KVWarden：单GPU多租户公平调度，无需Kubernetes的LLM推理编排层

KVWarden is a lightweight orchestration layer (≈3500 lines of code) running on vLLM/SGLang. It addresses multi-tenant fairness issues in LLM inference by token-bucket rate limiting, supports single-GPU multi-model lifecycle management (frequency+recency strategy), and provides an OpenAI-compatible HTTP API. It eliminates the need for Kubernetes, making it ideal for small teams or edge deployments.

When multiple users/apps share a GPU for LLM inference, resource allocation fairness becomes critical. For example, a noisy neighbor (32 RPS) and quiet user (1 RPS) on Llama-3.1-8B: without fairness control, quiet user's TTFT p99 jumps from 53.9ms to 1585ms (29x increase). This is the problem KVWarden solves.

KVWarden adds three key features to vLLM/SGLang:

1. Tenant-level token-bucket rate limiting: Controls request rate at entry, reducing quiet user's TTFT p99 to 61.5ms (14% higher than no competition).
2. Single-GPU multi-model lifecycle management: Uses frequency+recency for model switching and cache eviction (smarter than LRU).
3. OpenAI-compatible HTTP API: Existing apps can integrate without code changes.

Core components:

• WorkloadRouter: Request analysis, length-aware scheduling, OpenAI API, streaming support.
• AdmissionController: Concurrency limit, priority queue (lower number = higher priority), Prometheus metrics.
• TenantManager: Tenant budget management, token-bucket implementation, DRR priority scoring.
• CacheManager: Model KV cache lifecycle, snapshot on unloading, layered eviction. Request flow: Client → WorkloadRouter → TenantManager (rate limit check) → AdmissionController (queue/admit) → CacheManager (model/KV) → vLLM/SGLang. Requests use X-Tenant-ID header for tenant identification.

Experiments on A100-SXM4 80GB with Llama-3.1-8B (vLLM):

1. Fairness test: 32 RPS noisy +1 RPS quiet. Token-bucket reduces quiet user's TTFT p99 from 1585ms (29x) to61.5ms (1.14x).
2. Admission cap test: Global concurrency limits don't improve single-model performance (vLLM's batching is efficient).
3. Benchmark framework validation: End-to-end tests with real vLLM ensure no systematic bias.

KVWarden fills a gap as a no-Kubernetes single-node multi-tenant fair scheduler:

System	Needs K8s	Multi-model	Tenant Fairness	Scenario
NVIDIA Dynamo	Yes	Yes	No	Data center
llm-d (CNCF)	Yes	Single pool	No	Cloud-native large scale
Mammoth	Yes	Yes	No	Multi-hardware
AIBrix	Yes	Yes	No	Enterprise
Ollama	No	LRU eviction	No	Local single node
vLLM/SGLang	No	Single model	No	Basic inference
KVWarden	No	Yes (freq+recency)	Yes (token-bucket+DRR)	Single node (1-4 GPUs)

Quick start: pip install kvwarden kvwarden serve --config configs/quickstart_fairness.yaml Wait for health check: until curl -fs localhost:8000/health > /dev/null; do sleep2; done Send requests as tenants: Noisy: curl localhost:8000/v1/completions -H "X-Tenant-ID: noisy" -d '{"model":"llama31-8b","prompt":"...","max_tokens":64}' Quiet: curl localhost:8000/v1/completions -H "X-Tenant-ID: quiet" -d '{"model":"llama31-8b","prompt":"...","max_tokens":64}' Config example (yaml): tenants with rate limits, models with engine/gpu memory. CLI tools: version, doctor, man, bench.

Limitations: Not a vLLM/SGLang replacement; not for K8s-scale deployments; no magic single-tenant TTFT improvement. 适用场景: Single node (1-4 GPUs), no K8s, multi-tenant fairness needed, edge/local. Future plans: 近期: 8-tenant test, Llama-3.1-70B on4×A100, Mixtral MoE fairness. v0.2.x: Multi-engine routing (vLLM ↔ SGLang). v0.3: KV cache分层 (LMCache), 32K context fairness.