Merge branch 'upstream' into concedo_experimental

# Conflicts:
#	.github/workflows/build.yml
#	ggml/CMakeLists.txt
#	ggml/cmake/ggml-config.cmake.in
#	ggml/src/CMakeLists.txt
#	models/templates/README.md
#	tools/imatrix/imatrix.cpp

Author: LostRuins

common/arg.cpp

@@ -26,6 +26,7 @@
 #include <cstdarg>
 #include <filesystem>
 #include <fstream>
+#include <list>
 #include <regex>
 #include <set>
 #include <string>
@@ -2377,20 +2378,35 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
                     }
                     throw std::invalid_argument("unknown buffer type");
                 }
-                // FIXME: this leaks memory
-                params.tensor_buft_overrides.push_back({strdup(tensor_name.c_str()), buft_list.at(buffer_type)});
+                // keep strings alive and avoid leaking memory by storing them in a static vector
+                static std::list<std::string> buft_overrides;
+                buft_overrides.push_back(tensor_name);
+                params.tensor_buft_overrides.push_back({buft_overrides.back().c_str(), buft_list.at(buffer_type)});
             }
         }
     ));
     add_opt(common_arg(
-        {"--cpu-moe"},
-        "use CPU for Mixture of Experts (MoE) weights",
+        {"--cpu-moe", "-cmoe"},
+        "keep all Mixture of Experts (MoE) weights in the CPU",
         [](common_params & params) {
-            params.tensor_buft_overrides.push_back({"\\.ffn_up_exps\\.weight$",   ggml_backend_cpu_buffer_type()});
-            params.tensor_buft_overrides.push_back({"\\.ffn_down_exps\\.weight$", ggml_backend_cpu_buffer_type()});
-            params.tensor_buft_overrides.push_back({"\\.ffn_gate_exps\\.weight$", ggml_backend_cpu_buffer_type()});
+            params.tensor_buft_overrides.push_back({"\\.ffn_(up|down|gate)_exps", ggml_backend_cpu_buffer_type()});
         }
     ).set_env("LLAMA_ARG_CPU_MOE"));
+    add_opt(common_arg(
+        {"--n-cpu-moe", "-ncmoe"}, "N",
+        "keep the Mixture of Experts (MoE) weights of the first N layers in the CPU",
+        [](common_params & params, int value) {
+            if (value < 0) {
+                throw std::invalid_argument("invalid value");
+            }
+            for (int i = 0; i < value; ++i) {
+                // keep strings alive and avoid leaking memory by storing them in a static vector
+                static std::list<std::string> buft_overrides;
+                buft_overrides.push_back(string_format("blk\\.%d\\.ffn_(up|down|gate)_exps", i));
+                params.tensor_buft_overrides.push_back({buft_overrides.back().c_str(), ggml_backend_cpu_buffer_type()});
+            }
+        }
+    ).set_env("LLAMA_ARG_N_CPU_MOE"));
     add_opt(common_arg(
         {"-ngl", "--gpu-layers", "--n-gpu-layers"}, "N",
         "number of layers to store in VRAM",
@@ -2651,10 +2667,10 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
     ).set_examples({LLAMA_EXAMPLE_IMATRIX}));
     add_opt(common_arg(
         {"--output-format"}, "{gguf,dat}",
-        string_format("output format for imatrix file (default: %s)", params.imat_dat ? "dat" : "gguf"),
+        string_format("output format for imatrix file (default: %s)", params.imat_dat > 0 ? "dat" : "gguf"),
         [](common_params & params, const std::string & value) {
-            /**/ if (value == "gguf") { params.imat_dat = false; }
-            else if (value == "dat")  { params.imat_dat = true;  }
+            /**/ if (value == "gguf") { params.imat_dat = -1; }
+            else if (value == "dat")  { params.imat_dat = 1;  }
             else { throw std::invalid_argument("invalid output format"); }
         }
     ).set_examples({LLAMA_EXAMPLE_IMATRIX}));

common/common.h

@@ -435,7 +435,7 @@ struct common_params {
     int32_t n_out_freq  = 10; // output the imatrix every n_out_freq iterations
     int32_t n_save_freq =  0; // save the imatrix every n_save_freq iterations
     int32_t i_chunk     =  0; // start processing from this chunk
-    bool    imat_dat    = false; // whether the legacy imatrix.dat format should be output
+    int8_t  imat_dat    =  0; // whether the legacy imatrix.dat format should be output (gguf <= 0 < dat)
 
     bool process_output  = false; // collect data for the output tensor
     bool compute_ppl     = true;  // whether to compute perplexity

convert_hf_to_gguf.py

@@ -678,6 +678,9 @@ def get_vocab_base_pre(self, tokenizer) -> str:
         if chkhsh == "a1336059768a55c99a734006ffb02203cd450fed003e9a71886c88acf24fdbc2":
             # ref: https://huggingface.co/THUDM/glm-4-9b-hf
             res = "glm4"
+        if chkhsh == "9ca2dd618e8afaf09731a7cf6e2105b373ba6a1821559f258b272fe83e6eb902":
+            # ref: https://huggingface.co/zai-org/GLM-4.5-Air
+            res = "glm4"
         if chkhsh == "1431a23e583c97432bc230bff598d103ddb5a1f89960c8f1d1051aaa944d0b35":
             # ref: https://huggingface.co/sapienzanlp/Minerva-7B-base-v1.0
             res = "minerva-7b"
@@ -6696,6 +6699,139 @@ def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iter
         return super().modify_tensors(data_torch, name, bid)
 
 
+@ModelBase.register("Glm4MoeForCausalLM")
+class Glm4MoeModel(TextModel):
+    model_arch = gguf.MODEL_ARCH.GLM4_MOE
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        # GLM4_MOE has num_hidden_layers + 1 actual layers (including NextN layer)
+        self.block_count = self.hparams["num_hidden_layers"] + self.hparams.get("num_nextn_predict_layers", 0)
+        self.tensor_map = gguf.get_tensor_name_map(self.model_arch, self.block_count)
+
+    def set_vocab(self):
+        from transformers import AutoTokenizer
+
+        tokenizer = AutoTokenizer.from_pretrained(self.dir_model)
+        special_vocab = gguf.SpecialVocab(self.dir_model, load_merges=True)
+        tokens, toktypes, tokpre = self.get_vocab_base()
+        self.gguf_writer.add_tokenizer_model("gpt2")
+        self.gguf_writer.add_tokenizer_pre(tokpre)
+        self.gguf_writer.add_token_list(tokens)
+        self.gguf_writer.add_token_types(toktypes)
+
+        # Special tokens
+        # Note: Using <|endoftext|> (151329) for eot causes endless generation
+        special_vocab._set_special_token("bos", tokenizer.get_added_vocab()["[gMASK]"])  # 151331
+        special_vocab._set_special_token("eot", tokenizer.get_added_vocab()["<|user|>"])  # 151336
+        special_vocab._set_special_token("unk", tokenizer.get_added_vocab()["<|endoftext|>"]) # 151329
+        special_vocab._set_special_token("eom", tokenizer.get_added_vocab()["<|observation|>"])  # 151338
+
+        # Patch broken chat template
+        if isinstance(special_vocab.chat_template, str) and "visible_text(m.content).endswith" in special_vocab.chat_template:
+            special_vocab.chat_template = special_vocab.chat_template.replace(
+                """{{ visible_text(m.content) }}\n{{- '/nothink' if (enable_thinking is defined and not enable_thinking and not visible_text(m.content).endswith("/nothink")) else '' -}}""",
+                """{% set content = visible_text(m.content) %}{{ content }}\n{{- '/nothink' if (enable_thinking is defined and not enable_thinking and not content.endswith("/nothink")) else '' -}}""")
+
+        special_vocab.add_to_gguf(self.gguf_writer)
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        if (rope_dim := self.hparams.get("head_dim")) is None:
+            rope_dim = (
+                self.hparams["hidden_size"] // self.hparams["num_attention_heads"]
+            )
+        self.gguf_writer.add_rope_dimension_count(
+            int(rope_dim * self.hparams.get("partial_rotary_factor", 0.5))
+        )
+
+        # MoE parameters - Use only routed expert count (shared experts handled separately)
+        if (n_routed_experts := self.hparams.get("n_routed_experts")) is not None:
+            self.gguf_writer.add_expert_count(n_routed_experts)
+        if (moe_intermediate_size := self.hparams.get("moe_intermediate_size")) is not None:
+            self.gguf_writer.add_expert_feed_forward_length(moe_intermediate_size)
+        if (n_shared_experts := self.hparams.get("n_shared_experts")) is not None:
+            self.gguf_writer.add_expert_shared_count(n_shared_experts)
+        if (first_k_dense_replace := self.hparams.get("first_k_dense_replace")) is not None:
+            self.gguf_writer.add_leading_dense_block_count(first_k_dense_replace)
+
+        # Expert gating function (sigmoid for GLM4_MOE)
+        self.gguf_writer.add_expert_gating_func(gguf.ExpertGatingFuncType.SIGMOID)
+
+        # Routed scaling factor
+        if (routed_scaling_factor := self.hparams.get("routed_scaling_factor")) is not None:
+            self.gguf_writer.add_expert_weights_scale(routed_scaling_factor)
+
+        # Normalise topk probabilities
+        if (norm_topk_prob := self.hparams.get("norm_topk_prob")) is not None:
+            self.gguf_writer.add_expert_weights_norm(norm_topk_prob)
+
+        # NextN/MTP prediction layers
+        if (num_nextn_predict_layers := self.hparams.get("num_nextn_predict_layers")) is not None:
+            self.gguf_writer.add_nextn_predict_layers(num_nextn_predict_layers)
+
+    _experts: list[dict[str, Tensor]] | None = None
+
+    def modify_tensors(
+        self, data_torch: Tensor, name: str, bid: int | None
+    ) -> Iterable[tuple[str, Tensor]]:
+        if name.startswith("model.visual."):  # ignore visual part
+            return []
+        elif name.startswith("model.language_model."):
+            name = name.replace("language_model.", "")  # for multimodal variants
+
+        # Handle main token embedding (but not layer-specific NextN embeddings)
+        if name == "model.embed_tokens.weight" and ".layers." not in name:
+            return [(self.map_tensor_name("token_embd.weight"), data_torch)]
+
+        # Handle routed experts
+        if name.find("mlp.experts") != -1:
+            n_experts = self.hparams["n_routed_experts"]
+            assert bid is not None
+
+            if self._experts is None:
+                self._experts = [{} for _ in range(self.block_count)]
+
+            self._experts[bid][name] = data_torch
+
+            if len(self._experts[bid]) >= n_experts * 3:
+                tensors: list[tuple[str, Tensor]] = []
+
+                # merge the experts into a single 3d tensor
+                for w_name in ["down_proj", "gate_proj", "up_proj"]:
+                    datas: list[Tensor] = []
+
+                    for xid in range(n_experts):
+                        ename = f"model.layers.{bid}.mlp.experts.{xid}.{w_name}.weight"
+                        datas.append(self._experts[bid][ename])
+                        del self._experts[bid][ename]
+
+                    data_torch = torch.stack(datas, dim=0)
+
+                    merged_name = f"model.layers.{bid}.mlp.experts.{w_name}.weight"
+
+                    new_name = self.map_tensor_name(merged_name)
+                    tensors.append((new_name, data_torch))
+                return tensors
+            else:
+                return []
+
+        if name.endswith("e_score_correction_bias"):
+            name = name.replace("e_score_correction_bias", "e_score_correction.bias")
+
+        new_name = self.map_tensor_name(name)
+
+        return [(new_name, data_torch)]
+
+    def prepare_tensors(self):
+        super().prepare_tensors()
+        if self._experts is not None:
+            # flatten `list[dict[str, Tensor]]` into `list[str]`
+            experts = [k for d in self._experts for k in d.keys()]
+            if len(experts) > 0:
+                raise ValueError(f"Unprocessed experts: {experts}")
+
+
 @ModelBase.register("GlmForCausalLM", "ChatGLMModel", "ChatGLMForConditionalGeneration")
 class ChatGLMModel(TextModel):
     model_arch = gguf.MODEL_ARCH.CHATGLM

convert_hf_to_gguf_update.py

@@ -147,6 +147,7 @@ class TOKENIZER_TYPE(IntEnum):
     {"name": "chatglm-bpe", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/THUDM/glm-4-9b-chat", "chkhsh": "b6e8e1518dc4305be2fe39c313ed643381c4da5db34a98f6a04c093f8afbe99b"},
     {"name": "chatglm-bpe", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/THUDM/glm-4-9b-chat", "chkhsh": "81d72c7348a9f0ebe86f23298d37debe0a5e71149e29bd283904c02262b27516"},
     {"name": "glm4", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/THUDM/glm-4-9b-hf", "chkhsh": "a1336059768a55c99a734006ffb02203cd450fed003e9a71886c88acf24fdbc2"},
+    {"name": "glm4", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/zai-org/GLM-4.5-Air", "chkhsh": "9ca2dd618e8afaf09731a7cf6e2105b373ba6a1821559f258b272fe83e6eb902"},
     {"name": "minerva-7b", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/sapienzanlp/Minerva-7B-base-v1.0", "chkhsh": "1431a23e583c97432bc230bff598d103ddb5a1f89960c8f1d1051aaa944d0b35"},
     {"name": "hunyuan", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/tencent/Hunyuan-A13B-Instruct", "chkhsh": "7e57df22b1fe23a7b1e1c7f3dc4e3f96d43a4eb0836d0c6bdc3436d7b2f1c664"},
     {"name": "hunyuan-dense", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/tencent/Hunyuan-4B-Instruct", "chkhsh": "bba3b3366b646dbdded5dbc42d59598b849371afc42f7beafa914afaa5b70aa6"},

ggml/src/ggml-backend-reg.cpp

@@ -499,6 +499,9 @@ static ggml_backend_reg_t ggml_backend_load_best(const char * name, bool silent,
 
     std::vector<fs::path> search_paths;
     if (user_search_path == nullptr) {
+#ifdef GGML_BACKEND_DIR
+        search_paths.push_back(fs::u8path(GGML_BACKEND_DIR));
+#endif
         // default search paths: executable directory, current directory
         search_paths.push_back(get_executable_path());
         search_paths.push_back(fs::current_path());