Skip to content

Conversation

@v-Golubev
Copy link
Contributor

@v-Golubev v-Golubev commented Nov 18, 2025

Details:

  • Fixed precision enforcement issue via workaround for precision enforcement pipeline: do not enforce precisions if precision enforcement breaks transpose fusion. Although this leads to non-optimal performance, such cases (Subgraph input precisions are f32, but bf16 is forced internally) don't occur in real scenarios.
  • The changes were validated on large model scope to confirm that there are no perf degradations on real models
  • Tests are extended to cover the fixed scenario (f32 in/out precision and bf16 inference precision)

Tickets:

@github-actions github-actions bot added category: IE Tests OpenVINO Test: plugins and common category: CPU OpenVINO CPU plugin category: snippets labels Nov 18, 2025
@v-Golubev v-Golubev force-pushed the vg/snippets/precision_enforcement_fix branch 2 times, most recently from 1a8a21c to 6a1f37e Compare November 19, 2025 15:14
@v-Golubev v-Golubev marked this pull request as ready for review November 20, 2025 09:18
@v-Golubev v-Golubev requested review from a team as code owners November 20, 2025 09:18
@v-Golubev v-Golubev force-pushed the vg/snippets/precision_enforcement_fix branch from 63cad4b to 7b6bf77 Compare November 20, 2025 09:19
@v-Golubev v-Golubev added this to the 2026.0 milestone Nov 21, 2025
@v-Golubev v-Golubev enabled auto-merge November 24, 2025 12:43
@v-Golubev v-Golubev added this pull request to the merge queue Nov 24, 2025
Merged via the queue into openvinotoolkit:master with commit 792ddf3 Nov 24, 2025
207 checks passed
@v-Golubev v-Golubev deleted the vg/snippets/precision_enforcement_fix branch November 24, 2025 21:36
ababushk added a commit that referenced this pull request Nov 25, 2025
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

category: CPU OpenVINO CPU plugin category: IE Tests OpenVINO Test: plugins and common category: snippets

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants