Skip to content

Latest commit

 

History

History
413 lines (411 loc) · 56.4 KB

File metadata and controls

413 lines (411 loc) · 56.4 KB

Inference-Perf CLI Flags

These command line flags are automatically generated from the CLI parser. The global flags at the top of the table control the tool itself; every other flag is generated from the internal Config schema and overrides that configuration directly from the CLI without using a yaml configuration file.

Flag Type Description
-c, --config_file str Config File
-a, --analyze list of str Path to a report directories to analyze
-u, --unified_analysis_dir str Unified analysis directory path
--log-level Enum (DEBUG, INFO, WARNING, ERROR, CRITICAL) Logging level (default: INFO)
--api.type Enum (completion, chat, anthropic_messages) API endpoint to benchmark: text completion or chat completion.
--api.streaming boolean Stream responses instead of waiting for the full response. Enables TTFT and TPOT metrics.
--api.headers JSON Additional HTTP headers to send with every request.
--api.slo_unit str Time unit for SLO header values: 's', 'ms' or 'us'. Defaults to 'ms'.
--api.slo_tpot_header str Request header carrying the per-request TPOT SLO threshold. Defaults to 'x-slo-tpot-<slo_unit>'.
--api.slo_ttft_header str Request header carrying the per-request TTFT SLO threshold. Defaults to 'x-slo-ttft-<slo_unit>'.
--api.response_format.type Enum (json_schema, json_object) Structured output mode: a full JSON schema or any JSON object.
--api.response_format.name str Name given to the JSON schema in the request payload.
--api.response_format.json_schema JSON JSON schema the model output must conform to when type is 'json_schema'.
--api.session_id_header_key str Header used to send the session ID with each request in multi-turn benchmarks.
--api.session_token_header_key str Response header carrying a server-assigned session token, replayed as a request header on later requests of the same session to keep router session affinity.
--data.type Enum (mock, shareGPT, synthetic, random, shared_prefix, cnn_dailymail, infinity_instruct, billsum_conversations, otel_trace_replay, weka_trace_replay, conversation_replay, visionarena, synthetic_agentic) Dataset or generator used to produce prompts.
--data.path str Path to the downloaded ShareGPT dataset. Only used by the 'shareGPT' type.
--data.corpus_file_path str Path to a text file to use as the prompt tokenization corpus instead of the default hardcoded sonnet
--data.input_distribution.min int Smallest value the distribution can produce; samples below are clamped.
--data.input_distribution.max int Largest value the distribution can produce; samples above are clamped.
--data.input_distribution.mean float Mean of the distribution.
--data.input_distribution.std_dev float Standard deviation of the distribution. Exclusive with 'variance'.
--data.input_distribution.total_count int Total number of values to sample from the distribution.
--data.input_distribution.type Enum (normal, skew_normal, lognormal, uniform, poisson, fixed) Shape of the distribution to sample values from.
--data.input_distribution.variance float Variance of the distribution. Exclusive with 'std_dev'.
--data.input_distribution.skew float Skewness of the distribution. Only used when type is 'skew_normal'.
--data.output_distribution.min int Smallest value the distribution can produce; samples below are clamped.
--data.output_distribution.max int Largest value the distribution can produce; samples above are clamped.
--data.output_distribution.mean float Mean of the distribution.
--data.output_distribution.std_dev float Standard deviation of the distribution. Exclusive with 'variance'.
--data.output_distribution.total_count int Total number of values to sample from the distribution.
--data.output_distribution.type Enum (normal, skew_normal, lognormal, uniform, poisson, fixed) Shape of the distribution to sample values from.
--data.output_distribution.variance float Variance of the distribution. Exclusive with 'std_dev'.
--data.output_distribution.skew float Skewness of the distribution. Only used when type is 'skew_normal'.
--data.shared_prefix.num_groups int Number of unique system prompts (shared prefix groups) to generate.
--data.shared_prefix.num_prompts_per_group int Number of prompts generated per shared system prompt.
--data.shared_prefix.system_prompt_len string Length of the shared system prompt in tokens: a fixed value or a distribution.
--data.shared_prefix.question_len string Length of the question part in tokens: a fixed value or a distribution.
--data.shared_prefix.output_len string Requested output length in tokens: a fixed value or a distribution.
--data.shared_prefix.seed int Random seed for reproducible prompt generation.
--data.shared_prefix.question_distribution.min int Smallest value the distribution can produce; samples below are clamped.
--data.shared_prefix.question_distribution.max int Largest value the distribution can produce; samples above are clamped.
--data.shared_prefix.question_distribution.mean float Mean of the distribution.
--data.shared_prefix.question_distribution.std_dev float Standard deviation of the distribution. Exclusive with 'variance'.
--data.shared_prefix.question_distribution.total_count int Total number of values to sample from the distribution.
--data.shared_prefix.question_distribution.type Enum (normal, skew_normal, lognormal, uniform, poisson, fixed) Shape of the distribution to sample values from.
--data.shared_prefix.question_distribution.variance float Variance of the distribution. Exclusive with 'std_dev'.
--data.shared_prefix.question_distribution.skew float Skewness of the distribution. Only used when type is 'skew_normal'.
--data.shared_prefix.output_distribution.min int Smallest value the distribution can produce; samples below are clamped.
--data.shared_prefix.output_distribution.max int Largest value the distribution can produce; samples above are clamped.
--data.shared_prefix.output_distribution.mean float Mean of the distribution.
--data.shared_prefix.output_distribution.std_dev float Standard deviation of the distribution. Exclusive with 'variance'.
--data.shared_prefix.output_distribution.total_count int Total number of values to sample from the distribution.
--data.shared_prefix.output_distribution.type Enum (normal, skew_normal, lognormal, uniform, poisson, fixed) Shape of the distribution to sample values from.
--data.shared_prefix.output_distribution.variance float Variance of the distribution. Exclusive with 'std_dev'.
--data.shared_prefix.output_distribution.skew float Skewness of the distribution. Only used when type is 'skew_normal'.
--data.shared_prefix.enable_multi_turn_chat boolean Send each group's prompts as consecutive turns of one chat conversation.
--data.shared_prefix.multimodal.image.count.min int Smallest value the distribution can produce; samples below are clamped.
--data.shared_prefix.multimodal.image.count.max int Largest value the distribution can produce; samples above are clamped.
--data.shared_prefix.multimodal.image.count.mean float Mean of the distribution.
--data.shared_prefix.multimodal.image.count.std_dev float Standard deviation of the distribution. Exclusive with 'variance'.
--data.shared_prefix.multimodal.image.count.total_count int Total number of values to sample from the distribution.
--data.shared_prefix.multimodal.image.count.type Enum (normal, skew_normal, lognormal, uniform, poisson, fixed) Shape of the distribution to sample values from.
--data.shared_prefix.multimodal.image.count.variance float Variance of the distribution. Exclusive with 'std_dev'.
--data.shared_prefix.multimodal.image.count.skew float Skewness of the distribution. Only used when type is 'skew_normal'.
--data.shared_prefix.multimodal.image.insertion_point string Placement of media within the text prompt. Float in range [0.0, 1.0] (0=start, 1=end), or a Distribution to sample from.
--data.shared_prefix.multimodal.image.resolutions JSON Resolution or list of weighted resolutions for generated images.
--data.shared_prefix.multimodal.image.representation Enum (png, jpeg, webp) Wire encoding for emitted image bytes: png (default, lossless) or jpeg (lossy, smaller payload). Some VLMs prefer one or the other; consult the model's spec sheet.
--data.shared_prefix.multimodal.video.count.min int Smallest value the distribution can produce; samples below are clamped.
--data.shared_prefix.multimodal.video.count.max int Largest value the distribution can produce; samples above are clamped.
--data.shared_prefix.multimodal.video.count.mean float Mean of the distribution.
--data.shared_prefix.multimodal.video.count.std_dev float Standard deviation of the distribution. Exclusive with 'variance'.
--data.shared_prefix.multimodal.video.count.total_count int Total number of values to sample from the distribution.
--data.shared_prefix.multimodal.video.count.type Enum (normal, skew_normal, lognormal, uniform, poisson, fixed) Shape of the distribution to sample values from.
--data.shared_prefix.multimodal.video.count.variance float Variance of the distribution. Exclusive with 'std_dev'.
--data.shared_prefix.multimodal.video.count.skew float Skewness of the distribution. Only used when type is 'skew_normal'.
--data.shared_prefix.multimodal.video.insertion_point string Placement of media within the text prompt. Float in range [0.0, 1.0] (0=start, 1=end), or a Distribution to sample from.
--data.shared_prefix.multimodal.video.profiles JSON Video profile or list of weighted video profiles for generated videos.
--data.shared_prefix.multimodal.video.representation Enum (mp4, png_frames, jpeg_frames) Wire-format strategy. mp4 sends one video_url block carrying an MP4 blob (measures full pipeline including server-side decode). png_frames and jpeg_frames send frames × image_url blocks at one insertion point in the named encoding (no decode dependency, useful for prefix-cache benchmarks and servers that don't accept video_url).
--data.shared_prefix.multimodal.audio.count.min int Smallest value the distribution can produce; samples below are clamped.
--data.shared_prefix.multimodal.audio.count.max int Largest value the distribution can produce; samples above are clamped.
--data.shared_prefix.multimodal.audio.count.mean float Mean of the distribution.
--data.shared_prefix.multimodal.audio.count.std_dev float Standard deviation of the distribution. Exclusive with 'variance'.
--data.shared_prefix.multimodal.audio.count.total_count int Total number of values to sample from the distribution.
--data.shared_prefix.multimodal.audio.count.type Enum (normal, skew_normal, lognormal, uniform, poisson, fixed) Shape of the distribution to sample values from.
--data.shared_prefix.multimodal.audio.count.variance float Variance of the distribution. Exclusive with 'std_dev'.
--data.shared_prefix.multimodal.audio.count.skew float Skewness of the distribution. Only used when type is 'skew_normal'.
--data.shared_prefix.multimodal.audio.insertion_point string Placement of media within the text prompt. Float in range [0.0, 1.0] (0=start, 1=end), or a Distribution to sample from.
--data.shared_prefix.multimodal.audio.durations JSON Duration or list of weighted durations for generated audio clips.
--data.multimodal.image.count.min int Smallest value the distribution can produce; samples below are clamped.
--data.multimodal.image.count.max int Largest value the distribution can produce; samples above are clamped.
--data.multimodal.image.count.mean float Mean of the distribution.
--data.multimodal.image.count.std_dev float Standard deviation of the distribution. Exclusive with 'variance'.
--data.multimodal.image.count.total_count int Total number of values to sample from the distribution.
--data.multimodal.image.count.type Enum (normal, skew_normal, lognormal, uniform, poisson, fixed) Shape of the distribution to sample values from.
--data.multimodal.image.count.variance float Variance of the distribution. Exclusive with 'std_dev'.
--data.multimodal.image.count.skew float Skewness of the distribution. Only used when type is 'skew_normal'.
--data.multimodal.image.insertion_point string Placement of media within the text prompt. Float in range [0.0, 1.0] (0=start, 1=end), or a Distribution to sample from.
--data.multimodal.image.resolutions JSON Resolution or list of weighted resolutions for generated images.
--data.multimodal.image.representation Enum (png, jpeg, webp) Wire encoding for emitted image bytes: png (default, lossless) or jpeg (lossy, smaller payload). Some VLMs prefer one or the other; consult the model's spec sheet.
--data.multimodal.video.count.min int Smallest value the distribution can produce; samples below are clamped.
--data.multimodal.video.count.max int Largest value the distribution can produce; samples above are clamped.
--data.multimodal.video.count.mean float Mean of the distribution.
--data.multimodal.video.count.std_dev float Standard deviation of the distribution. Exclusive with 'variance'.
--data.multimodal.video.count.total_count int Total number of values to sample from the distribution.
--data.multimodal.video.count.type Enum (normal, skew_normal, lognormal, uniform, poisson, fixed) Shape of the distribution to sample values from.
--data.multimodal.video.count.variance float Variance of the distribution. Exclusive with 'std_dev'.
--data.multimodal.video.count.skew float Skewness of the distribution. Only used when type is 'skew_normal'.
--data.multimodal.video.insertion_point string Placement of media within the text prompt. Float in range [0.0, 1.0] (0=start, 1=end), or a Distribution to sample from.
--data.multimodal.video.profiles JSON Video profile or list of weighted video profiles for generated videos.
--data.multimodal.video.representation Enum (mp4, png_frames, jpeg_frames) Wire-format strategy. mp4 sends one video_url block carrying an MP4 blob (measures full pipeline including server-side decode). png_frames and jpeg_frames send frames × image_url blocks at one insertion point in the named encoding (no decode dependency, useful for prefix-cache benchmarks and servers that don't accept video_url).
--data.multimodal.audio.count.min int Smallest value the distribution can produce; samples below are clamped.
--data.multimodal.audio.count.max int Largest value the distribution can produce; samples above are clamped.
--data.multimodal.audio.count.mean float Mean of the distribution.
--data.multimodal.audio.count.std_dev float Standard deviation of the distribution. Exclusive with 'variance'.
--data.multimodal.audio.count.total_count int Total number of values to sample from the distribution.
--data.multimodal.audio.count.type Enum (normal, skew_normal, lognormal, uniform, poisson, fixed) Shape of the distribution to sample values from.
--data.multimodal.audio.count.variance float Variance of the distribution. Exclusive with 'std_dev'.
--data.multimodal.audio.count.skew float Skewness of the distribution. Only used when type is 'skew_normal'.
--data.multimodal.audio.insertion_point string Placement of media within the text prompt. Float in range [0.0, 1.0] (0=start, 1=end), or a Distribution to sample from.
--data.multimodal.audio.durations JSON Duration or list of weighted durations for generated audio clips.
--data.trace.file str Path to the trace file to replay.
--data.trace.format Enum (AzurePublicDataset) Format of the trace file.
--data.otel_trace_replay.use_static_model boolean Use a single static model for all requests
--data.otel_trace_replay.static_model_name str Static model name (required if use_static_model=True)
--data.otel_trace_replay.model_mapping JSON Map recorded model names to target models
--data.otel_trace_replay.default_max_tokens int Default max_tokens if not specified in trace
--data.otel_trace_replay.override_tool_call_max_tokens boolean Override tool call max_tokens to 4096 instead of using trace recorded length
--data.otel_trace_replay.tool_choice_mode Enum (force_recorded, as_recorded) Whether to inject a tool_choice policy on recorded tool-call turns. 'force_recorded' (default) forces the recorded function, or 'required' when the recorded turn made several calls or named a tool absent from this turn's list. 'as_recorded' injects nothing, leaving the choice to the model, at the cost of turns where it answers in prose and the recorded tool results no longer match.
--data.otel_trace_replay.inject_random_session_id boolean Inject random string into unique segments to invalidate KV-cache between sessions
--data.otel_trace_replay.duplicate_sessions_target int Target number of sessions to reach by duplicating existing sessions. If None, no duplication occurs.
--data.otel_trace_replay.max_wait_ms int Maximum inter-event wait time in milliseconds. Caps the delay between predecessor completion and event dispatch to avoid reproducing unusually long tool/agent execution times from the original trace.
--data.otel_trace_replay.predecessor_wait_timeout_sec float Seconds to wait for predecessor events to complete before failing. 0 waits indefinitely; use with care because a genuinely stuck predecessor will then never time out and successors will wait forever.
--data.otel_trace_replay.include_errors boolean Include spans with error status
--data.otel_trace_replay.skip_invalid_files boolean Skip invalid trace files instead of failing
--data.otel_trace_replay.bad_tool_call_handling Enum (none, use_recorded) How to handle tool_calls whose function.arguments is not valid JSON. none (default): no mitigation, bytes propagate and vLLM may return HTTP 400 on the next turn. use_recorded: discard the live response and substitute the recorded assistant message at the affected slot; the recorded tool_call_id flows into the recorded role:tool successor unchanged.
--data.otel_trace_replay.trace_directory str Directory containing OTel JSON trace files
--data.otel_trace_replay.trace_files JSON List of paths to specific OTel JSON trace files
--data.otel_trace_replay.hf_dataset_path JSON HuggingFace dataset path. Can be:
  • String: 'username/dataset-name'
  • Dict: {'path': 'username/dataset-name', 'revision': 'main', 'split': 'train'} Any extra keys in the dict are passed as kwargs to datasets.load_dataset(). | | --data.otel_trace_replay.filter | str | Lambda expression to filter trace records. Applied uniformly to all data sources. Example: "lambda x: x['benchmark'] == 'gsm8k'" or "lambda x: 'spans' in x and len(x['spans']) > 5" Security: Filter expressions use eval() and should only contain trusted input. | | --data.otel_trace_replay.disable_output_substitution | boolean | When True, replay each call with its recorded assistant output (text and tool calls) instead of substituting the live output from predecessor calls. Dependency timing (waiting for predecessors) is still enforced. Default False preserves faithful live-output replay. | | --data.otel_trace_replay.attribute_to_header_map | JSON | Map OTel span attributes to HTTP headers | | --data.otel_trace_replay.attribute_to_label_map | JSON | Map OTel span attributes to metrics reporting labels | | --data.weka_trace_replay.use_static_model | boolean | Use a single static model for all requests | | --data.weka_trace_replay.static_model_name | str | Static model name (required if use_static_model=True) | | --data.weka_trace_replay.model_mapping | JSON | Map recorded model names to target models | | --data.weka_trace_replay.default_max_tokens | int | Default max_tokens if not specified in trace | | --data.weka_trace_replay.override_tool_call_max_tokens | boolean | Override tool call max_tokens to 4096 instead of using trace recorded length | | --data.weka_trace_replay.tool_choice_mode | Enum (force_recorded, as_recorded) | Whether to inject a tool_choice policy on recorded tool-call turns. 'force_recorded' (default) forces the recorded function, or 'required' when the recorded turn made several calls or named a tool absent from this turn's list. 'as_recorded' injects nothing, leaving the choice to the model, at the cost of turns where it answers in prose and the recorded tool results no longer match. | | --data.weka_trace_replay.inject_random_session_id | boolean | Inject random string into unique segments to invalidate KV-cache between sessions | | --data.weka_trace_replay.duplicate_sessions_target | int | Target number of sessions to reach by duplicating existing sessions. If None, no duplication occurs. | | --data.weka_trace_replay.max_wait_ms | int | Maximum inter-event wait time in milliseconds. Caps the delay between predecessor completion and event dispatch to avoid reproducing unusually long tool/agent execution times from the original trace. | | --data.weka_trace_replay.predecessor_wait_timeout_sec | float | Seconds to wait for predecessor events to complete before failing. 0 waits indefinitely; use with care because a genuinely stuck predecessor will then never time out and successors will wait forever. | | --data.weka_trace_replay.include_errors | boolean | Include spans with error status | | --data.weka_trace_replay.skip_invalid_files | boolean | Skip invalid trace files instead of failing | | --data.weka_trace_replay.bad_tool_call_handling | Enum (none, use_recorded) | How to handle tool_calls whose function.arguments is not valid JSON. none (default): no mitigation, bytes propagate and vLLM may return HTTP 400 on the next turn. use_recorded: discard the live response and substitute the recorded assistant message at the affected slot; the recorded tool_call_id flows into the recorded role:tool successor unchanged. | | --data.weka_trace_replay.trace_directory | str | Directory containing Weka JSON trace files | | --data.weka_trace_replay.trace_files | JSON | List of paths to specific Weka JSON trace files | | --data.weka_trace_replay.hf_dataset_path | JSON | HuggingFace dataset path. Can be:
  • String: 'username/dataset-name'
  • Dict: {'path': 'username/dataset-name', 'revision': 'main', 'split': 'train'} Any extra keys in the dict are passed as kwargs to datasets.load_dataset(). | | --data.weka_trace_replay.filter | str | Lambda expression to filter traces. Applied uniformly to all data sources. Receives the trace dict plus derived aggregates: max_tokens (largest single-request input+output), total_tokens, and num_turns, which counts flattened requests, so a subagent making three calls contributes three. Example: "lambda x: x['max_tokens'] < 262144" Security: Filter expressions use eval() and should only contain trusted input. | | --data.weka_trace_replay.trace_idle_gap_cap_seconds | float | Cap idle timing gaps between turns in seconds | | --data.weka_trace_replay.ignore_trace_delays | boolean | Ignore delays/delays from original trace and run back-to-back | | --data.weka_trace_replay.use_think_time_only | boolean | Only use think_time attribute instead of timestamps | | --data.weka_trace_replay.default_block_size | int | Default block size if not specified in trace | | --data.weka_trace_replay.num_dataset_entries | int | Max number of dataset traces to load from HuggingFace | | --data.weka_trace_replay.datagen_workers | int | Number of processes used to reconstruct trace sessions during data generation. Defaults to the number of CPUs available to the process (the minimum of the scheduling affinity and the cgroup CPU quota), capped at the trace count. Set to 1 to disable multiprocessing. Parallel building requires Linux; other platforms build serially. Output is deterministic and identical regardless of this value. | | --data.conversation_replay.seed | int | Random seed for deterministic generation | | --data.conversation_replay.num_conversations | int | Number of conversation blueprints to generate | | --data.conversation_replay.shared_system_prompt_len | int | Fixed shared system prompt length in tokens | | --data.conversation_replay.dynamic_system_prompt_len.min | int | Smallest value the distribution can produce; samples below are clamped. | | --data.conversation_replay.dynamic_system_prompt_len.max | int | Largest value the distribution can produce; samples above are clamped. | | --data.conversation_replay.dynamic_system_prompt_len.mean | float | Mean of the distribution. | | --data.conversation_replay.dynamic_system_prompt_len.std_dev | float | Standard deviation of the distribution. Exclusive with 'variance'. | | --data.conversation_replay.dynamic_system_prompt_len.total_count | int | Total number of values to sample from the distribution. | | --data.conversation_replay.dynamic_system_prompt_len.type | Enum (normal, skew_normal, lognormal, uniform, poisson, fixed) | Shape of the distribution to sample values from. | | --data.conversation_replay.dynamic_system_prompt_len.variance | float | Variance of the distribution. Exclusive with 'std_dev'. | | --data.conversation_replay.dynamic_system_prompt_len.skew | float | Skewness of the distribution. Only used when type is 'skew_normal'. | | --data.conversation_replay.turns_per_conversation.min | int | Smallest value the distribution can produce; samples below are clamped. | | --data.conversation_replay.turns_per_conversation.max | int | Largest value the distribution can produce; samples above are clamped. | | --data.conversation_replay.turns_per_conversation.mean | float | Mean of the distribution. | | --data.conversation_replay.turns_per_conversation.std_dev | float | Standard deviation of the distribution. Exclusive with 'variance'. | | --data.conversation_replay.turns_per_conversation.total_count | int | Total number of values to sample from the distribution. | | --data.conversation_replay.turns_per_conversation.type | Enum (normal, skew_normal, lognormal, uniform, poisson, fixed) | Shape of the distribution to sample values from. | | --data.conversation_replay.turns_per_conversation.variance | float | Variance of the distribution. Exclusive with 'std_dev'. | | --data.conversation_replay.turns_per_conversation.skew | float | Skewness of the distribution. Only used when type is 'skew_normal'. | | --data.conversation_replay.input_tokens_per_turn.min | int | Smallest value the distribution can produce; samples below are clamped. | | --data.conversation_replay.input_tokens_per_turn.max | int | Largest value the distribution can produce; samples above are clamped. | | --data.conversation_replay.input_tokens_per_turn.mean | float | Mean of the distribution. | | --data.conversation_replay.input_tokens_per_turn.std_dev | float | Standard deviation of the distribution. Exclusive with 'variance'. | | --data.conversation_replay.input_tokens_per_turn.total_count | int | Total number of values to sample from the distribution. | | --data.conversation_replay.input_tokens_per_turn.type | Enum (normal, skew_normal, lognormal, uniform, poisson, fixed) | Shape of the distribution to sample values from. | | --data.conversation_replay.input_tokens_per_turn.variance | float | Variance of the distribution. Exclusive with 'std_dev'. | | --data.conversation_replay.input_tokens_per_turn.skew | float | Skewness of the distribution. Only used when type is 'skew_normal'. | | --data.conversation_replay.output_tokens_per_turn.min | int | Smallest value the distribution can produce; samples below are clamped. | | --data.conversation_replay.output_tokens_per_turn.max | int | Largest value the distribution can produce; samples above are clamped. | | --data.conversation_replay.output_tokens_per_turn.mean | float | Mean of the distribution. | | --data.conversation_replay.output_tokens_per_turn.std_dev | float | Standard deviation of the distribution. Exclusive with 'variance'. | | --data.conversation_replay.output_tokens_per_turn.total_count | int | Total number of values to sample from the distribution. | | --data.conversation_replay.output_tokens_per_turn.type | Enum (normal, skew_normal, lognormal, uniform, poisson, fixed) | Shape of the distribution to sample values from. | | --data.conversation_replay.output_tokens_per_turn.variance | float | Variance of the distribution. Exclusive with 'std_dev'. | | --data.conversation_replay.output_tokens_per_turn.skew | float | Skewness of the distribution. Only used when type is 'skew_normal'. | | --data.conversation_replay.tool_call_latency_sec.min | int | Smallest value the distribution can produce; samples below are clamped. | | --data.conversation_replay.tool_call_latency_sec.max | int | Largest value the distribution can produce; samples above are clamped. | | --data.conversation_replay.tool_call_latency_sec.mean | float | Mean of the distribution. | | --data.conversation_replay.tool_call_latency_sec.std_dev | float | Standard deviation of the distribution. Exclusive with 'variance'. | | --data.conversation_replay.tool_call_latency_sec.total_count | int | Total number of values to sample from the distribution. | | --data.conversation_replay.tool_call_latency_sec.type | Enum (normal, skew_normal, lognormal, uniform, poisson, fixed) | Shape of the distribution to sample values from. | | --data.conversation_replay.tool_call_latency_sec.variance | float | Variance of the distribution. Exclusive with 'std_dev'. | | --data.conversation_replay.tool_call_latency_sec.skew | float | Skewness of the distribution. Only used when type is 'skew_normal'. | | --data.conversation_replay.max_model_len | int | Maximum model context length in tokens | | --data.visionarena.hf_dataset_name | str | HuggingFace dataset identifier; override only when mirroring the dataset elsewhere. | | --data.visionarena.hf_split | str | HuggingFace split to stream. | | --data.visionarena.hf_data_files | str | Optional data_files glob forwarded to load_dataset. | | --data.visionarena.num_rows | int | Number of usable rows to stream into the in-memory request pool at startup. Caps memory use; the benchmark cycles through this pool. | | --data.visionarena.max_images_per_request | int | Cap on images attached per request; truncates a row's image list. | | --data.visionarena.insertion_point | string | Placement of the image block(s) within the prompt text. Float in [0.0, 1.0] (0=start, 1=end), or a Distribution to sample per request. | | --data.synthetic_agentic.use_static_model | boolean | Use a single static model for all requests | | --data.synthetic_agentic.static_model_name | str | Static model name (required if use_static_model=True) | | --data.synthetic_agentic.model_mapping | JSON | Map recorded model names to target models | | --data.synthetic_agentic.default_max_tokens | int | Default max_tokens if not specified in trace | | --data.synthetic_agentic.override_tool_call_max_tokens | boolean | Override tool-call max_tokens to 4096 instead of using the generated call's own length. Defaults False here (unlike trace replay) because the generator sizes each tool call itself, so the generated length is already correct for this model. | | --data.synthetic_agentic.tool_choice_mode | Enum (force_recorded, as_recorded) | Whether to inject a tool_choice policy on recorded tool-call turns. 'force_recorded' (default) forces the recorded function, or 'required' when the recorded turn made several calls or named a tool absent from this turn's list. 'as_recorded' injects nothing, leaving the choice to the model, at the cost of turns where it answers in prose and the recorded tool results no longer match. | | --data.synthetic_agentic.inject_random_session_id | boolean | Not applicable to synthetic generation (pinned False): sessions are already generated with distinct content per session index, so there is no recorded session ID to randomize. | | --data.synthetic_agentic.duplicate_sessions_target | int | Not applicable to synthetic generation (pinned None): raise num_sessions to generate more sessions instead of duplicating existing ones. | | --data.synthetic_agentic.max_wait_ms | int | Maximum inter-event wait time in milliseconds. Caps the delay between predecessor completion and event dispatch to avoid reproducing unusually long tool/agent execution times from the original trace. | | --data.synthetic_agentic.predecessor_wait_timeout_sec | float | Seconds to wait for predecessor events to complete before failing. 0 waits indefinitely; use with care because a genuinely stuck predecessor will then never time out and successors will wait forever. | | --data.synthetic_agentic.include_errors | boolean | Not applicable to synthetic generation (pinned True): this filters recorded spans by error status when building a replay graph from a trace file; synthetic sessions have no recorded spans to filter. | | --data.synthetic_agentic.skip_invalid_files | boolean | Not applicable to synthetic generation (pinned False): sessions are generated in-memory, not loaded from trace files, so there are no invalid files to skip. | | --data.synthetic_agentic.bad_tool_call_handling | Enum (none, use_recorded) | How to handle tool_calls whose function.arguments is not valid JSON. none (default): no mitigation, bytes propagate and vLLM may return HTTP 400 on the next turn. use_recorded: discard the live response and substitute the recorded assistant message at the affected slot; the recorded tool_call_id flows into the recorded role:tool successor unchanged. | | --data.synthetic_agentic.num_sessions | int | Number of sessions (load volume) | | --data.synthetic_agentic.input_tokens_per_turn.min | int | Smallest value the distribution can produce; samples below are clamped. | | --data.synthetic_agentic.input_tokens_per_turn.max | int | Largest value the distribution can produce; samples above are clamped. | | --data.synthetic_agentic.input_tokens_per_turn.mean | float | Mean of the distribution. | | --data.synthetic_agentic.input_tokens_per_turn.std_dev | float | Standard deviation of the distribution. Exclusive with 'variance'. | | --data.synthetic_agentic.input_tokens_per_turn.total_count | int | Total number of values to sample from the distribution. | | --data.synthetic_agentic.input_tokens_per_turn.type | Enum (normal, skew_normal, lognormal, uniform, poisson, fixed) | Shape of the distribution to sample values from. | | --data.synthetic_agentic.input_tokens_per_turn.variance | float | Variance of the distribution. Exclusive with 'std_dev'. | | --data.synthetic_agentic.input_tokens_per_turn.skew | float | Skewness of the distribution. Only used when type is 'skew_normal'. | | --data.synthetic_agentic.output_tokens_per_turn.min | int | Smallest value the distribution can produce; samples below are clamped. | | --data.synthetic_agentic.output_tokens_per_turn.max | int | Largest value the distribution can produce; samples above are clamped. | | --data.synthetic_agentic.output_tokens_per_turn.mean | float | Mean of the distribution. | | --data.synthetic_agentic.output_tokens_per_turn.std_dev | float | Standard deviation of the distribution. Exclusive with 'variance'. | | --data.synthetic_agentic.output_tokens_per_turn.total_count | int | Total number of values to sample from the distribution. | | --data.synthetic_agentic.output_tokens_per_turn.type | Enum (normal, skew_normal, lognormal, uniform, poisson, fixed) | Shape of the distribution to sample values from. | | --data.synthetic_agentic.output_tokens_per_turn.variance | float | Variance of the distribution. Exclusive with 'std_dev'. | | --data.synthetic_agentic.output_tokens_per_turn.skew | float | Skewness of the distribution. Only used when type is 'skew_normal'. | | --data.synthetic_agentic.turns_per_session.min | int | Smallest value the distribution can produce; samples below are clamped. | | --data.synthetic_agentic.turns_per_session.max | int | Largest value the distribution can produce; samples above are clamped. | | --data.synthetic_agentic.turns_per_session.mean | float | Mean of the distribution. | | --data.synthetic_agentic.turns_per_session.std_dev | float | Standard deviation of the distribution. Exclusive with 'variance'. | | --data.synthetic_agentic.turns_per_session.total_count | int | Total number of values to sample from the distribution. | | --data.synthetic_agentic.turns_per_session.type | Enum (normal, skew_normal, lognormal, uniform, poisson, fixed) | Shape of the distribution to sample values from. | | --data.synthetic_agentic.turns_per_session.variance | float | Variance of the distribution. Exclusive with 'std_dev'. | | --data.synthetic_agentic.turns_per_session.skew | float | Skewness of the distribution. Only used when type is 'skew_normal'. | | --data.synthetic_agentic.fanout_probability | float | Probability an agent spawns sub-agents (instead of just answering), rolled fresh for each of the root's turns and once for each sub-agent. Default 0 = single-agent; 1 = always spawn (full tree to max_depth). | | --data.synthetic_agentic.theme_mix | JSON | theme name -> weight. Preferred form is an explicit block, {name: {weight: W}}; a bare float {name: W} is also accepted for brevity. Default is an equal mix of the four built-in themes. Use theme_weights() to read normalized {name: float}. | | --data.synthetic_agentic.seed | int | Base seed for stable per-session RNG | | --data.synthetic_agentic.shared_system_prompt_len | int | Tokens of a fixed system-prompt head that opens EVERY agent call (the standing 'system head' real agents carry: tool instructions, policies). Defaults to 1000 because virtually every agentic flow ships a non-trivial system prompt; set 0 only for a deliberately head-less baseline. | | --data.synthetic_agentic.tool_loop_depth.min | int | Smallest value the distribution can produce; samples below are clamped. | | --data.synthetic_agentic.tool_loop_depth.max | int | Largest value the distribution can produce; samples above are clamped. | | --data.synthetic_agentic.tool_loop_depth.mean | float | Mean of the distribution. | | --data.synthetic_agentic.tool_loop_depth.std_dev | float | Standard deviation of the distribution. Exclusive with 'variance'. | | --data.synthetic_agentic.tool_loop_depth.total_count | int | Total number of values to sample from the distribution. | | --data.synthetic_agentic.tool_loop_depth.type | Enum (normal, skew_normal, lognormal, uniform, poisson, fixed) | Shape of the distribution to sample values from. | | --data.synthetic_agentic.tool_loop_depth.variance | float | Variance of the distribution. Exclusive with 'std_dev'. | | --data.synthetic_agentic.tool_loop_depth.skew | float | Skewness of the distribution. Only used when type is 'skew_normal'. | | --data.synthetic_agentic.sub_agents_per_spawn.min | int | Smallest value the distribution can produce; samples below are clamped. | | --data.synthetic_agentic.sub_agents_per_spawn.max | int | Largest value the distribution can produce; samples above are clamped. | | --data.synthetic_agentic.sub_agents_per_spawn.mean | float | Mean of the distribution. | | --data.synthetic_agentic.sub_agents_per_spawn.std_dev | float | Standard deviation of the distribution. Exclusive with 'variance'. | | --data.synthetic_agentic.sub_agents_per_spawn.total_count | int | Total number of values to sample from the distribution. | | --data.synthetic_agentic.sub_agents_per_spawn.type | Enum (normal, skew_normal, lognormal, uniform, poisson, fixed) | Shape of the distribution to sample values from. | | --data.synthetic_agentic.sub_agents_per_spawn.variance | float | Variance of the distribution. Exclusive with 'std_dev'. | | --data.synthetic_agentic.sub_agents_per_spawn.skew | float | Skewness of the distribution. Only used when type is 'skew_normal'. | | --data.synthetic_agentic.max_depth | int | Hard recursion terminator | | --data.synthetic_agentic.max_events_per_session | int | Self-limiting event budget | | --data.synthetic_agentic.tool_catalog_size_per_agent.min | int | Smallest value the distribution can produce; samples below are clamped. | | --data.synthetic_agentic.tool_catalog_size_per_agent.max | int | Largest value the distribution can produce; samples above are clamped. | | --data.synthetic_agentic.tool_catalog_size_per_agent.mean | float | Mean of the distribution. | | --data.synthetic_agentic.tool_catalog_size_per_agent.std_dev | float | Standard deviation of the distribution. Exclusive with 'variance'. | | --data.synthetic_agentic.tool_catalog_size_per_agent.total_count | int | Total number of values to sample from the distribution. | | --data.synthetic_agentic.tool_catalog_size_per_agent.type | Enum (normal, skew_normal, lognormal, uniform, poisson, fixed) | Shape of the distribution to sample values from. | | --data.synthetic_agentic.tool_catalog_size_per_agent.variance | float | Variance of the distribution. Exclusive with 'std_dev'. | | --data.synthetic_agentic.tool_catalog_size_per_agent.skew | float | Skewness of the distribution. Only used when type is 'skew_normal'. | | --data.synthetic_agentic.parallel_tool_calls_per_step.min | int | Smallest value the distribution can produce; samples below are clamped. | | --data.synthetic_agentic.parallel_tool_calls_per_step.max | int | Largest value the distribution can produce; samples above are clamped. | | --data.synthetic_agentic.parallel_tool_calls_per_step.mean | float | Mean of the distribution. | | --data.synthetic_agentic.parallel_tool_calls_per_step.std_dev | float | Standard deviation of the distribution. Exclusive with 'variance'. | | --data.synthetic_agentic.parallel_tool_calls_per_step.total_count | int | Total number of values to sample from the distribution. | | --data.synthetic_agentic.parallel_tool_calls_per_step.type | Enum (normal, skew_normal, lognormal, uniform, poisson, fixed) | Shape of the distribution to sample values from. | | --data.synthetic_agentic.parallel_tool_calls_per_step.variance | float | Variance of the distribution. Exclusive with 'std_dev'. | | --data.synthetic_agentic.parallel_tool_calls_per_step.skew | float | Skewness of the distribution. Only used when type is 'skew_normal'. | | --data.synthetic_agentic.tool_call_latency_sec.min | int | Smallest value the distribution can produce; samples below are clamped. | | --data.synthetic_agentic.tool_call_latency_sec.max | int | Largest value the distribution can produce; samples above are clamped. | | --data.synthetic_agentic.tool_call_latency_sec.mean | float | Mean of the distribution. | | --data.synthetic_agentic.tool_call_latency_sec.std_dev | float | Standard deviation of the distribution. Exclusive with 'variance'. | | --data.synthetic_agentic.tool_call_latency_sec.total_count | int | Total number of values to sample from the distribution. | | --data.synthetic_agentic.tool_call_latency_sec.type | Enum (normal, skew_normal, lognormal, uniform, poisson, fixed) | Shape of the distribution to sample values from. | | --data.synthetic_agentic.tool_call_latency_sec.variance | float | Variance of the distribution. Exclusive with 'std_dev'. | | --data.synthetic_agentic.tool_call_latency_sec.skew | float | Skewness of the distribution. Only used when type is 'skew_normal'. | | --data.synthetic_agentic.user_think_time_sec.min | int | Smallest value the distribution can produce; samples below are clamped. | | --data.synthetic_agentic.user_think_time_sec.max | int | Largest value the distribution can produce; samples above are clamped. | | --data.synthetic_agentic.user_think_time_sec.mean | float | Mean of the distribution. | | --data.synthetic_agentic.user_think_time_sec.std_dev | float | Standard deviation of the distribution. Exclusive with 'variance'. | | --data.synthetic_agentic.user_think_time_sec.total_count | int | Total number of values to sample from the distribution. | | --data.synthetic_agentic.user_think_time_sec.type | Enum (normal, skew_normal, lognormal, uniform, poisson, fixed) | Shape of the distribution to sample values from. | | --data.synthetic_agentic.user_think_time_sec.variance | float | Variance of the distribution. Exclusive with 'std_dev'. | | --data.synthetic_agentic.user_think_time_sec.skew | float | Skewness of the distribution. Only used when type is 'skew_normal'. | | --data.synthetic_agentic.max_model_len | int | Fail-fast context-length ceiling (tokens). When set, a config whose single largest request -- worst-case inputs (system head + tool catalog + accumulated turns + tool loop) plus the output to generate -- would exceed this is rejected at load, instead of 400-ing mid-run. Uses each distribution's clip ceiling (max) as the worst case. Excludes the model's per-message chat-template wrapper (~10-15 tok/msg), so set this at or a little below your model's true window. Omit to skip the check. | | --data.synthetic_agentic.context_compaction.trigger_tokens.min | int | Smallest value the distribution can produce; samples below are clamped. | | --data.synthetic_agentic.context_compaction.trigger_tokens.max | int | Largest value the distribution can produce; samples above are clamped. | | --data.synthetic_agentic.context_compaction.trigger_tokens.mean | float | Mean of the distribution. | | --data.synthetic_agentic.context_compaction.trigger_tokens.std_dev | float | Standard deviation of the distribution. Exclusive with 'variance'. | | --data.synthetic_agentic.context_compaction.trigger_tokens.total_count | int | Total number of values to sample from the distribution. | | --data.synthetic_agentic.context_compaction.trigger_tokens.type | Enum (normal, skew_normal, lognormal, uniform, poisson, fixed) | Shape of the distribution to sample values from. | | --data.synthetic_agentic.context_compaction.trigger_tokens.variance | float | Variance of the distribution. Exclusive with 'std_dev'. | | --data.synthetic_agentic.context_compaction.trigger_tokens.skew | float | Skewness of the distribution. Only used when type is 'skew_normal'. | | --data.synthetic_agentic.context_compaction.target_tokens.min | int | Smallest value the distribution can produce; samples below are clamped. | | --data.synthetic_agentic.context_compaction.target_tokens.max | int | Largest value the distribution can produce; samples above are clamped. | | --data.synthetic_agentic.context_compaction.target_tokens.mean | float | Mean of the distribution. | | --data.synthetic_agentic.context_compaction.target_tokens.std_dev | float | Standard deviation of the distribution. Exclusive with 'variance'. | | --data.synthetic_agentic.context_compaction.target_tokens.total_count | int | Total number of values to sample from the distribution. | | --data.synthetic_agentic.context_compaction.target_tokens.type | Enum (normal, skew_normal, lognormal, uniform, poisson, fixed) | Shape of the distribution to sample values from. | | --data.synthetic_agentic.context_compaction.target_tokens.variance | float | Variance of the distribution. Exclusive with 'std_dev'. | | --data.synthetic_agentic.context_compaction.target_tokens.skew | float | Skewness of the distribution. Only used when type is 'skew_normal'. | | --data.use_chat_template | boolean | Wrap each generated prompt in the tokenizer's chat template as a single user turn before sending it on the completions path, reproducing the request shape of harnesses that benchmark with chat templating enabled. The input length distribution targets the fully templated prompt, so the server-side prefill token count still matches the configured length. Only supported by the 'random' type; setting it with any other type is a config error. | | --load.type | Enum (constant, poisson, trace_replay, concurrent, trace_session_replay) | Load pattern used to schedule requests. | | --load.interval | float | Seconds to wait between stages. | | --load.stages | JSON | Load stages to run in sequence. The stage fields depend on the load type. | | --load.sweep.type | Enum (geometric, linear) | How stage rates are spaced up to the saturation rate: 'geometric' or 'linear'. | | --load.sweep.num_requests | int | Number of requests sent in the initial burst used to find the saturation rate. | | --load.sweep.timeout | float | Time limit in seconds for the saturation probe stage. | | --load.sweep.num_stages | int | Number of load stages to generate. | | --load.sweep.stage_duration | int | Duration of each generated stage in seconds. | | --load.sweep.saturation_percentile | float | Percentile of observed request rates taken as the saturation point. | | --load.num_workers | int | Number of worker processes sending requests. Defaults to the CPU count. | | --load.worker_max_concurrency | int | Maximum concurrent in-flight requests per worker. | | --load.worker_max_tcp_connections | int | Maximum TCP connections per worker. | | --load.trace.file | str | Path to the trace file to replay. | | --load.trace.format | Enum (AzurePublicDataset) | Format of the trace file. | | --load.circuit_breakers | JSON | Names of configured circuit breakers to enable for the run. | | --load.request_timeout | float | Per-request timeout in seconds. | | --load.request_retries | int | How many extra attempts to make when a request fails before response headers were obtained -- a connection refused, reset, or dropped from the pool. Faults raised once a response was established are never retried, since re-sending could mix or discard token-level measurements; neither are timeouts or TLS configuration errors, which fail identically on every attempt. A failed attempt can still have reached the server, so a retry is cheap for the client but not free for the endpoint. Defaults to 0, preserving the historical behavior of failing on the first fault. request_timeout applies per attempt, so worst-case wall time per request becomes (1 + request_retries) * request_timeout plus backoff. | | --load.request_retry_backoff_sec | float | Base delay before the first retry, in seconds. Each further retry doubles it and adds jitter, so attempts do not resynchronize into bursts against a server that is already dropping connections. Ignored when request_retries is 0. | | --load.stage_teardown_grace_seconds | float | How long to let in-flight requests finish after a stage ends or times out, before they are cancelled. Teardown is bounded: once the grace (plus a fixed margin) expires, remaining work is force-cancelled and unresponsive workers are terminated and respawned, so report generation always runs. Set to 0 to cancel in-flight requests immediately at stage end. The teardown window is excluded from the stage's reported end_time and metrics windows; it is reported separately as teardown_duration. | | --load.lora_traffic_split | JSON | Traffic split across LoRA adapters. Splits must sum to 1.0. | | --load.base_seed | int | Base random seed for load generation. Defaults to the current time. | | --metrics.type | Enum (prometheus, default) | Metrics client used to collect server-side metrics. | | --metrics.prometheus.scrape_interval | int | Scrape interval of the Prometheus server in seconds. | | --metrics.prometheus.url | string | URL of the Prometheus server to query. | | --metrics.prometheus.filters | JSON | PromQL label matchers (e.g. 'namespace="default"') applied to every metric query. | | --metrics.prometheus.google_managed | boolean | Query Google Cloud Managed Service for Prometheus instead of a self-hosted server. | | --report.request_lifecycle.summary | boolean | Generate a summary report across the whole run. | | --report.request_lifecycle.per_stage | boolean | Generate a report for each load stage. | | --report.request_lifecycle.per_request | boolean | Generate a report with per-request details. | | --report.request_lifecycle.per_request_fields.request | boolean | Include the raw request payload in per-request report entries. | | --report.request_lifecycle.per_request_fields.response | boolean | Include the raw response payload in per-request report entries. | | --report.request_lifecycle.per_request_fields.info | boolean | Include response metadata in per-request report entries. | | --report.request_lifecycle.per_request_fields.response_chunks | boolean | Include streamed response chunks in per-request metadata. | | --report.request_lifecycle.per_request_fields.computed_metrics | boolean | Include computed per-request latency and token metrics in report entries. | | --report.request_lifecycle.per_adapter | boolean | Generate a report for each LoRA adapter. | | --report.request_lifecycle.per_adapter_stage | boolean | Generate a report for each LoRA adapter within each load stage. | | --report.request_lifecycle.percentiles | JSON | Percentiles reported for each metric. | | --report.request_lifecycle.use_server_output_tokens | boolean | Use the server-reported output token counts in metrics instead of tokenizing the response text. | | --report.request_lifecycle.max_error_messages | int | Cap on the number of distinct example error messages retained per error label in the failure report, and per substitution entry. | | --report.prometheus.summary | boolean | Generate a summary report across the whole run. | | --report.prometheus.per_stage | boolean | Generate a report for each load stage. | | --report.session_lifecycle.summary | boolean | Generate a summary report across the whole run. | | --report.session_lifecycle.per_stage | boolean | Generate a report for each load stage. | | --report.session_lifecycle.per_session | boolean | Generate a report with per-session details. | | --report.goodput.constraints | JSON | SLO thresholds in seconds that a request must meet to count as good. Keys: 'ttft', 'tpot', 'itl', 'ntpot', 'request_latency'. | | --storage.local_storage.path | str | Directory or object key prefix where report files are written. '{timestamp}' is replaced with the run's start time. | | --storage.local_storage.report_file_prefix | str | Prefix added to every report file name. | | --storage.google_cloud_storage.path | str | Directory or object key prefix where report files are written. '{timestamp}' is replaced with the run's start time. | | --storage.google_cloud_storage.report_file_prefix | str | Prefix added to every report file name. | | --storage.google_cloud_storage.bucket_name | str | Name of the Google Cloud Storage bucket where reports are uploaded. | | --storage.simple_storage_service.path | str | Directory or object key prefix where report files are written. '{timestamp}' is replaced with the run's start time. | | --storage.simple_storage_service.report_file_prefix | str | Prefix added to every report file name. | | --storage.simple_storage_service.bucket_name | str | Name of the S3 bucket where reports are uploaded. | | --storage.simple_storage_service.endpoint_url | str | Custom endpoint URL, for S3-compatible object stores. | | --storage.simple_storage_service.region_name | str | AWS region of the bucket. | | --storage.simple_storage_service.addressing_style | string | S3 addressing style: 'auto', 'virtual' (bucket in hostname) or 'path'. | | --server.type | Enum (vllm, sglang, tgi, mock) | Type of model server being benchmarked. | | --server.model_name | str | Model name sent in each request. Auto-detected from the server if unset. | | --server.base_url | str | Base URL of the model server, e.g. 'http://localhost:8000'. | | --server.ignore_eos | boolean | Ask the server to keep generating past the end-of-sequence token so outputs hit the requested length. | | --server.api_key | str | API key sent as a bearer token with each request. | | --server.cert_path | str | Path to a client TLS certificate file. | | --server.key_path | str | Path to the private key for the client TLS certificate. | | --tokenizer.pretrained_model_name_or_path | str | HuggingFace model name or local path of the tokenizer to load. | | --tokenizer.trust_remote_code | boolean | Allow the tokenizer to execute code from its repository when loading. | | --tokenizer.token | str | HuggingFace access token used to download the tokenizer. | | --tokenizer.load_timeout | float | Deadline in seconds for loading the tokenizer, including any download from Hugging Face Hub. Null disables the deadline. | | --circuit_breakers | JSON | Circuit breakers that stop the run when observed metrics cross configured thresholds. |