ROCm-on-WSL: support discrete Radeon (RDNA 3/4) in WSL, not just Strix Halo

install.sh

@@ -1483,6 +1483,81 @@ elif [ "$OS" = "macos" ]; then
 fi
 tauri_diag_marker "$_TAURI_INITIAL_GPU_BRANCH" "none"
 
+# AMD GPU name from the Windows host via WMI, or empty. Discrete cards aren't in
+# /proc/cpuinfo, so ask Windows. Cached ("-" = negative), self-contained, bounded
+# to 10s. Defined here so the reroute below can use it before _run_bounded exists.
+_WSL_AMD_GPU_NAME_CACHE=""
+_wsl_amd_gpu_name() {
+    if [ -n "$_WSL_AMD_GPU_NAME_CACHE" ]; then
+        [ "$_WSL_AMD_GPU_NAME_CACHE" = "-" ] && return 1
+        printf '%s' "$_WSL_AMD_GPU_NAME_CACHE"; return 0
+    fi
+    command -v powershell.exe >/dev/null 2>&1 || { _WSL_AMD_GPU_NAME_CACHE="-"; return 1; }
+    _wag_ps="(Get-CimInstance Win32_VideoController | Where-Object { \$_.Name -match 'AMD|Radeon' } | Select-Object -First 1).Name"
+    if command -v timeout >/dev/null 2>&1; then
+        _wag_n="$(timeout 10 powershell.exe -NoProfile -Command "$_wag_ps" 2>/dev/null | tr -d '\r\n\000')"
+    else
+        _wag_n="$(powershell.exe -NoProfile -Command "$_wag_ps" 2>/dev/null | tr -d '\r\n\000')"
+    fi
+    if [ -n "$_wag_n" ]; then _WSL_AMD_GPU_NAME_CACHE="$_wag_n"; printf '%s' "$_wag_n"; return 0; fi
+    _WSL_AMD_GPU_NAME_CACHE="-"; return 1
+}
+
+# ── Bounded command runner ──
+# Runs a command under a 10s timeout when the `timeout` binary is available,
+# otherwise runs it unbounded. Keeps a wedged nvidia-smi (blocking during
+# driver init or after a reset) from hanging the installer: a timed-out probe
+# exits nonzero and is treated exactly like a failed probe. No-op semantics on
+# hosts without `timeout` (e.g. macOS) or when the probe is healthy.
+_run_bounded() {
+    if command -v timeout >/dev/null 2>&1; then
+        timeout 10 "$@"
+    else
+        "$@"
+    fi
+}
+
+# Returns 0 (true) when CUDA_VISIBLE_DEVICES is set to "" or "-1", i.e. every
+# NVIDIA device is deliberately hidden (mixed AMD+NVIDIA hosts steering work to
+# the AMD card). Unset means all devices visible. nvidia-smi ignores this env
+# var, so the probes below cannot see the distinction on their own.
+_cvd_hides_nvidia() {
+    [ "${CUDA_VISIBLE_DEVICES+set}" = "set" ] || return 1
+    _cvd_trim=$(printf '%s' "$CUDA_VISIBLE_DEVICES" | tr -d '[:space:]')
+    [ -z "$_cvd_trim" ] || [ "$_cvd_trim" = "-1" ]
+}
+
+# ── NVIDIA usable-GPU helper ──
+# Returns 0 (true) if an NVIDIA GPU is present and usable.
+# Primary probe: nvidia-smi -L. Fallback: /proc/driver/nvidia/gpus/ sysfs,
+# which the NVIDIA driver populates on Linux regardless of nvidia-smi state
+# -- handles PATH gaps, subprocess timeouts, and driver init races that
+# could otherwise cause nvidia-smi to fail and silence NVIDIA detection.
+# A GPU hidden via CUDA_VISIBLE_DEVICES=""/-1 counts as NOT usable (matches
+# install_llama_prebuilt.py has_usable_nvidia), so AMD/CPU routing still runs.
+_has_usable_nvidia_gpu() {
+    if _cvd_hides_nvidia; then
+        return 1
+    fi
+    _nvsmi=""
+    if command -v nvidia-smi >/dev/null 2>&1; then
+        _nvsmi="nvidia-smi"
+    elif [ -x "/usr/bin/nvidia-smi" ]; then
+        _nvsmi="/usr/bin/nvidia-smi"
+    fi
+    if [ -n "$_nvsmi" ]; then
+        if _run_bounded "$_nvsmi" -L 2>/dev/null | awk '/^GPU[[:space:]]+[0-9]+:/{found=1} END{exit !found}'; then
+            return 0
+        fi
+    fi
+    # Fallback: NVIDIA driver exposes one subdir per GPU under this path.
+    if [ -d /proc/driver/nvidia/gpus ] && \
+       [ -n "$(ls -A /proc/driver/nvidia/gpus 2>/dev/null)" ]; then
+        return 0
+    fi
+    return 1
+}
+
 # Strix Halo ROCm-on-WSL only targets Ubuntu 24.04. On a newer distro (e.g. 26.04)
 # with a 24.04 distro present, re-run the install there and stop; else fall through
 # to CPU + the `wsl --install` hint below (never auto-create a distro). Runs before
@@ -1493,7 +1568,15 @@ _maybe_reroute_strixhalo_to_2404() {
     [ "${UNSLOTH_SKIP_ROCM_WSL_SETUP:-0}" = "1" ] && return 0
     [ "${UNSLOTH_WSL_REROUTED:-0}" = "1" ] && return 0
     [ -e /dev/dxg ] || return 0
-    grep -qiE 'Ryzen AI Max|Radeon 80[0-9]0S|Strix Halo' /proc/cpuinfo 2>/dev/null || return 0
+    # A usable NVIDIA GPU (common on hybrid AMD+NVIDIA hosts) means the CUDA path works on
+    # this distro, so don't reroute for AMD. _has_usable_nvidia_gpu (moved above) honors
+    # CUDA_VISIBLE_DEVICES=""/-1 and the /proc/driver/nvidia fallback for PATH/timeout gaps.
+    if _has_usable_nvidia_gpu; then return 0; fi
+    # Strix APUs show in /proc/cpuinfo; discrete cards don't, so also try WMI. Either reroutes.
+    if ! grep -qiE 'Ryzen AI Max|Radeon 80[0-9]0S|Strix Halo' /proc/cpuinfo 2>/dev/null \
+       && ! _wsl_amd_gpu_name >/dev/null 2>&1; then
+        return 0
+    fi
     # Already ROCm-on-WSL? leave a working GPU alone, whatever the version.
     if [ -e /opt/rocm/lib/librocdxg.so ] || [ -e /opt/rocm/lib64/librocdxg.so ]; then
         return 0
@@ -1959,61 +2042,6 @@ _has_amd_rocm_gpu() {
     return 1
 }
 
-# ── Bounded command runner ──
-# Runs a command under a 10s timeout when the `timeout` binary is available,
-# otherwise runs it unbounded. Keeps a wedged nvidia-smi (blocking during
-# driver init or after a reset) from hanging the installer: a timed-out probe
-# exits nonzero and is treated exactly like a failed probe. No-op semantics on
-# hosts without `timeout` (e.g. macOS) or when the probe is healthy.
-_run_bounded() {
-    if command -v timeout >/dev/null 2>&1; then
-        timeout 10 "$@"
-    else
-        "$@"
-    fi
-}
-
-# Returns 0 (true) when CUDA_VISIBLE_DEVICES is set to "" or "-1", i.e. every
-# NVIDIA device is deliberately hidden (mixed AMD+NVIDIA hosts steering work to
-# the AMD card). Unset means all devices visible. nvidia-smi ignores this env
-# var, so the probes below cannot see the distinction on their own.
-_cvd_hides_nvidia() {
-    [ "${CUDA_VISIBLE_DEVICES+set}" = "set" ] || return 1
-    _cvd_trim=$(printf '%s' "$CUDA_VISIBLE_DEVICES" | tr -d '[:space:]')
-    [ -z "$_cvd_trim" ] || [ "$_cvd_trim" = "-1" ]
-}
-
-# ── NVIDIA usable-GPU helper ──
-# Returns 0 (true) if an NVIDIA GPU is present and usable.
-# Primary probe: nvidia-smi -L. Fallback: /proc/driver/nvidia/gpus/ sysfs,
-# which the NVIDIA driver populates on Linux regardless of nvidia-smi state
-# -- handles PATH gaps, subprocess timeouts, and driver init races that
-# could otherwise cause nvidia-smi to fail and silence NVIDIA detection.
-# A GPU hidden via CUDA_VISIBLE_DEVICES=""/-1 counts as NOT usable (matches
-# install_llama_prebuilt.py has_usable_nvidia), so AMD/CPU routing still runs.
-_has_usable_nvidia_gpu() {
-    if _cvd_hides_nvidia; then
-        return 1
-    fi
-    _nvsmi=""
-    if command -v nvidia-smi >/dev/null 2>&1; then
-        _nvsmi="nvidia-smi"
-    elif [ -x "/usr/bin/nvidia-smi" ]; then
-        _nvsmi="/usr/bin/nvidia-smi"
-    fi
-    if [ -n "$_nvsmi" ]; then
-        if _run_bounded "$_nvsmi" -L 2>/dev/null | awk '/^GPU[[:space:]]+[0-9]+:/{found=1} END{exit !found}'; then
-            return 0
-        fi
-    fi
-    # Fallback: NVIDIA driver exposes one subdir per GPU under this path.
-    if [ -d /proc/driver/nvidia/gpus ] && \
-       [ -n "$(ls -A /proc/driver/nvidia/gpus 2>/dev/null)" ]; then
-        return 0
-    fi
-    return 1
-}
-
 # ── Detect GPU and choose PyTorch index URL ──
 # Mirrors Get-TorchIndexUrl in install.ps1.
 # On CPU-only machines this returns the cpu index, avoiding the solver
@@ -2327,19 +2355,19 @@ _persist_rocm_wsl_dropin() {
     fi
 }
 
+# _wsl_amd_gpu_name is defined earlier so both the reroute and this bootstrap can use it.
 _maybe_bootstrap_rocm_wsl() {
     [ "${OS:-}" = "wsl" ] || return 0
     [ "${SKIP_TORCH:-false}" = "false" ] || return 0
     [ "${UNSLOTH_SKIP_ROCM_WSL_SETUP:-0}" = "1" ] && return 0
     # Leave any already-usable GPU completely alone (NVIDIA, or working ROCm).
     if _has_usable_nvidia_gpu; then return 0; fi
-    # "Usable ROCm" here = rocminfo enumerates the gfx1151 agent. Don't use the
-    # generic _has_amd_rocm_gpu: its broad gfx match accepts "gfx11-generic" and
-    # would skip this bootstrap while the real GPU is still unusable. awk consumes
-    # all input, so rocminfo isn't SIGPIPE'd like `grep -q` would under pipefail.
+    # Usable ROCm = rocminfo enumerates a real GPU agent: gfx[1-9] (excludes gfx000,
+    # the CPU agent) and not the "gfx11-generic" fallback. awk consumes all input so
+    # rocminfo isn't SIGPIPE'd like `grep -q` under pipefail.
     _ensure_rocm_probe_env
     if command -v rocminfo >/dev/null 2>&1 && \
-       rocminfo 2>/dev/null | awk '/Name:[[:space:]]*gfx1151/{found=1} END{exit !found}'; then
+       rocminfo 2>/dev/null | awk '/Name:[[:space:]]*gfx[1-9]/ && !/generic/{found=1} END{exit !found}'; then
         # rocminfo may work only via the transient env _ensure_rocm_probe_env
         # just set, which dies with the installer. Persist the drop-in so login
         # shells (Studio, llama.cpp) inherit it -- else a reinstall over an
@@ -2349,9 +2377,12 @@ _maybe_bootstrap_rocm_wsl() {
     fi
     # WSL GPU passthrough device must exist (present on any WSL2 GPU host).
     [ -e /dev/dxg ] || return 0
-    # Only Strix Halo (gfx1151): rocminfo can't tell us the arch yet, so match
-    # the CPU model string WSL exposes (e.g. "AMD Ryzen AI Max+ ... Radeon 8060S").
-    grep -qiE 'Ryzen AI Max|Radeon 80[0-9]0S|Strix Halo' /proc/cpuinfo 2>/dev/null || return 0
+    # Strix APUs show in /proc/cpuinfo (the CPU model); discrete cards don't, so also
+    # ask the Windows host. Either signal suffices; the bootstrap detects arch from rocminfo.
+    if ! grep -qiE 'Ryzen AI Max|Radeon 80[0-9]0S|Strix Halo' /proc/cpuinfo 2>/dev/null \
+       && ! _wsl_amd_gpu_name >/dev/null 2>&1; then
+        return 0
+    fi
     command -v bash >/dev/null 2>&1 || return 0
 
     # Fast path: already configured (librocdxg present) but launched from a
@@ -2369,7 +2400,8 @@ _maybe_bootstrap_rocm_wsl() {
     fi
 
     echo ""
-    substep "Detected AMD Strix Halo (Radeon 8000S) in WSL with no ROCm runtime yet." "$C_WARN"
+    _rw_gpu="$(_wsl_amd_gpu_name 2>/dev/null || true)"; [ -n "$_rw_gpu" ] || _rw_gpu="an AMD GPU"
+    substep "Detected ${_rw_gpu} in WSL with no ROCm runtime yet." "$C_WARN"
     substep "Setting up ROCm-on-WSL (ROCm 7.2 + librocdxg) automatically to enable this GPU."
     substep "One-time, uses sudo and a large download. (skip: re-run with UNSLOTH_SKIP_ROCM_WSL_SETUP=1)"
 

scripts/install_rocm_wsl_strixhalo.sh

@@ -3,13 +3,14 @@
 # Copyright 2026-present the Unsloth AI Inc. team. All rights reserved.
 #
 # ──────────────────────────────────────────────────────────────────────────────
-# Enable ROCm-on-WSL for AMD Strix Halo (Radeon 8060S / gfx1151)
+# Enable ROCm-on-WSL for AMD GPUs (Strix Halo/Point APUs AND discrete Radeon RX
+# 7000/9000). Verified on gfx1151 (Radeon 8060S) and gfx1200 (Radeon RX 9060 XT).
 # ──────────────────────────────────────────────────────────────────────────────
-# install.sh already routes gfx1151 to the right ROCm wheels once a ROCm runtime
-# is present; what it does NOT do is install AMD's ROCm userspace + the WSL DXG
-# bridge. This helper automates that Linux-side prerequisite on Ubuntu 24.04
-# WSL2 and is invoked by install.sh when it sees a Strix Halo APU in WSL (via
-# /dev/dxg) but no ROCm runtime yet. Fully idempotent (re-run just re-verifies).
+# install.sh routes the detected arch to the right ROCm wheels once a runtime exists;
+# what it does NOT do is install AMD's ROCm userspace + the WSL DXG bridge (librocdxg).
+# This helper does that Linux-side prerequisite on Ubuntu 24.04 WSL2, invoked by
+# install.sh when it sees an AMD GPU via /dev/dxg but no ROCm yet. Arch-agnostic: the
+# arch is auto-detected from rocminfo (override UNSLOTH_WSL_GFX=gfx1200). Idempotent.
 #
 # Manual, admin-gated Windows prerequisite: an AMD Adrenalin driver with
 # production ROCDXG/WSL support (26.2.2+). install.ps1 offers to update it. Once
@@ -34,10 +35,12 @@ set -euo pipefail
 
 # ── Tunables (override via env) ──────────────────────────────────────────────
 ROCM_VER="${UNSLOTH_WSL_ROCM_VER:-7.2.1}"            # ROCm release to install
-GFX="gfx1151"
+# GPU arch: empty = auto-detect from rocminfo after install (override UNSLOTH_WSL_GFX=gfx1200).
+# The ROCm + librocdxg setup is arch-agnostic; only verify + the smoke test need the arch.
+GFX="${UNSLOTH_WSL_GFX:-}"
 LIBROCDXG_REF="${UNSLOTH_LIBROCDXG_REF:-develop}"    # ROCm/librocdxg git ref to build
-# AMD's gfx1151 wheel index (same one install.sh uses); only for the smoke test.
-TORCH_INDEX="${UNSLOTH_AMD_ROCM_MIRROR:-https://repo.amd.com/rocm/whl}/${GFX}/"
+# AMD's wheel index for the (optional) smoke test; resolved after arch detection.
+TORCH_INDEX=""
 # Optional torch smoke test (throwaway venv). OFF by default: install.sh installs
 # torch itself into the real venv right after, so a duplicate download is wasteful.
 SMOKE_TEST="${UNSLOTH_WSL_SMOKE_TEST:-0}"
@@ -220,12 +223,12 @@ $SUDO ldconfig
 say "Persisting ROCm-on-WSL environment"
 _envfile="/etc/profile.d/unsloth-rocm-wsl.sh"
 $SUDO tee "$_envfile" >/dev/null <<EOF
-# >>> Unsloth ROCm-on-WSL (gfx1151) >>>
+# >>> Unsloth ROCm-on-WSL >>>
 export HSA_ENABLE_DXG_DETECTION=1
 export TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL=1
 export PATH="${ROCM_DIR}/bin:\${PATH}"
 export LD_LIBRARY_PATH="${ROCM_DIR}/lib:\${LD_LIBRARY_PATH:-}"
-# <<< Unsloth ROCm-on-WSL (gfx1151) <<<
+# <<< Unsloth ROCm-on-WSL <<<
 EOF
 # also drop into ~/.bashrc for interactive shells
 if [ -n "${HOME:-}" ] && ! grep -q "Unsloth ROCm-on-WSL" "${HOME}/.bashrc" 2>/dev/null; then
@@ -237,32 +240,50 @@ export PATH="${ROCM_DIR}/bin:${PATH}"
 export LD_LIBRARY_PATH="${ROCM_DIR}/lib:${LD_LIBRARY_PATH:-}"
 
 # ── Step 5: verify the runtime enumerates the GPU ────────────────────────────
-say "Verifying rocminfo sees ${GFX}"
+say "Verifying rocminfo enumerates the GPU over DXG"
 # Capture rocminfo into a var BEFORE grepping: piping into `grep -q` SIGPIPEs
 # rocminfo on first match, which under `set -o pipefail` turns a successful match
-# into a pipeline failure. Match the gfx1151 ISA "Name:" agent exactly (not a
-# broad gfx1[0-9]) so a generic fallback ISA or unrelated RDNA GPU can't pass.
+# into a pipeline failure.
 _rocminfo_out="$(rocminfo 2>/dev/null || true)"
-if ! printf '%s\n' "$_rocminfo_out" | grep -qE "Name:[[:space:]]*${GFX}([^0-9]|$)"; then
+# GPU agents advertise an ISA "Name: gfxNNNN". Match gfx[1-9] (excludes gfx000, the CPU
+# agent), drop the "gfx*-generic" fallback ISA, and take the first real GPU arch.
+_detected_gfx="$(printf '%s\n' "$_rocminfo_out" | grep -E 'Name:[[:space:]]*gfx[1-9]' | grep -v 'generic' | grep -oE 'gfx[1-9][0-9a-z]*' | head -1 || true)"
+if [ -z "$_detected_gfx" ]; then
     printf '%s\n' "$_rocminfo_out" | head -25 >&2 || true
-    die "rocminfo did not enumerate a ${GFX} GPU agent. Most common cause: the Windows AMD driver predates production ROCDXG -- update Adrenalin (install.ps1 offers this), reboot, and re-run."
+    die "rocminfo did not enumerate any GPU agent. Most common cause: the Windows AMD driver predates production ROCDXG -- update Adrenalin (install.ps1 offers this), reboot, and re-run."
 fi
+# Honour a caller-pinned arch (sanity-check via a consuming grep, not grep -q: under
+# pipefail -q would SIGPIPE printf on large output and misreport the arch); else adopt.
+if [ -n "$GFX" ] && ! printf '%s\n' "$_rocminfo_out" | grep -E "Name:[[:space:]]*${GFX}([^0-9]|$)" >/dev/null; then
+    die "rocminfo enumerated '${_detected_gfx}' but not the requested UNSLOTH_WSL_GFX='${GFX}'."
+fi
+GFX="${GFX:-$_detected_gfx}"
 # Display-only summary: best-effort (|| true) so head's early pipe-close under
 # `set -o pipefail` can't fail the bootstrap after verification already passed.
 printf '%s\n' "$_rocminfo_out" | grep -E 'Marketing Name|Device Type|Compute Unit' | grep -iE "Radeon|GPU|Compute" | head -3 || true
 note "ROCm-on-WSL runtime is live for ${GFX}."
 
-# ── Step 6 (optional): torch smoke test from the gfx1151 index ───────────────
+# ── Step 6 (optional): torch smoke test from AMD's per-arch wheel index ───────
 if [ "$SMOKE_TEST" = "1" ]; then
     say "Smoke-testing PyTorch on ${GFX} (throwaway venv)"
+    # Map the detected arch to AMD's repo.amd.com wheel family index.
+    case "$GFX" in
+        gfx1200|gfx1201)                 _fam="gfx120X-all" ;;
+        gfx1100|gfx1101|gfx1102|gfx1103) _fam="gfx110X-all" ;;
+        *)                               _fam="$GFX" ;;   # gfx1150/gfx1151/gfx90a: own index
+    esac
+    TORCH_INDEX="${UNSLOTH_AMD_ROCM_MIRROR:-https://repo.amd.com/rocm/whl}/${_fam}/"
     _venv="${HOME}/.unsloth/rocm-smoketest"
     rm -rf "$_venv"; python3 -m venv "$_venv"
     "$_venv/bin/pip" install --quiet --upgrade pip
-    # gfx1151 index is primary (torch + triton); PyPI only an extra for pure-py
+    # AMD arch index is primary (torch + triton); PyPI only an extra for pure-py
     # deps. The constraint keeps pip on the ROCm wheel, not a newer PyPI CUDA torch.
     "$_venv/bin/pip" install --index-url "$TORCH_INDEX" \
         --extra-index-url https://pypi.org/simple "$TORCH_CONSTRAINT" || \
         die "torch install from ${TORCH_INDEX} failed."
+    # WSL: torch's bundled ROCr must load the DXG bridge -- drop librocdxg into torch/lib.
+    _tlib="$("$_venv/bin/python" -c 'import torch,os;print(os.path.join(os.path.dirname(torch.__file__),"lib"))' 2>/dev/null || true)"
+    [ -d "$_tlib" ] && cp -f "${ROCM_DIR}"/lib/librocdxg.so* "$_tlib"/ 2>/dev/null || true
     "$_venv/bin/python" - <<'PY'
 import torch
 ok = torch.cuda.is_available()