NemotronHBridge missing get_hf_tokenizer_kwargs causes checkpoint conversion failure

[shanecmoran]

Description

Nemotron Nano v2 models (nvidia/NVIDIA-Nemotron-Nano-9B-v2, nvidia/NVIDIA-Nemotron-Nano-12B-v2) fail during convert_checkpoints.py import at the tokenizer-save step:

AttributeError: 'bool' object has no attribute 'chat_template'

self.tokenizer is False instead of a HuggingFace tokenizer object.

Root Cause

Three factors combine:

1. The nvcr.io/nvidia/nemo:26.02 container's megatron-core (core_r0.16.0) defaults HuggingFaceTokenizer.__init__ to use_fast=False
2. Nemotron Nano v2 only ships a fast tokenizer (tokenizer.json, no tokenizer.model), so AutoTokenizer.from_pretrained(..., use_fast=False) silently returns False
3. NemotronHBridge does not implement get_hf_tokenizer_kwargs() to override this default

GLM45VBridge already solves this exact problem by implementing get_hf_tokenizer_kwargs() returning {"use_fast": True}.

Fix

Add get_hf_tokenizer_kwargs to NemotronHBridge, following the GLM45VBridge pattern. The call site in auto_bridge.py:768 already checks for this method via hasattr — no other changes needed.

[yaoyu-33]

Your fix in #2664 looks correct — we'll review and merge it.