[Unified Cache][4/N]: Add Mooncake backend for external linker - #37205
Conversation
Co-authored-by: hzh0425 <hzh0425@apache.org>
|
/tag-and-rerun-ci |
|
/rerun-test test/registered/unit/mem_cache/test_unified_cache_linker.py |
|
/rerun-group radix_cache/unified_radix_tree |
|
Results for 🚀 |
|
Results for 🚀 🚀 🚀 🚀 🚀 |
|
/rerun-group hicache |
|
Results for 🚀 🚀 🚀 🚀 ⛔ |
|
/rerun-test test/registered/unit/mem_cache/test_mooncake_direct_linker.py |
|
/rerun-group radix_cache/unified_radix_tree |
|
Results for 🚀 |
|
Results for 🚀 🚀 🚀 🚀 🚀 |
|
/rerun-failed-ci |
|
/rerun-test registered/radix_cache/unified_radix_tree/test_unified_radix_cache_kl_dsv4.py test_unified_radix_cache_kl_dsv4_pp.py |
|
Results for 🚀 🚀 |
Wires the Mooncake direct external linker (PR sgl-project#37205 backend, sgl-project#37307 translator passthrough) into the serving stack as an opt-in direct-L3 mode for UnifiedRadixCache, with no host cache tier: - server_args: --enable-unified-cache-external-linker (default FALSE, runtime stays off unless explicitly enabled) + --unified-cache-external-linker-backend {mooncake}. - arg_groups/hicache_hook: mutual-exclusion guard against --enable-hierarchical-cache and --hicache-storage-backend; linker mode skips hicache normalization entirely. - registry: route to _create_unified_radix_cache under the flag and attach MooncakeDirectLinker via init_cache_linker, registering its layer-done counter with the kv cache and tp worker. - scheduler: consolidate aborted-request cache-state release into _release_aborted_request (fires for hicache storage or the external linker); poll linker events, and gate NO_TOKEN / hicache-consumer batch behavior under the linker flag. - kv_cache_builder: forward attn_cp_rank / attn_cp_size into CacheInitParams (consumed by the storage config suffix). - kv_cache_configurator: keep the DSA indexer-K cache under the linker. - full_component: honor the PREPARE phase of ExternalLinkerLoadPhase (return the transfer unmodified; assert COMMIT after it). - tests: MooncakeTestServices lifecycle helper, DSV4-Flash and GLM-5.2 direct-linker KL E2E suites (registered as extra-b CUDA CI), plus the unit-test fixtures for the new scheduler/registry fields. Applied cleanly over the local branch (no textual conflicts; scheduler abort consolidation coexists with the local chunked-prefill abort re-dispatch and fail-soft changes). Default-off runtime behavior is preserved per the port-but-don't-switch decision. (cherry picked from commit d2ade53, PR sgl-project#37381)
…roject#37205) Co-authored-by: hzh0425 <hzh0425@apache.org>
Motivation
This is the fourth PR split from #35687.
Previous splits: #37091 and #37098 and #37151
Modifications
Accuracy Tests
Speed Tests and Profiling
Checklist
Review and Merge Process
/tag-and-rerun-ci,/tag-run-ci-label,/rerun-failed-ciCI States
Latest PR Test (Base): 🚫 Run #33383052227
Latest PR Test (Extra): ✅ Run #33383051768
Latest PR Test (AMD ROCm 7.2): ❌ Run #33383052171