[Mypy Fix] Mypy fix for "vllm/model_executor/models/[tT]" (#53466)
Signed-off-by: Jensen Chen <a1043904820@gmail.com> Co-authored-by: Claude Opus 5 <noreply@anthropic.com> Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
Jensen Chen committed
Aug 26, 2026 at 07:12 UTC
c47ca4aaebafbbcbe47d0ae1f3b9672bcf1b4ed8
3 files changed
+14
-9
tools/pre_commit/mypy.py
+1
-1
@@ -124,7 +124,7 @@ EXCLUDE = [
124
r"vllm/model_executor/models/[qQ]",
125
r"vllm/model_executor/models/[rR]",
126
r"vllm/model_executor/models/[sS]",
127
- r"vllm/model_executor/models/[tT]",
127
+ r"vllm/model_executor/models/transformers",
128
r"vllm/model_executor/models/[uU]",
129
r"vllm/model_executor/models/[vV]",
130
r"vllm/model_executor/models/[wW]",
vllm/model_executor/models/telechat2.py
+12
-7
@@ -28,7 +28,7 @@ import torch.nn as nn
28
from vllm.config import VllmConfig
29
from vllm.model_executor.models.llama import LlamaForCausalLM, LlamaModel
30
31
-from .llama import LlamaDecoderLayer
31
+from .llama import LlamaAttention, LlamaDecoderLayer, LlamaMLP
32
from .utils import AutoWeightsLoader, PPMissingLayer, WeightsMapper
33
34
@@ -46,7 +46,7 @@ class TeleChat2Model(LlamaModel):
46
def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
47
hf_config = vllm_config.model_config.hf_config
48
49
- vllm_config.model_config.hf_config.attribute_map = {
49
+ vllm_config.model_config.hf_config.attribute_map = { # type: ignore[misc]
50
"num_hidden_layers": "n_layer",
51
"num_attention_heads": "n_head",
52
"intermediate_size": "ffn_hidden_size",
@@ -63,11 +63,16 @@ class TeleChat2Model(LlamaModel):
63
# Telechat2's gate_up_proj and qkv_proj don't have bias
64
# see: https://github.com/vllm-project/vllm/pull/10311#issuecomment-2490297566
65
for layer in self.layers:
66
- if not isinstance(layer, PPMissingLayer):
67
- layer.self_attn.qkv_proj.bias = None
68
- layer.self_attn.qkv_proj.skip_bias_add = True
69
- layer.mlp.gate_up_proj.bias = None
70
- layer.mlp.gate_up_proj.skip_bias_add = True
66
+ if isinstance(layer, PPMissingLayer):
67
+ continue
68
+ self_attn = layer.self_attn
69
+ mlp = layer.mlp
70
+ assert isinstance(self_attn, LlamaAttention)
71
+ assert isinstance(mlp, LlamaMLP)
72
+ self_attn.qkv_proj.register_parameter("bias", None)
73
+ self_attn.qkv_proj.skip_bias_add = True
74
+ mlp.gate_up_proj.register_parameter("bias", None)
75
+ mlp.gate_up_proj.skip_bias_add = True
76
77
def _split_key_value(
78
self, weights: Iterable[tuple[str, torch.Tensor]]
vllm/model_executor/models/terratorch.py
+1
-1
@@ -289,7 +289,7 @@ class Terratorch(nn.Module, IsAttentionFree, SupportsMultiModal):
289
def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
290
params_list = []
291
model_buffers = dict(self.named_buffers())
292
- loaded_buffers = []
292
+ loaded_buffers: list[str] = []
293
for key, value in weights:
294
if isinstance(value, (dict, OrderedDict)):
295
if key == "state_dict":