nemo-mbridge-perf-moe-hardware-configs

oleh nvidia

Buku pedoman pelatihan MoE yang representatif berdasarkan platform perangkat keras dan keluarga model. Merangkum pita throughput yang dibulatkan, pola paralelisme, dan penyetelan umum…

npx skills add https://github.com/nvidia/skills --skill nemo-mbridge-perf-moe-hardware-configs

Unduh ZIP GitHub

3.1k

Coba skill iniDisponsori

MoE Hardware Configuration Reference

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

Quick Platform Playbook

These rows are search seeds, not hardware defaults or throughput promises.

Platform	Candidates to screen after `alltoall` bring-up	What usually matters most
H100	DeepEP or HybridEP, explicit overlap, supported FP8 modes	communication overlap, dispatcher/runtime compatibility, and PP efficiency
B200	DeepEP or HybridEP, supported FP8 modes, careful PP layout	container quality and tuned communication settings
GB200	HybridEP, then profile-driven graphs and CPU cleanup	host overhead, topology-aware dispatch, memory headroom
GB300	HybridEP and the target container's lower-precision/kernel stack	the same system interactions as GB200, with remeasurement required

First Answer Checklist

For hardware playbook questions, answer from these canonical rows before adding throughput caveats:

Workload	Hardware	Dispatcher	Layout
DSV3	H100	DeepEP	TP=2, EP=64, PP=8, VPP=4
DSV3	GB200/GB300	HybridEP	TP=1, EP=64, PP=4, VPP=4
Qwen3 235B	H100	`alltoall` + overlap in the current canonical recipe	TP=2, EP=32, PP=8, VPP=4
Qwen3 235B	GB200	HybridEP	TP=1 or 2, EP=32-64, PP=4, VPP=unspecified
Qwen3 30B	16×H100	HybridEP	TP=1, EP=16, PP=1, plain EP overlap

For Qwen3 235B on GB200, explicitly say VPP=unspecified; do not invent or extrapolate VPP=12 unless a measured row provides it. Treat TE-scoped CUDA graph scopes (attn, moe_router, moe_preprocess) as profile-driven candidates, CUDA_DEVICE_MAX_CONNECTIONS selection, PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True, NCCL_GRAPH_REGISTER=0, GB200/GB300 CPU-side tuning, and the warning not to cargo-cult tracker rows.

Rounded Performance Bands

These are intentionally rounded so the document stays durable as the tracker moves. Treat them as planning ranges, not exact promises.

Workload family	Hardware	Typical band	Representative shape
DSV3, large-scale	H100	low-to-mid hundreds TFLOPS/GPU, high-teens MFU	TP2, EP64, PP8, DeepEP
DSV3, large-scale	B200	high-hundreds TFLOPS/GPU, mid-teens MFU	TP1, EP32, PP8, DeepEP
DSV3, large-scale	GB200	around 1K TFLOPS/GPU, low-20s MFU	TP1, EP64, PP4, HybridEP
DSV3, large-scale	GB300	above the GB200 band, often mid-20s MFU	TP1, EP64, PP4, HybridEP
Qwen3 235B	H100	historical low-300s snapshots; remeasure the current recipe	TP2, EP32, PP8; current recipe uses `alltoall` + overlap
Qwen3 235B	GB200	high-hundreds TFLOPS/GPU in tuned runs	TP1 or TP2, EP32-64, PP4, HybridEP
Qwen3 30B	H100	about 300 TFLOPS/GPU on the validated 16-GPU shape	TP1, EP16, PP1, HybridEP + EP overlap
Qwen3-Next 80B	GB200	low-300s TFLOPS/GPU in BF16-class runs	TP1, EP32, PP2, HybridEP

Representative Config Families

DSV3 on H100

Dispatcher: DeepEP
TP=2  EP=64  PP=8  VPP=4
Routing: force balance
Recompute: light-to-moderate selective recompute
Priority: overlap communication and keep PP efficient

DSV3 on B200

Dispatcher: DeepEP
TP=1  EP=32  PP=8  VPP=2 or similar
Precision: MXFP8-class
Recompute: selective recompute around MLA up-projection and MLP-side modules
Priority: container quality, PP layout, and DeepEP SMS tuning

DSV3 on GB200 or GB300

Dispatcher: HybridEP
TP=1  EP=64  PP=4  VPP=4
Precision: MXFP8-class
CUDA Graph: attn + moe_router + moe_preprocess
Priority: HybridEP, CPU optimization, and graph-friendly static shapes

Qwen3 235B on H100

Dispatcher: alltoall in the current canonical recipe; re-screen flex backends on the target stack
TP=2  EP=32  PP=8  VPP=4
Recompute: none in the current canonical recipe
Priority: communication overlap and router-path cleanup

Qwen3 235B on GB200

Dispatcher: HybridEP
TP=1 or 2  EP=32 to 64  PP=4  VPP=unspecified unless measured
CUDA Graph: attn + moe_router + moe_preprocess
Recompute: moe_act, mlp, or norm depending on memory pressure
Priority: balance throughput against memory headroom

Qwen3 30B-A3B on 16 H100

Dispatcher: HybridEP
TP=1  EP=16  PP=1  CP=1
Precision: BF16
Sequence: 4096
Batch: MBS1 GBS1024
Routing: force balance
EP overlap: enabled
Delayed wgrad: disabled
CUDA Graph: moe_router + moe_preprocess
HybridEP: permute fusion, 32 SMs, 64-token combine chunks
Measured: 20.14729s/step, 299.352 model TFLOPS/GPU over iterations 41-50
Rank-0 peak allocated memory: 62.166 GiB

The current number is the final multi-knob canonical recipe result. An earlier matched A/B isolated plain EP overlap: 244.039 to 287.305 TFLOPS/GPU, with communication hidden by GEMM/attention increasing from 0.11% to 36.55%. Do not attribute the later 299.352 result entirely to overlap.

Qwen3-Next 80B on GB200

Dispatcher: HybridEP
TP=1  EP=32  PP=2  VPP around 4
CUDA Graph: attn + moe_router + moe_preprocess
Priority: pipeline layout and grouped GEMM quality

Cross-Cutting Patterns

PP layout

E = embedding
t = transformer
m = MTP
L = loss
| = stage boundary

The biggest platform difference is usually not just the dispatcher. It is the combination of dispatcher, PP shape, and whether VPP keeps each stage balanced.

Recompute strategy

Memory pressure	Starting point
low	none or a very narrow selective set
moderate	`moe_act`, `mlp`, `norm`, or similar selective modules
high	model-specific up-projection plus selective MoE and MLP modules
extreme or long-context	full recompute only if the selective path still does not fit

Environment variables

CUDA_DEVICE_MAX_CONNECTIONS=1
CUDA_DEVICE_MAX_CONNECTIONS=32   # common when EP overlap and CUDA graphs are combined
PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True
NCCL_GRAPH_REGISTER=0

CPU-side tuning

On GB200 and GB300, CPU affinity and general host-overhead cleanup can move the needle almost as much as a dispatcher swap. Treat them as first-class tuning work, not as afterthoughts.

Pitfalls

Do not cargo-cult a tracker row: the winning config usually depends on routing mode, container, and PP layout as much as on hardware name.
Container quality matters: large regressions can come from the software stack rather than the model recipe.
VPP must be intentional: a bad VPP split can erase the gain from a better dispatcher.
Compare absolute throughput, not only MFU: MFU can mislead when switching between BF16, FP8, and other precision modes.
Force-balance routing is benchmark-only: it can control routing variance, but it changes semantics. Keep routing fixed within an A/B and validate natural routing separately for training acceptance.
Do not treat the dispatcher table as a hard platform rule: HybridEP is the validated winner for the canonical 16×H100 Qwen3 30B shape, while the current 256×H100 Qwen3 235B recipe uses alltoall. Benchmark backend compatibility and throughput in the production container.
Separate screening, causality, and acceptance: short runs reject weak candidates, matched one-variable A/Bs explain a mechanism, and a 50-step final run validates the complete winner.