feat: Adapter für neue GGUF-Metadata-Keys (QAT + MTP Q4_0 Draft)

Cherry-pick von 673629f4d (archive/cherry-dflash) + Erweiterungen:
- LLM_KV arch_name-Override: nutzt originalen arch_name aus GGUF als
  Key-Prefix (z.B. 'gemma4-assistant' mit Bindestrich)
- Arch-Aliase vereinheitlicht: gemma4-assistant, gemma4_assistant,
  gemma4_mtp alle -> LLM_ARCH_GEMMA4_ASSISTANT
- embedding_length_out als Fallback für n_embd_backbone (upstream PR ggml-org#23398
  nutzt embedding_length_out statt n_embd_backbone)
- n_layer_nextn Default auf n_layer_all falls Key fehlt

Ermöglicht Laden von neuen QAT-Modellen und MTP Q4_0 Draft-Modellen die
mit upstream llama.cpp PR ggml-org#23398 (b9549) konvertiert wurden.

Author: fukuro

src/llama-arch.cpp

@@ -823,8 +823,12 @@ static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = {
 
 LLM_KV::LLM_KV(llm_arch arch, const char * suffix) : arch(arch), suffix(suffix) {}
 
+LLM_KV::LLM_KV(llm_arch arch, const std::string & arch_name, const char * suffix)
+    : arch(arch), arch_name(arch_name), suffix(suffix) {}
+
 std::string LLM_KV::operator()(llm_kv kv) const {
-    std::string name = ::format(LLM_KV_NAMES.at(kv), LLM_ARCH_NAMES.at(arch));
+    const char * prefix = arch_name.empty() ? LLM_ARCH_NAMES.at(arch) : arch_name.c_str();
+    std::string name    = ::format(LLM_KV_NAMES.at(kv), prefix);
 
     if (suffix != nullptr) {
         name += ".";
@@ -875,9 +879,17 @@ llm_arch llm_arch_from_string(const std::string & name) {
         }
     }
 
-    // Aliases for compatibility with older GGUF files
-    if (name == "gemma4-assistant" || name == "gemma4_mtp") {
-        return LLM_ARCH_GEMMA4_ASSISTANT;
+    // Aliases for compatibility with different GGUF converters
+    // - Upstream PR #23398 uses "gemma4-assistant" (hyphen)
+    // - Older converters use "gemma4_assistant" (underscore)
+    static const std::map<std::string, llm_arch> LLM_ARCH_ALIASES = {
+        { "gemma4-assistant",  LLM_ARCH_GEMMA4_ASSISTANT },
+        { "gemma4_assistant",  LLM_ARCH_GEMMA4_ASSISTANT },
+        { "gemma4_mtp",        LLM_ARCH_GEMMA4_ASSISTANT },
+    };
+    auto it = LLM_ARCH_ALIASES.find(name);
+    if (it != LLM_ARCH_ALIASES.end()) {
+        return it->second;
     }
 
     return LLM_ARCH_UNKNOWN;

src/llama-arch.h

@@ -604,8 +604,10 @@ enum llm_tensor_layer {
 
 struct LLM_KV {
     LLM_KV(llm_arch arch, const char * suffix = nullptr);
+    LLM_KV(llm_arch arch, const std::string & arch_name, const char * suffix = nullptr);
 
-    llm_arch arch;
+    llm_arch   arch;
+    std::string arch_name; // override key prefix (empty = use LLM_ARCH_NAMES.at(arch))
     const char * suffix;
 
     std::string operator()(llm_kv kv) const;

src/llama-model-loader.cpp

@@ -553,7 +553,7 @@ llama_model_loader::llama_model_loader(
         }
 
         get_key(llm_kv(LLM_KV_GENERAL_ARCHITECTURE), arch_name, false);
-        llm_kv = LLM_KV(llm_arch_from_string(arch_name));
+        llm_kv = LLM_KV(llm_arch_from_string(arch_name), arch_name);
 
         files.emplace_back(new llama_file(fname.c_str(), "rb", use_direct_io));
         contexts.emplace_back(ctx);
@@ -679,7 +679,7 @@ llama_model_loader::llama_model_loader(
         }
 
         get_key(llm_kv(LLM_KV_GENERAL_ARCHITECTURE), arch_name, false);
-        llm_kv = LLM_KV(llm_arch_from_string(arch_name));
+        llm_kv = LLM_KV(llm_arch_from_string(arch_name), arch_name);
 
         files.emplace_back(new llama_file(file));
         contexts.emplace_back(ctx);
@@ -697,7 +697,7 @@ llama_model_loader::llama_model_loader(
         }
     } else {
         get_key(llm_kv(LLM_KV_GENERAL_ARCHITECTURE), arch_name, false);
-        llm_kv = LLM_KV(llm_arch_from_string(arch_name));
+        llm_kv = LLM_KV(llm_arch_from_string(arch_name), arch_name);
     }
 
     n_kv      = gguf_get_n_kv(metadata);

src/models/gemma4-assistant.cpp

@@ -4,8 +4,15 @@
 
 void llama_model_gemma4_assistant::load_arch_hparams(llama_model_loader & ml) {
     // Read backbone embedding length (target model hidden size)
-    // Older GGUF files use n_embd_backbone, newer ones use embedding_length_out
+    // Three sources, in priority order:
+    //   1. n_embd_backbone (older GGUF converters, e.g. TheTom's)
+    //   2. embedding_length_out (upstream PR #23398, newer converters)
+    //   3. n_embd (fallback — wrong for MTP but avoids crash)
     ml.get_key(LLM_KV_EMBEDDING_LENGTH_BACKBONE, hparams.n_embd_backbone, false);
+    if (hparams.n_embd_backbone == 0) {
+        // n_embd_out_impl was already set from embedding_length_out in llama-model.cpp
+        hparams.n_embd_backbone = hparams.n_embd_out_impl;
+    }
     if (hparams.n_embd_backbone == 0) {
         hparams.n_embd_backbone = hparams.n_embd;
     }