Skip to content

Commit 2d87c9c

Browse files
monatismonatis
authored
llama : refactor tensor names (#2622)
* gguf: update tensor names searched in quantization * gguf : define tensor names as constants
1 parent da424b6 commit 2d87c9c

File tree

1 file changed

+33
-22
lines changed

1 file changed

+33
-22
lines changed

‎gguf-llama.cpp‎

Lines changed: 33 additions & 22 deletions
Original file line numberDiff line numberDiff line change
@@ -56,6 +56,20 @@
5656
#pragma warning(disable: 4244 4267) // possible loss of data
5757
#endif
5858

59+
// tensor names
60+
#define TN_TOKEN_EMBD "token_embd.weight"
61+
#define TN_OUTPUT_NORM "output_norm.weight"
62+
#define TN_OUTPUT "output.weight"
63+
#define TN_ATTN_NORM "blk.%d.attn_norm.weight"
64+
#define TN_ATTN_Q "blk.%d.attn_q.weight"
65+
#define TN_ATTN_K "blk.%d.attn_k.weight"
66+
#define TN_ATTN_V "blk.%d.attn_v.weight"
67+
#define TN_ATTN_OUTPUT "blk.%d.attn_output.weight"
68+
#define TN_FFN_NORM "blk.%d.ffn_norm.weight"
69+
#define TN_FFN_GATE "blk.%d.ffn_gate.weight"
70+
#define TN_FFN_DOWN "blk.%d.ffn_down.weight"
71+
#define TN_FFN_UP "blk.%d.ffn_up.weight"
72+
5973
static void llama_log_internal(llama_log_level level, const char* format, ...);
6074
static void llama_log_callback_default(llama_log_level level, const char * text, void * user_data);
6175
#define LLAMA_LOG_INFO(...) llama_log_internal(LLAMA_LOG_LEVEL_INFO , __VA_ARGS__)
@@ -1310,7 +1324,7 @@ static void llama_model_load_internal(
13101324

13111325
ml->ggml_ctx = ctx;
13121326

1313-
model.tok_embeddings = ml->get_tensor("token_embd.weight", {n_embd, n_vocab}, GGML_BACKEND_CPU);
1327+
model.tok_embeddings = ml->get_tensor(TN_TOKEN_EMBD, {n_embd, n_vocab}, GGML_BACKEND_CPU);
13141328

13151329
// "output" tensor
13161330
{
@@ -1331,8 +1345,8 @@ static void llama_model_load_internal(
13311345
backend_output = GGML_BACKEND_CPU;
13321346
}
13331347

1334-
model.norm = ml->get_tensor("output_norm.weight", {n_embd}, backend_norm);
1335-
model.output = ml->get_tensor("output.weight", {n_embd, n_vocab}, backend_output);
1348+
model.norm = ml->get_tensor(TN_OUTPUT_NORM, {n_embd}, backend_norm);
1349+
model.output = ml->get_tensor(TN_OUTPUT, {n_embd, n_vocab}, backend_output);
13361350
if (backend_norm == GGML_BACKEND_GPU) {
13371351
vram_weights += ggml_nbytes(model.norm);
13381352
}
@@ -1349,21 +1363,18 @@ static void llama_model_load_internal(
13491363
const ggml_backend backend_split = int(i) < i_gpu_start ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD_SPLIT; // NOLINT
13501364

13511365
auto & layer = model.layers[i];
1366+
layer.attention_norm = ml->get_tensor(format(TN_ATTN_NORM, i), {n_embd}, backend);
13521367

1353-
std::string layers_i = "blk." + std::to_string(i);
1354-
1355-
layer.attention_norm = ml->get_tensor(layers_i + ".attn_norm.weight", {n_embd}, backend);
1356-
1357-
layer.wq = ml->get_tensor(layers_i + ".attn_q.weight", {n_embd, n_embd}, backend_split);
1358-
layer.wk = ml->get_tensor(layers_i + ".attn_k.weight", {n_embd, n_embd_gqa}, backend_split);
1359-
layer.wv = ml->get_tensor(layers_i + ".attn_v.weight", {n_embd, n_embd_gqa}, backend_split);
1360-
layer.wo = ml->get_tensor(layers_i + ".attn_output.weight", {n_embd, n_embd}, backend_split);
1368+
layer.wq = ml->get_tensor(format(TN_ATTN_Q, i), {n_embd, n_embd}, backend_split);
1369+
layer.wk = ml->get_tensor(format(TN_ATTN_K, i), {n_embd, n_embd_gqa}, backend_split);
1370+
layer.wv = ml->get_tensor(format(TN_ATTN_V, i), {n_embd, n_embd_gqa}, backend_split);
1371+
layer.wo = ml->get_tensor(format(TN_ATTN_OUTPUT, i), {n_embd, n_embd}, backend_split);
13611372

1362-
layer.ffn_norm = ml->get_tensor(layers_i + ".ffn_norm.weight", {n_embd}, backend);
1373+
layer.ffn_norm = ml->get_tensor(format(TN_FFN_NORM, i), {n_embd}, backend);
13631374

1364-
layer.w1 = ml->get_tensor(layers_i + ".ffn_gate.weight", {n_embd, n_ff}, backend_split);
1365-
layer.w2 = ml->get_tensor(layers_i + ".ffn_down.weight", { n_ff, n_embd}, backend_split);
1366-
layer.w3 = ml->get_tensor(layers_i + ".ffn_up.weight", {n_embd, n_ff}, backend_split);
1375+
layer.w1 = ml->get_tensor(format(TN_FFN_GATE, i), {n_embd, n_ff}, backend_split);
1376+
layer.w2 = ml->get_tensor(format(TN_FFN_DOWN, i), { n_ff, n_embd}, backend_split);
1377+
layer.w3 = ml->get_tensor(format(TN_FFN_UP, i), {n_embd, n_ff}, backend_split);
13671378

13681379
if (backend == GGML_BACKEND_GPU) {
13691380
vram_weights +=
@@ -3240,10 +3251,10 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
32403251
int n_attention_wv = 0;
32413252
int n_feed_forward_w2 = 0;
32423253
for (auto& tensor : model_loader->tensors_map.tensors) {
3243-
if (tensor.name.find("attention.wv.weight") != std::string::npos) {
3254+
if (tensor.name.find("attn_v.weight") != std::string::npos) {
32443255
++n_attention_wv;
32453256
}
3246-
else if (tensor.name.find("feed_forward.w2.weight") != std::string::npos) {
3257+
else if (tensor.name.find("ffn_down.weight") != std::string::npos) {
32473258
++n_feed_forward_w2;
32483259
}
32493260
}
@@ -3298,13 +3309,13 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
32983309
} else {
32993310
new_type = quantized_type;
33003311
#ifdef GGML_USE_K_QUANTS
3301-
if (tensor.name == "output.weight") {
3312+
if (tensor.name == TN_OUTPUT) {
33023313
int nx = tensor.ne.at(0);
33033314
int ny = tensor.ne.at(1);
33043315
if (nx % QK_K == 0 && ny % QK_K == 0) {
33053316
new_type = GGML_TYPE_Q6_K;
33063317
}
3307-
} else if (tensor.name.find("attention.wv.weight") != std::string::npos) {
3318+
} else if (tensor.name.find("attn_v.weight") != std::string::npos) {
33083319
if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M || ftype == LLAMA_FTYPE_MOSTLY_Q2_K) new_type = GGML_TYPE_Q4_K;
33093320
else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_L) new_type = GGML_TYPE_Q5_K;
33103321
else if ((ftype == LLAMA_FTYPE_MOSTLY_Q4_K_M || ftype == LLAMA_FTYPE_MOSTLY_Q5_K_M) &&
@@ -3319,7 +3330,7 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
33193330
use_more_bits(i_feed_forward_w2, n_feed_forward_w2)) new_type = GGML_TYPE_Q6_K;
33203331
//else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_S && i_feed_forward_w2 < n_feed_forward_w2/8) new_type = GGML_TYPE_Q6_K;
33213332
++i_feed_forward_w2;
3322-
} else if (tensor.name.find("attention.wo.weight") != std::string::npos) {
3333+
} else if (tensor.name.find("attn_output.weight") != std::string::npos) {
33233334
if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M || ftype == LLAMA_FTYPE_MOSTLY_Q2_K) new_type = GGML_TYPE_Q4_K;
33243335
else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_L) new_type = GGML_TYPE_Q5_K;
33253336
}
@@ -3334,10 +3345,10 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
33343345
}
33353346
}
33363347
if (convert_incompatible_tensor) {
3337-
if (tensor.name == "output.weight") {
3348+
if (tensor.name == TN_OUTPUT) {
33383349
new_type = GGML_TYPE_F16; //fall back to F16 instead of just failing.
33393350
LLAMA_LOG_WARN("F16 will be used for this tensor instead.\n");
3340-
} else if (tensor.name == "tok_embeddings.weight") {
3351+
} else if (tensor.name == TN_TOKEN_EMBD) {
33413352
new_type = GGML_TYPE_Q4_0; //fall back to Q4_0 instead of just failing.
33423353
LLAMA_LOG_WARN("Q4_0 will be used for this tensor instead.\n");
33433354
} else {

0 commit comments

Comments
 (0)