Qwen 2.5 VL#2868

Draft

albert-inflection wants to merge 64 commits intometa-pytorch:mainfrom

InflectionAI:albert/qwen2.5-vl

Commits on Jul 3, 2025

qwen 2.5 vl code skeleton

albert-inflection
authored and
Albert Luo
committed
model builder progress

albert-inflection
authored and
Albert Luo
committed
more model building progress

albert-inflection
authored and
Albert Luo
committed
airplane update

albert-inflection
authored and
Albert Luo
committed
WIP transform + rope

Albert Luo
authored and
Albert Luo
committed
image transform progress

Albert Luo
authored and
Albert Luo
committed
image transform progress

albert-inflection
authored and
Albert Luo
committed
Qwen2_5_VLImageTransform complete

lawrence-inflection
authored and
Albert Luo
committed
remove context.md from tracking

lawrence-inflection
authored and
Albert Luo
committed
Qwen2_5_VLTransform implemented

lawrence-inflection
authored and
Albert Luo
committed
module progress

Albert Luo
authored and
Albert Luo
committed
batch size in ViT forward

Albert Luo
authored and
Albert Luo
committed
rehaul modules, start from near HF

Albert Luo
authored and
Albert Luo
committed
Rope + Window attn attempt 1

Albert Luo
authored and
Albert Luo
committed
_positional_embeddings.py implementation

lawrencefeng17
authored and
Albert Luo
committed
progress on _component_builders.py for decoder

lawrencefeng17
authored and
Albert Luo
committed
upstream cleanup

Albert Luo
authored and
Albert Luo
committed
more cleanup

Albert Luo
authored and
Albert Luo
committed
merge temp branch onto albert/qwen2.5-vl

lawrencefeng17
authored and
Albert Luo
committed
refactored Qwen25VLRotaryPositionalEmbeddings; passed test cases

lawrencefeng17
authored and
Albert Luo
committed
refactored Qwen25VLRotaryPositionalEmbeddings; added summary context.md

lawrencefeng17
authored and
Albert Luo
committed
feat: Qwen25VLEarlyFusionModel wrapper class

lawrencefeng17
authored and
Albert Luo
committed
rebase

Albert Luo
authored and
Albert Luo
committed
clean up mlps

Albert Luo
authored and
Albert Luo
committed
clean up encoder builder

Albert Luo
authored and
Albert Luo
committed
fix: removed raise condition; decoder bias fix

lawrencefeng17
authored and
Albert Luo
committed
checkpointing + edits

Albert Luo
authored and
Albert Luo
committed
init

Albert Luo
authored and
Albert Luo
committed
convert weights final

Albert Luo
authored and
Albert Luo
committed
model builder slight fix

Albert Luo
authored and
Albert Luo
committed
fixes: minor changes, early end-to-end testing

lawrencefeng17
authored and
Albert Luo
committed
fix: completely rewrote mrope

lawrencefeng17
authored and
Albert Luo
committed
fix: minor fixes to mrope

lawrencefeng17
authored and
Albert Luo
committed
transform edits

Albert Luo
authored and
Albert Luo
committed
feat: mrope cache implemented for decoder (#2)

authored and
Albert Luo
committed
encoder forward pass edits

Albert Luo
authored and
Albert Luo
committed
bug fixes, training works now

albert-inflection
authored and
Albert Luo
committed
tested and fixed _transform

authored and
Albert Luo
committed
weight saving fix + import

albert-inflection
authored and
Albert Luo
committed
Lawrence/qwen2.5 vl/encoder tests

authored and
Albert Luo
committed
feat: added other qwen variants in model builders

lawrencefeng17
authored and
Albert Luo
committed
custom collation + init edits

albert-inflection
authored and
Albert Luo
committed
fix: removed default args to transform

lawrencefeng17
authored and
Albert Luo
committed
nits

albert-inflection
authored and
Albert Luo
committed
7B config

albert-inflection
authored and
Albert Luo
committed
config nit

albert-inflection
authored and
Albert Luo
committed
added test cases in torchtune style

lawrencefeng17
authored and
Albert Luo
committed
cleanup

albert-inflection
authored and
Albert Luo
committed
rm uv.lock

albert-inflection
authored and
Albert Luo
committed
trainable params

albert-inflection
authored and
Albert Luo
committed
updated model builders

albert-inflection
authored and
Albert Luo
committed
rename rope

albert-inflection
authored and
Albert Luo
committed
cleanup

lawrencefeng17
authored and
Albert Luo
committed
fix

lawrencefeng17
authored and
Albert Luo
committed
cleanup:

lawrencefeng17
authored and
Albert Luo
committed
3B recipe and model builder edit

albert-inflection
authored and
Albert Luo
committed
32B config and modelbuilder changes'

albert-inflection
authored and
Albert Luo
committed
72B config

albert-inflection
authored and
Albert Luo
committed
nit diffs

Albert Luo
authored and
Albert Luo
committed
fix padding token

Albert Luo
authored and
Albert Luo
committed

Commits on Jul 7, 2025

Commits on Jul 8, 2025

fixes to pass linter and all unit tests
lawrencefeng17
committed