[wenh06]

What

Adds the GPU leg of the executor benchmark that the parallel-execution docs page left open (larger-model and GPU tiers to be answered by the benchmark), and folds its findings into that page.

Configuration: FedAvg on FedSynthetic with the 1.19M-parameter mlp_d3 candidate model (dimension=512), 64 clients, 3 iterations, 5 local epochs, batch 256, best of 3, on a 12-core host with one consumer GPU.

Findings (now in the docs page)

device	sequential	threads at 8 workers	speedup
CPU	14.9 s	10.3-10.6 s	1.40-1.45x
GPU (single)	5.1 s	4.9-5.1 s	1.01-1.04x (none)

• CPU with a compute-heavier model: the speedup grows over the tiny-model configuration (about 1.2x) - the torch kernels release the GIL, so the benefit tracks the compute share of each client task.
• Single GPU: no benefit from the threads backend. All clients submit kernels to the same device, where they execute serially regardless of the submitting thread; the sequential loop already saturates one card (the measured GPU run was itself about 3x faster than the CPU run for this workload). Scaling beyond one GPU comes from multi-GPU client placement (ServerConfig.visiable_gpus) or CUDA streams - not from concurrent submission to a single device.

The docs page now states these as observations on the measured configurations (not ceilings), and keeps the pointer to the planned process-based backend for GIL-bound CPU workloads.

Verification

• the script was executed end-to-end twice (once from the working copy, once from the committed path) with consistent numbers;
• pre-commit (black / isort / flake8) green.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 82.30%. Comparing base (967b45b) to head (0ee7476).

Additional details and impacted files

@@            Coverage Diff             @@
##           master      #55      +/-   ##
==========================================
+ Coverage   82.28%   82.30%   +0.01%     
==========================================
  Files          55       55              
  Lines        5640     5640              
==========================================
+ Hits         4641     4642       +1     
+ Misses        999      998       -1     

see 2 files with indirect coverage changes

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.