Dataset Viewer
Auto-converted to Parquet Duplicate
The dataset viewer is not available for this split.
Cannot load the dataset split (in streaming mode) to extract the first rows.
Error code:   StreamingRowsError
Exception:    CastError
Message:      Couldn't cast
schema_version: int64
run_id: string
group_id: string
run_index: int64
label: string
started_at: string
source: string
pr_number: int64
branch: string
commit_id: string
commit_date: string
commit_subject: string
ab_id: string
ab_side: string
ab_round: int64
model_id: string
bench_args: string
scenario: string
attn_implementation: string
compile_mode: string
kernelize: bool
continuous_batching: bool
tp: bool
batch_size: int64
sequence_length: int64
num_tokens_to_generate: int64
warmup_iterations: int64
measurement_iterations: int64
success: bool
n_samples: int64
e2e_ms_median: double
e2e_ms_mean: double
e2e_ms_min: double
e2e_ms_max: double
e2e_ms_p90: double
e2e_ms_cov_pct: double
ttft_ms_median: double
ttft_ms_mean: double
ttft_ms_min: double
ttft_ms_max: double
ttft_ms_p90: double
ttft_ms_cov_pct: double
itl_ms_median: double
itl_ms_mean: double
itl_ms_min: double
itl_ms_max: double
itl_ms_p90: double
itl_ms_cov_pct: double
itl_first_half_ms_median: double
itl_first_half_ms_mean: double
itl_first_half_ms_min: double
itl_first_half_ms_max: double
itl_first_half_ms_p90: double
itl_first_half_ms_cov_pct: double
itl_second_half_ms_median: double
itl_second_half_ms_mean: double
itl_second_half_ms_min: double
itl_second_half_ms_max: double
itl_second_half_ms_p90: double
itl_second_half_ms_cov_pct: double
tokens_per_s_median: double
tokens_per_s_mean: double
tokens_per_s_min: double
tokens_per_s_max: double
tokens_per_s_p90: double
tokens_per_s_cov_pct: double
startup_load_time_s: double
startup_first_generate_ms: double
startup_warmup_first_ms: double
startup_warmup_last_ms: double
startup_first_generate_peak_memory_mb: double
startup_measurement_peak_memory_mb: double
startup_first_generate_compiled_graphs: int64
startup_first_generate_graph_breaks: int64
startup_first_generate_fxgraph_cache_hits: int64
startup_first_generate_fxgraph_cache_misses: int64
startup_warmup_compiled_graphs: int64
startup_warmup_graph_breaks: int64
startup_warmup_fxgraph_cache_hits: int64
startup_warmup_fxgraph_cache_misses: int64
startup_measurement_compiled_graphs: int64
startup_measurement_graph_breaks: int64
startup_measurement_fxgraph_cache_hits: int64
startup_measurement_fxgraph_cache_misses: int64
series_key: string
instance_type: string
cpu_model: string
kernel: string
gpu_index: int64
numa_node: int64
cpus: string
gpu_name: string
driver_version: string
sm_clock_mhz: int64
mem_clock_mhz: int64
power_limit_w: double
gpu_temp_c_before: int64
python_version: string
torch_version: string
torch_cuda: string
transformers_version: string
kernels_version: string
numa_balancing: string
aslr: string
probe_exit_code: int64
probe_elapsed_s: double
probe_command_cpu_s: double
probe_foreign_cpu_s: double
probe_runqueue_wait_ms: double
probe_voluntary_ctx_switches: int64
probe_involuntary_ctx_switches: int64
probe_interrupts_on_cpus: int64
probe_gpu_util_mean_pct: double
probe_gpu_avg_power_w: double
probe_power_throttle_ms: double
probe_thermal_throttle_ms: double
to
{'schema_version': Value('int64'), 'run_id': Value('string'), 'group_id': Value('string'), 'run_index': Value('int64'), 'label': Value('string'), 'started_at': Value('string'), 'source': Value('string'), 'pr_number': Value('int64'), 'branch': Value('string'), 'commit_id': Value('string'), 'commit_date': Value('string'), 'commit_subject': Value('string'), 'model_id': Value('string'), 'bench_args': Value('string'), 'scenario': Value('string'), 'attn_implementation': Value('string'), 'compile_mode': Value('string'), 'kernelize': Value('bool'), 'continuous_batching': Value('bool'), 'tp': Value('bool'), 'batch_size': Value('int64'), 'sequence_length': Value('int64'), 'num_tokens_to_generate': Value('int64'), 'warmup_iterations': Value('int64'), 'measurement_iterations': Value('int64'), 'success': Value('bool'), 'n_samples': Value('int64'), 'e2e_ms_median': Value('float64'), 'e2e_ms_mean': Value('float64'), 'e2e_ms_min': Value('float64'), 'e2e_ms_max': Value('float64'), 'e2e_ms_p90': Value('float64'), 'e2e_ms_cov_pct': Value('float64'), 'ttft_ms_median': Value('float64'), 'ttft_ms_mean': Value('float64'), 'ttft_ms_min': Value('float64'), 'ttft_ms_max': Value('float64'), 'ttft_ms_p90': Value('float64'), 'ttft_ms_cov_pct': Value('float64'), 'itl_ms_median': Value('float64'), 'itl_ms_mean': Value('float64'), 'itl_ms_min': Value('float64'), 'itl_ms_max': Value('float64'), 'itl_ms_p90': Value('float64'), 'itl_ms_cov_pct': Value('float64'), 'tokens_per_s_median': Value('float64'), 'tokens_per_s_mean': Value('float64'), 'tokens_per_s_min': Value('float64'), 'tokens_per_s_max': Value('float64'), 'tokens_per_s_p90': Value('float64'), 'tokens_per_s_cov_pct': Value('float64'), 'series_key': Value('string'), 'instance_type': Value('string'), 'cpu_model': Value('string'), 'kernel': Value('string'), 'gpu_index': Value('int64'), 'numa_node': Value('int64'), 'cpus': Value('string'), 'gpu_name': Value('string'), 'driver_version': Value('string'), 'sm_clock_mhz': Value('int64'), 'mem_clock_mhz': Value('int64'), 'power_limit_w': Value('float64'), 'gpu_temp_c_before': Value('int64'), 'python_version': Value('string'), 'torch_version': Value('string'), 'torch_cuda': Value('string'), 'transformers_version': Value('string'), 'kernels_version': Value('string'), 'numa_balancing': Value('string'), 'aslr': Value('string'), 'probe_exit_code': Value('int64'), 'probe_elapsed_s': Value('float64'), 'probe_command_cpu_s': Value('float64'), 'probe_foreign_cpu_s': Value('float64'), 'probe_runqueue_wait_ms': Value('float64'), 'probe_voluntary_ctx_switches': Value('int64'), 'probe_involuntary_ctx_switches': Value('int64'), 'probe_interrupts_on_cpus': Value('int64'), 'probe_gpu_util_mean_pct': Value('float64'), 'probe_gpu_avg_power_w': Value('float64'), 'probe_power_throttle_ms': Value('float64'), 'probe_thermal_throttle_ms': Value('float64')}
because column names don't match
Traceback:    Traceback (most recent call last):
                File "/src/services/worker/src/worker/utils.py", line 149, in get_rows_or_raise
                  return get_rows(
                      dataset=dataset,
                  ...<4 lines>...
                      column_names=column_names,
                  )
                File "/src/libs/libcommon/src/libcommon/utils.py", line 274, in decorator
                  return func(*args, **kwargs)
                File "/src/services/worker/src/worker/utils.py", line 129, in get_rows
                  rows_plus_one = list(itertools.islice(safe_iter(ds, dataset=dataset), rows_max_number + 1))
                File "/src/services/worker/src/worker/utils.py", line 544, in safe_iter
                  yield from ds.decode(False) if ds.features else ds
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 3074, in __iter__
                  for key, example in ex_iterable:
                                      ^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2578, in __iter__
                  for key, pa_table in self._iter_arrow():
                                       ~~~~~~~~~~~~~~~~^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2603, in _iter_arrow
                  for key, pa_table in self.ex_iterable._iter_arrow():
                                       ~~~~~~~~~~~~~~~~~~~~~~~~~~~~^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 577, in _iter_arrow
                  for key, pa_table in iterator:
                                       ^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 459, in _iter_arrow
                  for key, pa_table in self.generate_tables_fn(**gen_kwags):
                                       ~~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/parquet/parquet.py", line 220, in _generate_tables
                  yield Key(file_idx, batch_idx), self._cast_table(pa_table)
                                                  ~~~~~~~~~~~~~~~~^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/parquet/parquet.py", line 156, in _cast_table
                  pa_table = table_cast(pa_table, self.info.features.arrow_schema)
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2394, in table_cast
                  return cast_table_to_schema(table, schema)
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2322, in cast_table_to_schema
                  raise CastError(
                  ...<3 lines>...
                  )
              datasets.table.CastError: Couldn't cast
              schema_version: int64
              run_id: string
              group_id: string
              run_index: int64
              label: string
              started_at: string
              source: string
              pr_number: int64
              branch: string
              commit_id: string
              commit_date: string
              commit_subject: string
              ab_id: string
              ab_side: string
              ab_round: int64
              model_id: string
              bench_args: string
              scenario: string
              attn_implementation: string
              compile_mode: string
              kernelize: bool
              continuous_batching: bool
              tp: bool
              batch_size: int64
              sequence_length: int64
              num_tokens_to_generate: int64
              warmup_iterations: int64
              measurement_iterations: int64
              success: bool
              n_samples: int64
              e2e_ms_median: double
              e2e_ms_mean: double
              e2e_ms_min: double
              e2e_ms_max: double
              e2e_ms_p90: double
              e2e_ms_cov_pct: double
              ttft_ms_median: double
              ttft_ms_mean: double
              ttft_ms_min: double
              ttft_ms_max: double
              ttft_ms_p90: double
              ttft_ms_cov_pct: double
              itl_ms_median: double
              itl_ms_mean: double
              itl_ms_min: double
              itl_ms_max: double
              itl_ms_p90: double
              itl_ms_cov_pct: double
              itl_first_half_ms_median: double
              itl_first_half_ms_mean: double
              itl_first_half_ms_min: double
              itl_first_half_ms_max: double
              itl_first_half_ms_p90: double
              itl_first_half_ms_cov_pct: double
              itl_second_half_ms_median: double
              itl_second_half_ms_mean: double
              itl_second_half_ms_min: double
              itl_second_half_ms_max: double
              itl_second_half_ms_p90: double
              itl_second_half_ms_cov_pct: double
              tokens_per_s_median: double
              tokens_per_s_mean: double
              tokens_per_s_min: double
              tokens_per_s_max: double
              tokens_per_s_p90: double
              tokens_per_s_cov_pct: double
              startup_load_time_s: double
              startup_first_generate_ms: double
              startup_warmup_first_ms: double
              startup_warmup_last_ms: double
              startup_first_generate_peak_memory_mb: double
              startup_measurement_peak_memory_mb: double
              startup_first_generate_compiled_graphs: int64
              startup_first_generate_graph_breaks: int64
              startup_first_generate_fxgraph_cache_hits: int64
              startup_first_generate_fxgraph_cache_misses: int64
              startup_warmup_compiled_graphs: int64
              startup_warmup_graph_breaks: int64
              startup_warmup_fxgraph_cache_hits: int64
              startup_warmup_fxgraph_cache_misses: int64
              startup_measurement_compiled_graphs: int64
              startup_measurement_graph_breaks: int64
              startup_measurement_fxgraph_cache_hits: int64
              startup_measurement_fxgraph_cache_misses: int64
              series_key: string
              instance_type: string
              cpu_model: string
              kernel: string
              gpu_index: int64
              numa_node: int64
              cpus: string
              gpu_name: string
              driver_version: string
              sm_clock_mhz: int64
              mem_clock_mhz: int64
              power_limit_w: double
              gpu_temp_c_before: int64
              python_version: string
              torch_version: string
              torch_cuda: string
              transformers_version: string
              kernels_version: string
              numa_balancing: string
              aslr: string
              probe_exit_code: int64
              probe_elapsed_s: double
              probe_command_cpu_s: double
              probe_foreign_cpu_s: double
              probe_runqueue_wait_ms: double
              probe_voluntary_ctx_switches: int64
              probe_involuntary_ctx_switches: int64
              probe_interrupts_on_cpus: int64
              probe_gpu_util_mean_pct: double
              probe_gpu_avg_power_w: double
              probe_power_throttle_ms: double
              probe_thermal_throttle_ms: double
              to
              {'schema_version': Value('int64'), 'run_id': Value('string'), 'group_id': Value('string'), 'run_index': Value('int64'), 'label': Value('string'), 'started_at': Value('string'), 'source': Value('string'), 'pr_number': Value('int64'), 'branch': Value('string'), 'commit_id': Value('string'), 'commit_date': Value('string'), 'commit_subject': Value('string'), 'model_id': Value('string'), 'bench_args': Value('string'), 'scenario': Value('string'), 'attn_implementation': Value('string'), 'compile_mode': Value('string'), 'kernelize': Value('bool'), 'continuous_batching': Value('bool'), 'tp': Value('bool'), 'batch_size': Value('int64'), 'sequence_length': Value('int64'), 'num_tokens_to_generate': Value('int64'), 'warmup_iterations': Value('int64'), 'measurement_iterations': Value('int64'), 'success': Value('bool'), 'n_samples': Value('int64'), 'e2e_ms_median': Value('float64'), 'e2e_ms_mean': Value('float64'), 'e2e_ms_min': Value('float64'), 'e2e_ms_max': Value('float64'), 'e2e_ms_p90': Value('float64'), 'e2e_ms_cov_pct': Value('float64'), 'ttft_ms_median': Value('float64'), 'ttft_ms_mean': Value('float64'), 'ttft_ms_min': Value('float64'), 'ttft_ms_max': Value('float64'), 'ttft_ms_p90': Value('float64'), 'ttft_ms_cov_pct': Value('float64'), 'itl_ms_median': Value('float64'), 'itl_ms_mean': Value('float64'), 'itl_ms_min': Value('float64'), 'itl_ms_max': Value('float64'), 'itl_ms_p90': Value('float64'), 'itl_ms_cov_pct': Value('float64'), 'tokens_per_s_median': Value('float64'), 'tokens_per_s_mean': Value('float64'), 'tokens_per_s_min': Value('float64'), 'tokens_per_s_max': Value('float64'), 'tokens_per_s_p90': Value('float64'), 'tokens_per_s_cov_pct': Value('float64'), 'series_key': Value('string'), 'instance_type': Value('string'), 'cpu_model': Value('string'), 'kernel': Value('string'), 'gpu_index': Value('int64'), 'numa_node': Value('int64'), 'cpus': Value('string'), 'gpu_name': Value('string'), 'driver_version': Value('string'), 'sm_clock_mhz': Value('int64'), 'mem_clock_mhz': Value('int64'), 'power_limit_w': Value('float64'), 'gpu_temp_c_before': Value('int64'), 'python_version': Value('string'), 'torch_version': Value('string'), 'torch_cuda': Value('string'), 'transformers_version': Value('string'), 'kernels_version': Value('string'), 'numa_balancing': Value('string'), 'aslr': Value('string'), 'probe_exit_code': Value('int64'), 'probe_elapsed_s': Value('float64'), 'probe_command_cpu_s': Value('float64'), 'probe_foreign_cpu_s': Value('float64'), 'probe_runqueue_wait_ms': Value('float64'), 'probe_voluntary_ctx_switches': Value('int64'), 'probe_involuntary_ctx_switches': Value('int64'), 'probe_interrupts_on_cpus': Value('int64'), 'probe_gpu_util_mean_pct': Value('float64'), 'probe_gpu_avg_power_w': Value('float64'), 'probe_power_throttle_ms': Value('float64'), 'probe_thermal_throttle_ms': Value('float64')}
              because column names don't match

Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.

transformers performance tracking

Results of the transformers-perf benchmarks of transformers on a dedicated, tuned host (one GPU per run, clocks locked, CPUs pinned on the GPU's NUMA node, background services masked). Each run is published by tperf record publish (generation) or tperf record publish-processors.

Layout

  • data/<YYYY-MM>/<run_id>.parquet: one row per (run, scenario)
  • processors/<YYYY-MM>/<group>.parquet: multimodal processor and encoder runs, one row per (run, scenario)
  • raw/<YYYY-MM>/<run_id>/: the files the rows were built from (manifest.json, probe.json, harness results)

A run is one run_benchmarks.py invocation. Runs repeated back to back share a group_id.

Columns

Column(s) Meaning
run_id, group_id, run_index, label, started_at run identity
source (main, pr, manual), pr_number, branch, commit_id, commit_date, commit_subject code under test
model_id, scenario, attn_implementation, compile_mode, kernelize, continuous_batching, tp, batch_size, sequence_length, num_tokens_to_generate, warmup_iterations, measurement_iterations, bench_args workload
harness_commit harness version: the last commit of huggingface/transformers-perf that changed it (-dirty: local changes)
success, n_samples a failed scenario has success = false and no metrics
e2e_ms_*, ttft_ms_*, itl_ms_*, tokens_per_s_* median, mean, min, max, p90, cov_pct (coefficient of variation, %) over the measurement iterations
instance_type, cpu_model, kernel, gpu_name, gpu_index, numa_node, cpus, driver_version, sm_clock_mhz, mem_clock_mhz, power_limit_w, gpu_temp_c_before hardware and clocks
python_version, torch_version, torch_cuda, transformers_version, kernels_version, numa_balancing, aslr software
probe_* interference over the whole run, read from counters before and after (nothing runs during measurements): CPU time on the benchmark's CPUs used by others (foreign_cpu_s), run-queue wait, context switches, interrupts, GPU utilization, average power, power/thermal throttling
series_key hash of model, scenario, instance type, GPU and SM clock: rows with the same key are comparable

Querying

import duckdb

duckdb.sql("""
    SELECT started_at, source, branch, commit_id[:10] AS commit, scenario, e2e_ms_median, ttft_ms_median, e2e_ms_cov_pct
    FROM 'hf://datasets/hf-internal-testing/transformers-perf/data/*/*.parquet'
    WHERE success
    ORDER BY series_key, started_at
""").show()

Compare a PR with main: take the median of e2e_ms_median per series_key for source = 'pr' AND pr_number = N and for source = 'main' at the PR's base commit, and treat differences smaller than the cov_pct of either side as noise.

Downloads last month
25