System Info

• transformers version: 5.8.0 (also reproduced on 5.0.0 through 5.7.0)
• Platform: Linux-5.14.0-503.11.1.el9_5.x86_64-x86_64-with-glibc2.34
• Python version: 3.12.13
• Huggingface_hub version: 1.14.0
• Safetensors version: 0.7.0
• Tokenizers version: 0.22.2
• PyTorch version: not installed (tokenizer-only reproduction)
• Using distributed or parallel set-up in script?: No

Who can help?

@ArthurZucker and @itazap

Information

• The official example scripts
• My own modified scripts

Tasks

• An officially supported task in the examples folder (such as GLUE/SQuAD, ...)
• My own task or dataset (give details below)

Reproduction

from transformers import AutoTokenizer, PreTrainedTokenizerFast

model_id = "ibm-granite/granite-4.1-8b"  # any Granite 4+ model

tok_auto = AutoTokenizer.from_pretrained(model_id)
tok_correct = PreTrainedTokenizerFast.from_pretrained(model_id)

# Numeric strings are most visibly affected (digit splitting differs)
print(tok_auto.encode("2023", add_special_tokens=False))       # [508, 1419] ← WRONG
print(tok_correct.encode("2023", add_special_tokens=False))    # [2366, 18]  ← correct

print(tok_auto.encode("650841823", add_special_tokens=False))      # [13655, 5833, 972, 1419] ← WRONG
print(tok_correct.encode("650841823", add_special_tokens=False))   # [13655, 25496, 23848]    ← correct

print(tok_auto.encode("ISO 9001:2015", add_special_tokens=False))      # [25141, 220, 24, 4119, 25, 679, 20] ← WRONG
print(tok_correct.encode("ISO 9001:2015", add_special_tokens=False))   # [25141, 220, 7467, 16, 25, 679, 20] ← correct

Pre-tokenizer mismatch:

print(tok_auto.backend_tokenizer.pre_tokenizer)
# ByteLevel(add_prefix_space=False, trim_offsets=True, use_regex=True)  ← WRONG

print(tok_correct.backend_tokenizer.pre_tokenizer)
# Sequence([Split(regex, ...), ByteLevel(..., use_regex=False)])  ← correct (matches tokenizer.json)

Expected behavior

Granite models ship a tokenizer.json with a Sequence(Split(\p{N}{1,3}) + ByteLevel) pre-tokenizer (tiktoken/cl100k style — splits digit runs into ≤3-character chunks before BPE). But AutoTokenizer routes Granite to GPT2Tokenizer, whose __init__ hardcodes ByteLevel(use_regex=True) — which uses the GPT-2 regex (\p{N}+, keeps all digits together).

Since BPE was trained with \p{N}{1,3} splitting, using a different pre-tokenizer at inference produces wrong merges and wrong token IDs.

Workaround

Use PreTrainedTokenizerFast directly (works in both v4 and v5):

from transformers import PreTrainedTokenizerFast
tok = PreTrainedTokenizerFast.from_pretrained("ibm-granite/granite-4.1-8b")

Proposed Fix

Change TOKENIZER_MAPPING_NAMES for Granite model types from "GPT2Tokenizer" to "TokenizersBackend":

-        ("granite", "GPT2Tokenizer"),
-        ("granitemoe", "GPT2Tokenizer"),
-        ("granitemoehybrid", "GPT2Tokenizer"),
-        ("granitemoeshared", "GPT2Tokenizer"),
+        ("granite", "TokenizersBackend" if is_tokenizers_available() else None),
+        ("granitemoe", "TokenizersBackend" if is_tokenizers_available() else None),
+        ("granitemoehybrid", "TokenizersBackend" if is_tokenizers_available() else None),
+        ("granitemoeshared", "TokenizersBackend" if is_tokenizers_available() else None),

This triggers the existing mismatch detection (mapping says "TokenizersBackend" ≠ hub's "GPT2Tokenizer"), which falls through to TokenizersBackend.from_pretrained() — loading tokenizer.json faithfully.

Why not MODELS_WITH_INCORRECT_HUB_TOKENIZER_CLASS? That override set is only consulted inside the mismatch block (line 733) — which requires TOKENIZER_MAPPING_NAMES[model_type] to disagree with the hub's tokenizer_class. For Granite, both say "GPT2Tokenizer", so no mismatch is detected and the override set is never checked.

Related Issues

• #45488 — Same bug class: LlamaTokenizer hardcodes Metaspace, breaks DeepSeek V3/R1 (GSM8K drops 63.7% → 26.1%)
• #44462 — Same bug class: AutoTokenizer ignores tokenizer.json pre-tokenizer for deepseek-coder
• #43122 — Same bug class: different tokenization between v4.57.3 and v5.0 (MiniMax)
• #45701 — Same bug class: character-level tokenization in v5.7.0 instead of subword (camembert v2)