llama : respect no_alloc in all memory caches (#31087)

* llama : respect no_alloc in memory caches

Assisted-by: dsh:llama.cpp/ggml-org/DeepSeek-V4-Flash-Vision-Exp-GGUF (Mac Studio, M2 Ultra, 192GB)

* llama : TODO: move no_alloc out of llama_hparams

Assisted-by: dsh:llama.cpp/ggml-org/DeepSeek-V4-Flash-Vision-Exp-GGUF (Mac Studio, M2 Ultra, 192GB)

* llama : TODO: move no_alloc out of llama_hparams

Assisted-by: dsh:llama.cpp/ggml-org/DeepSeek-V4-Flash-Vision-Exp-GGUF (Mac Studio, M2 Ultra, 192GB)

src/llama-hparams.h

@@ -61,7 +61,7 @@ struct llama_hparams {
     //       for example: n_embd_out() vs n_embd_out_impl
 
     bool vocab_only;
-    bool no_alloc;
+    bool no_alloc; // TODO: no_alloc is not a model hyperparameter; move it out of llama_hparams
     bool rope_finetuned;
     bool use_par_res;
     bool swin_norm;

src/llama-kv-cache-dsv4.cpp

@@ -946,13 +946,12 @@ llama_dsv4_comp_state::llama_dsv4_comp_state(
             uint32_t        n_rs_seq,
         const char    * name,
         const llama_memory_i::layer_filter_cb & filter) :
+    hparams(model.hparams),
     ratio(ratio),
     state_size(state_size),
     n_embd_state(n_embd_state),
     n_stream(unified ? 1 : n_seq_max),
     n_rs_seq(n_rs_seq) {
-    const llama_hparams & hparams = model.hparams;
-
     struct ggml_backend_buft_comparator {
         bool operator()(const ggml_backend_buffer_type_t & lhs, const ggml_backend_buffer_type_t & rhs) const {
             return strcmp(ggml_backend_buft_name(lhs), ggml_backend_buft_name(rhs)) < 0;
@@ -1027,7 +1026,15 @@ llama_dsv4_comp_state::llama_dsv4_comp_state(
     }
 
     for (auto & [buft, ctx] : ctx_map) {
-        ggml_backend_buffer_t buf = ggml_backend_alloc_ctx_tensors_from_buft(ctx.get(), buft);
+        ggml_backend_buffer_t buf;
+        if (hparams.no_alloc) {
+            buf = ggml_backend_buft_alloc_buffer(buft, /*size =*/ 0); // dummy buffer
+            for (ggml_tensor * t = ggml_get_first_tensor(ctx.get()); t != nullptr; t = ggml_get_next_tensor(ctx.get(), t)) {
+                t->buffer = buf; // set dummy buffer for DSV4 state so that the backend scheduler won't try to allocate it
+            }
+        } else {
+            buf = ggml_backend_alloc_ctx_tensors_from_buft(ctx.get(), buft); // real buffer
+        }
         if (!buf) {
             throw std::runtime_error("failed to allocate buffer for DSV4 compressor state");
         }
@@ -1119,9 +1126,15 @@ uint32_t llama_dsv4_comp_state::get_n_rows() const {
 
 std::map<ggml_backend_buffer_type_t, size_t> llama_dsv4_comp_state::memory_breakdown() const {
     std::map<ggml_backend_buffer_type_t, size_t> ret;
-    for (const auto & [_, buf] : ctxs_bufs) {
+    for (const auto & [ctx, buf] : ctxs_bufs) {
         ggml_backend_buffer_type_t buft = ggml_backend_buffer_get_type(buf.get());
-        ret[buft] += ggml_backend_buffer_get_size(buf.get());
+
+        if (hparams.no_alloc) {
+            GGML_ASSERT(ggml_backend_buffer_get_base(buf.get()) == nullptr);
+            ret[buft] += ggml_backend_alloc_ctx_tensors_from_buft_size(ctx.get(), buft);
+        } else {
+            ret[buft] += ggml_backend_buffer_get_size(buf.get());
+        }
     }
     return ret;
 }

src/llama-kv-cache-dsv4.h

@@ -61,6 +61,8 @@ private:
         std::vector<ggml_tensor *> score_stream;
     };
 
+    const llama_hparams & hparams;
+
     const uint32_t ratio;
     const uint32_t state_size;
     const uint32_t n_embd_state;

src/llama-memory-recurrent.cpp

@@ -116,7 +116,15 @@ llama_memory_recurrent::llama_memory_recurrent(
 
     // allocate tensors and initialize the buffers to avoid NaNs in the padding
     for (auto & [buft, ctx] : ctx_map) {
-        ggml_backend_buffer_t buf = ggml_backend_alloc_ctx_tensors_from_buft(ctx.get(), buft);
+        ggml_backend_buffer_t buf;
+        if (hparams.no_alloc) {
+            buf = ggml_backend_buft_alloc_buffer(buft, /*size =*/ 0); // dummy buffer
+            for (ggml_tensor * t = ggml_get_first_tensor(ctx.get()); t != nullptr; t = ggml_get_next_tensor(ctx.get(), t)) {
+                t->buffer = buf; // set dummy buffer for RS cache so that the backend scheduler won't try to allocate it
+            }
+        } else {
+            buf = ggml_backend_alloc_ctx_tensors_from_buft(ctx.get(), buft); // real buffer
+        }
         if (!buf) {
             throw std::runtime_error("failed to allocate buffer for rs cache");
         }
@@ -435,8 +443,15 @@ void llama_memory_recurrent::set_rs_idx(llama_seq_id seq_id, uint32_t idx) {
 
 std::map<ggml_backend_buffer_type_t, size_t> llama_memory_recurrent::memory_breakdown() const {
     std::map<ggml_backend_buffer_type_t, size_t> ret;
-    for (const auto & [_, buf] : ctxs_bufs) {
-        ret[ggml_backend_buffer_get_type(buf.get())] += ggml_backend_buffer_get_size(buf.get());
+    for (const auto & [ctx, buf] : ctxs_bufs) {
+        ggml_backend_buffer_type_t buft = ggml_backend_buffer_get_type(buf.get());
+
+        if (hparams.no_alloc) {
+            GGML_ASSERT(ggml_backend_buffer_get_base(buf.get()) == nullptr);
+            ret[buft] += ggml_backend_alloc_ctx_tensors_from_buft_size(ctx.get(), buft);
+        } else {
+            ret[buft] += ggml_backend_buffer_get_size(buf.get());
+        }
     }
     return ret;
 }