Fix LMCache offload reload handoff#1026
Draft
yhl-amd wants to merge 11 commits into
Draft
Conversation
…WIP) New atom/kv_transfer/offload/ package: a standalone (non-vLLM) KV offload connector that reuses LMCache as a storage tier (CPU LRU + NVMe L3) via StorageManager + ChunkedTokenDatabase, with an opaque per-block byte codec (ATOMKVByteCodec) that bypasses engine.store/retrieve to preserve AITER's swizzled KV layout. Load/save run daemon-after-forward off the RPC thread; cross-process hit lookup via LMCache's ZMQ LookupClient/LookupServer. Engine hooks: - disaggregation/factory.py: register the "lmcache_offload" connector. - model_engine/scheduler.py: offload-wake branch (a parked WAITING_FOR_REMOTE_KVS seq resumes as a suffix prefill, not the P/D decode-jump) + offload_resume guard against re-allocate-on-populated-blocks. TP=2 fixes (offload was non-functional at TP>1): - config.py: lookup_server_worker_ids=[0]. The cross-rank ZMQ lookup took min() across ranks and rank!=0 returned 0 despite having stored the chunk (contains()=True) -> min(0,hit)=0 -> the load never fired. Only rank 0 answers lookup now (both ranks save in lockstep, so rank 0 is authoritative). - connector.py: split load/save into separate executors so a latency-critical reload never queues behind the fire-and-forget save backlog. Verified end-to-end at TP=2 (MiniMax-M2.5 FP8, 2x MI325X): an evicted 32K prompt reloads 32000 tokens from CPU and recomputes only the 5-token suffix (Scheduled prefill cached:[32000], new:[5]); was a full recompute before. Known issues (WIP): - Reload latency high (~131s in the micro-bench): the per-block Python copy path is slow (~88ms/chunk) and the load waits on the storage_manager lock held by a save burst. Needs a bulk/batched copy rewrite of ATOMKVByteCodec. See ../OFFLOAD_TP2_FIXES.md and ../PHASE4_RESULTS.md. - Verbose [OFFLOAD-*] diagnostic logging + an engine.lookup monkeypatch are still present and must be removed/demoted before benchmark/production use. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
This reverts commit 77dbd02.
Apply Scheme A load semantics after allocation: skip small or unsafe CPU reloads, hand off unaligned HBM floors by prefill-to-boundary when enabled, and promote completed remote-KV waiters before fresh admissions. Add scheduler and connector tests for aligned reloads, unaligned handoff, min-load skips, and deferred-output cleanup.
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.
Summary
OFFLOAD_UNALIGNED_HANDOFFflow to prefill up to the next LMCache chunk boundary before issuing the CPU loadValidation
git diff --checkpython3 -m pytest tests/test_lmcache_offload_connector.py tests/test_scheduler.py(49 passed, 11 skipped)Notes