Upload teacher_code/builder.py with huggingface_hub

teacher_code/builder.py

@@ -0,0 +1,81 @@
+import os
+from .clip_encoder import CLIPVisionTower
+from .imagebind import ImageBindWrapper
+from .open_clip_encoder import OpenCLIPVisionTower
+from .siglip_encoder import SigLipVisionTower
+from .clip_encoder import CLIPVisionTower, CLIPVisionTowerS2
+
+from .eva_clip.eva_clip_encoder import EvaClipVisionTower
+from .dev_eva_clip.eva_vit import EvaViTWrapper
+
+from blip3o.model.nextdit_crossattn import NextDiTCrossAttnConfig, NextDiTCrossAttn
+from blip3o.model.sana_crossattn import SanaCrossAttnConfig, SanaCrossAttn
+
+from diffusers.models import AutoencoderKL
+from diffusers.schedulers import FlowMatchEulerDiscreteScheduler, DPMSolverMultistepScheduler
+from diffusers import SanaTransformer2DModel
+import torch
+
+def build_vision_tower(vision_tower_cfg, **kwargs):
+    vision_tower = getattr(vision_tower_cfg, 'mm_vision_tower', getattr(vision_tower_cfg, 'vision_tower', None))
+    is_absolute_path_exists = os.path.exists(vision_tower)
+    use_s2 = getattr(vision_tower_cfg, 's2', False)
+    if "siglip" in vision_tower:
+        return SigLipVisionTower(vision_tower, vision_tower_cfg=vision_tower_cfg, **kwargs)
+    if "eva" in vision_tower:
+        return EvaClipVisionTower(vision_tower, args=vision_tower_cfg, **kwargs)
+    if is_absolute_path_exists or vision_tower.startswith("openai") or vision_tower.startswith("laion") or "ShareGPT4V" in vision_tower:
+        if use_s2:
+            return CLIPVisionTowerS2(vision_tower, args=vision_tower_cfg, **kwargs)
+        else:
+            return CLIPVisionTower(vision_tower, args=vision_tower_cfg, **kwargs)
+
+    raise ValueError(f'Unknown vision tower: {vision_tower}')
+
+
+
+
+def build_gen_vision_tower(vision_tower_cfg, **kwargs):
+    vision_tower = getattr(vision_tower_cfg, 'gen_vision_tower')
+    is_absolute_path_exists = os.path.exists(vision_tower)
+    use_s2 = getattr(vision_tower_cfg, 's2', False)
+    if "siglip" in vision_tower:
+        return SigLipVisionTower(vision_tower, vision_tower_cfg=vision_tower_cfg, **kwargs)
+    if "eva" in vision_tower:
+        return EvaClipVisionTower(vision_tower, args=vision_tower_cfg, **kwargs)
+    if is_absolute_path_exists or vision_tower.startswith("openai") or vision_tower.startswith("laion") or "ShareGPT4V" in vision_tower:
+        if use_s2:
+            return CLIPVisionTowerS2(vision_tower, args=vision_tower_cfg, **kwargs)
+        else:
+            return CLIPVisionTower(vision_tower, args=vision_tower_cfg, **kwargs)
+
+    raise ValueError(f'Unknown vision tower: {vision_tower}')
+
+
+
+def build_dit(vision_tower_cfg, **kwargs):
+    vision_tower_cfg.hidden_size = 896
+    print("="*20, "vision tower config", vision_tower_cfg, "="*20)
+    # if not hasattr(vision_tower_cfg, "hidden_size"):
+    #     if "3B" in vision_tower_cfg.model_name_or_path:
+    #         vision_tower_cfg.hidden_size = 2048
+    #     elif "7B" in vision_tower_cfg.model_name_or_path:
+    #         vision_tower_cfg.hidden_size = 3584
+    #     else:
+    #         vision_tower_cfg.hidden_size = 3072
+    print("="*20, "Building SANA with hidden size", vision_tower_cfg.hidden_size, "="*20)
+    
+
+    #dit = NextDiTCrossAttn(NextDiTCrossAttnConfig(latent_embedding_size=vision_tower_cfg.hidden_size))
+    #noise_scheduler = FlowMatchEulerDiscreteScheduler.from_pretrained("Alpha-VLLM/Lumina-Next-SFT-diffusers", subfolder="scheduler")
+    #dit = SanaTransformer2DModel.from_pretrained("Efficient-Large-Model/Sana_600M_512px_diffusers", subfolder="transformer", torch_dtype=torch.float16,)
+
+    # dit = SanaCrossAttn(SanaCrossAttnConfig()) #cross_attention_dim=vision_tower_cfg.hidden_size))
+    #dit = SanaTransformer2DModel.from_pretrained("Efficient-Large-Model/Sana_600M_512px_diffusers", subfolder="transformer")
+    dit = SanaTransformer2DModel.from_pretrained("Efficient-Large-Model/Sana_600M_1024px_diffusers",device_map="cpu", subfolder="transformer")
+    dit_teacher = SanaTransformer2DModel.from_pretrained("Efficient-Large-Model/SANA1.5_4.8B_1024px_diffusers", device_map="cpu", subfolder="transformer")
+    noise_scheduler = DPMSolverMultistepScheduler.from_pretrained("Efficient-Large-Model/Sana_600M_512px_diffusers",subfolder="scheduler")
+
+    return dit, dit_teacher, noise_scheduler
+
+