| """ |
| FLUX.2-klein-4B GGUF โ CPU-only Inference Space |
| ็กฌไปถ๏ผๆ GPU ยท CPU ยท 18 GB RAM |
| """ |
|
|
| import os |
| import sys |
| import json |
| import torch |
| import numpy as np |
| from huggingface_hub import login, hf_hub_download, list_repo_files |
|
|
| |
| os.environ["TOKENIZERS_PARALLELISM"] = "false" |
| torch.set_num_threads(os.cpu_count() or 4) |
|
|
| GGUF_REPO = "unsloth/FLUX-2-klein-4B-GGUF" |
| TOKEN = os.environ.get("HF_TOKEN", "") |
|
|
| |
| print("โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ") |
| print("โ FLUX.2-klein-4B GGUF โ CPU Inference โ") |
| print("โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ") |
|
|
| if TOKEN: |
| login(token=TOKEN, add_to_git_credential=False) |
| print(" โ HF_TOKEN authenticated") |
| else: |
| print(" โ HF_TOKEN not set!") |
| print(" โ Go to Space Settings โ Repository Secrets โ Add HF_TOKEN") |
| print(" โ Get token: https://huggingface.co/settings/tokens") |
| print(" โ Also accept model license on the model page first!\n") |
|
|
| print(f" Model : {GGUF_REPO}") |
| print(f" Threads : {torch.get_num_threads()}") |
| print(f" PyTorch : {torch.__version__}") |
|
|
| |
| from diffusers import FluxPipeline, FluxTransformer2DModel |
|
|
| pipe = None |
|
|
| |
| |
| |
| if pipe is None: |
| try: |
| print("\n[A] Trying direct diffusers loading...") |
| pipe = FluxPipeline.from_pretrained( |
| GGUF_REPO, |
| torch_dtype=torch.float32, |
| token=TOKEN or None, |
| ) |
| pipe.to("cpu") |
| print("[A] โ Direct loading succeeded!\n") |
| except Exception as e: |
| print(f"[A] โ Failed: {e}\n") |
|
|
| |
| |
| |
| if pipe is None: |
| try: |
| print("[B] Manual GGUF loading...") |
| from gguf import GGUFReader |
|
|
| |
| print(" Listing repo files...") |
| repo_files = list_repo_files(GGUF_REPO, token=TOKEN or None) |
| gguf_candidates = [f for f in repo_files if f.endswith(".gguf")] |
| if not gguf_candidates: |
| raise FileNotFoundError("No .gguf file found in repo") |
|
|
| |
| preferred = [f for f in gguf_candidates if "Q4_K_M" in f.upper()] |
| gguf_fname = preferred[0] if preferred else gguf_candidates[0] |
| print(f" Selected: {gguf_fname}") |
|
|
| |
| print(" Downloading GGUF file...") |
| gguf_path = hf_hub_download( |
| GGUF_REPO, gguf_fname, token=TOKEN or None |
| ) |
| print(f" Saved to: {gguf_path}") |
|
|
| |
| print(" Reading GGUF tensors...") |
| reader = GGUFReader(gguf_path) |
| gguf_sd = {} |
| for tensor in reader.tensors: |
| name = tensor.name |
| if isinstance(name, bytes): |
| name = name.decode("utf-8") |
| data = np.array(tensor.data, dtype=np.float32) |
| data = data.reshape(tensor.shape).copy() |
| gguf_sd[name] = torch.from_numpy(data) |
|
|
| print(f" Loaded {len(gguf_sd)} tensors") |
| sample_keys = list(gguf_sd.keys())[:5] |
| print(f" Sample keys: {sample_keys}") |
|
|
| |
| print(" Loading transformer config...") |
| config_path = hf_hub_download( |
| GGUF_REPO, "transformer/config.json", token=TOKEN or None |
| ) |
| with open(config_path, "r") as f: |
| tf_config = json.load(f) |
|
|
| |
| print(" Building transformer and loading weights...") |
| transformer = FluxTransformer2DModel.from_config(tf_config) |
|
|
| |
| loaded_ok = False |
|
|
| |
| try: |
| info = transformer.load_state_dict(gguf_sd, strict=False) |
| missing = [k for k in info.missing_keys if "bias" not in k.lower()] |
| if len(missing) < 5: |
| loaded_ok = True |
| print(f" B1 โ Direct mapping (missing={len(missing)})") |
| except Exception: |
| pass |
|
|
| |
| if not loaded_ok: |
| try: |
| mapped = {"transformer." + k: v for k, v in gguf_sd.items()} |
| info = transformer.load_state_dict(mapped, strict=False) |
| missing = [k for k in info.missing_keys if "bias" not in k.lower()] |
| if len(missing) < 5: |
| loaded_ok = True |
| print(f" B2 โ Prefix mapping (missing={len(missing)})") |
| except Exception: |
| pass |
|
|
| |
| if not loaded_ok: |
| try: |
| mapped = {} |
| for k, v in gguf_sd.items(): |
| new_key = k |
| for prefix in ["transformer.", "model."]: |
| if k.startswith(prefix): |
| new_key = k[len(prefix):] |
| break |
| mapped[new_key] = v |
| info = transformer.load_state_dict(mapped, strict=False) |
| missing = [k for k in info.missing_keys if "bias" not in k.lower()] |
| if len(missing) < 5: |
| loaded_ok = True |
| print(f" B3 โ Strip-prefix mapping (missing={len(missing)})") |
| except Exception: |
| pass |
|
|
| if not loaded_ok: |
| |
| model_keys = set(transformer.state_dict().keys()) |
| gguf_keys = set(gguf_sd.keys()) |
| print(f"\n โโ Key mismatch diagnostics โโ") |
| print(f" Model expects {len(model_keys)} keys") |
| print(f" GGUF has {len(gguf_keys)} keys") |
| print(f" Model key sample : {list(model_keys)[:3]}") |
| print(f" GGUF key sample : {list(gguf_keys)[:3]}") |
| raise RuntimeError("Could not map GGUF keys to model keys") |
|
|
| |
| print(" Building full pipeline...") |
| pipe = FluxPipeline.from_pretrained( |
| GGUF_REPO, |
| transformer=transformer, |
| torch_dtype=torch.float32, |
| token=TOKEN or None, |
| ) |
| pipe.to("cpu") |
| print("[B] โ Manual GGUF loading succeeded!\n") |
|
|
| except Exception as e: |
| print(f"[B] โ Failed: {e}") |
| import traceback |
| traceback.print_exc() |
| print() |
|
|
| |
| if pipe is None: |
| print("=" * 58) |
| print("FATAL: Could not load model.") |
| print("") |
| print("Please check:") |
| print(" 1. HF_TOKEN is set as a Space secret") |
| print(" 2. You have accepted the model license at:") |
| print(f" https://huggingface.co/{GGUF_REPO}") |
| print("=" * 58) |
| sys.exit(1) |
|
|
| print("โ Model ready for inference!\n") |
|
|
| |
| @torch.no_grad() |
| def generate( |
| prompt: str, |
| negative_prompt: str, |
| width: int, |
| height: int, |
| steps: int, |
| guidance: float, |
| seed: int, |
| progress=gr.Progress(track_tqdm=True), |
| ): |
| if not prompt.strip(): |
| raise gr.Error("่ฏท่พๅ
ฅๆ็คบ่ฏ") |
|
|
| seed = seed if seed >= 0 else int(time.time() * 1000) % (2**32) |
| generator = torch.Generator("cpu").manual_seed(seed) |
|
|
| t0 = time.time() |
| result = pipe( |
| prompt=prompt, |
| negative_prompt=negative_prompt if negative_prompt.strip() else None, |
| width=width, |
| height=height, |
| num_inference_steps=steps, |
| guidance_scale=guidance, |
| generator=generator, |
| ) |
| dt = time.time() - t0 |
|
|
| return result.images[0], seed, f"ๅฎๆ ยท ่ๆถ {dt:.1f}s ยท seed {seed}" |
|
|
|
|
| |
| CSS = """ |
| :root{--bg:#0a0a0f;--card:#13131a;--border:#23233a;--fg:#e4e4ec;--muted:#6b6b80;--accent:#7c5cfc;--glow:rgba(124,92,252,.25)} |
| .gradio-container{background:var(--bg)!important;color:var(--fg)!important;max-width:1080px!important;padding-top:1rem!important} |
| .gr-block,.gr-panel{background:var(--card)!important;border:1px solid var(--border)!important;border-radius:14px!important} |
| .gr-input,textarea.gr-textbox,input[type=number].gr-number{background:#0e0e16!important;border:1px solid var(--border)!important;color:var(--fg)!important;border-radius:10px!important;font-size:.92rem!important} |
| textarea.gr-textbox{resize:vertical!important} |
| .gr-label{color:var(--muted)!important;font-size:.78rem!important;text-transform:uppercase!important;letter-spacing:.06em!important;font-weight:600!important} |
| .gr-button-primary{background:linear-gradient(135deg,#7c5cfc,#a78bfa)!important;border:none!important;font-weight:700!important;font-size:1rem!important;border-radius:10px!important;padding:.7rem 1.5rem!important;box-shadow:0 4px 24px var(--glow)!important;transition:all .2s ease!important} |
| .gr-button-primary:hover{transform:translateY(-1px)!important;box-shadow:0 6px 32px var(--glow)!important;filter:brightness(1.1)!important} |
| .gr-slider input[type=range]{accent-color:var(--accent)!important} |
| .gr-image{border-radius:12px!important;overflow:hidden!important} |
| footer{display:none!important} |
| ::-webkit-scrollbar{width:6px}::-webkit-scrollbar-track{background:transparent}::-webkit-scrollbar-thumb{background:#333;border-radius:3px} |
| .warn-bar{background:linear-gradient(90deg,#1c1a2e,#1a1428);border:1px solid #2e2848;border-radius:10px;padding:.7rem 1rem;margin-bottom:1.2rem;font-size:.82rem;color:#a0a0b8;line-height:1.5} |
| .warn-bar strong{color:#c4b5fd} |
| .info-bar{background:#0e0e16;border:1px solid var(--border);border-radius:8px;padding:.5rem .8rem;font-size:.82rem;color:var(--muted);font-family:'SF Mono','Fira Code',monospace} |
| """ |
|
|
| import time |
| import gradio as gr |
|
|
| with gr.Blocks(title="FLUX.2-klein-4B GGUF", css=CSS, theme=gr.themes.Soft(primary_hue="violet")) as demo: |
|
|
| gr.HTML(""" |
| <div style="text-align:center;padding:2rem 0 .6rem"> |
| <h1 style="font-size:2.4rem;font-weight:800;letter-spacing:-.03em; |
| background:linear-gradient(135deg,#7c5cfc 0%,#c084fc 50%,#f0abfc 100%); |
| -webkit-background-clip:text;-webkit-text-fill-color:transparent; |
| margin:0;line-height:1.2">FLUX.2-klein-4B</h1> |
| <p style="color:#555;font-size:.88rem;margin-top:.5rem;font-weight:500; |
| letter-spacing:.08em;text-transform:uppercase"> |
| GGUF Quantized ยท CPU Inference ยท 4B Parameters</p> |
| </div>""") |
|
|
| gr.HTML(""" |
| <div class="warn-bar"> |
| ⚠ ็บฏ CPU ๆจ็๏ผ้ๅบฆ่พๆ
ขใๅปบ่ฎฎ <strong>512ร512</strong> ไปฅไธ + |
| <strong>8~12 ๆญฅ</strong>ใ้ฆๆฌก็ๆๅ JIT ็ผ่ฏๆดๆ
ข๏ผๅ็ปญๅ ้ใ |
| ่ถ
ๆถ่ฏท้ไฝๅ่พจ็ๆๆญฅๆฐใ |
| </div>""") |
|
|
| with gr.Row(equal_height=False): |
| with gr.Column(scale=3): |
| prompt = gr.Textbox(label="Prompt", placeholder="A cyberpunk city at night with neon signs reflecting in rain-soaked streets...", lines=3) |
| negative_prompt = gr.Textbox(label="Negative Prompt", placeholder="blurry, low quality, distorted...", lines=1) |
| with gr.Row(): |
| w = gr.Slider(256, 768, value=512, step=64, label="Width") |
| h = gr.Slider(256, 768, value=512, step=64, label="Height") |
| with gr.Row(): |
| steps = gr.Slider(4, 30, value=10, step=1, label="Steps") |
| guidance = gr.Slider(1.0, 10.0, value=3.5, step=0.5, label="Guidance Scale") |
| seed = gr.Number(label="Seed๏ผ-1 ้ๆบ๏ผ", value=-1, precision=0) |
| btn = gr.Button("โฆ Generate", variant="primary", size="lg") |
|
|
| gr.Examples(examples=[ |
| ["A serene Japanese garden with cherry blossoms over a stone bridge, koi pond, golden hour, 4K", ""], |
| ["Futuristic space station interior, holographic star maps, bioluminescent plants, cinematic wide shot", ""], |
| ["Oil painting of a cozy wooden cabin in snowy mountains, warm light from windows, impressionist style", ""], |
| ["Macro photograph of a dewdrop on a spider web at sunrise, prismatic light, shallow depth of field", ""], |
| ["Whimsical treehouse library at twilight, floating books with glowing pages, fairy tale atmosphere", "blurry"], |
| ["Underwater coral reef city, bioluminescent architecture, tropical fish, volumetric god rays", ""], |
| ], inputs=[prompt, negative_prompt], label="Quick Examples") |
|
|
| with gr.Column(scale=4): |
| output_img = gr.Image(label="Output", show_label=False, height=540) |
| info_text = gr.Textbox(label="Status", interactive=False, max_lines=1, elem_classes="info-bar", value="Ready") |
|
|
| btn.click(fn=generate, inputs=[prompt, negative_prompt, w, h, steps, guidance, seed], outputs=[output_img, seed, info_text]) |
| prompt.submit(fn=generate, inputs=[prompt, negative_prompt, w, h, steps, guidance, seed], outputs=[output_img, seed, info_text]) |
|
|
| demo.launch(server_name="0.0.0.0", server_port=7860) |
