[sabira-shaik]

Summary

Three fixes to serving-llms-on-instinct, found while testing the skill on an MI300X.

1. Docker check never fails (scripts/validate.py)

docker ps -q 2>&1 | head -1 returns head's exit code (always 0), and 2>&1 moves Docker's error into stdout, so validate.py reported ready: true with Docker unreachable. It now runs plain docker ps -q.

$ DOCKER_HOST=unix:///nope docker ps -q 2>&1 | head -1; echo "exit=$?"
failed to connect to the docker API at unix:///nope ...
exit=0
$ DOCKER_HOST=unix:///nope docker ps -q; echo "exit=$?"
failed to connect to the docker API at unix:///nope ...
exit=1

2. gfx950 gpt-oss override never applied (SKILL.md)

data/gpu_overrides.json sets VLLM_ROCM_USE_AITER_MOE=0 for openai/gpt-oss on gfx950 (the AITER MoE path corrupts its output), but SKILL.md never mentions model_overrides, so the agent never applies it. Step 4 now merges model_overrides[].env_set when the model ID matches.

3. Fit check ignores memory already in use (scripts/estimate_vram.py, SKILL.md)

The fit check assumed the whole GPU was free. vLLM's --gpu-memory-utilization is a fraction of total VRAM and vLLM refuses to start if that much is not free, so with another vLLM container running, the skill said a model fits and the launch failed. This also makes two models on one GPU a trial-and-error loop.

estimate_vram.py now reads free VRAM from amd-smi metric --mem-usage (or --free-vram-gb for remote hosts), caps gpu_memory_utilization to what is free (2 GB margin), and reports free_vram_gb, min_needed_gb (weights + 4 GB overhead + 2 GB KV) and can_serve. SKILL.md stops before launch when can_serve is false and otherwise uses the computed utilization.

MI300X, with another vLLM container (Qwen3.5-9B at util 0.9) holding 175 GB:

Model	Free	Before	After
Qwen3-0.6B	20.5 GB	fits, util 0.9	can_serve: true, util 0.09
openai/gpt-oss-20b	20.5 GB	fits, util 0.9, 156 GB for KV	can_serve: false (needs 18.8 GB)
Qwen3-32B	20.5 GB	fits, util 0.9	can_serve: false (needs 67 GB)
openai/gpt-oss-20b, --free-vram-gb 191.7	191.7 GB	fits	can_serve: true, util 0.9 (unchanged)

Testing

• Reproduced the Docker check failure above.
• Ran estimate_vram.py on an MI300X (ROCm 7.2) for the cases in the table.
• ./.github/scripts/check.sh passes.

[danielholanda]

@TwelveNights