Detach all grads during backprop. (#1243)

* Detach all grads during backprop.

* Add an environment variable to select the backprop behavior.

* Update the comment.

candle-core/src/backprop.rs

@@ -15,6 +15,17 @@ fn broadcast_back(arg: &Tensor, node: &Tensor, reduced_dims: &[usize]) -> Result
     }
 }
 
+thread_local! {
+    static CANDLE_GRAD_DO_NOT_DETACH: bool = {
+        match std::env::var("CANDLE_GRAD_DO_NOT_DETACH") {
+            Ok(s) => {
+                !s.is_empty() && s != "0"
+            },
+            Err(_) => false,
+        }
+    }
+}
+
 impl Tensor {
     /// Return all the nodes that lead to this value in a topologically sorted vec, the first
     /// elements having dependencies on the latter ones, e.g. the first element if any is the
@@ -155,10 +166,16 @@ impl Tensor {
             if node.is_variable() {
                 continue;
             }
-            let grad = grads.remove(node).unwrap();
+            let grad = grads
-            // TODO: We should perform all these operations in place (or at least not track the
+                .remove(node)
-            // whole graph). The only drawback would be if we wanted to support grad of grad but
+                .expect("candle internal error - grad not populated");
-            // this is out of scope.
+            // https://github.com/huggingface/candle/issues/1241
+            // Ideally, we would make these operations in place where possible to ensure that we
+            // do not have to allocate too often. Here we just call `.detach` to avoid computing
+            // the backprop graph of the backprop itself. This would be an issue for second order
+            // derivatives but these are out of scope at the moment.
+            let do_not_detach = CANDLE_GRAD_DO_NOT_DETACH.with(|b| *b);
+            let grad = if do_not_detach { grad } else { grad.detach()? };
             if let Some(op) = node.op() {
                 match op {
                     Op::Binary(lhs, rhs, BinaryOp::Add) => {