add force dtype on load

Signed-off-by: Vladimir Mandic <mandic00@live.com>

modules/devices.py

@@ -377,7 +377,7 @@ def test_bf16():
             else:
                 from modules.zluda_installer import default_agent
                 agent = default_agent
-            if agent is not None and agent.gfx_version < 0x1100 and agent.arch != rocm.MicroArchitecture.CDNA: # all cards before RDNA 3 except for CDNA cards
+            if (agent is not None) and (agent.gfx_version < 0x1100) and (agent.arch != rocm.MicroArchitecture.CDNA): # all cards before RDNA 3 except for CDNA cards
                 bf16_ok = False
                 return bf16_ok
     try:

modules/sd_hijack_te.py

@@ -40,6 +40,8 @@ def hijack_encode_prompt(*args, **kwargs):
     timer.process.add('te', t1-t0)
     shared.sd_model = sd_models.apply_balanced_offload(shared.sd_model)
     shared.state.end(jobid)
+    from modules import memstats
+    log.debug(f'Encode: memory={memstats.memory_stats()}')
     return res
 
 

modules/ui_definitions.py

@@ -220,6 +220,7 @@ def create_settings(cmd_opts):
         "math_sep": OptionInfo("<h2>Execution Precision</h2>", "", gr.HTML),
         "precision": OptionInfo("Autocast", "Precision type", gr.Radio, {"choices": ["Autocast", "Full"], "visible": False}),
         "cuda_dtype": OptionInfo("Auto", "Device precision type", gr.Radio, {"choices": ["Auto", "FP32", "FP16", "BF16"]}),
+        "force_dtype": OptionInfo(False, "Force dtype on load", None, None, None),
         "no_half": OptionInfo(False, "Force full precision (--no-half)", None, None, None),
         "upcast_sampling": OptionInfo(False if sys.platform != "darwin" else True, "Upcast sampling", gr.Checkbox, {"visible": False}),