fix(testing): resolve tokenizer loading error and UnboundLocalError in logit checker

[chiajunglien]

Description

This PR fixes two regression bugs that cause the mistral-7bnightly E2E test to crash in forward_pass_logit_checker.py:

DAG Error Log

1. UnicodeDecodeError / OSError during tokenizer loading when reading local binary SentencePiece .model files.
2. UnboundLocalError due to model_prefix being referenced outside its defined conditional block for text-only models.

Changes

1. Added mistral-7b Online ID Mapping

• File Modified: src/maxtext/utils/globals.py
• Details: Added "mistral-7b": "mistralai/Mistral-7B-v0.1" to the HF_IDS dictionary. This enables the logit checker to fall back to the correct online Hugging Face repository if local loading fails.

2. Implemented Robust Tokenizer Loading Fallback (Fixes Unicode Crash)

• File Modified: tests/utils/forward_pass_logit_checker.py
• Details:
  1. Imported HF_IDS from globals.py.
  2. Enhanced the tokenizer loading logic in main(): If loading from the local config.tokenizer_path fails (which occurs because tokenizer.mistral-v1 is a binary SentencePiece .model file instead of JSON), the script now catches the exception and falls back to loading the mapped Hugging Face ID from HF_IDS.

3. Fixed model_prefix Variable Scope (Fixes UnboundLocalError Crash)

• File Modified: tests/utils/forward_pass_logit_checker.py
• Details:
  1. Moved the initialization of model_prefix from inside the if config.use_multimodal: block to the very top of get_data().
  2. This ensures model_prefix is always defined, preventing text-only models (like mistral-7b) from crashing when the script checks the Qwen3 RoPE 3D position ID logic.

Tests

1. Ran pytest tests/unit/configs_test.py locally: All 76 tests passed successfully.

python3 -m tests.utils.forward_pass_logit_checker \
    src/maxtext/configs/base.yml \
    base_output_directory=gs://runner-maxtext-logs/2026-07-12 \
    load_parameters_path=gs://runner-maxtext-logs/2026-07-12/7b/scanned_ckpt/0/items \
    run_name=matmul_forward_pass_test \
    per_device_batch_size=1 \
    model_name=mistral-7b \
    tokenizer_path=/deps/src/maxtext/assets/tokenizer.mistral-v1 \
    max_prefill_predict_length=11 \
    max_target_length=11 \
    dataset_type=synthetic \
    dtype=float32 \
    megablox=False \
    sparse_matmul=False \
    --atol=3 \
    --rtol=1 \
    --token_size=4

Log: https://paste.googleplex.com/4641417803726848

Checklist

Before submitting this PR, please make sure (put X in square brackets):

• I have performed a self-review of my code. For an optional AI review, add the gemini-review label.
• I have necessary comments in my code, particularly in hard-to-understand areas.
• I have run end-to-end tests tests and provided workload links above if applicable.
• I have made or will make corresponding changes to the doc if needed, including adding new documentation pages to the relevant Table of Contents (toctree directive) as explained in our documentation.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!