whispercpp 1.3.2 → 1.3.4

data/ext/sources/examples/talk-llama/llama-memory-recurrent.h

@@ -0,0 +1,182 @@
+#pragma once
+
+#include "llama-batch.h"
+#include "llama-graph.h"
+#include "llama-memory.h"
+
+#include <map>
+#include <set>
+#include <vector>
+
+//
+// llama_memory_recurrent
+//
+
+// TODO: extract the cache state used for graph computation into llama_memory_recurrent_context_i
+//       see the implementation of llama_kv_cache_context_i for an example how to do it
+class llama_memory_recurrent : public llama_memory_i {
+public:
+    llama_memory_recurrent(
+            const llama_model & model,
+                    ggml_type   type_r,
+                    ggml_type   type_s,
+                         bool   offload,
+                     uint32_t   mem_size,
+                     uint32_t   n_seq_max,
+        const layer_filter_cb & filter);
+
+    ~llama_memory_recurrent() = default;
+
+    //
+    // llama_memory_i
+    //
+
+    llama_memory_context_ptr init_batch(
+            llama_batch_allocr & balloc,
+            uint32_t n_ubatch,
+            bool embd_all) override;
+
+    llama_memory_context_ptr init_full() override;
+
+    llama_memory_context_ptr init_update(llama_context * lctx, bool optimize) override;
+
+    void clear(bool data) override;
+
+    bool seq_rm  (llama_seq_id seq_id,                              llama_pos p0, llama_pos p1) override;
+    void seq_cp  (llama_seq_id seq_id_src, llama_seq_id seq_id_dst, llama_pos p0, llama_pos p1) override;
+    void seq_keep(llama_seq_id seq_id)                                                          override;
+    void seq_add (llama_seq_id seq_id,                              llama_pos p0, llama_pos p1, llama_pos shift) override;
+    void seq_div (llama_seq_id seq_id,                              llama_pos p0, llama_pos p1, int d) override;
+
+    llama_pos seq_pos_min(llama_seq_id seq_id) const override;
+    llama_pos seq_pos_max(llama_seq_id seq_id) const override;
+
+    std::map<ggml_backend_buffer_type_t, size_t> memory_breakdown() const override;
+
+    bool prepare(const std::vector<llama_ubatch> & ubatches);
+
+    // find a contiguous slot of memory cells and emplace the ubatch there
+    bool find_slot(const llama_ubatch & ubatch);
+
+    bool get_can_shift() const override;
+
+    // state write/load
+
+    void state_write(llama_io_write_i & io, llama_seq_id seq_id = -1, llama_state_seq_flags flags = 0) const override;
+    void state_read (llama_io_read_i  & io, llama_seq_id seq_id = -1, llama_state_seq_flags flags = 0) override;
+
+    uint32_t head = 0; // the location where the batch will be placed in the cache (see find_slot())
+    uint32_t size = 0; // total number of cells, shared across all sequences
+    uint32_t used = 0; // used cells (i.e. at least one seq_id)
+
+    // computed before each graph build
+    uint32_t n = 0;
+
+    // first zero-ed state
+    int32_t rs_z = -1;
+
+    // TODO: optimize for recurrent state needs
+    struct mem_cell {
+        llama_pos pos  = -1;
+        int32_t   src  = -1; // used to know where states should be copied from
+        int32_t   src0 = -1; // like src, but only used when setting the inputs (allowing to copy once)
+        int32_t   tail = -1;
+
+        std::set<llama_seq_id> seq_id;
+
+        bool has_seq_id(const llama_seq_id & id) const {
+            return seq_id.find(id) != seq_id.end();
+        }
+
+        bool is_empty() const {
+            return seq_id.empty();
+        }
+
+        bool is_same_seq(const mem_cell & other) const {
+            return seq_id == other.seq_id;
+        }
+    };
+
+    std::vector<mem_cell> cells;
+
+    // per layer
+    std::vector<ggml_tensor *> r_l;
+    std::vector<ggml_tensor *> s_l;
+
+private:
+    //const llama_model & model;
+    const llama_hparams & hparams;
+
+    const uint32_t n_seq_max = 1;
+
+    std::vector<ggml_context_ptr>        ctxs;
+    std::vector<ggml_backend_buffer_ptr> bufs;
+
+    size_t total_size() const;
+
+    size_t size_r_bytes() const;
+    size_t size_s_bytes() const;
+
+    void state_write_meta(llama_io_write_i & io, const std::vector<std::pair<uint32_t, uint32_t>> & cell_ranges, llama_seq_id seq_id = -1) const;
+    void state_write_data(llama_io_write_i & io, const std::vector<std::pair<uint32_t, uint32_t>> & cell_ranges) const;
+
+    bool state_read_meta(llama_io_read_i & io, uint32_t cell_count, llama_seq_id dest_seq_id = -1);
+    bool state_read_data(llama_io_read_i & io, uint32_t cell_count);
+};
+
+class llama_memory_recurrent_context : public llama_memory_context_i {
+public:
+    // used for errors
+    llama_memory_recurrent_context(llama_memory_status status);
+
+    // used to create a full-cache or update context
+    llama_memory_recurrent_context(
+            llama_memory_recurrent * mem);
+
+    // used to create a batch processing context from a batch
+    llama_memory_recurrent_context(
+            llama_memory_recurrent * mem,
+            std::vector<llama_ubatch> ubatches);
+
+    virtual ~llama_memory_recurrent_context();
+
+    //
+    // llama_memory_context_i
+    //
+
+    bool next()  override;
+    bool apply() override;
+
+    llama_memory_status  get_status() const override;
+    const llama_ubatch & get_ubatch() const override;
+
+    //
+    // llama_memory_recurrent_context specific API
+    //
+
+    uint32_t get_n_rs() const;
+    uint32_t get_head() const;
+    int32_t  get_rs_z() const;
+    uint32_t get_size() const;
+
+    ggml_tensor * get_r_l(int32_t il) const;
+    ggml_tensor * get_s_l(int32_t il) const;
+
+    int32_t s_copy(int i) const;
+
+private:
+    const llama_memory_status status;
+
+    llama_memory_recurrent * mem;
+
+    size_t i_next = 0;
+
+    std::vector<llama_ubatch> ubatches;
+
+    //
+    // data needed for building the compute graph for the current ubatch:
+    // TODO: extract all the state like `head` and `n` here
+    //
+
+    const bool is_full = false;
+};

data/ext/sources/examples/talk-llama/llama-memory.cpp

@@ -1 +1,59 @@
 #include "llama-memory.h"
+
+llama_memory_status llama_memory_status_combine(llama_memory_status s0, llama_memory_status s1) {
+    bool has_update = false;
+
+    switch (s0) {
+        case LLAMA_MEMORY_STATUS_SUCCESS:
+            {
+                has_update = true;
+                break;
+            }
+        case LLAMA_MEMORY_STATUS_NO_UPDATE:
+            {
+                break;
+            }
+        case LLAMA_MEMORY_STATUS_FAILED_PREPARE:
+        case LLAMA_MEMORY_STATUS_FAILED_COMPUTE:
+            {
+                return s0;
+            }
+    }
+
+    switch (s1) {
+        case LLAMA_MEMORY_STATUS_SUCCESS:
+            {
+                has_update = true;
+                break;
+            }
+        case LLAMA_MEMORY_STATUS_NO_UPDATE:
+            {
+                break;
+            }
+        case LLAMA_MEMORY_STATUS_FAILED_PREPARE:
+        case LLAMA_MEMORY_STATUS_FAILED_COMPUTE:
+            {
+                return s1;
+            }
+    }
+
+    // if either status has an update, then the combined status has an update
+    return has_update ? LLAMA_MEMORY_STATUS_SUCCESS : LLAMA_MEMORY_STATUS_NO_UPDATE;
+}
+
+bool llama_memory_status_is_fail(llama_memory_status status) {
+    switch (status) {
+        case LLAMA_MEMORY_STATUS_SUCCESS:
+        case LLAMA_MEMORY_STATUS_NO_UPDATE:
+            {
+                return false;
+            }
+        case LLAMA_MEMORY_STATUS_FAILED_PREPARE:
+        case LLAMA_MEMORY_STATUS_FAILED_COMPUTE:
+            {
+                return true;
+            }
+    }
+
+    return false;
+}

data/ext/sources/examples/talk-llama/llama-memory.h

@@ -2,6 +2,17 @@
 
 #include "llama.h"
 
+#include <map>
+#include <memory>
+#include <functional>
+
+struct llama_ubatch;
+
+class llama_batch_allocr;
+
+class llama_io_write_i;
+class llama_io_read_i;
+
 struct llama_memory_params {
     // kv cache
     ggml_type type_k;
@@ -11,13 +22,83 @@ struct llama_memory_params {
     bool swa_full;
 };
 
+enum llama_memory_status {
+    LLAMA_MEMORY_STATUS_SUCCESS = 0,
+    LLAMA_MEMORY_STATUS_NO_UPDATE,
+    LLAMA_MEMORY_STATUS_FAILED_PREPARE,
+    LLAMA_MEMORY_STATUS_FAILED_COMPUTE,
+};
+
+// helper function for combining the status of two memory contexts
+// useful for implementing hybrid memory types (e.g. iSWA)
+llama_memory_status llama_memory_status_combine(llama_memory_status s0, llama_memory_status s1);
+
+// helper function for checking if a memory status indicates a failure
+bool llama_memory_status_is_fail(llama_memory_status status);
+
+// the interface for managing the memory context during batch processing
+// this interface is implemented per memory type. see:
+//   - llama_kv_cache_context
+//   - llama_kv_cache_iswa_context
+//   ...
+//
+// the only method that should mutate the memory and the memory context is llama_memory_i::apply()
+struct llama_memory_context_i {
+    virtual ~llama_memory_context_i() = default;
+
+    // consume the current ubatch from the context and proceed to the next one
+    // return false if we are done
+    virtual bool next() = 0;
+
+    // apply the memory state for the current ubatch to the memory object
+    // return false on failure
+    virtual bool apply() = 0;
+
+    // get the current ubatch
+    virtual const llama_ubatch & get_ubatch() const = 0;
+
+    // get the status of the memory context - used for error handling and checking if any updates would be applied
+    virtual llama_memory_status get_status() const = 0;
+};
+
+using llama_memory_context_ptr = std::unique_ptr<llama_memory_context_i>;
+
 // general concept of LLM memory
 // the KV cache is a type of LLM memory, but there can be other types
-class llama_memory_i {
-public:
+struct llama_memory_i {
+    // this callback is used to filter out layers that should not be included in the cache
+    using layer_filter_cb = std::function<bool(int32_t il)>;
+
+    // this callback is used to specify which layers should reuse memory from other layers
+    // return negative value to indicate that the layer il should not reuse memory
+    using layer_reuse_cb = std::function<int32_t(int32_t il)>;
+
     virtual ~llama_memory_i() = default;
 
-    virtual void clear() = 0;
+    // split the input batch into a set of ubatches and verify that they can fit into the cache
+    // return a context object containing the ubatches and memory state required to process them
+    // check the llama_memory_context_i::get_status() for the result
+    virtual llama_memory_context_ptr init_batch(
+            llama_batch_allocr & balloc,
+            uint32_t n_ubatch,
+            bool embd_all) = 0;
+
+    // simulate full cache, used for allocating worst-case compute buffers
+    virtual llama_memory_context_ptr init_full() = 0;
+
+    // prepare for any pending memory updates, such as shifts, copies, etc.
+    // status == LLAMA_MEMORY_STATUS_NO_UPDATE if there is nothing to update
+    virtual llama_memory_context_ptr init_update(llama_context * lctx, bool optimize) = 0;
+
+    // getters
+    virtual bool get_can_shift() const = 0;
+
+    //
+    // ops
+    //
+
+    // if data == true, the data buffers will also be cleared together with the metadata
+    virtual void clear(bool data) = 0;
 
     virtual bool seq_rm  (llama_seq_id seq_id,                              llama_pos p0, llama_pos p1) = 0;
     virtual void seq_cp  (llama_seq_id seq_id_src, llama_seq_id seq_id_dst, llama_pos p0, llama_pos p1) = 0;
@@ -28,5 +109,14 @@ public:
     virtual llama_pos seq_pos_min(llama_seq_id seq_id) const = 0;
     virtual llama_pos seq_pos_max(llama_seq_id seq_id) const = 0;
 
-    virtual bool get_can_edit() const = 0;
+    virtual std::map<ggml_backend_buffer_type_t, size_t> memory_breakdown() const = 0;
+
+    //
+    // state write/read
+    //
+
+    virtual void state_write(llama_io_write_i & io, llama_seq_id seq_id = -1, llama_state_seq_flags flags = 0) const = 0;
+    virtual void state_read (llama_io_read_i  & io, llama_seq_id seq_id = -1, llama_state_seq_flags flags = 0) = 0;
 };
+
+using llama_memory_ptr = std::unique_ptr<llama_memory_i>;

data/ext/sources/examples/talk-llama/llama-mmap.cpp

@@ -401,7 +401,7 @@ struct llama_mmap::impl {
                 }
             }
 #else
-            throw std::runtime_error("PrefetchVirtualMemory unavailable");
+            LLAMA_LOG_DEBUG("skipping PrefetchVirtualMemory because _WIN32_WINNT < 0x602\n");
 #endif
         }
     }

data/ext/sources/examples/talk-llama/llama-model-loader.cpp

@@ -35,6 +35,7 @@ static std::string llama_model_ftype_name(llama_ftype ftype) {
         case LLAMA_FTYPE_MOSTLY_Q5_0:     return "Q5_0";
         case LLAMA_FTYPE_MOSTLY_Q5_1:     return "Q5_1";
         case LLAMA_FTYPE_MOSTLY_Q8_0:     return "Q8_0";
+        case LLAMA_FTYPE_MOSTLY_MXFP4_MOE: return "MXFP4 MoE";
         case LLAMA_FTYPE_MOSTLY_Q2_K:     return "Q2_K - Medium";
         case LLAMA_FTYPE_MOSTLY_Q2_K_S:   return "Q2_K - Small";
         case LLAMA_FTYPE_MOSTLY_Q3_K_S:   return "Q3_K - Small";
@@ -288,9 +289,10 @@ namespace GGUFMeta {
 
     template<typename T>
     bool llama_model_loader::get_arr(const std::string & key, std::vector<T> & result, bool required) {
-        const int kid = gguf_find_key(meta.get(), key.c_str());
+        const gguf_context * ctx = meta.get();
+        const int kid = gguf_find_key(ctx, key.c_str());
 
-        if (kid < 0 || gguf_get_kv_type(meta.get(), kid) != GGUF_TYPE_ARRAY) {
+        if (kid < 0 || gguf_get_kv_type(ctx, kid) != GGUF_TYPE_ARRAY) {
             if (required) {
                 throw std::runtime_error(format("array key not found in model: %s", key.c_str()));
             }
@@ -298,28 +300,40 @@ namespace GGUFMeta {
         }
 
         struct GGUFMeta::ArrayInfo arr_info =
-            GGUFMeta::GKV<GGUFMeta::ArrayInfo>::get_kv(meta.get(), kid);
+            GGUFMeta::GKV<GGUFMeta::ArrayInfo>::get_kv(ctx, kid);
 
         switch (arr_info.gt) {
             case GGUF_TYPE_UINT32:
-            case GGUF_TYPE_INT32:   GGML_ASSERT((std::is_same<T,  int32_t>::value) ||
-                                                (std::is_same<T, uint32_t>::value)); break;
-            case GGUF_TYPE_FLOAT32: GGML_ASSERT((std::is_same<T,    float>::value)); break;
+            case GGUF_TYPE_INT32:   GGML_ASSERT((std::is_same<T,     int32_t>::value) ||
+                                                (std::is_same<T,    uint32_t>::value)); break;
+            case GGUF_TYPE_FLOAT32: GGML_ASSERT((std::is_same<T,       float>::value)); break;
+            case GGUF_TYPE_STRING:  GGML_ASSERT((std::is_same<T, std::string>::value)); break;
             default:
-                throw std::runtime_error(format("%s is not a float32/uint32/int32 array", key.c_str()));
+                throw std::runtime_error(format("%s is not a string/float32/uint32/int32 array", key.c_str()));
         }
 
-        result.resize(arr_info.length);
-        result.assign((const T*)arr_info.data, (const T *)arr_info.data + arr_info.length);
+        if constexpr (std::is_same<T, std::string>::value) {
+            const size_t n_items = gguf_get_arr_n(ctx, kid);
+            result.clear();
+
+            for (size_t i = 0; i < n_items; i++) {
+                const T value = gguf_get_arr_str(ctx, kid, i);
+                result.emplace_back(value);
+            }
+        } else {
+            result.resize(arr_info.length);
+            result.assign((const T*)arr_info.data, (const T *)arr_info.data + arr_info.length);
+        }
 
         return true;
     }
 
     template<typename T, size_t N_MAX>
     bool llama_model_loader::get_arr(const std::string & key, std::array<T, N_MAX> & result, bool required) {
-        const int kid = gguf_find_key(meta.get(), key.c_str());
+        const gguf_context * ctx = meta.get();
+        const int kid = gguf_find_key(ctx, key.c_str());
 
-        if (kid < 0 || gguf_get_kv_type(meta.get(), kid) != GGUF_TYPE_ARRAY) {
+        if (kid < 0 || gguf_get_kv_type(ctx, kid) != GGUF_TYPE_ARRAY) {
             if (required) {
                 throw std::runtime_error(format("array key not found in model: %s", key.c_str()));
             }
@@ -327,22 +341,32 @@ namespace GGUFMeta {
         }
 
         struct GGUFMeta::ArrayInfo arr_info =
-            GGUFMeta::GKV<GGUFMeta::ArrayInfo>::get_kv(meta.get(), kid);
+            GGUFMeta::GKV<GGUFMeta::ArrayInfo>::get_kv(ctx, kid);
 
         switch (arr_info.gt) {
             case GGUF_TYPE_UINT32:
-            case GGUF_TYPE_INT32:   GGML_ASSERT((std::is_same<T,  int32_t>::value) ||
-                                                (std::is_same<T, uint32_t>::value)); break;
-            case GGUF_TYPE_FLOAT32: GGML_ASSERT((std::is_same<T,    float>::value)); break;
+            case GGUF_TYPE_INT32:   GGML_ASSERT((std::is_same<T,     int32_t>::value) ||
+                                                (std::is_same<T,    uint32_t>::value)); break;
+            case GGUF_TYPE_FLOAT32: GGML_ASSERT((std::is_same<T,       float>::value)); break;
+            case GGUF_TYPE_STRING:  GGML_ASSERT((std::is_same<T, std::string>::value)); break;
             default:
-                throw std::runtime_error(format("%s is not a float32/uint32/int32 array", key.c_str()));
+                throw std::runtime_error(format("%s is not a string/float32/uint32/int32 array", key.c_str()));
         }
 
         if (arr_info.length > N_MAX) {
             throw std::runtime_error(format("array length %u for key %s exceeds max %u", (uint32_t) arr_info.length, key.c_str(), (uint32_t) N_MAX));
         }
 
-        std::copy((const T*)arr_info.data, (const T *)arr_info.data + arr_info.length, result.begin());
+        if constexpr (std::is_same<T, std::string>::value) {
+            const size_t n_items = gguf_get_arr_n(ctx, kid);
+
+            for (size_t i = 0; i < n_items; i++) {
+                const T value = gguf_get_arr_str(ctx, kid, i);
+                result[i] = value;
+            }
+        } else {
+            std::copy((const T*)arr_info.data, (const T *)arr_info.data + arr_info.length, result.begin());
+        }
 
         return true;
     }
@@ -352,6 +376,8 @@ namespace GGUFMeta {
         return get_arr(llm_kv(kid), result, required);
     }
 
+    template bool llama_model_loader::get_arr<std::vector<std::string>>(enum llm_kv kid, std::vector<std::string> & result, bool required);
+
     template<typename T>
     bool llama_model_loader::get_key(const std::string & key, T & result, bool required) {
         auto it = kv_overrides.find(key);
@@ -762,6 +788,7 @@ const struct ggml_tensor * llama_model_loader::check_tensor_dims(const std::stri
 }
 
 struct ggml_tensor * llama_model_loader::create_tensor(struct ggml_context * ctx, const std::string & name, const std::initializer_list<int64_t> & ne, int flags) {
+    LLAMA_LOG_DEBUG("%s: loading tensor %s\n", __func__, name.c_str());
     const struct ggml_tensor * cur = check_tensor_dims(name, ne, !(flags & TENSOR_NOT_REQUIRED));
 
     if (cur == NULL) {

data/ext/sources/examples/talk-llama/llama-model-loader.h

@@ -58,8 +58,9 @@ struct llama_model_loader {
         }
     };
 
-    static const int TENSOR_NOT_REQUIRED = 1;
-    static const int TENSOR_DUPLICATED   = 2;
+    static const int TENSOR_NOT_REQUIRED = 1 << 0;
+    static const int TENSOR_DUPLICATED   = 1 << 1;
+    static const int TENSOR_SKIP         = 1 << 2;
 
     int n_kv      = 0;
     int n_tensors = 0;

data/ext/sources/examples/talk-llama/llama-model-saver.cpp

@@ -228,6 +228,7 @@ void llama_model_saver::add_kv_from_model() {
     // add_kv(LLM_KV_TOKENIZER_MASK_ID,                 ???);
     add_kv(LLM_KV_TOKENIZER_ADD_BOS,                 vocab.get_add_bos());
     add_kv(LLM_KV_TOKENIZER_ADD_EOS,                 vocab.get_add_eos());
+    add_kv(LLM_KV_TOKENIZER_ADD_SEP,                 vocab.get_add_sep());
     add_kv(LLM_KV_TOKENIZER_ADD_PREFIX,              vocab.get_add_space_prefix());
     add_kv(LLM_KV_TOKENIZER_REMOVE_EXTRA_WS,         vocab.get_remove_extra_whitespaces());
     add_kv(LLM_KV_TOKENIZER_PRECOMPILED_CHARSMAP,    vocab.get_precompiled_charsmap());