implementation detail · filed under inference & serving
Asynchronous cache offloading and restoration
Runs cache offloading and restoration on side CUDA streams rather than the compute stream to avoid stalling generation.
- source
- 1
- models
- 2
- lab adopt it
- 1
- strongest
- used
How sources treat it
One count per evidence span, weakest treatment to strongest.
used 1
Documented in
Evidence
1 span quoted from the sources, strongest treatment first.
Offloading and restoration run on side CUDA streams rather than the compute stream, never stalling generation.
usedinference servingin MiMo-V2.6Xiaomi
Filed alongside
Other methods under inference & serving :: KV cache management.
Chunked prefillPrefix cachingPrompt cachingSWA Bounded ReplayLanguage-model-only serving modePeriodic cache checkpointingCompressed KV cachingKDA-aware prefix-cache managementRadixCacheZero SWA caching8-bit Mamba cache quantizationAutomatic cacheBlock-based KV cacheCross-layer KV-cache reuseCustomized heterogeneous KV-cache layoutDisable prefix caching for benchmarkingDynamic allocation of fixed-size state-cache poolsEncoder SWA bounded replayExact SWA KV reconstruction via full multi-layer replayFine-grained prefix hashingHierarchical KV cachingInference-side KV-cache reset on weight synchronizationKDA with prefill cacheKey-value reuse in global attention layers