Nemo Mbridge Perf Moe Dispatcher Selection

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

Hardware narrows the candidate set; it does not select the winner. Hold the model, routing, batch shape, parallelism, overlap, graph scope, container, and timing window fixed during the comparison.

By EP degree

On one NVL8 domain in BF16, treat alltoall and HybridEP as matched candidates: their throughput can be close once the full stack is held fixed. HybridEP is a high-priority tuning path, not a reason to skip the correctness baseline.

Model-Family Patterns

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

The current canonical 16×H100 BF16 performance recipe uses HybridEP, 32 HybridEP SMs, 64-token combine chunks, plain expert-parallel communication overlap, delayed weight-gradient compute disabled, and TE graphs over moe_router and moe_preprocess. Its verified 50-step run averaged 20.14729 s and 299.352 model TFLOPS/GPU over steps 41–50. This proves HybridEP can win on NVL8 H100; it does not prove HybridEP is universal.

An earlier matched overlap A/B on the same broad shape isolated a rise from 244.039 to 287.305 TFLOPS/GPU. Keep that causal result separate from the later multi-knob canonical winner.

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

Forced load balancing is a benchmark-only control that can reduce routing variance across dispatcher backends. It changes routing semantics, so keep it fixed within the dispatcher A/B and do not use it to accept a training-equivalent or convergence-sensitive result. Validate the production winner again with natural routing.

Key Interactions

When To Use Each

alltoall

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

DeepEP

• any supported target runtime where DeepEP imports successfully
• cross-node EP is clearly visible in profiles
• a matched steady-state A/B beats the alternatives

HybridEP

• NVL72 systems, where the topology makes it a high-priority candidate
• NVL8 systems when the package supports the topology and a matched A/B wins
• large EP degrees
• memory headroom matters in addition to throughput
• an NVL8 BF16 matched A/B beats or materially improves headroom over alltoall; a small or negative delta is a valid reason to keep alltoall

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: configure the actual NVLink domain and do not infer support or performance from the GPU SKU alone.

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.

7. Forced routing is not training equivalence: use it only as a disclosed benchmark control, then validate natural routing separately.

8. Config selection is not backend proof: require runtime evidence that the requested flex backend initialized and completed steady iterations.