fix: scale hot KV allocation to hot_pct of total ctx

The tiered cache percentages define hard capacity limits per tier that
sum to the full ctx budget. Previously the R9700 KV buffer was still
allocated at full n_ctx despite the hot tier being only a fraction.

- Add kv_tier_total_ctx to common_params to preserve the full budget
- Before common_init_from_params, reduce n_ctx to hot_pct * n_ctx so
  the GPU KV buffer only allocates for the hot tier
- server-tiered-cache.cpp uses kv_tier_total_ctx for config.total_ctx
  so warm/cold capacity calculations remain correct

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

Author: kmbandy

common/common.h

@@ -583,6 +583,7 @@ struct common_params {
     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)
+    int   kv_tier_total_ctx     = 0;     // full ctx budget across all tiers (set at load time)
 
     std::string hostname      = "127.0.0.1";
     std::string public_path   = "";                                                                         // NOLINT

tools/server/server-context.cpp

@@ -750,6 +750,11 @@ struct server_context_impl {
         SRV_INF("loading model '%s'\n", params.model.path.c_str());
 
         params_base = params;
+        if (params_base.kv_tiered_enabled && params_base.kv_tier_hot_pct > 0.0f && params_base.kv_tier_hot_pct < 100.0f) {
+            params_base.kv_tier_total_ctx = params_base.n_ctx;
+            params_base.n_ctx = std::max(512, (int)(params_base.n_ctx * params_base.kv_tier_hot_pct / 100.0f));
+            SRV_INF("tiered KV: total ctx=%d, hot ctx=%d (%.0f%%)\n", params_base.kv_tier_total_ctx, params_base.n_ctx, params_base.kv_tier_hot_pct);
+        }
 
         llama_init = common_init_from_params(params_base);
 

tools/server/server-tiered-cache.cpp

@@ -60,7 +60,7 @@ bool server_tiered_cache::init_slot(int slot_id, const llama_model& model) {
     config.hot_percent = params.kv_tier_hot_pct;
     config.warm_percent = params.kv_tier_warm_pct;
     config.cold_percent = params.kv_tier_cold_pct;
-    config.total_ctx = params.n_ctx;
+    config.total_ctx = params.kv_tier_total_ctx > 0 ? params.kv_tier_total_ctx : params.n_ctx;
     config.warm_device = params.kv_warm_device;
 
     // Create tiered cache instance