added fixes for handling multiple shape warmup #13
Conversation
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
…dded tests for multiple shape warmup Signed-off-by: Joshua Rosenkranz <jmrosenk@us.ibm.com>
ani300
reviewed
Apr 1, 2025
|
|
||
| for layer in kwargs["past_key_value_states"]: | ||
| for tensor in layer: | ||
| torch._dynamo.mark_static(tensor, 0) |
There was a problem hiding this comment.
we could move the mark kv cache sequence dimension as dynamic code here as well
ani300
reviewed
Apr 1, 2025
Comment on lines
27
to
28
| torch._dynamo.mark_dynamic(kwargs["mask"], 1) | ||
| torch._dynamo.mark_dynamic(kwargs["mask"], 2) |
There was a problem hiding this comment.
we probably only need to mark the dim 2 as dynamic here
ani300
reviewed
Apr 1, 2025
Comment on lines
+17
to
+23
| torch._dynamo.mark_static(input_ids, 0) | ||
| torch._dynamo.mark_static(input_ids, 1) | ||
| torch._dynamo.mark_static(kwargs["mask"], 0) | ||
| torch._dynamo.mark_static(kwargs["mask"], 1) | ||
| torch._dynamo.mark_static(kwargs["mask"], 2) | ||
| torch._dynamo.mark_static(kwargs["position_ids"], 0) | ||
| torch._dynamo.mark_static(kwargs["position_ids"], 1) |
There was a problem hiding this comment.
do we need to mark all the sequence dimensions as static or is just the batch dimensions enough?
There was a problem hiding this comment.
It probably is enough, however I marked everything as static to ensure we get a static prefill -- I believe symbolic ints can cause changes in the graph in prefill that we may not want to introduce.
Signed-off-by: Joshua Rosenkranz <jmrosenk@us.ibm.com>
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Outdated suggestions cannot be applied.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 PR is to address issues with warming up multiple shapes on the AIU. This PR introduces the use of a
prepare_model_inputs_hookwhich will mark certain dimensions as static/dynamic prior to forward pass. This relies on the following PR foundation-model-stack/foundation-model-stack#388