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

[CI] Revert PRs 34818 and 33600 (#34979)

Browse Source
This commit is contained in:
Lucas Wilkinson
2026-02-20 16:25:50 -05:00
committed by GitHub
parent f24b2de3d3
commit aaefc58ee0
16 changed files with 249 additions and 301 deletions
Download Patch File Download Diff File Expand all files Collapse all files

9
tests/models/utils.py
View File

@@ -457,9 +457,6 @@ def dummy_hf_overrides(
# Kimi uses `num_expert_group` instead of `n_group`.
if n_group is None:
n_group = getattr(text_config, "num_expert_group", None)
# InternS1Pro uses `router_n_groups` instead of `n_group`.
if n_group is None:
n_group = getattr(text_config, "router_n_groups", None)
num_experts = n_group * 2 if n_group is not None else 2
# we use three layers for Gemma-3n to check
@@ -489,14 +486,12 @@ def dummy_hf_overrides(
# Only set MoE related config when the model has MoE layers.
# Otherwise all models detected as MoE by _get_transformers_backend_cls.
if model_arch_config.num_experts > 0:
orig_topk = getattr(text_config, "num_experts_per_tok", 2)
topk = min(orig_topk, 2)
update_dict.update(
{
"num_experts": num_experts,
"num_experts_per_tok": topk,
"num_experts_per_tok": 2,
# Kimi uses `num_experts_per_token`.
"num_experts_per_token": topk,
"num_experts_per_token": 2,
"num_local_experts": num_experts,
# Otherwise there will not be any expert layers
"first_k_dense_replace": 0,