SDNQ show a warining for no Triton instead and stop overriding matmul to false

modules/model_quant.py

@@ -98,11 +98,6 @@ def create_sdnq_config(kwargs = None, allow: bool = True, module: str = 'Model',
     from modules import shared
     if allow and (shared.opts.sdnq_quantize_mode in {'pre', 'auto'}) and (module == 'any' or module in shared.opts.sdnq_quantize_weights):
         from modules.sdnq import SDNQConfig
-        from modules.sdnq.common import use_torch_compile as sdnq_use_torch_compile
-
-        if shared.opts.sdnq_use_quantized_matmul and not sdnq_use_torch_compile:
-            log.warning('SDNQ Quantized MatMul requires a working Triton install. Disabling Quantized MatMul.')
-            shared.opts.sdnq_use_quantized_matmul = False
 
         if weights_dtype is None:
             if module in {"TE", "LLM"} and shared.opts.sdnq_quantize_weights_mode_te not in {"Same as model", "default"}:
@@ -354,7 +349,6 @@ def sdnq_quantize_model(model, op=None, sd_model=None, do_gc: bool = True, weigh
     global quant_last_model_name, quant_last_model_device # pylint: disable=global-statement
     from modules import devices, shared, timer
     from modules.sdnq import sdnq_post_load_quant
-    from modules.sdnq.common import use_torch_compile as sdnq_use_torch_compile
 
     if (
         hasattr(model, "quantization_config")
@@ -364,10 +358,6 @@ def sdnq_quantize_model(model, op=None, sd_model=None, do_gc: bool = True, weigh
         log.warning(f'Quantization: Trying to quantize a pre-quantized model. Skipping quantization of module="{op if op is not None else model.__class__}"')
         return model
 
-    if shared.opts.sdnq_use_quantized_matmul and not sdnq_use_torch_compile:
-        log.warning('SDNQ Quantized MatMul requires a working Triton install. Disabling Quantized MatMul.')
-        shared.opts.sdnq_use_quantized_matmul = False
-
     if weights_dtype is None:
         if (op is not None) and ("text_encoder" in op or op in {"TE", "LLM"}) and (shared.opts.sdnq_quantize_weights_mode_te not in {"Same as model", "default"}):
             weights_dtype = shared.opts.sdnq_quantize_weights_mode_te

modules/sd_models.py

@@ -191,10 +191,6 @@ def set_diffuser_options(sd_model, vae=None, op:str='model', offload:bool=True,
     for module_name in get_module_names(sd_model):
         module = getattr(sd_model, module_name, None)
         if hasattr(module, "quantization_config") and getattr(module.quantization_config, "quant_method", None) == "sdnq":
-            from modules.sdnq.common import use_torch_compile as sdnq_use_torch_compile
-            if shared.opts.sdnq_use_quantized_matmul and not sdnq_use_torch_compile:
-                log.warning('SDNQ Quantized MatMul requires a working Triton install. Disabling Quantized MatMul.')
-                shared.opts.sdnq_use_quantized_matmul = False
             if module.quantization_config.use_quantized_matmul != shared.opts.sdnq_use_quantized_matmul:
                 from modules.sdnq.loader import apply_sdnq_options_to_model
                 # log.debug(f'Setting {op} {module_name}: sdnq_use_quantized_matmul={shared.opts.sdnq_use_quantized_matmul}')

modules/sdnq/common.py

@@ -6,7 +6,7 @@ import torch
 
 from modules import shared, devices
 
-sdnq_version = "0.2.0"
+sdnq_version = "0.2.1"
 sdnq_keys = {"weight", "scale", "zero_point", "svd_up", "svd_down"}
 
 torch_version = torch.__version__[:4]

modules/sdnq/loader.py

@@ -3,6 +3,7 @@ import json
 import torch
 from diffusers.models.modeling_utils import ModelMixin
 
+from modules import shared
 from .common import dtype_dict, is_fp8_mm_supported, use_tensorwise_fp8_matmul, check_torch_compile, linear_types
 from .quantizer import QuantizationMethod, SDNQConfig, SDNQQuantizer, sdnq_post_load_quant
 from .quant_utils import prepare_weight_for_matmul, prepare_svd_for_matmul
@@ -288,7 +289,7 @@ def apply_sdnq_options_to_module(model, quantization_config: SDNQConfig, dtype:
 
 def apply_sdnq_options_to_model(model, dtype: torch.dtype | None = None, dequantize_fp32: bool | None = None, use_quantized_matmul: bool | None = None):
     if use_quantized_matmul and not check_torch_compile():
-        raise RuntimeError("SDNQ Quantized MatMul requires a working Triton install.")
+        shared.log.warning("SDNQ: Quantized MatMul requires a working Triton install for best performance.")
     model = apply_sdnq_options_to_module(model, model.quantization_config, dtype=dtype, dequantize_fp32=dequantize_fp32, use_quantized_matmul=use_quantized_matmul)
     if hasattr(model, "quantization_config"):
         if use_quantized_matmul is not None:

modules/sdnq/quantizer.py

@@ -920,7 +920,7 @@ class SDNQConfig(QuantizationConfigMixin):
         Safety checker that arguments are correct
         """
         if self.use_quantized_matmul and not check_torch_compile():
-            raise RuntimeError("SDNQ Quantized MatMul requires a working Triton install.")
+            shared.log.warning("SDNQ: Quantized MatMul requires a working Triton install for best performance.")
         if self.weights_dtype not in accepted_weight_dtypes:
             raise ValueError(f"SDNQ only support weight dtypes in {accepted_weight_dtypes} but found {self.weights_dtype}")
         if self.quantized_matmul_dtype is not None and self.quantized_matmul_dtype not in accepted_matmul_dtypes: