Bump llama.cpp b11069 → b11080, drop patch 0010

.github/verify-patches-applied.sh

@@ -7,21 +7,26 @@
 # Asserts that every llama/patches/*.patch really reached the fetched llama.cpp tree.
 #
 # WHY THIS EXISTS. The patch applier (llama/cmake/apply-llama-patches.cmake) is fail-loud on
-# "does not apply", so a *stale* patch cannot ship silently. What it cannot detect is a patch
-# that stops having an effect while still applying, and most patches do not need this check
-# because they have a runnable guard that reds CI on every platform if they go missing:
+# "does not apply", so a *stale* patch cannot ship silently. What it cannot detect is the
+# applier never having run at all, or a patched tree being reverted after the fact — the stamp
+# bookkeeping and the tree's dirty state are the only evidence of that, and this script asserts
+# both. It runs in the always-on `C++ Tests` job, needs no model, and costs milliseconds.
+#
+# Every patch in the set does also have a runnable guard that reds CI on every platform if it
+# goes missing, so these checks are a second line rather than the only one:
 #
 #   0003, 0006, 0007, 0008  -> jllama.cpp / native_server.cpp call the symbols they add,
 #                              so dropping one is a compile or link error.
 #   0012                    -> src/test/cpp/test_model_split.cpp.
+#   0014                    -> src/test/cpp/test_common_log_callback.cpp (link error).
 #   0001, 0002              -> model-gated Java jobs (Windows argv, LoadProgressCallbackTest).
 #
-# `0010` is the exception and the reason for this script. It casts one enum to int inside
-# upstream's `get_res_model_info()`, which is `static` in server-context.cpp and therefore
-# unreachable from jllama_test; reverting it leaves `ctest` completely green. Its only guard is
-# NativeServerAttachIntegrationTest.models_reportNumericVocabType, which is model-gated — so the
-# day a platform stops downloading models, the regression ships. This check runs in the
-# always-on `C++ Tests` job, needs no model, and costs milliseconds.
+# It used to carry a third, patch-specific check for `0010`, the one patch with no runnable
+# guard (it cast an enum inside upstream's `static get_res_model_info()`, unreachable from
+# jllama_test). That patch was DROPPED at the b11080 bump — upstream #28518 gave
+# `common_json_value` an enum constructor, fixing the defect at its root — so the check retired
+# with it. If a future patch is ever added that likewise cannot be reached from `ctest`, add a
+# check for it here rather than relying on a model-gated Java test.
 #
 # Usage: .github/verify-patches-applied.sh [<llama.cpp-src-dir>]
 # Exit codes: 0 all good, 1 a check failed.
@@ -67,14 +72,4 @@ if git -C "$SRC" rev-parse --git-dir >/dev/null 2>&1; then
     fi
 fi
 
-# --- 3. the one patch with no runnable guard ------------------------------------------------------
-VOCAB_CAST='(int) meta.model_vocab_type'
-SERVER_CONTEXT="$SRC/tools/server/server-context.cpp"
-[ -f "$SERVER_CONTEXT" ] || fail "not found: $SERVER_CONTEXT"
-grep -qF "$VOCAB_CAST" "$SERVER_CONTEXT" \
-    || fail "patches/0010 is not present in $SERVER_CONTEXT: expected '$VOCAB_CAST'.
-         Without the cast, common_json binds the unscoped enum to its bool constructor and
-         GET /models + GET /v1/models report vocab_type as true/false instead of a number.
-         If upstream added the cast themselves, DROP patch 0010 and update this check."
-
-echo "patches verified: $on_disk applied, tree dirty, patches/0010 cast present"
+echo "patches verified: $on_disk applied, tree dirty"

.github/workflows/publish.yml

@@ -2356,10 +2356,10 @@ jobs:
         run: |
           mvn -q --no-transfer-progress -f llama/pom.xml compile
           .github/build.sh -DBUILD_TESTING=ON
-      # Most patches have a runnable guard that reds this job if they go missing (a link error, or
-      # test_utils.cpp / test_model_split.cpp). patches/0010 has none — it casts one enum inside a
-      # `static` function unreachable from jllama_test, so reverting it leaves ctest fully green and
-      # only a model-gated Java test notices. This is the always-on, model-free check for it.
+      # Every patch has a runnable guard that reds this job if it goes missing (a link error, or
+      # test_utils.cpp / test_model_split.cpp / test_common_log_callback.cpp). This is the second
+      # line: it asserts the applier actually ran and nothing reverted the patched tree, which no
+      # per-patch guard covers directly. Model-free, milliseconds.
       - name: Verify llama.cpp patches are applied
         run: .github/verify-patches-applied.sh
       - name: Run C++ unit tests

CHANGELOG.md

@@ -101,6 +101,33 @@ from version 5.0.0 onward. Pre-fork releases (`1.x`–`4.2.0`) were authored by
   where the backend cannot provide it, `OFF` disables it.
 
 ### Changed
+- **llama.cpp `b11069` → `b11080`, and local patch `0010` dropped — upstream fixed the
+  enum-to-JSON-boolean trap at its root.** Eleven upstream commits, 1244 KiB, no project-source
+  change. The size is one commit that does not concern this project (llama.cpp #29197 rewrites 46
+  files under `ggml/src/ggml-hexagon/`; no hexagon classifier is built here); the rest of `ggml/src`
+  is additive ARM repack kernels, a Metal fusion-list simplification and a SYCL softmax tweak, and
+  `ggml/include` is byte-identical. The one that matters is llama.cpp #28518 ("json: Fixed json enum
+  handling"): `common_json_value` gains an `std::is_enum`-gated constructor delegating to the
+  underlying type, so an unscoped enum no longer binds to `common_json_value(bool)` and serialises
+  as `true`/`false`. That is exactly the defect `patches/0010` cast around in upstream's own
+  `get_res_model_info()`, so the patch became a redundant carry and was dropped rather than kept
+  (the `0009`/`0011`/`0013` precedent). **Nothing observable changes for consumers** —
+  `GET /models` and `GET /v1/models` reported a numeric `vocab_type` with the patch and still do
+  without it — but the drop is worth flagging because `0010` *still applied cleanly*: the fail-loud
+  applier can only detect "does not apply", never "upstream already fixed this", which is why that
+  patch carries a by-hand drop-check on every bump. Its guard was kept and re-pointed: the
+  `CommonJsonEnumTrap` pair in `test_json_helpers.cpp` is now the `CommonJsonEnum` trio and pins
+  upstream's contract (uncast enum is numeric, an explicit cast is equivalent, a real `bool` is
+  still a boolean), so a bump that loses the constructor reds `C++ Tests` everywhere instead of
+  shipping a boolean. `jllama.cpp` keeps its own two `"vocab_type"` casts — correct either way.
+  Also in range: six existing sampling flags gained environment defaults (#27380 —
+  `LLAMA_ARG_TEMPERATURE`, `_TOP_P`, `_MIN_P`, `_REPEAT_PENALTY`, `_PRESENCE_PENALTY`,
+  `_FREQUENCY_PENALTY`), which adds no option but does mean a host with those variables set now
+  inherits them; and a router no longer forwards `LLAMA_ARG_API_KEY_FILE` to spawned children
+  (#28938). `tools/server/`'s schema, task and context translation units are byte-identical, and
+  the request-field set (68), bounded-field set (23) and response-key set (142) were all verified
+  unchanged mechanically, so the server wire contract cannot have moved. The other eight patches
+  apply unchanged and all four remaining drop-checks still say "still required".
 - **llama.cpp `b11062` → `b11069`, and local patch `0011` dropped — malformed UTF-8 in a
   completion is now replaced, not truncated.** Seven upstream commits, 57 KiB, no project-source
   change; the one that matters is llama.cpp #29161 ("common/peg : handle invalid utf-8 sequences in