From c0c329a2f111ea829862ceba9c598e4cadba202a Mon Sep 17 00:00:00 2001
From: yesMohanHere <mohanasriram.e@gmail.com>
Date: Fri, 17 Jul 2026 15:53:36 +0530
Subject: [PATCH 01/10] Add support for NVIDIA Blackwell GPUs (sm_120, RTX
 50-series)

- Bump gsplat to >=1.5.3 (sm_120-capable release line) and nerfacc to >=0.5.3
- Drop EOL Python 3.8/3.9: requires-python >=3.10, CI publish/doc workflows
- torch.load(..., weights_only=False) for torch>=2.6 checkpoint compatibility
- Use torch.amp custom_fwd/custom_bwd with device_type="cuda" (torch>=2.4 API)
- Replace deprecated PIL private encoder API in pil_to_numpy with np.asarray
  (fixes Pillow>=11 compatibility)
- Docs: add Torch 2.8+ / CUDA 12.8 install tab for Blackwell GPUs

Tested on RTX 5060 (sm_120) with torch 2.11.0+cu128: splatfacto and
nerfacto train end-to-end and ns-render produces correct output.
---
 .github/workflows/doc.yml                     |  2 +-
 .github/workflows/publish.yml                 |  2 +-
 docs/quickstart/installation.md               | 28 +++++++++++++++++--
 nerfstudio/data/utils/data_utils.py           | 19 +------------
 nerfstudio/engine/trainer.py                  |  4 +--
 nerfstudio/field_components/activations.py    |  6 ++--
 nerfstudio/scripts/downloads/download_data.py |  2 +-
 nerfstudio/utils/eval_utils.py                |  2 +-
 pyproject.toml                                |  6 ++--
 9 files changed, 38 insertions(+), 33 deletions(-)

diff --git a/.github/workflows/doc.yml b/.github/workflows/doc.yml
index d2637429ca..3974bd4cac 100644
--- a/.github/workflows/doc.yml
+++ b/.github/workflows/doc.yml
@@ -15,7 +15,7 @@ jobs:
       - uses: actions/checkout@v3
       - uses: actions/setup-python@v3
         with:
-          python-version: '3.9'
+          python-version: '3.10'
       - name: Install dependencies
         run: |
           pip install uv
diff --git a/.github/workflows/publish.yml b/.github/workflows/publish.yml
index b75984535a..e9039a2aa6 100644
--- a/.github/workflows/publish.yml
+++ b/.github/workflows/publish.yml
@@ -16,7 +16,7 @@ jobs:
       - name: Set up Python
         uses: actions/setup-python@v1
         with:
-          python-version: '3.8'
+          python-version: '3.10'
       - name: Install dependencies
         run: |
           python -m pip install build twine
diff --git a/docs/quickstart/installation.md b/docs/quickstart/installation.md
index de79064213..c4e79a19fc 100644
--- a/docs/quickstart/installation.md
+++ b/docs/quickstart/installation.md
@@ -5,7 +5,7 @@
 ::::::{tab-set}
 :::::{tab-item} Linux
 
-Nerfstudio requires `python >= 3.8`. We recommend using conda to manage dependencies. Make sure to install [Conda](https://docs.conda.io/en/latest/miniconda.html) before proceeding.
+Nerfstudio requires `python >= 3.10`. We recommend using conda to manage dependencies. Make sure to install [Conda](https://docs.conda.io/en/latest/miniconda.html) before proceeding.
 
 :::::
 :::::{tab-item} Windows
@@ -51,7 +51,7 @@ For example:
 When updating, or if you close your terminal before you finish the installation and run your first `splatfacto`, you have to re-do this environment activation step.
 :::
 
-Nerfstudio requires `python >= 3.8`. We recommend using conda to manage dependencies. Make sure to install [Conda](https://docs.conda.io/en/latest/miniconda.html) before proceeding.
+Nerfstudio requires `python >= 3.10`. We recommend using conda to manage dependencies. Make sure to install [Conda](https://docs.conda.io/en/latest/miniconda.html) before proceeding.
 
 :::::
 ::::::
@@ -59,7 +59,7 @@ Nerfstudio requires `python >= 3.8`. We recommend using conda to manage dependen
 ## Create environment
 
 ```bash
-conda create --name nerfstudio -y python=3.8
+conda create --name nerfstudio -y python=3.10
 conda activate nerfstudio
 python -m pip install --upgrade pip
 
@@ -79,6 +79,28 @@ pip uninstall torch torchvision functorch tinycudann
 ```
 
 ::::{tab-set}
+:::{tab-item} Torch 2.8+ with CUDA 12.8 (Blackwell / RTX 50-series)
+
+Required for NVIDIA Blackwell GPUs (RTX 50-series, compute capability sm_120), which need CUDA 12.8 or newer:
+
+```bash
+pip install torch torchvision --index-url https://download.pytorch.org/whl/cu128
+```
+
+To build the necessary CUDA extensions, `cuda-toolkit` 12.8 is also required. We
+recommend installing with conda:
+
+```bash
+conda install -c "nvidia/label/cuda-12.8.0" cuda-toolkit
+```
+
+When building CUDA extensions (gsplat, tiny-cuda-nn, ...) for a Blackwell GPU, make sure the architecture list includes sm_120:
+
+```bash
+export TORCH_CUDA_ARCH_LIST="12.0+PTX"
+```
+
+:::
 :::{tab-item} Torch 2.1.2 with CUDA 11.8 (recommended)
 
 Install PyTorch 2.1.2 with CUDA 11.8:
diff --git a/nerfstudio/data/utils/data_utils.py b/nerfstudio/data/utils/data_utils.py
index d1a6721844..109c848cad 100644
--- a/nerfstudio/data/utils/data_utils.py
+++ b/nerfstudio/data/utils/data_utils.py
@@ -36,24 +36,7 @@ def pil_to_numpy(im: PILImage) -> np.ndarray:
     # Load in image completely (PIL defaults to lazy loading)
     im.load()
 
-    # Unpack data
-    e = Image._getencoder(im.mode, "raw", im.mode)
-    e.setimage(im.im)
-
-    # NumPy buffer for the result
-    shape, typestr = Image._conv_type_shape(im)
-    data = np.empty(shape, dtype=np.dtype(typestr))
-    mem = data.data.cast("B", (data.data.nbytes,))
-
-    bufsize, s, offset = 65536, 0, 0
-    while not s:
-        _, s, d = e.encode(bufsize)
-        mem[offset : offset + len(d)] = d
-        offset += len(d)
-    if s < 0:
-        raise RuntimeError("encoder error %d in tobytes" % s)
-
-    return data
+    return np.asarray(im)
 
 
 def get_image_mask_tensor_from_path(filepath: Union[Path, IO[bytes]], scale_factor: float = 1.0) -> torch.Tensor:
diff --git a/nerfstudio/engine/trainer.py b/nerfstudio/engine/trainer.py
index a653e1de8d..7f2f0d34d0 100644
--- a/nerfstudio/engine/trainer.py
+++ b/nerfstudio/engine/trainer.py
@@ -429,7 +429,7 @@ def _load_checkpoint(self) -> None:
                 load_step = sorted(int(x[x.find("-") + 1 : x.find(".")]) for x in os.listdir(load_dir))[-1]
             load_path: Path = load_dir / f"step-{load_step:09d}.ckpt"
             assert load_path.exists(), f"Checkpoint {load_path} does not exist"
-            loaded_state = torch.load(load_path, map_location="cpu")
+            loaded_state = torch.load(load_path, map_location="cpu", weights_only=False)
             self._start_step = loaded_state["step"] + 1
             # load the checkpoints for pipeline, optimizers, and gradient scalar
             self.pipeline.load_pipeline(loaded_state["pipeline"], loaded_state["step"])
@@ -440,7 +440,7 @@ def _load_checkpoint(self) -> None:
             CONSOLE.print(f"Done loading Nerfstudio checkpoint from {load_path}")
         elif load_checkpoint is not None:
             assert load_checkpoint.exists(), f"Checkpoint {load_checkpoint} does not exist"
-            loaded_state = torch.load(load_checkpoint, map_location="cpu")
+            loaded_state = torch.load(load_checkpoint, map_location="cpu", weights_only=False)
             self._start_step = loaded_state["step"] + 1
             # load the checkpoints for pipeline, optimizers, and gradient scalar
             self.pipeline.load_pipeline(loaded_state["pipeline"], loaded_state["step"])
diff --git a/nerfstudio/field_components/activations.py b/nerfstudio/field_components/activations.py
index e413965e24..3719990949 100644
--- a/nerfstudio/field_components/activations.py
+++ b/nerfstudio/field_components/activations.py
@@ -21,21 +21,21 @@
 import torch
 from jaxtyping import Float
 from torch import Tensor
+from torch.amp import custom_bwd, custom_fwd
 from torch.autograd import Function
-from torch.cuda.amp import custom_bwd, custom_fwd
 
 
 class _TruncExp(Function):
     # Implementation from torch-ngp:
     # https://github.com/ashawkey/torch-ngp/blob/93b08a0d4ec1cc6e69d85df7f0acdfb99603b628/activation.py
     @staticmethod
-    @custom_fwd(cast_inputs=torch.float32)
+    @custom_fwd(cast_inputs=torch.float32, device_type="cuda")
     def forward(ctx, x):
         ctx.save_for_backward(x)
         return torch.exp(x)
 
     @staticmethod
-    @custom_bwd
+    @custom_bwd(device_type="cuda")
     def backward(ctx, g):
         x = ctx.saved_tensors[0]
         return g * torch.exp(x.clamp(-15, 15))
diff --git a/nerfstudio/scripts/downloads/download_data.py b/nerfstudio/scripts/downloads/download_data.py
index a2c9d27584..653fbf94d0 100644
--- a/nerfstudio/scripts/downloads/download_data.py
+++ b/nerfstudio/scripts/downloads/download_data.py
@@ -514,7 +514,7 @@ def download(self, save_dir: Path) -> None:
             split_filepaths = []
             for image_path, new_image_path in copied_images.items():
                 metadata_path = image_path.parent.parent / "metadata" / f"{image_path.stem}.pt"
-                metadata = torch.load(metadata_path, map_location="cpu")
+                metadata = torch.load(metadata_path, map_location="cpu", weights_only=False)
                 c2w = torch.eye(4)
                 c2w[:3] = metadata["c2w"]
                 file_path = str(Path("images") / f"{new_image_path.name}")
diff --git a/nerfstudio/utils/eval_utils.py b/nerfstudio/utils/eval_utils.py
index 11a8b23416..8d07fe842a 100644
--- a/nerfstudio/utils/eval_utils.py
+++ b/nerfstudio/utils/eval_utils.py
@@ -59,7 +59,7 @@ def eval_load_checkpoint(config: TrainerConfig, pipeline: Pipeline) -> Tuple[Pat
         load_step = config.load_step
     load_path = config.load_dir / f"step-{load_step:09d}.ckpt"
     assert load_path.exists(), f"Checkpoint {load_path} does not exist"
-    loaded_state = torch.load(load_path, map_location="cpu")
+    loaded_state = torch.load(load_path, map_location="cpu", weights_only=False)
     pipeline.load_pipeline(loaded_state["pipeline"], loaded_state["step"])
     CONSOLE.print(f":white_check_mark: Done loading checkpoint from {load_path}")
     return load_path, load_step
diff --git a/pyproject.toml b/pyproject.toml
index 585eee4f9c..fd1b60bb76 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -8,7 +8,7 @@ version = "1.1.5"
 description = "All-in-one repository for state-of-the-art NeRFs"
 readme = "README.md"
 license = { text="Apache 2.0"}
-requires-python = ">=3.8.0"
+requires-python = ">=3.10"
 classifiers = [
     "Development Status :: 3 - Alpha",
     "Programming Language :: Python",
@@ -31,7 +31,7 @@ dependencies = [
     "mediapy>=1.1.0",
     "msgpack>=1.0.4",
     "msgpack_numpy>=0.4.8",
-    "nerfacc==0.5.2",
+    "nerfacc>=0.5.3",
     "open3d>=0.16.0",
     "opencv-python-headless==4.10.0.84",
     "Pillow>=10.3.0",
@@ -64,7 +64,7 @@ dependencies = [
     "xatlas",
     "trimesh>=3.20.2",
     "timm==0.6.7",
-    "gsplat==1.4.0",
+    "gsplat>=1.5.3",
     "pytorch-msssim",
     "pathos",
     "packaging",

From 962602929c16f1dfc5388c46f0e7b6b0d9cea2ee Mon Sep 17 00:00:00 2001
From: yesMohanHere <mohanasriram.e@gmail.com>
Date: Fri, 17 Jul 2026 16:27:47 +0530
Subject: [PATCH 02/10] Apply ruff 0.12.2 formatting to render.py and
 render_state_machine.py

Pre-existing formatting drift on main: the pinned ruff 0.12.2 reformats
parenthesized with-statements, which fails the 'Run Ruff Formatter' CI
step for any PR branched from main.
---
 nerfstudio/scripts/render.py              | 7 ++++---
 nerfstudio/viewer/render_state_machine.py | 8 +++++---
 2 files changed, 9 insertions(+), 6 deletions(-)

diff --git a/nerfstudio/scripts/render.py b/nerfstudio/scripts/render.py
index fc035c4a9f..4f25301d78 100644
--- a/nerfstudio/scripts/render.py
+++ b/nerfstudio/scripts/render.py
@@ -186,9 +186,10 @@ def _render_trajectory_video(
                         max_idx = true_max_idx
 
                 if crop_data is not None:
-                    with renderers.background_color_override_context(
-                        crop_data.background_color.to(pipeline.device)
-                    ), torch.no_grad():
+                    with (
+                        renderers.background_color_override_context(crop_data.background_color.to(pipeline.device)),
+                        torch.no_grad(),
+                    ):
                         outputs = pipeline.model.get_outputs_for_camera(
                             cameras[camera_idx : camera_idx + 1], obb_box=obb_box
                         )
diff --git a/nerfstudio/viewer/render_state_machine.py b/nerfstudio/viewer/render_state_machine.py
index ced692d453..33b2e5ac5f 100644
--- a/nerfstudio/viewer/render_state_machine.py
+++ b/nerfstudio/viewer/render_state_machine.py
@@ -160,9 +160,11 @@ def _render_img(self, camera_state: CameraState):
                                 [color[0] / 255.0, color[1] / 255.0, color[2] / 255.0],
                                 device=self.viewer.get_model().device,
                             )
-                        with background_color_override_context(
-                            background_color
-                        ), torch.no_grad(), viewer_utils.SetTrace(self.check_interrupt):
+                        with (
+                            background_color_override_context(background_color),
+                            torch.no_grad(),
+                            viewer_utils.SetTrace(self.check_interrupt),
+                        ):
                             outputs = self.viewer.get_model().get_outputs_for_camera(camera, obb_box=obb)
                     else:
                         with torch.no_grad(), viewer_utils.SetTrace(self.check_interrupt):

From 76694778db36e87d76719062498f0630b0d833a7 Mon Sep 17 00:00:00 2001
From: yesMohanHere <mohanasriram.e@gmail.com>
Date: Fri, 17 Jul 2026 16:46:20 +0530
Subject: [PATCH 03/10] Fix write_next_bytes for numpy>=2.0

struct.pack no longer implicitly converts 1-element numpy arrays to
scalars in numpy 2.x, breaking COLMAP binary writers (e.g.
write_points3D_binary with error=np.array([0])) and failing
tests/process_data/test_process_images.py on current numpy.
---
 nerfstudio/data/utils/colmap_parsing_utils.py | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/nerfstudio/data/utils/colmap_parsing_utils.py b/nerfstudio/data/utils/colmap_parsing_utils.py
index 0eeaf6911b..8eabee9caf 100644
--- a/nerfstudio/data/utils/colmap_parsing_utils.py
+++ b/nerfstudio/data/utils/colmap_parsing_utils.py
@@ -95,6 +95,8 @@ def write_next_bytes(fid, data, format_char_sequence, endian_character="<"):
     """
     if isinstance(data, (list, tuple)):
         bytes = struct.pack(endian_character + format_char_sequence, *data)
+    elif isinstance(data, np.ndarray):
+        bytes = struct.pack(endian_character + format_char_sequence, *data.tolist())
     else:
         bytes = struct.pack(endian_character + format_char_sequence, data)
     fid.write(bytes)

From aa6e6dc12d44781cb78bd91ab07a6e462ef925ef Mon Sep 17 00:00:00 2001
From: yesMohanHere <mohanasriram.e@gmail.com>
Date: Fri, 17 Jul 2026 17:54:27 +0530
Subject: [PATCH 04/10] Dockerfile: build tiny-cuda-nn with
 --no-build-isolation
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

pip's isolated build environment pulls the latest setuptools, which no
longer ships pkg_resources (removed in recent setuptools), failing
tcnn's setup.py. The image already pins setuptools<70, which provides
pkg_resources — disable build isolation so it is actually used.
---
 Dockerfile | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/Dockerfile b/Dockerfile
index 6cd4e058f0..06729c6b7d 100644
--- a/Dockerfile
+++ b/Dockerfile
@@ -85,7 +85,7 @@ RUN pip install --no-cache-dir --upgrade pip 'setuptools<70.0.0' && \
     pip install --no-cache-dir torch==2.1.2+cu118 torchvision==0.16.2+cu118 'numpy<2.0.0' --extra-index-url https://download.pytorch.org/whl/cu118 && \
     git clone --branch master --recursive https://github.com/cvg/Hierarchical-Localization.git /opt/hloc && \
     cd /opt/hloc && git checkout v1.4 && python3.10 -m pip install --no-cache-dir . && cd ~ && \
-    TCNN_CUDA_ARCHITECTURES="${CUDA_ARCHITECTURES}" pip install --no-cache-dir "git+https://github.com/NVlabs/tiny-cuda-nn.git@b3473c81396fe927293bdfd5a6be32df8769927c#subdirectory=bindings/torch" && \
+    TCNN_CUDA_ARCHITECTURES="${CUDA_ARCHITECTURES}" pip install --no-cache-dir --no-build-isolation "git+https://github.com/NVlabs/tiny-cuda-nn.git@b3473c81396fe927293bdfd5a6be32df8769927c#subdirectory=bindings/torch" && \
     pip install --no-cache-dir pycolmap==0.6.1 pyceres==2.1 omegaconf==2.3.0
 
 # Install gsplat and nerfstudio.

From b23314d63845e5c9b257ab761475bf6ef778d448 Mon Sep 17 00:00:00 2001
From: yesMohanHere <mohanasriram.e@gmail.com>
Date: Fri, 17 Jul 2026 17:54:27 +0530
Subject: [PATCH 05/10] Pin gsplat==1.5.3
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Keep the exact-pin convention (was ==1.4.0) — the Dockerfile extracts
the gsplat version from pyproject.toml with a 'gsplat==' regex to fetch
the matching git tag, which a >= specifier would break.
---
 pyproject.toml | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/pyproject.toml b/pyproject.toml
index fd1b60bb76..ed0a9ae6f1 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -64,7 +64,7 @@ dependencies = [
     "xatlas",
     "trimesh>=3.20.2",
     "timm==0.6.7",
-    "gsplat>=1.5.3",
+    "gsplat==1.5.3",
     "pytorch-msssim",
     "pathos",
     "packaging",

From 0c2174e3a77677c2b24cd1240d77d6e1cb2c59d0 Mon Sep 17 00:00:00 2001
From: yesMohanHere <mohanasriram.e@gmail.com>
Date: Fri, 17 Jul 2026 19:38:09 +0530
Subject: [PATCH 06/10] Dockerfile: build gsplat with --no-build-isolation

gsplat's setup.py imports torch at build time; pip's isolated build
environment hides the image's torch and fails with
ModuleNotFoundError: No module named 'torch'.
---
 Dockerfile | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/Dockerfile b/Dockerfile
index 06729c6b7d..8ed113d9eb 100644
--- a/Dockerfile
+++ b/Dockerfile
@@ -98,7 +98,7 @@ COPY --from=source /tmp/nerfstudio/ /tmp/nerfstudio
 RUN export TORCH_CUDA_ARCH_LIST="$(echo "$CUDA_ARCHITECTURES" | tr ';' '\n' | awk '$0 > 70 {print substr($0,1,1)"."substr($0,2)}' | tr '\n' ' ' | sed 's/ $//')" && \
     export MAX_JOBS=4 && \
     GSPLAT_VERSION="$(sed -n 's/.*gsplat==\s*\([^," '"'"']*\).*/\1/p' /tmp/nerfstudio/pyproject.toml)" && \
-    pip install --no-cache-dir git+https://github.com/nerfstudio-project/gsplat.git@v${GSPLAT_VERSION} && \
+    pip install --no-cache-dir --no-build-isolation git+https://github.com/nerfstudio-project/gsplat.git@v${GSPLAT_VERSION} && \
     pip install --no-cache-dir /tmp/nerfstudio 'numpy<2.0.0' && \
     rm -rf /tmp/nerfstudio
 

From 63a6aca37da02511d540ebfb6d2c0d37ecaa4f06 Mon Sep 17 00:00:00 2001
From: yesMohanHere <mohanasriram.e@gmail.com>
Date: Fri, 17 Jul 2026 21:56:00 +0530
Subject: [PATCH 07/10] Keep custom_fwd/custom_bwd compatible with torch < 2.4

torch.amp only exports custom_fwd/custom_bwd (with mandatory
device_type) from torch 2.4; older torch only has them in
torch.cuda.amp without device_type. Fall back to the legacy import and
drop device_type there so nerfstudio keeps working with its declared
torch>=1.13.1 floor (e.g. the cu118 Docker image, torch 2.1.2).
---
 nerfstudio/field_components/activations.py | 16 +++++++++++++---
 1 file changed, 13 insertions(+), 3 deletions(-)

diff --git a/nerfstudio/field_components/activations.py b/nerfstudio/field_components/activations.py
index 3719990949..a7a54c319f 100644
--- a/nerfstudio/field_components/activations.py
+++ b/nerfstudio/field_components/activations.py
@@ -21,21 +21,31 @@
 import torch
 from jaxtyping import Float
 from torch import Tensor
-from torch.amp import custom_bwd, custom_fwd
 from torch.autograd import Function
 
+try:
+    # torch >= 2.4: custom_fwd/custom_bwd live in torch.amp and require device_type
+    from torch.amp import custom_bwd, custom_fwd
+
+    _CUSTOM_AMP_KWARGS = {"device_type": "cuda"}
+except ImportError:
+    # torch < 2.4: custom_fwd/custom_bwd live in torch.cuda.amp (no device_type argument)
+    from torch.cuda.amp import custom_bwd, custom_fwd
+
+    _CUSTOM_AMP_KWARGS = {}
+
 
 class _TruncExp(Function):
     # Implementation from torch-ngp:
     # https://github.com/ashawkey/torch-ngp/blob/93b08a0d4ec1cc6e69d85df7f0acdfb99603b628/activation.py
     @staticmethod
-    @custom_fwd(cast_inputs=torch.float32, device_type="cuda")
+    @custom_fwd(cast_inputs=torch.float32, **_CUSTOM_AMP_KWARGS)
     def forward(ctx, x):
         ctx.save_for_backward(x)
         return torch.exp(x)
 
     @staticmethod
-    @custom_bwd(device_type="cuda")
+    @custom_bwd(**_CUSTOM_AMP_KWARGS)
     def backward(ctx, g):
         x = ctx.saved_tensors[0]
         return g * torch.exp(x.clamp(-15, 15))

From 9e623f2ff680b1cc6422c89b644100dd6ffb8b4c Mon Sep 17 00:00:00 2001
From: yesMohanHere <mohanasriram.e@gmail.com>
Date: Sat, 18 Jul 2026 00:09:21 +0530
Subject: [PATCH 08/10] Fix custom_bwd decorator usage across torch versions

torch.cuda.amp.custom_bwd (torch < 2.4) is applied bare, while
torch.amp.custom_bwd (torch >= 2.4) must be called with device_type to
yield the decorator. Use small wrappers so both work.
---
 nerfstudio/field_components/activations.py | 13 +++++++------
 1 file changed, 7 insertions(+), 6 deletions(-)

diff --git a/nerfstudio/field_components/activations.py b/nerfstudio/field_components/activations.py
index a7a54c319f..b8edebc7fa 100644
--- a/nerfstudio/field_components/activations.py
+++ b/nerfstudio/field_components/activations.py
@@ -25,27 +25,28 @@
 
 try:
     # torch >= 2.4: custom_fwd/custom_bwd live in torch.amp and require device_type
-    from torch.amp import custom_bwd, custom_fwd
+    from torch.amp import custom_bwd as _custom_bwd, custom_fwd as _custom_fwd
 
-    _CUSTOM_AMP_KWARGS = {"device_type": "cuda"}
+    def custom_fwd(*args, **kwargs):
+        return _custom_fwd(*args, device_type="cuda", **kwargs)
+
+    custom_bwd = _custom_bwd(device_type="cuda")
 except ImportError:
     # torch < 2.4: custom_fwd/custom_bwd live in torch.cuda.amp (no device_type argument)
     from torch.cuda.amp import custom_bwd, custom_fwd
 
-    _CUSTOM_AMP_KWARGS = {}
-
 
 class _TruncExp(Function):
     # Implementation from torch-ngp:
     # https://github.com/ashawkey/torch-ngp/blob/93b08a0d4ec1cc6e69d85df7f0acdfb99603b628/activation.py
     @staticmethod
-    @custom_fwd(cast_inputs=torch.float32, **_CUSTOM_AMP_KWARGS)
+    @custom_fwd(cast_inputs=torch.float32)
     def forward(ctx, x):
         ctx.save_for_backward(x)
         return torch.exp(x)
 
     @staticmethod
-    @custom_bwd(**_CUSTOM_AMP_KWARGS)
+    @custom_bwd
     def backward(ctx, g):
         x = ctx.saved_tensors[0]
         return g * torch.exp(x.clamp(-15, 15))

From 4c8c27c11add83c8869d6dd8f2599ec40c6df776 Mon Sep 17 00:00:00 2001
From: yesMohanHere <mohanasriram.e@gmail.com>
Date: Sat, 18 Jul 2026 00:23:27 +0530
Subject: [PATCH 09/10] Suppress pyright redefinition error on legacy amp
 fallback import

---
 nerfstudio/field_components/activations.py | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/nerfstudio/field_components/activations.py b/nerfstudio/field_components/activations.py
index b8edebc7fa..29afaf1f03 100644
--- a/nerfstudio/field_components/activations.py
+++ b/nerfstudio/field_components/activations.py
@@ -33,7 +33,7 @@ def custom_fwd(*args, **kwargs):
     custom_bwd = _custom_bwd(device_type="cuda")
 except ImportError:
     # torch < 2.4: custom_fwd/custom_bwd live in torch.cuda.amp (no device_type argument)
-    from torch.cuda.amp import custom_bwd, custom_fwd
+    from torch.cuda.amp import custom_bwd, custom_fwd  # type: ignore
 
 
 class _TruncExp(Function):

From 88ee118755b4c18c60b996a3937bf72977ffa9b2 Mon Sep 17 00:00:00 2001
From: yesMohanHere <mohanasriram.e@gmail.com>
Date: Sat, 18 Jul 2026 18:55:55 +0530
Subject: [PATCH 10/10] Add Dockerfile.blackwell for NVIDIA Blackwell (sm_120)
 GPUs

CUDA 12.8-based image variant that can build for and run on sm_120:
- nvidia/cuda:12.8.1 base, PyTorch 2.8.0+cu128 (ships sm_120 kernels)
- CUDA_ARCHITECTURES defaults to 120; arch-list awk handles 3-digit archs
- COLMAP 3.12.6 / GLOMAP 1.2.0 (CUDA 12.8-compatible releases)
- tiny-cuda-nn pinned to a Blackwell-capable master commit
- tcnn/gsplat built with --no-build-isolation
- nerfacc JIT kernels pre-compiled at build time (runtime stage has no nvcc)
- apt sandbox workaround for seccomp-related GPG verification failures
- docs: mention the variant in the docker install section

Validated on RTX 5060: image builds clean; splatfacto and nerfacto both
train to completion inside the container.
---
 Dockerfile.blackwell            | 181 ++++++++++++++++++++++++++++++++
 docs/quickstart/installation.md |   8 ++
 2 files changed, 189 insertions(+)
 create mode 100644 Dockerfile.blackwell

diff --git a/Dockerfile.blackwell b/Dockerfile.blackwell
new file mode 100644
index 0000000000..2443209569
--- /dev/null
+++ b/Dockerfile.blackwell
@@ -0,0 +1,181 @@
+# syntax=docker/dockerfile:1
+# Blackwell (sm_120, RTX 50-series) variant of the nerfstudio Docker image.
+#
+# Differences from the stock Dockerfile:
+# - CUDA 12.8 base images (sm_120 requires CUDA >= 12.8)
+# - PyTorch 2.8 with cu128 wheels (ships sm_120 kernels)
+# - COLMAP 3.12.6 / GLOMAP 1.2.0 (CUDA 12.8-compatible releases)
+# - tiny-cuda-nn pinned to a Blackwell-capable master commit
+# - CUDA_ARCHITECTURES defaults to "120"
+# - nerfacc's JIT CUDA kernels are pre-compiled during the build (the
+#   runtime stage has no nvcc, so they could not be built on first use)
+ARG UBUNTU_VERSION=22.04
+ARG NVIDIA_CUDA_VERSION=12.8.1
+# CUDA architectures, required by Colmap and tiny-cuda-nn. Use >= 8.0 for faster TCNN.
+ARG CUDA_ARCHITECTURES="120"
+ARG NERFSTUDIO_VERSION=""
+
+# Pull source either provided or from git.
+FROM scratch as source_copy
+ONBUILD COPY . /tmp/nerfstudio
+FROM alpine/git as source_no_copy
+ARG NERFSTUDIO_VERSION
+ONBUILD RUN git clone --branch ${NERFSTUDIO_VERSION} --recursive https://github.com/nerfstudio-project/nerfstudio.git /tmp/nerfstudio
+ARG NERFSTUDIO_VERSION
+FROM source_${NERFSTUDIO_VERSION:+no_}copy as source
+
+FROM nvidia/cuda:${NVIDIA_CUDA_VERSION}-devel-ubuntu${UBUNTU_VERSION} as builder
+ARG CUDA_ARCHITECTURES
+ARG NVIDIA_CUDA_VERSION
+ARG UBUNTU_VERSION
+
+ENV DEBIAN_FRONTEND=noninteractive
+ENV QT_XCB_GL_INTEGRATION=xcb_egl
+# Work around apt _apt sandbox verification failing under docker's seccomp
+# profile on newer kernels (spurious "invalid signature" GPG errors).
+RUN echo 'APT::Sandbox::User "root";' > /etc/apt/apt.conf.d/99docker-sandbox && \
+    apt-get update && \
+    apt-get install -y --no-install-recommends --no-install-suggests \
+        git \
+        wget \
+        ninja-build \
+        build-essential \
+        libboost-program-options-dev \
+        libboost-filesystem-dev \
+        libboost-graph-dev \
+        libboost-system-dev \
+        libeigen3-dev \
+        libflann-dev \
+        libfreeimage-dev \
+        libmetis-dev \
+        libgoogle-glog-dev \
+        libgtest-dev \
+        libsqlite3-dev \
+        libglew-dev \
+        qtbase5-dev \
+        libqt5opengl5-dev \
+        libcgal-dev \
+        libceres-dev \
+        python3.10-dev \
+        python3-pip
+
+# Build and install CMake
+RUN wget https://github.com/Kitware/CMake/releases/download/v3.31.3/cmake-3.31.3-linux-x86_64.sh \
+    -q -O /tmp/cmake-install.sh \
+    && chmod u+x /tmp/cmake-install.sh \
+    && mkdir /opt/cmake-3.31.3 \
+    && /tmp/cmake-install.sh --skip-license --prefix=/opt/cmake-3.31.3 \
+    && rm /tmp/cmake-install.sh \
+    && ln -s /opt/cmake-3.31.3/bin/* /usr/local/bin
+
+# Build and install GLOMAP.
+RUN git clone https://github.com/colmap/glomap.git && \
+    cd glomap && \
+    git checkout "1.2.0" && \
+    mkdir build && \
+    cd build && \
+    mkdir -p /build && \
+    cmake .. -GNinja "-DCMAKE_CUDA_ARCHITECTURES=${CUDA_ARCHITECTURES}" \
+        -DCMAKE_INSTALL_PREFIX=/build/glomap && \
+    ninja install -j4 && \
+    cd ~
+
+# Build and install COLMAP.
+RUN git clone https://github.com/colmap/colmap.git && \
+    cd colmap && \
+    git checkout "3.12.6" && \
+    mkdir build && \
+    cd build && \
+    mkdir -p /build && \
+    cmake .. -GNinja "-DCMAKE_CUDA_ARCHITECTURES=${CUDA_ARCHITECTURES}" \
+        -DCMAKE_INSTALL_PREFIX=/build/colmap && \
+    ninja install -j4 && \
+    cd ~
+
+# Upgrade pip and install dependencies (torch with sm_120 kernels via cu128 wheels).
+RUN pip install --no-cache-dir --upgrade pip 'setuptools<70.0.0' && \
+    pip install --no-cache-dir torch==2.8.0 torchvision==0.23.0 --index-url https://download.pytorch.org/whl/cu128 && \
+    git clone --branch master --recursive https://github.com/cvg/Hierarchical-Localization.git /opt/hloc && \
+    cd /opt/hloc && git checkout v1.4 && python3.10 -m pip install --no-cache-dir . && cd ~ && \
+    TCNN_CUDA_ARCHITECTURES="${CUDA_ARCHITECTURES}" pip install --no-cache-dir --no-build-isolation "git+https://github.com/NVlabs/tiny-cuda-nn.git@749dd70c5afc5a9dadb85e5652ed65d55e0ba187#subdirectory=bindings/torch" && \
+    pip install --no-cache-dir pycolmap==0.6.1 pyceres==2.1 omegaconf==2.3.0
+
+# Install gsplat and nerfstudio.
+# NOTE: both are installed jointly in order to prevent docker cache with latest
+# gsplat version (we do not expliticly specify the commit hash).
+#
+# We set MAX_JOBS to reduce resource usage for GH actions:
+# - https://github.com/nerfstudio-project/gsplat/blob/db444b904976d6e01e79b736dd89a1070b0ee1d0/setup.py#L13-L23
+COPY --from=source /tmp/nerfstudio/ /tmp/nerfstudio
+RUN export TORCH_CUDA_ARCH_LIST="$(echo "$CUDA_ARCHITECTURES" | tr ';' '\n' | awk '{if ($0 > 99) print substr($0,1,2)"."substr($0,3); else if ($0 > 70) print substr($0,1,1)"."substr($0,2)}' | tr '\n' ' ' | sed 's/ $//')" && \
+    export MAX_JOBS=4 && \
+    GSPLAT_VERSION="$(sed -n 's/.*gsplat==\s*\([^," '"'"']*\).*/\1/p' /tmp/nerfstudio/pyproject.toml)" && \
+    pip install --no-cache-dir --no-build-isolation git+https://github.com/nerfstudio-project/gsplat.git@v${GSPLAT_VERSION} && \
+    pip install --no-cache-dir /tmp/nerfstudio 'numpy<2.0.0' && \
+    rm -rf /tmp/nerfstudio
+
+# Pre-compile nerfacc's JIT CUDA kernels. nerfacc compiles its kernels on first
+# use, but the runtime stage has no nvcc, so build them now and copy the torch
+# extension cache into the runtime image. The build container has no GPU, so
+# use the CUDA driver stub to satisfy the import after compilation.
+RUN export TORCH_CUDA_ARCH_LIST="$(echo "$CUDA_ARCHITECTURES" | tr ';' '\n' | awk '{if ($0 > 99) print substr($0,1,2)"."substr($0,3); else if ($0 > 70) print substr($0,1,1)"."substr($0,2)}' | tr '\n' ' ' | sed 's/ $//')" && \
+    export MAX_JOBS=4 && \
+    ln -sf /usr/local/cuda/lib64/stubs/libcuda.so /usr/local/cuda/lib64/stubs/libcuda.so.1 && \
+    LD_LIBRARY_PATH=/usr/local/cuda/lib64/stubs:${LD_LIBRARY_PATH} \
+        python3.10 -c "from nerfacc.cuda._backend import _C; print('nerfacc CUDA kernels pre-compiled')" && \
+    find /root/.cache/torch_extensions -name "nerfacc_cuda.so" | grep .
+
+# Fix permissions
+RUN chmod -R go=u /usr/local/lib/python3.10 && \
+    chmod -R go=u /build
+
+#
+# Docker runtime stage.
+#
+FROM nvidia/cuda:${NVIDIA_CUDA_VERSION}-runtime-ubuntu${UBUNTU_VERSION} as runtime
+ARG CUDA_ARCHITECTURES
+ARG NVIDIA_CUDA_VERSION
+ARG UBUNTU_VERSION
+
+LABEL org.opencontainers.image.source = "https://github.com/nerfstudio-project/nerfstudio"
+LABEL org.opencontainers.image.licenses = "Apache License 2.0"
+LABEL org.opencontainers.image.base.name="docker.io/library/nvidia/cuda:${NVIDIA_CUDA_VERSION}-devel-ubuntu${UBUNTU_VERSION}"
+LABEL org.opencontainers.image.documentation = "https://docs.nerf.studio/"
+
+# Minimal dependencies to run COLMAP binary compiled in the builder stage.
+# Note: this reduces the size of the final image considerably, since all the
+# build dependencies are not needed.
+RUN echo 'APT::Sandbox::User "root";' > /etc/apt/apt.conf.d/99docker-sandbox && \
+    apt-get update && \
+    apt-get install -y --no-install-recommends --no-install-suggests \
+        libboost-filesystem1.74.0 \
+        libboost-program-options1.74.0 \
+        libc6 \
+        libceres2 \
+        libfreeimage3 \
+        libgcc-s1 \
+        libgl1 \
+        libglew2.2 \
+        libgoogle-glog0v5 \
+        libqt5core5a \
+        libqt5gui5 \
+        libqt5widgets5 \
+        python3.10 \
+        python3.10-dev \
+        build-essential \
+        python-is-python3 \
+        ffmpeg
+
+# Copy packages from builder stage.
+COPY --from=builder /build/colmap/ /usr/local/
+COPY --from=builder /build/glomap/ /usr/local/
+COPY --from=builder /usr/local/lib/python3.10/dist-packages/ /usr/local/lib/python3.10/dist-packages/
+COPY --from=builder /usr/local/bin/ns* /usr/local/bin/
+# Pre-compiled nerfacc JIT kernels (see builder stage).
+COPY --from=builder /root/.cache/torch_extensions /root/.cache/torch_extensions
+
+# Install nerfstudio cli auto completion
+RUN /bin/bash -c 'ns-install-cli --mode install'
+
+# Bash as default entrypoint.
+CMD /bin/bash -l
diff --git a/docs/quickstart/installation.md b/docs/quickstart/installation.md
index c4e79a19fc..22bb7403d6 100644
--- a/docs/quickstart/installation.md
+++ b/docs/quickstart/installation.md
@@ -267,6 +267,14 @@ docker build \
     --file Dockerfile .
 ```
 
+For NVIDIA Blackwell GPUs (RTX 50-series, compute capability sm_120), use the CUDA 12.8-based
+`Dockerfile.blackwell` variant instead — the stock image targets CUDA 11.8, which does not
+support sm_120:
+
+```bash
+docker build --tag nerfstudio-blackwell -f Dockerfile.blackwell .
+```
+
 ### Using an interactive container
 
 The docker container can be launched with an interactive terminal where nerfstudio commands can be entered as usual. Some parameters are required and some are strongly recommended for usage as following: