only enable query key scaling during fp16 #7946
Merged
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
Signed-off-by: Gerald Shen <[email protected]>
|
jenkins |
Signed-off-by: Gerald Shen <[email protected]>
|
jenkins |
ericharper
reviewed
Nov 29, 2023
| @@ -1544,6 +1544,11 @@ def build_transformer_config(self) -> TransformerConfig: | |||
|
|
|||
| attention_softmax_in_fp32 = False # not currently used in NeMo unless apply_query_key_layer_scaling is True | |||
| apply_query_key_layer_scaling = self.cfg.get('apply_query_key_layer_scaling', False) | |||
|
|
|||
| if apply_query_key_layer_scaling and not model_parallel_config.fp16: | |||
| logging.warning("apply_query_key_layer_scaling is only enabled when using FP16, setting it to False") | |||
There was a problem hiding this comment.
I don't think model_parallel_config.fp16 is the right check though. That arg is for fp16 + megatron_amp_O2.
Maybe we should just check trainer.precision?
There was a problem hiding this comment.
can I check against self.torch_dtype?
Signed-off-by: Gerald Shen <[email protected]>
|
jenkins |
Signed-off-by: Gerald Shen <[email protected]>
|
jenkins |
Signed-off-by: Gerald Shen <[email protected]>
|
jenkins |
ericharper
reviewed
Dec 1, 2023
| if fp16_enabled: | ||
| os.environ["NVTE_APPLY_QK_LAYER_SCALING"] = "1" | ||
| else: | ||
| logging.warning("apply_query_key_layer_scaling is only enabled when using FP16, setting it to False") |
There was a problem hiding this comment.
Should we set the env var here as well to "0" ?
There was a problem hiding this comment.
good point, otherwise it will error 366cf61
Signed-off-by: Gerald Shen <[email protected]>
|
jenkins |
pzelasko
pushed a commit
to pzelasko/NeMo
that referenced
this pull request
Jan 3, 2024
* only enable query key scaling during fp16 Signed-off-by: Gerald Shen <[email protected]> * add warning Signed-off-by: Gerald Shen <[email protected]> * fixup! only enable query key scaling during fp16 Signed-off-by: Gerald Shen <[email protected]> * remove var from jenkens file Signed-off-by: Gerald Shen <[email protected]> * fix test by setting TE var Signed-off-by: Gerald Shen <[email protected]> * set to 0 if disabled Signed-off-by: Gerald Shen <[email protected]> --------- Signed-off-by: Gerald Shen <[email protected]> Signed-off-by: Piotr Żelasko <[email protected]>
ssh-meister
pushed a commit
to ssh-meister/NeMo
that referenced
this pull request
Feb 15, 2024
* only enable query key scaling during fp16 Signed-off-by: Gerald Shen <[email protected]> * add warning Signed-off-by: Gerald Shen <[email protected]> * fixup! only enable query key scaling during fp16 Signed-off-by: Gerald Shen <[email protected]> * remove var from jenkens file Signed-off-by: Gerald Shen <[email protected]> * fix test by setting TE var Signed-off-by: Gerald Shen <[email protected]> * set to 0 if disabled Signed-off-by: Gerald Shen <[email protected]> --------- Signed-off-by: Gerald Shen <[email protected]> Signed-off-by: Sasha Meister <[email protected]>
rohitrango
pushed a commit
to rohitrango/NeMo
that referenced
this pull request
Jun 25, 2024
* only enable query key scaling during fp16 Signed-off-by: Gerald Shen <[email protected]> * add warning Signed-off-by: Gerald Shen <[email protected]> * fixup! only enable query key scaling during fp16 Signed-off-by: Gerald Shen <[email protected]> * remove var from jenkens file Signed-off-by: Gerald Shen <[email protected]> * fix test by setting TE var Signed-off-by: Gerald Shen <[email protected]> * set to 0 if disabled Signed-off-by: Gerald Shen <[email protected]> --------- Signed-off-by: Gerald Shen <[email protected]>
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.
What does this PR do ?
Add a one line overview of what this PR aims to accomplish.
Enable query key scaling only during fp16, since this is when TE enables query key scaling. https://github.com/NVIDIA/TransformerEngine/blob/666539f36275fa9c0fbc99f9ea50f2d6e29e336f/transformer_engine/pytorch/attention.py#L940