# Design Doc: Large Model

## Abstract

We propose an approach to support the large parameter.
For embedding layer, the parameter may very large and could
not be stored in one trainer's memory. In this approach, a Trainer would
prefetch a sliced parameter from different Parameter Server instances
according to the input `Ids`, and then run forward, backward and send
the gradient to Parameter Server to execute the optimize program.

## Design

**NOTE**: this approach is a feature of Fluid distributed trianing, maybe you want
to know [Distributed Architecture](./distributed_architecture.md) and
[Parameter Server](./parameter_server.md) before reading the following content.

Fluid large model distributed training use 
[Distributed Transpiler](./parameter_server.md#distributed-transpiler) to split
a large parameter into multiple parameters which stored on Parameter Server, and
the Trainer would prefetch them by `RPC` interface.

### Split Large Parameter

<img src="src/split_parameter.png" width="400" />

**Distributed Transpiler** would split the large parameter
(weight) into some sliced parameters (weight_0, weight_1, weight_2) as the 
figure above.

### Prefetch Parameters from Parameter Servers

<img src="src/prefetch_parameters.png" width="400" />

- `PrefetchRpc` operator would send the rows index the multiple Parameter Servers,
  and then receive the SelctedRows.
- The different with normal Fluid distributed training, we only prefetch the rows

## TODO

- Async Update

  To avoid slow-node, Async update is important for distributed training,
  we need an design doc and implement it in future.