spec: Add benchmark-only synthetic speculative acceptance options (#27711)

* Add benchmark-only synthetic speculative acceptance to llama-server and llama-cli

* Address review comments

* Address review comments

* Add some comments in the code

tools/cli/README.md

@@ -200,6 +200,8 @@
 | `--spec-draft-n-cpu-moe, --spec-draft-ncmoe, -ncmoed, --n-cpu-moe-draft N` | keep the Mixture of Experts (MoE) weights of the first N layers in the CPU for the draft model<br/>(env: LLAMA_ARG_SPEC_DRAFT_N_CPU_MOE) |
 | `--spec-draft-n-max N` | number of tokens to draft for speculative decoding (default: 3)<br/>(env: LLAMA_ARG_SPEC_DRAFT_N_MAX) |
 | `--spec-draft-n-min N` | minimum number of draft tokens to use for speculative decoding (default: 0)<br/>(env: LLAMA_ARG_SPEC_DRAFT_N_MIN) |
+| `--spec-synth-len L` | target mean synthetic acceptance length, including the target token (benchmarking only)<br/>(env: LLAMA_ARG_SPEC_SYNTH_LEN) |
+| `--spec-synth-rates P0,P1,...` | comma-separated unconditional per-position synthetic acceptance probabilities (benchmarking only)<br/>(env: LLAMA_ARG_SPEC_SYNTH_RATES) |
 | `--spec-draft-p-split, --draft-p-split P` | speculative decoding split probability (default: 0.10)<br/>(env: LLAMA_ARG_SPEC_DRAFT_P_SPLIT) |
 | `--spec-draft-p-min, --draft-p-min P` | minimum speculative decoding probability (greedy) (default: 0.00)<br/>(env: LLAMA_ARG_SPEC_DRAFT_P_MIN) |
 | `--spec-draft-backend-sampling, --no-spec-draft-backend-sampling` | offload draft sampling to the backend (default: enabled)<br/>(env: LLAMA_ARG_SPEC_DRAFT_BACKEND_SAMPLING) |

tools/server/README.md

@@ -259,6 +259,8 @@ For the full list of features, please refer to [server's changelog](https://gith
 | `--spec-draft-n-cpu-moe, --spec-draft-ncmoe, -ncmoed, --n-cpu-moe-draft N` | keep the Mixture of Experts (MoE) weights of the first N layers in the CPU for the draft model<br/>(env: LLAMA_ARG_SPEC_DRAFT_N_CPU_MOE) |
 | `--spec-draft-n-max N` | number of tokens to draft for speculative decoding (default: 3)<br/>(env: LLAMA_ARG_SPEC_DRAFT_N_MAX) |
 | `--spec-draft-n-min N` | minimum number of draft tokens to use for speculative decoding (default: 0)<br/>(env: LLAMA_ARG_SPEC_DRAFT_N_MIN) |
+| `--spec-synth-len L` | target mean synthetic acceptance length, including the target token (benchmarking only)<br/>(env: LLAMA_ARG_SPEC_SYNTH_LEN) |
+| `--spec-synth-rates P0,P1,...` | comma-separated unconditional per-position synthetic acceptance probabilities (benchmarking only)<br/>(env: LLAMA_ARG_SPEC_SYNTH_RATES) |
 | `--spec-draft-p-split, --draft-p-split P` | speculative decoding split probability (default: 0.10)<br/>(env: LLAMA_ARG_SPEC_DRAFT_P_SPLIT) |
 | `--spec-draft-p-min, --draft-p-min P` | minimum speculative decoding probability (greedy) (default: 0.00)<br/>(env: LLAMA_ARG_SPEC_DRAFT_P_MIN) |
 | `--spec-draft-backend-sampling, --no-spec-draft-backend-sampling` | offload draft sampling to the backend (default: enabled)<br/>(env: LLAMA_ARG_SPEC_DRAFT_BACKEND_SAMPLING) |

tools/server/server-context.cpp

@@ -23,6 +23,7 @@
 #include <exception>
 #include <memory>
 #include <filesystem>
+#include <random>
 #include <utility>
 #include <fstream>
 
@@ -51,6 +52,50 @@ static common_speculative_output_limits server_output_limits(const common_params
     return result;
 }
 
+// synthetic draft verification for benchmarking - accept draft tokens at random instead of by match with the target
+// on replay the draft was already accepted before a context checkpoint restore, so repeat the same decisions
+static std::vector<llama_token> server_sample_and_accept_synth(
+        common_sampler * smpl,
+        llama_context * ctx,
+        const std::vector<int32_t> & idxs,
+        const llama_tokens & draft,
+        const std::vector<double> & synth_probs,
+        std::mt19937 & rng,
+        bool is_replay) {
+    GGML_ASSERT(idxs.size() == draft.size() + 1);
+    GGML_ASSERT(synth_probs.size() >= draft.size());
+
+    std::vector<llama_token> result;
+    result.reserve(idxs.size());
+
+    const llama_vocab * vocab = llama_model_get_vocab(llama_get_model(ctx));
+    std::uniform_real_distribution<double> dist(0.0, 1.0);
+    for (size_t i = 0; i < draft.size(); ++i) {
+        const llama_token id = common_sampler_sample(smpl, ctx, idxs[i]);
+        const bool accept = is_replay || dist(rng) < synth_probs[i];
+        // do not accept a drafted EOG token - it would end the generation early
+        // on replay the last token is from the target and can be EOG, so skip this check
+        if (accept && (is_replay || !llama_vocab_is_eog(vocab, draft[i]))) {
+            // synthetic draft tokens do not advance grammar or reasoning state
+            // the last replay token is from the target and must advance both
+            const bool is_replay_target = is_replay && i + 1 == draft.size();
+            common_sampler_accept(smpl, draft[i], is_replay_target);
+            result.push_back(draft[i]);
+            continue;
+        }
+
+        common_sampler_accept(smpl, id, true);
+        result.push_back(id);
+        return result;
+    }
+
+    const llama_token id = common_sampler_sample(smpl, ctx, idxs[draft.size()]);
+    common_sampler_accept(smpl, id, true);
+    result.push_back(id);
+
+    return result;
+}
+
 // state diagram: https://github.com/ggml-org/llama.cpp/pull/9283
 enum slot_state {
     SLOT_STATE_IDLE,
@@ -211,6 +256,7 @@ struct server_slot {
     std::vector<int32_t> spec_i_batch;
     common_prompt_checkpoint spec_ckpt;
     bool spec_is_replay = false;
+    std::mt19937 spec_synth_rng;
 
     // TODO: move members that belong to the task (such as `generated_text`, `has_new_line`) to task_results_state
     //       see https://github.com/ggml-org/llama.cpp/pull/18283#issuecomment-3710175837
@@ -1194,6 +1240,9 @@ private:
                 spec.reset(common_speculative_init(params_base.speculative, params_base.n_parallel));
             } catch (const std::exception & e) {
                 SRV_ERR("failed to initialize speculative decoding context: %s\n", e.what());
+                if (params_base.speculative.has_synth()) {
+                    return false;
+                }
             }
         }
 
@@ -1209,6 +1258,11 @@ private:
             model_dft = nullptr;
         }
 
+        if (!spec && params_base.speculative.has_synth()) {
+            SRV_ERR("%s", "synthetic acceptance requires an initialized speculative decoding context\n");
+            return false;
+        }
+
         for (int i = 0; i < params_base.n_parallel; i++) {
             server_slot & slot = slots[i];
 
@@ -1717,6 +1771,13 @@ private:
 
             SLT_TRC(slot, "sampler chain: %s\n", common_sampler_print(slot.smpl.get()).c_str());
             SLT_TRC(slot, "sampler params: \n%s\n", task.params.sampling.print().c_str());
+
+            if (spec && !common_speculative_get_synth_probs(spec.get()).empty()) {
+                const uint32_t seed = task.params.sampling.seed == LLAMA_DEFAULT_SEED
+                    ? std::random_device{}()
+                    : task.params.sampling.seed;
+                slot.spec_synth_rng.seed(seed);
+            }
         } else {
             slot.smpl.reset();
         }
@@ -3802,7 +3863,12 @@ private:
                 common_sampler_ptr smpl_save(common_sampler_clone(slot.smpl.get()));
 
                 GGML_ASSERT(slot.spec_i_batch.size() == n_draft + 1);
-                auto accepted = common_sampler_sample_and_accept_n(slot.smpl.get(), slot.ctx_tgt, slot.spec_i_batch, slot.spec_draft);
+                const auto & synth_probs = common_speculative_get_synth_probs(spec.get());
+                auto accepted = synth_probs.empty()
+                    ? common_sampler_sample_and_accept_n(slot.smpl.get(), slot.ctx_tgt, slot.spec_i_batch, slot.spec_draft)
+                    : server_sample_and_accept_synth(
+                            slot.smpl.get(), slot.ctx_tgt, slot.spec_i_batch, slot.spec_draft,
+                            synth_probs, slot.spec_synth_rng, slot.spec_is_replay);
                 slot.spec_i_batch.clear();
 
                 GGML_ASSERT(accepted.size() >= 1);
@@ -3868,7 +3934,7 @@ private:
 
             auto & n_accepted_per_pos = slot.n_accepted_per_pos;
             if (n_accepted_per_pos.empty()) {
-                n_accepted_per_pos.resize(common_speculative_n_max(&params_base.speculative), 0);
+                n_accepted_per_pos.resize(common_speculative_n_max(spec.get()), 0);
             }
             for (size_t i = 0; i < n_accepted && i < n_accepted_per_pos.size(); ++i) {
                 n_accepted_per_pos[i]++;

tools/server/tests/unit/test_speculative.py

@@ -52,6 +52,18 @@ def test_with_and_without_draft():
 
     assert tokens_no_draft == tokens_draft
 
+    server.stop()
+    create_server()
+    assert server.spec_draft_n_max is not None
+    server.spec_synth_rates = [0.0] * server.spec_draft_n_max
+    server.start()
+    res = server.make_request("POST", "/completion", data=request)
+
+    assert res.status_code == 200
+    assert res.body["timings"]["draft_n"] > 0
+    assert res.body["timings"]["draft_n_accepted"] == 0
+    assert res.body["tokens"] == tokens_no_draft
+
 
 def test_different_draft_min_draft_max():
     global server
@@ -80,6 +92,66 @@ def test_different_draft_min_draft_max():
         last_content = res.body["content"]
 
 
+def test_synth_is_deterministic():
+    global server
+    assert server.spec_draft_n_max is not None
+    server.spec_synth_rates = [0.75 ** (i + 1) for i in range(server.spec_draft_n_max)]
+    server.start()
+
+    request = {
+        "prompt": "I believe the meaning of life is",
+        "temperature": 0.2,
+        "top_k": 5,
+        "seed": 4242,
+        "n_predict": 32,
+    }
+    responses = [server.make_request("POST", "/completion", data=request) for _ in range(2)]
+
+    for res in responses:
+        assert res.status_code == 200
+        assert res.body["timings"]["draft_n"] > 0
+    assert responses[0].body["timings"]["draft_n"] == responses[1].body["timings"]["draft_n"]
+    assert responses[0].body["timings"]["draft_n_accepted"] == responses[1].body["timings"]["draft_n_accepted"]
+
+
+def test_synth_ignores_target_tokens():
+    global server
+    assert server.spec_draft_n_max is not None
+    server.spec_synth_rates = [1.0] * server.spec_draft_n_max
+    server.start()
+
+    res = server.make_request("POST", "/completion", data={
+        "prompt": "I believe the meaning of life is",
+        "temperature": 0.0,
+        "seed": 4242,
+        "n_predict": 32,
+    })
+
+    assert res.status_code == 200
+    assert res.body["timings"]["draft_n"] > 0
+    assert res.body["timings"]["draft_n_accepted"] == res.body["timings"]["draft_n"]
+
+    res = server.make_request("POST", "/completion", data={
+        "prompt": "I believe the meaning of life is",
+        "temperature": 0.0,
+        "seed": 4242,
+        "n_predict": 6,
+        "grammar": 'root ::= "a"{5,5}',
+    })
+    assert res.status_code == 200, res.body
+
+    res = server.make_request("POST", "/completion", data={
+        "prompt": "Respond with only: OK",
+        "temperature": 0.0,
+        "seed": 4242,
+        "n_predict": 64,
+        "ignore_eos": True,
+    })
+    assert res.status_code == 200, res.body
+    assert res.body["tokens_predicted"] == 64
+    assert res.body["stop_type"] == "limit"
+
+
 def test_slot_ctx_not_exceeded():
     global server
     server.n_ctx = 256

tools/server/tests/utils.py

@@ -99,6 +99,8 @@ class ServerProcess:
     spec_type: str | None = None
     spec_draft_n_min: int | None = None
     spec_draft_n_max: int | None = None
+    spec_synth_len: float | None = None
+    spec_synth_rates: List[float] | None = None
     no_ui: bool | None = None
     jinja: bool | None = None
     reasoning_format: Literal['deepseek', 'none', 'nothink'] | None = None
@@ -245,6 +247,11 @@ class ServerProcess:
             server_args.extend(["--spec-draft-n-max", self.spec_draft_n_max])
         if self.spec_draft_n_min:
             server_args.extend(["--spec-draft-n-min", self.spec_draft_n_min])
+        if self.spec_synth_len is not None:
+            server_args.extend(["--spec-synth-len", self.spec_synth_len])
+        if self.spec_synth_rates is not None:
+            rates = ",".join(str(rate) for rate in self.spec_synth_rates)
+            server_args.extend(["--spec-synth-rates", rates])
         if self.no_ui:
             server_args.append("--no-ui")
         if self.no_models_autoload: