HRM_sudoku / GUIDE_SPEED_TEST.md
Code2aum's picture
Upload folder using huggingface_hub
5dc80b3 verified
|
Raw
History Blame Contribute Delete
1.79 kB

Speed Test Guide: HRM SRAM/DRAM Benchmarking

Use this guide to measure the hardware performance (latency, throughput, memory overhead) of the memory-tiered architecture.

This guide uses synthetic data. No external datasets or prior training are required. You can run this immediately after setup.


1. Environment Setup

Create a clean virtual environment and install dependencies:

python -m venv venv
source venv/bin/activate
pip install --upgrade pip
pip install -r requirements.txt
pip install triton matplotlib

2. Run Hardware Comparison

This command compares the Baseline (Original) model vs. the Tiered (Optimized) model across multiple configurations.

python run_benchmark.py --mode compare --plot

What this does:

  1. Generates random tensors to simulate a reasoning workload.
  2. Compiles Triton kernels for the L-level (SRAM) and H-level (DRAM) paths.
  3. Records precise GPU timing using torch.cuda.Event.
  4. Saves charts to benchmark_results/.

3. Deep Dive into Tiered Metrics

To get a detailed breakdown of memory tier behavior:

python run_benchmark.py --mode tiered --iterations 50 --batch-sizes 8 --seq-lens 128

Key Metrics to watch:

  • L_lat(μs): Speed of the fast-updating L-level module in SRAM.
  • H_lat(μs): Speed of the planning H-level module in DRAM.
  • H/L Ratio: Shows the latency multiplier between memory tiers.
  • Memory Efficiency: Percentage of time spent on math vs. memory transfer.

4. Custom Hardware Stress Test

Test the limits of your GPU by increasing batch sizes or sequence lengths:

python run_benchmark.py \
    --mode compare \
    --batch-sizes 64,128 \
    --seq-lens 256,512 \
    --hidden-size 1024