llama: update llama.cpp vendor code to commit d7cfe1ff (#9356)

llama/llama.cpp/src/llama-chat.h

@@ -15,6 +15,7 @@ enum llm_chat_template {
     LLM_CHAT_TEMPLATE_MISTRAL_V3_TEKKEN,
     LLM_CHAT_TEMPLATE_MISTRAL_V7,
     LLM_CHAT_TEMPLATE_PHI_3,
+    LLM_CHAT_TEMPLATE_PHI_4,
     LLM_CHAT_TEMPLATE_FALCON_3,
     LLM_CHAT_TEMPLATE_ZEPHYR,
     LLM_CHAT_TEMPLATE_MONARCH,
@@ -30,6 +31,7 @@ enum llm_chat_template {
     LLM_CHAT_TEMPLATE_LLAMA_3,
     LLM_CHAT_TEMPLATE_CHATGML_3,
     LLM_CHAT_TEMPLATE_CHATGML_4,
+    LLM_CHAT_TEMPLATE_GLMEDGE,
     LLM_CHAT_TEMPLATE_MINICPM,
     LLM_CHAT_TEMPLATE_EXAONE_3,
     LLM_CHAT_TEMPLATE_RWKV_WORLD,

llama/llama.cpp/src/llama-context.cpp

 #include "llama-context.h"
 
+#include "llama-impl.h"
+#include "llama-mmap.h"
+
 #include <cassert>
 #include <cmath>
 #include <cstring>
@@ -513,7 +516,7 @@ size_t llama_output_reserve(struct llama_context & lctx, size_t n_outputs) {
 
         auto * buft = ggml_backend_cpu_buffer_type();
         // try to use the host buffer of the device where the output tensor is allocated for faster transfer to system memory
-        auto * output_dev = lctx.model.dev_output.dev;
+        auto * output_dev = lctx.model.dev_output();
         auto * output_dev_host_buft = output_dev ? ggml_backend_dev_host_buffer_type(output_dev) : nullptr;
         if (output_dev_host_buft) {
             buft = output_dev_host_buft;

llama/llama.cpp/src/llama-context.h

@@ -22,12 +22,12 @@ struct llama_context {
 
     const struct llama_model & model;
 
-    struct llama_cparams        cparams;
-    struct llama_sbatch         sbatch;  // TODO: revisit if needed
-    struct llama_kv_cache       kv_self;
-    struct llama_control_vector cvec;
+    struct llama_cparams      cparams;
+    struct llama_sbatch       sbatch;  // TODO: revisit if needed
+    struct llama_kv_cache     kv_self;
+    struct llama_adapter_cvec cvec;
 
-    std::unordered_map<struct llama_lora_adapter *, float> lora_adapters;
+    std::unordered_map<struct llama_adapter_lora *, float> lora;
 
     std::vector<ggml_backend_ptr> backends;
     std::vector<std::pair<ggml_backend_t, ggml_backend_set_n_threads_t>> set_n_threads_fns;

llama/llama.cpp/src/llama-grammar.h

@@ -114,6 +114,15 @@ struct llama_grammar {
 
     // buffer for partially generated UTF-8 sequence from accepted tokens
     llama_partial_utf8 partial_utf8;
+
+    // lazy grammars wait for trigger words or tokens before constraining the sampling.
+    // we still have trigger_tokens for non-lazy grammars to force printing of special trigger tokens.
+    // (useful e.g. for tool_choice=required)
+    bool                     lazy             = false;
+    bool                     awaiting_trigger = false; // Initialized to true for lazy grammars only
+    std::string              trigger_buffer;           // Output buffered by lazy grammar. Will be cleared once trigger is found.
+    std::vector<llama_token> trigger_tokens;           // Tokens that trigger a lazy grammar, or tokens to force printing of (even if special).
+    std::vector<std::string> trigger_words;
 };
 
 //
@@ -127,7 +136,15 @@ struct llama_grammar * llama_grammar_init_impl(
         size_t n_rules,
         size_t start_rule_index);
 
-struct llama_grammar * llama_grammar_init_impl(const struct llama_vocab * vocab, const char * grammar_str, const char * grammar_root);
+struct llama_grammar * llama_grammar_init_impl(
+        const struct llama_vocab * vocab,
+                      const char * grammar_str,
+                      const char * grammar_root,
+                              bool lazy,
+                     const char ** trigger_words,
+                            size_t num_trigger_words,
+               const llama_token * trigger_tokens,
+                            size_t num_trigger_tokens);
 
 void llama_grammar_free_impl(struct llama_grammar * grammar);
 
@@ -141,3 +158,7 @@ void llama_grammar_apply_impl(
 void llama_grammar_accept_impl(
               struct llama_grammar & grammar,
                        llama_token   token);
+
+void llama_grammar_accept_str(
+              struct llama_grammar & grammar,
+                 const std::string & piece);

llama/llama.cpp/src/llama-hparams.cpp

@@ -54,7 +54,7 @@ uint32_t llama_hparams::n_embd_v_gqa(uint32_t il) const {
 uint32_t llama_hparams::n_embd_k_s() const {
     if (wkv_head_size != 0) {
         // for RWKV models
-        return 2 * n_embd;
+        return token_shift_count * n_embd;
     }
 
     // TODO: maybe support other convolution strides than 1
@@ -82,4 +82,4 @@ bool llama_hparams::n_bskcn(uint32_t n, uint32_t il) const {
 
 bool llama_hparams::cross_attention_layers(uint32_t il) const {
     return std::find(cross_attn_layers.begin(), cross_attn_layers.end(), il) != cross_attn_layers.end();
-}
+}
\ No newline at end of file

llama/llama.cpp/src/llama-hparams.h

@@ -30,7 +30,6 @@ struct llama_hparams {
     bool use_par_res;
     bool swin_norm;
 
-    uint32_t n_vocab = 0;
     uint32_t n_ctx_train; // context size the model was trained on
     uint32_t n_embd;
     uint32_t n_embd_features = 0;
@@ -41,8 +40,8 @@ struct llama_hparams {
     uint32_t n_embd_head_v; // dimension of values (d_v) aka n_embd_head
     uint32_t n_expert = 0;
     uint32_t n_expert_used = 0;
-    uint32_t n_vocab_type = 0; // for BERT-style token types
     uint32_t n_rel_attn_bkts = 0;
+    uint32_t n_vocab = 0;
 
     // for WavTokenizer
     struct llama_hparams_posnet   posnet;
@@ -79,6 +78,7 @@ struct llama_hparams {
     uint32_t time_mix_extra_dim     = 0;
     uint32_t time_decay_extra_dim   = 0;
     uint32_t wkv_head_size          = 0;
+    uint32_t token_shift_count      = 2;
 
     float    rope_attn_factor = 1.0f;
     float    rope_freq_base_train;
@@ -141,7 +141,7 @@ struct llama_hparams {
     // Block skip connection
     bool n_bskcn(uint32_t n, uint32_t il) const;
 
-    // cross attention layers   
+    // cross attention layers
     bool cross_attention_layers(uint32_t il) const;
 };
 

llama/llama.cpp/src/llama-impl.cpp

 #include "llama-impl.h"
 
+#include "gguf.h"
 #include "llama.h"
 
 #include <cinttypes>
@@ -138,7 +139,7 @@ std::string gguf_kv_to_str(const struct gguf_context * ctx_gguf, int i) {
             {
                 const enum gguf_type arr_type = gguf_get_arr_type(ctx_gguf, i);
                 int arr_n = gguf_get_arr_n(ctx_gguf, i);
-                const void * data = gguf_get_arr_data(ctx_gguf, i);
+                const void * data = arr_type == GGUF_TYPE_STRING ? nullptr : gguf_get_arr_data(ctx_gguf, i);
                 std::stringstream ss;
                 ss << "[";
                 for (int j = 0; j < arr_n; j++) {

llama/llama.cpp/src/llama-impl.h

@@ -6,13 +6,13 @@
 #include <vector>
 
 #ifdef __GNUC__
-#ifdef __MINGW32__
-#define LLAMA_ATTRIBUTE_FORMAT(...) __attribute__((format(gnu_printf, __VA_ARGS__)))
+#    if defined(__MINGW32__) && !defined(__clang__)
+#        define LLAMA_ATTRIBUTE_FORMAT(...) __attribute__((format(gnu_printf, __VA_ARGS__)))
+#    else
+#        define LLAMA_ATTRIBUTE_FORMAT(...) __attribute__((format(printf, __VA_ARGS__)))
+#    endif
 #else
-#define LLAMA_ATTRIBUTE_FORMAT(...) __attribute__((format(printf, __VA_ARGS__)))
-#endif
-#else
-#define LLAMA_ATTRIBUTE_FORMAT(...)
+#    define LLAMA_ATTRIBUTE_FORMAT(...)
 #endif
 
 //

llama/llama.cpp/src/llama-kv-cache.cpp

@@ -72,39 +72,6 @@ bool llama_kv_cache_init(
     cache.v_l.reserve(n_layer);
 
     for (int i = 0; i < n_layer; i++) {
-        // for cross attention layers
-        if (model.arch == LLM_ARCH_MLLAMA && hparams.cross_attention_layers(i)) {
-            const uint32_t n_embd_k_gqa = hparams.n_embd_k_gqa(i) + hparams.n_embd_k_s();
-            const llama_model::buft_list_t * buft_list;
-            if (offload) {
-                buft_list = model.dev_layer.at(i).buft_list;
-            } else {
-                buft_list = &model.cpu_buft_list;
-            }
-            ggml_backend_buffer_type_t buft = select_buft(*buft_list,
-                [&](ggml_context * ctx) {
-                    ggml_tensor * k = ggml_new_tensor_1d(ctx, type_k, n_embd_k_gqa*kv_size);
-                    if (hparams.rope_type == LLAMA_ROPE_TYPE_NONE) {
-                        return k;
-                    }
-                    ggml_tensor * p = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, 1);
-                    return ggml_rope(ctx, k, p, hparams.n_rot, hparams.rope_type);
-                });
-            ggml_context * ctx = ctx_for_buft(buft);
-
-            if (!ctx) {
-                LLAMA_LOG_ERROR("%s: failed to create ggml context for kv cache\n", __func__);
-                return false;
-            }
-            ggml_tensor * k = ggml_new_tensor_3d(ctx, GGML_TYPE_F32, hparams.n_embd_head_k, 6404, hparams.n_head_kv(i));
-            ggml_tensor * v = ggml_new_tensor_3d(ctx, GGML_TYPE_F32, hparams.n_embd_head_v, 6404, hparams.n_head_kv(i));
-            ggml_format_name(k, "cache_k_l%d", i);
-            ggml_format_name(v, "cache_v_l%d", i);
-            cache.k_l.push_back(k);
-            cache.v_l.push_back(v);
-            continue;
-        }
-
         const uint32_t n_embd_k_gqa = hparams.n_embd_k_gqa(i) + hparams.n_embd_k_s();
         const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(i) + hparams.n_embd_v_s();
 
@@ -112,7 +79,7 @@ bool llama_kv_cache_init(
 
         ggml_backend_buffer_type_t buft;
         if (offload) {
-            auto * dev = model.dev_layer.at(i).dev;
+            auto * dev = model.dev_layer(i);
             buft = ggml_backend_dev_buffer_type(dev);
         } else {
             buft = ggml_backend_cpu_buffer_type();
@@ -124,8 +91,17 @@ bool llama_kv_cache_init(
             return false;
         }
 
-        ggml_tensor * k = ggml_new_tensor_1d(ctx, type_k, n_embd_k_gqa*kv_size);
-        ggml_tensor * v = ggml_new_tensor_1d(ctx, type_v, n_embd_v_gqa*kv_size);
+        ggml_tensor * k, *v;
+
+        // for cross attention layers
+        if (model.arch == LLM_ARCH_MLLAMA && hparams.cross_attention_layers(i)) {
+            k = ggml_new_tensor_3d(ctx, GGML_TYPE_F32, hparams.n_embd_head_k, 6404, hparams.n_head_kv(i));
+            v = ggml_new_tensor_3d(ctx, GGML_TYPE_F32, hparams.n_embd_head_v, 6404, hparams.n_head_kv(i));
+        } else {
+            k = ggml_new_tensor_1d(ctx, type_k, n_embd_k_gqa*kv_size);
+            v = ggml_new_tensor_1d(ctx, type_v, n_embd_v_gqa*kv_size);
+        }
+
         ggml_format_name(k, "cache_k_l%d", i);
         ggml_format_name(v, "cache_v_l%d", i);
         cache.k_l.push_back(k);
@@ -152,10 +128,10 @@ bool llama_kv_cache_init(
 
 struct llama_kv_cache_slot_info llama_kv_cache_find_slot(
            struct llama_kv_cache & cache,
-       const struct llama_ubatch & batch) {
-    const uint32_t n_tokens = batch.n_tokens;
-    const uint32_t n_seqs   = batch.n_seqs;
-    const uint32_t n_seq_tokens = batch.n_seq_tokens;
+       const struct llama_ubatch & ubatch) {
+    const uint32_t n_tokens = ubatch.n_tokens;
+    const uint32_t n_seqs   = ubatch.n_seqs;
+    const uint32_t n_seq_tokens = ubatch.n_seq_tokens;
 
     if (cache.recurrent) {
         // For recurrent state architectures (like Mamba or RWKV),
@@ -163,16 +139,16 @@ struct llama_kv_cache_slot_info llama_kv_cache_find_slot(
         // A slot should be always be contiguous.
 
         // can only process batches with an equal number of new tokens in each sequence
-        GGML_ASSERT(batch.equal_seqs);
+        GGML_ASSERT(ubatch.equal_seqs);
 
         int32_t min = cache.size - 1;
         int32_t max = 0;
 
         // everything should fit if all seq_ids are smaller than the max
         for (uint32_t s = 0; s < n_seqs; ++s) {
-            const uint32_t n_seq_id = batch.n_seq_id[s];
+            const uint32_t n_seq_id = ubatch.n_seq_id[s];
             for (uint32_t j = 0; j < n_seq_id; ++j) {
-                const llama_seq_id seq_id = batch.seq_id[s][j];
+                const llama_seq_id seq_id = ubatch.seq_id[s][j];
 
                 if (seq_id < 0 || (uint32_t) seq_id >= cache.size) {
                     // too big seq_id
@@ -231,7 +207,7 @@ struct llama_kv_cache_slot_info llama_kv_cache_find_slot(
 
         // find usable cell range
         for (uint32_t s = 0; s < n_seqs; ++s) {
-            const llama_seq_id seq_id = batch.seq_id[s][0];
+            const llama_seq_id seq_id = ubatch.seq_id[s][0];
             llama_kv_cell & seq_meta = cache.cells[seq_id];
             bool has_cell = false;
             if (seq_meta.tail >= 0) {
@@ -270,7 +246,7 @@ struct llama_kv_cache_slot_info llama_kv_cache_find_slot(
         // gather and re-order
         for (uint32_t s = 0; s < n_seqs; ++s) {
             int32_t dst_id = s + min;
-            int32_t src_id = cache.cells[batch.seq_id[s][0]].tail;
+            int32_t src_id = cache.cells[ubatch.seq_id[s][0]].tail;
             if (dst_id != src_id) {
                 llama_kv_cell & dst_cell = cache.cells[dst_id];
                 llama_kv_cell & src_cell = cache.cells[src_id];
@@ -291,7 +267,7 @@ struct llama_kv_cache_slot_info llama_kv_cache_find_slot(
 
         // update the pos of the used seqs
         for (uint32_t s = 0; s < n_seqs; ++s) {
-            const llama_pos last_pos = batch.pos[n_seq_tokens * s + n_seq_tokens - 1];
+            const llama_pos last_pos = ubatch.pos[n_seq_tokens * s + n_seq_tokens - 1];
             int32_t cell_id = s + min;
             llama_kv_cell & cell = cache.cells[cell_id];
 
@@ -299,12 +275,12 @@ struct llama_kv_cache_slot_info llama_kv_cache_find_slot(
                 // What should happen when the pos backtracks or skips a value?
                 // Clearing the state mid-batch would require special-casing which isn't done.
                 LLAMA_LOG_WARN("%s: non-consecutive token position %d after %d for sequence %d with %u new tokens\n",
-                    __func__, last_pos, cell.pos, batch.seq_id[s][0], n_seq_tokens);
+                    __func__, last_pos, cell.pos, ubatch.seq_id[s][0], n_seq_tokens);
             }
             cell.pos = last_pos;
             cell.seq_id.clear();
-            for (int32_t j = 0; j < batch.n_seq_id[s]; ++j) {
-                const llama_seq_id seq_id = batch.seq_id[s][j];
+            for (int32_t j = 0; j < ubatch.n_seq_id[s]; ++j) {
+                const llama_seq_id seq_id = ubatch.seq_id[s][j];
                 cell.seq_id.insert(seq_id);
                 cache.cells[seq_id].tail = cell_id;
             }
@@ -358,10 +334,10 @@ struct llama_kv_cache_slot_info llama_kv_cache_find_slot(
     for (uint32_t s = 0; s < n_seqs; s++) {
         for (uint32_t i = 0; i < n_seq_tokens; ++i) {
             uint32_t k = s*n_seq_tokens + i;
-            cache.cells[cache.head + k].pos = batch.pos[k];
+            cache.cells[cache.head + k].pos = ubatch.pos[k];
 
-            for (int32_t j = 0; j < batch.n_seq_id[s]; j++) {
-                cache.cells[cache.head + k].seq_id.insert(batch.seq_id[s][j]);
+            for (int32_t j = 0; j < ubatch.n_seq_id[s]; j++) {
+                cache.cells[cache.head + k].seq_id.insert(ubatch.seq_id[s][j]);
             }
         }
     }

llama/llama.cpp/src/llama-kv-cache.h

@@ -37,7 +37,7 @@ struct llama_kv_cache {
     bool can_shift = false;
 
     // Note: The value of head isn't only used to optimize searching
-    // for a free KV slot. llama_decode_internal also uses it, so it
+    // for a free KV slot. llama_decode_impl also uses it, so it
     // cannot be freely changed after a slot has been allocated.
     uint32_t head = 0;
     uint32_t size = 0;

llama/llama.cpp/src/llama-mmap.cpp

@@ -7,6 +7,7 @@
 #include <cstring>
 #include <climits>
 #include <stdexcept>
+#include <cerrno>
 
 #ifdef __has_include
     #if __has_include(<unistd.h>)
@@ -35,7 +36,7 @@
 
 // TODO: consider moving to llama-impl.h if needed in more places
 #if defined(_WIN32)
-std::string llama_format_win_err(DWORD err) {
+static std::string llama_format_win_err(DWORD err) {
     LPSTR buf;
     size_t size = FormatMessageA(FORMAT_MESSAGE_ALLOCATE_BUFFER | FORMAT_MESSAGE_FROM_SYSTEM | FORMAT_MESSAGE_IGNORE_INSERTS,
                                  NULL, err, MAKELANGID(LANG_NEUTRAL, SUBLANG_DEFAULT), (LPSTR)&buf, 0, NULL);
@@ -241,12 +242,16 @@ llama_file::~llama_file() = default;
 size_t llama_file::tell() const { return pimpl->tell(); }
 size_t llama_file::size() const { return pimpl->size; }
 
-int llama_file::fileno() const {
+int llama_file::file_id() const {
 #ifdef _WIN32
     return _fileno(pimpl->fp);
+#else
+#if defined(fileno)
+    return fileno(pimpl->fp);
 #else
     return ::fileno(pimpl->fp);
 #endif
+#endif
 }
 
 void llama_file::seek(size_t offset, int whence) const { pimpl->seek(offset, whence); }
@@ -265,7 +270,7 @@ struct llama_mmap::impl {
 
     impl(struct llama_file * file, size_t prefetch, bool numa) {
         size = file->size();
-        int fd = file->fileno();
+        int fd = file->file_id();
         int flags = MAP_SHARED;
         if (numa) { prefetch = 0; }
 #ifdef __linux__
@@ -357,7 +362,7 @@ struct llama_mmap::impl {
 
         size = file->size();
 
-        HANDLE hFile = (HANDLE) _get_osfhandle(file->fileno());
+        HANDLE hFile = (HANDLE) _get_osfhandle(file->file_id());
 
         HANDLE hMapping = CreateFileMappingA(hFile, NULL, PAGE_READONLY, 0, 0, NULL);
 

llama/llama.cpp/src/llama-mmap.h

 #pragma once
 
+#include <cstdint>
 #include <memory>
 #include <vector>
 
@@ -18,7 +19,7 @@ struct llama_file {
     size_t tell() const;
     size_t size() const;
 
-    int fileno() const;
+    int file_id() const; // fileno overload
 
     void seek(size_t offset, int whence) const;
 

llama/llama.cpp/src/llama-model-loader.cpp

@@ -7,6 +7,10 @@
 #include <cstring>
 #include <future>
 
+static const size_t kiB = 1024;
+static const size_t MiB = 1024*kiB;
+static const size_t GiB = 1024*MiB;
+
 const char * llama_file_version_name(llama_fver version) {
     switch (version) {
         case GGUF_FILE_VERSION_V1: return "GGUF V1 (support until nov 2023)";
@@ -17,8 +21,78 @@ const char * llama_file_version_name(llama_fver version) {
     return "unknown";
 }
 
+static std::string llama_model_ftype_name(llama_ftype ftype) {
+    if (ftype & LLAMA_FTYPE_GUESSED) {
+        return llama_model_ftype_name((enum llama_ftype) (ftype & ~LLAMA_FTYPE_GUESSED)) + " (guessed)";
+    }
+
+    switch (ftype) {
+        case LLAMA_FTYPE_ALL_F32:         return "all F32";
+        case LLAMA_FTYPE_MOSTLY_F16:      return "F16";
+        case LLAMA_FTYPE_MOSTLY_BF16:     return "BF16";
+        case LLAMA_FTYPE_MOSTLY_Q4_0:     return "Q4_0";
+        case LLAMA_FTYPE_MOSTLY_Q4_1:     return "Q4_1";
+        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_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";
+        case LLAMA_FTYPE_MOSTLY_Q3_K_M:   return "Q3_K - Medium";
+        case LLAMA_FTYPE_MOSTLY_Q3_K_L:   return "Q3_K - Large";
+        case LLAMA_FTYPE_MOSTLY_Q4_K_S:   return "Q4_K - Small";
+        case LLAMA_FTYPE_MOSTLY_Q4_K_M:   return "Q4_K - Medium";
+        case LLAMA_FTYPE_MOSTLY_Q5_K_S:   return "Q5_K - Small";
+        case LLAMA_FTYPE_MOSTLY_Q5_K_M:   return "Q5_K - Medium";
+        case LLAMA_FTYPE_MOSTLY_Q6_K:     return "Q6_K";
+        case LLAMA_FTYPE_MOSTLY_TQ1_0:    return "TQ1_0 - 1.69 bpw ternary";
+        case LLAMA_FTYPE_MOSTLY_TQ2_0:    return "TQ2_0 - 2.06 bpw ternary";
+        case LLAMA_FTYPE_MOSTLY_IQ2_XXS:  return "IQ2_XXS - 2.0625 bpw";
+        case LLAMA_FTYPE_MOSTLY_IQ2_XS:   return "IQ2_XS - 2.3125 bpw";
+        case LLAMA_FTYPE_MOSTLY_IQ2_S:    return "IQ2_S - 2.5 bpw";
+        case LLAMA_FTYPE_MOSTLY_IQ2_M:    return "IQ2_M - 2.7 bpw";
+        case LLAMA_FTYPE_MOSTLY_IQ3_XS:   return "IQ3_XS - 3.3 bpw";
+        case LLAMA_FTYPE_MOSTLY_IQ3_XXS:  return "IQ3_XXS - 3.0625 bpw";
+        case LLAMA_FTYPE_MOSTLY_IQ1_S:    return "IQ1_S - 1.5625 bpw";
+        case LLAMA_FTYPE_MOSTLY_IQ1_M:    return "IQ1_M - 1.75 bpw";
+        case LLAMA_FTYPE_MOSTLY_IQ4_NL:   return "IQ4_NL - 4.5 bpw";
+        case LLAMA_FTYPE_MOSTLY_IQ4_XS:   return "IQ4_XS - 4.25 bpw";
+        case LLAMA_FTYPE_MOSTLY_IQ3_S:    return "IQ3_S - 3.4375 bpw";
+        case LLAMA_FTYPE_MOSTLY_IQ3_M:    return "IQ3_S mix - 3.66 bpw";
+
+        default: return "unknown, may not work";
+    }
+}
+
+// return a list of splits for a given path
+// for example, given "<name>-00002-of-00004.gguf", returns list of all 4 splits
+static std::vector<std::string> llama_get_list_splits(const std::string & path, const int idx, const int n_split) {
+    std::vector<std::string> paths;
+    std::string split_prefix;
+    std::vector<char> buf(llama_path_max(), 0);
+
+    {
+        int ret = llama_split_prefix(buf.data(), buf.size(), path.c_str(), idx, n_split);
+        if (!ret) {
+            throw std::runtime_error(format("invalid split file name: %s", path.c_str()));
+        }
+        split_prefix = std::string(buf.data(), ret);
+    }
+
+    if (split_prefix.empty()) {
+        throw std::runtime_error(format("invalid split file: %s", path.c_str()));
+    }
+
+    for (int idx = 0; idx < n_split; ++idx) {
+        int ret = llama_split_path(buf.data(), buf.size(), split_prefix.c_str(), idx, n_split);
+        paths.push_back(std::string(buf.data(), ret));
+    }
+
+    return paths;
+}
+
 namespace GGUFMeta {
-    template <typename T, gguf_type gt_, T (*gfun)(const gguf_context *, const int)>
+    template <typename T, gguf_type gt_, T (*gfun)(const gguf_context *, const int64_t)>
     struct GKV_Base_Type {
         static constexpr gguf_type gt = gt_;
 
@@ -60,10 +134,11 @@ namespace GGUFMeta {
         public:
         static constexpr gguf_type gt = GGUF_TYPE_ARRAY;
         static ArrayInfo getter(const gguf_context *ctx, const int k) {
+            const enum gguf_type arr_type = gguf_get_arr_type(ctx, k);
             return ArrayInfo {
-                gguf_get_arr_type(ctx, k),
+                arr_type,
                 size_t(gguf_get_arr_n(ctx, k)),
-                gguf_get_arr_data(ctx, k),
+                arr_type == GGUF_TYPE_STRING ? nullptr : gguf_get_arr_data(ctx, k),
             };
         }
     };
@@ -368,7 +443,12 @@ namespace GGUFMeta {
     template bool llama_model_loader::get_key_or_arr<std::array<uint32_t, 512>>(enum llm_kv kid, std::array<uint32_t, 512> & result, uint32_t n, bool required);
     template bool llama_model_loader::get_key_or_arr<uint32_t>(const std::string & key, std::array<uint32_t, 512> & result, uint32_t n, bool required);
 
-llama_model_loader::llama_model_loader(const std::string & fname, bool use_mmap, bool check_tensors, const struct llama_model_kv_override * param_overrides_p) {
+llama_model_loader::llama_model_loader(
+        const std::string & fname,
+        std::vector<std::string> & splits,
+        bool use_mmap,
+        bool check_tensors,
+        const struct llama_model_kv_override * param_overrides_p) {
     int trace = 0;
     if (getenv("LLAMA_TRACE")) {
         trace = atoi(getenv("LLAMA_TRACE"));
@@ -380,6 +460,7 @@ llama_model_loader::llama_model_loader(const std::string & fname, bool use_mmap,
         }
     }
 
+    // Load the main GGUF
     struct ggml_context * ctx = NULL;
     struct gguf_init_params params = {
         /*.no_alloc = */ true,
@@ -415,35 +496,54 @@ llama_model_loader::llama_model_loader(const std::string & fname, bool use_mmap,
 
     // Load additional GGML contexts
     if (n_split > 1) {
+        // make sure the main file is loaded first
         uint16_t idx = 0;
-        get_key(llm_kv(LLM_KV_SPLIT_NO), idx);
+        const std::string kv_split_no = llm_kv(LLM_KV_SPLIT_NO);
+        get_key(kv_split_no, idx);
         if (idx != 0) {
-            throw std::runtime_error(format("illegal split file: %d, model must be loaded with the first split", idx));
+            throw std::runtime_error(format("illegal split file idx: %d (file: %s), model must be loaded with the first split", idx, fname.c_str()));
+        }
+
+        // generate list of splits if needed
+        if (splits.empty()) {
+            splits = llama_get_list_splits(fname, idx, n_split);
         }
 
-        std::vector<char> split_prefix(llama_path_max(), 0);
-        if (!llama_split_prefix(split_prefix.data(), split_prefix.size(), fname.c_str(), idx, n_split)) {
-            throw std::runtime_error(format("invalid split file: %s", fname.c_str()));
+        // in case user give a custom list of splits, check if it matches the expected number
+        if (n_split != (uint16_t)splits.size()) {
+            throw std::runtime_error(format("invalid split count, given: %zu splits, but expected %d", splits.size(), n_split));
         }
 
         if (trace > 0) {
             LLAMA_LOG_INFO("%s: loading additional %d GGUFs\n", __func__, n_split);
         }
 
-        std::vector<char> split_path(llama_path_max(), 0);
+        // load other splits
         for (idx = 1; idx < n_split; idx++) {
-            llama_split_path(split_path.data(), split_path.size(), split_prefix.data(), idx, n_split);
+            const char * fname_split = splits[idx].c_str();
 
             struct gguf_init_params split_params = {
                 /*.no_alloc = */ true,
                 /*.ctx      = */ &ctx,
             };
-            gguf_context_ptr ctx_gguf { gguf_init_from_file(split_path.data(), split_params) };
+            gguf_context_ptr ctx_gguf { gguf_init_from_file(fname_split, split_params) };
             if (!ctx_gguf) {
-                throw std::runtime_error(format("%s: failed to load GGUF split from %s\n", __func__, split_path.data()));
+                throw std::runtime_error(format("%s: failed to load GGUF split from %s\n", __func__, fname_split));
+            }
+
+            // check idx
+            {
+                const int kid = gguf_find_key(ctx_gguf.get(), kv_split_no.c_str());
+                if (kid < 0) {
+                    throw std::runtime_error(format("missing key %s in GGUF split %s", kv_split_no.c_str(), fname_split));
+                }
+                int idx_gguf = gguf_get_val_u16(ctx_gguf.get(), kid);
+                if (idx_gguf != idx) {
+                    throw std::runtime_error(format("invalid split file idx: %d (file: %s), expected %d", idx_gguf, fname_split, idx));
+                }
             }
 
-            files.emplace_back(new llama_file(split_path.data(), "rb"));
+            files.emplace_back(new llama_file(fname_split, "rb"));
             contexts.emplace_back(ctx);
 
             // Save tensors data offset info of the shard.
@@ -556,7 +656,7 @@ llama_model_loader::llama_model_loader(const std::string & fname, bool use_mmap,
             const enum gguf_type type   = gguf_get_kv_type(meta.get(), i);
             const std::string type_name =
                 type == GGUF_TYPE_ARRAY
-                ? format("%s[%s,%d]", gguf_type_name(type), gguf_type_name(gguf_get_arr_type(meta.get(), i)), gguf_get_arr_n(meta.get(), i))
+                ? format("%s[%s,%zu]", gguf_type_name(type), gguf_type_name(gguf_get_arr_type(meta.get(), i)), gguf_get_arr_n(meta.get(), i))
                 : gguf_type_name(type);
 
             std::string value          = gguf_kv_to_str(meta.get(), i);
@@ -722,7 +822,7 @@ void llama_model_loader::init_mappings(bool prefetch, llama_mlocks * mlock_mmaps
         for (const auto & file : files) {
             auto * reg = ggml_backend_dev_backend_reg(ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU));
             auto * is_numa_fn = (decltype(ggml_is_numa) *) ggml_backend_reg_get_proc_address(reg, "ggml_backend_cpu_is_numa");
-            std::unique_ptr<llama_mmap> mapping(new llama_mmap(file.get(), prefetch ? -1 : 0, is_numa_fn()));
+            std::unique_ptr<llama_mmap> mapping = std::make_unique<llama_mmap>(file.get(), prefetch ? -1 : 0, is_numa_fn());
             mmaps_used.emplace_back(mapping->size(), 0);
             if (mlock_mmaps) {
                 std::unique_ptr<llama_mlock> mlock_mmap(new llama_mlock());
@@ -1011,3 +1111,17 @@ bool llama_model_loader::load_all_data(
 
     return true;
 }
+
+std::string llama_model_loader::ftype_name() const {
+    return llama_model_ftype_name(ftype);
+}
+
+void llama_model_loader::print_info() const {
+    LLAMA_LOG_INFO("%s: file format = %s\n", __func__, llama_file_version_name(fver));
+    LLAMA_LOG_INFO("%s: file type   = %s\n", __func__, llama_model_ftype_name(ftype).c_str());
+    if (n_bytes < GiB) {
+        LLAMA_LOG_INFO("%s: file size   = %.2f MiB (%.2f BPW) \n", __func__, n_bytes/1024.0/1024.0,        n_bytes*8.0/n_elements);
+    } else {
+        LLAMA_LOG_INFO("%s: file size   = %.2f GiB (%.2f BPW) \n", __func__, n_bytes/1024.0/1024.0/1024.0, n_bytes*8.0/n_elements);
+    }
+}

llama/llama.cpp/src/llama-model-loader.h

@@ -90,7 +90,12 @@ struct llama_model_loader {
     size_t size_data = 0;
     std::vector<std::pair<size_t, size_t>> mmaps_used;
 
-    llama_model_loader(const std::string & fname, bool use_mmap, bool check_tensors, const struct llama_model_kv_override * param_overrides_p);
+    llama_model_loader(
+        const std::string & fname,
+        std::vector<std::string> & splits, // optional, only need if the split does not follow naming scheme
+        bool use_mmap,
+        bool check_tensors,
+        const struct llama_model_kv_override * param_overrides_p);
 
     template<typename T>
     typename std::enable_if<std::is_integral<T>::value, bool>::type
@@ -155,4 +160,8 @@ struct llama_model_loader {
             llama_mlocks * lmlocks,
             llama_progress_callback progress_callback,
             void * progress_callback_user_data);
+
+    std::string ftype_name() const;
+
+    void print_info() const;
 };

llama/llama.cpp/src/llama-model.h

@@ -4,81 +4,83 @@
 #include "llama-arch.h"
 #include "llama-hparams.h"
 #include "llama-vocab.h"
-#include "llama-mmap.h"
-
-#include "ggml-cpp.h"
 
+#include <memory>
+#include <string>
+#include <unordered_map>
 #include <vector>
 #include <stdexcept>
 
+struct llama_model_loader;
+
 // available models
-// TODO: this enum does not follow the enum naming convention
 enum llm_type {
-    MODEL_UNKNOWN,
-    MODEL_14M,
-    MODEL_17M,
-    MODEL_22M,
-    MODEL_33M,
-    MODEL_60M,
-    MODEL_70M,
-    MODEL_80M,
-    MODEL_109M,
-    MODEL_137M,
-    MODEL_160M,
-    MODEL_220M,
-    MODEL_250M,
-    MODEL_270M,
-    MODEL_335M,
-    MODEL_410M,
-    MODEL_450M,
-    MODEL_770M,
-    MODEL_780M,
-    MODEL_0_5B,
-    MODEL_1B,
-    MODEL_1_3B,
-    MODEL_1_4B,
-    MODEL_1_5B,
-    MODEL_1_6B,
-    MODEL_2B,
-    MODEL_2_8B,
-    MODEL_3B,
-    MODEL_4B,
-    MODEL_6B,
-    MODEL_6_9B,
-    MODEL_7B,
-    MODEL_8B,
-    MODEL_9B,
-    MODEL_11B,
-    MODEL_12B,
-    MODEL_13B,
-    MODEL_14B,
-    MODEL_15B,
-    MODEL_16B,
-    MODEL_20B,
-    MODEL_22B,
-    MODEL_30B,
-    MODEL_32B,
-    MODEL_34B,
-    MODEL_35B,
-    MODEL_40B,
-    MODEL_65B,
-    MODEL_70B,
-    MODEL_90B,
-    MODEL_236B,
-    MODEL_314B,
-    MODEL_671B,
-    MODEL_SMALL,
-    MODEL_MEDIUM,
-    MODEL_LARGE,
-    MODEL_XL,
-    MODEL_A1_7B,
-    MODEL_A2_7B,
-    MODEL_8x7B,
-    MODEL_8x22B,
-    MODEL_16x12B,
-    MODEL_10B_128x3_66B,
-    MODEL_57B_A14B,
-    MODEL_27B,
+    LLM_TYPE_UNKNOWN,
+    LLM_TYPE_14M,
+    LLM_TYPE_17M,
+    LLM_TYPE_22M,
+    LLM_TYPE_33M,
+    LLM_TYPE_60M,
+    LLM_TYPE_70M,
+    LLM_TYPE_80M,
+    LLM_TYPE_109M,
+    LLM_TYPE_137M,
+    LLM_TYPE_160M,
+    LLM_TYPE_220M,
+    LLM_TYPE_250M,
+    LLM_TYPE_270M,
+    LLM_TYPE_335M,
+    LLM_TYPE_410M,
+    LLM_TYPE_450M,
+    LLM_TYPE_770M,
+    LLM_TYPE_780M,
+    LLM_TYPE_0_5B,
+    LLM_TYPE_1B,
+    LLM_TYPE_1_3B,
+    LLM_TYPE_1_4B,
+    LLM_TYPE_1_5B,
+    LLM_TYPE_1_6B,
+    LLM_TYPE_2B,
+    LLM_TYPE_2_8B,
+    LLM_TYPE_3B,
+    LLM_TYPE_4B,
+    LLM_TYPE_6B,
+    LLM_TYPE_6_9B,
+    LLM_TYPE_7B,
+    LLM_TYPE_8B,
+    LLM_TYPE_9B,
+    LLM_TYPE_11B,
+    LLM_TYPE_12B,
+    LLM_TYPE_13B,
+    LLM_TYPE_14B,
+    LLM_TYPE_15B,
+    LLM_TYPE_16B,
+    LLM_TYPE_20B,
+    LLM_TYPE_22B,
+    LLM_TYPE_30B,
+    LLM_TYPE_32B,
+    LLM_TYPE_34B,
+    LLM_TYPE_35B,
+    LLM_TYPE_40B,
+    LLM_TYPE_65B,
+    LLM_TYPE_70B,
+    LLM_TYPE_90B,
+    LLM_TYPE_236B,
+    LLM_TYPE_314B,
+    LLM_TYPE_671B,
+    LLM_TYPE_SMALL,
+    LLM_TYPE_MEDIUM,
+    LLM_TYPE_LARGE,
+    LLM_TYPE_XL,
+    LLM_TYPE_A1_7B,
+    LLM_TYPE_A2_7B,
+    LLM_TYPE_8x7B,
+    LLM_TYPE_8x22B,
+    LLM_TYPE_16x12B,
+    LLM_TYPE_16x3_8B,
+    LLM_TYPE_10B_128x3_66B,
+    LLM_TYPE_57B_A14B,
+    LLM_TYPE_27B,
 };
 
 struct llama_layer_posnet {
@@ -243,15 +245,19 @@ struct llama_layer {
     struct ggml_tensor * time_mix_lerp_v     = nullptr;
     struct ggml_tensor * time_mix_lerp_r     = nullptr;
     struct ggml_tensor * time_mix_lerp_g     = nullptr;
-
-    struct ggml_tensor * time_mix_first      = nullptr;
-    struct ggml_tensor * time_mix_decay      = nullptr;
-    struct ggml_tensor * time_mix_decay_w1   = nullptr;
-    struct ggml_tensor * time_mix_decay_w2   = nullptr;
-    struct ggml_tensor * time_mix_key        = nullptr;
-    struct ggml_tensor * time_mix_value      = nullptr;
-    struct ggml_tensor * time_mix_receptance = nullptr;
-    struct ggml_tensor * time_mix_gate       = nullptr;
+    struct ggml_tensor * time_mix_lerp_fused = nullptr;
+
+    struct ggml_tensor * time_mix_first        = nullptr;
+    struct ggml_tensor * time_mix_decay        = nullptr;
+    struct ggml_tensor * time_mix_decay_w1     = nullptr;
+    struct ggml_tensor * time_mix_decay_w2     = nullptr;
+    struct ggml_tensor * time_mix_key          = nullptr;
+    struct ggml_tensor * time_mix_key_b        = nullptr;
+    struct ggml_tensor * time_mix_value        = nullptr;
+    struct ggml_tensor * time_mix_value_b      = nullptr;
+    struct ggml_tensor * time_mix_receptance   = nullptr;
+    struct ggml_tensor * time_mix_receptance_b = nullptr;
+    struct ggml_tensor * time_mix_gate         = nullptr;
 
     struct ggml_tensor * time_mix_ln     = nullptr;
     struct ggml_tensor * time_mix_ln_b   = nullptr;
@@ -280,7 +286,7 @@ struct llama_layer {
 
     struct ggml_tensor * bskcn_tv = nullptr;
 
-     // cross attention
+    // cross attention
     struct ggml_tensor * cross_attn_k_norm = nullptr;
     struct ggml_tensor * cross_attn_k_proj = nullptr;
     struct ggml_tensor * cross_attn_o_proj = nullptr;
@@ -296,11 +302,9 @@ struct llama_layer {
 };
 
 struct llama_model {
-    llm_type type = MODEL_UNKNOWN;
+    llm_type type = LLM_TYPE_UNKNOWN;
     llm_arch arch = LLM_ARCH_UNKNOWN;
 
-    llama_ftype ftype = LLAMA_FTYPE_ALL_F32;
-
     std::string name = "n/a";
 
     llama_hparams hparams = {};
@@ -329,117 +333,53 @@ struct llama_model {
 
     std::vector<llama_layer> layers;
 
+    llama_model_params params;
+
     // gguf metadata
     std::unordered_map<std::string, std::string> gguf_kv;
 
-    llama_split_mode split_mode;
-    int main_gpu;
-    int n_gpu_layers;
-
-    std::vector<std::string> rpc_servers;
-
     // list of devices used in this model
     std::vector<ggml_backend_dev_t> devices;
 
+    // for quantize-stats only
+    std::vector<std::pair<std::string, struct ggml_tensor *>> tensors_by_name;
 
-    // lists of buffer types used for each layer
-    using buft_list_t = std::vector<std::pair<ggml_backend_dev_t, ggml_backend_buffer_type_t>>;
-    buft_list_t cpu_buft_list;
-    std::map<ggml_backend_dev_t, buft_list_t> gpu_buft_list;
+    int64_t t_load_us  = 0;
+    int64_t t_start_us = 0;
 
-    struct layer_dev {
-        ggml_backend_dev_t dev;
-        buft_list_t * buft_list;
-    };
+    explicit llama_model(const struct llama_model_params & params);
+    ~llama_model();
 
-    layer_dev dev_input = {};
-    layer_dev dev_output = {};
-    std::vector<layer_dev> dev_layer;
+    void load_stats  (llama_model_loader & ml);
+    void load_arch   (llama_model_loader & ml);
+    void load_hparams(llama_model_loader & ml);
+    void load_vocab  (llama_model_loader & ml);
+    bool load_tensors(llama_model_loader & ml); // returns false if cancelled by progress_callback
 
-    // contexts where the model tensors metadata is stored
-    std::vector<ggml_context_ptr> ctxs;
+    std::string arch_name() const;
+    std::string type_name() const;
 
-    // the model memory buffers for the tensor data
-    std::vector<ggml_backend_buffer_ptr> bufs;
+    std::string desc() const;
 
-    // model memory mapped files
-    llama_mmaps mappings;
+    size_t size() const;
+    size_t max_nodes() const;
+    size_t n_devices() const;
 
-    // objects representing data potentially being locked in memory
-    llama_mlocks mlock_bufs;
-    llama_mlocks mlock_mmaps;
+    // total number of parameters in the model
+    uint64_t n_elements() const;
 
-    // for quantize-stats only
-    std::vector<std::pair<std::string, struct ggml_tensor *>> tensors_by_name;
+    void print_info() const;
 
-    int64_t t_load_us  = 0;
-    int64_t t_start_us = 0;
+    ggml_backend_dev_t dev_layer(int il) const;
+    ggml_backend_dev_t dev_output() const;
 
-    // total number of parameters in the model
-    uint64_t n_elements = 0;
+    ggml_backend_buffer_type_t select_buft(int il) const;
 
-    // total size of all the tensors in the model in bytes
-    size_t  n_bytes     = 0;
+    const struct ggml_tensor * get_tensor(const char * name) const;
+
+private:
+    struct impl;
+    std::unique_ptr<impl> pimpl;
 };
 
 const char * llm_type_name(llm_type type);
-
-std::string llama_model_arch_name (const llama_model & model);
-std::string llama_model_type_name (const llama_model & model);
-std::string llama_model_ftype_name(const llama_model & model);
-
-template<typename F>
-bool buft_supported(ggml_backend_buffer_type_t buft, ggml_backend_dev_t dev, F & fn) {
-    ggml_init_params params = {
-        /*.mem_size   =*/ ggml_tensor_overhead()*8,
-        /*.mem_buffer =*/ NULL,
-        /*.no_alloc   =*/ true,
-    };
-
-    ggml_context_ptr ctx { ggml_init(params) };
-    if (!ctx) {
-        throw std::runtime_error("failed to create ggml context");
-    }
-
-    ggml_backend_buffer_ptr buf { ggml_backend_buft_alloc_buffer(buft, 0) };
-    ggml_tensor * op_tensor = fn(ctx.get());
-    for (int i = 0; i < GGML_MAX_SRC; i++) {
-        if (op_tensor->src[i] != nullptr) {
-            op_tensor->src[i]->buffer = buf.get();
-        }
-    }
-
-    bool op_supported = ggml_backend_dev_supports_op(dev, op_tensor);
-
-    return op_supported;
-}
-
-template<typename F>
-ggml_backend_buffer_type_t select_buft(const llama_model::buft_list_t & buft_list, const F & fn) {
-    for (const auto & cur : buft_list) {
-        ggml_backend_dev_t cur_dev = cur.first;
-        ggml_backend_buffer_type_t cur_buft = cur.second;
-        if (buft_supported(cur_buft, cur_dev, fn)) {
-            return cur_buft;
-        }
-    }
-
-    throw std::runtime_error("no suitable buffer type found");
-}
-
-// used by llama_adapter_cvec
-ggml_backend_buffer_type_t llama_model_select_buft(const llama_model & model, int il);
-
-// used by llama_adapter_lora
-struct ggml_tensor * llama_model_get_tensor(const struct llama_model & model, const char * name);
-
-size_t llama_model_max_nodes(const llama_model & model);
-
-struct llama_model_loader;
-
-// TODO: become llama_model methods
-void llm_load_stats     (llama_model_loader & ml, llama_model & model);
-void llm_load_arch      (llama_model_loader & ml, llama_model & model);
-void llm_load_hparams   (llama_model_loader & ml, llama_model & model);
-void llm_load_vocab     (llama_model_loader & ml, llama_model & model);
-void llm_load_print_meta(llama_model_loader & ml, llama_model & model);

llama/llama.cpp/src/llama-quant.cpp

@@ -7,14 +7,12 @@
 #include <algorithm>
 #include <cmath>
 #include <cstring>
+#include <cinttypes>
 #include <fstream>
 #include <mutex>
 #include <thread>
 #include <unordered_map>
 
-// TODO: replace with ggml API call
-#define QK_K 256
-
 static void zeros(std::ofstream & file, size_t n) {
     char zero = 0;
     for (size_t i = 0; i < n; ++i) {
@@ -22,7 +20,7 @@ static void zeros(std::ofstream & file, size_t n) {
     }
 }
 
-struct quantize_state_internal {
+struct quantize_state_impl {
     const llama_model                 & model;
     const llama_model_quantize_params * params;
 
@@ -43,13 +41,13 @@ struct quantize_state_internal {
     // used to figure out if a model shares tok_embd with the output weight
     bool has_output = false;
 
-    quantize_state_internal(const llama_model & model, const llama_model_quantize_params * params)
+    quantize_state_impl(const llama_model & model, const llama_model_quantize_params * params)
         : model(model)
         , params(params)
         {}
 };
 
-static void llama_tensor_dequantize_internal(
+static void llama_tensor_dequantize_impl(
     struct ggml_tensor * tensor, std::vector<no_init<float>> & output, std::vector<std::thread> & workers,
     const size_t nelements, const int nthread
 ) {
@@ -121,7 +119,7 @@ static void llama_tensor_dequantize_internal(
     workers.clear();
 }
 
-static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type new_type, const ggml_tensor * tensor, llama_ftype ftype) {
+static ggml_type llama_tensor_get_type(quantize_state_impl & qs, ggml_type new_type, const ggml_tensor * tensor, llama_ftype ftype) {
     const std::string name = ggml_get_name(tensor);
 
     // TODO: avoid hardcoded tensor names - use the TN_* constants
@@ -154,8 +152,10 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n
         if (qs.params->output_tensor_type < GGML_TYPE_COUNT) {
             new_type = qs.params->output_tensor_type;
         } else {
-            int nx = tensor->ne[0];
-            if (arch == LLM_ARCH_FALCON || nx % QK_K != 0) {
+            const int64_t nx = tensor->ne[0];
+            const int64_t qk_k = ggml_blck_size(new_type);
+
+            if (arch == LLM_ARCH_FALCON || nx % qk_k != 0) {
                 new_type = GGML_TYPE_Q8_0;
             }
             else if (ftype == LLAMA_FTYPE_MOSTLY_IQ2_XXS || ftype == LLAMA_FTYPE_MOSTLY_IQ2_XS || ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS ||
@@ -235,7 +235,7 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n
         else if ((ftype == LLAMA_FTYPE_MOSTLY_Q4_K_M || ftype == LLAMA_FTYPE_MOSTLY_Q5_K_M) &&
                 use_more_bits(qs.i_attention_wv, qs.n_attention_wv)) new_type = GGML_TYPE_Q6_K;
         else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_S && qs.i_attention_wv < 4) new_type = GGML_TYPE_Q5_K;
-        if (qs.model.type == MODEL_70B) {
+        if (qs.model.type == LLM_TYPE_70B) {
             // In the 70B model we have 8 heads sharing the same attn_v weights. As a result, the attn_v.weight tensor is
             // 8x smaller compared to attn_q.weight. Hence, we can get a nice boost in quantization accuracy with
             // nearly negligible increase in model size by quantizing this tensor with more bits:
@@ -367,20 +367,19 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n
     //    if (ftype == LLAMA_FTYPE_MOSTLY_Q5_K_S) new_type = GGML_TYPE_Q4_K;
     //}
     bool convert_incompatible_tensor = false;
-    if (new_type == GGML_TYPE_Q2_K    || new_type == GGML_TYPE_Q3_K    || new_type == GGML_TYPE_Q4_K   ||
-        new_type == GGML_TYPE_Q5_K    || new_type == GGML_TYPE_Q6_K    || new_type == GGML_TYPE_IQ4_XS ||
-        new_type == GGML_TYPE_IQ2_XS  || new_type == GGML_TYPE_IQ2_XXS || new_type == GGML_TYPE_IQ2_S  ||
-        new_type == GGML_TYPE_IQ3_XXS || new_type == GGML_TYPE_IQ1_S   || new_type == GGML_TYPE_IQ3_S  ||
-        new_type == GGML_TYPE_IQ1_M) {
-        int nx = tensor->ne[0];
-        int ny = tensor->ne[1];
-        if (nx % QK_K != 0) {
-            LLAMA_LOG_WARN("\n\n%s : tensor cols %d x %d are not divisible by %d, required for %s", __func__, nx, ny, QK_K, ggml_type_name(new_type));
+    {
+        const int64_t nx = tensor->ne[0];
+        const int64_t ny = tensor->ne[1];
+        const int64_t qk_k = ggml_blck_size(new_type);
+
+        if (nx % qk_k != 0) {
+            LLAMA_LOG_WARN("\n\n%s : tensor cols %" PRId64 " x %" PRId64 " are not divisible by %" PRId64 ", required for %s", __func__, nx, ny, qk_k, ggml_type_name(new_type));
             convert_incompatible_tensor = true;
         } else {
             ++qs.n_k_quantized;
         }
     }
+
     if (convert_incompatible_tensor) {
         switch (new_type) {
             case GGML_TYPE_TQ1_0:
@@ -410,7 +409,7 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n
     return new_type;
 }
 
-static size_t llama_tensor_quantize_internal(enum ggml_type new_type, const float * f32_data, void * new_data, const int64_t chunk_size, int64_t nrows, int64_t n_per_row, const float * imatrix, std::vector<std::thread> & workers, const int nthread) {
+static size_t llama_tensor_quantize_impl(enum ggml_type new_type, const float * f32_data, void * new_data, const int64_t chunk_size, int64_t nrows, int64_t n_per_row, const float * imatrix, std::vector<std::thread> & workers, const int nthread) {
     if (nthread < 2) {
         // single-thread
         size_t new_size = ggml_quantize_chunk(new_type, f32_data, new_data, 0, nrows, n_per_row, imatrix);
@@ -464,7 +463,7 @@ static size_t llama_tensor_quantize_internal(enum ggml_type new_type, const floa
     return new_size;
 }
 
-static void llama_model_quantize_internal(const std::string & fname_inp, const std::string & fname_out, const llama_model_quantize_params * params) {
+static void llama_model_quantize_impl(const std::string & fname_inp, const std::string & fname_out, const llama_model_quantize_params * params) {
     ggml_type default_type;
     llama_ftype ftype = params->ftype;
 
@@ -526,18 +525,21 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
         auto v = (std::vector<llama_model_kv_override>*)params->kv_overrides;
         kv_overrides = v->data();
     }
-    llama_model_loader ml(fname_inp, use_mmap, /*check_tensors*/ true, kv_overrides);
+
+    std::vector<std::string> splits = {};
+    llama_model_loader ml(fname_inp, splits, use_mmap, /*check_tensors*/ true, kv_overrides);
     ml.init_mappings(false); // no prefetching
 
-    llama_model model;
-    llm_load_arch   (ml, model);
-    llm_load_hparams(ml, model);
-    llm_load_stats  (ml, model);
+    llama_model model(llama_model_default_params());
+
+    model.load_arch   (ml);
+    model.load_hparams(ml);
+    model.load_stats  (ml);
 
-    struct quantize_state_internal qs(model, params);
+    struct quantize_state_impl qs(model, params);
 
     if (params->only_copy) {
-        ftype = model.ftype;
+        ftype = ml.ftype;
     }
     const std::unordered_map<std::string, std::vector<float>> * imatrix_data = nullptr;
     if (params->imatrix) {
@@ -621,7 +623,8 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
 
     qs.n_ffn_down = qs.n_ffn_gate = qs.n_ffn_up = (int)model.hparams.n_layer;
 
-    // sanity checks
+    // sanity checks for models that have attention layers
+    if (qs.n_attention_wv != 0)
     {
         const auto & n_head_kv_iter = model.hparams.n_head_kv_arr.begin();
         // attention layers have a non-zero number of kv heads
@@ -761,6 +764,7 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
         quantize &= name.find("time_mix_w2.weight") == std::string::npos;
         quantize &= name.find("time_mix_decay_w1.weight") == std::string::npos;
         quantize &= name.find("time_mix_decay_w2.weight") == std::string::npos;
+        quantize &= name.find("time_mix_lerp_fused.weight") == std::string::npos;
 
         // do not quantize relative position bias (T5)
         quantize &= name.find("attn_rel_b.weight") == std::string::npos;
@@ -839,7 +843,7 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
             } else if (ggml_is_quantized(tensor->type) && !params->allow_requantize) {
                 throw std::runtime_error(format("requantizing from type %s is disabled", ggml_type_name(tensor->type)));
             } else {
-                llama_tensor_dequantize_internal(tensor, f32_conv_buf, workers, nelements, nthread);
+                llama_tensor_dequantize_impl(tensor, f32_conv_buf, workers, nelements, nthread);
                 f32_data = (float *) f32_conv_buf.data();
             }
 
@@ -868,7 +872,7 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
                 void * new_data_03 = (char *)new_data + ggml_row_size(new_type, n_per_row) * i03 * nrows;
                 const float * imatrix_03 = imatrix ? imatrix + i03 * n_per_row : nullptr;
 
-                new_size += llama_tensor_quantize_internal(new_type, f32_data_03, new_data_03, chunk_size, nrows, n_per_row, imatrix_03, workers, nthread_use);
+                new_size += llama_tensor_quantize_impl(new_type, f32_data_03, new_data_03, chunk_size, nrows, n_per_row, imatrix_03, workers, nthread_use);
             }
             LLAMA_LOG_INFO("size = %8.2f MiB -> %8.2f MiB\n", ggml_nbytes(tensor)/1024.0/1024.0, new_size/1024.0/1024.0);
         }
@@ -877,7 +881,8 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
 
         // update the gguf meta data as we go
         gguf_set_tensor_type(ctx_outs[cur_split].get(), name.c_str(), new_type);
-        gguf_set_tensor_data(ctx_outs[cur_split].get(), name.c_str(), new_data, new_size);
+        GGML_ASSERT(gguf_get_tensor_size(ctx_outs[cur_split].get(), gguf_find_tensor(ctx_outs[cur_split].get(), name.c_str())) == new_size);
+        gguf_set_tensor_data(ctx_outs[cur_split].get(), name.c_str(), new_data);
 
         // write tensor data + padding
         fout.write((const char *) new_data, new_size);
@@ -921,7 +926,7 @@ uint32_t llama_model_quantize(
         const char * fname_out,
         const llama_model_quantize_params * params) {
     try {
-        llama_model_quantize_internal(fname_inp, fname_out, params);
+        llama_model_quantize_impl(fname_inp, fname_out, params);
     } catch (const std::exception & err) {
         LLAMA_LOG_ERROR("%s: failed to quantize: %s\n", __func__, err.what());
         return 1;

llama/llama.cpp/src/llama-sampling.h

@@ -2,7 +2,9 @@
 
 // TODO: rename llama-sampling.h/.cpp to llama-sampler.h/.cpp ?
 
-#include "llama-grammar.h"
+#include "llama.h"
+
+#include <vector>
 
 struct llama_vocab;
 struct llama_grammar;
@@ -21,24 +23,6 @@ struct llama_sampler_chain {
     mutable int32_t n_sample;
 };
 
-struct llama_sampler * llama_sampler_init_grammar_impl(
-        const struct llama_vocab & vocab,
-                      const char * grammar_str,
-                      const char * grammar_root);
-
-struct llama_sampler * llama_sampler_init_infill_impl(
-        const struct llama_vocab & vocab);
-
-struct llama_sampler * llama_sampler_init_dry_impl(
-        const struct llama_vocab &  vocab,
-                         int32_t    context_size,
-                           float    dry_multiplier,
-                           float    dry_base,
-                         int32_t    dry_allowed_length,
-                         int32_t    dry_penalty_last_n,
-                      const char ** seq_breakers,
-                          size_t    num_breakers);
-
 struct llama_sampler * llama_sampler_init_dry_testing(
                          int32_t   context_size,
                            float   dry_multiplier,