[Fix] Fix gpt oss triton kernels and upgrade flashinfer back to 0.6.11.post1 (#25335)

Co-authored-by: sglang-bot <sglang-bot@users.noreply.github.com>
Co-authored-by: b8zhong <b8zhong@users.noreply.github.com>
Co-authored-by: mmangkad <mmangkad@users.noreply.github.com>

test/registered/dsv4/test_deepseek_v4_flash_fp4_h200.py

@@ -91,7 +91,6 @@ class TestDSV4FlashFP4H200(ServerSanityMixin, CustomTestCase):
         self.assertGreater(metrics["score"], 0.93)
 
 
-@unittest.skip("broken on main, see #24816")
 @unittest.skipUnless(
     _flashinfer_has_sm90_cutlass_mxfp4(),
     "FlashInfer build lacks SM90 mixed-input MXFP4 helpers (PR #3084, >= 0.6.11)",

test/registered/moe/test_cutedsl_moe.py

@@ -899,13 +899,14 @@ class TestCuteDslV1(unittest.TestCase):
                         masked_m.to(hidden_states.device),
                     )
 
+                    a_global_scale = input_global_scale[:1]
                     a_fp4, a_scale_interleaved = fp4_quantize(
-                        hidden_states, input_global_scale
+                        hidden_states, a_global_scale
                     )
                     a_in_dtype = dequantize_nvfp4_to_dtype(
                         a_fp4,
                         a_scale_interleaved,
-                        input_global_scale,
+                        a_global_scale,
                         dtype=hidden_states.dtype,
                         device=hidden_states.device,
                         block_size=16,
@@ -1077,11 +1078,12 @@ class TestCuteDslV1(unittest.TestCase):
                 masked_m.to(device),
             )
 
-            a_fp4, a_scale_interleaved = fp4_quantize(hidden_states, input_global_scale)
+            a_global_scale = input_global_scale[:1]
+            a_fp4, a_scale_interleaved = fp4_quantize(hidden_states, a_global_scale)
             a_in_dtype = dequantize_nvfp4_to_dtype(
                 a_fp4,
                 a_scale_interleaved,
-                input_global_scale,
+                a_global_scale,
                 dtype=hidden_states.dtype,
                 device=device,
                 block_size=16,
@@ -1251,11 +1253,12 @@ class TestCuteDslV1(unittest.TestCase):
             )
 
             # PyTorch reference (same as the bf16 input test)
-            a_fp4, a_scale = fp4_quantize(hidden_states, input_gs)
+            a_gs = input_gs[:1]
+            a_fp4, a_scale = fp4_quantize(hidden_states, a_gs)
             a_deq = dequantize_nvfp4_to_dtype(
                 a_fp4,
                 a_scale,
-                input_gs,
+                a_gs,
                 dtype=torch.bfloat16,
                 device=device,
                 block_size=16,

test/registered/unit/layers/quantization/test_mxfp4_sm90_cutlass.py

@@ -20,12 +20,7 @@ import torch
 
 from sglang.test.ci.ci_register import register_cuda_ci
 
-register_cuda_ci(
-    est_time=120,
-    stage="stage-b",
-    runner_config="1-gpu-large",
-    disabled="broken on main, see #24816",
-)
+register_cuda_ci(est_time=120, stage="stage-b", runner_config="1-gpu-large")
 
 flashinfer_fused_moe = pytest.importorskip("flashinfer.fused_moe")