convert : update Falcon script for new HF config

Author: cebtenzzre

convert-falcon-hf-to-gguf.py

@@ -99,7 +99,7 @@ def parse_args() -> argparse.Namespace:
 with open(dir_model / "config.json", "r", encoding="utf-8") as f:
     hparams = json.load(f)
 
-if hparams["architectures"][0] != "RWForCausalLM":
+if hparams["architectures"][0] != "FalconForCausalLM":
     print("Model architecture not supported: " + hparams["architectures"][0])
 
     sys.exit(1)
@@ -112,15 +112,15 @@ def parse_args() -> argparse.Namespace:
 
 print("gguf: get model metadata")
 
-block_count = hparams["n_layer"]
+block_count = hparams["num_hidden_layers"]
 
 gguf_writer.add_name("Falcon")
 gguf_writer.add_context_length(2048) # not in config.json
 gguf_writer.add_tensor_data_layout("jploski") # qkv tensor transform
 gguf_writer.add_embedding_length(hparams["hidden_size"])
 gguf_writer.add_feed_forward_length(4 * hparams["hidden_size"])
 gguf_writer.add_block_count(block_count)
-gguf_writer.add_head_count(hparams["n_head"])
+gguf_writer.add_head_count(hparams["num_attention_heads"])
 if "n_head_kv" in hparams:
     gguf_writer.add_head_count_kv(hparams["n_head_kv"])
 else:
@@ -179,7 +179,7 @@ def parse_args() -> argparse.Namespace:
 tensor_map = gguf.get_tensor_name_map(ARCH,block_count)
 
 # params for qkv transform
-n_head    = hparams["n_head"]
+n_head    = hparams["num_attention_heads"]
 n_head_kv = hparams["n_head_kv"] if "n_head_kv" in hparams else 1
 
 head_dim = hparams["hidden_size"] // n_head