server : simplify

ggml-ci

Author: ggerganov

examples/server/server.cpp

@@ -2251,6 +2251,8 @@ struct server_context {
 
                     id = common_sampler_sample(slot.smpl, ctx, slot.i_batch - i);
 
+                    slot.i_batch = -1;
+
                     common_sampler_accept(slot.smpl, id, true);
 
                     slot.n_decoded += 1;
@@ -2277,73 +2279,64 @@ struct server_context {
                         slot.print_timings();
                         send_final_response(slot);
                         metrics.on_prediction(slot);
+                        continue;
                     }
                 }
 
-                slot.i_batch = -1;
-
-                if (slot.ctx_dft) {
-                    struct common_speculative_params params_spec;
-                    params_spec.n_draft   = params.n_draft;
-                    params_spec.n_reuse   = 256;
-                    params_spec.p_min     = 0.9f;
-
-                    llama_tokens draft = common_speculative_gen_draft(slot.spec, params_spec, slot.cache_tokens, id);
+                // check if the slot supports speculative decoding
+                if (!slot.ctx_dft) {
+                    continue;
+                }
 
-                    if (draft.size() > params.n_draft_min) {
-                        common_batch_clear(slot.batch_spec);
-                        common_batch_add(slot.batch_spec, id, slot.n_past++, { slot.id }, true);
+                // TODO: configurable through requests
+                struct common_speculative_params params_spec;
+                params_spec.n_draft   = params.n_draft;
+                params_spec.n_reuse   = 256;
+                params_spec.p_min     = 0.9f;
 
-                        for (size_t i = 0; i < draft.size(); ++i) {
-                            common_batch_add(slot.batch_spec, draft[i], slot.n_past + i, { slot.id }, true);
-                        }
+                llama_tokens draft = common_speculative_gen_draft(slot.spec, params_spec, slot.cache_tokens, id);
 
-                        llama_decode(ctx, slot.batch_spec);
-
-                        const auto ids = common_sampler_sample_n(slot.smpl, ctx, draft);
+                if (params.n_draft_min > (int) draft.size()) {
+                    continue;
+                }
 
-                        slot.n_past += ids.size() - 1;
+                // construct the speculation batch
+                common_batch_clear(slot.batch_spec);
+                common_batch_add  (slot.batch_spec, id, slot.n_past, { slot.id }, true);
 
-                        slot.cache_tokens.push_back(id);
+                for (size_t i = 0; i < draft.size(); ++i) {
+                    common_batch_add(slot.batch_spec, draft[i], slot.n_past + 1 + i, { slot.id }, true);
+                }
 
-                        for (size_t i = 0; i < ids.size(); ++i) {
-                            completion_token_output result;
+                llama_decode(ctx, slot.batch_spec);
 
-                            id = ids[i];
+                // the accepted tokens from the speculation
+                const auto ids = common_sampler_sample_n(slot.smpl, ctx, draft);
 
-                            common_sampler_accept(slot.smpl, id, true);
+                slot.n_past    += ids.size();
+                slot.n_decoded += ids.size();
 
-                            slot.n_decoded += 1;
-                            if (slot.n_decoded == 1) {
-                                slot.t_start_generation = ggml_time_us();
-                                slot.t_prompt_processing = (slot.t_start_generation - slot.t_start_process_prompt) / 1e3;
-                                metrics.on_prompt_eval(slot);
-                            }
+                slot.cache_tokens.push_back(id);
+                slot.cache_tokens.insert(slot.cache_tokens.end(), ids.begin(), ids.end() - 1);
 
-                            result.tok = id;
+                llama_kv_cache_seq_rm(ctx, slot.id, slot.n_past, -1);
 
-                            const auto * cur_p = common_sampler_get_candidates(slot.smpl);
+                for (size_t i = 0; i < ids.size(); ++i) {
+                    completion_token_output result;
 
-                            for (size_t i = 0; i < (size_t) slot.sparams.n_probs; ++i) {
-                                result.probs.push_back({
-                                    cur_p->data[i].id,
-                                        i >= cur_p->size ? 0.0f : cur_p->data[i].p,
-                                });
-                            }
+                    id = ids[i];
 
-                            if (!process_token(result, slot)) {
-                                // release slot because of stop condition
-                                slot.release();
-                                slot.print_timings();
-                                send_final_response(slot);
-                                metrics.on_prediction(slot);
-                                break;
-                            }
-                        }
+                    common_sampler_accept(slot.smpl, id, true);
 
-                        llama_kv_cache_seq_rm(ctx, slot.id, slot.n_past, -1);
+                    result.tok = id;
 
-                        slot.cache_tokens.insert(slot.cache_tokens.end(), ids.begin(), ids.end() - 1);
+                    if (!process_token(result, slot)) {
+                        // release slot because of stop condition
+                        slot.release();
+                        slot.print_timings();
+                        send_final_response(slot);
+                        metrics.on_prediction(slot);
+                        break;
                     }
                 }
             }