fix: un-blocklist deepseek_v3 in transformers tokenizer class set

nemo_rl/models/megatron/setup.py

@@ -1206,11 +1206,10 @@ def setup_model_and_optimizer(
     # Context used for persisting some state between checkpoint saves.
     checkpointing_context = init_checkpointing_context(megatron_cfg.checkpoint)
 
-    # Tokenizer
-    if megatron_cfg.tokenizer.hf_tokenizer_kwargs is None:
-        megatron_cfg.tokenizer.hf_tokenizer_kwargs = {}
-    megatron_cfg.tokenizer.hf_tokenizer_kwargs["trust_remote_code"] = True
-    megatron_cfg.tokenizer.hf_tokenizer_kwargs["use_fast"] = True
+    # Set the attribute directly instead of updating hf_tokenizer_kwargs, because
+    # Megatron-Bridge's TokenizerConfig snapshots hf_tokenizer_kwargs into plain
+    # attributes at __post_init__ and never re-reads the dict afterwards.
+    megatron_cfg.tokenizer.trust_remote_code = True
     build_tokenizer(
         megatron_cfg.tokenizer,
         make_vocab_size_divisible_by=megatron_cfg.model.make_vocab_size_divisible_by

nemo_rl/models/policy/__init__.py

@@ -18,6 +18,55 @@
 from nemo_rl.utils.checkpoint import PretrainedCheckpointConfig
 
 
+def _patch_transformers_tokenizer_class_set():
+    """Undo the transformers block on deepseek_v3 tokenizers.
+
+    Root cause: transformers 5.4-5.11 lists "deepseek_v3" in two internal
+    registries -- MODELS_WITH_INCORRECT_HUB_TOKENIZER_CLASS (a set) and
+    TOKENIZER_MAPPING_NAMES (a dict pinning it to "TokenizersBackend"). Together
+    they force the fast tokenizer backend and suppress trust_remote_code, so
+    AutoTokenizer can only load via a local tokenizer.json. Models like
+    Moonlight-16B-A3B ship no tokenizer.json (only tiktoken.model + a remote-code
+    TikTokenTokenizer), so offline loading fails.
+
+    Removing both entries restores the trust_remote_code / auto_map path.
+    discard/pop-with-default are no-ops when the entries are absent, so this is
+    safe on any transformers version in the currently-supported range.
+
+    Placed here (nemo_rl/models/policy/__init__.py) so it fires exactly once
+    per process the first time any policy code is imported -- covers the driver
+    (via nemo_rl.algorithms.grpo) and every policy worker (Megatron / DTensor /
+    DTensor v2 all import from nemo_rl.models.policy) without polluting nemo_rl
+    consumers that don't touch tokenizers.
+    """
+    import transformers
+    from packaging.version import Version as PkgVersion
+
+    # This whole patch exists only because Megatron-Bridge caps the transformers
+    # upper bound below 5.9 today, which forces us onto a transformers version
+    # that still has the deepseek_v3 tokenizer-blocklist bug. Once MBridge relaxes
+    # its transformers upper bound to >=5.12, we can drop this workaround.
+    # TODO: remove this patch (and the assert below) once MBridge relaxes its
+    # transformers upper bound past the deepseek_v3 fix (~transformers 5.12).
+    # https://github.com/NVIDIA-NeMo/RL/issues/2764
+    assert PkgVersion(transformers.__version__) < PkgVersion("5.12.0"), (
+        f"transformers {transformers.__version__} detected. "
+        "The deepseek_v3 tokenizer-blocklist patch was written for <5.12. "
+        "Check if the upstream fix now applies and remove this patch if so."
+    )
+
+    from transformers.models.auto.tokenization_auto import (
+        MODELS_WITH_INCORRECT_HUB_TOKENIZER_CLASS,
+        TOKENIZER_MAPPING_NAMES,
+    )
+
+    MODELS_WITH_INCORRECT_HUB_TOKENIZER_CLASS.discard("deepseek_v3")
+    TOKENIZER_MAPPING_NAMES.pop("deepseek_v3", None)
+
+
+_patch_transformers_tokenizer_class_set()
+
+
 class LoRAConfigDisabled(TypedDict):
     enabled: Literal[False]