diff --git a/comfy/model_management.py b/comfy/model_management.py
index ff963eb8e..980865a77 100644
--- a/comfy/model_management.py
+++ b/comfy/model_management.py
@@ -475,13 +475,16 @@ except:
 SUPPORT_FP8_OPS = args.supports_fp8_compute
 
 AMD_RDNA2_AND_OLDER_ARCH = ["gfx1030", "gfx1031", "gfx1035", "gfx1010", "gfx1011", "gfx1012", "gfx906", "gfx900", "gfx803"]
+# MIOpen fp16 conv kernels segfault on RDNA1/RDNA2 (e.g. gfx1030 with ROCm 6.x), so
+# keep MIOpen only on the oldest GCN/Vega archs where it was the previous default.
+AMD_MIOPEN_KEEP_ENABLED_ARCH = ("gfx906", "gfx900", "gfx803")
 AMD_ENABLE_MIOPEN_ENV = 'COMFYUI_ENABLE_MIOPEN'
 
 try:
     if is_amd():
         arch = torch.cuda.get_device_properties(get_torch_device()).gcnArchName.split(':')[0]
-        if not (any((a in arch) for a in AMD_RDNA2_AND_OLDER_ARCH)):
-            if os.getenv(AMD_ENABLE_MIOPEN_ENV) != '1':
+        if os.getenv(AMD_ENABLE_MIOPEN_ENV) != '1':
+            if not any(a in arch for a in AMD_MIOPEN_KEEP_ENABLED_ARCH):
                 torch.backends.cudnn.enabled = False  # Seems to improve things a lot on AMD
                 logging.info("Set: torch.backends.cudnn.enabled = False for better AMD performance.")