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 2, 04:45 UTC
[BugFix] [GDN] Read linear_key_head_dim from hf_text_config for multimodal models (#43978)
Signed-off-by: IdoAtadTD <ido.atad@twodelta.com>
IdoAtadTD committed
Jun 2, 2026 at 16:17 UTC
c91a87f01a2b3947d087706241fe6348c978324d
1 file changed
+2
-2
vllm/model_executor/layers/mamba/gdn/qwen_gdn_linear_attn.py
+2
-2
@@ -177,7 +177,7 @@ def _resolve_gdn_prefill_backend(
177
return backend, "triton"
178
179
head_k_dim = getattr(
180
- vllm_config.model_config.hf_config, "linear_key_head_dim", None
180
+ vllm_config.model_config.hf_text_config, "linear_key_head_dim", None
181
)
182
183
supports_flashinfer = False
@@ -218,7 +218,7 @@ def _log_gdn_backend_decision(
218
) -> None:
219
"""Log the GDN prefill backend choice in the attention-selector style."""
220
head_k_dim = getattr(
221
- vllm_config.model_config.hf_config, "linear_key_head_dim", None
221
+ vllm_config.model_config.hf_text_config, "linear_key_head_dim", None
222
)
223
chosen = {
224
"flashinfer": "FlashInfer",