# benchmark details
benchmark_name: gpt-oss-120b
scenario: offline
test_mode: performance
engine_version: sync

vllm_env_config:
  VLLM_LOGGING_LEVEL: "INFO"
  HARNESS_GC_LIMIT: 100000
  VLLM_ROCM_SHUFFLE_KV_CACHE_LAYOUT: 1
  VLLM_ROCM_USE_AITER: 1
  AITER_LOG_LEVEL: "INFO"
  AMDGCN_USE_BUFFER_OPS: 0

vllm_engine_config:
  model: /model/gpt-oss-120b
  tensor_parallel_size: 1
  quantization: mxfp4
  gpu_memory_utilization: 0.96
  max_num_batched_tokens: 65536
  max_model_len: 32768
  max_num_seqs: 1792
  enforce_eager: False
  enable_chunked_prefill: False
  block_size: 16
  enable_prefix_caching: False
  disable_log_stats: True
  dtype: auto
  async_scheduling: True
  attention_backend: ROCM_AITER_FA
  kv_cache_dtype: fp8
  compilation_config:
    cudagraph_mode: FULL_AND_PIECEWISE
    compile_sizes: [2, 4, 8, 16, 32, 64, 128, 256, 512, 1024, 1536, 1792, 4096, 8192, 32768]
    cudagraph_capture_range: [[2048, 0, -8], 4, 2, 1]
    pass_config:
      fuse_rope_kvcache: true
    use_inductor_graph_partition: true

# configuration related to the sampling params
vllm_sampling_config:
  temperature: 1
  min_tokens: 1
  max_tokens: 10240
  top_p: 1
  top_k: -1
  ignore_eos: False
  detokenize: False

# configuration related to the harness tests.
harness_config:
  dataset_path: /data/gptoss-dataset/perf/perf_eval_ref.parquet
  accuracy_dataset_path: /data/gptoss-dataset/acc/acc_eval_ref.parquet
  mlperf_conf_path: /lab-mlperf-inference/mlperf_inference/mlperf.conf
  user_conf_path: /lab-mlperf-inference/code/gpt-oss-120b/user_mi355x.conf
  target_qps: -1
  total_sample_count: 6396
  accuracy_sample_count: 4395
  output_log_dir: /lab-mlperf-inference/logs
  enable_log_trace: False
  enable_warmup: True
  device_count: 8
  sorting:
    strategy: ignore
    buckets: [
      12.45692710500971,
      12.542803033483215,
      12.540764079874208,
      12.533327896123708,
      12.486551901564114,
      12.481994240555746,
      12.472878918539005,
      12.484752824850286,
    ]