Nemo Mbridge Perf Moe Dispatcher Selection

Choose the right MoE token dispatcher (`alltoall`, DeepEP, or HybridEP) for the hardware, EP degree, and optimization stage. Summarizes patterns from DSV3, Qwen3, Qwen3-Next, and VLM bring-up work.

Published by @NVIDIA·0 agent reads / 30d·0 saves·

MoE Dispatcher Selection Guide

Stable docs: @docs/training/moe-optimization.md Card: @skills/nemo-mbridge-perf-moe-dispatcher-selection/card.yaml

Quick Decision

By hardware

HardwareFirst choiceWhy
H100DeepEP, if the runtime package is installedStrong default for cross-node EP on Hopper
B200DeepEP, if the runtime package is installedGood first choice unless a platform-specific HybridEP path is available
GB200 / GB300 NVL72HybridEP, if the runtime package is installedBest fit for NVLink-domain-aware dispatch and lower memory pressure
Unknown or first bring-upalltoallEasiest path for correctness and debugging

By EP degree

EP sizeGuidance
Small EPDispatcher choice is usually second-order; start with alltoall or DeepEP
Medium EPDeepEP often becomes worthwhile
Large EPHybridEP is usually the best target on NVL72 systems

Model-Family Patterns

WorkloadCommon best pathNotes
DSV3 at large scaleHybridEP on GB200 or GB300, DeepEP on H100Dispatcher choice matters more as EP and PP both grow
Qwen3 235BDeepEP on H100, HybridEP on GB200HybridEP usually wins on GB200 and often uses less memory
Qwen3 30BDeepEPSmaller models still benefit, but the absolute gap is smaller
Qwen3-NextClose race in BF16, HybridEP stronger in FP8 or memory-tight runsGood reminder to test, not assume
MoE VLMsStart simple, then test HybridEP on GB200-class systemsVision workloads are sensitive to both memory and host overhead

Rounded Evidence Summary

Backend availability gate

Do not interpret a dispatcher timing until the container has proven that the selected backend package is available. --moe_flex_dispatcher_backend None selects the standard alltoall dispatcher, while deepep and hybridep select moe_token_dispatcher_type="flex" and then require their corresponding runtime packages at model construction time. If DeepEP or HybridEP is missing, record the import failure as an environment limitation and treat alltoall as the only measured correctness fallback for that run.

Qwen3 30B A3B on H100

A short 2026-05-17 H100 smoke run used Qwen3 30B A3B BF16, 16 GPUs, EP=16, the recipe's Transformer Engine CUDA graph scopes (moe_router, moe_preprocess), and model.moe_permute_fusion=false due to a Triton JIT compatibility issue in the run container. The alltoall fallback completed five steps with 45.65 s mean step time after warmup, 132.9 mean TFLOP/s/GPU after warmup, final loss 11.44050, and 61.351 GB peak max allocated memory. DeepEP and HybridEP selected the requested flex backend in the dumped configs but failed before the first iteration because the packages were not installed. This confirms the availability gate; it is not a throughput ranking for flex dispatchers on H100.

DSV3 on GB200 or GB300

The broad trend is more important than any single row in the tracker:

  • plain alltoall is usually the conservative baseline
  • DeepEP improves that baseline once EP communication becomes visible
  • HybridEP adds another step up on NVL72 systems, especially after CUDA graphs, routing improvements, and CPU-side cleanup are already in place

In practice, the stack often moves from roughly "low-teens MFU" territory with an untuned baseline into "high-teens to low-20s MFU" territory after the full dispatcher and kernel stack is tuned.

Qwen3 235B on GB200

For Qwen3 235B, the practical ordering is usually:

  1. alltoall for initial bring-up
  2. DeepEP if you want a familiar tuned path
  3. HybridEP for the strongest steady-state result on GB200

HybridEP is usually modestly faster than alltoall on this workload and often has noticeably better memory headroom.

Qwen3-Next on GB200

This family is a good reminder that dispatcher wins are workload-dependent:

  • in BF16, alltoall and HybridEP can be close
  • in FP8 or memory-constrained settings, HybridEP tends to look better
  • pipeline layout and grouped-GEMM changes can matter almost as much as the dispatcher itself

Tuning Parameters

DeepEP

DeepEP is selected by setting moe_token_dispatcher_type="flex" and moe_flex_dispatcher_backend="deepep".

--moe-deepep-num-sms 20

Tune the SM count allocated to DeepEP communication kernels (default 20). The optimal value depends on the workload and EP degree. First confirm the DeepEP package imports in the target container; a missing package fails during model construction, before any dispatcher timing is available.

HybridEP

HybridEP is selected by setting moe_token_dispatcher_type="flex" and moe_flex_dispatcher_backend="hybridep".

--moe-hybridep-num-sms 16

Tune the SM count allocated to HybridEP communication (default 16). The performance harness uses 32 for HybridEP workloads. Sweep between 16 and 32 for the target hardware. Set NUM_OF_HYBRID_EP_RANKS_PER_NVLINK_DOMAIN to match the NVLink domain size of the deployment. If it does not match the actual topology, performance and sometimes correctness will suffer. First confirm the HybridEP package imports in the target container; a missing package fails during model construction, before any dispatcher timing is available.

Routing mode

--moe-router-force-load-balancing

For performance benchmarking, force-balance routing is the safer default. It usually outperforms dropless routing in large-scale benchmarks and makes results more comparable across dispatcher backends.

Key Interactions

FeatureInteraction
CUDA graphsBest paired with attn moe_router moe_preprocess on dropless MoE
EP overlapHelps when dispatcher time is still visible after backend tuning
FP8Often increases the relative importance of communication and host overhead
CPU affinityCan matter as much as dispatcher choice on GB200 or GB300
Pipeline layoutPoor PP or VPP layout can erase dispatcher gains

When To Use Each

alltoall

  • first correctness bring-up
  • small EP configurations
  • debugging communication regressions

DeepEP

  • Hopper or B200 deployments
  • cross-node EP is clearly visible in profiles
  • you want a mature intermediate step before testing HybridEP

HybridEP

  • GB200 or GB300 NVL72 systems
  • large EP degrees
  • memory headroom matters in addition to throughput

Pitfalls

  1. Do not compare dispatchers on different stacks: container, routing mode, PP layout, and CUDA-graph scope can move the result as much as the dispatcher.

  2. HybridEP is topology-sensitive: it is not a universal win outside the hardware it was designed for.

  3. Both dispatchers need SM tuning: default moe_deepep_num_sms (20) and moe_hybridep_num_sms (16) are reasonable starting points but rarely optimal.

  4. Force-balance and dropless are not interchangeable baselines: keep the routing mode fixed when comparing dispatcher backends.

  5. Memory and throughput can trade off differently by model: Qwen3-style runs may show a smaller speed delta than DSV3, but still justify HybridEP for memory headroom.

  6. Backend import failures are not performance data: if DeepEP or HybridEP is missing from the container, do not compare its failed job against a completed alltoall job. Fix the environment first, then rerun the same stack.

Bundled with this artifact

5 files

Reference files that ship alongside this artifact. Agents pull these in only when the task needs them.

More on the bench

SKILL0

Tensorflow And Deep Learning Rules

TensorFlow and deep learning rules for building, training, evaluating, and deploying neural network models

data-science-ml+1
0
SKILL0

Fortran Programming Guidelines

Modern Fortran rules for scientific computing, modules, explicit interfaces, kind parameters, memory safety, and testing

software-engineering+1
0
SKILL0

Automl And Hyperparameter Optimization Rules

AutoML and hyperparameter optimization rules for Python ML projects using Ray Tune, Optuna, PyCaret, and time-series AutoML libraries

data-science-ml+1
0