[Kernel] Drop the vendored dense BF16 GEMM port in favor of FlashInfer 0.6.18 (#38124)

Co-authored-by: Mohammad Angkad <mohammad.angkad@radixark.ai>

test/registered/unit/layers/quantization/test_bf16_splitk_gemm.py

@@ -2,25 +2,25 @@ import pytest
 
 from sglang.srt.environ import envs
 from sglang.srt.layers.quantization.unquant import (
-    _FLASHINFER_PR4266_TUNED_TACTICS,
+    _BF16_SPLITK_TUNED_TACTICS,
     Bf16GemmBackend,
     should_enable_bf16_splitk_gemm,
-    use_flashinfer_pr4266_bf16_gemm,
+    use_bf16_splitk_gemm,
 )
 from sglang.test.ci.ci_register import register_cpu_ci
 
 register_cpu_ci(est_time=11, suite="base-a-test-cpu")
 
 
-@pytest.mark.parametrize("m,n,k", _FLASHINFER_PR4266_TUNED_TACTICS)
-def test_flashinfer_pr4266_selects_tuned_oakhaven_shape(m: int, n: int, k: int):
-    assert use_flashinfer_pr4266_bf16_gemm(m, n, k)
+@pytest.mark.parametrize("m,n,k", _BF16_SPLITK_TUNED_TACTICS)
+def test_splitk_selects_tuned_oakhaven_shape(m: int, n: int, k: int):
+    assert use_bf16_splitk_gemm(m, n, k)
 
 
 @pytest.mark.parametrize("m", [0, 33, 64])
 @pytest.mark.parametrize("n,k", [(256, 8192), (512, 8192), (2304, 8192), (2560, 8192)])
-def test_flashinfer_pr4266_keeps_large_m_on_existing_path(m: int, n: int, k: int):
-    assert not use_flashinfer_pr4266_bf16_gemm(m, n, k)
+def test_splitk_keeps_large_m_on_existing_path(m: int, n: int, k: int):
+    assert not use_bf16_splitk_gemm(m, n, k)
 
 
 @pytest.mark.parametrize(
@@ -32,12 +32,8 @@ def test_flashinfer_pr4266_keeps_large_m_on_existing_path(m: int, n: int, k: int
         (32, 4096, 8192),
     ],
 )
-def test_flashinfer_pr4266_rejects_unmeasured_shapes(shape: tuple[int, int, int]):
-    assert not use_flashinfer_pr4266_bf16_gemm(*shape)
-
-
-def test_flashinfer_pr4266_backend_is_explicit():
-    assert Bf16GemmBackend.FLASHINFER_PR4266.value == "flashinfer_pr4266"
+def test_splitk_rejects_unmeasured_shapes(shape: tuple[int, int, int]):
+    assert not use_bf16_splitk_gemm(*shape)
 
 
 def test_bf16_splitk_is_enabled_by_default():

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

@@ -168,15 +168,11 @@ class TestApplyWithAddend(CustomTestCase):
                 )
             elif route == "splitk":
                 enter(patch.object(unquant, "_enable_bf16_splitk_gemm", True))
-                enter(
-                    patch.object(
-                        unquant, "use_flashinfer_pr4266_bf16_gemm", lambda *a: True
-                    )
-                )
+                enter(patch.object(unquant, "use_bf16_splitk_gemm", lambda *a: True))
                 enter(
                     patch.object(
                         unquant,
-                        "_flashinfer_pr4266_bf16_gemm",
+                        "_bf16_splitk_gemm",
                         _fake_kernel(kernel_calls, route),
                     )
                 )