2.75 GB

Ctrl+K

1 contributor

History: 12 commits

qmichalwarda

Correct: benchmarks guided checkpoint choice and some training data shares benchmark formats; release is SG50

cb077b4 verified about 2 hours ago

q8
Image input: vision tower + merger in MLX (vision.safetensors, bf16, also in q8/), qev_mlx.py images=, card 1 day ago
.gitattributes

1.62 kB
qev-0.8b-MLX: initial release 2 days ago
README.md

40.8 kB
Correct: benchmarks guided checkpoint choice and some training data shares benchmark formats; release is SG50 about 2 hours ago
SHA256SUMS

1.71 kB
Card: rename the Images section to 'Image input' (the Hub drops id=images, so #images links did not resolve) 1 day ago
chat_template.jinja

7.76 kB
qev-0.8b-MLX: initial release 2 days ago
config.json

2.42 kB
qev-0.8b-MLX: initial release 2 days ago
model.safetensors

1.5 GB
xet

qev-0.8b-MLX: initial release 2 days ago
model.safetensors.index.json

28 kB
qev-0.8b-MLX: initial release 2 days ago
preprocessor_config.json

390 Bytes
Image input: vision tower + merger in MLX (vision.safetensors, bf16, also in q8/), qev_mlx.py images=, card 1 day ago
qev_config.json

5.15 kB
qev-0.8b-MLX: initial release 2 days ago
qev_mlx.py

32.7 kB
Images on the fast path: compiled + batched image requests in decide_batch, vision tower batched by grid size, per_question / cache_mb image prefixes (cache keyed by image content), image-embedding LRU; card speed (Linux) and numbers 1 day ago
qev_mlx_vision.py

13.7 kB
Images on the fast path: compiled + batched image requests in decide_batch, vision tower batched by grid size, per_question / cache_mb image prefixes (cache keyed by image content), image-embedding LRU; card speed (Linux) and numbers 1 day ago
tokenizer.json

20 MB
xet

qev-0.8b-MLX: initial release 2 days ago
tokenizer_config.json

1.51 kB
qev-0.8b-MLX: initial release 2 days ago
vision.safetensors

201 MB
xet

Image input: vision tower + merger in MLX (vision.safetensors, bf16, also in q8/), qev_mlx.py images=, card 1 day ago