flumina.py

# flumina.py
import torch
import io
import json
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 logging
import os
import re
import PIL.Image as Image
from typing import Optional, Set, Tuple

from flux_pipeline import FluxPipeline
from util import load_config, ModelVersion

_ASPECT_RATIOS = [
    (1, 1),
    (21, 9),
    (16, 9),
    (3, 2),
    (5, 4),
    (4, 5),
    (2, 3),
    (9, 16),
    (9, 21),
    (4, 3),
    (3, 4),
]

# 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"
        )

    if (w, h) not in _ASPECT_RATIOS:
        raise ValueError(
            f"Invalid aspect ratio: {aspect_ratio}. Aspect ratio must be one of {_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 request and response schemata
class Text2ImageRequest(BaseModel):
    prompt: str
    aspect_ratio: str = "16:9"
    guidance_scale: float = 3.5
    num_inference_steps: int = 30
    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 = False


class FluminaModule(FluminaModule):
    def __init__(self):
        super().__init__()

        # Read configuration from config.json
        with open('config.json', 'r') as f:
            config_data = json.load(f)

        # Now, we need to construct the config and load the model
        if 'config_path' in config_data:
            self.pipeline = FluxPipeline.load_pipeline_from_config_path(
                config_data['config_path'],
                flow_model_path=config_data.get('flow_model_path', None)
            )
        else:
            model_version = (
                ModelVersion.flux_dev
                if config_data.get('model_version', 'flux-dev') == "flux-dev"
                else ModelVersion.flux_schnell
            )
            config = load_config(
                model_version,
                flux_path=config_data.get('flow_model_path', None),
                flux_device=config_data.get('flux_device', 'cuda:0'),
                ae_path=config_data.get('autoencoder_path', None),
                ae_device=config_data.get('autoencoder_device', 'cuda:0'),
                text_enc_path=config_data.get('text_enc_path', None),
                text_enc_device=config_data.get('text_enc_device', 'cuda:0'),
                flow_dtype="float16",
                text_enc_dtype="bfloat16",
                ae_dtype="bfloat16",
                num_to_quant=config_data.get('num_to_quant', 20),
                compile_extras=config_data.get('compile', False),
                compile_blocks=config_data.get('compile', False),
                quant_text_enc=(
                    None
                    if config_data.get('quant_text_enc', 'qfloat8') == "bf16"
                    else config_data.get('quant_text_enc', 'qfloat8')
                ),
                quant_ae=config_data.get('quant_ae', False),
                offload_flow=config_data.get('offload_flow', False),
                offload_ae=config_data.get('offload_ae', True),
                offload_text_enc=config_data.get('offload_text_enc', True),
                prequantized_flow=config_data.get('prequantized_flow', False),
                quantize_modulation=config_data.get('quantize_modulation', True),
                quantize_flow_embedder_layers=config_data.get(
                    'quantize_flow_embedder_layers', False
                ),
            )
            self.pipeline = FluxPipeline.load_pipeline_from_config(config)

        # Initialize LoRA adapters
        self.lora_adapters: Set[str] = set()
        self.active_lora_adapter: Optional[str] = None

        # Warm-up
        self._warm_up()

        # Testing
        self._test_return_sync_response = False


    def _warm_up(self):
        for f, s in _ASPECT_RATIOS:
            print(f"Warm-up for aspect ratio {f}:{s}")
            width, height = _aspect_ratio_to_width_height(f"{f}:{s}")
            self.pipeline.generate(
                prompt="a quick brown fox",
                height=height,
                width=width,
                guidance=3.5,
                num_steps=1,
                seed=0,
            )

    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, image_bytes: bytes, mime_type: str, billing_info: BillingInfo
    ):
        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)
        img_bio = self.pipeline.generate(
            prompt=body.prompt,
            height=height,
            width=width,
            guidance=body.guidance_scale,
            num_steps=body.num_inference_steps,
            seed=body.seed,
        )

        billing_info = BillingInfo(
            steps=body.num_inference_steps,
            height=height,
            width=width,
        )
        return self._image_response(img_bio.getvalue(), mime_type, billing_info)

    @property
    def supported_addon_types(self):
        return ['lora']

    # Addon interface methods adjusted to remove ControlNet support
    def load_addon(
        self,
        addon_account_id: str,
        addon_model_id: str,
        addon_type: str,
        addon_data_path: os.PathLike,
    ):
        if addon_type not in self.supported_addon_types:
            raise ValueError(
                f"Invalid addon type {addon_type}. Supported types: {self.supported_addon_types}"
            )

        qualname = f"accounts/{addon_account_id}/models/{addon_model_id}"

        if addon_type == 'lora':
            self.pipeline.load_lora_weights(addon_data_path, adapter_name=qualname)
            self.lora_adapters.add(qualname)
        else:
            raise NotImplementedError(
                f'Addon support for type {addon_type} not implemented'
            )

    def unload_addon(
        self, addon_account_id: str, addon_model_id: str, addon_type: str
    ):
        qualname = f"accounts/{addon_account_id}/models/{addon_model_id}"

        if addon_type == 'lora':
            assert qualname in self.lora_adapters
            self.pipeline.delete_adapters([qualname])
            self.lora_adapters.remove(qualname)
        else:
            raise NotImplementedError(
                f'Addon support for type {addon_type} not implemented'
            )

    def activate_addon(self, addon_account_id: str, addon_model_id: str):
        qualname = f"accounts/{addon_account_id}/models/{addon_model_id}"

        if qualname in self.lora_adapters:
            if self.active_lora_adapter is not None:
                raise ValueError(
                    f"LoRA adapter {self.active_lora_adapter} already active. Multi-LoRA not yet supported"
                )

            self.active_lora_adapter = qualname
            return

        raise ValueError(f"Unknown addon {qualname}")

    def deactivate_addon(self, addon_account_id: str, addon_model_id: str):
        qualname = f"accounts/{addon_account_id}/models/{addon_model_id}"

        if self.active_lora_adapter == qualname:
            self.active_lora_adapter = None
        else:
            raise AssertionError(f'Addon {qualname} not loaded!')


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

if __name__ == "__main__":
    f = FluminaModule()
    f._test_return_sync_response = True
    import asyncio
    out = asyncio.run(f.text_to_image(
        body=Text2ImageRequest(
            prompt="test",
            aspect_ratio="4:3",
        ),
        accept="*/*"
    ))
    with open("out_image.png", "wb") as f:
        f.write(out)