Merge 24bad77ebf into 9ba399dfa7

server : add support for "encoding_format": "base64" to the */embeddings endpoints (#10967 )
* add support for base64 * fix base64 test * improve test --------- Co-authored-by: Xuan Son Nguyen <son@huggingface.co>
2024-12-25 10:54:36 +00:00 · 2024-12-24 12:55:08 -08:00 · 2024-12-24 21:33:04 +01:00 · 2024-12-24 18:54:49 +01:00 · 2024-12-24 17:39:49 +01:00 · 2024-12-24 09:44:20 +02:00
27 changed files with 1111 additions and 405 deletions
--- a/README.md
+++ b/README.md
@ -91,7 +91,7 @@ Instructions for adding support for new models: [HOWTO-add-model.md](docs/develo
 - [x] [Bitnet b1.58 models](https://huggingface.co/1bitLLM)
 - [x] [Flan T5](https://huggingface.co/models?search=flan-t5)
 - [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] [ChatGLM3-6b](https://huggingface.co/THUDM/chatglm3-6b) + [ChatGLM4-9b](https://huggingface.co/THUDM/glm-4-9b) + [GLMEdge-1.5b](https://huggingface.co/THUDM/glm-edge-1.5b-chat) + [GLMEdge-4b](https://huggingface.co/THUDM/glm-edge-4b-chat)
 - [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)
 - [x] [FalconMamba Models](https://huggingface.co/collections/tiiuae/falconmamba-7b-66b9a580324dd1598b0f6d4a)
@ -111,6 +111,7 @@ Instructions for adding support for new models: [HOWTO-add-model.md](docs/develo
 - [x] [Mini CPM](https://huggingface.co/models?search=MiniCPM)
 - [x] [Moondream](https://huggingface.co/vikhyatk/moondream2)
 - [x] [Bunny](https://github.com/BAAI-DCAI/Bunny)
+- [x] [GLM-EDGE](https://huggingface.co/models?search=glm-edge)
 - [x] [Qwen2-VL](https://huggingface.co/collections/Qwen/qwen2-vl-66cee7455501d7126940800d)

 </details>
--- a/convert_hf_to_gguf.py
+++ b/convert_hf_to_gguf.py
@ -642,7 +642,7 @@ class Model:
        if chkhsh == "7967bfa498ade6b757b064f31e964dddbb80f8f9a4d68d4ba7998fcf281c531a":
            # ref: https://huggingface.co/jinaai/jina-embeddings-v2-base-code
            res = "jina-v2-code"
-        if chkhsh == "b6e8e1518dc4305be2fe39c313ed643381c4da5db34a98f6a04c093f8afbe99b":
+        if chkhsh == "b6e8e1518dc4305be2fe39c313ed643381c4da5db34a98f6a04c093f8afbe99b" or chkhsh == "81d72c7348a9f0ebe86f23298d37debe0a5e71149e29bd283904c02262b27516":
            # ref: https://huggingface.co/THUDM/glm-4-9b-chat
            res = "chatglm-bpe"
        if chkhsh == "7fc505bd3104ca1083b150b17d088b59534ede9bde81f0dd2090967d7fe52cee":
@ -4280,7 +4280,7 @@ class JaisModel(Model):
        self.gguf_writer.add_max_alibi_bias(self.max_alibi_bias)


-@Model.register("ChatGLMModel", "ChatGLMForConditionalGeneration")
+@Model.register("GlmForCausalLM", "ChatGLMModel", "ChatGLMForConditionalGeneration")
 class ChatGLMModel(Model):
    model_arch = gguf.MODEL_ARCH.CHATGLM

@ -4386,47 +4386,56 @@ class ChatGLMModel(Model):

        from transformers import AutoTokenizer
        tokenizer = AutoTokenizer.from_pretrained(dir_model, trust_remote_code=True)
-        vocab_size = hparams["padded_vocab_size"]
+        vocab_size = hparams.get("padded_vocab_size",hparams["vocab_size"])
        assert max(tokenizer.get_vocab().values()) < vocab_size

-        tokpre = self.get_vocab_base_pre(tokenizer)
+        if(hparams["partial_rotary_factor"] == 1.0):
+            # only for glm-edge series
+            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_vocab = gguf.SpecialVocab(self.dir_model, load_merges=True)
+        else:
+            # for glm4 series
+            tokpre = self.get_vocab_base_pre(tokenizer)
+            merges = []
+            vocab = {}
+            mergeable_ranks = tokenizer._mergeable_ranks
+            for token, rank in mergeable_ranks.items():
+                vocab[ChatGLMModel.token_bytes_to_string(token)] = rank
+                if len(token) == 1:
+                    continue
+                merged = ChatGLMModel.bpe(mergeable_ranks, token, max_rank=rank)
+                assert len(merged) >= 2 and len(merged) <= 7
+                merges.append(' '.join(map(ChatGLMModel.token_bytes_to_string, merged)))

-        merges = []
-        vocab = {}
-        mergeable_ranks = tokenizer.mergeable_ranks
-        for token, rank in mergeable_ranks.items():
-            vocab[ChatGLMModel.token_bytes_to_string(token)] = rank
-            if len(token) == 1:
-                continue
-            merged = ChatGLMModel.bpe(mergeable_ranks, token, max_rank=rank)
-            assert len(merged) >= 2 and len(merged) <= 7
-            merges.append(' '.join(map(ChatGLMModel.token_bytes_to_string, merged)))
+            # for this kind of tokenizer, added_vocab is not a subset of vocab, so they need to be combined
+            added_vocab = tokenizer.get_added_vocab()
+            reverse_vocab = {id_ : encoded_tok for encoded_tok, id_ in {**vocab, **added_vocab}.items()}

-        # for this kind of tokenizer, added_vocab is not a subset of vocab, so they need to be combined
-        added_vocab = tokenizer.get_added_vocab()
-        reverse_vocab = {id_ : encoded_tok for encoded_tok, id_ in {**vocab, **added_vocab}.items()}
-
-        for i in range(vocab_size):
-            if i not in reverse_vocab:
-                tokens.append(f"[PAD{i}]")
-                toktypes.append(gguf.TokenType.UNUSED)
-            elif reverse_vocab[i] in added_vocab:
-                tokens.append(reverse_vocab[i])
-                if tokenizer.added_tokens_decoder[i].special:
-                    toktypes.append(gguf.TokenType.CONTROL)
+            for i in range(vocab_size):
+                if i not in reverse_vocab:
+                    tokens.append(f"[PAD{i}]")
+                    toktypes.append(gguf.TokenType.UNUSED)
+                elif reverse_vocab[i] in added_vocab:
+                    tokens.append(reverse_vocab[i])
+                    if tokenizer.added_tokens_decoder[i].special:
+                        toktypes.append(gguf.TokenType.CONTROL)
+                    else:
+                        toktypes.append(gguf.TokenType.USER_DEFINED)
                else:
-                    toktypes.append(gguf.TokenType.USER_DEFINED)
-            else:
-                tokens.append(reverse_vocab[i])
-                toktypes.append(gguf.TokenType.NORMAL)
+                    tokens.append(reverse_vocab[i])
+                    toktypes.append(gguf.TokenType.NORMAL)

-        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)
+            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_vocab = gguf.SpecialVocab(dir_model, load_merges=False)
-        special_vocab.merges = merges
+            special_vocab = gguf.SpecialVocab(dir_model, load_merges=False)
+            special_vocab.merges = merges
        # only add special tokens when they were not already loaded from config.json
        special_vocab._set_special_token("eos", tokenizer.get_added_vocab()["<|endoftext|>"])
        special_vocab._set_special_token("eot", tokenizer.get_added_vocab()["<|user|>"])
@ -4437,16 +4446,20 @@ class ChatGLMModel(Model):
    def set_gguf_parameters(self):
        n_embed = self.hparams.get("hidden_size", self.hparams.get("n_embed"))
        n_head = self.hparams.get("n_head", self.hparams.get("num_attention_heads"))
-        n_head_kv = self.hparams.get("multi_query_group_num", n_head)
+        n_head_kv = self.hparams.get("multi_query_group_num", self.hparams.get("num_key_value_heads", n_head))
        self.gguf_writer.add_context_length(self.hparams.get("seq_length", n_embed))
        self.gguf_writer.add_embedding_length(n_embed)
-        self.gguf_writer.add_feed_forward_length(self.hparams.get("ffn_hidden_size", 4 * n_embed))
-        self.gguf_writer.add_block_count(self.hparams["num_layers"])
+        self.gguf_writer.add_feed_forward_length(self.hparams.get("ffn_hidden_size", self.hparams.get("intermediate_size", 4 * n_embed)))
+        self.gguf_writer.add_block_count(self.hparams.get("num_layers", self.hparams["num_hidden_layers"]))
        self.gguf_writer.add_head_count(n_head)
        self.gguf_writer.add_head_count_kv(n_head_kv)
-        self.gguf_writer.add_layer_norm_rms_eps(self.hparams["layernorm_epsilon"])
+        self.gguf_writer.add_layer_norm_rms_eps(self.hparams.get("layernorm_epsilon",1e-5))
        self.gguf_writer.add_file_type(self.ftype)
-        self.gguf_writer.add_rope_dimension_count(64)
+        if "attention_dim" in self.hparams:
+            rope_dim = self.hparams["attention_dim"]
+        else:
+            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)))
        self.gguf_writer.add_add_bos_token(False)
        rope_freq = 10000
        if "rope_ratio" in self.hparams:
@ -4456,7 +4469,7 @@ class ChatGLMModel(Model):
    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
        del bid  # unused

-        if name.endswith(".rotary_pos_emb.inv_freq"):
+        if name.endswith(".rotary_pos_emb.inv_freq") or name.startswith("model.vision."):
            return []

        name = name.removeprefix("transformer.")
--- a/examples/llava/README-glmedge.md
+++ b/examples/llava/README-glmedge.md
@ -0,0 +1,43 @@
+# GLMV-EDGE
+
+Currently this implementation supports [glm-edge-v-2b](https://huggingface.co/THUDM/glm-edge-v-2b) and [glm-edge-v-5b](https://huggingface.co/THUDM/glm-edge-v-5b).
+
+## Usage
+Build with cmake or run `make llama-llava-cli` to build it.
+
+After building, run: `./llama-llava-cli` to see the usage. For example:
+
+```sh
+./llama-llava-cli -m model_path/ggml-model-f16.gguf --mmproj model_path/mmproj-model-f16.gguf --image img_path/image.jpg -p "<|system|>\n system prompt <image><|user|>\n prompt <|assistant|>\n"
+```
+
+**note**: A lower temperature like 0.1 is recommended for better quality. add `--temp 0.1` to the command to do so.
+**note**: For GPU offloading ensure to use the `-ngl` flag just like usual
+
+## GGUF conversion
+
+1. Clone a GLMV-EDGE model ([2B](https://huggingface.co/THUDM/glm-edge-v-2b) or [5B](https://huggingface.co/THUDM/glm-edge-v-5b)). For example:
+
+```sh
+git clone https://huggingface.co/THUDM/glm-edge-v-5b or https://huggingface.co/THUDM/glm-edge-v-2b
+```
+
+2. Use `glmedge-surgery.py` to split the GLMV-EDGE model to LLM and multimodel projector constituents:
+
+```sh
+python ./examples/llava/glmedge-surgery.py -m ../model_path
+```
+
+4. Use `glmedge-convert-image-encoder-to-gguf.py` to convert the GLMV-EDGE image encoder to GGUF:
+
+```sh
+python ./examples/llava/glmedge-convert-image-encoder-to-gguf.py -m ../model_path --llava-projector ../model_path/glm.projector --output-dir ../model_path
+```
+
+5. Use `examples/convert_hf_to_gguf.py` to convert the LLM part of GLMV-EDGE to GGUF:
+
+```sh
+python convert_hf_to_gguf.py ../model_path
+```
+
+Now both the LLM part and the image encoder are in the `model_path` directory.
--- a/examples/llava/clip.cpp
+++ b/examples/llava/clip.cpp
@ -101,6 +101,7 @@ static std::string format(const char * fmt, ...) {
 #define KEY_HAS_VIS_ENC         "clip.has_vision_encoder"
 #define KEY_HAS_LLAVA_PROJ      "clip.has_llava_projector"
 #define KEY_HAS_MINICPMV_PROJ   "clip.has_minicpmv_projector"
+#define KEY_HAS_GLM_PROJ        "clip.has_glm_projector"
 #define KEY_MINICPMV_VERSION    "clip.minicpmv_version"
 #define KEY_HAS_QWEN2VL_MERGER  "clip.has_qwen2vl_merger"
 #define KEY_USE_GELU            "clip.use_gelu"
@ -159,6 +160,15 @@ static std::string format(const char * fmt, ...) {
 #define TN_MINICPMV_ATTN "resampler.attn.%s.%s"
 #define TN_MINICPMV_LN "resampler.ln_%s.%s"

+#define TN_GLM_ADAPER_CONV "adapter.conv.%s"
+#define TN_GLM_ADAPTER_LINEAR "adapter.linear.linear.%s"
+#define TN_GLM_ADAPTER_NORM_1 "adapter.linear.norm1.%s"
+#define TN_GLM_ADAPTER_D_H_2_4H "adapter.linear.dense_h_to_4h.%s"
+#define TN_GLM_ADAPTER_GATE "adapter.linear.gate.%s"
+#define TN_GLM_ADAPTER_D_4H_2_H "adapter.linear.dense_4h_to_h.%s"
+#define TN_GLM_BOI_W "adapter.boi"
+#define TN_GLM_EOI_W "adapter.eoi"
+

 enum projector_type {
    PROJECTOR_TYPE_MLP,
@ -166,6 +176,7 @@ enum projector_type {
    PROJECTOR_TYPE_LDP,
    PROJECTOR_TYPE_LDPV2,
    PROJECTOR_TYPE_RESAMPLER,
+    PROJECTOR_TYPE_ADAPTER,
    PROJECTOR_TYPE_MERGER,
    PROJECTOR_TYPE_UNKNOWN,
 };
@ -175,6 +186,7 @@ static std::map<projector_type, std::string> PROJECTOR_TYPE_NAMES = {
    { PROJECTOR_TYPE_LDP, "ldp" },
    { PROJECTOR_TYPE_LDPV2, "ldpv2"},
    { PROJECTOR_TYPE_RESAMPLER, "resampler"},
+    { PROJECTOR_TYPE_ADAPTER, "adapter"},
    { PROJECTOR_TYPE_MERGER, "qwen2vl_merger"},
 };

@ -499,6 +511,12 @@ struct clip_vision_model {
    struct ggml_tensor * mm_4_w = NULL;
    struct ggml_tensor * mm_4_b = NULL;

+    //GLMV-Edge projection
+    struct ggml_tensor * mm_model_adapter_conv_w;
+    struct ggml_tensor * mm_model_adapter_conv_b;
+    struct ggml_tensor * boi_w;
+    struct ggml_tensor * eoi_w;
+
    // MobileVLM projection
    struct ggml_tensor * mm_model_mlp_1_w;
    struct ggml_tensor * mm_model_mlp_1_b;
@ -559,6 +577,7 @@ struct clip_ctx {
    bool has_vision_encoder  = false;
    bool has_llava_projector = false;
    bool has_minicpmv_projector = false;
+    bool has_glm_projector = false;
    bool has_qwen2vl_merger = false;
    int minicpmv_version = 2;

@ -637,7 +656,7 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32

    const int batch_size = imgs->size;

-    if (ctx->has_llava_projector || ctx->has_minicpmv_projector) {
+    if (ctx->has_llava_projector || ctx->has_minicpmv_projector || ctx->has_glm_projector) {
        GGML_ASSERT(batch_size == 1);
    }

@ -730,8 +749,7 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
    }

    // loop over layers
-    if (ctx->has_minicpmv_projector || ctx->has_qwen2vl_merger) {
-        // TODO: figure out why we doing thing in this way ???
+    if (ctx->has_minicpmv_projector || ctx->has_glm_projector || ctx->has_qwen2vl_merger) {
        n_layer += 1;
    }
    for (int il = 0; il < n_layer - 1; il++) {
@ -1086,7 +1104,33 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
            GGML_ASSERT(false);
        }
    }
-    else if (ctx->proj_type == PROJECTOR_TYPE_MERGER) {
+    // glm projector
+    else if(ctx->has_glm_projector){
+        if (ctx->proj_type == PROJECTOR_TYPE_ADAPTER){
+            size_t gridsz = (size_t)sqrt(embeddings->ne[1]);
+            embeddings = ggml_cont(ctx0, ggml_permute(ctx0,embeddings,1,0,2,3));
+            embeddings = ggml_reshape_3d(ctx0,embeddings,gridsz,gridsz,embeddings->ne[1]);
+            embeddings = ggml_conv_2d(ctx0, model.mm_model_adapter_conv_w, embeddings, 2, 2, 0, 0, 1, 1);
+            embeddings = ggml_reshape_3d(ctx0, embeddings,embeddings->ne[0]*embeddings->ne[1] , embeddings->ne[2], batch_size);
+            embeddings = ggml_cont(ctx0, ggml_permute(ctx0,embeddings,1,0,2,3));
+            embeddings = ggml_add(ctx0, embeddings, model.mm_model_adapter_conv_b);
+            //GLU
+            {
+                embeddings = ggml_mul_mat(ctx0, model.mm_model_mlp_0_w, embeddings);
+                embeddings = ggml_norm(ctx0, embeddings, eps);
+                embeddings = ggml_add(ctx0, ggml_mul(ctx0, embeddings, model.mm_model_ln_q_w), model.mm_model_ln_q_b);
+                embeddings = ggml_gelu_inplace(ctx0, embeddings);
+                struct ggml_tensor * x = embeddings;
+                embeddings = ggml_mul_mat(ctx0, model.mm_model_mlp_2_w, embeddings);
+                x = ggml_mul_mat(ctx0,model.mm_model_mlp_1_w,x);
+                embeddings = ggml_silu_inplace(ctx0,embeddings);
+                embeddings = ggml_mul(ctx0,embeddings,x);
+                embeddings = ggml_mul_mat(ctx0,model.mm_model_mlp_3_w,embeddings);
+            }
+        }else{
+            GGML_ABORT("fatel error");
+        }
+    }else if (ctx->proj_type == PROJECTOR_TYPE_MERGER) {
        embeddings = ggml_reshape_3d(ctx0, embeddings, hidden_size * 4, num_positions / 4, batch_size);

        embeddings = ggml_mul_mat(ctx0, model.mm_0_w, embeddings);
@ -1275,6 +1319,11 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
            new_clip->minicpmv_version = gguf_get_val_i32(ctx, idx);
        }

+        idx = gguf_find_key(ctx, KEY_HAS_GLM_PROJ);
+        if (idx != -1) {
+            new_clip->has_glm_projector = gguf_get_val_bool(ctx, idx);
+        }
+
        idx = gguf_find_key(ctx, KEY_HAS_QWEN2VL_MERGER);
        if (idx != -1) {
            new_clip->has_qwen2vl_merger = gguf_get_val_bool(ctx, idx);
@ -1299,6 +1348,7 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
            LOG_INF("%s: vision_encoder: %d\n", __func__, new_clip->has_vision_encoder);
            LOG_INF("%s: llava_projector:  %d\n", __func__, new_clip->has_llava_projector);
            LOG_INF("%s: minicpmv_projector:  %d\n", __func__, new_clip->has_minicpmv_projector);
+            LOG_INF("%s: glm_projector:  %d\n", __func__, new_clip->has_glm_projector);
            LOG_INF("%s: model size:     %.2f MB\n", __func__, model_size / 1024.0 / 1024.0);
            LOG_INF("%s: metadata size:  %.2f MB\n", __func__, ggml_get_mem_size(meta) / 1024.0 / 1024.0);
        }
@ -1566,6 +1616,18 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
            vision_model.mm_model_ln_post_w = get_tensor(new_clip->ctx_data, format(TN_MINICPMV_LN, "post", "weight"));
            vision_model.mm_model_ln_post_b = get_tensor(new_clip->ctx_data, format(TN_MINICPMV_LN, "post", "bias"));
        }
+        else if(new_clip->proj_type == PROJECTOR_TYPE_ADAPTER){
+            vision_model.mm_model_adapter_conv_w = get_tensor(new_clip->ctx_data, format(TN_GLM_ADAPER_CONV, "weight"));
+            vision_model.mm_model_adapter_conv_b = get_tensor(new_clip->ctx_data, format(TN_GLM_ADAPER_CONV, "bias"));
+            vision_model.mm_model_mlp_0_w = get_tensor(new_clip->ctx_data, format(TN_GLM_ADAPTER_LINEAR,"weight"));
+            vision_model.mm_model_ln_q_w = get_tensor(new_clip->ctx_data, format(TN_GLM_ADAPTER_NORM_1,"weight"));
+            vision_model.mm_model_ln_q_b = get_tensor(new_clip->ctx_data, format(TN_GLM_ADAPTER_NORM_1,"bias"));
+            vision_model.mm_model_mlp_1_w =  get_tensor(new_clip->ctx_data, format(TN_GLM_ADAPTER_D_H_2_4H,"weight"));
+            vision_model.mm_model_mlp_2_w =  get_tensor(new_clip->ctx_data, format(TN_GLM_ADAPTER_GATE,"weight"));
+            vision_model.mm_model_mlp_3_w =  get_tensor(new_clip->ctx_data, format(TN_GLM_ADAPTER_D_4H_2_H,"weight"));
+            vision_model.boi_w = get_tensor(new_clip->ctx_data, TN_GLM_BOI_W);
+            vision_model.eoi_w = get_tensor(new_clip->ctx_data, TN_GLM_EOI_W);
+        }
        else if (new_clip->proj_type == PROJECTOR_TYPE_MERGER) {
            vision_model.mm_0_w = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 0, "weight"));
            vision_model.mm_0_b = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 0, "bias"));
@ -2098,6 +2160,20 @@ bool clip_image_preprocess(struct clip_ctx * ctx, const clip_image_u8 * img, cli
        return true;
    }

+    if(ctx->has_glm_projector){
+        res_imgs->size = 1;
+        res_imgs->data = new clip_image_f32[res_imgs->size];
+        clip_image_u8 resized_image;
+        int32_t sz=ctx->vision_model.hparams.image_size;
+        bicubic_resize(*img, resized_image,sz,sz);
+        clip_image_f32 * res = clip_image_f32_init();
+        //clip_image_save_to_bmp(resized_image, "resized.bmp");
+        normalize_image_u8_to_f32(&resized_image, res, ctx->image_mean, ctx->image_std);
+        res_imgs->data[0] = *res;
+        clip_image_f32_free(res);
+        return true;
+    }
+
    bool pad_to_square = true;
    if (!ctx->has_vision_encoder) {
        LOG_ERR("This gguf file seems to have no vision encoder\n");
@ -2283,6 +2359,8 @@ void clip_free(clip_ctx * ctx) {
 }

 size_t clip_embd_nbytes(const struct clip_ctx * ctx) {
+    if(ctx->has_glm_projector)
+        return (clip_n_patches(ctx)+2) * clip_n_mmproj_embd(ctx) * sizeof(float);
    return clip_n_patches(ctx) * clip_n_mmproj_embd(ctx) * sizeof(float);
 }

@ -2325,7 +2403,7 @@ int clip_n_patches_by_img(const struct clip_ctx * ctx, struct clip_image_f32 * i

    int n_patches = (params.image_size / params.patch_size) * (params.image_size / params.patch_size);

-    if (ctx->proj_type == PROJECTOR_TYPE_LDP || ctx->proj_type == PROJECTOR_TYPE_LDPV2) {
+    if (ctx->proj_type == PROJECTOR_TYPE_LDP || ctx->proj_type == PROJECTOR_TYPE_LDPV2 || ctx->proj_type == PROJECTOR_TYPE_ADAPTER) {
        n_patches /= 4;
    } else if (ctx->proj_type == PROJECTOR_TYPE_RESAMPLER) {
        if (ctx->minicpmv_version == 2) {
@ -2455,6 +2533,12 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
    if (ctx->has_minicpmv_projector) {
        GGML_ASSERT(batch_size == 1);
    }
+    if(ctx->has_glm_projector){
+        GGML_ASSERT(batch_size == 1);
+        ggml_tensor * boi = ctx->vision_model.boi_w;
+        ggml_backend_tensor_get(boi,vec,0,ggml_nbytes(boi));
+        vec=(float*)(vec+ggml_nelements(boi)); //offset for boi
+    }

    // build the inference graph
    ggml_cgraph * gf = clip_image_build_graph(ctx, imgs, ctx->load_image_size, true);
@ -2604,7 +2688,7 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
            ggml_backend_tensor_set(positions, positions_data, 0, ggml_nbytes(positions));
            free(positions_data);

-            {
+            if (!ctx->has_glm_projector){
                struct ggml_tensor * patches = ggml_graph_get_tensor(gf, "patches");
                int* patches_data = (int*)malloc(ggml_nbytes(patches));
                for (int i = 0; i < num_patches; i++) {
@ -2628,6 +2712,13 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
    // copy the embeddings to the location passed by the user
    ggml_backend_tensor_get(embeddings, vec, 0, ggml_nbytes(embeddings));

+    if(ctx->has_glm_projector){
+        //eoi
+        ggml_tensor * eoi = ctx->vision_model.eoi_w;
+        int offset=ggml_nelements(eoi)*clip_n_patches(ctx);
+        ggml_backend_tensor_get(eoi,vec+offset,0,ggml_nbytes(eoi));
+    }
+
    return true;
 }

@ -2785,6 +2876,9 @@ int clip_n_mmproj_embd(const struct clip_ctx * ctx) {
            return 3584;
        }
    }
+    if (ctx->proj_type == PROJECTOR_TYPE_ADAPTER){
+        return ctx->vision_model.mm_model_mlp_3_w->ne[1];
+    }
    if (ctx->proj_type == PROJECTOR_TYPE_MERGER) {
        return ctx->vision_model.mm_1_b->ne[0];
    }
@ -2800,6 +2894,9 @@ int clip_is_minicpmv(const struct clip_ctx * ctx) {
    return 0;
 }

+bool clip_is_glm(const struct clip_ctx * ctx) {
+    return ctx->has_glm_projector;
+}
 bool clip_is_qwen2vl(const struct clip_ctx * ctx) {
    return ctx->has_qwen2vl_merger;
 }
--- a/examples/llava/clip.h
+++ b/examples/llava/clip.h
@ -93,6 +93,8 @@ CLIP_API bool clip_is_qwen2vl(const struct clip_ctx * ctx);

 CLIP_API bool clip_encode_float_image (struct clip_ctx * ctx, int n_threads, float * img, int h, int w, float * vec);

+CLIP_API bool clip_is_glm(const struct clip_ctx * ctx);
+
 #ifdef __cplusplus
 }
 #endif
--- a/examples/llava/glmedge-convert-image-encoder-to-gguf.py
+++ b/examples/llava/glmedge-convert-image-encoder-to-gguf.py
@ -0,0 +1,280 @@
+import argparse
+import os
+import json
+import re
+
+import torch
+import numpy as np
+from gguf import *
+
+TEXT = "clip.text"
+VISION = "clip.vision"
+from transformers import SiglipVisionModel, SiglipVisionConfig
+
+def k(raw_key: str, arch: str) -> str:
+    return raw_key.format(arch=arch)
+
+
+def should_skip_tensor(name: str, has_text: bool, has_vision: bool, has_llava: bool) -> bool:
+    if name in (
+        "logit_scale",
+        "text_model.embeddings.position_ids",
+        "vision_model.embeddings.position_ids",
+    ):
+        return True
+
+    if name in (
+        "vision_model.head.probe",
+        "vision_model.head.attention.in_proj_weight",
+        "vision_model.head.attention.in_proj_bias",
+        "vision_model.head.attention.out_proj.weight",
+        "vision_model.head.attention.out_proj.bias",
+        "vision_model.head.layernorm.weight",
+        "vision_model.head.layernorm.bias",
+        "vision_model.head.mlp.fc1.weight",
+        "vision_model.head.mlp.fc1.bias",
+        "vision_model.head.mlp.fc2.weight",
+        "vision_model.head.mlp.fc2.bias"
+    ):
+        return True
+
+    if name.startswith("v") and not has_vision:
+        return True
+
+    if name.startswith("t") and not has_text:
+        return True
+
+    return False
+
+
+def get_tensor_name(name: str) -> str:
+    if "projection" in name:
+        return name
+    if "mm_projector" in name:
+        name = name.replace("model.mm_projector", "mm")
+        name = re.sub(r'mm\.mlp\.mlp', 'mm.model.mlp', name, count=1)
+        name = re.sub(r'mm\.peg\.peg', 'mm.model.peg', name, count=1)
+        return name
+
+    return name.replace("text_model", "t").replace("vision_model", "v").replace("encoder.layers", "blk").replace("embeddings.", "").replace("_proj", "").replace("self_attn.", "attn_").replace("layer_norm", "ln").replace("layernorm", "ln").replace("mlp.fc1", "ffn_down").replace("mlp.fc2", "ffn_up").replace("embedding", "embd").replace("final", "post").replace("layrnorm", "ln")
+
+
+def bytes_to_unicode():
+    """
+    Returns list of utf-8 byte and a corresponding list of unicode strings.
+    The reversible bpe codes work on unicode strings.
+    This means you need a large # of unicode characters in your vocab if you want to avoid UNKs.
+    When you're at something like a 10B token dataset you end up needing around 5K for decent coverage.
+    This is a significant percentage of your normal, say, 32K bpe vocab.
+    To avoid that, we want lookup tables between utf-8 bytes and unicode strings.
+    And avoids mapping to whitespace/control characters the bpe code barfs on.
+    """
+    bs = (
+        list(range(ord("!"), ord("~") + 1))
+        + list(range(ord("¡"), ord("¬") + 1))
+        + list(range(ord("®"), ord("ÿ") + 1))
+    )
+    cs = bs[:]
+    n = 0
+    for b in range(2**8):
+        if b not in bs:
+            bs.append(b)
+            cs.append(2**8 + n)
+            n += 1
+    cs = [chr(n) for n in cs]
+    return dict(zip(bs, cs))
+
+
+ap = argparse.ArgumentParser()
+ap.add_argument("-m", "--model-dir", help="Path to model directory cloned from HF Hub", required=True)
+ap.add_argument("--use-f32", action="store_true", default=False, help="Use f32 instead of f16")
+ap.add_argument("--text-only", action="store_true", required=False,
+                help="Save a text-only model. It can't be used to encode images")
+ap.add_argument("--vision-only", action="store_true", required=False,
+                help="Save a vision-only model. It can't be used to encode texts")
+ap.add_argument("--clip-model-is-vision", action="store_true", required=False,
+                help="The clip model is a pure vision model (ShareGPT4V vision extract for example)")
+ap.add_argument("--clip-model-is-openclip", action="store_true", required=False,
+                help="The clip model is from openclip (for ViT-SO400M type))")
+ap.add_argument("--llava-projector", help="Path to llava.projector file. If specified, save an image encoder for LLaVA models.")
+ap.add_argument("--projector-type", help="Type of projector. Possible values: mlp, ldp, ldpv2", choices=["mlp", "ldp", "ldpv2","adapter"], default="adapter")
+ap.add_argument("-o", "--output-dir", help="Directory to save GGUF files. Default is the original model directory", default=None)
+# Example --image_mean 0.48145466 0.4578275 0.40821073 --image_std 0.26862954 0.26130258 0.27577711
+# Example --image_mean 0.5 0.5 0.5 --image_std 0.5 0.5 0.5
+default_image_mean = [0.5, 0.5, 0.5]
+default_image_std = [0.5, 0.5, 0.5]
+ap.add_argument('--image-mean', type=float, nargs='+', help='Mean of the images for normalization (overrides processor) ', default=None)
+ap.add_argument('--image-std', type=float, nargs='+', help='Standard deviation of the images for normalization (overrides processor)', default=None)
+
+# with proper
+args = ap.parse_args()
+
+
+if args.text_only and args.vision_only:
+    print("--text-only and --image-only arguments cannot be specified at the same time.")
+    exit(1)
+
+if args.use_f32:
+    print("WARNING: Weights for the convolution op is always saved in f16, as the convolution op in GGML does not support 32-bit kernel weights yet.")
+
+# output in the same directory as the model if output_dir is None
+dir_model = args.model_dir
+
+if args.clip_model_is_vision or not os.path.exists(dir_model + "/vocab.json") or args.clip_model_is_openclip:
+    vocab = None
+    tokens = None
+else:
+    with open(dir_model + "/vocab.json", "r", encoding="utf-8") as f:
+        vocab = json.load(f)
+        tokens = [key for key in vocab]
+
+with open(dir_model + "/config.json", "r", encoding="utf-8") as f:
+    config = json.load(f)
+    if args.clip_model_is_vision:
+        v_hparams = config
+        t_hparams = None
+    else:
+        v_hparams = config["vision_config"]
+        t_hparams = None
+
+# possible data types
+#   ftype == 0 -> float32
+#   ftype == 1 -> float16
+#
+# map from ftype to string
+ftype_str = ["f32", "f16"]
+
+ftype = 1
+if args.use_f32:
+    ftype = 0
+
+vision_config = SiglipVisionConfig(**v_hparams)
+model = SiglipVisionModel(vision_config)
+model.load_state_dict(torch.load(os.path.join(dir_model, "glm.clip")))
+
+fname_middle = None
+has_text_encoder = False
+has_vision_encoder = True
+has_glm_projector = True
+if args.text_only:
+    fname_middle = "text-"
+    has_vision_encoder = False
+elif args.llava_projector is not None:
+    fname_middle = "mmproj-"
+    has_text_encoder = False
+    has_glm_projector = True
+elif args.vision_only:
+    fname_middle = "vision-"
+    has_text_encoder = False
+else:
+    fname_middle = ""
+
+output_dir = args.output_dir if args.output_dir is not None else dir_model
+os.makedirs(output_dir, exist_ok=True)
+output_prefix = os.path.basename(output_dir).replace("ggml_", "")
+fname_out = os.path.join(output_dir, f"{fname_middle}model-{ftype_str[ftype]}.gguf")
+fout = GGUFWriter(path=fname_out, arch="clip")
+
+fout.add_bool("clip.has_text_encoder", has_text_encoder)
+fout.add_bool("clip.has_vision_encoder", has_vision_encoder)
+fout.add_bool("clip.has_glm_projector", has_glm_projector)
+fout.add_file_type(ftype)
+model_name = config["_name_or_path"] if "_name_or_path" in config else os.path.basename(dir_model)
+fout.add_name(model_name)
+if has_glm_projector:
+    fout.add_description("image encoder for glm4v")
+    fout.add_string("clip.projector_type", "adapter")
+else:
+    fout.add_description("two-tower CLIP model")
+
+if has_text_encoder:
+    assert t_hparams is not None
+    assert tokens is not None
+    # text_model hparams
+    fout.add_uint32(k(KEY_CONTEXT_LENGTH, TEXT), t_hparams["max_position_embeddings"])
+    fout.add_uint32(k(KEY_EMBEDDING_LENGTH, TEXT), t_hparams["hidden_size"])
+    fout.add_uint32(k(KEY_FEED_FORWARD_LENGTH, TEXT), t_hparams["intermediate_size"])
+    fout.add_uint32("clip.text.projection_dim", t_hparams.get("projection_dim", config["projection_dim"]))
+    fout.add_uint32(k(KEY_ATTENTION_HEAD_COUNT, TEXT), t_hparams["num_attention_heads"])
+    fout.add_float32(k(KEY_ATTENTION_LAYERNORM_EPS, TEXT), t_hparams["layer_norm_eps"])
+    fout.add_uint32(k(KEY_BLOCK_COUNT, TEXT), t_hparams["num_hidden_layers"])
+    fout.add_token_list(tokens)
+
+if has_vision_encoder:
+    # vision_model hparams
+    fout.add_uint32("clip.vision.image_size", v_hparams["image_size"])
+    fout.add_uint32("clip.vision.patch_size", v_hparams["patch_size"])
+    fout.add_uint32(k(KEY_EMBEDDING_LENGTH, VISION), v_hparams["hidden_size"])
+    fout.add_uint32(k(KEY_FEED_FORWARD_LENGTH, VISION), v_hparams["intermediate_size"])
+    fout.add_uint32("clip.vision.projection_dim", 0)
+    fout.add_uint32(k(KEY_ATTENTION_HEAD_COUNT, VISION), v_hparams["num_attention_heads"])
+    fout.add_float32(k(KEY_ATTENTION_LAYERNORM_EPS, VISION), 1e-6)
+    fout.add_uint32(k(KEY_BLOCK_COUNT, VISION), v_hparams["num_hidden_layers"])
+
+    image_mean = args.image_mean if args.image_mean is not None else default_image_mean
+    image_std = args.image_std if args.image_std is not None else default_image_std
+    fout.add_array("clip.vision.image_mean", image_mean)
+    fout.add_array("clip.vision.image_std", image_std)
+
+fout.add_bool("clip.use_gelu", True)
+
+
+if has_glm_projector:
+    # model.vision_model.encoder.layers.pop(-1)  # pyright: ignore[reportAttributeAccessIssue]
+    projector = torch.load(args.llava_projector)
+    for name, data in projector.items():
+        name = get_tensor_name(name)
+        # pw and dw conv ndim==4
+        if data.ndim == 2 or data.ndim == 4:
+            data = data.squeeze().numpy().astype(np.float16)
+        else:
+            data = data.squeeze().numpy().astype(np.float32)
+        if name.startswith("vision."):
+            name=name.replace("vision.","")
+        fout.add_tensor(name, data)
+        print(f"Projector {name} - {data.dtype} - shape = {data.shape}")
+        # print(f"Projector {name} tensors added\n")
+
+state_dict = model.state_dict()  # pyright: ignore[reportAttributeAccessIssue]
+for name, data in state_dict.items():
+    if should_skip_tensor(name, has_text_encoder, has_vision_encoder, has_glm_projector):
+        # we don't need this
+        print(f"skipping parameter: {name}")
+        continue
+
+    name = get_tensor_name(name)
+    data = data.squeeze().numpy()
+
+    n_dims = len(data.shape)
+
+    # ftype == 0 -> float32, ftype == 1 -> float16
+    ftype_cur = 0
+    if n_dims == 4:
+        print(f"tensor {name} is always saved in f16")
+        data = data.astype(np.float16)
+        ftype_cur = 1
+    elif ftype == 1:
+        if name[-7:] == ".weight" and n_dims == 2:
+            # print("  Converting to float16")
+            data = data.astype(np.float16)
+            ftype_cur = 1
+        else:
+            # print("  Converting to float32")
+            data = data.astype(np.float32)
+            ftype_cur = 0
+    else:
+        if data.dtype != np.float32:
+            # print("  Converting to float32")
+            data = data.astype(np.float32)
+            ftype_cur = 0
+    print(f"siglip {name} - {data.dtype} - shape = {data.shape}")
+    # print(f"{name} - {ftype_str[ftype_cur]} - shape = {data.shape}")
+    fout.add_tensor(name, data)
+
+
+fout.write_header_to_file()
+fout.write_kv_data_to_file()
+fout.write_tensors_to_file()
+fout.close()
+
+print("Done. Output file: " + fname_out)
--- a/examples/llava/glmedge-surgery.py
+++ b/examples/llava/glmedge-surgery.py
@ -0,0 +1,33 @@
+import argparse
+import os
+import torch
+from transformers import AutoModel
+
+ap = argparse.ArgumentParser()
+ap.add_argument("-m", "--model", help="Path to GLM model")
+args = ap.parse_args()
+
+# find the model part that includes the the multimodal projector weights
+model = AutoModel.from_pretrained(args.model, trust_remote_code=True, local_files_only=True)
+checkpoint = model.state_dict()
+
+# get a list of mm tensor names
+mm_tensors = [k for k, v in checkpoint.items() if k.startswith("vision.adapter.")]
+
+# store these tensors in a new dictionary and torch.save them
+projector = {name: checkpoint[name].float() for name in mm_tensors}
+torch.save(projector, f"{args.model}/glm.projector")
+
+clip_tensors = [k for k, v in checkpoint.items() if k.startswith("vision.vit.model.vision_model.")]
+if len(clip_tensors) > 0:
+    clip = {name.replace("vision.vit.model.", ""): checkpoint[name].float() for name in clip_tensors}
+    torch.save(clip, f"{args.model}/glm.clip")
+
+    # added tokens should be removed to be able to convert Mistral models
+    if os.path.exists(f"{args.model}/added_tokens.json"):
+        with open(f"{args.model}/added_tokens.json", "w") as f:
+            f.write("{}\n")
+
+print("Done!")
+print(f"Now you can convert {args.model} to a regular LLaMA GGUF file.")
+print(f"Also, use {args.model}glm.projector to prepare a glm-encoder.gguf file.")
--- a/examples/llava/llava.cpp
+++ b/examples/llava/llava.cpp
@ -314,6 +314,20 @@ static bool encode_image_with_clip(clip_ctx * ctx_clip, int n_threads, const cli
        clip_add_load_image_size(ctx_clip, load_image_size);
        LOG_INF("%s: load_image_size %d %d\n", __func__, load_image_size->width, load_image_size->height);
    }
+    else if (clip_is_glm(ctx_clip)){
+        struct clip_image_size * load_image_size = clip_image_size_init();
+        load_image_size->width = img_res_v.data[0].nx;
+        load_image_size->height = img_res_v.data[0].ny;
+        clip_add_load_image_size(ctx_clip, load_image_size);
+
+        bool encoded = clip_image_encode(ctx_clip, n_threads, &img_res_v.data[0], image_embd);
+        int pos = int(load_image_size->width/clip_patch_size(ctx_clip)/2);
+        *n_img_pos = (pos * pos + 2);
+        if (!encoded){
+            LOG_ERR("Unable to encode image \n");
+            return false;
+        }
+    }
    else if (strcmp(mm_patch_merge_type, "spatial_unpad") != 0) {
        // flat / default llava-1.5 type embedding
        *n_img_pos = clip_n_patches(ctx_clip);
@ -398,6 +412,9 @@ bool llava_image_embed_make_with_clip_img(clip_ctx * ctx_clip, int n_threads, co
    if (clip_is_minicpmv(ctx_clip)) {
        num_max_patches = 10;
    }
+    if (clip_is_glm(ctx_clip)) {
+        num_max_patches = 1;
+    }
    float * image_embd;
    if (clip_is_qwen2vl(ctx_clip)) {
        // qwen2vl don't split image into chunks, so `num_max_patches` is not needed.
--- a/examples/server/CMakeLists.txt
+++ b/examples/server/CMakeLists.txt
@ -34,6 +34,7 @@ endforeach()
 add_executable(${TARGET} ${TARGET_SRCS})
 install(TARGETS ${TARGET} RUNTIME)

+target_include_directories(${TARGET} PRIVATE ${CMAKE_SOURCE_DIR})
 target_link_libraries(${TARGET} PRIVATE common ${CMAKE_THREAD_LIBS_INIT})

 if (LLAMA_SERVER_SSL)
--- a/examples/server/README.md
+++ b/examples/server/README.md
@ -450,6 +450,8 @@ These words will not be included in the completion, so make sure to add them to

 `post_sampling_probs`: Returns the probabilities of top `n_probs` tokens after applying sampling chain.

+`response_fields`: A list of response fields, for example: `"response_fields": ["content", "generation_settings/n_predict"]`. If the specified field is missing, it will simply be omitted from the response without triggering an error.
+
 **Response format**

 - Note: In streaming mode (`stream`), only `content`, `tokens` and `stop` will be returned until end of completion. Responses are sent using the [Server-sent events](https://html.spec.whatwg.org/multipage/server-sent-events.html) standard. Note: the browser's `EventSource` interface cannot be used due to its lack of `POST` request support.
--- a/examples/server/server.cpp
+++ b/examples/server/server.cpp
@ -92,6 +92,7 @@ struct slot_params {
    int64_t t_max_predict_ms = -1; // if positive, limit the generation phase to this time limit

    std::vector<std::string> antiprompt;
+    std::vector<std::string> response_fields;
    bool timings_per_token = false;
    bool post_sampling_probs = false;
    bool ignore_eos = false;
@ -209,6 +210,7 @@ struct server_task {
        params.n_discard        = json_value(data, "n_discard",          defaults.n_discard);
      //params.t_max_prompt_ms  = json_value(data, "t_max_prompt_ms",    defaults.t_max_prompt_ms); // TODO: implement
        params.t_max_predict_ms = json_value(data, "t_max_predict_ms",   defaults.t_max_predict_ms);
+        params.response_fields  = json_value(data, "response_fields",   std::vector<std::string>());

        params.sampling.top_k              = json_value(data, "top_k",              defaults.sampling.top_k);
        params.sampling.top_p              = json_value(data, "top_p",              defaults.sampling.top_p);
@ -522,6 +524,7 @@ struct server_task_result_cmpl_final : server_task_result {

    bool post_sampling_probs;
    std::vector<completion_token_output> probs_output;
+    std::vector<std::string>  response_fields;

    slot_params generation_params;

@ -568,7 +571,7 @@ struct server_task_result_cmpl_final : server_task_result {
        if (!stream && !probs_output.empty()) {
            res["completion_probabilities"] = completion_token_output::probs_vector_to_json(probs_output, post_sampling_probs);
        }
-        return res;
+        return response_fields.empty() ? res : json_get_nested_values(response_fields, res);
    }

    json to_json_oaicompat_chat() {
@ -2066,6 +2069,7 @@ struct server_context {
        res->tokens          = slot.generated_tokens;
        res->timings         = slot.get_timings();
        res->prompt          = common_detokenize(ctx, slot.prompt_tokens, true);
+        res->response_fields = slot.params.response_fields;

        res->truncated           = slot.truncated;
        res->n_decoded           = slot.n_decoded;
@ -3786,6 +3790,17 @@ int main(int argc, char ** argv) {
            return;
        }

+        bool use_base64 = false;
+        if (body.count("encoding_format") != 0) {
+            const std::string& format = body.at("encoding_format");
+            if (format == "base64") {
+                use_base64 = true;
+            } else if (format != "float") {
+                res_error(res, format_error_response("The format to return the embeddings in. Can be either float or base64", ERROR_TYPE_INVALID_REQUEST));
+                return;
+            }
+        }
+
        std::vector<llama_tokens> tokenized_prompts = tokenize_input_prompts(ctx_server.ctx, prompt, true, true);
        for (const auto & tokens : tokenized_prompts) {
            // this check is necessary for models that do not add BOS token to the input
@ -3837,7 +3852,7 @@ int main(int argc, char ** argv) {
        }

        // write JSON response
-        json root = oaicompat ? format_embeddings_response_oaicompat(body, responses) : json(responses);
+        json root = oaicompat ? format_embeddings_response_oaicompat(body, responses, use_base64) : json(responses);
        res_ok(res, root);
    };

--- a/examples/server/tests/unit/test_completion.py
+++ b/examples/server/tests/unit/test_completion.py
@ -95,7 +95,7 @@ def test_consistent_result_same_seed(n_slots: int):
        res = server.make_request("POST", "/completion", data={
            "prompt": "I believe the meaning of life is",
            "seed": 42,
-            "temperature": 1.0,
+            "temperature": 0.0,
            "cache_prompt": False,  # TODO: remove this once test_cache_vs_nocache_prompt is fixed
        })
        if last_res is not None:
@ -120,9 +120,10 @@ def test_different_result_different_seed(n_slots: int):
            assert res.body["content"] != last_res.body["content"]
        last_res = res

-
+# TODO figure why it don't work with temperature = 1
+# @pytest.mark.parametrize("temperature", [0.0, 1.0])
@pytest.mark.parametrize("n_batch", [16, 32])
-@pytest.mark.parametrize("temperature", [0.0, 1.0])
+@pytest.mark.parametrize("temperature", [0.0])
 def test_consistent_result_different_batch_size(n_batch: int, temperature: float):
    global server
    server.n_batch = n_batch
@ -257,6 +258,40 @@ def test_completion_parallel_slots(n_slots: int, n_requests: int):
        # assert match_regex(re_content, res.body["content"])


+@pytest.mark.parametrize(
+    "prompt,n_predict,response_fields",
+    [
+        ("I believe the meaning of life is", 8, []),
+        ("I believe the meaning of life is", 32, ["content", "generation_settings/n_predict", "prompt"]),
+    ],
+)
+def test_completion_response_fields(
+    prompt: str, n_predict: int, response_fields: list[str]
+):
+    global server
+    server.start()
+    res = server.make_request(
+        "POST",
+        "/completion",
+        data={
+            "n_predict": n_predict,
+            "prompt": prompt,
+            "response_fields": response_fields,
+        },
+    )
+    assert res.status_code == 200
+    assert "content" in res.body
+    assert len(res.body["content"])
+    if len(response_fields):
+        assert res.body["generation_settings/n_predict"] == n_predict
+        assert res.body["prompt"] == "<s> " + prompt
+        assert isinstance(res.body["content"], str)
+        assert len(res.body) == len(response_fields)
+    else:
+        assert len(res.body)
+        assert "generation_settings" in res.body
+
+
 def test_n_probs():
    global server
    server.start()
--- a/examples/server/tests/unit/test_embedding.py
+++ b/examples/server/tests/unit/test_embedding.py
@ -1,3 +1,5 @@
+import base64
+import struct
 import pytest
 from openai import OpenAI
 from utils import *
@ -194,3 +196,42 @@ def test_embedding_usage_multiple():
    assert res.status_code == 200
    assert res.body['usage']['prompt_tokens'] == res.body['usage']['total_tokens']
    assert res.body['usage']['prompt_tokens'] == 2 * 9
+
+
+def test_embedding_openai_library_base64():
+    server.start()
+    test_input = "Test base64 embedding output"
+
+    # get embedding in default format
+    res = server.make_request("POST", "/v1/embeddings", data={
+        "input": test_input
+    })
+    assert res.status_code == 200
+    vec0 = res.body["data"][0]["embedding"]
+
+    # get embedding in base64 format
+    res = server.make_request("POST", "/v1/embeddings", data={
+        "input": test_input,
+        "encoding_format": "base64"
+    })
+
+    assert res.status_code == 200
+    assert "data" in res.body
+    assert len(res.body["data"]) == 1
+
+    embedding_data = res.body["data"][0]
+    assert "embedding" in embedding_data
+    assert isinstance(embedding_data["embedding"], str)
+
+    # Verify embedding is valid base64
+    decoded = base64.b64decode(embedding_data["embedding"])
+    # Verify decoded data can be converted back to float array
+    float_count = len(decoded) // 4  # 4 bytes per float
+    floats = struct.unpack(f'{float_count}f', decoded)
+    assert len(floats) > 0
+    assert all(isinstance(x, float) for x in floats)
+    assert len(floats) == len(vec0)
+
+    # make sure the decoded data is the same as the original
+    for x, y in zip(floats, vec0):
+        assert abs(x - y) < EPSILON
--- a/examples/server/utils.hpp
+++ b/examples/server/utils.hpp
@ -3,6 +3,7 @@
 #include "common.h"
 #include "log.h"
 #include "llama.h"
+#include "common/base64.hpp"

 #ifndef NDEBUG
 // crash the server in debug mode, otherwise send an http 500 error
@ -90,6 +91,28 @@ static bool json_is_array_of_mixed_numbers_strings(const json & data) {
    return false;
 }

+// get value by path(key1 / key2)
+static json json_get_nested_values(const std::vector<std::string> & paths, const json & js) {
+    json result = json::object();
+
+    for (const std::string & path : paths) {
+        json current = js;
+        const auto keys = string_split<std::string>(path, /*separator*/ '/');
+        bool valid_path = true;
+        for (const std::string & k : keys) {
+            if (valid_path && current.is_object() && current.contains(k)) {
+                current = current[k];
+            } else {
+                valid_path = false;
+            }
+        }
+        if (valid_path) {
+            result[path] = current;
+        }
+    }
+    return result;
+}
+
 /**
 * this handles 2 cases:
 * - only string, example: "string"
@ -591,16 +614,31 @@ static json oaicompat_completion_params_parse(
    return llama_params;
 }

-static json format_embeddings_response_oaicompat(const json & request, const json & embeddings) {
+static json format_embeddings_response_oaicompat(const json & request, const json & embeddings, bool use_base64 = false) {
    json data = json::array();
    int32_t n_tokens = 0;
    int i = 0;
    for (const auto & elem : embeddings) {
-        data.push_back(json{
-            {"embedding", json_value(elem, "embedding", json::array())},
-            {"index",     i++},
-            {"object",    "embedding"}
-        });
+        json embedding_obj;
+
+        if (use_base64) {
+            const auto& vec = json_value(elem, "embedding", json::array()).get<std::vector<float>>();
+            const char* data_ptr = reinterpret_cast<const char*>(vec.data());
+            size_t data_size = vec.size() * sizeof(float);
+            embedding_obj = {
+                {"embedding", base64::encode(data_ptr, data_size)},
+                {"index", i++},
+                {"object", "embedding"},
+                {"encoding_format", "base64"}
+            };
+        } else {
+            embedding_obj = {
+                {"embedding", json_value(elem, "embedding", json::array())},
+                {"index", i++},
+                {"object", "embedding"}
+            };
+        }
+        data.push_back(embedding_obj);

        n_tokens += json_value(elem, "tokens_evaluated", 0);
    }
--- a/ggml/src/CMakeLists.txt
+++ b/ggml/src/CMakeLists.txt
@ -234,6 +234,7 @@ function(ggml_add_backend_library backend)
        # write the shared library to the output directory
        set_target_properties(${backend} PROPERTIES LIBRARY_OUTPUT_DIRECTORY ${CMAKE_RUNTIME_OUTPUT_DIRECTORY})
        target_compile_definitions(${backend} PRIVATE GGML_BACKEND_DL)
+        add_dependencies(ggml ${backend})
    else()
        add_library(${backend} ${ARGN})
        target_link_libraries(ggml PUBLIC ${backend})
--- a/ggml/src/ggml-backend-reg.cpp
+++ b/ggml/src/ggml-backend-reg.cpp
@ -66,6 +66,26 @@
 #include "ggml-kompute.h"
 #endif

+// disable C++17 deprecation warning for std::codecvt_utf8
+#if defined(__clang__)
+#    pragma clang diagnostic push
+#    pragma clang diagnostic ignored "-Wdeprecated-declarations"
+#endif
+
+static std::wstring utf8_to_utf16(const std::string & str) {
+    std::wstring_convert<std::codecvt_utf8_utf16<wchar_t>> converter;
+    return converter.from_bytes(str);
+}
+
+static std::string utf16_to_utf8(const std::wstring & str) {
+    std::wstring_convert<std::codecvt_utf8_utf16<wchar_t>> converter;
+    return converter.to_bytes(str);
+}
+
+#if defined(__clang__)
+#    pragma clang diagnostic pop
+#endif
+
 #ifdef _WIN32

 using dl_handle = std::remove_pointer_t<HMODULE>;
@ -88,11 +108,6 @@ static dl_handle * dl_load_library(const std::wstring & path) {
    return handle;
 }

-static dl_handle * dl_load_library(const std::string & path) {
-    std::wstring_convert<std::codecvt_utf8_utf16<wchar_t>> converter;
-    return dl_load_library(converter.from_bytes(path));
-}
-
 static void * dl_get_sym(dl_handle * handle, const char * name) {
    DWORD old_mode = SetErrorMode(SEM_FAILCRITICALERRORS);
    SetErrorMode(old_mode | SEM_FAILCRITICALERRORS);
@ -114,8 +129,8 @@ struct dl_handle_deleter {
    }
 };

-static void * dl_load_library(const std::string & path) {
-    dl_handle * handle = dlopen(path.c_str(), RTLD_NOW | RTLD_LOCAL);
+static void * dl_load_library(const std::wstring & path) {
+    dl_handle * handle = dlopen(utf16_to_utf8(path).c_str(), RTLD_NOW | RTLD_LOCAL);

    return handle;
 }
@ -202,11 +217,11 @@ struct ggml_backend_registry {
        devices.push_back(device);
    }

-    ggml_backend_reg_t load_backend(const char * path, bool silent) {
+    ggml_backend_reg_t load_backend(const std::wstring & path, bool silent) {
        dl_handle_ptr handle { dl_load_library(path) };
        if (!handle) {
            if (!silent) {
-                GGML_LOG_ERROR("%s: failed to load %s\n", __func__, path);
+                GGML_LOG_ERROR("%s: failed to load %s\n", __func__, utf16_to_utf8(path).c_str());
            }
            return nullptr;
        }
@ -214,7 +229,7 @@ struct ggml_backend_registry {
        auto score_fn = (ggml_backend_score_t) dl_get_sym(handle.get(), "ggml_backend_score");
        if (score_fn && score_fn() == 0) {
            if (!silent) {
-                GGML_LOG_INFO("%s: backend %s is not supported on this system\n", __func__, path);
+                GGML_LOG_INFO("%s: backend %s is not supported on this system\n", __func__, utf16_to_utf8(path).c_str());
            }
            return nullptr;
        }
@ -222,7 +237,7 @@ struct ggml_backend_registry {
        auto backend_init_fn = (ggml_backend_init_t) dl_get_sym(handle.get(), "ggml_backend_init");
        if (!backend_init_fn) {
            if (!silent) {
-                GGML_LOG_ERROR("%s: failed to find ggml_backend_init in %s\n", __func__, path);
+                GGML_LOG_ERROR("%s: failed to find ggml_backend_init in %s\n", __func__, utf16_to_utf8(path).c_str());
            }
            return nullptr;
        }
@ -231,16 +246,16 @@ struct ggml_backend_registry {
        if (!reg || reg->api_version != GGML_BACKEND_API_VERSION) {
            if (!silent) {
                if (!reg) {
-                    GGML_LOG_ERROR("%s: failed to initialize backend from %s: ggml_backend_init returned NULL\n", __func__, path);
+                    GGML_LOG_ERROR("%s: failed to initialize backend from %s: ggml_backend_init returned NULL\n", __func__, utf16_to_utf8(path).c_str());
                } else {
                    GGML_LOG_ERROR("%s: failed to initialize backend from %s: incompatible API version (backend: %d, current: %d)\n",
-                        __func__, path, reg->api_version, GGML_BACKEND_API_VERSION);
+                        __func__, utf16_to_utf8(path).c_str(), reg->api_version, GGML_BACKEND_API_VERSION);
                }
            }
            return nullptr;
        }

-        GGML_LOG_INFO("%s: loaded %s backend from %s\n", __func__, ggml_backend_reg_name(reg), path);
+        GGML_LOG_INFO("%s: loaded %s backend from %s\n", __func__, ggml_backend_reg_name(reg), utf16_to_utf8(path).c_str());

        register_backend(reg, std::move(handle));

@ -376,14 +391,14 @@ ggml_backend_t ggml_backend_init_best(void) {

 // Dynamic loading
 ggml_backend_reg_t ggml_backend_load(const char * path) {
-    return get_reg().load_backend(path, false);
+    return get_reg().load_backend(utf8_to_utf16(path), false);
 }

 void ggml_backend_unload(ggml_backend_reg_t reg) {
    get_reg().unload_backend(reg, true);
 }

-static std::string get_executable_path() {
+static std::wstring get_executable_path() {
 #if defined(__APPLE__)
    // get executable path
    std::vector<char> path;
@ -401,7 +416,7 @@ static std::string get_executable_path() {
    if (last_slash != std::string::npos) {
        base_path = base_path.substr(0, last_slash);
    }
-    return base_path + "/";
+    return utf8_to_utf16(base_path + "/");
 #elif defined(__linux__) || defined(__FreeBSD__)
    std::string base_path = ".";
    std::vector<char> path(1024);
@ -427,57 +442,63 @@ static std::string get_executable_path() {
        path.resize(path.size() * 2);
    }

-    return base_path + "/";
+    return utf8_to_utf16(base_path + "/");
 #elif defined(_WIN32)
-    std::vector<char> path(MAX_PATH);
-    DWORD len = GetModuleFileNameA(NULL, path.data(), path.size());
+    std::vector<wchar_t> path(MAX_PATH);
+    DWORD len = GetModuleFileNameW(NULL, path.data(), path.size());
    if (len == 0) {
-        return "";
+        return {};
    }
-    std::string base_path(path.data(), len);
+    std::wstring base_path(path.data(), len);
    // remove executable name
    auto last_slash = base_path.find_last_of('\\');
    if (last_slash != std::string::npos) {
        base_path = base_path.substr(0, last_slash);
    }
-    return base_path + "\\";
+    return base_path + L"\\";
+#else
+    return {};
 #endif
 }

-static std::string backend_filename_prefix() {
+static std::wstring backend_filename_prefix() {
 #ifdef _WIN32
-    return "ggml-";
+    return L"ggml-";
 #else
-    return "libggml-";
+    return L"libggml-";
 #endif
 }

-static std::string backend_filename_suffix() {
+static std::wstring backend_filename_suffix() {
 #ifdef _WIN32
-    return ".dll";
+    return L".dll";
 #else
-    return ".so";
+    return L".so";
+#endif
+}
+
+static std::wstring path_separator() {
+#ifdef _WIN32
+    return L"\\";
+#else
+    return L"/";
 #endif
 }

 static ggml_backend_reg_t ggml_backend_load_best(const char * name, bool silent, const char * user_search_path) {
    // enumerate all the files that match [lib]ggml-name-*.[so|dll] in the search paths
     // TODO: search system paths
-    std::string file_prefix = backend_filename_prefix() + name + "-";
-    std::vector<std::string> search_paths;
+    std::wstring file_prefix = backend_filename_prefix() + utf8_to_utf16(name) + L"-";
+    std::vector<std::wstring> search_paths;
    if (user_search_path == nullptr) {
-        search_paths.push_back("./");
+        search_paths.push_back(L"." + path_separator());
        search_paths.push_back(get_executable_path());
    } else {
-#if defined(_WIN32)
-        search_paths.push_back(std::string(user_search_path) + "\\");
-#else
-        search_paths.push_back(std::string(user_search_path) + "/");
-#endif
+        search_paths.push_back(utf8_to_utf16(user_search_path) + path_separator());
    }

    int best_score = 0;
-    std::string best_path;
+    std::wstring best_path;

    namespace fs = std::filesystem;
    for (const auto & search_path : search_paths) {
@ -487,27 +508,27 @@ static ggml_backend_reg_t ggml_backend_load_best(const char * name, bool silent,
        fs::directory_iterator dir_it(search_path, fs::directory_options::skip_permission_denied);
        for (const auto & entry : dir_it) {
            if (entry.is_regular_file()) {
-                std::string filename = entry.path().filename().string();
-                std::string ext = entry.path().extension().string();
+                std::wstring filename = entry.path().filename().wstring();
+                std::wstring ext = entry.path().extension().wstring();
                if (filename.find(file_prefix) == 0 && ext == backend_filename_suffix()) {
-                    dl_handle_ptr handle { dl_load_library(entry.path().c_str()) };
+                    dl_handle_ptr handle { dl_load_library(entry.path().wstring()) };
                    if (!handle && !silent) {
-                        GGML_LOG_ERROR("%s: failed to load %s\n", __func__, entry.path().string().c_str());
+                        GGML_LOG_ERROR("%s: failed to load %s\n", __func__, utf16_to_utf8(entry.path().wstring()).c_str());
                    }
                    if (handle) {
                        auto score_fn = (ggml_backend_score_t) dl_get_sym(handle.get(), "ggml_backend_score");
                        if (score_fn) {
                            int s = score_fn();
 #ifndef NDEBUG
-                            GGML_LOG_DEBUG("%s: %s score: %d\n", __func__, entry.path().string().c_str(), s);
+                            GGML_LOG_DEBUG("%s: %s score: %d\n", __func__, utf16_to_utf8(entry.path().wstring()).c_str(), s);
 #endif
                            if (s > best_score) {
                                best_score = s;
-                                best_path = entry.path().string();
+                                best_path = entry.path().wstring();
                            }
                        } else {
                            if (!silent) {
-                                GGML_LOG_INFO("%s: failed to find ggml_backend_score in %s\n", __func__, entry.path().string().c_str());
+                                GGML_LOG_INFO("%s: failed to find ggml_backend_score in %s\n", __func__, utf16_to_utf8(entry.path().wstring()).c_str());
                            }
                        }
                    }
@ -519,15 +540,15 @@ static ggml_backend_reg_t ggml_backend_load_best(const char * name, bool silent,
    if (best_score == 0) {
        // try to load the base backend
        for (const auto & search_path : search_paths) {
-            std::string path = search_path + backend_filename_prefix() + name + backend_filename_suffix();
+            std::wstring path = search_path + backend_filename_prefix() + utf8_to_utf16(name) + backend_filename_suffix();
            if (fs::exists(path)) {
-                return get_reg().load_backend(path.c_str(), silent);
+                return get_reg().load_backend(path, silent);
            }
        }
        return nullptr;
    }

-    return get_reg().load_backend(best_path.c_str(), silent);
+    return get_reg().load_backend(best_path, silent);
 }

 void ggml_backend_load_all() {
--- a/ggml/src/ggml-cpu/CMakeLists.txt
+++ b/ggml/src/ggml-cpu/CMakeLists.txt
@ -135,14 +135,20 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
            endif()

            # show enabled features
+            if (CMAKE_HOST_SYSTEM_NAME STREQUAL "Windows")
+                set(FEAT_INPUT_FILE "NUL")
+            else()
+                set(FEAT_INPUT_FILE "/dev/null")
+            endif()
+
            execute_process(
                COMMAND ${CMAKE_C_COMPILER} ${ARCH_FLAGS} -dM -E -
-                INPUT_FILE "/dev/null"
+                INPUT_FILE ${FEAT_INPUT_FILE}
                OUTPUT_VARIABLE ARM_FEATURE
                RESULT_VARIABLE ARM_FEATURE_RESULT
            )
            if (ARM_FEATURE_RESULT)
-                message(FATAL_ERROR "Failed to get ARM features")
+                message(WARNING "Failed to get ARM features")
            else()
                foreach(feature DOTPROD SVE MATMUL_INT8 FMA FP16_VECTOR_ARITHMETIC)
                    string(FIND "${ARM_FEATURE}" "__ARM_FEATURE_${feature} 1" feature_pos)
@ -317,6 +323,11 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
    target_compile_definitions(${GGML_CPU_NAME} PRIVATE ${ARCH_DEFINITIONS})

    if (GGML_BACKEND_DL)
+        if (GGML_NATIVE)
+            # the feature check relies on ARCH_DEFINITIONS, but it is not set with GGML_NATIVE
+            message(FATAL_ERROR "GGML_NATIVE is not compatible with GGML_BACKEND_DL, consider using GGML_CPU_ALL_VARIANTS")
+        endif()
+
        # The feature detection code is compiled as a separate target so that
        # it can be built without the architecture flags
        # Since multiple variants of the CPU backend may be included in the same
--- a/ggml/src/ggml-cpu/ggml-cpu.c
+++ b/ggml/src/ggml-cpu/ggml-cpu.c
@ -7419,14 +7419,14 @@ static void ggml_compute_forward_mul_mat(
    if (src1_cont) {
        for (int64_t i13 = 0; i13 < ne13; i13++)
            for (int64_t i12 = 0; i12 < ne12; i12++)
-                if (!llamafile_sgemm(ne01, ne11, ne00/ggml_blck_size(src0->type),
+                if (!llamafile_sgemm(params,
+                                     ne01, ne11, ne00/ggml_blck_size(src0->type),
                                     (const char *)src0->data + i12/r2*nb02 + i13/r3*nb03,
                                     nb01/ggml_type_size(src0->type),
                                     (const char *)src1->data + i12*nb12 + i13*nb13,
                                     nb11/ggml_type_size(src1->type),
                                     (char *)dst->data + i12*nb2 + i13*nb3,
                                     nb1/ggml_type_size(dst->type),
-                                     ith, nth,
                                     src0->type,
                                     src1->type,
                                     dst->type))
@ -7471,14 +7471,14 @@ UseGgmlGemm1:;

        for (int64_t i13 = 0; i13 < ne13; i13++)
            for (int64_t i12 = 0; i12 < ne12; i12++)
-                if (!llamafile_sgemm(ne01, ne11, ne00/ggml_blck_size(src0->type),
+                if (!llamafile_sgemm(params,
+                                     ne01, ne11, ne00/ggml_blck_size(src0->type),
                                     (const char *)src0->data + i12/r2*nb02 + i13/r3*nb03,
                                     nb01/ggml_type_size(src0->type),
                                     (const char *)wdata + (i12*ne11 + i13*ne12*ne11)*row_size,
                                     row_size/ggml_type_size(vec_dot_type),
                                     (char *)dst->data + i12*nb2 + i13*nb3,
                                     nb1/ggml_type_size(dst->type),
-                                     ith, nth,
                                     src0->type,
                                     vec_dot_type,
                                     dst->type))
--- a/ggml/src/ggml-cpu/llamafile/sgemm.cpp
+++ b/ggml/src/ggml-cpu/llamafile/sgemm.cpp
@ -53,6 +53,8 @@
 #include "ggml-cpu-impl.h"
 #include "ggml-quants.h"

+#include <atomic>
+
 #ifdef _MSC_VER
 #define NOINLINE __declspec(noinline)
 #else
@ -134,6 +136,16 @@ inline __m512 madd(__m512 a, __m512 b, __m512 c) {
    return _mm512_fmadd_ps(a, b, c);
 }
 #endif
+#if defined(__AVX512BF16__)
+template <>
+inline __m512 madd(__m512bh a, __m512bh b, __m512 c) {
+    return _mm512_dpbf16_ps(c, a, b);
+}
+template <>
+inline __m256 madd(__m256bh a, __m256bh b, __m256 c) {
+    return _mm256_dpbf16_ps(c, a, b);
+}
+#endif
 #endif

 #if defined(__ARM_FEATURE_FMA)
@ -226,6 +238,13 @@ template <> inline __m256 load(const float *p) {
 }
 #endif // __AVX__

+#if defined(__AVX2__) || defined(__AVX512F__)
+template <> inline __m256 load(const ggml_bf16_t *p) {
+    return _mm256_castsi256_ps(
+        _mm256_slli_epi32(_mm256_cvtepu16_epi32(_mm_loadu_si128((const __m128i *)p)), 16));
+}
+#endif // __AVX2__
+
 #if defined(__F16C__)
 template <> inline __m256 load(const ggml_fp16_t *p) {
    return _mm256_cvtph_ps(_mm_loadu_si128((const __m128i *)p));
@ -239,8 +258,27 @@ template <> inline __m512 load(const float *p) {
 template <> inline __m512 load(const ggml_fp16_t *p) {
    return _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i *)p));
 }
+template <> inline __m512 load(const ggml_bf16_t *p) {
+    return _mm512_castsi512_ps(
+        _mm512_slli_epi32(_mm512_cvtepu16_epi32(_mm256_loadu_si256((const __m256i *)p)), 16));
+}
 #endif // __AVX512F__

+#if defined(__AVX512BF16__)
+template <> inline __m512bh load(const ggml_bf16_t *p) {
+    return (__m512bh)_mm512_loadu_ps((const float *)p);
+}
+template <> inline __m256bh load(const ggml_bf16_t *p) {
+    return (__m256bh)_mm256_loadu_ps((const float *)p);
+}
+template <> inline __m512bh load(const float *p) {
+    return _mm512_cvtne2ps_pbh(_mm512_loadu_ps(p + 16), _mm512_loadu_ps(p));
+}
+template <> inline __m256bh load(const float *p) {
+    return _mm512_cvtneps_pbh(_mm512_loadu_ps(p));
+}
+#endif
+
 ////////////////////////////////////////////////////////////////////////////////////////////////////
 // CONSTANTS

@ -252,199 +290,170 @@ static const __m128i iq4nlt = _mm_loadu_si128((const __m128i *) kvalues_iq4nl);
 ////////////////////////////////////////////////////////////////////////////////////////////////////
 // FLOATING POINT MATRIX MULTIPLICATION

+template <int M>
+static inline int64_t BLOCK_SIZE(size_t m) {
+    const int64_t NB_BLOC_M = (m + M - 1) / M;
+    return (m % NB_BLOC_M == 0) ? m / NB_BLOC_M : (m / NB_BLOC_M) + 1;
+}
+
+static constexpr inline int64_t BLOC_POS(int64_t ib, int64_t ibN, int64_t bloc_size) {
+    return ib < ibN ? ib * bloc_size : ibN * bloc_size + (ib - ibN) * (bloc_size - 1);
+}
+
 template <int KN, typename D, typename V, typename TA, typename TB, typename TC>
 class tinyBLAS {
  public:
-    tinyBLAS(int64_t k,
+    tinyBLAS(const ggml_compute_params * params, int64_t k,
             const TA *A, int64_t lda,
             const TB *B, int64_t ldb,
-             TC *C, int64_t ldc,
-             int ith, int nth)
-        : A(A), B(B), C(C), k(k), lda(lda), ldb(ldb), ldc(ldc), ith(ith), nth(nth) {
+             TC *C, int64_t ldc)
+        : params(params), A(A), B(B), C(C), k(k), lda(lda), ldb(ldb), ldc(ldc) {
    }

-    void matmul(int64_t m, int64_t n) {
-        mnpack(0, m, 0, n);
+    bool matmul(int64_t m, int64_t n) {
+        if (k % KN != 0)
+            return false;
+        // compute RM for only need tile with size RM&RM-1
+#if VECTOR_REGISTERS == 32
+        if (m % 16 == 0 && (m/16 >= params->nth)) {
+            const int64_t SIZE_N = BLOCK_SIZE<6>(n);
+            mnpack<4, 6, 4>(m, n, SIZE_N, 12);
+            return true;
+        }
+        if (m % 8 == 0 ) {
+            const int64_t SIZE_N = BLOCK_SIZE<6>(n);
+            mnpack<4, 6, 2>(m, n, SIZE_N, 12);
+            return true;
+        }
+        if (m % 4 == 0) {
+            const int64_t SIZE_N = BLOCK_SIZE<6>(n);
+            mnpack<4, 6, 1>(m, n, SIZE_N, 12);
+            return true;
+        }
+#else  // VECTOR_REGISTERS == 16
+        if (m % 16 == 0 && (m/16 >= params->nth)) {
+            const int64_t SIZE_N = BLOCK_SIZE<3>(n);
+            mnpack<4, 3, 4>(m, n, SIZE_N, 24);
+            return true;
+        }
+        if (m % 8 == 0 ) {
+            const int64_t SIZE_N = BLOCK_SIZE<3>(n);
+            mnpack<4, 3, 2>(m, n, SIZE_N, 24);
+            return true;
+        }
+        if (m % 4 == 0) {
+            const int64_t SIZE_N = BLOCK_SIZE<3>(n);
+            mnpack<4, 3, 1>(m, n, SIZE_N, 24);
+            return true;
+        }
+#endif
+        return false;
    }

  private:
-    NOINLINE void mnpack(int64_t m0, int64_t m, int64_t n0, int64_t n) {
-        int64_t mc, nc, mp, np;
-        switch ((MIN(m - m0, 5) << 4) | MIN(n - n0, 5)) {
-#if VECTOR_REGISTERS == 32
-        case 0x55:
-            mc = 5;
-            nc = 5;
-            gemm<5, 5>(m0, m, n0, n);
-            break;
-        case 0x45:
-            mc = 4;
-            nc = 5;
-            gemm<4, 5>(m0, m, n0, n);
-            break;
-        case 0x54:
-            mc = 5;
-            nc = 4;
-            gemm<5, 4>(m0, m, n0, n);
-            break;
-        case 0x44:
-            mc = 4;
-            nc = 4;
-            gemm<4, 4>(m0, m, n0, n);
-            break;
-        case 0x53:
-            mc = 5;
-            nc = 3;
-            gemm<5, 3>(m0, m, n0, n);
-            break;
-        case 0x35:
-            mc = 3;
-            nc = 5;
-            gemm<3, 5>(m0, m, n0, n);
-            break;
-        case 0x43:
-            mc = 4;
-            nc = 3;
-            gemm<4, 3>(m0, m, n0, n);
-            break;
-#else
-        case 0x55:
-        case 0x54:
-        case 0x53:
-        case 0x45:
-        case 0x44:
-        case 0x43:
-            mc = 4;
-            nc = 3;
-            gemm<4, 3>(m0, m, n0, n);
-            break;
-        case 0x35:
-#endif
-        case 0x34:
-            mc = 3;
-            nc = 4;
-            gemm<3, 4>(m0, m, n0, n);
-            break;
-        case 0x52:
-            mc = 5;
-            nc = 2;
-            gemm<5, 2>(m0, m, n0, n);
-            break;
-        case 0x33:
-            mc = 3;
-            nc = 3;
-            gemm<3, 3>(m0, m, n0, n);
-            break;
-        case 0x25:
-            mc = 2;
-            nc = 5;
-            gemm<2, 5>(m0, m, n0, n);
-            break;
-        case 0x42:
-            mc = 4;
-            nc = 2;
-            gemm<4, 2>(m0, m, n0, n);
-            break;
-        case 0x24:
-            mc = 2;
-            nc = 4;
-            gemm<2, 4>(m0, m, n0, n);
-            break;
-        case 0x32:
-            mc = 3;
-            nc = 2;
-            gemm<3, 2>(m0, m, n0, n);
-            break;
-        case 0x23:
-            mc = 2;
-            nc = 3;
-            gemm<2, 3>(m0, m, n0, n);
-            break;
-        case 0x51:
-            mc = 5;
-            nc = 1;
-            gemm<5, 1>(m0, m, n0, n);
-            break;
-        case 0x41:
-            mc = 4;
-            nc = 1;
-            gemm<4, 1>(m0, m, n0, n);
-            break;
-        case 0x22:
-            mc = 2;
-            nc = 2;
-            gemm<2, 2>(m0, m, n0, n);
-            break;
-        case 0x15:
-            mc = 1;
-            nc = 5;
-            gemm<1, 5>(m0, m, n0, n);
-            break;
-        case 0x14:
-            mc = 1;
-            nc = 4;
-            gemm<1, 4>(m0, m, n0, n);
-            break;
-        case 0x31:
-            mc = 3;
-            nc = 1;
-            gemm<3, 1>(m0, m, n0, n);
-            break;
-        case 0x13:
-            mc = 1;
-            nc = 3;
-            gemm<1, 3>(m0, m, n0, n);
-            break;
-        case 0x21:
-            mc = 2;
-            nc = 1;
-            gemm<2, 1>(m0, m, n0, n);
-            break;
-        case 0x12:
-            mc = 1;
-            nc = 2;
-            gemm<1, 2>(m0, m, n0, n);
-            break;
-        case 0x11:
-            mc = 1;
-            nc = 1;
-            gemm<1, 1>(m0, m, n0, n);
-            break;
-        default:
-            return;
+    template <int RM, int RN, int BM>
+    inline void mnpack(int64_t m, int64_t n, int64_t SIZE_N, int64_t BN) {
+        if (SIZE_N == RN) {
+            return gemm<RM, RN, BM>(m, n, BN);
+        }
+        if constexpr (RN > 1) {
+            return mnpack<RM, RN-1, BM>(m, n, SIZE_N, BN);
+        } else {
+            GGML_LOG_ERROR("mnpack<%d, %d> bloc size not supported\n", RM, (int)SIZE_N);
+            GGML_ASSERT(false); // we have miss something.
        }
-        mp = m0 + (m - m0) / mc * mc;
-        np = n0 + (n - n0) / nc * nc;
-        mnpack(mp, m, n0, np);
-        mnpack(m0, m, np, n);
    }

    template <int RM, int RN>
-    NOINLINE void gemm(int64_t m0, int64_t m, int64_t n0, int64_t n) {
-        int64_t ytiles = (m - m0) / RM;
-        int64_t xtiles = (n - n0) / RN;
-        int64_t tiles = xtiles * ytiles;
-        int64_t duty = (tiles + nth - 1) / nth;
-        int64_t start = duty * ith;
-        int64_t end = start + duty;
-        if (end > tiles)
-            end = tiles;
-        for (int64_t job = start; job < end; ++job) {
-            int64_t ii = m0 + job / xtiles * RM;
-            int64_t jj = n0 + job % xtiles * RN;
-            D Cv[RN][RM] = {};
-            for (int64_t l = 0; l < k; l += KN)
-                for (int64_t j = 0; j < RN; ++j)
-                    for (int64_t i = 0; i < RM; ++i)
-                        Cv[j][i] = madd(load<V>(A + lda * (ii + i) + l),
-                                        load<V>(B + ldb * (jj + j) + l),
-                                        Cv[j][i]);
-            for (int64_t j = 0; j < RN; ++j)
-                for (int64_t i = 0; i < RM; ++i)
-                    C[ldc * (jj + j) + (ii + i)] = hsum(Cv[j][i]);
+    inline void gemm_bloc(int64_t ii, int64_t jj) {
+        D Cv[RN][RM] = {};
+        for (int64_t l = 0; l < k; l += KN) {
+            // help compiler for op order.
+            if constexpr (RM <= RN) {
+                V Av[RM];
+                for (int64_t i = 0; i < RM; ++i) {
+                    Av[i] = load<V>(A + lda * (ii + i) + l);
+                }
+                for (int64_t j = 0; j < RN; ++j) {
+                    V Bv = load<V>(B + ldb * (jj + j) + l);
+                    for (int64_t i = 0; i < RM; ++i) {
+                        Cv[j][i] = madd(Av[i], Bv, Cv[j][i]);
+                    }
+                }
+            } else {
+                V Bv[RN];
+                for (int64_t j = 0; j < RN; ++j) {
+                    Bv[j] = load<V>(B + ldb * (jj + j) + l);
+                }
+                for (int64_t i = 0; i < RM; ++i) {
+                    V Av = load<V>(A + lda * (ii + i) + l);
+                    for (int64_t j = 0; j < RN; ++j) {
+                        Cv[j][i] = madd(Av, Bv[j], Cv[j][i]);
+                    }
+                }
+            }
        }
+        for (int64_t j = 0; j < RN; ++j)
+            for (int64_t i = 0; i < RM; ++i)
+                C[ldc * (jj + j) + (ii + i)] = hsum(Cv[j][i]);
    }

+    template <int RM, int RN, int BM>
+    NOINLINE void gemm(int64_t m, int64_t n, int64_t BN) {
+        static std::atomic<int64_t> current_chunk;
+
+        GGML_ASSERT(m % (RM * BM) == 0);
+        const int64_t ytiles = m / (RM * BM);
+        const int64_t xtiles = (n + RN -1) / RN;
+        const int64_t jj_RN = (xtiles - (xtiles * RN - n));
+
+        // "round" bloc_size to "nearest" BN
+        const int64_t NB_BN = xtiles < BN ? 1 : (xtiles + BN / 2) / BN;
+        const int64_t SIZE_BN = xtiles % NB_BN == 0 ? xtiles / NB_BN : xtiles / NB_BN + 1;
+        const int64_t jj_BN = (NB_BN - (NB_BN * SIZE_BN - xtiles));
+        const int64_t nb_job = ytiles * NB_BN;
+
+        if (params->ith == 0) {
+            GGML_ASSERT( jj_BN * SIZE_BN + (NB_BN - jj_BN) * (SIZE_BN - 1) == xtiles);
+            // Every thread starts at ith, so the first unprocessed chunk is nth.  This save a bit of coordination right at the start.
+            std::atomic_store_explicit(&current_chunk, (int64_t)params->nth, std::memory_order_relaxed);
+        }
+
+        ggml_barrier(params->threadpool);
+
+        int64_t job = params->ith;
+        while (job < nb_job) {
+            const int64_t ii = (job % ytiles) * RM * BM;
+            const int64_t jb =  job / ytiles;
+            const int64_t jr0 = BLOC_POS(jb  , jj_BN, SIZE_BN);
+            const int64_t jrN = BLOC_POS(jb+1, jj_BN, SIZE_BN);
+
+            const int64_t jj0 = BLOC_POS(jr0, jj_RN, RN);
+            const int64_t jj2 = BLOC_POS(jrN, jj_RN, RN);
+            const int64_t jj1 = jj2 < jj_RN * RN ? jj2 : jj_RN * RN;
+
+            for (int64_t bi = 0; bi < BM * RM; bi += RM) {
+                int64_t jj = jj0;
+                for (; jj < jj1; jj += RN) {
+                    gemm_bloc<RM, RN>(ii + bi, jj);
+                }
+                if constexpr (RN > 1) {
+                    for (; jj < jj2; jj += RN - 1) {
+                        gemm_bloc<RM, RN-1>(ii + bi, jj);
+                    }
+                }
+                GGML_ASSERT(jj == jj2);
+            }
+
+            // next step.
+            job = std::atomic_fetch_add_explicit(&current_chunk, (int64_t)1, std::memory_order_relaxed);
+        }
+
+        ggml_barrier(params->threadpool);
+        return;
+    }
+
+    const ggml_compute_params * params;
    const TA *const A;
    const TB *const B;
    TC *const C;
@ -452,8 +461,6 @@ class tinyBLAS {
    const int64_t lda;
    const int64_t ldb;
    const int64_t ldc;
-    const int ith;
-    const int nth;
 };

 //////////////////////////////////////////////////////////////////////////////////////////
@ -1657,8 +1664,9 @@ class tinyBLAS_PPC {
 * @param Ctype is GGML data type of `C`
 * @return true if this function was able to service the matmul request
 */
-bool llamafile_sgemm(int64_t m, int64_t n, int64_t k, const void *A, int64_t lda, const void *B, int64_t ldb, void *C,
-                     int64_t ldc, int ith, int nth, int Atype, int Btype, int Ctype) {
+bool llamafile_sgemm(const struct ggml_compute_params * params, int64_t m, int64_t n, int64_t k,
+                     const void *A, int64_t lda, const void *B, int64_t ldb, void *C,
+                     int64_t ldc, int Atype, int Btype, int Ctype) {

    assert(m >= 0);
    assert(n >= 0);
@ -1666,8 +1674,8 @@ bool llamafile_sgemm(int64_t m, int64_t n, int64_t k, const void *A, int64_t lda
    assert(lda >= k);
    assert(ldb >= k);
    assert(ldc >= m);
-    assert(nth > 0);
-    assert(ith < nth);
+    assert(params->nth > 0);
+    assert(params->ith < params->nth);

    // only enable sgemm for prompt processing
    if (n < 2)
@ -1682,37 +1690,25 @@ bool llamafile_sgemm(int64_t m, int64_t n, int64_t k, const void *A, int64_t lda
        if (Btype != GGML_TYPE_F32)
            return false;
 #if defined(__AVX512F__)
-        if (k % 16)
-            return false;
-        tinyBLAS<16, __m512, __m512, float, float, float> tb{
+        tinyBLAS<16, __m512, __m512, float, float, float> tb{ params,
            k, (const float *)A, lda,
            (const float *)B, ldb,
-            (float *)C, ldc,
-            ith, nth};
-        tb.matmul(m, n);
-        return true;
+            (float *)C, ldc};
+        return tb.matmul(m, n);
 #elif defined(__AVX__) || defined(__AVX2__)
-        if (k % 8)
-            return false;
-        tinyBLAS<8, __m256, __m256, float, float, float> tb{
+        tinyBLAS<8, __m256, __m256, float, float, float> tb{ params,
            k, (const float *)A, lda,
            (const float *)B, ldb,
-            (float *)C, ldc,
-            ith, nth};
-        tb.matmul(m, n);
-        return true;
+            (float *)C, ldc};
+        return tb.matmul(m, n);
 #elif defined(__ARM_NEON)
        if (n < 4)
            return false;
-        if (k % 4)
-            return false;
-        tinyBLAS<4, float32x4_t, float32x4_t, float, float, float> tb{
+        tinyBLAS<4, float32x4_t, float32x4_t, float, float, float> tb{ params,
            k, (const float *)A, lda,
            (const float *)B, ldb,
-            (float *)C, ldc,
-            ith, nth};
-        tb.matmul(m, n);
-        return true;
+            (float *)C, ldc};
+        return tb.matmul(m, n);
 #elif defined(__MMA__)
        if (k % 8)
            return false;
@ -1720,7 +1716,7 @@ bool llamafile_sgemm(int64_t m, int64_t n, int64_t k, const void *A, int64_t lda
            k, (const float *)A, lda,
            (const float *)B, ldb,
            (float *)C, ldc,
-            ith, nth};
+            params->ith, params->nth};
        tb.matmul(m, n);
        return true;
 #else
@ -1728,60 +1724,71 @@ bool llamafile_sgemm(int64_t m, int64_t n, int64_t k, const void *A, int64_t lda
 #endif
    }

+    case GGML_TYPE_BF16: {
+#if defined(__AVX512BF16__)
+        if (Btype == GGML_TYPE_BF16) {
+            tinyBLAS<32, __m512, __m512bh, ggml_bf16_t, ggml_bf16_t, float> tb{ params, k,
+                (const ggml_bf16_t *)A, lda,
+                (const ggml_bf16_t *)B, ldb,
+                (float *)C, ldc};
+            return tb.matmul(m, n);
+        }
+#elif defined(__AVX512F__)
+        if (Btype == GGML_TYPE_BF16) {
+            tinyBLAS<16, __m512, __m512, ggml_bf16_t, ggml_bf16_t, float> tb{ params, k,
+                (const ggml_bf16_t *)A, lda,
+                (const ggml_bf16_t *)B, ldb,
+                (float *)C, ldc};
+            return tb.matmul(m, n);
+        }
+#elif defined(__AVX2__)
+        if (Btype == GGML_TYPE_BF16) {
+            tinyBLAS<8, __m256, __m256, ggml_bf16_t, ggml_bf16_t, float> tb{ params, k,
+                (const ggml_bf16_t *)A, lda,
+                (const ggml_bf16_t *)B, ldb,
+                (float *)C, ldc};
+            return tb.matmul(m, n);
+        }
+#endif
+        return false;
+    }
    case GGML_TYPE_F16: {
 #if defined(__AVX512F__)
-        if (k % 16)
-            return false;
-        if (Btype != GGML_TYPE_F32)
-            return false;
-        tinyBLAS<16, __m512, __m512, ggml_fp16_t, float, float> tb{
-            k, (const ggml_fp16_t *)A, lda,
-            (const float *)B, ldb,
-            (float *)C, ldc,
-            ith, nth};
-        tb.matmul(m, n);
-        return true;
+        if (Btype == GGML_TYPE_F16) {
+            tinyBLAS<16, __m512, __m512, ggml_fp16_t, ggml_fp16_t, float> tb{ params, k,
+                (const ggml_fp16_t *)A, lda,
+                (const ggml_fp16_t *)B, ldb,
+                (float *)C, ldc};
+            return tb.matmul(m, n);
+        }
 #elif (defined(__AVX__) || defined(__AVX2__)) && defined(__F16C__)
-        if (k % 8)
-            return false;
-        if (Btype != GGML_TYPE_F32)
-            return false;
-        tinyBLAS<8, __m256, __m256, ggml_fp16_t, float, float> tb{
-            k, (const ggml_fp16_t *)A, lda,
-            (const float *)B, ldb,
-            (float *)C, ldc,
-            ith, nth};
-        tb.matmul(m, n);
-        return true;
+        if (Btype == GGML_TYPE_F16) {
+            tinyBLAS<8, __m256, __m256, ggml_fp16_t, ggml_fp16_t, float> tb{ params, k,
+                (const ggml_fp16_t *)A, lda,
+                (const ggml_fp16_t *)B, ldb,
+                (float *)C, ldc};
+            return tb.matmul(m, n);
+        }
 #elif defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && !defined(_MSC_VER)
        if (n < 8)
            return false;
-        if (k % 8)
-            return false;
-        if (Btype != GGML_TYPE_F16)
-            return false;
-        tinyBLAS<8, float16x8_t, float16x8_t, ggml_fp16_t, ggml_fp16_t, float> tb{
-            k, (const ggml_fp16_t *)A, lda,
-            (const ggml_fp16_t *)B, ldb,
-            (float *)C, ldc,
-            ith, nth};
-        tb.matmul(m, n);
-        return true;
+        if (Btype == GGML_TYPE_F16) {
+            tinyBLAS<8, float16x8_t, float16x8_t, ggml_fp16_t, ggml_fp16_t, float> tb{ params,
+                k, (const ggml_fp16_t *)A, lda,
+                (const ggml_fp16_t *)B, ldb,
+                (float *)C, ldc};
+            return tb.matmul(m, n);
+        }
 #elif defined(__ARM_NEON) && !defined(_MSC_VER)
-        if (k % 4)
-            return false;
-        if (Btype != GGML_TYPE_F32)
-            return false;
-        tinyBLAS<4, float32x4_t, float32x4_t, ggml_fp16_t, float, float> tb{
-            k, (const ggml_fp16_t *)A, lda,
-            (const float *)B, ldb,
-            (float *)C, ldc,
-            ith, nth};
-        tb.matmul(m, n);
-        return true;
-#else
-        return false;
+        if (Btype == GGML_TYPE_F32) {
+            tinyBLAS<4, float32x4_t, float32x4_t, ggml_fp16_t, float, float> tb{ params,
+                k, (const ggml_fp16_t *)A, lda,
+                (const float *)B, ldb,
+                (float *)C, ldc};
+            return tb.matmul(m, n);
+        }
 #endif
+        return false;
    }

    case GGML_TYPE_Q8_0: {
@ -1792,7 +1799,7 @@ bool llamafile_sgemm(int64_t m, int64_t n, int64_t k, const void *A, int64_t lda
            k, (const block_q8_0 *)A, lda,
            (const block_q8_0 *)B, ldb,
            (float *)C, ldc,
-            ith, nth};
+            params->ith, params->nth};
        tb.matmul(m, n);
        return true;
 #elif defined(__ARM_FEATURE_DOTPROD)
@ -1800,7 +1807,7 @@ bool llamafile_sgemm(int64_t m, int64_t n, int64_t k, const void *A, int64_t lda
            k, (const block_q8_0 *)A, lda,
            (const block_q8_0 *)B, ldb,
            (float *)C, ldc,
-            ith, nth};
+            params->ith, params->nth};
        tb.matmul(m, n);
        return true;
 #else
@ -1816,7 +1823,7 @@ bool llamafile_sgemm(int64_t m, int64_t n, int64_t k, const void *A, int64_t lda
            k, (const block_q4_0 *)A, lda,
            (const block_q8_0 *)B, ldb,
            (float *)C, ldc,
-            ith, nth};
+            params->ith, params->nth};
        tb.matmul(m, n);
        return true;
 #elif defined(__ARM_FEATURE_DOTPROD)
@ -1824,7 +1831,7 @@ bool llamafile_sgemm(int64_t m, int64_t n, int64_t k, const void *A, int64_t lda
            k, (const block_q4_0 *)A, lda,
            (const block_q8_0 *)B, ldb,
            (float *)C, ldc,
-            ith, nth};
+            params->ith, params->nth};
        tb.matmul(m, n);
        return true;
 #else
@ -1840,7 +1847,7 @@ bool llamafile_sgemm(int64_t m, int64_t n, int64_t k, const void *A, int64_t lda
            k, (const block_q5_0 *)A, lda,
            (const block_q8_0 *)B, ldb,
            (float *)C, ldc,
-            ith, nth};
+            params->ith, params->nth};
        tb.matmul(m, n);
        return true;
 #else
@ -1856,7 +1863,7 @@ bool llamafile_sgemm(int64_t m, int64_t n, int64_t k, const void *A, int64_t lda
            k, (const block_iq4_nl *)A, lda,
            (const block_q8_0 *)B, ldb,
            (float *)C, ldc,
-            ith, nth};
+            params->ith, params->nth};
        tb.matmul(m, n);
        return true;
 #else
@ -1868,6 +1875,7 @@ bool llamafile_sgemm(int64_t m, int64_t n, int64_t k, const void *A, int64_t lda
        return false;
    }

+    (void)params;
    (void)m;
    (void)n;
    (void)k;
@ -1877,8 +1885,6 @@ bool llamafile_sgemm(int64_t m, int64_t n, int64_t k, const void *A, int64_t lda
    (void)ldb;
    (void)C;
    (void)ldc;
-    (void)ith;
-    (void)nth;
    (void)Atype;
    (void)Btype;
    (void)Ctype;
--- a/ggml/src/ggml-cpu/llamafile/sgemm.h
+++ b/ggml/src/ggml-cpu/llamafile/sgemm.h
@ -5,8 +5,8 @@
 extern "C" {
 #endif

-bool llamafile_sgemm(int64_t, int64_t, int64_t, const void *, int64_t,
-                     const void *, int64_t, void *, int64_t, int, int,
+bool llamafile_sgemm(const struct ggml_compute_params * params, int64_t, int64_t, int64_t,
+                     const void *, int64_t, const void *, int64_t, void *, int64_t,
                     int, int, int);

 #ifdef __cplusplus
--- a/gguf-py/gguf/constants.py
+++ b/gguf-py/gguf/constants.py
@ -1284,6 +1284,9 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
        MODEL_TENSOR.OUTPUT,
        MODEL_TENSOR.ATTN_NORM,
        MODEL_TENSOR.ATTN_QKV,
+        MODEL_TENSOR.ATTN_Q,
+        MODEL_TENSOR.ATTN_K,
+        MODEL_TENSOR.ATTN_V,
        MODEL_TENSOR.ATTN_OUT,
        MODEL_TENSOR.FFN_NORM,
        MODEL_TENSOR.FFN_DOWN,
--- a/scripts/compare-llama-bench.py
+++ b/scripts/compare-llama-bench.py
@ -126,6 +126,8 @@ connection = sqlite3.connect(input_file)
 cursor = connection.cursor()
 builds = cursor.execute("SELECT DISTINCT build_commit FROM test;").fetchall()

+commit_short_len = len(builds[0][0])
+
 try:
    repo = git.Repo(".", search_parent_directories=True)
 except git.InvalidGitRepositoryError:
@ -138,11 +140,11 @@ def find_parent_in_data(commit: git.Commit):
    seen_hexsha8 = set()
    while heap:
        depth, current_commit = heapq.heappop(heap)
-        current_hexsha8 = commit.hexsha[:8]
+        current_hexsha8 = commit.hexsha[:commit_short_len]
        if (current_hexsha8,) in builds:
            return current_hexsha8
        for parent in commit.parents:
-            parent_hexsha8 = parent.hexsha[:8]
+            parent_hexsha8 = parent.hexsha[:commit_short_len]
            if parent_hexsha8 not in seen_hexsha8:
                seen_hexsha8.add(parent_hexsha8)
                heapq.heappush(heap, (depth + 1, parent))
@ -156,9 +158,9 @@ def get_all_parent_hexsha8s(commit: git.Commit):

    while unvisited:
        current_commit = unvisited.pop(0)
-        visited.append(current_commit.hexsha[:8])
+        visited.append(current_commit.hexsha[:commit_short_len])
        for parent in current_commit.parents:
-            if parent.hexsha[:8] not in visited:
+            if parent.hexsha[:commit_short_len] not in visited:
                unvisited.append(parent)

    return visited
@ -169,10 +171,10 @@ def get_commit_name(hexsha8):
    if repo is None:
        return hexsha8
    for h in repo.heads:
-        if h.commit.hexsha[:8] == hexsha8:
+        if h.commit.hexsha[:commit_short_len] == hexsha8:
            return h.name
    for t in repo.tags:
-        if t.commit.hexsha[:8] == hexsha8:
+        if t.commit.hexsha[:commit_short_len] == hexsha8:
            return t.name
    return hexsha8

@ -183,13 +185,13 @@ def get_commit_hexsha8(name):
        return None
    for h in repo.heads:
        if h.name == name:
-            return h.commit.hexsha[:8]
+            return h.commit.hexsha[:commit_short_len]
    for t in repo.tags:
        if t.name == name:
-            return t.commit.hexsha[:8]
+            return t.commit.hexsha[:commit_short_len]
    for c in repo.iter_commits("--all"):
-        if c.hexsha[:8] == name[:8]:
-            return c.hexsha[:8]
+        if c.hexsha[:commit_short_len] == name[:commit_short_len]:
+            return c.hexsha[:commit_short_len]
    return None


--- a/scripts/hf.sh
+++ b/scripts/hf.sh
@ -26,7 +26,7 @@ function has_cmd {
 }

 if has_cmd wget; then
-    cmd="wget -q --show-progress -c -O %s/%s %s"
+    cmd="wget -q -c -O %s/%s %s"
 elif has_cmd curl; then
    cmd="curl -C - -f --output-dir %s -o %s -L %s"
 else
--- a/src/llama-vocab.cpp
+++ b/src/llama-vocab.cpp
@ -1657,7 +1657,7 @@ bool llama_token_is_control_impl(const struct llama_vocab & vocab, llama_token t
 }

 llama_token llama_token_bos_impl(const struct llama_vocab & vocab) {
-    return vocab.special_bos_id;
+    return vocab.type != LLAMA_VOCAB_TYPE_WPM ? vocab.special_bos_id : vocab.special_cls_id;
 }

 llama_token llama_token_eos_impl(const struct llama_vocab & vocab) {
--- a/src/llama-vocab.h
+++ b/src/llama-vocab.h
@ -45,7 +45,7 @@ struct llama_vocab {
    id special_unk_id  = 0;
    id special_sep_id  = LLAMA_TOKEN_NULL;
    id special_pad_id  = LLAMA_TOKEN_NULL;
-    id special_cls_id  = LLAMA_TOKEN_NULL;
+    id special_cls_id  = LLAMA_TOKEN_NULL; // TODO: revisit if this is really needed https://github.com/ggerganov/llama.cpp/pull/10930
    id special_mask_id = LLAMA_TOKEN_NULL;

    id linefeed_id = 13;
--- a/src/llama.cpp
+++ b/src/llama.cpp
@ -1440,6 +1440,9 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
            { LLM_TENSOR_OUTPUT,          "output" },
            { LLM_TENSOR_ATTN_NORM,       "blk.%d.attn_norm" },
            { LLM_TENSOR_ATTN_QKV,        "blk.%d.attn_qkv" },
+            { LLM_TENSOR_ATTN_Q,          "blk.%d.attn_q" },
+            { LLM_TENSOR_ATTN_K,          "blk.%d.attn_k" },
+            { LLM_TENSOR_ATTN_V,          "blk.%d.attn_v" },
            { LLM_TENSOR_ATTN_OUT,        "blk.%d.attn_output" },
            { LLM_TENSOR_FFN_NORM,        "blk.%d.ffn_norm" },
            { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
@ -1715,6 +1718,7 @@ enum llm_chat_template {
    LLM_CHAT_TEMPLATE_LLAMA_3,
    LLM_CHAT_TEMPLATE_CHATGML_3,
    LLM_CHAT_TEMPLATE_CHATGML_4,
+    LLM_CHAT_TEMPLATE_GLMEDGE,
    LLM_CHAT_TEMPLATE_MINICPM,
    LLM_CHAT_TEMPLATE_EXAONE_3,
    LLM_CHAT_TEMPLATE_RWKV_WORLD,
@ -1749,6 +1753,7 @@ static const std::map<std::string, llm_chat_template> LLM_CHAT_TEMPLATES = {
    { "llama3",            LLM_CHAT_TEMPLATE_LLAMA_3           },
    { "chatglm3",          LLM_CHAT_TEMPLATE_CHATGML_3         },
    { "chatglm4",          LLM_CHAT_TEMPLATE_CHATGML_4         },
+    { "glmedge",           LLM_CHAT_TEMPLATE_GLMEDGE           },
    { "minicpm",           LLM_CHAT_TEMPLATE_MINICPM           },
    { "exaone3",           LLM_CHAT_TEMPLATE_EXAONE_3          },
    { "rwkv-world",        LLM_CHAT_TEMPLATE_RWKV_WORLD        },
@ -6300,8 +6305,20 @@ static void llm_load_hparams(
            {
                ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
                switch (hparams.n_layer) {
-                    case 28: model.type = e_model::MODEL_6B; break;
-                    case 40: model.type = e_model::MODEL_9B; break;
+                    case 28: {
+                        if(hparams.n_head(0)==16){
+                            model.type = e_model::MODEL_1_6B;
+                        }else{
+                            model.type = e_model::MODEL_6B;
+                        }
+                    }break;
+                    case 40:{
+                        if(hparams.n_head(0)==24){
+                            model.type = e_model::MODEL_4B;
+                        }else{
+                            model.type = e_model::MODEL_9B;
+                        }
+                    } break;
                    default: model.type = e_model::MODEL_UNKNOWN;
                }
            } break;
@ -9434,9 +9451,14 @@ static bool llm_load_tensors(
                        auto & layer = model.layers[i];

                        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
-
-                        layer.wqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa}, 0);
-                        layer.bqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "bias", i),   {n_embd + 2*n_embd_gqa}, 0);
+                        if(model.type == e_model::MODEL_1_6B || model.type == e_model::MODEL_4B){
+                            layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q,   "weight", i), {n_embd, n_embd_head_k * n_head}, 0);
+                            layer.wk = create_tensor(tn(LLM_TENSOR_ATTN_K,   "weight", i), {n_embd, n_embd_k_gqa}, 0);
+                            layer.wv = create_tensor(tn(LLM_TENSOR_ATTN_V,   "weight", i), {n_embd, n_embd_v_gqa}, 0);
+                        }else{
+                            layer.wqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa}, 0);
+                            layer.bqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "bias", i),   {n_embd + 2*n_embd_gqa}, llama_model_loader::TENSOR_NOT_REQUIRED);
+                        }

                        layer.wo   = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, 0);

@ -16827,20 +16849,28 @@ struct llm_build_context {
                struct ggml_tensor * Qcur = nullptr;
                struct ggml_tensor * Kcur = nullptr;
                struct ggml_tensor * Vcur = nullptr;
+                if(model.type == e_model::MODEL_1_6B || model.type == e_model::MODEL_4B){
+                    Qcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wq, cur);
+                    cb(Qcur, "Qcur", il);
+                    Kcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wk, cur);
+                    cb(Kcur, "Kcur", il);
+                    Vcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wv, cur);
+                    cb(Vcur, "Vcur", il);
+                }else{
+                    cur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wqkv, cur);
+                    cb(cur, "wqkv", il);
+                    if(model.layers[il].bqkv){
+                        cur = ggml_add(ctx0, cur, model.layers[il].bqkv);
+                        cb(cur, "bqkv", il);
+                    }
+                    Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd,     n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd)));
+                    Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd)));
+                    Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa)));
+                    cb(Qcur, "Qcur", il);
+                    cb(Kcur, "Kcur", il);
+                    cb(Vcur, "Vcur", il);
+                }

-                cur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wqkv, cur);
-                cb(cur, "wqkv", il);
-
-                cur = ggml_add(ctx0, cur, model.layers[il].bqkv);
-                cb(cur, "bqkv", il);
-
-                Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd,     n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd)));
-                Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd)));
-                Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa)));
-
-                cb(Qcur, "Qcur", il);
-                cb(Kcur, "Kcur", il);
-                cb(Vcur, "Vcur", il);
                //printf("freq_base: %f freq_scale: %f ext_factor: %f attn_factor: %f\n", freq_base, freq_scale, ext_factor, attn_factor);
                Qcur = ggml_rope_ext(
                    ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, nullptr,
@ -22921,6 +22951,8 @@ static llm_chat_template llama_chat_detect_template(const std::string & tmpl) {
        return LLM_CHAT_TEMPLATE_CHATGML_3;
    } else if (tmpl_contains("[gMASK]<sop>")) {
        return LLM_CHAT_TEMPLATE_CHATGML_4;
+    } else if (tmpl_contains("<|user|>") && tmpl_contains("<|assistant|>") && !tmpl_contains("<|end|>") && !tmpl_contains("</s>")) {
+        return LLM_CHAT_TEMPLATE_GLMEDGE;
    } else if (tmpl_contains(LU8("<用户>"))) {
        // MiniCPM-3B-OpenHermes-2.5-v2-GGUF
        return LLM_CHAT_TEMPLATE_MINICPM;
@ -23204,6 +23236,14 @@ static int32_t llama_chat_apply_template_internal(
        if (add_ass) {
            ss << "<|assistant|>";
        }
+    } else if(tmpl == LLM_CHAT_TEMPLATE_GLMEDGE){
+        for (auto message : chat) {
+            std::string role(message->role);
+            ss << "<|" << role << "|>" << "\n" << message->content;
+        }
+        if (add_ass) {
+            ss << "<|assistant|>";
+        }
    } else if (tmpl == LLM_CHAT_TEMPLATE_MINICPM) {
        // MiniCPM-3B-OpenHermes-2.5-v2-GGUF
        for (auto message : chat) {
--- a/tests/test-chat-template.cpp
+++ b/tests/test-chat-template.cpp
@ -61,6 +61,8 @@ int main(void) {
        "{% for message in messages %}{% if loop.first %}[gMASK]sop<|{{ message['role'] }}|>\n {{ message['content'] }}{% else %}<|{{ message['role'] }}|>\n {{ message['content'] }}{% endif %}{% endfor %}{% if add_generation_prompt %}<|assistant|>{% endif %}",
        // ChatGLM4
        u8"[gMASK]<sop>{% for item in messages %}{% if item['tools'] is defined %}<|system|>\n你是一个名为 ChatGLM 的人工智能助手。你是基于智谱AI训练的语言模型 GLM-4 模型开发的，你的任务是针对用户的问题和要求提供适当的答复和支持。\n\n# 可用工具{% set tools = item['tools'] %}{% for tool in tools %}{% if tool['type'] == 'function' %}\n\n## {{ tool['function']['name'] }}\n\n{{ tool['function'] | tojson(indent=4) }}\n......{% endif %}{% endfor %}{% endif %}{% if item['content'] %}<|{{ item['role'] }}|>{{ item['metadata'] }}\n{{ item['content'] }}{% endif %}{% endfor %}{% if add_generation_prompt %}<|assistant|>{% endif %}",
+        // GLM-edge
+        "{% for item in messages %}{% if item['role'] == 'system' %}<|system|>\n{{ item['content'] }}{% elif item['role'] == 'user' %}<|user|>\n{{ item['content'] }}{% elif item['role'] == 'assistant' %}<|assistant|>\n{{ item['content'] }}{% endif %}{% endfor %}<|assistant|>\n",
        // MiniCPM-3B-OpenHermes-2.5-v2-GGUF
        u8"{% for message in messages %}{% if message['role'] == 'user' %}{{'<用户>' + message['content'].strip() + '<AI>'}}{% else %}{{message['content'].strip()}}{% endif %}{% endfor %}",
        // DeepSeek-V2
@ -119,6 +121,8 @@ int main(void) {
        "[gMASK]sop<|system|>\n You are a helpful assistant<|user|>\n Hello<|assistant|>\n Hi there<|user|>\n Who are you<|assistant|>\n    I am an assistant   <|user|>\n Another question<|assistant|>",
        // ChatGLM4
        "[gMASK]<sop><|system|>\nYou are a helpful assistant<|user|>\nHello<|assistant|>\nHi there<|user|>\nWho are you<|assistant|>\n   I am an assistant   <|user|>\nAnother question<|assistant|>",
+        // GLM-Edge
+        "<|system|>\nYou are a helpful assistant<|user|>\nHello<|assistant|>\nHi there<|user|>\nWho are you<|assistant|>\n   I am an assistant   <|user|>\nAnother question<|assistant|>",
        // MiniCPM-3B-OpenHermes-2.5-v2-GGUF
        u8"You are a helpful assistant<用户>Hello<AI>Hi there<用户>Who are you<AI>I am an assistant<用户>Another question<AI>",
        // DeepSeek-V2
Author	SHA1	Message	Date
piDack	fa2f378326	Merge `24bad77ebf` into `9ba399dfa7`	2024-12-24 12:55:08 -08:00
Reza Kakhki	9ba399dfa7	server : add support for "encoding_format": "base64" to the /embeddings endpoints (#10967 ) Some checks are pending flake8 Lint / Lint (push) Waiting to run Details Python Type-Check / pyright type-check (push) Waiting to run Details add support for base64 * fix base64 test * improve test --------- Co-authored-by: Xuan Son Nguyen <son@huggingface.co>	2024-12-24 21:33:04 +01:00
Djip007	2cd43f4900	ggml : more perfo with llamafile tinyblas on x86_64 (#10714 ) Some checks are pending flake8 Lint / Lint (push) Waiting to run Details Python Type-Check / pyright type-check (push) Waiting to run Details * more perfo with llamafile tinyblas on x86_64. - add bf16 suport - change dispache strategie (thanks: https://github.com/ikawrakow/ik_llama.cpp/pull/71 ) - reduce memory bandwidth simple tinyblas dispache and more cache freindly * tinyblas dynamic dispaching * sgemm: add M blocs. * - git 2.47 use short id of len 9. - show-progress is not part of GNU Wget2 * remove not stable test	2024-12-24 18:54:49 +01:00
NeverLucky	09fe2e7613	server: allow filtering llama server response fields (#10940 ) * llama_server_response_fields * llama_server_response_fields_fix_issues * params fixes * fix * clarify docs * change to "response_fields" --------- Co-authored-by: Xuan Son Nguyen <son@huggingface.co>	2024-12-24 17:39:49 +01:00
Georgi Gerganov	30caac3a68	llama : the WPM vocabs use the CLS token as BOS (#10930 ) * llama : the WPM vocabs use the CLS token as BOS ggml-ci * llama : add comment	2024-12-24 09:44:20 +02:00
Diego Devesa	60cfa728e2	ggml : use wstring for backend search paths (#10960 ) ggml-ci	2024-12-24 04:05:27 +01:00
Diego Devesa	3327bb0f8d	ggml : fix arm enabled features check (#10961 )	2024-12-24 04:05:17 +01:00
pcdack	24bad77ebf	fix confict	2024-12-19 09:02:35 +00:00
piDack	f91cf62b89	Merge branch 'master' into support_glm_edge_model	2024-12-19 16:50:23 +08:00
piDack	bc93d2a44e	Merge branch 'ggerganov:master' into support_glm_edge_model	2024-12-04 11:50:29 +08:00
liyuhang	6e9fdb0b52	fix chat template	2024-12-04 11:27:01 +08:00
liyuhang	3b409c1e92	add edge template	2024-12-03 19:23:11 +08:00
piDack	82cbfda7b9	Merge branch 'master' into support_glm_edge_model	2024-12-03 13:27:23 +08:00
piDack	5ff563257c	Update src/llama.cpp for not contain <\|end\|> or </s> Co-authored-by: Xuan Son Nguyen <thichthat@gmail.com>	2024-11-30 10:29:58 +08:00
piDack	692880535a	remove unused AutoTokenizer	2024-11-29 08:15:15 +00:00
piDack	816d93db75	Merge branch 'master' of https://github.com/piDack/llama.cpp into support_glm_edge_model	2024-11-29 08:11:31 +00:00
piDack	6c50e9caca	llava.cpp trailing whitespace	2024-11-29 06:11:12 +00:00
piDack	7d80a4aa97	fix format	2024-11-29 06:05:05 +00:00
piDack	55a6f951ca	remove debug info	2024-11-29 05:49:13 +00:00
piDack	3b27041727	vision model support	2024-11-26 12:36:48 +00:00
piDack	ae41d3efed	Merge branch 'master' of https://github.com/piDack/llama.cpp into support_glm_edge_model	2024-11-26 10:01:09 +00:00
liyuhang	6fc90cb727	support for glm edge model	2024-11-26 09:17:30 +00:00
liyuhang	4f696624a4	use config partial_rotary_factor as rope ratio	2024-11-26 05:29:42 +00:00
liyuhang	a249dc0fbb	Merge branch 'master' of https://github.com/piDack/llama.cpp into support_glm_edge_model	2024-11-08 03:53:05 +00:00
liyuhang	677058f470	add glm edge chat model	2024-11-08 03:33:43 +00:00