clserve 0.1.1

CLI tool for serving LLM models on Alps

clserve

CLI tool for serving LLM models on Alps with sglang.

Installation

pip install -e /path/to/clserve

Or from the clserve directory:

pip install -e .

Quick Start

# Serve a model using predefined config
clserve serve deepseek-v3

# Check status of all clserve jobs
clserve status

# Get the endpoint URL by model name
clserve url deepseek-v3

# Stop the serving job by model name
clserve stop deepseek-v3

Commands

clserve serve

Start serving a model.

# Serve with predefined configuration
clserve serve deepseek-v3
clserve serve llama-405b
clserve serve qwen3-235b

# Serve with multiple workers
clserve serve deepseek-v3 --workers 2 --use-router

# Serve a custom model
clserve serve my-org/my-model --tp-size 4 --nodes-per-worker 1

# Serve a small model with 4 instances per node
clserve serve llama-8b --num-gpus-per-worker 1 --use-router

Options:

• --workers, -w: Number of workers (default: 1)
• --nodes-per-worker, -n: Nodes per worker (default: 1)
• --partition, -p: SLURM partition (default: normal)
• --environment, -e: Container environment (default: sglang_gb200)
• --tp-size: Tensor parallel size (default: 1)
• --dp-size: Data parallel size (default: 1)
• --ep-size: Expert parallel size (default: 1)
• --num-gpus-per-worker: GPUs per worker process (1, 2, or 4)
• --use-router/--no-router: Enable load balancer router
• --router-policy: Router policy (cache_aware, random, round_robin)
• --time-limit, -t: Job time limit in HH:MM:SS (default: 04:00:00)
• --job-name, -j: Custom job name

clserve status

Show status of serving jobs.

# Show all running jobs
clserve status

# Show status for a specific job
clserve status 12345

# Show status for jobs serving a model
clserve status deepseek-v3

clserve url

Get the endpoint URL for a serving job by model name. If multiple jobs are serving the same model, you'll be prompted to select one.

# Get URL by model name
clserve url deepseek-v3

# Get URL by full model path
clserve url deepseek-ai/DeepSeek-V3.1

clserve stop

Stop serving jobs by model name. If multiple jobs are serving the same model, you'll be prompted to select one.

# Stop by model name (selector if multiple)
clserve stop deepseek-v3

# Stop all jobs for a model
clserve stop deepseek-v3 --all

# Stop all running jobs
clserve stop --all

clserve models

List available predefined model configurations.

clserve models

clserve logs

Get the log file path for a job by model name. If multiple jobs are serving the same model, you'll be prompted to select one. Logs are stored in ~/.clserve/logs/<job_id>/.

clserve logs deepseek-v3
tail -f $(clserve logs deepseek-v3)/log.out

Predefined Model Configurations

The following models have optimized configurations:

Alias	Model	TP Size	Nodes/Worker	Description
deepseek-v3	deepseek-ai/DeepSeek-V3.1	16	4	DeepSeek V3.1 MoE (FP8)
deepseek-r1	deepseek-ai/DeepSeek-R1	16	4	DeepSeek R1 reasoning model
llama-405b	meta-llama/Llama-3.1-405B-Instruct	16	4	Llama 3.1 405B
llama-70b	meta-llama/Llama-3.1-70B-Instruct	4	1	Llama 3.1 70B
llama-8b	meta-llama/Llama-3.1-8B-Instruct	1	1	Llama 3.1 8B (4x per node)
qwen3-235b	Qwen/Qwen3-235B-A22B-Instruct-2507	8	2	Qwen3 235B MoE
qwen3-coder-480b	Qwen/Qwen3-Coder-480B-A35B-Instruct	16	4	Qwen3 Coder 480B MoE
qwen3-32b	Qwen/Qwen3-32B	2	1	Qwen3 32B (2x per node)
qwen3-8b	Qwen/Qwen3-8B	1	1	Qwen3 8B (4x per node)
apertus-8b	swiss-ai/Apertus-8B-Instruct-2509	1	1	Apertus 8B (4x per node)

Examples

Serve DeepSeek V3 with default config

clserve serve deepseek-v3

This will:

• Use 4 nodes with TP=16
• Start the model on the cluster
• Print the job ID and endpoint URL instructions

Serve with multiple workers and router

clserve serve deepseek-v3 --workers 2 --use-router

This doubles capacity with load balancing.

Serve a small model efficiently

clserve serve llama-8b

Predefined config runs 4 instances per node with a router for high throughput.

Full workflow example

# Start serving
clserve serve deepseek-v3
# Output: Job ID: 12345

# Wait for startup, then get URL
clserve url 12345
# Output: http://10.0.0.1:30000

# Use the API
curl http://10.0.0.1:30000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{"model": "deepseek-ai/DeepSeek-V3.1", "messages": [{"role": "user", "content": "Hello!"}]}'

# When done, stop the job
clserve stop 12345

Architecture

clserve unifies single-node and multi-node deployments into a single template:

• Single node, full GPU: --nodes-per-worker 1 --num-gpus-per-worker 4
• Multi-node distributed: --nodes-per-worker 4 --tp-size 16
• Multiple instances per node: --num-gpus-per-worker 1 --use-router

The router is automatically configured when needed for load balancing across multiple worker processes.

Job Naming

Jobs are automatically named with a clserve_ prefix followed by a random ID:

• clserve_abc123
• clserve_xyz789

This prefix allows clserve to filter and show only its own jobs in clserve status, without interference from other SLURM jobs.

You can override with --job-name (but should keep the clserve_ prefix for filtering).