lookup: fibonacci hashing, fix crashes (#8548)

diff --git a/common/ngram-cache.h b/common/ngram-cache.h
index e4fa4cbd12f11eea481a3d22dd975f64f5c7a6f2..ab4c9b37665460281683a6bb09bcddd444eb3634 100644 (file)
--- a/common/ngram-cache.h
+++ b/common/ngram-cache.h
@@ -37,11 +37,18 @@ struct llama_ngram {
     }
 };
 
+struct llama_token_hash_function {
+    size_t operator()(const llama_token token) const {
+        // see https://probablydance.com/2018/06/16/fibonacci-hashing-the-optimization-that-the-world-forgot-or-a-better-alternative-to-integer-modulo/
+        return token * 11400714819323198485llu;
+    }
+};
+
 struct llama_ngram_hash_function {
     size_t operator()(const llama_ngram & ngram) const {
-        size_t hash = 0;
-        for (int i = 0; i < LLAMA_NGRAM_MAX; ++i) {
-            hash ^= std::hash<llama_token>{}(ngram.tokens[i]);
+        size_t hash = llama_token_hash_function{}(ngram.tokens[0]);
+        for (int i = 1; i < LLAMA_NGRAM_MAX; ++i) {
+            hash ^= llama_token_hash_function{}(ngram.tokens[i]);
         }
         return hash;
     }

diff --git a/examples/lookup/lookup-stats.cpp b/examples/lookup/lookup-stats.cpp
index 0b171c87273d19d58e8197040b6443ef71f66ed2..2fe67100e6c03b7be94be899cfef868c6bbd60c2 100644 (file)
--- a/examples/lookup/lookup-stats.cpp
+++ b/examples/lookup/lookup-stats.cpp
@@ -31,7 +31,6 @@ int main(int argc, char ** argv){
 
     // load the model
     std::tie(model, ctx) = llama_init_from_gpt_params(params);
-    GGML_ASSERT(llama_n_vocab(model) < (1 << 16));
 
     // tokenize the prompt
     std::vector<llama_token> inp;
@@ -65,7 +64,7 @@ int main(int argc, char ** argv){
     }
 
     const int n_input = inp.size();
-    const int n_ctx = params.n_ctx;
+    const int n_ctx = llama_n_ctx(ctx);
 
     int n_drafted = 0;
     int n_accept  = 0;

diff --git a/examples/lookup/lookup.cpp b/examples/lookup/lookup.cpp
index 80ecd925d5962766ba80d1a1c4679ed43fa8c79e..bb571bac4d7781ff87428bc8c52abd673c479b04 100644 (file)
--- a/examples/lookup/lookup.cpp
+++ b/examples/lookup/lookup.cpp
@@ -39,7 +39,6 @@ int main(int argc, char ** argv){
 
     // load the model
     std::tie(model, ctx) = llama_init_from_gpt_params(params);
-    GGML_ASSERT(llama_n_vocab(model) < (1 << 16));
 
     // tokenize the prompt
     std::vector<llama_token> inp;