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

llama : add support for llama 3.1 rope scaling factors (#8676)

Browse source 
* Add llama 3.1 rope scaling factors to llama conversion and inference

This commit generates the rope factors on conversion and adds them to the resulting model as a tensor. At inference time, these factors are passed to the `ggml_rope_ext` rope oepration, improving results for context windows above 8192

* Update convert_hf_to_gguf.py

Co-authored-by: compilade <git@compilade.net>

* address comments

* address comments

* Update src/llama.cpp

Co-authored-by: compilade <git@compilade.net>

* Update convert_hf_to_gguf.py

Co-authored-by: compilade <git@compilade.net>

---------

Co-authored-by: compilade <git@compilade.net>
This commit is contained in:
Jeffrey Morgan 2024-07-27 05:03:45 -07:00 • committed by GitHub
parent 92090eca21
commit b5e95468b1
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
2 changed files with 40 additions and 2 deletions
Download patch file Download diff file Expand all files Collapse all files

14
src/llama.cpp
View file

@ -2451,6 +2451,7 @@ struct llama_layer {
// long rope factors
struct ggml_tensor * rope_long = nullptr;
struct ggml_tensor * rope_short = nullptr;
struct ggml_tensor * rope_freqs = nullptr;
// bitnet scale
struct ggml_tensor * wq_scale;
@ -6059,6 +6060,8 @@ static bool llm_load_tensors(
layer.ffn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd});
layer.rope_freqs = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ROPE_FREQS, "weight"), {n_embd/n_head/2}, llama_model_loader::TENSOR_NOT_REQUIRED | (i != 0 ? llama_model_loader::TENSOR_DUPLICATED : 0));
if (n_expert == 0) {
layer.ffn_gate = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff});
layer.ffn_down = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd});
@ -8536,6 +8539,10 @@ struct llm_build_context {
// choose long/short freq factors based on the context size
const auto n_ctx_pre_seq = cparams.n_ctx / cparams.n_seq_max;
if (model.layers[il].rope_freqs != nullptr) {
return model.layers[il].rope_freqs;
}
if (n_ctx_pre_seq > hparams.n_ctx_orig_yarn) {
return model.layers[il].rope_long;
}
@ -8730,6 +8737,9 @@ struct llm_build_context {
// self-attention
{
// rope freq factors for llama3; may return nullptr for llama2 and other models
struct ggml_tensor * rope_factors = build_rope_factors(il);
// compute Q and K and RoPE them
struct ggml_tensor * Qcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wq, cur);
cb(Qcur, "Qcur", il);
@ -8753,14 +8763,14 @@ struct llm_build_context {
}
Qcur = ggml_rope_ext(
ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, nullptr,
ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, rope_factors,
n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
ext_factor, attn_factor, beta_fast, beta_slow
);
cb(Qcur, "Qcur", il);
Kcur = ggml_rope_ext(
ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr,
ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, rope_factors,
n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
ext_factor, attn_factor, beta_fast, beta_slow
);