The eviction-triggered D2h offload reads KV blocks straight from the#235
Open
copybara-service[bot] wants to merge 1 commit into
Open
The eviction-triggered D2h offload reads KV blocks straight from the#235copybara-service[bot] wants to merge 1 commit into
copybara-service[bot] wants to merge 1 commit into
Conversation
copybara-service
Bot
force-pushed
the
test_940532973
branch
2 times, most recently
from
ab73301 to
c122f9f
Compare
device buffer; without ordering it can read a block mid-write while the forward is still updating it in place, corrupting the offloaded copy. Add a wait_ready parameter to D2h: when set, the torch KVCacheManager re-resolves each involved layer/shard's live KV buffer and blocks on its PJRT readiness event before the copy, so the store sees settled bytes. AwaitBufferReady uses the stable PJRT C ABI (PJRT_Buffer_ReadyEvent + PJRT_Event_Await), not xla::Future, whose C++ ABI mismatches across the independent raiden/torch_tpu builds. The base D2h ignores wait_ready (its registration holds carry a retired handle); only the torch override, which retains the live device tensors, honors it. PiperOrigin-RevId: 940532973
copybara-service
Bot
force-pushed
the
test_940532973
branch
from
c122f9f to
4195164
Compare
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
1 participant
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
The eviction-triggered D2h offload reads KV blocks straight from the
device buffer; without ordering it can read a block mid-write while the
forward is still updating it in place, corrupting the offloaded copy.
Add a wait_ready parameter to D2h: when set, the torch KVCacheManager
re-resolves each involved layer/shard's live KV buffer and blocks on its
PJRT readiness event before the copy, so the store sees settled bytes.
AwaitBufferReady uses the stable PJRT C ABI (PJRT_Buffer_ReadyEvent +
PJRT_Event_Await), not xla::Future, whose C++ ABI mismatches across the
independent raiden/torch_tpu builds. The base D2h ignores wait_ready (its
registration holds carry a retired handle); only the torch override,
which retains the live device tensors, honors it.