Distributed inference

When a model doesn’t fit on a single GPU, distributed inference with tensor parallelism can help. Tensor parallelism shards a model onto multiple GPUs and parallelizes computations such as matrix multiplication. It enables fitting larger model sizes into memory and is faster because each GPU can process a tensor slice.

However, tensor parallelism adds communication overhead and should be used on single machine setups with multiple GPUs to take advantage of fast intra-node communication. For multi-node training, it may be more efficient to use pipeline or data parallelism depending on your use case.

Refer to the Ultra-Scale Playbook section on tensor parallelism to learn more.

Check the list below for models that natively support tensor parallelism. Open a GitHub issue or pull request to add support for a model.

Show supported models

This guide shows how to enable tensor parallelism with Transformers and different partitioning strategies.

Partitioning a model

Transformers supports tensor parallelism if a model has a tp_plan. There are two plans to partition a model.

The auto tensor parallelism plan partitions a model (see the supported models above) based on a predefined configuration.
You can also manually specify your own partitioning plan and pass it to the tp_plan parameter in from_pretrained().

auto plan

manual plan

Partitioning strategies

All partitioning strategies are defined in the ParallelInterface class which maps a string to the strategy implementation. You don’t need to interact with this class directly since all the strategies are set with tp_plan in from_pretrained(), but it is useful for checking what strategies are available.

class ParallelInterface(MutableMapping):
    """
    Dict-like object keeping track of allowed attention functions. You can easily add a new attention function
    with a call to `register()`. If a model needs to locally overwrite an existing attention function, say `sdpa`,
    it needs to declare a new instance of this class inside the `modeling_<model>.py`, and declare it on that instance.
    """
    _global_mapping = {
        "colwise": ColwiseParallel(),
        "rowwise": RowwiseParallel(),
        "colwise_rep": ColwiseParallel(output_layouts=Replicate()),
        "rowwise_rep": RowwiseParallel(input_layouts=Replicate()),
        "local_colwise": ColwiseParallel(use_dtensor=False),
        "local_rowwise": RowwiseParallel(use_dtensor=False),
        "local": IsolatedParallel(),
        "gather": GatherParallel(),
        "local_packed_rowwise": PackedRowwiseParallel(use_dtensor=False),
        "sequence_parallel": SequenceParallel(),
        "replicate": ReplicateParallel(),
    }

Refer to the table below to learn more about each strategy.

Strategy	Description
`ColwiseParallel`	Column-wise partitioning of weights and biases.
`RowwiseParallel`	Row-wise partitioning of weights and biases. Also supports partitioning `nn.Embedding` modules.
`SequenceParallel`	Sequence parallel implementation to support `LayerNorm` and `Dropout` layers. Also supports Python implementation of RMSNorm.
`PackedColwiseParallel`	Variant of `ColwiseParallel` to support packed weights (for example, packing `up_proj` and `gate_proj` together). Refer to the code for more details.
`PackedRowwiseParallel`	Variant of `RowwiseParallel` to support packed weights (refer to the code for more details).
`GatherParallel`	Gather outputs of the module across devices.
`IsolatedParallel`	Used for Experts in Mixture-of-Experts (MoE) layers to isolates module from other devices.
`ReplicateParallel`	Replicate modules across all devices to prevent `torch.distributed` APIs from breaking due to a partially sharded model.

Packed strategies

Weight packing packs multiple linear layers into a single, bigger layer. Packed strategies, PackedColwiseParallel and PackedRowwiseParallel, are used to shard packed weights. The more basic ColwiseParallel or RowwiseParallel will incorrectly shard the packed weights.

The example below packs up_proj and gate_proj into a single gate_up_proj module and requires the PackedRowwiseParallel strategy to shard gate_up_proj.

class Llama4TextExperts(nn.Module):
    ...
    self.gate_up_proj = nn.Parameter(torch.empty(self.num_experts, self.hidden_size, 2 * self.expert_dim))

Batch matrix multiplication can be used in the forward pass to compute the output of the gate_up_proj module.

def forward(self, hidden_states):
    ...
    gate_up = torch.bmm(hidden_states, self.gate_up_proj) # Compute the output of the gate_up_proj module
    gate, up = gate_up.chunk(2, dim=-1) # Split the output into gate and up

Refer to this comment for an visual representation of why Packed* needs to be used.

Local strategies

Local strategies (local_colwise, local_rowwise, local_packed_rowwise) don’t use DTensor because it isn’t supported for some operations such as torch.chunk. Instead, local strategies use the basic torch.Tensor and performs some of the distributed logic manually.

Custom partitioning strategies

A custom partitioning strategy should inherit from TensorParallelLayer and implement partition_tensor, _prepare_input_fn and _prepare_output_fn.

Then it needs to be registered in the ParallelInterface mapping so the dispatching logic can find it when specified in tp_plan.

The example below shows how to implement ColwiseParallel with this workflow.

Inherit from TensorParallelLayer. In the __init__ method, define input_layouts and output_layouts to describe how the input and output tensors should be placed on devices. The desired_input_layouts attribute is used to specify how the input should be placed on devices.

class ColwiseParallel(TensorParallelLayer):
    def __init__(
        self,
        *,
        input_layouts: Optional[Placement] = None, # The input layout coming from the previous layer
        output_layouts: Optional[Placement] = None, # The output layout we want to achieve
        use_local_output: bool = True, # Whether to use local output or not
        use_dtensor=True, # Whether to use DTensor or not
    ):
        self.input_layouts = (input_layouts or Replicate(),) # The input sharding coming from the previous layer
        self.output_layouts = (output_layouts or Shard(-1),) # Desired output sharding
        self.desired_input_layouts = (Replicate(),) # Desired input sharding, inputs should be replicated across GPUs
        self.use_local_output = use_local_output
        self.use_dtensor = use_dtensor

Implement the partition_tensor, _prepare_input_fn and _prepare_output_fn methods.

The partition_tensor method partitions the tensor and fills empty_param with the partitioned tensor. Use the utility function get_tensor_shard to help you get the correct shard of the original parameter for a given rank and get_packed_weights to help with packed weights.

def partition_tensor(
    self,
    param, # Full tensor of the parameter
    empty_param, # Empty tensor of the parameter, will be filled with the partitioned tensor
    param_type, # Type of the parameter, `bias` or `weight`
    param_casting_dtype, # The type to cast the parameter to
    to_contiguous, # Whether to convert the tensor to a contiguous memory layout
    rank, # The rank of the current device
    device_mesh, # The device mesh
) -> nn.Parameter: # Return the partitioned parameter
    ...

The _prepare_input_fn and _prepare_output_fn methods are used in the pre-forward and forward hooks. They redistribute the inputs and outputs to the desired layout as specified in the __init__.

def _prepare_input_fn(input_layouts, desired_input_layouts, mod, inputs, device_mesh):
    ...
    # Do some custom logic, cast to DTensor etc.
    ...
    return inputs.redistribute(placements=desired_input_layouts, device_mesh=device_mesh)
def _prepare_output_fn(output_layouts, use_local_output, mod, outputs, device_mesh):
    ...
    # Do some custom logic, cast to DTensor etc.
    ...
    return outputs.redistribute(placements=output_layouts, device_mesh=device_mesh)

from transformers.integrations.tensor_parallel import ParallelInterface

ParallelInterface.register_strategy("colwise_custom", ColwiseParallel)
tp_plan = {
    "model.layers.*.self_attn.q_proj": "colwise_custom",
    ...
}
model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype=torch.bfloat16, tp_plan=tp_plan)

Benchmarks

Tensor parallelism can considerably speedup inference, especially for inputs with large batch sizes or long sequences.

Refer to the chart below for the expected speedup for a single forward pass on Llama with a sequence length of 512.

Design implementation

The Transformers tensor parallelism implementation is framework-agnostic, but for specific implementations, we rely on DeviceMesh and DTensor from torch.distributed to provide a simple and extensible interface.

DeviceMesh

Imagine DeviceMesh as a multi-dimensional grid of devices that communicate together. Different parallelization strategies require different types of communication patterns, so you can create a DeviceMesh with multiple sub-meshes.

from torch.distributed.device_mesh import init_device_mesh

# Create a 1D mesh of 4 GPUs
device_mesh = init_device_mesh("cuda", (4,), mesh_dim_names=["tp"])

Most of the torch.distributed defined parallelization strategies can be applied to the mesh itself, or its sub-mesh, and it automatically handles the communication patterns.

DTensor

DTensor (Distributed Tensor) is a tensor subclass that handles the distributed logic on top of the usual tensor operations. Most of the model weights in tensor parallelism are stored as DTensors.

The most important part of DTensor is the placement attribute because it tells PyTorch how a tensor is placed on the devices in DeviceMesh. The placement attribute can take the following values.

Shard(dimension) - Indicates how a DTensor is sharded across a given dimension, over the DeviceMesh it was constructed under. The example below demonstrates how to shard weights over different dimensions for column-wise partitioning.

weight = ...
weight = DTensor.from_local(weight, device_mesh["tp"], placements=[Shard(0)]) # Shard across the 1st (column-wise) dimension
bias = ...
bias = DTensor.from_local(bias, device_mesh["tp"], placements=[Shard(-1)]) # Shard across the ONLY dimension

This example demonstrates how to shard weights over different dimensions for row-wise partitioning.

weight = ...
weight = DTensor.from_local(weight, device_mesh["tp"], placements=[Shard(1)]) # Shard across the 2nd (row-wise) dimension
bias = ...
bias = DTensor.from_local(bias, device_mesh["tp"], placements=[Replicate()]) # Replicate bias across all GPUs

Replicate() - Indicates a DTensor is replicated across the DeviceMesh. It only creates a full copy of the tensor on each device.
```
bias = ...
bias = DTensor.from_local(bias, device_mesh["tp"], placements=[Replicate()]) # Replicate bias across all GPUs
```
Partial() - Indicates a tensor is pending a reduction operation (not typically relevant for usage in Transformers).

< > Update on GitHub