OverviewHistoryStatsSecurity
npx skills add ...
Documentation
SKILL.md
npx skills add nvidia/skills --skill 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.
npx skills add nvidia/skills --skill nemo-mbridge-perf-moe-dispatcher-selection
Stable docs: @docs/training/moe-optimization.md Card: @skills/nemo-mbridge-perf-moe-dispatcher-selection/card.yaml
| Hardware | First choice | Why |
|---|---|---|
| H100 | DeepEP, if the runtime package is installed | Strong default for cross-node EP on Hopper |
| B200 | DeepEP, if the runtime package is installed | Good first choice unless a platform-specific HybridEP path is available |
| GB200 / GB300 NVL72 | HybridEP, if the runtime package is installed | Best fit for NVLink-domain-aware dispatch and lower memory pressure |
| Unknown or first bring-up | alltoall | Easiest path for correctness and debugging |
| EP size | Guidance |
|---|---|
| Small EP | Dispatcher choice is usually second-order; start with alltoall or DeepEP |
| Medium EP | DeepEP often becomes worthwhile |
| Large EP | HybridEP is usually the best target on NVL72 systems |
| Workload | Common best path | Notes |
|---|---|---|
| DSV3 at large scale | HybridEP on GB200 or GB300, DeepEP on H100 | Dispatcher choice matters more as EP and PP both grow |
| Qwen3 235B | DeepEP on H100, HybridEP on GB200 | HybridEP usually wins on GB200 and often uses less memory |
| Qwen3 30B | DeepEP | Smaller models still benefit, but the absolute gap is smaller |
| Qwen3-Next | Close race in BF16, HybridEP stronger in FP8 or memory-tight runs | Good reminder to test, not assume |
| MoE VLMs | Start simple, then test HybridEP on GB200-class systems | Vision workloads are sensitive to both memory and host overhead |
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.
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.
The broad trend is more important than any single row in the tracker:
alltoall is usually the conservative baselineIn 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.
For Qwen3 235B, the practical ordering is usually:
alltoall for initial bring-upHybridEP is usually modestly faster than alltoall on this workload and often
has noticeably better memory headroom.
This family is a good reminder that dispatcher wins are workload-dependent:
alltoall and HybridEP can be closeDeepEP is selected by setting
moe_token_dispatcher_type="flex" and moe_flex_dispatcher_backend="deepep".
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 is selected by setting
moe_token_dispatcher_type="flex" and moe_flex_dispatcher_backend="hybridep".
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.
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.
| Feature | Interaction |
|---|---|
| CUDA graphs | Best paired with attn moe_router moe_preprocess on dropless MoE |
| EP overlap | Helps when dispatcher time is still visible after backend tuning |
| FP8 | Often increases the relative importance of communication and host overhead |
| CPU affinity | Can matter as much as dispatcher choice on GB200 or GB300 |
| Pipeline layout | Poor PP or VPP layout can erase dispatcher gains |
alltoallDo not compare dispatchers on different stacks: container, routing mode, PP layout, and CUDA-graph scope can move the result as much as the dispatcher.
HybridEP is topology-sensitive: it is not a universal win outside the hardware it was designed for.
Both dispatchers need SM tuning: default moe_deepep_num_sms (20) and
moe_hybridep_num_sms (16) are reasonable starting points but rarely optimal.
Force-balance and dropless are not interchangeable baselines: keep the routing mode fixed when comparing dispatcher backends.
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.
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.