Pipeline parallelism and gradient checkpointing (edit: and ZeRO 2!) don’t work together about gpt-neox

[StellaAthena]

Update: we can now use any two of the following three options: ZeRO Stage 2, Parallel Pipelining, and Activation Checkpointing. If all three are enabled, it throws the following error:

/root/anaconda3/envs/ds/lib/python3.7/site-packages/deepspeed/runtime/pipe/engine.py:993: UserWarning: The .grad attribute of a Tensor that is not a leaf Tensor is being accessed. Its .grad attribute won't be populated during autograd.backward(). If you indeed want the gradient for a non-leaf Tensor, use .retain_grad() on the non-leaf Tensor. If you access the non-leaf Tensor by mistake, make sure you access the leaf Tensor instead. See github.com/pytorch/pytorch/pull/30531 for more informations.

if you open the Stella branch you can replicate this. If you run sh scripts/train_enwik8_pipeline.sh it will have all three enabled and error.

To turn off activation checkpointing, set “number_checkpoints”: null in configs/deepspeed_zero2.json.

To turn off pipelining, run sh scripts/train_enwik8.sh.

To turn off ZeRO Stage 2, use configs/deepspeed_zero1.json as your config file.

[leogao2]

I ran all pairs of 3 and the results are as follows.

Zero2+pipeline: does not work (contiguous gradients both on and off)
Checkpoint+pipeline: does work (contiguous gradients both on and off)
Zero2+checkpoint: does work (contiguous gradients on; didn't test off)

[leogao2]

All of the errors and warnings that occur for zero2+pipeline:

Traceback (most recent call last):
  File "train_enwik8_pipeline.py", line 109, in <module>
    loss = model_engine.train_batch()
  File "/root/anaconda3/envs/ds/lib/python3.7/site-packages/deepspeed/runtime/pipe/engine.py", line 273, in train_batch
    self._exec_schedule(sched)
  File "/root/anaconda3/envs/ds/lib/python3.7/site-packages/deepspeed/runtime/pipe/engine.py", line 1162, in _exec_schedule
    self._exec_instr(**cmd.kwargs)
  File "/root/anaconda3/envs/ds/lib/python3.7/site-packages/deepspeed/runtime/pipe/engine.py", line 952, in _exec_optimizer_step
    self._take_model_step(lr_kwargs)
  File "/root/anaconda3/envs/ds/lib/python3.7/site-packages/deepspeed/runtime/engine.py", line 916, in _take_model_step
    self.optimizer.step()
  File "/root/anaconda3/envs/ds/lib/python3.7/site-packages/deepspeed/runtime/zero/stage2.py", line 1341, in step
    self.check_overflow()
  File "/root/anaconda3/envs/ds/lib/python3.7/site-packages/deepspeed/runtime/zero/stage2.py", line 1612, in check_overflow
    self._check_overflow(partition_gradients)
  File "/root/anaconda3/envs/ds/lib/python3.7/site-packages/deepspeed/runtime/zero/stage2.py", line 1516, in _check_overflow
    self.overflow = self.has_overflow(partition_gradients)
  File "/root/anaconda3/envs/ds/lib/python3.7/site-packages/deepspeed/runtime/zero/stage2.py", line 1535, in has_overflow
    overflow = self.local_overflow if self.cpu_offload else self.has_overflow_partitioned_grads_serial(
  File "/root/anaconda3/envs/ds/lib/python3.7/site-packages/deepspeed/runtime/zero/stage2.py", line 1528, in has_overflow_partitioned_grads_serial
    for j, grad in enumerate(self.averaged_gradients[i]):
KeyError: 0

(this one shows up 4 times)

Traceback (most recent call last):
  File "train_enwik8_pipeline.py", line 109, in <module>
    self._exec_schedule(sched)
  File "/root/anaconda3/envs/ds/lib/python3.7/site-packages/deepspeed/runtime/pipe/engine.py", line 1162, in _exec_schedule
    loss = model_engine.train_batch()
  File "/root/anaconda3/envs/ds/lib/python3.7/site-packages/deepspeed/runtime/pipe/engine.py", line 273, in train_batch
    self._exec_schedule(sched)
  File "/root/anaconda3/envs/ds/lib/python3.7/site-packages/deepspeed/runtime/pipe/engine.py", line 1162, in _exec_schedule
    self._exec_instr(**cmd.kwargs)
  File "/root/anaconda3/envs/ds/lib/python3.7/site-packages/deepspeed/runtime/pipe/engine.py", line 602, in _exec_backward_pass
    torch.autograd.backward(tensors=(outputs, ), grad_tensors=(grad_tensors, ))
  File "/root/anaconda3/envs/ds/lib/python3.7/site-packages/torch/autograd/__init__.py", line 132, in backward
    self._exec_instr(**cmd.kwargs)
  File "/root/anaconda3/envs/ds/lib/python3.7/site-packages/deepspeed/runtime/pipe/engine.py", line 602, in _exec_backward_pass
    allow_unreachable=True)  # allow_unreachable flag
  File "/root/anaconda3/envs/ds/lib/python3.7/site-packages/deepspeed/runtime/zero/stage2.py", line 594, in reduce_partition_and_remove_grads
    torch.autograd.backward(tensors=(outputs, ), grad_tensors=(grad_tensors, ))
  File "/root/anaconda3/envs/ds/lib/python3.7/site-packages/torch/autograd/__init__.py", line 132, in backward
    self.reduce_ready_partitions_and_remove_grads(param, i)
  File "/root/anaconda3/envs/ds/lib/python3.7/site-packages/deepspeed/runtime/zero/stage2.py", line 984, in reduce_ready_partitions_and_remove_grads
    allow_unreachable=True)  # allow_unreachable flag
  File "/root/anaconda3/envs/ds/lib/python3.7/site-packages/deepspeed/runtime/zero/stage2.py", line 594, in reduce_partition_and_remove_grads
    self.reduce_independent_p_g_buckets_and_remove_grads(param, i)
  File "/root/anaconda3/envs/ds/lib/python3.7/site-packages/deepspeed/runtime/zero/stage2.py", line 633, in reduce_independent_p_g_buckets_and_remove_grads
    self.reduce_ready_partitions_and_remove_grads(param, i)
  File "/root/anaconda3/envs/ds/lib/python3.7/site-packages/deepspeed/runtime/zero/stage2.py", line 984, in reduce_ready_partitions_and_remove_grads
    new_grad_tensor = self.ipg_buffer[self.ipg_index].narrow(
AttributeError: 'FP16_DeepSpeedZeroOptimizer' object has no attribute 'ipg_index'

(this one pops up 4 times and I think 2 of them got mangled together here)

As well as the warning that stella mentioned.

[leogao2]

With contiguous gradients off, the FP16_DeepSpeedZeroOptimizer error no longer happens and I get 8 KeyErrors.

[leogao2]

Checkpoint+pipeline works with both continuous gradients on and off. Therefore, I don't think it's a major factor for zero2 breaking, but I'll keep it off for the remainder of my tests.

[leogao2]

Focusing on the KeyError now.

The only place where self.averaged_gradients is written to within stage2.py is in the independent_gradient_partition_epilogue function (https://github.com/microsoft/DeepSpeed/blob/master/deepspeed/runtime/zero/stage2.py#L485). So either this function just isn't being called, or it is being called but L485 is never being reached.

[leogao2]

So the only place where independent_gradient_partition_epilogue is called is https://github.com/microsoft/DeepSpeed/blob/master/deepspeed/runtime/zero/stage2.py#L580. Which is only called at https://github.com/microsoft/DeepSpeed/blob/master/deepspeed/runtime/engine.py#L835. Which is only called at https://github.com/microsoft/DeepSpeed/blob/master/deepspeed/runtime/engine.py#L915.

[leogao2]

The reason the pipeline code is problematic is because it disables the backward_allreduce
https://github.com/microsoft/DeepSpeed/blob/81aeea361da3936b875a678b9cb44596800510b5/deepspeed/runtime/pipe/engine.py#L56
which means allreduce_gradients in the non-Pipelined engine never runs
https://github.com/microsoft/DeepSpeed/blob/master/deepspeed/runtime/engine.py#L914
which means independent_gradient_partition_epilogue never gets called (see previous comment)

The pipeline code pushes a ReduceTiedGrads and then a ReduceGrads here: https://github.com/microsoft/DeepSpeed/blob/81aeea361da3936b875a678b9cb44596800510b5/deepspeed/runtime/pipe/schedule.py#L235

Execution of that ReduceTiedGrads op:
https://github.com/microsoft/DeepSpeed/blob/81aeea361da3936b875a678b9cb44596800510b5/deepspeed/runtime/pipe/engine.py#L1139
https://github.com/microsoft/DeepSpeed/blob/81aeea361da3936b875a678b9cb44596800510b5/deepspeed/runtime/pipe/engine.py#L208
https://github.com/microsoft/DeepSpeed/blob/81aeea361da3936b875a678b9cb44596800510b5/deepspeed/runtime/pipe/module.py#L405

Execution of ReduceGrads:
https://github.com/microsoft/DeepSpeed/blob/81aeea361da3936b875a678b9cb44596800510b5/deepspeed/runtime/pipe/engine.py#L211
calls buffered_allreduce_fallback, but only if dataparallel is enabled.
https://github.com/microsoft/DeepSpeed/blob/865104be85902ca398038045ad9cf94ec7d48745/deepspeed/runtime/engine.py#L1156

[leogao2]

With the patch applied,

Zero2+pipeline now works
Checkpoint+pipeline now works
Zero2+checkpoint now works

Zero2+checkpoint+pipeline now works

[leogao2]

Profiling results:

patched, zero2+checkpoint+pipeline: samples/sec: 1159.741, max vram: 3245MiB
patched: zero2+checkpoint: samples/sec: 1120.8568733324405, max vram: 1704MiB

[StellaAthena]

With DeepSpeed's updates this seems to run just fine. The question of if it runs efficiently is still open though.

microsoft/DeepSpeed#677

[StellaAthena]

Turns out we weren’t using gradient checkpointing at all! You can add checkpointing to the params without initializing the checkpointer and you can initialize the checkpointer without actually using it! #90 should actually implement gradient checkpointing.