Add MiniCPM-1B model support

[Belugaaaa]

Before submitting

• Lint code. If there are lint issues, please format the code first.

Validation passed locally:

• pre-commit run --files ...

• bash -n scripts/minicpm/run_ms_swift_gsm8k.sh scripts/minicpm/run_sft_gsm8k.sh tests/integration_test/minicpm_sft_single_card.sh

• python3 -m py_compile scripts/minicpm/*.py paddleformers/nn/norm.py paddleformers/transformers/minicpm/*.py tests/transformers/minicpm/*.py

• Add test cases into tests folder. If there are codecov issues, please add tests cases first.

Added test coverage:

• tests/transformers/minicpm/test_modeling.py
• tests/config/ci/minicpm_sft_single.yaml
• tests/integration_test/minicpm_sft_single_card.sh

PR types

New features

PR changes

Models, Docs

Description

This PR adds PaddleFormers support for openbmb/MiniCPM-1B-sft-bf16, including:

• MiniCPM config and model implementation.
• AutoConfig and AutoModelForCausalLM registration.
• HF-to-Paddle native checkpoint conversion helper.
• Forward/generation alignment helpers.
• GSM8K SFT and ms-swift baseline helper scripts.
• Inference compiler benchmark helper.
• Tiny random checkpoint generation helper for CI/CE.
• Unit tests and single-card SFT CI smoke config.
• PR-facing acceptance status document.

Local validation uses the current best path: full-depth, full-width MiniCPM-1B. The HF pytorch_model.bin checkpoint is converted to Paddle native model_state.pdparams, then loaded through PaddleFormers for alignment and training validation.

Validation results:

• Full MiniCPM-1B FP32 logits alignment:
  • max_diff=4.57763671875e-05
  • mean_diff=8.224584234994836e-06
  • last_max_diff=3.0517578125e-05
  • last_mean_diff=7.700375135755166e-06
• Generation alignment:
  • first_10_tokens_match=True
  • HF and Paddle first 10 generated token ids: [11225, 72, 5, 2219, 8107, 1379, 8360, 1410, 11225, 72]
• GSM8K 300-step PaddleFormers SFT completed:
  • Eval losses: step 50 0.5329, step 100 0.5421, step 150 0.5408, step 200 0.5424, step 250 0.5483, step 300 0.5435
  • Final train metrics: train_loss=0.1506, train_steps_per_second=0.151
• GSM8K 300-step ms-swift baseline completed:
  • Eval losses: step 50 0.58973086, step 100 0.56501245, step 150 0.55724013, step 200 0.55477822, step 250 0.55431318, step 300 0.55419004
  • PaddleFormers and ms-swift eval curves converge into the same range; step-300 eval loss delta is about 0.0107
• Unit tests passed locally:
  • Ran 33 tests ... OK (skipped=5)
• Tiny SFT smoke completed locally:
  • train_loss=11.2081

Known remaining acceptance items:

• CE tiny checkpoint still needs to be uploaded to the official tiny model location used by tests/integration_test/minicpm_sft_single_card.sh.
• Local compiler inference benchmark did not reach the 20% speedup target. Logs indicate missing symbolic shape support for fused_rms_norm_ext; this should be rerun in the official compiler/runtime acceptance environment.