vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

mpt : remove tabs, trailing whitespace

Browse source
This commit is contained in:
Jan Ploski 2023-10-02 21:55:22 +02:00
parent 90e7d6de28
commit 470801292d
2 changed files with 5 additions and 5 deletions
Download patch file Download diff file Expand all files Collapse all files

2
convert-mpt-hf-to-gguf.py
View file

@ -121,7 +121,7 @@ gguf_writer.add_embedding_length(hparams["d_model"])
gguf_writer.add_block_count(block_count) gguf_writer.add_block_count(block_count)
gguf_writer.add_feed_forward_length(4 * hparams["d_model"]) gguf_writer.add_feed_forward_length(4 * hparams["d_model"])
gguf_writer.add_head_count(hparams["n_heads"]) gguf_writer.add_head_count(hparams["n_heads"])
gguf_writer.add_layer_norm_eps(1e-05) gguf_writer.add_layer_norm_eps(1e-05)
if hparams["attn_config"]["clip_qkv"] is not None: if hparams["attn_config"]["clip_qkv"] is not None:
gguf_writer.add_clamp_kqv(hparams["attn_config"]["clip_qkv"]) gguf_writer.add_clamp_kqv(hparams["attn_config"]["clip_qkv"])
gguf_writer.add_max_alibi_bias(hparams["attn_config"]["alibi_bias_max"]) gguf_writer.add_max_alibi_bias(hparams["attn_config"]["alibi_bias_max"])

8
llama.cpp
View file

@ -4304,11 +4304,11 @@ static struct ggml_cgraph * llm_build_mpt(
offload_func_kq(KQ_scaled); offload_func_kq(KQ_scaled);
ggml_set_name(KQ_scaled, "KQ_scaled"); ggml_set_name(KQ_scaled, "KQ_scaled");
// TODO: replace with ggml_add() // TODO: replace with ggml_add()
struct ggml_tensor * KQ_scaled_alibi = struct ggml_tensor * KQ_scaled_alibi =
ggml_alibi(ctx0, KQ_scaled, std::max(kv_head, n_kv - n_tokens), n_head, max_alibi_bias); ggml_alibi(ctx0, KQ_scaled, std::max(kv_head, n_kv - n_tokens), n_head, max_alibi_bias);
offload_func_kq(KQ_scaled_alibi); offload_func_kq(KQ_scaled_alibi);
ggml_set_name(KQ_scaled_alibi, "KQ_scaled_alibi"); ggml_set_name(KQ_scaled_alibi, "KQ_scaled_alibi");
struct ggml_tensor * KQ_masked = ggml_add(ctx0, KQ_scaled_alibi, KQ_mask); struct ggml_tensor * KQ_masked = ggml_add(ctx0, KQ_scaled_alibi, KQ_mask);
offload_func_kq(KQ_masked); offload_func_kq(KQ_masked);