Allow using sockets to distribute work to GPUs.

[Menkib64]

If server has many GPUs, test can suffer from concurrent compile for another GPU. We want use shared client and binaries for all GPUs to avoid compilation interfering with testing. This change allows running client with -N <number of gpus> to use all GPUs using one client.

[Copilot]

🟡 Changes recommended

Fix the critical workload round-count mismatch before approval.

Get a fresh assessment by requesting another Copilot review.

Pull request overview

This PR enables one client to distribute work across multiple GPUs while sharing binaries.

Changes:

• Adds multi-GPU runner distribution.
• Routes engine and benchmark processes to selected GPUs.
• Coordinates runner completion and result tracking.

File summaries

File	Summary
OpenBench/workloads/get_workload.py	Enables multiple runner assignments.
Client/worker.py	Adds GPU routing and coordination; contains a critical -rounds mismatch at lines 413 and 1414.
Client/bench.py	Passes GPU indices to benchmark processes.

Review details

Suppressed comments (1)

Client/worker.py:1414

• A runner clears results['done'] and continues after emitting its own completion marker, but batch_done is not set until every runner has reported. If one GPU is faster, it can enqueue additional result batches while another GPU is still finishing; those batches are then drained and uploaded, exceeding the allocated rounds-per-runner and potentially consuming another runner's opening range. Stop this runner when its own done result is emitted, not only when the global event is set.

            if batch_done.is_set():

• Files reviewed: 3/3 changed files
• Comments generated: 1
• Review effort level: Lite

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.