llama : align KV cache tensor-split granularity to cache block size

When a quantized KV cache is used with SPLIT_MODE_TENSOR the per-device
slice size must be a multiple of the cache type's block size (e.g. 32 for
Q4_0).  Previously the granularity was derived solely from the weight
tensor's block size (1 for F16), which happened to satisfy the constraint
only when n_embd_head_k was already a multiple of the KV block size.

Make the alignment explicit by taking the lcm of granularity_kv and
ggml_blck_size(tensor->type) for cache tensors.  For F16/BF16 cache
(blck_size=1) the behaviour is unchanged.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Author: yrougy

src/llama-model.cpp

@@ -592,11 +592,17 @@ struct ggml_backend_meta_split_state llama_meta_device_get_split_state(const str
 
             const int64_t granularity_kv = granularity_q / n_gqa;
             if (std::regex_match(tensor_name, pattern_kv_weight) ||
-                std::regex_match(tensor_name, pattern_kv_bias) ||
-                std::regex_match(tensor_name, pattern_kv_cache)) {
+                std::regex_match(tensor_name, pattern_kv_bias)) {
                 GGML_ASSERT(segments.size() == 1);
                 return {granularity_kv};
             }
+            if (std::regex_match(tensor_name, pattern_kv_cache)) {
+                GGML_ASSERT(segments.size() == 1);
+                // For quantized KV cache the split boundary must also align to
+                // the cache type's block size (e.g. 32 for Q4_0).
+                const int64_t kv_blck_size = ggml_blck_size(tensor->type);
+                return {std::lcm(granularity_kv, kv_blck_size)};
+            }
             if (std::regex_match(tensor_name, pattern_qkv_weight) || std::regex_match(tensor_name, pattern_qkv_bias)) {
                 GGML_ASSERT(segments.size() == 3);
                 return {granularity_q, granularity_kv, granularity_kv};