[WANDY666]

Initializing a DeepSeek-V3.2 tokenizer currently imports deepseek3_2.model, which also loads model weights, inference layers, and GPU kernel dependencies. This unnecessarily exposes CPU-only PD masters and other tokenizer-only processes to GPU import failures.

Move the existing DeepSeekV32Tokenizer class unchanged into deepseek3_2/tokenizer.py and update the tokenizer factory. Replace the fake model-module stub in the import-isolation regression with the real wrapper, covering text encoding, thinking, and tool prompt construction while rejecting GPU model implementation imports.

Validation:

• The strengthened regression fails on the original main import path and passes with this change.
• Executed the regression subprocesses directly for both Llama and DeepSeek-V3.2 with CUDA_VISIBLE_DEVICES='' and HF_HUB_OFFLINE=1; pytest is unavailable in the validation environment.
• Confirmed the moved tokenizer class is AST-identical to the original.
• Repository Black and Flake8 hooks passed.

Summary by CodeRabbit

• New Features
  • DeepSeek V3.2 supports chat prompts in regular and thinking modes, including prompts with tools.
  • Prompts can be returned as text or encoded tokens, with generation prompts supported.
  • Chat requests can provide messages using either the conversation or messages format.
• Bug Fixes
  • Tool definitions are formatted and attached to prompts for inclusion in chat requests.

[coderabbitai]

No actionable comments were generated in the recent review.

Recent review info

Run configuration

• Source mode: Auto
• Repository: Unavailable
• Branch: Unavailable
• File: No configuration file from this repository applied
• Commit: Unavailable
• Layers: defaults
• Review profile: CHILL
• Plan: Advanced
• Run ID: c28f01bf-f61a-4da6-9daf-ce734686d841

Commits

Reviewing files that changed from the base of the PR and between 4c1002c and 4c35bbf.

Files selected for processing: 3

• lightllm/models/deepseek3_2/model.py
• lightllm/server/tokenizer.py
• unit_tests/server/test_tokenizer.py

Files with no reviewable changes: 1

• lightllm/models/deepseek3_2/model.py

Included review availability: This review used your included allowance. Your plan provides up to 8 included reviews per hour; 6 remain after this review.

Walkthrough

The DeepSeek V3.2 tokenizer wrapper moves from model.py to tokenizer.py. The server import and tokenizer tests are updated to use and verify the new module location.

Changes: DeepSeek V3.2 tokenizer

Layer / File(s)	Summary
Move the tokenizer wrapper lightllm/models/deepseek3_2/model.py, lightllm/models/deepseek3_2/tokenizer.py	The wrapper and its chat-template handling move from the model module to the tokenizer module.
Wire and verify tokenizer selection lightllm/server/tokenizer.py, unit_tests/server/test_tokenizer.py	The server imports the wrapper from the tokenizer module. Tests check regular, thinking, and tool template output, and verify that the model module is not loaded.

Priority: Normal

Estimated code review effort: 2 (Simple) | ~10 minutes

Change: Bug fix

Possibly related PRs

• ModelTC/LightLLM#1593: Adds a tokenizer-only path and import-isolation tests for DeepSeek-V4; this PR applies similar import-isolation testing to DeepSeek-V3.2.

Suggested reviewers: sufubao

Merge Risk: Minimal · up to 4c35b

No concrete merge-blocking risk was identified in the tokenizer relocation.

Pre-merge checks: 5

Check name	Status	Explanation
Description Check		Check skipped - CodeRabbit’s high-level summary is enabled.
Title check		The title clearly describes the main change: preventing GPU model imports when initializing the DeepSeek-V3.2 tokenizer.
Docstring Coverage		Docstring coverage is inconclusive: files from earlier reviews still need to be rechecked within the check's analysis limits. Not blocking.
Linked Issues check		Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check		Check skipped because no linked issues were found for this pull request.

Finishing Touches

📝 Generate docstrings

• Commit to this branch
• Create a new PR

🧪 Generate unit tests (beta)

• Commit to this branch
• Create a new PR

• Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}