[Fix] Reject online weight updates while the HPC-Ops router GEMM split cache is active (#31943)

Co-authored-by: Halcyon <56064364+VAthree@users.noreply.github.com>
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>

test/registered/gemm/test_linear_bf16_fp32_hpc.py

@@ -13,7 +13,9 @@ import torch
 from sglang.kernels.ops.attention.dsv4.gemm import (
     _hpc_gemm_bf16xfp32_available,
     _linear_bf16_fp32_hpc,
+    hpc_bf16xfp32_gemm_enabled,
     linear_bf16_fp32,
+    mark_hpc_bf16xfp32_gemm_enabled,
 )
 from sglang.test.ci.ci_register import register_cuda_ci
 from sglang.test.test_utils import CustomTestCase
@@ -32,6 +34,7 @@ class TestLinearBf16Fp32Hpc(CustomTestCase):
 
     @classmethod
     def setUpClass(cls):
+        mark_hpc_bf16xfp32_gemm_enabled()
         torch.manual_seed(0)
 
     def test_matches_fp32_reference(self):
@@ -70,6 +73,48 @@ class TestLinearBf16Fp32Hpc(CustomTestCase):
         # The kernel leaves the cached split-K workspace zeroed.
         self.assertTrue((cache[3] == 0).all().item())
 
+    def test_online_weight_updates_rejected_when_enabled(self):
+        from sglang.srt.model_executor.model_runner_components.weight_updater import (
+            _unsupported_derived_weight_cache_error,
+        )
+
+        self.assertTrue(hpc_bf16xfp32_gemm_enabled())
+        self.assertIsNotNone(_unsupported_derived_weight_cache_error())
+
+    def test_split_buffers_stable_for_cuda_graph(self):
+        # Captured graphs replay the split buffers by address; an in-place
+        # weight write must never reallocate them.
+        k, n = _ROUTER_SHAPES[1]
+        x = torch.randn(16, k, dtype=torch.bfloat16, device="cuda")
+        w = torch.randn(n, k, dtype=torch.float32, device="cuda")
+        w_orig = w.clone()
+        _linear_bf16_fp32_hpc(x, w)  # populate the cache outside the graph
+        cache_before = getattr(w, "_sglang_bf16xfp32_weight_cache")
+
+        graph = torch.cuda.CUDAGraph()
+        stream = torch.cuda.Stream()
+        stream.wait_stream(torch.cuda.current_stream())
+        with torch.cuda.stream(stream):
+            _linear_bf16_fp32_hpc(x, w)  # warmup on the side stream
+        torch.cuda.current_stream().wait_stream(stream)
+        with torch.cuda.graph(graph):
+            out = _linear_bf16_fp32_hpc(x, w)
+
+        graph.replay()
+        torch.testing.assert_close(
+            out, torch.mm(x.float(), w_orig.t()), rtol=0.08, atol=0.01
+        )
+
+        w.data.copy_(torch.randn_like(w))
+        out_eager = _linear_bf16_fp32_hpc(x, w)
+        cache_after = getattr(w, "_sglang_bf16xfp32_weight_cache")
+        self.assertIs(cache_before, cache_after)
+        graph.replay()
+        torch.testing.assert_close(
+            out, torch.mm(x.float(), w_orig.t()), rtol=0.08, atol=0.01
+        )
+        torch.testing.assert_close(out_eager, out)
+
 
 if __name__ == "__main__":
     unittest.main()