main : add token healing

mare5x · mare5x · commit cec3120cfde3 · 2024-07-01T14:50:22.000+02:00
diff --git a/common/common.cpp b/common/common.cpp
@@ -1058,6 +1058,25 @@ bool gpt_params_find_arg(int argc, char ** argv, const std::string & arg, gpt_pa
         sparams.grammar = json_schema_to_grammar(json::parse(argv[i]));
         return true;
     }
+    if (arg == "-th" || arg == "--token-healing") {
+        CHECK_ARG
+        sparams.token_healing_enabled = true;
+        auto & th_type = sparams.token_healing_type;
+        auto & th_n_rollback = sparams.token_healing_n_rollback;
+        std::string value(argv[i]);
+        /**/ if (value    == "0" ) { sparams.token_healing_enabled = false; }
+        else if (value    == "1" ) { th_type = llama_token_healing_type::ROLLBACK_LAST; }
+        else if (value    == "d1") { th_type = llama_token_healing_type::DYNAMIC_ONCE; }
+        else if (value    == "d" ) { th_type = llama_token_healing_type::DYNAMIC_MULTI; }
+        else if (value[0] == 'r' ) {
+            th_type = llama_token_healing_type::ROLLBACK_MULTI;
+            th_n_rollback = std::stoi(value.substr(1));
+            if (th_n_rollback <= 0) {
+                sparams.token_healing_enabled = false;
+            }
+        } else { invalid_param = true; }
+        return true;
+    }
     if (arg == "--override-kv") {
         CHECK_ARG
         if (!string_parse_kv_override(argv[i], params.kv_overrides)) {
@@ -1455,6 +1474,10 @@ void gpt_params_print_usage(int /*argc*/, char ** argv, const gpt_params & param
                                                                         "set custom jinja chat template (default: template taken from model's metadata)\n"
                                                                         "only commonly used templates are accepted:\n"
                                                                         "https://github.com/ggerganov/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template" });
+
+    options.push_back({ "main",        "-th,   --token-healing {0,1,d1,d,r{N}}",
+                                                                        "Token healing type. (default: 0, disabled)\n"
+                                                                        "1: replace one token, d1: replace longest suffix with one token, d: replace longest suffix, r{N}: roll back N tokens" });
     options.push_back({ "grammar" });
     options.push_back({ "*",           "       --grammar GRAMMAR",      "BNF-like grammar to constrain generations (see samples in grammars/ dir) (default: '%s')", sparams.grammar.c_str() });
     options.push_back({ "*",           "       --grammar-file FNAME",   "file to read grammar from" });
diff --git a/common/sampling.cpp b/common/sampling.cpp
@@ -2,6 +2,112 @@
 #include "sampling.h"
 #include <random>
 
+//
+// Token healing (internal)
+//
+
+static bool startswith(const std::string & str, const std::string & prefix) {
+    return str.rfind(prefix, 0) != std::string::npos;
+}
+
+static bool token_healing_prefix_exists(const llama_context * ctx_main, const std::string & prefix) {
+    const int32_t n_vocab = llama_n_vocab(llama_get_model(ctx_main));
+    for (llama_token token_id = 0; token_id < n_vocab; ++token_id) {
+        if (startswith(llama_token_to_piece(ctx_main, token_id), prefix)) {
+            return true;
+        }
+    }
+    return false;
+}
+
+static std::vector<llama_token> token_healing_find_prefix(
+                                const llama_context * ctx_main,
+                                const std::string & prefix,
+                                const bool include_partial_prefix) {
+    // Example: prefix=" world" -> " world", " worldwide", ...
+    // If `include_partial_prefix`, include also: " w", " wo", ...
+    std::vector<llama_token> candidates;
+    const int32_t n_vocab = llama_n_vocab(llama_get_model(ctx_main));
+    for (llama_token token_id = 0; token_id < n_vocab; ++token_id) {
+        std::string token = llama_token_to_piece(ctx_main, token_id);
+        if (startswith(token, prefix) ||
+                (include_partial_prefix && startswith(prefix, token))) {
+            candidates.push_back(token_id);
+        }
+    }
+    return candidates;
+}
+
+//
+// Token healing (external)
+//
+
+std::string llama_token_healing_rollback(
+            const llama_context * ctx_main,
+            llama_token_healing_type th_type,
+            std::vector<llama_token> & tokens,
+            int max_to_remove,
+            int * n_removed) {
+    // NB. To avoid returning empty `tokens`, at least 1 token will remain in `tokens` after rolling back.
+    //     It is the caller's responsibility to add BOS to the start of the prompt if they want to roll back the whole prompt.
+    if (n_removed != nullptr) {
+        *n_removed = 0;
+    }
+    if (tokens.size() <= 1) {
+        return "";
+    }
+    const llama_model * model = llama_get_model(ctx_main);
+    const bool is_dynamic = th_type == llama_token_healing_type::DYNAMIC_ONCE || th_type == llama_token_healing_type::DYNAMIC_MULTI;
+    const int n_ctx = tokens.size();
+    max_to_remove = th_type == llama_token_healing_type::ROLLBACK_LAST ? 1 : max_to_remove;
+    max_to_remove = max_to_remove < 0 ? n_ctx - 1 : std::min(max_to_remove, n_ctx - 1);  // 1 token must remain
+    int removed = 0;
+    std::string prefix;
+    // Roll back tokens a fixed amount or until there does not exist a token that can cover the prompt
+    // and stop early if a special token is encountered.
+    // NB. This doesn't handle cases where a long token is split many times,
+    //     e.g. if "abc" is tokenized into ["a", "b", "c"] but "bc" is not a token (hypothetically),
+    //     then "abc" will not be returned even if "abcd" exists in the vocab.
+    while (removed < max_to_remove) {
+        const llama_token next_token_id = tokens[n_ctx - removed - 1];
+        if (llama_token_is_control(model, next_token_id) || llama_token_is_eog(model, next_token_id)) {
+            break;  // Don't roll back e.g. <|endoftext|>
+        }
+        std::string new_prefix = llama_token_to_piece(ctx_main, next_token_id) + prefix;
+        if (is_dynamic && !token_healing_prefix_exists(ctx_main, new_prefix)) {
+            break;
+        }
+        removed += 1;
+        prefix = new_prefix;
+    }
+    if (removed == 0) {  // E.g. if the last token is a special token
+        return "";
+    }
+    // If constrained decoding would give back the original prompt, there is no need to modify the context
+    const bool is_multi_step = th_type == llama_token_healing_type::ROLLBACK_MULTI ||
+                               th_type == llama_token_healing_type::DYNAMIC_MULTI;
+    const std::vector<llama_token> candidates = token_healing_find_prefix(ctx_main, prefix, is_multi_step);
+    LOG("token_healing: prefix = '%s' (%d tokens)\n", prefix.c_str(), removed);
+    if (removed == 1 && candidates.size() == 1) {
+        LOG("token_healing: nothing to heal\n");
+        return "";
+    }
+    // Finalize outputs
+    if (n_removed != nullptr) {
+        *n_removed = removed;
+    }
+    tokens.resize(n_ctx - removed);
+    return prefix;
+}
+
+void llama_token_healing_set_prefix(llama_sampling_context * ctx_sampling, const std::string & prefix) {
+    ctx_sampling->token_healing_prefix = prefix;
+}
+
+//
+// Sampling
+//
+
 struct llama_sampling_context * llama_sampling_init(const struct llama_sampling_params & params) {
     struct llama_sampling_context * result = new llama_sampling_context();
 
@@ -72,6 +178,8 @@ void llama_sampling_reset(llama_sampling_context * ctx) {
         ctx->grammar = grammar;
     }
 
+    ctx->token_healing_prefix.clear();
+
     std::fill(ctx->prev.begin(), ctx->prev.end(), 0);
     ctx->cur.clear();
     ctx->n_valid = 0;
@@ -130,7 +238,7 @@ std::string llama_sampling_print(const llama_sampling_params & params) {
 }
 
 std::string llama_sampling_order_print(const llama_sampling_params & params) {
-    std::string result = "CFG -> Penalties ";
+    std::string result = "(Token healing) -> CFG -> Penalties ";
     if (params.mirostat == 0) {
         for (auto sampler_type : params.samplers_sequence) {
             const auto sampler_type_name = llama_sampling_type_to_str(sampler_type);
@@ -392,8 +500,27 @@ static llama_token_data_array llama_sampling_prepare_impl(
 
     cur.clear();
 
-    for (llama_token token_id = 0; token_id < n_vocab; token_id++) {
-        cur.emplace_back(llama_token_data{token_id, logits[token_id], 0.0f});
+    // Constrain tokens based on the remaining token healing prefix (if any)
+    const auto & th_type = params.token_healing_type;
+    const auto & th_prefix = ctx_sampling->token_healing_prefix;
+    if (params.token_healing_enabled && !th_prefix.empty()) {
+        const bool is_multi_step = th_type == llama_token_healing_type::ROLLBACK_MULTI ||
+                                   th_type == llama_token_healing_type::DYNAMIC_MULTI;
+        std::vector<llama_token> th_candidates = token_healing_find_prefix(ctx_main, th_prefix, is_multi_step);
+
+        LOG("token_healing: prefix = '%s'\n", th_prefix.c_str());
+        for (const llama_token token_id : th_candidates) {
+            LOG(" [%6d] '%s'\n", token_id, llama_token_to_piece(ctx_main, token_id).c_str());
+        }
+
+        // N.B. We could also set token constraints by setting rejected tokens' logits to -inf
+        for (const llama_token token_id : th_candidates) {
+            cur.emplace_back(llama_token_data{ token_id, logits[token_id], 0.0f });
+        }
+    } else {
+        for (llama_token token_id = 0; token_id < n_vocab; token_id++) {
+            cur.emplace_back(llama_token_data{ token_id, logits[token_id], 0.0f });
+        }
     }
 
     llama_token_data_array cur_p = { cur.data(), cur.size(), false };
@@ -456,4 +583,19 @@ void llama_sampling_accept(
     if (ctx_sampling->grammar != NULL && apply_grammar) {
         llama_grammar_accept_token(ctx_main, ctx_sampling->grammar, id);
     }
+
+    if (ctx_sampling->params.token_healing_enabled && apply_grammar) {
+        std::string & th_prefix = ctx_sampling->token_healing_prefix;
+        if (!th_prefix.empty()) {
+            const std::string new_token_piece = llama_token_to_piece(ctx_main, id);
+            if (new_token_piece.size() < th_prefix.size()) {
+                // Shift prefix constraint (for multi step token healing)
+                th_prefix = th_prefix.substr(new_token_piece.size());
+            } else {
+                // Prefix has been generated => no more constrained generation
+                th_prefix.clear();
+                LOG("token_healing: done\n");
+            }
+        }
+    }
 }
diff --git a/common/sampling.h b/common/sampling.h
@@ -19,6 +19,13 @@ enum class llama_sampler_type : char {
     TEMPERATURE = 't'
 };
 
+enum class llama_token_healing_type : uint8_t {
+    ROLLBACK_LAST,   // roll back last token with a single constrained decoding step
+    ROLLBACK_MULTI,  // roll back a fixed amount of tokens, multiple constrained decoding steps
+    DYNAMIC_ONCE,    // dynamic roll back, single constrained decoding step
+    DYNAMIC_MULTI    // dynamic roll back, multiple constrained decoding steps
+};
+
 // sampling parameters
 typedef struct llama_sampling_params {
     int32_t     n_prev                = 64;                 // number of previous tokens to remember
@@ -62,6 +69,10 @@ typedef struct llama_sampling_params {
 
     std::vector<llama_token> penalty_prompt_tokens;
     bool                     use_penalty_prompt_tokens = false;
+
+    llama_token_healing_type token_healing_type       = llama_token_healing_type::ROLLBACK_LAST;
+    bool                     token_healing_enabled    = false;
+    int                      token_healing_n_rollback = -1;  // number of tokens to roll back
 } llama_sampling_params;
 
 // general sampler context
@@ -78,6 +89,8 @@ struct llama_sampling_context {
     // internal
     grammar_parser::parse_state parsed_grammar;
 
+    std::string token_healing_prefix;  // remaining prefix to constrain sampling
+
     // TODO: replace with ring-buffer
     std::vector<llama_token>      prev;
     std::vector<llama_token_data> cur;
@@ -158,3 +171,18 @@ void llama_sampling_accept(
         struct llama_context * ctx_main,
         llama_token id,
         bool apply_grammar);
+
+//
+// Token healing
+//
+
+// Roll back `tokens` for constrained generation according to the token healing
+// strategy. Returns the prefix for constrained generation.
+std::string llama_token_healing_rollback(
+            const llama_context * ctx_main,
+            llama_token_healing_type th_type,
+            std::vector<llama_token> & tokens,
+            int max_to_remove = -1,
+            int * n_removed = nullptr);
+
+void llama_token_healing_set_prefix(llama_sampling_context * ctx_sampling, const std::string & prefix);
diff --git a/examples/main/README.md b/examples/main/README.md
@@ -246,6 +246,19 @@ A more practical use case might be to prevent the generation of `\code{begin}` a
 
 Example usage: `--logit-bias 29905-inf`
 
+### Token healing
+
+-   `-th {0,1,d1,d,r{N}}, --token-healing {0,1,d1,d,r{N}}`: Set the token healing strategy (default: 0, 0 = disabled).
+
+Token healing (a.k.a. token alignment) alleviates tokenization artifacts for text completion.
+
+- `-th 1`: Roll back the last token and constrain the bytes of the next token to start with the chopped off last token [0, 2].
+- `-th d1`: Roll back multiple tokens until there doesn't exist a token which can cover the prompt's suffix and do a single constrained decoding step [2].
+- `-th d`: Like `d1` but allow multiple decoding steps until the removed suffix is generated.
+- `-th r{N}`: Like `d` but roll back `N` tokens, where `-th r3` is recommended [1].
+
+Sources: [0](https://github.com/guidance-ai/guidance/blob/main/notebooks/art_of_prompt_design/prompt_boundaries_and_token_healing.ipynb), [1](https://arxiv.org/abs/2403.08688), [2](https://arxiv.org/abs/2402.01035).
+
 ### RNG Seed
 
 -   `-s SEED, --seed SEED`: Set the random number generator (RNG) seed (default: -1, -1 = random seed).
diff --git a/examples/main/main.cpp b/examples/main/main.cpp
@@ -276,6 +276,17 @@ int main(int argc, char ** argv) {
         LOG("tokens: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd_inp).c_str());
     }
 
+    if (sparams.token_healing_enabled && (params.conversation || !params.input_suffix.empty())) {
+        sparams.token_healing_enabled = false;
+        LOG("token_healing: disabled due to custom suffix/conversation mode");
+    }
+    std::string token_healing_prefix;
+    int token_healing_n_removed = 0;
+    if (!params.interactive_first && sparams.token_healing_enabled) {
+        token_healing_prefix = llama_token_healing_rollback(ctx, sparams.token_healing_type, embd_inp,
+                                                                 sparams.token_healing_n_rollback, &token_healing_n_removed);
+    }
+
     // Should not run without any tokens
     if (embd_inp.empty()) {
         embd_inp.push_back(llama_token_bos(model));
@@ -295,7 +306,7 @@ int main(int argc, char ** argv) {
         std::vector<llama_token> original_inp = ::llama_tokenize(ctx, params.prompt, true, true);
         LOG("original_inp tokenized: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, original_inp).c_str());
 
-        original_prompt_len = original_inp.size();
+        original_prompt_len = original_inp.size() - token_healing_n_removed;
         guidance_offset = (int)guidance_inp.size() - original_prompt_len;
         LOG("original_prompt_len: %s", log_tostr(original_prompt_len));
         LOG("guidance_offset:     %s", log_tostr(guidance_offset));
@@ -490,6 +501,7 @@ int main(int argc, char ** argv) {
     int n_consumed         = 0;
     int n_session_consumed = 0;
     int n_past_guidance    = 0;
+    int n_bytes_to_skip    = 0;  // to skip printing when generating token healing prefix
 
     std::vector<int>   input_tokens;  g_input_tokens  = &input_tokens;
     std::vector<int>   output_tokens; g_output_tokens = &output_tokens;
@@ -516,6 +528,7 @@ int main(int argc, char ** argv) {
         fprintf(stderr, "%s: failed to initialize sampling subsystem\n", __func__);
         exit(1);
     }
+    llama_token_healing_set_prefix(ctx_sampling, token_healing_prefix);
 
     while ((n_remain != 0 && !is_antiprompt) || params.interactive) {
         // predict
@@ -732,7 +745,15 @@ int main(int argc, char ** argv) {
                 const std::string token_str = llama_token_to_piece(ctx, id, params.special);
 
                 // Console/Stream Output
-                fprintf(stdout, "%s", token_str.c_str());
+                // Suppress printing while generating token healing prefix
+                if (n_bytes_to_skip > 0 && n_bytes_to_skip < (int)token_str.size()) {
+                    fprintf(stdout, "%s", token_str.substr(n_bytes_to_skip).c_str());
+                    n_bytes_to_skip = 0;
+                } else if (n_bytes_to_skip > 0) {
+                    n_bytes_to_skip -= token_str.size();
+                } else {
+                    fprintf(stdout, "%s", token_str.c_str());
+                }
 
                 // Record Displayed Tokens To Log
                 // Note: Generated tokens are created one by one hence this check
@@ -824,6 +845,7 @@ int main(int argc, char ** argv) {
                 assistant_ss << llama_token_to_piece(ctx, id, false);
             }
 
+            token_healing_n_removed = 0;
             if (n_past > 0 && is_interacting) {
                 LOG("waiting for user input\n");
 
@@ -889,6 +911,17 @@ int main(int argc, char ** argv) {
                     embd_inp.insert(embd_inp.end(), line_inp.begin(), line_inp.end());
                     embd_inp.insert(embd_inp.end(), line_sfx.begin(), line_sfx.end());
 
+                    if (sparams.token_healing_enabled) {
+                        // Limit token healing rollback to new tokens only (otherwise would need to shift everything)
+                        const int n_new_tokens = embd_inp.size() - original_size;
+                        const int max_to_remove = sparams.token_healing_n_rollback < 0
+                                                   ? n_new_tokens
+                                                   : std::min(sparams.token_healing_n_rollback, n_new_tokens);
+                        token_healing_prefix = llama_token_healing_rollback(ctx, sparams.token_healing_type, embd_inp,
+                                                                            max_to_remove, &token_healing_n_removed);
+                        n_bytes_to_skip = token_healing_prefix.size();
+                    }
+
                     for (size_t i = original_size; i < embd_inp.size(); ++i) {
                         const llama_token token = embd_inp[i];
                         output_tokens.push_back(token);
@@ -898,7 +931,7 @@ int main(int argc, char ** argv) {
                     // reset assistant message
                     assistant_ss.str("");
 
-                    n_remain -= line_inp.size();
+                    n_remain -= line_inp.size() + token_healing_n_removed;
                     LOG("n_remain: %d\n", n_remain);
                 } else {
                     LOG("empty line, passing control back\n");
@@ -910,6 +943,10 @@ int main(int argc, char ** argv) {
             if (n_past > 0) {
                 if (is_interacting) {
                     llama_sampling_reset(ctx_sampling);
+                    if (token_healing_n_removed > 0) {
+                        // Set new prefix after an interaction
+                        llama_token_healing_set_prefix(ctx_sampling, token_healing_prefix);
+                    }
                 }
                 is_interacting = false;
             }
