Warn and disable quantized matmul if triton is not available

modules/model_quant.py

@@ -166,6 +166,11 @@ 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:
+            shared.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"}:
@@ -492,6 +497,11 @@ 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 shared.opts.sdnq_use_quantized_matmul and not sdnq_use_torch_compile:
+        shared.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"}):

modules/sd_models.py

@@ -167,6 +167,10 @@ 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:
+                shared.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
                 shared.log.debug(f'Setting {op} {module_name}: sdnq_use_quantized_matmul={shared.opts.sdnq_use_quantized_matmul}')

modules/sdnq/common.py

@@ -5,7 +5,7 @@ import torch
 
 from modules import shared, devices
 
-sdnq_version = "0.1.0"
+sdnq_version = "0.1.1"
 
 dtype_dict = {
     "int32": {"min": -2147483648, "max": 2147483647, "num_bits": 32, "sign": 1, "exponent": 0, "mantissa": 31, "target_dtype": torch.int32, "torch_dtype": torch.int32, "storage_dtype": torch.int32, "is_unsigned": False, "is_integer": True, "is_packed": False},

modules/sdnq/loader.py

@@ -3,7 +3,7 @@ import json
 import torch
 from diffusers.models.modeling_utils import ModelMixin
 
-from .common import dtype_dict, use_tensorwise_fp8_matmul
+from .common import dtype_dict, use_tensorwise_fp8_matmul, use_torch_compile
 from .quantizer import SDNQConfig, sdnq_post_load_quant, prepare_weight_for_matmul, prepare_svd_for_matmul, get_quant_args_from_config
 from .forward import get_forward_func
 from .file_loader import load_files
@@ -206,6 +206,8 @@ def apply_sdnq_options_to_module(model, dtype: torch.dtype = None, dequantize_fp
 
 
 def apply_sdnq_options_to_model(model, dtype: torch.dtype = None, dequantize_fp32: bool = None, use_quantized_matmul: bool = None):
+    if use_quantized_matmul and not use_torch_compile:
+        raise ValueError("SDNQ Quantized MatMul requires a working Triton install.")
     model = apply_sdnq_options_to_module(model, 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

@@ -909,7 +909,7 @@ class SDNQConfig(QuantizationConfigMixin):
         Safety checker that arguments are correct
         """
         if self.use_quantized_matmul and not use_torch_compile:
-            raise ValueError("Quantized MatMul requires a working Triton install.")
+            raise ValueError("SDNQ Quantized MatMul requires a working Triton install.")
         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: