Fix adaptive p sampler bug with string ban (#1287)

* adaptive p: upadte internal state only if not rewinding

* adaptive p: conditional update for speculative decoding

* adaptive p: refactor to rewind instead of update

* adaptive p fix: better comments

* fix rewind check

* add record to handle multi-token rewind

* better comment

common/sampling.cpp

@@ -142,6 +142,16 @@ void common_sampler_reset(common_sampler * ctx) {
     llama_sampler_dry_reset(ctx->smpl);
 }
 
+void common_sampler_review(common_sampler * ctx) {
+    const bool record = ctx->record_samplers;
+    const bool rewind = ctx->rewind_samplers;
+
+    llama_review_adaptive_p(ctx->adapt_p_ctx, record, rewind);
+
+    ctx->record_samplers = false;
+    ctx->rewind_samplers = false;
+}
+
 void llama_sampling_set_rng_seed(struct common_sampler * ctx, uint32_t seed) {
     if (seed == LLAMA_DEFAULT_SEED) {
         seed = std::random_device{}();

common/sampling.h

@@ -134,6 +134,9 @@ struct common_sampler {
     llama_token_data_array cur_p; // current candidates
 
     std::mt19937 rng;
+
+    bool record_samplers = false;   // record current state for stateful samplers
+    bool rewind_samplers = false;   // rewind stateful samplers to last recorded
 };
 
 
@@ -148,6 +151,11 @@ void common_sampler_free(struct common_sampler * ctx);
 // - reset grammar
 void common_sampler_reset(common_sampler * ctx);
 
+// Review stateful samplers
+// | record current state for rewinding
+// | rewind to last recorded state
+void common_sampler_review(common_sampler * ctx);
+
 // Set the sampler seed
 void llama_sampling_set_rng_seed(struct common_sampler * ctx, uint32_t seed);