OpenNMT/CTranslate2

falcon3 support please?

Opened this issue · 1 comments

BBC-Esq commented

Falcon3 was just released today:

https://huggingface.co/collections/tiiuae/falcon3-67605ae03578be86e4e87026

More details:

https://huggingface.co/blog/falcon3

More stuff:

details and stuff 
model |   |  
-- | -- | --
model.embed_tokens.weight | [131 072, 3 072] | BF16
model.layers.0 |   |  
model.layers.0.input_layernorm.weight | [3 072] | BF16
model.layers.0.mlp.down_proj.weight | [3 072, 23 040] | BF16
model.layers.0.mlp.gate_proj.weight | [23 040, 3 072] | BF16
model.layers.0.mlp.up_proj.weight | [23 040, 3 072] | BF16
model.layers.0.post_attention_layernorm.weight | [3 072] | BF16
model.layers.0.self_attn.k_proj.weight | [1 024, 3 072] | BF16
model.layers.0.self_attn.o_proj.weight | [3 072, 3 072] | BF16
model.layers.0.self_attn.q_proj.weight | [3 072, 3 072] | BF16
model.layers.0.self_attn.v_proj.weight | [1 024, 3 072] | BF16
model.layers.1 |   |  
model.layers.1.input_layernorm.weight | [3 072] | BF16
model.layers.1.mlp.down_proj.weight | [3 072, 23 040] | BF16
model.layers.1.mlp.gate_proj.weight | [23 040, 3 072] | BF16
model.layers.1.mlp.up_proj.weight | [23 040, 3 072] | BF16
model.layers.1.post_attention_layernorm.weight | [3 072] | BF16
model.layers.1.self_attn.k_proj.weight | [1 024, 3 072] | BF16
model.layers.1.self_attn.o_proj.weight | [3 072, 3 072] | BF16
model.layers.1.self_attn.q_proj.weight | [3 072, 3 072] | BF16
model.layers.1.self_attn.v_proj.weight | [1 024, 3 072] | BF16
model.layers.2 |   |  
model.layers.2.input_layernorm.weight | [3 072] | BF16
model.layers.2.mlp.down_proj.weight | [3 072, 23 040] | BF16
model.layers.2.mlp.gate_proj.weight | [23 040, 3 072] | BF16
model.layers.2.mlp.up_proj.weight | [23 040, 3 072] | BF16
model.layers.2.post_attention_layernorm.weight | [3 072] | BF16
model.layers.2.self_attn.k_proj.weight | [1 024, 3 072] | BF16
model.layers.2.self_attn.o_proj.weight | [3 072, 3 072] | BF16
model.layers.2.self_attn.q_proj.weight | [3 072, 3 072] | BF16
model.layers.2.self_attn.v_proj.weight | [1 024, 3 072] | BF16
model.layers.3 |   |  
model.layers.3.input_layernorm.weight | [3 072] | BF16
model.layers.3.mlp.down_proj.weight | [3 072, 23 040] | BF16
model.layers.3.mlp.gate_proj.weight | [23 040, 3 072] | BF16
model.layers.3.mlp.up_proj.weight | [23 040, 3 072] | BF16
model.layers.3.post_attention_layernorm.weight | [3 072] | BF16
model.layers.3.self_attn.k_proj.weight | [1 024, 3 072] | BF16
model.layers.3.self_attn.o_proj.weight | [3 072, 3 072] | BF16
model.layers.3.self_attn.q_proj.weight | [3 072, 3 072] | BF16
model.layers.3.self_attn.v_proj.weight | [1 024, 3 072] | BF16
model.layers.4 |   |  
model.layers.4.input_layernorm.weight | [3 072] | BF16
model.layers.4.mlp.down_proj.weight | [3 072, 23 040] | BF16
model.layers.4.mlp.gate_proj.weight | [23 040, 3 072] | BF16
model.layers.4.mlp.up_proj.weight | [23 040, 3 072] | BF16
model.layers.4.post_attention_layernorm.weight | [3 072] | BF16
model.layers.4.self_attn.k_proj.weight | [1 024, 3 072] | BF16
model.layers.4.self_attn.o_proj.weight | [3 072, 3 072] | BF16
model.layers.4.self_attn.q_proj.weight | [3 072, 3 072] | BF16
model.layers.4.self_attn.v_proj.weight | [1 024, 3 072] | BF16
model.layers.5 |   |  
model.layers.5.input_layernorm.weight | [3 072] | BF16
model.layers.5.mlp.down_proj.weight | [3 072, 23 040] | BF16
model.layers.5.mlp.gate_proj.weight | [23 040, 3 072] | BF16
model.layers.5.mlp.up_proj.weight | [23 040, 3 072] | BF16
model.layers.5.post_attention_layernorm.weight | [3 072] | BF16
model.layers.5.self_attn.k_proj.weight | [1 024, 3 072] | BF16
model.layers.5.self_attn.o_proj.weight | [3 072, 3 072] | BF16
model.layers.5.self_attn.q_proj.weight | [3 072, 3 072] | BF16
model.layers.5.self_attn.v_proj.weight | [1 024, 3 072] | BF16
model.layers.6 |   |  
model.layers.6.input_layernorm.weight | [3 072] | BF16
model.layers.6.mlp.down_proj.weight | [3 072, 23 040] | BF16
model.layers.6.mlp.gate_proj.weight | [23 040, 3 072] | BF16
model.layers.6.mlp.up_proj.weight | [23 040, 3 072] | BF16
model.layers.6.post_attention_layernorm.weight | [3 072] | BF16
model.layers.6.self_attn.k_proj.weight | [1 024, 3 072] | BF16
model.layers.6.self_attn.o_proj.weight | [3 072, 3 072] | BF16
model.layers.6.self_attn.q_proj.weight | [3 072, 3 072] | BF16
model.layers.6.self_attn.v_proj.weight | [1 024, 3 072] | BF16
model.layers.7 |   |  
model.layers.7.input_layernorm.weight | [3 072] | BF16
model.layers.7.mlp.down_proj.weight | [3 072, 23 040] | BF16
model.layers.7.mlp.gate_proj.weight | [23 040, 3 072] | BF16
model.layers.7.mlp.up_proj.weight | [23 040, 3 072] | BF16
model.layers.7.post_attention_layernorm.weight | [3 072] | BF16
model.layers.7.self_attn.k_proj.weight | [1 024, 3 072] | BF16
model.layers.7.self_attn.o_proj.weight | [3 072, 3 072] | BF16
model.layers.7.self_attn.q_proj.weight | [3 072, 3 072] | BF16
model.layers.7.self_attn.v_proj.weight | [1 024, 3 072] | BF16
model.layers.8 |   |  
model.layers.8.input_layernorm.weight | [3 072] | BF16
model.layers.8.mlp.down_proj.weight | [3 072, 23 040] | BF16
model.layers.8.mlp.gate_proj.weight | [23 040, 3 072] | BF16
model.layers.8.mlp.up_proj.weight | [23 040, 3 072] | BF16
model.layers.8.post_attention_layernorm.weight | [3 072] | BF16
model.layers.8.self_attn.k_proj.weight | [1 024, 3 072] | BF16
model.layers.8.self_attn.o_proj.weight | [3 072, 3 072] | BF16
model.layers.8.self_attn.q_proj.weight | [3 072, 3 072] | BF16
model.layers.8.self_attn.v_proj.weight | [1 024, 3 072] | BF16
model.layers.9 |   |  
model.layers.9.input_layernorm.weight | [3 072] | BF16
model.layers.9.mlp.down_proj.weight | [3 072, 23 040] | BF16
model.layers.9.mlp.gate_proj.weight | [23 040, 3 072] | BF16
model.layers.9.mlp.up_proj.weight | [23 040, 3 072] | BF16
model.layers.9.post_attention_layernorm.weight | [3 072] | BF16
model.layers.9.self_attn.k_proj.weight | [1 024, 3 072] | BF16
model.layers.9.self_attn.o_proj.weight | [3 072, 3 072] | BF16
model.layers.9.self_attn.q_proj.weight | [3 072, 3 072] | BF16
model.layers.9.self_attn.v_proj.weight | [1 024, 3 072] | BF16
model.layers.10 |   |  
model.layers.10.input_layernorm.weight | [3 072] | BF16
model.layers.10.mlp.down_proj.weight | [3 072, 23 040] | BF16
model.layers.10.mlp.gate_proj.weight | [23 040, 3 072] | BF16
model.layers.10.mlp.up_proj.weight | [23 040, 3 072] | BF16
model.layers.10.post_attention_layernorm.weight | [3 072] | BF16
model.layers.10.self_attn.k_proj.weight | [1 024, 3 072] | BF16
model.layers.10.self_attn.o_proj.weight | [3 072, 3 072] | BF16
model.layers.10.self_attn.q_proj.weight | [3 072, 3 072] | BF16
model.layers.10.self_attn.v_proj.weight | [1 024, 3 072] | BF16
model.layers.11 |   |  
model.layers.11.input_layernorm.weight | [3 072] | BF16
model.layers.11.mlp.down_proj.weight | [3 072, 23 040] | BF16
model.layers.11.mlp.gate_proj.weight | [23 040, 3 072] | BF16
model.layers.11.mlp.up_proj.weight | [23 040, 3 072] | BF16
model.layers.11.post_attention_layernorm.weight | [3 072] | BF16
model.layers.11.self_attn.k_proj.weight | [1 024, 3 072] | BF16
model.layers.11.self_attn.o_proj.weight | [3 072, 3 072] | BF16
model.layers.11.self_attn.q_proj.weight | [3 072, 3 072] | BF16
model.layers.11.self_attn.v_proj.weight | [1 024, 3 072] | BF16
model.layers.12 |   |  
model.layers.12.input_layernorm.weight | [3 072] | BF16
model.layers.12.mlp.down_proj.weight | [3 072, 23 040] | BF16
model.layers.12.mlp.gate_proj.weight | [23 040, 3 072] | BF16
model.layers.12.mlp.up_proj.weight | [23 040, 3 072] | BF16
model.layers.12.post_attention_layernorm.weight | [3 072] | BF16
model.layers.12.self_attn.k_proj.weight | [1 024, 3 072] | BF16
model.layers.12.self_attn.o_proj.weight | [3 072, 3 072] | BF16
model.layers.12.self_attn.q_proj.weight | [3 072, 3 072] | BF16
model.layers.12.self_attn.v_proj.weight | [1 024, 3 072] | BF16
model.layers.13 |   |  
model.layers.13.input_layernorm.weight | [3 072] | BF16
model.layers.13.mlp.down_proj.weight | [3 072, 23 040] | BF16
model.layers.13.mlp.gate_proj.weight | [23 040, 3 072] | BF16
model.layers.13.mlp.up_proj.weight | [23 040, 3 072] | BF16
model.layers.13.post_attention_layernorm.weight | [3 072] | BF16
model.layers.13.self_attn.k_proj.weight | [1 024, 3 072] | BF16
model.layers.13.self_attn.o_proj.weight | [3 072, 3 072] | BF16
model.layers.13.self_attn.q_proj.weight | [3 072, 3 072] | BF16
model.layers.13.self_attn.v_proj.weight | [1 024, 3 072] | BF16
model.layers.14 |   |  
model.layers.14.input_layernorm.weight | [3 072] | BF16
model.layers.14.mlp.down_proj.weight | [3 072, 23 040] | BF16
model.layers.14.mlp.gate_proj.weight | [23 040, 3 072] | BF16
model.layers.14.mlp.up_proj.weight | [23 040, 3 072] | BF16
model.layers.14.post_attention_layernorm.weight | [3 072] | BF16
model.layers.14.self_attn.k_proj.weight | [1 024, 3 072] | BF16
model.layers.14.self_attn.o_proj.weight | [3 072, 3 072] | BF16
model.layers.14.self_attn.q_proj.weight | [3 072, 3 072] | BF16
model.layers.14.self_attn.v_proj.weight | [1 024, 3 072] | BF16
model.layers.15 |   |  
model.layers.15.input_layernorm.weight | [3 072] | BF16
model.layers.15.mlp.down_proj.weight | [3 072, 23 040] | BF16
model.layers.15.mlp.gate_proj.weight | [23 040, 3 072] | BF16
model.layers.15.mlp.up_proj.weight | [23 040, 3 072] | BF16
model.layers.15.post_attention_layernorm.weight | [3 072] | BF16
model.layers.15.self_attn.k_proj.weight | [1 024, 3 072] | BF16
model.layers.15.self_attn.o_proj.weight | [3 072, 3 072] | BF16
model.layers.15.self_attn.q_proj.weight | [3 072, 3 072] | BF16
model.layers.15.self_attn.v_proj.weight | [1 024, 3 072] | BF16
model.layers.16 |   |  
model.layers.16.input_layernorm.weight | [3 072] | BF16
model.layers.16.mlp.down_proj.weight | [3 072, 23 040] | BF16
model.layers.16.mlp.gate_proj.weight | [23 040, 3 072] | BF16
model.layers.16.mlp.up_proj.weight | [23 040, 3 072] | BF16
model.layers.16.post_attention_layernorm.weight | [3 072] | BF16
model.layers.16.self_attn.k_proj.weight | [1 024, 3 072] | BF16
model.layers.16.self_attn.o_proj.weight | [3 072, 3 072] | BF16
model.layers.16.self_attn.q_proj.weight | [3 072, 3 072] | BF16
model.layers.16.self_attn.v_proj.weight | [1 024, 3 072] | BF16
model.layers.17 |   |  
model.layers.17.input_layernorm.weight | [3 072] | BF16
model.layers.17.mlp.down_proj.weight | [3 072, 23 040] | BF16
model.layers.17.mlp.gate_proj.weight | [23 040, 3 072] | BF16
model.layers.17.mlp.up_proj.weight | [23 040, 3 072] | BF16
model.layers.17.post_attention_layernorm.weight | [3 072] | BF16
model.layers.17.self_attn.k_proj.weight | [1 024, 3 072] | BF16
model.layers.17.self_attn.o_proj.weight | [3 072, 3 072] | BF16
model.layers.17.self_attn.q_proj.weight | [3 072, 3 072] | BF16
model.layers.17.self_attn.v_proj.weight | [1 024, 3 072] | BF16
model.layers.18 |   |  
model.layers.18.input_layernorm.weight | [3 072] | BF16
model.layers.18.mlp.down_proj.weight | [3 072, 23 040] | BF16
model.layers.18.mlp.gate_proj.weight | [23 040, 3 072] | BF16
model.layers.18.mlp.up_proj.weight | [23 040, 3 072] | BF16
model.layers.18.post_attention_layernorm.weight | [3 072] | BF16
model.layers.18.self_attn.k_proj.weight | [1 024, 3 072] | BF16
model.layers.18.self_attn.o_proj.weight | [3 072, 3 072] | BF16
model.layers.18.self_attn.q_proj.weight | [3 072, 3 072] | BF16
model.layers.18.self_attn.v_proj.weight | [1 024, 3 072] | BF16
model.layers.19 |   |  
model.layers.19.input_layernorm.weight | [3 072] | BF16
model.layers.19.mlp.down_proj.weight | [3 072, 23 040] | BF16
model.layers.19.mlp.gate_proj.weight | [23 040, 3 072] | BF16
model.layers.19.mlp.up_proj.weight | [23 040, 3 072] | BF16
model.layers.19.post_attention_layernorm.weight | [3 072] | BF16
model.layers.19.self_attn.k_proj.weight | [1 024, 3 072] | BF16
model.layers.19.self_attn.o_proj.weight | [3 072, 3 072] | BF16
model.layers.19.self_attn.q_proj.weight | [3 072, 3 072] | BF16
model.layers.19.self_attn.v_proj.weight | [1 024, 3 072] | BF16
model.layers.20 |   |  
model.layers.20.input_layernorm.weight | [3 072] | BF16
model.layers.20.mlp.down_proj.weight | [3 072, 23 040] | BF16
model.layers.20.mlp.gate_proj.weight | [23 040, 3 072] | BF16
model.layers.20.mlp.up_proj.weight | [23 040, 3 072] | BF16
model.layers.20.post_attention_layernorm.weight | [3 072] | BF16
model.layers.20.self_attn.k_proj.weight | [1 024, 3 072] | BF16
model.layers.20.self_attn.o_proj.weight | [3 072, 3 072] | BF16
model.layers.20.self_attn.q_proj.weight | [3 072, 3 072] | BF16
model.layers.20.self_attn.v_proj.weight | [1 024, 3 072] | BF16
model.layers.21 |   |  
model.layers.21.input_layernorm.weight | [3 072] | BF16
model.layers.21.mlp.down_proj.weight | [3 072, 23 040] | BF16
model.layers.21.mlp.gate_proj.weight | [23 040, 3 072] | BF16
model.layers.21.mlp.up_proj.weight | [23 040, 3 072] | BF16
model.layers.21.post_attention_layernorm.weight | [3 072] | BF16
model.layers.21.self_attn.k_proj.weight | [1 024, 3 072] | BF16
model.layers.21.self_attn.o_proj.weight | [3 072, 3 072] | BF16
model.layers.21.self_attn.q_proj.weight | [3 072, 3 072] | BF16
model.layers.21.self_attn.v_proj.weight | [1 024, 3 072] | BF16
model.layers.22 |   |  
model.layers.22.input_layernorm.weight | [3 072] | BF16
model.layers.22.mlp.down_proj.weight | [3 072, 23 040] | BF16
model.layers.22.mlp.gate_proj.weight | [23 040, 3 072] | BF16
model.layers.22.mlp.up_proj.weight | [23 040, 3 072] | BF16
model.layers.22.post_attention_layernorm.weight | [3 072] | BF16
model.layers.22.self_attn.k_proj.weight | [1 024, 3 072] | BF16
model.layers.22.self_attn.o_proj.weight | [3 072, 3 072] | BF16
model.layers.22.self_attn.q_proj.weight | [3 072, 3 072] | BF16
model.layers.22.self_attn.v_proj.weight | [1 024, 3 072] | BF16
model.layers.23 |   |  
model.layers.23.input_layernorm.weight | [3 072] | BF16
model.layers.23.mlp.down_proj.weight | [3 072, 23 040] | BF16
model.layers.23.mlp.gate_proj.weight | [23 040, 3 072] | BF16
model.layers.23.mlp.up_proj.weight | [23 040, 3 072] | BF16
model.layers.23.post_attention_layernorm.weight | [3 072] | BF16
model.layers.23.self_attn.k_proj.weight | [1 024, 3 072] | BF16
model.layers.23.self_attn.o_proj.weight | [3 072, 3 072] | BF16
model.layers.23.self_attn.q_proj.weight | [3 072, 3 072] | BF16
model.layers.23.self_attn.v_proj.weight | [1 024, 3 072] | BF16
model.layers.24 |   |  
model.layers.24.input_layernorm.weight | [3 072] | BF16
model.layers.24.mlp.down_proj.weight | [3 072, 23 040] | BF16
model.layers.24.mlp.gate_proj.weight | [23 040, 3 072] | BF16
model.layers.24.mlp.up_proj.weight | [23 040, 3 072] | BF16
model.layers.24.post_attention_layernorm.weight | [3 072] | BF16
model.layers.24.self_attn.k_proj.weight | [1 024, 3 072] | BF16
model.layers.24.self_attn.o_proj.weight | [3 072, 3 072] | BF16
model.layers.24.self_attn.q_proj.weight | [3 072, 3 072] | BF16
model.layers.24.self_attn.v_proj.weight | [1 024, 3 072] | BF16
model.layers.25 |   |  
model.layers.25.input_layernorm.weight | [3 072] | BF16
model.layers.25.mlp.down_proj.weight | [3 072, 23 040] | BF16
model.layers.25.mlp.gate_proj.weight | [23 040, 3 072] | BF16
model.layers.25.mlp.up_proj.weight | [23 040, 3 072] | BF16
model.layers.25.post_attention_layernorm.weight | [3 072] | BF16
model.layers.25.self_attn.k_proj.weight | [1 024, 3 072] | BF16
model.layers.25.self_attn.o_proj.weight | [3 072, 3 072] | BF16
model.layers.25.self_attn.q_proj.weight | [3 072, 3 072] | BF16
model.layers.25.self_attn.v_proj.weight | [1 024, 3 072] | BF16
model.layers.26 |   |  
model.layers.26.input_layernorm.weight | [3 072] | BF16
model.layers.26.mlp.down_proj.weight | [3 072, 23 040] | BF16
model.layers.26.mlp.gate_proj.weight | [23 040, 3 072] | BF16
model.layers.26.mlp.up_proj.weight | [23 040, 3 072] | BF16
model.layers.26.post_attention_layernorm.weight | [3 072] | BF16
model.layers.26.self_attn.k_proj.weight | [1 024, 3 072] | BF16
model.layers.26.self_attn.o_proj.weight | [3 072, 3 072] | BF16
model.layers.26.self_attn.q_proj.weight | [3 072, 3 072] | BF16
model.layers.26.self_attn.v_proj.weight | [1 024, 3 072] | BF16
model.layers.27 |   |  
model.layers.27.input_layernorm.weight | [3 072] | BF16
model.layers.27.mlp.down_proj.weight | [3 072, 23 040] | BF16
model.layers.27.mlp.gate_proj.weight | [23 040, 3 072] | BF16
model.layers.27.mlp.up_proj.weight | [23 040, 3 072] | BF16
model.layers.27.post_attention_layernorm.weight | [3 072] | BF16
model.layers.27.self_attn.k_proj.weight | [1 024, 3 072] | BF16
model.layers.27.self_attn.o_proj.weight | [3 072, 3 072] | BF16
model.layers.27.self_attn.q_proj.weight | [3 072, 3 072] | BF16
model.layers.27.self_attn.v_proj.weight | [1 024, 3 072] | BF16
model.norm.weight | [3 072] | BF16
lm_head.weight | [131 072, 3 072] | BF16
jncraton commented

The transformer-based Falcon 3 models use the Llama architecture, so they are already supported. I played around with the 1B, 3B, and 7B models yesterday.

I did run into an issue because the Faclon 3 models did not specify a bos token. Once I set this manually, the model appeared to work correctly.