Kunlun Graph

Why we need Kunlun Graph?

When in LLM inference, each token requires nearly thousand operator executions, and when host launching operators are slower than device, it will cause host bound. In severe cases, the device will be idle for more than half of the time. To solve this problem, we use graph in LLM inference.

eager mode:

host:   |  launch op1  |  launch op2  |  launch op3  |  launch op4  |  launch op5  |

device:                | run op1 |free| run op2 |free| run op3 |free| run op4 |free| run op5 |

        | <-----                           total time                                 -----> |

graph mode:

host:   |  launch graph  |

device:                  | run op1 | run op2 | run op3 | run op4 | run op5 |

        | <-----                    total time                      -----> |

How to use Kunlun Graph?

Kunlun Graph is enabled by default in V1 Engine, just need to check that enforce_eager is not set to True.

How it works?

In short, graph mode works in two steps: capture and replay. When engine starts, we will capture all of the ops in model forward and save it as a graph, and when req come in, we just replay the graph on devices, and waiting for result.

But in reality, graph mode is not that simple.

Padding and Bucketing

Due to graph can only replay the ops captured before, without doing tiling and checking graph input, we need to ensure the consistency of the graph input, but we know that model input's shape depends on the request scheduled by Scheduler, we can't ensure the consistency.

Obviously, we can solve this problem by capturing the biggest shape and padding all of the model input to it. But it will bring a lot of redundant computing and make performance worse. So we can capture multiple graphs with different shape, and pad the model input to the nearest graph, which will greatly reduce redundant computing. But when max_num_batched_tokens is very large, the number of graphs that need to be captured will also become very large. But we know that when intensor's shape is large, the computing time will be very long, and graph mode is not necessary in this case. So all of things we need to do is:

Set a threshold;
When num_scheduled_tokens is bigger than the threshold, use eager_mode;
Capture multiple graphs within a range below the threshold;

|    graph1    |
|           graph2           |
|                    graph3                    |
|                              graph4                              |    # the threshold

| input1 | pad |    # use graph1
|           input2           |  # don't need pad
|                      input3                      |      pad      |    # use graph4
|                                    input4                                    |    # use eager mode

Piecewise and Full graph

Due to the increasing complexity of the attention layer in current LLM, we can't ensure all types of attention can run in graph. In MLA, prefill_tokens and decode_tokens have different calculation method, so when a batch has both prefills and decodes in MLA, graph mode is difficult to handle this situation.

vLLM solves this problem with piecewise graph mode. We use eager mode to launch attention's ops, and use graph to deal with others. But it also bring some problems: The cost of launching ops has become large again, although much smaller than eager mode, but it will also lead to host bound when cpu is poor or num_tokens is small.

How it be implemented?

vLLM has already implemented most of the modules in graph mode. You can see more details at: CUDA Graphs

When in graph mode, vLLM will call current_platform.get_static_graph_wrapper_cls to get current device's graph model wrapper, so what we need to do is to implement the graph mode wrapper on Kunlun: Kunlun Graph Wrapper.

vLLM has added support_torch_compile decorator to all models, this decorator will replace the __init__ and forward interface of the model class, and when forward called, the code inside the vllm_kunlun.compilation will be executed, and it will do capture or replay as mentioned above.

Limitation

FULL and FULL_AND_PIECEWISE are not supported now;
use_inductor is not supported now;

4.2 KiB Raw Permalink Blame History