fix: route Granite models to TokenizersBackend to preserve tokenizer.json pre-tokenizer by kndtran · Pull Request #45813 · huggingface/transformers

kndtran · 2026-05-06T18:49:09Z

What does this PR do?

Change TOKENIZER_MAPPING_NAMES for Granite model types from "GPT2Tokenizer" to "TokenizersBackend" so that AutoTokenizer loads tokenizer.json faithfully instead of routing through GPT2Tokenizer.__init__ which hardcodes a wrong pre-tokenizer.

Fixes #45812

Code Agent Policy

The Transformers repo is currently being overwhelmed by a large number of PRs and issue comments written by
code agents. We are currently bottlenecked by our ability to review and respond to them. As a result,
we ask that new users do not submit pure code agent PRs at this time.
You may use code agents in drafting or to help you diagnose issues. We'd also ask autonomous "OpenClaw"-like agents
not to open any PRs or issues for the moment.

PRs that appear to be fully agent-written will probably be closed without review, and we may block users who do this
repeatedly or maliciously.

This is a rapidly-evolving situation that's causing significant shockwaves in the open-source community. As a result,
this policy is likely to be updated regularly in the near future. For more information, please read CONTRIBUTING.md.

I confirm that this is not a pure code agent PR.

Before submitting

This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
Did you read the contributor guideline,
Pull Request section?
Was this discussed/approved via a Github issue or the forum? Please add a link
to it if that's the case.
Did you make sure to update the documentation with your changes? Here are the
documentation guidelines, and
here are tips on formatting docstrings.
Did you write any new necessary tests?

Who can review?

Anyone in the community is free to review the PR once the tests have passed. Feel free to tag
members/contributors who may be interested in your PR.

tokenizers: @ArthurZucker and @itazap

Fixes huggingface#45812

…json pre-tokenizer previous fix was not triggering Fixes huggingface#45812

itazap · 2026-05-07T04:49:09Z

Thank you @kndtran for finding this and producing a fix! I'm looking into why existing tests didn't find this so we should also add / update a test for granite models. Sorry it slipped through the cracks!

ArthurZucker

fix is valid tho! as @itazap said we need to probably add tests (integration) to cover that case

HuggingFaceDocBuilderDev · 2026-05-11T06:36:16Z

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

itazap · 2026-05-12T03:26:34Z

transformers/tests/models/granitemoehybrid/test_modeling_granitemoehybrid.py

Lines 373 to 384 in d4b9110

    
           def test_model_generation(self, device): 
        
               EXPECTED_TEXT_COMPLETION = "Simply put, the theory of relativity states that 1) the laws of physics are the same for all observers in uniform motion relative" 
        
               prompt = "Simply put, the theory of relativity states that " 
        
               tokenizer = AutoTokenizer.from_pretrained("ibm-granite/granite-4.0-h-tiny") 
        
               model = GraniteMoeHybridForCausalLM.from_pretrained("ibm-granite/granite-4.0-h-tiny", device_map=device) 
        
               model_inputs = tokenizer(prompt, return_tensors="pt").to(model.device) 
        
               # greedy generation outputs 
        
               generated_ids = model.generate(**model_inputs, max_new_tokens=16, do_sample=False) 
        
               text = tokenizer.decode(generated_ids[0], skip_special_tokens=True) 
        
               self.assertEqual(EXPECTED_TEXT_COMPLETION, text)

this is the only test we had for a granite-4 model and it did no test digit strings 😢 We can add a simple test with digits to this file - let me know if you'd like to add one or I can!

something like

self.assertEqual(tokenizer.encode("650841823", add_special_tokens=False), EXPECTED_OUTPUT)
etc.

itazap · 2026-05-18T01:44:13Z

@kndtran please let me if you'd like to add a test otherwise I can!

kndtran · 2026-05-18T02:33:30Z

@itazap Ah, I thought you were replying to Arthur. I will add a few tests. There were a few other odd strings too.

Verify that AutoTokenizer produces correct token IDs for digit strings, punctuation, and mixed alphanumeric inputs.

kndtran · 2026-05-18T03:17:42Z

@itazap I added a new test class as the torch gating seemed unnecessary. Please modify as needed.

Here are some tables showing the tokenization effects on the test strings.

Decoded Token Comparison

String	v5 (broken)	v5 (fixed)
`2023`	`['20', '23']`	`['202', '3']`
`650841823`	`['650', '84', '18', '23']`	`['650', '841', '823']`
`60-138-3818`	`['60', '-', '138', '-', '38', '18']`	`['60', '-', '138', '-', '381', '8']`
`d.o.o`	`['d', '.', 'o', '.', 'o']`	`['d', '.o', '.o']`
`FY2023`	`['FY', '20', '23']`	`['FY', '202', '3']`
`ISO 9001:2015`	`['ISO', ' ', '9', '001', ':', '201', '5']`	`['ISO', ' ', '900', '1', ':', '201', '5']`

Token ID Comparison

String	v5 IDs (broken)	v5 IDs (fixed)
`2023`	`[508, 1419]`	`[2366, 18]`
`650841823`	`[13655, 5833, 972, 1419]`	`[13655, 25496, 23848]`
`60-138-3818`	`[1399, 12, 10350, 12, 1987, 972]`	`[1399, 12, 10350, 12, 19162, 23]`
`d.o.o`	`[67, 13, 78, 13, 78]`	`[67, 14778, 14778]`
`FY2023`	`[82029, 508, 1419]`	`[82029, 2366, 18]`
`ISO 9001:2015`	`[25141, 220, 24, 4119, 25, 679, 20]`	`[25141, 220, 7467, 16, 25, 679, 20]`

Reproduce

from transformers import AutoTokenizer, PreTrainedTokenizerFast

MODEL = "ibm-granite/granite-4.0-h-tiny"
STRINGS = ["2023", "650841823", "60-138-3818", "d.o.o", "FY2023", "ISO 9001:2015"]

broken = AutoTokenizer.from_pretrained(MODEL)
fixed = PreTrainedTokenizerFast.from_pretrained(MODEL, use_fast=True)

print(f"{'String':20s} {'v5 (broken)':45s} {'v5 (fixed)'}")
print("─" * 110)
for s in STRINGS:
    b_ids = broken.encode(s, add_special_tokens=False)
    f_ids = fixed.encode(s, add_special_tokens=False)
    b_tok = [broken.decode([i]) for i in b_ids]
    f_tok = [fixed.decode([i]) for i in f_ids]
    print(f"{s:20s} {str(b_tok):45s} {str(f_tok)}")

print()
print(f"{'String':20s} {'v5 IDs (broken)':45s} {'v5 IDs (fixed)'}")
print("─" * 110)
for s in STRINGS:
    b_ids = broken.encode(s, add_special_tokens=False)
    f_ids = fixed.encode(s, add_special_tokens=False)
    print(f"{s:20s} {str(b_ids):45s} {str(f_ids)}")

itazap · 2026-05-18T06:18:16Z

Thank you @kndtran ! Appreciate your thorough investigation on this 🙌

github-actions · 2026-05-18T06:40:04Z

[For maintainers] Suggested jobs to run (before merge)

run-slow: auto, granitemoehybrid

itazap · 2026-05-18T06:41:55Z

run-slow: auto, granitemoehybrid

github-actions · 2026-05-18T06:43:18Z

Workflow Run ⚙️

This comment contains run-slow, running the specified jobs:

models: ["models/auto", "models/granitemoehybrid"]
quantizations: []

github-actions · 2026-05-18T06:54:09Z

CI Results

Workflow Run ⚙️

Commit Info

Context	Commit	Description
RUN	a95b663e	workflow commit (merge commit)
PR	e10d0757	branch commit (from PR)
main	1ae5aaae	base commit (on `main`)

✅ No failing test specific to this PR 🎉 👏 !

…json pre-tokenizer (huggingface#45813) * fix: add Granite models to MODELS_WITH_INCORRECT_HUB_TOKENIZER_CLASS Fixes huggingface#45812 * fix: route Granite models to TokenizersBackend to preserve tokenizer.json pre-tokenizer previous fix was not triggering Fixes huggingface#45812 * test: add tokenizer encoding test for Granite 4+ Verify that AutoTokenizer produces correct token IDs for digit strings, punctuation, and mixed alphanumeric inputs. --------- Co-authored-by: Khoi-Nguyen Tran <kndtran@ibm.com> Co-authored-by: Ita Zaporozhets <31893021+itazap@users.noreply.github.com>

fix: add Granite models to MODELS_WITH_INCORRECT_HUB_TOKENIZER_CLASS

f0a9a73

Fixes huggingface#45812

kndtran mentioned this pull request May 6, 2026

AutoTokenizer produces wrong token IDs for all Granite models (silent v4→v5 regression) #45812

Closed

4 tasks

fix: route Granite models to TokenizersBackend to preserve tokenizer.…

2c4ed38

…json pre-tokenizer previous fix was not triggering Fixes huggingface#45812

kndtran changed the title ~~fix: add Granite models to MODELS_WITH_INCORRECT_HUB_TOKENIZER_CLASS~~ fix: route Granite models to TokenizersBackend to preserve tokenizer.json pre-tokenizer May 6, 2026

ArthurZucker approved these changes May 11, 2026

View reviewed changes

avinash2692 mentioned this pull request May 12, 2026

fix: Bumps to transformers==5.0.0 generative-computing/mellea#418

Merged

11 tasks

kndtran mentioned this pull request May 12, 2026

AutoTokenizer produces wrong token IDs for OLMo2, HyperClovaX, DeepSeek-R1-Distill-Llama, Yi, and others (v5 regression) #45920

Open

4 tasks

itazap mentioned this pull request May 14, 2026

Fix models for which we don't have a dedicated tokenizer class, and the listed one is incorrect #45936

Open

test: add tokenizer encoding test for Granite 4+

e8d7ffe

Verify that AutoTokenizer produces correct token IDs for digit strings, punctuation, and mixed alphanumeric inputs.

itazap added this pull request to the merge queue May 18, 2026

github-merge-queue Bot removed this pull request from the merge queue due to failed status checks May 18, 2026

Merge branch 'main' into fix/granite-tokenizer-pretokenizer

e10d075

itazap added this pull request to the merge queue May 18, 2026

Merged via the queue into huggingface:main with commit b8eea41 May 18, 2026
30 checks passed

Conversation

kndtran commented May 6, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

What does this PR do?

Code Agent Policy

Before submitting

Who can review?

Uh oh!

itazap commented May 7, 2026

Uh oh!

ArthurZucker left a comment

Choose a reason for hiding this comment

Uh oh!

HuggingFaceDocBuilderDev commented May 11, 2026

Uh oh!

itazap commented May 12, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

itazap commented May 18, 2026

Uh oh!

kndtran commented May 18, 2026

Uh oh!

kndtran commented May 18, 2026

Decoded Token Comparison

Token ID Comparison

Uh oh!

itazap commented May 18, 2026

Uh oh!

Uh oh!

github-actions Bot commented May 18, 2026

Uh oh!

itazap commented May 18, 2026

Uh oh!

github-actions Bot commented May 18, 2026

Uh oh!

github-actions Bot commented May 18, 2026

CI Results

Commit Info

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

5 participants

kndtran commented May 6, 2026 •

edited

Loading

itazap commented May 12, 2026 •

edited

Loading