Generate: multi-device support for contrastive search #24635
Merged
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Outdated suggestions cannot be applied.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?
Fixes #24634
In multi-gpu settings, the past KV cache may be scattered across devices -- the cache corresponding to a layer sits in the same device as the layer itself, and different layers may be in different devices.
In contrastive search, we must apply indexing operations on the past KV cache. The indexes are in a tensor, which sits on the same device as the model outputs by default. Applying these indexes on the past KV cache currently results in an exception if the model is split across devices (see the issue linked above).
This means we either move the indexing tensor to all possible devices or keep the tensor on CPU. Indexing is typically CPU-heavy on PyTorch, so the benchmarks on my end indicate that moving the indexing tensor to the CPU enables multi-device contrastive search without noticeable throughput degradation 🙌