Implement online nvfp4 quantization (#26083)

test/registered/backends/test_flashinfer_trtllm_gen_moe_backend.py

@@ -12,7 +12,7 @@ from sglang.test.test_utils import (
     popen_launch_server,
 )
 
-register_cuda_ci(est_time=600, suite="nightly-4-gpu-b200", nightly=True)
+register_cuda_ci(est_time=800, suite="nightly-4-gpu-b200", nightly=True)
 
 
 class FlashinferTrtllmGenMoeBackendFP8Base:
@@ -243,6 +243,58 @@ class FlashinferTrtllmGenMoeBackendNVFP4Base:
         self.assertGreater(metrics["score"], 0.89)
 
 
+class FlashinferTrtllmGenMoeBackendNvFp4OnlineBase:
+    backend = None
+    extra_env = {}
+
+    @classmethod
+    def setUpClass(cls):
+        cls.model = "Qwen/Qwen3-Next-80B-A3B-Instruct-FP8"
+        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, **cls.extra_env, "SGLANG_ENABLE_JIT_DEEPGEMM": "False"},
+            other_args=[
+                "--attention-backend",
+                "triton",
+                "--moe-runner-backend",
+                cls.backend,
+                "--cuda-graph-max-bs",
+                "128",
+                "--tp-size",
+                "4",
+                "--ep-size",
+                "2",
+                "--quantization",
+                "nvfp4_online",
+                "--mem-fraction-static",
+                "0.7",
+                "--mamba-ssm-dtype",
+                "bfloat16",
+            ],
+        )
+
+    @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"], 0.90)
+
+
 class TestFlashinferTrtllmGenMoeBackendFP8(
     FlashinferTrtllmGenMoeBackendFP8Base, CustomTestCase
 ):
@@ -273,12 +325,28 @@ class TestFlashinferTrtllmGenMoeBackendBF16Routed(
     backend = "flashinfer_trtllm_routed"
 
 
-class TestFlashinferTrtllmGenMoeBackendPerTokenNVFP4Routed(
+class TestFlashinferTrtllmGenMoeBackendNvFp4PerTokenActivationRouted(
     FlashinferTrtllmGenMoeBackendNVFP4Base, CustomTestCase
 ):
     extra_env = {"SGLANG_FLASHINFER_NVFP4_PER_TOKEN_ACTIVATION": "1"}
     backend = "flashinfer_trtllm_routed"
 
 
+class TestFlashinferTrtllmGenMoeBackendNvFp4Online(
+    FlashinferTrtllmGenMoeBackendNvFp4OnlineBase, CustomTestCase
+):
+    extra_env = {
+        "FLASHINFER_NVFP4_4OVER6": "1",
+        "FLASHINFER_NVFP4_4OVER6_ERR_MODE": "MSE",
+        "FLASHINFER_NVFP4_4OVER6_ERR_USE_FAST_MATH": "1",
+        "FLASHINFER_NVFP4_4OVER6_E4M3_USE_256": "1",
+        "SGLANG_FP4_IGNORED_LAYERS": ",".join(
+            ["shared_expert"]
+            + [f"model.layers.{layer_id}" for layer_id in range(40, 48)]
+        ),
+    }
+    backend = "flashinfer_trtllm"
+
+
 if __name__ == "__main__":
     unittest.main()