# SPDX-License-Identifier: Apache-2.0 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project import math from collections.abc import Iterable, Mapping, Sequence from typing import Annotated, Any, Literal, TypeAlias import numpy as np import torch import torch.nn as nn import torch.nn.functional as F from transformers import ( BatchFeature, Phi4MultimodalAudioConfig, Phi4MultimodalConfig, Phi4MultimodalFeatureExtractor, Phi4MultimodalImageProcessorFast, ) from transformers import Phi4MultimodalProcessor as Phi4MMProcessor from transformers.models.phi4_multimodal.modeling_phi4_multimodal import ( Phi4MultimodalAudioConvModule, Phi4MultimodalAudioNemoConvSubsampling, Phi4MultimodalAudioRelativeAttentionBias, adaptive_enc_mask, unfold_tensor, ) from vllm.config import VllmConfig from vllm.config.multimodal import BaseDummyOptions from vllm.distributed import ( divide, get_tensor_model_parallel_rank, get_tensor_model_parallel_world_size, ) from vllm.model_executor.layers.activation import MulAndSilu, get_act_fn from vllm.model_executor.layers.linear import ( ColumnParallelLinear, MergedColumnParallelLinear, QKVParallelLinear, RowParallelLinear, ) from vllm.model_executor.layers.quantization import QuantizationConfig from vllm.model_executor.model_loader.weight_utils import default_weight_loader from vllm.model_executor.models.module_mapping import MultiModelKeys from vllm.multimodal import MULTIMODAL_REGISTRY from vllm.multimodal.inputs import ( MultiModalDataDict, MultiModalFieldConfig, MultiModalKwargsItems, NestedTensors, ) from vllm.multimodal.parse import ( AudioProcessorItems, ImageEmbeddingItems, ImageProcessorItems, ImageSize, MultiModalDataItems, MultiModalDataParser, ) from vllm.multimodal.processing import ( BaseMultiModalProcessor, BaseProcessingInfo, PromptReplacement, PromptUpdate, ) from vllm.multimodal.profiling import BaseDummyInputsBuilder from vllm.sequence import IntermediateTensors from vllm.utils.tensor_schema import TensorSchema, TensorShape from .idefics2_vision_model import Idefics2VisionTransformer from .interfaces import MultiModalEmbeddings, SupportsLoRA, SupportsMultiModal from .utils import ( AutoWeightsLoader, WeightsMapper, init_vllm_registered_model, maybe_prefix, ) _AUDIO_MAX_SOUNDFILE_SIZE = 241_000 def _get_padding_size( orig_width: int, orig_height: int, target_height: int, target_width: int ): ratio_width = target_width / orig_width ratio_height = target_height / orig_height if ratio_width < ratio_height: padding_width = 0 padding_height = target_height - int(orig_height * ratio_width) else: padding_width = target_width - int(orig_width * ratio_height) padding_height = 0 return padding_height, padding_width class Phi4MMProjector(nn.Module): def __init__(self, input_size: int, hidden_size: int): super().__init__() self.up = ColumnParallelLinear(input_size, hidden_size) self.down = RowParallelLinear(hidden_size, hidden_size) self.act = get_act_fn("gelu") def forward(self, x: torch.Tensor) -> torch.Tensor: x, _ = self.up(x) x = self.act(x) x, _ = self.down(x) return x class Phi4MMImageEmbedding(nn.Module): """Image embedding.""" def __init__(self, config: Phi4MultimodalConfig): super().__init__() self.config = config self.layer_idx = config.vision_config.feature_layer self.crop_size = config.vision_config.crop_size self.image_dim_out = config.vision_config.hidden_size n_patches = config.vision_config.image_size // config.vision_config.patch_size if n_patches % 2 != 0: self.img_processor_padding = nn.ReflectionPad2d((0, 1, 0, 1)) n_patches += 1 self.num_img_tokens = (n_patches // 2) ** 2 num_hidden_layers = ( config.vision_config.num_hidden_layers + self.layer_idx + 1 if self.layer_idx < 0 else self.layer_idx + 1 ) self.img_processor = Idefics2VisionTransformer( config.vision_config, require_post_norm=False, num_hidden_layers_override=num_hidden_layers, ) self.image_token_compression = nn.AvgPool2d(kernel_size=2, stride=2) self.img_projection = Phi4MMProjector(self.image_dim_out, config.hidden_size) self.global_img_feature_extensor = nn.Parameter( torch.zeros([1, 1, self.image_dim_out]) ) self.sub_img_feature_extensor = nn.Parameter( torch.zeros([1, 1, 1, self.image_dim_out]) ) def get_img_features( self, img_embeds: torch.FloatTensor, attention_mask: torch.Tensor | None = None, ) -> torch.FloatTensor: img_feature = self.img_processor( img_embeds, patch_attention_mask=attention_mask ) patch_feature = img_feature # reshape to 2D tensor width = int(math.sqrt(patch_feature.size(1))) patch_feature = patch_feature.view(-1, width, width, patch_feature.size(-1)) # convert to NCHW patch_feature = patch_feature.permute(0, 3, 1, 2) if getattr(self, "img_processor_padding", None) is not None: patch_feature = self.img_processor_padding(patch_feature) patch_feature = self.image_token_compression(patch_feature) # convert to NHWC patch_feature = patch_feature.permute(0, 2, 3, 1) patch_feature = patch_feature.view( -1, patch_feature.size(1) * patch_feature.size(2), patch_feature.size(-1) ) return patch_feature def forward( self, image_pixel_values: torch.FloatTensor, image_sizes: torch.Tensor | None = None, image_attention_mask: torch.Tensor | None = None, ) -> torch.FloatTensor: image_pixel_values = image_pixel_values.to( self.img_processor.embeddings.patch_embedding.weight.dtype ) target_device = self.img_projection.up.bias.device target_dtype = self.img_projection.up.bias.dtype batch_size = image_pixel_values.shape[0] img_features = self.get_img_features( image_pixel_values.flatten(0, 1), attention_mask=image_attention_mask.flatten(0, 1).to( dtype=bool, device=target_device ), ) base_feat_size = int(np.sqrt(img_features.shape[1])) img_features = img_features.view( batch_size, -1, base_feat_size**2, self.image_dim_out ) image_sizes = image_sizes.view(-1, 2) output_imgs = [] for idx in range(batch_size): height, width = image_sizes[idx] height_ratio = height // self.crop_size width_ratio = width // self.crop_size area_ratio = height_ratio * width_ratio global_img = img_features[idx, :1] global_img = global_img.reshape( 1, base_feat_size, base_feat_size, self.image_dim_out ).contiguous() temporary_extensor = self.sub_img_feature_extensor.repeat( 1, base_feat_size, 1, 1 ) global_img = torch.cat([global_img, temporary_extensor], dim=2).reshape( 1, -1, self.image_dim_out ) sub_img = img_features[idx, 1:] sub_img = sub_img[:area_ratio] sub_img = ( sub_img.reshape( height_ratio, width_ratio, base_feat_size, base_feat_size, self.image_dim_out, ) .transpose(1, 2) .reshape( 1, height_ratio * base_feat_size, width_ratio * base_feat_size, self.image_dim_out, ) .contiguous() ) if image_attention_mask is not None: reshaped_image_attention_mask = ( image_attention_mask[idx, 1 : area_ratio + 1, 0::2, 0::2] .reshape(height_ratio, width_ratio, base_feat_size, base_feat_size) .transpose(1, 2) .reshape( 1, height_ratio * base_feat_size, width_ratio * base_feat_size ) ) useful_height = int(reshaped_image_attention_mask[0, :, 0].sum().item()) useful_width = int(reshaped_image_attention_mask[0, 0, :].sum().item()) sub_img = sub_img[:, :useful_height, :useful_width] temporary_extensor = self.sub_img_feature_extensor.repeat( 1, useful_height, 1, 1 ) else: temporary_extensor = self.sub_img_feature_extensor.repeat( 1, height_ratio * base_feat_size, 1, 1 ) sub_img = torch.cat([sub_img, temporary_extensor], dim=2).reshape( 1, -1, self.image_dim_out ) # Merge global and sub output_imgs.append( torch.cat( [sub_img, self.global_img_feature_extensor, global_img], dim=1 ) ) img_set_tensor = [] for output_img in output_imgs: output_img = output_img.to(device=target_device, dtype=target_dtype) img_feature_proj = self.img_projection(output_img) img_set_tensor.append(img_feature_proj.flatten(0, 1)) return img_set_tensor class Phi4MultimodalAudioMLP(nn.Module): def __init__( self, config: Phi4MultimodalAudioConfig, quant_config: QuantizationConfig | None = None, prefix: str = "", ): super().__init__() self.layer_norm = nn.LayerNorm(config.hidden_size) self.act_fn = MulAndSilu() self.gate_up_proj = MergedColumnParallelLinear( config.hidden_size, [config.intermediate_size] * 2, bias=True, quant_config=quant_config, prefix=f"{prefix}.gate_up_proj", ) self.down_proj = RowParallelLinear( config.intermediate_size, config.hidden_size, bias=True, quant_config=quant_config, prefix=f"{prefix}.down_proj", ) def forward(self, hidden_states: torch.Tensor) -> torch.Tensor: hidden_states = self.layer_norm(hidden_states) hidden_states, _ = self.gate_up_proj(hidden_states) hidden_states = self.act_fn(hidden_states) hidden_states, _ = self.down_proj(hidden_states) return hidden_states class Phi4MultimodalAudioAttention(nn.Module): def __init__( self, config: Phi4MultimodalAudioConfig, quant_config: QuantizationConfig | None = None, prefix: str = "", ): super().__init__() self.config = config self.embed_dim = config.hidden_size self.total_num_heads = config.num_attention_heads self.head_dim = self.embed_dim // self.total_num_heads if self.head_dim * self.total_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.qkv_proj = QKVParallelLinear( hidden_size=self.embed_dim, head_size=self.head_dim, total_num_heads=self.total_num_heads, quant_config=quant_config, prefix=f"{prefix}.qkv_proj", ) self.o_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.tp_rank = get_tensor_model_parallel_rank() self.num_heads = divide(self.total_num_heads, self.tp_size) def split_attn_mask(self, attention_mask: torch.Tensor) -> torch.Tensor: start_idx = self.num_heads * self.tp_rank end_idx = self.num_heads * (self.tp_rank + 1) return attention_mask[:, start_idx:end_idx] def forward( self, hidden_states: torch.Tensor, attention_mask: torch.Tensor, ) -> torch.Tensor: qkv_states, _ = self.qkv_proj(hidden_states) query, key, value = qkv_states.chunk(3, dim=-1) bsz, seq_len, _ = query.size() query = query.view(bsz, seq_len, self.num_heads, self.head_dim) key = key.view(bsz, seq_len, self.num_heads, self.head_dim) value = value.view(bsz, seq_len, self.num_heads, self.head_dim) query, key, value = (x.transpose(1, 2) for x in (query, key, value)) attention_mask = self.split_attn_mask(attention_mask) out = F.scaled_dot_product_attention( query, key, value, scale=self.scale, attn_mask=attention_mask, ) out = out.transpose(1, 2).reshape(bsz, seq_len, -1) attn_output, _ = self.o_proj(out) return attn_output class Phi4MultimodalAudioConformerEncoderLayer(nn.Module): def __init__(self, config: Phi4MultimodalAudioConfig): super().__init__() self.feed_forward_in = Phi4MultimodalAudioMLP(config) self.self_attn = Phi4MultimodalAudioAttention(config) self.conv = Phi4MultimodalAudioConvModule(config) self.feed_forward_out = Phi4MultimodalAudioMLP(config) self.layer_norm_att = nn.LayerNorm(config.hidden_size) self.layer_norm = nn.LayerNorm(config.hidden_size) def forward( self, hidden_states: torch.Tensor, attention_mask: torch.Tensor, ) -> torch.Tensor: residual = hidden_states + 0.5 * self.feed_forward_in(hidden_states) hidden_states = self.layer_norm_att(residual) hidden_states = residual + self.self_attn(hidden_states, attention_mask) hidden_states = hidden_states + self.conv(hidden_states) hidden_states = hidden_states + 0.5 * self.feed_forward_out(hidden_states) out = self.layer_norm(hidden_states) return out class Phi4MMAudioMeanVarianceNormLayer(nn.Module): """Mean/variance normalization layer. Will subtract mean and multiply input by inverted standard deviation. Typically used as a very first layer in a model. Args: config: [Phi4MultimodalAudioConfig](https://huggingface.co/docs/transformers/model_doc/phi4_multimodal#transformers.Phi4MultimodalAudioConfig) object containing model parameters. """ def __init__(self, config: Phi4MultimodalAudioConfig): super().__init__() self.global_mean = nn.Parameter(torch.zeros(config.input_size)) self.global_invstd = nn.Parameter(torch.ones(config.input_size)) def forward(self, input_: torch.Tensor) -> torch.Tensor: """MeanVarianceNormLayer Forward Args: input_: torch.Tensor input tensor. """ return (input_ - self.global_mean) * self.global_invstd class Phi4MultimodalAudioModel(nn.Module): def __init__(self, config: Phi4MultimodalAudioConfig): super().__init__() self.config = config self.encoder_embedding = Phi4MMAudioMeanVarianceNormLayer(config) self.embed = Phi4MultimodalAudioNemoConvSubsampling(config) self.relative_attention_bias_layer = Phi4MultimodalAudioRelativeAttentionBias( config ) self.encoders = nn.ModuleList( [ Phi4MultimodalAudioConformerEncoderLayer(config) for _ in range(config.num_blocks) ] ) def _streaming_mask( self, seq_len: int, batch_size: int, chunk_size: int, left_chunk: int, ): # Create mask matrix for streaming # S stores start index. if chunksize is 18, s is [0,18,36,....] chunk_start_idx = np.arange(0, seq_len, chunk_size) enc_streaming_mask = ( adaptive_enc_mask(seq_len, chunk_start_idx, left_window=left_chunk) .unsqueeze(0) .expand([batch_size, -1, -1]) ) return enc_streaming_mask def forward_embeddings( self, hidden_states: torch.Tensor, masks: torch.Tensor, ): """Forwarding the inputs through the top embedding layers""" seq_len = math.ceil(hidden_states.shape[1] / self.config.time_reduction) if seq_len <= 0: raise ValueError( f"Sequence length after time reduction is invalid: {seq_len}." "Your input feature is too short." ) batch_size = hidden_states.shape[0] enc_streaming_mask = self._streaming_mask( seq_len, batch_size, self.config.chunk_size, self.config.left_chunk ) enc_streaming_mask = enc_streaming_mask.to(hidden_states.device) hidden_states, masks = self.embed(hidden_states, masks) streaming_mask = enc_streaming_mask if streaming_mask is not None and masks is not None: hs_mask = masks & streaming_mask elif masks is not None: hs_mask = masks else: hs_mask = streaming_mask return hidden_states, hs_mask, masks def calculate_hs_mask( self, hidden_states: torch.Tensor, device: torch.device, mask: torch.Tensor ): max_audio_length = hidden_states.shape[1] batch_size = hidden_states.shape[0] enc_streaming_mask = self._streaming_mask( max_audio_length, batch_size, self.config.chunk_size, self.config.left_chunk ) enc_streaming_mask = enc_streaming_mask.to(device) if mask is None: return enc_streaming_mask feature_lens = mask.sum(1) padding_length = feature_lens pad_mask = torch.arange(0, max_audio_length, device=device).expand( padding_length.size(0), -1 ) < padding_length.unsqueeze(1) pad_mask = pad_mask.unsqueeze(1) pad_mask = pad_mask & enc_streaming_mask return pad_mask def forward(self, hidden_states: torch.Tensor, mask: torch.Tensor | None = None): hidden_states = self.encoder_embedding(hidden_states) hidden_states, hs_mask, mask = self.forward_embeddings(hidden_states, mask) unfolded = False bs, seq_len, _ = hidden_states.shape max_seq_len = 500 # maximum position for absolute positional encoding if seq_len > max_seq_len: # audio sequence is longer than max_seq_len, # unfold it into chunks of max_seq_len unfolded = True # the unfold op will drop residual frames, # pad it to the multiple of max_seq_len if seq_len % max_seq_len > 0: chunk_pad_size = max_seq_len - (seq_len % max_seq_len) else: chunk_pad_size = 0 if chunk_pad_size > 0: hidden_states_pad = F.pad( hidden_states, (0, 0, 0, chunk_pad_size), "constant", 0 ) hidden_states = hidden_states_pad.to(hidden_states.device) hidden_states = unfold_tensor(hidden_states, max_seq_len) masks_unfold = None if mask is not None: # revise hs_mask here because the previous calculated hs_mask # did not consider extra pad subsampled_pad_mask = mask.squeeze(1) # [bz, subsampled_unmask_seq_len] extra_padded_subsamlped_pad_mask = F.pad( subsampled_pad_mask, (0, chunk_pad_size), "constant", False ) # extra padding to the pad mask extra_padded_subsamlped_pad_mask = ( extra_padded_subsamlped_pad_mask.unsqueeze(-1).float() ) masks_unfold = unfold_tensor( extra_padded_subsamlped_pad_mask, max_seq_len ) # unfold the pad mask like we did to the input tensor masks_unfold = masks_unfold.squeeze( -1 ).bool() # unfold op does not support bool tensor hs_mask = self.calculate_hs_mask( hidden_states, hidden_states.device, masks_unfold ) # calculate hs_mask based on the unfolded pad mask relative_attention_bias = self.relative_attention_bias_layer(hidden_states) attention_mask = hs_mask.unsqueeze(1) + relative_attention_bias for layer in self.encoders: hidden_states = layer(hidden_states, attention_mask) if unfolded: embed_dim = hidden_states.shape[-1] hidden_states = hidden_states.reshape(bs, -1, embed_dim) # if we ever padded before unfolding, we need to remove the padding if chunk_pad_size > 0: hidden_states = hidden_states[:, :-chunk_pad_size, :] return hidden_states class Phi4MMAudioEmbedding(nn.Module): def __init__(self, config: Phi4MultimodalConfig): super().__init__() self.config = config self.layer_idx = config.audio_config.feature_layer self.encoder = Phi4MultimodalAudioModel(config.audio_config) audio_config = config.audio_config proj_input_size = audio_config.hidden_size * audio_config.downsample_rate self.vision_speech_projection = Phi4MMProjector( proj_input_size, config.hidden_size ) self.speech_projection = Phi4MMProjector(proj_input_size, config.hidden_size) def get_projection( self, audio_projection_mode: Literal["speech", "vision"], ) -> Phi4MMProjector: if audio_projection_mode == "speech": return self.speech_projection elif audio_projection_mode == "vision": return self.vision_speech_projection def forward( self, audio_input_features: torch.FloatTensor, audio_embed_sizes=None, audio_attention_mask=None, audio_projection_mode="speech", ) -> torch.FloatTensor: audio_projection = self.get_projection(audio_projection_mode) target_device = audio_projection.up.bias.device target_dtype = audio_projection.up.bias.dtype audio_input_features = audio_input_features.to( device=target_device, dtype=target_dtype ) audio_encoder_hidden_states = self.encoder( audio_input_features, audio_attention_mask ) audio_embeds = audio_projection(audio_encoder_hidden_states) return audio_embeds.flatten(0, 1) 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 class Phi4MMImagePixelInputs(TensorSchema): """ Dimensions: - bn: Batch size * number of images - p: Number of patches (1 + num_patches) - c: Number of channels (3) - h: Height of each image patch - w: Width of each image patch - nc: Number of crops - H_mask: Height of attention mask - W_mask: Width of attention mask """ type: Literal["pixel_values"] pixel_values: Annotated[ torch.Tensor | list[torch.Tensor], TensorShape( "bn", "p", 3, "h", "w", dynamic_dims={"p"} ), # may be different per batch and image ] image_sizes: Annotated[ torch.Tensor, TensorShape("bn", 2), # (height, width) ] num_img_tokens: Annotated[ list[int], TensorShape("bn"), ] image_attention_mask: Annotated[ torch.Tensor, TensorShape("bn", "nc", 32, 32), # H_mask, W_mask ] class Phi4MMImageEmbeddingInputs(TensorSchema): """ Dimensions: - bn: Batch size * number of images - f: Image feature size - h: Hidden size (must match language model backbone) """ type: Literal["image_embeds"] data: Annotated[ torch.Tensor | list[torch.Tensor], TensorShape("bn", "f", "h"), ] class Phi4MMAudioFeatureInputs(TensorSchema): """ Dimensions: - bn: Batch size * number of audios - f: Number of Mel filterbank bins (80) - t: Time frames (M) """ type: Literal["audio_features"] audio_features: Annotated[ torch.Tensor | list[torch.Tensor], TensorShape("bn", "t", 80, dynamic_dims={"t"}), ] class Phi4MMAudioEmbeddingInputs(TensorSchema): """ Dimensions: - b: Batch size - n: Number of audios - f: Audio feature size - h: Hidden size (must match language model backbone) """ type: Literal["audio_embeds"] data: Annotated[ NestedTensors, TensorShape("b", "n", "f", "h"), ] Phi4MMImageInput: TypeAlias = Phi4MMImagePixelInputs | Phi4MMImageEmbeddingInputs Phi4MMAudioInputs: TypeAlias = Phi4MMAudioFeatureInputs | Phi4MMAudioEmbeddingInputs def cat_with_pad(tensors, dim, padding_value=0): """ cat along dim, while pad to max for all other dims """ ndim = tensors[0].dim() assert all(t.dim() == ndim for t in tensors[1:]), ( "All tensors must have the same number of dimensions" ) out_size = [max(t.shape[i] for t in tensors) for i in range(ndim)] out_size[dim] = sum(t.shape[dim] for t in tensors) output = tensors[0].new_full(out_size, padding_value) index = 0 for t in tensors: # Create a slice list where every dimension except dim is full slice slices = [slice(0, t.shape[d]) for d in range(ndim)] # Update only the concat dimension slice slices[dim] = slice(index, index + t.shape[dim]) output[slices] = t index += t.shape[dim] return output class Phi4MMProcessingInfo(BaseProcessingInfo): def get_hf_config(self) -> Phi4MultimodalConfig: return self.ctx.get_hf_config(Phi4MultimodalConfig) def get_hf_processor(self, **kwargs: object) -> Phi4MMProcessor: return self.ctx.get_hf_processor(Phi4MMProcessor, **kwargs) def get_feature_extractor(self, **kwargs: object) -> Phi4MultimodalFeatureExtractor: return self.get_hf_processor(**kwargs).audio_processor def get_image_processor( self, processor: Phi4MMProcessor | None = None, ) -> Phi4MultimodalImageProcessorFast: if processor is None: processor = self.get_hf_processor() return processor.image_processor def get_dynamic_hd( self, processor: Phi4MMProcessor | None = None, ) -> int: return self.get_image_processor(processor).dynamic_hd def get_supported_mm_limits(self) -> Mapping[str, int | None]: return {"audio": None, "image": None} def _find_target_aspect_ratio( self, orig_width: int, orig_height: int, image_size: int, max_num: int, min_num: int, ): w_crop_num = math.ceil(orig_width / float(image_size)) h_crop_num = math.ceil(orig_height / float(image_size)) if w_crop_num * h_crop_num > max_num: aspect_ratio = orig_width / orig_height # calculate the existing image aspect ratio target_ratios = set( (i, j) for i in range(1, max_num + 1) for j in range(1, max_num + 1) if i * j <= max_num and i * j >= min_num ) target_ratios = sorted(target_ratios, key=lambda x: x[0] * x[1]) # find the closest aspect ratio to the target image_processor = self.get_image_processor() target_aspect_ratio = image_processor.find_closest_aspect_ratio( aspect_ratio, target_ratios, orig_width, orig_height, image_size, ) # calculate the target width and height target_width = image_size * target_aspect_ratio[0] target_height = image_size * target_aspect_ratio[1] else: target_width = image_size * w_crop_num target_height = image_size * h_crop_num target_aspect_ratio = (w_crop_num, h_crop_num) return target_aspect_ratio, target_height, target_width def _compute_num_image_tokens( self, orig_width: int, orig_height: int, dynamic_hd_size: int, vit_image_size: int, vit_patch_size: int, token_compression_factor: int = 2, ): """ compute the number of tokens an image is expected to take up considering the image encoder architecture and exclude output features containing only padding pixels for siglip, vit_image_size=448, vit_patch_size=14, so output will be 32x32 feature map NOTE right now, Phi4MM uses hard-coded token_compression_factor=2 """ assert vit_image_size % vit_patch_size == 0, ( "vit_image_size must be divisible by vit_patch_size" ) assert vit_image_size // vit_patch_size % token_compression_factor == 0, ( "vit_image_size // vit_patch_size must be divisible by " "token_compression_factor" ) target_aspect_ratio, target_height, target_width = ( self._find_target_aspect_ratio( orig_width, orig_height, vit_image_size, dynamic_hd_size, min_num=1 ) ) assert target_aspect_ratio[0] * vit_image_size == target_width, ( f"{target_aspect_ratio[0]} * {vit_image_size} != {target_width}" ) assert target_aspect_ratio[1] * vit_image_size == target_height, ( f"{target_aspect_ratio[1]} * {vit_image_size} != {target_height}" ) assert ( target_height % vit_image_size == 0 and target_width % vit_image_size == 0 ) padding_height, padding_width = _get_padding_size( orig_width, orig_height, target_height, target_width ) assert padding_width == 0 or padding_height == 0, ( "padding_width or padding_height must be 0" ) target_feat_width = target_width // vit_patch_size target_feat_height = target_height // vit_patch_size if padding_width >= vit_patch_size: assert padding_height == 0, "padding_height not 0" non_pad_feat_width = target_feat_width - math.floor( padding_width / vit_patch_size ) non_pad_feat_height = target_feat_height elif padding_height >= vit_patch_size: assert padding_width == 0, "padding_width not 0" non_pad_feat_height = target_feat_height - math.floor( padding_height / vit_patch_size ) non_pad_feat_width = target_feat_width else: # small padding shorter than a vit patch non_pad_feat_width = target_feat_width non_pad_feat_height = target_feat_height feat_width = non_pad_feat_width // token_compression_factor feat_height = non_pad_feat_height // token_compression_factor # NOTE it's possible that the non-padding feature is not divisible if non_pad_feat_width % token_compression_factor != 0: feat_width += 1 if non_pad_feat_height % token_compression_factor != 0: feat_height += 1 num_hd_patch_tokens = feat_width * feat_height num_hd_newline_tokens = feat_height vit_feature_size = vit_image_size // vit_patch_size num_global_image_tokens = (vit_feature_size // token_compression_factor) ** 2 num_sep_tokens = 1 num_global_image_newline_tokens = vit_feature_size // token_compression_factor return ( num_global_image_tokens + num_sep_tokens + num_hd_patch_tokens + num_hd_newline_tokens + num_global_image_newline_tokens ) def get_num_image_tokens( self, *, image_width: int, image_height: int, processor: Phi4MMProcessor | None = None, ) -> int: hf_config = self.get_hf_config() vision_config = hf_config.vision_config vit_image_size = vision_config.image_size vit_patch_size = vision_config.patch_size dynamic_hd_size = self.get_dynamic_hd(processor=processor) # we use default `token_compression_factor=2`, # since it's not in HF vision config. image_num_tokens = self._compute_num_image_tokens( image_width, image_height, dynamic_hd_size=dynamic_hd_size, vit_image_size=vit_image_size, vit_patch_size=vit_patch_size, ) return image_num_tokens def get_image_size_with_most_features( self, processor: Phi4MMProcessor | None = None, ) -> ImageSize: vit_image_size = self.get_hf_config().vision_config.image_size max_side = vit_image_size * self.get_dynamic_hd(processor=processor) return ImageSize(height=max_side, width=vit_image_size) def get_audio_num_frames(self, audio_len: int, sr: float) -> int: """ Compute the output size of the `extract_features` method. Args: audio_len (int): Length of the input waveform in samples. sr (float): Sampling rate of the waveform, either 16000 or 8000. Returns: tuple (int, int): Output size as (T, D), where: T: Number of time frames. D: Number of Mel filterbank bins (80). """ # Resample to 16000 or 8000 if needed if sr > 16000: audio_len //= sr // 16000 elif 8000 <= sr < 16000: # We'll resample to 16K from 8K audio_len *= 2 elif sr < 8000: raise RuntimeError(f"Unsupported sample rate {sr}") # Spectrogram parameters for 16 kHz win_length = 400 # Frame length in samples hop_length = 160 # Frame shift in samples # Calculate number of frames (T) num_frames = (audio_len - win_length) // hop_length + 1 if num_frames < 1: raise ValueError("Waveform too short for given parameters.") # Return time frames (T) return num_frames def _compute_audio_embed_size(self, audio_frames: int) -> int: """ Compute the size of audio embeddings from the number of audio frames. """ # `_compute_audio_embed_size` in audio_processor use torch for # computation, therefore we re-implement it to use pythonic # numeric computation to avoid extra tensor conversion. audio_processor = self.get_feature_extractor() audio_compression_rate = audio_processor.audio_compression_rate audio_downsample_rate = audio_processor.audio_downsample_rate integer = audio_frames // audio_compression_rate remainder = audio_frames % audio_compression_rate result = integer + int(remainder > 0) integer = result // audio_downsample_rate remainder = result % audio_downsample_rate result = integer + int(remainder > 0) # qformer compression return result class Phi4MMDummyInputsBuilder(BaseDummyInputsBuilder[Phi4MMProcessingInfo]): def get_dummy_text(self, mm_counts: Mapping[str, int]) -> str: num_audios = mm_counts.get("audio", 0) num_images = mm_counts.get("image", 0) tokenizer = self.info.get_tokenizer() image_tokens: str = tokenizer.image_token * num_images audio_tokens: str = tokenizer.audio_token * num_audios return image_tokens + audio_tokens def get_dummy_mm_data( self, seq_len: int, mm_counts: Mapping[str, int], mm_options: Mapping[str, BaseDummyOptions] | None = None, ) -> MultiModalDataDict: num_audios = mm_counts.get("audio", 0) num_images = mm_counts.get("image", 0) target_width, target_height = self.info.get_image_size_with_most_features() image_overrides = mm_options.get("image") if mm_options else None audio_overrides = mm_options.get("audio") if mm_options else None mm_data = { "image": self._get_dummy_images( width=target_width, height=target_height, num_images=num_images, overrides=image_overrides, ), "audio": self._get_dummy_audios( length=_AUDIO_MAX_SOUNDFILE_SIZE, num_audios=num_audios, overrides=audio_overrides, ), } return mm_data class Phi4MMMultiModalProcessor(BaseMultiModalProcessor[Phi4MMProcessingInfo]): def _get_data_parser(self) -> MultiModalDataParser: feature_extractor = self.info.get_feature_extractor() return MultiModalDataParser(target_sr=feature_extractor.sampling_rate) def _call_hf_processor( self, prompt: str, mm_data: Mapping[str, object], mm_kwargs: Mapping[str, object], tok_kwargs: Mapping[str, object], ) -> BatchFeature: if not mm_data: prompt_ids = self.info.get_tokenizer().encode(prompt) prompt_ids = self._apply_hf_processor_tokens_only(prompt_ids) return BatchFeature(dict(input_ids=[prompt_ids]), tensor_type="pt") audio_data = mm_data.pop("audios", []) if audio_data: mm_data["audio"] = audio_data processed_outputs = super()._call_hf_processor( prompt, mm_data, mm_kwargs, tok_kwargs ) if "image_pixel_values" in processed_outputs: num_img_tokens = [ self.info.get_num_image_tokens( image_width=img_size[0], image_height=img_size[1] ) for img_size in processed_outputs["image_sizes"] ] processed_outputs["num_img_tokens"] = num_img_tokens if audio_data: audio_features = processed_outputs["audio_input_features"] sr = self.info.get_feature_extractor(**mm_kwargs).sampling_rate feature_sizes = [ self.info.get_audio_num_frames(len(audio), sr) for audio in audio_data ] processed_outputs["audio_input_features"] = [ audio_features[idx, :size] for idx, size in enumerate(feature_sizes) ] return processed_outputs def _get_mm_fields_config( self, hf_inputs: BatchFeature, hf_processor_mm_kwargs: Mapping[str, object], ) -> Mapping[str, MultiModalFieldConfig]: return dict( image_pixel_values=MultiModalFieldConfig.batched("image"), image_attention_mask=MultiModalFieldConfig.batched("image"), image_sizes=MultiModalFieldConfig.batched("image"), num_img_tokens=MultiModalFieldConfig.batched("image"), audio_input_features=MultiModalFieldConfig.batched("audio"), ) def _get_prompt_updates( self, mm_items: MultiModalDataItems, hf_processor_mm_kwargs: Mapping[str, Any], out_mm_kwargs: MultiModalKwargsItems, ) -> Sequence[PromptUpdate]: tokenizer = self.info.get_tokenizer() image_token_id: int = tokenizer.vocab[tokenizer.image_token] audio_token_id: int = tokenizer.vocab[tokenizer.audio_token] hf_processor = self.info.get_hf_processor(**hf_processor_mm_kwargs) audio_processor = self.info.get_feature_extractor(**hf_processor_mm_kwargs) def get_image_replacement_phi4mm(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, processor=hf_processor, ) return [image_token_id] * num_image_tokens def get_audio_replacement_phi4mm(item_idx: int): audios = mm_items.get_items("audio", AudioProcessorItems) # TODO(Isotr0py): support embedding inputs audio_len = audios.get_audio_length(item_idx) audio_frames = self.info.get_audio_num_frames( audio_len, audio_processor.sampling_rate ) audio_embed_size = self.info._compute_audio_embed_size(audio_frames) return [audio_token_id] * audio_embed_size return [ PromptReplacement( modality="audio", target=[audio_token_id], replacement=get_audio_replacement_phi4mm, ), PromptReplacement( modality="image", target=[image_token_id], replacement=get_image_replacement_phi4mm, ), ] @MULTIMODAL_REGISTRY.register_processor( Phi4MMMultiModalProcessor, info=Phi4MMProcessingInfo, dummy_inputs=Phi4MMDummyInputsBuilder, ) class Phi4MultimodalForCausalLM(nn.Module, SupportsLoRA, SupportsMultiModal): """ Implements the Phi-4-multimodal-instruct model in vLLM. """ merge_by_field_config = True packed_modules_mapping = { "qkv_proj": [ "qkv_proj", ], "gate_up_proj": [ "gate_up_proj", ], } hf_to_vllm_mapper = WeightsMapper( orig_to_new_prefix={ # Multimodal embedding "model.embed_tokens_extend.": "", # LLM backbone "model.": "language_model.model.", }, orig_to_new_substr={ # projection ".img_projection_": ".img_projection.", ".up_proj_for_speech.": ".speech_projection.up.", ".up_proj_for_vision_speech.": ".vision_speech_projection.up.", ".down_proj_for_speech.": ".speech_projection.down.", ".down_proj_for_vision_speech.": ".vision_speech_projection.down.", }, ) @classmethod def get_placeholder_str(cls, modality: str, i: int) -> str | None: if modality.startswith("image"): return "<|image|>" if modality.startswith("audio"): return "<|audio|>" raise ValueError("Only image or audio modality is supported") def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): super().__init__() config = vllm_config.model_config.hf_config multimodal_config = vllm_config.model_config.multimodal_config self.config = config self.multimodal_config = multimodal_config # TODO: Optionally initializes these for supporting input embeddings. self.image_embed = Phi4MMImageEmbedding( config, # prefix=maybe_prefix(prefix, "image_embed"), ) self.audio_embed = Phi4MMAudioEmbedding( config, # prefix=maybe_prefix(prefix, "audio_embed"), ) self.language_model = init_vllm_registered_model( vllm_config=vllm_config, prefix=maybe_prefix(prefix, "language_model"), architectures=["Phi3ForCausalLM"], ) self.make_empty_intermediate_tensors = ( self.language_model.make_empty_intermediate_tensors ) def _parse_and_validate_audio_input( self, **kwargs: object ) -> Phi4MMAudioInputs | None: """ Parse and validate the audio input to the model. This handles both audio features and audio embeddings, but only the former is used for now. Args: kwargs (object): Keyword arguments. Returns: Optional[Phi4MMAudioInputs]: Parsed and validated audio inputs. """ audio_features = kwargs.pop("audio_input_features", None) audio_embeds = kwargs.pop("audio_embeds", None) if audio_features is None and audio_embeds is None: return None if audio_features is not None: return Phi4MMAudioFeatureInputs( type="audio_features", audio_features=audio_features, ) if audio_embeds is not None: return Phi4MMAudioEmbeddingInputs(type="audio_embeds", data=audio_embeds) raise AssertionError("This line should be unreachable.") def _process_audio_input( self, audio_input: Phi4MMAudioInputs, audio_projection_mode: str ) -> NestedTensors: """ Create the audio embeddings from the audio input, where the audio input is pairs of audio features and audio embed lengths. The audio input is created by `input_mapper_for_phi4mm_audio`. Args: audio_input (Phi4MMAudioInputs): Audio input. Returns: NestedTensors: Audio embeddings """ if audio_input["type"] == "audio_embeds": return audio_input["data"] audio_features = audio_input["audio_features"] # (e.g. multiple examples) and the second dim is the multi-audio dim # (e.g. multiple audios in the same example) dtype = next(self.audio_embed.parameters()).dtype audio_embeds = [ self.audio_embed( features.unsqueeze(0).to(dtype), audio_projection_mode=audio_projection_mode, ) for features in audio_features ] return audio_embeds def _parse_and_validate_image_input( self, **kwargs: object ) -> Phi4MMImagePixelInputs | None: pixel_values = kwargs.get("image_pixel_values") if pixel_values is None: return None image_sizes = kwargs.get("image_sizes") image_attention_mask = kwargs.get("image_attention_mask") num_img_tokens = kwargs.get("num_img_tokens") assert ( image_sizes is not None and image_attention_mask is not None and num_img_tokens is not None ), "Missing image inputs" return Phi4MMImagePixelInputs( type="pixel_values", pixel_values=pixel_values, image_sizes=image_sizes, image_attention_mask=image_attention_mask, num_img_tokens=num_img_tokens, ) def _parse_and_validate_multimodal_inputs(self, **kwargs: object) -> dict: modalities = {} # Preserve the order of modalities if there are multiple of them # from the order of kwargs. for input_key in kwargs: if ( input_key in ("image_pixel_values", "image_embeds") and "images" not in modalities ): modalities["images"] = self._parse_and_validate_image_input(**kwargs) if ( input_key in ("audio_input_features", "audio_embeds") and "audios" not in modalities ): modalities["audios"] = self._parse_and_validate_audio_input(**kwargs) return modalities def _process_image_input( self, image_input: Phi4MMImagePixelInputs ) -> list[torch.Tensor]: if image_input["type"] == "image_embeds": image_embeds = image_input["image_embeds"].type(self.visual.dtype) else: dtype = next(self.image_embed.parameters()).dtype pixel_values = image_input["pixel_values"].to(dtype) image_sizes = image_input["image_sizes"] image_attention_mask = image_input["image_attention_mask"] image_embeds = self.image_embed( pixel_values, image_sizes, image_attention_mask ) return image_embeds def embed_multimodal(self, **kwargs: object) -> MultiModalEmbeddings: modalities = self._parse_and_validate_multimodal_inputs(**kwargs) if not modalities: return [] # The result multimodal_embeddings is tuple of tensors, with each # tensor corresponding to a multimodal data item (image or video). multimodal_embeddings: tuple[torch.Tensor, ...] = () # NOTE: It is important to iterate over the keys in this dictionary # to preserve the order of the modalities. audio_projection_mode = "speech" for modality in modalities: # make sure process images first if modality == "images": audio_projection_mode = "vision" image_input = modalities["images"] image_embeddings = self._process_image_input(image_input) multimodal_embeddings += tuple(image_embeddings) if modality == "audios": audio_input = modalities["audios"] audio_embeddings = self._process_audio_input( audio_input, audio_projection_mode=audio_projection_mode ) multimodal_embeddings += tuple(audio_embeddings) return multimodal_embeddings def forward( self, input_ids: torch.Tensor, positions: torch.Tensor, intermediate_tensors: IntermediateTensors | None = None, inputs_embeds: torch.Tensor | None = None, **kwargs: object, ) -> torch.Tensor: if intermediate_tensors is not None: inputs_embeds = None hidden_states = self.language_model( input_ids, positions, intermediate_tensors, inputs_embeds=inputs_embeds, ) return hidden_states def compute_logits( self, hidden_states: torch.Tensor, ) -> torch.Tensor | None: return self.language_model.compute_logits(hidden_states) def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]: loader = AutoWeightsLoader(self) return loader.load_weights(weights, mapper=self.hf_to_vllm_mapper) def get_mm_mapping(self) -> MultiModelKeys: """ Get the module prefix in multimodal models """ return MultiModelKeys.from_string_field( language_model="language_model.", connector=[ "img_projection", "vision_speech_projection", "speech_projection", ], tower_model=["image_embed", "audio_embed"], ) def get_language_model(self) -> torch.nn.Module: return self.language_model