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/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/harbor/harbor.py", line 171, in _split_generators
raise DataFilesNotFoundError("No task.toml or instruction.md files found")
datasets.exceptions.DataFilesNotFoundError: No task.toml or instruction.md files found
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 68, 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.
FlashTopK full-prefill benchmark inputs
The trained (glm51) and untrained (init) input sets of FlashTopK's reproduction scripts (reproduce/ of the FlashTopK repository): GLM-5.1's DSA indexer inputs at layers 24 and 48 over the first 1,048,576 tokens of the InfiniteBench document longbook_qa_chn #123, from SGLang serving GLM-5.1.
| set | model |
|---|---|
glm51 |
GLM-5.1 in BF16, zai-org/GLM-5.1 at 26e1bd6e011feb778d25ae34b09b07074139d92d |
init |
the same architecture with its weights drawn as Megatron-LM initializes a model (as the --init-method-std 6e-3 run of the released measurements): N(0, 0.006^2) matrices, the attention and MLP output projections N(0, (0.006 / sqrt(2 * 78))^2), unit norm weights, zero biases, each tensor and each tensor-parallel shard from its own seed |
Files
full_bf16/<set>_1m/call<layer>.pt, atorch.savedict:q[L, 32, 128]andkv[L, 128]in BF16, as the indexer scores them (after its Hadamard rotation);weights[L, 32]FP32, the head weights its logits take (weights_proj(x) * 32^-0.5 * 128^-0.5, without FP8 query scales);starts,ends[L]int32, the causal windows[0, r + 1);meta. L = 1,048,576.SHA256SUMS.dump/: the scripts that wrote the files, asreproduce/of the FlashTopK repository has them (dump_sglang.py,dump_hook/,get_inputs.pyand the pins they read). They run with SGLang 0.5.18 and huggingface-hub 1.33.0 on Python 3.11 to 3.13.
Every shorter length of the benchmark (128K to 768K) is the first L rows of the 1M file: one causal prefill wrote them all. get_inputs.py hf in reproduce/ downloads this dataset at the revision pins.toml names, cuts the shorter lengths and writes the FP8 files (power-of-two scales per key and per query head) locally.
The document's text is not redistributed here. get_inputs.py text downloads it from InfiniteBench at the revision pins.toml names and checks its sha256.
How they were made
SGLang 0.5.18 on four nodes of 8 H800 (tensor parallel 8, pipeline parallel 4), the model cut after layer 48 (the layers before it compute the same), one request of the document's first 1,048,576 tokens tokenized by GLM-5.1's tokenizer, chunked prefill of 16,384 tokens, FP8 KV cache, FlashMLA attention (flashmla_kv), Triton MoE and GEMM, the indexer's unfused path; a hook saved each chunk's indexer inputs on tensor-parallel rank 0, joined in position order. Positions past GLM-5.1's 202,752-token context take its extrapolated RoPE.
License
The tensors are derived from GLM-5.1 (MIT) and from the InfiniteBench text, which is published without a license; the text belongs to its authors and is not included here.
- Downloads last month
- 120