llama : dynamic temperature sampling (#4972)

* implemented dynamic temperature sampling from koboldcpp

* removed trailing whitespace

* removed unused temp parameter in llama_sample_entropy

* exposed exponent_val in dynamic temp sampler

* added debug check for printf statements

* use nullptr in llama_sample_softmax call during llama_sample_entropy

this avoids counting the time taken stats twice

Co-authored-by: Georgi Gerganov <redacted>
* return earlier if there is only 1 candiate (i.e. max_entropy == 0)

* reformat 't' case in llama_sample_queue

Co-authored-by: Jared Van Bortel <redacted>
* check for one or zero candidates case in llama_sample_entropy

---------

Co-authored-by: Georgi Gerganov <redacted>
Co-authored-by: Jared Van Bortel <redacted>

diff --git a/common/sampling.cpp b/common/sampling.cpp
index dd1ffeb1b808352e22a7d367ee3b92dd60fc2458..efd7eab6e50b829bae1a34cfe1293c7e6574fe0b 100644 (file)
--- a/common/sampling.cpp
+++ b/common/sampling.cpp
@@ -129,6 +129,8 @@ static void sampler_queue(
     const int n_vocab = llama_n_vocab(llama_get_model(ctx_main));
 
     const float         temp              = params.temp;
+    const float         dynatemp_range    = params.dynatemp_range;
+    const float         dynatemp_exponent = params.dynatemp_exponent;
     const int32_t       top_k             = params.top_k <= 0 ? n_vocab : params.top_k;
     const float         top_p             = params.top_p;
     const float         min_p             = params.min_p;
@@ -143,7 +145,15 @@ static void sampler_queue(
             case 'y': llama_sample_typical  (ctx_main, &cur_p, typical_p, min_keep); break;
             case 'p': llama_sample_top_p    (ctx_main, &cur_p, top_p,     min_keep); break;
             case 'm': llama_sample_min_p    (ctx_main, &cur_p, min_p,     min_keep); break;
-            case 't': llama_sample_temp     (ctx_main, &cur_p, temp); break;
+            case 't':
+                if (dynatemp_range > 0) {
+                    float dynatemp_min = std::max(0.0f, temp - dynatemp_range);
+                    float dynatemp_max = std::max(0.0f, temp + dynatemp_range);
+                    llama_sample_entropy(ctx_main, &cur_p, dynatemp_min, dynatemp_max, dynatemp_exponent);
+                } else {
+                    llama_sample_temp(ctx_main, &cur_p, temp);
+                }
+                break;
             default : break;
         }
     }

diff --git a/common/sampling.h b/common/sampling.h
index 2ee1803761aa24be03b080abf9f6eb3eaa9673a8..88899c094866eff2dfb199b7daa3360475758b61 100644 (file)
--- a/common/sampling.h
+++ b/common/sampling.h
@@ -18,6 +18,8 @@ typedef struct llama_sampling_params {
     float       tfs_z                 = 1.00f;    // 1.0 = disabled
     float       typical_p             = 1.00f;    // 1.0 = disabled
     float       temp                  = 0.80f;    // <= 0.0 to sample greedily, 0.0 to not output probabilities
+    float       dynatemp_range        = 0.00f;    // 0.0 = disabled
+    float       dynatemp_exponent     = 1.00f;    // controls how entropy maps to temperature in dynamic temperature sampler
     int32_t     penalty_last_n        = 64;       // last n tokens to penalize (0 = disable penalty, -1 = context size)
     float       penalty_repeat        = 1.10f;    // 1.0 = disabled
     float       penalty_freq          = 0.00f;    // 0.0 = disabled

diff --git a/llama.cpp b/llama.cpp
index 6a7506e85f8ec6013cbd0f1042eb6240655a4ea6..823d42d7fb50ab087192057b4ffb93a077abc474 100644 (file)
--- a/llama.cpp
+++ b/llama.cpp
@@ -8151,6 +8151,73 @@ void llama_sample_typical(struct llama_context * ctx, llama_token_data_array * c
     }
 }
 
+void llama_sample_entropy(struct llama_context * ctx, llama_token_data_array * candidates_p, float min_temp, float max_temp, float exponent_val) {
+    const int64_t t_start_sample_us = ggml_time_us();
+
+    // no need to do anything if there is only one (or zero) candidates
+    if(candidates_p->size <= 1) {
+        return;
+    }
+
+    // Calculate maximum possible entropy
+    float max_entropy = -logf(1.0f / candidates_p->size);
+
+    llama_sample_softmax(nullptr, candidates_p);
+
+    // Calculate entropy of the softmax probabilities
+    float entropy = 0.0f;
+    for (size_t i = 0; i < candidates_p->size; ++i) {
+        float prob = candidates_p->data[i].p;
+        if (prob > 0.0f) { // Ensure no log(0)
+            entropy -= prob * logf(prob);
+        }
+    }
+
+    // Normalize the entropy (max_entropy cannot be 0 here because we checked candidates_p->size != 1 above)
+    float normalized_entropy = entropy / max_entropy;
+
+    // Map the normalized entropy to the desired temperature range using the power function
+    float dyn_temp = min_temp + (max_temp - min_temp) * powf(normalized_entropy, exponent_val);
+
+#ifdef DEBUG
+    LLAMA_LOG_INFO("Your text maxtemp value is: %f\n", max_temp);
+    LLAMA_LOG_INFO("Entropy: %f\n", entropy);
+    LLAMA_LOG_INFO("Max Possible Entropy: %f\n", max_entropy);
+    LLAMA_LOG_INFO("Normalized Entropy: %f\n", normalized_entropy);
+    LLAMA_LOG_INFO("Exponent: %f\n", exponent_val);
+    LLAMA_LOG_INFO("Dynamic Temperature (dyn_temp): %f\n", dyn_temp);
+#endif
+
+    // Apply the dynamically calculated temperature scaling
+    for (size_t i = 0; i < candidates_p->size; ++i) {
+        candidates_p->data[i].logit /= dyn_temp;
+    }
+
+    // Re-compute softmax probabilities after scaling logits with dynamic temperature
+    double max_l_double = candidates_p->data[0].logit;
+    double cum_sum_double = 0.0;
+    for (size_t i = 0; i < candidates_p->size; ++i) {
+        double p = exp(candidates_p->data[i].logit - max_l_double);
+        candidates_p->data[i].p = p; // Store the scaled probability
+        cum_sum_double += p;
+    }
+    for (size_t i = 0; i < candidates_p->size; ++i) {
+        candidates_p->data[i].p /= cum_sum_double; // Re-normalize the probabilities
+    }
+
+#ifdef DEBUG
+    // Print the updated top 25 probabilities after temperature scaling
+    LLAMA_LOG_INFO("\nUpdated Top 25 Probabilities After Dynamic Temperature Scaling (in percentages):\n");
+    for (size_t i = 0; i < 25 && i < candidates_p->size; ++i) {
+        LLAMA_LOG_INFO("Token %zu: %f%%\n", i + 1, candidates_p->data[i].p * 100.0f);
+    }
+#endif
+
+    if (ctx) {
+        ctx->t_sample_us += ggml_time_us() - t_start_sample_us;
+    }
+}
+
 void llama_sample_temp(struct llama_context * ctx, llama_token_data_array * candidates_p, float temp) {
     const int64_t t_start_sample_us = ggml_time_us();
 

diff --git a/llama.h b/llama.h
index bb6054557593228ec66e9531141e2a796c3f43b9..7b3634aa68505cfeaf77790aef66af74ae67b394 100644 (file)
--- a/llama.h
+++ b/llama.h
@@ -775,6 +775,14 @@ extern "C" {
                            float   p,
                           size_t   min_keep);
 
+    /// @details Dynamic temperature implementation described in the paper https://arxiv.org/abs/2309.02772.
+    LLAMA_API void llama_sample_entropy(
+            struct llama_context * ctx,
+          llama_token_data_array * candidates_p,
+                           float   min_temp,
+                           float   max_temp,
+                           float   exponent_val);
+
     LLAMA_API void llama_sample_temp(
             struct llama_context * ctx,
           llama_token_data_array * candidates,