rotate_half
rotate_half: 24 implementations from 96 papers ran on one shared input (census 2026-09-22, battery b986f7e04d79); they produced 4 distinct outputs across 2 buckets, one shared input per bucket.
Identical values to six decimals (the recorded digest) on the shared input are agreement on those inputs, not equivalence. Implementations are compared only within one bucket, the positional (rank, kind, dtype) of each array argument; the argument name is not part of the key because the harness draws the shared array from (rank, kind) and casts it to the dtype, whatever the name; each bucket's shared input is fixed by that key, so members of one bucket saw bitwise-identical inputs under their own scalar arguments. A cluster is the set of members whose recorded output digest is identical. Nothing here says which computation a paper's method intended, and nothing reproduces a paper's results.
Not compared, and not in the tables or the counts above:
- Every examined implementation of this name that took an array argument ran on the shared input.
- No implementation of this name ran on its own fixture arguments only: every one that ran took an array argument.
- Every compared output was digested.
- The harness also ran torch's own computation of this name under 6 named conventions; those rows are not implementations, are not counted, and only label a cluster whose digest they share.
Bucket 1 of 2: arg 1: rank 2, kind float, dtype float32
21 implementations from 93 papers share this bucket (rank, kind, dtype of each array argument, positional; each member's recorded signature, argument name included, is shown under it); 2 distinct outputs, largest cluster first. Values are the first 8 of the recorded output, flattened.
| Cluster (same digest to six decimals) | Members | Shared output on this bucket's input |
|---|---|---|
| 15 implementations 83 papers e9dcbf18fea6 recorded values identical same digest as torch: halves |
modeling_xalma.py b99eea63 recorded x:2/float/float32
time_moe/models/modeling_time_moe.py e03d53ba recorded x:2/float/float32
src/language/transformer.py 437c7011 recorded x:2/float/float32
src/gemma.py cec833a9 recorded x:2/float/float32
retro_pytorch/retro_pytorch.py fdd93453 recorded x:2/float/float32
palm_pytorch/palm_pytorch.py 08213c66 recorded x:2/float/float32 one code sha held from 2 papers' repositories modeling_mole.py 0ad8d888 recorded x:2/float/float32
models/dit_gumbel.py 8526ce17 recorded x:2/float/float32 perceiver_ar_pytorch/perceiver_ar_pytorch.py 1964789e recorded x:2/float/float32 SVDLLM.py 2a2393c4 recorded x:2/float/float32
model/cart.py 4da1a1be recorded x:2/float/float32
models/air/air_1net_L2x_H2x_input_token_prepend.py 5d0ae162 recorded x:2/float/float32
src/genrec/models/model_seqrec/sasrec_sprint.py 91bc58ef recorded x:2/float/float32 server/block_generate_server.py be9b02c5 recorded x:2/float/float32
src/flas/model.py e0782939 recorded x:2/float/float32 |
[0.902849, 1.03047, 0.662761, -0.572889, 1.80263, -1.53378, 0.476053, -0.690887, …] shape [4, 8] · float32 · Tensor |
| 6 implementations 10 papers ccd07400c61f recorded values identical same digest as torch: interleaved |
fairseq/modules/rotary_embedding.py 58823d94 recorded x:2/float/float32
lightglue/lightglue.py a3ec43c1 recorded x:2/float/float32
LatentMDM/model/latent_mdm.py 0c3c32d2 recorded x:2/float/float32 src/models/encoder.py 7082304f recorded x:2/float/float32 lam/lam/modules/lam.py a61b4182 recorded x:2/float/float32
models/FM/EEGPT/Model_EEGPT.py c1f873b7 recorded x:2/float/float32 |
[1.53378, 1.80263, 0.690887, 0.476053, 1.03047, -0.902849, -0.572889, -0.662761, …] shape [4, 8] · float32 · Tensor |
Bucket 2 of 2: arg 1: rank 3, kind float, dtype float32
3 implementations from 3 papers share this bucket (rank, kind, dtype of each array argument, positional; each member's recorded signature, argument name included, is shown under it); 2 distinct outputs, largest cluster first. Values are the first 8 of the recorded output, flattened.
| Cluster (same digest to six decimals) | Members | Shared output on this bucket's input |
|---|---|---|
| 2 implementations 2 papers 6dd639b65787 recorded values identical same digest as torch: interleaved |
fastvideo/layers/rotary_embedding_3d.py d727d58b recorded x:3/float/float32 hyvideo/modules/models.py dda8bc93 recorded x:3/float/float32 |
[-1.68285, -0.893127, 0.258206, -1.2588, 1.43476, 0.150996, 0.381126, -0.340146, …] shape [2, 4, 8] · float32 · Tensor |
| 1 implementation 1 paper 68fbd6edefdc same digest as torch: halves |
block_recurrent_transformer/transformer.py 6b01a575 recorded x:3/float/float32 |
[-0.150996, 1.43476, 0.340146, 0.381126, -0.893127, 1.68285, -1.2588, -0.258206, …] shape [2, 4, 8] · float32 · Tensor |
Identical values to six decimals (the recorded digest) on the shared input are agreement on those inputs, not equivalence; where a cluster's members carry recorded values, the largest difference among them is shown under the cluster. Paper titles are the archive's archive 2025-07-28 where the paper is in the archive and the graph's where it was added by Syntology; papers with no page here are shown by their recorded paper id only. A paper count above the implementation count means one implementation (one code sha) is held from several papers' repositories and counts once. Per-sample status, licence and fingerprint records for each paper are on its paper page. JSON twin: /census/rotate-half.json.
Report a problem or propose a change · a person checks every report against the paper or source before anything changes; decisions are listed on /corrections