falcon : write file type

convert-falcon-hf-to-gguf.py

@@ -107,6 +107,7 @@ if "n_head_kv" in hparams:
 else:
     gguf_writer.add_head_count_kv(1)
 gguf_writer.add_layer_norm_eps(hparams["layer_norm_epsilon"])
+gguf_writer.add_file_type(ftype)
 
 # TOKENIZATION
 

scripts/run-all-ppl.sh

@@ -1,7 +1,7 @@
 #!/bin/bash
 
 qnt=(f16 q8_0 q6_k q5_k q5_1 q5_0 q4_k q4_1 q4_0 q3_k q2_k)
-args="--no-mmap -ngl 999 -t 8"
+args="-ngl 999 -t 8"
 
 if [ -z "$1" ]; then
     echo "usage: $0 <model> [qnt] [args]"