Skip to main content

study-config

Configuration Reference

Full reference for all ExperimentConfig fields. All fields except model are optional and have sensible defaults.

Sections:

Top-Level Fields

FieldTypeDefaultDescription
taskTaskConfig(see section)Task configuration: model, dataset, workload shape
engineEngine(see section)Inference engine
measurementMeasurementConfig(see section)Measurement methodology: warmup, baseline, energy sampling
sampling_preset'deterministic''standard''creative'
transformersConfigNonenull
vllmConfigNonenull
tensorrtConfigNonenull
harnessHarnessConfigNonenull
passthrough_kwargsdictNonenull

Warmup (warmup:)

FieldTypeDefaultDescription
enabledbooleantrueEnable warmup phase
n_promptsinteger5Number of full-length warmup prompts in fixed mode
thermal_floor_secondsnumber60.0Minimum seconds to wait after warmup before measuring (thermal stabilisation). Minimum 30s enforced.
convergence_detectionbooleanfalseEnable CV-based adaptive convergence (governed by min_prompts, max_prompts, cv_threshold, window_size)
cv_thresholdnumber0.05CV target for convergence (only used when convergence_detection=True)
max_promptsinteger20Maximum warmup prompts when CV mode is on (safety cap)
window_sizeinteger3Sliding window size for CV calculation (3 balances responsiveness and stability)
min_promptsinteger5Minimum prompts before checking convergence (warm start)

Baseline (baseline:)

FieldTypeDefaultDescription
enabledbooleantrueEnable baseline power measurement
duration_secondsnumber30.0Baseline measurement duration in seconds
strategy'cached''validated''fresh'
cache_ttl_secondsnumber7200.0How long a cached baseline remains valid before re-measurement, in seconds. Only used with strategy='cached' or 'validated'.
validation_intervalinteger5Re-validate baseline every N experiments. Only used with strategy='validated'.
drift_thresholdnumber0.1Power drift threshold (fraction) to trigger re-measurement. Only used with strategy='validated'.

Transformers Engine (transformers:)

FieldTypeDefaultDescription
engine_paramsEngineParamsNonenull
sampling_paramsSamplingParamsNonenull

Transformers Engine Params (transformers.engine_params:)

FieldTypeDefaultDescription
dtypeanyNonenull
attn_implementationanyNonenull
load_in_4bitanyNonenull
load_in_8bitanyNonenull
bnb_4bit_compute_dtypeanyNonenull
bnb_4bit_quant_typeanyNonenull
bnb_4bit_use_double_quantanyNonenull
use_cachebooleanNonenull
cache_implementationstringNonenull
num_beamsintegerNonenull
early_stoppingbooleanNonenull
length_penaltynumberNonenull
no_repeat_ngram_sizeintegerNonenull
prompt_lookup_num_tokensintegerNonenull
device_mapanyNonenull
max_memoryanyNonenull
low_cpu_mem_usageanyNonenull
tp_plananyNonenull
tp_sizeanyNonenull

Transformers Sampling Params (transformers.sampling_params:)

FieldTypeDefaultDescription
temperaturenumberNonenull
do_samplebooleanNonenull
top_kintegerNonenull
top_pnumberNonenull
repetition_penaltynumberNonenull
min_pnumberNonenull
min_new_tokensintegerNonenull

vLLM Engine (vllm:)

FieldTypeDefaultDescription
engine_paramsEngineParamsNonenull
sampling_paramsSamplingParamsNonenull

vLLM Engine Params (vllm.engine_params:)

FieldTypeDefaultDescription
dtype'auto''half''float16'
gpu_memory_utilizationnumberNone0.9
cpu_offload_gbnumberNone0
block_sizeintegerNonenull
kv_cache_dtype'auto''float16''bfloat16'
enforce_eagerbooleanNonefalse
enable_chunked_prefillbooleanNonenull
max_num_seqsintegerNonenull
max_num_batched_tokensintegerNonenull
max_model_lenintegerNonenull
tensor_parallel_sizeintegerNone1
pipeline_parallel_sizeintegerNone1
distributed_executor_backendanyNonenull
enable_prefix_cachingbooleanNonenull
quantizationanyNonenull
speculative_configSpeculativeConfigNonenull
offload_group_sizeintegerNone0
offload_num_in_groupintegerNone1
offload_prefetch_stepintegerNone1
offload_paramsanyNone[]
disable_custom_all_reducebooleanNonefalse
kv_cache_memory_bytesintegerNonenull
compilation_configCompilationConfigNonenull
attentionanyNonenull
beam_searchanyNonenull

vLLM Sampling Params (vllm.sampling_params:)

FieldTypeDefaultDescription
temperaturenumberNone1.0
top_kintegerNone0
top_pnumberNone1.0
repetition_penaltynumberNone1.0
min_pnumberNone0.0
min_tokensintegerNone0
presence_penaltynumberNone0.0
frequency_penaltynumberNone0.0
ignore_eosbooleanNonefalse
nintegerNone1

TensorRT-LLM Engine (tensorrt:)

FieldTypeDefaultDescription
engine_paramsEngineParamsNonenull
sampling_paramsSamplingParamsNonenull

TensorRT-LLM Engine Params (tensorrt.engine_params:)

FieldTypeDefaultDescription
max_batch_sizeintegerNonenull
tensor_parallel_sizeintegerNone1
pipeline_parallel_sizeintegerNone1
max_input_lenintegerNonenull
max_seq_lenintegerNonenull
max_num_tokensintegerNonenull
dtypestringNoneauto
fast_buildbooleanNonefalse
backendstringNonenull
quant_configanyNonenull
kv_cache_configanyNonenull
scheduler_configanyNonenull

TensorRT-LLM Sampling Params (tensorrt.sampling_params:)

FieldTypeDefaultDescription
temperaturenumberNonenull
top_kintegerNonenull
top_pnumberNonenull
repetition_penaltynumberNonenull
min_pnumberNonenull
min_tokensintegerNonenull
nintegerNone1
ignore_eosbooleanNonefalse

Harness Overrides (harness:)

FieldTypeDefaultDescription
transformersTransformersHarnessNonenull

Transformers Harness (harness.transformers:)

FieldTypeDefaultDescription
batch_sizeintegerNonenull
torch_compilebooleanNonenull
torch_compile_modestringNonenull
torch_compile_backendstringNonenull
allow_tf32booleanNonenull
autocast_enabledbooleanNonenull
autocast_dtype'float16''bfloat16'None