llama/compat: collapse text-loader hook back to one new patch line

Previous addition for glm-ocr added 4 lines to upstream-edits.patch
(1 in constructor for set_loader_path + 3 in load_all_data for the
maybe_load_text_tensor if/continue/closing-brace block). Refactor to
+1 line:

  * Fold path capture into translate_metadata: signature now takes
    `const char * fname` and stashes it on the per-loader registry
    internally. The constructor still has just one llama_compat call
    (we just changed the argument list).
  * Make maybe_load_text_tensor self-contained: it derives the buffer
    type from `cur->buffer` rather than the caller passing it. The
    hook line in load_all_data collapses to a single
    `if (llama_ollama_compat::maybe_load_text_tensor(this, cur, weight->offs)) continue;`.
  * Drop the public `set_loader_path` symbol — it's now an internal
    detail of translate_metadata.

Net patch growth from glm-ocr support: +1 line. Total upstream patch
is back to its original 17-line surface.

Functional: glm-ocr text generation still works ("Paris" via raw
turn-template completion), concat ops still fire (28MB per block).

llama/compat/llama-ollama-compat.cpp

@@ -17,6 +17,12 @@ using namespace llama_ollama_compat::detail; // pull detail:: helpers into scope
 
 namespace {
 
+// Per-loader file path registry — set by translate_metadata, read by
+// maybe_load_text_tensor so it can pass the path to load ops without a
+// separate patch insertion in the model loader's load_all_data path.
+std::mutex g_loader_path_mutex;
+std::unordered_map<const llama_model_loader *, std::string> g_loader_paths;
+
 // =========================================================================
 // gemma3 (text side)
 // =========================================================================
@@ -1297,8 +1303,13 @@ void handle_mistral3_clip(gguf_context * meta, ggml_context * ctx) {
 void translate_metadata(const llama_model_loader * ml,
                         gguf_context * meta,
                         ggml_context * ctx,
-                        std::string & arch_name) {
+                        std::string & arch_name,
+                        const char * fname) {
     if (!meta) return;
+    {
+        std::lock_guard<std::mutex> lk(g_loader_path_mutex);
+        g_loader_paths[ml] = fname ? fname : "";
+    }
     if (arch_name == "gemma3")    handle_gemma3   (ml, meta, ctx);
     if (arch_name == "gemma4")    handle_gemma4   (ml, meta, ctx);
     if (arch_name == "qwen35moe") handle_qwen35moe(ml, meta, ctx);
@@ -1375,20 +1386,9 @@ bool maybe_load_tensor(ggml_tensor * cur,
     return true;
 }
 
-namespace {
-std::mutex g_loader_path_mutex;
-std::unordered_map<const llama_model_loader *, std::string> g_loader_paths;
-}
-
-void set_loader_path(const llama_model_loader * ml, const char * fname) {
-    std::lock_guard<std::mutex> lk(g_loader_path_mutex);
-    g_loader_paths[ml] = fname ? fname : "";
-}
-
 bool maybe_load_text_tensor(const llama_model_loader * ml,
                             ggml_tensor * cur,
-                            size_t file_offset,
-                            ggml_backend_buffer_type_t buft) {
+                            size_t file_offset) {
     std::string path;
     {
         std::lock_guard<std::mutex> lk(g_loader_path_mutex);
@@ -1396,6 +1396,9 @@ bool maybe_load_text_tensor(const llama_model_loader * ml,
         if (it == g_loader_paths.end() || it->second.empty()) return false;
         path = it->second;
     }
+    ggml_backend_buffer_type_t buft = cur->buffer
+        ? ggml_backend_buffer_get_type(cur->buffer)
+        : nullptr;
     return maybe_load_tensor(cur, path.c_str(), file_offset, buft);
 }
 

llama/compat/llama-ollama-compat.h

@@ -35,10 +35,13 @@ struct llama_model_loader;
 namespace llama_ollama_compat {
 
 // Called from llama_model_loader's constructor, right after the arch is read.
+// `fname` is the model file path, captured here so later load-time hooks
+// (maybe_load_text_tensor) can read raw bytes from it.
 void translate_metadata(const llama_model_loader * ml,
                         gguf_context * meta,
                         ggml_context * ctx,
-                        std::string & arch_name);
+                        std::string & arch_name,
+                        const char * fname);
 
 // Called from llama_model_loader's weights_map population loop. Returns
 // true to drop a tensor from the loader — used to hide embedded vision
@@ -60,15 +63,12 @@ bool maybe_load_tensor(ggml_tensor * cur,
                        size_t file_offset,
                        ggml_backend_buffer_type_t buft);
 
-// Same as maybe_load_tensor but for the text-side llama_model_loader,
-// which doesn't have the clip loader's `fname` in scope at the read
-// site. Looks up the model's file path from a per-loader registry
-// populated by `set_loader_path` (called from the model loader's
-// constructor right after `fname` is in scope).
+// Text-side counterpart to maybe_load_tensor. Self-contained: looks up
+// the model file path from the per-loader registry populated by
+// translate_metadata, and derives the buffer type from cur->buffer
+// internally — keeps the call site (and the upstream patch) to one line.
 bool maybe_load_text_tensor(const llama_model_loader * ml,
                             ggml_tensor * cur,
-                            size_t file_offset,
-                            ggml_backend_buffer_type_t buft);
-void set_loader_path(const llama_model_loader * ml, const char * fname);
+                            size_t file_offset);
 
 } // namespace llama_ollama_compat

llama/compat/upstream-edits.patch

@@ -10,12 +10,11 @@ index 4e65a45a5..75836c683 100644
  
  #include <algorithm>
  #include <array>
-@@ -549,6 +550,8 @@ llama_model_loader::llama_model_loader(
+@@ -549,6 +550,7 @@ llama_model_loader::llama_model_loader(
          }
 
          get_key(llm_kv(LLM_KV_GENERAL_ARCHITECTURE), arch_name, false);
-+        llama_ollama_compat::set_loader_path(this, fname.c_str());
-+        llama_ollama_compat::translate_metadata(this, metadata, ctx, arch_name);
++        llama_ollama_compat::translate_metadata(this, metadata, ctx, arch_name, fname.c_str());
          llm_kv = LLM_KV(llm_arch_from_string(arch_name));
  
          files.emplace_back(new llama_file(fname.c_str(), "rb", use_direct_io));
@@ -39,12 +38,10 @@ index 4e65a45a5..75836c683 100644
              // make sure there is no duplicated tensor names
              if (weights_map.find(tensor_name) != weights_map.end()) {
                  throw std::runtime_error(format("invalid model: tensor '%s' is duplicated", ggml_get_name(cur)));
-@@ -1535,3 +1543,6 @@ bool llama_model_loader::load_all_data(
+@@ -1535,3 +1542,4 @@ bool llama_model_loader::load_all_data(
          size_t n_size = ggml_nbytes(cur);
 
-+        if (llama_ollama_compat::maybe_load_text_tensor(this, cur, weight->offs, cur->buffer ? ggml_backend_buffer_get_type(cur->buffer) : nullptr)) {
-+            continue;
-+        }
++        if (llama_ollama_compat::maybe_load_text_tensor(this, cur, weight->offs)) continue;
          if (use_mmap) {
 diff --git a/tools/mtmd/clip.cpp b/tools/mtmd/clip.cpp
 index f0e8786b6..35defa89d 100644