Skip to content

Commit dd42a71

Browse files
author
Joan Martinez
committed
fix: do not complicate things
1 parent cc0ac09 commit dd42a71

File tree

3 files changed

+7
-33
lines changed

3 files changed

+7
-33
lines changed

‎gguf-py/gguf/constants.py

Lines changed: 1 addition & 6 deletions
Original file line numberDiff line numberDiff line change
@@ -186,8 +186,6 @@ class MODEL_TENSOR(IntEnum):
186186
ATTN_Q_NORM = auto()
187187
ATTN_K_NORM = auto()
188188
LAYER_OUT_NORM = auto()
189-
LAYER_NORM_1 = auto()
190-
LAYER_NORM_2 = auto()
191189
SSM_IN = auto()
192190
SSM_CONV1D = auto()
193191
SSM_X = auto()
@@ -276,8 +274,6 @@ class MODEL_TENSOR(IntEnum):
276274
MODEL_TENSOR.FFN_DOWN_EXP: "blk.{bid}.ffn_down_exps",
277275
MODEL_TENSOR.FFN_UP_EXP: "blk.{bid}.ffn_up_exps",
278276
MODEL_TENSOR.LAYER_OUT_NORM: "blk.{bid}.layer_output_norm",
279-
MODEL_TENSOR.LAYER_NORM_1: "blk.{bid}.layer_norm_1",
280-
MODEL_TENSOR.LAYER_NORM_2: "blk.{bid}.layer_norm_2",
281277
MODEL_TENSOR.SSM_IN: "blk.{bid}.ssm_in",
282278
MODEL_TENSOR.SSM_CONV1D: "blk.{bid}.ssm_conv1d",
283279
MODEL_TENSOR.SSM_X: "blk.{bid}.ssm_x",
@@ -430,8 +426,7 @@ class MODEL_TENSOR(IntEnum):
430426
MODEL_TENSOR.FFN_GATE,
431427
MODEL_TENSOR.FFN_DOWN,
432428
MODEL_TENSOR.LAYER_OUT_NORM,
433-
MODEL_TENSOR.LAYER_NORM_1,
434-
MODEL_TENSOR.LAYER_NORM_2,
429+
MODEL_TENSOR.ATTN_NORM_2,
435430
],
436431
MODEL_ARCH.MPT: [
437432
MODEL_TENSOR.TOKEN_EMBD,

‎gguf-py/gguf/tensor_mapping.py

Lines changed: 1 addition & 11 deletions
Original file line numberDiff line numberDiff line change
@@ -102,6 +102,7 @@ class TensorNameMap:
102102
# Attention norm 2
103103
MODEL_TENSOR.ATTN_NORM_2: (
104104
"transformer.h.{bid}.ln_attn", # falcon40b
105+
"encoder.layer.{bid}.layer_norm_1", # jina-v2-code
105106
),
106107

107108
# Attention query-key-value
@@ -351,20 +352,9 @@ class TensorNameMap:
351352
"encoder.layers.{bid}.norm2", # nomic-bert
352353
"transformer.decoder_layer.{bid}.rms_norm_3", # Grok
353354
"encoder.layer.{bid}.mlp.layernorm", # jina-bert-v2
354-
"encoder.layer.{bid}.layer_norm_1", # jina-v2-code
355355
"encoder.layer.{bid}.layer_norm_2" # jina-v2-code
356356
),
357357

358-
359-
MODEL_TENSOR.LAYER_NORM_1: (
360-
"encoder.layer.{bid}.layer_norm_1", # jina-v2-code
361-
),
362-
363-
364-
MODEL_TENSOR.LAYER_NORM_2: (
365-
"encoder.layer.{bid}.layer_norm_2", # jina-v2-code
366-
),
367-
368358
MODEL_TENSOR.SSM_IN: (
369359
"model.layers.{bid}.in_proj",
370360
"backbone.layers.{bid}.mixer.in_proj",

‎llama.cpp

Lines changed: 5 additions & 16 deletions
Original file line numberDiff line numberDiff line change
@@ -496,8 +496,6 @@ enum llm_tensor {
496496
LLM_TENSOR_ATTN_KV_B,
497497
LLM_TENSOR_ATTN_Q_A_NORM,
498498
LLM_TENSOR_ATTN_KV_A_NORM,
499-
LLM_TENSOR_LAYER_NORM_1,
500-
LLM_TENSOR_LAYER_NORM_2,
501499
};
502500

503501
static const std::map<llm_arch, std::map<llm_tensor, std::string>> LLM_TENSOR_NAMES = {
@@ -719,8 +717,7 @@ static const std::map<llm_arch, std::map<llm_tensor, std::string>> LLM_TENSOR_NA
719717
{ LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" },
720718
{ LLM_TENSOR_FFN_GATE, "blk.%d.ffn_gate" },
721719
{ LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" },
722-
{ LLM_TENSOR_LAYER_NORM_1, "blk.%d.layer_norm_1" },
723-
{ LLM_TENSOR_LAYER_NORM_2, "blk.%d.layer_norm_2" },
720+
{ LLM_TENSOR_ATTN_NORM_2, "blk.%d.attn_norm_2" },
724721
},
725722
},
726723
{
@@ -2017,8 +2014,6 @@ struct llama_layer {
20172014
// extra normalization layers needed by `jina-embeddings-v2-base-code`
20182015
struct ggml_tensor * layer_norm_1;
20192016
struct ggml_tensor * layer_norm_1_b;
2020-
struct ggml_tensor * layer_norm_2;
2021-
struct ggml_tensor * layer_norm_2_b;
20222017

20232018
// ff
20242019
struct ggml_tensor * ffn_gate; // w1
@@ -4680,7 +4675,8 @@ static void llm_load_vocab(
46804675
tokenizer_pre == "jina-es" ||
46814676
tokenizer_pre == "jina-de" ||
46824677
tokenizer_pre == "jina-v2-es" ||
4683-
tokenizer_pre == "jina-v2-de") {
4678+
tokenizer_pre == "jina-v2-de" ||
4679+
tokenizer_pre == "jina-v2-code") {
46844680
vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_GPT2;
46854681
} else if (
46864682
tokenizer_pre == "refact") {
@@ -5547,12 +5543,9 @@ static bool llm_load_tensors(
55475543
layer.attn_out_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_OUT_NORM, "weight", i), {n_embd}); //output_norm
55485544
layer.attn_out_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_OUT_NORM, "bias", i), {n_embd});
55495545

5550-
layer.layer_norm_1 = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_LAYER_NORM_1, "weight", i), {n_embd}, llama_model_loader::TENSOR_NOT_REQUIRED);
5551-
layer.layer_norm_1_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_LAYER_NORM_1, "bias", i), {n_embd}, llama_model_loader::TENSOR_NOT_REQUIRED);
5546+
layer.layer_norm_1 = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM_2, "weight", i), {n_embd}, llama_model_loader::TENSOR_NOT_REQUIRED);
5547+
layer.layer_norm_1_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM_2, "bias", i), {n_embd}, llama_model_loader::TENSOR_NOT_REQUIRED);
55525548

5553-
layer.layer_norm_2 = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_LAYER_NORM_2, "weight", i), {n_embd}, llama_model_loader::TENSOR_NOT_REQUIRED);
5554-
layer.layer_norm_2_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_LAYER_NORM_2, "bias", i), {n_embd}, llama_model_loader::TENSOR_NOT_REQUIRED);
5555-
55565549
layer.ffn_up = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff});
55575550
layer.ffn_gate = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff});
55585551

@@ -8520,10 +8513,6 @@ struct llm_build_context {
85208513
cur = llm_build_norm(ctx0, cur, hparams, model.layers[il].layer_norm_1, model.layers[il].layer_norm_1_b, LLM_NORM, cb, il);
85218514
}
85228515

8523-
if (model.layers[il].layer_norm_2 != nullptr) {
8524-
cur = llm_build_norm(ctx0, cur, hparams, model.layers[il].layer_norm_2, model.layers[il].layer_norm_2_b, LLM_NORM, cb, il);
8525-
}
8526-
85278516
struct ggml_tensor * ffn_inp = cur;
85288517
cb(ffn_inp, "ffn_inp", il);
85298518

0 commit comments

Comments
 (0)