benchmarks / qwen3-5-35b-a3b--dgx-spark-x1--vllm--gptq-int4--p2048-o256-c1--2026-07-06
Qwen3.5 35B-A3B on NVIDIA DGX Spark
compare38.9tok/s
341msttft
6.0kprefill
vLLM 0.24.0GPTQ-Int4Jul 6, 2026
Run configuration
nodes
1
batch size
—
concurrency
1
context
10240
prompt / output
2039 / 256
weights Qwen/Qwen3.5-35B-A3B-GPTQ-Int4 ↗
Metrics
decode tok/s
38.9
prefill tok/s
6.0k
ttft
341ms
total tok/s
37.0
inter-token
26ms
e2e latency
6.90s
req/s
0.144
peak memory
69.4 GB
avg power
30.5 W
energy/req
211.1 J
Reproduce
PATH="$HOME/venvs/vllm/bin:$PATH" VLLM_USE_DEEP_GEMM=0 TORCHINDUCTOR_COMPILE_THREADS=2 MAX_JOBS=4 vllm serve ~/models/hf/Qwen3.5-35B-A3B-GPTQ-Int4 --served-model-name Qwen/Qwen3.5-35B-A3B-GPTQ-Int4 --max-model-len 10240 --gpu-memory-utilization 0.5 --max-num-seqs 8 --limit-mm-per-prompt '{"image": 0, "video": 0}' --port 8000
python3 bench/harness.py bench/configs/qwen3-5-35b-a3b-gptq-int4--vllm.json --rows-out data/benchmarks/qwen3-5-35b-a3b.jsonOfficial Qwen GPTQ-Int4 checkpoint, vLLM on GB10. Same-precision counterpart to the llama.cpp Q4_K_M rows (engine comparison).