fix tensor collect

[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 updates the tinker dependency to version 0.16.1 and refactors the model output processing logic in common.py to better handle pipeline parallel stages. Specifically, _tinker_build_output now correctly identifies and skips stages that produce no outputs, and _normalize_tensor_output has been updated to handle empty lists and tensor concatenation. Feedback was provided to improve memory efficiency by moving individual tensors to the CPU before concatenation to prevent potential GPU OOM issues.

[Copilot]

Pull request overview

Fixes tensor collection/normalization for Tinker-compatible server backends (especially when some pipeline stages produce no logits/logps), and updates the Docker image to use a newer tinker version.

Changes:

• Normalize logits/logps outputs in a single path and treat empty lists as “missing output” to avoid downstream collection issues.
• Return an empty per-datum output list when neither logits nor logps are produced (e.g., non-last PP stage).
• Bump tinker in the Dockerfile from 0.14.0 to 0.16.1.

Reviewed changes

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

File	Description
src/twinkle/server/model/backends/common.py	Adjusts output extraction/normalization to handle None and empty-list outputs cleanly for distributed/pipeline scenarios.
Dockerfile	Updates the container dependency pin for tinker to a newer version.