diff --git a/llamafile/cuda.sh b/llamafile/cuda.sh
--- a/llamafile/cuda.sh
+++ b/llamafile/cuda.sh
@@ -259,6 +259,22 @@
     COMMON_FLAGS="$COMMON_FLAGS -DGGML_CUDA_FA_ALL_QUANTS"
 fi
 
+# opencoti-hook: cuda-build-parallelism (#531) — nvcc intra-TU arch threading. Each nvcc TU compiles
+# the gencode arches (sm_75/80/86/89/90/120f); without --threads they run serially, so any TU left
+# alone at the build tail uses one core while the rest idle. --threads compiles the arches
+# concurrently. To avoid oversubscribing the bulk (JOBS nvcc already saturate the cores), we divide
+# the outer job-slot count by NVCC_THREADS so PEAK concurrency stays ≈ JOBS (slots × threads). The
+# Windows cuda.bat already passes --threads 5; the Linux path lacked it. Tunable: NVCC_THREADS=1
+# disables (restores pre-#531 behavior). See docs/protocols/LLAMAFILE_UPGRADE.md.
+NVCC_THREADS="${NVCC_THREADS:-2}"
+JOBS_EFF="$JOBS"
+if [ "${NVCC_THREADS}" -gt 1 ] 2>/dev/null; then
+    COMMON_FLAGS="$COMMON_FLAGS --threads ${NVCC_THREADS}"
+    JOBS_EFF=$(( JOBS / NVCC_THREADS ))
+    [ "$JOBS_EFF" -lt 1 ] && JOBS_EFF=1
+fi
+echo "  nvcc --threads ${NVCC_THREADS}  →  ${JOBS_EFF} job-slots × ${NVCC_THREADS} arch-threads (#531)"
+
 # Collect sources
 collect_gpu_sources "$GGML_CUDA_DIR" "$EXTRA_SOURCES" "$NO_IQ_QUANTS" "$FA_ALL_QUANTS"
 echo "  Sources: $NUM_SOURCES .cu files"
@@ -267,7 +283,7 @@
 START_TIME=$(date +%s)
 
 # Compile GPU sources
-compile_gpu_sources_parallel "$NVCC" "$ARCH_FLAGS" "$COMMON_FLAGS" "$BUILD_DIR" "$JOBS"
+compile_gpu_sources_parallel "$NVCC" "$ARCH_FLAGS" "$COMMON_FLAGS" "$BUILD_DIR" "$JOBS_EFF"
 
 COMPILE_TIME=$(date +%s)
 echo "Compilation took $((COMPILE_TIME - START_TIME)) seconds"
diff --git a/llama.cpp/ggml/src/ggml-cuda/fattn-mma-f16.cuh b/llama.cpp/ggml/src/ggml-cuda/fattn-mma-f16.cuh
--- a/llama.cpp/ggml/src/ggml-cuda/fattn-mma-f16.cuh
+++ b/llama.cpp/ggml/src/ggml-cuda/fattn-mma-f16.cuh
@@ -2600,3 +2600,41 @@
 extern DECL_FATTN_MMA_F16_CASE(576, 512, 16,  4);
 extern DECL_FATTN_MMA_F16_CASE(576, 512,  1, 32);
 extern DECL_FATTN_MMA_F16_CASE(576, 512,  2, 32);
+
+// opencoti-hook: cuda-build-parallelism (#531) — shard the DCA-fused MMA case into parallel
+// instance TUs (template-instances/fattn-mma-f16-dca-fused-instance-*.cu), mirroring the stock
+// DECL_FATTN_MMA_F16_CASE sharding above. Before this, fattn.cu's dca-fused dispatch
+// (ggml_cuda_flash_attn_ext_mma_f16_dca_fused_switch) instantiated all 48
+// ggml_cuda_flash_attn_ext_mma_f16_dca_fused_case<DKQ,DV,ncols1,ncols2> INLINE — DKQ==DV ∈ {128,256,512}
+// × ncols2 ∈ {1,2,4,8} × ncols1 ∈ {8,16,32,64}/ncols2 → 576 device kernels in ONE TU = a ~52-min
+// SOLO compile tail (fattn.o was 29 MB, the largest object, finishing ~52 min after every other TU).
+// These extern decls move the instantiation to the parallel bulk; fattn.cu now just references the
+// symbols (the regular MMA/TILE/VEC cases were already sharded this way; only our DCA kernel wasn't).
+// The instance set matches the dispatch 1:1 — every reachable _case<DKQ,DV,ncols1,ncols2> has a
+// definition, none extra — so the linked DSO is byte-for-byte equivalent (same 576 kernels, relocated).
+#define DECL_FATTN_MMA_F16_DCA_FUSED_CASE(DKQ, DV, ncols1, ncols2)                 \
+    template void ggml_cuda_flash_attn_ext_mma_f16_dca_fused_case                  \
+    <DKQ, DV, ncols1, ncols2>(ggml_backend_cuda_context & ctx, ggml_tensor * dst)  \
+
+// DCA always uses DKQ==DV ∈ {128,256,512}; one extern per (ncols1,ncols2) across all three head dims.
+#define EXTERN_DECL_FATTN_MMA_F16_DCA_FUSED_ALL_DKQ(ncols1, ncols2)      \
+    extern DECL_FATTN_MMA_F16_DCA_FUSED_CASE(128, 128, ncols1, ncols2);  \
+    extern DECL_FATTN_MMA_F16_DCA_FUSED_CASE(256, 256, ncols1, ncols2);  \
+    extern DECL_FATTN_MMA_F16_DCA_FUSED_CASE(512, 512, ncols1, ncols2);  \
+
+EXTERN_DECL_FATTN_MMA_F16_DCA_FUSED_ALL_DKQ( 8, 1)
+EXTERN_DECL_FATTN_MMA_F16_DCA_FUSED_ALL_DKQ(16, 1)
+EXTERN_DECL_FATTN_MMA_F16_DCA_FUSED_ALL_DKQ(32, 1)
+EXTERN_DECL_FATTN_MMA_F16_DCA_FUSED_ALL_DKQ(64, 1)
+EXTERN_DECL_FATTN_MMA_F16_DCA_FUSED_ALL_DKQ( 4, 2)
+EXTERN_DECL_FATTN_MMA_F16_DCA_FUSED_ALL_DKQ( 8, 2)
+EXTERN_DECL_FATTN_MMA_F16_DCA_FUSED_ALL_DKQ(16, 2)
+EXTERN_DECL_FATTN_MMA_F16_DCA_FUSED_ALL_DKQ(32, 2)
+EXTERN_DECL_FATTN_MMA_F16_DCA_FUSED_ALL_DKQ( 2, 4)
+EXTERN_DECL_FATTN_MMA_F16_DCA_FUSED_ALL_DKQ( 4, 4)
+EXTERN_DECL_FATTN_MMA_F16_DCA_FUSED_ALL_DKQ( 8, 4)
+EXTERN_DECL_FATTN_MMA_F16_DCA_FUSED_ALL_DKQ(16, 4)
+EXTERN_DECL_FATTN_MMA_F16_DCA_FUSED_ALL_DKQ( 1, 8)
+EXTERN_DECL_FATTN_MMA_F16_DCA_FUSED_ALL_DKQ( 2, 8)
+EXTERN_DECL_FATTN_MMA_F16_DCA_FUSED_ALL_DKQ( 4, 8)
+EXTERN_DECL_FATTN_MMA_F16_DCA_FUSED_ALL_DKQ( 8, 8)
diff --git a/llama.cpp/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-dca-fused-instance-ncols1_16-ncols2_1.cu b/llama.cpp/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-dca-fused-instance-ncols1_16-ncols2_1.cu
new file mode 100644
--- /dev/null
+++ b/llama.cpp/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-dca-fused-instance-ncols1_16-ncols2_1.cu
@@ -0,0 +1,10 @@
+// opencoti-hook: cuda-build-parallelism (#531) — DCA-fused MMA case instance, relocated out of
+// fattn.cu so it compiles in the parallel bulk instead of inflating fattn.cu's solo tail.
+// Mirrors template-instances/fattn-mma-f16-instance-*.cu. extern decls live in ../fattn-mma-f16.cuh.
+// DKQ==DV ∈ {128,256,512} (the only head dims DCA dispatches: Gemma-4 512, qwen35moe 256, qwen* 128).
+
+#include "../fattn-mma-f16.cuh"
+
+DECL_FATTN_MMA_F16_DCA_FUSED_CASE(128, 128, 16, 1);
+DECL_FATTN_MMA_F16_DCA_FUSED_CASE(256, 256, 16, 1);
+DECL_FATTN_MMA_F16_DCA_FUSED_CASE(512, 512, 16, 1);
diff --git a/llama.cpp/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-dca-fused-instance-ncols1_16-ncols2_2.cu b/llama.cpp/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-dca-fused-instance-ncols1_16-ncols2_2.cu
new file mode 100644
--- /dev/null
+++ b/llama.cpp/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-dca-fused-instance-ncols1_16-ncols2_2.cu
@@ -0,0 +1,10 @@
+// opencoti-hook: cuda-build-parallelism (#531) — DCA-fused MMA case instance, relocated out of
+// fattn.cu so it compiles in the parallel bulk instead of inflating fattn.cu's solo tail.
+// Mirrors template-instances/fattn-mma-f16-instance-*.cu. extern decls live in ../fattn-mma-f16.cuh.
+// DKQ==DV ∈ {128,256,512} (the only head dims DCA dispatches: Gemma-4 512, qwen35moe 256, qwen* 128).
+
+#include "../fattn-mma-f16.cuh"
+
+DECL_FATTN_MMA_F16_DCA_FUSED_CASE(128, 128, 16, 2);
+DECL_FATTN_MMA_F16_DCA_FUSED_CASE(256, 256, 16, 2);
+DECL_FATTN_MMA_F16_DCA_FUSED_CASE(512, 512, 16, 2);
diff --git a/llama.cpp/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-dca-fused-instance-ncols1_16-ncols2_4.cu b/llama.cpp/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-dca-fused-instance-ncols1_16-ncols2_4.cu
new file mode 100644
--- /dev/null
+++ b/llama.cpp/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-dca-fused-instance-ncols1_16-ncols2_4.cu
@@ -0,0 +1,10 @@
+// opencoti-hook: cuda-build-parallelism (#531) — DCA-fused MMA case instance, relocated out of
+// fattn.cu so it compiles in the parallel bulk instead of inflating fattn.cu's solo tail.
+// Mirrors template-instances/fattn-mma-f16-instance-*.cu. extern decls live in ../fattn-mma-f16.cuh.
+// DKQ==DV ∈ {128,256,512} (the only head dims DCA dispatches: Gemma-4 512, qwen35moe 256, qwen* 128).
+
+#include "../fattn-mma-f16.cuh"
+
+DECL_FATTN_MMA_F16_DCA_FUSED_CASE(128, 128, 16, 4);
+DECL_FATTN_MMA_F16_DCA_FUSED_CASE(256, 256, 16, 4);
+DECL_FATTN_MMA_F16_DCA_FUSED_CASE(512, 512, 16, 4);
diff --git a/llama.cpp/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-dca-fused-instance-ncols1_1-ncols2_8.cu b/llama.cpp/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-dca-fused-instance-ncols1_1-ncols2_8.cu
new file mode 100644
--- /dev/null
+++ b/llama.cpp/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-dca-fused-instance-ncols1_1-ncols2_8.cu
@@ -0,0 +1,10 @@
+// opencoti-hook: cuda-build-parallelism (#531) — DCA-fused MMA case instance, relocated out of
+// fattn.cu so it compiles in the parallel bulk instead of inflating fattn.cu's solo tail.
+// Mirrors template-instances/fattn-mma-f16-instance-*.cu. extern decls live in ../fattn-mma-f16.cuh.
+// DKQ==DV ∈ {128,256,512} (the only head dims DCA dispatches: Gemma-4 512, qwen35moe 256, qwen* 128).
+
+#include "../fattn-mma-f16.cuh"
+
+DECL_FATTN_MMA_F16_DCA_FUSED_CASE(128, 128, 1, 8);
+DECL_FATTN_MMA_F16_DCA_FUSED_CASE(256, 256, 1, 8);
+DECL_FATTN_MMA_F16_DCA_FUSED_CASE(512, 512, 1, 8);
diff --git a/llama.cpp/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-dca-fused-instance-ncols1_2-ncols2_4.cu b/llama.cpp/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-dca-fused-instance-ncols1_2-ncols2_4.cu
new file mode 100644
--- /dev/null
+++ b/llama.cpp/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-dca-fused-instance-ncols1_2-ncols2_4.cu
@@ -0,0 +1,10 @@
+// opencoti-hook: cuda-build-parallelism (#531) — DCA-fused MMA case instance, relocated out of
+// fattn.cu so it compiles in the parallel bulk instead of inflating fattn.cu's solo tail.
+// Mirrors template-instances/fattn-mma-f16-instance-*.cu. extern decls live in ../fattn-mma-f16.cuh.
+// DKQ==DV ∈ {128,256,512} (the only head dims DCA dispatches: Gemma-4 512, qwen35moe 256, qwen* 128).
+
+#include "../fattn-mma-f16.cuh"
+
+DECL_FATTN_MMA_F16_DCA_FUSED_CASE(128, 128, 2, 4);
+DECL_FATTN_MMA_F16_DCA_FUSED_CASE(256, 256, 2, 4);
+DECL_FATTN_MMA_F16_DCA_FUSED_CASE(512, 512, 2, 4);
diff --git a/llama.cpp/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-dca-fused-instance-ncols1_2-ncols2_8.cu b/llama.cpp/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-dca-fused-instance-ncols1_2-ncols2_8.cu
new file mode 100644
--- /dev/null
+++ b/llama.cpp/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-dca-fused-instance-ncols1_2-ncols2_8.cu
@@ -0,0 +1,10 @@
+// opencoti-hook: cuda-build-parallelism (#531) — DCA-fused MMA case instance, relocated out of
+// fattn.cu so it compiles in the parallel bulk instead of inflating fattn.cu's solo tail.
+// Mirrors template-instances/fattn-mma-f16-instance-*.cu. extern decls live in ../fattn-mma-f16.cuh.
+// DKQ==DV ∈ {128,256,512} (the only head dims DCA dispatches: Gemma-4 512, qwen35moe 256, qwen* 128).
+
+#include "../fattn-mma-f16.cuh"
+
+DECL_FATTN_MMA_F16_DCA_FUSED_CASE(128, 128, 2, 8);
+DECL_FATTN_MMA_F16_DCA_FUSED_CASE(256, 256, 2, 8);
+DECL_FATTN_MMA_F16_DCA_FUSED_CASE(512, 512, 2, 8);
diff --git a/llama.cpp/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-dca-fused-instance-ncols1_32-ncols2_1.cu b/llama.cpp/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-dca-fused-instance-ncols1_32-ncols2_1.cu
new file mode 100644
--- /dev/null
+++ b/llama.cpp/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-dca-fused-instance-ncols1_32-ncols2_1.cu
@@ -0,0 +1,10 @@
+// opencoti-hook: cuda-build-parallelism (#531) — DCA-fused MMA case instance, relocated out of
+// fattn.cu so it compiles in the parallel bulk instead of inflating fattn.cu's solo tail.
+// Mirrors template-instances/fattn-mma-f16-instance-*.cu. extern decls live in ../fattn-mma-f16.cuh.
+// DKQ==DV ∈ {128,256,512} (the only head dims DCA dispatches: Gemma-4 512, qwen35moe 256, qwen* 128).
+
+#include "../fattn-mma-f16.cuh"
+
+DECL_FATTN_MMA_F16_DCA_FUSED_CASE(128, 128, 32, 1);
+DECL_FATTN_MMA_F16_DCA_FUSED_CASE(256, 256, 32, 1);
+DECL_FATTN_MMA_F16_DCA_FUSED_CASE(512, 512, 32, 1);
diff --git a/llama.cpp/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-dca-fused-instance-ncols1_32-ncols2_2.cu b/llama.cpp/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-dca-fused-instance-ncols1_32-ncols2_2.cu
new file mode 100644
--- /dev/null
+++ b/llama.cpp/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-dca-fused-instance-ncols1_32-ncols2_2.cu
@@ -0,0 +1,10 @@
+// opencoti-hook: cuda-build-parallelism (#531) — DCA-fused MMA case instance, relocated out of
+// fattn.cu so it compiles in the parallel bulk instead of inflating fattn.cu's solo tail.
+// Mirrors template-instances/fattn-mma-f16-instance-*.cu. extern decls live in ../fattn-mma-f16.cuh.
+// DKQ==DV ∈ {128,256,512} (the only head dims DCA dispatches: Gemma-4 512, qwen35moe 256, qwen* 128).
+
+#include "../fattn-mma-f16.cuh"
+
+DECL_FATTN_MMA_F16_DCA_FUSED_CASE(128, 128, 32, 2);
+DECL_FATTN_MMA_F16_DCA_FUSED_CASE(256, 256, 32, 2);
+DECL_FATTN_MMA_F16_DCA_FUSED_CASE(512, 512, 32, 2);
diff --git a/llama.cpp/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-dca-fused-instance-ncols1_4-ncols2_2.cu b/llama.cpp/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-dca-fused-instance-ncols1_4-ncols2_2.cu
new file mode 100644
--- /dev/null
+++ b/llama.cpp/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-dca-fused-instance-ncols1_4-ncols2_2.cu
@@ -0,0 +1,10 @@
+// opencoti-hook: cuda-build-parallelism (#531) — DCA-fused MMA case instance, relocated out of
+// fattn.cu so it compiles in the parallel bulk instead of inflating fattn.cu's solo tail.
+// Mirrors template-instances/fattn-mma-f16-instance-*.cu. extern decls live in ../fattn-mma-f16.cuh.
+// DKQ==DV ∈ {128,256,512} (the only head dims DCA dispatches: Gemma-4 512, qwen35moe 256, qwen* 128).
+
+#include "../fattn-mma-f16.cuh"
+
+DECL_FATTN_MMA_F16_DCA_FUSED_CASE(128, 128, 4, 2);
+DECL_FATTN_MMA_F16_DCA_FUSED_CASE(256, 256, 4, 2);
+DECL_FATTN_MMA_F16_DCA_FUSED_CASE(512, 512, 4, 2);
diff --git a/llama.cpp/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-dca-fused-instance-ncols1_4-ncols2_4.cu b/llama.cpp/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-dca-fused-instance-ncols1_4-ncols2_4.cu
new file mode 100644
--- /dev/null
+++ b/llama.cpp/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-dca-fused-instance-ncols1_4-ncols2_4.cu
@@ -0,0 +1,10 @@
+// opencoti-hook: cuda-build-parallelism (#531) — DCA-fused MMA case instance, relocated out of
+// fattn.cu so it compiles in the parallel bulk instead of inflating fattn.cu's solo tail.
+// Mirrors template-instances/fattn-mma-f16-instance-*.cu. extern decls live in ../fattn-mma-f16.cuh.
+// DKQ==DV ∈ {128,256,512} (the only head dims DCA dispatches: Gemma-4 512, qwen35moe 256, qwen* 128).
+
+#include "../fattn-mma-f16.cuh"
+
+DECL_FATTN_MMA_F16_DCA_FUSED_CASE(128, 128, 4, 4);
+DECL_FATTN_MMA_F16_DCA_FUSED_CASE(256, 256, 4, 4);
+DECL_FATTN_MMA_F16_DCA_FUSED_CASE(512, 512, 4, 4);
diff --git a/llama.cpp/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-dca-fused-instance-ncols1_4-ncols2_8.cu b/llama.cpp/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-dca-fused-instance-ncols1_4-ncols2_8.cu
new file mode 100644
--- /dev/null
+++ b/llama.cpp/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-dca-fused-instance-ncols1_4-ncols2_8.cu
@@ -0,0 +1,10 @@
+// opencoti-hook: cuda-build-parallelism (#531) — DCA-fused MMA case instance, relocated out of
+// fattn.cu so it compiles in the parallel bulk instead of inflating fattn.cu's solo tail.
+// Mirrors template-instances/fattn-mma-f16-instance-*.cu. extern decls live in ../fattn-mma-f16.cuh.
+// DKQ==DV ∈ {128,256,512} (the only head dims DCA dispatches: Gemma-4 512, qwen35moe 256, qwen* 128).
+
+#include "../fattn-mma-f16.cuh"
+
+DECL_FATTN_MMA_F16_DCA_FUSED_CASE(128, 128, 4, 8);
+DECL_FATTN_MMA_F16_DCA_FUSED_CASE(256, 256, 4, 8);
+DECL_FATTN_MMA_F16_DCA_FUSED_CASE(512, 512, 4, 8);
diff --git a/llama.cpp/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-dca-fused-instance-ncols1_64-ncols2_1.cu b/llama.cpp/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-dca-fused-instance-ncols1_64-ncols2_1.cu
new file mode 100644
--- /dev/null
+++ b/llama.cpp/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-dca-fused-instance-ncols1_64-ncols2_1.cu
@@ -0,0 +1,10 @@
+// opencoti-hook: cuda-build-parallelism (#531) — DCA-fused MMA case instance, relocated out of
+// fattn.cu so it compiles in the parallel bulk instead of inflating fattn.cu's solo tail.
+// Mirrors template-instances/fattn-mma-f16-instance-*.cu. extern decls live in ../fattn-mma-f16.cuh.
+// DKQ==DV ∈ {128,256,512} (the only head dims DCA dispatches: Gemma-4 512, qwen35moe 256, qwen* 128).
+
+#include "../fattn-mma-f16.cuh"
+
+DECL_FATTN_MMA_F16_DCA_FUSED_CASE(128, 128, 64, 1);
+DECL_FATTN_MMA_F16_DCA_FUSED_CASE(256, 256, 64, 1);
+DECL_FATTN_MMA_F16_DCA_FUSED_CASE(512, 512, 64, 1);
diff --git a/llama.cpp/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-dca-fused-instance-ncols1_8-ncols2_1.cu b/llama.cpp/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-dca-fused-instance-ncols1_8-ncols2_1.cu
new file mode 100644
--- /dev/null
+++ b/llama.cpp/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-dca-fused-instance-ncols1_8-ncols2_1.cu
@@ -0,0 +1,10 @@
+// opencoti-hook: cuda-build-parallelism (#531) — DCA-fused MMA case instance, relocated out of
+// fattn.cu so it compiles in the parallel bulk instead of inflating fattn.cu's solo tail.
+// Mirrors template-instances/fattn-mma-f16-instance-*.cu. extern decls live in ../fattn-mma-f16.cuh.
+// DKQ==DV ∈ {128,256,512} (the only head dims DCA dispatches: Gemma-4 512, qwen35moe 256, qwen* 128).
+
+#include "../fattn-mma-f16.cuh"
+
+DECL_FATTN_MMA_F16_DCA_FUSED_CASE(128, 128, 8, 1);
+DECL_FATTN_MMA_F16_DCA_FUSED_CASE(256, 256, 8, 1);
+DECL_FATTN_MMA_F16_DCA_FUSED_CASE(512, 512, 8, 1);
diff --git a/llama.cpp/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-dca-fused-instance-ncols1_8-ncols2_2.cu b/llama.cpp/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-dca-fused-instance-ncols1_8-ncols2_2.cu
new file mode 100644
--- /dev/null
+++ b/llama.cpp/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-dca-fused-instance-ncols1_8-ncols2_2.cu
@@ -0,0 +1,10 @@
+// opencoti-hook: cuda-build-parallelism (#531) — DCA-fused MMA case instance, relocated out of
+// fattn.cu so it compiles in the parallel bulk instead of inflating fattn.cu's solo tail.
+// Mirrors template-instances/fattn-mma-f16-instance-*.cu. extern decls live in ../fattn-mma-f16.cuh.
+// DKQ==DV ∈ {128,256,512} (the only head dims DCA dispatches: Gemma-4 512, qwen35moe 256, qwen* 128).
+
+#include "../fattn-mma-f16.cuh"
+
+DECL_FATTN_MMA_F16_DCA_FUSED_CASE(128, 128, 8, 2);
+DECL_FATTN_MMA_F16_DCA_FUSED_CASE(256, 256, 8, 2);
+DECL_FATTN_MMA_F16_DCA_FUSED_CASE(512, 512, 8, 2);
diff --git a/llama.cpp/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-dca-fused-instance-ncols1_8-ncols2_4.cu b/llama.cpp/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-dca-fused-instance-ncols1_8-ncols2_4.cu
new file mode 100644
--- /dev/null
+++ b/llama.cpp/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-dca-fused-instance-ncols1_8-ncols2_4.cu
@@ -0,0 +1,10 @@
+// opencoti-hook: cuda-build-parallelism (#531) — DCA-fused MMA case instance, relocated out of
+// fattn.cu so it compiles in the parallel bulk instead of inflating fattn.cu's solo tail.
+// Mirrors template-instances/fattn-mma-f16-instance-*.cu. extern decls live in ../fattn-mma-f16.cuh.
+// DKQ==DV ∈ {128,256,512} (the only head dims DCA dispatches: Gemma-4 512, qwen35moe 256, qwen* 128).
+
+#include "../fattn-mma-f16.cuh"
+
+DECL_FATTN_MMA_F16_DCA_FUSED_CASE(128, 128, 8, 4);
+DECL_FATTN_MMA_F16_DCA_FUSED_CASE(256, 256, 8, 4);
+DECL_FATTN_MMA_F16_DCA_FUSED_CASE(512, 512, 8, 4);
diff --git a/llama.cpp/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-dca-fused-instance-ncols1_8-ncols2_8.cu b/llama.cpp/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-dca-fused-instance-ncols1_8-ncols2_8.cu
new file mode 100644
--- /dev/null
+++ b/llama.cpp/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-dca-fused-instance-ncols1_8-ncols2_8.cu
@@ -0,0 +1,10 @@
+// opencoti-hook: cuda-build-parallelism (#531) — DCA-fused MMA case instance, relocated out of
+// fattn.cu so it compiles in the parallel bulk instead of inflating fattn.cu's solo tail.
+// Mirrors template-instances/fattn-mma-f16-instance-*.cu. extern decls live in ../fattn-mma-f16.cuh.
+// DKQ==DV ∈ {128,256,512} (the only head dims DCA dispatches: Gemma-4 512, qwen35moe 256, qwen* 128).
+
+#include "../fattn-mma-f16.cuh"
+
+DECL_FATTN_MMA_F16_DCA_FUSED_CASE(128, 128, 8, 8);
+DECL_FATTN_MMA_F16_DCA_FUSED_CASE(256, 256, 8, 8);
+DECL_FATTN_MMA_F16_DCA_FUSED_CASE(512, 512, 8, 8);