[KERNELS] no longer enforce persistent when is used (#7214)

Author: ptillet

python/triton_kernels/tests/test_matmul.py

@@ -448,9 +448,8 @@ def test_set_idle_sms():
     from triton_kernels.matmul_ogs_details.opt_flags import make_opt_flags
     num_idle_sms = 24
     matmul_ogs_set_idle_sms(num_idle_sms)
-    flags = make_opt_flags(torch.float32, torch.float32, torch.float32, PrecisionConfig(), 0, 0, 0, None, False, False,
-                           1)
-    assert flags.is_persistent
+    flags = make_opt_flags(torch.float32, torch.float32, torch.float32, PrecisionConfig(), \
+                           1024, 1024, 1024, None, True, False, 1)
     assert flags.idle_sms == num_idle_sms
 
 

python/triton_kernels/triton_kernels/matmul_ogs.py

@@ -614,7 +614,10 @@ def _create_tma_descriptors(
     return x_tensor_or_desc, w_desc_and_transpose, mx_desc_and_transpose
 
 def matmul_ogs_set_idle_sms(num_idle_sms):
-    update_opt_flags_constraints({"is_persistent": True, "idle_sms": num_idle_sms})
+    """
+    persistent kernels will leave `num_idle_sms` idle
+    """
+    update_opt_flags_constraints({"idle_sms": num_idle_sms})
 
 def matmul_ogs(x, w, bias,
                routing_data: RoutingData | None = None,