Skip to content

WIP: add Qwen3.5 Megatron reference receipts - #6

Draft
zrr1999 wants to merge 1 commit into
PFCCLab:mainfrom
zrr1999:qwen35-bit-exact-alignment
Draft

WIP: add Qwen3.5 Megatron reference receipts#6
zrr1999 wants to merge 1 commit into
PFCCLab:mainfrom
zrr1999:qwen35-bit-exact-alignment

Conversation

@zrr1999

@zrr1999 zrr1999 commented Aug 5, 2026

Copy link
Copy Markdown

Status

Work-in-progress native ms-swift/Megatron reference and receipt path for Qwen3.5 raw bit-exact alignment.

The Torch reference is deterministic, but cross-framework text loss is still 0/10 exact.

Scope

  • local/eager accuracy-compatible model path
  • compile-disabled determinism
  • native first-step/internal boundary receipts
  • Qwen3.5 argument and callback integration

Validation

  • formal Torch world4 EP4 CUDA/BF16 10-step transaction: same-side raw exact x3
  • local path contains zero TransformerEngine modules

Not ready for merge; cross-framework forward/loss, backward, optimizer, checkpoint, and multimodal gates remain open.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant