docs: add comprehensive pipeline parallelism comparison guide

[yurekami]

Summary

This PR adds a comprehensive comparison guide (docs/COMPARISON.md) for pipeline parallelism methods, directly addressing the questions raised in issues #15 and #20.

Motivation

Several users have asked for clarification on:

• Issue Comparison between VPP and DualPipe #15: How does VPP compare to DualPipe? When should each be used?
• Issue What is the benifit of DualPipeV? #20: What are the benefits of DualPipeV?

This documentation provides detailed answers with tables, formulas, and decision guidelines.

Documentation Contents

Methods Compared

• 1F1B - Traditional pipeline parallelism
• VPP - Megatron-LM's Virtual Pipeline Parallelism
• ZB1P - Zero Bubble scheduling
• DualPipe - DeepSeek's bidirectional pipeline
• DualPipeV - V-shape half-device variant

Sections Included

1. Overview Table - Quick comparison of all methods
2. Method Descriptions - Detailed explanation of each approach
3. Bubble Ratio Analysis - Formulas and example calculations
4. Memory Trade-offs - Parameter and activation memory comparison
5. Communication Patterns - Network requirements for each method
6. When to Use Each Method - Practical guidance
7. MoE with EP Considerations - Special notes for Expert Parallelism
8. Decision Flowchart - Step-by-step guide for choosing

Key Insights Documented

• When vpp > 3, VPP's bubble ratio becomes competitive with DualPipe
• DualPipe achieves low bubble through overlap rather than interleaving
• DualPipeV provides DualPipe-like benefits with half the devices
• For MoE+EP, compare methods using the same number of PP stages

Test Plan

• Verified markdown renders correctly
• Cross-referenced with README comparison table
• Validated formulas against DeepSeek-V3 paper
• Checked references are accurate

Closes #15
Closes #20

🤖 Generated with Claude Code

[hiSandog]

This comparison doc would be very useful for the recurring VPP vs DualPipe questions.

One thing I would tighten before merge is the assumptions behind the bubble formulas: whether acc means gradient accumulation microbatches, whether forward/backward costs are assumed balanced, and whether communication time is excluded or hidden. Stating those assumptions near the overview table would prevent the ratios from being read as exact throughput predictions.

[yurekami]

@hiSandog Good point, added in 81e55b7. The note under the Overview table now states all three assumptions: acc is gradient-accumulation microbatches per optimizer step, forward and backward chunk costs are treated as equal (backward is roughly 2x forward in practice, which shifts the constants but not the trends), and communication time is excluded, assumed fully overlapped with computation. The Bubble Ratio Analysis section references the same assumptions, and the ratios are now labeled scheduling-model estimates rather than throughput predictions.