Optimizing Dynamic Batching for LLMs: Boost GPU Throughput

If your LLM service experiences high load with many concurrent users, each request is processed sequentially—without batching, throughput drops significantly and latency becomes unacceptable. We configure dynamic batching to drive GPU utilization to 80%+ instead of 5%. Our engineers have 4+ years of LLM production experience and have delivered over 20 projects on vLLM, TensorRT-LLM, and custom solutions. Dynamic batching merges multiple parallel requests into a single forward pass through the GPU. This is a key mechanism for high LLM throughput: the GPU is parallel and handles matrix multiplications more efficiently with larger batches. Proper batching tuning can reduce the number of required GPUs by a factor of 3–5, saving between 150,000 and 500,000 rubles per month on infrastructure.

Why Batching Is Critical for LLMs

Without batching, even a powerful A100 80GB GPU delivers only 30 tokens/sec for a Llama-3-8B model. With batch=16, it jumps to 300 tokens/sec; with batch=64, it reaches 900 tokens/sec—a 30x improvement. However, p99 latency rises from 200 ms to 400 ms, still acceptable for most real-time scenarios. If you have 100 concurrent users, without batching each waits in queue—total response time can exceed a minute. With continuous batching, all requests are processed in parallel, reducing response time to seconds.

Why Continuous Batching Outperforms Static Batching

Static batching fixes the batch size and waits for it to fill, increasing latency under low load. Continuous batching (in-flight batching) dynamically adds requests as soon as the GPU is free, reducing wait time and improving utilization.

Continuous (In-Flight) Batching in vLLM

According to the official vLLM documentation, continuous batching is implemented automatically. Key parameters: max-num-seqs—maximum number of requests per batch, max-num-batched-tokens—total tokens per batch, scheduler-delay-factor—delay before forming a batch. Example configuration:

python -m vllm.entrypoints.openai.api_server \
  --model meta-llama/Llama-3-8b-instruct \
  --max-num-seqs 256 \
  --max-num-batched-tokens 32768 \
  --scheduler-delay-factor 0.5 \
  --use-v2-block-manager \
  --enable-chunked-prefill

Chunked prefill splits long prefill into chunks to avoid blocking decode of other requests:

--enable-chunked-prefill
--max-num-batched-tokens 8192

How to Tune Dynamic Batching for a Specific GPU

Follow these steps:

1. Determine model and GPU. For example, Llama-3-8B on A100-80GB.
2. Choose a framework. vLLM for a quick start, TensorRT-LLM for maximum performance.
3. Run benchmarks. Load test with varying numbers of concurrent users.
4. Tune parameters. max-num-seqs, max-num-batched-tokens, scheduler-delay-factor.
5. Monitor. Track metrics like num_requests_running, avg_prompt_throughput_toks_per_s.

Which Monitoring Metrics Matter for Batching?

vLLM exports metrics via Prometheus: num_requests_running (requests in active batch), num_requests_waiting (queued), avg_prompt_throughput_toks_per_s, avg_generation_throughput_toks_per_s. Use these to balance throughput and latency. For comprehensive monitoring, use Grafana.

Common batching configuration mistakes:

• Too large max-num-seqs: increases p99 latency due to KV cache memory contention.
• Ignoring chunked prefill: long prompts block decode, lowering utilization.
• No real-load benchmarking: parameters tuned on synthetic data often fail in production.

Configuring Dynamic Batching in TensorRT-LLM / Triton

# tensorrt_llm/config.pbtxt
parameters {
  key: "max_tokens_in_paged_kv_cache"
  value: { string_value: "40000" }
}
parameters {
  key: "batch_scheduler_policy"
  value: { string_value: "guaranteed_no_evict" }
}
parameters {
  key: "executor_static_batch_size"
  value: { string_value: "-1" }
}

Manual Batching Implementation (Example: DynamicBatchInferenceServer)

If using a custom inference server:

import asyncio
from dataclasses import dataclass
from collections import deque
import time

@dataclass
class PendingRequest:
    id: str
    prompt: str
    max_tokens: int
    future: asyncio.Future
    enqueued_at: float

class DynamicBatchInferenceServer:
    def __init__(
        self,
        model,
        max_batch_size: int = 64,
        max_wait_ms: float = 20.0,
        max_tokens_per_batch: int = 16384
    ):
        self.model = model
        self.max_batch_size = max_batch_size
        self.max_wait_ms = max_wait_ms
        self.max_tokens_per_batch = max_tokens_per_batch
        self.queue: deque[PendingRequest] = deque()
        self.lock = asyncio.Lock()
        self._batch_worker_task = None

    async def start(self):
        self._batch_worker_task = asyncio.create_task(self._batch_worker())

    async def predict(self, prompt: str, max_tokens: int = 512) -> str:
        future = asyncio.get_event_loop().create_future()
        request = PendingRequest(
            id=str(time.time()),
            prompt=prompt,
            max_tokens=max_tokens,
            future=future,
            enqueued_at=time.time()
        )
        async with self.lock:
            self.queue.append(request)
        return await future

    async def _batch_worker(self):
        while True:
            await asyncio.sleep(self.max_wait_ms / 1000)
            async with self.lock:
                if not self.queue:
                    continue
                batch: list[PendingRequest] = []
                total_tokens = 0
                while (self.queue
                       and len(batch) < self.max_batch_size
                       and total_tokens + self.queue[0].max_tokens <= self.max_tokens_per_batch):
                    req = self.queue.popleft()
                    batch.append(req)
                    total_tokens += len(req.prompt.split()) + req.max_tokens
            if not batch:
                continue
            prompts = [req.prompt for req in batch]
            max_tokens_list = [req.max_tokens for req in batch]
            try:
                outputs = self.model.generate_batch(prompts, max(max_tokens_list))
                for req, output in zip(batch, outputs):
                    if not req.future.done():
                        req.future.set_result(output)
            except Exception as e:
                for req in batch:
                    if not req.future.done():
                        req.future.set_exception(e)

Thanks to dynamic batching tuning, our clients reduce GPU infrastructure costs by a factor of 3–10, achieving project payback within 2–3 months. Savings start at 150,000 rubles per month.

What's Included in the Configuration

• Inference server configuration (vLLM, TensorRT-LLM, or custom)
• Benchmarking and batching parameter tuning
• Integration of batching monitoring metrics
• Deployment and maintenance documentation
• Team training (optional)

Estimated timeline: 2 to 10 working days, depending on complexity. Pricing is individual, determined after analysis.

Get a consultation on optimizing your LLM throughput. Contact us — we'll evaluate your project in 1 day. Order an audit of your current batching configuration — we'll identify bottlenecks and suggest improvements with cost savings estimates.

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.