[yxf0314]

Publish device mappings only after detection completes and retry empty results on the next lookup. Reject Docker workloads when their requested accelerator mapping is missing so a privileged container cannot start with an empty device mask.

Ref gpustack/gpustack#6755

Validation

• 273 deployer tests passed. make lint passed.
• On Ascend 910B2 with the real Docker API, the forced overlap reproduced the empty prefill mask in 5 of 5 runs on Runtime 0.2.7.post1. All 5 runs with the candidate wheel produced backend masks 2,3 for prefill and 0,1 for decode.
• An empty discovery result blocked container creation. A subsequent request on the same deployer succeeded once discovery recovered.

Hardware validation covered Docker container configuration. Full vLLM PD inference is pending.

[gemini-code-assist]

Code Review

This pull request introduces thread-safe device material detection and caching across the deployer modules. It adds a locking mechanism to prevent race conditions during device detection, updates the Docker container creation logic to handle device mapping more robustly, and includes comprehensive tests to verify concurrent device detection and error handling. I have no feedback to provide as there were no review comments.

[Copilot]

🟡 Changes recommended

Empty device discovery allows privileged accelerator requests to pass with fallback-only UNKNOWN_* masks, permitting container creation without enforcing the requested device selection.

1 open issue

• Reject fallback-only material for auto-mapped accelerator requests · New

What changed in this PR

This PR addresses missing accelerator mappings in GPUStack Runtime’s deployment layer, targeting the incorrect Ascend device selection reported in gpustack/gpustack#6755.

Changes:

• Publish completed device mappings under a lock and retry discovery after empty results.
• Reject empty Docker device mappings and adapt Kubernetes’ policy-specific cache.
• Add tests for missing Ascend mappings, discovery recovery, and concurrent lookups.

File	Description
tests/​gpustack_runtime/​deployer/​test_visible_devices_ordering.py	Adds rejection, retry, and concurrent mapping tests.
gpustack_runtime/​deployer/​kubernetes.py	Adapts policy-specific caching to the new material lookup flow.
gpustack_runtime/​deployer/​docker.py	Consolidates device requests and rejects empty mappings.
gpustack_runtime/​deployer/​__types__.py	Synchronizes cache initialization and leaves empty discovery results uncached.

_{Review effort: Balanced}

[thxCode]

LGTM