Datasets:

hlky
/

model-operator-workloads

pretty_name: Model Operator Workloads
license: cc-by-4.0
language:
  - en
size_categories:
  - 10K<n<100K
tags:
  - benchmarking
  - gpu
  - operator-workloads
  - gemm
  - model-architecture
configs:
  - config_name: gemm_runtime
    default: true
    data_files:
      - split: train
        path: data/gemm_runtime/train-00000-of-00001.parquet
  - config_name: gemm_weights
    data_files:
      - split: train
        path: data/gemm_weights/train-00000-of-00001.parquet
  - config_name: gemm_benchmark
    data_files:
      - split: train
        path: data/gemm_benchmark/train-00000-of-00001.parquet
  - config_name: convolution
    data_files:
      - split: train
        path: data/convolution/train-00000-of-00001.parquet
  - config_name: convolution_lowerings
    data_files:
      - split: train
        path: data/convolution_lowerings/train-00000-of-00001.parquet
  - config_name: attention
    data_files:
      - split: train
        path: data/attention/train-00000-of-00001.parquet
  - config_name: embedding
    data_files:
      - split: train
        path: data/embedding/train-00000-of-00001.parquet
  - config_name: moe
    data_files:
      - split: train
        path: data/moe/train-00000-of-00001.parquet
  - config_name: normalization
    data_files:
      - split: train
        path: data/normalization/train-00000-of-00001.parquet
  - config_name: pooling
    data_files:
      - split: train
        path: data/pooling/train-00000-of-00001.parquet
  - config_name: reduction
    data_files:
      - split: train
        path: data/reduction/train-00000-of-00001.parquet
  - config_name: transform
    data_files:
      - split: train
        path: data/transform/train-00000-of-00001.parquet

Model Operator Workloads

Unique neural-network operator shapes and selected reference workloads for kernel benchmarking, workload selection, and architecture coverage analysis. Created by hlky.

What is included

The nine primary operator configurations contain 43,150 rows. Three supplemental configurations expose static GEMM weight shapes, synthetic GEMM benchmark shapes, and symbolic convolution lowerings. All 12 configurations total 60,690 rows. This sum is across different views and is not a claim of global uniqueness across operators or views. Each configuration is unique on its documented workload identity, excluding ranks and provenance counts. Module instances, expanded checkpoint instances, and symbolic module inventories are excluded. The duplicate unranked GEMM weight list is also excluded.

Configuration	Unique rows	Evidence/workload category
`gemm_runtime`	9,418	reference_workloads
`gemm_weights`	1,961	architecture_shapes
`gemm_benchmark`	13,727	synthetic_benchmark
`convolution`	2,920	architecture_shapes
`convolution_lowerings`	1,852	symbolic_lowerings
`attention`	3,415	reference_workloads
`embedding`	8,182	reference_workloads
`moe`	4,626	reference_workloads
`normalization`	4,416	reference_workloads
`pooling`	342	reference_workloads
`reduction`	8,843	reference_workloads
`transform`	988	reference_workloads

Each configuration has a single train split for loading convenience. This is an inventory, not a training corpus; no independent train/test partition is implied.

Loading

from datasets import load_dataset

repo_id = "hlky/model-operator-workloads"
gemm = load_dataset(repo_id, "gemm_runtime", split="train")
attention = load_dataset(repo_id, "attention", split="train")

Workload meaning

gemm_runtime: unique M,N,K tuples from selected linear and per-head attention reference workloads. A tuple can occur in multiple execution_forms. Keep those forms and attention_batch_groups when interpreting it; batch-head groups are not folded into M. The aggregates do not preserve every pairing between a workload, execution form, and batch-head count.
gemm_weights: oriented architecture K,N pairs. Header evidence can assume ordinary PyTorch N,K weight storage; custom/packed layouts need interpretation.
gemm_benchmark: synthetic Cartesian expansion over selected M buckets (1, 16, 64, 256, 1024, 4096, 16384). These are benchmark choices, not traced model shapes.
convolution: unique Conv/ConvTranspose 1D–3D module geometries, with groups, stride, padding, dilation, bias, and symbolic lowering geometry explicit. Ordinary lowering uses M=B*output_positions, N=Cout/groups, K=(Cin/groups)kernel_volume. Transposed lowering uses M=Binput_positions, N=(Cout/groups)*kernel_volume, K=Cin/groups.
convolution_lowerings: unique per-group lowering signatures. Summed checkpoint and family counts may count a source repeatedly across different convolution shapes.
attention: B,heads,Q,K,head_dim with KV grouping, masks, dtype, prefill/decode, and cross-attention behavior retained.
embedding: lookup/gather and vocabulary projection, with table geometry, token counts, tied-weight state, dtype, and output scope.
moe: expert GEMM dimensions, selected routed-token profiles, expert grouping, routing assumptions, capacity policy and dtype. Routing distributions are selected estimates, not captured routing traffic.
normalization: LayerNorm/RMSNorm/GroupNorm/BatchNorm row and reduction geometry, epsilon, affine state, groups, and dtype.
pooling: full selected tensor geometry, pooling parameters and semantics.
reduction: outer rows and reduction width with mask, top-k, dtype, execution mode and fused context. Some vocabulary operations are consumer-selected.
transform: tensor transforms/resampling with selected input/output geometry, operation parameters, layout and dtype.

Schema and missing values

See SCHEMA.md and schema.json for every column, type, null count and identity key. Source column names and row order are retained except for the audited MoE consolidation below. Unambiguous integers become int64; true/false columns become boolean. Epsilon remains a string because numeric values coexist with not_exposed. Empty cells become null. Explicit markers such as not_applicable, not_configured, not_observed, runtime_selected, and literal null strings remain strings. Mixed numeric/marker columns remain strings. Geometry strings and JSON parameters retain their original encoding. Semicolon-delimited examples and aggregates remain source strings: examples can be truncated and must not be treated as exhaustive membership lists. Unknown family markers are preserved.

Limitations

Counts and rankings represent this source snapshot's coverage, not deployment frequency, market share, hardware performance, or complete production layer counts. Source families can overlap. Source counting conventions and treatment of missing checkpoint IDs can vary by operator. No new frequency estimates are introduced. Reference batch/sequence/spatial sizes, routing distributions and FLOPs are selected or derived rather than observed execution. Unsupported or unresolved workloads in the wider inventories are absent from these unique resolved lists. Configured compute dtype and observed weight dtype are distinct; unpinned dtype stays explicit. Different dtype, mask, grouping, layout or operation semantics remain distinct even when numerical dimensions coincide. Do not concatenate configurations and deduplicate solely by dimensions. Validate inferred shapes and assumptions for the target model and backend before using them as kernel correctness fixtures.

License and attribution

Copyright 2026 hlky. The dataset compilation and original contributions are licensed under Creative Commons Attribution 4.0 International. Credit hlky, link to the dataset and license, and indicate changes when applicable. Upstream model identifiers and architecture metadata retain their respective origins; this license does not relicense upstream model weights, code, or documents.

Suggested attribution: “Model Operator Workloads by hlky (https://github.com/hlky), CC BY 4.0.” Include the dataset repository URL when citing the uploaded dataset. See ATTRIBUTION.md.