From 8a1743712a1956e3b6e20d890ee6945faade3668 Mon Sep 17 00:00:00 2001
From: Disty0 <disty@disty.xyz>
Date: Wed, 15 Apr 2026 23:28:20 +0300
Subject: [PATCH] SDNQ disable FP8 MM on dynamic quant with unsupported GPUs

---
 modules/sdnq/loader.py    |  2 +-
 modules/sdnq/quantizer.py | 12 +++++++++---
 2 files changed, 10 insertions(+), 4 deletions(-)

diff --git a/modules/sdnq/loader.py b/modules/sdnq/loader.py
index 66fb3f638..8ad8e1530 100644
--- a/modules/sdnq/loader.py
+++ b/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, is_fp8_mm_supported, check_torch_compile, conv_types, linear_types
+from modules import dtype_dict, is_fp8_mm_supported, use_tensorwise_fp8_matmul, check_torch_compile, conv_types, linear_types
 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
diff --git a/modules/sdnq/quantizer.py b/modules/sdnq/quantizer.py
index 4911f4c5b..69bb2149d 100644
--- a/modules/sdnq/quantizer.py
+++ b/modules/sdnq/quantizer.py
@@ -15,7 +15,7 @@ from diffusers.utils import get_module_from_name
 from accelerate import init_empty_weights
 
 from modules import devices, shared
-from .common import sdnq_version, dtype_dict, common_skip_keys, module_skip_keys_dict, accepted_weight_dtypes, accepted_matmul_dtypes, weights_dtype_order, allowed_types, linear_types, embedding_types, conv_types, conv_transpose_types, compile_func, use_tensorwise_fp8_matmul, use_contiguous_mm, check_torch_compile
+from .common import sdnq_version, dtype_dict, common_skip_keys, module_skip_keys_dict, accepted_weight_dtypes, accepted_matmul_dtypes, weights_dtype_order, allowed_types, linear_types, embedding_types, conv_types, conv_transpose_types, compile_func, is_fp8_mm_supported, use_tensorwise_fp8_matmul, use_contiguous_mm, check_torch_compile
 from .dequantizer import SDNQDequantizer, dequantize_sdnq_model
 from .packed_int import pack_int
 from .packed_float import pack_float
@@ -437,10 +437,16 @@ def sdnq_quantize_layer_weight_dynamic(weight, layer_class_name=None, weights_dt
     quantization_loss = None
     svd_is_transposed = False
     for i in range(weights_dtype_order.index(weights_dtype), len(weights_dtype_order)):
+        current_weights_dtype = weights_dtype_order[i]
+        if quantized_matmul_dtype is None and not is_fp8_mm_supported and not dtype_dict[current_weights_dtype]["is_integer"] and dtype_dict[current_weights_dtype]["num_bits"] < 16:
+            current_use_quantized_matmul = False
+        else:
+            current_use_quantized_matmul = use_quantized_matmul
+
         quantized_weight, scale, zero_point, _, _, sdnq_dequantizer = sdnq_quantize_layer_weight(
             svd_weight,
             layer_class_name=layer_class_name,
-            weights_dtype=weights_dtype_order[i],
+            weights_dtype=current_weights_dtype,
             quantized_matmul_dtype=quantized_matmul_dtype,
             torch_dtype=torch_dtype,
             group_size=group_size,
@@ -448,7 +454,7 @@ def sdnq_quantize_layer_weight_dynamic(weight, layer_class_name=None, weights_dt
             svd_steps=svd_steps,
             use_svd=False,
             using_pre_calculated_svd=use_svd,
-            use_quantized_matmul=use_quantized_matmul,
+            use_quantized_matmul=current_use_quantized_matmul,
             use_stochastic_rounding=use_stochastic_rounding,
             dequantize_fp32=dequantize_fp32,
             param_name=param_name,