37 lines
1.1 KiB
Python
37 lines
1.1 KiB
Python
import logging
|
|
import re
|
|
from functools import lru_cache
|
|
|
|
import torch
|
|
|
|
logger = logging.getLogger(__name__)
|
|
|
|
|
|
def get_layer_id(weight_name):
|
|
# example weight name: model.layers.10.self_attn.qkv_proj.weight
|
|
match = re.search(r"layers\.(\d+)\.", weight_name)
|
|
if match:
|
|
return int(match.group(1))
|
|
return None
|
|
|
|
|
|
class PPMissingLayer(torch.nn.Identity):
|
|
# Adapted from
|
|
# https://github.com/vllm-project/vllm/blob/18ed3132d2bfe1df9a74729457b69243955221e8/vllm/model_executor/models/utils.py#L468C1-L486C1
|
|
"""
|
|
A placeholder layer for missing layers in a pipeline parallel model.
|
|
"""
|
|
|
|
def __init__(self, *args, **kwargs):
|
|
super().__init__()
|
|
self.return_tuple = kwargs.get("return_tuple", False)
|
|
|
|
def forward(self, *args, **kwargs):
|
|
"""
|
|
Return the first arg from args or the first value from kwargs.
|
|
|
|
Wraps the input in a tuple if `self.return_tuple` is True.
|
|
"""
|
|
input = args[0] if args else next(iter(kwargs.values()))
|
|
return (input,) if self.return_tuple else input
|