import math from typing import Iterable, List, Optional, Set, Tuple, TypedDict, Union import torch import torch.nn as nn from torch.nn.init import trunc_normal_ from transformers import LlamaConfig from vllm.attention import AttentionMetadata from vllm.config import CacheConfig, QuantizationConfig, VllmConfig from vllm.distributed import get_tensor_model_parallel_rank from vllm.inputs import INPUT_REGISTRY, token_inputs 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.quantization.compressed_tensors.utils import ( get_compressed_tensors_cache_scale) from vllm.model_executor.layers.sampler import (Sampler, SamplerOutput, SamplingMetadata) from vllm.model_executor.layers.vocab_parallel_embedding import ParallelLMHead from vllm.model_executor.model_loader.weight_utils import ( default_weight_loader, maybe_remap_kv_scale_name) from vllm.model_executor.models.idefics2_vision_model import ( Idefics2VisionTransformer) from vllm.model_executor.models.interfaces import SupportsMultiModal from vllm.model_executor.models.llama import (LlamaDecoderLayer, LlamaMLP, LlamaModel) from vllm.model_executor.models.utils import (AutoWeightsLoader, WeightsMapper, is_pp_missing_parameter, maybe_prefix, merge_multimodal_embeddings) from vllm.multimodal import MULTIMODAL_REGISTRY from vllm.multimodal.base import MultiModalInputs from vllm.multimodal.image import cached_get_image_processor from vllm.multimodal.inputs import NestedTensors from vllm.multimodal.utils import (cached_get_tokenizer, repeat_and_pad_placeholder_tokens) from vllm.sequence import IntermediateTensors from vllm.transformers_utils.configs.aria import (AriaMoELMConfig, AriaVisionConfig) from .utils import flatten_bn 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(Idefics2VisionTransformer): """ AriaVisionTransformer is a modified version of Idefics2VisionTransformer that replaces the post-layernorm with an identity layer. """ def __init__( self, config: AriaVisionConfig, quant_config: Optional[QuantizationConfig] = None, prefix: str = "", ) -> None: super().__init__(config, quant_config, prefix) self.post_layernorm = nn.Identity() class AriaVisionModel(nn.Module): config_class = AriaVisionConfig def __init__( self, config: AriaVisionConfig, quant_config: Optional[QuantizationConfig] = None, *, prefix: str = "", ) -> None: super().__init__() self.vision_model = AriaVisionTransformer( config, quant_config, prefix=f"{prefix}.vision_model", ) def forward( self, pixel_values: torch.Tensor, pixel_mask: Optional[torch.BoolTensor] = None, ) -> Tuple[torch.Tensor, Optional[torch.BoolTensor]]: patch_attention_mask = self._create_patch_attention_mask(pixel_mask) vit_oup = self.vision_model( pixel_values=pixel_values, patch_attention_mask=patch_attention_mask, ) image_atts = self._create_image_attention_mask(patch_attention_mask) return vit_oup, image_atts def _create_patch_attention_mask(self, pixel_mask): if pixel_mask is None: return None patches_subgrid = pixel_mask.unfold( dimension=1, size=self.vision_model.config.patch_size, step=self.vision_model.config.patch_size, ).unfold( dimension=2, size=self.vision_model.config.patch_size, step=self.vision_model.config.patch_size, ) return (patches_subgrid.sum(dim=(-1, -2)) > 0).bool() def _create_image_attention_mask(self, patch_attention_mask): if patch_attention_mask is None: return None flattened_mask = patch_attention_mask.flatten(1) return torch.logical_not(flattened_mask) class FFN(nn.Module): def __init__(self, embed_dim, ff_dim, output_dim): super().__init__() self.linear_in = ColumnParallelLinear(embed_dim, ff_dim, bias=False) self.linear_out = RowParallelLinear(ff_dim, output_dim, bias=False) self.act = get_act_fn("gelu_new") def forward(self, hidden_states): hidden_states, _ = self.linear_in(hidden_states) hidden_states = self.act(hidden_states) hidden_states, _ = self.linear_out(hidden_states) return hidden_states class CrossAttention(nn.Module): def __init__(self, kv_dim, embed_dim, num_heads, drop_out_rate=0): super().__init__() self.num_heads = num_heads self.q_proj = nn.Linear(embed_dim, embed_dim, bias=False) self.k_proj = nn.Linear(kv_dim, embed_dim, bias=False) self.v_proj = nn.Linear(kv_dim, embed_dim, bias=False) self.multihead_attn = nn.MultiheadAttention(embed_dim, num_heads) self.linear = nn.Linear(embed_dim, embed_dim) self.dropout = nn.Dropout(drop_out_rate) self.layer_norm = nn.LayerNorm(embed_dim) self.ln_kv = nn.LayerNorm(kv_dim) def forward(self, x, hidden_states, attn_mask=None, add_residual=False): normed_hidden_states = self.layer_norm(hidden_states) query = self.q_proj(normed_hidden_states).permute(1, 0, 2) x = self.ln_kv(x) key = self.k_proj(x).permute(1, 0, 2) value = self.v_proj(x).permute(1, 0, 2) attn_output, _ = self.multihead_attn(query, key, value, attn_mask=attn_mask) attn_output = attn_output.permute(1, 0, 2) if add_residual: attn_output = hidden_states + self.dropout( self.linear(attn_output)) else: attn_output = self.dropout(self.linear(attn_output)) return attn_output 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, patch_to_query_dict, embed_dim, num_heads, kv_dim, ff_dim, output_dim, norm_layer=nn.LayerNorm, ): super().__init__() self.patch_to_query_dict = patch_to_query_dict self.embed_dim = embed_dim self.num_heads = num_heads self.query = nn.Parameter( torch.zeros(max(patch_to_query_dict.values()), self.embed_dim)) trunc_normal_(self.query, std=0.02) self.cross_attn = CrossAttention(kv_dim, embed_dim, num_heads) self.ln_ffn = norm_layer(embed_dim) self.ffn = FFN(embed_dim, ff_dim, output_dim) def forward(self, x, attn_mask=None): bs = x.shape[0] queries = self.query.unsqueeze(0).repeat(bs, 1, 1) query_num = self.patch_to_query_dict.get(x.shape[1], None) assert (query_num is not None ), f"Query number for {x.shape[1]} patches is not provided" queries = queries[:, :query_num, :] 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.ffn(self.ln_ffn(attention_out)) return out class AriaFusedMoE(FusedMoE): def weight_loader(self, param: nn.Parameter, loaded_weight: torch.Tensor, shard_id: str) -> Set[str]: # 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 MoELayer(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: AriaMoELMConfig, 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.moe_intermediate_size, quant_config=quant_config, reduce_results=True, ) self.shared_experts = LlamaMLP( config.hidden_size, config.moe_intermediate_size * config.moe_num_shared_experts, "silu", quant_config=quant_config, ) 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) shared_expert_output = self.shared_experts(hidden_states) sparse_expert_output = self.experts(hidden_states, router_output) return sparse_expert_output + shared_expert_output class MoEDecoderLayer(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: LlamaConfig, cache_config: Optional[CacheConfig] = None, quant_config: Optional[QuantizationConfig] = None, prefix: str = "", ) -> None: super().__init__(config, cache_config, quant_config, prefix) self.mlp = MoELayer(config, quant_config=quant_config) class AriaMoELMModel(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=MoEDecoderLayer) # 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 scale_name := get_compressed_tensors_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, 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 def build_mm_projector(config): return AriaProjector( patch_to_query_dict=config.projector_patch_to_query_dict, embed_dim=config.vision_config.hidden_size, num_heads=config.vision_config.num_attention_heads, kv_dim=config.vision_config.hidden_size, ff_dim=config.text_config.hidden_size, output_dim=config.text_config.hidden_size, ) def get_max_multimodal_tokens(ctx): return max(ctx.model_config.hf_config.image_size2tokens.values()) def input_mapper_for_aria(ctx, data): return MultiModalInputs(data) def input_processor(ctx, llm_inputs): multi_modal_data = llm_inputs.get("multi_modal_data") # if it is pure text input, use it as is if multi_modal_data is None or "image" not in multi_modal_data: return llm_inputs model_config = ctx.model_config tokenizer = cached_get_tokenizer(model_config.tokenizer) image_processor = cached_get_image_processor( model_config.model, trust_remote_code=model_config.trust_remote_code) hf_config = model_config.hf_config # prepare image tokens, the max_image_size is used to determine the number # of patch_size for every image max_image_size = multi_modal_data.pop("max_image_size", 980) _split_image = multi_modal_data.pop("split_image", False) assert isinstance(max_image_size, (int, float)), "max_image_size should be float or int" images = (multi_modal_data["image"] if isinstance( multi_modal_data["image"], list) else [multi_modal_data["image"]]) image_inputs = image_processor.preprocess(images, max_image_size=max_image_size, split_image=_split_image, return_tensors="pt").data image_inputs['pixel_values'] = image_inputs['pixel_values'].to( ctx.model_config.dtype) num_crops = image_inputs.pop("num_crops") prompt_token_ids = llm_inputs["prompt_token_ids"] if num_crops.sum().item() > 0: _, prompt_token_ids, _ = repeat_and_pad_placeholder_tokens( tokenizer, None, prompt_token_ids, placeholder_token_id=hf_config.image_token_index, repeat_count=num_crops, ) repeat_count = [hf_config.image_size2tokens[max_image_size] ] * sum(num_crops).item() new_prompt, new_token_ids, _ = repeat_and_pad_placeholder_tokens( tokenizer, None, prompt_token_ids, placeholder_token_id=hf_config.image_token_index, repeat_count=repeat_count, ) return token_inputs( prompt_token_ids=new_token_ids, prompt=new_prompt, multi_modal_data={"image": image_inputs}, ) @MULTIMODAL_REGISTRY.register_max_image_tokens(get_max_multimodal_tokens) @MULTIMODAL_REGISTRY.register_image_input_mapper(input_mapper_for_aria) @INPUT_REGISTRY.register_input_processor(input_processor) 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. """ def __init__( self, vllm_config: VllmConfig, prefix: str = "", ): super().__init__() config = vllm_config.model_config.hf_config quant_config = vllm_config.quant_config # prepare the image_size to tokens mapping for the image preprocess, see # input_processor config.image_size2tokens = { int(math.sqrt(k) * config.vision_config.patch_size): v for k, v in config.projector_patch_to_query_dict.items() } self.config = config self.vision_tower = AriaVisionModel(config.vision_config) self.multi_modal_projector = build_mm_projector(config) self.vocab_size = config.text_config.vocab_size self.language_model = AriaMoELMModel( 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 = 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: pixel_mask = flatten_bn(pixel_mask, concat=True) return AriaImagePixelInputs( pixel_values=pixel_values, pixel_mask=pixel_mask, ) 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'] image_feature, image_attn_mask = self.vision_tower( pixel_values, pixel_mask=pixel_mask) return self.multi_modal_projector(image_feature, 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]]): 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", }, ) loader = AutoWeightsLoader(self) loader.load_weights(weights, mapper=hf_to_vllm_mapper)