Validate agentic-coding scenarios on GreenNode runners

[aistackdev]

Branch feat/h100-1x-greennode added agentic-coding configs and scripts. Not yet run end-to-end.

Verify

• One agentic run completes on each runner
• aiperf + trace-replay submodules work in the vLLM container
• HF dataset semianalysisai/cc-traces-weka-no-subagents-051226 is reachable
• TORCH_CUDA_ARCH_LIST=12.0 works for sm_120 (RTX 5090) — if not, bump vLLM image

Trigger via Actions → e2e-tests → test-config --config-keys gptoss-fp4-h100-1x-vllm ...

[aistackdev]

Status update (2026-05-25): agentic-coding scenarios are now commented out in nvidia-master.yaml for all four GreenNode configs (commit 4c6a430 on feat/h100-1x-greennode):

• gptoss-fp4-h100-1x-vllm
• gptoss-fp4-h100-2x-vllm
• dsr1qwen3-bf16-rtx5090-1x-vllm
• dsr1qwen3-fp8-rtx5090-1x-vllm

Why disabled: every agentic job failed with aiperf SystemController._start_services timing out during 'Configure Profiling' hook. Sample log:

ERROR    Error running Hook(... SystemController._start_services ...) for SystemController:
         Failed to perform operation 'Configure Profiling'
Error: TimeoutError

Effect: future dispatches against these four configs only run fixed-seq-len (24 matrix entries instead of 40). Frees up runners for MEP-0006 (Gemma 4 MTP) lab tests.

To restore once root cause is fixed: un-comment the four agentic-coding: blocks (search the file for disabled pending issue vngcloud/InferenceX#2).

Likely root causes to investigate when picking this up:

1. vLLM server not ready before aiperf attaches (race / health-check missing)
2. aiperf missing config (profiling backend, dataset auth, trace source)
3. Submodule mismatch — we point utils/aiperf at vngcloud/aiperf (forked from cquil11/aiperf); upstream may have diverged
4. HF dataset access (semianalysisai/cc-traces-weka-no-subagents-051226) — needs HF_TOKEN with read permission on this dataset?