Update on "aot autograd: dont allow symint outputs to get tangents in the bw graph"

bdhirsh · bdhirsh · commit e6858d605510 · 2023-03-07T22:58:22.000Z
Previously, if dynamic shapes were turned on and we had a forward graph that returns a symint, then we would generate a backward graph that takes in a tangent input for that symint fwd output. This causes problems for downstream - inductor will see an input that it expects to be a symint, but it gets a `None` from autograd.

Confirmed that this repro now passes:
```
benchmarks/dynamo/torchbench.py --devices cuda --inductor --dynamic-shapes --unspecialize-int --accuracy --training --only drq
```

cc anijain2305 




[ghstack-poisoned]
diff --git a/test/functorch/test_aotdispatch.py b/test/functorch/test_aotdispatch.py
@@ -2092,7 +2092,7 @@ def f(a, b, c, d):
         bw_graph = bw_graph_cell[0]
 
         self.assertEqual(get_num_ins_outs(fw_graph), (4, 13))
-        self.assertEqual(get_num_ins_outs(bw_graph), (12, 4))
+        self.assertEqual(get_num_ins_outs(bw_graph), (10, 4))
         _, fw_graph_out_nodes = get_ins_outs(fw_graph)
         self.assertEqual(
             # fw outputs include b.size() which expands to 2 symints,
