LLM Inference Optimization with TensorRT-LLM

Your LLaMA-70B answers in 5 seconds, but you need 1.5? Load is growing, GPU utilization is 80%, and you pay for every hour of downtime. A typical scenario: LLaMA-2 13B on A100-80GB yields 50 tokens/sec at batch=8, latency p99 2.5 sec. After optimization with TensorRT-LLM, enabling FP8 quantization and in-flight batching, throughput increases to 180 tokens/sec, latency drops to 0.8 sec, GPU utilization rises from 65% to 95%. We do this systematically.

TensorRT-LLM is not just a library—it's a way to squeeze the last percentages of performance out of NVIDIA GPUs. Over several years, we have performed more than 50 LLM inference optimizations for companies of various sizes. We are NVIDIA-certified for TensorRT-LLM deployment. Average GPU utilization after optimization is 95%, latency p99 decreases 2–4x. On average, clients save 40% on GPU instances after optimization.

How TensorRT-LLM works

TensorRT-LLM compiles the model into an optimized TensorRT engine. Graph compilation: the model graph is compiled considering the specific GPU (architecture, VRAM, tensor cores). Kernel fusion: multiple operations are combined into a single CUDA kernel (LayerNorm + Linear, Flash Attention). Quantization: FP8, INT8, INT4 with precise calibration methods. In-flight batching: the most advanced continuous batching implementation. According to the NVIDIA TensorRT-LLM technical report, FP8 quantization reduces quality by less than 0.5%.

Why TensorRT-LLM is faster than vLLM

TensorRT-LLM does what vLLM cannot: it compiles the model into machine code specific to your GPU. On H100 with FP8 quantization, throughput grows 2–3x without noticeable quality degradation (<0.5% on benchmarks). Hardware tensor cores run at full capacity—GPU utilization reaches 95%. If vLLM is a universal server, TensorRT-LLM is a racing car for NVIDIA.

How TensorRT-LLM differs from vLLM

If you need to quickly launch a prototype or work with non-NVIDIA GPUs—choose vLLM. If the goal is to squeeze maximum from each GPU and reduce costs—TensorRT-LLM gives 2–4x gain for the same money.

What is included in the optimization work?

We don't just run scripts. The delivery includes:

• Audit: measurement of current latency p99, throughput, GPU utilization, tokens/sec.
• Configuration selection: choice of TensorRT-LLM version, quantization type (FP8/INT8/INT4), batch and context window parameters.
• Compilation: building the engine with kernel fusion, in-flight batching, PagedAttention.
• Integration with Triton: setting up an ensemble of tokenization, inference, post-processing.
• Load testing: stability check, latency p99, throughput under peak load (up to 10k requests).
• Documentation and training: delivery of configs, scripts, monitoring recommendations, one-hour team training.

We guarantee at least 2x acceleration or money back for the work.

How FP8 quantization works on H100

H100 has hardware support for FP8—the greatest performance boost:

from tensorrt_llm.quantization import QuantAlgo

build_config_fp8 = BuildConfig(
    max_batch_size=128,
    max_input_len=4096,
    max_output_len=1024,
    quant_config=QuantConfig(
        quant_algo=QuantAlgo.FP8,
        kv_cache_quant_algo=QuantAlgo.FP8,
    ),
    plugin_config={
        "use_fp8_context_fmha": True,
        "gemm_plugin": "float16",
    }
)

FP8 on H100: roughly 2x throughput gain compared to BF16, quality degradation < 0.5% on standard benchmarks.

Table: Typical optimization results

Integration with Triton Inference Server

TensorRT-LLM natively integrates with NVIDIA Triton:

model_repository/
├── ensemble/
│   └── config.pbtxt
├── preprocessing/
│   ├── config.pbtxt
│   └── 1/model.py
├── tensorrt_llm/
│   ├── config.pbtxt
│   └── 1/
│       ├── model.engine
│       └── config.json
└── postprocessing/
    ├── config.pbtxt
    └── 1/model.py

name: "tensorrt_llm"
backend: "tensorrtllm"
max_batch_size: 128
parameters {
  key: "max_beam_width"
  value: { string_value: "1" }
}
parameters {
  key: "executor_worker_path"
  value: { string_value: "/opt/tritonserver/backends/tensorrtllm/trtllmExecutorWorker" }
}
parameters {
  key: "decoding_mode"
  value: { string_value: "top_p_top_k" }
}

Multi-GPU with Tensor Parallelism

build_config_tp4 = BuildConfig(
    max_batch_size=64,
    max_input_len=8192,
    max_output_len=2048,
    auto_parallel_config=AutoParallelConfig(
        world_size=4,
        gpus_per_node=4,
        shards_along_head=4,
    )
)

Timelines and implementation process

1. Day 1–3: installation of TRT-LLM, compilation of first model, measurement of baseline metrics.
2. Week 1–2: selection of quantization and fusion parameters, integration with Triton.
3. Week 3–4: load testing, tuning, deployment to production.
4. Month 2: optimization for specific scenarios (latency vs throughput), multi-model deployment.

Ready to accelerate your LLM? Order an inference audit. Get a consultation on TensorRT-LLM implementation from certified NVIDIA engineers.

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.