TE integration via full TransformerLayer #1297
Open
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
|
Will merge this after the finegrained TE PR |
| "The mask will be discarded") | ||
| hidden_states, attention_mask = args | ||
|
|
||
| fp8_format = Format.HYBRID # E4M3 during forward pass, E5M2 during backward pass |
There was a problem hiding this comment.
should instead accept neox_arg from the new te_fp8_format
| @@ -271,6 +272,24 @@ def init_specs(self): | |||
| layer_number=i, | |||
| ) | |||
| ) | |||
| elif layer_type in ["TE"]: | |||
There was a problem hiding this comment.
needs tested with PP and TP, since we'd be relying on two external codebases (deepspeed for PP, TE for TP) whose topologies probably don't play nicely together.
|
|
||
| RUN DS_BUILD_FUSED_LAMB=1 DS_BUILD_FUSED_ADAM=1 DS_BUILD_TRANSFORMER=1 DS_BUILD_STOCHASTIC_TRANSFORMER=1 DS_BUILD_UTILS=1 \ | ||
| TORCH_CUDA_ARCH_LIST="8.0 9.0+PTX" \ | ||
| python -m pip install git+https://github.com/microsoft/[email protected] |
There was a problem hiding this comment.
this should probably instead be latest deeperspeed. Can't hardcode the arch list.
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
This is a sketch of using the attention picking mechanism ("global", "flash", NEW: "TE") to use the high level TransformerLayer from TransformerEngine. This is more of a prototype to show that integration with deepspeed is possible and what perf to expect.
Things that work:
Many aspects are hardcoded, e.g. RoPE and activation checkpointing can not be reconfigured from the config files. #1282 is much more elaborate in that it exposes TE layers on a much lower level. Meanwhile this PR could serve as a benchmark, showing what is possible with TE on a classic GPT2 style network.
I kept the implementation as minimal as possible, there is room for further performance depending on the workload. There is e.g. sequence parallelism and different memory layouts.
The dockerfile now uses a later ngc pytorch container and installs a later deepspeed tag from source for compatibility.