Add optional Windows ROCm backend for DiffusionGemma

[Yasei-no-otoko]

Add an optional Windows ROCm backend for the original DiffusionGemma checkpoint, validated on Radeon 8060S (gfx1151). LocalJev continues to use its OpenAI-compatible upstream interface: users can connect an existing ROCm server through the normal environment variables, or choose the bundled PyTorch/Transformers adapter. The oMLX defaults and decision/evaluation logic stay unchanged.

The implementation includes:

• A text Chat Completions adapter with explicit HIP placement, SDPA/eager MoE, dynamic cache, bounded serialized inference, request validation and optional authentication. JSON schemas are prompt instructions; LocalJev retains output validation and corrective retries.
• A Windows PowerShell launcher with dependency/GPU/port checks, readiness waiting, explicit environment handling and cleanup of its owned backend process tree.
• An HTTP smoke command covering all three Jev question types, plus a per-request HTTP timeout adjustment so long inference can finish under the Engine's existing upstream timeout policy.

ROCmFPX is not a dependency. It was used only as an external runner for supplementary Q8 validation. This PR includes no ROCmFPX patches, build scripts, models or evaluation dumps. The guide links that historical result separately and does not present its 240/240 coverage as an FP16-adapter result. Python dependencies are optional and deliberately do not install/replace the user's HIP PyTorch wheel.

Validation for this PR:

• Frozen Bun install; 27 Bun tests; TypeScript type check; 12 CPU-only Python tests: passed.
• start-rocm.ps1 -CheckOnly passed with Windows PowerShell 5.1 and PowerShell 7 on the gfx1151 host, without loading model weights.
• Implementation and test files match the previously GPU-validated ROCm change. The guide records two earlier real FP16 /ready → /v1/systemone smoke requests on September 20; both validated choice, score and noul outputs.
• Diff/scope review confirms no changes to the Engine, configuration defaults, evaluator, scoring or lockfile.

OpenAI Codex assisted with the implementation, validation and PR preparation at the contributor's request.

[Copilot]

Copilot review overview

🔵 Needs a closer look

Four moderate issues remain in model loading, rejected-request handling, numeric validation, and request-wide timeout enforcement.

Review effort: Lite
Findings: None

What changed in this PR

Adds an optional Windows ROCm/PyTorch backend for DiffusionGemma, with launch tooling, smoke tests, and documentation.

Changes:

• Adds ROCm server, optional dependencies, and PowerShell lifecycle management.
• Adds HTTP smoke tooling and backend tests.
• Adjusts request timeout handling for long inference.

File	Description
test/​test_rocm_server.py	Backend validation tests
test/​index.test.ts	Timeout regression tests
test/​http-smoke.test.ts	HTTP smoke tests
src/​index.ts	Inference timeout handling
scripts/​start-rocm.ps1	Windows launcher and cleanup
scripts/​rocm_server.py	ROCm-compatible inference backend
scripts/​http-smoke.ts	End-to-end smoke command
requirements-rocm.txt	Optional Python dependencies
README.md	ROCm overview and commands
package.json	Bun configuration and scripts
docs/​gfx1151.md	Windows ROCm setup guide
.gitignore	ROCm runtime exclusions

---

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