Skip to main content

Qwen3.6 35B A3B vLLM recipes

helm recipes

The recipes are written for Helm (mixa3607/charts/vllm), but you can easily rewrite them for Docker.

A. Qwen/Qwen3.6-35B-A3B​

yaml setup
image:
  registry: docker.io
  repository: mixa3607/vllm-gfx906
  tag: 0.19.1-rocm-7.2.1-aiinfos

extraEnvVars:
  - name: FLASH_ATTENTION_TRITON_AMD_ENABLE
    value: "TRUE"
  - name: VLLM_MEMORY_PROFILER_ESTIMATE_CUDAGRAPHS
    value: "1"
  - name: RECIPE
    value: "Setup A"

app-configuration:
  vllm.yaml: |-
    model: Qwen/Qwen3.6-35B-A3B
    served-model-name:
      - Qwen3.6-35B-A3B
      - Qwen/Qwen3.6-35B-A3B
    async-scheduling: true
    trust-remote-code: true
    enable-auto-tool-choice: true
    reasoning-parser: qwen3
    tool-call-parser: qwen3_coder
    max-model-len: 200K
    tensor-parallel-size: 4
    data-parallel-size: 1
    dtype: float16
    gpu-memory-utilization: 0.95
    limit-mm-per-prompt.image: 20
    limit-mm-per-prompt.video: 4
    max-num-seqs: 16
    enable-expert-parallel: true
    enable-prefix-caching: true

B. cyankiwi/Qwen3.6-35B-A3B-AWQ-4bit​

yaml setup
image:
  registry: docker.io
  repository: mixa3607/vllm-gfx906
  tag: 0.19.1-rocm-7.2.1-aiinfos

extraEnvVars:
  - name: FLASH_ATTENTION_TRITON_AMD_ENABLE
    value: "TRUE"
  - name: VLLM_MEMORY_PROFILER_ESTIMATE_CUDAGRAPHS
    value: "1"
  - name: RECIPE
    value: "Setup A"

app-configuration:
  vllm.yaml: |-
    model: cyankiwi/Qwen3.6-35B-A3B-AWQ-4bit
    served-model-name:
      - Qwen3.6-35B-A3B
      - cyankiwi/Qwen3.6-35B-A3B-AWQ-4bit
    async-scheduling: true
    trust-remote-code: true
    enable-auto-tool-choice: true
    reasoning-parser: qwen3
    tool-call-parser: qwen3_coder
    max-model-len: 200K
    tensor-parallel-size: 4
    data-parallel-size: 1
    dtype: float16
    gpu-memory-utilization: 0.95
    limit-mm-per-prompt.image: 20
    limit-mm-per-prompt.video: 4
    max-num-seqs: 16
    enable-expert-parallel: true
    enable-prefix-caching: true