fix short math grpo cookbook

[Yunnglin]

PR type

• Bug Fix
• New Feature
• Document Updates
• More Models or Datasets Support

PR information

Write the detail information belongs to this PR.

Experiment results

Paste your experiment result here(if needed).

[gemini-code-assist]

Code Review

This pull request transitions several math training examples to use the GSM8K dataset and GRPO, incorporating a new brevity reward and SwanLab logging. Key improvements include fixing a potential crash in the template base class when handling multi-modal token types as numpy arrays and correcting a logic error in how default padding values are applied. Additionally, feedback was provided to ensure consistency in the 'enable_thinking' configuration across different cookbook implementations.

[Copilot]

Pull request overview

This PR updates the “short math GRPO” cookbooks and related serving/config helpers, aligning examples with the newer twinkle.* APIs and GSM8K-focused reward/preprocessing while making small robustness tweaks in the template layer.

Changes:

• Make Template.concat_input_feature() more tolerant of non-tensor mm_token_type_ids, and adjust _apply_chat_template() defaults via processor_kwargs.
• Revise GRPO cookbook examples to use GSM8K processing + a brevity reward, update imports to twinkle modules, and add SwanLab logging in the Twinkle self-host example.
• Update Megatron cookbook launch/config: run server via python -m twinkle.server, add save-dir/server-config flags, and extend sampler engine args.

Reviewed changes

Copilot reviewed 11 out of 11 changed files in this pull request and generated 2 comments.

Show a summary per file

File	Description
src/twinkle/template/base.py	Robustifies mm token-type concatenation and adjusts chat-template padding defaults via processor_kwargs.
src/twinkle/server/model/tinker_handlers.py	Adds a TODO note around LoRA config flexibility.
cookbook/client/twinkle/self_host/short_math_grpo.py	Switches to GSM8K processor + brevity reward; adds SwanLab logging; updates imports to twinkle.*.
cookbook/client/twinkle/self_host/self_congnition.py	Updates imports from twinkle_client.* to twinkle.*.
cookbook/client/twinkle/modelscope/self_congnition.py	Updates imports from twinkle_client.* to twinkle.*.
cookbook/client/tinker/self_host/short_math_grpo.py	Uses Qwen3_5Template explicitly and tweaks training config defaults.
cookbook/client/tinker/modelscope/short_math_grpo.py	Reworks example from “Math” to GSM8K with brevity reward + GSM8K processor and template updates.
cookbook/client/server/megatron/server.py	Removes legacy launcher script (replaced by module CLI usage in run.sh).
cookbook/client/server/megatron/server_config.yaml	Adds enable_tower_connector_lora to sampler engine args.
cookbook/client/server/megatron/server_config_4b.yaml	Adjusts sampler nproc_per_node and adds enable_tower_connector_lora.
cookbook/client/server/megatron/run.sh	Adds save-dir/server-config flags, exports TWINKLE_DEFAULT_SAVE_DIR, switches to python -m twinkle.server, and expands cleanup logic.

Comments suppressed due to low confidence (1)

cookbook/client/twinkle/self_host/short_math_grpo.py:125

• swanlab.login(api_key=os.environ.get('SWANLAB_API_KEY', '')) will attempt to log in with an empty API key when the env var is missing, which is likely to fail with a confusing error and prevents the script from running with USE_SWANLAB=True. Consider either (a) requiring the env var (like other cookbook scripts) and raising a clear error, or (b) skipping SwanLab initialization when the key is absent.