CPU Offloading

References

• Stable docs: @docs/training/cpu-offloading.md
• Structured metadata: @skills/nemo-mbridge-perf-cpu-offloading/card.yaml

What It Is

Two independent mechanisms to move data from GPU to CPU memory:

Quick Decision

Enablement

Optimizer CPU offloading (recommended for large models)

cfg.optimizer.optimizer_cpu_offload = True
cfg.optimizer.optimizer_offload_fraction = 1.0
cfg.optimizer.overlap_cpu_optimizer_d2h_h2d = True

CLI overrides:

optimizer.optimizer_cpu_offload=True \
optimizer.optimizer_offload_fraction=0.5 \
optimizer.overlap_cpu_optimizer_d2h_h2d=True

Activation CPU offloading (small/medium models only)

cfg.model.cpu_offloading = True
cfg.model.cpu_offloading_num_layers = 16
cfg.model.cpu_offloading_activations = True
cfg.model.cpu_offloading_weights = False

cfg.model.pipeline_model_parallel_size = 1
cfg.model.recompute_granularity = None
cfg.model.cuda_graph_impl = "none"

Config Parameter Reference

Optimizer offloading

Activation offloading

Compatibility And Constraints

Activation offloading

• pipeline_model_parallel_size must be 1
• recompute_granularity must be None
• Cannot combine with fine_grained_activation_offloading
• Cannot combine with CUDA graphs
• cpu_offloading_num_layers must be in [0, num_layers-1)

Optimizer offloading

• Requires use_distributed_optimizer = True (default in most recipes)
• No PP, recompute, or CUDA graph restrictions
• optimizer_offload_fraction must be in [0.0, 1.0]

Practical: large MoE models

Activation offloading is blocked for Qwen3-30B-A3B and similar large MoE models. The PP=1 constraint means each GPU holds all 48 layers; model weights + optimizer states alone (~70 GB) exceed H100 80 GB capacity.

Minimal Runnable Command

uv run python scripts/training/run_recipe.py \
  --recipe qwen3_30b_a3b_pretrain_config \
  optimizer.optimizer_cpu_offload=True \
  optimizer.optimizer_offload_fraction=0.5 \
  train.train_iters=20 \
  train.global_batch_size=8 \
  train.micro_batch_size=1

Verification

Unit tests

uv run python -m pytest \
  tests/unit_tests/models/test_gpt_full_te_layer_autocast_spec.py -k "cpu_offload" \
  tests/unit_tests/peft/test_utils.py -k "cpu_offload" -q

Success criteria

• Config validation passes for the selected offloading mode
• Training completes without OOM or NCCL errors
• Loss matches the non-offloaded baseline (max delta < 0.001)
• Memory usage drops proportionally to offload fraction

Code Anchors

MCore activation offload constraints

        if self.cpu_offloading and (
            self.cpu_offloading_num_layers < 0 or self.cpu_offloading_num_layers >= self.num_layers
        ):
            raise ValueError(...)

        if self.cpu_offloading and self.pipeline_model_parallel_size > 1:
            raise ValueError(
                "Currently there is no support for Pipeline parallelism with CPU offloading"
            )

        if self.cpu_offloading and self.recompute_granularity is not None:
            raise ValueError(
                "CPU offloading does not work when activation recomputation is enabled"
            )

MCore CUDA graph incompatibility

            if self.cpu_offloading:
                raise ValueError("CUDA graphs not supported with CPU offloading.")

MCore fine-grained offloading mutual exclusion

        if self.fine_grained_activation_offloading:
            assert (
                not self.cpu_offloading
            ), "fine_grained_activation_offloading cannot be enabled with cpu_offloading."

MCore HybridDeviceOptimizer instantiation

        if config.optimizer_cpu_offload:
            # ... setup cpu/gpu optimizer classes ...
            optimizer = HybridDeviceOptimizer(
                param_groups,
                offload_fraction=config.optimizer_offload_fraction,
                cpu_optimizer_cls=cpu_optimizer_cls,
                gpu_optimizer_cls=gpu_optimizer_cls,
                overlap_cpu_optimizer_d2h_h2d=config.overlap_cpu_optimizer_d2h_h2d,
                pin_cpu_grads=config.pin_cpu_grads,
                pin_cpu_params=config.pin_cpu_params,
            )

Bridge CUDA graph guard

        assert not config.cpu_offloading and config.recompute_granularity is None, "Cudagraphs not supported"

Bridge activation offloading in PEFT

        if self.config.cpu_offloading and self.config.cpu_offloading_activations:
            x.activation_offloading = True
        x, _ = self.linear_in(x)
        x = self.activation(x)
        if self.config.cpu_offloading and self.config.cpu_offloading_activations:
            x.activation_offloading = True
        x, _ = self.linear_out(x)

Failure Diagnosis

Known Limitations

• Activation offloading requires PP=1, making it impractical for large models (30B+ MoE) that need pipeline parallelism.
• Optimizer offloading throughput penalty scales linearly (~1.9x at 25%, ~4.2x at 100% for Qwen3-30B-A3B).
• D2H/H2D overlap provides only ~7% speedup because CPU Adam compute is the dominant bottleneck.
• fine_grained_activation_offloading is a separate module-level approach that works with PP > 1 but cannot be combined with layer-level cpu_offloading.