convert scripts: load model as late as possible

gpt4all-backend/scripts/convert_bert_hf_to_gguf.py

@@ -7,7 +7,7 @@ from pathlib import Path
 
 import gguf
 import numpy as np
-from transformers import AutoModel, AutoTokenizer
+from transformers import AutoConfig, AutoModel, AutoTokenizer
 
 
 if not 2 <= len(sys.argv) < 4:
@@ -44,17 +44,15 @@ gguf_writer = gguf.GGUFWriter(fname_out, gguf.MODEL_ARCH_NAMES[ARCH])
 
 print("gguf: get model metadata")
 
-model = AutoModel.from_pretrained(dir_model, low_cpu_mem_usage=True)
-hparams = model.config
-print(model)
+config = AutoConfig(dir_model)
 
-block_count = hparams.num_hidden_layers
+block_count = config.num_hidden_layers
 gguf_writer.add_name("BERT")
-gguf_writer.add_context_length(hparams.max_position_embeddings)
-gguf_writer.add_embedding_length(hparams.hidden_size)
-gguf_writer.add_feed_forward_length(hparams.intermediate_size)
+gguf_writer.add_context_length(config.max_position_embeddings)
+gguf_writer.add_embedding_length(config.hidden_size)
+gguf_writer.add_feed_forward_length(config.intermediate_size)
 gguf_writer.add_block_count(block_count)
-gguf_writer.add_head_count(hparams.num_attention_heads)
+gguf_writer.add_head_count(config.num_attention_heads)
 gguf_writer.add_file_type(ftype)
 
 print("gguf: get tokenizer metadata")
@@ -76,7 +74,7 @@ reverse_vocab = {id: encoded_tok for encoded_tok, id in tokenizer.vocab.items()}
 
 # The number of tokens in tokenizer.json can differ from the expected vocab size.
 # This causes downstream issues with mismatched tensor sizes when running the inference
-for i in range(hparams.vocab_size):
+for i in range(config.vocab_size):
     try:
         text = reverse_vocab[i]
     except KeyError:
@@ -94,6 +92,9 @@ special_vocab.add_to_gguf(gguf_writer)
 
 print("gguf: get tensor metadata")
 
+model = AutoModel.from_pretrained(dir_model, config=config, low_cpu_mem_usage=True)
+print(model)
+
 tensor_map = gguf.get_tensor_name_map(ARCH, block_count)
 
 list_vars = model.state_dict()

gpt4all-backend/scripts/convert_mpt_hf_to_gguf.py

@@ -80,12 +80,6 @@ gguf_writer = gguf.GGUFWriter(fname_out, gguf.MODEL_ARCH_NAMES[ARCH])
 print("gguf: get model metadata")
 
 config = AutoConfig.from_pretrained(model_name)
-print("Loading model:", model_name)
-model = AutoModelForCausalLM.from_pretrained(
-    model_name, config=config, torch_dtype=torch.float16 if ftype == 1 else torch.float32, low_cpu_mem_usage=True,
-)
-config = model.config
-print("Model loaded:", model_name)
 
 block_count = config.n_layers
 gguf_writer.add_name("MPT")
@@ -129,6 +123,12 @@ gguf_writer.add_token_types(toktypes)
 
 print("gguf: get tensor metadata")
 
+print("Loading model:", model_name)
+model = AutoModelForCausalLM.from_pretrained(
+    model_name, config=config, torch_dtype=torch.float16 if ftype == 1 else torch.float32, low_cpu_mem_usage=True,
+)
+print("Model loaded:", model_name)
+
 tensor_map = gguf.get_tensor_name_map(ARCH, block_count)
 
 list_vars = model.state_dict()

gpt4all-backend/scripts/convert_replit_hf_to_gguf.py

@@ -8,7 +8,7 @@ from pathlib import Path
 import gguf
 import numpy as np
 from sentencepiece import SentencePieceProcessor
-from transformers import AutoModelForCausalLM, AutoTokenizer
+from transformers import AutoConfig, AutoModelForCausalLM, AutoTokenizer
 
 
 if not 2 <= len(sys.argv) < 4:
@@ -42,9 +42,7 @@ gguf_writer = gguf.GGUFWriter(fname_out, gguf.MODEL_ARCH_NAMES[ARCH])
 
 print("gguf: get model metadata")
 
-model = AutoModelForCausalLM.from_pretrained(dir_model, low_cpu_mem_usage=True)
-config = model.config
-#print(model)
+config = AutoConfig(dir_model)
 
 block_count = config.n_layers
 gguf_writer.add_name("Replit")
@@ -95,6 +93,9 @@ special_vocab.add_to_gguf(gguf_writer)
 
 print("gguf: get tensor metadata")
 
+model = AutoModelForCausalLM.from_pretrained(dir_model, config=config, low_cpu_mem_usage=True)
+#print(model)
+
 tensor_map = gguf.get_tensor_name_map(ARCH, block_count)
 
 list_vars = model.state_dict()