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

Fix editorconfig formatting

Browse source
This commit is contained in:
Phillip Kravtsov 2023-10-03 13:18:23 -07:00
parent 7a279fe5a8
commit c90ed9f16b
3 changed files with 14 additions and 14 deletions
Download patch file Download diff file Expand all files Collapse all files

6
convert-persimmon-to-gguf.py
View file

@ -24,7 +24,7 @@ def _flatten_dict(dct, tensors, prefix=None):
def _get_sentencepiece_tokenizer_info(dir_model: Path): def _get_sentencepiece_tokenizer_info(dir_model: Path):
tokenizer_path = dir_model / 'adept_vocab.model' tokenizer_path = dir_model / 'adept_vocab.model'
print('gguf: getting sentencepiece tokenizer from', tokenizer_path) print('gguf: getting sentencepiece tokenizer from', tokenizer_path)
tokenizer = SentencePieceProcessor(str(tokenizer_path)) tokenizer = SentencePieceProcessor(str(tokenizer_path))
print('gguf: adding tokens') print('gguf: adding tokens')
tokens: list[bytes] = [] tokens: list[bytes] = []
scores: list[float] = [] scores: list[float] = []
@ -70,7 +70,7 @@ def main():
arch = gguf.MODEL_ARCH.PERSIMMON arch = gguf.MODEL_ARCH.PERSIMMON
gguf_writer = gguf.GGUFWriter(args.outfile, gguf.MODEL_ARCH_NAMES[arch]) gguf_writer = gguf.GGUFWriter(args.outfile, gguf.MODEL_ARCH_NAMES[arch])
block_count = hparams.num_layers block_count = hparams.num_layers
head_count = hparams.num_attention_heads head_count = hparams.num_attention_heads
head_count_kv = head_count head_count_kv = head_count
@ -127,4 +127,4 @@ def main():
if __name__ == '__main__': if __name__ == '__main__':
main() main()

2
ggml-metal.metal
View file

@ -1129,7 +1129,7 @@ kernel void kernel_concat(
uint3 tgpig[[threadgroup_position_in_grid]], uint3 tgpig[[threadgroup_position_in_grid]],
uint3 tpitg[[thread_position_in_threadgroup]], uint3 tpitg[[thread_position_in_threadgroup]],
uint3 ntg[[threads_per_threadgroup]]) { uint3 ntg[[threads_per_threadgroup]]) {
const int64_t i03 = tgpig.z; const int64_t i03 = tgpig.z;
const int64_t i02 = tgpig.y; const int64_t i02 = tgpig.y;
const int64_t i01 = tgpig.x; const int64_t i01 = tgpig.x;

20
llama.cpp
View file

@ -2476,7 +2476,7 @@ static void llm_load_tensors(
} }
} break; } break;
case LLM_ARCH_PERSIMMON: case LLM_ARCH_PERSIMMON:
{ {
model.tok_embeddings = ml.create_tensor(ctx, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, GGML_BACKEND_CPU); model.tok_embeddings = ml.create_tensor(ctx, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, GGML_BACKEND_CPU);
model.output_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, GGML_BACKEND_CPU); model.output_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, GGML_BACKEND_CPU);
model.output_norm_b = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT_NORM, "bias"), {n_embd}, GGML_BACKEND_CPU); model.output_norm_b = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT_NORM, "bias"), {n_embd}, GGML_BACKEND_CPU);
@ -4162,8 +4162,8 @@ static struct ggml_cgraph * llm_build_persimmon(
cur = ggml_add(ctx0, cur, model.layers[il].attn_norm_b); cur = ggml_add(ctx0, cur, model.layers[il].attn_norm_b);
offload_func(cur); offload_func(cur);
ggml_format_name(cur, "input_layernorm_%d", il); ggml_format_name(cur, "input_layernorm_%d", il);
} }
// self attention // self attention
{ {
cur = ggml_mul_mat(ctx0, model.layers[il].wqkv, cur); cur = ggml_mul_mat(ctx0, model.layers[il].wqkv, cur);
offload_func_kq(cur); offload_func_kq(cur);
@ -4206,7 +4206,7 @@ static struct ggml_cgraph * llm_build_persimmon(
offload_func_v(tmpk); offload_func_v(tmpk);
tmpk = ggml_add(ctx0, tmpk, model.layers[il].attn_k_norm_b); tmpk = ggml_add(ctx0, tmpk, model.layers[il].attn_k_norm_b);
offload_func_v(tmpk); offload_func_v(tmpk);
// RoPE the first n_rot of q/k, pass the other half, and concat. // RoPE the first n_rot of q/k, pass the other half, and concat.
struct ggml_tensor * qrot = ggml_view_3d( struct ggml_tensor * qrot = ggml_view_3d(
ctx0, tmpq, n_rot, n_head, n_tokens, ctx0, tmpq, n_rot, n_head, n_tokens,
@ -4227,7 +4227,7 @@ static struct ggml_cgraph * llm_build_persimmon(
// get the second half of tmpq, e.g tmpq[n_rot:, :, :] // get the second half of tmpq, e.g tmpq[n_rot:, :, :]
struct ggml_tensor * qpass = ggml_view_3d( struct ggml_tensor * qpass = ggml_view_3d(
ctx0, tmpq, n_rot, n_head, n_tokens, ctx0, tmpq, n_rot, n_head, n_tokens,
ggml_element_size(tmpq) * n_embd_head, ggml_element_size(tmpq) * n_embd_head,
ggml_element_size(tmpq) * n_embd_head * n_head, ggml_element_size(tmpq) * n_embd_head * n_head,
ggml_element_size(tmpq) * n_rot ggml_element_size(tmpq) * n_rot
@ -4328,9 +4328,9 @@ static struct ggml_cgraph * llm_build_persimmon(
offload_func_kq(KQ_soft_max); offload_func_kq(KQ_soft_max);
ggml_set_name(KQ_soft_max, "KQ_soft_max"); ggml_set_name(KQ_soft_max, "KQ_soft_max");
struct ggml_tensor * V = struct ggml_tensor * V =
ggml_view_3d(ctx0, kv_self.v, ggml_view_3d(ctx0, kv_self.v,
n_kv, n_embd_head, n_head_kv, n_kv, n_embd_head, n_head_kv,
ggml_element_size(kv_self.v)*n_ctx, ggml_element_size(kv_self.v)*n_ctx,
ggml_element_size(kv_self.v)*n_ctx*n_embd_head, ggml_element_size(kv_self.v)*n_ctx*n_embd_head,
ggml_element_size(kv_self.v)*n_ctx*n_embd_gqa*il); ggml_element_size(kv_self.v)*n_ctx*n_embd_gqa*il);
@ -4361,11 +4361,11 @@ static struct ggml_cgraph * llm_build_persimmon(
ggml_set_name(inpFF, "inpFF"); ggml_set_name(inpFF, "inpFF");
{ {
// MLP // MLP
{ {
// Norm // Norm
cur = ggml_norm(ctx0, inpFF, norm_eps); cur = ggml_norm(ctx0, inpFF, norm_eps);
offload_func(cur); offload_func(cur);
cur = ggml_add(ctx0, cur = ggml_add(ctx0,
ggml_mul(ctx0, cur, model.layers[il].ffn_norm), ggml_mul(ctx0, cur, model.layers[il].ffn_norm),
model.layers[il].ffn_norm_b model.layers[il].ffn_norm_b
); );
@ -4386,7 +4386,7 @@ static struct ggml_cgraph * llm_build_persimmon(
cur = ggml_mul_mat(ctx0, model.layers[il].w2, cur); cur = ggml_mul_mat(ctx0, model.layers[il].w2, cur);
offload_func(cur); offload_func(cur);
cur = ggml_add(ctx0, cur = ggml_add(ctx0,
cur, cur,
model.layers[il].b2); model.layers[il].b2);
offload_func(cur); offload_func(cur);