benchmarks / qwen3-6-35b-a3b--dgx-spark-x1--vllm--nvfp4--dflash-greedy-baseline--p2048-o256-c1--2026-07-06
Qwen3.6 35B-A3B on NVIDIA DGX Spark
compare76.8tok/s
320msttft
6.4kprefill
vLLM 0.24.0NVFP4spec-decodingJul 6, 2026
Run configuration
nodes
1
batch size
—
concurrency
1
context
10240
prompt / output
2038 / 256
weights nvidia/Qwen3.6-35B-A3B-NVFP4 ↗
Metrics
decode tok/s
76.8
prefill tok/s
6.4k
ttft
320ms
total tok/s
70.3
inter-token
13ms
e2e latency
3.64s
req/s
0.275
peak memory
70.2 GB
avg power
29.9 W
energy/req
108.9 J
temperature
0
Reproduce
PATH="$HOME/venvs/vllm/bin:$PATH" VLLM_USE_DEEP_GEMM=0 VLLM_ATTENTION_BACKEND=FLASH_ATTN TORCHINDUCTOR_COMPILE_THREADS=2 MAX_JOBS=4 vllm serve ~/models/hf/Qwen3.6-35B-A3B-NVFP4 --served-model-name Qwen3.6-35B-A3B-NVFP4 --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/qwen3-6-35b-a3b--dflash-baseline--vllm.json --rows-out data/benchmarks/qwen3-6-35b-a3b.jsonspec-decoding profile: greedy no-draft NVFP4 baseline for the DFlash comparison (official target).