just a repo where i store information about the UT3G in the future, the idea would be to have multiple GPU, with different eGPU board, that shard model. why not a 9070xtx with 32GB and a tinygrad chesnut
Start by setting up Tinygrad
Then we want to install TinyGPU
With the 7900xtx we use ENV
Since commit 33cd373ad all the hardware specifique part of my setup has been depreceted
Mooved from system.py to extra/hcq1/
git checkout 33cd373ad keeps our tinygrad version working with tinygpu
or git checkout -b $NEW_BRANCH 33cd373ad to create a new branch from this commit
we want to understand each part that is involved in our setup and make it work on (master) to use VIZ, SHARD, and other tools
One load directly the weight in the card(tinyagent), master is representing it as a UOps. i do not understand how we can represent the parameters/weight on DISK correctly, has UOps.
│ Allocator's internal staging ring (self.b) + _copyin/ │ Staging as a graph node: stage = UOp.alloc(..., device=dev.host)
│ _copyout │ (hcq2.py:170)
├───────────────────────────────────────────────────────────┼──────────────────────────────────────────────────────────────────
│ CPU memcpy through .cpu_view() window │ Enqueued SDMA copy reading the stage via GPUVM
├───────────────────────────────────────────────────────────┼──────────────────────────────────────────────────────────────────
│ TCP-remote peer: CPU:host:port with .remote = │ Same HostAllocator._alloc hook (device.py:302)
│ RemotePCIDevice │
The actuall DEV for our setup is DEV=PCI+AMD:HIP:gfx1100
PCI which is how we talk with the AMD device, HIP is the renderer(compiler), gfx1100 which is the target architecture
DEV=AMD:HIP uv run python3 -m tinygrad.device
uv run python -c "from tinygrad import Device; print(Device.DEFAULT)"list mac driver extension
systemextensionsctl listthe TinyGPU should appear has enabled and activated
Theses are the Mac specifique commands to get info about the hardware
system_profiler SPMemoryDataType
system_profiler SPUSBDataType
system_profiler SPPCIDataType
system_profiler SPThunderboltDataType