Hugging Face's logo

Datasets:
sjakek
/
qwen36-27b-mtp-long-context-decay

Tags:
benchmark
llama-cpp
gguf
qwen3.6
mtp
apple-silicon
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
quant: string
ctx: int64
kv: string
repeat: int64
config_id: string
ok: bool
started_at: timestamp[s]
early_stop: bool
cumulative_tokens: int64
cumulative_wall_s: double
cumulative_tok_s: double
active_mactop: struct<n: int64, DRAM_BW_Combined_GBs_median: double, DRAM_BW_Combined_GBs_mean: double, DRAM_BW_Com (... 438 chars omitted)
  child 0, n: int64
  child 1, DRAM_BW_Combined_GBs_median: double
  child 2, DRAM_BW_Combined_GBs_mean: double
  child 3, DRAM_BW_Combined_GBs_max: double
  child 4, DRAM_Read_BW_GBs_median: double
  child 5, DRAM_Read_BW_GBs_mean: double
  child 6, DRAM_Read_BW_GBs_max: double
  child 7, DRAM_Write_BW_GBs_median: double
  child 8, DRAM_Write_BW_GBs_mean: double
  child 9, DRAM_Write_BW_GBs_max: double
  child 10, GPU_Usage_median: double
  child 11, GPU_Usage_mean: double
  child 12, GPU_Usage_max: double
  child 13, CPU_Usage_median: double
  child 14, CPU_Usage_mean: double
  child 15, CPU_Usage_max: double
  child 16, Total_Power_median: double
  child 17, Total_Power_mean: double
  child 18, Total_Power_max: double
ended_at: timestamp[s]
power_cohort: string
power_cohort_reason: string
median_total_power_w: double
repo_id: string
charts_added_at: timestamp[s]
source_run_dir: string
charts: list<item: string>
  child 0, item: string
entrypoint: string
created_at: timestamp[s]
primary_summary: string
to
{'created_at': Value('timestamp[s]'), 'source_run_dir': Value('string'), 'repo_id': Value('string'), 'entrypoint': Value('string'), 'primary_summary': Value('string'), 'charts_added_at': Value('timestamp[s]'), 'charts': List(Value('string'))}
because column names don't match
Traceback:    Traceback (most recent call last):
                File "/src/services/worker/src/worker/utils.py", line 99, in get_rows_or_raise
                  return get_rows(
                         ^^^^^^^^^
                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 77, in get_rows
                  rows_plus_one = list(itertools.islice(ds, rows_max_number + 1))
                                  ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.12/site-packages/datasets/iterable_dataset.py", line 2690, in __iter__
                  for key, example in ex_iterable:
                                      ^^^^^^^^^^^
                File "/usr/local/lib/python3.12/site-packages/datasets/iterable_dataset.py", line 2227, in __iter__
                  for key, pa_table in self._iter_arrow():
                                       ^^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.12/site-packages/datasets/iterable_dataset.py", line 2251, in _iter_arrow
                  for key, pa_table in self.ex_iterable._iter_arrow():
                                       ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.12/site-packages/datasets/iterable_dataset.py", line 494, in _iter_arrow
                  for key, pa_table in iterator:
                                       ^^^^^^^^
                File "/usr/local/lib/python3.12/site-packages/datasets/iterable_dataset.py", line 384, in _iter_arrow
                  for key, pa_table in self.generate_tables_fn(**gen_kwags):
                                       ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.12/site-packages/datasets/packaged_modules/json/json.py", line 299, in _generate_tables
                  self._cast_table(pa_table, json_field_paths=json_field_paths),
                  ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.12/site-packages/datasets/packaged_modules/json/json.py", line 128, in _cast_table
                  pa_table = table_cast(pa_table, self.info.features.arrow_schema)
                             ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.12/site-packages/datasets/table.py", line 2321, in table_cast
                  return cast_table_to_schema(table, schema)
                         ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.12/site-packages/datasets/table.py", line 2249, in cast_table_to_schema
                  raise CastError(
              datasets.table.CastError: Couldn't cast
              quant: string
              ctx: int64
              kv: string
              repeat: int64
              config_id: string
              ok: bool
              started_at: timestamp[s]
              early_stop: bool
              cumulative_tokens: int64
              cumulative_wall_s: double
              cumulative_tok_s: double
              active_mactop: struct<n: int64, DRAM_BW_Combined_GBs_median: double, DRAM_BW_Combined_GBs_mean: double, DRAM_BW_Com (... 438 chars omitted)
                child 0, n: int64
                child 1, DRAM_BW_Combined_GBs_median: double
                child 2, DRAM_BW_Combined_GBs_mean: double
                child 3, DRAM_BW_Combined_GBs_max: double
                child 4, DRAM_Read_BW_GBs_median: double
                child 5, DRAM_Read_BW_GBs_mean: double
                child 6, DRAM_Read_BW_GBs_max: double
                child 7, DRAM_Write_BW_GBs_median: double
                child 8, DRAM_Write_BW_GBs_mean: double
                child 9, DRAM_Write_BW_GBs_max: double
                child 10, GPU_Usage_median: double
                child 11, GPU_Usage_mean: double
                child 12, GPU_Usage_max: double
                child 13, CPU_Usage_median: double
                child 14, CPU_Usage_mean: double
                child 15, CPU_Usage_max: double
                child 16, Total_Power_median: double
                child 17, Total_Power_mean: double
                child 18, Total_Power_max: double
              ended_at: timestamp[s]
              power_cohort: string
              power_cohort_reason: string
              median_total_power_w: double
              repo_id: string
              charts_added_at: timestamp[s]
              source_run_dir: string
              charts: list<item: string>
                child 0, item: string
              entrypoint: string
              created_at: timestamp[s]
              primary_summary: string
              to
              {'created_at': Value('timestamp[s]'), 'source_run_dir': Value('string'), 'repo_id': Value('string'), 'entrypoint': Value('string'), 'primary_summary': Value('string'), 'charts_added_at': Value('timestamp[s]'), 'charts': List(Value('string'))}
              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.

Qwen3.6-27B MTP Long-Context Decay Benchmark

This artifact contains a local Apple Silicon benchmark of Unsloth Qwen3.6-27B GGUF quantizations running with llama.cpp MTP draft-2 speculative decoding. The benchmark measured generation speed over sequential 1K-token windows up to 16K generated tokens, across context caps and KV cache precision.

Hub repo: sjakek/qwen36-27b-mtp-long-context-decay

Generated locally: 2026-05-14T08:51:11
Run directory on source machine: /Users/jkooker/Documents/Codex/2026-05-12/download-install-this-model-and-set/runs/qwen36_long_context_decay_20260513-232142

Visual Summary

Clustered bar chart showing token/s decay by generated-token depth

Line chart showing continuous token/s decay across all 1K windows

Six-line chart showing continuous token/s decay split by model and KV cache

Clustered bar chart comparing context sensitivity at 1K, 8K, and 16K output depths

Line chart showing q8_0 KV speed ratio versus f16 KV over generated-token depth

Important Power-Cohort Note

The Mac battery drained overnight while plugged in. To avoid mixing power-limited results into the primary averages, the raw run has been split into:

  • nominal_power: primary cohort for averages and comparisons.
  • suspected_power_limited: late-run tail after the observed power breakpoint or below the median-power threshold.
  • smoke: smoke-gate attempts, excluded from full-run averages.

Classification rule: attempts are tagged suspected_power_limited when started_at >= 2026-05-14T07:30:41 or active median total power is below 88.0 W.

Cohort Sizes

Cohort Attempts Successful Attempts Full Windows Generated Tokens
nominal_power 29 29 464 463784
suspected_power_limited 5 4 64 64000
smoke 3 3 0 0

Nominal Power Results

quant ctx kv completed_repeats attempt_tok_s_median first_window_tok_s_median mid_window_tok_s_median final_window_tok_s_median acceptance_median
Q4_K_M 16384 f16 2 17.171 21.454 17.816 11.440 1.000
Q4_K_M 16384 q8_0 2 16.414 20.285 16.706 11.326 1.000
Q4_K_M 32768 f16 2 17.949 20.406 17.988 16.272 1.000
Q4_K_M 32768 q8_0 2 17.006 20.761 16.621 15.768 1.000
Q4_K_M 65536 f16 2 17.696 20.213 17.664 16.133 1.000
Q4_K_M 65536 q8_0 2 17.167 20.797 16.792 15.827 1.000
Q6_K 16384 f16 1 15.417 18.298 15.505 13.928 1.000
Q6_K 16384 q8_0 1 14.600 17.541 14.774 13.727 0.998
Q6_K 32768 f16 1 15.276 17.478 14.825 13.769 1.000
Q6_K 32768 q8_0 1 14.536 17.617 14.696 13.713 0.998
Q6_K 65536 f16 1 15.281 17.666 15.492 13.660 1.000
Q6_K 65536 q8_0 1 14.484 17.368 14.699 13.679 0.998
UD-Q4_K_XL 16384 f16 2 15.582 20.329 17.051 7.117 0.998
UD-Q4_K_XL 16384 q8_0 2 16.059 19.506 17.037 11.482 1.000
UD-Q4_K_XL 32768 f16 2 16.940 20.225 17.216 16.258 0.998
UD-Q4_K_XL 32768 q8_0 2 16.547 19.727 16.601 15.739 1.000
UD-Q4_K_XL 65536 f16 2 17.164 20.416 17.520 16.446 0.998
UD-Q4_K_XL 65536 q8_0 1 16.551 20.013 16.582 15.721 1.000

Nominal q8_0 KV vs f16 KV

quant ctx f16_attempt_tok_s q8_attempt_tok_s q8_vs_f16_ratio
Q4_K_M 16384 17.171 16.414 0.956
Q4_K_M 32768 17.949 17.006 0.947
Q4_K_M 65536 17.696 17.167 0.970
Q6_K 16384 15.417 14.600 0.947
Q6_K 32768 15.276 14.536 0.952
Q6_K 65536 15.281 14.484 0.948
UD-Q4_K_XL 16384 15.582 16.059 1.031
UD-Q4_K_XL 32768 16.940 16.547 0.977
UD-Q4_K_XL 65536 17.164 16.551 0.964

Suspected Power-Limited Attempts

config_id started_at ok cumulative_tok_s median_total_power_w power_cohort_reason
UD-Q4_K_XL_ctx65536_q8_0_r2 2026-05-14T07:30:41 True 15.568 83.590 started_at >= 2026-05-14T07:30:41
Q6_K_ctx16384_f16_r2 2026-05-14T07:48:18 True 14.274 85.435 started_at >= 2026-05-14T07:30:41
Q6_K_ctx16384_q8_0_r2 2026-05-14T08:07:34 True 13.830 85.640 started_at >= 2026-05-14T07:30:41
Q6_K_ctx32768_f16_r2 2026-05-14T08:27:17 True 14.060 84.450 started_at >= 2026-05-14T07:30:41
Q6_K_ctx32768_q8_0_r2 2026-05-14T08:46:41 False started_at >= 2026-05-14T07:30:41

Hardware And Runtime

  • Apple M4 Max
  • 16 CPU cores, 40 GPU cores
  • 64 GB unified memory
  • Nominal memory bandwidth: 546 GB/s
  • Runtime: llama.cpp MTP branch local binary
  • Serving mode: OpenAI-compatible llama-server
  • MTP flags: --spec-type mtp --spec-draft-n-max 2
  • Flash attention: -fa on
  • Context caps: 16,384, 32,768, 65,536
  • KV cache types: f16, q8_0

See manifests/hardware_manifest.json, manifests/model_manifest.json, and manifests/benchmark_config.json for the machine-readable run metadata.

Files

  • index.html: polished report for quick visual inspection.
  • reports/report_power_cohorts.html: same report under a report path.
  • data/summary_power_cohorts.json: cohort summary.
  • data/attempts_power_cohorts.jsonl: per-config attempts with cohort labels.
  • data/windows_power_cohorts.jsonl: per-1K-window rows with cohort labels.
  • data/windows_nominal_power.jsonl: primary window dataset excluding the suspected power-limited tail.
  • data/windows_suspected_power_limited.jsonl: late-run/power-constrained window dataset.
  • raw/attempts.jsonl, raw/windows.jsonl, raw/events.jsonl: original run logs.
  • raw/mactop_samples.csv.gz: compressed system-level DRAM/GPU/CPU/power samples.
  • raw/server_logs.tar.gz: compressed llama-server logs.

Caveats

  • mactop bandwidth is system-level telemetry, not per-process attribution.
  • The suspected power-limited cohort should be used for sensitivity checks, not primary averages.
  • This is a local-hardware inference benchmark, not a model quality benchmark.
Downloads last month
102
Total file size:
14.3 MB