feat: NVMe->VRAM demand weight paging via SAM (Phases 1-4)

- llama-weight-pager: page table, VRAM slot pool, LRU eviction
- page_in via pread into hipExtMallocWithFlags fine-grained VRAM (SAM path)
- ggml eval callback hook: redirects tensor data pointers before each op
- async io_uring prefetch: pipelines next layer NVMe read with GPU compute
- CLI: --weight-paging, --weight-paging-slots, --weight-paging-prefetch
- metrics: weight pager stats exposed at /metrics endpoint

Fixes: hipMalloc pool alloc, strcmp tensor name comparison, slot eviction
page invalidation, llama-model.h include for server-context.

Author: kmbandy

common/arg.cpp

@@ -2430,6 +2430,27 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             }
         }
     ).set_env("LLAMA_ARG_N_GPU_LAYERS"));
+    add_opt(common_arg(
+        {"--weight-paging"}, "",
+        "enable NVMe→VRAM demand paging for model weights (allows models larger than VRAM)",
+        [](common_params & params, const std::string & value) {
+            params.weight_paging_enabled = true;
+        }
+    ).set_examples({LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_CLI}).set_env("LLAMA_ARG_WEIGHT_PAGING"));
+    add_opt(common_arg(
+        {"--weight-paging-slots"}, "N",
+        "number of VRAM slots for weight paging (-1 = auto, default: -1)",
+        [](common_params & params, const std::string & value) {
+            params.weight_paging_slots = std::stoi(value);
+        }
+    ).set_examples({LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_CLI}).set_env("LLAMA_ARG_WEIGHT_PAGING_SLOTS"));
+    add_opt(common_arg(
+        {"--weight-paging-prefetch"}, "",
+        "enable async prefetch of next layer (default: enabled)",
+        [](common_params & params, const std::string & value) {
+            params.weight_paging_prefetch = true;
+        }
+    ).set_examples({LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_CLI}).set_env("LLAMA_ARG_WEIGHT_PAGING_PREFETCH"));
     add_opt(common_arg(
         {"-sm", "--split-mode"}, "{none,layer,row,tensor}",
         "how to split the model across multiple GPUs, one of:\n"

common/common.h

@@ -533,6 +533,11 @@ struct common_params {
     bool no_extra_bufts    = false; // disable extra buffer types (used for weight repacking)
     bool no_host           = false; // bypass host buffer allowing extra buffers to be used
 
+    // weight paging parameters
+    bool    weight_paging_enabled = false;  // enable NVMe→VRAM demand paging for model weights
+    int32_t weight_paging_slots   = -1;     // number of VRAM slots for weight paging (-1 = auto)
+    bool    weight_paging_prefetch = true;  // enable async prefetch of next layer
+
     bool single_turn       = false; // single turn chat conversation
 
     ggml_type cache_type_k = GGML_TYPE_F16; // KV cache data type for the K

src/CMakeLists.txt

@@ -34,6 +34,7 @@ add_library(llama
             llama-io-uring.cpp
             llama-model-saver.cpp
             llama-model.cpp
+            llama-weight-pager.cpp
             llama-quant.cpp
             llama-sampler.cpp
             llama-vocab.cpp

src/llama-model-loader.h

@@ -6,6 +6,7 @@
 #include "llama-arch.h"
 #include "llama-hparams.h"
 #include "llama-mmap.h"
+#include "llama-weight-pager.h"
 
 #include "ggml-cpp.h"
 
@@ -28,6 +29,13 @@ enum llama_fver {
 
 const char * llama_file_version_name(llama_fver version);
 
+// Information about a weight tensor for the weight pager
+struct llama_weight_page_info {
+    std::string name;   // tensor name
+    size_t      offset; // file offset in GGUF
+    size_t      size;   // tensor size in bytes
+};
+
 struct llama_model_loader {
     // Holds information on a model weight
     struct llama_tensor_weight {
@@ -90,6 +98,9 @@ struct llama_model_loader {
     std::unordered_map<std::string, llama_model_kv_override> kv_overrides;
     const llama_model_tensor_buft_override * tensor_buft_overrides;
 
+    // weight pager info (Phase 2)
+    std::vector<llama_weight_page_info> weight_page_infos;
+
     gguf_context_ptr metadata_ptr;
     struct gguf_context * metadata; // either metadata_ptr.get() or externally set
     llama_model_set_tensor_data_t set_tensor_data;

src/llama-model.h

@@ -6,6 +6,7 @@
 #include "llama-hparams.h"
 #include "llama-memory.h"
 #include "llama-vocab.h"
+#include "llama-weight-pager.h"
 
 #include <map>
 #include <memory>
@@ -571,6 +572,9 @@ struct llama_model {
     // for keeping track of associated LoRA adapters
     std::unordered_set<llama_adapter_lora *> loras;
 
+    // weight paging for NVMe→VRAM demand paging
+    std::unique_ptr<llama_weight_pager> weight_pager;
+
     // statically allocated context for assigning
     struct llama_meta_device_get_split_state_userdata get_split_state_ud;