bottleneck
p2-etf-thermo-bottleneck-resultsbottleneck-oracle-graphsnla-bottleneck-resultsqwen36-coding-layer-bottleneck-profiles
Qwen3.6 Coding Layer Bottleneck Profiles
This dataset packages a local llama.cpp/ATX profiling campaign for identifying which whole transformer layers are the strongest candidates to keep hot for coding and agentic workloads.
The goal is to compare a learned top-layer policy against architecture heuristics such as the actual full-attention layers, first-10, and last-10. The included results are timing-attribution measurements, not CUDA speedup claims. They are intended to seed… See the full description on the dataset page: https://huggingface.co/datasets/jakeatx/qwen36-coding-layer-bottleneck-profiles.bottleneck-oracle-graphsDataset Card — bottleneck-oracle-graphs
Overview
A synthetic graph dataset of PyTorch profiler execution traces converted into heterogeneous graphs, designed for GNN-based bottleneck prediction in transformer model inference. Each graph represents one forward pass of a transformer variant, with nodes labelled by whether they lie on the critical execution path.
Dataset Construction
Traces were generated by profiling three transformer configurations (tiny, small, medium) using torch.profiler… See the full description on the dataset page: https://huggingface.co/datasets/archi829/bottleneck-oracle-graphs.open-bottleneck-ranklong27b-slurm-364982-rollouts
Open Bottleneck RankLong 27B — Slurm array 364982
Compact rollout evidence archived from completed Slurm array 364982.
Config
Files / steps
Records
JSONL bytes
Note
rank_a40
60 (1–60)
15,360
66,445,670
Complete local rollout evidence
rank_a80
54 (1–54)
13,824
59,386,451
Includes the cancelled arm's final dumped step (54.jsonl)
Each JSONL record contains input, output, gts, score, acc,
response_length, grouprel_reward, and step.
Only rollout evidence is archived… See the full description on the dataset page: https://huggingface.co/datasets/ryankim17920/open-bottleneck-ranklong27b-slurm-364982-rollouts.