Server: fused single-GPU decode for --batched-session (co-decode 2-4 sessions in one graph pass)

Makefile

@@ -344,6 +344,18 @@ test-cuda-mixed-batch: tests/test_cuda_mixed_batch
 	DS4_TEST_MODEL="$(DS4_TEST_MODEL)" ./tests/test_cuda_mixed_batch
 endif
 
+tests/attention_multi_smoke.o: tests/attention_multi_smoke.c ds4_gpu.h
+	$(CC) $(CFLAGS) -I. -c -o $@ tests/attention_multi_smoke.c
+
+tests/attention_multi_smoke: tests/attention_multi_smoke.o ds4_cuda.o
+	$(NVCC) $(NVCCFLAGS) -o $@ $^ $(CUDA_LDLIBS)
+
+tests/rope_multi_bench.o: tests/rope_multi_bench.c ds4_gpu.h
+	$(CC) $(CFLAGS) -I. -c -o $@ tests/rope_multi_bench.c
+
+tests/rope_multi_bench: tests/rope_multi_bench.o ds4_cuda.o
+	$(NVCC) $(NVCCFLAGS) -o $@ $^ $(CUDA_LDLIBS)
+
 ds4_test: ds4_test.o ds4_help.o ds4_kvstore.o rax.o $(CORE_OBJS)
 ifeq ($(UNAME_S),Darwin)
 	$(CC) $(CFLAGS) -o $@ ds4_test.o ds4_help.o ds4_kvstore.o rax.o $(CORE_OBJS) $(METAL_LDLIBS)

ds4.h

@@ -160,6 +160,10 @@ typedef struct {
     int placement_ctx_hint;
     /* Server batch mode serializes execution and can share prefill scratch. */
     bool share_session_prefill_workspace;
+    /* Number of resident sessions the batched server will create after the
+     * engine opens; used to keep that much VRAM out of the lazy weight
+     * caches so session creation cannot fail.  0 = single-session server. */
+    int resident_sessions;
     bool first_token_test;
     bool metal_graph_test;
     bool load_slice;