[Bugfix] Fix k_proj's bias for GLM-ASR (#40160)

Signed-off-by: Rishapveer Singh <singhrishapveer@gmail.com>
This commit is contained in:
Rishapveer Singh
2026-04-17 22:34:33 -07:00
committed by GitHub
parent cda19ecf4d
commit aeee7ef939
+3 -1
View File
@@ -66,7 +66,7 @@ from .interfaces import (
SupportsTranscription,
)
from .utils import AutoWeightsLoader, init_vllm_registered_model, maybe_prefix
from .whisper import ISO639_1_SUPPORTED_LANGS
from .whisper import ISO639_1_SUPPORTED_LANGS, _create_fake_bias_for_k_proj
class GlmAsrEncoderRotaryEmbedding(nn.Module):
@@ -499,6 +499,8 @@ class GlmAsrEncoder(nn.Module):
"""Custom weight loading to handle q_proj/k_proj/v_proj -> qkv_proj mapping."""
from vllm.model_executor.model_loader.weight_utils import default_weight_loader
weights = _create_fake_bias_for_k_proj(weights, ".k_proj.weight")
stacked_params_mapping = [
# (param_name, shard_name, shard_id)
("qkv_proj", "q_proj", "q"),