Skip to content

Compiled Benchmarks

Eager vs torch.compile performance and compatibility. Graph-break counts come from torch._dynamo.explain and are PyTorch-version dependent.

Metadata

  • Level: reduced
  • Date: 2026-11-09T23:01:52.662427+00:00
  • PyTorch: 2.15.0+cu132
  • CUDA: 13.2
  • GPU: NVIDIA GeForce RTX 3090
  • Python: 3.13.3
  • Commit: db0c77d
  • Timing protocol: 5 warmup, 20 measured iterations

Compatibility Matrix

Operator Level Method Device Config Class Graph breaks Max rel diff Error Details
Add wrapper adj cpu default works 0 0 details
Add wrapper adj cuda default works 0 0 details
Add wrapper fwd cpu default works 0 0 details
Add wrapper fwd cuda default works 0 0 details
ArrayToBlocks wrapper adj cpu default works 0 0 details
ArrayToBlocks wrapper adj cuda default works 0 0 details
ArrayToBlocks wrapper fwd cpu default works 0 0 details
ArrayToBlocks wrapper fwd cuda default works 0 0 details
BlocksToArray wrapper adj cpu default works 0 0 details
BlocksToArray wrapper adj cuda default works 0 0 details
BlocksToArray wrapper fwd cpu default works 0 0 details
BlocksToArray wrapper fwd cuda default works 0 0 details
Chain wrapper adj cpu default works 0 0 details
Chain wrapper adj cuda default works 0 0 details
Chain wrapper fwd cpu default works 0 0 details
Chain wrapper fwd cuda default works 0 0 details
Concat wrapper adj cpu default works 0 0 details
Concat wrapper adj cuda default works_with_breaks 1 0 details
Concat wrapper fwd cpu default works 0 0 details
Concat wrapper fwd cuda default works_with_breaks 1 0 details
Convolution inner adj cpu default compiles_slower 0 0 details
Convolution inner adj cuda default compiles_slower 0 0 details
Convolution inner fwd cpu default compiles_slower 0 0 details
Convolution inner fwd cuda default compiles_slower 0 0 details
Convolution wrapper adj cpu default works 0 0 details
Convolution wrapper adj cuda default compiles_slower 0 0 details
Convolution wrapper fwd cpu default works 0 0 details
Convolution wrapper fwd cuda default compiles_slower 0 0 details
Dense wrapper adj cpu default works 0 0 details
Dense wrapper adj cuda default works 0 0 details
Dense wrapper fwd cpu default works 0 0 details
Dense wrapper fwd cuda default works 0 0 details
DenseSimple wrapper adj cpu default works 0 0 details
DenseSimple wrapper adj cuda default works 0 0 details
DenseSimple wrapper fwd cpu default works 0 0 details
DenseSimple wrapper fwd cuda default works 0 0 details
Diagonal wrapper adj cpu default compiles_slower 0 0 details
Diagonal wrapper adj cuda default compiles_slower 0 0 details
Diagonal wrapper fwd cpu default compiles_slower 0 0 details
Diagonal wrapper fwd cuda default compiles_slower 0 0 details
FFT inner adj cpu default compiles_slower 0 0 details
FFT inner adj cuda default compiles_slower 0 0 details
FFT inner fwd cpu default compiles_slower 0 0 details
FFT inner fwd cuda default compiles_slower 0 0 details
FFT wrapper adj cpu default compiles_slower 0 0 details
FFT wrapper adj cuda default compiles_slower 0 0 details
FFT wrapper fwd cpu default compiles_slower 0 0 details
FFT wrapper fwd cuda default compiles_slower 0 0 details
FastSamplingNUFFT inner adj cpu default works_with_breaks 3 0 details
FastSamplingNUFFT inner adj cuda default works_with_breaks 3 0 details
FastSamplingNUFFT inner fwd cpu default works 0 0 details
FastSamplingNUFFT inner fwd cuda default works 0 0 details
FastSamplingNUFFT wrapper adj cpu default works_with_breaks 3 0 details
FastSamplingNUFFT wrapper adj cuda default works_with_breaks 4 0 details
FastSamplingNUFFT wrapper fwd cpu default works 0 0 details
FastSamplingNUFFT wrapper fwd cuda default works 0 0 details
Identity wrapper adj cpu default compiles_slower 0 0 details
Identity wrapper adj cuda default compiles_slower 0 0 details
Identity wrapper fwd cpu default compiles_slower 0 0 details
Identity wrapper fwd cuda default compiles_slower 0 0 details
Interpolate inner adj cpu default works 0 0 details
Interpolate inner adj cuda default works 0 0 details
Interpolate inner fwd cpu default works 0 0 details
Interpolate inner fwd cuda default works 0 0 details
Interpolate wrapper adj cpu default works 0 0 details
Interpolate wrapper adj cuda default works 0 0 details
Interpolate wrapper fwd cpu default works 0 0 details
Interpolate wrapper fwd cuda default works 0 0 details
NUFFT inner adj cpu default works 0 0 details
NUFFT inner adj cuda default compiles_slower 0 0 details
NUFFT inner fwd cpu default works 0 0 details
NUFFT inner fwd cuda default compiles_slower 0 0 details
NUFFT wrapper adj cpu default works 0 0 details
NUFFT wrapper adj cuda default compiles_slower 0 0 details
NUFFT wrapper fwd cpu default works 0 0 details
NUFFT wrapper fwd cuda default silent_degradation 0 0 details
Pad wrapper adj cpu default compiles_slower 0 0 details
Pad wrapper adj cuda default compiles_slower 0 0 details
Pad wrapper fwd cpu default works 0 0 details
Pad wrapper fwd cuda default works 0 0 details
PadDim wrapper adj cpu default compiles_slower 0 0 details
PadDim wrapper adj cuda default compiles_slower 0 0 details
PadDim wrapper fwd cpu default works 0 0 details
PadDim wrapper fwd cuda default works 0 0 details
Rearrange wrapper adj cpu default compiles_slower 0 0 details
Rearrange wrapper adj cuda default compiles_slower 0 0 details
Rearrange wrapper fwd cpu default compiles_slower 0 0 details
Rearrange wrapper fwd cuda default compiles_slower 0 0 details
Repeat wrapper adj cpu default works 0 0 details
Repeat wrapper adj cuda default works 0 6.31e-07 details
Repeat wrapper fwd cpu default works 0 0 details
Repeat wrapper fwd cuda default works 0 0 details
Sampling wrapper adj cpu default works 0 0 details
Sampling wrapper adj cuda default works 0 0 details
Sampling wrapper fwd cpu default works 0 0 details
Sampling wrapper fwd cuda default works 0 0 details
SamplingNUFFT inner adj cpu default works_with_breaks 1 0 details
SamplingNUFFT inner adj cuda default works_with_breaks 1 0 details
SamplingNUFFT inner fwd cpu default works 0 0 details
SamplingNUFFT inner fwd cuda default compiles_slower 0 0 details
SamplingNUFFT wrapper adj cpu default works_with_breaks 1 0 details
SamplingNUFFT wrapper adj cuda default works_with_breaks 2 0 details
SamplingNUFFT wrapper fwd cpu default works 0 0 details
SamplingNUFFT wrapper fwd cuda default compiles_slower 0 0 details
Scalar wrapper adj cpu default compiles_slower 0 0 details
Scalar wrapper adj cuda default compiles_slower 0 0 details
Scalar wrapper fwd cpu default compiles_slower 0 0 details
Scalar wrapper fwd cuda default compiles_slower 0 0 details
ShapeSpec wrapper adj cpu default compiles_slower 0 0 details
ShapeSpec wrapper adj cuda default compiles_slower 0 0 details
ShapeSpec wrapper fwd cpu default compiles_slower 0 0 details
ShapeSpec wrapper fwd cuda default compiles_slower 0 0 details
SimpleLinop wrapper adj cpu default works 0 0 details
SimpleLinop wrapper adj cuda default compiles_slower 0 0 details
SimpleLinop wrapper fwd cpu default works 0 0 details
SimpleLinop wrapper fwd cuda default compiles_slower 0 0 details
Stack wrapper adj cpu default compiles_slower 0 0 details
Stack wrapper adj cuda default works_with_breaks 1 0 details
Stack wrapper fwd cpu default works 0 0 details
Stack wrapper fwd cuda default works_with_breaks 1 0 details
SumReduce wrapper adj cpu default works 0 0 details
SumReduce wrapper adj cuda default works 0 0 details
SumReduce wrapper fwd cpu default works 0 0 details
SumReduce wrapper fwd cuda default works 0 1.11e-07 details
ToDevice wrapper adj cpu default works 0 0 details
ToDevice wrapper adj cuda default works 0 0 details
ToDevice wrapper fwd cpu default works 0 0 details
ToDevice wrapper fwd cuda default works 0 0 details
ToeplitzPSF inner adj cpu default compiles_slower 0 0 details
ToeplitzPSF inner adj cuda default compiles_slower 0 0 details
ToeplitzPSF inner fwd cpu default compiles_slower 0 0 details
ToeplitzPSF inner fwd cuda default compiles_slower 0 0 details
ToeplitzPSF wrapper adj cpu default compiles_slower 0 0 details
ToeplitzPSF wrapper adj cuda default compiles_slower 0 0 details
ToeplitzPSF wrapper fwd cpu default compiles_slower 0 0 details
ToeplitzPSF wrapper fwd cuda default compiles_slower 0 0 details
Truncate wrapper adj cpu default works 0 0 details
Truncate wrapper adj cuda default works 0 0 details
Truncate wrapper fwd cpu default compiles_slower 0 0 details
Truncate wrapper fwd cuda default compiles_slower 0 0 details
Zero wrapper adj cpu default compiles_slower 0 0 details
Zero wrapper adj cuda default compiles_slower 0 0 details
Zero wrapper fwd cpu default compiles_slower 0 0 details
Zero wrapper fwd cuda default compiles_slower 0 0 details
functional.fold inner fwd cpu default works 0 0 details
functional.fold inner fwd cuda default works 0 0 details
functional.grid inner fwd cpu default works 0 0 details
functional.grid inner fwd cuda default compiles_slower 0 0 details
functional.unfold inner fwd cpu default works 0 0 details
functional.unfold inner fwd cuda default works 0 0 details
functional.ungrid inner fwd cpu default works 0 0 details
functional.ungrid inner fwd cuda default compiles_slower 0 0 details

Graph-break details

  • Concat (wrapper, adj, cuda, default) — 1 breaks
  • /home/mark/.local/share/opencode/worktree/d3d230/compiled-benchmarks/src/torchlinops/linops/namedlinop.py:151 in forward: Unsupported method call
  • Concat (wrapper, fwd, cpu, default) — 0 breaks
  • /home/mark/.local/share/opencode/worktree/d3d230/compiled-benchmarks/src/torchlinops/linops/concat.py:173 in _fn: generic_jump TensorVariable()
  • Concat (wrapper, fwd, cuda, default) — 1 breaks
  • /home/mark/.local/share/opencode/worktree/d3d230/compiled-benchmarks/src/torchlinops/linops/namedlinop.py:151 in forward: Data-dependent branching
  • /home/mark/.local/share/opencode/worktree/d3d230/compiled-benchmarks/src/torchlinops/linops/concat.py:173 in _fn: generic_jump TensorVariable()
  • FastSamplingNUFFT (wrapper, adj, cpu, default) — 3 breaks
  • FastSamplingNUFFT (inner, adj, cpu, default) — 3 breaks
  • FastSamplingNUFFT (wrapper, adj, cuda, default) — 4 breaks
  • /home/mark/.local/share/opencode/worktree/d3d230/compiled-benchmarks/src/torchlinops/linops/namedlinop.py:151 in forward: Data-dependent branching
  • FastSamplingNUFFT (inner, adj, cuda, default) — 3 breaks
  • SamplingNUFFT (wrapper, adj, cpu, default) — 1 breaks
  • SamplingNUFFT (inner, adj, cpu, default) — 1 breaks
  • SamplingNUFFT (wrapper, adj, cuda, default) — 2 breaks
  • /home/mark/.local/share/opencode/worktree/d3d230/compiled-benchmarks/src/torchlinops/linops/namedlinop.py:151 in forward: Data-dependent branching
  • SamplingNUFFT (inner, adj, cuda, default) — 1 breaks
  • Stack (wrapper, adj, cpu, default) — 0 breaks
  • /home/mark/.local/share/opencode/worktree/d3d230/compiled-benchmarks/src/torchlinops/linops/stack.py:213 in _fn: Unsupported method call
  • Stack (wrapper, adj, cuda, default) — 1 breaks
  • /home/mark/.local/share/opencode/worktree/d3d230/compiled-benchmarks/src/torchlinops/linops/namedlinop.py:151 in forward: Unsupported method call
  • /home/mark/.local/share/opencode/worktree/d3d230/compiled-benchmarks/src/torchlinops/linops/stack.py:213 in _fn: Unsupported method call
  • Stack (wrapper, fwd, cuda, default) — 1 breaks
  • /home/mark/.local/share/opencode/worktree/d3d230/compiled-benchmarks/src/torchlinops/linops/namedlinop.py:151 in forward: Unsupported method call

Summary

Effective classes

Class Rows
compiles_slower 62
silent_degradation 1
works 77
works_with_breaks 12

Rows by device

Device Rows
cpu 76
cuda 76

Recompilation

These sections require the full suite.

Sections

Full-only Phases

Autograd and pipeline checks require the full suite; this page was generated from a reduced run.


See Methodology for timing, graph-break, and level definitions.