006734910c
Production-tested recipe for running Qwen3.5-397B-A17B with INT4 AutoRound quantization across 4 DGX Spark nodes using tensor parallelism. Performance (4× DGX Spark, driver 580.126.09): - Single user: 37 tok/s - 4 concurrent: ~26 tok/s per user, ~103 tok/s aggregate The Marlin TP fix resolves the MIN_THREAD_N=64 constraint that breaks in_proj_ba layers at TP=4 (output_size=128/4=32 < 64). Solution: ReplicatedLinear for B/A projections, applied via diff patches. Key config: - VLLM_MARLIN_USE_ATOMIC_ADD=1 (required for Marlin correctness) - KV cache FP8, prefix caching enabled - gpu_memory_utilization 0.78 (UMA safe margin) - CUDAGraphs enabled (default, requires driver 580.x) Note: Driver 590.x has CUDAGraph capture deadlock on GB10 unified memory. Stay on driver 580.126.09.
53 lines
1.6 KiB
YAML
53 lines
1.6 KiB
YAML
# Recipe: Qwen3.5-397B-A17B-INT4-Autoround
|
||
# Qwen3.5-397B model in Intel INT4-Autoround quantization, TP=4 across 4 DGX Spark nodes
|
||
# Benchmarked at 37 tok/s single-user, 103 tok/s aggregate (4 concurrent) on 4× DGX Spark
|
||
# Requires NVIDIA driver 580.x (590.x has CUDAGraph deadlock bug on GB10)
|
||
|
||
recipe_version: "1"
|
||
name: Qwen3.5-397B-INT4-Autoround
|
||
description: Qwen3.5-397B with TP=4 across 4 DGX Spark nodes (Marlin fix applied)
|
||
|
||
# HuggingFace model to download (optional, for --download-model)
|
||
model: Intel/Qwen3.5-397B-A17B-int4-AutoRound
|
||
|
||
# Container image to use
|
||
container: vllm-node-tf5
|
||
|
||
build_args:
|
||
- --tf5
|
||
|
||
# Mods required: coder-next tool/reasoning parser + Marlin TP fix
|
||
mods:
|
||
- mods/fix-qwen3-coder-next
|
||
- mods/fix-qwen35-tp4-marlin
|
||
|
||
# Environment variables
|
||
env:
|
||
VLLM_MARLIN_USE_ATOMIC_ADD: 1
|
||
|
||
# Default settings (can be overridden via CLI, e.g. --tensor_parallel 2)
|
||
defaults:
|
||
port: 8000
|
||
host: 0.0.0.0
|
||
tensor_parallel: 4
|
||
gpu_memory_utilization: 0.78
|
||
max_model_len: 32768
|
||
max_num_batched_tokens: 8192
|
||
|
||
# The vLLM serve command template
|
||
command: |
|
||
vllm serve Intel/Qwen3.5-397B-A17B-int4-AutoRound \
|
||
--tool-call-parser qwen3_coder \
|
||
--reasoning-parser qwen3 \
|
||
--enable-auto-tool-choice \
|
||
--tensor-parallel-size {tensor_parallel} \
|
||
--distributed-executor-backend ray \
|
||
--kv-cache-dtype fp8 \
|
||
--gpu-memory-utilization {gpu_memory_utilization} \
|
||
--max-model-len {max_model_len} \
|
||
--max-num-batched-tokens {max_num_batched_tokens} \
|
||
--enable-prefix-caching \
|
||
--trust-remote-code \
|
||
--host {host} \
|
||
--port {port}
|