[quantization] AWQ Marlin doesn't work when dtype is bfloat16 (#11494)

Signed-off-by: Kai-Hsun Chen <khchen@x.ai>
Co-authored-by: Xinyuan Tong <115166877+JustinTong0323@users.noreply.github.com>

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

@@ -840,12 +840,9 @@ class AWQMoEMethod(FusedMoEMethodBase):
             self.moe_runner_config.activation == "silu"
         ), "Only SiLU activation is supported."
 
-        # The input must currently be float16
         x = dispatch_output.hidden_states
         topk_output = dispatch_output.topk_output
-
         orig_dtype = x.dtype
-        x = x.half()
 
         topk_weights, topk_ids, router_logits = topk_output
 

python/sglang/srt/model_executor/model_runner.py

@@ -527,7 +527,7 @@ class ModelRunner:
                 quantization_config := getattr(
                     self.model_config.hf_config, "quantization_config", None
                 )
-            ) is not None:
+            ) is not None and "weight_block_size" in quantization_config:
                 weight_block_size_n = quantization_config["weight_block_size"][0]
 
                 if self.tp_size % self.moe_ep_size != 0:

sgl-kernel/python/sgl_kernel/fused_moe.py

@@ -80,6 +80,12 @@ def fused_marlin_moe(
     assert w1.is_contiguous(), "Expert weights1 must be contiguous"
     assert w2.is_contiguous(), "Expert weights2 must be contiguous"
     assert hidden_states.dtype in [torch.float16, torch.bfloat16]
+    assert (
+        hidden_states.dtype == w1_scale.dtype
+    ), f"moe_wna16_marlin_gemm assumes hidden_states.dtype ({hidden_states.dtype}) == w1_scale.dtype ({w1_scale.dtype})"
+    assert (
+        hidden_states.dtype == w2_scale.dtype
+    ), f"moe_wna16_marlin_gemm assumes hidden_states.dtype ({hidden_states.dtype}) == w2_scale.dtype ({w2_scale.dtype})"
     assert num_bits in [4, 8]
 
     M, K = hidden_states.shape

test/srt/quant/test_awq.py

@@ -1,6 +1,8 @@
 import unittest
 from types import SimpleNamespace
 
+import requests
+
 from sglang.srt.utils import kill_process_tree
 from sglang.test.run_eval import run_eval
 from sglang.test.test_utils import (
@@ -41,5 +43,38 @@ class TestAWQ(CustomTestCase):
         self.assertGreater(metrics["score"], 0.64)
 
 
+class TestAWQMarlinBfloat16(CustomTestCase):
+    """
+    Verify that the model can be loaded with bfloat16 dtype and awq_marlin quantization
+    """
+
+    @classmethod
+    def setUpClass(cls):
+        cls.model = "QuantTrio/Qwen3-VL-30B-A3B-Instruct-AWQ"
+        cls.base_url = DEFAULT_URL_FOR_TEST
+        cls.process = popen_launch_server(
+            cls.model,
+            cls.base_url,
+            timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
+            other_args=["--dtype", "bfloat16", "--quantization", "awq_marlin"],
+        )
+
+    @classmethod
+    def tearDownClass(cls):
+        kill_process_tree(cls.process.pid)
+
+    def test_mmlu(self):
+        args = SimpleNamespace(
+            base_url=self.base_url,
+            model=self.model,
+            eval_name="mmlu",
+            num_examples=64,
+            num_threads=32,
+        )
+
+        metrics = run_eval(args)
+        self.assertGreater(metrics["score"], 0.88)
+
+
 if __name__ == "__main__":
     unittest.main()