Add bfloat16 support to DeepSpeedStrategy

[rohitgr7]

What does this PR do?

Add bf16 support to DeepSpeedStrategy.

More info: https://www.deepspeed.ai/docs/config-json/#bfloat16-training-options

Note that bf16 support was added in deepspeed==0.6.0. Do we maintain a minimum version for deepspeed? If not, then if someone has deepspeed < 0.6.0 and tries to use bf16, it won't fail. I can add a check in that case and throw an exception if we support deepspeed < 0.6.0.

Does your PR introduce any breaking changes? If yes, please list them.

Before submitting

• Was this discussed/approved via a GitHub issue? (not for typos and docs)
• Did you read the contributor guideline, Pull Request section?
• Did you make sure your PR does only one thing, instead of bundling different changes together?
• Did you make sure to update the documentation with your changes? (if necessary)
• Did you write any new necessary tests? (not for typos and docs)
• Did you verify new and existing tests pass locally with your changes?
• Did you list all the breaking changes introduced by this pull request?
• Did you update the CHANGELOG? (not for typos, docs, test updates, or internal minor changes/refactorings)

PR review

Anyone in the community is welcome to review the PR.
Before you start reviewing make sure you have read Review guidelines. In short, see the following bullet-list:

• Is this pull request ready for review? (if not, please submit in draft mode)
• Check that all items from Before submitting are resolved
• Make sure the title is self-explanatory and the description concisely explains the PR
• Add labels and milestones (and optionally projects) to the PR so it can be classified

Did you have fun?

Make sure you had fun coding 🙃

cc @Borda @SeanNaren @awaelchli @rohitgr7 @akihironitta

[leezu]

This PR misses implementation for model_sharded_context. You can apply the following patch to add support:

@@ -512,11 +521,12 @@ def _initialize_deepspeed_train(self, model):
     def model_sharded_context(self) -> Generator[None, None, None]:
         if self.zero_stage_3:
             assert self._config_initialized
-            dtype = (
-                torch.float16
-                if self.precision_plugin.precision in (PrecisionType.HALF, PrecisionType.MIXED)
-                else torch.float32
-            )
+            if self.precision_plugin.precision in (PrecisionType.HALF, PrecisionType.MIXED):
+                dtype = torch.float16
+            elif self.precision_plugin.precision == PrecisionType.BFLOAT:
+                dtype = torch.bfloat16
+            else:
+                dtype = torch.float32
             model_parallel_context = deepspeed.zero.Init(
                 remote_device=self.remote_device, pin_memory=True, config_dict_or_path=self.config, dtype=dtype
             )

[leezu]

If you include patch, please also add a test for ZeRO3+bf16 with model_sharded_context. Otherwise please ping me and I can open a new PR after this one is merged.

[rohitgr7]

This PR misses implementation for model_sharded_context. You can apply the following patch to add support:

yep, I was reading about why it is done like that for stage3, that's why kept this draft.

[toriving]

Can i use bf16 in the latest master branch?

[rohitgr7]

Can i use bf16 in the latest master branch?

@toriving , yes

[toriving]

Can i use bf16 in the latest master branch?

@toriving , yes

Thx. I can not use the latest release version (1.6.5) but can use it in master branch.

[rohitgr7]

Can i use bf16 in the latest master branch?

@toriving , yes

Thx. I can not use the latest release version (1.6.5) but can use it in the master branch.

yep... v1.6.5 is the bug fix branch and does not include any new features that were merged after v1.6 release. This is currently available only on master and will be available in v1.7 release.