Qwen3.6-35B-A3B AutoRound runtime check
Observed July 2026

Server configuration
  runtime: patched vLLM on XPU
  dtype: torch.bfloat16
  max_seq_len: 131072
  tensor_parallel_size: 1
  quantization: inc
  kv_cache_dtype: auto
  device_config: xpu
  compilation mode: VLLM_COMPILE
  graph mode: FULL_DECODE_ONLY
  graph capture sizes: [1, 2]
  max_num_batched_tokens: 8192
  max_num_seqs: 1
  gpu_memory_utilization: 0.9

Observed startup
  attention backend: XPU Flash Attention
  model allocation: 19.60 GiB
  available KV cache: 5.69 GiB
  KV cache capacity: 289626 tokens

Measured workload
  context: 131072
  prior tokens: 115404
  prompt tokens: 2048
  generated tokens: 512
  concurrency: 1
  measured runs: 1
  prompt processing: 3377.133868 tokens/second
  generation: 47.030922 tokens/second
  first response: 34.905 seconds
  timed benchmark phase: 50.784 seconds
  coherence gate: passed

The device check identifies the physical B70. This record shows that vLLM selected XPU, initialized the XPU attention backend, allocated the model and KV cache there, and completed the measured request.