Dataset Viewer
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: string
swapped_linears: int64
int4_bytes_gb: double
ppl_bf16: double
ppl_w4: double
decode_tok_s: double
sample: string
bf16: struct<tok_s: double, ppl: double>
child 0, tok_s: double
child 1, ppl: double
bnb_nf4: struct<tok_s: double, ppl: double>
child 0, tok_s: double
child 1, ppl: double
to
{'model': Value('string'), 'bf16': {'tok_s': Value('float64'), 'ppl': Value('float64')}, 'bnb_nf4': {'tok_s': Value('float64'), 'ppl': Value('float64')}}
because column names don't match
Traceback: Traceback (most recent call last):
File "/src/services/worker/src/worker/utils.py", line 147, 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 127, 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 483, 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 2951, in __iter__
for key, example in ex_iterable:
^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2461, in __iter__
for key, pa_table in self._iter_arrow():
~~~~~~~~~~~~~~~~^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2486, 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 547, in _iter_arrow
for key, pa_table in iterator:
^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 430, 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 2378, in table_cast
return cast_table_to_schema(table, schema)
File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2306, in cast_table_to_schema
raise CastError(
...<3 lines>...
)
datasets.table.CastError: Couldn't cast
model: string
swapped_linears: int64
int4_bytes_gb: double
ppl_bf16: double
ppl_w4: double
decode_tok_s: double
sample: string
bf16: struct<tok_s: double, ppl: double>
child 0, tok_s: double
child 1, ppl: double
bnb_nf4: struct<tok_s: double, ppl: double>
child 0, tok_s: double
child 1, ppl: double
to
{'model': Value('string'), 'bf16': {'tok_s': Value('float64'), 'ppl': Value('float64')}, 'bnb_nf4': {'tok_s': Value('float64'), 'ppl': Value('float64')}}
because column names don't matchNeed help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
Roofline-First Decoding: results
Result files for roofline-decoding: a fused 4-bit dequantise-plus-GEMV Triton kernel for batch-1 decoding, measured against the memory-bandwidth roofline of one RTX 3060 12GB.
Files
roofline.json: measured copy bandwidth (323.9 GB/s) and the ceilings table,ceiling = bandwidth / (weight bytes + KV-cache read at 2,048 tokens)baselines.json: bf16 and bitsandbytes NF4 decode speed and probe perplexity (Qwen2.5-1.5B-Instruct)decode.json: end-to-end decode with every linear swapped for the W4A16 kernel (197 layers)kernels.json: per-shape kernel latencies (reference, v1, v2, v3) and % of rooflinepct_roofline.png: the percent-of-roofline chart
Key numbers (Qwen2.5-1.5B-Instruct, batch-1 greedy)
| Setup | tok/s | Probe perplexity | % of its roofline |
|---|---|---|---|
| bf16 (HF) | 31.3 | 12.91 | 29.8% |
| bitsandbytes NF4 | 25.7 | 18.98 | 6.1% |
| own W4 kernel (v3) | 25.5 | 15.76 | 6.1% |
Caveats
- Probe perplexity uses four sentences (noisy), and bitsandbytes quantises different layers than the swap, so the perplexity gap is directional.
- The 6.1% figure is relative to the all-int4 ceiling (411.8 tok/s). The swapped model keeps the tied embedding / lm_head in bf16 and reads 1.263 GB of weights per token, so against bytes actually moved it is about 10% of measured bandwidth (model ceiling about 245 tok/s).
- In the per-shape microbenchmarks (
kernels.json) v3 is not faster than v2 on any of the five shapes measured; end-to-end decoding uses v3. - One GPU, one prompt for timing, batch-1 and greedy only. No Marlin or llama.cpp numbers.
- Downloads last month
- 16