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

mpt : addendum to changeset:84e30e8 - leave parameter clamp_kqv out from metadata rather than use 0.0 to indicate "no clamping" (more compliant with the current GGUF spec?)

Browse source
This commit is contained in:
Jan Ploski 2023-10-01 01:14:07 +02:00
parent 00e8c5c5f6
commit 1be89c4002
2 changed files with 7 additions and 2 deletions
Download patch file Download diff file Expand all files Collapse all files

3
convert-mpt-hf-to-gguf.py
View file

@ -122,7 +122,8 @@ 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)
gguf_writer.add_clamp_kqv(hparams["attn_config"]["clip_qkv"] or 0.0) if hparams["attn_config"]["clip_qkv"] is not None:
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"])
# TOKENIZATION # TOKENIZATION

4
llama.cpp
View file

@ -1926,7 +1926,11 @@ static void llm_load_hparams(
case LLM_ARCH_MPT: case LLM_ARCH_MPT:
{ {
GGUF_GET_KEY(ctx, hparams.f_norm_eps, gguf_get_val_f32, GGUF_TYPE_FLOAT32, true, kv(LLM_KV_ATTENTION_LAYERNORM_EPS)); GGUF_GET_KEY(ctx, hparams.f_norm_eps, gguf_get_val_f32, GGUF_TYPE_FLOAT32, true, kv(LLM_KV_ATTENTION_LAYERNORM_EPS));
if (gguf_find_key(ctx, kv(LLM_KV_ATTENTION_CLAMP_KQV).c_str()) >= 0) {
GGUF_GET_KEY(ctx, hparams.f_clamp_kqv, gguf_get_val_f32, GGUF_TYPE_FLOAT32, true, kv(LLM_KV_ATTENTION_CLAMP_KQV)); GGUF_GET_KEY(ctx, hparams.f_clamp_kqv, gguf_get_val_f32, GGUF_TYPE_FLOAT32, true, kv(LLM_KV_ATTENTION_CLAMP_KQV));
} else {
hparams.f_clamp_kqv = 0.0f;
}
GGUF_GET_KEY(ctx, hparams.f_max_alibi_bias, gguf_get_val_f32, GGUF_TYPE_FLOAT32, true, kv(LLM_KV_ATTENTION_MAX_ALIBI_BIAS)); GGUF_GET_KEY(ctx, hparams.f_max_alibi_bias, gguf_get_val_f32, GGUF_TYPE_FLOAT32, true, kv(LLM_KV_ATTENTION_MAX_ALIBI_BIAS));
switch (hparams.n_layer) { switch (hparams.n_layer) {