diff --git a/CHANGELOG.md b/CHANGELOG.md
index cf6acaff2..fb797f9e3 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -3,7 +3,9 @@
 ## Update for 2026-06-20
 
 - **Features**
-  - **SDNQ** support for NPU during quantization and inference
+  - **SDNQ** support for NPU during quantization and inference  
+  - add option: force dtype on load  
+    use to force model components to override loading with desired dtype regardless of component config  
 - **Fixes**
   - `sdnq` warn instead of error for triton
   - `insightface` missing dependencies
diff --git a/data/reference-quant.json b/data/reference-quant.json
index a12cc7f95..596959ee8 100644
--- a/data/reference-quant.json
+++ b/data/reference-quant.json
@@ -129,16 +129,6 @@
     "size": 17.27,
     "date": "2025 December"
   },
-  "Qwen-Image-2512 sdnq-dynamic-uint4": {
-    "path": "Disty0/Qwen-Image-2512-SDNQ-4bit-dynamic",
-    "preview": "Disty0--Qwen-Image-2512-SDNQ-uint4-svd-r32.jpg",
-    "desc": "Quantization of Qwen/Qwen-Image-2512 using SDNQ: sdnq-dynamic 4-bit uint",
-    "skip": true,
-    "tags": "quantized",
-    "extras": "",
-    "size": 18.51,
-    "date": "2026 January"
-  },
   "Qwen-Image-Edit sdnq-svd-uint4": {
     "path": "Disty0/Qwen-Image-Edit-SDNQ-uint4-svd-r32",
     "preview": "Qwen--Qwen-Image-Edit.jpg",
diff --git a/modules/devices.py b/modules/devices.py
index 438c12028..43dc25888 100644
--- a/modules/devices.py
+++ b/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:
diff --git a/modules/sd_hijack_te.py b/modules/sd_hijack_te.py
index c1163eabb..172b3f863 100644
--- a/modules/sd_hijack_te.py
+++ b/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
 
 
diff --git a/modules/ui_definitions.py b/modules/ui_definitions.py
index 679bcb4fc..27f29970a 100644
--- a/modules/ui_definitions.py
+++ b/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}),
 
diff --git a/pipelines/generic_text_encoder.py b/pipelines/generic_text_encoder.py
index 2bb94e58d..8da8039d5 100644
--- a/pipelines/generic_text_encoder.py
+++ b/pipelines/generic_text_encoder.py
@@ -1,5 +1,6 @@
 import os
 import json
+import torch
 import transformers
 from modules import shared, devices, errors, sd_models, sd_offload, model_quant
 from modules.logger import log
@@ -147,4 +148,14 @@ def load_text_encoder(repo_id, cls_name, load_config=None, subfolder="text_encod
         module_memory = sd_offload.get_module_memory(text_encoder)
         log.debug(f'Load model: text_encoder="{repo_id}" quant="{quant_type}" size={module_size:.3f} params={param_num:.3f} memory={module_memory}')
 
+    try:
+        actual_dtype = text_encoder.dtype
+        if isinstance(actual_dtype, torch.dtype) and isinstance(dtype, torch.dtype) and actual_dtype != dtype:
+            force = shared.opts.force_dtype
+            log.warning(f'Load model: text_encoder="{repo_id}" dtype desired={dtype} actual={actual_dtype} force={force}')
+            if force:
+                text_encoder = text_encoder.to(dtype)
+    except Exception:
+        pass
+
     return text_encoder
diff --git a/pipelines/generic_transformer.py b/pipelines/generic_transformer.py
index f7ee42c8d..8cc177cb8 100644
--- a/pipelines/generic_transformer.py
+++ b/pipelines/generic_transformer.py
@@ -1,4 +1,5 @@
 import os
+import torch
 from modules import shared, devices, errors, sd_models, sd_offload, model_quant
 from modules.logger import log
 from pipelines.generic_util import get_loader
@@ -133,4 +134,14 @@ def load_transformer(repo_id, cls_name, load_config=None, subfolder="transformer
         module_memory = sd_offload.get_module_memory(transformer)
         log.debug(f'Load model: transformer="{repo_id}" quant="{quant_type}" size={module_size:.3f} params={param_num:.3f} memory={module_memory}')
 
+    try:
+        actual_dtype = transformer.dtype
+        if isinstance(actual_dtype, torch.dtype) and isinstance(dtype, torch.dtype) and actual_dtype != dtype:
+            force = shared.opts.force_dtype
+            log.warning(f'Load model: transformer="{repo_id}" dtype desired={dtype} actual={actual_dtype} force={force}')
+            if force:
+                transformer = transformer.to(dtype)
+    except Exception:
+        pass
+
     return transformer