[RL] Fix FP8 skip matching for trailing-dot prefixes (#26287)

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

@@ -50,6 +50,8 @@ def _module_path_match(ignored: str, prefix: str) -> bool:
     # match `mlp.gate_up_proj`. Needed for quant configs (e.g. Qwen3.6-FP8)
     # whose `modules_to_not_convert` lists MoE-template names like `mlp.gate`
     # that collide with fused dense MLP names by plain substring.
+    ignored = ignored.rstrip(".")
+    prefix = prefix.rstrip(".")
     if ignored == prefix:
         return True
     if prefix.startswith(ignored + "."):

test/registered/backends/test_flashinfer_trtllm_gen_moe_backend.py

@@ -155,6 +155,50 @@ class FlashinferTrtllmGenMoeBackendMXFP8Base:
         self.assertGreater(metrics["score"], 0.93)
 
 
+class FlashinferTrtllmGenMoeBackendMXFP8MixedBF16Base:
+    backend = None
+
+    @classmethod
+    def setUpClass(cls):
+        cls.model = "zianglih/JoyAI-LLM-Flash-MXFP8-last-6-BF16"
+        cls.base_url = DEFAULT_URL_FOR_TEST
+        cls.process = popen_launch_server(
+            cls.model,
+            cls.base_url,
+            timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
+            env={**os.environ, "SGLANG_ENABLE_JIT_DEEPGEMM": "False"},
+            other_args=[
+                "--kv-cache-dtype",
+                "bf16",
+                "--fp8-gemm-backend",
+                "flashinfer_cutlass",
+                "--moe-runner-backend",
+                cls.backend,
+                "--tp-size",
+                "4",
+                "--trust-remote-code",
+            ],
+        )
+
+    @classmethod
+    def tearDownClass(cls):
+        kill_process_tree(cls.process.pid)
+
+    def test_gsm8k(self):
+        args = SimpleNamespace(
+            base_url=self.base_url,
+            model=self.model,
+            eval_name="gsm8k",
+            api="completion",
+            max_tokens=512,
+            num_examples=200,
+            num_threads=128,
+        )
+        metrics = run_eval(args)
+        print(f"{metrics=}")
+        self.assertGreater(metrics["score"], 0.92)
+
+
 class FlashinferTrtllmGenMoeBackendNVFP4Base:
     backend = None
     extra_env = {}
@@ -217,6 +261,12 @@ class TestFlashinferTrtllmGenMoeBackendMXFP8Routed(
     backend = "flashinfer_trtllm_routed"
 
 
+class TestFlashinferTrtllmRoutedMxfp8MixedBF16(
+    FlashinferTrtllmGenMoeBackendMXFP8MixedBF16Base, CustomTestCase
+):
+    backend = "flashinfer_trtllm_routed"
+
+
 class TestFlashinferTrtllmGenMoeBackendBF16Routed(
     FlashinferTrtllmGenMoeBackendBF16Base, CustomTestCase
 ):

test/registered/quant/test_is_layer_skipped.py

@@ -47,6 +47,17 @@ class TestIsLayerSkipped(CustomTestCase):
         )
         self.assertTrue(is_layer_skipped("model.layers.0.mlp.gate", ignored, {}))
 
+    def test_trailing_dot_prefix_matches_child_modules(self):
+        # Mixed-precision checkpoints may use a trailing-dot layer prefix to keep
+        # every module under the layer in higher precision.
+        ignored = ["model.layers.34."]
+        self.assertTrue(
+            is_layer_skipped("model.layers.34.mlp.experts.0.down_proj", ignored, {})
+        )
+        self.assertFalse(
+            is_layer_skipped("model.layers.340.mlp.experts.0.down_proj", ignored, {})
+        )
+
 
 if __name__ == "__main__":
     unittest.main()