More detailed documentation for recipes

[pggPL]

Description

This PR adds detailed documentation for the Low Precision Training feature in Transformer Engine, covering FP8, MXFP8, NVFP4, and other quantization recipes for both PyTorch and JAX frameworks.

Type of change

• Documentation change (change only to the documentation, either a fix or a new content)
• Bug fix (non-breaking change which fixes an issue)
• New feature (non-breaking change which adds functionality)
• Breaking change (fix or feature that would cause existing functionality to not work as expected)
• Infra/Build change
• Code refactoring

Checklist:

• I have read and followed the contributing guidelines
• The functionality is complete
• I have commented my code, particularly in hard-to-understand areas
• I have made corresponding changes to the documentation
• My changes generate no new warnings
• I have added tests that prove my fix is effective or that my feature works
• New and existing unit tests pass locally with my changes

[greptile-apps]

Greptile Overview

Greptile Summary

This PR adds comprehensive documentation for Transformer Engine's low precision training capabilities, covering FP8, MXFP8, and NVFP4 quantization recipes for both PyTorch and JAX frameworks.

Key Additions

• Introduction documentation explaining mixed precision training fundamentals, BF16/FP16 formats, master weights, and the autocast API for both frameworks
• Recipe-specific documentation for five quantization approaches: FP8 Current Scaling, FP8 Delayed Scaling, FP8 Block Scaling, MXFP8, and NVFP4, each with detailed technical explanations and framework-specific examples
• Performance optimization guide covering transpose handling, memory efficiency, fused layers, and architecture-specific considerations (Hopper vs Blackwell)
• 28 SVG diagrams illustrating concepts like scaling factors, data formats, tensor layouts, and distributed training flows
• Code examples for both PyTorch and JAX demonstrating recipe usage, distributed training, and performance measurement
• CSS styling for diagram colors, responsive SVGs, and tab navigation

Documentation Quality

The documentation is well-structured with clear progression from basic concepts to advanced optimization. Technical explanations are thorough, diagrams effectively illustrate complex concepts, and code examples cover both single-GPU and distributed scenarios.

Minor Issues

• One arXiv reference appears to use a future date (line 10 in nvfp4.rst) - should be verified
• One example could be more explicit about which recipe is being used (memory_usage_2_pytorch.py)

All previously identified issues from earlier review rounds have been addressed.

Confidence Score: 5/5

• This PR is safe to merge with minimal risk as it only adds documentation
• Documentation-only PR with comprehensive content covering low precision training recipes. All code examples are non-executable documentation snippets. Previous review issues have been addressed. Only minor style suggestions remain (arXiv reference verification and explicit recipe parameter).
• No files require special attention

Important Files Changed

Filename	Overview
docs/features/low_precision_training/introduction/introduction.rst	Comprehensive introduction to mixed precision and FP8 training concepts
docs/features/low_precision_training/fp8_current_scaling/fp8_current_scaling.rst	Detailed documentation for FP8 Current Scaling recipe with examples
docs/features/low_precision_training/fp8_delayed_scaling/fp8_delayed_scaling.rst	Explains FP8 Delayed Scaling with historical amax management
docs/features/low_precision_training/performance_considerations/performance_considerations.rst	In-depth performance optimization guidance for transpose handling and memory
docs/features/low_precision_training/nvfp4/nvfp4.rst	Documents NVFP4 4-bit precision recipe with hierarchical scaling

Sequence Diagram

sequenceDiagram
    participant User as User/Developer
    participant Docs as Documentation
    participant Intro as Introduction
    participant Recipes as Recipe Docs<br/>(FP8/MXFP8/NVFP4)
    participant Perf as Performance Guide
    participant Examples as Code Examples

    User->>Docs: Read low precision training docs
    Docs->>Intro: 1. Learn mixed precision basics
    Note over Intro: BF16/FP16 concepts<br/>Master weights<br/>Autocast usage
    
    Intro->>Recipes: 2. Choose quantization recipe
    Note over Recipes: FP8 Current Scaling<br/>FP8 Delayed Scaling<br/>FP8 Block Scaling<br/>MXFP8<br/>NVFP4
    
    Recipes->>Examples: 3. Review framework examples
    Note over Examples: PyTorch examples<br/>JAX examples<br/>Distributed training
    
    Examples->>Perf: 4. Optimize performance
    Note over Perf: Transpose handling<br/>Memory optimization<br/>Fused layers
    
    Perf->>User: 5. Implement in production
    Note over User: Apply recipe with<br/>te.autocast(recipe=...)

Loading

[greptile-apps]

_{46 files reviewed, no comments}

_{Edit Code Review Agent Settings | Greptile}

[greptile-apps]

_{2 files reviewed, 2 comments}

_{Edit Code Review Agent Settings | Greptile}

[greptile-apps]

_{1 file reviewed, 1 comment}

_{Edit Code Review Agent Settings | Greptile}

[greptile-apps]

_{2 files reviewed, 2 comments}

_{Edit Code Review Agent Settings | Greptile}

[greptile-apps]

_{3 files reviewed, 3 comments}

_{Edit Code Review Agent Settings | Greptile}

[greptile-apps]

_{1 file reviewed, 1 comment}

_{Edit Code Review Agent Settings | Greptile}

[greptile-apps]

_{1 file reviewed, 1 comment}

_{Edit Code Review Agent Settings | Greptile}

[greptile-apps]

_{No files reviewed, no comments}

_{Edit Code Review Agent Settings | Greptile}

[greptile-apps]

_{3 files reviewed, 3 comments}

_{Edit Code Review Agent Settings | Greptile}

[jberchtold-nvidia]

Fantastic work! The diagrams and explanations are great and along with the code examples will hopefully make things a lot clearer to users

[pggPL]

• emulated = power of 2 scaling factor only
• remove the compatibility part.

[pggPL]

Comment about SM 12

[greptile-apps]

_{5 files reviewed, no comments}

_{Edit Code Review Agent Settings | Greptile}

[greptile-apps]

_{5 files reviewed, 2 comments}

_{Edit Code Review Agent Settings | Greptile}

[greptile-apps]

arXiv ID 2509.25149 uses prefix 2509 (September 2025), which is in the future. Verify this is the correct reference or use a placeholder format like [arXiv link pending] until the paper is published.

[greptile-apps]

Missing recipe parameter. te.autocast(enabled=True) uses a default recipe, but for documentation clarity, explicitly specify which recipe is being used (e.g., recipe=DelayedScaling()) to match other examples.

_{Note: If this suggestion doesn't match your team's coding style, reply to this and let me know. I'll remember it for next time!}