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

[geunhwan]

For resolving performance regression in many models.

[hyunback]

LGTM

[github-actions]

CI Doctor — Merge Queue failure on this PR

Pipeline: Linux (Ubuntu 22.04, Python 3.11)
Failure: Docker build fails - Azure Container Registry OAuth token fetch refused
Automatic restart: ✅ Re-run of failed jobs requested (reason: Transient Azure Container Registry connection refused — TCP connection to openvinogithubactions.azurecr.io failed, typically resolves on retry)

Possible remedy

1. Wait for the automatic re-run of the failed Docker job to complete — this is a transient network failure and should pass on the next attempt.
2. If the failure repeats, check the availability of openvinogithubactions.azurecr.io and the runner network routing to 172.16.0.197:443.
3. Consider adding retry logic (3 attempts with backoff) to .github/actions/handle_docker/images_api.py to handle transient ACR authentication failures gracefully.

What happened

The Docker job failed because the runner could not fetch an OAuth token from the internal Azure Container Registry mirror (openvinogithubactions.azurecr.io) — the TCP connection to 172.16.0.197:443 was refused during the docker buildx build step for ubuntu_22_04_x64_igpu/Dockerfile. This is a transient external service disruption unrelated to the PR's code changes.

Generated by CI Failure Doctor — Merge Queue · 168.6 AIC · ⌖ 6.42 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: Docker buildx builder boot fails - buildkit image not found in Azure Container Registry
Automatic restart: ✅ Re-run of failed jobs requested (reason: transient Docker buildx infrastructure failure; buildkit image was momentarily unavailable in the Azure Container Registry mirror)

Possible remedy

1. The re-run should resolve this automatically — the failure is not related to your code changes.
2. If the re-run also fails with the same error, check the health of openvinogithubactions.azurecr.io and the availability of dockerio/moby/buildkit:buildx-stable-1.
3. Consider asking the infrastructure team to add retry logic to the docker buildx create --bootstrap step in .github/actions/handle_docker.

What happened

The Docker_arm64 job failed when docker buildx inspect --bootstrap reported No such image: openvinogithubactions.azurecr.io/dockerio/moby/buildkit:buildx-stable-1 immediately after a seemingly successful pull (0.2s). This is a transient race condition in the Azure Container Registry mirror — the image pull completed but the image was not available to the container creation step. This is entirely unrelated to your GPU RMS normalization revert changes.

Generated by CI Failure Doctor — Merge Queue · 122.9 AIC · ⌖ 7.87 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_f32 Expected : ref_f32
Automatic restart: ❌ Not triggered — deterministic test assertion failure unlikely to be fixed by a restart

Possible remedy

1. Check whether smoke_ConvertCPULayerTest_4D_Dynamic with primitive=acl was already failing on master before this PR was queued.
2. Run in isolation: ./ov_cpu_func_tests --gtest_filter=*smoke_ConvertCPULayerTest_4D_Dynamic*primitive=acl* on an ARM64 cross-compiled build.
3. If the failure is pre-existing and unrelated to this revert, re-run the failed job or add the test to tests/functional_test_utils/layer_tests_summary/github/skip_configs/CPU/expected_failures_OP.csv.
4. If the failure is reproduced by this revert, investigate ACL dispatch for Convert f32→i8 nchw format on ARM64 (src/plugins/intel_cpu/tests/functional/utils/cpu_test_utils.cpp:248).

What happened

The CPU functional tests job failed on a single test: ConvertCPULayerTest (f32→i8, nchw, ACL primitive) reported primitive type unknown_f32 instead of the expected ref_f32. This PR reverts GPU RMS normalization changes and does not modify any CPU plugin or ACL code, so the failure is likely pre-existing or caused by a different concurrent merge-queue entry. (This was run attempt 2; attempt 1 failed with a Docker infrastructure issue.)

Generated by CI Failure Doctor — Merge Queue · 124.4 AIC · ⌖ 7.05 AIC · ⊞ 18.6K · ◷