diff --git a/tinygrad/runtime/graph/hcq.py b/tinygrad/runtime/graph/hcq.py index be51643b49..671a12d567 100644 --- a/tinygrad/runtime/graph/hcq.py +++ b/tinygrad/runtime/graph/hcq.py @@ -107,11 +107,11 @@ class HCQGraph(MultiGraphRunner): self.copy_to_devs[Device[dest.device]].add(Device[src.device]) self.op_cmd_idx[j] = (enqueue_queue, len(enqueue_queue) - 1) - if signal_val is not None: enqueue_queue.signal(signal, signal_val) - # Encode finish profile timestamp (if needed). if prof_info: enqueue_queue.timestamp(prof_info[1]) + if signal_val is not None: enqueue_queue.signal(signal, signal_val) + for dev in self.devices: for dep_dev in list(self.copy_to_devs[dev]) + [dev]: if (last_j:=self.last_ji[self.copy_queues[dep_dev]]) is None: continue