Merge commit '34df42f7bef5a711b2b40f5d2b6b78254def99c3' into concedo_experimental

# Conflicts:
#	README.md
#	ggml/src/ggml-hexagon/ggml-hexagon.cpp
#	ggml/src/ggml-hexagon/htp/CMakeLists.txt
#	ggml/src/ggml-hexagon/htp/act-ops.c
#	ggml/src/ggml-hexagon/htp/binary-ops.c
#	ggml/src/ggml-hexagon/htp/cpy-ops.c
#	ggml/src/ggml-hexagon/htp/get-rows-ops.c
#	ggml/src/ggml-hexagon/htp/htp-msg.h
#	ggml/src/ggml-hexagon/htp/htp-ops.h
#	ggml/src/ggml-hexagon/htp/hvx-arith.h
#	ggml/src/ggml-hexagon/htp/hvx-base.h
#	ggml/src/ggml-hexagon/htp/hvx-inverse.h
#	ggml/src/ggml-hexagon/htp/hvx-utils.h
#	ggml/src/ggml-hexagon/htp/main.c
#	ggml/src/ggml-hexagon/htp/rope-ops.c
#	ggml/src/ggml-hexagon/htp/set-rows-ops.c
#	ggml/src/ggml-hexagon/htp/softmax-ops.c
#	ggml/src/ggml-hexagon/htp/unary-ops.c
#	ggml/src/ggml-opencl/CMakeLists.txt
#	ggml/src/ggml-opencl/ggml-opencl.cpp
#	tests/test-backend-ops.cpp
#	tools/cli/cli.cpp
#	tools/server/webui/src/lib/components/app/chat/ChatScreen/ChatScreen.svelte

src/llama-batch.cpp

@@ -394,11 +394,13 @@ llama_ubatch llama_batch_allocr::ubatch_reserve(uint32_t n_seq_tokens, uint32_t
     clear();
     split_reset();
 
+    const int64_t n_pos_all = (int64_t) n_tokens*n_pos_per_embd;
+
     auto udata = std::make_shared<llama_ubatch::data_t>();
 
     udata->token     .resize(n_tokens);
     udata->embd      .clear();
-    udata->pos       .resize(n_tokens);
+    udata->pos       .resize(n_pos_all);
     udata->n_seq_id  .resize(n_tokens);
     udata->seq_id    .resize(n_tokens);
     udata->seq_id_unq.resize(0);

src/llama-context.cpp

@@ -1049,11 +1049,15 @@ void llama_context::set_adapters_lora(llama_adapter_lora ** adapters, size_t n_a
 bool llama_context::adapters_lora_are_same(llama_adapter_lora ** adapters, size_t n_adapters, float * scales) {
     LLAMA_LOG_DEBUG("%s: adapters = %p\n", __func__, (void *) adapters);
 
-    if (n_adapters != loras->size()) {
-        return false;
-    }
+    // Adapters with a zero scale are never added to `loras`, so also ignore them for the comparison.
+    size_t n_non_zero = 0;
 
     for (size_t i = 0; i < n_adapters; i ++) {
+        if (scales[i] == 0.0f) {
+            continue;
+        }
+        n_non_zero++;
+
         auto it = loras->find(adapters[i]);
 
         if (it == loras->end() || it->second != scales[i]) {
@@ -1061,6 +1065,10 @@ bool llama_context::adapters_lora_are_same(llama_adapter_lora ** adapters, size_
         }
     }
 
+    if (n_non_zero != loras->size()) {
+        return false;
+    }
+
     return true;
 }
 

src/llama-kv-cache.cpp

@@ -1760,8 +1760,10 @@ void llama_kv_cache::state_write_meta(llama_io_write_i & io, const cell_ranges_t
             io.write(&pos,      sizeof(pos));
             io.write(&n_seq_id, sizeof(n_seq_id));
 
-            // TODO: we also need to save llama_kv_cell_ext when apply_ubatch() support loading it
-            //       see: https://github.com/ggml-org/llama.cpp/pull/16825#issuecomment-3460868350
+            if (hparams.n_pos_per_embd() > 1) {
+                const llama_kv_cell_ext ext = cells.ext_get(i);
+                io.write(&ext, sizeof(ext));
+            }
 
             for (const auto & seq_id : seq_ids) {
                 io.write(&seq_id, sizeof(seq_id));
@@ -1895,6 +1897,14 @@ bool llama_kv_cache::state_read_meta(llama_io_read_i & io, uint32_t strm, uint32
                 return false;
             }
 
+            if (hparams.n_pos_per_embd() > 1) {
+                llama_kv_cell_ext ext;
+                io.read_to(&ext, sizeof(ext));
+
+                ubatch.pos[i + ubatch.n_tokens]   = ext.y;
+                ubatch.pos[i + ubatch.n_tokens*2] = ext.x;
+            }
+
             // read the sequence id, but directly discard it - we will use dest_seq_id instead
             {
                 llama_seq_id seq_id;

src/models/delta-net-base.cpp

@@ -1,7 +1,5 @@
 #include "models.h"
 
-#define CHUNK_SIZE 64
-
 // utility to get one slice from the third dimension
 // input dim:  [x, y, c, b]
 // output dim: [x, y, 1, b]
@@ -57,7 +55,7 @@ std::pair<ggml_tensor *, ggml_tensor *> llm_build_delta_net_base::build_delta_ne
     g = ggml_permute(ctx0, g, 0, 2, 1, 3); // [g_0, n_tokens, H_v, n_seqs]
     b = ggml_permute(ctx0, b, 0, 2, 1, 3); // [  1, n_tokens, H_v, n_seqs]
 
-    const int CS = CHUNK_SIZE;
+    const int CS = kda ? 16 : 64; // chunk size
 
     const int pad = (CS - n_tokens % CS) % CS;
     const int n_chunks = (n_tokens + pad) / CS;