LLM Inference with TGI: Reduce Latency and Save VRAM

Problem: LLM inference with unpredictable latency and high memory consumption

When deploying LLM inference in production, latency jumps from 200 ms to 5 seconds, GPU memory overflows under peak load, and each new request requires restarting the pipeline. Teams spend weeks tuning inference, yet the result remains unstable. Text Generation Inference (TGI) from HuggingFace solves these problems at the production-server level: it is written in Rust and Python, natively integrated with HuggingFace Hub, and supports advanced techniques—continuous batching, Flash Attention 2, tensor parallelism, and speculative decoding.

How TGI reduces latency and boosts throughput

TGI uses continuous batching (in-flight batching): new requests are added to the active batch without waiting for previous ones to finish. This achieves GPU utilization above 95% and reduces average queue wait time. Flash Attention 2 provides O(n) memory instead of O(n²)—critical for long contexts. In practice, we achieved p99 latency under 300 ms for Llama-3-8B with 100 concurrent requests. Adopting TGI pays off by cutting GPU infrastructure costs up to 60%.

Why TGI over a custom implementation?

Custom inference requires manual memory management, batching, and parallelization. TGI provides a production-ready server with continuous batching, tensor parallelism, and quantization built-in. This lowers the barrier: Docker images deploy in minutes. The trade-off in flexibility is offset by stability and reduced debugging time—for instance, speculative decoding accelerates generation by 20–30% without model changes.

Quick start

# Docker (recommended)
docker run --gpus all \
  -p 8080:80 \
  -v /data/models:/data \
  ghcr.io/huggingface/text-generation-inference:2.1 \
  --model-id meta-llama/Llama-3-8b-instruct \
  --max-input-length 4096 \
  --max-total-tokens 8192 \
  --max-batch-prefill-tokens 32768 \
  --num-shard 1 \
  --dtype bfloat16 \
  --huggingface-hub-token $HF_TOKEN

# Client via official package
from huggingface_hub import InferenceClient

client = InferenceClient(model="http://localhost:8080")

response = client.text_generation(
    prompt="Explain transformer attention in simple terms",
    max_new_tokens=512,
    temperature=0.7,
    repetition_penalty=1.1,
    stream=False
)

# Streaming
for token in client.text_generation(prompt, stream=True):
    print(token, end="", flush=True)

Key TGI capabilities

• Continuous batching (in-flight batching): new requests join the batch during ongoing generation.
• Flash Attention 2: efficient self‑attention with O(n) memory vs O(n²).
• Tensor Parallelism: distribute model across multiple GPUs via --num-shard.
• Speculative Decoding: via --speculate N—draft model generates N tokens, target verifies.
• Quantization: built‑in support for GPTQ, AWQ, EETQ, BitsAndBytes for LLM quantization.

Configuration for different scenarios

Custom Handlers

TGI allows adding preprocessing/postprocessing via a custom handler:

# custom_handler.py
class CustomHandler:
    def __init__(self):
        self.tokenizer = AutoTokenizer.from_pretrained(...)

    def preprocess(self, inputs: dict) -> dict:
        """Transform incoming request before inference."""
        prompt = inputs.get("inputs", "")
        full_prompt = f"<|system|>You are a helpful assistant.<|end|>\n<|user|>{prompt}<|end|>\n<|assistant|>"
        return {"inputs": full_prompt, **{k: v for k, v in inputs.items() if k != "inputs"}}

    def postprocess(self, model_output: dict) -> dict:
        """Post-process model output."""
        generated = model_output["generated_text"]
        return {"generated_text": generated.split("<|assistant|>")[-1].strip()}

Monitoring and metrics

TGI exports Prometheus metrics at /metrics:

tgi_request_duration_seconds_bucket  # latency histogram
tgi_batch_inference_duration_seconds  # batch inference time
tgi_request_input_length              # input lengths
tgi_request_generated_tokens          # generated token lengths
tgi_batch_current_size                # current batch size
tgi_queue_size                        # queue size

Which to choose: TGI or vLLM?

For most use cases, both offer similar performance. TGI is more convenient when operating in the HF ecosystem.

Process and what’s included

We offer turnkey TGI implementation. Stages:

1. Infrastructure audit—we assess load, latency, VRAM usage, and existing pipeline.
2. Configuration selection—we choose the model, quantization (INT4 vs FP16), number of shards, and continuous batching parameters.
3. Deployment—we configure the Docker image, integrate with your API, and set up monitoring via Prometheus + Grafana.
4. Optimization—we tune p99 latency, throughput, and memory footprint. We use speculative decoding to accelerate generation by 20–30%.
5. Documentation and training—we deliver operation instructions, configuration templates, and dashboards. We conduct a workshop for your team.

Estimated implementation time: 2 to 4 weeks depending on infrastructure complexity.

Results and guarantees

Our MLOps engineers have 5+ years of experience in MLOps and have completed over 20 LLM inference projects for chatbots, RAG systems, and assistants. We guarantee stable operation—average uptime 99.9% after deployment. Specific metrics: p99 latency reduction of 30%, VRAM savings up to 50% on 7B models with INT4 quantization. This translates to GPU-hour cost reductions—in some projects savings reach 60%. Get a consultation on TGI setup—we’ll recommend the best configuration for your case. Contact us for a project assessment.

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.