diff --git a/install.ps1 b/install.ps1
index d2312c98ec..af51ae3604 100644
--- a/install.ps1
+++ b/install.ps1
@@ -48,8 +48,8 @@ function Install-UnslothStudio {
         }
     }
 
-    # Ref for fetching install assets (provision_llama_cuda.sh, the .ico) from
-    # raw.githubusercontent.com; UNSLOTH_INSTALL_REF overrides 'main' for pre-merge testing.
+    # raw.githubusercontent.com ref for install assets (provision_llama_cuda.sh, .ico).
+    # UNSLOTH_INSTALL_REF overrides 'main' for pre-merge testing.
     function Get-UnslothInstallRef {
         if ($env:UNSLOTH_INSTALL_REF -and $env:UNSLOTH_INSTALL_REF.Trim()) { return $env:UNSLOTH_INSTALL_REF.Trim() }
         return 'main'
@@ -97,9 +97,8 @@ function Install-UnslothStudio {
         if ($TauriMode) {
             exit $Code
         }
-        # -File runs exit 0 on a plain return regardless of $LASTEXITCODE, so `exit`
-        # must carry the code there; under `irm | iex` (no $PSCommandPath) `exit`
-        # would kill the user's shell, so fall through.
+        # -File ignores $LASTEXITCODE on plain return, so `exit` must carry the code;
+        # under `irm | iex` (no $PSCommandPath) `exit` would kill the user's shell.
         if ($PSCommandPath) {
             exit $Code
         }
@@ -1774,13 +1773,13 @@ shell.Run cmd, 0, False
     $TorchIndexUrl = Get-TorchIndexUrl
 
     # ===== Windows-on-ARM + NVIDIA GPU -> automatic WSL2 fallback (N1X "RTX Spark" / DGX Spark-class) =====
-    # win_arm64 has no CUDA PyTorch/Triton wheel, so run the Linux installer inside WSL2 (full GPU) and
-    # add a Windows `unsloth` shim that forwards into it. x86_64 / ARM64-without-NVIDIA unaffected; if a
-    # win_arm64 CUDA torch wheel ever ships, the probe below keeps the native install automatically.
-    # Opt out: UNSLOTH_NO_WSL_FALLBACK=1; choose the distro with UNSLOTH_WSL_DISTRO.
+    # win_arm64 has no CUDA PyTorch/Triton wheel, so run the Linux installer inside WSL2 (full GPU) plus
+    # a Windows `unsloth` shim that forwards into it. x86_64 / ARM64-without-NVIDIA unaffected; the probe
+    # below keeps the native install if a win_arm64 CUDA torch wheel ever ships.
+    # Opt out: UNSLOTH_NO_WSL_FALLBACK=1; pick distro with UNSLOTH_WSL_DISTRO.
     try { $_winArm64 = ([System.Runtime.InteropServices.RuntimeInformation]::OSArchitecture.ToString() -ieq 'Arm64') } catch { $_winArm64 = $false }
-    # x64-emulated PS on ARM reports X64/AMD64 via .NET and $env:; Win32_Processor.Architecture
-    # (12=ARM64) and machine-level PROCESSOR_ARCHITECTURE read the true arch. Only ever turns $_winArm64 ON.
+    # x64-emulated PS on ARM reports X64/AMD64; Win32_Processor.Architecture (12=ARM64) and machine-level
+    # PROCESSOR_ARCHITECTURE read the true arch. Only ever turns $_winArm64 ON.
     if (-not $_winArm64) {
         try { if ((@(Get-CimInstance Win32_Processor -ErrorAction Stop))[0].Architecture -eq 12) { $_winArm64 = $true } } catch {}
     }
@@ -1793,10 +1792,10 @@ shell.Run cmd, 0, False
     $_nativeCudaTorchOk = $false
     if ($_winArm64 -and $HasNvidiaSmi -and (-not $SkipTorch)) {
         # Probe with the SAME spec as the real install ("torch>=2.4,<2.11.0"): a bare `torch` probe
-        # can match an out-of-range wheel, skipping WSL only to fail the real pinned install.
+        # could match an out-of-range wheel, skipping WSL only to fail the real pinned install.
         $prevEapProbe = $ErrorActionPreference; $ErrorActionPreference = "Continue"
-        # --reinstall: an already-installed (e.g. CPU-only) torch must not satisfy the probe --
-        # it has to prove a native win_arm64 CUDA wheel exists on the index.
+        # --reinstall: an installed (e.g. CPU-only) torch mustn't satisfy the probe -- it must
+        # prove a native win_arm64 CUDA wheel exists on the index.
         $global:LASTEXITCODE = -1
         try {
             & uv pip install --python $VenvPython --dry-run --reinstall "torch>=2.4,<2.11.0" --index-url $TorchIndexUrl *> $null
@@ -1809,15 +1808,14 @@ shell.Run cmd, 0, False
         substep "no win_arm64 CUDA PyTorch/Triton yet; WSL2 delivers full GPU (DGX Spark / RTX Spark path)." "Yellow"
 
         # The Tauri desktop app launches its backend from a Windows venv (resolve_backend_binary),
-        # not WSL, so a WSL-only install would report complete yet fail to start -- send those
-        # users to the CLI installer.
+        # not WSL, so a WSL-only install would start nothing -- send those users to the CLI installer.
         if ($TauriMode) {
             return (Exit-InstallFailure "Windows-on-ARM + NVIDIA GPU needs the WSL2 GPU install, which the desktop app can't launch yet. Install from PowerShell instead:  irm https://unsloth.ai/install.ps1 | iex" 1)
         }
 
         $wslReady = $false
         if (Get-Command wsl.exe -ErrorAction SilentlyContinue) {
-            # Reset first: a stale 0 in $LASTEXITCODE would wrongly mark WSL ready if wsl.exe fails to start.
+            # Reset: a stale 0 would wrongly mark WSL ready if wsl.exe fails to start.
             $global:LASTEXITCODE = -1
             try { & wsl.exe --status *> $null; if ($LASTEXITCODE -eq 0) { $wslReady = $true } } catch {}
         }
@@ -1835,8 +1833,8 @@ shell.Run cmd, 0, False
                 substep "reboot, then re-run:                 irm https://unsloth.ai/install.ps1 | iex" "Cyan"
             }
             # Deferred until reboot: restore any rolled-aside previous venv and signal not-complete.
-            # A plain return exits 0 for -File runs regardless of $LASTEXITCODE, so `exit 1` there;
-            # under `irm | iex` ($PSCommandPath empty) exit would kill the user's shell, so return.
+            # `exit 1` for -File (plain return exits 0); under `irm | iex` (no $PSCommandPath) return
+            # instead, since exit would kill the user's shell.
             Restore-StudioVenvRollback
             $global:LASTEXITCODE = 1
             if ($PSCommandPath) { exit 1 }
@@ -1844,9 +1842,8 @@ shell.Run cmd, 0, False
         }
 
         $distro = if ($env:UNSLOTH_WSL_DISTRO) { $env:UNSLOTH_WSL_DISTRO } else { "Ubuntu-24.04" }
-        # For cmd-context uses (.cmd shim, copy-paste hints): wsl.exe rejects a QUOTED space-free
-        # name (WSL_E_DISTRO_NOT_FOUND, verified on 2.x) yet splits a bare spaced one after -d --
-        # so quote ONLY when the name contains whitespace.
+        # For cmd-context uses (.cmd shim, copy-paste hints): wsl.exe rejects a QUOTED space-free name
+        # (WSL_E_DISTRO_NOT_FOUND on 2.x) yet splits a bare spaced one after -d, so quote ONLY when spaced.
         $_distroArg = if ($distro -match '\s') { '"' + $distro + '"' } else { $distro }
         # Detect the distro by exit code (encoding-proof; wsl --list emits UTF-16 that PS mis-parses).
         $haveDistro = $false
@@ -1857,9 +1854,8 @@ shell.Run cmd, 0, False
             try { & wsl.exe --install -d $distro --no-launch } catch {}
         } else {
             # A PRE-EXISTING distro may be WSL1 (no GPU passthrough; would only fail at the final
-            # torch.cuda check). Detect from inside the distro (encoding-proof, unlike UTF-16
-            # `wsl -l -v`) and convert in place -- `wsl --set-version` preserves the files.
-            # Fresh installs default to WSL2, so only the pre-existing case needs this.
+            # torch.cuda check). Detect from inside (encoding-proof, unlike UTF-16 `wsl -l -v`) and
+            # convert in place -- `wsl --set-version` preserves files. Fresh installs default to WSL2.
             $_wsl2Probe = 'grep -qiE ''microsoft-standard|WSL2'' /proc/version 2>/dev/null || test -e /usr/lib/wsl/lib/libcuda.so'
             $_isWsl2 = $false
             $global:LASTEXITCODE = -1
@@ -1878,14 +1874,14 @@ shell.Run cmd, 0, False
             }
         }
         substep "installing Unsloth Studio inside WSL '$distro' with full GPU (this downloads PyTorch)..." "Cyan"
-        # Non-main ref: fetch + export THAT ref so the WSL venv gets the branch's setup.sh +
-        # patches (else install.sh pulls PyPI unsloth). main == plain unsloth.ai/install.sh.
+        # Non-main ref: fetch + export THAT ref so the WSL venv gets the branch's setup.sh + patches
+        # (else install.sh pulls PyPI unsloth). main == plain unsloth.ai/install.sh.
         $_instRef = Get-UnslothInstallRef
-        # UNSLOTH_WSL_LLAMA_DEFERRED=1: setup.sh skips its foreground CUDA llama.cpp build because
-        # install.ps1 builds it in the background (a DIRECT install.sh run in WSL doesn't set it).
-        # apt stderr stays visible (only stdout -> /dev/null) so network/repo failures are diagnosable.
-        # Forward UNSLOTH_NO_LLAMA_CUDA into WSL: the same opt-out skips the dispatch below, so
-        # unforwarded, setup.sh would defer to a background builder that never starts (no llama-server).
+        # UNSLOTH_WSL_LLAMA_DEFERRED=1: setup.sh skips its foreground CUDA llama.cpp build since we build
+        # it in the background (a DIRECT install.sh run in WSL doesn't set it). apt stderr stays visible
+        # (only stdout -> /dev/null) so network/repo failures are diagnosable.
+        # Forward UNSLOTH_NO_LLAMA_CUDA into WSL: it also skips the dispatch below, so unforwarded
+        # setup.sh would defer to a background builder that never starts (no llama-server).
         $_fwdEnv = ''
         if ($env:UNSLOTH_NO_LLAMA_CUDA -eq '1') { $_fwdEnv = 'export UNSLOTH_NO_LLAMA_CUDA=1; ' }
         if ($_instRef -eq 'main') {
@@ -1894,7 +1890,7 @@ shell.Run cmd, 0, False
             $wslInstall = $_fwdEnv + 'export DEBIAN_FRONTEND=noninteractive UNSLOTH_WSL_LLAMA_DEFERRED=1; export UNSLOTH_INSTALL_REF=' + $_instRef + '; apt-get update -y >/dev/null; apt-get install -y build-essential cmake git curl pciutils libcurl4-openssl-dev >/dev/null; curl -fsSL https://raw.githubusercontent.com/unslothai/unsloth/' + $_instRef + '/install.sh | sh'
         }
         # install.sh may exit non-zero on the optional llama.cpp prebuilt step (no aarch64 prebuilt)
-        # though torch + unsloth + Studio still install, so lower EAP so it doesn't abort under Stop.
+        # though torch + unsloth + Studio still install; lower EAP so it doesn't abort under Stop.
         $prevEapWsl = $ErrorActionPreference
         $ErrorActionPreference = "Continue"
         $global:LASTEXITCODE = -1
@@ -1909,15 +1905,15 @@ shell.Run cmd, 0, False
         $torchOk = $false
         $prevEapChk = $ErrorActionPreference
         $ErrorActionPreference = "Continue"
-        # Reset first so a stale 0 from a prior command can't mark torch OK if this fails to launch.
+        # Reset so a stale 0 can't mark torch OK if this fails to launch.
         $global:LASTEXITCODE = -1
         try {
             & wsl.exe -d $distro --cd /root -u root -- /root/.unsloth/studio/unsloth_studio/bin/python -c "import torch,sys; sys.exit(0 if torch.cuda.is_available() else 3)" *> $null
             $torchOk = ($LASTEXITCODE -eq 0)
         } catch {} finally { $ErrorActionPreference = $prevEapChk }
         # Self-heal web-server deps: a cut-short install.sh "studio deps" step leaves torch + unsloth
-        # but no fastapi/uvicorn/structlog/starlette (`unsloth studio` dies). Reinstall them without
-        # pinning huggingface-hub/transformers/datasets so the verified GPU torch stack stays intact.
+        # but no fastapi/uvicorn/structlog/starlette (`unsloth studio` dies). Reinstall them unpinned
+        # (no huggingface-hub/transformers/datasets) so the verified GPU torch stack stays intact.
         if ($torchOk) {
             $_studioPy = "/root/.unsloth/studio/unsloth_studio/bin/python"
             $_serverOk = $false
@@ -1928,8 +1924,8 @@ shell.Run cmd, 0, False
             } catch {} finally { $ErrorActionPreference = $prevEapS }
             if (-not $_serverOk) {
                 substep "Studio web-server deps incomplete (install.sh step cut short) -- installing them now..." "Cyan"
-                # studio.txt minus the huggingface-hub pin; uv preferred, pip fallback. Bare names
-                # only: `>=` becomes a redirection through PowerShell -> wsl.exe -> bash -lc, and
+                # studio.txt minus the huggingface-hub pin; uv preferred, pip fallback. Bare names only:
+                # `>=` would become a redirection through PowerShell -> wsl.exe -> bash -lc, and
                 # latest-of-each satisfies the studio.txt minimums anyway.
                 $_deps = 'typer fastapi uvicorn matplotlib pandas nest_asyncio pyjwt easydict addict structlog diceware ddgs cryptography httpx fastmcp'
                 $_repair = 'PY=/root/.unsloth/studio/unsloth_studio/bin/python; UV="$(command -v uv 2>/dev/null || echo /root/.local/bin/uv)"; if [ -x "$UV" ] || command -v uv >/dev/null 2>&1; then "$UV" pip install --python "$PY" ' + $_deps + '; else "$PY" -m pip install ' + $_deps + '; fi'
@@ -1943,8 +1939,8 @@ shell.Run cmd, 0, False
                 if ($_serverOk) { substep "Studio web-server deps installed." "Green" }
                 else { substep "(could not auto-install Studio server deps; 'unsloth studio' may fail to start)" "Yellow" }
             }
-            # The uv-managed venv ships no `pip`, but unsloth-zoo's exporter calls check_pip() and only
-            # finds `uv pip` when uv is on PATH. Seed pip so `save_pretrained_gguf` works regardless.
+            # The uv-managed venv ships no `pip`, but unsloth-zoo's exporter's check_pip() finds `uv pip`
+            # only when uv is on PATH. Seed pip so `save_pretrained_gguf` works regardless.
             $prevEapP = $ErrorActionPreference; $ErrorActionPreference = "Continue"
             try {
                 & wsl.exe -d $distro --cd /root -u root -- $_studioPy -m pip --version *> $null
@@ -1956,7 +1952,7 @@ shell.Run cmd, 0, False
         if ($torchOk) {
             step "done" "Unsloth Studio installed in WSL '$distro' -- GPU ready (torch.cuda available)." "Green"
             # Native Windows `unsloth` shim forwards every `unsloth ...` into the WSL GPU env so the user
-            # never touches WSL. WSL2 forwards 127.0.0.1, so http://localhost:8888 opens in the Windows browser.
+            # never touches WSL. WSL2 forwards 127.0.0.1, so http://localhost:8888 opens in Windows.
             try {
                 $shimDir = Join-Path $env:LOCALAPPDATA "Unsloth\bin"
                 New-Item -ItemType Directory -Force -Path $shimDir *> $null
@@ -1966,10 +1962,10 @@ shell.Run cmd, 0, False
                     "wsl.exe -d $_distroArg -u root -- /root/.unsloth/studio/unsloth_studio/bin/unsloth %*"
                 )
                 Set-Content -LiteralPath (Join-Path $shimDir "unsloth.cmd") -Value $shimLines -Encoding ASCII
-                # Record the distro so the uninstaller can clean a custom UNSLOTH_WSL_DISTRO
-                # install without the env var being set again.
+                # Record the distro so the uninstaller can clean a custom UNSLOTH_WSL_DISTRO install
+                # without the env var set again.
                 try { Set-Content -LiteralPath (Join-Path (Split-Path $shimDir -Parent) "wsl-distro.txt") -Value $distro -Encoding ASCII } catch {}
-                # A fresh profile may have no HKCU 'Path' at all; null would make TrimEnd() throw.
+                # A fresh profile may have no HKCU 'Path'; null would make TrimEnd() throw.
                 $userPath = [Environment]::GetEnvironmentVariable("Path", "User")
                 if (-not $userPath) { $userPath = "" }
                 if (($userPath -split ';') -notcontains $shimDir) {
@@ -1997,14 +1993,14 @@ shell.Run cmd, 0, False
                     'wsl.exe -d $distro --cd /root -u root -- bash -lic "unsloth studio -p 8888"'
                 )
                 Set-Content -LiteralPath $launcher -Value $L -Encoding UTF8
-                # Icon must live OUTSIDE %LOCALAPPDATA%: on WoA the shell's sandboxed icon broker
-                # can't read a .ico under AppData\Local, so the shortcut renders BLANK -- the same
-                # file under the user profile renders fine (verified on N1X). Only the icon moves.
+                # Icon must live OUTSIDE %LOCALAPPDATA%: on WoA the sandboxed icon broker can't read a
+                # .ico under AppData\Local, so the shortcut renders BLANK; under the user profile it
+                # renders fine (verified on N1X). Only the icon moves.
                 $iconDir = Join-Path $env:USERPROFILE ".unsloth"
                 New-Item -ItemType Directory -Force -Path $iconDir *> $null
                 $icon = Join-Path $iconDir "unsloth.ico"
-                # Prefer the bundled icon; fall back to a GitHub download. Validate the ICO header
-                # (00 00 01 00) before attaching, so a partial/HTML-404 download never makes a blank icon.
+                # Prefer the bundled icon, else download from GitHub. Validate the ICO header (00 00 01 00)
+                # before attaching, so a partial/HTML-404 download never makes a blank icon.
                 $bundledIcon = $null
                 if ($PSScriptRoot -and $PSScriptRoot.Trim()) { $bundledIcon = Join-Path $PSScriptRoot "studio\frontend\public\unsloth.ico" }
                 if ($bundledIcon -and (Test-Path -LiteralPath $bundledIcon)) {
@@ -2050,21 +2046,21 @@ shell.Run cmd, 0, False
             } catch {
                 substep "(could not create shortcuts: $($_.Exception.Message))" "Yellow"
             }
-            # GGUF *inference* needs a CUDA llama-server (no aarch64+CUDA prebuilt exists), so build one
-            # into ~/.unsloth/llama.cpp in the BACKGROUND. Best-effort; opt out: UNSLOTH_NO_LLAMA_CUDA=1.
+            # GGUF *inference* needs a CUDA llama-server (no aarch64+CUDA prebuilt), so build one into
+            # ~/.unsloth/llama.cpp in the BACKGROUND. Best-effort; opt out: UNSLOTH_NO_LLAMA_CUDA=1.
             if ($env:UNSLOTH_NO_LLAMA_CUDA -ne '1') {
                 $prevEapL = $ErrorActionPreference; $ErrorActionPreference = "Continue"
                 try {
                     $_llamaUrl = "https://raw.githubusercontent.com/unslothai/unsloth/$(Get-UnslothInstallRef)/studio/scripts/provision_llama_cuda.sh"
                     # Step 1: fetch the provision script + write a runner (base64 to dodge quoting layers).
-                    # The runner restores PATH (non-login shells miss /usr/lib/wsl/lib nvidia-smi ->
+                    # The runner restores PATH (non-login shells miss /usr/lib/wsl/lib nvidia-smi, so
                     # provision early-exits) and exports the env knobs below (Windows env vars don't cross
                     # into WSL). A runner FILE lets the detached launcher pass only space-free args,
                     # avoiding Start-Process mis-splitting `bash -lc <str>`.
                     $_pathLine = 'export PATH="/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/sbin:/bin:/usr/lib/wsl/lib:$PATH"' + "`n"
                     $_jobsLine = if ($env:UNSLOTH_LLAMA_BUILD_JOBS) { "export UNSLOTH_LLAMA_BUILD_JOBS=$($env:UNSLOTH_LLAMA_BUILD_JOBS)`n" } else { "" }
-                    # Bridge UNSLOTH_LLAMA_TAG / UNSLOTH_LLAMA_PR pins into WSL too, else the deferred
-                    # build silently ignores them. sh-single-quoted (tags/PRs are simple tokens).
+                    # Bridge UNSLOTH_LLAMA_TAG / UNSLOTH_LLAMA_PR pins into WSL, else the deferred build
+                    # ignores them. sh-single-quoted (tags/PRs are simple tokens).
                     $_tagLine = if ($env:UNSLOTH_LLAMA_TAG) { "export UNSLOTH_LLAMA_TAG='$($env:UNSLOTH_LLAMA_TAG)'`n" } else { "" }
                     $_prLine = if ($env:UNSLOTH_LLAMA_PR) { "export UNSLOTH_LLAMA_PR='$($env:UNSLOTH_LLAMA_PR)'`n" } else { "" }
                     $_runner = "#!/usr/bin/env bash`n" + $_pathLine + $_jobsLine + $_tagLine + $_prLine + "exec bash /root/.unsloth/provision_llama_cuda.sh > /root/.unsloth/llama_cuda_build.log 2>&1`n"
@@ -2073,9 +2069,9 @@ shell.Run cmd, 0, False
                     $_fetchOut = & wsl.exe -d $distro --cd /root -u root -- bash -lc $_fetchCmd 2>$null
                     if ("$_fetchOut" -match 'PROV_FETCHED') {
                         # Step 2: a detached Windows-side wsl.exe keeps the WSL VM up for the whole build
-                        # (a WSL-side `nohup &` dies: WSL stops the VM when the launching session exits).
-                        # PS 5.1 Start-Process joins -ArgumentList WITHOUT quoting, so pass $_distroArg
-                        # (pre-quoted only when spaced); all other tokens are space-free.
+                        # (a WSL-side `nohup &` dies when the launching session exits). PS 5.1 Start-Process
+                        # joins -ArgumentList WITHOUT quoting, so pass $_distroArg (pre-quoted only when
+                        # spaced); all other tokens are space-free.
                         Start-Process -WindowStyle Hidden -FilePath 'wsl.exe' -ArgumentList @('-d', $_distroArg, '--cd', '/root', '-u', 'root', '--', 'bash', '/root/.unsloth/run_llama_build.sh') | Out-Null
                         step "llama.cpp" "building CUDA llama.cpp for GGUF inference in the background (a few min); log: ~/.unsloth/llama_cuda_build.log" "Green"
                     } else {
@@ -2088,15 +2084,15 @@ shell.Run cmd, 0, False
             substep "retry, or launch manually:  wsl -d $_distroArg -u root -- bash -lic 'unsloth studio -p 8888'" "Cyan"
         }
         if ($torchOk) {
-            # Success: the Windows venv is vestigial here (everything runs in WSL), so drop the
+            # Success: the Windows venv is vestigial (everything runs in WSL), so drop the
             # rolled-aside previous-venv backup instead of orphaning it.
             Complete-StudioVenvRollback
             substep "GPU training + GGUF export run inside WSL. (GGUF *inference* additionally needs a CUDA llama.cpp build.)" "Yellow"
             $global:LASTEXITCODE = 0
             return
         }
-        # Failed (torch.cuda unavailable): restore any rolled-aside previous venv and report
-        # non-zero (plain return exits 0 for -File; under iex `exit` would kill the caller's shell).
+        # Failed (torch.cuda unavailable): restore any rolled-aside previous venv and report non-zero
+        # (plain return exits 0 for -File; under iex `exit` would kill the caller's shell).
         Restore-StudioVenvRollback
         $global:LASTEXITCODE = 1
         if ($PSCommandPath) { exit 1 }
diff --git a/install.sh b/install.sh
index 4f062b9f44..13d93531e7 100755
--- a/install.sh
+++ b/install.sh
@@ -2649,9 +2649,9 @@ elif [ -n "$TORCH_INDEX_URL" ]; then
             "unsloth-zoo @ git+https://github.com/unslothai/unsloth-zoo"
     elif [ -n "${UNSLOTH_INSTALL_REF:-}" ] && [ "${UNSLOTH_INSTALL_REF}" != "main" ] && [ "$PACKAGE_NAME" = "unsloth" ]; then
         # Pre-merge testing: install unsloth from a git ref (install.ps1 sets
-        # UNSLOTH_INSTALL_REF) so the branch's setup.sh + patches run. unsloth-zoo
-        # is not a base dep and SKIP_STUDIO_BASE skips studio base.txt, so name it
-        # explicitly or it never gets installed.
+        # UNSLOTH_INSTALL_REF) so the branch's setup.sh + patches run. Name
+        # unsloth-zoo explicitly: it's not a base dep and SKIP_STUDIO_BASE skips
+        # base.txt, so otherwise it never installs.
         substep "installing unsloth from git ref '$UNSLOTH_INSTALL_REF'..."
         run_install_cmd "install unsloth (@$UNSLOTH_INSTALL_REF)" uv pip install --python "$_VENV_PY" \
             --upgrade-package unsloth --upgrade-package unsloth-zoo \
@@ -2660,11 +2660,11 @@ elif [ -n "$TORCH_INDEX_URL" ]; then
         run_install_cmd "install unsloth" uv pip install --python "$_VENV_PY" \
             --upgrade-package unsloth -- "$PACKAGE_NAME"
     fi
-    # aarch64 + NVIDIA (DGX Spark / GB10 / N1X): base unsloth lacks bitsandbytes
-    # (the cuXXX extras are x86_64-oriented), so 4-bit QLoRA fails out of the box.
-    # aarch64 manylinux wheels work (verified on sm_121 via PTX JIT); best-effort,
-    # no wheel just keeps 16-bit LoRA / full finetuning. SKIP_TORCH gate: a
-    # --no-torch (GGUF-only) install must not let bitsandbytes drag torch back in.
+    # aarch64 + NVIDIA (DGX Spark / GB10 / N1X): unsloth's cuXXX extras are
+    # x86_64-oriented, so 4-bit QLoRA fails out of the box. aarch64 manylinux
+    # wheels work (verified on sm_121 via PTX JIT); best-effort, no wheel just
+    # keeps 16-bit LoRA / full finetuning. SKIP_TORCH gate: a --no-torch
+    # (GGUF-only) install must not let bitsandbytes drag torch back in.
     if [ "$SKIP_TORCH" = false ] \
             && { [ "$(uname -m)" = "aarch64" ] || [ "$(uname -m)" = "arm64" ]; } \
             && command -v nvidia-smi >/dev/null 2>&1 \
diff --git a/scripts/uninstall.ps1 b/scripts/uninstall.ps1
index e70b8d413f..287531fd71 100644
--- a/scripts/uninstall.ps1
+++ b/scripts/uninstall.ps1
@@ -418,12 +418,12 @@ function Uninstall-UnslothStudio {
     } catch { }
 
     # ── Windows-on-Arm WSL-fallback artifacts ──
-    # The ARM64+NVIDIA fallback puts Studio inside WSL plus a native shim + launcher under
-    # %LOCALAPPDATA%\Unsloth (not "Unsloth Studio") with a PATH entry -- all missed by the cleanup above.
+    # The ARM64+NVIDIA fallback puts Studio in WSL plus a native shim + launcher under
+    # %LOCALAPPDATA%\Unsloth (not "Unsloth Studio") with a PATH entry -- all missed above.
     _Step "Removing WSL-fallback artifacts (shim, launcher, PATH entry, WSL install)..."
     $unslothDir = if ($env:LOCALAPPDATA) { Join-Path $env:LOCALAPPDATA "Unsloth" } else { $null }
-    # wsl-distro.txt records a custom UNSLOTH_WSL_DISTRO install so it is cleanable
-    # without the env var set; read it BEFORE the directory is removed below.
+    # wsl-distro.txt records a custom UNSLOTH_WSL_DISTRO install so it's cleanable without the
+    # env var set; read it BEFORE the directory is removed below.
     $_recordedDistro = $null
     if ($unslothDir) {
         try {
@@ -460,19 +460,19 @@ function Uninstall-UnslothStudio {
     # Remove the Studio install inside each WSL distro (the real GPU install + any CUDA llama.cpp build).
     if (Get-Command wsl.exe -ErrorAction SilentlyContinue) {
         try {
-            # `wsl --list` emits UTF-16 PS mis-parses, so probe candidates by exit code instead
-            # ('' = default distro). rm runs FIRST (the kills could SIGKILL this shell) and also
-            # drops the dangling /root/.local/bin/unsloth symlink. Scope STRICTLY to /root (where
-            # the fallback installs): /home/*/.unsloth may be an unrelated user's. The port-8888
-            # kill is gated on an Unsloth install existing (checked BEFORE rm deletes the marker)
-            # so an unrelated 8888 listener survives; pkill matches argv containing /root/.unsloth/
-            # rather than bare names that would kill a user's own llama-server, and the backslash
-            # + [h]-bracket in '/root/\.unslot[h]/' keep it from matching this command's own argv.
+            # Probe candidates by exit code ('' = default distro) since `wsl --list` emits UTF-16 PS
+            # mis-parses. rm runs FIRST (the kills could SIGKILL this shell) and drops the dangling
+            # /root/.local/bin/unsloth symlink. Scope STRICTLY to /root (where the fallback installs);
+            # /home/*/.unsloth may be another user's. The 8888 kill is gated on an Unsloth install
+            # existing (checked BEFORE rm deletes the marker) so an unrelated listener survives; pkill
+            # matches argv containing /root/.unsloth/ (not bare names that would hit a user's own
+            # llama-server), and the backslash + [h]-bracket in '/root/\.unslot[h]/' keep it from
+            # matching this command's own argv.
             $_clean = '_had=0; if [ -d /root/.unsloth ] || [ -L /root/.local/bin/unsloth ]; then _had=1; fi; rm -rf /root/.unsloth /root/llama-cuda /root/provision_llama_cuda.sh /root/llama_cuda_build.log 2>/dev/null; rm -f /root/.local/bin/unsloth 2>/dev/null; if [ $_had -eq 1 ]; then fuser -k 8888/tcp 2>/dev/null; fi; pkill -9 -f ''/root/\.unslot[h]/'' 2>/dev/null; true'
-            # Clean only distros with evidence of a fallback install: the wsl-distro.txt marker
-            # or an explicit UNSLOTH_WSL_DISTRO. The broad candidate probe is only for legacy
-            # marker-less installs, which exist only on ARM64 hosts -- on x86 it would delete
-            # distros this installer never touched (e.g. a ROCm-on-WSL Studio under /root).
+            # Clean only distros with evidence of a fallback install: the wsl-distro.txt marker or an
+            # explicit UNSLOTH_WSL_DISTRO. The broad candidate probe is only for legacy marker-less
+            # installs, which exist only on ARM64 -- on x86 it would delete distros this installer
+            # never touched (e.g. a ROCm-on-WSL Studio under /root).
             $_cands = @()
             if ($env:UNSLOTH_WSL_DISTRO) { $_cands += $env:UNSLOTH_WSL_DISTRO }
             if ($_recordedDistro) { $_cands += $_recordedDistro }
@@ -505,8 +505,8 @@ function Uninstall-UnslothStudio {
         Write-Host "  `$env:UNSLOTH_STUDIO_HOME = 'C:\your\path'; irm https://raw.githubusercontent.com/unslothai/unsloth/main/scripts/uninstall.ps1 | iex"
     }
 
-    # The distro probes leave a failing $LASTEXITCODE; reset it so success exits 0. Set the
-    # var rather than `exit 0` so `irm ... | iex` doesn't terminate the caller's shell.
+    # The distro probes leave a failing $LASTEXITCODE; reset so success exits 0. Set the var
+    # rather than `exit 0` so `irm ... | iex` doesn't terminate the caller's shell.
     $global:LASTEXITCODE = 0
 }
 
diff --git a/scripts/uninstall.sh b/scripts/uninstall.sh
index 8abc8bf3b9..20285da856 100755
--- a/scripts/uninstall.sh
+++ b/scripts/uninstall.sh
@@ -303,9 +303,9 @@ case "$_os" in
                             } catch { }
                         }
                     }
-                    # WoA WSL-fallback (install.ps1) native shim/launcher dir
-                    # (%LOCALAPPDATA%\Unsloth) + its PATH entry. install.ps1 created the
-                    # shim; clean it here too so a WSL-side bash uninstall is complete.
+                    # Remove the WoA WSL-fallback native shim/launcher dir
+                    # (%LOCALAPPDATA%\Unsloth) + its PATH entry that install.ps1
+                    # created, so a WSL-side bash uninstall is complete.
                     $ud = if ($env:LOCALAPPDATA) { Join-Path $env:LOCALAPPDATA "Unsloth" } else { $null };
                     if ($ud) {
                         $shim = (Join-Path $ud "bin").TrimEnd("\","/");
diff --git a/studio/backend/core/training/worker.py b/studio/backend/core/training/worker.py
index 3ed83024c8..0d1e51343d 100644
--- a/studio/backend/core/training/worker.py
+++ b/studio/backend/core/training/worker.py
@@ -738,11 +738,10 @@ def _nvidia_classify_spark_unified_memory(props: Any) -> tuple[str, bool]:
     Returns ``(marker, is_unified)``; marker is ``"is_integrated"`` or the matched
     device-name token, else ``""``. Spark-class parts (DGX Spark / GB10, N1X "RTX
     Spark") share one memory pool with the OS, so like the ROCm APUs they need a
-    ``set_per_process_memory_fraction`` cap -- exhausting the pool can stall the
-    box instead of raising a catchable OutOfMemoryError.
+    ``set_per_process_memory_fraction`` cap -- exhausting the pool can stall the box.
 
     ``is_integrated`` is authoritative on native Linux, but WSL2 paravirtualization
-    masks it to 0 and renames the device (the N1X reports ``JMJWOA-Generic-GPU``,
+    masks it to 0 and renames the device (N1X reports ``JMJWOA-Generic-GPU``,
     verified on hardware) -- hence the name-token fallback. Tokens mirror
     ``_DGX_SPARK_DEVICE_TOKENS`` in ``unsloth/models/_utils.py`` (duplicated
     because this guard runs before any ML import).
@@ -2432,16 +2431,16 @@ def run_training_process(*, event_queue: Any, stop_queue: Any, config: dict) ->
             logger.debug("Could not set GPU memory fraction: %s", _oom_guard_err)
 
     # ── 1h. NVIDIA Spark-class unified-memory OOM guard ──
-    # NVIDIA flavor of the ROCm APU guard above: Spark-class parts share one
-    # memory pool with the OS, so over-allocation can stall the box instead of
-    # raising a catchable OutOfMemoryError. Cap at 0.80 like Strix Halo (20%
-    # headroom stays with the OS/page cache). UNSLOTH_SPARK_MEM_FRACTION
-    # overrides; outside (0, 1] disables. Discrete NVIDIA GPUs untouched.
+    # NVIDIA flavor of the ROCm APU guard above: Spark-class parts share one pool
+    # with the OS, so over-allocation can stall the box instead of raising a
+    # catchable OutOfMemoryError. Cap at 0.80 like Strix Halo (20% headroom for
+    # OS/page cache). UNSLOTH_SPARK_MEM_FRACTION overrides; outside (0, 1] disables.
+    # Discrete NVIDIA GPUs untouched.
     else:
         try:
-            # Set PYTORCH_CUDA_ALLOC_CONF before get_device_properties below inits
-            # the CUDA allocator -- the later `import unsloth` patch is too late for
-            # THIS worker process. CUDA-free nvidia-smi sniff (mirrors
+            # Set PYTORCH_CUDA_ALLOC_CONF before get_device_properties below inits the
+            # CUDA allocator -- the later `import unsloth` patch is too late for THIS
+            # worker process. CUDA-free nvidia-smi sniff (mirrors
             # _is_dgx_spark_no_cuda_init), same append-don't-override and
             # UNSLOTH_NO_EXPANDABLE_SEGMENTS opt-out as the library patch.
             try:
diff --git a/studio/backend/tests/test_spark_oom_guard.py b/studio/backend/tests/test_spark_oom_guard.py
index 74610d4369..4e61dfd3af 100644
--- a/studio/backend/tests/test_spark_oom_guard.py
+++ b/studio/backend/tests/test_spark_oom_guard.py
@@ -3,13 +3,11 @@
 
 """Unit tests for _nvidia_classify_spark_unified_memory (Spark OOM-guard classifier).
 
-Two paths: (1) the ``is_integrated`` device property (authoritative on native
-Linux), (2) device-name token match — needed because WSL2's GPU
-paravirtualization masks ``is_integrated`` to 0 and renames the device (the N1X
-reports ``JMJWOA-Generic-GPU``; verified on hardware).
-
-Mirrors test_rocm_oom_guard.py for the ROCm/Strix-Halo classifier the NVIDIA
-guard was modeled on.
+Two paths: (1) ``is_integrated`` device property (authoritative on native Linux),
+(2) device-name token match -- needed because WSL2's GPU paravirtualization masks
+``is_integrated`` to 0 and renames the device (N1X reports ``JMJWOA-Generic-GPU``;
+verified on hardware). Mirrors test_rocm_oom_guard.py for the ROCm/Strix-Halo
+classifier the NVIDIA guard was modeled on.
 """
 
 from __future__ import annotations
diff --git a/studio/scripts/provision_llama_cuda.sh b/studio/scripts/provision_llama_cuda.sh
index 8c593e7950..ef4c63bb50 100644
--- a/studio/scripts/provision_llama_cuda.sh
+++ b/studio/scripts/provision_llama_cuda.sh
@@ -13,9 +13,9 @@ LLAMA_DIR="${UNSLOTH_LLAMA_CPP_PATH:-$HOME/.unsloth/llama.cpp}"
 SERVER="$LLAMA_DIR/build/bin/llama-server"
 log() { printf '  - %s\n' "$*"; }
 
-# CUDA shows up two ways: old monolithic (libggml-cuda in ldd) or current split
-# build (dlopen-ed libggml-cuda.so* beside the binary, missed by ldd). CPU-only
-# builds ship no libggml-cuda.so, so its presence is the reliable signal.
+# CUDA shows up two ways: monolithic (libggml-cuda in ldd) or split (dlopen-ed
+# libggml-cuda.so* beside the binary, missed by ldd). CPU-only builds ship no
+# libggml-cuda.so, so its presence is the reliable signal.
 is_cuda_server() {
     [ -x "$1" ] || return 1
     ldd "$1" 2>/dev/null | grep -qi 'libggml-cuda' && return 0
@@ -165,8 +165,8 @@ if ! _cmake_configure; then
     _cmake_configure || { log "cmake configure failed"; cd /; _restore_prev; exit 0; }
 fi
 # Also builds the targets unsloth-zoo's GGUF exporter needs (llama-mtmd-cli,
-# llama-gguf-split). Jobs default to ~half the cores -- full -j(nproc) CUDA builds
-# trip thermal shutdowns on NVIDIA-ARM laptops (N1X "RTX Spark") -- and are
+# llama-gguf-split). Jobs default to ~half the cores (full -j(nproc) CUDA builds
+# trip thermal shutdowns on NVIDIA-ARM laptops like the N1X "RTX Spark") and are
 # RAM-capped (~1.5 GB/nvcc job). Tune: UNSLOTH_LLAMA_BUILD_JOBS=N; re-runs resume.
 _ncpu="$(nproc 2>/dev/null || echo 4)"
 # Honor a valid positive-int override; ignore junk/0 (cmake reads -j0 as "all cores").
@@ -198,7 +198,7 @@ _cmake_build_extras() {
     done
 }
 if ! _cmake_build; then
-    # An interrupted build (thermal/power shutdown -- this machine class is prone)
+    # An interrupted build (thermal/power shutdown, common on this machine class)
     # can leave a half-linked libggml-cuda.so that breaks the resume link
     # (undefined ggml_cuda_op_* refs); wipe and rebuild clean once.
     log "build failed (likely interrupted/partial); wiping build dir and rebuilding clean"
diff --git a/studio/setup.sh b/studio/setup.sh
index 3a55c1607a..c8d9a630d8 100755
--- a/studio/setup.sh
+++ b/studio/setup.sh
@@ -974,9 +974,9 @@ LLAMA_CPP_DIR="$UNSLOTH_HOME/llama.cpp"
 LLAMA_SERVER_BIN="$LLAMA_CPP_DIR/build/bin/llama-server"
 _NEED_LLAMA_SOURCE_BUILD=false
 _LLAMA_CPP_DEGRADED=false
-# Deferred != degraded: on WSL2 aarch64+NVIDIA install.ps1 builds the real CUDA
-# server in the background, so a temporarily-absent server is success and must not
-# trip the arm64 CPU-prebuilt last-resort or the exit 1.
+# Deferred != degraded: on WSL2 aarch64+NVIDIA install.ps1 builds the CUDA server
+# in the background, so an absent server is success and must not trip the arm64
+# CPU-prebuilt last-resort or exit 1.
 _LLAMA_CPP_DEFERRED=false
 _LLAMA_FORCE_COMPILE="${UNSLOTH_LLAMA_FORCE_COMPILE:-0}"
 _REQUESTED_LLAMA_TAG="${UNSLOTH_LLAMA_TAG:-${_DEFAULT_LLAMA_TAG}}"
@@ -1161,10 +1161,10 @@ if [ "$_NEED_LLAMA_SOURCE_BUILD" = true ] && \
 fi
 
 # ── WSL2 aarch64 + NVIDIA, no nvcc yet: defer to the background CUDA build ──
-# install.ps1 builds the real CUDA llama-server in the background after install;
-# without nvcc, section 9 could only make a slow CPU server that build discards.
-# With nvcc we fall through to section 9; opted out (UNSLOTH_NO_LLAMA_CUDA=1) the
-# CPU build is kept as the only server.
+# install.ps1 builds the CUDA llama-server in the background; without nvcc,
+# section 9 could only make a slow CPU server that build discards. With nvcc we
+# fall through to section 9; opted out (UNSLOTH_NO_LLAMA_CUDA=1) the CPU build is
+# kept as the only server.
 if [ "$_NEED_LLAMA_SOURCE_BUILD" = true ] \
         && [ "$_LLAMA_FORCE_COMPILE" != "1" ] \
         && [ -z "$_LLAMA_PR" ] \
@@ -1454,10 +1454,10 @@ else
                     if [ "$_CUDA_TOOLKIT_ALLOWED" = true ]; then
                         CMAKE_ARGS="$CMAKE_ARGS -DGGML_CUDA=ON"
 
-                        # glibc >= 2.41 + CUDA < 13.3: rsqrt/rsqrtf header clash fails every .cu
-                        # ("exception specification is incompatible") -> CPU fallback; only fix is
-                        # CUDA >= 13.3. Diagnostic only -- never changes flags or aborts. Checked
-                        # against the final _NVCC_VER (after the driver-compat swap above).
+                        # glibc >= 2.41 + CUDA < 13.3: rsqrt/rsqrtf header clash fails
+                        # every .cu -> CPU fallback; only fix is CUDA >= 13.3. Diagnostic
+                        # only (never changes flags or aborts). Checks the final _NVCC_VER
+                        # (after the driver-compat swap above).
                         _GLIBC_VER="$(getconf GNU_LIBC_VERSION 2>/dev/null | awk '{print $2}')" || _GLIBC_VER=""
                         if [ -n "$_GLIBC_VER" ]; then
                             _GLIBC_MAJ="${_GLIBC_VER%%.*}"; _GLIBC_MIN="${_GLIBC_VER#*.}"; _GLIBC_MIN="${_GLIBC_MIN%%.*}"
@@ -1683,12 +1683,11 @@ fi  # end _SKIP_GGUF_BUILD check
 #    llama.cpp when the source build above could not (no CUDA toolkit found) ──
 # No aarch64+CUDA prebuilt exists and a fresh Spark ships only driver + nvidia-smi,
 # so the build above fell back to CPU; mirror the Windows/WSL fix
-# (provision_llama_cuda.sh) for native Linux. Gated to Linux aarch64 + NVIDIA with
-# no CUDA server yet (opt out: UNSLOTH_NO_LLAMA_CUDA=1); best-effort -- provision
-# always exits 0 and on failure the prior CPU/degraded state stands.
-# CUDA detection covers both layouts: old monolithic (libggml-cuda in ldd) and
-# split build (dlopen-ed libggml-cuda.so* beside the binary, missed by ldd --
-# CPU-only builds ship no libggml-cuda.so, so its presence is the signal).
+# (provision_llama_cuda.sh) for native Linux. Best-effort: provision always exits
+# 0, and on failure the prior CPU/degraded state stands.
+# CUDA detection covers both layouts: monolithic (libggml-cuda in ldd) and split
+# (dlopen-ed libggml-cuda.so* beside the binary, missed by ldd). CPU-only builds
+# ship no libggml-cuda.so, so its presence is the signal.
 _have_cuda_llama_server() {
     [ -x "$LLAMA_SERVER_BIN" ] || return 1
     ldd "$LLAMA_SERVER_BIN" 2>/dev/null | grep -qi 'libggml-cuda' && return 0
@@ -1702,11 +1701,11 @@ if [ "$_HOST_SYSTEM" = "Linux" ] \
         && command -v nvidia-smi >/dev/null 2>&1 \
         && nvidia-smi -L 2>/dev/null | awk '/^GPU[[:space:]]+[0-9]+:/{found=1} END{exit !found}' \
         && ! _have_cuda_llama_server; then
-    # Under WSL this runs ONLY for a DIRECT `install.sh` run: install.ps1 exports
-    # UNSLOTH_WSL_LLAMA_DEFERRED=1 and builds CUDA llama.cpp in the background, but
-    # a direct run has no background builder, so provision here.
-    # Resolve provision_llama_cuda.sh: copy beside setup.sh, then local-dev repo,
-    # else fetch from GitHub (so `curl | sh` works on an older wheel without it).
+    # Under WSL this runs ONLY for a DIRECT `install.sh` run: install.ps1 sets
+    # UNSLOTH_WSL_LLAMA_DEFERRED=1 and builds in the background, but a direct run
+    # has no background builder, so provision here.
+    # Resolve provision_llama_cuda.sh: beside setup.sh, then local-dev repo, else
+    # fetch from GitHub (so `curl | sh` works on an older wheel without it).
     _PROV_SH=""
     if [ -f "$SCRIPT_DIR/scripts/provision_llama_cuda.sh" ]; then
         _PROV_SH="$SCRIPT_DIR/scripts/provision_llama_cuda.sh"
diff --git a/unsloth/kernels/flex_attention.py b/unsloth/kernels/flex_attention.py
index 044e8911f2..8cd1e82141 100644
--- a/unsloth/kernels/flex_attention.py
+++ b/unsloth/kernels/flex_attention.py
@@ -27,10 +27,9 @@ torch_compile_options = {
 
 
 def _flex_is_dgx_spark():
-    # Local CUDA-free copy of _utils._is_dgx_spark_no_cuda_init() (avoids a circular
-    # import). Runs at module import, before ._utils -- touching torch.cuda here would
-    # init the allocator before patch_dgx_spark_memory_config() can set
-    # PYTORCH_CUDA_ALLOC_CONF on the very Spark hosts this targets.
+    # CUDA-free copy of _utils._is_dgx_spark_no_cuda_init() (avoids a circular import).
+    # Runs at module import, before ._utils -- touching torch.cuda here would init the
+    # allocator before patch_dgx_spark_memory_config() can set PYTORCH_CUDA_ALLOC_CONF.
     _force = os.environ.get("UNSLOTH_FORCE_DGX_SPARK")
     if _force == "1":
         return True
diff --git a/unsloth/models/_utils.py b/unsloth/models/_utils.py
index ab7f19133c..7cbe5e1aba 100644
--- a/unsloth/models/_utils.py
+++ b/unsloth/models/_utils.py
@@ -993,8 +993,8 @@ from transformers.modeling_utils import logger as transformers_logger
 
 
 # ---- NVIDIA DGX Spark (GB10) / N1X "RTX Spark" unified-memory support ----
-# Device names vary ("NVIDIA GB10" on DGX Spark, "JMJWOA-Generic-GPU" on N1X);
-# the aarch64 + CUDA gate keeps every Spark workaround a no-op elsewhere.
+# Device names vary ("NVIDIA GB10", "JMJWOA-Generic-GPU" on N1X); the
+# aarch64 + CUDA gate keeps every Spark workaround a no-op elsewhere.
 _DGX_SPARK_DEVICE_TOKENS = ("GB10", "JMJWOA", "N1X", "DGX SPARK", "GB110")
 
 
@@ -1024,9 +1024,8 @@ def is_dgx_spark():
 
 @functools.lru_cache(maxsize = None)
 def _is_dgx_spark_no_cuda_init():
-    """Spark detection that never initializes CUDA: reads device names via
-    `nvidia-smi` instead of torch, so allocator-init-time settings
-    (PYTORCH_CUDA_ALLOC_CONF) can still be set after calling it. Same
+    """Spark detection that never inits CUDA: reads device names via `nvidia-smi`,
+    not torch, so PYTORCH_CUDA_ALLOC_CONF can still be set afterwards. Same
     UNSLOTH_FORCE_DGX_SPARK override; False on any error."""
     _force = os.environ.get("UNSLOTH_FORCE_DGX_SPARK")
     if _force == "1":
@@ -1055,10 +1054,10 @@ def _is_dgx_spark_no_cuda_init():
 def patch_dgx_spark_caching_allocator_warmup():
     """No-op `transformers.modeling_utils.caching_allocator_warmup` on Spark UMA.
 
-    `cudaMemGetInfo()` undercounts free memory on Spark unified memory, so HF's
-    warmup `torch.empty(...)` raises `AcceleratorError: invalid argument` and
-    aborts bitsandbytes 4/8-bit loads. The warmup is only a speed hint, so skip
-    it. Gated by `is_dgx_spark()`; idempotent (`_unsloth_spark_noop` marker).
+    `cudaMemGetInfo()` undercounts free UMA memory, so HF's warmup
+    `torch.empty(...)` raises `AcceleratorError: invalid argument` and aborts
+    bitsandbytes 4/8-bit loads. The warmup is only a speed hint, so skip it.
+    Gated by `is_dgx_spark()`; idempotent (`_unsloth_spark_noop` marker).
     """
     if not is_dgx_spark():
         return
@@ -1085,7 +1084,7 @@ def patch_dgx_spark_memory_config():
     Appends to PYTORCH_CUDA_ALLOC_CONF only when absent; opt out with
     UNSLOTH_NO_EXPANDABLE_SEGMENTS=1. Must run before the first CUDA allocation,
     hence the CUDA-free `_is_dgx_spark_no_cuda_init()` gate -- `is_dgx_spark()`
-    would initialize the allocator before the env var could take effect.
+    would init the allocator before the env var could take effect.
     """
     if not _is_dgx_spark_no_cuda_init():
         return
@@ -1104,7 +1103,7 @@ def patch_dgx_spark_runtime_defaults():
 
     - UNSLOTH_DISABLE_DOUBLE_BUFFER=1 (setdefault): zoo's grad-checkpointing
       double-buffer gates on a mem_get_info check that UNDERCOUNTS on UMA, and
-      its extra staging buffer is pure waste on a shared pool.
+      its staging buffer is pure waste on a shared pool.
     - UNSLOTH_SPARK_MEM_FRACTION=<0..1> (opt-in, default NO cap): caps the
       allocator so over-allocation raises OutOfMemoryError early instead of
       wedging the box (untracked UMA allocations may never trip a catchable OOM).
@@ -1126,9 +1125,9 @@ def patch_dgx_spark_runtime_defaults():
 def patch_dgx_spark_dataloader_defaults():
     """Default `dataloader_pin_memory` to False on Spark UMA (accuracy-neutral).
 
-    With one shared memory pool, pinning only reserves non-pageable RAM and adds
-    a staging copy (mirrors transformers' own use_cpu precedent). Wrapping the
-    base `TrainingArguments.__post_init__` covers SFT + every TRL trainer in one
+    On one shared pool, pinning only reserves non-pageable RAM and adds a staging
+    copy (mirrors transformers' own use_cpu precedent). Wrapping the base
+    `TrainingArguments.__post_init__` covers SFT + every TRL trainer in one
     idempotent patch. Opt out: UNSLOTH_SPARK_KEEP_PIN_MEMORY=1. No-op off-Spark.
     """
     if not is_dgx_spark():