feat(issue8): MoE combine three-mode precision/traffic trade-off analysis

[xyaohubery]

Summary

Analytical model and benchmark framework for DeepEP V2's three combine reduction modes, addressing Issue #8.

Three Modes Analyzed

Mode	Template Config	Trade-off
A (No-Expand)	kUseExpandedLayout=false, kAllowMultipleReduction=false	Minimal SM, best precision, 1:1 token mapping
B (Local Reduce)	kUseExpandedLayout=true, kAllowMultipleReduction=true	Reduced network traffic via shared-memory BF16 reduction
C (Expanded Send)	kUseExpandedLayout=true, kAllowMultipleReduction=false	Highest traffic, FP32 epilogue accuracy

What's Included

• combine_mode_analysis.py: Analytical model with configurable hardware/workload parameters, based on the actual DeepEP kernel implementations
• test_compare_modes.py: Benchmark runner supporting both simulation mode and real GPU hardware
• combine_decision_table.json: Pre-computed decision table across 231 configurations (repetition rate × tokens × hidden dim)
• README.md: Documentation and quick start guide

Verification

MoE Combine Mode Analysis — DeepEP V2
========================================================================
             Mode A               Mode B               Mode C
             (No Expand)          (Local Reduce)       (Expanded Send)
------------------------------------------------------------------------
Net Traffic    0.0587 GB           0.0433 GB           0.4698 GB
Net Latency       65.2 μs             48.1 μs            522.0 μs
Precision         7.0 bits             6.2 bits            23.0 bits

Key Finding

For typical MoE workloads (repetition_rate ~0.3), Mode B is recommended:

• 26% less network traffic than Mode A
• 26% lower latency than Mode A
• Acceptable precision loss (< 2 ULP in BF16)

🤖 Generated with Claude Code