DeepQMC-Survey — Profiling Neural-Network Quantum Monte Carlo
Project Description
DeepQMC-Survey (a.k.a. NNVMC-Survey) is a Linux command-line companion for
profiling and post-processing DeepQMC-style electronic-structure workloads. It
wraps NVIDIA Nsight Systems and Nsight Compute, records GPU memory timelines,
and renders roofline, benchmark, and kernel-breakdown figures from the resulting
traces.
The project targets neural-network variational Monte Carlo (NNVMC) methods for
solving the Schrödinger equation — FermiNet and PauliNet (via the `deepqmc`
runtime) and Orbformer and Psiformer (via a `oneqmc`-style checkout). A key
design principle: **the tool does not ship training code or weights; it is a
reproducible, hardware-aware profiler that turns raw Nsight traces into
comparable performance artefacts** (memory summaries, roofline plots, kernel
breakdowns) across methods and GPUs (A5000 / A100 / H100).
This project aims to (1) extend the profiling coverage to additional NNVMC
targets and hardware presets, and (2) build a reproducible benchmark suite that
compares runtime, GPU-memory footprint, and compute-efficiency (roofline)
across the supported quantum-chemistry models.
profiling and post-processing DeepQMC-style electronic-structure workloads. It
wraps NVIDIA Nsight Systems and Nsight Compute, records GPU memory timelines,
and renders roofline, benchmark, and kernel-breakdown figures from the resulting
traces.
The project targets neural-network variational Monte Carlo (NNVMC) methods for
solving the Schrödinger equation — FermiNet and PauliNet (via the `deepqmc`
runtime) and Orbformer and Psiformer (via a `oneqmc`-style checkout). A key
design principle: **the tool does not ship training code or weights; it is a
reproducible, hardware-aware profiler that turns raw Nsight traces into
comparable performance artefacts** (memory summaries, roofline plots, kernel
breakdowns) across methods and GPUs (A5000 / A100 / H100).
This project aims to (1) extend the profiling coverage to additional NNVMC
targets and hardware presets, and (2) build a reproducible benchmark suite that
compares runtime, GPU-memory footprint, and compute-efficiency (roofline)
across the supported quantum-chemistry models.
Supervisor
CHENG, Lixue
Quota
2
Course type
UROP1000
UROP1100
UROP2100
UROP3100
UROP3200
UROP4100
Applicant's Roles
1. Study and adapt the DeepQMC-Survey codebase (`deepqmc_survey/`) to understand
the CLI surface, the Nsight wrappers, and the plotting/roofline pipeline.
2. Help add profiling support for new targets or hardware presets and validate
the external DeepQMC / oneqmc runtime integration.
3. Build a benchmark suite of NNVMC workloads and run comparative experiments
measuring runtime, GPU-memory, and roofline efficiency across models.
4. Analyze results, document findings, and draft a workshop paper.
the CLI surface, the Nsight wrappers, and the plotting/roofline pipeline.
2. Help add profiling support for new targets or hardware presets and validate
the external DeepQMC / oneqmc runtime integration.
3. Build a benchmark suite of NNVMC workloads and run comparative experiments
measuring runtime, GPU-memory, and roofline efficiency across models.
4. Analyze results, document findings, and draft a workshop paper.
Applicant's Learning Objectives
1. Design and run a reproducible, hardware-aware performance-profiling workflow
for scientific ML workloads.
2. Connect quantum-chemistry knowledge (variational Monte Carlo, many-electron
wavefunctions) with GPU performance engineering.
3. Gain hands-on experience with NVIDIA Nsight tooling, roofline analysis, and
NNVMC models (FermiNet / PauliNet / Orbformer / Psiformer).
4. Develop and evaluate a domain-specific benchmark for AI4Science compute.
for scientific ML workloads.
2. Connect quantum-chemistry knowledge (variational Monte Carlo, many-electron
wavefunctions) with GPU performance engineering.
3. Gain hands-on experience with NVIDIA Nsight tooling, roofline analysis, and
NNVMC models (FermiNet / PauliNet / Orbformer / Psiformer).
4. Develop and evaluate a domain-specific benchmark for AI4Science compute.
Complexity of the project
Challenging