[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
 

test/registered/unit/layers/quantization/test_quark_utils.py

@@ -115,5 +115,46 @@ class TestE8M0ToF32(CustomTestCase):
         self.assertTrue(torch.isnan(out[2]).item())
 
 
+QKV_MAPPING = {"qkv_proj": ["q_proj", "k_proj", "v_proj"]}
+
+
+class TestShouldIgnoreLayerFusedNames(CustomTestCase):
+    """An `exclude` entry naming an already-fused module, or naming experts
+    individually, must exclude the fused module SGLang builds; otherwise an
+    MXFP4-packed parameter is allocated for a BF16 tensor and loading aborts."""
+
+    # ---- Bug-catchers: must FAIL on unfixed code ---------------------------
+
+    def test_directly_excluded_fused_qkv_is_ignored(self):
+        name = "visual.blocks.0.attn.qkv_proj"
+        self.assertTrue(
+            should_ignore_layer(name, ignore=[name], fused_mapping=QKV_MAPPING)
+        )
+
+    def test_per_expert_excludes_ignore_the_fused_moe_module(self):
+        layer = "model.layers.6.mlp.experts"
+        ignore = [
+            f"{layer}.{i}.{proj}"
+            for i in range(3)
+            for proj in ("down_proj", "gate_proj", "up_proj")
+        ]
+        self.assertTrue(
+            should_ignore_layer(layer, ignore=ignore, fused_mapping=QKV_MAPPING)
+        )
+
+    # ---- Guards: behavior that must NOT change -----------------------------
+
+    def test_unrelated_moe_layer_is_not_ignored(self):
+        # a prefix match must not bleed into a neighboring layer index
+        ignore = ["model.layers.6.mlp.experts.0.down_proj"]
+        self.assertFalse(
+            should_ignore_layer(
+                "model.layers.7.mlp.experts",
+                ignore=ignore,
+                fused_mapping=QKV_MAPPING,
+            )
+        )
+
+
 if __name__ == "__main__":
     unittest.main()