Efficient MoE Training for Biological Foundation Models

As language models grow, scaling dense architectures becomes increasingly expensive. In a dense transformer, every token passes through every layer, so adding capabilities increases computation for both training and inference.

Mixture-of-experts (MoE) architectures take a different approach to scaling by using many subnetworks, or experts, while activating only a small subset for each token.

This tradeoff has made MoE architectures increasingly attractive to the large language model (LLM) community. They can scale model capacity more efficiently, but the benefits depend heavily on implementation. Fragmented expert computation can reduce GPU utilization. Routing adds communication overhead, and larger parameter footprints create memory and distributed-training challenges. NVIDIA Transformer Engine (TE) helps address these bottlenecks with optimized primitives for grouped expert computation, kernel fusion, and low-precision training. As biological foundation models grow in parameter count and sequence length, these primitives can improve GPU efficiency while expanding model capacity.

This tutorial shows how to put these techniques into practice with the NVIDIA BioNeMo MoE recipe and TE. You’ll see how GroupedLinear improves expert computation, MXFP8 reduces memory use, and the GroupedMLP kernel fuses quantization, SwiGLU, and routing-weight scaling. Together, these capabilities provide a practical reference for efficiently training MoE-based biological foundation models.

Prerequisites #

Before you begin, you need:

• Familiarity with Python, PyTorch, and distributed training concepts
• An NVIDIA CUDA-enabled environment—you can use the linked Dockerfile or install the recipe requirements
• At least two GPUs for expert parallelism; NVIDIA Blackwell GPUs are required to use the fused MXFP8 GroupedMLP kernel

Challenge 1: Fragmented expert kernels #

MoE models replace a single dense feed-forward block with multiple expert networks. A naive implementation, however, can trigger excessive kernel launches. For example, the Hugging Face baseline implementation iterates over all experts in a Python loop, with each expert triggering separate kernel launches.

for expert_idx, expert_layer in enumerate(self.experts):
    idx, top_x = torch.where(expert_mask[expert_idx])
    current_state = hidden_states[None, top_x].reshape(-1, hidden_dim)
    current_hidden = expert_layer(current_state) * routing_weights[top_x, idx, None]
    final_hidden_states.index_add_(0, top_x, current_hidden)

Grouped execution preserves the individual expert matrices but submits their work together. TE’s GroupedLinear applies multiple linear transformations in one call by gathering the expert weights and input tokens. Because each expert can receive a different number of tokens, GroupedLinear accepts per-expert token counts (split_sizes). It submits the local experts through the TE grouped GEMM path instead of launching one PyTorch Linear operation per expert, reducing launch and scheduling overhead.

Use GroupedLinear as follows. Each expert retains its own weight tensor (weight0, weight1, and so on), and the call accepts the per-expert token counts as an additional positional argument:

from transformer_engine.pytorch.ops import GroupedLinear

experts_gate_up = GroupedLinear(
    num_groups=num_local_experts,
    in_features=hidden_size,
    out_features=2 * intermediate_size,
    bias=False,
    dtype=torch.bfloat16,
    device="cuda",
)

gate_up_output = experts_gate_up(tokens, split_sizes)

Compared with the Python loop, this approach submits the gate-up projections as one grouped operation instead of multiple separate calls.

Hugging Face Transformers also provides grouped_mm. However, TE can fuse GroupedLinear with MXFP8 quantization, activation, routing-weight scaling, and intermediate data movement into a GroupedMLP kernel, as shown in the later sections.

Challenge 2: Large model size and activation memory #

MoE architectures increase total parameter capacity, and genomics workloads often use long sequences, which puts pressure on activation memory during training. BF16 uses 16 bits to represent each model weight and activation.

The BioNeMo recipe uses TE to support FP8 and MXFP8 training, reducing memory use. Both formats represent weight and activation values with 8 bits instead of 16. The main difference between FP8 and MXFP8 is scaling granularity: MXFP8 assigns a scaling factor to each block of 32 consecutive values, helping preserve numerical range and accuracy. On NVIDIA Blackwell GPUs, MXFP8 is hardware-accelerated, enabling MXFP8 GEMMs to use specialized Tensor Core instructions. For details about MXFP8 and block scaling, see the Transformer Engine FP8 primer.

Challenge 3: Quantization overhead in low-precision training #

Although most training computation uses 8-bit precision, the model retains its master weights in 16 bits. The training framework therefore adds quantization and dequantization steps to convert between formats. Quantization converts BF16 weights and activations to MXFP8 before the low-precision GEMM; dequantization converts the result back to the higher-precision format. A naive path performs these steps as separate operations, motivating the fused MLP path described next.

fp8_recipe = te_recipe.MXFP8BlockScaling()
model = TEMixtralMXFP8ForCausalLM(config, fp8_recipe=fp8_recipe, dispatcher=dispatcher)

The TE autocast API enables MXFP8 precision for the model’s forward and backward passes:

with te.autocast(enabled=True, recipe=self._fp8_recipe):
    for decoder_layer in self.layers:
        hidden_states = decoder_layer(hidden_states)

See the BioNeMo recipe for the complete code.

To use the fused MLP, import the Transformer Engine Sequential API to chain together gate_up, ScaledSwiGLU, and down. The API also folds dequantization into the fused path. ScaledSwiGLU combines the routing probabilities (“scales”) with the expert feed-forward network computations.

from transformer_engine.pytorch.ops import GroupedLinear, ScaledSwiGLU, Sequential

experts_ffn = Sequential(GroupedLinear(gate_up), ScaledSwiGLU(), GroupedLinear(down))

The TE Sequential API scans the operations and, when the pattern matches, replaces the GroupedLinear → ScaledSwiGLU → GroupedLinear sequence with a fused operation object: ForwardGroupedMLP_CuTeGEMMSwiGLU_MXFP8 for the forward pass and a matching fused backward operation. This reduces framework overhead, fuses the SwiGLU and probability-scaling work into the grouped MLP path, and avoids materializing some intermediate results.

Results #

These are several of the optimizations in the BioNeMo recipe. In our training benchmark on eight NVIDIA B200 Tensor Core GPUs, the recipe delivered up to 2.21x the throughput of the Hugging Face baseline.

Run the recipe #

Start with the two-GPU L0_sanity configuration to confirm that expert parallelism and the training environment work correctly:

torchrun --nproc_per_node=2 train_fsdp2_ep.py --config-name L0_sanity

After validation, scale to the Mixtral-8x7B configuration with expert parallelism (EP=8) and MXFP8 precision across eight GPUs:

torchrun --nproc_per_node=8 train_fsdp2_ep.py --config-name L1_8x7B_ep checkpoint.ckpt_dir=/path/to/ckpt

Select BF16 or MXFP8 based on your GPU and memory requirements and set the data-parallel and expert-parallel sizes so their product equals the total GPU count. The recipe README includes launch, checkpoint, and benchmark commands.

Try the Mixtral Native Transformer Engine recipe in BioNeMo Recipes and learn more about the optimized MoE kernels in the NVIDIA Transformer Engine documentation.

Acknowledgments #

Sudhakar Singh US, Varun Thumbe US, Santosh Santosh US, Timur Rvachov US, Chris Hoge US,