Repository navigation
Releases
Releases · sameer-sde/phantom-kv
v0.7.0: Phantom complete
All six phases of the roadmap are done.
- Paper (PDF and LaTeX source in
docs/paper/): the method, the
evaluation with planted and unplanted results kept separate, the two
LMCache case studies, limitations and related work. - Blog post (
docs/blog/phantom.md). - PyPI page now uses a short description written for PyPI
(docs/pypi.md); the README's Mermaid diagrams showed as code there. - README: a "Results at a glance" table at the top; the opening line now
names only the engines actually tested (vLLM and LMCache). - CI and the publish workflow moved to current action versions (Node 24),
which removes GitHub's Node 20 deprecation warnings.
v0.6.1: first PyPI release
- On PyPI:
pip install phantom-kv, thenphantomkv fuzz .... No
dependencies. - Published from GitHub Actions with PyPI Trusted Publishing (no stored
token) when a GitHub release is created; the workflow checks that the
version matches the tag, runs the tests andtwine checkfirst. - Package metadata (keywords, classifiers, project links); README links are
absolute so they work on the PyPI page.
v0.6.0: bug hunting on vLLM offloading and LMCache
Three targets, each run with its own code unmodified. vLLM's two CPU
offloading connectors: clean, planted faults caught, storage failures
recovered. LMCache: two real, open bugs found independently and shrunk to
minimal reproductions (both already reported upstream, so no duplicates
filed); Phantom's A/B tests show both open fixes work. SGLang, llm-d and
the NIXL fault classes were not covered (they need GPUs).
- vLLM's
SimpleCPUOffloadConnectoron the CPU backend: its manager runs
unmodified, only the CUDA copy engine is replaced. Sixvllm-fuzz
profiles (cpu-offload*); 0 violations so far, 3,753 blocks stored and
317 loaded back and verified. - Offload copy engine with separate load/store queues (like vLLM's two CUDA
streams); plantedstore-reported-earlycaught; offloading in CI. phantomkv vllm-diff: run a seed twice, show the first step where the
runs differ. Found that the harness stopped stepping while vLLM still had
pending connector work; fixed.- vLLM's
OffloadingConnector(target 2) through its pluggable
OffloadingSpec, no vLLM code patched:offload2*profiles, planted
store-reported-earlycaught, and storage failures (load-fails,
store-fails) recovered correctly by vLLM; in CI. - Partial-chunk evidence, revisit-heavy chunk profiles and
scripts/offload_campaign.sh: 50 seeds per run, 8 profiles clean, planted
fault caught 24/50, storage failures recovered. README Phase 5 section. - LMCache (target 3):
scripts/lmcache_smoke.py, vLLM + LMCache's
LMCacheMPConnector+ a CPU-only LMCache server; 928 tokens stored and
loaded back through LMCache, verified by Phantom. - LMCache under
vllm-fuzz(lmcache*profiles, one server per campaign,
restart fault, planted bug in LMCache's scatter path caught 10/10).
Lead: after a fast server restart vLLM never re-registers with LMCache. lmcache-restart-settleprofile. The restart lead is a known LMCache bug
(issue #4642), found independently; an A/B test of the open fix
(PR #6106) shows caching recovers after every restart.- LMCache lock and object counts after every run (from
/status),
lmcache-only-noabort, and three scripts: a one-abort probe, a lock-leak
shrinker (ddmin with a custom failure test) and a replayer for the saved
minimal cases. The Day 22 replay differences are a read-lock leak on
aborted waiting requests (known upstream, issue #5656); A/B of the open
fix (PR #6540): 0 leaked locks, all seeds identical on replay.
v0.5.0: two-engine prefill/decode with transfer faults
- Two real vLLM engines (prefill + decode) exchanging KV through vLLM's
KV-connector API; Phantom verifies every transferred slot on the decode side. PhantomFaultConnector: planted transfer faultsswap(KV of the wrong
request) andsilent-drop(load reported done, nothing written). Both caught.PhantomAsyncFaultConnector: asynchronous loads that land steps later,
one-sided like RDMA.abort-free-raceplant (the SGLang Oct-2026 bug class): an aborted request's
blocks are freed while its KV is still in flight. Unmodified vLLM stays clean
(4 late landings handled); with the plant, 54misplaced-blockviolations and
5 corrupted requests.- Seeded PD fuzzing (
phantomkv pd-fuzz,pd-shrink): each seed picks the
load delay, transfers, abort timings and fillers. Unmodified vLLM: 0/20
failing with 35 aborts while KV was in flight.abort-free-race: 14/20
caught; a failing seed shrinks from 24 events to 2. - Fourth transfer fault,
duplicate: a stale copy of a KV write lands after
the request's blocks may have been reused. scripts/pd_campaign.sh, 200 seeds per mode: unmodified vLLM 0/200;
abort-free-race115,swap199,silent-drop199,duplicate184
caught. The seed swap and silent-drop missed read no transferred KV at all.- All two-engine runs are part of the
real-vllmCI job.
v0.4.0: profiles: chunked prefill, preemption, block sizes
- vLLM configuration profiles:
baseline,chunked-prefill,preemption,
block64, with engine counters proving each stressed path was exercised. - Second planted bug class,
missing-touch(refcount). Found that vLLM's own
assertions catch bugs that make its bookkeeping inconsistent; Phantom is
needed for bugs that leave it consistent but wrong. - In-process engine for exact step ordering and replay; honest verdicts
(PASS / CAUGHT / VLLM-ASSERT / MISSED / CRASH) invllm_campaign.sh. - Engine errors recorded with file, line and source.
v0.3.0: Phantom inside real vLLM, with CI
PhantomForCausalLM: a weightless model registered through
vllm.general_pluginsthat writes provenance tags into vLLM's real KV cache
through its block tables and slot mappings, and verifies every read.- Float32-exact tag encoding.
- Planted
stale-prefix-cachebug in vLLM's block pool, caught on 15/20 seeds. - Seeded step-by-step harness (
phantomkv vllm-fuzz) and shrinking of
real-vLLM failures (phantomkv vllm-shrink). real-vllmCI job running Phantom against a pinned vLLM CPU wheel.
v0.2.0: shrinking failing runs to minimal repros
- Delta-debugging minimiser (ddmin): a failing seed becomes a minimal,
1-minimal event list that still fails. phantomkv shrinkcommand.
v0.1.0: oracle and deterministic simulator
- Provenance tags and prefix chain hashes; an exact oracle with four
violation kinds (unwritten-read,misplaced-block,chain-break,
slot-mismatch). - Toy paged-KV engine with prefix cache, preemption, PD transfer and aborts.
- Deterministic simulator driven by a single seed;
fuzzandreplaycommands. - CI running unit tests and a 1,000-seed clean fuzz on every push.