From 947aa71eeb4e9e58d9cfd0de690708e40f908d57 Mon Sep 17 00:00:00 2001 From: Roland Tannous Date: Fri, 27 Mar 2026 19:39:37 +0000 Subject: [PATCH 01/25] add Docker support: skip venv, install only missing deps --- studio/install_python_stack.py | 241 +++++++++++++++++++-------------- studio/setup.sh | 41 +++++- unsloth_cli/commands/studio.py | 14 +- 3 files changed, 187 insertions(+), 109 deletions(-) diff --git a/studio/install_python_stack.py b/studio/install_python_stack.py index 3fd1e6af66..9273c70272 100644 --- a/studio/install_python_stack.py +++ b/studio/install_python_stack.py @@ -907,9 +907,9 @@ def install_python_stack() -> int: base_total += 3 _TOTAL = (base_total - 1) if skip_base else base_total - # 1. Try to use uv for faster installs (must happen before pip upgrade - # because uv venvs don't include pip by default) - USE_UV = _bootstrap_uv() + # # 1. Try to use uv for faster installs (must happen before pip upgrade + # # because uv venvs don't include pip by default) + # USE_UV = _bootstrap_uv() # 2. Ensure pip is available (uv venvs created by install.sh don't include pip) _progress("pip bootstrap") @@ -938,17 +938,115 @@ def install_python_stack() -> int: ).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: @@ -1048,96 +1146,39 @@ def install_python_stack() -> int: "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, + # ) if not IS_WINDOWS and not IS_MACOS and not NO_TORCH: _progress("flash-attn") diff --git a/studio/setup.sh b/studio/setup.sh index 3e875eed30..c396c71d9e 100755 --- a/studio/setup.sh +++ b/studio/setup.sh @@ -417,7 +417,7 @@ 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="$STUDIO_HOME/.venv_t5_550" @@ -430,7 +430,12 @@ VENV_T5_550_DIR="$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 [ "$IS_COLAB" = true ]; then # On Colab there is no Studio venv -- install backend deps into system Python. # Strip all version constraints so pip keeps Colab's pre-installed @@ -495,6 +500,38 @@ if [ "$_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 140940209f..ca641bde54 100644 --- a/unsloth_cli/commands/studio.py +++ b/unsloth_cli/commands/studio.py @@ -430,9 +430,11 @@ 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() @@ -478,10 +480,8 @@ def studio_default( ) raise typer.Exit(rc) 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 3a6643609f81f2309b0dd28ef1e2547c3cd5d1e3 Mon Sep 17 00:00:00 2001 From: Roland Tannous Date: Thu, 2 Apr 2026 23:24:47 +0000 Subject: [PATCH 02/25] 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 fc91178d88..a2453dd8d0 100644 --- a/unsloth/models/loader.py +++ b/unsloth/models/loader.py @@ -1559,7 +1559,10 @@ 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 2ee7c2e1e0d1da3992c017e97ff677892c6d8e12 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 03/25] [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 a2453dd8d0..c413de95dc 100644 --- a/unsloth/models/loader.py +++ b/unsloth/models/loader.py @@ -1559,10 +1559,6 @@ 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 a29cba3ecf334f5d31cf4bad2250d9e697c0affb Mon Sep 17 00:00:00 2001 From: Roland Tannous Date: Sun, 5 Apr 2026 04:50:43 +0000 Subject: [PATCH 04/25] 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 c396c71d9e..88982678a5 100755 --- a/studio/setup.sh +++ b/studio/setup.sh @@ -610,6 +610,9 @@ 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="$UNSLOTH_HOME/llama.cpp" @@ -1093,6 +1096,7 @@ else fi } fi # end _SKIP_GGUF_BUILD check +fi # end non-Docker llama.cpp block # ── Footer ── if [ "$_LLAMA_ONLY" = "1" ]; then From 6aca1b403036d8080f46ec298639b5a7bb799780 Mon Sep 17 00:00:00 2001 From: Roland Tannous Date: Mon, 6 Apr 2026 16:25:07 +0000 Subject: [PATCH 05/25] 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 f77b1966c4..1be456d4d4 100644 --- a/studio/backend/core/export/worker.py +++ b/studio/backend/core/export/worker.py @@ -181,11 +181,8 @@ def _setup_log_capture(resp_queue: Any) -> None: 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 60b9e994ab..e9ed9e23a1 100644 --- a/studio/backend/core/training/worker.py +++ b/studio/backend/core/training/worker.py @@ -385,6 +385,7 @@ 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 @@ -396,6 +397,22 @@ 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 3d082aa70d..28ea242208 100644 --- a/studio/setup.ps1 +++ b/studio/setup.ps1 @@ -1750,6 +1750,89 @@ 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 = $prevEAP_t5 +step "transformers" "5.5.0 pre-installed" + } else { step "python" "dependencies up to date" # Restore ErrorActionPreference (was lowered for pip/python section) diff --git a/studio/setup.sh b/studio/setup.sh index 88982678a5..e96e0999ed 100755 --- a/studio/setup.sh +++ b/studio/setup.sh @@ -566,6 +566,29 @@ fi if [ "$_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 348bea59ba03108df9327654eee9d0232b8a2e68 Mon Sep 17 00:00:00 2001 From: Roland Tannous Date: Mon, 6 Apr 2026 17:28:55 +0000 Subject: [PATCH 06/25] 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 e9ed9e23a1..60b9e994ab 100644 --- a/studio/backend/core/training/worker.py +++ b/studio/backend/core/training/worker.py @@ -385,7 +385,6 @@ 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 @@ -397,22 +396,6 @@ 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 c413de95dc..8aa1161b8c 100644 --- a/unsloth/models/loader.py +++ b/unsloth/models/loader.py @@ -1371,7 +1371,9 @@ 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 435b6994be4bda70d6c9d06cdf413052c2b01496 Mon Sep 17 00:00:00 2001 From: Roland Tannous Date: Mon, 6 Apr 2026 17:29:22 +0000 Subject: [PATCH 07/25] 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 8aa1161b8c..c413de95dc 100644 --- a/unsloth/models/loader.py +++ b/unsloth/models/loader.py @@ -1371,9 +1371,7 @@ 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 6d157ee0957376df7a7578b9ac4ec41000b62d3d Mon Sep 17 00:00:00 2001 From: Roland Tannous Date: Mon, 6 Apr 2026 18:20:08 +0000 Subject: [PATCH 08/25] restrict trust_remote_code auto-enable to Nemotron models only --- studio/backend/core/inference/worker.py | 52 +++++++-- studio/backend/core/training/worker.py | 139 ++++++++++++++++++++++-- 2 files changed, 173 insertions(+), 18 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 60b9e994ab..d8967445bb 100644 --- a/studio/backend/core/training/worker.py +++ b/studio/backend/core/training/worker.py @@ -75,6 +75,97 @@ def _model_wants_causal_conv1d(model_name: str) -> bool: ) +def _causal_conv1d_platform_tag() -> str | None: + machine = platform.machine().lower() + if sys.platform.startswith("linux"): + if machine in {"x86_64", "amd64"}: + return "linux_x86_64" + if machine in {"aarch64", "arm64"}: + return "linux_aarch64" + return None + # No prebuilt wheels published for macOS or Windows + return None + + +def _probe_causal_conv1d_env() -> dict[str, str] | None: + try: + probe = _sp.run( + [ + sys.executable, + "-c", + ( + "import json, sys, re, torch; " + "parts = torch.__version__.split('+', 1)[0].split('.')[:2]; " + "minor = re.sub(r'[^0-9].*', '', parts[1]) if len(parts) > 1 else '0'; " + "torch_mm = parts[0] + '.' + minor; " + "print(json.dumps({" + "'python_tag': f'cp{sys.version_info.major}{sys.version_info.minor}', " + "'torch_mm': torch_mm, " + "'cuda_major': str(int(str(torch.version.cuda).split('.', 1)[0])) if torch.version.cuda else '', " + "'cxx11abi': str(torch._C._GLIBCXX_USE_CXX11_ABI).upper()" + "}))" + ), + ], + stdout = _sp.PIPE, + stderr = _sp.PIPE, + text = True, + timeout = 30, + ) + except _sp.TimeoutExpired: + logger.warning("Torch environment probe timed out after 30s") + return None + if probe.returncode != 0: + logger.warning( + "Failed to probe torch environment for causal-conv1d wheel:\n%s", + probe.stdout, + ) + return None + + try: + return json.loads(probe.stdout.strip()) + except json.JSONDecodeError: + logger.warning( + "Failed to parse torch environment probe output: %s", probe.stdout + ) + return None + + +def _direct_wheel_url( + *, + filename_prefix: str, + package_version: str, + release_tag: str, + release_base_url: str, + env: dict[str, str] | None = None, +) -> str | None: + env = env or _probe_causal_conv1d_env() + platform_tag = _causal_conv1d_platform_tag() + if env is None or platform_tag is None or not env.get("cuda_major"): + return None + + filename = ( + f"{filename_prefix}-{package_version}" + f"+cu{env['cuda_major']}torch{env['torch_mm']}" + f"cxx11abi{env['cxx11abi']}-{env['python_tag']}-{env['python_tag']}-{platform_tag}.whl" + ) + return f"{release_base_url}/{release_tag}/{filename}" + + +def _url_exists(url: str) -> bool: + try: + request = urllib.request.Request(url, method = "HEAD") + with urllib.request.urlopen(request, timeout = 10): + return True + except urllib.error.HTTPError as exc: + if exc.code == 404: + return False + logger.warning("Unexpected HTTP error while probing %s: %s", url, exc) + return False + except Exception as exc: + logger.warning("Failed to probe %s: %s", url, exc) + return False + + def _install_package_wheel_first( *, event_queue: Any, @@ -327,15 +418,50 @@ def _ensure_flash_attn_for_long_context(event_queue: Any, max_seq_length: int) - 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( @@ -385,17 +511,14 @@ 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 5c44ead041a11872e5414e5b008a5c2a130b01c3 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 09/25] [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 d8967445bb..55a53814cc 100644 --- a/studio/backend/core/training/worker.py +++ b/studio/backend/core/training/worker.py @@ -517,10 +517,7 @@ 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 639ac277f19793facc6a6d4febfeb1795559c4d8 Mon Sep 17 00:00:00 2001 From: Roland Tannous Date: Mon, 6 Apr 2026 19:40:11 +0000 Subject: [PATCH 10/25] 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 55a53814cc..1d4b9a2ee7 100644 --- a/studio/backend/core/training/worker.py +++ b/studio/backend/core/training/worker.py @@ -517,7 +517,11 @@ 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 17af40f663..0c4bb4020b 100644 --- a/studio/backend/utils/transformers_version.py +++ b/studio/backend/utils/transformers_version.py @@ -67,14 +67,25 @@ TRANSFORMERS_550_MODEL_SUBSTRINGS: tuple[str, ...] = ( "qwen3.6", ) -# 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] = { @@ -84,15 +95,14 @@ _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 @@ -102,45 +112,6 @@ _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. @@ -266,43 +237,28 @@ 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) @@ -314,21 +270,26 @@ 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: @@ -338,19 +299,35 @@ 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 4e93ce0024b8e115e78049a64287051719107023 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 11/25] [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 0c4bb4020b..a9fb9e8086 100644 --- a/studio/backend/utils/transformers_version.py +++ b/studio/backend/utils/transformers_version.py @@ -273,9 +273,7 @@ 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 679b3740a85eee4342c8135ad5f05f025c0d8b68 Mon Sep 17 00:00:00 2001 From: Roland Tannous Date: Mon, 6 Apr 2026 20:26:32 +0000 Subject: [PATCH 12/25] 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 a9fb9e8086..0c4bb4020b 100644 --- a/studio/backend/utils/transformers_version.py +++ b/studio/backend/utils/transformers_version.py @@ -273,7 +273,9 @@ 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 47dc808afe7d2706f753f735604731b1c67871e5 Mon Sep 17 00:00:00 2001 From: Roland Tannous Date: Mon, 6 Apr 2026 20:26:32 +0000 Subject: [PATCH 13/25] 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 1d4b9a2ee7..55a53814cc 100644 --- a/studio/backend/core/training/worker.py +++ b/studio/backend/core/training/worker.py @@ -517,11 +517,7 @@ 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 0c4bb4020b..90c9b7e981 100644 --- a/studio/backend/utils/transformers_version.py +++ b/studio/backend/utils/transformers_version.py @@ -67,25 +67,14 @@ TRANSFORMERS_550_MODEL_SUBSTRINGS: tuple[str, ...] = ( "qwen3.6", ) -# 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] = { @@ -95,8 +84,8 @@ _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" @@ -237,28 +226,43 @@ 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) @@ -270,26 +274,21 @@ 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: @@ -299,35 +298,19 @@ 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 3f62c7bfbd41d061ae04649fc21f62742fab1b64 Mon Sep 17 00:00:00 2001 From: Roland Tannous Date: Mon, 6 Apr 2026 20:27:15 +0000 Subject: [PATCH 14/25] 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 55a53814cc..1d4b9a2ee7 100644 --- a/studio/backend/core/training/worker.py +++ b/studio/backend/core/training/worker.py @@ -517,7 +517,11 @@ 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 b36558335a4fa7dcb5e7dc75ed08bfa5208a28fd 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 15/25] [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 d9c5725481f35aa25f647a059b137bb3d8becf56 Mon Sep 17 00:00:00 2001 From: Roland Tannous Date: Mon, 6 Apr 2026 20:31:06 +0000 Subject: [PATCH 16/25] 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 90c9b7e981..5b7648ae52 100644 --- a/studio/backend/utils/transformers_version.py +++ b/studio/backend/utils/transformers_version.py @@ -302,15 +302,15 @@ 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 60fa0e61a6cee6b41bc8ed7ae35d6de69458e462 Mon Sep 17 00:00:00 2001 From: Roland Tannous Date: Mon, 6 Apr 2026 22:35:50 +0000 Subject: [PATCH 17/25] 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 1be456d4d4..80b05aca79 100644 --- a/studio/backend/core/export/worker.py +++ b/studio/backend/core/export/worker.py @@ -181,8 +181,11 @@ def _setup_log_capture(resp_queue: Any) -> None: 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) @@ -203,19 +206,6 @@ 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 1d4b9a2ee7..b621def9e3 100644 --- a/studio/backend/core/training/worker.py +++ b/studio/backend/core/training/worker.py @@ -418,50 +418,15 @@ def _ensure_flash_attn_for_long_context(event_queue: Any, max_seq_length: int) - 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 5b7648ae52..9f0e87c611 100644 --- a/studio/backend/utils/transformers_version.py +++ b/studio/backend/utils/transformers_version.py @@ -101,6 +101,49 @@ _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 b122a570993688ff84723eff7cd7dffed5c868ed Mon Sep 17 00:00:00 2001 From: Roland Tannous Date: Mon, 6 Apr 2026 22:45:22 +0000 Subject: [PATCH 18/25] 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 80b05aca79..f77b1966c4 100644 --- a/studio/backend/core/export/worker.py +++ b/studio/backend/core/export/worker.py @@ -206,6 +206,19 @@ 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 b621def9e3..a583fc2626 100644 --- a/studio/backend/core/training/worker.py +++ b/studio/backend/core/training/worker.py @@ -476,14 +476,16 @@ 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 9e91b427d74d31e937da10a721b86978c68c8a81 Mon Sep 17 00:00:00 2001 From: Roland Tannous Date: Mon, 6 Apr 2026 22:46:14 +0000 Subject: [PATCH 19/25] 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 9f0e87c611..08771af298 100644 --- a/studio/backend/utils/transformers_version.py +++ b/studio/backend/utils/transformers_version.py @@ -91,7 +91,8 @@ _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 e96e0999ed..17f4ecad51 100755 --- a/studio/setup.sh +++ b/studio/setup.sh @@ -576,6 +576,7 @@ 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 "$VENV_T5_530_DIR" --no-deps "huggingface_hub==1.8.0" @@ -583,6 +584,7 @@ if [ "$_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 c395d7699ed804bd94bd6ba882b51415e0c27351 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 20/25] [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 08771af298..17af40f663 100644 --- a/studio/backend/utils/transformers_version.py +++ b/studio/backend/utils/transformers_version.py @@ -125,9 +125,7 @@ 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( @@ -138,9 +136,7 @@ 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 9161a8c2f733b2766c712096ae59fe16d53deddb Mon Sep 17 00:00:00 2001 From: Roland Tannous Date: Tue, 7 Apr 2026 07:24:47 +0000 Subject: [PATCH 21/25] update --- studio/setup.sh | 28 +++++++++++++++++++++------- 1 file changed, 21 insertions(+), 7 deletions(-) diff --git a/studio/setup.sh b/studio/setup.sh index 17f4ecad51..e8990fb489 100755 --- a/studio/setup.sh +++ b/studio/setup.sh @@ -515,15 +515,29 @@ 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 e7574903b1d5e3479f43f94e8f0ef4db624760a5 Mon Sep 17 00:00:00 2001 From: Roland Tannous Date: Sun, 5 Apr 2026 05:22:43 +0000 Subject: [PATCH 22/25] 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 b52609b06b..396c440ac4 100644 --- a/studio/backend/utils/paths/storage_roots.py +++ b/studio/backend/utils/paths/storage_roots.py @@ -193,10 +193,11 @@ def _setup_cache_env() -> None: os.environ.get("XDG_CACHE_HOME", Path.home() / ".cache") ).expanduser() hf_default = xdg_cache / "huggingface" + hf_home = Path(os.environ.get("HF_HOME", str(hf_default))) defaults: dict[str, str] = { "HF_HOME": str(hf_default), - "HF_HUB_CACHE": str(hf_default / "hub"), - "HF_XET_CACHE": str(hf_default / "xet"), + "HF_HUB_CACHE": str(hf_home / "hub"), + "HF_XET_CACHE": str(hf_home / "xet"), "UV_CACHE_DIR": str(root / "uv"), "VLLM_CACHE_ROOT": str(root / "vllm"), } From 4b248abc9bc4e031a1c13f3c50d97e7898a31750 Mon Sep 17 00:00:00 2001 From: Roland Tannous Date: Thu, 16 Apr 2026 00:48:14 +0400 Subject: [PATCH 23/25] spark studio v0.13.6 2026.4.5 --- .worktreeinclude | 1 + 1 file changed, 1 insertion(+) create mode 100644 .worktreeinclude diff --git a/.worktreeinclude b/.worktreeinclude new file mode 100644 index 0000000000..ceb2b988dc --- /dev/null +++ b/.worktreeinclude @@ -0,0 +1 @@ +CLAUDE.md From 77da18d649914d8021e13cd87b8838d824de4bb6 Mon Sep 17 00:00:00 2001 From: Roland Tannous Date: Wed, 15 Apr 2026 21:03:05 +0000 Subject: [PATCH 24/25] docker container v0.13.6 beta 2026.4.6 --- studio/install_python_stack.py | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-) diff --git a/studio/install_python_stack.py b/studio/install_python_stack.py index 9273c70272..fd3115e0ad 100644 --- a/studio/install_python_stack.py +++ b/studio/install_python_stack.py @@ -1180,9 +1180,9 @@ def install_python_stack() -> int: # constrain = False, # ) - if not IS_WINDOWS and not IS_MACOS and not NO_TORCH: - _progress("flash-attn") - _ensure_flash_attn() + #if not IS_WINDOWS and not IS_MACOS and not NO_TORCH: + # _progress("flash-attn") + # _ensure_flash_attn() # # 6. Patch: override llama_cpp.py with fix from unsloth-zoo feature/llama-cpp-windows-support branch # patch_package_file( From 0d2b10f0cd5e809864c7b2a9dcf188f12b00c8e0 Mon Sep 17 00:00:00 2001 From: Roland Tannous Date: Thu, 23 Apr 2026 20:48:58 +0000 Subject: [PATCH 25/25] fix failed to start on docker --- unsloth_cli/commands/studio.py | 421 ++++----------------------------- 1 file changed, 51 insertions(+), 370 deletions(-) diff --git a/unsloth_cli/commands/studio.py b/unsloth_cli/commands/studio.py index ca641bde54..2c81125588 100644 --- a/unsloth_cli/commands/studio.py +++ b/unsloth_cli/commands/studio.py @@ -1,19 +1,11 @@ # SPDX-License-Identifier: AGPL-3.0-only # Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0 -import importlib.util -import hashlib -import json import os import platform -import secrets -import sqlite3 import subprocess import sys -import tempfile import time -import types -from datetime import datetime, timezone from pathlib import Path from typing import List, Optional import typer @@ -21,52 +13,12 @@ import typer studio_app = typer.Typer(help = "Unsloth Studio commands.") STUDIO_HOME = Path.home() / ".unsloth" / "studio" -BOOTSTRAP_PASSWORD_FILE = ".bootstrap_password" -DESKTOP_SECRET_FILE = ".desktop_secret" -DEFAULT_ADMIN_USERNAME = "unsloth" -DESKTOP_SECRET_PREFIX = "desktop-" -API_KEY_PBKDF2_SALT_KEY = "api_key_pbkdf2_salt" -DESKTOP_SECRET_HASH_KEY = "desktop_secret_hash" -DESKTOP_SECRET_CREATED_AT_KEY = "desktop_secret_created_at" -PBKDF2_ITERATIONS = 100_000 # __file__ is unsloth_cli/commands/studio.py -- two parents up is the package root # (either site-packages or the repo root for editable installs). _PACKAGE_ROOT = Path(__file__).resolve().parent.parent.parent -def _should_hide_windows_subprocesses() -> bool: - """Hide child console windows only for non-interactive Windows launches.""" - if platform.system() != "Windows": - return False - try: - return not sys.stdout.isatty() - except (AttributeError, OSError, ValueError): - return True - - -def _windows_hidden_subprocess_kwargs() -> dict[str, object]: - """Return Windows-only Popen kwargs that suppress transient console windows.""" - if not _should_hide_windows_subprocesses(): - return {} - - kwargs: dict[str, object] = {} - create_no_window = getattr(subprocess, "CREATE_NO_WINDOW", 0) - if create_no_window: - kwargs["creationflags"] = create_no_window - - startupinfo_factory = getattr(subprocess, "STARTUPINFO", None) - startf_use_showwindow = getattr(subprocess, "STARTF_USESHOWWINDOW", 0) - sw_hide = getattr(subprocess, "SW_HIDE", 0) - if startupinfo_factory is not None and startf_use_showwindow: - startupinfo = startupinfo_factory() - startupinfo.dwFlags |= startf_use_showwindow - startupinfo.wShowWindow = sw_hide - kwargs["startupinfo"] = startupinfo - - return kwargs - - def _studio_venv_python() -> Optional[Path]: """Return the studio venv Python binary, or None if not set up.""" if platform.system() == "Windows": @@ -145,228 +97,12 @@ def _create_api_key_inprocess(name: str) -> str: ``POST /api/auth/api-keys`` on fresh installs. Safe because the CLI already has filesystem access to ``~/.unsloth/studio``. """ - storage = _load_backend_auth_storage() + from auth.storage import create_api_key, DEFAULT_ADMIN_USERNAME - raw_key, _row = storage.create_api_key( - username = storage.DEFAULT_ADMIN_USERNAME, - name = name, - ) + raw_key, _row = create_api_key(username = DEFAULT_ADMIN_USERNAME, name = name) return raw_key -def _load_backend_auth_storage(): - run_py = _find_run_py() - backend_dir = ( - run_py.parent if run_py is not None else _PACKAGE_ROOT / "studio" / "backend" - ) - if backend_dir.is_dir() and str(backend_dir) not in sys.path: - sys.path.insert(0, str(backend_dir)) - - auth_dir = backend_dir / "auth" - storage_py = auth_dir / "storage.py" - loaded = sys.modules.get("auth.storage") - loaded_path = Path(getattr(loaded, "__file__", "")).resolve() - if loaded is not None and loaded_path == storage_py: - return loaded - - package = sys.modules.get("auth") - package_paths = [Path(path).resolve() for path in getattr(package, "__path__", [])] - if package is None or auth_dir.resolve() not in package_paths: - package = types.ModuleType("auth") - package.__path__ = [str(auth_dir)] - package.__package__ = "auth" - package.__file__ = str(auth_dir / "__init__.py") - sys.modules["auth"] = package - - spec = importlib.util.spec_from_file_location("auth.storage", storage_py) - if spec is None or spec.loader is None: - raise ImportError(f"Could not load backend auth storage from {storage_py}") - storage = importlib.util.module_from_spec(spec) - sys.modules["auth.storage"] = storage - spec.loader.exec_module(storage) - - return storage - - -def _write_auth_secret(path: Path, secret: str) -> None: - path.parent.mkdir(parents = True, exist_ok = True) - fd, tmp_name = tempfile.mkstemp(prefix = f".{path.name}.", dir = path.parent) - tmp_path = Path(tmp_name) - try: - try: - os.chmod(tmp_path, 0o600) - except OSError: - pass - with os.fdopen(fd, "w") as f: - fd = -1 - f.write(secret) - os.replace(tmp_path, path) - except Exception: - if fd >= 0: - os.close(fd) - tmp_path.unlink(missing_ok = True) - raise - try: - os.chmod(path, 0o600) - except OSError: - pass - - -def _connect_auth_db() -> sqlite3.Connection: - auth_dir = STUDIO_HOME / "auth" - auth_dir.mkdir(parents = True, exist_ok = True) - conn = sqlite3.connect(auth_dir / "auth.db") - conn.execute( - """ - CREATE TABLE IF NOT EXISTS auth_user ( - id INTEGER PRIMARY KEY, - username TEXT UNIQUE NOT NULL, - password_salt TEXT NOT NULL, - password_hash TEXT NOT NULL, - jwt_secret TEXT NOT NULL, - must_change_password INTEGER NOT NULL DEFAULT 0 - ); - """ - ) - conn.execute( - """ - CREATE TABLE IF NOT EXISTS refresh_tokens ( - id INTEGER PRIMARY KEY, - token_hash TEXT NOT NULL, - username TEXT NOT NULL, - expires_at TEXT NOT NULL, - is_desktop INTEGER NOT NULL DEFAULT 0 - ); - """ - ) - conn.execute( - """ - CREATE TABLE IF NOT EXISTS api_keys ( - id INTEGER PRIMARY KEY AUTOINCREMENT, - username TEXT NOT NULL, - key_prefix TEXT NOT NULL, - key_hash TEXT NOT NULL UNIQUE, - name TEXT NOT NULL DEFAULT '', - created_at TEXT NOT NULL, - last_used_at TEXT, - expires_at TEXT, - is_active INTEGER NOT NULL DEFAULT 1 - ); - """ - ) - conn.execute( - """ - CREATE TABLE IF NOT EXISTS app_secrets ( - key TEXT PRIMARY KEY, - value TEXT NOT NULL - ); - """ - ) - auth_columns = {row[1] for row in conn.execute("PRAGMA table_info(auth_user)")} - if "must_change_password" not in auth_columns: - conn.execute( - "ALTER TABLE auth_user ADD COLUMN must_change_password INTEGER NOT NULL DEFAULT 0" - ) - refresh_columns = { - row[1] for row in conn.execute("PRAGMA table_info(refresh_tokens)") - } - if "is_desktop" not in refresh_columns: - conn.execute( - "ALTER TABLE refresh_tokens ADD COLUMN is_desktop INTEGER NOT NULL DEFAULT 0" - ) - conn.commit() - return conn - - -def _pbkdf2_hex(value: str, salt: bytes) -> str: - return hashlib.pbkdf2_hmac( - "sha256", - value.encode("utf-8"), - salt, - PBKDF2_ITERATIONS, - ).hex() - - -def _hash_password(password: str) -> tuple[str, str]: - salt = secrets.token_hex(16) - pwd_hash = _pbkdf2_hex(password, salt.encode("utf-8")) - return salt, pwd_hash - - -def _get_or_create_api_key_pbkdf2_salt(conn: sqlite3.Connection) -> bytes: - row = conn.execute( - "SELECT value FROM app_secrets WHERE key = ?", - (API_KEY_PBKDF2_SALT_KEY,), - ).fetchone() - if row is None: - salt_hex = secrets.token_hex(32) - conn.execute( - "INSERT OR IGNORE INTO app_secrets (key, value) VALUES (?, ?)", - (API_KEY_PBKDF2_SALT_KEY, salt_hex), - ) - row = conn.execute( - "SELECT value FROM app_secrets WHERE key = ?", - (API_KEY_PBKDF2_SALT_KEY,), - ).fetchone() - return bytes.fromhex(row[0]) - - -def _ensure_cli_default_admin(conn: sqlite3.Connection) -> None: - row = conn.execute( - "SELECT 1 FROM auth_user WHERE username = ?", - (DEFAULT_ADMIN_USERNAME,), - ).fetchone() - if row is not None: - return - - bootstrap_password = secrets.token_urlsafe(32) - password_salt, password_hash = _hash_password(bootstrap_password) - conn.execute( - """ - INSERT INTO auth_user ( - username, - password_salt, - password_hash, - jwt_secret, - must_change_password - ) - VALUES (?, ?, ?, ?, ?) - """, - ( - DEFAULT_ADMIN_USERNAME, - password_salt, - password_hash, - secrets.token_urlsafe(64), - 1, - ), - ) - _write_auth_secret( - STUDIO_HOME / "auth" / BOOTSTRAP_PASSWORD_FILE, - bootstrap_password, - ) - - -def _create_desktop_secret_in_cli() -> str: - raw_secret = DESKTOP_SECRET_PREFIX + secrets.token_urlsafe(48) - now = datetime.now(timezone.utc).isoformat() - conn = _connect_auth_db() - try: - _ensure_cli_default_admin(conn) - secret_hash = _pbkdf2_hex(raw_secret, _get_or_create_api_key_pbkdf2_salt(conn)) - conn.execute( - "INSERT OR REPLACE INTO app_secrets (key, value) VALUES (?, ?)", - (DESKTOP_SECRET_HASH_KEY, secret_hash), - ) - conn.execute( - "INSERT OR REPLACE INTO app_secrets (key, value) VALUES (?, ?)", - (DESKTOP_SECRET_CREATED_AT_KEY, now), - ) - conn.commit() - return raw_secret - finally: - conn.close() - - def _load_model_via_http( port: int, api_key: str, @@ -420,11 +156,6 @@ def studio_default( host: str = typer.Option("127.0.0.1", "--host", "-H"), frontend: Optional[Path] = typer.Option(None, "--frontend", "-f"), silent: bool = typer.Option(False, "--silent", "-q"), - api_only: bool = typer.Option( - False, - "--api-only", - help = "Run API server only, no frontend serving (for Tauri desktop app)", - ), ): """Launch the Unsloth Studio server.""" if ctx.invoked_subcommand is not None: @@ -436,49 +167,49 @@ def studio_default( 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") - if api_only: - args.append("--api-only") - # 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, **_windows_hidden_subprocess_kwargs()) - 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: typer.echo("Studio not set up. Run install.sh first.") raise typer.Exit(1) @@ -491,7 +222,7 @@ def studio_default( display_host = _resolve_external_ip() if host == "0.0.0.0" else host typer.echo(f"Starting Unsloth Studio on http://{display_host}:{port}") - run_kwargs = dict(host = host, port = port, silent = silent, api_only = api_only) + run_kwargs = dict(host = host, port = port, silent = silent) if frontend is not None: run_kwargs["frontend_path"] = frontend run_server(**run_kwargs) @@ -925,16 +656,9 @@ def _run_setup_script(*, verbose: bool = False) -> None: env = {**os.environ, "UNSLOTH_VERBOSE": "1"} if verbose else None if platform.system() == "Windows": - powershell_args = ["powershell.exe"] - if _should_hide_windows_subprocesses(): - powershell_args.extend( - ["-NoLogo", "-NoProfile", "-NonInteractive", "-WindowStyle", "Hidden"] - ) - powershell_args.extend(["-ExecutionPolicy", "Bypass", "-File", str(script)]) result = subprocess.run( - powershell_args, + ["powershell", "-ExecutionPolicy", "Bypass", "-File", str(script)], env = env, - **_windows_hidden_subprocess_kwargs(), ) else: result = subprocess.run(["bash", str(script)], env = env) @@ -990,44 +714,6 @@ def update( # ── unsloth studio reset-password ──────────────────────────────────── -@studio_app.command("desktop-capabilities", hidden = True) -def desktop_capabilities( - json_output: bool = typer.Option( - False, - "--json", - help = "Emit machine-readable JSON.", - ), -): - payload = { - "desktop_protocol_version": 1, - "supports_provision_desktop_auth": True, - "supports_api_only": True, - "version": "unknown", - } - try: - from importlib.metadata import version as package_version - - payload["version"] = package_version("unsloth") - except Exception: - pass - - if json_output: - typer.echo(json.dumps(payload, sort_keys = True)) - return - - for key, value in payload.items(): - typer.echo(f"{key}: {value}") - - -@studio_app.command("provision-desktop-auth", hidden = True) -def provision_desktop_auth(): - """Create/repair desktop auth state for the local machine.""" - auth_dir = STUDIO_HOME / "auth" - secret = _create_desktop_secret_in_cli() - _write_auth_secret(auth_dir / DESKTOP_SECRET_FILE, secret) - typer.echo("Desktop auth ready.") - - @studio_app.command("reset-password") def reset_password(): """Reset the Studio admin password. @@ -1038,18 +724,13 @@ def reset_password(): """ auth_dir = STUDIO_HOME / "auth" db_file = auth_dir / "auth.db" - stale_files = [ - auth_dir / BOOTSTRAP_PASSWORD_FILE, - auth_dir / DESKTOP_SECRET_FILE, - ] - had_db = db_file.exists() + pw_file = auth_dir / ".bootstrap_password" - db_file.unlink(missing_ok = True) - for path in stale_files: - path.unlink(missing_ok = True) - - if not had_db: + if not db_file.exists(): typer.echo("No auth database found -- nothing to reset.") raise typer.Exit(0) + db_file.unlink(missing_ok = True) + pw_file.unlink(missing_ok = True) + typer.echo("Auth database deleted. Restart Unsloth Studio to get a new password.")