|
| 1 | +from typing import Any |
| 2 | + |
| 3 | +import torch |
| 4 | +from PIL.Image import Image |
| 5 | +from pydantic import field_validator |
| 6 | + |
| 7 | +from invokeai.app.invocations.baseinvocation import BaseInvocation, invocation |
| 8 | +from invokeai.app.invocations.fields import FieldDescriptions, ImageField, InputField, UIComponent, UIType |
| 9 | +from invokeai.app.invocations.model import ModelIdentifierField |
| 10 | +from invokeai.app.invocations.primitives import StringOutput |
| 11 | +from invokeai.app.services.shared.invocation_context import InvocationContext |
| 12 | +from invokeai.backend.llava_onevision_model import LlavaOnevisionModel |
| 13 | +from invokeai.backend.util.devices import TorchDevice |
| 14 | + |
| 15 | + |
| 16 | +@invocation("llava_onevision_vllm", title="LLaVA OneVision VLLM", tags=["vllm"], category="vllm", version="1.0.0") |
| 17 | +class LlavaOnevisionVllmInvocation(BaseInvocation): |
| 18 | + """Run a LLaVA OneVision VLLM model.""" |
| 19 | + |
| 20 | + images: list[ImageField] | ImageField | None = InputField(default=None, max_length=3, description="Input image.") |
| 21 | + prompt: str = InputField( |
| 22 | + default="", |
| 23 | + description="Input text prompt.", |
| 24 | + ui_component=UIComponent.Textarea, |
| 25 | + ) |
| 26 | + vllm_model: ModelIdentifierField = InputField( |
| 27 | + title="LLaVA Model Type", |
| 28 | + description=FieldDescriptions.vllm_model, |
| 29 | + ui_type=UIType.LlavaOnevisionModel, |
| 30 | + ) |
| 31 | + |
| 32 | + @field_validator("images", mode="before") |
| 33 | + def listify_images(cls, v: Any) -> list: |
| 34 | + if v is None: |
| 35 | + return v |
| 36 | + if not isinstance(v, list): |
| 37 | + return [v] |
| 38 | + return v |
| 39 | + |
| 40 | + def _get_images(self, context: InvocationContext) -> list[Image]: |
| 41 | + if self.images is None: |
| 42 | + return [] |
| 43 | + |
| 44 | + image_fields = self.images if isinstance(self.images, list) else [self.images] |
| 45 | + return [context.images.get_pil(image_field.image_name, "RGB") for image_field in image_fields] |
| 46 | + |
| 47 | + @torch.no_grad() |
| 48 | + def invoke(self, context: InvocationContext) -> StringOutput: |
| 49 | + images = self._get_images(context) |
| 50 | + |
| 51 | + with context.models.load(self.vllm_model) as vllm_model: |
| 52 | + assert isinstance(vllm_model, LlavaOnevisionModel) |
| 53 | + output = vllm_model.run( |
| 54 | + prompt=self.prompt, |
| 55 | + images=images, |
| 56 | + device=TorchDevice.choose_torch_device(), |
| 57 | + dtype=TorchDevice.choose_torch_dtype(), |
| 58 | + ) |
| 59 | + |
| 60 | + return StringOutput(value=output) |
0 commit comments