![\"MoE](\"https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/blog/moe-transformers/moe_routing.png\")
Over the past few years, scaling dense language models has driven most progress in LLMs. From early models like the original ULMFiT (~30M parameters) or GPT-2 (1.5B parameters, which at the time was considered "too dangerous to release" 🧌), and eventually to today’s hundred-billion–parameter systems, the recipe was simple:
\n\n\nMore data + more parameters gives better performance.
\n
Scaling laws reinforced this trend, but dense scaling has practical limits:
\nThis is where Mixture of Experts (MoEs) enter the picture.
\n\n\nIf you're already familiar with MoEs and want to jump straight into the engineering work done in transformers, you can head directly to Transformers and MoEs.
\n
A Mixture of Experts model keeps the Transformer backbone, but replaces certain dense feed-forward layers with a set of experts. An “expert” is not a topic-specialized module (e.g., "math expert", "code expert"). It is simply a learnable sub-network. For each token, a router selects a small subset of experts to process it.
\n | \n
|---|
| Figure 1: Expert 1 among 4 experts is activated (Source: Maarten Grootendorst) | \n
Different tokens activate different experts, based on their hidden representations.
\n\n\nModel capacity depends on total parameters, but inference speed depends on active parameters.
\n
This is the key idea.
\nFor example, take gpt-oss-20b. It has 21B total parameters, but uses 4 active experts per token, out of a total of 32 experts. Considering the shared components plus the active experts, this model uses ~3.6B active parameters per token. Running this model on an M3 Ultra Mac, which has a memory bandwidth of about 800 GB, we could estimate generation speed as ~ 800 / (3.6 * 2) in bfloat16, where each parameter takes 2 bytes. This yields about 111 tokens per second. The actual performance number we get is ~115 tok/s, which is very close to the back-of-the-envelope calculation.
Your browser does not support the video tag.
\nThis super fast speed confirms the model works approximately as a 3.6B parameter one, but it has the same capacity (or quality) as a 21B parameter model.
\n(Note: speed would be even faster if we used kernels for the native mxfp4 quantization the model uses).
\nMoEs are attractive for these reasons:
\nBetter Compute Efficiency
\nGiven a fixed training FLOP budget, MoEs often outperform dense counterparts.
\n![\"MoE](\"https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/blog/moe-transformers/faster_training.png\") | \n
|---|
| Figure 2: Dense vs. MoE training curves (Source: OLMoE: Open Mixture-of-Experts Language Models) | \n
This means faster iteration and better scaling efficiency.
\nA Natural Parallelization Axis
\nExperts provide a structural boundary in the computation graph. Since different tokens engage different experts, we can parallelize across experts (we discuss this later in Expert Parallelism).
\nIndustry Adoption
\nRecent major MoE releases of open models that happened in the past few weeks include Qwen 3.5, MiniMax M2, GLM-5, or Kimi K2.5.
\nThe trend accelerated after the success of DeepSeek R1 in January 2025, building on earlier systems like DeepSeek V2. Another early MoE was Mixtral-8x7B, released in December 2023.
\n![\"2-year](\"https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/blog/moe-transformers/moe_2y_timeline.png\") | \n
|---|
Figure 3: 2-year timeline of MoE model addition to the transformers library. DeepSeek R1 marks a clear inflection point. | \n
Closed labs use MoEs too. ChatGPT has long been rumored to use a sparse architecture, and the open gpt-oss models certainly do.
\n\n\nIf you want to learn more about MoEs in general, we strongly suggest reading this blog and watching our recent YouTube video on routing.
\n
Most tooling in the ecosystem, including model loading, device placement, quantization, and backend execution was originally designed for dense models. MoEs challenge these assumptions.
\nMaking MoEs first-class citizens in transformers means redesigning parts of the loading pipeline, execution model, and distributed abstractions, not just adding new model classes. We’ll focus on how the transformers library has evolved to support sparse architectures across:
AutoModelForCausalLM.from_pretrained("model_id") downloads and loads model weights into a PyTorch model. For dense models, loading is relatively straightforward where each tensor in the checkpoint maps one-to-one to a parameter in the runtime module.
For MoEs, it’s more complicated. In most MoE checkpoints, each expert is serialized independently. If you peek inside the DeepSeek-V3 checkpoint index, you’ll see keys like:
\n\n\n\nEach expert has its own set of weight matrices, essentially 256 (0 to 255 total, taking DeepSeek-V3 as an example) small feed-forward networks saved side by side. At runtime, however, GPUs execute optimized kernels. Modern MoE kernels such as grouped GEMMs and fused MoE implementations are designed to process all experts in a single operation, not by looping over them one at a time.
\nTo do that efficiently, they require expert weights to be packed into a single contiguous tensor.
\nSo we have a mismatch:
\nBridging this gap systematically is what the weight loading refactor enables.
\nWith the introduction of a generic WeightConverter, the mental model shifted from:
\n\n\nA checkpoint already matches my runtime layout; loading is mostly a key-by-key copy.
\n
to:
\n\n\nA checkpoint is just a serialized source of tensors. Loading is a conversion pipeline that transforms them into the runtime layout we want.
\n
WeightConverterThe central abstraction introduced by this refactor is dynamic weight loading via a WeightConverter.
WeightConverter lets us define:
source key patterns → target key(s) + operations\nPrimitive operations (chunk, concatenate, etc.) are composable. Two that are particularly useful for MoEs:
\nMergeModulelist merges a list of tensors into a single tensor. For example, you can compose MergeModulelist with Concatenate to stack the experts in a MoE and pack them into one tensor.
\nSplitModulelist splits a tensor back into a list of tensors. For example, you can split a stack of experts back into individual experts.
\nThe refactor improves not just what conversions exist, but how they’re scheduled.
\nThe loader scans checkpoint keys once, matches them against converter patterns, and groups tensors per converter. Once a key is identified as needed, it’s registered as a future and materialized via a thread pool. Conversion operations run only once their dependencies are ready. For example, MergeModulelist waits until all experts for a layer are loaded.
This avoids repeated scans and reduces memory peaks.
\nTo evaluate the improvements introduced by the new weight-loading pipeline, we benchmarked the v4 vs v5 versions of transformers. The focus is on loading speed of large MoE models, which is often a bottleneck in training and inference.
We benchmarked v4 vs v5 using:
\nExample:
\n\n\n = \n = \nTwo relevant environment variables:
\nHF_ENABLE_PARALLEL_LOADING: Enables parallel shard loading via threads.
HF_DEACTIVATE_ASYNC_LOAD:Disables the new async pipeline (v5 escape hatch).
Model: Qwen/Qwen1.5-110B-Chat GPU: 1Ă— A100 (80GB)
| Version | \nStrategy | \nLoading Mode | \nTime | \n
|---|---|---|---|
| v4.57.6 | \ndevice_map="auto" | \nThreadpool | \n66.24s | \n
| v4.57.6 | \ndevice_map="auto" | \nSequential | \n67.29s | \n
| v4.57.6 | \nTP | \n— | \nOOM | \n
| v5 | \ndevice_map="auto" | \nAsync (default) | \n20.71s | \n
| v5 | \ndevice_map="auto" | \nSync | \n45.3s | \n
| v5 | \nTP | \nAsync | \n10.1s | \n
| v5 | \nTP | \nSync | \n19.28s | \n
![\"Loading](\"https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/blog/moe-transformers/loading_benchmark.png\") | \n
|---|
| Figure 4: Loading benchmarks (v4 vs v5) | \n
The speedup is not just “more threads.”
\nIt’s the combination of Single-pass routing, Async materialization, and Conversion-aware scheduling which together avoid unnecessary materialization and memory peaks while enabling expert packing and projection fusion at load time.
\nWith this refactor we can now create the runtime module structure first and then convert the weights into the structure. We can now optionally attach quantization within the conversion pipeline, making quantization part of the weight loading pipeline itself. This is crucial because quantizing “per expert” only makes sense once experts exist in a predictable packed layout.
\nThis end to end pipeline was not possible earlier and now it comes to the users as an exposed API.
\nOnce experts are packed into a single runtime tensor, another question arises:
\n\n\nHow do you actually route through them efficiently?
\n
In a Mixture of Experts model, each token is routed to different experts. This means the runtime must dispatch tokens to their selected expert weights, execute the projections efficiently, apply the routing weights and then collect and reorder the results.
\nThis is what the Experts Backend system (introduced in PR #42697) addresses. The Experts Backend introduces a pluggable execution architecture that decouples expert computation from the model implementation. Instead of hardcoding one dispatch strategy inside each MoE model, the system allows expert layers to dynamically select a backend at runtime.
\nThis is implemented via a decorator pattern:
\n\nThe decorator wraps expert classes and dispatches computation to the selected backend automatically.
\nThree backends are currently provided:
\neager which loops over the selected experts and applies projections per expert. This is used for correctness reference and debugging.
batched_mm uses the torch.bmm API. This duplicate selected expert weights per token and performs a single batched GEMM. This backend is very well suited for small batch, GPU-heavy workloads where memory is available.
grouped_mm uses torch._grouped_mm API. Here we sort tokens by expert ID, group them, and then perform a single grouped GEMM. This backend shines with large batches or memory-constrained setups.
![\"\"](\"https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/blog/moe-transformers/expert_backend.png\") | \n
|---|
| Figure: Expert backend illustration | \n
Mixture of Experts (MoE) models can have hundreds of billions of parameters (far more than what fits on a single GPU). Expert parallelism (EP) addresses this by distributing experts across multiple devices. Each device loads only its assigned subset of experts, computes for those experts and then participates in result aggregation. This approach scales models to far larger parameter counts without increasing computation cost because each token activates only a few experts.
\nExpert parallelism is enabled via enable_expert_parallel:
\n\n\n\n = \n\n = \nLaunch with:
\n\nWhere N evenly divides the total number of experts, and possibly matches the number of GPUs in your node.
When enable_expert_parallel=True, the model switches from the standard tensor-parallel (TP) plan to an expert-parallel (EP) plan with specialized sharding strategies.
Core components of EP lie in:
\nGroupedGemmParallel: This splits the expert weights along the expert dimension (dim=0). Here each device loads only num_experts / num_devices.
RouterParallel: This remaps global expert indices to local indices, masks out experts not assigned to the current rank, ensures each device computes only with its local experts and uses an all-reduce to combine partial outputs across devices.
MoEs are excellent for scaling inference, but training them is significantly more complex.
\nMoEs have a Massive parameter count, the distributed expert communication is complicated, there are routing in-stabilities that need to be handled. To address this, we collaborated with Unsloth to enable significantly faster Mixture-of-Experts training:
\nWe leverage the Expert Backend abstraction, standardize around PyTorch’s torch._grouped_mm API and use custom Triton grouped-GEMM + LoRA kernels. Unsloth builds on top of the Transformers (and TRL) optimizations to push performance further.
\n\nFor full details, we recommend reading: Unsloth’s official guide
\n
As sparse architectures continue to evolve, we want the transformers library to evolve with them. If you’re building with MoEs or experimenting with new sparse ideas, we’d love to hear from you. Let us know what abstractions, kernels, or workflows you’d like to see next in transformers.