[FlashInfer v0.6.18] add FlashInfer CuTe DSL NVFP4 W4A16 mode (#35120)

test/registered/backends/test_flashinfer_nvfp4_online_moe_backend.py

@@ -14,7 +14,7 @@ from sglang.test.test_utils import (
     popen_launch_server,
 )
 
-register_cuda_ci(est_time=800, stage="nightly", runner_config="4-gpu-b200")
+register_cuda_ci(est_time=1200, stage="nightly", runner_config="4-gpu-b200")
 
 
 class FlashinferNvFp4OnlineMoeBackendBase:
@@ -124,5 +124,47 @@ class TestFlashinferCuteDSLMoeBackendNvFp4Online(
     }
 
 
+class TestFlashinferCuteDSLMoeBackendNvFp4OnlineW4A16(
+    FlashinferNvFp4OnlineMoeBackendBase, CustomTestCase
+):
+    backend = "flashinfer_cutedsl"
+    model = "nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-FP8"
+    extra_args = [
+        "--reasoning-parser",
+        "nemotron_3",
+        "--tool-call-parser",
+        "qwen3_coder",
+        "--speculative-algorithm",
+        "EAGLE",
+        "--speculative-num-steps",
+        "3",
+        "--speculative-eagle-topk",
+        "1",
+        "--speculative-num-draft-tokens",
+        "4",
+        "--dp-size",
+        "4",
+        "--enable-dp-attention",
+        "--enable-dp-lm-head",
+        "--moe-a2a-backend",
+        "flashinfer",
+        "--cuda-graph-backend-prefill",
+        "disabled",
+    ]
+    eval_args = {"max_tokens": 16000, "temperature": 1.0, "top_p": 0.95}
+    spec_accept_length_threshold = 2.5
+    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_FLASHINFER_CUTEDSL_NVFP4_W4A16": "1",
+        "SGLANG_FLASHINFER_NVFP4_PER_TOKEN_ACTIVATION": "0",
+        "SGLANG_FLASHINFER_MOE_FUSED_FINALIZE": "1",
+        "SGLANG_MOE_NVFP4_DISPATCH": "0",
+        "SGLANG_FLASHINFER_NUM_MAX_DISPATCH_TOKENS_PER_RANK": "4096",
+    }
+
+
 if __name__ == "__main__":
     unittest.main()

test/registered/rl/test_update_weights_from_disk_blackwell.py

@@ -1,6 +1,6 @@
 from sglang.test.ci.ci_register import register_cuda_ci
 
-register_cuda_ci(est_time=320, stage="extra-b", runner_config="4-gpu-b200")
+register_cuda_ci(est_time=420, stage="extra-b", runner_config="4-gpu-b200")
 
 import time
 import unittest
@@ -269,5 +269,37 @@ class TestServerUpdateWeightsFromDiskNVFP4CuteDSL(
     )
 
 
+class TestServerUpdateWeightsFromDiskNVFP4W4A16CuteDSL(
+    UpdateWeightsFromDiskBase, CustomTestCase
+):
+    model = "nvidia/Qwen3-30B-A3B-NVFP4"
+    decode_payload = {**UpdateWeightsFromDiskBase.decode_payload, "routed_dp_rank": 0}
+    launch_env = {
+        "SGLANG_FLASHINFER_CUTEDSL_NVFP4_W4A16": "1",
+        "SGLANG_FLASHINFER_NVFP4_PER_TOKEN_ACTIVATION": "0",
+        "SGLANG_MOE_NVFP4_DISPATCH": "0",
+        "SGLANG_FLASHINFER_NUM_MAX_DISPATCH_TOKENS_PER_RANK": "4096",
+    }
+    backend_test_suites = (
+        {
+            "name": "flashinfer_cutedsl_nvfp4_w4a16",
+            "other_args": (
+                "--tp-size",
+                "4",
+                "--dp-size",
+                "4",
+                "--enable-dp-attention",
+                "--ep-size",
+                "4",
+                "--fp4-gemm-backend",
+                "flashinfer_cutedsl",
+                "--moe-runner-backend",
+                "flashinfer_cutedsl",
+                "--enable-deterministic-inference",
+            ),
+        },
+    )
+
+
 if __name__ == "__main__":
     unittest.main()

test/registered/unit/layers/moe/test_flashinfer_cutedsl_dispatch.py

@@ -31,6 +31,7 @@ def test_flashinfer_prefill_returns_standard_combine_input():
     wrapper.run.return_value = expected_output
     quant_info = SimpleNamespace(
         wrapper=wrapper,
+        quant_mode="w4a4",
         use_per_token_activation=False,
         a1_scale=torch.tensor(1.0),
         a2_scale=torch.tensor(1.0),

test/registered/unit/model_loader/test_modelopt_loader.py

@@ -962,6 +962,20 @@ class TestModelOptMixedPrecisionConfig(CustomTestCase):
             )
         )
 
+    def test_lm_head_guard_accepts_modelopt_fp4_cutedsl_w4a16_runtime_state(self):
+        lm_head = nn.Module()
+        lm_head.weight = nn.Parameter(
+            torch.empty(128, 1024, dtype=torch.uint8), requires_grad=False
+        )
+        lm_head.weight_scale_interleaved = nn.Parameter(torch.empty(1))
+        lm_head.alpha = nn.Parameter(torch.empty(1))
+        lm_head.input_size_per_partition = 2048
+        lm_head.output_size_per_partition = 128
+        quant_method = ModelOptFp4LinearMethod(ModelOptFp4Config())
+        quant_method.quant_mode = "w4a16"
+
+        self.assertTrue(should_apply_lm_head_quant_method(lm_head, quant_method))
+
     def test_lm_head_guard_rejects_stale_modelopt_fp4_method_on_dense_head(self):
         lm_head = nn.Module()
         lm_head.weight = nn.Parameter(torch.empty(128000, 2048))