Skip to content

some fixes to make training work #5

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
sfc-gh-rsamdani wants to merge 13 commits into
base: optim-linear
Choose a base branch
from
Open

some fixes to make training work #5

Changes from all commits
Commits
Show all changes
13 commits
Select commit Hold shift + click to select a range
d6d1bfe
ensure capacity does not exceed number of tokens
jeffra Apr 2, 2024
c96e350
auto convert moe param groups
jeffra Apr 2, 2024
e942e76
remove unrelated test
jeffra Apr 2, 2024
2c42ef1
Update test_moe.py
jeffra Apr 2, 2024
5905f9b
Merge branch 'master' into auto-param-group
jeffra Apr 2, 2024
a876db7
Merge pull request #3 from Snowflake-Labs/optim-linear
sfc-gh-jrasley Apr 4, 2024
955e37c
Merge branch 'master' into auto-param-group
jeffra Apr 4, 2024
fb67b34
Merge pull request #4 from Snowflake-Labs/auto-param-group
sfc-gh-jrasley Apr 4, 2024
8a34783
some fixes to make lora training work
sfc-gh-rsamdani Apr 17, 2024
5adb928
small
sfc-gh-rsamdani Apr 17, 2024
f8fc11f
small
sfc-gh-rsamdani Apr 21, 2024
ebcdead
small
sfc-gh-rsamdani Apr 21, 2024
35af944
small
sfc-gh-rsamdani Apr 21, 2024
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
6 changes: 3 additions & 3 deletions deepspeed/linear/optimized_linear.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -86,11 +86,11 @@ def __init__(self,
self.zero_shards = self.lora_config.base_weight_sharding
self.sharded_weight_size = int(float(self.input_dim) // self.zero_shards)
w = torch.nn.Parameter(torch.empty((self.output_dim, self.sharded_weight_size), dtype=dtype))
torch.nn.init.xavier_uniform_(w)

if self.quantization_config is not None:
assert dtype == torch.bfloat16, "only bfloat16 is supported when using quantization"
self.base_weight = QuantizedParameter(w, quantization_config=quantization_config)
self.base_weight = QuantizedParameter(w,
requires_grad=False, # quantized weights must be frozen
quantization_config=quantization_config)
else:
self.base_weight = w

Expand Down