vxcontrol--pentagi/examples/configs/vllm-qwen3.5-27b-fp8.provider.yml

# Qwen3.5-27B FP8 Provider Configuration - THINKING MODE (default)
# Based on official Qwen recommendations for vLLM inference
# Architecture: Hybrid 75% DeltaNet + 25% Full Attention (48+16 layers)
# Context: 262K native, expandable to 1M with YaRN
# Vision: VLM with Vision Encoder (uses VRAM even for text-only tasks)
#
# Thinking mode is enabled by default (no extra_body needed)
# Recommended sampling parameters:
# - General tasks: temp=1.0, top_p=0.95, top_k=20, min_p=0.0, pp=1.5, rp=1.0
# - Precise coding: temp=0.6, top_p=0.95, top_k=20, min_p=0.0, pp=0.0, rp=1.0
#
# Non-thinking mode is disabled via extra_body parameter
# Recommended sampling parameters:
# - General tasks: temp=0.7, top_p=0.8, top_k=20, min_p=0.0, pp=1.5, rp=1.0
# - Reasoning tasks: temp=1.0, top_p=0.95, top_k=20, min_p=0.0, pp=1.5, rp=1.0

simple:
  model: "Qwen/Qwen3.5-27B-FP8"
  temperature: 0.7
  top_k: 20
  top_p: 0.8
  min_p: 0.0
  presence_penalty: 1.5
  repetition_penalty: 1.0
  n: 1
  max_tokens: 32768
  extra_body:
    chat_template_kwargs:
      enable_thinking: false

simple_json:
  model: "Qwen/Qwen3.5-27B-FP8"
  temperature: 0.7
  top_k: 20
  top_p: 0.8
  min_p: 0.0
  presence_penalty: 1.5
  repetition_penalty: 1.0
  n: 1
  max_tokens: 32768
  json: true
  extra_body:
    chat_template_kwargs:
      enable_thinking: false

primary_agent:
  model: "Qwen/Qwen3.5-27B-FP8"
  temperature: 1.0
  top_k: 20
  top_p: 0.95
  min_p: 0.0
  presence_penalty: 1.5
  repetition_penalty: 1.0
  n: 1
  max_tokens: 32768

assistant:
  model: "Qwen/Qwen3.5-27B-FP8"
  temperature: 1.0
  top_k: 20
  top_p: 0.95
  min_p: 0.0
  presence_penalty: 1.5
  repetition_penalty: 1.0
  n: 1
  max_tokens: 32768

generator:
  model: "Qwen/Qwen3.5-27B-FP8"
  temperature: 1.0
  top_k: 20
  top_p: 0.95
  min_p: 0.0
  presence_penalty: 1.5
  repetition_penalty: 1.0
  n: 1
  max_tokens: 32768

refiner:
  model: "Qwen/Qwen3.5-27B-FP8"
  temperature: 1.0
  top_k: 20
  top_p: 0.95
  min_p: 0.0
  presence_penalty: 1.5
  repetition_penalty: 1.0
  n: 1
  max_tokens: 32768

adviser:
  model: "Qwen/Qwen3.5-27B-FP8"
  temperature: 1.0
  top_k: 20
  top_p: 0.95
  min_p: 0.0
  presence_penalty: 1.5
  repetition_penalty: 1.0
  n: 1
  max_tokens: 32768

reflector:
  model: "Qwen/Qwen3.5-27B-FP8"
  temperature: 1.0
  top_k: 20
  top_p: 0.95
  min_p: 0.0
  presence_penalty: 1.5
  repetition_penalty: 1.0
  n: 1
  max_tokens: 32768
  extra_body:
    chat_template_kwargs:
      enable_thinking: false

searcher:
  model: "Qwen/Qwen3.5-27B-FP8"
  temperature: 0.7
  top_k: 20
  top_p: 0.8
  min_p: 0.0
  presence_penalty: 1.5
  repetition_penalty: 1.0
  n: 1
  max_tokens: 32768
  extra_body:
    chat_template_kwargs:
      enable_thinking: false

enricher:
  model: "Qwen/Qwen3.5-27B-FP8"
  temperature: 0.7
  top_k: 20
  top_p: 0.8
  min_p: 0.0
  presence_penalty: 1.5
  repetition_penalty: 1.0
  n: 1
  max_tokens: 32768
  extra_body:
    chat_template_kwargs:
      enable_thinking: false

coder:
  model: "Qwen/Qwen3.5-27B-FP8"
  temperature: 0.6
  top_k: 20
  top_p: 0.95
  min_p: 0.0
  presence_penalty: 0.0
  repetition_penalty: 1.0
  n: 1
  max_tokens: 32768

installer:
  model: "Qwen/Qwen3.5-27B-FP8"
  temperature: 0.6
  top_k: 20
  top_p: 0.95
  min_p: 0.0
  presence_penalty: 0.0
  repetition_penalty: 1.0
  n: 1
  max_tokens: 32768

pentester:
  model: "Qwen/Qwen3.5-27B-FP8"
  temperature: 0.6
  top_k: 20
  top_p: 0.95
  min_p: 0.0
  presence_penalty: 0.0
  repetition_penalty: 1.0
  n: 1
  max_tokens: 32768