22a020e4c7
bitsandbytes quantizes on-the-fly: loads full bf16 weights (~54GB RAM peak) before compressing to int4. Kills the 40GB OptiPlex on torch.compile warmup. Fix in next commit: switch serve phase to llama.cpp + GGUF Q4_K_M. Pre-quantized weights load directly — peak RAM ~16GB, no compile overhead.
44 lines
1.3 KiB
YAML
44 lines
1.3 KiB
YAML
---
|
|
# ------------------------------------------------------------------------------
|
|
# FILE: roles/llm-inference/tasks/vllm.yml
|
|
# DESCRIPTION: Phase 3 — Python venv + vLLM install.
|
|
# Idempotent: venv creation and pip install only run if the
|
|
# venv binary or vllm package is absent.
|
|
# Already completed manually on 2026-08-03 — will no-op.
|
|
# ------------------------------------------------------------------------------
|
|
|
|
- name: Create Python venv for vLLM
|
|
ansible.builtin.command:
|
|
cmd: python3 -m venv {{ llm_venv_path }}
|
|
creates: "{{ llm_venv_path }}/bin/python"
|
|
become: true
|
|
become_user: "{{ llm_venv_owner }}"
|
|
|
|
- name: Upgrade pip inside venv
|
|
ansible.builtin.pip:
|
|
name: pip
|
|
state: latest
|
|
virtualenv: "{{ llm_venv_path }}"
|
|
become: true
|
|
become_user: "{{ llm_venv_owner }}"
|
|
|
|
- name: Install vLLM and bitsandbytes
|
|
ansible.builtin.pip:
|
|
name:
|
|
- vllm
|
|
- bitsandbytes
|
|
state: present
|
|
virtualenv: "{{ llm_venv_path }}"
|
|
become: true
|
|
become_user: "{{ llm_venv_owner }}"
|
|
|
|
- name: Verify vLLM is importable
|
|
ansible.builtin.command:
|
|
cmd: "{{ llm_venv_path }}/bin/python -c 'import vllm; print(vllm.__version__)'"
|
|
register: vllm_version
|
|
changed_when: false
|
|
|
|
- name: Print vLLM version
|
|
ansible.builtin.debug:
|
|
msg: "vLLM version: {{ vllm_version.stdout }}"
|