Transformers documentation
Gemma 3
Gemma 3
Gemma 3 is a multimodal model with pretrained and instruction-tuned variants, available in 1B, 13B, and 27B parameters. The architecture is mostly the same as the previous Gemma versions. The key differences are alternating 5 local sliding window self-attention layers for every global self-attention layer, support for a longer context length of 128K tokens, and a SigLip encoder that can βpan & scanβ high-resolution images to prevent information from disappearing in high resolution images or images with non-square aspect ratios.
The instruction-tuned variant was post-trained with knowledge distillation and reinforcement learning.
You can find all the original Gemma 3 checkpoints under the Gemma 3 release.
Click on the Gemma 3 models in the right sidebar for more examples of how to apply Gemma to different vision and language tasks.
The example below demonstrates how to generate text based on an image with Pipeline or the AutoModel class.
import torch
from transformers import pipeline
pipeline = pipeline(
task="image-text-to-text",
model="google/gemma-3-4b-pt",
device=0,
torch_dtype=torch.bfloat16
)
pipeline(
"https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/pipeline-cat-chonk.jpeg",
text="<start_of_image> What is shown in this image?"
)
Quantization reduces the memory burden of large models by representing the weights in a lower precision. Refer to the Quantization overview for more available quantization backends.
The example below uses torchao to only quantize the weights to int4.
# pip install torchao
import torch
from transformers import TorchAoConfig, Gemma3ForConditionalGeneration, AutoProcessor
quantization_config = TorchAoConfig("int4_weight_only", group_size=128)
model = Gemma3ForConditionalGeneration.from_pretrained(
"google/gemma-3-27b-it",
torch_dtype=torch.bfloat16,
device_map="auto",
quantization_config=quantization_config
)
processor = AutoProcessor.from_pretrained(
"google/gemma-3-27b-it",
padding_side="left"
)
messages = [
{
"role": "system",
"content": [
{"type": "text", "text": "You are a helpful assistant."}
]
},
{
"role": "user", "content": [
{"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/pipeline-cat-chonk.jpeg"},
{"type": "text", "text": "What is shown in this image?"},
]
},
]
inputs = processor.apply_chat_template(
messages,
tokenize=True,
return_dict=True,
return_tensors="pt",
add_generation_prompt=True,
).to("cuda")
output = model.generate(**inputs, max_new_tokens=50, cache_implementation="static")
print(processor.decode(output[0], skip_special_tokens=True))
Use the AttentionMaskVisualizer to better understand what tokens the model can and cannot attend to.
from transformers.utils.attention_visualizer import AttentionMaskVisualizer
visualizer = AttentionMaskVisualizer("google/gemma-3-4b-it")
visualizer("<img>What is shown in this image?")

Notes
Use Gemma3ForConditionalGeneration for image-and-text and image-only inputs.
Gemma 3 supports multiple input images, but make sure the images are correctly batched before passing them to the processor. Each batch should be a list of one or more images.
url_cow = "https://media.istockphoto.com/id/1192867753/photo/cow-in-berchida-beach-siniscola.jpg?s=612x612&w=0&k=20&c=v0hjjniwsMNfJSuKWZuIn8pssmD5h5bSN1peBd1CmH4=" url_cat = "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/pipeline-cat-chonk.jpeg" messages =[ { "role": "system", "content": [ {"type": "text", "text": "You are a helpful assistant."} ] }, { "role": "user", "content": [ {"type": "image", "url": url_cow}, {"type": "image", "url": url_cat}, {"type": "text", "text": "Which image is cuter?"}, ] }, ]
Text passed to the processor should have a
<start_of_image>
token wherever an image should be inserted.The processor has its own apply_chat_template() method to convert chat messages to model inputs.
By default, images arenβt cropped and only the base image is forwarded to the model. In high resolution images or images with non-square aspect ratios, artifacts can result because the vision encoder uses a fixed resolution of 896x896. To prevent these artifacts and improve performance during inference, set
do_pan_and_scan=True
to crop the image into multiple smaller patches and concatenate them with the base image embedding. You can disable pan and scan for faster inference.inputs = processor.apply_chat_template( messages, tokenize=True, return_dict=True, return_tensors="pt", add_generation_prompt=True, + do_pan_and_scan=True, ).to("cuda")
For Gemma-3 1B checkpoint trained in text-only mode, use AutoModelForCausalLM instead.
import torch from transformers import AutoModelForCausalLM, AutoTokenizer tokenizer = AutoTokenizer.from_pretrained( "google/gemma-3-1b-pt", ) model = AutoModelForCausalLM.from_pretrained( "google/gemma-3-1b-pt", torch_dtype=torch.bfloat16, device_map="auto", attn_implementation="sdpa" ) input_ids = tokenizer("Plants create energy through a process known as", return_tensors="pt").to("cuda") output = model.generate(**input_ids, cache_implementation="static") print(tokenizer.decode(output[0], skip_special_tokens=True))
Gemma3ImageProcessor
class transformers.Gemma3ImageProcessor
< source >( do_resize: bool = True size: typing.Dict[str, int] = None resample: Resampling = <Resampling.BILINEAR: 2> do_rescale: bool = True rescale_factor: typing.Union[int, float] = 0.00392156862745098 do_normalize: bool = True image_mean: typing.Union[float, typing.List[float], NoneType] = None image_std: typing.Union[float, typing.List[float], NoneType] = None do_convert_rgb: typing.Optional[bool] = None do_pan_and_scan: typing.Optional[bool] = None pan_and_scan_min_crop_size: typing.Optional[int] = None pan_and_scan_max_num_crops: typing.Optional[int] = None pan_and_scan_min_ratio_to_activate: typing.Optional[float] = None **kwargs )
Parameters
- do_resize (
bool
, optional, defaults toTrue
) β Whether to resize the imageβs (height, width) dimensions to the specifiedsize
. Can be overridden bydo_resize
in thepreprocess
method. - size (
Dict[str, int]
optional, defaults to{"height" -- 224, "width": 224}
): Size of the image after resizing. Can be overridden bysize
in thepreprocess
method. - resample (
PILImageResampling
, optional, defaults toResampling.BILINEAR
) β Resampling filter to use if resizing the image. Can be overridden byresample
in thepreprocess
method. - do_rescale (
bool
, optional, defaults toTrue
) β Whether to rescale the image by the specified scalerescale_factor
. Can be overridden bydo_rescale
in thepreprocess
method. - rescale_factor (
int
orfloat
, optional, defaults to1/255
) β Scale factor to use if rescaling the image. Can be overridden byrescale_factor
in thepreprocess
method. - do_normalize (
bool
, optional, defaults toTrue
) β Whether to normalize the image by the specified mean and standard deviation. Can be overridden bydo_normalize
in thepreprocess
method. - image_mean (
float
orList[float]
, optional, defaults to[0.5, 0.5, 0.5]
) β Mean to use if normalizing the image. This is a float or list of floats the length of the number of channels in the image. Can be overridden by theimage_mean
parameter in thepreprocess
method. - image_std (
float
orList[float]
, optional, defaults to[0.5, 0.5, 0.5]
) β Standard deviation to use if normalizing the image. This is a float or list of floats the length of the number of channels in the image. Can be overridden by theimage_std
parameter in thepreprocess
method. Can be overridden by theimage_std
parameter in thepreprocess
method. - do_convert_rgb (
bool
, optional, defaults toTrue
) β Whether to convert the image to RGB. - do_pan_and_scan (
bool
, optional) β Whether to applypan_and_scan
to images. - pan_and_scan_min_crop_size (
int
, optional) β Minimum size of each crop in pan and scan. - pan_and_scan_max_num_crops (
int
, optional) β Maximum number of crops per image in pan and scan. - pan_and_scan_min_ratio_to_activate (
float
, optional) β Minimum aspect ratio to activate pan and scan.
Constructs a SigLIP image processor.
pan_and_scan
< source >( image: ndarray pan_and_scan_min_crop_size: int pan_and_scan_max_num_crops: int pan_and_scan_min_ratio_to_activate: float data_format: typing.Union[str, transformers.image_utils.ChannelDimension, NoneType] = None input_data_format: typing.Union[str, transformers.image_utils.ChannelDimension, NoneType] = None )
Parameters
- image (
np.ndarray
) β Image to resize. - pan_and_scan_min_crop_size (
int
, optional) β Minimum size of each crop in pan and scan. - pan_and_scan_max_num_crops (
int
, optional) β Maximum number of crops per image in pan and scan. - pan_and_scan_min_ratio_to_activate (
float
, optional) β Minimum aspect ratio to activate pan and scan. - data_format (
str
orChannelDimension
, optional) β The channel dimension format of the image. If not provided, it will be the same as the input image. - input_data_format (
ChannelDimension
orstr
, optional) β The channel dimension format of the input image. If not provided, it will be inferred.
Pan and Scan and image, by cropping into smaller images when the aspect ratio exceeds minumum allowed ratio.
preprocess
< source >( images: typing.Union[ForwardRef('PIL.Image.Image'), numpy.ndarray, ForwardRef('torch.Tensor'), list['PIL.Image.Image'], list[numpy.ndarray], list['torch.Tensor']] do_resize: typing.Optional[bool] = None size: typing.Dict[str, int] = None resample: Resampling = None do_rescale: typing.Optional[bool] = None rescale_factor: typing.Optional[float] = None do_normalize: typing.Optional[bool] = None image_mean: typing.Union[float, typing.List[float], NoneType] = None image_std: typing.Union[float, typing.List[float], NoneType] = None return_tensors: typing.Union[str, transformers.utils.generic.TensorType, NoneType] = None data_format: typing.Optional[transformers.image_utils.ChannelDimension] = <ChannelDimension.FIRST: 'channels_first'> input_data_format: typing.Union[str, transformers.image_utils.ChannelDimension, NoneType] = None do_convert_rgb: typing.Optional[bool] = None do_pan_and_scan: typing.Optional[bool] = None pan_and_scan_min_crop_size: typing.Optional[int] = None pan_and_scan_max_num_crops: typing.Optional[int] = None pan_and_scan_min_ratio_to_activate: typing.Optional[float] = None )
Parameters
- images (
ImageInput
) β 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, setdo_rescale=False
. - do_resize (
bool
, optional, defaults toself.do_resize
) β Whether to resize the image. - size (
Dict[str, int]
, optional, defaults toself.size
) β Size of the image after resizing. - resample (
int
, optional, defaults toself.resample
) β Resampling filter to use if resizing the image. This can be one of the enumPILImageResampling
. Only has an effect ifdo_resize
is set toTrue
. - do_rescale (
bool
, optional, defaults toself.do_rescale
) β Whether to rescale the image. - rescale_factor (
float
, optional, defaults toself.rescale_factor
) β Rescale factor to rescale the image by ifdo_rescale
is set toTrue
. - do_normalize (
bool
, optional, defaults toself.do_normalize
) β Whether to normalize the image. - image_mean (
float
orList[float]
, optional, defaults toself.image_mean
) β Image mean to use for normalization. Only has an effect ifdo_normalize
is set toTrue
. - image_std (
float
orList[float]
, optional, defaults toself.image_std
) β Image standard deviation to use for normalization. Only has an effect ifdo_normalize
is set toTrue
. - return_tensors (
str
orTensorType
, optional) β The type of tensors to return. Can be one of:- Unset: Return a list of
np.ndarray
. TensorType.TENSORFLOW
or'tf'
: Return a batch of typetf.Tensor
.TensorType.PYTORCH
or'pt'
: Return a batch of typetorch.Tensor
.TensorType.NUMPY
or'np'
: Return a batch of typenp.ndarray
.TensorType.JAX
or'jax'
: Return a batch of typejax.numpy.ndarray
.
- Unset: Return a list of
- data_format (
ChannelDimension
orstr
, optional, defaults toChannelDimension.FIRST
) β The channel dimension format for the output image. Can be one of:"channels_first"
orChannelDimension.FIRST
: image in (num_channels, height, width) format."channels_last"
orChannelDimension.LAST
: image in (height, width, num_channels) format.- Unset: Use the channel dimension format of the input image.
- input_data_format (
ChannelDimension
orstr
, optional) β 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"
orChannelDimension.FIRST
: image in (num_channels, height, width) format."channels_last"
orChannelDimension.LAST
: image in (height, width, num_channels) format."none"
orChannelDimension.NONE
: image in (height, width) format.
- do_convert_rgb (
bool
, optional, defaults toself.do_convert_rgb
) β Whether to convert the image to RGB. - do_pan_and_scan (
bool
, optional, defaults toself.do_convert_rgb
) β Whether to applypan_and_scan
to images. - pan_and_scan_min_crop_size (
int
, optional, defaults toself.pan_and_scan_min_crop_size
) β Minimum size of each crop in pan and scan. - pan_and_scan_max_num_crops (
int
, optional, defaults toself.pan_and_scan_max_num_crops
) β Maximum number of crops per image in pan and scan. - pan_and_scan_min_ratio_to_activate (
float
, optional, defaults toself.pan_and_scan_min_ratio_to_activate
) β Minimum aspect ratio to activate pan and scan.
Preprocess an image or batch of images.
Gemma3ImageProcessorFast
class transformers.Gemma3ImageProcessorFast
< source >( **kwargs: typing_extensions.Unpack[transformers.models.gemma3.image_processing_gemma3_fast.Gemma3FastImageProcessorKwargs] )
Parameters
- do_resize (
bool
, optional, defaults toself.do_resize
) β Whether to resize the imageβs (height, width) dimensions to the specifiedsize
. Can be overridden by thedo_resize
parameter in thepreprocess
method. - size (
dict
, optional, defaults toself.size
) β Size of the output image after resizing. Can be overridden by thesize
parameter in thepreprocess
method. - default_to_square (
bool
, optional, defaults toself.default_to_square
) β Whether to default to a square image when resizing, if size is an int. - resample (
PILImageResampling
, optional, defaults toself.resample
) β Resampling filter to use if resizing the image. Only has an effect ifdo_resize
is set toTrue
. Can be overridden by theresample
parameter in thepreprocess
method. - do_center_crop (
bool
, optional, defaults toself.do_center_crop
) β Whether to center crop the image to the specifiedcrop_size
. Can be overridden bydo_center_crop
in thepreprocess
method. - crop_size (
Dict[str, int]
optional, defaults toself.crop_size
) β Size of the output image after applyingcenter_crop
. Can be overridden bycrop_size
in thepreprocess
method. - do_rescale (
bool
, optional, defaults toself.do_rescale
) β Whether to rescale the image by the specified scalerescale_factor
. Can be overridden by thedo_rescale
parameter in thepreprocess
method. - rescale_factor (
int
orfloat
, optional, defaults toself.rescale_factor
) β Scale factor to use if rescaling the image. Only has an effect ifdo_rescale
is set toTrue
. Can be overridden by therescale_factor
parameter in thepreprocess
method. - do_normalize (
bool
, optional, defaults toself.do_normalize
) β Whether to normalize the image. Can be overridden by thedo_normalize
parameter in thepreprocess
method. Can be overridden by thedo_normalize
parameter in thepreprocess
method. - image_mean (
float
orList[float]
, optional, defaults toself.image_mean
) β Mean to use if normalizing the image. This is a float or list of floats the length of the number of channels in the image. Can be overridden by theimage_mean
parameter in thepreprocess
method. Can be overridden by theimage_mean
parameter in thepreprocess
method. - image_std (
float
orList[float]
, optional, defaults toself.image_std
) β Standard deviation to use if normalizing the image. This is a float or list of floats the length of the number of channels in the image. Can be overridden by theimage_std
parameter in thepreprocess
method. Can be overridden by theimage_std
parameter in thepreprocess
method. - do_convert_rgb (
bool
, optional, defaults toself.do_convert_rgb
) β Whether to convert the image to RGB. - return_tensors (
str
orTensorType
, optional, defaults toself.return_tensors
) β Returns stacked tensors if set to `pt, otherwise returns a list of tensors. - data_format (
ChannelDimension
orstr
, optional, defaults toself.data_format
) β OnlyChannelDimension.FIRST
is supported. Added for compatibility with slow processors. - input_data_format (
ChannelDimension
orstr
, optional, defaults toself.input_data_format
) β 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"
orChannelDimension.FIRST
: image in (num_channels, height, width) format."channels_last"
orChannelDimension.LAST
: image in (height, width, num_channels) format."none"
orChannelDimension.NONE
: image in (height, width) format.
- device (
torch.device
, optional, defaults toself.device
) β The device to process the images on. If unset, the device is inferred from the input images. - do_pan_and_scan (
bool
, optional) β Whether to applypan_and_scan
to images. - pan_and_scan_min_crop_size (
int
, optional) β Minimum size of each crop in pan and scan. - pan_and_scan_max_num_crops (
int
, optional) β Maximum number of crops per image in pan and scan. - pan_and_scan_min_ratio_to_activate (
float
, optional) β Minimum aspect ratio to activate pan and scan.
Constructs a fast ConvNeXT image processor. Based on SiglipImageProcessor with incorporation of Pan adn Scan cropping method.
pan_and_scan_batched
< source >( images: torch.Tensor pan_and_scan_min_crop_size: int pan_and_scan_max_num_crops: int pan_and_scan_min_ratio_to_activate: float )
Parameters
- image (
torch.Tensor
) β Image to resize. - pan_and_scan_min_crop_size (
int
, optional) β Minimum size of each crop in pan and scan. - pan_and_scan_max_num_crops (
int
, optional) β Maximum number of crops per image in pan and scan. - pan_and_scan_min_ratio_to_activate (
float
, optional) β Minimum aspect ratio to activate pan and scan.
Pan and Scan an image, by cropping into smaller images when the aspect ratio exceeds minumum allowed ratio.
preprocess
< source >( images: typing.Union[ForwardRef('PIL.Image.Image'), numpy.ndarray, ForwardRef('torch.Tensor'), list['PIL.Image.Image'], list[numpy.ndarray], list['torch.Tensor']] **kwargs: typing_extensions.Unpack[transformers.models.gemma3.image_processing_gemma3_fast.Gemma3FastImageProcessorKwargs] )
Parameters
- images (
ImageInput
) β 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, setdo_rescale=False
. - do_resize (
bool
, optional, defaults toself.do_resize
) β Whether to resize the image. - size (
Dict[str, int]
, optional, defaults toself.size
) β Describes the maximum input dimensions to the model. - resample (
PILImageResampling
orInterpolationMode
, optional, defaults toself.resample
) β Resampling filter to use if resizing the image. This can be one of the enumPILImageResampling
. Only has an effect ifdo_resize
is set toTrue
. - do_center_crop (
bool
, optional, defaults toself.do_center_crop
) β Whether to center crop the image. - crop_size (
Dict[str, int]
, optional, defaults toself.crop_size
) β Size of the output image after applyingcenter_crop
. - do_rescale (
bool
, optional, defaults toself.do_rescale
) β Whether to rescale the image. - rescale_factor (
float
, optional, defaults toself.rescale_factor
) β Rescale factor to rescale the image by ifdo_rescale
is set toTrue
. - do_normalize (
bool
, optional, defaults toself.do_normalize
) β Whether to normalize the image. - image_mean (
float
orList[float]
, optional, defaults toself.image_mean
) β Image mean to use for normalization. Only has an effect ifdo_normalize
is set toTrue
. - image_std (
float
orList[float]
, optional, defaults toself.image_std
) β Image standard deviation to use for normalization. Only has an effect ifdo_normalize
is set toTrue
. - do_convert_rgb (
bool
, optional, defaults toself.do_convert_rgb
) β Whether to convert the image to RGB. - return_tensors (
str
orTensorType
, optional, defaults toself.return_tensors
) β Returns stacked tensors if set to `pt, otherwise returns a list of tensors. - data_format (
ChannelDimension
orstr
, optional, defaults toself.data_format
) β OnlyChannelDimension.FIRST
is supported. Added for compatibility with slow processors. - input_data_format (
ChannelDimension
orstr
, optional, defaults toself.input_data_format
) β 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"
orChannelDimension.FIRST
: image in (num_channels, height, width) format."channels_last"
orChannelDimension.LAST
: image in (height, width, num_channels) format."none"
orChannelDimension.NONE
: image in (height, width) format.
- device (
torch.device
, optional, defaults toself.device
) β The device to process the images on. If unset, the device is inferred from the input images. do_pan_and_scan (bool
, optional): Whether to applypan_and_scan
to images. pan_and_scan_min_crop_size (int
, optional): Minimum size of each crop in pan and scan. pan_and_scan_max_num_crops (int
, optional): Maximum number of crops per image in pan and scan. pan_and_scan_min_ratio_to_activate (float
, optional): Minimum aspect ratio to activate pan and scan.
Preprocess an image or batch of images.
Gemma3Processor
class transformers.Gemma3Processor
< source >( image_processor tokenizer chat_template = None image_seq_length: int = 256 **kwargs )
This method forwards all its arguments to GemmaTokenizerFastβs batch_decode(). Please refer to the docstring of this method for more information.
This method forwards all its arguments to GemmaTokenizerFastβs decode(). Please refer to the docstring of this method for more information.
Gemma3TextConfig
class transformers.Gemma3TextConfig
< source >( vocab_size = 262208 hidden_size = 2304 intermediate_size = 9216 num_hidden_layers = 26 num_attention_heads = 8 num_key_value_heads = 4 head_dim = 256 hidden_activation = 'gelu_pytorch_tanh' max_position_embeddings = 131072 initializer_range = 0.02 rms_norm_eps = 1e-06 use_cache = True pad_token_id = 0 eos_token_id = 1 bos_token_id = 2 tie_word_embeddings = True rope_theta = 1000000.0 attention_bias = False attention_dropout = 0.0 query_pre_attn_scalar = 256 sliding_window = 4096 final_logit_softcapping = None attn_logit_softcapping = None cache_implementation = 'hybrid' rope_scaling = None rope_local_base_freq = 10000.0 sliding_window_pattern = 6 **kwargs )
Parameters
- vocab_size (
int
, optional, defaults to 262208) β Vocabulary size of the Gemma3Text model. Defines the number of different tokens that can be represented by theinputs_ids
passed when calling Gemma3TextModel - hidden_size (
int
, optional, defaults to 2304) β Dimension of the hidden representations. - intermediate_size (
int
, optional, defaults to 9216) β Dimension of the MLP representations. - num_hidden_layers (
int
, optional, defaults to 26) β Number of hidden layers in the Transformer decoder. - num_attention_heads (
int
, optional, defaults to 8) β Number of attention heads for each attention layer in the Transformer decoder. - num_key_value_heads (
int
, optional, defaults to 4) β This is the number of key_value heads that should be used to implement Grouped Query Attention. Ifnum_key_value_heads=num_attention_heads
, the model will use Multi Head Attention (MHA), ifnum_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 checkout this paper. If it is not specified, will default tonum_attention_heads
. - head_dim (
int
, optional, defaults to 256) β The attention head dimension. - hidden_activation (
str
orfunction
, optional, defaults to"gelu_pytorch_tanh"
) β The non-linear activation function (function or string) in the decoder. Will default to"gelu_pytorch_tanh"
if not specified."gelu_pytorch_tanh"
uses an approximation of the"gelu"
activation function. - max_position_embeddings (
int
, optional, defaults to 131072) β The maximum sequence length that this model might ever be used with. - initializer_range (
float
, optional, defaults to 0.02) β The standard deviation of the truncated_normal_initializer for initializing all weight matrices. - rms_norm_eps (
float
, optional, defaults to 1e-06) β The epsilon used by the rms normalization layers. - use_cache (
bool
, optional, defaults toTrue
) β Whether or not the model should return the last key/values attentions (not used by all models). Only relevant ifconfig.is_decoder=True
. - pad_token_id (
int
, optional, defaults to 0) β Padding token id. - eos_token_id (
int
, optional, defaults to 1) β End of stream token id. - bos_token_id (
int
, optional, defaults to 2) β Beginning of stream token id. - tie_word_embeddings (
bool
, optional, defaults toTrue
) β Whether to tie weight embeddings - rope_theta (
float
, optional, defaults to 1000000.0) β The base period of the RoPE embeddings. - attention_bias (
bool
, defaults toFalse
, optional, defaults toFalse
) β Whether to use a bias in the query, key, value and output projection layers during self-attention. - attention_dropout (
float
, optional, defaults to 0.0) β The dropout ratio for the attention probabilities. - query_pre_attn_scalar (
float
, optional, defaults to 256) β Scaling factor used on the attention scores - sliding_window (
int
, optional, defaults to 4096) β in Gemma3Text, every other layer uses sliding window attention. This is the size of the sliding window. - final_logit_softcapping (
float
, optional) β Scaling factor when applying tanh softcapping on the logits. - attn_logit_softcapping (
float
, optional) β Scaling factor when applying tanh softcapping on the attention scores. - cache_implementation (
str
, optional, defaults to"hybrid"
) β the cache type to be used withgenerate
. - rope_scaling (
Dict
, optional) β Dictionary containing the scaling configuration for the RoPE embeddings used in gloabl attention. NOTE: if you apply new rope type and you expect the model to work on longermax_position_embeddings
, we recommend you to update this value accordingly. Expected contents:rope_type
(str
): The sub-variant of RoPE to use. Can be one of [βdefaultβ, βlinearβ, βdynamicβ, βyarnβ, βlongropeβ, βllama3β], with βdefaultβ being the original RoPE implementation.factor
(float
, optional): Used with all rope types except βdefaultβ. The scaling factor to apply to the RoPE embeddings. In most scaling types, afactor
of x will enable the model to handle sequences of length x original maximum pre-trained length.original_max_position_embeddings
(int
, optional): Used with βdynamicβ, βlongropeβ and βllama3β. The original max position embeddings used during pretraining.attention_factor
(float
, optional): Used with βyarnβ and βlongropeβ. The scaling factor to be applied on the attention computation. If unspecified, it defaults to value recommended by the implementation, using thefactor
field to infer the suggested value.beta_fast
(float
, optional): Only used with βyarnβ. Parameter to set the boundary for extrapolation (only) in the linear ramp function. If unspecified, it defaults to 32.beta_slow
(float
, optional): Only used with βyarnβ. Parameter to set the boundary for interpolation (only) in the linear ramp function. If unspecified, it defaults to 1.short_factor
(List[float]
, optional): Only used with βlongropeβ. The scaling factor to be applied to short contexts (<original_max_position_embeddings
). Must be a list of numbers with the same length as the hidden size divided by the number of attention heads divided by 2long_factor
(List[float]
, optional): Only used with βlongropeβ. The scaling factor to be applied to long contexts (<original_max_position_embeddings
). Must be a list of numbers with the same length as the hidden size divided by the number of attention heads divided by 2low_freq_factor
(float
, optional): Only used with βllama3β. Scaling factor applied to low frequency components of the RoPEhigh_freq_factor
(float
, optional*): Only used with βllama3β. Scaling factor applied to high frequency components of the RoPE - rope_local_base_freq (float, optional, defaults to 10000.0) β The base period of the RoPE embeddings for local attention.
- sliding_window_pattern (
int
, optional, defaults to 6) β Pattern for the sliding window attention.
This is the configuration class to store the configuration of a Gemma3TextModel. It is used to instantiate an Gemma3Text 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 Gemma3Text-7B. e.g. google/gemma3_text-7b Configuration objects inherit from PretrainedConfig and can be used to control the model outputs. Read the documentation from PretrainedConfig for more information.
>>> from transformers import Gemma3TextModel, Gemma3TextConfig
>>> # Initializing a Gemma3Text gemma3_text-7b style configuration
>>> configuration = Gemma3TextConfig()
>>> # Initializing a model from the gemma3_text-7b style configuration
>>> model = Gemma3TextModel(configuration)
>>> # Accessing the model configuration
>>> configuration = model.config
Gemma3Config
class transformers.Gemma3Config
< source >( text_config: typing.Union[transformers.models.gemma3.configuration_gemma3.Gemma3TextConfig, typing.Dict[str, typing.Any], NoneType] = None vision_config: typing.Union[transformers.models.siglip.configuration_siglip.SiglipVisionConfig, typing.Dict[str, typing.Any], NoneType] = None mm_tokens_per_image: int = 256 boi_token_index: int = 255999 eoi_token_index: int = 256000 image_token_index: int = 262144 initializer_range: float = 0.02 **kwargs )
Parameters
- text_config (
Union[Gemma3TextConfig, dict]
, optional) β The config object of the text backbone. - vision_config (
Union[AutoConfig, dict]
, optional) β Custom vision config or dict. - mm_tokens_per_image (
int
, optional, defaults to 256) β The number of tokens per image embedding. - boi_token_index (
int
, optional, defaults to 255999) β The begin-of-image token index to wrap the image prompt. - eoi_token_index (
int
, optional, defaults to 256000) β The end-of-image token index to wrap the image prompt. - image_token_index (
int
, optional, defaults to 262144) β The image token index to encode the image prompt. - initializer_range (
float
, optional, defaults to 0.02) β The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
This is the configuration class to store the configuration of a Gemma3ForConditionalGeneration. It is used to instantiate an Gemma3ForConditionalGeneration according to the specified arguments, defining the model architecture. Instantiating a configuration with the defaults will yield a similar configuration to that of the PaliGemma-2B.
e.g. google/gemma-3-4b
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 Gemma3ForConditionalGeneration, Gemma3Config, SiglipVisionConfig, Gemma3TextConfig
>>> # Initializing a Siglip-like vision config
>>> vision_config = SiglipVisionConfig()
>>> # Initializing a Gemma3 Text config
>>> text_config = Gemma3TextConfig()
>>> # Initializing a Gemma3 gemma-3-4b style configuration
>>> configuration = Gemma3Config(vision_config, text_config)
>>> # Initializing a model from the gemma-3-4b style configuration
>>> model = Gemma3TextConfig(configuration)
>>> # Accessing the model configuration
>>> configuration = model.config
Gemma3TextModel
class transformers.Gemma3TextModel
< source >( config: Gemma3TextConfig )
Parameters
- config (Gemma3Config) β 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.
- config β Gemma3TextConfig
The bare Gemma3Text 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.
Transformer decoder consisting of config.num_hidden_layers layers. Each layer is a Gemma3TextDecoderLayer
forward
< source >( input_ids: typing.Optional[torch.LongTensor] = None attention_mask: typing.Optional[torch.Tensor] = None position_ids: typing.Optional[torch.LongTensor] = None past_key_values: typing.Optional[transformers.cache_utils.HybridCache] = None inputs_embeds: typing.Optional[torch.FloatTensor] = None use_cache: typing.Optional[bool] = None output_attentions: typing.Optional[bool] = None output_hidden_states: typing.Optional[bool] = None cache_position: typing.Optional[torch.LongTensor] = None last_cache_position: typing.Optional[int] = None **flash_attn_kwargs: typing_extensions.Unpack[transformers.modeling_flash_attention_utils.FlashAttentionKwargs] )
Parameters
- input_ids (
torch.LongTensor
of shape(batch_size, sequence_length)
) β Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide it.Indices can be obtained using AutoTokenizer. See PreTrainedTokenizer.encode() and PreTrainedTokenizer.call() for details.
- attention_mask (
torch.Tensor
of shape(batch_size, sequence_length)
, optional) β 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.
Indices can be obtained using AutoTokenizer. See PreTrainedTokenizer.encode() and PreTrainedTokenizer.call() for details.
If
past_key_values
is used, optionally only the lastinput_ids
have to be input (seepast_key_values
).If you want to change padding behavior, you should read
modeling_opt._prepare_decoder_attention_mask
and modify to your needs. See diagram 1 in the paper for more information on the default strategy.- 1 indicates the head is not masked,
- 0 indicates the head is masked.
- position_ids (
torch.LongTensor
of shape(batch_size, sequence_length)
, optional) β Indices of positions of each input sequence tokens in the position embeddings. Selected in the range[0, config.n_positions - 1]
. - past_key_values (
Cache
, optional) β 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 thepast_key_values
returned by the model at a previous stage of decoding, whenuse_cache=True
orconfig.use_cache=True
.It is a Cache instance. For more details, see our kv cache guide.
If
past_key_values
are used, the user can optionally input only the lastinput_ids
(those that donβt have their past key value states given to this model) of shape(batch_size, 1)
instead of allinput_ids
of shape(batch_size, sequence_length)
. - inputs_embeds (
torch.FloatTensor
of shape(batch_size, sequence_length, hidden_size)
, optional) β Optionally, instead of passinginput_ids
you can choose to directly pass an embedded representation. This is useful if you want more control over how to convertinput_ids
indices into associated vectors than the modelβs internal embedding lookup matrix. - use_cache (
bool
, optional) β If set toTrue
,past_key_values
key value states are returned and can be used to speed up decoding (seepast_key_values
). - output_attentions (
bool
, optional) β Whether or not to return the attentions tensors of all attention layers. Seeattentions
under returned tensors for more detail. - output_hidden_states (
bool
, optional) β Whether or not to return the hidden states of all layers. Seehidden_states
under returned tensors for more detail. - return_dict (
bool
, optional) β Whether or not to return a ModelOutput instead of a plain tuple. - cache_position (
torch.LongTensor
of shape(sequence_length)
, optional) β Indices depicting the position of the input sequence tokens in the sequence. Contrarily toposition_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.
The Gemma3TextModel 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.
Gemma3ForCausalLM
forward
< source >( input_ids: typing.Optional[torch.LongTensor] = None attention_mask: typing.Optional[torch.Tensor] = None position_ids: typing.Optional[torch.LongTensor] = None past_key_values: typing.Optional[transformers.cache_utils.HybridCache] = None inputs_embeds: typing.Optional[torch.FloatTensor] = None labels: typing.Optional[torch.LongTensor] = None use_cache: typing.Optional[bool] = None output_attentions: typing.Optional[bool] = None output_hidden_states: typing.Optional[bool] = None cache_position: typing.Optional[torch.LongTensor] = None logits_to_keep: typing.Union[int, torch.Tensor] = 0 **loss_kwargs ) β transformers.modeling_outputs.CausalLMOutputWithPast or tuple(torch.FloatTensor)
Parameters
- input_ids (
torch.LongTensor
of shape(batch_size, sequence_length)
) β Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide it.Indices can be obtained using AutoTokenizer. See PreTrainedTokenizer.encode() and PreTrainedTokenizer.call() for details.
- attention_mask (
torch.Tensor
of shape(batch_size, sequence_length)
, optional) β 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.
Indices can be obtained using AutoTokenizer. See PreTrainedTokenizer.encode() and PreTrainedTokenizer.call() for details.
If
past_key_values
is used, optionally only the lastinput_ids
have to be input (seepast_key_values
).If you want to change padding behavior, you should read
modeling_opt._prepare_decoder_attention_mask
and modify to your needs. See diagram 1 in the paper for more information on the default strategy.- 1 indicates the head is not masked,
- 0 indicates the head is masked.
- position_ids (
torch.LongTensor
of shape(batch_size, sequence_length)
, optional) β Indices of positions of each input sequence tokens in the position embeddings. Selected in the range[0, config.n_positions - 1]
. - past_key_values (
Cache
, optional) β 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 thepast_key_values
returned by the model at a previous stage of decoding, whenuse_cache=True
orconfig.use_cache=True
.It is a Cache instance. For more details, see our kv cache guide.
If
past_key_values
are used, the user can optionally input only the lastinput_ids
(those that donβt have their past key value states given to this model) of shape(batch_size, 1)
instead of allinput_ids
of shape(batch_size, sequence_length)
. - inputs_embeds (
torch.FloatTensor
of shape(batch_size, sequence_length, hidden_size)
, optional) β Optionally, instead of passinginput_ids
you can choose to directly pass an embedded representation. This is useful if you want more control over how to convertinput_ids
indices into associated vectors than the modelβs internal embedding lookup matrix. - use_cache (
bool
, optional) β If set toTrue
,past_key_values
key value states are returned and can be used to speed up decoding (seepast_key_values
). - output_attentions (
bool
, optional) β Whether or not to return the attentions tensors of all attention layers. Seeattentions
under returned tensors for more detail. - output_hidden_states (
bool
, optional) β Whether or not to return the hidden states of all layers. Seehidden_states
under returned tensors for more detail. - return_dict (
bool
, optional) β Whether or not to return a ModelOutput instead of a plain tuple. - cache_position (
torch.LongTensor
of shape(sequence_length)
, optional) β Indices depicting the position of the input sequence tokens in the sequence. Contrarily toposition_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. - labels (
torch.LongTensor
of shape(batch_size, sequence_length)
, optional) β Labels for computing the masked language modeling loss. Indices should either be in[0, ..., config.vocab_size]
or -100 (seeinput_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]
. - logits_to_keep (
int
ortorch.Tensor
, optional) β If anint
, compute logits for the lastlogits_to_keep
tokens. If0
, calculate logits for allinput_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 atorch.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).
Returns
transformers.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 (Gemma3Config) and inputs.
-
loss (
torch.FloatTensor
of shape(1,)
, optional, returned whenlabels
is provided) β Language modeling loss (for next-token prediction). -
logits (
torch.FloatTensor
of shape(batch_size, sequence_length, config.vocab_size)
) β Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax). -
past_key_values (
tuple(tuple(torch.FloatTensor))
, optional, returned whenuse_cache=True
is passed or whenconfig.use_cache=True
) β Tuple oftuple(torch.FloatTensor)
of lengthconfig.n_layers
, with each tuple having 2 tensors of shape(batch_size, num_heads, sequence_length, embed_size_per_head)
)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 whenoutput_hidden_states=True
is passed or whenconfig.output_hidden_states=True
) β Tuple oftorch.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 whenoutput_attentions=True
is passed or whenconfig.output_attentions=True
) β Tuple oftorch.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 Gemma3ForCausalLM 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, Gemma3ForCausalLM
>>> model = Gemma3ForCausalLM.from_pretrained("google/gemma-2-9b")
>>> tokenizer = AutoTokenizer.from_pretrained("google/gemma-2-9b")
>>> prompt = "What is your favorite condiment?"
>>> 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]
"What is your favorite condiment?"
Gemma3ForConditionalGeneration
class transformers.Gemma3ForConditionalGeneration
< source >( config: Gemma3Config )
Parameters
- config (Gemma3Config) β 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.
The GEMMA3 model which consists of a vision backbone and a language model. 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.
forward
< source >( input_ids: typing.Optional[torch.LongTensor] = None pixel_values: typing.Optional[torch.FloatTensor] = None attention_mask: typing.Optional[torch.Tensor] = None position_ids: typing.Optional[torch.LongTensor] = None past_key_values: typing.Union[typing.List[torch.FloatTensor], transformers.cache_utils.Cache, NoneType] = None token_type_ids: typing.Optional[torch.LongTensor] = None cache_position: typing.Optional[torch.LongTensor] = None inputs_embeds: typing.Optional[torch.FloatTensor] = None labels: typing.Optional[torch.LongTensor] = None use_cache: typing.Optional[bool] = None output_attentions: typing.Optional[bool] = None output_hidden_states: typing.Optional[bool] = None logits_to_keep: typing.Union[int, torch.Tensor] = 0 **lm_kwargs ) β transformers.models.gemma3.modeling_gemma3.Gemma3CausalLMOutputWithPast
or tuple(torch.FloatTensor)
Parameters
- input_ids (
torch.LongTensor
of shape(batch_size, sequence_length)
) β Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide it.Indices can be obtained using AutoTokenizer. See PreTrainedTokenizer.encode() and PreTrainedTokenizer.call() for details.
- attention_mask (
torch.Tensor
of shape(batch_size, sequence_length)
, optional) β 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.
Indices can be obtained using AutoTokenizer. See PreTrainedTokenizer.encode() and PreTrainedTokenizer.call() for details.
If
past_key_values
is used, optionally only the lastinput_ids
have to be input (seepast_key_values
).If you want to change padding behavior, you should read
modeling_opt._prepare_decoder_attention_mask
and modify to your needs. See diagram 1 in the paper for more information on the default strategy.- 1 indicates the head is not masked,
- 0 indicates the head is masked.
- position_ids (
torch.LongTensor
of shape(batch_size, sequence_length)
, optional) β Indices of positions of each input sequence tokens in the position embeddings. Selected in the range[0, config.n_positions - 1]
. - past_key_values (
Cache
, optional) β 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 thepast_key_values
returned by the model at a previous stage of decoding, whenuse_cache=True
orconfig.use_cache=True
.It is a Cache instance. For more details, see our kv cache guide.
If
past_key_values
are used, the user can optionally input only the lastinput_ids
(those that donβt have their past key value states given to this model) of shape(batch_size, 1)
instead of allinput_ids
of shape(batch_size, sequence_length)
. - inputs_embeds (
torch.FloatTensor
of shape(batch_size, sequence_length, hidden_size)
, optional) β Optionally, instead of passinginput_ids
you can choose to directly pass an embedded representation. This is useful if you want more control over how to convertinput_ids
indices into associated vectors than the modelβs internal embedding lookup matrix. - use_cache (
bool
, optional) β If set toTrue
,past_key_values
key value states are returned and can be used to speed up decoding (seepast_key_values
). - output_attentions (
bool
, optional) β Whether or not to return the attentions tensors of all attention layers. Seeattentions
under returned tensors for more detail. - output_hidden_states (
bool
, optional) β Whether or not to return the hidden states of all layers. Seehidden_states
under returned tensors for more detail. - return_dict (
bool
, optional) β Whether or not to return a ModelOutput instead of a plain tuple. - cache_position (
torch.LongTensor
of shape(sequence_length)
, optional) β Indices depicting the position of the input sequence tokens in the sequence. Contrarily toposition_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. - labels (
torch.LongTensor
of shape(batch_size, sequence_length)
, optional) β Labels for computing the masked language modeling loss. Indices should either be in[0, ..., config.text_config.vocab_size]
or -100 (seeinput_ids
docstring). Tokens with indices set to-100
are ignored (masked), the loss is only computed for the tokens with labels in[0, ..., config.text_config.vocab_size]
. - logits_to_keep (
int
ortorch.Tensor
, optional) β If anint
, compute logits for the lastlogits_to_keep
tokens. If0
, calculate logits for allinput_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 atorch.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).
Returns
transformers.models.gemma3.modeling_gemma3.Gemma3CausalLMOutputWithPast
or tuple(torch.FloatTensor)
A transformers.models.gemma3.modeling_gemma3.Gemma3CausalLMOutputWithPast
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 (Gemma3Config) and inputs.
-
loss (
torch.FloatTensor
of shape(1,)
, optional, returned whenlabels
is provided) β Language modeling loss (for next-token prediction). -
logits (
torch.FloatTensor
of shape(batch_size, sequence_length, config.text_config.vocab_size)
) β Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax). -
past_key_values (
tuple(tuple(torch.FloatTensor))
, optional, returned whenuse_cache=True
is passed or whenconfig.use_cache=True
) β Tuple oftuple(torch.FloatTensor)
of lengthconfig.n_layers
, with each tuple having 2 tensors of shape(batch_size, num_heads, sequence_length, embed_size_per_head)
)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 whenoutput_hidden_states=True
is passed or whenconfig.output_hidden_states=True
) β Tuple oftorch.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 whenoutput_attentions=True
is passed or whenconfig.output_attentions=True
) β Tuple oftorch.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.
-
image_hidden_states (
torch.FloatTensor
, optional) β Atorch.FloatTensor
of size(batch_size, sequence_length, hidden_size)
. image_hidden_states of the model produced by the vision encoder after projecting last hidden state.
The Gemma3ForConditionalGeneration 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 PIL import Image
>>> import requests
>>> from transformers import AutoProcessor, Gemma3ForConditionalGeneration
>>> model = Gemma3ForConditionalGeneration.from_pretrained("google/gemma-3-4b-it")
>>> processor = AutoProcessor.from_pretrained("google/gemma-3-4b-it")
>>> messages = [
... {
... "role": "system",
... "content": [
... {"type": "text", "text": "You are a helpful assistant."}
... ]
... },
... {
... "role": "user", "content": [
... {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/pipeline-cat-chonk.jpeg"},
... {"type": "text", "text": "Where is the cat standing?"},
... ]
... },
... ]
>>> inputs = processor.apply_chat_template(
... messages,
... tokenizer=True,
... return_dict=True,
... return_tensors="pt",
... add_generation_prompt=True
... )
>>> # Generate
>>> generate_ids = model.generate(**inputs)
>>> processor.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
"user\nYou are a helpful assistant.\n\n\n\n\n\nWhere is the cat standing?\nmodel\nBased on the image, the cat is standing in a snowy area, likely outdoors. It appears to"