Add pydocstyle to the ruff rules (#52136)
Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com> Signed-off-by: Harry Mellor <hej.mellor@gmail.com> Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
This commit is contained in:
co-authored by
Claude Opus 5
parent
804e5377aa
commit
8c1557a79c
@@ -88,8 +88,7 @@ def _normalize_expert_bias(
|
||||
|
||||
|
||||
class Param2MoEAttention(nn.Module):
|
||||
"""
|
||||
Grouped-Query Attention (GQA) for Param2MoE.
|
||||
"""Grouped-Query Attention (GQA) for Param2MoE.
|
||||
|
||||
Notable differences from a vanilla GQA layer:
|
||||
* The checkpoint fuses Q, K, V into a single ``query_key_value`` weight.
|
||||
@@ -252,8 +251,7 @@ class Param2MoEMLP(nn.Module):
|
||||
|
||||
|
||||
class Param2MoEMoEBlock(nn.Module):
|
||||
"""
|
||||
Mixture-of-Experts block for Param2MoE.
|
||||
"""Mixture-of-Experts block for Param2MoE.
|
||||
|
||||
Routing:
|
||||
* Sigmoid scoring (config.score_function = "sigmoid")
|
||||
@@ -371,8 +369,7 @@ class Param2MoEMoEBlock(nn.Module):
|
||||
|
||||
|
||||
class Param2MoEDecoderLayer(nn.Module):
|
||||
"""
|
||||
Single transformer decoder block.
|
||||
"""Single transformer decoder block.
|
||||
|
||||
Dense for the first ``first_k_dense_replace`` layers; MoE thereafter.
|
||||
"""
|
||||
@@ -605,8 +602,7 @@ class Param2MoEMixtureOfExperts(MixtureOfExperts):
|
||||
class Param2MoEForCausalLM(
|
||||
nn.Module, SupportsPP, SupportsLoRA, Param2MoEMixtureOfExperts
|
||||
):
|
||||
"""
|
||||
vLLM-native Param2MoE CausalLM.
|
||||
"""vLLM-native Param2MoE CausalLM.
|
||||
|
||||
Uses Grouped-Query Attention (GQA) with a Sigmoid-scored,
|
||||
grouped-topk Mixture-of-Experts MLP.
|
||||
|
||||
Reference in New Issue
Block a user