nemo-mbridge-perf-moe-long-context

MoE Long-Context Training

What Changes At Long Context

Once sequence length moves well past the 4K-class regime, attention memory and activation residency become the dominant constraints. For MoE models, that usually means you need some combination of:

• context parallelism
• selective recompute
• lower precision
• CPU offload for optimizer state
• a dispatcher and PP layout that do not waste the smaller remaining DP budget

Rounded Scaling Patterns

DSV3 on H100

The DSV3 long-context runs show a stable pattern:

• selective recompute works better than full recompute once you move past the shortest contexts
• throughput stays in a fairly narrow band from mid-length through very long contexts if CP is increased appropriately
• the trade shifts from "memory fit" to "GPU-count feasibility" as CP grows

In other words, long context does not immediately collapse utilization if the layout is chosen well, but it does consume the DP budget very quickly.

Qwen3-Next on GB200

Qwen3-Next behaves more like a memory-sensitive medium-scale model:

• 8K and 32K remain practical with moderate CP
• 64K is possible, but the throughput drop is noticeable and memory becomes much tighter
• pipeline layout and grouped-GEMM improvements matter almost as much as CP

Qwen3 235B on GB200

Qwen3 235B shows that long context can still be efficient on NVL72 systems when TP, CP, and HybridEP are coordinated. The best 128K-class configurations are not just "fit-only" recipes; they can remain highly efficient if routing, parallelism, and recompute are balanced.

CP Sizing Rules Of Thumb

1. Start from a 4K shard target: a good first guess is CP ~= seq_len / 4096, then round to a practical power-of-two layout.

2. Keep DP alive if possible: long-context scaling becomes brittle once CP, EP, TP, and PP together squeeze DP down to the floor.

3. Prefer selective recompute: recompute modules such as up_proj, norm, moe, moe_act, or mlp before reaching for full recompute.

4. Avoid SDPA-heavy recompute at very long context: recomputing attention internals can add a lot of work for less memory benefit than recomputing smaller MoE and MLP-side modules.

5. Use TP as another lever on NVL72 systems: GB200 and GB300 runs can sometimes trade some CP for TP while still staying efficient.

6. Assume GBS will need to shrink: as CP rises and DP falls, you may need to reduce global batch size or accept higher GA.

Representative Config Families

DSV3 at 128K on H100

TP=1  CP=32  EP=32  PP=8  VPP=4
Precision: FP8-class
Dispatcher: DeepEP
Recompute: up_proj, norm, moe, mlp
Extra memory help: optimizer CPU offload

DSV3 at 256K on H100

TP=1  CP=64  EP=32  PP=8  EDP=2  VPP=4
Precision: FP8-class
Dispatcher: DeepEP
Recompute: up_proj, norm, moe, mlp
Extra memory help: optimizer CPU offload

Qwen3 235B at 128K on GB200

TP=4  CP=4  EP=32  PP=4  VPP=12
Precision: BF16 or MXFP8
Dispatcher: HybridEP
Recompute: moe_act, norm
CUDA Graph: attn + moe_router + moe_preprocess

Recompute And CUDA Graph Guidance

For long-context MoE training:

• start with selective recompute
• add CUDA graphs only after the shapes and routing path are stable
• keep sequence length and MBS fixed when using CUDA graphs
• if the run depends on highly dynamic batches, prefer eager execution

Pitfalls

1. CP does not replace EP or PP: it adds another dimension; it does not make the others disappear.

2. A good 4K baseline can still be a bad long-context baseline: routing mode, recompute choice, and offload strategy often need to change.

3. GPU-count feasibility becomes the real constraint: very long context can look fine in a single recipe, then become impossible once EP and PP are added honestly across the full model.

4. CUDA graphs need static shapes: variable-length batches and opportunistic padding strategies can silently break the path.

5. Container and kernel support matters more at 128K+: long-context paths tend to rely on newer kernels and bug fixes than short-context bring-up does.