From 970219a3035d3cc67e9953204f5a3ab58924bbf0 Mon Sep 17 00:00:00 2001
From: Roland Tannous <rolandtannous@gravityq.ai>
Date: Mon, 6 Apr 2026 16:25:07 +0000
Subject: [PATCH 01/17] split venv_t5 into venv_t5_530 and venv_t5_550 for
 tiered transformers 5.x support

---
 studio/backend/core/export/worker.py          |  39 ++-
 studio/backend/core/inference/worker.py       |  39 ++-
 studio/backend/core/training/worker.py        |  47 ++-
 .../tests/test_transformers_version.py        | 134 ++++++++
 studio/backend/utils/models/model_config.py   |   5 +-
 studio/backend/utils/transformers_version.py  | 294 ++++++++++++++----
 studio/setup.ps1                              |  74 ++++-
 studio/setup.sh                               |  37 ++-
 8 files changed, 534 insertions(+), 135 deletions(-)

diff --git a/studio/backend/core/export/worker.py b/studio/backend/core/export/worker.py
index 6af6ff1193..4bdfd13b78 100644
--- a/studio/backend/core/export/worker.py
+++ b/studio/backend/core/export/worker.py
@@ -32,8 +32,8 @@
 def _activate_transformers_version(model_name: str) -> None:
     """Activate the correct transformers version BEFORE any ML imports.
 
-    If the model needs transformers 5.x, prepend the pre-installed .venv_t5/
-    directory to sys.path. Otherwise do nothing (default 4.57.x in .venv/).
+    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)
@@ -41,24 +41,37 @@ def _activate_transformers_version(model_name: str) -> None:
         sys.path.insert(0, backend_path)
 
     from utils.transformers_version import (
-        needs_transformers_5,
+        get_transformers_tier,
         _resolve_base_model,
-        _ensure_venv_t5_exists,
-        _VENV_T5_DIR,
+        _ensure_venv_t5_530_exists,
+        _ensure_venv_t5_550_exists,
+        _VENV_T5_530_DIR,
+        _VENV_T5_550_DIR,
     )
 
     resolved = _resolve_base_model(model_name)
-    if needs_transformers_5(resolved):
-        if not _ensure_venv_t5_exists():
+    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.x: .venv_t5 missing at {_VENV_T5_DIR}"
+                f"Cannot activate transformers 5.3.0: .venv_t5_530 missing at {_VENV_T5_530_DIR}"
             )
-        if _VENV_T5_DIR not in sys.path:
-            sys.path.insert(0, _VENV_T5_DIR)
-        logger.info("Activated transformers 5.x from %s", _VENV_T5_DIR)
-        # Propagate to child subprocesses (e.g. GGUF converter)
+        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_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)
 
diff --git a/studio/backend/core/inference/worker.py b/studio/backend/core/inference/worker.py
index 7f7291a56d..3c5363cb77 100644
--- a/studio/backend/core/inference/worker.py
+++ b/studio/backend/core/inference/worker.py
@@ -36,8 +36,8 @@
 def _activate_transformers_version(model_name: str) -> None:
     """Activate the correct transformers version BEFORE any ML imports.
 
-    If the model needs transformers 5.x, prepend the pre-installed .venv_t5/
-    directory to sys.path. Otherwise do nothing (default 4.57.x in .venv/).
+    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)
@@ -45,24 +45,37 @@ def _activate_transformers_version(model_name: str) -> None:
         sys.path.insert(0, backend_path)
 
     from utils.transformers_version import (
-        needs_transformers_5,
+        get_transformers_tier,
         _resolve_base_model,
-        _ensure_venv_t5_exists,
-        _VENV_T5_DIR,
+        _ensure_venv_t5_530_exists,
+        _ensure_venv_t5_550_exists,
+        _VENV_T5_530_DIR,
+        _VENV_T5_550_DIR,
     )
 
     resolved = _resolve_base_model(model_name)
-    if needs_transformers_5(resolved):
-        if not _ensure_venv_t5_exists():
+    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.x: .venv_t5 missing at {_VENV_T5_DIR}"
+                f"Cannot activate transformers 5.3.0: .venv_t5_530 missing at {_VENV_T5_530_DIR}"
             )
-        if _VENV_T5_DIR not in sys.path:
-            sys.path.insert(0, _VENV_T5_DIR)
-        logger.info("Activated transformers 5.x from %s", _VENV_T5_DIR)
-        # Propagate to child subprocesses (e.g. GGUF converter)
+        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_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)
 
diff --git a/studio/backend/core/training/worker.py b/studio/backend/core/training/worker.py
index 0454eada89..db5cf8e948 100644
--- a/studio/backend/core/training/worker.py
+++ b/studio/backend/core/training/worker.py
@@ -308,8 +308,8 @@ def _ensure_mamba_ssm(event_queue: Any, model_name: str) -> None:
 def _activate_transformers_version(model_name: str) -> None:
     """Activate the correct transformers version BEFORE any ML imports.
 
-    If the model needs transformers 5.x, prepend the pre-installed .venv_t5/
-    directory to sys.path. Otherwise do nothing (default 4.57.x in .venv/).
+    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)
@@ -317,24 +317,37 @@ def _activate_transformers_version(model_name: str) -> None:
         sys.path.insert(0, backend_path)
 
     from utils.transformers_version import (
-        needs_transformers_5,
+        get_transformers_tier,
         _resolve_base_model,
-        _ensure_venv_t5_exists,
-        _VENV_T5_DIR,
+        _ensure_venv_t5_530_exists,
+        _ensure_venv_t5_550_exists,
+        _VENV_T5_530_DIR,
+        _VENV_T5_550_DIR,
     )
 
     resolved = _resolve_base_model(model_name)
-    if needs_transformers_5(resolved):
-        if not _ensure_venv_t5_exists():
+    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.x: .venv_t5 missing at {_VENV_T5_DIR}"
+                f"Cannot activate transformers 5.3.0: .venv_t5_530 missing at {_VENV_T5_530_DIR}"
             )
-        if _VENV_T5_DIR not in sys.path:
-            sys.path.insert(0, _VENV_T5_DIR)
-        logger.info("Activated transformers 5.x from %s", _VENV_T5_DIR)
-        # Propagate to child subprocesses (e.g. GGUF converter)
+        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_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)
 
@@ -392,14 +405,14 @@ def run_training_process(
     # 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 needs_transformers_5
+    from utils.transformers_version import get_transformers_tier
 
     _lowered = model_name.lower()
-    _is_native_t5 = any(x in _lowered for x in ("gemma-4", "gemma4"))
+    _tier = get_transformers_tier(model_name)
     if (
-        needs_transformers_5(model_name)
+        _tier != "default"
         and _lowered.startswith("unsloth/")
-        and not _is_native_t5
+        and _tier != "550"  # Gemma 4 is native t5.5 — trust_remote_code bypasses compiler
         and not config.get("trust_remote_code", False)
     ):
         config["trust_remote_code"] = True
diff --git a/studio/backend/tests/test_transformers_version.py b/studio/backend/tests/test_transformers_version.py
index f3dae537c7..387c1c6f13 100644
--- a/studio/backend/tests/test_transformers_version.py
+++ b/studio/backend/tests/test_transformers_version.py
@@ -31,8 +31,11 @@
 from utils.transformers_version import (
     _resolve_base_model,
     _check_tokenizer_config_needs_v5,
+    _check_config_needs_550,
     _tokenizer_class_cache,
+    _config_needs_550_cache,
     needs_transformers_5,
+    get_transformers_tier,
 )
 
 
@@ -188,3 +191,134 @@ def test_local_checkpoint_resolved_via_config(self, tmp_path: Path):
         # We test the full resolution chain here:
         resolved = _resolve_base_model(str(tmp_path))
         assert needs_transformers_5(resolved) is True
+
+
+# ---------------------------------------------------------------------------
+# _check_config_needs_550 — config.json architecture/model_type check
+# ---------------------------------------------------------------------------
+
+
+class TestCheckConfigNeeds550:
+    """Tests for _check_config_needs_550() local config.json checks."""
+
+    def setup_method(self):
+        _config_needs_550_cache.clear()
+
+    def test_gemma4_architecture(self, tmp_path: Path):
+        """config.json with Gemma4ForConditionalGeneration should return True."""
+        cfg = {"architectures": ["Gemma4ForConditionalGeneration"], "model_type": "gemma4"}
+        (tmp_path / "config.json").write_text(json.dumps(cfg))
+
+        assert _check_config_needs_550(str(tmp_path)) is True
+
+    def test_gemma4_model_type_only(self, tmp_path: Path):
+        """config.json with model_type=gemma4 (no architectures) should return True."""
+        cfg = {"model_type": "gemma4"}
+        (tmp_path / "config.json").write_text(json.dumps(cfg))
+
+        assert _check_config_needs_550(str(tmp_path)) is True
+
+    def test_llama_architecture(self, tmp_path: Path):
+        """config.json with LlamaForCausalLM should return False."""
+        cfg = {"architectures": ["LlamaForCausalLM"], "model_type": "llama"}
+        (tmp_path / "config.json").write_text(json.dumps(cfg))
+
+        assert _check_config_needs_550(str(tmp_path)) is False
+
+    def test_no_config_json(self, tmp_path: Path):
+        """Missing config.json should return False (fail-open)."""
+        # Patch network call to avoid real fetch
+        with patch("urllib.request.urlopen") as mock_urlopen:
+            mock_urlopen.side_effect = Exception("no network")
+            assert _check_config_needs_550(str(tmp_path)) is False
+
+    def test_result_is_cached(self, tmp_path: Path):
+        """Subsequent calls should use the cache."""
+        cfg = {"architectures": ["Gemma4ForConditionalGeneration"]}
+        (tmp_path / "config.json").write_text(json.dumps(cfg))
+
+        key = str(tmp_path)
+        _check_config_needs_550(key)
+        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."""
+        cfg = {"architectures": ["LlamaForCausalLM"]}
+        (tmp_path / "config.json").write_text(json.dumps(cfg))
+
+        with patch("urllib.request.urlopen") as mock_urlopen:
+            _check_config_needs_550(str(tmp_path))
+            mock_urlopen.assert_not_called()
+
+
+# ---------------------------------------------------------------------------
+# get_transformers_tier — tier detection
+# ---------------------------------------------------------------------------
+
+
+class TestGetTransformersTier:
+    """Tests for get_transformers_tier() tiered version detection."""
+
+    def setup_method(self):
+        _tokenizer_class_cache.clear()
+        _config_needs_550_cache.clear()
+
+    def test_gemma4_substring_returns_550(self):
+        assert get_transformers_tier("google/gemma-4-E2B-it") == "550"
+
+    def test_gemma4_alt_substring_returns_550(self):
+        assert get_transformers_tier("unsloth/gemma4-E4B-it") == "550"
+
+    def test_gemma4_config_json_returns_550(self, tmp_path: Path):
+        """Local checkpoint with Gemma4 architecture → 550."""
+        cfg = {"architectures": ["Gemma4ForConditionalGeneration"], "model_type": "gemma4"}
+        (tmp_path / "config.json").write_text(json.dumps(cfg))
+
+        assert get_transformers_tier(str(tmp_path)) == "550"
+
+    def test_qwen35_returns_530(self):
+        with patch(
+            "utils.transformers_version._check_config_needs_550",
+            return_value = False,
+        ):
+            assert get_transformers_tier("Qwen/Qwen3.5-9B") == "530"
+
+    def test_ministral_returns_530(self):
+        with patch(
+            "utils.transformers_version._check_config_needs_550",
+            return_value = False,
+        ):
+            assert get_transformers_tier("mistralai/Ministral-3-8B-Instruct-2512") == "530"
+
+    def test_llama_returns_default(self):
+        with patch(
+            "utils.transformers_version._check_config_needs_550",
+            return_value = False,
+        ), patch(
+            "utils.transformers_version._check_tokenizer_config_needs_v5",
+            return_value = False,
+        ):
+            assert get_transformers_tier("meta-llama/Llama-3-8B") == "default"
+
+    def test_550_checked_before_530(self):
+        """Ensure 5.5.0 is checked first — a model matching both should get 550."""
+        # This shouldn't happen in practice, but verifies priority
+        assert get_transformers_tier("gemma-4-model") == "550"
+
+    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
+        with patch(
+            "utils.transformers_version._check_config_needs_550",
+            return_value = False,
+        ):
+            assert needs_transformers_5("Qwen/Qwen3.5-9B") is True
+        with patch(
+            "utils.transformers_version._check_config_needs_550",
+            return_value = False,
+        ), patch(
+            "utils.transformers_version._check_tokenizer_config_needs_v5",
+            return_value = False,
+        ):
+            assert needs_transformers_5("meta-llama/Llama-3-8B") is False
diff --git a/studio/backend/utils/models/model_config.py b/studio/backend/utils/models/model_config.py
index 61226e52cb..5be0b183d5 100644
--- a/studio/backend/utils/models/model_config.py
+++ b/studio/backend/utils/models/model_config.py
@@ -493,8 +493,9 @@ def load_model_config(
     "minicpmv",
 }
 
-# Pre-computed .venv_t5 path and backend dir for subprocess version switching.
-_VENV_T5_DIR = str(Path.home() / ".unsloth" / "studio" / ".venv_t5")
+# Pre-computed .venv_t5 paths and backend dir for subprocess version switching.
+# Vision check uses 5.5.0 (newest, recognizes all architectures).
+_VENV_T5_DIR = str(Path.home() / ".unsloth" / "studio" / ".venv_t5_550")
 _BACKEND_DIR = str(Path(__file__).resolve().parent.parent.parent)
 
 # Inline script executed in a subprocess with transformers 5.x activated.
diff --git a/studio/backend/utils/transformers_version.py b/studio/backend/utils/transformers_version.py
index 07e4a5c000..a9a026895b 100644
--- a/studio/backend/utils/transformers_version.py
+++ b/studio/backend/utils/transformers_version.py
@@ -5,20 +5,25 @@
 Automatic transformers version switching.
 
 Some newer model architectures (Ministral-3, GLM-4.7-Flash, Qwen3-30B-A3B MoE,
-tiny_qwen3_moe) require transformers>=5.3.0, while everything else needs the
-default 4.57.x that ships with Unsloth.
+tiny_qwen3_moe) require transformers>=5.3.0, while Gemma 4 models require
+transformers>=5.5.0.  Everything else needs the default 4.57.x that ships
+with Unsloth.
+
+Two separate target directories are maintained:
+  - .venv_t5_530/  — transformers 5.3.0 (Ministral-3, GLM, Qwen3 MoE, etc.)
+  - .venv_t5_550/  — transformers 5.5.0 (Gemma 4)
 
 When loading a LoRA adapter with a custom name, we resolve the base model from
 ``adapter_config.json`` and check *that* against the model list.
 
 Strategy:
   Training and inference run in subprocesses that activate the correct version
-  via sys.path (prepending .venv_t5/ for 5.x models). See:
+  via sys.path (prepending the appropriate .venv_t5_*/ directory). See:
     - core/training/worker.py
     - core/inference/worker.py
 
   For export (still in-process), ensure_transformers_version() does a lightweight
-  sys.path swap using the same .venv_t5/ directory pre-installed by setup.sh.
+  sys.path swap using the same directories pre-installed by setup.sh.
 """
 
 import importlib
@@ -39,7 +44,7 @@
 # ---------------------------------------------------------------------------
 
 # Lowercase substrings — if ANY appears anywhere in the lowered model name,
-# we need transformers 5.x.
+# we need transformers 5.3.0.
 TRANSFORMERS_5_MODEL_SUBSTRINGS: tuple[str, ...] = (
     "ministral-3-",  # Ministral-3-{3,8,14}B-{Instruct,Reasoning,Base}-2512
     "glm-4.7-flash",  # GLM-4.7-Flash
@@ -47,10 +52,23 @@
     "qwen3.5",  # Qwen3.5 family (35B-A3B, etc.)
     "qwen3-next",  # Qwen3-Next and variants
     "tiny_qwen3_moe",  # imdatta0/tiny_qwen3_moe_2.8B_0.7B
+)
+
+# Lowercase substrings for models that require transformers 5.5.0 (checked first).
+TRANSFORMERS_550_MODEL_SUBSTRINGS: tuple[str, ...] = (
     "gemma-4",  # Gemma-4 (E2B-it, E4B-it, 31B-it, 26B-A4B-it)
     "gemma4",  # Gemma-4 alternate naming
 )
 
+# Architecture classes / model_type values that require transformers 5.5.0.
+# Checked via config.json (local or HuggingFace).
+_TRANSFORMERS_550_ARCHITECTURES: set[str] = {
+    "Gemma4ForConditionalGeneration",
+}
+_TRANSFORMERS_550_MODEL_TYPES: set[str] = {
+    "gemma4",
+}
+
 # Tokenizer classes that only exist in transformers>=5.x
 _TRANSFORMERS_5_TOKENIZER_CLASSES: set[str] = {
     "TokenizersBackend",
@@ -59,12 +77,21 @@
 # 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] = {}
+
 # Versions
-TRANSFORMERS_5_VERSION = "5.5.0"
+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
+TRANSFORMERS_5_VERSION = TRANSFORMERS_550_VERSION
 
-# Pre-installed directory for transformers 5.x — created by setup.sh / setup.ps1
-_VENV_T5_DIR = str(Path.home() / ".unsloth" / "studio" / ".venv_t5")
+# Pre-installed directories — created by setup.sh / setup.ps1
+_VENV_T5_530_DIR = str(Path.home() / ".unsloth" / "studio" / ".venv_t5_530")
+_VENV_T5_550_DIR = str(Path.home() / ".unsloth" / "studio" / ".venv_t5_550")
+# Backwards-compat alias
+_VENV_T5_DIR = _VENV_T5_550_DIR
 
 
 def _resolve_base_model(model_name: str) -> str:
@@ -192,18 +219,105 @@ def _check_tokenizer_config_needs_v5(model_name: str) -> bool:
         return False
 
 
-def needs_transformers_5(model_name: str) -> bool:
-    """Return True if *model_name* belongs to an architecture that requires
-    ``transformers>=5.3.0``.
+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).
 
-    First checks the hardcoded substring list for known models, then
-    dynamically fetches ``tokenizer_config.json`` from HuggingFace to check
-    if the tokenizer_class (e.g. ``TokenizersBackend``) requires v5.
+    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).
+    """
+    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_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)
+            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)
+
+    # --- Fall back to fetching from HuggingFace ---------------------------
+    import urllib.request
+
+    url = f"https://huggingface.co/{model_name}/raw/main/config.json"
+    try:
+        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
+    except Exception as exc:
+        logger.debug(
+            "Could not fetch config.json for '%s': %s", model_name, exc
+        )
+        _config_needs_550_cache[model_name] = False
+        return False
+
+
+def get_transformers_tier(model_name: str) -> str:
+    """Return the transformers tier required for *model_name*.
+
+    Returns ``"550"`` for models needing transformers 5.5.0 (e.g. Gemma 4),
+    ``"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.
     """
     lowered = model_name.lower()
+
+    # --- 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 True
-    return _check_tokenizer_config_needs_v5(model_name)
+        return "530"
+    if _check_tokenizer_config_needs_v5(model_name):
+        return "530"
+
+    return "default"
+
+
+def needs_transformers_5(model_name: str) -> bool:
+    """Return True if *model_name* requires any transformers 5.x version.
+
+    Convenience wrapper around :func:`get_transformers_tier`.
+    """
+    return get_transformers_tier(model_name) != "default"
 
 
 # ---------------------------------------------------------------------------
@@ -258,27 +372,36 @@ def _purge_modules() -> int:
     return len(to_remove)
 
 
-_VENV_T5_PACKAGES = (
-    f"transformers=={TRANSFORMERS_5_VERSION}",
+_VENV_T5_530_PACKAGES = (
+    f"transformers=={TRANSFORMERS_530_VERSION}",
+    "huggingface_hub==1.8.0",
+    "hf_xet==1.4.2",
+    "tiktoken",
+)
+
+_VENV_T5_550_PACKAGES = (
+    f"transformers=={TRANSFORMERS_550_VERSION}",
     "huggingface_hub==1.8.0",
     "hf_xet==1.4.2",
     "tiktoken",
 )
 
+# Backwards-compat alias
+_VENV_T5_PACKAGES = _VENV_T5_550_PACKAGES
 
-def _venv_t5_is_valid() -> bool:
-    """Return True if .venv_t5/ has all required packages at the correct versions."""
-    if not os.path.isdir(_VENV_T5_DIR) or not os.listdir(_VENV_T5_DIR):
+
+def _venv_dir_is_valid(venv_dir: str, packages: tuple[str, ...]) -> bool:
+    """Return True if *venv_dir* has all *packages* at the correct versions."""
+    if not os.path.isdir(venv_dir) or not os.listdir(venv_dir):
         return False
-    # Check that the key package directories exist AND match the required version
-    for pkg_spec in _VENV_T5_PACKAGES:
+    for pkg_spec in packages:
         parts = pkg_spec.split("==")
         pkg_name = parts[0]
         pkg_version = parts[1] if len(parts) > 1 else None
         pkg_name_norm = pkg_name.replace("-", "_")
         # Check directory exists
         if not any(
-            (Path(_VENV_T5_DIR) / d).is_dir()
+            (Path(venv_dir) / d).is_dir()
             for d in (pkg_name_norm, pkg_name_norm.replace("_", "-"))
         ):
             return False
@@ -287,7 +410,7 @@ def _venv_t5_is_valid() -> bool:
             continue
         # Check version via .dist-info metadata
         dist_info_found = False
-        for di in Path(_VENV_T5_DIR).glob(f"{pkg_name_norm}-*.dist-info"):
+        for di in Path(venv_dir).glob(f"{pkg_name_norm}-*.dist-info"):
             metadata = di / "METADATA"
             if not metadata.is_file():
                 continue
@@ -296,7 +419,8 @@ def _venv_t5_is_valid() -> bool:
                     installed_ver = line.split(":", 1)[1].strip()
                     if installed_ver != pkg_version:
                         logger.info(
-                            ".venv_t5 has %s==%s but need %s",
+                            "%s has %s==%s but need %s",
+                            venv_dir,
                             pkg_name,
                             installed_ver,
                             pkg_version,
@@ -311,8 +435,13 @@ def _venv_t5_is_valid() -> bool:
     return True
 
 
-def _install_to_venv_t5(pkg: str) -> bool:
-    """Install a single package into .venv_t5/, preferring uv then pip."""
+def _venv_t5_is_valid() -> bool:
+    """Backwards-compat: check the 5.5.0 venv."""
+    return _venv_dir_is_valid(_VENV_T5_550_DIR, _VENV_T5_550_PACKAGES)
+
+
+def _install_to_dir(pkg: str, target_dir: str) -> bool:
+    """Install a single package into *target_dir*, preferring uv then pip."""
     # Try uv first (faster) if already on PATH -- do NOT install uv at runtime
     if shutil.which("uv"):
         result = subprocess.run(
@@ -323,7 +452,7 @@ def _install_to_venv_t5(pkg: str) -> bool:
                 "--python",
                 sys.executable,
                 "--target",
-                _VENV_T5_DIR,
+                target_dir,
                 "--no-deps",
                 "--upgrade",
                 pkg,
@@ -344,7 +473,7 @@ def _install_to_venv_t5(pkg: str) -> bool:
             "pip",
             "install",
             "--target",
-            _VENV_T5_DIR,
+            target_dir,
             "--no-deps",
             "--upgrade",
             pkg,
@@ -359,47 +488,58 @@ def _install_to_venv_t5(pkg: str) -> bool:
     return True
 
 
-def _ensure_venv_t5_exists() -> bool:
-    """Ensure .venv_t5/ exists with all required packages. Install if missing."""
-    if _venv_t5_is_valid():
+def _ensure_venv_dir(venv_dir: str, packages: tuple[str, ...], label: str) -> bool:
+    """Ensure *venv_dir* exists with all *packages*. Install if missing."""
+    if _venv_dir_is_valid(venv_dir, packages):
         return True
 
     logger.warning(
-        ".venv_t5 not found or incomplete at %s -- installing at runtime", _VENV_T5_DIR
+        "%s not found or incomplete at %s -- installing at runtime", label, venv_dir
     )
-    shutil.rmtree(_VENV_T5_DIR, ignore_errors = True)
-    os.makedirs(_VENV_T5_DIR, exist_ok = True)
-    for pkg in _VENV_T5_PACKAGES:
-        if not _install_to_venv_t5(pkg):
+    shutil.rmtree(venv_dir, ignore_errors = True)
+    os.makedirs(venv_dir, exist_ok = True)
+    for pkg in packages:
+        if not _install_to_dir(pkg, venv_dir):
             return False
-    logger.info("Installed transformers 5.x to %s", _VENV_T5_DIR)
+    logger.info("Installed %s to %s", label, venv_dir)
     return True
 
 
-def _activate_5x() -> None:
-    """Prepend .venv_t5/ to sys.path, purge stale modules, reimport."""
-    if not _ensure_venv_t5_exists():
-        raise RuntimeError(
-            f"Cannot activate transformers 5.x: .venv_t5 missing at {_VENV_T5_DIR}"
-        )
+def _ensure_venv_t5_530_exists() -> bool:
+    """Ensure .venv_t5_530/ exists with transformers 5.3.0."""
+    return _ensure_venv_dir(_VENV_T5_530_DIR, _VENV_T5_530_PACKAGES, "transformers 5.3.0")
+
+
+def _ensure_venv_t5_550_exists() -> bool:
+    """Ensure .venv_t5_550/ exists with transformers 5.5.0."""
+    return _ensure_venv_dir(_VENV_T5_550_DIR, _VENV_T5_550_PACKAGES, "transformers 5.5.0")
 
-    if _VENV_T5_DIR not in sys.path:
-        sys.path.insert(0, _VENV_T5_DIR)
-        logger.info("Prepended %s to sys.path", _VENV_T5_DIR)
+
+def _ensure_venv_t5_exists() -> bool:
+    """Backwards-compat: ensure the 5.5.0 venv exists."""
+    return _ensure_venv_t5_550_exists()
+
+
+def _activate_venv(venv_dir: str, label: str) -> None:
+    """Prepend *venv_dir* to sys.path, purge stale modules, reimport."""
+    if venv_dir not in sys.path:
+        sys.path.insert(0, venv_dir)
+        logger.info("Prepended %s to sys.path", venv_dir)
 
     count = _purge_modules()
     logger.info("Purged %d cached modules", count)
 
     import transformers
 
-    logger.info("Loaded transformers %s", transformers.__version__)
+    logger.info("Loaded transformers %s (%s)", transformers.__version__, label)
 
 
 def _deactivate_5x() -> None:
-    """Remove .venv_t5/ from sys.path, purge stale modules, reimport."""
-    while _VENV_T5_DIR in sys.path:
-        sys.path.remove(_VENV_T5_DIR)
-    logger.info("Removed %s from sys.path", _VENV_T5_DIR)
+    """Remove all .venv_t5_*/ dirs from sys.path, purge stale modules, reimport."""
+    for d in (_VENV_T5_530_DIR, _VENV_T5_550_DIR):
+        while d in sys.path:
+            sys.path.remove(d)
+    logger.info("Removed venv_t5 dirs from sys.path")
 
     count = _purge_modules()
     logger.info("Purged %d cached modules", count)
@@ -412,9 +552,10 @@ def _deactivate_5x() -> None:
 def ensure_transformers_version(model_name: str) -> None:
     """Ensure the correct ``transformers`` version is active for *model_name*.
 
-    Uses sys.path with .venv_t5/ (pre-installed by setup.sh):
-      • Need 5.x → prepend .venv_t5/ to sys.path, purge modules.
-      • Need 4.x → remove .venv_t5/ from sys.path, purge modules.
+    Uses sys.path with .venv_t5_530/ or .venv_t5_550/ (pre-installed by setup.sh):
+      • Need 5.5.0 → prepend .venv_t5_550/ to sys.path, purge modules.
+      • Need 5.3.0 → prepend .venv_t5_530/ to sys.path, purge modules.
+      • Need 4.x  → remove all .venv_t5_*/ from sys.path, purge modules.
 
     For LoRA adapters with custom names, the base model is resolved from
     ``adapter_config.json`` before checking.
@@ -424,8 +565,21 @@ def ensure_transformers_version(model_name: str) -> None:
     """
     # Resolve LoRA adapters to their base model for accurate detection
     resolved = _resolve_base_model(model_name)
-    want_5 = needs_transformers_5(resolved)
-    target_version = TRANSFORMERS_5_VERSION if want_5 else TRANSFORMERS_DEFAULT_VERSION
+    tier = get_transformers_tier(resolved)
+
+    if tier == "550":
+        target_version = TRANSFORMERS_550_VERSION
+        venv_dir = _VENV_T5_550_DIR
+        ensure_fn = _ensure_venv_t5_550_exists
+    elif tier == "530":
+        target_version = TRANSFORMERS_530_VERSION
+        venv_dir = _VENV_T5_530_DIR
+        ensure_fn = _ensure_venv_t5_530_exists
+    else:
+        target_version = TRANSFORMERS_DEFAULT_VERSION
+        venv_dir = None
+        ensure_fn = None
+
     target_major = int(target_version.split(".")[0])
 
     # Check what's actually loaded in memory
@@ -441,8 +595,17 @@ def ensure_transformers_version(model_name: str) -> None:
 
     # --- Already correct? ---------------------------------------------------
     if in_memory is not None:
+        if in_memory == target_version:
+            logger.info(
+                "transformers %s already loaded — correct for '%s'",
+                in_memory,
+                model_name,
+            )
+            return
+        # Different 5.x → need to switch (e.g. 5.3.0 loaded but need 5.5.0)
         in_memory_major = int(in_memory.split(".")[0])
-        if in_memory_major == target_major:
+        if in_memory_major == target_major and venv_dir is None:
+            # Both are default (4.x) — close enough
             logger.info(
                 "transformers %s already loaded — correct for '%s'",
                 in_memory,
@@ -451,9 +614,16 @@ def ensure_transformers_version(model_name: str) -> None:
             return
 
     # --- Switch version -----------------------------------------------------
-    if want_5:
-        logger.info("Activating transformers %s via .venv_t5…", TRANSFORMERS_5_VERSION)
-        _activate_5x()
+    if venv_dir is not None:
+        # First remove any other 5.x venv from sys.path
+        _deactivate_5x()
+        if not ensure_fn():
+            raise RuntimeError(
+                f"Cannot activate transformers {target_version}: "
+                f"venv missing at {venv_dir}"
+            )
+        logger.info("Activating transformers %s…", target_version)
+        _activate_venv(venv_dir, f"transformers {target_version}")
     else:
         logger.info(
             "Reverting to default transformers %s…", TRANSFORMERS_DEFAULT_VERSION
diff --git a/studio/setup.ps1 b/studio/setup.ps1
index 60dfce0661..fa853cdadf 100644
--- a/studio/setup.ps1
+++ b/studio/setup.ps1
@@ -1579,48 +1579,88 @@ if ($stackExit -ne 0) {
     exit 1
 }
 
-# ── Pre-install transformers 5.x into .venv_t5/ ──
-# Models like GLM-4.7-Flash need transformers>=5.3.0. Instead of pip-installing
-# at runtime (slow, ~10-15s), we pre-install into a separate directory.
-# The training subprocess just prepends .venv_t5/ to sys.path -- instant switch.
+# ── 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 ""
-substep "pre-installing transformers 5.x for newer model support..."
-$VenvT5Dir = Join-Path $env:USERPROFILE ".unsloth\studio\.venv_t5"
-if (Test-Path $VenvT5Dir) { Remove-Item -Recurse -Force $VenvT5Dir }
-New-Item -ItemType Directory -Path $VenvT5Dir -Force | Out-Null
+
+# Clean up legacy single .venv_t5 directory
+$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 $VenvT5Dir --no-deps $pkg
+        Fast-Install --target $VenvT5_550Dir --no-deps $pkg
         $t5PkgExit = $LASTEXITCODE
         $output = ""
     } else {
-        $output = Fast-Install --target $VenvT5Dir --no-deps $pkg | Out-String
+        $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/" -ForegroundColor Red
+        Write-Host "[FAIL] Could not install $pkg into .venv_t5_550/" -ForegroundColor Red
         Write-Host $output -ForegroundColor Red
         $ErrorActionPreference = $prevEAP_t5
         exit 1
     }
 }
-# tiktoken is needed by Qwen-family tokenizers -- install with deps since
-# regex/requests may be missing on Windows
 if ($script:UnslothVerbose) {
-    Fast-Install --target $VenvT5Dir tiktoken
+    Fast-Install --target $VenvT5_550Dir tiktoken
     $tiktokenInstallExit = $LASTEXITCODE
     $output = ""
 } else {
-    $output = Fast-Install --target $VenvT5Dir tiktoken | Out-String
+    $output = Fast-Install --target $VenvT5_550Dir tiktoken | Out-String
     $tiktokenInstallExit = $LASTEXITCODE
 }
 if ($tiktokenInstallExit -ne 0) {
-    substep "Could not install tiktoken into .venv_t5/ -- Qwen tokenizers may fail" "Yellow"
+    substep "Could not install tiktoken into .venv_t5_550/ -- Qwen tokenizers may fail" "Yellow"
 }
 $ErrorActionPreference = $prevEAP_t5
-step "transformers" "5.x pre-installed"
+step "transformers" "5.5.0 pre-installed"
 
 } else {
     step "python" "dependencies up to date"
diff --git a/studio/setup.sh b/studio/setup.sh
index e3ff2da35c..72df82f601 100755
--- a/studio/setup.sh
+++ b/studio/setup.sh
@@ -394,11 +394,14 @@ fi
 # ── Python venv + deps ──
 STUDIO_HOME="$HOME/.unsloth/studio"
 VENV_DIR="$STUDIO_HOME/unsloth_studio"
-VENV_T5_DIR="$STUDIO_HOME/.venv_t5"
+VENV_T5_530_DIR="$STUDIO_HOME/.venv_t5_530"
+VENV_T5_550_DIR="$STUDIO_HOME/.venv_t5_550"
 
 [ -d "$REPO_ROOT/.venv" ] && rm -rf "$REPO_ROOT/.venv"
 [ -d "$REPO_ROOT/.venv_overlay" ] && rm -rf "$REPO_ROOT/.venv_overlay"
 [ -d "$REPO_ROOT/.venv_t5" ] && rm -rf "$REPO_ROOT/.venv_t5"
+[ -d "$REPO_ROOT/.venv_t5_530" ] && rm -rf "$REPO_ROOT/.venv_t5_530"
+[ -d "$REPO_ROOT/.venv_t5_550" ] && rm -rf "$REPO_ROOT/.venv_t5_550"
 # Note: do NOT delete $STUDIO_HOME/.venv here — install.sh handles migration
 
 _COLAB_NO_VENV=false
@@ -502,16 +505,28 @@ fi
 if [ "$_SKIP_PYTHON_DEPS" = false ]; then
     install_python_stack
 
-    # ── 6b. Pre-install transformers 5.x into .venv_t5/ ──
-    # Models like GLM-4.7-Flash need transformers>=5.3.0. Instead of pip-installing
-    # at runtime (slow, ~10-15s), we pre-install into a separate directory.
-    # The training subprocess just prepends .venv_t5/ to sys.path -- instant switch.
-    mkdir -p "$VENV_T5_DIR"
-    run_quiet "install transformers 5.x" fast_install --target "$VENV_T5_DIR" --no-deps "transformers==5.5.0"
-    run_quiet "install huggingface_hub for t5" fast_install --target "$VENV_T5_DIR" --no-deps "huggingface_hub==1.8.0"
-    run_quiet "install hf_xet for t5" fast_install --target "$VENV_T5_DIR" --no-deps "hf_xet==1.4.2"
-    run_quiet "install tiktoken for t5" fast_install --target "$VENV_T5_DIR" "tiktoken"
-    step "transformers" "5.x pre-installed"
+    # ── 6b. Pre-install transformers 5.x into .venv_t5_530/ and .venv_t5_550/ ──
+    # Models like GLM-4.7-Flash, Qwen3 MoE need transformers>=5.3.0.
+    # Gemma 4 models need transformers>=5.5.0.
+    # Pre-install into separate directories to avoid runtime pip overhead.
+    # The training subprocess prepends the appropriate dir to sys.path.
+
+    # Clean up legacy single .venv_t5 directory
+    [ -d "$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 e8328c11de8c14f8fd361bf15d8a98a1cbef83fc Mon Sep 17 00:00:00 2001
From: Roland Tannous <rolandtannous@gravityq.ai>
Date: Mon, 6 Apr 2026 17:28:55 +0000
Subject: [PATCH 02/17] 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 | 8 ++++----
 unsloth/models/loader.py               | 4 +++-
 2 files changed, 7 insertions(+), 5 deletions(-)

diff --git a/studio/backend/core/training/worker.py b/studio/backend/core/training/worker.py
index db5cf8e948..9815e9d7b7 100644
--- a/studio/backend/core/training/worker.py
+++ b/studio/backend/core/training/worker.py
@@ -402,9 +402,9 @@ def run_training_process(
     # ── 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).
+    # Only auto-enable for models that genuinely need it (set in YAML defaults).
+    # Native transformers 5.x models (Qwen3.5, Gemma 4, etc.) do NOT need it
+    # and enabling it can bypass the compiler (disabling fused CE).
     from utils.transformers_version import get_transformers_tier
 
     _lowered = model_name.lower()
@@ -412,7 +412,7 @@ def run_training_process(
     if (
         _tier != "default"
         and _lowered.startswith("unsloth/")
-        and _tier != "550"  # Gemma 4 is native t5.5 — trust_remote_code bypasses compiler
+        and _tier not in ("530", "550")  # Native t5 models don't need trust_remote_code
         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 df97c5c7df..a34744dba8 100644
--- a/unsloth/models/loader.py
+++ b/unsloth/models/loader.py
@@ -1357,7 +1357,9 @@ def from_pretrained(
                 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 efd11915b8efdefcdea5bf372a4e2e3b19ce3417 Mon Sep 17 00:00:00 2001
From: Roland Tannous <rolandtannous@gravityq.ai>
Date: Mon, 6 Apr 2026 17:29:22 +0000
Subject: [PATCH 03/17] 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 a34744dba8..df97c5c7df 100644
--- a/unsloth/models/loader.py
+++ b/unsloth/models/loader.py
@@ -1357,9 +1357,7 @@ def from_pretrained(
                 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 3400654afc95a26dee687b35e15ea229aa16a10d Mon Sep 17 00:00:00 2001
From: Roland Tannous <rolandtannous@gravityq.ai>
Date: Mon, 6 Apr 2026 18:20:08 +0000
Subject: [PATCH 04/17] restrict trust_remote_code auto-enable to Nemotron
 models only

---
 studio/backend/core/inference/worker.py | 13 +++++--------
 studio/backend/core/training/worker.py  | 20 +++++++-------------
 2 files changed, 12 insertions(+), 21 deletions(-)

diff --git a/studio/backend/core/inference/worker.py b/studio/backend/core/inference/worker.py
index 3c5363cb77..1010e56dac 100644
--- a/studio/backend/core/inference/worker.py
+++ b/studio/backend/core/inference/worker.py
@@ -322,19 +322,16 @@ 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 unsloth/* transformers 5.x models
-        # (matches the training worker logic in core/training/worker.py)
+        # 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.
         trust_remote_code = config.get("trust_remote_code", False)
         if not trust_remote_code:
-            from utils.transformers_version import needs_transformers_5
-
             model_name = config["model_name"]
-            if needs_transformers_5(model_name) and model_name.lower().startswith(
-                "unsloth/"
-            ):
+            if "nemotron" in model_name.lower():
                 trust_remote_code = True
                 logger.info(
-                    "Auto-enabled trust_remote_code for unsloth/* transformers 5.x model: %s",
+                    "Auto-enabled trust_remote_code for Nemotron model: %s",
                     model_name,
                 )
 
diff --git a/studio/backend/core/training/worker.py b/studio/backend/core/training/worker.py
index 9815e9d7b7..b412a05cc9 100644
--- a/studio/backend/core/training/worker.py
+++ b/studio/backend/core/training/worker.py
@@ -399,25 +399,19 @@ def run_training_process(
         )
         return
 
-    # ── 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 models that genuinely need it (set in YAML defaults).
-    # Native transformers 5.x models (Qwen3.5, Gemma 4, etc.) do NOT need it
-    # and enabling it can bypass the compiler (disabling fused CE).
-    from utils.transformers_version import get_transformers_tier
-
+    # ── 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).
     _lowered = model_name.lower()
-    _tier = get_transformers_tier(model_name)
     if (
-        _tier != "default"
-        and _lowered.startswith("unsloth/")
-        and _tier not in ("530", "550")  # Native t5 models don't need trust_remote_code
+        "nemotron" in _lowered
         and not config.get("trust_remote_code", False)
     ):
         config["trust_remote_code"] = True
         logger.info(
-            "Auto-enabled trust_remote_code for unsloth/* transformers 5.x model: %s",
+            "Auto-enabled trust_remote_code for Nemotron model: %s",
             model_name,
         )
 

From 41b103adee260fa944e5dbf76a5688361c2b8164 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 05/17] [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 +--
 .../tests/test_transformers_version.py        | 44 ++++++++++++-------
 studio/backend/utils/transformers_version.py  | 12 ++---
 3 files changed, 37 insertions(+), 24 deletions(-)

diff --git a/studio/backend/core/training/worker.py b/studio/backend/core/training/worker.py
index b412a05cc9..54667dd7af 100644
--- a/studio/backend/core/training/worker.py
+++ b/studio/backend/core/training/worker.py
@@ -405,10 +405,7 @@ def run_training_process(
     # (Qwen3.5, Gemma 4, etc.) are native and do NOT need it — enabling it
     # bypasses the compiler (disabling fused CE).
     _lowered = model_name.lower()
-    if (
-        "nemotron" in _lowered
-        and not config.get("trust_remote_code", False)
-    ):
+    if "nemotron" in _lowered and not config.get("trust_remote_code", False):
         config["trust_remote_code"] = True
         logger.info(
             "Auto-enabled trust_remote_code for Nemotron model: %s",
diff --git a/studio/backend/tests/test_transformers_version.py b/studio/backend/tests/test_transformers_version.py
index 387c1c6f13..c031c2fea3 100644
--- a/studio/backend/tests/test_transformers_version.py
+++ b/studio/backend/tests/test_transformers_version.py
@@ -206,7 +206,10 @@ def setup_method(self):
 
     def test_gemma4_architecture(self, tmp_path: Path):
         """config.json with Gemma4ForConditionalGeneration should return True."""
-        cfg = {"architectures": ["Gemma4ForConditionalGeneration"], "model_type": "gemma4"}
+        cfg = {
+            "architectures": ["Gemma4ForConditionalGeneration"],
+            "model_type": "gemma4",
+        }
         (tmp_path / "config.json").write_text(json.dumps(cfg))
 
         assert _check_config_needs_550(str(tmp_path)) is True
@@ -272,7 +275,10 @@ def test_gemma4_alt_substring_returns_550(self):
 
     def test_gemma4_config_json_returns_550(self, tmp_path: Path):
         """Local checkpoint with Gemma4 architecture → 550."""
-        cfg = {"architectures": ["Gemma4ForConditionalGeneration"], "model_type": "gemma4"}
+        cfg = {
+            "architectures": ["Gemma4ForConditionalGeneration"],
+            "model_type": "gemma4",
+        }
         (tmp_path / "config.json").write_text(json.dumps(cfg))
 
         assert get_transformers_tier(str(tmp_path)) == "550"
@@ -289,15 +295,20 @@ def test_ministral_returns_530(self):
             "utils.transformers_version._check_config_needs_550",
             return_value = False,
         ):
-            assert get_transformers_tier("mistralai/Ministral-3-8B-Instruct-2512") == "530"
+            assert (
+                get_transformers_tier("mistralai/Ministral-3-8B-Instruct-2512") == "530"
+            )
 
     def test_llama_returns_default(self):
-        with patch(
-            "utils.transformers_version._check_config_needs_550",
-            return_value = False,
-        ), patch(
-            "utils.transformers_version._check_tokenizer_config_needs_v5",
-            return_value = False,
+        with (
+            patch(
+                "utils.transformers_version._check_config_needs_550",
+                return_value = False,
+            ),
+            patch(
+                "utils.transformers_version._check_tokenizer_config_needs_v5",
+                return_value = False,
+            ),
         ):
             assert get_transformers_tier("meta-llama/Llama-3-8B") == "default"
 
@@ -314,11 +325,14 @@ def test_needs_transformers_5_compat(self):
             return_value = False,
         ):
             assert needs_transformers_5("Qwen/Qwen3.5-9B") is True
-        with patch(
-            "utils.transformers_version._check_config_needs_550",
-            return_value = False,
-        ), patch(
-            "utils.transformers_version._check_tokenizer_config_needs_v5",
-            return_value = False,
+        with (
+            patch(
+                "utils.transformers_version._check_config_needs_550",
+                return_value = False,
+            ),
+            patch(
+                "utils.transformers_version._check_tokenizer_config_needs_v5",
+                return_value = False,
+            ),
         ):
             assert needs_transformers_5("meta-llama/Llama-3-8B") is False
diff --git a/studio/backend/utils/transformers_version.py b/studio/backend/utils/transformers_version.py
index a9a026895b..7ba20a366b 100644
--- a/studio/backend/utils/transformers_version.py
+++ b/studio/backend/utils/transformers_version.py
@@ -279,9 +279,7 @@ def _check_cfg(cfg: dict) -> bool:
         _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
-        )
+        logger.debug("Could not fetch config.json for '%s': %s", model_name, exc)
         _config_needs_550_cache[model_name] = False
         return False
 
@@ -507,12 +505,16 @@ def _ensure_venv_dir(venv_dir: str, packages: tuple[str, ...], label: str) -> bo
 
 def _ensure_venv_t5_530_exists() -> bool:
     """Ensure .venv_t5_530/ exists with transformers 5.3.0."""
-    return _ensure_venv_dir(_VENV_T5_530_DIR, _VENV_T5_530_PACKAGES, "transformers 5.3.0")
+    return _ensure_venv_dir(
+        _VENV_T5_530_DIR, _VENV_T5_530_PACKAGES, "transformers 5.3.0"
+    )
 
 
 def _ensure_venv_t5_550_exists() -> bool:
     """Ensure .venv_t5_550/ exists with transformers 5.5.0."""
-    return _ensure_venv_dir(_VENV_T5_550_DIR, _VENV_T5_550_PACKAGES, "transformers 5.5.0")
+    return _ensure_venv_dir(
+        _VENV_T5_550_DIR, _VENV_T5_550_PACKAGES, "transformers 5.5.0"
+    )
 
 
 def _ensure_venv_t5_exists() -> bool:

From fc49ae24531780a658049e6c238f146960f466b6 Mon Sep 17 00:00:00 2001
From: Roland Tannous <rolandtannous@gravityq.ai>
Date: Mon, 6 Apr 2026 19:40:11 +0000
Subject: [PATCH 06/17] 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  | 121 ++++++++++--------
 4 files changed, 99 insertions(+), 59 deletions(-)

diff --git a/studio/backend/core/inference/worker.py b/studio/backend/core/inference/worker.py
index 1010e56dac..506e631c51 100644
--- a/studio/backend/core/inference/worker.py
+++ b/studio/backend/core/inference/worker.py
@@ -328,7 +328,11 @@ def _handle_load(backend, config: dict, resp_queue: Any) -> None:
         trust_remote_code = config.get("trust_remote_code", False)
         if not trust_remote_code:
             model_name = config["model_name"]
-            if "nemotron" in model_name.lower():
+            _mn_lower = model_name.lower()
+            if (
+                "nemotron" in _mn_lower
+                and (_mn_lower.startswith("unsloth/") or _mn_lower.startswith("nvidia/"))
+            ):
                 trust_remote_code = True
                 logger.info(
                     "Auto-enabled trust_remote_code for Nemotron model: %s",
diff --git a/studio/backend/core/training/worker.py b/studio/backend/core/training/worker.py
index 54667dd7af..fa05817697 100644
--- a/studio/backend/core/training/worker.py
+++ b/studio/backend/core/training/worker.py
@@ -405,7 +405,11 @@ def run_training_process(
     # (Qwen3.5, Gemma 4, etc.) are native and do NOT need it — enabling it
     # bypasses the compiler (disabling fused CE).
     _lowered = model_name.lower()
-    if "nemotron" in _lowered and not config.get("trust_remote_code", False):
+    if (
+        "nemotron" in _lowered
+        and (_lowered.startswith("unsloth/") or _lowered.startswith("nvidia/"))
+        and not config.get("trust_remote_code", False)
+    ):
         config["trust_remote_code"] = True
         logger.info(
             "Auto-enabled trust_remote_code for Nemotron model: %s",
diff --git a/studio/backend/tests/test_transformers_version.py b/studio/backend/tests/test_transformers_version.py
index c031c2fea3..609a154a9a 100644
--- a/studio/backend/tests/test_transformers_version.py
+++ b/studio/backend/tests/test_transformers_version.py
@@ -32,8 +32,9 @@
     _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 @@ def test_result_is_cached(self, tmp_path: Path):
 
         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 @@ def test_550_checked_before_530(self):
         # 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 7ba20a366b..308f416940 100644
--- a/studio/backend/utils/transformers_version.py
+++ b/studio/backend/utils/transformers_version.py
@@ -60,14 +60,25 @@
     "gemma4",  # Gemma-4 alternate naming
 )
 
-# Architecture classes / model_type values that require transformers 5.5.0.
-# Checked via config.json (local or HuggingFace).
+# Architecture classes that require transformers 5.5.0.
 _TRANSFORMERS_550_ARCHITECTURES: set[str] = {
     "Gemma4ForConditionalGeneration",
 }
+
+# model_type values (from config.json) → tier mapping.
 _TRANSFORMERS_550_MODEL_TYPES: set[str] = {
     "gemma4",
 }
+_TRANSFORMERS_530_MODEL_TYPES: set[str] = {
+    "qwen3_moe",
+    "qwen3_5_moe",
+    "qwen3_vl_moe",
+    "qwen3_next",
+    "deepseek_v3_moe",
+    "glm4_moe",
+    "glm4_moe_lite",
+    "ministral",
+}
 
 # Tokenizer classes that only exist in transformers>=5.x
 _TRANSFORMERS_5_TOKENIZER_CLASSES: set[str] = {
@@ -77,8 +88,8 @@
 # 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"
@@ -219,43 +230,28 @@ def _check_tokenizer_config_needs_v5(model_name: str) -> bool:
         return False
 
 
-def _check_config_needs_550(model_name: str) -> bool:
-    """Check ``config.json`` for architectures or model_type that require
-    transformers 5.5.0 (e.g. Gemma 4).
+_SENTINEL = object()  # distinguishes "not cached" from "cached as None"
 
-    Checks locally first, then falls back to fetching from HuggingFace.
-    Results are cached in ``_config_needs_550_cache``.
-    Returns False on any error (fail-open to lower tier).
-    """
-    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
+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``.
+    """
+    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)
 
@@ -267,21 +263,26 @@ def _check_cfg(cfg: dict) -> 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:
@@ -291,19 +292,35 @@ def get_transformers_tier(model_name: str) -> str:
     ``"530"`` for models needing transformers 5.3.0 (e.g. Ministral-3, Qwen3 MoE),
     or ``"default"`` for everything else (4.57.x).
 
-    The 5.5.0 check runs first, then 5.3.0.
+    Fast path: substring checks (no I/O) for both tiers run first.
+    Slow path: single config.json fetch (cached) checks model_type for
+    both tiers, then tokenizer_config.json as final fallback.
     """
     lowered = model_name.lower()
 
-    # --- 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"
+
+    # --- 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 fb43d468e25379f28dd2477e6c24dd60cf55c099 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 07/17] [pre-commit.ci] auto fixes from pre-commit.com hooks

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

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

From 81d2581d2d58fbe7c50aae7c47739a0c9ad07db4 Mon Sep 17 00:00:00 2001
From: Roland Tannous <rolandtannous@gravityq.ai>
Date: Mon, 6 Apr 2026 20:26:32 +0000
Subject: [PATCH 08/17] Revert "[pre-commit.ci] auto fixes from pre-commit.com
 hooks"

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

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

From 4df9e68e079c28aea834714577676f796bd85177 Mon Sep 17 00:00:00 2001
From: Roland Tannous <rolandtannous@gravityq.ai>
Date: Mon, 6 Apr 2026 20:26:32 +0000
Subject: [PATCH 09/17] Revert "use config.json model_type for tier detection,
 add unsloth/nvidia namespace guard"

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

diff --git a/studio/backend/core/inference/worker.py b/studio/backend/core/inference/worker.py
index 506e631c51..1010e56dac 100644
--- a/studio/backend/core/inference/worker.py
+++ b/studio/backend/core/inference/worker.py
@@ -328,11 +328,7 @@ def _handle_load(backend, config: dict, resp_queue: Any) -> None:
         trust_remote_code = config.get("trust_remote_code", False)
         if not trust_remote_code:
             model_name = config["model_name"]
-            _mn_lower = model_name.lower()
-            if (
-                "nemotron" in _mn_lower
-                and (_mn_lower.startswith("unsloth/") or _mn_lower.startswith("nvidia/"))
-            ):
+            if "nemotron" in model_name.lower():
                 trust_remote_code = True
                 logger.info(
                     "Auto-enabled trust_remote_code for Nemotron model: %s",
diff --git a/studio/backend/core/training/worker.py b/studio/backend/core/training/worker.py
index fa05817697..54667dd7af 100644
--- a/studio/backend/core/training/worker.py
+++ b/studio/backend/core/training/worker.py
@@ -405,11 +405,7 @@ def run_training_process(
     # (Qwen3.5, Gemma 4, etc.) are native and do NOT need it — enabling it
     # bypasses the compiler (disabling fused CE).
     _lowered = model_name.lower()
-    if (
-        "nemotron" in _lowered
-        and (_lowered.startswith("unsloth/") or _lowered.startswith("nvidia/"))
-        and not config.get("trust_remote_code", False)
-    ):
+    if "nemotron" in _lowered and not config.get("trust_remote_code", False):
         config["trust_remote_code"] = True
         logger.info(
             "Auto-enabled trust_remote_code for Nemotron model: %s",
diff --git a/studio/backend/tests/test_transformers_version.py b/studio/backend/tests/test_transformers_version.py
index 609a154a9a..c031c2fea3 100644
--- a/studio/backend/tests/test_transformers_version.py
+++ b/studio/backend/tests/test_transformers_version.py
@@ -32,9 +32,8 @@
     _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 @@ def test_result_is_cached(self, tmp_path: Path):
 
         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 @@ def test_550_checked_before_530(self):
         # This shouldn't happen in practice, but verifies priority
         assert get_transformers_tier("gemma-4-model") == "550"
 
-    def test_config_json_model_type_530(self, tmp_path: Path):
-        """Local checkpoint with qwen3_moe model_type → 530."""
-        cfg = {"model_type": "qwen3_moe", "architectures": ["Qwen3MoeForCausalLM"]}
-        (tmp_path / "config.json").write_text(json.dumps(cfg))
-
-        assert get_transformers_tier(str(tmp_path)) == "530"
-
-    def test_config_json_model_type_glm4_moe(self, tmp_path: Path):
-        """Local checkpoint with glm4_moe model_type → 530."""
-        cfg = {"model_type": "glm4_moe", "architectures": ["Glm4MoeForCausalLM"]}
-        (tmp_path / "config.json").write_text(json.dumps(cfg))
-
-        assert get_transformers_tier(str(tmp_path)) == "530"
-
     def test_needs_transformers_5_compat(self):
         """needs_transformers_5 should return True for both 530 and 550 models."""
         assert needs_transformers_5("google/gemma-4-E2B-it") is True
diff --git a/studio/backend/utils/transformers_version.py b/studio/backend/utils/transformers_version.py
index 308f416940..7ba20a366b 100644
--- a/studio/backend/utils/transformers_version.py
+++ b/studio/backend/utils/transformers_version.py
@@ -60,25 +60,14 @@
     "gemma4",  # Gemma-4 alternate naming
 )
 
-# Architecture classes that require transformers 5.5.0.
+# Architecture classes / model_type values that require transformers 5.5.0.
+# Checked via config.json (local or HuggingFace).
 _TRANSFORMERS_550_ARCHITECTURES: set[str] = {
     "Gemma4ForConditionalGeneration",
 }
-
-# model_type values (from config.json) → tier mapping.
 _TRANSFORMERS_550_MODEL_TYPES: set[str] = {
     "gemma4",
 }
-_TRANSFORMERS_530_MODEL_TYPES: set[str] = {
-    "qwen3_moe",
-    "qwen3_5_moe",
-    "qwen3_vl_moe",
-    "qwen3_next",
-    "deepseek_v3_moe",
-    "glm4_moe",
-    "glm4_moe_lite",
-    "ministral",
-}
 
 # Tokenizer classes that only exist in transformers>=5.x
 _TRANSFORMERS_5_TOKENIZER_CLASSES: set[str] = {
@@ -88,8 +77,8 @@
 # Cache for dynamic tokenizer_config.json lookups to avoid repeated fetches
 _tokenizer_class_cache: dict[str, bool] = {}
 
-# Cache for config.json lookups (returns the parsed dict or None)
-_config_json_cache: dict[str, dict | None] = {}
+# Cache for dynamic config.json lookups (architecture/model_type checks)
+_config_needs_550_cache: dict[str, bool] = {}
 
 # Versions
 TRANSFORMERS_550_VERSION = "5.5.0"
@@ -230,28 +219,43 @@ def _check_tokenizer_config_needs_v5(model_name: str) -> bool:
         return False
 
 
-_SENTINEL = object()  # distinguishes "not cached" from "cached as None"
-
-
-def _get_config_json(model_name: str) -> dict | None:
-    """Read and cache ``config.json`` for *model_name*.
+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).
 
-    Checks local path first, then fetches from HuggingFace.
-    Returns the parsed dict, or ``None`` on any error (fail-open).
-    The result is cached in ``_config_json_cache``.
+    Checks locally first, then falls back to fetching from HuggingFace.
+    Results are cached in ``_config_needs_550_cache``.
+    Returns False on any error (fail-open to lower tier).
     """
-    cached = _config_json_cache.get(model_name, _SENTINEL)
-    if cached is not _SENTINEL:
-        return cached
+    if model_name in _config_needs_550_cache:
+        return _config_needs_550_cache[model_name]
+
+    def _check_cfg(cfg: dict) -> bool:
+        archs = cfg.get("architectures", [])
+        if any(a in _TRANSFORMERS_550_ARCHITECTURES for a in archs):
+            return True
+        if cfg.get("model_type") in _TRANSFORMERS_550_MODEL_TYPES:
+            return True
+        return False
 
     # --- Check local config.json first ------------------------------------
-    local_cfg = Path(model_name) / "config.json"
+    local_path = Path(model_name)
+    local_cfg = local_path / "config.json"
     if local_cfg.is_file():
         try:
             with open(local_cfg) as f:
                 cfg = json.load(f)
-            _config_json_cache[model_name] = cfg
-            return cfg
+            result = _check_cfg(cfg)
+            if result:
+                logger.info(
+                    "Local config.json check: %s needs transformers 5.5.0 "
+                    "(architectures=%s, model_type=%s)",
+                    model_name,
+                    cfg.get("architectures", []),
+                    cfg.get("model_type"),
+                )
+            _config_needs_550_cache[model_name] = result
+            return result
         except Exception as exc:
             logger.debug("Could not read %s: %s", local_cfg, exc)
 
@@ -263,26 +267,21 @@ def _get_config_json(model_name: str) -> dict | None:
         req = urllib.request.Request(url, headers = {"User-Agent": "unsloth-studio"})
         with urllib.request.urlopen(req, timeout = 10) as resp:
             cfg = json.loads(resp.read().decode())
-        _config_json_cache[model_name] = cfg
-        return cfg
+        result = _check_cfg(cfg)
+        if result:
+            logger.info(
+                "Dynamic config.json check: %s needs transformers 5.5.0 "
+                "(architectures=%s, model_type=%s)",
+                model_name,
+                cfg.get("architectures", []),
+                cfg.get("model_type"),
+            )
+        _config_needs_550_cache[model_name] = result
+        return result
     except Exception as exc:
-        logger.debug(
-            "Could not fetch config.json for '%s': %s", model_name, exc
-        )
-        _config_json_cache[model_name] = None
-        return None
-
-
-def _check_config_needs_550(model_name: str) -> bool:
-    """Check ``config.json`` for architectures or model_type that require
-    transformers 5.5.0.  Uses the shared ``_get_config_json`` cache."""
-    cfg = _get_config_json(model_name)
-    if cfg is None:
+        logger.debug("Could not fetch config.json for '%s': %s", model_name, exc)
+        _config_needs_550_cache[model_name] = False
         return False
-    archs = cfg.get("architectures", [])
-    if any(a in _TRANSFORMERS_550_ARCHITECTURES for a in archs):
-        return True
-    return cfg.get("model_type") in _TRANSFORMERS_550_MODEL_TYPES
 
 
 def get_transformers_tier(model_name: str) -> str:
@@ -292,35 +291,19 @@ def get_transformers_tier(model_name: str) -> str:
     ``"530"`` for models needing transformers 5.3.0 (e.g. Ministral-3, Qwen3 MoE),
     or ``"default"`` for everything else (4.57.x).
 
-    Fast path: substring checks (no I/O) for both tiers run first.
-    Slow path: single config.json fetch (cached) checks model_type for
-    both tiers, then tokenizer_config.json as final fallback.
+    The 5.5.0 check runs first, then 5.3.0.
     """
     lowered = model_name.lower()
 
-    # --- Fast substring checks (no I/O) -----------------------------------
+    # --- Check 5.5.0 first ------------------------------------------------
     if any(sub in lowered for sub in TRANSFORMERS_550_MODEL_SUBSTRINGS):
         return "550"
+    if _check_config_needs_550(model_name):
+        return "550"
+
+    # --- Check 5.3.0 ------------------------------------------------------
     if any(sub in lowered for sub in TRANSFORMERS_5_MODEL_SUBSTRINGS):
         return "530"
-
-    # --- config.json model_type / architecture check (single fetch) -------
-    cfg = _get_config_json(model_name)
-    if cfg is not None:
-        model_type = cfg.get("model_type", "")
-        archs = cfg.get("architectures", [])
-
-        # Check 5.5.0 first
-        if model_type in _TRANSFORMERS_550_MODEL_TYPES:
-            return "550"
-        if any(a in _TRANSFORMERS_550_ARCHITECTURES for a in archs):
-            return "550"
-
-        # Check 5.3.0
-        if model_type in _TRANSFORMERS_530_MODEL_TYPES:
-            return "530"
-
-    # --- Final fallback: tokenizer_config.json for 5.3.0 ------------------
     if _check_tokenizer_config_needs_v5(model_name):
         return "530"
 

From 96dd82fdf5c8cf1cdbbb929f3815aee42906bb54 Mon Sep 17 00:00:00 2001
From: Roland Tannous <rolandtannous@gravityq.ai>
Date: Mon, 6 Apr 2026 20:27:15 +0000
Subject: [PATCH 10/17] add unsloth/nvidia namespace guard to Nemotron
 trust_remote_code auto-enable

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

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

From 35226e67f90825f443a6e4030236a6308c66ac9f 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 11/17] [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 82dbf70910c37593e11d49c9c62dc2447412ca3c Mon Sep 17 00:00:00 2001
From: Roland Tannous <rolandtannous@gravityq.ai>
Date: Mon, 6 Apr 2026 20:31:06 +0000
Subject: [PATCH 12/17] reorder tier checks: all substring matches before
 config.json fetches

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

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

From 30bc604cadcd830dcd89d3ebac620a2600199c0e Mon Sep 17 00:00:00 2001
From: Roland Tannous <rolandtannous@gravityq.ai>
Date: Mon, 6 Apr 2026 22:35:50 +0000
Subject: [PATCH 13/17] extract shared activate_transformers_for_subprocess
 into transformers_version.py

---
 studio/backend/core/export/worker.py         | 41 ++-----------------
 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(+), 114 deletions(-)

diff --git a/studio/backend/core/export/worker.py b/studio/backend/core/export/worker.py
index 4bdfd13b78..c38c925476 100644
--- a/studio/backend/core/export/worker.py
+++ b/studio/backend/core/export/worker.py
@@ -30,50 +30,15 @@
 
 
 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,
-    )
-
-    resolved = _resolve_base_model(model_name)
-    tier = get_transformers_tier(resolved)
+    from utils.transformers_version import activate_transformers_for_subprocess
 
-    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 _send_response(resp_queue: Any, response: dict) -> None:
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 @@
 
 
 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,
-    )
-
-    resolved = _resolve_base_model(model_name)
-    tier = get_transformers_tier(resolved)
+    from utils.transformers_version import activate_transformers_for_subprocess
 
-    if tier == "550":
-        if not _ensure_venv_t5_550_exists():
-            raise RuntimeError(
-                f"Cannot activate transformers 5.5.0: .venv_t5_550 missing at {_VENV_T5_550_DIR}"
-            )
-        if _VENV_T5_550_DIR not in sys.path:
-            sys.path.insert(0, _VENV_T5_550_DIR)
-        logger.info("Activated transformers 5.5.0 from %s", _VENV_T5_550_DIR)
-        _pp = os.environ.get("PYTHONPATH", "")
-        os.environ["PYTHONPATH"] = _VENV_T5_550_DIR + (os.pathsep + _pp if _pp else "")
-    elif tier == "530":
-        if not _ensure_venv_t5_530_exists():
-            raise RuntimeError(
-                f"Cannot activate transformers 5.3.0: .venv_t5_530 missing at {_VENV_T5_530_DIR}"
-            )
-        if _VENV_T5_530_DIR not in sys.path:
-            sys.path.insert(0, _VENV_T5_530_DIR)
-        logger.info("Activated transformers 5.3.0 from %s", _VENV_T5_530_DIR)
-        _pp = os.environ.get("PYTHONPATH", "")
-        os.environ["PYTHONPATH"] = _VENV_T5_530_DIR + (os.pathsep + _pp if _pp else "")
-    else:
-        logger.info("Using default transformers (4.57.x) for %s", model_name)
+    activate_transformers_for_subprocess(model_name)
 
 
 def _decode_image(image_base64: str):
diff --git a/studio/backend/core/training/worker.py b/studio/backend/core/training/worker.py
index fa05817697..bfff599001 100644
--- a/studio/backend/core/training/worker.py
+++ b/studio/backend/core/training/worker.py
@@ -306,50 +306,15 @@ def _ensure_mamba_ssm(event_queue: Any, model_name: str) -> None:
 
 
 def _activate_transformers_version(model_name: str) -> None:
-    """Activate the correct transformers version BEFORE any ML imports.
-
-    Uses get_transformers_tier() to decide between .venv_t5_550/ (5.5.0),
-    .venv_t5_530/ (5.3.0), or the default 4.57.x.
-    """
+    """Activate the correct transformers version BEFORE any ML imports."""
     # Ensure backend is on path for utils imports
     backend_path = str(Path(__file__).resolve().parent.parent.parent)
     if backend_path not in sys.path:
         sys.path.insert(0, backend_path)
 
-    from utils.transformers_version import (
-        get_transformers_tier,
-        _resolve_base_model,
-        _ensure_venv_t5_530_exists,
-        _ensure_venv_t5_550_exists,
-        _VENV_T5_530_DIR,
-        _VENV_T5_550_DIR,
-    )
-
-    resolved = _resolve_base_model(model_name)
-    tier = get_transformers_tier(resolved)
+    from utils.transformers_version import activate_transformers_for_subprocess
 
-    if tier == "550":
-        if not _ensure_venv_t5_550_exists():
-            raise RuntimeError(
-                f"Cannot activate transformers 5.5.0: .venv_t5_550 missing at {_VENV_T5_550_DIR}"
-            )
-        if _VENV_T5_550_DIR not in sys.path:
-            sys.path.insert(0, _VENV_T5_550_DIR)
-        logger.info("Activated transformers 5.5.0 from %s", _VENV_T5_550_DIR)
-        _pp = os.environ.get("PYTHONPATH", "")
-        os.environ["PYTHONPATH"] = _VENV_T5_550_DIR + (os.pathsep + _pp if _pp else "")
-    elif tier == "530":
-        if not _ensure_venv_t5_530_exists():
-            raise RuntimeError(
-                f"Cannot activate transformers 5.3.0: .venv_t5_530 missing at {_VENV_T5_530_DIR}"
-            )
-        if _VENV_T5_530_DIR not in sys.path:
-            sys.path.insert(0, _VENV_T5_530_DIR)
-        logger.info("Activated transformers 5.3.0 from %s", _VENV_T5_530_DIR)
-        _pp = os.environ.get("PYTHONPATH", "")
-        os.environ["PYTHONPATH"] = _VENV_T5_530_DIR + (os.pathsep + _pp if _pp else "")
-    else:
-        logger.info("Using default transformers (4.57.x) for %s", model_name)
+    activate_transformers_for_subprocess(model_name)
 
 
 def run_training_process(
diff --git a/studio/backend/utils/transformers_version.py b/studio/backend/utils/transformers_version.py
index 8403d9cc2f..0517a5f378 100644
--- a/studio/backend/utils/transformers_version.py
+++ b/studio/backend/utils/transformers_version.py
@@ -94,6 +94,49 @@
 _VENV_T5_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 18db38b0a275ee018b8cd05a074f1de41a53efa1 Mon Sep 17 00:00:00 2001
From: Roland Tannous <rolandtannous@gravityq.ai>
Date: Mon, 6 Apr 2026 22:45:22 +0000
Subject: [PATCH 14/17] narrow Nemotron trust_remote_code to
 nemotron_h/nemotron-3-nano, add to export worker

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

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

From 1cc14f41d1a6e989dbdb74665a93a496c81662c6 Mon Sep 17 00:00:00 2001
From: Roland Tannous <rolandtannous@gravityq.ai>
Date: Mon, 6 Apr 2026 22:46:14 +0000
Subject: [PATCH 15/17] clean venv_t5 dirs before re-install in setup.sh,
 clarify version alias comment

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

diff --git a/studio/backend/utils/transformers_version.py b/studio/backend/utils/transformers_version.py
index 0517a5f378..62176dc9b4 100644
--- a/studio/backend/utils/transformers_version.py
+++ b/studio/backend/utils/transformers_version.py
@@ -84,7 +84,8 @@
 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 72df82f601..3040569f54 100755
--- a/studio/setup.sh
+++ b/studio/setup.sh
@@ -514,6 +514,7 @@ 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 "$VENV_T5_530_DIR" --no-deps "huggingface_hub==1.8.0"
@@ -521,6 +522,7 @@ if [ "$_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 5445fa6e16e9233179480bf19c5fd7fdc3459938 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 16/17] [pre-commit.ci] auto fixes from pre-commit.com hooks

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

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

From 45e3a02cc5cbaa7871a5588a84a27191a0ff2eb0 Mon Sep 17 00:00:00 2001
From: Roland Tannous <rolandtannous@gravityq.ai>
Date: Tue, 7 Apr 2026 07:16:26 +0000
Subject: [PATCH 17/17] run venv_t5 migration outside deps fast-path gate in
 both setup scripts

---
 studio/setup.ps1 | 39 ++++++++++++++++++++++++---------------
 studio/setup.sh  | 34 +++++++++++++++++++++++-----------
 2 files changed, 47 insertions(+), 26 deletions(-)

diff --git a/studio/setup.ps1 b/studio/setup.ps1
index fa853cdadf..c3a8cd71ca 100644
--- a/studio/setup.ps1
+++ b/studio/setup.ps1
@@ -1579,23 +1579,37 @@ 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 ""
+} else {
+    step "python" "dependencies up to date"
+    # Restore ErrorActionPreference (was lowered for pip/python section)
+    $ErrorActionPreference = $prevEAP
+}
 
-# Clean up legacy single .venv_t5 directory
+# ── Pre-install transformers 5.x into .venv_t5_530/ and .venv_t5_550/ ──
+# Runs outside the deps fast-path gate so that upgrades from the legacy
+# single .venv_t5 are always migrated to the tiered layout.
+$VenvT5_530Dir = Join-Path $env:USERPROFILE ".unsloth\studio\.venv_t5_530"
+$VenvT5_550Dir = Join-Path $env:USERPROFILE ".unsloth\studio\.venv_t5_550"
 $VenvT5Legacy = Join-Path $env:USERPROFILE ".unsloth\studio\.venv_t5"
-if (Test-Path $VenvT5Legacy) { Remove-Item -Recurse -Force $VenvT5Legacy }
+
+$_NeedT5Install = $false
+if (Test-Path $VenvT5Legacy) {
+    Remove-Item -Recurse -Force $VenvT5Legacy
+    $_NeedT5Install = $true
+}
+if (-not (Test-Path $VenvT5_530Dir)) { $_NeedT5Install = $true }
+if (-not (Test-Path $VenvT5_550Dir)) { $_NeedT5Install = $true }
+# Also reinstall when python deps were updated
+if (-not $SkipPythonDeps) { $_NeedT5Install = $true }
+
+if ($_NeedT5Install) {
+Write-Host ""
 
 $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")) {
@@ -1629,7 +1643,6 @@ 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")) {
@@ -1662,11 +1675,7 @@ if ($tiktokenInstallExit -ne 0) {
 $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)
-    $ErrorActionPreference = $prevEAP
-}
+} # end $_NeedT5Install
 
 # ==========================================================================
 #  PHASE 3.4: Prefer prebuilt llama.cpp bundles before source build
diff --git a/studio/setup.sh b/studio/setup.sh
index 3040569f54..eb89e583ca 100755
--- a/studio/setup.sh
+++ b/studio/setup.sh
@@ -504,16 +504,31 @@ fi
 
 if [ "$_SKIP_PYTHON_DEPS" = false ]; then
     install_python_stack
+else
+    step "python" "dependencies up to date"
+    verbose_substep "python deps check: installed=$_PKG_NAME@${INSTALLED_VER:-unknown} latest=${LATEST_VER:-unknown}"
+fi
 
-    # ── 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"
+# ── 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.
+#
+# Runs outside the _SKIP_PYTHON_DEPS gate so that upgrades from legacy
+# single .venv_t5 are always migrated to the tiered layout.
+_NEED_T5_INSTALL=false
+if [ -d "$STUDIO_HOME/.venv_t5" ]; then
+    # Legacy layout — migrate
+    rm -rf "$STUDIO_HOME/.venv_t5"
+    _NEED_T5_INSTALL=true
+fi
+[ ! -d "$VENV_T5_530_DIR" ] && _NEED_T5_INSTALL=true
+[ ! -d "$VENV_T5_550_DIR" ] && _NEED_T5_INSTALL=true
+# Also reinstall when python deps were updated (packages may need rebuild)
+[ "$_SKIP_PYTHON_DEPS" = false ] && _NEED_T5_INSTALL=true
 
+if [ "$_NEED_T5_INSTALL" = true ]; then
     [ -d "$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 "$VENV_T5_530_DIR" --no-deps "transformers==5.3.0"
@@ -529,9 +544,6 @@ if [ "$_SKIP_PYTHON_DEPS" = false ]; then
     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}"
 fi
 fi