server: fix report progress for loading spec models, add "stages" list (#24870)

* server: fix report progress for loading spec models, add "stages" list

* improve

* nits

* nits 2

tools/mtmd/clip.cpp

@@ -2796,7 +2796,7 @@ struct clip_model_loader {
         }
 
         // load data
-        if (!ctx_clip.no_alloc) {
+        {
             std::vector<uint8_t> read_buf;
 
             // start loading event
@@ -2814,38 +2814,42 @@ struct clip_model_loader {
             ggml_backend_buffer_type_t buft = ggml_backend_get_default_buffer_type(ctx_clip.backend);
             ctx_clip.buf.reset(ggml_backend_alloc_ctx_tensors_from_buft(ctx_clip.ctx_data.get(), buft));
             ggml_backend_buffer_set_usage(ctx_clip.buf.get(), GGML_BACKEND_BUFFER_USAGE_WEIGHTS);
-            size_t data_loaded = 0;
-            for (auto & t : tensors_to_load) {
-                ggml_tensor * cur = ggml_get_tensor(ctx_clip.ctx_data.get(), t->name);
-                GGML_ASSERT(cur && "tensor not found in ctx_data");
-                auto it_off = tensor_offset.find(t->name);
-                GGML_ASSERT(it_off != tensor_offset.end() && "no offset for tensor");
-                const size_t offset = it_off->second;
-                fin.seekg(offset, std::ios::beg);
-                if (!fin) {
-                    throw std::runtime_error(string_format("%s: failed to seek for tensor %s\n", __func__, t->name));
-                }
-                size_t num_bytes = ggml_nbytes(cur);
-                if (ggml_backend_buft_is_host(buft)) {
-                    // for the CPU and Metal backend, we can read directly into the tensor
-                    fin.read(reinterpret_cast<char *>(cur->data), num_bytes);
-                } else {
-                    // read into a temporary buffer first, then copy to device memory
-                    read_buf.resize(num_bytes);
-                    fin.read(reinterpret_cast<char *>(read_buf.data()), num_bytes);
-                    ggml_backend_tensor_set(cur, read_buf.data(), 0, num_bytes);
-                }
-                data_loaded += num_bytes;
-                if (progress_callback && total_data_size > 0) {
-                    const float progress = (float)data_loaded / (float)total_data_size;
-                    if (!progress_callback(progress, progress_callback_user_data)) {
-                        throw std::runtime_error(string_format("%s: model loading cancelled by progress_callback\n", __func__));
+            // read the weight from file
+            if (!ctx_clip.no_alloc) {
+                size_t data_loaded = 0;
+                for (auto & t : tensors_to_load) {
+                    ggml_tensor * cur = ggml_get_tensor(ctx_clip.ctx_data.get(), t->name);
+                    GGML_ASSERT(cur && "tensor not found in ctx_data");
+                    auto it_off = tensor_offset.find(t->name);
+                    GGML_ASSERT(it_off != tensor_offset.end() && "no offset for tensor");
+                    const size_t offset = it_off->second;
+                    fin.seekg(offset, std::ios::beg);
+                    if (!fin) {
+                        throw std::runtime_error(string_format("%s: failed to seek for tensor %s\n", __func__, t->name));
+                    }
+                    size_t num_bytes = ggml_nbytes(cur);
+                    if (ggml_backend_buft_is_host(buft)) {
+                        // for the CPU and Metal backend, we can read directly into the tensor
+                        fin.read(reinterpret_cast<char *>(cur->data), num_bytes);
+                    } else {
+                        // read into a temporary buffer first, then copy to device memory
+                        read_buf.resize(num_bytes);
+                        fin.read(reinterpret_cast<char *>(read_buf.data()), num_bytes);
+                        ggml_backend_tensor_set(cur, read_buf.data(), 0, num_bytes);
+                    }
+                    data_loaded += num_bytes;
+                    if (progress_callback && total_data_size > 0) {
+                        const float progress = (float)data_loaded / (float)total_data_size;
+                        if (!progress_callback(progress, progress_callback_user_data)) {
+                            throw std::runtime_error(string_format("%s: model loading cancelled by progress_callback\n", __func__));
+                        }
                     }
                 }
+                LOG_DBG("%s: loaded %zu tensors from %s\n", __func__, tensors_to_load.size(), fname.c_str());
+            } else {
+                LOG_DBG("%s: no_alloc is set, skipping tensor data loading (%zu tensors)\n", __func__, tensors_to_load.size());
             }
             fin.close();
-
-            LOG_DBG("%s: loaded %zu tensors from %s\n", __func__, tensors_to_load.size(), fname.c_str());
         }
 
     }

tools/mtmd/mtmd.cpp

@@ -2142,8 +2142,7 @@ std::map<ggml_backend_dev_t, size_t> mtmd_get_memory_usage(const char * mmproj_f
 
     try {
         mtmd_log_set(stub_log_callback, nullptr); // suppress logging
-        // TODO @ngxson : fix no_alloc here
-        ctx.reset(new mtmd_context(mmproj_fname, nullptr, ctx_params));
+        ctx.reset(new mtmd_context(mmproj_fname, nullptr, ctx_params, true));
         mtmd_log_set(saved_log_callback, saved_log_user_data); // restore log callback
         std::map<ggml_backend_dev_t, size_t> total_mem;
         auto merge = [&](const struct clip_ctx * c) {

tools/server/README.md

@@ -1863,11 +1863,15 @@ Example events:
   "data": {
     "status": "loading",
     "progress": {
-      "stage": "fit_params",
-      "value": 0.5 // from 0.0 to 1.0 ; note: not all stages have this "value"
+      "stages": ["text_model", "spec_model", "mmproj_model"],
+      "current": "text_model",
+      "value": 0.5
     }
   }
 }
+// note for "loading" status:
+// - subsequent events will follow the same order of "stages" list
+// - mmap is may report incorrect progress on some platforms; if you need exact progress, use --no-mmap
 
 {
   "model": "...",