tests: re-enable MiniMax M3 in test-llama-archs (#26633)

* mtmd: support multi-row batching for deepseek-ocr

* mtmd: weave deepseek-ocr rows in one shot instead of per row (#26615)

---------

Co-authored-by: Saba Fallah <sabafallah@gmail.com>

SECURITY.md

@@ -21,11 +21,18 @@ Please disclose it as a private [security advisory](https://github.com/ggml-org/
 
 A team of volunteers on a reasonable-effort basis maintains this project. As such, please give us at least 90 days to work on a fix before public exposure.
 
+### AI-powered code scan
+
+llama.cpp has an AI security scanner that scans the code periodically. The full prompts and tool set can be found in [ggml-org/security-scan-prompt](https://github.com/ggml-org/security-scan-prompt).
+
+We greatly appreciate reports that reflect genuine research effort, and we are happy to spend our time reviewing them. Findings that an autonomous AI agent can surface on its own add little on top of the scans we already run.
+
 ### Requirements
 
 Before submitting your report, ensure you meet the following requirements:
 
 - You have read this policy and fully understand it.
+- You have searched for existing discussions of the issue. If it has already been reported, your report will likely be rejected as a duplicate.
 - AI is only permitted in an assistive capacity as stated in [AGENTS.md](AGENTS.md). We do not accept reports that are written exclusively by AI.
 - Your report must include a working Proof-of-Concept in the form of a script and/or attached files.
 
@@ -46,6 +53,8 @@ Only vulnerabilities that fall within these parts of the project are considered
 
 Note that none of the topics under [Using llama.cpp securely](#using-llamacpp-securely) are considered vulnerabilities in LLaMA C++.
 
+Denial-of-Service (DoS) bugs are generally not treated as vulnerabilities. We don't reject them outright, but we look at them case-by-case and only accept those that are genuinely worth fixing.
+
 For vulnerabilities that fall within the `vendor` directory, please report them directly to the third-party project.
 
 ## Using llama.cpp securely

common/fit.cpp

@@ -136,7 +136,10 @@ static std::vector<llama_device_memory_data> common_get_device_memory_data_impl(
         devs.push_back(llama_model_get_device(model, i));
     }
 
-    hp_ngl         = llama_model_n_layer(model) + llama_model_n_layer_nextn(model);
+    hp_ngl         = llama_model_n_layer(model);
+    if (mparams->load_mtp) {
+        hp_ngl    += llama_model_n_layer_nextn(model);
+    }
     hp_n_ctx_train = llama_model_n_ctx_train(model);
     hp_n_expert    = llama_model_n_expert(model);
 

conversion/deepseek.py

@@ -17,8 +17,11 @@ from .base import LazyTorchTensor, MmprojModel, ModelBase, TextModel, gguf, logg
 from .qwen import QwenModel
 
 
-@ModelBase.register("DeepseekOCRForCausalLM", "UnlimitedOCRForCausalLM")
+@ModelBase.register("DeepseekOCRForCausalLM")
 class DeepseekOCRVisionModel(MmprojModel):
+    # HF dynamic_preprocess() max_num, which differs per model
+    preproc_max_tiles = 9
+
     def __init__(self, *args, **kwargs):
         super().__init__(*args, **kwargs)
         self.clip_projector_type = gguf.VisionProjectorType.DEEPSEEKOCR
@@ -43,6 +46,9 @@ class DeepseekOCRVisionModel(MmprojModel):
         # @bluebread: there's no window_size in config but just add it here anyway
         self.gguf_writer.add_vision_window_size(self.hparams.get("window_size", 14))
 
+        self.gguf_writer.add_vision_preproc_min_tiles(2)
+        self.gguf_writer.add_vision_preproc_max_tiles(self.preproc_max_tiles)
+
         # SAM configuration
         sam_hparams = hparams['sam']
         self.gguf_writer.add_vision_sam_layers_count(sam_hparams['layers'])
@@ -93,8 +99,15 @@ class DeepseekOCRVisionModel(MmprojModel):
         return super().filter_tensors((name, gen))
 
 
+@ModelBase.register("UnlimitedOCRForCausalLM")
+class UnlimitedOCRVisionModel(DeepseekOCRVisionModel):
+    preproc_max_tiles = 32
+
+
 @ModelBase.register("DeepseekOCR2ForCausalLM")
 class DeepseekOCR2VisionModel(DeepseekOCRVisionModel):
+    preproc_max_tiles = 6
+
     def __init__(self, *args, **kwargs):
         super().__init__(*args, **kwargs)
         self.clip_projector_type = gguf.VisionProjectorType.DEEPSEEKOCR2

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,
 }

src/llama-grammar.cpp

@@ -648,10 +648,12 @@ const char * llama_grammar_parser::parse_sequence(
             } else {
                 throw std::runtime_error(std::string("expecting ',' at ") + pos);
             }
-            bool has_max = max_times != UINT64_MAX;
-            if (min_times > MAX_REPETITION_THRESHOLD || (has_max && max_times > MAX_REPETITION_THRESHOLD)) {
+            if (min_times > MAX_REPETITION_THRESHOLD) {
                 throw std::runtime_error(std::string("number of repetitions exceeds sane defaults, please reduce the number of repetitions"));
             }
+            if (max_times != UINT64_MAX && max_times > MAX_REPETITION_THRESHOLD) {
+                max_times = UINT64_MAX;
+            }
             handle_repetitions(min_times, max_times);
         } else {
             break;

src/llama-model.cpp

@@ -2890,6 +2890,21 @@ void llama_model_base::create_tensor_qkv(llama_layer & layer, int bid,
         int64_t n_embd_, int64_t n_embd_q_, int64_t n_embd_k_, int64_t n_embd_v_,
         int flags) {
     const int64_t n_embd_qkv = n_embd_q_ + n_embd_k_ + n_embd_v_;
+
+    if (flags & TENSOR_SKIP) {
+        const int skip = TENSOR_NOT_REQUIRED | TENSOR_SKIP;
+
+        create_tensor(tn(LLM_TENSOR_ATTN_QKV, "weight", bid), {n_embd_, n_embd_qkv}, skip | TENSOR_SKIP_IF_VIRTUAL);
+        create_tensor(tn(LLM_TENSOR_ATTN_QKV, "bias",   bid), {n_embd_qkv},          skip | TENSOR_SKIP_IF_VIRTUAL);
+        create_tensor(tn(LLM_TENSOR_ATTN_Q,   "weight", bid), {n_embd_, n_embd_q_},  skip);
+        create_tensor(tn(LLM_TENSOR_ATTN_K,   "weight", bid), {n_embd_, n_embd_k_},  skip);
+        create_tensor(tn(LLM_TENSOR_ATTN_V,   "weight", bid), {n_embd_, n_embd_v_},  skip);
+        create_tensor(tn(LLM_TENSOR_ATTN_Q,   "bias",   bid), {n_embd_q_},           skip);
+        create_tensor(tn(LLM_TENSOR_ATTN_K,   "bias",   bid), {n_embd_k_},           skip);
+        create_tensor(tn(LLM_TENSOR_ATTN_V,   "bias",   bid), {n_embd_v_},           skip);
+        return;
+    }
+
     layer.wqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "weight", bid), {n_embd_, n_embd_qkv}, TENSOR_NOT_REQUIRED | TENSOR_SKIP_IF_VIRTUAL);
     if (layer.wqkv) {
         layer.wqkv_b = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "bias", bid), {n_embd_qkv}, TENSOR_NOT_REQUIRED | TENSOR_SKIP_IF_VIRTUAL);

tests/test-grammar-parser.cpp

@@ -153,6 +153,53 @@ int main()
         root ::= "a"{,10}"
     )""");
 
+    verify_failure(R"""(
+        root ::= "a"{5000}
+    )""");
+
+    verify_failure(R"""(
+        root ::= "a"{5000,}
+    )""");
+
+    verify_failure(R"""(
+        root ::= "a"{5000,6000}
+    )""");
+
+    verify_parsing(R"""(
+        root ::= "a"{0,5000}
+    )""", {
+        {"root", 0},
+        {"root_1", 1},
+    }, {
+        // root (index 0)
+        {LLAMA_GRETYPE_RULE_REF, /* root_1 */ 1},
+        {LLAMA_GRETYPE_END, 0},
+        // root_1 (index 1)
+        {LLAMA_GRETYPE_CHAR, 'a'},
+        {LLAMA_GRETYPE_RULE_REF, /* root_1 */ 1},
+        {LLAMA_GRETYPE_ALT, 0},
+        {LLAMA_GRETYPE_END, 0},
+    });
+
+    verify_parsing(R"""(
+        root ::= "a"{3,5000}
+    )""", {
+        {"root", 0},
+        {"root_1", 1},
+    }, {
+        // root (index 0)
+        {LLAMA_GRETYPE_CHAR, 'a'},
+        {LLAMA_GRETYPE_CHAR, 'a'},
+        {LLAMA_GRETYPE_CHAR, 'a'},
+        {LLAMA_GRETYPE_RULE_REF, /* root_1 */ 1},
+        {LLAMA_GRETYPE_END, 0},
+        // root_1 (index 1)
+        {LLAMA_GRETYPE_CHAR, 'a'},
+        {LLAMA_GRETYPE_RULE_REF, /* root_1 */ 1},
+        {LLAMA_GRETYPE_ALT, 0},
+        {LLAMA_GRETYPE_END, 0},
+    });
+
     verify_parsing(R"""(
         root  ::= "a"
     )""", {

tests/test-llama-archs.cpp

@@ -432,7 +432,7 @@ static bool arch_supported(const llm_arch arch) {
 
     // FIXME: these hit scheduler/view-backed-output issues with WebGPU on CI.
 #ifdef GGML_USE_WEBGPU
-    if (arch == LLM_ARCH_DEEPSEEK32 || arch == LLM_ARCH_GLM_DSA || arch == LLM_ARCH_MINIMAX_M3) {
+    if (arch == LLM_ARCH_DEEPSEEK32 || arch == LLM_ARCH_GLM_DSA) {
         return false;
     }
 #endif // GGML_USE_WEBGPU

tools/mtmd/clip.cpp

@@ -1761,6 +1761,10 @@ struct clip_model_loader {
                             // qwen2 encoder is GQA, requires KEY_N_HEAD_KV
                             get_u32(string_format(KEY_N_HEAD_KV, "vision"), hparams.n_head_kv);
                         }
+                        // unlimited-ocr shares the v1 projector but tiles up to 32
+                        get_u32(KEY_PREPROC_MIN_TILES, hparams.preproc_min_tiles, false);
+                        get_u32(KEY_PREPROC_MAX_TILES, hparams.preproc_max_tiles, false);
+                        GGML_ASSERT(hparams.preproc_min_tiles <= hparams.preproc_max_tiles);
                      } break;
                 case PROJECTOR_TYPE_HUNYUANVL:
                     {
@@ -1909,6 +1913,9 @@ struct clip_model_loader {
                 if (hparams.image_max_pixels > 0) {
                     LOG_INF("%s: image_max_pixels:   %d%s\n", __func__, hparams.image_max_pixels, hparams.custom_image_max_tokens > 0 ? " (custom value)" : "");
                 }
+                if (hparams.preproc_max_tiles > 0) {
+                    LOG_INF("%s: preproc_tiles:      %d - %d\n", __func__, hparams.preproc_min_tiles, hparams.preproc_max_tiles);
+                }
             } else if (is_audio) {
                 LOG_INF("\n--- audio hparams ---\n");
                 LOG_INF("%s: n_mel_bins:         %d\n", __func__, hparams.n_mel_bins);

tools/mtmd/models/deepseekocr.cpp

@@ -253,6 +253,9 @@ ggml_cgraph * clip_graph_deepseekocr::build() {
 
     bool is_overview = img.add_viewsep;
     int n_tiles_per_row = 0;
+    // number of separate "row" images batched together in this graph call
+    // (captured now, before n_batch below gets repurposed as the SAM/ViT batch size)
+    const int n_rows_batch = n_batch;
 
     // note: we expect either a batch of rows or a batch of overviews, but not a mix of both
 
@@ -272,16 +275,18 @@ ggml_cgraph * clip_graph_deepseekocr::build() {
         GGML_ASSERT(img.ny() % img.nx() == 0);
         n_tiles_per_row = img.ny() / img.nx();
 
-        // input shape: [tile_size, tile_size * n_tiles_per_row, 3]
-        // we want to reshape it to [tile_size, tile_size, 3, n_tiles_per_row]
-        inp_raw = ggml_reshape_4d(ctx0, inp_raw, img.nx(), img.nx(), n_tiles_per_row, 3);
-        inp_raw = ggml_cont(ctx0, ggml_permute(ctx0, inp_raw, 0, 1, 3, 2));
+        // each entry is one "row" image of shape [tile_size, tile_size * n_tiles_per_row, 3];
+        // merge the tile axis into the batch axis, giving a combined SAM input of shape
+        // [tile_size, tile_size, 3, n_tiles_per_row * n_rows_batch] (tile fast, row slow)
+        inp_raw = ggml_reshape_4d(ctx0, inp_raw, img.nx() * img.nx(), n_tiles_per_row, 3, n_rows_batch);
+        inp_raw = ggml_cont(ctx0, ggml_permute(ctx0, inp_raw, 0, 2, 1, 3));
+        inp_raw = ggml_reshape_4d(ctx0, inp_raw, img.nx(), img.nx(), 3, n_tiles_per_row * n_rows_batch);
     }
 
     ggml_tensor * sam_out = build_sam(inp_raw);
 
     if (!is_overview) {
-        n_batch = n_tiles_per_row;
+        n_batch = n_tiles_per_row * n_rows_batch;
     }
 
     const int clip_n_patches = sam_out->ne[0] * sam_out->ne[1];
@@ -354,34 +359,36 @@ ggml_cgraph * clip_graph_deepseekocr::build() {
         const auto w     = h;
         const auto n_dim = cur->ne[0];
 
-        ggml_tensor * imgnl = ggml_repeat_4d(ctx0, model.image_newline, n_dim, 1, h, 1);
-        cur = ggml_reshape_3d(ctx0, cur, n_dim, w, h);
-        cur = ggml_reshape_2d(ctx0, ggml_concat(ctx0, cur, imgnl, 1), n_dim, (w + 1) * h);
-        cur = ggml_concat(ctx0, cur, model.view_seperator, 1);  // (n_dim, h*(w+1) + 1)
+        ggml_tensor * imgnl = ggml_repeat_4d(ctx0, model.image_newline, n_dim, 1, h, n_batch);
+        cur = ggml_reshape_4d(ctx0, cur, n_dim, w, h, n_batch);
+        cur = ggml_reshape_3d(ctx0, ggml_concat(ctx0, cur, imgnl, 1), n_dim, (w + 1) * h, n_batch);
+        ggml_tensor * vs = ggml_repeat_4d(ctx0, model.view_seperator, n_dim, 1, n_batch, 1);
+        cur = ggml_concat(ctx0, cur, vs, 1);  // (n_dim, h*(w+1) + 1, n_batch)
     } else {
         // tile row: interleave tiles within each row, add newline per row
-        const int grid_x      = static_cast<int>(std::sqrt(static_cast<float>(clip_n_patches)));
-        const int grid_y      = grid_x;
-        const auto n_dim      = cur->ne[0];
+        const int  grid_x = static_cast<int>(std::sqrt(static_cast<float>(clip_n_patches)));
+        const int  grid_y = grid_x;
+        const auto n_dim  = cur->ne[0];
 
-        // (n_dim, clip_n_patches, n_batch) -> (n_dim, grid_x, grid_y, n_batch)
-        cur = ggml_reshape_4d(ctx0, cur, n_dim, grid_x, grid_y, n_batch);
+        // merge n_dim into the grid_x axis, freeing the 4th axis for n_rows_batch
+        // (n_dim, clip_n_patches, n_tiles_per_row * n_rows_batch) -> (n_dim*grid_x, grid_y, n_tiles_per_row, n_rows_batch)
+        cur = ggml_reshape_4d(ctx0, cur, n_dim * grid_x, grid_y, n_tiles_per_row, n_rows_batch);
 
         // tiles: re-order from A.row0 A.row1 B.row0 B.row1 ...
         //        to A.row0 B.row0 A.row1 B.row1 ...
         //        then add nl: A.row0 B.row0 [nl] A.row1 B.row1 [nl] ...
-        // interleave tiles: (n_dim, grid_x, grid_y, n_batch) -> (n_dim, grid_x, n_batch, grid_y)
-        cur = ggml_cont(ctx0, ggml_permute(ctx0, cur, 0, 1, 3, 2));
+        // interleave tiles: -> (n_dim*grid_x, n_tiles_per_row, grid_y, n_rows_batch)
+        cur = ggml_cont(ctx0, ggml_permute(ctx0, cur, 0, 2, 1, 3));
 
-        // merge: (n_dim, grid_x, n_batch, grid_y) -> (n_dim, grid_x*n_batch, grid_y, 1)
-        cur = ggml_reshape_4d(ctx0, cur, n_dim, grid_x * n_batch, grid_y, 1);
+        // merge: -> (n_dim, grid_x*n_tiles_per_row, grid_y, n_rows_batch)
+        cur = ggml_reshape_4d(ctx0, cur, n_dim, grid_x * n_tiles_per_row, grid_y, n_rows_batch);
 
-        // append newline per row: (n_dim, grid_x*n_batch+1, grid_y, 1)
-        ggml_tensor * imgnl = ggml_repeat_4d(ctx0, model.image_newline, n_dim, 1, grid_y, 1);
+        // append newline per row: (n_dim, grid_x*n_tiles_per_row+1, grid_y, n_rows_batch)
+        ggml_tensor * imgnl = ggml_repeat_4d(ctx0, model.image_newline, n_dim, 1, grid_y, n_rows_batch);
         cur = ggml_concat(ctx0, cur, imgnl, 1);
 
-        // flatten: (n_dim, (grid_x*n_batch+1)*grid_y)
-        cur = ggml_reshape_2d(ctx0, cur, n_dim, (grid_x * n_batch + 1) * grid_y);
+        // flatten: (n_dim, (grid_x*n_tiles_per_row+1)*grid_y, n_rows_batch)
+        cur = ggml_reshape_3d(ctx0, cur, n_dim, (grid_x * n_tiles_per_row + 1) * grid_y, n_rows_batch);
     }
 
     cb(cur, "dsocr_output", -1);

tools/mtmd/models/deepseekocr2.cpp

@@ -14,8 +14,9 @@ ggml_cgraph * clip_graph_deepseekocr2::build() {
     {
         ggml_tensor * inp;
 
-        inp = ggml_reshape_2d(ctx0, sam_out, sam_out->ne[0] * sam_out->ne[1], sam_out->ne[2]); // H*W, C
-        inp = ggml_cont(ctx0, ggml_permute(ctx0, inp, 1, 0, 2, 3));
+        // H*W, C, B
+        inp = ggml_reshape_3d(ctx0, sam_out, sam_out->ne[0] * sam_out->ne[1], sam_out->ne[2], sam_out->ne[3]);
+        inp = ggml_cont(ctx0, ggml_permute(ctx0, inp, 1, 0, 2, 3)); // C, H*W, B
 
         auto num_image_tokens = inp->ne[1]; // H*W
         GGML_ASSERT(num_image_tokens == 144 || num_image_tokens == 256);
@@ -32,8 +33,10 @@ ggml_cgraph * clip_graph_deepseekocr2::build() {
             num_queries = 144;
         }
 
-        // (B, num_image_tokens + num_queries, C)
-        inp = ggml_concat(ctx0, inp, ggml_cast(ctx0, query_embed, inp->type), 1);
+        // repeat the query embedding per batch item, then append: (C, num_image_tokens + num_queries, B)
+        query_embed = ggml_cast(ctx0, query_embed, inp->type);
+        query_embed = ggml_repeat_4d(ctx0, query_embed, query_embed->ne[0], num_queries, inp->ne[2], 1);
+        inp = ggml_concat(ctx0, inp, query_embed, 1);
 
         auto seq_len = inp->ne[1];
 
@@ -57,7 +60,7 @@ ggml_cgraph * clip_graph_deepseekocr2::build() {
                                       /* learned_pos_embd */ nullptr, add_rope, vit_opts);
 
         cur = ggml_cont(ctx0,
-                        ggml_view_2d(ctx0, cur, cur->ne[0], num_queries, cur->nb[1],
+                        ggml_view_3d(ctx0, cur, cur->ne[0], num_queries, cur->ne[2], cur->nb[1], cur->nb[2],
                                      cur->nb[1] * (cur->ne[1] - num_queries))); // only take query tokens for output
 
         ggml_build_forward_expand(gf, cur);
@@ -71,7 +74,8 @@ ggml_cgraph * clip_graph_deepseekocr2::build() {
 
     // view_seperator only after the global view
     if (img.add_viewsep) {
-        cur = ggml_concat(ctx0, cur, model.view_seperator, 1); // (n_dim, 257)
+        ggml_tensor * vs = ggml_repeat_4d(ctx0, model.view_seperator, model.view_seperator->ne[0], 1, cur->ne[2], 1);
+        cur = ggml_concat(ctx0, cur, vs, 1); // (n_dim, 257, n_batch)
     }
 
     cb(cur, "dsocr2_output", -1);

tools/mtmd/models/models.h

@@ -138,12 +138,13 @@ struct clip_graph_deepseekocr : clip_graph {
     clip_graph_deepseekocr(clip_ctx * ctx, const clip_image_f32 & img) : clip_graph(ctx, img) {}
     ggml_cgraph * build() override;
     ggml_tensor * build_sam(ggml_tensor * inp); // build the SAM model
-    // bool support_batch() const override { return true; } // TODO: support batch for DeepSeek-OCR v1
+    bool support_batch() const override { return true; }
 };
 
 struct clip_graph_deepseekocr2 : clip_graph_deepseekocr {
     clip_graph_deepseekocr2(clip_ctx * ctx, const clip_image_f32 & img) : clip_graph_deepseekocr(ctx, img) {}
     ggml_cgraph * build() override; // reuses build_sam() from base
+    bool support_batch() const override { return true; }
 };
 
 struct clip_graph_conformer : clip_graph {

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();