vserve 0.6.4b1

CLI for managing LLM inference on GPU workstations

vserve

A CLI for managing LLM inference on GPU workstations.

Download models. Auto-tune limits. Serve with one command. Multiple backends.

---

Release: 0.6.3

vserve is now stable — first non-beta release of the 0.6 line.

Highlights in 0.6.3:

• vLLM 0.22 support: range >=0.20,<0.23 (pinned stable runtime stays 0.21.0 — see caveats)
• migrates the deprecated VLLM_USE_FLASHINFER_MOE_* env vars to vLLM 0.22's hardware-aware backend auto-selection, with a --moe-backend escape hatch
• emits default-chat-template-kwargs on 0.22 (the flag was renamed upstream); thinking toggles keep working across 0.20–0.22
• spec-decode with fp8 KV cache keeps CUDA graphs on 0.22 (DFlash fp8 fix verified upstream); TurboQuant stays conservatively un-graphed
• fixes Qwen 3.5 / 3.6 tool calling — these emit the XML tool format and now route to the qwen3_coder parser instead of hermes (which silently dropped tool calls into message content)
• the 0.6.1–0.6.3b3 line: research bundle, vserve bench, de-spaghetti refactor, arch-registry canonicalization — GPU-verified on 0.21 and re-verified on 0.22

Known caveats:

• Gemma-4 NVFP4 on vLLM 0.22: startup OOMs from multimodal video-encoder profiling (vLLM #43169 batched encoder; the V1 profiler allocates max-size dummy videos). Serve it on 0.21, or pass --limit-mm-per-prompt '{"image":1,"video":0}' plus --mm-processor-kwargs '{"max_soft_tokens":560}'. Automatic multimodal caps land in 0.6.4. This is why the pinned stable runtime stays 0.21.0.
• vLLM 0.20.0–0.22.0 compute RMSNorm weights in FP32 upstream (vllm#42325, fixed after the 0.22.0 cut) — accuracy-sensitive users should adopt the next vLLM patch
• non-interactive startup remains intentionally strict: if the backend never reaches a healthy API state within the timeout window, run exits nonzero even if the service is still warming
• multi-user coordination is best-effort operational safety, not a security boundary

---

Install

Install from PyPI:

uv tool install vserve
pip install vserve

For llama.cpp GGUF tuning support:

pip install 'vserve[llamacpp]'

---

Quick Start

vserve init                        # scan GPU, backends, CUDA, systemd — write config
vserve runtime check vllm          # verify the external vLLM runtime
vserve add                         # search HuggingFace, pick variant, download
vserve run <model>                 # auto-tune + interactive config + serve
vserve run <model> --tools         # enable tool calling (auto-detected)
vserve run <model> --backend llamacpp  # force a specific backend

Scriptable serving:

vserve run qwen fp8 --yes --context 32768 --slots 4 --kv-cache-dtype fp8 --port 8888
vserve run qwen fp8 --yes --replace              # safe non-interactive restart
vserve run qwen fp8 --save-profile fast --yes
vserve run qwen fp8 --profile fast
vserve run --profile /opt/vllm/configs/models/provider--Model.fast.yaml --yes

Runtime repair and GGUF-only setup:

vserve runtime check vllm
vserve runtime upgrade vllm --stable
vserve add TheBloke/some-model-GGUF
vserve run some model q4 --backend llamacpp --yes --gpu-layers 999

Automation:

vserve run qwen fp8 --profile fast --yes
vserve status
vserve stop

---

Backends

vserve auto-detects the right backend from the model format:

Format	Backend	Engine
safetensors, GPTQ, AWQ, FP8	vLLM	PagedAttention, continuous batching
GGUF	llama.cpp	CPU/GPU offload, quantized inference

No configuration needed — download a model and vserve run picks the right engine.

vLLM

The default for transformer models in safetensors format. Optimized for high-throughput serving with PagedAttention, KV cache management, and automatic batching.

• Auto-tunes --max-model-len, --max-num-seqs, --kv-cache-dtype based on your GPU
• Calculates PagedAttention block-rounded capacity for native, FP8, and TurboQuant KV-cache dtypes
• Recommends scheduler profiles with chunked-prefill-oriented token budgets and vLLM 0.20 optimization knobs
• Tool calling with parser auto-detection (Qwen, Llama, Mistral, DeepSeek, Gemma, GPT-OSS)
• Systemd service management via vllm.service

llama.cpp

For GGUF quantized models. Serves via llama-server with an OpenAI-compatible API.

• Auto-calculates --n-gpu-layers, --ctx-size, --parallel based on VRAM
• Reads GGUF metadata without the optional gguf package and accounts for layerwise KV heads, sliding-window attention, and recurrent state
• Partial GPU offload — serve models that don't fully fit in VRAM
• Tool calling via --jinja (no parser configuration needed)
• Systemd service management via llama-cpp.service

---

What It Does

vserve manages the full lifecycle of serving LLMs on a GPU workstation:

• Download — search HuggingFace, see available weight variants (FP8, NVFP4, BF16, GGUF) with sizes, download only one backend format at a time, and materialize each runnable variant into its own model root
• Auto-tune — calculate exactly what context lengths and concurrency your GPU can handle, based on model architecture and available VRAM
• Benchmark — opt into bounded backend microbenchmarks with vserve tune --bench
• Tool calling — auto-detects the correct parser from the model's chat template (vLLM) or uses --jinja (llama.cpp)
• Run/Stop — interactive config wizard, systemd service management, health check with timeout
• Fan control — temperature-based curve daemon with quiet hours, or hold a fixed speed
• Multi-user — best-effort session coordination warns other vserve users before they disrupt your running model
• Doctor — diagnose GPU, CUDA, backend, systemd issues with actionable fix suggestions

---

Commands

Command	Description
vserve	Dashboard — GPU, models, status
vserve init	Auto-discover backends and write config
vserve list [name]	List models with backend, tools, and limits
vserve add [model]	Search and download from HuggingFace with variant picker
vserve rm <name>	Remove a downloaded model
vserve tune [model]	Calculate context/concurrency limits
vserve tune [model] --bench	Run bounded benchmarks for tuned vLLM or llama.cpp profiles
vserve run [model]	Configure and start serving (auto-tunes if needed)
vserve run MODEL... --yes --context N --slots N	Non-interactive serving from flags
vserve run MODEL... --yes --replace	Non-interactive restart; without --replace, running backends are refused
vserve run MODEL... --profile NAME_OR_PATH	Serve a saved profile by name or explicit path
vserve run MODEL... --tools --tool-parser hermes --reasoning-parser qwen3	Start with explicit parsers
vserve run MODEL... --trust-remote-code	Opt in to vLLM remote model code execution
vserve run MODEL... --backend llamacpp --gpu-layers 999	Force llama.cpp for GGUF
vserve profile list|show|rm	Manage saved serving profiles
vserve stop	Stop the running server
vserve status [--json]	Show current serving config and probe uncertainty
vserve fan [auto|off|30-100]	GPU fan control with temp-based curve
vserve doctor [--json] [--strict]	Check system readiness; strict exits nonzero on failures
vserve cache clean [--dry-run] [--all] [--yes]	Preview or clean stale sockets and JIT caches
vserve runtime check vllm	Check vLLM version/dependency compatibility
vserve runtime upgrade vllm --stable	Reinstall vserve's pinned stable vLLM runtime
vserve version	Show current version and check for updates
vserve update [--nightly]	Update vserve, optionally allowing pre-releases

Model-taking commands support fuzzy matching — vserve run qwen fp8 finds the right model.

Profile rules: names saved with --save-profile must match [A-Za-z0-9._-]+ and cannot be ., .., or include path separators. Profile names resolve inside configured vserve profile roots. Explicit external --profile paths are accepted only by run and infer backend from YAML/JSON when possible. profile show and profile rm never read or delete arbitrary external paths, even with --force.

Automation note: run --yes is fully non-interactive. If it needs to stop or start systemd services it uses non-prompting service operations; configure passwordless service control for the vserve operator or run without --yes.

---

Tool Calling

vLLM

Auto-detects the correct vLLM parser by reading the model's chat template:

Model Family	Tool Parser	Reasoning Parser
Qwen 2.5	hermes	—
Qwen 3	hermes	qwen3
Qwen 3.5	qwen3_coder	qwen3
Llama 3.1 / 3.2 / 3.3	llama3_json	—
Llama 4	llama4_pythonic	—
Mistral / Mixtral	mistral	mistral
DeepSeek V3 / R1	deepseek_v3	deepseek_r1
Gemma 4	gemma4	gemma4
GPT-OSS	openai	openai_gptoss

Detection is template-based (not model-name regex), so it works for fine-tunes and community uploads.

Remote model code is disabled by default. Use --trust-remote-code only for repositories you trust; generated profiles include trust-remote-code only when that flag is explicitly set.

llama.cpp

Uses --jinja to read the model's chat template directly. No parser selection needed — one flag covers all model families.

---

Prerequisites

Requirement	Check	Install
NVIDIA GPU + drivers	nvidia-smi	nvidia.com/drivers
CUDA toolkit	nvcc --version	sudo apt install nvidia-cuda-toolkit
systemd	(most Linux servers)	See troubleshooting
sudo access	for systemctl, fan control

For vLLM backend:

Requirement	Check	Install
stable vLLM 0.20.x–0.22.x	vserve runtime check vllm	vserve runtime upgrade vllm --stable or docs.vllm.ai

For llama.cpp backend:

Requirement	Check	Install
llama-server	llama-server --version	github.com/ggml-org/llama.cpp

---

Configuration

Auto-discovered on first run. Override at ~/.config/vserve/config.yaml:

schema_version: 2
cuda_home: /usr/local/cuda
gpu:
  index: 0
  memory_utilization: 0.91
backends:
  vllm:
    root: /opt/vllm
    service_name: vllm
    service_user: vllm
    port: 8888
  llamacpp:
    root: /opt/llama-cpp
    service_name: llama-cpp
    service_user: llama-cpp

Legacy top-level vllm_root, service_name, llamacpp_root, and GPU memory keys still load, but newly saved config uses the backend-indexed schema above.

gpu.index is part of runtime truth, not only a tuning hint. vserve records it in active manifests and tuning fingerprints. llama.cpp launch scripts export CUDA_VISIBLE_DEVICES=<index>. vLLM writes configs/.env with the same value and doctor expects the systemd unit to load that environment file.

---

Directory Layout

/opt/vllm/                     # vLLM backend
├── venv/bin/vllm              # Python venv
├── .venv/bin/vllm             # alternate Python venv location
├── models/                    # safetensors models
├── configs/
│   ├── .env                   # service environment
│   ├── active.yaml            # active profile symlink
│   └── models/                # limits + YAML profiles
├── tmp/                       # RPC sockets / runtime temp files
├── .cache/
│   ├── flashinfer/            # FlashInfer JIT cache
│   ├── torch_extensions/      # torch extension cache
│   └── vllm/                  # vLLM/torch.compile cache
├── run/
│   └── active-manifest.json   # active backend state
└── logs/

/opt/llama-cpp/                # llama.cpp backend
├── bin/llama-server           # compiled binary
├── models/                    # GGUF models
├── configs/
│   ├── active.sh              # active launch script symlink
│   ├── active.json            # active config symlink
│   └── models/                # JSON profiles
├── run/
│   └── active-manifest.json   # active backend state
└── logs/

GGUF downloads create one runnable model root per selected quant/subdirectory, so Q4_K_M and Q8_0 variants do not share a directory. Source roots left only for materialization are ignored by model scanning.

---

Fan Control

vserve fan              # show status, interactive menu
vserve fan auto         # temp-based curve with quiet hours
vserve fan 80           # hold at 80% (persistent daemon)
vserve fan off          # stop daemon, restore NVIDIA auto

The auto curve ramps with temperature and caps fan speed during quiet hours (configurable). Emergency override at 88C ignores quiet hours.

---

Architecture

vserve uses a Backend Protocol pattern. Each inference engine implements the same interface:

Backend Protocol
├── VllmBackend        — safetensors, AWQ, FP8, GPTQ
├── LlamaCppBackend    — GGUF
└── (future: SGLang, etc.)

The registry auto-detects the right backend from the model format. Runtime checks, tuning fingerprints, profile/config generation, service lifecycle, active manifests, and status summaries live behind the backend protocol so the command layer can stay focused on user workflows.

---

Development

git clone https://github.com/Gavin-Qiao/vserve.git
cd vserve
uv sync --dev
./scripts/install-hooks.sh        # wire up .githooks/ pre-commit + pre-push
uv run pytest tests/              # full suite
uv run ruff check src/ tests/     # lint
uv run mypy src/vserve/ --ignore-missing-imports --check-untyped-defs

The install-hooks.sh step points core.hooksPath at the tracked .githooks/ directory. Once set, every git commit runs the same gates CI runs (ruff + mypy + pytest under CI=true GITHUB_ACTIONS=true and COLUMNS=80, so that anything green locally is green in CI), and every git push re-runs those plus a version-sync check and an uv build dry-run. Bypass with --no-verify only when you've already verified CI separately.

---

License

MIT