Merge pull request #4579 from vladmandic/dev

merge dev

CHANGELOG.md

@@ -1,5 +1,14 @@
 # Change Log for SD.Next
 
+## Update for 2026-01-21
+
+- add `SD_DEVICE_DEBUG` env variable to trace rocm/xpu/directml init failures  
+- fix lora load when using peft/diffusers loader  
+- fix rocm hipblaslt detection  
+- fix image delete  
+- fix `all_seeds` error  
+- improve `wrap_gradio` error handling  
+
 ## Update for 2026-01-20
 
 ### Highlights for 2026-01-20

modules/call_queue.py

@@ -37,7 +37,6 @@ def wrap_gradio_gpu_call(func, extra_outputs=None, name=None):
             id_task = None
         with get_lock():
             progress.start_task(id_task)
-            res = [None, '', '', '']
             try:
                 res = func(*args, **kwargs)
                 progress.record_results(id_task, res)
@@ -45,7 +44,8 @@ def wrap_gradio_gpu_call(func, extra_outputs=None, name=None):
                 shared.log.error(f"Exception: {e}")
                 shared.log.error(f"Arguments: args={str(args)[:10240]} kwargs={str(kwargs)[:10240]}")
                 errors.display(e, 'gradio call')
-                res[-1] = f"<div class='error'>{html.escape(str(e))}</div>"
+                res = extra_outputs or []
+                res.append(f"<div class='error'>{html.escape(str(e))}</div>")
             finally:
                 progress.finish_task(id_task)
         return res
@@ -70,7 +70,8 @@ def wrap_gradio_call(func, extra_outputs=None, add_stats=False, name=None):
             if res is None:
                 msg = "No result returned from function"
                 shared.log.warning(msg)
-                res = [None, '', '', f"<div class='error'>{html.escape(msg)}</div>"]
+                res = extra_outputs_array or []
+                res.append(f"<div class='error'>{html.escape(msg)}</div>")
             else:
                 res = list(res)
             if shared.cmd_opts.profile:
@@ -78,9 +79,8 @@ def wrap_gradio_call(func, extra_outputs=None, add_stats=False, name=None):
                 errors.profile(pr, 'Wrap')
         except Exception as e:
             errors.display(e, 'gradio call')
-            if extra_outputs_array is None:
-                extra_outputs_array = [None, '']
-            res = extra_outputs_array + [f"<div class='error'>{html.escape(type(e).__name__+': '+str(e))}</div>"]
+            res = extra_outputs_array or []
+            res.append(f"<div class='error'>{html.escape(type(e).__name__+': '+str(e))}</div>")
         shared.state.end(jobid)
         if not add_stats:
             return tuple(res)

modules/lora/extra_networks_lora.py

@@ -173,7 +173,7 @@ class ExtraNetworkLora(extra_networks.ExtraNetwork):
     def signature(self, names: List[str], te_multipliers: List, unet_multipliers: List):
         return [f'{name}:{te}:{unet}' for name, te, unet in zip(names, te_multipliers, unet_multipliers)]
 
-    def changed(self, requested: List[str], include: List[str], exclude: List[str]):
+    def changed(self, requested: List[str], include: List[str] = None, exclude: List[str] = None) -> bool:
         if shared.opts.lora_force_reload:
             return True
         sd_model = shared.sd_model.pipe if hasattr(shared.sd_model, 'pipe') else shared.sd_model
@@ -213,15 +213,17 @@ class ExtraNetworkLora(extra_networks.ExtraNetwork):
             debug_log(f'Network load: type=LoRA include={include} exclude={exclude} method={load_method} requested={requested} fn={fn}')
 
         if load_method == 'diffusers':
-            has_changed = False # diffusers handles its own loading
-            if len(exclude) == 0:
+            has_changed = self.changed(requested)
+            if has_changed:
                 jobid = shared.state.begin('LoRA')
                 lora_load.network_load(names, te_multipliers, unet_multipliers, dyn_dims, lora_modules) # load only on first call
                 sd_models.set_diffuser_offload(shared.sd_model, op="model")
                 shared.state.end(jobid)
+
         elif load_method == 'nunchaku':
             from modules.lora import lora_nunchaku
             has_changed = lora_nunchaku.load_nunchaku(names, unet_multipliers)
+
         else: # native
             lora_load.network_load(names, te_multipliers, unet_multipliers, dyn_dims) # load
             has_changed = self.changed(requested, include, exclude)

modules/lora/lora_load.py

@@ -304,7 +304,7 @@ def network_load(names, te_multipliers=None, unet_multipliers=None, dyn_dims=Non
             shared.log.error(f'Network load: type=LoRA action=fuse {str(e)}')
             if l.debug:
                 errors.display(e, 'LoRA')
-        shared.sd_model = sd_models.apply_balanced_offload(shared.sd_model, force=True) # some layers may end up on cpu without hook
+        shared.sd_model = sd_models.apply_balanced_offload(shared.sd_model, force=True, silent=True) # some layers may end up on cpu without hook
 
     if len(l.loaded_networks) > 0 and l.debug:
         shared.log.debug(f'Network load: type=LoRA loaded={[n.name for n in l.loaded_networks]} cache={list(lora_cache)} fuse={shared.opts.lora_fuse_native}:{shared.opts.lora_fuse_diffusers}')

modules/processing_helpers.py

@@ -227,7 +227,7 @@ def decode_first_stage(model, x):
 
 
 def get_fixed_seed(seed):
-    if seed is None or seed == '' or seed == -1:
+    if (seed is None) or (seed == '') or (seed == -1):
         random.seed()
         seed = int(random.randrange(4294967294))
     return seed
@@ -236,10 +236,16 @@ def get_fixed_seed(seed):
 def fix_seed(p):
     p.seed = get_fixed_seed(p.seed)
     p.subseed = get_fixed_seed(p.subseed)
-    for i in range(len(p.all_seeds)):
-        p.all_seeds[i] = get_fixed_seed(p.all_seeds[i])
-    for i in range(len(p.all_subseeds)):
-        p.all_subseeds[i] = get_fixed_seed(p.all_subseeds[i])
+    if p.all_seeds is None or len(p.all_seeds) == 0:
+        p.all_seeds = [p.seed]
+    else:
+        for i in range(len(p.all_seeds)):
+            p.all_seeds[i] = get_fixed_seed(p.all_seeds[i])
+    if p.all_subseeds is None or len(p.all_subseeds) == 0:
+        p.all_subseeds = [p.subseed]
+    else:
+        for i in range(len(p.all_subseeds)):
+            p.all_subseeds[i] = get_fixed_seed(p.all_subseeds[i])
 
 
 def old_hires_fix_first_pass_dimensions(width, height):

modules/rocm.py

@@ -110,7 +110,7 @@ class Agent:
         else:
             self.arch = MicroArchitecture.GCN
         self.is_apu = (self.gfx_version & 0xFFF0 == 0x1150) or self.gfx_version in (0x801, 0x902, 0x90c, 0x1013, 0x1033, 0x1035, 0x1036, 0x1103,)
-        self.blaslt_supported = False if blaslt_tensile_libpath is None else os.path.exists(os.path.join(blaslt_tensile_libpath, f"Kernels.so-000-{self.name}.hsaco" if sys.platform == "win32" else f"extop_{self.name}.co"))
+        self.blaslt_supported = False if blaslt_tensile_libpath is None else os.path.exists(os.path.join(blaslt_tensile_libpath, f"TensileLibrary_lazy_{self.name}.dat"))
 
     def __str__(self) -> str:
         return self.name
@@ -138,6 +138,8 @@ class Agent:
         return None
 
     def get_gfx_version(self) -> Union[str, None]:
+        if self.gfx_version is None:
+            return None
         if self.gfx_version >= 0x1100 and self.gfx_version < 0x1200:
             return "11.0.0"
         elif self.gfx_version != 0x1030 and self.gfx_version >= 0x1000 and self.gfx_version < 0x1100:
@@ -293,12 +295,14 @@ if sys.platform == "win32":
 
             build_targets = torch.cuda.get_arch_list()
             agents = get_agents()
+            log.debug(f'ROCm: agents={agents}')
             if all(available.name not in build_targets for available in agents):
-                log.warning('ROCm: torch-rocm is installed, but none of build targets is available')
+                log.warning('ROCm: torch-rocm is installed, but none of build targets are available')
                 # use cpu instead of crashing
                 torch.cuda.is_available = lambda: False
 
             agent = get_hip_agent()
+            log.debug(f'ROCm: selected={agents}')
             if not agent.blaslt_supported:
                 log.warning(f'ROCm: hipBLASLt unavailable agent={agent}')
             if (agent.gfx_version & 0xFFF0) == 0x1200:

modules/rocm_triton_windows.py

@@ -59,7 +59,7 @@ if sys.platform == "win32":
         return zluda.core.to_hip_stream(_cuda_getCurrentRawStream(device))
 
     def get_default_agent() -> Union[Agent, None]:
-        if shared.devices.backend == "rocm":
+        if shared.devices.has_rocm():
             return devices.get_hip_agent()
         else:
             from modules import zluda

modules/sd_offload.py

@@ -466,7 +466,7 @@ def report_model_stats(module_name, module):
         shared.log.error(f'Module stats: name={module_name} {e}')
 
 
-def apply_balanced_offload(sd_model=None, exclude=None, force=False):
+def apply_balanced_offload(sd_model=None, exclude:list[str]=None, force:bool=False, silent:bool=False):
     global offload_hook_instance # pylint: disable=global-statement
     if shared.opts.diffusers_offload_mode != "balanced":
         return sd_model
@@ -499,7 +499,8 @@ def apply_balanced_offload(sd_model=None, exclude=None, force=False):
             module.module_name = module_name
             module.offload_dir = os.path.join(shared.opts.accelerate_offload_path, checkpoint_name, module_name)
             apply_balanced_offload_to_module(module, op='apply')
-            report_model_stats(module_name, module)
+            if not silent:
+                report_model_stats(module_name, module)
 
     set_accelerate(sd_model)
     t = time.time() - t0

modules/shared.py

@@ -87,16 +87,22 @@ elif cmd_opts.use_ipex or devices.has_xpu():
     ok, e = ipex_init()
     if not ok:
         log.error(f'IPEX initialization failed: {e}')
+        if os.environ.get('SD_DEVICE_DEBUG', None) is not None:
+            errors.display(e, 'IPEX')
 elif cmd_opts.use_directml:
     from modules.dml import directml_init
     ok, e = directml_init()
     if not ok:
         log.error(f'DirectML initialization failed: {e}')
+        if os.environ.get('SD_DEVICE_DEBUG', None) is not None:
+            errors.display(e, 'DirectML')
 elif cmd_opts.use_rocm or devices.has_rocm():
     from modules.rocm import rocm_init
     ok, e = rocm_init()
     if not ok:
         log.error(f'ROCm initialization failed: {e}')
+        if os.environ.get('SD_DEVICE_DEBUG', None) is not None:
+            errors.display(e, 'ROCm')
 devices.backend = devices.get_backend(cmd_opts)
 devices.device = devices.get_optimal_device()
 mem_stat = memory_stats()

modules/ui_common.py

@@ -88,7 +88,7 @@ def delete_files(js_data, files, all_files, index):
                 continue
             if os.path.exists(fn) and os.path.isfile(fn):
                 deleted.append(fn)
-                # os.remove(fn)
+                os.remove(fn)
                 if fn in all_files:
                     all_files.remove(fn)
                     shared.log.info(f'Delete: image="{fn}"')

modules/video_models/video_run.py

@@ -11,6 +11,7 @@ debug = shared.log.trace if os.environ.get('SD_VIDEO_DEBUG', None) is not None e
 
 def generate(*args, **kwargs):
     task_id, ui_state, engine, model, prompt, negative, styles, width, height, frames, steps, sampler_index, sampler_shift, dynamic_shift, seed, guidance_scale, guidance_true, init_image, init_strength, last_image, vae_type, vae_tile_frames, mp4_fps, mp4_interpolate, mp4_codec, mp4_ext, mp4_opt, mp4_video, mp4_frames, mp4_sf, vlm_enhance, vlm_model, vlm_system_prompt, override_settings = args
+
     if engine is None or model is None or engine == 'None' or model == 'None':
         return video_utils.queue_err('model not selected')
     # videojob = shared.state.begin('Video')

modules/video_models/video_ui.py

@@ -190,7 +190,7 @@ def create_ui(prompt, negative, styles, overrides, init_image, init_strength, la
     ]
 
     video_dict = dict(
-        fn=call_queue.wrap_gradio_gpu_call(video_run.generate, extra_outputs=[None, '', ''], name='Video'),
+        fn=call_queue.wrap_gradio_gpu_call(video_run.generate, extra_outputs=[gr.update(), gr.update(), gr.update(), gr.update()], name='Video'),
         _js="submit_video",
         inputs=state_inputs + video_inputs,
         outputs=video_outputs,