llama : add llama_arch_* query API and refactor test-llama-archs to use it

src/llama-arch.cpp

@@ -1,5 +1,6 @@
 #include "llama-arch.h"
 
+#include "llama-ext.h"
 #include "llama-impl.h"
 
 #include <map>
@@ -1348,3 +1349,43 @@ bool llm_arch_supports_sm_tensor(const llm_arch & arch) {
             return true;
     }
 }
+
+bool llama_arch_supports_save(llm_arch arch) {
+    switch (arch) {
+        case LLM_ARCH_SENSENOVA_U1:
+        case LLM_ARCH_GEMMA3N:
+        case LLM_ARCH_BITNET:
+        case LLM_ARCH_T5:
+        case LLM_ARCH_APERTUS:
+        case LLM_ARCH_STEP35:
+        case LLM_ARCH_CLEF: // the head tensors are not saved
+        case LLM_ARCH_VLLM_DECISION2:
+            return false;
+        default:
+            return true;
+    }
+}
+
+std::vector<llm_arch> llama_arch_all() {
+    return llm_arch_all();
+}
+
+const char * llama_arch_name(llm_arch arch) {
+    return llm_arch_name(arch);
+}
+
+llm_arch llama_arch_from_string(const std::string & name) {
+    return llm_arch_from_string(name);
+}
+
+std::string llama_arch_kv_name(llm_arch arch, llm_kv key) {
+    return LLM_KV(arch)(key);
+}
+
+bool llama_arch_supports_sm_tensor(llm_arch arch) {
+    return llm_arch_supports_sm_tensor(arch);
+}
+
+bool llama_arch_supports_mixed_batch(llm_arch arch) {
+    return llm_arch_supports_mixed_batch(arch);
+}

src/llama-ext.h

@@ -5,11 +5,36 @@
 // try as much as possible to not include this header in the rest of the codebase
 
 #include "llama.h"
+#include "llama-arch.h"
 
 #include <cstdint>
 #include <map>
+#include <string>
 #include <vector>
 
+//
+// llama architecture queries
+//
+
+// all the LLM architectures known to the library
+LLAMA_API std::vector<llm_arch> llama_arch_all();
+
+// the canonical name of an architecture (e.g. "llama"), or "unknown"
+LLAMA_API const char * llama_arch_name(llm_arch arch);
+
+// the architecture with the given canonical name, or LLM_ARCH_UNKNOWN
+LLAMA_API llm_arch llama_arch_from_string(const std::string & name);
+
+// the arch-resolved GGUF key of a KV id, e.g. LLM_KV_CONTEXT_LENGTH -> "llama.context_length"
+LLAMA_API std::string llama_arch_kv_name(llm_arch arch, llm_kv key);
+
+LLAMA_API bool llama_arch_supports_sm_tensor(llm_arch arch);
+
+LLAMA_API bool llama_arch_supports_mixed_batch(llm_arch arch);
+
+// FIXME temporary function for better error messages
+LLAMA_API bool llama_arch_supports_save(llm_arch arch);
+
 // Reserve a new compute graph. It is valid until the next call to llama_graph_reserve.
 LLAMA_API struct ggml_cgraph * llama_graph_reserve(
         struct llama_context * ctx,

src/llama-model-saver.cpp

@@ -4,6 +4,7 @@
 #include "gguf.h"
 
 #include "llama-arch.h"
+#include "llama-ext.h"
 #include "llama.h"
 #include "llama-hparams.h"
 #include "llama-model.h"
@@ -14,34 +15,13 @@
 #include <string>
 #include <unordered_set>
 
-bool llama_model_saver_supports_arch(llm_arch arch) {
-    switch (arch) {
-        case LLM_ARCH_SENSENOVA_U1:
-        case LLM_ARCH_GEMMA3N:
-        case LLM_ARCH_BITNET:
-        case LLM_ARCH_T5:
-        case LLM_ARCH_APERTUS:
-        case LLM_ARCH_STEP35:
-        case LLM_ARCH_CLEF: // the head tensors are not saved
-        case LLM_ARCH_VLLM_DECISION2:
-            return false;
-        default:
-            return true;
-    }
-}
-
 llama_model_saver::llama_model_saver(const struct llama_model * model) :
-        gguf_ctx(gguf_init_empty()), gguf_ctx_owned(true), model(model), llm_kv(model->arch) {
-    GGML_ASSERT(llama_model_saver_supports_arch(model->arch));
+        gguf_ctx(gguf_init_empty()), model(model), llm_kv(model->arch) {
+    GGML_ASSERT(llama_arch_supports_save(model->arch));
 }
 
-llama_model_saver::llama_model_saver(enum llm_arch arch, struct gguf_context * gguf_ctx) :
-        gguf_ctx(gguf_ctx == nullptr ? gguf_init_empty() : gguf_ctx), gguf_ctx_owned(gguf_ctx == nullptr), model(nullptr), llm_kv(arch) {}
-
 llama_model_saver::~llama_model_saver() {
-    if (gguf_ctx_owned) {
-        gguf_free(gguf_ctx);
-    }
+    gguf_free(gguf_ctx);
 }
 
 void llama_model_saver::add_kv(const enum llm_kv key, const uint32_t value) {
@@ -119,11 +99,6 @@ void llama_model_saver::add_kv(const enum llm_kv key, const Container & value, c
         GGML_ABORT("fatal error");
     }
 }
-// instantiate for external usage:
-template void llama_model_saver::add_kv<std::vector<uint32_t>>(const enum llm_kv, const std::vector<uint32_t> &, const bool);
-template void llama_model_saver::add_kv<std::vector<float>>(const enum llm_kv, const std::vector<float> &, const bool);
-template void llama_model_saver::add_kv<std::vector<uint64_t>>(const enum llm_kv, const std::vector<uint64_t> &, const bool);
-template void llama_model_saver::add_kv<std::vector<int32_t>>(const enum llm_kv, const std::vector<int32_t> &, const bool);
 
 void llama_model_saver::add_kv(const enum llm_kv key, const std::vector<std::string> & value) {
     std::vector<const char *> tmp(value.size());
@@ -551,7 +526,3 @@ void llama_model_saver::add_tensors_from_model() {
 void llama_model_saver::save(const std::string & path_model) {
     gguf_write_to_file(gguf_ctx, path_model.c_str(), false);
 }
-
-void llama_model_saver::save(FILE * file) {
-    gguf_write_to_file_ptr(gguf_ctx, file, false);
-}

src/llama-model-saver.h

@@ -6,17 +6,12 @@
 
 #include <vector>
 
-// FIXME temporary function for better error messages
-bool llama_model_saver_supports_arch(llm_arch arch);
-
 struct llama_model_saver {
     struct gguf_context * gguf_ctx = nullptr;
-    const bool gguf_ctx_owned;
     const struct llama_model * model;
     const struct LLM_KV llm_kv;
 
     llama_model_saver(const struct llama_model * model);
-    llama_model_saver(enum llm_arch arch, struct gguf_context * gguf_ctx);
     ~llama_model_saver();
 
     void add_kv(enum llm_kv key, uint32_t     value);
@@ -41,5 +36,4 @@ struct llama_model_saver {
     void add_tensors_from_model();
 
     void save(const std::string & path_model);
-    void save(FILE * file);
 };

tests/CMakeLists.txt

@@ -223,57 +223,56 @@ if (NOT WIN32 OR NOT BUILD_SHARED_LIBS)
 
     # llama_build_and_test(test-double-float.cpp) # SLOW
 
-    # TODO: after `test-llama-archs` is refactored to not link internal symbols, we can move all related tests out of
-    #       this branch and run them on all platforms
-    llama_build(test-llama-archs.cpp)
-
-    set(MODEL_DIR "${CMAKE_CURRENT_BINARY_DIR}/test-models/")
-    file(MAKE_DIRECTORY "${MODEL_DIR}")
-
-    llama_test(
-        test-llama-archs
-        NAME test-generate-models
-        LABEL "main behavior backend"
-        ARGS -o "${MODEL_DIR}"
-    )
-    set_tests_properties(test-generate-models PROPERTIES
-        FIXTURES_SETUP generate-models
-    )
+endif()
 
-    # Target model for test-speculative-rejection-sampling
-    set(SPEC_TARGET_DIR "${CMAKE_CURRENT_BINARY_DIR}/test-speculative-models/target")
-    file(MAKE_DIRECTORY "${SPEC_TARGET_DIR}")
-    llama_test(
-        test-llama-archs
-        NAME test-generate-speculative-target
-        LABEL main
-        ARGS -a llama -s 999 -o "${SPEC_TARGET_DIR}"
-    )
-    set_tests_properties(
-        test-generate-speculative-target
-        PROPERTIES FIXTURES_SETUP speculative-model
-    )
+llama_build(test-llama-archs.cpp)
 
-    # To avoid recreating the model use the fixture exclude any option:
-    # -FA "^speculative-model$"
-    llama_build_and_test(test-speculative-rejection-sampling.cpp ARGS
-        -m "${SPEC_TARGET_DIR}/llama-dense.gguf")
-    set_tests_properties(test-speculative-rejection-sampling PROPERTIES FIXTURES_REQUIRED speculative-model)
-
-    # Test context state handling across all architectures, using the generated dummy models
-    llama_test(
-        test-llama-context
-        LABEL "main behavior backend"
-        ARGS --models "${MODEL_DIR}"
-    )
-    # abort on graph reallocations that the initial reserve should have covered
-    set_tests_properties(test-llama-context PROPERTIES
-        FIXTURES_REQUIRED generate-models
-        ENVIRONMENT "GGML_SCHED_DEBUG_REALLOC=1"
-    )
+set(MODEL_DIR "${CMAKE_CURRENT_BINARY_DIR}/test-models/")
+file(MAKE_DIRECTORY "${MODEL_DIR}")
 
-    llama_build(test-fusion.cpp)
-endif()
+llama_test(
+    test-llama-archs
+    NAME test-generate-models
+    LABEL "main behavior backend"
+    ARGS -o "${MODEL_DIR}"
+)
+set_tests_properties(test-generate-models PROPERTIES
+    FIXTURES_SETUP generate-models
+)
+
+# Target model for test-speculative-rejection-sampling
+set(SPEC_TARGET_DIR "${CMAKE_CURRENT_BINARY_DIR}/test-speculative-models/target")
+file(MAKE_DIRECTORY "${SPEC_TARGET_DIR}")
+llama_test(
+    test-llama-archs
+    NAME test-generate-speculative-target
+    LABEL main
+    ARGS -a llama -s 999 -o "${SPEC_TARGET_DIR}"
+)
+set_tests_properties(
+    test-generate-speculative-target
+    PROPERTIES FIXTURES_SETUP speculative-model
+)
+
+# To avoid recreating the model use the fixture exclude any option:
+# -FA "^speculative-model$"
+llama_build_and_test(test-speculative-rejection-sampling.cpp ARGS
+    -m "${SPEC_TARGET_DIR}/llama-dense.gguf")
+set_tests_properties(test-speculative-rejection-sampling PROPERTIES FIXTURES_REQUIRED speculative-model)
+
+# Test context state handling across all architectures, using the generated dummy models
+llama_test(
+    test-llama-context
+    LABEL "main behavior backend"
+    ARGS --models "${MODEL_DIR}"
+)
+# abort on graph reallocations that the initial reserve should have covered
+set_tests_properties(test-llama-context PROPERTIES
+    FIXTURES_REQUIRED generate-models
+    ENVIRONMENT "GGML_SCHED_DEBUG_REALLOC=1"
+)
+
+llama_build(test-fusion.cpp)
 
 llama_build_and_test(test-chat-peg-parser.cpp peg-parser/simple-tokenize.cpp  LABEL behavior)
 llama_build_and_test(test-jinja.cpp  LABEL behavior)