graph : fix equal_seq() check (#14986)

ggml-ci

.devops/cann.Dockerfile

@@ -0,0 +1,130 @@
+# ==============================================================================
+# ARGUMENTS
+# ==============================================================================
+
+# Define the CANN base image for easier version updates later
+ARG CANN_BASE_IMAGE=quay.io/ascend/cann:8.1.rc1-910b-openeuler22.03-py3.10
+
+# ==============================================================================
+# BUILD STAGE
+# Compile all binary files and libraries
+# ==============================================================================
+FROM ${CANN_BASE_IMAGE} AS build
+
+# Define the Ascend chip model for compilation. Default is Ascend910B3
+ARG ASCEND_SOC_TYPE=Ascend910B3
+
+# -- Install build dependencies --
+RUN yum install -y gcc g++ cmake make git libcurl-devel python3 python3-pip && \
+    yum clean all && \
+    rm -rf /var/cache/yum
+
+# -- Set the working directory --
+WORKDIR /app
+
+# -- Copy project files --
+COPY . .
+
+# -- Set CANN environment variables (required for compilation) --
+# Using ENV instead of `source` allows environment variables to persist across the entire image layer
+ENV ASCEND_TOOLKIT_HOME=/usr/local/Ascend/ascend-toolkit/latest
+ENV LD_LIBRARY_PATH=${ASCEND_TOOLKIT_HOME}/lib64:${LD_LIBRARY_PATH}
+ENV PATH=${ASCEND_TOOLKIT_HOME}/bin:${PATH}
+ENV ASCEND_OPP_PATH=${ASCEND_TOOLKIT_HOME}/opp
+ENV LD_LIBRARY_PATH=${ASCEND_TOOLKIT_HOME}/runtime/lib64/stub:$LD_LIBRARY_PATH
+# ... You can add other environment variables from the original file as needed ...
+# For brevity, only core variables are listed here. You can paste the original ENV list here.
+
+# -- Build llama.cpp --
+# Use the passed ASCEND_SOC_TYPE argument and add general build options
+RUN source /usr/local/Ascend/ascend-toolkit/set_env.sh --force \
+    && \
+    cmake -B build \
+        -DGGML_CANN=ON \
+        -DCMAKE_BUILD_TYPE=Release \
+        -DSOC_TYPE=${ASCEND_SOC_TYPE} \
+        . && \
+    cmake --build build --config Release -j$(nproc)
+
+# -- Organize build artifacts for copying in later stages --
+# Create a lib directory to store all .so files
+RUN mkdir -p /app/lib && \
+    find build -name "*.so" -exec cp {} /app/lib \;
+
+# Create a full directory to store all executables and Python scripts
+RUN mkdir -p /app/full && \
+    cp build/bin/* /app/full/ && \
+    cp *.py /app/full/ && \
+    cp -r gguf-py /app/full/ && \
+    cp -r requirements /app/full/ && \
+    cp requirements.txt /app/full/
+    # If you have a tools.sh script, make sure it is copied here
+    # cp .devops/tools.sh /app/full/tools.sh
+
+# ==============================================================================
+# BASE STAGE
+# Create a minimal base image with CANN runtime and common libraries
+# ==============================================================================
+FROM ${CANN_BASE_IMAGE} AS base
+
+# -- Install runtime dependencies --
+RUN yum install -y libgomp curl && \
+    yum clean all && \
+    rm -rf /var/cache/yum
+
+# -- Set CANN environment variables (required for runtime) --
+ENV ASCEND_TOOLKIT_HOME=/usr/local/Ascend/ascend-toolkit/latest
+ENV LD_LIBRARY_PATH=/app:${ASCEND_TOOLKIT_HOME}/lib64:${LD_LIBRARY_PATH}
+ENV PATH=${ASCEND_TOOLKIT_HOME}/bin:${PATH}
+ENV ASCEND_OPP_PATH=${ASCEND_TOOLKIT_HOME}/opp
+# ... You can add other environment variables from the original file as needed ...
+
+WORKDIR /app
+
+# Copy compiled .so files from the build stage
+COPY --from=build /app/lib/ /app
+
+# ==============================================================================
+# FINAL STAGES (TARGETS)
+# ==============================================================================
+
+### Target: full
+# Complete image with all tools, Python bindings, and dependencies
+# ==============================================================================
+FROM base AS full
+
+COPY --from=build /app/full /app
+
+# Install Python dependencies
+RUN yum install -y git python3 python3-pip && \
+    pip3 install --no-cache-dir --upgrade pip setuptools wheel && \
+    pip3 install --no-cache-dir -r requirements.txt && \
+    yum clean all && \
+    rm -rf /var/cache/yum
+
+# You need to provide a tools.sh script as the entrypoint
+ENTRYPOINT ["/app/tools.sh"]
+# If there is no tools.sh, you can set the default to start the server
+# ENTRYPOINT ["/app/llama-server"]
+
+### Target: light
+# Lightweight image containing only llama-cli
+# ==============================================================================
+FROM base AS light
+
+COPY --from=build /app/full/llama-cli /app
+
+ENTRYPOINT [ "/app/llama-cli" ]
+
+### Target: server
+# Dedicated server image containing only llama-server
+# ==============================================================================
+FROM base AS server
+
+ENV LLAMA_ARG_HOST=0.0.0.0
+
+COPY --from=build /app/full/llama-server /app
+
+HEALTHCHECK --interval=5m CMD [ "curl", "-f", "http://localhost:8080/health" ]
+
+ENTRYPOINT [ "/app/llama-server" ]

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()) {
@@ -2091,6 +2095,13 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             params.no_kv_offload = true;
         }
     ).set_env("LLAMA_ARG_NO_KV_OFFLOAD"));
+    add_opt(common_arg(
+        {"-nr", "--no-repack"},
+        "disable weight repacking",
+        [](common_params & params) {
+            params.no_extra_bufts = true;
+        }
+    ).set_env("LLAMA_ARG_NO_REPACK"));
     add_opt(common_arg(
         {"-ctk", "--cache-type-k"}, "TYPE",
         string_format(
@@ -2369,6 +2380,15 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             }
         }
     ));
+    add_opt(common_arg(
+        {"--cpu-moe"},
+        "use CPU for Mixture of Experts (MoE) weights",
+        [](common_params & params) {
+            params.tensor_buft_overrides.push_back({"\\.ffn_up_exps\\.weight$",   ggml_backend_cpu_buffer_type()});
+            params.tensor_buft_overrides.push_back({"\\.ffn_down_exps\\.weight$", ggml_backend_cpu_buffer_type()});
+            params.tensor_buft_overrides.push_back({"\\.ffn_gate_exps\\.weight$", ggml_backend_cpu_buffer_type()});
+        }
+    ).set_env("LLAMA_ARG_CPU_MOE"));
     add_opt(common_arg(
         {"-ngl", "--gpu-layers", "--n-gpu-layers"}, "N",
         "number of layers to store in VRAM",
@@ -3249,6 +3269,13 @@ 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.cpp

@@ -1122,6 +1122,7 @@ struct llama_model_params common_model_params_to_llama(common_params & params) {
     mparams.use_mmap        = params.use_mmap;
     mparams.use_mlock       = params.use_mlock;
     mparams.check_tensors   = params.check_tensors;
+    mparams.use_extra_bufts = !params.no_extra_bufts;
 
     if (params.kv_overrides.empty()) {
         mparams.kv_overrides = NULL;

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
@@ -224,7 +225,7 @@ struct common_params_diffusion {
     bool    visual_mode   = false;
 
     float   eps           = 0;        // epsilon for timesteps
-    int32_t block_length  = 32;       // block length for generation
+    int32_t block_length  = 0;        // block length for generation
 
     int32_t algorithm     = 4;        // default algorithm: low-confidence
     float   alg_temp      = 0.0f;     // algorithm temperature
@@ -358,6 +359,7 @@ struct common_params {
     bool warmup            = true;  // warmup run
     bool check_tensors     = false; // validate tensor data
     bool no_op_offload     = false; // globally disable offload host tensor operations to device
+    bool no_extra_bufts    = false; // disable extra buffer types (used for weight repacking)
 
     bool single_turn       = false; // single turn chat conversation
 

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-cann/ggml-cann.cpp

@@ -2016,6 +2016,9 @@ static bool ggml_backend_cann_cpy_tensor_async(
         (ggml_backend_cann_context*)backend_dst->context;
 
     size_t copy_size = ggml_nbytes(dst);
+    if (copy_size == 0) {
+        return true;
+    }
     if (backend_src != backend_dst) {
         ggml_backend_cann_buffer_context* buf_ctx_src =
             (ggml_backend_cann_buffer_context*)buf_src->context;

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

include/llama.h

@@ -284,10 +284,11 @@ extern "C" {
         const struct llama_model_kv_override * kv_overrides;
 
         // Keep the booleans together to avoid misalignment during copy-by-value.
-        bool vocab_only;    // only load the vocabulary, no weights
-        bool use_mmap;      // use mmap if possible
-        bool use_mlock;     // force system to keep model in RAM
-        bool check_tensors; // validate model tensor data
+        bool vocab_only;      // only load the vocabulary, no weights
+        bool use_mmap;        // use mmap if possible
+        bool use_mlock;       // force system to keep model in RAM
+        bool check_tensors;   // validate model tensor data
+        bool use_extra_bufts; // use extra buffer types (used for weight repacking)
     };
 
     // NOTE: changing the default values of parameters marked as [EXPERIMENTAL] may cause crashes or incorrect results in certain configurations

scripts/compare-commits.sh

@@ -1,19 +1,41 @@
 #!/usr/bin/env bash
 
 if [ $# -lt 2 ]; then
-    echo "usage: ./scripts/compare-commits.sh <commit1> <commit2> [additional llama-bench arguments]"
+    echo "usage: ./scripts/compare-commits.sh <commit1> <commit2> [tool] [additional arguments]"
+    echo "  tool: 'llama-bench' (default) or 'test-backend-ops'"
+    echo "  additional arguments: passed to the selected tool"
     exit 1
 fi
 
 set -e
 set -x
 
+# Parse arguments
+commit1=$1
+commit2=$2
+tool=${3:-llama-bench}
+additional_args="${@:4}"
+
+# Validate tool argument
+if [ "$tool" != "llama-bench" ] && [ "$tool" != "test-backend-ops" ]; then
+    echo "Error: tool must be 'llama-bench' or 'test-backend-ops'"
+    exit 1
+fi
+
 # verify at the start that the compare script has all the necessary dependencies installed
 ./scripts/compare-llama-bench.py --check
 
-bench_args="${@:3}"
+if [ "$tool" = "llama-bench" ]; then
+    db_file="llama-bench.sqlite"
+    target="llama-bench"
+    run_args="-o sql -oe md $additional_args"
+else  # test-backend-ops
+    db_file="test-backend-ops.sqlite"
+    target="test-backend-ops"
+    run_args="perf --output sql $additional_args"
+fi
 
-rm -f llama-bench.sqlite > /dev/null
+rm -f "$db_file" > /dev/null
 
 # to test a backend, call the script with the corresponding environment variable (e.g. GGML_CUDA=1 ./scripts/compare-commits.sh ...)
 if [ -n "$GGML_CUDA" ]; then
@@ -25,14 +47,14 @@ dir="build-bench"
 function run {
     rm -fr ${dir} > /dev/null
     cmake -B ${dir} -S . ${CMAKE_OPTS} > /dev/null
-    cmake --build ${dir} -t llama-bench > /dev/null
-    ${dir}/bin/llama-bench -o sql -oe md $bench_args | sqlite3 llama-bench.sqlite
+    cmake --build ${dir} -t $target -j $(nproc) > /dev/null
+    ${dir}/bin/$target $run_args | sqlite3 "$db_file"
 }
 
-git checkout $1 > /dev/null
+git checkout $commit1 > /dev/null
 run
 
-git checkout $2 > /dev/null
+git checkout $commit2 > /dev/null
 run
 
-./scripts/compare-llama-bench.py -b $1 -c $2
+./scripts/compare-llama-bench.py -b $commit1 -c $commit2 --tool $tool -i "$db_file"

scripts/compare-llama-bench.py

@@ -1,16 +1,16 @@
 #!/usr/bin/env python3
 
-import logging
 import argparse
-import heapq
-import sys
-import os
-from glob import glob
-import sqlite3
-import json
 import csv
-from typing import Optional, Union
+import heapq
+import json
+import logging
+import os
+import sqlite3
+import sys
 from collections.abc import Iterator, Sequence
+from glob import glob
+from typing import Any, Optional, Union
 
 try:
     import git
@@ -23,7 +23,7 @@ except ImportError as e:
 logger = logging.getLogger("compare-llama-bench")
 
 # All llama-bench SQL fields
-DB_FIELDS = [
+LLAMA_BENCH_DB_FIELDS = [
     "build_commit", "build_number", "cpu_info",       "gpu_info",   "backends",     "model_filename",
     "model_type",   "model_size",   "model_n_params", "n_batch",    "n_ubatch",     "n_threads",
     "cpu_mask",     "cpu_strict",   "poll",           "type_k",     "type_v",       "n_gpu_layers",
@@ -33,7 +33,7 @@ DB_FIELDS = [
     "test_time",    "avg_ns",       "stddev_ns",      "avg_ts",     "stddev_ts",
 ]
 
-DB_TYPES = [
+LLAMA_BENCH_DB_TYPES = [
     "TEXT",    "INTEGER", "TEXT",    "TEXT",    "TEXT",    "TEXT",
     "TEXT",    "INTEGER", "INTEGER", "INTEGER", "INTEGER", "INTEGER",
     "TEXT",    "INTEGER", "INTEGER", "TEXT",    "TEXT",    "INTEGER",
@@ -42,20 +42,41 @@ DB_TYPES = [
     "INTEGER", "INTEGER", "INTEGER", "INTEGER", "INTEGER", "INTEGER",
     "TEXT",    "INTEGER", "INTEGER", "REAL",    "REAL",
 ]
-assert len(DB_FIELDS) == len(DB_TYPES)
 
-# Properties by which to differentiate results per commit:
-KEY_PROPERTIES = [
+# All test-backend-ops SQL fields
+TEST_BACKEND_OPS_DB_FIELDS = [
+    "test_time", "build_commit", "backend_name",  "op_name", "op_params", "test_mode",
+    "supported", "passed",       "error_message", "time_us", "flops",     "bandwidth_gb_s",
+    "memory_kb", "n_runs"
+]
+
+TEST_BACKEND_OPS_DB_TYPES = [
+    "TEXT",    "TEXT",    "TEXT", "TEXT", "TEXT", "TEXT",
+    "INTEGER", "INTEGER", "TEXT", "REAL", "REAL", "REAL",
+    "INTEGER", "INTEGER"
+]
+
+assert len(LLAMA_BENCH_DB_FIELDS) == len(LLAMA_BENCH_DB_TYPES)
+assert len(TEST_BACKEND_OPS_DB_FIELDS) == len(TEST_BACKEND_OPS_DB_TYPES)
+
+# Properties by which to differentiate results per commit for llama-bench:
+LLAMA_BENCH_KEY_PROPERTIES = [
     "cpu_info", "gpu_info", "backends", "n_gpu_layers", "tensor_buft_overrides", "model_filename", "model_type",
     "n_batch", "n_ubatch", "embeddings", "cpu_mask", "cpu_strict", "poll", "n_threads", "type_k", "type_v",
     "use_mmap", "no_kv_offload", "split_mode", "main_gpu", "tensor_split", "flash_attn", "n_prompt", "n_gen", "n_depth"
 ]
 
-# Properties that are boolean and are converted to Yes/No for the table:
-BOOL_PROPERTIES = ["embeddings", "cpu_strict", "use_mmap", "no_kv_offload", "flash_attn"]
+# Properties by which to differentiate results per commit for test-backend-ops:
+TEST_BACKEND_OPS_KEY_PROPERTIES = [
+    "backend_name", "op_name", "op_params", "test_mode"
+]
 
-# Header names for the table:
-PRETTY_NAMES = {
+# Properties that are boolean and are converted to Yes/No for the table:
+LLAMA_BENCH_BOOL_PROPERTIES = ["embeddings", "cpu_strict", "use_mmap", "no_kv_offload", "flash_attn"]
+TEST_BACKEND_OPS_BOOL_PROPERTIES = ["supported", "passed"]
+
+# Header names for the table (llama-bench):
+LLAMA_BENCH_PRETTY_NAMES = {
     "cpu_info": "CPU", "gpu_info": "GPU", "backends": "Backends", "n_gpu_layers": "GPU layers",
     "tensor_buft_overrides": "Tensor overrides", "model_filename": "File", "model_type": "Model", "model_size": "Model size [GiB]",
     "model_n_params": "Num. of par.", "n_batch": "Batch size", "n_ubatch": "Microbatch size", "embeddings": "Embeddings",
@@ -64,21 +85,42 @@ PRETTY_NAMES = {
     "flash_attn": "FlashAttention",
 }
 
-DEFAULT_SHOW = ["model_type"]  # Always show these properties by default.
-DEFAULT_HIDE = ["model_filename"]  # Always hide these properties by default.
+# Header names for the table (test-backend-ops):
+TEST_BACKEND_OPS_PRETTY_NAMES = {
+    "backend_name": "Backend", "op_name": "GGML op", "op_params": "Op parameters", "test_mode": "Mode",
+    "supported": "Supported", "passed": "Passed", "error_message": "Error",
+    "flops": "FLOPS", "bandwidth_gb_s": "Bandwidth (GB/s)", "memory_kb": "Memory (KB)", "n_runs": "Runs"
+}
+
+DEFAULT_SHOW_LLAMA_BENCH = ["model_type"]  # Always show these properties by default.
+DEFAULT_HIDE_LLAMA_BENCH = ["model_filename"]  # Always hide these properties by default.
+
+DEFAULT_SHOW_TEST_BACKEND_OPS = ["backend_name", "op_name"]  # Always show these properties by default.
+DEFAULT_HIDE_TEST_BACKEND_OPS = ["error_message"]  # Always hide these properties by default.
+
 GPU_NAME_STRIP = ["NVIDIA GeForce ", "Tesla ", "AMD Radeon "]  # Strip prefixes for smaller tables.
 MODEL_SUFFIX_REPLACE = {" - Small": "_S", " - Medium": "_M", " - Large": "_L"}
 
-DESCRIPTION = """Creates tables from llama-bench data written to multiple JSON/CSV files, a single JSONL file or SQLite database. Example usage (Linux):
+DESCRIPTION = """Creates tables from llama-bench or test-backend-ops data written to multiple JSON/CSV files, a single JSONL file or SQLite database. Example usage (Linux):
 
+For llama-bench:
 $ git checkout master
-$ make clean && make llama-bench
+$ cmake -B ${BUILD_DIR} ${CMAKE_OPTS} && cmake --build ${BUILD_DIR} -t llama-bench -j $(nproc)
 $ ./llama-bench -o sql | sqlite3 llama-bench.sqlite
 $ git checkout some_branch
-$ make clean && make llama-bench
+$ cmake -B ${BUILD_DIR} ${CMAKE_OPTS} && cmake --build ${BUILD_DIR} -t llama-bench -j $(nproc)
 $ ./llama-bench -o sql | sqlite3 llama-bench.sqlite
 $ ./scripts/compare-llama-bench.py
 
+For test-backend-ops:
+$ git checkout master
+$ cmake -B ${BUILD_DIR} ${CMAKE_OPTS} && cmake --build ${BUILD_DIR} -t test-backend-ops -j $(nproc)
+$ ./test-backend-ops perf --output sql | sqlite3 test-backend-ops.sqlite
+$ git checkout some_branch
+$ cmake -B ${BUILD_DIR} ${CMAKE_OPTS} && cmake --build ${BUILD_DIR} -t test-backend-ops -j $(nproc)
+$ ./test-backend-ops perf --output sql | sqlite3 test-backend-ops.sqlite
+$ ./scripts/compare-llama-bench.py --tool test-backend-ops -i test-backend-ops.sqlite
+
 Performance numbers from multiple runs per commit are averaged WITHOUT being weighted by the --repetitions parameter of llama-bench.
 """
 
@@ -96,6 +138,13 @@ help_c = (
     "Defaults to the non-master commit for which llama-bench was run most recently."
 )
 parser.add_argument("-c", "--compare", help=help_c)
+help_t = (
+    "The tool whose data is being compared. "
+    "Either 'llama-bench' or 'test-backend-ops'. "
+    "This determines the database schema and comparison logic used. "
+    "If left unspecified, try to determine from the input file."
+)
+parser.add_argument("-t", "--tool", help=help_t, default=None, choices=[None, "llama-bench", "test-backend-ops"])
 help_i = (
     "JSON/JSONL/SQLite/CSV files for comparing commits. "
     "Specify multiple times to use multiple input files (JSON/CSV only). "
@@ -114,7 +163,8 @@ parser.add_argument("-o", "--output", help=help_o, default="pipe")
 help_s = (
     "Columns to add to the table. "
     "Accepts a comma-separated list of values. "
-    f"Legal values: {', '.join(KEY_PROPERTIES[:-3])}. "
+    f"Legal values for test-backend-ops: {', '.join(TEST_BACKEND_OPS_KEY_PROPERTIES)}. "
+    f"Legal values for llama-bench: {', '.join(LLAMA_BENCH_KEY_PROPERTIES[:-3])}. "
     "Defaults to model name (model_type) and CPU and/or GPU name (cpu_info, gpu_info) "
     "plus any column where not all data points are the same. "
     "If the columns are manually specified, then the results for each unique combination of the "
@@ -142,8 +192,14 @@ if unknown_args:
     sys.exit(1)
 
 input_file = known_args.input
-if not input_file and os.path.exists("./llama-bench.sqlite"):
-    input_file = ["llama-bench.sqlite"]
+tool = known_args.tool
+
+if not input_file:
+    if tool == "llama-bench" and os.path.exists("./llama-bench.sqlite"):
+        input_file = ["llama-bench.sqlite"]
+    elif tool == "test-backend-ops" and os.path.exists("./test-backend-ops.sqlite"):
+        input_file = ["test-backend-ops.sqlite"]
+
 if not input_file:
     sqlite_files = glob("*.sqlite")
     if len(sqlite_files) == 1:
@@ -161,14 +217,23 @@ class LlamaBenchData:
     build_len_max: int
     build_len: int = 8
     builds: list[str] = []
-    check_keys = set(KEY_PROPERTIES + ["build_commit", "test_time", "avg_ts"])
+    tool: str = "llama-bench"  # Tool type: "llama-bench" or "test-backend-ops"
 
-    def __init__(self):
+    def __init__(self, tool: str = "llama-bench"):
+        self.tool = tool
         try:
             self.repo = git.Repo(".", search_parent_directories=True)
         except git.InvalidGitRepositoryError:
             self.repo = None
 
+        # Set schema-specific properties based on tool
+        if self.tool == "llama-bench":
+            self.check_keys = set(LLAMA_BENCH_KEY_PROPERTIES + ["build_commit", "test_time", "avg_ts"])
+        elif self.tool == "test-backend-ops":
+            self.check_keys = set(TEST_BACKEND_OPS_KEY_PROPERTIES + ["build_commit", "test_time"])
+        else:
+            assert False
+
     def _builds_init(self):
         self.build_len = self.build_len_min
 
@@ -252,52 +317,121 @@ class LlamaBenchData:
 class LlamaBenchDataSQLite3(LlamaBenchData):
     connection: sqlite3.Connection
     cursor: sqlite3.Cursor
+    table_name: str
 
-    def __init__(self):
-        super().__init__()
+    def __init__(self, tool: str = "llama-bench"):
+        super().__init__(tool)
         self.connection = sqlite3.connect(":memory:")
         self.cursor = self.connection.cursor()
-        self.cursor.execute(f"CREATE TABLE test({', '.join(' '.join(x) for x in zip(DB_FIELDS, DB_TYPES))});")
+
+        # Set table name and schema based on tool
+        if self.tool == "llama-bench":
+            self.table_name = "test"
+            db_fields = LLAMA_BENCH_DB_FIELDS
+            db_types = LLAMA_BENCH_DB_TYPES
+        elif self.tool == "test-backend-ops":
+            self.table_name = "test_backend_ops"
+            db_fields = TEST_BACKEND_OPS_DB_FIELDS
+            db_types = TEST_BACKEND_OPS_DB_TYPES
+        else:
+            assert False
+
+        self.cursor.execute(f"CREATE TABLE {self.table_name}({', '.join(' '.join(x) for x in zip(db_fields, db_types))});")
 
     def _builds_init(self):
         if self.connection:
-            self.build_len_min = self.cursor.execute("SELECT MIN(LENGTH(build_commit)) from test;").fetchone()[0]
-            self.build_len_max = self.cursor.execute("SELECT MAX(LENGTH(build_commit)) from test;").fetchone()[0]
+            self.build_len_min = self.cursor.execute(f"SELECT MIN(LENGTH(build_commit)) from {self.table_name};").fetchone()[0]
+            self.build_len_max = self.cursor.execute(f"SELECT MAX(LENGTH(build_commit)) from {self.table_name};").fetchone()[0]
 
             if self.build_len_min != self.build_len_max:
                 logger.warning("Data contains commit hashes of differing lengths. It's possible that the wrong commits will be compared. "
                                "Try purging the the database of old commits.")
-                self.cursor.execute(f"UPDATE test SET build_commit = SUBSTRING(build_commit, 1, {self.build_len_min});")
+                self.cursor.execute(f"UPDATE {self.table_name} SET build_commit = SUBSTRING(build_commit, 1, {self.build_len_min});")
 
-            builds = self.cursor.execute("SELECT DISTINCT build_commit FROM test;").fetchall()
+            builds = self.cursor.execute(f"SELECT DISTINCT build_commit FROM {self.table_name};").fetchall()
             self.builds = list(map(lambda b: b[0], builds))  # list[tuple[str]] -> list[str]
         super()._builds_init()
 
     def builds_timestamp(self, reverse: bool = False) -> Union[Iterator[tuple], Sequence[tuple]]:
         data = self.cursor.execute(
-            "SELECT build_commit, test_time FROM test ORDER BY test_time;").fetchall()
+            f"SELECT build_commit, test_time FROM {self.table_name} ORDER BY test_time;").fetchall()
         return reversed(data) if reverse else data
 
     def get_rows(self, properties: list[str], hexsha8_baseline: str, hexsha8_compare: str) -> Sequence[tuple]:
+        if self.tool == "llama-bench":
+            return self._get_rows_llama_bench(properties, hexsha8_baseline, hexsha8_compare)
+        elif self.tool == "test-backend-ops":
+            return self._get_rows_test_backend_ops(properties, hexsha8_baseline, hexsha8_compare)
+        else:
+            assert False
+
+    def _get_rows_llama_bench(self, properties: list[str], hexsha8_baseline: str, hexsha8_compare: str) -> Sequence[tuple]:
         select_string = ", ".join(
             [f"tb.{p}" for p in properties] + ["tb.n_prompt", "tb.n_gen", "tb.n_depth", "AVG(tb.avg_ts)", "AVG(tc.avg_ts)"])
         equal_string = " AND ".join(
-            [f"tb.{p} = tc.{p}" for p in KEY_PROPERTIES] + [
+            [f"tb.{p} = tc.{p}" for p in LLAMA_BENCH_KEY_PROPERTIES] + [
                 f"tb.build_commit = '{hexsha8_baseline}'", f"tc.build_commit = '{hexsha8_compare}'"]
         )
         group_order_string = ", ".join([f"tb.{p}" for p in properties] + ["tb.n_gen", "tb.n_prompt", "tb.n_depth"])
-        query = (f"SELECT {select_string} FROM test tb JOIN test tc ON {equal_string} "
+        query = (f"SELECT {select_string} FROM {self.table_name} tb JOIN {self.table_name} tc ON {equal_string} "
+                 f"GROUP BY {group_order_string} ORDER BY {group_order_string};")
+        return self.cursor.execute(query).fetchall()
+
+    def _get_rows_test_backend_ops(self, properties: list[str], hexsha8_baseline: str, hexsha8_compare: str) -> Sequence[tuple]:
+        # For test-backend-ops, we compare FLOPS and bandwidth metrics (prioritizing FLOPS over bandwidth)
+        select_string = ", ".join(
+            [f"tb.{p}" for p in properties] + [
+                "AVG(tb.flops)", "AVG(tc.flops)",
+                "AVG(tb.bandwidth_gb_s)", "AVG(tc.bandwidth_gb_s)"
+            ])
+        equal_string = " AND ".join(
+            [f"tb.{p} = tc.{p}" for p in TEST_BACKEND_OPS_KEY_PROPERTIES] + [
+                f"tb.build_commit = '{hexsha8_baseline}'", f"tc.build_commit = '{hexsha8_compare}'",
+                "tb.supported = 1", "tc.supported = 1", "tb.passed = 1", "tc.passed = 1"]  # Only compare successful tests
+        )
+        group_order_string = ", ".join([f"tb.{p}" for p in properties])
+        query = (f"SELECT {select_string} FROM {self.table_name} tb JOIN {self.table_name} tc ON {equal_string} "
                  f"GROUP BY {group_order_string} ORDER BY {group_order_string};")
         return self.cursor.execute(query).fetchall()
 
 
 class LlamaBenchDataSQLite3File(LlamaBenchDataSQLite3):
-    def __init__(self, data_file: str):
-        super().__init__()
+    def __init__(self, data_file: str, tool: Any):
+        super().__init__(tool)
 
         self.connection.close()
         self.connection = sqlite3.connect(data_file)
         self.cursor = self.connection.cursor()
+
+        # Check which table exists in the database
+        tables = self.cursor.execute("SELECT name FROM sqlite_master WHERE type='table';").fetchall()
+        table_names = [table[0] for table in tables]
+
+        # Tool selection logic
+        if tool is None:
+            if "test" in table_names:
+                self.table_name = "test"
+                self.tool = "llama-bench"
+            elif "test_backend_ops" in table_names:
+                self.table_name = "test_backend_ops"
+                self.tool = "test-backend-ops"
+            else:
+                raise RuntimeError(f"No suitable table found in database. Available tables: {table_names}")
+        elif tool == "llama-bench":
+            if "test" in table_names:
+                self.table_name = "test"
+                self.tool = "llama-bench"
+            else:
+                raise RuntimeError(f"Table 'test' not found for tool 'llama-bench'. Available tables: {table_names}")
+        elif tool == "test-backend-ops":
+            if "test_backend_ops" in table_names:
+                self.table_name = "test_backend_ops"
+                self.tool = "test-backend-ops"
+            else:
+                raise RuntimeError(f"Table 'test_backend_ops' not found for tool 'test-backend-ops'. Available tables: {table_names}")
+        else:
+            raise RuntimeError(f"Unknown tool: {tool}")
+
         self._builds_init()
 
     @staticmethod
@@ -317,20 +451,23 @@ class LlamaBenchDataSQLite3File(LlamaBenchDataSQLite3):
 
 
 class LlamaBenchDataJSONL(LlamaBenchDataSQLite3):
-    def __init__(self, data_file: str):
-        super().__init__()
+    def __init__(self, data_file: str, tool: str = "llama-bench"):
+        super().__init__(tool)
+
+        # Get the appropriate field list based on tool
+        db_fields = LLAMA_BENCH_DB_FIELDS if tool == "llama-bench" else TEST_BACKEND_OPS_DB_FIELDS
 
         with open(data_file, "r", encoding="utf-8") as fp:
             for i, line in enumerate(fp):
                 parsed = json.loads(line)
 
-                for k in parsed.keys() - set(DB_FIELDS):
+                for k in parsed.keys() - set(db_fields):
                     del parsed[k]
 
                 if (missing_keys := self._check_keys(parsed.keys())):
                     raise RuntimeError(f"Missing required data key(s) at line {i + 1}: {', '.join(missing_keys)}")
 
-                self.cursor.execute(f"INSERT INTO test({', '.join(parsed.keys())}) VALUES({', '.join('?' * len(parsed))});", tuple(parsed.values()))
+                self.cursor.execute(f"INSERT INTO {self.table_name}({', '.join(parsed.keys())}) VALUES({', '.join('?' * len(parsed))});", tuple(parsed.values()))
 
         self._builds_init()
 
@@ -349,21 +486,24 @@ class LlamaBenchDataJSONL(LlamaBenchDataSQLite3):
 
 
 class LlamaBenchDataJSON(LlamaBenchDataSQLite3):
-    def __init__(self, data_files: list[str]):
-        super().__init__()
+    def __init__(self, data_files: list[str], tool: str = "llama-bench"):
+        super().__init__(tool)
+
+        # Get the appropriate field list based on tool
+        db_fields = LLAMA_BENCH_DB_FIELDS if tool == "llama-bench" else TEST_BACKEND_OPS_DB_FIELDS
 
         for data_file in data_files:
             with open(data_file, "r", encoding="utf-8") as fp:
                 parsed = json.load(fp)
 
                 for i, entry in enumerate(parsed):
-                    for k in entry.keys() - set(DB_FIELDS):
+                    for k in entry.keys() - set(db_fields):
                         del entry[k]
 
                     if (missing_keys := self._check_keys(entry.keys())):
                         raise RuntimeError(f"Missing required data key(s) at entry {i + 1}: {', '.join(missing_keys)}")
 
-                    self.cursor.execute(f"INSERT INTO test({', '.join(entry.keys())}) VALUES({', '.join('?' * len(entry))});", tuple(entry.values()))
+                    self.cursor.execute(f"INSERT INTO {self.table_name}({', '.join(entry.keys())}) VALUES({', '.join('?' * len(entry))});", tuple(entry.values()))
 
         self._builds_init()
 
@@ -384,21 +524,24 @@ class LlamaBenchDataJSON(LlamaBenchDataSQLite3):
 
 
 class LlamaBenchDataCSV(LlamaBenchDataSQLite3):
-    def __init__(self, data_files: list[str]):
-        super().__init__()
+    def __init__(self, data_files: list[str], tool: str = "llama-bench"):
+        super().__init__(tool)
+
+        # Get the appropriate field list based on tool
+        db_fields = LLAMA_BENCH_DB_FIELDS if tool == "llama-bench" else TEST_BACKEND_OPS_DB_FIELDS
 
         for data_file in data_files:
             with open(data_file, "r", encoding="utf-8") as fp:
                 for i, parsed in enumerate(csv.DictReader(fp)):
                     keys = set(parsed.keys())
 
-                    for k in keys - set(DB_FIELDS):
+                    for k in keys - set(db_fields):
                         del parsed[k]
 
                     if (missing_keys := self._check_keys(keys)):
                         raise RuntimeError(f"Missing required data key(s) at line {i + 1}: {', '.join(missing_keys)}")
 
-                    self.cursor.execute(f"INSERT INTO test({', '.join(parsed.keys())}) VALUES({', '.join('?' * len(parsed))});", tuple(parsed.values()))
+                    self.cursor.execute(f"INSERT INTO {self.table_name}({', '.join(parsed.keys())}) VALUES({', '.join('?' * len(parsed))});", tuple(parsed.values()))
 
         self._builds_init()
 
@@ -419,21 +562,90 @@ class LlamaBenchDataCSV(LlamaBenchDataSQLite3):
         return True
 
 
+def format_flops(flops_value: float) -> str:
+    """Format FLOPS values with appropriate units for better readability."""
+    if flops_value == 0:
+        return "0.00"
+
+    # Define unit thresholds and names
+    units = [
+        (1e12, "T"),   # TeraFLOPS
+        (1e9, "G"),    # GigaFLOPS
+        (1e6, "M"),    # MegaFLOPS
+        (1e3, "k"),    # kiloFLOPS
+        (1, "")        # FLOPS
+    ]
+
+    for threshold, unit in units:
+        if abs(flops_value) >= threshold:
+            formatted_value = flops_value / threshold
+            if formatted_value >= 100:
+                return f"{formatted_value:.1f}{unit}"
+            else:
+                return f"{formatted_value:.2f}{unit}"
+
+    # Fallback for very small values
+    return f"{flops_value:.2f}"
+
+
+def format_flops_for_table(flops_value: float, target_unit: str) -> str:
+    """Format FLOPS values for table display without unit suffix (since unit is in header)."""
+    if flops_value == 0:
+        return "0.00"
+
+    # Define unit thresholds based on target unit
+    unit_divisors = {
+        "TFLOPS": 1e12,
+        "GFLOPS": 1e9,
+        "MFLOPS": 1e6,
+        "kFLOPS": 1e3,
+        "FLOPS": 1
+    }
+
+    divisor = unit_divisors.get(target_unit, 1)
+    formatted_value = flops_value / divisor
+
+    if formatted_value >= 100:
+        return f"{formatted_value:.1f}"
+    else:
+        return f"{formatted_value:.2f}"
+
+
+def get_flops_unit_name(flops_values: list) -> str:
+    """Determine the best FLOPS unit name based on the magnitude of values."""
+    if not flops_values or all(v == 0 for v in flops_values):
+        return "FLOPS"
+
+    # Find the maximum absolute value to determine appropriate unit
+    max_flops = max(abs(v) for v in flops_values if v != 0)
+
+    if max_flops >= 1e12:
+        return "TFLOPS"
+    elif max_flops >= 1e9:
+        return "GFLOPS"
+    elif max_flops >= 1e6:
+        return "MFLOPS"
+    elif max_flops >= 1e3:
+        return "kFLOPS"
+    else:
+        return "FLOPS"
+
+
 bench_data = None
 if len(input_file) == 1:
     if LlamaBenchDataSQLite3File.valid_format(input_file[0]):
-        bench_data = LlamaBenchDataSQLite3File(input_file[0])
+        bench_data = LlamaBenchDataSQLite3File(input_file[0], tool)
     elif LlamaBenchDataJSON.valid_format(input_file):
-        bench_data = LlamaBenchDataJSON(input_file)
+        bench_data = LlamaBenchDataJSON(input_file, tool)
     elif LlamaBenchDataJSONL.valid_format(input_file[0]):
-        bench_data = LlamaBenchDataJSONL(input_file[0])
+        bench_data = LlamaBenchDataJSONL(input_file[0], tool)
     elif LlamaBenchDataCSV.valid_format(input_file):
-        bench_data = LlamaBenchDataCSV(input_file)
+        bench_data = LlamaBenchDataCSV(input_file, tool)
 else:
     if LlamaBenchDataJSON.valid_format(input_file):
-        bench_data = LlamaBenchDataJSON(input_file)
+        bench_data = LlamaBenchDataJSON(input_file, tool)
     elif LlamaBenchDataCSV.valid_format(input_file):
-        bench_data = LlamaBenchDataCSV(input_file)
+        bench_data = LlamaBenchDataCSV(input_file, tool)
 
 if not bench_data:
     raise RuntimeError("No valid (or some invalid) input files found.")
@@ -504,12 +716,29 @@ else:
 
 name_compare = bench_data.get_commit_name(hexsha8_compare)
 
+# Get tool-specific configuration
+if tool == "llama-bench":
+    key_properties = LLAMA_BENCH_KEY_PROPERTIES
+    bool_properties = LLAMA_BENCH_BOOL_PROPERTIES
+    pretty_names = LLAMA_BENCH_PRETTY_NAMES
+    default_show = DEFAULT_SHOW_LLAMA_BENCH
+    default_hide = DEFAULT_HIDE_LLAMA_BENCH
+elif tool == "test-backend-ops":
+    key_properties = TEST_BACKEND_OPS_KEY_PROPERTIES
+    bool_properties = TEST_BACKEND_OPS_BOOL_PROPERTIES
+    pretty_names = TEST_BACKEND_OPS_PRETTY_NAMES
+    default_show = DEFAULT_SHOW_TEST_BACKEND_OPS
+    default_hide = DEFAULT_HIDE_TEST_BACKEND_OPS
+else:
+    assert False
+
 # If the user provided columns to group the results by, use them:
 if known_args.show is not None:
     show = known_args.show.split(",")
     unknown_cols = []
     for prop in show:
-        if prop not in KEY_PROPERTIES[:-3]:  # Last three values are n_prompt, n_gen, n_depth.
+        valid_props = key_properties if tool == "test-backend-ops" else key_properties[:-3]  # Exclude n_prompt, n_gen, n_depth for llama-bench
+        if prop not in valid_props:
             unknown_cols.append(prop)
     if unknown_cols:
         logger.error(f"Unknown values for --show: {', '.join(unknown_cols)}")
@@ -518,32 +747,54 @@ if known_args.show is not None:
     rows_show = bench_data.get_rows(show, hexsha8_baseline, hexsha8_compare)
 # Otherwise, select those columns where the values are not all the same:
 else:
-    rows_full = bench_data.get_rows(KEY_PROPERTIES, hexsha8_baseline, hexsha8_compare)
+    rows_full = bench_data.get_rows(key_properties, hexsha8_baseline, hexsha8_compare)
     properties_different = []
-    for i, kp_i in enumerate(KEY_PROPERTIES):
-        if kp_i in DEFAULT_SHOW or kp_i in ["n_prompt", "n_gen", "n_depth"]:
-            continue
-        for row_full in rows_full:
-            if row_full[i] != rows_full[0][i]:
-                properties_different.append(kp_i)
-                break
+
+    if tool == "llama-bench":
+        # For llama-bench, skip n_prompt, n_gen, n_depth from differentiation logic
+        check_properties = [kp for kp in key_properties if kp not in ["n_prompt", "n_gen", "n_depth"]]
+        for i, kp_i in enumerate(key_properties):
+            if kp_i in default_show or kp_i in ["n_prompt", "n_gen", "n_depth"]:
+                continue
+            for row_full in rows_full:
+                if row_full[i] != rows_full[0][i]:
+                    properties_different.append(kp_i)
+                    break
+    elif tool == "test-backend-ops":
+        # For test-backend-ops, check all key properties
+        for i, kp_i in enumerate(key_properties):
+            if kp_i in default_show:
+                continue
+            for row_full in rows_full:
+                if row_full[i] != rows_full[0][i]:
+                    properties_different.append(kp_i)
+                    break
+    else:
+        assert False
 
     show = []
-    # Show CPU and/or GPU by default even if the hardware for all results is the same:
-    if rows_full and "n_gpu_layers" not in properties_different:
-        ngl = int(rows_full[0][KEY_PROPERTIES.index("n_gpu_layers")])
 
-        if ngl != 99 and "cpu_info" not in properties_different:
-            show.append("cpu_info")
+    if tool == "llama-bench":
+        # Show CPU and/or GPU by default even if the hardware for all results is the same:
+        if rows_full and "n_gpu_layers" not in properties_different:
+            ngl = int(rows_full[0][key_properties.index("n_gpu_layers")])
 
-    show += properties_different
+            if ngl != 99 and "cpu_info" not in properties_different:
+                show.append("cpu_info")
 
-    index_default = 0
-    for prop in ["cpu_info", "gpu_info", "n_gpu_layers", "main_gpu"]:
-        if prop in show:
-            index_default += 1
-    show = show[:index_default] + DEFAULT_SHOW + show[index_default:]
-    for prop in DEFAULT_HIDE:
+        show += properties_different
+
+        index_default = 0
+        for prop in ["cpu_info", "gpu_info", "n_gpu_layers", "main_gpu"]:
+            if prop in show:
+                index_default += 1
+        show = show[:index_default] + default_show + show[index_default:]
+    elif tool == "test-backend-ops":
+        show = default_show + properties_different
+    else:
+        assert False
+
+    for prop in default_hide:
         try:
             show.remove(prop)
         except ValueError:
@@ -551,7 +802,7 @@ else:
 
     # Add plot_x parameter to parameters to show if it's not already present:
     if known_args.plot:
-        for k, v in PRETTY_NAMES.items():
+        for k, v in pretty_names.items():
             if v == known_args.plot_x and k not in show:
                 show.append(k)
                 break
@@ -563,60 +814,120 @@ if not rows_show:
     sys.exit(1)
 
 table = []
-for row in rows_show:
-    n_prompt = int(row[-5])
-    n_gen    = int(row[-4])
-    n_depth  = int(row[-3])
-    if n_prompt != 0 and n_gen == 0:
-        test_name = f"pp{n_prompt}"
-    elif n_prompt == 0 and n_gen != 0:
-        test_name = f"tg{n_gen}"
-    else:
-        test_name = f"pp{n_prompt}+tg{n_gen}"
-    if n_depth != 0:
-        test_name = f"{test_name}@d{n_depth}"
-    #           Regular columns    test name    avg t/s values              Speedup
-    #            VVVVVVVVVVVVV     VVVVVVVVV    VVVVVVVVVVVVVV              VVVVVVV
-    table.append(list(row[:-5]) + [test_name] + list(row[-2:]) + [float(row[-1]) / float(row[-2])])
+primary_metric = "FLOPS"  # Default to FLOPS for test-backend-ops
+
+if tool == "llama-bench":
+    # For llama-bench, create test names and compare avg_ts values
+    for row in rows_show:
+        n_prompt = int(row[-5])
+        n_gen    = int(row[-4])
+        n_depth  = int(row[-3])
+        if n_prompt != 0 and n_gen == 0:
+            test_name = f"pp{n_prompt}"
+        elif n_prompt == 0 and n_gen != 0:
+            test_name = f"tg{n_gen}"
+        else:
+            test_name = f"pp{n_prompt}+tg{n_gen}"
+        if n_depth != 0:
+            test_name = f"{test_name}@d{n_depth}"
+        #           Regular columns    test name    avg t/s values              Speedup
+        #            VVVVVVVVVVVVV     VVVVVVVVV    VVVVVVVVVVVVVV              VVVVVVV
+        table.append(list(row[:-5]) + [test_name] + list(row[-2:]) + [float(row[-1]) / float(row[-2])])
+elif tool == "test-backend-ops":
+    # Determine the primary metric by checking rows until we find one with valid data
+    if rows_show:
+        primary_metric = "FLOPS"  # Default to FLOPS
+        flops_values = []
+
+        # Collect all FLOPS values to determine the best unit
+        for sample_row in rows_show:
+            baseline_flops = float(sample_row[-4])
+            compare_flops = float(sample_row[-3])
+            baseline_bandwidth = float(sample_row[-2])
+
+            if baseline_flops > 0:
+                flops_values.extend([baseline_flops, compare_flops])
+            elif baseline_bandwidth > 0 and not flops_values:
+                primary_metric = "Bandwidth (GB/s)"
+
+        # If we have FLOPS data, determine the appropriate unit
+        if flops_values:
+            primary_metric = get_flops_unit_name(flops_values)
+
+    # For test-backend-ops, prioritize FLOPS > bandwidth for comparison
+    for row in rows_show:
+        # Extract metrics: flops, bandwidth_gb_s (baseline and compare)
+        baseline_flops = float(row[-4])
+        compare_flops = float(row[-3])
+        baseline_bandwidth = float(row[-2])
+        compare_bandwidth = float(row[-1])
+
+        # Determine which metric to use for comparison (prioritize FLOPS > bandwidth)
+        if baseline_flops > 0 and compare_flops > 0:
+            # Use FLOPS comparison (higher is better)
+            speedup = compare_flops / baseline_flops
+            baseline_str = format_flops_for_table(baseline_flops, primary_metric)
+            compare_str = format_flops_for_table(compare_flops, primary_metric)
+        elif baseline_bandwidth > 0 and compare_bandwidth > 0:
+            # Use bandwidth comparison (higher is better)
+            speedup = compare_bandwidth / baseline_bandwidth
+            baseline_str = f"{baseline_bandwidth:.2f}"
+            compare_str = f"{compare_bandwidth:.2f}"
+        else:
+            # Fallback if no valid data is available
+            baseline_str = "N/A"
+            compare_str = "N/A"
+            from math import nan
+            speedup = nan
+
+        table.append(list(row[:-4]) + [baseline_str, compare_str, speedup])
+else:
+    assert False
 
 # Some a-posteriori fixes to make the table contents prettier:
-for bool_property in BOOL_PROPERTIES:
+for bool_property in bool_properties:
     if bool_property in show:
         ip = show.index(bool_property)
         for row_table in table:
             row_table[ip] = "Yes" if int(row_table[ip]) == 1 else "No"
 
-if "model_type" in show:
-    ip = show.index("model_type")
-    for (old, new) in MODEL_SUFFIX_REPLACE.items():
+if tool == "llama-bench":
+    if "model_type" in show:
+        ip = show.index("model_type")
+        for (old, new) in MODEL_SUFFIX_REPLACE.items():
+            for row_table in table:
+                row_table[ip] = row_table[ip].replace(old, new)
+
+    if "model_size" in show:
+        ip = show.index("model_size")
         for row_table in table:
-            row_table[ip] = row_table[ip].replace(old, new)
+            row_table[ip] = float(row_table[ip]) / 1024 ** 3
 
-if "model_size" in show:
-    ip = show.index("model_size")
-    for row_table in table:
-        row_table[ip] = float(row_table[ip]) / 1024 ** 3
+    if "gpu_info" in show:
+        ip = show.index("gpu_info")
+        for row_table in table:
+            for gns in GPU_NAME_STRIP:
+                row_table[ip] = row_table[ip].replace(gns, "")
 
-if "gpu_info" in show:
-    ip = show.index("gpu_info")
-    for row_table in table:
-        for gns in GPU_NAME_STRIP:
-            row_table[ip] = row_table[ip].replace(gns, "")
+            gpu_names = row_table[ip].split(", ")
+            num_gpus = len(gpu_names)
+            all_names_the_same = len(set(gpu_names)) == 1
+            if len(gpu_names) >= 2 and all_names_the_same:
+                row_table[ip] = f"{num_gpus}x {gpu_names[0]}"
 
-        gpu_names = row_table[ip].split(", ")
-        num_gpus = len(gpu_names)
-        all_names_the_same = len(set(gpu_names)) == 1
-        if len(gpu_names) >= 2 and all_names_the_same:
-            row_table[ip] = f"{num_gpus}x {gpu_names[0]}"
-
-headers  = [PRETTY_NAMES[p] for p in show]
-headers += ["Test", f"t/s {name_baseline}", f"t/s {name_compare}", "Speedup"]
+headers  = [pretty_names.get(p, p) for p in show]
+if tool == "llama-bench":
+    headers += ["Test", f"t/s {name_baseline}", f"t/s {name_compare}", "Speedup"]
+elif tool == "test-backend-ops":
+    headers += [f"{primary_metric} {name_baseline}", f"{primary_metric} {name_compare}", "Speedup"]
+else:
+    assert False
 
 if known_args.plot:
-    def create_performance_plot(table_data: list[list[str]], headers: list[str], baseline_name: str, compare_name: str, output_file: str, plot_x_param: str, log_scale: bool = False):
+    def create_performance_plot(table_data: list[list[str]], headers: list[str], baseline_name: str, compare_name: str, output_file: str, plot_x_param: str, log_scale: bool = False, tool_type: str = "llama-bench", metric_name: str = "t/s"):
         try:
-            import matplotlib.pyplot as plt
             import matplotlib
+            import matplotlib.pyplot as plt
             matplotlib.use('Agg')
         except ImportError as e:
             logger.error("matplotlib is required for --plot.")
@@ -627,7 +938,7 @@ if known_args.plot:
         plot_x_label = plot_x_param
 
         if plot_x_param not in ["n_prompt", "n_gen", "n_depth"]:
-            pretty_name = PRETTY_NAMES.get(plot_x_param, plot_x_param)
+            pretty_name = LLAMA_BENCH_PRETTY_NAMES.get(plot_x_param, plot_x_param)
             if pretty_name in data_headers:
                 plot_x_index = data_headers.index(pretty_name)
                 plot_x_label = pretty_name
@@ -746,8 +1057,16 @@ if known_args.plot:
 
             title = ', '.join(title_parts) if title_parts else "Performance comparison"
 
+            # Determine y-axis label based on tool type
+            if tool_type == "llama-bench":
+                y_label = "Tokens per second (t/s)"
+            elif tool_type == "test-backend-ops":
+                y_label = metric_name
+            else:
+                assert False
+
             ax.set_xlabel(plot_x_label, fontsize=12, fontweight='bold')
-            ax.set_ylabel('Tokens per second (t/s)', fontsize=12, fontweight='bold')
+            ax.set_ylabel(y_label, fontsize=12, fontweight='bold')
             ax.set_title(title, fontsize=12, fontweight='bold')
             ax.legend(loc='best', fontsize=10)
             ax.grid(True, alpha=0.3)
@@ -765,7 +1084,7 @@ if known_args.plot:
         plt.savefig(output_file, dpi=300, bbox_inches='tight')
         plt.close()
 
-    create_performance_plot(table, headers, name_baseline, name_compare, known_args.plot, known_args.plot_x, known_args.plot_log_scale)
+    create_performance_plot(table, headers, name_baseline, name_compare, known_args.plot, known_args.plot_x, known_args.plot_log_scale, tool, primary_metric)
 
 print(tabulate( # noqa: NP100
     table,

src/llama-context.cpp

@@ -113,6 +113,15 @@ llama_context::llama_context(
         }
     }
 
+    {
+        const char * LLAMA_GRAPH_REUSE_DISABLE = getenv("LLAMA_GRAPH_REUSE_DISABLE");
+        graph_reuse_disable = LLAMA_GRAPH_REUSE_DISABLE ? (atoi(LLAMA_GRAPH_REUSE_DISABLE) != 0) : graph_reuse_disable;
+
+        if (graph_reuse_disable) {
+            LLAMA_LOG_WARN("%s: graph reuse disabled\n", __func__);
+        }
+    }
+
     const uint32_t n_ctx_per_seq = cparams.n_ctx / cparams.n_seq_max;
 
     LLAMA_LOG_INFO("%s: n_seq_max     = %u\n",   __func__, cparams.n_seq_max);
@@ -716,7 +725,7 @@ llm_graph_result * llama_context::process_ubatch(const llama_ubatch & ubatch, ll
     // in order to correctly reuse a graph, it's full topology has to be uniquely determined by these parameters
     const auto gparams = graph_params(res, ubatch, mctx, gtype);
 
-    if (res->can_reuse(gparams)) {
+    if (!graph_reuse_disable && res->can_reuse(gparams)) {
         //LLAMA_LOG_DEBUG("%s: reusing previous graph\n", __func__);
 
         n_reused++;

src/llama-context.h

@@ -291,6 +291,9 @@ private:
     // ref: https://github.com/ggml-org/llama.cpp/pull/14285
     bool supports_set_rows = false;
 
+    // env: LLAMA_GRAPH_REUSE_DISABLE
+    bool graph_reuse_disable = false;
+
     // perf
     mutable int64_t t_start_us  = 0;
     mutable int64_t t_load_us   = 0;

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

@@ -423,7 +423,9 @@ struct llm_graph_params {
                 (!ubatch.embd  && !other.ubatch.embd)
             );
 
-        if (can_reuse_ubatch && !ubatch.equal_seqs()) {
+        // when we split the batch using "equal_seqs" we have to verify that the participating sequences are the same
+        //   the reason is because the set of attention streams would be different for different sequences
+        if (can_reuse_ubatch && ubatch.equal_seqs()) {
             if (!ubatch.data) {
                 // if the old ubatch does not own it's data, then we cannot guarantee that it is still alive, and
                 //   therefore we cannot perform the sequence id check. normally should never happen
@@ -631,19 +633,8 @@ 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

@@ -290,7 +290,7 @@ static ggml_backend_buffer_type_t select_weight_buft(const llama_hparams & hpara
 }
 
 // CPU: ACCEL -> GPU host -> CPU extra -> CPU
-static buft_list_t make_cpu_buft_list(const std::vector<ggml_backend_dev_t> & devices) {
+static buft_list_t make_cpu_buft_list(const std::vector<ggml_backend_dev_t> & devices, bool use_extra_bufts) {
     buft_list_t buft_list;
 
     // add ACCEL buffer types
@@ -319,21 +319,22 @@ static buft_list_t make_cpu_buft_list(const std::vector<ggml_backend_dev_t> & de
         }
     }
 
-    // add extra buffer types, only if no GPU device is present
-    // ref: https://github.com/ggml-org/llama.cpp/issues/12481#issuecomment-2743136094
-    auto * cpu_dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU);
-    if (cpu_dev == nullptr) {
-        throw std::runtime_error(format("%s: no CPU backend found", __func__));
-    }
+    // add extra buffer types
+    if (use_extra_bufts) {
+        auto * cpu_dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU);
+        if (cpu_dev == nullptr) {
+            throw std::runtime_error(format("%s: no CPU backend found", __func__));
+        }
 
-    auto * cpu_reg = ggml_backend_dev_backend_reg(cpu_dev);
-    auto ggml_backend_dev_get_extra_bufts_fn = (ggml_backend_dev_get_extra_bufts_t)
-        ggml_backend_reg_get_proc_address(cpu_reg, "ggml_backend_dev_get_extra_bufts");
-    if (ggml_backend_dev_get_extra_bufts_fn) {
-        ggml_backend_buffer_type_t * extra_bufts = ggml_backend_dev_get_extra_bufts_fn(cpu_dev);
-        while (extra_bufts && *extra_bufts) {
-            buft_list.emplace_back(cpu_dev, *extra_bufts);
-            ++extra_bufts;
+        auto * cpu_reg = ggml_backend_dev_backend_reg(cpu_dev);
+        auto ggml_backend_dev_get_extra_bufts_fn = (ggml_backend_dev_get_extra_bufts_t)
+            ggml_backend_reg_get_proc_address(cpu_reg, "ggml_backend_dev_get_extra_bufts");
+        if (ggml_backend_dev_get_extra_bufts_fn) {
+            ggml_backend_buffer_type_t * extra_bufts = ggml_backend_dev_get_extra_bufts_fn(cpu_dev);
+            while (extra_bufts && *extra_bufts) {
+                buft_list.emplace_back(cpu_dev, *extra_bufts);
+                ++extra_bufts;
+            }
         }
     }
 
@@ -1839,7 +1840,7 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
     LLAMA_LOG_INFO("%s: loading model tensors, this can take a while... (mmap = %s)\n", __func__, ml.use_mmap ? "true" : "false");
 
     // build a list of buffer types for the CPU and GPU devices
-    pimpl->cpu_buft_list = make_cpu_buft_list(devices);
+    pimpl->cpu_buft_list = make_cpu_buft_list(devices, params.use_extra_bufts);
     for (auto * dev : devices) {
         buft_list_t buft_list = make_gpu_buft_list(dev, split_mode, tensor_split);
         // add CPU buffer types as a fallback
@@ -2044,7 +2045,13 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                 for (const auto * overrides = ml.tensor_buft_overrides; overrides->pattern != nullptr; ++overrides) {
                     std::regex pattern(overrides->pattern);
                     if (std::regex_search(tensor_name, pattern)) {
-                        buft = overrides->buft;
+                        if (overrides->buft == ggml_backend_cpu_buffer_type()) {
+                            // when overriding to a CPU buffer, consider the extra buffer types
+                            buft = select_weight_buft(hparams, t_meta, op, pimpl->cpu_buft_list);
+                        } else {
+                            buft = overrides->buft;
+                        }
+
                         LLAMA_LOG_DEBUG("tensor %s (%zu MiB %s) buffer type overridden to %s\n",
                                 tensor_name.c_str(),
                                 ggml_nbytes(t_meta) / 1024 / 1024, ggml_type_name(t_meta->type),
@@ -17320,10 +17327,18 @@ 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;
@@ -17831,6 +17846,7 @@ llama_model_params llama_model_default_params() {
         /*.use_mmap                    =*/ true,
         /*.use_mlock                   =*/ false,
         /*.check_tensors               =*/ false,
+        /*.use_extra_bufts             =*/ true,
     };
 
 #ifdef GGML_USE_METAL

src/llama-quant.cpp

@@ -875,9 +875,10 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
 
             // get more optimal quantization type based on the tensor shape, layer, etc.
             if (!params->pure && ggml_is_quantized(default_type)) {
+                int fallback = qs.n_fallback;
                 new_type = llama_tensor_get_type(qs, new_type, tensor, ftype);
-                // unless the user specifies a type
-                if (params->tensor_types) {
+                // unless the user specifies a type, and the tensor geometry will not require fallback quantisation
+                if (params->tensor_types && qs.n_fallback - fallback == 0) {
                     const std::vector<tensor_quantization> & tensor_types = *static_cast<const std::vector<tensor_quantization> *>(params->tensor_types);
                     const std::string tensor_name(tensor->name);
                     for (const auto & [tname, qtype] : tensor_types) {
@@ -890,7 +891,6 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
                     }
                 }
             }
-
             if (params->token_embedding_type < GGML_TYPE_COUNT && strcmp(tensor->name, "token_embd.weight") == 0) {
                 new_type = params->token_embedding_type;
             }

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