Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Fast LLM inference engine. PagedAttention, continuous batching, tensor parallelism, speculative decoding, and prefix caching. OpenAI-compatible API server. Supports Llama, Mistral, Qwen, DeepSeek, and hundreds of models.
.claude/skills/mkurman-vllm/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-04 | ✗→✓ | ▲ Improved | -39% | 0% |
| case-05 | ✗→✓ | ▲ Improved | -19% | 0% |
| case-10 | ✓→✗ | ▼ Worse | 17% | 0% |
| case-01 | ✓→✓ | ＝ Same ✓ | 47% | 0% |
| case-02 | ✓→✓ | ＝ Same ✓ | 75% | 0% |
vLLM is a high-throughput, memory-efficient LLM inference engine featuring PagedAttention (near-zero memory waste), continuous batching, tensor parallelism, speculative decoding, prefix caching, and an OpenAI-compatible API.
bashuv pip install vllm
pythonfrom vllm import LLM, SamplingParams llm = LLM(model="Qwen/Qwen2.5-1.5B-Instruct") params = SamplingParams(temperature=0.7, top_p=0.9, max_tokens=512) outputs = llm.generate(["What is the capital of France?"], params) for o in outputs: print(o.outputs[0].text)
bashvllm serve Qwen/Qwen2.5-1.5B-Instruct --port 8000 # OpenAI client: curl http://localhost:8000/v1/chat/completions \ -H "Content-Type: application/json" \ -d '{"model": "Qwen/Qwen2.5-1.5B-Instruct", "messages": [{"role": "user", "content": "Hello!"}]}'
pythonllm = LLM(model="meta-llama/Llama-3.1-8B", tensor_parallel_size=2)
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | pass→pass | 3,060 | 1,899 | -38% | 1 | 1 | 0% | 449 | 661 | +47% | 0 | 0 | — |
case-02 | pass→pass | 2,200 | 1,543 | -30% | 1 | 1 | 0% | 325 | 568 | +75% | 0 | 0 | — |
case-03 | pass→pass | 7,595 | 5,124 | -33% | 1 | 1 | 0% | 1,433 | 1,284 | -10% | 0 | 0 | — |
case-04 | fail→pass | 6,635 | 2,204 | -67% | 1 | 1 | 0% | 1,265 | 775 | -39% | 0 | 0 | — |
case-05 | fail→pass | 4,982 | 2,918 | -41% | 1 | 1 | 0% | 911 | 740 | -19% | 0 | 0 | — |
case-06 | pass→pass | 3,767 | 3,713 | -1% | 1 | 1 | 0% | 701 | 936 | +34% | 0 | 0 | — |
case-07 | pass→pass | 2,645 | 1,747 | -34% | 1 | 1 | 0% | 466 | 676 | +45% | 0 | 0 | — |
case-08 | pass→pass | 4,386 | 2,520 | -43% | 1 | 1 | 0% | 918 | 858 | -7% | 0 | 0 | — |
case-09 | pass→pass | 9,022 | 5,718 | -37% | 1 | 1 | 0% | 1,659 | 1,404 | -15% | 0 | 0 | — |
case-10 | pass→fail | 4,598 | 3,258 | -29% | 1 | 1 | 0% | 893 | 1,047 | +17% | 0 | 0 | — |
case-11 | pass→pass | 2,341 | 1,428 | -39% | 1 | 1 | 0% | 357 | 558 | +56% | 0 | 0 | — |
case-12 | pass→pass | 2,934 | 1,279 | -56% | 1 | 1 | 0% | 390 | 571 | +46% | 0 | 0 | — |
case-13 | pass→pass | 9,623 | 4,834 | -50% | 1 | 1 | 0% | 1,772 | 1,326 | -25% | 0 | 0 | — |
case-14 | pass→pass | 3,160 | 1,500 | -53% | 1 | 1 | 0% | 460 | 603 | +31% | 0 | 0 | — |
case-15 | pass→pass | 3,447 | 2,588 | -25% | 1 | 1 | 0% | 567 | 812 | +43% | 0 | 0 | — |
case-16 | pass→pass | 5,830 | 4,145 | -29% | 1 | 1 | 0% | 1,105 | 1,137 | +3% | 0 | 0 | — |
case-17 | pass→pass | 4,398 | 3,033 | -31% | 1 | 1 | 0% | 686 | 854 | +24% | 0 | 0 | — |
case-18 | pass→pass | 5,740 | 3,482 | -39% | 1 | 1 | 0% | 1,037 | 1,022 | -1% | 0 | 0 | — |
case-19 | pass→pass | 2,824 | 1,679 | -41% | 1 | 1 | 0% | 400 | 600 | +50% | 0 | 0 | — |
case-20 | pass→pass | 2,525 | 1,401 | -45% | 1 | 1 | 0% | 365 | 569 | +56% | 0 | 0 | — |
case-21 | pass→pass | 17,740 | 15,986 | -10% | 1 | 1 | 0% | 3,445 | 3,710 | +8% | 0 | 0 | — |
case-22 | pass→pass | 13,151 | 15,493 | +18% | 1 | 1 | 0% | 2,407 | 2,358 | -2% | 0 | 0 | — |
case-23 | pass→pass | 14,352 | 9,413 | -34% | 1 | 1 | 0% | 2,540 | 2,181 | -14% | 0 | 0 | — |
DecimalAI ran this skill against gemini-3.6-flash twice over the same eval suite — once with the skill loaded and once without — and compared the two runs case by case. 23 cases were attempted. The headline lift of +4 percentage points is the difference between those two pass rates over the 23 comparable cases. 1 case got worse with the skill loaded, and it is included in that figure.
Without the skill loaded, the model failed this case. With it loaded, the same prompt on the same model passed. This is one improved case from the latest verified run; every case, including any that regressed, is in the table above.
Other measured skills in the registry, with their headline benchmark lift.