Fix sampler and grpo

[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 implements a mechanism to bypass weight caching by using timestamped symlinks for sampler checkpoints and ensures the prefix cache is reset upon loading new weights. It also increases HTTP timeouts and simplifies the sampler initialization to exclusively use vLLMSampler. Feedback was provided regarding the removal of the sampler_type branching logic, which leaves a dead parameter, and suggestions were made to improve symlink robustness by using relative paths and higher-resolution timestamps.

[Copilot]

Pull request overview

Fixes issues around sampler weight saving/loading and GRPO-related sampling behavior by adjusting checkpoint naming, cache invalidation, and request timeouts.

Changes:

• Increase HTTP client request timeouts for Twinkle client requests.
• Change sampler checkpoint saving to always write weights under sampler_weights/latest while returning a per-save timestamp path (via symlink) to avoid path-based caching.
• Reset vLLM prefix cache when new sampler weights/adapters are used, and simplify sampler deployment initialization to always use vLLM.

Reviewed changes

Copilot reviewed 9 out of 9 changed files in this pull request and generated 9 comments.

Show a summary per file

File	Description
src/twinkle_client/http/http_utils.py	Increases default request timeout for GET/POST/DELETE helpers.
src/twinkle/server/utils/checkpoint_base.py	Adjusts sampler checkpoint naming/storage and adds timestamp symlink return path; improves sampler-weight cleanup for symlinks.
src/twinkle/server/sampler/twinkle_handlers.py	Resets sampler prefix cache when an adapter URI is provided for Twinkle sampling requests.
src/twinkle/server/sampler/tinker_handlers.py	Resets sampler prefix cache during Tinker sampling flow.
src/twinkle/server/sampler/app.py	Removes sampler-type branching and always initializes vLLMSampler.
src/twinkle/server/model/twinkle_handlers.py	Ensures sampler weight files are saved under latest/ to match checkpoint manager behavior.
src/twinkle/server/model/tinker_handlers.py	Ensures sampler weight files are saved under latest/ for Tinker sampler checkpoints.
src/twinkle/sampler/vllm_sampler/vllm_sampler.py	Adds logging when attempting to load LoRA from a path during sampling.
src/twinkle/sampler/vllm_sampler/vllm_engine.py	Adds logging when returning a cached LoRA request.