LLM Acceleration: KV-cache and Semantic Cache Tuning

Accelerating LLMs: Tuning KV-cache and Semantic Cache

Picture this: your LLM service processes queries, and 40% of them are repetitive questions. Each time the model runs a full autoregressive inference: loading weights, computing keys and values (self-attention), generating tokens one by one. GPU hours burn, latency spikes to 5 seconds. We solve this with two-level caching — Semantic Cache at the application level and KV-cache at the model level. The combination yields up to 70% savings on inference costs and reduces p99 latency 5x. Our proven methodology, backed by 15+ projects, guarantees a 5x latency reduction and up to 70% GPU cost savings. For example, a client saved $12,000/month after implementing our caching solution.

Caching is not an option but a necessity in production. Without it, every query is recomputed even if the answer was generated a minute ago. LLM inference caching is essential for production deployments. This caching strategy is a key part of LLM optimization. We offer a turnkey solution: from audit to deployment within 5-10 business days. Contact us for a free project estimate.

How does two-level caching work?

The first level — Semantic Cache. It uses vector embeddings to find semantically similar queries. When a new query arrives, it is encoded into a vector (768-dimensional embeddings), and nearest neighbors are searched in a vector database (Qdrant, pgvector). If a record with similarity above a threshold (usually 0.85–0.95) is found, the cached response is returned without inference. This is a form of LLM response caching. Here’s a Python implementation using Sentence Transformers, Redis, and Qdrant:

from sentence_transformers import SentenceTransformer
import numpy as np
import redis
import json

class SemanticCache:
    def __init__(self, similarity_threshold: float = 0.92):
        self.encoder = SentenceTransformer("paraphrase-multilingual-mpnet-base-v2")
        self.redis = redis.Redis(host="localhost", port=6379, db=1)
        self.threshold = similarity_threshold
        from qdrant_client import QdrantClient
        self.vector_db = QdrantClient("localhost", port=6333)

    def get(self, prompt: str, system_prompt: str = "") -> str | None:
        cache_key = self._make_key(prompt, system_prompt)
        exact = self.redis.get(cache_key)
        if exact:
            return json.loads(exact)["response"]
        embedding = self.encoder.encode(prompt)
        results = self.vector_db.search(
            collection_name="llm_cache",
            query_vector=embedding.tolist(),
            limit=1,
            score_threshold=self.threshold
        )
        if results:
            cached_response = json.loads(results[0].payload["response"])
            self.redis.expire(results[0].id, 3600)
            return cached_response
        return None

    def set(self, prompt: str, response: str, system_prompt: str = "", ttl: int = 3600):
        embedding = self.encoder.encode(prompt)
        cache_id = self._make_key(prompt, system_prompt)
        self.vector_db.upsert(
            collection_name="llm_cache",
            points=[{
                "id": abs(hash(cache_id)) % (2**31),
                "vector": embedding.tolist(),
                "payload": {"prompt": prompt, "response": json.dumps(response), "system_prompt": system_prompt}
            }]
        )
        self.redis.setex(cache_id, ttl, json.dumps({"response": response}))

    def _make_key(self, prompt: str, system_prompt: str) -> str:
        import hashlib
        return hashlib.sha256(f"{system_prompt}||{prompt}".encode()).hexdigest()

The second level — KV-cache at the model level. vLLM automatically caches keys and values for common prefixes, such as system prompt, using paged attention and block-level caching in the transformer architecture. With prefix caching enabled, hit rate reaches 60–80%, reducing latency 2–5x. Compared to no caching, KV-cache reduces generation latency from 4.2s to 0.3s—over 10x faster.

# vLLM automatically uses prefix caching
# system prompt should be identical across requests

python -m vllm.entrypoints.openai.api_server \
  --model meta-llama/Llama-3-8b-instruct \
  --enable-prefix-caching \
  --max-model-len 8192

# Metric: vllm:gpu_cache_usage_perc shows cache occupancy

GPTCache — ready-made solution

GPTCache is a library that implements Semantic Cache and manages the entire infrastructure: embeddings, vector search, TTL. Integration boils down to replacing the openai call with cached_openai:

from gptcache import cache
from gptcache.adapter import openai as cached_openai
from gptcache.embedding import Onnx
from gptcache.manager import CacheBase, VectorBase, get_data_manager
from gptcache.similarity_evaluation.distance import SearchDistanceEvaluation

embedding_model = Onnx()
data_manager = get_data_manager(
    CacheBase("sqlite"),
    VectorBase("qdrant", host="localhost", port=6333, dimension=512)
)

cache.init(
    embedding_func=embedding_model.to_embeddings,
    data_manager=data_manager,
    similarity_evaluation=SearchDistanceEvaluation(max_distance=0.3),
    cache_enable_func=lambda *args, **kwargs: True
)

response = cached_openai.ChatCompletion.create(
    model="gpt-4o",
    messages=[{"role": "user", "content": "What is Python?"}]
)

Comparison of caching types

Semantic cache is 90% faster than full inference; KV-cache reduces latency 2-5x compared to no caching.

How to choose similarity threshold for Semantic Cache?

The similarity threshold is a key parameter for LLM cache tuning. Too low (0.7) leads to false positives, too high (0.98) to frequent misses. The optimal value depends on the task. For FAQ bots, 0.85–0.90 works well; for RAG with fixed documents, 0.90–0.95; for classification, 0.95+. We tune the threshold based on analysis of your data: we take 1000 real queries, label semantically equivalent pairs, and pick the threshold by F1 score. Tuning LLM cache parameters like similarity threshold and TTL is crucial for optimal response caching.

When doesn’t caching give a gain?

Caching is pointless for personalized answers, queries with current time or date, financial data (rates, prices), code generation, and when temperature > 0.8. In such cases, better disable caching. The biggest effect comes from caching in FAQ bots, RAG with fixed documents, and classification tasks.

Why is caching not a panacea?

Caching does not fix model quality. If the base model hallucinates, caching only solidifies errors. You must monitor the staleness rate — the share of cached responses that become outdated. We implement A/B tests: periodically compare cached response with a fresh inference. If divergence exceeds a threshold, the cache is invalidated.

Integration with existing infrastructure

Caching integrates easily with popular MLOps tools. vLLM and TGI support prefix caching at the inference server level. For Semantic Cache, we use Redis as a fast exact-matching cache and Qdrant or pgvector for vector search. All components are deployed in Docker or Kubernetes, monitored via Prometheus and Grafana.

Caching metrics

What's included in our caching implementation

• Architecture audit and caching strategy document
• Semantic cache setup with vector database (Qdrant/pgvector)
• KV-cache and prefix caching configuration in vLLM
• Integration with your existing LLM API (OpenAI-compatible)
• Deployment scripts and Docker/Kubernetes manifests
• Monitoring dashboards (Prometheus/Grafana) with alerting on cache hit rate and latency
• Knowledge transfer session and documentation

Our engineers have experience implementing caching in 15+ projects. Average savings: 60% on inference costs. Contact us for a free audit of your architecture — it takes no more than an hour. We’ll discuss your cache parameters and provide a turnkey solution at a fixed price.

MLOps: Infrastructure for Training, Deploying, and Monitoring ML Models

The model is trained, metrics — F1 0.94 on validation. Three months later in production, quality drops by 12%. No one knows when — there is no monitoring. It's impossible to retrain quickly — the training script is in a Jupyter notebook of a data scientist who has already left. Data for retraining is collected manually from three disparate systems. About half of the projects come to us with this pain. We build a turnkey MLOps platform: from experiment tracking to automatic deployment and data drift monitoring. We will assess your infrastructure in 1–2 weeks, and in 4–6 weeks you will get a basic MLOps core running in production. Our team has 10+ years of experience in ML infrastructure, over 50 implementations.

How does MLOps infrastructure benefit your ML projects?

Experiment Tracking and Reproducibility

Without tracking, an ML project turns into chaos: it's unclear which checkpoint is better, which hyperparameters were used, which dataset. Reproducing a result a month later is a quest.

Why is experiment tracking the foundation of reproducibility?

MLflow is an open source standard for tracking. It logs parameters, metrics, artifacts (models, graphs), and code. MLflow Model Registry is a centralized model storage with versioning and lifecycle stages (Staging → Production → Archived). Deployment via MLflow Serving or integration with external systems.

Typical initialization in code:

import mlflow

mlflow.set_experiment("fraud-detection-v2")
with mlflow.start_run():
    mlflow.log_params({"learning_rate": 3e-4, "batch_size": 64, "epochs": 10})
    mlflow.log_metric("val_f1", val_f1, step=epoch)
    mlflow.pytorch.log_model(model, "model")

This is the minimum. In production, we add logging of system metrics (GPU utilization, memory), dataset (hash, version), code (git commit hash). Weights & Biases — richer UI, collaboration features, sweep for hyperparameter optimization. MLflow — for on-premise deployment without external dependencies.

DVC (Data Version Control) — versioning of data and models on top of git. Data is stored in S3/GCS/Azure Blob, only metadata (hashes) in git. dvc repro reproduces the entire pipeline from raw data to metrics.

To ensure reproducibility of training, fix random seeds (torch.manual_seed, numpy.random.seed, random.seed) and record them in experiment metadata. Without this, debugging irregular results is painful. Log the dataset version (DVC hash) and git commit — then any experiment can be reproduced down to the byte.

Pipeline Orchestration: Kubeflow, Airflow, Prefect

A pipeline orchestrator becomes necessary when: A 100-line training script in cron is fine for simple tasks. But as soon as you have a multi-step pipeline (data loading → preprocessing → feature engineering → training → validation → deployment if quality above threshold), you need an orchestrator with retry logic, visualization, and alerts.

Kubeflow — Kubernetes-native orchestrator for ML (see Kubeflow). Each step is a Docker container. Supports parallel steps, conditional branches, artifacts between steps. Integrates with Katib (AutoML), KServe (serving), Feast (feature store).

Apache Airflow — more general DAG orchestrator. Wide ecosystem of operators (S3, Spark, DBT, Kubernetes). Easier to deploy if Airflow already exists in the company.

Prefect / Metaflow — less boilerplate. Prefect 2.x with @flow and @task decorators — quick start for small teams.

Typical training pipeline architecture on Kubeflow:

1. Data ingestion component — fetches data from S3/DB, validates schema via Great Expectations
2. Preprocessing component — transformations, normalization, train/val/test split
3. Training component — training on GPU, logging to MLflow
4. Evaluation component — metric calculation, comparison with baseline in Model Registry
5. Conditional deployment — deploy only if new model is better than current by >2% F1

Each component is a separate Docker image. Pipeline is versioned in git. Scheduled run (retraining once a week on new data) or manual.

Model Registry and Lifecycle Management

Model Registry is not just a checkpoint store. It is a centralized system that knows:

• Which model is currently in production (and with what metrics)
• History of all versions with training parameters
• Metadata: dataset, git commit, validation results
• Lifecycle stage: None → Staging → Production → Archived

MLflow Model Registry — standard. For enterprise — Vertex AI Model Registry (GCP), SageMaker Model Registry (AWS), Azure ML Model Registry.

Model promotion through stages: automatically move model to Staging after successful eval, then manual or automatic (during A/B test) promotion to Production. Rollback — switch to previous Production version in seconds.

Serving: From FastAPI to Triton Inference Server

Simple case. FastAPI + PyTorch/ONNX on one server — 80% of production ML deployments are exactly that. Sufficient for most tasks with load up to 100 req/s.

from fastapi import FastAPI
import onnxruntime as ort

app = FastAPI()
session = ort.InferenceSession("model.onnx", providers=["CUDAExecutionProvider"])

@app.post("/predict")
async def predict(request: PredictRequest):
    inputs = preprocess(request.text)
    outputs = session.run(None, {"input_ids": inputs})
    return {"label": postprocess(outputs)}

Triton Inference Server — production standard for high loads (500+ req/s). Dynamic batching, concurrent model execution, model ensemble. Supports TensorRT, ONNX, PyTorch TorchScript, TensorFlow SavedModel.

KServe — Kubernetes-native ML serving with autoscaling, canary deployments, A/B testing out of the box. Scale-to-zero for inactive models — savings on infrastructure up to 40% annually for a project with 10 models.

Monitoring: Data Drift, Model Drift, Infrastructure Metrics

Monitoring — what is usually done last and regretted first. Three levels.

Infrastructure monitoring. Latency (P50/P95/P99), throughput (req/s), error rate (4xx, 5xx), GPU/CPU utilization. Prometheus + Grafana — standard. Alert when P99 latency > threshold or error rate > 1%.

Data drift monitoring. Distribution of input data changes over time. Detect via PSI (Population Stability Index) for numerical features: PSI > 0.2 — strong drift. Chi-squared test for categorical, Kolmogorov-Smirnov test for continuous. Evidently AI — open source library with ready-made drift tests.

Model drift monitoring. If ground truth is delayed (e.g., we know conversion after a week) — monitor real metrics. If not — surrogate metrics: distribution of prediction scores, proportion of confident predictions.

Alerting. Three levels: INFO (minor drift, log it), WARNING (significant, notify team), CRITICAL (quality dropped below threshold — automatic switch to fallback model).

Why is data drift monitoring important?

Without it, you learn about model degradation only from user complaints or ringing SLA. A drift alert allows you to retrain the model in advance, before errors start causing losses. In one of our projects, PSI monitoring detected drift 2 days after a data source change — this saved the campaign.

Feature Store

Feature Store solves the training-serving skew problem. If preprocessing during training and inference is implemented in two different places — divergence is inevitable.

A Feature Store is needed when:

• Several models use the same features
• Features are computed from streaming data (real-time)
• Large team with different people on feature engineering and model training

Feast — open source Feature Store. Offline store (S3 + Parquet) for training, online store (Redis, DynamoDB) for low-latency inference. Feature definitions as code, materialization job syncs offline → online.

Tecton (commercial), Vertex AI Feature Store (GCP), SageMaker Feature Store (AWS) — managed options with less ops overhead.

CI/CD for ML

ML CI/CD is regular CI/CD plus specific ML steps.

ML-specific checks in CI:

• Reproducibility check: run training with a fixed seed, result must match
• Data validation: Great Expectations or Pandera on schema/distribution checks
• Model performance check: automatic eval on holdout, block merge if degradation > threshold
• Latency regression test: inference must meet SLA

GitOps for deployment. Merge to main → CI triggers training → eval → if passes → automatic deployment to Staging → smoke tests → manual promotion to Production or automatic upon successful canary.

Tools: GitHub Actions / GitLab CI for CI, ArgoCD for GitOps deployment on Kubernetes.

What's Included in MLOps Platform Development

We provide a full cycle of work, documentation, and team training.

Total time from audit to full MLOps platform: 3–5 months. Also possible phased launch: basic level (tracking + serving) in 4–6 weeks.

Cost is calculated individually based on data volume, number of models, and infrastructure requirements. Order an MLOps infrastructure audit — get a roadmap in 1–2 weeks. Contact us for a project assessment — we will send a preliminary estimate within 2 business days.

Note: warranty on architectural solutions — 12 months. We provide integration certificates with major cloud providers (AWS, GCP, Azure). During our work, we have not lost a single client after the first implementation — the experience of 50+ successful MLOps projects speaks for itself. Get a consultation on building an MLOps platform today.