Merged full weights: base Qwen3.5-9B @ c202236 + LoRA adapter @ b585e34 (offline merge, scaling 2.0, 128 modules)

by bingran-you - opened 25 days ago

base: refs/heads/main

←

from: refs/pr/5

Discussion Files changed

+496572

-0

This PR is in draft mode

bingran-you

BenchFlow org 25 days ago

Runtime LoRA on the qwen3_5 hybrid arch is silently a no-op in vLLM (and crashes SGLang), so serving must use merged weights. Merge: W += (alpha/r) * B@A, alpha=64 r=32; tokenizer/chat_template/generation_config overlaid from the adapter repo. Produced on env0-h100-spot 2026-07-03, max|delta|=1.9e-03. Checkpoint-backup PR per repo convention - do not merge to main (main stays the adapter).

Merged full weights: base Qwen3.5-9B @ c202236 + LoRA adapter @ b585e34 (offline merge, scaling 2.0, 128 modules)7dcd5a3a

Upload images, audio, and videos by dragging in the text input, pasting, or clicking here.

Tap or paste here to upload images

Publish this branch

This branch is in draft mode, publish it to be able to merge.

· Sign up or log in to comment