Merge branch 'master' into Nexes_PR_Quants

Author: Nexesenex

.github/workflows/bench.yml renamed to .github/workflows/bench.yml.disabled

@@ -1,3 +1,6 @@
+# TODO: there have been some issues with the workflow, so disabling for now
+#       https://github.com/ggerganov/llama.cpp/issues/7893
+#
 # Benchmark
 name: Benchmark
 

README.md

@@ -105,6 +105,7 @@ Typically finetunes of the base models below are supported as well.
 - [x] [Open Elm models](https://huggingface.co/collections/apple/openelm-instruct-models-6619ad295d7ae9f868b759ca)
 - [x] [ChatGLM3-6b](https://huggingface.co/THUDM/chatglm3-6b) + [ChatGLM4-9b](https://huggingface.co/THUDM/glm-4-9b)
 - [x] [SmolLM](https://huggingface.co/collections/HuggingFaceTB/smollm-6695016cad7167254ce15966)
+- [x] [EXAONE-3.0-7.8B-Instruct](https://huggingface.co/LGAI-EXAONE/EXAONE-3.0-7.8B-Instruct)
 
 (instructions for supporting more models: [HOWTO-add-model.md](./docs/development/HOWTO-add-model.md))
 

common/common.cpp

@@ -110,8 +110,34 @@ int32_t cpu_get_num_physical_cores() {
     if (result == 0) {
         return num_physical_cores;
     }
-#elif defined(_WIN32)
-    //TODO: Implement
+#elif defined(_WIN32) && (_WIN32_WINNT >= 0x0601) && !defined(__MINGW64__) // windows 7 and later
+    // TODO: windows + arm64 + mingw64
+    unsigned int n_threads_win = std::thread::hardware_concurrency();
+    unsigned int default_threads = n_threads_win > 0 ? (n_threads_win <= 4 ? n_threads_win : n_threads_win / 2) : 4;
+
+    DWORD buffer_size = 0;
+    if (!GetLogicalProcessorInformationEx(RelationProcessorCore, nullptr, &buffer_size)) {
+        if (GetLastError() != ERROR_INSUFFICIENT_BUFFER) {
+            return default_threads;
+        }
+    }
+
+    std::vector<char> buffer(buffer_size);
+    if (!GetLogicalProcessorInformationEx(RelationProcessorCore, reinterpret_cast<PSYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX>(buffer.data()), &buffer_size)) {
+        return default_threads;
+    }
+
+    int32_t num_physical_cores = 0;
+    PSYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX info = reinterpret_cast<PSYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX>(buffer.data());
+    while (buffer_size > 0) {
+        if (info->Relationship == RelationProcessorCore) {
+            num_physical_cores += info->Processor.GroupCount;
+        }
+        buffer_size -= info->Size;
+        info = reinterpret_cast<PSYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX>(reinterpret_cast<char*>(info) + info->Size);
+    }
+
+    return num_physical_cores > 0 ? num_physical_cores : default_threads;
 #endif
     unsigned int n_threads = std::thread::hardware_concurrency();
     return n_threads > 0 ? (n_threads <= 4 ? n_threads : n_threads / 2) : 4;
@@ -1735,7 +1761,13 @@ std::string gpt_params_get_system_info(const gpt_params & params) {
     if (params.n_threads_batch != -1) {
         os << " (n_threads_batch = " << params.n_threads_batch << ")";
     }
+#if defined(_WIN32) && (_WIN32_WINNT >= 0x0601) && !defined(__MINGW64__) // windows 7 and later
+    // TODO: windows + arm64 + mingw64
+    DWORD logicalProcessorCount = GetActiveProcessorCount(ALL_PROCESSOR_GROUPS);
+    os << " / " << logicalProcessorCount << " | " << llama_print_system_info();
+#else
     os << " / " << std::thread::hardware_concurrency() << " | " << llama_print_system_info();
+#endif
 
     return os.str();
 }
@@ -2710,12 +2742,6 @@ std::string llama_detokenize(llama_context * ctx, const std::vector<llama_token>
     return text;
 }
 
-bool llama_should_add_bos_token(const llama_model * model) {
-    const int add_bos = llama_add_bos_token(model);
-
-    return add_bos != -1 ? bool(add_bos) : (llama_vocab_type(model) == LLAMA_VOCAB_TYPE_SPM);
-}
-
 //
 // Chat template utils
 //

common/common.h

@@ -382,10 +382,6 @@ std::string llama_detokenize(
         const std::vector<llama_token> & tokens,
                                   bool   special = true);
 
-// Uses the value from the model metadata if possible, otherwise
-// defaults to true when model type is SPM, otherwise false.
-bool llama_should_add_bos_token(const llama_model * model);
-
 //
 // Chat template utils
 //

convert_hf_to_gguf.py

@@ -590,6 +590,15 @@ def get_vocab_base_pre(self, tokenizer) -> str:
         if chkhsh == "855059429035d75a914d1eda9f10a876752e281a054a7a3d421ef0533e5b6249":
             # ref: https://huggingface.co/HuggingFaceTB/SmolLM-135M
             res = "smollm"
+        if chkhsh == "3c30d3ad1d6b64202cd222813e7736c2db6e1bd6d67197090fc1211fbc612ae7":
+            # ref: https://huggingface.co/bigscience/bloom
+            res = "bloom"
+        if chkhsh == "bc01ce58980e1db43859146dc51b1758b3b88729b217a74792e9f8d43e479d21":
+            # ref: https://huggingface.co/TurkuNLP/gpt3-finnish-small
+            res = "gpt3-finnish"
+        if chkhsh == "4e2b24cc4770243d65a2c9ec19770a72f08cffc161adbb73fcbb6b7dd45a0aae":
+            # ref: https://huggingface.co/LGAI-EXAONE/EXAONE-3.0-7.8B-Instruct
+            res = "exaone"
 
         if res is None:
             logger.warning("\n")
@@ -893,7 +902,7 @@ def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iter
         return tensors
 
 
-@Model.register("BloomForCausalLM")
+@Model.register("BloomForCausalLM", "BloomModel")
 class BloomModel(Model):
     model_arch = gguf.MODEL_ARCH.BLOOM
 
@@ -3734,6 +3743,118 @@ def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iter
         name = name.removeprefix("transformer.")
         return [(self.map_tensor_name(name), data_torch)]
 
+
+@Model.register("NemotronForCausalLM")
+class NemotronModel(Model):
+    model_arch = gguf.MODEL_ARCH.NEMOTRON
+
+    def set_vocab(self):
+        self._set_vocab_sentencepiece()
+        self.gguf_writer.add_pad_token_id(0)
+        self.gguf_writer.add_unk_token_id(1)
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        hparams = self.hparams
+        self.gguf_writer.add_vocab_size(hparams["vocab_size"])
+
+        f_norm_eps = self.find_hparam(["layer_norm_eps", "layer_norm_epsilon", "norm_epsilon", "norm_eps"])
+        self.gguf_writer.add_layer_norm_eps(f_norm_eps)
+
+        # * Partial RoPE
+        rot_pct = self.find_hparam(["partial_rotary_factor", "rope_pct", "rope_percent"])
+        n_embd = self.find_hparam(["hidden_size", "n_embd"])
+        n_head = self.find_hparam(["num_attention_heads", "n_head"])
+        self.gguf_writer.add_rope_dimension_count(int(rot_pct * n_embd) // n_head)
+
+        # * RopeScaling for Nemotron
+        if "rope_scaling" not in self.hparams or self.hparams["rope_scaling"] is None:
+            self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.NONE)
+        else:
+            self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LINEAR)
+            self.gguf_writer.add_rope_scaling_factor(self.hparams["factor"])
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        # * Adding +1 to LayerNorm's weights here to implement layernorm1p w/o changing anything on the GGML engine side
+        #   model.layers.{l}.input_layernorm.weight
+        #   model.layers.{l}.post_attention_layernorm.weight
+        #   model.norm.weight
+        if name.endswith("norm.weight"):
+            data_torch = data_torch + 1
+
+        return [(self.map_tensor_name(name), data_torch)]
+
+
+@Model.register("ExaoneForCausalLM")
+class ExaoneModel(Model):
+    model_arch = gguf.MODEL_ARCH.EXAONE
+
+    def set_gguf_parameters(self):
+        hparams = self.hparams
+
+        assert(hparams["activation_function"] == "silu")
+
+        max_position_embeddings = hparams["max_position_embeddings"]
+        embed_dim = hparams["hidden_size"]
+        num_heads = hparams["num_attention_heads"]
+        num_kv_heads = hparams.get("num_key_value_heads", num_heads)
+        layer_norm_eps = hparams["layer_norm_epsilon"]
+        intermediate_size = hparams["intermediate_size"] if "intermediate_size" in hparams else 4 * embed_dim
+        num_layers = hparams["num_layers"]
+        # ignore for now as EXAONE-3.0-7.8B-Instruct attentino_dropout is 0.0
+        # attention_dropout_rate = hparams["attention_dropout"]
+        # ignore for now as EXAONE-3.0-7.8B-Instruct embed_dropout is 0.0
+        # embed_dropout_rate = hparams["embed_dropout"]
+        self.gguf_writer.add_embedding_length(embed_dim)
+        self.gguf_writer.add_head_count(num_heads)
+        self.gguf_writer.add_head_count_kv(num_kv_heads)
+        self.gguf_writer.add_context_length(max_position_embeddings)
+        self.gguf_writer.add_layer_norm_rms_eps(layer_norm_eps)
+        self.gguf_writer.add_feed_forward_length(intermediate_size)
+        self.gguf_writer.add_block_count(num_layers)
+        self.gguf_writer.add_file_type(self.ftype)
+
+        if (rope_theta := self.hparams.get("rope_theta")) is not None:
+            self.gguf_writer.add_rope_freq_base(rope_theta)
+        rotary_factor = self.find_hparam(["partial_rotary_factor", "rope_pct"], optional=True)
+        rotary_factor = rotary_factor if rotary_factor is not None else 1.0
+        self.gguf_writer.add_rope_dimension_count(int(rotary_factor * (hparams["hidden_size"] // hparams["num_attention_heads"])))
+        if hparams.get("rope_scaling") is not None and "factor" in hparams["rope_scaling"]:
+            if hparams["rope_scaling"].get("type") == "linear":
+                self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LINEAR)
+                self.gguf_writer.add_rope_scaling_factor(hparams["rope_scaling"]["factor"])
+
+    def prepare_tensors(self):
+        if rope_scaling := self.find_hparam(["rope_scaling"], optional=True):
+            if rope_scaling.get("rope_type", '').lower() == "llama3":
+                base = self.hparams.get("rope_theta", 10000.0)
+                dim = self.hparams["hidden_size"] // self.hparams["num_attention_heads"]
+                freqs = 1.0 / (base ** (torch.arange(0, dim, 2, dtype=torch.float32) / dim))
+
+                factor = rope_scaling.get("factor", 8.0)
+                low_freq_factor = rope_scaling.get("low_freq_factor", 1.0)
+                high_freq_factor = rope_scaling.get("high_freq_factor", 4.0)
+                old_context_len = self.hparams.get("original_max_position_embeddings", 8192)
+
+                low_freq_wavelen = old_context_len / low_freq_factor
+                high_freq_wavelen = old_context_len / high_freq_factor
+                assert low_freq_wavelen != high_freq_wavelen
+
+                rope_factors = []
+                for freq in freqs:
+                    wavelen = 2 * math.pi / freq
+                    if wavelen < high_freq_wavelen:
+                        rope_factors.append(1)
+                    elif wavelen > low_freq_wavelen:
+                        rope_factors.append(factor)
+                    else:
+                        smooth = (old_context_len / wavelen - low_freq_factor) / (high_freq_factor - low_freq_factor)
+                        rope_factors.append(1 / ((1 - smooth) / factor + smooth))
+
+                self.gguf_writer.add_tensor(self.format_tensor_name(gguf.MODEL_TENSOR.ROPE_FREQS), np.array(rope_factors, dtype=np.float32))
+
+        super().prepare_tensors()
+
 ###### CONVERSION LOGIC ######
 
 

convert_hf_to_gguf_update.py

@@ -94,6 +94,9 @@ class TOKENIZER_TYPE(IntEnum):
     {"name": "codeshell",      "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/WisdomShell/CodeShell-7B", },
     {"name": "tekken",         "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/mistralai/Mistral-Nemo-Base-2407", },
     {"name": "smollm",         "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/HuggingFaceTB/SmolLM-135M", },
+    {'name': "bloom",          "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/bigscience/bloom", },
+    {'name': "gpt3-finnish",   "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/TurkuNLP/gpt3-finnish-small", },
+    {"name": "exaone",         "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/LGAI-EXAONE/EXAONE-3.0-7.8B-Instruct", },
 ]
 
 

examples/cvector-generator/cvector-generator.cpp

@@ -271,7 +271,7 @@ struct tokenized_prompt {
     size_t max_seq_len;
 
     tokenized_prompt(llama_context * ctx, std::string pos, std::string neg) {
-        const bool add_bos = llama_should_add_bos_token(llama_get_model(ctx));
+        const bool add_bos = llama_add_bos_token(llama_get_model(ctx));
         tokens_pos = ::llama_tokenize(ctx, pos, add_bos, true);
         tokens_neg = ::llama_tokenize(ctx, neg, add_bos, true);
         max_seq_len = std::max(tokens_pos.size(), tokens_neg.size());

examples/eval-callback/eval-callback.cpp

@@ -127,7 +127,7 @@ static bool ggml_debug(struct ggml_tensor * t, bool ask, void * user_data) {
 }
 
 static bool run(llama_context * ctx, const gpt_params & params) {
-    const bool add_bos = llama_should_add_bos_token(llama_get_model(ctx));
+    const bool add_bos = llama_add_bos_token(llama_get_model(ctx));
 
     std::vector<llama_token> tokens = ::llama_tokenize(ctx, params.prompt, add_bos);
 

examples/imatrix/imatrix.cpp

@@ -434,8 +434,8 @@ static void process_logits(
 }
 
 static bool compute_imatrix(llama_context * ctx, const gpt_params & params) {
-    const bool add_bos = llama_should_add_bos_token(llama_get_model(ctx));
-    GGML_ASSERT(llama_add_eos_token(llama_get_model(ctx)) != 1);
+    const bool add_bos = llama_add_bos_token(llama_get_model(ctx));
+    GGML_ASSERT(!llama_add_eos_token(llama_get_model(ctx)));
     const int n_ctx = llama_n_ctx(ctx);
 
     auto tim1 = std::chrono::high_resolution_clock::now();

examples/infill/infill.cpp

@@ -203,8 +203,8 @@ int main(int argc, char ** argv) {
         LOG_TEE("\n");
         LOG_TEE("%s\n", gpt_params_get_system_info(params).c_str());
     }
-    const bool add_bos = llama_should_add_bos_token(model);
-    GGML_ASSERT(llama_add_eos_token(model) != 1);
+    const bool add_bos = llama_add_bos_token(model);
+    GGML_ASSERT(!llama_add_eos_token(model));
     LOG("add_bos: %d\n", add_bos);
 
     std::vector<llama_token> embd_inp;