Skip to content

fix(normalization): harden dynamic 4PE solution validation - #114

Closed
wangyuascend-spec wants to merge 1 commit into
PTO-ISA:mainfrom
wangyuascend-spec:fix/normalization-solution-validation
Closed

wangyuascend-spec wants to merge 1 commit into
PTO-ISA:mainfrom
wangyuascend-spec:fix/normalization-solution-validation

Conversation

@wangyuascend-spec

Copy link
Copy Markdown
Collaborator

Summary

Follow-up correction for the normalization solution merged in #113.

  • repartition GroupNormGrad by complete groups so dependent stages stay on one PE and no software spin barrier is required
  • align all normalization test buffers to 4 KiB
  • add benchmark markers for gfsim performance scope
  • remove unused references to the nonexistent _start_multipe.s
  • fix generator/comparator default compare paths

Validation

4PE gfrun accuracy on SuperScalarModel bc7fae00988a3296a849b5e138b83db433d6c4c8:

  • rms_norm [512, 8192]: PASS, max_abs=0.001953125
  • rms_norm_binary [16, 16384]: PASS, max_abs=0.001953125
  • group_norm_grad N=32,C=16,G=8,HxW=8192: PASS (dx/dgamma/dbeta)
  • group_norm_grad_1d N=512,C=64,G=8,D=8: PASS (dx/dgamma/dbeta)

Forced-fourpe gfsim smoke:

  • group_norm_grad N=1,C=16,G=8,HxW=256: PASS, reached SuperScalar Report Stop

Note: #113 is already merged, so it cannot be closed; this PR is its corrective follow-up.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant