[CPU] Fix Qwen-Next SSM type for AMX GDN (#48073)
Signed-off-by: jiang1.li <jiang1.li@intel.com>
Li, Jiang committed
Jul 9, 2026 at 15:09 UTC
1171467e91b3a1f334ad99df934ce35b9fb7255c
1 file changed
+8
vllm/platforms/cpu.py
+8
@@ -125,6 +125,14 @@ class CpuPlatform(Platform):
125
"otherwise the performance is not optimized."
126
)
127
128
+ # AMX GDN requires float32 state
129
+ if (
130
+ torch.cpu._is_amx_tile_supported()
131
+ and cache_config.mamba_ssm_cache_dtype != "float32"
132
+ ):
133
+ cache_config.mamba_ssm_cache_dtype = "float32"
134
+ logger.warning("Reset SSM cache type to float32 for AMX mamba attention.")
135
+
136
# Lagecy setting
137
env_key = "VLLM_CPU_KVCACHE_SPACE"
138
if env_key in os.environ and os.environ[env_key] != "":