remove sdnq_dequantize_compile option

- advanced users should use the env var SDNQ_USE_TORCH_COMPILE instead

cli/sdnq-attention-benchmark.py

@@ -872,7 +872,8 @@ def probe_compiled_prep():
     # host c++ compiler (msvc on windows) fails every sdnq attention call at generation.
     # must run before any other sdnq_triton_atten call: a prior eager run masks the
     # cold-start failure the webui hits
-    if not shared.opts.sdnq_dequantize_compile:
+    from sdnq.common import use_torch_compile
+    if use_torch_compile:
         return "disabled", None
     q, k, v = make_qkv(1, 2, 256, 64, structured=False)
     with console.status("probing compiled input prep (compiles on first run, cached afterwards)"):

modules/devices.py

@@ -456,8 +456,6 @@ def test_triton(early: bool = False):
     t1 = time.time()
     fn = f'{sys._getframe(2).f_code.co_name}:{sys._getframe(1).f_code.co_name}' # pylint: disable=protected-access
     log.debug(f'Triton: pass={triton_ok} version={triton_version} fn={fn} time={t1-t0:.2f}')
-    if not triton_ok and opts is not None:
-        opts.sdnq_dequantize_compile = False
     return triton_ok
 
 

modules/processing.py

@@ -420,13 +420,12 @@ def print_stats():
     log.debug(f'Processed: timers={timer.process.dct()}')
     log.debug(f'Processed: memory={memstats.memory_stats()}')
 
-    if shared.opts.sdnq_dequantize_compile:
+    if devices.triton_ok:
         from modules.timer_sdnq import update_sdnq_attention_timers
         update_sdnq_attention_timers()
         if timer.autotune.get_total() > 0.1:
             log.debug(f'Processed: autotune={timer.autotune.dct(min_time=0)}')
 
-    if devices.triton_ok:
         from modules.sd_models_compile import update_compile_times
         update_compile_times()
         dynamo_dct = timer.dynamo.dct(min_time=2.0, no_total=True)

modules/ui_definitions.py

@@ -189,7 +189,6 @@ def create_settings(cmd_opts):
         "sdnq_use_dynamic_quantization": OptionInfo(False, "Use Dynamic quantization", gr.Checkbox),
         "sdnq_quantize_conv_layers": OptionInfo(False, "Quantize convolutional layers", gr.Checkbox),
         "sdnq_quantize_embedding_layers": OptionInfo(False, "Quantize embedding layers", gr.Checkbox),
-        "sdnq_dequantize_compile": OptionInfo(devices.has_triton(early=True), "Dequantize using torch.compile", gr.Checkbox),
         "sdnq_use_quantized_matmul_conv": OptionInfo(False, "Use quantized MatMul with conv", gr.Checkbox),
         "sdnq_quantize_with_gpu": OptionInfo(True, "Quantize using GPU", gr.Checkbox),
         "sdnq_dequantize_fp32": OptionInfo(True, "Dequantize using full precision", gr.Checkbox),