server : renamed spec checkpoints option

Author: srogmann

common/arg.cpp

@@ -3465,13 +3465,17 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         }
     ).set_examples({LLAMA_EXAMPLE_SERVER}));
     add_opt(common_arg(
-        {"--spec-ckpt-num-tries"}, "N",
-        string_format("number of tries for speculative decoding with recurrent memory (default: %d)", params.speculative.ckpt_num_tries),
-        [](common_params & params, int value) {
-            if (value < 0 || value > 10) {
-                throw std::invalid_argument("number of tries must be between 0 and 10 inclusive");
+        {"--spec-use-checkpoints"}, "[on|off|auto]",
+        string_format("use checkpoints to rewind token history in recurrent models ('on', 'off', or 'auto', default: %s)",
+                        params.speculative.use_checkpoints ? "on" : "off"),
+        [](common_params & params, const std::string & value) {
+            if (is_truthy(value) || is_autoy(value)) {
+                params.speculative.use_checkpoints = true;
+            } else if (is_falsey(value)) {
+                params.speculative.use_checkpoints = false;
+            } else {
+                throw std::invalid_argument("invalid value for --spec-use-checkpoints");
             }
-            params.speculative.ckpt_num_tries = value;
         }
     ).set_examples({LLAMA_EXAMPLE_SERVER}));
     add_opt(common_arg(

common/common.h

@@ -270,7 +270,8 @@ struct common_params_speculative {
     uint16_t ngram_size_n     = 12; // ngram size for lookup
     uint16_t ngram_size_m     = 48; // mgram size for speculative tokens
     uint16_t ngram_min_hits   =  1; // minimum hits at ngram/mgram lookup for mgram to be proposed
-    uint16_t ckpt_num_tries   =  0; // number of tries in case of recurrent memory
+    bool     use_checkpoints  =  false; // use checkpoints to rewind in token history of recurrent models
+
 
     std::shared_ptr<common_ngram_mod> ngram_mod;
 

common/speculative.cpp

@@ -1136,8 +1136,8 @@ struct common_speculative_session::impl {
             clear_draft();
             return draft;
         }
-        if (params_spec.ckpt_num_tries > 0
-                && spec_ckpt_n_denials >= params_spec.ckpt_num_tries) {
+        if (params_spec.use_checkpoints
+                && spec_ckpt_n_denials > 0) {
             clear_draft();
             return draft;
         }
@@ -1166,7 +1166,7 @@ struct common_speculative_session::impl {
             draft.resize(n_draft_max);
         }
 
-        bool do_checkpoint = !draft.empty() && params_spec.ckpt_num_tries > 0;
+        bool do_checkpoint = !draft.empty() && params_spec.use_checkpoints;
         if (do_checkpoint && cached_text_tokens.size() > 5) {
             LOG_DBG("draft.size = %zu, n_spec_denials = %d, do_checkpoint = %s, tokens=[..., %d, %d, %d]\n",
                 draft.size(), spec_ckpt_n_denials,
@@ -1235,8 +1235,6 @@ struct common_speculative_session::impl {
                     return common_speculative_accept_response(std::move(ids), n_draft, true);
                 }
 
-                //spec_ckpt_n_accepted = (spec_ckpt_n_denials < params_spec.ckpt_num_tries) ? (int) (ids.size() - 1) : 0;
-
                 callback.batch_clear();
             }
         }

tools/server/server-context.cpp

@@ -838,7 +838,7 @@ struct server_context_impl {
             slot.prompt.tokens.has_mtmd = mctx != nullptr;
 
             // try speculative decoding
-            if (can_spec || params_base.speculative.ckpt_num_tries > 0) {
+            if (can_spec || params_base.speculative.use_checkpoints) {
                 if (mctx) {
                     SRV_ERR("%s\n", "speculative decoding is not supported with multimodal");
                     return false;