[luckyq]

Series map (Phase 1 harness, one PR per step; each is based on the previous one)

PR	Step	Test cases
#137	Harness core: stack, client, telemetry, reference engine, CI unit tests; phase1 job in vllm-e2e.yml on the two-GPU runner (TC-01; #138/#139 extend --cases)	TC-01 integration generation correctness
#138	Local DRAM flow and hinted peer transfer, hybrid geometry flags	TC-02, TC-03
#139	Guard protection, takeover, replacement	TC-04
#140	Streaming client, performance arms, Guard recovery timing	TC-05, TC-06
#141	Sidecar kvcr_service, 2P+1D, two nodes, k8s / Slurm / Docker runners; phase1-2p1d CI job (2P+1D on the two-GPU runner, prefill workers share GPU 0)	all cases on those topologies
#142	docs/ci.md: what the two CI lanes verify, how to run them elsewhere, how to read evidence	—
#143	Test plan, run history, bare-vLLM probes	—
#180	--engine sglang: dynamo.sglang + HiCache KVCR storage backend (quick-start config), SGLang telemetry mapping, deterministic mode; no Guard/disagg	TC-01..03, TC-05 on SGLang

Replaces #103. All seven are drafts until the per-PR GPU runs on the current main image are done.

Problem

• Part 4 of the stacked Phase 1 series (on top of TC-04). The functional cases prove correctness; nothing yet measures what KVCR costs or how fast the Guard recovers.

What this adds

• tests/e2e/phase1/perf.py with streaming requests (Client.stream_completion: client-side TTFT, TPOT, per-chunk timing, worker identity) and the KVCR-disabled cold/warm baseline recorded in the reference phase.
• TC-05 performance arms: cold, KVCR local hit, DRAM-pool hit after evicting the GPU prefix cache (--perf-evict-filler), remote hit through a peer transfer, plus a concurrency sweep (--perf-concurrency-levels). Every stream must reproduce the reference tokens; transfer telemetry must reconcile (remote_deliver == source_write, bytes equal, exact pages for hybrid geometry). No latency thresholds in this round; deltas against the baseline are recorded.
• TC-06 Guard recovery timing: t0 engine kill, t1 first correct Guard-served response (must be covered by remote_deliver on the target; attempts rotate over prompts the target has not seen so a local re-commit cannot pass; a token mismatch fails), t2 replacement serving; post-recovery hits from the reclaimed pool.
• Unit tests for the Guard-source and attempt-outcome verdicts.

Validation

• ruff, pytest tests/ci: see CI.
• GPU: TC-05/06 passed on 2×H100 with KVCR v0.1.1 (dense Qwen3-8B: Guard-served 23 pages, t1 1.4 s; hybrid Qwen3.5-4B: 5 pages = 2 + 3 SSM, t1 1.2 s) in the test: KVCR Phase 1 end-to-end harness (CI image build, TC-01..06, dlcluster and K8s runners) #103 runs.

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.