[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M6MB4X4NH4BCRGAAT16XQFY9":3,"skill-same:01M6MB4X4NH4BCRGAAT16XQFY9":98},{"skillNo":4,"name":5,"nameZh":6,"description":7,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":16,"categories":17,"keywords":23,"publishedAt":45,"linkOnly":46,"markdownContent":47,"markdownBody":48,"frontMatter":49,"refName":56,"commitSha":57,"skillDirectory":58,"forksCount":59,"repositoryHtmlUrl":60,"repositoryHomepageUrl":61,"ownerHtmlUrl":62,"ownerBlogUrl":63,"ownerDescription":64,"updatedAt":65,"install":66,"resources":79},"01M6MB4X4NH4BCRGAAT16XQFY9","serving-llms-vllm",null,"Serves LLMs with high throughput using vLLM's PagedAttention and continuous batching. Use when deploying production LLM APIs, optimizing inference latency\u002Fthroughput, or serving models with limited GPU memory. Supports OpenAI-compatible endpoints, quantization (GPTQ\u002FAWQ\u002FFP8), and tensor parallelism.","hashgraph-online\u002Fawesome-codex-plugins","hashgraph-online","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F160371962?v=4","01M6CNB6NRJVNNGN14TR01Q6D3","HOL (Hashgraph Online)","01M6JT1HSKKG9V68KQAQGBA7XP",1566,"MIT","https:\u002F\u002Fgithub.com\u002Fhashgraph-online\u002Fawesome-codex-plugins\u002Fblob\u002Fmain\u002Fplugins\u002FKalarisLabs\u002Fresearch-agent-skills\u002Fskills\u002Fserving-llms-vllm\u002FSKILL.md",[18],{"categoryNo":19,"categoryName":20,"taxonomyType":21,"taxonomyTypeDesc":22},"01JBSEEDCAT000000000000010","后端与接口","SCENARIO","使用场景",[24,25,26,27,28,29,30,31,32,33,34,35,36,37,38,39,40,41,42,43,44],"serving","vLLM","llms","serving llms vllm","ops","when","batching","openai-compatible","throughput","GPTQ","AWQ","latency","use","api","Transformers","Inference Serving","FP8","openai api","openai","ml","memory","2026-11-10 10:12:30",false,"---\nname: serving-llms-vllm\ndescription: Serves LLMs with high throughput using vLLM's PagedAttention and continuous batching. Use when deploying production LLM APIs, optimizing inference latency\u002Fthroughput, or serving models with limited GPU memory. Supports OpenAI-compatible endpoints, quantization (GPTQ\u002FAWQ\u002FFP8), and tensor parallelism.\nlicense: MIT\nmetadata:\n  version: 1.0.0\n  category: ml-inference-and-ops\n  maintainer: Kalaris Labs\n  tags: vLLM, Inference Serving, PagedAttention, Continuous Batching, High Throughput, Production, OpenAI API, Quantization, Tensor Parallelism\n  dependencies: vllm, torch, transformers\n---\n\n# vLLM - High-Performance LLM Serving\n\n## Quick start\n\nvLLM achieves 24x higher throughput than standard transformers through PagedAttention (block-based KV cache) and continuous batching (mixing prefill\u002Fdecode requests).\n\n**Installation**:\n```bash\npip install vllm\n```\n\n**Basic offline inference**:\n```python\nfrom vllm import LLM, SamplingParams\n\nllm = LLM(model=\"meta-llama\u002FLlama-3-8B-Instruct\")\nsampling = SamplingParams(temperature=0.7, max_tokens=256)\n\noutputs = llm.generate([\"Explain quantum computing\"], sampling)\nprint(outputs[0].outputs[0].text)\n```\n\n**OpenAI-compatible server**:\n```bash\nvllm serve meta-llama\u002FLlama-3-8B-Instruct\n\n# Query with OpenAI SDK\npython -c \"\nfrom openai import OpenAI\nclient = OpenAI(base_url='http:\u002F\u002Flocalhost:8000\u002Fv1', api_key='EMPTY')\nprint(client.chat.completions.create(\n    model='meta-llama\u002FLlama-3-8B-Instruct',\n    messages=[{'role': 'user', 'content': 'Hello!'}]\n).choices[0].message.content)\n\"\n```\n\n## Common workflows\n\n### Workflow 1: Production API deployment\n\nCopy this checklist and track progress:\n\n```\nDeployment Progress:\n- [ ] Step 1: Configure server settings\n- [ ] Step 2: Test with limited traffic\n- [ ] Step 3: Enable monitoring\n- [ ] Step 4: Deploy to production\n- [ ] Step 5: Verify performance metrics\n```\n\n**Step 1: Configure server settings**\n\nChoose configuration based on your model size:\n\n```bash\n# For 7B-13B models on single GPU\nvllm serve meta-llama\u002FLlama-3-8B-Instruct \\\n  --gpu-memory-utilization 0.9 \\\n  --max-model-len 8192 \\\n  --port 8000\n\n# For 30B-70B models with tensor parallelism\nvllm serve meta-llama\u002FLlama-2-70b-hf \\\n  --tensor-parallel-size 4 \\\n  --gpu-memory-utilization 0.9 \\\n  --quantization awq \\\n  --port 8000\n\n# For production with caching and metrics\nvllm serve meta-llama\u002FLlama-3-8B-Instruct \\\n  --gpu-memory-utilization 0.9 \\\n  --enable-prefix-caching \\\n  --enable-metrics \\\n  --metrics-port 9090 \\\n  --port 8000 \\\n  --host 0.0.0.0\n```\n\n**Step 2: Test with limited traffic**\n\nRun load test before production:\n\n```bash\n# Install load testing tool\npip install locust\n\n# Create test_load.py with sample requests\n# Run: locust -f test_load.py --host http:\u002F\u002Flocalhost:8000\n```\n\nVerify TTFT (time to first token) \u003C 500ms and throughput > 100 req\u002Fsec.\n\n**Step 3: Enable monitoring**\n\nvLLM exposes Prometheus metrics on port 9090:\n\n```bash\ncurl http:\u002F\u002Flocalhost:9090\u002Fmetrics | grep vllm\n```\n\nKey metrics to monitor:\n- `vllm:time_to_first_token_seconds` - Latency\n- `vllm:num_requests_running` - Active requests\n- `vllm:gpu_cache_usage_perc` - KV cache utilization\n\n**Step 4: Deploy to production**\n\nUse Docker for consistent deployment:\n\n```bash\n# Run vLLM in Docker\ndocker run --gpus all -p 8000:8000 \\\n  vllm\u002Fvllm-openai:latest \\\n  --model meta-llama\u002FLlama-3-8B-Instruct \\\n  --gpu-memory-utilization 0.9 \\\n  --enable-prefix-caching\n```\n\n**Step 5: Verify performance metrics**\n\nCheck that deployment meets targets:\n- TTFT \u003C 500ms (for short prompts)\n- Throughput > target req\u002Fsec\n- GPU utilization > 80%\n- No OOM errors in logs\n\n### Workflow 2: Offline batch inference\n\nFor processing large datasets without server overhead.\n\nCopy this checklist:\n\n```\nBatch Processing:\n- [ ] Step 1: Prepare input data\n- [ ] Step 2: Configure LLM engine\n- [ ] Step 3: Run batch inference\n- [ ] Step 4: Process results\n```\n\n**Step 1: Prepare input data**\n\n```python\n# Load prompts from file\nprompts = []\nwith open(\"prompts.txt\") as f:\n    prompts = [line.strip() for line in f]\n\nprint(f\"Loaded {len(prompts)} prompts\")\n```\n\n**Step 2: Configure LLM engine**\n\n```python\nfrom vllm import LLM, SamplingParams\n\nllm = LLM(\n    model=\"meta-llama\u002FLlama-3-8B-Instruct\",\n    tensor_parallel_size=2,  # Use 2 GPUs\n    gpu_memory_utilization=0.9,\n    max_model_len=4096\n)\n\nsampling = SamplingParams(\n    temperature=0.7,\n    top_p=0.95,\n    max_tokens=512,\n    stop=[\"\u003C\u002Fs>\", \"\\n\\n\"]\n)\n```\n\n**Step 3: Run batch inference**\n\nvLLM automatically batches requests for efficiency:\n\n```python\n# Process all prompts in one call\noutputs = llm.generate(prompts, sampling)\n\n# vLLM handles batching internally\n# No need to manually chunk prompts\n```\n\n**Step 4: Process results**\n\n```python\n# Extract generated text\nresults = []\nfor output in outputs:\n    prompt = output.prompt\n    generated = output.outputs[0].text\n    results.append({\n        \"prompt\": prompt,\n        \"generated\": generated,\n        \"tokens\": len(output.outputs[0].token_ids)\n    })\n\n# Save to file\nimport json\nwith open(\"results.jsonl\", \"w\") as f:\n    for result in results:\n        f.write(json.dumps(result) + \"\\n\")\n\nprint(f\"Processed {len(results)} prompts\")\n```\n\n### Workflow 3: Quantized model serving\n\nFit large models in limited GPU memory.\n\n```\nQuantization Setup:\n- [ ] Step 1: Choose quantization method\n- [ ] Step 2: Find or create quantized model\n- [ ] Step 3: Launch with quantization flag\n- [ ] Step 4: Verify accuracy\n```\n\n**Step 1: Choose quantization method**\n\n- **AWQ**: Best for 70B models, minimal accuracy loss\n- **GPTQ**: Wide model support, good compression\n- **FP8**: Fastest on H100 GPUs\n\n**Step 2: Find or create quantized model**\n\nUse pre-quantized models from HuggingFace:\n\n```bash\n# Search for AWQ models\n# Example: TheBloke\u002FLlama-2-70B-AWQ\n```\n\n**Step 3: Launch with quantization flag**\n\n```bash\n# Using pre-quantized model\nvllm serve TheBloke\u002FLlama-2-70B-AWQ \\\n  --quantization awq \\\n  --tensor-parallel-size 1 \\\n  --gpu-memory-utilization 0.95\n\n# Results: 70B model in ~40GB VRAM\n```\n\n**Step 4: Verify accuracy**\n\nTest outputs match expected quality:\n\n```python\n# Compare quantized vs non-quantized responses\n# Verify task-specific performance unchanged\n```\n\n## When to use vs alternatives\n\n**Use vLLM when:**\n- Deploying production LLM APIs (100+ req\u002Fsec)\n- Serving OpenAI-compatible endpoints\n- Limited GPU memory but need large models\n- Multi-user applications (chatbots, assistants)\n- Need low latency with high throughput\n\n**Use alternatives instead:**\n- **llama.cpp**: CPU\u002Fedge inference, single-user\n- **HuggingFace transformers**: Research, prototyping, one-off generation\n- **TensorRT-LLM**: NVIDIA-only, need absolute maximum performance\n- **Text-Generation-Inference**: Already in HuggingFace ecosystem\n\n## Common issues\n\n**Issue: Out of memory during model loading**\n\nReduce memory usage:\n```bash\nvllm serve MODEL \\\n  --gpu-memory-utilization 0.7 \\\n  --max-model-len 4096\n```\n\nOr use quantization:\n```bash\nvllm serve MODEL --quantization awq\n```\n\n**Issue: Slow first token (TTFT > 1 second)**\n\nEnable prefix caching for repeated prompts:\n```bash\nvllm serve MODEL --enable-prefix-caching\n```\n\nFor long prompts, enable chunked prefill:\n```bash\nvllm serve MODEL --enable-chunked-prefill\n```\n\n**Issue: Model not found error**\n\nUse `--trust-remote-code` for custom models:\n```bash\nvllm serve MODEL --trust-remote-code\n```\n\n**Issue: Low throughput (\u003C50 req\u002Fsec)**\n\nIncrease concurrent sequences:\n```bash\nvllm serve MODEL --max-num-seqs 512\n```\n\nCheck GPU utilization with `nvidia-smi` - should be >80%.\n\n**Issue: Inference slower than expected**\n\nVerify tensor parallelism uses power of 2 GPUs:\n```bash\nvllm serve MODEL --tensor-parallel-size 4  # Not 3\n```\n\nEnable speculative decoding for faster generation:\n```bash\nvllm serve MODEL --speculative-model DRAFT_MODEL\n```\n\n## Advanced topics\n\n**Server deployment patterns**: See [references\u002Fserver-deployment.md](references\u002Fserver-deployment.md) for Docker, Kubernetes, and load balancing configurations.\n\n**Performance optimization**: See [references\u002Foptimization.md](references\u002Foptimization.md) for PagedAttention tuning, continuous batching details, and benchmark results.\n\n**Quantization guide**: See [references\u002Fquantization.md](references\u002Fquantization.md) for AWQ\u002FGPTQ\u002FFP8 setup, model preparation, and accuracy comparisons.\n\n**Troubleshooting**: See [references\u002Ftroubleshooting.md](references\u002Ftroubleshooting.md) for detailed error messages, debugging steps, and performance diagnostics.\n\n## Hardware requirements\n\n- **Small models (7B-13B)**: 1x A10 (24GB) or A100 (40GB)\n- **Medium models (30B-40B)**: 2x A100 (40GB) with tensor parallelism\n- **Large models (70B+)**: 4x A100 (40GB) or 2x A100 (80GB), use AWQ\u002FGPTQ\n\nSupported platforms: NVIDIA (primary), AMD ROCm, Intel GPUs, TPUs\n\n## Resources\n\n- Official docs: https:\u002F\u002Fdocs.vllm.ai\n- GitHub: https:\u002F\u002Fgithub.com\u002Fvllm-project\u002Fvllm\n- Paper: \"Efficient Memory Management for Large Language Model Serving with PagedAttention\" (SOSP 2023)\n- Community: https:\u002F\u002Fdiscuss.vllm.ai\n\n## Agent operating procedure\n\n1. **Check the environment.** Confirm hardware, framework and server versions, model format, and expected load.\n2. **Pin down the inputs.** Confirm formats, identifiers and parameters from the data or the user. Ask rather than guess any value that changes the result.\n3. **Run a small version first.** Serve or log a single request or run end to end before scaling.\n4. **Execute the full task** using the instructions and references above.\n5. **Validate the result.** Measure latency, throughput and output correctness against a reference; check resource usage and costs.\n6. **Report.** State what was run (versions, commands, parameters), what was checked, and what is still uncertain.\n\n| If this happens | Do this |\n|---|---|\n| The server fails to start or OOMs | Check model size versus memory, quantization and parallelism settings. |\n| A function, flag or endpoint in these instructions is missing in the installed version | Check the installed version's own documentation (`help()`, `--help`, official docs), adapt, and tell the user. Never invent an API. |\n| A required input, identifier or parameter is ambiguous | Ask the user, or state the assumption explicitly before running. |\n\n**Integrity rules**\n\n- Never fabricate results, parameters, identifiers, citations or statistics. If something cannot be run or verified, say so plainly.\n- Do not expose services or credentials publicly; confirm cloud costs before provisioning.\n- Treat version-specific details here as possibly outdated: confirm them against the official documentation for the installed version.\n- Ask before actions that cost money, consume shared GPUs or cloud quota, touch personal or patient data, or cannot be undone.\n\n## Related skills\n\n- `tensorrt-llm`: Optimizes LLM inference with NVIDIA TensorRT for maximum throughput and lowest latency.\n- `awq-quantization`: Activation-aware weight quantization for 4-bit LLM compression with 3x speedup and minimal accuracy loss.\n- `gptq`: Quantizes LLMs to 4-bit (also 3-bit) with GPTQ using group-wise quantization (group size 128 by default), via AutoGPTQ and transformers.\n","# vLLM - High-Performance LLM Serving\n\n## Quick start\n\nvLLM achieves 24x higher throughput than standard transformers through PagedAttention (block-based KV cache) and continuous batching (mixing prefill\u002Fdecode requests).\n\n**Installation**:\n```bash\npip install vllm\n```\n\n**Basic offline inference**:\n```python\nfrom vllm import LLM, SamplingParams\n\nllm = LLM(model=\"meta-llama\u002FLlama-3-8B-Instruct\")\nsampling = SamplingParams(temperature=0.7, max_tokens=256)\n\noutputs = llm.generate([\"Explain quantum computing\"], sampling)\nprint(outputs[0].outputs[0].text)\n```\n\n**OpenAI-compatible server**:\n```bash\nvllm serve meta-llama\u002FLlama-3-8B-Instruct\n\n# Query with OpenAI SDK\npython -c \"\nfrom openai import OpenAI\nclient = OpenAI(base_url='http:\u002F\u002Flocalhost:8000\u002Fv1', api_key='EMPTY')\nprint(client.chat.completions.create(\n    model='meta-llama\u002FLlama-3-8B-Instruct',\n    messages=[{'role': 'user', 'content': 'Hello!'}]\n).choices[0].message.content)\n\"\n```\n\n## Common workflows\n\n### Workflow 1: Production API deployment\n\nCopy this checklist and track progress:\n\n```\nDeployment Progress:\n- [ ] Step 1: Configure server settings\n- [ ] Step 2: Test with limited traffic\n- [ ] Step 3: Enable monitoring\n- [ ] Step 4: Deploy to production\n- [ ] Step 5: Verify performance metrics\n```\n\n**Step 1: Configure server settings**\n\nChoose configuration based on your model size:\n\n```bash\n# For 7B-13B models on single GPU\nvllm serve meta-llama\u002FLlama-3-8B-Instruct \\\n  --gpu-memory-utilization 0.9 \\\n  --max-model-len 8192 \\\n  --port 8000\n\n# For 30B-70B models with tensor parallelism\nvllm serve meta-llama\u002FLlama-2-70b-hf \\\n  --tensor-parallel-size 4 \\\n  --gpu-memory-utilization 0.9 \\\n  --quantization awq \\\n  --port 8000\n\n# For production with caching and metrics\nvllm serve meta-llama\u002FLlama-3-8B-Instruct \\\n  --gpu-memory-utilization 0.9 \\\n  --enable-prefix-caching \\\n  --enable-metrics \\\n  --metrics-port 9090 \\\n  --port 8000 \\\n  --host 0.0.0.0\n```\n\n**Step 2: Test with limited traffic**\n\nRun load test before production:\n\n```bash\n# Install load testing tool\npip install locust\n\n# Create test_load.py with sample requests\n# Run: locust -f test_load.py --host http:\u002F\u002Flocalhost:8000\n```\n\nVerify TTFT (time to first token) \u003C 500ms and throughput > 100 req\u002Fsec.\n\n**Step 3: Enable monitoring**\n\nvLLM exposes Prometheus metrics on port 9090:\n\n```bash\ncurl http:\u002F\u002Flocalhost:9090\u002Fmetrics | grep vllm\n```\n\nKey metrics to monitor:\n- `vllm:time_to_first_token_seconds` - Latency\n- `vllm:num_requests_running` - Active requests\n- `vllm:gpu_cache_usage_perc` - KV cache utilization\n\n**Step 4: Deploy to production**\n\nUse Docker for consistent deployment:\n\n```bash\n# Run vLLM in Docker\ndocker run --gpus all -p 8000:8000 \\\n  vllm\u002Fvllm-openai:latest \\\n  --model meta-llama\u002FLlama-3-8B-Instruct \\\n  --gpu-memory-utilization 0.9 \\\n  --enable-prefix-caching\n```\n\n**Step 5: Verify performance metrics**\n\nCheck that deployment meets targets:\n- TTFT \u003C 500ms (for short prompts)\n- Throughput > target req\u002Fsec\n- GPU utilization > 80%\n- No OOM errors in logs\n\n### Workflow 2: Offline batch inference\n\nFor processing large datasets without server overhead.\n\nCopy this checklist:\n\n```\nBatch Processing:\n- [ ] Step 1: Prepare input data\n- [ ] Step 2: Configure LLM engine\n- [ ] Step 3: Run batch inference\n- [ ] Step 4: Process results\n```\n\n**Step 1: Prepare input data**\n\n```python\n# Load prompts from file\nprompts = []\nwith open(\"prompts.txt\") as f:\n    prompts = [line.strip() for line in f]\n\nprint(f\"Loaded {len(prompts)} prompts\")\n```\n\n**Step 2: Configure LLM engine**\n\n```python\nfrom vllm import LLM, SamplingParams\n\nllm = LLM(\n    model=\"meta-llama\u002FLlama-3-8B-Instruct\",\n    tensor_parallel_size=2,  # Use 2 GPUs\n    gpu_memory_utilization=0.9,\n    max_model_len=4096\n)\n\nsampling = SamplingParams(\n    temperature=0.7,\n    top_p=0.95,\n    max_tokens=512,\n    stop=[\"\u003C\u002Fs>\", \"\\n\\n\"]\n)\n```\n\n**Step 3: Run batch inference**\n\nvLLM automatically batches requests for efficiency:\n\n```python\n# Process all prompts in one call\noutputs = llm.generate(prompts, sampling)\n\n# vLLM handles batching internally\n# No need to manually chunk prompts\n```\n\n**Step 4: Process results**\n\n```python\n# Extract generated text\nresults = []\nfor output in outputs:\n    prompt = output.prompt\n    generated = output.outputs[0].text\n    results.append({\n        \"prompt\": prompt,\n        \"generated\": generated,\n        \"tokens\": len(output.outputs[0].token_ids)\n    })\n\n# Save to file\nimport json\nwith open(\"results.jsonl\", \"w\") as f:\n    for result in results:\n        f.write(json.dumps(result) + \"\\n\")\n\nprint(f\"Processed {len(results)} prompts\")\n```\n\n### Workflow 3: Quantized model serving\n\nFit large models in limited GPU memory.\n\n```\nQuantization Setup:\n- [ ] Step 1: Choose quantization method\n- [ ] Step 2: Find or create quantized model\n- [ ] Step 3: Launch with quantization flag\n- [ ] Step 4: Verify accuracy\n```\n\n**Step 1: Choose quantization method**\n\n- **AWQ**: Best for 70B models, minimal accuracy loss\n- **GPTQ**: Wide model support, good compression\n- **FP8**: Fastest on H100 GPUs\n\n**Step 2: Find or create quantized model**\n\nUse pre-quantized models from HuggingFace:\n\n```bash\n# Search for AWQ models\n# Example: TheBloke\u002FLlama-2-70B-AWQ\n```\n\n**Step 3: Launch with quantization flag**\n\n```bash\n# Using pre-quantized model\nvllm serve TheBloke\u002FLlama-2-70B-AWQ \\\n  --quantization awq \\\n  --tensor-parallel-size 1 \\\n  --gpu-memory-utilization 0.95\n\n# Results: 70B model in ~40GB VRAM\n```\n\n**Step 4: Verify accuracy**\n\nTest outputs match expected quality:\n\n```python\n# Compare quantized vs non-quantized responses\n# Verify task-specific performance unchanged\n```\n\n## When to use vs alternatives\n\n**Use vLLM when:**\n- Deploying production LLM APIs (100+ req\u002Fsec)\n- Serving OpenAI-compatible endpoints\n- Limited GPU memory but need large models\n- Multi-user applications (chatbots, assistants)\n- Need low latency with high throughput\n\n**Use alternatives instead:**\n- **llama.cpp**: CPU\u002Fedge inference, single-user\n- **HuggingFace transformers**: Research, prototyping, one-off generation\n- **TensorRT-LLM**: NVIDIA-only, need absolute maximum performance\n- **Text-Generation-Inference**: Already in HuggingFace ecosystem\n\n## Common issues\n\n**Issue: Out of memory during model loading**\n\nReduce memory usage:\n```bash\nvllm serve MODEL \\\n  --gpu-memory-utilization 0.7 \\\n  --max-model-len 4096\n```\n\nOr use quantization:\n```bash\nvllm serve MODEL --quantization awq\n```\n\n**Issue: Slow first token (TTFT > 1 second)**\n\nEnable prefix caching for repeated prompts:\n```bash\nvllm serve MODEL --enable-prefix-caching\n```\n\nFor long prompts, enable chunked prefill:\n```bash\nvllm serve MODEL --enable-chunked-prefill\n```\n\n**Issue: Model not found error**\n\nUse `--trust-remote-code` for custom models:\n```bash\nvllm serve MODEL --trust-remote-code\n```\n\n**Issue: Low throughput (\u003C50 req\u002Fsec)**\n\nIncrease concurrent sequences:\n```bash\nvllm serve MODEL --max-num-seqs 512\n```\n\nCheck GPU utilization with `nvidia-smi` - should be >80%.\n\n**Issue: Inference slower than expected**\n\nVerify tensor parallelism uses power of 2 GPUs:\n```bash\nvllm serve MODEL --tensor-parallel-size 4  # Not 3\n```\n\nEnable speculative decoding for faster generation:\n```bash\nvllm serve MODEL --speculative-model DRAFT_MODEL\n```\n\n## Advanced topics\n\n**Server deployment patterns**: See [references\u002Fserver-deployment.md](references\u002Fserver-deployment.md) for Docker, Kubernetes, and load balancing configurations.\n\n**Performance optimization**: See [references\u002Foptimization.md](references\u002Foptimization.md) for PagedAttention tuning, continuous batching details, and benchmark results.\n\n**Quantization guide**: See [references\u002Fquantization.md](references\u002Fquantization.md) for AWQ\u002FGPTQ\u002FFP8 setup, model preparation, and accuracy comparisons.\n\n**Troubleshooting**: See [references\u002Ftroubleshooting.md](references\u002Ftroubleshooting.md) for detailed error messages, debugging steps, and performance diagnostics.\n\n## Hardware requirements\n\n- **Small models (7B-13B)**: 1x A10 (24GB) or A100 (40GB)\n- **Medium models (30B-40B)**: 2x A100 (40GB) with tensor parallelism\n- **Large models (70B+)**: 4x A100 (40GB) or 2x A100 (80GB), use AWQ\u002FGPTQ\n\nSupported platforms: NVIDIA (primary), AMD ROCm, Intel GPUs, TPUs\n\n## Resources\n\n- Official docs: https:\u002F\u002Fdocs.vllm.ai\n- GitHub: https:\u002F\u002Fgithub.com\u002Fvllm-project\u002Fvllm\n- Paper: \"Efficient Memory Management for Large Language Model Serving with PagedAttention\" (SOSP 2023)\n- Community: https:\u002F\u002Fdiscuss.vllm.ai\n\n## Agent operating procedure\n\n1. **Check the environment.** Confirm hardware, framework and server versions, model format, and expected load.\n2. **Pin down the inputs.** Confirm formats, identifiers and parameters from the data or the user. Ask rather than guess any value that changes the result.\n3. **Run a small version first.** Serve or log a single request or run end to end before scaling.\n4. **Execute the full task** using the instructions and references above.\n5. **Validate the result.** Measure latency, throughput and output correctness against a reference; check resource usage and costs.\n6. **Report.** State what was run (versions, commands, parameters), what was checked, and what is still uncertain.\n\n| If this happens | Do this |\n|---|---|\n| The server fails to start or OOMs | Check model size versus memory, quantization and parallelism settings. |\n| A function, flag or endpoint in these instructions is missing in the installed version | Check the installed version's own documentation (`help()`, `--help`, official docs), adapt, and tell the user. Never invent an API. |\n| A required input, identifier or parameter is ambiguous | Ask the user, or state the assumption explicitly before running. |\n\n**Integrity rules**\n\n- Never fabricate results, parameters, identifiers, citations or statistics. If something cannot be run or verified, say so plainly.\n- Do not expose services or credentials publicly; confirm cloud costs before provisioning.\n- Treat version-specific details here as possibly outdated: confirm them against the official documentation for the installed version.\n- Ask before actions that cost money, consume shared GPUs or cloud quota, touch personal or patient data, or cannot be undone.\n\n## Related skills\n\n- `tensorrt-llm`: Optimizes LLM inference with NVIDIA TensorRT for maximum throughput and lowest latency.\n- `awq-quantization`: Activation-aware weight quantization for 4-bit LLM compression with 3x speedup and minimal accuracy loss.\n- `gptq`: Quantizes LLMs to 4-bit (also 3-bit) with GPTQ using group-wise quantization (group size 128 by default), via AutoGPTQ and transformers.\n",{"name":5,"license":15,"metadata":50,"description":7},{"tags":51,"version":52,"category":53,"maintainer":54,"dependencies":55},"vLLM, Inference Serving, PagedAttention, Continuous Batching, High Throughput, Production, OpenAI API, Quantization, Tensor Parallelism","1.0.0","ml-inference-and-ops","Kalaris Labs","vllm, torch, transformers","main","2b63f856ce19c2a797929cacabd9e0fc936812fa","plugins\u002FKalarisLabs\u002Fresearch-agent-skills\u002Fskills\u002Fserving-llms-vllm",483,"https:\u002F\u002Fgithub.com\u002Fhashgraph-online\u002Fawesome-codex-plugins","https:\u002F\u002Fhol.org\u002Fplugins\u002Fbest-codex-plugins","https:\u002F\u002Fgithub.com\u002Fhashgraph-online","https:\u002F\u002Fhol.org","Open-source standards and developer tools for interoperable AI agents: discovery, identity, communication, payments, privacy and security.","2026-11-13 02:16:32",{"source":8,"skillName":5,"commands":67,"prompt":77,"zipAvailable":78},[68,71,74],{"tool":69,"command":70},"npx","npx skills add hashgraph-online\u002Fawesome-codex-plugins --skill serving-llms-vllm",{"tool":72,"command":73},"bunx","bunx skills add hashgraph-online\u002Fawesome-codex-plugins --skill serving-llms-vllm",{"tool":75,"command":76},"pnpm","pnpm dlx skills add hashgraph-online\u002Fawesome-codex-plugins --skill serving-llms-vllm","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002Fhashgraph-online\u002Fawesome-codex-plugins\u002Fblob\u002Fmain\u002Fplugins\u002FKalarisLabs\u002Fresearch-agent-skills\u002Fskills\u002Fserving-llms-vllm\u002FSKILL.md",true,[80,86,90,94],{"path":81,"resourceType":82,"resourceTypeDesc":83,"sizeBytes":84,"sourceHtmlUrl":85},"references\u002Foptimization.md","REFERENCE","参考文档",5769,"https:\u002F\u002Fgithub.com\u002Fhashgraph-online\u002Fawesome-codex-plugins\u002Fblob\u002Fmain\u002Fplugins\u002FKalarisLabs\u002Fresearch-agent-skills\u002Fskills\u002Fserving-llms-vllm\u002Freferences\u002Foptimization.md",{"path":87,"resourceType":82,"resourceTypeDesc":83,"sizeBytes":88,"sourceHtmlUrl":89},"references\u002Fquantization.md",6746,"https:\u002F\u002Fgithub.com\u002Fhashgraph-online\u002Fawesome-codex-plugins\u002Fblob\u002Fmain\u002Fplugins\u002FKalarisLabs\u002Fresearch-agent-skills\u002Fskills\u002Fserving-llms-vllm\u002Freferences\u002Fquantization.md",{"path":91,"resourceType":82,"resourceTypeDesc":83,"sizeBytes":92,"sourceHtmlUrl":93},"references\u002Fserver-deployment.md",5252,"https:\u002F\u002Fgithub.com\u002Fhashgraph-online\u002Fawesome-codex-plugins\u002Fblob\u002Fmain\u002Fplugins\u002FKalarisLabs\u002Fresearch-agent-skills\u002Fskills\u002Fserving-llms-vllm\u002Freferences\u002Fserver-deployment.md",{"path":95,"resourceType":82,"resourceTypeDesc":83,"sizeBytes":96,"sourceHtmlUrl":97},"references\u002Ftroubleshooting.md",9037,"https:\u002F\u002Fgithub.com\u002Fhashgraph-online\u002Fawesome-codex-plugins\u002Fblob\u002Fmain\u002Fplugins\u002FKalarisLabs\u002Fresearch-agent-skills\u002Fskills\u002Fserving-llms-vllm\u002Freferences\u002Ftroubleshooting.md",[99,108,115,122,147,172],{"skillNo":100,"name":101,"nameZh":6,"description":102,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":103,"licenseFileUrl":6,"sourceHtmlUrl":104,"categories":105,"keywords":106,"publishedAt":107},"01M746ZWACVKMZ1BJXJZHYJSSX","migrate-to-e2e","Moves an existing Playwright, Cypress, Selenium, Appium, Detox, Maestro, or TesterArmy suite to e2e one slice at a time, with the old suite running beside the new one until each ported test is proven. Keeps a MIGRATION.md progress file so any later session resumes where the last stopped. Optionally plants a defect to show both suites catch it. Use for \u002Fmigrate-to-e2e, \"migrate our tests to e2e\", \"port these Maestro flows\", \"move our TesterArmy tests\", or when a project has an e2e config and an older suite still running.","Apache-2.0","https:\u002F\u002Fgithub.com\u002Fhashgraph-online\u002Fawesome-codex-plugins\u002Fblob\u002Fmain\u002Fplugins\u002Ftester-army\u002Fe2e\u002Fskills\u002Fmigrate-to-e2e\u002FSKILL.md",[],[],"2026-11-11 15:11:14",{"skillNo":109,"name":110,"nameZh":6,"description":111,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":103,"licenseFileUrl":6,"sourceHtmlUrl":112,"categories":113,"keywords":114,"publishedAt":107},"01M746ZPTAX1Q0FBKAB1CYND7E","e2e","Writes, runs, and debugs end-to-end tests for web and mobile apps with e2e, the agentic test runner, which mixes agent.act goals with exact locator assertions on a Playwright browser or an iOS or Android device. Also drives a project's running app live over the e2e MCP server to look at a screen, check a UI change, take a screenshot or record a video, and try a locator; and runs bug bashes, parallel e2e explore runs with each bug proven by a repro test. Use when a project depends on e2e or has an e2e.config.ts, when asked for end-to-end, browser, mobile, or agentic UI tests, to check a change in the running app, to QA or bug bash an app or branch, or when an e2e run fails.","https:\u002F\u002Fgithub.com\u002Fhashgraph-online\u002Fawesome-codex-plugins\u002Fblob\u002Fmain\u002Fplugins\u002Ftester-army\u002Fe2e\u002Fskills\u002Fe2e\u002FSKILL.md",[],[],{"skillNo":116,"name":117,"nameZh":6,"description":118,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":103,"licenseFileUrl":6,"sourceHtmlUrl":119,"categories":120,"keywords":121,"publishedAt":107},"01M746ZJN784HVHMDP9MQGCSGG","create-verification-skill","Generate a project-local verify-\u003Capp> skill on top of e2e, so any coding agent can launch the app, check it is worth driving, exercise a feature the way a user does, keep evidence, and bug bash it. Writes the skill, a feature map with one explore charter per feature, and proves it once. Use for \u002Fcreate-verification-skill, \"make a verification skill for this repo\", or when a project has no scripted way to prove UI behavior.","https:\u002F\u002Fgithub.com\u002Fhashgraph-online\u002Fawesome-codex-plugins\u002Fblob\u002Fmain\u002Fplugins\u002Ftester-army\u002Fe2e\u002Fskills\u002Fcreate-verification-skill\u002FSKILL.md",[],[],{"skillNo":123,"name":124,"nameZh":6,"description":125,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":103,"licenseFileUrl":6,"sourceHtmlUrl":126,"categories":127,"keywords":128,"publishedAt":45},"01M716A6ETNB9B1SQ3GXZC75VR","deja-search","deja-vu memory — search the user's past AI coding sessions with the deja CLI. Holds what they already settled and never wrote into the repo (naming rules, banned or already-tried tools, required checks before merge, preferred values, who owns what). Use before your first edit in a task, before you call a change done or ready to merge, before you recommend a tool, library, design or reviewer, before re-debugging an error that may already be solved, when they say 'didn't we fix this before', 'what did we decide about X' or 'I already have that', and before telling them something on this machine does not exist.","https:\u002F\u002Fgithub.com\u002Fhashgraph-online\u002Fawesome-codex-plugins\u002Fblob\u002Fmain\u002Fplugins\u002Fvshulcz\u002Fdeja-vu\u002Fskills\u002Fdeja-search\u002FSKILL.md",[],[129,130,131,132,133,44,134,135,136,137,138,29,139,140,141,142,143,144,145,146],"search","deja search","never","error","ready","re","decide","fix","design","debugging","you","done","reviewer","io","change","cli","merge","past",{"skillNo":148,"name":149,"nameZh":6,"description":150,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":103,"licenseFileUrl":6,"sourceHtmlUrl":151,"categories":152,"keywords":153,"publishedAt":45},"01M716A3J2W5SENSE9VGMXB375","deja-reflect","End-of-session reflection and memory curation with deja-vu. Review what was decided, what commands broke and what repaired them, and record durable project rules or decisions with deja remember so the next session starts with them.","https:\u002F\u002Fgithub.com\u002Fhashgraph-online\u002Fawesome-codex-plugins\u002Fblob\u002Fmain\u002Fplugins\u002Fvshulcz\u002Fdeja-vu\u002Fskills\u002Fdeja-reflect\u002FSKILL.md",[],[154,155,156,142,157,158,159,160,161,162,163,164,165,166,44,167,168,169,170,171],"deja reflect","reflect","review","with","record","github","next","https","node","Rules","decisions","brew","remember","project rules","project","session","end","reflection",{"skillNo":173,"name":174,"nameZh":6,"description":175,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":103,"licenseFileUrl":6,"sourceHtmlUrl":176,"categories":177,"keywords":185,"publishedAt":45},"01M715T3RR4BAYSR57HPPTJ0PF","revid","Make, edit, export and post AI videos with Revid (revid.ai) through the Revid MCP server. Use whenever the user wants a short-form or long-form video made (from an idea, a script, a URL, an article, a PDF, a tweet, a product page, their own clips, or a song), wants to find what is going viral on TikTok, YouTube, Instagram or X before making something, wants to edit or re-export an existing Revid project, generate a standalone AI image or clip, make a series episode, or publish and schedule a finished video to connected social accounts. Covers setup (hosted MCP with OAuth, or an API key), the 36 tools, credits, polling and the rules for publishing. Routes to the more specific revid-* skills when one fits.","https:\u002F\u002Fgithub.com\u002Fhashgraph-online\u002Fawesome-codex-plugins\u002Fblob\u002Fmain\u002Fplugins\u002Ftmakersas\u002Frevid-skills\u002Fskills\u002Frevid\u002FSKILL.md",[178,181,184],{"categoryNo":179,"categoryName":180,"taxonomyType":21,"taxonomyTypeDesc":22},"01JBSEEDCAT000000000000016","图像与多媒体",{"categoryNo":182,"categoryName":183,"taxonomyType":21,"taxonomyTypeDesc":22},"01JBSEEDCAT000000000000023","Skill与智能体开发",{"categoryNo":19,"categoryName":20,"taxonomyType":21,"taxonomyTypeDesc":22},[174,186,187,188,189,190,191,168,192,134,193,194,195,196,29,197,198,199,37,200],"URL","Video","short-form","standalone","PDF","viral","tiktok","short","social","MCP","song","image","mcp server","schedule","publish"]