MILAB-6382: dual-mode operation via exec.hasGpu

[blackcat]

Summary

gpu-test now runs cleanly on both GPU-enabled and CPU-only clusters. The workflow gates .gpuMemory() on exec.hasGpu and passes a --mode {gpu,cpu} flag to the software so the Python side runs the matching path with no runtime detection. The CPU branch is first-class — the block returns a "no GPU available" report and exits successfully, never errors.

First of six per-block PRs in the GPU-on-GCP plan — pattern reference for the remaining blocks (clonotype-space, immuno-match, 3d-structure-prediction, paratope-clustering, compositional-analysis).

Why this matters

Calling .gpuMemory() against a backend that reports exec.hasGpu = false is a permanent runner error (SDK changelog: "Calling .gpuMemory() against a backend that reports no GPU now produces a permanent error from the runner driver"). Today's gpu-test workflow calls .gpuMemory() unconditionally when args.gpuMemory is set — fine on a GPU cluster, broken on a CPU-only one. This PR fixes the gate AND moves the GPU-vs-CPU decision out of Python runtime detection into tengo, where it belongs (per the project-wide GPU plan).

Changes

File	What
workflow/src/main.tpl.tengo	Wrap .gpuMemory() in if exec.hasGpu; always pass --mode gpu / --mode cpu so the software path matches
software/gpu-info/src/main.py	Add --mode {gpu,cpu}. CPU mode skips probes and emits a minimal report; GPU mode runs the existing flow unchanged. format_report() shows the active mode and refines the SUMMARY for CPU mode
model/src/index.ts	GpuReport.mode: 'gpu' | 'cpu'
ui/src/pages/GpuInfoPage.vue	Summary banner shows "CPU branch (exec.hasGpu = false)" alongside "GPU NOT AVAILABLE" when running on CPU
.changeset/milab-6382-gpu-support.md	Minor bump across workflow / gpu-info / model / ui

Existing user-controllable args (cpu, mem, gpuMemory, matrixSize, seed) keep working — gpuMemory is just ignored on the CPU branch.

Test plan

• python3 main.py --mode cpu --seed 42 → fast, no probes, SUMMARY: GPU NOT AVAILABLE (CPU branch, no probes attempted)
• python3 main.py --mode gpu --seed 42 → existing detection runs; on this dev host (no GPU) reports SUMMARY: GPU NOT AVAILABLE as before
• pnpm type-check (turbo) → 8/8 packages OK
• End-to-end on the GPU staging cluster after PR 1 (https://github.com/milaboratory/pl/pull/1934) deploys → pod lands on gpu-l4-24g, banner shows GPU AVAILABLE, benchmark reports speedup > 1
• End-to-end on a CPU-only cluster (or any cluster with PL_RUNNER_GPU_AVAILABLE=disabled) → block returns successfully, banner shows GPU NOT AVAILABLE + CPU branch (exec.hasGpu = false)

Pattern for follow-up block PRs

This block was the simplest case (already-passing gpuMemory arg, the Python diagnostic was easy to split into two branches). The remaining blocks need the same shape — workflow gate + explicit software flag + remove runtime detection. The harder ones:

• clonotype-space — switch the cuML auto backend to explicit cuml / sklearn per branch (no more try: import cuml); drop the fallback.
• 3d-structure-prediction — add a CUDA Dockerfile variant + --device cuda|cpu flag (currently CPU-pinned at the Dockerfile level).
• paratope-clustering — drop map_location="cpu", take device from --device.
• compositional-analysis — TF auto-detect → explicit CUDA_VISIBLE_DEVICES="" in CPU mode.
• immuno-match — HuggingFace Trainer no_cuda from --device.

🤖 Generated with Claude Code

[gemini-code-assist]

Code Review

This pull request introduces a dual-mode operation (GPU vs. CPU) for the GPU test workflow. The execution path is decided in the Tengo workflow using the exec.hasGpu gate, which passes a --mode flag to the Python backend. In CPU mode, the backend skips GPU probes and benchmarks to return a fast, schema-compatible report, and the UI is updated to display a CPU branch label. Feedback on the changes suggests adding the missing torch_cuda field to the TypeScript GpuReport interface to fully align with the backend's report schema.

Important

The consumer version of Gemini Code Assist on GitHub is being sunset. Starting June 18, 2026, new organization installations will be blocked, and all code review activity will officially cease on July 17, 2026.
For more details on the timeline and next steps, please review the Help Documentation.

[gemini-code-assist]

The torch_cuda field is populated by the Python backend (both in gpu and cpu modes) but is currently missing from the TypeScript GpuReport interface. Adding it ensures the TypeScript model fully and accurately represents the backend report schema.

Suggested change

	mode: 'gpu' | 'cpu';
	mode: 'gpu' | 'cpu';
	torch_cuda: {
	available: boolean;
	device_count: number;
	cuda_version: string | null;
	cudnn_version: string | null;
	devices: Array<{
	index: number;
	name: string;
	total_memory_mb: number;
	free_memory_mb?: number;
	major: number;
	minor: number;
	multi_processor_count: number;
	}>;
	error?: string;
	};