Mirrors/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/llama.cpp.git synced 2024-12-26 14:20:31 +01:00
Code Issues Packages Projects Releases Wiki Activity

llama : fix llama3.1 rope_freqs not respecting custom head_dim (#9141)

Browse Source 
* fix: llama3.1 rope_freqs not respecting custom head_dim

* fix: use potential head_dim for Exaone
This commit is contained in:
Carsten Kragelund Jørgensen 2024-08-27 08:53:40 +02:00 committed by GitHub
parent ad76569f8e
commit 75e1dbbaab
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
2 changed files with 5 additions and 4 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
convert_hf_to_gguf.py
View File

@ -1572,7 +1572,7 @@ class LlamaModel(Model):
if rope_scaling := self.find_hparam(["rope_scaling"], optional=True): if rope_scaling := self.find_hparam(["rope_scaling"], optional=True):
if rope_scaling.get("rope_type", '').lower() == "llama3": if rope_scaling.get("rope_type", '').lower() == "llama3":
base = self.hparams.get("rope_theta", 10000.0) base = self.hparams.get("rope_theta", 10000.0)
dim = self.hparams["hidden_size"] // self.hparams["num_attention_heads"] dim = self.hparams.get("head_dim", self.hparams["hidden_size"] // self.hparams["num_attention_heads"])
freqs = 1.0 / (base ** (torch.arange(0, dim, 2, dtype=torch.float32) / dim)) freqs = 1.0 / (base ** (torch.arange(0, dim, 2, dtype=torch.float32) / dim))
factor = rope_scaling.get("factor", 8.0) factor = rope_scaling.get("factor", 8.0)
@ -3820,7 +3820,7 @@ class ExaoneModel(Model):
if rope_scaling := self.find_hparam(["rope_scaling"], optional=True): if rope_scaling := self.find_hparam(["rope_scaling"], optional=True):
if rope_scaling.get("rope_type", '').lower() == "llama3": if rope_scaling.get("rope_type", '').lower() == "llama3":
base = self.hparams.get("rope_theta", 10000.0) base = self.hparams.get("rope_theta", 10000.0)
dim = self.hparams["hidden_size"] // self.hparams["num_attention_heads"] dim = self.hparams.get("head_dim", self.hparams["hidden_size"] // self.hparams["num_attention_heads"])
freqs = 1.0 / (base ** (torch.arange(0, dim, 2, dtype=torch.float32) / dim)) freqs = 1.0 / (base ** (torch.arange(0, dim, 2, dtype=torch.float32) / dim))
factor = rope_scaling.get("factor", 8.0) factor = rope_scaling.get("factor", 8.0)

5
src/llama.cpp
View File

@ -6605,6 +6605,7 @@ static bool llm_load_tensors(
const int64_t n_embd_gqa = n_embd_v_gqa; const int64_t n_embd_gqa = n_embd_v_gqa;
const int64_t n_vocab = hparams.n_vocab; const int64_t n_vocab = hparams.n_vocab;
const int64_t n_vocab_type = hparams.n_vocab_type; const int64_t n_vocab_type = hparams.n_vocab_type;
const int64_t n_rot = hparams.n_rot;
const int64_t n_expert = hparams.n_expert; const int64_t n_expert = hparams.n_expert;
const int64_t n_expert_used = hparams.n_expert_used; const int64_t n_expert_used = hparams.n_expert_used;
const int64_t n_ctx_train = hparams.n_ctx_train; const int64_t n_ctx_train = hparams.n_ctx_train;
@ -6662,7 +6663,7 @@ static bool llm_load_tensors(
layer.ffn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}); 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)); layer.rope_freqs = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ROPE_FREQS, "weight"), {n_rot/2}, llama_model_loader::TENSOR_NOT_REQUIRED | (i != 0 ? llama_model_loader::TENSOR_DUPLICATED : 0));
if (n_expert == 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_gate = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff});
@ -8193,7 +8194,7 @@ static bool llm_load_tensors(
layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd}); layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd});
layer.ffn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}); 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)); layer.rope_freqs = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ROPE_FREQS, "weight"), {n_rot/2}, llama_model_loader::TENSOR_NOT_REQUIRED | (i != 0 ? llama_model_loader::TENSOR_DUPLICATED : 0));
layer.ffn_gate = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}); 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}); layer.ffn_down = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd});
layer.ffn_up = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}); layer.ffn_up = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff});