Skip to content
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

Jump to bottom

Estimate adapter memory overhead in choose_num_blocks() #346

Merged
merged 6 commits into from
Jul 13, 2023
Merged

Estimate adapter memory overhead in choose_num_blocks() #346

Changes from 1 commit
Commits
Show all changes
6 commits
Select commit Hold shift + click to select a range
86d1afb
estimate adapter memory overhead
Jul 12, 2023
d18e2d1
Update src/petals/server/server.py
justheuristic Jul 13, 2023
58a92cb
Update src/petals/utils/peft.py
justheuristic Jul 13, 2023
65675d0
Merge branch 'main' into fix-num-blocks
justheuristic Jul 13, 2023
832cb51
black
Jul 13, 2023
49b6cc8
review
borzunov Jul 13, 2023
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
Prev Previous commit
Next Next commit
Update src/petals/server/server.py 
Co-authored-by: Alexander Borzunov <borzunov.alexander@gmail.com>
  • Loading branch information
@justheuristic @borzunov
justheuristic and borzunov committed Jul 13, 2023
commit d18e2d14fba6a50f66932ed07a0f850469a0a659
10 changes: 7 additions & 3 deletions src/petals/server/server.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -250,9 +250,13 @@ def _choose_num_blocks(self) -> int:
# Estimate of GPU memory used in rpc_backward (2 GiB for BLOOM, proportional for other models)
autograd_memory = 2 * gib * num_devices / 14336 * self.block_config.hidden_size

adapter_memory_per_block = estimate_adapter_memory_per_block(
self.block_config, self.torch_dtype, self.adapters, self.cache_dir
)
if adapters:
# Delay import of petals.utils.peft to avoid unnecessary import of bitsandbytes
from petals.utils.peft import estimate_adapter_memory_per_block

adapter_memory_per_block = estimate_adapter_memory_per_block(
self.block_config, self.torch_dtype, self.adapters, self.cache_dir
)
total_memory_per_block = block_size + adapter_memory_per_block + self._cache_bytes_per_block

num_blocks = math.floor((total_memory - autograd_memory) / total_memory_per_block)
Expand Down
Loading