Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →HuggingFace Text Generation Inference (TGI). High-performance LLM serving with continuous batching, tensor parallelism, watermarking, and OpenAI-compatible API. Native HF model hub integration.
.claude/skills/mkurman-huggingface-tgi/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-05 | ✗→✓ | ▲ Improved | 24% | 0% |
| case-07 | ✗→✓ | ▲ Improved | -11% | 0% |
| case-08 | ✗→✓ | ▲ Improved | -21% | 0% |
| case-12 | ✗→✓ | ▲ Improved | -20% | 0% |
| case-15 | ✗→✓ | ▲ Improved | -38% | 0% |
Text Generation Inference (TGI) is a production-ready LLM serving solution from Hugging Face. It provides optimized inference with continuous batching, quantization (GPTQ, AWQ), tensor parallelism, flash attention, and an OpenAI-compatible API.
bash# Docker deployment (recommended) docker run --gpus all -p 8080:80 -v $HOME/models:/data ghcr.io/huggingface/text-generation-inference:latest --model-id Qwen/Qwen2.5-1.5B-Instruct
pythonfrom openai import OpenAI client = OpenAI(base_url="http://localhost:8080/v1", api_key="none") response = client.chat.completions.create( model="tgi", messages=[{"role": "user", "content": "Hello!"}], ) print(response.choices[0].message.content)
pythonstream = client.chat.completions.create( model="tgi", messages=[{"role": "user", "content": "Write a poem"}], stream=True, ) for chunk in stream: print(chunk.choices[0].delta.content or "", end="")
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | fail→fail | 8,756 | 3,955 | -55% | 1 | 1 | 0% | 1,737 | 1,145 | -34% | 0 | 0 | — |
case-02 | pass→pass | 9,582 | 5,410 | -44% | 1 | 1 | 0% | 1,799 | 1,365 | -24% | 0 | 0 | — |
case-03 | fail→fail | 7,563 | 4,978 | -34% | 1 | 1 | 0% | 1,479 | 1,241 | -16% | 0 | 0 | — |
case-04 | pass→pass | 7,337 | 3,594 | -51% | 1 | 1 | 0% | 1,252 | 986 | -21% | 0 | 0 | — |
case-05 | fail→pass | 4,432 | 3,710 | -16% | 1 | 1 | 0% | 822 | 1,018 | +24% | 0 | 0 | — |
case-06 | pass→pass | 6,251 | 5,034 | -19% | 1 | 1 | 0% | 1,228 | 1,241 | +1% | 0 | 0 | — |
case-07 | fail→pass | 4,825 | 2,775 | -42% | 1 | 1 | 0% | 937 | 836 | -11% | 0 | 0 | — |
case-08 | fail→pass | 12,538 | 8,058 | -36% | 1 | 1 | 0% | 2,197 | 1,736 | -21% | 0 | 0 | — |
case-09 | pass→pass | 4,590 | 2,864 | -38% | 1 | 1 | 0% | 778 | 791 | +2% | 0 | 0 | — |
case-10 | pass→pass | 2,751 | 2,385 | -13% | 1 | 1 | 0% | 507 | 617 | +22% | 0 | 0 | — |
case-11 | pass→pass | 2,846 | 2,132 | -25% | 1 | 1 | 0% | 455 | 706 | +55% | 0 | 0 | — |
case-12 | fail→pass | 4,476 | 2,008 | -55% | 1 | 1 | 0% | 841 | 674 | -20% | 0 | 0 | — |
case-13 | pass→pass | 4,311 | 1,811 | -58% | 1 | 1 | 0% | 711 | 577 | -19% | 0 | 0 | — |
case-14 | pass→pass | 2,392 | 1,867 | -22% | 1 | 1 | 0% | 386 | 625 | +62% | 0 | 0 | — |
case-15 | fail→pass | 5,570 | 1,778 | -68% | 1 | 1 | 0% | 1,010 | 624 | -38% | 0 | 0 | — |
case-16 | pass→pass | 3,604 | 1,716 | -52% | 1 | 1 | 0% | 607 | 636 | +5% | 0 | 0 | — |
case-17 | pass→pass | 5,215 | 1,747 | -67% | 1 | 1 | 0% | 881 | 614 | -30% | 0 | 0 | — |
case-18 | pass→pass | 6,889 | 4,593 | -33% | 1 | 1 | 0% | 1,360 | 1,382 | +2% | 0 | 0 | — |
case-19 | pass→pass | 2,119 | 2,289 | +8% | 1 | 1 | 0% | 330 | 721 | +118% | 0 | 0 | — |
case-20 | pass→pass | 10,466 | 8,236 | -21% | 1 | 1 | 0% | 2,066 | 2,047 | -1% | 0 | 0 | — |
case-21 | pass→pass | 15,109 | 12,865 | -15% | 1 | 1 | 0% | 3,157 | 2,886 | -9% | 0 | 0 | — |
case-22 | pass→pass | 8,241 | 6,158 | -25% | 1 | 1 | 0% | 1,556 | 1,634 | +5% | 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. 22 cases were attempted. The headline lift of +23 percentage points is the difference between those two pass rates over the 22 comparable cases.
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.