[Feature] Add Qwen3.5 vision model support - #3474

Open

gnguralnick wants to merge 19 commits into

mlc-ai:mainfrom

gnguralnick:qwen35-vision-pr

Commits on Apr 2, 2026

[Feature] Add Qwen3.5 conversation templates
Gabriel Guralnick
committed
implementation of gemma 3 with vision
Gabriel Guralnick
committed
make attention bug workaround more general
Gabriel Guralnick
committed
fix image dtype for webgpu
Gabriel Guralnick
committed
allow quantization of vision encoder
Gabriel Guralnick
committed
fix lint: isort, black, mypy, and pylint issues
Gabriel Guralnick
committed
address code review: use op.clip, op.nn.avg_pool2d, read BOI/EOI from config
Gabriel Guralnick
committed
fix op.clip and op.nn.avg_pool2d: use relax.op with wrap_nested
Gabriel Guralnick
committed
address code review: use config for image_size, refactor normalize, fix type hint
Gabriel Guralnick
committed
fix gemma3v: prompt format, dead code, redundant template, simplify SigLIP
Gabriel Guralnick
committed
fix typing issues
Gabriel Guralnick
committed
remove gemma3v_quantization.py: use centralized factory instead
Gabriel Guralnick
committed
add tests for gemma3v, siglip vision encoder, and attention fallback
Gabriel Guralnick
committed
post-rebase: migrate tir→tirx, remove attention fallback
Gabriel Guralnick
committed
refactor gemma3v_loader to use make_standard_hf_loader
Gabriel Guralnick
committed
fix ImageData embed_size attribute error on TVM FFI objects
Gabriel Guralnick
committed
[Feature] Add Qwen3.5 vision support and conversation templates
Gabriel Guralnick
committed
[Fix] Qwen3.5V: require vision_config and fix import formatting
Gabriel Guralnick
committed
[Test] Add unit tests for Qwen3.5 vision-language model
Gabriel Guralnick
committed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

[Feature] Add Qwen3.5 vision model support - #3474

[Feature] Add Qwen3.5 vision model support#3474
gnguralnick wants to merge 19 commits into
mlc-ai:mainfrom
gnguralnick:qwen35-vision-pr

Commits on Apr 2, 2026

[Feature] Add Qwen3.5 conversation templates

implementation of gemma 3 with vision

make attention bug workaround more general

fix image dtype for webgpu

allow quantization of vision encoder

fix lint: isort, black, mypy, and pylint issues

address code review: use op.clip, op.nn.avg_pool2d, read BOI/EOI from config

fix op.clip and op.nn.avg_pool2d: use relax.op with wrap_nested

address code review: use config for image_size, refactor normalize, fix type hint

fix gemma3v: prompt format, dead code, redundant template, simplify SigLIP

fix typing issues

remove gemma3v_quantization.py: use centralized factory instead

add tests for gemma3v, siglip vision encoder, and attention fallback

post-rebase: migrate tir→tirx, remove attention fallback

refactor gemma3v_loader to use make_standard_hf_loader

fix ImageData embed_size attribute error on TVM FFI objects

[Feature] Add Qwen3.5 vision support and conversation templates

[Fix] Qwen3.5V: require vision_config and fix import formatting

[Test] Add unit tests for Qwen3.5 vision-language model