[Misc] Remove deprecated calculate_kv_scales runtime KV scale calculation (#49389)

Signed-off-by: wangxiyuan <wangxiyuan1007@gmail.com>
This commit is contained in:
wangxiyuan
2026-08-03 11:15:13 -04:00
committed by GitHub
parent 0cf49a5d15
commit dd11df04f3
19 changed files with 16 additions and 263 deletions
-17
View File
@@ -445,23 +445,6 @@ class HybridAttentionMambaModelConfig(VerifyAndUpdateConfig):
Args:
vllm_config: vLLM Config
"""
cache_config = vllm_config.cache_config
# Disable calculate_kv_scales for hybrid models: uninitialized
# recurrent state corrupts scales during the calibration pass.
# See issue: https://github.com/vllm-project/vllm/issues/37554
if cache_config.calculate_kv_scales:
logger.warning(
"Disabling calculate_kv_scales for hybrid model '%s'. "
"Hybrid models with recurrent layers (GDN, Mamba, SSM) "
"produce unreliable KV cache scales during the "
"calibration pass because recurrent state is "
"uninitialized. Using default scale of 1.0 instead.",
vllm_config.model_config.model,
)
cache_config.calculate_kv_scales = False
# Enable FULL_AND_PIECEWISE by default
MambaModelConfig.verify_and_update_config(vllm_config)