Merge 3e5c3318f2 into 82a973c043

README.md

@@ -3,6 +3,14 @@ A web interface for Stable Diffusion, implemented using Gradio library.
 
 ![](screenshot.png)
 
+## RTX 5090 Compatibility Fork
+
+This fork includes compatibility updates for running with RTX 5090 GPUs and recent PyTorch nightly builds.
+
+- Branch with changes: `feat/rtx5090-compat`
+- Key compatibility notes: optional guards for evolving `ldm`/CompVis APIs and startup robustness improvements
+- Full patch details: `docs/rtx5090-compatibility.md`
+
 ## Features
 [Detailed feature showcase with images](https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Features):
 - Original txt2img and img2img modes

docs/rtx5090-compatibility.md

@@ -0,0 +1,79 @@
+# RTX 5090 Compatibility Patch Notes
+
+## Summary
+This branch adds compatibility guards and dependency fixes so AUTOMATIC1111 WebUI can run with modern RTX 5090 + PyTorch nightly (`cu128`) while using CompVis Stable Diffusion repository layout.
+
+## Code Changes
+
+### 1) Guard missing `ATTENTION_MODES`
+- File: `modules/sd_hijack.py`
+- Change: wrap assignment to `BasicTransformerBlock.ATTENTION_MODES` with `hasattr(...)`.
+- Reason: some `ldm` variants do not define this attribute.
+
+### 2) Guard missing `use_linear` on `SpatialTransformer`
+- File: `modules/sd_hijack_unet.py`
+- Change: replace direct `self.use_linear` access with:
+  - `use_linear = getattr(self, "use_linear", False)`
+- Reason: CompVis transformer blocks may not expose `use_linear`.
+
+### 3) Make depth class import optional
+- File: `modules/processing.py`
+- Change: `LatentDepth2ImageDiffusion` import wrapped in `try/except ImportError`.
+- Change: `isinstance(...)` check now guarded by `LatentDepth2ImageDiffusion is not None`.
+- Reason: class may not exist in all `ldm` implementations.
+
+### 4) Make `AddMiDaS` import optional
+- File: `modules/processing.py`
+- Change: `AddMiDaS` import wrapped in `try/except` with a no-op fallback class.
+- Reason: `ldm.data.util` is absent in some CompVis repo states.
+
+### 5) Make MiDaS module optional
+- File: `modules/sd_models.py`
+- Change: wrap `import ldm.modules.midas as midas` in `try/except`.
+- Change: `enable_midas_autodownload()` returns early when `midas is None`.
+- Reason: prevents startup crash when MiDaS module is unavailable.
+
+### 6) Update default Stable Diffusion upstream for working clone path
+- File: `modules/launch_utils.py`
+- Change: `STABLE_DIFFUSION_REPO` default set to `https://github.com/CompVis/stable-diffusion.git`.
+- Change: default `STABLE_DIFFUSION_COMMIT_HASH` set to `21f890f9da3cfbeaba8e2ac3c425ee9e998d5229`.
+- Reason: previous default URL may be unreachable in current environment.
+
+## Environment / Install Actions
+- Installed GitHub CLI (`gh`) in user path: `~/.local/bin/gh`
+- Forked upstream: `AUTOMATIC1111/stable-diffusion-webui` -> `PiercingXX/stable-diffusion-webui`
+- Cloned into: `/media/Working-Storage/GitHub/stable-diffusion-webui`
+- Created branch: `feat/rtx5090-compat`
+
+Python environment:
+- Python: `3.10.14`
+- Torch: `2.12.0.dev20260310+cu128`
+- CUDA available: `True`
+- GPU detected: `NVIDIA GeForce RTX 5090`
+
+Installed dependencies:
+- `requirements_versions.txt`
+- `clip` (OpenAI CLIP zip, `--no-build-isolation`)
+- `taming-transformers-rom1504`
+- Additional runtime deps used by startup path (`pytorch_lightning`, `kornia`, `diskcache`, etc.)
+
+## Validation Performed
+
+### Runtime checks
+- `torch.cuda.is_available()` -> `True`
+- `torch.cuda.get_device_name(0)` -> `NVIDIA GeForce RTX 5090`
+
+### WebUI smoke startup
+Command used for startup validation:
+- `webui.py --skip-python-version-check --listen --server-name 0.0.0.0 --port 7860 --skip-load-model-at-start --no-download-sd-model`
+
+Observed in logs:
+- `Running on local URL:  http://0.0.0.0:7860`
+- `Startup time: 4.2s ...`
+
+HTTP response verified:
+- `curl http://0.0.0.0:7860/` returned HTML.
+
+## Notes / Remaining Validation
+- Full text-to-image generation requires a model checkpoint to be present or fully downloaded.
+- This branch has been validated for startup/runtime compatibility and serving UI on RTX 5090 stack.

modules/launch_utils.py

@@ -346,13 +346,13 @@ def prepare_environment():
     openclip_package = os.environ.get('OPENCLIP_PACKAGE', "https://github.com/mlfoundations/open_clip/archive/bb6e834e9c70d9c27d0dc3ecedeebeaeb1ffad6b.zip")
 
     assets_repo = os.environ.get('ASSETS_REPO', "https://github.com/AUTOMATIC1111/stable-diffusion-webui-assets.git")
-    stable_diffusion_repo = os.environ.get('STABLE_DIFFUSION_REPO', "https://github.com/Stability-AI/stablediffusion.git")
+    stable_diffusion_repo = os.environ.get('STABLE_DIFFUSION_REPO', "https://github.com/CompVis/stable-diffusion.git")
     stable_diffusion_xl_repo = os.environ.get('STABLE_DIFFUSION_XL_REPO', "https://github.com/Stability-AI/generative-models.git")
     k_diffusion_repo = os.environ.get('K_DIFFUSION_REPO', 'https://github.com/crowsonkb/k-diffusion.git')
     blip_repo = os.environ.get('BLIP_REPO', 'https://github.com/salesforce/BLIP.git')
 
     assets_commit_hash = os.environ.get('ASSETS_COMMIT_HASH', "6f7db241d2f8ba7457bac5ca9753331f0c266917")
-    stable_diffusion_commit_hash = os.environ.get('STABLE_DIFFUSION_COMMIT_HASH', "cf1d67a6fd5ea1aa600c4df58e5b47da45f6bdbf")
+    stable_diffusion_commit_hash = os.environ.get('STABLE_DIFFUSION_COMMIT_HASH', "21f890f9da3cfbeaba8e2ac3c425ee9e998d5229")
     stable_diffusion_xl_commit_hash = os.environ.get('STABLE_DIFFUSION_XL_COMMIT_HASH', "45c443b316737a4ab6e40413d7794a7f5657c19f")
     k_diffusion_commit_hash = os.environ.get('K_DIFFUSION_COMMIT_HASH', "ab527a9a6d347f364e3d185ba6d714e22d80cb3c")
     blip_commit_hash = os.environ.get('BLIP_COMMIT_HASH', "48211a1594f1321b00f14c9f7a5b4813144b2fb9")

modules/processing.py

@@ -28,8 +28,21 @@ import modules.images as images
 import modules.styles
 import modules.sd_models as sd_models
 import modules.sd_vae as sd_vae
-from ldm.data.util import AddMiDaS
-from ldm.models.diffusion.ddpm import LatentDepth2ImageDiffusion
+try:
+    from ldm.data.util import AddMiDaS
+except ImportError:
+    class AddMiDaS:
+        """Fallback no-op depth transform when ldm.data.util is unavailable."""
+
+        def __init__(self, *args, **kwargs):
+            pass
+
+        def __call__(self, sample):
+            return sample
+try:
+    from ldm.models.diffusion.ddpm import LatentDepth2ImageDiffusion
+except ImportError:
+    LatentDepth2ImageDiffusion = None
 
 from einops import repeat, rearrange
 from blendmodes.blend import blendLayers, BlendType
@@ -377,7 +390,7 @@ class StableDiffusionProcessing:
 
         # HACK: Using introspection as the Depth2Image model doesn't appear to uniquely
         # identify itself with a field common to all models. The conditioning_key is also hybrid.
-        if isinstance(self.sd_model, LatentDepth2ImageDiffusion):
+        if LatentDepth2ImageDiffusion is not None and isinstance(self.sd_model, LatentDepth2ImageDiffusion):
             return self.depth2img_image_conditioning(source_image)
 
         if self.sd_model.cond_stage_key == "edit":

modules/sd_hijack.py

@@ -27,7 +27,8 @@ diffusionmodules_model_AttnBlock_forward = ldm.modules.diffusionmodules.model.At
 # new memory efficient cross attention blocks do not support hypernets and we already
 # have memory efficient cross attention anyway, so this disables SD2.0's memory efficient cross attention
 ldm.modules.attention.MemoryEfficientCrossAttention = ldm.modules.attention.CrossAttention
-ldm.modules.attention.BasicTransformerBlock.ATTENTION_MODES["softmax-xformers"] = ldm.modules.attention.CrossAttention
+if hasattr(ldm.modules.attention.BasicTransformerBlock, "ATTENTION_MODES"):
+    ldm.modules.attention.BasicTransformerBlock.ATTENTION_MODES["softmax-xformers"] = ldm.modules.attention.CrossAttention
 
 # silence new console spam from SD2
 ldm.modules.attention.print = shared.ldm_print

modules/sd_hijack_unet.py

@@ -87,17 +87,18 @@ def spatial_transformer_forward(_, self, x: torch.Tensor, context=None):
     b, c, h, w = x.shape
     x_in = x
     x = self.norm(x)
-    if not self.use_linear:
+    use_linear = getattr(self, "use_linear", False)
+    if not use_linear:
         x = self.proj_in(x)
     x = x.permute(0, 2, 3, 1).reshape(b, h * w, c)
-    if self.use_linear:
+    if use_linear:
         x = self.proj_in(x)
     for i, block in enumerate(self.transformer_blocks):
         x = block(x, context=context[i])
-    if self.use_linear:
+    if use_linear:
         x = self.proj_out(x)
     x = x.view(b, h, w, c).permute(0, 3, 1, 2)
-    if not self.use_linear:
+    if not use_linear:
         x = self.proj_out(x)
     return x + x_in
 

modules/sd_models.py

@@ -10,7 +10,10 @@ import re
 import safetensors.torch
 from omegaconf import OmegaConf, ListConfig
 from urllib import request
-import ldm.modules.midas as midas
+try:
+    import ldm.modules.midas as midas
+except ImportError:
+    midas = None
 
 from modules import paths, shared, modelloader, devices, script_callbacks, sd_vae, sd_disable_initialization, errors, hashes, sd_models_config, sd_unet, sd_models_xl, cache, extra_networks, processing, lowvram, sd_hijack, patches
 from modules.timer import Timer
@@ -549,6 +552,9 @@ def enable_midas_autodownload():
     location automatically.
     """
 
+    if midas is None:
+        return
+
     midas_path = os.path.join(paths.models_path, 'midas')
 
     # stable-diffusion-stability-ai hard-codes the midas model path to