llama : add jina v2 base code (#7596)

* feat: add changes to handle jina v2 base code

* fix: do not complicate things

* fix: fix the usage of the code model

* fix: fix comments

* fix: fix linting issues

* fix: remove ollama patches

* style : minor

---------

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>

convert-hf-to-gguf-update.py

@@ -83,6 +83,7 @@ models = [
     {"name": "jina-v2-es",     "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/jinaai/jina-embeddings-v2-base-es", },
     {"name": "jina-v2-de",     "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/jinaai/jina-embeddings-v2-base-de", },
     {"name": "smaug-bpe",      "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/abacusai/Smaug-Llama-3-70B-Instruct", },
+    {"name": "jina-v2-code",   "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/jinaai/jina-embeddings-v2-base-code", },
 ]
 
 

convert-hf-to-gguf.py

@@ -475,6 +475,9 @@ class Model:
         if chkhsh == "c136ed14d01c2745d4f60a9596ae66800e2b61fa45643e72436041855ad4089d":
             # ref: https://huggingface.co/abacusai/Smaug-Llama-3-70B-Instruct
             res = "smaug-bpe"
+        if chkhsh == "7967bfa498ade6b757b064f31e964dddbb80f8f9a4d68d4ba7998fcf281c531a":
+            # ref: https://huggingface.co/jinaai/jina-embeddings-v2-base-code
+            res = "jina-v2-code"
 
         if res is None:
             logger.warning("\n")
@@ -2452,11 +2455,13 @@ class JinaBertV2Model(BertModel):
 
     def get_tensors(self):
         for name, data in super().get_tensors():
-            if 'gated_layers' in name:
+            if 'gated_layer' in name:
                 d1 = data[:self.intermediate_size, :]
                 name1 = name.replace('gated_layers', 'gated_layers_w')
+                name1 = name1.replace('up_gated_layer', 'gated_layers_v')
                 d2 = data[self.intermediate_size:, :]
                 name2 = name.replace('gated_layers', 'gated_layers_v')
+                name2 = name2.replace('up_gated_layer', 'gated_layers_w')
                 yield name1, d1
                 yield name2, d2
                 continue

gguf-py/gguf/constants.py

@@ -415,6 +415,7 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.TOKEN_EMBD,
         MODEL_TENSOR.TOKEN_EMBD_NORM,
         MODEL_TENSOR.TOKEN_TYPES,
+        MODEL_TENSOR.ATTN_NORM_2,
         MODEL_TENSOR.ATTN_OUT_NORM,
         MODEL_TENSOR.ATTN_Q,
         MODEL_TENSOR.ATTN_Q_NORM,

gguf-py/gguf/tensor_mapping.py

@@ -102,6 +102,7 @@ class TensorNameMap:
         # Attention norm 2
         MODEL_TENSOR.ATTN_NORM_2: (
             "transformer.h.{bid}.ln_attn",  # falcon40b
+            "encoder.layer.{bid}.layer_norm_1",             # jina-v2-code
         ),
 
         # Attention query-key-value
@@ -311,6 +312,7 @@ class TensorNameMap:
             "model.layers.{bid}.mlp.c_proj",                          # starcoder2
             "encoder.layer.{bid}.mlp.wo",                             # jina-bert-v2
             "model.layers.{bid}.residual_mlp.w2",                     # arctic
+            "encoder.layer.{bid}.mlp.down_layer",                     # jina-bert-v2
         ),
 
         MODEL_TENSOR.FFN_DOWN_EXP: (
@@ -350,6 +352,7 @@ class TensorNameMap:
             "encoder.layers.{bid}.norm2",                   # nomic-bert
             "transformer.decoder_layer.{bid}.rms_norm_3",   # Grok
             "encoder.layer.{bid}.mlp.layernorm",            # jina-bert-v2
+            "encoder.layer.{bid}.layer_norm_2"              # jina-v2-code
         ),
 
         MODEL_TENSOR.SSM_IN: (

llama.cpp

@@ -704,6 +704,7 @@ static const std::map<llm_arch, std::map<llm_tensor, std::string>> LLM_TENSOR_NA
             { LLM_TENSOR_TOKEN_EMBD,      "token_embd" },
             { LLM_TENSOR_TOKEN_EMBD_NORM, "token_embd_norm" },
             { LLM_TENSOR_TOKEN_TYPES,     "token_types" },
+            { LLM_TENSOR_ATTN_NORM_2,     "blk.%d.attn_norm_2" },
             { LLM_TENSOR_ATTN_OUT_NORM,   "blk.%d.attn_output_norm" },
             { LLM_TENSOR_ATTN_Q,          "blk.%d.attn_q" },
             { LLM_TENSOR_ATTN_Q_NORM,     "blk.%d.attn_q_norm" },
@@ -4653,8 +4654,7 @@ static void llm_load_vocab(
                 LLAMA_LOG_WARN("%s: ************************************        \n", __func__);
                 LLAMA_LOG_WARN("%s:                                             \n", __func__);
                 vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_DEFAULT;
-            } else if (
-                    tokenizer_pre == "default") {
+            } else if (tokenizer_pre == "default") {
                 vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_DEFAULT;
             } else if (
                     tokenizer_pre == "llama3"   ||
@@ -4681,7 +4681,8 @@ static void llm_load_vocab(
                     tokenizer_pre == "jina-es" ||
                     tokenizer_pre == "jina-de" ||
                     tokenizer_pre == "jina-v2-es" ||
-                    tokenizer_pre == "jina-v2-de") {
+                    tokenizer_pre == "jina-v2-de" ||
+                    tokenizer_pre == "jina-v2-code") {
                 vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_GPT2;
             } else if (
                     tokenizer_pre == "refact") {
@@ -5515,7 +5516,7 @@ static bool llm_load_tensors(
 
                             layer.ffn_down_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_DOWN, "bias", i),   {n_embd});
                         } else {
-                            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.layer_out_norm   = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_LAYER_OUT_NORM, "weight", i), {n_embd});
@@ -5556,6 +5557,9 @@ static bool llm_load_tensors(
                         layer.attn_out_norm   = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_OUT_NORM, "weight", i), {n_embd}); //output_norm
                         layer.attn_out_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_OUT_NORM, "bias", i),   {n_embd});
 
+                        layer.attn_norm_2   = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM_2, "weight", i), {n_embd}, llama_model_loader::TENSOR_NOT_REQUIRED);
+                        layer.attn_norm_2_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM_2, "bias", i),   {n_embd}, llama_model_loader::TENSOR_NOT_REQUIRED);
+
                         layer.ffn_up = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP,        "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});
 
@@ -8519,6 +8523,11 @@ struct llm_build_context {
             // attention layer norm
             cur = llm_build_norm(ctx0, cur, hparams, model.layers[il].attn_out_norm, model.layers[il].attn_out_norm_b, LLM_NORM, cb, il);
 
+            if (model.layers[il].attn_norm_2 != nullptr) {
+                cur = ggml_add(ctx0, cur, inpL); // re-add the layer input
+                cur = llm_build_norm(ctx0, cur, hparams, model.layers[il].attn_norm_2, model.layers[il].attn_norm_2_b, LLM_NORM, cb, il);
+            }
+
             struct ggml_tensor * ffn_inp = cur;
             cb(ffn_inp, "ffn_inp", il);