Add tokenizer longhaul cache

common/arg.cpp

@@ -817,6 +817,15 @@ static bool common_params_parse_ex(int argc, char ** argv, common_params_context
     if (params.load_mode == LLAMA_LOAD_MODE_LONGHAUL && params.longhaul_cache_bytes == 0) {
         throw std::invalid_argument("error: --longhaul requires --longhaul-cache N\n");
     }
+    if (params.tokenizer_longhaul != (params.tokenizer_longhaul_cache_bytes != 0)) {
+        throw std::invalid_argument(
+            params.tokenizer_longhaul
+                ? "error: --tokenizer-longhaul requires --tokenizer-longhaul-cache N\n"
+                : "error: --tokenizer-longhaul-cache requires --tokenizer-longhaul\n");
+    }
+    if (params.tokenizer_longhaul && params.token_cache_size_mib == 0) {
+        throw std::invalid_argument("error: --tokenizer-longhaul requires an enabled persistent token cache\n");
+    }
 
     postprocess_cpu_params(params.cpuparams,       nullptr);
     postprocess_cpu_params(params.cpuparams_batch, &params.cpuparams);
@@ -2631,7 +2640,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
     ).set_env("LLAMA_ARG_LOAD_MODE"));
     add_opt(common_arg(
         {"--longhaul"},
-        "stream routed Qwen3.5 MoE, Laguna, or Inkling experts through a bounded CPU or Metal cache",
+        "stream routed Qwen3.5 MoE experts through a bounded Metal cache",
         [](common_params & params) {
             params.load_mode = LLAMA_LOAD_MODE_LONGHAUL;
         }
@@ -2646,6 +2655,23 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             params.longhaul_cache_bytes = uint64_t(value) * 1024 * 1024 * 1024;
         }
     ).set_env("LLAMA_ARG_LONGHAUL_CACHE"));
+    add_opt(common_arg(
+        {"--tokenizer-longhaul"},
+        "build and page the Qwen3.5/Laguna tokenizer through a bounded cache",
+        [](common_params & params) {
+            params.tokenizer_longhaul = true;
+        }
+    ).set_env("LLAMA_ARG_TOKENIZER_LONGHAUL"));
+    add_opt(common_arg(
+        {"--tokenizer-longhaul-cache"}, "N",
+        "tokenizer-longhaul RAM cache size in MiB",
+        [](common_params & params, int value) {
+            if (value <= 0) {
+                throw std::invalid_argument("tokenizer-longhaul cache size must be positive");
+            }
+            params.tokenizer_longhaul_cache_bytes = uint64_t(value) * 1024 * 1024;
+        }
+    ).set_env("LLAMA_ARG_TOKENIZER_LONGHAUL_CACHE"));
     add_opt(common_arg(
         {"--numa"}, "TYPE",
         "attempt optimizations that help on some NUMA systems\n"

common/chat.cpp

@@ -2530,118 +2530,6 @@ static common_chat_params common_chat_params_init_minimax_m3(const common_chat_t
     return data;
 }
 
-// Inkling's TML format can emit multiple typed content blocks in one turn.
-// <|end_message|> closes a block while content_model_end_sampling closes the
-// generation, so the generic single-block parser is not sufficient.
-static common_chat_params common_chat_params_init_inkling(
-        const common_chat_template & tmpl,
-        const autoparser::generation_params & inputs) {
-    common_chat_params data;
-
-    const std::string MSG_MODEL    = "<|message_model|>";
-    const std::string MSG_USER     = "<|message_user|>";
-    const std::string MSG_SYSTEM   = "<|message_system|>";
-    const std::string MSG_TOOL     = "<|message_tool|>";
-    const std::string THINK        = "<|content_thinking|>";
-    const std::string TEXT         = "<|content_text|>";
-    const std::string END_MESSAGE  = "<|end_message|>";
-    const std::string END_SAMPLING = "<|content_model_end_sampling|>";
-    const std::string INVOKE_TOOL  = "<|content_invoke_tool_json|>";
-
-    data.prompt             = common_chat_template_direct_apply_impl(tmpl, inputs);
-    data.generation_prompt  = common_chat_template_generation_prompt_impl(tmpl, inputs);
-    data.format             = COMMON_CHAT_FORMAT_PEG_NATIVE;
-    data.supports_thinking  = true;
-    data.thinking_start_tag = THINK;
-    data.thinking_end_tags  = {END_MESSAGE};
-    data.preserved_tokens   = {
-        MSG_MODEL, MSG_USER, MSG_SYSTEM, MSG_TOOL,
-        THINK, TEXT, END_MESSAGE, END_SAMPLING, INVOKE_TOOL,
-    };
-    data.message_delimiters = {
-        { COMMON_CHAT_ROLE_ASSISTANT, MSG_MODEL },
-        { COMMON_CHAT_ROLE_USER,      MSG_USER },
-        { COMMON_CHAT_ROLE_SYSTEM,    MSG_SYSTEM },
-        { COMMON_CHAT_ROLE_TOOL,      MSG_TOOL },
-    };
-
-    const bool has_tools = inputs.tools.is_array() && !inputs.tools.empty();
-    const bool extract_reasoning =
-        inputs.reasoning_format != COMMON_REASONING_FORMAT_NONE;
-
-    if (inputs.has_continuation()) {
-        const auto & msg = inputs.continue_msg;
-        data.generation_prompt = MSG_MODEL + THINK + msg.reasoning_content;
-        if (inputs.continue_final_message == COMMON_CHAT_CONTINUATION_CONTENT) {
-            data.generation_prompt += END_MESSAGE + TEXT + msg.render_content();
-        }
-        data.prompt += data.generation_prompt;
-    }
-
-    auto parser = build_chat_peg_parser([&](common_chat_peg_builder & p) {
-        auto generation_prompt = p.literal(MSG_MODEL);
-        auto end = p.end();
-
-        common_peg_parser reasoning_block = p.eps();
-        if (extract_reasoning) {
-            reasoning_block =
-                p.literal(THINK) +
-                p.reasoning(p.until_one_of({ END_MESSAGE, TEXT, END_SAMPLING })) +
-                p.optional(p.literal(END_MESSAGE));
-        } else {
-            reasoning_block = p.content(
-                p.literal(THINK) +
-                p.until_one_of({ END_MESSAGE, TEXT, END_SAMPLING }) +
-                p.optional(p.literal(END_MESSAGE)));
-        }
-        auto reasoning = p.optional(reasoning_block);
-
-        auto text_block =
-            p.optional(p.literal(MSG_MODEL)) +
-            p.optional(p.literal(TEXT)) +
-            p.content(p.until_one_of({ THINK, END_MESSAGE, END_SAMPLING })) +
-            p.optional(p.literal(END_MESSAGE));
-        auto text_content =
-            p.one_or_more(p.choice({ reasoning_block, text_block }));
-
-        if (!has_tools || inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_NONE) {
-            return generation_prompt + reasoning + text_content +
-                p.optional(p.literal(END_SAMPLING)) + end;
-        }
-
-        auto tool_section = p.standard_json_tools(
-            INVOKE_TOOL, END_MESSAGE, inputs.tools,
-            /* parallel_tool_calls = */ false,
-            /* force_tool_calls    = */ true,
-            /* name_key            = */ "name",
-            /* args_key            = */ "args",
-            /* array_wrapped       = */ false,
-            /* function_is_key     = */ false,
-            /* call_id_key         = */ "",
-            /* gen_call_id_key     = */ "",
-            /* parameters_order    = */ {},
-            /* accept_openai_wrapper = */ false);
-        auto tool_block =
-            p.optional(p.literal(MSG_MODEL)) +
-            p.until_one_of({ INVOKE_TOOL, TEXT, THINK, END_MESSAGE, END_SAMPLING }) +
-            tool_section;
-        auto tool_calls = inputs.parallel_tool_calls
-            ? p.one_or_more(tool_block)
-            : tool_block;
-        auto mixed_body =
-            p.one_or_more(p.choice({ tool_block, reasoning_block, text_block }));
-        auto body = inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_REQUIRED
-            ? tool_calls
-            : mixed_body;
-
-        return generation_prompt + reasoning + body +
-            p.optional(p.literal(END_SAMPLING)) + end;
-    });
-
-    data.parser = parser.save();
-    return data;
-}
-
 namespace workaround {
 
 static void map_developer_role_to_system(json & messages) {
@@ -3059,13 +2947,6 @@ std::optional<common_chat_params> common_chat_try_specialized_template(
         return common_chat_params_init_cohere2moe(tmpl, params);
     }
 
-    if (src.find("<|content_thinking|>") != std::string::npos &&
-        src.find("<|content_text|>") != std::string::npos &&
-        src.find("<|message_model|>") != std::string::npos) {
-        LOG_DBG("Using specialized template: Inkling\n");
-        return common_chat_params_init_inkling(tmpl, params);
-    }
-
     if (is_lfm2_template(src)) {
         LOG_DBG("Using specialized template: LFM2\n");
         return common_chat_params_init_lfm2(tmpl, params, /* tool_list_tokens = */ true);

common/common.cpp

@@ -1592,6 +1592,8 @@ struct llama_model_params common_model_params_to_llama(common_params & params) {
     mparams.split_mode      = params.split_mode;
     mparams.load_mode       = params.load_mode;
     mparams.longhaul_cache_bytes = params.longhaul_cache_bytes;
+    mparams.tokenizer_longhaul = params.tokenizer_longhaul;
+    mparams.tokenizer_longhaul_cache_bytes = params.tokenizer_longhaul_cache_bytes;
     mparams.tensor_split    = params.tensor_split;
     mparams.check_tensors   = params.check_tensors;
     mparams.use_extra_bufts = !params.no_extra_bufts;

common/common.h

@@ -486,6 +486,8 @@ struct common_params {
     enum llama_split_mode split_mode = LLAMA_SPLIT_MODE_LAYER; // how to split the model across GPUs
     enum llama_load_mode  load_mode  = LLAMA_LOAD_MODE_MMAP; // how to load the model
     uint64_t longhaul_cache_bytes = 0;
+    bool tokenizer_longhaul = false;
+    uint64_t tokenizer_longhaul_cache_bytes = 0;
 
     common_cpu_params cpuparams;
     common_cpu_params cpuparams_batch;

docs/longhaul.md

@@ -11,17 +11,11 @@ llama-cli \
     --model /path/to/model.gguf \
     --longhaul \
     --longhaul-cache 2 \
-    --n-gpu-layers 0
+    --n-gpu-layers 99
 ```
 
 `--longhaul-cache` is the expert cache budget in GiB. It is required when `--longhaul` is used. The same options are accepted by `llama-server`.
 
-The command above runs every repeating layer on CPU and is supported on macOS,
-Linux, and Windows. On macOS, the existing all-Metal mode remains available by
-using `--n-gpu-layers 99` instead. Longhaul does not silently change device
-placement: an unsupported GPU or mixed CPU/GPU repeating-layer placement fails
-with guidance to use `--n-gpu-layers 0`.
-
 Longhaul may reduce `--ubatch-size` so that every expert selected by one graph segment can be present in the cache at the same time. The effective value is logged during context creation. If one token selects more experts than the cache has slots, the routed MoE computation is split into multiple stages and the partial results are summed. This permits smaller caches at the cost of additional graph work.
 
 The normal startup warmup is skipped automatically in longhaul mode. Routed expert weights are not read until the first real decode.
@@ -30,53 +24,23 @@ The normal startup warmup is skipped automatically in longhaul mode. Routed expe
 
 Longhaul currently requires:
 
-- all repeating layers on CPU, or all repeating layers on Metal
-- Qwen3.5 MoE, Laguna, or Inkling architecture
+- macOS with the Metal backend
+- Qwen3.5 MoE or Laguna architecture
+- all repeating model layers assigned to Metal
 - text generation without embeddings or LoRA adapters
 
-CPU mode is available wherever the CPU backend is supported. Metal mode requires
-macOS. Longhaul does not restrict the GGUF quantization type; individual tensor
-types must still be supported by the selected compute backend.
+Longhaul does not restrict the GGUF quantization type. Individual tensor types must still be supported by Metal.
 
-MTP/speculative decoding, tensor validation during loading, vocabulary-only
-loading, and mixed CPU/GPU repeating-layer placement are not supported. Both
-single-file and split GGUF models are supported; routed expert tensors are read
-from the shard that owns each tensor.
+MTP/speculative decoding, tensor validation during loading, vocabulary-only loading, and CPU or mixed CPU/Metal layer placement are not supported. Both single-file and split GGUF models are supported; routed expert tensors are read from the shard that owns each tensor.
 
-The cache budget covers the compact routed-expert tensors. It does not include
-dense weights, attention weights, the KV cache, graph allocations, or temporary
-staging buffers. CPU expert caches use the standard directly writable tensor
-layout so individual slots can be replaced; optimized CPU buffer selection
-continues to apply to all non-streamed weights.
-
-File-cache control is best effort and is outside the explicit cache budget.
-macOS disables caching for streamed files where supported, and Linux advises the
-kernel to discard completed reads. Windows may retain file data in its system
-cache. Windows reads from one GGUF shard are serialized to preserve positional
-read correctness, while POSIX systems retain concurrent `pread` operations.
+The cache budget covers the compact routed-expert tensors. It does not include dense weights, attention weights, the KV cache, graph allocations, or the temporary buffer used for one expert read. Disk reads bypass the macOS unified file cache where supported, avoiding a second long-lived copy of streamed weights in system RAM.
 
 This implementation synchronizes at each routed MoE layer to discover the selected experts, populate missing cache slots, and continue execution with cache-local expert IDs. Storage speed and expert reuse therefore have a large effect on generation speed.
 
 Expert IDs are planned as a batch at each synchronization point. Experts already
 needed by that batch are protected from eviction, duplicate IDs are loaded only
-once, and independent expert slices are read concurrently where the platform
-supports positional reads. CPU and shared Metal buffers are populated directly;
-private Metal buffers use a staged fallback.
-
-Inkling keeps its two shared experts resident and streams only the routed
-256-expert banks. Inkling-Small selects six routed experts per token. For the
-current three-shard UD-IQ1_S model, a 2 GiB cache provides seven routed-expert
-slots and therefore executes each MoE layer in one stage:
-
-```sh
-llama-cli \
-    --model /path/to/Inkling-Small-UD-IQ1_S-00001-of-00003.gguf \
-    --longhaul \
-    --longhaul-cache 2
-```
-
-Add `--n-gpu-layers 0` to run entirely on CPU; on macOS the default all-Metal
-placement is supported.
+once, and independent expert slices are read concurrently on shared-memory Metal
+devices. Private Metal buffers use a staged fallback.
 
 ## Benchmarking prompt processing
 
@@ -87,7 +51,7 @@ llama-bench \
     --model /path/to/model.gguf \
     --load-mode longhaul \
     --longhaul-cache 2 \
-    --n-gpu-layers 0 \
+    --n-gpu-layers 99 \
     --n-prompt 2048 \
     --n-gen 0
 ```

docs/token-cache.md

@@ -30,3 +30,24 @@ call `llama_token_cache_flush()`.
 
 The 5 GiB limit accounts for logical entry payloads. SQLite metadata and bounded
 journal files are not included.
+
+## Tokenizer longhaul
+
+Qwen3.5 MoE and Laguna models can move their reverse-token and BPE-merge indexes
+into a persistent, disk-backed index while keeping a bounded lookup working set:
+
+```text
+--tokenizer-longhaul
+--tokenizer-longhaul-cache N
+```
+
+`N` is the application-owned lookup cache budget in MiB and is required when
+the mode is enabled. The mode is independent of model `--longhaul`, so the two
+can be used together. The finite tokenizer index builds in the background while
+model loading continues. If tokenizer work arrives first, it waits for atomic
+index publication; build or storage failures are reported instead of silently
+falling back to an unbounded tokenizer.
+
+The persistent index uses the directory selected by `--token-cache-dir` and is
+subject to `--token-cache-size`. Tokenized-text results continue to be cached
+on demand and therefore do not have a finite completion point.

include/llama.h

@@ -360,6 +360,8 @@ extern "C" {
         bool no_alloc;        // only load metadata and simulate memory allocations
 
         uint64_t longhaul_cache_bytes; // routed expert cache size for LLAMA_LOAD_MODE_LONGHAUL
+        bool tokenizer_longhaul;       // page Qwen3.5/Laguna tokenizer indexes through a bounded cache
+        uint64_t tokenizer_longhaul_cache_bytes; // tokenizer-longhaul RAM cache budget
     };
 
     struct llama_sampler_seq_config {

src/CMakeLists.txt

@@ -38,6 +38,7 @@ add_library(llama
             llama-quant.cpp
             llama-sampler.cpp
             llama-token-cache.cpp
+            llama-tokenizer-longhaul.cpp
             llama-vocab.cpp
             unicode-data.cpp
             unicode.cpp

src/llama-arch.cpp

@@ -144,7 +144,6 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_TALKIE,           "talkie"           },
     { LLM_ARCH_MELLUM,           "mellum"           },
     { LLM_ARCH_NANBEIGE,         "nanbeige"         },
-    { LLM_ARCH_INKLING,          "inkling"          },
     { LLM_ARCH_UNKNOWN,          "(unknown)"        },
 };
 
@@ -321,15 +320,6 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
     { LLM_KV_NORM_BEFORE_FC,        "%s.norm_before_fc"       },
 
     { LLM_KV_SHORTCONV_L_CACHE, "%s.shortconv.l_cache" },
-    { LLM_KV_INKLING_D_REL,               "%s.d_rel"               },
-    { LLM_KV_INKLING_REL_EXTENT,          "%s.rel_extent"          },
-    { LLM_KV_INKLING_REL_EXTENT_SWA,      "%s.rel_extent_swa"      },
-    { LLM_KV_INKLING_SHORTCONV_KERNEL,    "%s.shortconv_kernel"    },
-    { LLM_KV_INKLING_DENSE_BLOCK_COUNT,   "%s.dense_block_count"   },
-    { LLM_KV_INKLING_LOGIT_SCALE_DENOM,   "%s.logit_scale_denom"   },
-    { LLM_KV_INKLING_LOG_SCALING_N_FLOOR, "%s.log_scaling_n_floor" },
-    { LLM_KV_INKLING_LOG_SCALING_ALPHA,   "%s.log_scaling_alpha"   },
-    { LLM_KV_INKLING_UNPADDED_VOCAB_SIZE, "%s.unpadded_vocab_size" },
     // sentence-transformers dense modules feature dims
     { LLM_KV_DENSE_2_FEAT_IN,        "%s.dense_2_feat_in"  },
     { LLM_KV_DENSE_2_FEAT_OUT,       "%s.dense_2_feat_out" },
@@ -602,19 +592,9 @@ static const std::map<llm_tensor, const char *> LLM_TENSOR_NAMES = {
     { LLM_TENSOR_SHORTCONV_CONV,                         "blk.%d.shortconv.conv" },
     { LLM_TENSOR_SHORTCONV_INPROJ,                       "blk.%d.shortconv.in_proj" },
     { LLM_TENSOR_SHORTCONV_OUTPROJ,                      "blk.%d.shortconv.out_proj" },
-    { LLM_TENSOR_ATTN_R,                                 "blk.%d.attn_r" },
-    { LLM_TENSOR_ATTN_REL_PROJ,                          "blk.%d.attn_rel_proj" },
-    { LLM_TENSOR_SHORTCONV_K,                            "blk.%d.shortconv_k" },
-    { LLM_TENSOR_SHORTCONV_V,                            "blk.%d.shortconv_v" },
-    { LLM_TENSOR_SHORTCONV_ATTN,                         "blk.%d.shortconv_attn" },
-    { LLM_TENSOR_SHORTCONV_MLP,                          "blk.%d.shortconv_mlp" },
-    { LLM_TENSOR_FFN_GSCALE,                             "blk.%d.ffn_gscale" },
     { LLM_TENSOR_FFN_GATE_CHEXPS,                        "blk.%d.ffn_gate_chexps" },
     { LLM_TENSOR_FFN_DOWN_CHEXPS,                        "blk.%d.ffn_down_chexps" },
     { LLM_TENSOR_FFN_UP_CHEXPS,                          "blk.%d.ffn_up_chexps" },
-    { LLM_TENSOR_FFN_GATE_SHEXPS,                        "blk.%d.ffn_gate_shexp" },
-    { LLM_TENSOR_FFN_DOWN_SHEXPS,                        "blk.%d.ffn_down_shexp" },
-    { LLM_TENSOR_FFN_UP_SHEXPS,                          "blk.%d.ffn_up_shexp" },
     { LLM_TENSOR_VISEXP_ATTN_QKV,                        "blk.%d.vis_attn_qkv" },
     { LLM_TENSOR_VISEXP_ATTN_OUT,                        "blk.%d.vis_attn_output" },
     { LLM_TENSOR_VISEXP_FFN_GATE,                        "blk.%d.vis_gate" },
@@ -815,9 +795,6 @@ static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = {
     {LLM_TENSOR_FFN_UP_EXPS,                {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT_ID}},
     {LLM_TENSOR_FFN_GATE_UP_EXPS,           {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT_ID}},
     {LLM_TENSOR_FFN_DOWN_CHEXPS,            {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT_ID}},
-    {LLM_TENSOR_FFN_DOWN_SHEXPS,            {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT_ID}},
-    {LLM_TENSOR_FFN_GATE_SHEXPS,            {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT_ID}},
-    {LLM_TENSOR_FFN_UP_SHEXPS,              {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT_ID}},
     {LLM_TENSOR_FFN_GATE_CHEXPS,            {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT_ID}},
     {LLM_TENSOR_FFN_UP_CHEXPS,              {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT_ID}},
     {LLM_TENSOR_FFN_EXP_PROBS_B,            {LLM_TENSOR_LAYER_REPEATING, GGML_OP_ADD}},
@@ -859,13 +836,6 @@ static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = {
     {LLM_TENSOR_SHORTCONV_CONV,             {LLM_TENSOR_LAYER_REPEATING, GGML_OP_SSM_CONV}},
     {LLM_TENSOR_SHORTCONV_INPROJ,           {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
     {LLM_TENSOR_SHORTCONV_OUTPROJ,          {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
-    {LLM_TENSOR_ATTN_R,                     {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
-    {LLM_TENSOR_ATTN_REL_PROJ,              {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
-    {LLM_TENSOR_SHORTCONV_K,                {LLM_TENSOR_LAYER_REPEATING, GGML_OP_SSM_CONV}},
-    {LLM_TENSOR_SHORTCONV_V,                {LLM_TENSOR_LAYER_REPEATING, GGML_OP_SSM_CONV}},
-    {LLM_TENSOR_SHORTCONV_ATTN,             {LLM_TENSOR_LAYER_REPEATING, GGML_OP_SSM_CONV}},
-    {LLM_TENSOR_SHORTCONV_MLP,              {LLM_TENSOR_LAYER_REPEATING, GGML_OP_SSM_CONV}},
-    {LLM_TENSOR_FFN_GSCALE,                 {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
     {LLM_TENSOR_VISEXP_ATTN_QKV,            {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
     {LLM_TENSOR_VISEXP_ATTN_OUT,            {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
     {LLM_TENSOR_VISEXP_FFN_GATE,            {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
@@ -998,7 +968,6 @@ bool llm_arch_is_hybrid(const llm_arch & arch) {
         case LLM_ARCH_KIMI_LINEAR:
         case LLM_ARCH_QWEN35:
         case LLM_ARCH_QWEN35MOE:
-        case LLM_ARCH_INKLING:
             return true;
         default:
             return false;
@@ -1055,7 +1024,6 @@ bool llm_arch_supports_sm_tensor(const llm_arch & arch) {
         case LLM_ARCH_MINIMAX_M3:
         case LLM_ARCH_MISTRAL4:
         case LLM_ARCH_KIMI_LINEAR:
-        case LLM_ARCH_INKLING:
             return false;
         default:
             return true;

src/llama-arch.h

@@ -149,7 +149,6 @@ enum llm_arch {
     LLM_ARCH_MINIMAX_M3,
     LLM_ARCH_DFLASH,
     LLM_ARCH_NANBEIGE,
-    LLM_ARCH_INKLING,
     LLM_ARCH_UNKNOWN,
 };
 
@@ -367,15 +366,6 @@ enum llm_kv {
     LLM_KV_NORM_BEFORE_FC,
 
     LLM_KV_SHORTCONV_L_CACHE,
-    LLM_KV_INKLING_D_REL,
-    LLM_KV_INKLING_REL_EXTENT,
-    LLM_KV_INKLING_REL_EXTENT_SWA,
-    LLM_KV_INKLING_SHORTCONV_KERNEL,
-    LLM_KV_INKLING_DENSE_BLOCK_COUNT,
-    LLM_KV_INKLING_LOGIT_SCALE_DENOM,
-    LLM_KV_INKLING_LOG_SCALING_N_FLOOR,
-    LLM_KV_INKLING_LOG_SCALING_ALPHA,
-    LLM_KV_INKLING_UNPADDED_VOCAB_SIZE,
 
     LLM_KV_XIELU_ALPHA_N,
     LLM_KV_XIELU_ALPHA_P,
@@ -444,9 +434,6 @@ enum llm_tensor {
     LLM_TENSOR_FFN_DOWN_CHEXPS,
     LLM_TENSOR_FFN_GATE_CHEXPS,
     LLM_TENSOR_FFN_UP_CHEXPS,
-    LLM_TENSOR_FFN_DOWN_SHEXPS,
-    LLM_TENSOR_FFN_GATE_SHEXPS,
-    LLM_TENSOR_FFN_UP_SHEXPS,
     LLM_TENSOR_FFN_EXP_PROBS_B,
     LLM_TENSOR_FFN_LATENT_DOWN,
     LLM_TENSOR_FFN_LATENT_UP,
@@ -608,13 +595,6 @@ enum llm_tensor {
     LLM_TENSOR_SHORTCONV_CONV,
     LLM_TENSOR_SHORTCONV_INPROJ,
     LLM_TENSOR_SHORTCONV_OUTPROJ,
-    LLM_TENSOR_ATTN_R,
-    LLM_TENSOR_ATTN_REL_PROJ,
-    LLM_TENSOR_SHORTCONV_K,
-    LLM_TENSOR_SHORTCONV_V,
-    LLM_TENSOR_SHORTCONV_ATTN,
-    LLM_TENSOR_SHORTCONV_MLP,
-    LLM_TENSOR_FFN_GSCALE,
     LLM_TENSOR_VISEXP_ATTN_QKV,
     LLM_TENSOR_VISEXP_ATTN_OUT,
     LLM_TENSOR_VISEXP_FFN_GATE,

src/llama-graph.cpp

@@ -1765,119 +1765,6 @@ ggml_tensor * llm_graph_context::build_ffn(
     return cur;
 }
 
-ggml_tensor * llm_graph_context::build_moe_experts(
-         ggml_tensor * cur,
-         ggml_tensor * selected_experts,
-         ggml_tensor * weights,
-         ggml_tensor * up_exps,
-         ggml_tensor * gate_exps,
-         ggml_tensor * down_exps,
-             int64_t   n_expert_used,
-     llm_ffn_op_type   type_op,
-                 int   il) const {
-    GGML_ASSERT(cur != nullptr);
-    GGML_ASSERT(selected_experts != nullptr);
-    GGML_ASSERT(weights != nullptr);
-    GGML_ASSERT(up_exps != nullptr && down_exps != nullptr);
-
-    const int64_t n_embd   = cur->ne[0];
-    const int64_t n_tokens = cur->ne[1];
-
-    // Ensure routing and weights are available before a longhaul cache
-    // callback resolves and loads the first expert chunk.
-    ggml_build_forward_expand(gf, weights);
-
-    ggml_tensor * moe_inp = ggml_reshape_3d(ctx0, cur, n_embd, 1, n_tokens);
-    const int64_t chunk_size = longhaul
-        ? std::min<int64_t>(n_expert_used, longhaul->capacity())
-        : n_expert_used;
-    GGML_ASSERT(chunk_size > 0);
-
-    ggml_tensor * moe_out = nullptr;
-    int chunk = 0;
-
-    for (int64_t first = 0; first < n_expert_used; first += chunk_size, ++chunk) {
-        const int64_t count = std::min<int64_t>(chunk_size, n_expert_used - first);
-
-        ggml_tensor * selected_chunk = selected_experts;
-        ggml_tensor * weights_chunk  = weights;
-        if (count != n_expert_used) {
-            selected_chunk = ggml_view_2d(
-                ctx0, selected_experts, count, n_tokens,
-                selected_experts->nb[1], first*selected_experts->nb[0]);
-            weights_chunk = ggml_view_3d(
-                ctx0, weights, 1, count, n_tokens,
-                weights->nb[1], weights->nb[2], first*weights->nb[1]);
-        }
-
-        ggml_tensor * selected_moe = selected_chunk;
-        if (longhaul) {
-            selected_moe = ggml_dup(ctx0, selected_chunk);
-            ggml_format_name(selected_moe, "longhaul.remap.%d.%d", il, chunk);
-            ggml_build_forward_expand(gf, selected_moe);
-        }
-
-        ggml_tensor * up = build_lora_mm_id(up_exps, moe_inp, selected_moe);
-        ggml_tensor * act = gate_exps
-            ? build_lora_mm_id(gate_exps, moe_inp, selected_moe)
-            : up;
-
-        switch (type_op) {
-            case LLM_FFN_SILU:
-                act = gate_exps
-                    ? ggml_swiglu_split(ctx0, act, up)
-                    : ggml_silu(ctx0, act);
-                break;
-            case LLM_FFN_GELU:
-                act = gate_exps
-                    ? ggml_geglu_split(ctx0, act, up)
-                    : ggml_gelu(ctx0, act);
-                break;
-            case LLM_FFN_RELU:
-                act = gate_exps
-                    ? ggml_reglu_split(ctx0, act, up)
-                    : ggml_relu(ctx0, act);
-                break;
-            default:
-                GGML_ABORT("unsupported routed expert activation");
-        }
-
-        ggml_tensor * experts = build_lora_mm_id(
-            down_exps, act, selected_moe);
-        experts = ggml_mul(ctx0, experts, weights_chunk);
-        ggml_build_forward_expand(gf, experts);
-
-        ggml_tensor * chunk_out = nullptr;
-        for (int64_t i = 0; i < count; ++i) {
-            ggml_tensor * expert = ggml_view_2d(
-                ctx0, experts, n_embd, n_tokens,
-                experts->nb[2], i*experts->nb[1]);
-            ggml_build_forward_expand(gf, expert);
-            chunk_out = chunk_out ? ggml_add(ctx0, chunk_out, expert) : expert;
-            ggml_build_forward_expand(gf, chunk_out);
-        }
-        if (count == 1) {
-            chunk_out = ggml_cont(ctx0, chunk_out);
-        }
-
-        ggml_tensor * contribution = chunk_out;
-        if (longhaul) {
-            ggml_tensor * release = ggml_cont(ctx0, chunk_out);
-            ggml_format_name(release, "longhaul.release.%d.%d", il, chunk);
-            ggml_build_forward_expand(gf, release);
-            if (chunk_size < n_expert_used) {
-                contribution = release;
-            }
-        }
-
-        moe_out = moe_out ? ggml_add(ctx0, moe_out, contribution) : contribution;
-        ggml_build_forward_expand(gf, moe_out);
-    }
-
-    cb(moe_out, "ffn_moe_out", il);
-    return moe_out;
-}
-
 ggml_tensor * llm_graph_context::build_moe_ffn(
          ggml_tensor * cur,
          ggml_tensor * gate_inp,

src/llama-graph.h

@@ -1000,20 +1000,6 @@ struct llm_graph_context {
        llm_ffn_gate_type   type_gate,
                      int   il) const;
 
-    // Execute experts for an architecture that supplies its own selected IDs
-    // and weights. In longhaul mode this transparently stages those experts
-    // through the bounded cache.
-    ggml_tensor * build_moe_experts(
-             ggml_tensor * cur,
-             ggml_tensor * selected_experts,
-             ggml_tensor * weights,
-             ggml_tensor * up_exps,
-             ggml_tensor * gate_exps,
-             ggml_tensor * down_exps,
-                 int64_t   n_expert_used,
-         llm_ffn_op_type   type_op,
-                     int   il) const;
-
     // build MoE FFN without bias tensors
     ggml_tensor * build_moe_ffn(
              ggml_tensor * cur,

src/llama-hparams.cpp

@@ -191,10 +191,6 @@ uint32_t llama_hparams::n_embd_k_idx(uint32_t il) const {
 }
 
 uint32_t llama_hparams::n_embd_r() const {
-    if (n_embd_r_impl != 0) {
-        return n_embd_r_impl;
-    }
-
     if (wkv_head_size != 0) {
         // for RWKV models
         return token_shift_count * n_embd;

src/llama-hparams.h

@@ -80,18 +80,6 @@ struct llama_hparams {
 
     uint32_t n_shortconv_l_cache  = 0;
 
-    // Explicit recurrent-state size override. Inkling packs four independent
-    // short-convolution histories into the recurrent cache for every layer.
-    uint32_t n_embd_r_impl = 0;
-
-    // Inkling relative-attention and output metadata.
-    uint32_t inkling_d_rel            = 0;
-    uint32_t inkling_rel_extent       = 0;
-    uint32_t inkling_rel_extent_swa   = 0;
-    uint32_t inkling_log_n_floor      = 0;
-    float    inkling_log_alpha        = 0.0f;
-    uint32_t inkling_unpadded_n_vocab = 0;
-
     std::array<uint32_t, LLAMA_MAX_LAYERS> n_head_arr;
     std::array<uint32_t, LLAMA_MAX_LAYERS> n_head_kv_arr;
     std::array<uint32_t, LLAMA_MAX_LAYERS> n_ff_arr;

src/llama-kv-cache.cpp

@@ -1931,37 +1931,6 @@ void llama_kv_cache::set_input_pos_bucket(ggml_tensor * dst, const llama_ubatch
     }
 }
 
-void llama_kv_cache::set_input_pos_rel_flat(
-        ggml_tensor * dst,
-        const llama_ubatch * ubatch,
-        uint32_t extent) const {
-    GGML_ASSERT(ggml_backend_buffer_is_host(dst->buffer));
-    GGML_ASSERT(dst->ne[1] == (int64_t) ubatch->n_tokens);
-
-    int32_t * data = (int32_t *) dst->data;
-    const int64_t n_kv = dst->ne[0];
-
-    // Each query owns an extent+1-row block in the flattened relative-logit
-    // table. The final row is the zero-bias sentinel for empty/out-of-band KV
-    // cells.
-    for (int64_t i = 0; i < (int64_t) ubatch->n_tokens; ++i) {
-        const llama_seq_id seq_id = ubatch->seq_id[i][0];
-        const auto & cells = v_cells[seq_to_stream[seq_id]];
-        const llama_pos p1 = ubatch->pos[i];
-
-        for (int64_t j = 0; j < n_kv; ++j) {
-            int32_t rel = (int32_t) extent;
-            if (!cells.is_empty(j)) {
-                const llama_pos d = p1 - cells.pos_get(j);
-                if (d >= 0 && d < (llama_pos) extent) {
-                    rel = (int32_t) d;
-                }
-            }
-            data[i*n_kv + j] = (int32_t) (i*(extent + 1)) + rel;
-        }
-    }
-}
-
 void llama_kv_cache::set_input_k_rot(ggml_tensor * dst) const {
     GGML_ASSERT(ggml_backend_buffer_is_host(dst->buffer));
 
@@ -2927,13 +2896,6 @@ void llama_kv_cache_context::set_input_pos_bucket(ggml_tensor * dst, const llama
     kv->set_input_pos_bucket(dst, ubatch);
 }
 
-void llama_kv_cache_context::set_input_pos_rel_flat(
-        ggml_tensor * dst,
-        const llama_ubatch * ubatch,
-        uint32_t extent) const {
-    kv->set_input_pos_rel_flat(dst, ubatch, extent);
-}
-
 void llama_kv_cache_context::set_input_k_rot(ggml_tensor * dst) const {
     kv->set_input_k_rot(dst);
 }

src/llama-kv-cache.h

@@ -215,8 +215,6 @@ public:
 
     void set_input_kq_mask   (ggml_tensor * dst, const llama_ubatch * ubatch, bool causal_attn) const;
     void set_input_pos_bucket(ggml_tensor * dst, const llama_ubatch * ubatch) const;
-    void set_input_pos_rel_flat(
-        ggml_tensor * dst, const llama_ubatch * ubatch, uint32_t extent) const;
 
     void set_input_k_rot(ggml_tensor * dst) const;
     void set_input_v_rot(ggml_tensor * dst) const;
@@ -407,8 +405,6 @@ public:
     void set_input_k_shift   (ggml_tensor * dst) const;
     void set_input_kq_mask   (ggml_tensor * dst, const llama_ubatch * ubatch, bool causal_attn) const;
     void set_input_pos_bucket(ggml_tensor * dst, const llama_ubatch * ubatch) const;
-    void set_input_pos_rel_flat(
-        ggml_tensor * dst, const llama_ubatch * ubatch, uint32_t extent) const;
 
     void set_input_k_rot(ggml_tensor * dst) const;
     void set_input_v_rot(ggml_tensor * dst) const;

src/llama-mmap.cpp

@@ -9,7 +9,6 @@
 #include <stdexcept>
 #include <cerrno>
 #include <algorithm>
-#include <mutex>
 
 #ifdef __has_include
     #if __has_include(<unistd.h>)
@@ -67,12 +66,8 @@ static std::string llama_format_win_err(DWORD err) {
 // llama_file
 
 struct llama_file::impl {
-    bool no_cache = false;
-
 #if defined(_WIN32)
     HANDLE fp_win32;
-    std::mutex read_at_mutex;
-
     std::string GetErrorMessageWin32(DWORD error_code) const {
         std::string ret;
         LPSTR lpMsgBuf = NULL;
@@ -438,10 +433,6 @@ void llama_file::read_raw_unsafe(void * ptr, size_t len) { pimpl->read_raw_unsaf
 
 void llama_file::read_at(void * ptr, size_t len, size_t offset) {
 #ifdef _WIN32
-    // The Windows FILE pointer is shared by all longhaul I/O workers. Keep the
-    // seek and read together so concurrent jobs cannot change each other's
-    // offsets. POSIX uses pread below and does not need this serialization.
-    std::lock_guard<std::mutex> lock(pimpl->read_at_mutex);
     seek(offset, SEEK_SET);
     read_raw(ptr, len);
 #else
@@ -457,13 +448,6 @@ void llama_file::read_at(void * ptr, size_t len, size_t offset) {
         }
         n_read += result;
     }
-#if defined(__linux__) && defined(POSIX_FADV_DONTNEED)
-    if (pimpl->no_cache) {
-        // Longhaul owns a bounded expert cache, so avoid retaining a second
-        // long-lived copy of completed reads in the kernel page cache.
-        (void) posix_fadvise(file_id(), (off_t) offset, (off_t) len, POSIX_FADV_DONTNEED);
-    }
-#endif
 #endif
 }
 
@@ -472,12 +456,6 @@ void llama_file::set_no_cache() const {
     if (fcntl(file_id(), F_NOCACHE, 1) != 0) {
         throw std::runtime_error(format("failed to disable file cache: %s", strerror(errno)));
     }
-#elif defined(__linux__) && defined(POSIX_FADV_RANDOM)
-    pimpl->no_cache = true;
-    const int err = posix_fadvise(file_id(), 0, 0, POSIX_FADV_RANDOM);
-    if (err != 0) {
-        LLAMA_LOG_WARN("%s: failed to set random-access file advice: %s\n", __func__, strerror(err));
-    }
 #endif
 }
 

src/llama-model-loader.cpp

@@ -1220,7 +1220,6 @@ struct ggml_tensor * llama_model_loader::create_tensor(
         }
 
         ggml_backend_buffer_type_t buft = nullptr;
-        ggml_backend_buffer_type_t requested_override = nullptr;
 
         // check overrides
         if (tensor_buft_overrides) {
@@ -1228,7 +1227,6 @@ struct ggml_tensor * llama_model_loader::create_tensor(
             for (const auto * overrides = tensor_buft_overrides; overrides->pattern != nullptr; ++overrides) {
                 std::regex pattern(overrides->pattern);
                 if (std::regex_search(tensor_name, pattern)) {
-                    requested_override = 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, buft_list_cpu);
@@ -1258,25 +1256,6 @@ struct ggml_tensor * llama_model_loader::create_tensor(
             }
         }
 
-        // Streamed CPU experts are updated one expert slice at a time. CPU
-        // repack and accelerator buffers only accept whole-tensor updates, so
-        // keep these cache tensors in the standard, directly writable layout.
-        if ((flags & TENSOR_LONGHAUL) && buft_list == buft_list_cpu) {
-            auto * cpu_dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU);
-            if (!cpu_dev) {
-                throw std::runtime_error("no CPU backend found");
-            }
-            ggml_backend_buffer_type_t cpu_buft = ggml_backend_dev_buffer_type(cpu_dev);
-            if (requested_override != nullptr &&
-                requested_override != ggml_backend_cpu_buffer_type() &&
-                requested_override != cpu_buft) {
-                throw std::runtime_error(format(
-                    "longhaul CPU expert tensor '%s' has an incompatible buffer override",
-                    tn.str().c_str()));
-            }
-            buft = cpu_buft;
-        }
-
         // avoid using a host buffer when using mmap
         auto * buft_dev = ggml_backend_buft_get_device(buft);
         if (use_mmap && buft_dev && buft == ggml_backend_dev_host_buffer_type(buft_dev)) {

src/llama-model-saver.cpp

@@ -29,7 +29,6 @@ bool llama_model_saver_supports_arch(llm_arch arch) {
         case LLM_ARCH_STEP35:
         case LLM_ARCH_MELLUM:
         case LLM_ARCH_LAGUNA:
-        case LLM_ARCH_INKLING:
             return false;
         default:
             return true;

src/llama-model.cpp

@@ -312,8 +312,6 @@ static llama_model * llama_model_mapping(llm_arch arch, const llama_model_params
             return new llama_model_kimi_linear(params);
         case LLM_ARCH_STEP35:
             return new llama_model_step35(params);
-        case LLM_ARCH_INKLING:
-            return new llama_model_inkling(params);
         default:
             throw std::runtime_error(std::string("unsupported model architecture: '") + llm_arch_name(arch) + "'");
     }
@@ -1251,7 +1249,7 @@ void llama_model_base::load_hparams(llama_model_loader & ml) {
 void llama_model_base::load_vocab(llama_model_loader & ml) {
     const auto kv = LLM_KV(arch);
 
-    vocab.load(ml, kv);
+    vocab.load(ml, kv, params.tokenizer_longhaul, params.tokenizer_longhaul_cache_bytes);
 }
 
 bool llama_model_base::load_tensors(llama_model_loader & ml) {
@@ -1358,8 +1356,11 @@ bool llama_model_base::load_tensors(llama_model_loader & ml) {
         }
 
         if (params.load_mode == LLAMA_LOAD_MODE_LONGHAUL) {
-            if (arch != LLM_ARCH_QWEN35MOE && arch != LLM_ARCH_LAGUNA && arch != LLM_ARCH_INKLING) {
-                throw std::runtime_error("longhaul currently requires a qwen35moe, laguna, or inkling model");
+#if !defined(__APPLE__)
+            throw std::runtime_error("longhaul is only supported on macOS");
+#endif
+            if (arch != LLM_ARCH_QWEN35MOE && arch != LLM_ARCH_LAGUNA) {
+                throw std::runtime_error("longhaul currently requires a qwen35moe or laguna model");
             }
             if (hparams.n_layer_nextn != 0) {
                 throw std::runtime_error("longhaul does not support MTP tensors");
@@ -1367,27 +1368,13 @@ bool llama_model_base::load_tensors(llama_model_loader & ml) {
             if (params.check_tensors || params.no_alloc || params.vocab_only) {
                 throw std::runtime_error("longhaul does not support check-tensors, no-alloc, or vocab-only loading");
             }
-            bool all_cpu = true;
-            bool all_metal = true;
             for (int il = 0; il < n_layer_all; ++il) {
                 ggml_backend_dev_t dev = pimpl->dev_layer[il].dev;
-                const char * backend = ggml_backend_reg_name(ggml_backend_dev_backend_reg(dev));
-                all_cpu = all_cpu && ggml_backend_dev_type(dev) == GGML_BACKEND_DEVICE_TYPE_CPU;
-                all_metal = all_metal &&
-                    ggml_backend_dev_type(dev) == GGML_BACKEND_DEVICE_TYPE_GPU &&
-                    strcmp(backend, "MTL") == 0;
+                if (ggml_backend_dev_type(dev) != GGML_BACKEND_DEVICE_TYPE_GPU ||
+                    strcmp(ggml_backend_reg_name(ggml_backend_dev_backend_reg(dev)), "MTL") != 0) {
+                    throw std::runtime_error("longhaul requires every model layer on Metal");
+                }
             }
-            if (!all_cpu && !all_metal) {
-                throw std::runtime_error(
-                    "longhaul requires every repeating model layer on CPU or Metal; "
-                    "use --n-gpu-layers 0 for CPU mode");
-            }
-#if !defined(__APPLE__)
-            if (all_metal) {
-                throw std::runtime_error("longhaul Metal mode is only supported on macOS");
-            }
-#endif
-            LLAMA_LOG_INFO("%s: longhaul using %s expert cache\n", __func__, all_cpu ? "CPU" : "Metal");
             ml.configure_longhaul(params.longhaul_cache_bytes, n_expert, n_layer_all);
             if (ml.longhaul_slots < (size_t) n_expert_used) {
                 LLAMA_LOG_WARN("%s: longhaul cache has %zu slots for %lld selected experts; "
@@ -2167,7 +2154,7 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params,
                     // layer filters, so pick the right one here
                     llama_memory_hybrid::layer_filter_cb filter_attn = nullptr;
                     llama_memory_hybrid::layer_filter_cb filter_recr = nullptr;
-                    if (arch == LLM_ARCH_FALCON_H1 || arch == LLM_ARCH_INKLING) {
+                    if (arch == LLM_ARCH_FALCON_H1) {
                         filter_attn = [&](uint32_t) { return true; };
                         filter_recr = [&](uint32_t) { return true; };
                     } else if (arch == LLM_ARCH_NEMOTRON_H || arch == LLM_ARCH_NEMOTRON_H_MOE) {
@@ -2393,6 +2380,8 @@ llama_model_params llama_model_default_params() {
         /*.no_host                     =*/ false,
         /*.no_alloc                    =*/ false,
         /*.longhaul_cache_bytes        =*/ 0,
+        /*.tokenizer_longhaul          =*/ false,
+        /*.tokenizer_longhaul_cache_bytes =*/ 0,
     };
 
     return result;
@@ -2508,7 +2497,6 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
         case LLM_ARCH_NEMOTRON_H:
         case LLM_ARCH_NEMOTRON_H_MOE:
         case LLM_ARCH_KIMI_LINEAR:
-        case LLM_ARCH_INKLING:
             return LLAMA_ROPE_TYPE_NONE;
 
         // use what we call a normal RoPE, operating on pairs of consecutive head values

src/llama-model.h

@@ -534,15 +534,6 @@ struct llama_layer {
     struct llama_layer_shortconv shortconv;
 
     struct llama_layer_nextn nextn;
-
-    // Inkling relative attention and short-convolution weights.
-    struct ggml_tensor * wr             = nullptr;
-    struct ggml_tensor * attn_rel_proj  = nullptr;
-    struct ggml_tensor * shortconv_k    = nullptr;
-    struct ggml_tensor * shortconv_v    = nullptr;
-    struct ggml_tensor * shortconv_attn = nullptr;
-    struct ggml_tensor * shortconv_mlp  = nullptr;
-    struct ggml_tensor * ffn_gscale     = nullptr;
 };
 
 struct llama_device {

src/llama-token-cache.cpp

@@ -214,6 +214,10 @@ public:
         if (!ensure_open()) return params.capacity_bytes == 0;
         std::lock_guard<std::mutex> db_lock(db_mutex);
         if (sqlite3_exec(db, "DELETE FROM entries; PRAGMA incremental_vacuum;", nullptr, nullptr, nullptr) != SQLITE_OK) return db_error();
+        std::error_code ec;
+        fs::remove_all(fs::path(path).parent_path() / "tokenizer-indexes", ec);
+        if (ec) return false;
+        sqlite3_exec(db, "INSERT OR REPLACE INTO meta(key,value) VALUES('external_bytes',0)", nullptr, nullptr, nullptr);
         return true;
     }
 
@@ -239,6 +243,11 @@ public:
                 result.bytes_used = sqlite3_column_int64(stmt, 1);
             }
             sqlite3_finalize(stmt);
+            if (sqlite3_prepare_v2(db, "SELECT value FROM meta WHERE key='external_bytes'", -1, &stmt, nullptr) == SQLITE_OK &&
+                    sqlite3_step(stmt) == SQLITE_ROW) {
+                result.bytes_used += static_cast<uint64_t>(sqlite3_column_int64(stmt, 0));
+            }
+            sqlite3_finalize(stmt);
         }
         return result;
     }
@@ -247,6 +256,45 @@ public:
         if (hit) ++stats.tokenizer_hits; else ++stats.tokenizer_misses;
     }
 
+    std::string database_path() {
+        return ensure_open() ? path : std::string();
+    }
+
+    uint64_t capacity() const {
+        return params.capacity_bytes;
+    }
+
+    void set_external_bytes(uint64_t bytes) {
+        if (!ensure_open()) return;
+        std::lock_guard<std::mutex> db_lock(db_mutex);
+        sqlite3_exec(db, "BEGIN IMMEDIATE", nullptr, nullptr, nullptr);
+        sqlite3_stmt * stmt = nullptr;
+        sqlite3_prepare_v2(db, "INSERT OR REPLACE INTO meta(key,value) VALUES('external_bytes',?1)", -1, &stmt, nullptr);
+        sqlite3_bind_int64(stmt, 1, static_cast<sqlite3_int64>(bytes));
+        sqlite3_step(stmt);
+        sqlite3_finalize(stmt);
+
+        uint64_t total = 0;
+        sqlite3_prepare_v2(db, "SELECT COALESCE(SUM(size),0) FROM entries", -1, &stmt, nullptr);
+        if (sqlite3_step(stmt) == SQLITE_ROW) total = sqlite3_column_int64(stmt, 0);
+        sqlite3_finalize(stmt);
+        const uint64_t entry_capacity = bytes >= params.capacity_bytes ? 0 : params.capacity_bytes - bytes;
+        while (total > entry_capacity) {
+            uint64_t victim_size = 0;
+            sqlite3_prepare_v2(db, "SELECT size FROM entries ORDER BY last_access LIMIT 1", -1, &stmt, nullptr);
+            if (sqlite3_step(stmt) == SQLITE_ROW) victim_size = sqlite3_column_int64(stmt, 0);
+            sqlite3_finalize(stmt);
+            if (victim_size == 0 || sqlite3_exec(db,
+                    "DELETE FROM entries WHERE rowid=(SELECT rowid FROM entries ORDER BY last_access LIMIT 1)",
+                    nullptr, nullptr, nullptr) != SQLITE_OK) {
+                break;
+            }
+            total -= std::min(total, victim_size);
+            ++stats.evictions;
+        }
+        sqlite3_exec(db, "COMMIT", nullptr, nullptr, nullptr);
+    }
+
 private:
     static int64_t now_tick() {
         return std::chrono::duration_cast<std::chrono::microseconds>(
@@ -311,13 +359,22 @@ private:
         sqlite3_bind_int64(stmt, 1, static_cast<sqlite3_int64>(params.capacity_bytes));
         sqlite3_step(stmt);
         sqlite3_finalize(stmt);
+        uint64_t external_bytes = 0;
+        sqlite3_prepare_v2(db, "SELECT value FROM meta WHERE key='external_bytes'", -1, &stmt, nullptr);
+        if (sqlite3_step(stmt) == SQLITE_ROW) {
+            external_bytes = static_cast<uint64_t>(sqlite3_column_int64(stmt, 0));
+        }
+        sqlite3_finalize(stmt);
+        const uint64_t entry_capacity = external_bytes >= params.capacity_bytes
+            ? 0
+            : params.capacity_bytes - external_bytes;
         sqlite3_exec(db, "BEGIN IMMEDIATE", nullptr, nullptr, nullptr);
         while (true) {
             sqlite3_prepare_v2(db, "SELECT COALESCE(SUM(size),0) FROM entries", -1, &stmt, nullptr);
             uint64_t total = 0;
             if (sqlite3_step(stmt) == SQLITE_ROW) total = sqlite3_column_int64(stmt, 0);
             sqlite3_finalize(stmt);
-            if (total <= params.capacity_bytes) break;
+            if (total <= entry_capacity) break;
             if (sqlite3_exec(db,
                     "DELETE FROM entries WHERE rowid=(SELECT rowid FROM entries ORDER BY last_access LIMIT 1)",
                     nullptr, nullptr, nullptr) != SQLITE_OK) {
@@ -355,12 +412,19 @@ private:
         std::lock_guard<std::mutex> db_lock(db_mutex);
         if (sqlite3_exec(db, "BEGIN IMMEDIATE", nullptr, nullptr, nullptr) != SQLITE_OK) { db_error(); return; }
         uint64_t capacity = params.capacity_bytes;
+        uint64_t external_bytes = 0;
         sqlite3_stmt * capacity_stmt = nullptr;
         if (sqlite3_prepare_v2(db, "SELECT value FROM meta WHERE key='capacity'", -1, &capacity_stmt, nullptr) == SQLITE_OK &&
                 sqlite3_step(capacity_stmt) == SQLITE_ROW) {
             capacity = static_cast<uint64_t>(sqlite3_column_int64(capacity_stmt, 0));
         }
         sqlite3_finalize(capacity_stmt);
+        if (sqlite3_prepare_v2(db, "SELECT value FROM meta WHERE key='external_bytes'", -1, &capacity_stmt, nullptr) == SQLITE_OK &&
+                sqlite3_step(capacity_stmt) == SQLITE_ROW) {
+            external_bytes = static_cast<uint64_t>(sqlite3_column_int64(capacity_stmt, 0));
+        }
+        sqlite3_finalize(capacity_stmt);
+        capacity = external_bytes >= capacity ? 0 : capacity - external_bytes;
         if (item.payload.size() > capacity) {
             sqlite3_exec(db, "ROLLBACK", nullptr, nullptr, nullptr);
             return;
@@ -492,6 +556,18 @@ void llama_token_cache_note_tokenizer_hit(bool hit) {
     cache().note_tokenizer(hit);
 }
 
+std::string llama_token_cache_database_path() {
+    return cache().database_path();
+}
+
+uint64_t llama_token_cache_capacity() {
+    return cache().capacity();
+}
+
+void llama_token_cache_set_external_bytes(uint64_t bytes) {
+    cache().set_external_bytes(bytes);
+}
+
 llama_token_cache_params llama_token_cache_default_params() {
     return {nullptr, DEFAULT_CAPACITY};
 }

src/llama-token-cache.h

@@ -27,3 +27,8 @@ void llama_token_cache_store(
 
 void llama_token_cache_note_tokenizer_hit(bool hit);
 
+// Internal helpers used by tokenizer-longhaul artifacts.
+// Opening the cache here also fixes the configured path for the process.
+std::string llama_token_cache_database_path();
+uint64_t llama_token_cache_capacity();
+void llama_token_cache_set_external_bytes(uint64_t bytes);

src/llama-tokenizer-longhaul.cpp

@@ -0,0 +1,471 @@
+#include "llama-tokenizer-longhaul.h"
+
+#include "llama-token-cache.h"
+
+#include <sqlite3.h>
+
+#include <algorithm>
+#include <chrono>
+#include <filesystem>
+#include <list>
+#include <mutex>
+#include <thread>
+#include <unordered_map>
+
+#if defined(_WIN32)
+#  include <process.h>
+#else
+#  include <unistd.h>
+#endif
+
+namespace fs = std::filesystem;
+
+namespace {
+
+constexpr int INDEX_VERSION = 1;
+
+int process_id() {
+#if defined(_WIN32)
+    return _getpid();
+#else
+    return getpid();
+#endif
+}
+
+bool exec_sql(sqlite3 * db, const char * sql, std::string & error) {
+    char * message = nullptr;
+    if (sqlite3_exec(db, sql, nullptr, nullptr, &message) == SQLITE_OK) {
+        return true;
+    }
+    error = message ? message : sqlite3_errmsg(db);
+    sqlite3_free(message);
+    return false;
+}
+
+struct lock_directory {
+    fs::path path;
+    bool owned = false;
+
+    ~lock_directory() {
+        if (owned) {
+            std::error_code ec;
+            fs::remove(path, ec);
+        }
+    }
+};
+
+uint64_t tokenizer_artifact_bytes(const fs::path & directory) {
+    uint64_t total = 0;
+    std::error_code ec;
+    for (const auto & entry : fs::directory_iterator(directory, ec)) {
+        if (ec) break;
+        if (!entry.is_regular_file(ec) || entry.path().extension() != ".sqlite3") continue;
+        total += entry.file_size(ec);
+        if (ec) {
+            ec.clear();
+        }
+    }
+    return total;
+}
+
+} // namespace
+
+struct llama_tokenizer_longhaul_index::impl {
+    struct cache_entry {
+        int value;
+        uint64_t bytes;
+        std::list<std::string>::iterator position;
+    };
+
+    std::string source_key;
+    uint64_t capacity;
+    fs::path artifact;
+    sqlite3 * db = nullptr;
+    std::string id;
+    uint64_t merge_count = 0;
+
+    mutable std::mutex mutex;
+    mutable uint64_t used = 0;
+    mutable std::list<std::string> lru;
+    mutable std::unordered_map<std::string, cache_entry> values;
+
+    impl(std::string source_key, uint64_t capacity)
+        : source_key(std::move(source_key)), capacity(capacity) {
+    }
+
+    ~impl() {
+        if (db) {
+            sqlite3_close(db);
+        }
+    }
+
+    bool open_existing(std::string & error) {
+        if (!fs::exists(artifact)) {
+            return false;
+        }
+        sqlite3 * opened = nullptr;
+        if (sqlite3_open_v2(artifact.string().c_str(), &opened,
+                SQLITE_OPEN_READONLY | SQLITE_OPEN_FULLMUTEX, nullptr) != SQLITE_OK) {
+            error = opened ? sqlite3_errmsg(opened) : "failed to open tokenizer index";
+            if (opened) sqlite3_close(opened);
+            return false;
+        }
+
+        sqlite3_stmt * stmt = nullptr;
+        const char * sql =
+            "SELECT version, tokenizer_id, n_merges FROM metadata WHERE complete=1 LIMIT 1";
+        bool valid = sqlite3_prepare_v2(opened, sql, -1, &stmt, nullptr) == SQLITE_OK &&
+            sqlite3_step(stmt) == SQLITE_ROW &&
+            sqlite3_column_int(stmt, 0) == INDEX_VERSION;
+        if (valid) {
+            const unsigned char * value = sqlite3_column_text(stmt, 1);
+            id = value ? reinterpret_cast<const char *>(value) : "";
+            merge_count = static_cast<uint64_t>(sqlite3_column_int64(stmt, 2));
+            valid = !id.empty();
+        }
+        sqlite3_finalize(stmt);
+        if (!valid) {
+            error = "tokenizer index is incomplete or has an unsupported version";
+            sqlite3_close(opened);
+            return false;
+        }
+        db = opened;
+        sqlite3_busy_timeout(db, 5000);
+        return true;
+    }
+
+    bool build(
+            const std::vector<llama_tokenizer_longhaul_reverse> & reverse,
+            const std::vector<llama_tokenizer_longhaul_merge> & merges,
+            const std::string & tokenizer_id,
+            std::string & error) {
+        const fs::path temp = artifact.string() + ".tmp." + std::to_string(process_id());
+        std::error_code ec;
+        fs::remove(temp, ec);
+
+        sqlite3 * output = nullptr;
+        if (sqlite3_open_v2(temp.string().c_str(), &output,
+                SQLITE_OPEN_READWRITE | SQLITE_OPEN_CREATE | SQLITE_OPEN_EXCLUSIVE, nullptr) != SQLITE_OK) {
+            error = output ? sqlite3_errmsg(output) : "failed to create tokenizer index";
+            if (output) sqlite3_close(output);
+            return false;
+        }
+        auto fail = [&](const std::string & message) {
+            error = message;
+            sqlite3_close(output);
+            output = nullptr;
+            std::error_code remove_ec;
+            fs::remove(temp, remove_ec);
+            return false;
+        };
+
+        if (!exec_sql(output,
+                "PRAGMA journal_mode=OFF;"
+                "PRAGMA synchronous=OFF;"
+                "PRAGMA locking_mode=EXCLUSIVE;"
+                "CREATE TABLE metadata("
+                " version INTEGER NOT NULL,complete INTEGER NOT NULL,tokenizer_id TEXT NOT NULL,"
+                " n_reverse INTEGER NOT NULL,n_merges INTEGER NOT NULL);"
+                "CREATE TABLE reverse_tokens(text BLOB PRIMARY KEY,id INTEGER NOT NULL) WITHOUT ROWID;"
+                "CREATE TABLE merges(left_text BLOB NOT NULL,right_text BLOB NOT NULL,rank INTEGER NOT NULL,"
+                " PRIMARY KEY(left_text,right_text)) WITHOUT ROWID;"
+                "BEGIN IMMEDIATE;", error)) {
+            return fail(error);
+        }
+
+        sqlite3_stmt * insert_reverse = nullptr;
+        sqlite3_stmt * insert_merge = nullptr;
+        if (sqlite3_prepare_v2(output,
+                "INSERT INTO reverse_tokens(text,id) VALUES(?1,?2)", -1, &insert_reverse, nullptr) != SQLITE_OK ||
+            sqlite3_prepare_v2(output,
+                "INSERT INTO merges(left_text,right_text,rank) VALUES(?1,?2,?3)",
+                -1, &insert_merge, nullptr) != SQLITE_OK) {
+            if (insert_reverse) sqlite3_finalize(insert_reverse);
+            if (insert_merge) sqlite3_finalize(insert_merge);
+            return fail(sqlite3_errmsg(output));
+        }
+
+        for (const auto & item : reverse) {
+            sqlite3_bind_blob(insert_reverse, 1, item.text.data(), int(item.text.size()), SQLITE_STATIC);
+            sqlite3_bind_int(insert_reverse, 2, item.id);
+            if (sqlite3_step(insert_reverse) != SQLITE_DONE) {
+                const std::string message = sqlite3_errmsg(output);
+                sqlite3_finalize(insert_reverse);
+                sqlite3_finalize(insert_merge);
+                return fail(message);
+            }
+            sqlite3_reset(insert_reverse);
+            sqlite3_clear_bindings(insert_reverse);
+        }
+        for (const auto & item : merges) {
+            sqlite3_bind_blob(insert_merge, 1, item.left.data(), int(item.left.size()), SQLITE_STATIC);
+            sqlite3_bind_blob(insert_merge, 2, item.right.data(), int(item.right.size()), SQLITE_STATIC);
+            sqlite3_bind_int(insert_merge, 3, item.rank);
+            if (sqlite3_step(insert_merge) != SQLITE_DONE) {
+                const std::string message = sqlite3_errmsg(output);
+                sqlite3_finalize(insert_reverse);
+                sqlite3_finalize(insert_merge);
+                return fail(message);
+            }
+            sqlite3_reset(insert_merge);
+            sqlite3_clear_bindings(insert_merge);
+        }
+        sqlite3_finalize(insert_reverse);
+        sqlite3_finalize(insert_merge);
+
+        sqlite3_stmt * metadata = nullptr;
+        if (sqlite3_prepare_v2(output,
+                "INSERT INTO metadata VALUES(?1,1,?2,?3,?4)", -1, &metadata, nullptr) != SQLITE_OK) {
+            return fail(sqlite3_errmsg(output));
+        }
+        sqlite3_bind_int(metadata, 1, INDEX_VERSION);
+        sqlite3_bind_text(metadata, 2, tokenizer_id.c_str(), -1, SQLITE_STATIC);
+        sqlite3_bind_int64(metadata, 3, static_cast<sqlite3_int64>(reverse.size()));
+        sqlite3_bind_int64(metadata, 4, static_cast<sqlite3_int64>(merges.size()));
+        const bool metadata_ok = sqlite3_step(metadata) == SQLITE_DONE;
+        sqlite3_finalize(metadata);
+        if (!metadata_ok || !exec_sql(output, "COMMIT;", error)) {
+            return fail(metadata_ok ? error : sqlite3_errmsg(output));
+        }
+        sqlite3_close(output);
+        output = nullptr;
+
+        const uint64_t persistent_capacity = llama_token_cache_capacity();
+        const uint64_t artifact_size = fs::file_size(temp, ec);
+        if (ec || artifact_size > persistent_capacity) {
+            fs::remove(temp, ec);
+            error = "completed tokenizer index does not fit in --token-cache-size";
+            return false;
+        }
+
+        struct artifact_candidate {
+            fs::path path;
+            fs::file_time_type modified;
+            uint64_t size;
+        };
+        std::vector<artifact_candidate> candidates;
+        uint64_t existing_bytes = 0;
+        for (const auto & entry : fs::directory_iterator(artifact.parent_path(), ec)) {
+            if (ec) break;
+            if (!entry.is_regular_file(ec) || entry.path().extension() != ".sqlite3" ||
+                    entry.path() == artifact || entry.path() == temp) {
+                continue;
+            }
+            const uint64_t size = entry.file_size(ec);
+            if (ec) {
+                ec.clear();
+                continue;
+            }
+            existing_bytes += size;
+            candidates.push_back({entry.path(), entry.last_write_time(ec), size});
+            if (ec) ec.clear();
+        }
+        std::sort(candidates.begin(), candidates.end(),
+            [](const artifact_candidate & a, const artifact_candidate & b) {
+                return a.modified < b.modified;
+            });
+        for (const auto & candidate : candidates) {
+            if (existing_bytes + artifact_size <= persistent_capacity) break;
+            fs::remove(candidate.path, ec);
+            if (!ec) existing_bytes -= std::min(existing_bytes, candidate.size);
+            ec.clear();
+        }
+        if (existing_bytes + artifact_size > persistent_capacity) {
+            fs::remove(temp, ec);
+            error = "tokenizer indexes do not fit in --token-cache-size";
+            return false;
+        }
+
+        fs::rename(temp, artifact, ec);
+        if (ec) {
+            const std::string rename_error = ec.message();
+            // A competing process may have published the same immutable artifact.
+            std::error_code remove_ec;
+            fs::remove(temp, remove_ec);
+            std::string ignored;
+            if (open_existing(ignored)) {
+                return true;
+            }
+            error = "failed to publish tokenizer index: " + rename_error;
+            return false;
+        }
+        const bool opened = open_existing(error);
+        if (opened) {
+            llama_token_cache_set_external_bytes(tokenizer_artifact_bytes(artifact.parent_path()));
+        }
+        return opened;
+    }
+
+    int lookup(const std::string & cache_key, const char * sql,
+            const std::string & first, const std::string * second, int missing) const {
+        std::lock_guard<std::mutex> lock(mutex);
+        auto found = values.find(cache_key);
+        if (found != values.end()) {
+            lru.splice(lru.begin(), lru, found->second.position);
+            return found->second.value;
+        }
+
+        sqlite3_stmt * stmt = nullptr;
+        int value = missing;
+        if (sqlite3_prepare_v2(db, sql, -1, &stmt, nullptr) == SQLITE_OK) {
+            sqlite3_bind_blob(stmt, 1, first.data(), int(first.size()), SQLITE_STATIC);
+            if (second) {
+                sqlite3_bind_blob(stmt, 2, second->data(), int(second->size()), SQLITE_STATIC);
+            }
+            if (sqlite3_step(stmt) == SQLITE_ROW) {
+                value = sqlite3_column_int(stmt, 0);
+            }
+        }
+        sqlite3_finalize(stmt);
+
+        const uint64_t bytes = cache_key.size() + sizeof(cache_entry) + sizeof(std::string);
+        if (bytes <= capacity) {
+            while (!lru.empty() && used + bytes > capacity) {
+                auto victim = values.find(lru.back());
+                if (victim != values.end()) {
+                    used -= victim->second.bytes;
+                    values.erase(victim);
+                }
+                lru.pop_back();
+            }
+            lru.push_front(cache_key);
+            values.emplace(cache_key, cache_entry{value, bytes, lru.begin()});
+            used += bytes;
+        }
+        return value;
+    }
+};
+
+llama_tokenizer_longhaul_index::llama_tokenizer_longhaul_index(
+        std::string source_key, uint64_t cache_bytes)
+    : pimpl(std::make_unique<impl>(std::move(source_key), cache_bytes)) {
+}
+
+llama_tokenizer_longhaul_index::~llama_tokenizer_longhaul_index() = default;
+
+bool llama_tokenizer_longhaul_index::prepare(
+        const std::vector<llama_tokenizer_longhaul_reverse> & reverse,
+        const std::vector<llama_tokenizer_longhaul_merge> & merges,
+        const std::string & tokenizer_id,
+        std::string & error) {
+    const std::string cache_db = llama_token_cache_database_path();
+    if (cache_db.empty()) {
+        error = "persistent token cache is disabled or unavailable";
+        return false;
+    }
+    const fs::path directory = fs::path(cache_db).parent_path() / "tokenizer-indexes";
+    std::error_code ec;
+    fs::create_directories(directory, ec);
+    if (ec) {
+        error = "failed to create tokenizer index directory: " + ec.message();
+        return false;
+    }
+    pimpl->artifact = directory / (llama_token_cache_hash(pimpl->source_key) + ".sqlite3");
+    if (pimpl->open_existing(error)) {
+        if (pimpl->id == tokenizer_id) {
+            llama_token_cache_set_external_bytes(tokenizer_artifact_bytes(directory));
+            return true;
+        }
+        sqlite3_close(pimpl->db);
+        pimpl->db = nullptr;
+        std::error_code stale_ec;
+        fs::remove(pimpl->artifact, stale_ec);
+    }
+
+    lock_directory lock{pimpl->artifact.string() + ".lock", false};
+    lock.owned = fs::create_directory(lock.path, ec);
+    if (!lock.owned) {
+        // Another process is building. Wait for its atomic publication.
+        for (int i = 0; i < 24000; ++i) {
+            std::this_thread::sleep_for(std::chrono::milliseconds(25));
+            std::string ignored;
+            if (pimpl->open_existing(ignored)) {
+                if (pimpl->id == tokenizer_id) {
+                    llama_token_cache_set_external_bytes(tokenizer_artifact_bytes(directory));
+                    return true;
+                }
+                sqlite3_close(pimpl->db);
+                pimpl->db = nullptr;
+            }
+            if (!fs::exists(lock.path)) {
+                lock.owned = fs::create_directory(lock.path, ec);
+                if (lock.owned) break;
+            }
+        }
+        if (!lock.owned) {
+            error = "timed out waiting for another tokenizer index builder";
+            return false;
+        }
+    }
+
+    error.clear();
+    if (pimpl->open_existing(error)) {
+        if (pimpl->id == tokenizer_id) {
+            llama_token_cache_set_external_bytes(tokenizer_artifact_bytes(directory));
+            return true;
+        }
+        sqlite3_close(pimpl->db);
+        pimpl->db = nullptr;
+    }
+    std::error_code remove_ec;
+    fs::remove(pimpl->artifact, remove_ec);
+    return pimpl->build(reverse, merges, tokenizer_id, error);
+}
+
+llama_token llama_tokenizer_longhaul_index::text_to_token(const std::string & text) const {
+    const std::string key = "t" + text;
+    return pimpl->lookup(key,
+        "SELECT id FROM reverse_tokens WHERE text=?1", text, nullptr, LLAMA_TOKEN_NULL);
+}
+
+int llama_tokenizer_longhaul_index::find_bpe_rank(
+        const std::string & left, const std::string & right) const {
+    std::string key;
+    key.reserve(left.size() + right.size() + 2);
+    key.push_back('m');
+    key += left;
+    key.push_back('\0');
+    key += right;
+    return pimpl->lookup(key,
+        "SELECT rank FROM merges WHERE left_text=?1 AND right_text=?2", left, &right, -1);
+}
+
+std::vector<std::string> llama_tokenizer_longhaul_index::get_bpe_merges() const {
+    std::lock_guard<std::mutex> lock(pimpl->mutex);
+    std::vector<std::string> result;
+    result.reserve(pimpl->merge_count);
+    sqlite3_stmt * stmt = nullptr;
+    if (sqlite3_prepare_v2(pimpl->db,
+            "SELECT left_text,right_text FROM merges ORDER BY rank", -1, &stmt, nullptr) != SQLITE_OK) {
+        return result;
+    }
+    while (sqlite3_step(stmt) == SQLITE_ROW) {
+        const char * left = static_cast<const char *>(sqlite3_column_blob(stmt, 0));
+        const int left_size = sqlite3_column_bytes(stmt, 0);
+        const char * right = static_cast<const char *>(sqlite3_column_blob(stmt, 1));
+        const int right_size = sqlite3_column_bytes(stmt, 1);
+        std::string item(left, left_size);
+        item.push_back(' ');
+        item.append(right, right_size);
+        result.push_back(std::move(item));
+    }
+    sqlite3_finalize(stmt);
+    return result;
+}
+
+uint64_t llama_tokenizer_longhaul_index::cache_bytes_used() const {
+    std::lock_guard<std::mutex> lock(pimpl->mutex);
+    return pimpl->used;
+}
+
+uint64_t llama_tokenizer_longhaul_index::cache_capacity() const {
+    return pimpl->capacity;
+}
+
+uint64_t llama_tokenizer_longhaul_index::n_merges() const {
+    return pimpl->merge_count;
+}
+
+const std::string & llama_tokenizer_longhaul_index::tokenizer_id() const {
+    return pimpl->id;
+}

src/llama-tokenizer-longhaul.h

@@ -0,0 +1,47 @@
+#pragma once
+
+#include "llama.h"
+
+#include <cstdint>
+#include <memory>
+#include <string>
+#include <vector>
+
+struct llama_tokenizer_longhaul_reverse {
+    std::string text;
+    llama_token id;
+};
+
+struct llama_tokenizer_longhaul_merge {
+    std::string left;
+    std::string right;
+    int rank;
+};
+
+class llama_tokenizer_longhaul_index {
+public:
+    llama_tokenizer_longhaul_index(std::string source_key, uint64_t cache_bytes);
+    ~llama_tokenizer_longhaul_index();
+
+    llama_tokenizer_longhaul_index(const llama_tokenizer_longhaul_index &) = delete;
+    llama_tokenizer_longhaul_index & operator=(const llama_tokenizer_longhaul_index &) = delete;
+
+    bool prepare(
+            const std::vector<llama_tokenizer_longhaul_reverse> & reverse,
+            const std::vector<llama_tokenizer_longhaul_merge> & merges,
+            const std::string & tokenizer_id,
+            std::string & error);
+
+    llama_token text_to_token(const std::string & text) const;
+    int find_bpe_rank(const std::string & left, const std::string & right) const;
+    std::vector<std::string> get_bpe_merges() const;
+
+    uint64_t cache_bytes_used() const;
+    uint64_t cache_capacity() const;
+    uint64_t n_merges() const;
+    const std::string & tokenizer_id() const;
+
+private:
+    struct impl;
+    std::unique_ptr<impl> pimpl;
+};

src/llama-vocab.cpp

@@ -1,5 +1,6 @@
 #include "llama-vocab.h"
 #include "llama-token-cache.h"
+#include "llama-tokenizer-longhaul.h"
 
 #include "ggml.h"
 #include "gguf.h"
@@ -13,13 +14,16 @@
 #include <cctype>
 #include <cfloat>
 #include <cmath>
+#include <condition_variable>
 #include <cstdarg>
 #include <cstring>
 #include <forward_list>
 #include <limits>
 #include <map>
+#include <mutex>
 #include <queue>
 #include <set>
+#include <thread>
 #include <unordered_map>
 
 //
@@ -434,13 +438,6 @@ struct llm_tokenizer_bpe : llm_tokenizer {
                     "[^\\r\\n\\p{L}\\p{N}]?((?=[\\p{L}])([^a-z]))*((?=[\\p{L}])([^A-Z]))+(?:'[sS]|'[tT]|'[rR][eE]|'[vV][eE]|'[mM]|'[lL][lL]|'[dD])?|[^\\r\\n\\p{L}\\p{N}]?((?=[\\p{L}])([^a-z]))+((?=[\\p{L}])([^A-Z]))*(?:'[sS]|'[tT]|'[rR][eE]|'[vV][eE]|'[mM]|'[lL][lL]|'[dD])?|\\p{N}{1,3}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n/]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+",
                 };
                 break;
-            case LLAMA_VOCAB_PRE_TYPE_INKLING:
-                // o200k-family expression with combining marks included in
-                // letter lookaheads, as used by the Inkling tokenizer.
-                regex_exprs = {
-                    "[^\\r\\n\\p{L}\\p{N}]?((?=[\\p{L}\\p{M}])([^a-z]))*((?=[\\p{L}\\p{M}])([^A-Z]))+(?:'[sS]|'[tT]|'[rR][eE]|'[vV][eE]|'[mM]|'[lL][lL]|'[dD])?|[^\\r\\n\\p{L}\\p{N}]?((?=[\\p{L}\\p{M}])([^a-z]))+((?=[\\p{L}\\p{M}])([^A-Z]))*(?:'[sS]|'[tT]|'[rR][eE]|'[vV][eE]|'[mM]|'[lL][lL]|'[dD])?|\\p{N}{1,3}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n/]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+",
-                };
-                break;
             case LLAMA_VOCAB_PRE_TYPE_GRANITE_EMB_MULTI:
                 // Same lookaheads as GPT4O but with \p{M} added so combining marks
                 // (diacritics) attach to their base letters. Avoids excessive
@@ -1853,12 +1850,37 @@ struct llama_vocab::impl {
 
     std::vector<char> precompiled_charsmap;
 
+    enum class longhaul_state {
+        disabled,
+        building,
+        ready,
+        failed,
+    };
+    bool tokenizer_longhaul = false;
+    uint64_t tokenizer_longhaul_cache_bytes = 0;
+    uint64_t tokenizer_longhaul_n_merges = 0;
+    mutable std::mutex tokenizer_longhaul_mutex;
+    mutable std::condition_variable tokenizer_longhaul_ready;
+    mutable longhaul_state tokenizer_longhaul_state = longhaul_state::disabled;
+    mutable std::string tokenizer_longhaul_error;
+    mutable std::unique_ptr<llama_tokenizer_longhaul_index> tokenizer_longhaul_index;
+    std::thread tokenizer_longhaul_worker;
+
     impl(const llama_vocab & vocab) : vocab(vocab) {
     }
 
-    ~impl() = default;
+    ~impl() {
+        if (tokenizer_longhaul_worker.joinable()) {
+            tokenizer_longhaul_worker.join();
+        }
+    }
 
-    void load(llama_model_loader & ml, const LLM_KV & kv);
+    void load(
+            llama_model_loader & ml,
+            const LLM_KV & kv,
+            bool tokenizer_longhaul,
+            uint64_t tokenizer_longhaul_cache_bytes);
+    void wait_for_tokenizer_longhaul() const;
 
     bool load_cache_snapshot(const std::vector<uint8_t> & snapshot);
     void reset_cache_snapshot_state();
@@ -1915,7 +1937,7 @@ struct llama_vocab::impl {
                          bool   special) const;
 
     // use cached data
-    const std::string & token_to_piece(llama_token token) const;
+    std::string token_to_piece(llama_token token) const;
 
     int32_t detokenize(
             const llama_token * tokens,
@@ -1935,11 +1957,29 @@ private:
     const llama_vocab & vocab;
 };
 
-void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
+void llama_vocab::impl::load(
+        llama_model_loader & ml,
+        const LLM_KV & kv,
+        bool tokenizer_longhaul_value,
+        uint64_t tokenizer_longhaul_cache_bytes_value) {
     struct gguf_context * ctx = ml.metadata;
 
+    tokenizer_longhaul = tokenizer_longhaul_value;
+    tokenizer_longhaul_cache_bytes = tokenizer_longhaul_cache_bytes_value;
+    if (tokenizer_longhaul) {
+        if (tokenizer_longhaul_cache_bytes == 0) {
+            throw std::runtime_error("tokenizer-longhaul cache budget must be positive");
+        }
+        if (ml.get_arch() != LLM_ARCH_QWEN35MOE && ml.get_arch() != LLM_ARCH_LAGUNA) {
+            throw std::runtime_error("tokenizer-longhaul currently requires a qwen35moe or laguna model");
+        }
+        if (ml.get_token_cache_source_key().empty()) {
+            throw std::runtime_error("tokenizer-longhaul requires a file-backed GGUF model");
+        }
+    }
+
     const std::string & source_cache_key = ml.get_token_cache_source_key();
-    if (!source_cache_key.empty()) {
+    if (!tokenizer_longhaul && !source_cache_key.empty()) {
         std::vector<uint8_t> snapshot;
         bool found = llama_token_cache_lookup(LLAMA_TOKEN_CACHE_KIND_TOKENIZER, source_cache_key, snapshot);
         if (found && snapshot.size() == 65 && snapshot[0] == 'A') {
@@ -1965,6 +2005,9 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
         ml.get_key(LLM_KV_TOKENIZER_TOKEN_TYPE_COUNT, n_token_types, false);
 
         if (tokenizer_model == "no_vocab" || tokenizer_model == "none") {
+            if (tokenizer_longhaul) {
+                throw std::runtime_error("tokenizer-longhaul requires a BPE tokenizer");
+            }
             type = LLAMA_VOCAB_TYPE_NONE;
 
             // default special tokens
@@ -2331,10 +2374,6 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
                 tokenizer_pre == "talkie") {
                 pre_type = LLAMA_VOCAB_PRE_TYPE_GPT4O;
                 clean_spaces = false;
-            } else if (
-                tokenizer_pre == "inkling") {
-                pre_type = LLAMA_VOCAB_PRE_TYPE_INKLING;
-                clean_spaces = false;
             } else if (
                 tokenizer_pre == "granite-embed-multi-97m") {
                 pre_type = LLAMA_VOCAB_PRE_TYPE_GRANITE_EMB_MULTI;
@@ -2618,12 +2657,7 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
             if (suppress_idx != -1) {
                 const int n = gguf_get_arr_n(ctx, suppress_idx);
                 const int32_t * data = (const int32_t *) gguf_get_arr_data(ctx, suppress_idx);
-                suppress_tokens.reserve(n);
-                for (int i = 0; i < n; ++i) {
-                    if (data[i] >= 0 && data[i] < (int32_t) id_to_token.size()) {
-                        suppress_tokens.push_back(data[i]);
-                    }
-                }
+                suppress_tokens.assign(data, data + n);
             }
         }
 
@@ -3079,6 +3113,69 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
         }
     }
 
+    if (tokenizer_longhaul) {
+        if (type != LLAMA_VOCAB_TYPE_BPE ||
+                (tokenizer_pre != "qwen35" && tokenizer_pre != "laguna")) {
+            throw std::runtime_error(
+                "tokenizer-longhaul currently supports only qwen35 and laguna BPE tokenizers");
+        }
+
+        std::vector<llama_tokenizer_longhaul_reverse> reverse;
+        reverse.reserve(token_to_id.size());
+        while (!token_to_id.empty()) {
+            auto node = token_to_id.extract(token_to_id.begin());
+            reverse.push_back({std::move(node.key()), node.mapped()});
+        }
+        token_to_id.rehash(0);
+
+        std::vector<llama_tokenizer_longhaul_merge> merges;
+        merges.reserve(bpe_ranks.size());
+        while (!bpe_ranks.empty()) {
+            auto node = bpe_ranks.extract(bpe_ranks.begin());
+            auto key = std::move(node.key());
+            merges.push_back({std::move(key.first), std::move(key.second), node.mapped()});
+        }
+        bpe_ranks.rehash(0);
+        tokenizer_longhaul_n_merges = merges.size();
+        cache_token_to_piece.clear();
+        cache_token_to_piece.shrink_to_fit();
+
+        token_cache_id = llama_token_cache_hash(
+            std::string("tokenizer-longhaul-v1\n") + source_cache_key);
+        {
+            std::lock_guard<std::mutex> lock(tokenizer_longhaul_mutex);
+            tokenizer_longhaul_state = longhaul_state::building;
+        }
+        const uint64_t cache_bytes = tokenizer_longhaul_cache_bytes;
+        const std::string source_key = source_cache_key;
+        const std::string tokenizer_id = token_cache_id;
+        tokenizer_longhaul_worker = std::thread(
+            [this, source_key, cache_bytes, tokenizer_id,
+             reverse = std::move(reverse), merges = std::move(merges)]() mutable {
+                auto index = std::make_unique<llama_tokenizer_longhaul_index>(source_key, cache_bytes);
+                std::string error;
+                const bool ok = index->prepare(reverse, merges, tokenizer_id, error);
+                {
+                    std::lock_guard<std::mutex> lock(tokenizer_longhaul_mutex);
+                    if (ok) {
+                        tokenizer_longhaul_index = std::move(index);
+                        tokenizer_longhaul_state = longhaul_state::ready;
+                        LLAMA_LOG_INFO("%s: tokenizer-longhaul index is ready\n", "llama_vocab");
+                    } else {
+                        tokenizer_longhaul_error = error.empty()
+                            ? "unknown tokenizer-longhaul build error"
+                            : std::move(error);
+                        tokenizer_longhaul_state = longhaul_state::failed;
+                        LLAMA_LOG_ERROR("%s: tokenizer-longhaul index failed: %s\n",
+                            "llama_vocab", tokenizer_longhaul_error.c_str());
+                    }
+                }
+                tokenizer_longhaul_ready.notify_all();
+            });
+        LLAMA_LOG_INFO("%s: tokenizer-longhaul index is building in the background\n", __func__);
+        return;
+    }
+
     // Persist a versioned canonical snapshot of the effective tokenizer state.
     // The snapshot also serves as the namespace for text-to-token cache entries.
     {
@@ -3167,6 +3264,23 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
     }
 }
 
+void llama_vocab::impl::wait_for_tokenizer_longhaul() const {
+    if (!tokenizer_longhaul) {
+        return;
+    }
+    std::unique_lock<std::mutex> lock(tokenizer_longhaul_mutex);
+    if (tokenizer_longhaul_state == longhaul_state::disabled) {
+        return;
+    }
+    tokenizer_longhaul_ready.wait(lock, [this] {
+        return tokenizer_longhaul_state == longhaul_state::ready ||
+               tokenizer_longhaul_state == longhaul_state::failed;
+    });
+    if (tokenizer_longhaul_state == longhaul_state::failed) {
+        throw std::runtime_error("tokenizer-longhaul failed: " + tokenizer_longhaul_error);
+    }
+}
+
 void llama_vocab::impl::reset_cache_snapshot_state() {
     n_token_types = 0;
     tokenizer_model.clear();
@@ -3298,36 +3412,43 @@ std::string llama_vocab::impl::type_name() const{
 }
 
 bool llama_vocab::impl::is_normal(llama_token id) const {
+    wait_for_tokenizer_longhaul();
     GGML_ASSERT(type != LLAMA_VOCAB_TYPE_NONE);
     return id_to_token[id].attr & LLAMA_TOKEN_ATTR_NORMAL;
 }
 
 bool llama_vocab::impl::is_unknown(llama_token id) const {
+    wait_for_tokenizer_longhaul();
     GGML_ASSERT(type != LLAMA_VOCAB_TYPE_NONE);
     return id_to_token[id].attr & LLAMA_TOKEN_ATTR_UNKNOWN;
 }
 
 bool llama_vocab::impl::is_control(llama_token id) const {
+    wait_for_tokenizer_longhaul();
     GGML_ASSERT(type != LLAMA_VOCAB_TYPE_NONE);
     return id_to_token[id].attr & LLAMA_TOKEN_ATTR_CONTROL;
 }
 
 bool llama_vocab::impl::is_byte(llama_token id) const {
+    wait_for_tokenizer_longhaul();
     GGML_ASSERT(type != LLAMA_VOCAB_TYPE_NONE);
     return id_to_token[id].attr & LLAMA_TOKEN_ATTR_BYTE;
 }
 
 bool llama_vocab::impl::is_user_defined(llama_token id) const {
+    wait_for_tokenizer_longhaul();
     GGML_ASSERT(type != LLAMA_VOCAB_TYPE_NONE);
     return id_to_token[id].attr & LLAMA_TOKEN_ATTR_USER_DEFINED;
 }
 
 bool llama_vocab::impl::is_unused(llama_token id) const {
+    wait_for_tokenizer_longhaul();
     GGML_ASSERT(type != LLAMA_VOCAB_TYPE_NONE);
     return id_to_token[id].attr & LLAMA_TOKEN_ATTR_UNUSED;
 }
 
 bool llama_vocab::impl::is_eog(llama_token id) const {
+    wait_for_tokenizer_longhaul();
     return id != LLAMA_TOKEN_NULL && special_eog_ids.count(id) > 0;
 }
 
@@ -3355,6 +3476,7 @@ uint8_t llama_vocab::impl::token_to_byte(llama_token id) const {
 }
 
 llama_token_attr llama_vocab::impl::token_get_attr(llama_token id) const {
+    wait_for_tokenizer_longhaul();
     GGML_ASSERT(type != LLAMA_VOCAB_TYPE_NONE);
     return id_to_token.at(id).attr;
 }
@@ -3765,6 +3887,7 @@ std::vector<llama_token> llama_vocab::impl::tokenize(
         const std::string & raw_text,
         bool add_special,
         bool parse_special) const {
+    wait_for_tokenizer_longhaul();
     if (raw_text.empty() || token_cache_id.empty()) {
         return tokenize_uncached(raw_text, add_special, parse_special);
     }
@@ -3808,6 +3931,7 @@ std::vector<llama_token> llama_vocab::impl::tokenize(
 }
 
 int32_t llama_vocab::impl::token_to_piece(llama_token token, char * buf, int32_t length, int32_t lstrip, bool special) const {
+    wait_for_tokenizer_longhaul();
     // ref: https://github.com/ggml-org/llama.cpp/pull/7587#discussion_r1620983843
     static const int attr_special = LLAMA_TOKEN_ATTR_UNKNOWN | LLAMA_TOKEN_ATTR_CONTROL;
     const llama_token_attr attr = token_get_attr(token);
@@ -3924,7 +4048,11 @@ int32_t llama_vocab::impl::token_to_piece(llama_token token, char * buf, int32_t
     return 0;
 }
 
-const std::string & llama_vocab::impl::token_to_piece(llama_token token) const {
+std::string llama_vocab::impl::token_to_piece(llama_token token) const {
+    wait_for_tokenizer_longhaul();
+    if (tokenizer_longhaul) {
+        return token_to_piece_for_cache(token, true);
+    }
     return cache_token_to_piece.at(token);
 }
 
@@ -4043,7 +4171,15 @@ int32_t llama_vocab::impl::detokenize(
 void llama_vocab::impl::print_info() const {
     LLAMA_LOG_INFO("%s: vocab type            = %s\n",     __func__, type_name().c_str());
     LLAMA_LOG_INFO("%s: n_vocab               = %u\n",     __func__, vocab.n_tokens());
-    LLAMA_LOG_INFO("%s: n_merges              = %u\n",     __func__, (uint32_t) bpe_ranks.size());
+    LLAMA_LOG_INFO("%s: n_merges              = %u\n",     __func__,
+        (uint32_t) (tokenizer_longhaul ? tokenizer_longhaul_n_merges : bpe_ranks.size()));
+    if (tokenizer_longhaul) {
+        std::lock_guard<std::mutex> lock(tokenizer_longhaul_mutex);
+        LLAMA_LOG_INFO("%s: tokenizer longhaul     = %s, cache = %.2f MiB\n", __func__,
+            tokenizer_longhaul_state == longhaul_state::ready ? "ready" :
+            tokenizer_longhaul_state == longhaul_state::failed ? "failed" : "building",
+            tokenizer_longhaul_cache_bytes / 1024.0 / 1024.0);
+    }
 
     // special tokens
     if (special_bos_id  != LLAMA_TOKEN_NULL)    { LLAMA_LOG_INFO( "%s: BOS token             = %d '%s'\n", __func__, special_bos_id,     id_to_token.at(special_bos_id).text.c_str() );  }
@@ -4076,8 +4212,12 @@ llama_vocab::llama_vocab() : pimpl(new impl(*this)) {
 
 llama_vocab::~llama_vocab() = default;
 
-void llama_vocab::load(llama_model_loader & ml, const LLM_KV & kv) {
-    pimpl->load(ml, kv);
+void llama_vocab::load(
+        llama_model_loader & ml,
+        const LLM_KV & kv,
+        bool tokenizer_longhaul,
+        uint64_t tokenizer_longhaul_cache_bytes) {
+    pimpl->load(ml, kv, tokenizer_longhaul, tokenizer_longhaul_cache_bytes);
 }
 
 std::string llama_vocab::get_tokenizer_model() const {
@@ -4147,23 +4287,29 @@ llama_token llama_vocab::byte_to_token(uint8_t ch) const {
         case LLAMA_VOCAB_TYPE_SPM:
         case LLAMA_VOCAB_TYPE_UGM: {
             const char buf[7] = { '<', '0', 'x', hex[ch >> 4], hex[ch & 15], '>', 0 };
-            auto token = pimpl->token_to_id.find(buf);
-            if (token != pimpl->token_to_id.end()) {
-                return (*token).second;
+            const auto token = text_to_token(buf);
+            if (token != LLAMA_TOKEN_NULL) {
+                return token;
             }
             // Try to fall back to just the byte as a string
             const char buf2[2] = { (char)ch, 0 };
-            return pimpl->token_to_id.at(buf2);
+            const auto fallback = text_to_token(buf2);
+            if (fallback == LLAMA_TOKEN_NULL) throw std::out_of_range("byte token not found");
+            return fallback;
         }
         case LLAMA_VOCAB_TYPE_WPM:
         case LLAMA_VOCAB_TYPE_BPE: {
-            return pimpl->token_to_id.at(unicode_byte_to_utf8(ch));
+            const auto token = text_to_token(unicode_byte_to_utf8(ch));
+            if (token == LLAMA_TOKEN_NULL) throw std::out_of_range("byte token not found");
+            return token;
         }
         case LLAMA_VOCAB_TYPE_PLAMO2: {
             // PLaMo-2 uses byte tokens in format <0xXX>
             char hex_str[8];
             snprintf(hex_str, sizeof(hex_str), "<0x%02X>", ch);
-            return pimpl->token_to_id.at(hex_str);
+            const auto token = text_to_token(hex_str);
+            if (token == LLAMA_TOKEN_NULL) throw std::out_of_range("byte token not found");
+            return token;
         }
         default:
             GGML_ABORT("fatal error");
@@ -4172,6 +4318,10 @@ llama_token llama_vocab::byte_to_token(uint8_t ch) const {
 
 llama_token llama_vocab::text_to_token(const std::string & text) const {
     GGML_ASSERT(pimpl->type != LLAMA_VOCAB_TYPE_NONE);
+    pimpl->wait_for_tokenizer_longhaul();
+    if (pimpl->tokenizer_longhaul_index) {
+        return pimpl->tokenizer_longhaul_index->text_to_token(text);
+    }
     auto it = pimpl->token_to_id.find(text);
     if (it != pimpl->token_to_id.end()) {
         return (*it).second;
@@ -4181,16 +4331,19 @@ llama_token llama_vocab::text_to_token(const std::string & text) const {
 
 const llama_vocab::token_data & llama_vocab::get_token_data(llama_token id) const {
     GGML_ASSERT(pimpl->type != LLAMA_VOCAB_TYPE_NONE);
+    pimpl->wait_for_tokenizer_longhaul();
     return pimpl->id_to_token.at(id);
 }
 
 const char * llama_vocab::token_get_text(llama_token id) const {
     GGML_ASSERT(pimpl->type != LLAMA_VOCAB_TYPE_NONE);
+    pimpl->wait_for_tokenizer_longhaul();
     return pimpl->id_to_token.at(id).text.c_str();
 }
 
 float llama_vocab::token_get_score(llama_token id) const {
     GGML_ASSERT(pimpl->type != LLAMA_VOCAB_TYPE_NONE);
+    pimpl->wait_for_tokenizer_longhaul();
     return pimpl->id_to_token.at(id).score;
 }
 
@@ -4322,6 +4475,10 @@ int llama_vocab::find_bpe_rank(const std::string & token_left, const std::string
     GGML_ASSERT(token_left.find(' ')   == std::string::npos);
     GGML_ASSERT(token_right.find(' ')  == std::string::npos);
 
+    pimpl->wait_for_tokenizer_longhaul();
+    if (pimpl->tokenizer_longhaul_index) {
+        return pimpl->tokenizer_longhaul_index->find_bpe_rank(token_left, token_right);
+    }
     auto it = pimpl->bpe_ranks.find(std::make_pair(token_left, token_right));
     if (it == pimpl->bpe_ranks.end()) {
         return -1;
@@ -4331,6 +4488,10 @@ int llama_vocab::find_bpe_rank(const std::string & token_left, const std::string
 }
 
 std::vector<std::string> llama_vocab::get_bpe_merges() const {
+    pimpl->wait_for_tokenizer_longhaul();
+    if (pimpl->tokenizer_longhaul_index) {
+        return pimpl->tokenizer_longhaul_index->get_bpe_merges();
+    }
     int max_rank = -1;
     for (const auto & pair : pimpl->bpe_ranks) {
         max_rank = std::max(max_rank, pair.second);
@@ -4380,7 +4541,7 @@ std::vector<llama_token> llama_vocab::tokenize(
     return pimpl->tokenize(raw_text, add_special, parse_special);
 }
 
-const std::string & llama_vocab::token_to_piece(llama_token token) const {
+std::string llama_vocab::token_to_piece(llama_token token) const {
     return pimpl->token_to_piece(token);
 }
 

src/llama-vocab.h

@@ -65,7 +65,6 @@ enum llama_vocab_pre_type {
     LLAMA_VOCAB_PRE_TYPE_GRANITE_EMB_MULTI = 54,
     LLAMA_VOCAB_PRE_TYPE_MELLUM2           = 55,
     LLAMA_VOCAB_PRE_TYPE_LAGUNA            = 56,
-    LLAMA_VOCAB_PRE_TYPE_INKLING           = 57,
 };
 
 struct LLM_KV;
@@ -87,7 +86,11 @@ struct llama_vocab {
     llama_vocab();
     ~llama_vocab();
 
-    void load(llama_model_loader & ml, const LLM_KV & kv);
+    void load(
+            llama_model_loader & ml,
+            const LLM_KV & kv,
+            bool tokenizer_longhaul = false,
+            uint64_t tokenizer_longhaul_cache_bytes = 0);
 
     std::string get_tokenizer_model() const;
     std::string get_tokenizer_pre() const;
@@ -182,7 +185,7 @@ struct llama_vocab {
                          bool   special) const;
 
     // use cached data
-    const std::string & token_to_piece(llama_token token) const;
+    std::string token_to_piece(llama_token token) const;
 
     int32_t detokenize(
             const llama_token * tokens,

src/models/inkling.cpp

@@ -1,489 +0,0 @@
-// Inkling: relative-position attention with per-layer packed short-convolution
-// state and a mixture of routed and shared experts.
-
-#include "models.h"
-
-#include "../llama-memory-hybrid-iswa.h"
-#include "../llama-memory-recurrent.h"
-
-#include <algorithm>
-#include <cmath>
-
-void llama_model_inkling::load_arch_hparams(llama_model_loader & ml) {
-    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
-
-    ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp);
-    ml.get_key(LLM_KV_EXPERT_SHARED_COUNT,        hparams.n_expert_shared);
-    ml.get_key(LLM_KV_EXPERT_WEIGHTS_SCALE,       hparams.expert_weights_scale);
-    ml.get_key(LLM_KV_EXPERT_GATING_FUNC,         hparams.expert_gating_func, false);
-
-    ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW, hparams.n_swa);
-    hparams.swa_type = LLAMA_SWA_TYPE_STANDARD;
-    ml.get_key_or_arr(LLM_KV_ATTENTION_SLIDING_WINDOW_PATTERN, hparams.is_swa_impl, hparams.n_layer());
-
-    for (uint32_t il = 0; il < hparams.n_layer(); ++il) {
-        hparams.is_recr_impl[il] = 1;
-    }
-
-    ml.get_key(LLM_KV_INKLING_D_REL,             hparams.inkling_d_rel);
-    ml.get_key(LLM_KV_INKLING_REL_EXTENT,        hparams.inkling_rel_extent);
-    ml.get_key(LLM_KV_INKLING_REL_EXTENT_SWA,    hparams.inkling_rel_extent_swa);
-    ml.get_key(LLM_KV_INKLING_SHORTCONV_KERNEL,  hparams.n_shortconv_l_cache);
-    ml.get_key(LLM_KV_INKLING_DENSE_BLOCK_COUNT, hparams.n_layer_dense_lead);
-
-    float logit_scale_denom = 0.0f;
-    ml.get_key(LLM_KV_INKLING_LOGIT_SCALE_DENOM, logit_scale_denom);
-    GGML_ASSERT(logit_scale_denom != 0.0f);
-    hparams.f_logit_scale = 1.0f/logit_scale_denom;
-
-    ml.get_key(LLM_KV_INKLING_LOG_SCALING_N_FLOOR, hparams.inkling_log_n_floor,      false);
-    ml.get_key(LLM_KV_INKLING_LOG_SCALING_ALPHA,   hparams.inkling_log_alpha,        false);
-    ml.get_key(LLM_KV_INKLING_UNPADDED_VOCAB_SIZE, hparams.inkling_unpadded_n_vocab, false);
-
-    GGML_ASSERT(hparams.n_shortconv_l_cache > 1);
-    GGML_ASSERT(hparams.inkling_d_rel > 0);
-    GGML_ASSERT(hparams.inkling_rel_extent > 0 && hparams.inkling_rel_extent_swa > 0);
-
-    // Four streams per layer: K, V, attention output, and MLP output.
-    const uint32_t d_conv = hparams.n_shortconv_l_cache - 1;
-    hparams.n_embd_r_impl = d_conv *
-        (hparams.n_embd_k_gqa_max() + hparams.n_embd_v_gqa_max() + 2*hparams.n_embd);
-
-    type = LLM_TYPE_UNKNOWN;
-}
-
-void llama_model_inkling::load_arch_tensors(llama_model_loader &) {
-    LLAMA_LOAD_LOCALS;
-
-    const int64_t head_dim = hparams.n_embd_head_k();
-    const int64_t d_rel    = hparams.inkling_d_rel;
-    const int64_t K        = hparams.n_shortconv_l_cache;
-    const int64_t n_ff_exp = hparams.n_ff_exp;
-    const int64_t n_shexp  = hparams.n_expert_shared;
-    const int expert_flags = params.load_mode == LLAMA_LOAD_MODE_LONGHAUL ? TENSOR_LONGHAUL : 0;
-
-    tok_embd    = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD,      "weight"), {n_embd, n_vocab}, 0);
-    tok_norm    = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD_NORM, "weight", 0), {n_embd}, 0);
-    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM,     "weight"), {n_embd}, 0);
-    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,          "weight"), {n_embd, n_vocab}, 0);
-
-    for (int i = 0; i < n_layer; ++i) {
-        auto & layer = layers[i];
-
-        const int64_t n_head_kv_i = hparams.n_head_kv(i);
-        const int64_t kvw         = n_head_kv_i*head_dim;
-        const int64_t rel_extent  = hparams.is_swa(i)
-            ? hparams.inkling_rel_extent_swa
-            : hparams.inkling_rel_extent;
-
-        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
-
-        layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q,   "weight", i), {n_embd, n_head*head_dim}, 0);
-        layer.wk = create_tensor(tn(LLM_TENSOR_ATTN_K,   "weight", i), {n_embd, kvw}, 0);
-        layer.wv = create_tensor(tn(LLM_TENSOR_ATTN_V,   "weight", i), {n_embd, kvw}, 0);
-        layer.wr = create_tensor(tn(LLM_TENSOR_ATTN_R,   "weight", i), {n_embd, n_head*d_rel}, 0);
-        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_head*head_dim, n_embd}, 0);
-
-        layer.attn_q_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {head_dim}, 0);
-        layer.attn_k_norm = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {head_dim}, 0);
-
-        layer.attn_rel_proj = create_tensor(
-            tn(LLM_TENSOR_ATTN_REL_PROJ, "weight", i), {rel_extent, d_rel}, 0);
-
-        layer.shortconv_k    = create_tensor(tn(LLM_TENSOR_SHORTCONV_K,    "weight", i), {K, kvw}, 0);
-        layer.shortconv_v    = create_tensor(tn(LLM_TENSOR_SHORTCONV_V,    "weight", i), {K, kvw}, 0);
-        layer.shortconv_attn = create_tensor(tn(LLM_TENSOR_SHORTCONV_ATTN, "weight", i), {K, n_embd}, 0);
-        layer.shortconv_mlp  = create_tensor(tn(LLM_TENSOR_SHORTCONV_MLP,  "weight", i), {K, n_embd}, 0);
-
-        layer.ffn_norm   = create_tensor(tn(LLM_TENSOR_FFN_NORM,   "weight", i), {n_embd}, 0);
-        layer.ffn_gscale = create_tensor(tn(LLM_TENSOR_FFN_GSCALE, "weight", i), {1}, 0);
-
-        if (i < (int) hparams.n_layer_dense_lead) {
-            const int64_t n_ff_i = hparams.n_ff(i);
-            layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff_i}, 0);
-            layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd, n_ff_i}, 0);
-            layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff_i, n_embd}, 0);
-        } else {
-            GGML_ASSERT(n_expert > 0 && n_expert_used > 0 && n_shexp > 0);
-
-            // The final rows are shared-expert sink logits.
-            layer.ffn_gate_inp = create_tensor(
-                tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), {n_embd, n_expert + n_shexp}, 0);
-            layer.ffn_exp_probs_b = create_tensor(
-                tn(LLM_TENSOR_FFN_EXP_PROBS_B, "bias", i), {n_expert}, 0);
-
-            layer.ffn_gate_exps = create_tensor(
-                tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i),
-                {n_embd, n_ff_exp, n_expert}, expert_flags);
-            layer.ffn_up_exps = create_tensor(
-                tn(LLM_TENSOR_FFN_UP_EXPS, "weight", i),
-                {n_embd, n_ff_exp, n_expert}, expert_flags);
-            layer.ffn_down_exps = create_tensor(
-                tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i),
-                {n_ff_exp, n_embd, n_expert}, expert_flags);
-
-            // Shared experts stay resident and are never part of the longhaul cache.
-            layer.ffn_gate_shexp = create_tensor(
-                tn(LLM_TENSOR_FFN_GATE_SHEXPS, "weight", i),
-                {n_embd, n_ff_exp, n_shexp}, 0);
-            layer.ffn_up_shexp = create_tensor(
-                tn(LLM_TENSOR_FFN_UP_SHEXPS, "weight", i),
-                {n_embd, n_ff_exp, n_shexp}, 0);
-            layer.ffn_down_shexp = create_tensor(
-                tn(LLM_TENSOR_FFN_DOWN_SHEXPS, "weight", i),
-                {n_ff_exp, n_embd, n_shexp}, 0);
-        }
-    }
-}
-
-class llm_graph_input_inkling : public llm_graph_input_i {
-public:
-    llm_graph_input_inkling(
-            const llama_hparams & hparams,
-            const llama_memory_hybrid_iswa_context * mctx) :
-        hparams(hparams), mctx(mctx) {
-    }
-
-    void set_input(const llama_ubatch * ubatch) override {
-        if (tau) {
-            GGML_ASSERT(ggml_backend_buffer_is_host(tau->buffer));
-            float * data = (float *) tau->data;
-            const float n_floor = (float) hparams.inkling_log_n_floor;
-            const float alpha   = hparams.inkling_log_alpha;
-            for (int64_t i = 0; i < (int64_t) ubatch->n_tokens; ++i) {
-                const float eff = (float) (ubatch->pos[i] + 1)/n_floor;
-                data[i] = 1.0f + alpha*logf(std::max(eff, 1.0f));
-            }
-        }
-
-        if (rel_idx) {
-            mctx->get_attn()->get_base()->set_input_pos_rel_flat(
-                rel_idx, ubatch, hparams.inkling_rel_extent);
-        }
-        if (rel_idx_swa) {
-            mctx->get_attn()->get_swa()->set_input_pos_rel_flat(
-                rel_idx_swa, ubatch, hparams.inkling_rel_extent_swa);
-        }
-
-        if (vocab_mask) {
-            GGML_ASSERT(ggml_backend_buffer_is_host(vocab_mask->buffer));
-            float * data = (float *) vocab_mask->data;
-            for (int64_t id = 0; id < vocab_mask->ne[0]; ++id) {
-                data[id] = id < (int64_t) hparams.inkling_unpadded_n_vocab ? 0.0f : -INFINITY;
-            }
-        }
-
-        if (shexp_idx) {
-            GGML_ASSERT(ggml_backend_buffer_is_host(shexp_idx->buffer));
-            int32_t * data = (int32_t *) shexp_idx->data;
-            for (int64_t j = 0; j < shexp_idx->ne[1]; ++j) {
-                for (int64_t s = 0; s < shexp_idx->ne[0]; ++s) {
-                    data[j*shexp_idx->ne[0] + s] = (int32_t) s;
-                }
-            }
-        }
-    }
-
-    ggml_tensor * tau         = nullptr;
-    ggml_tensor * rel_idx     = nullptr;
-    ggml_tensor * rel_idx_swa = nullptr;
-    ggml_tensor * vocab_mask  = nullptr;
-    ggml_tensor * shexp_idx   = nullptr;
-
-private:
-    const llama_hparams hparams;
-    const llama_memory_hybrid_iswa_context * mctx;
-};
-
-std::unique_ptr<llm_graph_context> llama_model_inkling::build_arch_graph(
-        const llm_graph_params & params) const {
-    return std::make_unique<graph>(*this, params);
-}
-
-llama_model_inkling::graph::graph(
-        const llama_model & model,
-        const llm_graph_params & params) :
-    llm_graph_context(params) {
-    const int64_t head_dim = hparams.n_embd_head_k();
-    const int64_t d_rel    = hparams.inkling_d_rel;
-    const int64_t d_conv   = hparams.n_shortconv_l_cache - 1;
-    const int64_t n_embd_r = hparams.n_embd_r();
-    const int64_t kw_max   = hparams.n_embd_k_gqa_max();
-    const int64_t vw_max   = hparams.n_embd_v_gqa_max();
-
-    const int64_t off_k    = 0;
-    const int64_t off_v    = d_conv*kw_max;
-    const int64_t off_attn = d_conv*(kw_max + vw_max);
-    const int64_t off_mlp  = d_conv*(kw_max + vw_max + n_embd);
-
-    const auto * mctx_hyb  = static_cast<const llama_memory_hybrid_iswa_context *>(mctx);
-    const auto * mctx_recr = mctx_hyb->get_recr();
-    const auto * mctx_attn = mctx_hyb->get_attn();
-    const uint32_t kv_head = mctx_recr->get_head();
-
-    const int64_t n_seq_tokens = ubatch.n_seq_tokens;
-    const int64_t n_seqs       = ubatch.n_seqs;
-    GGML_ASSERT(n_seqs != 0);
-    GGML_ASSERT(ubatch.equal_seqs());
-    GGML_ASSERT(ubatch.n_tokens == n_seq_tokens*n_seqs);
-
-    bool has_global = false;
-    bool has_local  = false;
-    for (int il = 0; il < n_layer; ++il) {
-        has_local  |= hparams.is_swa(il);
-        has_global |= !hparams.is_swa(il);
-    }
-
-    auto inp = std::make_unique<llm_graph_input_inkling>(hparams, mctx_hyb);
-    if (hparams.inkling_log_n_floor > 0 && has_global) {
-        inp->tau = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, 1, 1, n_tokens);
-        ggml_set_input(inp->tau);
-        ggml_set_name(inp->tau, "inkling_tau");
-    }
-    if (has_global) {
-        inp->rel_idx = ggml_new_tensor_2d(
-            ctx0, GGML_TYPE_I32, mctx_attn->get_base()->get_n_kv(), n_tokens);
-        ggml_set_input(inp->rel_idx);
-        ggml_set_name(inp->rel_idx, "inkling_rel_idx");
-    }
-    if (has_local) {
-        inp->rel_idx_swa = ggml_new_tensor_2d(
-            ctx0, GGML_TYPE_I32, mctx_attn->get_swa()->get_n_kv(), n_tokens);
-        ggml_set_input(inp->rel_idx_swa);
-        ggml_set_name(inp->rel_idx_swa, "inkling_rel_idx_swa");
-    }
-
-    const int64_t n_vocab = model.vocab.n_tokens();
-    if (!cparams.embeddings &&
-            hparams.inkling_unpadded_n_vocab > 0 &&
-            (int64_t) hparams.inkling_unpadded_n_vocab < n_vocab) {
-        inp->vocab_mask = ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, n_vocab);
-        ggml_set_input(inp->vocab_mask);
-        ggml_set_name(inp->vocab_mask, "inkling_vocab_mask");
-    }
-    if (hparams.n_expert_shared > 0 && (uint32_t) n_layer > hparams.n_layer_dense_lead) {
-        inp->shexp_idx = ggml_new_tensor_2d(
-            ctx0, GGML_TYPE_I32, hparams.n_expert_shared, n_tokens);
-        ggml_set_input(inp->shexp_idx);
-        ggml_set_name(inp->shexp_idx, "inkling_shexp_idx");
-    }
-
-    ggml_tensor * tau         = inp->tau;
-    ggml_tensor * rel_idx     = inp->rel_idx;
-    ggml_tensor * rel_idx_swa = inp->rel_idx_swa;
-    ggml_tensor * vocab_mask  = inp->vocab_mask;
-    ggml_tensor * shexp_idx   = inp->shexp_idx;
-    res->add_input(std::move(inp));
-
-    auto * inp_hybrid = build_inp_mem_hybrid_iswa();
-    ggml_tensor * conv_rs_cur = nullptr;
-
-    auto build_sconv = [&](ggml_tensor * x2d, ggml_tensor * kernel, int64_t off, int il) {
-        const int64_t w = x2d->ne[0];
-        ggml_tensor * x3 = ggml_reshape_3d(ctx0, x2d, w, n_seq_tokens, n_seqs);
-        ggml_tensor * xt = ggml_transpose(ctx0, x3);
-
-        ggml_tensor * conv_state = mctx_recr->get_r_l(il);
-        GGML_ASSERT(conv_rs_cur != nullptr);
-        const size_t sz = ggml_element_size(conv_rs_cur);
-
-        ggml_tensor * state = ggml_view_3d(
-            ctx0, conv_rs_cur, d_conv, w, n_seqs,
-            d_conv*sz, conv_rs_cur->nb[1], off*sz);
-        ggml_tensor * sx = ggml_concat(ctx0, state, xt, 0);
-
-        ggml_tensor * new_state = ggml_view_3d(
-            ctx0, sx, d_conv, w, n_seqs,
-            sx->nb[1], sx->nb[2], (sx->ne[0] - d_conv)*sx->nb[0]);
-        ggml_tensor * state_dst = ggml_view_3d(
-            ctx0, conv_state, d_conv, w, n_seqs,
-            d_conv*sz, n_embd_r*sz, (kv_head*n_embd_r + off)*sz);
-        ggml_build_forward_expand(gf, ggml_cpy(ctx0, new_state, state_dst));
-
-        ggml_tensor * conv_out = ggml_ssm_conv(ctx0, sx, kernel);
-        return ggml_reshape_2d(ctx0, ggml_add(ctx0, x3, conv_out), w, n_seq_tokens*n_seqs);
-    };
-
-    auto build_attn_block = [&](ggml_tensor * cur, int il) {
-        const auto & layer = model.layers[il];
-        const bool is_swa = hparams.is_swa(il);
-        const int64_t n_head_kv  = hparams.n_head_kv(il);
-        const int64_t rel_extent = is_swa
-            ? hparams.inkling_rel_extent_swa
-            : hparams.inkling_rel_extent;
-
-        ggml_tensor * q = build_lora_mm(layer.wq, cur);
-        ggml_tensor * k = build_lora_mm(layer.wk, cur);
-        ggml_tensor * v = build_lora_mm(layer.wv, cur);
-        ggml_tensor * r = build_lora_mm(layer.wr, cur);
-
-        k = build_sconv(k, layer.shortconv_k, off_k, il);
-        v = build_sconv(v, layer.shortconv_v, off_v, il);
-
-        q = ggml_reshape_3d(ctx0, q, head_dim, n_head, n_tokens);
-        k = ggml_reshape_3d(ctx0, k, head_dim, n_head_kv, n_tokens);
-        v = ggml_reshape_3d(ctx0, v, head_dim, n_head_kv, n_tokens);
-        q = build_norm(q, layer.attn_q_norm, nullptr, LLM_NORM_RMS, il);
-        k = build_norm(k, layer.attn_k_norm, nullptr, LLM_NORM_RMS, il);
-
-        if (tau && !is_swa) {
-            q = ggml_mul(ctx0, q, tau);
-        }
-
-        ggml_tensor * r2 = ggml_reshape_2d(ctx0, r, d_rel, n_head*n_tokens);
-        ggml_tensor * proj = ggml_cont(ctx0, ggml_transpose(ctx0, layer.attn_rel_proj));
-        ggml_tensor * rel = ggml_mul_mat(ctx0, proj, r2);
-        ggml_mul_mat_set_prec(rel, GGML_PREC_F32);
-        rel = ggml_reshape_3d(ctx0, rel, rel_extent, n_head, n_tokens);
-        if (tau && !is_swa) {
-            rel = ggml_mul(ctx0, rel, tau);
-        }
-
-        // Scale Q rather than build_attn's joint logits so the relative term
-        // remains unscaled, matching the reference implementation.
-        q = ggml_scale(ctx0, q, 1.0f/(float) head_dim);
-
-        rel = ggml_pad(ctx0, rel, 1, 0, 0, 0);
-        rel = ggml_cont(ctx0, ggml_permute(ctx0, rel, 1, 0, 2, 3));
-        rel = ggml_reshape_2d(ctx0, rel, n_head, (rel_extent + 1)*n_tokens);
-
-        ggml_tensor * idx = is_swa ? rel_idx_swa : rel_idx;
-        GGML_ASSERT(idx != nullptr);
-        const int64_t n_kv = idx->ne[0];
-        ggml_tensor * idx1 = ggml_reshape_1d(ctx0, idx, n_kv*n_tokens);
-        ggml_tensor * kq_b = ggml_get_rows(ctx0, rel, idx1);
-        kq_b = ggml_reshape_3d(ctx0, kq_b, n_head, n_kv, n_tokens);
-        kq_b = ggml_cont(ctx0, ggml_permute(ctx0, kq_b, 2, 0, 1, 3));
-
-        auto * inp_attn = inp_hybrid->get_attn();
-        const int64_t n_stream =
-            (is_swa ? inp_attn->get_kq_mask_swa() : inp_attn->get_kq_mask())->ne[3];
-        if (n_stream > 1) {
-            kq_b = ggml_view_4d(
-                ctx0, kq_b, n_kv, n_tokens/n_stream, n_head, n_stream,
-                kq_b->nb[1], kq_b->nb[2], (n_tokens/n_stream)*kq_b->nb[1], 0);
-        }
-
-        return build_attn(
-            inp_attn, layer.wo, nullptr, nullptr,
-            q, k, v, kq_b, nullptr, nullptr, 1.0f, il);
-    };
-
-    auto build_dense_ffn = [&](ggml_tensor * cur, int il) {
-        cur = build_ffn(
-            cur,
-            model.layers[il].ffn_up,   nullptr, nullptr,
-            model.layers[il].ffn_gate, nullptr, nullptr,
-            model.layers[il].ffn_down, nullptr, nullptr,
-            nullptr, LLM_FFN_SILU, LLM_FFN_PAR, il);
-        return ggml_mul(ctx0, cur, model.layers[il].ffn_gscale);
-    };
-
-    auto build_moe = [&](ggml_tensor * cur, int il) {
-        const auto & layer = model.layers[il];
-        const int64_t n_shexp = hparams.n_expert_shared;
-
-        ggml_tensor * logits = build_lora_mm(layer.ffn_gate_inp, cur);
-        ggml_mul_mat_set_prec(logits, GGML_PREC_F32);
-        const size_t lsz = ggml_element_size(logits);
-        ggml_tensor * routed = ggml_cont(ctx0, ggml_view_2d(
-            ctx0, logits, n_expert, n_tokens, logits->nb[1], 0));
-        ggml_tensor * shared_logits = ggml_view_2d(
-            ctx0, logits, n_shexp, n_tokens, logits->nb[1], n_expert*lsz);
-
-        ggml_tensor * scores = ggml_add(
-            ctx0, ggml_sigmoid(ctx0, routed), layer.ffn_exp_probs_b);
-        ggml_tensor * selected = ggml_argsort_top_k(
-            ctx0, scores, n_expert_used);
-
-        ggml_tensor * routed3 = ggml_reshape_3d(ctx0, routed, 1, n_expert, n_tokens);
-        ggml_tensor * topk_logits = ggml_reshape_2d(
-            ctx0, ggml_get_rows(ctx0, routed3, selected), n_expert_used, n_tokens);
-        ggml_tensor * all_logits = ggml_concat(ctx0, topk_logits, shared_logits, 0);
-
-        ggml_tensor * w = ggml_neg(
-            ctx0, ggml_softplus(ctx0, ggml_neg(ctx0, all_logits)));
-        w = ggml_soft_max(ctx0, w);
-        w = ggml_scale(ctx0, w, hparams.expert_weights_scale);
-        w = ggml_mul(ctx0, w, layer.ffn_gscale);
-
-        const size_t wsz = ggml_element_size(w);
-        ggml_tensor * weights = ggml_cont(ctx0, ggml_view_2d(
-            ctx0, w, n_expert_used, n_tokens, w->nb[1], 0));
-        weights = ggml_reshape_3d(ctx0, weights, 1, n_expert_used, n_tokens);
-
-        ggml_tensor * moe_out = build_moe_experts(
-            cur, selected, weights,
-            layer.ffn_up_exps, layer.ffn_gate_exps, layer.ffn_down_exps,
-            n_expert_used, LLM_FFN_SILU, il);
-
-        GGML_ASSERT(shexp_idx != nullptr);
-        ggml_tensor * xr = ggml_reshape_3d(ctx0, cur, n_embd, 1, n_tokens);
-        ggml_tensor * gs = build_lora_mm_id(layer.ffn_gate_shexp, xr, shexp_idx);
-        ggml_tensor * us = build_lora_mm_id(layer.ffn_up_shexp,   xr, shexp_idx);
-        ggml_tensor * hs = ggml_swiglu_split(ctx0, gs, us);
-
-        ggml_tensor * gammas = ggml_cont(ctx0, ggml_view_2d(
-            ctx0, w, n_shexp, n_tokens, w->nb[1], n_expert_used*wsz));
-        hs = ggml_mul(ctx0, hs, ggml_reshape_3d(ctx0, gammas, 1, n_shexp, n_tokens));
-        ggml_tensor * ds = build_lora_mm_id(layer.ffn_down_shexp, hs, shexp_idx);
-        for (int64_t s = 0; s < n_shexp; ++s) {
-            ggml_tensor * e = ggml_view_2d(
-                ctx0, ds, n_embd, n_tokens, ds->nb[2], s*ds->nb[1]);
-            moe_out = ggml_add(ctx0, moe_out, e);
-        }
-        return moe_out;
-    };
-
-    ggml_tensor * cur = build_inp_embd(model.tok_embd);
-    if (ubatch.token) {
-        cur = build_norm(cur, model.tok_norm, nullptr, LLM_NORM_RMS, -1);
-    }
-    ggml_build_forward_expand(gf, cur);
-
-    for (int il = 0; il < n_layer; ++il) {
-        conv_rs_cur = build_rs(
-            inp_hybrid->get_recr(), mctx_recr->get_r_l(il), n_embd_r, n_seqs);
-
-        ggml_tensor * attn_in = build_norm(
-            cur, model.layers[il].attn_norm, nullptr, LLM_NORM_RMS, il);
-        ggml_tensor * attn_out = build_attn_block(attn_in, il);
-        attn_out = build_sconv(
-            attn_out, model.layers[il].shortconv_attn, off_attn, il);
-        cur = ggml_add(ctx0, cur, attn_out);
-
-        ggml_tensor * ffn_in = build_norm(
-            cur, model.layers[il].ffn_norm, nullptr, LLM_NORM_RMS, il);
-        ggml_tensor * ffn_out = il < (int) hparams.n_layer_dense_lead
-            ? build_dense_ffn(ffn_in, il)
-            : build_moe(ffn_in, il);
-        ffn_out = build_sconv(
-            ffn_out, model.layers[il].shortconv_mlp, off_mlp, il);
-        cur = ggml_add(ctx0, cur, ffn_out);
-
-        cur = build_cvec(cur, il);
-        cb(cur, "l_out", il);
-    }
-
-    ggml_tensor * inp_out_ids = build_inp_out_ids();
-    if (inp_out_ids) {
-        cur = ggml_get_rows(ctx0, cur, inp_out_ids);
-    }
-
-    cur = build_norm(cur, model.output_norm, nullptr, LLM_NORM_RMS, -1);
-    res->t_embd = cur;
-
-    if (!cparams.embeddings) {
-        cur = ggml_scale(ctx0, cur, hparams.f_logit_scale);
-        cur = build_lora_mm(model.output, cur);
-        if (model.output->type == GGML_TYPE_F32) {
-            ggml_mul_mat_set_prec(cur, GGML_PREC_F32);
-        }
-        if (vocab_mask) {
-            cur = ggml_add(ctx0, cur, vocab_mask);
-        }
-        res->t_logits = cur;
-    }
-
-    ggml_build_forward_expand(gf, cur);
-}

src/models/models.h

@@ -1865,18 +1865,6 @@ struct llama_model_lfm2moe : public llama_model_base {
     std::unique_ptr<llm_graph_context> build_arch_graph(const llm_graph_params & params) const override;
 };
 
-struct llama_model_inkling : public llama_model_base {
-    llama_model_inkling(const struct llama_model_params & params) : llama_model_base(params) {}
-    void load_arch_hparams(llama_model_loader & ml) override;
-    void load_arch_tensors(llama_model_loader & ml) override;
-
-    struct graph : public llm_graph_context {
-        graph(const llama_model & model, const llm_graph_params & params);
-    };
-
-    std::unique_ptr<llm_graph_context> build_arch_graph(const llm_graph_params & params) const override;
-};
-
 
 struct llama_model_smallthinker : public llama_model_base {
     llama_model_smallthinker(const struct llama_model_params & params) : llama_model_base(params) {}

tests/CMakeLists.txt

@@ -259,40 +259,8 @@ set_tests_properties(test-thread-safety PROPERTIES FIXTURES_REQUIRED test-downlo
 
 llama_build_and_test(test-arg-parser.cpp)
 llama_build_and_test(test-longhaul.cpp)
-llama_test(
-    test-longhaul
-    NAME test-longhaul-cpu-qwen35moe
-    ARGS --cpu-model "${MODEL_DIR}/qwen35moe-moe.gguf" 3145728
-)
-set_tests_properties(test-longhaul-cpu-qwen35moe PROPERTIES
-    FIXTURES_REQUIRED generate-models
-)
-llama_test(
-    test-longhaul
-    NAME test-longhaul-cpu-laguna
-    ARGS --cpu-model "${MODEL_DIR}/laguna-moe.gguf" 2097152
-)
-set_tests_properties(test-longhaul-cpu-laguna PROPERTIES
-    FIXTURES_REQUIRED generate-models
-)
-llama_test(
-    test-longhaul
-    NAME test-longhaul-cpu-inkling
-    ARGS --cpu-model "${MODEL_DIR}/inkling-moe.gguf" 73728
-)
-set_tests_properties(test-longhaul-cpu-inkling PROPERTIES
-    FIXTURES_REQUIRED generate-models
-)
-if (APPLE AND GGML_METAL)
-    llama_test(
-        test-longhaul
-        NAME test-longhaul-metal-inkling
-        ARGS --metal-model "${MODEL_DIR}/inkling-moe.gguf" 73728
-    )
-    set_tests_properties(test-longhaul-metal-inkling PROPERTIES
-        FIXTURES_REQUIRED generate-models
-    )
-endif()
+llama_build_and_test(test-tokenizer-longhaul.cpp
+    ARGS ${PROJECT_SOURCE_DIR}/models/ggml-vocab-qwen35.gguf)
 llama_build_and_test(test-token-cache.cpp)
 target_include_directories(test-token-cache PRIVATE ${PROJECT_SOURCE_DIR}/src)
 

tests/test-arg-parser.cpp

@@ -160,6 +160,21 @@ static void test(void) {
     assert(params.load_mode == LLAMA_LOAD_MODE_LONGHAUL);
     assert(params.longhaul_cache_bytes == 2ULL * 1024 * 1024 * 1024);
 
+    params.tokenizer_longhaul = false;
+    params.tokenizer_longhaul_cache_bytes = 0;
+    argv = {"binary_name", "--tokenizer-longhaul"};
+    assert(false == common_params_parse(argv.size(), list_str_to_char(argv).data(), params, LLAMA_EXAMPLE_COMMON));
+
+    argv = {"binary_name", "--tokenizer-longhaul", "--tokenizer-longhaul-cache", "64"};
+    assert(true == common_params_parse(argv.size(), list_str_to_char(argv).data(), params, LLAMA_EXAMPLE_COMMON));
+    assert(params.tokenizer_longhaul);
+    assert(params.tokenizer_longhaul_cache_bytes == 64ULL * 1024 * 1024);
+
+    params.tokenizer_longhaul = false;
+    params.tokenizer_longhaul_cache_bytes = 0;
+    argv = {"binary_name", "--tokenizer-longhaul-cache", "64"};
+    assert(false == common_params_parse(argv.size(), list_str_to_char(argv).data(), params, LLAMA_EXAMPLE_COMMON));
+
     argv = {"binary_name", "--token-cache-size", "256", "--token-cache-dir", "/tmp/llama-token-cache-test.sqlite3"};
     assert(true == common_params_parse(argv.size(), list_str_to_char(argv).data(), params, LLAMA_EXAMPLE_COMMON));
     assert(params.token_cache_size_mib == 256);

tests/test-chat-auto-parser.cpp

@@ -91,7 +91,6 @@ static void test_normalize_quotes_with_embedded_quotes(testing & t);
 static void test_tagged_args_with_embedded_quotes(testing & t);
 
 static void test_role_markers_all_templates(testing & t);
-static void test_inkling_tml_parser(testing & t);
 
 int main(int argc, char * argv[]) {
     testing t(std::cout);
@@ -118,7 +117,6 @@ int main(int argc, char * argv[]) {
     t.test("standard_json_tools", test_standard_json_tools_formats);
     t.test("normalize_quotes_to_json", test_normalize_quotes_to_json);
     t.test("tagged_args_embedded_quotes", test_tagged_args_with_embedded_quotes);
-    t.test("inkling_tml", test_inkling_tml_parser);
     t.test("role_markers_all_templates", test_role_markers_all_templates);
 
     return t.summary();
@@ -2077,53 +2075,6 @@ static void test_role_markers_all_templates(testing & t) {
     }
 }
 
-static void test_inkling_tml_parser(testing & t) {
-    const std::string tmpl_src =
-        "{# <|content_thinking|><|content_text|> #}"
-        "{%- for message in messages -%}"
-        "<|message_user|><|content_text|>{{ message.content }}<|end_message|>"
-        "{%- endfor -%}"
-        "{%- if add_generation_prompt -%}<|message_model|>{%- endif -%}";
-
-    common_chat_templates_ptr tmpls =
-        common_chat_templates_init(nullptr, tmpl_src);
-    common_chat_templates_inputs inputs;
-    common_chat_msg user_message;
-    user_message.role = "user";
-    user_message.content = "Hello";
-    inputs.messages = { user_message };
-    inputs.add_generation_prompt = true;
-    inputs.reasoning_format = COMMON_REASONING_FORMAT_DEEPSEEK;
-    inputs.use_jinja = true;
-
-    const common_chat_params params =
-        common_chat_templates_apply(tmpls.get(), inputs);
-    t.assert_equal(
-        "Inkling uses the native PEG parser",
-        COMMON_CHAT_FORMAT_PEG_NATIVE,
-        params.format);
-    t.assert_true("Inkling exposes thinking blocks", params.supports_thinking);
-    t.assert_equal(
-        "Inkling generation marker",
-        "<|message_model|>",
-        params.generation_prompt);
-
-    common_peg_arena arena;
-    arena.load(params.parser);
-    common_chat_parser_params parser_params(params);
-    parser_params.reasoning_format = COMMON_REASONING_FORMAT_DEEPSEEK;
-
-    const common_chat_msg parsed = common_chat_peg_parse(
-        arena,
-        "<|content_thinking|>plan<|end_message|>"
-        "<|message_model|><|content_text|>answer<|end_message|>"
-        "<|content_model_end_sampling|>",
-        false,
-        parser_params);
-    t.assert_equal("reasoning block", "plan", parsed.reasoning_content);
-    t.assert_equal("text block", "answer", parsed.content);
-}
-
 // Test that reproduces the Seed-OSS template issue with embedded quotes
 static void test_tagged_args_with_embedded_quotes(testing & t) {
     json tools = build_edit_tool();
@@ -2241,3 +2192,4 @@ static void test_tagged_args_with_embedded_quotes(testing & t) {
         }
     }
 }
+

tests/test-llama-archs.cpp

@@ -9,7 +9,6 @@
 
 // TODO: replace with #include "llama-ext.h" in the future
 #include "../src/llama-arch.h"
-#include "../src/llama-model.h"
 #include "../src/llama-model-saver.h"
 
 #include <cinttypes>
@@ -114,11 +113,6 @@ static gguf_context_ptr get_gguf_ctx(const llm_arch arch, const bool moe) {
         n_layer = 3;
     } else if (arch == LLM_ARCH_CHAMELEON) {
         n_vocab = 10240;
-    } else if (arch == LLM_ARCH_INKLING) {
-        n_embd  = 64;
-        n_head  = 2;
-        n_ff    = 96;
-        n_layer = 2;
     }
 
     const uint32_t n_embd_head = n_embd / n_head;
@@ -203,10 +197,6 @@ static gguf_context_ptr get_gguf_ctx(const llm_arch arch, const bool moe) {
             pattern.push_back(il % 2);
         }
         ms.add_kv(LLM_KV_ATTENTION_SLIDING_WINDOW_PATTERN, pattern);
-    } else if (arch == LLM_ARCH_INKLING) {
-        ms.add_kv(
-            LLM_KV_ATTENTION_SLIDING_WINDOW_PATTERN,
-            std::vector<uint32_t>({1, 0}));
     } else {
         ms.add_kv(LLM_KV_ATTENTION_SLIDING_WINDOW_PATTERN, uint32_t(2));
     }
@@ -226,27 +216,12 @@ static gguf_context_ptr get_gguf_ctx(const llm_arch arch, const bool moe) {
     if (moe) {
         ms.add_kv(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, n_ff);
         ms.add_kv(LLM_KV_INTERLEAVE_MOE_LAYER_STEP,  uint32_t(2));
-        ms.add_kv(LLM_KV_EXPERT_COUNT,               arch == LLM_ARCH_INKLING ? uint32_t(8) : uint32_t(2));
-        ms.add_kv(LLM_KV_EXPERT_USED_COUNT,          arch == LLM_ARCH_INKLING ? uint32_t(6) : uint32_t(1));
-        ms.add_kv(LLM_KV_EXPERT_SHARED_COUNT,        arch == LLM_ARCH_INKLING ? uint32_t(2) : uint32_t(1));
+        ms.add_kv(LLM_KV_EXPERT_COUNT,               uint32_t(2));
+        ms.add_kv(LLM_KV_EXPERT_USED_COUNT,          uint32_t(1));
+        ms.add_kv(LLM_KV_EXPERT_SHARED_COUNT,        uint32_t(1));
         ms.add_kv(LLM_KV_EXPERT_GATING_FUNC,         uint32_t(2)); // sigmoid
         ms.add_kv(LLM_KV_EXPERT_GROUP_SCALE,         1.0f);
         ms.add_kv(LLM_KV_EXPERTS_PER_GROUP,          uint32_t(1));
-        if (arch == LLM_ARCH_INKLING) {
-            ms.add_kv(LLM_KV_EXPERT_WEIGHTS_SCALE, 1.0f);
-        }
-    }
-
-    if (arch == LLM_ARCH_INKLING) {
-        ms.add_kv(LLM_KV_INKLING_D_REL,               uint32_t(8));
-        ms.add_kv(LLM_KV_INKLING_REL_EXTENT,          uint32_t(32));
-        ms.add_kv(LLM_KV_INKLING_REL_EXTENT_SWA,      uint32_t(16));
-        ms.add_kv(LLM_KV_INKLING_SHORTCONV_KERNEL,    uint32_t(4));
-        ms.add_kv(LLM_KV_INKLING_DENSE_BLOCK_COUNT,   uint32_t(1));
-        ms.add_kv(LLM_KV_INKLING_LOGIT_SCALE_DENOM,   1.0f);
-        ms.add_kv(LLM_KV_INKLING_LOG_SCALING_N_FLOOR, uint32_t(16));
-        ms.add_kv(LLM_KV_INKLING_LOG_SCALING_ALPHA,   0.1f);
-        ms.add_kv(LLM_KV_INKLING_UNPADDED_VOCAB_SIZE, n_vocab);
     }
 
     ms.add_kv(LLM_KV_POSNET_EMBEDDING_LENGTH,   n_embd);
@@ -396,7 +371,6 @@ static bool moe_mandatory(const llm_arch arch) {
         case LLM_ARCH_MISTRAL4:
         case LLM_ARCH_MELLUM:
         case LLM_ARCH_LAGUNA:
-        case LLM_ARCH_INKLING:
             return true;
         default:
             return false;
@@ -502,11 +476,7 @@ static int save_models(const llm_arch target_arch, const size_t seed, const ggml
             if (!moe && moe_mandatory(arch)) {
                 continue;
             }
-            const bool can_save_fixture =
-                llama_model_saver_supports_arch(arch) ||
-                arch == LLM_ARCH_LAGUNA ||
-                arch == LLM_ARCH_INKLING;
-            if (!can_save_fixture || !arch_supported(arch)) {
+            if (!llama_model_saver_supports_arch(arch) || !arch_supported(arch)) {
                 LOG_INF("%s: %s model (%s) is unsupported, skipping\n", __func__, llm_arch_name(arch), moe ? "MoE" : "dense");
                 continue;
             }
@@ -514,26 +484,7 @@ static int save_models(const llm_arch target_arch, const size_t seed, const ggml
             auto model_and_ctx = get_model_and_ctx(gguf_ctx.get(), nullptr, seed, {});
             const std::string path = dir + "/" + llm_arch_name(arch) + (moe ? "-moe.gguf" : "-dense.gguf");
             LOG_INF("%s: Saving %s model (%s) to %s...\n", __func__, llm_arch_name(arch), moe ? "MoE" : "dense", path.c_str());
-            const bool private_fixture =
-                arch == LLM_ARCH_LAGUNA || arch == LLM_ARCH_INKLING;
-            if (llama_model_saver_supports_arch(arch) && !private_fixture) {
-                llama_model_save_to_file(model_and_ctx.first.get(), path.c_str());
-            } else {
-                // Laguna and Inkling are not supported by the production
-                // model saver yet.
-                // Preserve the exact synthetic fixture metadata and attach the
-                // initialized model tensors so longhaul can be tested from a
-                // real file without broadening the saver API.
-                gguf_context_ptr fixture(gguf_init_empty());
-                // Model initialization normalizes the caller's GGUF context,
-                // so regenerate the exact private metadata for the file.
-                gguf_context_ptr fixture_metadata = get_gguf_ctx(arch, moe);
-                gguf_set_kv(fixture.get(), fixture_metadata.get());
-                for (const auto & item : llama_internal_get_tensor_map(model_and_ctx.first.get())) {
-                    gguf_add_tensor(fixture.get(), item.second);
-                }
-                gguf_write_to_file(fixture.get(), path.c_str(), false);
-            }
+            llama_model_save_to_file(model_and_ctx.first.get(), path.c_str());
         }
     }
     llama_log_set(ud.original_logger.callback, ud.original_logger.user_data);

tests/test-longhaul.cpp

@@ -1,16 +1,11 @@
 #include "testing.h"
 
 #include "../src/llama-longhaul.h"
-#include "../src/llama-model.h"
 
 #include "ggml-backend.h"
-#include "llama-cpp.h"
 
-#include <cmath>
 #include <cstdio>
 #include <memory>
-#include <stdexcept>
-#include <string>
 #include <vector>
 
 struct longhaul_fixture {
@@ -133,32 +128,6 @@ static void test_multiple_sources(testing & t) {
     t.assert_equal(uint8_t(34), fixture.slot_value(fixture.weights_b, remapped[1]));
 }
 
-static void test_concurrent_source_reads(testing & t) {
-    longhaul_fixture fixture(2, 4, true);
-
-    for (int iteration = 0; iteration < 64; ++iteration) {
-        const int32_t first = iteration % 2 == 0 ? 0 : 2;
-        const int32_t second = first + 1;
-        const auto remapped = fixture.remap({first, second});
-
-        if (!t.assert_equal(2u, remapped.size())) {
-            return;
-        }
-        t.assert_equal(
-            uint8_t(1 + first),
-            fixture.slot_value(fixture.weights_a, remapped[0]));
-        t.assert_equal(
-            uint8_t(33 + first),
-            fixture.slot_value(fixture.weights_b, remapped[0]));
-        t.assert_equal(
-            uint8_t(1 + second),
-            fixture.slot_value(fixture.weights_a, remapped[1]));
-        t.assert_equal(
-            uint8_t(33 + second),
-            fixture.slot_value(fixture.weights_b, remapped[1]));
-    }
-}
-
 static void test_invalid_ids(testing & t) {
     longhaul_fixture fixture(2, 4);
 
@@ -183,116 +152,8 @@ static void test_read_failure_recovery(testing & t) {
     t.assert_equal(uint8_t(1), fixture.slot_value(fixture.weights_a, remapped[0]));
 }
 
-struct cpu_decode_result {
-    std::vector<float> logits;
-    uint32_t capacity = 0;
-    uint64_t misses = 0;
-    uint64_t bytes_read = 0;
-    bool expert_buffer_is_host = false;
-};
-
-static cpu_decode_result decode_model(
-        const std::string & path,
-        llama_load_mode load_mode,
-        uint64_t cache_bytes,
-        int32_t n_gpu_layers) {
-    llama_model_params model_params = llama_model_default_params();
-    model_params.n_gpu_layers = n_gpu_layers;
-    model_params.load_mode = load_mode;
-    model_params.longhaul_cache_bytes = cache_bytes;
-    model_params.use_extra_bufts = true;
-
-    llama_model_ptr model(llama_model_load_from_file(path.c_str(), model_params));
-    if (!model) {
-        throw std::runtime_error("failed to load CPU model: " + path);
-    }
-
-    llama_context_params context_params = llama_context_default_params();
-    context_params.n_ctx = 32;
-    context_params.n_batch = 8;
-    context_params.n_ubatch = 8;
-    context_params.n_threads = 2;
-    context_params.n_threads_batch = 2;
-
-    llama_context_ptr context(llama_init_from_model(model.get(), context_params));
-    if (!context) {
-        throw std::runtime_error("failed to create CPU context: " + path);
-    }
-
-    std::vector<llama_token> tokens = {1, 2, 3, 4, 5, 6, 7, 8};
-    llama_batch batch = llama_batch_get_one(tokens.data(), tokens.size());
-    if (llama_decode(context.get(), batch) != 0) {
-        throw std::runtime_error("failed to decode CPU model: " + path);
-    }
-
-    const int32_t n_vocab = llama_vocab_n_tokens(llama_model_get_vocab(model.get()));
-    const float * logits = llama_get_logits_ith(context.get(), -1);
-    if (!logits) {
-        throw std::runtime_error("CPU model produced no logits: " + path);
-    }
-
-    cpu_decode_result result;
-    result.logits.assign(logits, logits + n_vocab);
-
-    if (load_mode == LLAMA_LOAD_MODE_LONGHAUL) {
-        llama_longhaul_cache * cache = model->longhaul_cache();
-        if (!cache) {
-            throw std::runtime_error("longhaul cache was not created: " + path);
-        }
-        result.capacity = cache->capacity();
-        result.misses = cache->misses();
-        result.bytes_read = cache->bytes_read_count();
-
-        for (const auto & item : llama_internal_get_tensor_map(model.get())) {
-            const std::string & name = item.first;
-            if (name.find("ffn_") != std::string::npos &&
-                name.find("_exps.weight") != std::string::npos &&
-                item.second->ne[2] == result.capacity) {
-                result.expert_buffer_is_host = ggml_backend_buffer_is_host(item.second->buffer);
-                break;
-            }
-        }
-    }
-
-    return result;
-}
-
-static void test_model(
-        testing & t,
-        const std::string & path,
-        uint64_t cache_bytes,
-        int32_t n_gpu_layers) {
-    const cpu_decode_result regular =
-        decode_model(path, LLAMA_LOAD_MODE_NONE, 0, n_gpu_layers);
-    const cpu_decode_result longhaul =
-        decode_model(path, LLAMA_LOAD_MODE_LONGHAUL, cache_bytes, n_gpu_layers);
-
-    t.assert_equal(1u, longhaul.capacity);
-    t.assert_true("longhaul loaded at least one expert", longhaul.misses > 0);
-    t.assert_true("longhaul read expert bytes", longhaul.bytes_read > 0);
-    if (n_gpu_layers == 0) {
-        t.assert_true(
-            "streamed CPU experts use a directly writable host buffer",
-            longhaul.expert_buffer_is_host);
-    }
-    t.assert_equal(regular.logits.size(), longhaul.logits.size());
-
-    double squared_error = 0.0;
-    double squared_reference = 0.0;
-    for (size_t i = 0; i < regular.logits.size(); ++i) {
-        const double delta = double(regular.logits[i]) - double(longhaul.logits[i]);
-        squared_error += delta * delta;
-        squared_reference += double(regular.logits[i]) * double(regular.logits[i]);
-    }
-    const double nmse = squared_reference > 0.0 ? squared_error / squared_reference : squared_error;
-    t.assert_true(
-        "longhaul CPU logits NMSE = " + std::to_string(nmse),
-        std::isfinite(nmse) && nmse <= 1e-6);
-}
-
 int main(int argc, char ** argv) {
     testing t;
-    llama_backend_init();
 
     const char * verbose = getenv("LLAMA_TEST_VERBOSE");
     if (verbose) {
@@ -302,31 +163,14 @@ int main(int argc, char ** argv) {
         llama_log_set([](ggml_log_level, const char *, void *) {}, nullptr);
     }
 
-    if (argc == 4 &&
-            (std::string(argv[1]) == "--cpu-model" ||
-             std::string(argv[1]) == "--metal-model")) {
-        const std::string path = argv[2];
-        const uint64_t cache_bytes = std::stoull(argv[3]);
-        const bool metal = std::string(argv[1]) == "--metal-model";
-        t.test(metal ? "metal_model" : "cpu_model", [&](testing & current) {
-            test_model(current, path, cache_bytes, metal ? 99 : 0);
-        });
-        const int result = t.summary();
-        llama_backend_free();
-        return result;
-    }
-
     if (argc > 1) {
         t.set_filter(argv[1]);
     }
 
     t.test("batch_planning",  test_batch_planning);
     t.test("multiple_sources", test_multiple_sources);
-    t.test("concurrent_reads", test_concurrent_source_reads);
     t.test("invalid_ids",     test_invalid_ids);
     t.test("read_failure",    test_read_failure_recovery);
 
-    const int result = t.summary();
-    llama_backend_free();
-    return result;
+    return t.summary();
 }

tests/test-tokenizer-longhaul.cpp

@@ -0,0 +1,144 @@
+#include "llama.h"
+#include "../src/llama-token-cache.h"
+#include "../src/llama-tokenizer-longhaul.h"
+
+#include <chrono>
+#include <cstring>
+#include <filesystem>
+#include <string>
+#include <thread>
+#include <vector>
+
+#undef NDEBUG
+#include <cassert>
+
+static std::vector<llama_token> tokenize(const llama_vocab * vocab, const std::string & text) {
+    int32_t count = llama_tokenize(vocab, text.data(), int32_t(text.size()), nullptr, 0, false, true);
+    assert(count < 0);
+    std::vector<llama_token> result(size_t(-count));
+    count = llama_tokenize(vocab, text.data(), int32_t(text.size()), result.data(), result.size(), false, true);
+    assert(count == int32_t(result.size()));
+    return result;
+}
+
+int main(int argc, char ** argv) {
+    namespace fs = std::filesystem;
+
+    const fs::path root = fs::temp_directory_path() /
+        ("llama-tokenizer-longhaul-test-" + std::to_string(
+            std::chrono::high_resolution_clock::now().time_since_epoch().count()));
+    const std::string db = (root / "token-cache.sqlite3").string();
+    const llama_token_cache_params params = {
+        db.c_str(),
+        128 * 1024 * 1024,
+    };
+    assert(llama_token_cache_configure(params));
+    assert(llama_token_cache_clear());
+
+    std::vector<llama_tokenizer_longhaul_reverse> reverse;
+    std::vector<llama_tokenizer_longhaul_merge> merges;
+    for (int i = 0; i < 200; ++i) {
+        reverse.push_back({"token-" + std::to_string(i), i});
+        if (i > 0) {
+            merges.push_back({
+                "token-" + std::to_string(i - 1),
+                "token-" + std::to_string(i),
+                i - 1,
+            });
+        }
+    }
+
+    {
+        llama_tokenizer_longhaul_index index("source:test", 512);
+        std::string error;
+        assert(index.prepare(reverse, merges, "tokenizer-id", error));
+        assert(error.empty());
+        assert(index.n_merges() == merges.size());
+
+        for (int i = 0; i < 200; ++i) {
+            assert(index.text_to_token("token-" + std::to_string(i)) == i);
+        }
+        assert(index.text_to_token("missing") == LLAMA_TOKEN_NULL);
+        for (int i = 1; i < 200; ++i) {
+            assert(index.find_bpe_rank(
+                "token-" + std::to_string(i - 1),
+                "token-" + std::to_string(i)) == i - 1);
+        }
+        assert(index.find_bpe_rank("missing", "pair") == -1);
+        assert(index.cache_bytes_used() <= index.cache_capacity());
+
+        const auto loaded_merges = index.get_bpe_merges();
+        assert(loaded_merges.size() == merges.size());
+        assert(loaded_merges.front() == "token-0 token-1");
+        assert(loaded_merges.back() == "token-198 token-199");
+    }
+
+    // A warm open must use the completed artifact even without rebuild inputs.
+    {
+        llama_tokenizer_longhaul_index index("source:test", 128);
+        std::string error;
+        assert(index.prepare({}, {}, "tokenizer-id", error));
+        assert(index.text_to_token("token-42") == 42);
+
+        std::vector<std::thread> workers;
+        for (int thread = 0; thread < 4; ++thread) {
+            workers.emplace_back([&index, thread] {
+                for (int i = thread; i < 200; i += 4) {
+                    assert(index.text_to_token("token-" + std::to_string(i)) == i);
+                }
+            });
+        }
+        for (auto & worker : workers) worker.join();
+        assert(index.cache_bytes_used() <= index.cache_capacity());
+    }
+
+    if (argc > 1) {
+        llama_backend_init();
+
+        llama_model_params normal_params = llama_model_default_params();
+        normal_params.vocab_only = true;
+        llama_model * normal = llama_model_load_from_file(argv[1], normal_params);
+        assert(normal != nullptr);
+
+        llama_model_kv_override overrides[2] = {};
+        std::strcpy(overrides[0].key, "general.architecture");
+        overrides[0].tag = LLAMA_KV_OVERRIDE_TYPE_STR;
+        std::strcpy(overrides[0].val_str, "qwen35moe");
+
+        llama_model_params longhaul_params = llama_model_default_params();
+        longhaul_params.vocab_only = true;
+        longhaul_params.kv_overrides = overrides;
+        longhaul_params.tokenizer_longhaul = true;
+        longhaul_params.tokenizer_longhaul_cache_bytes = 1024;
+        llama_model * longhaul = llama_model_load_from_file(argv[1], longhaul_params);
+        assert(longhaul != nullptr);
+
+        const llama_vocab * normal_vocab = llama_model_get_vocab(normal);
+        const llama_vocab * longhaul_vocab = llama_model_get_vocab(longhaul);
+        for (const std::string & text : {
+                "Hello, world!",
+                " tokenizer longhaul",
+                "Qwen3.5: 12345\n",
+                "<|im_end|>",
+                "Unicode: Καλημέρα 世界 🚚",
+            }) {
+            const auto expected = tokenize(normal_vocab, text);
+            const auto actual = tokenize(longhaul_vocab, text);
+            assert(actual == expected);
+            for (const auto token : actual) {
+                assert(std::strcmp(
+                    llama_vocab_get_text(normal_vocab, token),
+                    llama_vocab_get_text(longhaul_vocab, token)) == 0);
+            }
+        }
+
+        llama_model_free(longhaul);
+        llama_model_free(normal);
+        llama_backend_free();
+    }
+
+    assert(llama_token_cache_clear());
+    std::error_code ec;
+    fs::remove_all(root, ec);
+    return 0;
+}

tools/cli/README.md

@@ -35,6 +35,8 @@
 | `--token-cache-size N` | shared tokenizer and tokenized-text cache size in MiB (default: 5120, 0 = disable)<br/>(env: LLAMA_ARG_TOKEN_CACHE_SIZE) |
 | `--token-cache-dir PATH` | path to the shared tokenizer and tokenized-text cache database<br/>(env: LLAMA_ARG_TOKEN_CACHE_DIR) |
 | `--token-cache-clear` | clear the shared tokenizer and tokenized-text cache before loading the model<br/>(env: LLAMA_ARG_TOKEN_CACHE_CLEAR) |
+| `--tokenizer-longhaul` | build and page the Qwen3.5/Laguna tokenizer through a bounded cache<br/>(env: LLAMA_ARG_TOKENIZER_LONGHAUL) |
+| `--tokenizer-longhaul-cache N` | tokenizer-longhaul RAM cache size in MiB<br/>(env: LLAMA_ARG_TOKENIZER_LONGHAUL_CACHE) |
 | `-fa, --flash-attn [on\|off\|auto]` | set Flash Attention use ('on', 'off', or 'auto', default: 'auto')<br/>(env: LLAMA_ARG_FLASH_ATTN) |
 | `-p, --prompt PROMPT` | prompt to start generation with; for system message, use -sys |
 | `--perf, --no-perf` | whether to enable internal libllama performance timings (default: false)<br/>(env: LLAMA_ARG_PERF) |
@@ -61,7 +63,7 @@
 | `--mmap, --no-mmap` | DEPRECATED in favor of `--load-mode`: whether to memory-map model. (if mmap disabled, slower load but may reduce pageouts if not using mlock)<br/>(env: LLAMA_ARG_MMAP) |
 | `-dio, --direct-io, -ndio, --no-direct-io` | DEPRECATED in favor of `--load-mode`: use DirectIO if available<br/>(env: LLAMA_ARG_DIO) |
 | `-lm, --load-mode MODE` | model loading mode (default: mmap)<br/>- none: no special loading mode<br/>- mmap: memory-map model (if mmap disabled, slower load but may reduce pageouts if not using mlock)<br/>- mlock: force system to keep model in RAM rather than swapping or compressing<br/>- mmap+mlock: mmap + force system to keep model in RAM rather than swapping or compressing<br/>- dio: use DirectIO if available<br/>- longhaul: stream routed MoE experts through a bounded cache<br/><br/>(env: LLAMA_ARG_LOAD_MODE) |
-| `--longhaul` | stream routed Qwen3.5 MoE, Laguna, or Inkling experts through a bounded CPU or Metal cache<br/>(env: LLAMA_ARG_LONGHAUL) |
+| `--longhaul` | stream routed Qwen3.5 MoE experts through a bounded Metal cache<br/>(env: LLAMA_ARG_LONGHAUL) |
 | `--longhaul-cache N` | longhaul expert cache size in GiB<br/>(env: LLAMA_ARG_LONGHAUL_CACHE) |
 | `--numa TYPE` | attempt optimizations that help on some NUMA systems<br/>- distribute: spread execution evenly over all nodes<br/>- isolate: only spawn threads on CPUs on the node that execution started on<br/>- numactl: use the CPU map provided by numactl<br/>if run without this previously, it is recommended to drop the system page cache before using this<br/>see https://github.com/ggml-org/llama.cpp/issues/1437<br/>(env: LLAMA_ARG_NUMA) |
 | `-dev, --device <dev1,dev2,..>` | comma-separated list of devices to use for offloading (none = don't offload)<br/>use --list-devices to see a list of available devices<br/>(env: LLAMA_ARG_DEVICE) |

tools/completion/README.md

@@ -118,6 +118,8 @@ llama-completion.exe -m models\gemma-1.1-7b-it.Q4_K_M.gguf --ignore-eos -n -1
 | `--token-cache-size N` | shared tokenizer and tokenized-text cache size in MiB (default: 5120, 0 = disable)<br/>(env: LLAMA_ARG_TOKEN_CACHE_SIZE) |
 | `--token-cache-dir PATH` | path to the shared tokenizer and tokenized-text cache database<br/>(env: LLAMA_ARG_TOKEN_CACHE_DIR) |
 | `--token-cache-clear` | clear the shared tokenizer and tokenized-text cache before loading the model<br/>(env: LLAMA_ARG_TOKEN_CACHE_CLEAR) |
+| `--tokenizer-longhaul` | build and page the Qwen3.5/Laguna tokenizer through a bounded cache<br/>(env: LLAMA_ARG_TOKENIZER_LONGHAUL) |
+| `--tokenizer-longhaul-cache N` | tokenizer-longhaul RAM cache size in MiB<br/>(env: LLAMA_ARG_TOKENIZER_LONGHAUL_CACHE) |
 | `-fa, --flash-attn [on\|off\|auto]` | set Flash Attention use ('on', 'off', or 'auto', default: 'auto')<br/>(env: LLAMA_ARG_FLASH_ATTN) |
 | `-p, --prompt PROMPT` | prompt to start generation with; for system message, use -sys |
 | `--perf, --no-perf` | whether to enable internal libllama performance timings (default: false)<br/>(env: LLAMA_ARG_PERF) |
@@ -144,7 +146,7 @@ llama-completion.exe -m models\gemma-1.1-7b-it.Q4_K_M.gguf --ignore-eos -n -1
 | `--mmap, --no-mmap` | DEPRECATED in favor of `--load-mode`: whether to memory-map model. (if mmap disabled, slower load but may reduce pageouts if not using mlock)<br/>(env: LLAMA_ARG_MMAP) |
 | `-dio, --direct-io, -ndio, --no-direct-io` | DEPRECATED in favor of `--load-mode`: use DirectIO if available<br/>(env: LLAMA_ARG_DIO) |
 | `-lm, --load-mode MODE` | model loading mode (default: mmap)<br/>- none: no special loading mode<br/>- mmap: memory-map model (if mmap disabled, slower load but may reduce pageouts if not using mlock)<br/>- mlock: force system to keep model in RAM rather than swapping or compressing<br/>- mmap+mlock: mmap + force system to keep model in RAM rather than swapping or compressing<br/>- dio: use DirectIO if available<br/>- longhaul: stream routed MoE experts through a bounded cache<br/><br/>(env: LLAMA_ARG_LOAD_MODE) |
-| `--longhaul` | stream routed Qwen3.5 MoE, Laguna, or Inkling experts through a bounded CPU or Metal cache<br/>(env: LLAMA_ARG_LONGHAUL) |
+| `--longhaul` | stream routed Qwen3.5 MoE experts through a bounded Metal cache<br/>(env: LLAMA_ARG_LONGHAUL) |
 | `--longhaul-cache N` | longhaul expert cache size in GiB<br/>(env: LLAMA_ARG_LONGHAUL_CACHE) |
 | `--numa TYPE` | attempt optimizations that help on some NUMA systems<br/>- distribute: spread execution evenly over all nodes<br/>- isolate: only spawn threads on CPUs on the node that execution started on<br/>- numactl: use the CPU map provided by numactl<br/>if run without this previously, it is recommended to drop the system page cache before using this<br/>see https://github.com/ggml-org/llama.cpp/issues/1437<br/>(env: LLAMA_ARG_NUMA) |
 | `-dev, --device <dev1,dev2,..>` | comma-separated list of devices to use for offloading (none = don't offload)<br/>use --list-devices to see a list of available devices<br/>(env: LLAMA_ARG_DEVICE) |

tools/llama-bench/README.md

@@ -57,6 +57,8 @@ test parameters:
   -ctk, --cache-type-k <t>                  (default: f16)
   -ctv, --cache-type-v <t>                  (default: f16)
   --longhaul-cache <GiB>                    expert cache size for longhaul mode
+  --tokenizer-longhaul                      page tokenizer indexes through a bounded cache
+  --tokenizer-longhaul-cache <MiB>          tokenizer lookup cache size
   -t, --threads <n>                         (default: system dependent)
   -C, --cpu-mask <hex,hex>                  (default: 0x0)
   --cpu-strict <0|1>                        (default: 0)

tools/llama-bench/llama-bench.cpp

@@ -342,6 +342,8 @@ struct cmd_params {
     std::vector<llama_split_mode>    split_mode;
     std::vector<llama_load_mode>     load_mode;
     uint64_t                         longhaul_cache_bytes;
+    bool                             tokenizer_longhaul;
+    uint64_t                         tokenizer_longhaul_cache_bytes;
     std::vector<int>                 main_gpu;
     std::vector<bool>                no_kv_offload;
     std::vector<llama_flash_attn_type> flash_attn;
@@ -387,6 +389,8 @@ static const cmd_params cmd_params_defaults = {
     /* split_mode           */ { LLAMA_SPLIT_MODE_LAYER },
     /* load_mode            */ { LLAMA_LOAD_MODE_MMAP },
     /* longhaul_cache_bytes */ 0,
+    /* tokenizer_longhaul   */ false,
+    /* tokenizer_longhaul_cache_bytes */ 0,
     /* main_gpu             */ { 0 },
     /* no_kv_offload        */ { false },
     /* flash_attn           */ { LLAMA_FLASH_ATTN_TYPE_AUTO },
@@ -463,6 +467,8 @@ static void print_usage(int /* argc */, char ** argv) {
     printf("  -dev, --device <dev0/dev1/...>                    (default: auto)\n");
     printf("  -lm, --load-mode <none|mmap|mlock|mmap+mlock|dio|longhaul> (default: %s)\n", join(transform_to_str(cmd_params_defaults.load_mode, llama_load_mode_name), ",").c_str());
     printf("  --longhaul-cache <GiB>                           expert cache size for longhaul mode\n");
+    printf("  --tokenizer-longhaul                             page tokenizer indexes through a bounded cache\n");
+    printf("  --tokenizer-longhaul-cache <MiB>                 tokenizer lookup cache size (required with --tokenizer-longhaul)\n");
     printf("  -mmp, --mmap <0|1>                                (DEPRECATED IN FAVOUR OF --load-mode)\n");
     printf("  -dio, --direct-io <0|1>                           (DEPRECATED IN FAVOUR OF --load-mode)\n");
     printf("  -embd, --embeddings <0|1>                         (default: %s)\n", join(cmd_params_defaults.embeddings, ",").c_str());
@@ -525,6 +531,8 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
     params.no_warmup            = cmd_params_defaults.no_warmup;
     params.offline              = cmd_params_defaults.offline;
     params.longhaul_cache_bytes = cmd_params_defaults.longhaul_cache_bytes;
+    params.tokenizer_longhaul = cmd_params_defaults.tokenizer_longhaul;
+    params.tokenizer_longhaul_cache_bytes = cmd_params_defaults.tokenizer_longhaul_cache_bytes;
 
     if (const char * env = getenv("HF_TOKEN")) {
         params.hf_token = env;
@@ -801,6 +809,19 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
                     break;
                 }
                 params.longhaul_cache_bytes = gib * 1024 * 1024 * 1024;
+            } else if (arg == "--tokenizer-longhaul") {
+                params.tokenizer_longhaul = true;
+            } else if (arg == "--tokenizer-longhaul-cache") {
+                if (++i >= argc) {
+                    invalid_param = true;
+                    break;
+                }
+                const uint64_t mib = std::stoull(argv[i]);
+                if (mib == 0 || mib > UINT64_MAX / 1024 / 1024) {
+                    invalid_param = true;
+                    break;
+                }
+                params.tokenizer_longhaul_cache_bytes = mib * 1024 * 1024;
             } else if (arg == "-mg" || arg == "--main-gpu") {
                 if (++i >= argc) {
                     invalid_param = true;
@@ -1157,6 +1178,10 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
         fprintf(stderr, "error: longhaul load mode requires --longhaul-cache N\n");
         exit(1);
     }
+    if (params.tokenizer_longhaul != (params.tokenizer_longhaul_cache_bytes != 0)) {
+        fprintf(stderr, "error: --tokenizer-longhaul and --tokenizer-longhaul-cache must be used together\n");
+        exit(1);
+    }
     if (params.main_gpu.empty()) {
         params.main_gpu = cmd_params_defaults.main_gpu;
     }
@@ -1224,6 +1249,8 @@ struct cmd_params_instance {
     llama_split_mode   split_mode;
     llama_load_mode    load_mode;
     uint64_t           longhaul_cache_bytes;
+    bool               tokenizer_longhaul;
+    uint64_t           tokenizer_longhaul_cache_bytes;
     int                main_gpu;
     bool               no_kv_offload;
     llama_flash_attn_type flash_attn;
@@ -1246,6 +1273,8 @@ struct cmd_params_instance {
         mparams.split_mode    = split_mode;
         mparams.load_mode     = load_mode;
         mparams.longhaul_cache_bytes = longhaul_cache_bytes;
+        mparams.tokenizer_longhaul = tokenizer_longhaul;
+        mparams.tokenizer_longhaul_cache_bytes = tokenizer_longhaul_cache_bytes;
         mparams.main_gpu      = main_gpu;
         mparams.tensor_split  = tensor_split.data();
         mparams.no_host       = no_host;
@@ -1294,6 +1323,8 @@ struct cmd_params_instance {
                split_mode == other.split_mode &&
                main_gpu == other.main_gpu && tensor_split == other.tensor_split &&
                load_mode == other.load_mode && longhaul_cache_bytes == other.longhaul_cache_bytes &&
+               tokenizer_longhaul == other.tokenizer_longhaul &&
+               tokenizer_longhaul_cache_bytes == other.tokenizer_longhaul_cache_bytes &&
                devices == other.devices && no_host == other.no_host &&
                vec_tensor_buft_override_equal(tensor_buft_overrides, other.tensor_buft_overrides);
     }
@@ -1368,6 +1399,8 @@ static std::vector<cmd_params_instance> get_cmd_params_instances(const cmd_param
                 /* .split_mode            = */ sm,
                 /* .load_mode             = */ lm,
                 /* .longhaul_cache_bytes  = */ params.longhaul_cache_bytes,
+                /* .tokenizer_longhaul    = */ params.tokenizer_longhaul,
+                /* .tokenizer_longhaul_cache_bytes = */ params.tokenizer_longhaul_cache_bytes,
                 /* .main_gpu              = */ mg,
                 /* .no_kv_offload         = */ nkvo,
                 /* .flash_attn            = */ fa,
@@ -1405,6 +1438,8 @@ static std::vector<cmd_params_instance> get_cmd_params_instances(const cmd_param
                 /* .split_mode            = */ sm,
                 /* .load_mode             = */ lm,
                 /* .longhaul_cache_bytes  = */ params.longhaul_cache_bytes,
+                /* .tokenizer_longhaul    = */ params.tokenizer_longhaul,
+                /* .tokenizer_longhaul_cache_bytes = */ params.tokenizer_longhaul_cache_bytes,
                 /* .main_gpu              = */ mg,
                 /* .no_kv_offload         = */ nkvo,
                 /* .flash_attn            = */ fa,
@@ -1442,6 +1477,8 @@ static std::vector<cmd_params_instance> get_cmd_params_instances(const cmd_param
                 /* .split_mode            = */ sm,
                 /* .load_mode             = */ lm,
                 /* .longhaul_cache_bytes  = */ params.longhaul_cache_bytes,
+                /* .tokenizer_longhaul    = */ params.tokenizer_longhaul,
+                /* .tokenizer_longhaul_cache_bytes = */ params.tokenizer_longhaul_cache_bytes,
                 /* .main_gpu              = */ mg,
                 /* .no_kv_offload         = */ nkvo,
                 /* .flash_attn            = */ fa,
@@ -1484,6 +1521,8 @@ struct test {
     llama_split_mode         split_mode;
     llama_load_mode          load_mode;
     uint64_t                 longhaul_cache_bytes;
+    bool                     tokenizer_longhaul;
+    uint64_t                 tokenizer_longhaul_cache_bytes;
     int                      main_gpu;
     bool                     no_kv_offload;
     llama_flash_attn_type    flash_attn;
@@ -1524,6 +1563,8 @@ struct test {
         split_mode     = inst.split_mode;
         load_mode      = inst.load_mode;
         longhaul_cache_bytes = inst.longhaul_cache_bytes;
+        tokenizer_longhaul = inst.tokenizer_longhaul;
+        tokenizer_longhaul_cache_bytes = inst.tokenizer_longhaul_cache_bytes;
         main_gpu       = inst.main_gpu;
         no_kv_offload  = inst.no_kv_offload;
         flash_attn     = inst.flash_attn;
@@ -1591,7 +1632,8 @@ struct test {
             "n_ubatch",       "n_threads",      "cpu_mask",      "cpu_strict",     "poll",
             "type_k",         "type_v",         "n_gpu_layers",  "n_cpu_moe",      "split_mode",
             "main_gpu",       "no_kv_offload",  "flash_attn",    "devices",        "tensor_split",
-            "tensor_buft_overrides",            "load_mode",     "longhaul_cache_bytes", "embeddings",
+            "tensor_buft_overrides",            "load_mode",     "longhaul_cache_bytes",
+            "tokenizer_longhaul", "tokenizer_longhaul_cache_bytes", "embeddings",
             "no_op_offload",  "no_host",        "fit_target",    "fit_min_ctx",
             "n_prompt",       "n_gen",          "n_depth",
             "test_time",      "avg_ns",         "stddev_ns",     "avg_ts",         "stddev_ts"
@@ -1607,11 +1649,11 @@ struct test {
             field == "main_gpu" || field == "n_prompt" || field == "n_gen" || field == "n_depth" || field == "avg_ns" ||
             field == "stddev_ns" || field == "no_op_offload" || field == "n_cpu_moe" ||
             field == "fit_target" || field == "fit_min_ctx" || field == "flash_attn" ||
-            field == "longhaul_cache_bytes") {
+            field == "longhaul_cache_bytes" || field == "tokenizer_longhaul_cache_bytes") {
             return INT;
         }
         if (field == "f16_kv" || field == "no_kv_offload" || field == "cpu_strict" ||
-            field == "embeddings" || field == "no_host") {
+            field == "embeddings" || field == "no_host" || field == "tokenizer_longhaul") {
             return BOOL;
         }
         if (field == "avg_ts" || field == "stddev_ts") {
@@ -1688,6 +1730,8 @@ struct test {
                                             tensor_buft_overrides_str,
                                             llama_load_mode_name(load_mode),
                                             std::to_string(longhaul_cache_bytes),
+                                            std::to_string(tokenizer_longhaul),
+                                            std::to_string(tokenizer_longhaul_cache_bytes),
                                             std::to_string(embeddings),
                                             std::to_string(no_op_offload),
                                             std::to_string(no_host),
@@ -2005,6 +2049,12 @@ struct markdown_printer : public printer {
         if (params.longhaul_cache_bytes != cmd_params_defaults.longhaul_cache_bytes) {
             fields.emplace_back("longhaul_cache_bytes");
         }
+        if (params.tokenizer_longhaul != cmd_params_defaults.tokenizer_longhaul) {
+            fields.emplace_back("tokenizer_longhaul");
+        }
+        if (params.tokenizer_longhaul_cache_bytes != cmd_params_defaults.tokenizer_longhaul_cache_bytes) {
+            fields.emplace_back("tokenizer_longhaul_cache_bytes");
+        }
         if (params.embeddings.size() > 1 || params.embeddings != cmd_params_defaults.embeddings) {
             fields.emplace_back("embeddings");
         }

tools/server/README.md

@@ -56,6 +56,8 @@ For the full list of features, please refer to [server's changelog](https://gith
 | `--token-cache-size N` | shared tokenizer and tokenized-text cache size in MiB (default: 5120, 0 = disable)<br/>(env: LLAMA_ARG_TOKEN_CACHE_SIZE) |
 | `--token-cache-dir PATH` | path to the shared tokenizer and tokenized-text cache database<br/>(env: LLAMA_ARG_TOKEN_CACHE_DIR) |
 | `--token-cache-clear` | clear the shared tokenizer and tokenized-text cache before loading the model<br/>(env: LLAMA_ARG_TOKEN_CACHE_CLEAR) |
+| `--tokenizer-longhaul` | build and page the Qwen3.5/Laguna tokenizer through a bounded cache<br/>(env: LLAMA_ARG_TOKENIZER_LONGHAUL) |
+| `--tokenizer-longhaul-cache N` | tokenizer-longhaul RAM cache size in MiB<br/>(env: LLAMA_ARG_TOKENIZER_LONGHAUL_CACHE) |
 | `-fa, --flash-attn [on\|off\|auto]` | set Flash Attention use ('on', 'off', or 'auto', default: 'auto')<br/>(env: LLAMA_ARG_FLASH_ATTN) |
 | `--perf, --no-perf` | whether to enable internal libllama performance timings (default: false)<br/>(env: LLAMA_ARG_PERF) |
 | `-e, --escape, --no-escape` | whether to process escapes sequences (\n, \r, \t, \', \", \\) (default: true) |
@@ -78,7 +80,7 @@ For the full list of features, please refer to [server's changelog](https://gith
 | `--mmap, --no-mmap` | DEPRECATED in favor of `--load-mode`: whether to memory-map model. (if mmap disabled, slower load but may reduce pageouts if not using mlock)<br/>(env: LLAMA_ARG_MMAP) |
 | `-dio, --direct-io, -ndio, --no-direct-io` | DEPRECATED in favor of `--load-mode`: use DirectIO if available<br/>(env: LLAMA_ARG_DIO) |
 | `-lm, --load-mode MODE` | model loading mode (default: mmap)<br/>- none: no special loading mode<br/>- mmap: memory-map model (if mmap disabled, slower load but may reduce pageouts if not using mlock)<br/>- mlock: force system to keep model in RAM rather than swapping or compressing<br/>- mmap+mlock: mmap + force system to keep model in RAM rather than swapping or compressing<br/>- dio: use DirectIO if available<br/>- longhaul: stream routed MoE experts through a bounded cache<br/><br/>(env: LLAMA_ARG_LOAD_MODE) |
-| `--longhaul` | stream routed Qwen3.5 MoE, Laguna, or Inkling experts through a bounded CPU or Metal cache<br/>(env: LLAMA_ARG_LONGHAUL) |
+| `--longhaul` | stream routed Qwen3.5 MoE experts through a bounded Metal cache<br/>(env: LLAMA_ARG_LONGHAUL) |
 | `--longhaul-cache N` | longhaul expert cache size in GiB<br/>(env: LLAMA_ARG_LONGHAUL_CACHE) |
 | `--numa TYPE` | attempt optimizations that help on some NUMA systems<br/>- distribute: spread execution evenly over all nodes<br/>- isolate: only spawn threads on CPUs on the node that execution started on<br/>- numactl: use the CPU map provided by numactl<br/>if run without this previously, it is recommended to drop the system page cache before using this<br/>see https://github.com/ggml-org/llama.cpp/issues/1437<br/>(env: LLAMA_ARG_NUMA) |
 | `-dev, --device <dev1,dev2,..>` | comma-separated list of devices to use for offloading (none = don't offload)<br/>use --list-devices to see a list of available devices<br/>(env: LLAMA_ARG_DEVICE) |