scan and apply_layers #7901

tengyifei · 2024-08-23T00:57:04Z

Add the lowering of scan to HLO While op.

Introduce apply_layers which can sequentially apply a bunch of layers
using scan underneath.

Beef up unit tests including linear layers and decoders.

JackCaoG · 2024-08-23T16:57:02Z

======================================================================
ERROR: test_decoder_model (__main__.ApplyLayersTest)
----------------------------------------------------------------------
Traceback (most recent call last):
  File "/__w/xla/xla/pytorch/xla/test/test_apply_layers.py", line 77, in test_decoder_model
    from decoder_only_model import DecoderOnlyConfig, DecoderOnlyModel  # type:ignore
ModuleNotFoundError: No module named 'decoder_only_model'

you can't just import it, you need to setup import dir correctly. Take a look at https://github.com/pytorch/xla/blob/master/test/dynamo/test_dynamo_dynamic_shape.py#L1-L6

tengyifei · 2024-08-23T18:37:08Z

@JackCaoG ty. i followed your example and got it working.

test/test_apply_layers.py

test/test_operations.py

JackCaoG · 2024-08-23T20:00:18Z

test/test_operations.py

+    import json
+    hlo_json = json.loads(ctx.hlo_json())
+    num_parameters = len(hlo_json["hostProgramShape"]["parameters"])
+    self.assertEqual(len(mapping), num_parameters)


so you expect both value to be 10?

Unfortunately not. It looks like some integer values (e.g. values <= 2) are shared when you put multiple copies into the HLO, but values above 2 are not shared. So we don't necessarily get 10. In any case, the precise number of parameters seems to be an implementation detail that we can't reliably test.

test/test_scan.py

JackCaoG · 2024-08-23T20:03:08Z

torch_xla/csrc/init_python_bindings.cpp

@@ -1077,7 +1076,9 @@ class PyLoweringContext {
      at::ScalarType dtype =
          MaybeUpcastToHostTorchType(literal.shape().element_type());
      at::Tensor input = MakeTensorFromXlaLiteral(literal, dtype);
-      results[param_ids[i]] = input;
+      std::optional param_id = lowering_ctx.GetParameterId(device_data[i]);
+      XLA_CHECK(param_id.has_value());


when would it not has value?

When GetParameterId receives a BackendData that is not a parameter in this lowering context, it will return std::nullopt. However, this loop is only iterating over parameters (line 1071, const std::vector<torch::lazy::BackendDataPtr>& device_data = lowering_ctx.GetParametersData();), so we will expect all BackendData there to have an ID. Seems good to enforce this invariant.

torch_xla/experimental/apply_layers.py

JackCaoG · 2024-08-23T20:07:09Z

torch_xla/experimental/apply_layers.py

+    return input_data
+
+  # Extract and stack the parameters into a pytree.
+  params = [_extract_weights_dict(layer) for layer in layers]


what if it is a dropout layer that parameters are more than just tensors?

If there is a dropout layer that references tensors other than model parameters (for example, the dropout probability), then those tensors will be captured as an additional HLO parameter to the XlaComputation object. As implemented now, apply_layers and scan will trace the first layer, and then use the same captured tensor for subsequent layers. This will be a problem if the user passes different dropout probabilities for say a sequence of dropout layers -- we'll instead incorrectly just keep using the first dropout's probability. I'll have to dig deeper and find a solution for this.

If there's a layer that references things other than tensors, then either that thing (e.g. a bool field) will impact the traced HLO computation, in which case I need to add a verification that all layers trace to equivalent computations. Or that thing won't impact the traced computation, in which case it won't matter to us.

JackCaoG · 2024-08-23T20:08:21Z

torch_xla/experimental/apply_layers.py

+  example_layer = deepcopy(next(iter(layers)))
+
+  # Hollow out the weights and biases in the example layer.
+  example_layer = example_layer.to_empty(device=None)


is this not going to impact the cloned arg?

Could you clarify this question -- I thought to_empty is going to destroy the value inside example_layer, so I deepcopy it before to backup.

torch_xla/experimental/scan.py

JackCaoG · 2024-08-26T22:34:13Z

torch_xla/experimental/scan.py

+  fn_output_carry_pytree, fn_output_y_pytree = flat_fn(*(fake_carry + fake_x))
+
+  # Later we'll use `fn_output_carry_spec` etc to turn flattened outputs back to a PyTree.
+  fn_output_carry, fn_output_carry_spec = tree_flatten(fn_output_carry_pytree)
+  assert fn_output_carry_spec == carry_spec
+  fn_output_y, fn_output_y_spec = tree_flatten(fn_output_y_pytree)
+  flat_y_len = len(fn_output_y)
+  fn_outputs = fn_output_carry + fn_output_y


what if there are in place updates to the tensor but it is not being return from the function?

I tested this and we give the wrong answer: https://github.com/tengyifei/playground/blob/master/scan_with_in_place_updates.ipynb

In the notebook, I wrote an approach to detect and prevent in place updates like that. TLDR is we'll have to trace every forward of each layer and verify that they're the same.

In the latest commit, this will now fail with an assertion error instead of silently giving wrong results.

torch_xla/experimental/scan.py

JackCaoG · 2024-08-26T22:39:13Z

torch_xla/experimental/scan.py

+
+    def step_fn(grad_carry, pytree: Tuple[torch.Tensor, torch.Tensor,
+                                          torch.Tensor]):
+      grad_y, carry, x = pytree


is this a typo?

I don't think so -- pytree is a tuple of the output grad at current step (grad_y), carry at the current step (carry), and input at current step (x)

torch_xla/experimental/scan.py

miladm · 2024-09-20T00:06:34Z

@tengyifei is this PR a 2.5 candidate?

tengyifei · 2024-09-20T05:47:29Z

@miladm yes, I'd like to backport this to 2.5 after addressing the comments etc.

torch_xla/experimental/scan.py

torch_xla/experimental/apply_layers.py

torch_xla/experimental/scan.py

Add the lowering of scan to HLO While op. Introduce apply_layers which can sequentially apply a bunch of layers using scan underneath. Beef up unit tests including linear layers and decoders. add regression test for parameter_id_tensor_mapping add test_apply_layers.py to test shell scripts correctly import decoder model from examples

tengyifei requested review from JackCaoG and alanwaketan August 23, 2024 18:37