Phase 2 expanded prefill CP persistent KV capacity by sharding KV at rest, but its compatibility runtime can still rebuild full-view KV/index buffers from sharded storage. This documents a narrower Phase 3 target: reuse current chunk KV/index that was already CP all-gathered and reranged, while leaving deeper history shard-aware work to Phase 4.
The new plan separates current-only and mixed current/history cases, defines fallback requirements to the Phase 2 full-view compatibility path, and records validation criteria before any runtime implementation starts.
Constraint: Phase 3 must reduce avoidable duplicate materialization without rewriting NSA topk or sparse attention kernels
Constraint: History KV/index access remains Phase 2 compatibility materialization until Phase 4
Rejected: Start with shard-aware topk/distributed attention | too broad for Phase 3 and harder to validate against the current Phase 2 baseline
Rejected: Remove all shared-KV runtime materialization immediately | history/prefix/chunked-prefill paths still need a compatibility fallback
Confidence: high
Scope-risk: narrow
Directive: Keep Phase 3 scoped to current chunk reuse; put selected-history and shard-aware algorithms in Phase 4 unless the design is explicitly revised
Tested: git diff --check for the two documentation files
Not-tested: Runtime behavior; documentation-only change
Prefill CP previously replicated NSA/MLA persistent KV on every CP rank, so CP8 consumed eight copies of KV memory while exposing only one rank of logical cache capacity. This change splits logical KV locs from per-rank physical storage, shards MLA latent KV and NSA index K/scale by deterministic page ownership, and keeps existing NSA attention kernels working through a full-view runtime materialization layer.
Mooncake PD transfer now sends each prefill CP rank's owned physical pages with explicit logical page positions so non-CP decode can reconstruct full-layout KV. The implementation is guarded by an explicit server flag and startup checks, and the design documentation records the implemented scope, debug environment, and Phase 3 boundary.
Constraint: Phase 2 must preserve existing NSA attention/index kernels via runtime full-view materialization
Constraint: Decode side remains non-CP and receives full KV through Mooncake
Rejected: Shard-aware NSA attention in this change | belongs to Phase 3 because it requires distributed topk/softmax/output contracts
Rejected: Request-contiguous CP ownership | unstable under chunked prefill and tied to attention split mode
Confidence: medium
Scope-risk: broad
Directive: Do not enable round-robin CP shared KV without wiring runtime materialization/PD transfer contracts for that split mode
Directive: Keep SGLANG_DEBUG_CP_SHARED_KV disabled for perf measurements; it intentionally enables CUDA-syncing diagnostics
Tested: Remote py_compile for shared-KV touched Python files in g0034 container
Tested: Remote pytest selected cp_shared/shared_kv/nsa suite: 37 passed, 34 deselected
Not-tested: Full GLM5 multi-node throughput/regression run after final doc update
Not-tested: Phase 3 shard-aware runtime, round-robin CP mode, and non-Mooncake PD backends
- nsa_prefill_cp_all_gather.md: explains why there are two all-gathers
in current NSA prefill CP (per-layer KV/index gather vs tail output
hidden gather), with data flow diagrams and involved file lists
- nsa_prefill_cp_phase1_narrow_output_collection.md: detailed Phase 1
design plan covering batch-level eligibility, fallback conditions,
affected modules, KV transfer impact, risks, and test matrix
Replace full hidden all-gather at prefill tail with per-request last-token
hidden collection, reducing communication from total_tokens x hidden_size
to cp_size x bs x hidden_size for both in-seq-split and round-robin modes.
- nsa/utils.py: add cp_collect_last_token_hidden() with mode-specific
narrow collection helpers that only gather the last token hidden
- deepseek_v2.py: add _should_use_narrow_output_path() gate on
DeepseekV2Model, fallback to full gather for EAGLE/return_logprob/
capture_hidden batches
- logits_processor.py: add _is_compact_hidden_states() to bypass
_get_pruned_states() when hidden is already compact
