From aeee7ef9391028939afd08e20c12a1e279efbdf1 Mon Sep 17 00:00:00 2001 From: Rishapveer Singh Date: Sat, 18 Apr 2026 07:34:33 +0200 Subject: [PATCH] [Bugfix] Fix k_proj's bias for GLM-ASR (#40160) Signed-off-by: Rishapveer Singh --- vllm/model_executor/models/glmasr.py | 4 +++- 1 file changed, 3 insertions(+), 1 deletion(-) diff --git a/vllm/model_executor/models/glmasr.py b/vllm/model_executor/models/glmasr.py index 0d54588cc2..96be79a309 100644 --- a/vllm/model_executor/models/glmasr.py +++ b/vllm/model_executor/models/glmasr.py @@ -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"),