app.py

import torch
import torch.nn as nn
import gradio as gr
from transformers import AutoTokenizer, Qwen2ForCausalLM

device = "cpu"
dtype = torch.float32

# Qwen 0.5B = hidden_size 896
tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen2-0.5B")
text_encoder = Qwen2ForCausalLM.from_pretrained(
    "Qwen/Qwen2-0.5B",
    torch_dtype=dtype,
)

# Projection 896 -> 2048 pour FLUX.1-Schnell
# Projection 896 -> 4096
proj_tokens = nn.Linear(896, 4096)

# Projection pooled 4096 -> 3072
proj_pooled = nn.Linear(4096, 3072)



def encode(prompt: str):
    if not prompt or prompt.strip() == "":
        prompt = tokenizer.eos_token or "."

    tokens = tokenizer(
        prompt,
        return_tensors="pt",
        padding=True,
        truncation=True,
        max_length=512,
    )

    out = text_encoder(
        **tokens,
        output_hidden_states=True,
        use_cache=False,
    )

    # Embeddings Qwen (896 dims)
    embeds_896 = out.hidden_states[-1]  # [1, L, 896]

    # Projection -> 2048 dims
    embeds_2048 = proj_tokens(embeds_896)  # [1, L, 2048]

    # pooled -> moyenne -> projection 768 dims
    pooled_2048 = embeds_2048.mean(dim=1)  # [1, 2048]
    pooled_768 = proj_pooled(pooled_2048)  # [1, 768]

    # Sauvegarde
    torch.save(embeds_2048, "embeds.pt")
    torch.save(pooled_768, "pooled.pt")

    return str(embeds_2048.shape), "embeds.pt", "pooled.pt"

demo = gr.Interface(
    fn=encode,
    inputs=gr.Textbox(label="Prompt"),
    outputs=[
        gr.Textbox(label="Shape"),
        gr.File(label="Embeddings 2048"),
        gr.File(label="Pooled 768")
    ],
    title="External Text Encoder — FLUX.1‑Schnell Compatible"
)

demo.launch()