removed flash attention

Previously HF_HUB_CACHE always defaulted to ~/.cache/huggingface/hub
even when HF_HOME was explicitly set (e.g. in Docker). This caused
models to download to the wrong location instead of the configured
HF_HOME path.

.worktreeinclude

@@ -0,0 +1 @@
+CLAUDE.md

studio/backend/core/training/worker.py

@@ -64,6 +64,97 @@ def _model_wants_causal_conv1d(model_name: str) -> bool:
     )
 
 
+def _causal_conv1d_platform_tag() -> str | None:
+    machine = platform.machine().lower()
+    if sys.platform.startswith("linux"):
+        if machine in {"x86_64", "amd64"}:
+            return "linux_x86_64"
+        if machine in {"aarch64", "arm64"}:
+            return "linux_aarch64"
+        return None
+    # No prebuilt wheels published for macOS or Windows
+    return None
+
+
+def _probe_causal_conv1d_env() -> dict[str, str] | None:
+    try:
+        probe = _sp.run(
+            [
+                sys.executable,
+                "-c",
+                (
+                    "import json, sys, re, torch; "
+                    "parts = torch.__version__.split('+', 1)[0].split('.')[:2]; "
+                    "minor = re.sub(r'[^0-9].*', '', parts[1]) if len(parts) > 1 else '0'; "
+                    "torch_mm = parts[0] + '.' + minor; "
+                    "print(json.dumps({"
+                    "'python_tag': f'cp{sys.version_info.major}{sys.version_info.minor}', "
+                    "'torch_mm': torch_mm, "
+                    "'cuda_major': str(int(str(torch.version.cuda).split('.', 1)[0])) if torch.version.cuda else '', "
+                    "'cxx11abi': str(torch._C._GLIBCXX_USE_CXX11_ABI).upper()"
+                    "}))"
+                ),
+            ],
+            stdout = _sp.PIPE,
+            stderr = _sp.PIPE,
+            text = True,
+            timeout = 30,
+        )
+    except _sp.TimeoutExpired:
+        logger.warning("Torch environment probe timed out after 30s")
+        return None
+    if probe.returncode != 0:
+        logger.warning(
+            "Failed to probe torch environment for causal-conv1d wheel:\n%s",
+            probe.stdout,
+        )
+        return None
+
+    try:
+        return json.loads(probe.stdout.strip())
+    except json.JSONDecodeError:
+        logger.warning(
+            "Failed to parse torch environment probe output: %s", probe.stdout
+        )
+        return None
+
+
+def _direct_wheel_url(
+    *,
+    filename_prefix: str,
+    package_version: str,
+    release_tag: str,
+    release_base_url: str,
+    env: dict[str, str] | None = None,
+) -> str | None:
+    env = env or _probe_causal_conv1d_env()
+    platform_tag = _causal_conv1d_platform_tag()
+    if env is None or platform_tag is None or not env.get("cuda_major"):
+        return None
+
+    filename = (
+        f"{filename_prefix}-{package_version}"
+        f"+cu{env['cuda_major']}torch{env['torch_mm']}"
+        f"cxx11abi{env['cxx11abi']}-{env['python_tag']}-{env['python_tag']}-{platform_tag}.whl"
+    )
+    return f"{release_base_url}/{release_tag}/{filename}"
+
+
+def _url_exists(url: str) -> bool:
+    try:
+        request = urllib.request.Request(url, method = "HEAD")
+        with urllib.request.urlopen(request, timeout = 10):
+            return True
+    except urllib.error.HTTPError as exc:
+        if exc.code == 404:
+            return False
+        logger.warning("Unexpected HTTP error while probing %s: %s", url, exc)
+        return False
+    except Exception as exc:
+        logger.warning("Failed to probe %s: %s", url, exc)
+        return False
+
+
 def _install_package_wheel_first(
     *,
     event_queue: Any,
@@ -397,10 +488,10 @@ def run_training_process(
     try:
         _ensure_causal_conv1d_fast_path(event_queue, model_name)
         _ensure_mamba_ssm(event_queue, model_name)
-        _ensure_flash_attn_for_long_context(
-            event_queue,
-            int(config.get("max_seq_length", 2048)),
-        )
+        #_ensure_flash_attn_for_long_context(
+        #    event_queue,
+        #    int(config.get("max_seq_length", 2048)),
+        #)
     except Exception as exc:
         event_queue.put(
             {

studio/backend/utils/paths/storage_roots.py

@@ -193,10 +193,11 @@ def _setup_cache_env() -> None:
         os.environ.get("XDG_CACHE_HOME", Path.home() / ".cache")
     ).expanduser()
     hf_default = xdg_cache / "huggingface"
+    hf_home = Path(os.environ.get("HF_HOME", str(hf_default)))
     defaults: dict[str, str] = {
         "HF_HOME": str(hf_default),
-        "HF_HUB_CACHE": str(hf_default / "hub"),
-        "HF_XET_CACHE": str(hf_default / "xet"),
+        "HF_HUB_CACHE": str(hf_home / "hub"),
+        "HF_XET_CACHE": str(hf_home / "xet"),
         "UV_CACHE_DIR": str(root / "uv"),
         "VLLM_CACHE_ROOT": str(root / "vllm"),
     }

studio/install_python_stack.py

@@ -841,220 +841,157 @@ def install_python_stack() -> int:
         base_total += 3
     _TOTAL = (base_total - 1) if skip_base else base_total
 
-    # 1. Try to use uv for faster installs (must happen before pip upgrade
-    #    because uv venvs don't include pip by default)
-    USE_UV = _bootstrap_uv()
+    # # 1. Try to use uv for faster installs (must happen before pip upgrade
+    # #    because uv venvs don't include pip by default)
+    # USE_UV = _bootstrap_uv()
 
-    # 2. Ensure pip is available (uv venvs created by install.sh don't include pip)
-    _progress("pip bootstrap")
-    if USE_UV:
-        run(
-            "Bootstrapping pip via uv",
-            [
-                "uv",
-                "pip",
-                "install",
-                "--python",
-                sys.executable,
-                "pip",
-            ],
-        )
-    else:
-        # pip may not exist yet (uv-created venvs omit it). Try ensurepip
-        # first, then upgrade. Only fall back to a direct upgrade when pip
-        # is already present.
-        _has_pip = (
-            subprocess.run(
-                [sys.executable, "-m", "pip", "--version"],
-                stdout = subprocess.DEVNULL,
-                stderr = subprocess.DEVNULL,
-            ).returncode
-            == 0
-        )
+    # # 2. Ensure pip is available (uv venvs created by install.sh don't include pip)
+    # _progress("pip bootstrap")
+    # if USE_UV:
+    #     run(
+    #         "Bootstrapping pip via uv",
+    #         [
+    #             "uv",
+    #             "pip",
+    #             "install",
+    #             "--python",
+    #             sys.executable,
+    #             "pip",
+    #         ],
+    #     )
+    # else:
+    #     # pip may not exist yet (uv-created venvs omit it). Try ensurepip
+    #     # first, then upgrade. Only fall back to a direct upgrade when pip
+    #     # is already present.
+    #     _has_pip = (
+    #         subprocess.run(
+    #             [sys.executable, "-m", "pip", "--version"],
+    #             stdout = subprocess.DEVNULL,
+    #             stderr = subprocess.DEVNULL,
+    #         ).returncode
+    #         == 0
+    #     )
+    #
+    #     if not _has_pip:
+    #         run(
+    #             "Bootstrapping pip via ensurepip",
+    #             [sys.executable, "-m", "ensurepip", "--upgrade"],
+    #         )
+    #     else:
+    #         run(
+    #             "Upgrading pip",
+    #             [sys.executable, "-m", "pip", "install", "--upgrade", "pip"],
+    #         )
 
-        if not _has_pip:
-            run(
-                "Bootstrapping pip via ensurepip",
-                [sys.executable, "-m", "ensurepip", "--upgrade"],
-            )
-        else:
-            run(
-                "Upgrading pip",
-                [sys.executable, "-m", "pip", "install", "--upgrade", "pip"],
-            )
+    # # 3. Core packages: unsloth-zoo + unsloth (or custom package name)
+    # if skip_base:
+    #     print(_green(f"✅ {package_name} already installed — skipping base packages"))
+    # elif NO_TORCH:
+    #     # No-torch update path: install unsloth + unsloth-zoo with --no-deps
+    #     # (current PyPI metadata still declares torch as a hard dep), then
+    #     # runtime deps with --no-deps (avoids transitive torch).
+    #     _progress("base packages (no torch)")
+    #     pip_install(
+    #         f"Updating {package_name} + unsloth-zoo (no-torch mode)",
+    #         "--no-cache-dir",
+    #         "--no-deps",
+    #         "--upgrade-package",
+    #         package_name,
+    #         "--upgrade-package",
+    #         "unsloth-zoo",
+    #         package_name,
+    #         "unsloth-zoo",
+    #     )
+    #     pip_install(
+    #         "Installing no-torch runtime deps",
+    #         "--no-cache-dir",
+    #         "--no-deps",
+    #         req = REQ_ROOT / "no-torch-runtime.txt",
+    #     )
+    #     if local_repo:
+    #         pip_install(
+    #             "Overlaying local repo (editable)",
+    #             "--no-cache-dir",
+    #             "--no-deps",
+    #             "-e",
+    #             local_repo,
+    #             constrain = False,
+    #         )
+    # elif local_repo:
+    #     _progress("base packages")
+    #     pip_install(
+    #         "Updating base packages",
+    #         "--no-cache-dir",
+    #         "--upgrade-package",
+    #         "unsloth",
+    #         "--upgrade-package",
+    #         "unsloth-zoo",
+    #         req = REQ_ROOT / "base.txt",
+    #     )
+    #     pip_install(
+    #         "Overlaying local repo (editable)",
+    #         "--no-cache-dir",
+    #         "--no-deps",
+    #         "-e",
+    #         local_repo,
+    #         constrain = False,
+    #     )
+    # elif package_name != "unsloth":
+    #     _progress("base packages")
+    #     pip_install(
+    #         f"Installing {package_name}",
+    #         "--no-cache-dir",
+    #         package_name,
+    #     )
+    # else:
+    #     _progress("base packages")
+    #     pip_install(
+    #         "Updating base packages",
+    #         "--no-cache-dir",
+    #         "--upgrade-package",
+    #         "unsloth",
+    #         "--upgrade-package",
+    #         "unsloth-zoo",
+    #         req = REQ_ROOT / "base.txt",
+    #     )
 
-    # 3. Core packages: unsloth-zoo + unsloth (or custom package name)
-    if skip_base:
-        pass
-    elif NO_TORCH:
-        # No-torch update path: install unsloth + unsloth-zoo with --no-deps
-        # (current PyPI metadata still declares torch as a hard dep), then
-        # runtime deps with --no-deps (avoids transitive torch).
-        _progress("base packages (no torch)")
-        pip_install(
-            f"Updating {package_name} + unsloth-zoo (no-torch mode)",
-            "--no-cache-dir",
-            "--no-deps",
-            "--upgrade-package",
-            package_name,
-            "--upgrade-package",
-            "unsloth-zoo",
-            package_name,
-            "unsloth-zoo",
-        )
-        pip_install(
-            "Installing no-torch runtime deps",
-            "--no-cache-dir",
-            "--no-deps",
-            req = REQ_ROOT / "no-torch-runtime.txt",
-        )
-        if local_repo:
-            pip_install(
-                "Overlaying local repo (editable)",
-                "--no-cache-dir",
-                "--no-deps",
-                "-e",
-                local_repo,
-                constrain = False,
-            )
-    elif local_repo:
-        # Local dev install: update deps from base.txt, then overlay the
-        # local checkout as an editable install (--no-deps so torch is
-        # never re-resolved).
-        _progress("base packages")
-        pip_install(
-            "Updating base packages",
-            "--no-cache-dir",
-            "--upgrade-package",
-            "unsloth",
-            "--upgrade-package",
-            "unsloth-zoo",
-            req = REQ_ROOT / "base.txt",
-        )
-        pip_install(
-            "Overlaying local repo (editable)",
-            "--no-cache-dir",
-            "--no-deps",
-            "-e",
-            local_repo,
-            constrain = False,
-        )
-    elif package_name != "unsloth":
-        # Custom package name (e.g. roland-sloth for testing) — install directly
-        _progress("base packages")
-        pip_install(
-            f"Installing {package_name}",
-            "--no-cache-dir",
-            package_name,
-        )
-    else:
-        # Update path: upgrade only unsloth + unsloth-zoo while preserving
-        # existing torch/CUDA installations.  Torch is pre-installed by
-        # install.sh / setup.ps1; --upgrade-package targets only base pkgs.
-        _progress("base packages")
-        pip_install(
-            "Updating base packages",
-            "--no-cache-dir",
-            "--upgrade-package",
-            "unsloth",
-            "--upgrade-package",
-            "unsloth-zoo",
-            req = REQ_ROOT / "base.txt",
-        )
+    # pip_install(
+    #     "Installing additional unsloth dependencies",
+    #     "--no-cache-dir",
+    #     req = REQ_ROOT / "extras.txt",
+    # )
 
-    # 2b. AMD ROCm: reinstall torch with HIP wheels if the host has ROCm but the
-    #     venv received CPU-only torch (common when pip resolves torch from PyPI).
-    #     Must come immediately after base packages so torch is present for inspection.
-    if not IS_WINDOWS and not IS_MACOS and not NO_TORCH:
-        _progress("ROCm torch check")
-        _ensure_rocm_torch()
+    # pip_install(
+    #     "Installing extras (no-deps)",
+    #     "--no-deps",
+    #     "--no-cache-dir",
+    #     req = REQ_ROOT / "extras-no-deps.txt",
+    # )
 
-    # Windows + AMD GPU: PyTorch does not publish ROCm wheels for Windows.
-    # Detect and warn so users know manual steps are needed for GPU training.
-    if IS_WINDOWS and not NO_TORCH and not _has_usable_nvidia_gpu():
-        # Validate actual AMD GPU presence (not just tool existence)
-        import re as _re_win
+    # # 4. Overrides (torchao, transformers) -- force-reinstall
+    # _progress("dependency overrides")
+    # pip_install(
+    #     "Installing dependency overrides",
+    #     "--force-reinstall",
+    #     "--no-cache-dir",
+    #     req = REQ_ROOT / "overrides.txt",
+    # )
 
-        def _win_amd_smi_has_gpu(stdout: str) -> bool:
-            return bool(_re_win.search(r"(?im)^gpu\s*[:\[]\s*\d", stdout))
+    # # 5. Triton kernels (no-deps, from source)
+    # #    Skip on Windows (no support) and macOS (no support).
+    # if not IS_WINDOWS and not IS_MACOS:
+    #     _progress("triton kernels")
+    #     pip_install(
+    #         "Installing triton kernels",
+    #         "--no-deps",
+    #         "--no-cache-dir",
+    #         req = REQ_ROOT / "triton-kernels.txt",
+    #         constrain = False,
+    #     )
 
-        _win_amd_gpu = False
-        for _wcmd, _check_fn in (
-            (["hipinfo"], lambda out: "gcnarchname" in out.lower()),
-            (["amd-smi", "list"], _win_amd_smi_has_gpu),
-        ):
-            _wexe = shutil.which(_wcmd[0])
-            if not _wexe:
-                continue
-            try:
-                _wr = subprocess.run(
-                    [_wexe, *_wcmd[1:]],
-                    stdout = subprocess.PIPE,
-                    stderr = subprocess.DEVNULL,
-                    text = True,
-                    timeout = 10,
-                )
-            except Exception:
-                continue
-            if _wr.returncode == 0 and _check_fn(_wr.stdout):
-                _win_amd_gpu = True
-                break
-        if _win_amd_gpu:
-            _safe_print(
-                _dim("  Note:"),
-                "AMD GPU detected on Windows. ROCm-enabled PyTorch must be",
-            )
-            _safe_print(
-                " " * 8,
-                "installed manually. See: https://docs.unsloth.ai/get-started/install-and-update/amd",
-            )
-
-    # 3. Extra dependencies
-    _progress("unsloth extras")
-    pip_install(
-        "Installing additional unsloth dependencies",
-        "--no-cache-dir",
-        req = REQ_ROOT / "extras.txt",
-    )
-
-    # 3b. Extra dependencies (no-deps) -- audio model support etc.
-    _progress("extra codecs")
-    pip_install(
-        "Installing extras (no-deps)",
-        "--no-deps",
-        "--no-cache-dir",
-        req = REQ_ROOT / "extras-no-deps.txt",
-    )
-
-    # 4. Overrides (torchao, transformers) -- force-reinstall
-    #    Skip entirely when torch is unavailable (e.g. Intel Mac GGUF-only mode)
-    #    because overrides.txt contains torchao which requires torch.
-    if NO_TORCH:
-        _progress("dependency overrides (skipped, no torch)")
-    else:
-        _progress("dependency overrides")
-        pip_install(
-            "Installing dependency overrides",
-            "--force-reinstall",
-            "--no-cache-dir",
-            req = REQ_ROOT / "overrides.txt",
-        )
-
-    # 5. Triton kernels (no-deps, from source)
-    #    Skip on Windows (no support) and macOS (no support).
-    if not IS_WINDOWS and not IS_MACOS:
-        _progress("triton kernels")
-        pip_install(
-            "Installing triton kernels",
-            "--no-deps",
-            "--no-cache-dir",
-            req = REQ_ROOT / "triton-kernels.txt",
-            constrain = False,
-        )
-
-    if not IS_WINDOWS and not IS_MACOS and not NO_TORCH:
-        _progress("flash-attn")
-        _ensure_flash_attn()
+    #if not IS_WINDOWS and not IS_MACOS and not NO_TORCH:
+    #    _progress("flash-attn")
+    #    _ensure_flash_attn()
 
     # # 6. Patch: override llama_cpp.py with fix from unsloth-zoo  feature/llama-cpp-windows-support branch
     # patch_package_file(

studio/setup.ps1

@@ -1581,6 +1581,89 @@ if ($stackExit -ne 0) {
     exit 1
 }
 
+# ── Pre-install transformers 5.x into .venv_t5_530/ and .venv_t5_550/ ──
+# Models like GLM-4.7-Flash, Qwen3 MoE need transformers>=5.3.0.
+# Gemma 4 models need transformers>=5.5.0.
+# Pre-install into separate directories to avoid runtime pip overhead.
+# The training subprocess prepends the appropriate dir to sys.path.
+Write-Host ""
+
+# Clean up legacy single .venv_t5 directory
+$VenvT5Legacy = Join-Path $env:USERPROFILE ".unsloth\studio\.venv_t5"
+if (Test-Path $VenvT5Legacy) { Remove-Item -Recurse -Force $VenvT5Legacy }
+
+$prevEAP_t5 = $ErrorActionPreference
+$ErrorActionPreference = "Continue"
+
+# --- .venv_t5_530 (transformers 5.3.0) ---
+substep "pre-installing transformers 5.3.0 for newer model support..."
+$VenvT5_530Dir = Join-Path $env:USERPROFILE ".unsloth\studio\.venv_t5_530"
+if (Test-Path $VenvT5_530Dir) { Remove-Item -Recurse -Force $VenvT5_530Dir }
+New-Item -ItemType Directory -Path $VenvT5_530Dir -Force | Out-Null
+foreach ($pkg in @("transformers==5.3.0", "huggingface_hub==1.8.0", "hf_xet==1.4.2")) {
+    if ($script:UnslothVerbose) {
+        Fast-Install --target $VenvT5_530Dir --no-deps $pkg
+        $t5PkgExit = $LASTEXITCODE
+        $output = ""
+    } else {
+        $output = Fast-Install --target $VenvT5_530Dir --no-deps $pkg | Out-String
+        $t5PkgExit = $LASTEXITCODE
+    }
+    if ($t5PkgExit -ne 0) {
+        Write-Host "[FAIL] Could not install $pkg into .venv_t5_530/" -ForegroundColor Red
+        Write-Host $output -ForegroundColor Red
+        $ErrorActionPreference = $prevEAP_t5
+        exit 1
+    }
+}
+if ($script:UnslothVerbose) {
+    Fast-Install --target $VenvT5_530Dir tiktoken
+    $tiktokenInstallExit = $LASTEXITCODE
+    $output = ""
+} else {
+    $output = Fast-Install --target $VenvT5_530Dir tiktoken | Out-String
+    $tiktokenInstallExit = $LASTEXITCODE
+}
+if ($tiktokenInstallExit -ne 0) {
+    substep "Could not install tiktoken into .venv_t5_530/ -- Qwen tokenizers may fail" "Yellow"
+}
+step "transformers" "5.3.0 pre-installed"
+
+# --- .venv_t5_550 (transformers 5.5.0) ---
+substep "pre-installing transformers 5.5.0 for Gemma 4 support..."
+$VenvT5_550Dir = Join-Path $env:USERPROFILE ".unsloth\studio\.venv_t5_550"
+if (Test-Path $VenvT5_550Dir) { Remove-Item -Recurse -Force $VenvT5_550Dir }
+New-Item -ItemType Directory -Path $VenvT5_550Dir -Force | Out-Null
+foreach ($pkg in @("transformers==5.5.0", "huggingface_hub==1.8.0", "hf_xet==1.4.2")) {
+    if ($script:UnslothVerbose) {
+        Fast-Install --target $VenvT5_550Dir --no-deps $pkg
+        $t5PkgExit = $LASTEXITCODE
+        $output = ""
+    } else {
+        $output = Fast-Install --target $VenvT5_550Dir --no-deps $pkg | Out-String
+        $t5PkgExit = $LASTEXITCODE
+    }
+    if ($t5PkgExit -ne 0) {
+        Write-Host "[FAIL] Could not install $pkg into .venv_t5_550/" -ForegroundColor Red
+        Write-Host $output -ForegroundColor Red
+        $ErrorActionPreference = $prevEAP_t5
+        exit 1
+    }
+}
+if ($script:UnslothVerbose) {
+    Fast-Install --target $VenvT5_550Dir tiktoken
+    $tiktokenInstallExit = $LASTEXITCODE
+    $output = ""
+} else {
+    $output = Fast-Install --target $VenvT5_550Dir tiktoken | Out-String
+    $tiktokenInstallExit = $LASTEXITCODE
+}
+if ($tiktokenInstallExit -ne 0) {
+    substep "Could not install tiktoken into .venv_t5_550/ -- Qwen tokenizers may fail" "Yellow"
+}
+$ErrorActionPreference = $prevEAP_t5
+step "transformers" "5.5.0 pre-installed"
+
 } else {
     step "python" "dependencies up to date"
     # Restore ErrorActionPreference (was lowered for pip/python section)

studio/setup.sh

@@ -392,7 +392,7 @@ if [ -d "$SCRIPT_DIR/backend/core/data_recipe/oxc-validator" ] && command -v npm
 fi
 
 # ── Python venv + deps ──
-STUDIO_HOME="$HOME/.unsloth/studio"
+STUDIO_HOME="${UNSLOTH_STUDIO_HOME:-$HOME/.unsloth/studio}"
 VENV_DIR="$STUDIO_HOME/unsloth_studio"
 VENV_T5_530_DIR="$STUDIO_HOME/.venv_t5_530"
 VENV_T5_550_DIR="$STUDIO_HOME/.venv_t5_550"
@@ -405,7 +405,12 @@ VENV_T5_550_DIR="$STUDIO_HOME/.venv_t5_550"
 # Note: do NOT delete $STUDIO_HOME/.venv here — install.sh handles migration
 
 _COLAB_NO_VENV=false
-if [ ! -x "$VENV_DIR/bin/python" ]; then
+_DOCKER_NO_VENV=false
+if [ -n "$UNSLOTH_DOCKER" ]; then
+    # Docker: packages already in /opt/conda — skip venv entirely.
+    # Only pre-install .venv_t5 for transformers 5.x switching (handled below).
+    _DOCKER_NO_VENV=true
+elif [ ! -x "$VENV_DIR/bin/python" ]; then
     if [ "$IS_COLAB" = true ]; then
         # On Colab there is no Studio venv -- install backend deps into system Python.
         # Strip all version constraints so pip keeps Colab's pre-installed
@@ -470,6 +475,52 @@ if [ "$_COLAB_NO_VENV" = true ]; then
     substep "continuing to llama.cpp install for GGUF inference support"
 fi
 
+# In Docker, packages are pre-installed in /opt/conda — only install missing
+# studio/data-designer deps and pre-install .venv_t5 for transformers 5.x.
+if [ "$_DOCKER_NO_VENV" = true ]; then
+    echo "   Docker detected — skipping venv activation."
+
+    # Install branch's unsloth/unsloth_cli/studio into /opt/conda
+    # (overwrites PyPI version with Docker-aware code)
+    echo "   Installing local unsloth from branch..."
+    pip install --force-reinstall --no-deps "$REPO_ROOT"
+
+    # Install only missing deps (studio, data-designer, plugin, metadata patch).
+    # Heavy packages (torch, unsloth, vllm, etc.) are already in /opt/conda.
+    # install_python_stack.py has steps 1-5 commented out for this branch.
+    python "$SCRIPT_DIR/install_python_stack.py"
+
+    # Pre-install transformers 5.x into .venv_t5_530/ and .venv_t5_550/
+    echo ""
+    echo "   Pre-installing transformers 5.x for newer model support..."
+
+    # Clean up legacy single .venv_t5 directory
+    [ -d "$STUDIO_HOME/.venv_t5" ] && rm -rf "$STUDIO_HOME/.venv_t5"
+
+    [ -d "$VENV_T5_530_DIR" ] && rm -rf "$VENV_T5_530_DIR"
+    mkdir -p "$VENV_T5_530_DIR"
+    pip install --target "$VENV_T5_530_DIR" --no-deps "transformers==5.3.0" 2>/dev/null
+    pip install --target "$VENV_T5_530_DIR" --no-deps "huggingface_hub==1.8.0" 2>/dev/null
+    pip install --target "$VENV_T5_530_DIR" --no-deps "hf_xet==1.4.2" 2>/dev/null
+    pip install --target "$VENV_T5_530_DIR" "tiktoken" 2>/dev/null
+
+    [ -d "$VENV_T5_550_DIR" ] && rm -rf "$VENV_T5_550_DIR"
+    mkdir -p "$VENV_T5_550_DIR"
+    pip install --target "$VENV_T5_550_DIR" --no-deps "transformers==5.5.0" 2>/dev/null
+    pip install --target "$VENV_T5_550_DIR" --no-deps "huggingface_hub==1.8.0" 2>/dev/null
+    pip install --target "$VENV_T5_550_DIR" --no-deps "hf_xet==1.4.2" 2>/dev/null
+    pip install --target "$VENV_T5_550_DIR" "tiktoken" 2>/dev/null
+
+    echo "✅ Transformers 5.3.0 pre-installed to $VENV_T5_530_DIR/"
+    echo "✅ Transformers 5.5.0 pre-installed to $VENV_T5_550_DIR/"
+
+    echo ""
+    echo "╔══════════════════════════════════════╗"
+    echo "║     Docker Studio Setup Complete!    ║"
+    echo "╚══════════════════════════════════════╝"
+    exit 0
+fi
+
 # ── Check if Python deps need updating ──
 # Compare installed package version against PyPI latest.
 # Skip all Python dependency work if versions match (fast update path).
@@ -504,6 +555,31 @@ fi
 
 if [ "$_SKIP_PYTHON_DEPS" = false ]; then
     install_python_stack
+
+    # ── 6b. Pre-install transformers 5.x into .venv_t5_530/ and .venv_t5_550/ ──
+    # Models like GLM-4.7-Flash, Qwen3 MoE need transformers>=5.3.0.
+    # Gemma 4 models need transformers>=5.5.0.
+    # Pre-install into separate directories to avoid runtime pip overhead.
+    # The training subprocess prepends the appropriate dir to sys.path.
+
+    # Clean up legacy single .venv_t5 directory
+    [ -d "$STUDIO_HOME/.venv_t5" ] && rm -rf "$STUDIO_HOME/.venv_t5"
+
+    [ -d "$VENV_T5_530_DIR" ] && rm -rf "$VENV_T5_530_DIR"
+    mkdir -p "$VENV_T5_530_DIR"
+    run_quiet "install transformers 5.3.0" fast_install --target "$VENV_T5_530_DIR" --no-deps "transformers==5.3.0"
+    run_quiet "install huggingface_hub for t5_530" fast_install --target "$VENV_T5_530_DIR" --no-deps "huggingface_hub==1.8.0"
+    run_quiet "install hf_xet for t5_530" fast_install --target "$VENV_T5_530_DIR" --no-deps "hf_xet==1.4.2"
+    run_quiet "install tiktoken for t5_530" fast_install --target "$VENV_T5_530_DIR" "tiktoken"
+    step "transformers" "5.3.0 pre-installed"
+
+    [ -d "$VENV_T5_550_DIR" ] && rm -rf "$VENV_T5_550_DIR"
+    mkdir -p "$VENV_T5_550_DIR"
+    run_quiet "install transformers 5.5.0" fast_install --target "$VENV_T5_550_DIR" --no-deps "transformers==5.5.0"
+    run_quiet "install huggingface_hub for t5_550" fast_install --target "$VENV_T5_550_DIR" --no-deps "huggingface_hub==1.8.0"
+    run_quiet "install hf_xet for t5_550" fast_install --target "$VENV_T5_550_DIR" --no-deps "hf_xet==1.4.2"
+    run_quiet "install tiktoken for t5_550" fast_install --target "$VENV_T5_550_DIR" "tiktoken"
+    step "transformers" "5.5.0 pre-installed"
 else
     step "python" "dependencies up to date"
     verbose_substep "python deps check: installed=$_PKG_NAME@${INSTALLED_VER:-unknown} latest=${LATEST_VER:-unknown}"
@@ -548,6 +624,9 @@ fi
 fi
 
 # ── 7. Prefer prebuilt llama.cpp bundles before any source build path ──
+if [ "$_DOCKER_NO_VENV" = true ]; then
+    step "llama.cpp" "skipped (Docker)"
+else  # begin non-Docker llama.cpp block
 UNSLOTH_HOME="$HOME/.unsloth"
 mkdir -p "$UNSLOTH_HOME"
 LLAMA_CPP_DIR="$UNSLOTH_HOME/llama.cpp"
@@ -1031,6 +1110,7 @@ else
     fi
 }
 fi  # end _SKIP_GGUF_BUILD check
+fi  # end non-Docker llama.cpp block
 
 # ── Footer ──
 if [ "$_LLAMA_ONLY" = "1" ]; then

unsloth/models/loader.py

@@ -1562,7 +1562,6 @@ class FastModel(FastBaseModel):
 
             if _clippable_linear_cls is not None:
                 from peft.tuners.lora.model import LoraModel as _LoraModel
-
                 _original_car = _LoraModel._create_and_replace
 
                 def _patched_car(

unsloth_cli/commands/studio.py

@@ -158,56 +158,58 @@ def studio_default(
     if ctx.invoked_subcommand is not None:
         return
 
-    # Always use the studio venv if it exists and we're not already in it
-    studio_venv_dir = STUDIO_HOME / "unsloth_studio"
-    in_studio_venv = sys.prefix.startswith(str(studio_venv_dir))
+    # In Docker, packages live in /opt/conda — skip venv re-exec entirely.
+    if not os.environ.get("UNSLOTH_DOCKER"):
+        # Always use the studio venv if it exists and we're not already in it
+        studio_venv_dir = STUDIO_HOME / "unsloth_studio"
+        in_studio_venv = sys.prefix.startswith(str(studio_venv_dir))
 
-    if not in_studio_venv:
-        studio_python = _studio_venv_python()
-        run_py = _find_run_py()
-        if studio_python and run_py:
-            if not silent:
-                typer.echo("Launching Unsloth Studio... Please wait...")
-            args = [
-                str(studio_python),
-                str(run_py),
-                "--host",
-                host,
-                "--port",
-                str(port),
-            ]
-            if frontend:
-                args.extend(["--frontend", str(frontend)])
-            if silent:
-                args.append("--silent")
-            # On Windows, os.execvp() spawns a child but the parent lingers,
-            # so Ctrl+C only kills the parent leaving the child orphaned.
-            # Use subprocess.run() on Windows so the parent waits for the child.
-            if sys.platform == "win32":
-                import subprocess as _sp
+        if not in_studio_venv:
+            studio_python = _studio_venv_python()
+            run_py = _find_run_py()
+            if studio_python and run_py:
+                if not silent:
+                    typer.echo("Launching Unsloth Studio... Please wait...")
+                args = [
+                    str(studio_python),
+                    str(run_py),
+                    "--host",
+                    host,
+                    "--port",
+                    str(port),
+                ]
+                if frontend:
+                    args.extend(["--frontend", str(frontend)])
+                if silent:
+                    args.append("--silent")
+                # On Windows, os.execvp() spawns a child but the parent lingers,
+                # so Ctrl+C only kills the parent leaving the child orphaned.
+                # Use subprocess.run() on Windows so the parent waits for the child.
+                if sys.platform == "win32":
+                    import subprocess as _sp
 
-                proc = _sp.Popen(args)
-                try:
-                    rc = proc.wait()
-                except KeyboardInterrupt:
-                    # Child has its own signal handler — let it finish
-                    rc = proc.wait()
-                if rc != 0:
-                    typer.echo(
-                        f"\nError: Studio server exited unexpectedly (code {rc}).",
-                        err = True,
-                    )
-                    typer.echo(
-                        "Check the error above. If a package is missing, "
-                        "re-run: unsloth studio setup",
-                        err = True,
-                    )
-                raise typer.Exit(rc)
+                    proc = _sp.Popen(args)
+                    try:
+                        rc = proc.wait()
+                    except KeyboardInterrupt:
+                        # Child has its own signal handler — let it finish
+                        rc = proc.wait()
+                    if rc != 0:
+                        typer.echo(
+                            f"\nError: Studio server exited unexpectedly (code {rc}).",
+                            err = True,
+                        )
+                        typer.echo(
+                            "Check the error above. If a package is missing, "
+                            "re-run: unsloth studio setup",
+                            err = True,
+                        )
+                    raise typer.Exit(rc)
+                else:
+                    os.execvp(str(studio_python), args)
             else:
-                os.execvp(str(studio_python), args)
-        else:
-            typer.echo("Studio not set up. Run install.sh first.")
-            raise typer.Exit(1)
+                typer.echo("Studio not set up. Run install.sh first.")
+                raise typer.Exit(1)
 
     from studio.backend.run import run_server