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
@@ -67,6 +67,7 @@ def compute_skip_topk_layers(config: PretrainedConfig) -> set[int]:
|
||||
Raises:
|
||||
ValueError: If ``indexer_types`` has the wrong length or an unknown
|
||||
entry, or if ``index_topk_freq`` is not a positive integer.
|
||||
|
||||
"""
|
||||
if not hasattr(config, "index_topk"):
|
||||
return set()
|
||||
@@ -115,6 +116,7 @@ def is_skip_topk_indexer_weight(weight_name: str, skip_topk_layers: set[int]) ->
|
||||
Returns:
|
||||
True when the weight is an indexer weight of a layer that has no
|
||||
indexer module and therefore must be dropped.
|
||||
|
||||
"""
|
||||
if ".indexer." not in weight_name or not skip_topk_layers:
|
||||
return False
|
||||
@@ -557,6 +559,7 @@ class HYV4MLAAttention(nn.Module):
|
||||
The backend class to bind, or None when no sink-capable backend is
|
||||
available; the caller then loads the sink weight but disables the
|
||||
bias.
|
||||
|
||||
"""
|
||||
head_size = self.kv_lora_rank + self.qk_rope_head_dim
|
||||
dtype = torch.get_default_dtype()
|
||||
|
||||
@@ -99,6 +99,7 @@ class HYV4FlashMLASparseImpl(FlashMLASparseImpl):
|
||||
Raises:
|
||||
ValueError: If the dtype is not float32 or the shape is not
|
||||
``(num_heads,)``.
|
||||
|
||||
"""
|
||||
if sinks is None:
|
||||
return
|
||||
@@ -134,6 +135,7 @@ class HYV4FlashMLASparseImpl(FlashMLASparseImpl):
|
||||
Raises:
|
||||
ValueError: If the sink and query head layouts disagree, or if they
|
||||
live on different devices.
|
||||
|
||||
"""
|
||||
sinks = self.sinks
|
||||
if sinks is None:
|
||||
|
||||
@@ -107,6 +107,7 @@ class HYV4HCPreLayer(nn.Module):
|
||||
Returns:
|
||||
A tuple of the pre-gated reduction ``[num_tokens, d]`` and the post
|
||||
gates ``[num_tokens, hc]`` consumed by `HYV4HCPostLayer`.
|
||||
|
||||
"""
|
||||
if self.hpc_op is not None:
|
||||
return self.hpc_op(x)
|
||||
@@ -186,6 +187,7 @@ class HYV4HCPostLayer(nn.Module):
|
||||
|
||||
Returns:
|
||||
The updated residual channels ``[num_tokens, hc, d]``.
|
||||
|
||||
"""
|
||||
if self.hpc_op is not None:
|
||||
return self.hpc_op(x, residual, post)
|
||||
@@ -272,6 +274,7 @@ class HYV4HCHeadLayer(nn.Module):
|
||||
|
||||
Returns:
|
||||
The merged hidden state ``[num_tokens, d]``.
|
||||
|
||||
"""
|
||||
if self.hpc_op is not None:
|
||||
return self.hpc_op(x)
|
||||
@@ -367,6 +370,7 @@ class HYV4HCLayer(nn.Module):
|
||||
A tuple of the reduced hidden states ``[num_tokens, d]``, the post
|
||||
gates ``[num_tokens, hc]`` (``None`` when iHC is disabled) and the
|
||||
residual (the untouched input).
|
||||
|
||||
"""
|
||||
if not self.enable_ihc:
|
||||
return hidden_states, None, hidden_states
|
||||
|
||||
@@ -189,7 +189,7 @@ class HYV4DecoderLayer(nn.Module):
|
||||
positions: torch.Tensor,
|
||||
hidden_states: torch.Tensor,
|
||||
) -> tuple[torch.Tensor, None]:
|
||||
"""iHC forward: each sub-block reduces and re-scatters the channels."""
|
||||
"""IHC forward: each sub-block reduces and re-scatters the channels."""
|
||||
hidden_states = self.hc_attn_layer.prepare_input(hidden_states)
|
||||
hidden_states, post_gates, residual = self.hc_attn_layer.pre(hidden_states)
|
||||
hidden_states = self.input_layernorm(hidden_states)
|
||||
|
||||
@@ -106,6 +106,7 @@ def _resolve_fused_expert_param(
|
||||
|
||||
Returns:
|
||||
The matching draft parameter name, or None when there is none.
|
||||
|
||||
"""
|
||||
if not ckpt_suffix:
|
||||
return param_base if param_base in params_dict else None
|
||||
@@ -164,6 +165,7 @@ def _create_mtp_quant_config(
|
||||
Returns:
|
||||
The quantization config to use for the MTP layers, or None when the MTP
|
||||
layers are unquantized.
|
||||
|
||||
"""
|
||||
mtp_quant_algo = getattr(hf_config, "mtp_quant_algo", None)
|
||||
|
||||
@@ -241,6 +243,7 @@ def _remap_mtp_quant_exclusions(
|
||||
|
||||
Returns:
|
||||
A shallow copy with the translated exclusions, or the input unchanged.
|
||||
|
||||
"""
|
||||
if quant_config is None:
|
||||
return None
|
||||
@@ -671,6 +674,7 @@ class HYV4MTP(nn.Module):
|
||||
Returns:
|
||||
True when the weight was consumed (even if this rank holds none of
|
||||
the addressed experts).
|
||||
|
||||
"""
|
||||
base = name.split(".experts.")[0]
|
||||
for ckpt_proj, tag in (
|
||||
|
||||
Reference in New Issue
Block a user