From 4115f87701f1d29cd7832adfb0e0b447ea8ed801 Mon Sep 17 00:00:00 2001 From: Lawrence Lane Date: Wed, 1 Apr 2026 11:02:02 -0400 Subject: [PATCH 1/2] docs: add AEGIS classifier GPU utilization note and release notes (Issue #878) Document confirmed full GPU utilization for the AEGIS safety classifier on multi-GPU setups and add performance expectations note about the LlamaGuard-7b generative model being slower than encoder-based classifiers. Signed-off-by: Lawrence Lane Co-Authored-By: Claude Opus 4.6 Signed-off-by: Lawrence Lane --- fern/versions/v26.04/pages/about/release-notes/index.mdx | 6 ++++++ .../quality-assessment/distributed-classifier.mdx | 8 +++++++- 2 files changed, 13 insertions(+), 1 deletion(-) diff --git a/fern/versions/v26.04/pages/about/release-notes/index.mdx b/fern/versions/v26.04/pages/about/release-notes/index.mdx index a85242fbbe..67388465c2 100644 --- a/fern/versions/v26.04/pages/about/release-notes/index.mdx +++ b/fern/versions/v26.04/pages/about/release-notes/index.mdx @@ -25,6 +25,12 @@ Upgraded Cosmos-Xenna from 0.1.2 to 0.2.0 with a simplified resource model and i - **Cosmos-Xenna**: Updated from 0.1.2 to 0.2.0 with simplified resource model - **Ray**: Updated to 2.54 +## Improvements + +### AEGIS Classifier GPU Utilization (Issue #878) + +Confirmed full GPU utilization for the AEGIS safety classifier when running on multi-GPU setups. The AEGIS classifier, which uses the LlamaGuard-7b generative model, now properly distributes inference across all available GPUs. Added a performance note to the [classifier documentation](/curate-text/process-data/quality-assessment/distributed-classifier) to set expectations for processing times relative to encoder-based classifiers. + ## Bug Fixes ### Audio Stage Name Propagation (PR #1470) diff --git a/fern/versions/v26.04/pages/curate-text/process-data/quality-assessment/distributed-classifier.mdx b/fern/versions/v26.04/pages/curate-text/process-data/quality-assessment/distributed-classifier.mdx index d5fbfedb42..8801634231 100644 --- a/fern/versions/v26.04/pages/curate-text/process-data/quality-assessment/distributed-classifier.mdx +++ b/fern/versions/v26.04/pages/curate-text/process-data/quality-assessment/distributed-classifier.mdx @@ -167,7 +167,13 @@ pipeline.add_stage(writer) results = pipeline.run() # Uses XennaExecutor by default ``` -The classifier adds a column with labels: "safe," "O1" through "O13" (each representing specific safety risks), or "unknown." For raw LLM output, use: +The classifier adds a column with labels: "safe," "O1" through "O13" (each representing specific safety risks), or "unknown." + + +The AEGIS classifier relies on the [LlamaGuard-7b](https://huggingface.co/meta-llama/LlamaGuard-7b) base model, which is a generative LLM. This makes it significantly slower than the other classifiers in NeMo Curator that use encoder-based models (such as DeBERTa). Full GPU utilization is confirmed when running AEGIS on multi-GPU setups, but expect longer processing times compared to non-generative classifiers due to the autoregressive nature of the underlying model. + + +For raw LLM output, use: ```python safety_classifier = AegisClassifier( From 585a50e19655dd9fc8b5a8884b9be9d5c863f0f9 Mon Sep 17 00:00:00 2001 From: Lawrence Lane Date: Wed, 1 Apr 2026 11:08:13 -0400 Subject: [PATCH 2/2] docs: address Greptile review feedback on AEGIS docs Remove "now" from release notes to clarify this confirms existing behavior rather than implying a bug fix. Link to the NVIDIA AEGIS model page instead of Meta's base LlamaGuard-7b model. Signed-off-by: Lawrence Lane Co-Authored-By: Claude Opus 4.6 Signed-off-by: Lawrence Lane --- fern/versions/v26.04/pages/about/release-notes/index.mdx | 2 +- .../process-data/quality-assessment/distributed-classifier.mdx | 2 +- 2 files changed, 2 insertions(+), 2 deletions(-) diff --git a/fern/versions/v26.04/pages/about/release-notes/index.mdx b/fern/versions/v26.04/pages/about/release-notes/index.mdx index 67388465c2..83a3919f20 100644 --- a/fern/versions/v26.04/pages/about/release-notes/index.mdx +++ b/fern/versions/v26.04/pages/about/release-notes/index.mdx @@ -29,7 +29,7 @@ Upgraded Cosmos-Xenna from 0.1.2 to 0.2.0 with a simplified resource model and i ### AEGIS Classifier GPU Utilization (Issue #878) -Confirmed full GPU utilization for the AEGIS safety classifier when running on multi-GPU setups. The AEGIS classifier, which uses the LlamaGuard-7b generative model, now properly distributes inference across all available GPUs. Added a performance note to the [classifier documentation](/curate-text/process-data/quality-assessment/distributed-classifier) to set expectations for processing times relative to encoder-based classifiers. +Confirmed full GPU utilization for the AEGIS safety classifier when running on multi-GPU setups. The AEGIS classifier, which uses the LlamaGuard-7b generative model, properly distributes inference across all available GPUs. Added a performance note to the [classifier documentation](/curate-text/process-data/quality-assessment/distributed-classifier) to set expectations for processing times relative to encoder-based classifiers. ## Bug Fixes diff --git a/fern/versions/v26.04/pages/curate-text/process-data/quality-assessment/distributed-classifier.mdx b/fern/versions/v26.04/pages/curate-text/process-data/quality-assessment/distributed-classifier.mdx index 8801634231..fdd7e97fc1 100644 --- a/fern/versions/v26.04/pages/curate-text/process-data/quality-assessment/distributed-classifier.mdx +++ b/fern/versions/v26.04/pages/curate-text/process-data/quality-assessment/distributed-classifier.mdx @@ -170,7 +170,7 @@ results = pipeline.run() # Uses XennaExecutor by default The classifier adds a column with labels: "safe," "O1" through "O13" (each representing specific safety risks), or "unknown." -The AEGIS classifier relies on the [LlamaGuard-7b](https://huggingface.co/meta-llama/LlamaGuard-7b) base model, which is a generative LLM. This makes it significantly slower than the other classifiers in NeMo Curator that use encoder-based models (such as DeBERTa). Full GPU utilization is confirmed when running AEGIS on multi-GPU setups, but expect longer processing times compared to non-generative classifiers due to the autoregressive nature of the underlying model. +The AEGIS classifier relies on the [LlamaGuard-7b](https://huggingface.co/nvidia/Aegis-AI-Content-Safety-LlamaGuard-Defensive-1.0) base model, which is a generative LLM. This makes it significantly slower than the other classifiers in NeMo Curator that use encoder-based models (such as DeBERTa). Full GPU utilization is confirmed when running AEGIS on multi-GPU setups, but expect longer processing times compared to non-generative classifiers due to the autoregressive nature of the underlying model. For raw LLM output, use: