ceos/unsloth
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions
unsloth/tests/sh/test_get_torch_index_url.sh
Daniel Han a3a0cb1606
 studio/setup.sh: cope with fresh CUDA toolkits like 13.3 (#5826) 
* Studio setup.sh: cope with fresh CUDA toolkits like 13.3

CUDA 13.3 shipped today. Three loose ends in studio/setup.sh surfaced
during the llama.cpp build path:

1. setup.ps1 already aborts cleanly when the CUDA toolkit is below
   llama.cpp's minimum (12.4) via #4517, but setup.sh still hit the
   generic cmake failure described in #4437. Added a min-version check
   that downgrades to a CPU build for nvcc < 12.4 with a clear message
   pointing to the toolkit archive.
2. The first day a new CUDA toolkit ships, its host-compiler whitelist
   lags whatever gcc/clang the distro is on, so nvcc rejects the host
   compiler with a wall of "#error -- unsupported GNU version" before
   any real compile runs. NVCC_PREPEND_FLAGS now carries
   -allow-unsupported-compiler so the build moves on instead.
3. The Linux CUDA/ROCm configure failure path had no symmetry with the
   macOS Metal fallback: a single nvcc failure left BUILD_OK=false and
   no llama.cpp at all. Generalised the existing Metal -> CPU fallback
   to cover any GPU_BACKEND, so a CUDA configure or build failure now
   transparently retries with the CPU args and the user still ends up
   with a working llama-server.

Pulled the version probe out into _nvcc_meets_llama_minimum so it can
be unit-tested. Added tests/sh/test_nvcc_meets_llama_minimum.sh and two
extra cases in tests/sh/test_get_torch_index_url.sh covering the legacy
"CUDA Version: 13.3" header (driver-reported) and the future 13.7
case. Wired the new test into tests/run_all.sh and the studio-backend
CI workflow.

* tests: relax pr4562 regression to allow generic GPU fallback label

* studio tests: assert setup.sh exports NVCC_PREPEND_FLAGS=-allow-unsupported-compiler

The -allow-unsupported-compiler flag is the core of the fresh-CUDA-toolkit fix
(it lets nvcc accept a host gcc/clang newer than its release-time whitelist, so
CUDA 13.3 day-one builds do not abort on '#error -- unsupported GNU version'),
but it had no automated coverage. Add a source-pattern test asserting the flag
is present, delivered via NVCC_PREPEND_FLAGS so it also covers cmake's CUDA
compiler-id probe, and kept out of CMAKE_ARGS for bash word-splitting safety.

* studio/setup.ps1: allow unsupported host compiler for CUDA build (Windows parity)

Mirror the Linux setup.sh headline fix from this PR on Windows. A freshly
released CUDA toolkit ships with a host-compiler whitelist that lags the
installed toolchain, so nvcc can reject the host with
"#error -- unsupported Microsoft Visual Studio version!" before any real
compile runs (the MSVC analogue of the gcc wall the Linux side hit on
CUDA 13.3). Set NVCC_PREPEND_FLAGS=-allow-unsupported-compiler in the CUDA
build branch so both cmake's configure-time CUDA compiler-id probe and the
cmake --build step proceed. The flag disables the host version check only and
is a no-op when the compiler is already supported.

Set via the process environment (not the $CmakeArgs array), after the
Refresh-Environment calls that re-sanitize CUDA env vars, and appended
idempotently to any value the user already set.

Validated with PowerShell 7.6.2: full setup.ps1 AST parse is clean and the
snippet is idempotent (empty -> set, existing -> append once, no duplicate).
Needs real Windows + CUDA CI to exercise the actual nvcc/MSVC build.

Adds test_setup_ps1_exports_allow_unsupported_compiler asserting the flag is
present, env-delivered, kept out of $CmakeArgs, and scoped to the CUDA-on branch.

* studio: tighten code comments added in this PR

Shorten the verbose multi-line comments and test docstrings introduced by
this PR (setup.sh, setup.ps1, and the shell/python tests) to be succinct
while preserving the rationale. No code or test-assertion changes.
2026-05-29 05:09:20 -07:00

341 lines
12 KiB
Bash
Executable file
Raw Permalink Blame History

#!/bin/bash
# Unit tests for get_torch_index_url() from install.sh
set -e
SCRIPT_DIR="$(cd "$(dirname "$0")" && pwd)"
INSTALL_SH="$SCRIPT_DIR/../../install.sh"
PASS=0
FAIL=0
# Extract get_torch_index_url and its helper functions from install.sh.
# Also replace the hardcoded /usr/bin/nvidia-smi fallback with a
# controllable path so we can test the "no GPU" scenario on GPU machines.
_FUNC_FILE=$(mktemp)
_FAKE_SMI_DIR=$(mktemp -d)
{
sed -n '/^_has_amd_rocm_gpu()/,/^}/p' "$INSTALL_SH"
echo ""
sed -n '/^_has_usable_nvidia_gpu()/,/^}/p' "$INSTALL_SH"
echo ""
sed -n '/^get_torch_index_url()/,/^}/p' "$INSTALL_SH"
} | sed "s|/usr/bin/nvidia-smi|$_FAKE_SMI_DIR/nvidia-smi-absent|g" \
> "$_FUNC_FILE"
# Save system PATH so we always have basic tools (uname, grep, head, etc.)
_SYS_PATH="/usr/local/bin:/usr/bin:/bin"
assert_eq() {
_label="$1"; _expected="$2"; _actual="$3"
if [ "$_actual" = "$_expected" ]; then
echo " PASS: $_label"
PASS=$((PASS + 1))
else
echo " FAIL: $_label (expected '$_expected', got '$_actual')"
FAIL=$((FAIL + 1))
fi
}
# Helper: create a mock nvidia-smi that prints a given CUDA version string.
# Handles both default output (version header) and -L (GPU listing) so that
# _has_usable_nvidia_gpu sees a valid GPU.
make_mock_smi() {
_dir=$(mktemp -d)
cat > "$_dir/nvidia-smi" <<MOCK
#!/bin/sh
case "\$1" in
-L)
echo "GPU 0: NVIDIA GeForce RTX 3090 (UUID: GPU-fake-uuid)"
;;
*)
cat <<'SMI_OUT'
+-----------------------------------------------------------------------------------------+
| NVIDIA-SMI 550.54.15 Driver Version: 550.54.15 CUDA Version: $1 |
+-----------------------------------------------------------------------------------------+
SMI_OUT
;;
esac
MOCK
chmod +x "$_dir/nvidia-smi"
echo "$_dir"
}
# Helper: create a mock nvidia-smi that prints the new "CUDA UMD Version" header
# layout used by newer NVIDIA drivers (e.g. 610.x on Windows). See issue #5812.
make_mock_smi_umd() {
_dir=$(mktemp -d)
cat > "$_dir/nvidia-smi" <<MOCK
#!/bin/sh
case "\$1" in
-L)
echo "GPU 0: NVIDIA GeForce RTX 5090 Laptop GPU (UUID: GPU-fake-uuid)"
;;
*)
cat <<'SMI_OUT'
+-----------------------------------------------------------------------------------------+
| NVIDIA-SMI 610.47 KMD Version: 610.47 CUDA UMD Version: $1 |
+-----------------------------------------------------------------------------------------+
SMI_OUT
;;
esac
MOCK
chmod +x "$_dir/nvidia-smi"
echo "$_dir"
}
# Helper: create a mock amd-smi that prints a given ROCm version string
# Supports both "amd-smi version" and "amd-smi list" subcommands so that
# the GPU presence check (amd-smi list) also succeeds in tests.
make_mock_amd_smi() {
_dir=$(mktemp -d)
cat > "$_dir/amd-smi" <<MOCK
#!/bin/sh
case "\$1" in
list)
printf 'GPU: 0\\n BDF: 0000:03:00.0\\n NAME: gfx1100\\n'
;;
*)
cat <<AMD_OUT
AMDSMI Tool: 25.0.1+2b74356 | AMDSMI Library version: 25.0.1.0 | ROCm version: $1
AMD_OUT
;;
esac
MOCK
chmod +x "$_dir/amd-smi"
echo "$_dir"
}
# Build a minimal tools directory with symlinks to essential commands
# (uname, grep, head, etc.) but WITHOUT nvidia-smi or amd-smi.
_TOOLS_DIR=$(mktemp -d)
for _cmd in uname grep sed head sh bash cat awk printf; do
_real=$(command -v "$_cmd" 2>/dev/null || true)
[ -n "$_real" ] && ln -sf "$_real" "$_TOOLS_DIR/$_cmd"
done
# Helper: run get_torch_index_url with a custom PATH
# $1 = directory with mock nvidia-smi (prepended to PATH), or "none" for no-GPU test
run_func() {
_mock_dir="$1"
if [ "$_mock_dir" = "none" ]; then
# Minimal PATH with only basic tools, no nvidia-smi anywhere
PATH="$_TOOLS_DIR" bash -c ". '$_FUNC_FILE'; get_torch_index_url" 2>/dev/null
else
# Put mock nvidia-smi dir first, then basic tools
PATH="$_mock_dir:$_TOOLS_DIR" bash -c ". '$_FUNC_FILE'; get_torch_index_url" 2>/dev/null
fi
}
echo "=== test_get_torch_index_url ==="
# 1) No nvidia-smi available -> cpu
_result=$(run_func "none")
assert_eq "no nvidia-smi -> cpu" "https://download.pytorch.org/whl/cpu" "$_result"
# 2) CUDA 12.6 -> cu126
_dir=$(make_mock_smi "12.6")
_result=$(run_func "$_dir")
assert_eq "CUDA 12.6 -> cu126" "https://download.pytorch.org/whl/cu126" "$_result"
rm -rf "$_dir"
# 3) CUDA 12.8 -> cu128
_dir=$(make_mock_smi "12.8")
_result=$(run_func "$_dir")
assert_eq "CUDA 12.8 -> cu128" "https://download.pytorch.org/whl/cu128" "$_result"
rm -rf "$_dir"
# 4) CUDA 13.0 -> cu130
_dir=$(make_mock_smi "13.0")
_result=$(run_func "$_dir")
assert_eq "CUDA 13.0 -> cu130" "https://download.pytorch.org/whl/cu130" "$_result"
rm -rf "$_dir"
# 5) CUDA 12.4 -> cu124
_dir=$(make_mock_smi "12.4")
_result=$(run_func "$_dir")
assert_eq "CUDA 12.4 -> cu124" "https://download.pytorch.org/whl/cu124" "$_result"
rm -rf "$_dir"
# 6) CUDA 11.8 -> cu118
_dir=$(make_mock_smi "11.8")
_result=$(run_func "$_dir")
assert_eq "CUDA 11.8 -> cu118" "https://download.pytorch.org/whl/cu118" "$_result"
rm -rf "$_dir"
# 7) CUDA 10.2 (too old) -> cpu
_dir=$(make_mock_smi "10.2")
_result=$(run_func "$_dir")
assert_eq "CUDA 10.2 -> cpu" "https://download.pytorch.org/whl/cpu" "$_result"
rm -rf "$_dir"
# 8) Unparseable nvidia-smi version but valid GPU listing -> cu126 default
_dir=$(mktemp -d)
cat > "$_dir/nvidia-smi" <<'MOCK'
#!/bin/sh
case "$1" in
-L) echo "GPU 0: NVIDIA GeForce RTX 3090 (UUID: GPU-fake-uuid)" ;;
*) echo "something completely unexpected" ;;
esac
MOCK
chmod +x "$_dir/nvidia-smi"
_result=$(run_func "$_dir")
assert_eq "unparseable -> cu126" "https://download.pytorch.org/whl/cu126" "$_result"
rm -rf "$_dir"
# 9) ROCm 6.3 (no nvidia-smi) -> rocm6.3
_dir=$(make_mock_amd_smi "6.3")
_result=$(run_func "$_dir")
assert_eq "ROCm 6.3 -> rocm6.3" "https://download.pytorch.org/whl/rocm6.3" "$_result"
rm -rf "$_dir"
# 10) ROCm 7.1 (no nvidia-smi) -> rocm7.1
_dir=$(make_mock_amd_smi "7.1")
_result=$(run_func "$_dir")
assert_eq "ROCm 7.1 -> rocm7.1" "https://download.pytorch.org/whl/rocm7.1" "$_result"
rm -rf "$_dir"
# 11) ROCm 7.2 (no nvidia-smi) -> rocm7.1 (capped due to torch <2.11.0)
_dir=$(make_mock_amd_smi "7.2")
_result=$(run_func "$_dir")
assert_eq "ROCm 7.2 -> rocm7.1 (capped)" "https://download.pytorch.org/whl/rocm7.1" "$_result"
rm -rf "$_dir"
# 12) Both nvidia-smi and amd-smi present -> CUDA takes precedence
_cuda_dir=$(make_mock_smi "12.6")
_amd_dir=$(make_mock_amd_smi "6.3")
_combined_dir=$(mktemp -d)
ln -sf "$_cuda_dir/nvidia-smi" "$_combined_dir/nvidia-smi"
ln -sf "$_amd_dir/amd-smi" "$_combined_dir/amd-smi"
_result=$(run_func "$_combined_dir")
assert_eq "CUDA+ROCm -> CUDA precedence" "https://download.pytorch.org/whl/cu126" "$_result"
rm -rf "$_cuda_dir" "$_amd_dir" "$_combined_dir"
# 13) No nvidia-smi, no amd-smi -> cpu (duplicate of test 1, confirms ROCm didn't break it)
_result=$(run_func "none")
assert_eq "no GPU -> cpu" "https://download.pytorch.org/whl/cpu" "$_result"
# 14) ROCm 6.1 (no nvidia-smi) -> rocm6.1
_dir=$(make_mock_amd_smi "6.1")
_result=$(run_func "$_dir")
assert_eq "ROCm 6.1 -> rocm6.1" "https://download.pytorch.org/whl/rocm6.1" "$_result"
rm -rf "$_dir"
# 15) ROCm 6.4 (no nvidia-smi) -> rocm6.4
_dir=$(make_mock_amd_smi "6.4")
_result=$(run_func "$_dir")
assert_eq "ROCm 6.4 -> rocm6.4" "https://download.pytorch.org/whl/rocm6.4" "$_result"
rm -rf "$_dir"
# 16) ROCm 7.0 (no nvidia-smi) -> rocm7.0
_dir=$(make_mock_amd_smi "7.0")
_result=$(run_func "$_dir")
assert_eq "ROCm 7.0 -> rocm7.0" "https://download.pytorch.org/whl/rocm7.0" "$_result"
rm -rf "$_dir"
# 17) ROCm 8.0 (future, no nvidia-smi) -> rocm7.1 (capped)
_dir=$(make_mock_amd_smi "8.0")
_result=$(run_func "$_dir")
assert_eq "ROCm 8.0 -> rocm7.1 (capped)" "https://download.pytorch.org/whl/rocm7.1" "$_result"
rm -rf "$_dir"
# 18) Malformed amd-smi output (empty version field) -> cpu
_dir=$(mktemp -d)
cat > "$_dir/amd-smi" <<'MOCK'
#!/bin/sh
echo "AMDSMI Tool: 25.0.1 | AMDSMI Library version: 25.0.1.0 | ROCm version: "
MOCK
chmod +x "$_dir/amd-smi"
_result=$(run_func "$_dir")
assert_eq "empty amd-smi version -> cpu" "https://download.pytorch.org/whl/cpu" "$_result"
rm -rf "$_dir"
# 19) amd-smi with "N/A" version -> cpu
_dir=$(mktemp -d)
cat > "$_dir/amd-smi" <<'MOCK'
#!/bin/sh
echo "AMDSMI Tool: 25.0.1 | AMDSMI Library version: 25.0.1.0 | ROCm version: N/A"
MOCK
chmod +x "$_dir/amd-smi"
_result=$(run_func "$_dir")
assert_eq "N/A amd-smi version -> cpu" "https://download.pytorch.org/whl/cpu" "$_result"
rm -rf "$_dir"
# 20) ROCm version with trailing text (e.g. "6.3.1-beta") -> rocm6.3
_dir=$(make_mock_amd_smi "6.3.1-beta")
_result=$(run_func "$_dir")
assert_eq "ROCm 6.3.1-beta -> rocm6.3" "https://download.pytorch.org/whl/rocm6.3" "$_result"
rm -rf "$_dir"
# 22) CUDA 12.6 still works after ROCm changes (regression check)
_dir=$(make_mock_smi "12.6")
_result=$(run_func "$_dir")
assert_eq "CUDA 12.6 regression -> cu126" "https://download.pytorch.org/whl/cu126" "$_result"
rm -rf "$_dir"
# 23) CUDA 13.0 still works after ROCm changes (regression check)
_dir=$(make_mock_smi "13.0")
_result=$(run_func "$_dir")
assert_eq "CUDA 13.0 regression -> cu130" "https://download.pytorch.org/whl/cu130" "$_result"
rm -rf "$_dir"
# 24) CUDA 12.8 still works after ROCm changes (regression check)
_dir=$(make_mock_smi "12.8")
_result=$(run_func "$_dir")
assert_eq "CUDA 12.8 regression -> cu128" "https://download.pytorch.org/whl/cu128" "$_result"
rm -rf "$_dir"
# 25) UNSLOTH_PYTORCH_MIRROR overrides base URL (CUDA case)
_dir=$(make_mock_smi "12.6")
_result=$(UNSLOTH_PYTORCH_MIRROR="https://mirror.example.com/whl" run_func "$_dir")
assert_eq "mirror env + CUDA 12.6 -> mirror/cu126" "https://mirror.example.com/whl/cu126" "$_result"
rm -rf "$_dir"
# 26) UNSLOTH_PYTORCH_MIRROR overrides base URL (CPU case)
_result=$(UNSLOTH_PYTORCH_MIRROR="https://mirror.example.com/whl" run_func "none")
assert_eq "mirror env + no GPU -> mirror/cpu" "https://mirror.example.com/whl/cpu" "$_result"
# 27) Empty UNSLOTH_PYTORCH_MIRROR falls back to official URL
_result=$(UNSLOTH_PYTORCH_MIRROR="" run_func "none")
assert_eq "empty mirror env -> official/cpu" "https://download.pytorch.org/whl/cpu" "$_result"
# 28) Trailing slash in UNSLOTH_PYTORCH_MIRROR is stripped
_result=$(UNSLOTH_PYTORCH_MIRROR="https://mirror.example.com/whl/" run_func "none")
assert_eq "trailing slash stripped -> mirror/cpu" "https://mirror.example.com/whl/cpu" "$_result"
# 29) "CUDA UMD Version: 13.3" header (newer NVIDIA driver layout, issue #5812)
# -> cu130, not the cu126 fallback.
_dir=$(make_mock_smi_umd "13.3")
_result=$(run_func "$_dir")
assert_eq "CUDA UMD Version 13.3 -> cu130" "https://download.pytorch.org/whl/cu130" "$_result"
rm -rf "$_dir"
# 30) "CUDA UMD Version: 12.8" header (newer layout on a 12.x driver) -> cu128
_dir=$(make_mock_smi_umd "12.8")
_result=$(run_func "$_dir")
assert_eq "CUDA UMD Version 12.8 -> cu128" "https://download.pytorch.org/whl/cu128" "$_result"
rm -rf "$_dir"
# 31) "CUDA UMD Version: 11.8" header (newer layout on an older driver) -> cu118
_dir=$(make_mock_smi_umd "11.8")
_result=$(run_func "$_dir")
assert_eq "CUDA UMD Version 11.8 -> cu118" "https://download.pytorch.org/whl/cu118" "$_result"
rm -rf "$_dir"
# 32) Driver-reported "CUDA Version: 13.3" (legacy header) -> cu130.
_dir=$(make_mock_smi "13.3")
_result=$(run_func "$_dir")
assert_eq "CUDA Version 13.3 -> cu130" "https://download.pytorch.org/whl/cu130" "$_result"
rm -rf "$_dir"
# 33) "CUDA Version: 13.7" -> cu130 (until a cu137 wheel index exists).
_dir=$(make_mock_smi "13.7")
_result=$(run_func "$_dir")
assert_eq "CUDA Version 13.7 -> cu130" "https://download.pytorch.org/whl/cu130" "$_result"
rm -rf "$_dir"
rm -f "$_FUNC_FILE"
rm -rf "$_FAKE_SMI_DIR"
rm -rf "$_TOOLS_DIR"
echo ""
echo "Results: $PASS passed, $FAIL failed"
[ "$FAIL" -eq 0 ] || exit 1
Reference in a new issue View git blame Copy permalink