Use bf16 text encoder + bf16 stock fallback: fp8 weights bypass ZeroGPU packing (comfy_kitchen QuantizedTensor) and produce NaN

app.py

@@ -82,13 +82,17 @@ except Exception as exc:  # pragma: no cover
 # Companion models always pulled from Comfy-Org/Krea-2 (the CivitAI checkpoint
 # is UNet-only): (repo_id, subfolder, filename). local_dir = models/ root, so
 # the repo subfolder (text_encoders/vae/loras) is replicated under models/.
+#
+# IMPORTANT: bf16 files only. fp8 models are comfy_kitchen QuantizedTensor
+# objects whose .to("cuda") bypasses ZeroGPU's torch patch, so they are never
+# packed/streamed to VRAM and inference produces NaN. bf16 packs fine.
 COMPANION_MODELS = [
-    ("Comfy-Org/Krea-2", "text_encoders", "qwen3vl_4b_fp8_scaled.safetensors"),
+    ("Comfy-Org/Krea-2", "text_encoders", "qwen3vl_4b_bf16.safetensors"),
     ("Comfy-Org/Krea-2", "vae", "qwen_image_vae.safetensors"),
     ("Comfy-Org/Krea-2", "loras", "krea2_darkbrush.safetensors"),
 ]
 
-STOCK_UNET = ("Comfy-Org/Krea-2", "diffusion_models", "krea2_turbo_fp8_scaled.safetensors")
+STOCK_UNET = ("Comfy-Org/Krea-2", "diffusion_models", "krea2_turbo_bf16.safetensors")
 
 # CivitAI checkpoint (the diffusion model):
 #   CIVIT_API_KEY        secret on the Space (required)
@@ -260,7 +264,7 @@ unet_loader = UNETLoader()
 UNET = unet_loader.load_unet(unet_name=UNET_NAME, weight_dtype="default")
 
 clip_loader = CLIPLoader()
-CLIP = clip_loader.load_clip(clip_name="qwen3vl_4b_fp8_scaled.safetensors", type="krea2")
+CLIP = clip_loader.load_clip(clip_name="qwen3vl_4b_bf16.safetensors", type="krea2")
 
 vae_loader = VAELoader()
 VAE = vae_loader.load_vae(vae_name="qwen_image_vae.safetensors")