server: Adding spec-decode counters to /metrics endpoint (#26389)

* * server: add spec-decode counters to /metrics endpoint

* server: fixed review comments and now aligned param names exactly with vLLM.

gguf-py/gguf/scripts/gguf_convert_endian.py

@@ -59,11 +59,29 @@ def byteswap_q6_k(tensor, block_offs):
     delta.byteswap(inplace=True)
 
 
+def byteswap_q1_0(tensor, block_offs):
+    # Each block_q1_0 consists of an f16 delta followed by 16 int8 quantizations.
+
+    # Byte-Swap f16 sized delta field
+    delta = tensor.data[block_offs:block_offs + 2].view(dtype=np.uint16)
+    delta.byteswap(inplace=True)
+
+
+def byteswap_tq2_0(tensor, block_offs):
+    # Each block_tq2_0 consists of 64 int8 values followed by 1 f16 value.
+
+    # Byte-Swap f16 sized field
+    delta = tensor.data[block_offs + 64:block_offs + 66].view(dtype=np.uint16)
+    delta.byteswap(inplace=True)
+
+
 byteswap_tensors = {
+    gguf.GGMLQuantizationType.Q1_0:  byteswap_q1_0,
     gguf.GGMLQuantizationType.Q4_0:  byteswap_q4_0,
     gguf.GGMLQuantizationType.Q8_0:  byteswap_q8_0,
     gguf.GGMLQuantizationType.Q4_K:  byteswap_q4_k,
     gguf.GGMLQuantizationType.Q6_K:  byteswap_q6_k,
+    gguf.GGMLQuantizationType.TQ2_0: byteswap_tq2_0,
     gguf.GGMLQuantizationType.MXFP4: byteswap_noop,
     gguf.GGMLQuantizationType.NVFP4: byteswap_noop,
 }

tools/server/README.md

@@ -1076,6 +1076,10 @@ In *router mode* the query param `?model={model_id}` has to be set. This endpoin
 | `llamacpp:n_tokens_max` | Counter | High watermark of the context size observed. |
 | `llamacpp:n_decode_total` | Counter | Total Number of llama_decode() calls. |
 | `llamacpp:n_busy_slots_per_decode` | Gauge | Average number of busy slots per llama_decode() call. |
+| `llamacpp:spec_decode_num_draft_tokens_total` | Counter | Total draft tokens generated (0 when spec-decode is off). |
+| `llamacpp:spec_decode_num_accepted_tokens_total` | Counter | Total draft tokens accepted by the target model (0 when spec-decode is off). |
+| `llamacpp:spec_decode_num_drafts_total` | Counter | Total speculative decoding verification steps (0 when spec-decode is off). |
+| `llamacpp:spec_decode_num_accepted_tokens_per_pos_total` | Counter | Accepted tokens per draft position (labeled `position="N"`; absent when spec-decode is off or before the first completed speculative request). |
 
 ### POST `/slots/{id_slot}?action=save`: Save the prompt cache of the specified slot to a file.
 

tools/server/server-context.cpp

@@ -845,6 +845,11 @@ struct server_metrics {
     uint64_t n_decode_total     = 0;
     uint64_t n_busy_slots_total = 0;
 
+    uint64_t n_draft_tokens_total      = 0;
+    uint64_t n_draft_accepted_total    = 0;
+    uint64_t n_draft_verif_steps_total = 0;
+    std::vector<uint64_t> n_accepted_per_pos_total;
+
     void init() {
         t_start = ggml_time_us();
     }
@@ -863,6 +868,17 @@ struct server_metrics {
         n_tokens_predicted         += slot.n_decoded;
         t_tokens_generation        += slot.t_token_generation;
         t_tokens_generation_total  += slot.t_token_generation;
+
+        n_draft_tokens_total      += slot.n_draft_total;
+        n_draft_accepted_total    += slot.n_draft_accepted;
+        n_draft_verif_steps_total += slot.n_draft_verif_steps;
+
+        if (n_accepted_per_pos_total.size() < slot.n_accepted_per_pos.size()) {
+            n_accepted_per_pos_total.resize(slot.n_accepted_per_pos.size(), 0);
+        }
+        for (size_t i = 0; i < slot.n_accepted_per_pos.size(); i++) {
+            n_accepted_per_pos_total[i] += slot.n_accepted_per_pos[i];
+        }
     }
 
     void on_decoded(const std::vector<server_slot> & slots) {
@@ -2552,6 +2568,11 @@ private:
                     res->n_decode_total          = metrics.n_decode_total;
                     res->n_busy_slots_total      = metrics.n_busy_slots_total;
 
+                    res->n_draft_tokens_total      = metrics.n_draft_tokens_total;
+                    res->n_draft_accepted_total    = metrics.n_draft_accepted_total;
+                    res->n_draft_verif_steps_total = metrics.n_draft_verif_steps_total;
+                    res->n_accepted_per_pos_total  = metrics.n_accepted_per_pos_total;
+
                     if (task.metrics_reset_bucket) {
                         metrics.reset_bucket();
                     }
@@ -4440,6 +4461,18 @@ void server_routes::init_routes() {
                     {"name",  "n_tokens_max"},
                     {"help",  "Largest observed n_tokens."},
                     {"value",  res_task->n_tokens_max}
+            }, {
+                    {"name",  "spec_decode_num_draft_tokens_total"},
+                    {"help",  "Total draft tokens generated"},
+                    {"value",  res_task->n_draft_tokens_total}
+            }, {
+                    {"name",  "spec_decode_num_accepted_tokens_total"},
+                    {"help",  "Total draft tokens accepted by the target model"},
+                    {"value",  res_task->n_draft_accepted_total}
+            }, {
+                    {"name",  "spec_decode_num_drafts_total"},
+                    {"help",  "Total speculative decoding verification steps"},
+                    {"value",  res_task->n_draft_verif_steps_total}
             }}},
             {"gauge", {{
                     {"name",  "prompt_tokens_seconds"},
@@ -4481,6 +4514,17 @@ void server_routes::init_routes() {
             }
         }
 
+        // labeled counter: one time series per draft position
+        if (!res_task->n_accepted_per_pos_total.empty()) {
+            prometheus << "# HELP llamacpp:spec_decode_num_accepted_tokens_per_pos_total"
+                          " Accepted tokens per draft position\n"
+                       << "# TYPE llamacpp:spec_decode_num_accepted_tokens_per_pos_total counter\n";
+            for (size_t i = 0; i < res_task->n_accepted_per_pos_total.size(); i++) {
+                prometheus << "llamacpp:spec_decode_num_accepted_tokens_per_pos_total{position=\""
+                           << i << "\"} " << res_task->n_accepted_per_pos_total[i] << "\n";
+            }
+        }
+
         res->headers["Process-Start-Time-Unix"] = std::to_string(res_task->t_start);
         res->content_type = "text/plain; version=0.0.4";
         res->status = 200;

tools/server/server-task.cpp

@@ -1560,6 +1560,11 @@ json server_task_result_metrics::to_json() {
         { "n_decode_total",                  n_decode_total },
         { "n_busy_slots_total",              n_busy_slots_total },
 
+        { "n_draft_tokens_total",            n_draft_tokens_total },
+        { "n_draft_accepted_total",          n_draft_accepted_total },
+        { "n_draft_verif_steps_total",       n_draft_verif_steps_total },
+        { "n_accepted_per_pos_total",        n_accepted_per_pos_total },
+
         { "slots",                           slots_data },
     };
 }

tools/server/server-task.h

@@ -532,6 +532,11 @@ struct server_task_result_metrics : server_task_result {
     uint64_t n_decode_total     = 0;
     uint64_t n_busy_slots_total = 0;
 
+    uint64_t n_draft_tokens_total      = 0;
+    uint64_t n_draft_accepted_total    = 0;
+    uint64_t n_draft_verif_steps_total = 0;
+    std::vector<uint64_t> n_accepted_per_pos_total;
+
     // while we can also use std::vector<server_slot> this requires copying the slot object which can be quite messy
     // therefore, we use json to temporarily store the slot.to_json() result
     json slots_data = json::array();