Add DGX Spark support to the card
Browse files
README.md
CHANGED
|
@@ -37,12 +37,25 @@ published build of this model carries.
|
|
| 37 |
|
| 38 |
## Requirements
|
| 39 |
|
|
|
|
|
|
|
| 40 |
| | |
|
| 41 |
| --- | --- |
|
| 42 |
| macOS | **15.0 or later** |
|
| 43 |
| chip | **Apple Silicon** (arm64). There is no Intel build. |
|
| 44 |
| Python | none -- the binary carries what it needs |
|
| 45 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 46 |
Memory is not a fixed figure for a paged build, and that is the point of one: it
|
| 47 |
fills what fits and streams the rest from disk. On a 512 GB Mac Studio with room
|
| 48 |
to spare this model settles at about **172 GB** resident. A smaller machine holds
|
|
@@ -62,6 +75,7 @@ the wait is the SSD's, not the model's.
|
|
| 62 |
# upgrading? clear the previous version's unpack directory first
|
| 63 |
rm -rf ~/.libra/cache/onefile/gbx_lm
|
| 64 |
|
|
|
|
| 65 |
curl -fL -o gbx_lm-darwin-arm64.tar.gz 'https://github.com/GreenBitAI/gbx-lm/releases/latest/download/gbx_lm-darwin-arm64.tar.gz' \
|
| 66 |
&& tar -xzf gbx_lm-darwin-arm64.tar.gz gbx_lm \
|
| 67 |
&& mkdir -p "$HOME/.local/bin" \
|
|
@@ -71,8 +85,8 @@ curl -fL -o gbx_lm-darwin-arm64.tar.gz 'https://github.com/GreenBitAI/gbx-lm/rel
|
|
| 71 |
gbx_lm -h
|
| 72 |
```
|
| 73 |
|
| 74 |
-
The build is signed with a Developer ID and notarised, so macOS runs it
|
| 75 |
-
the usual detour for a downloaded binary.
|
| 76 |
|
| 77 |
**`command not found`** -- `$HOME/.local/bin` is not on your `PATH`:
|
| 78 |
|
|
@@ -81,7 +95,7 @@ echo 'export PATH="$HOME/.local/bin:$PATH"' >> ~/.zshrc && source ~/.zshrc #
|
|
| 81 |
echo 'export PATH="$HOME/.local/bin:$PATH"' >> ~/.bash_profile && source ~/.bash_profile # bash
|
| 82 |
```
|
| 83 |
|
| 84 |
-
**`Killed: 9`** -- a previous version's files are still in the unpack directory,
|
| 85 |
and macOS refuses to mix two builds. Run the `rm -rf` line above, then try again.
|
| 86 |
|
| 87 |
## Run
|
|
|
|
| 37 |
|
| 38 |
## Requirements
|
| 39 |
|
| 40 |
+
**Mac**
|
| 41 |
+
|
| 42 |
| | |
|
| 43 |
| --- | --- |
|
| 44 |
| macOS | **15.0 or later** |
|
| 45 |
| chip | **Apple Silicon** (arm64). There is no Intel build. |
|
| 46 |
| Python | none -- the binary carries what it needs |
|
| 47 |
|
| 48 |
+
**Linux** (gbx-lm v0.8.1 or later)
|
| 49 |
+
|
| 50 |
+
| | |
|
| 51 |
+
| --- | --- |
|
| 52 |
+
| machine | **NVIDIA DGX Spark** (GB10). Other Linux machines are not supported yet. |
|
| 53 |
+
| system | arm64, glibc 2.39 or later (DGX OS / Ubuntu 24.04) |
|
| 54 |
+
| driver | an NVIDIA driver with CUDA 13 support. No CUDA toolkit is needed: the binary carries the CUDA libraries it uses. |
|
| 55 |
+
| Python | none |
|
| 56 |
+
|
| 57 |
+
On a DGX Spark (128 GB) the model does not fit, so its experts are paged from the SSD.
|
| 58 |
+
|
| 59 |
Memory is not a fixed figure for a paged build, and that is the point of one: it
|
| 60 |
fills what fits and streams the rest from disk. On a 512 GB Mac Studio with room
|
| 61 |
to spare this model settles at about **172 GB** resident. A smaller machine holds
|
|
|
|
| 75 |
# upgrading? clear the previous version's unpack directory first
|
| 76 |
rm -rf ~/.libra/cache/onefile/gbx_lm
|
| 77 |
|
| 78 |
+
# on a DGX Spark, use gbx_lm-linux-arm64.tar.gz in place of gbx_lm-darwin-arm64.tar.gz
|
| 79 |
curl -fL -o gbx_lm-darwin-arm64.tar.gz 'https://github.com/GreenBitAI/gbx-lm/releases/latest/download/gbx_lm-darwin-arm64.tar.gz' \
|
| 80 |
&& tar -xzf gbx_lm-darwin-arm64.tar.gz gbx_lm \
|
| 81 |
&& mkdir -p "$HOME/.local/bin" \
|
|
|
|
| 85 |
gbx_lm -h
|
| 86 |
```
|
| 87 |
|
| 88 |
+
The Mac build is signed with a Developer ID and notarised, so macOS runs it
|
| 89 |
+
without the usual detour for a downloaded binary.
|
| 90 |
|
| 91 |
**`command not found`** -- `$HOME/.local/bin` is not on your `PATH`:
|
| 92 |
|
|
|
|
| 95 |
echo 'export PATH="$HOME/.local/bin:$PATH"' >> ~/.bash_profile && source ~/.bash_profile # bash
|
| 96 |
```
|
| 97 |
|
| 98 |
+
**`Killed: 9`** (Mac) -- a previous version's files are still in the unpack directory,
|
| 99 |
and macOS refuses to mix two builds. Run the `rm -rf` line above, then try again.
|
| 100 |
|
| 101 |
## Run
|