Re-support kwargs at run time (#929)

## Description

Implements #928 

Users want the first pipeline stage to accept kwargs if the original
program does.
This is controlled by the `_codegen` field of the graph as @angelayi
suggests, so we make a copy from the traced program to submod0.


## Feature/Issue validation/testing

Added kwargs in test_fwd.py.
Also changed a few HF examples to directly kwargs.

Author: kwen2501

examples/hf/pippy_albert.py

@@ -46,7 +46,6 @@ def run(args):
     # Input configs
     example_inputs = generate_inputs_for_model(
         model_class, albert, model_name, args.batch_size, args.device)
-    input_ids = example_inputs["input_ids"]
 
     # Annotate split points
     add_split_points(albert, args.world_size)
@@ -55,7 +54,8 @@ def run(args):
     albert_pipe = Pipe.from_tracing(
         albert,
         num_chunks=args.chunks,
-        example_args=(input_ids, ),
+        example_args=(),
+        example_kwargs=example_inputs,
     )
     nstages = len(list(albert_pipe.split_gm.children()))
     assert nstages == args.world_size, f"nstages = {nstages} nranks = {args.world_size}"
@@ -72,7 +72,7 @@ def run(args):
 
     # Run
     if args.rank == 0:
-        stage(input_ids)
+        stage(**example_inputs)
     elif args.rank == args.world_size - 1:
         out = stage()
     else:

examples/hf/pippy_bart.py

@@ -43,7 +43,6 @@ def run(args):
     # Input configs
     example_inputs = generate_inputs_for_model(
         model_class, bart, model_name, args.batch_size, args.device)
-    input_ids = example_inputs["input_ids"]
 
     # Annotate split points
     add_split_points(bart, args.world_size)
@@ -52,7 +51,8 @@ def run(args):
     bart_pipe = Pipe.from_tracing(
         bart,
         num_chunks=args.chunks,
-        example_args=(input_ids, ),
+        example_args=(),
+        example_kwargs=example_inputs,
     )
     nstages = len(list(bart_pipe.split_gm.children()))
     assert nstages == args.world_size, f"nstages = {nstages} nranks = {args.world_size}"
@@ -69,7 +69,7 @@ def run(args):
 
     # Run
     if args.rank == 0:
-        stage(input_ids)
+        stage(**example_inputs)
     elif args.rank == args.world_size - 1:
         out = stage()
     else:

examples/hf/pippy_bert.py

@@ -43,7 +43,6 @@ def run(args):
     # Input configs
     example_inputs = generate_inputs_for_model(
         model_class, bert, model_name, args.batch_size, args.device)
-    input_ids = example_inputs["input_ids"]
 
     # Annotate split points
     add_split_points(bert, args.world_size)
@@ -52,7 +51,8 @@ def run(args):
     bert_pipe = Pipe.from_tracing(
         bert,
         num_chunks=args.chunks,
-        example_args=(input_ids, ),
+        example_args=(),
+        example_kwargs=example_inputs,
     )
     nstages = len(list(bert_pipe.split_gm.children()))
     assert nstages == args.world_size, f"nstages = {nstages} nranks = {args.world_size}"
@@ -69,7 +69,7 @@ def run(args):
 
     # Run
     if args.rank == 0:
-        stage(input_ids)
+        stage(**example_inputs)
     elif args.rank == args.world_size - 1:
         out = stage()
     else:

examples/hf/pippy_camemBert.py

@@ -43,7 +43,6 @@ def run(args):
     # Input configs
     example_inputs = generate_inputs_for_model(
         model_class, camembert, model_name, args.batch_size, args.device)
-    input_ids = example_inputs["input_ids"]
 
     # Annotate split points
     add_split_points(camembert, args.world_size)
@@ -52,7 +51,8 @@ def run(args):
     camembert_pipe = Pipe.from_tracing(
         camembert,
         num_chunks=args.chunks,
-        example_args=(input_ids, ),
+        example_args=(),
+        example_kwargs=example_inputs,
     )
     nstages = len(list(camembert_pipe.split_gm.children()))
     assert nstages == args.world_size, f"nstages = {nstages} nranks = {args.world_size}"
@@ -69,7 +69,7 @@ def run(args):
 
     # Run
     if args.rank == 0:
-        stage(input_ids)
+        stage(**example_inputs)
     elif args.rank == args.world_size - 1:
         out = stage()
     else:

examples/hf/pippy_gpt2.py

@@ -52,7 +52,6 @@ def run(args):
     # Input configs
     example_inputs = generate_inputs_for_model(
         model_class, gpt2, model_name, args.batch_size, args.device)
-    input_ids = example_inputs["input_ids"]
 
     # Annotate split points
     add_split_points(gpt2, args.world_size)
@@ -61,7 +60,8 @@ def run(args):
     gpt2_pipe = Pipe.from_tracing(
         gpt2,
         num_chunks=args.chunks,
-        example_args=(input_ids, ),
+        example_args=(),
+        example_kwargs=example_inputs,
     )
     assert len(list(gpt2_pipe.split_gm.children())) == args.world_size
     if args.rank == 0:
@@ -77,7 +77,7 @@ def run(args):
 
     # Run
     if args.rank == 0:
-        stage(input_ids)
+        stage(**example_inputs)
     elif args.rank == args.world_size - 1:
         out = stage()
     else:

examples/hf/pippy_gptNeo.py

@@ -43,7 +43,6 @@ def run(args):
     # Input configs
     example_inputs = generate_inputs_for_model(
         model_class, gptneo, model_name, args.batch_size, args.device)
-    input_ids = example_inputs["input_ids"]
 
     # Annotate split points
     add_split_points(gptneo, args.world_size)
@@ -52,7 +51,8 @@ def run(args):
     gptneo_pipe = Pipe.from_tracing(
         gptneo,
         num_chunks=args.chunks,
-        example_args=(input_ids, ),
+        example_args=(),
+        example_kwargs=example_inputs,
     )
     nstages = len(list(gptneo_pipe.split_gm.children()))
     assert nstages == args.world_size, f"nstages = {nstages} nranks = {args.world_size}"
@@ -69,7 +69,7 @@ def run(args):
 
     # Run
     if args.rank == 0:
-        stage(input_ids)
+        stage(**example_inputs)
     elif args.rank == args.world_size - 1:
         out = stage()
     else:

examples/hf/pippy_opt.py

@@ -43,7 +43,6 @@ def run(args):
     # Input configs
     example_inputs = generate_inputs_for_model(
         model_class, opt, model_name, args.batch_size, args.device)
-    input_ids = example_inputs["input_ids"]
 
     # Annotate split points
     add_split_points(opt, args.world_size)
@@ -52,7 +51,8 @@ def run(args):
     opt_pipe = Pipe.from_tracing(
         opt,
         num_chunks=args.chunks,
-        example_args=(input_ids, ),
+        example_args=(),
+        example_kwargs=example_inputs,
     )
     nstages = len(list(opt_pipe.split_gm.children()))
     assert nstages == args.world_size, f"nstages = {nstages} nranks = {args.world_size}"
@@ -69,7 +69,7 @@ def run(args):
 
     # Run
     if args.rank == 0:
-        stage(input_ids)
+        stage(**example_inputs)
     elif args.rank == args.world_size - 1:
         out = stage()
     else:

pippy/IR.py

@@ -3,6 +3,7 @@
 import logging
 import operator
 from enum import Enum
+from inspect import Parameter, signature, Signature
 from typing import Any, Callable, Dict, List, Optional, Tuple, Union
 
 import torch
@@ -655,8 +656,6 @@ def throw(self, *args, **kwargs):
     def forward(self, *args, **kwargs):
         executor_args = args
         if len(kwargs) > 0:
-            from inspect import Parameter, Signature
-
             parameters = []
             for node in self.split_gm.graph.nodes:
                 if node.op == "placeholder":
@@ -1005,6 +1004,34 @@ def move_param_to_callee(
 
         split.delete_all_unused_submodules()
 
+        # Users want the first pipeline stage to accept kwargs if the original
+        # program does. This is controlled by the `_codegen` field of the graph,
+        # so we make a copy here. Note: we only want the input spec and not the
+        # output spec, because the output spec is for the last stage. Maybe a
+        # TODO? Not sure yet.
+        submod0 = list(split.children())[0]
+        model_sign = signature(traced.forward)
+        model_num_args = len(model_sign.parameters)
+        submod0_sign = signature(submod0.forward)
+        submod0_num_args = len(submod0_sign.parameters)
+        if model_num_args != submod0_num_args:
+            # We don't change the signature of the first stage if it takes
+            # different number of args than original model
+            logger.info(
+                f"Original model takes {model_num_args} args but the first pipeline stage takes {submod0_num_args}. "
+                "Please provide args to respective pipeline stages."
+            )
+        else:
+            # Support kwargs for the first stage
+            submod0.graph._codegen = copy.deepcopy(traced.graph._codegen)
+            # `_replace` is actually not "private" or internal. based on this doc:
+            # To prevent conflicts with field names, the method and attribute names
+            # start with an underscore
+            submod0.graph._codegen.pytree_info = (
+                submod0.graph._codegen.pytree_info._replace(out_spec=None)
+            )
+            submod0.recompile()
+
         split.graph.lint()
         split.recompile()
 
@@ -1071,6 +1098,7 @@ def _trace_with_export(
                 example_kwargs,
                 constraints,
             )
+            logger.debug(f"Traced model: {traced}")
             if split_policy is not None:
                 traced = split_policy(traced)
         finally:

pippy/PipelineStage.py

@@ -7,6 +7,7 @@
 import torch.distributed as dist
 import torch.fx as fx
 from torch._subclasses.fake_tensor import FakeTensor
+from torch.fx.node import map_aggregate, map_arg
 from torch.nn.parallel import DistributedDataParallel
 
 from pippy.backward import stage_backward
@@ -47,6 +48,10 @@ class StageArgPlaceholder:
     pass
 
 
+class StageKwargPlaceholder:
+    pass
+
+
 class PipelineStage(torch.nn.Module):
     def __init__(
         self,
@@ -269,14 +274,15 @@ def create_recv_tensor(
 
         # `args` is a Tuple, hence we will have:
         # Tuple[RecvInfo]
-        args_recv_info = fx.node.map_arg(self.node.args, create_recv_tensor)
+        args_recv_info = map_arg(self.node.args, create_recv_tensor)
 
         # `kwargs` is a Dict, hence we will have:
         # Dict[keyword, RecvInfo]
-        kwargs_recv_info = fx.node.map_arg(self.node.kwargs, create_recv_tensor)
+        kwargs_recv_info = map_arg(self.node.kwargs, create_recv_tensor)
 
         logger.info(
-            f"[{self.group_rank}] " f"Activation recv info: {args_recv_info}"
+            f"[{self.group_rank}] "
+            f"Activation recv / args info: {args_recv_info}"
         )
         return args_recv_info, kwargs_recv_info
 
@@ -370,9 +376,9 @@ def map_recv_to_send(a):
                 grad_send_info.append(None)
                 return None
 
-        fx.node.map_aggregate(args_recv_info, map_recv_to_send)
+        map_aggregate(args_recv_info, map_recv_to_send)
 
-        fx.node.map_aggregate(kwargs_recv_info, map_recv_to_send)
+        map_aggregate(kwargs_recv_info, map_recv_to_send)
 
         logger.info(f"[{self.group_rank}] " f"Grad send info: {grad_send_info}")
         return grad_send_info
@@ -422,35 +428,43 @@ def _recv_and_fill_inputs(
 
         act_recv = self.recv_tensor_fn(recv_reqs)
 
+        chunk_args_list: List = []
         if self.args_split:
             chunk_args = self.args_split[chunk]
             chunk_args_list = list(chunk_args)
 
         def recv_args(info):
             if isinstance(info, RecvInfo):
+                # This is an activation to receive
                 return act_recv(info)
             else:
-                return chunk_args_list.pop(0)  # type: ignore[has-type]
+                # This is a pass-in argument
+                if len(chunk_args_list):
+                    return chunk_args_list.pop(0)  # type: ignore[has-type]
+                else:
+                    # kwargs were treated as args in graph phase. That's why
+                    # there are extra placeholders here. We mark them and filter
+                    # them out later.
+                    return StageKwargPlaceholder()
 
-        composite_args = fx.node.map_aggregate(
+        composite_args = map_aggregate(
             self.args_recv_info[chunk],
             recv_args,
         )
+        # Filter out kwarg placeholders
+        composite_args = tuple(
+            x
+            for x in composite_args
+            if not isinstance(x, StageKwargPlaceholder)
+        )
 
+        # Middle stages won't have incoming activations in kwargs form. So if
+        # kwargs_split is not empty, it must be model inputs for stage 0. We
+        # hence pass it as is to the interal submodule, without performing
+        # `recv_args` on it.
+        composite_kwargs: Dict = {}
         if self.kwargs_split:
-            chunk_kwargs = self.kwargs_split[chunk]
-
-        def recv_kwargs(info):
-            if isinstance(info, RecvInfo):
-                return act_recv(info)
-            else:
-                k = next(iter(chunk_kwargs))  # type: ignore[has-type]
-                return chunk_kwargs.pop(k)  # type: ignore[has-type]
-
-        composite_kwargs = fx.node.map_aggregate(
-            self.kwargs_recv_info[chunk],
-            recv_kwargs,
-        )
+            composite_kwargs = self.kwargs_split[chunk]
 
         # Wait for all recvs to finish
         for work in recv_reqs:
@@ -496,7 +510,7 @@ def _recv_grads(
         recv_grad = self.recv_tensor_fn(grad_recv_reqs)
 
         # Receive gradients
-        grads = fx.node.map_aggregate(
+        grads = map_aggregate(
             self.grad_recv_info[bwd_chunk],
             recv_grad,
         )