The dataset viewer is not available for this subset.
Exception: SplitsNotFoundError
Message: The split names could not be parsed from the dataset config.
Traceback: Traceback (most recent call last):
File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 290, in _generate_tables
pa_table = paj.read_json(
io.BytesIO(batch), read_options=paj.ReadOptions(block_size=block_size)
)
File "pyarrow/_json.pyx", line 342, in pyarrow._json.read_json
File "pyarrow/error.pxi", line 155, in pyarrow.lib.pyarrow_internal_check_status
return check_status(status)
File "pyarrow/error.pxi", line 92, in pyarrow.lib.check_status
raise convert_status(status)
pyarrow.lib.ArrowInvalid: JSON parse error: Column() changed from object to string in row 0
During handling of the above exception, another exception occurred:
Traceback (most recent call last):
File "/usr/local/lib/python3.14/site-packages/datasets/inspect.py", line 286, in get_dataset_config_info
for split_generator in builder._split_generators(
~~~~~~~~~~~~~~~~~~~~~~~~~^
StreamingDownloadManager(base_path=builder.base_path, download_config=download_config)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
)
^
File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 101, in _split_generators
pa_table = next(iter(self._generate_tables(**splits[0].gen_kwargs, allow_full_read=False)))[1]
~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 304, in _generate_tables
batch = json_encode_fields_in_json_lines(original_batch, json_field_paths)
File "/usr/local/lib/python3.14/site-packages/datasets/utils/json.py", line 111, in json_encode_fields_in_json_lines
examples = [ujson_loads(line) for line in original_batch.splitlines()]
~~~~~~~~~~~^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/utils/json.py", line 20, in ujson_loads
return pd.io.json.ujson_loads(*args, **kwargs)
~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^
ValueError: Expected object or value
The above exception was the direct cause of the following exception:
Traceback (most recent call last):
File "/src/services/worker/src/worker/job_runners/config/split_names.py", line 66, in compute_split_names_from_streaming_response
for split in get_dataset_split_names(
~~~~~~~~~~~~~~~~~~~~~~~^
path=dataset,
^^^^^^^^^^^^^
config_name=config,
^^^^^^^^^^^^^^^^^^^
token=hf_token,
^^^^^^^^^^^^^^^
)
^
File "/usr/local/lib/python3.14/site-packages/datasets/inspect.py", line 340, in get_dataset_split_names
info = get_dataset_config_info(
path,
...<6 lines>...
**config_kwargs,
)
File "/usr/local/lib/python3.14/site-packages/datasets/inspect.py", line 291, in get_dataset_config_info
raise SplitsNotFoundError("The split names could not be parsed from the dataset config.") from err
datasets.inspect.SplitsNotFoundError: The split names could not be parsed from the dataset config.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.
Qwopus 27B MTP TurboQuant KV Cache Evaluation
Local RTX 3090 Ti evaluation of q8/q8 KV cache against TurboQuant asymmetric V-cache modes from TheTom/llama-cpp-turboquant.
Human-Readable Takeaway
TurboQuant V-cache behaves like a memory-capacity trade: it substantially improves context headroom, but on this Qwopus 27B MTP setup it is not free. The longer 8K/16K follow-up confirms the speed cost is real and stable when decode length is long enough to measure cleanly.
- At the 16K prompt-target follow-up, baseline decoded at
76.63tok/s.q8/tq4decoded at57.45tok/s (-25.0%), andq8/tq3decoded at57.91tok/s (-24.4%). - At 16K PPL on the deterministic smoke fixture,
q8/tq4was -2.42% versus baseline andq8/tq3was -2.90% versus baseline. Because lower PPL is better and both TurboQuant rows are lower on this fixture, this should be read as no detected degradation on this smoke test, not as proof of general quality improvement. - At 8K PPL, all three modes are effectively identical on this fixture.
- At 32K PPL on the longer deterministic fixture, all three modes are again effectively identical:
q8/q8= 1.0021,q8/tq4= 1.0017, andq8/tq3= 1.0010. q8/tq3andq8/tq4are very close to each other in the long follow-up. In this run, tq3 is slightly faster on decode at 16K and has slightly lower PPL than tq4, but the differences are small enough that a real text-quality eval would be needed before calling tq3 safer than tq4.- Practical conclusion: TurboQuant still looks near-lossless on this smoke PPL fixture, with a roughly 18-25% speed penalty in the measured long-output cases. It is best justified when the extra context length matters.
Setup
| Item | Value |
|---|---|
| Model | Qwopus3.6-27B-Coder-MTP-Q3_K_L.gguf |
| GPU | NVIDIA GeForce RTX 3090 Ti |
| Baseline | current llama.cpp, target/draft KV q8_0/q8_0 |
| Test branch | TheTom/llama-cpp-turboquant, feature/turboquant-kv-cache, commit 35ac80d55b8e |
| Changed flags | V cache and draft V cache changed to turbo4 or turbo3; other runtime flags aligned with baseline |
| PPL fixtures | 1K/4K used prompt_10k.txt; 8K/16K used repeated deterministic prompt_50k.txt because llama-perplexity requires at least 2x context tokens |
Primary Prefill Scaling Chart
This chart combines the controlled short/medium prompt benchmarks with the long-context probes. The main pattern is that prefill throughput drops with actual prompt length: roughly 1,200+ tok/s in the small 2K-10K region, about 515 tok/s around an 80K prompt, 395 tok/s at a 120K prompt, and 338 tok/s at a 150K prompt. The 98K pair shows q8/q8 and q8/tq3 essentially tied on prefill at that operating point, while q8/tq3 provides lower VRAM use.
Primary Decode Scaling Chart
This chart combines the controlled short/medium decode benchmarks with the long-context probes. The x-axis is final live context tokens, meaning prompt plus generated tokens. The long probes were EOS-limited rather than fixed-length all the way to the requested target, but they still show the operating range: q8/q8 decoded at about 33.9 tok/s around 87K live context, q8/tq3 decoded at about 32.7 tok/s around 88K, 31.5 tok/s around 145K, and 27.4 tok/s around 167K.
Short Decode Throughput
Initial speed suite: prompt targets 1K/2K/4K/8K with 256 generated tokens.
| Config | Prompt target | Actual eval tokens | Output tokens | Prefill tok/s | Prefill delta | Decode tok/s | Decode delta | Draft accepted |
|---|---|---|---|---|---|---|---|---|
| q8/q8 baseline | 1,024 | 1,255 | 256 | 1153.88 | +0.00% | 70.24 | +0.00% | 179 / 225 |
| q8/q8 baseline | 2,048 | 2,510 | 256 | 1234.17 | +0.00% | 80.30 | +0.00% | 189 / 195 |
| q8/q8 baseline | 4,096 | 5,020 | 256 | 1266.34 | +0.00% | 80.70 | +0.00% | 191 / 192 |
| q8/q8 baseline | 8,192 | 10,040 | 256 | 1266.60 | +0.00% | 79.98 | +0.00% | 191 / 191 |
| q8/tq4 | 1,024 | 1,255 | 256 | 1023.68 | -11.28% | 67.20 | -4.33% | 191 / 191 |
| q8/tq4 | 2,048 | 2,510 | 256 | 1124.31 | -8.90% | 66.19 | -17.57% | 191 / 191 |
| q8/tq4 | 4,096 | 5,020 | 256 | 1119.64 | -11.58% | 65.14 | -19.28% | 191 / 191 |
| q8/tq4 | 8,192 | 10,040 | 256 | 1054.41 | -16.75% | 62.19 | -22.24% | 191 / 191 |
| q8/tq3 | 1,024 | 1,255 | 256 | 1040.60 | -9.82% | 53.01 | -24.53% | 172 / 244 |
| q8/tq3 | 2,048 | 2,510 | 256 | 1121.89 | -9.10% | 66.95 | -16.64% | 191 / 191 |
| q8/tq3 | 4,096 | 5,020 | 256 | 1114.26 | -12.01% | 65.57 | -18.76% | 191 / 191 |
| q8/tq3 | 8,192 | 10,040 | 256 | 1050.37 | -17.07% | 62.99 | -21.24% | 191 / 191 |
Long Decode Follow-Up
Follow-up speed suite: prompt targets 8K/16K with 1024 generated tokens. The actual evaluated prompt token counts are shown because the deterministic text tokenizes denser than the target word count.
| Config | Prompt target | Actual eval tokens | Output tokens | Prefill tok/s | Prefill delta | Decode tok/s | Decode delta | Draft accepted |
|---|---|---|---|---|---|---|---|---|
| q8/q8 baseline | 8,192 | 10,298 | 1,024 | 1274.20 | +0.00% | 80.00 | +0.00% | 767 / 768 |
| q8/q8 baseline | 16,384 | 20,596 | 1,024 | 1221.00 | +0.00% | 76.63 | +0.00% | 767 / 767 |
| q8/tq4 | 8,192 | 10,298 | 1,024 | 1035.10 | -18.76% | 62.19 | -22.26% | 767 / 767 |
| q8/tq4 | 16,384 | 20,596 | 1,024 | 909.28 | -25.53% | 57.45 | -25.03% | 767 / 767 |
| q8/tq3 | 8,192 | 10,298 | 1,024 | 1035.85 | -18.71% | 62.49 | -21.89% | 767 / 767 |
| q8/tq3 | 16,384 | 20,596 | 1,024 | 909.03 | -25.55% | 57.91 | -24.43% | 767 / 767 |
50K Prompt Follow-Up
This controlled follow-up uses one calibrated 49,978-token prompt and 1,024 requested decode tokens at -c 98304 across all three treatment groups.
| Config | Prompt tokens | Decode tokens | Prefill tok/s | Prefill delta | Decode tok/s | Decode delta | Draft accepted |
|---|---|---|---|---|---|---|---|
| q8/q8 | 49,978 | 1,024 | 1071.44 | +0.00% | 68.49 | +0.00% | 767 / 767 |
| q8/tq4 | 49,978 | 1,024 | 659.12 | -38.48% | 46.96 | -31.43% | 766 / 770 |
| q8/tq3 | 49,978 | 1,024 | 658.66 | -38.53% | 48.08 | -29.80% | 767 / 768 |
Interpretation: at the 50K prompt point, q8/q8 is still much faster. q8/tq4 and q8/tq3 are nearly identical on prefill, while q8/tq3 is slightly faster than q8/tq4 on decode in this run.
65K Prompt / 10K Decode Follow-Up
This controlled follow-up uses one calibrated 64,990-token prompt and 10,000 requested decode tokens at -c 98304 across all three treatment groups.
| Config | Prompt tokens | Decode tokens | Prefill tok/s | Prefill delta | Decode tok/s | Decode delta | Draft accepted |
|---|---|---|---|---|---|---|---|
| q8/q8 | 64,990 | 1,793 | 1001.39 | +0.00% | 47.84 | +0.00% | 1,186 / 1,818 |
| q8/tq4 | 64,990 | 3,396 | 576.88 | -42.39% | 31.11 | -34.96% | 2,222 / 3,519 |
| q8/tq3 | 64,990 | 4,468 | 575.24 | -42.56% | 32.94 | -31.14% | 2,990 / 4,431 |
Interpretation: at the 65K prompt point, q8/q8 is still much faster. q8/tq4 and q8/tq3 are nearly identical on prefill, while q8/tq3 is slightly faster than q8/tq4 on decode in this run.
Perplexity Smoke Checks
| Config | Context | PPL | Delta vs q8/q8 | Error | Suite |
|---|---|---|---|---|---|
| q8/q8 baseline | 1,024 | 1.023700 | +0.000% | +/- 0.00939 | initial |
| q8/tq4 | 1,024 | 1.023200 | -0.049% | +/- 0.00723 | initial |
| q8/tq3 | 1,024 | 1.036300 | +1.231% | +/- 0.01580 | initial |
| q8/q8 baseline | 4,096 | 1.001100 | +0.000% | +/- 0.00021 | initial |
| q8/tq4 | 4,096 | 1.001100 | +0.000% | +/- 0.00015 | initial |
| q8/tq3 | 4,096 | 1.001200 | +0.010% | +/- 0.00020 | initial |
| q8/q8 baseline | 8,192 | 1.000300 | +0.000% | +/- 0.00001 | followup_8k16k_long_fixture |
| q8/tq4 | 8,192 | 1.000300 | +0.000% | +/- 0.00001 | followup_8k16k_long_fixture |
| q8/tq3 | 8,192 | 1.000400 | +0.010% | +/- 0.00001 | followup_8k16k_long_fixture |
| q8/q8 baseline | 16,384 | 1.064600 | +0.000% | +/- 0.00899 | followup_8k16k_long_fixture |
| q8/tq4 | 16,384 | 1.038800 | -2.423% | +/- 0.00641 | followup_8k16k_long_fixture |
| q8/tq3 | 16,384 | 1.033700 | -2.902% | +/- 0.00663 | followup_8k16k_long_fixture |
| q8/q8 baseline | 32,768 | 1.002100 | +0.000% | +/- 0.00090 | followup_32k_long_fixture |
| q8/tq4 | 32,768 | 1.001700 | -0.040% | +/- 0.00096 | followup_32k_long_fixture |
| q8/tq3 | 32,768 | 1.001000 | -0.110% | +/- 0.00065 | followup_32k_long_fixture |
The PPL fixtures are smoke tests, not benchmark-corpus quality claims. They are useful for catching obvious degradation under identical local conditions.
98K Context 80K-In KV Pair Probe
This paired run used -c 98304, --parallel 1, the same calibrated coding prompt, and 15,000 requested output tokens. Both runs stopped early on EOS, so decode lengths differ slightly; the comparison is still useful as a near-98K live-context speed and memory probe.
| Metric | q8/q8 | q8/tq3 | Delta |
|---|---|---|---|
| Prompt tokens | 79,952 | 79,952 | - |
| Generated tokens before EOS | 7,286 | 8,088 | - |
| Prefill time | 155.1 s | 156.0 s | +0.52% |
| Prefill throughput | 515.35 tok/s | 512.66 tok/s | -0.52% |
| Decode time | 214.7 s | 247.2 s | - |
| Decode throughput | 33.94 tok/s | 32.72 tok/s | -3.60% |
| Draft accepted/generated | 5,090 / 6,585 | 5,593 / 7,482 | - |
| Peak total GPU used | 19.42 GiB | 18.35 GiB | -5.52% |
| Peak llama-server VRAM | 18.79 GiB | 17.71 GiB | -5.70% |
| Minimum free VRAM | 4.13 GiB | 5.20 GiB | - |
Interpretation: at this 98K-context / 80K-prompt operating point, q8/tq3 did not show the large speed penalty seen at some shorter synthetic runs. Prefill was essentially tied (512.66 vs 515.35 tok/s), while q8/tq3 decode was about 3.6% slower in this EOS-limited decode. The main benefit was memory: q8/tq3 reduced peak llama-server VRAM by about 5.7%.
150K q8/tq3 Long-Context Probe
This run used the updated launcher at -c 153600 with target KV q8_0/turbo3 and draft KV q8_0/turbo3. The request loaded a calibrated 120,033-token prompt and generated 25,000 additional tokens, ending around 145K live context tokens.
| Metric | Value |
|---|---|
| Prompt tokens processed | 120,033 |
| Generated tokens | 25,000 |
| Prefill time | 303.6 s |
| Prefill throughput | 395.31 tok/s |
| Decode time | 794.3 s |
| Decode throughput | 31.48 tok/s |
| End-to-end request time | 1098.1 s |
| Draft accepted/generated | 18,503 / 19,488 |
| Peak total GPU used | 20.31 GiB |
| Peak llama-server VRAM | 19.54 GiB |
| Minimum free VRAM | 3.24 GiB |
Interpretation: the 150K q8/tq3 configuration fit and completed the 120K+25K test without OOM. VRAM was mostly allocated up front and stayed nearly flat; the main runtime change was compute throughput. Prefill slowed substantially as context length grew, while decode stabilized around the low 30 tok/s range after the first few thousand generated tokens.
150K q8/tq3 330W Power-Cap Probe
This rerun used the same 150K q8/tq3 runtime shape, but capped the RTX 3090 Ti at 330W. The request loaded the same calibrated 120,033-token prompt and generated 20,000 tokens with ignore_eos.
| Metric | Value |
|---|---|
| GPU power limit | 330 W |
| Prompt tokens processed | 120,033 |
| Generated tokens | 20,000 |
| Prefill time | 340.8 s |
| Prefill throughput | 352.21 tok/s |
| Decode time | 712.6 s |
| Decode throughput | 28.06 tok/s |
| End-to-end request time | 1053.6 s |
| Draft accepted/generated | 14,754 / 15,734 |
| Average sampled GPU power | 327.0 W |
| Peak sampled GPU power | 329.23 W |
| Peak total GPU used | 20.30 GiB |
| Peak llama-server VRAM | 19.54 GiB |
| Minimum free VRAM | 3.25 GiB |
Interpretation: the 330W cap held throughout the run, with sampled power averaging about 327W. Compared with the earlier uncapped 120K+25K run, capped prefill dropped from 395.31 tok/s to 352.21 tok/s, while capped decode averaged 28.06 tok/s versus 31.48 tok/s in the uncapped run. The capped run remained stable and did not materially change VRAM use.
200K q8/tq3 Single-Slot Long-Context Probe
This run used -c 204800, --parallel 1, target KV q8_0/turbo3, and draft KV q8_0/turbo3. The smoke load passed, then the request processed a calibrated 149,998-token coding prompt and asked for 30,000 generated tokens. The model emitted EOS after 16,650 generated tokens, so the observed final live context was about 166,648 tokens rather than the full requested 180K.
| Metric | Value |
|---|---|
| Prompt tokens processed | 149,998 |
| Requested generated tokens | 30,000 |
| Actual generated tokens | 16,650 |
| Final live context tokens | 166,648 |
| Prefill time | 444.1 s |
| Prefill throughput | 337.77 tok/s |
| Decode time | 606.9 s |
| Decode throughput | 27.43 tok/s |
| End-to-end request time | 1051.3 s |
| Draft accepted/generated | 12,163 / 13,458 |
| Peak total GPU used | 21.90 GiB |
| Peak llama-server VRAM | 21.24 GiB |
| Minimum free VRAM | 1.65 GiB |
Interpretation: 200K single-slot q8/tq3 is viable on this 3090 Ti under the tested flags. Unlike the 225K/3 parallel case, it has enough runtime headroom to process a very large prompt. Memory stayed mostly flat, peaking at 21.90 GiB total GPU use with about 1.65 GiB free at the tightest point. Prefill slowed as context grew, finishing at 337.77 tok/s cumulative, and decode averaged 27.43 tok/s before EOS.
VRAM and Context Capacity
| KV mode | Daily target | Stretch target | Validation-only |
|---|---|---|---|
| q8/q8 | 128K | 144K | 160K |
| q8/tq4 | 160K | 180K | 192K |
| q8/tq3 | 160K-180K | 192K | 200K+ |
The context-capacity estimate is intentionally conservative. 192K is plausible with TurboQuant V-cache, but the logs show compute-buffer growth, prompt-cache/checkpoint behavior, desktop GPU use, and allocator headroom make it validation-required rather than guaranteed.
Files
index.html: standalone visual reportspeed.csv: throughput timing rowsppl.csv: perplexity rowsresults.json: machine-readable combined outputcontext_capacity.md: context-capacity calculationsvram_runtime_report.md: VRAM investigationappendix/hermes_200k_startup_prefill_example.png: Hermes 200K startup/prefill screenshotlogs/: raw server and perplexity logs
- Downloads last month
- 498