Update app.py

app.py

@@ -42,9 +42,9 @@ GPU_SIZE = os.environ.get("H3_GPU_SIZE", "xlarge")
 # default, which reproduces the official weights exactly. Confirmed diffusers-native key naming (not ComfyUI,
 # not pruned-AdaLN) via `xal2077/PinkCherry_MiniMax-H3-Demo`'s own working `load_state_dict(strict=True)` call
 # against an earlier release of the same lineage.
-CUSTOM_TRANSFORMER_REPO = os.environ.get("H3_CUSTOM_TRANSFORMER_REPO", "Comfy-Org/MiniMax-H3")
+CUSTOM_TRANSFORMER_REPO = os.environ.get("H3_CUSTOM_TRANSFORMER_REPO", "SexGod1979/PinkCherry_MiniMax-H3")
 CUSTOM_TRANSFORMER_FILE = os.environ.get(
-    "H3_CUSTOM_TRANSFORMER_FILE", "diffusion_models/minimax_h3_fl2va_bf16.safetensors"
+    "H3_CUSTOM_TRANSFORMER_FILE", "v1-final-fl2va/PinkCherry_v1_bf16_fla2va_H3.safetensors"
 )
 
 LORA_REPO = os.environ.get("H3_LORA_REPO", "dagloop5/LoRA")
@@ -511,6 +511,22 @@ def load_models() -> str | None:
             with safe_open(custom_path, framework="pt") as handle:
                 raw = {k: handle.get_tensor(k) for k in handle.keys()}
             converted = _convert_full_checkpoint(raw, base_shapes)
+            if os.environ.get("H3_DEBUG_COMPARE", "0") == "1":
+                from huggingface_hub import hf_hub_download as _dl
+                import json as _json
+            
+                index_path = _dl(MODEL_REPO, "transformer/diffusion_pytorch_model.safetensors.index.json")
+                with open(index_path) as handle:
+                    index = _json.load(handle)
+                probe_key = "transformer_blocks.0.norm1.weight"
+                shard_name = index["weight_map"][probe_key]
+                shard_path = _dl(MODEL_REPO, f"transformer/{shard_name}")
+                with safe_open(shard_path, framework="pt") as official_handle:
+                    official_tensor = official_handle.get_tensor(probe_key)
+                converted_tensor = converted[probe_key]
+                print(f"[debug-compare] official {probe_key}: mean={official_tensor.float().mean():.6f} std={official_tensor.float().std():.6f} first5={official_tensor.flatten()[:5].tolist()}", flush=True)
+                print(f"[debug-compare] converted {probe_key}: mean={converted_tensor.float().mean():.6f} std={converted_tensor.float().std():.6f} first5={converted_tensor.flatten()[:5].tolist()}", flush=True)
+                print(f"[debug-compare] allclose: {torch.allclose(official_tensor, converted_tensor)}", flush=True)
             custom_transformer.load_state_dict(converted, strict=True, assign=True)
 
             # `rope.inv_freq` is a *non-persistent* buffer (`persistent=False` in `MiniMaxH3RotaryPosEmbed`) —