Bump llama.cpp to b10178 and honour :use_mlock / :use_mmap independently

CHANGELOG.md

@@ -88,6 +88,16 @@ facade's streams emit errors instead of truncating silently.
   server cannot honour — and then raised an `ArgumentError` naming
   `LlamaCppEx.Server.complete_tokens/3`, a function the caller never called. They
   are now rejected up front, naming the function the caller did call.
+- **`use_mlock: true, use_mmap: false` no longer memory-maps the model.** Since
+  v0.8.39 the three load booleans have collapsed into llama.cpp's single
+  `load_mode` enum, where `LLAMA_LOAD_MODE_MLOCK` meant "mmap **and** mlock" — so
+  `:use_mlock` silently forced a memory map and `:use_mmap` was unreachable
+  whenever it was set. Upstream b10173 split that into `MLOCK` (mlock without
+  mmap) and `MMAP_MLOCK` (both), and `Model.load/2` now honours the two options
+  independently: both true selects `mmap_mlock`, `:use_mlock` alone selects
+  `mlock` and reads the weights into anonymous memory. `:use_direct_io` still
+  takes precedence over both, and the default path is unaffected because
+  `:use_mmap` defaults to `true`.
 
 ### Security
 
@@ -359,6 +369,48 @@ facade's streams emit errors instead of truncating silently.
   actually hits (`:dialyzer` `plt_local_path`/`plt_core_path` point at the
   `priv/plts` that CI caches; dialyxir was writing under `_build`).
 
+### Changed
+
+- **llama.cpp submodule** — Updated from ff067f76d to 992c32532 (45 commits, tag
+  b10178). Three binding-relevant headers changed, one of them behaviour-changing:
+  - `include/llama.h` — **behaviour change**: `enum llama_load_mode` gains
+    `LLAMA_LOAD_MODE_MMAP_MLOCK` and renumbers `_DIRECT_IO` from 3 to 4.
+    `LLAMA_LOAD_MODE_MLOCK` previously meant "mmap **and** mlock" and now means
+    mlock *without* mmap, with the new `_MMAP_MLOCK` covering the combination
+    (#26135). The NIF selects the mode by name so the renumbering is transparent,
+    but the redefinition is not: `model_load/10` now maps `:use_mlock` +
+    `:use_mmap` to `_MMAP_MLOCK` and `:use_mlock` alone to `_MLOCK`, which is what
+    the Breaking entry above describes. Without that change `use_mlock: true`
+    would have silently stopped memory-mapping.
+  - `common/common.h` — the free functions `common_context_seq_rm`,
+    `common_context_seq_add` and `common_context_seq_cp` became `static` and were
+    replaced by a `common_memory` struct carrying target/draft contexts and
+    exposing them as methods (#26221). The binding never called them — it drives
+    `llama_memory_*` directly and only reads `common_context_can_seq_rm`, whose
+    signature and `common_context_seq_rm_type` enum are both unchanged — so the
+    refactor is inert here. A new `COMMON_SPECULATIVE_TYPE_DRAFT_DSPARK` enum
+    value and its `need_n_rs_seq()` branch (#25173) are likewise inert: the
+    binding sets `types` explicitly for its MTP path.
+  - `common/chat.h` — additive only: a `COMMON_CHAT_FORMAT_PEG_MINIMAX_M3` format
+    for the new MiniMax-M3 parser (#26210).
+  - `ggml/include/ggml.h`, `ggml/include/ggml-backend.h`,
+    `common/json-schema-to-grammar.h`, `common/sampling.h` and
+    `common/speculative.h` are untouched in this range.
+  - **llama core / models**: MiniMax-M3 (MiniMax Sparse Attention) support
+    (#24908) and its indexer tensors kept at F32 for speed and accuracy (#26144);
+    Nanbeige4.2 (#25994); a Laguna-S-2.1 `LLM_TYPE` (#26233); NextN/MTP
+    speculative decoding for GLM_DSA / GLM-5.2 (#25980).
+  - **common / speculative**: DSpark speculative decoding (#25173); eagle3-v3
+    support for gpt-oss (#25794); `common_print_available_devices()` (#26170);
+    explicit `-md` precedence over draft sidecar resolution (#26165); a
+    `subproc.h` wrapper, disabled on android/ios (#26102); NextN (MTP) blocks now
+    counted in `n_gpu_layers` so front layers stay on GPU (#26177).
+  - **ggml / Metal**: an FWHT kernel for the Metal backend (#25924); view-src
+    output is now set (#25729); op offloading logic adjusted to prefer the
+    weight's backend (#25832).
+  - **mtmd**: GLM-5.2-Vision (#26126); MiniMax-M3 vision (#25113); Nemotron 3
+    Nano Omni / parakeet (#22520); MiMo-V2.5 RVQ audio input (#26190).
+
 ### Tests
 
 - **67 of 268 test cases never compiled.** Five compile-time

Makefile

@@ -22,7 +22,7 @@ endif
 # Pinned llama.cpp commit, used when vendor/llama.cpp has to be cloned. MUST
 # match the vendor/llama.cpp submodule; bump both together, see
 # docs/release-guide.md. Override to build the NIF against another revision.
-LLAMA_COMMIT ?= ff067f76dd8e9e05f0528056f1274adf01a54d70
+LLAMA_COMMIT ?= 992c325323f925cb82c86778e5e91a63de199063
 
 # The commit actually on disk. A submodule can be bumped without LLAMA_COMMIT
 # following it, and the build has to key off what is really there.

c_src/llama_cpp_ex/llama_nif.cpp

@@ -331,12 +331,16 @@ model_load(ErlNifEnv* env, std::string path, int64_t n_gpu_layers, bool use_mmap
     params.main_gpu = static_cast<int32_t>(main_gpu);
     params.split_mode = static_cast<enum llama_split_mode>(split_mode);
     // Upstream collapsed the use_mmap/use_mlock/use_direct_io booleans into a
-    // single llama_load_mode enum. Preserve the documented precedence of the
-    // Elixir options: direct I/O wins over mmap, and mlock implies mmap.
-    params.load_mode = use_direct_io ? LLAMA_LOAD_MODE_DIRECT_IO
-                     : use_mlock     ? LLAMA_LOAD_MODE_MLOCK
-                     : use_mmap      ? LLAMA_LOAD_MODE_MMAP
-                                     : LLAMA_LOAD_MODE_NONE;
+    // single llama_load_mode enum. Direct I/O still wins over everything else,
+    // but mlock no longer implies mmap: b10173 redefined LLAMA_LOAD_MODE_MLOCK
+    // as mlock *without* mmap and added LLAMA_LOAD_MODE_MMAP_MLOCK for the
+    // combination, so :use_mlock and :use_mmap are both honoured rather than
+    // the former silently forcing a memory map.
+    params.load_mode = use_direct_io         ? LLAMA_LOAD_MODE_DIRECT_IO
+                     : use_mlock && use_mmap ? LLAMA_LOAD_MODE_MMAP_MLOCK
+                     : use_mlock             ? LLAMA_LOAD_MODE_MLOCK
+                     : use_mmap              ? LLAMA_LOAD_MODE_MMAP
+                                             : LLAMA_LOAD_MODE_NONE;
     params.vocab_only = vocab_only;
     params.check_tensors = check_tensors;
 

lib/llama_cpp_ex/model.ex

@@ -62,11 +62,11 @@ defmodule LlamaCppEx.Model do
   > #### Load mode {: .info}
   >
   > llama.cpp collapsed its three loading booleans into one `load_mode` enum, so
-  > these options resolve to a single mode with `:use_direct_io` > `:use_mlock` >
-  > `:use_mmap` precedence — respectively `dio`, `mlock`, `mmap`, and `none` when
-  > all are false. Because `mlock` now implies mmap upstream, passing
-  > `use_mlock: true, use_mmap: false` memory-maps the file rather than reading
-  > it into anonymous memory.
+  > these options resolve to a single mode. `:use_direct_io` takes precedence
+  > over everything and selects `dio`; otherwise `:use_mlock` and `:use_mmap`
+  > combine — both true selects `mmap_mlock`, `:use_mlock` alone selects `mlock`
+  > (read into anonymous memory, no mapping), `:use_mmap` alone selects `mmap`,
+  > and all false selects `none`.
 
   > #### Untrusted models {: .warning}
   >

lib/llama_cpp_ex/server.ex

@@ -313,8 +313,8 @@ defmodule LlamaCppEx.Server do
     * Model loading options are forwarded to `LlamaCppEx.Model.load/2` —
       `:main_gpu`, `:split_mode`, `:tensor_split`, `:use_mmap`, `:use_mlock`,
       `:use_direct_io` and `:check_tensors`. The three load flags collapse into
-      llama.cpp's single `load_mode` with `:use_direct_io` > `:use_mlock` >
-      `:use_mmap` precedence; see `LlamaCppEx.Model.load/2`.
+      llama.cpp's single `load_mode`: `:use_direct_io` wins outright, otherwise
+      `:use_mlock` and `:use_mmap` combine; see `LlamaCppEx.Model.load/2`.
     * GenServer options like `:name`.
 
   """