Improves cli args handling

convert_to_quant/cli/main.py

@@ -71,11 +71,6 @@ def extract_filter_flags(args) -> dict:
         if getattr(args, name):
             flags[name] = True
 
-    # Text model aliases: these flags imply generic_text input_scale behavior
-    TEXT_MODEL_ALIASES = {"qwen35"}
-    if any(flags.get(alias) for alias in TEXT_MODEL_ALIASES):
-        flags["generic_text"] = True
-
     return flags
 
 
@@ -325,7 +320,7 @@ def run_conversion(args):
     # Handle NVFP4 quantization mode (separate workflow OR unified if mixing formats)
     if args.nvfp4:
         # Check if we need mixed format support
-        needs_mixing = args.custom_type or args.fallback
+        needs_mixing = args.custom_type or args.fallback or args.layer_config
 
         if needs_mixing:
             # Route through unified path for mixed format support
@@ -447,7 +442,7 @@ def run_conversion(args):
     # Handle MXFP8 quantization mode (separate workflow OR unified if mixing formats)
     if args.mxfp8:
         # Check if we need mixed format support
-        needs_mixing = args.custom_type or args.fallback
+        needs_mixing = args.custom_type or args.fallback or args.layer_config
 
         if needs_mixing:
             # Route through unified path for mixed format support
@@ -730,9 +725,9 @@ def run_conversion(args):
     # Call convert_to_fp8_scaled with explicit args (no **kwargs footgun)
     # Determine primary_format for NVFP4/MXFP8 mixed mode (when they fall through here)
     primary_format = None
-    if args.nvfp4 and (args.custom_type or args.fallback):
+    if args.nvfp4 and (args.custom_type or args.fallback or args.layer_config):
         primary_format = "nvfp4"
-    elif args.mxfp8 and (args.custom_type or args.fallback):
+    elif args.mxfp8 and (args.custom_type or args.fallback or args.layer_config):
         primary_format = "mxfp8"
 
     convert_to_fp8_scaled(

convert_to_quant/formats/fp8_conversion.py

@@ -330,6 +330,10 @@ def get_format_info(fmt: str) -> dict:
                     layer_format = "fp8"
                 elif fmt.startswith("int8"):
                     layer_format = "int8"
+                elif fmt == "mxfp8":
+                    layer_format = "mxfp8"
+                elif fmt == "nvfp4":
+                    layer_format = "nvfp4"
                 else:
                     layer_format = "fp8"  # fallback
 

tests/test_filter_flags.py

@@ -9,7 +9,7 @@
   - Excluded layers (via filter "exclude", "highprec", or "remove") are handled correctly
   - 1D/3D/4D tensors are never quantized regardless of layer name
   - AVOID_KEY_NAMES (norm, bias, embed_tokens…) are skipped in nvfp4/mxfp8 paths
-  - TEXT_MODEL_ALIASES (qwen35 -> generic_text) propagate correctly
+  - filter extraction uses only explicitly enabled filter flags
   - Quantized layers produce .weight_scale and .comfy_quant
   - Skipped layers do NOT produce .weight_scale / .comfy_quant / .input_scale
   - "remove" key deletes layers entirely from output (t5xxl decoder removal)
@@ -385,11 +385,11 @@ def test_mxfp8_qwen35_flag_skips_excluded_layers(self):
             self.assertIn(f"{base}.weight", out)
 
     # ------------------------------------------------------------------
-    # 6. TEXT_MODEL_ALIASES: qwen35 via extract_filter_flags injects generic_text
+    # 6. extract_filter_flags preserves only explicit filter flags
     # ------------------------------------------------------------------
 
-    def test_extract_filter_flags_qwen35_injects_generic_text(self):
-        """extract_filter_flags must set generic_text=True when qwen35 is set."""
+    def test_extract_filter_flags_qwen35_does_not_inject_generic_text(self):
+        """extract_filter_flags must not add generic_text when only qwen35 is set."""
         # Build a fake args namespace with all MODEL_FILTERS keys set to False
         # except qwen35
         import types
@@ -404,24 +404,25 @@ def test_extract_filter_flags_qwen35_injects_generic_text(self):
         flags = extract_filter_flags(ns)
 
         self.assertTrue(flags.get("qwen35"), "qwen35 must be True in flags")
-        self.assertTrue(flags.get("generic_text"), "generic_text must be auto-injected when qwen35 is set")
+        self.assertNotIn("generic_text", flags, "generic_text must not be auto-injected when qwen35 is set")
 
-    def test_extract_filter_flags_no_alias_without_qwen35(self):
-        """generic_text must NOT be injected when qwen35 is not set."""
+    def test_extract_filter_flags_keeps_explicit_generic_text(self):
+        """generic_text must be present only when explicitly enabled."""
         import types
 
         from convert_to_quant.constants import MODEL_FILTERS
 
         ns = types.SimpleNamespace()
         for name in MODEL_FILTERS.keys():
             setattr(ns, name, False)
+        setattr(ns, "generic_text", True)
 
         flags = extract_filter_flags(ns)
 
-        self.assertNotIn("generic_text", flags, "generic_text must not appear when no alias is active")
+        self.assertTrue(flags.get("generic_text"), "generic_text must be kept when explicitly active")
 
     # ------------------------------------------------------------------
-    # 7. --qwen35 input_scale behavior via generic_text injection (FP8)
+    # 7. Explicit qwen35 + generic_text input_scale behavior (FP8)
     # ------------------------------------------------------------------
 
     def test_fp8_qwen35_generic_text_produces_input_scale(self):