[Fix] FP8 Qwen3-Next quant error by removing fallback fused shards (#23973)

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

@@ -60,8 +60,6 @@ def _module_path_match(ignored: str, prefix: str) -> bool:
 _FALLBACK_FUSED_SHARDS: Mapping[str, List[str]] = {
     "qkv_proj": ["q_proj", "k_proj", "v_proj"],
     "gate_up_proj": ["gate_proj", "up_proj"],
-    "in_proj_ba": ["in_proj_b", "in_proj_a"],
-    "in_proj_qkvz": ["in_proj_qkv", "in_proj_z"],
 }
 
 

test/registered/quant/test_is_layer_skipped.py

@@ -0,0 +1,52 @@
+import unittest
+
+from sglang.srt.layers.quantization.utils import is_layer_skipped
+from sglang.test.ci.ci_register import register_cpu_ci
+from sglang.test.test_utils import CustomTestCase
+
+register_cpu_ci(est_time=5, suite="stage-a-test-cpu")
+
+
+# Qwen3-Next FP8 actually publishes the equivalent of this in
+# packed_modules_mapping (qwen3_next.py:908-911). in_proj_ba / in_proj_qkvz
+# are deliberately omitted because they are real, unified tensors.
+QWEN3_NEXT_FUSED_MAPPING = {
+    "qkv_proj": ["q_proj", "k_proj", "v_proj"],
+    "gate_up_proj": ["gate_proj", "up_proj"],
+}
+
+
+def _qwen3_next_ignored_layers(layer_idx: int, name: str) -> list:
+    # Mirrors the normalization in Fp8Config.from_config: each entry is kept in
+    # both "model.<...>" and bare "<...>" forms.
+    base = f"layers.{layer_idx}.linear_attn.{name}"
+    return [base, f"model.{base}"]
+
+
+class TestIsLayerSkipped(CustomTestCase):
+    def test_qwen3_next_in_proj_ba_is_skipped(self):
+        # Regression for #23467: in_proj_ba is a unified tensor in the FP8
+        # checkpoint. modules_to_not_convert lists it explicitly, so it must
+        # bypass FP8 quantization (otherwise validate_block_quant_shapes raises
+        # on output_partition_size=8 vs block_n=128 at tp=4).
+        prefix = "model.layers.0.linear_attn.in_proj_ba"
+        ignored = _qwen3_next_ignored_layers(0, "in_proj_ba")
+        self.assertTrue(is_layer_skipped(prefix, ignored, QWEN3_NEXT_FUSED_MAPPING))
+
+    def test_qwen3_next_in_proj_qkvz_is_skipped(self):
+        prefix = "model.layers.5.linear_attn.in_proj_qkvz"
+        ignored = _qwen3_next_ignored_layers(5, "in_proj_qkvz")
+        self.assertTrue(is_layer_skipped(prefix, ignored, QWEN3_NEXT_FUSED_MAPPING))
+
+    def test_mlp_gate_does_not_match_gate_up_proj(self):
+        # The motivation for #23467: an entry "mlp.gate" in
+        # modules_to_not_convert must NOT skip a sibling "mlp.gate_up_proj".
+        ignored = ["mlp.gate"]
+        self.assertFalse(
+            is_layer_skipped("model.layers.0.mlp.gate_up_proj", ignored, {})
+        )
+        self.assertTrue(is_layer_skipped("model.layers.0.mlp.gate", ignored, {}))
+
+
+if __name__ == "__main__":
+    unittest.main()