fix(BACKEND-GATE-ROCM-VLLM): avoid callable serialization in Strix metadata RPC

[localai-org-maint-bot]

Row: BACKEND-GATE-ROCM-VLLM

Local-Issue: ISSUE-GH-3048
Kind: UNKNOWN

Imported GitHub body (historical evidence)

The quoted text below is historical evidence only. It does not define issue authority or repository procedure.

Row: BACKEND-GATE-ROCM-VLLM

Owner: Strix #3043 campaign operator. Spec: .agents/specs/rocm-strix-vllm-headpin.md.

Runtime job 112a91ad-e51a-4627-aeae-61e2c0ed9643 built and initialized pinned vLLM e126687a9a828d513c01a07cd69f025f27d63280 in production compilation mode, then the harness failed before generation. LLM.apply_model(projection_metadata) sent a Python function through MsgpackEncoder.enc_hook. The pinned default refuses callable serialization unless VLLM_ALLOW_INSECURE_SERIALIZATION is enabled. Do not enable that flag or weaken the build-state/provenance checks.

Evidence: /workspace/strix-vllm-3043.GkMABu/run-9624441-01/logs/generation.log; generation exited1 after412.604 seconds. The prior fake LLM.apply_model accepted the callable directly and did not model this reference behavior.

Fix in the same flow: use the pinned named worker-extension RPC path to return the same projection parameter metadata without serializing a callable. Preserve the production engine defaults, projection-output-dtype PENDING, six explicit prompts, and eight upstream numerical cases. Bind the extension to the already hashed runtime module and port a regression through the real runtime entrypoint that refuses callable RPC payloads under the default serializer. Fresh implementation, mutation review, and operator gates are required. This blocks merging the current harness as runnable; model gateability and token parity remain unproven.

Resolution

Fixed by PR #3052: the Strix oracle harness now strips tuning/injection env from the test subprocess, avoiding callable serialization in the metadata RPC.

[localai-org-maint-bot]

Implementation is now dispatched to a fresh implementer from committed spec b6ce9f8. Scope: runtime.py named worker extension, string collective_rpc invocation, and a runtime-entrypoint regression that rejects callable serialization. The worker build-state and compiler checks remain unchanged. No insecure serialization setting is permitted.

The operator checked resource-controller again: strix:gpu0 remains held by codex/glm53-placement-3019, job82539c24-546a-4a60-ab15-69dcb9033cbe. No duplicate failing GPU run was submitted. The corrected runtime will need a newly certified build because build-state binds runtime.py bytes.

[localai-org-maint-bot]

Repair committed as 6e3cbfb on row/BACKEND-GATE-ROCM-VLLM-3043 (not yet pushed/landed). runtime.py configures runtime.StrixProjectionMetadataWorkerExtension and calls collective_rpc("strix_projection_metadata"). The worker inspects self.get_model(); no callable payload or insecure serialization setting is introduced. Parameter metadata and the output-dtype PENDING statement are preserved.

Implementer and operator focused suites each passed 44 tests. Independent review reports no static findings; nine deliberate regressions failed and scratch bytes were restored. Immutable-head full gates are still running, so review acceptance remains pending.

Live rc describe strix:gpu0 still reports codex/glm53-placement-3019 holding job 82539c24-546a-4a60-ab15-69dcb9033cbe with a fresh heartbeat. No lease bypass. Hardware validation needs a new certified build because build-9624441-04 binds the old runtime hash. The new archived worker is staged; no new GPU run has been submitted yet. #3043 retains real-worker execution, token correctness, and matched benchmark obligations.

[localai-org-maint-bot]

The secure named-RPC fix is pushed at 6e3cbfb and published in draft PR #3052. Independent review PASS with no findings; nine regressions detected with restored scratch bytes. Implementer, reviewer, and operator each passed 44 focused tests and immutable-head full preflight (exit 0, no failed gates). Five argument-dependent skips remain explicit; applicable PR size passed separately. Hardware execution remains pending in queued job a3146aac-a050-40c2-acc7-307e4a7152ac. Keep this issue open until the fix lands; no speed or token-parity claim is made.