fix: use rs for only MTP

Author: am17an

common/common.cpp

@@ -1496,9 +1496,10 @@ struct llama_context_params common_context_params_to_llama(const common_params &
     cparams.n_ctx             = params.n_ctx;
     cparams.n_seq_max         = params.n_parallel;
     {
-        const bool has_spec = (params.speculative.type != COMMON_SPECULATIVE_TYPE_NONE)
-                              || params.speculative.has_dft();
-        cparams.n_rs_seq = has_spec ? (uint32_t) params.speculative.draft.n_max : 0u;
+        // enable partial rollback only for MTP, each recurrent slot requires memory
+        // and MTP uses max 3-4 slots vs other techniques
+        const bool has_mtp_spec = params.speculative.type == COMMON_SPECULATIVE_TYPE_MTP;
+        cparams.n_rs_seq = has_mtp_spec ? (uint32_t) params.speculative.draft.n_max : 0u;
     }
     cparams.n_batch           = params.n_batch;
     cparams.n_ubatch          = params.n_ubatch;

common/common.h

@@ -370,10 +370,6 @@ struct common_params_speculative {
     bool has_dft() const {
         return !draft.mparams.path.empty() || !draft.mparams.hf_repo.empty();
     }
-
-    bool has_mtp() const {
-        return type == COMMON_SPECULATIVE_TYPE_MTP && mtp.model != nullptr;
-    }
 };
 
 struct common_params_vocoder {

common/speculative.cpp

@@ -1152,7 +1152,7 @@ common_speculative * common_speculative_init(
     }
 
     llama_context * ctx_mtp = nullptr;
-    if (params.has_mtp()) {
+    if (params.type == COMMON_SPECULATIVE_TYPE_MTP) {
         ctx_mtp = llama_init_from_model(params.mtp.model, params.mtp.cparams);
         if (ctx_mtp == nullptr) {
             LOG_ERR("%s", "failed to create MTP context\n");

tools/server/server-context.cpp

@@ -963,7 +963,7 @@ struct server_context_impl {
 
             // try speculative decoding
             if (ctx_seq_rm_type != COMMON_CONTEXT_SEQ_RM_TYPE_NO) {
-                slot.is_mtp_enabled = params_base.speculative.has_mtp();
+                slot.is_mtp_enabled = params_base.speculative.type == COMMON_SPECULATIVE_TYPE_MTP;
                 slot.spec.reset(common_speculative_init(params_base.speculative, slot.ctx));
 
                 if (slot.spec) {