split venv_t5 into tiered 5.3.0/5.5.0 and fix trust_remote_code

studio/backend/core/export/worker.py

@@ -30,37 +30,15 @@
 
 
 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/).
-    """
+    """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 (
-        needs_transformers_5,
-        _resolve_base_model,
-        _ensure_venv_t5_exists,
-        _VENV_T5_DIR,
-    )
+    from utils.transformers_version import activate_transformers_for_subprocess
 
-    resolved = _resolve_base_model(model_name)
-    if needs_transformers_5(resolved):
-        if not _ensure_venv_t5_exists():
-            raise RuntimeError(
-                f"Cannot activate transformers 5.x: .venv_t5 missing at {_VENV_T5_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)
-        _pp = os.environ.get("PYTHONPATH", "")
-        os.environ["PYTHONPATH"] = _VENV_T5_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:
@@ -78,6 +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,

studio/backend/core/inference/worker.py

@@ -34,37 +34,15 @@
 
 
 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/).
-    """
+    """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 (
-        needs_transformers_5,
-        _resolve_base_model,
-        _ensure_venv_t5_exists,
-        _VENV_T5_DIR,
-    )
+    from utils.transformers_version import activate_transformers_for_subprocess
 
-    resolved = _resolve_base_model(model_name)
-    if needs_transformers_5(resolved):
-        if not _ensure_venv_t5_exists():
-            raise RuntimeError(
-                f"Cannot activate transformers 5.x: .venv_t5 missing at {_VENV_T5_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)
-        _pp = os.environ.get("PYTHONPATH", "")
-        os.environ["PYTHONPATH"] = _VENV_T5_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):
@@ -309,19 +287,21 @@ 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 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:
-            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/"
+            _mn_lower = model_name.lower()
+            if any(sub in _mn_lower for sub in _NEMOTRON_TRUST_SUBSTRINGS) and (
+                _mn_lower.startswith("unsloth/") or _mn_lower.startswith("nvidia/")
             ):
                 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,
                 )
 

studio/backend/core/training/worker.py

@@ -306,37 +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.
-
-    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/).
-    """
+    """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 (
-        needs_transformers_5,
-        _resolve_base_model,
-        _ensure_venv_t5_exists,
-        _VENV_T5_DIR,
-    )
+    from utils.transformers_version import activate_transformers_for_subprocess
 
-    resolved = _resolve_base_model(model_name)
-    if needs_transformers_5(resolved):
-        if not _ensure_venv_t5_exists():
-            raise RuntimeError(
-                f"Cannot activate transformers 5.x: .venv_t5 missing at {_VENV_T5_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)
-        _pp = os.environ.get("PYTHONPATH", "")
-        os.environ["PYTHONPATH"] = _VENV_T5_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(
@@ -386,25 +364,22 @@ 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 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
-
+    # ── 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()
-    _is_native_t5 = any(x in _lowered for x in ("gemma-4", "gemma4"))
     if (
-        needs_transformers_5(model_name)
-        and _lowered.startswith("unsloth/")
-        and not _is_native_t5
+        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)
     ):
         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,
         )
 

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 +191,148 @@ 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

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.