sync flashinfer and update sgl-kernel tests (#3081)

sgl-kernel/tests/test_activation.py

@@ -35,4 +35,5 @@ def test_fused_gelu_mul(dim, batch_size, seq_len):
     torch.testing.assert_close(y_ref, y, rtol=1e-3, atol=1e-3)
 
 
-test_fused_silu_mul(128, 1, 1)
+if __name__ == "__main__":
+    pytest.main([__file__])

sgl-kernel/tests/test_lightning_attention_decode.py

@@ -82,3 +82,7 @@ def test_lightning_attention_decode(dtype, batch_size, num_heads, dim, embed_dim
         msg=f"New KV mismatch for batch_size={batch_size}, num_heads={num_heads}, "
         f"dim={dim}, embed_dim={embed_dim}, dtype={dtype}",
     )
+
+
+if __name__ == "__main__":
+    pytest.main([__file__])

sgl-kernel/tests/test_norm.py

@@ -127,3 +127,7 @@ def test_gemma_fused_add_rmsnorm(batch_size, hidden_size, dtype):
 
     torch.testing.assert_close(x_fused, x_native, rtol=1e-3, atol=1e-3)
     torch.testing.assert_close(residual_fused, residual_native, rtol=1e-3, atol=1e-3)
+
+
+if __name__ == "__main__":
+    pytest.main([__file__])