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

Fix llm_load_tensors: the asserts were not backcompat

Browse source
This commit is contained in:
Nam Nguyen 2023-12-31 16:38:58 -08:00
parent 522e534903
commit 4c1c0d68f2
1 changed files with 11 additions and 11 deletions
Download patch file Download diff file Expand all files Collapse all files

22
llama.cpp
View file

@ -3224,7 +3224,7 @@ static bool llm_load_tensors(
const uint32_t n_ff = hparams.n_ff; const uint32_t n_ff = hparams.n_ff;
const int64_t n_embd_gqa = n_embd_v_gqa; const int64_t n_embd_gqa = n_embd_v_gqa;
GGML_ASSERT(n_embd_gqa == n_embd); GGML_ASSERT(n_embd_gqa == n_embd / hparams.n_gqa());
GGML_ASSERT(n_embd_gqa == n_embd_k_gqa); GGML_ASSERT(n_embd_gqa == n_embd_k_gqa);
const int i_gpu_start = n_layer - n_gpu_layers; const int i_gpu_start = n_layer - n_gpu_layers;
@ -3295,7 +3295,7 @@ static bool llm_load_tensors(
const uint32_t n_ff = hparams.n_ff; const uint32_t n_ff = hparams.n_ff;
const int64_t n_embd_gqa = n_embd_v_gqa; const int64_t n_embd_gqa = n_embd_v_gqa;
GGML_ASSERT(n_embd_gqa == n_embd); GGML_ASSERT(n_embd_gqa == n_embd / hparams.n_gqa());
GGML_ASSERT(n_embd_gqa == n_embd_k_gqa); GGML_ASSERT(n_embd_gqa == n_embd_k_gqa);
const int i_gpu_start = n_layer - n_gpu_layers; const int i_gpu_start = n_layer - n_gpu_layers;
@ -3346,7 +3346,7 @@ static bool llm_load_tensors(
const uint32_t n_ff = hparams.n_ff; const uint32_t n_ff = hparams.n_ff;
const int64_t n_embd_gqa = n_embd_v_gqa; const int64_t n_embd_gqa = n_embd_v_gqa;
GGML_ASSERT(n_embd_gqa == n_embd); GGML_ASSERT(n_embd_gqa == n_embd / hparams.n_gqa());
GGML_ASSERT(n_embd_gqa == n_embd_k_gqa); GGML_ASSERT(n_embd_gqa == n_embd_k_gqa);
const int i_gpu_start = n_layer - n_gpu_layers; const int i_gpu_start = n_layer - n_gpu_layers;
@ -3399,7 +3399,7 @@ static bool llm_load_tensors(
const uint32_t n_ff = hparams.n_ff; const uint32_t n_ff = hparams.n_ff;
const int64_t n_embd_gqa = n_embd_v_gqa; const int64_t n_embd_gqa = n_embd_v_gqa;
GGML_ASSERT(n_embd_gqa == n_embd); GGML_ASSERT(n_embd_gqa == n_embd / hparams.n_gqa());
GGML_ASSERT(n_embd_gqa == n_embd_k_gqa); GGML_ASSERT(n_embd_gqa == n_embd_k_gqa);
const int i_gpu_start = n_layer - n_gpu_layers; const int i_gpu_start = n_layer - n_gpu_layers;
@ -3454,7 +3454,7 @@ static bool llm_load_tensors(
const uint32_t n_ff = hparams.n_ff; const uint32_t n_ff = hparams.n_ff;
const int64_t n_embd_gqa = n_embd_v_gqa; const int64_t n_embd_gqa = n_embd_v_gqa;
GGML_ASSERT(n_embd_gqa == n_embd); GGML_ASSERT(n_embd_gqa == n_embd / hparams.n_gqa());
GGML_ASSERT(n_embd_gqa == n_embd_k_gqa); GGML_ASSERT(n_embd_gqa == n_embd_k_gqa);
const int i_gpu_start = n_layer - n_gpu_layers; const int i_gpu_start = n_layer - n_gpu_layers;
@ -3507,7 +3507,7 @@ static bool llm_load_tensors(
const uint32_t n_ff = hparams.n_ff; const uint32_t n_ff = hparams.n_ff;
const int64_t n_embd_gqa = n_embd_v_gqa; const int64_t n_embd_gqa = n_embd_v_gqa;
GGML_ASSERT(n_embd_gqa == n_embd); GGML_ASSERT(n_embd_gqa == n_embd / hparams.n_gqa());
GGML_ASSERT(n_embd_gqa == n_embd_k_gqa); GGML_ASSERT(n_embd_gqa == n_embd_k_gqa);
const int i_gpu_start = n_layer - n_gpu_layers; const int i_gpu_start = n_layer - n_gpu_layers;
@ -3561,7 +3561,7 @@ static bool llm_load_tensors(
const uint32_t n_ff = hparams.n_ff; const uint32_t n_ff = hparams.n_ff;
const int64_t n_embd_gqa = n_embd_v_gqa; const int64_t n_embd_gqa = n_embd_v_gqa;
GGML_ASSERT(n_embd_gqa == n_embd); GGML_ASSERT(n_embd_gqa == n_embd / hparams.n_gqa());
GGML_ASSERT(n_embd_gqa == n_embd_k_gqa); GGML_ASSERT(n_embd_gqa == n_embd_k_gqa);
const int i_gpu_start = n_layer - n_gpu_layers; const int i_gpu_start = n_layer - n_gpu_layers;
@ -3611,7 +3611,7 @@ static bool llm_load_tensors(
const uint32_t n_ff = hparams.n_ff; const uint32_t n_ff = hparams.n_ff;
const int64_t n_embd_gqa = n_embd_v_gqa; const int64_t n_embd_gqa = n_embd_v_gqa;
GGML_ASSERT(n_embd_gqa == n_embd); GGML_ASSERT(n_embd_gqa == n_embd / hparams.n_gqa());
GGML_ASSERT(n_embd_gqa == n_embd_k_gqa); GGML_ASSERT(n_embd_gqa == n_embd_k_gqa);
const int i_gpu_start = n_layer - n_gpu_layers; const int i_gpu_start = n_layer - n_gpu_layers;
@ -3712,7 +3712,7 @@ static bool llm_load_tensors(
const uint32_t n_ff = hparams.n_ff; const uint32_t n_ff = hparams.n_ff;
const int64_t n_embd_gqa = n_embd_v_gqa; const int64_t n_embd_gqa = n_embd_v_gqa;
GGML_ASSERT(n_embd_gqa == n_embd); GGML_ASSERT(n_embd_gqa == n_embd / hparams.n_gqa());
GGML_ASSERT(n_embd_gqa == n_embd_k_gqa); GGML_ASSERT(n_embd_gqa == n_embd_k_gqa);
const int i_gpu_start = n_layer - n_gpu_layers; const int i_gpu_start = n_layer - n_gpu_layers;
@ -3764,7 +3764,7 @@ static bool llm_load_tensors(
const uint32_t n_ff = hparams.n_ff; const uint32_t n_ff = hparams.n_ff;
const int64_t n_embd_gqa = n_embd_v_gqa; const int64_t n_embd_gqa = n_embd_v_gqa;
GGML_ASSERT(n_embd_gqa == n_embd); GGML_ASSERT(n_embd_gqa == n_embd / hparams.n_gqa());
GGML_ASSERT(n_embd_gqa == n_embd_k_gqa); GGML_ASSERT(n_embd_gqa == n_embd_k_gqa);
const int i_gpu_start = n_layer - n_gpu_layers; const int i_gpu_start = n_layer - n_gpu_layers;
@ -3814,7 +3814,7 @@ static bool llm_load_tensors(
const uint32_t n_ff = hparams.n_ff; const uint32_t n_ff = hparams.n_ff;
const int64_t n_embd_gqa = n_embd_v_gqa; const int64_t n_embd_gqa = n_embd_v_gqa;
GGML_ASSERT(n_embd_gqa == n_embd); GGML_ASSERT(n_embd_gqa == n_embd / hparams.n_gqa());
GGML_ASSERT(n_embd_gqa == n_embd_k_gqa); GGML_ASSERT(n_embd_gqa == n_embd_k_gqa);
const int i_gpu_start = n_layer - n_gpu_layers; const int i_gpu_start = n_layer - n_gpu_layers;