ggml : add asserts (#14720)

* ggml : add asserts

ggml-ci

* cont : fix constant type

Co-authored-by: Diego Devesa <slarengh@gmail.com>

---------

Co-authored-by: Diego Devesa <slarengh@gmail.com>

common/common.cpp

@@ -1005,15 +1005,21 @@ struct common_init_result common_init_from_params(common_params & params) {
         params.sampling.ignore_eos = false;
     }
 
-    if (params.sampling.ignore_eos) {
-        for (llama_token i = 0; i < llama_vocab_n_tokens(vocab); i++) {
-            if (llama_vocab_is_eog(vocab, i)) {
-                LOG_INF("%s: added %s logit bias = %f\n", __func__, common_token_to_piece(lctx, i).c_str(), -INFINITY);
-                params.sampling.logit_bias.push_back({i, -INFINITY});
-            }
+    // initialize once
+    for (llama_token i = 0; i < llama_vocab_n_tokens(vocab); i++) {
+        if (llama_vocab_is_eog(vocab, i)) {
+            LOG_INF("%s: added %s logit bias = %f\n", __func__, common_token_to_piece(lctx, i).c_str(), -INFINITY);
+            params.sampling.logit_bias_eog.push_back({i, -INFINITY});
         }
     }
 
+    if (params.sampling.ignore_eos) {
+        // add EOG biases to the active set of logit biases
+        params.sampling.logit_bias.insert(
+                params.sampling.logit_bias.end(),
+                params.sampling.logit_bias_eog.begin(), params.sampling.logit_bias_eog.end());
+    }
+
     if (params.sampling.penalty_last_n == -1) {
         LOG_INF("%s: setting penalty_last_n to ctx_size = %d\n", __func__, llama_n_ctx(lctx));
         params.sampling.penalty_last_n = llama_n_ctx(lctx);

common/common.h

@@ -177,7 +177,8 @@ struct common_params_sampling {
     std::vector<common_grammar_trigger> grammar_triggers; // optional triggers (for lazy grammars)
     std::set<llama_token>               preserved_tokens;
 
-    std::vector<llama_logit_bias> logit_bias; // logit biases to apply
+    std::vector<llama_logit_bias> logit_bias;     // logit biases to apply
+    std::vector<llama_logit_bias> logit_bias_eog; // pre-calculated logit biases for EOG tokens
 
     // print the parameters into a string
     std::string print() const;

convert_hf_to_gguf.py

@@ -669,6 +669,36 @@ class TextModel(ModelBase):
         # NOTE: if you get an error here, you need to update the convert_hf_to_gguf_update.py script
         #       or pull the latest version of the model from Huggingface
         #       don't edit the hashes manually!
+        if chkhsh == "b6e8e1518dc4305be2fe39c313ed643381c4da5db34a98f6a04c093f8afbe99b":
+            # ref: https://huggingface.co/THUDM/glm-4-9b-chat
+            res = "chatglm-bpe"
+        if chkhsh == "81d72c7348a9f0ebe86f23298d37debe0a5e71149e29bd283904c02262b27516":
+            # ref: https://huggingface.co/THUDM/glm-4-9b-chat
+            res = "chatglm-bpe"
+        if chkhsh == "a1336059768a55c99a734006ffb02203cd450fed003e9a71886c88acf24fdbc2":
+            # ref: https://huggingface.co/THUDM/glm-4-9b-hf
+            res = "glm4"
+        if chkhsh == "1431a23e583c97432bc230bff598d103ddb5a1f89960c8f1d1051aaa944d0b35":
+            # ref: https://huggingface.co/sapienzanlp/Minerva-7B-base-v1.0
+            res = "minerva-7b"
+        if chkhsh == "7e57df22b1fe23a7b1e1c7f3dc4e3f96d43a4eb0836d0c6bdc3436d7b2f1c664":
+            # ref: https://huggingface.co/tencent/Hunyuan-A13B-Instruct
+            res = "hunyuan"
+        if chkhsh == "a6b57017d60e6edb4d88ecc2845188e0eb333a70357e45dcc9b53964a73bbae6":
+            # ref: https://huggingface.co/tiiuae/Falcon-H1-0.5B-Base
+            res = "falcon-h1"
+        if chkhsh == "60476e1243776c4fb1b993dbd7a5f15ac22f83c80afdf425fa5ae01c8d44ef86":
+            # ref: https://huggingface.co/tiiuae/Falcon-H1-1B-Base
+            res = "falcon-h1"
+        if chkhsh == "3eda48b4c4dc7de733d1a8b3e3b4a85243dbbf704da2ee9d42c6beced8897896":
+            # ref: https://huggingface.co/tiiuae/Falcon-H1-7B-Base
+            res = "falcon-h1"
+        if chkhsh == "48f8e02c0359c0bbdd82f26909171fac1c18a457bb47573ed1fe3bbb2c1cfd4b":
+            # ref: https://huggingface.co/tiiuae/Falcon-H1-34B-Base
+            res = "falcon-h1"
+        if chkhsh == "81212dc7cdb7e0c1074ca62c5aeab0d43c9f52b8a737be7b12a777c953027890":
+            # ref: https://huggingface.co/moonshotai/Kimi-K2-Base
+            res = "kimi-k2"
         if chkhsh == "0ef9807a4087ebef797fc749390439009c3b9eda9ad1a097abbe738f486c01e5":
             # ref: https://huggingface.co/meta-llama/Meta-Llama-3-8B
             res = "llama-bpe"
@@ -804,36 +834,9 @@ class TextModel(ModelBase):
         if chkhsh == "d5f1dd6f980fec569fb218a81a7658ac45fc56b38c5a0adeb1c232fbe04ef5ec":
             # ref: https://huggingface.co/ByteDance-Seed/Seed-Coder-8B-Base
             res = "seed-coder"
-        if chkhsh == "b6e8e1518dc4305be2fe39c313ed643381c4da5db34a98f6a04c093f8afbe99b":
-            # ref: https://huggingface.co/THUDM/glm-4-9b-chat
-            res = "chatglm-bpe"
-        if chkhsh == "81d72c7348a9f0ebe86f23298d37debe0a5e71149e29bd283904c02262b27516":
-            # ref: https://huggingface.co/THUDM/glm-4-9b-chat
-            res = "chatglm-bpe"
-        if chkhsh == "a1336059768a55c99a734006ffb02203cd450fed003e9a71886c88acf24fdbc2":
-            # ref: https://huggingface.co/THUDM/glm-4-9b-hf
-            res = "glm4"
-        if chkhsh == "1431a23e583c97432bc230bff598d103ddb5a1f89960c8f1d1051aaa944d0b35":
-            # ref: https://huggingface.co/sapienzanlp/Minerva-7B-base-v1.0
-            res = "minerva-7b"
-        if chkhsh == "7e57df22b1fe23a7b1e1c7f3dc4e3f96d43a4eb0836d0c6bdc3436d7b2f1c664":
-            # ref: https://huggingface.co/tencent/Hunyuan-A13B-Instruct
-            res = "hunyuan"
         if chkhsh == "b0a6b1c0bd5998ebd9df08611efde34a4ff03faed45ae09c43e6b31ebd4b94cf":
             # ref: https://huggingface.co/skt/A.X-4.0
             res = "a.x-4.0"
-        if chkhsh == "a6b57017d60e6edb4d88ecc2845188e0eb333a70357e45dcc9b53964a73bbae6":
-            # ref: https://huggingface.co/tiiuae/Falcon-H1-0.5B-Base
-            res = "falcon-h1"
-        if chkhsh == "60476e1243776c4fb1b993dbd7a5f15ac22f83c80afdf425fa5ae01c8d44ef86":
-            # ref: https://huggingface.co/tiiuae/Falcon-H1-1B-Base
-            res = "falcon-h1"
-        if chkhsh == "3eda48b4c4dc7de733d1a8b3e3b4a85243dbbf704da2ee9d42c6beced8897896":
-            # ref: https://huggingface.co/tiiuae/Falcon-H1-7B-Base
-            res = "falcon-h1"
-        if chkhsh == "48f8e02c0359c0bbdd82f26909171fac1c18a457bb47573ed1fe3bbb2c1cfd4b":
-            # ref: https://huggingface.co/tiiuae/Falcon-H1-34B-Base
-            res = "falcon-h1"
         if chkhsh == "f6791d196f87ce6b56a7d234be618e0d58f8cda3549416635b2bebcd22cd95c4":
             # ref: https://huggingface.co/K-intelligence/Midm-2.0-Base-Instruct
             res = "midm-2.0"
@@ -5739,7 +5742,58 @@ class DeepseekV2Model(TextModel):
     model_arch = gguf.MODEL_ARCH.DEEPSEEK2
 
     def set_vocab(self):
-        self._set_vocab_gpt2()
+        try:
+            self._set_vocab_gpt2()
+            return
+        except Exception:
+            pass
+
+        from transformers import AutoTokenizer
+        tokenizer = AutoTokenizer.from_pretrained(self.dir_model, trust_remote_code=True)
+        tokpre = self.get_vocab_base_pre(tokenizer)
+
+        if tokpre == "kimi-k2":
+            # Build merges list using the approach similar to HunYuanMoE
+            merges = []
+            vocab = {}
+            mergeable_ranks = tokenizer.model._mergeable_ranks
+            for token, rank in mergeable_ranks.items():
+                vocab[QwenModel.token_bytes_to_string(token)] = rank
+                if len(token) == 1:
+                    continue
+                merged = QwenModel.bpe(mergeable_ranks, token, max_rank=rank)
+                if len(merged) == 2:
+                    merges.append(' '.join(map(QwenModel.token_bytes_to_string, merged)))
+
+            # Build token list
+            vocab_size = self.hparams["vocab_size"]
+            special_tokens = tokenizer.special_tokens
+            reverse_vocab = {id_ : encoded_tok for encoded_tok, id_ in {**vocab, **special_tokens}.items()}
+            tokens: list[str] = []
+            toktypes: list[int] = []
+
+            for i in range(vocab_size):
+                if i not in reverse_vocab:
+                    tokens.append(f"[PAD{i}]")
+                    toktypes.append(gguf.TokenType.UNUSED)
+                else:
+                    token = reverse_vocab[i]
+                    tokens.append(token)
+                    if i in special_tokens.values():
+                        toktypes.append(gguf.TokenType.CONTROL)
+                    else:
+                        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_token_merges(merges)
+
+            special_vocab = gguf.SpecialVocab(self.dir_model, load_merges=False)
+            special_vocab.add_to_gguf(self.gguf_writer)
+        else:
+            raise NotImplementedError(f"Deepseek pre-tokenizer {tokpre!r} is not supported yet!")
 
     def set_gguf_parameters(self):
 

convert_hf_to_gguf_update.py

@@ -146,6 +146,7 @@ pre_computed_hashes = [
     {"name": "falcon-h1", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/tiiuae/Falcon-H1-1B-Base", "chkhsh": "60476e1243776c4fb1b993dbd7a5f15ac22f83c80afdf425fa5ae01c8d44ef86"},
     {"name": "falcon-h1", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/tiiuae/Falcon-H1-7B-Base", "chkhsh": "3eda48b4c4dc7de733d1a8b3e3b4a85243dbbf704da2ee9d42c6beced8897896"},
     {"name": "falcon-h1", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/tiiuae/Falcon-H1-34B-Base", "chkhsh": "48f8e02c0359c0bbdd82f26909171fac1c18a457bb47573ed1fe3bbb2c1cfd4b"},
+    {"name": "kimi-k2",   "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/moonshotai/Kimi-K2-Base",   "chkhsh": "81212dc7cdb7e0c1074ca62c5aeab0d43c9f52b8a737be7b12a777c953027890"},
 ]
 
 
@@ -231,7 +232,7 @@ for model in models:
 # generate the source code for the convert_hf_to_gguf.py:get_vocab_base_pre() function:
 
 src_ifs = ""
-for model in [*all_models, *pre_computed_hashes]:
+for model in [*pre_computed_hashes, *all_models]:
     name = model["name"]
     tokt = model["tokt"]
     chkhsh = model.get("chkhsh")
@@ -239,11 +240,6 @@ for model in [*all_models, *pre_computed_hashes]:
     if tokt == TOKENIZER_TYPE.SPM or tokt == TOKENIZER_TYPE.UGM:
         continue
 
-    # Skip if the tokenizer folder does not exist or there are other download issues previously
-    if not os.path.exists(f"models/tokenizers/{name}"):
-        logger.warning(f"Directory for tokenizer {name} not found. Skipping...")
-        continue
-
     # create the tokenizer
     if chkhsh is not None:
         # if the model has a pre-computed hash, use it
@@ -253,6 +249,12 @@ for model in [*all_models, *pre_computed_hashes]:
         chkhsh = existing_models[name]
     else:
         # otherwise, compute the hash of the tokenizer
+
+        # Skip if the tokenizer folder does not exist or there are other download issues previously
+        if not os.path.exists(f"models/tokenizers/{name}"):
+            logger.warning(f"Directory for tokenizer {name} not found. Skipping...")
+            continue
+
         try:
             logger.info(f"Loading tokenizer from {f'models/tokenizers/{name}'}...")
             if name == "t5":

ggml/src/ggml-cpu/ops.cpp

@@ -4015,6 +4015,9 @@ static void ggml_compute_forward_rms_norm_f32(
 
                 const float scale = 1.0f/sqrtf(mean + eps);
 
+                // if you hit this, likely you got an inf somewhere earlier
+                assert(scale > 0.0f);
+
                 ggml_vec_scale_f32(ne00, y, scale);
             }
         }

ggml/src/ggml-cpu/vec.cpp

@@ -221,6 +221,9 @@ void ggml_vec_dot_f16(int n, float * GGML_RESTRICT s, size_t bs, ggml_fp16_t * G
     for (int i = np; i < n; ++i) {
         sumf += (ggml_float)(GGML_CPU_FP16_TO_FP32(x[i])*GGML_CPU_FP16_TO_FP32(y[i]));
     }
+
+    // if you hit this, you are likely running outside the FP range
+    assert(!isnan(sumf) && !isinf(sumf));
 #else
     for (int i = 0; i < n; ++i) {
         sumf += (ggml_float)(GGML_CPU_FP16_TO_FP32(x[i])*GGML_CPU_FP16_TO_FP32(y[i]));

ggml/src/ggml-vulkan/ggml-vulkan.cpp

@@ -4922,7 +4922,7 @@ static bool ggml_vk_dim01_contiguous(const ggml_tensor * tensor) {
     return
         tensor->nb[0] == ggml_type_size(tensor->type) &&
         tensor->nb[1] == (tensor->nb[0]*tensor->ne[0])/ggml_blck_size(tensor->type) &&
-        tensor->nb[3] == tensor->nb[2]*tensor->ne[2];
+        (tensor->ne[3] == 1 || tensor->nb[3] == tensor->nb[2]*tensor->ne[2]);
 }
 
 static vk_pipeline ggml_vk_get_cpy_pipeline(ggml_backend_vk_context * ctx, const ggml_tensor * src, const ggml_tensor * dst, ggml_type to) {
@@ -10356,10 +10356,6 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
                         // If there's not enough shared memory for row_ids and the result tile, fallback to CPU
                         return false;
                     }
-                    // Check against size of shared memory variable
-                    if (op->src[2]->ne[0] > 4096) {
-                        return false;
-                    }
                 }
                 switch (src0_type) {
                     case GGML_TYPE_F32:

ggml/src/ggml-vulkan/vulkan-shaders/dequant_q2_k.comp

@@ -10,7 +10,7 @@ layout (binding = 1) writeonly buffer D {D_TYPE data_b[];};
 void main() {
     [[unroll]] for (uint wgy = 0; wgy < 256; wgy++) {
         const uint i = gl_WorkGroupID.x * 256 + wgy;
-        if (i >= p.M * p.K / QUANT_K) {
+        if (i >= p.nel / QUANT_K) {
             return;
         }
 

ggml/src/ggml-vulkan/vulkan-shaders/dequant_q3_k.comp

@@ -10,7 +10,7 @@ layout (binding = 1) writeonly buffer D {D_TYPE data_b[];};
 void main() {
     [[unroll]] for (uint wgy = 0; wgy < 256; wgy++) {
         const uint i = uint(gl_WorkGroupID.x * 256 + wgy);
-        if (i >= p.M * p.K / QUANT_K) {
+        if (i >= p.nel / QUANT_K) {
             return;
         }
 

ggml/src/ggml-vulkan/vulkan-shaders/dequant_q4_k.comp

@@ -10,7 +10,7 @@ layout (binding = 1) writeonly buffer D {D_TYPE data_b[];};
 void main() {
     [[unroll]] for (uint wgy = 0; wgy < 256; wgy++) {
         const uint ib = gl_WorkGroupID.x * 256 + wgy;
-        if (ib >= p.M * p.K / QUANT_K) {
+        if (ib >= p.nel / QUANT_K) {
             return;
         }
 

ggml/src/ggml-vulkan/vulkan-shaders/dequant_q5_k.comp

@@ -10,7 +10,7 @@ layout (binding = 1) writeonly buffer D {D_TYPE data_b[];};
 void main() {
     [[unroll]] for (uint wgy = 0; wgy < 256; wgy++) {
         const uint ib = gl_WorkGroupID.x * 256 + wgy;
-        if (ib >= p.M * p.K / QUANT_K) {
+        if (ib >= p.nel / QUANT_K) {
             return;
         }
 

ggml/src/ggml-vulkan/vulkan-shaders/dequant_q6_k.comp

@@ -10,7 +10,7 @@ layout (binding = 1) writeonly buffer D {D_TYPE data_b[];};
 void main() {
     [[unroll]] for (uint wgy = 0; wgy < 256; wgy++) {
         const uint i = gl_WorkGroupID.x * 256 + wgy;
-        if (i >= p.M * p.K / QUANT_K) {
+        if (i >= p.nel / QUANT_K) {
             return;
         }
         const uint tid = gl_LocalInvocationID.x;

gguf-py/gguf/scripts/gguf_dump.py

@@ -234,6 +234,8 @@ def dump_markdown_metadata(reader: GGUFReader, args: argparse.Namespace) -> None
     markdown_content += '## Key Value Metadata Store\n\n'
     markdown_content += f'There are {len(reader.fields)} key-value pairs in this file\n'
     markdown_content += '\n'
+    total_model_bytes = 0
+    total_model_elements = 0
 
     kv_dump_table: list[dict[str, str | int]] = []
     for n, field in enumerate(reader.fields.values(), 1):
@@ -377,6 +379,8 @@ def dump_markdown_metadata(reader: GGUFReader, args: argparse.Namespace) -> None
             tensors = tensor_groups[group]
             group_elements = sum(tensor.n_elements for tensor in tensors)
             group_percentage = group_elements / total_elements * 100
+            total_group_bytes = 0
+            total_group_elements = 0
             markdown_content += f"### <a name=\"{group.replace('.', '_')}\">{translate_tensor_name(group)} Tensor Group : {element_count_rounded_notation(group_elements)} Elements</a>\n\n"
 
             # Precalculate column sizing for visual consistency
@@ -397,7 +401,13 @@ def dump_markdown_metadata(reader: GGUFReader, args: argparse.Namespace) -> None
                 element_count_est = f"({element_count_rounded_notation(tensor.n_elements):>{prettify_element_est_count_size}})"
                 element_count_string = f"{element_count_est} {tensor.n_elements:>{prettify_element_count_size}}"
                 type_name_string = f"{tensor.tensor_type.name}"
-                tensor_dump_table.append({"t_id":tensor_name_to_key[tensor.name], "layer_name":tensor.name, "human_layer_name":human_friendly_name, "element_count":element_count_string, "pretty_dimension":pretty_dimension, "tensor_type":type_name_string})
+                if tensor.n_elements > 0:
+                    bpw = (tensor.n_bytes * 8) / tensor.n_elements
+                else:
+                    bpw = float('nan')
+                tensor_dump_table.append({"t_id":tensor_name_to_key[tensor.name], "layer_name":tensor.name, "human_layer_name":human_friendly_name, "element_count":element_count_string, "pretty_dimension":pretty_dimension, "tensor_type":type_name_string, "bpw": f"{bpw:.4f}"})
+                total_group_bytes += tensor.n_bytes
+                total_group_elements += tensor.n_elements
 
             tensor_dump_table_header_map = [
                 {'key_name':'t_id',             'header_name':'T_ID',                             'align':'right'},
@@ -406,6 +416,7 @@ def dump_markdown_metadata(reader: GGUFReader, args: argparse.Namespace) -> None
                 {'key_name':'element_count',    'header_name':'Elements',                         'align':'left'},
                 {'key_name':'pretty_dimension', 'header_name':'Shape',                            'align':'left'},
                 {'key_name':'tensor_type',      'header_name':'Type',                             'align':'left'},
+                {'key_name':'bpw',              'header_name':'BPW',                              'align':'right'},
             ]
 
             markdown_content += markdown_table_with_alignment_support(tensor_dump_table_header_map, tensor_dump_table)
@@ -413,8 +424,20 @@ def dump_markdown_metadata(reader: GGUFReader, args: argparse.Namespace) -> None
             markdown_content += "\n"
             markdown_content += f"- Total elements in {group}: ({element_count_rounded_notation(group_elements):>4}) {group_elements}\n"
             markdown_content += f"- Percentage of total elements: {group_percentage:.2f}%\n"
+            if total_group_elements > 0:
+                total_group_bpw = (total_group_bytes * 8) / total_group_elements
+                markdown_content += f"- Bits per Weight (BPW) for {group}: {total_group_bpw:.4f} bits\n"
+            else:
+                markdown_content += f"- Bits per Weight (BPW) for {group}: undefined (no elements)\n"
             markdown_content += "\n\n"
+            total_model_bytes += total_group_bytes
+            total_model_elements += total_group_elements
 
+    if total_model_elements > 0:
+        total_model_bpw = (total_model_bytes * 8) / total_model_elements
+        markdown_content += f"Total BPW for {os.path.basename(args.model)}: {total_model_bpw:.4f} bits"
+    else:
+        markdown_content += f"Total BPW for {os.path.basename(args.model)}: undefined (no elements)"
     print(markdown_content)  # noqa: NP100
 
 

include/llama.h

@@ -725,7 +725,7 @@ extern "C" {
     //   - lazily on next llama_decode()
     // p0 < 0 : [0,  p1]
     // p1 < 0 : [p0, inf)
-    DEPRECATED(void llama_kv_self_seq_div(
+    DEPRECATED(LLAMA_API void llama_kv_self_seq_div(
             struct llama_context * ctx,
                     llama_seq_id   seq_id,
                        llama_pos   p0,

models/templates/moonshotai-Kimi-K2.jinja

@@ -0,0 +1,43 @@
+{%- if tools -%}
+  <|im_system|>tool_declare<|im_middle|>{{ tools | tojson }}<|im_end|>
+{%- endif -%}
+{%- for message in messages -%}
+  {%- if loop.first and messages[0]['role'] != 'system' -%}
+    <|im_system|>system<|im_middle|>You are a helpful assistant<|im_end|>
+  {%- endif -%}
+  {%- if message['role'] == 'system' -%}
+    <|im_system|>system<|im_middle|>
+  {%- elif message['role'] == 'user' -%}
+    <|im_user|>user<|im_middle|>
+  {%- elif message['role'] == 'assistant' -%}
+    <|im_assistant|>assistant<|im_middle|>
+  {%- elif message['role'] == 'tool' -%}
+    <|im_system|>tool<|im_middle|>
+  {%- endif -%}
+  {%- if message['role'] == 'assistant' and message.get('tool_calls') -%}
+    {%- if message['content'] -%}{{ message['content'] }}{%- endif -%}
+    <|tool_calls_section_begin|>
+    {%- for tool_call in message['tool_calls'] -%}
+      {%- set func_name = tool_call['function']['name'] -%}
+      {%- set formatted_id = 'functions.' + func_name + ':' + loop.index0|string -%}
+      <|tool_call_begin|>{{ formatted_id }}<|tool_call_argument_begin|>{{ tool_call['function']['arguments'] | tojson}}<|tool_call_end|>
+    {%- endfor -%}
+    <|tool_calls_section_end|>
+  {%- elif message['role'] == 'tool' -%}
+    ## Return of {{ message.tool_call_id }}\n{{ message['content'] }}
+  {%- elif message['content'] is string -%}
+    {{ message['content'] }}
+  {%- elif message['content'] is not none -%}
+    {% for content in message['content'] -%}
+      {% if content['type'] == 'image' or 'image' in content or 'image_url' in content -%}
+        <|media_start|>image<|media_content|><|media_pad|><|media_end|>
+      {% else -%}
+        {{ content['text'] }}
+      {%- endif -%}
+    {%- endfor -%}
+  {%- endif -%}
+  <|im_end|>
+{%- endfor -%}
+{%- if add_generation_prompt -%}
+  <|im_assistant|>assistant<|im_middle|>
+{%- endif -%}

scripts/server-bench.py

@@ -2,9 +2,11 @@
 
 import argparse
 import json
+import os
+import random
 import subprocess
 from time import sleep, time
-from typing import Optional
+from typing import Optional, Union
 
 import datasets
 import logging
@@ -18,31 +20,39 @@ logging.basicConfig(level=logging.INFO, format='%(message)s')
 logger = logging.getLogger("server-bench")
 
 
-def get_prompts(n_prompts: int) -> list[str]:
-    logger.info("Loading MMLU dataset...")
-    ret = datasets.load_dataset("cais/mmlu", "all")["test"]["question"]  # type: ignore
+def get_prompts_text(dataset_name: str, n_prompts: int) -> Optional[list[str]]:
+    ret = []
+    if dataset_name.lower() == "mmlu":
+        logger.info("Loading MMLU dataset...")
+        ret = datasets.load_dataset("cais/mmlu", "all")["test"]["question"]  # type: ignore
+    else:
+        return None
     if n_prompts >= 0:
         ret = ret[:n_prompts]
     return ret
 
 
-def get_server(path_server: str, path_model: str, path_log: Optional[str], port: int, n_gpu_layers: int, parallel: int, ctx_size: int) -> dict:
-    logger.info("Starting the llama.cpp server...")
-    address = f"http://localhost:{port}"
+def get_prompt_lengths_rng(n_prompts: int, prompt_length_min: int, prompt_length_max: int) -> list[int]:
+    assert n_prompts >= 0
+    ret: list[int] = []
+    for i in range(n_prompts):
+        random.seed(13 * i + 0)
+        ret.append(random.randint(prompt_length_min, prompt_length_max))
+    return ret
 
-    popen_args: list[str] = [
-        path_server,
-        "--flash-attn",
-        "--n-gpu-layers", str(n_gpu_layers),
-        "--parallel", str(parallel),
-        "--ctx-size", str(parallel * ctx_size),
-        "--model", path_model,
-        "--port", str(port),
-        "--swa-full",  # FIXME performance bad otherwise
-        # "--attn-streams",
-    ]
-    fout = open("bench.log", "w") if path_log is not None else subprocess.DEVNULL
-    process = subprocess.Popen(popen_args, stdout=fout, stderr=subprocess.STDOUT)
+
+def get_prompts_rng(prompt_lengths: list[int]) -> list[list[int]]:
+    return [[random.randint(100, 10000) for _ in range(pl)] for pl in prompt_lengths]
+
+
+def get_server(path_server: str, path_log: Optional[str]) -> dict:
+    logger.info("Starting the llama.cpp server...")
+    hostname: str = os.environ.get("LLAMA_ARG_HOST", "127.0.0.1")
+    port: str = os.environ.get("LLAMA_ARG_PORT", "8080")
+    address: str = f"http://{hostname}:{port}"
+
+    fout = open(path_log, "w") if path_log is not None else subprocess.DEVNULL
+    process = subprocess.Popen([path_server], stdout=fout, stderr=subprocess.STDOUT)
 
     n_failures: int = 0
     while True:
@@ -50,14 +60,14 @@ def get_server(path_server: str, path_model: str, path_log: Optional[str], port:
             sleep(1.0)
             exit_code = process.poll()
             if exit_code is not None:
-                raise RuntimeError(f"llama.cpp server for {path_model} exited unexpectedly with exit code {exit_code}")
+                raise RuntimeError(f"llama.cpp server exited unexpectedly with exit code {exit_code}, see {path_log}")
             response = requests.get(f"{address}/health")
             if response.status_code == 200:
                 break
         except requests.ConnectionError:
             n_failures += 1
             if n_failures >= 10:
-                raise RuntimeError(f"llama.cpp server for {path_model} is not healthy after 10 seconds")
+                raise RuntimeError("llama.cpp server is not healthy after 10 seconds")
 
     return {"process": process, "address": address, "fout": fout}
 
@@ -87,58 +97,97 @@ def send_prompt(data: dict) -> tuple[float, list[float]]:
     session = data["session"]
     server_address: str = data["server_address"]
 
-    response = session.post(
-        f"{server_address}/apply-template",
-        json={"messages": [{"role": "user", "content": data["prompt"], "stream": True}]}
-    )
-    if response.status_code != 200:
-        raise RuntimeError(f"Server returned status code {response.status_code}: {response.text}")
-    prompt: str = json.loads(response.text)["prompt"]
+    t_submit = time()
+    if data["synthetic_prompt"]:
+        json_data: dict = {
+            "prompt": data["prompt"], "ignore_eos": True, "cache_prompt": False,
+            "seed": data["seed"], "n_predict": data["n_predict"], "stream": True}
+        response = session.post(f"{server_address}/completion", json=json_data, stream=True)
+    else:
+        response = session.post(
+            f"{server_address}/apply-template",
+            json={"messages": [{"role": "user", "content": data["prompt"], "stream": True}]}
+        )
+        if response.status_code != 200:
+            raise RuntimeError(f"Server returned status code {response.status_code}: {response.text}")
+        prompt: str = json.loads(response.text)["prompt"]
 
-    json_data: dict = {"prompt": prompt, "seed": data["seed"], "n_predict": data["n_predict"], "stream": True}
-    response = session.post(f"{server_address}/completion", json=json_data, stream=True)
+        json_data: dict = {"prompt": prompt, "seed": data["seed"], "n_predict": data["n_predict"], "stream": True}
+        response = session.post(f"{server_address}/completion", json=json_data, stream=True)
 
-    last_valid_line: str = ""
     token_arrival_times: list[float] = []
-    for line in response.iter_lines(decode_unicode=True):
-        if not line.startswith("data: "):
+    for line in response.iter_lines(decode_unicode=False):
+        if not line.startswith(b"data: "):
             continue
-        last_valid_line = line
         token_arrival_times.append(time())
     token_arrival_times = token_arrival_times[:-1]
 
     if response.status_code != 200:
         raise RuntimeError(f"Server returned status code {response.status_code}: {response.text}")
-    timings: dict = json.loads(last_valid_line[6:])["timings"]
 
-    return (timings["prompt_ms"], token_arrival_times)
+    return (t_submit, token_arrival_times)
 
 
-def benchmark(path_server: str, path_model: str, path_log: Optional[str], port: int, n_gpu_layers: int, parallel: int, ctx_size: int, n_prompts: int, n_predict: int):
-    num_workers: int = parallel + 1
-    prompts: list[str] = get_prompts(n_prompts)
+def benchmark(path_server: str, path_log: Optional[str], prompt_source: str, n_prompts: int, n_predict: int, n_predict_min: int):
+    if os.environ.get("LLAMA_ARG_N_PARALLEL") is None:
+        logger.info("LLAMA_ARG_N_PARALLEL not explicitly set, using 32")
+        os.environ["LLAMA_ARG_N_PARALLEL"] = "32"
+    if os.environ.get("LLAMA_ARG_N_GPU_LAYERS") is None:
+        logger.info("LLAMA_ARG_N_GPU_LAYERS not explicitly set, using 999")
+        os.environ["LLAMA_ARG_N_GPU_LAYERS"] = "999"
+    if os.environ.get("LLAMA_ARG_FLASH_ATTN") is None:
+        logger.info("LLAMA_ARG_FLASH_ATTN not explicitly set, using 'true'")
+        os.environ["LLAMA_ARG_FLASH_ATTN"] = "true"
+
+    parallel: int = int(os.environ.get("LLAMA_ARG_N_PARALLEL", 1))
+    prompts: Union[None, list[str], list[list[int]]] = get_prompts_text(prompt_source, n_prompts)
+    synthetic_prompts: bool = prompts is None
+    prompt_n = []
+
+    if synthetic_prompts:
+        prompt_source_split: list[str] = prompt_source.split("-")
+        assert len(prompt_source_split) == 3
+        assert prompt_source_split[0].lower() == "rng"
+        prompt_length_min: int = int(prompt_source_split[1])
+        prompt_length_max: int = int(prompt_source_split[2])
+        logger.info("Generating random prompts...")
+        prompt_n = get_prompt_lengths_rng(n_prompts, prompt_length_min, prompt_length_max)
+        prompts = get_prompts_rng(prompt_n)
+    else:
+        n_predict_min = n_predict
+
+    if os.environ.get("LLAMA_ARG_CTX_SIZE") is None:
+        context_per_slot: int = int(1.05 * (n_predict + (np.max(prompt_n) if synthetic_prompts else 2048)))
+        context_total: int = context_per_slot * parallel
+        os.environ["LLAMA_ARG_CTX_SIZE"] = str(context_total)
+        logger.info(f"LLAMA_ARG_CTX_SIZE not explicitly set, using {context_total} ({context_per_slot} per slot).")
 
     server: Optional[dict] = None
     session = None
     try:
-        server = get_server(path_server, path_model, path_log, port, n_gpu_layers, parallel, ctx_size)
+        server = get_server(path_server, path_log)
         server_address: str = server["address"]
 
-        adapter = requests.adapters.HTTPAdapter(pool_connections=num_workers, pool_maxsize=num_workers)  # type: ignore
+        adapter = requests.adapters.HTTPAdapter(pool_connections=parallel, pool_maxsize=parallel)  # type: ignore
         session = requests.Session()
         session.mount("http://", adapter)
         session.mount("https://", adapter)
 
         data: list[dict] = []
-        for i, p in enumerate(prompts):
-            data.append({"session": session, "server_address": server_address, "prompt": p, "n_predict": n_predict, "seed": i})
 
-        logger.info("Getting the prompt lengths...")
-        prompt_n = [get_prompt_length(d) for d in data]
+        for i, p in enumerate(prompts):
+            random.seed(13 * i + 1)
+            data.append({
+                "session": session, "server_address": server_address, "prompt": p, "synthetic_prompt": synthetic_prompts,
+                "n_predict": random.randint(n_predict_min, n_predict), "seed": 13 * i + 2})
+
+        if not synthetic_prompts:
+            logger.info("Getting the prompt lengths...")
+            prompt_n = [get_prompt_length(d) for d in data]
 
         logger.info("Starting the benchmark...\n")
         t0 = time()
-        results: list[tuple[int, list[float]]] = thread_map(send_prompt, data, max_workers=num_workers, chunksize=1)
+        results: list[tuple[float, list[float]]] = thread_map(send_prompt, data, max_workers=parallel, chunksize=1)
     finally:
         if server is not None:
             server["process"].terminate()
@@ -146,17 +195,18 @@ def benchmark(path_server: str, path_model: str, path_log: Optional[str], port:
         if session is not None:
             session.close()
 
-    prompt_ms = []
+    prompt_t = []
     token_t = []
     depth_sum: int = 0
-    for pn, (pms, tat) in zip(prompt_n, results):
-        prompt_ms.append(pms)
+    for pn, (t_submit, tat) in zip(prompt_n, results):
+        prompt_t.append(tat[0] - t_submit)
         token_t += tat
         n_tokens: int = len(tat)
         depth_sum += n_tokens * pn
         depth_sum += n_tokens * (n_tokens + 1) // 2
+    assert len(token_t) > 0
     prompt_n = np.array(prompt_n, dtype=np.int64)
-    prompt_ms = np.array(prompt_ms, dtype=np.float64)
+    prompt_t = np.array(prompt_t, dtype=np.float64)
     token_t = np.array(token_t, dtype=np.float64)
 
     token_t -= t0
@@ -167,18 +217,21 @@ def benchmark(path_server: str, path_model: str, path_log: Optional[str], port:
     logger.info(f"Request throughput:                {n_prompts / token_t_last:.2f} requests/s = {n_prompts / (token_t_last/60):.2f} requests/min")
     logger.info(f"Total prompt length:               {np.sum(prompt_n)} tokens")
     logger.info(f"Average prompt length:             {np.mean(prompt_n):.2f} tokens")
-    logger.info(f"Average prompt latency:            {np.mean(prompt_ms):.2f} ms")
-    logger.info(f"Average prompt speed:              {np.sum(prompt_n) / (1e-3 * np.sum(prompt_ms)):.2f} tokens/s")
+    logger.info(f"Average prompt latency:            {1e3 * np.mean(prompt_t):.2f} ms")
+    logger.info(f"Average prompt speed:              {np.sum(prompt_n) / np.sum(prompt_t):.2f} tokens/s")
     logger.info(f"Total generated tokens:            {token_t.shape[0]}")
     logger.info(f"Average generation depth:          {depth_sum / token_t.shape[0]:.2f} tokens")
     logger.info(f"Average total generation speed:    {token_t.shape[0] / token_t_last:.2f} tokens/s")
     logger.info(f"Average generation speed per slot: {token_t.shape[0] / (parallel * token_t_last):.2f} tokens/s / slot")
+    logger.info("")
+    logger.info(
+        "The above numbers are the speeds as observed by the Python script and may differ from the performance reported by the server, "
+        "particularly when the server is fast vs. the network or Python script (e.g. when serving a very small model).")
 
     plt.figure()
-    plt.scatter(prompt_n, prompt_ms, s=10.0, marker=".", alpha=0.25)
-    plt.xlim(0, 1.05 * np.max(prompt_n))
-    plt.ylim(0, 1.05 * np.max(prompt_ms))
-    plt.title(path_model)
+    plt.scatter(prompt_n, 1e3 * prompt_t, s=10.0, marker=".", alpha=0.25)
+    plt.xlim(0, 1.05e0 * np.max(prompt_n))
+    plt.ylim(0, 1.05e3 * np.max(prompt_t))
     plt.xlabel("Prompt length [tokens]")
     plt.ylabel("Time to first token [ms]")
     plt.savefig("prompt_time.png", dpi=240)
@@ -187,7 +240,6 @@ def benchmark(path_server: str, path_model: str, path_log: Optional[str], port:
     plt.figure()
     plt.hist(token_t, np.arange(0, bin_max))
     plt.xlim(0, bin_max + 1)
-    plt.title(path_model)
     plt.xlabel("Time [s]")
     plt.ylabel("Num. tokens generated per second")
     plt.savefig("gen_rate.png", dpi=240)
@@ -196,15 +248,18 @@ def benchmark(path_server: str, path_model: str, path_log: Optional[str], port:
 if __name__ == "__main__":
     parser = argparse.ArgumentParser(
         description="Tool for benchmarking the throughput of the llama.cpp HTTP server. "
-        "Results are printed to console and visualized as plots (saved to current working directory).")
+        "Results are printed to console and visualized as plots (saved to current working directory). "
+        "To pass arguments such as the model path to the server, set the corresponding environment variables (see llama-server --help).")
     parser.add_argument("--path_server", type=str, default="llama-server", help="Path to the llama.cpp server binary")
-    parser.add_argument("--path_model", type=str, required=True, help="Path to the model to use for the benchmark")
-    parser.add_argument("--path_log", type=str, default=None, help="Path to the model to use for the benchmark")
-    parser.add_argument("--port", type=int, default=18725, help="Port to use for the server during the benchmark")
-    parser.add_argument("--n_gpu_layers", type=int, default=999, help="Number of GPU layers for the server")
-    parser.add_argument("--parallel", type=int, default=16, help="Number of slots for the server")
-    parser.add_argument("--ctx_size", type=int, default=4096, help="Server context size per slot")
-    parser.add_argument("--n_prompts", type=int, default=1000, help="Number of prompts to evaluate")
+    parser.add_argument("--path_log", type=str, default="server-bench.log", help="Path to the model to use for the benchmark")
+    parser.add_argument(
+        "--prompt_source", type=str, default="rng-1024-2048",
+        help="How to get the prompts for the benchmark, either 'mmlu' for MMLU questions or "
+        "rng-MIN-MAX for synthetic prompts with random lengths in the interval [MIN, MAX]")
+    parser.add_argument("--n_prompts", type=int, default=100, help="Number of prompts to evaluate")
     parser.add_argument("--n_predict", type=int, default=2048, help="Max. number of tokens to predict per prompt")
+    parser.add_argument(
+        "--n_predict_min", type=int, default=1024,
+        help="Min. number of tokens to predict per prompt (supported for synthetic prompts only)")
     args = parser.parse_args()
     benchmark(**vars(args))

src/llama-chat.cpp

@@ -65,6 +65,7 @@ static const std::map<std::string, llm_chat_template> LLM_CHAT_TEMPLATES = {
     { "llama4",            LLM_CHAT_TEMPLATE_LLAMA4            },
     { "smolvlm",           LLM_CHAT_TEMPLATE_SMOLVLM           },
     { "hunyuan-moe",       LLM_CHAT_TEMPLATE_HUNYUAN_MOE       },
+    { "kimi-k2",           LLM_CHAT_TEMPLATE_KIMI_K2           },
 };
 
 llm_chat_template llm_chat_template_from_str(const std::string & name) {
@@ -188,6 +189,8 @@ llm_chat_template llm_chat_detect_template(const std::string & tmpl) {
         return LLM_CHAT_TEMPLATE_DOTS1;
     } else if (tmpl_contains("<|startoftext|>") && tmpl_contains("<|extra_4|>")) {
         return LLM_CHAT_TEMPLATE_HUNYUAN_MOE;
+    } else if (tmpl_contains("<|im_assistant|>assistant<|im_middle|>")) {
+        return LLM_CHAT_TEMPLATE_KIMI_K2;
     }
     return LLM_CHAT_TEMPLATE_UNKNOWN;
 }
@@ -680,6 +683,26 @@ int32_t llm_chat_apply_template(
                 ss << "<|startoftext|>" << message->content << "<|extra_0|>";
             }
         }
+    } else if (tmpl == LLM_CHAT_TEMPLATE_KIMI_K2) {
+        // moonshotai/Kimi-K2-Instruct
+        for (auto message : chat) {
+            std::string role(message->role);
+            if (role == "system") {
+                ss << "<|im_system|>system<|im_middle|>";
+            } else if (role == "user") {
+                ss << "<|im_user|>user<|im_middle|>";
+            } else if (role == "assistant") {
+                ss << "<|im_assistant|>assistant<|im_middle|>";
+            } else if (role == "tool") {
+                ss << "<|im_system|>tool<|im_middle|>";
+            }
+
+            ss << message->content << "<|im_end|>";
+
+            if (add_ass) {
+                ss << "<|im_assistant|>assistant<|im_middle|>";
+            }
+        }
     } else {
         // template not supported
         return -1;

src/llama-chat.h

@@ -45,6 +45,7 @@ enum llm_chat_template {
     LLM_CHAT_TEMPLATE_SMOLVLM,
     LLM_CHAT_TEMPLATE_DOTS1,
     LLM_CHAT_TEMPLATE_HUNYUAN_MOE,
+    LLM_CHAT_TEMPLATE_KIMI_K2,
     LLM_CHAT_TEMPLATE_UNKNOWN,
 };
 

src/llama-hparams.h

@@ -6,7 +6,7 @@
 
 // bump if necessary
 #define LLAMA_MAX_LAYERS  512
-#define LLAMA_MAX_EXPERTS 256  // DeepSeekV3
+#define LLAMA_MAX_EXPERTS 384  // Kimi-K2
 
 enum llama_expert_gating_func_type {
     LLAMA_EXPERT_GATING_FUNC_TYPE_NONE    = 0,

src/llama-model.cpp

@@ -15763,6 +15763,7 @@ private:
         cb(zx, "mamba_in_proj", il);
         // {8192, 5, 1, 1} -> {8192, 1, 5, 1}
         zx = ggml_permute(ctx0, zx, 0, 2, 1, 3);
+        zx = ggml_cont(ctx0, zx);
         zx = ggml_reshape_4d(ctx0, zx, head_dim * 2, n_heads, n_seq_tokens, n_seqs);
         cb(zx, "mamba_in_proj_out", il);
 
@@ -15780,7 +15781,6 @@ private:
         // conv1d
         {
             // => {d_conv - 1 + n_seq_tokens, d_inner, n_seqs}
-            x = ggml_view_2d(ctx0, x, d_inner, n_seq_tokens * n_seqs, d_inner * x->nb[0], 0);
             ggml_tensor * conv_x = ggml_concat(ctx0, conv, ggml_transpose(ctx0, x), 0);
             cb(conv_x, "mamba_conv1d_input", il);
 

src/llama-vocab.cpp

@@ -405,6 +405,13 @@ struct llm_tokenizer_bpe : llm_tokenizer {
                     "[^\\r\\n\\p{L}\\p{N}]?((?=[\\p{L}])([^a-z]))*((?=[\\p{L}])([^A-Z]))+(?:'[sS]|'[tT]|'[rR][eE]|'[vV][eE]|'[mM]|'[lL][lL]|'[dD])?|[^\\r\\n\\p{L}\\p{N}]?((?=[\\p{L}])([^a-z]))+((?=[\\p{L}])([^A-Z]))*(?:'[sS]|'[tT]|'[rR][eE]|'[vV][eE]|'[mM]|'[lL][lL]|'[dD])?|\\p{N}{1,3}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n/]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+",
                 };
                 break;
+            case LLAMA_VOCAB_PRE_TYPE_KIMI_K2:
+                regex_exprs = {
+                    // K2 trigger pattern - this will activate the custom K2 handler in unicode.cpp
+                    // The custom handler implements all K2 patterns with proper Han character exclusion
+                    "\\p{Han}+",
+                };
+                break;
             case LLAMA_VOCAB_PRE_TYPE_SUPERBPE:
                 regex_exprs = {
                     "\\p{N}+",
@@ -1954,6 +1961,10 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
                 tokenizer_pre == "hunyuan") {
                 pre_type = LLAMA_VOCAB_PRE_TYPE_HUNYUAN;
                 clean_spaces = false;
+            } else if (
+                tokenizer_pre == "kimi-k2") {
+                pre_type = LLAMA_VOCAB_PRE_TYPE_KIMI_K2;
+                clean_spaces = false;
             } else {
                 throw std::runtime_error(format("unknown pre-tokenizer type: '%s'", tokenizer_pre.c_str()));
             }

src/llama-vocab.h

@@ -45,6 +45,7 @@ enum llama_vocab_pre_type {
     LLAMA_VOCAB_PRE_TYPE_PIXTRAL        = 34,
     LLAMA_VOCAB_PRE_TYPE_SEED_CODER     = 35,
     LLAMA_VOCAB_PRE_TYPE_HUNYUAN        = 36,
+    LLAMA_VOCAB_PRE_TYPE_KIMI_K2        = 37,
 };
 
 struct LLM_KV;

src/unicode.cpp

@@ -557,6 +557,178 @@ static std::vector<size_t> unicode_regex_split_stl(const std::string & text, con
     return bpe_offsets;
 }
 
+// K2 system regex patterns (from tokenization_kimi.py):
+// [\p{Han}]+|[^\r\n\p{L}\p{N}]?[\p{Lu}\p{Lt}\p{Lm}\p{Lo}\p{M}&&[^\p{Han}]]*[\p{Ll}\p{Lm}\p{Lo}\p{M}&&[^\p{Han}]]+(?i:'s|'t|'re|'ve|'m|'ll|'d)?|[^\r\n\p{L}\p{N}]?[\p{Lu}\p{Lt}\p{Lm}\p{Lo}\p{M}&&[^\p{Han}]]+[\p{Ll}\p{Lm}\p{Lo}\p{M}&&[^\p{Han}]]*(?i:'s|'t|'re|'ve|'m|'ll|'d)?|\p{N}{1,3}| ?[^\s\p{L}\p{N}]+[\r\n]*|\s*[\r\n]+|\s+(?!\S)|\s+
+static std::vector<size_t> unicode_regex_split_custom_kimi_k2(const std::string & text, const std::vector<size_t> & offsets) {
+    std::vector<size_t> bpe_offsets;
+    bpe_offsets.reserve(offsets.size());
+
+    const auto cpts = unicode_cpts_from_utf8(text);
+
+    size_t start = 0;
+    for (auto offset : offsets) {
+        const size_t offset_ini = start;
+        const size_t offset_end = start + offset;
+        assert(offset_end <= cpts.size());
+        start = offset_end;
+
+        static const uint32_t OUT_OF_RANGE = 0xFFFFFFFF;
+        auto _get_cpt = [&] (const size_t pos) -> uint32_t {
+            return (offset_ini <= pos && pos < offset_end) ? cpts[pos] : OUT_OF_RANGE;
+        };
+
+        auto _get_flags = [&] (const size_t pos) -> unicode_cpt_flags {
+            return (offset_ini <= pos && pos < offset_end) ? unicode_cpt_flags_from_cpt(cpts[pos]) : unicode_cpt_flags{};
+        };
+
+        size_t _prev_end = offset_ini;
+        auto _add_token = [&] (const size_t end) -> size_t {
+            assert(_prev_end <= end && end <= offset_end);
+            size_t len = end - _prev_end;
+            if (len > 0) {
+                bpe_offsets.push_back(len);
+            }
+            _prev_end = end;
+            return len;
+        };
+
+        for (size_t pos = offset_ini; pos < offset_end; /*pos++*/ ) {
+            const uint32_t cpt = _get_cpt(pos);
+            const auto flags = _get_flags(pos);
+
+            // Pattern 1: [\p{Han}]+ (Chinese characters)
+            if (unicode_cpt_is_han(cpt)) {
+                while (unicode_cpt_is_han(_get_cpt(pos))) {
+                    pos++;
+                }
+                _add_token(pos);
+                continue;
+            }
+
+            // Pattern 2 & 3: Letter words excluding Han characters with optional contractions
+            // [^\r\n\p{L}\p{N}]?[\p{Lu}\p{Lt}\p{Lm}\p{Lo}\p{M}&&[^\p{Han}]]*[\p{Ll}\p{Lm}\p{Lo}\p{M}&&[^\p{Han}]]+(?:'s|'t|'re|'ve|'m|'ll|'d)?
+            // [^\r\n\p{L}\p{N}]?[\p{Lu}\p{Lt}\p{Lm}\p{Lo}\p{M}&&[^\p{Han}]]+[\p{Ll}\p{Lm}\p{Lo}\p{M}&&[^\p{Han}]]*(?:'s|'t|'re|'ve|'m|'ll|'d)?
+            // Check if current char is a letter OR if current char could be a leading char and next char is a letter
+            bool is_letter_pattern = (flags.is_letter && !unicode_cpt_is_han(cpt)) ||
+                                     (!(cpt == '\r' || cpt == '\n' || flags.is_letter || flags.is_number) &&
+                                      _get_flags(pos + 1).is_letter && !unicode_cpt_is_han(_get_cpt(pos + 1)));
+
+            if (is_letter_pattern) {
+                // Handle optional leading non-letter/non-number character
+                bool has_leading_char = false;
+                if (!(cpt == '\r' || cpt == '\n' || flags.is_letter || flags.is_number)) {
+                    has_leading_char = true;
+                    pos++;
+                }
+
+                // Match letter sequence (excluding Han characters)
+                bool has_letters = false;
+                while (_get_flags(pos).is_letter && !unicode_cpt_is_han(_get_cpt(pos))) {
+                    has_letters = true;
+                    pos++;
+                }
+
+                // Only proceed if we found letters (after potentially skipping leading char)
+                if (has_letters || (!has_leading_char && _get_flags(pos).is_letter && !unicode_cpt_is_han(_get_cpt(pos)))) {
+                    if (!has_letters) pos++; // consume the first letter if we didn't already
+
+                    // Continue consuming letters
+                    while (_get_flags(pos).is_letter && !unicode_cpt_is_han(_get_cpt(pos))) {
+                        pos++;
+                    }
+
+                    // Check for optional contractions (?:'s|'t|'re|'ve|'m|'ll|'d)
+                    if (_get_cpt(pos) == '\'' && pos + 1 < offset_end) {
+                        uint32_t cpt_next = unicode_tolower(_get_cpt(pos + 1));
+                        if (cpt_next == 's' || cpt_next == 't' || cpt_next == 'm' || cpt_next == 'd') {
+                            pos += 2;
+                        } else if (pos + 2 < offset_end) {
+                            uint32_t cpt_next_next = unicode_tolower(_get_cpt(pos + 2));
+                            if ((cpt_next == 'r' && cpt_next_next == 'e') ||
+                                (cpt_next == 'v' && cpt_next_next == 'e') ||
+                                (cpt_next == 'l' && cpt_next_next == 'l')) {
+                                pos += 3;
+                            }
+                        }
+                    }
+
+                    _add_token(pos);
+                    continue;
+                } else if (has_leading_char) {
+                    // We consumed a leading char but found no letters, backtrack
+                    pos--;
+                }
+            }
+
+            // Pattern 4: \p{N}{1,3} (numbers 1-3 digits)
+            if (flags.is_number) {
+                size_t ini = pos;
+                while (_get_flags(pos).is_number) {
+                    if (++pos - ini >= 3) {
+                        _add_token(pos);
+                        ini = pos;
+                    }
+                }
+                _add_token(pos);
+                continue;
+            }
+
+            // Pattern 5:  ?[^\s\p{L}\p{N}]+[\r\n]* (optional space + non-word chars + optional newlines)
+            auto flags2 = (cpt == ' ' ? _get_flags(pos + 1) : flags);
+            if (!(flags2.is_whitespace || flags2.is_letter || flags2.is_number) && flags2.as_uint()) {
+                pos += (cpt == ' ');
+                while (!(flags2.is_whitespace || flags2.is_letter || flags2.is_number) && flags2.as_uint()) {
+                    flags2 = _get_flags(++pos);
+                }
+                // Match optional [\r\n]*
+                uint32_t cpt2 = _get_cpt(pos);
+                while (cpt2 == '\r' || cpt2 == '\n') {
+                    cpt2 = _get_cpt(++pos);
+                }
+                _add_token(pos);
+                continue;
+            }
+
+            // Count whitespace characters
+            size_t num_whitespaces = 0;
+            size_t last_end_r_or_n = 0;
+            while (_get_flags(pos + num_whitespaces).is_whitespace) {
+                uint32_t cpt2 = _get_cpt(pos + num_whitespaces);
+                if (cpt2 == '\r' || cpt2 == '\n') {
+                    last_end_r_or_n = pos + num_whitespaces + 1;
+                }
+                num_whitespaces++;
+            }
+
+            // Pattern 6: \s*[\r\n]+ (whitespace with newlines)
+            if (last_end_r_or_n > 0) {
+                pos = last_end_r_or_n;
+                _add_token(pos);
+                continue;
+            }
+
+            // Pattern 7: \s+(?!\S) (trailing whitespace)
+            if (num_whitespaces > 1 && _get_cpt(pos + num_whitespaces) != OUT_OF_RANGE) {
+                pos += num_whitespaces - 1;
+                _add_token(pos);
+                continue;
+            }
+
+            // Pattern 8: \s+ (general whitespace)
+            if (num_whitespaces > 0) {
+                pos += num_whitespaces;
+                _add_token(pos);
+                continue;
+            }
+
+            // No matches - consume single character
+            _add_token(++pos);
+        }
+    }
+
+    return bpe_offsets;
+}
+
 static std::vector<size_t> unicode_regex_split_custom(const std::string & text, const std::string & regex_expr, const std::vector<size_t> & offsets) {
     std::vector<size_t> bpe_offsets;
 
@@ -567,6 +739,9 @@ static std::vector<size_t> unicode_regex_split_custom(const std::string & text,
             regex_expr == "(?:'[sS]|'[tT]|'[rR][eE]|'[vV][eE]|'[mM]|'[lL][lL]|'[dD])|[^\\r\\n\\p{L}\\p{N}]?\\p{L}+|\\p{N}{1,3}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+") {
 
         bpe_offsets = unicode_regex_split_custom_llama3(text, offsets);
+    } else if (regex_expr == "\\p{Han}+") {
+        // K2's first pattern - handle all K2 patterns together
+        bpe_offsets = unicode_regex_split_custom_kimi_k2(text, offsets);
     }
 
     return bpe_offsets;
@@ -672,6 +847,38 @@ uint32_t unicode_tolower(uint32_t cpt) {
     return cpt;  // Return the original code point if no lowercase mapping is found
 }
 
+bool unicode_cpt_is_han(uint32_t cpt) {
+    // Han character ranges (Chinese/CJK characters)
+    // CJK Unified Ideographs (most common)
+    if (cpt >= 0x4E00 && cpt <= 0x9FFF) return true;
+
+    // CJK Extension A
+    if (cpt >= 0x3400 && cpt <= 0x4DBF) return true;
+
+    // CJK Extension B
+    if (cpt >= 0x20000 && cpt <= 0x2A6DF) return true;
+
+    // CJK Extension C
+    if (cpt >= 0x2A700 && cpt <= 0x2B73F) return true;
+
+    // CJK Extension D
+    if (cpt >= 0x2B740 && cpt <= 0x2B81F) return true;
+
+    // CJK Extension E
+    if (cpt >= 0x2B820 && cpt <= 0x2CEAF) return true;
+
+    // CJK Extension F
+    if (cpt >= 0x2CEB0 && cpt <= 0x2EBEF) return true;
+
+    // CJK Compatibility Ideographs
+    if (cpt >= 0xF900 && cpt <= 0xFAFF) return true;
+
+    // CJK Compatibility Ideographs Supplement
+    if (cpt >= 0x2F800 && cpt <= 0x2FA1F) return true;
+
+    return false;
+}
+
 std::vector<std::string> unicode_regex_split(const std::string & text, const std::vector<std::string> & regex_exprs) {
     // unicode categories
     static const std::map<std::string, int> k_ucat_enum = {

src/unicode.h

@@ -63,4 +63,6 @@ uint8_t     unicode_utf8_to_byte(const std::string & utf8);
 
 uint32_t unicode_tolower(uint32_t cpt);
 
+bool unicode_cpt_is_han(uint32_t cpt);
+
 std::vector<std::string> unicode_regex_split(const std::string & text, const std::vector<std::string> & regex_exprs);

tools/server/README.md

@@ -7,7 +7,7 @@ Set of LLM REST APIs and a simple web front end to interact with llama.cpp.
 **Features:**
  * LLM inference of F16 and quantized models on GPU and CPU
  * [OpenAI API](https://github.com/openai/openai-openapi) compatible chat completions and embeddings routes
- * Reranking endoint (https://github.com/ggml-org/llama.cpp/pull/9510)
+ * Reranking endpoint (https://github.com/ggml-org/llama.cpp/pull/9510)
  * Parallel decoding with multi-user support
  * Continuous batching
  * Multimodal ([documentation](../../docs/multimodal.md)) / with OpenAI-compatible API support

tools/server/server.cpp

@@ -127,7 +127,6 @@ struct slot_params {
     std::vector<std::string> response_fields;
     bool timings_per_token = false;
     bool post_sampling_probs = false;
-    bool ignore_eos = false;
 
     struct common_params_sampling sampling;
     struct common_params_speculative speculative;
@@ -441,7 +440,6 @@ struct server_task {
 
         {
             params.sampling.logit_bias.clear();
-            params.ignore_eos = json_value(data, "ignore_eos", false);
 
             const auto & logit_bias = data.find("logit_bias");
             if (logit_bias != data.end() && logit_bias->is_array()) {
@@ -472,6 +470,13 @@ struct server_task {
                     }
                 }
             }
+
+            params.sampling.ignore_eos = json_value(data, "ignore_eos", params_base.sampling.ignore_eos);
+            if (params.sampling.ignore_eos) {
+                params.sampling.logit_bias.insert(
+                        params.sampling.logit_bias.end(),
+                        defaults.sampling.logit_bias_eog.begin(), defaults.sampling.logit_bias_eog.end());
+            }
         }
 
         {
@@ -1898,7 +1903,6 @@ struct server_context {
 
     bool clean_kv_cache = true;
     bool add_bos_token  = true;
-    bool has_eos_token  = false;
 
     int32_t n_ctx; // total context for all clients / slots
 
@@ -1957,7 +1961,6 @@ struct server_context {
         n_ctx = llama_n_ctx(ctx);
 
         add_bos_token = llama_vocab_get_add_bos(vocab);
-        has_eos_token = llama_vocab_eos(vocab) != LLAMA_TOKEN_NULL;
 
         if (!params_base.speculative.model.path.empty() || !params_base.speculative.model.hf_repo.empty()) {
             SRV_INF("loading draft model '%s'\n", params_base.speculative.model.path.c_str());
@@ -2217,10 +2220,6 @@ struct server_context {
             slot.params.n_predict = slot.n_predict;
         }
 
-        if (slot.params.ignore_eos && has_eos_token) {
-            slot.params.sampling.logit_bias.push_back({llama_vocab_eos(vocab), -INFINITY});
-        }
-
         {
             if (slot.smpl != nullptr) {
                 common_sampler_free(slot.smpl);