Skip to content

[Fix] Prevent FSDPModule from auto-converting position_embedding to bf16 in DecoderLayer #1462

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
HAOCHENYE wants to merge 3 commits into
base: gh/HAOCHENYE/13/base
Choose a base branch
from
+4 −1
Open

[Fix] Prevent FSDPModule from auto-converting position_embedding to bf16 in DecoderLayer #1462

Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
cd350aa
Update
HAOCHENYE Jan 29, 2026
23f10c5
Update
HAOCHENYE Jan 29, 2026
1edfb1d
Update
HAOCHENYE Jan 31, 2026
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
5 changes: 4 additions & 1 deletion xtuner/v1/model/compose/qwen3_vl/modeling_vision.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -316,6 +316,9 @@ def fully_shard(
mp_policy = MixedPrecisionPolicy(
param_dtype=fsdp_config.param_dtype, reduce_dtype=fsdp_config.reduce_dtype
)
decoder_layer_mp_policy = MixedPrecisionPolicy(
param_dtype=fsdp_config.param_dtype, reduce_dtype=fsdp_config.reduce_dtype, cast_forward_inputs=False
)

# NOTE: 在 cpu_offload 模式下，mesh 应该是 cuda 的，在 meta fully_shard 后在调用 .to_empty(device=cpu)
self.fsdp_mesh = init_world_mesh()
Expand Down Expand Up @@ -350,7 +353,7 @@ def fully_shard(
fully_shard(
layer,
mesh=self.fsdp_mesh,
mp_policy=mp_policy,
mp_policy=decoder_layer_mp_policy,
reshard_after_forward=True,
offload_policy=CPUOffloadPolicy()
if fsdp_config.cpu_offload
Expand Down