xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2026-06-02 02:31:19 +08:00
Code Issues Packages Projects Releases Wiki Activity

[Bugfix] Add logs for all model dtype casting (#4717)

Browse Source
This commit is contained in:
Michael Goin 2024-05-09 14:36:25 -04:00 committed by GitHub
parent cea64430f6
commit be0c5180ac
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 3 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

3
vllm/config.py
View File

@ -1063,6 +1063,7 @@ def _get_and_verify_dtype(
if config_dtype == torch.float32: if config_dtype == torch.float32:
# Following the common practice, we use float16 for float32 # Following the common practice, we use float16 for float32
# models. # models.
logger.info("Casting torch.float32 to torch.float16.")
torch_dtype = torch.float16 torch_dtype = torch.float16
else: else:
torch_dtype = config_dtype torch_dtype = config_dtype
@ -1087,9 +1088,11 @@ def _get_and_verify_dtype(
if torch_dtype != config_dtype: if torch_dtype != config_dtype:
if torch_dtype == torch.float32: if torch_dtype == torch.float32:
# Upcasting to float32 is allowed. # Upcasting to float32 is allowed.
logger.info("Upcasting %s to %s.", config_dtype, torch_dtype)
pass pass
elif config_dtype == torch.float32: elif config_dtype == torch.float32:
# Downcasting from float32 to float16 or bfloat16 is allowed. # Downcasting from float32 to float16 or bfloat16 is allowed.
logger.info("Downcasting %s to %s.", config_dtype, torch_dtype)
pass pass
else: else:
# Casting between float16 and bfloat16 is allowed with a warning. # Casting between float16 and bfloat16 is allowed with a warning.