--- dataset_info: - config_name: 1xA100_EXPERTS_2_QUANT_NAIVE features: - name: server_args struct: - name: allow_toppings_registration dtype: bool - name: api_key dtype: 'null' - name: attention_backend dtype: string - name: base_min_new_token_ratio dtype: float64 - name: chat_template dtype: 'null' - name: chunked_prefill_size dtype: int64 - name: constrained_json_whitespace_pattern dtype: 'null' - name: context_length dtype: int64 - name: controller_port dtype: int64 - name: debug dtype: bool - name: decode_log_interval dtype: int64 - name: detokenizer_ipc_name dtype: string - name: detokenizer_port dtype: int64 - name: device dtype: string - name: disable_cuda_graph dtype: bool - name: disable_cuda_graph_padding dtype: bool - name: disable_custom_all_reduce dtype: bool - name: disable_disk_cache dtype: bool - name: disable_flashinfer dtype: bool - name: disable_flashinfer_sampling dtype: bool - name: disable_mla dtype: bool - name: disable_nan_detection dtype: bool - name: disable_penalizer dtype: bool - name: disable_radix_cache dtype: bool - name: disable_regex_jump_forward dtype: bool - name: dist_init_addr dtype: 'null' - name: dp_size dtype: int64 - name: dtype dtype: string - name: enable_double_sparsity dtype: bool - name: enable_dp_attention dtype: bool - name: enable_mixed_chunk dtype: bool - name: enable_overlap_schedule dtype: bool - name: enable_p2p_check dtype: bool - name: enable_stats_logging dtype: bool - name: enable_system_controller dtype: bool - name: enable_toppings dtype: bool - name: enable_torch_compile dtype: bool - name: flashinfer_workspace_size dtype: int64 - name: host dtype: string - name: init_new_token_ratio dtype: float64 - name: init_number_of_deltas dtype: int64 - name: init_number_of_loras dtype: int64 - name: init_toppings dtype: 'null' - name: is_embedding dtype: bool - name: kv_cache_dtype dtype: string - name: load_balance_method dtype: string - name: load_format dtype: string - name: log_requests dtype: bool - name: lora_paths dtype: 'null' - name: max_lora_ranks dtype: int64 - name: max_prefill_tokens dtype: int64 - name: max_running_requests dtype: 'null' - name: max_toppings_per_batch dtype: int64 - name: max_torch_compile_bs dtype: int64 - name: max_total_tokens dtype: 'null' - name: mem_fraction_static dtype: float64 - name: model_path dtype: string - name: nccl_ports sequence: string - name: new_token_ratio_decay dtype: float64 - name: nnodes dtype: int64 - name: node_rank dtype: int64 - name: num_continue_decode_steps dtype: int64 - name: num_continuous_decode_steps dtype: int64 - name: port dtype: int64 - name: quantization dtype: 'null' - name: random_seed dtype: int64 - name: retract_decode_steps dtype: int64 - name: sampling_backend dtype: string - name: schedule_conservativeness dtype: float64 - name: schedule_policy dtype: string - name: scheduler_input_ipc_name dtype: string - name: scheduler_port dtype: int64 - name: served_model_name dtype: string - name: server_id dtype: string - name: show_time_cost dtype: bool - name: skip_tokenizer_init dtype: bool - name: stream_interval dtype: int64 - name: tokenizer_ipc_name dtype: string - name: tokenizer_mode dtype: string - name: tokenizer_path dtype: string - name: tokenizer_port dtype: int64 - name: torchao_config dtype: string - name: tp_size dtype: int64 - name: triton_attention_reduce_in_fp32 dtype: bool - name: trust_remote_code dtype: bool - name: use_heterogeneous_pool dtype: bool - name: watchdog_timeout dtype: float64 - name: client_args struct: - name: dataset dtype: string - name: endpoint dtype: string - name: goodput dtype: 'null' - name: metric_percentiles dtype: string - name: model dtype: string - name: num_prompts dtype: int64 - name: output dtype: string - name: percentile_metrics dtype: string - name: request_rate dtype: float64 - name: tokenizer dtype: string - name: wait_until_ready dtype: bool - name: metrics struct: - name: completed dtype: int64 - name: mean_e2el_ms dtype: float64 - name: mean_itl_ms dtype: float64 - name: mean_tpot_ms dtype: float64 - name: mean_ttft_ms dtype: float64 - name: median_e2el_ms dtype: float64 - name: median_itl_ms dtype: float64 - name: median_tpot_ms dtype: float64 - name: median_ttft_ms dtype: float64 - name: output_throughput dtype: float64 - name: percentiles_e2el_ms sequence: sequence: float64 - name: percentiles_itl_ms sequence: sequence: float64 - name: percentiles_tpot_ms sequence: sequence: float64 - name: percentiles_ttft_ms sequence: sequence: float64 - name: request_goodput dtype: float64 - name: request_throughput dtype: float64 - name: std_e2el_ms dtype: float64 - name: std_itl_ms dtype: float64 - name: std_tpot_ms dtype: float64 - name: std_ttft_ms dtype: float64 - name: total_input dtype: int64 - name: total_output dtype: int64 - name: total_token_throughput dtype: float64 - name: device dtype: string splits: - name: train num_bytes: 9732 num_examples: 7 download_size: 56303 dataset_size: 9732 - config_name: 1xA100_EXPERTS_2_QUANT_SBMM_FORLOOP features: - name: server_args struct: - name: allow_toppings_registration dtype: bool - name: api_key dtype: 'null' - name: attention_backend dtype: string - name: base_min_new_token_ratio dtype: float64 - name: chat_template dtype: 'null' - name: chunked_prefill_size dtype: int64 - name: constrained_json_whitespace_pattern dtype: 'null' - name: context_length dtype: int64 - name: controller_port dtype: int64 - name: debug dtype: bool - name: decode_log_interval dtype: int64 - name: detokenizer_ipc_name dtype: string - name: detokenizer_port dtype: int64 - name: device dtype: string - name: disable_cuda_graph dtype: bool - name: disable_cuda_graph_padding dtype: bool - name: disable_custom_all_reduce dtype: bool - name: disable_disk_cache dtype: bool - name: disable_flashinfer dtype: bool - name: disable_flashinfer_sampling dtype: bool - name: disable_mla dtype: bool - name: disable_nan_detection dtype: bool - name: disable_penalizer dtype: bool - name: disable_radix_cache dtype: bool - name: disable_regex_jump_forward dtype: bool - name: dist_init_addr dtype: 'null' - name: dp_size dtype: int64 - name: dtype dtype: string - name: enable_double_sparsity dtype: bool - name: enable_dp_attention dtype: bool - name: enable_mixed_chunk dtype: bool - name: enable_overlap_schedule dtype: bool - name: enable_p2p_check dtype: bool - name: enable_stats_logging dtype: bool - name: enable_system_controller dtype: bool - name: enable_toppings dtype: bool - name: enable_torch_compile dtype: bool - name: flashinfer_workspace_size dtype: int64 - name: host dtype: string - name: init_new_token_ratio dtype: float64 - name: init_number_of_deltas dtype: int64 - name: init_number_of_loras dtype: int64 - name: init_toppings dtype: 'null' - name: is_embedding dtype: bool - name: kv_cache_dtype dtype: string - name: load_balance_method dtype: string - name: load_format dtype: string - name: log_requests dtype: bool - name: lora_paths dtype: 'null' - name: max_lora_ranks dtype: int64 - name: max_prefill_tokens dtype: int64 - name: max_running_requests dtype: 'null' - name: max_toppings_per_batch dtype: int64 - name: max_torch_compile_bs dtype: int64 - name: max_total_tokens dtype: 'null' - name: mem_fraction_static dtype: float64 - name: model_path dtype: string - name: nccl_ports sequence: string - name: new_token_ratio_decay dtype: float64 - name: nnodes dtype: int64 - name: node_rank dtype: int64 - name: num_continue_decode_steps dtype: int64 - name: num_continuous_decode_steps dtype: int64 - name: port dtype: int64 - name: quantization dtype: 'null' - name: random_seed dtype: int64 - name: retract_decode_steps dtype: int64 - name: sampling_backend dtype: string - name: schedule_conservativeness dtype: float64 - name: schedule_policy dtype: string - name: scheduler_input_ipc_name dtype: string - name: scheduler_port dtype: int64 - name: served_model_name dtype: string - name: server_id dtype: string - name: show_time_cost dtype: bool - name: skip_tokenizer_init dtype: bool - name: stream_interval dtype: int64 - name: tokenizer_ipc_name dtype: string - name: tokenizer_mode dtype: string - name: tokenizer_path dtype: string - name: tokenizer_port dtype: int64 - name: torchao_config dtype: string - name: tp_size dtype: int64 - name: triton_attention_reduce_in_fp32 dtype: bool - name: trust_remote_code dtype: bool - name: use_heterogeneous_pool dtype: bool - name: watchdog_timeout dtype: float64 - name: client_args struct: - name: dataset dtype: string - name: endpoint dtype: string - name: goodput dtype: 'null' - name: metric_percentiles dtype: string - name: model dtype: string - name: num_prompts dtype: int64 - name: output dtype: string - name: percentile_metrics dtype: string - name: request_rate dtype: float64 - name: tokenizer dtype: string - name: wait_until_ready dtype: bool - name: metrics struct: - name: completed dtype: int64 - name: mean_e2el_ms dtype: float64 - name: mean_itl_ms dtype: float64 - name: mean_tpot_ms dtype: float64 - name: mean_ttft_ms dtype: float64 - name: median_e2el_ms dtype: float64 - name: median_itl_ms dtype: float64 - name: median_tpot_ms dtype: float64 - name: median_ttft_ms dtype: float64 - name: output_throughput dtype: float64 - name: percentiles_e2el_ms sequence: sequence: float64 - name: percentiles_itl_ms sequence: sequence: float64 - name: percentiles_tpot_ms sequence: sequence: float64 - name: percentiles_ttft_ms sequence: sequence: float64 - name: request_goodput dtype: float64 - name: request_throughput dtype: float64 - name: std_e2el_ms dtype: float64 - name: std_itl_ms dtype: float64 - name: std_tpot_ms dtype: float64 - name: std_ttft_ms dtype: float64 - name: total_input dtype: int64 - name: total_output dtype: int64 - name: total_token_throughput dtype: float64 - name: device dtype: string splits: - name: train num_bytes: 9942 num_examples: 7 download_size: 56472 dataset_size: 9942 - config_name: 1xA100_EXPERTS_2_QUANT_SBMM_MULTILAUNCH features: - name: server_args struct: - name: allow_toppings_registration dtype: bool - name: api_key dtype: 'null' - name: attention_backend dtype: string - name: base_min_new_token_ratio dtype: float64 - name: chat_template dtype: 'null' - name: chunked_prefill_size dtype: int64 - name: constrained_json_whitespace_pattern dtype: 'null' - name: context_length dtype: int64 - name: controller_port dtype: int64 - name: debug dtype: bool - name: decode_log_interval dtype: int64 - name: detokenizer_ipc_name dtype: string - name: detokenizer_port dtype: int64 - name: device dtype: string - name: disable_cuda_graph dtype: bool - name: disable_cuda_graph_padding dtype: bool - name: disable_custom_all_reduce dtype: bool - name: disable_disk_cache dtype: bool - name: disable_flashinfer dtype: bool - name: disable_flashinfer_sampling dtype: bool - name: disable_mla dtype: bool - name: disable_nan_detection dtype: bool - name: disable_penalizer dtype: bool - name: disable_radix_cache dtype: bool - name: disable_regex_jump_forward dtype: bool - name: dist_init_addr dtype: 'null' - name: dp_size dtype: int64 - name: dtype dtype: string - name: enable_double_sparsity dtype: bool - name: enable_dp_attention dtype: bool - name: enable_mixed_chunk dtype: bool - name: enable_overlap_schedule dtype: bool - name: enable_p2p_check dtype: bool - name: enable_stats_logging dtype: bool - name: enable_system_controller dtype: bool - name: enable_toppings dtype: bool - name: enable_torch_compile dtype: bool - name: flashinfer_workspace_size dtype: int64 - name: host dtype: string - name: init_new_token_ratio dtype: float64 - name: init_number_of_deltas dtype: int64 - name: init_number_of_loras dtype: int64 - name: init_toppings dtype: 'null' - name: is_embedding dtype: bool - name: kv_cache_dtype dtype: string - name: load_balance_method dtype: string - name: load_format dtype: string - name: log_requests dtype: bool - name: lora_paths dtype: 'null' - name: max_lora_ranks dtype: int64 - name: max_prefill_tokens dtype: int64 - name: max_running_requests dtype: 'null' - name: max_toppings_per_batch dtype: int64 - name: max_torch_compile_bs dtype: int64 - name: max_total_tokens dtype: 'null' - name: mem_fraction_static dtype: float64 - name: model_path dtype: string - name: nccl_ports sequence: string - name: new_token_ratio_decay dtype: float64 - name: nnodes dtype: int64 - name: node_rank dtype: int64 - name: num_continue_decode_steps dtype: int64 - name: num_continuous_decode_steps dtype: int64 - name: port dtype: int64 - name: quantization dtype: 'null' - name: random_seed dtype: int64 - name: retract_decode_steps dtype: int64 - name: sampling_backend dtype: string - name: schedule_conservativeness dtype: float64 - name: schedule_policy dtype: string - name: scheduler_input_ipc_name dtype: string - name: scheduler_port dtype: int64 - name: served_model_name dtype: string - name: server_id dtype: string - name: show_time_cost dtype: bool - name: skip_tokenizer_init dtype: bool - name: stream_interval dtype: int64 - name: tokenizer_ipc_name dtype: string - name: tokenizer_mode dtype: string - name: tokenizer_path dtype: string - name: tokenizer_port dtype: int64 - name: torchao_config dtype: string - name: tp_size dtype: int64 - name: triton_attention_reduce_in_fp32 dtype: bool - name: trust_remote_code dtype: bool - name: use_heterogeneous_pool dtype: bool - name: watchdog_timeout dtype: float64 - name: client_args struct: - name: dataset dtype: string - name: endpoint dtype: string - name: goodput dtype: 'null' - name: metric_percentiles dtype: string - name: model dtype: string - name: num_prompts dtype: int64 - name: output dtype: string - name: percentile_metrics dtype: string - name: request_rate dtype: float64 - name: tokenizer dtype: string - name: wait_until_ready dtype: bool - name: metrics struct: - name: completed dtype: int64 - name: mean_e2el_ms dtype: float64 - name: mean_itl_ms dtype: float64 - name: mean_tpot_ms dtype: float64 - name: mean_ttft_ms dtype: float64 - name: median_e2el_ms dtype: float64 - name: median_itl_ms dtype: float64 - name: median_tpot_ms dtype: float64 - name: median_ttft_ms dtype: float64 - name: output_throughput dtype: float64 - name: percentiles_e2el_ms sequence: sequence: float64 - name: percentiles_itl_ms sequence: sequence: float64 - name: percentiles_tpot_ms sequence: sequence: float64 - name: percentiles_ttft_ms sequence: sequence: float64 - name: request_goodput dtype: float64 - name: request_throughput dtype: float64 - name: std_e2el_ms dtype: float64 - name: std_itl_ms dtype: float64 - name: std_tpot_ms dtype: float64 - name: std_ttft_ms dtype: float64 - name: total_input dtype: int64 - name: total_output dtype: int64 - name: total_token_throughput dtype: float64 - name: device dtype: string splits: - name: train num_bytes: 9998 num_examples: 7 download_size: 56481 dataset_size: 9998 - config_name: 1xA100_EXPERTS_2_QUANT_SBMM_NAIVE features: - name: server_args struct: - name: allow_toppings_registration dtype: bool - name: api_key dtype: 'null' - name: attention_backend dtype: string - name: base_min_new_token_ratio dtype: float64 - name: chat_template dtype: 'null' - name: chunked_prefill_size dtype: int64 - name: constrained_json_whitespace_pattern dtype: 'null' - name: context_length dtype: int64 - name: controller_port dtype: int64 - name: debug dtype: bool - name: decode_log_interval dtype: int64 - name: detokenizer_ipc_name dtype: string - name: detokenizer_port dtype: int64 - name: device dtype: string - name: disable_cuda_graph dtype: bool - name: disable_cuda_graph_padding dtype: bool - name: disable_custom_all_reduce dtype: bool - name: disable_disk_cache dtype: bool - name: disable_flashinfer dtype: bool - name: disable_flashinfer_sampling dtype: bool - name: disable_mla dtype: bool - name: disable_nan_detection dtype: bool - name: disable_penalizer dtype: bool - name: disable_radix_cache dtype: bool - name: disable_regex_jump_forward dtype: bool - name: dist_init_addr dtype: 'null' - name: dp_size dtype: int64 - name: dtype dtype: string - name: enable_double_sparsity dtype: bool - name: enable_dp_attention dtype: bool - name: enable_mixed_chunk dtype: bool - name: enable_overlap_schedule dtype: bool - name: enable_p2p_check dtype: bool - name: enable_stats_logging dtype: bool - name: enable_system_controller dtype: bool - name: enable_toppings dtype: bool - name: enable_torch_compile dtype: bool - name: flashinfer_workspace_size dtype: int64 - name: host dtype: string - name: init_new_token_ratio dtype: float64 - name: init_number_of_deltas dtype: int64 - name: init_number_of_loras dtype: int64 - name: init_toppings dtype: 'null' - name: is_embedding dtype: bool - name: kv_cache_dtype dtype: string - name: load_balance_method dtype: string - name: load_format dtype: string - name: log_requests dtype: bool - name: lora_paths dtype: 'null' - name: max_lora_ranks dtype: int64 - name: max_prefill_tokens dtype: int64 - name: max_running_requests dtype: 'null' - name: max_toppings_per_batch dtype: int64 - name: max_torch_compile_bs dtype: int64 - name: max_total_tokens dtype: 'null' - name: mem_fraction_static dtype: float64 - name: model_path dtype: string - name: nccl_ports sequence: string - name: new_token_ratio_decay dtype: float64 - name: nnodes dtype: int64 - name: node_rank dtype: int64 - name: num_continue_decode_steps dtype: int64 - name: num_continuous_decode_steps dtype: int64 - name: port dtype: int64 - name: quantization dtype: 'null' - name: random_seed dtype: int64 - name: retract_decode_steps dtype: int64 - name: sampling_backend dtype: string - name: schedule_conservativeness dtype: float64 - name: schedule_policy dtype: string - name: scheduler_input_ipc_name dtype: string - name: scheduler_port dtype: int64 - name: served_model_name dtype: string - name: server_id dtype: string - name: show_time_cost dtype: bool - name: skip_tokenizer_init dtype: bool - name: stream_interval dtype: int64 - name: tokenizer_ipc_name dtype: string - name: tokenizer_mode dtype: string - name: tokenizer_path dtype: string - name: tokenizer_port dtype: int64 - name: torchao_config dtype: string - name: tp_size dtype: int64 - name: triton_attention_reduce_in_fp32 dtype: bool - name: trust_remote_code dtype: bool - name: use_heterogeneous_pool dtype: bool - name: watchdog_timeout dtype: float64 - name: client_args struct: - name: dataset dtype: string - name: endpoint dtype: string - name: goodput dtype: 'null' - name: metric_percentiles dtype: string - name: model dtype: string - name: num_prompts dtype: int64 - name: output dtype: string - name: percentile_metrics dtype: string - name: request_rate dtype: float64 - name: tokenizer dtype: string - name: wait_until_ready dtype: bool - name: metrics struct: - name: completed dtype: int64 - name: mean_e2el_ms dtype: float64 - name: mean_itl_ms dtype: float64 - name: mean_tpot_ms dtype: float64 - name: mean_ttft_ms dtype: float64 - name: median_e2el_ms dtype: float64 - name: median_itl_ms dtype: float64 - name: median_tpot_ms dtype: float64 - name: median_ttft_ms dtype: float64 - name: output_throughput dtype: float64 - name: percentiles_e2el_ms sequence: sequence: float64 - name: percentiles_itl_ms sequence: sequence: float64 - name: percentiles_tpot_ms sequence: sequence: float64 - name: percentiles_ttft_ms sequence: sequence: float64 - name: request_goodput dtype: float64 - name: request_throughput dtype: float64 - name: std_e2el_ms dtype: float64 - name: std_itl_ms dtype: float64 - name: std_tpot_ms dtype: float64 - name: std_ttft_ms dtype: float64 - name: total_input dtype: int64 - name: total_output dtype: int64 - name: total_token_throughput dtype: float64 - name: device dtype: string splits: - name: train num_bytes: 9914 num_examples: 7 download_size: 56450 dataset_size: 9914 - config_name: 1xA100_EXPERTS_2_UNQUANT features: - name: server_args struct: - name: allow_toppings_registration dtype: bool - name: api_key dtype: 'null' - name: attention_backend dtype: string - name: base_min_new_token_ratio dtype: float64 - name: chat_template dtype: 'null' - name: chunked_prefill_size dtype: int64 - name: constrained_json_whitespace_pattern dtype: 'null' - name: context_length dtype: int64 - name: controller_port dtype: int64 - name: debug dtype: bool - name: decode_log_interval dtype: int64 - name: detokenizer_ipc_name dtype: string - name: detokenizer_port dtype: int64 - name: device dtype: string - name: disable_cuda_graph dtype: bool - name: disable_cuda_graph_padding dtype: bool - name: disable_custom_all_reduce dtype: bool - name: disable_disk_cache dtype: bool - name: disable_flashinfer dtype: bool - name: disable_flashinfer_sampling dtype: bool - name: disable_mla dtype: bool - name: disable_nan_detection dtype: bool - name: disable_penalizer dtype: bool - name: disable_radix_cache dtype: bool - name: disable_regex_jump_forward dtype: bool - name: dist_init_addr dtype: 'null' - name: dp_size dtype: int64 - name: dtype dtype: string - name: enable_double_sparsity dtype: bool - name: enable_dp_attention dtype: bool - name: enable_mixed_chunk dtype: bool - name: enable_overlap_schedule dtype: bool - name: enable_p2p_check dtype: bool - name: enable_stats_logging dtype: bool - name: enable_system_controller dtype: bool - name: enable_toppings dtype: bool - name: enable_torch_compile dtype: bool - name: flashinfer_workspace_size dtype: int64 - name: host dtype: string - name: init_new_token_ratio dtype: float64 - name: init_number_of_deltas dtype: int64 - name: init_number_of_loras dtype: int64 - name: init_toppings dtype: 'null' - name: is_embedding dtype: bool - name: kv_cache_dtype dtype: string - name: load_balance_method dtype: string - name: load_format dtype: string - name: log_requests dtype: bool - name: lora_paths dtype: 'null' - name: max_lora_ranks dtype: int64 - name: max_prefill_tokens dtype: int64 - name: max_running_requests dtype: 'null' - name: max_toppings_per_batch dtype: int64 - name: max_torch_compile_bs dtype: int64 - name: max_total_tokens dtype: 'null' - name: mem_fraction_static dtype: float64 - name: model_path dtype: string - name: nccl_ports sequence: string - name: new_token_ratio_decay dtype: float64 - name: nnodes dtype: int64 - name: node_rank dtype: int64 - name: num_continue_decode_steps dtype: int64 - name: num_continuous_decode_steps dtype: int64 - name: port dtype: int64 - name: quantization dtype: 'null' - name: random_seed dtype: int64 - name: retract_decode_steps dtype: int64 - name: sampling_backend dtype: string - name: schedule_conservativeness dtype: float64 - name: schedule_policy dtype: string - name: scheduler_input_ipc_name dtype: string - name: scheduler_port dtype: int64 - name: served_model_name dtype: string - name: server_id dtype: string - name: show_time_cost dtype: bool - name: skip_tokenizer_init dtype: bool - name: stream_interval dtype: int64 - name: tokenizer_ipc_name dtype: string - name: tokenizer_mode dtype: string - name: tokenizer_path dtype: string - name: tokenizer_port dtype: int64 - name: torchao_config dtype: string - name: tp_size dtype: int64 - name: triton_attention_reduce_in_fp32 dtype: bool - name: trust_remote_code dtype: bool - name: use_heterogeneous_pool dtype: bool - name: watchdog_timeout dtype: float64 - name: client_args struct: - name: dataset dtype: string - name: endpoint dtype: string - name: goodput dtype: 'null' - name: metric_percentiles dtype: string - name: model dtype: string - name: num_prompts dtype: int64 - name: output dtype: string - name: percentile_metrics dtype: string - name: request_rate dtype: float64 - name: tokenizer dtype: string - name: wait_until_ready dtype: bool - name: metrics struct: - name: completed dtype: int64 - name: mean_e2el_ms dtype: float64 - name: mean_itl_ms dtype: float64 - name: mean_tpot_ms dtype: float64 - name: mean_ttft_ms dtype: float64 - name: median_e2el_ms dtype: float64 - name: median_itl_ms dtype: float64 - name: median_tpot_ms dtype: float64 - name: median_ttft_ms dtype: float64 - name: output_throughput dtype: float64 - name: percentiles_e2el_ms sequence: sequence: float64 - name: percentiles_itl_ms sequence: sequence: float64 - name: percentiles_tpot_ms sequence: sequence: float64 - name: percentiles_ttft_ms sequence: sequence: float64 - name: request_goodput dtype: float64 - name: request_throughput dtype: float64 - name: std_e2el_ms dtype: float64 - name: std_itl_ms dtype: float64 - name: std_tpot_ms dtype: float64 - name: std_ttft_ms dtype: float64 - name: total_input dtype: int64 - name: total_output dtype: int64 - name: total_token_throughput dtype: float64 - name: device dtype: string splits: - name: train num_bytes: 8920 num_examples: 7 download_size: 55709 dataset_size: 8920 - config_name: 1xA100_EXPERTS_3_QUANT_NAIVE features: - name: server_args struct: - name: allow_toppings_registration dtype: bool - name: api_key dtype: 'null' - name: attention_backend dtype: string - name: base_min_new_token_ratio dtype: float64 - name: chat_template dtype: 'null' - name: chunked_prefill_size dtype: int64 - name: constrained_json_whitespace_pattern dtype: 'null' - name: context_length dtype: int64 - name: controller_port dtype: int64 - name: debug dtype: bool - name: decode_log_interval dtype: int64 - name: detokenizer_ipc_name dtype: string - name: detokenizer_port dtype: int64 - name: device dtype: string - name: disable_cuda_graph dtype: bool - name: disable_cuda_graph_padding dtype: bool - name: disable_custom_all_reduce dtype: bool - name: disable_disk_cache dtype: bool - name: disable_flashinfer dtype: bool - name: disable_flashinfer_sampling dtype: bool - name: disable_mla dtype: bool - name: disable_nan_detection dtype: bool - name: disable_penalizer dtype: bool - name: disable_radix_cache dtype: bool - name: disable_regex_jump_forward dtype: bool - name: dist_init_addr dtype: 'null' - name: dp_size dtype: int64 - name: dtype dtype: string - name: enable_double_sparsity dtype: bool - name: enable_dp_attention dtype: bool - name: enable_mixed_chunk dtype: bool - name: enable_overlap_schedule dtype: bool - name: enable_p2p_check dtype: bool - name: enable_stats_logging dtype: bool - name: enable_system_controller dtype: bool - name: enable_toppings dtype: bool - name: enable_torch_compile dtype: bool - name: flashinfer_workspace_size dtype: int64 - name: host dtype: string - name: init_new_token_ratio dtype: float64 - name: init_number_of_deltas dtype: int64 - name: init_number_of_loras dtype: int64 - name: init_toppings dtype: 'null' - name: is_embedding dtype: bool - name: kv_cache_dtype dtype: string - name: load_balance_method dtype: string - name: load_format dtype: string - name: log_requests dtype: bool - name: lora_paths dtype: 'null' - name: max_lora_ranks dtype: int64 - name: max_prefill_tokens dtype: int64 - name: max_running_requests dtype: 'null' - name: max_toppings_per_batch dtype: int64 - name: max_torch_compile_bs dtype: int64 - name: max_total_tokens dtype: 'null' - name: mem_fraction_static dtype: float64 - name: model_path dtype: string - name: nccl_ports sequence: string - name: new_token_ratio_decay dtype: float64 - name: nnodes dtype: int64 - name: node_rank dtype: int64 - name: num_continue_decode_steps dtype: int64 - name: num_continuous_decode_steps dtype: int64 - name: port dtype: int64 - name: quantization dtype: 'null' - name: random_seed dtype: int64 - name: retract_decode_steps dtype: int64 - name: sampling_backend dtype: string - name: schedule_conservativeness dtype: float64 - name: schedule_policy dtype: string - name: scheduler_input_ipc_name dtype: string - name: scheduler_port dtype: int64 - name: served_model_name dtype: string - name: server_id dtype: string - name: show_time_cost dtype: bool - name: skip_tokenizer_init dtype: bool - name: stream_interval dtype: int64 - name: tokenizer_ipc_name dtype: string - name: tokenizer_mode dtype: string - name: tokenizer_path dtype: string - name: tokenizer_port dtype: int64 - name: torchao_config dtype: string - name: tp_size dtype: int64 - name: triton_attention_reduce_in_fp32 dtype: bool - name: trust_remote_code dtype: bool - name: use_heterogeneous_pool dtype: bool - name: watchdog_timeout dtype: float64 - name: client_args struct: - name: dataset dtype: string - name: endpoint dtype: string - name: goodput dtype: 'null' - name: metric_percentiles dtype: string - name: model dtype: string - name: num_prompts dtype: int64 - name: output dtype: string - name: percentile_metrics dtype: string - name: request_rate dtype: float64 - name: tokenizer dtype: string - name: wait_until_ready dtype: bool - name: metrics struct: - name: completed dtype: int64 - name: mean_e2el_ms dtype: float64 - name: mean_itl_ms dtype: float64 - name: mean_tpot_ms dtype: float64 - name: mean_ttft_ms dtype: float64 - name: median_e2el_ms dtype: float64 - name: median_itl_ms dtype: float64 - name: median_tpot_ms dtype: float64 - name: median_ttft_ms dtype: float64 - name: output_throughput dtype: float64 - name: percentiles_e2el_ms sequence: sequence: float64 - name: percentiles_itl_ms sequence: sequence: float64 - name: percentiles_tpot_ms sequence: sequence: float64 - name: percentiles_ttft_ms sequence: sequence: float64 - name: request_goodput dtype: float64 - name: request_throughput dtype: float64 - name: std_e2el_ms dtype: float64 - name: std_itl_ms dtype: float64 - name: std_tpot_ms dtype: float64 - name: std_ttft_ms dtype: float64 - name: total_input dtype: int64 - name: total_output dtype: int64 - name: total_token_throughput dtype: float64 - name: device dtype: string splits: - name: train num_bytes: 9732 num_examples: 7 download_size: 56291 dataset_size: 9732 - config_name: 1xA100_EXPERTS_3_QUANT_SBMM_FORLOOP features: - name: server_args struct: - name: allow_toppings_registration dtype: bool - name: api_key dtype: 'null' - name: attention_backend dtype: string - name: base_min_new_token_ratio dtype: float64 - name: chat_template dtype: 'null' - name: chunked_prefill_size dtype: int64 - name: constrained_json_whitespace_pattern dtype: 'null' - name: context_length dtype: int64 - name: controller_port dtype: int64 - name: debug dtype: bool - name: decode_log_interval dtype: int64 - name: detokenizer_ipc_name dtype: string - name: detokenizer_port dtype: int64 - name: device dtype: string - name: disable_cuda_graph dtype: bool - name: disable_cuda_graph_padding dtype: bool - name: disable_custom_all_reduce dtype: bool - name: disable_disk_cache dtype: bool - name: disable_flashinfer dtype: bool - name: disable_flashinfer_sampling dtype: bool - name: disable_mla dtype: bool - name: disable_nan_detection dtype: bool - name: disable_penalizer dtype: bool - name: disable_radix_cache dtype: bool - name: disable_regex_jump_forward dtype: bool - name: dist_init_addr dtype: 'null' - name: dp_size dtype: int64 - name: dtype dtype: string - name: enable_double_sparsity dtype: bool - name: enable_dp_attention dtype: bool - name: enable_mixed_chunk dtype: bool - name: enable_overlap_schedule dtype: bool - name: enable_p2p_check dtype: bool - name: enable_stats_logging dtype: bool - name: enable_system_controller dtype: bool - name: enable_toppings dtype: bool - name: enable_torch_compile dtype: bool - name: flashinfer_workspace_size dtype: int64 - name: host dtype: string - name: init_new_token_ratio dtype: float64 - name: init_number_of_deltas dtype: int64 - name: init_number_of_loras dtype: int64 - name: init_toppings dtype: 'null' - name: is_embedding dtype: bool - name: kv_cache_dtype dtype: string - name: load_balance_method dtype: string - name: load_format dtype: string - name: log_requests dtype: bool - name: lora_paths dtype: 'null' - name: max_lora_ranks dtype: int64 - name: max_prefill_tokens dtype: int64 - name: max_running_requests dtype: 'null' - name: max_toppings_per_batch dtype: int64 - name: max_torch_compile_bs dtype: int64 - name: max_total_tokens dtype: 'null' - name: mem_fraction_static dtype: float64 - name: model_path dtype: string - name: nccl_ports sequence: string - name: new_token_ratio_decay dtype: float64 - name: nnodes dtype: int64 - name: node_rank dtype: int64 - name: num_continue_decode_steps dtype: int64 - name: num_continuous_decode_steps dtype: int64 - name: port dtype: int64 - name: quantization dtype: 'null' - name: random_seed dtype: int64 - name: retract_decode_steps dtype: int64 - name: sampling_backend dtype: string - name: schedule_conservativeness dtype: float64 - name: schedule_policy dtype: string - name: scheduler_input_ipc_name dtype: string - name: scheduler_port dtype: int64 - name: served_model_name dtype: string - name: server_id dtype: string - name: show_time_cost dtype: bool - name: skip_tokenizer_init dtype: bool - name: stream_interval dtype: int64 - name: tokenizer_ipc_name dtype: string - name: tokenizer_mode dtype: string - name: tokenizer_path dtype: string - name: tokenizer_port dtype: int64 - name: torchao_config dtype: string - name: tp_size dtype: int64 - name: triton_attention_reduce_in_fp32 dtype: bool - name: trust_remote_code dtype: bool - name: use_heterogeneous_pool dtype: bool - name: watchdog_timeout dtype: float64 - name: client_args struct: - name: dataset dtype: string - name: endpoint dtype: string - name: goodput dtype: 'null' - name: metric_percentiles dtype: string - name: model dtype: string - name: num_prompts dtype: int64 - name: output dtype: string - name: percentile_metrics dtype: string - name: request_rate dtype: float64 - name: tokenizer dtype: string - name: wait_until_ready dtype: bool - name: metrics struct: - name: completed dtype: int64 - name: mean_e2el_ms dtype: float64 - name: mean_itl_ms dtype: float64 - name: mean_tpot_ms dtype: float64 - name: mean_ttft_ms dtype: float64 - name: median_e2el_ms dtype: float64 - name: median_itl_ms dtype: float64 - name: median_tpot_ms dtype: float64 - name: median_ttft_ms dtype: float64 - name: output_throughput dtype: float64 - name: percentiles_e2el_ms sequence: sequence: float64 - name: percentiles_itl_ms sequence: sequence: float64 - name: percentiles_tpot_ms sequence: sequence: float64 - name: percentiles_ttft_ms sequence: sequence: float64 - name: request_goodput dtype: float64 - name: request_throughput dtype: float64 - name: std_e2el_ms dtype: float64 - name: std_itl_ms dtype: float64 - name: std_tpot_ms dtype: float64 - name: std_ttft_ms dtype: float64 - name: total_input dtype: int64 - name: total_output dtype: int64 - name: total_token_throughput dtype: float64 - name: device dtype: string splits: - name: train num_bytes: 9942 num_examples: 7 download_size: 56445 dataset_size: 9942 - config_name: 1xA100_EXPERTS_3_QUANT_SBMM_MULTILAUNCH features: - name: server_args struct: - name: allow_toppings_registration dtype: bool - name: api_key dtype: 'null' - name: attention_backend dtype: string - name: base_min_new_token_ratio dtype: float64 - name: chat_template dtype: 'null' - name: chunked_prefill_size dtype: int64 - name: constrained_json_whitespace_pattern dtype: 'null' - name: context_length dtype: int64 - name: controller_port dtype: int64 - name: debug dtype: bool - name: decode_log_interval dtype: int64 - name: detokenizer_ipc_name dtype: string - name: detokenizer_port dtype: int64 - name: device dtype: string - name: disable_cuda_graph dtype: bool - name: disable_cuda_graph_padding dtype: bool - name: disable_custom_all_reduce dtype: bool - name: disable_disk_cache dtype: bool - name: disable_flashinfer dtype: bool - name: disable_flashinfer_sampling dtype: bool - name: disable_mla dtype: bool - name: disable_nan_detection dtype: bool - name: disable_penalizer dtype: bool - name: disable_radix_cache dtype: bool - name: disable_regex_jump_forward dtype: bool - name: dist_init_addr dtype: 'null' - name: dp_size dtype: int64 - name: dtype dtype: string - name: enable_double_sparsity dtype: bool - name: enable_dp_attention dtype: bool - name: enable_mixed_chunk dtype: bool - name: enable_overlap_schedule dtype: bool - name: enable_p2p_check dtype: bool - name: enable_stats_logging dtype: bool - name: enable_system_controller dtype: bool - name: enable_toppings dtype: bool - name: enable_torch_compile dtype: bool - name: flashinfer_workspace_size dtype: int64 - name: host dtype: string - name: init_new_token_ratio dtype: float64 - name: init_number_of_deltas dtype: int64 - name: init_number_of_loras dtype: int64 - name: init_toppings dtype: 'null' - name: is_embedding dtype: bool - name: kv_cache_dtype dtype: string - name: load_balance_method dtype: string - name: load_format dtype: string - name: log_requests dtype: bool - name: lora_paths dtype: 'null' - name: max_lora_ranks dtype: int64 - name: max_prefill_tokens dtype: int64 - name: max_running_requests dtype: 'null' - name: max_toppings_per_batch dtype: int64 - name: max_torch_compile_bs dtype: int64 - name: max_total_tokens dtype: 'null' - name: mem_fraction_static dtype: float64 - name: model_path dtype: string - name: nccl_ports sequence: string - name: new_token_ratio_decay dtype: float64 - name: nnodes dtype: int64 - name: node_rank dtype: int64 - name: num_continue_decode_steps dtype: int64 - name: num_continuous_decode_steps dtype: int64 - name: port dtype: int64 - name: quantization dtype: 'null' - name: random_seed dtype: int64 - name: retract_decode_steps dtype: int64 - name: sampling_backend dtype: string - name: schedule_conservativeness dtype: float64 - name: schedule_policy dtype: string - name: scheduler_input_ipc_name dtype: string - name: scheduler_port dtype: int64 - name: served_model_name dtype: string - name: server_id dtype: string - name: show_time_cost dtype: bool - name: skip_tokenizer_init dtype: bool - name: stream_interval dtype: int64 - name: tokenizer_ipc_name dtype: string - name: tokenizer_mode dtype: string - name: tokenizer_path dtype: string - name: tokenizer_port dtype: int64 - name: torchao_config dtype: string - name: tp_size dtype: int64 - name: triton_attention_reduce_in_fp32 dtype: bool - name: trust_remote_code dtype: bool - name: use_heterogeneous_pool dtype: bool - name: watchdog_timeout dtype: float64 - name: client_args struct: - name: dataset dtype: string - name: endpoint dtype: string - name: goodput dtype: 'null' - name: metric_percentiles dtype: string - name: model dtype: string - name: num_prompts dtype: int64 - name: output dtype: string - name: percentile_metrics dtype: string - name: request_rate dtype: float64 - name: tokenizer dtype: string - name: wait_until_ready dtype: bool - name: metrics struct: - name: completed dtype: int64 - name: mean_e2el_ms dtype: float64 - name: mean_itl_ms dtype: float64 - name: mean_tpot_ms dtype: float64 - name: mean_ttft_ms dtype: float64 - name: median_e2el_ms dtype: float64 - name: median_itl_ms dtype: float64 - name: median_tpot_ms dtype: float64 - name: median_ttft_ms dtype: float64 - name: output_throughput dtype: float64 - name: percentiles_e2el_ms sequence: sequence: float64 - name: percentiles_itl_ms sequence: sequence: float64 - name: percentiles_tpot_ms sequence: sequence: float64 - name: percentiles_ttft_ms sequence: sequence: float64 - name: request_goodput dtype: float64 - name: request_throughput dtype: float64 - name: std_e2el_ms dtype: float64 - name: std_itl_ms dtype: float64 - name: std_tpot_ms dtype: float64 - name: std_ttft_ms dtype: float64 - name: total_input dtype: int64 - name: total_output dtype: int64 - name: total_token_throughput dtype: float64 - name: device dtype: string splits: - name: train num_bytes: 9998 num_examples: 7 download_size: 56494 dataset_size: 9998 - config_name: 1xA100_EXPERTS_3_QUANT_SBMM_NAIVE features: - name: server_args struct: - name: allow_toppings_registration dtype: bool - name: api_key dtype: 'null' - name: attention_backend dtype: string - name: base_min_new_token_ratio dtype: float64 - name: chat_template dtype: 'null' - name: chunked_prefill_size dtype: int64 - name: constrained_json_whitespace_pattern dtype: 'null' - name: context_length dtype: int64 - name: controller_port dtype: int64 - name: debug dtype: bool - name: decode_log_interval dtype: int64 - name: detokenizer_ipc_name dtype: string - name: detokenizer_port dtype: int64 - name: device dtype: string - name: disable_cuda_graph dtype: bool - name: disable_cuda_graph_padding dtype: bool - name: disable_custom_all_reduce dtype: bool - name: disable_disk_cache dtype: bool - name: disable_flashinfer dtype: bool - name: disable_flashinfer_sampling dtype: bool - name: disable_mla dtype: bool - name: disable_nan_detection dtype: bool - name: disable_penalizer dtype: bool - name: disable_radix_cache dtype: bool - name: disable_regex_jump_forward dtype: bool - name: dist_init_addr dtype: 'null' - name: dp_size dtype: int64 - name: dtype dtype: string - name: enable_double_sparsity dtype: bool - name: enable_dp_attention dtype: bool - name: enable_mixed_chunk dtype: bool - name: enable_overlap_schedule dtype: bool - name: enable_p2p_check dtype: bool - name: enable_stats_logging dtype: bool - name: enable_system_controller dtype: bool - name: enable_toppings dtype: bool - name: enable_torch_compile dtype: bool - name: flashinfer_workspace_size dtype: int64 - name: host dtype: string - name: init_new_token_ratio dtype: float64 - name: init_number_of_deltas dtype: int64 - name: init_number_of_loras dtype: int64 - name: init_toppings dtype: 'null' - name: is_embedding dtype: bool - name: kv_cache_dtype dtype: string - name: load_balance_method dtype: string - name: load_format dtype: string - name: log_requests dtype: bool - name: lora_paths dtype: 'null' - name: max_lora_ranks dtype: int64 - name: max_prefill_tokens dtype: int64 - name: max_running_requests dtype: 'null' - name: max_toppings_per_batch dtype: int64 - name: max_torch_compile_bs dtype: int64 - name: max_total_tokens dtype: 'null' - name: mem_fraction_static dtype: float64 - name: model_path dtype: string - name: nccl_ports sequence: string - name: new_token_ratio_decay dtype: float64 - name: nnodes dtype: int64 - name: node_rank dtype: int64 - name: num_continue_decode_steps dtype: int64 - name: num_continuous_decode_steps dtype: int64 - name: port dtype: int64 - name: quantization dtype: 'null' - name: random_seed dtype: int64 - name: retract_decode_steps dtype: int64 - name: sampling_backend dtype: string - name: schedule_conservativeness dtype: float64 - name: schedule_policy dtype: string - name: scheduler_input_ipc_name dtype: string - name: scheduler_port dtype: int64 - name: served_model_name dtype: string - name: server_id dtype: string - name: show_time_cost dtype: bool - name: skip_tokenizer_init dtype: bool - name: stream_interval dtype: int64 - name: tokenizer_ipc_name dtype: string - name: tokenizer_mode dtype: string - name: tokenizer_path dtype: string - name: tokenizer_port dtype: int64 - name: torchao_config dtype: string - name: tp_size dtype: int64 - name: triton_attention_reduce_in_fp32 dtype: bool - name: trust_remote_code dtype: bool - name: use_heterogeneous_pool dtype: bool - name: watchdog_timeout dtype: float64 - name: client_args struct: - name: dataset dtype: string - name: endpoint dtype: string - name: goodput dtype: 'null' - name: metric_percentiles dtype: string - name: model dtype: string - name: num_prompts dtype: int64 - name: output dtype: string - name: percentile_metrics dtype: string - name: request_rate dtype: float64 - name: tokenizer dtype: string - name: wait_until_ready dtype: bool - name: metrics struct: - name: completed dtype: int64 - name: mean_e2el_ms dtype: float64 - name: mean_itl_ms dtype: float64 - name: mean_tpot_ms dtype: float64 - name: mean_ttft_ms dtype: float64 - name: median_e2el_ms dtype: float64 - name: median_itl_ms dtype: float64 - name: median_tpot_ms dtype: float64 - name: median_ttft_ms dtype: float64 - name: output_throughput dtype: float64 - name: percentiles_e2el_ms sequence: sequence: float64 - name: percentiles_itl_ms sequence: sequence: float64 - name: percentiles_tpot_ms sequence: sequence: float64 - name: percentiles_ttft_ms sequence: sequence: float64 - name: request_goodput dtype: float64 - name: request_throughput dtype: float64 - name: std_e2el_ms dtype: float64 - name: std_itl_ms dtype: float64 - name: std_tpot_ms dtype: float64 - name: std_ttft_ms dtype: float64 - name: total_input dtype: int64 - name: total_output dtype: int64 - name: total_token_throughput dtype: float64 - name: device dtype: string splits: - name: train num_bytes: 9914 num_examples: 7 download_size: 56440 dataset_size: 9914 - config_name: 1xA100_EXPERTS_3_UNQUANT features: - name: server_args struct: - name: allow_toppings_registration dtype: bool - name: api_key dtype: 'null' - name: attention_backend dtype: string - name: base_min_new_token_ratio dtype: float64 - name: chat_template dtype: 'null' - name: chunked_prefill_size dtype: int64 - name: constrained_json_whitespace_pattern dtype: 'null' - name: context_length dtype: int64 - name: controller_port dtype: int64 - name: debug dtype: bool - name: decode_log_interval dtype: int64 - name: detokenizer_ipc_name dtype: string - name: detokenizer_port dtype: int64 - name: device dtype: string - name: disable_cuda_graph dtype: bool - name: disable_cuda_graph_padding dtype: bool - name: disable_custom_all_reduce dtype: bool - name: disable_disk_cache dtype: bool - name: disable_flashinfer dtype: bool - name: disable_flashinfer_sampling dtype: bool - name: disable_mla dtype: bool - name: disable_nan_detection dtype: bool - name: disable_penalizer dtype: bool - name: disable_radix_cache dtype: bool - name: disable_regex_jump_forward dtype: bool - name: dist_init_addr dtype: 'null' - name: dp_size dtype: int64 - name: dtype dtype: string - name: enable_double_sparsity dtype: bool - name: enable_dp_attention dtype: bool - name: enable_mixed_chunk dtype: bool - name: enable_overlap_schedule dtype: bool - name: enable_p2p_check dtype: bool - name: enable_stats_logging dtype: bool - name: enable_system_controller dtype: bool - name: enable_toppings dtype: bool - name: enable_torch_compile dtype: bool - name: flashinfer_workspace_size dtype: int64 - name: host dtype: string - name: init_new_token_ratio dtype: float64 - name: init_number_of_deltas dtype: int64 - name: init_number_of_loras dtype: int64 - name: init_toppings dtype: 'null' - name: is_embedding dtype: bool - name: kv_cache_dtype dtype: string - name: load_balance_method dtype: string - name: load_format dtype: string - name: log_requests dtype: bool - name: lora_paths dtype: 'null' - name: max_lora_ranks dtype: int64 - name: max_prefill_tokens dtype: int64 - name: max_running_requests dtype: 'null' - name: max_toppings_per_batch dtype: int64 - name: max_torch_compile_bs dtype: int64 - name: max_total_tokens dtype: 'null' - name: mem_fraction_static dtype: float64 - name: model_path dtype: string - name: nccl_ports sequence: string - name: new_token_ratio_decay dtype: float64 - name: nnodes dtype: int64 - name: node_rank dtype: int64 - name: num_continue_decode_steps dtype: int64 - name: num_continuous_decode_steps dtype: int64 - name: port dtype: int64 - name: quantization dtype: 'null' - name: random_seed dtype: int64 - name: retract_decode_steps dtype: int64 - name: sampling_backend dtype: string - name: schedule_conservativeness dtype: float64 - name: schedule_policy dtype: string - name: scheduler_input_ipc_name dtype: string - name: scheduler_port dtype: int64 - name: served_model_name dtype: string - name: server_id dtype: string - name: show_time_cost dtype: bool - name: skip_tokenizer_init dtype: bool - name: stream_interval dtype: int64 - name: tokenizer_ipc_name dtype: string - name: tokenizer_mode dtype: string - name: tokenizer_path dtype: string - name: tokenizer_port dtype: int64 - name: torchao_config dtype: string - name: tp_size dtype: int64 - name: triton_attention_reduce_in_fp32 dtype: bool - name: trust_remote_code dtype: bool - name: use_heterogeneous_pool dtype: bool - name: watchdog_timeout dtype: float64 - name: client_args struct: - name: dataset dtype: string - name: endpoint dtype: string - name: goodput dtype: 'null' - name: metric_percentiles dtype: string - name: model dtype: string - name: num_prompts dtype: int64 - name: output dtype: string - name: percentile_metrics dtype: string - name: request_rate dtype: float64 - name: tokenizer dtype: string - name: wait_until_ready dtype: bool - name: metrics struct: - name: completed dtype: int64 - name: mean_e2el_ms dtype: float64 - name: mean_itl_ms dtype: float64 - name: mean_tpot_ms dtype: float64 - name: mean_ttft_ms dtype: float64 - name: median_e2el_ms dtype: float64 - name: median_itl_ms dtype: float64 - name: median_tpot_ms dtype: float64 - name: median_ttft_ms dtype: float64 - name: output_throughput dtype: float64 - name: percentiles_e2el_ms sequence: sequence: float64 - name: percentiles_itl_ms sequence: sequence: float64 - name: percentiles_tpot_ms sequence: sequence: float64 - name: percentiles_ttft_ms sequence: sequence: float64 - name: request_goodput dtype: float64 - name: request_throughput dtype: float64 - name: std_e2el_ms dtype: float64 - name: std_itl_ms dtype: float64 - name: std_tpot_ms dtype: float64 - name: std_ttft_ms dtype: float64 - name: total_input dtype: int64 - name: total_output dtype: int64 - name: total_token_throughput dtype: float64 - name: device dtype: string splits: - name: train num_bytes: 8920 num_examples: 7 download_size: 55721 dataset_size: 8920 - config_name: 1xA100_EXPERTS_4_QUANT_NAIVE features: - name: server_args struct: - name: allow_toppings_registration dtype: bool - name: api_key dtype: 'null' - name: attention_backend dtype: string - name: base_min_new_token_ratio dtype: float64 - name: chat_template dtype: 'null' - name: chunked_prefill_size dtype: int64 - name: constrained_json_whitespace_pattern dtype: 'null' - name: context_length dtype: int64 - name: controller_port dtype: int64 - name: debug dtype: bool - name: decode_log_interval dtype: int64 - name: detokenizer_ipc_name dtype: string - name: detokenizer_port dtype: int64 - name: device dtype: string - name: disable_cuda_graph dtype: bool - name: disable_cuda_graph_padding dtype: bool - name: disable_custom_all_reduce dtype: bool - name: disable_disk_cache dtype: bool - name: disable_flashinfer dtype: bool - name: disable_flashinfer_sampling dtype: bool - name: disable_mla dtype: bool - name: disable_nan_detection dtype: bool - name: disable_penalizer dtype: bool - name: disable_radix_cache dtype: bool - name: disable_regex_jump_forward dtype: bool - name: dist_init_addr dtype: 'null' - name: dp_size dtype: int64 - name: dtype dtype: string - name: enable_double_sparsity dtype: bool - name: enable_dp_attention dtype: bool - name: enable_mixed_chunk dtype: bool - name: enable_overlap_schedule dtype: bool - name: enable_p2p_check dtype: bool - name: enable_stats_logging dtype: bool - name: enable_system_controller dtype: bool - name: enable_toppings dtype: bool - name: enable_torch_compile dtype: bool - name: flashinfer_workspace_size dtype: int64 - name: host dtype: string - name: init_new_token_ratio dtype: float64 - name: init_number_of_deltas dtype: int64 - name: init_number_of_loras dtype: int64 - name: init_toppings dtype: 'null' - name: is_embedding dtype: bool - name: kv_cache_dtype dtype: string - name: load_balance_method dtype: string - name: load_format dtype: string - name: log_requests dtype: bool - name: lora_paths dtype: 'null' - name: max_lora_ranks dtype: int64 - name: max_prefill_tokens dtype: int64 - name: max_running_requests dtype: 'null' - name: max_toppings_per_batch dtype: int64 - name: max_torch_compile_bs dtype: int64 - name: max_total_tokens dtype: 'null' - name: mem_fraction_static dtype: float64 - name: model_path dtype: string - name: nccl_ports sequence: string - name: new_token_ratio_decay dtype: float64 - name: nnodes dtype: int64 - name: node_rank dtype: int64 - name: num_continue_decode_steps dtype: int64 - name: num_continuous_decode_steps dtype: int64 - name: port dtype: int64 - name: quantization dtype: 'null' - name: random_seed dtype: int64 - name: retract_decode_steps dtype: int64 - name: sampling_backend dtype: string - name: schedule_conservativeness dtype: float64 - name: schedule_policy dtype: string - name: scheduler_input_ipc_name dtype: string - name: scheduler_port dtype: int64 - name: served_model_name dtype: string - name: server_id dtype: string - name: show_time_cost dtype: bool - name: skip_tokenizer_init dtype: bool - name: stream_interval dtype: int64 - name: tokenizer_ipc_name dtype: string - name: tokenizer_mode dtype: string - name: tokenizer_path dtype: string - name: tokenizer_port dtype: int64 - name: torchao_config dtype: string - name: tp_size dtype: int64 - name: triton_attention_reduce_in_fp32 dtype: bool - name: trust_remote_code dtype: bool - name: use_heterogeneous_pool dtype: bool - name: watchdog_timeout dtype: float64 - name: client_args struct: - name: dataset dtype: string - name: endpoint dtype: string - name: goodput dtype: 'null' - name: metric_percentiles dtype: string - name: model dtype: string - name: num_prompts dtype: int64 - name: output dtype: string - name: percentile_metrics dtype: string - name: request_rate dtype: float64 - name: tokenizer dtype: string - name: wait_until_ready dtype: bool - name: metrics struct: - name: completed dtype: int64 - name: mean_e2el_ms dtype: float64 - name: mean_itl_ms dtype: float64 - name: mean_tpot_ms dtype: float64 - name: mean_ttft_ms dtype: float64 - name: median_e2el_ms dtype: float64 - name: median_itl_ms dtype: float64 - name: median_tpot_ms dtype: float64 - name: median_ttft_ms dtype: float64 - name: output_throughput dtype: float64 - name: percentiles_e2el_ms sequence: sequence: float64 - name: percentiles_itl_ms sequence: sequence: float64 - name: percentiles_tpot_ms sequence: sequence: float64 - name: percentiles_ttft_ms sequence: sequence: float64 - name: request_goodput dtype: float64 - name: request_throughput dtype: float64 - name: std_e2el_ms dtype: float64 - name: std_itl_ms dtype: float64 - name: std_tpot_ms dtype: float64 - name: std_ttft_ms dtype: float64 - name: total_input dtype: int64 - name: total_output dtype: int64 - name: total_token_throughput dtype: float64 - name: device dtype: string splits: - name: train num_bytes: 9732 num_examples: 7 download_size: 56289 dataset_size: 9732 - config_name: 1xA100_EXPERTS_4_QUANT_SBMM_FORLOOP features: - name: server_args struct: - name: allow_toppings_registration dtype: bool - name: api_key dtype: 'null' - name: attention_backend dtype: string - name: base_min_new_token_ratio dtype: float64 - name: chat_template dtype: 'null' - name: chunked_prefill_size dtype: int64 - name: constrained_json_whitespace_pattern dtype: 'null' - name: context_length dtype: int64 - name: controller_port dtype: int64 - name: debug dtype: bool - name: decode_log_interval dtype: int64 - name: detokenizer_ipc_name dtype: string - name: detokenizer_port dtype: int64 - name: device dtype: string - name: disable_cuda_graph dtype: bool - name: disable_cuda_graph_padding dtype: bool - name: disable_custom_all_reduce dtype: bool - name: disable_disk_cache dtype: bool - name: disable_flashinfer dtype: bool - name: disable_flashinfer_sampling dtype: bool - name: disable_mla dtype: bool - name: disable_nan_detection dtype: bool - name: disable_penalizer dtype: bool - name: disable_radix_cache dtype: bool - name: disable_regex_jump_forward dtype: bool - name: dist_init_addr dtype: 'null' - name: dp_size dtype: int64 - name: dtype dtype: string - name: enable_double_sparsity dtype: bool - name: enable_dp_attention dtype: bool - name: enable_mixed_chunk dtype: bool - name: enable_overlap_schedule dtype: bool - name: enable_p2p_check dtype: bool - name: enable_stats_logging dtype: bool - name: enable_system_controller dtype: bool - name: enable_toppings dtype: bool - name: enable_torch_compile dtype: bool - name: flashinfer_workspace_size dtype: int64 - name: host dtype: string - name: init_new_token_ratio dtype: float64 - name: init_number_of_deltas dtype: int64 - name: init_number_of_loras dtype: int64 - name: init_toppings dtype: 'null' - name: is_embedding dtype: bool - name: kv_cache_dtype dtype: string - name: load_balance_method dtype: string - name: load_format dtype: string - name: log_requests dtype: bool - name: lora_paths dtype: 'null' - name: max_lora_ranks dtype: int64 - name: max_prefill_tokens dtype: int64 - name: max_running_requests dtype: 'null' - name: max_toppings_per_batch dtype: int64 - name: max_torch_compile_bs dtype: int64 - name: max_total_tokens dtype: 'null' - name: mem_fraction_static dtype: float64 - name: model_path dtype: string - name: nccl_ports sequence: string - name: new_token_ratio_decay dtype: float64 - name: nnodes dtype: int64 - name: node_rank dtype: int64 - name: num_continue_decode_steps dtype: int64 - name: num_continuous_decode_steps dtype: int64 - name: port dtype: int64 - name: quantization dtype: 'null' - name: random_seed dtype: int64 - name: retract_decode_steps dtype: int64 - name: sampling_backend dtype: string - name: schedule_conservativeness dtype: float64 - name: schedule_policy dtype: string - name: scheduler_input_ipc_name dtype: string - name: scheduler_port dtype: int64 - name: served_model_name dtype: string - name: server_id dtype: string - name: show_time_cost dtype: bool - name: skip_tokenizer_init dtype: bool - name: stream_interval dtype: int64 - name: tokenizer_ipc_name dtype: string - name: tokenizer_mode dtype: string - name: tokenizer_path dtype: string - name: tokenizer_port dtype: int64 - name: torchao_config dtype: string - name: tp_size dtype: int64 - name: triton_attention_reduce_in_fp32 dtype: bool - name: trust_remote_code dtype: bool - name: use_heterogeneous_pool dtype: bool - name: watchdog_timeout dtype: float64 - name: client_args struct: - name: dataset dtype: string - name: endpoint dtype: string - name: goodput dtype: 'null' - name: metric_percentiles dtype: string - name: model dtype: string - name: num_prompts dtype: int64 - name: output dtype: string - name: percentile_metrics dtype: string - name: request_rate dtype: float64 - name: tokenizer dtype: string - name: wait_until_ready dtype: bool - name: metrics struct: - name: completed dtype: int64 - name: mean_e2el_ms dtype: float64 - name: mean_itl_ms dtype: float64 - name: mean_tpot_ms dtype: float64 - name: mean_ttft_ms dtype: float64 - name: median_e2el_ms dtype: float64 - name: median_itl_ms dtype: float64 - name: median_tpot_ms dtype: float64 - name: median_ttft_ms dtype: float64 - name: output_throughput dtype: float64 - name: percentiles_e2el_ms sequence: sequence: float64 - name: percentiles_itl_ms sequence: sequence: float64 - name: percentiles_tpot_ms sequence: sequence: float64 - name: percentiles_ttft_ms sequence: sequence: float64 - name: request_goodput dtype: float64 - name: request_throughput dtype: float64 - name: std_e2el_ms dtype: float64 - name: std_itl_ms dtype: float64 - name: std_tpot_ms dtype: float64 - name: std_ttft_ms dtype: float64 - name: total_input dtype: int64 - name: total_output dtype: int64 - name: total_token_throughput dtype: float64 - name: device dtype: string splits: - name: train num_bytes: 9942 num_examples: 7 download_size: 56447 dataset_size: 9942 - config_name: 1xA100_EXPERTS_4_QUANT_SBMM_MULTILAUNCH features: - name: server_args struct: - name: allow_toppings_registration dtype: bool - name: api_key dtype: 'null' - name: attention_backend dtype: string - name: base_min_new_token_ratio dtype: float64 - name: chat_template dtype: 'null' - name: chunked_prefill_size dtype: int64 - name: constrained_json_whitespace_pattern dtype: 'null' - name: context_length dtype: int64 - name: controller_port dtype: int64 - name: debug dtype: bool - name: decode_log_interval dtype: int64 - name: detokenizer_ipc_name dtype: string - name: detokenizer_port dtype: int64 - name: device dtype: string - name: disable_cuda_graph dtype: bool - name: disable_cuda_graph_padding dtype: bool - name: disable_custom_all_reduce dtype: bool - name: disable_disk_cache dtype: bool - name: disable_flashinfer dtype: bool - name: disable_flashinfer_sampling dtype: bool - name: disable_mla dtype: bool - name: disable_nan_detection dtype: bool - name: disable_penalizer dtype: bool - name: disable_radix_cache dtype: bool - name: disable_regex_jump_forward dtype: bool - name: dist_init_addr dtype: 'null' - name: dp_size dtype: int64 - name: dtype dtype: string - name: enable_double_sparsity dtype: bool - name: enable_dp_attention dtype: bool - name: enable_mixed_chunk dtype: bool - name: enable_overlap_schedule dtype: bool - name: enable_p2p_check dtype: bool - name: enable_stats_logging dtype: bool - name: enable_system_controller dtype: bool - name: enable_toppings dtype: bool - name: enable_torch_compile dtype: bool - name: flashinfer_workspace_size dtype: int64 - name: host dtype: string - name: init_new_token_ratio dtype: float64 - name: init_number_of_deltas dtype: int64 - name: init_number_of_loras dtype: int64 - name: init_toppings dtype: 'null' - name: is_embedding dtype: bool - name: kv_cache_dtype dtype: string - name: load_balance_method dtype: string - name: load_format dtype: string - name: log_requests dtype: bool - name: lora_paths dtype: 'null' - name: max_lora_ranks dtype: int64 - name: max_prefill_tokens dtype: int64 - name: max_running_requests dtype: 'null' - name: max_toppings_per_batch dtype: int64 - name: max_torch_compile_bs dtype: int64 - name: max_total_tokens dtype: 'null' - name: mem_fraction_static dtype: float64 - name: model_path dtype: string - name: nccl_ports sequence: string - name: new_token_ratio_decay dtype: float64 - name: nnodes dtype: int64 - name: node_rank dtype: int64 - name: num_continue_decode_steps dtype: int64 - name: num_continuous_decode_steps dtype: int64 - name: port dtype: int64 - name: quantization dtype: 'null' - name: random_seed dtype: int64 - name: retract_decode_steps dtype: int64 - name: sampling_backend dtype: string - name: schedule_conservativeness dtype: float64 - name: schedule_policy dtype: string - name: scheduler_input_ipc_name dtype: string - name: scheduler_port dtype: int64 - name: served_model_name dtype: string - name: server_id dtype: string - name: show_time_cost dtype: bool - name: skip_tokenizer_init dtype: bool - name: stream_interval dtype: int64 - name: tokenizer_ipc_name dtype: string - name: tokenizer_mode dtype: string - name: tokenizer_path dtype: string - name: tokenizer_port dtype: int64 - name: torchao_config dtype: string - name: tp_size dtype: int64 - name: triton_attention_reduce_in_fp32 dtype: bool - name: trust_remote_code dtype: bool - name: use_heterogeneous_pool dtype: bool - name: watchdog_timeout dtype: float64 - name: client_args struct: - name: dataset dtype: string - name: endpoint dtype: string - name: goodput dtype: 'null' - name: metric_percentiles dtype: string - name: model dtype: string - name: num_prompts dtype: int64 - name: output dtype: string - name: percentile_metrics dtype: string - name: request_rate dtype: float64 - name: tokenizer dtype: string - name: wait_until_ready dtype: bool - name: metrics struct: - name: completed dtype: int64 - name: mean_e2el_ms dtype: float64 - name: mean_itl_ms dtype: float64 - name: mean_tpot_ms dtype: float64 - name: mean_ttft_ms dtype: float64 - name: median_e2el_ms dtype: float64 - name: median_itl_ms dtype: float64 - name: median_tpot_ms dtype: float64 - name: median_ttft_ms dtype: float64 - name: output_throughput dtype: float64 - name: percentiles_e2el_ms sequence: sequence: float64 - name: percentiles_itl_ms sequence: sequence: float64 - name: percentiles_tpot_ms sequence: sequence: float64 - name: percentiles_ttft_ms sequence: sequence: float64 - name: request_goodput dtype: float64 - name: request_throughput dtype: float64 - name: std_e2el_ms dtype: float64 - name: std_itl_ms dtype: float64 - name: std_tpot_ms dtype: float64 - name: std_ttft_ms dtype: float64 - name: total_input dtype: int64 - name: total_output dtype: int64 - name: total_token_throughput dtype: float64 - name: device dtype: string splits: - name: train num_bytes: 9998 num_examples: 7 download_size: 56493 dataset_size: 9998 - config_name: 1xA100_EXPERTS_4_QUANT_SBMM_NAIVE features: - name: server_args struct: - name: allow_toppings_registration dtype: bool - name: api_key dtype: 'null' - name: attention_backend dtype: string - name: base_min_new_token_ratio dtype: float64 - name: chat_template dtype: 'null' - name: chunked_prefill_size dtype: int64 - name: constrained_json_whitespace_pattern dtype: 'null' - name: context_length dtype: int64 - name: controller_port dtype: int64 - name: debug dtype: bool - name: decode_log_interval dtype: int64 - name: detokenizer_ipc_name dtype: string - name: detokenizer_port dtype: int64 - name: device dtype: string - name: disable_cuda_graph dtype: bool - name: disable_cuda_graph_padding dtype: bool - name: disable_custom_all_reduce dtype: bool - name: disable_disk_cache dtype: bool - name: disable_flashinfer dtype: bool - name: disable_flashinfer_sampling dtype: bool - name: disable_mla dtype: bool - name: disable_nan_detection dtype: bool - name: disable_penalizer dtype: bool - name: disable_radix_cache dtype: bool - name: disable_regex_jump_forward dtype: bool - name: dist_init_addr dtype: 'null' - name: dp_size dtype: int64 - name: dtype dtype: string - name: enable_double_sparsity dtype: bool - name: enable_dp_attention dtype: bool - name: enable_mixed_chunk dtype: bool - name: enable_overlap_schedule dtype: bool - name: enable_p2p_check dtype: bool - name: enable_stats_logging dtype: bool - name: enable_system_controller dtype: bool - name: enable_toppings dtype: bool - name: enable_torch_compile dtype: bool - name: flashinfer_workspace_size dtype: int64 - name: host dtype: string - name: init_new_token_ratio dtype: float64 - name: init_number_of_deltas dtype: int64 - name: init_number_of_loras dtype: int64 - name: init_toppings dtype: 'null' - name: is_embedding dtype: bool - name: kv_cache_dtype dtype: string - name: load_balance_method dtype: string - name: load_format dtype: string - name: log_requests dtype: bool - name: lora_paths dtype: 'null' - name: max_lora_ranks dtype: int64 - name: max_prefill_tokens dtype: int64 - name: max_running_requests dtype: 'null' - name: max_toppings_per_batch dtype: int64 - name: max_torch_compile_bs dtype: int64 - name: max_total_tokens dtype: 'null' - name: mem_fraction_static dtype: float64 - name: model_path dtype: string - name: nccl_ports sequence: string - name: new_token_ratio_decay dtype: float64 - name: nnodes dtype: int64 - name: node_rank dtype: int64 - name: num_continue_decode_steps dtype: int64 - name: num_continuous_decode_steps dtype: int64 - name: port dtype: int64 - name: quantization dtype: 'null' - name: random_seed dtype: int64 - name: retract_decode_steps dtype: int64 - name: sampling_backend dtype: string - name: schedule_conservativeness dtype: float64 - name: schedule_policy dtype: string - name: scheduler_input_ipc_name dtype: string - name: scheduler_port dtype: int64 - name: served_model_name dtype: string - name: server_id dtype: string - name: show_time_cost dtype: bool - name: skip_tokenizer_init dtype: bool - name: stream_interval dtype: int64 - name: tokenizer_ipc_name dtype: string - name: tokenizer_mode dtype: string - name: tokenizer_path dtype: string - name: tokenizer_port dtype: int64 - name: torchao_config dtype: string - name: tp_size dtype: int64 - name: triton_attention_reduce_in_fp32 dtype: bool - name: trust_remote_code dtype: bool - name: use_heterogeneous_pool dtype: bool - name: watchdog_timeout dtype: float64 - name: client_args struct: - name: dataset dtype: string - name: endpoint dtype: string - name: goodput dtype: 'null' - name: metric_percentiles dtype: string - name: model dtype: string - name: num_prompts dtype: int64 - name: output dtype: string - name: percentile_metrics dtype: string - name: request_rate dtype: float64 - name: tokenizer dtype: string - name: wait_until_ready dtype: bool - name: metrics struct: - name: completed dtype: int64 - name: mean_e2el_ms dtype: float64 - name: mean_itl_ms dtype: float64 - name: mean_tpot_ms dtype: float64 - name: mean_ttft_ms dtype: float64 - name: median_e2el_ms dtype: float64 - name: median_itl_ms dtype: float64 - name: median_tpot_ms dtype: float64 - name: median_ttft_ms dtype: float64 - name: output_throughput dtype: float64 - name: percentiles_e2el_ms sequence: sequence: float64 - name: percentiles_itl_ms sequence: sequence: float64 - name: percentiles_tpot_ms sequence: sequence: float64 - name: percentiles_ttft_ms sequence: sequence: float64 - name: request_goodput dtype: float64 - name: request_throughput dtype: float64 - name: std_e2el_ms dtype: float64 - name: std_itl_ms dtype: float64 - name: std_tpot_ms dtype: float64 - name: std_ttft_ms dtype: float64 - name: total_input dtype: int64 - name: total_output dtype: int64 - name: total_token_throughput dtype: float64 - name: device dtype: string splits: - name: train num_bytes: 9914 num_examples: 7 download_size: 56440 dataset_size: 9914 - config_name: 1xA100_EXPERTS_4_UNQUANT features: - name: server_args struct: - name: allow_toppings_registration dtype: bool - name: api_key dtype: 'null' - name: attention_backend dtype: string - name: base_min_new_token_ratio dtype: float64 - name: chat_template dtype: 'null' - name: chunked_prefill_size dtype: int64 - name: constrained_json_whitespace_pattern dtype: 'null' - name: context_length dtype: int64 - name: controller_port dtype: int64 - name: debug dtype: bool - name: decode_log_interval dtype: int64 - name: detokenizer_ipc_name dtype: string - name: detokenizer_port dtype: int64 - name: device dtype: string - name: disable_cuda_graph dtype: bool - name: disable_cuda_graph_padding dtype: bool - name: disable_custom_all_reduce dtype: bool - name: disable_disk_cache dtype: bool - name: disable_flashinfer dtype: bool - name: disable_flashinfer_sampling dtype: bool - name: disable_mla dtype: bool - name: disable_nan_detection dtype: bool - name: disable_penalizer dtype: bool - name: disable_radix_cache dtype: bool - name: disable_regex_jump_forward dtype: bool - name: dist_init_addr dtype: 'null' - name: dp_size dtype: int64 - name: dtype dtype: string - name: enable_double_sparsity dtype: bool - name: enable_dp_attention dtype: bool - name: enable_mixed_chunk dtype: bool - name: enable_overlap_schedule dtype: bool - name: enable_p2p_check dtype: bool - name: enable_stats_logging dtype: bool - name: enable_system_controller dtype: bool - name: enable_toppings dtype: bool - name: enable_torch_compile dtype: bool - name: flashinfer_workspace_size dtype: int64 - name: host dtype: string - name: init_new_token_ratio dtype: float64 - name: init_number_of_deltas dtype: int64 - name: init_number_of_loras dtype: int64 - name: init_toppings dtype: 'null' - name: is_embedding dtype: bool - name: kv_cache_dtype dtype: string - name: load_balance_method dtype: string - name: load_format dtype: string - name: log_requests dtype: bool - name: lora_paths dtype: 'null' - name: max_lora_ranks dtype: int64 - name: max_prefill_tokens dtype: int64 - name: max_running_requests dtype: 'null' - name: max_toppings_per_batch dtype: int64 - name: max_torch_compile_bs dtype: int64 - name: max_total_tokens dtype: 'null' - name: mem_fraction_static dtype: float64 - name: model_path dtype: string - name: nccl_ports sequence: string - name: new_token_ratio_decay dtype: float64 - name: nnodes dtype: int64 - name: node_rank dtype: int64 - name: num_continue_decode_steps dtype: int64 - name: num_continuous_decode_steps dtype: int64 - name: port dtype: int64 - name: quantization dtype: 'null' - name: random_seed dtype: int64 - name: retract_decode_steps dtype: int64 - name: sampling_backend dtype: string - name: schedule_conservativeness dtype: float64 - name: schedule_policy dtype: string - name: scheduler_input_ipc_name dtype: string - name: scheduler_port dtype: int64 - name: served_model_name dtype: string - name: server_id dtype: string - name: show_time_cost dtype: bool - name: skip_tokenizer_init dtype: bool - name: stream_interval dtype: int64 - name: tokenizer_ipc_name dtype: string - name: tokenizer_mode dtype: string - name: tokenizer_path dtype: string - name: tokenizer_port dtype: int64 - name: torchao_config dtype: string - name: tp_size dtype: int64 - name: triton_attention_reduce_in_fp32 dtype: bool - name: trust_remote_code dtype: bool - name: use_heterogeneous_pool dtype: bool - name: watchdog_timeout dtype: float64 - name: client_args struct: - name: dataset dtype: string - name: endpoint dtype: string - name: goodput dtype: 'null' - name: metric_percentiles dtype: string - name: model dtype: string - name: num_prompts dtype: int64 - name: output dtype: string - name: percentile_metrics dtype: string - name: request_rate dtype: float64 - name: tokenizer dtype: string - name: wait_until_ready dtype: bool - name: metrics struct: - name: completed dtype: int64 - name: mean_e2el_ms dtype: float64 - name: mean_itl_ms dtype: float64 - name: mean_tpot_ms dtype: float64 - name: mean_ttft_ms dtype: float64 - name: median_e2el_ms dtype: float64 - name: median_itl_ms dtype: float64 - name: median_tpot_ms dtype: float64 - name: median_ttft_ms dtype: float64 - name: output_throughput dtype: float64 - name: percentiles_e2el_ms sequence: sequence: float64 - name: percentiles_itl_ms sequence: sequence: float64 - name: percentiles_tpot_ms sequence: sequence: float64 - name: percentiles_ttft_ms sequence: sequence: float64 - name: request_goodput dtype: float64 - name: request_throughput dtype: float64 - name: std_e2el_ms dtype: float64 - name: std_itl_ms dtype: float64 - name: std_tpot_ms dtype: float64 - name: std_ttft_ms dtype: float64 - name: total_input dtype: int64 - name: total_output dtype: int64 - name: total_token_throughput dtype: float64 - name: device dtype: string splits: - name: train num_bytes: 8920 num_examples: 7 download_size: 55711 dataset_size: 8920 - config_name: 1xA100_EXPERTS_5_QUANT_NAIVE features: - name: server_args struct: - name: allow_toppings_registration dtype: bool - name: api_key dtype: 'null' - name: attention_backend dtype: string - name: base_min_new_token_ratio dtype: float64 - name: chat_template dtype: 'null' - name: chunked_prefill_size dtype: int64 - name: constrained_json_whitespace_pattern dtype: 'null' - name: context_length dtype: int64 - name: controller_port dtype: int64 - name: debug dtype: bool - name: decode_log_interval dtype: int64 - name: detokenizer_ipc_name dtype: string - name: detokenizer_port dtype: int64 - name: device dtype: string - name: disable_cuda_graph dtype: bool - name: disable_cuda_graph_padding dtype: bool - name: disable_custom_all_reduce dtype: bool - name: disable_disk_cache dtype: bool - name: disable_flashinfer dtype: bool - name: disable_flashinfer_sampling dtype: bool - name: disable_mla dtype: bool - name: disable_nan_detection dtype: bool - name: disable_penalizer dtype: bool - name: disable_radix_cache dtype: bool - name: disable_regex_jump_forward dtype: bool - name: dist_init_addr dtype: 'null' - name: dp_size dtype: int64 - name: dtype dtype: string - name: enable_double_sparsity dtype: bool - name: enable_dp_attention dtype: bool - name: enable_mixed_chunk dtype: bool - name: enable_overlap_schedule dtype: bool - name: enable_p2p_check dtype: bool - name: enable_stats_logging dtype: bool - name: enable_system_controller dtype: bool - name: enable_toppings dtype: bool - name: enable_torch_compile dtype: bool - name: flashinfer_workspace_size dtype: int64 - name: host dtype: string - name: init_new_token_ratio dtype: float64 - name: init_number_of_deltas dtype: int64 - name: init_number_of_loras dtype: int64 - name: init_toppings dtype: 'null' - name: is_embedding dtype: bool - name: kv_cache_dtype dtype: string - name: load_balance_method dtype: string - name: load_format dtype: string - name: log_requests dtype: bool - name: lora_paths dtype: 'null' - name: max_lora_ranks dtype: int64 - name: max_prefill_tokens dtype: int64 - name: max_running_requests dtype: 'null' - name: max_toppings_per_batch dtype: int64 - name: max_torch_compile_bs dtype: int64 - name: max_total_tokens dtype: 'null' - name: mem_fraction_static dtype: float64 - name: model_path dtype: string - name: nccl_ports sequence: string - name: new_token_ratio_decay dtype: float64 - name: nnodes dtype: int64 - name: node_rank dtype: int64 - name: num_continue_decode_steps dtype: int64 - name: num_continuous_decode_steps dtype: int64 - name: port dtype: int64 - name: quantization dtype: 'null' - name: random_seed dtype: int64 - name: retract_decode_steps dtype: int64 - name: sampling_backend dtype: string - name: schedule_conservativeness dtype: float64 - name: schedule_policy dtype: string - name: scheduler_input_ipc_name dtype: string - name: scheduler_port dtype: int64 - name: served_model_name dtype: string - name: server_id dtype: string - name: show_time_cost dtype: bool - name: skip_tokenizer_init dtype: bool - name: stream_interval dtype: int64 - name: tokenizer_ipc_name dtype: string - name: tokenizer_mode dtype: string - name: tokenizer_path dtype: string - name: tokenizer_port dtype: int64 - name: torchao_config dtype: string - name: tp_size dtype: int64 - name: triton_attention_reduce_in_fp32 dtype: bool - name: trust_remote_code dtype: bool - name: use_heterogeneous_pool dtype: bool - name: watchdog_timeout dtype: float64 - name: client_args struct: - name: dataset dtype: string - name: endpoint dtype: string - name: goodput dtype: 'null' - name: metric_percentiles dtype: string - name: model dtype: string - name: num_prompts dtype: int64 - name: output dtype: string - name: percentile_metrics dtype: string - name: request_rate dtype: float64 - name: tokenizer dtype: string - name: wait_until_ready dtype: bool - name: metrics struct: - name: completed dtype: int64 - name: mean_e2el_ms dtype: float64 - name: mean_itl_ms dtype: float64 - name: mean_tpot_ms dtype: float64 - name: mean_ttft_ms dtype: float64 - name: median_e2el_ms dtype: float64 - name: median_itl_ms dtype: float64 - name: median_tpot_ms dtype: float64 - name: median_ttft_ms dtype: float64 - name: output_throughput dtype: float64 - name: percentiles_e2el_ms sequence: sequence: float64 - name: percentiles_itl_ms sequence: sequence: float64 - name: percentiles_tpot_ms sequence: sequence: float64 - name: percentiles_ttft_ms sequence: sequence: float64 - name: request_goodput dtype: float64 - name: request_throughput dtype: float64 - name: std_e2el_ms dtype: float64 - name: std_itl_ms dtype: float64 - name: std_tpot_ms dtype: float64 - name: std_ttft_ms dtype: float64 - name: total_input dtype: int64 - name: total_output dtype: int64 - name: total_token_throughput dtype: float64 - name: device dtype: string splits: - name: train num_bytes: 9732 num_examples: 7 download_size: 56283 dataset_size: 9732 - config_name: 1xA100_EXPERTS_5_QUANT_SBMM_FORLOOP features: - name: server_args struct: - name: allow_toppings_registration dtype: bool - name: api_key dtype: 'null' - name: attention_backend dtype: string - name: base_min_new_token_ratio dtype: float64 - name: chat_template dtype: 'null' - name: chunked_prefill_size dtype: int64 - name: constrained_json_whitespace_pattern dtype: 'null' - name: context_length dtype: int64 - name: controller_port dtype: int64 - name: debug dtype: bool - name: decode_log_interval dtype: int64 - name: detokenizer_ipc_name dtype: string - name: detokenizer_port dtype: int64 - name: device dtype: string - name: disable_cuda_graph dtype: bool - name: disable_cuda_graph_padding dtype: bool - name: disable_custom_all_reduce dtype: bool - name: disable_disk_cache dtype: bool - name: disable_flashinfer dtype: bool - name: disable_flashinfer_sampling dtype: bool - name: disable_mla dtype: bool - name: disable_nan_detection dtype: bool - name: disable_penalizer dtype: bool - name: disable_radix_cache dtype: bool - name: disable_regex_jump_forward dtype: bool - name: dist_init_addr dtype: 'null' - name: dp_size dtype: int64 - name: dtype dtype: string - name: enable_double_sparsity dtype: bool - name: enable_dp_attention dtype: bool - name: enable_mixed_chunk dtype: bool - name: enable_overlap_schedule dtype: bool - name: enable_p2p_check dtype: bool - name: enable_stats_logging dtype: bool - name: enable_system_controller dtype: bool - name: enable_toppings dtype: bool - name: enable_torch_compile dtype: bool - name: flashinfer_workspace_size dtype: int64 - name: host dtype: string - name: init_new_token_ratio dtype: float64 - name: init_number_of_deltas dtype: int64 - name: init_number_of_loras dtype: int64 - name: init_toppings dtype: 'null' - name: is_embedding dtype: bool - name: kv_cache_dtype dtype: string - name: load_balance_method dtype: string - name: load_format dtype: string - name: log_requests dtype: bool - name: lora_paths dtype: 'null' - name: max_lora_ranks dtype: int64 - name: max_prefill_tokens dtype: int64 - name: max_running_requests dtype: 'null' - name: max_toppings_per_batch dtype: int64 - name: max_torch_compile_bs dtype: int64 - name: max_total_tokens dtype: 'null' - name: mem_fraction_static dtype: float64 - name: model_path dtype: string - name: nccl_ports sequence: string - name: new_token_ratio_decay dtype: float64 - name: nnodes dtype: int64 - name: node_rank dtype: int64 - name: num_continue_decode_steps dtype: int64 - name: num_continuous_decode_steps dtype: int64 - name: port dtype: int64 - name: quantization dtype: 'null' - name: random_seed dtype: int64 - name: retract_decode_steps dtype: int64 - name: sampling_backend dtype: string - name: schedule_conservativeness dtype: float64 - name: schedule_policy dtype: string - name: scheduler_input_ipc_name dtype: string - name: scheduler_port dtype: int64 - name: served_model_name dtype: string - name: server_id dtype: string - name: show_time_cost dtype: bool - name: skip_tokenizer_init dtype: bool - name: stream_interval dtype: int64 - name: tokenizer_ipc_name dtype: string - name: tokenizer_mode dtype: string - name: tokenizer_path dtype: string - name: tokenizer_port dtype: int64 - name: torchao_config dtype: string - name: tp_size dtype: int64 - name: triton_attention_reduce_in_fp32 dtype: bool - name: trust_remote_code dtype: bool - name: use_heterogeneous_pool dtype: bool - name: watchdog_timeout dtype: float64 - name: client_args struct: - name: dataset dtype: string - name: endpoint dtype: string - name: goodput dtype: 'null' - name: metric_percentiles dtype: string - name: model dtype: string - name: num_prompts dtype: int64 - name: output dtype: string - name: percentile_metrics dtype: string - name: request_rate dtype: float64 - name: tokenizer dtype: string - name: wait_until_ready dtype: bool - name: metrics struct: - name: completed dtype: int64 - name: mean_e2el_ms dtype: float64 - name: mean_itl_ms dtype: float64 - name: mean_tpot_ms dtype: float64 - name: mean_ttft_ms dtype: float64 - name: median_e2el_ms dtype: float64 - name: median_itl_ms dtype: float64 - name: median_tpot_ms dtype: float64 - name: median_ttft_ms dtype: float64 - name: output_throughput dtype: float64 - name: percentiles_e2el_ms sequence: sequence: float64 - name: percentiles_itl_ms sequence: sequence: float64 - name: percentiles_tpot_ms sequence: sequence: float64 - name: percentiles_ttft_ms sequence: sequence: float64 - name: request_goodput dtype: float64 - name: request_throughput dtype: float64 - name: std_e2el_ms dtype: float64 - name: std_itl_ms dtype: float64 - name: std_tpot_ms dtype: float64 - name: std_ttft_ms dtype: float64 - name: total_input dtype: int64 - name: total_output dtype: int64 - name: total_token_throughput dtype: float64 - name: device dtype: string splits: - name: train num_bytes: 9942 num_examples: 7 download_size: 56451 dataset_size: 9942 - config_name: 1xA100_EXPERTS_5_QUANT_SBMM_MULTILAUNCH features: - name: server_args struct: - name: allow_toppings_registration dtype: bool - name: api_key dtype: 'null' - name: attention_backend dtype: string - name: base_min_new_token_ratio dtype: float64 - name: chat_template dtype: 'null' - name: chunked_prefill_size dtype: int64 - name: constrained_json_whitespace_pattern dtype: 'null' - name: context_length dtype: int64 - name: controller_port dtype: int64 - name: debug dtype: bool - name: decode_log_interval dtype: int64 - name: detokenizer_ipc_name dtype: string - name: detokenizer_port dtype: int64 - name: device dtype: string - name: disable_cuda_graph dtype: bool - name: disable_cuda_graph_padding dtype: bool - name: disable_custom_all_reduce dtype: bool - name: disable_disk_cache dtype: bool - name: disable_flashinfer dtype: bool - name: disable_flashinfer_sampling dtype: bool - name: disable_mla dtype: bool - name: disable_nan_detection dtype: bool - name: disable_penalizer dtype: bool - name: disable_radix_cache dtype: bool - name: disable_regex_jump_forward dtype: bool - name: dist_init_addr dtype: 'null' - name: dp_size dtype: int64 - name: dtype dtype: string - name: enable_double_sparsity dtype: bool - name: enable_dp_attention dtype: bool - name: enable_mixed_chunk dtype: bool - name: enable_overlap_schedule dtype: bool - name: enable_p2p_check dtype: bool - name: enable_stats_logging dtype: bool - name: enable_system_controller dtype: bool - name: enable_toppings dtype: bool - name: enable_torch_compile dtype: bool - name: flashinfer_workspace_size dtype: int64 - name: host dtype: string - name: init_new_token_ratio dtype: float64 - name: init_number_of_deltas dtype: int64 - name: init_number_of_loras dtype: int64 - name: init_toppings dtype: 'null' - name: is_embedding dtype: bool - name: kv_cache_dtype dtype: string - name: load_balance_method dtype: string - name: load_format dtype: string - name: log_requests dtype: bool - name: lora_paths dtype: 'null' - name: max_lora_ranks dtype: int64 - name: max_prefill_tokens dtype: int64 - name: max_running_requests dtype: 'null' - name: max_toppings_per_batch dtype: int64 - name: max_torch_compile_bs dtype: int64 - name: max_total_tokens dtype: 'null' - name: mem_fraction_static dtype: float64 - name: model_path dtype: string - name: nccl_ports sequence: string - name: new_token_ratio_decay dtype: float64 - name: nnodes dtype: int64 - name: node_rank dtype: int64 - name: num_continue_decode_steps dtype: int64 - name: num_continuous_decode_steps dtype: int64 - name: port dtype: int64 - name: quantization dtype: 'null' - name: random_seed dtype: int64 - name: retract_decode_steps dtype: int64 - name: sampling_backend dtype: string - name: schedule_conservativeness dtype: float64 - name: schedule_policy dtype: string - name: scheduler_input_ipc_name dtype: string - name: scheduler_port dtype: int64 - name: served_model_name dtype: string - name: server_id dtype: string - name: show_time_cost dtype: bool - name: skip_tokenizer_init dtype: bool - name: stream_interval dtype: int64 - name: tokenizer_ipc_name dtype: string - name: tokenizer_mode dtype: string - name: tokenizer_path dtype: string - name: tokenizer_port dtype: int64 - name: torchao_config dtype: string - name: tp_size dtype: int64 - name: triton_attention_reduce_in_fp32 dtype: bool - name: trust_remote_code dtype: bool - name: use_heterogeneous_pool dtype: bool - name: watchdog_timeout dtype: float64 - name: client_args struct: - name: dataset dtype: string - name: endpoint dtype: string - name: goodput dtype: 'null' - name: metric_percentiles dtype: string - name: model dtype: string - name: num_prompts dtype: int64 - name: output dtype: string - name: percentile_metrics dtype: string - name: request_rate dtype: float64 - name: tokenizer dtype: string - name: wait_until_ready dtype: bool - name: metrics struct: - name: completed dtype: int64 - name: mean_e2el_ms dtype: float64 - name: mean_itl_ms dtype: float64 - name: mean_tpot_ms dtype: float64 - name: mean_ttft_ms dtype: float64 - name: median_e2el_ms dtype: float64 - name: median_itl_ms dtype: float64 - name: median_tpot_ms dtype: float64 - name: median_ttft_ms dtype: float64 - name: output_throughput dtype: float64 - name: percentiles_e2el_ms sequence: sequence: float64 - name: percentiles_itl_ms sequence: sequence: float64 - name: percentiles_tpot_ms sequence: sequence: float64 - name: percentiles_ttft_ms sequence: sequence: float64 - name: request_goodput dtype: float64 - name: request_throughput dtype: float64 - name: std_e2el_ms dtype: float64 - name: std_itl_ms dtype: float64 - name: std_tpot_ms dtype: float64 - name: std_ttft_ms dtype: float64 - name: total_input dtype: int64 - name: total_output dtype: int64 - name: total_token_throughput dtype: float64 - name: device dtype: string splits: - name: train num_bytes: 9998 num_examples: 7 download_size: 56508 dataset_size: 9998 - config_name: 1xA100_EXPERTS_5_QUANT_SBMM_NAIVE features: - name: server_args struct: - name: allow_toppings_registration dtype: bool - name: api_key dtype: 'null' - name: attention_backend dtype: string - name: base_min_new_token_ratio dtype: float64 - name: chat_template dtype: 'null' - name: chunked_prefill_size dtype: int64 - name: constrained_json_whitespace_pattern dtype: 'null' - name: context_length dtype: int64 - name: controller_port dtype: int64 - name: debug dtype: bool - name: decode_log_interval dtype: int64 - name: detokenizer_ipc_name dtype: string - name: detokenizer_port dtype: int64 - name: device dtype: string - name: disable_cuda_graph dtype: bool - name: disable_cuda_graph_padding dtype: bool - name: disable_custom_all_reduce dtype: bool - name: disable_disk_cache dtype: bool - name: disable_flashinfer dtype: bool - name: disable_flashinfer_sampling dtype: bool - name: disable_mla dtype: bool - name: disable_nan_detection dtype: bool - name: disable_penalizer dtype: bool - name: disable_radix_cache dtype: bool - name: disable_regex_jump_forward dtype: bool - name: dist_init_addr dtype: 'null' - name: dp_size dtype: int64 - name: dtype dtype: string - name: enable_double_sparsity dtype: bool - name: enable_dp_attention dtype: bool - name: enable_mixed_chunk dtype: bool - name: enable_overlap_schedule dtype: bool - name: enable_p2p_check dtype: bool - name: enable_stats_logging dtype: bool - name: enable_system_controller dtype: bool - name: enable_toppings dtype: bool - name: enable_torch_compile dtype: bool - name: flashinfer_workspace_size dtype: int64 - name: host dtype: string - name: init_new_token_ratio dtype: float64 - name: init_number_of_deltas dtype: int64 - name: init_number_of_loras dtype: int64 - name: init_toppings dtype: 'null' - name: is_embedding dtype: bool - name: kv_cache_dtype dtype: string - name: load_balance_method dtype: string - name: load_format dtype: string - name: log_requests dtype: bool - name: lora_paths dtype: 'null' - name: max_lora_ranks dtype: int64 - name: max_prefill_tokens dtype: int64 - name: max_running_requests dtype: 'null' - name: max_toppings_per_batch dtype: int64 - name: max_torch_compile_bs dtype: int64 - name: max_total_tokens dtype: 'null' - name: mem_fraction_static dtype: float64 - name: model_path dtype: string - name: nccl_ports sequence: string - name: new_token_ratio_decay dtype: float64 - name: nnodes dtype: int64 - name: node_rank dtype: int64 - name: num_continue_decode_steps dtype: int64 - name: num_continuous_decode_steps dtype: int64 - name: port dtype: int64 - name: quantization dtype: 'null' - name: random_seed dtype: int64 - name: retract_decode_steps dtype: int64 - name: sampling_backend dtype: string - name: schedule_conservativeness dtype: float64 - name: schedule_policy dtype: string - name: scheduler_input_ipc_name dtype: string - name: scheduler_port dtype: int64 - name: served_model_name dtype: string - name: server_id dtype: string - name: show_time_cost dtype: bool - name: skip_tokenizer_init dtype: bool - name: stream_interval dtype: int64 - name: tokenizer_ipc_name dtype: string - name: tokenizer_mode dtype: string - name: tokenizer_path dtype: string - name: tokenizer_port dtype: int64 - name: torchao_config dtype: string - name: tp_size dtype: int64 - name: triton_attention_reduce_in_fp32 dtype: bool - name: trust_remote_code dtype: bool - name: use_heterogeneous_pool dtype: bool - name: watchdog_timeout dtype: float64 - name: client_args struct: - name: dataset dtype: string - name: endpoint dtype: string - name: goodput dtype: 'null' - name: metric_percentiles dtype: string - name: model dtype: string - name: num_prompts dtype: int64 - name: output dtype: string - name: percentile_metrics dtype: string - name: request_rate dtype: float64 - name: tokenizer dtype: string - name: wait_until_ready dtype: bool - name: metrics struct: - name: completed dtype: int64 - name: mean_e2el_ms dtype: float64 - name: mean_itl_ms dtype: float64 - name: mean_tpot_ms dtype: float64 - name: mean_ttft_ms dtype: float64 - name: median_e2el_ms dtype: float64 - name: median_itl_ms dtype: float64 - name: median_tpot_ms dtype: float64 - name: median_ttft_ms dtype: float64 - name: output_throughput dtype: float64 - name: percentiles_e2el_ms sequence: sequence: float64 - name: percentiles_itl_ms sequence: sequence: float64 - name: percentiles_tpot_ms sequence: sequence: float64 - name: percentiles_ttft_ms sequence: sequence: float64 - name: request_goodput dtype: float64 - name: request_throughput dtype: float64 - name: std_e2el_ms dtype: float64 - name: std_itl_ms dtype: float64 - name: std_tpot_ms dtype: float64 - name: std_ttft_ms dtype: float64 - name: total_input dtype: int64 - name: total_output dtype: int64 - name: total_token_throughput dtype: float64 - name: device dtype: string splits: - name: train num_bytes: 9942 num_examples: 7 download_size: 56451 dataset_size: 9942 - config_name: 1xA100_EXPERTS_5_UNQUANT features: - name: server_args struct: - name: allow_toppings_registration dtype: bool - name: api_key dtype: 'null' - name: attention_backend dtype: string - name: base_min_new_token_ratio dtype: float64 - name: chat_template dtype: 'null' - name: chunked_prefill_size dtype: int64 - name: constrained_json_whitespace_pattern dtype: 'null' - name: context_length dtype: int64 - name: controller_port dtype: int64 - name: debug dtype: bool - name: decode_log_interval dtype: int64 - name: detokenizer_ipc_name dtype: string - name: detokenizer_port dtype: int64 - name: device dtype: string - name: disable_cuda_graph dtype: bool - name: disable_cuda_graph_padding dtype: bool - name: disable_custom_all_reduce dtype: bool - name: disable_disk_cache dtype: bool - name: disable_flashinfer dtype: bool - name: disable_flashinfer_sampling dtype: bool - name: disable_mla dtype: bool - name: disable_nan_detection dtype: bool - name: disable_penalizer dtype: bool - name: disable_radix_cache dtype: bool - name: disable_regex_jump_forward dtype: bool - name: dist_init_addr dtype: 'null' - name: dp_size dtype: int64 - name: dtype dtype: string - name: enable_double_sparsity dtype: bool - name: enable_dp_attention dtype: bool - name: enable_mixed_chunk dtype: bool - name: enable_overlap_schedule dtype: bool - name: enable_p2p_check dtype: bool - name: enable_stats_logging dtype: bool - name: enable_system_controller dtype: bool - name: enable_toppings dtype: bool - name: enable_torch_compile dtype: bool - name: flashinfer_workspace_size dtype: int64 - name: host dtype: string - name: init_new_token_ratio dtype: float64 - name: init_number_of_deltas dtype: int64 - name: init_number_of_loras dtype: int64 - name: init_toppings dtype: 'null' - name: is_embedding dtype: bool - name: kv_cache_dtype dtype: string - name: load_balance_method dtype: string - name: load_format dtype: string - name: log_requests dtype: bool - name: lora_paths dtype: 'null' - name: max_lora_ranks dtype: int64 - name: max_prefill_tokens dtype: int64 - name: max_running_requests dtype: 'null' - name: max_toppings_per_batch dtype: int64 - name: max_torch_compile_bs dtype: int64 - name: max_total_tokens dtype: 'null' - name: mem_fraction_static dtype: float64 - name: model_path dtype: string - name: nccl_ports sequence: string - name: new_token_ratio_decay dtype: float64 - name: nnodes dtype: int64 - name: node_rank dtype: int64 - name: num_continue_decode_steps dtype: int64 - name: num_continuous_decode_steps dtype: int64 - name: port dtype: int64 - name: quantization dtype: 'null' - name: random_seed dtype: int64 - name: retract_decode_steps dtype: int64 - name: sampling_backend dtype: string - name: schedule_conservativeness dtype: float64 - name: schedule_policy dtype: string - name: scheduler_input_ipc_name dtype: string - name: scheduler_port dtype: int64 - name: served_model_name dtype: string - name: server_id dtype: string - name: show_time_cost dtype: bool - name: skip_tokenizer_init dtype: bool - name: stream_interval dtype: int64 - name: tokenizer_ipc_name dtype: string - name: tokenizer_mode dtype: string - name: tokenizer_path dtype: string - name: tokenizer_port dtype: int64 - name: torchao_config dtype: string - name: tp_size dtype: int64 - name: triton_attention_reduce_in_fp32 dtype: bool - name: trust_remote_code dtype: bool - name: use_heterogeneous_pool dtype: bool - name: watchdog_timeout dtype: float64 - name: client_args struct: - name: dataset dtype: string - name: endpoint dtype: string - name: goodput dtype: 'null' - name: metric_percentiles dtype: string - name: model dtype: string - name: num_prompts dtype: int64 - name: output dtype: string - name: percentile_metrics dtype: string - name: request_rate dtype: float64 - name: tokenizer dtype: string - name: wait_until_ready dtype: bool - name: metrics struct: - name: completed dtype: int64 - name: mean_e2el_ms dtype: float64 - name: mean_itl_ms dtype: float64 - name: mean_tpot_ms dtype: float64 - name: mean_ttft_ms dtype: float64 - name: median_e2el_ms dtype: float64 - name: median_itl_ms dtype: float64 - name: median_tpot_ms dtype: float64 - name: median_ttft_ms dtype: float64 - name: output_throughput dtype: float64 - name: percentiles_e2el_ms sequence: sequence: float64 - name: percentiles_itl_ms sequence: sequence: float64 - name: percentiles_tpot_ms sequence: sequence: float64 - name: percentiles_ttft_ms sequence: sequence: float64 - name: request_goodput dtype: float64 - name: request_throughput dtype: float64 - name: std_e2el_ms dtype: float64 - name: std_itl_ms dtype: float64 - name: std_tpot_ms dtype: float64 - name: std_ttft_ms dtype: float64 - name: total_input dtype: int64 - name: total_output dtype: int64 - name: total_token_throughput dtype: float64 - name: device dtype: string splits: - name: train num_bytes: 8920 num_examples: 7 download_size: 55721 dataset_size: 8920 - config_name: 1xA100_EXPERTS_6_QUANT_NAIVE features: - name: server_args struct: - name: allow_toppings_registration dtype: bool - name: api_key dtype: 'null' - name: attention_backend dtype: string - name: base_min_new_token_ratio dtype: float64 - name: chat_template dtype: 'null' - name: chunked_prefill_size dtype: int64 - name: constrained_json_whitespace_pattern dtype: 'null' - name: context_length dtype: int64 - name: controller_port dtype: int64 - name: debug dtype: bool - name: decode_log_interval dtype: int64 - name: detokenizer_ipc_name dtype: string - name: detokenizer_port dtype: int64 - name: device dtype: string - name: disable_cuda_graph dtype: bool - name: disable_cuda_graph_padding dtype: bool - name: disable_custom_all_reduce dtype: bool - name: disable_disk_cache dtype: bool - name: disable_flashinfer dtype: bool - name: disable_flashinfer_sampling dtype: bool - name: disable_mla dtype: bool - name: disable_nan_detection dtype: bool - name: disable_penalizer dtype: bool - name: disable_radix_cache dtype: bool - name: disable_regex_jump_forward dtype: bool - name: dist_init_addr dtype: 'null' - name: dp_size dtype: int64 - name: dtype dtype: string - name: enable_double_sparsity dtype: bool - name: enable_dp_attention dtype: bool - name: enable_mixed_chunk dtype: bool - name: enable_overlap_schedule dtype: bool - name: enable_p2p_check dtype: bool - name: enable_stats_logging dtype: bool - name: enable_system_controller dtype: bool - name: enable_toppings dtype: bool - name: enable_torch_compile dtype: bool - name: flashinfer_workspace_size dtype: int64 - name: host dtype: string - name: init_new_token_ratio dtype: float64 - name: init_number_of_deltas dtype: int64 - name: init_number_of_loras dtype: int64 - name: init_toppings dtype: 'null' - name: is_embedding dtype: bool - name: kv_cache_dtype dtype: string - name: load_balance_method dtype: string - name: load_format dtype: string - name: log_requests dtype: bool - name: lora_paths dtype: 'null' - name: max_lora_ranks dtype: int64 - name: max_prefill_tokens dtype: int64 - name: max_running_requests dtype: 'null' - name: max_toppings_per_batch dtype: int64 - name: max_torch_compile_bs dtype: int64 - name: max_total_tokens dtype: 'null' - name: mem_fraction_static dtype: float64 - name: model_path dtype: string - name: nccl_ports sequence: string - name: new_token_ratio_decay dtype: float64 - name: nnodes dtype: int64 - name: node_rank dtype: int64 - name: num_continue_decode_steps dtype: int64 - name: num_continuous_decode_steps dtype: int64 - name: port dtype: int64 - name: quantization dtype: 'null' - name: random_seed dtype: int64 - name: retract_decode_steps dtype: int64 - name: sampling_backend dtype: string - name: schedule_conservativeness dtype: float64 - name: schedule_policy dtype: string - name: scheduler_input_ipc_name dtype: string - name: scheduler_port dtype: int64 - name: served_model_name dtype: string - name: server_id dtype: string - name: show_time_cost dtype: bool - name: skip_tokenizer_init dtype: bool - name: stream_interval dtype: int64 - name: tokenizer_ipc_name dtype: string - name: tokenizer_mode dtype: string - name: tokenizer_path dtype: string - name: tokenizer_port dtype: int64 - name: torchao_config dtype: string - name: tp_size dtype: int64 - name: triton_attention_reduce_in_fp32 dtype: bool - name: trust_remote_code dtype: bool - name: use_heterogeneous_pool dtype: bool - name: watchdog_timeout dtype: float64 - name: client_args struct: - name: dataset dtype: string - name: endpoint dtype: string - name: goodput dtype: 'null' - name: metric_percentiles dtype: string - name: model dtype: string - name: num_prompts dtype: int64 - name: output dtype: string - name: percentile_metrics dtype: string - name: request_rate dtype: float64 - name: tokenizer dtype: string - name: wait_until_ready dtype: bool - name: metrics struct: - name: completed dtype: int64 - name: mean_e2el_ms dtype: float64 - name: mean_itl_ms dtype: float64 - name: mean_tpot_ms dtype: float64 - name: mean_ttft_ms dtype: float64 - name: median_e2el_ms dtype: float64 - name: median_itl_ms dtype: float64 - name: median_tpot_ms dtype: float64 - name: median_ttft_ms dtype: float64 - name: output_throughput dtype: float64 - name: percentiles_e2el_ms sequence: sequence: float64 - name: percentiles_itl_ms sequence: sequence: float64 - name: percentiles_tpot_ms sequence: sequence: float64 - name: percentiles_ttft_ms sequence: sequence: float64 - name: request_goodput dtype: float64 - name: request_throughput dtype: float64 - name: std_e2el_ms dtype: float64 - name: std_itl_ms dtype: float64 - name: std_tpot_ms dtype: float64 - name: std_ttft_ms dtype: float64 - name: total_input dtype: int64 - name: total_output dtype: int64 - name: total_token_throughput dtype: float64 - name: device dtype: string splits: - name: train num_bytes: 9732 num_examples: 7 download_size: 56293 dataset_size: 9732 - config_name: 1xA100_EXPERTS_6_QUANT_SBMM_FORLOOP features: - name: server_args struct: - name: allow_toppings_registration dtype: bool - name: api_key dtype: 'null' - name: attention_backend dtype: string - name: base_min_new_token_ratio dtype: float64 - name: chat_template dtype: 'null' - name: chunked_prefill_size dtype: int64 - name: constrained_json_whitespace_pattern dtype: 'null' - name: context_length dtype: int64 - name: controller_port dtype: int64 - name: debug dtype: bool - name: decode_log_interval dtype: int64 - name: detokenizer_ipc_name dtype: string - name: detokenizer_port dtype: int64 - name: device dtype: string - name: disable_cuda_graph dtype: bool - name: disable_cuda_graph_padding dtype: bool - name: disable_custom_all_reduce dtype: bool - name: disable_disk_cache dtype: bool - name: disable_flashinfer dtype: bool - name: disable_flashinfer_sampling dtype: bool - name: disable_mla dtype: bool - name: disable_nan_detection dtype: bool - name: disable_penalizer dtype: bool - name: disable_radix_cache dtype: bool - name: disable_regex_jump_forward dtype: bool - name: dist_init_addr dtype: 'null' - name: dp_size dtype: int64 - name: dtype dtype: string - name: enable_double_sparsity dtype: bool - name: enable_dp_attention dtype: bool - name: enable_mixed_chunk dtype: bool - name: enable_overlap_schedule dtype: bool - name: enable_p2p_check dtype: bool - name: enable_stats_logging dtype: bool - name: enable_system_controller dtype: bool - name: enable_toppings dtype: bool - name: enable_torch_compile dtype: bool - name: flashinfer_workspace_size dtype: int64 - name: host dtype: string - name: init_new_token_ratio dtype: float64 - name: init_number_of_deltas dtype: int64 - name: init_number_of_loras dtype: int64 - name: init_toppings dtype: 'null' - name: is_embedding dtype: bool - name: kv_cache_dtype dtype: string - name: load_balance_method dtype: string - name: load_format dtype: string - name: log_requests dtype: bool - name: lora_paths dtype: 'null' - name: max_lora_ranks dtype: int64 - name: max_prefill_tokens dtype: int64 - name: max_running_requests dtype: 'null' - name: max_toppings_per_batch dtype: int64 - name: max_torch_compile_bs dtype: int64 - name: max_total_tokens dtype: 'null' - name: mem_fraction_static dtype: float64 - name: model_path dtype: string - name: nccl_ports sequence: string - name: new_token_ratio_decay dtype: float64 - name: nnodes dtype: int64 - name: node_rank dtype: int64 - name: num_continue_decode_steps dtype: int64 - name: num_continuous_decode_steps dtype: int64 - name: port dtype: int64 - name: quantization dtype: 'null' - name: random_seed dtype: int64 - name: retract_decode_steps dtype: int64 - name: sampling_backend dtype: string - name: schedule_conservativeness dtype: float64 - name: schedule_policy dtype: string - name: scheduler_input_ipc_name dtype: string - name: scheduler_port dtype: int64 - name: served_model_name dtype: string - name: server_id dtype: string - name: show_time_cost dtype: bool - name: skip_tokenizer_init dtype: bool - name: stream_interval dtype: int64 - name: tokenizer_ipc_name dtype: string - name: tokenizer_mode dtype: string - name: tokenizer_path dtype: string - name: tokenizer_port dtype: int64 - name: torchao_config dtype: string - name: tp_size dtype: int64 - name: triton_attention_reduce_in_fp32 dtype: bool - name: trust_remote_code dtype: bool - name: use_heterogeneous_pool dtype: bool - name: watchdog_timeout dtype: float64 - name: client_args struct: - name: dataset dtype: string - name: endpoint dtype: string - name: goodput dtype: 'null' - name: metric_percentiles dtype: string - name: model dtype: string - name: num_prompts dtype: int64 - name: output dtype: string - name: percentile_metrics dtype: string - name: request_rate dtype: float64 - name: tokenizer dtype: string - name: wait_until_ready dtype: bool - name: metrics struct: - name: completed dtype: int64 - name: mean_e2el_ms dtype: float64 - name: mean_itl_ms dtype: float64 - name: mean_tpot_ms dtype: float64 - name: mean_ttft_ms dtype: float64 - name: median_e2el_ms dtype: float64 - name: median_itl_ms dtype: float64 - name: median_tpot_ms dtype: float64 - name: median_ttft_ms dtype: float64 - name: output_throughput dtype: float64 - name: percentiles_e2el_ms sequence: sequence: float64 - name: percentiles_itl_ms sequence: sequence: float64 - name: percentiles_tpot_ms sequence: sequence: float64 - name: percentiles_ttft_ms sequence: sequence: float64 - name: request_goodput dtype: float64 - name: request_throughput dtype: float64 - name: std_e2el_ms dtype: float64 - name: std_itl_ms dtype: float64 - name: std_tpot_ms dtype: float64 - name: std_ttft_ms dtype: float64 - name: total_input dtype: int64 - name: total_output dtype: int64 - name: total_token_throughput dtype: float64 - name: device dtype: string splits: - name: train num_bytes: 9942 num_examples: 7 download_size: 56450 dataset_size: 9942 - config_name: 1xA100_EXPERTS_6_QUANT_SBMM_MULTILAUNCH features: - name: server_args struct: - name: allow_toppings_registration dtype: bool - name: api_key dtype: 'null' - name: attention_backend dtype: string - name: base_min_new_token_ratio dtype: float64 - name: chat_template dtype: 'null' - name: chunked_prefill_size dtype: int64 - name: constrained_json_whitespace_pattern dtype: 'null' - name: context_length dtype: int64 - name: controller_port dtype: int64 - name: debug dtype: bool - name: decode_log_interval dtype: int64 - name: detokenizer_ipc_name dtype: string - name: detokenizer_port dtype: int64 - name: device dtype: string - name: disable_cuda_graph dtype: bool - name: disable_cuda_graph_padding dtype: bool - name: disable_custom_all_reduce dtype: bool - name: disable_disk_cache dtype: bool - name: disable_flashinfer dtype: bool - name: disable_flashinfer_sampling dtype: bool - name: disable_mla dtype: bool - name: disable_nan_detection dtype: bool - name: disable_penalizer dtype: bool - name: disable_radix_cache dtype: bool - name: disable_regex_jump_forward dtype: bool - name: dist_init_addr dtype: 'null' - name: dp_size dtype: int64 - name: dtype dtype: string - name: enable_double_sparsity dtype: bool - name: enable_dp_attention dtype: bool - name: enable_mixed_chunk dtype: bool - name: enable_overlap_schedule dtype: bool - name: enable_p2p_check dtype: bool - name: enable_stats_logging dtype: bool - name: enable_system_controller dtype: bool - name: enable_toppings dtype: bool - name: enable_torch_compile dtype: bool - name: flashinfer_workspace_size dtype: int64 - name: host dtype: string - name: init_new_token_ratio dtype: float64 - name: init_number_of_deltas dtype: int64 - name: init_number_of_loras dtype: int64 - name: init_toppings dtype: 'null' - name: is_embedding dtype: bool - name: kv_cache_dtype dtype: string - name: load_balance_method dtype: string - name: load_format dtype: string - name: log_requests dtype: bool - name: lora_paths dtype: 'null' - name: max_lora_ranks dtype: int64 - name: max_prefill_tokens dtype: int64 - name: max_running_requests dtype: 'null' - name: max_toppings_per_batch dtype: int64 - name: max_torch_compile_bs dtype: int64 - name: max_total_tokens dtype: 'null' - name: mem_fraction_static dtype: float64 - name: model_path dtype: string - name: nccl_ports sequence: string - name: new_token_ratio_decay dtype: float64 - name: nnodes dtype: int64 - name: node_rank dtype: int64 - name: num_continue_decode_steps dtype: int64 - name: num_continuous_decode_steps dtype: int64 - name: port dtype: int64 - name: quantization dtype: 'null' - name: random_seed dtype: int64 - name: retract_decode_steps dtype: int64 - name: sampling_backend dtype: string - name: schedule_conservativeness dtype: float64 - name: schedule_policy dtype: string - name: scheduler_input_ipc_name dtype: string - name: scheduler_port dtype: int64 - name: served_model_name dtype: string - name: server_id dtype: string - name: show_time_cost dtype: bool - name: skip_tokenizer_init dtype: bool - name: stream_interval dtype: int64 - name: tokenizer_ipc_name dtype: string - name: tokenizer_mode dtype: string - name: tokenizer_path dtype: string - name: tokenizer_port dtype: int64 - name: torchao_config dtype: string - name: tp_size dtype: int64 - name: triton_attention_reduce_in_fp32 dtype: bool - name: trust_remote_code dtype: bool - name: use_heterogeneous_pool dtype: bool - name: watchdog_timeout dtype: float64 - name: client_args struct: - name: dataset dtype: string - name: endpoint dtype: string - name: goodput dtype: 'null' - name: metric_percentiles dtype: string - name: model dtype: string - name: num_prompts dtype: int64 - name: output dtype: string - name: percentile_metrics dtype: string - name: request_rate dtype: float64 - name: tokenizer dtype: string - name: wait_until_ready dtype: bool - name: metrics struct: - name: completed dtype: int64 - name: mean_e2el_ms dtype: float64 - name: mean_itl_ms dtype: float64 - name: mean_tpot_ms dtype: float64 - name: mean_ttft_ms dtype: float64 - name: median_e2el_ms dtype: float64 - name: median_itl_ms dtype: float64 - name: median_tpot_ms dtype: float64 - name: median_ttft_ms dtype: float64 - name: output_throughput dtype: float64 - name: percentiles_e2el_ms sequence: sequence: float64 - name: percentiles_itl_ms sequence: sequence: float64 - name: percentiles_tpot_ms sequence: sequence: float64 - name: percentiles_ttft_ms sequence: sequence: float64 - name: request_goodput dtype: float64 - name: request_throughput dtype: float64 - name: std_e2el_ms dtype: float64 - name: std_itl_ms dtype: float64 - name: std_tpot_ms dtype: float64 - name: std_ttft_ms dtype: float64 - name: total_input dtype: int64 - name: total_output dtype: int64 - name: total_token_throughput dtype: float64 - name: device dtype: string splits: - name: train num_bytes: 9998 num_examples: 7 download_size: 56502 dataset_size: 9998 - config_name: 1xA100_EXPERTS_6_QUANT_SBMM_NAIVE features: - name: server_args struct: - name: allow_toppings_registration dtype: bool - name: api_key dtype: 'null' - name: attention_backend dtype: string - name: base_min_new_token_ratio dtype: float64 - name: chat_template dtype: 'null' - name: chunked_prefill_size dtype: int64 - name: constrained_json_whitespace_pattern dtype: 'null' - name: context_length dtype: int64 - name: controller_port dtype: int64 - name: debug dtype: bool - name: decode_log_interval dtype: int64 - name: detokenizer_ipc_name dtype: string - name: detokenizer_port dtype: int64 - name: device dtype: string - name: disable_cuda_graph dtype: bool - name: disable_cuda_graph_padding dtype: bool - name: disable_custom_all_reduce dtype: bool - name: disable_disk_cache dtype: bool - name: disable_flashinfer dtype: bool - name: disable_flashinfer_sampling dtype: bool - name: disable_mla dtype: bool - name: disable_nan_detection dtype: bool - name: disable_penalizer dtype: bool - name: disable_radix_cache dtype: bool - name: disable_regex_jump_forward dtype: bool - name: dist_init_addr dtype: 'null' - name: dp_size dtype: int64 - name: dtype dtype: string - name: enable_double_sparsity dtype: bool - name: enable_dp_attention dtype: bool - name: enable_mixed_chunk dtype: bool - name: enable_overlap_schedule dtype: bool - name: enable_p2p_check dtype: bool - name: enable_stats_logging dtype: bool - name: enable_system_controller dtype: bool - name: enable_toppings dtype: bool - name: enable_torch_compile dtype: bool - name: flashinfer_workspace_size dtype: int64 - name: host dtype: string - name: init_new_token_ratio dtype: float64 - name: init_number_of_deltas dtype: int64 - name: init_number_of_loras dtype: int64 - name: init_toppings dtype: 'null' - name: is_embedding dtype: bool - name: kv_cache_dtype dtype: string - name: load_balance_method dtype: string - name: load_format dtype: string - name: log_requests dtype: bool - name: lora_paths dtype: 'null' - name: max_lora_ranks dtype: int64 - name: max_prefill_tokens dtype: int64 - name: max_running_requests dtype: 'null' - name: max_toppings_per_batch dtype: int64 - name: max_torch_compile_bs dtype: int64 - name: max_total_tokens dtype: 'null' - name: mem_fraction_static dtype: float64 - name: model_path dtype: string - name: nccl_ports sequence: string - name: new_token_ratio_decay dtype: float64 - name: nnodes dtype: int64 - name: node_rank dtype: int64 - name: num_continue_decode_steps dtype: int64 - name: num_continuous_decode_steps dtype: int64 - name: port dtype: int64 - name: quantization dtype: 'null' - name: random_seed dtype: int64 - name: retract_decode_steps dtype: int64 - name: sampling_backend dtype: string - name: schedule_conservativeness dtype: float64 - name: schedule_policy dtype: string - name: scheduler_input_ipc_name dtype: string - name: scheduler_port dtype: int64 - name: served_model_name dtype: string - name: server_id dtype: string - name: show_time_cost dtype: bool - name: skip_tokenizer_init dtype: bool - name: stream_interval dtype: int64 - name: tokenizer_ipc_name dtype: string - name: tokenizer_mode dtype: string - name: tokenizer_path dtype: string - name: tokenizer_port dtype: int64 - name: torchao_config dtype: string - name: tp_size dtype: int64 - name: triton_attention_reduce_in_fp32 dtype: bool - name: trust_remote_code dtype: bool - name: use_heterogeneous_pool dtype: bool - name: watchdog_timeout dtype: float64 - name: client_args struct: - name: dataset dtype: string - name: endpoint dtype: string - name: goodput dtype: 'null' - name: metric_percentiles dtype: string - name: model dtype: string - name: num_prompts dtype: int64 - name: output dtype: string - name: percentile_metrics dtype: string - name: request_rate dtype: float64 - name: tokenizer dtype: string - name: wait_until_ready dtype: bool - name: metrics struct: - name: completed dtype: int64 - name: mean_e2el_ms dtype: float64 - name: mean_itl_ms dtype: float64 - name: mean_tpot_ms dtype: float64 - name: mean_ttft_ms dtype: float64 - name: median_e2el_ms dtype: float64 - name: median_itl_ms dtype: float64 - name: median_tpot_ms dtype: float64 - name: median_ttft_ms dtype: float64 - name: output_throughput dtype: float64 - name: percentiles_e2el_ms sequence: sequence: float64 - name: percentiles_itl_ms sequence: sequence: float64 - name: percentiles_tpot_ms sequence: sequence: float64 - name: percentiles_ttft_ms sequence: sequence: float64 - name: request_goodput dtype: float64 - name: request_throughput dtype: float64 - name: std_e2el_ms dtype: float64 - name: std_itl_ms dtype: float64 - name: std_tpot_ms dtype: float64 - name: std_ttft_ms dtype: float64 - name: total_input dtype: int64 - name: total_output dtype: int64 - name: total_token_throughput dtype: float64 - name: device dtype: string splits: - name: train num_bytes: 9914 num_examples: 7 download_size: 56442 dataset_size: 9914 - config_name: 1xA100_EXPERTS_6_UNQUANT features: - name: server_args struct: - name: allow_toppings_registration dtype: bool - name: api_key dtype: 'null' - name: attention_backend dtype: string - name: base_min_new_token_ratio dtype: float64 - name: chat_template dtype: 'null' - name: chunked_prefill_size dtype: int64 - name: constrained_json_whitespace_pattern dtype: 'null' - name: context_length dtype: int64 - name: controller_port dtype: int64 - name: debug dtype: bool - name: decode_log_interval dtype: int64 - name: detokenizer_ipc_name dtype: string - name: detokenizer_port dtype: int64 - name: device dtype: string - name: disable_cuda_graph dtype: bool - name: disable_cuda_graph_padding dtype: bool - name: disable_custom_all_reduce dtype: bool - name: disable_disk_cache dtype: bool - name: disable_flashinfer dtype: bool - name: disable_flashinfer_sampling dtype: bool - name: disable_mla dtype: bool - name: disable_nan_detection dtype: bool - name: disable_penalizer dtype: bool - name: disable_radix_cache dtype: bool - name: disable_regex_jump_forward dtype: bool - name: dist_init_addr dtype: 'null' - name: dp_size dtype: int64 - name: dtype dtype: string - name: enable_double_sparsity dtype: bool - name: enable_dp_attention dtype: bool - name: enable_mixed_chunk dtype: bool - name: enable_overlap_schedule dtype: bool - name: enable_p2p_check dtype: bool - name: enable_stats_logging dtype: bool - name: enable_system_controller dtype: bool - name: enable_toppings dtype: bool - name: enable_torch_compile dtype: bool - name: flashinfer_workspace_size dtype: int64 - name: host dtype: string - name: init_new_token_ratio dtype: float64 - name: init_number_of_deltas dtype: int64 - name: init_number_of_loras dtype: int64 - name: init_toppings dtype: 'null' - name: is_embedding dtype: bool - name: kv_cache_dtype dtype: string - name: load_balance_method dtype: string - name: load_format dtype: string - name: log_requests dtype: bool - name: lora_paths dtype: 'null' - name: max_lora_ranks dtype: int64 - name: max_prefill_tokens dtype: int64 - name: max_running_requests dtype: 'null' - name: max_toppings_per_batch dtype: int64 - name: max_torch_compile_bs dtype: int64 - name: max_total_tokens dtype: 'null' - name: mem_fraction_static dtype: float64 - name: model_path dtype: string - name: nccl_ports sequence: string - name: new_token_ratio_decay dtype: float64 - name: nnodes dtype: int64 - name: node_rank dtype: int64 - name: num_continue_decode_steps dtype: int64 - name: num_continuous_decode_steps dtype: int64 - name: port dtype: int64 - name: quantization dtype: 'null' - name: random_seed dtype: int64 - name: retract_decode_steps dtype: int64 - name: sampling_backend dtype: string - name: schedule_conservativeness dtype: float64 - name: schedule_policy dtype: string - name: scheduler_input_ipc_name dtype: string - name: scheduler_port dtype: int64 - name: served_model_name dtype: string - name: server_id dtype: string - name: show_time_cost dtype: bool - name: skip_tokenizer_init dtype: bool - name: stream_interval dtype: int64 - name: tokenizer_ipc_name dtype: string - name: tokenizer_mode dtype: string - name: tokenizer_path dtype: string - name: tokenizer_port dtype: int64 - name: torchao_config dtype: string - name: tp_size dtype: int64 - name: triton_attention_reduce_in_fp32 dtype: bool - name: trust_remote_code dtype: bool - name: use_heterogeneous_pool dtype: bool - name: watchdog_timeout dtype: float64 - name: client_args struct: - name: dataset dtype: string - name: endpoint dtype: string - name: goodput dtype: 'null' - name: metric_percentiles dtype: string - name: model dtype: string - name: num_prompts dtype: int64 - name: output dtype: string - name: percentile_metrics dtype: string - name: request_rate dtype: float64 - name: tokenizer dtype: string - name: wait_until_ready dtype: bool - name: metrics struct: - name: completed dtype: int64 - name: mean_e2el_ms dtype: float64 - name: mean_itl_ms dtype: float64 - name: mean_tpot_ms dtype: float64 - name: mean_ttft_ms dtype: float64 - name: median_e2el_ms dtype: float64 - name: median_itl_ms dtype: float64 - name: median_tpot_ms dtype: float64 - name: median_ttft_ms dtype: float64 - name: output_throughput dtype: float64 - name: percentiles_e2el_ms sequence: sequence: float64 - name: percentiles_itl_ms sequence: sequence: float64 - name: percentiles_tpot_ms sequence: sequence: float64 - name: percentiles_ttft_ms sequence: sequence: float64 - name: request_goodput dtype: float64 - name: request_throughput dtype: float64 - name: std_e2el_ms dtype: float64 - name: std_itl_ms dtype: float64 - name: std_tpot_ms dtype: float64 - name: std_ttft_ms dtype: float64 - name: total_input dtype: int64 - name: total_output dtype: int64 - name: total_token_throughput dtype: float64 - name: device dtype: string splits: - name: train num_bytes: 8920 num_examples: 7 download_size: 55715 dataset_size: 8920 configs: - config_name: 1xA100_EXPERTS_2_QUANT_NAIVE data_files: - split: train path: 1xA100_EXPERTS_2_QUANT_NAIVE/train-* - config_name: 1xA100_EXPERTS_2_QUANT_SBMM_FORLOOP data_files: - split: train path: 1xA100_EXPERTS_2_QUANT_SBMM_FORLOOP/train-* - config_name: 1xA100_EXPERTS_2_QUANT_SBMM_MULTILAUNCH data_files: - split: train path: 1xA100_EXPERTS_2_QUANT_SBMM_MULTILAUNCH/train-* - config_name: 1xA100_EXPERTS_2_QUANT_SBMM_NAIVE data_files: - split: train path: 1xA100_EXPERTS_2_QUANT_SBMM_NAIVE/train-* - config_name: 1xA100_EXPERTS_2_UNQUANT data_files: - split: train path: 1xA100_EXPERTS_2_UNQUANT/train-* - config_name: 1xA100_EXPERTS_3_QUANT_NAIVE data_files: - split: train path: 1xA100_EXPERTS_3_QUANT_NAIVE/train-* - config_name: 1xA100_EXPERTS_3_QUANT_SBMM_FORLOOP data_files: - split: train path: 1xA100_EXPERTS_3_QUANT_SBMM_FORLOOP/train-* - config_name: 1xA100_EXPERTS_3_QUANT_SBMM_MULTILAUNCH data_files: - split: train path: 1xA100_EXPERTS_3_QUANT_SBMM_MULTILAUNCH/train-* - config_name: 1xA100_EXPERTS_3_QUANT_SBMM_NAIVE data_files: - split: train path: 1xA100_EXPERTS_3_QUANT_SBMM_NAIVE/train-* - config_name: 1xA100_EXPERTS_3_UNQUANT data_files: - split: train path: 1xA100_EXPERTS_3_UNQUANT/train-* - config_name: 1xA100_EXPERTS_4_QUANT_NAIVE data_files: - split: train path: 1xA100_EXPERTS_4_QUANT_NAIVE/train-* - config_name: 1xA100_EXPERTS_4_QUANT_SBMM_FORLOOP data_files: - split: train path: 1xA100_EXPERTS_4_QUANT_SBMM_FORLOOP/train-* - config_name: 1xA100_EXPERTS_4_QUANT_SBMM_MULTILAUNCH data_files: - split: train path: 1xA100_EXPERTS_4_QUANT_SBMM_MULTILAUNCH/train-* - config_name: 1xA100_EXPERTS_4_QUANT_SBMM_NAIVE data_files: - split: train path: 1xA100_EXPERTS_4_QUANT_SBMM_NAIVE/train-* - config_name: 1xA100_EXPERTS_4_UNQUANT data_files: - split: train path: 1xA100_EXPERTS_4_UNQUANT/train-* - config_name: 1xA100_EXPERTS_5_QUANT_NAIVE data_files: - split: train path: 1xA100_EXPERTS_5_QUANT_NAIVE/train-* - config_name: 1xA100_EXPERTS_5_QUANT_SBMM_FORLOOP data_files: - split: train path: 1xA100_EXPERTS_5_QUANT_SBMM_FORLOOP/train-* - config_name: 1xA100_EXPERTS_5_QUANT_SBMM_MULTILAUNCH data_files: - split: train path: 1xA100_EXPERTS_5_QUANT_SBMM_MULTILAUNCH/train-* - config_name: 1xA100_EXPERTS_5_QUANT_SBMM_NAIVE data_files: - split: train path: 1xA100_EXPERTS_5_QUANT_SBMM_NAIVE/train-* - config_name: 1xA100_EXPERTS_5_UNQUANT data_files: - split: train path: 1xA100_EXPERTS_5_UNQUANT/train-* - config_name: 1xA100_EXPERTS_6_QUANT_NAIVE data_files: - split: train path: 1xA100_EXPERTS_6_QUANT_NAIVE/train-* - config_name: 1xA100_EXPERTS_6_QUANT_SBMM_FORLOOP data_files: - split: train path: 1xA100_EXPERTS_6_QUANT_SBMM_FORLOOP/train-* - config_name: 1xA100_EXPERTS_6_QUANT_SBMM_MULTILAUNCH data_files: - split: train path: 1xA100_EXPERTS_6_QUANT_SBMM_MULTILAUNCH/train-* - config_name: 1xA100_EXPERTS_6_QUANT_SBMM_NAIVE data_files: - split: train path: 1xA100_EXPERTS_6_QUANT_SBMM_NAIVE/train-* - config_name: 1xA100_EXPERTS_6_UNQUANT data_files: - split: train path: 1xA100_EXPERTS_6_UNQUANT/train-* ---