Pruning Neural Networks: Optimize and Accelerate Inference

Imagine: your trained GPT-like model takes 16 GB and produces a response in 5 seconds on a GPU, but on CPU or edge devices it's unacceptable. Clients complain about delays, infrastructure costs are rising. We constantly encounter this situation: the model is good, but you can't deploy it in production—it's expensive and slow. Pruning neural networks is one of the key tools to fix this.

Pruning—removing insignificant parameters (weights, neurons, attention heads, layers) from a trained neural network. The goal is to reduce model size and speed up inference with minimal quality loss. In the context of LLMs, pruning is often combined with model quantization and distillation for maximum compression.

Why is pruning important for LLMs?

Large language models (LLMs) contain billions of parameters, but many are redundant. Pruning can compress the model 2–4 times without significant degradation, which is critical for edge deployment, reducing GPU costs, and lowering latency. Without pruning, modern LLMs often remain research artifacts—too expensive for real-world application. Weight trimming is a synonym for unstructured pruning.

Types of pruning

• Unstructured pruning: individual weights are zeroed throughout the matrix. High compression, but requires sparse computation—standard GPUs do not accelerate sparse operations out of the box.
• Structured pruning: entire structural elements are removed—neurons, attention heads, layers. The result is a truly smaller dense model that runs faster on standard hardware.
• Semi-structured pruning (N:M sparsity): N weights are removed from each block of M. The 2:4 format is supported by NVIDIA Ampere and later at the hardware level (up to 2× acceleration).
• Depth pruning: entire layers are removed, reducing model depth.

Which pruning method to choose for your model?

The choice of pruning method depends on the target hardware and quality requirements. Below is a comparison of popular approaches.

Pruning methods with code examples

LLM-Pruner: structured pruning of LLMs

# Example usage of LLM-Pruner
# pip install llm-pruner

from LLMPruner.pruner import LlamaStructuredPruner
from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("meta-llama/Meta-Llama-3.1-7B")
tokenizer = AutoTokenizer.from_pretrained("meta-llama/Meta-Llama-3.1-7B")

pruner = LlamaStructuredPruner(
    model=model,
    tokenizer=tokenizer,
    pruning_ratio=0.25,  # Remove 25% of parameters
)

# Compute parameter importance on calibration data
calibration_data = ["Text for weight importance analysis...", ...]
pruner.get_mask(calibration_data, method="taylor")  # Taylor expansion importance

# Apply mask and prune
pruned_model = pruner.prune()

SparseGPT: efficient unstructured pruning without retraining

From the SparseGPT documentation: SparseGPT can prune 50% of the weights without retraining. — GitHub

# sparsegpt — library from the method authors
# Conceptual code example

from sparsegpt import SparseGPT

sparsegpt = SparseGPT(model)
sparsegpt.fasterprune(
    sparsity=0.5,         # 50% sparsity
    prunen=2,             # N in N:M
    prunem=4,             # M in N:M (2:4 — hardware supported)
    percdamp=0.01,
    blocksize=128,
)

With 2:4 sparsity (50%) on NVIDIA A100/H100, inference acceleration on Tensor Core is about 1.7–2×.

Wanda: simple and effective pruning

# Wanda is simpler than SparseGPT but comparable in quality
# Runs in minutes on a 7B model

def wanda_pruning(model, calibration_loader, sparsity=0.5):
    """Simplified Wanda implementation"""
    for name, module in model.named_modules():
        if isinstance(module, torch.nn.Linear):
            # Accumulate activation statistics
            activation_norms = get_activation_norms(module, calibration_loader)

            # Importance score = |W| * ||X||
            importance = module.weight.abs() * activation_norms

            # Pruning by threshold
            threshold = torch.quantile(importance, sparsity)
            mask = importance > threshold
            module.weight.data *= mask

    return model

Depth pruning: removing layers

For LLMs, middle layers are often less critical than first and last ones:

def depth_prune_llm(model, layers_to_remove: list[int]):
    """Remove specified decoder layers"""
    # For Llama architecture
    remaining_layers = [
        layer for i, layer in enumerate(model.model.layers)
        if i not in layers_to_remove
    ]
    model.model.layers = torch.nn.ModuleList(remaining_layers)
    return model

# Example: remove 8 middle layers out of 32 (25% depth reduction)
pruned_model = depth_prune_llm(model, layers_to_remove=list(range(12, 20)))
# Result: 24-layer model from 32-layer

Practical case: edge deployment optimization

Task: fine-tuned Llama 3.1 8B for an industrial controller (ARM server, 16GB RAM, no GPU). Requirement: inference < 2s per request.

Our optimization strategy:

1. GGUF Q4_K_M quantization: 8B → 4.1GB, 8 tok/s on CPU (insufficient)
2. Depth pruning (remove 8 layers out of 32): -25% latency, -3% quality
3. Width pruning of attention heads (remove 20% heads): -15% latency
4. Re-quantization: GGUF Q4_K_M on pruned model

Final pruned+quantized model characteristics:

In a typical project with a 7B parameter model, infrastructure savings reach $12,000/month (based on reduced GPU hours). Compare: pruning yields a 2.5× better cost-performance ratio than no optimization. This is 50% better than quantization alone.

# After pruning — brief fine-tuning for recovery
from trl import SFTTrainer, SFTConfig

# Use the same dataset as for fine-tuning, but with a lower LR
recovery_config = SFTConfig(
    num_train_epochs=1,       # 1 epoch for recovery
    learning_rate=5e-5,       # Lower than full fine-tuning
    gradient_checkpointing=True,
    bf16=True,
)
trainer = SFTTrainer(model=pruned_model, args=recovery_config, train_dataset=dataset)
trainer.train()

What is included in the work

• Full documentation of all changes, including pruning masks and recovery steps.
• Access to all pruned model variants (GGUF, ONNX, TensorRT) for immediate testing.
• Training session for your team on how to deploy and maintain the optimized model.
• Support for 3 months after deployment, including troubleshooting and performance monitoring.
• Additional deliverables: model analysis, pruning strategy, calibration data, benchmarking reports.

Work process: stages and timeline

1. Analysis: meet, discuss requirements, receive model or dataset. Identify bottlenecks (latency, size, quality).
2. Design: select pruning and recovery methods, prepare pipeline. Estimate expected effect.
3. Implementation: perform pruning, recovery fine-tuning using chosen tools.
4. Testing: compare with original, record metrics (perplexity, latency, throughput).
5. Deployment: package model, provide integration instructions.

Timeline:

• Pruning strategy selection: 3–5 days
• Calibration and pruning: 4–24 hours (depends on method and size)
• Recovery fine-tuning: 2–8 hours
• Benchmarking and evaluation: 3–5 days
• Total: 2–4 weeks

Cost is calculated individually—depends on model size, task complexity, and required quality. Our experience: 5+ years in ML, 10+ projects on model optimization for production. We guarantee transparent results and full documentation.

If you have a similar task, contact us. Request a consultation—we'll tell you how to speed up inference by 2–5× without significant quality loss.

LLM Development: Fine-Tuning, RAG, Agents, and Production Deployment

Using GPT‑4 or Claude 3.5 Sonnet through a public API is not a solution — it's just a tool. When the requirement is to "make it like ChatGPT, but on our data," there is a real engineering challenge behind it: from prompt engineering to training a 70B model on your own infrastructure. End-to-end LLM solution development is a complex stack, and we have been doing it for over 5 years. During this time, we have completed over 20 projects in generative AI: from RAG systems for legal departments to custom support agents. Where exactly your task falls depends on data, latency requirements, budget, and how critical confidentiality is.

A typical situation: the client has already tried ChatGPT, but results are unstable — sometimes accurate, sometimes hallucinating. Or they need integration into a corporate portal while complying with security policies. Let's break down each layer of the stack in detail — from RAG to production deployment.

Why Do RAG Systems Break and How to Fix It?

RAG (Retrieval-Augmented Generation) looks simple: find relevant documents, put them in context, get an answer. In practice, it fails in several places.

Chunking without overlap. Classic mistake: chunk_size=512, overlap=0. If the answer lies across two chunks, retrieval won't find either with sufficient confidence. Solution: overlap 15–25% of chunk_size, or better yet, sentence-aware splitting with spaCy or NLTK instead of naive character splitting.

Poor embedder. text-embedding-ada-002 is good for general use, but on legal or medical texts, specialized models like E5-large-v2, BGE-M3, or fine-tuned sentence-transformers on domain data outperform it. Recall@5 differences can be 15–25%.

No re-ranking. Vector search optimizes for speed, not relevance. A cross-encoder re-ranker (ms-marco-MiniLM-L-6-v2, bge-reranker-large) after initial retrieval improves top-3 accuracy with acceptable latency (+50–150ms). This is often more impactful than improving the embedding model.

Hybrid search. Dense vectors alone work poorly on exact queries: names, SKUs, codes. BM25 (sparse) finds exact matches but misses semantics. Hybrid via RRF (Reciprocal Rank Fusion) is the optimal compromise. Qdrant, Weaviate, and pgvector 0.7+ support hybrid search natively.

When to Fine-Tune Instead of Prompt Engineering?

Prompt engineering solves ~70% of LLM adaptation tasks for a domain. The remaining 30% require fine-tuning. Three indicators: the model ignores a specific output format even with detailed prompting; the task requires deep knowledge of specialized vocabulary (medicine, law); you need to significantly reduce token costs by replacing a large model with a smaller specialized one.

LoRA and QLoRA are the standard for SFT. LoRA adds trainable low-rank matrices to attention layers. A typical configuration for Llama-3 8B: r=64, lora_alpha=128, target_modules=["q_proj","v_proj","k_proj","o_proj"] yields ~0.8% trainable parameters, training on one A100 40GB. QLoRA adds 4-bit quantization (NF4) and allows fine-tuning 70B models on two A100 40GB, though speed drops by half compared to bf16.

DPO instead of RLHF. Direct Preference Optimization requires only (chosen, rejected) pairs, not scalar reward signals. DPOTrainer from the trl library (Hugging Face) implements it in a few dozen lines.

Common mistake. A dataset of 500 examples, 5 epochs, validation loss 0.8 — seems fine. But on test, the model degrades on general instructions. Cause: catastrophic forgetting. Solution: add 10–20% general instruction-following examples (Alpaca, FLAN) to the training set to preserve original capabilities.

How to Choose a Base Model: 8B or 70B?

For most tasks, fine-tuning an 8B model is sufficient. 70B is needed when deep reasoning is required or the 8B baseline does not reach the required quality even after fine-tuning. Inference cost for Llama-3 8B via vLLM on A100 is efficient; the exact cost depends on volume.

What Does PagedAttention Bring to Production?

vLLM is the first choice for serving open-source models. PagedAttention is the key technical innovation: KV-cache is managed like virtual memory in an OS, without fragmentation. This yields 2–4x higher throughput compared to naive HuggingFace Transformers inference. The vLLM documentation confirms that continuous batching and PagedAttention are the standard for high-load LLM services.

Typical numbers on A100 80GB for Llama-3 8B (bf16): 400–600 req/s, P50 latency 200–400ms, P99 latency 600–900ms at concurrency 64. For 70B on two A100 with tensor parallelism: 80–120 req/s, P99 latency 1.5–2.5s. AWQ or GPTQ quantization reduces memory consumption by 2x with quality loss within 1–3%.

Multi-Agent Systems

Agents are LLMs with access to tools: search, code execution, API calls, database interaction. Common patterns:

• ReAct (Reason + Act): the model reasons → chooses a tool → observes the result → reasons again. LangChain and LlamaIndex implement it out of the box.
• Multi-agent orchestration: multiple specialized agents with a coordinator on top. Example: coordinator → researcher (search + summarization) → coder (code generation and execution) → critic (verification). Tools: AutoGen (Microsoft), CrewAI, custom implementation on LangGraph.

In production, agent systems are non-deterministic. Essential: guardrails, step limits, logging of each step, human-in-the-loop for critical actions.

How We Work: Stages, Timeline, Deliverables

What Is Included

We deliver:

• Technical documentation (model card, configs, deployment instructions)
• Access to infrastructure (code repository, trained weights)
• 1 month of post-deployment support (consultations, bug fixes)
• Customer team training (2–3 sessions on system operation)

Timeline: basic RAG prototype — 1–2 weeks. Fine-tuning with customer data — 3–6 weeks (including data preparation). Production system with monitoring and retraining — 2–4 months. Cost is calculated individually based on data volume, model complexity, and infrastructure requirements.

We guarantee the quality of the final model with performance benchmarks and ongoing monitoring. Our engineers have hands‑on experience with dozens of production LLM systems.

Want to evaluate your project? Leave a request — we will prepare a preliminary summary within 1–2 business days. Or get a consultation on choosing the approach: RAG, fine-tuning, or hybrid — we will tell you what works best for you. Contact us to discuss your LLM development needs. Schedule a free consultation today.