From 2e8065050ee7cab67fcffc97fdb46d15e05fcddb Mon Sep 17 00:00:00 2001 From: Chang Liu <9713593+chang-l@users.noreply.github.com> Date: Thu, 10 Sep 2026 08:53:18 -0700 Subject: [PATCH 1/4] [None][doc] Group VisualGen feature guides under one navigation entry Signed-off-by: Chang Liu --- docs/source/features/visual-generation.rst | 20 ++++++++++++++++++++ docs/source/index.rst | 7 ++++--- 2 files changed, 24 insertions(+), 3 deletions(-) create mode 100644 docs/source/features/visual-generation.rst diff --git a/docs/source/features/visual-generation.rst b/docs/source/features/visual-generation.rst new file mode 100644 index 000000000000..0c2d1b072a91 --- /dev/null +++ b/docs/source/features/visual-generation.rst @@ -0,0 +1,20 @@ +.. SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. +.. SPDX-License-Identifier: Apache-2.0 + +Visual Generation (Beta) +======================== + +This section documents optimization features for VisualGen image and video +generation. Use the guides below for configuration, supported options, and +interactions between features. + +For supported models and an overview of VisualGen, see +:doc:`../models/visual-generation`. For usage examples, see +:doc:`../examples/visual_gen_examples`. + +.. toctree:: + :maxdepth: 1 + + CUDA Graphs + Quantized Attention + Sparse Attention diff --git a/docs/source/index.rst b/docs/source/index.rst index cd3bb7f95643..a9475b9b3f03 100644 --- a/docs/source/index.rst +++ b/docs/source/index.rst @@ -1,3 +1,6 @@ +.. SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. +.. SPDX-License-Identifier: Apache-2.0 + .. TensorRT LLM documentation master file, created by sphinx-quickstart on Wed Sep 20 08:35:21 2023. You can adapt this file completely to your liking, but it should at least @@ -89,9 +92,7 @@ Welcome to TensorRT LLM's Documentation! features/helix.md features/kv-cache-connector.md features/sparse-attention.md - features/visualgen-cuda-graph.md - features/visualgen-quantized-attention.md - features/visualgen-sparse-attention.md + features/visual-generation.rst .. toctree:: From ec48fe75986ed465317fbf31450cc38251e3744c Mon Sep 17 00:00:00 2001 From: Chang Liu <9713593+chang-l@users.noreply.github.com> Date: Thu, 10 Sep 2026 11:29:00 -0700 Subject: [PATCH 2/4] [None][doc] Link related VisualGen technical blogs from the feature overview Signed-off-by: Chang Liu --- docs/source/features/visual-generation.rst | 12 ++++++++++++ 1 file changed, 12 insertions(+) diff --git a/docs/source/features/visual-generation.rst b/docs/source/features/visual-generation.rst index 0c2d1b072a91..4216cf1727cc 100644 --- a/docs/source/features/visual-generation.rst +++ b/docs/source/features/visual-generation.rst @@ -18,3 +18,15 @@ For supported models and an overview of VisualGen, see CUDA Graphs Quantized Attention Sparse Attention + +Related technical blogs +----------------------- + +These articles explain the design choices and benchmark results behind +VisualGen optimizations: + +* :doc:`Scaling video generation across NVL72 <../blogs/tech_blog/blog25_Scaling_Video_Generation_Across_NVL72_Rack_with_TensorRT-LLM>` + covers multi-GPU parallelism and scaling results. +* :doc:`Quantization and sparse attention for video generation <../blogs/tech_blog/blog28_Accelerating_Video_Generation_with_GEMM_Quantization_Attention_Quantization_and_Skip_Softmax_Attention_in_TensorRT-LLM>` + covers GEMM quantization, quantized attention, and Skip Softmax Attention, + including performance and visual-quality trade-offs. From 5b286898c4cf842eda9eb2be972a966adae52d71 Mon Sep 17 00:00:00 2001 From: Chang Liu <9713593+chang-l@users.noreply.github.com> Date: Thu, 10 Sep 2026 12:32:29 -0700 Subject: [PATCH 3/4] [None][doc] Make the VisualGen overview readable in GitHub previews Signed-off-by: Chang Liu --- docs/source/conf.py | 11 ++++++++ docs/source/features/visual-generation.md | 32 ++++++++++++++++++++++ docs/source/features/visual-generation.rst | 32 ---------------------- docs/source/index.rst | 2 +- 4 files changed, 44 insertions(+), 33 deletions(-) create mode 100644 docs/source/features/visual-generation.md delete mode 100644 docs/source/features/visual-generation.rst diff --git a/docs/source/conf.py b/docs/source/conf.py index 3e5b9d942795..0a0ef3e1231b 100644 --- a/docs/source/conf.py +++ b/docs/source/conf.py @@ -11,6 +11,7 @@ # -- Project information ----------------------------------------------------- # https://www.sphinx-doc.org/en/master/usage/configuration.html#project-information import os +import re import subprocess import sys @@ -182,6 +183,15 @@ def tag_role(name, rawtext, text, lineno, inliner, options=None, content=None): return [node], [] +def _expand_visualgen_navigation(_app, docname: str, source: list[str]) -> None: + """Keep Sphinx navigation hidden in GitHub's Markdown preview.""" + if docname == 'features/visual-generation': + source[0] = re.sub(r'', + r'\1', + source[0], + flags=re.DOTALL) + + def setup(app): from helper import (check_llmapi_reference_size, compact_llmapi_search_signature, generate_examples, @@ -209,6 +219,7 @@ def setup(app): print(f"Warning: {msg}; skipping tag_llm_params") app.add_role('tag', tag_role) + app.connect('source-read', _expand_visualgen_navigation) app.connect('autodoc-process-docstring', strip_llmapi_search_docstrings) app.connect('autodoc-process-signature', compact_llmapi_search_signature) app.connect('build-finished', check_llmapi_reference_size) diff --git a/docs/source/features/visual-generation.md b/docs/source/features/visual-generation.md new file mode 100644 index 000000000000..ecaaf0c2b0d9 --- /dev/null +++ b/docs/source/features/visual-generation.md @@ -0,0 +1,32 @@ + + +# Visual Generation (Beta) + +This section documents optimization features for VisualGen image and video generation. Use the guides below for configuration, supported options, and interactions between features. + +For supported models and an overview of VisualGen, see the [model overview](../models/visual-generation.md). For usage examples, see [VisualGen Examples](https://nvidia.github.io/TensorRT-LLM/examples/visual_gen_examples.html). + +- [CUDA Graphs](visualgen-cuda-graph.md) +- [Quantized Attention](visualgen-quantized-attention.md) +- [Sparse Attention](visualgen-sparse-attention.md) + + + +## Related technical blogs + +These articles explain the design choices and benchmark results behind VisualGen optimizations: + +- [Scaling video generation across NVL72](../blogs/tech_blog/blog25_Scaling_Video_Generation_Across_NVL72_Rack_with_TensorRT-LLM.md) covers multi-GPU parallelism and scaling results. +- [Quantization and sparse attention for video generation](../blogs/tech_blog/blog28_Accelerating_Video_Generation_with_GEMM_Quantization_Attention_Quantization_and_Skip_Softmax_Attention_in_TensorRT-LLM.md) covers GEMM quantization, quantized attention, and Skip Softmax Attention, including performance and visual-quality trade-offs. diff --git a/docs/source/features/visual-generation.rst b/docs/source/features/visual-generation.rst deleted file mode 100644 index 4216cf1727cc..000000000000 --- a/docs/source/features/visual-generation.rst +++ /dev/null @@ -1,32 +0,0 @@ -.. SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. -.. SPDX-License-Identifier: Apache-2.0 - -Visual Generation (Beta) -======================== - -This section documents optimization features for VisualGen image and video -generation. Use the guides below for configuration, supported options, and -interactions between features. - -For supported models and an overview of VisualGen, see -:doc:`../models/visual-generation`. For usage examples, see -:doc:`../examples/visual_gen_examples`. - -.. toctree:: - :maxdepth: 1 - - CUDA Graphs - Quantized Attention - Sparse Attention - -Related technical blogs ------------------------ - -These articles explain the design choices and benchmark results behind -VisualGen optimizations: - -* :doc:`Scaling video generation across NVL72 <../blogs/tech_blog/blog25_Scaling_Video_Generation_Across_NVL72_Rack_with_TensorRT-LLM>` - covers multi-GPU parallelism and scaling results. -* :doc:`Quantization and sparse attention for video generation <../blogs/tech_blog/blog28_Accelerating_Video_Generation_with_GEMM_Quantization_Attention_Quantization_and_Skip_Softmax_Attention_in_TensorRT-LLM>` - covers GEMM quantization, quantized attention, and Skip Softmax Attention, - including performance and visual-quality trade-offs. diff --git a/docs/source/index.rst b/docs/source/index.rst index a9475b9b3f03..24c460d0e99e 100644 --- a/docs/source/index.rst +++ b/docs/source/index.rst @@ -92,7 +92,7 @@ Welcome to TensorRT LLM's Documentation! features/helix.md features/kv-cache-connector.md features/sparse-attention.md - features/visual-generation.rst + features/visual-generation.md .. toctree:: From e764ba4a1e3fa99d348537aed3ef936088666315 Mon Sep 17 00:00:00 2001 From: Chang Liu <9713593+chang-l@users.noreply.github.com> Date: Thu, 10 Sep 2026 13:38:38 -0700 Subject: [PATCH 4/4] [None][doc] Clarify the Skip Softmax blog link label Signed-off-by: Chang Liu --- docs/source/features/visual-generation.md | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/docs/source/features/visual-generation.md b/docs/source/features/visual-generation.md index ecaaf0c2b0d9..7a4bbcbd3110 100644 --- a/docs/source/features/visual-generation.md +++ b/docs/source/features/visual-generation.md @@ -29,4 +29,4 @@ Sparse Attention These articles explain the design choices and benchmark results behind VisualGen optimizations: - [Scaling video generation across NVL72](../blogs/tech_blog/blog25_Scaling_Video_Generation_Across_NVL72_Rack_with_TensorRT-LLM.md) covers multi-GPU parallelism and scaling results. -- [Quantization and sparse attention for video generation](../blogs/tech_blog/blog28_Accelerating_Video_Generation_with_GEMM_Quantization_Attention_Quantization_and_Skip_Softmax_Attention_in_TensorRT-LLM.md) covers GEMM quantization, quantized attention, and Skip Softmax Attention, including performance and visual-quality trade-offs. +- [Quantization and skipsoftmax attention for video generation](../blogs/tech_blog/blog28_Accelerating_Video_Generation_with_GEMM_Quantization_Attention_Quantization_and_Skip_Softmax_Attention_in_TensorRT-LLM.md) covers GEMM quantization, quantized attention, and Skip Softmax Attention, including performance and visual-quality trade-offs.