RELIABILITY LAYER · STATEFUL INFERENCE

Memory woven
to survive.

Inference became a distributed, stateful system. Minnesväv is its reliability layer. When a GPU node dies forty minutes into an agent run, the session re-weaves onto healthy hardware in seconds — no failed request, no re-paid prefill, no lost work.

<10s
resume · p99
<2%
serving overhead
0
tokens recomputed
50–70%
saved on spot GPUs

// design targets, enforced as release gates in our public chaos suite

The problem

When a node dies mid-session, everything burns.

Disaggregated serving, tiered KV caches, and hour-long agent runs turned every request into a multi-node, stateful pipeline. The stack moves bytes brilliantly — and defines nothing about what happens when a machine fails.

// sessions lost

The work dies with the node

A preemption or OOM destroys the entire accumulated context of an agent run. The request fails, the prefill is re-paid from zero, and forty minutes of compute evaporates.

// failures invisible

Dashboards stay green

TTFT spikes from cache eviction, tier thrash, and fragmentation — causes no app-layer trace or GPU counter can see. Diagnosis runs on vibes.

// capacity locked

Spot GPUs sit off-limits

Preemptible capacity is 50–70% cheaper and unusable for stateful serving. Everyone over-pays for on-demand because failure means starting over.

The product

One platform. Three guarantees.

RESUME

Sessions that survive

Append-only session state — KV blocks, generation cursor, sampler state — streams continuously to a buddy node over NIXL. Kill the machine; the session re-attaches and keeps decoding from the last committed token.

epoch fencing · zero re-prefill · exactly-once ownership
SEE

The missing trace layer

Every TTFT spike attributed to its real cause — eviction storms, tier misses, transfer stalls, fragmentation — joined per session, between the app traces and the GPU counters where nothing can see today.

OpenTelemetry-native · self-hosted store · per-tenant attribution
PREDICT

Caches warmed before demand

A behavioral engine learns who queries which documents and when — pre-warming context off-peak, inheriting document ACLs into the cache, and invalidating surgically when sources change.

phantom prefill · ACL projection · Merkle invalidation
How it works

Snapshot. Detect. Resume.

Inference state is append-only — strictly simpler to replicate than training state. We exploit that.

01

Snapshot

A per-node daemon streams finalized KV blocks and generation state to a topology-aware buddy over NVLink or RDMA — zero-copy, paced under 5% of link capacity, invisible to the engine.

02

Detect

Gossip heartbeats plus control-plane leases discover failure in under a second. Cloud preemption notices trigger graceful drain — planned handoff, effectively zero loss.

03

Resume

The session is fenced, the replica rehydrates into a healthy engine, the block table re-maps, and decoding continues mid-stream. Your client's connection never even closes.

Built on, not against
  • NVIDIA Dynamo
  • LMCache
  • NIXL
  • vLLM
  • SGLang
Who it's for

Everyone who serves models at scale.

Agent platforms

Long runs, brutal failure math

Ninety-minute autonomous runs with 100K+ tokens of context die on a single node fault. Minnesväv makes the run continue — and gives your framework hooks for tool-call reconciliation.

“The agent finished the task. Nobody noticed a node had died.”
Inference providers

SLAs you can actually sign

Session-level continuity and state-layer diagnosis turn p99 mysteries into attributed causes — and failed-request refunds into a line item you retire.

“TTFT spiked because tenant X's burst evicted tier one. Now you know.”
GPU clouds

Sell spot at a premium

Preemption becomes a drain-and-migrate, not an outage. Offer reliable spot inference as a new SKU — priced above raw spot, still far below on-demand.

“Higher utilization on hardware you already own.”
Regulated enterprise

On-prem, ACL-aware, air-gapped

Document caches inherit your DMS permissions, invalidate when documents change, and never leave your perimeter. Production behaves like the demo — at 5pm on month-end.

“Same cost, same latency, no matter how many people are querying.”
Open source first

Kill a node. Watch it not matter.

The resumability plugin for vLLM ships free and open source, with the chaos suite we test it with. Run the kill-and-recover demo on your own fleet in an afternoon — then decide if the platform is worth talking about.

# protect a vLLM fleet
$ vav install --engine vllm --policy intra-domain
# prove it
$ vav chaos kill-node gpu-07 --expect-resume
✓ 312 sessions fenced, rehydrated, resumed
✓ p99 resume 3.4s · tokens recomputed: 0
# see why anything was ever slow
$ vav session inspect 9f2c --attribute-latency
Pricing

Land free. Scale per GPU.

Open source

Resume plugin

free, forever

Session resumability for vLLM in a single failure domain, the resume-correctness oracle, and the chaos CLI. Apache-2.0. Run it on anything.

Get the plugin
Platform

Fleet reliability

per GPU / month

Cross-domain protection, graceful spot drain, state-layer observability with latency attribution, multi-tenant fencing and encryption, predictive cache orchestration, SLAs.

Talk to us
GPU clouds

Reliable-spot program

revenue share

Launch a reliable spot inference tier on your existing capacity. We integrate, you price the SKU, we share the uplift. Zero engineering lift on your side.

Partner with us

Inference that survives.

We're onboarding ten design partners — agent platforms, inference providers, and GPU clouds — ahead of the open-source release. Bring a fleet and a failure story.

[email protected] Request chaos-suite access