cua-lite/Fara-7B
A config-only fix of microsoft/Fara-7B.
The model weights are identical to the upstream release — only config.json
is changed to make the model load under transformers v5 + SGLang.
What changed
microsoft/Fara-7B ships a transformers-4.x config.json where rope_theta
and rope_scaling are separate top-level fields. Under transformers v5 (the
"untangle config" refactor), SGLang reads RoPE parameters via the computed
config.rope_parameters property and passes the top-level Qwen2_5_VLConfig
(not text_config) into Qwen2Model. For that top-level config,
rope_parameters returns rope_scaling without folding in rope_theta, so
SGLang's get_rope_config hits rope_params["rope_theta"] and dies at load:
KeyError: 'rope_theta'
This fork folds rope_theta into rope_scaling (at both the top level and
in text_config) so rope_parameters exposes it:
"rope_scaling": {
"mrope_section": [16, 24, 24],
"rope_type": "default",
"type": "default",
"rope_theta": 1000000.0
}
rope_theta is also kept as a top-level key for transformers-4.x backward
compatibility. Nothing else is modified — tokenizer, preprocessor, chat
template, and all safetensors shards are byte-identical to upstream.
Usage (SGLang)
python -m sglang.launch_server \
--model-path cua-lite/Fara-7B \
--host 0.0.0.0 --port 30088
Or, keeping microsoft/Fara-7B as the logical model id and only overriding the
weights/config path:
python scripts/serve_sglang.py \
--model-id microsoft/Fara-7B \
--model-path cua-lite/Fara-7B \
--port 30088
License & attribution
All rights, license, and model behavior follow the upstream model. See
microsoft/Fara-7B for the model
card, intended use, and limitations.