From f9deb87d20ceebcf6203ae859a20b88522d789b5 Mon Sep 17 00:00:00 2001 From: Cebtenzzre Date: Sat, 30 Sep 2023 18:03:23 -0400 Subject: [PATCH] convert scripts: add feed-forward length for better compatiblilty This GGUF key is used by all llama.cpp models with upstream support. --- gpt4all-backend/scripts/convert_gptj_to_gguf.py | 1 + gpt4all-backend/scripts/convert_mpt_hf_to_gguf.py | 1 + gpt4all-backend/scripts/convert_replit_hf_to_gguf.py | 1 + 3 files changed, 3 insertions(+) diff --git a/gpt4all-backend/scripts/convert_gptj_to_gguf.py b/gpt4all-backend/scripts/convert_gptj_to_gguf.py index 1377c9a7..9b8f3a62 100755 --- a/gpt4all-backend/scripts/convert_gptj_to_gguf.py +++ b/gpt4all-backend/scripts/convert_gptj_to_gguf.py @@ -70,6 +70,7 @@ gguf_writer.add_name("GPT-J") gguf_writer.add_context_length(config.n_positions) gguf_writer.add_embedding_length(config.n_embd) gguf_writer.add_block_count(block_count) +gguf_writer.add_feed_forward_length(4 * config.n_embd) gguf_writer.add_head_count(config.n_head) gguf_writer.add_rope_dimension_count(config.rotary_dim) gguf_writer.add_layer_norm_eps(config.layer_norm_epsilon) diff --git a/gpt4all-backend/scripts/convert_mpt_hf_to_gguf.py b/gpt4all-backend/scripts/convert_mpt_hf_to_gguf.py index d859348c..9db33c61 100755 --- a/gpt4all-backend/scripts/convert_mpt_hf_to_gguf.py +++ b/gpt4all-backend/scripts/convert_mpt_hf_to_gguf.py @@ -65,6 +65,7 @@ gguf_writer.add_name("MPT") gguf_writer.add_context_length(config.max_seq_len) gguf_writer.add_embedding_length(config.d_model) gguf_writer.add_block_count(block_count) +gguf_writer.add_feed_forward_length(4 * config.d_model) gguf_writer.add_head_count(config.n_heads) gguf_writer.add_max_alibi_bias(config.attn_config.alibi_bias_max) gguf_writer.add_layer_norm_eps(config.layer_norm_epsilon) diff --git a/gpt4all-backend/scripts/convert_replit_hf_to_gguf.py b/gpt4all-backend/scripts/convert_replit_hf_to_gguf.py index 38b504ad..125df6dd 100755 --- a/gpt4all-backend/scripts/convert_replit_hf_to_gguf.py +++ b/gpt4all-backend/scripts/convert_replit_hf_to_gguf.py @@ -50,6 +50,7 @@ gguf_writer.add_name("Replit") gguf_writer.add_context_length(config.max_seq_len) gguf_writer.add_embedding_length(config.d_model) gguf_writer.add_block_count(block_count) +gguf_writer.add_feed_forward_length(4 * config.d_model) gguf_writer.add_head_count(config.n_heads) gguf_writer.add_max_alibi_bias(config.attn_config.alibi_bias_max) gguf_writer.add_layer_norm_eps(config.layer_norm_epsilon)