[float8] Allow specifying arbitrary dtype for each tensor#1326
Draft
lw wants to merge 7 commits into
Draft
Conversation
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/ao/1326
Note: Links to docs will display an error until the docs builds have been completed. ❌ 1 New FailureAs of commit 97c9983 with merge base 1a0dbf1 ( NEW FAILURE - The following job has failed:
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
vkuzo
reviewed
Nov 26, 2024
| scaling_type: ScalingType = ScalingType.DYNAMIC | ||
| scaling_granularity: ScalingGranularity = ScalingGranularity.TENSORWISE | ||
| static_scale: Optional[torch.Tensor] = None | ||
| dtype: Optional[torch.dtype] = None |
Contributor
There was a problem hiding this comment.
nit:
- can we add a comment on what this is used for, and that
Nonemeans the default e4m3|e5m2 value will be used? - optional - thoughts about naming this in a more specific way such as
target_dtype,lowp_dtype, etc?dtypeis a bit ambiguous across torchao unfortunately :(
vkuzo
reviewed
Nov 26, 2024
| # grad_input_hp = grad_output_fp8_axiswise_dim0 @ weight_fp8_tensorwise | ||
| cc_go = CastConfig(scaling_granularity=ScalingGranularity.AXISWISE) | ||
| cc_go = CastConfig( | ||
| scaling_granularity=ScalingGranularity.AXISWISE, dtype=e4m3_dtype |
Contributor
There was a problem hiding this comment.
nit: maybe we can also add some context in the comments on L353:L363 that it also uses e4m3 for grads?
vkuzo
reviewed
Nov 26, 2024
| NoopFwToFloat8E5M2BwDelayed, | ||
| NoopFwToFloat8E5M2BwDynamic, | ||
| NoopFwToFloat8E5M2BwStatic, | ||
| NoopFwToFloat8BwDelayed, |
vkuzo
reviewed
Nov 26, 2024
| # Calculate the new scales from the updated history stacks | ||
| new_input_scales = amax_history_to_scale_stack( | ||
| fp8_input_amax_history_stack, e4m3_dtype, x_dtype, scale_fn_recipe | ||
| fp8_input_amax_history_stack, input_dtype, x_dtype, scale_fn_recipe |
Contributor
There was a problem hiding this comment.
will likely have to rebase on top of #1329 which changed this line
vkuzo
reviewed
Nov 26, 2024
| static_scale: Optional[torch.Tensor] = None | ||
| dtype: Optional[torch.dtype] = None | ||
|
|
||
| def short_str(self): |
Contributor
There was a problem hiding this comment.
can we also add the dtype here, so it appears when we print an instance of Float8Linear? Float8Linear.__extra_repr__ calls this method.
Contributor
|
This is great! LGTM, had some comments but all are pretty nitty. CI is green - ship it! |
Contributor
Author
|
Superseded by #1378 |
yanbing-j
pushed a commit
to yanbing-j/ao
that referenced
this pull request
Dec 9, 2024
This file contains hidden or 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
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
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.
Stack from ghstack (oldest at bottom):