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

Consolidate Handling of Phi Models in llama.cpp

Browse source 
- Replaced LLM_ARCH_PHI2 with LLM_ARCH_PHI to unify the handling of different Phi model variants (Phi-1, Phi-1.5, Phi-2).
- Updated architecture names map to reflect the consolidated architecture name from "phi2" to "phi".
- Adjusted the tensor names mapping to use the new architecture name "phi" for consistent tensor loading and processing.
- Modified hyperparameter loading to include a case for 24 layers under LLM_ARCH_PHI, classifying it as MODEL_1B. This change accommodates different layer counts for various Phi model variants.
- Updated tensor loading sections to use the new architecture enum, ensuring proper tensor creation based on the model architecture.
- Renamed build_phi2() to build_phi() in the graph building section, aligning with the new architecture name and ensuring correct computational graph construction for Phi models.
- Adjusted graph construction calls to use the renamed build_phi() function, ensuring seamless integration and functionality for different Phi model variants.

These changes aim to streamline the handling of various Phi models within `llama.cpp`, enhancing the application's capability to work effectively with these models while maintaining code clarity and consistency.
This commit is contained in:
teleprint-me 2023-12-20 15:34:56 -05:00
parent ea6ae8d04c
commit e290792ae4
No known key found for this signature in database
GPG key ID: B0D11345E65C4D48
1 changed files with 12 additions and 11 deletions
Download patch file Download diff file Expand all files Collapse all files

23
llama.cpp
View file

@ -195,7 +195,7 @@ enum llm_arch {
LLM_ARCH_BLOOM,
LLM_ARCH_STABLELM,
LLM_ARCH_QWEN,
LLM_ARCH_PHI2,
LLM_ARCH_PHI,
LLM_ARCH_UNKNOWN,
};
@ -213,7 +213,7 @@ static std::map<llm_arch, std::string> LLM_ARCH_NAMES = {
{ LLM_ARCH_BLOOM, "bloom" },
{ LLM_ARCH_STABLELM, "stablelm" },
{ LLM_ARCH_QWEN, "qwen" },
{ LLM_ARCH_PHI2, "phi2" },
{ LLM_ARCH_PHI, "phi" },
};
enum llm_kv {
@ -553,7 +553,7 @@ static std::map<llm_arch, std::map<llm_tensor, std::string>> LLM_TENSOR_NAMES =
},
},
{
LLM_ARCH_PHI2,
LLM_ARCH_PHI,
{
{ LLM_TENSOR_TOKEN_EMBD, "token_embd" },
{ LLM_TENSOR_OUTPUT_NORM, "output_norm" },
@ -2651,11 +2651,12 @@ static void llm_load_hparams(
default: model.type = e_model::MODEL_UNKNOWN;
}
} break;
case LLM_ARCH_PHI2:
case LLM_ARCH_PHI:
{
ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps);
switch (hparams.n_layer) {
case 24: model.type = e_model::MODEL_1B; break;
case 32: model.type = e_model::MODEL_3B; break;
default: model.type = e_model::MODEL_UNKNOWN;
}
@ -3655,7 +3656,7 @@ static void llm_load_tensors(
}
}
} break;
case LLM_ARCH_PHI2:
case LLM_ARCH_PHI:
{
model.tok_embd = ml.create_tensor(ctx, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, GGML_BACKEND_CPU);
@ -4117,7 +4118,7 @@ static struct ggml_tensor * llm_build_kqv(
struct ggml_tensor * kq = ggml_mul_mat(ctx, k, q);
cb(kq, "kq", il);
if (model.arch == LLM_ARCH_PHI2) {
if (model.arch == LLM_ARCH_PHI) {
// for this arch, we need to perform the KQ multiplication with F32 precision, otherwise we get NaNs
// ref: https://github.com/ggerganov/llama.cpp/pull/4490#issuecomment-1859055847
ggml_mul_mat_set_prec(kq, GGML_PREC_F32);
@ -5523,7 +5524,7 @@ struct llm_build_context {
return gf;
}
struct ggml_cgraph * build_phi2() {
struct ggml_cgraph * build_phi() {
struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
struct ggml_tensor * cur;
@ -5924,8 +5925,8 @@ static struct ggml_cgraph * llama_build_graph(
if (!ggml_allocr_is_measure(lctx.alloc)) {
const int64_t n_embd_head = model.hparams.n_embd_head();
if (model.arch == LLM_ARCH_PHI2) {
// with phi2, we scale the Q to avoid precision issues
if (model.arch == LLM_ARCH_PHI) {
// with phi, we scale the Q to avoid precision issues
// ref: https://github.com/ml-explore/mlx-examples/blob/08e862336ade809bc37d1035f94b359e7d1a5152/phi2/phi2.py#L64-L66
ggml_set_f32(cur, 1.0f);
} else {
@ -6157,9 +6158,9 @@ static struct ggml_cgraph * llama_build_graph(
{
result = llm.build_qwen();
} break;
case LLM_ARCH_PHI2:
case LLM_ARCH_PHI:
{
result = llm.build_phi2();
result = llm.build_phi();
} break;
default:
GGML_ASSERT(false);