[None][feat] Unified target/draft KV cache for DFlash (port of #20178, #20486 and #20709) by zheyuf · Pull Request #20806 · NVIDIA/TensorRT-LLM

[TRTLLM-16722][feat] Dflash Unified KVCache (port of NVIDIA/TensorRT-LLM#20486 incl. #20178 foundation)

allisonlim-nv

authored and

zheyuf

committed

1a1b710

[None][feat] Per View KV head transfer (port of draft NVIDIA/TensorRT-LLM#20709)

allisonlim-nv

authored and

zheyuf

committed

6b40260

[None][chore] Run the unified KV cache stack on the agentx KV cache bindings

zheyuf

committed

0e55a6e

[None][fix] Bound DFlash's shared context by the allocation's extent

zheyuf

committed

ff11364

[None][fix] Keep DFlash's dummy slot empty between steps

zheyuf

committed

7be69ca

[None][fix] Re-seed DFlash generation slots in CUDA-graph warmup

zheyuf

committed

2b23449

[None][test] Align the unified DFlash KV gate tests with per-view head transfer

zheyuf

committed

9862d63

[None][fix] Refuse an unshared windowed draft pool on disaggregated workers

zheyuf

committed

400ff30

[None][test] Expect the disaggregated DFlash warning only without a shared draft KV

zheyuf

committed

93bec95

[None][perf] Convert only the batch's rows of per-layer KV page tables

zheyuf

committed

70fbd34

[None][perf] Share attention-op page tables across layers of one KV pool

zheyuf

committed

87bb9b0

[None][feat] Let DFlash AUTO pick TRTLLM on workers that share the draft KV

zheyuf

committed

8abf578

[None][test] Use KVCacheManagerV2 in the unified draft KV disagg test

zheyuf

committed

f01094b

[None][fix] Fail a disaggregated request at once when its peers disagree on draft KV layers

zheyuf

committed

3abddc3

[None][fix] Share the DFlash draft KV under disaggregated_mpi_worker with the default AUTO backend

zheyuf

committed

d59b145

Navigation Menu

Repository navigation

[None][feat] Unified target/draft KV cache for DFlash (port of #20178, #20486 and #20709) - #20806

[None][feat] Unified target/draft KV cache for DFlash (port of #20178, #20486 and #20709)#20806
zheyuf merged 23 commits into
NVIDIA:feat/musespark-open-agentxfrom
zheyuf:musespark-open/dflash-unified-kv-stack

Commits on Oct 24, 2026

[TRTLLM-16722][feat] Dflash Unified KVCache (port of NVIDIA/TensorRT-LLM#20486 incl. #20178 foundation)

[None][feat] Per View KV head transfer (port of draft NVIDIA/TensorRT-LLM#20709)

[None][chore] Run the unified KV cache stack on the agentx KV cache bindings

[None][fix] Bound DFlash's shared context by the allocation's extent

[None][fix] Keep DFlash's dummy slot empty between steps

[None][fix] Re-seed DFlash generation slots in CUDA-graph warmup

[None][test] Align the unified DFlash KV gate tests with per-view head transfer

[None][fix] Refuse an unshared windowed draft pool on disaggregated workers

[None][test] Expect the disaggregated DFlash warning only without a shared draft KV

[None][perf] Convert only the batch's rows of per-layer KV page tables

[None][perf] Share attention-op page tables across layers of one KV pool

[None][feat] Let DFlash AUTO pick TRTLLM on workers that share the draft KV

[None][test] Use KVCacheManagerV2 in the unified draft KV disagg test

[None][fix] Fail a disaggregated request at once when its peers disagree on draft KV layers

[None][fix] Share the DFlash draft KV under disaggregated_mpi_worker with the default AUTO backend

Commits on Oct 26, 2026

[None][feat] Share the DFlash draft KV on disaggregated workers by default

[None][fix] Scatter all fan-in writers of a KV-bounce region in one batched copy

[None][fix] Allow TP fan-in KV bounce with non-uniform layer-group block sizes

[None][test] Run the DFlash and disaggregated KV transfer tests in the MuseSpark pre-merge stage

Commits on Oct 27, 2026

[None][test] Run the cpu_only DFlash and KV transfer tests in the MuseSpark CPU stage

[None][fix] Check the senders' head duplication before a fan-in KV bounce

[None][fix] Keep an explicit TRTLLM_DISAGG_SHARED_DRAFT_KV=0 under disaggregated_mpi_worker

[None][chore] Name BAD_PAGE_INDEX in the op-table kernel and walk warmup offsets once