Hugging Face's logo

Datasets:
LostGentoo
/
gentoo-small-model-throughput-vllm

Tasks:
Text Generation
Languages:
English
Tags:
benchmark
throughput
quality
lighteval
vllm
qwen3.5
License:
Dataset card Files
xet
 Community
Dataset Viewer
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
model_label: string
url: string
model: string
mode: string
max_tokens: int64
repeats: int64
context_sweep: list<item: struct<context_tier_tokens: int64, sample_prompt_tokens_qwen_flat: int64, concurrency: in (... 513 chars omitted)
  child 0, item: struct<context_tier_tokens: int64, sample_prompt_tokens_qwen_flat: int64, concurrency: int64, ok: in (... 501 chars omitted)
      child 0, context_tier_tokens: int64
      child 1, sample_prompt_tokens_qwen_flat: int64
      child 2, concurrency: int64
      child 3, ok: int64
      child 4, errors: int64
      child 5, error_rate: double
      child 6, total_wall_s: double
      child 7, avg_ttft_s: double
      child 8, sum_decode_tps_all_streams: double
      child 9, aggregate_tokens_per_s: null
      child 10, total_prompt_tokens: int64
      child 11, total_completion_tokens: int64
      child 12, per_request: list<item: struct<ok: bool, wall_s: double, error: null, prompt_tokens: int64, completion_tokens: in (... 94 chars omitted)
          child 0, item: struct<ok: bool, wall_s: double, error: null, prompt_tokens: int64, completion_tokens: int64, ttft_s (... 82 chars omitted)
              child 0, ok: bool
              child 1, wall_s: double
              child 2, error: null
              child 3, prompt_tokens: int64
              child 4, completion_tokens: int64
              child 5, ttft_s: double
              child 6, decode_tps: double
              child 7, tokens_per_s: null
              child 8, milestone_s
...
currency_sweep: list<item: struct<concurrency: int64, ok: int64, errors: int64, error_rate: double, total_wall_s: do (... 446 chars omitted)
  child 0, item: struct<concurrency: int64, ok: int64, errors: int64, error_rate: double, total_wall_s: double, avg_t (... 434 chars omitted)
      child 0, concurrency: int64
      child 1, ok: int64
      child 2, errors: int64
      child 3, error_rate: double
      child 4, total_wall_s: double
      child 5, avg_ttft_s: double
      child 6, sum_decode_tps_all_streams: double
      child 7, aggregate_tokens_per_s: null
      child 8, total_prompt_tokens: int64
      child 9, total_completion_tokens: int64
      child 10, per_request: list<item: struct<ok: bool, wall_s: double, error: null, prompt_tokens: int64, completion_tokens: in (... 94 chars omitted)
          child 0, item: struct<ok: bool, wall_s: double, error: null, prompt_tokens: int64, completion_tokens: int64, ttft_s (... 82 chars omitted)
              child 0, ok: bool
              child 1, wall_s: double
              child 2, error: null
              child 3, prompt_tokens: int64
              child 4, completion_tokens: int64
              child 5, ttft_s: double
              child 6, decode_tps: double
              child 7, tokens_per_s: null
              child 8, milestone_s: struct<50: double>
                  child 0, 50: double
      child 11, median_decode_tps: double
      child 12, p50_time_to_50_tokens_s: double
      child 13, mean_decode_tps: double
to
{'model_label': Value('string'), 'url': Value('string'), 'model': Value('string'), 'mode': Value('string'), 'max_tokens': Value('int64'), 'repeats': Value('int64'), 'concurrency_sweep': List({'concurrency': Value('int64'), 'ok': Value('int64'), 'errors': Value('int64'), 'error_rate': Value('float64'), 'total_wall_s': Value('float64'), 'avg_ttft_s': Value('float64'), 'sum_decode_tps_all_streams': Value('float64'), 'aggregate_tokens_per_s': Value('null'), 'total_prompt_tokens': Value('int64'), 'total_completion_tokens': Value('int64'), 'per_request': List({'ok': Value('bool'), 'wall_s': Value('float64'), 'error': Value('null'), 'prompt_tokens': Value('int64'), 'completion_tokens': Value('int64'), 'ttft_s': Value('float64'), 'decode_tps': Value('float64'), 'tokens_per_s': Value('null'), 'milestone_s': {'50': Value('float64')}}), 'median_decode_tps': Value('float64'), 'p50_time_to_50_tokens_s': Value('float64'), 'mean_decode_tps': 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 272, 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 489, 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 2818, in __iter__
                  for key, example in ex_iterable:
                                      ^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2355, in __iter__
                  for key, pa_table in self._iter_arrow():
                                       ~~~~~~~~~~~~~~~~^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2380, 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 536, in _iter_arrow
                  for key, pa_table in iterator:
                                       ^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 419, 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/json/json.py", line 343, in _generate_tables
                  self._cast_table(pa_table, json_field_paths=json_field_paths),
                  ~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 132, 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 2369, in table_cast
                  return cast_table_to_schema(table, schema)
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2297, in cast_table_to_schema
                  raise CastError(
                  ...<3 lines>...
                  )
              datasets.table.CastError: Couldn't cast
              model_label: string
              url: string
              model: string
              mode: string
              max_tokens: int64
              repeats: int64
              context_sweep: list<item: struct<context_tier_tokens: int64, sample_prompt_tokens_qwen_flat: int64, concurrency: in (... 513 chars omitted)
                child 0, item: struct<context_tier_tokens: int64, sample_prompt_tokens_qwen_flat: int64, concurrency: int64, ok: in (... 501 chars omitted)
                    child 0, context_tier_tokens: int64
                    child 1, sample_prompt_tokens_qwen_flat: int64
                    child 2, concurrency: int64
                    child 3, ok: int64
                    child 4, errors: int64
                    child 5, error_rate: double
                    child 6, total_wall_s: double
                    child 7, avg_ttft_s: double
                    child 8, sum_decode_tps_all_streams: double
                    child 9, aggregate_tokens_per_s: null
                    child 10, total_prompt_tokens: int64
                    child 11, total_completion_tokens: int64
                    child 12, per_request: list<item: struct<ok: bool, wall_s: double, error: null, prompt_tokens: int64, completion_tokens: in (... 94 chars omitted)
                        child 0, item: struct<ok: bool, wall_s: double, error: null, prompt_tokens: int64, completion_tokens: int64, ttft_s (... 82 chars omitted)
                            child 0, ok: bool
                            child 1, wall_s: double
                            child 2, error: null
                            child 3, prompt_tokens: int64
                            child 4, completion_tokens: int64
                            child 5, ttft_s: double
                            child 6, decode_tps: double
                            child 7, tokens_per_s: null
                            child 8, milestone_s
              ...
              currency_sweep: list<item: struct<concurrency: int64, ok: int64, errors: int64, error_rate: double, total_wall_s: do (... 446 chars omitted)
                child 0, item: struct<concurrency: int64, ok: int64, errors: int64, error_rate: double, total_wall_s: double, avg_t (... 434 chars omitted)
                    child 0, concurrency: int64
                    child 1, ok: int64
                    child 2, errors: int64
                    child 3, error_rate: double
                    child 4, total_wall_s: double
                    child 5, avg_ttft_s: double
                    child 6, sum_decode_tps_all_streams: double
                    child 7, aggregate_tokens_per_s: null
                    child 8, total_prompt_tokens: int64
                    child 9, total_completion_tokens: int64
                    child 10, per_request: list<item: struct<ok: bool, wall_s: double, error: null, prompt_tokens: int64, completion_tokens: in (... 94 chars omitted)
                        child 0, item: struct<ok: bool, wall_s: double, error: null, prompt_tokens: int64, completion_tokens: int64, ttft_s (... 82 chars omitted)
                            child 0, ok: bool
                            child 1, wall_s: double
                            child 2, error: null
                            child 3, prompt_tokens: int64
                            child 4, completion_tokens: int64
                            child 5, ttft_s: double
                            child 6, decode_tps: double
                            child 7, tokens_per_s: null
                            child 8, milestone_s: struct<50: double>
                                child 0, 50: double
                    child 11, median_decode_tps: double
                    child 12, p50_time_to_50_tokens_s: double
                    child 13, mean_decode_tps: double
              to
              {'model_label': Value('string'), 'url': Value('string'), 'model': Value('string'), 'mode': Value('string'), 'max_tokens': Value('int64'), 'repeats': Value('int64'), 'concurrency_sweep': List({'concurrency': Value('int64'), 'ok': Value('int64'), 'errors': Value('int64'), 'error_rate': Value('float64'), 'total_wall_s': Value('float64'), 'avg_ttft_s': Value('float64'), 'sum_decode_tps_all_streams': Value('float64'), 'aggregate_tokens_per_s': Value('null'), 'total_prompt_tokens': Value('int64'), 'total_completion_tokens': Value('int64'), 'per_request': List({'ok': Value('bool'), 'wall_s': Value('float64'), 'error': Value('null'), 'prompt_tokens': Value('int64'), 'completion_tokens': Value('int64'), 'ttft_s': Value('float64'), 'decode_tps': Value('float64'), 'tokens_per_s': Value('null'), 'milestone_s': {'50': Value('float64')}}), 'median_decode_tps': Value('float64'), 'p50_time_to_50_tokens_s': Value('float64'), 'mean_decode_tps': 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.

Small-model throughput + quality benchmarks (RTX 5060 Ti 16 GB)

Local vLLM throughput/scaling measurements and lighteval accuracy benchmarks for sub-2B chat models on a 3× RTX 5060 Ti (16 GB) host. Raw JSON outputs plus comparison tables.

Setup

Setting Value
Host RTX 5060 Ti 16 GB (CUDA device 2)
GPU CUDA device 2
Engine vLLM 0.24
max_model_len 4096
gpu_memory_utilization 0.75
max_num_seqs 128
Gen length 128 tokens
Warmup / repeats 2 / 2
Concurrency levels 1, 2, 4, 6, 8, 10, 12, 16, 20
Context tiers 128, 512, 1024, 2048, 4096
Prompts RAG-style chat scale cases

Models tested

  • Qwen/Qwen3.5-0.8B
  • Qwen/Qwen3.5-2B
  • LiquidAI/LFM2-1.2B-RAG
  • LiquidAI/LFM2.5-1.2B-Instruct
  • openbmb/MiniCPM5-1B

Key findings (throughput)

  1. Peak aggregate throughput @ conc=20: MiniCPM5-1B (~3700 tok/s) > Qwen3.5-0.8B (~3473) > LFM 1.2B (~3320) >> Qwen3.5-2B (~1787). All five models scale linearly to conc=20 with zero errors on the concurrency sweep.
  2. Single-stream decode: Qwen3.5-0.8B leads (~238 tok/s @ conc=1). MiniCPM5-1B ~200. LFM ~174. Qwen3.5-2B lags (~106) — likely hybrid DeltaNet + attention overhead at this size.
  3. Context scaling @ conc=10: TTFT grows roughly linearly with prompt tier. At 4096 tokens, LFM decode drops to ~80–86 tok/s/stream; MiniCPM similar. Qwen 4096 tier returns errors when prompt + 128 gen exceeds max_model_len (fixed null formatting in bench_scale.py; tier marked n/a in tables).
  4. LFM 1.2B RAG vs 2.5 1.2B Instruct: Nearly identical throughput profiles — pick by quality/RAG features, not speed.

Quality (lighteval chat_core suite)

Accuracy benchmarks for the same five models, served via the vLLM OpenAI-compatible endpoint and scored with lighteval. Suite chat_core = MMLU-Pro (0-shot, extractive_match), GPQA-Diamond (pass@1), and IFEval (instruction-following, strict/loose at prompt & instruction level).

Model MMLU-Pro GPQA-Diamond IFEval prompt (s/l) IFEval inst (s/l)
LiquidAI/LFM2-1.2B-RAG 0.274 0.258 0.610 / 0.625 0.722 / 0.734
LiquidAI/LFM2.5-1.2B-Instruct 0.447 0.308 0.832 / 0.852 0.880 / 0.894
Qwen/Qwen3.5-0.8B 0.372 0.273 0.527 / 0.569 0.633 / 0.675
Qwen/Qwen3.5-2B 0.563 0.460 0.675 / 0.730 0.764 / 0.807
openbmb/MiniCPM5-1B 0.390 0.202 0.754 / 0.787 0.823 / 0.844

s = strict, l = loose. Higher is better on all columns.

Quality vs. speed takeaways:

  • Qwen3.5-2B is the accuracy leader across the board (MMLU-Pro 0.563, GPQA 0.460) — but it is also the throughput laggard (~1,787 tok/s peak vs ~3,300–3,700 for the rest). The quality/speed trade is explicit: it roughly doubles GPQA over the 1B-class models at ~half the aggregate throughput.
  • LFM2.5-1.2B-Instruct is the instruction-following standout (IFEval prompt-strict 0.832) and clearly beats its RAG-tuned sibling LFM2-1.2B-RAG (0.610) on chat quality, despite near-identical speed profiles — confirming the throughput note: pick these two by quality, not speed.
  • MiniCPM5-1B pairs the best peak throughput (~3,700 tok/s) with strong IFEval (0.754), making it the best speed-for-instruction-following pick at the 1B tier.

Raw per-model lighteval dumps are in quality/; consolidated scores in quality/quality_summary.json.

Files

File Description
*_scale_conc.json Concurrency sweep per model
*_scale_ctx.json Context-length sweep per model
quality/quality_summary.json Consolidated lighteval scores (all 5 models)
quality/<model>.json Raw per-model lighteval results dump
SUMMARY.md Run log / pass-fail matrix
FINDINGS.md Auto-generated comparison tables

Reproduce 

# throughput
bash scripts/run_throughput_gentoo.sh
python scripts/compare_gentoo_throughput.py --dir data/results/gentoo-throughput-YYYYMMDD
# quality
bash evals/run_gauntlet_gentoo.sh --suite chat_core --gpu 1 --port 8011

Comparison tables

Concurrency sweep (short prompt, 128 tok gen)

Model conc TTFT (s) decode med (tok/s) agg decode (tok/s) err
LiquidAI/LFM2-1.2B-RAG 1 0.01 174.6 175 0
LiquidAI/LFM2-1.2B-RAG 2 0.02 175.0 350 0
LiquidAI/LFM2-1.2B-RAG 4 0.02 173.0 693 0
LiquidAI/LFM2-1.2B-RAG 6 0.03 170.3 1022 0
LiquidAI/LFM2-1.2B-RAG 8 0.04 170.1 1358 0
LiquidAI/LFM2-1.2B-RAG 10 0.05 169.1 1685 0
LiquidAI/LFM2-1.2B-RAG 12 0.06 169.1 2028 0
LiquidAI/LFM2-1.2B-RAG 16 0.07 168.5 2694 0
LiquidAI/LFM2-1.2B-RAG 20 0.08 166.4 3320 0
LiquidAI/LFM2.5-1.2B-Instruct 1 0.01 174.1 174 0
LiquidAI/LFM2.5-1.2B-Instruct 2 0.02 174.6 349 0
LiquidAI/LFM2.5-1.2B-Instruct 4 0.02 172.3 693 0
LiquidAI/LFM2.5-1.2B-Instruct 6 0.03 170.2 1023 0
LiquidAI/LFM2.5-1.2B-Instruct 8 0.04 170.2 1358 0
LiquidAI/LFM2.5-1.2B-Instruct 10 0.05 169.3 1690 0
LiquidAI/LFM2.5-1.2B-Instruct 12 0.05 168.7 2023 0
LiquidAI/LFM2.5-1.2B-Instruct 16 0.07 168.1 2690 0
LiquidAI/LFM2.5-1.2B-Instruct 20 0.08 166.4 3322 0
Qwen/Qwen3.5-0.8B 1 0.03 237.8 238 0
Qwen/Qwen3.5-0.8B 2 0.05 227.8 456 0
Qwen/Qwen3.5-0.8B 4 0.06 217.8 871 0
Qwen/Qwen3.5-0.8B 6 0.07 208.8 1253 0
Qwen/Qwen3.5-0.8B 8 0.07 201.3 1610 0
Qwen/Qwen3.5-0.8B 10 0.07 199.5 1993 0
Qwen/Qwen3.5-0.8B 12 0.07 193.7 2317 0
Qwen/Qwen3.5-0.8B 16 0.08 182.9 2906 0
Qwen/Qwen3.5-0.8B 20 0.09 173.6 3473 0
Qwen/Qwen3.5-2B 1 0.03 106.0 106 0
Qwen/Qwen3.5-2B 2 0.05 102.7 205 0
Qwen/Qwen3.5-2B 4 0.06 99.4 398 0
Qwen/Qwen3.5-2B 6 0.07 97.3 585 0
Qwen/Qwen3.5-2B 8 0.07 96.2 770 0
Qwen/Qwen3.5-2B 10 0.08 94.6 946 0
Qwen/Qwen3.5-2B 12 0.09 93.6 1122 0
Qwen/Qwen3.5-2B 16 0.11 91.6 1463 0
Qwen/Qwen3.5-2B 20 0.14 89.7 1787 0
openbmb/MiniCPM5-1B 1 0.01 200.3 200 0
openbmb/MiniCPM5-1B 2 0.01 200.6 401 0
openbmb/MiniCPM5-1B 4 0.02 200.5 803 0
openbmb/MiniCPM5-1B 6 0.02 201.8 1212 0
openbmb/MiniCPM5-1B 8 0.02 200.9 1608 0
openbmb/MiniCPM5-1B 10 0.03 197.7 1978 0
openbmb/MiniCPM5-1B 12 0.03 197.0 2365 0
openbmb/MiniCPM5-1B 16 0.04 194.1 3108 0
openbmb/MiniCPM5-1B 20 0.04 185.0 3700 0

Context sweep @ conc=10 (128 tok gen)

Model ctx tokens TTFT (s) decode med (tok/s) agg decode (tok/s) err
LiquidAI/LFM2-1.2B-RAG 128 0.07 165.5 1643 0
LiquidAI/LFM2-1.2B-RAG 512 0.15 149.6 1525 0
LiquidAI/LFM2-1.2B-RAG 1024 0.25 130.8 1347 0
LiquidAI/LFM2-1.2B-RAG 2048 0.44 114.1 1123 0
LiquidAI/LFM2-1.2B-RAG 4096 0.85 86.1 921 0
LiquidAI/LFM2.5-1.2B-Instruct 128 0.07 164.7 1634 0
LiquidAI/LFM2.5-1.2B-Instruct 512 0.15 146.4 1508 0
LiquidAI/LFM2.5-1.2B-Instruct 1024 0.25 128.8 1335 0
LiquidAI/LFM2.5-1.2B-Instruct 2048 0.44 105.5 1095 0
LiquidAI/LFM2.5-1.2B-Instruct 4096 0.86 80.1 854 0
Qwen/Qwen3.5-0.8B 128 0.08 190.9 1907 0
Qwen/Qwen3.5-0.8B 512 0.15 182.2 1766 0
Qwen/Qwen3.5-0.8B 1024 0.23 160.4 1597 0
Qwen/Qwen3.5-0.8B 2048 0.42 132.6 1347 0
Qwen/Qwen3.5-0.8B 4096 n/a n/a n/a 10
Qwen/Qwen3.5-2B 128 0.12 93.0 926 0
Qwen/Qwen3.5-2B 512 0.28 88.9 845 0
Qwen/Qwen3.5-2B 1024 0.49 77.2 760 0
Qwen/Qwen3.5-2B 2048 0.93 61.3 639 0
Qwen/Qwen3.5-2B 4096 n/a n/a n/a 10
openbmb/MiniCPM5-1B 128 0.03 195.1 1952 0
openbmb/MiniCPM5-1B 512 0.03 187.5 1876 0
openbmb/MiniCPM5-1B 1024 0.03 177.4 1776 0
openbmb/MiniCPM5-1B 2048 0.04 158.0 1584 0
openbmb/MiniCPM5-1B 4096 0.06 131.8 1320 0

Peak aggregate decode (concurrency sweep)

Model best conc agg decode (tok/s)
LiquidAI/LFM2-1.2B-RAG 20 3320
LiquidAI/LFM2.5-1.2B-Instruct 20 3322
Qwen/Qwen3.5-0.8B 20 3473
Qwen/Qwen3.5-2B 20 1787
openbmb/MiniCPM5-1B 20 3700
Downloads last month
61
Total file size:
427 kB