[fix] Add support for flashinfer MOE A2A to Qwen3 BF16 model path (#26255)

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

@@ -446,6 +446,12 @@ def should_skip_post_experts_all_reduce(
       - ``should_use_flashinfer_cutlass_moe_fp4_allgather()`` (TP path only):
         the flashinfer cutlass FP4 kernel performs an all-gather that absorbs
         the post-experts TP all-reduce. Not relevant to the EP all-reduce.
+      - ``get_moe_a2a_backend().is_flashinfer()``: the flashinfer A2A
+        dispatcher's ``MoeAlltoAll.combine`` already alltoall-reduces partial
+        MoE outputs back to the source rank, so any further EP/TP all-reduce
+        would double-count and overflow BF16. Mirrors TRTLLM's
+        ``not enable_alltoall`` gate
+        (``tensorrt_llm/_torch/modules/fused_moe/interface.py:879``).
 
     The first two args are layer-context flags from ``LayerCommunicator`` and
     default to ``False`` for models that don't use it. Pass ``is_tp_path=True``
@@ -457,6 +463,8 @@ def should_skip_post_experts_all_reduce(
         return True
     if is_tp_path and should_use_flashinfer_cutlass_moe_fp4_allgather():
         return True
+    if get_moe_a2a_backend().is_flashinfer():
+        return True
     return False
 
 

sgl-kernel/csrc/moe/moe_topk_softmax_kernels.cu

@@ -753,6 +753,14 @@ void topk_softmax(
   const int num_tokens = static_cast<int>(gating_output.size(0));
   const int topk = static_cast<int>(topk_weights.size(-1));
 
+  // No tokens on this DP rank, no need to do anything
+  if (num_tokens == 0) {
+    return;
+  }
+
+  TORCH_CHECK(num_experts > 0, "num_experts must be greater than 0");
+  TORCH_CHECK(topk > 0, "topk must be greater than 0");
+
   const bool is_pow_2 = (num_experts != 0) && ((num_experts & (num_experts - 1)) == 0);
   const bool needs_workspace = !is_pow_2 || num_experts > 512;
   const int64_t workspace_size = needs_workspace ? num_tokens * num_experts : 0;
@@ -822,4 +830,7 @@ void topk_softmax(
   } else {
     TORCH_CHECK(false, "Unsupported gating_output dtype: ", dtype);
   }
+
+  auto launch_error = cudaGetLastError();
+  TORCH_CHECK(launch_error == cudaSuccess, "topk_softmax launch error: ", cudaGetErrorString(launch_error));
 }

test/registered/moe/test_flashinfer_a2a_cutlass.py

@@ -0,0 +1,93 @@
+"""Test FlashInfer Cutlass BF16 MoE + FlashInfer alltoall on B200 with DP attention.
+
+Config: Qwen3-30B-A3B, B200x4, EP=4 DP=4, flashinfer cutlass + flashinfer a2a.
+"""
+
+import os
+import unittest
+from types import SimpleNamespace
+
+import torch
+
+from sglang.srt.utils import kill_process_tree
+from sglang.test.ci.ci_register import register_cuda_ci
+from sglang.test.run_eval import run_eval
+from sglang.test.test_utils import (
+    DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
+    DEFAULT_URL_FOR_TEST,
+    CustomTestCase,
+    popen_launch_server,
+)
+
+register_cuda_ci(
+    est_time=600,
+    stage="extra-b",
+    runner_config="4-gpu-b200",
+    disabled="Waived until sgl-kernel fix is released",
+)
+
+MODEL = os.environ.get("QWEN3_30B_A3B_MODEL_PATH", "Qwen/Qwen3-30B-A3B")
+
+SKIP_TEST = torch.cuda.get_device_capability() < (10, 0)
+SKIP_REASON = "Requires Blackwell (B200, sm_100a) or above."
+
+
+@unittest.skipIf(SKIP_TEST, SKIP_REASON)
+class TestFlashinferCutlassFlashinferA2A(CustomTestCase):
+    """FlashInfer Cutlass BF16 MoE + FlashInfer one-sided alltoall + DP4 EP4 on B200."""
+
+    @classmethod
+    def setUpClass(cls):
+        cls.model = MODEL
+        cls.base_url = DEFAULT_URL_FOR_TEST
+        cls.process = popen_launch_server(
+            cls.model,
+            cls.base_url,
+            timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH * 3,
+            other_args=[
+                "--trust-remote-code",
+                "--tp",
+                "4",
+                "--ep-size",
+                "4",
+                "--dp",
+                "4",
+                "--enable-dp-attention",
+                "--enable-dp-lm-head",
+                "--moe-runner-backend",
+                "flashinfer_cutlass",
+                "--moe-a2a-backend",
+                "flashinfer",
+                "--max-prefill-tokens",
+                "4096",
+                "--disable-radix-cache",
+                "--disable-flashinfer-autotune",
+                "--watchdog-timeout",
+                "900",
+            ],
+        )
+
+    @classmethod
+    def tearDownClass(cls):
+        kill_process_tree(cls.process.pid)
+
+    def test_gsm8k(self):
+        args = SimpleNamespace(
+            base_url=self.base_url,
+            eval_name="gsm8k",
+            num_examples=1319,
+            max_tokens=10240,
+            repeat=1,
+            num_threads=1319,
+            num_shots=8,
+            temperature=0.6,
+            top_p=0.95,
+            top_k=20,
+        )
+        metrics = run_eval(args)
+        print(metrics)
+        self.assertGreater(metrics["score"], 0.90)
+
+
+if __name__ == "__main__":
+    unittest.main()