Revert "[GPU] Keep RMS normalization nodes in FP32 to prevent NaN from FP16 overflow (#36504)"

[jade-cho]

Details:

• Revert PR [GPU] Keep RMS normalization nodes in FP32 to prevent NaN from FP16 overflow #36504 which forces all RMS normalization nodes to FP32
• This change causes significant global LLM performance regression.

Root Cause

PR #36504 was introduced to fix NaN in ParlerTTS Mini (CVS-187992) by keeping RMS nodes in FP32. However, it unconditionally marks ALL RMS nodes across all models, adding FP16↔FP32 conversion overhead on every transformer layer.

Tickets:

• 190812

[github-actions]

CI Doctor — Merge Queue failure on this PR

Pipeline: Windows (VS 2022, Python 3.11, Release)
Failure: JS API E2E Electron cleanup fails on Windows - Permission denied removing locked electron.exe/dll
Automatic restart: ✅ Re-run of failed jobs requested (reason: Windows file-locking flaky test — all E2E tests passed; only the Mocha after-all cleanup hook failed because electron.exe/DLLs were locked on the Windows runner at cleanup time)

Possible remedy

1. Fix the Mocha after all hook in electron-app.test.js to explicitly kill/wait for the Electron process before attempting rm -rf
2. Replace bare rm -rf with rimraf (has built-in Windows retry logic) for the cleanup step
3. Alternatively, add taskkill /f /im electron.exe (Windows) or a process kill step before directory removal

What happened

Both E2E tests passed (2 passing). The failure occurred only in the cleanup hook: rm -rf demo-electron-app-project was blocked by Windows file locking on electron.exe, d3dcompiler_47.dll, libEGL.dll, and libGLESv2.dll — the Electron process or its loaded DLLs were still held by the OS at cleanup time.

This is a known recurring flaky issue (5th occurrence since 2026-07-09, same pattern on PRs #36887, #36796, #36763, #36780). It is unrelated to the GPU RMS normalization revert in this PR.

Generated by CI Failure Doctor — Merge Queue · 157.1 AIC · ⌖ 9.42 AIC · ⊞ 18.6K · ◷

[github-actions]

CI Doctor — Merge Queue failure on this PR

Pipeline: Windows (VS 2022, Python 3.11, Release)
Failure: JS API E2E Electron cleanup fails on Windows - Permission denied removing locked electron.exe/dll
Automatic restart: ❌ Not triggered — run is already at attempt 2; re-running would risk a restart loop.

Possible remedy

1. This is a known transient flaky test (pattern seen 5 times since 2026-07-09). The actual E2E tests pass — only the "after all" cleanup hook fails.
2. Re-queue this PR in the merge queue to get a fresh CI run.
3. To fix permanently: replace rm -rf with rimraf (v4+) in src/bindings/js/node/tests/e2e/electron-app.test.js — it has built-in Windows retry logic for file-lock errors.
4. See prior occurrences: Run 29450212856 (PR [NPUW]MoE optimization - RoPE cache and unfold infer request. #36887), Run 29072781821 (PR [core] Add helpers for common operations in mmap tests #36763).

What happened

The JS API (22) job failed in the Mocha "after all" cleanup hook: rm -rf demo-electron-app-project could not delete node_modules/electron/dist/electron.exe because Windows held a file lock after the Electron process exited. Both functional E2E tests (should install dependencies, should run electron package and verify output) passed — only teardown failed.

This is a recurring Windows file-lock race (5 occurrences). It is unrelated to the GPU RMS normalization changes in this PR.

Generated by CI Failure Doctor — Merge Queue · 108.3 AIC · ⌖ 6 AIC · ⊞ 18.6K · ◷

[github-actions]

CI Doctor — Merge Queue failure on this PR

Pipeline: Linux (Ubuntu 22.04, ARM64 cross-compilation, Python 3.11)
Failure: smoke_ConvertCPULayerTest - primType is unexpected : unknown_i32 Expected : acl_i32
Automatic restart: ❌ Not triggered — deterministic test assertion failure that a restart cannot fix.

Possible remedy

1. Check whether this test was already failing on the releases/2026/3 base before this PR — a similar failure was seen on 2026-07-08 (run 28957097935) against a different PR, suggesting a pre-existing issue.
2. Inspect ACL Convert kernel type registration in src/plugins/intel_cpu/src/nodes/convert.cpp to confirm i32 is listed as a supported input type for the ACL path.
3. If pre-existing, add the failing test variant to tests/functional_test_utils/layer_tests_summary/github/skip_configs/CPU/expected_failures_OP.csv for the ARM64/ACL configuration.
4. If introduced by this revert, identify whether the reverted PR [GPU] Keep RMS normalization nodes in FP32 to prevent NaN from FP16 overflow #36504 incidentally enabled ACL i32 support and ensure that support is preserved separately.

What happened

The CPU functional test smoke_ConvertCPULayerTest_4D_Dynamic failed on the ARM64 runner: the test expects the ARM Compute Library (ACL) backend to handle an i32-to-i8 Convert op with primitive type acl_i32, but the runtime returned unknown_i32 (src/plugins/intel_cpu/tests/functional/utils/cpu_test_utils.cpp:248). This PR reverts a GPU-only change (RMS normalization FP32), so the failing CPU/ACL test is likely a pre-existing issue surfacing independently in the merge queue.

Generated by CI Failure Doctor — Merge Queue · 93.5 AIC · ⌖ 6.32 AIC · ⊞ 18.6K · ◷