xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2026-03-24 04:16:52 +08:00
Code Issues Packages Projects Releases Wiki Activity

[Docs] improve code formatting and comments for eliminate griffe build warning. (#25010)

Browse Source 
Signed-off-by: samzong <samzong.lu@gmail.com>
This commit is contained in:
samzong 2025-09-17 22:31:20 +08:00 committed by GitHub
parent dd6a910aac
commit 47f670b03b
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
3 changed files with 20 additions and 14 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
vllm/benchmarks/serve.py
View File

@ -139,7 +139,7 @@ async def get_request(
A lower burstiness value (0 < burstiness < 1) results
in more bursty requests, while a higher burstiness value
(burstiness > 1) results in a more uniform arrival of requests.
ramp_up_strategy (optional):
ramp_up_strategy (optional):
The ramp-up strategy. Can be "linear" or "exponential".
If None, uses constant request rate (specified by request_rate).
ramp_up_start_rps (optional):

9
vllm/distributed/eplb/eplb_state.py
View File

@ -337,11 +337,12 @@ class EplbState:
Args:
model (MixtureOfExperts): The MoE model.
is_dummy (bool): If `True`, this is a dummy step and the load
metrics recorded in this forward pass will not count. Defaults
to `False`.
metrics recorded in this forward pass will not count.
Defaults to `False`.
is_profile (bool): If `True`, perform a dummy rearrangement
with maximum communication cost. This is used in `profile_run`
to reserve enough memory for the communication buffer.
with maximum communication cost. This is used in
`profile_run` to reserve enough memory
for the communication buffer.
log_stats (bool): If `True`, log the expert load metrics.
# Stats

23
vllm/distributed/eplb/rebalance_algo.py
View File

@ -109,13 +109,16 @@ def rebalance_experts_hierarchical(
num_physical_experts: number of physical experts after replication
num_groups: number of expert groups
num_nodes: number of server nodes, where the intra-node network
(e.g, NVLink) is faster
(e.g., NVLink) is faster
num_gpus: number of GPUs, must be a multiple of `num_nodes`
Returns:
physical_to_logical_map: [num_moe_layers, num_physical_experts]
logical_to_physical_map: [num_moe_layers, num_logical_experts, X]
logical_count: [num_moe_layers, num_logical_experts]
physical_to_logical_map (torch.Tensor):
[num_moe_layers, num_physical_experts]
logical_to_physical_map (torch.Tensor):
[num_moe_layers, num_logical_experts, X]
logical_count (torch.Tensor):
[num_moe_layers, num_logical_experts]
"""
num_layers, num_logical_experts = weight.shape
assert num_logical_experts % num_groups == 0
@ -197,11 +200,13 @@ def rebalance_experts(
num_gpus: number of GPUs, must be a multiple of `num_nodes`
Returns:
physical_to_logical_map: [layers, num_replicas], the expert index of
each replica
logical_to_physical_map: [layers, num_logical_experts, X], the replica
indices for each expert
expert_count: [layers, num_logical_experts], number of physical
physical_to_logical_map:
[layers, num_replicas], the expert index of each replica
logical_to_physical_map:
[layers, num_logical_experts, X], the replica indices for each
expert
expert_count:
[layers, num_logical_experts], number of physical
replicas for each logical expert
"""
num_layers, num_logical_experts = weight.shape