[fix] Revert change that removed the option to run OffloadModel with out...

[fix] Revert change that removed the option to run OffloadModel with out activation checkpointing. (#608)

* revert change made

* add tests and revert sync shard changes

* add tests

* remove file checked in by error

* inine var

* fix lint errors

* add checkpoint activation

* fix mypy

* use a bigger model

* modify tests for now

* resolve conflicts
Co-authored-by: Anjali Sridhar <anj@devfair0443.h2.fair>

.circleci/config.yml

@@ -170,7 +170,7 @@ run_offload_benchmark: &run_offload_benchmark
   - run:
       name: Run Offload Benchmark
       command: |
-        python benchmarks/experimental/offload.py
+        python benchmarks/experimental/offload.py --checkpoint_activation
 
 run_pipe_benchmark: &run_pipe_benchmark
   - run:

benchmarks/experimental/offload.py

@@ -233,7 +233,7 @@ def train(model_config, model, benchmark_config, model_specs, args):
 
 
 def verify_peak_memory(golden_config, std_dev):
-    print("Peak allocated bytes on cuda:0: {:1d}".format(torch.cuda.memory_stats(0)["allocated_bytes.all.peak"]))
+
     current_device_usage = torch.cuda.memory_stats(0)["allocated_bytes.all.peak"]
     golden_ref = golden_config["peak_mem_usage"]
     if not current_device_usage < golden_ref * std_dev:
@@ -246,7 +246,6 @@ def verify_peak_memory(golden_config, std_dev):
 def verify_lm_throughput(wps, golden_config, args):
     """Verify that words per second for a given benchmark run matches the golden data."""
 
-    print("Throughput(wps) is {:.2f}.".format(wps))
     if not wps > (golden_config["avg_wps"] - (3 * golden_config["std_dev_wps"])):
         raise RuntimeError(
             "Throughput(wps):{:.2f} is below the golden threshold of an "
@@ -272,9 +271,12 @@ def benchmark_language_model(model_config, model, benchmark_config, model_specs,
         raise RuntimeError(
             f"Golden data verification is only supported for the Transformer(lm) model and not {args.model_name}"
         )
-    golden_config = get_golden_config(args.model_name, args)
-    verify_lm_throughput(wps, golden_config, args)
-    verify_peak_memory(golden_config, 1.1)
+    print("Throughput(wps) is {:.2f}.".format(wps))
+    print("Peak allocated bytes on cuda:0: {:1d}".format(torch.cuda.memory_stats(0)["allocated_bytes.all.peak"]))
+    if not args.dry_run:
+        golden_config = get_golden_config(args.model_name, args)
+        verify_lm_throughput(wps, golden_config, args)
+        verify_peak_memory(golden_config, 1.1)
 
 
 def get_synthetic_dataloaders(args, device, benchmark_config, model_specs):
@@ -343,11 +345,11 @@ def create_model_config(args, benchmark_config=None, model_specs=None):
         raise RuntimeError(f"Unrecognized args.model_mame {args.model_name}")
 
 
-def create_benchmark_config(model_name):
+def create_benchmark_config(args):
     """Return a dict with configurations required for benchmarking `model_name` model."""
 
     if args.model_name == "lm":
-        return lm_wikitext2.get_benchmark_config()
+        return lm_wikitext2.get_benchmark_config(checkpoint_activation=args.checkpoint_activation)
     elif args.model_name == "seq":
         return offload_seq.get_benchmark_config()
     else:
@@ -383,17 +385,15 @@ def run_benchmark(args):
     init_random_seed(0)
 
     if args.model_name == "lm":
-        benchmark_config = create_benchmark_config(args.model_name)
+        benchmark_config = create_benchmark_config(args)
         model_specs = get_model_specs(args.model_name)
         model_config = create_model_config(args, benchmark_config=benchmark_config, model_specs=model_specs)
         model = model_config["model"]
 
-        if args.dry_run:
-            train(model_config, model, benchmark_config, model_specs, args)
-        else:
-            benchmark_language_model(model_config, model, benchmark_config, model_specs, args)
+        benchmark_language_model(model_config, model, benchmark_config, model_specs, args)
+
     elif args.model_name == "seq":
-        benchmark_config = create_benchmark_config(args.model_name)
+        benchmark_config = create_benchmark_config(args)
         model_specs = get_model_specs(args.model_name)
         model_config = create_model_config(args, benchmark_config=benchmark_config, model_specs=model_specs)
         model = model_config["model"]
@@ -419,7 +419,7 @@ parser.add_argument(
     "--use_synthetic_data", default=True, action="store_true", help="Uses synthetic data for running benchmarks."
 )
 parser.add_argument("--use_fp16", action="store_true", default=False)
-parser.add_argument("--checkpoint_activation", action="store_true", default=True)
+parser.add_argument("--checkpoint_activation", action="store_true", default=False)
 parser.add_argument("--use_profiler", action="store_true", default=False)
 
 

benchmarks/golden_configs/lm_wikitext2.py

@@ -20,14 +20,14 @@ class Offload_Transformer:
             "seq_len": 32,
         }
 
-    def get_benchmark_config():
+    def get_benchmark_config(checkpoint_activation=True):
 
         return {
             "epochs": 1,
             "lr": 0.001,  # learning rate
             "batch_size": 8,
             "criterion": nn.CrossEntropyLoss(),
-            "checkpoint_activation": True,
+            "checkpoint_activation": checkpoint_activation,
             "num_microbatches": 1,
             "slices": 3,
         }
@@ -59,7 +59,7 @@ class Offload_Sequential:
             "criterion": nn.CrossEntropyLoss(),
             "slices": 3,
             "checkpoint_activation": True,
-            "num_microbatches": 4,
+            "num_microbatches": 1,
         }
 
 

fairscale/experimental/nn/offload.py

@@ -292,6 +292,75 @@ class OffloadFunction(torch.autograd.Function):
         return (None, None) + grads
 
 
+class ShardSyncLayer(torch.autograd.Function):
+    """
+     The shard sync layer is a synchronization point between model shards.
+     - In the forward pass, it drops parameters in the previous shard and
+     loads parameters for the next shard.
+     - In the backward pass, it does the reverse.
+     It does not change or create any outputs at all, instead it just
+     forwards the input as the output.
+     NOTE: see https://pytorch.org/docs/stable/autograd.html#torch.autograd.Function
+     """
+
+    @staticmethod
+    @_conditional_amp_fwd_decorator  # type: ignore
+    def forward(ctx: Any, inputs: Any, index: int, model_slices: Any, model_instance: Any) -> Any:
+        drop_index = index
+        load_index = index + 1
+        max_slices = len(model_slices)
+
+        if drop_index >= 0:
+            # Move shard from device to offload device.
+            model_slices[drop_index].forward_drop()
+
+        if load_index < max_slices:
+            # Load shard from offload device to device.
+            model_slices[load_index].forward_load()
+
+        ctx.index = index
+        ctx.model_slices = model_slices
+        ctx.model_instance = model_instance
+
+        return inputs if isinstance(inputs, tuple) else (inputs,)
+
+    @staticmethod
+    @_conditional_amp_bwd_decorator
+    def backward(ctx, *grad_outputs):  # type: ignore
+
+        load_index = ctx.index
+        drop_index = load_index + 1
+        model_slices = ctx.model_slices
+        model_instance = ctx.model_instance
+
+        # TODO(anj-s): Are these redundant in the backward pass?
+        if drop_index == len(model_slices):
+            # Drop the last activation since it is still on the CPU
+            # after the loss.backward() call.
+            model_instance._activations[-1] = tuple([a.cuda() for a in list(model_instance._activations[-1])])
+
+        if drop_index < len(model_slices):
+            # Move shard from device to offload device.
+            model_slices[drop_index].backward_drop()
+            model_instance._activations[drop_index] = tuple(
+                [a.cpu() for a in list(model_instance._activations[drop_index])]
+            )
+
+        if load_index >= 0:
+            # Load shard from offload device to device.
+            model_slices[load_index].backward_load()
+            model_instance._activations[load_index] = tuple(
+                [a.cuda() for a in list(model_instance._activations[load_index])]
+            )
+
+        # The returned variables need to mirror the forward inputs
+        # TODO(anj-s): Why do we need to do this?
+        if isinstance(grad_outputs, tuple):
+            return grad_outputs[0], None, None, None
+
+        return grad_outputs, None, None, None
+
+
 class OffloadModel(nn.Module):
     """Wraps an arbitrary :class:`nn.Sequential <torch.nn.Sequential>` module
     to train by offloading majority of the model parameters to the CPU.
@@ -405,4 +474,23 @@ class OffloadModel(nn.Module):
 
         # We need the second param to be a dummy input to enable the
         # backward pass to be triggered for integer inputs.
-        return OffloadFunction.apply(*inputs, torch.tensor([], requires_grad=True), self)
+        if self._checkpoint_activation:
+            return OffloadFunction.apply(*inputs, torch.tensor([], requires_grad=True), self)
+
+        self._activations = []
+        for index in range(-1, len(self.model_slices)):
+            if index >= 0:
+                # TODO(anj-s): This might be a redundant call since we have the previous
+                # activation on the device already.
+                self._activations[index] = tuple([a.cuda() for a in list(self._activations[index])])
+                inputs = self._activations[index]
+                inputs = self.model_slices[index](*inputs)
+            # Call the custom autograd hooks (discard/load slices FW and BW)
+            inputs = ShardSyncLayer.apply(inputs, index, self.model_slices, self)
+            self._activations.append(inputs)
+            if index >= 0:
+                self._activations[index] = tuple([a.cpu() for a in list(self._activations[index])])
+
+        result = self._activations[-1]
+        result = tuple([r.cuda() for r in result])
+        return result[0] if len(result) == 1 else result

tests/experimental/nn/test_offload.py

@@ -32,20 +32,38 @@ def test_single_run():
     device, offload_device = _init()
     model = _get_model()
 
-    offload_model = OffloadModel(model=model, device=device, offload_device=offload_device, num_slices=2,)
-    offload_optimizer = torch.optim.SGD(offload_model.parameters(), lr=0.001)
+    peak_mem = {}
+    for checkpoint_activation in [True, False]:
+        offload_model = OffloadModel(
+            model=model,
+            device=device,
+            offload_device=offload_device,
+            num_slices=2,
+            checkpoint_activation=checkpoint_activation,
+        )
+        offload_optimizer = torch.optim.SGD(offload_model.parameters(), lr=0.001)
+
+        input = torch.ones(1000, 2).to(device)
+        labels = torch.ones(1000, 2).to(device)
+        offload_model.train()
+        pred = offload_model(input)
+        loss_fn = torch.nn.MSELoss(reduction="sum")
+        loss = loss_fn(pred, labels)
+        loss.backward()
+        offload_optimizer.step()
+        key = "ca_" + str(checkpoint_activation)
+        peak_mem[key] = torch.cuda.memory_stats(0)["allocated_bytes.all.peak"]
+        print(
+            "Peak allocated bytes on cuda:0 for checkpoint_activation "
+            + str(checkpoint_activation)
+            + ": {:2f}".format(peak_mem[key])
+        )
 
-    input = torch.ones(2, 2).to(device)
-    labels = torch.ones(2, 2).to(device)
-    offload_model.train()
-    pred = offload_model(input)
-    loss_fn = torch.nn.MSELoss(reduction="sum")
-    loss = loss_fn(pred, labels)
-    loss.backward()
-    offload_optimizer.step()
+    # TODO(anj-s): We need a better requirement since this fails on CircleCI right now.
+    assert peak_mem["ca_True"] <= peak_mem["ca_False"]
 
 
-def _get_model(num_inputs=2, num_hidden=2, num_layers=1, num_outputs=2):
+def _get_model(num_inputs=2, num_hidden=20, num_layers=10, num_outputs=2):
     model = torch.nn.Sequential(
         torch.nn.Linear(num_inputs, num_hidden),
         *([torch.nn.Linear(num_hidden, num_hidden) for _ in range(num_layers)]),