[Gradient Compression] Check start_PowerSGD_iter > 1 and add guidance on tuning PowerSGD configs. (#51427)

Yi Wang · facebook-github-bot · commit 79e7544cb4ec · 2021-02-02T04:30:24.000-08:00
Summary: Pull Request resolved: #51427 A user reported that `start_PowerSGD_iter` failed when it's set as 1. This is because allocating memory for error tensors somehow overlap with bucket rebuilding process at iteration 1. Check `start_PowerSGD_iter > 1` instead of `start_PowerSGD_iter >= 1`. Also add a unit test of `test_invalid_powerSGD_state` and some guidance on tuning PowerSGD configs. Original PR issue: Investigate Applying PowerSGD to Communication Hook for Gradient Compression #47202 ghstack-source-id: 120834126 Test Plan: buck test mode/dev-nosan caffe2/test/distributed:c10d -- test_invalid_powerSGD_state Reviewed By: rohan-varma Differential Revision: D26166897 fbshipit-source-id: 34d5b64bb3dd43acb61d792626c70e6c8bb44a5d
diff --git a/test/distributed/test_c10d.py b/test/distributed/test_c10d.py
@@ -3826,6 +3826,22 @@ def test_powerSGD_ddp_comm_hook_nccl(self):
     def test_ddp_comm_hook_allreduce_hook_nccl_grad_is_view(self):
         self._test_ddp_comm_hook_allreduce_hook_nccl(gradient_as_bucket_view=True)
 
+    def test_invalid_powerSGD_state(self):
+        for start_powerSGD_iter, use_error_feedback, warm_start in product([0, 1], [True, False], [True, False]):
+            if not use_error_feedback and not warm_start:
+                continue
+            with self.assertRaisesRegex(
+                    ValueError,
+                    "Expect `start_powerSGD_iter` > 1 if `use_error_feedback` or `warm_start` is enabled, "
+                    "because PowerSGD can only be applied after the first two iterations in DDP."):
+                state = powerSGD.PowerSGDState(
+                    process_group=None,
+                    matrix_approximation_rank=1,
+                    start_powerSGD_iter=start_powerSGD_iter,
+                    use_error_feedback=use_error_feedback,
+                    warm_start=warm_start,
+                )
+
     @requires_nccl()
     @skip_if_lt_x_gpu(2)
     def test_default_ddp_comm_hooks_nccl_is_view(self):
diff --git a/torch/distributed/algorithms/ddp_comm_hooks/powerSGD_hook.py b/torch/distributed/algorithms/ddp_comm_hooks/powerSGD_hook.py
@@ -33,6 +33,30 @@ def _orthogonalize(matrix, epsilon=1e-8):
 
 
 class PowerSGDState(object):
+    """
+    Stores both the gradient compression configs and the internal states for all the gradients during the training.
+    Particularly, `matrix_approximation_rank` and `start_powerSGD_iter` are the main configs that need to be tuned by the user.
+    Although `use_error_feedback` and `warm_start` can also be tuned by the user,
+    they are typically turned on for performance.
+
+    Note [Guidance to Tune `matrix_approximation_rank` And `start_powerSGD_iter`]
+    ~~~~~~~~~~~~~~~~~~~~~~~~~~
+    1) To tune `matrix_approximation_rank`, the user can increase it from 1 by factors of 2,
+    until a satisfying accuracy can be reached.
+    The increase of `matrix_approximation_rank` can substantially increase the computation costs of the compression.
+    However, the accuracy may not be futher improved beyond a certain `matrix_approximation_rank` value.
+    2) To tune `start_powerSGD_iter`, the user can typically start with 10% of total training steps,
+    and increase it until a satisfying accuracy can be reached.
+    Deferrring PowerSGD can effectively improve the accuracy,
+    even a relatively small `matrix_approximation_rank` is used.
+    This is because that, the beginning of training phase is usually very sensitive to inaccurate gradients,
+    and compressing gradients too early may make the training quickly take a suboptimal trajectory,
+    which can result in an irrecoverable impact on the accuracy.
+    The minimum value allowed in DDP is 2, if error feedback or warm-up is enabled.
+    This is because there is another internal optimization that rebuilds buckets at iteration 1 in DDP,
+    and this can conflict with any tensor memorized before the rebuild process.
+    """
+
     __slots__ = [
         "process_group",
         # The two fields below are the configs that usually need to be tuned by the user.
@@ -58,6 +82,16 @@ def __init__(
         warm_start=True,
         random_seed=0,
     ):
+        logging.info(
+            "PowerSGD config: matrix_approximation_rank = {}; "
+            "start_powerSGD_iter = {}; use_error_feedback = {}; warm_start = {}.".format(
+                matrix_approximation_rank,
+                start_powerSGD_iter,
+                use_error_feedback,
+                warm_start,
+            )
+        )
+
         self.process_group = process_group
         # The low rank for matrix approximation controls the size of compressed low-rank tensors,
         # which determines the computation ratio.
@@ -80,7 +114,11 @@ def __init__(
         # However, this means that the shape of input bucketized tensors is subject to change,
         # which will complicate the implementations of error feedback and warm-up.
         # Running vanilla allreduce in the first few iterations can avoid this complexity.
-        assert start_powerSGD_iter >= 1
+        if (use_error_feedback or warm_start) and start_powerSGD_iter <= 1:
+            raise ValueError(
+                "Expect `start_powerSGD_iter` > 1 if `use_error_feedback` or `warm_start` is enabled, "
+                "because PowerSGD can only be applied after the first two iterations in DDP."
+            )
         self.start_powerSGD_iter = start_powerSGD_iter
         # Error feedback is usually crucial for both for convergence and generalization,
         # because PowerSGD is a biased compressor,
@@ -111,16 +149,6 @@ def __init__(
         # Iteration/step in the training loop.
         self.iter = 0
 
-        logging.info(
-            "PowerSGD config: matrix_approximation_rank = {}; "
-            "start_powerSGD_iter = {}; use_error_feedback = {}; warm_start = {}.".format(
-                self.matrix_approximation_rank,
-                self.start_powerSGD_iter,
-                self.use_error_feedback,
-                self.warm_start,
-            )
-        )
-
     def maybe_increase_iter(self, bucket):
         # Since bucket 0 is the last bucket to allreduce in an iteration.
         # Only increase `iter` when bucket 0 is processed.
@@ -165,6 +193,7 @@ def powerSGD_hook(state: PowerSGDState, bucket) -> torch.futures.Future:
 
     Args:
         state (PowerSGDState): State information to configure the compression rate and support error feedback, warm start, etc.
+            To tune the compression configs, see Note [Guidance to Tune `matrix_approximation_rank` And `start_powerSGD_iter`].
         bucket (dist._GradBucket): Bucket that stores a 1D flattened gradient tensor that batches multiple per-variable tensors.
             Note that since DDP comm hook only supports single process single device mode at this time,
             only exactly one tensor is stored in this bucket.
@@ -399,6 +428,13 @@ def batched_powerSGD_hook(state: PowerSGDState, bucket) -> torch.futures.Future:
     7) Computes M, which is approximately equal to PQ^T.
     8) Truncates the input tensor to the original length.
 
+    This variant is faster than `powerSGD_hook` that runs layer-wise gradient compression,
+    but it usually results in a much lower accuracy, unless `matrix_approximation_rank` in the state is 1.
+    Increasing `matrix_approximation_rank` may not necessarily increase the accuracy,
+    because batching per-parameter tensors without column/row alignment can destroy low-rank structure.
+    Therefore, the user shoud always consider `powerSGD_hook` first,
+    and only consider this variant when a satisfying accuracy can be achieved when `matrix_approximation_rank` is 1.
+
     Note that this communication hook enforces vanilla allreduce for the first `state.start_powerSGD_iter` iterations.
     This can not only allow the user to have a finer tuning over the tradeoff between speedup and accuracy,
     but also help abstract away some complexity of the internal optimization of DDP for future communication hook developers.
@@ -409,6 +445,7 @@ def batched_powerSGD_hook(state: PowerSGDState, bucket) -> torch.futures.Future:
 
     Args:
         state (PowerSGDState): State information to configure the compression rate and support error feedback, warm start, etc.
+            To tune the compression configs, see Note [Guidance to Tune `matrix_approximation_rank` And `start_powerSGD_iter`].
         bucket (dist._GradBucket): Bucket that stores a 1D flattened gradient tensor that batches multiple per-variable tensors.
             Note that since DDP comm hook only supports single process single device mode at this time,
             only exactly one tensor is stored in this bucket.
