Merge remote-tracking branch 'upstream/main' into fix/rocm-version-test-isolation

.github/workflows/consolidated-tests-ci.yml

@@ -7,7 +7,7 @@
 #
 # Why a separate workflow:
 #   - studio-backend-ci.yml's "Repo tests (CPU)" job already auto-discovers
-#     tests/ minus tests/qlora, tests/saving, tests/utils, tests/sh. The 16
+#     tests/ minus tests/qlora, tests/saving, tests/utils, tests/sh. The 17
 #     Bucket-A tests below live inside those --ignore dirs (CPU-runnable but
 #     historically excluded with their GPU siblings); pulling them out into
 #     a sibling job keeps the existing 760-passed baseline stable while we
@@ -274,6 +274,7 @@ jobs:
             tests/saving/test_export_dispatch.py \
             tests/saving/test_imatrix_export.py \
             tests/saving/test_gguf_single_pass_export.py \
+            tests/saving/test_offline_gguf_vlm_tokenizer_7481.py \
             tests/utils/test_attention_masks.py \
             tests/utils/test_trunc_normal_patch.py \
             tests/python/test_fast_language_model_text_only.py
@@ -365,6 +366,7 @@ jobs:
             tests/saving/test_export_dispatch.py \
             tests/saving/test_imatrix_export.py \
             tests/saving/test_gguf_single_pass_export.py \
+            tests/saving/test_offline_gguf_vlm_tokenizer_7481.py \
             tests/utils/test_attention_masks.py \
             tests/utils/test_trunc_normal_patch.py \
             tests/python/test_fast_language_model_text_only.py \
@@ -2129,7 +2131,7 @@ jobs:
           pip show unsloth_zoo
           echo "::endgroup::"
           echo "Consolidated job done. Coverage:"
-          echo "  - 16 unsloth Bucket-A tests under tests/saving/ + tests/utils/"
+          echo "  - 17 unsloth Bucket-A tests under tests/saving/ + tests/utils/"
           echo "  - unsloth_zoo @ ${UNSLOTH_ZOO_REF} pytest tests/ (5 GPU cases deselected)"
           echo "  - unsloth_zoo.compiler.test_apply_fused_lm_head"
 

.github/workflows/studio-backend-ci.yml

@@ -30,6 +30,13 @@ on:
       - 'unsloth/**'
       - 'unsloth_cli/**'
       - 'tests/**'
+      # The root installers: tests/sh/*.sh and tests/studio/install/* assert
+      # against these two files, so a change here must run the suite that
+      # covers it. Without them an install-only edit (the shape most AMD/ROCm
+      # routing fixes take) skipped Backend CI entirely.
+      - 'install.sh'
+      - 'install.ps1'
+      - 'scripts/**'
       - 'pyproject.toml'
       - '.github/workflows/studio-backend-ci.yml'
   push:
@@ -217,27 +224,32 @@ jobs:
             tests/studio/test_xpu_spoof_pipeline.py
 
       - name: Shell installer tests
-        # Subset that does not depend on a writable / pristine install.sh
+        # Auto-discovered rather than allowlisted. The old hardcoded list had
-        # tree; test_install_host_defaults.sh checks install.ps1 layout
+        # silently fallen seven files behind tests/run_all.sh, including
-        # which has drifted (separate followup).
+        # test_strixhalo_wsl_reroute.sh -- the only shell coverage of the ROCm
+        # WSL reroute -- so that suite never ran on a PR. Skips are explicit,
+        # each with a reason, and tests/studio/test_ci_shell_suite_coverage.py
+        # fails if this step stops discovering the directory or the skip list
+        # grows without one.
+        #
+        # Skipped:
+        #   test_install_host_defaults.sh: asserts an install.ps1 layout that
+        #     has drifted (separate followup).
+        #   test_install_rollback_lifecycle.sh: already runs on both platforms
+        #     in cross-platform-parity-ci.yml.
         run: |
           set -e
-          for s in \
+          skip="test_install_host_defaults.sh test_install_rollback_lifecycle.sh"
-              tests/sh/test_get_torch_index_url.sh \
+          found=0
-              tests/sh/test_mac_intel_compat.sh \
+          for s in tests/sh/test_*.sh; do
-              tests/sh/test_node_decision.sh \
+              case " $skip " in
-              tests/sh/test_studio_home_node_dir.sh \
+                  *" $(basename "$s") "*) echo "skipping $s (see workflow comment)"; continue ;;
-              tests/sh/test_system_node_readonly.sh \
+              esac
-              tests/sh/test_nvcc_meets_llama_minimum.sh \
+              found=$((found + 1))
-              tests/sh/test_resolve_cuda_archs.sh \
-              tests/sh/test_staged_validation_enabled.sh \
-              tests/sh/test_tauri_install_exit_order.sh \
-              tests/sh/test_torch_constraint.sh \
-              tests/sh/test_torch_flavor.sh \
-              tests/sh/test_with_llama_cpp_dir_flag.sh \
-              tests/sh/test_with_llama_cpp_dir_link_behavior.sh; do
               echo "::group::$s"
               bash "$s"
               echo "::endgroup::"
           done
+          [ "$found" -gt 0 ] || { echo "::error::no shell tests discovered under tests/sh"; exit 1; }
+          echo "ran $found shell installer test files"
 

.gitignore

@@ -238,4 +238,5 @@ package-lock.json
 !studio/package-lock.json
 llama.cpp/
 # Stray "~" dir some tools create from a literal ~ TMPDIR; never part of the repo.
-/~/
+~/
+/temp/

README.md

@@ -103,7 +103,7 @@ Unsloth Studio (Beta) works on **Windows, Linux, WSL** and **macOS**.
 * **NVIDIA:** Training works on RTX 30/40/50, Blackwell, DGX Spark, Station and more
 * **macOS:** Training, MLX and GGUF inference are ALL supported.
 * **AMD:** Training, RL, chat and deployment work on Windows, WSL and Linux. [Read the AMD guide](https://unsloth.ai/docs/basics/amd).
-* **Vulkan:** GGUF inference is supported on [compatible GPUs, including Intel GPUs](https://github.com/unslothai/unsloth/pull/5819).
+* **Vulkan:** GGUF inference is supported on [compatible GPUs, including Intel GPUs](https://github.com/unslothai/unsloth/pull/5819). Vulkan accelerates GGUF inference only; training still requires a supported PyTorch or MLX backend.
 * **Multi-GPU:** Available now, with a major upgrade on the way
 
 #### macOS, Linux, WSL:
@@ -112,12 +112,28 @@ curl -fsSL https://unsloth.ai/install.sh | sh
 ```
 Use the same command to update.
 
+To force the Vulkan llama.cpp backend, set `UNSLOTH_FORCE_VULKAN=1` **before installing or updating**. The setting selects the llama.cpp binary bundle, so setting it only when launching Studio cannot replace an existing CPU bundle:
+
+```bash
+export UNSLOTH_FORCE_VULKAN=1
+curl -fsSL https://unsloth.ai/install.sh | sh
+```
+
 #### Windows:
 ```powershell
 irm https://unsloth.ai/install.ps1 | iex
 ```
 Use the same command to update.
 
+To force the Vulkan llama.cpp backend, set the environment variable before running the installer or updater:
+
+```powershell
+$env:UNSLOTH_FORCE_VULKAN=1
+irm https://unsloth.ai/install.ps1 | iex
+```
+
+Re-running the current installer replaces a previously selected CPU bundle when the backend differs. A separate Vulkan SDK is not required; the GPU driver must provide a working Vulkan runtime.
+
 #### Launch
 ```bash
 unsloth studio -p 8888
@@ -263,6 +279,8 @@ unsloth studio -H 0.0.0.0 -p 8888
 ```
 The Cloudflare tunnel is **off by default**: `-H 0.0.0.0` exposes the raw port only, not a public internet URL. Pair the wildcard bind with `--cloudflare` (`unsloth studio -H 0.0.0.0 --cloudflare`) to also publish a public `https://*.trycloudflare.com` link, or prefer `--secure` (above), which keeps the raw port private. `--cloudflare` has no effect on a loopback bind.
 
+On a wildcard bind Unsloth works out the address to share by asking `ifconfig.me` for the public IP, then asks `check-host.net` whether that port is reachable so it can tell you if a firewall is in the way. Both contact a third party. Set `UNSLOTH_STUDIO_DISABLE_PUBLIC_CHECK=1` to skip them; the banner then shows the LAN address and no reachability line.
+
 The first time Unsloth is published on a public URL (`--secure` or `--cloudflare`) with the auto-generated admin password still in place, it asks for a new admin password in the terminal (masked input with confirmation) before the public link goes up. Without an attached terminal it warns instead and keeps the bootstrap deadline: Unsloth shuts down after `UNSLOTH_STUDIO_BOOTSTRAP_TIMEOUT` (default 1 hour) unless the password is changed in the web UI.
 
 For headless setups that cannot answer that prompt, set the initial admin password non-interactively with `--password` (only takes effect when no password is set yet; if one already exists it is a hard error, so rotate later with `unsloth studio reset-password`):

install.ps1

@@ -1917,12 +1917,14 @@ exit 0
             #    (gfx120X/110X/1151/1150/103X); unknown names fall back to CPU.
             elseif ($ROCmGpuLabel) {
                 $nameArchTable = @(
-                    @{ P = "9070 XT|9080";                                        A = "gfx1201" }  # RDNA 4 (RX 9070 XT / 9080)
+                    @{ P = "9070|9080";                                           A = "gfx1201" }  # RDNA 4 (Navi 48: RX 9070 XT / 9070 GRE / 9070 / 9080)
-                    @{ P = "9070|9060";                                           A = "gfx1200" }  # RDNA 4 (RX 9070 / 9060)
+                    @{ P = "9060";                                                A = "gfx1200" }  # RDNA 4 (Navi 44: RX 9060 XT / 9060)
                     @{ P = "8065S|8060S|8050S|8040S|Strix Halo|Ryzen AI Max|AI Max"; A = "gfx1151" }  # RDNA 3.5 (Strix Halo + Gorgon Halo: Radeon 8065S/8060S/8050S/8040S iGPU, Ryzen AI Max / Max+)
-                    @{ P = "890M|880M|860M|840M|Strix Point|Krackan|HX 37[05]|AI 9 HX|AI 9 36[05]|AI 7 35[05]|AI 5 34[05]|AI 7 PRO 35|AI 5 33"; A = "gfx1150" }  # RDNA 3.5 (Strix/Krackan Point: Radeon 890M/880M iGPU, Ryzen AI 9 HX 370/375)
+                    @{ P = "890M|880M|Strix Point|HX 37[05]|AI 9 HX|AI 9 36[05]"; A = "gfx1150" }  # RDNA 3.5 (Strix Point: Radeon 890M/880M, Ryzen AI 9 HX 370/375)
-                    @{ P = "RX 7900|RX 7800|RX 7700(?!S)|PRO W7900|PRO W7800|PRO W7700"; A = "gfx1100" }  # RDNA 3 desktop/workstation (Navi 31)
+                    @{ P = "860M|840M|Krackan|AI 7 35[05]|AI 5 34[05]|AI 7 PRO 35|AI 5 33"; A = "gfx1152" }  # RDNA 3.5 (Krackan Point: Radeon 860M/840M, Ryzen AI 7 350 / AI 5 340)
-                    @{ P = "RX 7600|RX 7700S|RX 7650|PRO W7600|PRO W7500|PRO V710"; A = "gfx1102" }  # RDNA 3 (Navi 33)
+                    @{ P = "RX 7900|PRO W7900|PRO W7800";                         A = "gfx1100" }  # RDNA 3 desktop/workstation (Navi 31)
+                    @{ P = "RX 7800|RX 7700(?!S)|PRO W7700|PRO V710";             A = "gfx1101" }  # RDNA 3 (Navi 32)
+                    @{ P = "RX 7600|RX 7700S|RX 7650|PRO W7600|PRO W7500";        A = "gfx1102" }  # RDNA 3 (Navi 33)
                     @{ P = "780M|760M|740M|Phoenix|Hawk Point|Z1 Extreme|Z2 Extreme"; A = "gfx1103" }  # RDNA 3 iGPU (Phoenix / Hawk Point)
                     @{ P = "RX 6900|RX 6800|RX 6750|RX 6700|PRO W6800|PRO W6900";  A = "gfx1030" }  # RDNA 2 (Navi 21) -- gfx103X family
                     @{ P = "RX 6650|RX 6600|PRO W6600|PRO W6650";                  A = "gfx1032" }  # RDNA 2 (Navi 23) -- gfx103X family
@@ -2203,6 +2205,7 @@ exit 0
         $archFamilyMap = @{
             "gfx1201" = "gfx120X-all"; "gfx1200" = "gfx120X-all"  # RDNA 4
             "gfx1151" = "gfx1151";     "gfx1150" = "gfx1150"       # RDNA 3.5 (Strix Halo/Point)
+            "gfx1152" = "gfx1152"                                  # RDNA 3.5 (Krackan Point)
             "gfx1103" = "gfx110X-all"; "gfx1102" = "gfx110X-all"   # RDNA 3
             "gfx1101" = "gfx110X-all"; "gfx1100" = "gfx110X-all"
             "gfx1036" = "gfx103X-all"; "gfx1035" = "gfx103X-all"   # RDNA 2 (RX 6000)
@@ -2224,6 +2227,7 @@ exit 0
         $torchFloorMap = @{
             "gfx1201" = "torch>=2.11.0,<2.12.0"; "gfx1200" = "torch>=2.11.0,<2.12.0"
             "gfx1151" = "torch>=2.11.0,<2.12.0"; "gfx1150" = "torch>=2.11.0,<2.12.0"
+            "gfx1152" = "torch>=2.11.0,<2.12.0"
         }
         # Companion ranges track the torch ceiling so pip resolves a consistent
         # trio on AMD's per-arch index (each published independently). Mirrors
@@ -2231,10 +2235,12 @@ exit 0
         $torchvisionFloorMap = @{
             "gfx1201" = "torchvision>=0.26.0,<0.27.0"; "gfx1200" = "torchvision>=0.26.0,<0.27.0"
             "gfx1151" = "torchvision>=0.26.0,<0.27.0"; "gfx1150" = "torchvision>=0.26.0,<0.27.0"
+            "gfx1152" = "torchvision>=0.26.0,<0.27.0"
         }
         $torchaudioFloorMap = @{
             "gfx1201" = "torchaudio>=2.11.0,<2.12.0"; "gfx1200" = "torchaudio>=2.11.0,<2.12.0"
             "gfx1151" = "torchaudio>=2.11.0,<2.12.0"; "gfx1150" = "torchaudio>=2.11.0,<2.12.0"
+            "gfx1152" = "torchaudio>=2.11.0,<2.12.0"
         }
         $archFamily = if ($ROCmGfxArch -and $archFamilyMap.ContainsKey($ROCmGfxArch)) { $archFamilyMap[$ROCmGfxArch] } else { $null }
         if ($archFamily) {
@@ -2264,7 +2270,7 @@ exit 0
             $_pinRocm211 = ([int]$Matches[1] -eq 7 -and [int]$Matches[2] -eq 2)
         }
         # Only the 2.11-allowlist gfx arches need the floor; others publish <2.11 and stay bare.
-        $_pinGfx211 = @('gfx120x-all', 'gfx1151', 'gfx1150') -contains $_pinLeaf
+        $_pinGfx211 = @('gfx120x-all', 'gfx1151', 'gfx1150', 'gfx1152') -contains $_pinLeaf
         if ($_pinGfx211 -or $_pinRocm211) {
             $ROCmIndexUrl = $TorchIndexUrl
             $ROCmTorchFloor = "torch>=2.11.0,<2.12.0"

install.sh

@@ -257,6 +257,51 @@ run_install_cmd_retry() {
     done
 }
 
+# True when the runtime target is gfx906 (MI50/Radeon VII): the prebuilt AMD
+# bitsandbytes wheel carries no gfx906 kernels, and force-reinstalling it would
+# clobber a user's source-built bnb (the only 4-bit path on this arch) on every
+# `studio update`. So skip the auto-install and leave whatever bnb is present.
+# _gfx906_target is set during torch-index resolution; also honor an explicit
+# UNSLOTH_ROCM_GFX_ARCH so a pinned-index install still skips. The override is
+# normalized (gfx906:sramecc-:xnack- -> gfx906) so a copied HIP gcnArchName counts.
+_is_gfx906_bnb_skip() {
+    [ "${_gfx906_target:-false}" = true ] && return 0
+    _bnb_gfx_env=$(printf '%s' "${UNSLOTH_ROCM_GFX_ARCH:-}" | tr '[:upper:]' '[:lower:]' | tr -d '[:space:]')
+    _bnb_gfx_env=${_bnb_gfx_env%%:*}
+    [ "$_bnb_gfx_env" = "gfx906" ] && return 0
+    # A pinned index (UNSLOTH_TORCH_INDEX_URL/_FAMILY) skips the reroute block that
+    # sets _gfx906_target, so a real gfx906 host with a pinned rocm6.3 index and no
+    # UNSLOTH_ROCM_GFX_ARCH would otherwise clobber a source-built bnb. Probe here
+    # in that gap; skip only when gfx906 is the SOLE distinct arch (mixed hosts
+    # opt in via the env var, mirroring the reroute block's de-dup rule).
+    if [ -z "$_bnb_gfx_env" ] && [ "${_torch_index_pinned:-false}" = true ]; then
+        _bnb_gfx_probe=$(_probe_amd_gfx_arch | awk 'NF && !seen[$0]++')
+        [ "$_bnb_gfx_probe" = "gfx906" ] && return 0
+    fi
+    return 1
+}
+
+# `pip install unsloth` resolves its unconditional bitsandbytes dep to a generic
+# CUDA wheel (no gfx906 kernels) once we skip the prebuilt one. Snapshot bnb before
+# the unsloth install, then drop a freshly pulled wheel afterwards while leaving a
+# pre-existing source build in place.
+_gfx906_bnb_installed() {
+    "$_VENV_PY" -c "import importlib.util as u, sys; sys.exit(0 if u.find_spec('bitsandbytes') else 1)" >/dev/null 2>&1
+}
+_gfx906_bnb_snapshot() {
+    _gfx906_bnb_absent_before=false
+    _is_gfx906_bnb_skip || return 0
+    _gfx906_bnb_installed || _gfx906_bnb_absent_before=true
+}
+_gfx906_bnb_prune() {
+    _is_gfx906_bnb_skip || return 0
+    [ "${_gfx906_bnb_absent_before:-false}" = true ] || return 0
+    _gfx906_bnb_installed || return 0
+    substep "gfx906: removing generic bitsandbytes pulled in as a dependency (no gfx906 kernels; build from source for 4-bit QLoRA)" "$C_WARN"
+    uv pip uninstall --python "$_VENV_PY" bitsandbytes >/dev/null 2>&1 \
+        || "$_VENV_PY" -m pip uninstall -y bitsandbytes >/dev/null 2>&1 || true
+}
+
 # Install bitsandbytes on AMD ROCm hosts. Uses the continuous-release_main
 # wheel for the ROCm 4-bit GEMV fix (bnb PR #1887, post-0.49.2); bnb <= 0.49.2
 # NaNs at decode shape on every AMD GPU. Falls back to PyPI >=0.49.1 if the
@@ -655,6 +700,15 @@ _apt_distro_description() {
     )
 }
 
+# ── Helper: can the controlling terminal actually be opened for reading? ──
+# `test -r` only checks permission bits, which look fine in containers and
+# systemd units where open() then fails with ENXIO. Probe with a real open.
+# The subshell is required: in dash a failed redirection on the special
+# builtin `:` exits the whole script.
+_can_read_tty() {
+    ( : </dev/tty ) >/dev/null 2>&1
+}
+
 # ── Helper: install packages via apt, escalating to sudo only if needed ──
 # Usage: _smart_apt_install pkg1 pkg2 pkg3 ...
 _smart_apt_install() {
@@ -695,24 +749,63 @@ _smart_apt_install() {
         echo "    from your distro's official repositories (not a third-party tarball)."
         echo "    !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!"
         echo ""
-        printf "    Accept? [Y/n] "
+        if _can_read_tty; then
-        if [ -r /dev/tty ]; then
+            printf "    Accept? [Y/n] "
-            read -r REPLY </dev/tty || REPLY="y"
+            # The device opened, so a failed read is EOF, not consent: decline,
-        else
+            # as the autostart prompt below does. Enter is still yes (a
-            REPLY="y"
+            # successful read of an empty line).
-        fi
+            read -r REPLY </dev/tty || REPLY="n"
-        case "$REPLY" in
+            case "$REPLY" in
-            [nN]*)
+                [nN]*)
+                    echo ""
+                    echo "    Please install these packages first, then re-run Unsloth Studio setup:"
+                    echo "    sudo apt-get update -y && sudo apt-get install -y $_STILL_MISSING"
+                    exit 1
+                    ;;
+            esac
+            # Mirror the headless branch: on a sudoers denial, a wrong password
+            # or an apt error, say what to run by hand instead of letting set -e
+            # abort on a bare sudo/apt message.
+            if sudo apt-get update -y </dev/null &&
+                sudo apt-get install -y $_STILL_MISSING </dev/null; then
+                :
+            else
                 echo ""
-                echo "    Please install these packages first, then re-run Unsloth Studio setup:"
+                echo "    Could not install these packages: $_STILL_MISSING"
+                echo "    See the error above."
+                echo "    Please install them first, then re-run Unsloth Studio setup:"
                 echo "    sudo apt-get update -y && sudo apt-get install -y $_STILL_MISSING"
                 exit 1
-                ;;
+            fi
-            *)
+        else
-                sudo apt-get update -y </dev/null
+            # Nobody can answer a prompt or type a password here. -n makes sudo
-                sudo apt-get install -y $_STILL_MISSING </dev/null
+            # refuse rather than prompt into a closed stdin, which is how #7307
-                ;;
+            # died. Probe with the real commands: `sudo -l` answers whether they
-        esac
+            # are *authorized*, not whether running them needs authentication.
+            # -k ignores any cached timestamp, so only a real NOPASSWD rule gets
+            # through, not someone's sudo in another shell minutes ago. Per
+            # sudo(8), -k alongside a command ignores the cached credentials and
+            # "will not update" them, so other sessions keep theirs.
+            echo "    No terminal to confirm on; trying passwordless sudo."
+            if sudo -n -k apt-get update -y </dev/null &&
+                sudo -n -k apt-get install -y $_STILL_MISSING </dev/null; then
+                echo "    Installed with passwordless sudo."
+            else
+                echo ""
+                echo "    Could not install these packages: $_STILL_MISSING"
+                echo "    Detected ${_ad_desc}."
+                # Either sudo refused, or apt failed on a bad repo, dpkg lock or
+                # network outage. sudo exits 1 on an auth/config problem and
+                # when the command cannot be executed, but otherwise passes the
+                # command's own status through, so state both causes.
+                echo "    Either sudo needs a password here, or apt-get itself"
+                echo "    failed; see the error above. With no terminal to"
+                echo "    authenticate on, this cannot be done unattended."
+                echo "    Please install them first, then re-run Unsloth Studio setup:"
+                echo "    sudo apt-get update -y && sudo apt-get install -y $_STILL_MISSING"
+                exit 1
+            fi
+        fi
     else
         echo ""
         echo "    sudo is not available on this system."
@@ -2260,6 +2353,7 @@ _amd_arch_index_family_for_gfx() {
         gfx1201|gfx1200) echo gfx120X-all ;;
         gfx1151) echo gfx1151 ;;
         gfx1150) echo gfx1150 ;;
+        gfx1152) echo gfx1152 ;;
         gfx1103|gfx1102|gfx1101|gfx1100) echo gfx110X-all ;;
         gfx1036|gfx1035|gfx1034|gfx1033|gfx1032|gfx1031|gfx1030) echo gfx103X-all ;;
         gfx90a) echo gfx90a ;;
@@ -2271,12 +2365,14 @@ _amd_arch_index_family_for_gfx() {
 # Map a GPU marketing name to gfx arch (kept in sync with install.ps1 nameArchTable).
 _infer_amd_gfx_arch_from_gpu_name() {
     case "$1" in
-        *"9070 XT"*|*9080*) echo gfx1201 ;;
+        *9070*|*9080*) echo gfx1201 ;;
-        *9070*|*9060*) echo gfx1200 ;;
+        *9060*) echo gfx1200 ;;
         *"8065S"*|*"8060S"*|*"8050S"*|*"8040S"*|*"Strix Halo"*|*"Ryzen AI Max"*|*"AI Max"*) echo gfx1151 ;;
-        *"890M"*|*"880M"*|*"860M"*|*"840M"*|*"Strix Point"*|*"Krackan"*|*"HX 37"*|*"AI 9 HX"*|*"AI 9 36"*|*"AI 7 35"*|*"AI 5 34"*|*"AI 7 PRO 35"*|*"AI 5 33"*) echo gfx1150 ;;
+        *"890M"*|*"880M"*|*"Strix Point"*|*"HX 37"*|*"AI 9 HX"*|*"AI 9 36"*) echo gfx1150 ;;
-        *"RX 7600"*|*"RX 7700S"*|*"RX 7650"*|*"PRO W7600"*|*"PRO W7500"*|*"PRO V710"*) echo gfx1102 ;;
+        *"860M"*|*"840M"*|*"Krackan"*|*"AI 7 35"*|*"AI 5 34"*|*"AI 7 PRO 35"*|*"AI 5 33"*) echo gfx1152 ;;
-        *"RX 7900"*|*"RX 7800"*|*"RX 7700"*|*"PRO W7900"*|*"PRO W7800"*|*"PRO W7700"*) echo gfx1100 ;;
+        *"RX 7600"*|*"RX 7700S"*|*"RX 7650"*|*"PRO W7600"*|*"PRO W7500"*) echo gfx1102 ;;
+        *"RX 7800"*|*"RX 7700"*|*"PRO W7700"*|*"PRO V710"*) echo gfx1101 ;;
+        *"RX 7900"*|*"PRO W7900"*|*"PRO W7800"*) echo gfx1100 ;;
         *"780M"*|*"760M"*|*"740M"*|*"Phoenix"*|*"Hawk Point"*|*"Z1 Extreme"*|*"Z2 Extreme"*) echo gfx1103 ;;
         *"RX 6900"*|*"RX 6800"*|*"RX 6750"*|*"RX 6700"*|*"PRO W6800"*|*"PRO W6900"*) echo gfx1030 ;;
         *"RX 6650"*|*"RX 6600"*|*"PRO W6600"*|*"PRO W6650"*) echo gfx1032 ;;
@@ -2316,10 +2412,14 @@ _infer_linux_amd_gfx_arch() {
         echo gfx1151
         return 0
     fi
-    if [ -n "$_gpu_evidence" ] && grep -qiE '890M|880M|860M|840M|Strix Point|Krackan|HX 37[05]|AI 9 HX|AI 9 36[05]|AI 7 35[05]|AI 5 34[05]|AI 7 PRO 35|AI 5 33' /proc/cpuinfo 2>/dev/null; then
+    if [ -n "$_gpu_evidence" ] && grep -qiE '890M|880M|Strix Point|HX 37[05]|AI 9 HX|AI 9 36[05]' /proc/cpuinfo 2>/dev/null; then
         echo gfx1150
         return 0
     fi
+    if [ -n "$_gpu_evidence" ] && grep -qiE '860M|840M|Krackan|AI 7 35[05]|AI 5 34[05]|AI 7 PRO 35|AI 5 33' /proc/cpuinfo 2>/dev/null; then
+        echo gfx1152
+        return 0
+    fi
     if command -v lspci >/dev/null 2>&1; then
         # A non-AMD controller can enumerate first (Intel/ASPEED before an AMD
         # dGPU), so scan every display-class line and take the first AMD one
@@ -3055,7 +3155,7 @@ if [ "$_torch_index_pinned" = false ] && [ "$SKIP_TORCH" = false ] && \
                     # whole handoff (a user-set override re-exports unchanged).
                     export UNSLOTH_ROCM_GFX_ARCH="$_linux_inferred_gfx"
                     case "$_linux_inferred_gfx" in
-                        gfx1201|gfx1200|gfx1151|gfx1150)
+                        gfx1201|gfx1200|gfx1151|gfx1150|gfx1152)
                             TORCH_CONSTRAINT="torch>=2.11.0,<2.12.0"
                             TORCHVISION_CONSTRAINT="torchvision>=0.26.0,<0.27.0"
                             TORCHAUDIO_CONSTRAINT="torchaudio>=2.11.0,<2.12.0"
@@ -3124,7 +3224,7 @@ fi
 # and a bare name can resolve a 2.12 ABI-mismatched wheel. Match on the FINAL leaf so a
 # custom mirror with a gfx/rocm7.2 path segment but a cu*/cpu family isn't forced.
 case "$_torch_index_leaf" in
-    rocm7.2|gfx120x-all|gfx1151|gfx1150)
+    rocm7.2|gfx120x-all|gfx1151|gfx1150|gfx1152)
         TORCH_CONSTRAINT="torch>=2.11.0,<2.12.0"
         TORCHVISION_CONSTRAINT="torchvision>=0.26.0,<0.27.0"
         TORCHAUDIO_CONSTRAINT="torchaudio>=2.11.0,<2.12.0"
@@ -3241,10 +3341,20 @@ case "$_torch_index_leaf" in
                     if (n > 0) print vals[idx]
                 }')
         fi
+        # An explicit UNSLOTH_ROCM_GFX_ARCH=gfx906 pins the runtime target to the
+        # MI50 / Radeon VII path and must win over Strix probe-order detection on a
+        # mixed Strix + MI50 host, so the Strix reroute is suppressed when it is set.
+        # Normalize a copied HIP gcnArchName (gfx906:sramecc-:xnack- -> gfx906) and
+        # trim whitespace (mirrors the Python .strip()) so the feature-flag suffix or
+        # a stray newline does not defeat the exact gfx906 comparisons below.
+        _gfx906_env=$(printf '%s' "${UNSLOTH_ROCM_GFX_ARCH:-}" | tr '[:upper:]' '[:lower:]' | tr -d '[:space:]')
+        _gfx906_env=${_gfx906_env%%:*}
         _strix_gfx=""
-        case "$_runtime_gfx" in
+        if [ "$_gfx906_env" != "gfx906" ]; then
-            gfx1151|gfx1150) _strix_gfx="$_runtime_gfx" ;;
+            case "$_runtime_gfx" in
-        esac
+                gfx1151|gfx1150|gfx1152) _strix_gfx="$_runtime_gfx" ;;
+            esac
+        fi
         # Skip rocm7.13+ generic indexes: they already ship the fixes, so the
         # arch build (rocm7.13) would be a downgrade rather than a rescue.
         if [ -n "$_strix_gfx" ] && _rocm_leaf_below "$_torch_index_leaf" 7 13; then
@@ -3272,6 +3382,57 @@ case "$_torch_index_leaf" in
             TORCHAUDIO_CONSTRAINT="torchaudio>=2.11.0,<2.12.0"
             _amd_gpu_radeon=false
         fi
+        # ── MI50 / Radeon VII (gfx906, Vega 20): legacy community-supported path ──
+        # Newer rocm wheel families bundle ROCm libraries whose Tensile kernels
+        # dropped gfx906 (rocBLAS "TensileLibrary.dat ... not read for gfx906",
+        # ROCm/TheRock#1844), so a rocm6.4+/7.x index installs a torch that fails
+        # at the first BLAS call. The rocm6.3 index is the last one whose wheels
+        # run on gfx906 (torch 2.7.0 verified on MI50 32GB; up to 2.9 in community
+        # use). Reroute any newer picked index; leave rocm6.0-6.3 alone.
+        #
+        # Target resolution: an explicit UNSLOTH_ROCM_GFX_ARCH wins (lets a host
+        # whose rocminfo/amd-smi emit no gfx token still opt in; _gfx906_env was
+        # lowercased above, before the Strix block it suppresses). Otherwise only
+        # treat gfx906 as the target when it is the SOLE distinct arch present:
+        # _gfx_all is de-duplicated by visible index, which loses per-device
+        # ordinals on a mixed host, so a non-gfx906 selection must never be
+        # downgraded to rocm6.3 -- such hosts set UNSLOTH_ROCM_GFX_ARCH to opt in.
+        _gfx906_target=false
+        if [ -n "$_gfx906_env" ]; then
+            [ "$_gfx906_env" = "gfx906" ] && _gfx906_target=true
+        elif [ -n "$_gfx_all" ]; then
+            _gfx906_uniq=$(printf '%s\n' "$_gfx_all" | awk 'NF && !seen[$0]++')
+            [ "$_gfx906_uniq" = "gfx906" ] && _gfx906_target=true
+        fi
+        # gfx906 always trains from the PyTorch rocm6.3 wheels, never the Radeon repo
+        # (repo.radeon.com wheels carry no gfx906 BLAS kernels). Clear the Radeon
+        # marketing-name flag as soon as gfx906 is the target -- even when the host
+        # already picks rocm6.0-6.3 and the reroute below is a no-op -- so a Radeon VII
+        # does not divert to the radeon branch on those versions.
+        if [ "$_gfx906_target" = true ]; then
+            _amd_gpu_radeon=false
+        fi
+        if [ "$_gfx906_target" = true ] && ! _rocm_leaf_below "$_torch_index_leaf" 6 4; then
+            echo "" >&2
+            echo "  [WARN] gfx906 (MI50 / Radeon VII / Vega 20) detected -- routing torch to the" >&2
+            echo "  [WARN] rocm6.3 index: it is the last wheel family that runs on gfx906 (newer" >&2
+            echo "  [WARN] rocm wheels ship without gfx906 BLAS kernels and fail at first use)." >&2
+            echo "  [WARN] gfx906 is a community-maintained legacy path: 16-bit LoRA and full" >&2
+            echo "  [WARN] finetuning work out of the box; bitsandbytes 4-bit QLoRA requires a" >&2
+            echo "  [WARN] source build of bitsandbytes for gfx906 (see docs.unsloth.ai/amd)." >&2
+            echo "" >&2
+            _amd_gfx906_base="${UNSLOTH_PYTORCH_MIRROR:-https://download.pytorch.org/whl}"
+            while [ "${_amd_gfx906_base%/}" != "$_amd_gfx906_base" ]; do
+                _amd_gfx906_base="${_amd_gfx906_base%/}"
+            done
+            TORCH_INDEX_URL="${_amd_gfx906_base}/rocm6.3"
+            # Reset to the default (<2.11) window: a rocm7.2 pick raised the floor
+            # to 2.11 above, which the rocm6.3 index (torch <= 2.9.x) cannot satisfy.
+            TORCH_CONSTRAINT="torch>=2.4,<2.11.0"
+            TORCHVISION_CONSTRAINT="torchvision>=0.19,<0.26.0"
+            TORCHAUDIO_CONSTRAINT="torchaudio>=2.4,<2.11.0"
+            # (_amd_gpu_radeon already cleared above for every gfx906 target.)
+        fi
         ;;
 esac
 fi  # _torch_index_pinned guard (Radeon + Strix reroute)
@@ -3339,12 +3500,14 @@ elif case "$TORCH_INDEX_URL" in */rocm*|*/gfx*) true ;; *) false ;; esac; then
         # gfx1102 matched BEFORE gfx1100 so the spaceless "RX 7700S" lands on
         # gfx1102 (bash case has no negative lookahead like the PS tables).
         case "$_gpu_disp_mkt" in
-            *"9070 XT"*|*9080*)                                                                            _gpu_disp_gfx="gfx1201" ;;  # RDNA 4
+            *9070*|*9080*)                                                                                 _gpu_disp_gfx="gfx1201" ;;  # RDNA 4 (Navi 48)
-            *9070*|*9060*)                                                                                 _gpu_disp_gfx="gfx1200" ;;  # RDNA 4
+            *9060*)                                                                                        _gpu_disp_gfx="gfx1200" ;;  # RDNA 4 (Navi 44)
             *"8065S"*|*"8060S"*|*"8050S"*|*"8040S"*|*"Strix Halo"*|*"Ryzen AI Max"*|*"AI Max"*) _gpu_disp_gfx="gfx1151" ;;  # RDNA 3.5 (Strix Halo + Gorgon Halo: Radeon 8065S/8060S/8050S/8040S iGPU, Ryzen AI Max / Max+)
-            *"890M"*|*"880M"*|*"860M"*|*"840M"*|*"Strix Point"*|*"Krackan"*|*"HX 37"*|*"AI 9 HX"*|*"AI 9 36"*|*"AI 7 35"*|*"AI 5 34"*|*"AI 7 PRO 35"*|*"AI 5 33"*) _gpu_disp_gfx="gfx1150" ;;  # RDNA 3.5 (Strix/Krackan Point: Radeon 890M/880M iGPU, Ryzen AI 9 HX 370/375)
+            *"890M"*|*"880M"*|*"Strix Point"*|*"HX 37"*|*"AI 9 HX"*|*"AI 9 36"*) _gpu_disp_gfx="gfx1150" ;;  # RDNA 3.5 (Strix Point: Radeon 890M/880M, Ryzen AI 9 HX 370/375)
-            *"RX 7600"*|*"RX 7700S"*|*"RX 7650"*|*"PRO W7600"*|*"PRO W7500"*|*"PRO V710"*)                  _gpu_disp_gfx="gfx1102" ;;  # RDNA 3 (Navi 33)
+            *"860M"*|*"840M"*|*"Krackan"*|*"AI 7 35"*|*"AI 5 34"*|*"AI 7 PRO 35"*|*"AI 5 33"*) _gpu_disp_gfx="gfx1152" ;;  # RDNA 3.5 (Krackan Point: Radeon 860M/840M, Ryzen AI 7 350 / AI 5 340)
-            *"RX 7900"*|*"RX 7800"*|*"RX 7700"*|*"PRO W7900"*|*"PRO W7800"*|*"PRO W7700"*)                  _gpu_disp_gfx="gfx1100" ;;  # RDNA 3 desktop / workstation (Navi 31)
+            *"RX 7600"*|*"RX 7700S"*|*"RX 7650"*|*"PRO W7600"*|*"PRO W7500"*)                              _gpu_disp_gfx="gfx1102" ;;  # RDNA 3 (Navi 33)
+            *"RX 7800"*|*"RX 7700"*|*"PRO W7700"*|*"PRO V710"*)                                            _gpu_disp_gfx="gfx1101" ;;  # RDNA 3 (Navi 32)
+            *"RX 7900"*|*"PRO W7900"*|*"PRO W7800"*)                                                       _gpu_disp_gfx="gfx1100" ;;  # RDNA 3 desktop / workstation (Navi 31)
             *"780M"*|*"760M"*|*"740M"*|*"Phoenix"*|*"Hawk Point"*|*"Z1 Extreme"*|*"Z2 Extreme"*)            _gpu_disp_gfx="gfx1103" ;;  # RDNA 3 iGPU (Phoenix / Hawk Point)
             *"RX 6900"*|*"RX 6800"*|*"RX 6750"*|*"RX 6700"*|*"PRO W6800"*|*"PRO W6900"*)                    _gpu_disp_gfx="gfx1030" ;;  # RDNA 2 (Navi 21)
             *"RX 6650"*|*"RX 6600"*|*"PRO W6600"*|*"PRO W6650"*)                                            _gpu_disp_gfx="gfx1032" ;;  # RDNA 2 (Navi 23)
@@ -3496,6 +3659,7 @@ for _p in ('torch', 'torchvision', 'torchaudio'):
 if [ "$_MIGRATED" = true ]; then
     # Migrated env: force-reinstall unsloth+unsloth-zoo for a clean state, preserving
     # existing torch/CUDA unless the ROCm repair below fires.
+    _gfx906_bnb_snapshot
     substep "upgrading unsloth in migrated environment..."
     if [ "$SKIP_TORCH" = true ]; then
         # No-torch: install unsloth + unsloth-zoo with --no-deps (current
@@ -3537,13 +3701,18 @@ if [ "$_MIGRATED" = true ]; then
     # existing ROCm installs gain the AMD bitsandbytes build without a
     # fresh reinstall.
     if [ "$SKIP_TORCH" = false ] && [ "$_torch_index_is_rocm_family" = true ]; then
-        _install_bnb_rocm "install bitsandbytes (AMD)" "$_VENV_PY"
+        if _is_gfx906_bnb_skip; then
+            substep "gfx906: skipping prebuilt bitsandbytes (no gfx906 kernels); build from source for 4-bit QLoRA -- https://docs.unsloth.ai/get-started/install-and-update/amd" "$C_WARN"
+        else
+            _install_bnb_rocm "install bitsandbytes (AMD)" "$_VENV_PY"
+        fi
         # Repair ROCm torch if overwritten during migrated install
         _has_hip=$("$_VENV_PY" -c "import torch; print(getattr(torch.version,'hip','') or '')" 2>/dev/null || true)
         if [ -z "$_has_hip" ]; then
             substep "repairing ROCm torch (overwritten by dependency resolution)..."
             _install_torch_default_index --force-reinstall
         fi
+        _gfx906_bnb_prune
     fi
 elif [ -n "$TORCH_INDEX_URL" ]; then
     # Fresh: Step 1 - install torch from explicit index (skip when --no-torch or Intel Mac)
@@ -3734,8 +3903,13 @@ elif [ -n "$TORCH_INDEX_URL" ]; then
     # host stays in GGUF-only mode rather than pulling in bitsandbytes,
     # which is only useful once torch is present for training.
     if [ "$SKIP_TORCH" = false ] && [ "$_torch_index_is_rocm_family" = true ]; then
-        _install_bnb_rocm "install bitsandbytes (AMD)" "$_VENV_PY"
+        if _is_gfx906_bnb_skip; then
+            substep "gfx906: skipping prebuilt bitsandbytes (no gfx906 kernels); build from source for 4-bit QLoRA -- https://docs.unsloth.ai/get-started/install-and-update/amd" "$C_WARN"
+        else
+            _install_bnb_rocm "install bitsandbytes (AMD)" "$_VENV_PY"
+        fi
     fi
+    _gfx906_bnb_snapshot
     # Fresh: Step 2 - install unsloth, preserving the torch Step 1 installed
     tauri_log "STEP" "Installing Unsloth"
     substep "installing unsloth (this may take a few minutes)..."
@@ -3786,6 +3960,7 @@ elif [ -n "$TORCH_INDEX_URL" ]; then
             substep "repairing ROCm torch (overwritten by dependency resolution)..."
             _install_torch_default_index --force-reinstall
         fi
+        _gfx906_bnb_prune
     fi
 else
     # Fallback: GPU detection failed to produce a URL -- let uv resolve torch
@@ -4046,9 +4221,11 @@ echo ""
 # In non-interactive environments (Docker, CI, cloud-init) just print instructions.
 if [ "$_SKIP_AUTOSTART" != true ] && [ -t 1 ]; then
     echo ""
-    printf "  Start Unsloth Studio now? [Y/n] "
     # No readable answer (closed/EOF tty) defaults to no; Enter is still yes.
-    if [ -r /dev/tty ]; then
+    # Prompt only when something can answer: `test -r` passes on the unopenable
+    # /dev/tty found in containers, leaving a dangling question in the log.
+    if _can_read_tty; then
+        printf "  Start Unsloth Studio now? [Y/n] "
         read -r _reply </dev/tty || _reply="n"
     else
         _reply="n"

scripts/lint_workflow_triggers.py

@@ -52,14 +52,14 @@ def _normalise_on(on_field):
 
 def _load_workflow(path: Path):
     try:
-        return yaml.safe_load(path.read_text())
+        return yaml.safe_load(path.read_text(encoding = "utf-8"))
     except Exception as exc:
         print(f"ERROR: failed to parse {path}: {exc}", file = sys.stderr)
         sys.exit(2)
 
 
 def _extract_cache_keys(path: Path) -> list[str]:
-    text = path.read_text()
+    text = path.read_text(encoding = "utf-8")
     keys: list[str] = []
     for m in re.finditer(r"(?:^|\n)\s*key:\s*([^\n]+)", text):
         keys.append(m.group(1).strip())
@@ -104,7 +104,7 @@ def main() -> int:
 
         for t in RESTRICTED_TRIGGERS:
             if t in triggers:
-                text = path.read_text()
+                text = path.read_text(encoding = "utf-8")
                 if "lint:workflow_triggers-allow-workflow_run" not in text:
                     findings.append(
                         f"{path.name}: RESTRICTED trigger '{t}' requires an "

scripts/scan_packages_baseline.json

@@ -98,6 +98,14 @@
       "evidence": "L587:                                 while True: sha256:06c2c7f15d73bf192e5e3272c5ff5fcaeff7f6774fef5f4eca6ef473ae50e2b3",
       "evidence_hash": "57acd497f404c203e4450d0580ad85aa8a33406e8d64ad06fbac6cf47d97b24d"
     },
+    {
+      "package": "fastapi",
+      "file": "fastapi/routing.py",
+      "check": "C2 polling/beaconing loop detected",
+      "severity": "CRITICAL",
+      "evidence": "L592:                                 while True: sha256:84283c09277ded3296998b2a6a838744457b606829cf5ab5d0da6f222ff020a0",
+      "evidence_hash": "a7295004315e26a8f3c64fb837521e9fdd7268219bb43e000fb0236ab0259223"
+    },
     {
       "package": "fastmcp-slim",
       "file": "fastmcp/cli/apps_dev.py",

studio/Unsloth_Studio_Colab.ipynb

@@ -1,134 +1,145 @@
 {
- "cells": [
+  "cells": [
-  {
+    {
-   "cell_type": "markdown",
+      "cell_type": "markdown",
-   "metadata": {
+      "metadata": {
-    "id": "view-in-github",
+        "id": "view-in-github",
-    "colab_type": "text"
+        "colab_type": "text"
-   },
+      },
-   "source": [
+      "source": [
-    "<a href=\"https://colab.research.google.com/github/unslothai/unsloth/blob/main/studio/Unsloth_Studio_Colab.ipynb\" target=\"_parent\"><img src=\"https://colab.research.google.com/assets/colab-badge.svg\" alt=\"Open In Colab\"/></a>"
+        "<a href=\"https://colab.research.google.com/github/unslothai/unsloth/blob/main/studio/Unsloth_Studio_Colab.ipynb\" target=\"_parent\"><img src=\"https://colab.research.google.com/assets/colab-badge.svg\" alt=\"Open In Colab\"/></a>"
-   ]
+      ]
+    },
+    {
+      "cell_type": "markdown",
+      "metadata": {
+        "id": "6b87de59"
+      },
+      "source": [
+        "To run this, press \"*Runtime*\" and press \"*Run all*\" on a **free** Tesla T4 Google Colab instance!\n",
+        "<div class=\"align-center\">\n",
+        "<a href=\"https://unsloth.ai/\"><img src=\"https://github.com/unslothai/unsloth/raw/main/images/unsloth%20new%20logo.png\" width=\"115\"></a>\n",
+        "<a href=\"https://discord.gg/unsloth\"><img src=\"https://github.com/unslothai/unsloth/raw/main/images/Discord button.png\" width=\"145\"></a>\n",
+        "<a href=\"https://unsloth.ai/docs/\"><img src=\"https://github.com/unslothai/unsloth/blob/main/images/documentation%20green%20button.png?raw=true\" width=\"125\"></a> Join Discord if you need help + ⭐ <i>Star us on <a href=\"https://github.com/unslothai/unsloth\">Github</a> </i> ⭐\n",
+        "</div>\n",
+        "\n",
+        "To install Unsloth Studio on your local device, follow [our guide](https://unsloth.ai/docs/new/unsloth-studio/install). Unsloth Studio is licensed [AGPL-3.0](https://github.com/unslothai/unsloth/blob/main/studio/LICENSE.AGPL-3.0).\n",
+        "\n",
+        "### Unsloth Studio\n",
+        "\n",
+        "Train and run open models with [**Unsloth Studio**](https://unsloth.ai/docs/new/unsloth-studio/start). NEW! Installation should now only take 2 mins!\n",
+        "\n",
+        "\n",
+        "We are actively working on making Unsloth Studio install on Colab T4 GPUs faster.\n",
+        "\n",
+        "[Features](https://unsloth.ai/docs/new/unsloth-studio#features) • [Quickstart](https://unsloth.ai/docs/new/unsloth-studio/start) • [Data Recipes](https://unsloth.ai/docs/new/unsloth-studio/data-recipe) • [Unsloth Chat](https://unsloth.ai/docs/new/unsloth-studio/chat) • [Export](https://unsloth.ai/docs/new/unsloth-studio/export)"
+      ],
+      "id": "6b87de59"
+    },
+    {
+      "cell_type": "markdown",
+      "metadata": {
+        "id": "e4206349"
+      },
+      "source": [
+        "<p align=\"left\"><img src=\"https://github.com/unslothai/unsloth/raw/main/studio/frontend/public/studio%20github%20landscape%20colab%20display.png\" width=\"600\"></p>"
+      ],
+      "id": "e4206349"
+    },
+    {
+      "cell_type": "markdown",
+      "metadata": {
+        "id": "27da2957"
+      },
+      "source": [
+        "### Setup: Clone repo and run setup"
+      ],
+      "id": "27da2957"
+    },
+    {
+      "cell_type": "code",
+      "metadata": {
+        "id": "27e68f91"
+      },
+      "source": "!git clone --depth 1 --branch main https://github.com/unslothai/unsloth.git\n%cd /content/unsloth\n!chmod +x studio/setup.sh && ./studio/setup.sh --local",
+      "execution_count": null,
+      "outputs": [],
+      "id": "27e68f91"
+    },
+    {
+      "cell_type": "markdown",
+      "metadata": {
+        "id": "3e1771a9"
+      },
+      "source": [
+        "### Start Unsloth Studio"
+      ],
+      "id": "3e1771a9"
+    },
+    {
+      "cell_type": "code",
+      "metadata": {
+        "id": "277e431e"
+      },
+      "source": [
+        "import sys\n",
+        "sys.path.insert(0, \"/content/unsloth/studio/backend\")\n",
+        "from colab import start\n",
+        "\n",
+        "# On Colab, start() auto-opens a Cloudflare link and prints admin login credentials.\n",
+        "# Use the Cloudflare link above the ready card to open Studio (in-cell iframes often stay blank).\n",
+        "start()\n",
+        "\n",
+        "# To skip the Cloudflare tunnel and try the in-notebook proxy iframe only:\n",
+        "# start(cloudflare=False)"
+      ],
+      "execution_count": null,
+      "outputs": [],
+      "id": "277e431e"
+    },
+    {
+      "cell_type": "markdown",
+      "metadata": {
+        "id": "f2b0c6a1"
+      },
+      "source": [
+        "And we're done! If you have any questions on Unsloth, we have a [Discord](https://discord.gg/unsloth) channel! If you find any bugs or want to keep updated with the latest LLM stuff, or need help, join projects etc, feel free to join our Discord!\n",
+        "\n",
+        "Some other resources:\n",
+        "1. Looking to use Unsloth locally? Read our [Installation Guide](https://unsloth.ai/docs/get-started/install) for details on installing Unsloth on Windows, Docker, AMD, Intel GPUs.\n",
+        "2. Learn how to do Reinforcement Learning with our [RL Guide and notebooks](https://unsloth.ai/docs/get-started/reinforcement-learning-rl-guide).\n",
+        "3. Read our guides and notebooks for [Text-to-speech (TTS)](https://unsloth.ai/docs/basics/text-to-speech-tts-fine-tuning) and [vision](https://unsloth.ai/docs/basics/vision-fine-tuning) model support.\n",
+        "4. Explore our [LLM Tutorials Directory](https://unsloth.ai/docs/models/tutorials-how-to-fine-tune-and-run-llms) to find dedicated guides for each model.\n",
+        "5. Need help with Inference? Read our [Inference & Deployment page](https://unsloth.ai/docs/basics/inference-and-deployment) for details on using vLLM, llama.cpp, Ollama etc.\n",
+        "\n",
+        "<div class=\"align-center\">\n",
+        "  <a href=\"https://unsloth.ai\"><img src=\"https://github.com/unslothai/unsloth/raw/main/images/unsloth%20new%20logo.png\" width=\"115\"></a>\n",
+        "  <a href=\"https://discord.gg/unsloth\"><img src=\"https://github.com/unslothai/unsloth/raw/main/images/Discord.png\" width=\"145\"></a>\n",
+        "  <a href=\"https://unsloth.ai/docs/\"><img src=\"https://github.com/unslothai/unsloth/blob/main/images/documentation%20green%20button.png?raw=true\" width=\"125\"></a>\n",
+        "\n",
+        "  Join Discord if you need help + ⭐️ <i>Star us on <a href=\"https://github.com/unslothai/unsloth\">Github</a> </i> ⭐️\n",
+        "\n",
+        "  <b>This notebook is licensed <a href=\"https://github.com/unslothai/unsloth/blob/main/studio/LICENSE.AGPL-3.0\">AGPL-3.0</a></b>\n",
+        "</div>"
+      ],
+      "id": "f2b0c6a1"
+    }
+  ],
+  "metadata": {
+    "accelerator": "GPU",
+    "colab": {
+      "gpuType": "T4",
+      "provenance": [],
+      "include_colab_link": true
+    },
+    "kernelspec": {
+      "display_name": "Python 3",
+      "name": "python3"
+    },
+    "language_info": {
+      "name": "python"
+    }
   },
-  {
+  "nbformat": 4,
-   "cell_type": "markdown",
+  "nbformat_minor": 5
-   "id": "6b87de59",
-   "metadata": {
-    "id": "6b87de59"
-   },
-   "source": [
-    "To run this, press \"*Runtime*\" and press \"*Run all*\" on a **free** Tesla T4 Google Colab instance!\n",
-    "<div class=\"align-center\">\n",
-    "<a href=\"https://unsloth.ai/\"><img src=\"https://github.com/unslothai/unsloth/raw/main/images/unsloth%20new%20logo.png\" width=\"115\"></a>\n",
-    "<a href=\"https://discord.gg/unsloth\"><img src=\"https://github.com/unslothai/unsloth/raw/main/images/Discord button.png\" width=\"145\"></a>\n",
-    "<a href=\"https://unsloth.ai/docs/\"><img src=\"https://github.com/unslothai/unsloth/blob/main/images/documentation%20green%20button.png?raw=true\" width=\"125\"></a> Join Discord if you need help + ⭐ <i>Star us on <a href=\"https://github.com/unslothai/unsloth\">Github</a> </i> ⭐\n",
-    "</div>\n",
-    "\n",
-    "To install Unsloth Studio on your local device, follow [our guide](https://unsloth.ai/docs/new/unsloth-studio/install). Unsloth Studio is licensed [AGPL-3.0](https://github.com/unslothai/unsloth/blob/main/studio/LICENSE.AGPL-3.0).\n",
-    "\n",
-    "### Unsloth Studio\n",
-    "\n",
-    "Train and run open models with [**Unsloth Studio**](https://unsloth.ai/docs/new/unsloth-studio/start). NEW! Installation should now only take 2 mins!\n",
-    "\n",
-    "\n",
-    "We are actively working on making Unsloth Studio install on Colab T4 GPUs faster.\n",
-    "\n",
-    "[Features](https://unsloth.ai/docs/new/unsloth-studio#features) • [Quickstart](https://unsloth.ai/docs/new/unsloth-studio/start) • [Data Recipes](https://unsloth.ai/docs/new/unsloth-studio/data-recipe) • [Unsloth Chat](https://unsloth.ai/docs/new/unsloth-studio/chat) • [Export](https://unsloth.ai/docs/new/unsloth-studio/export)"
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "id": "e4206349",
-   "metadata": {
-    "id": "e4206349"
-   },
-   "source": [
-    "<p align=\"left\"><img src=\"https://github.com/unslothai/unsloth/raw/main/studio/frontend/public/studio%20github%20landscape%20colab%20display.png\" width=\"600\"></p>"
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "id": "27da2957",
-   "metadata": {
-    "id": "27da2957"
-   },
-   "source": [
-    "### Setup: Clone repo and run setup"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": null,
-   "id": "27e68f91",
-   "metadata": {
-    "id": "27e68f91"
-   },
-   "outputs": [],
-   "source": "!git clone --depth 1 --branch main https://github.com/unslothai/unsloth.git\n%cd /content/unsloth\n!chmod +x studio/setup.sh && ./studio/setup.sh --local"
-  },
-  {
-   "cell_type": "markdown",
-   "id": "3e1771a9",
-   "metadata": {
-    "id": "3e1771a9"
-   },
-   "source": [
-    "### Start Unsloth Studio"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": null,
-   "id": "277e431e",
-   "metadata": {
-    "id": "277e431e"
-   },
-   "outputs": [],
-   "source": "import sys\nsys.path.insert(0, \"/content/unsloth/studio/backend\")\nfrom colab import start\n\n# Default: in-tab iframe only. start() blocks to keep the kernel alive.\nstart()\n\n# For a shareable Cloudflare link, replace start() above with:\n# start(cloudflare=True)"
-  },
-  {
-   "cell_type": "markdown",
-   "id": "f2b0c6a1",
-   "metadata": {
-    "id": "f2b0c6a1"
-   },
-   "source": [
-    "And we're done! If you have any questions on Unsloth, we have a [Discord](https://discord.gg/unsloth) channel! If you find any bugs or want to keep updated with the latest LLM stuff, or need help, join projects etc, feel free to join our Discord!\n",
-    "\n",
-    "Some other resources:\n",
-    "1. Looking to use Unsloth locally? Read our [Installation Guide](https://unsloth.ai/docs/get-started/install) for details on installing Unsloth on Windows, Docker, AMD, Intel GPUs.\n",
-    "2. Learn how to do Reinforcement Learning with our [RL Guide and notebooks](https://unsloth.ai/docs/get-started/reinforcement-learning-rl-guide).\n",
-    "3. Read our guides and notebooks for [Text-to-speech (TTS)](https://unsloth.ai/docs/basics/text-to-speech-tts-fine-tuning) and [vision](https://unsloth.ai/docs/basics/vision-fine-tuning) model support.\n",
-    "4. Explore our [LLM Tutorials Directory](https://unsloth.ai/docs/models/tutorials-how-to-fine-tune-and-run-llms) to find dedicated guides for each model.\n",
-    "5. Need help with Inference? Read our [Inference & Deployment page](https://unsloth.ai/docs/basics/inference-and-deployment) for details on using vLLM, llama.cpp, Ollama etc.\n",
-    "\n",
-    "<div class=\"align-center\">\n",
-    "  <a href=\"https://unsloth.ai\"><img src=\"https://github.com/unslothai/unsloth/raw/main/images/unsloth%20new%20logo.png\" width=\"115\"></a>\n",
-    "  <a href=\"https://discord.gg/unsloth\"><img src=\"https://github.com/unslothai/unsloth/raw/main/images/Discord.png\" width=\"145\"></a>\n",
-    "  <a href=\"https://unsloth.ai/docs/\"><img src=\"https://github.com/unslothai/unsloth/blob/main/images/documentation%20green%20button.png?raw=true\" width=\"125\"></a>\n",
-    "\n",
-    "  Join Discord if you need help + ⭐️ <i>Star us on <a href=\"https://github.com/unslothai/unsloth\">Github</a> </i> ⭐️\n",
-    "\n",
-    "  <b>This notebook is licensed <a href=\"https://github.com/unslothai/unsloth/blob/main/studio/LICENSE.AGPL-3.0\">AGPL-3.0</a></b>\n",
-    "</div>"
-   ]
-  }
- ],
- "metadata": {
-  "accelerator": "GPU",
-  "colab": {
-   "gpuType": "T4",
-   "provenance": [],
-   "include_colab_link": true
-  },
-  "kernelspec": {
-   "display_name": "Python 3",
-   "name": "python3"
-  },
-  "language_info": {
-   "name": "python"
-  }
- },
- "nbformat": 4,
- "nbformat_minor": 5
 }

studio/backend/assets/configs/full_finetune.yaml

@@ -30,6 +30,7 @@ lora:
   vision_all_linear: false
   use_rslora: false
   use_loftq: false
+  use_dora: false
   finetune_vision_layers: true
   finetune_language_layers: true
   finetune_attention_modules: true

studio/backend/assets/configs/lora_text.yaml

@@ -30,6 +30,7 @@ lora:
   vision_all_linear: false
   use_rslora: false
   use_loftq: false
+  use_dora: false
   finetune_vision_layers: true
   finetune_language_layers: true
   finetune_attention_modules: true

studio/backend/assets/configs/model_defaults/default.yaml

@@ -33,6 +33,7 @@ lora:
     - "down_proj"
   use_rslora: false
   use_loftq: false
+  use_dora: false
   finetune_vision_layers: true
   finetune_language_layers: true
   finetune_attention_modules: true

studio/backend/assets/configs/model_defaults/embedding/unsloth_Qwen3-Embedding-0.6B.yaml

@@ -34,6 +34,7 @@ lora:
     - "down_proj"
   use_rslora: false
   use_loftq: false
+  use_dora: false
 
 logging:
   enable_wandb: false

studio/backend/assets/configs/model_defaults/embedding/unsloth_all-MiniLM-L6-v2.yaml

@@ -30,6 +30,7 @@ lora:
     - "query"
   use_rslora: false
   use_loftq: false
+  use_dora: false
 
 logging:
   enable_wandb: false

studio/backend/assets/configs/model_defaults/embedding/unsloth_bge-m3.yaml

@@ -30,6 +30,7 @@ lora:
     - "value"
   use_rslora: false
   use_loftq: false
+  use_dora: false
 
 logging:
   enable_wandb: false

studio/backend/assets/configs/model_defaults/embedding/unsloth_embeddinggemma-300m.yaml

@@ -33,6 +33,7 @@ lora:
     - "down_proj"
   use_rslora: false
   use_loftq: false
+  use_dora: false
 
 logging:
   enable_wandb: false

studio/backend/assets/configs/model_defaults/embedding/unsloth_gte-modernbert-base.yaml

@@ -29,6 +29,7 @@ lora:
     - "Wqkv"
   use_rslora: false
   use_loftq: false
+  use_dora: false
 
 logging:
   enable_wandb: false

studio/backend/assets/configs/model_defaults/ernie/unsloth_ERNIE-4.5-21B-A3B-PT.yaml

@@ -34,6 +34,7 @@ lora:
     - "down_proj"
   use_rslora: false
   use_loftq: false
+  use_dora: false
 
 logging:
   enable_wandb: false

studio/backend/assets/configs/model_defaults/ernie/unsloth_ERNIE-4.5-VL-28B-A3B-PT.yaml

@@ -35,6 +35,7 @@ lora:
     - "down_proj"
   use_rslora: false
   use_loftq: false
+  use_dora: false
   finetune_vision_layers: true
   finetune_language_layers: true
   finetune_attention_modules: true

studio/backend/assets/configs/model_defaults/falcon/tiiuae_Falcon-H1-0.5B-Instruct.yaml

@@ -34,6 +34,7 @@ lora:
     - "down_proj"
   use_rslora: false
   use_loftq: false
+  use_dora: false
 
 logging:
   enable_wandb: false

studio/backend/assets/configs/model_defaults/gemma/unsloth_codegemma-7b-bnb-4bit.yaml

@@ -35,6 +35,7 @@ lora:
     - "down_proj"
   use_rslora: false
   use_loftq: false
+  use_dora: false
 
 logging:
   enable_wandb: false

studio/backend/assets/configs/model_defaults/gemma/unsloth_functiongemma-270m-it.yaml

@@ -35,6 +35,7 @@ lora:
     - "down_proj"
   use_rslora: false
   use_loftq: false
+  use_dora: false
 
 logging:
   enable_wandb: false

studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-2-27b-bnb-4bit.yaml

@@ -33,6 +33,7 @@ lora:
     - "down_proj"
   use_rslora: false
   use_loftq: false
+  use_dora: false
 
 logging:
   enable_wandb: false

studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-2-2b.yaml

@@ -34,6 +34,7 @@ lora:
     - "down_proj"
   use_rslora: false
   use_loftq: false
+  use_dora: false
 
 logging:
   enable_wandb: false

studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-3-270m-it.yaml

@@ -35,6 +35,7 @@ lora:
     - "down_proj"
   use_rslora: false
   use_loftq: false
+  use_dora: false
 
 logging:
   enable_wandb: false

studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-3-27b-it.yaml

@@ -29,6 +29,7 @@ lora:
     - "all-linear"
   use_rslora: false
   use_loftq: false
+  use_dora: false
   finetune_vision_layers: true
   finetune_language_layers: true
   finetune_attention_modules: true

studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-3-4b-it.yaml

@@ -29,6 +29,7 @@ lora:
     - "all-linear"
   use_rslora: false
   use_loftq: false
+  use_dora: false
   finetune_vision_layers: true
   finetune_language_layers: true
   finetune_attention_modules: true

studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-3-4b-pt.yaml

@@ -29,6 +29,7 @@ lora:
     - "all-linear"
   use_rslora: false
   use_loftq: false
+  use_dora: false
   finetune_vision_layers: true
   finetune_language_layers: true
   finetune_attention_modules: true

studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-3n-E4B-it.yaml

@@ -29,6 +29,7 @@ lora:
     - "all-linear"
   use_rslora: false
   use_loftq: false
+  use_dora: false
   finetune_vision_layers: true
   finetune_language_layers: true
   finetune_attention_modules: true

studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-3n-E4B.yaml

@@ -29,6 +29,7 @@ lora:
     - "all-linear"
   use_rslora: false
   use_loftq: false
+  use_dora: false
   finetune_vision_layers: true
   finetune_language_layers: true
   finetune_attention_modules: true

studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-4-26B-A4B-it.yaml

@@ -26,6 +26,7 @@ lora:
     - "all-linear"
   use_rslora: false
   use_loftq: false
+  use_dora: false
   finetune_vision_layers: true
   finetune_language_layers: true
   finetune_attention_modules: true

studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-4-26B-A4B.yaml

@@ -26,6 +26,7 @@ lora:
     - "all-linear"
   use_rslora: false
   use_loftq: false
+  use_dora: false
   finetune_vision_layers: true
   finetune_language_layers: true
   finetune_attention_modules: true

studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-4-31B-it.yaml

@@ -26,6 +26,7 @@ lora:
     - "all-linear"
   use_rslora: false
   use_loftq: false
+  use_dora: false
   finetune_vision_layers: true
   finetune_language_layers: true
   finetune_attention_modules: true

studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-4-31B.yaml

@@ -26,6 +26,7 @@ lora:
     - "all-linear"
   use_rslora: false
   use_loftq: false
+  use_dora: false
   finetune_vision_layers: true
   finetune_language_layers: true
   finetune_attention_modules: true

studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-4-E2B-it.yaml

@@ -26,6 +26,7 @@ lora:
     - "all-linear"
   use_rslora: false
   use_loftq: false
+  use_dora: false
   finetune_vision_layers: true
   finetune_language_layers: true
   finetune_attention_modules: true

studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-4-E2B.yaml

@@ -26,6 +26,7 @@ lora:
     - "all-linear"
   use_rslora: false
   use_loftq: false
+  use_dora: false
   finetune_vision_layers: true
   finetune_language_layers: true
   finetune_attention_modules: true

studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-4-E4B-it.yaml

@@ -26,6 +26,7 @@ lora:
     - "all-linear"
   use_rslora: false
   use_loftq: false
+  use_dora: false
   finetune_vision_layers: true
   finetune_language_layers: true
   finetune_attention_modules: true

studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-4-E4B.yaml

@@ -26,6 +26,7 @@ lora:
     - "all-linear"
   use_rslora: false
   use_loftq: false
+  use_dora: false
   finetune_vision_layers: true
   finetune_language_layers: true
   finetune_attention_modules: true

studio/backend/assets/configs/model_defaults/gpt-oss/unsloth_gpt-oss-120b.yaml

@@ -35,6 +35,7 @@ lora:
     - "down_proj"
   use_rslora: false
   use_loftq: false
+  use_dora: false
 
 logging:
   enable_wandb: false

studio/backend/assets/configs/model_defaults/gpt-oss/unsloth_gpt-oss-20b.yaml

@@ -35,6 +35,7 @@ lora:
     - "down_proj"
   use_rslora: false
   use_loftq: false
+  use_dora: false
 
 logging:
   enable_wandb: false

studio/backend/assets/configs/model_defaults/granite/unsloth_granite-4.0-350m-unsloth-bnb-4bit.yaml

@@ -37,6 +37,7 @@ lora:
     - "shared_mlp.output_linear"
   use_rslora: false
   use_loftq: false
+  use_dora: false
 
 logging:
   enable_wandb: false

studio/backend/assets/configs/model_defaults/granite/unsloth_granite-4.0-h-micro.yaml

@@ -37,6 +37,7 @@ lora:
     - "shared_mlp.output_linear"
   use_rslora: false
   use_loftq: false
+  use_dora: false
 
 logging:
   enable_wandb: false

studio/backend/assets/configs/model_defaults/llama/unsloth_Llama-3.2-11B-Vision-Instruct.yaml

@@ -29,6 +29,7 @@ lora:
     - "all-linear"
   use_rslora: false
   use_loftq: false
+  use_dora: false
   finetune_vision_layers: true
   finetune_language_layers: true
   finetune_attention_modules: true

studio/backend/assets/configs/model_defaults/llama/unsloth_Llama-3.2-1B-Instruct.yaml

@@ -34,6 +34,7 @@ lora:
     - "down_proj"
   use_rslora: false
   use_loftq: false
+  use_dora: false
 
 logging:
   enable_wandb: false

studio/backend/assets/configs/model_defaults/llama/unsloth_Llama-3.2-3B-Instruct.yaml

@@ -35,6 +35,7 @@ lora:
     - "down_proj"
   use_rslora: false
   use_loftq: false
+  use_dora: false
 
 logging:
   enable_wandb: false

studio/backend/assets/configs/model_defaults/llama/unsloth_Llama-3.3-70B-Instruct.yaml

@@ -35,6 +35,7 @@ lora:
     - "down_proj"
   use_rslora: false
   use_loftq: false
+  use_dora: false
 
 logging:
   enable_wandb: false

studio/backend/assets/configs/model_defaults/llama/unsloth_Meta-Llama-3.1-70B-bnb-4bit.yaml

@@ -34,6 +34,7 @@ lora:
     - "down_proj"
   use_rslora: false
   use_loftq: false
+  use_dora: false
 
 logging:
   enable_wandb: false

studio/backend/assets/configs/model_defaults/llama/unsloth_Meta-Llama-3.1-8B-Instruct-bnb-4bit.yaml

@@ -34,6 +34,7 @@ lora:
     - "down_proj"
   use_rslora: false
   use_loftq: false
+  use_dora: false
 
 logging:
   enable_wandb: false

studio/backend/assets/configs/model_defaults/llama/unsloth_llama-3-8b-Instruct-bnb-4bit.yaml

@@ -34,6 +34,7 @@ lora:
     - "down_proj"
   use_rslora: false
   use_loftq: false
+  use_dora: false
 
 logging:
   enable_wandb: false

studio/backend/assets/configs/model_defaults/llama/unsloth_llama-3-8b-bnb-4bit.yaml

@@ -34,6 +34,7 @@ lora:
     - "down_proj"
   use_rslora: false
   use_loftq: false
+  use_dora: false
 
 logging:
   enable_wandb: false

studio/backend/assets/configs/model_defaults/llasa/unsloth_Llasa-3B.yaml

@@ -30,6 +30,7 @@ lora:
     - "v_proj"
   use_rslora: false
   use_loftq: false
+  use_dora: false
 
 logging:
   enable_wandb: false

studio/backend/assets/configs/model_defaults/mistral/unsloth_Magistral-Small-2509-unsloth-bnb-4bit.yaml

@@ -35,6 +35,7 @@ lora:
     - "down_proj"
   use_rslora: false
   use_loftq: false
+  use_dora: false
   finetune_vision_layers: true
   finetune_language_layers: true
   finetune_attention_modules: true

studio/backend/assets/configs/model_defaults/mistral/unsloth_Ministral-3-3B-Instruct-2512.yaml

@@ -35,6 +35,7 @@ lora:
     - "down_proj"
   use_rslora: false
   use_loftq: false
+  use_dora: false
   finetune_vision_layers: true
   finetune_language_layers: true
   finetune_attention_modules: true

studio/backend/assets/configs/model_defaults/mistral/unsloth_Mistral-Nemo-Base-2407-bnb-4bit.yaml

@@ -34,6 +34,7 @@ lora:
     - "down_proj"
   use_rslora: false
   use_loftq: false
+  use_dora: false
 
 logging:
   enable_wandb: false

studio/backend/assets/configs/model_defaults/mistral/unsloth_Mistral-Small-Instruct-2409.yaml

@@ -34,6 +34,7 @@ lora:
     - "down_proj"
   use_rslora: false
   use_loftq: false
+  use_dora: false
 
 logging:
   enable_wandb: false

studio/backend/assets/configs/model_defaults/mistral/unsloth_Pixtral-12B-2409.yaml

@@ -29,6 +29,7 @@ lora:
     - "all-linear"
   use_rslora: false
   use_loftq: false
+  use_dora: false
   finetune_vision_layers: true
   finetune_language_layers: true
   finetune_attention_modules: false

studio/backend/assets/configs/model_defaults/mistral/unsloth_mistral-7b-instruct-v0.3-bnb-4bit.yaml

@@ -34,6 +34,7 @@ lora:
     - "down_proj"
   use_rslora: false
   use_loftq: false
+  use_dora: false
 
 logging:
   enable_wandb: false

studio/backend/assets/configs/model_defaults/mistral/unsloth_mistral-7b-v0.3-bnb-4bit.yaml

@@ -33,6 +33,7 @@ lora:
     - "down_proj"
   use_rslora: false
   use_loftq: false
+  use_dora: false
 
 logging:
   enable_wandb: false

studio/backend/assets/configs/model_defaults/other/OuteAI_Llama-OuteTTS-1.0-1B.yaml

@@ -33,6 +33,7 @@ lora:
     - "v_proj"
   use_rslora: false
   use_loftq: false
+  use_dora: false
 
 logging:
   enable_wandb: false

studio/backend/assets/configs/model_defaults/other/Spark-TTS-0.5B_LLM.yaml

@@ -38,6 +38,7 @@ lora:
     - "down_proj"
   use_rslora: false
   use_loftq: false
+  use_dora: false
 
 logging:
   enable_wandb: false

studio/backend/assets/configs/model_defaults/other/sesame_csm-1b.yaml

@@ -37,6 +37,7 @@ lora:
     - "down_proj"
   use_rslora: false
   use_loftq: false
+  use_dora: false
 
 logging:
   enable_wandb: false

studio/backend/assets/configs/model_defaults/other/unsloth_GLM-4.7-Flash.yaml

@@ -35,6 +35,7 @@ lora:
     - "out_proj"
   use_rslora: false
   use_loftq: false
+  use_dora: false
 
 logging:
   enable_wandb: false

studio/backend/assets/configs/model_defaults/other/unsloth_LFM2-1.2B.yaml

@@ -29,6 +29,7 @@ lora:
     - "all-linear"
   use_rslora: false
   use_loftq: false
+  use_dora: false
 
 logging:
   enable_wandb: false

studio/backend/assets/configs/model_defaults/other/unsloth_Nemotron-3-Nano-30B-A3B.yaml

@@ -37,6 +37,7 @@ lora:
     - "out_proj"
   use_rslora: false
   use_loftq: false
+  use_dora: false
 
 logging:
   enable_wandb: false

studio/backend/assets/configs/model_defaults/other/unsloth_PaddleOCR-VL.yaml

@@ -35,6 +35,7 @@ lora:
     - "down_proj"
   use_rslora: false
   use_loftq: false
+  use_dora: false
   finetune_vision_layers: true
   finetune_language_layers: true
   finetune_attention_modules: true

studio/backend/assets/configs/model_defaults/other/unsloth_answerdotai_ModernBERT-large.yaml

@@ -33,6 +33,7 @@ lora:
     - "down_proj"
   use_rslora: false
   use_loftq: false
+  use_dora: false
 
 logging:
   enable_wandb: false

studio/backend/assets/configs/model_defaults/other/unsloth_orpheus-3b-0.1-ft.yaml

@@ -38,6 +38,7 @@ lora:
     - "down_proj"
   use_rslora: false
   use_loftq: false
+  use_dora: false
 
 logging:
   enable_wandb: false

studio/backend/assets/configs/model_defaults/other/unsloth_tinyllama-bnb-4bit.yaml

@@ -34,6 +34,7 @@ lora:
     - "down_proj"
   use_rslora: false
   use_loftq: false
+  use_dora: false
 
 logging:
   enable_wandb: false

studio/backend/assets/configs/model_defaults/other/unsloth_whisper-large-v3.yaml

@@ -33,6 +33,7 @@ lora:
     - "v_proj"
   use_rslora: false
   use_loftq: false
+  use_dora: false
 
 logging:
   enable_wandb: false

studio/backend/assets/configs/model_defaults/phi/unsloth_Phi-3-medium-4k-instruct.yaml

@@ -34,6 +34,7 @@ lora:
     - "down_proj"
   use_rslora: false
   use_loftq: false
+  use_dora: false
 
 logging:
   enable_wandb: false

studio/backend/assets/configs/model_defaults/phi/unsloth_Phi-3.5-mini-instruct.yaml

@@ -34,6 +34,7 @@ lora:
     - "down_proj"
   use_rslora: false
   use_loftq: false
+  use_dora: false
 
 logging:
   enable_wandb: false

studio/backend/assets/configs/model_defaults/phi/unsloth_Phi-4.yaml

@@ -35,6 +35,7 @@ lora:
     - "down_proj"
   use_rslora: false
   use_loftq: false
+  use_dora: false
 
 logging:
   enable_wandb: false

studio/backend/assets/configs/model_defaults/qwen/imdatta0_tiny_qwen3_moe_2.8B_0.7B.yaml

@@ -36,6 +36,7 @@ lora:
     - "gate_up_proj"
   use_rslora: false
   use_loftq: false
+  use_dora: false
 
 logging:
   enable_wandb: false

studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen2-7B.yaml

@@ -34,6 +34,7 @@ lora:
     - "down_proj"
   use_rslora: false
   use_loftq: false
+  use_dora: false
 
 logging:
   enable_wandb: false

studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen2-VL-7B-Instruct.yaml

@@ -29,6 +29,7 @@ lora:
     - "all-linear"
   use_rslora: false
   use_loftq: false
+  use_dora: false
   finetune_vision_layers: true
   finetune_language_layers: true
   finetune_attention_modules: true

studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen2.5-1.5B-Instruct.yaml

@@ -34,6 +34,7 @@ lora:
     - "down_proj"
   use_rslora: false
   use_loftq: false
+  use_dora: false
 
 logging:
   enable_wandb: false

studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen2.5-7B.yaml

@@ -34,6 +34,7 @@ lora:
     - "down_proj"
   use_rslora: false
   use_loftq: false
+  use_dora: false
 
 logging:
   enable_wandb: false

studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen2.5-Coder-1.5B-Instruct.yaml

@@ -34,6 +34,7 @@ lora:
     - "down_proj"
   use_rslora: false
   use_loftq: false
+  use_dora: false
 
 logging:
   enable_wandb: false

studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen2.5-Coder-14B-Instruct.yaml

@@ -35,6 +35,7 @@ lora:
     - "down_proj"
   use_rslora: false
   use_loftq: false
+  use_dora: false
 
 logging:
   enable_wandb: false

studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen2.5-Coder-7B-Instruct-bnb-4bit.yaml

@@ -34,6 +34,7 @@ lora:
     - "down_proj"
   use_rslora: false
   use_loftq: false
+  use_dora: false
 
 logging:
   enable_wandb: false

studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen2.5-VL-7B-Instruct-bnb-4bit.yaml

@@ -29,6 +29,7 @@ lora:
     - "all-linear"
   use_rslora: false
   use_loftq: false
+  use_dora: false
   finetune_vision_layers: true
   finetune_language_layers: true
   finetune_attention_modules: true

studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen3-0.6B.yaml

@@ -35,6 +35,7 @@ lora:
     - "down_proj"
   use_rslora: false
   use_loftq: false
+  use_dora: false
 
 logging:
   enable_wandb: false

studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen3-14B-Base-unsloth-bnb-4bit.yaml

@@ -35,6 +35,7 @@ lora:
     - "down_proj"
   use_rslora: false
   use_loftq: false
+  use_dora: false
 
 logging:
   enable_wandb: false

studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen3-14B.yaml

@@ -35,6 +35,7 @@ lora:
     - "down_proj"
   use_rslora: false
   use_loftq: false
+  use_dora: false
 
 logging:
   enable_wandb: false

studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen3-30B-A3B-Instruct-2507.yaml

@@ -36,6 +36,7 @@ lora:
     - "gate_up_proj"
   use_rslora: false
   use_loftq: false
+  use_dora: false
 
 logging:
   enable_wandb: false

studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen3-32B.yaml

@@ -35,6 +35,7 @@ lora:
     - "down_proj"
   use_rslora: false
   use_loftq: false
+  use_dora: false
 
 logging:
   enable_wandb: false

studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen3-4B-Instruct-2507.yaml

@@ -35,6 +35,7 @@ lora:
     - "down_proj"
   use_rslora: false
   use_loftq: false
+  use_dora: false
 
 logging:
   enable_wandb: false

studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen3-4B-Thinking-2507.yaml

@@ -35,6 +35,7 @@ lora:
     - "down_proj"
   use_rslora: false
   use_loftq: false
+  use_dora: false
 
 logging:
   enable_wandb: false

studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen3-VL-8B-Instruct-unsloth-bnb-4bit.yaml

@@ -29,6 +29,7 @@ lora:
     - "all-linear"
   use_rslora: false
   use_loftq: false
+  use_dora: false
   finetune_vision_layers: true
   finetune_language_layers: true
   finetune_attention_modules: true

studio/backend/assets/configs/vision_lora.yaml

@@ -30,6 +30,7 @@ lora:
   vision_all_linear: true
   use_rslora: false
   use_loftq: false
+  use_dora: false
   finetune_vision_layers: true
   finetune_language_layers: true
   finetune_attention_modules: true

studio/backend/auth/authentication.py

@@ -164,6 +164,21 @@ async def get_current_subject_allow_password_change(
     )
 
 
+# The literal the examples ship with; pasted unedited more often than a revoked key.
+API_KEY_PLACEHOLDER = f"{API_KEY_PREFIX}YOUR_KEY"
+
+
+def _invalid_api_key_detail(token: str) -> str:
+    """Why the key failed. Only the example placeholder is called out; every real
+    key gets one indistinguishable message, so this leaks no key existence."""
+    if token == API_KEY_PLACEHOLDER:
+        return (
+            "This is the placeholder key from the example. Create an API key in "
+            f"Unsloth Studio under Settings > API and use it in place of {API_KEY_PLACEHOLDER}."
+        )
+    return "Invalid or expired API key"
+
+
 async def _get_current_subject(
     credentials: HTTPAuthorizationCredentials, *, allow_password_change: bool
 ) -> str:
@@ -176,7 +191,7 @@ async def _get_current_subject(
         if username is None:
             raise HTTPException(
                 status_code = status.HTTP_401_UNAUTHORIZED,
-                detail = "Invalid or expired API key",
+                detail = _invalid_api_key_detail(token),
             )
         return username
 

studio/backend/auth/storage.py

@@ -44,7 +44,7 @@ def generate_bootstrap_password() -> str:
 
     # Persisted from a previous run?
     if _BOOTSTRAP_PW_PATH.is_file():
-        _bootstrap_password = _BOOTSTRAP_PW_PATH.read_text().strip()
+        _bootstrap_password = _BOOTSTRAP_PW_PATH.read_text(encoding = "utf-8").strip()
         if _bootstrap_password:
             return _bootstrap_password
 
@@ -57,7 +57,7 @@ def generate_bootstrap_password() -> str:
 
     # Persist so the same passphrase survives restarts until password change.
     ensure_dir(_BOOTSTRAP_PW_PATH.parent)
-    _BOOTSTRAP_PW_PATH.write_text(_bootstrap_password)
+    _BOOTSTRAP_PW_PATH.write_text(_bootstrap_password, encoding = "utf-8")
     try:
         os.chmod(_BOOTSTRAP_PW_PATH, 0o600)
     except OSError:
@@ -76,7 +76,7 @@ def _load_bootstrap_password() -> Optional[str]:
     global _bootstrap_password
     _bootstrap_password = None
     if _BOOTSTRAP_PW_PATH.is_file():
-        bootstrap_password = _BOOTSTRAP_PW_PATH.read_text().strip()
+        bootstrap_password = _BOOTSTRAP_PW_PATH.read_text(encoding = "utf-8").strip()
         if bootstrap_password:
             _bootstrap_password = bootstrap_password
     return _bootstrap_password
@@ -99,7 +99,7 @@ def clear_bootstrap_password() -> None:
             # stale plaintext can't be re-seeded by generate_bootstrap_password()
             # if a later reset-password deletes auth.db and re-validates it.
             try:
-                _BOOTSTRAP_PW_PATH.write_text("")
+                _BOOTSTRAP_PW_PATH.write_text("", encoding = "utf-8")
                 cleared = True
             except OSError:
                 cleared = False

studio/backend/colab.py

@@ -1,9 +1,7 @@
 # SPDX-License-Identifier: AGPL-3.0-only
 # Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
-"""
+"""Colab helpers for Unsloth Studio. Uses Colab's built-in proxy."""
-Colab helpers for Unsloth Studio. Uses Colab's built-in proxy.
-"""
 
 from pathlib import Path
 import sys
@@ -22,11 +20,9 @@ logger = get_logger(__name__)
 
 
 def get_colab_url(port: int = 8888) -> str:
-    """
+    """Get the Colab proxy URL for a port.
-    Get the Colab proxy URL for a port.
 
-    Retries up to 3 times, validating the result is a real HTTPS Colab URL.
+    Retries 3x validating a real HTTPS Colab URL; falls back to localhost on failure.
-    Falls back to http://localhost:{port} only when all attempts fail.
     """
     import time as _time
 
@@ -55,28 +51,243 @@ def get_colab_url(port: int = 8888) -> str:
     return fallback
 
 
-def show_link(port: int = 8888, *, _url: "str | None" = None):
+def _short_colab_url(url: str, port: int) -> str:
-    """Display a styled clickable link to the UI.
+    """Truncated display form of a Colab proxy URL; falls back to the full URL."""
-
-    *_url* is an optional pre-fetched proxy URL; pass it to avoid a second eval_js round-trip.
-    """
-    from IPython.display import display, HTML
-
-    url = _url if _url is not None else get_colab_url(port)
-
-    # Truncated display URL; try/except so an odd URL shape still renders the link.
     try:
         port_prefix = f"{port}-"
         idx = url.index(port_prefix)
         next_dash = url.index("-", idx + len(port_prefix))
-        short_url = url[: next_dash + 1] + "..."
+        return url[: next_dash + 1] + "..."
     except (ValueError, IndexError):
-        short_url = url
+        return url
 
-    # Plain-text line so the URL shows even if HTML display fails.
-    logger.info(f"🌐 Unsloth Studio URL: {url}")
 
-    html = f"""
+def _is_colab_proxy_url(url: str, port: int) -> bool:
+    """True when *url* looks like a real Colab kernel proxy, not a localhost fallback."""
+    return bool(url and isinstance(url, str) and url.startswith("https://") and str(port) in url)
+
+
+def _is_colab_runtime() -> bool:
+    """True on a hosted Colab notebook kernel.
+
+    Reuses the backend's main Colab detector (``/content`` + Colab env / ``google.colab``)
+    instead of a single env var, which is not always present on hosted runtimes.
+    """
+    try:
+        from main import _IS_COLAB
+        return bool(_IS_COLAB)
+    except Exception:
+        return False
+
+
+def _colab_login_credentials_path() -> Path:
+    from auth.storage import DB_PATH
+    return DB_PATH.parent / ".colab_notebook_login"
+
+
+def _store_colab_login_credentials(username: str, password: str) -> None:
+    """Persist Colab admin credentials for notebook re-runs after interrupt."""
+    path = _colab_login_credentials_path()
+    try:
+        path.parent.mkdir(parents = True, exist_ok = True)
+        path.write_text(f"{username}\n{password}\n", encoding = "utf-8")
+        try:
+            import os
+            os.chmod(path, 0o600)
+        except OSError:
+            pass
+    except OSError as e:
+        logger.info(f"Could not persist Colab login credentials ({e}).")
+
+
+def _load_colab_login_credentials() -> "tuple[str, str] | None":
+    """Return stored Colab admin credentials from a previous ``start()`` run, if any."""
+    path = _colab_login_credentials_path()
+    try:
+        if not path.is_file():
+            return None
+        lines = path.read_text(encoding = "utf-8").splitlines()
+        if len(lines) >= 2 and lines[0] and lines[1]:
+            return lines[0], lines[1]
+    except (OSError, UnicodeDecodeError) as e:
+        logger.info(f"Could not load Colab login credentials ({e}).")
+    return None
+
+
+def _clear_colab_login_credentials() -> None:
+    """Drop the cached Colab credentials once they no longer authenticate."""
+    path = _colab_login_credentials_path()
+    try:
+        path.unlink(missing_ok = True)
+    except OSError as e:
+        logger.info(f"Could not clear Colab login credentials ({e}).")
+
+
+def _colab_credentials_still_valid(username: str, password: str) -> bool:
+    """True when *password* still matches the stored admin hash.
+
+    Guards against redisplaying a cached first-run password after the user has
+    changed the admin password through the app, which would print credentials
+    that no longer authenticate to the current Cloudflare tunnel.
+    """
+    try:
+        from auth.storage import get_user_and_secret
+        from auth.hashing import verify_password
+    except Exception as e:
+        logger.info(f"Could not load auth to validate cached Colab credentials ({e}).")
+        return False
+    try:
+        row = get_user_and_secret(username)
+        if not row:
+            return False
+        salt, pwd_hash = row[0], row[1]
+        return bool(verify_password(password, salt, pwd_hash))
+    except Exception as e:
+        logger.info(f"Could not validate cached Colab credentials ({e}).")
+        return False
+
+
+def _colab_wants_cloudflare(cloudflare: "bool | None") -> bool:
+    """Resolve whether to open a Cloudflare tunnel.
+
+    ``None`` auto-enables on real Colab (the in-cell proxy embed is often blank);
+    pass ``False`` to opt out.
+    """
+    if cloudflare is not None:
+        return cloudflare
+    return _is_colab_runtime()
+
+
+def _finalize_colab_admin_password() -> "tuple[str, str] | None":
+    """Clear the bootstrap-password gate on Colab so Cloudflare tunnels can start.
+
+    Returns ``(username, password)`` for display in the notebook. On first run the
+    random admin password is finalized; on later runs (e.g. after interrupt) the
+    stored credentials are re-displayed so the Cloudflare link stays usable.
+    Anyone who can read this cell already controls the runtime.
+    """
+    if not _is_colab_runtime():
+        return None
+    try:
+        from auth.storage import (
+            DEFAULT_ADMIN_USERNAME,
+            ensure_default_admin,
+            generate_bootstrap_password,
+            get_bootstrap_password,
+            requires_password_change,
+            update_password,
+        )
+    except Exception as e:
+        logger.warning(
+            f"Could not load auth for Colab setup ({e}); Cloudflare link may be blocked."
+        )
+        return None
+
+    try:
+        ensure_default_admin()
+        username = DEFAULT_ADMIN_USERNAME
+        if not requires_password_change(username):
+            creds = _load_colab_login_credentials()
+            if creds is not None and _colab_credentials_still_valid(username, creds[1]):
+                return creds
+            # The admin password was changed through the app after the first run,
+            # so the cached copy is stale; drop it instead of printing dead credentials.
+            _clear_colab_login_credentials()
+            return None
+        password = get_bootstrap_password() or generate_bootstrap_password()
+        if not update_password(username, password):
+            logger.warning(
+                "Could not finalize Colab admin password; Cloudflare link may be blocked."
+            )
+            return None
+        _store_colab_login_credentials(username, password)
+        return username, password
+    except Exception as e:
+        logger.warning(
+            f"Could not finalize Colab admin password ({e}); Cloudflare link may be blocked."
+        )
+        return None
+
+
+def _colab_login_html(username: str, password: str) -> str:
+    """Notebook card with Colab admin credentials (shown once after auto-finalize)."""
+    return f"""
+    <div style="display: inline-block; padding: 20px; background: #ffffff; border: 2px solid #000000;
+                border-radius: 12px; margin: 10px 0; font-family: system-ui, -apple-system, sans-serif;">
+        <h2 style="color: #000000; margin: 0 0 12px 0; font-size: 22px; font-weight: 800;">
+            Unsloth Studio Login (Colab)
+        </h2>
+        <p style="color: #333333; margin: 0 0 12px 0; font-size: 14px; font-weight: bold;">
+            Log in as <code>{username}</code> with this password. This cell is visible only in
+            your notebook session.
+        </p>
+        <p style="color: #333333; margin: 0; font-size: 14px; font-family: monospace; font-weight: bold;">
+            Password: <code>{password}</code>
+        </p>
+    </div>
+    """
+
+
+def _show_colab_login_credentials(username: str, password: str) -> None:
+    """Display Colab admin credentials in the notebook output."""
+    from IPython.display import HTML, display
+
+    logger.info(f"🔐 Unsloth Studio login — user: {username}")
+    display(HTML(_colab_login_html(username, password)))
+
+
+def _ready_card_html(
+    url: str,
+    port: int,
+    *,
+    has_cloudflare_link: bool = False,
+    cloudflare_requested: bool = False,
+) -> str:
+    """Branded ready card for the in-notebook Studio view.
+
+    Colab ``*.prod.colab.dev`` proxy URLs are session-scoped and 404 when opened as a
+    top-level tab or on another device, so never ``window.open`` them. On real Colab the
+    Cloudflare link is the supported entry point because in-cell proxy embeds often stay blank.
+    """
+    short_url = _short_colab_url(url, port)
+    if _is_colab_runtime() or _is_colab_proxy_url(url, port):
+        if has_cloudflare_link:
+            embed_note = (
+                "Open Studio with the Cloudflare link above. In-cell proxy previews on "
+                "current Colab often stay blank, so the tunnel link is the supported path."
+            )
+        elif cloudflare_requested:
+            embed_note = (
+                "Could not open a Cloudflare tunnel, so Studio may be unreachable on Colab. "
+                "Check the logs above and re-run this cell. Pass "
+                '<code style="background:#f3f3f3;padding:2px 6px;border-radius:4px;">'
+                "cloudflare=True</code> after fixing any tunnel errors."
+            )
+        else:
+            embed_note = (
+                "Colab proxy links cannot be opened in a new tab (they 404 outside this "
+                'notebook). Re-run with <code style="background:#f3f3f3;padding:2px 6px;'
+                'border-radius:4px;">start(cloudflare=True)</code> for a working link.'
+            )
+        return f"""
+    <div style="display: inline-block; padding: 20px; background: #ffffff; border: 2px solid #000000;
+                border-radius: 12px; margin: 10px 0; font-family: system-ui, -apple-system, sans-serif;">
+        <h2 style="color: #000000; margin: 0 0 12px 0; font-size: 26px; font-weight: 800;
+                   display: flex; align-items: center; gap: 12px;">
+            <img src="https://github.com/unslothai/unsloth/raw/main/studio/frontend/public/unsloth-gem.png"
+                 height="48" style="display:block;">
+            Unsloth Studio is Ready!
+        </h2>
+        <p style="color: #333333; margin: 0 0 8px 0; font-size: 15px; font-weight: bold;">
+            {embed_note}
+        </p>
+        <p style="color: #666666; margin: 16px 0 0 0; font-size: 13px; font-family: monospace; font-weight: bold;">
+            {short_url}
+        </p>
+    </div>
+    """
+
+    return f"""
     <div style="display: inline-block; padding: 20px; background: #ffffff; border: 2px solid #000000;
                 border-radius: 12px; margin: 10px 0; font-family: system-ui, -apple-system, sans-serif;">
         <h2 style="color: #000000; margin: 0 0 12px 0; font-size: 26px; font-weight: 800;
@@ -100,15 +311,52 @@ def show_link(port: int = 8888, *, _url: "str | None" = None):
         </p>
     </div>
     """
-    display(HTML(html))
+
+
+def show_link(
+    port: int = 8888,
+    *,
+    _url: "str | None" = None,
+    has_cloudflare_link: bool = False,
+    cloudflare_requested: bool = False,
+):
+    """Display a styled ready card for the UI.
+
+    Colab proxy URLs are informational only (no new-tab open; they 404 outside the cell);
+    non-proxy URLs keep a clickable open button. *_url* is an optional pre-fetched proxy
+    URL to avoid a second eval_js round-trip.
+    """
+    from IPython.display import display, HTML
+
+    url = _url if _url is not None else get_colab_url(port)
+    logger.info(f"🌐 Unsloth Studio URL: {url}")
+    display(
+        HTML(
+            _ready_card_html(
+                url,
+                port,
+                has_cloudflare_link = has_cloudflare_link,
+                cloudflare_requested = cloudflare_requested,
+            )
+        )
+    )
+
+
+def _warn_colab_cloudflare_missing(*, use_cloudflare: bool, cloudflare_url: "str | None") -> None:
+    """Log a prominent warning when Colab expected a tunnel but none was opened."""
+    if not use_cloudflare or cloudflare_url or not _is_colab_runtime():
+        return
+    logger.warning(
+        "Colab Cloudflare tunnel unavailable — Studio is unlikely to be reachable in this "
+        "notebook. Check the logs above for tunnel or auth errors, then re-run start()."
+    )
 
 
 def _bootstrap_password_pending() -> bool:
     """True while the default admin still owes a bootstrap-password change.
 
-    While pending, main.py injects that password into same-origin GETs, and a public
+    While pending, a public tunnel GET (no Origin) reads as same-origin and gets the
-    tunnel GET (no Origin) reads as same-origin, so sharing the link would leak admin
+    injected password, so sharing the link would leak admin access. Fails safe to pending.
-    access. Fails safe to pending if the state cannot be read.
     """
     try:
         from auth.storage import requires_password_change, DEFAULT_ADMIN_USERNAME
@@ -121,9 +369,8 @@ def _bootstrap_password_pending() -> bool:
 def start_cloudflare_tunnel(port: int) -> "str | None":
     """Open a shareable Cloudflare quick tunnel to localhost:*port*, or None.
 
-    run_server suppresses the tunnel on Colab by design, so we start it directly.
+    run_server suppresses the tunnel on Colab, so we start it directly. Refused while the
-    Refused while the bootstrap password is pending; any failure collapses to None
+    bootstrap password is pending; any failure collapses to None (Colab proxy still works).
-    and the Colab proxy still works.
     """
     if _bootstrap_password_pending():
         logger.warning(
@@ -152,9 +399,9 @@ def start_cloudflare_tunnel(port: int) -> "str | None":
 def _publish_cloudflare_url(cloudflare_url: "str | None") -> None:
     """Publish a directly-started tunnel URL onto app.state so /api/health advertises it.
 
-    run_server only sets this when it opens the tunnel itself, which it skips on Colab,
+    run_server sets this only when it opens the tunnel itself (skipped on Colab), so we
-    so we set it here. Otherwise the frontend's API examples fall back to an
+    set it here; otherwise the frontend's API examples fall back to an unreachable
-    unreachable server_url. Best-effort.
+    server_url. Best-effort.
     """
     if not cloudflare_url:
         return
@@ -183,8 +430,7 @@ def _stop_cloudflare_tunnel() -> None:
 def _is_studio_healthy(port: int, timeout: float = 2.0) -> bool:
     """True only if Unsloth Studio (not some other app) answers /api/health on *port*.
 
-    The service-marker check stops the reuse path reusing or tunneling a foreign
+    The service-marker check stops the reuse path reusing or tunneling a foreign process.
-    process that merely serves /api/health.
     """
     import json, urllib.request
     try:
@@ -194,8 +440,29 @@ def _is_studio_healthy(port: int, timeout: float = 2.0) -> bool:
         return False
 
 
-def _shareable_link_html(cloudflare_url: str) -> str:
+def _shareable_link_html(
-    """Branded card for the shareable Cloudflare link, styled like the show_link banner."""
+    cloudflare_url: str,
+    password: "str | None" = None,
+    username: "str | None" = None,
+) -> str:
+    """Branded card for the shareable Cloudflare link, styled like the show_link banner.
+
+    *password* renders under the link so the credential sits in the card with the button
+    it unlocks. The username is always the default admin, so it reads inline.
+    """
+    login_block = ""
+    if password:
+        login_block = f"""
+        <p style="color: #000000; margin: 16px 0 0 0; font-size: 20px; font-weight: 800;">
+            Password
+        </p>
+        <p style="margin: 6px 0 0 0;"><code style="display: inline-block; font-size: 24px;
+            font-weight: 800; text-decoration: underline; background: #f3f3f3;
+            padding: 4px 10px; border-radius: 6px;">{password}</code></p>
+        <p style="color: #666666; margin: 6px 0 0 0; font-size: 12px;">
+            Log in as <code>{username}</code> with this password. Shown only in your
+            notebook session, and never included in the shared link.
+        </p>"""
     return f"""
     <div style="display: inline-block; padding: 20px; background: #ffffff; border: 2px solid #000000;
                 border-radius: 12px; margin: 10px 0; font-family: system-ui, -apple-system, sans-serif;">
@@ -213,40 +480,55 @@ def _shareable_link_html(cloudflare_url: str) -> str:
             Open Unsloth Studio
         </a>
         <p style="color: #333333; margin: 12px 0 0 0; font-size: 14px; font-weight: bold;">
-            This Cloudflare HTTPS link works from any device — share it with anyone. The Colab view below only works in this tab.
+            This Cloudflare HTTPS link works from any device, so you can share it with anyone.
         </p>
         <p style="color: #333333; margin: 16px 0 0 0; font-size: 13px; font-family: monospace; font-weight: bold;">
-            🔗 {cloudflare_url}
+            🔗 <a href="{cloudflare_url}" onclick="var w=window.open(this.href,'_blank');if(!w){{return true;}}return false;"
-        </p>
+                  style="color: #000000; text-decoration: underline; cursor: pointer;">{cloudflare_url}</a>
+        </p>{login_block}
     </div>
     """
 
 
-def _show_and_embed(port: int, *, cloudflare_url: "str | None" = None):
+# Height for serve_kernel_port_as_iframe (~82vh on a 1080p screen, clamped).
-    """Render the Unsloth header + iframe for *port*, with a shareable-link card above
+_COLAB_IFRAME_HEIGHT = 900
-    when *cloudflare_url* is set. Falls back to serve_kernel_port_as_iframe."""
-    url = get_colab_url(port)
-    logger.info(f"🌐 Unsloth Studio URL: {url}")
-    if cloudflare_url:
-        logger.info(f"🔗 Shareable Cloudflare link: {cloudflare_url}")
 
+
+def _embed_kernel_port_iframe(port: int) -> bool:
+    """Embed Studio via Colab's native kernel-port iframe helper.
+
+    Only trusted on a real Colab runtime: colabtools can import ``google.colab`` and
+    queue browser-side JS without appending an iframe, so callers outside Colab must use
+    the HTML iframe path instead.
+    """
+    if not _is_colab_runtime():
+        return False
+    try:
+        from google.colab import output as colab_output
+    except ImportError:
+        return False
+    try:
+        colab_output.serve_kernel_port_as_iframe(
+            port,
+            height = _COLAB_IFRAME_HEIGHT,
+            width = "100%",
+        )
+        return True
+    except Exception as e:
+        logger.info(f"serve_kernel_port_as_iframe failed ({e}); trying HTML iframe.")
+        return False
+
+
+def _embed_html_iframe(url: str, port: int) -> bool:
+    """Fallback embed: raw HTML iframe when the Colab helper is unavailable."""
     try:
         from IPython.display import HTML, display
+    except ImportError:
+        return False
 
-        iframe_id = f"unsloth-studio-{port}"
+    short_url = _short_colab_url(url, port)
-
+    iframe_id = f"unsloth-studio-{port}"
-        # Truncated header URL — best-effort, falls back to full URL.
+    try:
-        try:
-            port_prefix = f"{port}-"
-            idx = url.index(port_prefix)
-            next_dash = url.index("-", idx + len(port_prefix))
-            short_url = url[: next_dash + 1] + "..."
-        except (ValueError, IndexError):
-            short_url = url
-
-        if cloudflare_url:
-            display(HTML(_shareable_link_html(cloudflare_url)))
-
         display(
             HTML(f"""
 <div style="font-family:system-ui,-apple-system,sans-serif;margin:8px 0;
@@ -266,41 +548,110 @@ def _show_and_embed(port: int, *, cloudflare_url: "str | None" = None):
 </div>
 """)
         )
-    except Exception:
+        return True
-        # Fallback: Colab's built-in helper.
+    except Exception as e:
+        logger.info(f"HTML iframe embed failed ({e}).")
+        return False
+
+
+def _show_and_embed(
+    port: int,
+    *,
+    cloudflare_url: "str | None" = None,
+    colab_login: "tuple[str, str] | None" = None,
+    cloudflare_requested: bool = False,
+):
+    """Render the Unsloth ready card + iframe for *port*.
+
+    Prefer Colab's ``serve_kernel_port_as_iframe`` on real Colab; raw HTML iframe is the
+    fallback. Cloudflare cards stay clickable.
+    """
+    url = get_colab_url(port)
+    logger.info(f"🌐 Unsloth Studio URL: {url}")
+    if cloudflare_url:
+        logger.info(f"🔗 Shareable Cloudflare link: {cloudflare_url}")
+
+    _warn_colab_cloudflare_missing(
+        use_cloudflare = cloudflare_requested,
+        cloudflare_url = cloudflare_url,
+    )
+
+    # Fold the credentials into the link card rather than a second card below it.
+    credentials_shown = False
+    if cloudflare_url:
         try:
-            from google.colab import output as colab_output
+            from IPython.display import HTML, display
-            colab_output.serve_kernel_port_as_iframe(port, height = 900, width = "100%")
+
-        except ImportError:
+            username, password = colab_login if colab_login else (None, None)
-            pass
+            display(HTML(_shareable_link_html(cloudflare_url, password, username)))
+            credentials_shown = bool(colab_login)
+        except Exception as e:
+            logger.info(f"Could not render Cloudflare link card ({e}).")
+
+    if colab_login and not credentials_shown:
+        try:
+            _show_colab_login_credentials(*colab_login)
+        except Exception as e:
+            logger.info(f"Could not render Colab login card ({e}).")
+
+    # With a tunnel up the embed below is skipped, so the ready card would only restate
+    # the link card and print a proxy URL that 404s outside this tab.
+    skip_ready_card = _is_colab_runtime() and bool(cloudflare_url)
+    if not skip_ready_card:
+        try:
+            show_link(
+                port,
+                _url = url,
+                has_cloudflare_link = bool(cloudflare_url),
+                cloudflare_requested = cloudflare_requested,
+            )
+        except Exception as e:
+            logger.info(f"Could not render Unsloth link card ({e}).")
+
+    # On Colab with a working tunnel, skip the in-cell proxy embed (often blank).
+    if _is_colab_runtime() and cloudflare_url:
+        return
+
+    # Real Colab: kernel helper needs only the port (works when eval_js failed).
+    if _is_colab_runtime():
+        if _embed_kernel_port_iframe(port):
+            return
+    _embed_html_iframe(url, port)
 
 
-def start(port: int = 8888, *, cloudflare: bool = False):
+def start(port: int = 8888, *, cloudflare: "bool | None" = None):
     """Start Unsloth Studio in Colab and display the URL.
 
     Args:
         port: Port to bind/serve on.
-        cloudflare: Opt in to a shareable Cloudflare HTTPS link reachable from any
+        cloudflare: Shareable Cloudflare HTTPS link. ``None`` (default) auto-enables on
-            device (default OFF). It exposes Unsloth's login page beyond Colab, so it
+            real Colab because the in-cell proxy embed is often blank; pass ``False`` to
-            stays an explicit opt-in; the default shows only the in-tab proxy iframe.
+            skip the tunnel or ``True`` to force it on other runtimes.
 
     Usage:
-        start()                    # Colab-proxy iframe only (default)
+        start()                    # Cloudflare link on Colab (auto); proxy iframe elsewhere
-        start(cloudflare=True)     # also open a shareable Cloudflare link
+        start(cloudflare=False)    # Colab proxy iframe only (often blank on current Colab)
+        start(cloudflare=True)     # force Cloudflare link on any runtime
     """
     import time
 
     logger.info("🦥 Starting Unsloth Studio...")
+    use_cloudflare = _colab_wants_cloudflare(cloudflare)
 
-    # Fast path: Unsloth already running (cell re-run). Re-launching would collide on
+    # Fast path: already running (cell re-run); re-show link/iframe instead of rebinding the port.
-    # the port, so just re-show the link and iframe.
     if _is_studio_healthy(port):
         logger.info(f"   Unsloth is already running on port {port} — reusing existing server.")
         # try/finally: tear the tunnel down even if interrupted mid-start/render.
         try:
-            cf_url = start_cloudflare_tunnel(port) if cloudflare else None
+            colab_login = _finalize_colab_admin_password() if use_cloudflare else None
+            cf_url = start_cloudflare_tunnel(port) if use_cloudflare else None
             _publish_cloudflare_url(cf_url)
-            _show_and_embed(port, cloudflare_url = cf_url)
+            _show_and_embed(
+                port,
+                cloudflare_url = cf_url,
+                colab_login = colab_login,
+                cloudflare_requested = use_cloudflare,
+            )
             for _ in range(10000):
                 time.sleep(300)
                 print("=", end = "", flush = True)
@@ -313,7 +664,6 @@ def start(port: int = 8888, *, cloudflare: bool = False):
     logger.info("   Loading backend...")
     from run import run_server
 
-    # Auto-detect frontend path
     repo_root = Path(__file__).parent.parent
     frontend_path = repo_root / "frontend" / "dist"
 
@@ -323,8 +673,7 @@ def start(port: int = 8888, *, cloudflare: bool = False):
 
     logger.info("   Starting server...")
     try:
-        # cloudflare=False: this helper owns the tunnel (Colab's own
+        # cloudflare=False: this helper owns the tunnel (via start(cloudflare=...)), so pin it off.
-        # start(cloudflare=...) drives it), so pin it off explicitly.
         app = run_server(
             host = "0.0.0.0",
             port = port,
@@ -339,14 +688,12 @@ def start(port: int = 8888, *, cloudflare: bool = False):
         logger.error(f"❌ Unsloth Studio failed to start: {exc}")
         return
 
-    # run_server auto-increments the port if in use; read back the bound port so the
+    # run_server may auto-increment the port; read back the bound port for the proxy URL/iframe.
-    # proxy URL and iframe point at the right place.
     actual_port: int = getattr(getattr(app, "state", None), "server_port", None) or port
 
     logger.info(f"   Server started on port {actual_port}!")
 
-    # Poll health endpoint before showing the link — avoids the race where ready_event
+    # Poll health before showing the link: avoids the race where ready_event fires pre-bind.
-    # fires but the process hasn't finished binding.
     import urllib.request
 
     server_ready = False
@@ -365,12 +712,17 @@ def start(port: int = 8888, *, cloudflare: bool = False):
         )
         return
 
-    # Open the tunnel now the server is healthy, publish its URL for /api/health, and
+    # Server healthy: finalize Colab auth, open the tunnel, publish URL, tear down on interrupt.
-    # tear it down on interrupt (try/finally) rather than orphan the process.
     try:
-        cf_url = start_cloudflare_tunnel(actual_port) if cloudflare else None
+        colab_login = _finalize_colab_admin_password() if use_cloudflare else None
+        cf_url = start_cloudflare_tunnel(actual_port) if use_cloudflare else None
         _publish_cloudflare_url(cf_url)
-        _show_and_embed(actual_port, cloudflare_url = cf_url)
+        _show_and_embed(
+            actual_port,
+            cloudflare_url = cf_url,
+            colab_login = colab_login,
+            cloudflare_requested = use_cloudflare,
+        )
 
         # Keep kernel alive so the daemon server thread runs.
         for _ in range(10000):

studio/backend/core/export/export.py

@@ -81,6 +81,82 @@ _PYTORCH_MISSING_MESSAGE = (
 _LLAMA_CPP_SCRIPTS_WARNING_EMITTED = False
 
 
+def _multi_gpu_device_map_kwargs() -> dict:
+    """``device_map`` kwargs for sharding a checkpoint across every visible GPU.
+
+    unsloth's ``from_pretrained`` defaults to ``device_map="sequential"``, which stacks
+    the whole model on GPU0 and OOMs multi-GPU hosts whose other GPUs sit empty (#7053).
+    Returns ``{"device_map": "balanced"}`` only on a real multi-GPU CUDA/ROCm host
+    (mirroring the inference loader's ``get_device_map``), else empty so single-GPU, CPU
+    and MLX loads keep the loader default."""
+    if _IS_MLX:
+        return {}
+    try:
+        from utils.hardware import get_device_map, get_parent_visible_gpu_ids
+
+        visible = get_parent_visible_gpu_ids()
+        if len(visible) > 1:
+            device_map = get_device_map(visible)
+        elif not visible:
+            # UUID/MIG masks resolve to no numeric ids; get_device_map(None) falls back
+            # to the visible-GPU count, so a multi-GPU UUID/MIG host still shards.
+            device_map = get_device_map(None)
+        else:
+            return {}
+        if device_map == "balanced":
+            return {"device_map": device_map}
+    except Exception as exc:
+        logger.debug(f"multi-GPU device_map resolution failed; using loader default: {exc}")
+    return {}
+
+
+def _is_oom_error(exc: BaseException) -> bool:
+    """True for an accelerator OOM, however it is spelled.
+
+    accelerate and transformers re-raise it as a plain ``RuntimeError`` on several paths
+    and ROCm/XPU use their own classes, so match the message too.
+    """
+    if torch is not None:
+        oom_types = tuple(
+            t
+            for t in (
+                getattr(torch, "OutOfMemoryError", None),
+                getattr(getattr(torch, "cuda", None), "OutOfMemoryError", None),
+                getattr(getattr(torch, "xpu", None), "OutOfMemoryError", None),
+            )
+            if isinstance(t, type)
+        )
+        if oom_types and isinstance(exc, oom_types):
+            return True
+    return "out of memory" in f"{type(exc).__name__}: {exc}".lower()
+
+
+def _is_cpu_spill_rejection(exc: BaseException) -> bool:
+    """bitsandbytes refuses a map that spills to CPU/disk with a plain ``ValueError``.
+
+    Busy secondary GPUs can make ``balanced`` spill to CPU even where the old sequential
+    load fit on GPU0, and that message says nothing about memory, so the retry has to
+    match it explicitly. See transformers ``quantizers/quantizer_bnb_4bit.py``.
+    """
+    return "dispatched on the cpu or the disk" in str(exc).lower()
+
+
+class _CpuSpillRetry(Exception):
+    """A multi-GPU load that succeeded but left modules offloaded to CPU/disk."""
+
+
+def _cpu_offloaded_modules(model) -> int:
+    """Count the modules a load parked on CPU or disk.
+
+    Only bitsandbytes refuses such a map; a full-precision load accepts it, leaves the
+    parameters on meta and dies much later in safetensors with "Cannot copy out of meta
+    tensor". Nothing raises at load time, so inspect the map directly. PEFT re-dispatches
+    when attaching an adapter, so in practice this catches merged checkpoints.
+    """
+    device_map = getattr(model, "hf_device_map", None) or {}
+    return sum(1 for target in device_map.values() if str(target) in ("cpu", "disk"))
+
+
 def _supports_kwarg(fn, name):
     """True if `fn` accepts keyword `name` directly or via **kwargs."""
     import inspect
@@ -165,7 +241,7 @@ def _offline_window_if(local_files_only):
 def _is_wsl():
     """Detect if running under Windows Subsystem for Linux."""
     try:
-        return "microsoft" in open("/proc/version").read().lower()
+        return "microsoft" in open("/proc/version", encoding = "utf-8").read().lower()
     except Exception:
         return False
 
@@ -271,6 +347,7 @@ class ExportBackend:
         load_in_4bit: bool = True,
         trust_remote_code: bool = False,
         hf_token: Optional[str] = None,
+        _device_map_override: Optional[dict] = None,
     ) -> Tuple[bool, str]:
         """
         Load a checkpoint for export.
@@ -303,6 +380,14 @@ class ExportBackend:
             # Skip the Hub when offline so a no-internet export uses the local cache.
             local_files_only = _hf_offline()
 
+            # Shard across every visible GPU instead of stacking on GPU0 (#7053); {} on
+            # single-GPU/CPU/MLX. _device_map_override is the single-device retry below.
+            _device_map_kw = (
+                _multi_gpu_device_map_kwargs()
+                if _device_map_override is None
+                else _device_map_override
+            )
+
             # Run the type-detection probes in the forced-offline window (else a gated
             # base 404s); it covers is_vision_model's Hub reads + the transformers-5
             # subprocess, and local_files_only makes detect_audio_type's requests.get skip.
@@ -328,6 +413,7 @@ class ExportBackend:
                     trust_remote_code = trust_remote_code,
                     token = token,
                     local_files_only = local_files_only,
+                    **_device_map_kw,
                 )
 
             elif self._audio_type == "whisper":
@@ -343,6 +429,7 @@ class ExportBackend:
                     trust_remote_code = trust_remote_code,
                     token = token,
                     local_files_only = local_files_only,
+                    **_device_map_kw,
                 )
 
             elif self._audio_type == "snac":
@@ -355,6 +442,7 @@ class ExportBackend:
                     trust_remote_code = trust_remote_code,
                     token = token,
                     local_files_only = local_files_only,
+                    **_device_map_kw,
                 )
 
             elif self._audio_type == "bicodec":
@@ -368,6 +456,7 @@ class ExportBackend:
                     trust_remote_code = trust_remote_code,
                     token = token,
                     local_files_only = local_files_only,
+                    **_device_map_kw,
                 )
 
             elif self._audio_type == "dac":
@@ -380,6 +469,7 @@ class ExportBackend:
                     trust_remote_code = trust_remote_code,
                     token = token,
                     local_files_only = local_files_only,
+                    **_device_map_kw,
                 )
 
             elif self.is_vision:
@@ -392,6 +482,7 @@ class ExportBackend:
                     trust_remote_code = trust_remote_code,
                     token = token,
                     local_files_only = local_files_only,
+                    **_device_map_kw,
                 )
                 tokenizer = processor  # vision: processor acts as tokenizer
 
@@ -405,8 +496,16 @@ class ExportBackend:
                     trust_remote_code = trust_remote_code,
                     token = token,
                     local_files_only = local_files_only,
+                    **_device_map_kw,
                 )
 
+            # Only when we asked for the multi-GPU map: a single-GPU host has no second
+            # placement to retry on, so leave its behaviour untouched.
+            _offloaded = _cpu_offloaded_modules(model) if _device_map_kw else 0
+            if _device_map_override is None and _offloaded:
+                del model
+                raise _CpuSpillRetry(f"{_offloaded} module(s) offloaded to CPU/disk")
+
             if _IS_MLX:
                 # MLX doesn't use PeftModel — detect LoRA via adapter_config.json
                 self.is_peft = adapter_config.exists()
@@ -429,11 +528,41 @@ class ExportBackend:
             return True, f"Loaded {model_type} model{peft_info} successfully"
 
         except Exception as e:
-            logger.error(f"Error loading checkpoint: {e}")
+            # Sharding is an optimisation, never a requirement. "balanced" budgets from the
-            import traceback
+            # free memory read BEFORE this process opens a CUDA context on each GPU, so when
+            # a training or chat job already owns the others the shard can OOM, or spill to
+            # CPU and be refused by bitsandbytes, where the old single-device load succeeded.
+            # Fall back once before giving up.
+            if (
+                _device_map_override is None
+                and (
+                    isinstance(e, _CpuSpillRetry) or _is_oom_error(e) or _is_cpu_spill_rejection(e)
+                )
+                and _multi_gpu_device_map_kwargs()
+            ):
+                # Retry outside this block: the live traceback pins the half-built model's
+                # frames, so an in-block retry inherits the exhausted device.
+                retry_reason = str(e)
+            else:
+                logger.error(f"Error loading checkpoint: {e}")
+                import traceback
 
-            logger.error(traceback.format_exc())
+                logger.error(traceback.format_exc())
-            return False, f"Failed to load checkpoint: {str(e)}"
+                return False, f"Failed to load checkpoint: {str(e)}"
+
+        logger.warning(
+            f"Multi-GPU export load unusable ({retry_reason}); retrying on "
+            f"the single-device loader default."
+        )
+        self.cleanup_memory()
+        return self.load_checkpoint(
+            checkpoint_path,
+            max_seq_length = max_seq_length,
+            load_in_4bit = load_in_4bit,
+            trust_remote_code = trust_remote_code,
+            hf_token = hf_token,
+            _device_map_override = {},
+        )
 
     def _write_export_metadata(self, save_directory: str):
         """Write export_metadata.json with base model info for Chat page discovery."""
@@ -445,7 +574,7 @@ class ExportBackend:
             )
             metadata = {"base_model": base_model}
             metadata_path = os.path.join(save_directory, "export_metadata.json")
-            with open(metadata_path, "w") as f:
+            with open(metadata_path, "w", encoding = "utf-8") as f:
                 json.dump(metadata, f, indent = 2)
             logger.info(f"Wrote export metadata to {metadata_path}")
         except Exception as e:
@@ -1048,6 +1177,21 @@ class ExportBackend:
                     "Use the safetensors adapter instead.",
                     None,
                 )
+            # llama.cpp's convert_lora_to_gguf.py has no concept of DoRA's
+            # lora_magnitude_vector tensors: it only reads the standard
+            # lora_A/lora_B delta, so exporting a DoRA adapter would silently
+            # drop the magnitude rescaling and produce a GGUF LoRA file that
+            # loads fine but no longer matches the trained model.
+            _peft_config = getattr(self.current_model, "peft_config", {}).get("default")
+            if getattr(_peft_config, "use_dora", False):
+                return (
+                    False,
+                    "GGUF LoRA export is not supported for DoRA adapters: the GGUF LoRA "
+                    "format has no way to represent DoRA's magnitude vectors, so the "
+                    "exported file would silently lose the DoRA behavior. Use the "
+                    "safetensors adapter instead, or merge to a full GGUF model.",
+                    None,
+                )
             outtype = str(gguf_outtype).lower()
             if outtype not in _GGUF_LORA_OUTTYPES:
                 return (

studio/backend/core/inference/_vulkan_probe.py

@@ -6,12 +6,14 @@
 Run in a short-lived subprocess (``python _vulkan_probe.py <bindir>``) so the
 Vulkan instance never lives in the long-running backend process. Loads the
 bundled ggml Vulkan backend from ``<bindir>`` and prints one
-``<idx>\\t<free_bytes>\\t<is_igpu>\\t<total_bytes>`` line per device to stdout.
+``<idx>\\t<free_bytes>\\t<is_igpu>\\t<total_bytes>\\t<name>`` line per device to
-Indices are ggml's own Vulkan device ordinals, which need not match nvidia-smi
+stdout. Indices are ggml's own Vulkan device ordinals, which need not match
-order. ``is_igpu`` (from ggml's device type) is ``1`` for an integrated GPU
+nvidia-smi order. ``is_igpu`` (from ggml's device type) is ``1`` for an
-sharing system RAM. ``total_bytes`` is the device-local heap; the reader uses
+integrated GPU sharing system RAM. ``total_bytes`` is the device-local heap;
-it to reserve absolute headroom on a discrete card (parity with the CUDA/ROCm
+the reader uses it to reserve absolute headroom on a discrete card (parity
-fit) and ignores it for an iGPU, whose "VRAM" is shared system RAM.
+with the CUDA/ROCm fit) and ignores it for an iGPU, whose "VRAM" is shared
+system RAM. ``name`` is ggml's device description (the marketing name, e.g.
+"AMD Radeon RX 9070 XT"); empty when the registry lookup fails.
 
 Uses only the standard library so it stays runnable as a bare script.
 """
@@ -24,15 +26,30 @@ import sys
 _GGML_BACKEND_DEVICE_TYPE_IGPU = 2
 
 
-def _igpu_flags(base, lib, count: int) -> list[bool]:
+def _igpu_flags_and_names(base, lib, count: int) -> tuple[list[bool], list[str]]:
-    """Per-device integrated-GPU flags via ggml's backend registry.
+    """Per-device integrated-GPU flags and descriptions via ggml's backend registry.
 
     The Vulkan reg enumerates devices in the same order as
     ``ggml_backend_vk_get_device_memory`` (each context uses ``ctx->device =
-    i``), so reg index == device ordinal. Returns all-False on any failure so
+    i``), so reg index == device ordinal. Returns all-False / empty-name on any
-    the reader never over-caps a discrete card.
+    failure so the reader never over-caps a discrete card and the memory
+    readings still get through.
     """
     flags = [False] * count
+    names = [""] * count
+
+    # The name lookup is bound OUTSIDE the type-detection try: a ggml-base
+    # without ggml_backend_dev_description (older/custom build) must degrade to
+    # unnamed devices, not abort before the iGPU flags are read (which would
+    # count an iGPU's shared RAM as VRAM).
+    describe = None
+    try:
+        base.ggml_backend_dev_description.restype = ctypes.c_char_p
+        base.ggml_backend_dev_description.argtypes = [ctypes.c_void_p]
+        describe = base.ggml_backend_dev_description
+    except Exception:
+        pass
+
     try:
         lib.ggml_backend_vk_reg.restype = ctypes.c_void_p
         lib.ggml_backend_vk_reg.argtypes = []
@@ -45,17 +62,31 @@ def _igpu_flags(base, lib, count: int) -> list[bool]:
 
         reg = lib.ggml_backend_vk_reg()
         if not reg:
-            return flags
+            return flags, names
         dev_count = base.ggml_backend_reg_dev_count(reg)
         for i in range(min(count, dev_count)):
             dev = base.ggml_backend_reg_dev_get(reg, i)
             if dev:
                 flags[i] = base.ggml_backend_dev_type(dev) == _GGML_BACKEND_DEVICE_TYPE_IGPU
+                if describe is not None:
+                    try:
+                        desc = describe(dev)
+                        if desc:
+                            # Tabs/newlines would corrupt the line protocol;
+                            # spaces are safe.
+                            names[i] = (
+                                desc.decode("utf-8", errors = "replace")
+                                .replace("\t", " ")
+                                .replace("\n", " ")
+                                .strip()
+                            )
+                    except Exception:
+                        pass
     except Exception:
-        # Best-effort: any failure degrades to "discrete" so the memory
+        # Best-effort: any failure degrades to "discrete"/"unnamed" so the
-        # readings still get through instead of crashing the probe.
+        # memory readings still get through instead of crashing the probe.
         pass
-    return flags
+    return flags, names
 
 
 def main() -> int:
@@ -63,6 +94,14 @@ def main() -> int:
         return 0
     bindir = sys.argv[1]
 
+    # Device names can be non-ASCII (localized drivers); the platform-default
+    # stdout encoding (e.g. cp1252) would raise on them and lose the whole
+    # inventory. The reader decodes UTF-8 with the same error mode.
+    try:
+        sys.stdout.reconfigure(encoding = "utf-8", errors = "replace")
+    except Exception:
+        pass
+
     # Hold add_dll_directory's handle for the rest of main() (the documented
     # idiom) so bindir stays on the search path while the sibling ggml DLLs
     # resolve below.
@@ -96,12 +135,12 @@ def main() -> int:
     ]
 
     count = lib.ggml_backend_vk_get_device_count()
-    igpu = _igpu_flags(base, lib, count)
+    igpu, names = _igpu_flags_and_names(base, lib, count)
     rows = []
     for i in range(count):
         free, total = ctypes.c_size_t(0), ctypes.c_size_t(0)
         lib.ggml_backend_vk_get_device_memory(i, ctypes.byref(free), ctypes.byref(total))
-        rows.append("%d\t%d\t%d\t%d" % (i, free.value, int(igpu[i]), total.value))
+        rows.append("%d\t%d\t%d\t%d\t%s" % (i, free.value, int(igpu[i]), total.value, names[i]))
     sys.stdout.write("\n".join(rows))
     return 0
 

studio/backend/core/inference/api_monitor.py

@@ -52,6 +52,13 @@ class ApiMonitorEntry:
     total_tokens: Optional[int] = None
     total_tokens_authoritative: bool = False
     error: Optional[str] = None
+    # "request" (HTTP call) or "lifecycle" (model load/unload: event/reason, not a prompt; shared).
+    kind: str = "request"
+    event: Optional[str] = None
+    reason: Optional[str] = None
+    shared: bool = False
+    # 0-100 for a running download row; None when not applicable.
+    progress: Optional[float] = None
 
     def snapshot(self, *, include_details: bool = True) -> dict[str, Any]:
         duration_ms = None
@@ -85,6 +92,10 @@ class ApiMonitorEntry:
             "completion_tokens": self.completion_tokens,
             "total_tokens": self.total_tokens,
             "error": self.error,
+            "kind": self.kind,
+            "event": self.event,
+            "reason": self.reason,
+            "progress": self.progress,
         }
         if include_details:
             payload["prompt"] = self.prompt
@@ -127,6 +138,73 @@ class ApiMonitor:
             self._trim_terminal_locked()
         return entry.id
 
+    def record_lifecycle(
+        self,
+        *,
+        event: str,
+        model: str,
+        reason: Optional[str] = None,
+        running: bool = False,
+    ) -> str:
+        """Record a model load/unload alongside the request traffic that caused it.
+
+        ``running=True`` opens the row for the caller to close with :meth:`finish` /
+        :meth:`fail`; an unload is terminal on arrival. Rows are shared (visible to
+        every subject) and share the request retention budget.
+        """
+        now = time.time()
+        entry = ApiMonitorEntry(
+            id = f"apievt_{uuid.uuid4().hex[:12]}",
+            endpoint = f"model.{event}",
+            method = "",
+            model = model or "default",
+            prompt = "",
+            status = "running" if running else "completed",
+            started_at = now,
+            updated_at = now,
+            started_monotonic = time.monotonic(),
+            finished_at = None if running else now,
+            finished_monotonic = None if running else time.monotonic(),
+            kind = "lifecycle",
+            event = event,
+            reason = reason,
+            shared = True,
+        )
+        with self._lock:
+            self._entries.appendleft(entry)
+            self._trim_terminal_locked()
+        return entry.id
+
+    def relabel(self, entry_id: Optional[str], model: str) -> None:
+        """Rename an open lifecycle row once the load resolves its real id: up front
+        the caller only has the load path, which may be an HF snapshot dir."""
+        if not entry_id or not model:
+            return
+        with self._lock:
+            entry = self._find_locked(entry_id)
+            if entry is not None:
+                entry.model = model
+                entry.updated_at = time.time()
+
+    def set_progress(self, entry_id: Optional[str], progress: Optional[float]) -> None:
+        """Update an open download row's percentage (clamped to 0-100)."""
+        if not entry_id or progress is None:
+            return
+        with self._lock:
+            entry = self._find_locked(entry_id)
+            if entry is not None and entry.status == "running":
+                entry.progress = min(100.0, max(0.0, float(progress)))
+                entry.updated_at = time.time()
+
+    def discard(self, entry_id: Optional[str]) -> None:
+        """Drop a row that turned out not to be an event (an already-satisfied load)."""
+        if not entry_id:
+            return
+        with self._lock:
+            entry = self._find_locked(entry_id)
+            if entry is not None:
+                self._entries.remove(entry)
+
     def append_reply(self, entry_id: Optional[str], text: str) -> None:
         if not entry_id or not text:
             return
@@ -212,6 +290,18 @@ class ApiMonitor:
             self._entries.appendleft(entry)
             self._trim_terminal_locked()
 
+    def fail_open(self, entry_id: Optional[str], error: str) -> None:
+        """Fail only a still-open row: unlike :meth:`fail`, a catch-all in a
+        ``finally`` cannot stamp an error onto a request that already succeeded."""
+        if not entry_id:
+            return
+        with self._lock:
+            entry = self._find_locked(entry_id)
+            if entry is None or entry.finished_at is not None:
+                return
+            # Same lock as the check, so a finish() cannot land in between.
+            self._fail_locked(entry, error)
+
     def fail(self, entry_id: Optional[str], error: str) -> None:
         if not entry_id:
             return
@@ -224,15 +314,18 @@ class ApiMonitor:
                 if error:
                     entry.error = _trim(error, 1000)
                 return
-            now = time.time()
+            self._fail_locked(entry, error)
-            entry.status = "error"
+
-            entry.error = _trim(error, 1000)
+    def _fail_locked(self, entry: ApiMonitorEntry, error: str) -> None:
-            entry.updated_at = now
+        now = time.time()
-            entry.finished_at = now
+        entry.status = "error"
-            entry.finished_monotonic = time.monotonic()
+        entry.error = _trim(error, 1000)
-            self._entries.remove(entry)
+        entry.updated_at = now
-            self._entries.appendleft(entry)
+        entry.finished_at = now
-            self._trim_terminal_locked()
+        entry.finished_monotonic = time.monotonic()
+        self._entries.remove(entry)
+        self._entries.appendleft(entry)
+        self._trim_terminal_locked()
 
     def snapshot(
         self,
@@ -244,7 +337,7 @@ class ApiMonitor:
             return [
                 entry.snapshot(include_details = include_details)
                 for entry in self._entries
-                if subject is None or entry.subject == subject
+                if self._visible(entry, subject)
             ]
 
     def get(
@@ -257,22 +350,29 @@ class ApiMonitor:
             entry = self._find_locked(entry_id)
             if entry is None:
                 return None
-            if subject is not None and entry.subject != subject:
+            if not self._visible(entry, subject):
                 return None
             return entry.snapshot(include_details = True)
 
     def active_count(self, *, subject: Optional[str] = None) -> int:
+        # Lifecycle rows show as "running" while loading but are not in-flight API requests.
         with self._lock:
             return sum(
                 1
                 for entry in self._entries
-                if entry.status == "running" and (subject is None or entry.subject == subject)
+                if entry.status == "running"
+                and entry.kind != "lifecycle"
+                and (subject is None or entry.subject == subject)
             )
 
     def clear(self) -> None:
         with self._lock:
             self._entries.clear()
 
+    @staticmethod
+    def _visible(entry: ApiMonitorEntry, subject: Optional[str]) -> bool:
+        return subject is None or entry.subject == subject or entry.shared
+
     def _find_locked(self, entry_id: str) -> Optional[ApiMonitorEntry]:
         for entry in self._entries:
             if entry.id == entry_id:

studio/backend/core/inference/chat_template_helpers.py

@@ -326,6 +326,58 @@ def _normalize_tool_call_arguments(messages: list) -> list:
     return out if mutated else messages
 
 
+def _take_tool_result(pending: list, call_id) -> Optional[dict]:
+    if call_id:
+        for i, result in enumerate(pending):
+            if result.get("tool_call_id") == call_id:
+                return pending.pop(i)
+    for i, result in enumerate(pending):
+        if not result.get("tool_call_id"):
+            return pending.pop(i)
+    return None
+
+
+def _split_parallel_tool_calls(messages: list) -> list:
+    """Llama 3.x templates render one call per message, so split parallel calls
+    into consecutive single-call messages, each followed by its own result."""
+    if not any(isinstance(m, dict) and len(m.get("tool_calls") or ()) > 1 for m in messages):
+        return messages
+
+    out: list = []
+    i = 0
+    total = len(messages)
+    while i < total:
+        msg = messages[i]
+        calls = msg.get("tool_calls") if isinstance(msg, dict) else None
+        if not calls or len(calls) <= 1:
+            out.append(msg)
+            i += 1
+            continue
+
+        # Tool results right after this message answer its calls.
+        j = i + 1
+        pending: list = []
+        while (
+            j < total
+            and isinstance(messages[j], dict)
+            and messages[j].get("role") in ("tool", "ipython")
+        ):
+            pending.append(messages[j])
+            j += 1
+
+        for idx, call in enumerate(calls):
+            piece = {**msg, "tool_calls": [call]}
+            if idx:
+                piece["content"] = ""
+            out.append(piece)
+            result = _take_tool_result(pending, call.get("id") if isinstance(call, dict) else None)
+            if result is not None:
+                out.append(result)
+        out.extend(pending)
+        i = j
+    return out
+
+
 def apply_chat_template_for_generation(
     tokenizer,
     messages: list,
@@ -378,13 +430,21 @@ def apply_chat_template_for_generation(
     try:
         return _render(messages)
     except Exception:
-        # Strict tool templates reject the JSON-string ``arguments`` form via
+        # Retry with repairs applied cumulatively. Originals render first, so
-        # TypeError or a broad Jinja raise_exception, so retry with dicts coerced.
+        # working templates stay byte-identical.
-        # Original messages render first, so working templates stay byte-identical.
+        candidates: list = []
         normalized = _normalize_tool_call_arguments(messages)
-        if normalized is messages:
+        if normalized is not messages:
-            raise
+            candidates.append(normalized)
-        return _render(normalized)
+        split = _split_parallel_tool_calls(normalized)
+        if split is not normalized:
+            candidates.append(split)
+        for candidate in candidates:
+            try:
+                return _render(candidate)
+            except Exception:
+                continue
+        raise
 
 
 def render_native_template(

studio/backend/core/inference/inference.py

@@ -567,7 +567,7 @@ class InferenceBackend:
                         _meta_path = Path(config.path) / "export_metadata.json"
                         try:
                             if _meta_path.exists():
-                                _meta = json.loads(_meta_path.read_text())
+                                _meta = json.loads(_meta_path.read_text(encoding = "utf-8"))
                                 if _meta.get("base_model"):
                                     processor_source = _meta["base_model"]
                         except Exception:

studio/backend/core/inference/llama_admission.py

@@ -13,37 +13,85 @@ from __future__ import annotations
 
 import asyncio
 import os
+import sys
 import threading
 from collections import deque
 from dataclasses import dataclass
 from typing import Deque, Optional
 
 
-ADMISSION_CONTROL_ENV = "UNSLOTH_OPENAI_COMPAT_ADMISSION_CONTROL"
+# dataclass(slots = True) halves per-instance overhead. Measured as perf-neutral
-ADMISSION_QUEUE_TIMEOUT_ENV = "UNSLOTH_OPENAI_COMPAT_ADMISSION_QUEUE_TIMEOUT"
+# here, not a speed win: it costs a little on construction and gains it back on
-ADMISSION_KEEPALIVE_INTERVAL_ENV = "UNSLOTH_OPENAI_COMPAT_ADMISSION_KEEPALIVE_INTERVAL"
+# access. It is 3.10+ and this package declares >=3.9, so gate it rather than
-ADMISSION_MAX_QUEUE_ENV = "UNSLOTH_OPENAI_COMPAT_ADMISSION_MAX_QUEUE"
+# dropping it outright. Empty on 3.9 means a plain dataclass.
+_SLOTS = {"slots": True} if sys.version_info >= (3, 10) else {}
+
+
+ADMISSION_CONTROL_ENV = "UNSLOTH_LLAMA_ADMISSION_CONTROL"
+ADMISSION_QUEUE_TIMEOUT_ENV = "UNSLOTH_LLAMA_ADMISSION_QUEUE_TIMEOUT"
+ADMISSION_KEEPALIVE_INTERVAL_ENV = "UNSLOTH_LLAMA_ADMISSION_KEEPALIVE_INTERVAL"
+ADMISSION_MAX_QUEUE_ENV = "UNSLOTH_LLAMA_ADMISSION_MAX_QUEUE"
+ADMISSION_QUEUE_PER_SLOT_ENV = "UNSLOTH_LLAMA_ADMISSION_QUEUE_PER_SLOT"
+
+# The UNSLOTH_OPENAI_COMPAT_* spellings predate this queue being shared with the
+# Anthropic /v1/messages route (same llama-server slots). Still honored; the
+# neutral name above wins when both are set.
+_LEGACY_ENV = {
+    ADMISSION_CONTROL_ENV: "UNSLOTH_OPENAI_COMPAT_ADMISSION_CONTROL",
+    ADMISSION_QUEUE_TIMEOUT_ENV: "UNSLOTH_OPENAI_COMPAT_ADMISSION_QUEUE_TIMEOUT",
+    ADMISSION_KEEPALIVE_INTERVAL_ENV: "UNSLOTH_OPENAI_COMPAT_ADMISSION_KEEPALIVE_INTERVAL",
+    ADMISSION_MAX_QUEUE_ENV: "UNSLOTH_OPENAI_COMPAT_ADMISSION_MAX_QUEUE",
+}
 
 DEFAULT_ADMISSION_ENABLED = True
+# None: a queued request waits for its slot indefinitely rather than timing out.
 DEFAULT_ADMISSION_QUEUE_TIMEOUT_S = None
 DEFAULT_ADMISSION_KEEPALIVE_INTERVAL_S = 5.0
-DEFAULT_ADMISSION_MAX_QUEUE = 64
+# None: no absolute cap, the wait line is sized from the pool instead.
+DEFAULT_ADMISSION_MAX_QUEUE = None
+# Wait line = 16 x the serving slots, so it tracks --parallel (4 slots -> 64
+# waiters, 8 -> 128). Purely a memory guard; waiting itself is never timed out.
+DEFAULT_ADMISSION_QUEUE_PER_SLOT = 16
+# Floor for the scaled line, so a 1-slot backend (plain `unsloth studio`, or any
+# load downshifted to fit VRAM) keeps the depth it had before scaling existed
+# rather than dropping to 16 and rejecting callers that used to queue.
+DEFAULT_ADMISSION_MIN_QUEUE = 64
 
 
-@dataclass(frozen = True)
+@dataclass(frozen = True, **_SLOTS)
 class LlamaAdmissionConfig:
     enabled: bool = DEFAULT_ADMISSION_ENABLED
     queue_timeout_s: Optional[float] = DEFAULT_ADMISSION_QUEUE_TIMEOUT_S
     keepalive_interval_s: float = DEFAULT_ADMISSION_KEEPALIVE_INTERVAL_S
     max_queue: Optional[int] = DEFAULT_ADMISSION_MAX_QUEUE
+    queue_per_slot: Optional[int] = DEFAULT_ADMISSION_QUEUE_PER_SLOT
+    # Unconditional floor on the scaled line. The env path clears it when the
+    # operator sets QUEUE_PER_SLOT, so only the default multiplier is floored.
+    min_queue: Optional[int] = DEFAULT_ADMISSION_MIN_QUEUE
+
+    def queue_limit(self, capacity: int) -> Optional[int]:
+        """How many callers may line up for a pool of ``capacity`` slots.
+
+        An explicit ``max_queue`` wins; otherwise the line scales with the slots
+        so it follows ``--parallel``. The default multiplier is floored, so a
+        1-slot backend does not end up shallower than it was before scaling. None
+        (or any non-positive setting) means an unbounded line.
+        """
+        if self.max_queue is not None:
+            return self.max_queue if self.max_queue > 0 else None
+        if not self.queue_per_slot or self.queue_per_slot <= 0:
+            return None
+        scaled = self.queue_per_slot * max(1, capacity)
+        return max(self.min_queue, scaled) if self.min_queue else scaled
 
 
-@dataclass(frozen = True)
+@dataclass(frozen = True, **_SLOTS)
 class LlamaAdmissionSnapshot:
     key: str
     capacity: int
     active: int
     queued: int
+    free: int = 0
 
 
 class LlamaAdmissionError(Exception):
@@ -69,8 +117,17 @@ class LlamaAdmissionCancelled(LlamaAdmissionError):
     pass
 
 
-def _bool_env(name: str, default: bool) -> bool:
+def _raw_env(name: str) -> Optional[str]:
+    """Value for a canonical name, falling back to its legacy spelling."""
     value = os.environ.get(name)
+    if value is None or not value.strip():
+        legacy = _LEGACY_ENV.get(name)
+        value = os.environ.get(legacy) if legacy else None
+    return value
+
+
+def _bool_env(name: str, default: bool) -> bool:
+    value = _raw_env(name)
     if value is None or not value.strip():
         return default
     value = value.strip().lower()
@@ -82,7 +139,7 @@ def _bool_env(name: str, default: bool) -> bool:
 
 
 def _optional_positive_float_env(name: str, default: Optional[float]) -> Optional[float]:
-    value = os.environ.get(name)
+    value = _raw_env(name)
     if value is None or not value.strip():
         return default
     try:
@@ -93,7 +150,7 @@ def _optional_positive_float_env(name: str, default: Optional[float]) -> Optiona
 
 
 def _positive_float_env(name: str, default: float) -> float:
-    value = os.environ.get(name)
+    value = _raw_env(name)
     if value is None or not value.strip():
         return default
     try:
@@ -103,19 +160,38 @@ def _positive_float_env(name: str, default: float) -> float:
     return parsed if parsed > 0 else default
 
 
-def _optional_positive_int_env(name: str, default: Optional[int]) -> Optional[int]:
+def _queue_limits_from_env() -> tuple[Optional[int], Optional[int], Optional[int]]:
-    value = os.environ.get(name)
+    """(max_queue, queue_per_slot, min_queue) from the environment.
-    if value is None or not value.strip():
+
-        return default
+    An absolute MAX_QUEUE wins outright; MAX_QUEUE=0 asks for an unbounded line.
+    Unset leaves the per-slot multiplier in charge (itself 0 for unbounded). The
+    floor applies only to the default multiplier: setting QUEUE_PER_SLOT means
+    the operator wants that exact depth, however shallow.
+    """
+    # Explicit means it parsed, not just that something was set: a typo falls back
+    # to the default multiplier, so it has to keep the default's floor too.
+    raw_per_slot = _raw_env(ADMISSION_QUEUE_PER_SLOT_ENV)
     try:
-        parsed = int(value.strip())
+        per_slot = int((raw_per_slot or "").strip())
     except ValueError:
-        return default
+        per_slot, min_queue = DEFAULT_ADMISSION_QUEUE_PER_SLOT, DEFAULT_ADMISSION_MIN_QUEUE
-    return parsed if parsed > 0 else None
+    else:
+        per_slot, min_queue = (per_slot if per_slot > 0 else None), None
+    raw = _raw_env(ADMISSION_MAX_QUEUE_ENV)
+    if raw is None or not raw.strip():
+        return None, per_slot, min_queue
+    try:
+        parsed = int(raw.strip())
+    except ValueError:
+        return None, per_slot, min_queue
+    return (parsed, None, None) if parsed > 0 else (None, None, None)
 
 
 def llama_admission_config_from_env() -> LlamaAdmissionConfig:
+    max_queue, queue_per_slot, min_queue = _queue_limits_from_env()
     return LlamaAdmissionConfig(
+        queue_per_slot = queue_per_slot,
+        min_queue = min_queue,
         enabled = _bool_env(ADMISSION_CONTROL_ENV, DEFAULT_ADMISSION_ENABLED),
         queue_timeout_s = _optional_positive_float_env(
             ADMISSION_QUEUE_TIMEOUT_ENV,
@@ -125,14 +201,11 @@ def llama_admission_config_from_env() -> LlamaAdmissionConfig:
             ADMISSION_KEEPALIVE_INTERVAL_ENV,
             DEFAULT_ADMISSION_KEEPALIVE_INTERVAL_S,
         ),
-        max_queue = _optional_positive_int_env(
+        max_queue = max_queue,
-            ADMISSION_MAX_QUEUE_ENV,
-            DEFAULT_ADMISSION_MAX_QUEUE,
-        ),
     )
 
 
-@dataclass
+@dataclass(**_SLOTS)
 class _Waiter:
     loop: asyncio.AbstractEventLoop
     future: asyncio.Future
@@ -141,11 +214,23 @@ class _Waiter:
 
 
 class LlamaAdmissionLease:
-    def __init__(self, queue: Optional["LlamaAdmissionQueue"]):
+    __slots__ = ("_queue", "_slot", "_released", "_release_lock")
+
+    def __init__(
+        self,
+        queue: Optional["LlamaAdmissionQueue"],
+        slot: Optional[int] = None,
+    ):
         self._queue = queue
+        self._slot = slot
         self._released = False
         self._release_lock = threading.Lock()
 
+    @property
+    def slot(self) -> Optional[int]:
+        """Pool slot this lease holds, or None when admission is disabled."""
+        return self._slot
+
     def release(self) -> None:
         queue = None
         with self._release_lock:
@@ -154,7 +239,7 @@ class LlamaAdmissionLease:
             self._released = True
             queue = self._queue
         if queue is not None:
-            queue.release()
+            queue.release(self._slot)
 
     async def __aenter__(self) -> "LlamaAdmissionLease":
         return self
@@ -164,6 +249,8 @@ class LlamaAdmissionLease:
 
 
 class LlamaAdmissionReservation:
+    __slots__ = ("_queue", "_lease", "_waiter", "snapshot")
+
     def __init__(
         self,
         *,
@@ -195,6 +282,13 @@ class LlamaAdmissionReservation:
         return self._lease
 
     async def wait(self, timeout_s: float) -> Optional[LlamaAdmissionLease]:
+        """Wait up to ``timeout_s`` for a slot.
+
+        A timeout leaves this reservation queued so the caller can poll again.
+        Any exit that abandons the wait for good must call ``cancel()``, or the
+        slot granted later is delivered to a future nobody reads and is never
+        released.
+        """
         lease = self.lease_nowait()
         if lease is not None:
             return lease
@@ -229,35 +323,80 @@ class LlamaAdmissionReservation:
 
 
 class LlamaAdmissionQueue:
+    """A fixed pool of generation slots for one llama-server, plus a FIFO wait line.
+
+    The pool mirrors llama-server's own ``--parallel`` slots: ``capacity`` slot ids
+    are each either free or held by exactly one caller. A caller that finds every
+    slot busy waits in arrival order and is handed the next slot to free, so no
+    caller is starved. This bounds only the callers that reserve: chat completions
+    and messages do, while /v1/completions, Studio's own chat endpoint and RAG
+    captioning all reach llama-server directly, so it is not a global cap.
+    Waiting is unbounded in time by default (``queue_timeout_s``
+    None); the wait line itself is bounded, and only how many may line up before
+    new arrivals are rejected. By default that is ``16 x slots`` floored at 64,
+    not unlimited: an unbounded line takes ``max_queue`` or ``queue_per_slot``
+    set to 0. See ``LlamaAdmissionConfig.queue_limit``.
+    """
+
+    __slots__ = ("key", "_lock", "_capacity", "_free", "_in_use", "_held", "_waiters")
+
     def __init__(self, key: str):
         self.key = key
         self._lock = threading.Lock()
-        self._active = 0
         self._capacity = 1
+        self._free: list[int] = [0]
+        # Held slots as a bitmask: one int instead of a set, so the pool costs the
+        # same whether it is idle or saturated. _held is its popcount, kept as a
+        # counter because int.bit_count() is 3.10+ and this package targets 3.9.
+        self._in_use = 0
+        self._held = 0
         self._waiters: Deque[_Waiter] = deque()
 
+    def _resize_pool_locked(self, capacity: int) -> None:
+        # Slots past a shrunk capacity retire when their holder releases them.
+        if capacity == self._capacity:
+            return
+        self._capacity = capacity
+        self._free = [slot for slot in range(capacity) if not self._in_use >> slot & 1]
+
+    def _can_admit_locked(self) -> bool:
+        # Slots still held above a shrunk capacity keep occupying the backend, so
+        # count every held slot against the ceiling, not just the ids below it.
+        return bool(self._free) and self._held < self._capacity
+
+    def _take_slot_locked(self) -> Optional[int]:
+        if not self._can_admit_locked():
+            return None
+        slot = self._free.pop()
+        self._in_use |= 1 << slot
+        self._held += 1
+        return slot
+
     def reserve(self, *, capacity: int, config: LlamaAdmissionConfig) -> LlamaAdmissionReservation:
         capacity = max(1, int(capacity or 1))
         if not config.enabled:
             return LlamaAdmissionReservation(
                 queue = None,
                 lease = LlamaAdmissionLease(None),
-                snapshot = LlamaAdmissionSnapshot(self.key, capacity, 0, 0),
+                snapshot = LlamaAdmissionSnapshot(self.key, capacity, 0, 0, capacity),
             )
 
         loop = asyncio.get_running_loop()
         with self._lock:
-            self._capacity = capacity
+            self._resize_pool_locked(capacity)
-            self._prune_waiters_locked()
             self._grant_waiters_locked()
-            if self._active < self._capacity and not self._waiters:
+            if not self._waiters:
-                self._active += 1
+                slot = self._take_slot_locked()
-                return LlamaAdmissionReservation(
+                if slot is not None:
-                    queue = self,
+                    # No snapshot here: callers read it through snapshot_now(),
-                    lease = LlamaAdmissionLease(self),
+                    # which re-reads the queue, so building one per admitted
-                    snapshot = self._snapshot_locked(),
+                    # request would be pure allocation on the hot path.
-                )
+                    return LlamaAdmissionReservation(
-            if config.max_queue is not None and len(self._waiters) >= config.max_queue:
+                        queue = self,
+                        lease = LlamaAdmissionLease(self, slot),
+                    )
+            limit = config.queue_limit(self._capacity)
+            if limit is not None and self._live_waiters_locked() >= limit:
                 raise LlamaAdmissionQueueFull(
                     "llama-server generation queue is full",
                     snapshot = self._snapshot_locked(),
@@ -270,13 +409,20 @@ class LlamaAdmissionQueue:
             return LlamaAdmissionReservation(
                 queue = self,
                 waiter = waiter,
-                snapshot = self._snapshot_locked(),
             )
 
-    def release(self) -> None:
+    def _release_slot_locked(self, slot: Optional[int]) -> None:
+        # A slot id at or past a shrunk capacity retires instead of returning.
+        if slot is None or not self._in_use >> slot & 1:
+            return
+        self._in_use &= ~(1 << slot)
+        self._held -= 1
+        if slot < self._capacity:
+            self._free.append(slot)
+
+    def release(self, slot: Optional[int]) -> None:
         with self._lock:
-            if self._active > 0:
+            self._release_slot_locked(slot)
-                self._active -= 1
             self._grant_waiters_locked()
 
     def cancel(self, waiter: _Waiter) -> None:
@@ -291,7 +437,13 @@ class LlamaAdmissionQueue:
                 lease_to_release = waiter.granted_lease
                 waiter.granted_lease = None
             if not waiter.future.done():
-                waiter.loop.call_soon_threadsafe(waiter.future.cancel)
+                try:
+                    waiter.loop.call_soon_threadsafe(waiter.future.cancel)
+                except RuntimeError:
+                    # Loop gone. Routes call cancel() from finally blocks, so
+                    # raising here would both mask their exception and skip the
+                    # release below, stranding the slot for the process lifetime.
+                    pass
         if lease_to_release is not None:
             lease_to_release.release()
 
@@ -303,20 +455,30 @@ class LlamaAdmissionQueue:
     def is_idle(self) -> bool:
         with self._lock:
             self._prune_waiters_locked()
-            return self._active == 0 and not self._waiters
+            return self._in_use == 0 and not self._waiters
 
     def _grant_waiters_locked(self) -> None:
-        self._prune_waiters_locked()
+        # Dead waiters are skipped as they are popped, so no prune is needed here.
-        while self._waiters and self._active < self._capacity:
+        while self._waiters and self._can_admit_locked():
             waiter = self._waiters.popleft()
             if waiter.cancelled or waiter.future.done():
                 continue
-            self._active += 1
+            slot = self._take_slot_locked()
-            lease = LlamaAdmissionLease(self)
+            lease = LlamaAdmissionLease(self, slot)
             waiter.granted_lease = lease
-            waiter.loop.call_soon_threadsafe(self._deliver_lease, waiter, lease)
+            try:
+                waiter.loop.call_soon_threadsafe(self._deliver_lease, waiter, lease)
+            except RuntimeError:
+                # Waiter's loop is gone. Reclaim the slot; leaving the bit set
+                # would strand it, since _free is rebuilt from the bitmask.
+                waiter.granted_lease = None
+                self._release_slot_locked(slot)
 
     def _deliver_lease(self, waiter: _Waiter, lease: LlamaAdmissionLease) -> None:
+        # Runs on the waiter's own loop thread, which is also the only thread that
+        # cancels that reservation, so waiter state is safe to touch unlocked here.
+        # release() may be called from any thread, but only reaches this via
+        # call_soon_threadsafe. Cancelling off-loop would need this under _lock.
         if waiter.cancelled or waiter.future.done():
             waiter.granted_lease = None
             if not waiter.future.done():
@@ -331,16 +493,32 @@ class LlamaAdmissionQueue:
             lease.release()
 
     def _prune_waiters_locked(self) -> None:
+        # Rebuilding the deque on every reserve/release dominated the hot path, so
+        # only pay it when a waiter actually died out of band (an externally
+        # cancelled future); cancel() already drops its own waiter eagerly.
+        for waiter in self._waiters:
+            if waiter.cancelled or waiter.future.done():
+                break
+        else:
+            return
         self._waiters = deque(
             waiter for waiter in self._waiters if not waiter.cancelled and not waiter.future.done()
         )
 
+    def _live_waiters_locked(self) -> int:
+        self._prune_waiters_locked()
+        return len(self._waiters)
+
     def _snapshot_locked(self) -> LlamaAdmissionSnapshot:
         return LlamaAdmissionSnapshot(
             key = self.key,
             capacity = self._capacity,
-            active = self._active,
+            active = self._held,
             queued = len(self._waiters),
+            # What another caller could actually take, so the admission log never
+            # shows free slots next to queued requests: after a shrink, ids below
+            # the new capacity can be free while holdovers still fill the ceiling.
+            free = min(len(self._free), max(0, self._capacity - self._held)),
         )