diff --git a/.github/workflows/ci.yml b/.github/workflows/ci.yml
index 17e21f7f..3bb942ed 100644
--- a/.github/workflows/ci.yml
+++ b/.github/workflows/ci.yml
@@ -144,6 +144,9 @@ jobs:
           - config_name: Manpages
             sos_config: --enable-manpages --enable-pmi-simple
             libfabric_version: v1.13.x
+          - config_name: Without OFI inject
+            sos_config: --disable-ofi-inject --enable-pmi-simple
+            libfabric_version: v1.13.x
                      
     name: OFI ${{ matrix.libfabric_version }} (${{ matrix.config_name }})
 
diff --git a/src/shmem_comm.h b/src/shmem_comm.h
index 8624f653..f10c6b0d 100644
--- a/src/shmem_comm.h
+++ b/src/shmem_comm.h
@@ -32,30 +32,45 @@
 
 static inline
 void
-shmem_internal_put_scalar(shmem_ctx_t ctx, void *target, const void *source, size_t len, int pe)
+shmem_internal_put_nb(shmem_ctx_t ctx, void *target, const void *source, size_t len, int pe,
+                      long *completion)
 {
-    shmem_internal_assert(len > 0);
+    if (len == 0)
+        return;
 
     if (shmem_shr_transport_use_write(ctx, target, source, len, pe)) {
-        shmem_shr_transport_put_scalar(ctx, target, source, len, pe);
+        shmem_shr_transport_put(ctx, target, source, len, pe);
     } else {
-        shmem_transport_put_scalar((shmem_transport_ctx_t *)ctx, target, source, len, pe);
+        shmem_transport_put_nb((shmem_transport_ctx_t *)ctx, target, source, len, pe, completion);
     }
 }
 
 
 static inline
 void
-shmem_internal_put_nb(shmem_ctx_t ctx, void *target, const void *source, size_t len, int pe,
-                      long *completion)
+shmem_internal_put_wait(shmem_ctx_t ctx, long *completion)
 {
-    if (len == 0)
-        return;
+    shmem_transport_put_wait((shmem_transport_ctx_t *)ctx, completion);
+    /* on-node is always blocking, so this is a no-op for them */
+}
+
+
+static inline
+void
+shmem_internal_put_scalar(shmem_ctx_t ctx, void *target, const void *source, size_t len, int pe)
+{
+    shmem_internal_assert(len > 0);
 
     if (shmem_shr_transport_use_write(ctx, target, source, len, pe)) {
-        shmem_shr_transport_put(ctx, target, source, len, pe);
+        shmem_shr_transport_put_scalar(ctx, target, source, len, pe);
     } else {
-        shmem_transport_put_nb((shmem_transport_ctx_t *)ctx, target, source, len, pe, completion);
+#ifndef DISABLE_OFI_INJECT
+        shmem_transport_put_scalar((shmem_transport_ctx_t *)ctx, target, source, len, pe);
+#else
+        long completion = 0;
+        shmem_transport_put_nb((shmem_transport_ctx_t *)ctx, target, source, len, pe, &completion);
+	shmem_internal_put_wait(ctx, &completion);
+#endif
     }
 }
 
@@ -106,15 +121,6 @@ shmem_internal_put_ct_nb(shmemx_ct_t ct, void *target, const void *source, size_
 }
 
 
-static inline
-void
-shmem_internal_put_wait(shmem_ctx_t ctx, long *completion)
-{
-    shmem_transport_put_wait((shmem_transport_ctx_t *)ctx, completion);
-    /* on-node is always blocking, so this is a no-op for them */
-}
-
-
 static inline
 void
 shmem_internal_get(shmem_ctx_t ctx, void *target, const void *source, size_t len, int pe)