cc @ArthurZucker for text models

@ArthurZucker please check this out! I've tested with MiniCPM3-4B and it worked well.

@ArthurZucker I've tested MiniCPM3-4B and our in-house model for the case q_lora_rank is None. Please review this when you are available and let me know if more needed! thanks a lot :) I hope this can be merged because we are going to open-weights one of models using this modeling in the near future.

End-to-end check on openbmb/MiniCPM3-4B:

import torch
from transformers import AutoTokenizer, MiniCPM3ForCausalLM

tokenizer = AutoTokenizer.from_pretrained("openbmb/MiniCPM3-4B")
model = MiniCPM3ForCausalLM.from_pretrained(
    "openbmb/MiniCPM3-4B",
    dtype=torch.float16,
    device_map="auto",
    attn_implementation="eager",
).eval()

messages = [{"role": "user", "content": "Tell me a fun fact about the Eiffel Tower."}]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to(model.device)
out = model.generate(**inputs, max_new_tokens=80, do_sample=False)
print(tokenizer.decode(out[0], skip_special_tokens=True))

Output:

user
Tell me a fun fact about the Eiffel Tower.
 assistant
The Eiffel Tower was originally intended to be a temporary structure, designed for the 1889 World's Fair to celebrate the 100th anniversary of the French Revolution. It was meant to be a symbol of the Industrial Revolution and a demonstration of France's engineering prowess. However, it was so popular that it was allowed to stand after the Fair. It

@aliyevaladddin Sorry I missed your email earlier. I just rebased this PR, would be happy to add you as coauthor since your message got me to come back to it.

@ArthurZucker @Cyrilvallez when you have a moment, would appreciate a review on this. Thanks!

Thanks for the rebase, @bzantium! I'm honored to be added as a co-author. I'm really excited to see MiniCPM3 support getting closer to being merged. Let me know if there's anything else I can help with regarding testing or further updates!

Will try to take a look today 🤗

View the CircleCI Test Summary for this PR:

https://huggingface.co/spaces/transformers-community/circle-ci-viz?pr=41116&sha=42b80e

We also need to sync with main and fix CI please

@vasqu

The security-gate / security-check failure comes from the "Enforce allowlist" step. The only offending file is docs/source/en/_toctree.yml, which the gate rejects because .yml is not allowed outside tests/ (only .py, .txt, .md are permitted there). That _toctree.yml entry is required to register MiniCPM3 in the docs navigation, so it cannot be dropped.

This gate only runs for fork PRs, and the failure comment notes that a maintainer can manually approve CI when a finding is a false positive. Could you approve CI when you get a chance? All 11 other changed files pass the allowlist. Thanks!

Thank you for your contribution 🤗!

CI Security Gate — automatic approval blocked

This PR was not automatically approved for CI because the security gate failed.

Possible reasons:

• A changed file is outside the allowed directories (src/, tests/, docs/, utils/) or has a disallowed extension (only .py, .txt, .md are permitted outside tests/)
• A new high-severity security issue was detected in the changed Python files (Bandit check)

See the workflow run for the exact violations.

A maintainer can review and manually approve CI if a finding is a false positive.

gentle ping @bzantium

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

run-slow: auto, minicpm3

@vasqu Thanks for the careful check! I verified on an A100 (compute capability 8.0, same hardware the ("cuda", 8) reference came from) using the current branch, which includes your modular cleanup. Nothing is broken.

bf16 reproduction (A100, cc 8.0):

slice : [0.765625, 3.640625, -0.189453, -0.835938, -0.835938]
ref   : [0.765625, 3.640625, -0.189453125, -0.8359375, -0.8359375]
diff  : [0, 0, 0, 0, 0]   # exact match

So your refactor (inheriting Gemma3TextScaledWordEmbedding and dropping the duplicated MiniCPM3Model.forward) is numerically a no op. That matches the code too: the new embedding base is identical to the old one, and the removed forward was a verbatim copy of the inherited LlamaModel.forward.

On the 8 vs 8.6 gap: both buckets run bf16 (the config torch_dtype is bfloat16, and the test loads with dtype="auto"), so the difference is purely cross GPU bf16 variation, not a dtype or code change. These particular logits sit near zero and the model is 62 layers deep, so bf16 rounding accumulates and the low magnitude entries wobble. To put a number on it, I also ran fp32 as a ground truth on the same A100:

fp32  : [0.535951, 3.610042, -0.174161, -1.174419, -1.174420]

The bf16 vs fp32 gap reaches about 0.34, which is larger than the 8.0 vs 8.6 gap of about 0.08. Most importantly, the argmax token is 11213 in bf16 and fp32 alike, so the actual prediction is stable across precision and only the near zero logit tail moves.

Net: your (8, 6) bucket is a legitimate per hardware entry and the existing (8) reference still reproduces exactly. No regression. Good to go from my side.

run-slow: minicpm3

Just sanity checking but should be good to go then 🫡

Workflow Run ⚙️

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

models: ["models/minicpm3"]
quantizations: []

CI Results

Workflow Run ⚙️

Commit Info

✅ No failing test specific to this PR 🎉 👏 !

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.

Nice thanks for sticking through and gz on the merge 🤗

Finaly!