You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Works on every platform with no additional drivers. **This is the recommended install for most users.** Inference runs on PyTorch's CPU backend and automatically upgrades to NVIDIA CUDA or Apple Silicon MPS if detected at runtime — no separate install needed for those.
17
+
18
+
**pip**
19
+
```bash
20
+
pip install "vizion3d[cpu]"
21
+
```
22
+
23
+
**uv**
24
+
```bash
25
+
uv add "vizion3d[cpu]"
26
+
```
27
+
28
+
> **Mac M-series users:** the standard CPU extra automatically includes Metal Performance Shaders (MPS) support — no separate install needed. vizion3d will use your GPU via MPS as long as you are on macOS 12.3 or later with PyTorch ≥ 2.0.
29
+
30
+
---
31
+
32
+
## NVIDIA CUDA
33
+
34
+
Delivers the highest throughput for depth estimation. On NVIDIA Ampere GPUs and newer (RTX 30xx / A100 and above), PyTorch additionally uses Tensor Cores for mixed-precision acceleration.
Install CUDA and cuDNN **before** installing vizion3d. The PyTorch wheel bundled with the `cuda` extra already includes its own CUDA runtime libraries, but the driver must be present on the host.
45
+
46
+
**pip**
47
+
```bash
48
+
pip install "vizion3d[cuda]"
49
+
```
50
+
51
+
**uv**
52
+
```bash
53
+
uv add "vizion3d[cuda]"
54
+
```
55
+
56
+
vizion3d detects CUDA via `torch.cuda.is_available()` at runtime and moves models and tensors to the GPU automatically — no configuration needed.
57
+
58
+
---
59
+
60
+
## AMD ROCm
61
+
62
+
Provides GPU-accelerated inference on supported AMD GPUs using the ROCm open-source compute stack. ROCm exposes itself through PyTorch's CUDA namespace (`torch.cuda.is_available()` returns `True`), so vizion3d uses it transparently with no code changes.
63
+
64
+
> **Platform:** ROCm is supported on **Linux only**. There is no ROCm support for Windows or macOS.
65
+
66
+
### Supported hardware
67
+
68
+
| Family | Examples |
69
+
|---|---|
70
+
| RDNA2 | RX 6700 XT, RX 6800, RX 6900 XT |
71
+
| RDNA3 | RX 7800 XT, RX 7900 XTX |
72
+
| CDNA | Instinct MI100 |
73
+
| CDNA2 | Instinct MI200 series |
74
+
| CDNA3 | Instinct MI300 series |
75
+
76
+
For the full supported GPU list see the [AMD ROCm hardware compatibility guide](https://rocm.docs.amd.com/projects/install-on-linux/en/latest/reference/system-requirements.html).
77
+
78
+
### Prerequisites
79
+
80
+
Install the ROCm stack on your system before installing the PyTorch ROCm wheel. Follow AMD's official guide:
81
+
82
+
-[ROCm installation for Linux](https://rocm.docs.amd.com/projects/install-on-linux/en/latest/)
83
+
84
+
### Install
85
+
86
+
Because the ROCm PyTorch wheel is hosted on PyTorch's own index (not PyPI), it must be installed **before** vizion3d — vizion3d's base install has no torch dependency and will not overwrite it.
For the full list of available ROCm wheel versions see [PyTorch ROCm install guide](https://rocm.docs.amd.com/projects/install-on-linux/en/latest/install/3rd-party/pytorch-install.html#using-wheels-package).
95
+
96
+
**Step 2 — install vizion3d (no extra needed)**
97
+
```bash
98
+
pip install vizion3d
99
+
```
100
+
101
+
Because `vizion3d` declares no torch dependency in its base install, pip will not touch the ROCm wheel you installed in step 1.
102
+
103
+
> **Warning:** do **not** run `pip install "vizion3d[cpu]"` or `pip install "vizion3d[cuda]"` after installing the ROCm wheel — those extras pull a standard PyPI torch build and will replace your ROCm installation.
104
+
105
+
### Limitations
106
+
107
+
- Linux only — ROCm does not run on Windows or macOS.
108
+
- Only GPUs on AMD's official support list are guaranteed to work; consumer RDNA1 cards (RX 5000 series) are not supported.
109
+
- Some PyTorch operations fall back to CPU on ROCm; performance for those ops will match CPU speed.
Copy file name to clipboardExpand all lines: docs/index.md
+4-107Lines changed: 4 additions & 107 deletions
Display the source diff
Display the rich diff
Original file line number
Diff line number
Diff line change
@@ -16,7 +16,7 @@ Every task is accessible through three consumption modes driven by one shared CQ
16
16
17
17
Requires **Python 3.12** (Open3D constraint).
18
18
19
-
PyTorch is **not bundled** in the base install — choose the extra that matches your hardware (see [Hardware acceleration](#hardware-acceleration) below). For most users the `cpu` extra is the right default; it installs the standard PyTorch wheel which covers CPU, NVIDIA CUDA, and Apple Silicon MPS automatically.
19
+
PyTorch is **not bundled** in the base install — choose the extra that matches your hardware (see [Hardware Acceleration](hardware_acceleration.md)). For most users the `cpu` extra is the right default; it installs the standard PyTorch wheel which covers CPU, NVIDIA CUDA, and Apple Silicon MPS automatically.
20
20
21
21
**pip**
22
22
```bash
@@ -38,113 +38,9 @@ uv add "vizion3d[cpu]"
38
38
39
39
## Hardware acceleration
40
40
41
-
vizion3d detects the best available device automatically at runtime — no code changes required. Choose the install extra that matches your hardware.
41
+
vizion3d detects the best available backend automatically at runtime — no code changes required. Supported backends are CPU, NVIDIA CUDA, Apple Silicon MPS, and AMD ROCm.
Works on every platform with no additional drivers. **This is the recommended install for most users.** Inference runs on PyTorch's CPU backend and automatically upgrades to NVIDIA CUDA or Apple Silicon MPS if detected at runtime — no separate install needed for those.
55
-
56
-
**pip**
57
-
```bash
58
-
pip install "vizion3d[cpu]"
59
-
```
60
-
61
-
**uv**
62
-
```bash
63
-
uv add "vizion3d[cpu]"
64
-
```
65
-
66
-
> **Mac M-series users:** the standard CPU extra automatically includes Metal Performance Shaders (MPS) support — no separate install needed. vizion3d will use your GPU via MPS as long as you are on macOS 12.3 or later with PyTorch ≥ 2.0.
67
-
68
-
---
69
-
70
-
### NVIDIA CUDA
71
-
72
-
Delivers the highest throughput for depth estimation. On NVIDIA Ampere GPUs and newer (RTX 30xx / A100 and above), PyTorch additionally uses Tensor Cores for mixed-precision acceleration.
Install CUDA and cuDNN **before** installing vizion3d. The PyTorch wheel bundled with the `cuda` extra already includes its own CUDA runtime libraries, but the driver must be present on the host.
83
-
84
-
**pip**
85
-
```bash
86
-
pip install "vizion3d[cuda]"
87
-
```
88
-
89
-
**uv**
90
-
```bash
91
-
uv add "vizion3d[cuda]"
92
-
```
93
-
94
-
vizion3d detects CUDA via `torch.cuda.is_available()` at runtime and moves models and tensors to the GPU automatically — no configuration needed.
95
-
96
-
---
97
-
98
-
### AMD ROCm
99
-
100
-
Provides GPU-accelerated inference on supported AMD GPUs using the ROCm open-source compute stack. ROCm exposes itself through PyTorch's CUDA namespace (`torch.cuda.is_available()` returns `True`), so vizion3d uses it transparently with no code changes.
101
-
102
-
> **Platform:** ROCm is supported on **Linux only**. There is no ROCm support for Windows or macOS.
103
-
104
-
#### Supported hardware
105
-
106
-
| Family | Examples |
107
-
|---|---|
108
-
| RDNA2 | RX 6700 XT, RX 6800, RX 6900 XT |
109
-
| RDNA3 | RX 7800 XT, RX 7900 XTX |
110
-
| CDNA | Instinct MI100 |
111
-
| CDNA2 | Instinct MI200 series |
112
-
| CDNA3 | Instinct MI300 series |
113
-
114
-
For the full supported GPU list see the [AMD ROCm hardware compatibility guide](https://rocm.docs.amd.com/projects/install-on-linux/en/latest/reference/system-requirements.html).
115
-
116
-
#### Prerequisites
117
-
118
-
Install the ROCm stack on your system before installing the PyTorch ROCm wheel. Follow AMD's official guide:
119
-
120
-
-[ROCm installation for Linux](https://rocm.docs.amd.com/projects/install-on-linux/en/latest/)
121
-
122
-
#### Install
123
-
124
-
Because the ROCm PyTorch wheel is hosted on PyTorch's own index (not PyPI), it must be installed **before** vizion3d — vizion3d's base install has no torch dependency and will not overwrite it.
For the full list of available ROCm wheel versions see [PyTorch ROCm install guide](https://rocm.docs.amd.com/projects/install-on-linux/en/latest/install/3rd-party/pytorch-install.html#using-wheels-package).
133
-
134
-
**Step 2 — install vizion3d (no extra needed)**
135
-
```bash
136
-
pip install vizion3d
137
-
```
138
-
139
-
Because `vizion3d` declares no torch dependency in its base install, pip will not touch the ROCm wheel you installed in step 1.
140
-
141
-
> **Warning:** do **not** run `pip install "vizion3d[cpu]"` or `pip install "vizion3d[cuda]"` after installing the ROCm wheel — those extras pull a standard PyPI torch build and will replace your ROCm installation.
142
-
143
-
#### Limitations
144
-
145
-
- Linux only — ROCm does not run on Windows or macOS.
146
-
- Only GPUs on AMD's official support list are guaranteed to work; consumer RDNA1 cards (RX 5000 series) are not supported.
147
-
- Some PyTorch operations fall back to CPU on ROCm; performance for those ops will match CPU speed.
43
+
For per-backend prerequisites, install commands, and platform notes, see the [Hardware Acceleration](hardware_acceleration.md) page.
148
44
149
45
---
150
46
@@ -215,4 +111,5 @@ Each task lives in its own module under `vizion3d/<category>/` and exposes exact
0 commit comments