Skip to content

Assuming the input and output have fixed lengths of p and d respectively, how should I perform profiling and auto-search? #46

Description

@zhanggj821

Assuming the input and output have fixed lengths of p and d respectively, how should I perform profiling and auto-search?

In the profile_one_cycle function within run_llama3.py, how should I set the total_batch_sizes for the prefill and decode phases?

total_batch_sizes = [128, 256, 384, 512, 640, 768, 896, 1024, 1152, 1280, 1408, 1536, 1664, 1792, 1920, 2048]

total_batch_sizes = [128, 256, 384, 512, 640]

When profiling the decode phase, should I set input_length to p and output_length to d ?

input_length = 1024
output_length = 0

In new_search.py, how do I determine the values for global_batch_size, decode_batch_size, and seq_len?

global_batch_size = 3072
decode_batch_size = 1280
# global_batch_size = 2048
# decode_batch_size = 640
>>>>>>> origin/master
# global_batch_size = 1024
# decode_batch_size = 384
seq_len = 1024

I'm looking forward to your reply very much.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions