backend: fix extra spaces in tokenization and a CUDA crash (#2778)

Also potentially improves accuracy of BOS insertion, token cache, and logit indexing.

Signed-off-by: Jared Van Bortel <jared@nomic.ai>

gpt4all-backend/llama.cpp-mainline

@@ -1 +1 @@
-Subproject commit c6546b0544ad2c01e8a1630b101e92336a68b036
+Subproject commit add387854ea73d83770a62282089dea666fa266f

gpt4all-backend/llamamodel.cpp

@@ -145,9 +145,8 @@ static int llama_sample_top_p_top_k(
         float top_p,
         float min_p,
         float temp,
-        float repeat_penalty,
+        float repeat_penalty) {
-        int32_t pos) {
+    auto logits = llama_get_logits_ith(ctx, -1);
-    auto logits = llama_get_logits_ith(ctx, pos);
     auto n_vocab = llama_n_vocab(llama_get_model(ctx));
     // Populate initial list of all candidates
     std::vector<llama_token_data> candidates;
@@ -529,13 +528,21 @@ size_t LLamaModel::restoreState(const uint8_t *src)
     return llama_set_state_data(d_ptr->ctx, const_cast<uint8_t*>(src));
 }
 
-std::vector<LLModel::Token> LLamaModel::tokenize(PromptContext &ctx, const std::string &str, bool special) const
+std::vector<LLModel::Token> LLamaModel::tokenize(PromptContext &ctx, const std::string &str, bool special)
 {
-    const bool wantBOS = ctx.n_past == 0 && ctx.tokens.empty();
+    bool atStart = m_tokenize_last_token == -1;
-    const bool useBOS = wantBOS && shouldAddBOS();
+    bool insertSpace = atStart || (
+        llama_token_get_attr(d_ptr->model, m_tokenize_last_token)
+        & (LLAMA_TOKEN_ATTR_CONTROL | LLAMA_TOKEN_ATTR_USER_DEFINED | LLAMA_TOKEN_ATTR_UNKNOWN)
+    );
     std::vector<LLModel::Token> fres(str.length() + 4);
-    auto fres_len = llama_tokenize(d_ptr->model, str.c_str(), str.length(), fres.data(), fres.size(), useBOS, special);
+    int32_t fres_len = llama_tokenize_gpt4all(
+        d_ptr->model, str.c_str(), str.length(), fres.data(), fres.size(), /*add_special*/ atStart,
+        /*parse_special*/ special, /*insert_space*/ insertSpace
+    );
     fres.resize(fres_len);
+    if (fres_len)
+        m_tokenize_last_token = fres.back();
     return fres;
 }
 
@@ -561,7 +568,7 @@ LLModel::Token LLamaModel::sampleToken(PromptContext &promptCtx) const
     return llama_sample_top_p_top_k(d_ptr->ctx,
         promptCtx.tokens.data() + promptCtx.tokens.size() - n_prev_toks,
         n_prev_toks, promptCtx.top_k, promptCtx.top_p, promptCtx.min_p, promptCtx.temp,
-        promptCtx.repeat_penalty, promptCtx.n_last_batch_tokens - 1);
+        promptCtx.repeat_penalty);
 }
 
 bool LLamaModel::evalTokens(PromptContext &ctx, const std::vector<int32_t> &tokens) const
@@ -571,7 +578,6 @@ bool LLamaModel::evalTokens(PromptContext &ctx, const std::vector<int32_t> &toke
     llama_batch batch = llama_batch_init(tokens.size(), 0, 1);
 
     batch.n_tokens = tokens.size();
-    ctx.n_last_batch_tokens = tokens.size();
 
     for (int32_t i = 0; i < batch.n_tokens; i++) {
         batch.token   [i] = tokens[i];
@@ -601,10 +607,7 @@ const std::vector<LLModel::Token> &LLamaModel::endTokens() const
 
 bool LLamaModel::shouldAddBOS() const
 {
-    int add_bos = llama_add_bos_token(d_ptr->model);
+    return llama_add_bos_token(d_ptr->model);
-    if (add_bos != -1) { return add_bos; }
-    auto vocab_type = llama_vocab_type(d_ptr->model);
-    return vocab_type == LLAMA_VOCAB_TYPE_SPM || vocab_type == LLAMA_VOCAB_TYPE_WPM;
 }
 
 int32_t LLamaModel::maxContextLength(std::string const &modelPath) const
@@ -946,7 +949,7 @@ void LLamaModel::embedInternal(
     const llama_token bos_token = llama_token_bos(d_ptr->model);
     const llama_token eos_token = llama_token_eos(d_ptr->model);
 
-    bool useBOS = shouldAddBOS();
+    bool useBOS = llama_add_bos_token(d_ptr->model);
     bool useEOS = llama_vocab_type(d_ptr->model) == LLAMA_VOCAB_TYPE_WPM;
 
     // no EOS, optional BOS
@@ -954,13 +957,16 @@ void LLamaModel::embedInternal(
         if (!text.empty() && text[0] != ' ') {
             text = ' ' + text; // normalize for SPM - our fork of llama.cpp doesn't add a space prefix
         }
-        wantBOS &= useBOS;
 
         tokens.resize(text.length()+4);
-        int32_t n_tokens = llama_tokenize(d_ptr->model, text.c_str(), text.length(), tokens.data(), tokens.size(), wantBOS, false);
+        int32_t n_tokens = llama_tokenize_gpt4all(
+            d_ptr->model, text.c_str(), text.length(), tokens.data(), tokens.size(), /*add_special*/ wantBOS,
+            /*parse_special*/ false, /*insert_space*/ false
+        );
         if (n_tokens) {
             (void)eos_token;
-            assert((useEOS && wantBOS) == (eos_token != -1 && tokens[n_tokens - 1] == eos_token));
+            (void)useBOS;
+            assert((useEOS && wantBOS && useBOS) == (eos_token != -1 && tokens[n_tokens - 1] == eos_token));
             if (useEOS && wantBOS)
                 n_tokens--; // erase EOS/SEP
         }

gpt4all-backend/llamamodel_impl.h

@@ -53,7 +53,7 @@ private:
     bool m_supportsCompletion = false;
 
 protected:
-    std::vector<Token> tokenize(PromptContext &ctx, const std::string &str, bool special) const override;
+    std::vector<Token> tokenize(PromptContext &ctx, const std::string &str, bool special) override;
     std::string tokenToString(Token id) const override;
     Token sampleToken(PromptContext &ctx) const override;
     bool evalTokens(PromptContext &ctx, const std::vector<int32_t> &tokens) const override;

gpt4all-backend/llmodel.h

@@ -14,11 +14,12 @@
 #include <utility>
 #include <vector>
 
+class Dlhandle;
+
 using namespace std::string_literals;
 
 #define LLMODEL_MAX_PROMPT_BATCH 128
 
-class Dlhandle;
 class LLModel {
 public:
     using Token = int32_t;
@@ -134,7 +135,6 @@ public:
         float   repeat_penalty = 1.10f;
         int32_t repeat_last_n = 64;     // last n tokens to penalize
         float   contextErase = 0.75f;   // percent of context to erase if we exceed the context window
-        int32_t n_last_batch_tokens = 0;
     };
 
     using ProgressCallback = std::function<bool(float progress)>;
@@ -212,7 +212,7 @@ public:
 protected:
     // These are pure virtual because subclasses need to implement as the default implementation of
     // 'prompt' above calls these functions
-    virtual std::vector<Token> tokenize(PromptContext &ctx, const std::string &str, bool special = false) const = 0;
+    virtual std::vector<Token> tokenize(PromptContext &ctx, const std::string &str, bool special = false) = 0;
     virtual std::string tokenToString(Token id) const = 0;
     virtual Token sampleToken(PromptContext &ctx) const = 0;
     virtual bool evalTokens(PromptContext &ctx, const std::vector<int32_t> &tokens) const = 0;
@@ -256,7 +256,8 @@ protected:
                           std::function<bool(bool)> recalculateCallback,
                           PromptContext &promptCtx);
 
-private:
+    Token m_tokenize_last_token = -1; // not serialized
+
     friend class LLMImplementation;
 };
 

gpt4all-backend/llmodel_c.cpp

@@ -117,9 +117,6 @@ void llmodel_prompt(llmodel_model model, const char *prompt,
         return response_callback(token_id, response.c_str());
     };
 
-    if (size_t(ctx->n_past) < wrapper->promptContext.tokens.size())
-        wrapper->promptContext.tokens.resize(ctx->n_past);
-
     // Copy the C prompt context
     wrapper->promptContext.n_past = ctx->n_past;
     wrapper->promptContext.n_ctx = ctx->n_ctx;

gpt4all-backend/llmodel_c.h

@@ -30,8 +30,6 @@ typedef void *llmodel_model;
  * behavior.
  */
 struct llmodel_prompt_context {
-    float *logits;          // logits of current context
-    size_t logits_size;     // the size of the raw logits vector
     int32_t *tokens;        // current tokens in the context window
     size_t tokens_size;     // the size of the raw tokens vector
     int32_t n_past;         // number of tokens in past conversation

gpt4all-backend/llmodel_shared.cpp

@@ -8,12 +8,16 @@
 #include <iostream>
 #include <optional>
 #include <regex>
+#include <sstream>
 #include <stdexcept>
 #include <string>
 #include <unordered_set>
 #include <vector>
 
 // TODO(cebtenzzre): replace this with llama_kv_cache_seq_shift for llamamodel (GPT-J needs this as-is)
+// FIXME(jared): if recalculate returns false, we leave n_past<tokens.size() and do not tell the caller to stop
+// FIXME(jared): if we get here during chat name or follow-up generation, bad things will happen when we try to restore
+// the old prompt context afterwards
 void LLModel::recalculateContext(PromptContext &promptCtx, std::function<bool(bool)> recalculate)
 {
     int n_keep = shouldAddBOS();
@@ -88,6 +92,16 @@ void LLModel::prompt(const std::string &prompt,
         return;
     }
 
+    // make sure token cache matches decode offset
+    if (promptCtx.tokens.size() < promptCtx.n_past) {
+        std::ostringstream ss;
+        ss << "expected n_past to be at most " << promptCtx.tokens.size() << ", got " << promptCtx.n_past;
+        throw std::out_of_range(ss.str());
+    }
+    if (promptCtx.n_past < promptCtx.tokens.size())
+        promptCtx.tokens.resize(promptCtx.n_past);
+    m_tokenize_last_token = promptCtx.tokens.empty() ? -1 : promptCtx.tokens.back(); // not serialized
+
     // parse the prompt template
     std::vector<std::smatch> placeholders;
     {
@@ -201,8 +215,6 @@ bool LLModel::decodePrompt(std::function<bool(int32_t)> promptCallback,
 
         size_t tokens = batch_end - i;
         for (size_t t = 0; t < tokens; ++t) {
-            if (int32_t(promptCtx.tokens.size()) == promptCtx.n_ctx)
-                promptCtx.tokens.erase(promptCtx.tokens.begin());
             promptCtx.tokens.push_back(batch.at(t));
             promptCtx.n_past += 1;
             if (!promptCallback(batch.at(t)))
@@ -270,8 +282,6 @@ void LLModel::generateResponse(std::function<bool(int32_t, const std::string&)>
 
         // Empty the cache
         for (auto t : cachedTokens) {
-            if (int32_t(promptCtx.tokens.size()) == promptCtx.n_ctx)
-                promptCtx.tokens.erase(promptCtx.tokens.begin());
             promptCtx.tokens.push_back(t);
             promptCtx.n_past += 1;
             //TODO: Conversion to std::string can be avoided here...

gpt4all-bindings/python/gpt4all/_pyllmodel.py

@@ -73,8 +73,6 @@ llmodel = load_llmodel_library()
 
 class LLModelPromptContext(ctypes.Structure):
     _fields_ = [
-        ("logits", ctypes.POINTER(ctypes.c_float)),
-        ("logits_size", ctypes.c_size_t),
         ("tokens", ctypes.POINTER(ctypes.c_int32)),
         ("tokens_size", ctypes.c_size_t),
         ("n_past", ctypes.c_int32),
@@ -351,7 +349,6 @@ class LLModel:
     ):
         if self.context is None:
             context = LLModelPromptContext(
-                logits_size=0,
                 tokens_size=0,
                 n_past=0,
                 n_ctx=0,

gpt4all-chat/chatapi.h

@@ -97,7 +97,7 @@ protected:
     // them as they are only called from the default implementation of 'prompt' which we override and
     // completely replace
 
-    std::vector<Token> tokenize(PromptContext &ctx, const std::string &str, bool special) const override {
+    std::vector<Token> tokenize(PromptContext &ctx, const std::string &str, bool special) override {
         (void)ctx;
         (void)str;
         (void)special;

gpt4all-chat/chatllm.cpp

@@ -611,6 +611,7 @@ std::string trim_whitespace(const std::string& input)
     return std::string(first_non_whitespace, last_non_whitespace);
 }
 
+// FIXME(jared): we don't actually have to re-decode the prompt to generate a new response
 void ChatLLM::regenerateResponse()
 {
     // ChatGPT uses a different semantic meaning for n_past than local models. For ChatGPT, the meaning