chore: QVAC-21914 address PR review — flush-cadence wording, nits

Non-behavioral cleanups from the PR tetherto#181 review pass:

- ggml-opencl graph_compute: correct the flush-cadence comment. The old
  "per-token decode hot path submission-free by construction" claim was
  false for multi-GB models — a decode step streams the whole model, so
  its estimated work crosses the default 512 MB budget a few times per
  token. Reworded to state that accurately (cost is negligible in
  practice since clFlush is non-blocking, and it is tunable/zeroable to
  make decode fully submission-free). Mechanism unchanged.
- ggml_cl_flash_attn: GGML_ASSERT(q->ne[1] <= INT32_MAX) before the int
  n_q truncation, since the q-chunk loop accumulates into an int and
  derives cl_ulong offsets from it (defensive; not reachable with real
  shapes).
- Consistency: normalize the ticket tag to bare `QVAC-21914` (drop the
  `qvac ` prefix) in clip.h and tests/CMakeLists.txt, matching the .cpp
  files and the fork's QVAC-21257 precedent.
- tests/CMakeLists.txt: move the unconditional test-opencl-fa-chunking
  registration up beside test-copy-tbq-subgroups (its self-skipping
  sibling) instead of sitting right after the LLAMA_MTMD endif() where it
  read as MTMD-gated; add a comment noting it deliberately does not link
  mtmd.

No functional change to the fix; local mtmd + both tests build,
test-clip-fa-cutoff passes.

Author: iancris

ggml/src/ggml-opencl/ggml-opencl.cpp

@@ -5798,10 +5798,14 @@ static ggml_status ggml_backend_opencl_graph_compute(ggml_backend_t backend, ggm
     // from cl_a8x_cmdbuf_mgr_submit_ibs (os_exit) — observed on Galaxy S25
     // Ultra / Adreno 830. Flushing whenever the estimated enqueued work
     // exceeds flush_work_budget hands the driver bounded batches instead.
-    // Gating on WORK (not a node count) keeps the per-token LLM decode hot
-    // path submission-free by construction: a decode step never accumulates
-    // anywhere near the budget, while the giant encode flushes dozens of
-    // times. clFlush only submits (no host stall).
+    // Gating on WORK (not a node count) scales the flush cadence to the batch
+    // size: the ~48 s encode flushes many times, while a per-token LLM decode
+    // (its work ~= the model size streamed once, i.e. a few flushes per token
+    // at the default budget for a multi-GB model) is far lighter. clFlush only
+    // submits (no host stall), so the decode-path cost is negligible in
+    // practice (measured GPU decode TPS within noise of pre-hardening), but it
+    // is NOT literally zero for large models — raise flush_work_budget past the
+    // model size, or set it to 0, to make decode fully submission-free.
     int64_t work_since_flush = 0;
 
     for (int i = 0; i < cgraph->n_nodes; i++) {
@@ -12945,6 +12949,11 @@ static void ggml_cl_flash_attn(ggml_backend_t backend, const ggml_tensor * q, co
 
     ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
 
+    // QVAC-21914: n_q is int; the q-chunk loop below accumulates into an int
+    // (q_base += chunk_rows) and derives cl_ulong byte offsets from it. Guard
+    // the int64->int truncation so a pathological q-row count can't wrap
+    // negative and produce an out-of-bounds device offset.
+    GGML_ASSERT(q->ne[1] <= INT32_MAX);
     const int n_q = q->ne[1];
     const int n_kv = k->ne[1];
     const int d_head_q = q->ne[0];

tests/CMakeLists.txt

@@ -254,9 +254,17 @@ llama_build_and_test(test-backend-ops.cpp)
 # GPU backend is available.
 llama_build_and_test(test-copy-tbq-subgroups.cpp)
 
-llama_build_and_test(test-model-load-cancel.cpp LABEL "model")
-llama_build_and_test(test-model-load-disk.cpp   LABEL "model")
-llama_build_and_test(test-model-load-memory.cpp LABEL "model")
+# QVAC-21914: ggml-opencl flash-attention q-chunking parity vs the CPU
+# backend. Unconditional and self-skipping at runtime when no OpenCL device is
+# present (like test-copy-tbq-subgroups above), so it exercises chunking
+# wherever OpenCL exists (Adreno devices, desktop OpenCL, PoCL). Does not link
+# mtmd, so it is intentionally outside the LLAMA_MTMD block below.
+llama_build_and_test(test-opencl-fa-chunking.cpp)
+
+llama_build_and_test(test-model-load-cancel.cpp       LABEL "model")
+llama_build_and_test(test-model-load-disk.cpp         LABEL "model")
+llama_build_and_test(test-model-load-memory.cpp       LABEL "model")
+target_include_directories(test-model-load-memory PRIVATE ${PROJECT_SOURCE_DIR}/common_test)
 llama_build_and_test(test-model-load-memory-split.cpp LABEL "model")
 llama_build_and_test(test-autorelease.cpp       LABEL "model")
 llama_build_and_test(test-backend-sampler.cpp   LABEL "model")
@@ -294,20 +302,14 @@ if (LLAMA_MTMD)
     target_link_libraries(${LLAMA_TEST_NAME} PRIVATE mtmd)
     unset(LLAMA_TEST_NAME)
 
-    # qvac QVAC-21914: clip flash-attention AUTO budget arithmetic (pure CPU).
+    # QVAC-21914: clip flash-attention AUTO budget arithmetic (pure CPU).
     set(LLAMA_TEST_NAME test-clip-fa-cutoff)
     llama_build_and_test(test-clip-fa-cutoff.cpp)
     target_link_libraries(${LLAMA_TEST_NAME} PRIVATE mtmd)
     target_include_directories(${LLAMA_TEST_NAME} PRIVATE ${PROJECT_SOURCE_DIR}/tools/mtmd)
     unset(LLAMA_TEST_NAME)
 endif()
 
-# qvac QVAC-21914: ggml-opencl flash-attention q-chunking parity vs the CPU
-# backend. Self-skipping at runtime when no OpenCL device is present, so it is
-# registered unconditionally and exercises chunking wherever OpenCL exists
-# (Adreno devices, desktop OpenCL, PoCL).
-llama_build_and_test(test-opencl-fa-chunking.cpp)
-
 # GGUF model data fetcher library for tests that need real model metadata
 # Only compile when cpp-httplib has SSL support (CPPHTTPLIB_OPENSSL_SUPPORT)
 if (TARGET cpp-httplib)

tools/mtmd/clip.h

@@ -122,7 +122,7 @@ int clip_model_n_temporal_merge(const struct clip_ctx * ctx); // TODO @ngxson :
 // (and ultimately by common/fit.cpp's heuristic). Restored from upstream b9341.
 std::map<ggml_backend_dev_t, size_t> clip_get_mem_usage(const struct clip_ctx * ctx);
 
-// qvac QVAC-21914: pure arithmetic of the flash-attention AUTO budget decision.
+// QVAC-21914: pure arithmetic of the flash-attention AUTO budget decision.
 // Returns the effective explicit-attention cutoff in n_patches given the
 // configured cutoff and the device memory probe:
 //