Initial batched inference using llama_batch

java/src/org/futo/inputmethod/latin/xlm/LanguageModel.java

@@ -82,16 +82,16 @@ public class LanguageModel extends Dictionary {
             @Override public void run() {
                 if(mNativeState != 0) return;
 
-                String modelPath = getPathToModelResource(context, R.raw.l2_steps_12k_w1_s1_1k, R.raw.l2_steps_12k_w1_s1_1k_tokenizer, false);
+                String modelPath = getPathToModelResource(context, R.raw.ml3, R.raw.ml3_tokenizer, true);
                 mNativeState = openNative(modelPath);
 
                 if(mNativeState == 0){
-                    modelPath = getPathToModelResource(context, R.raw.l2_steps_12k_w1_s1_1k, R.raw.l2_steps_12k_w1_s1_1k_tokenizer, true);
+                    modelPath = getPathToModelResource(context, R.raw.ml3, R.raw.ml3_tokenizer, true);
                     mNativeState = openNative(modelPath);
                 }
 
                 if(mNativeState == 0){
-                    throw new RuntimeException("Failed to load R.raw.l2_steps_12k_w1_s1_1k, R.raw.l2_steps_12k_w1_s1_1k_tokenizer model");
+                    throw new RuntimeException("Failed to load R.raw.ml3, R.raw.ml3_tokenizer model");
                 }
             }
         };
@@ -159,10 +159,12 @@ public class LanguageModel extends Dictionary {
 
             String word = outStrings[i].trim();
 
-            if(outProbabilities[i] > 150.0f) {
+            if(!partialWord.isEmpty()) {
                 kind = SuggestedWords.SuggestedWordInfo.KIND_WHITELIST | SuggestedWords.SuggestedWordInfo.KIND_FLAG_APPROPRIATE_FOR_AUTO_CORRECTION;
             }
 
+            Log.d("LanguageModel", "probability for word [" + word + "] is 100 * " + String.valueOf(outProbabilities[i]));
+
             suggestions.add(new SuggestedWords.SuggestedWordInfo( word, context, (int)(outProbabilities[i] * 100.0f), kind, this, 0, 0 ));
         }
 

native/jni/org_futo_inputmethod_latin_xlm_LanguageModel.cpp

@@ -39,10 +39,41 @@ static inline void sortProbabilityPairVectorDescending(std::vector<std::pair<flo
     std::partial_sort(vec.begin(), vec.begin() + partial, vec.end(), sortProbabilityPairDescending<T>);
 }
 
+typedef struct potential_sequence_data {
+    token_sequence tokens;
+    llama_seq_id seq_id;
+} potential_sequence_data;
+
+// P = P(tokens[0]) * P(tokens[1]) * [...]
+typedef std::pair<float, potential_sequence_data> potential_sequence;
+
+static void softmax(float * input, size_t input_len) {
+    float m = -INFINITY;
+    for (size_t i = 0; i < input_len; i++) {
+        if (input[i] > m) {
+            m = input[i];
+        }
+    }
+
+    float sum = 0.0;
+    for (size_t i = 0; i < input_len; i++) {
+        sum += expf(input[i] - m);
+    }
+
+    float offset = m + logf(sum);
+    for (size_t i = 0; i < input_len; i++) {
+        input[i] = expf(input[i] - offset);
+    }
+}
+
 struct LanguageModelState {
     LanguageModel *model;
 
     struct {
+        int SPACE;
+
+        std::vector<int> SAMPLING_BAD_TOKENS;
+
         int XBU;
         int XBC;
         int XEC;
@@ -57,10 +88,16 @@ struct LanguageModelState {
             return false;
         }
 
-        specialTokens.XBU = 104; //model->tokenToId("_XBU_");
-        specialTokens.XBC = 105; //model->tokenToId("_XBC_");
-        specialTokens.XEC = 106; //model->tokenToId("_XEC_");
-        specialTokens.LETTERS_TO_IDS[0] = 124; //model->tokenToId("_XU_LETTER_A_");
+        specialTokens.SPACE = 560; //model->tokenToId("▁");
+
+        // TODO: Don't hardcode these
+        // BOS, EOS, etc and some whitespace (linebreak, tab, carriage return)
+        specialTokens.SAMPLING_BAD_TOKENS = { 0, 1, 2, 3, 126, 127, 128, 129, 130 };
+
+        specialTokens.XBU = model->tokenToId("<XBU>");
+        specialTokens.XBC = model->tokenToId("<XBC>");
+        specialTokens.XEC = model->tokenToId("<XEC>");
+        specialTokens.LETTERS_TO_IDS[0] = model->tokenToId("<CHAR_A>");
 
         ASSERT(specialTokens.XBU != 0);
         ASSERT(specialTokens.XBC != 0);
@@ -74,8 +111,222 @@ struct LanguageModelState {
         return true;
     }
 
-    std::pair<float, token_sequence> Sample(){
-        float probability = 0.0f;
+    void transform_logits(float *logits, size_t n_vocab, bool allow_space){
+        softmax(logits, n_vocab);
+
+        logits[specialTokens.XBU] = -999.0f;
+
+        for(int x : specialTokens.SAMPLING_BAD_TOKENS) {
+            logits[x] = -999.0f;
+        }
+
+        if(!allow_space) {
+            logits[specialTokens.SPACE] = -999.0f;
+        }
+    }
+
+    std::vector<std::pair<float, token_sequence>> Sample(const token_sequence &prompt, int n_results) {
+        // TODO: Something seems wrong currently with kv_cache
+
+        llama_context *ctx = ((LlamaAdapter *) model->adapter)->context;
+        llama_batch batch = ((LlamaAdapter *) model->adapter)->batch;
+
+        size_t n_vocab = llama_n_vocab(llama_get_model(ctx));
+
+        std::vector<potential_sequence> sequences;
+
+        auto prompt_ff = transformer_context_fastforward(model->transformerContext, prompt);
+        llama_kv_cache_seq_rm(ctx, 0, prompt_ff.second, -1);
+
+        batch.n_tokens = prompt_ff.first.size();
+        for (int i = 0; i < prompt_ff.first.size(); i++) {
+            batch.token[i] = prompt_ff.first[i];
+            batch.pos[i] = prompt_ff.second + i;
+            batch.seq_id[i] = 0;
+            batch.logits[i] = false;
+        }
+
+        batch.logits[prompt_ff.first.size() - 1] = true;
+
+        if (llama_decode(ctx, batch) != 0) {
+            AKLOGE("llama_decode() failed");
+            return {};
+        }
+
+        transformer_context_apply(model->transformerContext, prompt_ff);
+
+        float *logits = llama_get_logits_ith(ctx, prompt_ff.first.size() - 1);
+        transform_logits(logits, n_vocab, false);
+
+        std::vector<std::pair<float, int>> index_value;
+        index_value.clear();
+        for (size_t i = 0; i < n_vocab; i++) {
+            index_value.emplace_back(logits[i], i);
+        }
+
+        sortProbabilityPairVectorDescending(index_value, n_results);
+
+        for (int i = 0; i < n_results; i++) {
+            sequences.emplace_back(
+                    index_value[i].first,
+                    potential_sequence_data{
+                            {index_value[i].second},
+                            i
+                    }
+            );
+        }
+
+        for (auto &sequence: sequences) {
+            if (sequence.second.seq_id == 0) continue;
+
+            llama_kv_cache_seq_cp(ctx, 0, sequence.second.seq_id, 0, prompt.size());
+        }
+
+        std::vector<potential_sequence> next_sequences;
+
+        std::vector<std::pair<float, token_sequence>> outputs;
+
+        while (true) {
+            next_sequences.clear();
+            for (auto sequence: std::move(sequences)) {
+                int next_token = sequence.second.tokens[sequence.second.tokens.size() - 1];
+
+                // Check if this is the end of correction
+                if (next_token == specialTokens.XEC) {
+                    token_sequence resulting_tokens = std::move(sequence.second.tokens);
+                    resulting_tokens.resize(resulting_tokens.size() - 1);
+                    outputs.emplace_back(sequence.first, resulting_tokens);
+                    continue;
+                }
+
+                // Check if this is the end of a word
+                std::string token = model->getToken(next_token);
+                if (token.size() >= 3 && (token[token.size() - 1] == '\x81') &&
+                    (token[token.size() - 2] == '\x96') && token[token.size() - 3] == '\xe2') {
+                    outputs.emplace_back(sequence.first, std::move(sequence.second.tokens));
+                    continue;
+                }
+
+                next_sequences.emplace_back(sequence);
+            }
+
+            sequences = next_sequences;
+            next_sequences.clear();
+
+            size_t remaining_count = n_results - outputs.size();
+            batch.n_tokens = 0;
+
+            for (auto &sequence: sequences) {
+                batch.token[batch.n_tokens] = sequence.second.tokens[sequence.second.tokens.size() -
+                                                                     1];
+                batch.pos[batch.n_tokens] = prompt.size() + (sequence.second.tokens.size() - 1);
+                batch.seq_id[batch.n_tokens] = sequence.second.seq_id;
+                batch.logits[batch.n_tokens] = true;
+
+                batch.n_tokens += 1;
+            }
+
+            ASSERT(batch.n_tokens == remaining_count); // usually 3
+
+            if (batch.n_tokens == 0) {
+                break;
+            }
+
+            llama_decode(ctx, batch);
+
+            for (int seq = 0; seq < remaining_count; seq++) {
+                const potential_sequence &parent_seq = sequences[seq];
+                logits = llama_get_logits_ith(ctx, seq);
+                transform_logits(logits, n_vocab, true);
+
+                index_value.clear();
+                for (size_t i = 0; i < n_vocab; i++) {
+                    index_value.emplace_back(logits[i], i);
+                }
+
+                sortProbabilityPairVectorDescending(index_value, remaining_count);
+
+                for (size_t i = 0; i < remaining_count; i++) {
+                    token_sequence new_sequence = parent_seq.second.tokens;
+                    new_sequence.push_back(index_value[i].second);
+
+                    if (index_value[i].first > 1.0f || index_value[i].first < 0.0f) {
+                        AKLOGE("Expected index_value to be probability [%.2f]",
+                               index_value[i].first);
+                    }
+
+                    if (sequences[i].first > 1.0f || sequences[i].first < 0.0f) {
+                        AKLOGE("Expected sequences value to be probability [%.2f]",
+                               sequences[i].first);
+                    }
+
+                    next_sequences.emplace_back(
+                            index_value[i].first * sequences[i].first,
+                            potential_sequence_data{
+                                    new_sequence,
+                                    parent_seq.second.seq_id
+                            }
+                    );
+                }
+            }
+
+            sortProbabilityPairVectorDescending(next_sequences, remaining_count);
+            next_sequences.resize(remaining_count);
+            sequences.clear();
+
+            // In some cases we may have picked a sequence from the same parent sequence
+            // We must re-assign the seq_id
+            int seq_id_use_count[n_results];
+            for (int i = 0; i < n_results; i++) seq_id_use_count[i] = 0;
+
+            for (auto &seq: next_sequences) seq_id_use_count[seq.second.seq_id] += 1;
+
+            for (auto &seq: next_sequences) {
+                if (seq_id_use_count[seq.second.seq_id] > 1) {
+                    int old_seq_id = seq.second.seq_id;
+
+                    int new_seq_id = -1;
+                    for (int i = 0; i < n_results; i++) {
+                        if (seq_id_use_count[i] == 0) {
+                            new_seq_id = i;
+                            break;
+                        }
+                    }
+
+                    if (new_seq_id == -1) {
+                        AKLOGE("Couldn't find an empty sequence id to use. This should never happen.");
+                        return {};
+                    }
+
+                    seq_id_use_count[old_seq_id]--;
+                    seq_id_use_count[new_seq_id]++;
+
+                    llama_kv_cache_seq_cp(
+                            ctx,
+                            old_seq_id,
+                            new_seq_id,
+                            0, // could start from prompt.size()
+                            prompt.size() + (seq.second.tokens.size() - 1)
+                    );
+
+                    seq.second.seq_id = new_seq_id;
+                }
+            }
+
+            sequences = next_sequences;
+        }
+
+        for (int i = 1; i < n_results; i++) {
+            llama_kv_cache_seq_rm(ctx, i, 0, -1);
+        }
+
+        return outputs;
+    }
+
+    std::vector<std::pair<float, token_sequence>> SampleOld(const token_sequence &prompt, int n_results) {
+        model->updateContext(prompt);
+
+        float probability = 1.0f;
         token_sequence sampled_sequence;
 
         std::vector<std::pair<float, int>> index_value;
@@ -84,6 +335,14 @@ struct LanguageModelState {
             std::vector<float> logits = model->infer();
             logits[specialTokens.XBU] = -999.0f;
 
+            for(int x : specialTokens.SAMPLING_BAD_TOKENS) {
+                logits[x] = -999.0f;
+            }
+
+            if(sampled_sequence.empty()) {
+                logits[specialTokens.SPACE] = -999.0f;
+            }
+
             index_value.clear();
             for (size_t i = 0; i < logits.size(); i++) {
                 index_value.emplace_back(logits[i], i);
@@ -99,7 +358,7 @@ struct LanguageModelState {
                 break;
             }
 
-            probability += index_value[0].first;
+            probability *= index_value[0].first;
             sampled_sequence.push_back(next_token);
 
 
@@ -110,19 +369,24 @@ struct LanguageModelState {
             }
         }
 
-        return {probability, std::move(sampled_sequence)};
+        return {{probability, std::move(sampled_sequence)}};
     }
 
-    std::string PredictNextWord(const std::string &context) {
+    std::vector<std::pair<float, std::string>> PredictNextWord(const std::string &context) {
         token_sequence next_context = model->tokenize(trim(context) + " ");
-        model->updateContext(next_context);
+        //model->updateContext(next_context);
 
-        auto result = Sample();
+        auto results = Sample(next_context, 3);
 
-        return model->decode(result.second);
+        std::vector<std::pair<float, std::string>> str_results;
+        for(const auto& result : results) {
+            str_results.emplace_back(result.first, model->decode(result.second));
+        }
+
+        return str_results;
     }
 
-    std::string PredictCorrection(const std::string &context, std::string &word) {
+    std::vector<std::pair<float, std::string>> PredictCorrection(const std::string &context, std::string &word) {
         token_sequence next_context = model->tokenize(trim(context) + " ");
         next_context.push_back(specialTokens.XBU);
 
@@ -137,11 +401,16 @@ struct LanguageModelState {
         }
         next_context.push_back(specialTokens.XBC);
 
-        model->updateContext(next_context);
+        //model->updateContext(next_context);
 
-        auto result = Sample();
+        auto results = Sample(next_context, 3);
 
-        return model->decode(result.second);
+        std::vector<std::pair<float, std::string>> str_results;
+        for(const auto& result : results) {
+            str_results.emplace_back(result.first, model->decode(result.second));
+        }
+
+        return str_results;
     }
 };
 
@@ -204,16 +473,20 @@ namespace latinime {
         AKLOGI("LanguageModel context [%s]", contextString.c_str());
 
         bool isAutoCorrect = false;
-        std::string result;
+        std::vector<std::pair<float, std::string>> results;
         if(partialWordString.empty()) {
-            result = state->PredictNextWord(contextString);
+            results = state->PredictNextWord(contextString);
 
-            AKLOGI("LanguageModel suggestion [%s]", result.c_str());
+            for(const auto &result : results) {
+                AKLOGI("LanguageModel suggestion %.2f [%s]", result.first, result.second.c_str());
+            }
         } else {
             isAutoCorrect = true;
-            result = state->PredictCorrection(contextString, partialWordString);
+            results = state->PredictCorrection(contextString, partialWordString);
 
-            AKLOGI("LanguageModel correction [%s] -> [%s]", partialWordString.c_str(), result.c_str());
+            for(const auto &result : results) {
+                AKLOGI("LanguageModel correction %.2f [%s] -> [%s]", result.first, partialWordString.c_str(), result.second.c_str());
+            }
         }
 
         // Output
@@ -222,10 +495,10 @@ namespace latinime {
         jfloat *probsArray = env->GetFloatArrayElements(outProbabilities, nullptr);
 
         // Output predictions for next word
-        for (int i = 0; i < 1; i++) {
-            jstring jstr = env->NewStringUTF(result.c_str());
+        for (int i = 0; i < results.size(); i++) {
+            jstring jstr = env->NewStringUTF(results[i].second.c_str());
             env->SetObjectArrayElement(outPredictions, i, jstr);
-            probsArray[i] = isAutoCorrect ? 200.0f : 100.0f;
+            probsArray[i] = results[i].first;
             env->DeleteLocalRef(jstr);
         }
 
@@ -250,8 +523,24 @@ namespace latinime {
             }
     };
 
+
+    static void llama_log_callback(ggml_log_level level, const char * text, void * user_data) {
+        switch(level) {
+            case GGML_LOG_LEVEL_ERROR:
+                AKLOGE("llama err:  %s", text);
+                break;
+            case GGML_LOG_LEVEL_WARN:
+                AKLOGI("llama warn: %s", text);
+                break;
+            case GGML_LOG_LEVEL_INFO:
+                AKLOGI("llama info: %s", text);
+                break;
+        }
+    }
+
     int register_LanguageModel(JNIEnv *env) {
         llama_backend_init(true /* numa??? */);
+        llama_log_set(llama_log_callback, nullptr);
 
         const char *const kClassPathName = "org/futo/inputmethod/latin/xlm/LanguageModel";
         return registerNativeMethods(env, kClassPathName, sMethods, NELEMS(sMethods));

native/jni/src/ggml/LanguageModel.cpp

@@ -22,15 +22,15 @@ const char *LlamaAdapter::getToken(int id) const {
 
 bool LlamaAdapter::eval(int nPast, token_sequence input, std::vector<float> &outLogits) {
     // TODO
-    ASSERT(nPast + input.size() < llama_model_n_ctx(model));
+    ASSERT(nPast + input.size() < LLAMA_CONTEXT_SIZE);
 
-    if(llama_eval(context, input.data(), input.size(), nPast, numThreads) != 0) {
+    if(llama_eval(context, input.data(), input.size(), nPast) != 0) {
         return false;
     }
 
     // TODO: Zero-copy
-    outLogits.resize(llama_n_vocab(context));
-    memcpy(outLogits.data(), llama_get_logits(context), llama_n_vocab(context) * sizeof(float));
+    outLogits.resize(llama_n_vocab(model));
+    memcpy(outLogits.data(), llama_get_logits(context), llama_n_vocab(model) * sizeof(float));
 
     return true;
 }
@@ -54,8 +54,11 @@ LanguageModel *LlamaAdapter::createLanguageModel(const std::string &paths) {
     auto adapter = new LlamaAdapter();
 
     llama_context_params ctx_params = llama_context_default_params();
+    ctx_params.n_ctx = LLAMA_CONTEXT_SIZE;
 
-    adapter->model = llama_load_model_from_file(modelPath.c_str(), ctx_params);
+    llama_model_params model_params = llama_model_default_params();
+
+    adapter->model = llama_load_model_from_file(modelPath.c_str(), model_params);
 
     if(adapter->model == nullptr) {
         delete adapter;
@@ -73,6 +76,8 @@ LanguageModel *LlamaAdapter::createLanguageModel(const std::string &paths) {
         return nullptr;
     }
 
+    adapter->batch = llama_batch_init(LLAMA_CONTEXT_SIZE, 0);
+
     return new LanguageModel(adapter);
 }
 

native/jni/src/ggml/LanguageModel.h

@@ -105,14 +105,15 @@ public:
     AK_FORCE_INLINE bool isPendingEvaluation() const {
         return pendingEvaluationSequence.size() > 0;
     }
+
+    LanguageModelAdapter *adapter;
+    transformer_context transformerContext;
 private:
     token_sequence pendingContext;
     token_sequence pendingEvaluationSequence;
     int pendingNPast = 0;
 
-    LanguageModelAdapter *adapter;
 
-    transformer_context transformerContext;
 
     std::vector<float> outLogits;
     std::vector<float> tmpOutLogits;
@@ -121,6 +122,7 @@ private:
 };
 
 
+#define LLAMA_CONTEXT_SIZE 2048
 class LlamaAdapter : public LanguageModelAdapter {
 public:
     int getVocabSize() const;
@@ -131,10 +133,13 @@ public:
     virtual std::string decode(const token_sequence &tokens) const;
 
     static LanguageModel *createLanguageModel(const std::string &paths);
-private:
-    LlamaAdapter();
     llama_context *context;
     llama_model *model;
+    llama_batch batch;
+private:
+    LlamaAdapter();
+
+
     sentencepiece::SentencePieceProcessor spm;
 };