Skip to content

Tighten rtol/atol from 1e-2 to 1e-3 for Helion kernels#132

Merged
msaroufim merged 1 commit intogpu-mode:mainfrom
yf225:tighten-helion-tolerances
Mar 14, 2026
Merged

Tighten rtol/atol from 1e-2 to 1e-3 for Helion kernels#132
msaroufim merged 1 commit intogpu-mode:mainfrom
yf225:tighten-helion-tolerances

Conversation

@yf225
Copy link
Copy Markdown
Contributor

@yf225 yf225 commented Mar 14, 2026 •
edited
Loading

Summary

  • Tighten tolerances from rtol=1e-2, atol=1e-2 to rtol=1e-3, atol=1e-3 for all four Helion kernels: causal_conv1d, gated_deltanet_chunk_fwd_h, chunk_fwd_o, and recompute_w_u
  • These kernels operate entirely in float32, so 1% tolerance was overly loose and could mask numerical bugs
  • Verified that existing submissions for causal_conv1d, chunk_fwd_o, and recompute_w_u pass at 1e-3

Test plan

  • causal_conv1d: all 5 tests pass at 1e-3
  • gated_deltanet_chunk_fwd_o: all tests pass at 1e-3
  • gated_deltanet_recompute_w_u: all tests pass at 1e-3

🤖 Generated with Claude Code

@yf225 @claude
Tighten rtol/atol from 1e-2 to 1e-3 for Helion kernels
ce0b6f7 
All four Helion kernels (causal_conv1d, gated_deltanet_chunk_fwd_h,
chunk_fwd_o, recompute_w_u) operate in float32 but used overly loose
tolerances of rtol=1e-2, atol=1e-2. Tighten to 1e-3 to better catch
numerical bugs while still allowing for accumulation chain error.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
@msaroufim msaroufim merged commit 05d42ec into gpu-mode:main Mar 14, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

@msaroufim msaroufim msaroufim approved these changes

Assignees

No one assigned

Labels

None yet

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

2 participants

@yf225 @msaroufim