flumina.py

import torch
import io
from fireworks.flumina import FluminaModule, main as flumina_main
from fireworks.flumina.route import post
import pydantic
from pydantic import BaseModel
from fastapi import  Header
from fastapi.responses import Response
import math
import re
import PIL.Image as Image
from typing import  Tuple
from tqdm import tqdm

from sd3_infer import SD3Inferencer, CONFIGS
from sd3_impls import  SD3LatentFormat


# Util
def _aspect_ratio_to_width_height(aspect_ratio: str) -> Tuple[int, int]:
    """
    Convert specified aspect ratio to a height/width pair.
    """
    if ":" not in aspect_ratio:
        raise ValueError(
            f"Invalid aspect ratio: {aspect_ratio}. Aspect ratio must be in w:h format, e.g. 16:9"
        )

    w, h = aspect_ratio.split(":")
    try:
        w, h = int(w), int(h)
    except ValueError:
        raise ValueError(
            f"Invalid aspect ratio: {aspect_ratio}. Aspect ratio must be in w:h format, e.g. 16:9"
        )

    valid_aspect_ratios = [
        (1, 1),
        (21, 9),
        (16, 9),
        (3, 2),
        (5, 4),
        (4, 5),
        (2, 3),
        (9, 16),
        (9, 21),
    ]
    if (w, h) not in valid_aspect_ratios:
        raise ValueError(
            f"Invalid aspect ratio: {aspect_ratio}. Aspect ratio must be one of {valid_aspect_ratios}"
        )

    # We consider megapixel not 10^6 pixels but 2^20 (1024x1024) pixels
    TARGET_SIZE_MP = 1
    target_size = TARGET_SIZE_MP * 2**20

    width = math.sqrt(target_size / (w * h)) * w
    height = math.sqrt(target_size / (w * h)) * h

    PAD_MULTIPLE = 64

    if PAD_MULTIPLE:
        width = width // PAD_MULTIPLE * PAD_MULTIPLE
        height = height // PAD_MULTIPLE * PAD_MULTIPLE

    return int(width), int(height)


def encode_image(
    image: Image.Image, mime_type: str, jpeg_quality: int = 95
) -> bytes:
    buffered = io.BytesIO()
    if mime_type == "image/jpeg":
        if jpeg_quality < 0 or jpeg_quality > 100:
            raise ValueError(
                f"jpeg_quality must be between 0 and 100, not {jpeg_quality}"
            )
        image.save(buffered, format="JPEG", quality=jpeg_quality)
    elif mime_type == "image/png":
        image.save(buffered, format="PNG")
    else:
        raise ValueError(f"invalid mime_type {mime_type}")
    return buffered.getvalue()


def parse_accept_header(accept: str) -> str:
    # Split the string into the comma-separated components
    parts = accept.split(",")
    weighted_types = []

    for part in parts:
        # Use a regular expression to extract the media type and the optional q-factor
        match = re.match(
            r"(?P<media_type>[^;]+)(;q=(?P<q_factor>\d+(\.\d+)?))?", part.strip()
        )
        if match:
            media_type = match.group("media_type").strip()
            q_factor = (
                float(match.group("q_factor")) if match.group("q_factor") else 1.0
            )
            weighted_types.append((media_type, q_factor))
        else:
            raise ValueError(f"Malformed Accept header value: {part.strip()}")

    # Sort the media types by q-factor, descending
    sorted_types = sorted(weighted_types, key=lambda x: x[1], reverse=True)

    # Define a list of supported MIME types
    supported_types = ["image/jpeg", "image/png"]

    for media_type, _ in sorted_types:
        if media_type in supported_types:
            return media_type
        elif media_type == "*/*":
            return supported_types[0]  # Default to the first supported type
        elif media_type == "image/*":
            # If "image/*" is specified, return the first matching supported image type
            return supported_types[0]

    raise ValueError(f"Accept header did not include any supported MIME types: {supported_types}")


# Define your request and response schemata here
class Text2ImageRequest(BaseModel):
        prompt: str
        aspect_ratio: str = "16:9"
        guidance_scale: float = 4.5
        num_inference_steps: int = 28
        seed: int = 0


class Error(BaseModel):
    object: str = "error"
    type: str = "invalid_request_error"
    message: str


class ErrorResponse(BaseModel):
    error: Error = pydantic.Field(default_factory=Error)


class BillingInfo(BaseModel):
    steps: int
    height: int
    width: int
    is_control_net: bool


MODEL = "models/sd3.5_medium.safetensors"
VERBOSE = True


class SD3InferencerInMemoryOutput(SD3Inferencer):
    def gen_image(
        self,
        prompts,
        width,
        height,
        steps,
        cfg_scale,
        sampler,
        seed,
        seed_type,
        init_image,
        denoise,
    ):
        latent = self.get_empty_latent(width, height)
        if init_image:
            image_data = Image.open(init_image)
            image_data = image_data.resize((width, height), Image.LANCZOS)
            latent = self.vae_encode(image_data)
            latent = SD3LatentFormat().process_in(latent)
        neg_cond = self.get_cond("")
        seed_num = None
        assert len(prompts) == 1
        pbar = tqdm(enumerate(prompts), total=len(prompts), position=0, leave=True)
        for i, prompt in pbar:
            if seed_type == "roll":
                seed_num = seed if seed_num is None else seed_num + 1
            elif seed_type == "rand":
                seed_num = torch.randint(0, 100000, (1,)).item()
            else:  # fixed
                seed_num = seed
            conditioning = self.get_cond(prompt)
            sampled_latent = self.do_sampling(
                latent,
                seed_num,
                conditioning,
                neg_cond,
                steps,
                cfg_scale,
                sampler,
                denoise if init_image else 1.0,
            )
            return self.vae_decode(sampled_latent)


class FluminaModule(FluminaModule):
    def __init__(self):
        super().__init__()
        self.inferencer = SD3InferencerInMemoryOutput()
        with torch.inference_mode():
            self.inferencer.load(model=MODEL, vae=MODEL, shift=CONFIGS["sd3.5_medium"]["shift"], verbose=VERBOSE)
            self.inferencer.clip_l.model.to("cuda")
            self.inferencer.clip_g.model.to("cuda")
            self.inferencer.t5xxl.model.to("cuda")
            self.inferencer.sd3.model.to("cuda")
            self.inferencer.vae.model.to("cuda")

        self._test_return_sync_response = False

    def _error_response(self, code: int, message: str) -> Response:
        response_json = ErrorResponse(
            error=Error(message=message),
        ).json()
        if self._test_return_sync_response:
            return response_json
        else:
            return Response(
                response_json,
                status_code=code,
                media_type="application/json",
            )

    def _image_response(self, img: Image.Image, mime_type: str, billing_info: BillingInfo):
        image_bytes = encode_image(img, mime_type)
        if self._test_return_sync_response:
            return image_bytes
        else:
            headers = {'Fireworks-Billing-Properties': billing_info.json()}
            return Response(image_bytes, status_code=200, media_type=mime_type, headers=headers)

    @post('/text_to_image')
    async def text_to_image(
        self,
        body: Text2ImageRequest,
        accept: str = Header("image/jpeg"),
     ):
        mime_type = parse_accept_header(accept)
        width, height = _aspect_ratio_to_width_height(body.aspect_ratio)
        with torch.inference_mode():
            img = self.inferencer.gen_image(
                prompts=[body.prompt],
                width=width,
                height=height,
                steps=body.num_inference_steps,
                cfg_scale=body.guidance_scale,
                sampler=CONFIGS['sd3.5_medium']['sampler'],
                seed=body.seed,
                seed_type="roll",
                init_image=None,
                denoise=0.0, # N/A with None init_image
            )

        billing_info = BillingInfo(
            steps=body.num_inference_steps,
            height=height,
            width=width,
            is_control_net=False,
        )
        return self._image_response(img, mime_type, billing_info)

    @property
    def supported_addon_types(self):
        return []


if __name__ == "__flumina_main__":
    f = FluminaModule()
    flumina_main(f)

if __name__ == "__main__":
    f = FluminaModule()
    f._test_return_sync_response = True

    import asyncio

    # Test text-to-image
    t2i_out = asyncio.run(f.text_to_image(
        Text2ImageRequest(
            prompt="A quick brown fox",
            aspect_ratio="16:9",
            guidance_scale=3.5,
            num_inference_steps=30,
            seed=0,
        ),
        accept="image/jpeg",
    ))
    assert isinstance(t2i_out, bytes), t2i_out
    with open('output.png', 'wb') as out_file:
        out_file.write(t2i_out)