scan and apply_layers: milestone 1

This commit adds the lowering of scan to HLO While op. It also
introduce apply_layers which can sequentially apply a bunch of layers
using scan underneath.

In this milestone we use AOTAutograd to obtain the backward of the
function being scanned. Users can either save the activations in
fn or recompute them by passing different graph partitioners to
AOTAutograd.

ALso give the lowered fn computation a more meaningful name

Author: tengyifei

examples/decoder_only_model.py

@@ -7,16 +7,16 @@
 from torch import nn
 
 
-# the default config is intentionally kept low to make it runable on a sigle tpu v2-8 core.
+# the default config is intentionally kept low to make it runnable on a single tpu v2-8 core.
 @dataclass
 class DecoderOnlyConfig:
   hidden_size: int = 1024
   num_hidden_layers: int = 2
   num_attention_heads: int = 8
   num_key_value_heads: int = 4
-  intermediate_size = 32 * 1024
-  vocab_size = 3200
-  use_flash_attention = False
+  intermediate_size: int = 32 * 1024
+  vocab_size: int = 3200
+  use_flash_attention: bool = False
 
 
 def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:

test/run_tests.sh

@@ -208,7 +208,8 @@ function run_xla_op_tests1 {
 function run_xla_op_tests2 {
   run_test "$CDIR/pjrt/test_dtypes.py"
   run_test "$CDIR/test_while_loop.py"
-  run_test "$CDIR/test_scan.py"
+  run_test "$CDIR/scan/test_scan.py"
+  run_test "$CDIR/scan/test_scan_layers.py"
   run_test "$CDIR/test_autocast.py"
   run_test "$CDIR/eager/test_eager.py"
   run_test "$CDIR/eager/test_eager_with_xla_compile.py"