feat: --kv-warm-device CLI flag + set_warm_elem_bytes wiring

Adds --kv-warm-device <N> flag (env: LLAMA_ARG_KV_WARM_DEVICE) to
specify a HIP device index for the warm KV cache tier. Wires through
common_params into llama_tier_config.warm_device and calls
set_warm_elem_bytes() before init() so hipMalloc sizing is correct.

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

Author: kmbandy

common/arg.cpp

@@ -1358,6 +1358,13 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             params.kv_tier_attention_threshold = float(value) / 100.0f;
         }
     ).set_env("LLAMA_ARG_TIER_ATTENTION_THRESHOLD").set_examples({LLAMA_EXAMPLE_SERVER}));
+    add_opt(common_arg(
+        {"--kv-warm-device"}, "DEVICE",
+        "HIP device index to use as warm KV cache tier (e.g. 1 for 6900XT eGPU); requires --kv-tiered",
+        [](common_params & params, int value) {
+            params.kv_warm_device = value;
+        }
+    ).set_env("LLAMA_ARG_KV_WARM_DEVICE").set_examples({LLAMA_EXAMPLE_SERVER}));
     add_opt(common_arg(
         {"-kvu", "--kv-unified"},
         {"-no-kvu", "--no-kv-unified"},

common/common.h

@@ -582,6 +582,7 @@ struct common_params {
     int kv_tier_eviction_policy = 3;     // 0=LRU, 1=LFU, 2=attention, 3=hybrid (default)
     int kv_tier_compression   = 1;       // 0=none, 1=int4, 2=int8, 3=lz4, 4=quantized
     float kv_tier_attention_threshold = 0.1f;  // attention threshold for eviction
+    int   kv_warm_device        = -1;    // HIP device index for warm KV tier (-1 = disabled)
 
     std::string hostname      = "127.0.0.1";
     std::string public_path   = "";                                                                         // NOLINT

src/llama-kv-cache-tiered.h

@@ -182,7 +182,11 @@ class llama_kv_cache_tiered {
     bool resize(uint32_t new_total_ctx);
     bool set_eviction_policy(llama_eviction_policy policy);
     bool set_attention_threshold(float threshold);
+#ifdef GGML_USE_HIP
     void set_warm_elem_bytes(size_t n) { warm_elem_bytes = n; }
+#else
+    void set_warm_elem_bytes(size_t) {}
+#endif
 
     // Eviction interface
     bool evict_tokens(uint32_t n_tokens_to_evict, llama_cache_tier from_tier);

tools/server/server-tiered-cache.cpp

@@ -61,6 +61,7 @@ bool server_tiered_cache::init_slot(int slot_id, const llama_model& model) {
     config.warm_percent = params.kv_tier_warm_pct;
     config.cold_percent = params.kv_tier_cold_pct;
     config.total_ctx = params.n_ctx;
+    config.warm_device = params.kv_warm_device;
 
     // Create tiered cache instance
     manager.tiered_cache = std::make_unique<llama_kv_cache_tiered>(
@@ -72,6 +73,14 @@ bool server_tiered_cache::init_slot(int slot_id, const llama_model& model) {
         attention_threshold
     );
 
+    // set bytes-per-token for warm device hipMalloc sizing
+    if (params.kv_warm_device >= 0) {
+        int32_t n_kv_heads = llama_model_n_head_kv(&model);
+        int32_t head_dim   = llama_model_n_embd(&model) / llama_model_n_head(&model);
+        manager.tiered_cache->set_warm_elem_bytes(
+            (size_t)n_kv_heads * head_dim * sizeof(ggml_fp16_t));
+    }
+
     // Initialize the tiered cache
     if (!manager.tiered_cache->init()) {
         return false;