biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[Feature] use --eplb_config to set eplb param (#20562)

Browse Source 
Signed-off-by: rongfu.leng <rongfu.leng@daocloud.io>
Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
Signed-off-by: rongfu.leng <lenronfu@gmail.com>
Co-authored-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
This commit is contained in:
rongfu.leng
2025-08-21 05:07:28 +08:00
committed by GitHub
parent 4e51fa8cba
commit 4fbda0b20c
9 changed files with 149 additions and 52 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
vllm/distributed/eplb/eplb_state.py
View File

@@ -244,7 +244,7 @@ class EplbState:
dtype=torch.int32,
device=device,
)
expert_load_window_size = parallel_config.eplb_window_size
expert_load_window_size = parallel_config.eplb_config.window_size
expert_load_window = torch.zeros(
(expert_load_window_size, model.num_moe_layers,
model.num_physical_experts),
@@ -253,7 +253,7 @@ class EplbState:
)
# Set the initial progress of rearrangement to 3/4
eplb_step_interval = parallel_config.eplb_step_interval
eplb_step_interval = parallel_config.eplb_config.step_interval
expert_rearrangement_step = max(
0, eplb_step_interval - eplb_step_interval // 4)