test: energy audit demo trigger CI workflow

[hongping-zh]

This PR adds a test Python file with intentional energy waste patterns to verify the EcoCompute Energy Audit GitHub Action.

Expected results:

• Critical: Default INT8 without threshold fix
• Warning: BS=1 loop pattern

[github-actions]

⚡ EcoCompute Energy Audit

🖥️ Hardware Environment

⚠️ No GPU detected. Energy measurement requires NVIDIA GPU. This run uses static analysis + estimation only.
Scanned 1 Python file(s). Found 1 critical, 1 warning(s).

🔴 Critical Issues

Default INT8 (bitsandbytes mixed-precision decomposition) — test_energy.py (line 9)

load_in_8bit=True without llm_int8_threshold=0.0 causes 17–147% energy waste due to INT8↔FP16 type conversion at every linear layer. Measured on RTX 4090D (+32.7%) and A800 (+122–147%).

Energy impact: +17–147% energy vs FP16

Fix: Add llm_int8_threshold=0.0 to your BitsAndBytesConfig:

config = BitsAndBytesConfig(
    load_in_8bit=True,
    llm_int8_threshold=0.0,  # Disables mixed-precision decomposition
)

🟡 Warnings

Sequential single-request processing (BS=1) — test_energy.py (line 19)

Processing prompts in a loop wastes up to 95.7% energy vs batched inference. Measured on A800: BS=1 → 1,768 J/request, BS=64 → 76 J/request. GPU utilization at BS=1 is only 45%.

Energy impact: Up to 95.7% energy waste vs batched

Fix: Batch your inputs or use a serving framework:

# Option 1: Batch with tokenizer
inputs = tokenizer(prompts, padding=True, return_tensors='pt').to('cuda')
outputs = model.generate(**inputs)

# Option 2: Use vLLM for production
from vllm import LLM
llm = LLM(model=model_name)
outputs = llm.generate(prompts)

📈 Relative Change (vs Baseline)

No baseline found. This run will be saved as the new baseline.

---

📊 Based on 93+ measurements across RTX 4090D / A800 / RTX 5090 · Full data · Install Bot · OpenClaw Skill

[hongping-zh]

Test completed successfully. All detection rules working as expected. Closing test PR.