Skip to content

Qwen3-VL

The Qwen3-VL model is the latest vision–language generation in the Qwen series. It improves text understanding and generation, visual perception and reasoning, spatial and temporal (video) understanding, and capabilities for visual agents, such as GUI understanding and tool use. It supports a long multimodal context (native 256K tokens, extendable up to 1M), expanded OCR (32 languages), and strong performance on multimodal reasoning including STEM-oriented tasks. The family is offered in dense and MoE sizes, with Instruct and reasoning-focused variants for different deployment needs. RBLN NPUs can accelerate Qwen3-VL inference using RBLN Optimum.

API Reference

RBLNQwen3VLVisionModel

Bases: RBLNModel

RBLN optimized Qwen3-VL vision transformer model.

This class provides hardware-accelerated inference for Qwen3-VL vision transformers on RBLN devices, supporting image and video encoding for multimodal vision-language tasks.

from_pretrained(model_id, export=None, rbln_config=None, **kwargs) classmethod

The from_pretrained() function is utilized in its standard form as in the HuggingFace transformers library. User can use this function to load a pre-trained model from the HuggingFace library and convert it to a RBLN model to be run on RBLN NPUs.

Parameters:

Name Type Description Default
model_id str | Path

The model id of the pre-trained model to be loaded. It can be downloaded from the HuggingFace model hub or a local path, or a model id of a compiled model using the RBLN Compiler.

 required
export bool | None

A boolean flag to indicate whether the model should be compiled. If None, it will be determined based on the existence of the compiled model files in the model_id.

 None
rbln_config dict | RBLNModelConfig | None

Configuration for RBLN model compilation and runtime. This can be provided as a dictionary or an instance of the model's configuration class (e.g., RBLNLlamaForCausalLMConfig for Llama models). For detailed configuration options, see the specific model's configuration class documentation.

 None
kwargs Any

Additional keyword arguments. Arguments with the prefix rbln_ are passed to rbln_config, while the remaining arguments are passed to the HuggingFace library.

 {}

Returns:

Type Description
RBLNModel

A RBLN model instance ready for inference on RBLN NPU devices.

get_hf_class() classmethod

Lazily load and cache the HuggingFace class this model wraps.

The class is looked up in hf_library_name under the name without the RBLN prefix (e.g. RBLNLlamaForCausalLM -> LlamaForCausalLM).

save_pretrained(save_directory, push_to_hub=False, **kwargs)

Saves a model and its configuration file to a directory, so that it can be re-loaded using the [~optimum.rbln.modeling_base.RBLNBaseModel.from_pretrained] class method.

Parameters:

Name Type Description Default
save_directory str | Path

Directory where to save the model file.

 required
push_to_hub bool

Whether or not to push your model to the HuggingFace model hub after saving it.

 False

from_model(model, config=None, rbln_config=None, model_save_dir=None, subfolder='', **kwargs) classmethod

Converts and compiles a pre-trained HuggingFace library model into a RBLN model. This method performs the actual model conversion and compilation process.

Parameters:

Name Type Description Default
model PreTrainedModel | ModelMixin

The PyTorch model to be compiled. The object must be a HuggingFace transformers PreTrainedModel or a Diffusers component based on diffusers.ModelMixin, such as a VAE, UNet or ControlNet.

 required
config PretrainedConfig | None

The configuration object associated with the model.

 None
rbln_config RBLNModelConfig | dict | None

Configuration for RBLN model compilation and runtime. This can be provided as a dictionary or an instance of the model's configuration class (e.g., RBLNLlamaForCausalLMConfig for Llama models). For detailed configuration options, see the specific model's configuration class documentation.

 None
kwargs Any

Additional keyword arguments. Arguments with the prefix rbln_ are passed to rbln_config, while the remaining arguments are passed to the HuggingFace library.

 {}

The method performs the following steps:

  1. Compiles the PyTorch model into an optimized RBLN graph
  2. Configures the model for the specified NPU device
  3. Creates the necessary runtime objects if requested
  4. Saves the compiled model and configurations

Returns:

Type Description
RBLNModel

A RBLN model instance ready for inference on RBLN NPU devices.

RBLNQwen3VLForConditionalGeneration

Bases: RBLNQwenVLBatchSortMixin, RBLNQwen3VLModel, RBLNDecoderOnlyModelForCausalLM

RBLNQwen3VLForConditionalGeneration is a multi-modal model that integrates vision and language processing capabilities, optimized for RBLN NPUs. It is designed for conditional generation tasks that involve both image and text inputs.

This model inherits from [RBLNDecoderOnlyModelForCausalLM]. Check the superclass documentation for the generic methods the library implements for all its models.

Important Note

This model includes a Large Language Model (LLM). For optimal performance, it is highly recommended to use tensor parallelism for the language model. This can be achieved by using the rbln_config parameter in the from_pretrained method. Refer to the from_pretrained documentation and the RBLNQwen3VLForConditionalGenerationConfig class for details.

Examples:

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
from optimum.rbln import RBLNQwen3VLForConditionalGeneration

model = RBLNQwen3VLForConditionalGeneration.from_pretrained(
    "Qwen/Qwen3-VL-4B-Instruct",
    export=True,
    rbln_config={
        "visual": {
            "max_seq_len": 6400,
            "device": 0,
        },
        "num_devices": 8,
        "kvcache_partition_len": 16_384,
        "max_seq_len": 262_144,
        "device": [0, 1, 2, 3, 4, 5, 6, 7],
    },
)

model.save_pretrained("compiled-qwen3-vl-4b-instruct")

generate(input_ids, attention_mask=None, generation_config=None, **kwargs)

The generate function is utilized in its standard form as in the HuggingFace transformers library. User can use this function to generate text from the model. Check the HuggingFace transformers documentation for more details.

Parameters:

Name Type Description Default
input_ids LongTensor

The input ids to the model.

 required
attention_mask LongTensor

The attention mask to the model.

 None
generation_config GenerationConfig

The generation configuration to be used as base parametrization for the generation call. **kwargs passed to generate matching the attributes of generation_config will override them. If generation_config is not provided, the default will be used, which had the following loading priority: 1) from the generation_config.json model file, if it exists; 2) from the model configuration. Please note that unspecified parameters will inherit GenerationConfig’s default values.

 None
kwargs dict[str, Any]

Additional arguments passed to the generate function. See the HuggingFace transformers documentation for more details.

 {}

Returns:

Type Description
ModelOutput | LongTensor

A ModelOutput (if return_dict_in_generate=True or when config.return_dict_in_generate=True) or a torch.LongTensor.

rescale_kvcache_num_blocks(compiled_models, rbln_config, target) classmethod

Resize an already-compiled artifact's kv-cache to target blocks.

The saved buffers hold per_block * current bytes, where current is the block count the artifact currently carries (rbln_config.kvcache_num_blocks): the count resolved at compile time, or the target of an earlier rescale. Scaling by the rational ratio target / current yields per_block * target exactly, so exp_rescale_buffer_size never rejects the ratio for a block-size-1 baseline.

from_model(model, config=None, rbln_config=None, model_save_dir=None, subfolder='', **kwargs) classmethod

Converts and compiles a pre-trained HuggingFace library model into a RBLN model. This method performs the actual model conversion and compilation process.

Parameters:

Name Type Description Default
model PreTrainedModel | ModelMixin

The PyTorch model to be compiled. The object must be a HuggingFace transformers PreTrainedModel or a Diffusers component based on diffusers.ModelMixin, such as a VAE, UNet or ControlNet.

 required
config PretrainedConfig | None

The configuration object associated with the model.

 None
rbln_config RBLNModelConfig | dict | None

Configuration for RBLN model compilation and runtime. This can be provided as a dictionary or an instance of the model's configuration class (e.g., RBLNLlamaForCausalLMConfig for Llama models). For detailed configuration options, see the specific model's configuration class documentation.

 None
kwargs Any

Additional keyword arguments. Arguments with the prefix rbln_ are passed to rbln_config, while the remaining arguments are passed to the HuggingFace library.

 {}

The method performs the following steps:

  1. Compiles the PyTorch model into an optimized RBLN graph
  2. Configures the model for the specified NPU device
  3. Creates the necessary runtime objects if requested
  4. Saves the compiled model and configurations

Returns:

Type Description
RBLNModel

A RBLN model instance ready for inference on RBLN NPU devices.

from_pretrained(model_id, export=None, rbln_config=None, **kwargs) classmethod

The from_pretrained() function is utilized in its standard form as in the HuggingFace transformers library. User can use this function to load a pre-trained model from the HuggingFace library and convert it to a RBLN model to be run on RBLN NPUs.

Parameters:

Name Type Description Default
model_id str | Path

The model id of the pre-trained model to be loaded. It can be downloaded from the HuggingFace model hub or a local path, or a model id of a compiled model using the RBLN Compiler.

 required
export bool | None

A boolean flag to indicate whether the model should be compiled. If None, it will be determined based on the existence of the compiled model files in the model_id.

 None
rbln_config dict | RBLNModelConfig | None

Configuration for RBLN model compilation and runtime. This can be provided as a dictionary or an instance of the model's configuration class (e.g., RBLNLlamaForCausalLMConfig for Llama models). For detailed configuration options, see the specific model's configuration class documentation.

 None
kwargs Any

Additional keyword arguments. Arguments with the prefix rbln_ are passed to rbln_config, while the remaining arguments are passed to the HuggingFace library.

 {}

Returns:

Type Description
RBLNModel

A RBLN model instance ready for inference on RBLN NPU devices.

get_hf_class() classmethod

Lazily load and cache the HuggingFace class this model wraps.

The class is looked up in hf_library_name under the name without the RBLN prefix (e.g. RBLNLlamaForCausalLM -> LlamaForCausalLM).

save_pretrained(save_directory, push_to_hub=False, **kwargs)

Saves a model and its configuration file to a directory, so that it can be re-loaded using the [~optimum.rbln.modeling_base.RBLNBaseModel.from_pretrained] class method.

Parameters:

Name Type Description Default
save_directory str | Path

Directory where to save the model file.

 required
push_to_hub bool

Whether or not to push your model to the HuggingFace model hub after saving it.

 False

set_adapter(adapter_name)

Sets the active adapter(s) for the model using adapter name(s).

Parameters:

Name Type Description Default
adapter_name str | list[str]

The name(s) of the adapter(s) to be activated. Can be a single adapter name or a list of adapter names.

 required

Raises:

Type Description
ValueError

If the model is not configured with LoRA or if the adapter name is not found.

RBLNQwen3VLForConditionalGenerationConfig

Bases: RBLNDecoderOnlyModelForCausalLMConfig

Configuration class for RBLNQwen3VLForConditionalGeneration.

This configuration class stores the configuration parameters specific to RBLN-optimized Qwen3-VL models for multimodal conditional generation tasks that combine vision and language processing capabilities.

__init__(use_inputs_embeds=True, visual=None, _load_visual_runtime=True, **kwargs)

Parameters:

Name Type Description Default
use_inputs_embeds bool

Whether or not to use inputs_embeds as input. Defaults to True.

 True
visual RBLNModelConfig | None

Configuration for the vision encoder component.

 None
_load_visual_runtime bool

Whether to create runtime for the visual encoder submodule. Set to False on decoder-only nodes in a disaggregated encoder setup to skip loading the visual encoder's compiled model (.rbln) and torch artifacts entirely. When False, pre-computed image_embeds / video_embeds must be provided to forward(). Defaults to True.

 True
kwargs Any

Additional arguments passed to the parent RBLNDecoderOnlyModelForCausalLMConfig.

 {}

Raises:

Type Description
ValueError

If use_inputs_embeds is False.

RBLNQwen3VLVisionModelConfig

Bases: RBLNModelConfig

Configuration class for RBLNQwen3VLVisionModel.

This configuration class stores the configuration parameters specific to RBLN-optimized Qwen3-VL vision transformer models for processing images and videos.

__init__(max_seq_len=None, batch_size=None, **kwargs)

Parameters:

Name Type Description Default
max_seq_len int | list[int] | None

Maximum sequence lengths for Vision Transformer attention. Can be an integer or list of integers, each indicating the number of patches in a sequence for an image or video. For example, an image of 512x512 pixels with patch size 16 yields (512/16) * (512/16) = 1024 patches; the ViT counts them before the 2x2 spatial merge, on the size after the processor's resize. RBLN optimization runs inference per image, or per temporal patch (2 frames) of a video, so set max_seq_len to match the maximum expected resolution to reduce computation. If not provided, a ValueError is raised.

 None
batch_size int | None

the vision encoder runs one image at a time (the parent config forces this by default), so only batch_size=1 is supported. Defaults to 1.

 None
kwargs Any

Additional arguments passed to the parent RBLNModelConfig.

 {}

Raises:

Type Description
ValueError

If max_seq_len is None or not provided.
ValueError

If batch_size is not 1.

from_pretrained(path, rbln_config=None, return_unused_kwargs=False, **kwargs) classmethod

Load a RBLNModelConfig from a path.

Parameters:

Name Type Description Default
path str

Path to the RBLNModelConfig file or directory containing the config file.

 required
rbln_config dict[str, Any] | None

Additional configuration to override.

 None
return_unused_kwargs bool

Whether to return unused kwargs.

 False
kwargs Any

Additional keyword arguments to override configuration values. Keys starting with 'rbln_' will have the prefix removed and be used to update the configuration.

 {}

Returns:

Name Type Description
RBLNModelConfig Self | tuple[Self, dict[str, Any]]

The loaded configuration instance.
Note

This method loads the configuration from the specified path and applies any provided overrides. If the loaded configuration class doesn't match the expected class, a warning will be logged.

Examples:

1
config = RBLNResNetForImageClassificationConfig.from_pretrained("/path/to/model")

initialize_submodule_config(submodule_config=None, force_kwargs=False, **kwargs)

Build the config of a submodule from the parent's runtime options and kwargs.

Returns the submodule config instance when its class is known (an instance was passed, or cls_name was given), otherwise the merged kwargs dict for the model class to resolve later.

get_load_overrides()

Extract the options that must cross the load boundary, for passing to from_pretrained.

Returns the explicitly-set runtime options (device, device_map, create_runtimes, activate_profiler, timeout) and the subclass-specific non-save attributes (load-behavior flags such as _load_visual_runtime, which are never serialized so they can only travel with the caller), recursively including submodules. Compile-time attributes are not included: when loading a compiled model they come from the artifact's rbln_config.json, and this dict is merged on top of it.

Returns:

Type Description
dict[str, Any]

Dictionary of load overrides, keyed like the rbln_config dict accepted by from_pretrained.