Add logit_bias to the OpenAI api (ggml-org#577)

* Add logit_bias to the OpenAI api

* Cleanup and refactor, test in swagger.

---------

Co-authored-by: Concedo <39025047+LostRuins@users.noreply.github.com>

Author: DebuggingLife46

expose.h

@@ -3,6 +3,7 @@
 const int stop_token_max = 16;
 const int ban_token_max = 16;
 const int tensor_split_max = 16;
+const int logit_bias_max = 16;
 // match kobold's sampler list and order
 enum samplers
 {
@@ -22,6 +23,10 @@ enum stop_reason
     EOS_TOKEN=1,
     CUSTOM_STOPPER=2,
 };
+struct logit_bias {
+    int32_t token_id;
+    float bias;
+};
 struct load_model_inputs
 {
     const int threads;
@@ -76,6 +81,7 @@ struct generation_inputs
     const char * grammar;
     const bool grammar_retain_state;
     const bool quiet = false;
+    const logit_bias logit_biases[logit_bias_max];
 };
 struct generation_outputs
 {

gpttype_adapter.cpp

@@ -101,6 +101,7 @@ static int stopper_unused_tokens = 0;
 static std::mutex concat_output_mtx;
 static std::string concat_output = "";
 static std::string concat_output_reader_copy = "";
+static std::vector<logit_bias> logit_biases;
 
 const int extra_context_handle_fragmentation = 80;
 
@@ -489,6 +490,12 @@ int mirostat, float mirostat_tau, float mirostat_eta, const std::vector<samplers
         candidates.emplace_back(llama_token_data{token_id, logits[token_id], 0.0f});
     }
 
+    for(int i=0;i<logit_biases.size();++i)
+    {
+        auto & itm = logit_biases[i];
+        candidates[itm.token_id].logit += itm.bias;
+    }
+
     llama_token_data_array candidates_p = { candidates.data(), candidates.size(), false };
 
     if (grammar != nullptr) {
@@ -1437,6 +1444,17 @@ generation_outputs gpttype_generate(const generation_inputs inputs, generation_o
         }
     }
 
+    logit_biases.clear();
+    for(int x=0;x<logit_bias_max;++x)
+    {
+        int32_t t_id = inputs.logit_biases[x].token_id;
+        float bias = inputs.logit_biases[x].bias;
+        if(t_id >= 0 && t_id < n_vocab && bias!=0)
+        {
+           logit_biases.push_back(inputs.logit_biases[x]);
+        }
+    }
+
     std::string addedmemory = inputs.memory;
     params.prompt = inputs.prompt;
     params.seed = inputs.seed;

kcpp_docs.embd

@@ -8,7 +8,7 @@
 
    <!-- schema -->
    <script>
-      let spec = {
+     let spec = {
             "components": {
                "schemas": {
                   "BasicError": {
@@ -176,7 +176,17 @@
                            "default": false,
                            "description": "KoboldCpp ONLY. If true, also removes detected stop_sequences from the output and truncates all text after them. Does not work with SSE streaming.",
                            "type": "boolean"
-                        }
+                        },
+                        "logit_bias": {
+                           "default": {},
+                           "description": "KoboldCpp ONLY. An dictionary of key-value pairs, which indicate the token IDs (int) and logit bias (float) to apply for that token. Up to 16 value can be provided.",
+                           "type": "object",
+                           "example": {
+                              "2": -20,
+                              "145": -1.4,
+                              "3105": 3.2
+                           },
+                        },
                      },
                      "required": [
                         "prompt"

koboldcpp.py

@@ -18,6 +18,9 @@
 stop_token_max = 16
 ban_token_max = 16
 tensor_split_max = 16
+logit_bias_max = 16
+bias_min_value = -100.0
+bias_max_value = 100.0
 
 class load_model_inputs(ctypes.Structure):
     _fields_ = [("threads", ctypes.c_int),
@@ -44,6 +47,10 @@ class load_model_inputs(ctypes.Structure):
                 ("banned_tokens", ctypes.c_char_p * ban_token_max),
                 ("tensor_split", ctypes.c_float * tensor_split_max)]
 
+class logit_bias(ctypes.Structure):
+    _fields_ = [("token_id", ctypes.c_int32),
+                ("bias", ctypes.c_float)]
+
 class generation_inputs(ctypes.Structure):
     _fields_ = [("seed", ctypes.c_int),
                 ("prompt", ctypes.c_char_p),
@@ -70,7 +77,8 @@ class generation_inputs(ctypes.Structure):
                 ("stream_sse", ctypes.c_bool),
                 ("grammar", ctypes.c_char_p),
                 ("grammar_retain_state", ctypes.c_bool),
-                ("quiet", ctypes.c_bool)]
+                ("quiet", ctypes.c_bool),
+                ("logit_biases", logit_bias * logit_bias_max)]
 
 class generation_outputs(ctypes.Structure):
     _fields_ = [("status", ctypes.c_int),
@@ -301,7 +309,7 @@ def load_model(model_filename):
     ret = handle.load_model(inputs)
     return ret
 
-def generate(prompt, memory="", max_length=32, max_context_length=512, temperature=0.7, top_k=100, top_a=0.0, top_p=0.92, min_p=0.0, typical_p=1.0, tfs=1.0, rep_pen=1.1, rep_pen_range=128, presence_penalty=0.0, mirostat=0, mirostat_tau=5.0, mirostat_eta=0.1, sampler_order=[6,0,1,3,4,2,5], seed=-1, stop_sequence=[], use_default_badwordsids=False, stream_sse=False, grammar='', grammar_retain_state=False, genkey='', trimstop=False, quiet=False):
+def generate(prompt, memory="", max_length=32, max_context_length=512, temperature=0.7, top_k=100, top_a=0.0, top_p=0.92, min_p=0.0, typical_p=1.0, tfs=1.0, rep_pen=1.1, rep_pen_range=128, presence_penalty=0.0, mirostat=0, mirostat_tau=5.0, mirostat_eta=0.1, sampler_order=[6,0,1,3,4,2,5], seed=-1, stop_sequence=[], use_default_badwordsids=False, stream_sse=False, grammar='', grammar_retain_state=False, genkey='', trimstop=False, quiet=False, logit_biases={}):
     global maxctx, args, currentusergenkey, totalgens
     inputs = generation_inputs()
     outputs = ctypes.create_unicode_buffer(ctypes.sizeof(generation_outputs))
@@ -355,6 +363,28 @@ def generate(prompt, memory="", max_length=32, max_context_length=512, temperatu
             inputs.stop_sequence[n] = "".encode("UTF-8")
         else:
             inputs.stop_sequence[n] = stop_sequence[n].encode("UTF-8")
+
+    bias_list = []
+    try:
+        if logit_biases and len(logit_biases) > 0:
+            bias_list = [{"key": key, "value": value} for key, value in logit_biases.items()]
+    except Exception as ex:
+        print(f"Logit bias dictionary is invalid: {ex}")
+
+    for n in range(logit_bias_max):
+        if n >= len(bias_list):
+            inputs.logit_biases[n] = logit_bias(-1, 0.0)
+        else:
+            try:
+                t_id = int(bias_list[n]['key'])
+                bias = float(bias_list[n]['value'])
+                t_id = -1 if t_id < 0 else t_id
+                bias = (bias_max_value if bias > bias_max_value else (bias_min_value if bias < bias_min_value else bias))
+                inputs.logit_biases[n] = logit_bias(t_id, bias)
+            except Exception as ex:
+                inputs.logit_biases[n] = logit_bias(-1, 0.0)
+                print(f"Skipped unparsable logit bias:{ex}")
+
     currentusergenkey = genkey
     totalgens += 1
     ret = handle.generate(inputs,outputs)
@@ -515,7 +545,8 @@ def run_blocking(): #api format 1=basic,2=kai,3=oai,4=oai-chat
                 grammar_retain_state = genparams.get('grammar_retain_state', False),
                 genkey=genparams.get('genkey', ''),
                 trimstop=genparams.get('trim_stop', False),
-                quiet=is_quiet)
+                quiet=is_quiet,
+                logit_biases=genparams.get('logit_bias', {}))
 
         recvtxt = ""
         if stream_flag: