Sync Ripple generation, training, and cache fixes

#1

Update Ripple generation, training, and cache handling in this patch repository.

Changes:

  • Fix cached token slicing, padding-aware positions, and static/sliding cache alignment.
  • Preserve requested outputs and caller attention masks; bound inference logits and MLP activation memory.
  • Fix gradient checkpointing, accumulation-aware loss, tensor-parallel head shapes, and tied embeddings.
  • Keep the default cache crop-compatible on Transformers 4 for assisted decoding.
  • Refresh the optional patch for older checkpoints and retain the README's file-copy instructions. No weights or configuration changes.

Validation: 242 CPU regression tests passed on Python 3.11.14, PyTorch 2.15.0, and Transformers 5.22.2. Patch application reproduces the shipped file exactly, and reverse application restores the baseline. No full 65B-weight evaluation was run.

zx-modelcloud changed pull request title from Sync GPT-QModel Ripple generation, training, and cache fixes to Sync Ripple generation, training, and cache fixes
Qubitium changed pull request status to merged

Sign up or log in to comment