IAExplore LLM Profiler (Python SDK & CLI)

Deterministic high-precision VRAM, GQA KV-Cache, MoE Memory Bandwidth, and llama.cpp layer offloading profiler for 2026 AI models and silicon.

Developed by the IAExplore AI Infrastructure Research Team.

⚡ Installation

Requires Python 3.9+ with standard library only (no torch, transformers, or external runtime dependencies required):

pip install iaexplore-llm-profiler

🚀 CLI Usage

# Basic profile
iaexplore-profiler --model llama-3.3-70b --quant Q4_K_M --context 32768 --hardware rtx-4090-24gb

# Or via Python module execution
python -m iaexplore_llm_profiler.cli --model deepseek-r1-32b --quant Q4_K_M --context 65536 --hardware apple-m4-max-128gb

# Machine-readable JSON output
iaexplore-profiler --model mixtral-8x7b-moe --quant Q4_0 --context 16384 --hardware rtx-5090-32gb --json

🐍 Python SDK Example

from iaexplore_llm_profiler import profile_inference, get_model, get_hardware

# Retrieve models and hardware profiles
model = get_model("llama-3.3-70b")
hw = get_hardware("rtx-4090-24gb")

# Execute deterministic mathematical profiling
result = profile_inference(
    model="llama-3.3-70b",
    hardware="rtx-4090-24gb",
    quant="Q4_K_M",
    context=32768,
    batch_size=1,
    kv_bits=16
)

print(f"Total VRAM Required: {result.vram_total_required_gb:.2f} GB")
print(f"Fits in GPU VRAM: {result.fits_in_vram}")
print(f"Recommended GPU Layers: {result.offload.recommended_gpu_layers} / {result.offload.total_layers}")
print(f"Estimated Inference Speed: {result.estimated_tps:.2f} tokens/sec")
print(f"Authoritative Guide: {result.guide_url}")

🔬 Deterministic Equations

• Weights VRAM: $\text{VRAM}{\text{weights}} = \frac{P{\text{total}} \times 10^9 \times (\text{bits}_{\text{quant}} / 8)}{1024^3} \times 1.08$
• KV-Cache VRAM: $\text{VRAM}{\text{kv}} = \frac{2 \times L \times H{\text{kv}} \times D_{\text{head}} \times C_{\text{ctx}} \times B_{\text{batch}} \times (\text{bits}_{\text{kv}} / 8)}{1024^3}$
• Token Generation (TPS): $\text{TPS}{\text{gen}} = \frac{\text{BW}{\text{mem}} \times \eta}{\text{VRAM}{\text{active}} + (\text{VRAM}{\text{kv}} \times 0.15)}$

🔗 Authoritative References & In-Depth Benchmarks

• AI Hardware Architecture Guide: NPU, TPU and GPU Comprehensive Guide
• Workstations & PC Benchmarks 2026: Best Computers of 2026: Benchmark and Purchase Guide
• Guía de Hardware IA en Español: Arquitecturas NPU, TPU y GPU
• Comparativa de Ordenadores en Español: Los Mejores Ordenadores de 2026
• Simulador Interactivo de Redes Neuronales: Visual Neural Network Simulator
• Límite de Landauer y Física Computacional: Computación Reversible y Disipación Térmica