design doc

src/client/TODO_design.py

@@ -33,5 +33,15 @@
 #         # [IMPORTANT] maybe first create an op for one batch, then a wrapper that split into batches
 #         return torch.cat(output_batches, dim=0)
 #
-#    def backward(ctx, grad_outputs):
-#         return TODO(ctx, )
+#    def backward(ctx, *grad_outputs: torch.Tensor):
+#         return TODO_think_through(ctx, *grad_outputs)
+
+# class RemoteSequentialInferenceSession:
+#    """An interface to a multi-step *inference* session for a sequence of remote transformer blocks"""
+#    # form sequence, maintain past inputs for all stages; if any stage breaks, fix the same way as in forward:
+#    # patch = RemoteSequenceManager[spans[BROKEN_INDEX].start : spans[BROKEN_INDEX].end].form_sequence()
+#    # feed_past_key_values(into=patch, inputs=past_key_values[BROKEN_INDEX])
+#    # self.TODO_replace_in_metadata(remove=BROKEN_INDEX, replace_with=patch)
+#    #
+#    TODO_think()
+#