                "residual": residual
+            })
+        hidden_states, _ = self.norm(hidden_states, residual)
+        return hidden_states
+
+
+class CohereForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
+    packed_modules_mapping = {
+        "qkv_proj": [
+            "q_proj",
+            "k_proj",
+            "v_proj",
+        ],
+        "gate_up_proj": [
+            "gate_proj",
+            "up_proj",
+        ],
+    }
+    # LoRA specific attributes
+    supported_lora_modules = [
+        "qkv_proj", "o_proj", "gate_up_proj", "down_proj", "embed_tokens"
+    ]
+    embedding_modules = {"embed_tokens": "input_embeddings"}
+    embedding_padding_modules = []
+
+    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
+        super().__init__()
+        config = vllm_config.model_config.hf_config
+        quant_config = vllm_config.quant_config
+        lora_config = vllm_config.lora_config
+        self.config = config
+        # currently all existing command R models have `tie_word_embeddings`
+        # enabled
+        assert config.tie_word_embeddings
+        self.unpadded_vocab_size = config.vocab_size
+        if lora_config:
+            self.unpadded_vocab_size += lora_config.lora_extra_vocab_size
+        self.quant_config = quant_config
+        self.logits_processor = LogitsProcessor(self.unpadded_vocab_size,
+                                                config.vocab_size,
+                                                scale=config.logit_scale)
+        self.model = CohereModel(vllm_config=vllm_config,
+                                 prefix=maybe_prefix(prefix, "model"))
+        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)
+
+    @torch.no_grad()
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        positions: torch.Tensor,
+        kv_caches: List[torch.Tensor],
+        attn_metadata: AttentionMetadata,
+        intermediate_tensors: Optional[IntermediateTensors] = None,
+        inputs_embeds: Optional[torch.Tensor] = None,
+    ) -> Union[torch.Tensor, IntermediateTensors]:
+        hidden_states = self.model(input_ids, positions, kv_caches,
+                                   attn_metadata, intermediate_tensors,
+                                   inputs_embeds)
+        return hidden_states
+
+    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)
+        else:
+            logits = self.logits_processor(self.model.embed_tokens.base_layer,
+                                           hidden_states, 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 (self.quant_config is not None and
+                (scale_name := self.quant_config.get_cache_scale(name))):
+                # Loading kv cache quantization scales
+                param = params_dict[scale_name]
+                weight_loader = getattr(param, "weight_loader",
+                                        default_weight_loader)
+                loaded_weight = (loaded_weight if loaded_weight.dim() == 0 else
+                                 loaded_weight[0])
+                weight_loader(param, loaded_weight)
+                loaded_params.add(scale_name)
+                continue
+
+            for param_name, shard_name, shard_id in stacked_params_mapping:
+                if shard_name not in name:
+                    continue
+                name = name.replace(shard_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:
+                # lm_head is not used in vllm as it is tied with embed_token.
+                # To prevent errors, skip loading lm_head.weight.
+                if "lm_head.weight" in name:
+                    continue
+                # Skip loading extra bias for GPTQ models.
+                if name.endswith(".bias") and name not in params_dict:
+                    continue
+                # Remapping the name of FP8 kv-scale.
+                name = maybe_remap_kv_scale_name(name, params_dict)
+                if name is None:
+                    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

.venv/lib/python3.11/site-packages/vllm/model_executor/models/dbrx.py

@@ -0,0 +1,496 @@
+# SPDX-License-Identifier: Apache-2.0
+
+from typing import Iterable, List, Optional, Set, Tuple, Union
+
+import torch
+import torch.nn as nn
+
+from vllm.attention import Attention, AttentionMetadata
+from vllm.config import CacheConfig, VllmConfig
+from vllm.distributed import (get_pp_group, get_tensor_model_parallel_rank,
+                              get_tensor_model_parallel_world_size)
+from vllm.model_executor.layers.fused_moe import FusedMoE
+from vllm.model_executor.layers.linear import (QKVParallelLinear,
+                                               ReplicatedLinear,
+                                               RowParallelLinear)
+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, maybe_remap_kv_scale_name)
+from vllm.model_executor.sampling_metadata import SamplingMetadata
+from vllm.sequence import IntermediateTensors
+from vllm.transformers_utils.configs.dbrx import DbrxConfig
+
+from .interfaces import SupportsPP
+from .utils import (is_pp_missing_parameter,
+                    make_empty_intermediate_tensors_factory, make_layers,
+                    maybe_prefix)
+
+
+class DbrxRouter(nn.Module):
+    """A Router implementation for DBRX that returns logits for each expert
+    per token.
+    """
+
+    def __init__(
+        self,
+        config: DbrxConfig,
+        params_dtype: Optional[torch.dtype] = None,
+    ):
+        super().__init__()
+        self.tp_size = get_tensor_model_parallel_world_size()
+        self.num_total_experts = config.ffn_config.moe_num_experts
+        self.d_model = config.d_model
+        self.layer = ReplicatedLinear(
+            self.d_model,
+            self.num_total_experts,
+            bias=False,
+            params_dtype=params_dtype,
+            quant_config=None,
+        )
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        router_logits, _ = self.layer(hidden_states)
+        return router_logits
+
+
+class DbrxExperts(FusedMoE):
+
+    def __init__(
+        self,
+        config: DbrxConfig,
+        quant_config: Optional[QuantizationConfig] = None,
+        params_dtype: Optional[torch.dtype] = None,
+    ):
+        super().__init__(
+            num_experts=config.ffn_config.moe_num_experts,
+            top_k=config.ffn_config.moe_top_k,
+            hidden_size=config.d_model,
+            intermediate_size=config.ffn_config.ffn_hidden_size,
+            params_dtype=params_dtype,
+            reduce_results=True,
+            renormalize=True,
+            quant_config=quant_config,
+            tp_size=get_tensor_model_parallel_world_size(),
+        )
+        self.config = config
+        self.tp_size = get_tensor_model_parallel_world_size()
+        self.d_model = config.d_model
+        self.intermediate_size = (self.config.ffn_config.ffn_hidden_size //
+                                  self.tp_size)
+
+    # Define custom weight loader for dbrx model
+    def weight_loader(self, param: nn.Parameter, loaded_weight: torch.Tensor,
+                      weight_name: str, param_name: str):
+        tp_rank = get_tensor_model_parallel_rank()
+        param_data = param.data
+        shard_size = self.intermediate_size
+        shard = slice(tp_rank * shard_size, (tp_rank + 1) * shard_size)
+        # DBRX uses GLU for each experts.
+        # GLU has 3 linear layers: w1, v1 and w2.
+        if weight_name.endswith("w1"):
+            if param_name.endswith("weight"):
+                loaded_weight = torch.reshape(
+                    loaded_weight,
+                    [-1, self.intermediate_size * self.tp_size, self.d_model],
+                )
+                param_data[:, 0:shard_size, :] = loaded_weight[:, shard, :]
+            elif param_name.endswith("weight_scale"):
+                param_data[:, 0] = loaded_weight
+            else:
+                param_data = loaded_weight
+        if weight_name.endswith("v1"):
+            if param_name.endswith("weight"):
+                loaded_weight = torch.reshape(
+                    loaded_weight,
+                    [-1, self.intermediate_size * self.tp_size, self.d_model],
+                )
+                param_data[:, shard_size:2 *
+                           shard_size, :] = loaded_weight[:, shard, :]
+            elif param_name.endswith("weight_scale"):
+                param_data[:, 1] = loaded_weight
+            else:
+                param_data[:] = loaded_weight
+        if weight_name.endswith("w2"):
+            if param_name.endswith("weight"):
+                loaded_weight = torch.reshape(
+                    loaded_weight,
+                    [-1, self.intermediate_size * self.tp_size, self.d_model],
+                ).transpose(1, 2)
+                param_data[:] = loaded_weight[:, :, shard]
+            else:
+                param_data[:] = loaded_weight
+
+
+class DbrxMoE(nn.Module):
+    """A tensor-parallel MoE implementation for DBRX.
+
+    Each expert's weights are sharded across all ranks and a fused MoE
+    kernel is used for the forward pass, and finally we reduce the outputs
+    across ranks.
+    """
+
+    def __init__(
+        self,
+        config: DbrxConfig,
+        quant_config: Optional[QuantizationConfig] = None,
+        params_dtype: Optional[torch.dtype] = None,
+    ):
+        super().__init__()
+        self.d_model = config.d_model
+        if params_dtype is None:
+            params_dtype = torch.get_default_dtype()
+        self.params_dtype = params_dtype
+
+        self.router = DbrxRouter(config, self.params_dtype)
+
+        self.experts = DbrxExperts(config=config,
+                                   quant_config=quant_config,
+                                   params_dtype=self.params_dtype)
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        orig_shape = hidden_states.shape
+        hidden_states = hidden_states.view(-1, self.d_model)
+        # router_logits: (num_tokens, n_experts)
+        router_logits = self.router(hidden_states)
+        final_hidden_states = self.experts(hidden_states, router_logits)
+        return final_hidden_states.view(orig_shape)
+
+
+class DbrxAttention(nn.Module):
+
+    def __init__(
+        self,
+        config: DbrxConfig,
+        cache_config: Optional[CacheConfig] = None,
+        quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
+    ):
+        super().__init__()
+        self.d_model = config.d_model
+        self.total_num_heads = config.n_heads
+        self.head_dim = self.d_model // self.total_num_heads
+        self.total_num_kv_heads = config.attn_config.kv_n_heads
+        self.clip_qkv = config.attn_config.clip_qkv
+        self.rope_theta = config.attn_config.rope_theta
+        self.max_position = config.max_seq_len
+
+        # pylint: disable=invalid-name
+        self.Wqkv = QKVParallelLinear(
+            self.d_model,
+            self.head_dim,
+            self.total_num_heads,
+            self.total_num_kv_heads,
+            bias=False,
+            quant_config=quant_config,
+        )
+        self.out_proj = RowParallelLinear(
+            self.d_model,
+            self.d_model,
+            bias=False,
+            quant_config=quant_config,
+        )
+        self.rotary_emb = get_rope(
+            self.head_dim,
+            rotary_dim=self.head_dim,
+            max_position=self.max_position,
+            base=int(self.rope_theta),
+            is_neox_style=True,
+        )
+
+        tp_world_size = get_tensor_model_parallel_world_size()
+        self.tp_size = tp_world_size
+        assert self.total_num_heads % tp_world_size == 0
+        self.num_heads = self.total_num_heads // tp_world_size
+        if self.total_num_kv_heads >= tp_world_size:
+            # Number of KV heads is greater than TP size, so we partition
+            # the KV heads across multiple tensor parallel GPUs.
+            assert self.total_num_kv_heads % tp_world_size == 0
+        else:
+            # Number of KV heads is less than TP size, so we replicate
+            # the KV heads across multiple tensor parallel GPUs.
+            assert tp_world_size % self.total_num_kv_heads == 0
+        self.num_kv_heads = max(1, self.total_num_kv_heads // tp_world_size)
+        self.q_size = self.num_heads * self.head_dim
+        self.kv_size = self.num_kv_heads * self.head_dim
+        self.scaling = self.head_dim**-0.5
+        self.attn = Attention(self.num_heads,
+                              self.head_dim,
+                              self.scaling,
+                              num_kv_heads=self.num_kv_heads,
+                              cache_config=cache_config,
+                              quant_config=quant_config,
+                              prefix=f"{prefix}.attn")
+
+    def forward(
+        self,
+        position_ids: torch.Tensor,
+        hidden_states: torch.Tensor,
+        kv_cache: torch.Tensor,
+        attn_metadata: AttentionMetadata,
+    ) -> torch.Tensor:
+        qkv, _ = self.Wqkv(hidden_states)
+        if self.clip_qkv is not None:
+            qkv.clamp_(min=-self.clip_qkv, max=self.clip_qkv)
+        q, k, v = qkv.split([self.q_size, self.kv_size, self.kv_size], dim=-1)
+        q, k = self.rotary_emb(position_ids, q, k)
+        attn_output = self.attn(q, k, v, kv_cache, attn_metadata)
+        hidden_states, _ = self.out_proj(attn_output)
+        return hidden_states
+
+
+class DbrxFusedNormAttention(nn.Module):
+
+    def __init__(
+        self,
+        config: DbrxConfig,
+        cache_config: Optional[CacheConfig] = None,
+        quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
+    ):
+        super().__init__()
+        self.d_model = config.d_model
+        self.attn = DbrxAttention(config,
+                                  cache_config,
+                                  quant_config,
+                                  prefix=f"{prefix}.attn")
+        self.norm_1 = nn.LayerNorm(self.d_model)
+        self.norm_2 = nn.LayerNorm(self.d_model)
+
+    def forward(
+        self,
+        position_ids: torch.Tensor,
+        hidden_states: torch.Tensor,
+        kv_cache: torch.Tensor,
+        attn_metadata: AttentionMetadata,
+    ) -> torch.Tensor:
+        residual = hidden_states
+        hidden_states = self.norm_1(hidden_states)
+        x = self.attn(
+            position_ids=position_ids,
+            hidden_states=hidden_states,
+            kv_cache=kv_cache,
+            attn_metadata=attn_metadata,
+        )
+        hidden_states = residual + x
+        residual = hidden_states
+        hidden_states = self.norm_2(hidden_states)
+        return hidden_states, residual
+
+
+class DbrxBlock(nn.Module):
+
+    def __init__(
+        self,
+        config: DbrxConfig,
+        cache_config: Optional[CacheConfig] = None,
+        quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
+    ):
+        super().__init__()
+        self.norm_attn_norm = DbrxFusedNormAttention(
+            config,
+            cache_config,
+            quant_config,
+            prefix=f"{prefix}.norm_attn_norm")
+        self.ffn = DbrxMoE(config, quant_config)
+
+    def forward(
+        self,
+        position_ids: torch.Tensor,
+        hidden_states: torch.Tensor,
+        kv_cache: torch.Tensor,
+        attn_metadata: AttentionMetadata,
+    ) -> torch.Tensor:
+        hidden_states, residual = self.norm_attn_norm(
+            position_ids=position_ids,
+            hidden_states=hidden_states,
+            kv_cache=kv_cache,
+            attn_metadata=attn_metadata,
+        )
+        hidden_states = self.ffn(hidden_states)
+        hidden_states = hidden_states + residual
+        return hidden_states
+
+
+class DbrxModel(nn.Module):
+
+    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
+        super().__init__()
+
+        config = vllm_config.model_config.hf_config
+        cache_config = vllm_config.cache_config
+        quant_config = vllm_config.quant_config
+
+        self.wte = VocabParallelEmbedding(
+            config.vocab_size,
+            config.d_model,
+        )
+        self.start_layer, self.end_layer, self.blocks = make_layers(
+            config.n_layers,
+            lambda prefix: DbrxBlock(
+                config, cache_config, quant_config, prefix=prefix),
+            prefix=f"{prefix}.blocks",
+        )
+        self.norm_f = nn.LayerNorm(config.d_model, eps=1e-5)
+        for module in self.modules():
+            if hasattr(module, "bias") and isinstance(module.bias,
+                                                      nn.Parameter):
+                # Remove the bias term in Linear and LayerNorm.
+                module.register_parameter("bias", None)
+        self.make_empty_intermediate_tensors = (
+            make_empty_intermediate_tensors_factory(["hidden_states"],
+                                                    config.d_model))
+
+    def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
+        return self.wte(input_ids)
+
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        position_ids: torch.Tensor,
+        kv_caches: List[torch.Tensor],
+        attn_metadata: AttentionMetadata,
+        intermediate_tensors: Optional[IntermediateTensors],
+        inputs_embeds: Optional[torch.Tensor] = None,
+    ) -> Union[torch.Tensor, IntermediateTensors]:
+        if get_pp_group().is_first_rank:
+            if inputs_embeds is not None:
+                hidden_states = inputs_embeds
+            else:
+                hidden_states = self.get_input_embeddings(input_ids)
+        else:
+            assert intermediate_tensors
+            hidden_states = intermediate_tensors["hidden_states"]
+        for i in range(self.start_layer, self.end_layer):
+            block = self.blocks[i]
+            hidden_states = block(
+                position_ids,
+                hidden_states,
+                kv_caches[i - self.start_layer],
+                attn_metadata,
+            )
+        if not get_pp_group().is_last_rank:
+            return IntermediateTensors({"hidden_states": hidden_states})
+        hidden_states = self.norm_f(hidden_states)
+        return hidden_states
+
+
+class DbrxForCausalLM(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
+        if config.tie_word_embeddings:
+            raise ValueError(
+                "tie_word_embeddings is not supported for Dbrx models.")
+        self.quant_config = quant_config
+        self.unpadded_vocab_size = config.vocab_size
+        self.transformer = DbrxModel(vllm_config=vllm_config,
+                                     prefix=maybe_prefix(
+                                         prefix, "transformer"))
+        self.lm_head = ParallelLMHead(
+            config.vocab_size,
+            config.d_model,
+            org_num_embeddings=config.vocab_size,
+            padding_size=DEFAULT_VOCAB_PADDING_SIZE,
+            quant_config=quant_config,
+        )
+        self.logits_processor = LogitsProcessor(self.unpadded_vocab_size,
+                                                config.vocab_size)
+        self.sampler = get_sampler()
+        self.make_empty_intermediate_tensors = (
+            self.transformer.make_empty_intermediate_tensors)
+
+    def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
+        return self.transformer.get_input_embeddings(input_ids)
+
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        positions: torch.Tensor,
+        kv_caches: List[torch.Tensor],
+        attn_metadata: AttentionMetadata,
+        intermediate_tensors: Optional[IntermediateTensors] = None,
+        inputs_embeds: Optional[torch.Tensor] = None,
+    ) -> Union[torch.Tensor, IntermediateTensors]:
+        hidden_states = self.transformer(input_ids, positions, kv_caches,
+                                         attn_metadata, intermediate_tensors,
+                                         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)
+        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]:
+        expert_params_mapping = [(
+            "w13" if weight_name in ["w1", "v1"] else "w2",
+            f"mlp.{weight_name}",
+        ) for weight_name in ["w1", "v1", "w2"]]
+        params_dict = dict(self.named_parameters(remove_duplicate=False))
+        loaded_params: Set[str] = set()
+
+        for name, loaded_weight in weights:
+            if (self.quant_config is not None and
+                (scale_name := self.quant_config.get_cache_scale(name))):
+                # Loading kv cache quantization scales
+                param = params_dict[scale_name]
+                weight_loader = getattr(param, "weight_loader",
+                                        default_weight_loader)
+                loaded_weight = (loaded_weight if loaded_weight.dim() == 0 else
+                                 loaded_weight[0])
+                weight_loader(param, loaded_weight)
+                loaded_params.add(scale_name)
+                continue
+
+            if name.endswith(("w1", "w2", "v1")):
+                name = name + "_weight"
+            for param_name, weight_name in expert_params_mapping:
+                if weight_name not in name:
+                    continue
+                name = name.replace(weight_name, param_name)
+                if is_pp_missing_parameter(name, self):
+                    continue
+                param = params_dict[name]
+                weight_loader = param.weight_loader
+                weight_loader(param, loaded_weight, weight_name, name)
+                break
+
+            else:
+                # Remapping the name of FP8 kv-scale.
+                name = maybe_remap_kv_scale_name(name, params_dict)
+                if name is None:
+                    continue
+
+                if is_pp_missing_parameter(name, self):
+                    continue
+                name = maybe_remap_kv_scale_name(name, params_dict)
+                if name is None:
+                    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

.venv/lib/python3.11/site-packages/vllm/model_executor/models/decilm.py

@@ -0,0 +1,124 @@
+# SPDX-License-Identifier: Apache-2.0
+
+# Adapted from
+# https://github.com/huggingface/transformers/blob/v4.28.0/src/transformers/models/llama/modeling_llama.py
+# Copyright 2023 DeciAI Research Team. All rights reserved.
+# Copyright 2023 The vLLM team.
+# Copyright 2022 EleutherAI and the HuggingFace Inc. team. All rights reserved.
+#
+# This code is based on MistralAI GPT-NeoX library and the GPT-NeoX
+# and OPT implementations in this library. It has been modified from its
+# original forms to accommodate minor architectural differences compared
+# to GPT-NeoX and OPT used by the Meta AI team that trained the model.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Inference-only DeciLM model compatible with HuggingFace weights."""
+
+from typing import Iterable, Set, Tuple
+
+import torch
+
+from vllm.config import VllmConfig
+from vllm.model_executor.model_loader.weight_utils import default_weight_loader
+from vllm.model_executor.models.llama import LlamaForCausalLM
+
+from .utils import is_pp_missing_parameter
+
+
+class DeciLMForCausalLM(LlamaForCausalLM):
+    """
+    Implementation for https://huggingface.co/Deci/DeciLM-7b-instruct.
+    Based on the llama executor.
+
+    The main difference is that DeciLM uses Variable Grouped Query Attention.
+    The constant number of GQA heads in the decoder is overridden with a value
+    per layer.
+
+    Usually, in the HuggingFace implementation, instead of
+    "config.num_key_value_heads", we use
+    "config.num_key_value_heads_per_layer[i]" which varies.
+
+    Currently, PagedAttention does not work well with variable GQA, so we
+    normalize the weights upon loading, and use uniform GQA with the max value
+    instead.
+    """
+
+    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
+        config = vllm_config.model_config.hf_config
+        config.num_key_value_heads = max(config.num_key_value_heads_per_layer)
+        delattr(config, "num_key_value_heads_per_layer")
+        super().__init__(vllm_config=vllm_config)
+
+    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 "k_proj" in name or "v_proj" in name:
+                loaded_weight = self._degroup_weight(loaded_weight)
+
+            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
+
+    def _degroup_weight(self, loaded_weight: torch.Tensor) -> torch.Tensor:
+        hidden_size = self.config.hidden_size
+        head_size = self.config.hidden_size // self.config.num_attention_heads
+        target_num_kv_heads = self.config.num_key_value_heads
+        num_kv_heads = loaded_weight.shape[0] // head_size
+        n_repeats = target_num_kv_heads / num_kv_heads
+        assert n_repeats == int(n_repeats)
+
+        n_repeats = int(n_repeats)
+        loaded_weight = loaded_weight.view(num_kv_heads, head_size,
+                                           hidden_size)
+        loaded_weight = torch.repeat_interleave(loaded_weight,
+                                                repeats=n_repeats,
+                                                dim=0)
+        loaded_weight = loaded_weight.reshape(target_num_kv_heads * head_size,
+                                              hidden_size)
+
+        return loaded_weight

.venv/lib/python3.11/site-packages/vllm/model_executor/models/deepseek_v2.py

@@ -0,0 +1,817 @@
+# SPDX-License-Identifier: Apache-2.0
+
+# Adapted from
+# https://github.com/huggingface/transformers/blob/v4.28.0/src/transformers/models/llama/modeling_llama.py
+# Copyright 2023 The vLLM team.
+# Copyright 2023 DeepSeek-AI and the HuggingFace Inc. team. All rights reserved.
+#
+# This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX
+# and OPT implementations in this library. It has been modified from its
+# original forms to accommodate minor architectural differences compared
+# to GPT-NeoX and OPT used by the Meta AI team that trained the model.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Inference-only DeepseekV2/DeepseekV3 model."""
+from typing import Any, Dict, Iterable, List, Optional, Set, Tuple, Union
+
+import torch
+from torch import nn
+from transformers import PretrainedConfig
+
+from vllm.attention import Attention, AttentionMetadata
+from vllm.compilation.decorators import support_torch_compile
+from vllm.config import CacheConfig, ModelConfig, VllmConfig
+from vllm.distributed import (get_pp_group,
+                              get_tensor_model_parallel_world_size,
+                              tensor_model_parallel_all_reduce)
+from vllm.model_executor.layers.activation import SiluAndMul
+from vllm.model_executor.layers.fused_moe import FusedMoE
+from vllm.model_executor.layers.layernorm import RMSNorm
+from vllm.model_executor.layers.linear import (ColumnParallelLinear,
+                                               MergedColumnParallelLinear,
+                                               ReplicatedLinear,
+                                               RowParallelLinear)
+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, maybe_remap_kv_scale_name)
+from vllm.model_executor.sampling_metadata import SamplingMetadata
+from vllm.sequence import IntermediateTensors
+
+from .interfaces import SupportsPP
+from .utils import (PPMissingLayer, is_pp_missing_parameter,
+                    make_empty_intermediate_tensors_factory, make_layers,
+                    maybe_prefix)
+
+
+class DeepseekV2MLP(nn.Module):
+
+    def __init__(
+        self,
+        hidden_size: int,
+        intermediate_size: int,
+        hidden_act: str,
+        quant_config: Optional[QuantizationConfig] = None,
+        reduce_results: bool = True,
+        prefix: str = "",
+    ) -> None:
+        super().__init__()
+        self.gate_up_proj = MergedColumnParallelLinear(
+            hidden_size, [intermediate_size] * 2,
+            bias=False,
+            quant_config=quant_config,
+            prefix=f"{prefix}.gate_up_proj")
+        self.down_proj = RowParallelLinear(intermediate_size,
+                                           hidden_size,
+                                           bias=False,
+                                           quant_config=quant_config,
+                                           reduce_results=reduce_results,
+                                           prefix=f"{prefix}.down_proj")
+        if hidden_act != "silu":
+            raise ValueError(f"Unsupported activation: {hidden_act}. "
+                             "Only silu is supported for now.")
+        self.act_fn = SiluAndMul()
+
+    def forward(self, x):
+        gate_up, _ = self.gate_up_proj(x)
+        x = self.act_fn(gate_up)
+        x, _ = self.down_proj(x)
+        return x
+
+
+class DeepseekV2MoE(nn.Module):
+
+    def __init__(
+        self,
+        config: PretrainedConfig,
+        quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
+    ):
+        super().__init__()
+        self.tp_size = get_tensor_model_parallel_world_size()
+        self.routed_scaling_factor = config.routed_scaling_factor
+        self.n_shared_experts = config.n_shared_experts
+        self.routed_scaling_factor = config.routed_scaling_factor
+        if self.tp_size > config.n_routed_experts:
+            raise ValueError(
+                f"Tensor parallel size {self.tp_size} is greater than "
+                f"the number of experts {config.n_routed_experts}.")
+
+        if config.hidden_act != "silu":
+            raise ValueError(f"Unsupported activation: {config.hidden_act}. "
+                             "Only silu is supported for now.")
+
+        self.gate = ReplicatedLinear(config.hidden_size,
+                                     config.n_routed_experts,
+                                     bias=False,
+                                     quant_config=None,
+                                     prefix=f"{prefix}.gate")
+        if config.topk_method == "noaux_tc":
+            self.gate.e_score_correction_bias = nn.Parameter(
+                torch.empty(config.n_routed_experts))
+        else:
+            self.gate.e_score_correction_bias = None
+
+        self.experts = FusedMoE(
+            num_experts=config.n_routed_experts,
+            top_k=config.num_experts_per_tok,
+            hidden_size=config.hidden_size,
+            intermediate_size=config.moe_intermediate_size,
+            reduce_results=False,
+            renormalize=config.norm_topk_prob,
+            quant_config=quant_config,
+            use_grouped_topk=True,
+            num_expert_group=config.n_group,
+            topk_group=config.topk_group,
+            prefix=f"{prefix}.experts",
+            scoring_func=config.scoring_func,
+            e_score_correction_bias=self.gate.e_score_correction_bias)
+
+        if config.n_shared_experts is not None:
+            intermediate_size = (config.moe_intermediate_size *
+                                 config.n_shared_experts)
+            self.shared_experts = DeepseekV2MLP(
+                hidden_size=config.hidden_size,
+                intermediate_size=intermediate_size,
+                hidden_act=config.hidden_act,
+                quant_config=quant_config,
+                reduce_results=False,
+            )
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        num_tokens, hidden_dim = hidden_states.shape
+        hidden_states = hidden_states.view(-1, hidden_dim)
+        if self.n_shared_experts is not None:
+            shared_output = self.shared_experts(hidden_states)
+        # router_logits: (num_tokens, n_experts)
+        router_logits, _ = self.gate(hidden_states)
+        final_hidden_states = self.experts(
+            hidden_states=hidden_states,
+            router_logits=router_logits) * self.routed_scaling_factor
+        if shared_output is not None:
+            final_hidden_states = final_hidden_states + shared_output
+        if self.tp_size > 1:
+            final_hidden_states = tensor_model_parallel_all_reduce(
+                final_hidden_states)
+
+        return final_hidden_states.view(num_tokens, hidden_dim)
+
+
+def yarn_get_mscale(scale: float = 1, mscale: float = 1) -> float:
+    import math
+    if scale <= 1:
+        return 1.0
+    return 0.1 * mscale * math.log(scale) + 1.0
+
+
+class DeepseekV2Attention(nn.Module):
+
+    def __init__(
+        self,
+        config: PretrainedConfig,
+        hidden_size: int,
+        num_heads: int,
+        qk_nope_head_dim: int,
+        qk_rope_head_dim: int,
+        v_head_dim: int,
+        q_lora_rank: int,
+        kv_lora_rank: int,
+        rope_theta: float = 10000,
+        rope_scaling: Optional[Dict[str, Any]] = None,
+        max_position_embeddings: int = 8192,
+        cache_config: Optional[CacheConfig] = None,
+        quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
+    ) -> None:
+        super().__init__()
+        self.hidden_size = hidden_size
+        self.qk_nope_head_dim = qk_nope_head_dim
+        self.qk_rope_head_dim = qk_rope_head_dim
+        self.qk_head_dim = qk_nope_head_dim + qk_rope_head_dim
+        self.v_head_dim = v_head_dim
+        self.q_lora_rank = q_lora_rank
+        self.kv_lora_rank = kv_lora_rank
+        self.num_heads = num_heads
+        tp_size = get_tensor_model_parallel_world_size()
+        assert num_heads % tp_size == 0
+        self.num_local_heads = num_heads // tp_size
+        self.scaling = self.qk_head_dim**-0.5
+        self.rope_theta = rope_theta
+        self.max_position_embeddings = max_position_embeddings
+
+        if self.q_lora_rank is not None:
+            self.q_a_proj = ReplicatedLinear(self.hidden_size,
+                                             self.q_lora_rank,
+                                             bias=False,
+                                             quant_config=quant_config,
+                                             prefix=f"{prefix}.q_a_proj")
+            self.q_a_layernorm = RMSNorm(self.q_lora_rank,
+                                         eps=config.rms_norm_eps)
+            self.q_b_proj = ColumnParallelLinear(q_lora_rank,
+                                                 self.num_heads *
+                                                 self.qk_head_dim,
+                                                 bias=False,
+                                                 quant_config=quant_config,
+                                                 prefix=f"{prefix}.q_b_proj")
+        else:
+            self.q_proj = ColumnParallelLinear(self.hidden_size,
+                                               self.num_heads *
+                                               self.qk_head_dim,
+                                               bias=False,
+                                               quant_config=quant_config,
+                                               prefix=f"{prefix}.q_proj")
+
+        self.kv_a_proj_with_mqa = ReplicatedLinear(
+            self.hidden_size,
+            self.kv_lora_rank + self.qk_rope_head_dim,
+            bias=False,
+            quant_config=quant_config,
+            prefix=f"{prefix}.kv_a_proj_with_mqa")
+        self.kv_a_layernorm = RMSNorm(self.kv_lora_rank,
+                                      eps=config.rms_norm_eps)
+        self.kv_b_proj = ColumnParallelLinear(
+            self.kv_lora_rank,
+            self.num_heads * (self.qk_nope_head_dim + self.v_head_dim),
+            bias=False,
+            quant_config=quant_config,
+            prefix=f"{prefix}.kv_b_proj")
+        # O projection.
+        self.o_proj = RowParallelLinear(self.num_heads * self.v_head_dim,
+                                        self.hidden_size,
+                                        bias=False,
+                                        quant_config=quant_config,
+                                        prefix=f"{prefix}.o_proj")
+        if rope_scaling:
+            rope_scaling["rope_type"] = 'deepseek_yarn'
+            self.use_normal_rope = False
+        else:
+            self.use_normal_rope = True
+        self.rotary_emb = get_rope(qk_rope_head_dim,
+                                   rotary_dim=qk_rope_head_dim,
+                                   max_position=max_position_embeddings,
+                                   base=rope_theta,
+                                   rope_scaling=rope_scaling,
+                                   is_neox_style=False)
+
+        if rope_scaling:
+            mscale_all_dim = rope_scaling.get("mscale_all_dim", False)
+            scaling_factor = rope_scaling["factor"]
+            mscale = yarn_get_mscale(scaling_factor, float(mscale_all_dim))
+            self.scaling = self.scaling * mscale * mscale
+
+        self.attn = Attention(self.num_local_heads,
+                              self.qk_head_dim,
+                              self.scaling,
+                              num_kv_heads=self.num_local_heads,
+                              cache_config=cache_config,
+                              quant_config=quant_config,
+                              prefix=f"{prefix}.attn")
+
+    def forward(
+        self,
+        positions: torch.Tensor,
+        hidden_states: torch.Tensor,
+        kv_cache: torch.Tensor,
+        attn_metadata: AttentionMetadata,
+    ) -> torch.Tensor:
+        if self.q_lora_rank is not None:
+            q = self.q_a_proj(hidden_states)[0]
+            q = self.q_a_layernorm(q)
+            q = self.q_b_proj(q)[0].view(-1, self.num_local_heads,
+                                         self.qk_head_dim)
+        else:
+            q = self.q_proj(hidden_states)[0].view(-1, self.num_local_heads,
+                                                   self.qk_head_dim)
+        q_nope, q_pe = q.split([self.qk_nope_head_dim, self.qk_rope_head_dim],
+                               dim=-1)
+        latent_cache = self.kv_a_proj_with_mqa(hidden_states)[0]
+        kv_a, _ = latent_cache.split(
+            [self.kv_lora_rank, self.qk_rope_head_dim], dim=-1)
+        latent_cache = latent_cache.unsqueeze(1)
+        kv_a = self.kv_a_layernorm(kv_a.contiguous())
+        kv = self.kv_b_proj(kv_a)[0]
+        kv = kv.view(-1, self.num_local_heads,
+                     self.qk_nope_head_dim + self.v_head_dim)
+        k_nope, v = kv.split([self.qk_nope_head_dim, self.v_head_dim], dim=-1)
+        k_pe = latent_cache[:, :, self.kv_lora_rank:]
+
+        if self.use_normal_rope:
+            seq_len = positions.size(0)
+            ori_q_pe_shape, ori_k_pe_shape = q_pe.shape, k_pe.shape
+            q_pe = q_pe.reshape(seq_len, -1)
+            k_pe = k_pe.reshape(seq_len, -1)
+
+        q_pe, k_pe = self.rotary_emb(positions, q_pe, k_pe)
+
+        if self.use_normal_rope:
+            q_pe, k_pe = q_pe.view(ori_q_pe_shape), k_pe.view(ori_k_pe_shape)
+
+        q[..., self.qk_nope_head_dim:] = q_pe
+        k = torch.empty_like(q)
+        k[..., :self.qk_nope_head_dim] = k_nope
+        k[..., self.qk_nope_head_dim:] = k_pe
+        # padding value to qk_head_dim for alignment
+        v = torch.nn.functional.pad(
+            v, [0, self.qk_head_dim - self.v_head_dim],
+            value=0).view(-1, self.num_local_heads * self.qk_head_dim)
+        attn_output = self.attn(q, k, v, kv_cache, attn_metadata)
+        attn_output = attn_output.view(
+            -1, self.num_local_heads,
+            self.qk_head_dim)[..., :self.v_head_dim].reshape(
+                -1, self.num_local_heads * self.v_head_dim)
+        output, _ = self.o_proj(attn_output)
+        return output
+
+
+class DeepseekV2MLAAttention(nn.Module):
+    """
+    Main reference: DeepseekV2 paper, and FlashInfer Implementation
+    (https://arxiv.org/abs/2405.04434 and https://github.com/flashinfer-ai/flashinfer/pull/551).
+    
+    For more info see MLACommonImpl in: vllm/attention/backends/mla/utils.py
+    """
+
+    def __init__(
+        self,
+        config: PretrainedConfig,
+        hidden_size: int,
+        num_heads: int,
+        qk_nope_head_dim: int,
+        qk_rope_head_dim: int,
+        v_head_dim: int,
+        q_lora_rank: Optional[int],
+        kv_lora_rank: int,
+        rope_theta: float = 10000,
+        rope_scaling: Optional[Dict[str, Any]] = None,
+        max_position_embeddings: int = 8192,
+        cache_config: Optional[CacheConfig] = None,
+        quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
+    ) -> None:
+        super().__init__()
+        self.hidden_size = hidden_size
+        self.qk_nope_head_dim = qk_nope_head_dim
+        self.qk_rope_head_dim = qk_rope_head_dim
+        self.qk_head_dim = qk_nope_head_dim + qk_rope_head_dim
+        self.v_head_dim = v_head_dim
+
+        self.q_lora_rank = q_lora_rank
+        self.kv_lora_rank = kv_lora_rank
+
+        self.num_heads = num_heads
+        tp_size = get_tensor_model_parallel_world_size()
+        assert num_heads % tp_size == 0
+        self.num_local_heads = num_heads // tp_size
+
+        self.scaling = self.qk_head_dim**-0.5
+        self.rope_theta = rope_theta
+        self.max_position_embeddings = max_position_embeddings
+
+        if self.q_lora_rank is not None:
+            self.q_a_proj = ReplicatedLinear(self.hidden_size,
+                                             self.q_lora_rank,
+                                             bias=False,
+                                             quant_config=quant_config,
+                                             prefix=f"{prefix}.q_a_proj")
+            self.q_a_layernorm = RMSNorm(self.q_lora_rank,
+                                         eps=config.rms_norm_eps)
+            self.q_b_proj = ColumnParallelLinear(q_lora_rank,
+                                                 self.num_heads *
+                                                 self.qk_head_dim,
+                                                 bias=False,
+                                                 quant_config=quant_config,
+                                                 prefix=f"{prefix}.q_b_proj")
+        else:
+            self.q_proj = ColumnParallelLinear(self.hidden_size,
+                                               self.num_heads *
+                                               self.qk_head_dim,
+                                               bias=False,
+                                               quant_config=quant_config,
+                                               prefix=f"{prefix}.q_proj")
+
+        self.kv_a_proj_with_mqa = ReplicatedLinear(
+            self.hidden_size,
+            self.kv_lora_rank + self.qk_rope_head_dim,
+            bias=False,
+            quant_config=quant_config,
+            prefix=f"{prefix}.kv_a_proj_with_mqa")
+        self.kv_a_layernorm = RMSNorm(self.kv_lora_rank,
+                                      eps=config.rms_norm_eps)
+        self.kv_b_proj = ColumnParallelLinear(
+            self.kv_lora_rank,
+            self.num_heads * (self.qk_nope_head_dim + self.v_head_dim),
+            bias=False,
+            quant_config=quant_config,
+            prefix=f"{prefix}.kv_b_proj")
+        self.o_proj = RowParallelLinear(self.num_heads * self.v_head_dim,
+                                        self.hidden_size,
+                                        bias=False,
+                                        quant_config=quant_config,
+                                        prefix=f"{prefix}.o_proj")
+
+        if rope_scaling:
+            rope_scaling["rope_type"] = 'deepseek_yarn'
+        self.rotary_emb = get_rope(qk_rope_head_dim,
+                                   rotary_dim=qk_rope_head_dim,
+                                   max_position=max_position_embeddings,
+                                   base=rope_theta,
+                                   rope_scaling=rope_scaling,
+                                   is_neox_style=False)
+        if rope_scaling:
+            mscale_all_dim = rope_scaling.get("mscale_all_dim", False)
+            scaling_factor = rope_scaling["factor"]
+            mscale = yarn_get_mscale(scaling_factor, float(mscale_all_dim))
+            self.scaling = self.scaling * mscale * mscale
+
+        self.mla_attn = Attention(
+            num_heads=self.num_local_heads,
+            head_size=self.kv_lora_rank,
+            scale=self.scaling,
+            num_kv_heads=1,
+            cache_config=cache_config,
+            quant_config=quant_config,
+            prefix=f"{prefix}.attn",
+            use_mla=True,
+            # MLA Args
+            q_lora_rank=self.q_lora_rank,
+            kv_lora_rank=self.kv_lora_rank,
+            qk_nope_head_dim=self.qk_nope_head_dim,
+            qk_rope_head_dim=self.qk_rope_head_dim,
+            qk_head_dim=self.qk_head_dim,
+            v_head_dim=self.v_head_dim,
+            rotary_emb=self.rotary_emb,
+            q_proj=self.q_proj if self.q_lora_rank is None else self.q_b_proj,
+            kv_b_proj=self.kv_b_proj,
+            o_proj=self.o_proj,
+        )
+
+        self.prefix = prefix
+        self.debug_layer_idx = int(self.prefix.split(".")[-2])
+
+    def forward(
+        self,
+        positions: torch.Tensor,
+        hidden_states: torch.Tensor,
+        kv_cache: torch.Tensor,
+        attn_metadata: AttentionMetadata,
+    ) -> torch.Tensor:
+        if self.q_lora_rank is not None:
+            ckq = self.q_a_proj(hidden_states)[0]
+            hidden_states_or_q_c = self.q_a_layernorm(ckq)
+        else:
+            hidden_states_or_q_c = hidden_states
+        kv_c, k_pe = self.kv_a_proj_with_mqa(hidden_states)[0].split(
+            [self.kv_lora_rank, self.qk_rope_head_dim], dim=-1)
+        kv_c_normed = self.kv_a_layernorm(kv_c.contiguous())
+        return self.mla_attn(hidden_states_or_q_c, kv_c_normed, k_pe, kv_cache,
+                             attn_metadata)
+
+
+class DeepseekV2DecoderLayer(nn.Module):
+
+    def __init__(
+        self,
+        config: PretrainedConfig,
+        prefix: str,
+        model_config: ModelConfig,
+        cache_config: Optional[CacheConfig] = None,
+        quant_config: Optional[QuantizationConfig] = None,
+    ) -> None:
+        super().__init__()
+        self.hidden_size = config.hidden_size
+        rope_theta = getattr(config, "rope_theta", 10000)
+        rope_scaling = getattr(config, "rope_scaling", None)
+        max_position_embeddings = getattr(config, "max_position_embeddings",
+                                          8192)
+        # DecoderLayers are created with `make_layers` which passes the prefix
+        # with the layer's index.
+        layer_idx = int(prefix.split(sep='.')[-1])
+        if model_config.use_mla:
+            attn_cls = DeepseekV2MLAAttention
+        else:
+            attn_cls = DeepseekV2Attention
+        self.self_attn = attn_cls(
+            config=config,
+            hidden_size=self.hidden_size,
+            num_heads=config.num_attention_heads,
+            qk_nope_head_dim=config.qk_nope_head_dim,
+            qk_rope_head_dim=config.qk_rope_head_dim,
+            v_head_dim=config.v_head_dim,
+            q_lora_rank=config.q_lora_rank
+            if hasattr(config, "q_lora_rank") else None,
+            kv_lora_rank=config.kv_lora_rank,
+            rope_theta=rope_theta,
+            rope_scaling=rope_scaling,
+            max_position_embeddings=max_position_embeddings,
+            cache_config=cache_config,
+            quant_config=quant_config,
+            prefix=f"{prefix}.self_attn",
+        )
+
+        if (config.n_routed_experts is not None
+                and layer_idx >= config.first_k_dense_replace
+                and layer_idx % config.moe_layer_freq == 0):
+            self.mlp = DeepseekV2MoE(
+                config=config,
+                quant_config=quant_config,
+                prefix=f"{prefix}.mlp",
+            )
+        else:
+            self.mlp = DeepseekV2MLP(
+                hidden_size=config.hidden_size,
+                intermediate_size=config.intermediate_size,
+                hidden_act=config.hidden_act,
+                quant_config=quant_config,
+                prefix=f"{prefix}.mlp",
+            )
+        self.input_layernorm = RMSNorm(config.hidden_size,
+                                       eps=config.rms_norm_eps)
+        self.post_attention_layernorm = RMSNorm(config.hidden_size,
+                                                eps=config.rms_norm_eps)
+
+    def forward(
+        self,
+        positions: torch.Tensor,
+        hidden_states: torch.Tensor,
+        kv_cache: torch.Tensor,
+        attn_metadata: AttentionMetadata,
+        residual: Optional[torch.Tensor],
+    ) -> torch.Tensor:
+        # Self Attention
+        if residual is None:
+            residual = hidden_states
+            hidden_states = self.input_layernorm(hidden_states)
+        else:
+            hidden_states, residual = self.input_layernorm(
+                hidden_states, residual)
+        hidden_states = self.self_attn(
+            positions=positions,
+            hidden_states=hidden_states,
+            kv_cache=kv_cache,
+            attn_metadata=attn_metadata,
+        )
+
+        # Fully Connected
+        hidden_states, residual = self.post_attention_layernorm(
+            hidden_states, residual)
+        hidden_states = self.mlp(hidden_states)
+        return hidden_states, residual
+
+
+@support_torch_compile
+class DeepseekV2Model(nn.Module):
+
+    fall_back_to_pt_during_load = False
+
+    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
+        super().__init__()
+
+        config = vllm_config.model_config.hf_config
+        model_config = vllm_config.model_config
+        cache_config = vllm_config.cache_config
+        quant_config = vllm_config.quant_config
+
+        self.padding_idx = config.pad_token_id
+        self.vocab_size = config.vocab_size
+
+        if get_pp_group().is_first_rank:
+            self.embed_tokens = VocabParallelEmbedding(
+                config.vocab_size,
+                config.hidden_size,
+            )
+        else:
+            self.embed_tokens = PPMissingLayer()
+
+        self.start_layer, self.end_layer, self.layers = make_layers(
+            config.num_hidden_layers,
+            lambda prefix: DeepseekV2DecoderLayer(
+                config,
+                prefix,
+                model_config=model_config,
+                cache_config=cache_config,
+                quant_config=quant_config,
+            ),
+            prefix=f"{prefix}.layers")
+
+        if get_pp_group().is_last_rank:
+            self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        else:
+            self.norm = PPMissingLayer()
+        self.make_empty_intermediate_tensors = (
+            make_empty_intermediate_tensors_factory(
+                ["hidden_states", "residual"], config.hidden_size))
+
+    def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
+        return self.embed_tokens(input_ids)
+
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        positions: torch.Tensor,
+        kv_caches: List[torch.Tensor],
+        attn_metadata: AttentionMetadata,
+        intermediate_tensors: Optional[IntermediateTensors],
+        inputs_embeds: Optional[torch.Tensor] = None,
+    ) -> Union[torch.Tensor, IntermediateTensors]:
+        if get_pp_group().is_first_rank:
+            if inputs_embeds is not None:
+                hidden_states = inputs_embeds
+            else:
+                hidden_states = self.get_input_embeddings(input_ids)
+            residual = None
+        else:
+            assert intermediate_tensors is not None
+            hidden_states = intermediate_tensors["hidden_states"]
+            residual = intermediate_tensors["residual"]
+
+        for i in range(self.start_layer, self.end_layer):
+            layer = self.layers[i]
+            hidden_states, residual = layer(positions, hidden_states,
+                                            kv_caches[i - self.start_layer],
+                                            attn_metadata, residual)
+
+        if not get_pp_group().is_last_rank:
+            return IntermediateTensors({
+                "hidden_states": hidden_states,
+                "residual": residual
+            })
+
+        hidden_states, _ = self.norm(hidden_states, residual)
+        return hidden_states
+
+
+class DeepseekV2ForCausalLM(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 = DeepseekV2Model(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,
+        kv_caches: List[torch.Tensor],
+        attn_metadata: AttentionMetadata,
+        intermediate_tensors: Optional[IntermediateTensors] = None,
+        inputs_embeds: Optional[torch.Tensor] = None,
+    ) -> Union[torch.Tensor, IntermediateTensors]:
+        hidden_states = self.model(input_ids, positions, kv_caches,
+                                   attn_metadata, intermediate_tensors,
+                                   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)
+        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 make_empty_intermediate_tensors(
+            self, batch_size: int, dtype: torch.dtype,
+            device: torch.device) -> IntermediateTensors:
+        return IntermediateTensors({
+            "hidden_states":
+            torch.zeros((batch_size, self.config.hidden_size),
+                        dtype=dtype,
+                        device=device),
+            "residual":
+            torch.zeros((batch_size, self.config.hidden_size),
+                        dtype=dtype,
+                        device=device),
+        })
+
+    def load_weights(self, weights: Iterable[Tuple[str,
+                                                   torch.Tensor]]) -> Set[str]:
+        stacked_params_mapping = [
+            # (param_name, shard_name, shard_id)
+            ("gate_up_proj", "gate_proj", 0),
+            ("gate_up_proj", "up_proj", 1),
+        ]
+
+        # Params for weights, fp8 weight scales, fp8 activation scales
+        # (param_name, weight_name, expert_id, shard_id)
+        expert_params_mapping = FusedMoE.make_expert_params_mapping(
+            ckpt_gate_proj_name="gate_proj",
+            ckpt_down_proj_name="down_proj",
+            ckpt_up_proj_name="up_proj",
+            num_experts=self.config.n_routed_experts)
+
+        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
+
+            # TODO(simon): support nextn predict layers
+            if hasattr(self.config, "num_nextn_predict_layers"
+                       ) and self.config.num_nextn_predict_layers > 0:
+                assert self.config.num_nextn_predict_layers == 1
+                layer_idx = self.config.num_hidden_layers
+                if name.startswith(f"model.layers.{layer_idx}"):
+                    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:
+                    continue
+                # We have mlp.experts[0].gate_proj in the checkpoint.
+                # Since we handle the experts below in expert_params_mapping,
+                # we need to skip here BEFORE we update the name, otherwise
+                # name will be updated to mlp.experts[0].gate_up_proj, which
+                # will then be updated below in expert_params_mapping
+                # for mlp.experts[0].gate_gate_up_proj, which breaks load.
+                if (("mlp.experts." in name) and name not in params_dict):
+                    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:
+                for mapping in expert_params_mapping:
+                    param_name, weight_name, expert_id, shard_id = mapping
+                    if weight_name not in name:
+                        continue
+                    name = name.replace(weight_name, param_name)
+
+                    if is_pp_missing_parameter(name, self):
+                        continue
+
+                    param = params_dict[name]
+                    weight_loader = param.weight_loader
+                    weight_loader(param,
+                                  loaded_weight,
+                                  name,
+                                  shard_id=shard_id,
+                                  expert_id=expert_id)
+                    break
+                else:
+                    # Skip loading extra bias for GPTQ models.
+                    if name.endswith(".bias") and name not in params_dict:
+                        continue
+
+                    # Remapping the name of FP8 kv-scale.
+                    name = maybe_remap_kv_scale_name(name, params_dict)
+                    if name is None:
+                        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 DeepseekV3ForCausalLM(DeepseekV2ForCausalLM):
+    pass

.venv/lib/python3.11/site-packages/vllm/model_executor/models/deepseek_vl2.py

@@ -0,0 +1,650 @@
+# SPDX-License-Identifier: Apache-2.0
+
+# adapted from https://github.com/deepseek-ai/DeepSeek-VL2/blob/faf18023f24b962b32d9f0a2d89e402a8d383a78/deepseek_vl2/models/modeling_deepseek_vl_v2.py
+"""Inference-only Deepseek-VL2 model compatible with HuggingFace weights."""
+import math
+from functools import cached_property
+from typing import (Iterable, List, Literal, Mapping, Optional, Set, Tuple,
+                    TypedDict, Union)
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from einops import rearrange, repeat
+from transformers import BatchFeature
+
+from vllm.attention import AttentionMetadata
+from vllm.config import VllmConfig
+from vllm.logger import init_logger
+from vllm.model_executor import SamplingMetadata
+from vllm.model_executor.layers.quantization import QuantizationConfig
+from vllm.model_executor.layers.sampler import SamplerOutput, get_sampler
+from vllm.model_executor.model_loader.utils import set_default_torch_dtype
+from vllm.multimodal import MULTIMODAL_REGISTRY
+from vllm.multimodal.inputs import (MultiModalFieldConfig, MultiModalKwargs,
+                                    NestedTensors)
+from vllm.multimodal.parse import (ImageEmbeddingItems, ImageProcessorItems,
+                                   ImageSize, MultiModalDataItems)
+from vllm.multimodal.processing import (BaseMultiModalProcessor,
+                                        BaseProcessingInfo, PromptReplacement)
+from vllm.multimodal.profiling import BaseDummyInputsBuilder, ProcessorInputs
+from vllm.multimodal.utils import cached_get_tokenizer
+from vllm.sequence import IntermediateTensors
+from vllm.transformers_utils.configs.deepseek_vl2 import (DeepseekVLV2Config,
+                                                          MlpProjectorConfig,
+                                                          VisionEncoderConfig)
+from vllm.transformers_utils.processors.deepseek_vl2 import (
+    DeepseekVLV2Processor)
+from vllm.utils import is_list_of
+
+from .interfaces import SupportsMultiModal, SupportsPP
+from .utils import (AutoWeightsLoader, WeightsMapper, flatten_bn,
+                    init_vllm_registered_model, maybe_prefix,
+                    merge_multimodal_embeddings)
+
+logger = init_logger(__name__)
+
+# The image token id may be various
+_IMAGE_TOKEN = "<image>"
+
+
+class DeepseekVL2ImagePixelInputs(TypedDict):
+    type: Literal["pixel_values"]
+    data: Union[torch.Tensor, List[torch.Tensor]]
+    """
+    Shape: `(batch_size * num_images, num_channels, height, width)`
+    """
+    images_spatial_crop: torch.Tensor
+    """
+    Shape: `(batch_size * num_images, 2)`
+    """
+
+
+class DeepseekVL2VImageEmbeddingInputs(TypedDict):
+    type: Literal["image_embeds"]
+    data: Union[torch.Tensor, List[torch.Tensor]]
+    """Shape: `(batch_size * num_images, image_feature_size, hidden_size)`
+
+    `hidden_size` must match the hidden size of language model backbone.
+    """
+
+
+DeepseekVL2ImageInputs = Union[DeepseekVL2ImagePixelInputs,
+                               DeepseekVL2VImageEmbeddingInputs]
+
+
+class MlpProjector(nn.Module):
+
+    def __init__(self, cfg: MlpProjectorConfig):
+
+        super().__init__()
+
+        self.cfg = cfg
+        assert not cfg.token_pooling, (
+            "Token pooling is not supported currently.")
+
+        if cfg.projector_type == "downsample_mlp_gelu":
+            mlp_depth = cfg.depth
+            mlp_ratio = cfg.mlp_ratio
+            modules = [
+                nn.Linear(
+                    cfg.input_dim * cfg.downsample_ratio *
+                    cfg.downsample_ratio, cfg.n_embed * mlp_ratio)
+            ]
+            for _ in range(1, mlp_depth - 1):
+                modules.append(nn.GELU())
+                modules.append(
+                    nn.Linear(cfg.n_embed * mlp_ratio,
+                              cfg.n_embed * mlp_ratio))
+            modules.append(nn.GELU())
+            modules.append(nn.Linear(cfg.n_embed * mlp_ratio, cfg.n_embed))
+            modules = nn.Sequential(*modules)
+
+        else:
+            raise NotImplementedError(
+                f"Unsupported projector type: {cfg.projector_type}")
+
+        self.layers = modules
+
+    def forward(self, x):
+        bs, hw, input_dim = x.shape
+        h = w = int((hw)**0.5)
+        """compute padding"""
+        if h % self.cfg.downsample_ratio:
+            pad = self.cfg.downsample_ratio - h % self.cfg.downsample_ratio
+        else:
+            pad = 0
+        x = x.reshape(bs, h, w, input_dim)
+        if pad > 0:
+            x = F.pad(x, (0, 0, 0, pad, 0, pad), "constant", 0)
+        """4 to 1 concat"""
+        x = x.permute(0, 3, 1, 2)  # B, C, H, W
+        x = F.unfold(x,
+                     kernel_size=self.cfg.downsample_ratio,
+                     stride=self.cfg.downsample_ratio,
+                     padding=0)  # B, C*4, HW // 4
+        x = x.permute(0, 2, 1)
+
+        return self.layers(x)
+
+
+class DeepseekVL2ProcessingInfo(BaseProcessingInfo):
+
+    def get_hf_config(self):
+        return self.ctx.get_hf_config(DeepseekVLV2Config)
+
+    def get_hf_processor(self) -> DeepseekVLV2Processor:
+        return self.ctx.get_hf_processor(DeepseekVLV2Processor)
+
+    def get_supported_mm_limits(self) -> Mapping[str, Optional[int]]:
+        return {"image": None}
+
+    def get_num_image_tokens(self, *, image_width: int,
+                             image_height: int) -> int:
+        hf_processor = self.get_hf_processor()
+        image_size = hf_processor.image_size
+        patch_size = hf_processor.patch_size
+        downsample_ratio = hf_processor.downsample_ratio
+
+        best_width, best_height = hf_processor.select_best_resolution(
+            (image_width, image_height))
+
+        num_width_tiles, num_height_tiles = (best_width // image_size,
+                                             best_height // image_size)
+        h = w = math.ceil((image_size // patch_size) / downsample_ratio)
+
+        global_views_tokens = h * (w + 1)
+        local_views_tokens = (num_height_tiles * h) * (num_width_tiles * w + 1)
+        return global_views_tokens + local_views_tokens + 1
+
+    def get_image_size_with_most_features(self) -> ImageSize:
+        hf_config = self.get_hf_config()
+        candidate_resolutions = hf_config.candidate_resolutions
+        height, width = max(candidate_resolutions,
+                            key=lambda x: self.get_num_image_tokens(
+                                image_width=x[1], image_height=x[0]))
+        return ImageSize(width=width, height=height)
+
+    def get_mm_max_tokens_per_item(
+        self,
+        seq_len: int,
+        mm_counts: Mapping[str, int],
+    ) -> Mapping[str, int]:
+        max_image_size = self.get_image_size_with_most_features()
+        max_image_tokens = self.get_num_image_tokens(
+            image_height=max_image_size.height,
+            image_width=max_image_size.width)
+
+        return {"image": max_image_tokens}
+
+
+class DeepseekVL2DummyInputsBuilder(
+        BaseDummyInputsBuilder[DeepseekVL2ProcessingInfo]):
+
+    def get_dummy_processor_inputs(
+        self,
+        seq_len: int,
+        mm_counts: Mapping[str, int],
+    ) -> ProcessorInputs:
+        num_images = mm_counts.get("image", 0)
+        hf_processor = self.info.get_hf_processor()
+        image_token: str = hf_processor.image_token
+
+        max_image_size = self.info.get_image_size_with_most_features()
+
+        mm_data = {
+            "image":
+            self._get_dummy_images(width=max_image_size.width,
+                                   height=max_image_size.height,
+                                   num_images=num_images)
+        }
+
+        return ProcessorInputs(
+            prompt_text=image_token * num_images,
+            mm_data=mm_data,
+        )
+
+
+class DeepseekVL2MultiModalProcessor(
+        BaseMultiModalProcessor[DeepseekVL2ProcessingInfo]):
+
+    def _call_hf_processor(
+        self,
+        prompt: str,
+        mm_data: Mapping[str, object],
+        mm_kwargs: Mapping[str, object],
+    ) -> BatchFeature:
+        if mm_data:
+            processed_outputs = self.info.ctx.call_hf_processor(
+                self.info.get_hf_processor(**mm_kwargs),
+                dict(prompt=prompt, **mm_data),
+                mm_kwargs,
+            )
+            target_dtype = self.info.ctx.model_config.dtype
+            pixel_values = processed_outputs.pop("pixel_values").to(
+                target_dtype)
+            # split pixel values into patches corresponding to each image
+            images_spatial_crop = processed_outputs["images_spatial_crop"]
+            patches_per_image = [
+                x.prod().item() + 1 for x in images_spatial_crop
+            ]
+            pixel_values = pixel_values.split(patches_per_image)
+            processed_outputs["pixel_values"] = pixel_values
+        else:
+            tokenizer = self.info.get_tokenizer()
+            processed_outputs = tokenizer(prompt,
+                                          add_special_tokens=True,
+                                          return_tensors="pt")
+
+        return processed_outputs
+
+    def _get_mm_fields_config(
+        self,
+        hf_inputs: BatchFeature,
+        hf_processor_mm_kwargs: Mapping[str, object],
+    ) -> Mapping[str, MultiModalFieldConfig]:
+        return dict(
+            pixel_values=MultiModalFieldConfig.batched("image"),
+            images_spatial_crop=MultiModalFieldConfig.batched("image"),
+            image_embeds=MultiModalFieldConfig.batched("image"),
+        )
+
+    def _get_prompt_replacements(
+        self,
+        mm_items: MultiModalDataItems,
+        hf_processor_mm_kwargs: Mapping[str, object],
+        out_mm_kwargs: MultiModalKwargs,
+    ) -> list[PromptReplacement]:
+        hf_processor = self.info.get_hf_processor(**hf_processor_mm_kwargs)
+
+        image_token_id = hf_processor.image_token_id
+        assert isinstance(image_token_id, int)
+
+        def get_replacement_deepseek_vl2(item_idx: int):
+            images = mm_items.get_items(
+                "image", (ImageEmbeddingItems, ImageProcessorItems))
+
+            if isinstance(images, ImageEmbeddingItems):
+                num_image_tokens = images.get_feature_size(item_idx)
+            else:
+                image_size = images.get_image_size(item_idx)
+
+                num_image_tokens = self.info.get_num_image_tokens(
+                    image_width=image_size.width,
+                    image_height=image_size.height,
+                )
+            return [image_token_id] * num_image_tokens
+
+        return [
+            PromptReplacement(
+                modality="image",
+                target=[image_token_id],
+                replacement=get_replacement_deepseek_vl2,
+            )
+        ]
+
+
+@MULTIMODAL_REGISTRY.register_processor(
+    DeepseekVL2MultiModalProcessor,
+    info=DeepseekVL2ProcessingInfo,
+    dummy_inputs=DeepseekVL2DummyInputsBuilder)
+class DeepseekVLV2ForCausalLM(nn.Module, SupportsMultiModal, SupportsPP):
+
+    hf_to_vllm_mapper = WeightsMapper(orig_to_new_prefix={
+        "language.": "language_model.",
+    })
+
+    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
+        super().__init__()
+        config: DeepseekVLV2Config = vllm_config.model_config.hf_config
+        quant_config = vllm_config.quant_config
+        multimodal_config = vllm_config.model_config.multimodal_config
+
+        self.config = config
+        self.multimodal_config = multimodal_config
+
+        self.vision_config = config.vision_config
+        self.projector_config = config.projector_config
+        self.text_config = config.text_config
+
+        model_config = vllm_config.model_config
+        tokenizer = cached_get_tokenizer(
+            model_config.tokenizer,
+            tokenizer_mode=model_config.tokenizer_mode,
+            tokenizer_revision=model_config.tokenizer_revision,
+            trust_remote_code=model_config.trust_remote_code,
+        )
+        self.image_token_id = tokenizer.vocab.get(_IMAGE_TOKEN)
+
+        self.vision = self._init_vision_module(self.vision_config,
+                                               quant_config,
+                                               maybe_prefix(prefix, "vision"))
+
+        self.projector = MlpProjector(self.projector_config)
+        self.tile_tag = config.tile_tag
+        self.global_view_pos = config.global_view_pos
+
+        # special token for image token sequence format
+        embed_std = 1 / torch.sqrt(
+            torch.tensor(self.projector_config.n_embed, dtype=torch.float32))
+        if self.tile_tag == "2D":
+            # <|view_separator|>, <|\n|>
+            self.image_newline = nn.Parameter(
+                torch.randn(self.projector_config.n_embed) * embed_std)
+            # This is a typo in original implementation
+            self.view_seperator = nn.Parameter(
+                torch.randn(self.projector_config.n_embed) * embed_std)
+        else:
+            raise ValueError(
+                f"Only 2D tile_tag is supported currently, got: {self.tile_tag}"
+            )
+
+        if self.text_config.topk_method == "noaux_tc":
+            architectures = ["DeepseekV3ForCausalLM"]
+        elif not self.text_config.use_mla:
+            architectures = ["DeepseekForCausalLM"]
+        else:
+            architectures = ["DeepseekV2ForCausalLM"]
+
+        self.language_model = init_vllm_registered_model(
+            vllm_config=vllm_config,
+            hf_config=self.text_config,
+            prefix=maybe_prefix(prefix, "language"),
+            architectures=architectures,
+        )
+
+        self.make_empty_intermediate_tensors = (
+            self.language_model.make_empty_intermediate_tensors)
+
+    def _init_vision_module(
+        self,
+        vision_config: VisionEncoderConfig,
+        quant_config: Optional[QuantizationConfig],
+        prefix: str = "",
+    ) -> nn.Module:
+        # TODO: refactor vision model through timm wrapper from transformers
+        try:
+            import timm
+        except ImportError:
+            raise ImportError("Please install timm") from ImportError
+
+        with set_default_torch_dtype(torch.float16):
+            model = timm.create_model(
+                "vit_so400m_patch14_siglip_384.webli",
+                pretrained=False,
+                num_classes=0,
+                dynamic_img_size=True,
+                dynamic_img_pad=True,
+            )
+
+        model = model.to(dtype=torch.get_default_dtype())
+        return model
+
+    @cached_property
+    def sampler(self):
+        if hasattr(self.language_model, "sampler"):
+            return self.language_model.sampler
+
+        return get_sampler()
+
+    def _validate_pixel_values(
+        self, data: Union[torch.Tensor, List[torch.Tensor]]
+    ) -> Union[torch.Tensor, List[torch.Tensor]]:
+
+        h = w = self.vision_config.image_size
+        expected_dims = (3, h, w)
+
+        def _validate_shape(d: torch.Tensor):
+            actual_dims = tuple(d.shape[1:])
+
+            if actual_dims != expected_dims:
+                expected_expr = ("num_patches", *map(str, expected_dims))
+                raise ValueError(
+                    "The expected shape of pixel values per image per batch "
+                    f"is {expected_expr}. You supplied {tuple(d.shape)}.")
+
+        for d in data:
+            _validate_shape(d)
+
+        return data
+
+    def _validate_images_spatial_crop(
+        self, data: Union[torch.Tensor, List[torch.Tensor]]
+    ) -> Union[torch.Tensor, List[torch.Tensor]]:
+        expected_dims = 2
+
+        def _validate_shape(d: torch.Tensor):
+            actual_dims = d.size(-1)
+
+            if actual_dims != expected_dims:
+                expected_expr = str(expected_dims)
+                raise ValueError(
+                    f"The expected shape of image sizes per image per batch "
+                    f"is {expected_expr}. You supplied {tuple(d.shape)}.")
+
+        for d in data:
+            _validate_shape(d)
+
+        return data
+
+    def _parse_and_validate_image_input(
+            self, **kwargs: object) -> Optional[DeepseekVL2ImageInputs]:
+        pixel_values = kwargs.pop("pixel_values", None)
+        images_spatial_crop = kwargs.pop("images_spatial_crop", None)
+        image_embeds = kwargs.pop("image_embeds", None)
+
+        if pixel_values is None and image_embeds is None:
+            return None
+
+        if pixel_values is not None:
+            if not isinstance(pixel_values, (torch.Tensor, list)):
+                raise ValueError("Incorrect type of pixel values. "
+                                 f"Got type: {type(pixel_values)}")
+
+            if not isinstance(images_spatial_crop, (torch.Tensor, list)):
+                raise ValueError("Incorrect type of image sizes. "
+                                 f"Got type: {type(images_spatial_crop)}")
+
+            return DeepseekVL2ImagePixelInputs(
+                type="pixel_values",
+                data=self._validate_pixel_values(flatten_bn(pixel_values)),
+                images_spatial_crop=self._validate_images_spatial_crop(
+                    flatten_bn(images_spatial_crop, concat=True)))
+
+        if image_embeds is not None:
+            if not isinstance(image_embeds, torch.Tensor):
+                raise ValueError("Incorrect type of image embeddings. "
+                                 f"Got type: {type(image_embeds)}")
+
+            return DeepseekVL2VImageEmbeddingInputs(
+                type="image_embeds",
+                data=flatten_bn(image_embeds),
+            )
+
+        raise AssertionError("This line should be unreachable.")
+
+    def _pixel_values_to_embedding(
+        self,
+        pixel_values: NestedTensors,
+        images_spatial_crop: torch.Tensor,
+    ) -> NestedTensors:
+        # Pixel_values: n_image * batch_size * [patch_per_img, 3, height, width]
+        total_tiles = [x for x in pixel_values]
+
+        # [batch_all_tiles, 3, height, width]
+        total_tiles = torch.cat(total_tiles, dim=0)
+
+        # [batch_all_tiles, vit_seq_len, c]
+        images_feature = self.vision.forward_features(total_tiles)
+
+        # [batch_all_tiles, hw, D]
+        images_embeds = self.projector(images_feature)
+
+        _, hw, n_dim = images_embeds.shape
+        h = w = int(hw**0.5)
+
+        # 根据self.tile_tag & self.global_view_pos填充image token sequence
+        tile_index = 0
+        vision_embeddings = []
+        for jdx in range(images_spatial_crop.size(0)):
+            # extra global & local features
+            num_width_tiles, num_height_tiles = images_spatial_crop[jdx]
+            if num_width_tiles == 0 or num_height_tiles == 0:
+                break
+            num_tiles_in_image = num_width_tiles * num_height_tiles
+
+            # [hw, D]
+            global_features = images_embeds[tile_index]
+
+            # [num_height_tiles * num_width_tiles, hw, D]
+            local_features = images_embeds[tile_index + 1:tile_index + 1 +
+                                           num_tiles_in_image]
+            tile_index += num_tiles_in_image + 1
+
+            # format global and local features
+            # ----------------- global view add newline -----------------
+            # [hw, D] -> [h, w, D]
+            global_features = global_features.view(h, w, n_dim)
+
+            # [D]     -> [h, 1, D]
+            new_lines_in_global = repeat(self.image_newline, "d -> h 1 d", h=h)
+
+            # cat([h, w, D], [h, 1, D], dim=1) -> [h, w + 1, D]
+            global_features = torch.cat([global_features, new_lines_in_global],
+                                        dim=1)
+
+            # [h, w + 1, D] -> [h * (w + 1), D]
+            global_features = global_features.view(-1, n_dim)
+
+            # ----------------- local view add newline -----------------
+            # [num_height_tiles * num_width_tiles, h * w, D] ->
+            # [num_height_tiles * h, num_width_tiles * w, D]
+            local_features = rearrange(local_features,
+                                       "(th tw) (h w) d -> (th h) (tw w) d",
+                                       th=num_height_tiles,
+                                       tw=num_width_tiles,
+                                       h=h,
+                                       w=w)
+
+            # [D] -> [num_height_tiles * h, 1, D]
+            new_lines_in_local = repeat(self.image_newline,
+                                        "d -> (th h) 1 d",
+                                        th=num_height_tiles,
+                                        h=h)
+
+            # [num_height_tiles * h, num_width_tiles * w + 1, D]
+            local_features = torch.cat([local_features, new_lines_in_local],
+                                       dim=1)
+
+            # [num_height_tiles * h, num_width_tiles * w + 1, D]
+            #   --> [(num_height_tiles * h) * (num_width_tiles * w + 1), D]
+            local_features = local_features.view(-1, n_dim)
+
+            # merge global and local tiles
+            if self.global_view_pos == "head":
+                global_local_features = torch.cat([
+                    global_features,
+                    self.view_seperator[None, :],
+                    local_features,
+                ])
+            else:
+                global_local_features = torch.cat([
+                    local_features,
+                    self.view_seperator[None, :],
+                    global_features,
+                ])
+
+            vision_embeddings.append(global_local_features)
+        return vision_embeddings
+
+    def _process_image_input(
+            self, image_input: DeepseekVL2ImageInputs) -> torch.Tensor:
+        if image_input["type"] == "image_embeds":
+            image_data = image_input["data"]
+            if is_list_of(image_data, torch.Tensor):
+                # it's already a list of tensors
+                return image_data
+            if len(image_data.shape) == 3:
+                # 3D tensor
+                return list(torch.unbind(image_data, dim=0))
+            raise ValueError(
+                "We expect batched 2D tensors;"
+                "this can be either a list of 2D tensors or a single 3D tensor."
+            )
+
+        pixel_values = image_input["data"]
+        images_spatial_crop = image_input["images_spatial_crop"]
+
+        return self._pixel_values_to_embedding(
+            pixel_values=pixel_values, images_spatial_crop=images_spatial_crop)
+
+    def get_multimodal_embeddings(self, **kwargs: object) -> torch.Tensor:
+        image_input = self._parse_and_validate_image_input(**kwargs)
+        if image_input is None:
+            return None
+        vision_embeddings = self._process_image_input(image_input)
+        return vision_embeddings
+
+    def get_input_embeddings(
+        self,
+        input_ids: torch.Tensor,
+        multimodal_embeddings: Optional[NestedTensors] = None,
+    ) -> torch.Tensor:
+        inputs_embeds = self.language_model.get_input_embeddings(input_ids)
+        if multimodal_embeddings is not None:
+            inputs_embeds = merge_multimodal_embeddings(
+                input_ids, inputs_embeds, multimodal_embeddings,
+                self.image_token_id)
+        return inputs_embeds
+
+    def forward(self,
+                input_ids: torch.Tensor,
+                positions: torch.Tensor,
+                kv_caches: List[torch.Tensor],
+                attn_metadata: AttentionMetadata,
+                intermediate_tensors: Optional[IntermediateTensors] = None,
+                inputs_embeds: Optional[torch.Tensor] = None,
+                **kwargs: object):
+
+        if intermediate_tensors is not None:
+            inputs_embeds = None
+
+        # NOTE: In v1, inputs_embeds is always generated at model runner, this
+        # condition is for v0 compatibility
+        elif inputs_embeds is None:
+            vision_embeddings = self.get_multimodal_embeddings(**kwargs)
+            inputs_embeds = self.get_input_embeddings(input_ids,
+                                                      vision_embeddings)
+            input_ids = None
+
+        hidden_states = self.language_model(input_ids,
+                                            positions,
+                                            kv_caches,
+                                            attn_metadata,
+                                            intermediate_tensors,
+                                            inputs_embeds=inputs_embeds)
+
+        return hidden_states
+
+    def compute_logits(
+        self,
+        hidden_states: torch.Tensor,
+        sampling_metadata: SamplingMetadata,
+    ) -> Optional[torch.Tensor]:
+        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)
+        autoloaded_weights = loader.load_weights(weights,
+                                                 mapper=self.hf_to_vllm_mapper)
+        return autoloaded_weights

.venv/lib/python3.11/site-packages/vllm/model_executor/models/exaone.py

@@ -0,0 +1,578 @@
+# SPDX-License-Identifier: Apache-2.0
+
+# Adapted from
+# https://huggingface.co/LGAI-EXAONE/EXAONE-3.0-7.8B-Instruct/blob/main/modeling_exaone.py
+# Copyright 2024 The LG U+ CTO AI Tech Lab.
+# Copyright 2021 The LG AI Research EXAONE Lab
+# Copyright (c) 2018, NVIDIA CORPORATION.  All rights reserved.
+#
+# This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX
+# and OPT implementations in this library. It has been modified from its
+# original forms to accommodate minor architectural differences compared
+# to GPT-NeoX and OPT used by the Meta AI team that trained the model.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Inference-only Exaone model compatible with HuggingFace weights."""
+
+from typing import Any, Dict, Iterable, List, Optional, Set, Tuple, Union
+
+import torch
+from torch import nn
+
+from vllm.attention import Attention, AttentionMetadata
+from vllm.compilation.decorators import support_torch_compile
+from vllm.config import CacheConfig, VllmConfig
+from vllm.distributed import get_pp_group, get_tensor_model_parallel_world_size
+from vllm.model_executor.layers.activation import SiluAndMul
+from vllm.model_executor.layers.layernorm import RMSNorm
+from vllm.model_executor.layers.linear import (MergedColumnParallelLinear,
+                                               QKVParallelLinear,
+                                               RowParallelLinear)
+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, maybe_remap_kv_scale_name)
+from vllm.model_executor.sampling_metadata import SamplingMetadata
+from vllm.sequence import IntermediateTensors
+from vllm.transformers_utils.configs.exaone import ExaoneConfig
+
+from .interfaces import SupportsLoRA, SupportsPP
+from .utils import (PPMissingLayer, is_pp_missing_parameter,
+                    make_empty_intermediate_tensors_factory, make_layers,
+                    maybe_prefix)
+
+
+class ExaoneGatedMLP(nn.Module):
+
+    def __init__(
+        self,
+        hidden_size: int,
+        intermediate_size: int,
+        hidden_act: str,
+        quant_config: Optional[QuantizationConfig] = None,
+        bias: bool = False,
+        prefix: str = "",
+    ) -> None:
+        super().__init__()
+        self.gate_up_proj = MergedColumnParallelLinear(
+            input_size=hidden_size,
+            output_sizes=[intermediate_size] * 2,
+            bias=bias,
+            quant_config=quant_config,
+            prefix=f"{prefix}.gate_up_proj",
+        )
+        self.c_proj = RowParallelLinear(
+            input_size=intermediate_size,
+            output_size=hidden_size,
+            bias=bias,
+            quant_config=quant_config,
+            prefix=f"{prefix}.c_proj",
+        )
+        if hidden_act != "silu":
+            raise ValueError(f"Unsupported activation: {hidden_act}. "
+                             "Only silu is supported for now.")
+        self.act_fn = SiluAndMul()
+
+    def forward(self, x):
+        gate_up, _ = self.gate_up_proj(x)
+        x = self.act_fn(gate_up)
+        x, _ = self.c_proj(x)
+        return x
+
+
+class ExaoneAttention(nn.Module):
+
+    def __init__(
+        self,
+        config: ExaoneConfig,
+        hidden_size: int,
+        num_heads: int,
+        num_kv_heads: int,
+        rope_theta: float = 10000,
+        rope_scaling: Optional[Dict[str, Any]] = None,
+        max_position_embeddings: int = 8192,
+        quant_config: Optional[QuantizationConfig] = None,
+        bias: bool = False,
+        cache_config: Optional[CacheConfig] = None,
+        prefix: str = "",
+    ) -> None:
+        super().__init__()
+        self.hidden_size = hidden_size
+        tp_size = get_tensor_model_parallel_world_size()
+        self.total_num_heads = num_heads
+        assert self.total_num_heads % tp_size == 0
+        self.num_heads = self.total_num_heads // tp_size
+        self.total_num_kv_heads = num_kv_heads
+        if self.total_num_kv_heads >= tp_size:
+            # Number of KV heads is greater than TP size, so we partition
+            # the KV heads across multiple tensor parallel GPUs.
+            assert self.total_num_kv_heads % tp_size == 0
+        else:
+            # Number of KV heads is less than TP size, so we replicate
+            # the KV heads across multiple tensor parallel GPUs.
+            assert tp_size % self.total_num_kv_heads == 0
+        self.num_kv_heads = max(1, self.total_num_kv_heads // tp_size)
+        # MistralConfig has an optional head_dim introduced by Mistral-Nemo
+        self.head_dim = getattr(config, "head_dim",
+                                self.hidden_size // self.total_num_heads)
+        self.q_size = self.num_heads * self.head_dim
+        self.kv_size = self.num_kv_heads * self.head_dim
+        self.scaling = self.head_dim**-0.5
+        self.rope_theta = rope_theta
+        self.max_position_embeddings = max_position_embeddings
+
+        self.qkv_proj = QKVParallelLinear(
+            hidden_size=hidden_size,
+            head_size=self.head_dim,
+            total_num_heads=self.total_num_heads,
+            total_num_kv_heads=self.total_num_kv_heads,
+            bias=bias,
+            quant_config=quant_config,
+            prefix=f"{prefix}.qkv_proj",
+        )
+
+        self.out_proj = RowParallelLinear(
+            input_size=self.total_num_heads * self.head_dim,
+            output_size=hidden_size,
+            bias=bias,
+            quant_config=quant_config,
+            prefix=f"{prefix}.out_proj",
+        )
+
+        is_neox_style = True
+        if quant_config is not None and quant_config.get_name() == "gguf":
+            is_neox_style = False
+
+        self.rotary_emb = get_rope(
+            self.head_dim,
+            rotary_dim=self.head_dim,
+            max_position=max_position_embeddings,
+            base=rope_theta,
+            rope_scaling=rope_scaling,
+            is_neox_style=is_neox_style,
+        )
+        self.attn = Attention(
+            self.num_heads,
+            self.head_dim,
+            self.scaling,
+            num_kv_heads=self.num_kv_heads,
+            cache_config=cache_config,
+            quant_config=quant_config,
+            prefix=f"{prefix}.attn",
+        )
+
+    def forward(
+        self,
+        positions: torch.Tensor,
+        hidden_states: torch.Tensor,
+        kv_cache: torch.Tensor,
+        attn_metadata: AttentionMetadata,
+    ) -> torch.Tensor:
+        qkv, _ = self.qkv_proj(hidden_states)
+        q, k, v = qkv.split([self.q_size, self.kv_size, self.kv_size], dim=-1)
+        q, k = self.rotary_emb(positions, q, k)
+        attn_output = self.attn(q, k, v, kv_cache, attn_metadata)
+        output, _ = self.out_proj(attn_output)
+        return output
+
+
+class ExaoneBlockAttention(nn.Module):
+
+    def __init__(
+        self,
+        config: ExaoneConfig,
+        hidden_size: int,
+        num_heads: int,
+        num_kv_heads: int,
+        rope_theta: float = 10000,
+        rope_scaling: Optional[Dict[str, Any]] = None,
+        max_position_embeddings: int = 8192,
+        quant_config: Optional[QuantizationConfig] = None,
+        bias: bool = False,
+        cache_config: Optional[CacheConfig] = None,
+        prefix: str = "",
+    ) -> None:
+        super().__init__()
+        self.attention = ExaoneAttention(
+            config=config,
+            hidden_size=hidden_size,
+            num_heads=num_heads,
+            num_kv_heads=num_kv_heads,
+            rope_theta=rope_theta,
+            rope_scaling=rope_scaling,
+            max_position_embeddings=max_position_embeddings,
+            quant_config=quant_config,
+            bias=bias,
+            cache_config=cache_config,
+            prefix=f"{prefix}.attention",
+        )
+
+    def forward(
+        self,
+        positions: torch.Tensor,
+        hidden_states: torch.Tensor,
+        kv_cache: torch.Tensor,
+        attn_metadata: AttentionMetadata,
+    ) -> torch.Tensor:
+        return self.attention(
+            positions=positions,
+            hidden_states=hidden_states,
+            kv_cache=kv_cache,
+            attn_metadata=attn_metadata,
+        )
+
+
+class ExaoneDecoderLayer(nn.Module):
+
+    def __init__(
+        self,
+        config: ExaoneConfig,
+        cache_config: Optional[CacheConfig] = None,
+        quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
+    ) -> None:
+        super().__init__()
+        self.hidden_size = config.hidden_size
+        rope_theta = getattr(config, "rope_theta", 10000)
+        rope_scaling = getattr(config, "rope_scaling", None)
+        if rope_scaling is not None and getattr(
+                config, "original_max_position_embeddings", None):
+            rope_scaling["original_max_position_embeddings"] = (
+                config.original_max_position_embeddings)
+        max_position_embeddings = getattr(config, "max_position_embeddings",
+                                          8192)
+        # Support abacusai/Smaug-72B-v0.1 with attention_bias
+        # Support internlm/internlm-7b with bias
+        attention_bias = getattr(config, "attention_bias", False) or getattr(
+            config, "bias", False)
+        self.attn = ExaoneBlockAttention(
+            config=config,
+            hidden_size=self.hidden_size,
+            num_heads=config.num_attention_heads,
+            num_kv_heads=getattr(config, "num_key_value_heads",
+                                 config.num_attention_heads),
+            rope_theta=rope_theta,
+            rope_scaling=rope_scaling,
+            max_position_embeddings=max_position_embeddings,
+            quant_config=quant_config,
+            bias=attention_bias,
+            cache_config=cache_config,
+            prefix=f"{prefix}.attn",
+        )
+        self.mlp = ExaoneGatedMLP(
+            hidden_size=self.hidden_size,
+            intermediate_size=config.intermediate_size,
+            hidden_act=config.activation_function,
+            quant_config=quant_config,
+            bias=getattr(config, "mlp_bias", False),
+            prefix=f"{prefix}.mlp",
+        )
+        self.ln_1 = RMSNorm(config.hidden_size, eps=config.layer_norm_epsilon)
+        self.ln_2 = RMSNorm(config.hidden_size, eps=config.layer_norm_epsilon)
+
+    def forward(
+        self,
+        positions: torch.Tensor,
+        hidden_states: torch.Tensor,
+        kv_cache: torch.Tensor,
+        attn_metadata: AttentionMetadata,
+        residual: Optional[torch.Tensor],
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        # Self Attention
+        if residual is None:
+            residual = hidden_states
+            hidden_states = self.ln_1(hidden_states)
+        else:
+            hidden_states, residual = self.ln_1(hidden_states, residual)
+        hidden_states = self.attn(
+            positions=positions,
+            hidden_states=hidden_states,
+            kv_cache=kv_cache,
+            attn_metadata=attn_metadata,
+        )
+
+        # Fully Connected
+        hidden_states, residual = self.ln_2(hidden_states, residual)
+        hidden_states = self.mlp(hidden_states)
+        return hidden_states, residual
+
+
+@support_torch_compile
+class ExaoneModel(nn.Module):
+
+    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
+        super().__init__()
+
+        config = vllm_config.model_config.hf_config
+        cache_config = vllm_config.cache_config
+        quant_config = vllm_config.quant_config
+        lora_config = vllm_config.lora_config
+
+        self.config = config
+        self.padding_idx = config.pad_token_id
+        lora_vocab = ((lora_config.lora_extra_vocab_size *
+                       (lora_config.max_loras or 1)) if lora_config else 0)
+        self.vocab_size = config.vocab_size + lora_vocab
+        self.wte = config.vocab_size
+        if get_pp_group().is_first_rank or (config.tie_word_embeddings
+                                            and get_pp_group().is_last_rank):
+            self.wte = VocabParallelEmbedding(
+                self.vocab_size,
+                config.hidden_size,
+                org_num_embeddings=config.vocab_size,
+                quant_config=quant_config,
+            )
+        else:
+            self.wte = PPMissingLayer()
+        self.start_layer, self.end_layer, self.h = make_layers(
+            config.num_hidden_layers,
+            lambda prefix: ExaoneDecoderLayer(
+                config=config,
+                cache_config=cache_config,
+                quant_config=quant_config,
+                prefix=prefix,
+            ),
+            prefix=f"{prefix}.h",
+        )
+        if get_pp_group().is_last_rank:
+            self.ln_f = RMSNorm(config.hidden_size,
+                                eps=config.layer_norm_epsilon)
+        else:
+            self.ln_f = PPMissingLayer()
+
+        self.make_empty_intermediate_tensors = (
+            make_empty_intermediate_tensors_factory(
+                ["hidden_states", "residual"], config.hidden_size))
+
+    def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
+        return self.wte(input_ids)
+
+    def forward(
+        self,
+        input_ids: Optional[torch.Tensor],
+        positions: torch.Tensor,
+        kv_caches: List[torch.Tensor],
+        attn_metadata: AttentionMetadata,
+        intermediate_tensors: Optional[IntermediateTensors],
+        inputs_embeds: Optional[torch.Tensor] = None,
+    ) -> Union[torch.Tensor, IntermediateTensors]:
+        if get_pp_group().is_first_rank:
+            if inputs_embeds is not None:
+                hidden_states = inputs_embeds
+            else:
+                hidden_states = self.get_input_embeddings(input_ids)
+            residual = None
+        else:
+            assert intermediate_tensors is not None
+            hidden_states = intermediate_tensors["hidden_states"]
+            residual = intermediate_tensors["residual"]
+
+        for i in range(self.start_layer, self.end_layer):
+            layer = self.h[i]
+            hidden_states, residual = layer(
+                positions,
+                hidden_states,
+                kv_caches[i - self.start_layer],
+                attn_metadata,
+                residual,
+            )
+
+        if not get_pp_group().is_last_rank:
+            return IntermediateTensors({
+                "hidden_states": hidden_states,
+                "residual": residual
+            })
+
+        hidden_states, _ = self.ln_f(hidden_states, residual)
+        return hidden_states
+
+
+class ExaoneForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
+    packed_modules_mapping = {
+        "qkv_proj": [
+            "q_proj",
+            "k_proj",
+            "v_proj",
+        ],
+        "gate_up_proj": [
+            "c_fc_0",
+            "c_fc_1",
+        ],
+    }
+
+    # LoRA specific attributes
+    supported_lora_modules = [
+        "qkv_proj",
+        "out_proj",
+        "gate_up_proj",
+        "c_proj",
+        "wte",
+        "lm_head",
+    ]
+    embedding_modules = {
+        "wte": "input_embeddings",
+        "lm_head": "output_embeddings",
+    }
+    embedding_padding_modules = ["lm_head"]
+
+    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
+        super().__init__()
+        config = vllm_config.model_config.hf_config
+        quant_config = vllm_config.quant_config
+        lora_config = vllm_config.lora_config
+
+        self.config = config
+        self.lora_config = lora_config
+        self.quant_config = quant_config
+
+        self.transformer = ExaoneModel(
+            vllm_config=vllm_config,
+            prefix=maybe_prefix(prefix, "model"),
+        )
+        if get_pp_group().is_last_rank:
+            self.unpadded_vocab_size = config.vocab_size
+            if lora_config:
+                self.unpadded_vocab_size += lora_config.lora_extra_vocab_size
+            self.lm_head = ParallelLMHead(
+                self.unpadded_vocab_size,
+                config.hidden_size,
+                org_num_embeddings=config.vocab_size,
+                padding_size=DEFAULT_VOCAB_PADDING_SIZE
+                # We need bigger padding if using lora for kernel
+                # compatibility
+                if not lora_config else lora_config.lora_vocab_padding_size,
+                quant_config=quant_config,
+            )
+            if config.tie_word_embeddings:
+                self.lm_head.weight = self.transformer.wte.weight
+
+            logit_scale = getattr(config, "logit_scale", 1.0)
+            self.logits_processor = LogitsProcessor(self.unpadded_vocab_size,
+                                                    config.vocab_size,
+                                                    logit_scale)
+        else:
+            self.lm_head = PPMissingLayer()
+
+        self.sampler = get_sampler()
+
+        self.make_empty_intermediate_tensors = (
+            self.transformer.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,
+        kv_caches: List[torch.Tensor],
+        attn_metadata: AttentionMetadata,
+        intermediate_tensors: Optional[IntermediateTensors] = None,
+        inputs_embeds: Optional[torch.Tensor] = None,
+    ) -> Union[torch.Tensor, IntermediateTensors]:
+        model_output = self.transformer(input_ids, positions, kv_caches,
+                                        attn_metadata, intermediate_tensors,
+                                        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)
+        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", ".c_fc_0", 0),
+            (".gate_up_proj", ".c_fc_1", 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):
+                # Models trained using ColossalAI may include these tensors in
+                # the checkpoint. Skip them.
+                continue
+            # With tie_word_embeddings, we can skip lm_head.weight
+            # The weight might appear unnecessarily in the files if the model is
+            # processed with quantization, LoRA, fine-tuning, etc.
+            if self.config.tie_word_embeddings and "lm_head.weight" in name:
+                continue
+            if (self.quant_config is not None and
+                (scale_name := self.quant_config.get_cache_scale(name))):
+                # Loading kv cache quantization scales
+                param = params_dict[scale_name]
+                weight_loader = getattr(param, "weight_loader",
+                                        default_weight_loader)
+                loaded_weight = (loaded_weight if loaded_weight.dim() == 0 else
+                                 loaded_weight[0])
+                weight_loader(param, loaded_weight)
+                loaded_params.add(scale_name)
+                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
+                # Remapping the name of FP8 kv-scale.
+                name = maybe_remap_kv_scale_name(name, params_dict)
+                if name is None:
+                    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

.venv/lib/python3.11/site-packages/vllm/model_executor/models/fairseq2_llama.py

@@ -0,0 +1,153 @@
+# SPDX-License-Identifier: Apache-2.0
+
+# Copyright 2024 The vLLM team.
+# Copyright 2024 Meta Platforms, Inc. and affiliates. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Llama model for fairseq2 weights."""
+
+from typing import Iterable, Set, Tuple
+
+import torch
+from torch.nn import Parameter
+
+from vllm.config import VllmConfig
+from vllm.distributed import (get_tensor_model_parallel_rank,
+                              get_tensor_model_parallel_world_size)
+from vllm.model_executor.layers.linear import set_weight_attrs
+from vllm.model_executor.models.llama import LlamaForCausalLM
+
+from .utils import AutoWeightsLoader, WeightsMapper
+
+
+class Fairseq2LlamaForCausalLM(LlamaForCausalLM):
+
+    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
+        super().__init__(vllm_config=vllm_config, prefix=prefix)
+        self.tp_rank = get_tensor_model_parallel_rank()
+        self.tp_size = get_tensor_model_parallel_world_size()
+        # For the model loader to read only the relevant checkpoint files
+        self.allow_patterns_overrides = [
+            # either the full checkpoint
+            "model.pt",
+            # or the tp-sharded checkpoint of the current rank
+            f"model.{self.tp_rank}.pt",
+        ]
+
+    def load_weights(self, weights: Iterable[Tuple[str,
+                                                   torch.Tensor]]) -> Set[str]:
+        # fairseq2's serialization adds a wrapper to usual .pt state_dict's:
+        # { "model_key": my_model_name, "my_model_name": state_dict }
+        # which we first need to unpack
+        weights_wrapped = dict(weights)
+        weights = weights_wrapped[
+            weights_wrapped["model_key"]].items()  # type: ignore
+
+        # remap keys
+        fs2_to_vllm_mapper = WeightsMapper(
+            orig_to_new_prefix={
+                "decoder_frontend.embed.": "model.embed_tokens.",
+                "decoder.": "model.",
+                "final_proj.": "lm_head.",
+            },
+            orig_to_new_substr={
+                ".self_attn_layer_norm.": ".input_layernorm.",
+                ".ffn_layer_norm.": ".post_attention_layernorm.",
+                ".self_attn.output_proj.": ".self_attn.o_proj.",
+                ".ffn.gate_proj.": ".mlp.gate_proj.",
+                ".ffn.inner_proj.": ".mlp.up_proj.",
+                ".ffn.output_proj.": ".mlp.down_proj.",
+                ".layer_norm.": ".norm.",
+            },
+        )
+        weights = fs2_to_vllm_mapper.apply(weights)
+
+        params = dict(self.named_parameters())
+
+        loader = AutoWeightsLoader(
+            self,
+            skip_prefixes=(["lm_head."]
+                           if self.config.tie_word_embeddings else None),
+        )
+        return loader.load_weights(
+            (self.reshape_fairseq2_weights(name, loaded_weight, params)
+             for name, loaded_weight in weights))
+
+    def flag_sharded_weights(self, params: dict[str, Parameter]):
+        """Sets the `is_sharded_weight` flag to True for all sharded weights"""
+        for name, param in params.items():
+            modules = name.split(".")
+            if "norm" in name and len(param.size()) < 2:
+                # layer norms are not sharded
+                continue
+            elif any(emb in modules for emb in ["embed_tokens", "lm_head"]):
+                # for now we repeat embedding layers for compatibility
+                continue
+            else:
+                # all other layers are sharded
+                set_weight_attrs(param, {"is_sharded_weight": True})
+
+    def reshape_fairseq2_weights(
+        self,
+        name: str,
+        loaded_weight: torch.Tensor,
+        params: dict[str, Parameter],
+    ) -> Tuple[str, torch.Tensor]:
+        """Reshape fairseq2's weights."""
+
+        def permute(w: torch.Tensor, n_heads: int) -> torch.Tensor:
+            attn_in = self.config.head_dim * n_heads
+            # check for a sharded weight on dim 0
+            if attn_in // self.tp_size == w.size()[0]:
+                attn_in //= self.tp_size
+                n_heads //= self.tp_size
+            attn_out = self.config.hidden_size
+            return (w.view(n_heads, attn_in // n_heads // 2, 2,
+                           attn_out).transpose(1,
+                                               2).reshape(attn_in, attn_out))
+
+        modules = name.split(".")
+
+        # rotary embeds should be sliced
+        if "k_proj" in modules:
+            loaded_weight = permute(loaded_weight,
+                                    self.config.num_key_value_heads)
+
+        elif "q_proj" in modules:
+            loaded_weight = permute(loaded_weight,
+                                    self.config.num_attention_heads)
+
+        # We make the loaded weights compatible with both
+        # full checkpoints and tp sharded checkpoints.
+        # Embeddings are repeated to fit the vocab size.
+        # Other weights are flagged for the weight_loader calls.
+        if any(emb in modules for emb in ["embed_tokens", "lm_head"]):
+            # Embeddings are sharded on dim 0
+            dim = 0
+            # In fairseq2, vocab size has to be divisible by tp_size
+            # so we don't worry about padding
+            if self.tp_size > 1 and loaded_weight.shape[
+                    dim] < self.config.vocab_size:
+                assert loaded_weight.shape[
+                    dim] * self.tp_size == self.config.vocab_size, \
+                        "vocab_size should be divisible by tp_size."
+                repeats = [1] * len(loaded_weight.size())
+                repeats[dim] = self.tp_size
+                # repeat to match vocab size and to be easily 'narrow'able
+                loaded_weight = loaded_weight.repeat(repeats)
+                set_weight_attrs(params[name], {"is_sharded_weight": False})
+                # if embeddings are sharded, the rest is too
+                if "embed_tokens" in modules:
+                    self.flag_sharded_weights(params)
+
+        return name, loaded_weight

.venv/lib/python3.11/site-packages/vllm/model_executor/models/gemma2.py

@@ -0,0 +1,463 @@
+# SPDX-License-Identifier: Apache-2.0
+
+# Copyright 2024 The vLLM team.
+# Copyright 2024 Google Inc. HuggingFace Inc. team. All rights reserved.
+#
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from typing import Iterable, List, Optional, Set, Tuple, Union
+
+import torch
+from torch import nn
+from transformers import Gemma2Config
+
+from vllm.attention import Attention, AttentionMetadata
+from vllm.compilation.decorators import support_torch_compile
+from vllm.config import CacheConfig, VllmConfig
+from vllm.distributed import get_pp_group, get_tensor_model_parallel_world_size
+from vllm.logger import init_logger
+from vllm.model_executor.layers.activation import GeluAndMul
+from vllm.model_executor.layers.layernorm import GemmaRMSNorm
+from vllm.model_executor.layers.linear import (MergedColumnParallelLinear,
+                                               QKVParallelLinear,
+                                               RowParallelLinear)
+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 (
+    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
+from .utils import (AutoWeightsLoader, extract_layer_index,
+                    is_pp_missing_parameter,
+                    make_empty_intermediate_tensors_factory, make_layers,
+                    maybe_prefix)
+
+logger = init_logger(__name__)
+
+
+class Gemma2MLP(nn.Module):
+
+    def __init__(
+        self,
+        hidden_size: int,
+        intermediate_size: int,
+        hidden_act: str,
+        hidden_activation: str,
+        quant_config: Optional[QuantizationConfig] = None,
+    ) -> None:
+        super().__init__()
+        self.gate_up_proj = MergedColumnParallelLinear(
+            hidden_size, [intermediate_size] * 2,
+            bias=False,
+            quant_config=quant_config)
+        self.down_proj = RowParallelLinear(intermediate_size,
+                                           hidden_size,
+                                           bias=False,
+                                           quant_config=quant_config)
+        if not (hidden_act == hidden_activation == "gelu_pytorch_tanh"):
+            raise ValueError(
+                "Gemma2 uses `gelu_pytorch_tanh` as the hidden activation "
+                "function. Please set `hidden_act` and `hidden_activation` to "
+                "`gelu_pytorch_tanh`.")
+        self.act_fn = GeluAndMul(approximate="tanh")
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        gate_up, _ = self.gate_up_proj(x)
+        x = self.act_fn(gate_up)
+        x, _ = self.down_proj(x)
+        return x
+
+
+class Gemma2Attention(nn.Module):
+
+    def __init__(self,
+                 config: Gemma2Config,
+                 hidden_size: int,
+                 num_heads: int,
+                 num_kv_heads: int,
+                 head_dim: int,
+                 max_position_embeddings: int,
+                 rope_theta: float,
+                 cache_config: Optional[CacheConfig] = None,
+                 quant_config: Optional[QuantizationConfig] = None,
+                 attn_logits_soft_cap: Optional[float] = None,
+                 prefix: str = "") -> None:
+        super().__init__()
+        self.config = config
+        self.hidden_size = hidden_size
+        tp_size = get_tensor_model_parallel_world_size()
+        self.total_num_heads = num_heads
+        assert self.total_num_heads % tp_size == 0
+        self.num_heads = self.total_num_heads // tp_size
+        self.total_num_kv_heads = num_kv_heads
+        if self.total_num_kv_heads >= tp_size:
+            # Number of KV heads is greater than TP size, so we partition
+            # the KV heads across multiple tensor parallel GPUs.
+            assert self.total_num_kv_heads % tp_size == 0
+        else:
+            # Number of KV heads is less than TP size, so we replicate
+            # the KV heads across multiple tensor parallel GPUs.
+            assert tp_size % self.total_num_kv_heads == 0
+        self.num_kv_heads = max(1, self.total_num_kv_heads // tp_size)
+        self.head_dim = head_dim
+        self.q_size = self.num_heads * self.head_dim
+        self.kv_size = self.num_kv_heads * self.head_dim
+        self.scaling = config.query_pre_attn_scalar**-0.5
+        self.rope_theta = rope_theta
+
+        self.qkv_proj = QKVParallelLinear(
+            hidden_size,
+            self.head_dim,
+            self.total_num_heads,
+            self.total_num_kv_heads,
+            bias=config.attention_bias,
+            quant_config=quant_config,
+        )
+        self.o_proj = RowParallelLinear(
+            self.total_num_heads * self.head_dim,
+            hidden_size,
+            bias=config.attention_bias,
+            quant_config=quant_config,
+        )
+        self.rotary_emb = get_rope(
+            self.head_dim,
+            rotary_dim=self.head_dim,
+            max_position=max_position_embeddings,
+            base=self.rope_theta,
+            is_neox_style=True,
+        )
+
+        # reference:
+        # https://github.com/huggingface/transformers/blob/54be2d7ae87e873482b984cc956e165ca4dc0ba3/src/transformers/models/gemma2/modeling_gemma2.py#L312 # noqa
+        layer_idx = extract_layer_index(prefix)
+        use_sliding_window = (layer_idx % 2 == 0 and
+                              config.interleaved_sliding_window is not None)
+        sliding_window = config.interleaved_sliding_window if \
+            use_sliding_window else None
+        self.attn = Attention(self.num_heads,
+                              self.head_dim,
+                              self.scaling,
+                              num_kv_heads=self.num_kv_heads,
+                              cache_config=cache_config,
+                              quant_config=quant_config,
+                              logits_soft_cap=attn_logits_soft_cap,
+                              per_layer_sliding_window=sliding_window,
+                              prefix=f"{prefix}.attn")
+
+    def forward(
+        self,
+        positions: torch.Tensor,
+        hidden_states: torch.Tensor,
+        kv_cache: torch.Tensor,
+        attn_metadata: AttentionMetadata,
+    ) -> torch.Tensor:
+        qkv, _ = self.qkv_proj(hidden_states)
+        q, k, v = qkv.split([self.q_size, self.kv_size, self.kv_size], dim=-1)
+        q, k = self.rotary_emb(positions, q, k)
+        attn_output = self.attn(q, k, v, kv_cache, attn_metadata)
+        output, _ = self.o_proj(attn_output)
+        return output
+
+
+class Gemma2DecoderLayer(nn.Module):
+
+    def __init__(
+        self,
+        config: Gemma2Config,
+        cache_config: Optional[CacheConfig] = None,
+        quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
+    ) -> None:
+        super().__init__()
+        self.hidden_size = config.hidden_size
+        self.self_attn = Gemma2Attention(
+            config=config,
+            hidden_size=self.hidden_size,
+            num_heads=config.num_attention_heads,
+            num_kv_heads=config.num_key_value_heads,
+            head_dim=config.head_dim,
+            max_position_embeddings=config.max_position_embeddings,
+            rope_theta=config.rope_theta,
+            cache_config=cache_config,
+            quant_config=quant_config,
+            attn_logits_soft_cap=config.attn_logit_softcapping,
+            prefix=f"{prefix}.self_attn",
+        )
+        self.hidden_size = config.hidden_size
+        self.mlp = Gemma2MLP(
+            hidden_size=self.hidden_size,
+            intermediate_size=config.intermediate_size,
+            hidden_act=config.hidden_act,
+            hidden_activation=config.hidden_activation,
+            quant_config=quant_config,
+        )
+        self.input_layernorm = GemmaRMSNorm(config.hidden_size,
+                                            eps=config.rms_norm_eps)
+        self.post_attention_layernorm = GemmaRMSNorm(config.hidden_size,
+                                                     eps=config.rms_norm_eps)
+        self.pre_feedforward_layernorm = GemmaRMSNorm(config.hidden_size,
+                                                      eps=config.rms_norm_eps)
+        self.post_feedforward_layernorm = GemmaRMSNorm(config.hidden_size,
+                                                       eps=config.rms_norm_eps)
+
+    def forward(
+        self,
+        positions: torch.Tensor,
+        hidden_states: torch.Tensor,
+        kv_cache: torch.Tensor,
+        attn_metadata: AttentionMetadata,
+        residual: Optional[torch.Tensor],
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        if residual is None:
+            residual = hidden_states
+            hidden_states = self.input_layernorm(hidden_states)
+        else:
+            hidden_states, residual = self.input_layernorm(
+                hidden_states, residual)
+        hidden_states = self.self_attn(
+            positions=positions,
+            hidden_states=hidden_states,
+            kv_cache=kv_cache,
+            attn_metadata=attn_metadata,
+        )
+        hidden_states = self.post_attention_layernorm(hidden_states)
+
+        hidden_states, residual = self.pre_feedforward_layernorm(
+            hidden_states, residual)
+        hidden_states = self.mlp(hidden_states)
+        hidden_states = self.post_feedforward_layernorm(hidden_states)
+        return hidden_states, residual
+
+
+@support_torch_compile
+class Gemma2Model(nn.Module):
+
+    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
+        super().__init__()
+        config = vllm_config.model_config.hf_config
+        cache_config = vllm_config.cache_config
+        quant_config = vllm_config.quant_config
+        self.config = config
+        self.quant_config = quant_config
+
+        self.embed_tokens = VocabParallelEmbedding(
+            config.vocab_size,
+            config.hidden_size,
+        )
+        self.start_layer, self.end_layer, self.layers = make_layers(
+            config.num_hidden_layers,
+            lambda prefix: Gemma2DecoderLayer(
+                config, cache_config, quant_config, prefix=prefix),
+            prefix=f"{prefix}.layers")
+        self.norm = GemmaRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+
+        # Normalize the embedding by sqrt(hidden_size)
+        # The normalizer's data type should be downcasted to the model's
+        # data type such as bfloat16, not float32.
+        # See https://github.com/huggingface/transformers/pull/29402
+        normalizer = self.config.hidden_size**0.5
+        self.register_buffer("normalizer", torch.tensor(normalizer))
+        self.make_empty_intermediate_tensors = (
+            make_empty_intermediate_tensors_factory(
+                ["hidden_states", "residual"], config.hidden_size))
+
+    def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
+        return self.embed_tokens(input_ids)
+
+    def forward(
+        self,
+        input_ids: Optional[torch.Tensor],
+        positions: torch.Tensor,
+        kv_caches: List[torch.Tensor],
+        attn_metadata: AttentionMetadata,
+        intermediate_tensors: Optional[IntermediateTensors],
+        inputs_embeds: Optional[torch.Tensor] = None,
+    ) -> Union[torch.Tensor, IntermediateTensors]:
+        if get_pp_group().is_first_rank:
+            if inputs_embeds is not None:
+                hidden_states = inputs_embeds
+            else:
+                hidden_states = self.get_input_embeddings(input_ids)
+            hidden_states *= self.normalizer
+            residual = None
+        else:
+            assert intermediate_tensors is not None
+            hidden_states = intermediate_tensors["hidden_states"]
+            residual = intermediate_tensors["residual"]
+        for i in range(self.start_layer, self.end_layer):
+            layer = self.layers[i]
+            hidden_states, residual = layer(
+                positions,
+                hidden_states,
+                kv_caches[i - self.start_layer],
+                attn_metadata,
+                residual,
+            )
+        if not get_pp_group().is_last_rank:
+            return IntermediateTensors({
+                "hidden_states": hidden_states,
+                "residual": residual
+            })
+        hidden_states, _ = self.norm(hidden_states, residual)
+        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:
+            if (self.quant_config is not None and
+                (scale_name := self.quant_config.get_cache_scale(name))):
+                # Loading kv cache scales for compressed-tensors quantization
+                param = params_dict[scale_name]
+                weight_loader = getattr(param, "weight_loader",
+                                        default_weight_loader)
+                loaded_weight = loaded_weight[0]
+                weight_loader(param, loaded_weight)
+                loaded_params.add(scale_name)
+                continue
+            for (param_name, shard_name, shard_id) in stacked_params_mapping:
+                if shard_name not in name:
+                    continue
+                name = name.replace(shard_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
+                # Remapping the name of FP8 kv-scale.
+                name = maybe_remap_kv_scale_name(name, params_dict)
+                if name is None:
+                    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)
+
+        unloaded_params = params_dict.keys() - loaded_params
+        if unloaded_params:
+            logger.warning(
+                "Some weights are not initialized from checkpoints: %s",
+                unloaded_params)
+        return loaded_params
+
+
+class Gemma2ForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
+    packed_modules_mapping = {
+        "qkv_proj": [
+            "q_proj",
+            "k_proj",
+            "v_proj",
+        ],
+        "gate_up_proj": [
+            "gate_proj",
+            "up_proj",
+        ],
+    }
+
+    # LoRA specific attributes
+    supported_lora_modules = [
+        "qkv_proj",
+        "o_proj",
+        "gate_up_proj",
+        "down_proj",
+    ]
+    # Gemma does not apply LoRA to the embedding layer.
+    embedding_modules = {}
+    embedding_padding_modules = []
+
+    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
+        config = vllm_config.model_config.hf_config
+        quant_config = vllm_config.quant_config
+        lora_config = vllm_config.lora_config
+        del lora_config  # Unused.
+        super().__init__()
+        self.config = config
+        # currently all existing Gemma models have `tie_word_embeddings` enabled
+        assert config.tie_word_embeddings
+        self.quant_config = quant_config
+        self.model = Gemma2Model(vllm_config=vllm_config,
+                                 prefix=maybe_prefix(prefix, "model"))
+        self.logits_processor = LogitsProcessor(
+            config.vocab_size, soft_cap=config.final_logit_softcapping)
+        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,
+        kv_caches: List[torch.Tensor],
+        attn_metadata: AttentionMetadata,
+        intermediate_tensors: Optional[IntermediateTensors] = None,
+        inputs_embeds: Optional[torch.Tensor] = None,
+    ) -> Union[torch.Tensor, IntermediateTensors]:
+        hidden_states = self.model(input_ids, positions, kv_caches,
+                                   attn_metadata, intermediate_tensors,
+                                   inputs_embeds)
+        return hidden_states
+
+    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)
+        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,
+            skip_prefixes=(["lm_head."]
+                           if self.config.tie_word_embeddings else None),
+        )
+        return loader.load_weights(weights)

.venv/lib/python3.11/site-packages/vllm/model_executor/models/glm.py

@@ -0,0 +1,22 @@
+# SPDX-License-Identifier: Apache-2.0
+"""Inference-only HF format GLM-4 model compatible with THUDM weights."""
+from vllm.config import VllmConfig
+from vllm.model_executor.models.llama import LlamaForCausalLM
+
+from .utils import PPMissingLayer
+
+
+class GlmForCausalLM(LlamaForCausalLM):
+
+    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
+        super().__init__(vllm_config=vllm_config, prefix=prefix)
+        # Hack Llama model to fit HF format GLM implementation
+        # Attention difference between GLM and Llama:
+        # 1. Half partial rotary_dim and no Neox style.
+        # 2. There is no bias for o_proj in attention
+        for layer in self.model.layers:
+            if not isinstance(layer, PPMissingLayer):
+                layer.self_attn.rotary_emb.rotary_dim //= 2
+                layer.self_attn.rotary_emb.is_neox_style = False
+                layer.self_attn.o_proj.bias = None
+                layer.self_attn.o_proj.skip_bias_add = True

.venv/lib/python3.11/site-packages/vllm/model_executor/models/gpt_j.py

@@ -0,0 +1,358 @@
+# SPDX-License-Identifier: Apache-2.0
+
+# Adapted from
+# https://github.com/huggingface/transformers/blob/v4.28.0/src/transformers/models/gptj/modeling_gptj.py
+# Copyright 2023 The vLLM team.
+# Copyright 2021 The EleutherAI and HuggingFace Teams. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Inference-only GPT-J model compatible with HuggingFace weights."""
+from typing import Iterable, List, Optional, Set, Tuple, Union
+
+import torch
+from torch import nn
+from transformers import GPTJConfig
+
+from vllm.attention import Attention, AttentionMetadata
+from vllm.compilation.decorators import support_torch_compile
+from vllm.config import CacheConfig, VllmConfig
+from vllm.distributed import get_pp_group, get_tensor_model_parallel_world_size
+from vllm.model_executor.layers.activation import get_act_fn
+from vllm.model_executor.layers.linear import (ColumnParallelLinear,
+                                               QKVParallelLinear,
+                                               RowParallelLinear)
+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, maybe_remap_kv_scale_name)
+from vllm.model_executor.sampling_metadata import SamplingMetadata
+from vllm.sequence import IntermediateTensors
+
+from .interfaces import SupportsPP
+from .utils import (is_pp_missing_parameter,
+                    make_empty_intermediate_tensors_factory, make_layers,
+                    maybe_prefix)
+
+
+class GPTJAttention(nn.Module):
+
+    def __init__(
+        self,
+        config: GPTJConfig,
+        cache_config: Optional[CacheConfig] = None,
+        quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
+    ):
+        super().__init__()
+        self.total_num_heads = config.num_attention_heads
+        self.hidden_size = config.hidden_size
+        self.head_size = self.hidden_size // self.total_num_heads
+
+        self.qkv_proj = QKVParallelLinear(
+            config.hidden_size,
+            self.head_size,
+            self.total_num_heads,
+            bias=False,
+            quant_config=quant_config,
+        )
+        self.out_proj = RowParallelLinear(
+            config.hidden_size,
+            config.hidden_size,
+            bias=False,
+            quant_config=quant_config,
+        )
+
+        tp_world_size = get_tensor_model_parallel_world_size()
+        assert self.total_num_heads % tp_world_size == 0
+        self.num_heads = self.total_num_heads // tp_world_size
+
+        scaling = self.head_size**-0.5
+        assert getattr(config, "rotary", True)
+        assert config.rotary_dim % 2 == 0
+        rope_theta = getattr(config, "rope_theta", 10000)
+        max_position_embeddings = getattr(config, "max_position_embeddings",
+                                          8192)
+        self.rotary_emb = get_rope(
+            self.head_size,
+            rotary_dim=config.rotary_dim,
+            max_position=max_position_embeddings,
+            base=rope_theta,
+            is_neox_style=False,
+        )
+        self.attn = Attention(self.num_heads,
+                              self.head_size,
+                              scaling,
+                              cache_config=cache_config,
+                              quant_config=quant_config,
+                              prefix=f"{prefix}.attn")
+
+    def forward(
+        self,
+        position_ids: torch.Tensor,
+        hidden_states: torch.Tensor,
+        kv_cache: torch.Tensor,
+        attn_metadata: AttentionMetadata,
+    ) -> torch.Tensor:
+        qkv, _ = self.qkv_proj(hidden_states)
+        q, k, v = qkv.chunk(chunks=3, dim=-1)
+        q, k = self.rotary_emb(position_ids, q, k)
+        attn_output = self.attn(q, k, v, kv_cache, attn_metadata)
+        attn_output, _ = self.out_proj(attn_output)
+        return attn_output
+
+
+class GPTJMLP(nn.Module):
+
+    def __init__(
+        self,
+        intermediate_size: int,
+        config: GPTJConfig,
+        quant_config: Optional[QuantizationConfig] = None,
+    ):
+        super().__init__()
+        hidden_size = config.n_embd
+        self.fc_in = ColumnParallelLinear(
+            hidden_size,
+            intermediate_size,
+            quant_config=quant_config,
+        )
+        self.fc_out = RowParallelLinear(
+            intermediate_size,
+            hidden_size,
+            quant_config=quant_config,
+        )
+        self.act = get_act_fn(config.activation_function)
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        hidden_states, _ = self.fc_in(hidden_states)
+        hidden_states = self.act(hidden_states)
+        hidden_states, _ = self.fc_out(hidden_states)
+        return hidden_states
+
+
+class GPTJBlock(nn.Module):
+
+    def __init__(
+        self,
+        config: GPTJConfig,
+        cache_config: Optional[CacheConfig] = None,
+        quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
+    ):
+        super().__init__()
+        inner_dim = (4 * config.n_embd
+                     if config.n_inner is None else config.n_inner)
+        self.ln_1 = nn.LayerNorm(config.n_embd, eps=config.layer_norm_epsilon)
+        self.attn = GPTJAttention(config,
+                                  cache_config,
+                                  quant_config,
+                                  prefix=f"{prefix}.attn")
+        self.mlp = GPTJMLP(inner_dim, config, quant_config)
+
+    def forward(
+        self,
+        position_ids: torch.Tensor,
+        hidden_states: torch.Tensor,
+        kv_cache: torch.Tensor,
+        attn_metadata: AttentionMetadata,
+    ) -> torch.Tensor:
+        residual = hidden_states
+        hidden_states = self.ln_1(hidden_states)
+        attn_output = self.attn(
+            position_ids=position_ids,
+            hidden_states=hidden_states,
+            kv_cache=kv_cache,
+            attn_metadata=attn_metadata,
+        )
+        mlp_output = self.mlp(hidden_states)
+        hidden_states = attn_output + mlp_output + residual
+        return hidden_states
+
+
+@support_torch_compile
+class GPTJModel(nn.Module):
+
+    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
+        super().__init__()
+
+        config = vllm_config.model_config.hf_config
+        cache_config = vllm_config.cache_config
+        quant_config = vllm_config.quant_config
+
+        self.config = config
+        self.embed_dim = config.n_embd
+        self.wte = VocabParallelEmbedding(
+            config.vocab_size,
+            self.embed_dim,
+        )
+        self.start_layer, self.end_layer, self.h = make_layers(
+            config.n_layer,
+            lambda prefix: GPTJBlock(
+                config, cache_config, quant_config, prefix=prefix),
+            prefix=f"{prefix}.h",
+        )
+        self.ln_f = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_epsilon)
+        self.make_empty_intermediate_tensors = (
+            make_empty_intermediate_tensors_factory(["hidden_states"],
+                                                    config.n_embd))
+
+    def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
+        return self.wte(input_ids)
+
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        position_ids: torch.Tensor,
+        kv_caches: List[torch.Tensor],
+        attn_metadata: AttentionMetadata,
+        intermediate_tensors: Optional[IntermediateTensors],
+        inputs_embeds: Optional[torch.Tensor] = None,
+    ) -> Union[torch.Tensor, IntermediateTensors]:
+        if get_pp_group().is_first_rank:
+            if inputs_embeds is not None:
+                hidden_states = inputs_embeds
+            else:
+                hidden_states = self.get_input_embeddings(input_ids)
+        else:
+            hidden_states = intermediate_tensors["hidden_states"]
+        for i in range(self.start_layer, self.end_layer):
+            layer = self.h[i]
+            hidden_states = layer(
+                position_ids,
+                hidden_states,
+                kv_caches[i - self.start_layer],
+                attn_metadata,
+            )
+        if not get_pp_group().is_last_rank:
+            return IntermediateTensors({"hidden_states": hidden_states})
+        hidden_states = self.ln_f(hidden_states)
+        return hidden_states
+
+
+class GPTJForCausalLM(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
+        assert not config.tie_word_embeddings
+        self.transformer = GPTJModel(vllm_config=vllm_config,
+                                     prefix=maybe_prefix(
+                                         prefix, "transformer"))
+        self.lm_head = ParallelLMHead(
+            config.vocab_size,
+            config.n_embd,
+            bias=True,
+            quant_config=quant_config,
+        )
+        self.logits_processor = LogitsProcessor(config.vocab_size)
+        self.sampler = get_sampler()
+        self.make_empty_intermediate_tensors = (
+            self.transformer.make_empty_intermediate_tensors)
+
+    def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
+        return self.transformer.get_input_embeddings(input_ids)
+
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        positions: torch.Tensor,
+        kv_caches: List[torch.Tensor],
+        attn_metadata: AttentionMetadata,
+        intermediate_tensors: Optional[IntermediateTensors] = None,
+        inputs_embeds: Optional[torch.Tensor] = None,
+    ) -> Union[torch.Tensor, IntermediateTensors]:
+        hidden_states = self.transformer(input_ids, positions, kv_caches,
+                                         attn_metadata, intermediate_tensors,
+                                         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, 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]:
+        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 "attn.bias" in name or "attn.masked_bias" in name:
+                continue
+
+            if (self.quant_config is not None and
+                (scale_name := self.quant_config.get_cache_scale(name))):
+                # Loading kv cache quantization scales
+                param = params_dict[scale_name]
+                weight_loader = getattr(param, "weight_loader",
+                                        default_weight_loader)
+                loaded_weight = (loaded_weight if loaded_weight.dim() == 0 else
+                                 loaded_weight[0])
+                weight_loader(param, loaded_weight)
+                loaded_params.add(scale_name)
+                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:
+                name = maybe_remap_kv_scale_name(name, params_dict)
+                if name is None:
+                    continue
+                # 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

.venv/lib/python3.11/site-packages/vllm/model_executor/models/gpt_neox.py

@@ -0,0 +1,352 @@
+# SPDX-License-Identifier: Apache-2.0
+
+# Adapted from
+# https://github.com/huggingface/transformers/blob/v4.28.0/src/transformers/models/gpt_neox/modeling_gpt_neox.py
+# Copyright 2023 The vLLM team.
+# Copyright 2022 EleutherAI The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Inference-only GPT-NeoX model compatible with HuggingFace weights."""
+from typing import Iterable, List, Optional, Set, Tuple, Union
+
+import torch
+from torch import nn
+from transformers import GPTNeoXConfig
+
+from vllm.attention import Attention, AttentionMetadata
+from vllm.compilation.decorators import support_torch_compile
+from vllm.config import CacheConfig, VllmConfig
+from vllm.distributed import get_pp_group, get_tensor_model_parallel_world_size
+from vllm.model_executor.layers.activation import get_act_fn
+from vllm.model_executor.layers.linear import (ColumnParallelLinear,
+                                               QKVParallelLinear,
+                                               RowParallelLinear)
+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
+from vllm.model_executor.sampling_metadata import SamplingMetadata
+from vllm.sequence import IntermediateTensors
+
+from .interfaces import SupportsPP
+from .utils import (is_pp_missing_parameter,
+                    make_empty_intermediate_tensors_factory, make_layers,
+                    maybe_prefix)
+
+
+class GPTNeoXAttention(nn.Module):
+
+    def __init__(
+        self,
+        config: GPTNeoXConfig,
+        cache_config: Optional[CacheConfig] = None,
+        quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
+    ):
+        super().__init__()
+        self.total_num_heads = config.num_attention_heads
+        self.hidden_size = config.hidden_size
+        self.head_size = self.hidden_size // self.total_num_heads
+        self.bias = getattr(config, "attention_bias", True)
+
+        tensor_model_parallel_world_size = (
+            get_tensor_model_parallel_world_size())
+        assert self.total_num_heads % tensor_model_parallel_world_size == 0
+        self.num_heads = (self.total_num_heads //
+                          tensor_model_parallel_world_size)
+
+        self.query_key_value = QKVParallelLinear(
+            config.hidden_size,
+            self.head_size,
+            self.total_num_heads,
+            bias=self.bias,
+            quant_config=quant_config,
+        )
+        self.dense = RowParallelLinear(
+            config.hidden_size,
+            config.hidden_size,
+            bias=self.bias,
+            quant_config=quant_config,
+        )
+        scaling = self.head_size**-0.5
+        rotary_dim = int(self.head_size * config.rotary_pct)
+        assert rotary_dim % 2 == 0
+        rope_theta = getattr(config, "rope_theta", 10000)
+        max_position_embeddings = getattr(config, "max_position_embeddings",
+                                          8192)
+        self.rotary_emb = get_rope(
+            self.head_size,
+            rotary_dim=rotary_dim,
+            max_position=max_position_embeddings,
+            base=rope_theta,
+        )
+        self.attn = Attention(self.num_heads,
+                              self.head_size,
+                              scaling,
+                              cache_config=cache_config,
+                              quant_config=quant_config,
+                              prefix=f"{prefix}.attn")
+
+    def forward(
+        self,
+        position_ids: torch.Tensor,
+        hidden_states: torch.Tensor,
+        kv_cache: torch.Tensor,
+        attn_metadata: AttentionMetadata,
+    ) -> torch.Tensor:
+        qkv, _ = self.query_key_value(hidden_states)
+        q, k, v = qkv.chunk(chunks=3, dim=-1)
+        q, k = self.rotary_emb(position_ids, q, k)
+        attn_output = self.attn(q, k, v, kv_cache, attn_metadata)
+        output, _ = self.dense(attn_output)
+        return output
+
+
+class GPTNeoXMLP(nn.Module):
+
+    def __init__(
+        self,
+        config: GPTNeoXConfig,
+        quant_config: Optional[QuantizationConfig] = None,
+    ):
+        super().__init__()
+        self.dense_h_to_4h = ColumnParallelLinear(
+            config.hidden_size,
+            config.intermediate_size,
+            quant_config=quant_config,
+        )
+        self.dense_4h_to_h = RowParallelLinear(
+            config.intermediate_size,
+            config.hidden_size,
+            quant_config=quant_config,
+        )
+        self.act = get_act_fn(config.hidden_act)
+
+    def forward(self, hidden_states):
+        hidden_states, _ = self.dense_h_to_4h(hidden_states)
+        hidden_states = self.act(hidden_states)
+        hidden_states, _ = self.dense_4h_to_h(hidden_states)
+        return hidden_states
+
+
+class GPTNeoXLayer(nn.Module):
+
+    def __init__(
+        self,
+        config: GPTNeoXConfig,
+        cache_config: Optional[CacheConfig] = None,
+        quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
+    ):
+        super().__init__()
+        self.use_parallel_residual = config.use_parallel_residual
+        self.input_layernorm = nn.LayerNorm(config.hidden_size,
+                                            eps=config.layer_norm_eps)
+        self.post_attention_layernorm = nn.LayerNorm(config.hidden_size,
+                                                     eps=config.layer_norm_eps)
+        self.attention = GPTNeoXAttention(config,
+                                          cache_config,
+                                          quant_config,
+                                          prefix=f"{prefix}.attention")
+        self.mlp = GPTNeoXMLP(config, quant_config)
+
+    def forward(
+        self,
+        position_ids: torch.Tensor,
+        hidden_states: torch.Tensor,
+        kv_cache: torch.Tensor,
+        attn_metadata: AttentionMetadata,
+    ) -> torch.Tensor:
+        attn_input = self.input_layernorm(hidden_states)
+        attn_output = self.attention(
+            position_ids=position_ids,
+            hidden_states=attn_input,
+            kv_cache=kv_cache,
+            attn_metadata=attn_metadata,
+        )
+
+        if self.use_parallel_residual:
+            # pseudocode:
+            # x = x + attn(ln1(x)) + mlp(ln2(x))
+            mlp_input = self.post_attention_layernorm(hidden_states)
+            mlp_output = self.mlp(mlp_input)
+            hidden_states = mlp_output + attn_output + hidden_states
+        else:
+            # pseudocode:
+            # x = x + attn(ln1(x))
+            # x = x + mlp(ln2(x))
+            attn_output = attn_output + hidden_states
+            mlp_input = self.post_attention_layernorm(attn_output)
+            mlp_output = self.mlp(mlp_input)
+            hidden_states = mlp_output + attn_output
+        return hidden_states
+
+
+@support_torch_compile
+class GPTNeoXModel(nn.Module):
+
+    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
+        super().__init__()
+
+        config = vllm_config.model_config.hf_config
+        cache_config = vllm_config.cache_config
+        quant_config = vllm_config.quant_config
+
+        self.config = config
+
+        self.embed_in = VocabParallelEmbedding(
+            config.vocab_size,
+            config.hidden_size,
+        )
+        self.start_layer, self.end_layer, self.layers = make_layers(
+            config.num_hidden_layers,
+            lambda prefix: GPTNeoXLayer(
+                config, cache_config, quant_config, prefix=prefix),
+            prefix=f"{prefix}.layers",
+        )
+        self.final_layer_norm = nn.LayerNorm(config.hidden_size,
+                                             eps=config.layer_norm_eps)
+        self.make_empty_intermediate_tensors = (
+            make_empty_intermediate_tensors_factory(["hidden_states"],
+                                                    config.hidden_size))
+
+    def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
+        return self.embed_in(input_ids)
+
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        position_ids: torch.Tensor,
+        kv_caches: List[torch.Tensor],
+        attn_metadata: AttentionMetadata,
+        intermediate_tensors: Optional[IntermediateTensors],
+        inputs_embeds: Optional[torch.Tensor] = None,
+    ) -> Union[torch.Tensor, IntermediateTensors]:
+        if get_pp_group().is_first_rank:
+            if inputs_embeds is not None:
+                hidden_states = inputs_embeds
+            else:
+                hidden_states = self.get_input_embeddings(input_ids)
+        else:
+            hidden_states = intermediate_tensors["hidden_states"]
+        for i in range(self.start_layer, self.end_layer):
+            layer = self.layers[i]
+            hidden_states = layer(
+                position_ids,
+                hidden_states,
+                kv_caches[i - self.start_layer],
+                attn_metadata,
+            )
+        if not get_pp_group().is_last_rank:
+            return IntermediateTensors({"hidden_states": hidden_states})
+        hidden_states = self.final_layer_norm(hidden_states)
+        return hidden_states
+
+
+class GPTNeoXForCausalLM(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.gpt_neox = GPTNeoXModel(vllm_config=vllm_config,
+                                     prefix=maybe_prefix(prefix, "gpt_neox"))
+        self.embed_out = ParallelLMHead(
+            config.vocab_size,
+            config.hidden_size,
+            quant_config=quant_config,
+        )
+        if self.config.tie_word_embeddings:
+            self.embed_out.weight = self.gpt_neox.embed_in.weight
+        self.logits_processor = LogitsProcessor(config.vocab_size)
+        self.sampler = get_sampler()
+        self.make_empty_intermediate_tensors = (
+            self.gpt_neox.make_empty_intermediate_tensors)
+
+    def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
+        return self.gpt_neox.get_input_embeddings(input_ids)
+
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        positions: torch.Tensor,
+        kv_caches: List[torch.Tensor],
+        attn_metadata: AttentionMetadata,
+        intermediate_tensors: Optional[IntermediateTensors] = None,
+        inputs_embeds: Optional[torch.Tensor] = None,
+    ) -> Union[torch.Tensor, IntermediateTensors]:
+        hidden_states = self.gpt_neox(input_ids, positions, kv_caches,
+                                      attn_metadata, intermediate_tensors,
+                                      inputs_embeds)
+        return hidden_states
+
+    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)
+        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())
+        loaded_params: Set[str] = set()
+        for name, loaded_weight in weights:
+            if ("attention.bias" in name or "attention.masked_bias" in name
+                    or "rotary_emb.inv_freq" in name):
+                continue
+            if ("rotary_emb.cos_cached" in name
+                    or "rotary_emb.sin_cached" in name):
+                # Models trained using OpenRLHF 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:
+                # NOTE: GPT-NeoX'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

.venv/lib/python3.11/site-packages/vllm/model_executor/models/granite.py

@@ -0,0 +1,520 @@
+# SPDX-License-Identifier: Apache-2.0
+
+# Adapted from
+# https://github.com/huggingface/transformers/blob/v4.28.0/src/transformers/models/llama/modeling_llama.py
+# Copyright 2023 The vLLM team.
+# Copyright 2022 EleutherAI and the HuggingFace Inc. team. All rights reserved.
+#
+# This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX
+# and OPT implementations in this library. It has been modified from its
+# original forms to accommodate minor architectural differences compared
+# to GPT-NeoX and OPT used by the Meta AI team that trained the model.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Inference-only IBM Granite model compatible with HuggingFace weights."""
+from typing import Any, Dict, Iterable, List, Optional, Set, Tuple, Union
+
+import torch
+from torch import nn
+from transformers import GraniteConfig
+
+from vllm.attention import Attention, AttentionMetadata
+from vllm.compilation.decorators import support_torch_compile
+from vllm.config import CacheConfig, VllmConfig
+from vllm.distributed import get_pp_group, get_tensor_model_parallel_world_size
+from vllm.model_executor.layers.activation import SiluAndMul
+from vllm.model_executor.layers.layernorm import RMSNorm
+from vllm.model_executor.layers.linear import (MergedColumnParallelLinear,
+                                               QKVParallelLinear,
+                                               RowParallelLinear)
+from vllm.model_executor.layers.logits_processor import LogitsProcessor
+from vllm.model_executor.layers.quantization.base_config 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, maybe_remap_kv_scale_name)
+from vllm.model_executor.sampling_metadata import SamplingMetadata
+from vllm.sequence import IntermediateTensors
+
+from .interfaces import SupportsLoRA, SupportsPP
+from .utils import (PPMissingLayer, is_pp_missing_parameter, make_layers,
+                    maybe_prefix)
+
+
+class GraniteMLP(nn.Module):
+
+    def __init__(
+        self,
+        hidden_size: int,
+        intermediate_size: int,
+        hidden_act: str,
+        quant_config: Optional[QuantizationConfig] = None,
+        bias: bool = False,
+        prefix: str = "",
+    ) -> None:
+        super().__init__()
+        self.gate_up_proj = MergedColumnParallelLinear(
+            input_size=hidden_size,
+            output_sizes=[intermediate_size] * 2,
+            bias=bias,
+            quant_config=quant_config,
+            prefix=f"{prefix}.gate_up_proj")
+        self.down_proj = RowParallelLinear(input_size=intermediate_size,
+                                           output_size=hidden_size,
+                                           bias=bias,
+                                           quant_config=quant_config,
+                                           prefix=f"{prefix}.down_proj")
+        if hidden_act != "silu":
+            raise ValueError(f"Unsupported activation: {hidden_act}. "
+                             "Only silu is supported for now.")
+        self.act_fn = SiluAndMul()
+
+    def forward(self, x):
+        gate_up, _ = self.gate_up_proj(x)
+        x = self.act_fn(gate_up)
+        x, _ = self.down_proj(x)
+        return x
+
+
+class GraniteAttention(nn.Module):
+
+    def __init__(
+        self,
+        config: GraniteConfig,
+        hidden_size: int,
+        num_heads: int,
+        num_kv_heads: int,
+        rope_theta: float = 10000,
+        rope_scaling: Optional[Dict[str, Any]] = None,
+        max_position_embeddings: int = 8192,
+        quant_config: Optional[QuantizationConfig] = None,
+        bias: bool = False,
+        cache_config: Optional[CacheConfig] = None,
+        prefix: str = "",
+    ) -> None:
+        super().__init__()
+        self.hidden_size = hidden_size
+        tp_size = get_tensor_model_parallel_world_size()
+        self.total_num_heads = num_heads
+        assert self.total_num_heads % tp_size == 0
+        self.num_heads = self.total_num_heads // tp_size
+        self.total_num_kv_heads = num_kv_heads
+        if self.total_num_kv_heads >= tp_size:
+            # Number of KV heads is greater than TP size, so we partition
+            # the KV heads across multiple tensor parallel GPUs.
+            assert self.total_num_kv_heads % tp_size == 0
+        else:
+            # Number of KV heads is less than TP size, so we replicate
+            # the KV heads across multiple tensor parallel GPUs.
+            assert tp_size % self.total_num_kv_heads == 0
+        self.num_kv_heads = max(1, self.total_num_kv_heads // tp_size)
+        # MistralConfig has an optional head_dim introduced by Mistral-Nemo
+        self.head_dim = getattr(config, "head_dim",
+                                self.hidden_size // self.total_num_heads)
+        self.q_size = self.num_heads * self.head_dim
+        self.kv_size = self.num_kv_heads * self.head_dim
+        self.scaling = config.attention_multiplier
+        self.rope_theta = rope_theta
+        self.max_position_embeddings = max_position_embeddings
+
+        self.qkv_proj = QKVParallelLinear(
+            hidden_size=hidden_size,
+            head_size=self.head_dim,
+            total_num_heads=self.total_num_heads,
+            total_num_kv_heads=self.total_num_kv_heads,
+            bias=bias,
+            quant_config=quant_config,
+            prefix=f"{prefix}.qkv_proj",
+        )
+        self.o_proj = RowParallelLinear(
+            input_size=self.total_num_heads * self.head_dim,
+            output_size=hidden_size,
+            bias=bias,
+            quant_config=quant_config,
+            prefix=f"{prefix}.o_proj",
+        )
+
+        self.rotary_emb = get_rope(
+            self.head_dim,
+            rotary_dim=self.head_dim,
+            max_position=max_position_embeddings,
+            base=rope_theta,
+            rope_scaling=rope_scaling,
+        )
+        self.attn = Attention(self.num_heads,
+                              self.head_dim,
+                              self.scaling,
+                              num_kv_heads=self.num_kv_heads,
+                              cache_config=cache_config,
+                              quant_config=quant_config,
+                              prefix=f"{prefix}.attn")
+
+    def forward(
+        self,
+        positions: torch.Tensor,
+        hidden_states: torch.Tensor,
+        kv_cache: torch.Tensor,
+        attn_metadata: AttentionMetadata,
+    ) -> torch.Tensor:
+        qkv, _ = self.qkv_proj(hidden_states)
+        q, k, v = qkv.split([self.q_size, self.kv_size, self.kv_size], dim=-1)
+        q, k = self.rotary_emb(positions, q, k)
+        attn_output = self.attn(q, k, v, kv_cache, attn_metadata)
+        output, _ = self.o_proj(attn_output)
+        return output
+
+
+class GraniteDecoderLayer(nn.Module):
+
+    def __init__(
+        self,
+        config: GraniteConfig,
+        cache_config: Optional[CacheConfig] = None,
+        quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
+    ) -> None:
+        super().__init__()
+        self.hidden_size = config.hidden_size
+        self.residual_multiplier = config.residual_multiplier
+        rope_theta = getattr(config, "rope_theta", 10000)
+        rope_scaling = getattr(config, "rope_scaling", None)
+        if rope_scaling is not None and getattr(
+                config, "original_max_position_embeddings", None):
+            rope_scaling["original_max_position_embeddings"] = (
+                config.original_max_position_embeddings)
+        max_position_embeddings = getattr(config, "max_position_embeddings",
+                                          8192)
+        # Support abacusai/Smaug-72B-v0.1 with attention_bias
+        # Support internlm/internlm-7b with bias
+        attention_bias = getattr(config, "attention_bias", False) or getattr(
+            config, "bias", False)
+        self.self_attn = GraniteAttention(
+            config=config,
+            hidden_size=self.hidden_size,
+            num_heads=config.num_attention_heads,
+            num_kv_heads=getattr(config, "num_key_value_heads",
+                                 config.num_attention_heads),
+            rope_theta=rope_theta,
+            rope_scaling=rope_scaling,
+            max_position_embeddings=max_position_embeddings,
+            quant_config=quant_config,
+            bias=attention_bias,
+            cache_config=cache_config,
+            prefix=f"{prefix}.self_attn",
+        )
+
+        self.mlp = GraniteMLP(
+            hidden_size=self.hidden_size,
+            intermediate_size=config.intermediate_size,
+            hidden_act=config.hidden_act,
+            quant_config=quant_config,
+            bias=getattr(config, "mlp_bias", False),
+            prefix=f"{prefix}.mlp",
+        )
+        self.input_layernorm = RMSNorm(config.hidden_size,
+                                       eps=config.rms_norm_eps)
+        self.post_attention_layernorm = RMSNorm(config.hidden_size,
+                                                eps=config.rms_norm_eps)
+
+    def forward(
+        self,
+        positions: torch.Tensor,
+        hidden_states: torch.Tensor,
+        kv_cache: torch.Tensor,
+        attn_metadata: AttentionMetadata,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        # Self Attention
+        residual = hidden_states
+        hidden_states = self.input_layernorm(hidden_states)
+        hidden_states = self.self_attn(
+            positions=positions,
+            hidden_states=hidden_states,
+            kv_cache=kv_cache,
+            attn_metadata=attn_metadata,
+        )
+        hidden_states = residual + hidden_states * self.residual_multiplier
+        # Fully Connected
+        residual = hidden_states
+        hidden_states = self.post_attention_layernorm(hidden_states)
+        hidden_states = self.mlp(hidden_states)
+        hidden_states = residual + hidden_states * self.residual_multiplier
+        return hidden_states
+
+
+@support_torch_compile
+class GraniteModel(nn.Module):
+
+    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
+        super().__init__()
+
+        config = vllm_config.model_config.hf_config
+        cache_config = vllm_config.cache_config
+        quant_config = vllm_config.quant_config
+        lora_config = vllm_config.lora_config
+
+        self.config = config
+        self.padding_idx = config.pad_token_id
+        lora_vocab = (lora_config.lora_extra_vocab_size *
+                      (lora_config.max_loras or 1)) if lora_config else 0
+        self.vocab_size = config.vocab_size + lora_vocab
+        self.org_vocab_size = config.vocab_size
+        if get_pp_group().is_first_rank or (config.tie_word_embeddings
+                                            and get_pp_group().is_last_rank):
+            self.embed_tokens = VocabParallelEmbedding(
+                self.vocab_size,
+                config.hidden_size,
+                org_num_embeddings=config.vocab_size,
+                quant_config=quant_config,
+            )
+        else:
+            self.embed_tokens = PPMissingLayer()
+        self.start_layer, self.end_layer, self.layers = make_layers(
+            config.num_hidden_layers,
+            lambda prefix: GraniteDecoderLayer(config=config,
+                                               cache_config=cache_config,
+                                               quant_config=quant_config,
+                                               prefix=prefix),
+            prefix=f"{prefix}.layers")
+        if get_pp_group().is_last_rank:
+            self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        else:
+            self.norm = PPMissingLayer()
+
+    def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
+        return self.embed_tokens(input_ids)
+
+    def forward(
+        self,
+        input_ids: Optional[torch.Tensor],
+        positions: torch.Tensor,
+        kv_caches: List[torch.Tensor],
+        attn_metadata: AttentionMetadata,
+        intermediate_tensors: Optional[IntermediateTensors],
+        inputs_embeds: Optional[torch.Tensor] = None,
+    ) -> Union[torch.Tensor, IntermediateTensors]:
+        if get_pp_group().is_first_rank:
+            if inputs_embeds is not None:
+                hidden_states = inputs_embeds
+            else:
+                hidden_states = self.get_input_embeddings(input_ids)
+            residual = None
+
+            hidden_states *= self.config.embedding_multiplier
+        else:
+            assert intermediate_tensors is not None
+            hidden_states = intermediate_tensors["hidden_states"]
+            residual = intermediate_tensors["residual"]
+
+        for i in range(self.start_layer, self.end_layer):
+            layer = self.layers[i]
+            hidden_states = layer(
+                positions,
+                hidden_states,
+                kv_caches[i - self.start_layer],
+                attn_metadata,
+            )
+
+        if not get_pp_group().is_last_rank:
+            return IntermediateTensors({
+                "hidden_states": hidden_states,
+                "residual": residual
+            })
+
+        hidden_states = self.norm(hidden_states)
+        return hidden_states
+
+
+class GraniteForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
+    packed_modules_mapping = {
+        "qkv_proj": [
+            "q_proj",
+            "k_proj",
+            "v_proj",
+        ],
+        "gate_up_proj": [
+            "gate_proj",
+            "up_proj",
+        ],
+    }
+
+    # LoRA specific attributes
+    supported_lora_modules = [
+        "qkv_proj", "o_proj", "gate_up_proj", "down_proj", "embed_tokens",
+        "lm_head"
+    ]
+    embedding_modules = {
+        "embed_tokens": "input_embeddings",
+        "lm_head": "output_embeddings",
+    }
+    embedding_padding_modules = ["lm_head"]
+
+    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
+        super().__init__()
+        config = vllm_config.model_config.hf_config
+        quant_config = vllm_config.quant_config
+        lora_config = vllm_config.lora_config
+
+        self.config = config
+        self.lora_config = lora_config
+        self.quant_config = quant_config
+
+        self.model = GraniteModel(vllm_config=vllm_config,
+                                  prefix=maybe_prefix(prefix, "model"))
+        if get_pp_group().is_last_rank:
+            self.unpadded_vocab_size = config.vocab_size
+            if lora_config:
+                self.unpadded_vocab_size += lora_config.lora_extra_vocab_size
+            self.lm_head = ParallelLMHead(
+                self.unpadded_vocab_size,
+                config.hidden_size,
+                org_num_embeddings=config.vocab_size,
+                padding_size=DEFAULT_VOCAB_PADDING_SIZE
+                # We need bigger padding if using lora for kernel
+                # compatibility
+                if not lora_config else lora_config.lora_vocab_padding_size,
+                quant_config=quant_config,
+            )
+            if config.tie_word_embeddings:
+                self.lm_head.weight = self.model.embed_tokens.weight
+
+            logit_scale = getattr(config, "logit_scale", 1.0)
+            if hasattr(config, "logits_scaling"):
+                logit_scale /= config.logits_scaling
+
+            self.logits_processor = LogitsProcessor(self.unpadded_vocab_size,
+                                                    config.vocab_size,
+                                                    scale=logit_scale)
+        else:
+            self.lm_head = PPMissingLayer()
+
+        self.sampler = get_sampler()
+
+    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,
+        kv_caches: List[torch.Tensor],
+        attn_metadata: AttentionMetadata,
+        intermediate_tensors: Optional[IntermediateTensors] = None,
+        inputs_embeds: Optional[torch.Tensor] = None,
+    ) -> Union[torch.Tensor, IntermediateTensors]:
+        model_output = self.model(input_ids, positions, kv_caches,
+                                  attn_metadata, intermediate_tensors,
+                                  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)
+        return logits
+
+    def sample(
+        self,
+        logits: torch.Tensor,
+        sampling_metadata: SamplingMetadata,
+    ) -> Optional[SamplerOutput]:
+        next_tokens = self.sampler(logits, sampling_metadata)
+        return next_tokens
+
+    def make_empty_intermediate_tensors(
+            self, batch_size: int, dtype: torch.dtype,
+            device: torch.device) -> IntermediateTensors:
+        return IntermediateTensors({
+            "hidden_states":
+            torch.zeros((batch_size, self.config.hidden_size),
+                        dtype=dtype,
+                        device=device),
+            "residual":
+            torch.zeros((batch_size, self.config.hidden_size),
+                        dtype=dtype,
+                        device=device),
+        })
+
+    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):
+                # Models trained using ColossalAI may include these tensors in
+                # the checkpoint. Skip them.
+                continue
+            # With tie_word_embeddings, we can skip lm_head.weight
+            # The weight might appear unnecessarily in the files if the model is
+            # processed with quantization, LoRA, fine-tuning, etc.
+            if self.config.tie_word_embeddings and "lm_head.weight" in name:
+                continue
+            if (self.quant_config is not None and
+                (scale_name := self.quant_config.get_cache_scale(name))):
+                # Loading kv cache quantization scales
+                param = params_dict[scale_name]
+                weight_loader = getattr(param, "weight_loader",
+                                        default_weight_loader)
+                loaded_weight = (loaded_weight if loaded_weight.dim() == 0 else
+                                 loaded_weight[0])
+                weight_loader(param, loaded_weight)
+                loaded_params.add(scale_name)
+                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
+                # Remapping the name of FP8 kv-scale.
+                name = maybe_remap_kv_scale_name(name, params_dict)
+                if name is None:
+                    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

.venv/lib/python3.11/site-packages/vllm/model_executor/models/gritlm.py

@@ -0,0 +1,250 @@
+# SPDX-License-Identifier: Apache-2.0
+
+from array import array
+from typing import List, Optional, Union
+
+import torch
+import torch.nn as nn
+from xformers.ops.fmha.attn_bias import BlockDiagonalMask
+
+from vllm.attention import AttentionMetadata
+from vllm.attention.backends.xformers import XFormersImpl
+from vllm.config import ModelConfig, VllmConfig
+from vllm.logger import init_logger
+from vllm.model_executor.layers.pooler import PoolerHead
+from vllm.model_executor.models.llama import LlamaForCausalLM
+from vllm.model_executor.pooling_metadata import (PoolingMetadata,
+                                                  PoolingTensors)
+from vllm.multimodal.utils import cached_get_tokenizer
+from vllm.sequence import (IntermediateTensors, PoolerOutput,
+                           PoolingSequenceGroupOutput)
+
+logger = init_logger(__name__)
+
+
+class GritLMPooler(nn.Module):
+
+    def __init__(self, model_config: ModelConfig):
+        super().__init__()
+
+        self.model_config = model_config
+
+        tokenizer = cached_get_tokenizer(
+            self.model_config.tokenizer,
+            tokenizer_mode=self.model_config.tokenizer_mode,
+            tokenizer_revision=self.model_config.tokenizer_revision,
+            trust_remote_code=self.model_config.trust_remote_code,
+        )
+
+        # Collect the tokens needed for pattern matching.
+        # "▁<" is different from "_<". The former uses "▁" to indicate that
+        # the next token is the start of a word.
+        # "<0x0A>" is the newline token (i.e. "\n")."
+        self.token_ids = {
+            tok: tokenizer.convert_tokens_to_ids([tok])[0]
+            for tok in ["<s>", "▁<", "<", "|", "embed", ">", "<0x0A>", "user"]
+        }
+
+        def tokens_to_ids(tokens: list[str]) -> array:
+            return array("i", [self.token_ids[token] for token in tokens])
+
+        self.user_pattern_ids = tokens_to_ids(
+            ["▁<", "|", "user", "|", ">", "<0x0A>"])
+        self.embed_newline_pattern_ids = tokens_to_ids(
+            ["<0x0A>", "<", "|", "embed", "|", ">", "<0x0A>"])
+        self.embed_pattern_ids = tokens_to_ids(
+            ["▁<", "|", "embed", "|", ">", "<0x0A>"])
+
+        self.head = PoolerHead(normalize=True, softmax=False)
+
+    def _find_array(self, arr: array, target: array, start_idx: int) -> int:
+        """
+        Find the first occurrence of target in arr starting from start_idx.
+
+        Args:
+        arr: The array to search within
+        target: The consecutive subsequence to find
+        start_idx: The starting index to search from
+
+        Returns:
+        int: The index of the first occurrence of target in arr.
+        """
+        if start_idx < 0:
+            raise ValueError("start_idx must be non-negative")
+        if not target or not arr:
+            raise ValueError("Empty arr or target not allowed")
+
+        target_len = len(target)
+        for i in range(start_idx, len(arr) - target_len + 1):
+            if arr[i:i + target_len] == target:
+                return i
+        return -1
+
+    def _get_instruction_len(self, prompt_token_ids: array) -> int:
+        """
+        Get the length of the instruction in the prompt.
+
+        We do a pattern matching to find the instruction in the prompt,
+        and then return the length of the instruction.
+
+        The pattern matching is done using integers instead of strings
+        because the prompt is given as a list of token IDs.
+        """
+
+        instruction_len = 0
+
+        # Return no instruction in case of missing BOS token.
+        if prompt_token_ids[0] != self.token_ids["<s>"]:
+            logger.warning("BOS token not found in prompt,"
+                           "thus using empty string for instruction."
+                           "GritLM requires BOS token in prompt.")
+            return instruction_len
+
+        # If user pattern is found in the prompt, that means there should be
+        # a newline token before the embed pattern.
+        embed_pattern_ids = self.embed_pattern_ids
+        if self._find_array(prompt_token_ids,
+                            self.user_pattern_ids,
+                            start_idx=1) == 1:
+            embed_pattern_ids = self.embed_newline_pattern_ids
+
+        # Find the embed pattern in the prompt.
+        found_embed_pattern_idx = self._find_array(prompt_token_ids,
+                                                   embed_pattern_ids,
+                                                   start_idx=1)
+
+        if found_embed_pattern_idx != -1:
+            instruction_len = found_embed_pattern_idx + len(embed_pattern_ids)
+        else:
+            logger.warning("Query instruction not found in prompt,"
+                           "thus using BOS token as instruction instead."
+                           "GritLM requires query instruction in prompt.")
+            instruction_len = 1
+
+        return instruction_len
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        pooling_metadata: PoolingMetadata,
+    ) -> PoolerOutput:
+        """
+        Pool the hidden states by summing the embeddings of
+        non-instruction tokens.
+        """
+        prompts_token_ids = [
+            token_ids.prompt_token_ids_array
+            for _, token_ids in pooling_metadata.seq_data.items()
+        ]
+
+        instruction_lens = torch.tensor(
+            [
+                self._get_instruction_len(prompt_token_ids)
+                for prompt_token_ids in prompts_token_ids
+            ],
+            device=hidden_states.device,
+        )
+
+        prompt_lens = PoolingTensors.from_pooling_metadata(
+            pooling_metadata, hidden_states.device).prompt_lens
+
+        mask = torch.zeros_like(hidden_states, dtype=torch.bool)
+
+        start_idx = 0
+        for prompt_len, instruction_len in zip(prompt_lens, instruction_lens):
+            end_idx = start_idx + prompt_len
+            mask[start_idx + instruction_len:end_idx] = True
+            start_idx = end_idx
+
+        masked_hidden_states = hidden_states.masked_fill(~mask, 0.0)
+
+        sum_embeddings = torch.zeros(len(prompt_lens),
+                                     hidden_states.size(1),
+                                     device=hidden_states.device)
+
+        start_idx = 0
+        for i, prompt_len in enumerate(prompt_lens):
+            end_idx = start_idx + prompt_len
+            sum_embeddings[i] = masked_hidden_states[start_idx:end_idx].sum(
+                dim=0)
+            start_idx = end_idx
+
+        num_non_instruction_tokens = prompt_lens - instruction_lens
+        mean_embeddings = sum_embeddings / num_non_instruction_tokens.unsqueeze(
+            1)
+
+        pooled_data = self.head(mean_embeddings)
+
+        pooled_outputs = [
+            PoolingSequenceGroupOutput(data) for data in pooled_data
+        ]
+
+        return PoolerOutput(outputs=pooled_outputs)
+
+
+class GritLM(LlamaForCausalLM):
+    """This class implements the embedding model for parasail-ai/GritLM-7B-vllm.
+
+    The class inherits from LlamaForCausalLM and provides a custom pooling
+    layer.
+
+    The main difference between the pooling layer in GritLM and the one in
+    LlamaForCausalLM is that GritLM ignores the query instruction in the prompt
+    when pooling the hidden states.
+
+    Embedding prompts should be in the following format:
+    - With instruction: "<|user|>\nINSTRUCTION\n<|embed|>\nPROMPT".
+    - Without instruction: "<|embed|>\nPROMPT".
+
+    Generation prompts should be in the following format:
+    - "<|user|>\nPROMPT\n<|assistant|>\n"
+    """
+
+    def __init__(
+        self,
+        vllm_config: VllmConfig,
+        prefix: str = "",
+        **kwargs,
+    ) -> None:
+        super().__init__(vllm_config=vllm_config, prefix=prefix, **kwargs)
+
+        self.runner_type = vllm_config.model_config.runner_type
+
+        self._pooler = GritLMPooler(vllm_config.model_config)
+
+        for layer in self.model.layers:
+            if self.runner_type == "pooling" and hasattr(layer, "self_attn"):
+                assert isinstance(layer.self_attn.attn.impl, XFormersImpl), (
+                    "GritLM embedding is only supported by XFormers backend, "
+                    "which can be forced by VLLM_ATTENTION_BACKEND=XFORMERS")
+
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        positions: torch.Tensor,
+        kv_caches: List[torch.Tensor],
+        attn_metadata: AttentionMetadata,
+        **kwargs,
+    ) -> Union[torch.Tensor, IntermediateTensors]:
+
+        # Change attention to non-causal for pooling tasks.
+        if self.runner_type == "pooling":
+            assert attn_metadata.prefill_metadata.attn_bias is None
+            attn_metadata.prefill_metadata.attn_bias = [
+                BlockDiagonalMask.from_seqlens(attn_metadata.seq_lens)
+            ]
+
+        return super().forward(
+            input_ids=input_ids,
+            positions=positions,
+            kv_caches=kv_caches,
+            attn_metadata=attn_metadata,
+            **kwargs,
+        )
+
+    def pooler(
+        self,
+        hidden_states: torch.Tensor,
+        pooling_metadata: PoolingMetadata,
+    ) -> Optional[PoolerOutput]:
+        return self._pooler(hidden_states, pooling_metadata)

.venv/lib/python3.11/site-packages/vllm/model_executor/models/idefics2_vision_model.py

@@ -0,0 +1,346 @@
+# SPDX-License-Identifier: Apache-2.0
+
+# adapted from https://github.com/huggingface/transformers/blob/v4.43.2/src/transformers/models/idefics2/modeling_idefics2.py
+# Copyright 2024 The vLLM team.
+# Copyright 2024 the HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""PyTorch Idefics2 model."""
+
+from typing import Iterable, Optional, Set, Tuple
+
+import torch
+from torch import nn
+from transformers.models.idefics2.configuration_idefics2 import (
+    Idefics2Config, Idefics2VisionConfig)
+
+from vllm.attention.layer import MultiHeadAttention
+from vllm.distributed import divide, get_tensor_model_parallel_world_size
+from vllm.model_executor.layers.activation import get_act_fn
+from vllm.model_executor.layers.linear import (ColumnParallelLinear,
+                                               QKVParallelLinear,
+                                               RowParallelLinear)
+from vllm.model_executor.layers.quantization import QuantizationConfig
+from vllm.model_executor.model_loader.weight_utils import default_weight_loader
+
+
+class Idefics2VisionEmbeddings(nn.Module):
+    """
+    This is a modified version of `siglip.modelign_siglip.SiglipVisionEmbeddings
+    ` to enable images of variable
+    resolution.
+
+    The modifications are adapted from [Patch n' Pack: NaViT, a Vision
+    Transformer for any Aspect Ratio and Resolution](https://arxiv.org/abs/2307.06304)
+    which allows treating images in their native aspect ratio and without the
+    need to resize them to the same fixed size. In particular, we start from the
+    original pre-trained SigLIP model(which uses images of fixed-size square
+    images) and adapt it by training on images of variable resolutions.
+    """
+
+    def __init__(self, config: Idefics2VisionConfig):
+        super().__init__()
+        self.embed_dim = config.hidden_size
+        self.image_size = config.image_size
+        self.patch_size = config.patch_size
+        self.patch_embedding = nn.Conv2d(
+            in_channels=config.num_channels,
+            out_channels=self.embed_dim,
+            kernel_size=self.patch_size,
+            stride=self.patch_size,
+            padding="valid",
+        )
+        self.num_patches_per_side = self.image_size // self.patch_size
+        self.num_patches = self.num_patches_per_side**2
+        self.num_positions = self.num_patches
+        self.position_embedding = nn.Embedding(self.num_positions,
+                                               self.embed_dim)
+
+    def forward(self,
+                pixel_values: torch.FloatTensor,
+                patch_attention_mask: torch.BoolTensor,
+                tgt_sizes: Optional[torch.IntTensor] = None) -> torch.Tensor:
+        batch_size, _, max_im_h, max_im_w = pixel_values.shape
+        target_dtype = self.patch_embedding.weight.dtype
+        patch_embeds = self.patch_embedding(pixel_values.to(target_dtype))
+        embeddings = patch_embeds.flatten(2).transpose(1, 2)
+        max_nb_patches_h, max_nb_patches_w = (
+            max_im_h // self.patch_size,
+            max_im_w // self.patch_size,
+        )
+        boundaries = torch.arange(1 / self.num_patches_per_side, 1.0,
+                                  1 / self.num_patches_per_side)
+        position_ids = torch.full(size=(batch_size,
+                                        max_nb_patches_h * max_nb_patches_w),
+                                  fill_value=0)
+
+        for batch_idx, p_attn_mask in enumerate(patch_attention_mask):
+
+            if tgt_sizes is not None:
+                nb_patches_h = tgt_sizes[batch_idx][0]
+                nb_patches_w = tgt_sizes[batch_idx][1]
+            else:
+                nb_patches_h = p_attn_mask[:, 0].sum()
+                nb_patches_w = p_attn_mask[0].sum()
+            fractional_coords_h = torch.arange(0, 1 - 1e-6, 1 / nb_patches_h)
+            fractional_coords_w = torch.arange(0, 1 - 1e-6, 1 / nb_patches_w)
+            bucket_coords_h = torch.bucketize(fractional_coords_h,
+                                              boundaries,
+                                              right=True)
+            bucket_coords_w = torch.bucketize(fractional_coords_w,
+                                              boundaries,
+                                              right=True)
+            pos_ids = (bucket_coords_h[:, None] * self.num_patches_per_side +
+                       bucket_coords_w).flatten()
+            position_ids[batch_idx][p_attn_mask.view(-1).cpu()] = pos_ids
+        position_ids = position_ids.to(self.position_embedding.weight.device)
+        embeddings = embeddings + self.position_embedding(position_ids)
+        return embeddings
+
+
+class Idefics2VisionAttention(nn.Module):
+    """Multi-headed attention from 'Attention Is All You Need' paper"""
+
+    def __init__(
+        self,
+        config: Idefics2Config,
+        quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
+    ) -> None:
+        super().__init__()
+        self.config = config
+        self.embed_dim = config.hidden_size
+        self.num_heads = config.num_attention_heads
+        self.head_dim = self.embed_dim // self.num_heads
+        if self.head_dim * self.num_heads != self.embed_dim:
+            raise ValueError(
+                f"embed_dim must be divisible by num_heads (got `embed_dim`: {self.embed_dim} and `num_heads`:"  # noqa: E501
+                f" {self.num_heads}).")
+        self.scale = self.head_dim**-0.5
+        self.dropout = config.attention_dropout
+        self.qkv_proj = QKVParallelLinear(
+            self.embed_dim,
+            self.head_dim,
+            self.num_heads,
+            quant_config=quant_config,
+            prefix=f"{prefix}.qkv_proj",
+        )
+        self.out_proj = RowParallelLinear(
+            self.embed_dim,
+            self.embed_dim,
+            bias=True,
+            quant_config=quant_config,
+            prefix=f"{prefix}.out_proj",
+        )
+        self.tp_size = get_tensor_model_parallel_world_size()
+        self.num_heads_per_partition = divide(self.num_heads, self.tp_size)
+        self.attn = MultiHeadAttention(self.num_heads_per_partition,
+                                       self.head_dim, self.scale)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+    ) -> torch.Tensor:
+        qkv, _ = self.qkv_proj(
+            hidden_states
+        )  # batch_size, q_len, 3 * num_heads_per_partition * head_dim
+        query_states, key_states, value_states = qkv.chunk(3, dim=-1)
+        out = self.attn(query_states, key_states, value_states)
+        attn_output, _ = self.out_proj(out)
+        return attn_output
+
+
+class Idefics2VisionMLP(nn.Module):
+
+    def __init__(
+        self,
+        config: Idefics2Config,
+        quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
+    ) -> None:
+        super().__init__()
+        self.config = config
+        self.activation_fn = get_act_fn(config.hidden_act)
+        self.fc1 = ColumnParallelLinear(
+            config.hidden_size,
+            config.intermediate_size,
+            bias=True,
+            quant_config=quant_config,
+            prefix=f"{prefix}.fc1",
+        )
+        self.fc2 = RowParallelLinear(
+            config.intermediate_size,
+            config.hidden_size,
+            bias=True,
+            quant_config=quant_config,
+            prefix=f"{prefix}.fc2",
+        )
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        hidden_states, _ = self.fc1(hidden_states)
+        hidden_states = self.activation_fn(hidden_states)
+        hidden_states, _ = self.fc2(hidden_states)
+        return hidden_states
+
+
+class Idefics2EncoderLayer(nn.Module):
+
+    def __init__(
+        self,
+        config: Idefics2Config,
+        quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
+    ) -> None:
+        super().__init__()
+        self.embed_dim = config.hidden_size
+        self.self_attn = Idefics2VisionAttention(config,
+                                                 quant_config=quant_config,
+                                                 prefix=f"{prefix}.self_attn")
+        self.layer_norm1 = nn.LayerNorm(self.embed_dim,
+                                        eps=config.layer_norm_eps)
+        self.mlp = Idefics2VisionMLP(config,
+                                     quant_config=quant_config,
+                                     prefix=f"{prefix}.mlp")
+        self.layer_norm2 = nn.LayerNorm(self.embed_dim,
+                                        eps=config.layer_norm_eps)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+    ) -> torch.Tensor:
+        """
+        Args:
+            hidden_states (`torch.FloatTensor`):
+                Input to the layer of shape `(batch, seq_len, embed_dim)`.
+
+        """
+        residual = hidden_states
+        hidden_states = self.layer_norm1(hidden_states)
+        hidden_states = self.self_attn(hidden_states)
+        hidden_states = residual + hidden_states
+        residual = hidden_states
+        hidden_states = self.layer_norm2(hidden_states)
+        hidden_states = self.mlp(hidden_states)
+        hidden_states = residual + hidden_states
+        return hidden_states
+
+
+class Idefics2Encoder(nn.Module):
+    """
+    Transformer encoder consisting of `config.num_hidden_layers` self attention
+    layers. Each layer is a
+    [`Idefics2EncoderLayer`].
+
+    Args:
+        config: Idefics2Config
+    """
+
+    def __init__(
+        self,
+        config: Idefics2Config,
+        quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
+    ) -> None:
+        super().__init__()
+
+        self.config = config
+        self.layers = nn.ModuleList([
+            Idefics2EncoderLayer(config,
+                                 quant_config=quant_config,
+                                 prefix=f"{prefix}.layers.{layer_idx}")
+            for layer_idx in range(config.num_hidden_layers)
+        ])
+
+    def forward(
+        self,
+        inputs_embeds: torch.Tensor,
+    ) -> torch.Tensor:
+        r"""
+        Args:
+            inputs_embeds (torch.Tensor):
+                Optionally, instead of passing `input_ids` you can choose to
+                directly pass an embedded representation.
+                This is useful if you want more control over how to convert
+                `input_ids` indices into associated vectorsthan the model's
+                internal embedding lookup matrix.
+        """
+        hidden_states = inputs_embeds
+        for encoder_layer in self.layers:
+            layer_outputs = encoder_layer(hidden_states)
+            hidden_states = layer_outputs
+        return hidden_states
+
+
+class Idefics2VisionTransformer(nn.Module):
+
+    def __init__(
+        self,
+        config: Idefics2VisionConfig,
+        quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
+    ) -> None:
+        super().__init__()
+
+        embed_dim = config.hidden_size
+        self.config = config
+        self.embeddings = Idefics2VisionEmbeddings(config)
+        self.encoder = Idefics2Encoder(config,
+                                       quant_config=quant_config,
+                                       prefix=f"{prefix}.encoder")
+        self.post_layernorm = nn.LayerNorm(embed_dim,
+                                           eps=config.layer_norm_eps)
+
+    def get_input_embeddings(self):
+        return self.embeddings
+
+    def forward(
+        self,
+        pixel_values,
+        patch_attention_mask: Optional[torch.BoolTensor] = None,
+        tgt_sizes: Optional[torch.IntTensor] = None,
+    ) -> torch.Tensor:
+        hidden_states = self.embeddings(
+            pixel_values=pixel_values,
+            patch_attention_mask=patch_attention_mask,
+            tgt_sizes=tgt_sizes,
+        )
+        encoder_outputs = self.encoder(hidden_states)
+        last_hidden_state = self.post_layernorm(encoder_outputs)
+        return last_hidden_state
+
+    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())
+        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)
+                param = params_dict[name]
+                weight_loader = param.weight_loader
+                weight_loader(param, loaded_weight, shard_id)
+                break
+            else:
+                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

.venv/lib/python3.11/site-packages/vllm/model_executor/models/interfaces.py

@@ -0,0 +1,443 @@
+# SPDX-License-Identifier: Apache-2.0
+
+from typing import (TYPE_CHECKING, ClassVar, Dict, List, Literal, Optional,
+                    Protocol, Type, Union, overload, runtime_checkable)
+
+import torch
+from typing_extensions import TypeIs, TypeVar
+
+from vllm.logger import init_logger
+from vllm.utils import supports_kw
+
+from .interfaces_base import is_pooling_model
+
+if TYPE_CHECKING:
+    from vllm.attention import AttentionMetadata
+    from vllm.multimodal.inputs import NestedTensors  # noqa: F401
+    from vllm.sequence import IntermediateTensors
+
+logger = init_logger(__name__)
+
+T = TypeVar("T", default="NestedTensors")
+
+
+@runtime_checkable
+class SupportsMultiModal(Protocol):
+    """The interface required for all multi-modal models."""
+
+    supports_multimodal: ClassVar[Literal[True]] = True
+    """
+    A flag that indicates this model supports multi-modal inputs.
+
+    Note:
+        There is no need to redefine this flag if this class is in the
+        MRO of your model class.
+    """
+
+    def get_multimodal_embeddings(self, **kwargs) -> Optional[T]:
+        """
+        Returns multimodal embeddings generated from multimodal kwargs 
+        to be merged with text embeddings.
+
+        The output embeddings must be one of the following formats:
+    
+        - A list or tuple of 2D tensors, where each tensor corresponds to
+          each input multimodal data item (e.g, image).
+        - A single 3D tensor, with the batch dimension grouping the 2D tensors.
+
+        Note:
+            The returned multimodal embeddings must be in the same order as
+            the appearances of their corresponding multimodal data item in the
+            input prompt.
+        """
+        ...
+
+    # Only for models that support v0 chunked prefill
+    # TODO(ywang96): Remove this overload once v0 is deprecated
+    @overload
+    def get_input_embeddings(
+        self,
+        input_ids: torch.Tensor,
+        multimodal_embeddings: Optional[T] = None,
+        attn_metadata: Optional["AttentionMetadata"] = None,
+    ) -> torch.Tensor:
+        ...
+
+    @overload
+    def get_input_embeddings(
+        self,
+        input_ids: torch.Tensor,
+        multimodal_embeddings: Optional[T] = None,
+    ) -> torch.Tensor:
+        """
+        Returns the input embeddings merged from the text embeddings from 
+        input_ids and the multimodal embeddings generated from multimodal 
+        kwargs.
+        """
+        ...
+
+
+# We can't use runtime_checkable with ClassVar for issubclass checks
+# so we need to treat the class as an instance and use isinstance instead
+@runtime_checkable
+class _SupportsMultiModalType(Protocol):
+    supports_multimodal: Literal[True]
+
+
+@overload
+def supports_multimodal(
+        model: Type[object]) -> TypeIs[Type[SupportsMultiModal]]:
+    ...
+
+
+@overload
+def supports_multimodal(model: object) -> TypeIs[SupportsMultiModal]:
+    ...
+
+
+def supports_multimodal(
+    model: Union[Type[object], object],
+) -> Union[TypeIs[Type[SupportsMultiModal]], TypeIs[SupportsMultiModal]]:
+    if isinstance(model, type):
+        return isinstance(model, _SupportsMultiModalType)
+
+    return isinstance(model, SupportsMultiModal)
+
+
+@runtime_checkable
+class SupportsLoRA(Protocol):
+    """The interface required for all models that support LoRA."""
+
+    supports_lora: ClassVar[Literal[True]] = True
+    """
+    A flag that indicates this model supports LoRA.
+
+    Note:
+        There is no need to redefine this flag if this class is in the
+        MRO of your model class.
+    """
+
+    packed_modules_mapping: ClassVar[Dict[str, List[str]]]
+    supported_lora_modules: ClassVar[List[str]]
+    embedding_modules: ClassVar[Dict[str, str]]
+    embedding_padding_modules: ClassVar[List[str]]
+
+
+# We can't use runtime_checkable with ClassVar for issubclass checks
+# so we need to treat the class as an instance and use isinstance instead
+@runtime_checkable
+class _SupportsLoRAType(Protocol):
+    supports_lora: Literal[True]
+
+    packed_modules_mapping: Dict[str, List[str]]
+    supported_lora_modules: List[str]
+    embedding_modules: Dict[str, str]
+    embedding_padding_modules: List[str]
+
+
+@overload
+def supports_lora(model: Type[object]) -> TypeIs[Type[SupportsLoRA]]:
+    ...
+
+
+@overload
+def supports_lora(model: object) -> TypeIs[SupportsLoRA]:
+    ...
+
+
+def supports_lora(
+    model: Union[Type[object], object],
+) -> Union[TypeIs[Type[SupportsLoRA]], TypeIs[SupportsLoRA]]:
+    result = _supports_lora(model)
+
+    if not result:
+        lora_attrs = (
+            "packed_modules_mapping",
+            "supported_lora_modules",
+            "embedding_modules",
+            "embedding_padding_modules",
+        )
+        missing_attrs = tuple(attr for attr in lora_attrs
+                              if not hasattr(model, attr))
+
+        if getattr(model, "supports_lora", False):
+            if missing_attrs:
+                logger.warning(
+                    "The model (%s) sets `supports_lora=True`, "
+                    "but is missing LoRA-specific attributes: %s",
+                    model,
+                    missing_attrs,
+                )
+        else:
+            if not missing_attrs:
+                logger.warning(
+                    "The model (%s) contains all LoRA-specific attributes, "
+                    "but does not set `supports_lora=True`.", model)
+
+    return result
+
+
+def _supports_lora(model: Union[Type[object], object]) -> bool:
+    if isinstance(model, type):
+        return isinstance(model, _SupportsLoRAType)
+
+    return isinstance(model, SupportsLoRA)
+
+
+@runtime_checkable
+class SupportsPP(Protocol):
+    """The interface required for all models that support pipeline parallel."""
+
+    supports_pp: ClassVar[Literal[True]] = True
+    """
+    A flag that indicates this model supports pipeline parallel.
+
+    Note:
+        There is no need to redefine this flag if this class is in the
+        MRO of your model class.
+    """
+
+    def make_empty_intermediate_tensors(
+        self,
+        batch_size: int,
+        dtype: torch.dtype,
+        device: torch.device,
+    ) -> "IntermediateTensors":
+        """Called when PP rank > 0 for profiling purposes."""
+        ...
+
+    def forward(
+        self,
+        *,
+        intermediate_tensors: Optional["IntermediateTensors"],
+    ) -> Union[torch.Tensor, "IntermediateTensors"]:
+        """
+        Accept :class:`IntermediateTensors` when PP rank > 0.
+
+        Return :class:`IntermediateTensors` only for the last PP rank.
+        """
+        ...
+
+
+# We can't use runtime_checkable with ClassVar for issubclass checks
+# so we need to treat the class as an instance and use isinstance instead
+@runtime_checkable
+class _SupportsPPType(Protocol):
+    supports_pp: Literal[True]
+
+    def make_empty_intermediate_tensors(
+        self,
+        batch_size: int,
+        dtype: torch.dtype,
+        device: torch.device,
+    ) -> "IntermediateTensors":
+        ...
+
+    def forward(
+        self,
+        *,
+        intermediate_tensors: Optional["IntermediateTensors"],
+    ) -> Union[torch.Tensor, "IntermediateTensors"]:
+        ...
+
+
+@overload
+def supports_pp(model: Type[object]) -> TypeIs[Type[SupportsPP]]:
+    ...
+
+
+@overload
+def supports_pp(model: object) -> TypeIs[SupportsPP]:
+    ...
+
+
+def supports_pp(
+    model: Union[Type[object], object],
+) -> Union[bool, TypeIs[Type[SupportsPP]], TypeIs[SupportsPP]]:
+    supports_attributes = _supports_pp_attributes(model)
+    supports_inspect = _supports_pp_inspect(model)
+
+    if supports_attributes and not supports_inspect:
+        logger.warning(
+            "The model (%s) sets `supports_pp=True`, but does not accept "
+            "`intermediate_tensors` in its `forward` method", model)
+
+    if not supports_attributes:
+        pp_attrs = ("make_empty_intermediate_tensors", )
+        missing_attrs = tuple(attr for attr in pp_attrs
+                              if not hasattr(model, attr))
+
+        if getattr(model, "supports_pp", False):
+            if missing_attrs:
+                logger.warning(
+                    "The model (%s) sets `supports_pp=True`, "
+                    "but is missing PP-specific attributes: %s",
+                    model,
+                    missing_attrs,
+                )
+        else:
+            if not missing_attrs:
+                logger.warning(
+                    "The model (%s) contains all PP-specific attributes, "
+                    "but does not set `supports_pp=True`.", model)
+
+    return supports_attributes and supports_inspect
+
+
+def _supports_pp_attributes(model: Union[Type[object], object]) -> bool:
+    if isinstance(model, type):
+        return isinstance(model, _SupportsPPType)
+
+    return isinstance(model, SupportsPP)
+
+
+def _supports_pp_inspect(model: Union[Type[object], object]) -> bool:
+    model_forward = getattr(model, "forward", None)
+    if not callable(model_forward):
+        return False
+
+    return supports_kw(model_forward, "intermediate_tensors")
+
+
+@runtime_checkable
+class HasInnerState(Protocol):
+    """The interface required for all models that has inner state."""
+
+    has_inner_state: ClassVar[Literal[True]] = True
+    """
+        A flag that indicates this model has inner state.
+        Models that has inner state usually need access to the scheduler_config
+        for max_num_seqs, etc. True for e.g. both Mamba and Jamba.
+    """
+
+
+@runtime_checkable
+class _HasInnerStateType(Protocol):
+    has_inner_state: ClassVar[Literal[True]]
+
+
+@overload
+def has_inner_state(model: object) -> TypeIs[HasInnerState]:
+    ...
+
+
+@overload
+def has_inner_state(model: Type[object]) -> TypeIs[Type[HasInnerState]]:
+    ...
+
+
+def has_inner_state(
+    model: Union[Type[object], object]
+) -> Union[TypeIs[Type[HasInnerState]], TypeIs[HasInnerState]]:
+    if isinstance(model, type):
+        return isinstance(model, _HasInnerStateType)
+
+    return isinstance(model, HasInnerState)
+
+
+@runtime_checkable
+class IsAttentionFree(Protocol):
+    """The interface required for all models like Mamba that lack attention,
+    but do have state whose size is constant wrt the number of tokens."""
+
+    is_attention_free: ClassVar[Literal[True]] = True
+    """
+        A flag that indicates this model has no attention.
+        Used for block manager and attention backend selection.
+        True for Mamba but not Jamba.
+    """
+
+
+@runtime_checkable
+class _IsAttentionFreeType(Protocol):
+    is_attention_free: ClassVar[Literal[True]]
+
+
+@overload
+def is_attention_free(model: object) -> TypeIs[IsAttentionFree]:
+    ...
+
+
+@overload
+def is_attention_free(model: Type[object]) -> TypeIs[Type[IsAttentionFree]]:
+    ...
+
+
+def is_attention_free(
+    model: Union[Type[object], object]
+) -> Union[TypeIs[Type[IsAttentionFree]], TypeIs[IsAttentionFree]]:
+    if isinstance(model, type):
+        return isinstance(model, _IsAttentionFreeType)
+
+    return isinstance(model, IsAttentionFree)
+
+
+@runtime_checkable
+class IsHybrid(Protocol):
+    """The interface required for all models like Jamba that have both
+    attention and mamba blocks, indicates that 
+    hf_config has 'layers_block_type'"""
+
+    is_hybrid: ClassVar[Literal[True]] = True
+    """
+        A flag that indicates this model has both mamba and attention blocks
+        , also indicates that the model's hf_config has 
+        'layers_block_type' """
+
+
+@runtime_checkable
+class _IsHybridType(Protocol):
+    is_hybrid: ClassVar[Literal[True]]
+
+
+@overload
+def is_hybrid(model: object) -> TypeIs[IsHybrid]:
+    ...
+
+
+@overload
+def is_hybrid(model: Type[object]) -> TypeIs[Type[IsHybrid]]:
+    ...
+
+
+def is_hybrid(
+    model: Union[Type[object], object]
+) -> Union[TypeIs[Type[IsHybrid]], TypeIs[IsHybrid]]:
+    if isinstance(model, type):
+        return isinstance(model, _IsHybridType)
+
+    return isinstance(model, IsHybrid)
+
+
+@runtime_checkable
+class SupportsCrossEncoding(Protocol):
+    """The interface required for all models that support cross encoding."""
+
+    supports_cross_encoding: ClassVar[Literal[True]] = True
+
+
+@overload
+def supports_cross_encoding(
+        model: Type[object]) -> TypeIs[Type[SupportsCrossEncoding]]:
+    ...
+
+
+@overload
+def supports_cross_encoding(model: object) -> TypeIs[SupportsCrossEncoding]:
+    ...
+
+
+def _supports_cross_encoding(
+    model: Union[Type[object], object],
+) -> Union[TypeIs[Type[SupportsCrossEncoding]], TypeIs[SupportsCrossEncoding]]:
+
+    if isinstance(model, type):
+        return isinstance(model, SupportsCrossEncoding)
+
+    return isinstance(model, SupportsCrossEncoding)
+
+
+def supports_cross_encoding(
+    model: Union[Type[object], object],
+) -> Union[TypeIs[Type[SupportsCrossEncoding]], TypeIs[SupportsCrossEncoding]]:
+    return is_pooling_model(model) and _supports_cross_encoding(model)

.venv/lib/python3.11/site-packages/vllm/model_executor/models/interfaces_base.py

@@ -0,0 +1,175 @@
+# SPDX-License-Identifier: Apache-2.0
+
+from typing import (TYPE_CHECKING, List, Optional, Protocol, Type, Union,
+                    overload, runtime_checkable)
+
+import torch
+import torch.nn as nn
+from typing_extensions import TypeIs, TypeVar
+
+from vllm.logger import init_logger
+from vllm.utils import supports_kw
+
+if TYPE_CHECKING:
+    from vllm.attention import AttentionMetadata
+    from vllm.config import VllmConfig
+    from vllm.model_executor.layers.pooler import PoolerOutput
+    from vllm.model_executor.layers.sampler import SamplerOutput
+    from vllm.model_executor.pooling_metadata import PoolingMetadata
+    from vllm.model_executor.sampling_metadata import SamplingMetadata
+
+logger = init_logger(__name__)
+
+# The type of hidden states
+# Currently, T = torch.Tensor for all models except for Medusa
+# which has T = List[torch.Tensor]
+T = TypeVar("T", default=torch.Tensor)
+T_co = TypeVar("T_co", default=torch.Tensor, covariant=True)
+
+# NOTE: Unlike those in `interfaces.py`, we don't define `ClassVar` tags
+# for the base interfaces to avoid breaking OOT registration for existing models
+# that don't inherit from the base interface classes
+
+
+@runtime_checkable
+class VllmModel(Protocol[T_co]):
+    """The interface required for all models in vLLM."""
+
+    def __init__(
+        self,
+        vllm_config: "VllmConfig",
+        prefix: str = "",
+    ) -> None:
+        ...
+
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        positions: torch.Tensor,
+        kv_caches: List[torch.Tensor],
+        attn_metadata: "AttentionMetadata",
+    ) -> T_co:
+        ...
+
+
+def _check_vllm_model_init(model: Union[Type[object], object]) -> bool:
+    model_init = model.__init__
+    return supports_kw(model_init, "vllm_config")
+
+
+def _check_vllm_model_forward(model: Union[Type[object], object]) -> bool:
+    model_forward = getattr(model, "forward", None)
+    if not callable(model_forward):
+        return False
+
+    vllm_kws = ("input_ids", "positions", "kv_caches", "attn_metadata")
+    missing_kws = tuple(kw for kw in vllm_kws
+                        if not supports_kw(model_forward, kw))
+
+    if missing_kws and (isinstance(model, type)
+                        and issubclass(model, nn.Module)):
+        logger.warning(
+            "The model (%s) is missing "
+            "vLLM-specific keywords from its `forward` method: %s",
+            model,
+            missing_kws,
+        )
+
+    return len(missing_kws) == 0
+
+
+@overload
+def is_vllm_model(model: Type[object]) -> TypeIs[Type[VllmModel]]:
+    ...
+
+
+@overload
+def is_vllm_model(model: object) -> TypeIs[VllmModel]:
+    ...
+
+
+def is_vllm_model(
+    model: Union[Type[object], object],
+) -> Union[TypeIs[Type[VllmModel]], TypeIs[VllmModel]]:
+    return _check_vllm_model_init(model) and _check_vllm_model_forward(model)
+
+
+@runtime_checkable
+class VllmModelForTextGeneration(VllmModel[T], Protocol[T]):
+    """The interface required for all generative models in vLLM."""
+
+    def compute_logits(
+        self,
+        hidden_states: T,
+        sampling_metadata: "SamplingMetadata",
+    ) -> Optional[T]:
+        """Return `None` if TP rank > 0."""
+        ...
+
+    def sample(
+        self,
+        logits: T,
+        sampling_metadata: "SamplingMetadata",
+    ) -> "SamplerOutput":
+        """Only called on TP rank 0."""
+        ...
+
+
+@overload
+def is_text_generation_model(
+        model: Type[object]) -> TypeIs[Type[VllmModelForTextGeneration]]:
+    ...
+
+
+@overload
+def is_text_generation_model(
+        model: object) -> TypeIs[VllmModelForTextGeneration]:
+    ...
+
+
+def is_text_generation_model(
+    model: Union[Type[object], object],
+) -> Union[TypeIs[Type[VllmModelForTextGeneration]],
+           TypeIs[VllmModelForTextGeneration]]:
+    if not is_vllm_model(model):
+        return False
+
+    if isinstance(model, type):
+        return isinstance(model, VllmModelForTextGeneration)
+
+    return isinstance(model, VllmModelForTextGeneration)
+
+
+@runtime_checkable
+class VllmModelForPooling(VllmModel[T], Protocol[T]):
+    """The interface required for all pooling models in vLLM."""
+
+    def pooler(
+        self,
+        hidden_states: T,
+        pooling_metadata: "PoolingMetadata",
+    ) -> "PoolerOutput":
+        """Only called on TP rank 0."""
+        ...
+
+
+@overload
+def is_pooling_model(model: Type[object]) -> TypeIs[Type[VllmModelForPooling]]:
+    ...
+
+
+@overload
+def is_pooling_model(model: object) -> TypeIs[VllmModelForPooling]:
+    ...
+
+
+def is_pooling_model(
+    model: Union[Type[object], object],
+) -> Union[TypeIs[Type[VllmModelForPooling]], TypeIs[VllmModelForPooling]]:
+    if not is_vllm_model(model):
+        return False
+
+    if isinstance(model, type):
+        return isinstance(model, VllmModelForPooling)
+
+    return isinstance(model, VllmModelForPooling)

.venv/lib/python3.11/site-packages/vllm/model_executor/models/intern_vit.py

@@ -0,0 +1,476 @@
+# SPDX-License-Identifier: Apache-2.0
+
+# adapted from https://huggingface.co/OpenGVLab/InternVL2-4B/blob/main/modeling_intern_vit.py
+# --------------------------------------------------------
+# InternVL
+# Copyright (c) 2023 OpenGVLab
+# Licensed under The MIT License [see LICENSE for details]
+# --------------------------------------------------------
+from functools import partial
+from typing import Iterable, Optional, Set, Tuple
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from transformers import PretrainedConfig
+
+from vllm.attention.layer import MultiHeadAttention
+from vllm.distributed import (divide, 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.layers.activation import get_act_fn
+from vllm.model_executor.layers.layernorm import RMSNorm
+from vllm.model_executor.layers.linear import (ColumnParallelLinear,
+                                               QKVParallelLinear,
+                                               RowParallelLinear)
+from vllm.model_executor.layers.quantization import QuantizationConfig
+from vllm.model_executor.model_loader.weight_utils import default_weight_loader
+
+NORM2FN = {
+    'rms_norm': RMSNorm,
+    'layer_norm': nn.LayerNorm,
+}
+
+
+class InternVisionEmbeddings(nn.Module):
+
+    def __init__(self, config: PretrainedConfig):
+        super().__init__()
+        self.config = config
+        self.embed_dim = config.hidden_size
+        self.image_size = config.image_size
+        self.patch_size = config.patch_size
+
+        self.class_embedding = nn.Parameter(torch.randn(1, 1, self.embed_dim))
+
+        self.patch_embedding = nn.Conv2d(in_channels=3,
+                                         out_channels=self.embed_dim,
+                                         kernel_size=self.patch_size,
+                                         stride=self.patch_size)
+
+        self.num_patches = (self.image_size // self.patch_size)**2
+        self.num_positions = self.num_patches + 1
+
+        self.position_embedding = nn.Parameter(
+            torch.randn(1, self.num_positions, self.embed_dim))
+
+    def _get_pos_embed(self, pos_embed: torch.Tensor, H: int, W: int):
+        target_dtype = pos_embed.dtype
+        pos_embed = pos_embed.float().reshape(
+            1, self.image_size // self.patch_size,
+            self.image_size // self.patch_size, -1).permute(0, 3, 1, 2)
+        pos_embed = F.interpolate(pos_embed,
+                                  size=(H, W),
+                                  mode='bicubic',
+                                  align_corners=False)
+        return pos_embed.reshape(1, -1, H * W).permute(0, 2,
+                                                       1).to(target_dtype)
+
+    def _get_position_embedding(self, H: int, W: int) -> torch.Tensor:
+        position_embedding = self.position_embedding
+        if self.num_patches == H * W:
+            return position_embedding
+
+        return torch.cat(
+            [
+                position_embedding[:, :1, :],
+                self._get_pos_embed(position_embedding[:, 1:, :], H, W),
+            ],
+            dim=1,
+        )
+
+    def forward(self, pixel_values: torch.FloatTensor) -> torch.Tensor:
+        target_dtype = self.patch_embedding.weight.dtype
+        patch_embeds = self.patch_embedding(pixel_values.to(
+            target_dtype))  # shape = [*, channel, width, height]
+        batch_size, _, height, width = patch_embeds.shape
+        patch_embeds = patch_embeds.flatten(2).transpose(1, 2)
+        class_embeds = self.class_embedding.expand(batch_size, 1,
+                                                   -1).to(target_dtype)
+        embeddings = torch.cat([class_embeds, patch_embeds], dim=1)
+        position_embedding = self._get_position_embedding(height, width)
+        embeddings = embeddings + position_embedding.to(target_dtype)
+        return embeddings
+
+
+class InternVisionPatchModel(nn.Module):
+
+    def __init__(self, config: PretrainedConfig):
+        super().__init__()
+        self.config = config
+        self.embeddings = InternVisionEmbeddings(config)
+
+    def get_input_embeddings(self):
+        return self.embeddings
+
+    def forward(
+        self,
+        pixel_values: Optional[torch.Tensor] = None,
+        pixel_embeds: Optional[torch.Tensor] = None,
+    ) -> torch.FloatTensor:
+        if pixel_values is None and pixel_embeds is None:
+            raise ValueError(
+                'You have to specify pixel_values or pixel_embeds')
+
+        if pixel_embeds is not None:
+            hidden_states = pixel_embeds
+        elif pixel_values is not None:
+            if pixel_values.ndim == 4:
+                hidden_states = self.embeddings(pixel_values)
+            else:
+                raise ValueError(
+                    f'wrong pixel_values size: {pixel_values.shape}')
+
+        return hidden_states
+
+
+class InternParallelAttention(nn.Module):
+    """Multi-headed attention from 'Attention Is All You Need' paper"""
+
+    def __init__(
+        self,
+        config: PretrainedConfig,
+        quant_config: Optional[QuantizationConfig] = None,
+        *,
+        num_dummy_heads: int = 0,
+        prefix: str = "",
+    ) -> None:
+        super().__init__()
+
+        self.config = config
+        self.embed_dim = config.hidden_size
+        self.num_heads = config.num_attention_heads
+        self.head_dim = self.embed_dim // self.num_heads
+        if self.head_dim * self.num_heads != self.embed_dim:
+            raise ValueError(
+                f'embed_dim must be divisible by num_heads '
+                f'(got `embed_dim`: {self.embed_dim} and `num_heads`:'
+                f' {self.num_heads}).')
+
+        self.tp_size = get_tensor_model_parallel_world_size()
+        self.tp_rank = get_tensor_model_parallel_rank()
+
+        # Additional dummy heads are used to enable TP for common GPU counts.
+        self.dummy_dim = (num_dummy_heads + self.num_heads) * self.head_dim
+        self.num_heads_per_partition = divide(num_dummy_heads + self.num_heads,
+                                              self.tp_size)
+
+        self.scale = self.head_dim**-0.5
+        self.qkv = QKVParallelLinear(
+            self.embed_dim,
+            self.head_dim,
+            num_dummy_heads + self.num_heads,
+            bias=config.qkv_bias,
+            quant_config=quant_config,
+            prefix=f"{prefix}.qkv",
+        )
+
+        self.qk_normalization = config.qk_normalization
+
+        if self.qk_normalization:
+            self.q_norm = RMSNorm(self.dummy_dim,
+                                  eps=config.layer_norm_eps,
+                                  var_hidden_size=self.embed_dim)
+            self.k_norm = RMSNorm(self.dummy_dim,
+                                  eps=config.layer_norm_eps,
+                                  var_hidden_size=self.embed_dim)
+
+        self.proj = RowParallelLinear(
+            self.dummy_dim,
+            self.embed_dim,
+            quant_config=quant_config,
+            prefix=f"{prefix}.proj",
+        )
+
+        self.attn = MultiHeadAttention(self.num_heads_per_partition,
+                                       self.head_dim, self.scale)
+
+    def _apply_qk_norm(self, q: torch.Tensor, k: torch.Tensor):
+        if self.tp_size > 1:
+            q = tensor_model_parallel_all_gather(q.contiguous())
+            k = tensor_model_parallel_all_gather(k.contiguous())
+        q = self.q_norm.forward_native(q)
+        k = self.k_norm.forward_native(k)
+        if self.tp_size > 1:
+            splitter = partial(split_tensor_along_last_dim,
+                               num_partitions=self.tp_size)
+            q = splitter(q)[self.tp_rank]
+            k = splitter(k)[self.tp_rank]
+        return q, k
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        B, N, _ = x.shape
+        qkv, _ = self.qkv(x)
+        q, k, v = qkv.chunk(3, dim=-1)
+
+        if self.qk_normalization:
+            q, k = self._apply_qk_norm(q, k)
+
+        out = self.attn(q, k, v)
+        out, _ = self.proj(out)
+        return out
+
+
+class InternSdpaAttention(nn.Module):
+    """Multi-headed attention from 'Attention Is All You Need' paper"""
+
+    def __init__(
+        self,
+        config: PretrainedConfig,
+        *,
+        num_dummy_heads: int = 0,
+    ) -> None:
+        super().__init__()
+
+        self.config = config
+        self.embed_dim = config.hidden_size
+        self.num_heads = config.num_attention_heads
+        self.head_dim = self.embed_dim // self.num_heads
+        if self.head_dim * self.num_heads != self.embed_dim:
+            raise ValueError(
+                f'embed_dim must be divisible by num_heads '
+                f'(got `embed_dim`: {self.embed_dim} and `num_heads`:'
+                f' {self.num_heads}).')
+
+        # Additional dummy heads are used to enable TP for common GPU counts.
+        self.dummy_dim = (num_dummy_heads + self.num_heads) * self.head_dim
+
+        self.scale = self.head_dim**-0.5
+        self.qkv = nn.Linear(self.embed_dim,
+                             3 * self.dummy_dim,
+                             bias=config.qkv_bias)
+
+        self.qk_normalization = config.qk_normalization
+
+        if self.qk_normalization:
+            self.q_norm = RMSNorm(self.dummy_dim,
+                                  eps=config.layer_norm_eps,
+                                  var_hidden_size=self.embed_dim)
+            self.k_norm = RMSNorm(self.dummy_dim,
+                                  eps=config.layer_norm_eps,
+                                  var_hidden_size=self.embed_dim)
+
+        self.proj = nn.Linear(self.dummy_dim, self.embed_dim)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        B, N, C = x.shape
+        qkv = self.qkv(x)
+        q, k, v = qkv.chunk(3, dim=-1)
+
+        q = q.view(B, N, self.num_heads, self.head_dim)
+        k = k.view(B, N, self.num_heads, self.head_dim)
+        v = v.view(B, N, self.num_heads, self.head_dim)
+
+        if self.qk_normalization:
+            B_, N_, H_, D_ = q.shape
+            q = self.q_norm.forward_native(q.flatten(-2,
+                                                     -1)).view(B_, N_, H_, D_)
+            k = self.k_norm.forward_native(k.flatten(-2,
+                                                     -1)).view(B_, N_, H_, D_)
+        q = q.transpose(1, 2)
+        k = k.transpose(1, 2)
+        v = v.transpose(1, 2)
+
+        x = F.scaled_dot_product_attention(q, k, v, scale=self.scale)
+        x = x.transpose(1, 2).reshape(B, N, -1)
+
+        x = self.proj(x)
+        return x
+
+
+class InternMLP(nn.Module):
+
+    def __init__(
+        self,
+        config: PretrainedConfig,
+        quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
+    ) -> None:
+        super().__init__()
+
+        self.config = config
+        self.activation_fn = get_act_fn(config.hidden_act)
+        self.fc1 = ColumnParallelLinear(config.hidden_size,
+                                        config.intermediate_size,
+                                        bias=True,
+                                        quant_config=quant_config,
+                                        prefix=f"{prefix}.fc1")
+        self.fc2 = RowParallelLinear(config.intermediate_size,
+                                     config.hidden_size,
+                                     bias=True,
+                                     quant_config=quant_config,
+                                     prefix=f"{prefix}.fc2")
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        hidden_states, _ = self.fc1(hidden_states)
+        hidden_states = self.activation_fn(hidden_states)
+        hidden_states, _ = self.fc2(hidden_states)
+
+        return hidden_states
+
+
+class InternVisionEncoderLayer(nn.Module):
+
+    def __init__(
+        self,
+        config: PretrainedConfig,
+        quant_config: Optional[QuantizationConfig] = None,
+        *,
+        num_dummy_heads: int = 0,
+        prefix: str = "",
+    ) -> None:
+        super().__init__()
+
+        self.embed_dim = config.hidden_size
+        self.intermediate_size = config.intermediate_size
+        self.norm_type = config.norm_type
+
+        self.attn = self._init_attn(config,
+                                    quant_config,
+                                    num_dummy_heads=num_dummy_heads,
+                                    prefix=f"{prefix}.attn")
+
+        self.mlp = InternMLP(config,
+                             quant_config=quant_config,
+                             prefix=f"{prefix}.mlp")
+        self.norm1 = NORM2FN[self.norm_type](self.embed_dim,
+                                             eps=config.layer_norm_eps)
+        self.norm2 = NORM2FN[self.norm_type](self.embed_dim,
+                                             eps=config.layer_norm_eps)
+
+        self.ls1 = nn.Parameter(config.initializer_factor *
+                                torch.ones(self.embed_dim))
+        self.ls2 = nn.Parameter(config.initializer_factor *
+                                torch.ones(self.embed_dim))
+
+    def _init_attn(
+        self,
+        config: PretrainedConfig,
+        quant_config: Optional[QuantizationConfig],
+        *,
+        num_dummy_heads: int,
+        prefix: str = "",
+    ):
+        # fallback to sdpa attention if tp unavailable
+        tp_size = get_tensor_model_parallel_world_size()
+        num_heads = config.num_attention_heads
+
+        if (num_heads + num_dummy_heads) % tp_size == 0:
+            return InternParallelAttention(config,
+                                           quant_config=quant_config,
+                                           num_dummy_heads=num_dummy_heads,
+                                           prefix=prefix)
+
+        return InternSdpaAttention(config, num_dummy_heads=num_dummy_heads)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+    ):
+        hidden_states = hidden_states + self.attn(
+            self.norm1(hidden_states)) * self.ls1
+
+        hidden_states = hidden_states + self.mlp(
+            self.norm2(hidden_states)) * self.ls2
+
+        return hidden_states
+
+
+class InternVisionEncoder(nn.Module):
+
+    def __init__(
+        self,
+        config: PretrainedConfig,
+        quant_config: Optional[QuantizationConfig] = None,
+        *,
+        num_hidden_layers_override: Optional[int] = None,
+        num_dummy_heads: int = 0,
+        prefix: str = "",
+    ):
+        super().__init__()
+
+        self.config = config
+
+        if num_hidden_layers_override is None:
+            num_hidden_layers = config.num_hidden_layers
+        else:
+            num_hidden_layers = num_hidden_layers_override
+
+        self.layers = nn.ModuleList([
+            InternVisionEncoderLayer(config,
+                                     quant_config,
+                                     num_dummy_heads=num_dummy_heads,
+                                     prefix=f"{prefix}.layers.{layer_idx}")
+            for layer_idx in range(num_hidden_layers)
+        ])
+
+    def forward(self, inputs_embeds: torch.Tensor):
+
+        hidden_states = inputs_embeds
+        for encoder_layer in self.layers:
+            hidden_states = encoder_layer(hidden_states)
+
+        return hidden_states
+
+
+class InternVisionModel(nn.Module):
+
+    def __init__(
+        self,
+        config: PretrainedConfig,
+        quant_config: Optional[QuantizationConfig] = None,
+        *,
+        num_hidden_layers_override: Optional[int] = None,
+        num_dummy_heads: int = 0,
+        prefix: str = "",
+    ) -> None:
+        super().__init__()
+
+        self.config = config
+
+        self.embeddings = InternVisionEmbeddings(config)
+        self.encoder = InternVisionEncoder(
+            config=config,
+            quant_config=quant_config,
+            num_hidden_layers_override=num_hidden_layers_override,
+            num_dummy_heads=num_dummy_heads,
+            prefix=f"{prefix}.encoder",
+        )
+
+    def get_input_embeddings(self):
+        return self.embeddings
+
+    def forward(
+        self,
+        pixel_values: Optional[torch.Tensor] = None,
+        pixel_embeds: Optional[torch.Tensor] = None,
+    ) -> torch.FloatTensor:
+        if pixel_values is None and pixel_embeds is None:
+            raise ValueError(
+                'You have to specify pixel_values or pixel_embeds')
+
+        if pixel_embeds is not None:
+            hidden_states = pixel_embeds
+        elif pixel_values is not None:
+            if pixel_values.ndim == 4:
+                hidden_states = self.embeddings(pixel_values)
+            else:
+                raise ValueError(
+                    f'wrong pixel_values size: {pixel_values.shape}')
+
+        encoder_outputs = self.encoder(inputs_embeds=hidden_states)
+
+        return encoder_outputs
+
+    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:
+            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

.venv/lib/python3.11/site-packages/vllm/model_executor/models/internlm2_ve.py

@@ -0,0 +1,156 @@
+# SPDX-License-Identifier: Apache-2.0
+
+from typing import List, Optional, Tuple, Union
+
+import torch
+from torch import nn
+from transformers import PretrainedConfig
+
+from vllm.attention import AttentionMetadata
+from vllm.config import CacheConfig, VllmConfig
+from vllm.distributed import get_pp_group
+from vllm.model_executor.layers.layernorm import RMSNorm
+from vllm.model_executor.layers.quantization import QuantizationConfig
+from vllm.model_executor.models.internlm2 import (InternLM2Attention,
+                                                  InternLM2ForCausalLM,
+                                                  InternLM2MLP, InternLM2Model)
+from vllm.sequence import IntermediateTensors
+
+
+class InternLM2VEDecoderLayer(nn.Module):
+
+    def __init__(
+        self,
+        config: PretrainedConfig,
+        cache_config: Optional[CacheConfig] = None,
+        quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
+    ) -> None:
+        super().__init__()
+        self.hidden_size = config.hidden_size
+        rope_theta = getattr(config, "rope_theta", 10000)
+        rope_scaling = getattr(config, "rope_scaling", None)
+        max_position_embeddings = getattr(config, "max_position_embeddings",
+                                          8192)
+        self.attention = InternLM2Attention(
+            hidden_size=self.hidden_size,
+            num_heads=config.num_attention_heads,
+            num_kv_heads=config.num_key_value_heads,
+            rope_theta=rope_theta,
+            rope_scaling=rope_scaling,
+            max_position_embeddings=max_position_embeddings,
+            cache_config=cache_config,
+            quant_config=quant_config,
+            prefix=f"{prefix}.attention",
+        )
+        self.feed_forward = InternLM2MLP(
+            hidden_size=self.hidden_size,
+            intermediate_size=config.intermediate_size,
+            hidden_act=config.hidden_act,
+            quant_config=quant_config,
+            prefix=f"{prefix}.feed_forward",
+        )
+        self.feed_forward_ve = InternLM2MLP(
+            hidden_size=self.hidden_size,
+            intermediate_size=config.intermediate_size,
+            hidden_act=config.hidden_act,
+            quant_config=quant_config,
+            prefix=f"{prefix}.feed_forward_ve",
+        )
+        self.attention_norm = RMSNorm(config.hidden_size,
+                                      eps=config.rms_norm_eps)
+        self.ffn_norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+
+    def forward(
+        self,
+        positions: torch.Tensor,
+        hidden_states: torch.Tensor,
+        kv_cache: torch.Tensor,
+        attn_metadata: AttentionMetadata,
+        residual: Optional[torch.Tensor],
+        visual_token_mask: Optional[torch.Tensor] = None,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        # Self Attention
+        if residual is None:
+            residual = hidden_states
+            hidden_states = self.attention_norm(hidden_states)
+        else:
+            hidden_states, residual = self.attention_norm(
+                hidden_states, residual)
+        hidden_states = self.attention(
+            positions=positions,
+            hidden_states=hidden_states,
+            kv_cache=kv_cache,
+            attn_metadata=attn_metadata,
+        )
+
+        # Fully Connected
+        hidden_states, residual = self.ffn_norm(hidden_states, residual)
+        if visual_token_mask is not None and visual_token_mask.any():
+            visual_token_mask = visual_token_mask.repeat(
+                1, self.hidden_size).bool()
+            text_token_mask = ~visual_token_mask
+            hidden_states[visual_token_mask] = self.feed_forward_ve(
+                hidden_states[visual_token_mask].reshape(
+                    -1, self.hidden_size)).flatten()
+            if text_token_mask.any():
+                hidden_states[text_token_mask] = self.feed_forward(
+                    hidden_states[text_token_mask].reshape(
+                        -1, self.hidden_size)).flatten()
+        else:
+            hidden_states = self.feed_forward(hidden_states)
+        return hidden_states, residual
+
+
+class InternLM2VEModel(InternLM2Model):
+
+    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
+        super().__init__(vllm_config=vllm_config,
+                         prefix=prefix,
+                         layer_type=InternLM2VEDecoderLayer)
+
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        positions: torch.Tensor,
+        kv_caches: List[torch.Tensor],
+        attn_metadata: AttentionMetadata,
+        intermediate_tensors: Optional[IntermediateTensors] = None,
+        inputs_embeds: Optional[torch.Tensor] = None,
+        visual_token_mask: Optional[torch.Tensor] = None,
+    ) -> Union[torch.Tensor, IntermediateTensors]:
+        if get_pp_group().is_first_rank:
+            if inputs_embeds is not None:
+                hidden_states = inputs_embeds
+            else:
+                hidden_states = self.tok_embeddings(input_ids)
+            residual = None
+        else:
+            assert intermediate_tensors is not None
+            hidden_states = intermediate_tensors["hidden_states"]
+            residual = intermediate_tensors["residual"]
+        for i in range(self.start_layer, self.end_layer):
+            layer = self.layers[i]
+            hidden_states, residual = layer(
+                positions,
+                hidden_states,
+                kv_caches[i - self.start_layer],
+                attn_metadata,
+                residual,
+                visual_token_mask=visual_token_mask,
+            )
+        if not get_pp_group().is_last_rank:
+            return IntermediateTensors({
+                "hidden_states": hidden_states,
+                "residual": residual
+            })
+        hidden_states, _ = self.norm(hidden_states, residual)
+        return hidden_states
+
+
+class InternLM2VEForCausalLM(InternLM2ForCausalLM):
+
+    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
+        super().__init__(vllm_config=vllm_config,
+                         prefix=prefix,
+                         model_type=InternLM2VEModel)

.venv/lib/python3.11/site-packages/vllm/model_executor/models/jais.py

@@ -0,0 +1,397 @@
+# SPDX-License-Identifier: Apache-2.0
+
+# Adapted from
+# https://huggingface.co/inceptionai/jais-30b-chat-v3/blob/main/modeling_jais.py
+# Copyright 2023 The vLLM team.
+# Copyright 2023 the Jais authors and HuggingFace Inc. team.  All rights
+# reserved.
+# Copyright 2023 Cerebras Systems.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Inference-only Jais model compatible with HuggingFace weights."""
+
+import math
+from typing import Iterable, List, Optional, Set, Tuple, Union
+
+import torch
+from torch import nn
+
+from vllm.attention import Attention, AttentionMetadata
+from vllm.compilation.decorators import support_torch_compile
+from vllm.config import CacheConfig, VllmConfig
+from vllm.distributed import (get_pp_group, get_tensor_model_parallel_rank,
+                              get_tensor_model_parallel_world_size)
+from vllm.model_executor.layers.linear import (ColumnParallelLinear,
+                                               QKVParallelLinear,
+                                               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
+from vllm.model_executor.sampling_metadata import SamplingMetadata
+from vllm.sequence import IntermediateTensors
+from vllm.transformers_utils.configs import JAISConfig
+
+from .interfaces import SupportsPP
+from .utils import (is_pp_missing_parameter,
+                    make_empty_intermediate_tensors_factory, make_layers,
+                    maybe_prefix)
+
+
+class SwiGLUActivation(nn.Module):
+
+    def forward(self, x1: torch.Tensor, x2: torch.Tensor) -> torch.Tensor:
+        return x1 * nn.functional.silu(x2)
+
+
+def _get_alibi_slopes(n):
+
+    def get_slopes_power_of_2(n):
+        start = 2**(-(2**-(math.log2(n) - 3)))
+        ratio = start
+        return [start * ratio**i for i in range(n)]
+
+    if math.log2(n).is_integer():
+        return get_slopes_power_of_2(n)
+    else:
+        closest_power_of_2 = 2**math.floor(math.log2(n))
+        return (get_slopes_power_of_2(closest_power_of_2) + _get_alibi_slopes(
+            2 * closest_power_of_2)[0::2][:n - closest_power_of_2])
+
+
+class JAISAttention(nn.Module):
+
+    def __init__(
+        self,
+        config: JAISConfig,
+        cache_config: Optional[CacheConfig] = None,
+        quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
+    ):
+        super().__init__()
+        self.hidden_size = config.hidden_size
+        total_num_heads = config.num_attention_heads
+        tensor_model_parallel_world_size = (
+            get_tensor_model_parallel_world_size())
+        assert total_num_heads % tensor_model_parallel_world_size == 0
+        self.num_heads = total_num_heads // tensor_model_parallel_world_size
+        self.head_dim = self.hidden_size // total_num_heads
+        if hasattr(config, "scale_qk_dot_by_d"):
+            config.mup_scale_qk_dot_by_d = config.scale_qk_dot_by_d
+        self.attn_scale_power = 1.0 if config.mup_scale_qk_dot_by_d else 0.5
+        self.scale = self.head_dim**-self.attn_scale_power
+
+        self.c_attn = QKVParallelLinear(
+            self.hidden_size,
+            self.head_dim,
+            total_num_heads,
+            bias=True,
+            quant_config=quant_config,
+        )
+        self.c_proj = RowParallelLinear(
+            self.hidden_size,
+            self.hidden_size,
+            bias=True,
+            quant_config=quant_config,
+        )
+
+        tp_rank = get_tensor_model_parallel_rank()
+        head_start = tp_rank * self.num_heads
+        head_end = (tp_rank + 1) * self.num_heads
+        alibi_slopes = _get_alibi_slopes(total_num_heads)
+        alibi_slopes = alibi_slopes[head_start:head_end]
+        self.attn = Attention(self.num_heads,
+                              self.head_dim,
+                              scale=self.scale,
+                              alibi_slopes=alibi_slopes,
+                              cache_config=cache_config,
+                              quant_config=quant_config,
+                              prefix=f"{prefix}.attn")
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        kv_cache: torch.Tensor,
+        attn_metadata: AttentionMetadata,
+    ) -> torch.Tensor:
+        qkv, _ = self.c_attn(hidden_states)
+        q, k, v = qkv.chunk(chunks=3, dim=-1)
+        attn_output = self.attn(q, k, v, kv_cache, attn_metadata)
+        attn_output, _ = self.c_proj(attn_output)
+        return attn_output
+
+
+class JAISMLP(nn.Module):
+
+    def __init__(
+        self,
+        intermediate_size: int,
+        config: JAISConfig,
+        quant_config: Optional[QuantizationConfig] = None,
+    ):
+        super().__init__()
+        hidden_size = config.hidden_size
+        self.swiglu = config.activation_function == "swiglu"
+        self.c_fc = ColumnParallelLinear(
+            hidden_size,
+            intermediate_size,
+            bias=True,
+            quant_config=quant_config,
+        )
+        self.c_fc2 = (ColumnParallelLinear(
+            hidden_size,
+            intermediate_size,
+            bias=True,
+            quant_config=quant_config,
+        ) if self.swiglu else None)
+        self.c_proj = RowParallelLinear(
+            intermediate_size,
+            hidden_size,
+            bias=True,
+            quant_config=quant_config,
+        )
+
+        self.act = SwiGLUActivation()
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        if self.swiglu:
+            hidden_states2, _ = self.c_fc2(hidden_states)
+        hidden_states, _ = self.c_fc(hidden_states)
+        hidden_states = (self.act(hidden_states, hidden_states2)
+                         if self.swiglu else self.act(hidden_states))
+        hidden_states, _ = self.c_proj(hidden_states)
+        return hidden_states
+
+
+class JAISBlock(nn.Module):
+
+    def __init__(
+        self,
+        config: JAISConfig,
+        cache_config: Optional[CacheConfig] = None,
+        quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
+    ):
+        super().__init__()
+        hidden_size = config.hidden_size
+        inner_dim = (config.n_inner if config.n_inner is not None else 4 *
+                     hidden_size)
+
+        self.ln_1 = nn.LayerNorm(hidden_size, eps=config.layer_norm_epsilon)
+        self.attn = JAISAttention(config,
+                                  cache_config,
+                                  quant_config,
+                                  prefix=f"{prefix}.attn")
+        self.ln_2 = nn.LayerNorm(hidden_size, eps=config.layer_norm_epsilon)
+        self.mlp = JAISMLP(inner_dim, config, quant_config)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        kv_cache: torch.Tensor,
+        attn_metadata: AttentionMetadata,
+    ) -> torch.Tensor:
+        residual = hidden_states
+        hidden_states = self.ln_1(hidden_states)
+        attn_output = self.attn(
+            hidden_states=hidden_states,
+            kv_cache=kv_cache,
+            attn_metadata=attn_metadata,
+        )
+        # residual connection
+        hidden_states = attn_output + residual
+
+        residual = hidden_states
+        hidden_states = self.ln_2(hidden_states)
+        feed_forward_hidden_states = self.mlp(hidden_states)
+        # residual connection
+        hidden_states = residual + feed_forward_hidden_states
+        return hidden_states
+
+
+@support_torch_compile
+class JAISModel(nn.Module):
+
+    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
+        super().__init__()
+
+        config = vllm_config.model_config.hf_config
+        cache_config = vllm_config.cache_config
+        quant_config = vllm_config.quant_config
+
+        self.config = config
+        assert not config.add_cross_attention
+        assert not config.scale_attn_by_inverse_layer_idx
+        assert not config.reorder_and_upcast_attn
+        self.embed_dim = config.hidden_size
+        self.wte = VocabParallelEmbedding(config.vocab_size, self.embed_dim)
+        self.wpe = (nn.Embedding(config.max_position_embeddings,
+                                 self.embed_dim)
+                    if config.position_embedding_type != "alibi" else None)
+        if hasattr(config, "embeddings_scale"):
+            self.embeddings_scale = config.embeddings_scale
+        else:
+            self.embeddings_scale = config.mup_embeddings_scale
+
+        self.start_layer, self.end_layer, self.h = make_layers(
+            config.num_hidden_layers,
+            lambda prefix: JAISBlock(config=config,
+                                     cache_config=cache_config,
+                                     quant_config=quant_config,
+                                     prefix=prefix),
+            prefix=f"{prefix}.h",
+        )
+
+        self.ln_f = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_epsilon)
+        self.make_empty_intermediate_tensors = (
+            make_empty_intermediate_tensors_factory(["hidden_states"],
+                                                    config.n_embd))
+
+    def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
+        return self.wte(input_ids)
+
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        position_ids: torch.Tensor,
+        kv_caches: List[torch.Tensor],
+        attn_metadata: AttentionMetadata,
+        intermediate_tensors: Optional[IntermediateTensors] = None,
+        inputs_embeds: Optional[torch.Tensor] = None,
+    ) -> Union[IntermediateTensors, torch.Tensor]:
+        if get_pp_group().is_first_rank:
+            if inputs_embeds is None:
+                inputs_embeds = self.get_input_embeddings(input_ids)
+            if self.wpe is not None:
+                position_embeds = self.wpe(position_ids)
+                hidden_states = inputs_embeds + position_embeds
+            else:
+                hidden_states = inputs_embeds
+            hidden_states *= torch.tensor(float(self.embeddings_scale),
+                                          dtype=hidden_states.dtype)
+        else:
+            assert intermediate_tensors is not None
+            hidden_states = intermediate_tensors["hidden_states"]
+
+        for i in range(self.start_layer, self.end_layer):
+            layer = self.h[i]
+            hidden_states = layer(hidden_states,
+                                  kv_caches[i - self.start_layer],
+                                  attn_metadata)
+
+        if not get_pp_group().is_last_rank:
+            return IntermediateTensors({"hidden_states": hidden_states})
+
+        hidden_states = self.ln_f(hidden_states)
+        return hidden_states
+
+
+class JAISLMHeadModel(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.transformer = JAISModel(vllm_config=vllm_config,
+                                     prefix=maybe_prefix(
+                                         prefix, "transformer"))
+        if self.config.tie_word_embeddings:
+            self.lm_head = self.transformer.wte
+        else:
+            self.lm_head = ParallelLMHead(self.config.vocab_size,
+                                          self.config.hidden_size)
+        if hasattr(config, "width_scale"):
+            self.output_logits_scale = config.width_scale
+        else:
+            self.output_logits_scale = (config.mup_output_alpha *
+                                        config.mup_width_scale)
+        self.logits_processor = LogitsProcessor(vocab_size=config.vocab_size,
+                                                scale=self.output_logits_scale)
+        self.sampler = get_sampler()
+        self.make_empty_intermediate_tensors = (
+            self.transformer.make_empty_intermediate_tensors)
+
+    def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
+        return self.transformer.get_input_embeddings(input_ids)
+
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        positions: torch.Tensor,
+        kv_caches: List[torch.Tensor],
+        attn_metadata: AttentionMetadata,
+        intermediate_tensors: Optional[IntermediateTensors] = None,
+        inputs_embeds: Optional[torch.Tensor] = None,
+    ) -> Union[IntermediateTensors, torch.Tensor]:
+        hidden_states = self.transformer(input_ids, positions, kv_caches,
+                                         attn_metadata, intermediate_tensors,
+                                         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)
+        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 "lm_head.weight" in name:
+                # GPT-2 ties the weights of the embedding layer and the final
+                # linear layer.
+                continue
+            if ".attn.bias" in name or ".attn.masked_bias" in name:
+                # Skip attention mask.
+                # NOTE: "c_attn.bias" should not be skipped.
+                continue
+            if "relative_pe" in name:
+                continue
+            if not name.startswith("transformer."):
+                name = "transformer." + name
+
+            if is_pp_missing_parameter(name, self):
+                continue
+
+            param = params_dict[name]
+            # The HF's GPT-2 implementation uses Conv1D instead of Linear.
+            # Because of this, we need to transpose the weights.
+            # Note(zhuohan): the logic below might break quantized models.
+            for conv1d_weight_name in ["c_attn", "c_proj", "c_fc"]:
+                if conv1d_weight_name not in name:
+                    continue
+                if not name.endswith(".weight"):
+                    continue
+                loaded_weight = loaded_weight.t()
+            weight_loader = getattr(param, "weight_loader",
+                                    default_weight_loader)
+            weight_loader(param, loaded_weight)
+            loaded_params.add(name)
+        return loaded_params

.venv/lib/python3.11/site-packages/vllm/model_executor/models/llava_next_video.py

@@ -0,0 +1,500 @@
+# SPDX-License-Identifier: Apache-2.0
+
+import math
+from functools import cached_property
+from typing import (Iterable, List, Literal, Mapping, Optional, Set, Tuple,
+                    TypedDict, Union)
+
+import torch
+import torch.nn as nn
+from transformers import (BatchFeature, LlavaNextVideoConfig,
+                          LlavaNextVideoProcessor)
+
+from vllm.attention import AttentionMetadata
+from vllm.config import VllmConfig
+from vllm.model_executor.layers.activation import get_act_fn
+from vllm.model_executor.layers.sampler import SamplerOutput, get_sampler
+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 (MultiModalFieldConfig, MultiModalKwargs,
+                                    NestedTensors)
+from vllm.multimodal.parse import (ImageSize, MultiModalDataItems,
+                                   VideoEmbeddingItems, VideoProcessorItems)
+from vllm.multimodal.processing import (BaseMultiModalProcessor,
+                                        BaseProcessingInfo, PromptReplacement)
+from vllm.multimodal.profiling import BaseDummyInputsBuilder, ProcessorInputs
+from vllm.sequence import IntermediateTensors
+from vllm.utils import is_list_of
+
+from .interfaces import SupportsMultiModal, SupportsPP
+from .llava import init_vision_tower_for_llava
+from .siglip import SiglipVisionModel
+from .utils import (AutoWeightsLoader, init_vllm_registered_model,
+                    maybe_prefix, merge_multimodal_embeddings)
+from .vision import get_vision_encoder_info
+
+
+class LlavaNextVideoPixelInputs(TypedDict):
+    type: Literal["pixel_values_videos"]
+    data: Union[torch.Tensor, List[torch.Tensor]]
+    """
+    Shape: `(batch_size, num_frames, num_channels, height, width)`
+
+    Note that `num_frames` may be different for each batch, in which case
+    the data is passed as a list instead of a batched tensor.
+
+    Note that it only supports one video input for one batch.
+    """
+
+
+class LlavaNextVideoProcessingInfo(BaseProcessingInfo):
+
+    def get_hf_config(self):
+        return self.ctx.get_hf_config(LlavaNextVideoConfig)
+
+    def get_vision_encoder_info(self):
+        return get_vision_encoder_info(self.get_hf_config())
+
+    def get_hf_processor(self):
+        return self.ctx.get_hf_processor(LlavaNextVideoProcessor)
+
+    def get_supported_mm_limits(self) -> Mapping[str, Optional[int]]:
+        return {"video": 1}
+
+    def get_mm_max_tokens_per_item(
+        self,
+        seq_len: int,
+        mm_counts: Mapping[str, int],
+    ) -> Mapping[str, int]:
+        target_width, target_height = self.get_image_size_with_most_features()
+
+        max_video_tokens = self.get_num_video_tokens(
+            image_width=target_width,
+            image_height=target_height,
+            num_frames=self.get_num_frames_with_most_features(seq_len),
+        )
+
+        return {"video": max_video_tokens}
+
+    def get_image_size_with_most_features(self) -> ImageSize:
+        vision_encoder_info = self.get_vision_encoder_info()
+        width = height = vision_encoder_info.get_image_size()
+        return ImageSize(width=width, height=height)
+
+    def _get_num_frame_tokens(
+        self,
+        *,
+        image_width: int,
+        image_height: int,
+    ) -> int:
+        hf_config = self.get_hf_config()
+        spatial_pool_stride = hf_config.spatial_pool_stride
+
+        vision_encoder_info = self.get_vision_encoder_info()
+        patch_grid_length = vision_encoder_info.get_patch_grid_length()
+        pooled_grid_length = math.ceil(patch_grid_length / spatial_pool_stride)
+
+        return pooled_grid_length * pooled_grid_length
+
+    def get_num_video_tokens(
+        self,
+        *,
+        image_width: int,
+        image_height: int,
+        num_frames: int,
+    ) -> int:
+        num_frame_tokens = self._get_num_frame_tokens(
+            image_width=image_width,
+            image_height=image_height,
+        )
+
+        return num_frame_tokens * num_frames
+
+    def _get_max_video_frames(self, max_tokens: int) -> int:
+        target_width, target_height = self.get_image_size_with_most_features()
+
+        num_frames = 0
+
+        while True:
+            next_num_frames = num_frames + 1
+            next_max_tokens = self.get_num_video_tokens(
+                image_width=target_width,
+                image_height=target_height,
+                num_frames=next_num_frames,
+            )
+
+            if next_max_tokens > max_tokens:
+                break
+
+            num_frames = next_num_frames
+
+        return num_frames
+
+    def get_num_frames_with_most_features(self, seq_len: int) -> int:
+        mm_config = self.ctx.get_mm_config()
+        max_videos = mm_config.limit_per_prompt.get("video", 1)
+
+        max_total_frames = self._get_max_video_frames(seq_len)
+
+        return max(max_total_frames // max(max_videos, 1), 1)
+
+
+class LlavaNextVideoDummyInputsBuilder(
+        BaseDummyInputsBuilder[LlavaNextVideoProcessingInfo]):
+
+    def get_dummy_processor_inputs(
+        self,
+        seq_len: int,
+        mm_counts: Mapping[str, int],
+    ) -> ProcessorInputs:
+        num_videos = mm_counts.get("video", 0)
+
+        processor = self.info.get_hf_processor()
+        video_token = processor.video_token
+
+        target_width, target_height = \
+            self.info.get_image_size_with_most_features()
+        target_num_frames = \
+            self.info.get_num_frames_with_most_features(seq_len)
+
+        mm_data = {
+            "video":
+            self._get_dummy_videos(
+                width=target_width,
+                height=target_height,
+                num_frames=target_num_frames,
+                num_videos=num_videos,
+            )
+        }
+
+        return ProcessorInputs(
+            prompt_text=video_token * num_videos,
+            mm_data=mm_data,
+        )
+
+
+class LlavaNextVideoMultiModalProcessor(
+        BaseMultiModalProcessor[LlavaNextVideoProcessingInfo]):
+
+    def _get_mm_fields_config(
+        self,
+        hf_inputs: BatchFeature,
+        hf_processor_mm_kwargs: Mapping[str, object],
+    ) -> Mapping[str, MultiModalFieldConfig]:
+        return dict(pixel_values_videos=MultiModalFieldConfig.batched("video"))
+
+    def _get_prompt_replacements(
+        self,
+        mm_items: MultiModalDataItems,
+        hf_processor_mm_kwargs: Mapping[str, object],
+        out_mm_kwargs: MultiModalKwargs,
+    ) -> list[PromptReplacement]:
+        hf_config = self.info.get_hf_config()
+        video_token_id = hf_config.video_token_index
+
+        def get_replacement(item_idx: int):
+            videos = mm_items.get_items(
+                "video", (VideoEmbeddingItems, VideoProcessorItems))
+
+            if isinstance(videos, VideoEmbeddingItems):
+                num_video_tokens = videos.get_feature_size(item_idx)
+            else:
+                image_size = videos.get_frame_size(item_idx)
+                num_video_tokens = self.info.get_num_video_tokens(
+                    image_width=image_size.width,
+                    image_height=image_size.height,
+                    num_frames=videos.get_num_frames(item_idx),
+                )
+
+            return [video_token_id] * num_video_tokens
+
+        return [
+            PromptReplacement(
+                modality="video",
+                target=[video_token_id],
+                replacement=get_replacement,
+            ),
+        ]
+
+
+# adopted from transformers modeling_llava_next_video.py
+class LlavaNextVideoPooler(nn.Module):
+
+    def __init__(self, config: LlavaNextVideoConfig):
+        super().__init__()
+
+        mode = config.spatial_pool_mode
+        stride = config.spatial_pool_stride
+        image_size = config.vision_config.image_size
+        patch_size = config.vision_config.patch_size
+        self.image_size = image_size // patch_size**2
+
+        if mode == "average":
+            self.pool = nn.AvgPool2d(kernel_size=stride, stride=stride)
+        elif mode == "max":
+            self.pool = nn.MaxPool2d(kernel_size=stride, stride=stride)
+        else:
+            # TODO: Support Conv2d pooling layer, need to load weights
+            raise ValueError(
+                f"Unknown pooling mode: {mode}. Expected [`average`, `max`]")
+
+    def forward(self, image_features: torch.Tensor):
+        ori_width = int(
+            math.sqrt(image_features.shape[1] * self.image_size //
+                      self.image_size))
+        ori_height = int(ori_width * self.image_size // self.image_size)
+
+        batch_size, _, dim = image_features.shape
+        image_features_spatial = image_features \
+            .view(batch_size, ori_height, ori_height, dim) \
+            .permute(0, 3, 1, 2)
+        image_features_spatial = self.pool(image_features_spatial)
+
+        return image_features_spatial.flatten(2).transpose(1, 2).contiguous()
+
+
+class LlavaNextMultiModalProjector(nn.Module):
+
+    def __init__(self, vision_hidden_size: int, text_hidden_size: int,
+                 projector_hidden_act: str, multimodal_projector_bias: bool):
+        super().__init__()
+
+        self.linear_1 = nn.Linear(vision_hidden_size,
+                                  text_hidden_size,
+                                  bias=multimodal_projector_bias)
+        self.act = get_act_fn(projector_hidden_act)
+        self.linear_2 = nn.Linear(text_hidden_size,
+                                  text_hidden_size,
+                                  bias=multimodal_projector_bias)
+
+    def forward(self, image_features: torch.Tensor) -> torch.Tensor:
+        hidden_states = self.linear_1(image_features)
+        hidden_states = self.act(hidden_states)
+        hidden_states = self.linear_2(hidden_states)
+        return hidden_states
+
+
+@MULTIMODAL_REGISTRY.register_processor(
+    LlavaNextVideoMultiModalProcessor,
+    info=LlavaNextVideoProcessingInfo,
+    dummy_inputs=LlavaNextVideoDummyInputsBuilder,
+)
+class LlavaNextVideoForConditionalGeneration(nn.Module, SupportsMultiModal,
+                                             SupportsPP):
+
+    def __init__(self, *, vllm_config: VllmConfig, prefix: str = "") -> None:
+        super().__init__()
+        config = vllm_config.model_config.hf_config
+        quant_config = vllm_config.quant_config
+        multimodal_config = vllm_config.model_config.multimodal_config
+
+        self.config = config
+        self.multimodal_config = multimodal_config
+
+        # Initialize the vision tower only up to the required feature layer
+        self.vision_tower = init_vision_tower_for_llava(
+            config,
+            quant_config,
+            require_post_norm=False,
+            prefix=maybe_prefix(prefix, "vision_tower"))
+        self.vision_resampler = LlavaNextVideoPooler(config)
+        self.multi_modal_projector = LlavaNextMultiModalProjector(
+            vision_hidden_size=config.vision_config.hidden_size,
+            text_hidden_size=config.text_config.hidden_size,
+            projector_hidden_act=config.projector_hidden_act,
+            multimodal_projector_bias=config.multimodal_projector_bias)
+        self.language_model = init_vllm_registered_model(
+            vllm_config=vllm_config,
+            hf_config=config.text_config,
+            prefix=maybe_prefix(prefix, "language_model"),
+        )
+
+        self.make_empty_intermediate_tensors = (
+            self.language_model.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_video_pixel_values(
+        self, data: Union[torch.Tensor, List[torch.Tensor]]
+    ) -> Union[torch.Tensor, List[torch.Tensor]]:
+
+        h = w = self.config.vision_config.image_size
+        expected_dims = (3, h, w)
+
+        def _validate_shape(d: torch.Tensor):
+            actual_dims = tuple(d.shape[2:])
+
+            if actual_dims != expected_dims:
+                expected_expr = ("num_frames", *map(str, expected_dims))
+                raise ValueError(
+                    "The expected shape of pixel values in each video frame "
+                    f"is {expected_expr}. You supplied {tuple(d.shape)}.")
+
+        for d in data:
+            _validate_shape(d)
+
+        return data
+
+    def _parse_and_validate_video_input(
+            self, **kwargs: object) -> Optional[LlavaNextVideoPixelInputs]:
+        """
+        A legal video input should have the following dimensions:
+        {
+            "pixel_values_videos" : 
+                List[b, Tensor(nb_frames, nb_channels, height, width)]
+        }
+        """
+        pixel_values = kwargs.pop("pixel_values_videos", None)
+
+        if pixel_values is None:
+            return None
+
+        if not (is_list_of(pixel_values,
+                           (torch.Tensor))  # different shape videos 
+                or isinstance(pixel_values,
+                              torch.Tensor)):  # same shape videos
+            raise ValueError("Incorrect type of pixel values. "
+                             f"Got type: {type(pixel_values)}")
+
+        return LlavaNextVideoPixelInputs(
+            type="pixel_values_videos",
+            data=pixel_values,
+        )
+
+    def _select_image_features(self, image_features: torch.Tensor, *,
+                               strategy: str) -> torch.Tensor:
+        if strategy == "default":
+            return image_features[:, 1:]
+        elif strategy == "full":
+            return image_features
+
+        raise ValueError(f"Unexpected select feature strategy: {strategy}")
+
+    def _video_pixels_to_features(
+        self,
+        vision_tower: Union[CLIPVisionModel, SiglipVisionModel],
+        pixel_values: torch.Tensor,
+    ) -> torch.Tensor:
+
+        # NOTE: we skip the step to select the vision feature layer since
+        # this is already done inside the vision tower
+        image_features = vision_tower(pixel_values)
+        image_features = self._select_image_features(
+            image_features,
+            strategy=self.config.vision_feature_select_strategy,
+        )
+        image_features = self.vision_resampler(image_features)
+        image_features = self.multi_modal_projector(image_features)
+        return image_features
+
+    def _process_video_pixels(self, inputs: LlavaNextVideoPixelInputs):
+        assert self.vision_tower is not None
+
+        video_pixels = inputs["data"]
+
+        if isinstance(video_pixels, torch.Tensor):
+            # TODO: support multiple videos per input
+            b, num_videos, num_frames, c, h, w = video_pixels.shape
+            assert (num_videos == 1)
+            stacked_pixels = video_pixels.view(b * num_videos * num_frames, c,
+                                               h, w)
+            stacked_embeddings = self._video_pixels_to_features(
+                self.vision_tower, stacked_pixels)
+            return stacked_embeddings.view(b, num_frames,
+                                           *stacked_embeddings.shape[1:])
+
+        elif is_list_of(video_pixels, torch.Tensor):
+            frames_per_videos = [v.shape[0] for v in video_pixels]
+            stacked_pixels = torch.cat(video_pixels, dim=0)
+            stacked_embeddings = self._video_pixels_to_features(
+                self.vision_tower, stacked_pixels)
+            return torch.split(stacked_embeddings, frames_per_videos, dim=0)
+
+        else:
+            raise ValueError(
+                f"Unsupported type of video input {type(video_pixels)}")
+
+    def get_multimodal_embeddings(self, **kwargs) -> Optional[NestedTensors]:
+        video_input = self._parse_and_validate_video_input(**kwargs)
+        if video_input is None:
+            return None
+        vision_embeddings = self._process_video_pixels(video_input)
+        return vision_embeddings
+
+    def get_input_embeddings(
+        self,
+        input_ids: torch.Tensor,
+        multimodal_embeddings: Optional[NestedTensors] = None,
+    ) -> torch.Tensor:
+        inputs_embeds = self.language_model.get_input_embeddings(input_ids)
+        if multimodal_embeddings is not None:
+            inputs_embeds = merge_multimodal_embeddings(
+                input_ids, inputs_embeds, multimodal_embeddings,
+                self.config.video_token_index)
+        return inputs_embeds
+
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        positions: torch.Tensor,
+        kv_caches: List[torch.Tensor],
+        attn_metadata: AttentionMetadata,
+        intermediate_tensors: Optional[IntermediateTensors] = None,
+        inputs_embeds: Optional[torch.Tensor] = None,
+        **kwargs: object,
+    ) -> Union[torch.Tensor, IntermediateTensors]:
+        """Run forward pass for LlaVA-NeXT-Video.
+        Args:
+            input_ids: Flattened (concatenated) input_ids corresponding to a
+                batch.
+            pixel_values_videos: Pixels in each frames for each input videos.
+        """
+        if intermediate_tensors is not None:
+            inputs_embeds = None
+
+        # NOTE: In v1, inputs_embeds is always generated at model runner, this
+        # condition is for v0 compatibility.
+        elif inputs_embeds is None:
+            vision_embeddings = self.get_multimodal_embeddings(**kwargs)
+            inputs_embeds = self.get_input_embeddings(input_ids,
+                                                      vision_embeddings)
+            input_ids = None
+
+        hidden_states = self.language_model.model(input_ids,
+                                                  positions,
+                                                  kv_caches,
+                                                  attn_metadata,
+                                                  intermediate_tensors,
+                                                  inputs_embeds=inputs_embeds)
+
+        return hidden_states
+
+    def compute_logits(
+        self,
+        hidden_states: torch.Tensor,
+        sampling_metadata: SamplingMetadata,
+    ) -> Optional[torch.Tensor]:
+        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,
+            # This model doesn't support images for now
+            ignore_unexpected_prefixes=["image_newline"],
+        )
+        return loader.load_weights(weights)