Delete pixeldit/pipeline.py with huggingface_hub

pixeldit/pipeline.py

@@ -1,81 +0,0 @@
-"""
-PixelDiT T2I Pipeline — thin orchestrator.
-
-Usage:
-    from pipeline import PixelDiTPipeline
-    pipe = PixelDiTPipeline()
-    images = pipe("a viking at sunset", height=512, width=512)
-    images[0].save("out.jpg")
-"""
-
-import torch
-from PIL import Image
-from .modeling_pixeldit import load_pixeldit
-from .modeling_pixeldit_hf import PixelDiTModel
-from .text_encoder_gemma import GemmaEncoder
-from .text_encoder_qwen import QwenEncoder
-from .scheduling_flow import FlowScheduler
-
-
-class PixelDiTPipeline:
-    def __init__(
-        self,
-        text_encoder="gemma",   # "gemma" | "qwen"
-        qwen_proj=None,
-        device="cuda",
-        dtype=torch.bfloat16,
-        cfg=3.5,
-        flow_shift=4.0,
-        pretrained=None,        # HF dir or repo id — loads via from_pretrained instead of .pth
-    ):
-        self.device = torch.device(device)
-        self.dtype  = dtype
-
-        if text_encoder == "qwen":
-            self.encoder = QwenEncoder(proj_path=qwen_proj, output_device=device, output_dtype=dtype)
-        else:
-            self.encoder = GemmaEncoder(output_device=device, output_dtype=dtype)
-
-        if pretrained is not None:
-            print(f"[pipeline] loading from HF: {pretrained}")
-            self.model = (
-                PixelDiTModel.from_pretrained(pretrained)
-                .to(device).to(dtype).eval()
-            )
-        else:
-            self.model = load_pixeldit(device=device, dtype=dtype)
-
-        self.scheduler = FlowScheduler(self.model, cfg=cfg, flow_shift=flow_shift)
-
-    @torch.no_grad()
-    def __call__(
-        self,
-        prompt,
-        negative_prompt="",
-        height=512,
-        width=512,
-        steps=20,
-        cfg=None,
-        seed=None,
-    ):
-        if isinstance(prompt, str):
-            prompts = [prompt]
-        else:
-            prompts = list(prompt)
-        B = len(prompts)
-
-        if cfg is not None:
-            self.scheduler.cfg = cfg
-        if seed is not None:
-            torch.manual_seed(seed)
-
-        cond   = self.encoder.encode(prompts)
-        uncond = (self.encoder.encode_null(B) if not negative_prompt
-                  else self.encoder.encode([negative_prompt] * B))
-
-        noise = torch.randn(B, 3, height, width, device=self.device, dtype=self.dtype)
-        imgs  = self.scheduler.sample(noise, cond, uncond, steps=steps)
-
-        imgs = (imgs.clamp(-1, 1) + 1) / 2
-        imgs = (imgs * 255).byte().permute(0, 2, 3, 1).cpu().numpy()
-        return [Image.fromarray(img) for img in imgs]