rerank : use [SEP] token instead of [BOS] (#9737)

* rerank : use [SEP] token instead of [BOS]

ggml-ci

* common : sanity check for non-NULL tokens

ggml-ci

* ci : adjust rank score interval

ggml-ci

* ci : add shebang to run.sh

ggml-ci

ci/run.sh

@@ -1,4 +1,4 @@
-#/bin/bash
+#!/bin/bash
 #
 # sample usage:
 #
@@ -751,7 +751,8 @@ function gg_run_rerank_tiny {
 
     model_f16="${path_models}/ggml-model-f16.gguf"
 
-    (time ./bin/llama-embedding --model ${model_f16}  -p "what is panda?</s><s>hi\nwhat is panda?</s><s>it's a bear\nwhat is panda?</s><s>The giant panda (Ailuropoda melanoleuca), sometimes called a panda bear or simply panda, is a bear species endemic to China." --pooling rank --embd-normalize -1 --verbose-prompt) 2>&1 | tee -a $OUT/${ci}-rk-f16.log
+    # for this model, the SEP token is "</s>"
+    (time ./bin/llama-embedding --model ${model_f16}  -p "what is panda?</s></s>hi\nwhat is panda?</s></s>it's a bear\nwhat is panda?</s></s>The giant panda (Ailuropoda melanoleuca), sometimes called a panda bear or simply panda, is a bear species endemic to China." --pooling rank --embd-normalize -1 --verbose-prompt) 2>&1 | tee -a $OUT/${ci}-rk-f16.log
 
     # sample output
     # rerank score 0:    0.029
@@ -774,7 +775,7 @@ function gg_run_rerank_tiny {
 
     check_score "rerank score 0" "$(cat $OUT/${ci}-rk-f16.log | grep "rerank score 0")" "0.00" "0.05" | tee -a $OUT/${ci}-rk-f16.log
     check_score "rerank score 1" "$(cat $OUT/${ci}-rk-f16.log | grep "rerank score 1")" "0.00" "0.05" | tee -a $OUT/${ci}-rk-f16.log
-    check_score "rerank score 2" "$(cat $OUT/${ci}-rk-f16.log | grep "rerank score 2")" "0.10" "0.15" | tee -a $OUT/${ci}-rk-f16.log
+    check_score "rerank score 2" "$(cat $OUT/${ci}-rk-f16.log | grep "rerank score 2")" "0.10" "0.30" | tee -a $OUT/${ci}-rk-f16.log
 
     set +e
 }

common/common.cpp

@@ -838,6 +838,31 @@ struct llama_init_result llama_init_from_gpt_params(gpt_params & params) {
         return iparams;
     }
 
+    if (params.reranking) {
+        bool ok = true;
+
+        if (llama_token_bos(model) == LLAMA_TOKEN_NULL) {
+            LOG_WRN("%s: warning: model does not have a  BOS token, reranking will not work\n", __func__);
+            ok = false;
+        }
+
+        if (llama_token_eos(model) == LLAMA_TOKEN_NULL) {
+            LOG_WRN("%s: warning: model does not have an EOS token, reranking will not work\n", __func__);
+            ok = false;
+        }
+
+        if (llama_token_sep(model) == LLAMA_TOKEN_NULL) {
+            LOG_WRN("%s: warning: model does not have a  SEP token, reranking will not work\n", __func__);
+            ok = false;
+        }
+
+        if (!ok) {
+            llama_free_model(model);
+
+            return iparams;
+        }
+    }
+
     auto cparams = llama_context_params_from_gpt_params(params);
 
     llama_context * lctx = llama_new_context_with_model(model, cparams);
@@ -855,6 +880,7 @@ struct llama_init_result llama_init_from_gpt_params(gpt_params & params) {
         if (cvec.n_embd == -1) {
             llama_free(lctx);
             llama_free_model(model);
+
             return iparams;
         }
 
@@ -867,6 +893,7 @@ struct llama_init_result llama_init_from_gpt_params(gpt_params & params) {
         if (err) {
             llama_free(lctx);
             llama_free_model(model);
+
             return iparams;
         }
     }
@@ -889,7 +916,7 @@ struct llama_init_result llama_init_from_gpt_params(gpt_params & params) {
         llama_lora_adapters_apply(lctx, iparams.lora_adapters);
     }
 
-    if (params.sparams.ignore_eos && llama_token_eos(model) == -1) {
+    if (params.sparams.ignore_eos && llama_token_eos(model) == LLAMA_TOKEN_NULL) {
         LOG_WRN("%s: warning: model does not have an EOS token, ignoring --ignore-eos\n", __func__);
         params.sparams.ignore_eos = false;
     }
@@ -930,6 +957,7 @@ struct llama_init_result llama_init_from_gpt_params(gpt_params & params) {
 
     iparams.model   = model;
     iparams.context = lctx;
+
     return iparams;
 }
 

examples/server/server.cpp

@@ -2027,7 +2027,7 @@ struct server_context {
                                 continue;
                             }
 
-                            // prompt: <s>query</s><s>doc</s>
+                            // prompt: [BOS]query[EOS][SEP]doc[EOS]
                             prompt_tokens.clear();
                             prompt_tokens.push_back(llama_token_bos(model));
                             {
@@ -2035,7 +2035,7 @@ struct server_context {
                                 prompt_tokens.insert(prompt_tokens.end(), part.begin(), part.end());
                             }
                             prompt_tokens.push_back(llama_token_eos(model));
-                            prompt_tokens.push_back(llama_token_bos(model));
+                            prompt_tokens.push_back(llama_token_sep(model));
                             {
                                 const auto part = tokenize(slot.prompt[1], false);
                                 prompt_tokens.insert(prompt_tokens.end(), part.begin(), part.end());

src/llama-vocab.h

@@ -40,17 +40,17 @@ struct llama_vocab {
     id special_bos_id  = 1;
     id special_eos_id  = 2;
     id special_unk_id  = 0;
-    id special_sep_id  = -1;
+    id special_sep_id  = LLAMA_TOKEN_NULL;
-    id special_pad_id  = -1;
+    id special_pad_id  = LLAMA_TOKEN_NULL;
-    id special_cls_id  = -1;
+    id special_cls_id  = LLAMA_TOKEN_NULL;
-    id special_mask_id = -1;
+    id special_mask_id = LLAMA_TOKEN_NULL;
 
     id linefeed_id       = 13;
-    id special_prefix_id = -1;
+    id special_prefix_id = LLAMA_TOKEN_NULL;
-    id special_suffix_id = -1;
+    id special_suffix_id = LLAMA_TOKEN_NULL;
-    id special_middle_id = -1;
+    id special_middle_id = LLAMA_TOKEN_NULL;
-    id special_eot_id    = -1; // TODO: move above after "eos_id", and here add "file separator" token
+    id special_eot_id    = LLAMA_TOKEN_NULL; // TODO: move above after "eos_id", and here add "file separator" token
-    id special_eom_id    = -1;
+    id special_eom_id    = LLAMA_TOKEN_NULL;
 
     // set of all tokens that cause "end of generation"
     std::set<id> special_eog_ids;

src/llama.cpp

@@ -2412,7 +2412,7 @@ struct llama_hparams {
 
     // needed by encoder-decoder models (e.g. T5, FLAN-T5)
     // ref: https://github.com/ggerganov/llama.cpp/pull/8141
-    llama_token dec_start_token_id = -1;
+    llama_token dec_start_token_id = LLAMA_TOKEN_NULL;
 
     enum llama_pooling_type      pooling_type            = LLAMA_POOLING_TYPE_NONE;
     enum llama_rope_type         rope_type               = LLAMA_ROPE_TYPE_NONE;