transformers/pr_33962/en/model_doc/phi4_multimodal.md

Phi4 Multimodal

Phi4 Multimodal is a multimodal model capable of text, image, and speech and audio inputs or any combination of these. It features a mixture of LoRA adapters for handling different inputs, and each input is routed to the appropriate encoder.

You can find all the original Phi4 Multimodal checkpoints under the Phi4 collection.

This model was contributed by cyrilvallez.

Click on the Phi-4 Multimodal in the right sidebar for more examples of how to apply Phi-4 Multimodal to different tasks.

The example below demonstrates how to generate text based on an image with Pipeline or the AutoModel class.

from transformers import pipeline
generator = pipeline("text-generation", model="microsoft/Phi-4-multimodal-instruct", dtype="auto", device=0)

prompt = "Explain the concept of multimodal AI in simple terms."

result = generator(prompt, max_length=50)
print(result[0]['generated_text'])

import torch
from transformers import AutoModelForCausalLM, AutoProcessor, GenerationConfig
from accelerate import Accelerator

model_path = "microsoft/Phi-4-multimodal-instruct"
device = Accelerator().device

processor = AutoProcessor.from_pretrained(model_path)
model = AutoModelForCausalLM.from_pretrained(model_path, device_map=device, dtype=torch.float16)

model.load_adapter(model_path, adapter_name="vision", device_map=device, adapter_kwargs={"subfolder": 'vision-lora'})

messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://www.ilankelman.org/stopsigns/australia.jpg"},
            {"type": "text", "text": "What is shown in this image?"},
        ],
    },
]

model.set_adapter("vision")
inputs = processor.apply_chat_template(
    messages,
    add_generation_prompt=True,
    tokenize=True,
    return_dict=True,
    return_tensors="pt",
).to(model.device)

generate_ids = model.generate(
    **inputs,
    max_new_tokens=1000,
    do_sample=False,
)
generate_ids = generate_ids[:, inputs['input_ids'].shape[1]:]
response = processor.batch_decode(
    generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False
)[0]
print(f'>>> Response\n{response}')

Notes

The example below demonstrates inference with an audio and text input.

import torch
from transformers import AutoModelForCausalLM, AutoProcessor, GenerationConfig
from accelerate import Accelerator

model_path = "microsoft/Phi-4-multimodal-instruct"
device = Accelerator().device

processor = AutoProcessor.from_pretrained(model_path)
model = AutoModelForCausalLM.from_pretrained(model_path, device_map=device,  dtype=torch.float16)

model.load_adapter(model_path, adapter_name="speech", device_map=device, adapter_kwargs={"subfolder": 'speech-lora'})
model.set_adapter("speech")
audio_url = "https://upload.wikimedia.org/wikipedia/commons/b/b0/Barbara_Sahakian_BBC_Radio4_The_Life_Scientific_29_May_2012_b01j5j24.flac"
messages = [
    {
        "role": "user",
        "content": [
            {"type": "audio", "url": audio_url},
            {"type": "text", "text": "Transcribe the audio to text, and then translate the audio to French. Use <sep> as a separator between the origina transcript and the translation."},
        ],
    },
]

inputs = processor.apply_chat_template(
    messages,
    add_generation_prompt=True,
    tokenize=True,
    return_dict=True,
    return_tensors="pt",
).to(model.device)

generate_ids = model.generate(
    **inputs,
    max_new_tokens=1000,
    do_sample=False,
)
generate_ids = generate_ids[:, inputs['input_ids'].shape[1]:]
response = processor.batch_decode(
    generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False
)[0]
print(f'>>> Response\n{response}')

Phi4MultimodalFeatureExtractor[[transformers.Phi4MultimodalFeatureExtractor]]

class transformers.Phi4MultimodalFeatureExtractortransformers.Phi4MultimodalFeatureExtractorhttps://github.com/huggingface/transformers/blob/vr_33962/src/transformers/models/phi4_multimodal/feature_extraction_phi4_multimodal.py#L36[{"name": "feature_size", "val": ": int = 80"}, {"name": "sampling_rate", "val": ": int = 16000"}, {"name": "hop_length", "val": ": int = 160"}, {"name": "n_fft", "val": ": int = 512"}, {"name": "win_length", "val": ": int = 400"}, {"name": "preemphasis", "val": ": float = 0.97"}, {"name": "padding_value", "val": ": float = 0.0"}, {"name": "audio_compression_rate", "val": ": int = 8"}, {"name": "audio_downsample_rate", "val": ": int = 1"}, {"name": "audio_feat_stride", "val": ": int = 1"}, {"name": "mel_min_frequency", "val": ": float = 0"}, {"name": "mel_max_frequency", "val": ": float = 7690"}, {"name": "**kwargs", "val": ""}]

Phi4MultimodalImageProcessorFast[[transformers.Phi4MultimodalImageProcessorFast]]

class transformers.Phi4MultimodalImageProcessorFasttransformers.Phi4MultimodalImageProcessorFasthttps://github.com/huggingface/transformers/blob/vr_33962/src/transformers/models/phi4_multimodal/image_processing_phi4_multimodal_fast.py#L51[{"name": "**kwargs", "val": ": typing_extensions.Unpack[transformers.models.phi4_multimodal.image_processing_phi4_multimodal_fast.Phi4MultimodalImageProcessorKwargs]"}]

Constructs a fast Phi4 Multimodal image processor.

pad_to_max_num_cropstransformers.Phi4MultimodalImageProcessorFast.pad_to_max_num_cropshttps://github.com/huggingface/transformers/blob/vr_33962/src/transformers/models/phi4_multimodal/image_processing_phi4_multimodal_fast.py#L140[{"name": "images", "val": ""}, {"name": "max_crops", "val": " = 5"}]

images: B x 3 x H x W, B<=max_crops

preprocesstransformers.Phi4MultimodalImageProcessorFast.preprocesshttps://github.com/huggingface/transformers/blob/vr_33962/src/transformers/models/phi4_multimodal/image_processing_phi4_multimodal_fast.py#L157[{"name": "images", "val": ": typing.Union[ForwardRef('PIL.Image.Image'), numpy.ndarray, ForwardRef('torch.Tensor'), list['PIL.Image.Image'], list[numpy.ndarray], list['torch.Tensor']]"}, {"name": "**kwargs", "val": ": typing_extensions.Unpack[transformers.models.phi4_multimodal.image_processing_phi4_multimodal_fast.Phi4MultimodalImageProcessorKwargs]"}]- images (Union[PIL.Image.Image, numpy.ndarray, torch.Tensor, list['PIL.Image.Image'], list[numpy.ndarray], list['torch.Tensor']]) -- Image to preprocess. Expects a single or batch of images with pixel values ranging from 0 to 255. If passing in images with pixel values between 0 and 1, set do_rescale=False.

• do_convert_rgb (bool, optional) -- Whether to convert the image to RGB.
• do_resize (bool, optional) -- Whether to resize the image.
• size (Annotated[Union[int, list[int], tuple[int, ...], dict[str, int], NoneType], None]) -- Describes the maximum input dimensions to the model.
• crop_size (Annotated[Union[int, list[int], tuple[int, ...], dict[str, int], NoneType], None]) -- Size of the output image after applying center_crop.
• resample (Annotated[Union[PILImageResampling, int, NoneType], None]) -- Resampling filter to use if resizing the image. This can be one of the enum PILImageResampling. Only has an effect if do_resize is set to True.
• do_rescale (bool, optional) -- Whether to rescale the image.
• rescale_factor (float, optional) -- Rescale factor to rescale the image by if do_rescale is set to True.
• do_normalize (bool, optional) -- Whether to normalize the image.
• image_mean (Union[float, list[float], tuple[float, ...], NoneType]) -- Image mean to use for normalization. Only has an effect if do_normalize is set to True.
• image_std (Union[float, list[float], tuple[float, ...], NoneType]) -- Image standard deviation to use for normalization. Only has an effect if do_normalize is set to True.
• do_pad (bool, optional) -- Whether to pad the image. Padding is done either to the largest size in the batch or to a fixed square size per image. The exact padding strategy depends on the model.
• pad_size (Annotated[Union[int, list[int], tuple[int, ...], dict[str, int], NoneType], None]) -- The size in {"height": int, "width" int} to pad the images to. Must be larger than any image size provided for preprocessing. If pad_size is not provided, images will be padded to the largest height and width in the batch. Applied only when do_pad=True.
• do_center_crop (bool, optional) -- Whether to center crop the image.
• data_format (Union[str, ~image_utils.ChannelDimension, NoneType]) -- Only ChannelDimension.FIRST is supported. Added for compatibility with slow processors.
• input_data_format (Union[str, ~image_utils.ChannelDimension, NoneType]) -- The channel dimension format for the input image. If unset, the channel dimension format is inferred from the input image. Can be one of:
  • "channels_first" or ChannelDimension.FIRST: image in (num_channels, height, width) format.
  • "channels_last" or ChannelDimension.LAST: image in (height, width, num_channels) format.
  • "none" or ChannelDimension.NONE: image in (height, width) format.
• device (Annotated[str, None], optional) -- The device to process the images on. If unset, the device is inferred from the input images.
• return_tensors (Annotated[Union[str, ~utils.generic.TensorType, NoneType], None]) -- Returns stacked tensors if set to `pt, otherwise returns a list of tensors.
• disable_grouping (bool, optional) -- Whether to disable grouping of images by size to process them individually and not in batches. If None, will be set to True if the images are on CPU, and False otherwise. This choice is based on empirical observations, as detailed here: https://github.com/huggingface/transformers/pull/38157
• patch_size (int, optional) -- The size of the patch.
• dynamic_hd (int, optional) -- The maximum number of crops per image.0<class 'transformers.image_processing_base.BatchFeature'>- data (dict) -- Dictionary of lists/arrays/tensors returned by the call method ('pixel_values', etc.).
• tensor_type (Union[None, str, TensorType], optional) -- You can give a tensor_type here to convert the lists of integers in PyTorch/Numpy Tensors at initialization.

Phi4MultimodalProcessor[[transformers.Phi4MultimodalProcessor]]

class transformers.Phi4MultimodalProcessortransformers.Phi4MultimodalProcessorhttps://github.com/huggingface/transformers/blob/vr_33962/src/transformers/models/phi4_multimodal/processing_phi4_multimodal.py#L41[{"name": "image_processor", "val": ""}, {"name": "audio_processor", "val": ""}, {"name": "tokenizer", "val": ""}, {"name": "**kwargs", "val": ""}]- image_processor (Phi4MultimodalImageProcessorFast) -- The image processor to use for images.

• audio_processor (Phi4MultimodalFeatureExtractor) -- The audio processor to use for audio inputs.
• tokenizer (GPT2TokenizerFast) -- The tokenizer to use for text.
• fake_image_token_pattern (str, optional, defaults to r"<\|image_\d+\|>") -- The fake image token pattern.
• fake_audio_token_pattern (str, optional, defaults to r"<\|audio_\d+\|>") -- The fake audio token pattern.0

Constructs a Phi4Multimodal processor which raps an image processor, a audio processor, and a GPT tokenizer into a single processor.

Phi4MultimodalProcessor offers all the functionalities of Phi4MultimodalImageProcessorFast and GPT2Tokenizer. See the __call__() and decode() for more information.

Phi4MultimodalAudioConfig[[transformers.Phi4MultimodalAudioConfig]]

class transformers.Phi4MultimodalAudioConfigtransformers.Phi4MultimodalAudioConfighttps://github.com/huggingface/transformers/blob/vr_33962/src/transformers/models/phi4_multimodal/configuration_phi4_multimodal.py#L111[{"name": "hidden_size", "val": ": int = 1024"}, {"name": "intermediate_size", "val": ": int = 1536"}, {"name": "num_blocks", "val": ": int = 24"}, {"name": "num_attention_heads", "val": ": int = 16"}, {"name": "activation", "val": ": str = 'swish'"}, {"name": "chunk_size", "val": ": int = -1"}, {"name": "left_chunk", "val": ": int = 18"}, {"name": "dropout_rate", "val": ": float = 0.0"}, {"name": "ext_pw_out_channel", "val": ": int = 1024"}, {"name": "depthwise_separable_out_channel", "val": ": int = 1024"}, {"name": "depthwise_multiplier", "val": ": int = 1"}, {"name": "kernel_size", "val": ": int = 3"}, {"name": "conv_activation", "val": ": str = 'swish'"}, {"name": "input_size", "val": ": int = 80"}, {"name": "conv_glu_type", "val": ": str = 'swish'"}, {"name": "time_reduction", "val": ": int = 8"}, {"name": "bias_max_distance", "val": ": int = 1000"}, {"name": "bias_symmetric", "val": ": bool = False"}, {"name": "nemo_activation", "val": ": str = 'relu'"}, {"name": "nemo_conv_channels", "val": ": int = 1024"}, {"name": "downsample_rate", "val": ": int = 1"}, {"name": "initializer_range", "val": ": float = 0.02"}, {"name": "audio_token_id", "val": ": int = 200011"}, {"name": "feature_layer", "val": ": int = -2"}, {"name": "**kwargs", "val": ""}]- hidden_size (int, optional, defaults to 1024) -- Dimensionality of the encoder layers.

• intermediate_size (int, optional, defaults to 1536) -- Dimensionality of the "intermediate" (i.e., feed-forward) layer in the Transformer encoder.
• num_blocks (int, optional, defaults to 24) -- Number of hidden layers in the Transformer encoder.
• num_attention_heads (int, optional, defaults to 16) -- Number of attention heads for each attention layer in the Transformer encoder.
• activation (str, optional, defaults to "swish") -- The non-linear activation function in the MLPs.
• chunk_size (int, optional, defaults to -1) -- The chunk size to create the masks.
• left_chunk (int, optional, defaults to 18) -- The left chunk to create the masks.
• dropout_rate (float, optional, defaults to 0.0) -- The dropout ratio.
• ext_pw_out_channel (int, optional, defaults to 1024) -- Number of out channels in the point-wise conv modules.
• depthwise_separable_out_channel (int, optional, defaults to 1024) -- Number of out channels in the depth-wise separable conv modules.
• depthwise_multiplier (int, optional, defaults to 1) -- Input size multiplier for the depth-wise separable conv modules.
• kernel_size (int, optional, defaults to 3) -- Kernel size for the depth-wise separable conv modules.
• conv_activation (str, optional, defaults to "swish") -- The non-linear activation function in the conv modules.
• input_size (int, optional, defaults to 80) -- Input size for the audio model.
• conv_glu_type (str, optional, defaults to "swish") -- The non-linear activation function in the point-wise conv modules.
• time_reduction (int, optional, defaults to 8) -- Time reduction (subsampling factor).
• bias_max_distance (int, optional, defaults to 1000) -- Max distance for the relative attention bias module.
• bias_symmetric (bool, optional, defaults to False) -- Whether the relative attention bias should be symmetric or not.
• nemo_activation (str, optional, defaults to "relu") -- The non-linear activation function in the nemo conv modules.
• nemo_conv_channels (int, optional, defaults to 1024) -- Number of channels in the nemo conv modules.
• downsample_rate (int, optional, defaults to 1) -- Downsample rate for the audio feature extractor.
• initializer_range (float, optional, defaults to 0.02) -- The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
• audio_token_id (int, optional, defaults to 200011) -- The audio token id.
• feature_layer (int, optional, defaults to -2) -- The index of the layer of the encoder from which to extract audio features.0

This is the configuration class to store the configuration of a Phi4MultimodalAudioModel. It is used to instantiate a Phi4Multimodal audio encoder according to the specified arguments, defining the model architecture. Instantiating a configuration with the defaults will yield a similar configuration to that of the audio encoder of microsoft/Phi-4-multimodal-instruct architecture.

Configuration objects inherit from PreTrainedConfig and can be used to control the model outputs. Read the documentation from PreTrainedConfig for more information.

Example:

>>> from transformers import Phi4MultimodalAudioConfig

>>> # Initializing a Phi4MultimodalAudioConfig with microsoft/Phi-4-multimodal-instruct style configuration
>>> configuration = Phi4MultimodalAudioConfig()

Phi4MultimodalVisionConfig[[transformers.Phi4MultimodalVisionConfig]]

class transformers.Phi4MultimodalVisionConfigtransformers.Phi4MultimodalVisionConfighttps://github.com/huggingface/transformers/blob/vr_33962/src/transformers/models/phi4_multimodal/configuration_phi4_multimodal.py#L26[{"name": "hidden_size", "val": " = 1152"}, {"name": "intermediate_size", "val": " = 4304"}, {"name": "num_hidden_layers", "val": " = 27"}, {"name": "num_attention_heads", "val": " = 16"}, {"name": "num_channels", "val": " = 3"}, {"name": "image_size", "val": " = 448"}, {"name": "patch_size", "val": " = 14"}, {"name": "hidden_act", "val": " = 'gelu_pytorch_tanh'"}, {"name": "layer_norm_eps", "val": " = 1e-06"}, {"name": "attention_dropout", "val": " = 0.0"}, {"name": "crop_size", "val": ": int = 448"}, {"name": "image_token_id", "val": ": int = 200010"}, {"name": "feature_layer", "val": ": int = -2"}, {"name": "**kwargs", "val": ""}]- hidden_size (int, optional, defaults to 1152) -- Dimensionality of the encoder layers and the pooler layer.

• intermediate_size (int, optional, defaults to 4304) -- Dimensionality of the "intermediate" (i.e., feed-forward) layer in the Transformer encoder.
• num_hidden_layers (int, optional, defaults to 27) -- Number of hidden layers in the Transformer encoder.
• num_attention_heads (int, optional, defaults to 16) -- Number of attention heads for each attention layer in the Transformer encoder.
• num_channels (int, optional, defaults to 3) -- Number of channels in the input images.
• image_size (int, optional, defaults to 448) -- The size (resolution) of each image.
• patch_size (int, optional, defaults to 14) -- The size (resolution) of each patch.
• hidden_act (str or function, optional, defaults to "gelu_pytorch_tanh") -- The non-linear activation function (function or string) in the encoder and pooler. If string, "gelu", "relu", "selu" and "gelu_new" "quick_gelu" are supported.
• layer_norm_eps (float, optional, defaults to 1e-06) -- The epsilon used by the layer normalization layers.
• attention_dropout (float, optional, defaults to 0.0) -- The dropout ratio for the attention probabilities.
• crop_size (int, optional, defaults to 448) -- Crop size for the input images.
• image_token_id (int, optional, defaults to 200010) -- The image token id.
• feature_layer (int, optional, defaults to -2) -- The index of the layer of the encoder from which to extract image features.0

This is the configuration class to store the configuration of a Phi4MultimodalVisionModel. It is used to instantiate a Phi4Multimodal vision encoder according to the specified arguments, defining the model architecture. Instantiating a configuration with the defaults will yield a similar configuration to that of the vision encoder of microsoft/Phi-4-multimodal-instruct architecture.

Configuration objects inherit from PreTrainedConfig and can be used to control the model outputs. Read the documentation from PreTrainedConfig for more information.

Example:

>>> from transformers import Phi4MultimodalVisionConfig

>>> # Initializing a Phi4MultimodalVisionConfig with microsoft/Phi-4-multimodal-instruct style configuration
>>> configuration = Phi4MultimodalVisionConfig()

Phi4MultimodalConfig[[transformers.Phi4MultimodalConfig]]

class transformers.Phi4MultimodalConfigtransformers.Phi4MultimodalConfighttps://github.com/huggingface/transformers/blob/vr_33962/src/transformers/models/phi4_multimodal/configuration_phi4_multimodal.py#L244[{"name": "vocab_size", "val": " = 200064"}, {"name": "hidden_size", "val": " = 3072"}, {"name": "intermediate_size", "val": " = 8192"}, {"name": "num_hidden_layers", "val": " = 32"}, {"name": "num_attention_heads", "val": " = 32"}, {"name": "num_key_value_heads", "val": " = 8"}, {"name": "resid_pdrop", "val": " = 0.0"}, {"name": "embd_pdrop", "val": " = 0.0"}, {"name": "attention_dropout", "val": " = 0.0"}, {"name": "hidden_act", "val": " = 'silu'"}, {"name": "max_position_embeddings", "val": " = 131072"}, {"name": "initializer_range", "val": " = 0.02"}, {"name": "rms_norm_eps", "val": " = 1e-05"}, {"name": "use_cache", "val": " = True"}, {"name": "tie_word_embeddings", "val": " = False"}, {"name": "rope_theta", "val": " = 10000.0"}, {"name": "rope_scaling", "val": " = None"}, {"name": "partial_rotary_factor", "val": " = 1"}, {"name": "bos_token_id", "val": " = 199999"}, {"name": "eos_token_id", "val": " = [199999, 200020]"}, {"name": "pad_token_id", "val": " = 199999"}, {"name": "original_max_position_embeddings", "val": " = 4096"}, {"name": "sliding_window", "val": " = None"}, {"name": "vision_config", "val": " = None"}, {"name": "audio_config", "val": " = None"}, {"name": "**kwargs", "val": ""}]- vocab_size (int, optional, defaults to 200064) -- Vocabulary size of the Phi-3 model. Defines the number of different tokens that can be represented by the inputs_ids passed when calling Phi3Model.

• hidden_size (int, optional, defaults to 3072) -- Dimension of the hidden representations.
• intermediate_size (int, optional, defaults to 8192) -- Dimension of the MLP representations.
• num_hidden_layers (int, optional, defaults to 32) -- Number of hidden layers in the Transformer decoder.
• num_attention_heads (int, optional, defaults to 32) -- Number of attention heads for each attention layer in the Transformer decoder.
• num_key_value_heads (int, optional, defaults to 8) -- This is the number of key_value heads that should be used to implement Grouped Query Attention. If num_key_value_heads=num_attention_heads, the model will use Multi Head Attention (MHA), if num_key_value_heads=1 the model will use Multi Query Attention (MQA) otherwise GQA is used. When converting a multi-head checkpoint to a GQA checkpoint, each group key and value head should be constructed by meanpooling all the original heads within that group. For more details, check out this paper. If it is not specified, will default to num_attention_heads.
• resid_pdrop (float, optional, defaults to 0.0) -- Dropout probability for mlp outputs.
• embd_pdrop (int, optional, defaults to 0.0) -- The dropout ratio for the embeddings.
• attention_dropout (float, optional, defaults to 0.0) -- The dropout ratio after computing the attention scores.
• hidden_act (str or function, optional, defaults to "silu") -- The non-linear activation function (function or string) in the decoder.
• max_position_embeddings (int, optional, defaults to 131072) -- The maximum sequence length that this model might ever be used with.
• initializer_range (float, optional, defaults to 0.02) -- The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
• rms_norm_eps (float, optional, defaults to 1e-05) -- The epsilon value used for the RMSNorm.
• use_cache (bool, optional, defaults to True) -- Whether or not the model should return the last key/values attentions (not used by all models). Only relevant if config.is_decoder=True. Whether to tie weight embeddings or not.
• tie_word_embeddings (bool, optional, defaults to False) -- Whether to tie weight embeddings
• rope_theta (float, optional, defaults to 10000.0) -- The base period of the RoPE embeddings.
• rope_scaling (dict, optional) -- The scaling strategy for the RoPE embeddings. If None, no scaling is applied. If a dictionary, it must contain the following keys: type, short_factor and long_factor. The type must be longrope and the short_factor and long_factor must be lists of numbers with the same length as the hidden size divided by the number of attention heads divided by 2.
• partial_rotary_factor (float, optional, defaults to 1.0) -- Percentage of the query and keys which will have rotary embedding. Must be between 0.0 and 1.0.
• bos_token_id (int, optional, defaults to 199999) -- The id of the "beginning-of-sequence" token.
• eos_token_id (int or list[int], optional, defaults to [199999, 200020]) -- The id of the "end-of-sequence" token.
• pad_token_id (int, optional, defaults to 199999) -- The id of the padding token.
• original_max_position_embeddings (int, optional, defaults to 4096) -- The maximum sequence length that this model was trained with. This is used to determine the size of the original RoPE embeddings when using long scaling.
• sliding_window (int, optional) -- Sliding window attention window size. If None, no sliding window is applied.
• vision_config (Phi4MultimodalVisionConfig or dict, optional) -- The vision config for the underlying image embedding model. If not provided, will default to the configuration used to instantiate a model similar in architecture as microsoft/Phi-4-multimodal-instruct.
• audio_config (Phi4MultimodalAudioConfig or dict, optional) -- The audio config for the underlying audio embedding model. If not provided, will default to the configuration used to instantiate a model similar in architecture as microsoft/Phi-4-multimodal-instruct.0

This is the configuration class to store the configuration of a Phi4MultimodalModel. It is used to instantiate a Phi4Multimodal model according to the specified arguments, defining the model architecture. Instantiating a configuration with the defaults will yield a similar configuration to that of the microsoft/Phi-4-multimodal-instruct architecture.

Configuration objects inherit from PreTrainedConfig and can be used to control the model outputs. Read the documentation from PreTrainedConfig for more information.

Example:

>>> from transformers import Phi4MultimodalModel, Phi4MultimodalConfig

>>> # Initializing a Phi4Multimodal style configuration
>>> configuration = Phi4MultimodalConfig.from_pretrained("microsoft/Phi-4-multimodal-instruct")

>>> # Initializing a model from the configuration
>>> model = Phi4MultimodalModel(configuration)

>>> # Accessing the model configuration
>>> configuration = model.config

Phi4MultimodalAudioModel[[transformers.Phi4MultimodalAudioModel]]

class transformers.Phi4MultimodalAudioModeltransformers.Phi4MultimodalAudioModelhttps://github.com/huggingface/transformers/blob/vr_33962/src/transformers/models/phi4_multimodal/modeling_phi4_multimodal.py#L998[{"name": "config", "val": ": Phi4MultimodalAudioConfig"}]

forward_embeddingstransformers.Phi4MultimodalAudioModel.forward_embeddingshttps://github.com/huggingface/transformers/blob/vr_33962/src/transformers/models/phi4_multimodal/modeling_phi4_multimodal.py#L1034[{"name": "hidden_states", "val": ""}, {"name": "masks", "val": ""}] Forwarding the inputs through the top embedding layers

Phi4MultimodalVisionModel[[transformers.Phi4MultimodalVisionModel]]

class transformers.Phi4MultimodalVisionModeltransformers.Phi4MultimodalVisionModelhttps://github.com/huggingface/transformers/blob/vr_33962/src/transformers/models/phi4_multimodal/modeling_phi4_multimodal.py#L472[{"name": "config", "val": ": Phi4MultimodalVisionConfig"}]

Phi4MultimodalModel[[transformers.Phi4MultimodalModel]]

class transformers.Phi4MultimodalModeltransformers.Phi4MultimodalModelhttps://github.com/huggingface/transformers/blob/vr_33962/src/transformers/models/phi4_multimodal/modeling_phi4_multimodal.py#L1541[{"name": "config", "val": ": Phi4MultimodalConfig"}]- config (Phi4MultimodalConfig) -- Model configuration class with all the parameters of the model. Initializing with a config file does not load the weights associated with the model, only the configuration. Check out the from_pretrained() method to load the model weights.0

The bare Phi4 Multimodal Model outputting raw hidden-states without any specific head on top.

This model inherits from PreTrainedModel. Check the superclass documentation for the generic methods the library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads etc.)

This model is also a PyTorch torch.nn.Module subclass. Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage and behavior.

forwardtransformers.Phi4MultimodalModel.forwardhttps://github.com/huggingface/transformers/blob/vr_33962/src/transformers/models/phi4_multimodal/modeling_phi4_multimodal.py#L1563[{"name": "input_ids", "val": ": typing.Optional[torch.LongTensor] = None"}, {"name": "attention_mask", "val": ": typing.Optional[torch.Tensor] = None"}, {"name": "position_ids", "val": ": typing.Optional[torch.LongTensor] = None"}, {"name": "past_key_values", "val": ": typing.Optional[transformers.cache_utils.Cache] = None"}, {"name": "inputs_embeds", "val": ": typing.Optional[torch.FloatTensor] = None"}, {"name": "image_pixel_values", "val": ": typing.Optional[torch.FloatTensor] = None"}, {"name": "image_sizes", "val": ": typing.Optional[torch.LongTensor] = None"}, {"name": "image_attention_mask", "val": " = None"}, {"name": "audio_input_features", "val": ": typing.Optional[torch.FloatTensor] = None"}, {"name": "audio_embed_sizes", "val": " = None"}, {"name": "audio_attention_mask", "val": " = None"}, {"name": "use_cache", "val": ": typing.Optional[bool] = None"}, {"name": "output_attentions", "val": ": typing.Optional[bool] = None"}, {"name": "output_hidden_states", "val": ": typing.Optional[bool] = None"}, {"name": "cache_position", "val": ": typing.Optional[torch.LongTensor] = None"}, {"name": "**kwargs", "val": ""}]

image_pixel_values (torch.FloatTensor, optional): If the input contains images, these correspond to the pixel values after transformations (as returned by the Processor) image_sizes (torch.LongTensor, optional): If the input contains images, these correspond to size of each image. image_attention_mask (torch.LongTensor, optional): Attention mask for the images. audio_input_features (torch.FloatTensor, optional): If the input contains audio samples, these correspond to the values after transformation (as returned by the Processor). audio_embed_sizes (torch.Tensor, optional): Size of the audio inputs. audio_attention_mask (`torch.Tensor, optional): Attention mask for the audio inputs.

Phi4MultimodalForCausalLM[[transformers.Phi4MultimodalForCausalLM]]

class transformers.Phi4MultimodalForCausalLMtransformers.Phi4MultimodalForCausalLMhttps://github.com/huggingface/transformers/blob/vr_33962/src/transformers/models/phi4_multimodal/modeling_phi4_multimodal.py#L1659[{"name": "config", "val": ""}]- config (Phi4MultimodalForCausalLM) -- Model configuration class with all the parameters of the model. Initializing with a config file does not load the weights associated with the model, only the configuration. Check out the from_pretrained() method to load the model weights.0

The Phi4 Multimodal Model for causal language modeling.

This model inherits from PreTrainedModel. Check the superclass documentation for the generic methods the library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads etc.)

This model is also a PyTorch torch.nn.Module subclass. Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage and behavior.

forwardtransformers.Phi4MultimodalForCausalLM.forwardhttps://github.com/huggingface/transformers/blob/vr_33962/src/transformers/models/phi4_multimodal/modeling_phi4_multimodal.py#L1673[{"name": "input_ids", "val": ": typing.Optional[torch.LongTensor] = None"}, {"name": "attention_mask", "val": ": typing.Optional[torch.Tensor] = None"}, {"name": "position_ids", "val": ": typing.Optional[torch.LongTensor] = None"}, {"name": "past_key_values", "val": ": typing.Optional[transformers.cache_utils.Cache] = None"}, {"name": "inputs_embeds", "val": ": typing.Optional[torch.FloatTensor] = None"}, {"name": "image_pixel_values", "val": ": typing.Optional[torch.FloatTensor] = None"}, {"name": "image_sizes", "val": ": typing.Optional[torch.LongTensor] = None"}, {"name": "image_attention_mask", "val": " = None"}, {"name": "audio_input_features", "val": ": typing.Optional[torch.FloatTensor] = None"}, {"name": "audio_embed_sizes", "val": " = None"}, {"name": "audio_attention_mask", "val": " = None"}, {"name": "labels", "val": ": typing.Optional[torch.LongTensor] = None"}, {"name": "use_cache", "val": ": typing.Optional[bool] = None"}, {"name": "output_attentions", "val": ": typing.Optional[bool] = None"}, {"name": "output_hidden_states", "val": ": typing.Optional[bool] = None"}, {"name": "cache_position", "val": ": typing.Optional[torch.LongTensor] = None"}, {"name": "logits_to_keep", "val": ": typing.Union[int, torch.Tensor] = 0"}, {"name": "**kwargs", "val": ""}]- input_ids (torch.LongTensor of shape (batch_size, sequence_length), optional) -- Indices of input sequence tokens in the vocabulary. Padding will be ignored by default.

Indices can be obtained using AutoTokenizer. See PreTrainedTokenizer.encode() and PreTrainedTokenizer.call() for details.

What are input IDs?

• attention_mask (torch.Tensor of shape (batch_size, sequence_length), optional) -- Mask to avoid performing attention on padding token indices. Mask values selected in [0, 1]:

  • 1 for tokens that are not masked,
  • 0 for tokens that are masked.

  What are attention masks?

• position_ids (torch.LongTensor of shape (batch_size, sequence_length), optional) -- Indices of positions of each input sequence tokens in the position embeddings. Selected in the range [0, config.n_positions - 1].

  What are position IDs?

• past_key_values (~cache_utils.Cache, optional) -- Pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention blocks) that can be used to speed up sequential decoding. This typically consists in the past_key_values returned by the model at a previous stage of decoding, when use_cache=True or config.use_cache=True.

  Only Cache instance is allowed as input, see our kv cache guide. If no past_key_values are passed, DynamicCache will be initialized by default.

  The model will output the same cache format that is fed as input.

  If past_key_values are used, the user is expected to input only unprocessed input_ids (those that don't have their past key value states given to this model) of shape (batch_size, unprocessed_length) instead of all input_ids of shape (batch_size, sequence_length).

• inputs_embeds (torch.FloatTensor of shape (batch_size, sequence_length, hidden_size), optional) -- 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 vectors than the model's internal embedding lookup matrix.

• image_pixel_values (torch.FloatTensor, optional) -- If the input contains images, these correspond to the pixel values after transformations (as returned by the Processor)

• image_sizes (torch.LongTensor, optional) -- If the input contains images, these correspond to size of each image.

• image_attention_mask (torch.LongTensor, optional) -- Attention mask for the images.

• audio_input_features (torch.FloatTensor, optional) -- If the input contains audio samples, these correspond to the values after transformation (as returned by the Processor).

• audio_embed_sizes (torch.Tensor, optional) -- Size of the audio inputs.

• audio_attention_mask (`torch.Tensor, optional) -- Attention mask for the audio inputs.

• labels (torch.LongTensor of shape (batch_size, sequence_length), optional) -- Labels for computing the masked language modeling loss. Indices should either be in [0, ..., config.vocab_size] or -100 (see input_ids docstring). Tokens with indices set to -100 are ignored (masked), the loss is only computed for the tokens with labels in [0, ..., config.vocab_size].

• use_cache (bool, optional) -- If set to True, past_key_values key value states are returned and can be used to speed up decoding (see past_key_values).

• output_attentions (bool, optional) -- Whether or not to return the attentions tensors of all attention layers. See attentions under returned tensors for more detail.

• output_hidden_states (bool, optional) -- Whether or not to return the hidden states of all layers. See hidden_states under returned tensors for more detail.

• cache_position (torch.LongTensor of shape (sequence_length), optional) -- Indices depicting the position of the input sequence tokens in the sequence. Contrarily to position_ids, this tensor is not affected by padding. It is used to update the cache in the correct position and to infer the complete sequence length.

• logits_to_keep (Union[int, torch.Tensor], defaults to 0) -- If an int, compute logits for the last logits_to_keep tokens. If 0, calculate logits for all input_ids (special case). Only last token logits are needed for generation, and calculating them only for that token can save memory, which becomes pretty significant for long sequences or large vocabulary size. If a torch.Tensor, must be 1D corresponding to the indices to keep in the sequence length dimension. This is useful when using packed tensor format (single dimension for batch and sequence length).0transformers.modeling_outputs.CausalLMOutputWithPast or tuple(torch.FloatTensor)A transformers.modeling_outputs.CausalLMOutputWithPast or a tuple of torch.FloatTensor (if return_dict=False is passed or when config.return_dict=False) comprising various elements depending on the configuration (Phi4MultimodalConfig) and inputs.

• loss (torch.FloatTensor of shape (1,), optional, returned when labels is provided) -- Language modeling loss (for next-token prediction).

• logits (torch.FloatTensor of shape (batch_size, sequence_length, config.vocab_size)) -- Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).

• past_key_values (Cache, optional, returned when use_cache=True is passed or when config.use_cache=True) -- It is a Cache instance. For more details, see our kv cache guide.

  Contains pre-computed hidden-states (key and values in the self-attention blocks) that can be used (see past_key_values input) to speed up sequential decoding.

• hidden_states (tuple(torch.FloatTensor), optional, returned when output_hidden_states=True is passed or when config.output_hidden_states=True) -- Tuple of torch.FloatTensor (one for the output of the embeddings, if the model has an embedding layer, + one for the output of each layer) of shape (batch_size, sequence_length, hidden_size).

  Hidden-states of the model at the output of each layer plus the optional initial embedding outputs.

• attentions (tuple(torch.FloatTensor), optional, returned when output_attentions=True is passed or when config.output_attentions=True) -- Tuple of torch.FloatTensor (one for each layer) of shape (batch_size, num_heads, sequence_length, sequence_length).

  Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads.

The Phi4MultimodalForCausalLM forward method, overrides the __call__ special method.

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the pre and post processing steps while the latter silently ignores them.

Example:

>>> from transformers import AutoTokenizer, Phi4MultimodalForCausalLM
>>> model = Phi4MultimodalForCausalLM.from_pretrained("TBA")
>>> tokenizer = AutoTokenizer.from_pretrained("TBA")
>>> prompt = "This is an example script ."
>>> inputs = tokenizer(prompt, return_tensors="pt")
>>> # Generate
>>> generate_ids = model.generate(inputs.input_ids, max_length=30)
>>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
'This is an example script .\n Certainly! Below is a sample script that demonstrates a simple task, such as calculating the sum'