---

Paper: "Characterizing State Space Model and Hybrid Language Model Performance with Long Context" Saptarshi Mitra, Rachid Karami, Haocheng Xu, Sitao Huang, Hyoukjun Kwon · ISPASS 2026

---

🗺️ Overview

SSM-Scope provides comprehensive benchmarking and profiling tools to evaluate three families of language models across long context lengths on both a desktop GPU (RTX 4090) and an edge device (NVIDIA Jetson Nano Orin):

Key Features

• Computational Performance Tracking — TTFT, TPOT, and overall throughput across generation stages.
• Detailed Memory Analysis — peak GPU memory decomposed into model weights, activations, and KV cache.
• Operator-Level Profiling — latency breakdowns separating GEMM, non-GEMM, and SSM-specific kernels.
• Energy Consumption Metrics — joules-per-prefill from nvidia-smi power logs, critical for edge evaluation.

_{Fig 5: GPU memory footprint vs sequence length across Transformer, SSM, and Hybrid model families — RTX 4090 (top) & Jetson Nano (bottom).}

---

📊 Paper Figures at a Glance

🔷 Motivation & Inference Performance

🔷 Memory Footprint

🔷 Operator-level Performance Breakdown

---

🗂️ Repository Structure

SSM-characterization/
├── src/                          # Core profiling framework
│   ├── profiling/                #   PyTorch profiler engine (TTFT, TPOT, energy, op shapes)
│   ├── models/                   #   Model loaders and per-model profiling entry points
│   ├── memory/                   #   Memory footprint & vLLM OOM sweep
│   └── visualization/            #   Figure generation from profiling CSVs
│
├── profile_data/                 # Pre-collected profiling CSVs — desktop GPU
├── profile_data_jetson/          # Pre-collected profiling CSVs — Jetson Nano Orin
│
└── ispass_ae/
    ├── notebooks/                # Interactive Jupyter notebooks (plot from pre-collected data)
    └── scripts/
        ├── env_setup/            # Virtual environment setup instructions
        │   └── README.md  ◄──── START HERE for environment setup
        └── paper_figures/        # End-to-end scripts per figure
            ├── Fig_1/  Fig_3/  Fig_5a/  Fig_5b/
            └── Fig_6a/ Fig_6b/ Fig_7/   Fig_8/  Fig_9a/  Fig_9b/

---

⚡ Quick Start

1 — Set Up Environments

Three virtual environments cover all models. 👉 See ispass_ae/scripts/env_setup/README.md for full instructions.

2 — Reproduce Any Figure

Every figure has a one-command end-to-end script:

# Example: reproduce Figure 7 (SSM op-breakdown)
bash ispass_ae/scripts/paper_figures/Fig_7/gen_fig7.sh

Or Follow the instructions in the target figure's README.md for more details and tips.

Or open any notebook in ispass_ae/notebooks/ to plot directly from pre-collected data — no GPU required.

---

🖥️ Hardware Requirements

---

📦 Pre-collected Data

Raw profiling CSVs are included — figures can be reproduced without re-running inference:

• Desktop GPU: profile_data/
• Jetson Nano Orin: profile_data_jetson/

---

🔍 Chrome Trace Visualization

Several profiling runs export Chrome trace files (.json) alongside the CSV summaries, located under profile_data/ and profile_data_jetson/ (e.g. profile_data/gpt-neo-125m_cuda_1_1024/gpt-neo-125m_cuda_1_1024.json).

To inspect them:

1. Open Perfetto UI in your browser — no installation required.
2. Click Open trace file and select any .json trace.
3. Alternatively, navigate to chrome://tracing in a Chromium-based browser and load the file there.

The traces show per-kernel GPU timelines and are useful for understanding operator overlap, launch overhead, and memory transfer patterns beyond what the CSV summaries capture.

---

Running the Profiling Framework

All scripts are run from within the src/ directory:

# Operator-level profiling (prefill / TTFT)
cd src
python -m models.profile_runner --model_name mamba2 --batch_size 1 --seq_len 1024 --device cuda

# Memory footprint sweep
python -m memory.mem_footprint

---

📄 Citation

If you use this framework in your research, please cite:

@article{mitra2025characterizing,
  title   = {Characterizing State Space Model (SSM) and SSM-Transformer Hybrid
             Language Model Performance with Long Context Length},
  author  = {Mitra, Saptarshi and Karami, Rachid and Xu, Haocheng and
             Huang, Sitao and Kwon, Hyoukjun},
  journal = {arXiv preprint arXiv:2507.12442},
  year    = {2025}
}

---