Vulkan: Fix minor debug mode issues (#14899)

* vulkan: fix debug mode issues

* vulkan: remove broken check_results GGML_OP_SET_ROWS support

common/arg.cpp

@@ -977,6 +977,10 @@ static bool common_params_parse_ex(int argc, char ** argv, common_params_context
         for (auto & seq_breaker : params.sampling.dry_sequence_breakers) {
             string_process_escapes(seq_breaker);
         }
+        for (auto & pair : params.speculative.replacements) {
+            string_process_escapes(pair.first);
+            string_process_escapes(pair.second);
+        }
     }
 
     if (!params.kv_overrides.empty()) {
@@ -3249,6 +3253,13 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             params.speculative.model.path = value;
         }
     ).set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_MODEL_DRAFT"));
+    add_opt(common_arg(
+        {"--spec-replace"}, "TARGET", "DRAFT",
+        "translate the string in TARGET into DRAFT if the draft model and main model are not compatible",
+        [](common_params & params, const std::string & tgt, const std::string & dft) {
+            params.speculative.replacements.push_back({ tgt, dft });
+        }
+    ).set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_SERVER}));
     add_opt(common_arg(
         {"-ctkd", "--cache-type-k-draft"}, "TYPE",
         string_format(

common/common.h

@@ -201,6 +201,7 @@ struct common_params_speculative {
     int32_t n_gpu_layers =    -1; // number of layers to store in VRAM for the draft model (-1 - use default)
     float   p_split      =  0.1f; // speculative decoding split probability
     float   p_min        = 0.75f; // minimum speculative decoding probability (greedy)
+    std::vector<std::pair<std::string, std::string>> replacements; // main to speculative model replacements
 
     ggml_type cache_type_k = GGML_TYPE_F16; // KV cache data type for the K
     ggml_type cache_type_v = GGML_TYPE_F16; // KV cache data type for the V

common/speculative.cpp

@@ -1,30 +1,39 @@
 #include "speculative.h"
 
+#include "ggml.h"
+#include "llama.h"
 #include "log.h"
 #include "common.h"
 #include "sampling.h"
 
 #include <cstring>
 #include <algorithm>
+#include <map>
 
 #define SPEC_VOCAB_MAX_SIZE_DIFFERENCE  128
 #define SPEC_VOCAB_CHECK_START_TOKEN_ID 5
 
 struct common_speculative {
-    struct llama_context * ctx;
+    struct llama_context * ctx_tgt; // only used for retokenizing from ctx_dft
+    struct llama_context * ctx_dft;
     struct common_sampler * smpl;
 
     llama_batch batch;
-    llama_tokens prompt;
+    llama_tokens prompt_dft;
+    bool vocab_dft_compatible = true; // whether retokenization is needed
+    std::map<std::string, std::string> tgt_dft_replacements = {};
 };
 
 struct common_speculative * common_speculative_init(
+        struct llama_context * ctx_tgt,
         struct llama_context * ctx_dft) {
     auto * result = new common_speculative {
-        /* .ctx    = */ ctx_dft,
-        /* .smpl   = */ nullptr,
-        /* .batch  = */ llama_batch_init(llama_n_batch(ctx_dft), 0, 1),
-        /* .prompt = */ {},
+        /* .ctx_tgt    = */ ctx_tgt,
+        /* .ctx_dft    = */ ctx_dft,
+        /* .smpl       = */ nullptr,
+        /* .batch      = */ llama_batch_init(llama_n_batch(ctx_dft), 0, 1),
+        /* .prompt_dft = */ {},
+        /* .vocab_dft_compatible = */ false,
     };
 
     // TODO: optimize or pass from outside?
@@ -59,6 +68,9 @@ struct common_speculative * common_speculative_init(
     }
 #endif
 
+    result->vocab_dft_compatible = common_speculative_are_compatible(ctx_tgt, ctx_dft);
+    LOG_DBG("vocab_dft_compatible = %d\n", result->vocab_dft_compatible);
+
     return result;
 }
 
@@ -75,8 +87,8 @@ void common_speculative_free(struct common_speculative * spec) {
 }
 
 bool common_speculative_are_compatible(
-        const struct llama_context * ctx_tgt,
-        const struct llama_context * ctx_dft) {
+    const struct llama_context * ctx_tgt,
+    const struct llama_context * ctx_dft) {
     const struct llama_model * model_tgt = llama_get_model(ctx_tgt);
     const struct llama_model * model_dft = llama_get_model(ctx_dft);
 
@@ -90,31 +102,32 @@ bool common_speculative_are_compatible(
     LOG_DBG("%s: vocab_type dft: %d\n", __func__, vocab_type_dft);
 
     if (vocab_type_tgt != vocab_type_dft) {
-        LOG_ERR("%s: draft model vocab type must match target model to use speculation but "
-                     "vocab_type_dft = %d while vocab_type_tgt = %d\n", __func__, vocab_type_dft, vocab_type_tgt);
+        LOG_DBG("%s: draft model vocab type must match target model to use speculation but ", __func__);
+        LOG_DBG("vocab_type_dft = %d while vocab_type_tgt = %d\n", vocab_type_dft, vocab_type_tgt);
         return false;
     }
 
-    if (llama_vocab_get_add_bos(vocab_tgt) != llama_vocab_get_add_bos(vocab_dft) ||
+    if (
+        llama_vocab_get_add_bos(vocab_tgt) != llama_vocab_get_add_bos(vocab_dft) ||
         llama_vocab_get_add_eos(vocab_tgt) != llama_vocab_get_add_eos(vocab_dft) ||
         llama_vocab_bos(vocab_tgt) != llama_vocab_bos(vocab_dft) ||
-        llama_vocab_eos(vocab_tgt) != llama_vocab_eos(vocab_dft)) {
-        LOG_ERR("%s: draft vocab special tokens must match target vocab to use speculation\n", __func__);
-        LOG_ERR("%s: tgt: bos = %d (%d), eos = %d (%d)\n", __func__, llama_vocab_bos(vocab_tgt), llama_vocab_get_add_bos(vocab_tgt), llama_vocab_eos(vocab_tgt), llama_vocab_get_add_eos(vocab_tgt));
-        LOG_ERR("%s: dft: bos = %d (%d), eos = %d (%d)\n", __func__, llama_vocab_bos(vocab_dft), llama_vocab_get_add_bos(vocab_dft), llama_vocab_eos(vocab_dft), llama_vocab_get_add_eos(vocab_dft));
+        llama_vocab_eos(vocab_tgt) != llama_vocab_eos(vocab_dft)
+    ) {
+        LOG_DBG("%s: draft model special tokens must match target model to use speculation\n", __func__);
         return false;
     }
 
     {
         const int n_vocab_tgt = llama_vocab_n_tokens(vocab_tgt);
         const int n_vocab_dft = llama_vocab_n_tokens(vocab_dft);
-
-        const int vocab_diff = std::abs(n_vocab_tgt - n_vocab_dft);
+        const int vocab_diff  = n_vocab_tgt > n_vocab_dft
+            ? n_vocab_tgt - n_vocab_dft
+            : n_vocab_dft - n_vocab_tgt;
 
         if (vocab_diff > SPEC_VOCAB_MAX_SIZE_DIFFERENCE) {
-            LOG_ERR("%s: draft model vocab must closely match target model to use speculation but "
-                         "target vocab size %d does not match draft vocab size %d - difference %d, max allowed %d\n",
-                    __func__, n_vocab_tgt, llama_vocab_n_tokens(vocab_dft), vocab_diff, SPEC_VOCAB_MAX_SIZE_DIFFERENCE);
+            LOG_DBG("%s: draft model vocab must closely match target model to use speculation but ", __func__);
+            LOG_DBG("target vocab size %d does not match draft vocab size %d - difference %d, max allowed %d\n",
+                    n_vocab_tgt, llama_vocab_n_tokens(vocab_dft), vocab_diff, SPEC_VOCAB_MAX_SIZE_DIFFERENCE);
             return false;
         }
 
@@ -122,8 +135,8 @@ bool common_speculative_are_compatible(
             const char * token_text_tgt = llama_vocab_get_text(vocab_tgt, i);
             const char * token_text_dft = llama_vocab_get_text(vocab_dft, i);
             if (std::strcmp(token_text_tgt, token_text_dft) != 0) {
-                LOG_ERR("%s: draft vocab vocab must match target vocab to use speculation but "
-                             "token %d content differs - target '%s', draft '%s'\n", __func__, i,
+                LOG_DBG("%s: draft model vocab must match target model to use speculation but ", __func__);
+                LOG_DBG("token %d content differs - target '%s', draft '%s'\n", i,
                         common_token_to_piece(ctx_tgt, i).c_str(),
                         common_token_to_piece(ctx_dft, i).c_str());
                 return false;
@@ -134,32 +147,93 @@ bool common_speculative_are_compatible(
     return true;
 }
 
+void common_speculative_add_replacement_tgt_dft(
+        struct common_speculative * spec,
+        const char *source, const char *dest) {
+    spec->tgt_dft_replacements[source] = dest;
+}
+
+static std::string replace_to_dft(
+        struct common_speculative * spec,
+        const std::string& input) {
+    std::string result = input;
+    for (const auto & pair : spec->tgt_dft_replacements) {
+        size_t pos = result.find(pair.first);
+        while (pos != std::string::npos) {
+            result.replace(pos, pair.first.length(), pair.second);
+            pos = result.find(pair.first, pos + pair.second.length());
+        }
+    }
+    return result;
+}
+
+static std::string replace_to_tgt(
+        struct common_speculative * spec,
+        const std::string& input) {
+    std::string result = input;
+    for (const auto& pair : spec->tgt_dft_replacements) {
+        size_t pos = result.find(pair.second);
+        while (pos != std::string::npos) {
+            result.replace(pos, pair.second.length(), pair.first);
+            pos = result.find(pair.second, pos + pair.first.length());
+        }
+    }
+    return result;
+}
+
+
 llama_tokens common_speculative_gen_draft(
         struct common_speculative * spec,
         struct common_speculative_params params,
-        const llama_tokens & prompt_tgt,
+        const llama_tokens & prompt_tgt_main_model, // specified in target model vocab
         llama_token id_last) {
     auto & batch  = spec->batch;
-    auto & ctx    = spec->ctx;
+    auto & ctx_tgt = spec->ctx_tgt;
+    auto & ctx_dft = spec->ctx_dft;
     auto & smpl   = spec->smpl;
-    auto & prompt = spec->prompt;
+    auto & prompt_dft = spec->prompt_dft;
 
-    auto * mem = llama_get_memory(ctx);
+    auto * mem_dft = llama_get_memory(ctx_dft);
 
     int reuse_i = 0;
     int reuse_n = 0;
 
-    const int n_ctx = llama_n_ctx(ctx) - params.n_draft;
+    const int n_ctx = llama_n_ctx(ctx_dft) - params.n_draft;
+
+    llama_tokens prompt_tgt_draft_model;
+    if (!spec->vocab_dft_compatible) {
+        std::string text;
+        text = common_detokenize(ctx_tgt, prompt_tgt_main_model, true);
+        text = replace_to_dft(spec, text);
+        LOG_DBG("%s: main->draft detokenized string: '%s'\n", __func__, text.c_str());
+        prompt_tgt_draft_model = common_tokenize(ctx_dft, text, false, true);
+
+        // convert id_last to draft vocab. llama_detokenize is called directly to avoid an allocation
+        const auto * model_tgt = llama_get_model(ctx_tgt);
+        const auto * vocab_tgt = llama_model_get_vocab(model_tgt);
+
+        int32_t n_chars = llama_detokenize(vocab_tgt, &id_last, 1, nullptr, 0, false, false);
+        GGML_ASSERT(n_chars < 0 && "failed to detokenize id_last");
+        text.resize(-n_chars);
+        llama_detokenize(vocab_tgt, &id_last, 1, text.data(), text.size(), false, false);
+        text = replace_to_dft(spec, text);
+
+        LOG_DBG("main->draft detokenized id_last(%d): '%s'\n", id_last, text.c_str());
+        id_last = common_tokenize(ctx_dft, text, false, true)[0];
+    }
+    // prompt_tgt's tokens will always be compatible with ctx_dft
+    const llama_tokens &prompt_tgt =
+        spec->vocab_dft_compatible ? prompt_tgt_main_model : prompt_tgt_draft_model;
 
     const int i_start = std::max<int>(0, (int) prompt_tgt.size() - n_ctx);
 
     // reuse as much as possible from the old draft context
     // ideally, the draft context should be as big as the target context and we will always reuse the entire prompt
-    for (int i = 0; i < (int) prompt.size(); ++i) {
+    for (int i = 0; i < (int) prompt_dft.size(); ++i) {
         int cur = 0;
         while (i_start + cur < (int) prompt_tgt.size() &&
-               i       + cur < (int) prompt.size() &&
-               prompt_tgt[i_start + cur] == prompt[i + cur]) {
+               i       + cur < (int) prompt_dft.size() &&
+               prompt_tgt[i_start + cur] == prompt_dft[i + cur]) {
             cur++;
         }
 
@@ -169,21 +243,20 @@ llama_tokens common_speculative_gen_draft(
         }
     }
 
-    LOG_DBG("%s: reuse_i = %d, reuse_n = %d, prompt = %d\n", __func__, reuse_i, reuse_n, (int) prompt.size());
+    LOG_DBG("%s: reuse_i = %d, reuse_n = %d, prompt = %d\n", __func__, reuse_i, reuse_n, (int) prompt_dft.size());
 
     llama_tokens result;
     result.reserve(params.n_draft);
 
     if (reuse_n == 0) {
-        llama_memory_clear(mem, false);
-
-        prompt.clear();
+        llama_memory_clear(mem_dft, false);
+        prompt_dft.clear();
     } else {
         // this happens when a previous draft has been discarded (for example, due to being too small), but the
         // target model agreed with it. in this case, we simply pass back the previous results to save compute
-        if (reuse_i + reuse_n < (int) prompt.size() && prompt[reuse_i + reuse_n] == id_last) {
-            for (int i = reuse_i + reuse_n + 1; i < (int) prompt.size(); ++i) {
-                result.push_back(prompt[i]);
+        if (reuse_i + reuse_n < (int) prompt_dft.size() && prompt_dft[reuse_i + reuse_n] == id_last) {
+            for (int i = reuse_i + reuse_n + 1; i < (int) prompt_dft.size(); ++i) {
+                result.push_back(prompt_dft[i]);
 
                 if (params.n_draft <= (int) result.size()) {
                     break;
@@ -194,16 +267,15 @@ llama_tokens common_speculative_gen_draft(
         }
 
         if (reuse_i > 0) {
-            llama_memory_seq_rm (mem, 0, 0, reuse_i);
-            llama_memory_seq_add(mem, 0, reuse_i, -1, -reuse_i);
+            llama_memory_seq_rm (mem_dft, 0, 0, reuse_i);
+            llama_memory_seq_add(mem_dft, 0, reuse_i, -1, -reuse_i);
 
-            prompt.erase(prompt.begin(), prompt.begin() + reuse_i);
+            prompt_dft.erase(prompt_dft.begin(), prompt_dft.begin() + reuse_i);
         }
 
-        if (reuse_n < (int) prompt.size()) {
-            llama_memory_seq_rm (mem, 0, reuse_n, -1);
-
-            prompt.erase(prompt.begin() + reuse_n, prompt.end());
+        if (reuse_n < (int) prompt_dft.size()) {
+            llama_memory_seq_rm (mem_dft, 0, reuse_n, -1);
+            prompt_dft.erase(prompt_dft.begin() + reuse_n, prompt_dft.end());
         }
     }
 
@@ -214,28 +286,28 @@ llama_tokens common_speculative_gen_draft(
         //LOG_DBG("i = %d, i_start = %d, reuse_n = %d, i - i_start = %d, id = %6d\n", i, i_start, reuse_n, i - i_start, prompt_tgt[i]);
         common_batch_add(batch, prompt_tgt[i], i - i_start, { 0 }, false);
 
-        prompt.push_back(prompt_tgt[i]);
+        prompt_dft.push_back(prompt_tgt[i]);
     }
 
     // we should rarely end-up here during normal decoding
     if (batch.n_tokens > 0) {
         //LOG_DBG("%s: draft prompt batch: %s\n", __func__, string_from(ctx, batch).c_str());
 
-        llama_decode(ctx, batch);
+        llama_decode(ctx_dft, batch);
     }
 
-    const llama_pos n_past = prompt.size();
+    const llama_pos n_past = prompt_dft.size();
 
     LOG_DBG("%s: n_past = %d\n", __func__, n_past);
 
     common_batch_clear(batch);
     common_batch_add  (batch, id_last, n_past, { 0 }, true);
 
-    prompt.push_back(id_last);
+    prompt_dft.push_back(id_last);
 
-    //LOG_DBG("%s: draft prompt: %s\n", __func__, string_from(ctx, prompt).c_str());
+    LOG_DBG("%s: draft prompt: %s\n", __func__, string_from(ctx_dft, prompt_dft).c_str());
 
-    llama_decode(ctx, batch);
+    llama_decode(ctx_dft, batch);
 
     common_sampler_reset(smpl);
 
@@ -243,13 +315,13 @@ llama_tokens common_speculative_gen_draft(
     for (int i = 0; i < params.n_draft; ++i) {
         common_batch_clear(batch);
 
-        common_sampler_sample(smpl, ctx, 0, true);
+        common_sampler_sample(smpl, ctx_dft, 0, true);
 
         const auto * cur_p = common_sampler_get_candidates(smpl);
 
         for (int k = 0; k < std::min(3, (int) cur_p->size); ++k) {
             LOG_DBG(" - draft candidate %3d, pos %3d: %6d (%8.3f) '%s'\n",
-                    k, i, cur_p->data[k].id, cur_p->data[k].p, common_token_to_piece(ctx, cur_p->data[k].id).c_str());
+                    k, i, cur_p->data[k].id, cur_p->data[k].p, common_token_to_piece(ctx_dft, cur_p->data[k].id).c_str());
         }
 
         // add drafted token for each sequence
@@ -271,10 +343,19 @@ llama_tokens common_speculative_gen_draft(
         common_batch_add(batch, id, n_past + i + 1, { 0 }, true);
 
         // evaluate the drafted tokens on the draft model
-        llama_decode(ctx, batch);
+        llama_decode(ctx_dft, batch);
 
-        prompt.push_back(id);
+        prompt_dft.push_back(id);
     }
 
+    if (!spec->vocab_dft_compatible) {
+        std::string detokenized = common_detokenize(ctx_dft, result, true);
+        detokenized = replace_to_tgt(spec, detokenized);
+        LOG_DBG("draft->main detokenized string: '%s'\n", detokenized.c_str());
+        result = common_tokenize(ctx_tgt, detokenized, false, true);
+        if (result.size() > (size_t)params.n_draft) {
+            result.resize(params.n_draft);
+        }
+    }
     return result;
 }

common/speculative.h

@@ -12,7 +12,10 @@ struct common_speculative_params {
     float p_min = 0.75f; // min probability required to accept a token in the draft
 };
 
-struct common_speculative * common_speculative_init(struct llama_context * ctx_dft);
+struct common_speculative * common_speculative_init(
+        struct llama_context * ctx_tgt,
+        struct llama_context * ctx_dft
+);
 
 void common_speculative_free(struct common_speculative * spec);
 
@@ -20,6 +23,10 @@ bool common_speculative_are_compatible(
         const struct llama_context * ctx_tgt,
         const struct llama_context * ctx_dft);
 
+void common_speculative_add_replacement_tgt_dft(
+        struct common_speculative * spec,
+        const char *source, const char *dest);
+
 // sample up to n_draft tokens and add them to the batch using the draft model
 llama_tokens common_speculative_gen_draft(
                struct common_speculative * spec,

docs/multimodal/minicpmo2.6.md

@@ -29,8 +29,8 @@ cmake --build build --config Release
 Convert PyTorch model to gguf files (You can also download the converted [gguf](https://huggingface.co/openbmb/MiniCPM-o-2_6-gguf) by us)
 
 ```bash
-python ./tools/mtmd/minicpmv-surgery.py -m ../MiniCPM-o-2_6
-python ./tools/mtmd/minicpmv-convert-image-encoder-to-gguf.py -m ../MiniCPM-o-2_6 --minicpmv-projector ../MiniCPM-o-2_6/minicpmv.projector --output-dir ../MiniCPM-o-2_6/ --image-mean 0.5 0.5 0.5 --image-std 0.5 0.5 0.5 --minicpmv_version 4
+python ./tools/mtmd/legacy-models/minicpmv-surgery.py -m ../MiniCPM-o-2_6
+python ./tools/mtmd/legacy-models/minicpmv-convert-image-encoder-to-gguf.py -m ../MiniCPM-o-2_6 --minicpmv-projector ../MiniCPM-o-2_6/minicpmv.projector --output-dir ../MiniCPM-o-2_6/ --minicpmv_version 4
 python ./convert_hf_to_gguf.py ../MiniCPM-o-2_6/model
 
 # quantize int4 version

docs/multimodal/minicpmo4.0.md

@@ -0,0 +1,47 @@
+## MiniCPM-o 4
+
+### Prepare models and code
+
+Download [MiniCPM-o-4](https://huggingface.co/openbmb/MiniCPM-o-4) PyTorch model from huggingface to "MiniCPM-o-4" folder.
+
+
+### Build llama.cpp
+Readme modification time: 20250206
+
+If there are differences in usage, please refer to the official build [documentation](https://github.com/ggerganov/llama.cpp/blob/master/docs/build.md)
+
+Clone llama.cpp:
+```bash
+git clone https://github.com/ggerganov/llama.cpp
+cd llama.cpp
+```
+
+Build llama.cpp using `CMake`:
+```bash
+cmake -B build
+cmake --build build --config Release
+```
+
+
+### Usage of MiniCPM-o 4
+
+Convert PyTorch model to gguf files (You can also download the converted [gguf](https://huggingface.co/openbmb/MiniCPM-o-4-gguf) by us)
+
+```bash
+python ./tools/mtmd/legacy-models/minicpmv-surgery.py -m ../MiniCPM-o-4
+python ./tools/mtmd/legacy-models/minicpmv-convert-image-encoder-to-gguf.py -m ../MiniCPM-o-4 --minicpmv-projector ../MiniCPM-o-4/minicpmv.projector --output-dir ../MiniCPM-o-4/ --minicpmv_version 6
+python ./convert_hf_to_gguf.py ../MiniCPM-o-4/model
+
+# quantize int4 version
+./build/bin/llama-quantize ../MiniCPM-o-4/model/ggml-model-f16.gguf ../MiniCPM-o-4/model/ggml-model-Q4_K_M.gguf Q4_K_M
+```
+
+
+Inference on Linux or Mac
+```bash
+# run in single-turn mode
+./build/bin/llama-mtmd-cli -m ../MiniCPM-o-4/model/ggml-model-f16.gguf --mmproj ../MiniCPM-o-4/mmproj-model-f16.gguf -c 4096 --temp 0.7 --top-p 0.8 --top-k 100 --repeat-penalty 1.05 --image xx.jpg -p "What is in the image?"
+
+# run in conversation mode
+./build/bin/llama-mtmd-cli -m ../MiniCPM-o-4/model/ggml-model-Q4_K_M.gguf --mmproj ../MiniCPM-o-4/mmproj-model-f16.gguf
+```

docs/multimodal/minicpmv2.5.md

@@ -28,8 +28,8 @@ cmake --build build --config Release
 Convert PyTorch model to gguf files (You can also download the converted [gguf](https://huggingface.co/openbmb/MiniCPM-Llama3-V-2_5-gguf) by us)
 
 ```bash
-python ./tools/mtmd/minicpmv-surgery.py -m ../MiniCPM-Llama3-V-2_5
-python ./tools/mtmd/minicpmv-convert-image-encoder-to-gguf.py -m ../MiniCPM-Llama3-V-2_5 --minicpmv-projector ../MiniCPM-Llama3-V-2_5/minicpmv.projector --output-dir ../MiniCPM-Llama3-V-2_5/ --image-mean 0.5 0.5 0.5 --image-std 0.5 0.5 0.5 --minicpmv_version 2
+python ./tools/mtmd/legacy-models/minicpmv-surgery.py -m ../MiniCPM-Llama3-V-2_5
+python ./tools/mtmd/legacy-models/minicpmv-convert-image-encoder-to-gguf.py -m ../MiniCPM-Llama3-V-2_5 --minicpmv-projector ../MiniCPM-Llama3-V-2_5/minicpmv.projector --output-dir ../MiniCPM-Llama3-V-2_5/ --minicpmv_version 2
 python ./convert_hf_to_gguf.py ../MiniCPM-Llama3-V-2_5/model
 
 # quantize int4 version

docs/multimodal/minicpmv2.6.md

@@ -28,8 +28,8 @@ cmake --build build --config Release
 Convert PyTorch model to gguf files (You can also download the converted [gguf](https://huggingface.co/openbmb/MiniCPM-V-2_6-gguf) by us)
 
 ```bash
-python ./tools/mtmd/minicpmv-surgery.py -m ../MiniCPM-V-2_6
-python ./tools/mtmd/minicpmv-convert-image-encoder-to-gguf.py -m ../MiniCPM-V-2_6 --minicpmv-projector ../MiniCPM-V-2_6/minicpmv.projector --output-dir ../MiniCPM-V-2_6/ --image-mean 0.5 0.5 0.5 --image-std 0.5 0.5 0.5 --minicpmv_version 3
+python ./tools/mtmd/legacy-models/minicpmv-surgery.py -m ../MiniCPM-V-2_6
+python ./tools/mtmd/legacy-models/minicpmv-convert-image-encoder-to-gguf.py -m ../MiniCPM-V-2_6 --minicpmv-projector ../MiniCPM-V-2_6/minicpmv.projector --output-dir ../MiniCPM-V-2_6/ --minicpmv_version 3
 python ./convert_hf_to_gguf.py ../MiniCPM-V-2_6/model
 
 # quantize int4 version

docs/multimodal/minicpmv4.0.md

@@ -0,0 +1,47 @@
+## MiniCPM-V 4
+
+### Prepare models and code
+
+Download [MiniCPM-V-4](https://huggingface.co/openbmb/MiniCPM-V-4) PyTorch model from huggingface to "MiniCPM-V-4" folder.
+
+
+### Build llama.cpp
+Readme modification time: 20250206
+
+If there are differences in usage, please refer to the official build [documentation](https://github.com/ggerganov/llama.cpp/blob/master/docs/build.md)
+
+Clone llama.cpp:
+```bash
+git clone https://github.com/ggerganov/llama.cpp
+cd llama.cpp
+```
+
+Build llama.cpp using `CMake`:
+```bash
+cmake -B build
+cmake --build build --config Release
+```
+
+
+### Usage of MiniCPM-V 4
+
+Convert PyTorch model to gguf files (You can also download the converted [gguf](https://huggingface.co/openbmb/MiniCPM-V-4-gguf) by us)
+
+```bash
+python ./tools/mtmd/legacy-models/minicpmv-surgery.py -m ../MiniCPM-V-4
+python ./tools/mtmd/legacy-models/minicpmv-convert-image-encoder-to-gguf.py -m ../MiniCPM-V-4 --minicpmv-projector ../MiniCPM-V-4/minicpmv.projector --output-dir ../MiniCPM-V-4/ --minicpmv_version 5
+python ./convert_hf_to_gguf.py ../MiniCPM-V-4/model
+
+# quantize int4 version
+./build/bin/llama-quantize ../MiniCPM-V-4/model/ggml-model-f16.gguf ../MiniCPM-V-4/model/ggml-model-Q4_K_M.gguf Q4_K_M
+```
+
+
+Inference on Linux or Mac
+```bash
+# run in single-turn mode
+./build/bin/llama-mtmd-cli -m ../MiniCPM-V-4/model/ggml-model-f16.gguf --mmproj ../MiniCPM-V-4/mmproj-model-f16.gguf -c 4096 --temp 0.7 --top-p 0.8 --top-k 100 --repeat-penalty 1.05 --image xx.jpg -p "What is in the image?"
+
+# run in conversation mode
+./build/bin/llama-mtmd-cli -m ../MiniCPM-V-4/model/ggml-model-Q4_K_M.gguf --mmproj ../MiniCPM-V-4/mmproj-model-f16.gguf
+```

examples/speculative-simple/speculative-simple.cpp

@@ -65,7 +65,7 @@ int main(int argc, char ** argv) {
     ctx_dft   = llama_init_dft.context.get();
 
     if (!common_speculative_are_compatible(ctx_tgt, ctx_dft)) {
-        return 1;
+        LOG_INF("the draft model '%s' is not compatible with the target model '%s'. tokens will be translated between the draft and target models.\n", params.speculative.model.path.c_str(), params.model.path.c_str());
     }
 
     // Tokenize the prompt
@@ -130,7 +130,10 @@ int main(int argc, char ** argv) {
     params_spec.n_reuse = llama_n_ctx(ctx_dft) - n_draft;
     params_spec.p_min   = p_min;
 
-    struct common_speculative * spec = common_speculative_init(ctx_dft);
+    struct common_speculative * spec = common_speculative_init(ctx_tgt, ctx_dft);
+    for (auto &pair : params.speculative.replacements) {
+        common_speculative_add_replacement_tgt_dft(spec, pair.first.c_str(), pair.second.c_str());
+    }
 
     llama_batch batch_tgt = llama_batch_init(llama_n_batch(ctx_tgt), 0, 1);
 

ggml/src/ggml-vulkan/ggml-vulkan.cpp

@@ -5225,9 +5225,9 @@ static void ggml_vk_quantize_q8_1(ggml_backend_vk_context * ctx, vk_context& sub
 }
 
 static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) {
-    VK_LOG_DEBUG("ggml_vk_mul_mat_q_f16((" << src0 << ", name=" << src0->name << ", type=" << src0->type << ", ne0=" << src0->ne[0] << ", ne1=" << src0->ne[1] << ", ne2=" << src0->ne[2] << ", ne3=" << src0->ne[3] << ", nb0=" << src0->nb[0] << ", nb1=" << src0->nb[1] << ", nb2=" << src0->nb[2] << ", nb3=" << src0->nb[3];
-    std::cerr << "), (" << src1 << ", name=" << src1->name << ", type=" << src1->type << ", ne0=" << src1->ne[0] << ", ne1=" << src1->ne[1] << ", ne2=" << src1->ne[2] << ", ne3=" << src1->ne[3] << ", nb0=" << src1->nb[0] << ", nb1=" << src1->nb[1] << ", nb2=" << src1->nb[2] << ", nb3=" << src1->nb[3];
-    std::cerr << "), (" << dst << ", name=" << dst->name << ", type=" << dst->type << ", ne0=" << dst->ne[0] << ", ne1=" << dst->ne[1] << ", ne2=" << dst->ne[2] << ", ne3=" << dst->ne[3] << ", nb0=" << dst->nb[0] << ", nb1=" << dst->nb[1] << ", nb2=" << dst->nb[2] << ", nb3=" << dst->nb[3];
+    VK_LOG_DEBUG("ggml_vk_mul_mat_q_f16((" << src0 << ", name=" << src0->name << ", type=" << ggml_type_name(src0->type) << ", ne0=" << src0->ne[0] << ", ne1=" << src0->ne[1] << ", ne2=" << src0->ne[2] << ", ne3=" << src0->ne[3] << ", nb0=" << src0->nb[0] << ", nb1=" << src0->nb[1] << ", nb2=" << src0->nb[2] << ", nb3=" << src0->nb[3];
+    std::cerr << "), (" << src1 << ", name=" << src1->name << ", type=" << ggml_type_name(src1->type) << ", ne0=" << src1->ne[0] << ", ne1=" << src1->ne[1] << ", ne2=" << src1->ne[2] << ", ne3=" << src1->ne[3] << ", nb0=" << src1->nb[0] << ", nb1=" << src1->nb[1] << ", nb2=" << src1->nb[2] << ", nb3=" << src1->nb[3];
+    std::cerr << "), (" << dst << ", name=" << dst->name << ", type=" << ggml_type_name(dst->type) << ", ne0=" << dst->ne[0] << ", ne1=" << dst->ne[1] << ", ne2=" << dst->ne[2] << ", ne3=" << dst->ne[3] << ", nb0=" << dst->nb[0] << ", nb1=" << dst->nb[1] << ", nb2=" << dst->nb[2] << ", nb3=" << dst->nb[3];
     std::cerr << "), " << (dryrun ? "dryrun" : "") << ")");
     GGML_ASSERT(ggml_vk_dim01_contiguous(src0) || src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16 || src0->type == GGML_TYPE_BF16);  // NOLINT
     GGML_ASSERT(ggml_vk_dim01_contiguous(src1) || src1->type == GGML_TYPE_F32 || src1->type == GGML_TYPE_F16);  // NOLINT
@@ -11168,7 +11168,7 @@ size_t comp_nb[GGML_MAX_DIMS];
 size_t check_counter = 0;
 static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_cgraph * cgraph, int tensor_idx) {
     ggml_tensor * tensor = cgraph->nodes[tensor_idx];
-    if (tensor->op == GGML_OP_TRANSPOSE) {
+    if (tensor->op == GGML_OP_TRANSPOSE || tensor->op == GGML_OP_SET_ROWS) {
         return;
     }
 
@@ -11288,7 +11288,7 @@ static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_cgraph *
         tensor_clone = ggml_upscale_ext(ggml_ctx, src_clone[0], tensor->ne[0], tensor->ne[1], tensor->ne[2], tensor->ne[3], (ggml_scale_mode) tensor->op_params[0]);
     } else if (tensor->op == GGML_OP_SCALE) {
         const float * params = (const float *)tensor->op_params;
-        tensor_clone = ggml_scale(ggml_ctx, src_clone[0], params[0]);
+        tensor_clone = ggml_scale_bias(ggml_ctx, src_clone[0], params[0], params[1]);
     } else if (tensor->op == GGML_OP_SQR) {
         tensor_clone = ggml_sqr(ggml_ctx, src_clone[0]);
     } else if (tensor->op == GGML_OP_SIN) {
@@ -11399,8 +11399,6 @@ static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_cgraph *
         } else {
             tensor_clone = ggml_cpy(ggml_ctx, src_clone[0], src_clone[1]);
         }
-    } else if (tensor->op == GGML_OP_SET_ROWS) {
-        tensor_clone = ggml_set_rows(ggml_ctx, src_clone[0], src_clone[1]);
     } else if (tensor->op == GGML_OP_CONT) {
         tensor_clone = ggml_cont_4d(ggml_ctx, src_clone[0], tensor->ne[0], tensor->ne[1], tensor->ne[2], tensor->ne[3]);
     } else if (tensor->op == GGML_OP_RESHAPE) {
@@ -11508,7 +11506,7 @@ static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_cgraph *
 
 static void ggml_vk_check_results_1(ggml_backend_vk_context * ctx, ggml_cgraph * cgraph, int tensor_idx) {
     ggml_tensor * tensor = cgraph->nodes[tensor_idx];
-    if (tensor->op == GGML_OP_TRANSPOSE) {
+    if (tensor->op == GGML_OP_TRANSPOSE || tensor->op == GGML_OP_SET_ROWS) {
         return;
     }
     bool fused_rms_norm_mul = false;
@@ -11568,6 +11566,9 @@ static void ggml_vk_check_results_1(ggml_backend_vk_context * ctx, ggml_cgraph *
                         } else if (tensor->type == GGML_TYPE_F16) {
                             correct = ggml_fp16_to_fp32(*(ggml_fp16_t *) ((char *) comp_result + i3*comp_nb[3] + i2*comp_nb[2] + i1*comp_nb[1] + i0*comp_nb[0]));
                             result  = ggml_fp16_to_fp32(*(ggml_fp16_t *) ((char *) tensor_data + i3*tensor->nb[3] + i2*tensor->nb[2] + i1*tensor->nb[1] + i0*tensor->nb[0]));
+                        } else if (tensor->type == GGML_TYPE_BF16) {
+                            correct = ggml_bf16_to_fp32(*(ggml_bf16_t *) ((char *) comp_result + i3*comp_nb[3] + i2*comp_nb[2] + i1*comp_nb[1] + i0*comp_nb[0]));
+                            result  = ggml_bf16_to_fp32(*(ggml_bf16_t *) ((char *) tensor_data + i3*tensor->nb[3] + i2*tensor->nb[2] + i1*tensor->nb[1] + i0*tensor->nb[0]));
                         } else if (tensor->type == GGML_TYPE_I32) {
                             correct = *(int32_t *) ((char *) comp_result + i3*comp_nb[3] + i2*comp_nb[2] + i1*comp_nb[1] + i0*comp_nb[0]);
                             result  = *(int32_t *) ((char *) tensor_data + i3*tensor->nb[3] + i2*tensor->nb[2] + i1*tensor->nb[1] + i0*tensor->nb[0]);

gguf-py/gguf/tensor_mapping.py

@@ -616,6 +616,7 @@ class TensorNameMap:
         ),
 
         MODEL_TENSOR.SSM_DT_NORM: (
+            "model.layers.layers.{bid}.mixer.dt_norm.weight",  # plamo2
             "model.layers.{bid}.mamba.dt_layernorm",  # jamba
         ),
 
@@ -645,10 +646,6 @@ class TensorNameMap:
             "model.layers.layers.{bid}.mixer.D",  # plamo2
         ),
 
-        MODEL_TENSOR.SSM_DT_NORM: (
-            "model.layers.layers.{bid}.mixer.dt_norm.weight",  # plamo2
-        ),
-
         MODEL_TENSOR.SSM_NORM: (
             "model.layers.{bid}.mamba.norm", # falcon-h1 granite-hybrid
             "backbone.layers.{bid}.mixer.norm",  # mamba2

src/llama-graph.cpp

@@ -785,13 +785,20 @@ ggml_tensor * llm_graph_context::build_moe_ffn(
                 bool   scale_w,
                float   w_scale,
          llama_expert_gating_func_type gating_op,
-                 int   il) const {
+                 int   il,
+         ggml_tensor * probs_in) const {
     const int64_t n_embd   = cur->ne[0];
     const int64_t n_tokens = cur->ne[1];
     const bool weight_before_ffn = arch == LLM_ARCH_LLAMA4; // for llama4, we apply the sigmoid-ed weights before the FFN
 
-    ggml_tensor * logits = build_lora_mm(gate_inp, cur); // [n_expert, n_tokens]
-    cb(logits, "ffn_moe_logits", il);
+    ggml_tensor * logits = nullptr;
+
+    if (probs_in == nullptr) {
+        logits = build_lora_mm(gate_inp, cur); // [n_expert, n_tokens]
+        cb(logits, "ffn_moe_logits", il);
+    } else {
+        logits = probs_in;
+    }
 
     ggml_tensor * probs = nullptr;
     switch (gating_op) {
@@ -884,6 +891,14 @@ ggml_tensor * llm_graph_context::build_moe_ffn(
                 cur = ggml_gelu(ctx0, cur);
                 cb(cur, "ffn_moe_gelu", il);
             } break;
+        case LLM_FFN_RELU:
+            if (gate_exps) {
+                cur = ggml_reglu_split(ctx0, cur, up);
+                cb(cur, "ffn_moe_reglu", il);
+            } else {
+                cur = ggml_relu(ctx0, cur);
+                cb(cur, "ffn_moe_relu", il);
+            } break;
         default:
             GGML_ABORT("fatal error");
     }
@@ -927,100 +942,6 @@ ggml_tensor * llm_graph_context::build_moe_ffn(
     return moe_out;
 }
 
-ggml_tensor * llm_graph_context::build_moe_ffn_from_probs(
-         ggml_tensor * cur,
-         ggml_tensor * probs,
-         ggml_tensor * up_exps,
-         ggml_tensor * gate_exps,
-         ggml_tensor * down_exps,
-         ggml_tensor * exp_probs_b,
-             int64_t   n_expert,
-             int64_t   n_expert_used,
-             llama_expert_gating_func_type gating_op,
-                 int   il) const {
-    const int64_t n_embd   = cur->ne[0];
-    const int64_t n_tokens = cur->ne[1];
-
-    // add experts selection bias - introduced in DeepSeek V3
-    // leave probs unbiased as it's later used to get expert weights
-    ggml_tensor * selection_probs = probs;
-    if (exp_probs_b != nullptr) {
-        selection_probs = ggml_add(ctx0, probs, exp_probs_b);
-        cb(selection_probs, "ffn_moe_probs_biased", il);
-    }
-
-    // select experts
-    ggml_tensor * selected_experts = ggml_top_k(ctx0, selection_probs, n_expert_used); // [n_expert_used, n_tokens]
-    cb(selected_experts->src[0], "ffn_moe_argsort", il);
-    cb(selected_experts, "ffn_moe_topk", il);
-
-    ggml_tensor * weights = ggml_get_rows(ctx0,
-            ggml_reshape_3d(ctx0, probs, 1, n_expert, n_tokens), selected_experts); // [1, n_expert_used, n_tokens]
-    cb(weights, "ffn_moe_weights", il);
-
-    weights = ggml_reshape_2d(ctx0, weights, n_expert_used, n_tokens);
-     if (gating_op == LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX) {
-        weights = ggml_soft_max(ctx0, weights);
-    } else {
-        weights = ggml_sigmoid(ctx0, weights);
-        ggml_tensor * weights_sum = ggml_sum_rows(ctx0, weights); // [1, n_tokens]
-        cb(weights_sum, "ffn_moe_weights_sum", il);
-
-        weights = ggml_div(ctx0, weights, weights_sum); // [n_expert_used, n_tokens]
-        cb(weights, "ffn_moe_weights_norm", il);
-    }
-
-    weights = ggml_reshape_3d(ctx0, weights, 1, n_expert_used, n_tokens);
-
-    cur = ggml_reshape_3d(ctx0, cur, n_embd, 1, n_tokens);
-
-    ggml_tensor * up = build_lora_mm_id(up_exps, cur, selected_experts); // [n_ff, n_expert_used, n_tokens]
-    cb(up, "ffn_moe_up", il);
-
-    ggml_tensor * experts = nullptr;
-    cur = build_lora_mm_id(gate_exps, cur, selected_experts); // [n_ff, n_expert_used, n_tokens]
-    cb(cur, "ffn_moe_gate", il);
-
-    cur = ggml_reglu_split(ctx0, cur, up);
-    cb(cur, "ffn_moe_reglu", il);
-
-    experts = build_lora_mm_id(down_exps, cur, selected_experts); // [n_embd, n_expert_used, n_tokens]
-    cb(experts, "ffn_moe_down", il);
-
-    experts = ggml_mul(ctx0, experts, weights);
-    cb(cur, "ffn_moe_weighted", il);
-
-    ggml_tensor * cur_experts[LLAMA_MAX_EXPERTS] = { nullptr };
-
-    assert(n_expert_used > 0);
-
-    // order the views before the adds
-    for (uint32_t i = 0; i < hparams.n_expert_used; ++i) {
-        cur_experts[i] = ggml_view_2d(ctx0, experts, n_embd, n_tokens, experts->nb[2], i*experts->nb[1]);
-
-        ggml_build_forward_expand(gf, cur_experts[i]);
-    }
-
-    // aggregate experts
-    // note: here we explicitly use hparams.n_expert_used instead of n_expert_used
-    //       to avoid potentially a large number of add nodes during warmup
-    //       ref: https://github.com/ggml-org/llama.cpp/pull/14753
-    ggml_tensor * moe_out = cur_experts[0];
-
-    for (uint32_t i = 1; i < hparams.n_expert_used; ++i) {
-        moe_out = ggml_add(ctx0, moe_out, cur_experts[i]);
-    }
-
-    if (n_expert_used == 1) {
-        // avoid returning a non-contiguous tensor
-        moe_out = ggml_cont(ctx0, moe_out);
-    }
-
-    cb(moe_out, "ffn_moe_out", il);
-
-    return moe_out;
-}
-
 // input embeddings with optional lora
 ggml_tensor * llm_graph_context::build_inp_embd(ggml_tensor * tok_embd) const {
     const int64_t n_embd = hparams.n_embd;

src/llama-graph.h

@@ -631,19 +631,8 @@ struct llm_graph_context {
                     bool   scale_w,
                    float   w_scale,
             llama_expert_gating_func_type gating_op,
-                     int   il) const;
-
-    ggml_tensor * build_moe_ffn_from_probs(
-             ggml_tensor * cur,
-             ggml_tensor * probs,
-             ggml_tensor * up_exps,
-             ggml_tensor * gate_exps,
-             ggml_tensor * down_exps,
-             ggml_tensor * exp_probs_b,
-                 int64_t   n_expert,
-                 int64_t   n_expert_used,
-            llama_expert_gating_func_type gating_op,
-                     int   il) const;
+                     int   il,
+             ggml_tensor * probs_in = nullptr) const;
 
     //
     // inputs

src/llama-model.cpp

@@ -17320,10 +17320,18 @@ struct llm_build_smallthinker : public llm_graph_context{
             cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
             cb(cur, "ffn_norm", il);
 
-            ggml_tensor * ffn_out = build_moe_ffn_from_probs(cur, probs, model.layers[il].ffn_up_exps,
-                                                model.layers[il].ffn_gate_exps, model.layers[il].ffn_down_exps,
-                                                nullptr, n_expert, n_expert_used,
-                                                static_cast<llama_expert_gating_func_type>(hparams.expert_gating_func), il);
+            ggml_tensor * ffn_out =
+                build_moe_ffn(cur,
+                        nullptr,
+                        model.layers[il].ffn_up_exps,
+                        model.layers[il].ffn_gate_exps,
+                        model.layers[il].ffn_down_exps,
+                        nullptr,
+                        n_expert, n_expert_used,
+                        LLM_FFN_RELU, true,
+                        false, 0.0,
+                        static_cast<llama_expert_gating_func_type>(hparams.expert_gating_func),
+                        il, probs);
 
             cb(ffn_out, "ffn_out", il);
             cur = ffn_out;

tools/mtmd/clip.cpp

@@ -868,10 +868,16 @@ struct clip_graph {
             int n_head = n_embd/d_head;
             int num_query = 96;
             if (ctx->model.hparams.minicpmv_version == 2) {
+                // MiniCPM-V 2.5
                 num_query = 96;
             } else if (ctx->model.hparams.minicpmv_version == 3) {
+                // MiniCPM-V 2.6
                 num_query = 64;
             } else if (ctx->model.hparams.minicpmv_version == 4) {
+                // MiniCPM-o 2.6
+                num_query = 64;
+            } else if (ctx->model.hparams.minicpmv_version == 5) {
+                // MiniCPM-V 4.0
                 num_query = 64;
             }
 
@@ -3551,10 +3557,16 @@ int clip_n_output_tokens(const struct clip_ctx * ctx, struct clip_image_f32 * im
         case PROJECTOR_TYPE_MINICPMV:
             {
                 if (params.minicpmv_version == 2) {
+                    // MiniCPM-V 2.5
                     n_patches_sq = 96;
                 } else if (params.minicpmv_version == 3) {
+                    // MiniCPM-V 2.6
                     n_patches_sq = 64;
                 } else if (params.minicpmv_version == 4) {
+                    // MiniCPM-o 2.6
+                    n_patches_sq = 64;
+                } else if (params.minicpmv_version == 5) {
+                    // MiniCPM-V 4.0
                     n_patches_sq = 64;
                 } else {
                     GGML_ABORT("Unknown minicpmv version");
@@ -4103,11 +4115,17 @@ int clip_n_mmproj_embd(const struct clip_ctx * ctx) {
             return ctx->model.mm_3_b->ne[0];
         case PROJECTOR_TYPE_MINICPMV:
             if (hparams.minicpmv_version == 2) {
+                // MiniCPM-V 2.5
                 return 4096;
             } else if (hparams.minicpmv_version == 3) {
+                // MiniCPM-V 2.6
                 return 3584;
             } else if (hparams.minicpmv_version == 4) {
+                // MiniCPM-o 2.6
                 return 3584;
+            } else if (hparams.minicpmv_version == 5) {
+                // MiniCPM-V 4.0
+                return 2560;
             }
             GGML_ABORT("Unknown minicpmv version");
         case PROJECTOR_TYPE_GLM_EDGE:

tools/mtmd/legacy-models/minicpmv-convert-image-encoder-to-gguf.py

@@ -497,11 +497,11 @@ ap.add_argument("--projector-type", help="Type of projector. Possible values: ml
 ap.add_argument("-o", "--output-dir", help="Directory to save GGUF files. Default is the original model directory", default=None)
 # Example --image_mean 0.48145466 0.4578275 0.40821073 --image_std 0.26862954 0.26130258 0.27577711
 # Example --image_mean 0.5 0.5 0.5 --image_std 0.5 0.5 0.5
-default_image_mean = [0.48145466, 0.4578275, 0.40821073]
-default_image_std = [0.26862954, 0.26130258, 0.27577711]
+default_image_mean = [0.5, 0.5, 0.5]
+default_image_std = [0.5, 0.5, 0.5]
 ap.add_argument('--image-mean', type=float, nargs='+', help='Mean of the images for normalization (overrides processor) ', default=None)
 ap.add_argument('--image-std', type=float, nargs='+', help='Standard deviation of the images for normalization (overrides processor)', default=None)
-ap.add_argument('--minicpmv_version', type=int, help='minicpmv_version: MiniCPM-V-2 use 1; MiniCPM-V-2.5 use 2; MiniCPM-V-2.6 use 3; MiniCPM-o-2.6 use 4', default=2)
+ap.add_argument('--minicpmv_version', type=int, help='minicpmv_version: MiniCPM-V-2 use 1; MiniCPM-V-2.5 use 2; MiniCPM-V-2.6 use 3; MiniCPM-o-2.6 use 4; MiniCPM-V 4.0 use 5; MiniCPM-o-4.0 use 6', default=2)
 
 # with proper
 args = ap.parse_args()
@@ -517,6 +517,17 @@ if args.use_f32:
 # output in the same directory as the model if output_dir is None
 dir_model = args.model_dir
 
+# If minicpmv_projector is not specified but the default path exists, use the default path
+if args.minicpmv_projector is None:
+    default_projector_path = os.path.join(dir_model, "minicpmv.projector")
+    if os.path.isfile(default_projector_path):
+        args.minicpmv_projector = default_projector_path
+        print(f"Found default projector file: {default_projector_path}")
+
+# If output_dir is not specified, use model_dir as the default value
+if args.output_dir is None:
+    args.output_dir = dir_model
+
 if args.clip_model_is_vision or not os.path.exists(dir_model + "/vocab.json") or args.clip_model_is_openclip:
     vocab = None
     tokens = None
@@ -546,18 +557,21 @@ if args.use_f32:
 minicpmv_version = args.minicpmv_version
 emb_dim = 4096
 block_count = 26
-if minicpmv_version == 1:
+if minicpmv_version == 1:  # MiniCPM-V 2.0
     emb_dim = 2304
     block_count = 26
-elif minicpmv_version == 2:
+elif minicpmv_version == 2:  # MiniCPM-V 2.5
     emb_dim = 4096
     block_count = 27
-elif minicpmv_version == 3:
+elif minicpmv_version == 3:  # MiniCPM-V 2.6
     emb_dim = 3584
     block_count = 27
-elif minicpmv_version == 4:
+elif minicpmv_version == 4:  # MiniCPM-o 2.6
     emb_dim = 3584
     block_count = 27
+elif minicpmv_version == 5:  # MiniCPM-V 4.0
+    emb_dim = 2560
+    block_count = 27
 
 default_vision_config = {
         "hidden_size": 1152,
@@ -577,6 +591,10 @@ if minicpmv_version == 3:
 elif minicpmv_version == 4:
     vision_config = SiglipVisionConfig(**default_vision_config)
     model = SiglipVisionTransformer(vision_config)
+elif minicpmv_version == 5:
+    default_vision_config["model_type"] = "siglip_vision_model"
+    vision_config = SiglipVisionConfig(**default_vision_config)
+    model = SiglipVisionTransformer(vision_config)
 
 processor = None
 # if model.attn_pool is not None:
@@ -603,7 +621,7 @@ elif args.vision_only:
 else:
     fname_middle = ""
 
-output_dir = args.output_dir if args.output_dir is not None else dir_model
+output_dir = args.output_dir
 os.makedirs(output_dir, exist_ok=True)
 output_prefix = os.path.basename(output_dir).replace("ggml_", "")
 fname_out = os.path.join(output_dir, f"{fname_middle}model-{ftype_str[ftype]}.gguf")

tools/mtmd/mtmd.cpp

@@ -207,7 +207,7 @@ struct mtmd_context {
             tok_row_end_trail = false; // no trailing end-of-row token
             ov_img_first      = true;
 
-        } else if (minicpmv_version == 3 || minicpmv_version == 4) {
+        } else if (minicpmv_version == 3 || minicpmv_version == 4 || minicpmv_version == 5) {
             // minicpmv 2.6 format:
             // <image> (overview) </image><slice> (slice) </slice><slice> (slice) </slice>\n ...
             slice_tmpl        = MTMD_SLICE_TMPL_MINICPMV_2_6;

tools/server/README.md

@@ -469,7 +469,7 @@ These words will not be included in the completion, so make sure to add them to
 
 `ignore_eos`: Ignore end of stream token and continue generating.  Default: `false`
 
-`logit_bias`: Modify the likelihood of a token appearing in the generated text completion. For example, use `"logit_bias": [[15043,1.0]]` to increase the likelihood of the token 'Hello', or `"logit_bias": [[15043,-1.0]]` to decrease its likelihood. Setting the value to false, `"logit_bias": [[15043,false]]` ensures that the token `Hello` is never produced. The tokens can also be represented as strings, e.g. `[["Hello, World!",-0.5]]` will reduce the likelihood of all the individual tokens that represent the string `Hello, World!`, just like the `presence_penalty` does. Default: `[]`
+`logit_bias`: Modify the likelihood of a token appearing in the generated text completion. For example, use `"logit_bias": [[15043,1.0]]` to increase the likelihood of the token 'Hello', or `"logit_bias": [[15043,-1.0]]` to decrease its likelihood. Setting the value to false, `"logit_bias": [[15043,false]]` ensures that the token `Hello` is never produced. The tokens can also be represented as strings, e.g. `[["Hello, World!",-0.5]]` will reduce the likelihood of all the individual tokens that represent the string `Hello, World!`, just like the `presence_penalty` does. For compatibility with the OpenAI API, a JSON object {"<string or token id>": bias, ...} can also be passed. Default: `[]`
 
 `n_probs`: If greater than 0, the response also contains the probabilities of top N tokens for each generated token given the sampling settings. Note that for temperature < 0 the tokens are sampled greedily but token probabilities are still being calculated via a simple softmax of the logits without considering any other sampler settings. Default: `0`
 

tools/server/server.cpp

@@ -473,6 +473,33 @@ struct server_task {
                         }
                     }
                 }
+           } else if (logit_bias != data.end() && logit_bias->is_object()) {
+                const int n_vocab = llama_vocab_n_tokens(vocab);
+                for (const auto & el : logit_bias->items()) {
+                    float bias;
+                    const auto & key = el.key();
+                    const auto & value = el.value();
+                    if (value.is_number()) {
+                        bias = value.get<float>();
+                    } else if (value.is_boolean() && !value.get<bool>()) {
+                        bias = -INFINITY;
+                    } else {
+                        continue;
+                    }
+
+                    char *end;
+                    llama_token tok = strtol(key.c_str(), &end, 10);
+                    if (*end == 0) {
+                        if (tok >= 0 && tok < n_vocab) {
+                            params.sampling.logit_bias.push_back({tok, bias});
+                        }
+                    } else {
+                        auto toks = common_tokenize(vocab, key, false);
+                        for (auto tok : toks) {
+                            params.sampling.logit_bias.push_back({tok, bias});
+                        }
+                    }
+                }
             }
 
             params.sampling.ignore_eos = json_value(data, "ignore_eos", params_base.sampling.ignore_eos);
@@ -1902,6 +1929,7 @@ struct server_context {
     mtmd_context * mctx = nullptr;
 
     const llama_vocab * vocab = nullptr;
+    bool vocab_dft_compatible = true;
 
     llama_model * model_dft = nullptr;
 
@@ -1992,10 +2020,9 @@ struct server_context {
                 return false;
             }
 
-            if (!common_speculative_are_compatible(ctx, llama_init_dft.context.get())) {
-                SRV_ERR("the draft model '%s' is not compatible with the target model '%s'\n", params_base.speculative.model.path.c_str(), params_base.model.path.c_str());
-
-                return false;
+            vocab_dft_compatible = common_speculative_are_compatible(ctx, llama_init_dft.context.get());
+            if (!vocab_dft_compatible) {
+                SRV_INF("the draft model '%s' is not compatible with the target model '%s'. tokens will be translated between the draft and target models.\n", params_base.speculative.model.path.c_str(), params_base.model.path.c_str());
             }
 
             const int n_ctx_dft = llama_n_ctx(llama_init_dft.context.get());
@@ -2085,11 +2112,14 @@ struct server_context {
                     return;
                 }
 
-                slot.spec = common_speculative_init(slot.ctx_dft);
+                slot.spec = common_speculative_init(slot.ctx, slot.ctx_dft);
                 if (slot.spec == nullptr) {
                     SRV_ERR("%s", "failed to create speculator\n");
                     return;
                 }
+                for (auto &pair : params_base.speculative.replacements) {
+                    common_speculative_add_replacement_tgt_dft(slot.spec, pair.first.c_str(), pair.second.c_str());
+                }
             }
 
             SLT_INF(slot, "new slot n_ctx_slot = %d\n", slot.n_ctx);

tools/server/tests/unit/test_chat_completion.py

@@ -351,3 +351,32 @@ def test_logprobs_stream():
                     assert token.top_logprobs is not None
                     assert len(token.top_logprobs) > 0
     assert aggregated_text == output_text
+
+
+def test_logit_bias():
+    global server
+    server.start()
+
+    exclude = ["i", "I", "the", "The", "to", "a", "an", "be", "is", "was", "but", "But", "and", "And", "so", "So", "you", "You", "he", "He", "she", "She", "we", "We", "they", "They", "it", "It", "his", "His", "her", "Her", "book", "Book"]
+
+    res = server.make_request("POST", "/tokenize", data={
+        "content": " " + " ".join(exclude) + " ",
+    })
+    assert res.status_code == 200
+    tokens = res.body["tokens"]
+    logit_bias = {tok: -100 for tok in tokens}
+
+    client = OpenAI(api_key="dummy", base_url=f"http://{server.server_host}:{server.server_port}/v1")
+    res = client.chat.completions.create(
+        model="gpt-3.5-turbo-instruct",
+        temperature=0.0,
+        messages=[
+            {"role": "system", "content": "Book"},
+            {"role": "user", "content": "What is the best book"},
+        ],
+        max_tokens=64,
+        logit_bias=logit_bias
+    )
+    output_text = res.choices[0].message.content
+    assert output_text
+    assert all(output_text.find(" " + tok + " ") == -1 for tok in exclude)

tools/server/tests/unit/test_completion.py

@@ -444,6 +444,39 @@ def test_n_probs_post_sampling():
         assert any(prob["prob"] == 1.0 for prob in tok["top_probs"])
 
 
+@pytest.mark.parametrize("tokenize,openai_style", [(False, False), (False, True), (True, False), (True, True)])
+def test_logit_bias(tokenize, openai_style):
+    global server
+    server.start()
+
+    exclude = ["i", "I", "the", "The", "to", "a", "an", "be", "is", "was", "but", "But", "and", "And", "so", "So", "you", "You", "he", "He", "she", "She", "we", "We", "they", "They", "it", "It", "his", "His", "her", "Her", "book", "Book"]
+
+    logit_bias = []
+    if tokenize:
+        res = server.make_request("POST", "/tokenize", data={
+            "content": " " + " ".join(exclude) + " ",
+        })
+        assert res.status_code == 200
+        tokens = res.body["tokens"]
+        logit_bias = [[tok, -100] for tok in tokens]
+
+    else:
+        logit_bias = [[" " + tok + " ", -100] for tok in exclude]
+
+    if openai_style:
+        logit_bias = {el[0]: -100 for el in logit_bias}
+
+    res = server.make_request("POST", "/completion", data={
+        "n_predict": 64,
+        "prompt": "What is the best book",
+        "logit_bias": logit_bias,
+        "temperature": 0.0
+    })
+    assert res.status_code == 200
+    output_text = res.body["content"]
+    assert all(output_text.find(" " + tok + " ") == -1 for tok in exclude)
+
+
 def test_cancel_request():
     global server
     server.n_ctx = 4096