[NVIDIA] Fix broken fp8 MoE of deepseek v3 (#13264)

Co-authored-by: Baizhou Zhang <sobereddiezhang@gmail.com>
Co-authored-by: Kangyan-Zhou <zky314343421@gmail.com>

test/srt/run_suite.py

@@ -176,6 +176,7 @@ suites = {
         # TODO: Add it back after the bug is fixed
         # TestFile("test_eagle_infer_beta_dp_attention.py", 200),
     ],
+    "per-commit-8-gpu-b200": [],
     "per-commit-4-gpu-gb200": [
         TestFile("test_cutedsl_moe.py", 300),
         TestFile("test_deepseek_v3_fp4_4gpu.py", 1800),
@@ -215,7 +216,9 @@ suites = {
         TestFile("test_deepseek_v3_fp4_cutlass_moe.py", 900),
         TestFile("test_fp4_moe.py", 300),
     ],
-    "nightly-8-gpu-b200": [],
+    "nightly-8-gpu-b200": [
+        TestFile("test_deepseek_r1_fp8_trtllm_backend.py", 3600),
+    ],
     "nightly-4-gpu": [
         TestFile("test_qwen3_next_deterministic.py", 200),
     ],

test/srt/test_deepseek_r1_fp8_trtllm_backend.py

@@ -0,0 +1,88 @@
+import os
+import unittest
+from types import SimpleNamespace
+
+from sglang.srt.utils import kill_process_tree
+from sglang.test.few_shot_gsm8k import run_eval as run_eval_few_shot_gsm8k
+from sglang.test.test_utils import (
+    DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
+    DEFAULT_URL_FOR_TEST,
+    CustomTestCase,
+    popen_launch_server,
+    try_cached_model,
+)
+
+FULL_DEEPSEEK_V3_MODEL_PATH = "deepseek-ai/DeepSeek-V3-0324"
+
+
+class TestDeepseekR1Fp8Flashinfer(CustomTestCase):
+    @classmethod
+    def setUpClass(cls):
+        cls.model = try_cached_model(FULL_DEEPSEEK_V3_MODEL_PATH)
+        cls.base_url = DEFAULT_URL_FOR_TEST
+        other_args = [
+            "--trust-remote-code",
+            "--disable-radix-cache",
+            "--max-running-requests",
+            "512",
+            "--chunked-prefill-size",
+            "8192",
+            "--mem-fraction-static",
+            "0.9",
+            "--cuda-graph-max-bs",
+            "128",
+            "--max-prefill-tokens",
+            "8192",
+            "--kv-cache-dtype",
+            "fp8_e4m3",
+            "--quantization",
+            "fp8",
+            "--tensor-parallel-size",
+            "8",
+            "--data-parallel-size",
+            "1",
+            "--expert-parallel-size",
+            "1",
+            "--scheduler-recv-interval",
+            "10",
+            "--stream-interval",
+            "10",
+            "--attention-backend",
+            "trtllm_mla",
+            "--moe-runner-backend",
+            "flashinfer_trtllm",
+            "--enable-symm-mem",
+        ]
+        cls.process = popen_launch_server(
+            cls.model,
+            cls.base_url,
+            timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
+            other_args=other_args,
+            env={
+                **os.environ,
+                "SGLANG_ENABLE_FLASHINFER_FP8_GEMM": "1",
+            },
+        )
+
+    @classmethod
+    def tearDownClass(cls):
+        kill_process_tree(cls.process.pid)
+
+    def test_gsm8k(self):
+        args = SimpleNamespace(
+            num_shots=5,
+            data_path=None,
+            num_questions=512,
+            parallel=512,
+            max_new_tokens=512,
+            host="http://127.0.0.1",
+            port=int(self.base_url.split(":")[-1]),
+        )
+        metrics = run_eval_few_shot_gsm8k(args)
+        print(f"Eval accuracy of GSM8K: {metrics=}")
+
+        self.assertGreater(metrics["accuracy"], 0.92)
+
+
+if __name__ == "__main__":
+    unittest.main()