Upload 2 files

quantized/combine.py

@@ -37,16 +37,16 @@ if os.path.isfile(metadata_file):
     try:
         with open(metadata_file, "r", encoding="utf-8") as f:
             metadata = json.load(f)
-        # Convert nested objects to JSON strings as required by safetensors
+        # Convert nested objects to JSON strings
         for k, v in metadata.items():
             if isinstance(v, dict):
                 metadata[k] = json.dumps(v)
     except Exception as e:
         print(f"⚠️ Failed to load metadata from {metadata_file}: {e}")
         print("⏳ Falling back to hardcoded metadata...")
-        metadata = {}  # Will populate below
+        metadata = {}
 
-# Fallback metadata (if file load failed or was missing)
+# Hardcoded FP4 metadata fallback
 if not metadata:
     metadata = {
         "model_class": "NunchakuFluxTransformer2dModel",
@@ -71,11 +71,6 @@ if not metadata:
                 "vec_in_dim": 768
             }
         }),
-        "quantization_config": json.dumps({
-            "method": "svdquant",
-            "weight": {"dtype": "int4", "scale_dtype": None, "group_size": 64},
-            "activation": {"dtype": "int4", "scale_dtype": None, "group_size": 64}
-        }),
         "config": json.dumps({
             "_class_name": "FluxTransformer2DModel",
             "_diffusers_version": "0.34.0.dev0",
@@ -91,10 +86,23 @@ if not metadata:
             "out_channels": None,
             "patch_size": 1,
             "pooled_projection_dim": 768
+        }),
+        "quantization_config": json.dumps({
+            "method": "svdquant",
+            "weight": {
+                "dtype": "fp4_e2m1_all",
+                "scale_dtype": [None, "fp8_e4m3_nan"],
+                "group_size": 16
+            },
+            "activation": {
+                "dtype": "fp4_e2m1_all",
+                "scale_dtype": "fp8_e4m3_nan",
+                "group_size": 16
+            }
         })
     }
 
-# Save the combined safetensors file
+# Save the combined file
 safetensors.torch.save_file(combined_state_dict, output_file, metadata=metadata)
 
 print(f"\n✅ Combined model saved to: {output_file}")

quantized/metadata.json

@@ -21,11 +21,6 @@
       "vec_in_dim": 768
     }
   },
-  "quantization_config": {
-    "method": "svdquant",
-    "weight": {"dtype": "int4", "scale_dtype": null, "group_size": 64},
-    "activation": {"dtype": "int4", "scale_dtype": null, "group_size": 64}
-  },
   "config": {
     "_class_name": "FluxTransformer2DModel",
     "_diffusers_version": "0.34.0.dev0",
@@ -41,5 +36,18 @@
     "out_channels": null,
     "patch_size": 1,
     "pooled_projection_dim": 768
+  },
+  "quantization_config": {
+    "method": "svdquant",
+    "weight": {
+      "dtype": "fp4_e2m1_all",
+      "scale_dtype": [null, "fp8_e4m3_nan"],
+      "group_size": 16
+    },
+    "activation": {
+      "dtype": "fp4_e2m1_all",
+      "scale_dtype": "fp8_e4m3_nan",
+      "group_size": 16
+    }
   }
-}
+}