A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 5, 01:00 UTC
[Bugfix] Fix k_proj's bias for GLM-ASR (#40160)
Signed-off-by: Rishapveer Singh <singhrishapveer@gmail.com>
Rishapveer Singh committed
Apr 18, 2026 at 07:34 UTC
aeee7ef9391028939afd08e20c12a1e279efbdf1
1 file changed
+3
-1
vllm/model_executor/models/glmasr.py
+3
-1
@@ -66,7 +66,7 @@ from .interfaces import (
66
SupportsTranscription,
67
)
68
from .utils import AutoWeightsLoader, init_vllm_registered_model, maybe_prefix
69
-from .whisper import ISO639_1_SUPPORTED_LANGS
69
+from .whisper import ISO639_1_SUPPORTED_LANGS, _create_fake_bias_for_k_proj
70
71
72
class GlmAsrEncoderRotaryEmbedding(nn.Module):
@@ -499,6 +499,8 @@ class GlmAsrEncoder(nn.Module):
499
"""Custom weight loading to handle q_proj/k_proj/v_proj -> qkv_proj mapping."""
500
from vllm.model_executor.model_loader.weight_utils import default_weight_loader
501
502
+ weights = _create_fake_bias_for_k_proj(weights, ".k_proj.weight")
503
+
504
stacked_params_mapping = [
505
# (param_name, shard_name, shard_id)
506
("qkv_proj", "q_proj", "q"),