Support plan for Falcon-E series

[younesbelkada]

Great work on building this library !
I was wondering if there is any plan to support Falcon-E series: https://huggingface.co/tiiuae/Falcon-E-3B-Instruct to this repository? The architecture is similar to the first version of Bitnet models - thanks a lot !

[SujanMishra]

Idea is to . Convert crate have atm idea to convert, any llama based should work . If native trained just like official bitnet will work out of box . Other will be quantized repacked to either a stream able bin for bigger model or a single safetensors as official one does.

[nimishchaudhari]

I tried running it with https://huggingface.co/tiiuae/Falcon3-3B-Instruct-1.58bit but ran into a tensor size mismatch error.

`[PROFILE] Total load_safetensors_mmap: 9.3507625s
[CONVERT] Loaded safetensors in 9.55s
[CONVERT] Packing weights...
19:48:12 [WARN] No lm_head.weight or output.weight found; using tied embeddings
19:48:30 [ERROR] Error during conversion: Invalid tensor shape for self_attn.q_proj.weight: expected [], got []

❌ Error during conversion: Invalid tensor shape for self_attn.q_proj.weight: expected [], got []
error: process didn't exit successfully: target\debug\bitnet-converter.exe --input-dir ../bitnet-b1.58-2B-4T --output-dir models/Converted/bitnet-b1.58-2B-4T (exit code: 1)
`

Even bitnet for me didn't work :/