import os
import sys
import torch
from diffusers import Flux2Transformer2DModel
from diffusers.pipelines.flux2.pipeline_flux2_klein import Flux2KleinPipeline as P
from safetensors.torch import load_file, save_file
model, out = sys.argv[1], sys.argv[2]
torch.set_grad_enabled(False)
ref = load_file(f"{out}/flux2.safetensors")
transformer = Flux2Transformer2DModel.from_pretrained(
f"{model}/transformer", num_layers=2, num_single_layers=2,
torch_dtype=torch.bfloat16).float().eval()
noise = ref["noise"][None]
embeds = ref["prompt_embeds"][None]
pred = transformer(hidden_states=P._pack_latents(noise), timestep=torch.ones(1),
guidance=None, encoder_hidden_states=embeds,
txt_ids=P._prepare_text_ids(embeds), img_ids=P._prepare_latent_ids(noise),
return_dict=False)[0]
dumps = {"step0_pred_blocks": pred[0].contiguous()}
if "ref_latents" in ref:
tokens, ids = ref["ref_latents"][None], ref["ref_ids"][None]
pred = transformer(hidden_states=torch.cat([P._pack_latents(noise), tokens], dim=1),
timestep=torch.ones(1), guidance=None, encoder_hidden_states=embeds,
txt_ids=P._prepare_text_ids(embeds),
img_ids=torch.cat([P._prepare_latent_ids(noise), ids], dim=1),
return_dict=False)[0][:, :1024]
dumps["step0_pred_blocks_ref"] = pred[0].contiguous()
save_file(dumps, f"{out}/flux2_blocks.safetensors")
print("wrote", out, flush=True)
os._exit(0)