[AMD] [GLM-5.3-Flash Day 0] Honor fused and per-expert names in quark exclude (#39317)

Co-authored-by: Yikai Zhang <ykzhang12@gmail.com>
Co-authored-by: Thomas Wang <thomawan@amd.com>
Co-authored-by: Kevin Mi <45493463+kevin-mii@users.noreply.github.com>
Co-authored-by: Kevin Mi <mikevin920@yahoo.com>
Co-authored-by: Claude Fable 5.1 <noreply@anthropic.com>

python/sglang/srt/layers/quantization/quark/utils.py

@@ -55,6 +55,19 @@ def should_ignore_layer(
     # proj_name = qkv_proj
     proj_name = layer_name.split(".")[-1]
 
+    # a fused module can be excluded under its fused name, so match it before expanding
+    if check_equal_or_regex_match(layer_name=layer_name, targets=ignore):
+        return True
+
+    # excludes may name experts individually, so an excluded expert excludes the module
+    if layer_name.endswith(".experts"):
+        expert_prefix = layer_name + "."
+        if any(
+            isinstance(target, str) and target.startswith(expert_prefix)
+            for target in ignore
+        ):
+            return True
+
     # Fused layers like gate_up_proj or qkv_proj will not be fused
     # in the safetensors checkpoint. So, we convert the name
     # from the fused version to unfused + check to make sure that
@@ -87,12 +100,9 @@ def should_ignore_layer(
                     "requires all to use the same scheme."
                 )
 
-    # Unfused layers like down_proj and o_proj will match
-    # the safetensors checkpoint already.
+    # an unfused name was already tried by the direct check above
     else:
-        should_ignore_layer = check_equal_or_regex_match(
-            layer_name=layer_name, targets=ignore
-        )
+        should_ignore_layer = False
 
     assert should_ignore_layer is not None