app.py

import gradio as gr
import torch
from torch import Tensor, nn
import spaces
import numpy as np
import io
import base64
from flax import nnx
import jax.numpy as jnp
from jax import Array as Tensor

from transformers import (FlaxCLIPTextModel, CLIPTokenizer, FlaxT5EncoderModel,
                          T5Tokenizer)

models = {}
class HFEmbedder(nnx.Module):
    def __init__(self, version: str, max_length: int, **hf_kwargs):
        self.is_clip = version.startswith("openai")
        self.max_length = max_length
        self.output_key = "pooler_output" if self.is_clip else "last_hidden_state"
        dtype = hf_kwargs.get("dtype", jnp.float32)
        if self.is_clip:
            self.tokenizer: CLIPTokenizer = CLIPTokenizer.from_pretrained(version, max_length=max_length)
            # self.hf_module: CLIPTextModel = CLIPTextModel.from_pretrained(version, **hf_kwargs)
            self.hf_module, params = FlaxCLIPTextModel.from_pretrained(version, _do_init=False, **hf_kwargs)
        else:
            self.tokenizer: T5Tokenizer = T5Tokenizer.from_pretrained(version, max_length=max_length)
            # self.hf_module: T5EncoderModel = T5EncoderModel.from_pretrained(version, **hf_kwargs)
            self.hf_module, params = FlaxT5EncoderModel.from_pretrained(version, _do_init=False,**hf_kwargs)
        self.hf_module._is_initialized = True
        import jax
        self.hf_module.params = jax.tree.map(lambda x: jax.device_put(x, jax.devices("cuda")[0]), params)
        # if dtype==jnp.bfloat16:

    def tokenize(self, text: list[str]) -> Tensor:
        batch_encoding = self.tokenizer(
            text,
            truncation=True,
            max_length=self.max_length,
            return_length=False,
            return_overflowing_tokens=False,
            padding="max_length",
            return_tensors="jax",
        )
        return batch_encoding["input_ids"]
    
    def __call__(self, input_ids: Tensor) -> Tensor:
        # outputs = self.hf_module(
        #     input_ids=batch_encoding["input_ids"].to(self.hf_module.device),
        #     attention_mask=None,
        #     output_hidden_states=False,
        # )
        outputs = self.hf_module(
            input_ids=input_ids,
            attention_mask=None,
            output_hidden_states=False,
            train=False,
        )
        return outputs[self.output_key]
    # def __call__(self, text: list[str]) -> Tensor:
    #     batch_encoding = self.tokenizer(
    #         text,
    #         truncation=True,
    #         max_length=self.max_length,
    #         return_length=False,
    #         return_overflowing_tokens=False,
    #         padding="max_length",
    #         return_tensors="jax",
    #     )

    #     # outputs = self.hf_module(
    #     #     input_ids=batch_encoding["input_ids"].to(self.hf_module.device),
    #     #     attention_mask=None,
    #     #     output_hidden_states=False,
    #     # )
    #     outputs = self.hf_module(
    #         input_ids=batch_encoding["input_ids"],
    #         attention_mask=None,
    #         output_hidden_states=False,
    #         train=False,
    #     )
    #     return outputs[self.output_key]




def load_t5(device: str | torch.device = "cuda", max_length: int = 512) -> HFEmbedder:
    # max length 64, 128, 256 and 512 should work (if your sequence is short enough)
    return HFEmbedder("lnyan/t5-v1_1-xxl-encoder", max_length=max_length, dtype=jnp.bfloat16)


def load_clip(device: str | torch.device = "cuda") -> HFEmbedder:
    return HFEmbedder("openai/clip-vit-large-patch14", max_length=77, dtype=jnp.bfloat16)

@spaces.GPU(duration=60)
def load_encoders():
    is_schnell = True
    t5 = load_t5("cuda", max_length=256 if is_schnell else 512)
    clip = load_clip("cuda")
    return t5, clip

import numpy as np
def b64(txt,vec):
    buffer = io.BytesIO()
    jnp.savez(buffer, txt=txt, vec=vec)
    buffer.seek(0)
    encoded = base64.b64encode(buffer.getvalue()).decode('utf-8')
    return encoded

# t5,clip=load_encoders()

@spaces.GPU(duration=20)
def convert(prompt):
    t5,clip=models["t5"],models["clip"]
    if isinstance(prompt, str):
        prompt = [prompt]
    txt = t5.tokenize(prompt)
    txt = t5(txt)
    vec = clip.tokenize(prompt)
    vec = clip(vec)
    return b64(txt,vec)
import jax
def _to_embed(t5, clip, txt, vec):
    t5=nnx.merge(*t5)
    clip=nnx.merge(*clip)
    return t5(txt), clip(vec)

to_embed=jax.jit(_to_embed)

# t5_tuple=nnx.split(t5)
# clip_tuple=nnx.split(clip)

@spaces.GPU(duration=120)
def compile(prompt):
    t5,clip,t5_tuple,clip_tuple=models["t5"],models["clip"],models["t5_tuple"],models["clip_tuple"]
    if isinstance(prompt, str):
        prompt = [prompt]
    txt = t5.tokenize(prompt)
    vec = clip.tokenize(prompt)
    text,vec=to_embed(t5_tuple,clip_tuple,txt,vec)
    return b64(txt,vec)

@spaces.GPU(duration=120)
def load(prompt):
    is_schnell = True
    t5 = load_t5("cuda", max_length=256 if is_schnell else 512)
    clip = load_clip("cuda")
    models["t5"]=t5
    models["clip"]=clip
    models["t5_tuple"]=nnx.split(t5)
    models["clip_tuple"]=nnx.split(clip)
    return "Loaded"

print(load(""))

with gr.Blocks() as demo:
    gr.Markdown("""A workaround for flux-flax to fit into 40G VRAM""")
    with gr.Row():
        with gr.Column():
            prompt = gr.Textbox(label="prompt")
            convert_btn = gr.Button(value="Convert")
            compile_btn = gr.Button(value="Compile")
            load_btn = gr.Button(value="Load")
        with gr.Column():
            output = gr.Textbox(label="output")
    load_btn.click(load, inputs=prompt, outputs=output, api_name="load")
    convert_btn.click(convert, inputs=prompt, outputs=output, api_name="convert")
    compile_btn.click(compile, inputs=prompt, outputs=output, api_name="compile")




demo.launch()