Summary

Delete the standalone GPU conformance workflow. Training and conformance are mostly identical today, so removing the standalone workflow cuts H100x2 runner usage without changing the existing training or inference workflow shape.

Parent: #541

Problem

• Training and standalone conformance overlap heavily in coverage.
• Both use linux-amd64-gpu-h100-latest-2, so a single PR consumes two H100x2 jobs.
• The standalone conformance workflow adds runner contention for limited additional signal.
• The current low-risk win is removing the third workflow, not redesigning training and inference execution.

Proposal

• Delete .github/workflows/gpu-h100-conformance-test.yaml.
• Keep the training workflow unchanged.
• Keep the inference workflow unchanged.
• Preserve any trigger coverage that would otherwise be lost by deleting the standalone conformance workflow.
• Fold only the required conformance-specific paths into the remaining workflow filters.
• Explicitly preserve .github/actions/setup-build-tools/** coverage, since the deleted workflow currently carries that trigger coverage for GPU training.

Acceptance Criteria

• The standalone conformance workflow is removed.
• Training continues to run its current conformance coverage.
• Inference continues to run its current conformance coverage.
• Required trigger coverage from the removed workflow is preserved.
• .github/actions/setup-build-tools/** remains covered by the remaining GPU workflow triggers.
• Inference remains on its current runner class rather than moving onto H100x2.

Component

Multiple components

Priority

Important (would improve my workflow)

Out of Scope

• Combining training and inference into one workflow or one job.
• Intent-aware triggers.
• Broader path-filter redesign beyond what is required for deleting the standalone conformance workflow.
• Local registry migration from kind load.
• Inference wait and retry hardening.

Related

• GPU CI: optimize build/deploy performance, deduplicate workflows, and simplify behavioral tests #541
• PR fix(ci): improve GPU test reliability and deploy timeout handling #539