From 04bca978b4616d33e9adf9aefea09783be8c758f Mon Sep 17 00:00:00 2001
From: Roland Tannous <rolandtannous@gravityq.ai>
Date: Fri, 27 Mar 2026 19:39:37 +0000
Subject: [PATCH 01/22] add Docker support: skip venv, install only missing
 deps

---
 studio/install_python_stack.py | 345 ++++++++++++++-------------------
 studio/setup.sh                |  41 +++-
 unsloth_cli/commands/studio.py |  96 ++++-----
 3 files changed, 229 insertions(+), 253 deletions(-)

diff --git a/studio/install_python_stack.py b/studio/install_python_stack.py
index a046f8f892..c63e8f6c2c 100644
--- a/studio/install_python_stack.py
+++ b/studio/install_python_stack.py
@@ -768,216 +768,153 @@ def install_python_stack() -> int:
         base_total += 2
     _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))
-
-        _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,
-        )
+    # # 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,
+    #     )
 
     # # 6. Patch: override llama_cpp.py with fix from unsloth-zoo  feature/llama-cpp-windows-support branch
     # patch_package_file(
diff --git a/studio/setup.sh b/studio/setup.sh
index c2044b5141..d2bd9485ba 100755
--- a/studio/setup.sh
+++ b/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 +475,38 @@ 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 "$SCRIPT_DIR/install_python_stack.py"
+
+    # Pre-install transformers 5.x into .venv_t5
+    echo ""
+    echo "   Pre-installing transformers 5.x for newer model support..."
+    mkdir -p "$VENV_T5_DIR"
+    pip install --target "$VENV_T5_DIR" --no-deps "transformers==5.3.0" 2>/dev/null
+    pip install --target "$VENV_T5_DIR" --no-deps "huggingface_hub==1.7.1" 2>/dev/null
+    pip install --target "$VENV_T5_DIR" --no-deps "hf_xet==1.4.2" 2>/dev/null
+    pip install --target "$VENV_T5_DIR" "tiktoken" 2>/dev/null
+    echo "✅ Transformers 5.x pre-installed to $VENV_T5_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).
diff --git a/unsloth_cli/commands/studio.py b/unsloth_cli/commands/studio.py
index 2fecb9d6b1..58b16736b5 100644
--- a/unsloth_cli/commands/studio.py
+++ b/unsloth_cli/commands/studio.py
@@ -84,56 +84,58 @@ 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
 

From 14f8daeaf38a8169d62ae4968faeb5bdec19efa1 Mon Sep 17 00:00:00 2001
From: Roland Tannous <rolandtannous@gravityq.ai>
Date: Thu, 2 Apr 2026 23:24:47 +0000
Subject: [PATCH 02/22] fix: patch PEFT for Gemma4ClippableLinear in loader
 checkpoint path

The same Gemma4ClippableLinear monkey-patch that exists in vision.py
for training is needed in loader.py for loading existing checkpoints
(used by export and inference).

Gemma4ClippableLinear wraps nn.Linear but does not subclass it, so
PEFT's LoRA injection fails with "Target module not supported".
The patch redirects PEFT to target the inner .linear child instead.

Applied only to the vision model PeftModel.from_pretrained path.
Temporary fix until PEFT adds native support (peft#3129).
---
 unsloth/models/loader.py | 3 +++
 1 file changed, 3 insertions(+)

diff --git a/unsloth/models/loader.py b/unsloth/models/loader.py
index df97c5c7df..2f933c7599 100644
--- a/unsloth/models/loader.py
+++ b/unsloth/models/loader.py
@@ -1545,7 +1545,10 @@ class FastModel(FastBaseModel):
 
             if _clippable_linear_cls is not None:
                 from peft.tuners.lora.model import LoraModel as _LoraModel
+<<<<<<< HEAD
 
+=======
+>>>>>>> 35ebf398 (fix: patch PEFT for Gemma4ClippableLinear in loader checkpoint path)
                 _original_car = _LoraModel._create_and_replace
 
                 def _patched_car(

From b9b829bf5985a5961287668f53c5536cd6327a1d Mon Sep 17 00:00:00 2001
From: "pre-commit-ci[bot]"
 <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Date: Thu, 2 Apr 2026 23:37:48 +0000
Subject: [PATCH 03/22] [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci
---
 unsloth/models/loader.py | 4 ----
 1 file changed, 4 deletions(-)

diff --git a/unsloth/models/loader.py b/unsloth/models/loader.py
index 2f933c7599..3bc810e4d2 100644
--- a/unsloth/models/loader.py
+++ b/unsloth/models/loader.py
@@ -1545,10 +1545,6 @@ class FastModel(FastBaseModel):
 
             if _clippable_linear_cls is not None:
                 from peft.tuners.lora.model import LoraModel as _LoraModel
-<<<<<<< HEAD
-
-=======
->>>>>>> 35ebf398 (fix: patch PEFT for Gemma4ClippableLinear in loader checkpoint path)
                 _original_car = _LoraModel._create_and_replace
 
                 def _patched_car(

From 47939e63fd33054fbe32e39cfcf0fc6936ae1d7f Mon Sep 17 00:00:00 2001
From: Roland Tannous <rolandtannous@gravityq.ai>
Date: Sun, 5 Apr 2026 04:50:43 +0000
Subject: [PATCH 04/22] Skip llama.cpp install in Docker mode

---
 studio/setup.sh | 4 ++++
 1 file changed, 4 insertions(+)

diff --git a/studio/setup.sh b/studio/setup.sh
index d2bd9485ba..20cc5b6ce7 100755
--- a/studio/setup.sh
+++ b/studio/setup.sh
@@ -585,6 +585,9 @@ 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"
@@ -1068,6 +1071,7 @@ else
     fi
 }
 fi  # end _SKIP_GGUF_BUILD check
+fi  # end non-Docker llama.cpp block
 
 # ── Footer ──
 if [ "$_LLAMA_ONLY" = "1" ]; then

From d0f15deab7be71ceec7007887885a1350e263c32 Mon Sep 17 00:00:00 2001
From: Roland Tannous <rolandtannous@gravityq.ai>
Date: Mon, 6 Apr 2026 16:25:07 +0000
Subject: [PATCH 05/22] split venv_t5 into venv_t5_530 and venv_t5_550 for
 tiered transformers 5.x support

---
 studio/backend/core/export/worker.py    |  3 -
 studio/backend/core/inference/worker.py |  2 -
 studio/backend/core/training/worker.py  | 17 +++++
 studio/setup.ps1                        | 83 +++++++++++++++++++++++++
 studio/setup.sh                         | 23 +++++++
 5 files changed, 123 insertions(+), 5 deletions(-)

diff --git a/studio/backend/core/export/worker.py b/studio/backend/core/export/worker.py
index 3f3dc955fa..26c3fa5455 100644
--- a/studio/backend/core/export/worker.py
+++ b/studio/backend/core/export/worker.py
@@ -31,11 +31,8 @@ logger = get_logger(__name__)
 
 def _activate_transformers_version(model_name: str) -> None:
     """Activate the correct transformers version BEFORE any ML imports."""
-    # Ensure backend is on path for utils imports
-    backend_path = str(Path(__file__).resolve().parent.parent.parent)
     if backend_path not in sys.path:
         sys.path.insert(0, backend_path)
-
     from utils.transformers_version import activate_transformers_for_subprocess
 
     activate_transformers_for_subprocess(model_name)
diff --git a/studio/backend/core/inference/worker.py b/studio/backend/core/inference/worker.py
index fbcce276ba..4f86f99ded 100644
--- a/studio/backend/core/inference/worker.py
+++ b/studio/backend/core/inference/worker.py
@@ -35,8 +35,6 @@ from utils.hardware import apply_gpu_ids
 
 def _activate_transformers_version(model_name: str) -> None:
     """Activate the correct transformers version BEFORE any ML imports."""
-    # Ensure backend is on path for utils imports
-    backend_path = str(Path(__file__).resolve().parent.parent.parent)
     if backend_path not in sys.path:
         sys.path.insert(0, backend_path)
 
diff --git a/studio/backend/core/training/worker.py b/studio/backend/core/training/worker.py
index c3d3daf07e..90236b3ded 100644
--- a/studio/backend/core/training/worker.py
+++ b/studio/backend/core/training/worker.py
@@ -448,6 +448,7 @@ def run_training_process(
         )
         return
 
+<<<<<<< HEAD
     # ── 1a. Auto-enable trust_remote_code for NemotronH/Nano models ──
     # NemotronH has config parsing bugs in transformers that require
     # trust_remote_code=True as a workaround. Other transformers 5.x models
@@ -459,6 +460,22 @@ def run_training_process(
     if (
         any(sub in _lowered for sub in _NEMOTRON_TRUST_SUBSTRINGS)
         and (_lowered.startswith("unsloth/") or _lowered.startswith("nvidia/"))
+=======
+    # ── 1a. Auto-enable trust_remote_code for unsloth/* transformers 5.x models ──
+    # Some newer architectures (e.g. NemotronH) have config parsing bugs in
+    # transformers that require trust_remote_code=True as a workaround.
+    # Only auto-enable for unsloth/* prefixed models (trusted source).
+    # Exclude Gemma 4 since it is a native transformers 5.5 model and
+    # trust_remote_code=True would bypass the compiler (disabling fused CE).
+    from utils.transformers_version import get_transformers_tier
+
+    _lowered = model_name.lower()
+    _tier = get_transformers_tier(model_name)
+    if (
+        _tier != "default"
+        and _lowered.startswith("unsloth/")
+        and _tier != "550"  # Gemma 4 is native t5.5 — trust_remote_code bypasses compiler
+>>>>>>> 970219a3 (split venv_t5 into venv_t5_530 and venv_t5_550 for tiered transformers 5.x support)
         and not config.get("trust_remote_code", False)
     ):
         config["trust_remote_code"] = True
diff --git a/studio/setup.ps1 b/studio/setup.ps1
index c3a8cd71ca..8d52358941 100644
--- a/studio/setup.ps1
+++ b/studio/setup.ps1
@@ -1579,6 +1579,89 @@ 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)
diff --git a/studio/setup.sh b/studio/setup.sh
index 20cc5b6ce7..0b9ca8383b 100755
--- a/studio/setup.sh
+++ b/studio/setup.sh
@@ -541,6 +541,29 @@ 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 "$STUDIO_HOME/.venv_t5"
+
+    mkdir -p "$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"
+
+    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}"

From 245f70d9ad9631da3dc9d3069d739e61324541d7 Mon Sep 17 00:00:00 2001
From: Roland Tannous <rolandtannous@gravityq.ai>
Date: Mon, 6 Apr 2026 17:28:55 +0000
Subject: [PATCH 06/22] fix bfloat16 crash on T4 for FORCE_FLOAT32 models and
 disable trust_remote_code auto-enable for native t5 models

---
 studio/backend/core/training/worker.py | 17 -----------------
 unsloth/models/loader.py               |  4 +++-
 2 files changed, 3 insertions(+), 18 deletions(-)

diff --git a/studio/backend/core/training/worker.py b/studio/backend/core/training/worker.py
index 90236b3ded..c3d3daf07e 100644
--- a/studio/backend/core/training/worker.py
+++ b/studio/backend/core/training/worker.py
@@ -448,7 +448,6 @@ def run_training_process(
         )
         return
 
-<<<<<<< HEAD
     # ── 1a. Auto-enable trust_remote_code for NemotronH/Nano models ──
     # NemotronH has config parsing bugs in transformers that require
     # trust_remote_code=True as a workaround. Other transformers 5.x models
@@ -460,22 +459,6 @@ def run_training_process(
     if (
         any(sub in _lowered for sub in _NEMOTRON_TRUST_SUBSTRINGS)
         and (_lowered.startswith("unsloth/") or _lowered.startswith("nvidia/"))
-=======
-    # ── 1a. Auto-enable trust_remote_code for unsloth/* transformers 5.x models ──
-    # Some newer architectures (e.g. NemotronH) have config parsing bugs in
-    # transformers that require trust_remote_code=True as a workaround.
-    # Only auto-enable for unsloth/* prefixed models (trusted source).
-    # Exclude Gemma 4 since it is a native transformers 5.5 model and
-    # trust_remote_code=True would bypass the compiler (disabling fused CE).
-    from utils.transformers_version import get_transformers_tier
-
-    _lowered = model_name.lower()
-    _tier = get_transformers_tier(model_name)
-    if (
-        _tier != "default"
-        and _lowered.startswith("unsloth/")
-        and _tier != "550"  # Gemma 4 is native t5.5 — trust_remote_code bypasses compiler
->>>>>>> 970219a3 (split venv_t5 into venv_t5_530 and venv_t5_550 for tiered transformers 5.x support)
         and not config.get("trust_remote_code", False)
     ):
         config["trust_remote_code"] = True
diff --git a/unsloth/models/loader.py b/unsloth/models/loader.py
index 3bc810e4d2..af46299e88 100644
--- a/unsloth/models/loader.py
+++ b/unsloth/models/loader.py
@@ -1357,7 +1357,9 @@ class FastModel(FastBaseModel):
                 or disable_name.lower() in model_types_all
             ) and ((dtype == torch.float16) or not SUPPORTS_BFLOAT16):
                 os.environ["UNSLOTH_FORCE_FLOAT32"] = "1"
-                dtype = torch.bfloat16  # Change to bfloat16 loading
+                # Use bfloat16 storage where supported; fall back to float32 on
+                # older GPUs (e.g. T4) that lack native bfloat16 support.
+                dtype = torch.bfloat16 if SUPPORTS_BFLOAT16 else torch.float32
                 break
         # Apply gradient checkpointing with smart heuristics
         use_gradient_checkpointing = apply_unsloth_gradient_checkpointing(

From f3ff84fc5cdb9325f462e33105ceb707a410ae2d Mon Sep 17 00:00:00 2001
From: Roland Tannous <rolandtannous@gravityq.ai>
Date: Mon, 6 Apr 2026 17:29:22 +0000
Subject: [PATCH 07/22] revert FORCE_FLOAT32 dtype change

---
 unsloth/models/loader.py | 4 +---
 1 file changed, 1 insertion(+), 3 deletions(-)

diff --git a/unsloth/models/loader.py b/unsloth/models/loader.py
index af46299e88..3bc810e4d2 100644
--- a/unsloth/models/loader.py
+++ b/unsloth/models/loader.py
@@ -1357,9 +1357,7 @@ class FastModel(FastBaseModel):
                 or disable_name.lower() in model_types_all
             ) and ((dtype == torch.float16) or not SUPPORTS_BFLOAT16):
                 os.environ["UNSLOTH_FORCE_FLOAT32"] = "1"
-                # Use bfloat16 storage where supported; fall back to float32 on
-                # older GPUs (e.g. T4) that lack native bfloat16 support.
-                dtype = torch.bfloat16 if SUPPORTS_BFLOAT16 else torch.float32
+                dtype = torch.bfloat16  # Change to bfloat16 loading
                 break
         # Apply gradient checkpointing with smart heuristics
         use_gradient_checkpointing = apply_unsloth_gradient_checkpointing(

From 833b38375ab07d859a4e0a9ff908494c3073d56d Mon Sep 17 00:00:00 2001
From: Roland Tannous <rolandtannous@gravityq.ai>
Date: Mon, 6 Apr 2026 18:20:08 +0000
Subject: [PATCH 08/22] restrict trust_remote_code auto-enable to Nemotron
 models only

---
 studio/backend/core/inference/worker.py |  52 ++++++--
 studio/backend/core/training/worker.py  | 170 ++++++++----------------
 2 files changed, 101 insertions(+), 121 deletions(-)

diff --git a/studio/backend/core/inference/worker.py b/studio/backend/core/inference/worker.py
index 4f86f99ded..1010e56dac 100644
--- a/studio/backend/core/inference/worker.py
+++ b/studio/backend/core/inference/worker.py
@@ -34,13 +34,50 @@ from utils.hardware import apply_gpu_ids
 
 
 def _activate_transformers_version(model_name: str) -> None:
-    """Activate the correct transformers version BEFORE any ML imports."""
+    """Activate the correct transformers version BEFORE any ML imports.
+
+    Uses get_transformers_tier() to decide between .venv_t5_550/ (5.5.0),
+    .venv_t5_530/ (5.3.0), or the default 4.57.x.
+    """
+    # Ensure backend is on path for utils imports
+    backend_path = str(Path(__file__).resolve().parent.parent.parent)
     if backend_path not in sys.path:
         sys.path.insert(0, backend_path)
 
-    from utils.transformers_version import activate_transformers_for_subprocess
+    from utils.transformers_version import (
+        get_transformers_tier,
+        _resolve_base_model,
+        _ensure_venv_t5_530_exists,
+        _ensure_venv_t5_550_exists,
+        _VENV_T5_530_DIR,
+        _VENV_T5_550_DIR,
+    )
 
-    activate_transformers_for_subprocess(model_name)
+    resolved = _resolve_base_model(model_name)
+    tier = get_transformers_tier(resolved)
+
+    if tier == "550":
+        if not _ensure_venv_t5_550_exists():
+            raise RuntimeError(
+                f"Cannot activate transformers 5.5.0: .venv_t5_550 missing at {_VENV_T5_550_DIR}"
+            )
+        if _VENV_T5_550_DIR not in sys.path:
+            sys.path.insert(0, _VENV_T5_550_DIR)
+        logger.info("Activated transformers 5.5.0 from %s", _VENV_T5_550_DIR)
+        _pp = os.environ.get("PYTHONPATH", "")
+        os.environ["PYTHONPATH"] = _VENV_T5_550_DIR + (os.pathsep + _pp if _pp else "")
+    elif tier == "530":
+        if not _ensure_venv_t5_530_exists():
+            raise RuntimeError(
+                f"Cannot activate transformers 5.3.0: .venv_t5_530 missing at {_VENV_T5_530_DIR}"
+            )
+        if _VENV_T5_530_DIR not in sys.path:
+            sys.path.insert(0, _VENV_T5_530_DIR)
+        logger.info("Activated transformers 5.3.0 from %s", _VENV_T5_530_DIR)
+        _pp = os.environ.get("PYTHONPATH", "")
+        os.environ["PYTHONPATH"] = _VENV_T5_530_DIR + (os.pathsep + _pp if _pp else "")
+    else:
+        logger.info("Using default transformers (4.57.x) for %s", model_name)
 
 
 def _decode_image(image_base64: str):
@@ -285,18 +322,13 @@ def _handle_load(backend, config: dict, resp_queue: Any) -> None:
                 except Exception as e:
                     logger.warning("Could not read adapter_config.json: %s", e)
 
-        # Auto-enable trust_remote_code for NemotronH/Nano models only.
+        # Auto-enable trust_remote_code for Nemotron models only.
         # NemotronH has config parsing bugs requiring trust_remote_code=True.
         # Other transformers 5.x models are native and do NOT need it.
-        # NOTE: Must NOT match Llama-Nemotron (standard Llama architecture).
-        _NEMOTRON_TRUST_SUBSTRINGS = ("nemotron_h", "nemotron-h", "nemotron-3-nano")
         trust_remote_code = config.get("trust_remote_code", False)
         if not trust_remote_code:
             model_name = config["model_name"]
-            _mn_lower = model_name.lower()
-            if any(sub in _mn_lower for sub in _NEMOTRON_TRUST_SUBSTRINGS) and (
-                _mn_lower.startswith("unsloth/") or _mn_lower.startswith("nvidia/")
-            ):
+            if "nemotron" in model_name.lower():
                 trust_remote_code = True
                 logger.info(
                     "Auto-enabled trust_remote_code for Nemotron model: %s",
diff --git a/studio/backend/core/training/worker.py b/studio/backend/core/training/worker.py
index c3d3daf07e..b412a05cc9 100644
--- a/studio/backend/core/training/worker.py
+++ b/studio/backend/core/training/worker.py
@@ -86,7 +86,6 @@ def _probe_causal_conv1d_env() -> dict[str, str] | None:
                     "'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 '', "
-                    "'hip_version': str(torch.version.hip) if getattr(torch.version, 'hip', None) else '', "
                     "'cxx11abi': str(torch._C._GLIBCXX_USE_CXX11_ABI).upper()"
                     "}))"
                 ),
@@ -238,111 +237,28 @@ def _install_package_wheel_first(
         else:
             logger.info("No published %s wheel found: %s", display_name, wheel_url)
 
-    is_hip = env and env.get("hip_version")
-    if is_hip and not shutil.which("hipcc"):
-        logger.error(
-            "%s requires hipcc for source compilation on ROCm. "
-            "Install the ROCm HIP SDK: https://rocm.docs.amd.com",
-            display_name,
-        )
-        _send_status(
-            event_queue,
-            f"{display_name}: hipcc not found (ROCm HIP SDK required)",
-        )
-        return
-
-    if is_hip:
-        _send_status(
-            event_queue,
-            f"Compiling {display_name} from source for ROCm "
-            "(this may take several minutes)...",
-        )
-    else:
-        _send_status(event_queue, f"Installing {display_name} from PyPI...")
-
-    # Prefer uv for faster dependency resolution when available
-    if shutil.which("uv"):
-        pypi_cmd = [
-            "uv",
-            "pip",
-            "install",
-            "--python",
-            sys.executable,
-            "--no-build-isolation",
-            "--no-deps",
-        ]
-        # Avoid stale cache artifacts from partial HIP source builds
-        if is_hip:
-            pypi_cmd.append("--no-cache")
-        pypi_cmd.append(f"{pypi_name}=={pypi_version}")
-    else:
-        pypi_cmd = [
-            sys.executable,
-            "-m",
-            "pip",
-            "install",
-            "--no-build-isolation",
-            "--no-deps",
-            "--no-cache-dir",
-            f"{pypi_name}=={pypi_version}",
-        ]
-
-    # Source compilation on ROCm can take 10-30 minutes; use a generous
-    # timeout. Non-HIP installs preserve the pre-existing "no timeout"
-    # behaviour so unrelated slow installs (e.g. causal-conv1d source
-    # build on Linux aarch64 or unsupported torch/CUDA combinations)
-    # are not aborted at 5 minutes by this PR.
-    _run_kwargs: dict[str, Any] = {
-        "stdout": _sp.PIPE,
-        "stderr": _sp.STDOUT,
-        "text": True,
-    }
-    if is_hip:
-        _run_kwargs["timeout"] = 1800
-
-    try:
-        result = _sp.run(pypi_cmd, **_run_kwargs)
-    except _sp.TimeoutExpired:
-        logger.error(
-            "%s installation timed out after %ds",
-            display_name,
-            _run_kwargs.get("timeout"),
-        )
-        _send_status(
-            event_queue,
-            f"{display_name} installation timed out after "
-            f"{_run_kwargs.get('timeout')}s",
-        )
-        return
-
+    _send_status(event_queue, f"Installing {display_name} from PyPI...")
+    pypi_cmd = [
+        sys.executable,
+        "-m",
+        "pip",
+        "install",
+        "--no-build-isolation",
+        "--no-deps",
+        "--no-cache-dir",
+        f"{pypi_name}=={pypi_version}",
+    ]
+    result = _sp.run(
+        pypi_cmd,
+        stdout = _sp.PIPE,
+        stderr = _sp.STDOUT,
+        text = True,
+    )
     if result.returncode != 0:
-        if is_hip:
-            # Surface a clear error for ROCm source build failures
-            error_lines = (result.stdout or "").strip().splitlines()
-            snippet = "\n".join(error_lines[-5:]) if error_lines else "(no output)"
-            logger.error(
-                "Failed to compile %s for ROCm:\n%s",
-                display_name,
-                result.stdout,
-            )
-            _send_status(
-                event_queue,
-                f"Failed to compile {display_name} for ROCm. "
-                "Check that hipcc and ROCm development headers are installed.\n"
-                f"{snippet}",
-            )
-        else:
-            logger.error(
-                "Failed to install %s from PyPI:\n%s",
-                display_name,
-                result.stdout,
-            )
+        logger.error("Failed to install %s from PyPI:\n%s", display_name, result.stdout)
         return
 
-    if is_hip:
-        logger.info("Compiled and installed %s from source for ROCm", display_name)
-    else:
-        logger.info("Installed %s from PyPI", display_name)
+    logger.info("Installed %s from PyPI", display_name)
 
 
 def _ensure_causal_conv1d_fast_path(event_queue: Any, model_name: str) -> None:
@@ -390,15 +306,50 @@ def _ensure_mamba_ssm(event_queue: Any, model_name: str) -> None:
 
 
 def _activate_transformers_version(model_name: str) -> None:
-    """Activate the correct transformers version BEFORE any ML imports."""
+    """Activate the correct transformers version BEFORE any ML imports.
+
+    Uses get_transformers_tier() to decide between .venv_t5_550/ (5.5.0),
+    .venv_t5_530/ (5.3.0), or the default 4.57.x.
+    """
     # Ensure backend is on path for utils imports
     backend_path = str(Path(__file__).resolve().parent.parent.parent)
     if backend_path not in sys.path:
         sys.path.insert(0, backend_path)
 
-    from utils.transformers_version import activate_transformers_for_subprocess
+    from utils.transformers_version import (
+        get_transformers_tier,
+        _resolve_base_model,
+        _ensure_venv_t5_530_exists,
+        _ensure_venv_t5_550_exists,
+        _VENV_T5_530_DIR,
+        _VENV_T5_550_DIR,
+    )
 
-    activate_transformers_for_subprocess(model_name)
+    resolved = _resolve_base_model(model_name)
+    tier = get_transformers_tier(resolved)
+
+    if tier == "550":
+        if not _ensure_venv_t5_550_exists():
+            raise RuntimeError(
+                f"Cannot activate transformers 5.5.0: .venv_t5_550 missing at {_VENV_T5_550_DIR}"
+            )
+        if _VENV_T5_550_DIR not in sys.path:
+            sys.path.insert(0, _VENV_T5_550_DIR)
+        logger.info("Activated transformers 5.5.0 from %s", _VENV_T5_550_DIR)
+        _pp = os.environ.get("PYTHONPATH", "")
+        os.environ["PYTHONPATH"] = _VENV_T5_550_DIR + (os.pathsep + _pp if _pp else "")
+    elif tier == "530":
+        if not _ensure_venv_t5_530_exists():
+            raise RuntimeError(
+                f"Cannot activate transformers 5.3.0: .venv_t5_530 missing at {_VENV_T5_530_DIR}"
+            )
+        if _VENV_T5_530_DIR not in sys.path:
+            sys.path.insert(0, _VENV_T5_530_DIR)
+        logger.info("Activated transformers 5.3.0 from %s", _VENV_T5_530_DIR)
+        _pp = os.environ.get("PYTHONPATH", "")
+        os.environ["PYTHONPATH"] = _VENV_T5_530_DIR + (os.pathsep + _pp if _pp else "")
+    else:
+        logger.info("Using default transformers (4.57.x) for %s", model_name)
 
 
 def run_training_process(
@@ -448,17 +399,14 @@ def run_training_process(
         )
         return
 
-    # ── 1a. Auto-enable trust_remote_code for NemotronH/Nano models ──
+    # ── 1a. Auto-enable trust_remote_code for Nemotron models ──
     # NemotronH has config parsing bugs in transformers that require
     # trust_remote_code=True as a workaround. Other transformers 5.x models
     # (Qwen3.5, Gemma 4, etc.) are native and do NOT need it — enabling it
     # bypasses the compiler (disabling fused CE).
-    # NOTE: Must NOT match Llama-Nemotron (standard Llama architecture).
-    _NEMOTRON_TRUST_SUBSTRINGS = ("nemotron_h", "nemotron-h", "nemotron-3-nano")
     _lowered = model_name.lower()
     if (
-        any(sub in _lowered for sub in _NEMOTRON_TRUST_SUBSTRINGS)
-        and (_lowered.startswith("unsloth/") or _lowered.startswith("nvidia/"))
+        "nemotron" in _lowered
         and not config.get("trust_remote_code", False)
     ):
         config["trust_remote_code"] = True

From db5a044c91929bc5310147e5030222ed82e45dfa Mon Sep 17 00:00:00 2001
From: "pre-commit-ci[bot]"
 <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Date: Mon, 6 Apr 2026 19:09:39 +0000
Subject: [PATCH 09/22] [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci
---
 studio/backend/core/training/worker.py | 5 +----
 1 file changed, 1 insertion(+), 4 deletions(-)

diff --git a/studio/backend/core/training/worker.py b/studio/backend/core/training/worker.py
index b412a05cc9..54667dd7af 100644
--- a/studio/backend/core/training/worker.py
+++ b/studio/backend/core/training/worker.py
@@ -405,10 +405,7 @@ def run_training_process(
     # (Qwen3.5, Gemma 4, etc.) are native and do NOT need it — enabling it
     # bypasses the compiler (disabling fused CE).
     _lowered = model_name.lower()
-    if (
-        "nemotron" in _lowered
-        and not config.get("trust_remote_code", False)
-    ):
+    if "nemotron" in _lowered and not config.get("trust_remote_code", False):
         config["trust_remote_code"] = True
         logger.info(
             "Auto-enabled trust_remote_code for Nemotron model: %s",

From 567df493aba62c03136f6361fcf7d2291e3aedb7 Mon Sep 17 00:00:00 2001
From: Roland Tannous <rolandtannous@gravityq.ai>
Date: Mon, 6 Apr 2026 19:40:11 +0000
Subject: [PATCH 10/22] use config.json model_type for tier detection, add
 unsloth/nvidia namespace guard

---
 studio/backend/core/inference/worker.py       |   6 +-
 studio/backend/core/training/worker.py        |   6 +-
 .../tests/test_transformers_version.py        |  25 ++-
 studio/backend/utils/transformers_version.py  | 161 ++++++++----------
 4 files changed, 99 insertions(+), 99 deletions(-)

diff --git a/studio/backend/core/inference/worker.py b/studio/backend/core/inference/worker.py
index 1010e56dac..506e631c51 100644
--- a/studio/backend/core/inference/worker.py
+++ b/studio/backend/core/inference/worker.py
@@ -328,7 +328,11 @@ def _handle_load(backend, config: dict, resp_queue: Any) -> None:
         trust_remote_code = config.get("trust_remote_code", False)
         if not trust_remote_code:
             model_name = config["model_name"]
-            if "nemotron" in model_name.lower():
+            _mn_lower = model_name.lower()
+            if (
+                "nemotron" in _mn_lower
+                and (_mn_lower.startswith("unsloth/") or _mn_lower.startswith("nvidia/"))
+            ):
                 trust_remote_code = True
                 logger.info(
                     "Auto-enabled trust_remote_code for Nemotron model: %s",
diff --git a/studio/backend/core/training/worker.py b/studio/backend/core/training/worker.py
index 54667dd7af..fa05817697 100644
--- a/studio/backend/core/training/worker.py
+++ b/studio/backend/core/training/worker.py
@@ -405,7 +405,11 @@ def run_training_process(
     # (Qwen3.5, Gemma 4, etc.) are native and do NOT need it — enabling it
     # bypasses the compiler (disabling fused CE).
     _lowered = model_name.lower()
-    if "nemotron" in _lowered and not config.get("trust_remote_code", False):
+    if (
+        "nemotron" in _lowered
+        and (_lowered.startswith("unsloth/") or _lowered.startswith("nvidia/"))
+        and not config.get("trust_remote_code", False)
+    ):
         config["trust_remote_code"] = True
         logger.info(
             "Auto-enabled trust_remote_code for Nemotron model: %s",
diff --git a/studio/backend/tests/test_transformers_version.py b/studio/backend/tests/test_transformers_version.py
index c031c2fea3..609a154a9a 100644
--- a/studio/backend/tests/test_transformers_version.py
+++ b/studio/backend/tests/test_transformers_version.py
@@ -32,8 +32,9 @@ from utils.transformers_version import (
     _resolve_base_model,
     _check_tokenizer_config_needs_v5,
     _check_config_needs_550,
+    _get_config_json,
     _tokenizer_class_cache,
-    _config_needs_550_cache,
+    _config_json_cache,
     needs_transformers_5,
     get_transformers_tier,
 )
@@ -202,7 +203,7 @@ class TestCheckConfigNeeds550:
     """Tests for _check_config_needs_550() local config.json checks."""
 
     def setup_method(self):
-        _config_needs_550_cache.clear()
+        _config_json_cache.clear()
 
     def test_gemma4_architecture(self, tmp_path: Path):
         """config.json with Gemma4ForConditionalGeneration should return True."""
@@ -242,8 +243,8 @@ class TestCheckConfigNeeds550:
 
         key = str(tmp_path)
         _check_config_needs_550(key)
-        assert key in _config_needs_550_cache
-        assert _config_needs_550_cache[key] is True
+        assert key in _config_json_cache
+        assert _config_json_cache[key] is not None
 
     def test_local_file_skips_network(self, tmp_path: Path):
         """When local config.json exists, no network request should be made."""
@@ -265,7 +266,7 @@ class TestGetTransformersTier:
 
     def setup_method(self):
         _tokenizer_class_cache.clear()
-        _config_needs_550_cache.clear()
+        _config_json_cache.clear()
 
     def test_gemma4_substring_returns_550(self):
         assert get_transformers_tier("google/gemma-4-E2B-it") == "550"
@@ -317,6 +318,20 @@ class TestGetTransformersTier:
         # This shouldn't happen in practice, but verifies priority
         assert get_transformers_tier("gemma-4-model") == "550"
 
+    def test_config_json_model_type_530(self, tmp_path: Path):
+        """Local checkpoint with qwen3_moe model_type → 530."""
+        cfg = {"model_type": "qwen3_moe", "architectures": ["Qwen3MoeForCausalLM"]}
+        (tmp_path / "config.json").write_text(json.dumps(cfg))
+
+        assert get_transformers_tier(str(tmp_path)) == "530"
+
+    def test_config_json_model_type_glm4_moe(self, tmp_path: Path):
+        """Local checkpoint with glm4_moe model_type → 530."""
+        cfg = {"model_type": "glm4_moe", "architectures": ["Glm4MoeForCausalLM"]}
+        (tmp_path / "config.json").write_text(json.dumps(cfg))
+
+        assert get_transformers_tier(str(tmp_path)) == "530"
+
     def test_needs_transformers_5_compat(self):
         """needs_transformers_5 should return True for both 530 and 550 models."""
         assert needs_transformers_5("google/gemma-4-E2B-it") is True
diff --git a/studio/backend/utils/transformers_version.py b/studio/backend/utils/transformers_version.py
index 0c13b5455b..308f416940 100644
--- a/studio/backend/utils/transformers_version.py
+++ b/studio/backend/utils/transformers_version.py
@@ -60,14 +60,25 @@ TRANSFORMERS_550_MODEL_SUBSTRINGS: tuple[str, ...] = (
     "gemma4",  # Gemma-4 alternate naming
 )
 
-# Architecture classes / model_type values that require transformers 5.5.0.
-# Checked via config.json (local or HuggingFace).
+# Architecture classes that require transformers 5.5.0.
 _TRANSFORMERS_550_ARCHITECTURES: set[str] = {
     "Gemma4ForConditionalGeneration",
 }
+
+# model_type values (from config.json) → tier mapping.
 _TRANSFORMERS_550_MODEL_TYPES: set[str] = {
     "gemma4",
 }
+_TRANSFORMERS_530_MODEL_TYPES: set[str] = {
+    "qwen3_moe",
+    "qwen3_5_moe",
+    "qwen3_vl_moe",
+    "qwen3_next",
+    "deepseek_v3_moe",
+    "glm4_moe",
+    "glm4_moe_lite",
+    "ministral",
+}
 
 # Tokenizer classes that only exist in transformers>=5.x
 _TRANSFORMERS_5_TOKENIZER_CLASSES: set[str] = {
@@ -77,15 +88,14 @@ _TRANSFORMERS_5_TOKENIZER_CLASSES: set[str] = {
 # Cache for dynamic tokenizer_config.json lookups to avoid repeated fetches
 _tokenizer_class_cache: dict[str, bool] = {}
 
-# Cache for dynamic config.json lookups (architecture/model_type checks)
-_config_needs_550_cache: dict[str, bool] = {}
+# Cache for config.json lookups (returns the parsed dict or None)
+_config_json_cache: dict[str, dict | None] = {}
 
 # Versions
 TRANSFORMERS_550_VERSION = "5.5.0"
 TRANSFORMERS_530_VERSION = "5.3.0"
 TRANSFORMERS_DEFAULT_VERSION = "4.57.6"
-# Backwards-compat alias — points to 5.5.0 (the highest 5.x tier).
-# Consumers should prefer TRANSFORMERS_530_VERSION / TRANSFORMERS_550_VERSION.
+# Backwards-compat alias used by other modules
 TRANSFORMERS_5_VERSION = TRANSFORMERS_550_VERSION
 
 # Pre-installed directories — created by setup.sh / setup.ps1
@@ -95,45 +105,6 @@ _VENV_T5_550_DIR = str(Path.home() / ".unsloth" / "studio" / ".venv_t5_550")
 _VENV_T5_DIR = _VENV_T5_550_DIR
 
 
-def activate_transformers_for_subprocess(model_name: str) -> None:
-    """Activate the correct transformers version in a subprocess worker.
-
-    Call this BEFORE any ML imports. Resolves LoRA adapters to their base
-    model, determines the required tier, and prepends the appropriate
-    ``.venv_t5_*`` directory to ``sys.path``.  Also propagates the path
-    via ``PYTHONPATH`` for child processes (e.g. GGUF converter).
-
-    Used by training, inference, and export workers.
-    """
-    resolved = _resolve_base_model(model_name)
-    tier = get_transformers_tier(resolved)
-
-    if tier == "550":
-        if not _ensure_venv_t5_550_exists():
-            raise RuntimeError(
-                f"Cannot activate transformers 5.5.0: "
-                f".venv_t5_550 missing at {_VENV_T5_550_DIR}"
-            )
-        if _VENV_T5_550_DIR not in sys.path:
-            sys.path.insert(0, _VENV_T5_550_DIR)
-        logger.info("Activated transformers 5.5.0 from %s", _VENV_T5_550_DIR)
-        _pp = os.environ.get("PYTHONPATH", "")
-        os.environ["PYTHONPATH"] = _VENV_T5_550_DIR + (os.pathsep + _pp if _pp else "")
-    elif tier == "530":
-        if not _ensure_venv_t5_530_exists():
-            raise RuntimeError(
-                f"Cannot activate transformers 5.3.0: "
-                f".venv_t5_530 missing at {_VENV_T5_530_DIR}"
-            )
-        if _VENV_T5_530_DIR not in sys.path:
-            sys.path.insert(0, _VENV_T5_530_DIR)
-        logger.info("Activated transformers 5.3.0 from %s", _VENV_T5_530_DIR)
-        _pp = os.environ.get("PYTHONPATH", "")
-        os.environ["PYTHONPATH"] = _VENV_T5_530_DIR + (os.pathsep + _pp if _pp else "")
-    else:
-        logger.info("Using default transformers (4.57.x) for %s", model_name)
-
-
 def _resolve_base_model(model_name: str) -> str:
     """If *model_name* points to a LoRA adapter, return its base model.
 
@@ -259,43 +230,28 @@ def _check_tokenizer_config_needs_v5(model_name: str) -> bool:
         return False
 
 
-def _check_config_needs_550(model_name: str) -> bool:
-    """Check ``config.json`` for architectures or model_type that require
-    transformers 5.5.0 (e.g. Gemma 4).
+_SENTINEL = object()  # distinguishes "not cached" from "cached as None"
 
-    Checks locally first, then falls back to fetching from HuggingFace.
-    Results are cached in ``_config_needs_550_cache``.
-    Returns False on any error (fail-open to lower tier).
+
+def _get_config_json(model_name: str) -> dict | None:
+    """Read and cache ``config.json`` for *model_name*.
+
+    Checks local path first, then fetches from HuggingFace.
+    Returns the parsed dict, or ``None`` on any error (fail-open).
+    The result is cached in ``_config_json_cache``.
     """
-    if model_name in _config_needs_550_cache:
-        return _config_needs_550_cache[model_name]
-
-    def _check_cfg(cfg: dict) -> bool:
-        archs = cfg.get("architectures", [])
-        if any(a in _TRANSFORMERS_550_ARCHITECTURES for a in archs):
-            return True
-        if cfg.get("model_type") in _TRANSFORMERS_550_MODEL_TYPES:
-            return True
-        return False
+    cached = _config_json_cache.get(model_name, _SENTINEL)
+    if cached is not _SENTINEL:
+        return cached
 
     # --- Check local config.json first ------------------------------------
-    local_path = Path(model_name)
-    local_cfg = local_path / "config.json"
+    local_cfg = Path(model_name) / "config.json"
     if local_cfg.is_file():
         try:
             with open(local_cfg) as f:
                 cfg = json.load(f)
-            result = _check_cfg(cfg)
-            if result:
-                logger.info(
-                    "Local config.json check: %s needs transformers 5.5.0 "
-                    "(architectures=%s, model_type=%s)",
-                    model_name,
-                    cfg.get("architectures", []),
-                    cfg.get("model_type"),
-                )
-            _config_needs_550_cache[model_name] = result
-            return result
+            _config_json_cache[model_name] = cfg
+            return cfg
         except Exception as exc:
             logger.debug("Could not read %s: %s", local_cfg, exc)
 
@@ -307,21 +263,26 @@ def _check_config_needs_550(model_name: str) -> bool:
         req = urllib.request.Request(url, headers = {"User-Agent": "unsloth-studio"})
         with urllib.request.urlopen(req, timeout = 10) as resp:
             cfg = json.loads(resp.read().decode())
-        result = _check_cfg(cfg)
-        if result:
-            logger.info(
-                "Dynamic config.json check: %s needs transformers 5.5.0 "
-                "(architectures=%s, model_type=%s)",
-                model_name,
-                cfg.get("architectures", []),
-                cfg.get("model_type"),
-            )
-        _config_needs_550_cache[model_name] = result
-        return result
+        _config_json_cache[model_name] = cfg
+        return cfg
     except Exception as exc:
-        logger.debug("Could not fetch config.json for '%s': %s", model_name, exc)
-        _config_needs_550_cache[model_name] = False
+        logger.debug(
+            "Could not fetch config.json for '%s': %s", model_name, exc
+        )
+        _config_json_cache[model_name] = None
+        return None
+
+
+def _check_config_needs_550(model_name: str) -> bool:
+    """Check ``config.json`` for architectures or model_type that require
+    transformers 5.5.0.  Uses the shared ``_get_config_json`` cache."""
+    cfg = _get_config_json(model_name)
+    if cfg is None:
         return False
+    archs = cfg.get("architectures", [])
+    if any(a in _TRANSFORMERS_550_ARCHITECTURES for a in archs):
+        return True
+    return cfg.get("model_type") in _TRANSFORMERS_550_MODEL_TYPES
 
 
 def get_transformers_tier(model_name: str) -> str:
@@ -331,19 +292,35 @@ def get_transformers_tier(model_name: str) -> str:
     ``"530"`` for models needing transformers 5.3.0 (e.g. Ministral-3, Qwen3 MoE),
     or ``"default"`` for everything else (4.57.x).
 
-    The 5.5.0 check runs first, then 5.3.0.
+    Fast path: substring checks (no I/O) for both tiers run first.
+    Slow path: single config.json fetch (cached) checks model_type for
+    both tiers, then tokenizer_config.json as final fallback.
     """
     lowered = model_name.lower()
 
-    # --- Fast substring checks (no I/O) ------------------------------------
+    # --- Fast substring checks (no I/O) -----------------------------------
     if any(sub in lowered for sub in TRANSFORMERS_550_MODEL_SUBSTRINGS):
         return "550"
     if any(sub in lowered for sub in TRANSFORMERS_5_MODEL_SUBSTRINGS):
         return "530"
 
-    # --- Slow config fallbacks (local file first, then network) -----------
-    if _check_config_needs_550(model_name):
-        return "550"
+    # --- config.json model_type / architecture check (single fetch) -------
+    cfg = _get_config_json(model_name)
+    if cfg is not None:
+        model_type = cfg.get("model_type", "")
+        archs = cfg.get("architectures", [])
+
+        # Check 5.5.0 first
+        if model_type in _TRANSFORMERS_550_MODEL_TYPES:
+            return "550"
+        if any(a in _TRANSFORMERS_550_ARCHITECTURES for a in archs):
+            return "550"
+
+        # Check 5.3.0
+        if model_type in _TRANSFORMERS_530_MODEL_TYPES:
+            return "530"
+
+    # --- Final fallback: tokenizer_config.json for 5.3.0 ------------------
     if _check_tokenizer_config_needs_v5(model_name):
         return "530"
 

From a2aba7a2be78cecbb5fc0a1a2e35cfee6d9f9a2a Mon Sep 17 00:00:00 2001
From: "pre-commit-ci[bot]"
 <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Date: Mon, 6 Apr 2026 19:42:48 +0000
Subject: [PATCH 11/22] [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci
---
 studio/backend/core/inference/worker.py      | 5 ++---
 studio/backend/utils/transformers_version.py | 4 +---
 2 files changed, 3 insertions(+), 6 deletions(-)

diff --git a/studio/backend/core/inference/worker.py b/studio/backend/core/inference/worker.py
index 506e631c51..db504bac3a 100644
--- a/studio/backend/core/inference/worker.py
+++ b/studio/backend/core/inference/worker.py
@@ -329,9 +329,8 @@ def _handle_load(backend, config: dict, resp_queue: Any) -> None:
         if not trust_remote_code:
             model_name = config["model_name"]
             _mn_lower = model_name.lower()
-            if (
-                "nemotron" in _mn_lower
-                and (_mn_lower.startswith("unsloth/") or _mn_lower.startswith("nvidia/"))
+            if "nemotron" in _mn_lower and (
+                _mn_lower.startswith("unsloth/") or _mn_lower.startswith("nvidia/")
             ):
                 trust_remote_code = True
                 logger.info(
diff --git a/studio/backend/utils/transformers_version.py b/studio/backend/utils/transformers_version.py
index 308f416940..643a0a827a 100644
--- a/studio/backend/utils/transformers_version.py
+++ b/studio/backend/utils/transformers_version.py
@@ -266,9 +266,7 @@ def _get_config_json(model_name: str) -> dict | None:
         _config_json_cache[model_name] = cfg
         return cfg
     except Exception as exc:
-        logger.debug(
-            "Could not fetch config.json for '%s': %s", model_name, exc
-        )
+        logger.debug("Could not fetch config.json for '%s': %s", model_name, exc)
         _config_json_cache[model_name] = None
         return None
 

From 370b41c4fc9eb1c191a222b760387ff6b5d37b59 Mon Sep 17 00:00:00 2001
From: Roland Tannous <rolandtannous@gravityq.ai>
Date: Mon, 6 Apr 2026 20:26:32 +0000
Subject: [PATCH 12/22] Revert "[pre-commit.ci] auto fixes from pre-commit.com
 hooks"

This reverts commit fb43d468e25379f28dd2477e6c24dd60cf55c099.
---
 studio/backend/core/inference/worker.py      | 5 +++--
 studio/backend/utils/transformers_version.py | 4 +++-
 2 files changed, 6 insertions(+), 3 deletions(-)

diff --git a/studio/backend/core/inference/worker.py b/studio/backend/core/inference/worker.py
index db504bac3a..506e631c51 100644
--- a/studio/backend/core/inference/worker.py
+++ b/studio/backend/core/inference/worker.py
@@ -329,8 +329,9 @@ def _handle_load(backend, config: dict, resp_queue: Any) -> None:
         if not trust_remote_code:
             model_name = config["model_name"]
             _mn_lower = model_name.lower()
-            if "nemotron" in _mn_lower and (
-                _mn_lower.startswith("unsloth/") or _mn_lower.startswith("nvidia/")
+            if (
+                "nemotron" in _mn_lower
+                and (_mn_lower.startswith("unsloth/") or _mn_lower.startswith("nvidia/"))
             ):
                 trust_remote_code = True
                 logger.info(
diff --git a/studio/backend/utils/transformers_version.py b/studio/backend/utils/transformers_version.py
index 643a0a827a..308f416940 100644
--- a/studio/backend/utils/transformers_version.py
+++ b/studio/backend/utils/transformers_version.py
@@ -266,7 +266,9 @@ def _get_config_json(model_name: str) -> dict | None:
         _config_json_cache[model_name] = cfg
         return cfg
     except Exception as exc:
-        logger.debug("Could not fetch config.json for '%s': %s", model_name, exc)
+        logger.debug(
+            "Could not fetch config.json for '%s': %s", model_name, exc
+        )
         _config_json_cache[model_name] = None
         return None
 

From 4f01b4569d10852e22962a45b98b6282e138a1c1 Mon Sep 17 00:00:00 2001
From: Roland Tannous <rolandtannous@gravityq.ai>
Date: Mon, 6 Apr 2026 20:26:32 +0000
Subject: [PATCH 13/22] Revert "use config.json model_type for tier detection,
 add unsloth/nvidia namespace guard"

This reverts commit fc49ae24531780a658049e6c238f146960f466b6.
---
 studio/backend/core/inference/worker.py       |   6 +-
 studio/backend/core/training/worker.py        |   6 +-
 .../tests/test_transformers_version.py        |  25 +---
 studio/backend/utils/transformers_version.py  | 121 ++++++++----------
 4 files changed, 59 insertions(+), 99 deletions(-)

diff --git a/studio/backend/core/inference/worker.py b/studio/backend/core/inference/worker.py
index 506e631c51..1010e56dac 100644
--- a/studio/backend/core/inference/worker.py
+++ b/studio/backend/core/inference/worker.py
@@ -328,11 +328,7 @@ def _handle_load(backend, config: dict, resp_queue: Any) -> None:
         trust_remote_code = config.get("trust_remote_code", False)
         if not trust_remote_code:
             model_name = config["model_name"]
-            _mn_lower = model_name.lower()
-            if (
-                "nemotron" in _mn_lower
-                and (_mn_lower.startswith("unsloth/") or _mn_lower.startswith("nvidia/"))
-            ):
+            if "nemotron" in model_name.lower():
                 trust_remote_code = True
                 logger.info(
                     "Auto-enabled trust_remote_code for Nemotron model: %s",
diff --git a/studio/backend/core/training/worker.py b/studio/backend/core/training/worker.py
index fa05817697..54667dd7af 100644
--- a/studio/backend/core/training/worker.py
+++ b/studio/backend/core/training/worker.py
@@ -405,11 +405,7 @@ def run_training_process(
     # (Qwen3.5, Gemma 4, etc.) are native and do NOT need it — enabling it
     # bypasses the compiler (disabling fused CE).
     _lowered = model_name.lower()
-    if (
-        "nemotron" in _lowered
-        and (_lowered.startswith("unsloth/") or _lowered.startswith("nvidia/"))
-        and not config.get("trust_remote_code", False)
-    ):
+    if "nemotron" in _lowered and not config.get("trust_remote_code", False):
         config["trust_remote_code"] = True
         logger.info(
             "Auto-enabled trust_remote_code for Nemotron model: %s",
diff --git a/studio/backend/tests/test_transformers_version.py b/studio/backend/tests/test_transformers_version.py
index 609a154a9a..c031c2fea3 100644
--- a/studio/backend/tests/test_transformers_version.py
+++ b/studio/backend/tests/test_transformers_version.py
@@ -32,9 +32,8 @@ from utils.transformers_version import (
     _resolve_base_model,
     _check_tokenizer_config_needs_v5,
     _check_config_needs_550,
-    _get_config_json,
     _tokenizer_class_cache,
-    _config_json_cache,
+    _config_needs_550_cache,
     needs_transformers_5,
     get_transformers_tier,
 )
@@ -203,7 +202,7 @@ class TestCheckConfigNeeds550:
     """Tests for _check_config_needs_550() local config.json checks."""
 
     def setup_method(self):
-        _config_json_cache.clear()
+        _config_needs_550_cache.clear()
 
     def test_gemma4_architecture(self, tmp_path: Path):
         """config.json with Gemma4ForConditionalGeneration should return True."""
@@ -243,8 +242,8 @@ class TestCheckConfigNeeds550:
 
         key = str(tmp_path)
         _check_config_needs_550(key)
-        assert key in _config_json_cache
-        assert _config_json_cache[key] is not None
+        assert key in _config_needs_550_cache
+        assert _config_needs_550_cache[key] is True
 
     def test_local_file_skips_network(self, tmp_path: Path):
         """When local config.json exists, no network request should be made."""
@@ -266,7 +265,7 @@ class TestGetTransformersTier:
 
     def setup_method(self):
         _tokenizer_class_cache.clear()
-        _config_json_cache.clear()
+        _config_needs_550_cache.clear()
 
     def test_gemma4_substring_returns_550(self):
         assert get_transformers_tier("google/gemma-4-E2B-it") == "550"
@@ -318,20 +317,6 @@ class TestGetTransformersTier:
         # This shouldn't happen in practice, but verifies priority
         assert get_transformers_tier("gemma-4-model") == "550"
 
-    def test_config_json_model_type_530(self, tmp_path: Path):
-        """Local checkpoint with qwen3_moe model_type → 530."""
-        cfg = {"model_type": "qwen3_moe", "architectures": ["Qwen3MoeForCausalLM"]}
-        (tmp_path / "config.json").write_text(json.dumps(cfg))
-
-        assert get_transformers_tier(str(tmp_path)) == "530"
-
-    def test_config_json_model_type_glm4_moe(self, tmp_path: Path):
-        """Local checkpoint with glm4_moe model_type → 530."""
-        cfg = {"model_type": "glm4_moe", "architectures": ["Glm4MoeForCausalLM"]}
-        (tmp_path / "config.json").write_text(json.dumps(cfg))
-
-        assert get_transformers_tier(str(tmp_path)) == "530"
-
     def test_needs_transformers_5_compat(self):
         """needs_transformers_5 should return True for both 530 and 550 models."""
         assert needs_transformers_5("google/gemma-4-E2B-it") is True
diff --git a/studio/backend/utils/transformers_version.py b/studio/backend/utils/transformers_version.py
index 308f416940..7ba20a366b 100644
--- a/studio/backend/utils/transformers_version.py
+++ b/studio/backend/utils/transformers_version.py
@@ -60,25 +60,14 @@ TRANSFORMERS_550_MODEL_SUBSTRINGS: tuple[str, ...] = (
     "gemma4",  # Gemma-4 alternate naming
 )
 
-# Architecture classes that require transformers 5.5.0.
+# Architecture classes / model_type values that require transformers 5.5.0.
+# Checked via config.json (local or HuggingFace).
 _TRANSFORMERS_550_ARCHITECTURES: set[str] = {
     "Gemma4ForConditionalGeneration",
 }
-
-# model_type values (from config.json) → tier mapping.
 _TRANSFORMERS_550_MODEL_TYPES: set[str] = {
     "gemma4",
 }
-_TRANSFORMERS_530_MODEL_TYPES: set[str] = {
-    "qwen3_moe",
-    "qwen3_5_moe",
-    "qwen3_vl_moe",
-    "qwen3_next",
-    "deepseek_v3_moe",
-    "glm4_moe",
-    "glm4_moe_lite",
-    "ministral",
-}
 
 # Tokenizer classes that only exist in transformers>=5.x
 _TRANSFORMERS_5_TOKENIZER_CLASSES: set[str] = {
@@ -88,8 +77,8 @@ _TRANSFORMERS_5_TOKENIZER_CLASSES: set[str] = {
 # Cache for dynamic tokenizer_config.json lookups to avoid repeated fetches
 _tokenizer_class_cache: dict[str, bool] = {}
 
-# Cache for config.json lookups (returns the parsed dict or None)
-_config_json_cache: dict[str, dict | None] = {}
+# Cache for dynamic config.json lookups (architecture/model_type checks)
+_config_needs_550_cache: dict[str, bool] = {}
 
 # Versions
 TRANSFORMERS_550_VERSION = "5.5.0"
@@ -230,28 +219,43 @@ def _check_tokenizer_config_needs_v5(model_name: str) -> bool:
         return False
 
 
-_SENTINEL = object()  # distinguishes "not cached" from "cached as None"
+def _check_config_needs_550(model_name: str) -> bool:
+    """Check ``config.json`` for architectures or model_type that require
+    transformers 5.5.0 (e.g. Gemma 4).
 
-
-def _get_config_json(model_name: str) -> dict | None:
-    """Read and cache ``config.json`` for *model_name*.
-
-    Checks local path first, then fetches from HuggingFace.
-    Returns the parsed dict, or ``None`` on any error (fail-open).
-    The result is cached in ``_config_json_cache``.
+    Checks locally first, then falls back to fetching from HuggingFace.
+    Results are cached in ``_config_needs_550_cache``.
+    Returns False on any error (fail-open to lower tier).
     """
-    cached = _config_json_cache.get(model_name, _SENTINEL)
-    if cached is not _SENTINEL:
-        return cached
+    if model_name in _config_needs_550_cache:
+        return _config_needs_550_cache[model_name]
+
+    def _check_cfg(cfg: dict) -> bool:
+        archs = cfg.get("architectures", [])
+        if any(a in _TRANSFORMERS_550_ARCHITECTURES for a in archs):
+            return True
+        if cfg.get("model_type") in _TRANSFORMERS_550_MODEL_TYPES:
+            return True
+        return False
 
     # --- Check local config.json first ------------------------------------
-    local_cfg = Path(model_name) / "config.json"
+    local_path = Path(model_name)
+    local_cfg = local_path / "config.json"
     if local_cfg.is_file():
         try:
             with open(local_cfg) as f:
                 cfg = json.load(f)
-            _config_json_cache[model_name] = cfg
-            return cfg
+            result = _check_cfg(cfg)
+            if result:
+                logger.info(
+                    "Local config.json check: %s needs transformers 5.5.0 "
+                    "(architectures=%s, model_type=%s)",
+                    model_name,
+                    cfg.get("architectures", []),
+                    cfg.get("model_type"),
+                )
+            _config_needs_550_cache[model_name] = result
+            return result
         except Exception as exc:
             logger.debug("Could not read %s: %s", local_cfg, exc)
 
@@ -263,26 +267,21 @@ def _get_config_json(model_name: str) -> dict | None:
         req = urllib.request.Request(url, headers = {"User-Agent": "unsloth-studio"})
         with urllib.request.urlopen(req, timeout = 10) as resp:
             cfg = json.loads(resp.read().decode())
-        _config_json_cache[model_name] = cfg
-        return cfg
+        result = _check_cfg(cfg)
+        if result:
+            logger.info(
+                "Dynamic config.json check: %s needs transformers 5.5.0 "
+                "(architectures=%s, model_type=%s)",
+                model_name,
+                cfg.get("architectures", []),
+                cfg.get("model_type"),
+            )
+        _config_needs_550_cache[model_name] = result
+        return result
     except Exception as exc:
-        logger.debug(
-            "Could not fetch config.json for '%s': %s", model_name, exc
-        )
-        _config_json_cache[model_name] = None
-        return None
-
-
-def _check_config_needs_550(model_name: str) -> bool:
-    """Check ``config.json`` for architectures or model_type that require
-    transformers 5.5.0.  Uses the shared ``_get_config_json`` cache."""
-    cfg = _get_config_json(model_name)
-    if cfg is None:
+        logger.debug("Could not fetch config.json for '%s': %s", model_name, exc)
+        _config_needs_550_cache[model_name] = False
         return False
-    archs = cfg.get("architectures", [])
-    if any(a in _TRANSFORMERS_550_ARCHITECTURES for a in archs):
-        return True
-    return cfg.get("model_type") in _TRANSFORMERS_550_MODEL_TYPES
 
 
 def get_transformers_tier(model_name: str) -> str:
@@ -292,35 +291,19 @@ def get_transformers_tier(model_name: str) -> str:
     ``"530"`` for models needing transformers 5.3.0 (e.g. Ministral-3, Qwen3 MoE),
     or ``"default"`` for everything else (4.57.x).
 
-    Fast path: substring checks (no I/O) for both tiers run first.
-    Slow path: single config.json fetch (cached) checks model_type for
-    both tiers, then tokenizer_config.json as final fallback.
+    The 5.5.0 check runs first, then 5.3.0.
     """
     lowered = model_name.lower()
 
-    # --- Fast substring checks (no I/O) -----------------------------------
+    # --- Check 5.5.0 first ------------------------------------------------
     if any(sub in lowered for sub in TRANSFORMERS_550_MODEL_SUBSTRINGS):
         return "550"
+    if _check_config_needs_550(model_name):
+        return "550"
+
+    # --- Check 5.3.0 ------------------------------------------------------
     if any(sub in lowered for sub in TRANSFORMERS_5_MODEL_SUBSTRINGS):
         return "530"
-
-    # --- config.json model_type / architecture check (single fetch) -------
-    cfg = _get_config_json(model_name)
-    if cfg is not None:
-        model_type = cfg.get("model_type", "")
-        archs = cfg.get("architectures", [])
-
-        # Check 5.5.0 first
-        if model_type in _TRANSFORMERS_550_MODEL_TYPES:
-            return "550"
-        if any(a in _TRANSFORMERS_550_ARCHITECTURES for a in archs):
-            return "550"
-
-        # Check 5.3.0
-        if model_type in _TRANSFORMERS_530_MODEL_TYPES:
-            return "530"
-
-    # --- Final fallback: tokenizer_config.json for 5.3.0 ------------------
     if _check_tokenizer_config_needs_v5(model_name):
         return "530"
 

From c77294bdc9662e7e5e1c5a289f85cca95b289acc Mon Sep 17 00:00:00 2001
From: Roland Tannous <rolandtannous@gravityq.ai>
Date: Mon, 6 Apr 2026 20:27:15 +0000
Subject: [PATCH 14/22] add unsloth/nvidia namespace guard to Nemotron
 trust_remote_code auto-enable

---
 studio/backend/core/inference/worker.py | 6 +++++-
 studio/backend/core/training/worker.py  | 6 +++++-
 2 files changed, 10 insertions(+), 2 deletions(-)

diff --git a/studio/backend/core/inference/worker.py b/studio/backend/core/inference/worker.py
index 1010e56dac..506e631c51 100644
--- a/studio/backend/core/inference/worker.py
+++ b/studio/backend/core/inference/worker.py
@@ -328,7 +328,11 @@ def _handle_load(backend, config: dict, resp_queue: Any) -> None:
         trust_remote_code = config.get("trust_remote_code", False)
         if not trust_remote_code:
             model_name = config["model_name"]
-            if "nemotron" in model_name.lower():
+            _mn_lower = model_name.lower()
+            if (
+                "nemotron" in _mn_lower
+                and (_mn_lower.startswith("unsloth/") or _mn_lower.startswith("nvidia/"))
+            ):
                 trust_remote_code = True
                 logger.info(
                     "Auto-enabled trust_remote_code for Nemotron model: %s",
diff --git a/studio/backend/core/training/worker.py b/studio/backend/core/training/worker.py
index 54667dd7af..fa05817697 100644
--- a/studio/backend/core/training/worker.py
+++ b/studio/backend/core/training/worker.py
@@ -405,7 +405,11 @@ def run_training_process(
     # (Qwen3.5, Gemma 4, etc.) are native and do NOT need it — enabling it
     # bypasses the compiler (disabling fused CE).
     _lowered = model_name.lower()
-    if "nemotron" in _lowered and not config.get("trust_remote_code", False):
+    if (
+        "nemotron" in _lowered
+        and (_lowered.startswith("unsloth/") or _lowered.startswith("nvidia/"))
+        and not config.get("trust_remote_code", False)
+    ):
         config["trust_remote_code"] = True
         logger.info(
             "Auto-enabled trust_remote_code for Nemotron model: %s",

From da68764ef1fdf1dfb41565742c94e53249da3470 Mon Sep 17 00:00:00 2001
From: "pre-commit-ci[bot]"
 <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Date: Mon, 6 Apr 2026 20:27:25 +0000
Subject: [PATCH 15/22] [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci
---
 studio/backend/core/inference/worker.py | 5 ++---
 1 file changed, 2 insertions(+), 3 deletions(-)

diff --git a/studio/backend/core/inference/worker.py b/studio/backend/core/inference/worker.py
index 506e631c51..db504bac3a 100644
--- a/studio/backend/core/inference/worker.py
+++ b/studio/backend/core/inference/worker.py
@@ -329,9 +329,8 @@ def _handle_load(backend, config: dict, resp_queue: Any) -> None:
         if not trust_remote_code:
             model_name = config["model_name"]
             _mn_lower = model_name.lower()
-            if (
-                "nemotron" in _mn_lower
-                and (_mn_lower.startswith("unsloth/") or _mn_lower.startswith("nvidia/"))
+            if "nemotron" in _mn_lower and (
+                _mn_lower.startswith("unsloth/") or _mn_lower.startswith("nvidia/")
             ):
                 trust_remote_code = True
                 logger.info(

From 5375c2628a6e2f3fd3ed079afbc5c70e19d1fb64 Mon Sep 17 00:00:00 2001
From: Roland Tannous <rolandtannous@gravityq.ai>
Date: Mon, 6 Apr 2026 20:31:06 +0000
Subject: [PATCH 16/22] reorder tier checks: all substring matches before
 config.json fetches

---
 studio/backend/utils/transformers_version.py | 10 +++++-----
 1 file changed, 5 insertions(+), 5 deletions(-)

diff --git a/studio/backend/utils/transformers_version.py b/studio/backend/utils/transformers_version.py
index 7ba20a366b..8403d9cc2f 100644
--- a/studio/backend/utils/transformers_version.py
+++ b/studio/backend/utils/transformers_version.py
@@ -295,15 +295,15 @@ def get_transformers_tier(model_name: str) -> str:
     """
     lowered = model_name.lower()
 
-    # --- Check 5.5.0 first ------------------------------------------------
+    # --- Fast substring checks (no I/O) ------------------------------------
     if any(sub in lowered for sub in TRANSFORMERS_550_MODEL_SUBSTRINGS):
         return "550"
-    if _check_config_needs_550(model_name):
-        return "550"
-
-    # --- Check 5.3.0 ------------------------------------------------------
     if any(sub in lowered for sub in TRANSFORMERS_5_MODEL_SUBSTRINGS):
         return "530"
+
+    # --- Slow config fallbacks (local file first, then network) -----------
+    if _check_config_needs_550(model_name):
+        return "550"
     if _check_tokenizer_config_needs_v5(model_name):
         return "530"
 

From c512c36cf01cbdf840d2c78ce8bf632d02281be9 Mon Sep 17 00:00:00 2001
From: Roland Tannous <rolandtannous@gravityq.ai>
Date: Mon, 6 Apr 2026 22:35:50 +0000
Subject: [PATCH 17/22] extract shared activate_transformers_for_subprocess
 into transformers_version.py

---
 studio/backend/core/export/worker.py         | 16 ++------
 studio/backend/core/inference/worker.py      | 41 ++-----------------
 studio/backend/core/training/worker.py       | 41 ++-----------------
 studio/backend/utils/transformers_version.py | 43 ++++++++++++++++++++
 4 files changed, 52 insertions(+), 89 deletions(-)

diff --git a/studio/backend/core/export/worker.py b/studio/backend/core/export/worker.py
index 26c3fa5455..c38c925476 100644
--- a/studio/backend/core/export/worker.py
+++ b/studio/backend/core/export/worker.py
@@ -31,8 +31,11 @@ logger = get_logger(__name__)
 
 def _activate_transformers_version(model_name: str) -> None:
     """Activate the correct transformers version BEFORE any ML imports."""
+    # Ensure backend is on path for utils imports
+    backend_path = str(Path(__file__).resolve().parent.parent.parent)
     if backend_path not in sys.path:
         sys.path.insert(0, backend_path)
+
     from utils.transformers_version import activate_transformers_for_subprocess
 
     activate_transformers_for_subprocess(model_name)
@@ -53,19 +56,6 @@ def _handle_load(backend, cmd: dict, resp_queue: Any) -> None:
     load_in_4bit = cmd.get("load_in_4bit", True)
     trust_remote_code = cmd.get("trust_remote_code", False)
 
-    # Auto-enable trust_remote_code for NemotronH/Nano models.
-    if not trust_remote_code:
-        _NEMOTRON_TRUST_SUBSTRINGS = ("nemotron_h", "nemotron-h", "nemotron-3-nano")
-        _cp_lower = checkpoint_path.lower()
-        if any(sub in _cp_lower for sub in _NEMOTRON_TRUST_SUBSTRINGS) and (
-            _cp_lower.startswith("unsloth/") or _cp_lower.startswith("nvidia/")
-        ):
-            trust_remote_code = True
-            logger.info(
-                "Auto-enabled trust_remote_code for Nemotron model: %s",
-                checkpoint_path,
-            )
-
     try:
         _send_response(
             resp_queue,
diff --git a/studio/backend/core/inference/worker.py b/studio/backend/core/inference/worker.py
index db504bac3a..85293162f5 100644
--- a/studio/backend/core/inference/worker.py
+++ b/studio/backend/core/inference/worker.py
@@ -34,50 +34,15 @@ from utils.hardware import apply_gpu_ids
 
 
 def _activate_transformers_version(model_name: str) -> None:
-    """Activate the correct transformers version BEFORE any ML imports.
-
-    Uses get_transformers_tier() to decide between .venv_t5_550/ (5.5.0),
-    .venv_t5_530/ (5.3.0), or the default 4.57.x.
-    """
+    """Activate the correct transformers version BEFORE any ML imports."""
     # Ensure backend is on path for utils imports
     backend_path = str(Path(__file__).resolve().parent.parent.parent)
     if backend_path not in sys.path:
         sys.path.insert(0, backend_path)
 
-    from utils.transformers_version import (
-        get_transformers_tier,
-        _resolve_base_model,
-        _ensure_venv_t5_530_exists,
-        _ensure_venv_t5_550_exists,
-        _VENV_T5_530_DIR,
-        _VENV_T5_550_DIR,
-    )
+    from utils.transformers_version import activate_transformers_for_subprocess
 
-    resolved = _resolve_base_model(model_name)
-    tier = get_transformers_tier(resolved)
-
-    if tier == "550":
-        if not _ensure_venv_t5_550_exists():
-            raise RuntimeError(
-                f"Cannot activate transformers 5.5.0: .venv_t5_550 missing at {_VENV_T5_550_DIR}"
-            )
-        if _VENV_T5_550_DIR not in sys.path:
-            sys.path.insert(0, _VENV_T5_550_DIR)
-        logger.info("Activated transformers 5.5.0 from %s", _VENV_T5_550_DIR)
-        _pp = os.environ.get("PYTHONPATH", "")
-        os.environ["PYTHONPATH"] = _VENV_T5_550_DIR + (os.pathsep + _pp if _pp else "")
-    elif tier == "530":
-        if not _ensure_venv_t5_530_exists():
-            raise RuntimeError(
-                f"Cannot activate transformers 5.3.0: .venv_t5_530 missing at {_VENV_T5_530_DIR}"
-            )
-        if _VENV_T5_530_DIR not in sys.path:
-            sys.path.insert(0, _VENV_T5_530_DIR)
-        logger.info("Activated transformers 5.3.0 from %s", _VENV_T5_530_DIR)
-        _pp = os.environ.get("PYTHONPATH", "")
-        os.environ["PYTHONPATH"] = _VENV_T5_530_DIR + (os.pathsep + _pp if _pp else "")
-    else:
-        logger.info("Using default transformers (4.57.x) for %s", model_name)
+    activate_transformers_for_subprocess(model_name)
 
 
 def _decode_image(image_base64: str):
diff --git a/studio/backend/core/training/worker.py b/studio/backend/core/training/worker.py
index fa05817697..bfff599001 100644
--- a/studio/backend/core/training/worker.py
+++ b/studio/backend/core/training/worker.py
@@ -306,50 +306,15 @@ def _ensure_mamba_ssm(event_queue: Any, model_name: str) -> None:
 
 
 def _activate_transformers_version(model_name: str) -> None:
-    """Activate the correct transformers version BEFORE any ML imports.
-
-    Uses get_transformers_tier() to decide between .venv_t5_550/ (5.5.0),
-    .venv_t5_530/ (5.3.0), or the default 4.57.x.
-    """
+    """Activate the correct transformers version BEFORE any ML imports."""
     # Ensure backend is on path for utils imports
     backend_path = str(Path(__file__).resolve().parent.parent.parent)
     if backend_path not in sys.path:
         sys.path.insert(0, backend_path)
 
-    from utils.transformers_version import (
-        get_transformers_tier,
-        _resolve_base_model,
-        _ensure_venv_t5_530_exists,
-        _ensure_venv_t5_550_exists,
-        _VENV_T5_530_DIR,
-        _VENV_T5_550_DIR,
-    )
+    from utils.transformers_version import activate_transformers_for_subprocess
 
-    resolved = _resolve_base_model(model_name)
-    tier = get_transformers_tier(resolved)
-
-    if tier == "550":
-        if not _ensure_venv_t5_550_exists():
-            raise RuntimeError(
-                f"Cannot activate transformers 5.5.0: .venv_t5_550 missing at {_VENV_T5_550_DIR}"
-            )
-        if _VENV_T5_550_DIR not in sys.path:
-            sys.path.insert(0, _VENV_T5_550_DIR)
-        logger.info("Activated transformers 5.5.0 from %s", _VENV_T5_550_DIR)
-        _pp = os.environ.get("PYTHONPATH", "")
-        os.environ["PYTHONPATH"] = _VENV_T5_550_DIR + (os.pathsep + _pp if _pp else "")
-    elif tier == "530":
-        if not _ensure_venv_t5_530_exists():
-            raise RuntimeError(
-                f"Cannot activate transformers 5.3.0: .venv_t5_530 missing at {_VENV_T5_530_DIR}"
-            )
-        if _VENV_T5_530_DIR not in sys.path:
-            sys.path.insert(0, _VENV_T5_530_DIR)
-        logger.info("Activated transformers 5.3.0 from %s", _VENV_T5_530_DIR)
-        _pp = os.environ.get("PYTHONPATH", "")
-        os.environ["PYTHONPATH"] = _VENV_T5_530_DIR + (os.pathsep + _pp if _pp else "")
-    else:
-        logger.info("Using default transformers (4.57.x) for %s", model_name)
+    activate_transformers_for_subprocess(model_name)
 
 
 def run_training_process(
diff --git a/studio/backend/utils/transformers_version.py b/studio/backend/utils/transformers_version.py
index 8403d9cc2f..0517a5f378 100644
--- a/studio/backend/utils/transformers_version.py
+++ b/studio/backend/utils/transformers_version.py
@@ -94,6 +94,49 @@ _VENV_T5_550_DIR = str(Path.home() / ".unsloth" / "studio" / ".venv_t5_550")
 _VENV_T5_DIR = _VENV_T5_550_DIR
 
 
+def activate_transformers_for_subprocess(model_name: str) -> None:
+    """Activate the correct transformers version in a subprocess worker.
+
+    Call this BEFORE any ML imports. Resolves LoRA adapters to their base
+    model, determines the required tier, and prepends the appropriate
+    ``.venv_t5_*`` directory to ``sys.path``.  Also propagates the path
+    via ``PYTHONPATH`` for child processes (e.g. GGUF converter).
+
+    Used by training, inference, and export workers.
+    """
+    resolved = _resolve_base_model(model_name)
+    tier = get_transformers_tier(resolved)
+
+    if tier == "550":
+        if not _ensure_venv_t5_550_exists():
+            raise RuntimeError(
+                f"Cannot activate transformers 5.5.0: "
+                f".venv_t5_550 missing at {_VENV_T5_550_DIR}"
+            )
+        if _VENV_T5_550_DIR not in sys.path:
+            sys.path.insert(0, _VENV_T5_550_DIR)
+        logger.info("Activated transformers 5.5.0 from %s", _VENV_T5_550_DIR)
+        _pp = os.environ.get("PYTHONPATH", "")
+        os.environ["PYTHONPATH"] = (
+            _VENV_T5_550_DIR + (os.pathsep + _pp if _pp else "")
+        )
+    elif tier == "530":
+        if not _ensure_venv_t5_530_exists():
+            raise RuntimeError(
+                f"Cannot activate transformers 5.3.0: "
+                f".venv_t5_530 missing at {_VENV_T5_530_DIR}"
+            )
+        if _VENV_T5_530_DIR not in sys.path:
+            sys.path.insert(0, _VENV_T5_530_DIR)
+        logger.info("Activated transformers 5.3.0 from %s", _VENV_T5_530_DIR)
+        _pp = os.environ.get("PYTHONPATH", "")
+        os.environ["PYTHONPATH"] = (
+            _VENV_T5_530_DIR + (os.pathsep + _pp if _pp else "")
+        )
+    else:
+        logger.info("Using default transformers (4.57.x) for %s", model_name)
+
+
 def _resolve_base_model(model_name: str) -> str:
     """If *model_name* points to a LoRA adapter, return its base model.
 

From f9923b398c7be08ba3f32ebad674bb970db22eae Mon Sep 17 00:00:00 2001
From: Roland Tannous <rolandtannous@gravityq.ai>
Date: Mon, 6 Apr 2026 22:45:22 +0000
Subject: [PATCH 18/22] narrow Nemotron trust_remote_code to
 nemotron_h/nemotron-3-nano, add to export worker

---
 studio/backend/core/export/worker.py    | 13 +++++++++++++
 studio/backend/core/inference/worker.py |  6 ++++--
 studio/backend/core/training/worker.py  |  6 ++++--
 3 files changed, 21 insertions(+), 4 deletions(-)

diff --git a/studio/backend/core/export/worker.py b/studio/backend/core/export/worker.py
index c38c925476..3f3dc955fa 100644
--- a/studio/backend/core/export/worker.py
+++ b/studio/backend/core/export/worker.py
@@ -56,6 +56,19 @@ def _handle_load(backend, cmd: dict, resp_queue: Any) -> None:
     load_in_4bit = cmd.get("load_in_4bit", True)
     trust_remote_code = cmd.get("trust_remote_code", False)
 
+    # Auto-enable trust_remote_code for NemotronH/Nano models.
+    if not trust_remote_code:
+        _NEMOTRON_TRUST_SUBSTRINGS = ("nemotron_h", "nemotron-h", "nemotron-3-nano")
+        _cp_lower = checkpoint_path.lower()
+        if any(sub in _cp_lower for sub in _NEMOTRON_TRUST_SUBSTRINGS) and (
+            _cp_lower.startswith("unsloth/") or _cp_lower.startswith("nvidia/")
+        ):
+            trust_remote_code = True
+            logger.info(
+                "Auto-enabled trust_remote_code for Nemotron model: %s",
+                checkpoint_path,
+            )
+
     try:
         _send_response(
             resp_queue,
diff --git a/studio/backend/core/inference/worker.py b/studio/backend/core/inference/worker.py
index 85293162f5..fbcce276ba 100644
--- a/studio/backend/core/inference/worker.py
+++ b/studio/backend/core/inference/worker.py
@@ -287,14 +287,16 @@ def _handle_load(backend, config: dict, resp_queue: Any) -> None:
                 except Exception as e:
                     logger.warning("Could not read adapter_config.json: %s", e)
 
-        # Auto-enable trust_remote_code for Nemotron models only.
+        # Auto-enable trust_remote_code for NemotronH/Nano models only.
         # NemotronH has config parsing bugs requiring trust_remote_code=True.
         # Other transformers 5.x models are native and do NOT need it.
+        # NOTE: Must NOT match Llama-Nemotron (standard Llama architecture).
+        _NEMOTRON_TRUST_SUBSTRINGS = ("nemotron_h", "nemotron-h", "nemotron-3-nano")
         trust_remote_code = config.get("trust_remote_code", False)
         if not trust_remote_code:
             model_name = config["model_name"]
             _mn_lower = model_name.lower()
-            if "nemotron" in _mn_lower and (
+            if any(sub in _mn_lower for sub in _NEMOTRON_TRUST_SUBSTRINGS) and (
                 _mn_lower.startswith("unsloth/") or _mn_lower.startswith("nvidia/")
             ):
                 trust_remote_code = True
diff --git a/studio/backend/core/training/worker.py b/studio/backend/core/training/worker.py
index bfff599001..a461972eca 100644
--- a/studio/backend/core/training/worker.py
+++ b/studio/backend/core/training/worker.py
@@ -364,14 +364,16 @@ def run_training_process(
         )
         return
 
-    # ── 1a. Auto-enable trust_remote_code for Nemotron models ──
+    # ── 1a. Auto-enable trust_remote_code for NemotronH/Nano models ──
     # NemotronH has config parsing bugs in transformers that require
     # trust_remote_code=True as a workaround. Other transformers 5.x models
     # (Qwen3.5, Gemma 4, etc.) are native and do NOT need it — enabling it
     # bypasses the compiler (disabling fused CE).
+    # NOTE: Must NOT match Llama-Nemotron (standard Llama architecture).
+    _NEMOTRON_TRUST_SUBSTRINGS = ("nemotron_h", "nemotron-h", "nemotron-3-nano")
     _lowered = model_name.lower()
     if (
-        "nemotron" in _lowered
+        any(sub in _lowered for sub in _NEMOTRON_TRUST_SUBSTRINGS)
         and (_lowered.startswith("unsloth/") or _lowered.startswith("nvidia/"))
         and not config.get("trust_remote_code", False)
     ):

From 44d0d69cb28d015b7ee2a5b21abd6280afb8b5b7 Mon Sep 17 00:00:00 2001
From: Roland Tannous <rolandtannous@gravityq.ai>
Date: Mon, 6 Apr 2026 22:46:14 +0000
Subject: [PATCH 19/22] clean venv_t5 dirs before re-install in setup.sh,
 clarify version alias comment

---
 studio/backend/utils/transformers_version.py | 3 ++-
 studio/setup.sh                              | 2 ++
 2 files changed, 4 insertions(+), 1 deletion(-)

diff --git a/studio/backend/utils/transformers_version.py b/studio/backend/utils/transformers_version.py
index 0517a5f378..62176dc9b4 100644
--- a/studio/backend/utils/transformers_version.py
+++ b/studio/backend/utils/transformers_version.py
@@ -84,7 +84,8 @@ _config_needs_550_cache: dict[str, bool] = {}
 TRANSFORMERS_550_VERSION = "5.5.0"
 TRANSFORMERS_530_VERSION = "5.3.0"
 TRANSFORMERS_DEFAULT_VERSION = "4.57.6"
-# Backwards-compat alias used by other modules
+# Backwards-compat alias — points to 5.5.0 (the highest 5.x tier).
+# Consumers should prefer TRANSFORMERS_530_VERSION / TRANSFORMERS_550_VERSION.
 TRANSFORMERS_5_VERSION = TRANSFORMERS_550_VERSION
 
 # Pre-installed directories — created by setup.sh / setup.ps1
diff --git a/studio/setup.sh b/studio/setup.sh
index 0b9ca8383b..4e6153d0ae 100755
--- a/studio/setup.sh
+++ b/studio/setup.sh
@@ -551,6 +551,7 @@ if [ "$_SKIP_PYTHON_DEPS" = false ]; then
     # Clean up legacy single .venv_t5 directory
     [ -d "$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 "$VENV_T5_530_DIR" --no-deps "huggingface_hub==1.8.0"
@@ -558,6 +559,7 @@ if [ "$_SKIP_PYTHON_DEPS" = false ]; then
     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"

From 940223ea6b14e6f9afb5e1a48edf99119f43d001 Mon Sep 17 00:00:00 2001
From: "pre-commit-ci[bot]"
 <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Date: Mon, 6 Apr 2026 22:46:58 +0000
Subject: [PATCH 20/22] [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci
---
 studio/backend/utils/transformers_version.py | 8 ++------
 1 file changed, 2 insertions(+), 6 deletions(-)

diff --git a/studio/backend/utils/transformers_version.py b/studio/backend/utils/transformers_version.py
index 62176dc9b4..0c13b5455b 100644
--- a/studio/backend/utils/transformers_version.py
+++ b/studio/backend/utils/transformers_version.py
@@ -118,9 +118,7 @@ def activate_transformers_for_subprocess(model_name: str) -> None:
             sys.path.insert(0, _VENV_T5_550_DIR)
         logger.info("Activated transformers 5.5.0 from %s", _VENV_T5_550_DIR)
         _pp = os.environ.get("PYTHONPATH", "")
-        os.environ["PYTHONPATH"] = (
-            _VENV_T5_550_DIR + (os.pathsep + _pp if _pp else "")
-        )
+        os.environ["PYTHONPATH"] = _VENV_T5_550_DIR + (os.pathsep + _pp if _pp else "")
     elif tier == "530":
         if not _ensure_venv_t5_530_exists():
             raise RuntimeError(
@@ -131,9 +129,7 @@ def activate_transformers_for_subprocess(model_name: str) -> None:
             sys.path.insert(0, _VENV_T5_530_DIR)
         logger.info("Activated transformers 5.3.0 from %s", _VENV_T5_530_DIR)
         _pp = os.environ.get("PYTHONPATH", "")
-        os.environ["PYTHONPATH"] = (
-            _VENV_T5_530_DIR + (os.pathsep + _pp if _pp else "")
-        )
+        os.environ["PYTHONPATH"] = _VENV_T5_530_DIR + (os.pathsep + _pp if _pp else "")
     else:
         logger.info("Using default transformers (4.57.x) for %s", model_name)
 

From c26d909265bbe27a0c1fb1f1daf942acead0191c Mon Sep 17 00:00:00 2001
From: Roland Tannous <rolandtannous@gravityq.ai>
Date: Tue, 7 Apr 2026 07:24:47 +0000
Subject: [PATCH 21/22] update

---
 studio/setup.sh | 28 +++++++++++++++++++++-------
 1 file changed, 21 insertions(+), 7 deletions(-)

diff --git a/studio/setup.sh b/studio/setup.sh
index 4e6153d0ae..420076ff3a 100755
--- a/studio/setup.sh
+++ b/studio/setup.sh
@@ -490,15 +490,29 @@ if [ "$_DOCKER_NO_VENV" = true ]; then
     # install_python_stack.py has steps 1-5 commented out for this branch.
     python "$SCRIPT_DIR/install_python_stack.py"
 
-    # Pre-install transformers 5.x into .venv_t5
+    # Pre-install transformers 5.x into .venv_t5_530/ and .venv_t5_550/
     echo ""
     echo "   Pre-installing transformers 5.x for newer model support..."
-    mkdir -p "$VENV_T5_DIR"
-    pip install --target "$VENV_T5_DIR" --no-deps "transformers==5.3.0" 2>/dev/null
-    pip install --target "$VENV_T5_DIR" --no-deps "huggingface_hub==1.7.1" 2>/dev/null
-    pip install --target "$VENV_T5_DIR" --no-deps "hf_xet==1.4.2" 2>/dev/null
-    pip install --target "$VENV_T5_DIR" "tiktoken" 2>/dev/null
-    echo "✅ Transformers 5.x pre-installed to $VENV_T5_DIR/"
+
+    # Clean up legacy single .venv_t5 directory
+    [ -d "$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 "╔══════════════════════════════════════╗"

From b362c60719dffd0db3646587b272ba2c2e44f0ef Mon Sep 17 00:00:00 2001
From: Roland Tannous <rolandtannous@gravityq.ai>
Date: Sun, 5 Apr 2026 05:22:43 +0000
Subject: [PATCH 22/22] fix: derive HF_HUB_CACHE from HF_HOME when set

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.
---
 studio/backend/utils/paths/storage_roots.py | 5 +++--
 1 file changed, 3 insertions(+), 2 deletions(-)

diff --git a/studio/backend/utils/paths/storage_roots.py b/studio/backend/utils/paths/storage_roots.py
index 4841c5d0a3..530aba250e 100644
--- a/studio/backend/utils/paths/storage_roots.py
+++ b/studio/backend/utils/paths/storage_roots.py
@@ -148,10 +148,11 @@ 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"),
     }