* fix(install.ps1): split torch+unsloth install to fix non-NVIDIA package resolution
--torch-backend=auto on a non-NVIDIA Windows machine causes uv to resolve
unsloth==2024.8 (pre-CLI, no unsloth.exe). Fix: detect GPU robustly (PATH +
hardcoded fallback paths, mirrors setup.ps1), install torch first with an
explicit --index-url (CUDA variant for NVIDIA, CPU for everyone else), then
install unsloth separately without --torch-backend so the solver always picks
a modern release that ships the Studio CLI.
Closes the remaining gap flagged in #4478.
* fix(install.ps1): align warning with setup.ps1, add --upgrade, handle CUDA 11.x
- Match the no-GPU warning message to studio/setup.ps1 wording
(chat-only GGUF mode, driver download link)
- Add CUDA 11.x floor check in Get-TorchIndexUrl so old drivers
fall back to CPU wheels instead of silently getting cu124
- Log a warning when nvidia-smi output cannot be parsed
- Add --upgrade to both uv pip install calls so re-runs pick up
newer package versions
* revert --upgrade from uv pip install calls
uv pip install already resolves to the latest satisfying version;
--upgrade is unnecessary and could force unwanted re-installs.
* fix: replace frozen cu124 fallbacks with cu126, guard CUDA 11.x
cu124 wheels are frozen at torch 2.6.0 -- falling back to them pins
users to an outdated PyTorch. Three issues fixed in both install.ps1
and setup.ps1:
1. CUDA 12.0-12.5 now maps to cu126 (was cu124).
2. CUDA 11.x and older now falls back to cpu (was cu124, which would
silently install incompatible GPU wheels).
3. Parse-failure and no-nvidia-smi fallbacks updated to cu126/cpu.
Adds tests/test_cuda_wheel_mapping.py covering the mapping logic,
nvidia-smi parsing, PS1 file sync, PyTorch index URL validation,
and sandbox torch installs.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* remove test file from PR branch
Test file kept locally, not needed in the PR.
* fix: map CUDA 11.x to cu118 instead of cpu
PyTorch still publishes cu118 wheels (up to torch 2.7.1), so CUDA 11.x
users get GPU-accelerated torch rather than being forced to CPU-only.
Only CUDA 10.x and older fall back to cpu.
* fix: revert CUDA 12.0-12.5 to cu124, handle cpu tag in setup.ps1
CUDA 12.0-12.5 drivers only support up to their reported CUDA version,
so cu126 wheels (built with CUDA 12.6) fail to load. Revert the catch-
all for 12.0-12.5 back to cu124.
Also fix setup.ps1 caller: when Get-PytorchCudaTag returns "cpu" (e.g.
CUDA 10.x driver), the installer now correctly skips Triton and prints
"CPU-only" instead of "CUDA support (cpu)".
* fix: add --upgrade to unsloth install for stale venv repair
On reruns against an existing venv, uv pip install unsloth makes no
changes if unsloth==2024.8 is already installed (it satisfies the
constraint). Adding --upgrade only to the unsloth install ensures
stale installs get repaired without forcing a multi-GB torch
re-download.
* fix: use --upgrade-package to avoid clobbering torch CUDA wheels
`--upgrade unsloth` re-resolves torch from default PyPI, stripping the
+cuXXX suffix installed in step 1. `--upgrade-package unsloth unsloth`
upgrades only unsloth (and pulls missing deps like transformers, trl)
while preserving the pinned torch from the CUDA-specific index.
* docs: explain why split-install and --upgrade-package are needed
Expand the inline comment block to document both design decisions:
1. Why torch is installed separately (solver fallback to 2024.8)
2. Why --upgrade-package is used instead of --upgrade (preserves CUDA wheels)
---------
Co-authored-by: LeoBorcherding <LeoBorcherding@users.noreply.github.com>
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
71c77d4e96