diff --git a/src/llama-context.cpp b/src/llama-context.cpp
index a6d2f264fc..a920c42311 100644
--- a/src/llama-context.cpp
+++ b/src/llama-context.cpp
@@ -2907,17 +2907,28 @@ public:
             }
 
             if (mbuf_cur.n_tensors == mbuf.n_tensors) {
-                // same chunking: copy 1:1 by index
+                // an equal tensor count does not imply the same chunking, e.g. save ranges [2,1] vs restore runs [1,2]
+                bool same_chunking = true;
                 for (size_t i = 0; i < mbuf_cur.org.size(); ++i) {
-                    GGML_ASSERT(ggml_nbytes(mbuf_cur.cpy[i]) == ggml_nbytes(mbuf.org[i]));
-                    ggml_backend_tensor_copy(mbuf_cur.cpy[i], mbuf.org[i]);
+                    if (ggml_nbytes(mbuf_cur.cpy[i]) != ggml_nbytes(mbuf.org[i])) {
+                        same_chunking = false;
+                        break;
+                    }
+                }
+
+                if (same_chunking) {
+                    // same chunking: copy 1:1 by index
+                    for (size_t i = 0; i < mbuf_cur.org.size(); ++i) {
+                        ggml_backend_tensor_copy(mbuf_cur.cpy[i], mbuf.org[i]);
+                    }
+                    continue;
                 }
-                continue;
             }
 
             // different chunking: copy the write-side data (mbuf_cur.cpy) into the read-side targets (mbuf.org)
             // with a byte cursor. Write and read enumerate the same logical data in the same order but may chunk
-            // it differently, so copy across tensor boundaries rather than 1:1 by index.
+            // it differently (even with an equal number of tensors), so copy across tensor boundaries rather than
+            // 1:1 by index.
             const size_t total = mbuf_cur.total_size;
 
             ggml_init_params params_scratch = {
diff --git a/src/llama-kv-cache.cpp b/src/llama-kv-cache.cpp
index 65afbd8c37..0e095f8b6b 100644
--- a/src/llama-kv-cache.cpp
+++ b/src/llama-kv-cache.cpp
@@ -2533,6 +2533,24 @@ bool llama_kv_cache::state_read_meta(llama_io_read_i & io, uint32_t strm, uint32
 bool llama_kv_cache::state_read_data(llama_io_read_i & io, uint32_t strm, uint32_t cell_count, const slot_info & sinfo) {
     auto & cells = v_cells[strm];
 
+    // batch the scatter reads per contiguous run of destination indices
+    // from inclusive, to exclusive - same convention as cell_ranges_t
+    // contiguous cells yield a single run covering the whole block
+    struct cell_run { uint32_t from; uint32_t to; };
+    std::vector<cell_run> runs;
+    if (cell_count > 0) {
+        const auto & idxs = sinfo.idxs[0];
+        uint32_t i0 = 0;
+        while (i0 < cell_count) {
+            uint32_t i1 = i0 + 1;
+            while (i1 < cell_count && idxs[i1] == idxs[i1 - 1] + 1) {
+                ++i1;
+            }
+            runs.push_back({idxs[i0], idxs[i1 - 1] + 1});
+            i0 = i1;
+        }
+    }
+
     uint32_t v_trans;
     uint32_t n_layer;
 
@@ -2580,17 +2598,8 @@ bool llama_kv_cache::state_read_data(llama_io_read_i & io, uint32_t strm, uint32
             return false;
         }
 
-        if (cell_count) {
-            if (sinfo.is_contiguous()) {
-                // Fast path: contiguous cells, single memcpy
-                io.read_tensor(k, sinfo.head() * k_size_row, cell_count * k_size_row);
-            } else {
-                // Slow path: scatter to non-contiguous positions
-                for (uint32_t i = 0; i < cell_count; ++i) {
-                    const size_t dst_offset = sinfo.idxs[0][i] * k_size_row;
-                    io.read_tensor(k, dst_offset, k_size_row);
-                }
-            }
+        for (const auto & r : runs) {
+            io.read_tensor(k, (size_t) r.from * k_size_row, (size_t) (r.to - r.from) * k_size_row);
         }
     }
 
@@ -2623,17 +2632,8 @@ bool llama_kv_cache::state_read_data(llama_io_read_i & io, uint32_t strm, uint32
                 return false;
             }
 
-            if (cell_count) {
-                if (sinfo.is_contiguous()) {
-                    // Fast path: contiguous cells, single memcpy
-                    io.read_tensor(v, sinfo.head() * v_size_row, cell_count * v_size_row);
-                } else {
-                    // Slow path: scatter to non-contiguous positions
-                    for (uint32_t i = 0; i < cell_count; ++i) {
-                        const size_t dst_offset = sinfo.idxs[0][i] * v_size_row;
-                        io.read_tensor(v, dst_offset, v_size_row);
-                    }
-                }
+            for (const auto & r : runs) {
+                io.read_tensor(v, (size_t) r.from * v_size_row, (size_t) (r.to - r.from) * v_size_row);
             }
         }
     } else {
@@ -2674,22 +2674,10 @@ bool llama_kv_cache::state_read_data(llama_io_read_i & io, uint32_t strm, uint32
                 return false;
             }
 
-            if (cell_count) {
-                if (sinfo.is_contiguous()) {
-                    // Fast path: contiguous cells
-                    const uint32_t h = sinfo.head();
-                    for (uint32_t j = 0; j < n_embd_v_gqa; ++j) {
-                        const size_t dst_offset = (h + j * cells.size()) * v_size_el;
-                        io.read_tensor(v, dst_offset, cell_count * v_size_el);
-                    }
-                } else {
-                    // Slow path: scatter to non-contiguous positions
-                    for (uint32_t j = 0; j < n_embd_v_gqa; ++j) {
-                        for (uint32_t i = 0; i < cell_count; ++i) {
-                            const size_t dst_offset = (sinfo.idxs[0][i] + j * cells.size()) * v_size_el;
-                            io.read_tensor(v, dst_offset, v_size_el);
-                        }
-                    }
+            for (uint32_t j = 0; j < n_embd_v_gqa; ++j) {
+                for (const auto & r : runs) {
+                    const size_t dst_offset = ((size_t) r.from + j * cells.size()) * v_size_el;
+                    io.read_tensor(v, dst_offset, (size_t) (r.to - r.from) * v_size_el);
                 }
             }
         }
diff --git a/tests/test-save-load-state.cpp b/tests/test-save-load-state.cpp
index 0ceab7c545..35c7690587 100644
--- a/tests/test-save-load-state.cpp
+++ b/tests/test-save-load-state.cpp
@@ -355,7 +355,101 @@ static bool test_seq_cp_device(struct llama_model * model, const struct common_p
 }
 
 
-// Run the full save/load test suite (tests 1-5) for a single model.
+// Test 6/7: seq copy (scatter)
+// - decode the same prefix on two sequences, interleaving seq 0 cells between the seq 1 cells
+// - save the seq 1 state, free the interleaved seq 0 cells, and restore via the given io path
+// - the restore destination is non-contiguous: scatter reads are batched per contiguous run
+// - save again on the host and compare the two blobs byte for byte
+static bool test_seq_cp_scatter(struct llama_model * model, const struct common_params & params, const llama_tokens & tokens, int test_num, bool on_device) {
+    auto params_ctx = common_context_params_to_llama(params);
+    params_ctx.n_ctx      = 256;
+    params_ctx.n_seq_max  = 2;
+    params_ctx.kv_unified = true;
+    auto ctx = llama_context_ptr{llama_init_from_model(model, params_ctx)};
+
+    LOG("\n=== Test %d: seq copy (%s, scatter) ===\n", test_num, on_device ? "device" : "host");
+
+    const uint32_t flags = on_device ? LLAMA_STATE_SEQ_FLAGS_ON_DEVICE : LLAMA_STATE_SEQ_FLAGS_NONE;
+
+    auto decode_one = [&](llama_token tok, int pos, llama_seq_id seq) {
+        llama_batch_ptr batch(1, 0, 1);
+        common_batch_add(batch.get(), tok, pos, { seq }, false);
+        return llama_decode(ctx.get(), batch.get()) == 0;
+    };
+
+    // seq 0 cells 0,1,4 interleave the seq 1 cells 2,3,5
+    if (!decode_one(tokens[0], 0, 0) ||
+        !decode_one(tokens[1], 1, 0) ||
+        !decode_one(tokens[0], 0, 1) ||
+        !decode_one(tokens[1], 1, 1) ||
+        !decode_one(tokens[2], 2, 0) ||
+        !decode_one(tokens[2], 2, 1)) {
+        LOG_ERR("%s: failed to build interleaved state\n", __func__);
+        return false;
+    }
+
+    const auto get_seq_state = [&](llama_seq_id seq_id, uint32_t fl, std::vector<uint8_t> & state) {
+        const size_t state_size = llama_state_seq_get_size_ext(ctx.get(), seq_id, fl);
+        if (state_size == 0) {
+            LOG_ERR("%s: sequence state is empty\n", __func__);
+            return false;
+        }
+
+        state.resize(state_size);
+        const size_t ncopy = llama_state_seq_get_data_ext(ctx.get(), state.data(), state.size(), seq_id, fl);
+        if (ncopy != state.size()) {
+            LOG_ERR("%s: sequence state length %zu does not match expected length %zu\n",
+                    __func__, ncopy, state.size());
+            return false;
+        }
+
+        return true;
+    };
+
+    // host blob: contains the KV data, used for the byte-for-byte comparison
+    std::vector<uint8_t> state_before;
+    if (!get_seq_state(1, LLAMA_STATE_SEQ_FLAGS_NONE, state_before)) {
+        return false;
+    }
+
+    // save via the io path under test
+    std::vector<uint8_t> state_save;
+    if (!get_seq_state(1, flags, state_save)) {
+        return false;
+    }
+    LOG_TRC("%s: seq 1 saved via %s, %zu bytes\n", __func__, on_device ? "device" : "host", state_save.size());
+
+    // free seq 0's cells so the ring is fragmented: the restore destination (seq 1's interleaved cells) stays non-contiguous
+    if (!llama_memory_seq_rm(llama_get_memory(ctx.get()), 0, -1, -1)) {
+        LOG_ERR("%s: failed to remove sequence 0\n", __func__);
+        return false;
+    }
+
+    // restore via the io path under test
+    const size_t nset = llama_state_seq_set_data_ext(ctx.get(), state_save.data(), state_save.size(), 1, flags);
+    if (nset != state_save.size()) {
+        LOG_ERR("%s: seq set data length %zu does not match expected length %zu\n", __func__, nset, state_save.size());
+        return false;
+    }
+    LOG_TRC("%s: seq 1 restored via %s, %zu bytes\n", __func__, on_device ? "device" : "host", nset);
+
+    std::vector<uint8_t> state_after;
+    if (!get_seq_state(1, LLAMA_STATE_SEQ_FLAGS_NONE, state_after)) {
+        return false;
+    }
+
+    // the blob is serialized in sequence cell order, so identical bytes iff the restore wrote the same KV
+    if (state_before.size() != state_after.size() || memcmp(state_before.data(), state_after.data(), state_before.size()) != 0) {
+        LOG_ERR("\n%s: error: restored KV state is not byte-identical to the saved state\n", __func__);
+        return false;
+    }
+
+    LOG("\nPASS\n");
+    return true;
+}
+
+
+// Run the full save/load test suite (tests 1-7) for a single model.
 // Returns true if all tests pass, false otherwise.
 static bool run_save_load_tests_for_model(const std::string & model_path, const struct common_params & base_params) {
     struct common_params params = base_params;
@@ -422,6 +516,16 @@ static bool run_save_load_tests_for_model(const std::string & model_path, const
         return false;
     }
 
+    // Test 6: seq copy (host, scatter)
+    if (!test_seq_cp_scatter(model, params, tokens, 6, false)) {
+        return false;
+    }
+
+    // Test 7: seq copy (device, scatter)
+    if (!test_seq_cp_scatter(model, params, tokens, 7, true)) {
+        return false;
+    }
+
     LOG("\nAll tests passed.\n");
 
     return true;