common : merge qwen3-coder and nemotron nano 3 parsers (#19765)

* common : migrate qwen3-coder to PEG parsing variant

* cont : add JSON parameter test

common/chat-parser.cpp

@@ -893,23 +893,6 @@ static void common_chat_parse_minimax_m2(common_chat_msg_parser & builder) {
     builder.consume_reasoning_with_xml_tool_calls(form, "<think>", "</think>");
 }
 
-static void common_chat_parse_qwen3_coder_xml(common_chat_msg_parser & builder) {
-    static const xml_tool_call_format form = ([]() {
-        xml_tool_call_format form {};
-        form.scope_start = "<tool_call>";
-        form.tool_start  = "<function=";
-        form.tool_sep    = ">";
-        form.key_start   = "<parameter=";
-        form.key_val_sep = ">";
-        form.val_end     = "</parameter>";
-        form.tool_end    = "</function>";
-        form.scope_end   = "</tool_call>";
-        form.trim_raw_argval = true;
-        return form;
-    })();
-    builder.consume_reasoning_with_xml_tool_calls(form);
-}
-
 static void common_chat_parse_kimi_k2(common_chat_msg_parser & builder) {
     static const xml_tool_call_format form = ([]() {
         xml_tool_call_format form {};
@@ -1590,9 +1573,6 @@ static void common_chat_parse(common_chat_msg_parser & builder) {
         case COMMON_CHAT_FORMAT_KIMI_K2:
             common_chat_parse_kimi_k2(builder);
             break;
-        case COMMON_CHAT_FORMAT_QWEN3_CODER_XML:
-            common_chat_parse_qwen3_coder_xml(builder);
-            break;
         case COMMON_CHAT_FORMAT_APRIEL_1_5:
             common_chat_parse_apriel_1_5(builder);
             break;

common/chat.cpp

@@ -65,14 +65,25 @@ json common_chat_msg::to_json_oaicompat(bool concat_typed_text) const {
     } else if (!content_parts.empty()) {
         if (concat_typed_text) {
             std::string text;
+            bool last_was_media_marker = false;
+            // join parts with newline, do not add newline before or after media markers
             for (const auto & part : content_parts) {
-                if (part.type != "text") {
+                bool add_new_line = true;
+                if (part.type == "text") {
+                    add_new_line = !last_was_media_marker && !text.empty();
+                    last_was_media_marker = false;
+                } else if (part.type == "media_marker") {
+                    add_new_line = false;
+                    last_was_media_marker = true;
+                } else {
                     LOG_WRN("Ignoring content part type: %s\n", part.type.c_str());
                     continue;
                 }
-                if (!text.empty()) {
+
+                if (add_new_line) {
                     text += '\n';
                 }
+
                 text += part.text;
             }
             jmsg["content"] = text;
@@ -319,7 +330,7 @@ std::vector<common_chat_msg> common_chat_msgs_parse_oaicompat(const json & messa
                             throw std::invalid_argument("Missing content part type: " + part.dump());
                         }
                         const auto & type = part.at("type");
-                        if (type != "text") {
+                        if (type != "text" && type != "media_marker") {
                             throw std::invalid_argument("Unsupported content part type: " + type.dump());
                         }
                         common_chat_msg_content_part msg_part;
@@ -725,7 +736,6 @@ const char * common_chat_format_name(common_chat_format format) {
         case COMMON_CHAT_FORMAT_MINIMAX_M2: return "MiniMax-M2";
         case COMMON_CHAT_FORMAT_GLM_4_5: return "GLM 4.5";
         case COMMON_CHAT_FORMAT_KIMI_K2: return "Kimi K2";
-        case COMMON_CHAT_FORMAT_QWEN3_CODER_XML: return "Qwen3 Coder";
         case COMMON_CHAT_FORMAT_APRIEL_1_5: return "Apriel 1.5";
         case COMMON_CHAT_FORMAT_XIAOMI_MIMO: return "Xiaomi MiMo";
         case COMMON_CHAT_FORMAT_SOLAR_OPEN: return "Solar Open";
@@ -1511,14 +1521,17 @@ static common_chat_params common_chat_params_init_nemotron_v2(const common_chat_
     return data;
 }
 
-static common_chat_params common_chat_params_init_nemotron_v3(const common_chat_template & tmpl, const struct templates_params & inputs) {
+static common_chat_params common_chat_params_init_qwen3_coder(const common_chat_template & tmpl, const struct templates_params & inputs) {
     common_chat_params data;
 
     data.prompt = apply(tmpl, inputs);
     data.format = COMMON_CHAT_FORMAT_PEG_CONSTRUCTED;
 
+    // Nemotron Nano 3 and Step-3.5-Flash use the Qwen3 Coder tool calling with thinking
+    bool supports_reasoning = (tmpl.source().find("<think>") != std::string::npos);
+
     // Handle thinking tags appropriately based on inputs.enable_thinking
-    if (string_ends_with(data.prompt, "<think>\n")) {
+    if (supports_reasoning && string_ends_with(data.prompt, "<think>\n")) {
         if (!inputs.enable_thinking) {
             data.prompt += "</think>";
         } else {
@@ -1527,19 +1540,21 @@ static common_chat_params common_chat_params_init_nemotron_v3(const common_chat_
     }
 
     data.preserved_tokens = {
-        "<think>",
-        "</think>",
         "<tool_call>",
         "</tool_call>",
     };
 
+    if (supports_reasoning) {
+        data.preserved_tokens.insert(data.preserved_tokens.end(), {"<think>", "</think>"});
+    }
+
     auto has_tools = inputs.tools.is_array() && !inputs.tools.empty();
     auto extract_reasoning = inputs.reasoning_format != COMMON_REASONING_FORMAT_NONE;
     auto include_grammar = true;
 
     auto parser = build_chat_peg_constructed_parser([&](auto & p) {
         auto reasoning = p.eps();
-        if (inputs.enable_thinking && extract_reasoning) {
+        if (supports_reasoning && inputs.enable_thinking && extract_reasoning) {
             auto reasoning_content = p.reasoning(p.until("</think>")) + ("</think>" | p.end());
             if (data.thinking_forced_open) {
                 reasoning = reasoning_content;
@@ -1877,38 +1892,6 @@ static common_chat_params common_chat_params_init_minimax_m2(const common_chat_t
     return data;
 }
 
-static common_chat_params common_chat_params_init_qwen3_coder_xml(const common_chat_template & tmpl, const struct templates_params & params) {
-    common_chat_params data;
-    data.grammar_lazy = params.tools.is_array() && !params.tools.empty() && params.tool_choice != COMMON_CHAT_TOOL_CHOICE_REQUIRED;
-
-    data.prompt = apply(tmpl, params);
-    data.format = COMMON_CHAT_FORMAT_QWEN3_CODER_XML;
-
-    data.preserved_tokens = {
-        "<tool_call>",
-        "</tool_call>",
-        "<function=",
-        "</function>",
-        "<parameter=",
-        "</parameter>",
-    };
-
-    // build grammar for tool call
-    static const xml_tool_call_format form {
-        /* form.scope_start = */ "<tool_call>\n",
-        /* form.tool_start  = */ "<function=",
-        /* form.tool_sep    = */ ">\n",
-        /* form.key_start   = */ "<parameter=",
-        /* form.key_val_sep = */ ">\n",
-        /* form.val_end     = */ "\n</parameter>\n",
-        /* form.tool_end    = */ "</function>\n",
-        /* form.scope_end   = */ "</tool_call>",
-    };
-    build_grammar_xml_tool_call(data, params.tools, form);
-
-    return data;
-}
-
 static common_chat_params common_chat_params_init_kimi_k2(const common_chat_template & tmpl, const struct templates_params & params) {
     common_chat_params data;
     data.grammar_lazy = params.tools.is_array() && !params.tools.empty() && params.tool_choice != COMMON_CHAT_TOOL_CHOICE_REQUIRED;
@@ -2032,6 +2015,7 @@ static common_chat_params common_chat_params_init_gpt_oss(const common_chat_temp
         if (has_reasoning_content && has_tool_calls) {
             auto adjusted_message = msg;
             adjusted_message["thinking"] = msg.at("reasoning_content");
+            adjusted_message.erase("content");
             adjusted_messages.push_back(adjusted_message);
         } else {
             adjusted_messages.push_back(msg);
@@ -3129,19 +3113,13 @@ static common_chat_params common_chat_templates_apply_jinja(
     }
 
     // Qwen3-Coder XML format detection (must come before Hermes 2 Pro)
-    // Detect via explicit XML markers unique to Qwen3-Coder to avoid false positives in other templates.
-    // Require presence of <tool_call>, <function=...>, and <parameter=...> blocks.
+    // Detect via XML markers: <tool_call>, <function=...>, and <parameter=...> blocks.
+    // Also matches Step-3.5-Flash and Nemotron 3 Nano which use the same output format.
     if (src.find("<tool_call>") != std::string::npos &&
-        src.find("<function>") != std::string::npos &&
         src.find("<function=") != std::string::npos &&
-        src.find("<parameters>") != std::string::npos &&
         src.find("<parameter=") != std::string::npos) {
         workaround::func_args_not_string(params.messages);
-        // Nemotron 3 Nano 30B A3B
-        if (src.find("<think>") != std::string::npos) {
-            return common_chat_params_init_nemotron_v3(tmpl, params);
-        }
-        return common_chat_params_init_qwen3_coder_xml(tmpl, params);
+        return common_chat_params_init_qwen3_coder(tmpl, params);
     }
 
     // Xiaomi MiMo format detection (must come before Hermes 2 Pro)
@@ -3307,7 +3285,7 @@ static common_chat_params common_chat_templates_apply_legacy(
     for (const auto & msg : inputs.messages) {
         auto content = msg.content;
         for (const auto & part : msg.content_parts) {
-            if (part.type != "text") {
+            if (part.type != "text" && part.type != "media_marker") {
                 LOG_WRN("Ignoring non-text content part: %s\n", part.type.c_str());
                 continue;
             }

common/chat.h

@@ -128,7 +128,6 @@ enum common_chat_format {
     COMMON_CHAT_FORMAT_GLM_4_5,
     COMMON_CHAT_FORMAT_MINIMAX_M2,
     COMMON_CHAT_FORMAT_KIMI_K2,
-    COMMON_CHAT_FORMAT_QWEN3_CODER_XML,
     COMMON_CHAT_FORMAT_APRIEL_1_5,
     COMMON_CHAT_FORMAT_XIAOMI_MIMO,
     COMMON_CHAT_FORMAT_SOLAR_OPEN,

common/jinja/value.cpp

@@ -4,6 +4,7 @@
 // for converting from JSON to jinja values
 #include <nlohmann/json.hpp>
 
+#include <sstream>
 #include <string>
 #include <cctype>
 #include <vector>
@@ -715,8 +716,46 @@ const func_builtins & value_string_t::get_builtins() const {
             return args.get_pos(0);
         }},
         {"tojson", tojson},
-        {"indent", [](const func_args &) -> value {
-            throw not_implemented_exception("String indent builtin not implemented");
+        {"indent", [](const func_args &args) -> value {
+            args.ensure_count(1, 4);
+            value val_input  = args.get_pos(0);
+            value val_width  = args.get_kwarg_or_pos("width", 1);
+            const bool first = args.get_kwarg_or_pos("first", 2)->as_bool(); // undefined == false
+            const bool blank = args.get_kwarg_or_pos("blank", 3)->as_bool(); // undefined == false
+            if (!is_val<value_string>(val_input)) {
+                throw raised_exception("indent() first argument must be a string");
+            }
+            std::string indent;
+            if (is_val<value_int>(val_width)) {
+                indent.assign(val_width->as_int(), ' ');
+            } else if (is_val<value_string>(val_width)) {
+                indent = val_width->as_string().str();
+            } else {
+                indent = "    ";
+            }
+            std::string indented;
+            std::string input = val_input->as_string().str();
+            std::istringstream iss = std::istringstream(input);
+            std::string line;
+            while (std::getline(iss, line)) {
+                if (!indented.empty()) {
+                    indented.push_back('\n');
+                }
+                if ((indented.empty() ? first : (!line.empty() || blank))) {
+                    indented += indent;
+                }
+                indented += line;
+            }
+            if (!input.empty() && input.back() == '\n') {
+                indented.push_back('\n');
+                if (blank) {
+                    indented += indent;
+                }
+            }
+
+            auto res = mk_val<value_string>(indented);
+            res->val_str.mark_input_based_on(val_input->as_string());
+            return res;
         }},
         {"join", [](const func_args &) -> value {
             throw not_implemented_exception("String join builtin not implemented");

convert_hf_to_gguf.py

@@ -1163,6 +1163,9 @@ class TextModel(ModelBase):
         if chkhsh == "b53802fb28e26d645c3a310b34bfe07da813026ec7c7716883404d5e0f8b1901":
             # ref: https://huggingface.co/core42/jais-13b
             res = "jais"
+        if chkhsh == "bc5108ee1eb6a3d600cadd065f63190fbd0554dbc9e4bbd6a0d977970afc8d2a":
+            # ref: https://huggingface.co/inceptionai/Jais-2-8B-Chat
+            res = "jais-2"
         if chkhsh == "7b3e7548e4308f52a76e8229e4e6cc831195d0d1df43aed21ac6c93da05fec5f":
             # ref: https://huggingface.co/WisdomShell/CodeShell-7B
             res = "codeshell"
@@ -3730,6 +3733,13 @@ class Ernie4_5Model(TextModel):
     def set_vocab(self):
         self._set_vocab_sentencepiece()
 
+        tokenizer_config_file = self.dir_model / 'tokenizer_config.json'
+        if tokenizer_config_file.is_file():
+            with open(tokenizer_config_file, "r", encoding="utf-8") as f:
+                tokenizer_config_json = json.load(f)
+                if "add_prefix_space" in tokenizer_config_json:
+                    self.gguf_writer.add_add_space_prefix(tokenizer_config_json["add_prefix_space"])
+
     def set_gguf_parameters(self):
         super().set_gguf_parameters()
 
@@ -3739,6 +3749,10 @@ class Ernie4_5Model(TextModel):
         if (head_dim := self.hparams.get("head_dim")) is None:
             head_dim = self.hparams["hidden_size"] // num_heads
 
+        if "mlp_AR" in name or "vision_model" in name:
+            # skip vision model and projector tensors
+            return
+
         if "ernie." in name:
             name = name.replace("ernie.", "model.")
         # split the qkv weights
@@ -3848,6 +3862,48 @@ class Ernie4_5MoeModel(Ernie4_5Model):
                 raise ValueError(f"Unprocessed experts: {experts}")
 
 
+@ModelBase.register("PaddleOCRVLForConditionalGeneration")
+class PaddleOCRModel(Ernie4_5Model):
+    model_arch = gguf.MODEL_ARCH.PADDLEOCR
+
+
+@ModelBase.register("PaddleOCRVisionModel")
+class PaddleOCRVisionModel(MmprojModel):
+    # PaddleOCR-VL uses a modified version of Siglip
+    min_pixels: int = 0
+    max_pixels: int = 0
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        assert self.hparams_vision is not None
+        self.min_pixels = self.preprocessor_config["min_pixels"]
+        self.max_pixels = self.preprocessor_config["max_pixels"]
+        self.hparams_vision["image_size"] = int(math.sqrt(self.max_pixels))
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        assert self.hparams_vision is not None
+        hparams = self.hparams_vision
+        self.gguf_writer.add_clip_projector_type(gguf.VisionProjectorType.PADDLEOCR)
+        self.gguf_writer.add_vision_max_pixels(self.max_pixels)
+        self.gguf_writer.add_vision_min_pixels(self.min_pixels)
+        self.gguf_writer.add_vision_use_gelu(True)
+        self.gguf_writer.add_vision_attention_layernorm_eps(hparams.get("rms_norm_eps", 1e-6))
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        name = name.replace("visual.", "model.")
+
+        if "vision_model" in name or "mlp_AR" in name:
+            if "packing_position_embedding" in name:
+                return # unused
+            elif "vision_model.head" in name:
+                # we don't yet support image embeddings for this model
+                return
+            else:
+                yield from super().modify_tensors(data_torch, name, bid)
+        return # skip other tensors
+
+
 @ModelBase.register(
     "Qwen2VLModel",
     "Qwen2VLForConditionalGeneration",
@@ -8633,6 +8689,17 @@ class T5EncoderModel(TextModel):
         yield from super().modify_tensors(data_torch, name, bid)
 
 
+@ModelBase.register("Jais2ForCausalLM")
+class Jais2Model(TextModel):
+    model_arch = gguf.MODEL_ARCH.JAIS2
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        hparams = self.hparams
+        head_dim = hparams.get("head_dim", hparams["hidden_size"] // hparams["num_attention_heads"])
+        self.gguf_writer.add_rope_dimension_count(head_dim)
+
+
 @ModelBase.register("JAISLMHeadModel")
 class JaisModel(TextModel):
     model_arch = gguf.MODEL_ARCH.JAIS
@@ -10726,7 +10793,7 @@ class LFM2Model(TextModel):
     def set_gguf_parameters(self):
         # set num_key_value_heads only for attention layers
         self.hparams["num_key_value_heads"] = [
-            self.hparams["num_key_value_heads"] if layer_type == "full_attention" else 0
+            self.hparams["num_key_value_heads"] if layer_type != "conv" else 0
             for layer_type in self.hparams["layer_types"]
         ]
 
@@ -10912,6 +10979,28 @@ class LFM2AudioModel(ConformerAudioModel):
         yield from super().modify_tensors(data_torch, name, bid)
 
 
+@ModelBase.register("Lfm25AudioTokenizer")
+class LFM25AudioTokenizer(LFM2Model):
+    model_arch = gguf.MODEL_ARCH.LFM2
+
+    def set_vocab(self):
+        self._set_vocab_none()
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        self.gguf_writer.add_sliding_window(self.hparams["sliding_window"])
+        self.gguf_writer.add_embedding_length_out(self.hparams["output_size"])
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        if name == "istft.window" or name.startswith("emb.emb"):
+            return
+
+        if name.startswith("lin"):
+            name = name.replace("lin", "dense_2_out")
+
+        yield from super().modify_tensors(data_torch, name, bid)
+
+
 @ModelBase.register("SmallThinkerForCausalLM")
 class SmallThinkerModel(TextModel):
     model_arch = gguf.MODEL_ARCH.SMALLTHINKER
@@ -11003,13 +11092,17 @@ class ModernBertModel(BertModel):
         self.gguf_writer.add_vocab_size(self.hparams["vocab_size"])
 
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
-        # these layers act as MLM head, so we don't need them
-        if name.startswith("decoder."):
-            return
-
         if name.startswith("model."):
             name = name[6:]
 
+        if self.cls_out_labels:
+            # For BertForSequenceClassification (direct projection layer)
+            if name == "classifier.weight":
+                name = "classifier.out_proj.weight"
+
+            if name == "classifier.bias":
+                name = "classifier.out_proj.bias"
+
         yield from super().modify_tensors(data_torch, name, bid)
 
 

convert_hf_to_gguf_update.py

@@ -114,6 +114,7 @@ models = [
     {"name": "gemma",            "tokt": TOKENIZER_TYPE.SPM, "repo": "https://huggingface.co/google/gemma-2b", },
     {"name": "gemma-2",          "tokt": TOKENIZER_TYPE.SPM, "repo": "https://huggingface.co/google/gemma-2-9b", },
     {"name": "jais",             "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/core42/jais-13b", },
+    {"name": "jais-2",           "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/inceptionai/Jais-2-8B-Chat", },
     {"name": "t5",               "tokt": TOKENIZER_TYPE.UGM, "repo": "https://huggingface.co/google-t5/t5-small", },
     {"name": "codeshell",        "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/WisdomShell/CodeShell-7B", },
     {"name": "tekken",           "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/mistralai/Mistral-Nemo-Base-2407", },

docs/ops.md

@@ -31,7 +31,7 @@ Legend:
 |                          CONV_3D | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
 |                CONV_TRANSPOSE_1D | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                CONV_TRANSPOSE_2D | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ |
-|                              COS | ❌ | ✅ | ✅ | ✅ | 🟡 | ❌ | ✅ | 🟡 | ❌ | ❌ | ❌ |
+|                              COS | ❌ | ✅ | ✅ | ✅ | 🟡 | ❌ | ✅ | 🟡 | ✅ | ❌ | ❌ |
 |                      COUNT_EQUAL | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                              CPY | ❌ | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | ❌ | ❌ |
 |               CROSS_ENTROPY_LOSS | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
@@ -96,13 +96,13 @@ Legend:
 |                          SIGMOID | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | 🟡 | ✅ | ❌ | ❌ |
 |                             SILU | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | 🟡 | ✅ | ❌ | ❌ |
 |                        SILU_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ |
-|                              SIN | ❌ | ✅ | ✅ | ✅ | 🟡 | ❌ | ✅ | 🟡 | ❌ | ❌ | ❌ |
+|                              SIN | ❌ | ✅ | ✅ | ✅ | 🟡 | ❌ | ✅ | 🟡 | ✅ | ❌ | ❌ |
 |                         SOFTPLUS | ❌ | ❌ | ✅ | 🟡 | 🟡 | ❌ | ✅ | 🟡 | ✅ | ❌ | ❌ |
 |                         SOFT_MAX | ❌ | 🟡 | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                    SOFT_MAX_BACK | ❌ | ❌ | 🟡 | 🟡 | ❌ | ❌ | 🟡 | ✅ | ❌ | ❌ | ❌ |
 |                        SOLVE_TRI | ❌ | ❌ | ✅ | 🟡 | ❌ | ❌ | ❌ | 🟡 | ❌ | ❌ | ❌ |
-|                              SQR | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 | ❌ | ❌ | ❌ |
-|                             SQRT | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 | ❌ | ❌ | ❌ |
+|                              SQR | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 | ✅ | ❌ | ❌ |
+|                             SQRT | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 | ✅ | ❌ | ❌ |
 |                         SSM_CONV | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                         SSM_SCAN | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | 🟡 | ❌ | ❌ | ❌ |
 |                             STEP | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | ✅ | 🟡 | ✅ | ❌ | ❌ |

docs/ops/WebGPU.csv

@@ -8760,22 +8760,14 @@
 "WebGPU: WebGPU","ADD_ID","type_a=f32,type_b=f32,n_embd=129,n_experts=8,n_experts_used=4,n_token=1","support","0","no","WebGPU"
 "WebGPU: WebGPU","ADD_ID","type_a=f32,type_b=f32,n_embd=129,n_experts=8,n_experts_used=4,n_token=32","support","0","no","WebGPU"
 "WebGPU: WebGPU","ADD_ID","type_a=f32,type_b=f32,n_embd=129,n_experts=8,n_experts_used=4,n_token=129","support","0","no","WebGPU"
-"WebGPU: WebGPU","SQR","type=f16,ne=[10,5,4,3]","support","0","no","WebGPU"
-"WebGPU: WebGPU","SQRT","type=f16,ne=[10,3,3,2]","support","0","no","WebGPU"
 "WebGPU: WebGPU","LOG","type=f16,ne=[10,5,4,3]","support","1","yes","WebGPU"
-"WebGPU: WebGPU","SIN","type=f16,ne=[10,2,2,2]","support","0","no","WebGPU"
-"WebGPU: WebGPU","COS","type=f16,ne=[10,2,2,2]","support","0","no","WebGPU"
 "WebGPU: WebGPU","CLAMP","type=f16,ne=[10,5,4,3],min=-0.500000,max=0.500000","support","1","yes","WebGPU"
 "WebGPU: WebGPU","LEAKY_RELU","type=f16,ne_a=[10,5,4,3],negative_slope=0.100000","support","0","no","WebGPU"
 "WebGPU: WebGPU","FLOOR","type=f16,ne=[10,2,2,2]","support","1","yes","WebGPU"
 "WebGPU: WebGPU","CEIL","type=f16,ne=[10,2,2,2]","support","1","yes","WebGPU"
 "WebGPU: WebGPU","ROUND","type=f16,ne=[10,2,2,2]","support","1","yes","WebGPU"
 "WebGPU: WebGPU","TRUNC","type=f16,ne=[10,2,2,2]","support","1","yes","WebGPU"
-"WebGPU: WebGPU","SQR","type=f16,ne=[7,1,5,3]","support","0","no","WebGPU"
-"WebGPU: WebGPU","SQRT","type=f16,ne=[7,1,5,3]","support","0","no","WebGPU"
 "WebGPU: WebGPU","LOG","type=f16,ne=[7,1,5,3]","support","1","yes","WebGPU"
-"WebGPU: WebGPU","SIN","type=f16,ne=[7,1,5,3]","support","0","no","WebGPU"
-"WebGPU: WebGPU","COS","type=f16,ne=[7,1,5,3]","support","0","no","WebGPU"
 "WebGPU: WebGPU","CLAMP","type=f16,ne=[7,1,5,3],min=-0.500000,max=0.500000","support","1","yes","WebGPU"
 "WebGPU: WebGPU","LEAKY_RELU","type=f16,ne_a=[7,1,5,3],negative_slope=0.100000","support","0","no","WebGPU"
 "WebGPU: WebGPU","FLOOR","type=f16,ne=[7,1,5,3]","support","1","yes","WebGPU"
@@ -8786,22 +8778,14 @@
 "WebGPU: WebGPU","ROUND","type=f16,ne=[1024,1024,1,1]","support","1","yes","WebGPU"
 "WebGPU: WebGPU","TRUNC","type=f16,ne=[7,1,5,3]","support","1","yes","WebGPU"
 "WebGPU: WebGPU","TRUNC","type=f16,ne=[1024,1024,1,1]","support","1","yes","WebGPU"
-"WebGPU: WebGPU","SQR","type=f32,ne=[10,5,4,3]","support","0","no","WebGPU"
-"WebGPU: WebGPU","SQRT","type=f32,ne=[10,3,3,2]","support","0","no","WebGPU"
 "WebGPU: WebGPU","LOG","type=f32,ne=[10,5,4,3]","support","1","yes","WebGPU"
-"WebGPU: WebGPU","SIN","type=f32,ne=[10,2,2,2]","support","0","no","WebGPU"
-"WebGPU: WebGPU","COS","type=f32,ne=[10,2,2,2]","support","0","no","WebGPU"
 "WebGPU: WebGPU","CLAMP","type=f32,ne=[10,5,4,3],min=-0.500000,max=0.500000","support","1","yes","WebGPU"
 "WebGPU: WebGPU","LEAKY_RELU","type=f32,ne_a=[10,5,4,3],negative_slope=0.100000","support","0","no","WebGPU"
 "WebGPU: WebGPU","FLOOR","type=f32,ne=[10,2,2,2]","support","1","yes","WebGPU"
 "WebGPU: WebGPU","CEIL","type=f32,ne=[10,2,2,2]","support","1","yes","WebGPU"
 "WebGPU: WebGPU","ROUND","type=f32,ne=[10,2,2,2]","support","1","yes","WebGPU"
 "WebGPU: WebGPU","TRUNC","type=f32,ne=[10,2,2,2]","support","1","yes","WebGPU"
-"WebGPU: WebGPU","SQR","type=f32,ne=[7,1,5,3]","support","0","no","WebGPU"
-"WebGPU: WebGPU","SQRT","type=f32,ne=[7,1,5,3]","support","0","no","WebGPU"
 "WebGPU: WebGPU","LOG","type=f32,ne=[7,1,5,3]","support","1","yes","WebGPU"
-"WebGPU: WebGPU","SIN","type=f32,ne=[7,1,5,3]","support","0","no","WebGPU"
-"WebGPU: WebGPU","COS","type=f32,ne=[7,1,5,3]","support","0","no","WebGPU"
 "WebGPU: WebGPU","CLAMP","type=f32,ne=[7,1,5,3],min=-0.500000,max=0.500000","support","1","yes","WebGPU"
 "WebGPU: WebGPU","LEAKY_RELU","type=f32,ne_a=[7,1,5,3],negative_slope=0.100000","support","0","no","WebGPU"
 "WebGPU: WebGPU","FLOOR","type=f32,ne=[7,1,5,3]","support","1","yes","WebGPU"
@@ -18901,3 +18885,27 @@
 "WebGPU: WebGPU","CROSS_ENTROPY_LOSS_BACK","type=f32,ne=[30000,1,1,1]","support","0","no","WebGPU"
 "WebGPU: WebGPU","OPT_STEP_ADAMW","type=f32,ne=[10,5,4,3]","support","0","no","WebGPU"
 "WebGPU: WebGPU","OPT_STEP_SGD","type=f32,ne=[10,5,4,3]","support","0","no","WebGPU"
+"WebGPU: WebGPU","SQR","type=f16,ne=[10,5,4,3]","support","1","yes","WebGPU"
+"WebGPU: WebGPU","SQRT","type=f16,ne=[10,3,3,2]","support","1","yes","WebGPU"
+"WebGPU: WebGPU","SIN","type=f16,ne=[10,2,2,2]","support","1","yes","WebGPU"
+"WebGPU: WebGPU","COS","type=f16,ne=[10,2,2,2]","support","1","yes","WebGPU"
+"WebGPU: WebGPU","SQR","type=f16,ne=[7,1,5,3]","support","1","yes","WebGPU"
+"WebGPU: WebGPU","SQR","type=f16,ne=[1024,1024,1,1]","support","1","yes","WebGPU"
+"WebGPU: WebGPU","SQRT","type=f16,ne=[7,1,5,3]","support","1","yes","WebGPU"
+"WebGPU: WebGPU","SQRT","type=f16,ne=[1024,1024,1,1]","support","1","yes","WebGPU"
+"WebGPU: WebGPU","SIN","type=f16,ne=[7,1,5,3]","support","1","yes","WebGPU"
+"WebGPU: WebGPU","SIN","type=f16,ne=[1024,1024,1,1]","support","1","yes","WebGPU"
+"WebGPU: WebGPU","COS","type=f16,ne=[7,1,5,3]","support","1","yes","WebGPU"
+"WebGPU: WebGPU","COS","type=f16,ne=[1024,1024,1,1]","support","1","yes","WebGPU"
+"WebGPU: WebGPU","SQR","type=f32,ne=[10,5,4,3]","support","1","yes","WebGPU"
+"WebGPU: WebGPU","SQRT","type=f32,ne=[10,3,3,2]","support","1","yes","WebGPU"
+"WebGPU: WebGPU","SIN","type=f32,ne=[10,2,2,2]","support","1","yes","WebGPU"
+"WebGPU: WebGPU","COS","type=f32,ne=[10,2,2,2]","support","1","yes","WebGPU"
+"WebGPU: WebGPU","SQR","type=f32,ne=[7,1,5,3]","support","1","yes","WebGPU"
+"WebGPU: WebGPU","SQR","type=f32,ne=[1024,1024,1,1]","support","1","yes","WebGPU"
+"WebGPU: WebGPU","SQRT","type=f32,ne=[7,1,5,3]","support","1","yes","WebGPU"
+"WebGPU: WebGPU","SQRT","type=f32,ne=[1024,1024,1,1]","support","1","yes","WebGPU"
+"WebGPU: WebGPU","SIN","type=f32,ne=[7,1,5,3]","support","1","yes","WebGPU"
+"WebGPU: WebGPU","SIN","type=f32,ne=[1024,1024,1,1]","support","1","yes","WebGPU"
+"WebGPU: WebGPU","COS","type=f32,ne=[7,1,5,3]","support","1","yes","WebGPU"
+"WebGPU: WebGPU","COS","type=f32,ne=[1024,1024,1,1]","support","1","yes","WebGPU"

ggml/src/ggml-cpu/arch-fallback.h

@@ -171,15 +171,9 @@
 #elif defined(__riscv)
 // quants.c
 #define quantize_row_q8_K_generic quantize_row_q8_K
-#define ggml_vec_dot_tq1_0_q8_K_generic ggml_vec_dot_tq1_0_q8_K
-#define ggml_vec_dot_tq2_0_q8_K_generic ggml_vec_dot_tq2_0_q8_K
 #define ggml_vec_dot_iq2_xxs_q8_K_generic ggml_vec_dot_iq2_xxs_q8_K
 #define ggml_vec_dot_iq2_xs_q8_K_generic ggml_vec_dot_iq2_xs_q8_K
-#define ggml_vec_dot_iq2_s_q8_K_generic ggml_vec_dot_iq2_s_q8_K
 #define ggml_vec_dot_iq3_xxs_q8_K_generic ggml_vec_dot_iq3_xxs_q8_K
-#define ggml_vec_dot_iq3_s_q8_K_generic ggml_vec_dot_iq3_s_q8_K
-#define ggml_vec_dot_iq1_s_q8_K_generic ggml_vec_dot_iq1_s_q8_K
-#define ggml_vec_dot_iq1_m_q8_K_generic ggml_vec_dot_iq1_m_q8_K
 #define ggml_vec_dot_iq4_nl_q8_0_generic ggml_vec_dot_iq4_nl_q8_0
 #define ggml_vec_dot_iq4_xs_q8_K_generic ggml_vec_dot_iq4_xs_q8_K
 #define ggml_vec_dot_mxfp4_q8_0_generic ggml_vec_dot_mxfp4_q8_0

ggml/src/ggml-cpu/arch/riscv/quants.c

@@ -1954,3 +1954,773 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
 #endif
 }
 
+static const uint8_t sign_gather_indices_arr[64] = {
+    0,0,0,0,0,0,0,0, 1,1,1,1,1,1,1,1, 2,2,2,2,2,2,2,2, 3,3,3,3,3,3,3,3,
+    4,4,4,4,4,4,4,4, 5,5,5,5,5,5,5,5, 6,6,6,6,6,6,6,6, 7,7,7,7,7,7,7,7
+};
+
+static const uint8_t sign_bit_masks_arr[64] = {
+    1,2,4,8,16,32,64,128, 1,2,4,8,16,32,64,128, 1,2,4,8,16,32,64,128, 1,2,4,8,16,32,64,128,
+    1,2,4,8,16,32,64,128, 1,2,4,8,16,32,64,128, 1,2,4,8,16,32,64,128, 1,2,4,8,16,32,64,128
+};
+
+static void ggml_vec_dot_iq2_s_q8_K_vl256(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
+    assert(n % QK_K == 0);
+    UNUSED(nrc); UNUSED(bx); UNUSED(by); UNUSED(bs);
+
+    const block_iq2_s * GGML_RESTRICT x = vx;
+    const block_q8_K  * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK_K;
+    const uint64_t * grid64 = (const uint64_t *)iq2s_grid;
+
+    // --- Pre-load Constants ---
+    uint16_t gather_qh_arr[8] = {0, 0, 0, 0, 1, 1, 1, 1};
+    vuint16mf2_t v_gather_qh = __riscv_vle16_v_u16mf2(gather_qh_arr, 8);
+    uint16_t shift_qh_arr[8] = {11, 9, 7, 5, 11, 9, 7, 5};
+    vuint16mf2_t v_shift_qh = __riscv_vle16_v_u16mf2(shift_qh_arr, 8);
+
+    // Constants for sign extraction
+    vuint8m2_t v_sign_gather_indices = __riscv_vle8_v_u8m2(sign_gather_indices_arr, 64);
+    vuint8m2_t v_sign_masks = __riscv_vle8_v_u8m2(sign_bit_masks_arr, 64);
+
+    float sumf = 0.0f;
+
+    for (int i = 0; i < nb; ++i) {
+        const float combined_scale = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
+
+        const uint8_t * GGML_RESTRICT qs = x[i].qs;
+        const uint8_t * GGML_RESTRICT qh = x[i].qh;
+        const uint8_t * GGML_RESTRICT scales = x[i].scales;
+        const int8_t  * GGML_RESTRICT q8 = y[i].qs;
+
+        const uint8_t * signs_ptr = qs + 32;
+
+        float sum_block = 0.0f;
+
+        for (int ib = 0; ib < 4; ++ib) {
+            // Combine low + high bits
+            vuint8mf4_t v_qs_u8 = __riscv_vle8_v_u8mf4(qs, 8);
+            qs += 8;
+            uint16_t qh_val;
+            memcpy(&qh_val, qh, 2);
+            qh += 2;
+            vuint8mf8_t v_qh_raw = __riscv_vle8_v_u8mf8((const uint8_t*)&qh_val, 2);
+            vuint16mf4_t v_qh_u16 = __riscv_vwcvtu_x_x_v_u16mf4(v_qh_raw, 2);
+            vuint16mf2_t v_qh_u16_ext = __riscv_vlmul_ext_v_u16mf4_u16mf2(v_qh_u16);
+            vuint16mf2_t v_qh_expanded = __riscv_vrgather_vv_u16mf2(v_qh_u16_ext, v_gather_qh, 8);
+            v_qh_expanded = __riscv_vsll_vv_u16mf2(v_qh_expanded, v_shift_qh, 8);
+
+            // Mask: We want bits 11-12. 0x1800 = 0001 1000 0000 0000
+            v_qh_expanded = __riscv_vand_vx_u16mf2(v_qh_expanded, 0x1800, 8);
+            vuint16mf2_t v_qs_u16 = __riscv_vwcvtu_x_x_v_u16mf2(v_qs_u8, 8);
+
+            // Multiply by 8 to get byte offset, instead of element offset
+            v_qs_u16 = __riscv_vsll_vx_u16mf2(v_qs_u16, 3, 8);
+            vuint16mf2_t v_grid_offsets = __riscv_vor_vv_u16mf2(v_qs_u16, v_qh_expanded, 8);
+
+            // Lookup Grid using Byte Offsets
+            vuint64m2_t v_grid_vals = __riscv_vluxei16_v_u64m2(grid64, v_grid_offsets, 8);
+
+            vuint8m2_t v_grid_u8 = __riscv_vreinterpret_v_u64m2_u8m2(v_grid_vals);
+            vint8m2_t v_grid_i8 = __riscv_vreinterpret_v_u8m2_i8m2(v_grid_u8);
+
+            // Load signs and generate sign mask
+            vuint8mf4_t v_signs_raw = __riscv_vle8_v_u8mf4(signs_ptr, 8);
+            signs_ptr += 8;
+
+            vuint8m2_t v_signs_source = __riscv_vlmul_ext_v_u8mf4_u8m2(v_signs_raw);
+            vuint8m2_t v_signs_bcast = __riscv_vrgather_vv_u8m2(v_signs_source, v_sign_gather_indices, 64);
+
+            vuint8m2_t v_sign_bits = __riscv_vand_vv_u8m2(v_signs_bcast, v_sign_masks, 64);
+            vbool4_t m_negative = __riscv_vmsne_vx_u8m2_b4(v_sign_bits, 0, 64);
+
+            vint8m2_t v_q8 = __riscv_vle8_v_i8m2(q8, 64);
+            q8 += 64;
+
+            vint8m2_t v_q8_signed = __riscv_vrsub_vx_i8m2_mu(m_negative, v_q8, v_q8, 0, 64);
+            vint16m4_t v_dot = __riscv_vwmul_vv_i16m4(v_grid_i8, v_q8_signed, 64);
+
+            vint32m1_t v_zero = __riscv_vmv_v_x_i32m1(0, 1);
+
+            int32_t s0 = __riscv_vmv_x_s_i32m1_i32(__riscv_vwredsum_vs_i16m1_i32m1(
+                __riscv_vget_v_i16m4_i16m1(v_dot, 0), v_zero, 16));
+            int32_t s1 = __riscv_vmv_x_s_i32m1_i32(__riscv_vwredsum_vs_i16m1_i32m1(
+                __riscv_vget_v_i16m4_i16m1(v_dot, 1), v_zero, 16));
+            int32_t s2 = __riscv_vmv_x_s_i32m1_i32(__riscv_vwredsum_vs_i16m1_i32m1(
+                __riscv_vget_v_i16m4_i16m1(v_dot, 2), v_zero, 16));
+            int32_t s3 = __riscv_vmv_x_s_i32m1_i32(__riscv_vwredsum_vs_i16m1_i32m1(
+                __riscv_vget_v_i16m4_i16m1(v_dot, 3), v_zero, 16));
+
+            uint8_t sc0 = scales[0];
+            uint8_t sc1 = scales[1];
+            scales += 2;
+
+            sum_block += s0 * (2 * (sc0 & 0xF) + 1);
+            sum_block += s1 * (2 * (sc0 >> 4)  + 1);
+            sum_block += s2 * (2 * (sc1 & 0xF) + 1);
+            sum_block += s3 * (2 * (sc1 >> 4)  + 1);
+        }
+        sumf += sum_block * combined_scale;
+    }
+    *s = 0.125f * sumf;
+}
+
+static void ggml_vec_dot_iq2_s_q8_K_vl128(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
+    assert(n % QK_K == 0);
+    UNUSED(nrc); UNUSED(bx); UNUSED(by); UNUSED(bs);
+
+    const block_iq2_s * GGML_RESTRICT x = vx;
+    const block_q8_K  * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK_K;
+    const uint64_t * grid64 = (const uint64_t *)iq2s_grid;
+
+    // Pre-load Constants
+    vuint8m2_t v_ids = __riscv_vid_v_u8m2(32);
+    vuint8m2_t v_sign_gather_indices = __riscv_vsrl_vx_u8m2(v_ids, 3, 32);
+    vuint8m2_t v_ones = __riscv_vmv_v_x_u8m2(1, 32);
+    vuint8m2_t v_shift_amts = __riscv_vand_vx_u8m2(v_ids, 7, 32);
+    vuint8m2_t v_sign_masks = __riscv_vsll_vv_u8m2(v_ones, v_shift_amts, 32);
+    uint16_t shift_qh_arr[4] = {11, 9, 7, 5};
+    vuint16mf2_t v_shift_qh = __riscv_vle16_v_u16mf2(shift_qh_arr, 4);
+
+    float sumf = 0.0f;
+
+    for (int i = 0; i < nb; ++i) {
+        const float combined_scale = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
+
+        const uint8_t * GGML_RESTRICT qs = x[i].qs;
+        const uint8_t * GGML_RESTRICT qh = x[i].qh;
+        const uint8_t * GGML_RESTRICT scales = x[i].scales;
+        const int8_t  * GGML_RESTRICT q8 = y[i].qs;
+
+        const uint8_t * signs_ptr = qs + 32;
+        float sum_block = 0.0f;
+
+        for (int ib = 0; ib < 8; ++ib) {
+
+            // Load Low Bits [4 bytes]
+            vuint8mf4_t v_qs_u8 = __riscv_vle8_v_u8mf4(qs, 4);
+            qs += 4;
+
+            // Load 1 byte. It contains bits for 4 mini-blocks.
+            uint8_t qh_val = *qh++;
+
+            // Combine Low + High bits of 10bit indices
+            vuint8mf4_t v_qh_raw = __riscv_vmv_v_x_u8mf4(qh_val, 4);
+            vuint16mf2_t v_qh_u16 = __riscv_vwcvtu_x_x_v_u16mf2(v_qh_raw, 4);
+            vuint16mf2_t v_qh_mf2 = __riscv_vsll_vv_u16mf2(v_qh_u16, v_shift_qh, 4);
+            v_qh_mf2 = __riscv_vand_vx_u16mf2(v_qh_mf2, 0x1800, 4);
+            vuint16mf2_t v_qs_u16_mf2 = __riscv_vwcvtu_x_x_v_u16mf2(v_qs_u8, 4);
+            vuint16mf2_t v_qs_u16 = __riscv_vsll_vx_u16mf2(v_qs_u16_mf2, 3, 4);
+            vuint16mf2_t v_grid_offsets = __riscv_vor_vv_u16mf2(v_qs_u16, v_qh_mf2, 4);
+
+            // Lookup Grid
+            vint8m2_t v_grid_i8 = __riscv_vreinterpret_v_u8m2_i8m2(__riscv_vreinterpret_v_u64m2_u8m2(__riscv_vluxei16_v_u64m2(grid64, v_grid_offsets, 4)));
+
+            vuint8mf4_t v_signs_raw = __riscv_vle8_v_u8mf4(signs_ptr, 4);
+            signs_ptr += 4;
+            vuint8m2_t v_signs_source = __riscv_vlmul_ext_v_u8mf4_u8m2(v_signs_raw);
+            vuint8m2_t v_signs_bcast = __riscv_vrgather_vv_u8m2(v_signs_source, v_sign_gather_indices, 32);
+
+            // generating sign mask
+            vuint8m2_t v_sign_bits = __riscv_vand_vv_u8m2(v_signs_bcast, v_sign_masks, 32);
+            vbool4_t m_negative = __riscv_vmsne_vx_u8m2_b4(v_sign_bits, 0, 32);
+
+            vint8m2_t v_q8 = __riscv_vle8_v_i8m2(q8, 32);
+            q8 += 32;
+
+            // apply signs
+            vint8m2_t v_q8_signed = __riscv_vrsub_vx_i8m2_mu(m_negative,v_q8, v_q8, 0, 32);
+            vint16m4_t v_dot = __riscv_vwmul_vv_i16m4(v_grid_i8, v_q8_signed, 32);
+
+            // Reduction
+            vint32m1_t v_zero = __riscv_vmv_v_x_i32m1(0, 1);
+
+            // Reduce 0-15 (First Half)
+            int32_t s0 = __riscv_vmv_x_s_i32m1_i32(__riscv_vwredsum_vs_i16m2_i32m1(
+                __riscv_vget_v_i16m4_i16m2(v_dot, 0), v_zero, 16));
+
+            // Reduce 16-31 (Second Half)
+            int32_t s1 = __riscv_vmv_x_s_i32m1_i32(__riscv_vwredsum_vs_i16m2_i32m1(
+                __riscv_vget_v_i16m4_i16m2(v_dot, 1), v_zero, 16));
+
+            // Apply sub Scales
+            uint8_t sc = *scales++;
+
+            sum_block += s0 * (2 * (sc & 0xF) + 1);
+            sum_block += s1 * (2 * (sc >> 4)  + 1);
+        }
+        sumf += sum_block * combined_scale;
+    }
+    *s = 0.125f * sumf;
+}
+
+void ggml_vec_dot_iq2_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
+#if defined __riscv_v_intrinsic
+    switch (__riscv_vlenb() * 8) {
+        case 128:
+            ggml_vec_dot_iq2_s_q8_K_vl128(n, s, bs, vx, bx, vy, by, nrc);
+            break;
+        case 256:
+            ggml_vec_dot_iq2_s_q8_K_vl256(n, s, bs, vx, bx, vy, by, nrc);
+            break;
+        default:
+            ggml_vec_dot_iq2_s_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
+            break;
+    }
+#else
+    ggml_vec_dot_iq2_s_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
+#endif
+}
+
+static void ggml_vec_dot_iq3_s_q8_K_vl256(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
+    assert(n % QK_K == 0);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_iq3_s * GGML_RESTRICT x = vx;
+    const block_q8_K  * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK_K;
+
+    const uint64_t * grid64 = (const uint64_t *)iq3s_grid;
+
+    // --- Pre-load Constants ---
+    const uint16_t qh_bit_shifts_arr[16] = {
+        0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15
+    };
+    vuint8m2_t v_sign_gather_indices = __riscv_vle8_v_u8m2(sign_gather_indices_arr, 64);
+    vuint8m2_t v_sign_masks = __riscv_vle8_v_u8m2(sign_bit_masks_arr, 64);
+    vuint16m1_t v_qh_shifts = __riscv_vle16_v_u16m1(qh_bit_shifts_arr, 16);
+
+    float sumf = 0.0f;
+
+    for (int i = 0; i < nb; ++i) {
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d);
+        const float combined_scale = d * y[i].d;
+
+        const uint8_t * GGML_RESTRICT qs = x[i].qs;
+        const uint8_t * GGML_RESTRICT qh = x[i].qh;
+        const uint8_t * GGML_RESTRICT scales = x[i].scales;
+        const uint8_t * GGML_RESTRICT signs = x[i].signs;
+        const int8_t  * GGML_RESTRICT q8 = y[i].qs;
+
+        float sum_block = 0.0f;
+
+        // Loop: Process 64 weights (16 mini-blocks of 4) per iteration
+        for (int ib = 0; ib < 4; ++ib) {
+
+            vuint8mf2_t v_qs_u8 = __riscv_vle8_v_u8mf2(qs, 16);
+            qs += 16;
+
+            uint16_t qh_val;
+            memcpy(&qh_val, qh, 2);
+            qh += 2;
+
+            vuint16m1_t v_qh_val = __riscv_vmv_v_x_u16m1(qh_val, 16);
+            // Extract bits: (qh >> i) & 1
+            v_qh_val = __riscv_vsrl_vv_u16m1(v_qh_val, v_qh_shifts, 16);
+            v_qh_val = __riscv_vand_vx_u16m1(v_qh_val, 1, 16);
+
+            vuint16m1_t v_qs_u16 = __riscv_vwcvtu_x_x_v_u16m1(v_qs_u8, 16);
+            v_qs_u16 = __riscv_vsll_vx_u16m1(v_qs_u16, 2, 16);
+            v_qh_val = __riscv_vsll_vx_u16m1(v_qh_val, 10, 16);
+            vuint16m1_t v_grid_offsets = __riscv_vor_vv_u16m1(v_qs_u16, v_qh_val, 16);
+
+            // Grid value is 4xuint8
+            vuint32m2_t v_grid_packed = __riscv_vluxei16_v_u32m2((const uint32_t *)grid64, v_grid_offsets, 16);
+            vuint8m2_t v_grid_u8 = __riscv_vreinterpret_v_u32m2_u8m2(v_grid_packed);
+            vuint8mf4_t v_signs_raw = __riscv_vle8_v_u8mf4(signs, 8);
+            signs += 8;
+
+            // Generate sign mask
+            vuint8m2_t v_signs_source = __riscv_vlmul_ext_v_u8mf4_u8m2(v_signs_raw);
+            vuint8m2_t v_signs_bcast = __riscv_vrgather_vv_u8m2(v_signs_source, v_sign_gather_indices, 64);
+            vuint8m2_t v_sign_bits = __riscv_vand_vv_u8m2(v_signs_bcast, v_sign_masks, 64);
+            vbool4_t m_negative = __riscv_vmsne_vx_u8m2_b4(v_sign_bits, 0, 64);
+
+            vint8m2_t v_q8 = __riscv_vle8_v_i8m2(q8, 64);
+            q8 += 64;
+
+            // Apply Signs
+            vint8m2_t v_q8_signed = __riscv_vrsub_vx_i8m2_mu(m_negative, v_q8, v_q8, 0, 64);
+            vint16m4_t v_dot = __riscv_vwmulsu_vv_i16m4(v_q8_signed, v_grid_u8, 64);
+
+            // Reduction
+            vint16m2_t v_dot_lo = __riscv_vget_v_i16m4_i16m2(v_dot, 0);
+            vint16m2_t v_dot_hi = __riscv_vget_v_i16m4_i16m2(v_dot, 1);
+            vint32m1_t v_zero = __riscv_vmv_v_x_i32m1(0, 1);
+
+            int32_t s_lo = __riscv_vmv_x_s_i32m1_i32(__riscv_vwredsum_vs_i16m2_i32m1(v_dot_lo, v_zero, 32));
+            int32_t s_hi = __riscv_vmv_x_s_i32m1_i32(__riscv_vwredsum_vs_i16m2_i32m1(v_dot_hi, v_zero, 32));
+
+            // Apply sub-scales
+            uint8_t sc_byte = *scales++;
+            int sc_lo = (sc_byte & 0xF) * 2 + 1;
+            int sc_hi = (sc_byte >> 4)  * 2 + 1;
+
+            sum_block += s_lo * sc_lo + s_hi * sc_hi;
+        }
+        sumf += sum_block * combined_scale;
+    }
+    *s = 0.125f * sumf;
+}
+
+void ggml_vec_dot_iq3_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
+#if defined __riscv_v_intrinsic
+    switch (__riscv_vlenb() * 8) {
+        case 256:
+            ggml_vec_dot_iq3_s_q8_K_vl256(n, s, bs, vx, bx, vy, by, nrc);
+            break;
+        default:
+            ggml_vec_dot_iq3_s_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
+            break;
+    }
+#else
+    ggml_vec_dot_iq3_s_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
+#endif
+}
+
+static void ggml_vec_dot_tq1_0_q8_K_vl256(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_tq1_0 * GGML_RESTRICT x = vx;
+    const block_q8_K  * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK_K;
+
+    float sumf = 0.0f;
+    uint8_t pow[16] = {1, 1, 1, 1, 3, 3, 3, 3, 9, 9, 9, 9, 27, 27, 27, 27};
+
+    for (int i = 0; i < nb; i++) {
+        // First loop.
+        vint32m4_t suml1;
+        {
+            const int vl = 32;
+            vuint8m1_t tq = __riscv_vle8_v_u8m1(x[i].qs, vl);
+
+            vuint16m2_t tq0 = __riscv_vsrl_vx_u16m2(__riscv_vwmulu_vx_u16m2(tq, 3, vl), 8, vl);
+            vuint16m2_t tq1 = __riscv_vsrl_vx_u16m2(__riscv_vwmulu_vx_u16m2(__riscv_vmul_vx_u8m1(tq, 3, vl), 3, vl), 8, vl);
+            vuint16m2_t tq2 = __riscv_vsrl_vx_u16m2(__riscv_vwmulu_vx_u16m2(__riscv_vmul_vx_u8m1(tq, 9, vl), 3, vl), 8, vl);
+            vuint16m2_t tq3 = __riscv_vsrl_vx_u16m2(__riscv_vwmulu_vx_u16m2(__riscv_vmul_vx_u8m1(tq, 27, vl), 3, vl), 8, vl);
+            vuint16m2_t tq4 = __riscv_vsrl_vx_u16m2(__riscv_vwmulu_vx_u16m2(__riscv_vmul_vx_u8m1(tq, 81, vl), 3, vl), 8, vl);
+
+            vint16m2_t q80 = __riscv_vwcvt_x_x_v_i16m2(__riscv_vle8_v_i8m1(y[i].qs + 0, vl), vl);
+            vint16m2_t q81 = __riscv_vwcvt_x_x_v_i16m2(__riscv_vle8_v_i8m1(y[i].qs + 32, vl), vl);
+            vint16m2_t q82 = __riscv_vwcvt_x_x_v_i16m2(__riscv_vle8_v_i8m1(y[i].qs + 64, vl), vl);
+            vint16m2_t q83 = __riscv_vwcvt_x_x_v_i16m2(__riscv_vle8_v_i8m1(y[i].qs + 96, vl), vl);
+            vint16m2_t q84 = __riscv_vwcvt_x_x_v_i16m2(__riscv_vle8_v_i8m1(y[i].qs + 128, vl), vl);
+
+            vint16m2_t sum0 = __riscv_vmul_vv_i16m2(__riscv_vreinterpret_v_u16m2_i16m2(__riscv_vsub_vx_u16m2(tq0, 1, vl)), q80, vl);
+            vint16m2_t sum1 = __riscv_vmul_vv_i16m2(__riscv_vreinterpret_v_u16m2_i16m2(__riscv_vsub_vx_u16m2(tq1, 1, vl)), q81, vl);
+            vint16m2_t sum2 = __riscv_vmul_vv_i16m2(__riscv_vreinterpret_v_u16m2_i16m2(__riscv_vsub_vx_u16m2(tq2, 1, vl)), q82, vl);
+            vint16m2_t sum3 = __riscv_vmul_vv_i16m2(__riscv_vreinterpret_v_u16m2_i16m2(__riscv_vsub_vx_u16m2(tq3, 1, vl)), q83, vl);
+            vint16m2_t sum4 = __riscv_vmul_vv_i16m2(__riscv_vreinterpret_v_u16m2_i16m2(__riscv_vsub_vx_u16m2(tq4, 1, vl)), q84, vl);
+
+            vint32m4_t sumi0 = __riscv_vwadd_vv_i32m4(sum0, sum1, vl);
+            vint32m4_t sumi1 = __riscv_vwadd_vv_i32m4(sum2, sum3, vl);
+            suml1 = __riscv_vadd_vv_i32m4(__riscv_vwcvt_x_x_v_i32m4(sum4, vl), __riscv_vadd_vv_i32m4(sumi0, sumi1, vl), vl);
+        }
+
+        // Second loop.
+        vint32m2_t suml2;
+        {
+            const int vl = 16;
+            vuint8mf2_t tq = __riscv_vle8_v_u8mf2(x[i].qs + 32, vl);
+
+            vuint16m1_t tq0 = __riscv_vsrl_vx_u16m1(__riscv_vwmulu_vx_u16m1(tq, 3 * 1, vl), 8, vl);
+            vuint16m1_t tq1 = __riscv_vsrl_vx_u16m1(__riscv_vwmulu_vx_u16m1(__riscv_vmul_vx_u8mf2(tq, 3, vl), 3, vl), 8, vl);
+            vuint16m1_t tq2 = __riscv_vsrl_vx_u16m1(__riscv_vwmulu_vx_u16m1(__riscv_vmul_vx_u8mf2(tq, 9, vl), 3, vl), 8, vl);
+            vuint16m1_t tq3 = __riscv_vsrl_vx_u16m1(__riscv_vwmulu_vx_u16m1(__riscv_vmul_vx_u8mf2(tq, 27, vl), 3, vl), 8, vl);
+            vuint16m1_t tq4 = __riscv_vsrl_vx_u16m1(__riscv_vwmulu_vx_u16m1(__riscv_vmul_vx_u8mf2(tq, 81, vl), 3, vl), 8, vl);
+
+            vint16m1_t q80 = __riscv_vwcvt_x_x_v_i16m1(__riscv_vle8_v_i8mf2(y[i].qs + 160, vl), vl);
+            vint16m1_t q81 = __riscv_vwcvt_x_x_v_i16m1(__riscv_vle8_v_i8mf2(y[i].qs + 176, vl), vl);
+            vint16m1_t q82 = __riscv_vwcvt_x_x_v_i16m1(__riscv_vle8_v_i8mf2(y[i].qs + 192, vl), vl);
+            vint16m1_t q83 = __riscv_vwcvt_x_x_v_i16m1(__riscv_vle8_v_i8mf2(y[i].qs + 208, vl), vl);
+            vint16m1_t q84 = __riscv_vwcvt_x_x_v_i16m1(__riscv_vle8_v_i8mf2(y[i].qs + 224, vl), vl);
+
+            vint16m1_t sum0 = __riscv_vmul_vv_i16m1(__riscv_vreinterpret_v_u16m1_i16m1(__riscv_vsub_vx_u16m1(tq0, 1, vl)), q80, vl);
+            vint16m1_t sum1 = __riscv_vmul_vv_i16m1(__riscv_vreinterpret_v_u16m1_i16m1(__riscv_vsub_vx_u16m1(tq1, 1, vl)), q81, vl);
+            vint16m1_t sum2 = __riscv_vmul_vv_i16m1(__riscv_vreinterpret_v_u16m1_i16m1(__riscv_vsub_vx_u16m1(tq2, 1, vl)), q82, vl);
+            vint16m1_t sum3 = __riscv_vmul_vv_i16m1(__riscv_vreinterpret_v_u16m1_i16m1(__riscv_vsub_vx_u16m1(tq3, 1, vl)), q83, vl);
+            vint16m1_t sum4 = __riscv_vmul_vv_i16m1(__riscv_vreinterpret_v_u16m1_i16m1(__riscv_vsub_vx_u16m1(tq4, 1, vl)), q84, vl);
+
+            vint32m2_t sumi0 = __riscv_vwadd_vv_i32m2(sum0, sum1, vl);
+            vint32m2_t sumi1 = __riscv_vwadd_vv_i32m2(sum2, sum3, vl);
+            suml2 = __riscv_vadd_vv_i32m2(__riscv_vwcvt_x_x_v_i32m2(sum4, vl), __riscv_vadd_vv_i32m2(sumi0, sumi1, vl), vl);
+        }
+
+        // Third loop.
+        vint32m2_t suml3;
+        {
+            const int vl = 16;
+
+            uint32_t qh;
+            memcpy(&qh, &x[i].qh[0], 4);
+            // Prevent fusion with vmv.
+            __asm__ __volatile__("" : "+r"(qh));
+            vuint8mf2_t tq = __riscv_vreinterpret_v_u32mf2_u8mf2(__riscv_vmv_v_x_u32mf2(qh, vl / 4));
+
+            vuint8mf2_t p = __riscv_vle8_v_u8mf2(pow, vl);
+
+            vuint16m1_t tq0 = __riscv_vsrl_vx_u16m1(__riscv_vwmulu_vx_u16m1(__riscv_vmul_vv_u8mf2(tq, p, vl), 3, vl), 8, vl);
+
+            vint16m1_t q80 = __riscv_vwcvt_x_x_v_i16m1(__riscv_vle8_v_i8mf2(y[i].qs + 240, vl), vl);
+
+            vint16m1_t sum0 = __riscv_vmul_vv_i16m1(__riscv_vreinterpret_v_u16m1_i16m1(__riscv_vsub_vx_u16m1(tq0, 1, vl)), q80, vl);
+            suml3 = __riscv_vwcvt_x_x_v_i32m2(sum0, vl);
+        }
+
+        vint32m2_t sumb = __riscv_vadd_vv_i32m2(__riscv_vget_v_i32m4_i32m2(suml1, 0), __riscv_vget_v_i32m4_i32m2(suml1, 1), 16);
+        sumb = __riscv_vadd_vv_i32m2(sumb, suml2, 16);
+        sumb = __riscv_vadd_vv_i32m2(sumb, suml3, 16);
+
+        vint32m1_t sum = __riscv_vredsum_vs_i32m2_i32m1(sumb, __riscv_vmv_v_x_i32m1(0, 1), 16);
+        sumf += __riscv_vmv_x_s_i32m1_i32(sum) * y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+    }
+
+    *s = sumf;
+}
+
+void ggml_vec_dot_tq1_0_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
+#if defined __riscv_v_intrinsic
+    switch (__riscv_vlenb() * 8) {
+        case 256:
+            ggml_vec_dot_tq1_0_q8_K_vl256(n, s, bs, vx, bx, vy, by, nrc);
+            break;
+        default:
+            ggml_vec_dot_tq1_0_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
+            break;
+    }
+#else
+    ggml_vec_dot_tq1_0_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
+#endif
+}
+
+static void ggml_vec_dot_tq2_0_q8_K_vl256(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
+    assert(n % QK_K == 0);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_tq2_0 * GGML_RESTRICT x = vx;
+    const block_q8_K  * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK_K;
+
+    float sumf = 0.0f;
+    for (int i = 0; i < nb; ++i) {
+        int32_t sumi = 0;
+
+        for (size_t j = 0; j < sizeof(x[0].qs); j += 32) {
+            const int8_t * py0 = &y[i].qs[j * 4 + 0 * 32];
+            const int8_t * py1 = &y[i].qs[j * 4 + 1 * 32];
+            const int8_t * py2 = &y[i].qs[j * 4 + 2 * 32];
+            const int8_t * py3 = &y[i].qs[j * 4 + 3 * 32];
+            const uint8_t* px  = &x[i].qs[j];
+
+            size_t vlmax_16m2 = __riscv_vsetvl_e16m2(32);
+            vint16m2_t vacc16 = __riscv_vmv_v_x_i16m2(0, vlmax_16m2);
+
+            size_t vl = __riscv_vsetvl_e8m1(32);
+
+            vuint8m1_t vx_u8 = __riscv_vle8_v_u8m1(px, vl);
+
+            vint8m1_t vy0 = __riscv_vle8_v_i8m1(py0 , vl);
+            vint8m1_t vy1 = __riscv_vle8_v_i8m1(py1, vl);
+            vint8m1_t vy2 = __riscv_vle8_v_i8m1(py2, vl);
+            vint8m1_t vy3 = __riscv_vle8_v_i8m1(py3, vl);
+
+            // l=0 (bits 1:0)
+            vuint8m1_t t0 = __riscv_vand_vx_u8m1(vx_u8, 0x03, vl);
+            vint8m1_t vq0 = __riscv_vsub_vx_i8m1(__riscv_vreinterpret_v_u8m1_i8m1(t0), 1, vl);
+
+            // l=1 (bits 3:2)
+            vuint8m1_t t1 = __riscv_vand_vx_u8m1(__riscv_vsrl_vx_u8m1(vx_u8, 2, vl), 0x03, vl);
+            vint8m1_t vq1 = __riscv_vsub_vx_i8m1(__riscv_vreinterpret_v_u8m1_i8m1(t1), 1, vl);
+
+            // l=2 (bits 5:4)
+            vuint8m1_t t2 = __riscv_vand_vx_u8m1(__riscv_vsrl_vx_u8m1(vx_u8, 4, vl), 0x03, vl);
+            vint8m1_t vq2 = __riscv_vsub_vx_i8m1(__riscv_vreinterpret_v_u8m1_i8m1(t2), 1, vl);
+
+            // l=3 (bits 7:6)
+            vuint8m1_t t3 = __riscv_vsrl_vx_u8m1(vx_u8, 6, vl); // No final AND needed as vsrl shifts in zeros
+            vint8m1_t vq3 = __riscv_vsub_vx_i8m1(__riscv_vreinterpret_v_u8m1_i8m1(t3), 1, vl);
+
+            // 4. Multiply and accumulate
+            vacc16 = __riscv_vwmacc_vv_i16m2(vacc16, vq0, vy0, vl);
+            vacc16 = __riscv_vwmacc_vv_i16m2(vacc16, vq1, vy1, vl);
+            vacc16 = __riscv_vwmacc_vv_i16m2(vacc16, vq2, vy2, vl);
+            vacc16 = __riscv_vwmacc_vv_i16m2(vacc16, vq3, vy3, vl);
+
+            vlmax_16m2 = __riscv_vsetvl_e16m2(32);
+            vint32m1_t vzero32 = __riscv_vmv_v_x_i32m1(0, 1);
+            vint32m1_t vred32 = __riscv_vwredsum_vs_i16m2_i32m1(vacc16, vzero32, vlmax_16m2);
+
+            sumi += __riscv_vmv_x_s_i32m1_i32(vred32);
+        }
+        const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+        sumf += (float)sumi * d;
+    }
+
+    *s = sumf;
+}
+
+void ggml_vec_dot_tq2_0_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
+#if defined __riscv_v_intrinsic
+    switch (__riscv_vlenb() * 8) {
+        case 256:
+            ggml_vec_dot_tq2_0_q8_K_vl256(n, s, bs, vx, bx, vy, by, nrc);
+            break;
+        default:
+            ggml_vec_dot_tq2_0_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
+            break;
+    }
+#else
+    ggml_vec_dot_tq2_0_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
+#endif
+}
+
+static void ggml_vec_dot_iq1_s_q8_K_vl256(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
+    assert(n % QK_K == 0);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_iq1_s * GGML_RESTRICT x = vx;
+    const block_q8_K  * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK_K;
+
+    float sumf = 0;
+    for (int i = 0; i < nb; ++i) {
+        // Load qh once for the entire superblock.
+        vuint16mf2_t qh = __riscv_vle16_v_u16mf2(x[i].qh, 8);
+
+        // Calculate ls.
+        vuint16mf2_t temp = __riscv_vsrl_vx_u16mf2(qh, 12, 8);
+        temp = __riscv_vand_vx_u16mf2(temp, 7, 8);
+        vint32m1_t ls = __riscv_vreinterpret_v_u32m1_i32m1(__riscv_vwmulu_vx_u32m1(temp, 2, 8));
+        ls = __riscv_vadd_vx_i32m1(ls, 1, 8);
+
+        // Calculate delta.
+        vbool32_t mask = __riscv_vmseq_vx_u16mf2_b32(__riscv_vand_vx_u16mf2(qh, 0x8000, 8), 0, 8);
+        vint32m1_t delta_neg = __riscv_vmv_v_x_i32m1(-1, 8);
+        vint32m1_t delta_pos = __riscv_vmv_v_x_i32m1(1, 8);
+        vint32m1_t delta = __riscv_vmerge_vvm_i32m1(delta_neg, delta_pos, mask, 8);
+
+        // Load qs.
+        vuint8m1_t qs = __riscv_vle8_v_u8m1(x[i].qs, 32);
+
+        // Prepare the indices.
+        const uint64_t shift = 0x0009000600030000;
+        vuint16m2_t qh_shift = __riscv_vreinterpret_v_u64m2_u16m2(__riscv_vmv_v_x_u64m2(shift, 8));
+        vuint16m2_t qh_gather_index = __riscv_vreinterpret_v_i16m2_u16m2(
+            __riscv_vdiv_vx_i16m2(__riscv_vreinterpret_v_u16m2_i16m2(__riscv_vid_v_u16m2(32)), 4, 32));
+        vuint16m2_t qh_ext = __riscv_vlmul_ext_v_u16m1_u16m2(__riscv_vlmul_ext_v_u16mf2_u16m1(qh));
+        vuint16m2_t qh_index = __riscv_vrgather_vv_u16m2(qh_ext, qh_gather_index, 32);
+        qh_index = __riscv_vsrl_vv_u16m2(qh_index, qh_shift, 32);
+        qh_index = __riscv_vand_vx_u16m2(qh_index, 7, 32);
+        qh_index = __riscv_vsll_vx_u16m2(qh_index, 8, 32);
+        qh_index = __riscv_vor_vv_u16m2(qh_index, __riscv_vzext_vf2_u16m2(qs, 32), 32);
+        vuint16m2_t index = __riscv_vsll_vx_u16m2(qh_index, 3, 32);
+
+        // Final lsums.
+        int32_t lsums_s[8];
+        vint32m1_t one_scalar = __riscv_vmv_v_x_i32m1(0, 1);
+
+        // Sub-blocks 1-4
+        {
+            vuint16m1_t grid_index0 = __riscv_vget_v_u16m2_u16m1(index, 0);
+            vint8m4_t grid0 = __riscv_vreinterpret_v_i64m4_i8m4(__riscv_vluxei16_v_i64m4((const int64_t*)iq1s_grid, grid_index0, 16));
+            vint8m4_t q80 = __riscv_vle8_v_i8m4(y[i].qs, 128);
+            vint16m8_t lsum0 = __riscv_vwmul_vv_i16m8(grid0, q80, 128);
+            lsums_s[0] = __riscv_vmv_x_s_i32m1_i32(__riscv_vwredsum_vs_i16m2_i32m1(__riscv_vget_v_i16m8_i16m2(lsum0, 0), one_scalar, 32));
+            lsums_s[1] = __riscv_vmv_x_s_i32m1_i32(__riscv_vwredsum_vs_i16m2_i32m1(__riscv_vget_v_i16m8_i16m2(lsum0, 1), one_scalar, 32));
+            lsums_s[2] = __riscv_vmv_x_s_i32m1_i32(__riscv_vwredsum_vs_i16m2_i32m1(__riscv_vget_v_i16m8_i16m2(lsum0, 2), one_scalar, 32));
+            lsums_s[3] = __riscv_vmv_x_s_i32m1_i32(__riscv_vwredsum_vs_i16m2_i32m1(__riscv_vget_v_i16m8_i16m2(lsum0, 3), one_scalar, 32));
+        }
+        __asm__ __volatile__("" ::: "memory");
+        // Sub-blocks 5-8
+        {
+            vuint16m1_t grid_index1 = __riscv_vget_v_u16m2_u16m1(index, 1);
+            vint8m4_t grid1 = __riscv_vreinterpret_v_i64m4_i8m4(__riscv_vluxei16_v_i64m4((const int64_t*)iq1s_grid, grid_index1, 16));
+            vint8m4_t q81 = __riscv_vle8_v_i8m4(&y[i].qs[128], 128);
+            vint16m8_t lsum1 = __riscv_vwmul_vv_i16m8(grid1, q81, 128);
+            lsums_s[4] = __riscv_vmv_x_s_i32m1_i32(__riscv_vwredsum_vs_i16m2_i32m1(__riscv_vget_v_i16m8_i16m2(lsum1, 0), one_scalar, 32));
+            lsums_s[5] = __riscv_vmv_x_s_i32m1_i32(__riscv_vwredsum_vs_i16m2_i32m1(__riscv_vget_v_i16m8_i16m2(lsum1, 1), one_scalar, 32));
+            lsums_s[6] = __riscv_vmv_x_s_i32m1_i32(__riscv_vwredsum_vs_i16m2_i32m1(__riscv_vget_v_i16m8_i16m2(lsum1, 2), one_scalar, 32));
+            lsums_s[7] = __riscv_vmv_x_s_i32m1_i32(__riscv_vwredsum_vs_i16m2_i32m1(__riscv_vget_v_i16m8_i16m2(lsum1, 3), one_scalar, 32));
+        }
+        __asm__ __volatile__("" ::: "memory");
+        vint32m1_t lsums = __riscv_vle32_v_i32m1(&lsums_s[0], 8);
+
+        // Calculate the bsums.
+        vint16m1_t bsums_0 = __riscv_vle16_v_i16m1(y[i].bsums, 16);
+        const vuint32m1_t bsums_i32 = __riscv_vreinterpret_v_u16m1_u32m1(__riscv_vreinterpret_v_i16m1_u16m1(bsums_0));
+        const vint16mf2_t bsums_i32_0 = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vnsrl_wx_u16mf2(bsums_i32, 0, 8));
+        const vint16mf2_t bsums_i32_1 = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vnsrl_wx_u16mf2(bsums_i32, 16, 8));
+        const vint32m1_t bsums = __riscv_vwadd_vv_i32m1(bsums_i32_0, bsums_i32_1, 8);
+
+        // Accumulation.
+        vint32m1_t sumi_v = __riscv_vmul_vv_i32m1(ls, lsums, 8);
+        vint32m1_t sumi1_v = __riscv_vmul_vv_i32m1(__riscv_vmul_vv_i32m1(ls, delta, 8), bsums, 8);
+
+        // Update sumf.
+        int sumi = __riscv_vmv_x_s_i32m1_i32(__riscv_vredsum_vs_i32m1_i32m1(sumi_v, __riscv_vmv_v_x_i32m1(0.0f, 1), 8));
+        int sumi1 = __riscv_vmv_x_s_i32m1_i32(__riscv_vredsum_vs_i32m1_i32m1(sumi1_v, __riscv_vmv_v_x_i32m1(0.0f, 1), 8));
+        sumf += GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d * (sumi + IQ1S_DELTA * sumi1);
+    }
+
+    *s = sumf;
+}
+
+void ggml_vec_dot_iq1_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
+#if defined __riscv_v_intrinsic
+    switch (__riscv_vlenb() * 8) {
+        case 256:
+            ggml_vec_dot_iq1_s_q8_K_vl256(n, s, bs, vx, bx, vy, by, nrc);
+            break;
+        default:
+            ggml_vec_dot_iq1_s_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
+            break;
+    }
+#else
+    ggml_vec_dot_iq1_s_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
+#endif
+}
+
+static void ggml_vec_dot_iq1_m_q8_K_vl256(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
+    assert(n % QK_K == 0);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_iq1_m * GGML_RESTRICT x = vx;
+    const block_q8_K  * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK_K;
+
+    iq1m_scale_t scale;
+    float sumf = 0.0f;
+    for (int i = 0; i < nb; ++i) {
+        const int8_t   * q8 = y[i].qs;
+        const uint8_t  * qs = x[i].qs;
+        const uint8_t  * qh = x[i].qh;
+        const uint16_t * sc = (const uint16_t *)x[i].scales;
+
+        scale.u16 = (sc[0] >> 12) | ((sc[1] >> 8) & 0x00f0) | ((sc[2] >> 4) & 0x0f00) | (sc[3] & 0xf000);
+
+        // Accumulators.
+        vint32m2_t acc1 = __riscv_vmv_v_x_i32m2(0, 16);
+        vint32m2_t acc2 = __riscv_vmv_v_x_i32m2(0, 16);
+
+        // We process 4 sub-blocks together.
+        for (int ib = 0; ib < QK_K/128; ib++) {
+            // Load qh for 4 sub-blocks.
+            const vuint8mf4_t qh_8 = __riscv_vle8_v_u8mf4(qh, 8);
+            const vuint16mf2_t qh_16_lo = __riscv_vzext_vf2_u16mf2(qh_8, 8);
+            const vuint16mf2_t qh_16_hi = __riscv_vsll_vx_u16mf2(qh_16_lo, 8, 8);
+            const vuint16m1_t qhb = __riscv_vzext_vf2_u16m1(
+                __riscv_vreinterpret_v_u16mf2_u8mf2(__riscv_vor_vv_u16mf2(qh_16_lo, qh_16_hi, 8)), 16);
+            qh += 8;
+
+            // Prepare grid indices.
+            const vuint16m1_t qsb = __riscv_vzext_vf2_u16m1(__riscv_vle8_v_u8mf2(&qs[0], 16), 16);
+            const vuint16m1_t shift = __riscv_vreinterpret_v_u32m1_u16m1(__riscv_vmv_v_x_u32m1(0x00040008, 8));
+            vuint16m1_t index = __riscv_vor_vv_u16m1(qsb, __riscv_vand_vx_u16m1(__riscv_vsll_vv_u16m1(qhb, shift, 16), 0x700, 16), 16);
+            index = __riscv_vsll_vx_u16m1(index, 3, 16);
+            qs += 16;
+
+            // Load the grid.
+            const vint8m4_t iq1b = __riscv_vreinterpret_v_i64m4_i8m4(__riscv_vreinterpret_v_u64m4_i64m4(
+                __riscv_vluxei16_v_u64m4(iq1s_grid, index, 16)));
+
+            // Prepare the deltas.
+            const vbool16_t mask = __riscv_vmsgtu_vx_u16m1_b16(
+                __riscv_vand_vv_u16m1(qhb, __riscv_vreinterpret_v_u32m1_u16m1(__riscv_vmv_v_x_u32m1(0x00800008, 8)), 16), 0, 16);
+            const vint64m4_t delta_pos = __riscv_vmv_v_x_i64m4(0x0101010101010101, 16);
+            const vint64m4_t delta_neg = __riscv_vmv_v_x_i64m4(0xffffffffffffffff, 16);
+            const vint8m4_t delta = __riscv_vreinterpret_v_i64m4_i8m4(
+                __riscv_vmerge_vvm_i64m4(delta_pos, delta_neg, mask, 16));
+
+            // Load q8 for sub-blocks.
+            const vint8m4_t q8b = __riscv_vle8_v_i8m4(q8, 128);
+            q8 += 128;
+
+            // Calculate the lsums.
+            const vint16m8_t lsum1 = __riscv_vwmul_vv_i16m8(iq1b, q8b, 128);
+            const vint16m8_t lsum2 = __riscv_vwmul_vv_i16m8(delta, q8b, 128);
+
+            // Prepare the scales.
+            const int16_t ls_0_0 = 2*((sc[0] >> 0) & 0x7) + 1;
+            const int16_t ls_0_1 = 2*((sc[0] >> 3) & 0x7) + 1;
+            const int16_t ls_1_0 = 2*((sc[0] >> 6) & 0x7) + 1;
+            const int16_t ls_1_1 = 2*((sc[0] >> 9) & 0x7) + 1;
+            const int16_t ls_2_0 = 2*((sc[1] >> 0) & 0x7) + 1;
+            const int16_t ls_2_1 = 2*((sc[1] >> 3) & 0x7) + 1;
+            const int16_t ls_3_0 = 2*((sc[1] >> 6) & 0x7) + 1;
+            const int16_t ls_3_1 = 2*((sc[1] >> 9) & 0x7) + 1;
+            sc += 2;
+
+            // Accumulate in acc0 and acc1 for each sub-block.
+            acc1 = __riscv_vwmacc_vx_i32m2(acc1, ls_0_0, __riscv_vget_v_i16m8_i16m1(lsum1, 0), 16);
+            acc1 = __riscv_vwmacc_vx_i32m2(acc1, ls_0_1, __riscv_vget_v_i16m8_i16m1(lsum1, 1), 16);
+            acc2 = __riscv_vwmacc_vx_i32m2(acc2, ls_0_0, __riscv_vget_v_i16m8_i16m1(lsum2, 0), 16);
+            acc2 = __riscv_vwmacc_vx_i32m2(acc2, ls_0_1, __riscv_vget_v_i16m8_i16m1(lsum2, 1), 16);
+            //
+            acc1 = __riscv_vwmacc_vx_i32m2(acc1, ls_1_0, __riscv_vget_v_i16m8_i16m1(lsum1, 2), 16);
+            acc1 = __riscv_vwmacc_vx_i32m2(acc1, ls_1_1, __riscv_vget_v_i16m8_i16m1(lsum1, 3), 16);
+            acc2 = __riscv_vwmacc_vx_i32m2(acc2, ls_1_0, __riscv_vget_v_i16m8_i16m1(lsum2, 2), 16);
+            acc2 = __riscv_vwmacc_vx_i32m2(acc2, ls_1_1, __riscv_vget_v_i16m8_i16m1(lsum2, 3), 16);
+            //
+            acc1 = __riscv_vwmacc_vx_i32m2(acc1, ls_2_0, __riscv_vget_v_i16m8_i16m1(lsum1, 4), 16);
+            acc1 = __riscv_vwmacc_vx_i32m2(acc1, ls_2_1, __riscv_vget_v_i16m8_i16m1(lsum1, 5), 16);
+            acc2 = __riscv_vwmacc_vx_i32m2(acc2, ls_2_0, __riscv_vget_v_i16m8_i16m1(lsum2, 4), 16);
+            acc2 = __riscv_vwmacc_vx_i32m2(acc2, ls_2_1, __riscv_vget_v_i16m8_i16m1(lsum2, 5), 16);
+            //
+            acc1 = __riscv_vwmacc_vx_i32m2(acc1, ls_3_0, __riscv_vget_v_i16m8_i16m1(lsum1, 6), 16);
+            acc1 = __riscv_vwmacc_vx_i32m2(acc1, ls_3_1, __riscv_vget_v_i16m8_i16m1(lsum1, 7), 16);
+            acc2 = __riscv_vwmacc_vx_i32m2(acc2, ls_3_0, __riscv_vget_v_i16m8_i16m1(lsum2, 6), 16);
+            acc2 = __riscv_vwmacc_vx_i32m2(acc2, ls_3_1, __riscv_vget_v_i16m8_i16m1(lsum2, 7), 16);
+        }
+
+        // Reduce and accumulate in `sumf`.
+        vint32m1_t one = __riscv_vmv_v_x_i32m1(0, 1);
+        int sumi1 = __riscv_vmv_x_s_i32m1_i32(__riscv_vredsum_vs_i32m2_i32m1(acc1, one, 16));
+        int sumi2 = __riscv_vmv_x_s_i32m1_i32(__riscv_vredsum_vs_i32m2_i32m1(acc2, one, 16));
+        sumf += y[i].d * GGML_CPU_FP16_TO_FP32(scale.f16) * (sumi1 + IQ1M_DELTA * sumi2);
+    }
+
+    *s = sumf;
+}
+
+void ggml_vec_dot_iq1_m_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
+#if defined __riscv_v_intrinsic
+    switch (__riscv_vlenb() * 8) {
+        case 256:
+            ggml_vec_dot_iq1_m_q8_K_vl256(n, s, bs, vx, bx, vy, by, nrc);
+            break;
+        default:
+            ggml_vec_dot_iq1_m_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
+            break;
+    }
+#else
+    ggml_vec_dot_iq1_m_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
+#endif
+}

ggml/src/ggml-cpu/llamafile/sgemm-ppc.h

@@ -1,333 +0,0 @@
-#pragma once
-
-typedef vector unsigned char vec_t;
-typedef __vector_quad acc_t;
-
-template <typename TA>
-class tinyBLAS_Q0_PPC {
-  public:
-    tinyBLAS_Q0_PPC(int64_t k,
-                    const TA *A, int64_t lda,
-                    const block_q8_0 *B, int64_t ldb,
-                    float *C, int64_t ldc,
-                    int ith, int nth);
-
-    void matmul(int64_t m, int64_t n);
-    void matmul_tiled_q0(int64_t m, int64_t n, int64_t mc, int64_t nc, int64_t kc) {
-        vec_t A_pack[mc*kc*2];
-        vec_t B_pack[nc*kc*2];
-        int comparray[mc*kc];
-        constexpr bool is_Ablock_q4 = std::is_same_v<TA, block_q4_0>;
-        int64_t ytiles = m / mc;
-        int64_t xtiles = n / nc;
-        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 = (job / xtiles) * mc;
-            int64_t jj = (job % xtiles) * nc;
-            for (int64_t kk = 0; kk < k; kk += kc) {
-                if constexpr(is_Ablock_q4) {
-                    packNormalInt4_large(A + ii*lda + kk, lda, mc, 4, (int8_t*)A_pack, comparray);
-                } else {
-                    packNormal_large<int8_t, vector signed char>(A + ii*lda + kk, lda, mc, 8, (int8_t*)A_pack, false, comparray);
-                }
-                packNormal_large<uint8_t, vector unsigned char>(B + jj*ldb + kk, ldb, nc, 8, (uint8_t*)B_pack, true);
-                KERNEL_Q0(ii, jj, mc, nc, kc, kk, A_pack, B_pack, comparray);
-            }
-        }
-    }
-
-  private:
-    inline void save_res(int ii, int jj, int idx, vector float* fin_res, int RM=4, int RN=4) {
-        for (int I = 0; I < RM; I++) {
-            for (int J = 0; J < RN; J++) {
-                *((float*)(C+ii+((jj+J)*ldc)+I)) = *((float*)&fin_res[idx+I]+J);
-            }
-        }
-    }
-
-    inline void add_save_res(int ii, int jj, int idx, vector float* fin_res, int RM=4, int RN=4) {
-        for (int I = 0; I < RM; I++) {
-            for (int J = 0; J < RN; J++) {
-                float * c_ptr = (float *)(C+ii+((jj+J)*ldc)+I);
-                *c_ptr += *((float*)&fin_res[idx+I]+J);
-            }
-        }
-    }
-
-    template<typename ArrayType>
-    inline void compute(acc_t* ACC, int c_idx, int s_idx, ArrayType& comparray, vector float* vs, vector float* fin_res) {
-        vector signed int vec_C[4];
-        vector float CA[4] = {0};
-        vector float res[4] = {0};
-        __builtin_mma_disassemble_acc(vec_C, ACC);
-        for (int i = 0; i < 4; i++) {
-            CA[i] = vec_splats((float)(((double)comparray[c_idx+i]) * -128.0));
-            res[i] = vec_add(vec_ctf(vec_C[i], 0), CA[i]);
-            fin_res[s_idx+i] = vec_madd(res[i], vs[s_idx+i], fin_res[s_idx+i]);
-        }
-    }
-
-    inline void process_q4_elements(vector signed char (&c)[2], int* ca) {
-        const vector signed char lowMask = vec_splats((signed char)0xF);
-        const vector unsigned char v4 = vec_splats((unsigned char)0x4);
-        const vector signed char v8 = vec_splats((signed char)0x8);
-        vector signed int vsum = {0};
-        vector signed int vsum2 = {0};
-        c[0] = vec_and(c[1], lowMask);
-        c[1] = vec_sr(c[1], v4);
-        c[0] = vec_sub(c[0], v8);
-        c[1] = vec_sub(c[1], v8);
-        vsum = vec_sum4s(c[0], vsum);
-        vsum2 = vec_sum4s(c[1], vsum2);
-        vsum = vec_add(vsum, vsum2);
-        *(ca) = vsum[0] + vsum[1] + vsum[2] + vsum[3];
-    }
-
-    template <typename V1, typename V2>
-    inline void vector_permute_store(V2 &s1, V2 &s2, V2 &s3, V2 &s4, V1 *vecOffset, bool flip) {
-        vector unsigned char swiz1 = {0, 1, 2, 3, 4, 5, 6, 7, 16, 17, 18, 19, 20, 21, 22, 23};
-        vector unsigned char swiz2 = {8, 9, 10, 11, 12, 13, 14, 15, 24, 25, 26, 27, 28, 29, 30, 31};
-        vector unsigned char swiz3 = {0, 1, 2, 3, 8, 9, 10, 11, 16, 17, 18, 19, 24, 25, 26, 27};
-        vector unsigned char swiz4 = {4, 5, 6, 7, 12, 13, 14, 15, 20, 21, 22, 23, 28, 29, 30, 31};
-        V2 t1, t2, t3, t4, t5, t6, t7, t8;
-        vector unsigned char xor_vector;
-        uint8_t flip_vec = 0x80;
-        xor_vector = vec_splats(flip_vec);
-        t1 = vec_perm(s1, s2, swiz1);
-        t2 = vec_perm(s1, s2, swiz2);
-        t3 = vec_perm(s3, s4, swiz1);
-        t4 = vec_perm(s3, s4, swiz2);
-        t5 = vec_perm(t1, t3, swiz3);
-        t6 = vec_perm(t1, t3, swiz4);
-        t7 = vec_perm(t2, t4, swiz3);
-        t8 = vec_perm(t2, t4, swiz4);
-        if (flip == true) {
-            t5 = vec_xor(t5, xor_vector);
-            t6 = vec_xor(t6, xor_vector);
-            t7 = vec_xor(t7, xor_vector);
-            t8 = vec_xor(t8, xor_vector);
-        }
-        vec_xst(t5, 0, vecOffset);
-        vec_xst(t6, 0, vecOffset+16);
-        vec_xst(t7, 0, vecOffset+32);
-        vec_xst(t8, 0, vecOffset+48);
-    }
-
-    template<int RM, int RN>
-    inline void kernel(int64_t ii, int64_t jj) {
-        if constexpr(RM == 4 && RN == 8) {
-            KERNEL_4x8(ii,jj);
-        } else if constexpr(RM == 8 && RN == 4) {
-            KERNEL_8x4(ii,jj);
-        } else if constexpr(RM == 8 && RN == 8) {
-            KERNEL_8x8(ii,jj);
-        } else {
-            assert(false && "RN/RM values not supported");
-        }
-    }
-    template<int size>
-    void packNormalInt4(const TA* a, int64_t lda, int rows, int cols, int8_t* vec, std::array<int, size>& comparray);
-    template<typename VA, typename VB>
-    void packNormal(const block_q8_0* a, int64_t lda, int rows, int cols, VA* vec, bool flip);
-    void mnpack(int64_t m0, int64_t m, int64_t n0, int64_t n);
-    void KERNEL_4x8(int64_t ii, int64_t jj);
-    void KERNEL_8x4(int64_t ii, int64_t jj);
-    void KERNEL_8x8(int64_t ii, int64_t jj);
-    void gemm_small(int64_t m0, int64_t m, int64_t n0, int64_t n, int RM, int RN);
-    template <int RM, int RN>
-    void gemm(int64_t m0, int64_t m, int64_t n0, int64_t n);
-
-    void compute_scale(int64_t ii, int64_t jj, int blk, vector float* vs){
-        for (int I = 0; I<8; I++) {
-            float a_scale = unhalf((A+((ii+I)*lda)+blk)->d);
-            for (int J = 0; J<4; J++) {
-                *((float*)&vs[I]+J) = (a_scale * unhalf((B+((jj+J)*ldb)+blk)->d));
-                *((float*)&vs[I+8]+J) = (a_scale * unhalf((B+((jj+J+4)*ldb)+blk)->d));
-             }
-         }
-    }
-
-    inline void process_q8_elements(const int8_t *qs, int *ca) {
-        vector signed char c1 = vec_xl(0, qs);
-        vector signed char c2 = vec_xl(16, qs);
-        vector signed int vsum1 = {0};
-        vector signed int vsum2 = {0};
-        vsum1 = vec_sum4s(c1, vsum1);
-        vsum2 = vec_sum4s(c2, vsum2);
-        vector signed int vsum = vec_add(vsum1, vsum2);
-        *ca = vsum[0] + vsum[1] + vsum[2] + vsum[3];
-    }
-
-    template<typename VA, typename VB>
-    void packNormal_large(const block_q8_0* a, int64_t lda, int rows, int cols, VA* vec, bool flip, int* comparray=nullptr) {
-        int64_t i, j;
-        block_q8_0 *aoffset = NULL;
-        VA *vecOffset = NULL;
-        block_q8_0* aoffsets[8];
-        __vector_pair arr[8];
-        VB c[8][2] = {0};
-        VB c1[8] = {0}; VB c2[8] = {0};
-        aoffset = const_cast<block_q8_0*>(a);
-        vecOffset = vec;
-        j = (rows >> 3);
-        int index = 0;
-        if (j > 0) {
-            do {
-                for (int it = 0; it < 8; it++)
-                    aoffsets[it] = aoffset + it*lda;
-                aoffset += 8 * lda;
-                for (int blk = 0; blk < kc; blk++) {
-                    for (int it = 0; it < 8; it++) {
-                        arr[it] = __builtin_vsx_lxvp(0, (__vector_pair*)(aoffsets[it]+blk)->qs);
-                        __builtin_vsx_disassemble_pair(c[it], &arr[it]);
-                        c1[it] = c[it][0];
-                        c2[it] = c[it][1];
-                        if (comparray){
-                            process_q8_elements((aoffsets[it]+ blk)->qs, &comparray[index + 8*blk + it]);
-                        }
-                    }
-                    vector_permute_store<VA, VB>(c1[0], c1[1], c1[2], c1[3], vecOffset, flip);
-                    vector_permute_store<VA, VB>(c2[0], c2[1], c2[2], c2[3], vecOffset+64, flip);
-                    vector_permute_store<VA, VB>(c1[4], c1[5], c1[6], c1[7], vecOffset+128, flip);
-                    vector_permute_store<VA, VB>(c2[4], c2[5], c2[6], c2[7], vecOffset+192, flip);
-                    vecOffset += 256;
-                }
-                j--;
-                index += 8*kc;
-            } while(j > 0);
-        }
-
-    }
-
-    void packNormalInt4_large(const TA* a, int64_t lda, int rows, int cols, int8_t* vec, int*comparray) {
-        int64_t i, j;
-        TA *aoffset = NULL;
-        int8_t *vecOffset = NULL;
-        TA *aoffset1 = NULL, *aoffset2 = NULL, *aoffset3 = NULL, *aoffset4 = NULL;
-        TA *aoffset5 = NULL, *aoffset6 = NULL, *aoffset7 = NULL, *aoffset8 = NULL;
-        vector signed char c1[2] = {0}, c2[2] = {0}, c3[2] = {0}, c4[2] = {0};
-        vector signed char c5[2] = {0}, c6[2] = {0}, c7[2] = {0}, c8[2] = {0};
-        aoffset = const_cast<TA*>(a);
-        vecOffset = vec;
-        int index = 0;
-        j = (rows >> 3);
-        if (j > 0) {
-            do {
-                aoffset1 = aoffset;
-                aoffset2 = aoffset1 + lda;
-                aoffset3 = aoffset2 + lda;
-                aoffset4 = aoffset3 + lda;
-                aoffset5 = aoffset4 + lda;
-                aoffset6 = aoffset5 + lda;
-                aoffset7 = aoffset6 + lda;
-                aoffset8 = aoffset7 + lda;
-                aoffset += 8 * lda;
-                for (int blk = 0; blk < kc; blk++) {
-                    c1[1] = reinterpret_cast<vector signed char>(vec_xl(0, (aoffset1+blk)->qs));
-                    c2[1] = reinterpret_cast<vector signed char>(vec_xl(0, (aoffset2+blk)->qs));
-                    c3[1] = reinterpret_cast<vector signed char>(vec_xl(0, (aoffset3+blk)->qs));
-                    c4[1] = reinterpret_cast<vector signed char>(vec_xl(0, (aoffset4+blk)->qs));
-                    c5[1] = reinterpret_cast<vector signed char>(vec_xl(0, (aoffset5+blk)->qs));
-                    c6[1] = reinterpret_cast<vector signed char>(vec_xl(0, (aoffset6+blk)->qs));
-                    c7[1] = reinterpret_cast<vector signed char>(vec_xl(0, (aoffset7+blk)->qs));
-                    c8[1] = reinterpret_cast<vector signed char>(vec_xl(0, (aoffset8+blk)->qs));
-
-                    process_q4_elements(c1, &comparray[index + 8*blk+0]);
-                    process_q4_elements(c2, &comparray[index + 8*blk+1]);
-                    process_q4_elements(c3, &comparray[index + 8*blk+2]);
-                    process_q4_elements(c4, &comparray[index + 8*blk+3]);
-                    process_q4_elements(c5, &comparray[index + 8*blk+4]);
-                    process_q4_elements(c6, &comparray[index + 8*blk+5]);
-                    process_q4_elements(c7, &comparray[index + 8*blk+6]);
-                    process_q4_elements(c8, &comparray[index + 8*blk+7]);
-                    vector_permute_store<int8_t, vector signed char>(c1[0], c2[0], c3[0], c4[0], vecOffset, false);
-                    vector_permute_store<int8_t, vector signed char>(c1[1], c2[1], c3[1], c4[1], vecOffset+64, false);
-                    vector_permute_store<int8_t, vector signed char>(c5[0], c6[0], c7[0], c8[0], vecOffset+128, false);
-                    vector_permute_store<int8_t, vector signed char>(c5[1], c6[1], c7[1], c8[1], vecOffset+192, false);
-                    vecOffset += 256;
-                }
-                j--;
-                index += 8*kc;
-            } while (j > 0);
-        }
-    }
-
-    void KERNEL_Q0(int64_t ii, int64_t jj, int64_t mc, int64_t nc, int64_t kc, int64_t l, vec_t *vec_A, vec_t *vec_B, int *comparray) {
-        acc_t acc[8];
-        for (int i = 0; i < mc ; i += 8) {
-            for (int j = 0; j < nc; j += 8) {
-                vector float fin_res[16] = {0};
-                vector float vs[16] = {0};
-                for (int64_t kk = 0; kk < kc; kk+=2) {
-                    for (int x = 0; x < 8; x++) {
-                        __builtin_mma_xxsetaccz(&acc[x]);
-                    }
-                    int A_block_idx = (i/8)*(16*kc) + kk*16;
-                    int B_block_idx = (j/8)*(16*kc)+ kk*16;
-                    vec_t *A_block = &vec_A[A_block_idx];
-                    vec_t *B_block = &vec_B[B_block_idx];
-                    for (int x = 0; x < 8; x++) {
-                        __builtin_mma_xvi8ger4pp(&acc[0], A_block[x],     B_block[x]);
-                        __builtin_mma_xvi8ger4pp(&acc[1], A_block[x + 8], B_block[x]);
-                        __builtin_mma_xvi8ger4pp(&acc[2], A_block[x],     B_block[x+8]);
-                        __builtin_mma_xvi8ger4pp(&acc[3], A_block[x+8],   B_block[x+8]);
-                    }
-                    compute_scale(ii+i, jj+j, l+kk, vs);
-                    int c_index = (i/8)*(8*kc)+ kk*8;
-                    int* c_block = &comparray[c_index];
-                    compute(&acc[0], 0,  0,  c_block, vs, fin_res);
-                    compute(&acc[1], 4,  4,  c_block, vs, fin_res);
-                    compute(&acc[2], 0,  8,  c_block, vs, fin_res);
-                    compute(&acc[3], 4, 12,  c_block, vs, fin_res);
-
-                    A_block_idx = (i/8)*(16*kc) + (kk+1)*16;
-                    B_block_idx = (j/8)*(16*kc)+ (kk+1)*16;
-                    A_block = &vec_A[A_block_idx];
-                    B_block = &vec_B[B_block_idx];
-                    for (int x = 0; x < 8; x++) {
-                        __builtin_mma_xvi8ger4pp(&acc[4], A_block[x],     B_block[x]);
-                        __builtin_mma_xvi8ger4pp(&acc[5], A_block[x + 8], B_block[x]);
-                        __builtin_mma_xvi8ger4pp(&acc[6], A_block[x],     B_block[x+8]);
-                        __builtin_mma_xvi8ger4pp(&acc[7], A_block[x+8],   B_block[x+8]);
-                    }
-                    compute_scale(ii+i, jj+j, l+kk+1, vs);
-                    c_index = (i/8)*(8*kc)+ (kk+1)*8;
-                    c_block = &comparray[c_index];
-                    compute(&acc[4], 0,  0,  c_block, vs, fin_res);
-                    compute(&acc[5], 4,  4,  c_block, vs, fin_res);
-                    compute(&acc[6], 0,  8,  c_block, vs, fin_res);
-                    compute(&acc[7], 4, 12,  c_block, vs, fin_res);
-
-                }
-                if (l == 0) {
-                    save_res(ii+i,   jj+j,    0,  fin_res);
-                    save_res(ii+i+4, jj+j,    4,  fin_res);
-                    save_res(ii+i,   jj+j+4,  8,  fin_res);
-                    save_res(ii+i+4, jj+j+4, 12,  fin_res);
-                } else {
-                    add_save_res(ii+i,   jj+j,    0,  fin_res);
-                    add_save_res(ii+i+4, jj+j,    4,  fin_res);
-                    add_save_res(ii+i,   jj+j+4,  8,  fin_res);
-                    add_save_res(ii+i+4, jj+j+4, 12,  fin_res);
-                }
-            }
-        }
-    }
-
-    const TA *const A;
-    const block_q8_0 *const B;
-    float *C;
-    const int64_t k;
-    int64_t kc;
-    const int64_t lda;
-    const int64_t ldb;
-    const int64_t ldc;
-    const int ith;
-    const int nth;
-};

ggml/src/ggml-cpu/llamafile/sgemm.cpp

@@ -121,7 +121,8 @@ inline float32x4_t mul(float32x4_t x, float32x4_t y) { return vec_mul(x, y); }
 #endif
 
 #if defined(__MMA__)
-#include "sgemm-ppc.h"
+typedef vector unsigned char vec_t;
+typedef __vector_quad acc_t;
 #endif
 ////////////////////////////////////////////////////////////////////////////////////////////////////
 // VECTORIZED FUSED MULTIPLY ADD
@@ -2153,7 +2154,7 @@ class tinyBLAS_HP16_PPC {
             packNormal((B+(jj*ldb)+l), ldb, 8, 4, (uint8_t*)vec_B);
             for (int x = 0; x < 4; x++) {
                 mma_instr<TA>::outer_product(&acc_0, vec_A[x], vec_B[x]);
-                mma_instr<TA>::outer_product(&acc_1, vec_A[x], vec_B[x+4]);
+                mma_instr<TA>::outer_product(&acc_1, vec_A[x+4], vec_B[x]);
             }
         }
         SAVE_ACC(&acc_0, ii, jj);
@@ -2301,43 +2302,299 @@ class tinyBLAS_HP16_PPC {
     const int nth;
 };
 
-    template <typename TA>
-    tinyBLAS_Q0_PPC<TA>::tinyBLAS_Q0_PPC(int64_t k,
-        const TA *A, int64_t lda,
-        const block_q8_0 *B, int64_t ldb,
-        float *C, int64_t ldc,
-        int ith, int nth)
+template <typename TA>
+class tinyBLAS_Q0_PPC {
+  public:
+    tinyBLAS_Q0_PPC(int64_t k,
+             const TA * A, int64_t lda,
+             const block_q8_0 * B, int64_t ldb,
+             float * 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) {
-                kc = 64;
     }
 
-    template<typename TA>
-    void tinyBLAS_Q0_PPC<TA>::matmul(int64_t m, int64_t n) {
-        int mc = 64; int nc = 64;
-        if (n % 8 == 0 && n < nc) {
-                nc = n;
-                mc = 32 ;
-                kc = 32;
+    void matmul(int64_t m, int64_t n) {
+        const int64_t mc = 64;
+        const int64_t kc = 64;
+        int64_t nc = 64;
+        int64_t n_aligned = 0;
+        if (n % 64 == 0) {
+            n_aligned = n;
+        } else if (n == 4) {
+            n_aligned = 4;
+        } else if (n < 64) {
+            n_aligned = (n / 8) * 8;
+        } else {
+            n_aligned = (n / 64) * 64;
         }
-        const bool is_aligned = ((m & (mc - 1)) == 0) & ((n & (nc - 1)) == 0) & ((k & (kc - 1)) == 0);
-        if (is_aligned) {
-            this->matmul_tiled_q0(m, n, mc, nc, kc);
+
+        if (n_aligned > 0) {
+            if (n_aligned % 64 == 0)      nc = 64;
+            else if (n_aligned == n)      nc = n;
+            else if (n_aligned % 32 == 0) nc = 32;
+            else if (n_aligned % 24 == 0) nc = 24;
+            else if (n_aligned % 16 == 0) nc = 16;
+            else                          nc = 8;
+        }
+        bool can_use_tiled = n_aligned > 0 && (m % mc == 0) && (k % kc == 0);
+        if (can_use_tiled) {
+            matmul_tiled(m, n_aligned, mc, nc, kc);
+            if (n > n_aligned) {
+                mnpack(0, m, n_aligned, n);
+            }
         } else {
             mnpack(0, m, 0, n);
         }
     }
 
-   template<typename TA>
-   template<int size>
-   void tinyBLAS_Q0_PPC<TA>::packNormalInt4(const TA* a, int64_t lda, int rows, int cols, int8_t* vec, std::array<int, size>& comparray) {
+  private:
+    inline void save_res(int ii, int jj, int idx, vector float * fin_res, int RM = 4, int RN = 4) {
+        for (int I = 0; I < RM; I++) {
+            for (int J = 0; J < RN; J++) {
+                *((float *)(C + ii + ((jj + J) * ldc) + I)) = *((float *)&fin_res[idx + I] + J);
+            }
+        }
+    }
+
+    inline void save_acc(acc_t * ACC, int64_t ii, int64_t jj) {
+        vec_t vec_C[4];
+        __builtin_mma_disassemble_acc(vec_C, ACC);
+        for (int I = 0; I < 4; I++) {
+            for (int J = 0; J < 4; J++) {
+                *((float *)(C + ii + ((jj + J) * ldc) + I)) = *((float *)&vec_C[I] + J);
+            }
+        }
+    }
+
+    inline void add_save_acc(acc_t * ACC, int64_t ii, int64_t jj) {
+        vec_t vec_C[4];
+        __builtin_mma_disassemble_acc(vec_C, ACC);
+        for (int I = 0; I < 4; I++) {
+            for (int J = 0; J < 4; J++) {
+                float * c_ptr = (float *)(C + ii+ ((jj + J) * ldc) + I);
+                *c_ptr += *((float *)&vec_C[I] + J);
+            }
+        }
+    }
+
+    template<typename ArrayType>
+    inline void compute(acc_t * ACC, int c_idx, int s_idx, ArrayType & comparray, vector float * vs, vector float * fin_res) {
+        vector signed int vec_C[4];
+        vector float CA[4] = {0};
+        vector float res[4] = {0};
+        __builtin_mma_disassemble_acc(vec_C, ACC);
+        for (int i = 0; i < 4; i++) {
+            CA[i] = vec_splats((float)(((double)comparray[c_idx + i]) * -128.0));
+            res[i] = vec_add(vec_ctf(vec_C[i], 0), CA[i]);
+            fin_res[s_idx + i] = vec_madd(res[i], vs[s_idx + i], fin_res[s_idx + i]);
+        }
+    }
+
+    inline void process_q4_elements(vector signed char (&c)[2], int * ca) {
+        const vector signed char lowMask = vec_splats((signed char)0xF);
+        const vector unsigned char v4 = vec_splats((unsigned char)0x4);
+        const vector signed char v8 = vec_splats((signed char)0x8);
+        vector signed int vsum = {0};
+        vector signed int vsum2 = {0};
+        c[0] = vec_and(c[1], lowMask);
+        c[1] = vec_sr(c[1], v4);
+        c[0] = vec_sub(c[0], v8);
+        c[1] = vec_sub(c[1], v8);
+        vsum = vec_sum4s(c[0], vsum);
+        vsum2 = vec_sum4s(c[1], vsum2);
+        vsum = vec_add(vsum, vsum2);
+        *(ca) = vsum[0] + vsum[1] + vsum[2] + vsum[3];
+    }
+
+    template <typename V1, typename V2>
+    inline void vector_permute_store(V2 & s1, V2 & s2, V2 & s3, V2 & s4, V1 * vecOffset, bool flip) {
+        vector unsigned char swiz1 = {0, 1, 2, 3, 4, 5, 6, 7, 16, 17, 18, 19, 20, 21, 22, 23};
+        vector unsigned char swiz2 = {8, 9, 10, 11, 12, 13, 14, 15, 24, 25, 26, 27, 28, 29, 30, 31};
+        vector unsigned char swiz3 = {0, 1, 2, 3, 8, 9, 10, 11, 16, 17, 18, 19, 24, 25, 26, 27};
+        vector unsigned char swiz4 = {4, 5, 6, 7, 12, 13, 14, 15, 20, 21, 22, 23, 28, 29, 30, 31};
+        V2 t1, t2, t3, t4, t5, t6, t7, t8;
+        vector unsigned char xor_vector;
+        uint8_t flip_vec = 0x80;
+        xor_vector = vec_splats(flip_vec);
+        t1 = vec_perm(s1, s2, swiz1);
+        t2 = vec_perm(s1, s2, swiz2);
+        t3 = vec_perm(s3, s4, swiz1);
+        t4 = vec_perm(s3, s4, swiz2);
+        t5 = vec_perm(t1, t3, swiz3);
+        t6 = vec_perm(t1, t3, swiz4);
+        t7 = vec_perm(t2, t4, swiz3);
+        t8 = vec_perm(t2, t4, swiz4);
+        if (flip == true) {
+            t5 = vec_xor(t5, xor_vector);
+            t6 = vec_xor(t6, xor_vector);
+            t7 = vec_xor(t7, xor_vector);
+            t8 = vec_xor(t8, xor_vector);
+        }
+        vec_xst(t5, 0, vecOffset);
+        vec_xst(t6, 0, vecOffset + 16);
+        vec_xst(t7, 0, vecOffset + 32);
+        vec_xst(t8, 0, vecOffset + 48);
+    }
+
+    inline void unpack_q4_to_q8(vector signed char packed, vector signed char & lo, vector signed char & hi) {
+        const vector signed char lowMask = vec_splats((signed char)0x0F);
+        const vector signed char v8      = vec_splats((signed char)0x08);
+        const vector unsigned char v4    = vec_splats((unsigned char)4);
+        lo = vec_and(packed, lowMask);
+        hi = vec_sr(packed, v4);
+        lo = vec_sub(lo, v8);
+        hi = vec_sub(hi, v8);
+    }
+
+    inline void vector_permute_store_fp16(vec_t * c, unsigned char * vecOffset) {
+        vec_t t[8], s[8];
+        vec_t swiz1 = {0, 1, 2, 3, 16, 17, 18, 19, 4, 5, 6, 7, 20, 21, 22, 23};
+        vec_t swiz2 = {8, 9, 10, 11, 24, 25, 26, 27, 12, 13, 14, 15, 28, 29, 30, 31};
+        vec_t swiz3 = {0, 1, 2, 3, 4, 5, 6, 7, 16, 17, 18, 19, 20, 21, 22, 23};
+        vec_t swiz4 = {8, 9, 10, 11, 12, 13, 14, 15, 24, 25, 26, 27, 28, 29, 30, 31};
+        for (int i = 0; i < 4; i += 2) {
+            t[i + 0] = vec_perm(c[i + 0], c[i + 1], swiz1);
+            t[i + 1] = vec_perm(c[i + 0], c[i + 1], swiz2);
+        }
+        for (int i = 4; i < 8; i += 2) {
+            t[i + 0] = vec_perm(c[i + 0], c[i + 1], swiz1);
+            t[i + 1] = vec_perm(c[i + 0], c[i + 1], swiz2);
+        }
+        s[0] = vec_perm(t[0], t[2], swiz3);
+        s[1] = vec_perm(t[0], t[2], swiz4);
+        s[2] = vec_perm(t[1], t[3], swiz3);
+        s[3] = vec_perm(t[1], t[3], swiz4);
+        s[4] = vec_perm(t[4], t[6], swiz3);
+        s[5] = vec_perm(t[4], t[6], swiz4);
+        s[6] = vec_perm(t[5], t[7], swiz3);
+        s[7] = vec_perm(t[5], t[7], swiz4);
+        for (int i = 0; i < 8; ++i) {
+            vec_xst(s[i], 0, (vec_t *)(vecOffset + i * 16));
+        }
+    }
+
+    static inline void convert_and_scale_q8(vector signed char raw, vector float v_scale, vector unsigned short & out_hi, vector unsigned short & out_lo) {
+        vector signed short i16_hi = vec_unpackh(raw);
+        vector signed short i16_lo = vec_unpackl(raw);
+
+        vector float f_hi_h = vec_ctf(vec_unpackh(i16_hi), 0);
+        vector float f_hi_l = vec_ctf(vec_unpackl(i16_hi), 0);
+        vector float f_lo_h = vec_ctf(vec_unpackh(i16_lo), 0);
+        vector float f_lo_l = vec_ctf(vec_unpackl(i16_lo), 0);
+        out_hi = vec_pack_to_short_fp32(vec_mul(f_hi_h, v_scale), vec_mul(f_hi_l, v_scale));
+        out_lo = vec_pack_to_short_fp32(vec_mul(f_lo_h, v_scale), vec_mul(f_lo_l, v_scale));
+    }
+
+    void packNormal_q4_fp16(const block_q4_0 * a, int64_t lda, int rows, int blocks, unsigned char * vec) {
+        unsigned char * vecOffset = vec;
+        for (int i = 0; i < rows; i += 8) {
+            const block_q4_0 * rows_base[8];
+            for (int r = 0; r < 8; r++) {
+                rows_base[r] = a + (i + r) * lda;
+            }
+            for (int blk = 0; blk < blocks; blk++) {
+                vector unsigned short hp_res[8][4];
+                for (int r = 0; r < 8; r++) {
+                    const block_q4_0 * current_blk = rows_base[r] + blk;
+                    vector float v_scale = vec_extract_fp32_from_shorth(vec_splats(current_blk->d));
+                    vector signed char v_qs = reinterpret_cast<vector signed char>(vec_xl(0, current_blk->qs));
+                    vector signed char c1, c2;
+                    unpack_q4_to_q8(v_qs, c1, c2);
+                    convert_and_scale_q8(c1, v_scale, hp_res[r][0], hp_res[r][1]);
+                    convert_and_scale_q8(c2, v_scale, hp_res[r][2], hp_res[r][3]);
+                }
+                for (int c = 0; c < 4; c++) {
+                    vector unsigned char c_arr[8];
+                    for (int r = 0; r < 8; r++) {
+                        c_arr[r] = (vector unsigned char)hp_res[r][c];
+                    }
+                    vector_permute_store_fp16((vec_t *)c_arr, vecOffset);
+                    vecOffset += 128;
+                }
+            }
+        }
+    }
+
+    template <int chunk_size>
+    static inline void pack_q8_block(const block_q8_0 * a, int64_t lda, int rows, int blocks, unsigned char * vec) {
+        unsigned char * vecOffset = vec;
+        const vec_t swiz1 = {0, 1, 2, 3, 16, 17, 18, 19, 4, 5, 6, 7, 20, 21, 22, 23};
+        const vec_t swiz2 = {8, 9, 10, 11, 24, 25, 26, 27, 12, 13, 14, 15, 28, 29, 30, 31};
+        const vec_t swiz3 = {0, 1, 2, 3, 4, 5, 6, 7, 16, 17, 18, 19, 20, 21, 22, 23};
+        const vec_t swiz4 = {8, 9, 10, 11, 12, 13, 14, 15, 24, 25, 26, 27, 28, 29, 30, 31};
+
+        for (int i = 0; i < rows; i += chunk_size) {
+            const block_q8_0 * rows_base[chunk_size];
+            for (int r = 0; r < chunk_size; r++) {
+                rows_base[r] = a + (i + r) * lda;
+            }
+            for (int blk = 0; blk < blocks; blk++) {
+                vector unsigned short hp_res[chunk_size][4];
+                for (int r = 0; r < chunk_size; r++) {
+                    const block_q8_0 * b = rows_base[r] + blk;
+                    vector float v_scale = vec_extract_fp32_from_shorth(vec_splats(b->d));
+                    vector signed char c[2];
+                    __vector_pair pair = __builtin_vsx_lxvp(0, (__vector_pair *)b->qs);
+                    __builtin_vsx_disassemble_pair(c, & pair);
+                    convert_and_scale_q8(c[0], v_scale, hp_res[r][0], hp_res[r][1]);
+                    convert_and_scale_q8(c[1], v_scale, hp_res[r][2], hp_res[r][3]);
+                }
+                for (int col = 0; col < 4; col++) {
+                    if constexpr (chunk_size == 8) {
+                        vec_t t[8];
+                        t[0] = vec_perm((vec_t)hp_res[0][col], (vec_t)hp_res[1][col], swiz1);
+                        t[1] = vec_perm((vec_t)hp_res[0][col], (vec_t)hp_res[1][col], swiz2);
+                        t[2] = vec_perm((vec_t)hp_res[2][col], (vec_t)hp_res[3][col], swiz1);
+                        t[3] = vec_perm((vec_t)hp_res[2][col], (vec_t)hp_res[3][col], swiz2);
+                        t[4] = vec_perm((vec_t)hp_res[4][col], (vec_t)hp_res[5][col], swiz1);
+                        t[5] = vec_perm((vec_t)hp_res[4][col], (vec_t)hp_res[5][col], swiz2);
+                        t[6] = vec_perm((vec_t)hp_res[6][col], (vec_t)hp_res[7][col], swiz1);
+                        t[7] = vec_perm((vec_t)hp_res[6][col], (vec_t)hp_res[7][col], swiz2);
+
+                        vec_xst(vec_perm(t[0], t[2], swiz3), 0, (vec_t *)(vecOffset + 0));
+                        vec_xst(vec_perm(t[0], t[2], swiz4), 0, (vec_t *)(vecOffset + 16));
+                        vec_xst(vec_perm(t[1], t[3], swiz3), 0, (vec_t *)(vecOffset + 32));
+                        vec_xst(vec_perm(t[1], t[3], swiz4), 0, (vec_t *)(vecOffset + 48));
+                        vec_xst(vec_perm(t[4], t[6], swiz3), 0, (vec_t *)(vecOffset + 64));
+                        vec_xst(vec_perm(t[4], t[6], swiz4), 0, (vec_t *)(vecOffset + 80));
+                        vec_xst(vec_perm(t[5], t[7], swiz3), 0, (vec_t *)(vecOffset + 96));
+                        vec_xst(vec_perm(t[5], t[7], swiz4), 0, (vec_t *)(vecOffset + 112));
+                        vecOffset += 128;
+                    } else {
+                        vec_t t0 = vec_perm((vec_t)hp_res[0][col], (vec_t)hp_res[1][col], swiz1);
+                        vec_t t1 = vec_perm((vec_t)hp_res[0][col], (vec_t)hp_res[1][col], swiz2);
+                        vec_t t2 = vec_perm((vec_t)hp_res[2][col], (vec_t)hp_res[3][col], swiz1);
+                        vec_t t3 = vec_perm((vec_t)hp_res[2][col], (vec_t)hp_res[3][col], swiz2);
+
+                        vec_xst(vec_perm(t0, t2, swiz3), 0, (vec_t *)(vecOffset + 0));
+                        vec_xst(vec_perm(t0, t2, swiz4), 0, (vec_t *)(vecOffset + 16));
+                        vec_xst(vec_perm(t1, t3, swiz3), 0, (vec_t *)(vecOffset + 32));
+                        vec_xst(vec_perm(t1, t3, swiz4), 0, (vec_t *)(vecOffset + 48));
+                        vecOffset += 64;
+                    }
+                }
+            }
+        }
+    }
+
+    void packNormal_q8_fp16(const block_q8_0 * a, int64_t lda, int rows, int blocks, unsigned char * vec) {
+        if (rows == 4) {
+            pack_q8_block<4>(a, lda, rows, blocks, vec);
+        } else {
+            pack_q8_block<8>(a, lda, rows, blocks, vec);
+        }
+    }
+
+    template<int size>
+    void packNormalInt4(const TA * a, int64_t lda, int rows, int cols, int8_t * vec, std::array<int, size> & comparray) {
         int64_t i, j;
-        TA *aoffset = NULL;
-        int8_t *vecOffset = NULL;
-        TA *aoffset1 = NULL, *aoffset2 = NULL, *aoffset3 = NULL, *aoffset4 = NULL;
-        TA *aoffset5 = NULL, *aoffset6 = NULL, *aoffset7 = NULL, *aoffset8 = NULL;
+        TA * aoffset = NULL;
+        int8_t * vecOffset = NULL;
+        TA * aoffset1 = NULL, * aoffset2 = NULL, * aoffset3 = NULL, * aoffset4 = NULL;
+        TA * aoffset5 = NULL, * aoffset6 = NULL, * aoffset7 = NULL, * aoffset8 = NULL;
         vector signed char c1[2] = {0}, c2[2] = {0}, c3[2] = {0}, c4[2] = {0};
         vector signed char c5[2] = {0}, c6[2] = {0}, c7[2] = {0}, c8[2] = {0};
-        aoffset = const_cast<TA*>(a);
+        aoffset = const_cast<TA *>(a);
         vecOffset = vec;
         j = (rows >> 3);
         if (j > 0) {
@@ -2363,18 +2620,18 @@ class tinyBLAS_HP16_PPC {
                         c7[1] = reinterpret_cast<vector signed char>(vec_xl(0, aoffset7->qs));
                         c8[1] = reinterpret_cast<vector signed char>(vec_xl(0, aoffset8->qs));
 
-                        process_q4_elements(c1, &comparray[0]);
-                        process_q4_elements(c2, &comparray[1]);
-                        process_q4_elements(c3, &comparray[2]);
-                        process_q4_elements(c4, &comparray[3]);
-                        process_q4_elements(c5, &comparray[4]);
-                        process_q4_elements(c6, &comparray[5]);
-                        process_q4_elements(c7, &comparray[6]);
-                        process_q4_elements(c8, &comparray[7]);
+                        process_q4_elements(c1, & comparray[0]);
+                        process_q4_elements(c2, & comparray[1]);
+                        process_q4_elements(c3, & comparray[2]);
+                        process_q4_elements(c4, & comparray[3]);
+                        process_q4_elements(c5, & comparray[4]);
+                        process_q4_elements(c6, & comparray[5]);
+                        process_q4_elements(c7, & comparray[6]);
+                        process_q4_elements(c8, & comparray[7]);
                         vector_permute_store<int8_t, vector signed char>(c1[0], c2[0], c3[0], c4[0], vecOffset, false);
-                        vector_permute_store<int8_t, vector signed char>(c1[1], c2[1], c3[1], c4[1], vecOffset+64, false);
-                        vector_permute_store<int8_t, vector signed char>(c5[0], c6[0], c7[0], c8[0], vecOffset+128, false);
-                        vector_permute_store<int8_t, vector signed char>(c5[1], c6[1], c7[1], c8[1], vecOffset+192, false);
+                        vector_permute_store<int8_t, vector signed char>(c1[1], c2[1], c3[1], c4[1], vecOffset + 64, false);
+                        vector_permute_store<int8_t, vector signed char>(c5[0], c6[0], c7[0], c8[0], vecOffset + 128, false);
+                        vector_permute_store<int8_t, vector signed char>(c5[1], c6[1], c7[1], c8[1], vecOffset + 192, false);
                         aoffset1 += lda;
                         aoffset2 += lda;
                         aoffset3 += lda;
@@ -2405,12 +2662,12 @@ class tinyBLAS_HP16_PPC {
                     c3[1] = reinterpret_cast<vector signed char>(vec_xl(0, aoffset3->qs));
                     c4[1] = reinterpret_cast<vector signed char>(vec_xl(0, aoffset4->qs));
 
-                    process_q4_elements(c1, &comparray[0]);
-                    process_q4_elements(c2, &comparray[1]);
-                    process_q4_elements(c3, &comparray[2]);
-                    process_q4_elements(c4, &comparray[3]);
+                    process_q4_elements(c1, & comparray[0]);
+                    process_q4_elements(c2, & comparray[1]);
+                    process_q4_elements(c3, & comparray[2]);
+                    process_q4_elements(c4, & comparray[3]);
                     vector_permute_store<int8_t, vector signed char>(c1[0], c2[0], c3[0], c4[0], vecOffset, false);
-                    vector_permute_store<int8_t, vector signed char>(c1[1], c2[1], c3[1], c4[1], vecOffset+64, false);
+                    vector_permute_store<int8_t, vector signed char>(c1[1], c2[1], c3[1], c4[1], vecOffset + 64, false);
                     aoffset1 += lda;
                     aoffset2 += lda;
                     aoffset3 += lda;
@@ -2434,12 +2691,12 @@ class tinyBLAS_HP16_PPC {
                         case 1: c1[1] = reinterpret_cast<vector signed char>(vec_xl(0, aoffset1->qs));
                             break;
                     }
-                    process_q4_elements(c1, &comparray[0]);
-                    process_q4_elements(c2, &comparray[1]);
-                    process_q4_elements(c3, &comparray[2]);
-                    process_q4_elements(c4, &comparray[3]);
+                    process_q4_elements(c1, & comparray[0]);
+                    process_q4_elements(c2, & comparray[1]);
+                    process_q4_elements(c3, & comparray[2]);
+                    process_q4_elements(c4, & comparray[3]);
                     vector_permute_store<int8_t, vector signed char>(c1[0], c2[0], c3[0], c4[0], vecOffset, false);
-                    vector_permute_store<int8_t, vector signed char>(c1[1], c2[1], c3[1], c4[1], vecOffset+64, false);
+                    vector_permute_store<int8_t, vector signed char>(c1[1], c2[1], c3[1], c4[1], vecOffset + 64, false);
                     aoffset1 += lda;
                     aoffset2 += lda;
                     aoffset3 += lda;
@@ -2450,39 +2707,38 @@ class tinyBLAS_HP16_PPC {
         }
     }
 
-    template<typename TA>
     template<typename VA, typename VB>
-    void tinyBLAS_Q0_PPC<TA>::packNormal(const block_q8_0* a, int64_t lda, int rows, int cols, VA* vec, bool flip) {
+    void packNormal(const block_q8_0 * a, int64_t lda, int rows, int cols, VA * vec, bool flip) {
         int64_t i, j;
-        block_q8_0 *aoffset = NULL;
-        VA *vecOffset = NULL;
-        block_q8_0* aoffsets[8];
+        block_q8_0 * aoffset = NULL;
+        VA * vecOffset = NULL;
+        block_q8_0 * aoffsets[8];
         __vector_pair arr[8];
         VB c[8][2] = {0};
         VB c1[8] = {0}; VB c2[8] = {0};
-        aoffset = const_cast<block_q8_0*>(a);
+        aoffset = const_cast<block_q8_0 *>(a);
         vecOffset = vec;
         j = (rows >> 3);
         if (j > 0) {
             do {
                 aoffsets[0] = aoffset;
                 for (int it = 1; it < 8; it++)
-                    aoffsets[it] = aoffsets[it-1] + lda;
+                    aoffsets[it] = aoffsets[it - 1] + lda;
                 aoffset += 8 * lda;
 
                 i = (cols >> 3);
                 if (i > 0) {
                 do {
                     for (int it = 0; it < 8; it++) {
-                        arr[it] = __builtin_vsx_lxvp(0, (__vector_pair*)aoffsets[it]->qs);
-                        __builtin_vsx_disassemble_pair(c[it], &arr[it]);
+                        arr[it] = __builtin_vsx_lxvp(0, (__vector_pair *)aoffsets[it]->qs);
+                        __builtin_vsx_disassemble_pair(c[it], & arr[it]);
                         c1[it] = c[it][0];
                         c2[it] = c[it][1];
                     }
                     vector_permute_store<VA, VB>(c1[0], c1[1], c1[2], c1[3], vecOffset, flip);
-                    vector_permute_store<VA, VB>(c2[0], c2[1], c2[2], c2[3], vecOffset+64, flip);
-                    vector_permute_store<VA, VB>(c1[4], c1[5], c1[6], c1[7], vecOffset+128, flip);
-                    vector_permute_store<VA, VB>(c2[4], c2[5], c2[6], c2[7], vecOffset+192, flip);
+                    vector_permute_store<VA, VB>(c2[0], c2[1], c2[2], c2[3], vecOffset + 64, flip);
+                    vector_permute_store<VA, VB>(c1[4], c1[5], c1[6], c1[7], vecOffset + 128, flip);
+                    vector_permute_store<VA, VB>(c2[4], c2[5], c2[6], c2[7], vecOffset + 192, flip);
                     for (int it = 0; it < 8; it++)
                         aoffsets[it] += lda;
                     vecOffset += 256;
@@ -2501,13 +2757,13 @@ class tinyBLAS_HP16_PPC {
             if (i > 0) {
                do {
                     for (int it = 0; it < 4; it++) {
-                        arr[it] = __builtin_vsx_lxvp(0, (__vector_pair*)aoffsets[it]->qs);
-                        __builtin_vsx_disassemble_pair(c[it], &arr[it]);
+                        arr[it] = __builtin_vsx_lxvp(0, (__vector_pair *)aoffsets[it]->qs);
+                        __builtin_vsx_disassemble_pair(c[it], & arr[it]);
                         c1[it] = c[it][0];
                         c2[it] = c[it][1];
                     }
                     vector_permute_store<VA, VB>(c1[0], c1[1], c1[2], c1[3], vecOffset, flip);
-                    vector_permute_store<VA, VB>(c2[0], c2[1], c2[2], c2[3], vecOffset+64, flip);
+                    vector_permute_store<VA, VB>(c2[0], c2[1], c2[2], c2[3], vecOffset + 64, flip);
                     for (int it = 0; it < 4; it++) {
                         aoffsets[it] += lda;
                     }
@@ -2520,24 +2776,24 @@ class tinyBLAS_HP16_PPC {
         if (rows & 3) {
             aoffsets[0]  = aoffset;
             for (int it = 1; it < 3; it++ )
-                aoffsets[it] = aoffsets[it-1] + lda;
+                aoffsets[it] = aoffsets[it - 1] + lda;
             i = (cols >> 3);
             if (i > 0) {
                 do {
                     switch(rows) {
-                        case 3: arr[2] = __builtin_vsx_lxvp(0, (__vector_pair*)aoffsets[2]->qs);
-                                __builtin_vsx_disassemble_pair(c[2], &arr[2]);
+                        case 3: arr[2] = __builtin_vsx_lxvp(0, (__vector_pair *)aoffsets[2]->qs);
+                                __builtin_vsx_disassemble_pair(c[2], & arr[2]);
                                 c1[2] = c[2][0]; c2[2] = c[2][1];
-                        case 2: arr[1] = __builtin_vsx_lxvp(0, (__vector_pair*)aoffsets[1]->qs);
-                                __builtin_vsx_disassemble_pair(c[1], &arr[1]);
+                        case 2: arr[1] = __builtin_vsx_lxvp(0, (__vector_pair *)aoffsets[1]->qs);
+                                __builtin_vsx_disassemble_pair(c[1], & arr[1]);
                                 c1[1] = c[1][0]; c2[1] = c[1][1];
-                        case 1: arr[0] = __builtin_vsx_lxvp(0, (__vector_pair*)aoffsets[0]->qs);
-                                __builtin_vsx_disassemble_pair(c[0], &arr[0]);
+                        case 1: arr[0] = __builtin_vsx_lxvp(0, (__vector_pair *)aoffsets[0]->qs);
+                                __builtin_vsx_disassemble_pair(c[0], & arr[0]);
                                 c1[0] = c[0][0]; c2[0] = c[0][1];
                                 break;
                     }
                     vector_permute_store<VA, VB>(c1[0], c1[1], c1[2], c1[3], vecOffset, flip);
-                    vector_permute_store<VA, VB>(c2[0], c2[1], c2[2], c2[3], vecOffset+64, flip);
+                    vector_permute_store<VA, VB>(c2[0], c2[1], c2[2], c2[3], vecOffset + 64, flip);
                     for (int it = 0; it < 3; it++)
                          aoffsets[it] += lda;
                     vecOffset += 128;
@@ -2547,8 +2803,7 @@ class tinyBLAS_HP16_PPC {
         }
     }
 
-    template<typename TA>
-    void tinyBLAS_Q0_PPC<TA>::mnpack(int64_t m0, int64_t m, int64_t n0, int64_t n) {
+    void mnpack(int64_t m0, int64_t m, int64_t n0, int64_t n) {
         int m_rem = MIN(m - m0, 16);
         int n_rem = MIN(n - n0, 16);
 
@@ -2585,8 +2840,7 @@ class tinyBLAS_HP16_PPC {
     }
 
 
-    template<typename TA>
-    void tinyBLAS_Q0_PPC<TA>::KERNEL_4x8(int64_t ii, int64_t jj) {
+    void KERNEL_4x8(int64_t ii, int64_t jj) {
         vec_t vec_A[8], vec_B[16] = {0};
         acc_t acc_0, acc_1;
         std::array<int, 4> comparray {};
@@ -2594,26 +2848,26 @@ class tinyBLAS_HP16_PPC {
         vector float vs[8] = {0};
         bool isAblock_q4 = std::is_same_v<TA, block_q4_0>;
         for (int l = 0; l < k; l++) {
-            __builtin_mma_xxsetaccz(&acc_0);
-            __builtin_mma_xxsetaccz(&acc_1);
+            __builtin_mma_xxsetaccz(& acc_0);
+            __builtin_mma_xxsetaccz(& acc_1);
             if (std::is_same_v<TA, block_q4_0>) {
-               packNormalInt4<4>((A+(ii*lda)+l), lda, 4, 4, (int8_t*)vec_A, comparray);
+               packNormalInt4<4>((A + (ii * lda) + l), lda, 4, 4, (int8_t *)vec_A, comparray);
             } else {
-               packNormal<int8_t, vector signed char>((const block_q8_0*)(A+(ii*lda)+l), lda, 4, 8, (int8_t*)vec_A, false);
+               packNormal<int8_t, vector signed char>((const block_q8_0 *)(A + (ii * lda) + l), lda, 4, 8, (int8_t *)vec_A, false);
             }
-            packNormal<uint8_t, vector unsigned char>((B+(jj*ldb)+l), ldb, 8, 8, (uint8_t*)vec_B, true);
+            packNormal<uint8_t, vector unsigned char>((B + (jj * ldb) + l), ldb, 8, 8, (uint8_t *)vec_B, true);
             for(int x = 0; x < 8; x++) {
-                __builtin_mma_xvi8ger4pp(&acc_0, vec_A[x], vec_B[x]);
-                __builtin_mma_xvi8ger4pp(&acc_1, vec_A[x], vec_B[x+8]);
+                __builtin_mma_xvi8ger4pp(& acc_0, vec_A[x], vec_B[x]);
+                __builtin_mma_xvi8ger4pp(& acc_1, vec_A[x], vec_B[x+8]);
             }
             for (int I = 0; I<4; I++) {
                 for (int J = 0; J<4; J++) {
-                    *((float*)&vs[I]+J) = (unhalf((A+((ii+I)*lda)+l)->d) * unhalf((B+((jj+J)*ldb)+l)->d));
-                    *((float*)&vs[I+4]+J) = (unhalf((A+((ii+I)*lda)+l)->d) * unhalf((B+((jj+J+4)*ldb)+l)->d));
+                    *((float *)& vs[I] + J) = (unhalf((A + ((ii + I) * lda) + l)->d) * unhalf((B + ((jj + J) * ldb) + l)->d));
+                    *((float *)& vs[I + 4] + J) = (unhalf((A +((ii + I) * lda) + l)->d) * unhalf((B + ((jj + J + 4) * ldb) + l)->d));
                 }
             }
             if (!isAblock_q4) {
-                auto aoffset = A+(ii*lda)+l;
+                auto aoffset = A + (ii * lda) + l;
                 for (int i = 0; i < 4; i++) {
                     comparray[i] = 0;
                     int ca = 0;
@@ -2624,15 +2878,14 @@ class tinyBLAS_HP16_PPC {
                     aoffset += lda;
                 }
             }
-            compute(&acc_0, 0, 0, comparray, vs, fin_res);
-            compute(&acc_1, 0, 4, comparray, vs, fin_res);
+            compute(& acc_0, 0, 0, comparray, vs, fin_res);
+            compute(& acc_1, 0, 4, comparray, vs, fin_res);
         }
         save_res(ii, jj, 0, fin_res);
-        save_res(ii, jj+4, 4, fin_res);
+        save_res(ii, jj + 4, 4, fin_res);
     }
 
-    template<typename TA>
-    void tinyBLAS_Q0_PPC<TA>::KERNEL_8x4(int64_t ii, int64_t jj) {
+    void KERNEL_8x4(int64_t ii, int64_t jj) {
         vec_t vec_A[16], vec_B[8] = {0};
         acc_t acc_0, acc_1;
         std::array<int, 8> comparray {};
@@ -2640,25 +2893,25 @@ class tinyBLAS_HP16_PPC {
         vector float vs[8] = {0};
         bool isAblock_q4 = std::is_same_v<TA, block_q4_0>;
         for (int l = 0; l < k; l++) {
-            __builtin_mma_xxsetaccz(&acc_0);
-            __builtin_mma_xxsetaccz(&acc_1);
+            __builtin_mma_xxsetaccz(& acc_0);
+            __builtin_mma_xxsetaccz(& acc_1);
             if (std::is_same_v<TA, block_q4_0>) {
-               packNormalInt4<8>((A+(ii*lda)+l), lda, 8, 4, (int8_t*)vec_A, comparray);
+               packNormalInt4<8>((A + (ii * lda) + l), lda, 8, 4, (int8_t *)vec_A, comparray);
             } else {
-               packNormal<int8_t, vector signed char>((const block_q8_0*)(A+(ii*lda)+l), lda, 8, 8, (int8_t*)vec_A, false);
+               packNormal<int8_t, vector signed char>((const block_q8_0 *)(A + (ii * lda) + l), lda, 8, 8, (int8_t *)vec_A, false);
             }
-            packNormal<uint8_t, vector unsigned char>((B+(jj*ldb)+l), ldb, 4, 8, (uint8_t*)vec_B, true);
+            packNormal<uint8_t, vector unsigned char>((B + (jj * ldb) + l), ldb, 4, 8, (uint8_t *)vec_B, true);
             for(int x = 0; x < 8; x++) {
-                __builtin_mma_xvi8ger4pp(&acc_0, vec_A[x], vec_B[x]);
-                __builtin_mma_xvi8ger4pp(&acc_1, vec_A[x+8], vec_B[x]);
+                __builtin_mma_xvi8ger4pp(& acc_0, vec_A[x], vec_B[x]);
+                __builtin_mma_xvi8ger4pp(& acc_1, vec_A[x + 8], vec_B[x]);
             }
-            for (int I = 0; I<8; I++) {
-                for (int J = 0; J<4; J++) {
-                    *((float*)&vs[I]+J) = (unhalf((A+((ii+I)*lda)+l)->d) * unhalf((B+((jj+J)*ldb)+l)->d));
+            for (int I = 0; I < 8; I++) {
+                for (int J = 0; J < 4; J++) {
+                    *((float *)&vs[I] + J) = (unhalf((A + ((ii + I) * lda) + l)->d) * unhalf((B + ((jj + J) * ldb) + l)->d));
                 }
             }
             if (!isAblock_q4) {
-                auto aoffset = A+(ii*lda)+l;
+                auto aoffset = A + (ii * lda) + l;
                 for (int i = 0; i < 8; i++) {
                     comparray[i] = 0;
                     int ca = 0;
@@ -2669,15 +2922,14 @@ class tinyBLAS_HP16_PPC {
                     aoffset += lda;
                 }
             }
-            compute(&acc_0, 0, 0, comparray, vs, fin_res);
-            compute(&acc_1, 4, 4, comparray, vs, fin_res);
+            compute(& acc_0, 0, 0, comparray, vs, fin_res);
+            compute(& acc_1, 4, 4, comparray, vs, fin_res);
         }
         save_res(ii, jj, 0, fin_res);
-        save_res(ii+4, jj, 4, fin_res);
+        save_res(ii + 4, jj, 4, fin_res);
     }
 
-    template<typename TA>
-    void tinyBLAS_Q0_PPC<TA>::KERNEL_8x8(int64_t ii, int64_t jj) {
+    void KERNEL_8x8(int64_t ii, int64_t jj) {
         vec_t vec_A[16], vec_B[16] = {0};
         acc_t acc_0, acc_1, acc_2, acc_3;
         acc_t acc_4, acc_5, acc_6, acc_7;
@@ -2686,30 +2938,30 @@ class tinyBLAS_HP16_PPC {
         vector float vs[16] = {0};
         bool isAblock_q4 = std::is_same_v<TA, block_q4_0>;
         for (int l = 0; l < k; l++) {
-            __builtin_mma_xxsetaccz(&acc_0);
-            __builtin_mma_xxsetaccz(&acc_1);
-            __builtin_mma_xxsetaccz(&acc_2);
-            __builtin_mma_xxsetaccz(&acc_3);
+            __builtin_mma_xxsetaccz(& acc_0);
+            __builtin_mma_xxsetaccz(& acc_1);
+            __builtin_mma_xxsetaccz(& acc_2);
+            __builtin_mma_xxsetaccz(& acc_3);
             if (std::is_same_v<TA, block_q4_0>) {
-               packNormalInt4<8>((A+(ii*lda)+l), lda, 8, 4, (int8_t*)vec_A, comparray);
+               packNormalInt4<8>((A + (ii * lda) + l), lda, 8, 4, (int8_t *)vec_A, comparray);
             } else {
-               packNormal<int8_t, vector signed char>((const block_q8_0*)(A+(ii*lda)+l), lda, 8, 8, (int8_t*)vec_A, false);
+               packNormal<int8_t, vector signed char>((const block_q8_0 *)(A + (ii * lda) + l), lda, 8, 8, (int8_t *)vec_A, false);
             }
-            packNormal<uint8_t, vector unsigned char>((B+(jj*ldb)+l), ldb, 8, 8, (uint8_t*)vec_B, true);
+            packNormal<uint8_t, vector unsigned char>((B + (jj * ldb) + l), ldb, 8, 8, (uint8_t *)vec_B, true);
             for(int x = 0; x < 8; x++) {
-                __builtin_mma_xvi8ger4pp(&acc_0, vec_A[x], vec_B[x]);
-                __builtin_mma_xvi8ger4pp(&acc_1, vec_A[x+8], vec_B[x]);
-                __builtin_mma_xvi8ger4pp(&acc_2, vec_A[x], vec_B[x+8]);
-                __builtin_mma_xvi8ger4pp(&acc_3, vec_A[x+8], vec_B[x+8]);
+                __builtin_mma_xvi8ger4pp(& acc_0, vec_A[x], vec_B[x]);
+                __builtin_mma_xvi8ger4pp(& acc_1, vec_A[x + 8], vec_B[x]);
+                __builtin_mma_xvi8ger4pp(& acc_2, vec_A[x], vec_B[x + 8]);
+                __builtin_mma_xvi8ger4pp(& acc_3, vec_A[x + 8], vec_B[x + 8]);
             }
-            for (int I = 0; I<8; I++) {
-                for (int J = 0; J<4; J++) {
-                    *((float*)&vs[I]+J) = (unhalf((A+((ii+I)*lda)+l)->d) * unhalf((B+((jj+J)*ldb)+l)->d));
-                    *((float*)&vs[I+8]+J) = (unhalf((A+((ii+I)*lda)+l)->d) * unhalf((B+((jj+J+4)*ldb)+l)->d));
+            for (int I = 0; I < 8 ; I++) {
+                for (int J = 0; J < 4; J++) {
+                    *((float *)& vs[I] + J) = (unhalf((A + ((ii + I) * lda) + l)->d) * unhalf((B + ((jj + J) * ldb) + l)->d));
+                    *((float *)& vs[I + 8] + J) = (unhalf((A + ((ii + I) * lda) + l)->d) * unhalf((B + ((jj + J + 4) * ldb) + l)->d));
                 }
             }
             if (!isAblock_q4) {
-                auto aoffset = A+(ii*lda)+l;
+                auto aoffset = A + (ii * lda) + l;
                 for (int i = 0; i < 8; i++) {
                     comparray[i] = 0;
                     int ca = 0;
@@ -2720,19 +2972,99 @@ class tinyBLAS_HP16_PPC {
                     aoffset += lda;
                 }
             }
-            compute(&acc_0, 0, 0, comparray, vs, fin_res);
-            compute(&acc_1, 4, 4, comparray, vs, fin_res);
-            compute(&acc_2, 0, 8, comparray, vs, fin_res);
-            compute(&acc_3, 4, 12, comparray, vs, fin_res);
+            compute(& acc_0, 0, 0, comparray, vs, fin_res);
+            compute(& acc_1, 4, 4, comparray, vs, fin_res);
+            compute(& acc_2, 0, 8, comparray, vs, fin_res);
+            compute(& acc_3, 4, 12, comparray, vs, fin_res);
         }
         save_res(ii, jj, 0, fin_res);
-        save_res(ii+4, jj, 4, fin_res);
-        save_res(ii, jj+4, 8, fin_res);
-        save_res(ii+4, jj+4, 12, fin_res);
+        save_res(ii + 4, jj, 4, fin_res);
+        save_res(ii, jj + 4, 8, fin_res);
+        save_res(ii + 4, jj + 4, 12, fin_res);
     }
 
-    template<typename TA>
-    void tinyBLAS_Q0_PPC<TA>::gemm_small(int64_t m0, int64_t m, int64_t n0, int64_t n, int RM, int RN) {
+    void KERNEL_Q0(int64_t ii, int64_t jj, int64_t mc, int64_t nc, int64_t kc, int64_t l, vec_t * vec_A, vec_t * vec_B) {
+        acc_t acc[8];
+        for (int i = 0; i < mc ; i += 16) {
+            for (int j = 0; j < nc; j += 8) {
+                int A0_base = (i / 16) * (2 * 32 * kc);
+                int B0_base = (j / 8) * (32 * kc);
+                for (int x = 0; x < 8; x++) {
+                     __builtin_mma_xxsetaccz(&acc[x]);
+                }
+                for (int64_t kk = 0; kk < kc; kk++) {
+                    int A0_block_idx = A0_base + kk * 32;
+                    int B0_block_idx = B0_base + kk * 32;
+                    int A1_block_idx = A0_block_idx + 32 * kc;
+                    int B1_block_idx = B0_block_idx + 32 * kc;
+                    vec_t * A0_block = & vec_A[A0_block_idx];
+                    vec_t * B0_block = & vec_B[B0_block_idx];
+                    vec_t * A1_block = & vec_A[A1_block_idx];
+                    for (int it = 0; it < 4; it++) {
+                        for (int x = 0; x < 4; x++) {
+                            __builtin_mma_xvf16ger2pp(& acc[0], A0_block[8 * it + x], B0_block[8 * it + x]);
+                            __builtin_mma_xvf16ger2pp(& acc[1], A0_block[8 * it + x], B0_block[8 * it + x + 4]);
+                            __builtin_mma_xvf16ger2pp(& acc[2], A0_block[8 * it + x + 4], B0_block[8 * it + x]);
+                            __builtin_mma_xvf16ger2pp(& acc[3], A0_block[8 * it + x + 4], B0_block[8 * it + x + 4]);
+                            __builtin_mma_xvf16ger2pp(& acc[4], A1_block[8 * it + x], B0_block[8 * it + x]);
+                            __builtin_mma_xvf16ger2pp(& acc[5], A1_block[8 * it + x], B0_block[8 * it+ x + 4]);
+                            __builtin_mma_xvf16ger2pp(& acc[6], A1_block[8 * it + x + 4], B0_block[8 * it + x]);
+                            __builtin_mma_xvf16ger2pp(& acc[7], A1_block[8 * it + x + 4], B0_block[8 * it + x + 4]);
+                        }
+                    }
+                }
+                if (l == 0) {
+                    save_acc(& acc[0], ii + i, jj + j);
+                    save_acc(& acc[1], ii + i, jj + j + 4);
+                    save_acc(& acc[2], ii + i + 4, jj + j);
+                    save_acc(& acc[3], ii + i + 4, jj + j + 4);
+                    save_acc(& acc[4], ii + i + 8, jj + j);
+                    save_acc(& acc[5], ii + i + 8, jj + j + 4);
+                    save_acc(& acc[6], ii + i + 12, jj + j);
+                    save_acc(& acc[7], ii + i + 12, jj + j + 4);
+                } else {
+                    add_save_acc(& acc[0], ii + i, jj + j);
+                    add_save_acc(& acc[1], ii + i, jj + j + 4);
+                    add_save_acc(& acc[2], ii + i + 4, jj + j);
+                    add_save_acc(& acc[3], ii + i + 4, jj + j + 4);
+                    add_save_acc(& acc[4], ii + i + 8, jj + j);
+                    add_save_acc(& acc[5], ii + i + 8, jj + j + 4);
+                    add_save_acc(& acc[6], ii + i + 12, jj + j);
+                    add_save_acc(& acc[7], ii + i + 12, jj + j + 4);
+                }
+            }
+        }
+    }
+
+    void matmul_tiled(int64_t m, int64_t n, int64_t mc, int64_t nc, int64_t kc) {
+        vec_t A_pack[mc * kc * 4];
+        vec_t B_pack[nc * kc * 4];
+        constexpr bool is_Ablock_q4 = std::is_same_v<TA, block_q4_0>;
+        int64_t ytiles = m / mc;
+        int64_t xtiles = n / nc;
+        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 = (job / xtiles) * mc;
+            int64_t jj = (job % xtiles) * nc;
+            for (int64_t kk = 0; kk < k; kk += kc) {
+                if constexpr(is_Ablock_q4) {
+                    packNormal_q4_fp16(A + ii * lda + kk, lda, mc, kc, (uint8_t *)A_pack);
+                } else {
+                    packNormal_q8_fp16(A + ii * lda + kk, lda, mc, kc, (uint8_t *)A_pack);
+                }
+                packNormal_q8_fp16(B + jj * ldb + kk, ldb, nc, kc, (uint8_t *)B_pack);
+                KERNEL_Q0(ii, jj, mc, nc, kc, kk, A_pack, B_pack);
+            }
+        }
+    }
+
+    void gemm_small(int64_t m0, int64_t m, int64_t n0, int64_t n, int RM, int RN) {
         int64_t ytiles = (m - m0) / RM;
         int64_t xtiles = (n - n0) / RN;
         int64_t tiles = xtiles * ytiles;
@@ -2754,32 +3086,32 @@ class tinyBLAS_HP16_PPC {
             vector float fin_res[4] = {0};
             vector float vs[4] = {0};
             vector float CA[4] = {0};
-            __builtin_prefetch((A+(ii*lda)+0)->qs, 0, 1); // prefetch first value
-            __builtin_prefetch((B+(jj*ldb)+0)->qs, 0, 1); // prefetch first value
+            __builtin_prefetch((A + (ii * lda) + 0)->qs, 0, 1); // prefetch first value
+            __builtin_prefetch((B + (jj * ldb) + 0)->qs, 0, 1); // prefetch first value
             for (int l = 0; l < k; l++) {
-                __builtin_prefetch((A+(ii*lda)+(l+1))->qs, 0, 1); // prefetch one loop ahead
-                __builtin_prefetch((B+(jj*ldb)+(l+1))->qs, 0, 1); // prefetch one loop ahead
-                __builtin_mma_xxsetaccz(&acc_0);
+                __builtin_prefetch((A + (ii * lda) + (l + 1))->qs, 0, 1); // prefetch one loop ahead
+                __builtin_prefetch((B + (jj * ldb) + (l + 1))->qs, 0, 1); // prefetch one loop ahead
+                __builtin_mma_xxsetaccz(& acc_0);
                 if (isAblock_q4) {
-                   packNormalInt4<4>((A+(ii*lda)+l), lda, RM, 4, (int8_t*)vec_A, comparray);
+                    packNormalInt4<4>((A + (ii * lda) + l), lda, RM, 4, (int8_t *)vec_A, comparray);
                 } else {
-                   packNormal<int8_t, vector signed char>((const block_q8_0*)(A+(ii*lda)+l), lda, RM, 8, (int8_t*)vec_A, false);
+                    packNormal<int8_t, vector signed char>((const block_q8_0 *)(A + (ii * lda) + l), lda, RM, 8, (int8_t *)vec_A, false);
                 }
-                packNormal<uint8_t, vector unsigned char>((B+(jj*ldb)+l), ldb, RN, 8, (uint8_t*)vec_B, true);
-                for(int x = 0; x < 8; x+=4) {
-                    __builtin_mma_xvi8ger4pp(&acc_0, vec_A[x], vec_B[x]);
-                    __builtin_mma_xvi8ger4pp(&acc_0, vec_A[x+1], vec_B[x+1]);
-                    __builtin_mma_xvi8ger4pp(&acc_0, vec_A[x+2], vec_B[x+2]);
-                    __builtin_mma_xvi8ger4pp(&acc_0, vec_A[x+3], vec_B[x+3]);
+                packNormal<uint8_t, vector unsigned char>((B + (jj * ldb) + l), ldb, RN, 8, (uint8_t *)vec_B, true);
+                for (int x = 0; x < 8; x += 4) {
+                    __builtin_mma_xvi8ger4pp(& acc_0, vec_A[x], vec_B[x]);
+                    __builtin_mma_xvi8ger4pp(& acc_0, vec_A[x + 1], vec_B[x + 1]);
+                    __builtin_mma_xvi8ger4pp(& acc_0, vec_A[x + 2], vec_B[x + 2]);
+                    __builtin_mma_xvi8ger4pp(& acc_0, vec_A[x + 3], vec_B[x + 3]);
                 }
-                for (int I = 0; I<RM; I++) {
-                    for (int J = 0; J<RN; J++) {
-                        *((float*)&vs[I]+J) = (unhalf((A+((ii+I)*lda)+l)->d) * unhalf((B+((jj+J)*ldb)+l)->d));
+                for (int I = 0; I < RM; I++) {
+                    for (int J = 0; J < RN; J++) {
+                        *((float*)&vs[I] + J) = (unhalf((A + ((ii + I) * lda) + l)->d) * unhalf((B + ((jj + J) * ldb) + l)->d));
                     }
                 }
-                __builtin_mma_disassemble_acc(vec_C, &acc_0);
+                __builtin_mma_disassemble_acc(vec_C, & acc_0);
                 if (!isAblock_q4) {
-                    auto aoffset = A+(ii*lda)+l;
+                    auto aoffset = A + (ii * lda) + l;
                     for (int i = 0; i < RM; i++) {
                         comparray[i] = 0;
                         int ca = 0;
@@ -2800,9 +3132,21 @@ class tinyBLAS_HP16_PPC {
         }
     }
 
-    template<typename TA>
+    template<int RM, int RN>
+    inline void kernel(int64_t ii, int64_t jj) {
+        if constexpr(RM == 4 && RN == 8) {
+            KERNEL_4x8(ii,jj);
+        } else if constexpr(RM == 8 && RN == 4) {
+            KERNEL_8x4(ii,jj);
+        } else if constexpr(RM == 8 && RN == 8) {
+            KERNEL_8x8(ii,jj);
+        } else {
+            assert(false && "RN/RM values not supported");
+        }
+    }
+
     template <int RM, int RN>
-    NOINLINE void tinyBLAS_Q0_PPC<TA>::gemm(int64_t m0, int64_t m, int64_t n0, int64_t n) {
+    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;
@@ -2814,12 +3158,20 @@ class tinyBLAS_HP16_PPC {
         for (int64_t job = start; job < end; ++job) {
             int64_t ii = m0 + job / xtiles * RM;
             int64_t jj = n0 + job % xtiles * RN;
-            this->kernel<RM, RN>(ii, jj);
+            kernel<RM, RN>(ii, jj);
         }
     }
-
-template class tinyBLAS_Q0_PPC<block_q4_0>;
-template class tinyBLAS_Q0_PPC<block_q8_0>;
+    const TA * const A;
+    const block_q8_0 * const B;
+    float * C;
+    const int64_t k;
+    int64_t kc;
+    const int64_t lda;
+    const int64_t ldb;
+    const int64_t ldc;
+    const int ith;
+    const int nth;
+};
 
 class tinyBLAS_PPC {
   public:

ggml/src/ggml-cuda/fattn-tile.cuh

@@ -1186,8 +1186,10 @@ static void launch_fattn_tile_switch_ncols2(ggml_backend_cuda_context & ctx, ggm
     GGML_ASSERT(Q->ne[2] % K->ne[2] == 0);
     const int gqa_ratio = Q->ne[2] / K->ne[2];
 
+    // On NVIDIA (Pascal and older) the GQA optimizations seem to be detrimental in some cases.
+    // However, for DKQ == 576, DV == 512 only the kernel variant with GQA optimizations is implemented.
     const bool nvidia = GGML_CUDA_CC_IS_NVIDIA(ggml_cuda_info().devices[ggml_cuda_get_device()].cc);
-    const int gqa_limit = nvidia && gqa_ratio <= 4 ? 16 : INT_MAX;
+    const int gqa_limit = nvidia && gqa_ratio <= 4 && DV <= 256 ? 16 : INT_MAX;
     const bool use_gqa_opt = mask && max_bias == 0.0f && Q->ne[1] <= gqa_limit && K->ne[1] % FATTN_KQ_STRIDE == 0;
 
     if constexpr (DV == 512) {

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

@@ -57,6 +57,8 @@ layout (push_constant) uniform parameter
     uint nbi1;
     uint ne11;
 #else
+    uint base_work_group_z;
+    uint num_batches;
     uint k_split;
     uint ne02;
     uint ne12;
@@ -108,7 +110,7 @@ void main() {
     const uint ic = gl_WorkGroupID.y;
 
 #ifdef MUL_MAT_ID
-    const uint expert_idx = gl_GlobalInvocationID.z;
+    const uint expert_idx = gl_WorkGroupID.z;
     if (ic * BN >= data_expert_count[expert_idx]) {
         return;
     }
@@ -118,7 +120,7 @@ void main() {
 #endif
 
 #ifndef MUL_MAT_ID
-    const uint batch_idx = gl_GlobalInvocationID.z;
+    const uint batch_idx = gl_WorkGroupID.z + p.base_work_group_z;
 
     const uint i13 = batch_idx / p.ne12;
     const uint i12 = batch_idx % p.ne12;
@@ -276,7 +278,7 @@ void main() {
     const uint dc = ic * BN + warp_c * WN;
 
 #ifndef MUL_MAT_ID
-    const uint offsets = batch_idx * p.batch_stride_d + ik * p.batch_stride_d * gl_NumWorkGroups.z;
+    const uint offsets = batch_idx * p.batch_stride_d + ik * p.batch_stride_d * p.num_batches;
 #endif
 
     [[unroll]] for (uint wsic = 0; wsic < WNITER; wsic++) {

ggml/src/ggml-webgpu/ggml-webgpu.cpp

@@ -2008,6 +2008,14 @@ static std::optional<webgpu_command> ggml_webgpu_encode_node(webgpu_context ctx,
             return ggml_webgpu_unary_op(ctx, src0, node);
         case GGML_OP_LOG:
             return ggml_webgpu_unary_op(ctx, src0, node);
+        case GGML_OP_SQR:
+            return ggml_webgpu_unary_op(ctx, src0, node);
+        case GGML_OP_SQRT:
+            return ggml_webgpu_unary_op(ctx, src0, node);
+        case GGML_OP_SIN:
+            return ggml_webgpu_unary_op(ctx, src0, node);
+        case GGML_OP_COS:
+            return ggml_webgpu_unary_op(ctx, src0, node);
         case GGML_OP_PAD:
             return ggml_webgpu_pad(ctx, src0, node);
         case GGML_OP_ARGMAX:
@@ -2967,6 +2975,18 @@ static bool ggml_backend_webgpu_device_supports_op(ggml_backend_dev_t dev, const
         case GGML_OP_LOG:
             supports_op = (op->type == GGML_TYPE_F32 || op->type == GGML_TYPE_F16) && (src0->type == op->type);
             break;
+        case GGML_OP_SQR:
+            supports_op = (op->type == GGML_TYPE_F32 || op->type == GGML_TYPE_F16) && (src0->type == op->type);
+            break;
+        case GGML_OP_SQRT:
+            supports_op = (op->type == GGML_TYPE_F32 || op->type == GGML_TYPE_F16) && (src0->type == op->type);
+            break;
+        case GGML_OP_SIN:
+            supports_op = (op->type == GGML_TYPE_F32 || op->type == GGML_TYPE_F16) && (src0->type == op->type);
+            break;
+        case GGML_OP_COS:
+            supports_op = (op->type == GGML_TYPE_F32 || op->type == GGML_TYPE_F16) && (src0->type == op->type);
+            break;
         case GGML_OP_PAD:
             supports_op = op->type == GGML_TYPE_F32 && src0->type == GGML_TYPE_F32;
             break;

ggml/src/ggml-webgpu/wgsl-shaders/unary.wgsl

@@ -170,6 +170,20 @@ fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
 #ifdef TRUNC
     let res = trunc(src[params.offset_src + src_idx]);
 #endif
+#ifdef SQR
+    let res = src[params.offset_src + src_idx] * src[params.offset_src + src_idx];
+#endif
+#ifdef SQRT
+    let res = sqrt(src[params.offset_src + src_idx]);
+#endif
+#ifdef SIN
+    let res_f32 = sin(f32(src[params.offset_src + src_idx]));
+    let res = TYPE(res_f32);
+#endif
+#ifdef COS
+    let res_f32 = cos(f32(src[params.offset_src + src_idx]));
+    let res = TYPE(res_f32);
+#endif
 
 #ifdef INPLACE
     src[params.offset_src + src_idx] = res;

gguf-py/gguf/constants.py

@@ -435,6 +435,7 @@ class MODEL_ARCH(IntEnum):
     T5               = auto()
     T5ENCODER        = auto()
     JAIS             = auto()
+    JAIS2            = auto()
     NEMOTRON         = auto()
     NEMOTRON_H       = auto()
     NEMOTRON_H_MOE   = auto()
@@ -472,6 +473,7 @@ class MODEL_ARCH(IntEnum):
     RND1             = auto()
     PANGU_EMBED      = auto()
     MISTRAL3         = auto()
+    PADDLEOCR        = auto()
     MIMO2            = auto()
     STEP35           = auto()
     LLAMA_EMBED      = auto()
@@ -652,6 +654,7 @@ class MODEL_TENSOR(IntEnum):
     ENC_OUTPUT_NORM      = auto()
     CLS                  = auto() # classifier
     CLS_OUT              = auto() # classifier output projection
+    CLS_NORM             = auto()
     CONV1D               = auto()
     CONVNEXT_DW          = auto()
     CONVNEXT_NORM        = auto()
@@ -873,6 +876,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
     MODEL_ARCH.T5:               "t5",
     MODEL_ARCH.T5ENCODER:        "t5encoder",
     MODEL_ARCH.JAIS:             "jais",
+    MODEL_ARCH.JAIS2:            "jais2",
     MODEL_ARCH.NEMOTRON:         "nemotron",
     MODEL_ARCH.NEMOTRON_H:       "nemotron_h",
     MODEL_ARCH.NEMOTRON_H_MOE:   "nemotron_h_moe",
@@ -911,6 +915,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
     MODEL_ARCH.RND1:             "rnd1",
     MODEL_ARCH.PANGU_EMBED:      "pangu-embedded",
     MODEL_ARCH.MISTRAL3:         "mistral3",
+    MODEL_ARCH.PADDLEOCR:        "paddleocr",
     MODEL_ARCH.MIMO2:            "mimo2",
     MODEL_ARCH.STEP35:           "step35",
     MODEL_ARCH.LLAMA_EMBED:      "llama-embed",
@@ -1088,6 +1093,7 @@ TENSOR_NAMES: dict[MODEL_TENSOR, str] = {
     MODEL_TENSOR.ENC_OUTPUT_NORM:           "enc.output_norm",
     MODEL_TENSOR.CLS:                       "cls",
     MODEL_TENSOR.CLS_OUT:                   "cls.output",
+    MODEL_TENSOR.CLS_NORM:                  "cls.norm",
     MODEL_TENSOR.CONV1D:                    "conv1d",
     MODEL_TENSOR.CONVNEXT_DW:               "convnext.{bid}.dw",
     MODEL_TENSOR.CONVNEXT_NORM:             "convnext.{bid}.norm",
@@ -1507,6 +1513,7 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.FFN_NORM,
         MODEL_TENSOR.CLS,
         MODEL_TENSOR.CLS_OUT,
+        MODEL_TENSOR.CLS_NORM,
     ],
     MODEL_ARCH.NOMIC_BERT: [
         MODEL_TENSOR.TOKEN_EMBD,
@@ -2814,6 +2821,19 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.FFN_GATE,
         MODEL_TENSOR.FFN_UP,
     ],
+    MODEL_ARCH.JAIS2: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.OUTPUT,
+        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_Q,
+        MODEL_TENSOR.ATTN_K,
+        MODEL_TENSOR.ATTN_V,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.FFN_NORM,
+        MODEL_TENSOR.FFN_DOWN,
+        MODEL_TENSOR.FFN_UP,
+    ],
     MODEL_ARCH.NEMOTRON: [
         MODEL_TENSOR.TOKEN_EMBD,
         MODEL_TENSOR.OUTPUT_NORM,
@@ -3168,6 +3188,20 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.FFN_DOWN,
         MODEL_TENSOR.FFN_UP,
     ],
+    MODEL_ARCH.PADDLEOCR: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.OUTPUT,
+        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_Q,
+        MODEL_TENSOR.ATTN_K,
+        MODEL_TENSOR.ATTN_V,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.FFN_NORM,
+        MODEL_TENSOR.FFN_GATE,
+        MODEL_TENSOR.FFN_DOWN,
+        MODEL_TENSOR.FFN_UP,
+    ],
     MODEL_ARCH.FALCON_H1: [
         # Token embedding
         MODEL_TENSOR.TOKEN_EMBD,
@@ -3829,6 +3863,7 @@ class VisionProjectorType:
     VOXTRAL = "voxtral"
     LFM2 = "lfm2"
     KIMIVL = "kimivl"
+    PADDLEOCR = "paddleocr"
     KIMIK25 = "kimik25"
     LIGHTONOCR = "lightonocr"
     COGVLM = "cogvlm"

gguf-py/gguf/tensor_mapping.py

@@ -1240,6 +1240,10 @@ class TensorNameMap:
         MODEL_TENSOR.CLS_OUT: (
             "classifier.out_proj", # roberta
         ),
+
+        MODEL_TENSOR.CLS_NORM: (
+            "head.norm", # modern-bert
+        ),
         #############################################################################
 
         MODEL_TENSOR.CONVNEXT_DW: (
@@ -1321,6 +1325,7 @@ class TensorNameMap:
             "multi_modal_projector.linear_{bid}",
             "mm_projector.proj.linear_{bid}", # Kimi-K2.5
             "visual.merger.mlp.{bid}", # qwen2vl
+            "mlp_AR.linear_{bid}", # PaddleOCR-VL
             "merger.mlp.{bid}",
         ),
 
@@ -1570,6 +1575,7 @@ class TensorNameMap:
             "mm_projector.pre_norm", # Kimi-K2.5
             "pre_mm_projector_norm",
             "model.vision.linear_proj.norm1", # cogvlm
+            "mlp_AR.pre_norm", # PaddleOCR-VL
             "merger.ln_q",
         ),
 
@@ -1595,6 +1601,7 @@ class TensorNameMap:
 
         MODEL_TENSOR.V_RESMPL_ATTN_OUT: (
             "resampler.attn.out_proj",
+            "model.vision_model.head.attention.out_proj",
         ),
 
         MODEL_TENSOR.V_RESMPL_KV: (

include/llama.h

@@ -389,6 +389,7 @@ extern "C" {
         bool only_copy;                       // only copy tensors - ftype, allow_requantize and quantize_output_tensor are ignored
         bool pure;                            // quantize all tensors to the default type
         bool keep_split;                      // quantize to the same number of shards
+        bool dry_run;                         // calculate and show the final quantization size without performing quantization
         void * imatrix;                       // pointer to importance matrix data
         void * kv_overrides;                  // pointer to vector containing overrides
         void * tensor_types;                  // pointer to vector containing tensor types

models/templates/stepfun-ai-Step-3.5-Flash.jinja

@@ -0,0 +1,80 @@
+{% macro render_content(content) %}{% if content is none %}{{- '' }}{% elif content is string %}{{- content }}{% elif content is mapping %}{{- content['value'] if 'value' in content else content['text'] }}{% elif content is iterable %}{% for item in content %}{% if item.type == 'text' %}{{- item['value'] if 'value' in item else item['text'] }}{% elif item.type == 'image' %}<im_patch>{% endif %}{% endfor %}{% endif %}{% endmacro %}
+{{bos_token}}{%- if tools %}
+    {{- '<|im_start|>system\n' }}
+    {%- if messages[0].role == 'system' %}
+        {{- render_content(messages[0].content) + '\n\n' }}
+    {%- endif %}
+    {{- "# Tools\n\nYou have access to the following functions in JSONSchema format:\n\n<tools>" }}
+    {%- for tool in tools %}
+        {{- "\n" }}
+        {{- tool | tojson(ensure_ascii=False) }}
+    {%- endfor %}
+    {{- "\n</tools>\n\nIf you choose to call a function ONLY reply in the following format with NO suffix:\n\n<tool_call>\n<function=example_function_name>\n<parameter=example_parameter_1>\nvalue_1\n</parameter>\n<parameter=example_parameter_2>\nThis is the value for the second parameter\nthat can span\nmultiple lines\n</parameter>\n</function>\n</tool_call>\n\n<IMPORTANT>\nReminder:\n- Function calls MUST follow the specified format: an inner <function=...>\n...\n</function> block must be nested within <tool_call>\n...\n</tool_call> XML tags\n- Required parameters MUST be specified\n</IMPORTANT><|im_end|>\n" }}
+{%- else %}
+    {%- if messages[0].role == 'system' %}
+        {{- '<|im_start|>system\n' + render_content(messages[0].content) + '<|im_end|>\n' }}
+    {%- endif %}
+{%- endif %}
+{%- set ns = namespace(multi_step_tool=true, last_query_index=messages|length - 1) %}
+{%- for message in messages[::-1] %}
+    {%- set index = (messages|length - 1) - loop.index0 %}
+    {%- if ns.multi_step_tool and message.role == "user" and render_content(message.content) is string and not(render_content(message.content).startswith('<tool_response>') and render_content(message.content).endswith('</tool_response>')) %}
+        {%- set ns.multi_step_tool = false %}
+        {%- set ns.last_query_index = index %}
+    {%- endif %}
+{%- endfor %}
+{%- for message in messages %}
+    {%- set content = render_content(message.content) %}
+    {%- if (message.role == "user") or (message.role == "system" and not loop.first) %}
+        {%- set role_name = 'observation' if (message.role == "system" and not loop.first and message.name == 'observation') else message.role %}
+        {{- '<|im_start|>' + role_name + '\n' + content + '<|im_end|>' + '\n' }}
+    {%- elif message.role == "assistant" %}
+        {%- if message.reasoning_content is string %}
+            {%- set reasoning_content = render_content(message.reasoning_content) %}
+        {%- else %}
+            {%- if '</think>' in content %}
+                {%- set reasoning_content = content.split('</think>')[0].rstrip('\n').split('<think>')[-1].lstrip('\n') %}
+                {%- set content = content.split('</think>')[-1].lstrip('\n') %}
+            {%- else %}
+                {%- set reasoning_content = '' %}
+            {%- endif %}
+        {%- endif %}
+        {%- if loop.index0 > ns.last_query_index %}
+            {{- '<|im_start|>' + message.role + '\n<think>\n' + reasoning_content + '\n</think>\n' + content }}
+        {%- else %}
+            {{- '<|im_start|>' + message.role + '\n' + content }}
+        {%- endif %}
+        {%- if message.tool_calls %}
+            {%- for tool_call in message.tool_calls %}
+                {%- if tool_call.function is defined %}
+                    {%- set tool_call = tool_call.function %}
+                {%- endif %}
+                {{- '<tool_call>\n<function=' + tool_call.name + '>\n' }}
+                {%- if tool_call.arguments is defined %}
+                    {%- set arguments = tool_call.arguments %}
+                    {%- for args_name, args_value in arguments|items %}
+                        {{- '<parameter=' + args_name + '>\n' }}
+                        {%- set args_value = args_value | tojson(ensure_ascii=False) | safe if args_value is mapping or (args_value is sequence and args_value is not string) else args_value | string %}
+                        {{- args_value }}
+                        {{- '\n</parameter>\n' }}
+                    {%- endfor %}
+                {%- endif %}
+                {{- '</function>\n</tool_call>' }}
+            {%- endfor %}
+        {%- endif %}
+        {{- '<|im_end|>\n' }}
+    {%- elif message.role == "tool" %}
+        {%- if loop.first or (messages[loop.index0 - 1].role != "tool") %}
+            {{- '<|im_start|>tool_response\n' }}
+        {%- endif %}
+        {{- '<tool_response>' }}
+        {{- content }}
+        {{- '</tool_response>' }}
+        {%- if loop.last or (messages[loop.index0 + 1].role != "tool") %}
+            {{- '<|im_end|>\n' }}
+        {%- endif %}
+    {%- endif %}
+{%- endfor %}
+{%- if add_generation_prompt %}
+    {{- '<|im_start|>assistant\n<think>\n' }}
+{%- endif %}

src/CMakeLists.txt

@@ -84,6 +84,7 @@ add_library(llama
             models/hunyuan-moe.cpp
             models/internlm2.cpp
             models/jais.cpp
+            models/jais2.cpp
             models/jamba.cpp
             models/kimi-linear.cpp
             models/lfm2.cpp
@@ -109,6 +110,7 @@ add_library(llama
             models/openai-moe-iswa.cpp
             models/openelm.cpp
             models/orion.cpp
+            models/paddleocr.cpp
             models/pangu-embedded.cpp
             models/phi2.cpp
             models/phi3.cpp

src/llama-arch.cpp

@@ -79,6 +79,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_T5,               "t5"               },
     { LLM_ARCH_T5ENCODER,        "t5encoder"        },
     { LLM_ARCH_JAIS,             "jais"             },
+    { LLM_ARCH_JAIS2,            "jais2"            },
     { LLM_ARCH_NEMOTRON,         "nemotron"         },
     { LLM_ARCH_NEMOTRON_H,       "nemotron_h"       },
     { LLM_ARCH_NEMOTRON_H_MOE,   "nemotron_h_moe"   },
@@ -120,6 +121,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_RND1,             "rnd1"             },
     { LLM_ARCH_PANGU_EMBED,      "pangu-embedded"   },
     { LLM_ARCH_MISTRAL3,         "mistral3"         },
+    { LLM_ARCH_PADDLEOCR,        "paddleocr"        },
     { LLM_ARCH_MIMO2,            "mimo2"            },
     { LLM_ARCH_STEP35,           "step35"           },
     { LLM_ARCH_LLAMA_EMBED,      "llama-embed"      },
@@ -367,6 +369,7 @@ static const std::map<llm_tensor, const char *> LLM_TENSOR_NAMES = {
     { LLM_TENSOR_TOKEN_TYPES,                            "token_types" },
     { LLM_TENSOR_CLS,                                    "cls" },
     { LLM_TENSOR_CLS_OUT,                                "cls.output" },
+    { LLM_TENSOR_CLS_NORM,                               "cls.norm" },
     { LLM_TENSOR_ENC_OUTPUT_NORM,                        "enc.output_norm" },
     { LLM_TENSOR_FFN_GATE_INP_SHEXP,                     "blk.%d.ffn_gate_inp_shexp" },
     { LLM_TENSOR_SSM_A_NOSCAN,                           "blk.%d.ssm_a" },
@@ -737,6 +740,7 @@ static std::set<llm_tensor> llm_get_tensor_names(llm_arch arch) {
         case LLM_ARCH_INTERNLM2:
         case LLM_ARCH_GRANITE:
         case LLM_ARCH_ERNIE4_5:
+        case LLM_ARCH_PADDLEOCR:
         case LLM_ARCH_SMOLLM3:
         case LLM_ARCH_DREAM:
         case LLM_ARCH_LLADA:
@@ -828,6 +832,7 @@ static std::set<llm_tensor> llm_get_tensor_names(llm_arch arch) {
                 LLM_TENSOR_FFN_NORM,
                 LLM_TENSOR_CLS,
                 LLM_TENSOR_CLS_OUT,
+                LLM_TENSOR_CLS_NORM,
             };
         case LLM_ARCH_JINA_BERT_V2:
             return {
@@ -1789,6 +1794,20 @@ static std::set<llm_tensor> llm_get_tensor_names(llm_arch arch) {
                 LLM_TENSOR_FFN_GATE,
                 LLM_TENSOR_FFN_DOWN,
             };
+        case LLM_ARCH_JAIS2:
+            return {
+                LLM_TENSOR_TOKEN_EMBD,
+                LLM_TENSOR_OUTPUT_NORM,
+                LLM_TENSOR_OUTPUT,
+                LLM_TENSOR_ATTN_NORM,
+                LLM_TENSOR_ATTN_Q,
+                LLM_TENSOR_ATTN_K,
+                LLM_TENSOR_ATTN_V,
+                LLM_TENSOR_ATTN_OUT,
+                LLM_TENSOR_FFN_NORM,
+                LLM_TENSOR_FFN_UP,
+                LLM_TENSOR_FFN_DOWN,
+            };
         case LLM_ARCH_NEMOTRON_H:
             return {
                 LLM_TENSOR_TOKEN_EMBD,
@@ -2518,6 +2537,7 @@ static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = {
     {LLM_TENSOR_OUTPUT,                     {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL_MAT}},
     {LLM_TENSOR_CLS,                        {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL_MAT}},
     {LLM_TENSOR_CLS_OUT,                    {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL_MAT}},
+    {LLM_TENSOR_CLS_NORM,                   {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL}},
     {LLM_TENSOR_DENSE_2_OUT,                {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL_MAT}}, // Dense layer output
     {LLM_TENSOR_DENSE_3_OUT,                {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL_MAT}}, // Dense layer output
     {LLM_TENSOR_OUTPUT_NORM,                {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL}},

src/llama-arch.h

@@ -83,6 +83,7 @@ enum llm_arch {
     LLM_ARCH_T5,
     LLM_ARCH_T5ENCODER,
     LLM_ARCH_JAIS,
+    LLM_ARCH_JAIS2,
     LLM_ARCH_NEMOTRON,
     LLM_ARCH_NEMOTRON_H,
     LLM_ARCH_NEMOTRON_H_MOE,
@@ -124,6 +125,7 @@ enum llm_arch {
     LLM_ARCH_RND1,
     LLM_ARCH_PANGU_EMBED,
     LLM_ARCH_MISTRAL3,
+    LLM_ARCH_PADDLEOCR,
     LLM_ARCH_MIMO2,
     LLM_ARCH_STEP35,
     LLM_ARCH_LLAMA_EMBED,
@@ -497,6 +499,7 @@ enum llm_tensor {
     LLM_TENSOR_ENC_OUTPUT_NORM,
     LLM_TENSOR_CLS,
     LLM_TENSOR_CLS_OUT,
+    LLM_TENSOR_CLS_NORM,
     LLM_TENSOR_CONV1D,
     LLM_TENSOR_CONVNEXT_DW,
     LLM_TENSOR_CONVNEXT_NORM,

src/llama-context.cpp

@@ -712,8 +712,6 @@ int64_t llama_context::output_resolve_row(int32_t i) const {
 }
 
 float * llama_context::get_logits_ith(int32_t i) {
-    int64_t j = -1;
-
     output_reorder();
 
     try {
@@ -721,26 +719,7 @@ float * llama_context::get_logits_ith(int32_t i) {
             throw std::runtime_error("no logits");
         }
 
-        // TODO: use output_resolve_row()
-        if (i < 0) {
-            j = n_outputs + i;
-            if (j < 0) {
-                throw std::runtime_error(format("negative index out of range [0, %d)", n_outputs));
-            }
-        } else if ((size_t) i >= output_ids.size()) {
-            throw std::runtime_error(format("out of range [0, %zu)", output_ids.size()));
-        } else {
-            j = output_ids[i];
-        }
-
-        if (j < 0) {
-            throw std::runtime_error(format("batch.logits[%d] != true", i));
-        }
-        if (j >= n_outputs) {
-            // This should not happen
-            throw std::runtime_error(format("corrupt output buffer (j=%" PRId64 ", n_outputs=%d)", j, n_outputs));
-        }
-
+        const int64_t j = output_resolve_row(i);
         return logits.data + j*model.vocab.n_tokens();
     } catch (const std::exception & err) {
         LLAMA_LOG_ERROR("%s: invalid logits id %d, reason: %s\n", __func__, i, err.what());
@@ -763,8 +742,6 @@ llama_token * llama_context::get_sampled_tokens()  const{
 }
 
 float * llama_context::get_embeddings_ith(int32_t i) {
-    int64_t j = -1;
-
     output_reorder();
 
     try {
@@ -772,26 +749,7 @@ float * llama_context::get_embeddings_ith(int32_t i) {
             throw std::runtime_error("no embeddings");
         }
 
-        // TODO: use output_resolve_row()
-        if (i < 0) {
-            j = n_outputs + i;
-            if (j < 0) {
-                throw std::runtime_error(format("negative index out of range [0, %d)", n_outputs));
-            }
-        } else if ((size_t) i >= output_ids.size()) {
-            throw std::runtime_error(format("out of range [0, %zu)", output_ids.size()));
-        } else {
-            j = output_ids[i];
-        }
-
-        if (j < 0) {
-            throw std::runtime_error(format("batch.logits[%d] != true", i));
-        }
-        if (j >= n_outputs) {
-            // This should not happen
-            throw std::runtime_error(format("corrupt output buffer (j=%" PRId64 ", n_outputs=%d)", j, n_outputs));
-        }
-
+        const int64_t j = output_resolve_row(i);
         const uint32_t n_embd_out = model.hparams.n_embd_out();
         return embd.data + j*n_embd_out;
     } catch (const std::exception & err) {
@@ -2761,6 +2719,7 @@ void llama_context::opt_init(struct llama_model * model, struct llama_opt_params
     llama_set_param(model->cls_b,           param_filter, param_filter_ud);
     llama_set_param(model->cls_out,         param_filter, param_filter_ud);
     llama_set_param(model->cls_out_b,       param_filter, param_filter_ud);
+    llama_set_param(model->cls_norm,        param_filter, param_filter_ud);
 
     for (struct llama_layer & layer : model->layers) {
         for (size_t i = 0; i < sizeof(layer)/sizeof(struct ggml_tensor *); ++i) {

src/llama-graph.cpp

@@ -185,7 +185,10 @@ bool llm_graph_input_out_ids::can_reuse(const llm_graph_params & params) {
 }
 
 void llm_graph_input_mean::set_input(const llama_ubatch * ubatch) {
-    if (cparams.embeddings && cparams.pooling_type == LLAMA_POOLING_TYPE_MEAN) {
+    if (cparams.embeddings   &&
+       (cparams.pooling_type == LLAMA_POOLING_TYPE_MEAN ||
+        cparams.pooling_type == LLAMA_POOLING_TYPE_RANK )) {
+
         const int64_t n_tokens     = ubatch->n_tokens;
         const int64_t n_seq_tokens = ubatch->n_seq_tokens;
         const int64_t n_seqs_unq   = ubatch->n_seqs_unq;
@@ -1125,8 +1128,8 @@ ggml_tensor * llm_graph_context::build_ffn(
 
     if (down) {
         cur = build_lora_mm(down, cur);
-        if (arch == LLM_ARCH_GLM4 || arch == LLM_ARCH_GLM4_MOE) {
-            // GLM4 and GLM4_MOE seem to have numerical issues with half-precision accumulators
+        if (arch == LLM_ARCH_GLM4 || arch == LLM_ARCH_GLM4_MOE || arch == LLM_ARCH_JAIS2) {
+            // GLM4, GLM4_MOE, and JAIS2 seem to have numerical issues with half-precision accumulators
             ggml_mul_mat_set_prec(cur, GGML_PREC_F32);
         }
     }
@@ -1721,7 +1724,8 @@ ggml_tensor * llm_graph_context::build_attn_mha(
 
     ggml_tensor * cur;
 
-    if (cparams.flash_attn && kq_b == nullptr) {
+    const bool use_flash_attn = cparams.flash_attn && kq_b == nullptr;
+    if (use_flash_attn) {
         GGML_ASSERT(kq_b == nullptr && "Flash attention does not support KQ bias yet");
 
         if (v_trans) {
@@ -1981,8 +1985,8 @@ ggml_tensor * llm_graph_context::build_attn(
 
     if (wo) {
         cur = build_lora_mm(wo, cur);
-        if (arch == LLM_ARCH_GLM4 || arch == LLM_ARCH_GLM4_MOE) {
-            // GLM4 and GLM4_MOE seem to have numerical issues with half-precision accumulators
+        if (arch == LLM_ARCH_GLM4 || arch == LLM_ARCH_GLM4_MOE || arch == LLM_ARCH_JAIS2) {
+            // GLM4, GLM4_MOE, and JAIS2 seem to have numerical issues with half-precision accumulators
             ggml_mul_mat_set_prec(cur, GGML_PREC_F32);
         }
     }
@@ -2414,8 +2418,9 @@ llm_graph_input_mem_hybrid_iswa * llm_graph_context::build_inp_mem_hybrid_iswa()
 
 void llm_graph_context::build_dense_out(
     ggml_tensor * dense_2,
+    ggml_tensor * dense_2_b,
     ggml_tensor * dense_3) const {
-    if (!cparams.embeddings || !(dense_2 || dense_3)) {
+    if (!cparams.embeddings || !(dense_2 || dense_2_b || dense_3)) {
         return;
     }
     ggml_tensor * cur = res->t_embd_pooled != nullptr ? res->t_embd_pooled : res->t_embd;
@@ -2424,6 +2429,9 @@ void llm_graph_context::build_dense_out(
     if (dense_2) {
         cur = ggml_mul_mat(ctx0, dense_2, cur);
     }
+    if (dense_2_b) {
+        cur = ggml_add(ctx0, cur, dense_2_b);
+    }
     if (dense_3) {
         cur = ggml_mul_mat(ctx0, dense_3, cur);
     }
@@ -2437,7 +2445,8 @@ void llm_graph_context::build_pooling(
         ggml_tensor * cls,
         ggml_tensor * cls_b,
         ggml_tensor * cls_out,
-        ggml_tensor * cls_out_b) const {
+        ggml_tensor * cls_out_b,
+        ggml_tensor * cls_norm) const {
     if (!cparams.embeddings) {
         return;
     }
@@ -2476,8 +2485,15 @@ void llm_graph_context::build_pooling(
             } break;
         case LLAMA_POOLING_TYPE_RANK:
             {
-                ggml_tensor * inp_cls = build_inp_cls();
-                cur = ggml_get_rows(ctx0, inp, inp_cls);
+                if (arch == LLM_ARCH_MODERN_BERT) {
+                    // modern bert gte reranker builds mean first then applies prediction head and classifier
+                    // https://github.com/huggingface/transformers/blob/main/src/transformers/models/modernbert/modular_modernbert.py#L1404-1411
+                    ggml_tensor * inp_mean = build_inp_mean();
+                    cur = ggml_mul_mat(ctx0, ggml_cont(ctx0, ggml_transpose(ctx0, inp)), inp_mean);
+                } else {
+                    ggml_tensor * inp_cls = build_inp_cls();
+                    cur = ggml_get_rows(ctx0, inp, inp_cls);
+                }
 
                 // classification head
                 // https://github.com/huggingface/transformers/blob/5af7d41e49bbfc8319f462eb45253dcb3863dfb7/src/transformers/models/roberta/modeling_roberta.py#L1566
@@ -2486,7 +2502,15 @@ void llm_graph_context::build_pooling(
                     if (cls_b) {
                         cur = ggml_add(ctx0, cur, cls_b);
                     }
-                    cur = ggml_tanh(ctx0, cur);
+                    if (arch == LLM_ARCH_MODERN_BERT) {
+                        cur = ggml_gelu(ctx0, cur);
+                    } else {
+                        cur = ggml_tanh(ctx0, cur);
+                    }
+                    if (cls_norm) {
+                        // head norm
+                        cur = build_norm(cur, cls_norm, NULL, LLM_NORM, -1);
+                    }
                 }
 
                 // some models don't have `cls_out`, for example: https://huggingface.co/jinaai/jina-reranker-v1-tiny-en

src/llama-graph.h

@@ -1000,7 +1000,8 @@ struct llm_graph_context {
             ggml_tensor * cls,
             ggml_tensor * cls_b,
             ggml_tensor * cls_out,
-            ggml_tensor * cls_out_b) const;
+            ggml_tensor * cls_out_b,
+            ggml_tensor * cls_norm) const;
 
     //
     // sampling (backend sampling)
@@ -1014,6 +1015,7 @@ struct llm_graph_context {
 
     void build_dense_out(
             ggml_tensor * dense_2,
+            ggml_tensor * dense_2_b,
             ggml_tensor * dense_3) const;
 };
 

src/llama-impl.cpp

@@ -109,9 +109,9 @@ std::string llama_format_tensor_shape(const std::vector<int64_t> & ne) {
 
 std::string llama_format_tensor_shape(const struct ggml_tensor * t) {
     char buf[256];
-    snprintf(buf, sizeof(buf), "%5" PRId64, t->ne[0]);
+    snprintf(buf, sizeof(buf), "%6" PRId64, t->ne[0]);
     for (int i = 1; i < GGML_MAX_DIMS; i++) {
-        snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf), ", %5" PRId64, t->ne[i]);
+        snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf), ", %6" PRId64, t->ne[i]);
     }
     return buf;
 }

src/llama-model-saver.cpp

@@ -271,6 +271,7 @@ void llama_model_saver::add_tensors_from_model() {
     add_tensor(model.cls_b);
     add_tensor(model.cls_out);
     add_tensor(model.cls_out_b);
+    add_tensor(model.cls_norm);
 
     for (const struct llama_layer & layer : model.layers) {
         for (size_t i = 0; i < sizeof(layer)/sizeof(struct ggml_tensor *); ++i) {

src/llama-model.cpp

@@ -908,7 +908,7 @@ void llama_model::load_hparams(llama_model_loader & ml) {
 
                     ml.get_key(LLM_KV_ROPE_FREQ_BASE_SWA, hparams.rope_freq_base_train_swa);
                     ml.get_key_or_arr(LLM_KV_ATTENTION_SLIDING_WINDOW_PATTERN, swa_period, false);
-                    hparams.set_swa_pattern(swa_period);
+                    hparams.set_swa_pattern(swa_period, true);
                 } else {
                     hparams.swa_type = LLAMA_SWA_TYPE_NONE;
                 }
@@ -1937,6 +1937,16 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                     default: type = LLM_TYPE_UNKNOWN;
                 }
             } break;
+        case LLM_ARCH_JAIS2:
+            {
+                ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps);
+
+                switch (hparams.n_layer) {
+                    case 32: type = LLM_TYPE_8B; break;
+                    case 68: type = LLM_TYPE_70B; break;
+                    default: type = LLM_TYPE_UNKNOWN;
+                }
+            } break;
         case LLM_ARCH_NEMOTRON:
             {
                 ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps);
@@ -2234,7 +2244,11 @@ void llama_model::load_hparams(llama_model_loader & ml) {
             } break;
         case LLM_ARCH_ERNIE4_5:
         case LLM_ARCH_ERNIE4_5_MOE:
+        case LLM_ARCH_PADDLEOCR:
             {
+                // paddleocr need mrope_section
+                ml.get_key_or_arr(LLM_KV_ROPE_DIMENSION_SECTIONS, hparams.rope_sections, 4, false);
+
                 ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
                 if (arch == LLM_ARCH_ERNIE4_5_MOE) {
                     ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH,        hparams.n_ff_exp);
@@ -2348,6 +2362,12 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                     case 10752: type = LLM_TYPE_2_6B; break;
                     default:    type = LLM_TYPE_UNKNOWN;
                 }
+                if (const auto is_swa = ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW, hparams.n_swa, false); is_swa && hparams.n_swa > 0) {
+                    hparams.swa_type = LLAMA_SWA_TYPE_STANDARD;
+                    for (uint32_t il = 0; il < hparams.n_layer; ++il) {
+                        hparams.swa_layers[il] = !hparams.recurrent_layer_arr[il];
+                    }
+                }
             } break;
         case LLM_ARCH_LFM2MOE:
             {
@@ -3513,9 +3533,10 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                         layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
                     }
 
-                    cls       = create_tensor(tn(LLM_TENSOR_CLS,     "weight"), {n_embd, n_embd}, TENSOR_NOT_REQUIRED);
-                    cls_out   = create_tensor(tn(LLM_TENSOR_CLS_OUT, "weight"), {n_embd, hparams.n_cls_out}, TENSOR_NOT_REQUIRED);
-                    cls_out_b = create_tensor(tn(LLM_TENSOR_CLS_OUT, "bias"),   {hparams.n_cls_out},         TENSOR_NOT_REQUIRED);
+                    cls_out   = create_tensor(tn(LLM_TENSOR_CLS_OUT,  "weight"), {n_embd, hparams.n_cls_out}, TENSOR_NOT_REQUIRED);
+                    cls_out_b = create_tensor(tn(LLM_TENSOR_CLS_OUT,  "bias"),   {hparams.n_cls_out},         TENSOR_NOT_REQUIRED);
+                    cls       = create_tensor(tn(LLM_TENSOR_CLS,      "weight"), {n_embd, n_embd},            TENSOR_NOT_REQUIRED);
+                    cls_norm  = create_tensor(tn(LLM_TENSOR_CLS_NORM, "weight"), {n_embd},                    TENSOR_NOT_REQUIRED);
 
                 } break;
             case LLM_ARCH_NEO_BERT:
@@ -5368,6 +5389,45 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                         layer.ffn_up_b   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "bias", i),   {n_ff}, 0);
                     }
                 } break;
+            case LLM_ARCH_JAIS2:
+                {
+                    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
+
+                    // output
+                    output_norm   = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
+                    output_norm_b = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "bias"),   {n_embd}, 0);
+                    output        = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
+                    if (!output) {
+                        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
+                    }
+
+                    for (int i = 0; i < n_layer; ++i) {
+                        auto & layer = layers[i];
+
+                        layer.attn_norm   = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
+                        layer.attn_norm_b = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "bias", i),   {n_embd}, 0);
+
+                        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);
+                        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd}, 0);
+
+                        // attention biases - all have shape n_embd (output dimension of projections)
+                        layer.bq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "bias", i), {n_embd}, 0);
+                        layer.bk = create_tensor(tn(LLM_TENSOR_ATTN_K, "bias", i), {n_embd}, 0);
+                        layer.bv = create_tensor(tn(LLM_TENSOR_ATTN_V, "bias", i), {n_embd}, 0);
+                        layer.bo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}, 0);
+
+                        layer.ffn_norm   = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
+                        layer.ffn_norm_b = create_tensor(tn(LLM_TENSOR_FFN_NORM, "bias", i),   {n_embd}, 0);
+
+                        // Jais-2 uses simple MLP (no gate) with biases
+                        layer.ffn_up     = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd, n_ff}, 0);
+                        layer.ffn_up_b   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "bias", i),   {n_ff}, 0);
+                        layer.ffn_down   = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}, 0);
+                        layer.ffn_down_b = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "bias", i),   {n_embd}, 0);
+                    }
+                } break;
             case LLM_ARCH_CHATGLM:
                 {
                     tok_embd   = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD,      "weight"), {n_embd, n_vocab}, 0);
@@ -6575,6 +6635,7 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                 } break;
             case LLM_ARCH_ERNIE4_5:
             case LLM_ARCH_ERNIE4_5_MOE:
+            case LLM_ARCH_PADDLEOCR:
                 {
                     tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
 
@@ -6895,7 +6956,8 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                     }
 
                     // for LFM2-ColBert-350M
-                    dense_2_out_layers = create_tensor(tn(LLM_TENSOR_DENSE_2_OUT, "weight"), {n_embd, hparams.n_embd_out()}, TENSOR_NOT_REQUIRED);
+                    dense_2_out_layers   = create_tensor(tn(LLM_TENSOR_DENSE_2_OUT, "weight"), {n_embd, hparams.n_embd_out()}, TENSOR_NOT_REQUIRED);
+                    dense_2_out_layers_b = create_tensor(tn(LLM_TENSOR_DENSE_2_OUT, "bias"),   {hparams.n_embd_out()        }, TENSOR_NOT_REQUIRED);
                 } break;
             case LLM_ARCH_SMALLTHINKER:
                 {
@@ -8553,6 +8615,10 @@ ggml_cgraph * llama_model::build_graph(const llm_graph_params & params) const {
             {
                 llm = std::make_unique<llm_build_jais>(*this, params);
             } break;
+        case LLM_ARCH_JAIS2:
+            {
+                llm = std::make_unique<llm_build_jais2>(*this, params);
+            } break;
         case LLM_ARCH_NEMOTRON:
             {
                 llm = std::make_unique<llm_build_nemotron>(*this, params);
@@ -8648,6 +8714,10 @@ ggml_cgraph * llama_model::build_graph(const llm_graph_params & params) const {
             {
                 llm = std::make_unique<llm_build_ernie4_5_moe>(*this, params);
             } break;
+        case LLM_ARCH_PADDLEOCR:
+            {
+                llm = std::make_unique<llm_build_paddleocr>(*this, params);
+            } break;
         case LLM_ARCH_HUNYUAN_MOE:
             {
                 llm = std::make_unique<llm_build_hunyuan_moe>(*this, params);
@@ -8671,7 +8741,11 @@ ggml_cgraph * llama_model::build_graph(const llm_graph_params & params) const {
         case LLM_ARCH_LFM2:
         case LLM_ARCH_LFM2MOE:
             {
-                llm = std::make_unique<llm_build_lfm2>(*this, params);
+                if (hparams.swa_type == LLAMA_SWA_TYPE_STANDARD) {
+                    llm = std::make_unique<llm_build_lfm2<true>>(*this, params);
+                } else {
+                    llm = std::make_unique<llm_build_lfm2<false>>(*this, params);
+                }
             } break;
         case LLM_ARCH_SMALLTHINKER:
             {
@@ -8734,7 +8808,7 @@ ggml_cgraph * llama_model::build_graph(const llm_graph_params & params) const {
     }
 
     // add on pooling layer
-    llm->build_pooling(cls, cls_b, cls_out, cls_out_b);
+    llm->build_pooling(cls, cls_b, cls_out, cls_out_b, cls_norm);
 
     // add backend sampling layers (if any)
     llm->build_sampling();
@@ -8743,7 +8817,7 @@ ggml_cgraph * llama_model::build_graph(const llm_graph_params & params) const {
     // there will be two additional dense projection layers
     // dense linear projections are applied after pooling
     // TODO: move reranking logic here and generalize
-    llm->build_dense_out(dense_2_out_layers, dense_3_out_layers);
+    llm->build_dense_out(dense_2_out_layers, dense_2_out_layers_b, dense_3_out_layers);
 
     llm->res->set_outputs();
 
@@ -8961,6 +9035,7 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
         case LLM_ARCH_BAILINGMOE2:
         case LLM_ARCH_DOTS1:
         case LLM_ARCH_HUNYUAN_MOE:
+        case LLM_ARCH_JAIS2:
         case LLM_ARCH_OPENAI_MOE:
         case LLM_ARCH_HUNYUAN_DENSE:
         case LLM_ARCH_LFM2:
@@ -8979,6 +9054,7 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
             return LLAMA_ROPE_TYPE_NEOX;
 
         case LLM_ARCH_QWEN2VL:
+        case LLM_ARCH_PADDLEOCR:
             return LLAMA_ROPE_TYPE_MROPE;
         case LLM_ARCH_QWEN3VL:
         case LLM_ARCH_QWEN3VLMOE:

src/llama-model.h

@@ -475,6 +475,7 @@ struct llama_model {
     struct ggml_tensor * cls_b     = nullptr;
     struct ggml_tensor * cls_out   = nullptr;
     struct ggml_tensor * cls_out_b = nullptr;
+    struct ggml_tensor * cls_norm  = nullptr;
 
     struct ggml_tensor * conv1d   = nullptr;
     struct ggml_tensor * conv1d_b = nullptr;
@@ -491,8 +492,9 @@ struct llama_model {
     //Dense linear projections for SentenceTransformers models like embeddinggemma
     // For Sentence Transformers models structure see
     // https://sbert.net/docs/sentence_transformer/usage/custom_models.html#structure-of-sentence-transformer-models
-    struct ggml_tensor * dense_2_out_layers = nullptr;
-    struct ggml_tensor * dense_3_out_layers = nullptr;
+    struct ggml_tensor * dense_2_out_layers   = nullptr;
+    struct ggml_tensor * dense_2_out_layers_b = nullptr;
+    struct ggml_tensor * dense_3_out_layers   = nullptr;
 
     // gguf metadata
     std::unordered_map<std::string, std::string> gguf_kv;

src/llama-quant.cpp

@@ -479,6 +479,17 @@ static size_t llama_tensor_quantize_impl(enum ggml_type new_type, const float *
     return new_size;
 }
 
+static bool tensor_type_requires_imatrix(const ggml_tensor * t, const ggml_type dst_type, const llama_ftype ftype) {
+    return (
+        dst_type == GGML_TYPE_IQ2_XXS || dst_type == GGML_TYPE_IQ2_XS ||
+        dst_type == GGML_TYPE_IQ3_XXS || dst_type == GGML_TYPE_IQ1_S  ||
+        dst_type == GGML_TYPE_IQ2_S   || dst_type == GGML_TYPE_IQ1_M  ||
+        (   // Q2_K_S is the worst k-quant type - only allow it without imatrix for token embeddings
+            dst_type == GGML_TYPE_Q2_K && ftype == LLAMA_FTYPE_MOSTLY_Q2_K_S && strcmp(t->name, "token_embd.weight") != 0
+        )
+    );
+}
+
 static void llama_model_quantize_impl(const std::string & fname_inp, const std::string & fname_out, const llama_model_quantize_params * params) {
     ggml_type default_type;
     llama_ftype ftype = params->ftype;
@@ -735,24 +746,36 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
     };
 
     const auto tn = LLM_TN(model.arch);
-    new_ofstream(0);
+
+    // no output file for --dry-run
+    if (!params->dry_run) {
+        new_ofstream(0);
+    }
+
+    // flag for `--dry-run`, to let the user know if imatrix will be required for a real
+    // quantization, as a courtesy
+    bool will_require_imatrix = false;
+
     for (const auto * it : tensors) {
         const auto & weight = *it;
         ggml_tensor * tensor = weight.tensor;
-        if (weight.idx != cur_split && params->keep_split) {
+        if (!params->dry_run && (weight.idx != cur_split && params->keep_split)) {
             close_ofstream();
             new_ofstream(weight.idx);
         }
 
         const std::string name = ggml_get_name(tensor);
+        const size_t tensor_size = ggml_nbytes(tensor);
 
-        if (!ml.use_mmap) {
-            if (read_data.size() < ggml_nbytes(tensor)) {
-                read_data.resize(ggml_nbytes(tensor));
+        if (!params->dry_run) {
+            if (!ml.use_mmap) {
+                if (read_data.size() < tensor_size) {
+                    read_data.resize(tensor_size);
+                }
+                tensor->data = read_data.data();
             }
-            tensor->data = read_data.data();
+            ml.load_data_for(tensor);
         }
-        ml.load_data_for(tensor);
 
         LLAMA_LOG_INFO("[%4d/%4d] %36s - [%s], type = %6s, ",
                ++idx, ml.n_tensors,
@@ -900,129 +923,155 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
             quantize = tensor->type != new_type;
         }
 
-        if (!quantize) {
-            new_type = tensor->type;
-            new_data = tensor->data;
-            new_size = ggml_nbytes(tensor);
-            LLAMA_LOG_INFO("size = %8.3f MiB\n", ggml_nbytes(tensor)/1024.0/1024.0);
-        } else {
-            const int64_t nelements = ggml_nelements(tensor);
-
-            const float * imatrix = nullptr;
-            if (imatrix_data) {
-                auto it = imatrix_data->find(remap_imatrix(tensor->name, mapped));
-                if (it == imatrix_data->end()) {
-                    LLAMA_LOG_INFO("\n====== %s: did not find weights for %s\n", __func__, tensor->name);
-                } else {
-                    if (it->second.size() == (size_t)tensor->ne[0]*tensor->ne[2]) {
-                        imatrix = it->second.data();
-                    } else {
-                        LLAMA_LOG_INFO("\n====== %s: imatrix size %d is different from tensor size %d for %s\n", __func__,
-                                int(it->second.size()), int(tensor->ne[0]*tensor->ne[2]), tensor->name);
-
-                        // this can happen when quantizing an old mixtral model with split tensors with a new incompatible imatrix
-                        // this is a significant error and it may be good idea to abort the process if this happens,
-                        // since many people will miss the error and not realize that most of the model is being quantized without an imatrix
-                        // tok_embd should be ignored in this case, since it always causes this warning
-                        if (name != tn(LLM_TENSOR_TOKEN_EMBD, "weight")) {
-                            throw std::runtime_error(format("imatrix size %d is different from tensor size %d for %s",
-                                    int(it->second.size()), int(tensor->ne[0]*tensor->ne[2]), tensor->name));
-                        }
-                    }
+        // we have now decided on the target type for this tensor
+        if (params->dry_run) {
+            // the --dry-run option calculates the final quantization size without quantizting
+            if (quantize) {
+                new_size = ggml_nrows(tensor) * ggml_row_size(new_type, tensor->ne[0]);
+                LLAMA_LOG_INFO("size = %8.2f MiB -> %8.2f MiB (%s)\n",
+                               tensor_size/1024.0/1024.0,
+                               new_size/1024.0/1024.0,
+                               ggml_type_name(new_type));
+                if (!will_require_imatrix && tensor_type_requires_imatrix(tensor, new_type, params->ftype)) {
+                    will_require_imatrix = true;
                 }
-            }
-            if ((new_type == GGML_TYPE_IQ2_XXS ||
-                 new_type == GGML_TYPE_IQ2_XS  ||
-                 new_type == GGML_TYPE_IQ2_S   ||
-                 new_type == GGML_TYPE_IQ1_S   ||
-                (new_type == GGML_TYPE_IQ1_M && strcmp(tensor->name, "token_embd.weight") && strcmp(tensor->name, "output.weight"))  ||
-                (new_type == GGML_TYPE_Q2_K && params->ftype == LLAMA_FTYPE_MOSTLY_Q2_K_S && strcmp(tensor->name, "token_embd.weight") != 0)) && !imatrix) {
-                LLAMA_LOG_ERROR("\n\n============================================================\n");
-                LLAMA_LOG_ERROR("Missing importance matrix for tensor %s in a very low-bit quantization\n", tensor->name);
-                LLAMA_LOG_ERROR("The result will be garbage, so bailing out\n");
-                LLAMA_LOG_ERROR("============================================================\n\n");
-                throw std::runtime_error(format("Missing importance matrix for tensor %s in a very low-bit quantization", tensor->name));
-            }
-
-            float * f32_data;
-
-            if (tensor->type == GGML_TYPE_F32) {
-                f32_data = (float *) tensor->data;
-            } else if (ggml_is_quantized(tensor->type) && !params->allow_requantize) {
-                throw std::runtime_error(format("requantizing from type %s is disabled", ggml_type_name(tensor->type)));
             } else {
-                llama_tensor_dequantize_impl(tensor, f32_conv_buf, workers, nelements, nthread);
-                f32_data = (float *) f32_conv_buf.data();
+                new_size = tensor_size;
+                LLAMA_LOG_INFO("size = %8.3f MiB\n", new_size/1024.0/1024.0);
             }
+            total_size_org += tensor_size;
+            total_size_new += new_size;
+            continue;
+        } else {
+            // no --dry-run, perform quantization
+            if (!quantize) {
+                new_type = tensor->type;
+                new_data = tensor->data;
+                new_size = tensor_size;
+                LLAMA_LOG_INFO("size = %8.3f MiB\n", tensor_size/1024.0/1024.0);
+            } else {
+                const int64_t nelements = ggml_nelements(tensor);
 
-            LLAMA_LOG_INFO("converting to %s .. ", ggml_type_name(new_type));
-            fflush(stdout);
+                const float * imatrix = nullptr;
+                if (imatrix_data) {
+                    auto it = imatrix_data->find(remap_imatrix(tensor->name, mapped));
+                    if (it == imatrix_data->end()) {
+                        LLAMA_LOG_INFO("\n====== %s: did not find weights for %s\n", __func__, tensor->name);
+                    } else {
+                        if (it->second.size() == (size_t)tensor->ne[0]*tensor->ne[2]) {
+                            imatrix = it->second.data();
+                        } else {
+                            LLAMA_LOG_INFO("\n====== %s: imatrix size %d is different from tensor size %d for %s\n", __func__,
+                                    int(it->second.size()), int(tensor->ne[0]*tensor->ne[2]), tensor->name);
 
-            if (work.size() < (size_t)nelements * 4) {
-                work.resize(nelements * 4); // upper bound on size
-            }
-            new_data = work.data();
-
-            const int64_t n_per_row = tensor->ne[0];
-            const int64_t nrows = tensor->ne[1];
-
-            static const int64_t min_chunk_size = 32 * 512;
-            const int64_t chunk_size = (n_per_row >= min_chunk_size ? n_per_row : n_per_row * ((min_chunk_size + n_per_row - 1)/n_per_row));
-
-            const int64_t nelements_matrix = tensor->ne[0] * tensor->ne[1];
-            const int64_t nchunk = (nelements_matrix + chunk_size - 1)/chunk_size;
-            const int64_t nthread_use = nthread > 1 ? std::max((int64_t)1, std::min((int64_t)nthread, nchunk)) : 1;
-
-            // quantize each expert separately since they have different importance matrices
-            new_size = 0;
-            for (int64_t i03 = 0; i03 < tensor->ne[2]; ++i03) {
-                const float * f32_data_03 = f32_data + i03 * nelements_matrix;
-                void * new_data_03 = (char *)new_data + ggml_row_size(new_type, n_per_row) * i03 * nrows;
-                const float * imatrix_03 = imatrix ? imatrix + i03 * n_per_row : nullptr;
-
-                new_size += llama_tensor_quantize_impl(new_type, f32_data_03, new_data_03, chunk_size, nrows, n_per_row, imatrix_03, workers, nthread_use);
-
-                // TODO: temporary sanity check that the F16 -> MXFP4 is lossless
-#if 0
-                if (new_type == GGML_TYPE_MXFP4) {
-                    auto * x = f32_data_03;
-
-                    //LLAMA_LOG_INFO("nrows = %d, n_per_row = %d\n", nrows, n_per_row);
-                    std::vector<float> deq(nrows*n_per_row);
-                    const ggml_type_traits * qtype = ggml_get_type_traits(new_type);
-                    qtype->to_float(new_data_03, deq.data(), deq.size());
-
-                    double err = 0.0f;
-                    for (int i = 0; i < (int) deq.size(); ++i) {
-                        err += fabsf(deq[i] - x[i]);
-                        //if (fabsf(deq[i] - x[i]) > 0.00001 && i < 256) {
-                        if (deq[i] != x[i]) {
-                            LLAMA_LOG_INFO("deq[%d] = %f, x[%d] = %f\n", i, deq[i], i, x[i]);
+                            // this can happen when quantizing an old mixtral model with split tensors with a new incompatible imatrix
+                            // this is a significant error and it may be good idea to abort the process if this happens,
+                            // since many people will miss the error and not realize that most of the model is being quantized without an imatrix
+                            // tok_embd should be ignored in this case, since it always causes this warning
+                            if (name != tn(LLM_TENSOR_TOKEN_EMBD, "weight")) {
+                                throw std::runtime_error(format("imatrix size %d is different from tensor size %d for %s",
+                                        int(it->second.size()), int(tensor->ne[0]*tensor->ne[2]), tensor->name));
+                            }
                         }
                     }
-                    //LLAMA_LOG_INFO("err = %f\n", err);
-                    GGML_ASSERT(err == 0.00000);
                 }
+                if (!imatrix && tensor_type_requires_imatrix(tensor, new_type, params->ftype)) {
+                    LLAMA_LOG_ERROR("\n\n============================================================\n");
+                    LLAMA_LOG_ERROR("Missing importance matrix for tensor %s in a very low-bit quantization\n", tensor->name);
+                    LLAMA_LOG_ERROR("The result will be garbage, so bailing out\n");
+                    LLAMA_LOG_ERROR("============================================================\n\n");
+                    throw std::runtime_error(format("Missing importance matrix for tensor %s in a very low-bit quantization", tensor->name));
+                }
+
+                float * f32_data;
+
+                if (tensor->type == GGML_TYPE_F32) {
+                    f32_data = (float *) tensor->data;
+                } else if (ggml_is_quantized(tensor->type) && !params->allow_requantize) {
+                    throw std::runtime_error(format("requantizing from type %s is disabled", ggml_type_name(tensor->type)));
+                } else {
+                    llama_tensor_dequantize_impl(tensor, f32_conv_buf, workers, nelements, nthread);
+                    f32_data = (float *) f32_conv_buf.data();
+                }
+
+                LLAMA_LOG_INFO("converting to %s .. ", ggml_type_name(new_type));
+                fflush(stdout);
+
+                if (work.size() < (size_t)nelements * 4) {
+                    work.resize(nelements * 4); // upper bound on size
+                }
+                new_data = work.data();
+
+                const int64_t n_per_row = tensor->ne[0];
+                const int64_t nrows = tensor->ne[1];
+
+                static const int64_t min_chunk_size = 32 * 512;
+                const int64_t chunk_size = (n_per_row >= min_chunk_size ? n_per_row : n_per_row * ((min_chunk_size + n_per_row - 1)/n_per_row));
+
+                const int64_t nelements_matrix = tensor->ne[0] * tensor->ne[1];
+                const int64_t nchunk = (nelements_matrix + chunk_size - 1)/chunk_size;
+                const int64_t nthread_use = nthread > 1 ? std::max((int64_t)1, std::min((int64_t)nthread, nchunk)) : 1;
+
+                // quantize each expert separately since they have different importance matrices
+                new_size = 0;
+                for (int64_t i03 = 0; i03 < tensor->ne[2]; ++i03) {
+                    const float * f32_data_03 = f32_data + i03 * nelements_matrix;
+                    void * new_data_03 = (char *)new_data + ggml_row_size(new_type, n_per_row) * i03 * nrows;
+                    const float * imatrix_03 = imatrix ? imatrix + i03 * n_per_row : nullptr;
+
+                    new_size += llama_tensor_quantize_impl(new_type, f32_data_03, new_data_03, chunk_size, nrows, n_per_row, imatrix_03, workers, nthread_use);
+
+                    // TODO: temporary sanity check that the F16 -> MXFP4 is lossless
+#if 0
+                    if (new_type == GGML_TYPE_MXFP4) {
+                        auto * x = f32_data_03;
+
+                        //LLAMA_LOG_INFO("nrows = %d, n_per_row = %d\n", nrows, n_per_row);
+                        std::vector<float> deq(nrows*n_per_row);
+                        const ggml_type_traits * qtype = ggml_get_type_traits(new_type);
+                        qtype->to_float(new_data_03, deq.data(), deq.size());
+
+                        double err = 0.0f;
+                        for (int i = 0; i < (int) deq.size(); ++i) {
+                            err += fabsf(deq[i] - x[i]);
+                            //if (fabsf(deq[i] - x[i]) > 0.00001 && i < 256) {
+                            if (deq[i] != x[i]) {
+                                LLAMA_LOG_INFO("deq[%d] = %f, x[%d] = %f\n", i, deq[i], i, x[i]);
+                            }
+                        }
+                        //LLAMA_LOG_INFO("err = %f\n", err);
+                        GGML_ASSERT(err == 0.00000);
+                    }
 #endif
+                }
+                LLAMA_LOG_INFO("size = %8.2f MiB -> %8.2f MiB\n", tensor_size/1024.0/1024.0, new_size/1024.0/1024.0);
             }
-            LLAMA_LOG_INFO("size = %8.2f MiB -> %8.2f MiB\n", ggml_nbytes(tensor)/1024.0/1024.0, new_size/1024.0/1024.0);
-        }
-        total_size_org += ggml_nbytes(tensor);
-        total_size_new += new_size;
+            total_size_org += tensor_size;
+            total_size_new += new_size;
 
-        // update the gguf meta data as we go
-        gguf_set_tensor_type(ctx_outs[cur_split].get(), name.c_str(), new_type);
-        GGML_ASSERT(gguf_get_tensor_size(ctx_outs[cur_split].get(), gguf_find_tensor(ctx_outs[cur_split].get(), name.c_str())) == new_size);
-        gguf_set_tensor_data(ctx_outs[cur_split].get(), name.c_str(), new_data);
+            // update the gguf meta data as we go
+            gguf_set_tensor_type(ctx_outs[cur_split].get(), name.c_str(), new_type);
+            GGML_ASSERT(gguf_get_tensor_size(ctx_outs[cur_split].get(), gguf_find_tensor(ctx_outs[cur_split].get(), name.c_str())) == new_size);
+            gguf_set_tensor_data(ctx_outs[cur_split].get(), name.c_str(), new_data);
 
-        // write tensor data + padding
-        fout.write((const char *) new_data, new_size);
-        zeros(fout, GGML_PAD(new_size, align) - new_size);
+            // write tensor data + padding
+            fout.write((const char *) new_data, new_size);
+            zeros(fout, GGML_PAD(new_size, align) - new_size);
+        } // no --dry-run
+    } // iterate over tensors
+
+    if (!params->dry_run) {
+        close_ofstream();
     }
-    close_ofstream();
 
-    LLAMA_LOG_INFO("%s: model size  = %8.2f MiB\n", __func__, total_size_org/1024.0/1024.0);
-    LLAMA_LOG_INFO("%s: quant size  = %8.2f MiB\n", __func__, total_size_new/1024.0/1024.0);
+    LLAMA_LOG_INFO("%s: model size  = %8.2f MiB (%.2f BPW)\n", __func__, total_size_org/1024.0/1024.0, total_size_org*8.0/ml.n_elements);
+    LLAMA_LOG_INFO("%s: quant size  = %8.2f MiB (%.2f BPW)\n", __func__, total_size_new/1024.0/1024.0, total_size_new*8.0/ml.n_elements);
+
+    if (!params->imatrix && params->dry_run && will_require_imatrix) {
+        LLAMA_LOG_WARN("%s: WARNING: dry run completed successfully, but actually completing this quantization will require an imatrix!\n",
+                       __func__
+        );
+    }
 
     if (qs.n_fallback > 0) {
         LLAMA_LOG_WARN("%s: WARNING: %d of %d tensor(s) required fallback quantization\n",
@@ -1045,6 +1094,7 @@ llama_model_quantize_params llama_model_quantize_default_params() {
         /*.only_copy                   =*/ false,
         /*.pure                        =*/ false,
         /*.keep_split                  =*/ false,
+        /*.dry_run                     =*/ false,
         /*.imatrix                     =*/ nullptr,
         /*.kv_overrides                =*/ nullptr,
         /*.tensor_type                 =*/ nullptr,

src/llama-vocab.cpp

@@ -289,6 +289,15 @@ struct llm_tokenizer_bpe : llm_tokenizer {
                     "(?:'[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+",
                 };
                 break;
+            case LLAMA_VOCAB_PRE_TYPE_JAIS2:
+                regex_exprs = {
+                    // original regex from tokenizer.json
+                    //"(?i:'s|'t|'re|'ve|'m|'ll|'d)|[^\\r\\n\\p{L}\\p{N}]?\\p{L}+|\\p{N}{1,3}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s{512}(?!\\S)|\\s{256}(?!\\S)|\\s{128}(?!\\S)|\\s{64}(?!\\S)|\\s{32}(?!\\S)|\\s{16}(?!\\S)|\\s{8}(?!\\S)|\\s{4}(?!\\S)|\\s{1,2}(?!\\S)|\\s{1}",
+
+                    // adapted: same as llama3 but with cascading whitespace pattern
+                    "(?:'[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{512}(?!\\S)|\\s{256}(?!\\S)|\\s{128}(?!\\S)|\\s{64}(?!\\S)|\\s{32}(?!\\S)|\\s{16}(?!\\S)|\\s{8}(?!\\S)|\\s{4}(?!\\S)|\\s{1,2}(?!\\S)|\\s{1}",
+                };
+                break;
             case LLAMA_VOCAB_PRE_TYPE_DBRX:
             case LLAMA_VOCAB_PRE_TYPE_SMAUG:
                 regex_exprs = {
@@ -1921,8 +1930,11 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
                     tokenizer_pre == "jina-v2-de" ||
                     tokenizer_pre == "a.x-4.0" ||
                     tokenizer_pre == "mellum"  ||
-                    tokenizer_pre == "modern-bert" ) {
+                    tokenizer_pre == "modern-bert") {
                 pre_type = LLAMA_VOCAB_PRE_TYPE_GPT2;
+            } else if (
+                    tokenizer_pre == "jais-2") {
+                pre_type = LLAMA_VOCAB_PRE_TYPE_JAIS2;
             } else if (
                     tokenizer_pre == "jina-v1-en" ||
                     tokenizer_pre == "jina-v2-code" ||
@@ -2458,6 +2470,7 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
                     || t.first == "<|calls|>"  // solar-open
                     || t.first == "<end_of_turn>"
                     || t.first == "<|endoftext|>"
+                    || t.first == "</s>"      // paddleocr
                     || t.first == "<|eom_id|>"
                     || t.first == "<EOT>"
                     || t.first == "_<EOT>"

src/llama-vocab.h

@@ -57,6 +57,7 @@ enum llama_vocab_pre_type {
     LLAMA_VOCAB_PRE_TYPE_QWEN35          = 46,
     LLAMA_VOCAB_PRE_TYPE_TINY_AYA        = 47,
     LLAMA_VOCAB_PRE_TYPE_JOYAI_LLM       = 48,
+    LLAMA_VOCAB_PRE_TYPE_JAIS2           = 49,
 };
 
 struct LLM_KV;

src/models/delta-net-base.cpp

@@ -27,6 +27,7 @@ std::pair<ggml_tensor *, ggml_tensor *> llm_build_delta_net_base::build_delta_ne
 
     const int64_t S_v = v->ne[0];
     const int64_t H_v = v->ne[1];
+    const bool kda = (g->ne[0] == S_k && g->ne[1] == H_k);
 
     GGML_ASSERT(S_k == S_v);
     GGML_ASSERT(H_v % H_k == 0);
@@ -35,9 +36,10 @@ std::pair<ggml_tensor *, ggml_tensor *> llm_build_delta_net_base::build_delta_ne
     GGML_ASSERT(k->ne[0] == S_k && k->ne[1] == H_k && k->ne[2] == n_tokens && k->ne[3] == n_seqs);
     GGML_ASSERT(v->ne[0] == S_v && v->ne[1] == H_v && v->ne[2] == n_tokens && v->ne[3] == n_seqs);
 
-    GGML_ASSERT(g->ne[0] == H_v && g->ne[1] == n_tokens && g->ne[2] == n_seqs);
-    GGML_ASSERT(b->ne[0] == H_v && b->ne[2] == n_tokens && b->ne[3] == n_seqs);
-    GGML_ASSERT(s->ne[0] == S_v && s->ne[1] == S_v && s->ne[2] == H_v && s->ne[3] == n_seqs);
+    GGML_ASSERT(g->ne[0] == 1   || g->ne[0] == S_v);
+    GGML_ASSERT(                   g->ne[1] == H_v && g->ne[2] == n_tokens && g->ne[3] == n_seqs);
+    GGML_ASSERT(b->ne[0] == 1   && b->ne[1] == H_v && b->ne[2] == n_tokens && b->ne[3] == n_seqs);
+    GGML_ASSERT(s->ne[0] == S_v && s->ne[1] == S_v && s->ne[2] == H_v      && s->ne[3] == n_seqs);
 
     const float scale = 1.0f / sqrtf(S_k);
 
@@ -52,8 +54,8 @@ std::pair<ggml_tensor *, ggml_tensor *> llm_build_delta_net_base::build_delta_ne
     q = ggml_permute(ctx0, q, 0, 2, 1, 3); // [S_k, n_tokens, H_k, n_seqs]
     k = ggml_permute(ctx0, k, 0, 2, 1, 3); // [S_k, n_tokens, H_k, n_seqs]
     v = ggml_permute(ctx0, v, 0, 2, 1, 3); // [S_v, n_tokens, H_v, n_seqs]
-    g = ggml_permute(ctx0, g, 2, 1, 3, 0); // [  1, n_tokens, H_v, n_seqs]
-    b = ggml_permute(ctx0, b, 2, 0, 1, 3); // [  1, n_tokens, H_v, n_seqs]
+    g = ggml_permute(ctx0, g, 0, 2, 1, 3); // [g_0, n_tokens, H_v, n_seqs]
+    b = ggml_permute(ctx0, b, 0, 2, 1, 3); // [  1, n_tokens, H_v, n_seqs]
 
     const int CS = CHUNK_SIZE;
 
@@ -78,33 +80,76 @@ std::pair<ggml_tensor *, ggml_tensor *> llm_build_delta_net_base::build_delta_ne
     v   = ggml_reshape_4d(ctx0, v,   S_v, CS, n_chunks, H_v * n_seqs);
     v_b = ggml_reshape_4d(ctx0, v_b, S_v, CS, n_chunks, H_v * n_seqs);
 
-    g = ggml_reshape_4d(ctx0, g, CS, 1, n_chunks, H_v * n_seqs);
-    b = ggml_reshape_4d(ctx0, b, 1, CS, n_chunks, H_v * n_seqs);
+    g = ggml_reshape_4d(ctx0, g, g->ne[0], CS, n_chunks, H_v * n_seqs);
+    b = ggml_reshape_4d(ctx0, b, 1,        CS, n_chunks, H_v * n_seqs);
 
-    // [CS, 1, n_chunks, H_v * n_seqs]
-    ggml_tensor * g_cs = ggml_cumsum(ctx0, g);
+    // [CS, g_0, n_chunks, H_v * n_seqs]
+    // TODO: extend ggml_cumsum with axis parameter to avoid transpose
+    ggml_tensor * g_cs = ggml_cumsum(ctx0, ggml_cont(ctx0, ggml_transpose(ctx0, g)));
     cb(g_cs, "g_cs", il);
 
-    ggml_tensor * g_cs_i = g_cs;
-    ggml_tensor * g_cs_j = ggml_reshape_4d(ctx0, g_cs, 1, CS, n_chunks, H_v * n_seqs);
+    ggml_tensor * kb = nullptr;
+    ggml_tensor * kq = nullptr;
+    if (kda) {
+        const int64_t CHB = n_chunks * H_k * n_seqs;
 
-    g_cs_j = ggml_repeat_4d(ctx0, g_cs_j, CS, CS, n_chunks, H_v * n_seqs);
+        ggml_tensor * g_cs_i = ggml_reshape_4d(ctx0, g_cs, CS, 1, S_k, CHB);  // [chunk_size, 1, S_k, CHB]
+        ggml_tensor * g_cs_j = ggml_reshape_4d(ctx0, g_cs, 1, CS, S_k, CHB);  // [1, chunk_size, S_k, CHB]
 
-    // [CS, CS, n_chunks, H_v * n_seqs]
-    ggml_tensor * decay_mask;
-    decay_mask = ggml_sub(ctx0, g_cs_j, g_cs_i);
-    decay_mask = ggml_tri(ctx0, decay_mask, GGML_TRI_TYPE_LOWER_DIAG);
-    decay_mask = ggml_exp(ctx0, decay_mask);
-    cb(decay_mask, "decay_mask", il);
+        g_cs_j = ggml_repeat_4d(ctx0, g_cs_j, CS, CS, S_k, CHB);  // [1, chunk_size, S_k, CHB] -> [chunk_size, chunk_size, S_k, CHB]
 
-    // [CS, CS, n_chunks, H_k * n_seqs]
-    ggml_tensor * kb;
-    kb = ggml_mul_mat(ctx0, k,  k_b);
-    kb = ggml_mul    (ctx0, kb, decay_mask);
+        // decay_mask [chunk_size,chunk_size,S_k,CHB]
+        ggml_tensor * decay_mask;
+        decay_mask = ggml_sub(ctx0, g_cs_j, g_cs_i);
+        decay_mask = ggml_tri(ctx0, decay_mask, GGML_TRI_TYPE_LOWER_DIAG);
+        decay_mask = ggml_exp(ctx0, decay_mask);
+        cb(decay_mask, "decay_mask", il);
+
+        // decay_mask [S_k,BT_j,BT_i,CHB] *Note* second and third chunk_sizes are switched
+        decay_mask = ggml_cont_4d(ctx0, ggml_permute(ctx0, decay_mask, 2, 1, 0, 3), S_k, CS, CS, CHB);
+
+        ggml_tensor * k_b_i = ggml_reshape_4d(ctx0, k_b, S_k, CS,  1, CHB);
+        ggml_tensor * k_j   = ggml_reshape_4d(ctx0, k,   S_k,  1, CS, CHB);
+        ggml_tensor * q_i   = ggml_reshape_4d(ctx0, q,   S_k, CS,  1, CHB);
+
+        ggml_tensor * decay_k_b_i = ggml_mul(ctx0, decay_mask, k_b_i);
+        ggml_tensor * decay_q_i   = ggml_mul(ctx0, decay_mask, q_i);
+
+        // decay_k_b_i [S,BT,BT,CHB] @ k_j [S,1,BT,CHB] = Akk [BT,1,BT,CHB]
+        kb = ggml_mul_mat(ctx0, decay_k_b_i, k_j);
+        kq = ggml_mul_mat(ctx0, decay_q_i,   k_j);
+
+        kb = ggml_cont(ctx0, ggml_transpose(ctx0, ggml_reshape_4d(ctx0, kb, CS, CS, n_chunks, H_v * n_seqs)));
+        kq = ggml_cont(ctx0, ggml_transpose(ctx0, ggml_reshape_4d(ctx0, kq, CS, CS, n_chunks, H_v * n_seqs)));
+    } else {
+        ggml_tensor * g_cs_i = g_cs;
+        ggml_tensor * g_cs_j = ggml_reshape_4d(ctx0, g_cs, 1, CS, n_chunks, H_v * n_seqs);
+
+        g_cs_j = ggml_repeat_4d(ctx0, g_cs_j, CS, CS, n_chunks, H_v * n_seqs);
+
+        // [CS, CS, n_chunks, H_v * n_seqs]
+        ggml_tensor * decay_mask;
+        decay_mask = ggml_sub(ctx0, g_cs_j, g_cs_i);
+        decay_mask = ggml_tri(ctx0, decay_mask, GGML_TRI_TYPE_LOWER_DIAG);
+        decay_mask = ggml_exp(ctx0, decay_mask);
+        cb(decay_mask, "decay_mask", il);
+
+        // [CS, CS, n_chunks, H_k * n_seqs]
+        kb = ggml_mul_mat(ctx0, k,  k_b);
+        kb = ggml_mul    (ctx0, kb, decay_mask);
+
+        // [CS, CS, n_chunks, H_k * n_seqs]
+        kq = ggml_mul_mat(ctx0, k, q);
+        kq = ggml_mul(ctx0, kq, decay_mask);
+    }
+
+    kq = ggml_tri(ctx0, kq, GGML_TRI_TYPE_LOWER_DIAG);
+    cb(kq, "kq", il);
 
     // [CS, CS, n_chunks, H_k * n_seqs]
     ggml_tensor * attn;
     attn = ggml_tri(ctx0, kb, GGML_TRI_TYPE_LOWER);
+    cb(attn, "attn", il);
 
     ggml_tensor * identity;
     identity = ggml_view_1d(ctx0, attn, CS, 0);
@@ -115,6 +160,7 @@ std::pair<ggml_tensor *, ggml_tensor *> llm_build_delta_net_base::build_delta_ne
     cb(lhs, "dnet_add_ch_lhs", il);
 
     attn = ggml_neg(ctx0, attn);
+    cb(attn, "attn_pre_solve", il);
 
     ggml_tensor * lin_solve = ggml_solve_tri(ctx0, lhs, attn, true, true, false);
     attn = ggml_add(ctx0, lin_solve, identity);
@@ -123,7 +169,7 @@ std::pair<ggml_tensor *, ggml_tensor *> llm_build_delta_net_base::build_delta_ne
     // [S_v, CS, n_chunks, H_v * n_seqs]
     v = ggml_mul_mat(ctx0, ggml_cont(ctx0, ggml_transpose(ctx0, v_b)), attn);
 
-    // [CS, 1, n_chunks, H_v * n_seqs]
+    // [CS, 1, n_chunks, H_v * n_seqs] KDA: [CS, S_k, n_chunks, H_v * n_seqs]
     ggml_tensor * g_exp = ggml_exp(ctx0, g_cs);
 
     k_b = ggml_cont(ctx0, ggml_transpose(ctx0, k_b));
@@ -136,16 +182,10 @@ std::pair<ggml_tensor *, ggml_tensor *> llm_build_delta_net_base::build_delta_ne
     ggml_tensor * k_cd = ggml_mul_mat(ctx0, kbg, attn);
     cb(k_cd, "k_cumdecay", il);
 
-    // [S_k, CS, n_chunks, H_k * n_seqs]
-    ggml_tensor * g_exp_t = ggml_transpose(ctx0, g_exp);
+    // [1, CS, n_chunks, H_k * n_seqs] KDA: [S_k, CS, n_chunks, H_k * n_seqs]
+    ggml_tensor * g_exp_t = ggml_cont(ctx0, ggml_transpose(ctx0, g_exp));
     ggml_tensor * q_g_exp = ggml_mul(ctx0, q, g_exp_t);
 
-    // [CS, CS, n_chunks, H_k * n_seqs]
-    ggml_tensor * kq = ggml_mul_mat(ctx0, k, q);
-    kq = ggml_mul(ctx0, kq, decay_mask);
-    kq = ggml_tri(ctx0, kq, GGML_TRI_TYPE_LOWER_DIAG);
-    cb(kq, "kq", il);
-
     // vectorized calculation of key_gdiff
     // improved from the chunked version:
     //   g_last = torch.clamp(g_cum[:, :, -1], max=50.0).exp().unsqueeze(-1).unsqueeze(-1)
@@ -156,8 +196,8 @@ std::pair<ggml_tensor *, ggml_tensor *> llm_build_delta_net_base::build_delta_ne
 
     // get last element in g_cumsum along CS dimension (ne0)
     // example: [[x, y, z, ..., last], ...] -> [[last], ...]
-    // [1, 1, n_chunks, H_v * n_seqs]
-    ggml_tensor * g_last = ggml_view_4d(ctx0, g_cs, 1, 1, g_cs->ne[2], g_cs->ne[3],
+    // [1, 1, n_chunks, H_v * n_seqs] KDA: [1, S_k, n_chunks, H_v * n_seqs]
+    ggml_tensor * g_last = ggml_view_4d(ctx0, g_cs, 1, g_cs->ne[1], g_cs->ne[2], g_cs->ne[3],
             g_cs->nb[1],
             g_cs->nb[2],
             g_cs->nb[3],
@@ -167,16 +207,15 @@ std::pair<ggml_tensor *, ggml_tensor *> llm_build_delta_net_base::build_delta_ne
     // TODO: remove this cont when CUDA supports non-cont unary ops
     g_last = ggml_cont(ctx0, g_last);
 
-    // [1, 1, n_chunks, H_v * n_seqs]
-    ggml_tensor * g_last_exp = ggml_exp(ctx0, g_last);
-    cb(g_last_exp, "g_last_exp", il);
+    // [1, 1, n_chunks, H_v * n_seqs] KDA: [S_k, 1, n_chunks, H_v * n_seqs]
+    ggml_tensor * g_last_exp_t = ggml_transpose(ctx0, ggml_exp(ctx0, g_last));
+    cb(g_last_exp_t, "g_last_exp_t", il);
 
-    // [CS, 1, n_chunks, H_v * n_seqs]
+    // [CS, 1, n_chunks, H_v * n_seqs] KDA: [CS, S_k, n_chunks, H_v * n_seqs]
     ggml_tensor * g_diff = ggml_neg(ctx0, ggml_sub(ctx0, g_cs, g_last));
     cb(g_diff, "g_diff", il);
 
-    ggml_tensor * g_diff_exp   = ggml_exp(ctx0, g_diff);
-    ggml_tensor * g_diff_exp_t = ggml_transpose(ctx0, g_diff_exp);
+    ggml_tensor * g_diff_exp_t = ggml_cont(ctx0, ggml_transpose(ctx0, ggml_exp(ctx0, g_diff)));
 
     // [S_k, CS, n_chunks, H_v * n_seqs]
     ggml_tensor * kg = ggml_mul(ctx0, k, g_diff_exp_t);
@@ -227,8 +266,9 @@ std::pair<ggml_tensor *, ggml_tensor *> llm_build_delta_net_base::build_delta_ne
         ggml_tensor * kgv = ggml_mul_mat(ctx0, ch_kg_t, v_t_new); // [S_k, S_v, 1, H_k * n_seqs]
 
         // last_recurrent_state = last_recurrent_state * g_last + kgdmulvnew
-        ggml_tensor * ch_g_last_exp = get_slice_2d(ctx0, g_last_exp, chunk);
-        s_t = ggml_mul(ctx0, s_t, ch_g_last_exp);
+        ggml_tensor * ch_g_last_exp_t = get_slice_2d(ctx0, g_last_exp_t, chunk);
+
+        s_t = ggml_mul(ctx0, s_t, ch_g_last_exp_t);
         s_t = ggml_add(ctx0, s_t, kgv);
         cb(s_t, "dnet_add_ch_state", il);
     }
@@ -241,9 +281,9 @@ std::pair<ggml_tensor *, ggml_tensor *> llm_build_delta_net_base::build_delta_ne
             ggml_row_size(v->type, S_v),
             ggml_row_size(v->type, S_v * CS * n_chunks),
             ggml_row_size(v->type, S_v * CS * n_chunks * H_v), 0);
-
     o = ggml_permute  (ctx0, o, 0, 2, 1, 3); // [S_v, H_v, n_tokens, n_seqs]
-    s = ggml_transpose(ctx0, s_t);           // [S_v, S_v, H_v, n_seqs]
+    s = ggml_transpose(ctx0, s_t);
+    cb(s, "output_state", il);
 
     return {o, s};
 }
@@ -273,9 +313,10 @@ std::pair<ggml_tensor *, ggml_tensor *> llm_build_delta_net_base::build_delta_ne
     GGML_ASSERT(k->ne[0] == S_k && k->ne[1] == H_k && k->ne[2] == n_tokens && k->ne[3] == n_seqs);
     GGML_ASSERT(v->ne[0] == S_v && v->ne[1] == H_v && v->ne[2] == n_tokens && v->ne[3] == n_seqs);
 
-    GGML_ASSERT(g->ne[0] == H_v && g->ne[1] == n_tokens && g->ne[2] == n_seqs);
-    GGML_ASSERT(b->ne[0] == H_v && b->ne[2] == n_tokens && b->ne[3] == n_seqs);
-    GGML_ASSERT(s->ne[0] == S_v && s->ne[1] == S_v && s->ne[2] == H_v && s->ne[3] == n_seqs);
+    GGML_ASSERT(g->ne[0] == 1   || g->ne[0] == S_v);
+    GGML_ASSERT(                   g->ne[1] == H_v && g->ne[2] == n_tokens && g->ne[3] == n_seqs);
+    GGML_ASSERT(b->ne[0] == 1   && b->ne[1] == H_v && b->ne[2] == n_tokens && b->ne[3] == n_seqs);
+    GGML_ASSERT(s->ne[0] == S_v && s->ne[1] == S_v && s->ne[2] == H_v      && s->ne[3] == n_seqs);
 
     const float scale = 1.0f / sqrtf(S_k);
 
@@ -291,8 +332,10 @@ std::pair<ggml_tensor *, ggml_tensor *> llm_build_delta_net_base::build_delta_ne
     cb(b, "b_in", il);
     cb(g, "g_in", il);
 
-    g = ggml_reshape_4d(ctx0, g, 1, 1, H_v, n_seqs);
-    b = ggml_reshape_4d(ctx0, b, 1, 1, H_v, n_seqs);
+    // GDA: [1,  1,  H_v, n_seqs]
+    // KDA: [1, S_k, H_v, n_seqs]
+    g = ggml_reshape_4d(ctx0, g, 1, g->ne[0], H_v, n_seqs);
+    b = ggml_reshape_4d(ctx0, b, 1,        1, H_v, n_seqs);
 
     // [S_v, S_v, H_v, n_seqs]
     g = ggml_exp(ctx0, g);

src/models/jais2.cpp

@@ -0,0 +1,123 @@
+#include "models.h"
+
+// JAIS-2 model graph builder
+// Uses: LayerNorm (not RMSNorm), relu2 activation, separate Q/K/V, RoPE embeddings
+llm_build_jais2::llm_build_jais2(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    // KV input for attention
+    auto * inp_attn = build_attn_inp_kv();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        // Pre-attention LayerNorm
+        cur = build_norm(inpL,
+                model.layers[il].attn_norm,
+                model.layers[il].attn_norm_b,
+                LLM_NORM, il);
+        cb(cur, "attn_norm", il);
+
+        // Self-attention with separate Q, K, V projections
+        {
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            cb(Qcur, "Qcur", il);
+            Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+            cb(Qcur, "Qcur_bias", il);
+
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            cb(Kcur, "Kcur", il);
+            Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+            cb(Kcur, "Kcur_bias", il);
+
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            cb(Vcur, "Vcur", il);
+            Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+            cb(Vcur, "Vcur_bias", il);
+
+            // Reshape for attention
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            // Apply RoPE
+            Qcur = ggml_rope_ext(
+                ctx0, Qcur, inp_pos, nullptr,
+                n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                ext_factor, attn_factor, beta_fast, beta_slow
+            );
+
+            Kcur = ggml_rope_ext(
+                ctx0, Kcur, inp_pos, nullptr,
+                n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                ext_factor, attn_factor, beta_fast, beta_slow
+            );
+
+            cb(Qcur, "Qcur_rope", il);
+            cb(Kcur, "Kcur_rope", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, model.layers[il].bo,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+        }
+
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur  = ggml_get_rows(ctx0,  cur, inp_out_ids);
+            inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
+        }
+
+        // Residual connection
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL);
+        cb(ffn_inp, "ffn_inp", il);
+
+        // Pre-FFN LayerNorm
+        cur = build_norm(ffn_inp,
+                model.layers[il].ffn_norm,
+                model.layers[il].ffn_norm_b,
+                LLM_NORM, il);
+        cb(cur, "ffn_norm", il);
+
+        // FFN with relu2 activation (ReLU squared) - no gate projection
+        // up -> relu2 -> down
+        cur = build_ffn(cur,
+                model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
+                NULL, NULL, NULL,  // no gate
+                model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
+                NULL,
+                LLM_FFN_RELU_SQR, LLM_FFN_SEQ, il);
+        cb(cur, "ffn_out", il);
+
+        // Residual connection
+        inpL = ggml_add(ctx0, cur, ffn_inp);
+        inpL = build_cvec(inpL, il);
+        cb(inpL, "l_out", il);
+    }
+
+    // Final LayerNorm
+    cur = build_norm(inpL,
+            model.output_norm,
+            model.output_norm_b,
+            LLM_NORM, -1);
+    cb(cur, "result_norm", -1);
+
+    res->t_embd = cur;
+
+    // Output projection
+    cur = build_lora_mm(model.output, cur);
+    cb(cur, "result_output", -1);
+
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}

src/models/kimi-linear.cpp

@@ -3,8 +3,6 @@
 
 #include "llama-memory-recurrent.h"
 
-#define CHUNK_SIZE 64
-
 // Causal Conv1d function for Q,K,V
 // When qkv is 0, it is Q, 1 is K, 2 is V
 static ggml_tensor * causal_conv1d(ggml_cgraph * gf, ggml_context * ctx0, ggml_tensor * conv_states_all, ggml_tensor * conv_state_all, int64_t qkv, ggml_tensor * x, ggml_tensor * proj_w, ggml_tensor * conv_w, int64_t d_conv, int64_t head_dim, int64_t n_head, int64_t n_seq_tokens, int64_t n_seqs, int64_t n_tokens, int64_t kv_head) {
@@ -67,7 +65,7 @@ static ggml_tensor * causal_conv1d(ggml_cgraph * gf, ggml_context * ctx0, ggml_t
 }
 
 llm_build_kimi_linear::llm_build_kimi_linear(const llama_model & model, const llm_graph_params & params) :
-    llm_build_mamba_base(params), model(model) {
+    llm_build_delta_net_base(params), model(model) {
     ggml_tensor * cur;
     ggml_tensor * inpL;
 
@@ -86,17 +84,6 @@ llm_build_kimi_linear::llm_build_kimi_linear(const llama_model & model, const ll
     // Output ids for selecting which tokens to output
     ggml_tensor * inp_out_ids = build_inp_out_ids();
 
-    ggml_tensor * chunked_causal_mask =
-        ggml_tri(ctx0, ggml_fill_inplace(ctx0, ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, CHUNK_SIZE, CHUNK_SIZE), 1.0f),
-                    GGML_TRI_TYPE_LOWER);
-
-    ggml_tensor * chunked_identity = ggml_diag(ctx0, ggml_fill_inplace(ctx0, ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, CHUNK_SIZE), 1.0f));
-    ggml_tensor * chunked_diag_mask = ggml_add(ctx0, chunked_causal_mask, chunked_identity);
-
-    ggml_build_forward_expand(gf, chunked_causal_mask);
-    ggml_build_forward_expand(gf, chunked_identity);
-    ggml_build_forward_expand(gf, chunked_diag_mask);
-
     // Kimi dimension constants
     const int64_t n_head = hparams.n_head();
     const int64_t head_dim = hparams.n_embd_head_kda;
@@ -162,27 +149,35 @@ llm_build_kimi_linear::llm_build_kimi_linear(const llama_model & model, const ll
             g1 = ggml_mul(ctx0, g1, A);
             cb(g1, "kda_g1", il);
 
+            g1 = ggml_reshape_4d(ctx0, g1, head_dim, n_head, n_seq_tokens, n_seqs);
+
             // Compute beta (mixing coefficient)
             ggml_tensor * beta = ggml_mul_mat(ctx0, layer.ssm_beta, cur);
-            beta = ggml_reshape_4d(ctx0, beta, n_head, 1, n_seq_tokens, n_seqs);
+            beta = ggml_reshape_4d(ctx0, beta, 1, n_head, n_seq_tokens, n_seqs);
             cb(beta, "kda_beta", il);
 
+            beta = ggml_sigmoid(ctx0, beta);
+
             // Reshape for KDA recurrence
             // {n_embd, n_tokens} -> {n_embd, n_seq_tokens, n_seqs}
             cur = ggml_reshape_3d(ctx0, cur, cur->ne[0], n_seq_tokens, n_seqs);
 
-            g1 = ggml_reshape_4d(ctx0, g1, head_dim, n_head, n_seq_tokens, n_seqs);
-
             // Get SSM state and compute KDA recurrence using ggml_kda_scan
             ggml_tensor * ssm_states_all = mctx_cur->get_s_l(il);
             ggml_tensor * state = build_rs(inp_rs, ssm_states_all, hparams.n_embd_s(), n_seqs);
             state = ggml_reshape_4d(ctx0, state, head_dim, head_dim, n_head, n_seqs);
-            // Choose between build_kda_chunking and build_kda_recurrent based on n_tokens
-            std::pair<ggml_tensor *, ggml_tensor *> attn_out = n_seq_tokens == 1 ?
-                build_kda_autoregressive(Qcur, Kcur, Vcur, g1, beta, state, il) :
-                build_kda_chunking(Qcur, Kcur, Vcur, g1, beta, state, chunked_causal_mask, chunked_identity, chunked_diag_mask, il);
 
-            ggml_tensor * output = attn_out.first;
+            const float eps_norm = hparams.f_norm_rms_eps;
+
+            Qcur = ggml_l2_norm(ctx0, Qcur, eps_norm);
+            Kcur = ggml_l2_norm(ctx0, Kcur, eps_norm);
+
+            // Choose between build_delta_net_chunking and build_delta_net_recurrent based on n_tokens
+            std::pair<ggml_tensor *, ggml_tensor *> attn_out = n_seq_tokens == 1 ?
+                build_delta_net_autoregressive(Qcur, Kcur, Vcur, g1, beta, state, il) :
+                build_delta_net_chunking(Qcur, Kcur, Vcur, g1, beta, state, il);
+
+            ggml_tensor * output = ggml_cont(ctx0, attn_out.first);
             ggml_tensor * new_state = attn_out.second;
             cb(output, "attn_output", il);
             cb(new_state, "new_state", il);
@@ -393,385 +388,3 @@ llm_build_kimi_linear::llm_build_kimi_linear(const llama_model & model, const ll
 
     ggml_build_forward_expand(gf, cur);
 }
-
-/*
-    This is a ggml implementation of the naive_chunk_kda function of
-    https://github.com/fla-org/flash-linear-attention/blob/main/fla/ops/kda/naive.py
-*/
-std::pair<ggml_tensor *, ggml_tensor *> llm_build_kimi_linear::build_kda_chunking(
-        ggml_tensor * q,
-        ggml_tensor * k,
-        ggml_tensor * v,
-        ggml_tensor * gk,
-        ggml_tensor * beta,
-        ggml_tensor * state,
-        ggml_tensor * causal_mask,
-        ggml_tensor * identity,
-        ggml_tensor * diag_mask,
-        int           il) {
-    GGML_ASSERT(ggml_is_contiguous(state));
-
-    const int64_t S_k      = q->ne[0];
-    const int64_t H_k      = q->ne[1];
-    const int64_t n_tokens = q->ne[2];
-    const int64_t n_seqs   = q->ne[3];
-
-    const int64_t S_v = v->ne[0];
-    const int64_t H_v = v->ne[1];
-
-    GGML_ASSERT(v->ne[2] == n_tokens);
-    GGML_ASSERT(k->ne[2] == n_tokens);
-    GGML_ASSERT(gk->ne[0] == S_v && gk->ne[1] == H_v && gk->ne[2] == n_tokens && gk->ne[3] == n_seqs);
-    GGML_ASSERT(beta->ne[0] == H_v && beta->ne[2] == n_tokens && beta->ne[3] == n_seqs);
-    GGML_ASSERT(state->ne[0] == S_v && state->ne[1] == S_v && state->ne[2] == H_v && state->ne[3] == n_seqs);
-
-    GGML_ASSERT(q->ne[0] == S_k && q->ne[1] == H_k && q->ne[2] == n_tokens && q->ne[3] == n_seqs);
-    GGML_ASSERT(k->ne[0] == S_k && k->ne[1] == H_k && k->ne[2] == n_tokens && k->ne[3] == n_seqs);
-
-    GGML_ASSERT(H_k == H_v);  // we did a repeat to make sure this is the case
-
-    // TODO: can this ever be false?
-    const bool use_qk_l2norm = true;
-
-    if (use_qk_l2norm) {
-        const float eps_norm = hparams.f_norm_rms_eps;
-
-        q = ggml_l2_norm(ctx0, q, eps_norm);
-        k = ggml_l2_norm(ctx0, k, eps_norm);
-    }
-
-    const float scale = 1.0f / sqrtf(S_v);
-
-    beta = ggml_sigmoid(ctx0, beta);
-
-    cb(q, "q_in", il);
-    cb(k, "k_in", il);
-    cb(v, "v_in", il);
-    cb(beta, "beta_in", il);
-    cb(gk, "gk_in", il);
-
-    q = ggml_cont_4d(ctx0, ggml_permute(ctx0, q, 0, 2, 1, 3), S_k, n_tokens, H_k, n_seqs);
-    k = ggml_cont_4d(ctx0, ggml_permute(ctx0, k, 0, 2, 1, 3), S_k, n_tokens, H_k, n_seqs);
-    v = ggml_cont_4d(ctx0, ggml_permute(ctx0, v, 0, 2, 1, 3), S_v, n_tokens, H_v, n_seqs);
-    gk = ggml_cont_4d(ctx0, ggml_permute(ctx0, gk, 0, 2, 1, 3), S_v, n_tokens, H_v, n_seqs);
-
-    beta  = ggml_cont(ctx0, ggml_permute(ctx0, beta, 2, 0, 1, 3));
-    state = ggml_reshape_4d(ctx0, state, S_v, S_v, H_v, n_seqs);
-
-    cb(q, "q_perm", il);
-    cb(k, "k_perm", il);
-    cb(v, "v_perm", il);
-    cb(beta, "beta_perm", il);
-    cb(gk, "gk_perm", il);
-    cb(state, "state_in", il);
-
-    GGML_ASSERT(q->ne[1] == n_tokens && q->ne[0] == S_k && q->ne[2] == H_k && q->ne[3] == n_seqs);
-    GGML_ASSERT(k->ne[1] == n_tokens && k->ne[0] == S_k && k->ne[2] == H_k && k->ne[3] == n_seqs);
-    GGML_ASSERT(v->ne[1] == n_tokens && v->ne[0] == S_v && v->ne[2] == H_k && v->ne[3] == n_seqs);
-    GGML_ASSERT(beta->ne[1] == n_tokens && beta->ne[2] == H_k && beta->ne[0] == 1 && beta->ne[3] == n_seqs);
-
-    // Do padding
-    const int64_t chunk_size = CHUNK_SIZE;
-
-    const int64_t pad = (chunk_size - n_tokens % chunk_size) % chunk_size;
-    const int64_t n_chunks = (n_tokens + pad) / chunk_size;
-
-    q = ggml_pad(ctx0, q, 0, pad, 0, 0);
-    k = ggml_pad(ctx0, k, 0, pad, 0, 0);
-    v = ggml_pad(ctx0, v, 0, pad, 0, 0);
-    gk = ggml_pad(ctx0, gk, 0, pad, 0, 0);
-    beta = ggml_pad(ctx0, beta, 0, pad, 0, 0);
-
-    cb(q, "q_pad", il);
-    cb(k, "k_pad", il);
-    cb(v, "v_pad", il);
-    cb(beta, "beta_pad", il);
-    cb(gk, "gk_pad", il);
-
-    ggml_tensor * v_beta = ggml_mul(ctx0, v, beta);
-    ggml_tensor * k_beta = ggml_mul(ctx0, k, beta);
-
-    cb(v_beta, "v_beta", il);
-    cb(k_beta, "k_beta", il);
-
-    const int64_t HB = H_k * n_seqs;
-
-    q      = ggml_cont_4d(ctx0, q,      S_k, chunk_size, n_chunks, HB);
-    k      = ggml_cont_4d(ctx0, k,      S_k, chunk_size, n_chunks, HB);
-    k_beta = ggml_cont_4d(ctx0, k_beta, S_k, chunk_size, n_chunks, HB);
-    v      = ggml_cont_4d(ctx0, v,      S_v, chunk_size, n_chunks, HB);
-    v_beta = ggml_cont_4d(ctx0, v_beta, S_v, chunk_size, n_chunks, HB);
-
-    gk    = ggml_cont_4d(ctx0, gk, S_k, chunk_size, n_chunks, HB);
-    beta = ggml_cont_4d(ctx0, beta, 1, chunk_size, n_chunks, HB);
-
-    // switch for cumsum
-    gk = ggml_cont_4d(ctx0, ggml_permute(ctx0, gk, 1, 0, 2, 3), chunk_size, S_k, n_chunks, HB);
-    cb(gk, "gk", il);
-    ggml_tensor * gk_cumsum = ggml_cumsum(ctx0, gk);
-    cb(gk_cumsum, "gk_cumsum", il);
-
-/*
-    Compute Akk and Aqk loop together
-    Akk loop:
-    for i in range(BT):
-        k_i = k[..., i, :] # k_i [B,H,NT,S]
-        g_i = g[..., i:i+1, :] # g_i [B,H,NT,1,S]
-        A[..., i] = torch.einsum('... c d, ... d -> ... c', k * (g - g_i).exp(), k_i)
-    Aqk loop:
-    for j in range(BT):
-        k_j = k[:, :, i, j]
-        g_j = g[:, :, i, j:j+1, :]
-        A[..., j] = torch.einsum('... c d, ... d -> ... c', q_i * (g_i - g_j).exp(), k_j)
-*/
-    const int64_t CHB = n_chunks * H_k * n_seqs;
-    ggml_tensor * gkcs_i = ggml_reshape_4d(ctx0, gk_cumsum, chunk_size, 1, S_k, CHB);  // [chunk_size, 1, S_k, CHB]
-    ggml_tensor * gkcs_j = ggml_reshape_4d(ctx0, gkcs_i, 1, chunk_size, S_k, CHB);  // [1, chunk_size, S_k, CHB]
-
-    ggml_tensor * gkcs_j_bc = ggml_repeat_4d(ctx0, gkcs_j, chunk_size, chunk_size, S_k, CHB);  // [1, chunk_size, S_k, CHB] -> [chunk_size, chunk_size, S_k, CHB]
-    // decay_mask [chunk_size,chunk_size,S_k,CHB]
-    ggml_tensor * decay_mask = ggml_sub(ctx0, gkcs_j_bc, gkcs_i);
-    cb(decay_mask, "decay_mask", il);
-
-    decay_mask = ggml_mul(ctx0, decay_mask, diag_mask);
-    cb(decay_mask, "decay_masked", il);
-    decay_mask = ggml_exp(ctx0, decay_mask);
-    decay_mask = ggml_mul(ctx0, decay_mask, diag_mask);
-
-    // decay_mask [S_k,BT_j,BT_i,CHB] *Note* second and third chunk_sizes are switched
-    decay_mask = ggml_cont_4d(ctx0, ggml_permute(ctx0, decay_mask, 2, 1, 0, 3), S_k, chunk_size, chunk_size, CHB);
-
-    ggml_tensor * k_i = ggml_reshape_4d(ctx0, k, S_k, chunk_size, 1, CHB);
-    ggml_tensor * k_j = ggml_reshape_4d(ctx0, k, S_k, 1, chunk_size, CHB);
-    ggml_tensor * q_i = ggml_reshape_4d(ctx0, q, S_k, chunk_size, 1, CHB);
-
-    ggml_tensor * decay_k_i = ggml_mul(ctx0, decay_mask, k_i);
-    ggml_tensor * decay_q_i = ggml_mul(ctx0, decay_mask, q_i);
-
-    // decay_k_i [S.BT,BT,CHB] @ k_j [S,1,BT,CHB] = Akk [BT,1,BT,CHB]
-    ggml_tensor * Akk = ggml_mul_mat(ctx0, decay_k_i, k_j);
-    ggml_tensor * Aqk = ggml_mul_mat(ctx0, decay_q_i, k_j);
-    Akk = ggml_cont(ctx0, ggml_transpose(ctx0, ggml_reshape_4d(ctx0, Akk, chunk_size, chunk_size, n_chunks, HB)));
-    Aqk = ggml_cont(ctx0, ggml_transpose(ctx0, ggml_reshape_4d(ctx0, Aqk, chunk_size, chunk_size, n_chunks, HB)));
-    cb(Akk, "Akk", il);
-    cb(Aqk, "Aqk", il);
-
-    Akk = ggml_mul(ctx0, Akk, beta);
-    Akk = ggml_neg(ctx0, ggml_mul(ctx0, Akk, causal_mask));
-    cb(Akk, "attn_pre_solve", il);
-
-    Aqk = ggml_mul(ctx0, Aqk, diag_mask);
-    Aqk = ggml_scale(ctx0, Aqk, scale); // scale q
-    cb(Aqk, "Aqk_masked", il);
-
-    // for i in range(1, chunk_size):
-    //          row = attn[..., i, :i].clone()
-    //          sub = attn[..., :i, :i].clone()
-    //          attn[..., i, :i] = row + (row.unsqueeze(-1) * sub).sum(-2)
-    // attn = attn + torch.eye(chunk_size, dtype=attn.dtype, device=attn.device)
-    //
-    // We reduce this to a linear triangular solve: AX = B, where B = attn, A = I - tril(A)
-    ggml_tensor * attn_lower = ggml_mul(ctx0, Akk, causal_mask);
-    ggml_tensor * lhs        = ggml_sub(ctx0, ggml_repeat(ctx0, identity, attn_lower), attn_lower);
-
-    ggml_tensor * lin_solve  = ggml_solve_tri(ctx0, lhs, Akk, true, true, false);
-    Akk                      = ggml_mul(ctx0, lin_solve, causal_mask);
-    Akk                      = ggml_add(ctx0, Akk, identity);
-
-    cb(Akk, "attn_solved", il);
-
-    // switch back for downstream
-    gk_cumsum = ggml_cont_4d(ctx0, ggml_permute(ctx0, gk_cumsum, 1, 0, 2, 3), S_k, chunk_size, n_chunks, HB);
-    ggml_tensor * gkexp      = ggml_exp(ctx0, gk_cumsum);
-    cb(gk_cumsum, "gk_cumsum", il);
-
-    // u = (A*beta[..., None, :]) @ v  aka U_[t]
-    ggml_tensor * vb = ggml_mul_mat(ctx0, ggml_cont(ctx0, ggml_transpose(ctx0, v_beta)), Akk);
-
-    ggml_tensor * kbeta_gkexp = ggml_mul(ctx0, k_beta, gkexp);
-    cb(kbeta_gkexp, "kbeta_gkexp", il);
-
-    ggml_tensor * k_cumdecay = ggml_mul_mat(ctx0, ggml_cont(ctx0, ggml_transpose(ctx0, kbeta_gkexp)), Akk);
-    cb(k_cumdecay, "k_cumdecay", il);
-
-    ggml_tensor * core_attn_out = nullptr;
-    ggml_tensor * new_state = ggml_dup(ctx0, state);
-
-    cb(new_state, "new_state", il);
-
-    for (int64_t chunk = 0; chunk < n_chunks; chunk++) {
-// extract one chunk worth of data
-        auto chunkify = [=](ggml_tensor * t) {
-                    return ggml_cont(ctx0, ggml_view_4d(ctx0, t, t->ne[0], chunk_size, 1, t->ne[3],
-                t->nb[1], t->nb[2], t->nb[3], t->nb[2] * chunk));
-        };
-        auto chunkify_A = [=](ggml_tensor * t) {
-                    return ggml_cont(ctx0, ggml_view_4d(ctx0, t, chunk_size, chunk_size, 1, t->ne[3],
-                t->nb[1], t->nb[2], t->nb[3], t->nb[2] * chunk));
-        };
-
-
-// k [S,BT,NT,H*B] => k_chunk [S,BT,1,H*B]
-        ggml_tensor * k_chunk = chunkify(k);
-        ggml_tensor * q_chunk = chunkify(q);
-        ggml_tensor * vb_chunk = chunkify(vb);
-
-// gk_cumsum [S,BT,NT,H*B] => gk_cs_chunk [S,BT,1,H*B]
-        ggml_tensor * gk_cs_chunk = chunkify(gk_cumsum);
-        ggml_tensor * k_cumdecay_chunk = chunkify(k_cumdecay);
-        ggml_tensor * gkexp_chunk = ggml_exp(ctx0, gk_cs_chunk);
-        ggml_tensor * Aqk_chunk = chunkify_A(Aqk);
-
-        ggml_tensor * state_t = ggml_cont_4d(ctx0, ggml_permute(ctx0, new_state, 1, 0, 2, 3), S_v, S_v, 1, H_v * n_seqs);
-
-        // new_state [S,S,1,H*B] k_cumdecay_chunk [S,BT,1,H*B]
-        // v_prime = (k_cumdecay[:, :, i]) @ last_recurrent_state or W_[t] @ S_[t]
-        ggml_tensor * v_prime = ggml_mul_mat(ctx0, state_t, k_cumdecay_chunk);
-
-        // v_new = v_i - v_prime or U_[t] - W_[t]*S_[t]
-        ggml_tensor * v_new = ggml_sub(ctx0, ggml_repeat(ctx0, vb_chunk, v_prime), v_prime);
-        ggml_tensor * v_new_t = ggml_cont(ctx0, ggml_transpose(ctx0, v_new));
-
-        // q_chunk [S,BT,1,H*B] gkexp_chunk [S,BT,1,H*B]
-        // attn_inter = (q_i * g[:, :, i, :, None].exp()) @ last_recurrent_state
-        // or Gamma_[t]*Q_]t] @ S
-        ggml_tensor * q_gk_exp   = ggml_mul(ctx0, q_chunk, gkexp_chunk);
-        ggml_tensor * attn_inter = ggml_mul_mat(ctx0, state_t, q_gk_exp);
-        attn_inter = ggml_scale(ctx0, attn_inter, scale); // scale q
-
-        // v_new_t [S,BT,1,H*B] Aqk [BT,BT,1,H*B]
-        // core_attn_out[:, :, i] = attn_inter + attn @ v_new or A' @ (U_[t] - W_[t]*S_[t])
-        ggml_tensor * v_attn = ggml_mul_mat(ctx0, v_new_t, Aqk_chunk);
-
-        // o[:, :, i] = (q_i * g_i.exp()) @ S + A @ v_i
-        ggml_tensor * core_attn_out_chunk = ggml_add(ctx0, attn_inter, v_attn);
-
-        core_attn_out = core_attn_out == nullptr ? core_attn_out_chunk : ggml_concat(ctx0, core_attn_out, core_attn_out_chunk, 1);
-
-        ggml_tensor * gk_cum_last =
-            ggml_cont(ctx0, ggml_view_4d(ctx0, gk_cs_chunk, gk_cs_chunk->ne[0], 1, gk_cs_chunk->ne[2], gk_cs_chunk->ne[3],
-                                        gk_cs_chunk->nb[1], gk_cs_chunk->nb[2], gk_cs_chunk->nb[3],
-                                        gk_cs_chunk->nb[1] * (gk_cs_chunk->ne[1] - 1)));
-
-        ggml_tensor * gkexp_last = ggml_exp(ctx0, ggml_cont(ctx0, ggml_transpose(ctx0, gk_cum_last)));
-
-        ggml_tensor * gk_diff = ggml_neg(ctx0, ggml_sub(ctx0, gk_cs_chunk, gk_cum_last));
-
-        ggml_tensor * gk_diff_exp = ggml_exp(ctx0, gk_diff);
-
-        ggml_tensor * key_gkdiff = ggml_mul(ctx0, k_chunk, gk_diff_exp);
-
-        // rearrange((g_i[:,:,-1:] - g_i).exp()*k_i, 'b h c k -> b h k c') @ (U_[t] - W_[t] @ S)
-        ggml_tensor * kgdmulvnew = ggml_mul_mat(ctx0, v_new_t, ggml_cont(ctx0, ggml_transpose(ctx0, key_gkdiff)));
-
-        new_state = ggml_add(ctx0,
-            ggml_mul(ctx0, new_state, ggml_reshape_4d(ctx0, gkexp_last, gkexp_last->ne[0], gkexp_last->ne[1], H_v, n_seqs)),
-            ggml_reshape_4d(ctx0, kgdmulvnew, kgdmulvnew->ne[0], kgdmulvnew->ne[1], H_v, n_seqs));
-    }
-
-    core_attn_out = ggml_cont_4d(ctx0, core_attn_out, S_v, chunk_size * n_chunks, H_v, n_seqs);
-
-    // truncate padded tokens
-    ggml_tensor * output_tokens = ggml_view_4d(ctx0, core_attn_out,
-            S_v, n_tokens, H_v, n_seqs,
-            ggml_row_size(core_attn_out->type, S_v),
-            ggml_row_size(core_attn_out->type, S_v * chunk_size * n_chunks),
-            ggml_row_size(core_attn_out->type, S_v * chunk_size * n_chunks * H_v), 0);
-    output_tokens = ggml_cont(ctx0, output_tokens);
-    // permute back to (S_v, H_v, n_tokens, n_seqs)
-    output_tokens = ggml_permute(ctx0, output_tokens, 0, 2, 1, 3);
-    output_tokens = ggml_cont(ctx0, output_tokens);
-
-    cb(new_state, "output_state", il);
-
-    return {output_tokens, new_state};
-}
-
-std::pair<ggml_tensor *, ggml_tensor *> llm_build_kimi_linear::build_kda_autoregressive(
-    ggml_tensor * q,
-    ggml_tensor * k,
-    ggml_tensor * v,
-    ggml_tensor * gk,
-    ggml_tensor * beta,
-    ggml_tensor * state,
-    int il) {
-    GGML_ASSERT(ggml_is_contiguous(v));
-    GGML_ASSERT(ggml_is_contiguous(gk));
-
-    const int64_t S_k      = q->ne[0];
-    const int64_t H_k      = q->ne[1];
-    const int64_t n_tokens = q->ne[2];
-    const int64_t n_seqs   = q->ne[3];
-
-    const int64_t S_v = v->ne[0];
-    const int64_t H_v = v->ne[1];
-
-    GGML_ASSERT(n_tokens == 1);
-    GGML_ASSERT(v->ne[2] == n_tokens);
-    GGML_ASSERT(k->ne[2] == n_tokens);
-    GGML_ASSERT(gk->ne[0] == S_k && gk->ne[1] == H_k && gk->ne[2] == n_tokens && gk->ne[3] == n_seqs);
-    GGML_ASSERT(beta->ne[0] == H_v && beta->ne[2] == n_tokens && beta->ne[3] == n_seqs);
-    GGML_ASSERT(state->ne[0] == S_v && state->ne[1] == S_k && state->ne[2] == H_v && state->ne[3] == n_seqs);
-
-    GGML_ASSERT(q->ne[0] == S_k && q->ne[1] == H_k && q->ne[2] == n_tokens && q->ne[3] == n_seqs);
-    GGML_ASSERT(k->ne[0] == S_k && k->ne[1] == H_k && k->ne[2] == n_tokens && k->ne[3] == n_seqs);
-
-    GGML_ASSERT(H_k == H_v);  // we did a repeat to make sure this is the case
-
-    const float eps_norm = hparams.f_norm_rms_eps;
-
-    q = ggml_l2_norm(ctx0, q, eps_norm);
-    k = ggml_l2_norm(ctx0, k, eps_norm);
-
-    const float scale = 1.0f / sqrtf(S_v);
-
-    q    = ggml_scale(ctx0, q, scale);
-    beta = ggml_sigmoid(ctx0, beta);
-
-    cb(q, "q_in", il);
-    cb(k, "k_in", il);
-    cb(v, "v_in", il);
-    cb(beta, "beta_in", il);
-    cb(gk, "gk_in", il);
-
-// g [H,1,B,1] g_t [1,H,B,1] => [1,1,H,B]
-// gk [S,H,1,B] => [S,1,H,B] gk_t [1,S,H,B]
-// beta [H,1,1,B] beta_t [1,H,1,B] => [1,1,H,B]
-    gk = ggml_reshape_4d(ctx0, gk, S_k, 1, H_k, n_seqs);
-    ggml_tensor * gk_t = ggml_cont(ctx0, ggml_transpose(ctx0, gk));
-    ggml_tensor * beta_t = ggml_reshape_4d(ctx0, ggml_transpose(ctx0, beta), 1, 1, H_k, n_seqs);
-
-    // Apply exponential to gk_t
-    gk_t = ggml_exp(ctx0, gk_t);
-    // Apply the gated delta rule for the single timestep
-    // last_recurrent_state = last_recurrent_state * gk_t
-    // S = S * g_i[..., None].exp()
-    state = ggml_mul(ctx0, state, gk_t);
-
-    ggml_tensor * state_t = ggml_cont(ctx0, ggml_transpose(ctx0, state));
-
-// state [S,S,H,B] k [S,1,H,B] k_state [S_v,1,H,B]
-    k = ggml_reshape_4d(ctx0, k, S_k, 1, H_k, n_seqs);
-    ggml_tensor * k_state = ggml_mul_mat(ctx0, state_t, k);
-
-    // v_i - (k_i[..., None] * S).sum(-2)
-    v = ggml_reshape_4d(ctx0, v, S_v, 1, H_v, n_seqs);
-    ggml_tensor * v_diff = ggml_sub(ctx0, v, k_state);
-
-    // b_i[..., None] * k_i
-    ggml_tensor * k_beta = ggml_mul(ctx0, k, beta_t);
-
-    // S = S + torch.einsum('b h k, b h v -> b h k v', b_i[..., None] * k_i, v_i - (k_i[..., None] * S).sum(-2))
-    // v_diff_t [1,S_v,H,B] k_beta_t [1,S_k,H,B] state [S_v,S_k,H,B]
-    state = ggml_add(ctx0, state, ggml_mul_mat(ctx0, ggml_cont(ctx0, ggml_transpose(ctx0, v_diff)), ggml_cont(ctx0, ggml_transpose(ctx0, k_beta))));
-
-    q = ggml_reshape_4d(ctx0, q, S_k, 1, H_k, n_seqs);
-    state_t = ggml_cont(ctx0, ggml_transpose(ctx0, state));
-    ggml_tensor * core_attn_out = ggml_mul_mat(ctx0, state_t, q);
-    // core_attn_out should be [S_v, 1, H_v, n_seqs] after this
-    cb(core_attn_out, "output_tokens", il);
-    cb(state, "new_state", il);
-
-    return {core_attn_out, state};
-}
-

src/models/lfm2.cpp

@@ -1,18 +1,149 @@
 #include "models.h"
 
+#include "../llama-memory-hybrid-iswa.h"
 #include "../llama-memory-hybrid.h"
 
+template <bool iswa>
+llm_build_lfm2<iswa>::llm_build_lfm2(const llama_model & model, const llm_graph_params & params) :
+    llm_graph_context(params) {
+    using inp_hybrid_type = std::conditional_t<iswa, llm_graph_input_mem_hybrid_iswa,  llm_graph_input_mem_hybrid>;
+    using inp_attn_type   = std::conditional_t<iswa, llm_graph_input_attn_kv_iswa,     llm_graph_input_attn_kv>;
+    using mem_hybrid_ctx  = std::conditional_t<iswa, llama_memory_hybrid_iswa_context, llama_memory_hybrid_context>;
 
-llm_build_lfm2::llm_build_lfm2(const llama_model & model, const llm_graph_params & params) :
-    llm_graph_context(params),
-    model(model) {
+    // lambda helpers for readability
+    auto build_dense_feed_forward = [&model, this](ggml_tensor * cur, int il) -> ggml_tensor * {
+        GGML_ASSERT(!model.layers[il].ffn_up_b);
+        GGML_ASSERT(!model.layers[il].ffn_gate_b);
+        GGML_ASSERT(!model.layers[il].ffn_down_b);
+        return build_ffn(cur,
+            model.layers[il].ffn_up, NULL, NULL,
+            model.layers[il].ffn_gate, NULL, NULL,
+            model.layers[il].ffn_down, NULL, NULL,
+            NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
+    };
+    auto build_moe_feed_forward = [&model, this](ggml_tensor * cur, int il) -> ggml_tensor * {
+        return build_moe_ffn(cur,
+                            model.layers[il].ffn_gate_inp, model.layers[il].ffn_up_exps,
+                            model.layers[il].ffn_gate_exps, model.layers[il].ffn_down_exps,
+                            model.layers[il].ffn_exp_probs_b, n_expert, n_expert_used, LLM_FFN_SILU, true, false, 0.0,
+                            static_cast<llama_expert_gating_func_type>(hparams.expert_gating_func), il);
+    };
+    auto build_attn_block = [&model, this](ggml_tensor *   cur,
+                                           ggml_tensor *   inp_pos,
+                                           inp_attn_type * inp_attn,
+                                           int             il) -> ggml_tensor * {
+        GGML_ASSERT(hparams.n_embd_v_gqa(il) == hparams.n_embd_k_gqa(il));
+        const auto n_embd_head = hparams.n_embd_head_v;
+        const auto n_head_kv   = hparams.n_head_kv(il);
+
+        auto * q = build_lora_mm(model.layers[il].wq, cur);
+        cb(q, "model.layers.{}.self_attn.q_proj", il);
+        auto * k = build_lora_mm(model.layers[il].wk, cur);
+        cb(k, "model.layers.{}.self_attn.k_proj", il);
+        auto * v = build_lora_mm(model.layers[il].wv, cur);
+        cb(v, "model.layers.{}.self_attn.v_proj", il);
+
+        q = ggml_reshape_3d(ctx0, q, n_embd_head, n_head, n_tokens);
+        k = ggml_reshape_3d(ctx0, k, n_embd_head, n_head_kv, n_tokens);
+        v = ggml_reshape_3d(ctx0, v, n_embd_head, n_head_kv, n_tokens);
+
+        // qk norm
+        q = build_norm(q, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);
+        cb(q, "model.layers.{}.self_attn.q_layernorm", il);
+        k = build_norm(k, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il);
+        cb(k, "model.layers.{}.self_attn.k_layernorm", il);
+
+        // RoPE
+        q = ggml_rope_ext(ctx0, q, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, ext_factor,
+                          attn_factor, beta_fast, beta_slow);
+        k = ggml_rope_ext(ctx0, k, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, ext_factor,
+                          attn_factor, beta_fast, beta_slow);
+
+        cur = build_attn(inp_attn,
+                model.layers[il].wo, NULL,
+                q, k, v, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il);
+
+        cb(cur, "model.layers.{}.self_attn.out_proj", il);
+
+        return cur;
+    };
+    auto build_shortconv_block = [&model, this](ggml_tensor *        cur,
+                                                llm_graph_input_rs * inp_recr,
+                                                int                  il) -> ggml_tensor * {
+        const auto * mctx_cur = static_cast<const mem_hybrid_ctx *>(mctx)->get_recr();
+        const uint32_t kv_head      = mctx_cur->get_head();
+        const int64_t  n_seq_tokens = ubatch.n_seq_tokens;
+        const int64_t  n_seqs       = ubatch.n_seqs;
+        GGML_ASSERT(n_seqs != 0);
+        GGML_ASSERT(ubatch.equal_seqs());
+        GGML_ASSERT(ubatch.n_tokens == n_seq_tokens * n_seqs);
+
+        GGML_ASSERT(hparams.n_shortconv_l_cache > 1);
+        const uint32_t d_conv = hparams.n_shortconv_l_cache - 1;
+
+        // {n_embd, n_tokens} => {n_embd, n_seq_tokens, n_seqs}
+        cur = ggml_reshape_3d(ctx0, cur, cur->ne[0], n_seq_tokens, n_seqs);
+
+        auto * bcx = build_lora_mm(model.layers[il].shortconv.in_proj, cur);
+        cb(bcx, "model.layers.{}.conv.in_proj", il);
+
+        constexpr auto n_chunks = 3;
+        GGML_ASSERT(bcx->ne[0] % n_chunks == 0);
+        const auto chunk_size = bcx->ne[0] / n_chunks;
+        auto *     b          = ggml_view_3d(ctx0, bcx, chunk_size, bcx->ne[1], bcx->ne[2], bcx->nb[1], bcx->nb[2],
+                                             0 * chunk_size * ggml_element_size(bcx));
+        auto *     c          = ggml_view_3d(ctx0, bcx, chunk_size, bcx->ne[1], bcx->ne[2], bcx->nb[1], bcx->nb[2],
+                                             1 * chunk_size * ggml_element_size(bcx));
+        auto *     x          = ggml_view_3d(ctx0, bcx, chunk_size, bcx->ne[1], bcx->ne[2], bcx->nb[1], bcx->nb[2],
+                                             2 * chunk_size * ggml_element_size(bcx));
+
+        auto * bx = ggml_transpose(ctx0, ggml_mul(ctx0, b, x));
+
+        // read conv state
+        auto * conv_state = mctx_cur->get_r_l(il);
+        auto * conv_rs    = build_rs(inp_recr, conv_state, hparams.n_embd_r(), n_seqs);
+        auto * conv       = ggml_reshape_3d(ctx0, conv_rs, d_conv, hparams.n_embd, n_seqs);
+
+        bx = ggml_concat(ctx0, conv, bx, 0);
+        GGML_ASSERT(bx->ne[0] > conv->ne[0]);
+
+        // last d_conv columns is a new conv state
+        auto * new_conv = ggml_view_3d(ctx0, bx, conv->ne[0], bx->ne[1], bx->ne[2], bx->nb[1], bx->nb[2],
+                                       (bx->ne[0] - conv->ne[0]) * ggml_element_size(bx));
+        GGML_ASSERT(ggml_are_same_shape(conv, new_conv));
+
+        // write new conv conv state
+        ggml_build_forward_expand(gf, ggml_cpy(ctx0, new_conv,
+                                               ggml_view_1d(ctx0, conv_state, ggml_nelements(new_conv),
+                                                            kv_head * d_conv * n_embd * ggml_element_size(new_conv))));
+
+        auto * conv_kernel = model.layers[il].shortconv.conv;
+        auto * conv_out    = ggml_ssm_conv(ctx0, bx, conv_kernel);
+        cb(conv_out, "model.layers.{}.conv.conv", il);
+
+        auto * y = ggml_mul(ctx0, c, conv_out);
+        y        = build_lora_mm(model.layers[il].shortconv.out_proj, y);
+        cb(y, "model.layers.{}.conv.out_proj", il);
+        // {n_embd, n_seq_tokens, n_seqs} => {n_embd, n_tokens}
+        y = ggml_reshape_2d(ctx0, y, y->ne[0], n_seq_tokens * n_seqs);
+
+        return y;
+    };
+
+    // actual graph construction starts here
     ggml_tensor * cur = build_inp_embd(model.tok_embd);
     cb(cur, "model.embed_tokens", -1);
 
     ggml_build_forward_expand(gf, cur);
 
+    inp_hybrid_type * inp_hybrid = nullptr;
+    if constexpr (iswa) {
+        inp_hybrid = build_inp_mem_hybrid_iswa();
+    } else {
+        inp_hybrid = build_inp_mem_hybrid();
+    }
+
     ggml_tensor * inp_pos     = build_inp_pos();
-    auto *        inp_hybrid  = build_inp_mem_hybrid();
     ggml_tensor * inp_out_ids = build_inp_out_ids();
 
     for (int il = 0; il < n_layer; ++il) {
@@ -54,122 +185,6 @@ llm_build_lfm2::llm_build_lfm2(const llama_model & model, const llm_graph_params
     ggml_build_forward_expand(gf, cur);
 }
 
-ggml_tensor * llm_build_lfm2::build_moe_feed_forward(ggml_tensor * cur, int il) const {
-    return build_moe_ffn(cur,
-                        model.layers[il].ffn_gate_inp, model.layers[il].ffn_up_exps,
-                        model.layers[il].ffn_gate_exps, model.layers[il].ffn_down_exps,
-                        model.layers[il].ffn_exp_probs_b, n_expert, n_expert_used, LLM_FFN_SILU, true, false, 0.0,
-                        static_cast<llama_expert_gating_func_type>(hparams.expert_gating_func), il);
-}
-
-ggml_tensor * llm_build_lfm2::build_dense_feed_forward(ggml_tensor * cur, int il) const {
-    GGML_ASSERT(!model.layers[il].ffn_up_b);
-    GGML_ASSERT(!model.layers[il].ffn_gate_b);
-    GGML_ASSERT(!model.layers[il].ffn_down_b);
-    return build_ffn(cur,
-        model.layers[il].ffn_up, NULL, NULL,
-        model.layers[il].ffn_gate, NULL, NULL,
-        model.layers[il].ffn_down, NULL, NULL,
-        NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
-}
-
-ggml_tensor * llm_build_lfm2::build_attn_block(ggml_tensor *             cur,
-                                               ggml_tensor *             inp_pos,
-                                               llm_graph_input_attn_kv * inp_attn,
-                                               int                       il) const {
-    GGML_ASSERT(hparams.n_embd_v_gqa(il) == hparams.n_embd_k_gqa(il));
-    const auto n_embd_head = hparams.n_embd_head_v;
-    const auto n_head_kv   = hparams.n_head_kv(il);
-
-    auto * q = build_lora_mm(model.layers[il].wq, cur);
-    cb(q, "model.layers.{}.self_attn.q_proj", il);
-    auto * k = build_lora_mm(model.layers[il].wk, cur);
-    cb(k, "model.layers.{}.self_attn.k_proj", il);
-    auto * v = build_lora_mm(model.layers[il].wv, cur);
-    cb(v, "model.layers.{}.self_attn.v_proj", il);
-
-    q = ggml_reshape_3d(ctx0, q, n_embd_head, n_head, n_tokens);
-    k = ggml_reshape_3d(ctx0, k, n_embd_head, n_head_kv, n_tokens);
-    v = ggml_reshape_3d(ctx0, v, n_embd_head, n_head_kv, n_tokens);
-
-    // qk norm
-    q = build_norm(q, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);
-    cb(q, "model.layers.{}.self_attn.q_layernorm", il);
-    k = build_norm(k, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il);
-    cb(k, "model.layers.{}.self_attn.k_layernorm", il);
-
-    // RoPE
-    q = ggml_rope_ext(ctx0, q, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, ext_factor,
-                      attn_factor, beta_fast, beta_slow);
-    k = ggml_rope_ext(ctx0, k, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, ext_factor,
-                      attn_factor, beta_fast, beta_slow);
-
-    cur = build_attn(inp_attn,
-            model.layers[il].wo, NULL,
-            q, k, v, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il);
-
-    cb(cur, "model.layers.{}.self_attn.out_proj", il);
-
-    return cur;
-}
-
-ggml_tensor * llm_build_lfm2::build_shortconv_block(ggml_tensor * cur, llm_graph_input_rs * inp_recr, int il) {
-    const auto *   mctx_cur     = static_cast<const llama_memory_hybrid_context *>(mctx)->get_recr();
-    const uint32_t kv_head      = mctx_cur->get_head();
-    const int64_t  n_seq_tokens = ubatch.n_seq_tokens;
-    const int64_t  n_seqs       = ubatch.n_seqs;
-    GGML_ASSERT(n_seqs != 0);
-    GGML_ASSERT(ubatch.equal_seqs());
-    GGML_ASSERT(ubatch.n_tokens == n_seq_tokens * n_seqs);
-
-    GGML_ASSERT(hparams.n_shortconv_l_cache > 1);
-    const uint32_t d_conv = hparams.n_shortconv_l_cache - 1;
-
-    // {n_embd, n_tokens} => {n_embd, n_seq_tokens, n_seqs}
-    cur = ggml_reshape_3d(ctx0, cur, cur->ne[0], n_seq_tokens, n_seqs);
-
-    auto * bcx = build_lora_mm(model.layers[il].shortconv.in_proj, cur);
-    cb(bcx, "model.layers.{}.conv.in_proj", il);
-
-    constexpr auto n_chunks = 3;
-    GGML_ASSERT(bcx->ne[0] % n_chunks == 0);
-    const auto chunk_size = bcx->ne[0] / n_chunks;
-    auto *     b          = ggml_view_3d(ctx0, bcx, chunk_size, bcx->ne[1], bcx->ne[2], bcx->nb[1], bcx->nb[2],
-                                         0 * chunk_size * ggml_element_size(bcx));
-    auto *     c          = ggml_view_3d(ctx0, bcx, chunk_size, bcx->ne[1], bcx->ne[2], bcx->nb[1], bcx->nb[2],
-                                         1 * chunk_size * ggml_element_size(bcx));
-    auto *     x          = ggml_view_3d(ctx0, bcx, chunk_size, bcx->ne[1], bcx->ne[2], bcx->nb[1], bcx->nb[2],
-                                         2 * chunk_size * ggml_element_size(bcx));
-
-    auto * bx = ggml_transpose(ctx0, ggml_mul(ctx0, b, x));
-
-    // read conv state
-    auto * conv_state = mctx_cur->get_r_l(il);
-    auto * conv_rs    = build_rs(inp_recr, conv_state, hparams.n_embd_r(), n_seqs);
-    auto * conv       = ggml_reshape_3d(ctx0, conv_rs, d_conv, hparams.n_embd, n_seqs);
-
-    bx = ggml_concat(ctx0, conv, bx, 0);
-    GGML_ASSERT(bx->ne[0] > conv->ne[0]);
-
-    // last d_conv columns is a new conv state
-    auto * new_conv = ggml_view_3d(ctx0, bx, conv->ne[0], bx->ne[1], bx->ne[2], bx->nb[1], bx->nb[2],
-                                   (bx->ne[0] - conv->ne[0]) * ggml_element_size(bx));
-    GGML_ASSERT(ggml_are_same_shape(conv, new_conv));
-
-    // write new conv conv state
-    ggml_build_forward_expand(gf, ggml_cpy(ctx0, new_conv,
-                                           ggml_view_1d(ctx0, conv_state, ggml_nelements(new_conv),
-                                                        kv_head * d_conv * n_embd * ggml_element_size(new_conv))));
-
-    auto * conv_kernel = model.layers[il].shortconv.conv;
-    auto * conv_out    = ggml_ssm_conv(ctx0, bx, conv_kernel);
-    cb(conv_out, "model.layers.{}.conv.conv", il);
-
-    auto * y = ggml_mul(ctx0, c, conv_out);
-    y        = build_lora_mm(model.layers[il].shortconv.out_proj, y);
-    cb(y, "model.layers.{}.conv.out_proj", il);
-    // {n_embd, n_seq_tokens, n_seqs} => {n_embd, n_tokens}
-    y = ggml_reshape_2d(ctx0, y, y->ne[0], n_seq_tokens * n_seqs);
-
-    return y;
-}
+// Explicit template instantiations
+template struct llm_build_lfm2<true>;
+template struct llm_build_lfm2<false>;

src/models/models.h

@@ -190,6 +190,10 @@ struct llm_build_ernie4_5_moe : public llm_graph_context {
     llm_build_ernie4_5_moe(const llama_model & model, const llm_graph_params & params);
 };
 
+struct llm_build_paddleocr : public llm_graph_context {
+    llm_build_paddleocr(const llama_model & model, const llm_graph_params & params);
+};
+
 template <bool iswa>
 struct llm_build_exaone4 : public llm_graph_context {
     llm_build_exaone4(const llama_model & model, const llm_graph_params & params);
@@ -316,12 +320,15 @@ struct llm_build_jais : public llm_graph_context {
     llm_build_jais(const llama_model & model, const llm_graph_params & params);
 };
 
+struct llm_build_jais2 : public llm_graph_context {
+    llm_build_jais2(const llama_model & model, const llm_graph_params & params);
+};
+
 struct llm_build_jamba : public llm_build_mamba_base {
     llm_build_jamba(const llama_model & model, const llm_graph_params & params);
 };
 
-// TODO: derive llm_build_delta_net_base instead
-struct llm_build_kimi_linear : public llm_build_mamba_base {
+struct llm_build_kimi_linear : public llm_build_delta_net_base {
     llm_build_kimi_linear(const llama_model & model, const llm_graph_params & params);
 
     std::pair<ggml_tensor *, ggml_tensor *> build_kda_autoregressive(
@@ -348,15 +355,9 @@ struct llm_build_kimi_linear : public llm_build_mamba_base {
     const llama_model & model;
 };
 
+template <bool iswa>
 struct llm_build_lfm2 : public llm_graph_context {
-    const llama_model & model;
-
     llm_build_lfm2(const llama_model & model, const llm_graph_params & params);
-    ggml_tensor * build_moe_feed_forward(ggml_tensor * cur, int il) const;
-    ggml_tensor * build_dense_feed_forward(ggml_tensor * cur, int il) const;
-    ggml_tensor * build_attn_block(ggml_tensor * cur, ggml_tensor * inp_pos, llm_graph_input_attn_kv * inp_attn, int il) const;
-    ggml_tensor * build_shortconv_block(ggml_tensor * cur, llm_graph_input_rs * inp_recr, int il);
-
 };
 
 struct llm_build_llada : public llm_graph_context {
@@ -542,8 +543,7 @@ private:
     const llama_model & model;
 };
 
-// TODO: derive llm_build_delta_net_base instead
-struct llm_build_qwen35 : public llm_graph_context {
+struct llm_build_qwen35 : public llm_build_delta_net_base {
     llm_build_qwen35(const llama_model & model, const llm_graph_params & params);
 private:
     ggml_tensor * build_layer_attn(
@@ -556,39 +556,12 @@ private:
     ggml_tensor * build_layer_attn_linear(
          llm_graph_input_rs * inp,
                 ggml_tensor * cur,
-                ggml_tensor * causal_mask,
-                ggml_tensor * identity,
-                ggml_tensor * diag_mask,
                         int   il);
 
-
     ggml_tensor * build_layer_ffn(
                 ggml_tensor * cur,
                         int   il);
 
-    // returns pair of output and new state
-    std::pair<ggml_tensor *, ggml_tensor *> build_delta_net_chunking(
-                ggml_tensor * q,
-                ggml_tensor * k,
-                ggml_tensor * v,
-                ggml_tensor * g,
-                ggml_tensor * beta,
-                ggml_tensor * state,
-                ggml_tensor * causal_mask,
-                ggml_tensor * identity,
-                ggml_tensor * diag_mask,
-                        int   il);
-
-    // returns pair of output and new state
-    std::pair<ggml_tensor *, ggml_tensor *> build_delta_net_autoregressive(
-                ggml_tensor * q,
-                ggml_tensor * k,
-                ggml_tensor * v,
-                ggml_tensor * g,
-                ggml_tensor * beta,
-                ggml_tensor * state,
-                int           il);
-
     ggml_tensor * build_norm_gated(
                 ggml_tensor * input,
                 ggml_tensor * weights,
@@ -604,7 +577,7 @@ private:
 };
 
 // TODO: derive llm_build_delta_net_base instead
-struct llm_build_qwen35moe : public llm_graph_context {
+struct llm_build_qwen35moe : public llm_build_delta_net_base {
     llm_build_qwen35moe(const llama_model & model, const llm_graph_params & params);
 private:
     ggml_tensor * build_layer_attn(
@@ -617,38 +590,12 @@ private:
     ggml_tensor * build_layer_attn_linear(
          llm_graph_input_rs * inp,
                 ggml_tensor * cur,
-                ggml_tensor * causal_mask,
-                ggml_tensor * identity,
-                ggml_tensor * diag_mask,
                         int   il);
 
     ggml_tensor * build_layer_ffn(
                 ggml_tensor * cur,
                         int   il);
 
-    // returns pair of output and new state
-    std::pair<ggml_tensor *, ggml_tensor *> build_delta_net_chunking(
-                ggml_tensor * q,
-                ggml_tensor * k,
-                ggml_tensor * v,
-                ggml_tensor * g,
-                ggml_tensor * beta,
-                ggml_tensor * state,
-                ggml_tensor * causal_mask,
-                ggml_tensor * identity,
-                ggml_tensor * diag_mask,
-                        int   il);
-
-    // returns pair of output and new state
-    std::pair<ggml_tensor *, ggml_tensor *> build_delta_net_autoregressive(
-                ggml_tensor * q,
-                ggml_tensor * k,
-                ggml_tensor * v,
-                ggml_tensor * g,
-                ggml_tensor * beta,
-                ggml_tensor * state,
-                int           il);
-
     ggml_tensor * build_norm_gated(
                 ggml_tensor * input,
                 ggml_tensor * weights,

src/models/modern-bert.cpp

@@ -104,13 +104,6 @@ llm_build_modern_bert::llm_build_modern_bert(const llama_model & model, const ll
             LLM_NORM, -1);
     cb(cur, "final_norm_out", -1);
 
-    if (hparams.pooling_type == LLAMA_POOLING_TYPE_CLS) {
-        // extracting cls token
-        cur = ggml_view_1d(ctx0, cur, hparams.n_embd, 0);
-        cb(cur, "cls_pooled_embd", -1);
-    }
-
-    cb(cur, "res_embd", -1);
     res->t_embd = cur;
     ggml_build_forward_expand(gf, cur);
 }

src/models/paddleocr.cpp

@@ -0,0 +1,122 @@
+#include "models.h"
+
+llm_build_paddleocr::llm_build_paddleocr(const llama_model & model, const llm_graph_params & params) :
+    llm_graph_context(params) {
+
+    // NOTE: same with qwen2vl.cpp, but bias tensors are optional
+
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    int sections[4];
+    std::copy(std::begin(hparams.rope_sections), std::begin(hparams.rope_sections) + 4, sections);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        // norm
+        {
+            cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
+            cb(cur, "attn_norm", il);
+        }
+        // self-attention
+        {
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            cb(Qcur, "Qcur", il);
+            if (model.layers[il].bq) {
+                Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+                cb(Qcur, "Qcur", il);
+            }
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            cb(Kcur, "Kcur", il);
+            if (model.layers[il].bk) {
+                Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+                cb(Kcur, "Kcur", il);
+            }
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            cb(Vcur, "Vcur", il);
+            if (model.layers[il].bv) {
+                Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+                cb(Vcur, "Vcur", il);
+            }
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            Qcur = ggml_rope_multi(
+                    ctx0, Qcur, inp_pos, nullptr,
+                    n_rot, sections, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            Kcur = ggml_rope_multi(
+                    ctx0, Kcur, inp_pos, nullptr,
+                    n_rot, sections, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, model.layers[il].bo,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+        }
+        if (il == n_layer - 1) {
+            // skip computing output for unused tokens
+            cur   = ggml_get_rows(ctx0, cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        // feed-forward network
+        {
+            cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
+            cb(cur, "ffn_norm", il);
+
+            cur = build_ffn(cur,
+                    model.layers[il].ffn_up, NULL, NULL,
+                    model.layers[il].ffn_gate, NULL, NULL,
+                    model.layers[il].ffn_down, NULL, NULL,
+                    NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
+            cb(cur, "ffn_out", il);
+        }
+        cur = ggml_add(ctx0, cur, ffn_inp);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+    cur = inpL;
+
+    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}

src/models/qwen35.cpp

@@ -2,10 +2,8 @@
 
 #include "llama-memory-recurrent.h"
 
-#define CHUNK_SIZE 64
-
 llm_build_qwen35::llm_build_qwen35(const llama_model & model, const llm_graph_params & params) :
-    llm_graph_context(params), model(model) {
+    llm_build_delta_net_base(params), model(model) {
     const int64_t n_embd_head = hparams.n_embd_head_v;
 
     GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
@@ -25,17 +23,6 @@ llm_build_qwen35::llm_build_qwen35(const llama_model & model, const llm_graph_pa
     ggml_tensor * inp_pos     = build_inp_pos();
     ggml_tensor * inp_out_ids = build_inp_out_ids();
 
-    ggml_tensor * causal_mask =
-        ggml_tri(ctx0, ggml_fill(ctx0, ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, CHUNK_SIZE, CHUNK_SIZE), 1.0f),
-                    GGML_TRI_TYPE_LOWER);
-
-    ggml_tensor * identity = ggml_diag(ctx0, ggml_fill(ctx0, ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, CHUNK_SIZE), 1.0f));
-    ggml_tensor * diag_mask = ggml_add(ctx0, causal_mask, identity);
-
-    ggml_build_forward_expand(gf, causal_mask);
-    ggml_build_forward_expand(gf, identity);
-    ggml_build_forward_expand(gf, diag_mask);
-
     for (int il = 0; il < n_layer; ++il) {
         ggml_tensor * inpSA = inpL;
 
@@ -45,7 +32,7 @@ llm_build_qwen35::llm_build_qwen35(const llama_model & model, const llm_graph_pa
         // Determine layer type and build appropriate attention mechanism
         if (hparams.is_recurrent(il)) {
             // Linear attention layer (gated delta net)
-            cur = build_layer_attn_linear(inp->get_recr(), cur, causal_mask, identity, diag_mask, il);
+            cur = build_layer_attn_linear(inp->get_recr(), cur, il);
         } else {
             // Full attention layer
             cur = build_layer_attn(inp->get_attn(), cur, inp_pos, sections, il);
@@ -95,361 +82,6 @@ llm_build_qwen35::llm_build_qwen35(const llama_model & model, const llm_graph_pa
     ggml_build_forward_expand(gf, cur);
 }
 
-// utility to get one slice from the third dimension
-// input dim:  [x, y, c, b]
-// output dim: [x, y, 1, b]
-static ggml_tensor * get_slice_2d(ggml_context * ctx0, ggml_tensor * t, int64_t c) {
-    return ggml_view_4d(ctx0, t, t->ne[0], t->ne[1], 1, t->ne[3],
-        t->nb[1], t->nb[2], t->nb[3], t->nb[2] * c);
-}
-
-std::pair<ggml_tensor *, ggml_tensor *> llm_build_qwen35::build_delta_net_chunking(
-        ggml_tensor * q,
-        ggml_tensor * k,
-        ggml_tensor * v,
-        ggml_tensor * g,
-        ggml_tensor * beta,
-        ggml_tensor * state,
-        ggml_tensor * causal_mask,
-        ggml_tensor * identity,
-        ggml_tensor * diag_mask,
-        int           il) {
-    const int64_t S_k      = q->ne[0];
-    const int64_t H_k      = q->ne[1];
-    const int64_t n_tokens = q->ne[2];
-    const int64_t n_seqs   = q->ne[3];
-
-    const int64_t S_v = v->ne[0];
-    const int64_t H_v = v->ne[1];
-
-    GGML_ASSERT(v->ne[2] == n_tokens);
-    GGML_ASSERT(k->ne[2] == n_tokens);
-    GGML_ASSERT(g->ne[0] == H_v && g->ne[1] == n_tokens && g->ne[2] == n_seqs);
-    GGML_ASSERT(beta->ne[0] == H_v && beta->ne[2] == n_tokens && beta->ne[3] == n_seqs);
-    GGML_ASSERT(state->ne[0] == S_v && state->ne[1] == S_v * H_v && state->ne[2] == 1 && state->ne[3] == n_seqs);
-
-    GGML_ASSERT(q->ne[0] == S_k && q->ne[1] == H_k && q->ne[2] == n_tokens && q->ne[3] == n_seqs);
-    GGML_ASSERT(k->ne[0] == S_k && k->ne[1] == H_k && k->ne[2] == n_tokens && k->ne[3] == n_seqs);
-
-    GGML_ASSERT(H_k == H_v);  // we did a repeat to make sure this is the case
-
-    const float eps_norm = hparams.f_norm_rms_eps;
-
-    q = ggml_l2_norm(ctx0, q, eps_norm);
-    k = ggml_l2_norm(ctx0, k, eps_norm);
-
-    const float scale = 1.0f / sqrtf(S_v);
-
-    q = ggml_scale(ctx0, q, scale);
-
-    beta = ggml_sigmoid(ctx0, beta);
-
-    cb(q, "q_in", il);
-    cb(k, "k_in", il);
-    cb(v, "v_in", il);
-    cb(beta, "beta_in", il);
-    cb(g, "g_in", il);
-
-    q = ggml_cont_4d(ctx0, ggml_permute(ctx0, q, 0, 2, 1, 3), S_v, n_tokens, H_v, n_seqs);
-    k = ggml_cont_4d(ctx0, ggml_permute(ctx0, k, 0, 2, 1, 3), S_v, n_tokens, H_v, n_seqs);
-    v = ggml_cont_4d(ctx0, ggml_permute(ctx0, v, 0, 2, 1, 3), S_v, n_tokens, H_v, n_seqs);
-    g = ggml_cont_4d(ctx0, ggml_permute(ctx0, g, 2, 0, 3, 1), n_tokens, 1, H_k, n_seqs);
-
-    beta  = ggml_cont(ctx0, ggml_permute(ctx0, beta, 2, 0, 1, 3));
-    state = ggml_reshape_4d(ctx0, state, S_v, S_v, H_v, n_seqs);
-
-    cb(q, "q_perm", il);
-    cb(k, "k_perm", il);
-    cb(v, "v_perm", il);
-    cb(beta, "beta_perm", il);
-    cb(g, "g_perm", il);
-    cb(state, "state_in", il);
-
-    GGML_ASSERT(q->ne[1] == n_tokens && q->ne[0] == S_k && q->ne[2] == H_k && q->ne[3] == n_seqs);
-    GGML_ASSERT(k->ne[1] == n_tokens && k->ne[0] == S_k && k->ne[2] == H_k && k->ne[3] == n_seqs);
-    GGML_ASSERT(v->ne[1] == n_tokens && v->ne[0] == S_v && v->ne[2] == H_k && v->ne[3] == n_seqs);
-    GGML_ASSERT(beta->ne[1] == n_tokens && beta->ne[2] == H_k && beta->ne[0] == 1 && beta->ne[3] == n_seqs);
-
-    // Do padding
-    const int64_t chunk_size = CHUNK_SIZE;
-
-    const int64_t pad = (chunk_size - n_tokens % chunk_size) % chunk_size;
-    const int64_t n_chunks = (n_tokens + pad) / chunk_size;
-
-    q = ggml_pad(ctx0, q, 0, pad, 0, 0);
-    k = ggml_pad(ctx0, k, 0, pad, 0, 0);
-    v = ggml_pad(ctx0, v, 0, pad, 0, 0);
-    g = ggml_pad(ctx0, g, pad, 0, 0, 0);
-    beta = ggml_pad(ctx0, beta, 0, pad, 0, 0);
-
-    cb(q, "q_pad", il);
-    cb(k, "k_pad", il);
-    cb(v, "v_pad", il);
-    cb(beta, "beta_pad", il);
-    cb(g, "g_pad", il);
-
-    ggml_tensor * v_beta = ggml_mul(ctx0, v, beta);
-    ggml_tensor * k_beta = ggml_mul(ctx0, k, beta);
-
-    cb(v_beta, "v_beta", il);
-    cb(k_beta, "k_beta", il);
-
-    q      = ggml_reshape_4d(ctx0, q,      S_k, chunk_size, n_chunks, H_k * n_seqs);
-    k      = ggml_reshape_4d(ctx0, k,      S_k, chunk_size, n_chunks, H_k * n_seqs);
-    k_beta = ggml_reshape_4d(ctx0, k_beta, S_k, chunk_size, n_chunks, H_k * n_seqs);
-    v      = ggml_reshape_4d(ctx0, v,      S_v, chunk_size, n_chunks, H_v * n_seqs);
-    v_beta = ggml_reshape_4d(ctx0, v_beta, S_v, chunk_size, n_chunks, H_v * n_seqs);
-
-    g    = ggml_reshape_4d(ctx0, g, chunk_size, 1, n_chunks, H_k * n_seqs);
-    beta = ggml_reshape_4d(ctx0, beta, 1, chunk_size, n_chunks, H_k * n_seqs);
-
-    ggml_tensor * g_cumsum = ggml_cumsum(ctx0, g);
-    cb(g_cumsum, "g_cumsum", il); // shape: (chunk_size, 1, n_chunks, H_v * n_seqs)
-
-    ggml_tensor * gcs_i = g_cumsum; // ggml_reshape_4d(ctx0, g_cumsum, chunk_size, 1, n_chunks, H_v * n_seqs);
-    ggml_tensor * gcs_j = ggml_reshape_4d(ctx0, g_cumsum, 1, chunk_size, n_chunks, H_v * n_seqs);
-
-    ggml_tensor * gcs_j_broadcast =
-        ggml_repeat_4d(ctx0, gcs_j, chunk_size, chunk_size, n_chunks, H_v * n_seqs);
-
-    ggml_tensor * decay_mask = ggml_sub(ctx0, gcs_j_broadcast, gcs_i);
-    cb(decay_mask, "decay_mask", il); // shape: (chunk_size, chunk_size, n_chunks, H_v * n_seqs)
-
-    decay_mask = ggml_mul(ctx0, decay_mask, diag_mask);
-    decay_mask = ggml_exp(ctx0, decay_mask);
-    decay_mask = ggml_mul(ctx0, decay_mask, diag_mask);
-
-    ggml_tensor * kmulkbeta = ggml_mul_mat(ctx0, k, k_beta);
-
-    ggml_tensor * k_decay = ggml_mul(ctx0, kmulkbeta, decay_mask);
-    ggml_tensor * attn    = ggml_neg(ctx0, ggml_mul(ctx0, k_decay, causal_mask));
-    cb(attn, "attn_pre_solve", il); // shape: (chunk_size, chunk_size, n_chunks, H_v * n_seqs)
-
-    ggml_tensor * attn_lower = ggml_mul(ctx0, attn, causal_mask);
-    ggml_tensor * lhs        = ggml_sub(ctx0, ggml_repeat(ctx0, identity, attn_lower), attn_lower);
-
-    ggml_tensor * lin_solve  = ggml_solve_tri(ctx0, lhs, attn, true, true, false);
-    attn                     = ggml_mul(ctx0, lin_solve, causal_mask);
-    attn                     = ggml_add(ctx0, attn, identity);
-    cb(attn, "attn_solved", il); // shape: (chunk_size, chunk_size, n_chunks, H_v * n_seqs)
-
-    v = ggml_mul_mat(ctx0, ggml_cont(ctx0, ggml_transpose(ctx0, v_beta)), attn);
-
-    ggml_tensor * g_cumsum_t = ggml_cont(ctx0, ggml_transpose(ctx0, g_cumsum));
-    ggml_tensor * gexp       = ggml_exp(ctx0, g_cumsum_t);
-
-    ggml_tensor * kbeta_gexp = ggml_mul(ctx0, k_beta, gexp);
-    cb(kbeta_gexp, "kbeta_gexp", il); // shape: (S_k, chunk_size, n_chunks, H_v * n_seqs)
-
-    ggml_tensor * k_cumdecay =
-        ggml_cont(ctx0, ggml_transpose(ctx0, ggml_mul_mat(ctx0, attn, ggml_cont(ctx0, ggml_transpose(ctx0, kbeta_gexp)))));
-    cb(k_cumdecay, "k_cumdecay", il); // shape: (chunk_size, chunk_size, n_chunks, H_v * n_seqs)
-
-    ggml_tensor * attn_kq = ggml_mul_mat(ctx0, k, q);
-    attn_kq = ggml_mul(ctx0, attn_kq, decay_mask);
-    attn_kq = ggml_mul(ctx0, attn_kq, diag_mask);
-    cb(attn_kq, "attn_kq", il); // shape: (chunk_size, chunk_size, n_chunks, H_v * n_seqs)
-
-
-    // vectorized calculation of key_gdiff
-    // improved from the chunked version:
-    //   g_last = torch.clamp(g_cum[:, :, -1], max=50.0).exp().unsqueeze(-1).unsqueeze(-1)
-    //   g_diff = torch.clamp(g_cum[:, :, -1:] - g_cum, max=50.0).exp()
-    //   key_gdiff = key * g_diff.unsqueeze(-1)
-    //   kgdmulvnew = (key_gdiff).transpose(-1, -2) @ v_new
-    //   last_recurrent_state = last_recurrent_state * g_last + kgdmulvnew
-
-    // get last element in g_cumsum along chunk_size dimension (ne0)
-    // example: [[x, y, z, ..., last], ...] -> [[last], ...]
-    ggml_tensor * g_last = ggml_view_4d(ctx0, g_cumsum, 1, 1, g_cumsum->ne[2], g_cumsum->ne[3],
-                                        g_cumsum->nb[1], g_cumsum->nb[2], g_cumsum->nb[3],
-                                        (g_cumsum->ne[0] - 1) * ggml_element_size(g_cumsum));
-    g_last = ggml_cont(ctx0, g_last);
-    cb(g_last, "g_last", il); // shape: (1, 1, n_chunks, H_v * n_seqs)
-
-    ggml_tensor * g_last_exp = ggml_exp(ctx0, g_last);
-    cb(g_last_exp, "g_last_exp", il); // shape: (1, 1, n_chunks, H_v * n_seqs)
-
-    ggml_tensor * g_diff = ggml_neg(ctx0, ggml_sub(ctx0, g_cumsum, g_last));
-    cb(g_diff, "g_diff", il); // shape: (chunk_size, 1, n_chunks, H_v * n_seqs)
-
-    ggml_tensor * g_diff_exp = ggml_exp(ctx0, g_diff);
-    ggml_tensor * g_diff_exp_t = ggml_reshape_4d(ctx0, g_diff_exp,
-                                                 1, chunk_size, n_chunks, g_diff_exp->ne[3]);
-
-    ggml_tensor * key_gdiff = ggml_mul(ctx0, k, g_diff_exp_t);
-    cb(key_gdiff, "key_gdiff", il); // shape: (S_k, chunk_size, n_chunks, H_v * n_seqs)
-
-    ggml_tensor * key_gdiff_t = ggml_cont(ctx0, ggml_transpose(ctx0, key_gdiff));
-    cb(key_gdiff_t, "key_gdiff_t", il); // shape: (chunk_size, S_k, n_chunks, H_v * n_seqs)
-
-    // state to be updated per chunk
-    ggml_tensor * new_state = state; // ggml_dup(ctx0, state);
-    cb(new_state, "new_state", il); // shape: (S_v, S_v, H_v, n_seqs)
-
-    // shape after loop of chunks: (S_v, chunk_size, n_chunks, H_v * n_seqs)
-    ggml_tensor * core_attn_out = nullptr;
-
-    for (int64_t chunk = 0; chunk < n_chunks; chunk++) {
-        // shape: (S_k, chunk_size, 1, H_k * n_seqs)
-        ggml_tensor * q_chunk = get_slice_2d(ctx0, q, chunk); // (no cont), next op: ggml_mul
-
-        // shape: (S_v, chunk_size, 1, H_v * n_seqs)
-        ggml_tensor * v_chunk = get_slice_2d(ctx0, v, chunk); // (no cont), next op: ggml_repeat
-
-        // shape: (chunk_size, 1, n_chunks, H_v * n_seqs)
-        ggml_tensor * gexp_chunk = get_slice_2d(ctx0, gexp, chunk); // (no cont), next op: ggml_mul
-
-        // shape: (chunk_size, 1, H_v * n_seqs)
-        ggml_tensor * k_cumdecay_chunk = get_slice_2d(ctx0, k_cumdecay, chunk); // (no cont), next op: ggml_mul_mat
-
-        // attn = (q_i @ k_i.transpose(-1, -2) * decay_mask[:, :, i]).masked_fill_(mask, 0)
-        // replaced by precomputed attn_kq
-        ggml_tensor * attn_chunk = get_slice_2d(ctx0, attn_kq, chunk);
-        cb(attn_chunk, "attn_chunk", il);
-
-        ggml_tensor * state_t = ggml_cont_4d(ctx0, ggml_permute(ctx0, new_state, 1, 0, 2, 3), S_v, S_v, 1, H_v * n_seqs);
-
-        // v_prime = (k_cumdecay[:, :, i]) @ last_recurrent_state
-        ggml_tensor * v_prime = ggml_mul_mat(ctx0, state_t, k_cumdecay_chunk);
-        cb(v_prime, "v_prime_chunk", il); // shape: (S_v, 1, H_v * n_seqs)
-
-        // v_new = v_i - v_prime
-        ggml_tensor * v_new = ggml_sub(ctx0, ggml_repeat(ctx0, v_chunk, v_prime), v_prime);
-        ggml_tensor * v_new_t = ggml_cont(ctx0, ggml_transpose(ctx0, v_new));
-        cb(v_new, "v_new_chunk", il);
-
-        // attn_inter = (q_i * g[:, :, i, :, None].exp()) @ last_recurrent_state
-        ggml_tensor * q_g_exp    = ggml_mul(ctx0, q_chunk, gexp_chunk);
-        ggml_tensor * attn_inter = ggml_mul_mat(ctx0, state_t, q_g_exp);
-        cb(attn_inter, "attn_inter_chunk", il);
-
-        // core_attn_out[:, :, i] = attn_inter + attn @ v_new
-        ggml_tensor * v_attn = ggml_mul_mat(ctx0, v_new_t, attn_chunk);
-        cb(v_attn, "v_attn_chunk", il);
-
-        ggml_tensor * core_attn_out_chunk = ggml_add(ctx0, attn_inter, v_attn);
-        cb(core_attn_out_chunk, "core_attn_out_chunk", il); // shape: (S_v, chunk_size, 1, H_v * n_seqs)
-
-        core_attn_out = core_attn_out == nullptr
-            ? core_attn_out_chunk
-            : ggml_concat(ctx0, core_attn_out, core_attn_out_chunk, 2);
-
-        // kgdmulvnew = (key_gdiff).transpose(-1, -2) @ v_new
-        ggml_tensor * k_gdiff_t = get_slice_2d(ctx0, key_gdiff_t, chunk);
-        //ggml_tensor * kgdmulvnew = ggml_mul_mat(ctx0, k_gdiff, v_new); // this is slower on metal, why?
-        ggml_tensor * kgdmulvnew = ggml_mul_mat(ctx0, v_new_t, k_gdiff_t);
-
-        // last_recurrent_state = last_recurrent_state * g_last + kgdmulvnew
-        ggml_tensor * gexp_last_chunk = ggml_cont(ctx0, get_slice_2d(ctx0, g_last_exp, chunk));
-        new_state = ggml_add(ctx0,
-            ggml_mul(ctx0, new_state, ggml_reshape_4d(ctx0, gexp_last_chunk, gexp_last_chunk->ne[0], gexp_last_chunk->ne[1], H_v, n_seqs)),
-            ggml_reshape_4d(ctx0, kgdmulvnew, kgdmulvnew->ne[0], kgdmulvnew->ne[1], H_v, n_seqs));
-    }
-
-    // truncate padded tokens
-    ggml_tensor * output_tokens = ggml_view_4d(ctx0, core_attn_out,
-            S_v, n_tokens, H_v, n_seqs,
-            ggml_row_size(core_attn_out->type, S_v),
-            ggml_row_size(core_attn_out->type, S_v * chunk_size * n_chunks),
-            ggml_row_size(core_attn_out->type, S_v * chunk_size * n_chunks * H_v), 0);
-    output_tokens = ggml_cont(ctx0, output_tokens);
-    cb(output_tokens, "output_tokens", il);
-
-    // permute back to (S_v, H_v, n_tokens, n_seqs)
-    output_tokens = ggml_permute(ctx0, output_tokens, 0, 2, 1, 3);
-    output_tokens = ggml_cont(ctx0, output_tokens);
-
-    return {output_tokens, new_state};
-}
-
-std::pair<ggml_tensor *, ggml_tensor *> llm_build_qwen35::build_delta_net_autoregressive(
-        ggml_tensor * q,
-        ggml_tensor * k,
-        ggml_tensor * v,
-        ggml_tensor * g,
-        ggml_tensor * beta,
-        ggml_tensor * state,
-        int           il) {
-    const int64_t S_k      = q->ne[0];
-    const int64_t H_k      = q->ne[1];
-    const int64_t n_tokens = q->ne[2];
-    const int64_t n_seqs   = q->ne[3];
-
-    const int64_t S_v = v->ne[0];
-    const int64_t H_v = v->ne[1];
-
-    GGML_ASSERT(n_tokens == 1);  // This function is optimized for single token processing
-    GGML_ASSERT(v->ne[2] == n_tokens);
-    GGML_ASSERT(k->ne[2] == n_tokens);
-    GGML_ASSERT(g->ne[0] == H_v && g->ne[1] == n_tokens && g->ne[2] == n_seqs);
-    GGML_ASSERT(beta->ne[0] == H_v && beta->ne[2] == n_tokens && beta->ne[3] == n_seqs);
-    GGML_ASSERT(state->ne[0] == S_v && state->ne[1] == S_v * H_v && state->ne[2] == 1 && state->ne[3] == n_seqs);
-
-    GGML_ASSERT(q->ne[0] == S_k && q->ne[1] == H_k && q->ne[2] == n_tokens && q->ne[3] == n_seqs);
-    GGML_ASSERT(k->ne[0] == S_k && k->ne[1] == H_k && k->ne[2] == n_tokens && k->ne[3] == n_seqs);
-
-    GGML_ASSERT(H_k == H_v);  // we did a repeat to make sure this is the case
-
-    const float eps_norm = hparams.f_norm_rms_eps;
-
-    q = ggml_l2_norm(ctx0, q, eps_norm);
-    k = ggml_l2_norm(ctx0, k, eps_norm);
-
-    const float scale = 1.0f / sqrtf(S_v);
-
-    q    = ggml_scale(ctx0, q, scale);
-    beta = ggml_sigmoid(ctx0, beta);
-
-    cb(q, "q_in", il);
-    cb(k, "k_in", il);
-    cb(v, "v_in", il);
-    cb(beta, "beta_in", il);
-    cb(g, "g_in", il);
-
-    state = ggml_reshape_4d(ctx0, state, S_v, S_v, H_v, n_seqs);
-
-    ggml_tensor * g_t    = ggml_reshape_4d(ctx0, ggml_transpose(ctx0, g), 1, 1, H_k, n_seqs);
-    ggml_tensor * beta_t = ggml_reshape_4d(ctx0, ggml_transpose(ctx0, beta), 1, 1, H_k, n_seqs);
-
-    // Apply exponential to g_t
-    g_t = ggml_exp(ctx0, g_t);
-
-    // Apply the gated delta rule for the single timestep
-    // last_recurrent_state = last_recurrent_state * g_t
-    state = ggml_mul(ctx0, state, g_t);
-
-    // kv_mem = (last_recurrent_state * k_t.unsqueeze(-1)).sum(dim=-2)
-    ggml_tensor * k_t_unsqueezed = ggml_reshape_4d(ctx0, k, 1, S_v, H_v, n_seqs);
-    ggml_tensor * kv_mem         = ggml_mul(ctx0, state, k_t_unsqueezed);
-    // we need to sum over dim=-2, so we transpose, sum, then transpose again
-    kv_mem = ggml_transpose(ctx0, ggml_sum_rows(ctx0, ggml_cont(ctx0, ggml_transpose(ctx0, kv_mem))));
-
-    // v_t = v.unsqueeze(2) (we insert the singleton dimension after n_seqs and H_v)
-    ggml_tensor * v_t    = ggml_reshape_4d(ctx0, v, S_v, 1, H_v, n_seqs);
-    // delta = (v_t - kv_mem) * beta_t
-    ggml_tensor * v_diff = ggml_sub(ctx0, v_t, kv_mem);  // both should be [S_v, 1, H_v, n_seqs]
-    ggml_tensor * delta  = ggml_mul(ctx0, v_diff, beta_t);
-
-    // last_recurrent_state = last_recurrent_state + k_t.unsqueeze(-1) * delta
-    ggml_tensor * k_t_delta = ggml_mul(ctx0, ggml_repeat_4d(ctx0, k_t_unsqueezed, S_v, S_v, H_v, n_seqs), delta);
-    state                   = ggml_add(ctx0, state, k_t_delta);
-
-    // Compute the attention output
-    // core_attn_out = (last_recurrent_state * q_t.unsqueeze(-1)).sum(dim=-2)
-    ggml_tensor * q_t_unsqueezed = ggml_reshape_4d(ctx0, q, 1, S_v, H_v, n_seqs);  // unsqueeze q_t
-    ggml_tensor * state_q        = ggml_mul(ctx0, state, q_t_unsqueezed);
-    // again, since it's over dim = -2, transpose, sum, transpose back
-    ggml_tensor * core_attn_out =
-        ggml_transpose(ctx0, ggml_sum_rows(ctx0, ggml_cont(ctx0, ggml_transpose(ctx0, state_q))));
-
-    // core_attn_out should be [S_v, 1, H_v, n_seqs] after this
-    cb(core_attn_out, "output_tokens", il);
-    cb(state, "new_state", il);
-
-    return {core_attn_out, state};
-}
-
 std::pair<ggml_tensor *, ggml_tensor *> llm_build_qwen35::build_qkvz(
                 ggml_tensor * input,
                         int   il) {
@@ -561,9 +193,6 @@ ggml_tensor * llm_build_qwen35::build_layer_attn(
 ggml_tensor * llm_build_qwen35::build_layer_attn_linear(
         llm_graph_input_rs * inp,
         ggml_tensor *        cur,
-        ggml_tensor *        causal_mask,
-        ggml_tensor *        identity,
-        ggml_tensor *        diag_mask,
         int                  il) {
     const auto * mctx_cur = inp->mctx;
 
@@ -587,8 +216,11 @@ ggml_tensor * llm_build_qwen35::build_layer_attn_linear(
     ggml_tensor * z         = qkvz.second;
 
     ggml_tensor * beta = build_lora_mm(model.layers[il].ssm_beta, cur);
-    beta  = ggml_reshape_4d(ctx0, beta, num_v_heads, 1, n_seq_tokens, n_seqs);
+    beta = ggml_reshape_4d(ctx0, beta, 1, num_v_heads, n_seq_tokens, n_seqs);
     cb(beta, "beta", il);
+
+    beta = ggml_sigmoid(ctx0, beta);
+
     ggml_tensor * alpha = build_lora_mm(model.layers[il].ssm_alpha, cur);
     alpha = ggml_cont_3d(ctx0, alpha, num_v_heads, n_seq_tokens, n_seqs);
     cb(alpha, "alpha", il);
@@ -596,15 +228,16 @@ ggml_tensor * llm_build_qwen35::build_layer_attn_linear(
     ggml_tensor * alpha_biased   = ggml_add(ctx0, alpha, model.layers[il].ssm_dt);
     ggml_tensor * alpha_softplus = ggml_softplus(ctx0, alpha_biased);
     cb(alpha_softplus, "a_softplus", il);
+
     ggml_tensor * gate = ggml_mul(ctx0, alpha_softplus, model.layers[il].ssm_a);  // -A_log.exp() * softplus
     cb(gate, "gate", il);
 
+    gate = ggml_reshape_4d(ctx0, gate, 1, num_v_heads, n_seq_tokens, n_seqs);
+
     // Get convolution states from cache
     ggml_tensor * conv_states_all = mctx_cur->get_r_l(il);
     ggml_tensor * ssm_states_all  = mctx_cur->get_s_l(il);
 
-    // bool use_precomputed_states = n_seq_tokens == 1 && mctx_cur->has_previous_state();
-
     // Build the convolution states tensor
     ggml_tensor * conv_states = build_rs(inp, conv_states_all, hparams.n_embd_r(), n_seqs);
     cb(conv_states, "conv_states", il);
@@ -613,11 +246,12 @@ ggml_tensor * llm_build_qwen35::build_layer_attn_linear(
     ggml_tensor * conv_kernel      = model.layers[il].ssm_conv1d;
     const int64_t conv_kernel_size = conv_kernel->ne[0];
     const int64_t conv_channels    = d_inner + 2 * hparams.ssm_n_group * hparams.ssm_d_state;
-    conv_states                    = ggml_reshape_3d(ctx0, conv_states, conv_kernel_size - 1, conv_channels, n_seqs);
+
+    conv_states = ggml_reshape_3d(ctx0, conv_states, conv_kernel_size - 1, conv_channels, n_seqs);
     cb(conv_states, "conv_states_reshaped", il);
 
-    qkv_mixed = ggml_permute(ctx0, qkv_mixed, 1, 0, 2, 3);
-    cb(qkv_mixed, "qkv_mixed_permuted", il);
+    qkv_mixed = ggml_transpose(ctx0, qkv_mixed);
+    cb(qkv_mixed, "qkv_mixed_transposed", il);
 
     ggml_tensor * conv_input = ggml_concat(ctx0, conv_states, qkv_mixed, 0);
     cb(conv_input, "conv_input", il);
@@ -637,7 +271,10 @@ ggml_tensor * llm_build_qwen35::build_layer_attn_linear(
     ggml_build_forward_expand(gf, ggml_cpy(ctx0, last_conv_states, state_update_target));
     cb(conv_states_all, "conv_states_updated", il);
 
-    // Apply SSM convolution
+    ggml_tensor * state = build_rs(inp, ssm_states_all, hparams.n_embd_s(), n_seqs);
+    state = ggml_reshape_4d(ctx0, state, head_v_dim, head_v_dim, num_v_heads, n_seqs);
+    cb(state, "state_predelta", il);
+
     ggml_tensor * conv_output_proper = ggml_ssm_conv(ctx0, conv_input, conv_kernel);
     cb(conv_output_proper, "conv_output_raw", il);
 
@@ -651,31 +288,41 @@ ggml_tensor * llm_build_qwen35::build_layer_attn_linear(
     int64_t nb1_qkv = ggml_row_size(conv_qkv_mix->type, qkv_dim);
 
     // Extract the convolved Q, K, V from conv_output
-    ggml_tensor * q_conv =
-        ggml_view_2d(ctx0, conv_qkv_mix, head_k_dim * num_k_heads, n_seq_tokens * n_seqs, nb1_qkv, 0);
+    ggml_tensor * q_conv = ggml_view_4d(ctx0, conv_qkv_mix, head_k_dim, num_k_heads, n_seq_tokens, n_seqs,
+            ggml_row_size(conv_qkv_mix->type, head_k_dim),
+            nb1_qkv,
+            nb1_qkv * n_seq_tokens,
+            0);
+
+    ggml_tensor * k_conv = ggml_view_4d(ctx0, conv_qkv_mix, head_k_dim, num_k_heads, n_seq_tokens, n_seqs,
+            ggml_row_size(conv_qkv_mix->type, head_k_dim),
+            nb1_qkv,
+            nb1_qkv * n_seq_tokens,
+            head_k_dim * num_k_heads * ggml_element_size(conv_qkv_mix));
+
+    ggml_tensor * v_conv = ggml_view_4d(ctx0, conv_qkv_mix, head_v_dim, num_v_heads, n_seq_tokens, n_seqs,
+            ggml_row_size(conv_qkv_mix->type, head_v_dim),
+            nb1_qkv,
+            nb1_qkv * n_seq_tokens,
+            ggml_row_size(conv_qkv_mix->type, 2 * head_k_dim * num_k_heads));
+
     cb(q_conv, "q_conv", il);
-    ggml_tensor * k_conv =
-        ggml_view_2d(ctx0, conv_qkv_mix, head_k_dim * num_k_heads, n_seq_tokens * n_seqs, nb1_qkv,
-                     head_k_dim * num_k_heads * ggml_element_size(conv_qkv_mix));
     cb(k_conv, "k_conv", il);
-    ggml_tensor * v_conv =
-        ggml_view_2d(ctx0, conv_qkv_mix, head_v_dim * num_v_heads, n_seq_tokens * n_seqs, nb1_qkv,
-                     2 * head_k_dim * num_k_heads * ggml_element_size(conv_qkv_mix));
     cb(v_conv, "v_conv", il);
 
-    // Unsqueeze them
-    q_conv = ggml_cont_4d(ctx0, q_conv, head_k_dim, num_k_heads, n_seq_tokens, n_seqs);
-    k_conv = ggml_cont_4d(ctx0, k_conv, head_k_dim, num_k_heads, n_seq_tokens, n_seqs);
-    v_conv = ggml_cont_4d(ctx0, v_conv, head_v_dim, num_v_heads, n_seq_tokens, n_seqs);
+    const float eps_norm = hparams.f_norm_rms_eps;
 
-    ggml_tensor * state = build_rs(inp, ssm_states_all, hparams.n_embd_s(), n_seqs);
-    state               = ggml_reshape_4d(ctx0, state, head_v_dim, head_v_dim * num_v_heads, 1, n_seqs);
-    cb(state, "state_predelta", il);
+    q_conv = ggml_l2_norm(ctx0, q_conv, eps_norm);
+    k_conv = ggml_l2_norm(ctx0, k_conv, eps_norm);
 
-    // if head keys and value keys are different, repeat Q/K to match V's head count
-    // V heads are in tiled order (from conversion), so simple tiled repeat works
+    //q_conv = ggml_cont_4d(ctx0, q_conv, head_k_dim, num_k_heads, n_seq_tokens, n_seqs);
+    //k_conv = ggml_cont_4d(ctx0, k_conv, head_k_dim, num_k_heads, n_seq_tokens, n_seqs);
+    //v_conv = ggml_cont_4d(ctx0, v_conv, head_v_dim, num_v_heads, n_seq_tokens, n_seqs);
+
+    // if head keys and value keys are different, repeat to force tensors into matching shapes
     if (num_k_heads != num_v_heads) {
         GGML_ASSERT(num_v_heads % num_k_heads == 0);
+        // TODO: try to avoid these explicit repeats by utilizing op broadcast
         q_conv = ggml_repeat_4d(ctx0, q_conv, head_k_dim, num_v_heads, n_seq_tokens, n_seqs);
         k_conv = ggml_repeat_4d(ctx0, k_conv, head_k_dim, num_v_heads, n_seq_tokens, n_seqs);
     }
@@ -689,7 +336,7 @@ ggml_tensor * llm_build_qwen35::build_layer_attn_linear(
     if (n_seq_tokens == 1) {
         attn_out = build_delta_net_autoregressive(q_conv, k_conv, v_conv, gate, beta, state, il);
     } else {
-        attn_out = build_delta_net_chunking(q_conv, k_conv, v_conv, gate, beta, state, causal_mask, identity, diag_mask, il);
+        attn_out = build_delta_net_chunking(q_conv, k_conv, v_conv, gate, beta, state, il);
     }
     ggml_tensor * output    = attn_out.first;
     ggml_tensor * new_state = attn_out.second;
@@ -698,19 +345,15 @@ ggml_tensor * llm_build_qwen35::build_layer_attn_linear(
 
     // Update the recurrent states
     ggml_build_forward_expand(gf,
-                              ggml_cpy(ctx0, new_state,
-                                       ggml_view_1d(ctx0, ssm_states_all, hparams.n_embd_s() * n_seqs,
-                                                    kv_head * hparams.n_embd_s() * ggml_element_size(ssm_states_all))));
-
-    // Reshape both attn_out_final and z to 2D tensors for normalization
-    // attn_out_final: [head_dim, n_heads, n_tokens, n_seqs] -> [n_heads * n_tokens * n_seqs, head_dim]
-    ggml_tensor * attn_out_2d_final = ggml_reshape_2d(ctx0, output, head_v_dim, num_v_heads * n_seq_tokens * n_seqs);
+            ggml_cpy(ctx0, new_state,
+                ggml_view_1d(ctx0, ssm_states_all, hparams.n_embd_s() * n_seqs,
+                    kv_head * hparams.n_embd_s() * ggml_element_size(ssm_states_all))));
 
     // z: [head_dim, n_heads, n_tokens, n_seqs] -> [n_heads * n_tokens * n_seqs, head_dim]
-    ggml_tensor * z_2d = ggml_reshape_2d(ctx0, z, head_v_dim, num_v_heads * n_seq_tokens * n_seqs);
+    ggml_tensor * z_2d = ggml_reshape_4d(ctx0, z, head_v_dim, num_v_heads, n_seq_tokens, n_seqs);
 
     // Apply gated normalization: self.norm(core_attn_out, z)
-    ggml_tensor * attn_out_norm = build_norm_gated(attn_out_2d_final, model.layers[il].ssm_norm, z_2d, il);
+    ggml_tensor * attn_out_norm = build_norm_gated(output, model.layers[il].ssm_norm, z_2d, il);
 
     // Final reshape: [head_dim, n_heads, n_tokens, n_seqs] -> [n_tokens, n_seqs, n_heads * head_dim]
     ggml_tensor * final_output = ggml_reshape_3d(ctx0, attn_out_norm, head_v_dim * num_v_heads, n_seq_tokens, n_seqs);
@@ -721,7 +364,8 @@ ggml_tensor * llm_build_qwen35::build_layer_attn_linear(
     cb(cur, "linear_attn_out", il);
 
     // Reshape back to original dimensions
-    cur = ggml_cont_2d(ctx0, cur, n_embd, n_seq_tokens * n_seqs);
+    cur = ggml_reshape_2d(ctx0, cur, n_embd, n_seq_tokens * n_seqs);
+
     return cur;
 }
 

src/models/qwen35moe.cpp

@@ -2,10 +2,8 @@
 
 #include "llama-memory-recurrent.h"
 
-#define CHUNK_SIZE 64
-
 llm_build_qwen35moe::llm_build_qwen35moe(const llama_model & model, const llm_graph_params & params) :
-    llm_graph_context(params), model(model) {
+    llm_build_delta_net_base(params), model(model) {
     const int64_t n_embd_head = hparams.n_embd_head_v;
 
     GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
@@ -25,17 +23,6 @@ llm_build_qwen35moe::llm_build_qwen35moe(const llama_model & model, const llm_gr
     ggml_tensor * inp_pos     = build_inp_pos();
     ggml_tensor * inp_out_ids = build_inp_out_ids();
 
-    ggml_tensor * causal_mask =
-        ggml_tri(ctx0, ggml_fill(ctx0, ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, CHUNK_SIZE, CHUNK_SIZE), 1.0f),
-                    GGML_TRI_TYPE_LOWER);
-
-    ggml_tensor * identity = ggml_diag(ctx0, ggml_fill(ctx0, ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, CHUNK_SIZE), 1.0f));
-    ggml_tensor * diag_mask = ggml_add(ctx0, causal_mask, identity);
-
-    ggml_build_forward_expand(gf, causal_mask);
-    ggml_build_forward_expand(gf, identity);
-    ggml_build_forward_expand(gf, diag_mask);
-
     for (int il = 0; il < n_layer; ++il) {
         ggml_tensor * inpSA = inpL;
 
@@ -45,7 +32,7 @@ llm_build_qwen35moe::llm_build_qwen35moe(const llama_model & model, const llm_gr
         // Determine layer type and build appropriate attention mechanism
         if (hparams.is_recurrent(il)) {
             // Linear attention layer (gated delta net)
-            cur = build_layer_attn_linear(inp->get_recr(), cur, causal_mask, identity, diag_mask, il);
+            cur = build_layer_attn_linear(inp->get_recr(), cur, il);
         } else {
             // Full attention layer
             cur = build_layer_attn(inp->get_attn(), cur, inp_pos, sections, il);
@@ -95,362 +82,6 @@ llm_build_qwen35moe::llm_build_qwen35moe(const llama_model & model, const llm_gr
     ggml_build_forward_expand(gf, cur);
 }
 
-// utility to get one slice from the third dimension
-// input dim:  [x, y, c, b]
-// output dim: [x, y, 1, b]
-static ggml_tensor * get_slice_2d(ggml_context * ctx0, ggml_tensor * t, int64_t c) {
-    return ggml_view_4d(ctx0, t, t->ne[0], t->ne[1], 1, t->ne[3],
-        t->nb[1], t->nb[2], t->nb[3], t->nb[2] * c);
-}
-
-std::pair<ggml_tensor *, ggml_tensor *> llm_build_qwen35moe::build_delta_net_chunking(
-        ggml_tensor * q,
-        ggml_tensor * k,
-        ggml_tensor * v,
-        ggml_tensor * g,
-        ggml_tensor * beta,
-        ggml_tensor * state,
-        ggml_tensor * causal_mask,
-        ggml_tensor * identity,
-        ggml_tensor * diag_mask,
-        int           il) {
-    const int64_t S_k      = q->ne[0];
-    const int64_t H_k      = q->ne[1];
-    const int64_t n_tokens = q->ne[2];
-    const int64_t n_seqs   = q->ne[3];
-
-    const int64_t S_v = v->ne[0];
-    const int64_t H_v = v->ne[1];
-
-    GGML_ASSERT(v->ne[2] == n_tokens);
-    GGML_ASSERT(k->ne[2] == n_tokens);
-    GGML_ASSERT(g->ne[0] == H_v && g->ne[1] == n_tokens && g->ne[2] == n_seqs);
-    GGML_ASSERT(beta->ne[0] == H_v && beta->ne[2] == n_tokens && beta->ne[3] == n_seqs);
-    GGML_ASSERT(state->ne[0] == S_v && state->ne[1] == S_v * H_v && state->ne[2] == 1 && state->ne[3] == n_seqs);
-
-    GGML_ASSERT(q->ne[0] == S_k && q->ne[1] == H_k && q->ne[2] == n_tokens && q->ne[3] == n_seqs);
-    GGML_ASSERT(k->ne[0] == S_k && k->ne[1] == H_k && k->ne[2] == n_tokens && k->ne[3] == n_seqs);
-
-    GGML_ASSERT(H_k == H_v);  // we did a repeat to make sure this is the case
-
-    const float eps_norm = hparams.f_norm_rms_eps;
-
-    q = ggml_l2_norm(ctx0, q, eps_norm);
-    k = ggml_l2_norm(ctx0, k, eps_norm);
-
-    const float scale = 1.0f / sqrtf(S_v);
-
-    q = ggml_scale(ctx0, q, scale);
-
-    beta = ggml_sigmoid(ctx0, beta);
-
-    cb(q, "q_in", il);
-    cb(k, "k_in", il);
-    cb(v, "v_in", il);
-    cb(beta, "beta_in", il);
-    cb(g, "g_in", il);
-
-    q = ggml_cont_4d(ctx0, ggml_permute(ctx0, q, 0, 2, 1, 3), S_v, n_tokens, H_v, n_seqs);
-    k = ggml_cont_4d(ctx0, ggml_permute(ctx0, k, 0, 2, 1, 3), S_v, n_tokens, H_v, n_seqs);
-    v = ggml_cont_4d(ctx0, ggml_permute(ctx0, v, 0, 2, 1, 3), S_v, n_tokens, H_v, n_seqs);
-    g = ggml_cont_4d(ctx0, ggml_permute(ctx0, g, 2, 0, 3, 1), n_tokens, 1, H_k, n_seqs);
-
-    beta  = ggml_cont(ctx0, ggml_permute(ctx0, beta, 2, 0, 1, 3));
-    state = ggml_reshape_4d(ctx0, state, S_v, S_v, H_v, n_seqs);
-
-    cb(q, "q_perm", il);
-    cb(k, "k_perm", il);
-    cb(v, "v_perm", il);
-    cb(beta, "beta_perm", il);
-    cb(g, "g_perm", il);
-    cb(state, "state_in", il);
-
-    GGML_ASSERT(q->ne[1] == n_tokens && q->ne[0] == S_k && q->ne[2] == H_k && q->ne[3] == n_seqs);
-    GGML_ASSERT(k->ne[1] == n_tokens && k->ne[0] == S_k && k->ne[2] == H_k && k->ne[3] == n_seqs);
-    GGML_ASSERT(v->ne[1] == n_tokens && v->ne[0] == S_v && v->ne[2] == H_k && v->ne[3] == n_seqs);
-    GGML_ASSERT(beta->ne[1] == n_tokens && beta->ne[2] == H_k && beta->ne[0] == 1 && beta->ne[3] == n_seqs);
-
-    // Do padding
-    const int64_t chunk_size = CHUNK_SIZE;
-
-    const int64_t pad = (chunk_size - n_tokens % chunk_size) % chunk_size;
-    const int64_t n_chunks = (n_tokens + pad) / chunk_size;
-
-    q = ggml_pad(ctx0, q, 0, pad, 0, 0);
-    k = ggml_pad(ctx0, k, 0, pad, 0, 0);
-    v = ggml_pad(ctx0, v, 0, pad, 0, 0);
-    g = ggml_pad(ctx0, g, pad, 0, 0, 0);
-    beta = ggml_pad(ctx0, beta, 0, pad, 0, 0);
-
-    cb(q, "q_pad", il);
-    cb(k, "k_pad", il);
-    cb(v, "v_pad", il);
-    cb(beta, "beta_pad", il);
-    cb(g, "g_pad", il);
-
-    ggml_tensor * v_beta = ggml_mul(ctx0, v, beta);
-    ggml_tensor * k_beta = ggml_mul(ctx0, k, beta);
-
-    cb(v_beta, "v_beta", il);
-    cb(k_beta, "k_beta", il);
-
-    q      = ggml_reshape_4d(ctx0, q,      S_k, chunk_size, n_chunks, H_k * n_seqs);
-    k      = ggml_reshape_4d(ctx0, k,      S_k, chunk_size, n_chunks, H_k * n_seqs);
-    k_beta = ggml_reshape_4d(ctx0, k_beta, S_k, chunk_size, n_chunks, H_k * n_seqs);
-    v      = ggml_reshape_4d(ctx0, v,      S_v, chunk_size, n_chunks, H_v * n_seqs);
-    v_beta = ggml_reshape_4d(ctx0, v_beta, S_v, chunk_size, n_chunks, H_v * n_seqs);
-
-    g    = ggml_reshape_4d(ctx0, g, chunk_size, 1, n_chunks, H_k * n_seqs);
-    beta = ggml_reshape_4d(ctx0, beta, 1, chunk_size, n_chunks, H_k * n_seqs);
-
-    ggml_tensor * g_cumsum = ggml_cumsum(ctx0, g);
-    cb(g_cumsum, "g_cumsum", il); // shape: (chunk_size, 1, n_chunks, H_v * n_seqs)
-
-    ggml_tensor * gcs_i = g_cumsum; // ggml_reshape_4d(ctx0, g_cumsum, chunk_size, 1, n_chunks, H_v * n_seqs);
-    ggml_tensor * gcs_j = ggml_reshape_4d(ctx0, g_cumsum, 1, chunk_size, n_chunks, H_v * n_seqs);
-
-    ggml_tensor * gcs_j_broadcast =
-        ggml_repeat_4d(ctx0, gcs_j, chunk_size, chunk_size, n_chunks, H_v * n_seqs);
-
-    ggml_tensor * decay_mask = ggml_sub(ctx0, gcs_j_broadcast, gcs_i);
-    cb(decay_mask, "decay_mask", il); // shape: (chunk_size, chunk_size, n_chunks, H_v * n_seqs)
-
-    decay_mask = ggml_mul(ctx0, decay_mask, diag_mask);
-    decay_mask = ggml_exp(ctx0, decay_mask);
-    decay_mask = ggml_mul(ctx0, decay_mask, diag_mask);
-
-    ggml_tensor * kmulkbeta = ggml_mul_mat(ctx0, k, k_beta);
-
-    ggml_tensor * k_decay = ggml_mul(ctx0, kmulkbeta, decay_mask);
-    ggml_tensor * attn    = ggml_neg(ctx0, ggml_mul(ctx0, k_decay, causal_mask));
-    cb(attn, "attn_pre_solve", il); // shape: (chunk_size, chunk_size, n_chunks, H_v * n_seqs)
-
-    ggml_tensor * attn_lower = ggml_mul(ctx0, attn, causal_mask);
-    ggml_tensor * lhs        = ggml_sub(ctx0, ggml_repeat(ctx0, identity, attn_lower), attn_lower);
-
-    ggml_tensor * lin_solve  = ggml_solve_tri(ctx0, lhs, attn, true, true, false);
-    attn                     = ggml_mul(ctx0, lin_solve, causal_mask);
-    attn                     = ggml_add(ctx0, attn, identity);
-    cb(attn, "attn_solved", il); // shape: (chunk_size, chunk_size, n_chunks, H_v * n_seqs)
-
-    v = ggml_mul_mat(ctx0, ggml_cont(ctx0, ggml_transpose(ctx0, v_beta)), attn);
-
-    ggml_tensor * g_cumsum_t = ggml_cont(ctx0, ggml_transpose(ctx0, g_cumsum));
-    ggml_tensor * gexp       = ggml_exp(ctx0, g_cumsum_t);
-
-    ggml_tensor * kbeta_gexp = ggml_mul(ctx0, k_beta, gexp);
-    cb(kbeta_gexp, "kbeta_gexp", il); // shape: (S_k, chunk_size, n_chunks, H_v * n_seqs)
-
-    ggml_tensor * k_cumdecay =
-        ggml_cont(ctx0, ggml_transpose(ctx0, ggml_mul_mat(ctx0, attn, ggml_cont(ctx0, ggml_transpose(ctx0, kbeta_gexp)))));
-    cb(k_cumdecay, "k_cumdecay", il); // shape: (chunk_size, chunk_size, n_chunks, H_v * n_seqs)
-
-    ggml_tensor * attn_kq = ggml_mul_mat(ctx0, k, q);
-    attn_kq = ggml_mul(ctx0, attn_kq, decay_mask);
-    attn_kq = ggml_mul(ctx0, attn_kq, diag_mask);
-    cb(attn_kq, "attn_kq", il); // shape: (chunk_size, chunk_size, n_chunks, H_v * n_seqs)
-
-
-    // vectorized calculation of key_gdiff
-    // improved from the chunked version:
-    //   g_last = torch.clamp(g_cum[:, :, -1], max=50.0).exp().unsqueeze(-1).unsqueeze(-1)
-    //   g_diff = torch.clamp(g_cum[:, :, -1:] - g_cum, max=50.0).exp()
-    //   key_gdiff = key * g_diff.unsqueeze(-1)
-    //   kgdmulvnew = (key_gdiff).transpose(-1, -2) @ v_new
-    //   last_recurrent_state = last_recurrent_state * g_last + kgdmulvnew
-
-    // get last element in g_cumsum along chunk_size dimension (ne0)
-    // example: [[x, y, z, ..., last], ...] -> [[last], ...]
-    ggml_tensor * g_last = ggml_view_4d(ctx0, g_cumsum, 1, 1, g_cumsum->ne[2], g_cumsum->ne[3],
-                                        g_cumsum->nb[1], g_cumsum->nb[2], g_cumsum->nb[3],
-                                        (g_cumsum->ne[0] - 1) * ggml_element_size(g_cumsum));
-    g_last = ggml_cont(ctx0, g_last);
-    cb(g_last, "g_last", il); // shape: (1, 1, n_chunks, H_v * n_seqs)
-
-    ggml_tensor * g_last_exp = ggml_exp(ctx0, g_last);
-    cb(g_last_exp, "g_last_exp", il); // shape: (1, 1, n_chunks, H_v * n_seqs)
-
-    ggml_tensor * g_diff = ggml_neg(ctx0, ggml_sub(ctx0, g_cumsum, g_last));
-    cb(g_diff, "g_diff", il); // shape: (chunk_size, 1, n_chunks, H_v * n_seqs)
-
-    ggml_tensor * g_diff_exp = ggml_exp(ctx0, g_diff);
-    ggml_tensor * g_diff_exp_t = ggml_reshape_4d(ctx0, g_diff_exp,
-                                                 1, chunk_size, n_chunks, g_diff_exp->ne[3]);
-
-    ggml_tensor * key_gdiff = ggml_mul(ctx0, k, g_diff_exp_t);
-    cb(key_gdiff, "key_gdiff", il); // shape: (S_k, chunk_size, n_chunks, H_v * n_seqs)
-
-    ggml_tensor * key_gdiff_t = ggml_cont(ctx0, ggml_transpose(ctx0, key_gdiff));
-    cb(key_gdiff_t, "key_gdiff_t", il); // shape: (chunk_size, S_k, n_chunks, H_v * n_seqs)
-
-
-    // state to be updated per chunk
-    ggml_tensor * new_state = state; // ggml_dup(ctx0, state);
-    cb(new_state, "new_state", il); // shape: (S_v, S_v, H_v, n_seqs)
-
-    // shape after loop of chunks: (S_v, chunk_size, n_chunks, H_v * n_seqs)
-    ggml_tensor * core_attn_out = nullptr;
-
-    for (int64_t chunk = 0; chunk < n_chunks; chunk++) {
-        // shape: (S_k, chunk_size, 1, H_k * n_seqs)
-        ggml_tensor * q_chunk = get_slice_2d(ctx0, q, chunk); // (no cont), next op: ggml_mul
-
-        // shape: (S_v, chunk_size, 1, H_v * n_seqs)
-        ggml_tensor * v_chunk = get_slice_2d(ctx0, v, chunk); // (no cont), next op: ggml_repeat
-
-        // shape: (chunk_size, 1, n_chunks, H_v * n_seqs)
-        ggml_tensor * gexp_chunk = get_slice_2d(ctx0, gexp, chunk); // (no cont), next op: ggml_mul
-
-        // shape: (chunk_size, 1, H_v * n_seqs)
-        ggml_tensor * k_cumdecay_chunk = get_slice_2d(ctx0, k_cumdecay, chunk); // (no cont), next op: ggml_mul_mat
-
-        // attn = (q_i @ k_i.transpose(-1, -2) * decay_mask[:, :, i]).masked_fill_(mask, 0)
-        // replaced by precomputed attn_kq
-        ggml_tensor * attn_chunk = get_slice_2d(ctx0, attn_kq, chunk);
-        cb(attn_chunk, "attn_chunk", il);
-
-        ggml_tensor * state_t = ggml_cont_4d(ctx0, ggml_permute(ctx0, new_state, 1, 0, 2, 3), S_v, S_v, 1, H_v * n_seqs);
-
-        // v_prime = (k_cumdecay[:, :, i]) @ last_recurrent_state
-        ggml_tensor * v_prime = ggml_mul_mat(ctx0, state_t, k_cumdecay_chunk);
-        cb(v_prime, "v_prime_chunk", il); // shape: (S_v, 1, H_v * n_seqs)
-
-        // v_new = v_i - v_prime
-        ggml_tensor * v_new = ggml_sub(ctx0, ggml_repeat(ctx0, v_chunk, v_prime), v_prime);
-        ggml_tensor * v_new_t = ggml_cont(ctx0, ggml_transpose(ctx0, v_new));
-        cb(v_new, "v_new_chunk", il);
-
-        // attn_inter = (q_i * g[:, :, i, :, None].exp()) @ last_recurrent_state
-        ggml_tensor * q_g_exp    = ggml_mul(ctx0, q_chunk, gexp_chunk);
-        ggml_tensor * attn_inter = ggml_mul_mat(ctx0, state_t, q_g_exp);
-        cb(attn_inter, "attn_inter_chunk", il);
-
-        // core_attn_out[:, :, i] = attn_inter + attn @ v_new
-        ggml_tensor * v_attn = ggml_mul_mat(ctx0, v_new_t, attn_chunk);
-        cb(v_attn, "v_attn_chunk", il);
-
-        ggml_tensor * core_attn_out_chunk = ggml_add(ctx0, attn_inter, v_attn);
-        cb(core_attn_out_chunk, "core_attn_out_chunk", il); // shape: (S_v, chunk_size, 1, H_v * n_seqs)
-
-        core_attn_out = core_attn_out == nullptr
-            ? core_attn_out_chunk
-            : ggml_concat(ctx0, core_attn_out, core_attn_out_chunk, 2);
-
-        // kgdmulvnew = (key_gdiff).transpose(-1, -2) @ v_new
-        ggml_tensor * k_gdiff_t = get_slice_2d(ctx0, key_gdiff_t, chunk);
-        //ggml_tensor * kgdmulvnew = ggml_mul_mat(ctx0, k_gdiff, v_new); // this is slower on metal, why?
-        ggml_tensor * kgdmulvnew = ggml_mul_mat(ctx0, v_new_t, k_gdiff_t);
-
-        // last_recurrent_state = last_recurrent_state * g_last + kgdmulvnew
-        ggml_tensor * gexp_last_chunk = ggml_cont(ctx0, get_slice_2d(ctx0, g_last_exp, chunk));
-        new_state = ggml_add(ctx0,
-            ggml_mul(ctx0, new_state, ggml_reshape_4d(ctx0, gexp_last_chunk, gexp_last_chunk->ne[0], gexp_last_chunk->ne[1], H_v, n_seqs)),
-            ggml_reshape_4d(ctx0, kgdmulvnew, kgdmulvnew->ne[0], kgdmulvnew->ne[1], H_v, n_seqs));
-    }
-
-    // truncate padded tokens
-    ggml_tensor * output_tokens = ggml_view_4d(ctx0, core_attn_out,
-            S_v, n_tokens, H_v, n_seqs,
-            ggml_row_size(core_attn_out->type, S_v),
-            ggml_row_size(core_attn_out->type, S_v * chunk_size * n_chunks),
-            ggml_row_size(core_attn_out->type, S_v * chunk_size * n_chunks * H_v), 0);
-    output_tokens = ggml_cont(ctx0, output_tokens);
-    cb(output_tokens, "output_tokens", il);
-
-    // permute back to (S_v, H_v, n_tokens, n_seqs)
-    output_tokens = ggml_permute(ctx0, output_tokens, 0, 2, 1, 3);
-    output_tokens = ggml_cont(ctx0, output_tokens);
-
-    return {output_tokens, new_state};
-}
-
-std::pair<ggml_tensor *, ggml_tensor *> llm_build_qwen35moe::build_delta_net_autoregressive(
-        ggml_tensor * q,
-        ggml_tensor * k,
-        ggml_tensor * v,
-        ggml_tensor * g,
-        ggml_tensor * beta,
-        ggml_tensor * state,
-        int           il) {
-    const int64_t S_k      = q->ne[0];
-    const int64_t H_k      = q->ne[1];
-    const int64_t n_tokens = q->ne[2];
-    const int64_t n_seqs   = q->ne[3];
-
-    const int64_t S_v = v->ne[0];
-    const int64_t H_v = v->ne[1];
-
-    GGML_ASSERT(n_tokens == 1);  // This function is optimized for single token processing
-    GGML_ASSERT(v->ne[2] == n_tokens);
-    GGML_ASSERT(k->ne[2] == n_tokens);
-    GGML_ASSERT(g->ne[0] == H_v && g->ne[1] == n_tokens && g->ne[2] == n_seqs);
-    GGML_ASSERT(beta->ne[0] == H_v && beta->ne[2] == n_tokens && beta->ne[3] == n_seqs);
-    GGML_ASSERT(state->ne[0] == S_v && state->ne[1] == S_v * H_v && state->ne[2] == 1 && state->ne[3] == n_seqs);
-
-    GGML_ASSERT(q->ne[0] == S_k && q->ne[1] == H_k && q->ne[2] == n_tokens && q->ne[3] == n_seqs);
-    GGML_ASSERT(k->ne[0] == S_k && k->ne[1] == H_k && k->ne[2] == n_tokens && k->ne[3] == n_seqs);
-
-    GGML_ASSERT(H_k == H_v);  // we did a repeat to make sure this is the case
-
-    const float eps_norm = hparams.f_norm_rms_eps;
-
-    q = ggml_l2_norm(ctx0, q, eps_norm);
-    k = ggml_l2_norm(ctx0, k, eps_norm);
-
-    const float scale = 1.0f / sqrtf(S_v);
-
-    q    = ggml_scale(ctx0, q, scale);
-    beta = ggml_sigmoid(ctx0, beta);
-
-    cb(q, "q_in", il);
-    cb(k, "k_in", il);
-    cb(v, "v_in", il);
-    cb(beta, "beta_in", il);
-    cb(g, "g_in", il);
-
-    state = ggml_reshape_4d(ctx0, state, S_v, S_v, H_v, n_seqs);
-
-    ggml_tensor * g_t    = ggml_reshape_4d(ctx0, ggml_transpose(ctx0, g), 1, 1, H_k, n_seqs);
-    ggml_tensor * beta_t = ggml_reshape_4d(ctx0, ggml_transpose(ctx0, beta), 1, 1, H_k, n_seqs);
-
-    // Apply exponential to g_t
-    g_t = ggml_exp(ctx0, g_t);
-
-    // Apply the gated delta rule for the single timestep
-    // last_recurrent_state = last_recurrent_state * g_t
-    state = ggml_mul(ctx0, state, g_t);
-
-    // kv_mem = (last_recurrent_state * k_t.unsqueeze(-1)).sum(dim=-2)
-    ggml_tensor * k_t_unsqueezed = ggml_reshape_4d(ctx0, k, 1, S_v, H_v, n_seqs);
-    ggml_tensor * kv_mem         = ggml_mul(ctx0, state, k_t_unsqueezed);
-    // we need to sum over dim=-2, so we transpose, sum, then transpose again
-    kv_mem = ggml_transpose(ctx0, ggml_sum_rows(ctx0, ggml_cont(ctx0, ggml_transpose(ctx0, kv_mem))));
-
-    // v_t = v.unsqueeze(2) (we insert the singleton dimension after n_seqs and H_v)
-    ggml_tensor * v_t    = ggml_reshape_4d(ctx0, v, S_v, 1, H_v, n_seqs);
-    // delta = (v_t - kv_mem) * beta_t
-    ggml_tensor * v_diff = ggml_sub(ctx0, v_t, kv_mem);  // both should be [S_v, 1, H_v, n_seqs]
-    ggml_tensor * delta  = ggml_mul(ctx0, v_diff, beta_t);
-
-    // last_recurrent_state = last_recurrent_state + k_t.unsqueeze(-1) * delta
-    ggml_tensor * k_t_delta = ggml_mul(ctx0, ggml_repeat_4d(ctx0, k_t_unsqueezed, S_v, S_v, H_v, n_seqs), delta);
-    state                   = ggml_add(ctx0, state, k_t_delta);
-
-    // Compute the attention output
-    // core_attn_out = (last_recurrent_state * q_t.unsqueeze(-1)).sum(dim=-2)
-    ggml_tensor * q_t_unsqueezed = ggml_reshape_4d(ctx0, q, 1, S_v, H_v, n_seqs);  // unsqueeze q_t
-    ggml_tensor * state_q        = ggml_mul(ctx0, state, q_t_unsqueezed);
-    // again, since it's over dim = -2, transpose, sum, transpose back
-    ggml_tensor * core_attn_out =
-        ggml_transpose(ctx0, ggml_sum_rows(ctx0, ggml_cont(ctx0, ggml_transpose(ctx0, state_q))));
-
-    // core_attn_out should be [S_v, 1, H_v, n_seqs] after this
-    cb(core_attn_out, "output_tokens", il);
-    cb(state, "new_state", il);
-
-    return {core_attn_out, state};
-}
-
 std::pair<ggml_tensor *, ggml_tensor *> llm_build_qwen35moe::build_qkvz(
                 ggml_tensor * input,
                         int   il) {
@@ -562,9 +193,6 @@ ggml_tensor * llm_build_qwen35moe ::build_layer_attn(
 ggml_tensor * llm_build_qwen35moe ::build_layer_attn_linear(
         llm_graph_input_rs * inp,
         ggml_tensor *        cur,
-        ggml_tensor *        causal_mask,
-        ggml_tensor *        identity,
-        ggml_tensor *        diag_mask,
         int                  il) {
     const auto * mctx_cur = inp->mctx;
 
@@ -588,8 +216,11 @@ ggml_tensor * llm_build_qwen35moe ::build_layer_attn_linear(
     ggml_tensor * z         = qkvz.second;
 
     ggml_tensor * beta = build_lora_mm(model.layers[il].ssm_beta, cur);
-    beta  = ggml_reshape_4d(ctx0, beta, num_v_heads, 1, n_seq_tokens, n_seqs);
+    beta = ggml_reshape_4d(ctx0, beta, 1, num_v_heads, n_seq_tokens, n_seqs);
     cb(beta, "beta", il);
+
+    beta = ggml_sigmoid(ctx0, beta);
+
     ggml_tensor * alpha = build_lora_mm(model.layers[il].ssm_alpha, cur);
     alpha = ggml_cont_3d(ctx0, alpha, num_v_heads, n_seq_tokens, n_seqs);
     cb(alpha, "alpha", il);
@@ -597,15 +228,16 @@ ggml_tensor * llm_build_qwen35moe ::build_layer_attn_linear(
     ggml_tensor * alpha_biased   = ggml_add(ctx0, alpha, model.layers[il].ssm_dt);
     ggml_tensor * alpha_softplus = ggml_softplus(ctx0, alpha_biased);
     cb(alpha_softplus, "a_softplus", il);
+
     ggml_tensor * gate = ggml_mul(ctx0, alpha_softplus, model.layers[il].ssm_a);  // -A_log.exp() * softplus
     cb(gate, "gate", il);
 
+    gate = ggml_reshape_4d(ctx0, gate, 1, num_v_heads, n_seq_tokens, n_seqs);
+
     // Get convolution states from cache
     ggml_tensor * conv_states_all = mctx_cur->get_r_l(il);
     ggml_tensor * ssm_states_all  = mctx_cur->get_s_l(il);
 
-    // bool use_precomputed_states = n_seq_tokens == 1 && mctx_cur->has_previous_state();
-
     // Build the convolution states tensor
     ggml_tensor * conv_states = build_rs(inp, conv_states_all, hparams.n_embd_r(), n_seqs);
     cb(conv_states, "conv_states", il);
@@ -614,11 +246,12 @@ ggml_tensor * llm_build_qwen35moe ::build_layer_attn_linear(
     ggml_tensor * conv_kernel      = model.layers[il].ssm_conv1d;
     const int64_t conv_kernel_size = conv_kernel->ne[0];
     const int64_t conv_channels    = d_inner + 2 * hparams.ssm_n_group * hparams.ssm_d_state;
-    conv_states                    = ggml_reshape_3d(ctx0, conv_states, conv_kernel_size - 1, conv_channels, n_seqs);
+
+    conv_states = ggml_reshape_3d(ctx0, conv_states, conv_kernel_size - 1, conv_channels, n_seqs);
     cb(conv_states, "conv_states_reshaped", il);
 
-    qkv_mixed = ggml_permute(ctx0, qkv_mixed, 1, 0, 2, 3);
-    cb(qkv_mixed, "qkv_mixed_permuted", il);
+    qkv_mixed = ggml_transpose(ctx0, qkv_mixed);
+    cb(qkv_mixed, "qkv_mixed_transposed", il);
 
     ggml_tensor * conv_input = ggml_concat(ctx0, conv_states, qkv_mixed, 0);
     cb(conv_input, "conv_input", il);
@@ -638,7 +271,10 @@ ggml_tensor * llm_build_qwen35moe ::build_layer_attn_linear(
     ggml_build_forward_expand(gf, ggml_cpy(ctx0, last_conv_states, state_update_target));
     cb(conv_states_all, "conv_states_updated", il);
 
-    // Apply SSM convolution
+    ggml_tensor * state = build_rs(inp, ssm_states_all, hparams.n_embd_s(), n_seqs);
+    state = ggml_reshape_4d(ctx0, state, head_v_dim, head_v_dim, num_v_heads, n_seqs);
+    cb(state, "state_predelta", il);
+
     ggml_tensor * conv_output_proper = ggml_ssm_conv(ctx0, conv_input, conv_kernel);
     cb(conv_output_proper, "conv_output_raw", il);
 
@@ -652,31 +288,41 @@ ggml_tensor * llm_build_qwen35moe ::build_layer_attn_linear(
     int64_t nb1_qkv = ggml_row_size(conv_qkv_mix->type, qkv_dim);
 
     // Extract the convolved Q, K, V from conv_output
-    ggml_tensor * q_conv =
-        ggml_view_2d(ctx0, conv_qkv_mix, head_k_dim * num_k_heads, n_seq_tokens * n_seqs, nb1_qkv, 0);
+    ggml_tensor * q_conv = ggml_view_4d(ctx0, conv_qkv_mix, head_k_dim, num_k_heads, n_seq_tokens, n_seqs,
+            ggml_row_size(conv_qkv_mix->type, head_k_dim),
+            nb1_qkv,
+            nb1_qkv * n_seq_tokens,
+            0);
+
+    ggml_tensor * k_conv = ggml_view_4d(ctx0, conv_qkv_mix, head_k_dim, num_k_heads, n_seq_tokens, n_seqs,
+            ggml_row_size(conv_qkv_mix->type, head_k_dim),
+            nb1_qkv,
+            nb1_qkv * n_seq_tokens,
+            head_k_dim * num_k_heads * ggml_element_size(conv_qkv_mix));
+
+    ggml_tensor * v_conv = ggml_view_4d(ctx0, conv_qkv_mix, head_v_dim, num_v_heads, n_seq_tokens, n_seqs,
+            ggml_row_size(conv_qkv_mix->type, head_v_dim),
+            nb1_qkv,
+            nb1_qkv * n_seq_tokens,
+            ggml_row_size(conv_qkv_mix->type, 2 * head_k_dim * num_k_heads));
+
     cb(q_conv, "q_conv", il);
-    ggml_tensor * k_conv =
-        ggml_view_2d(ctx0, conv_qkv_mix, head_k_dim * num_k_heads, n_seq_tokens * n_seqs, nb1_qkv,
-                     head_k_dim * num_k_heads * ggml_element_size(conv_qkv_mix));
     cb(k_conv, "k_conv", il);
-    ggml_tensor * v_conv =
-        ggml_view_2d(ctx0, conv_qkv_mix, head_v_dim * num_v_heads, n_seq_tokens * n_seqs, nb1_qkv,
-                     2 * head_k_dim * num_k_heads * ggml_element_size(conv_qkv_mix));
     cb(v_conv, "v_conv", il);
 
-    // Unsqueeze them
-    q_conv = ggml_cont_4d(ctx0, q_conv, head_k_dim, num_k_heads, n_seq_tokens, n_seqs);
-    k_conv = ggml_cont_4d(ctx0, k_conv, head_k_dim, num_k_heads, n_seq_tokens, n_seqs);
-    v_conv = ggml_cont_4d(ctx0, v_conv, head_v_dim, num_v_heads, n_seq_tokens, n_seqs);
+    const float eps_norm = hparams.f_norm_rms_eps;
 
-    ggml_tensor * state = build_rs(inp, ssm_states_all, hparams.n_embd_s(), n_seqs);
-    state               = ggml_reshape_4d(ctx0, state, head_v_dim, head_v_dim * num_v_heads, 1, n_seqs);
-    cb(state, "state_predelta", il);
+    q_conv = ggml_l2_norm(ctx0, q_conv, eps_norm);
+    k_conv = ggml_l2_norm(ctx0, k_conv, eps_norm);
 
-    // if head keys and value keys are different, repeat Q/K to match V's head count
-    // V heads are in tiled order (from conversion), so simple tiled repeat works
+    //q_conv = ggml_cont_4d(ctx0, q_conv, head_k_dim, num_k_heads, n_seq_tokens, n_seqs);
+    //k_conv = ggml_cont_4d(ctx0, k_conv, head_k_dim, num_k_heads, n_seq_tokens, n_seqs);
+    //v_conv = ggml_cont_4d(ctx0, v_conv, head_v_dim, num_v_heads, n_seq_tokens, n_seqs);
+
+    // if head keys and value keys are different, repeat to force tensors into matching shapes
     if (num_k_heads != num_v_heads) {
         GGML_ASSERT(num_v_heads % num_k_heads == 0);
+        // TODO: try to avoid these explicit repeats by utilizing op broadcast
         q_conv = ggml_repeat_4d(ctx0, q_conv, head_k_dim, num_v_heads, n_seq_tokens, n_seqs);
         k_conv = ggml_repeat_4d(ctx0, k_conv, head_k_dim, num_v_heads, n_seq_tokens, n_seqs);
     }
@@ -690,7 +336,7 @@ ggml_tensor * llm_build_qwen35moe ::build_layer_attn_linear(
     if (n_seq_tokens == 1) {
         attn_out = build_delta_net_autoregressive(q_conv, k_conv, v_conv, gate, beta, state, il);
     } else {
-        attn_out = build_delta_net_chunking(q_conv, k_conv, v_conv, gate, beta, state, causal_mask, identity, diag_mask, il);
+        attn_out = build_delta_net_chunking(q_conv, k_conv, v_conv, gate, beta, state, il);
     }
     ggml_tensor * output    = attn_out.first;
     ggml_tensor * new_state = attn_out.second;
@@ -699,19 +345,15 @@ ggml_tensor * llm_build_qwen35moe ::build_layer_attn_linear(
 
     // Update the recurrent states
     ggml_build_forward_expand(gf,
-                              ggml_cpy(ctx0, new_state,
-                                       ggml_view_1d(ctx0, ssm_states_all, hparams.n_embd_s() * n_seqs,
-                                                    kv_head * hparams.n_embd_s() * ggml_element_size(ssm_states_all))));
-
-    // Reshape both attn_out_final and z to 2D tensors for normalization
-    // attn_out_final: [head_dim, n_heads, n_tokens, n_seqs] -> [n_heads * n_tokens * n_seqs, head_dim]
-    ggml_tensor * attn_out_2d_final = ggml_reshape_2d(ctx0, output, head_v_dim, num_v_heads * n_seq_tokens * n_seqs);
+            ggml_cpy(ctx0, new_state,
+                ggml_view_1d(ctx0, ssm_states_all, hparams.n_embd_s() * n_seqs,
+                    kv_head * hparams.n_embd_s() * ggml_element_size(ssm_states_all))));
 
     // z: [head_dim, n_heads, n_tokens, n_seqs] -> [n_heads * n_tokens * n_seqs, head_dim]
-    ggml_tensor * z_2d = ggml_reshape_2d(ctx0, z, head_v_dim, num_v_heads * n_seq_tokens * n_seqs);
+    ggml_tensor * z_2d = ggml_reshape_4d(ctx0, z, head_v_dim, num_v_heads, n_seq_tokens, n_seqs);
 
     // Apply gated normalization: self.norm(core_attn_out, z)
-    ggml_tensor * attn_out_norm = build_norm_gated(attn_out_2d_final, model.layers[il].ssm_norm, z_2d, il);
+    ggml_tensor * attn_out_norm = build_norm_gated(output, model.layers[il].ssm_norm, z_2d, il);
 
     // Final reshape: [head_dim, n_heads, n_tokens, n_seqs] -> [n_tokens, n_seqs, n_heads * head_dim]
     ggml_tensor * final_output = ggml_reshape_3d(ctx0, attn_out_norm, head_v_dim * num_v_heads, n_seq_tokens, n_seqs);
@@ -722,7 +364,8 @@ ggml_tensor * llm_build_qwen35moe ::build_layer_attn_linear(
     cb(cur, "linear_attn_out", il);
 
     // Reshape back to original dimensions
-    cur = ggml_cont_2d(ctx0, cur, n_embd, n_seq_tokens * n_seqs);
+    cur = ggml_reshape_2d(ctx0, cur, n_embd, n_seq_tokens * n_seqs);
+
     return cur;
 }
 

src/models/qwen3next.cpp

@@ -306,8 +306,6 @@ ggml_tensor * llm_build_qwen3next::build_layer_attn_linear(
 
     ggml_tensor * beta = ggml_sigmoid(ctx0, b);
 
-    beta = ggml_reshape_4d(ctx0, beta, num_v_heads, 1, n_seq_tokens, n_seqs);
-
     // Reshape a to merge head dimensions: [batch, seq_len, num_k_heads, num_v_heads/num_k_heads] -> [batch, seq_len, num_v_heads]
     ggml_tensor * alpha = ggml_cont_3d(ctx0, a, num_v_heads, n_seq_tokens, n_seqs);
 
@@ -318,6 +316,9 @@ ggml_tensor * llm_build_qwen3next::build_layer_attn_linear(
     ggml_tensor * gate = ggml_mul(ctx0, alpha_softplus, model.layers[il].ssm_a);  // -A_log.exp() * softplus
     cb(gate, "gate", il);
 
+    beta = ggml_reshape_4d(ctx0, beta, 1, num_v_heads, n_seq_tokens, n_seqs);
+    gate = ggml_reshape_4d(ctx0, gate, 1, num_v_heads, n_seq_tokens, n_seqs);
+
     // Get convolution states from cache
     ggml_tensor * conv_states_all = mctx_cur->get_r_l(il);
     ggml_tensor * ssm_states_all  = mctx_cur->get_s_l(il);

tests/test-backend-ops.cpp

@@ -7791,6 +7791,10 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
     test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F16, GGML_TYPE_F32, 1056, 1, 67,  {1,  1}, {4, 1}, {0, 2, 1, 3}));
     test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F32, GGML_TYPE_F32, 16, 32, 32, { 1,  1}, {1, 1}, {0, 1, 2, 3}, 64, 3));
     test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F32, GGML_TYPE_F32, 64, 77, 77, {12,1}, {1,1}));
+    test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F16, GGML_TYPE_F32, 2, 1, 3, {128, 1024}, {1, 1}));
+    test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F16, GGML_TYPE_F32, 2, 3, 4, {128, 1024}, {1, 1}));
+    test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F16, GGML_TYPE_F32, 2, 1, 3, {128*1024, 1}, {1, 1}, {0, 2, 1, 3}));
+    test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F16, GGML_TYPE_F32, 2, 1, 3, {128*1024, 1}, {1, 1}, {0, 1, 2, 3}, 64));
 
     test_cases.emplace_back(new test_mul_mat(GGML_TYPE_Q4_0, GGML_TYPE_F32, 576, 512, 576, {1,1}, {1,1}));
     test_cases.emplace_back(new test_mul_mat(GGML_TYPE_Q4_0, GGML_TYPE_F32, 1, 2048, 8192, {1,  1}, {1, 1}));

tests/test-chat.cpp

@@ -229,6 +229,20 @@ common_chat_tool python_tool {
         "required": ["code"]
     })",
 };
+common_chat_tool todo_list_tool {
+    /* .name = */ "todo_list",
+    /* .description = */ "Create or update the todo list",
+    /* .parameters = */ R"({
+        "type": "object",
+        "properties": {
+            "todos": {
+                "type": "array",
+                "description": "List of TODO list items"
+            }
+        },
+        "required": ["todos"]
+    })",
+};
 common_chat_tool code_interpreter_tool {
     /* .name = */ "code_interpreter",
     /* .description = */ "an ipython interpreter",
@@ -3018,540 +3032,26 @@ Hey there!<|im_end|>
         );
     }
 
-    // Test Qwen3-Coder XML format
     {
-        // Basic XML tool call parsing
-        assert_msg_equals(
-            message_assist_call,
-            test_chat_parse(
-                "<tool_call>\n"
-                "  <function=special_function>\n"
-                "    <parameter=arg1>\n"
-                "      1\n"
-                "    </parameter>\n"
-                "  </function>\n"
-                "</tool_call>",
-                /* is_partial= */ false,
-                {COMMON_CHAT_FORMAT_QWEN3_CODER_XML}));
+        // Step-3.5-Flash template: uses same XML output format as Qwen3-Coder and Nemotron v3,
+        // but with <think> support. Routes to the Nemotron v3 PEG parser for streaming and
+        // schema-aware parameter parsing.
+        auto tmpls = read_templates("models/templates/stepfun-ai-Step-3.5-Flash.jinja");
+        assert_equals(COMMON_CHAT_FORMAT_PEG_CONSTRUCTED, common_chat_templates_apply(tmpls.get(), inputs_tools).format);
 
-        // Multiple parameters with different types
-        common_chat_msg expected_multi_param;
-        expected_multi_param.role = "assistant";
-        expected_multi_param.tool_calls = {
-            { "complex_function", "{\"name\":\"John Doe\",\"age\":30,\"active\":true,\"score\":95.5}", "" }
-        };
-
-        test_parser_with_streaming(expected_multi_param,
-                "<tool_call>\n"
-                "  <function=complex_function>\n"
-                "    <parameter=name>\n"
-                "      John Doe\n"
-                "    </parameter>\n"
-                "    <parameter=age>\n"
-                "      30\n"
-                "    </parameter>\n"
-                "    <parameter=active>\n"
-                "      true\n"
-                "    </parameter>\n"
-                "    <parameter=score>\n"
-                "      95.5\n"
-                "    </parameter>\n"
-                "  </function>\n"
-                "</tool_call>",
-            [&](const std::string &msg) { return test_chat_parse(msg, /* is_partial= */ true, {COMMON_CHAT_FORMAT_QWEN3_CODER_XML}); });
-
-        // Special characters and Unicode
-        common_chat_msg expected_special_chars;
-        expected_special_chars.role = "assistant";
-        expected_special_chars.tool_calls = {
-            { "unicode_function", "{\"message\":\"Hello 世界! 🌍 Special chars: @#$%^&*()\"}", "" }
-        };
-
-        test_parser_with_streaming(expected_special_chars,
-                "<tool_call>\n"
-                "  <function=unicode_function>\n"
-                "    <parameter=message>\n"
-                "      Hello 世界! 🌍 Special chars: @#$%^&*()\n"
-                "    </parameter>\n"
-                "  </function>\n"
-                "</tool_call>",
-            [&](const std::string &msg) { return test_chat_parse(msg, /* is_partial= */ true, {COMMON_CHAT_FORMAT_QWEN3_CODER_XML}); });
-
-        // Multiline content with newlines and indentation
-        common_chat_msg expected_multiline;
-        expected_multiline.role = "assistant";
-        expected_multiline.tool_calls = {
-            { "code_function", "{\"code\":\"def hello():\\n    print(\\\"Hello, World!\\\")\\n    return True\"}", "" }
-        };
-
-        test_parser_with_streaming(expected_multiline,
-                "<tool_call>\n"
-                "  <function=code_function>\n"
-                "    <parameter=code>\n"
-                "def hello():\n"
-                "    print(\"Hello, World!\")\n"
-                "    return True\n"
-                "    </parameter>\n"
-                "  </function>\n"
-                "</tool_call>",
-            [&](const std::string &msg) { return test_chat_parse(msg, /* is_partial= */ true, {COMMON_CHAT_FORMAT_QWEN3_CODER_XML}); });
-
-        // JSON object as parameter value
-        common_chat_msg expected_json_param;
-        expected_json_param.role = "assistant";
-        expected_json_param.tool_calls = {
-            { "json_function", "{\"config\":{\"host\":\"localhost\",\"port\":8080,\"ssl\":false}}", "" }
-        };
-
-        test_parser_with_streaming(
-            expected_json_param,
-                "<tool_call>\n"
-                "  <function=json_function>\n"
-                "    <parameter=config>\n"
-                "      {\"host\": \"localhost\", \"port\": 8080, \"ssl\": false}\n"
-                "    </parameter>\n"
-                "  </function>\n"
-                "</tool_call>",
-            [&](const std::string &msg) { return test_chat_parse(msg, /* is_partial= */ true, {COMMON_CHAT_FORMAT_QWEN3_CODER_XML}); });
-
-        // Array as parameter value
-        common_chat_msg expected_array_param;
-        expected_array_param.role = "assistant";
-        expected_array_param.tool_calls = {
-            { "array_function", "{\"items\":[\"apple\",\"banana\",\"cherry\"]}", "" }
-        };
-
-        test_parser_with_streaming(
-            expected_array_param,
-                "<tool_call>\n"
-                "  <function=array_function>\n"
-                "    <parameter=items>\n"
-                "      [\"apple\", \"banana\", \"cherry\"]\n"
-                "    </parameter>\n"
-                "  </function>\n"
-                "</tool_call>",
-            [&](const std::string &msg) { return test_chat_parse(msg, /* is_partial= */ true, {COMMON_CHAT_FORMAT_QWEN3_CODER_XML}); });
-
-        // Empty parameter
-        common_chat_msg expected_empty_param;
-        expected_empty_param.role = "assistant";
-        expected_empty_param.tool_calls = {
-            { "empty_function", "{\"empty_param\":\"\"}", "" }
-        };
-
-        test_parser_with_streaming(
-            expected_empty_param,
-                "<tool_call>\n"
-                "  <function=empty_function>\n"
-                "    <parameter=empty_param>\n"
-                "    </parameter>\n"
-                "  </function>\n"
-                "</tool_call>",
-            [&](const std::string &msg) { return test_chat_parse(msg, /* is_partial= */ true, {COMMON_CHAT_FORMAT_QWEN3_CODER_XML}); });
-
-        // Boolean values (true/false)
-        common_chat_msg expected_boolean;
-        expected_boolean.role = "assistant";
-        expected_boolean.tool_calls = {
-            { "boolean_function", "{\"enabled\":true,\"debug\":false}", "" }
-        };
-
-        test_parser_with_streaming(
-            expected_boolean,
-                "<tool_call>\n"
-                "  <function=boolean_function>\n"
-                "    <parameter=enabled>\n"
-                "      true\n"
-                "    </parameter>\n"
-                "    <parameter=debug>\n"
-                "      false\n"
-                "    </parameter>\n"
-                "  </function>\n"
-                "</tool_call>",
-            [&](const std::string &msg) { return test_chat_parse(msg, /* is_partial= */ true, {COMMON_CHAT_FORMAT_QWEN3_CODER_XML}); });
-
-        // Null value
-        common_chat_msg expected_null;
-        expected_null.role = "assistant";
-        expected_null.tool_calls = {
-            { "null_function", "{\"optional_param\":null}", "" }
-        };
-
-        test_parser_with_streaming(
-            expected_null,
-                "<tool_call>\n"
-                "  <function=null_function>\n"
-                "    <parameter=optional_param>\n"
-                "      null\n"
-                "    </parameter>\n"
-                "  </function>\n"
-                "</tool_call>",
-            [&](const std::string &msg) { return test_chat_parse(msg, /* is_partial= */ true, {COMMON_CHAT_FORMAT_QWEN3_CODER_XML}); });
-
-        // Negative numbers and scientific notation
-        common_chat_msg expected_numbers;
-        expected_numbers.role = "assistant";
-        expected_numbers.tool_calls = {
-            { "math_function", "{\"negative\":-42,\"decimal\":-3.14,\"scientific\":1.23e-4}", "" }
-        };
-
-        test_parser_with_streaming(
-            expected_numbers,
-                "<tool_call>\n"
-                "  <function=math_function>\n"
-                "    <parameter=negative>\n"
-                "      -42\n"
-                "    </parameter>\n"
-                "    <parameter=decimal>\n"
-                "      -3.14\n"
-                "    </parameter>\n"
-                "    <parameter=scientific>\n"
-                "      1.23e-4\n"
-                "    </parameter>\n"
-                "  </function>\n"
-                "</tool_call>",
-            [&](const std::string &msg) { return test_chat_parse(msg, /* is_partial= */ true, {COMMON_CHAT_FORMAT_QWEN3_CODER_XML}); });
-
-        // XML-like content in parameters (should be escaped)
-        common_chat_msg expected_xml_content;
-        expected_xml_content.role = "assistant";
-        expected_xml_content.tool_calls = {
-            { "xml_function", "{\"xml_content\":\"<root><item>value</item></root>\"}", "" }
-        };
-
-        test_parser_with_streaming(
-            expected_xml_content,
-                "<tool_call>\n"
-                "  <function=xml_function>\n"
-                "    <parameter=xml_content>\n"
-                "      <root><item>value</item></root>\n"
-                "    </parameter>\n"
-                "  </function>\n"
-                "</tool_call>",
-            [&](const std::string &msg) { return test_chat_parse(msg, /* is_partial= */ true, {COMMON_CHAT_FORMAT_QWEN3_CODER_XML}); });
-
-        // Quotes and escape characters
-        common_chat_msg expected_quotes;
-        expected_quotes.role = "assistant";
-        expected_quotes.tool_calls = {
-            { "quote_function", "{\"message\":\"She said \\\"Hello!\\\" and left.\"}", "" }
-        };
-
-        test_parser_with_streaming(
-            expected_quotes,
-                "<tool_call>\n"
-                "  <function=quote_function>\n"
-                "    <parameter=message>\n"
-                "      She said \"Hello!\" and left.\n"
-                "    </parameter>\n"
-                "  </function>\n"
-                "</tool_call>",
-            [&](const std::string &msg) { return test_chat_parse(msg, /* is_partial= */ true, {COMMON_CHAT_FORMAT_QWEN3_CODER_XML}); });
-
-        // Long parameter value (simplified)
-        std::string long_text = "This is a long text parameter that should test the parser's ability to handle larger amounts of text data.";
-
-        common_chat_msg expected_long_text;
-        expected_long_text.role = "assistant";
-        expected_long_text.tool_calls = {
-            { "long_function", "{\"long_text\":\"" + long_text + "\"}", "" }
-        };
-
-        test_parser_with_streaming(
-            expected_long_text,
-                "<tool_call>\n"
-                "  <function=long_function>\n"
-                "    <parameter=long_text>\n"
-                "      " + long_text + "\n"
-                "    </parameter>\n"
-                "  </function>\n"
-                "</tool_call>",
-            [&](const std::string &msg) { return test_chat_parse(msg, /* is_partial= */ true, {COMMON_CHAT_FORMAT_QWEN3_CODER_XML}); });
-
-        // Mixed content with text before and after tool call
-        common_chat_msg expected_mixed_content;
-        expected_mixed_content.role = "assistant";
-        expected_mixed_content.content = "I'll help you search for products. ";
-        expected_mixed_content.tool_calls = {
-            { "search_function", "{\"query\":\"laptops\"}", "" }
-        };
-
-        test_parser_with_streaming(
-            expected_mixed_content,
-                "I'll help you search for products. <tool_call>\n"
-                "  <function=search_function>\n"
-                "    <parameter=query>\n"
-                "      laptops\n"
-                "    </parameter>\n"
-                "  </function>\n"
-                "</tool_call>",
-            [&](const std::string &msg) { return test_chat_parse(msg, /* is_partial= */ true, {COMMON_CHAT_FORMAT_QWEN3_CODER_XML}); });
-
-        // Compact format (no extra whitespace)
-        common_chat_msg expected_compact;
-        expected_compact.role = "assistant";
-        expected_compact.tool_calls = {
-            { "compact_function", "{\"param\":\"value\"}", "" }
-        };
-
-        test_parser_with_streaming(
-            expected_compact,
-                "<tool_call><function=compact_function><parameter=param>value</parameter></function></tool_call>",
-            [&](const std::string &msg) { return test_chat_parse(msg, /* is_partial= */ true, {COMMON_CHAT_FORMAT_QWEN3_CODER_XML}); });
-
-        // Function name with underscores and numbers
-        common_chat_msg expected_complex_name;
-        expected_complex_name.role = "assistant";
-        expected_complex_name.tool_calls = {
-            { "get_user_data_v2", "{\"user_id\":12345}", "" }
-        };
-
-        test_parser_with_streaming(
-            expected_complex_name,
-                "<tool_call>\n"
-                "  <function=get_user_data_v2>\n"
-                "    <parameter=user_id>\n"
-                "      12345\n"
-                "    </parameter>\n"
-                "  </function>\n"
-                "</tool_call>",
-            [&](const std::string &msg) { return test_chat_parse(msg, /* is_partial= */ true, {COMMON_CHAT_FORMAT_QWEN3_CODER_XML}); });
-
-        // Parameter names with underscores and numbers
-        common_chat_msg expected_complex_params;
-        expected_complex_params.role = "assistant";
-        expected_complex_params.tool_calls = {
-            { "test_function", "{\"param_1\":\"value1\",\"param_2_name\":\"value2\",\"param3\":123}", "" }
-        };
-
-        test_parser_with_streaming(
-            expected_complex_params,
-                "<tool_call>\n"
-                "  <function=test_function>\n"
-                "    <parameter=param_1>\n"
-                "      value1\n"
-                "    </parameter>\n"
-                "    <parameter=param_2_name>\n"
-                "      value2\n"
-                "    </parameter>\n"
-                "    <parameter=param3>\n"
-                "      123\n"
-                "    </parameter>\n"
-                "  </function>\n"
-                "</tool_call>",
-            [&](const std::string &msg) { return test_chat_parse(msg, /* is_partial= */ true, {COMMON_CHAT_FORMAT_QWEN3_CODER_XML}); });
-
-        // Very deeply nested XML content in parameter
-        common_chat_msg expected_deep_xml;
-        expected_deep_xml.role = "assistant";
-        expected_deep_xml.tool_calls = {
-            { "xml_parser", "{\"xml\":\"<root><level1><level2><level3>deep content</level3></level2></level1></root>\"}", "" }
-        };
-
-        test_parser_with_streaming(
-            expected_deep_xml,
-                "<tool_call>\n"
-                "  <function=xml_parser>\n"
-                "    <parameter=xml>\n"
-                "      <root><level1><level2><level3>deep content</level3></level2></level1></root>\n"
-                "    </parameter>\n"
-                "  </function>\n"
-                "</tool_call>",
-            [&](const std::string &msg) { return test_chat_parse(msg, /* is_partial= */ true, {COMMON_CHAT_FORMAT_QWEN3_CODER_XML}); });
-
-        // Parameter with only whitespace
-        common_chat_msg expected_whitespace_param;
-        expected_whitespace_param.role = "assistant";
-        expected_whitespace_param.tool_calls = {
-            { "whitespace_function", "{\"spaces\":\"\"}", "" }
-        };
-
-        test_parser_with_streaming(
-            expected_whitespace_param,
-                "<tool_call>\n"
-                "  <function=whitespace_function>\n"
-                "    <parameter=spaces>\n"
-                "      \n"
-                "    </parameter>\n"
-                "  </function>\n"
-                "</tool_call>",
-            [&](const std::string &msg) { return test_chat_parse(msg, /* is_partial= */ true, {COMMON_CHAT_FORMAT_QWEN3_CODER_XML}); });
-
-        // Parameter with tabs and mixed whitespace
-        common_chat_msg expected_mixed_whitespace;
-        expected_mixed_whitespace.role = "assistant";
-        expected_mixed_whitespace.tool_calls = {
-            { "tab_function", "{\"content\":\"line1\\n\\tindented line\\n    spaces\"}", "" }
-        };
-
-        test_parser_with_streaming(
-            expected_mixed_whitespace,
-                "<tool_call>\n"
-                "  <function=tab_function>\n"
-                "    <parameter=content>\n"
-                "line1\n"
-                "\tindented line\n"
-                "    spaces\n"
-                "    </parameter>\n"
-                "  </function>\n"
-                "</tool_call>",
-            [&](const std::string &msg) { return test_chat_parse(msg, /* is_partial= */ true, {COMMON_CHAT_FORMAT_QWEN3_CODER_XML}); });
-
-        // Control characters and special Unicode
-        common_chat_msg expected_control_chars;
-        expected_control_chars.role = "assistant";
-        expected_control_chars.tool_calls = {
-            { "control_function", "{\"text\":\"Line1\\nLine2\\tTabbed\\rCarriage return\"}", "" }
-        };
-
-        test_parser_with_streaming(
-            expected_control_chars,
-                "<tool_call>\n"
-                "  <function=control_function>\n"
-                "    <parameter=text>\n"
-                "Line1\nLine2\tTabbed\rCarriage return\n"
-                "    </parameter>\n"
-                "  </function>\n"
-                "</tool_call>",
-            [&](const std::string &msg) { return test_chat_parse(msg, /* is_partial= */ true, {COMMON_CHAT_FORMAT_QWEN3_CODER_XML}); });
-
-        // Emoji and extended Unicode characters
-        common_chat_msg expected_emoji;
-        expected_emoji.role = "assistant";
-        expected_emoji.tool_calls = {
-            { "emoji_function", "{\"message\":\"Hello! 👋 🌟 🚀 Testing emojis: 😀😃😄😁 and symbols: ∑∏∆∇\"}", "" }
-        };
-
-        test_parser_with_streaming(
-            expected_emoji,
-                "<tool_call>\n"
-                "  <function=emoji_function>\n"
-                "    <parameter=message>\n"
-                "      Hello! 👋 🌟 🚀 Testing emojis: 😀😃😄😁 and symbols: ∑∏∆∇\n"
-                "    </parameter>\n"
-                "  </function>\n"
-                "</tool_call>",
-            [&](const std::string &msg) { return test_chat_parse(msg, /* is_partial= */ true, {COMMON_CHAT_FORMAT_QWEN3_CODER_XML}); });
-
-        // Mathematical expressions and formulas
-        common_chat_msg expected_math;
-        expected_math.role = "assistant";
-        expected_math.tool_calls = {
-            { "math_function", "{\"formula\":\"E = mc² and ∫f(x)dx = F(x) + C\"}", "" }
-        };
-
-        test_parser_with_streaming(
-            expected_math,
-                "<tool_call>\n"
-                "  <function=math_function>\n"
-                "    <parameter=formula>\n"
-                "      E = mc² and ∫f(x)dx = F(x) + C\n"
-                "    </parameter>\n"
-                "  </function>\n"
-                "</tool_call>",
-            [&](const std::string &msg) { return test_chat_parse(msg, /* is_partial= */ true, {COMMON_CHAT_FORMAT_QWEN3_CODER_XML}); });
-
-        // SQL injection-like content (should be safely escaped)
-        common_chat_msg expected_sql;
-        expected_sql.role = "assistant";
-        expected_sql.tool_calls = {
-            { "sql_function", "{\"query\":\"SELECT * FROM users WHERE id = 1; DROP TABLE users; --\"}", "" }
-        };
-
-        test_parser_with_streaming(
-            expected_sql,
-                "<tool_call>\n"
-                "  <function=sql_function>\n"
-                "    <parameter=query>\n"
-                "      SELECT * FROM users WHERE id = 1; DROP TABLE users; --\n"
-                "    </parameter>\n"
-                "  </function>\n"
-                "</tool_call>",
-            [&](const std::string &msg) { return test_chat_parse(msg, /* is_partial= */ true, {COMMON_CHAT_FORMAT_QWEN3_CODER_XML}); });
-
-        // HTML/XML injection content
-        common_chat_msg expected_html;
-        expected_html.role = "assistant";
-        expected_html.tool_calls = {
-            { "html_function", "{\"content\":\"<script>alert('xss')</script><img src=x onerror=alert(1)>\"}", "" }
-        };
-
-        test_parser_with_streaming(
-            expected_html,
-                "<tool_call>\n"
-                "  <function=html_function>\n"
-                "    <parameter=content>\n"
-                "      <script>alert('xss')</script><img src=x onerror=alert(1)>\n"
-                "    </parameter>\n"
-                "  </function>\n"
-                "</tool_call>",
-            [&](const std::string &msg) { return test_chat_parse(msg, /* is_partial= */ true, {COMMON_CHAT_FORMAT_QWEN3_CODER_XML}); });
-
-        // Binary-like content (base64)
-        common_chat_msg expected_binary;
-        expected_binary.role = "assistant";
-        expected_binary.tool_calls = {
-            { "binary_function", "{\"data\":\"SGVsbG8gV29ybGQhIFRoaXMgaXMgYmFzZTY0IGVuY29kZWQgdGV4dC4=\"}", "" }
-        };
-
-        test_parser_with_streaming(
-            expected_binary,
-                "<tool_call>\n"
-                "  <function=binary_function>\n"
-                "    <parameter=data>\n"
-                "      SGVsbG8gV29ybGQhIFRoaXMgaXMgYmFzZTY0IGVuY29kZWQgdGV4dC4=\n"
-                "    </parameter>\n"
-                "  </function>\n"
-                "</tool_call>",
-            [&](const std::string &msg) { return test_chat_parse(msg, /* is_partial= */ true, {COMMON_CHAT_FORMAT_QWEN3_CODER_XML}); });
-
-        // Very large numbers (should be parsed as scientific notation)
-        common_chat_msg expected_large_numbers;
-        expected_large_numbers.role = "assistant";
-        expected_large_numbers.tool_calls = {
-            { "number_function", "{\"big_int\":1e+60}", "" }  // Large number becomes scientific notation
-        };
-
-        test_parser_with_streaming(
-            expected_large_numbers,
-                "<tool_call>\n"
-                "  <function=number_function>\n"
-                "    <parameter=big_int>\n"
-                "      999999999999999999999999999999999999999999999999999999999999\n"
-                "    </parameter>\n"
-                "  </function>\n"
-                "</tool_call>",
-            [&](const std::string &msg) { return test_chat_parse(msg, /* is_partial= */ true, {COMMON_CHAT_FORMAT_QWEN3_CODER_XML}); });
-    }
-
-    {
-        // Qwen3-Coder template
-        auto tmpls = read_templates("models/templates/Qwen3-Coder.jinja");
-        common_chat_templates_inputs inputs;
-        inputs.messages = { message_user };
-
-        common_chat_tool qwen_union_tool {
-            /* .name = */ "qwen_union",
-            /* .description = */ "Test tool for union/anyOf handling",
-            /* .parameters = */ R"({
-                "type": "object",
-                "properties": {
-                    "priority": { "type": ["number", "null"] },
-                    "maybe_text": { "anyOf": [ { "type": "string" } ] },
-                    "config": { "anyOf": [ { "type": "object" }, { "type": "null" } ] }
-                },
-                "required": []
-            })",
-        };
-        inputs.tools = { qwen_union_tool };
-
-        auto params = common_chat_templates_apply(tmpls.get(), inputs);
-        assert_equals(COMMON_CHAT_FORMAT_QWEN3_CODER_XML, params.format);
-        assert_equals(false, params.grammar.empty());
-
-        // Grammar should compile successfully
-        auto grammar = build_grammar(params.grammar);
-        GGML_ASSERT(grammar && "Failed to build Qwen3-Coder grammar with union types");
+        // Grammar and PEG parser should be generated with thinking_forced_open
+        {
+            common_chat_templates_inputs inputs;
+            inputs.messages = { message_user };
+            inputs.tools = { special_function_tool };
+            auto params = common_chat_templates_apply(tmpls.get(), inputs);
+            assert_equals(COMMON_CHAT_FORMAT_PEG_CONSTRUCTED, params.format);
+            assert_equals(true, params.thinking_forced_open);
+            assert_equals(false, params.grammar.empty());
+            assert_equals(false, params.parser.empty());
+            auto grammar = build_grammar(params.grammar);
+            GGML_ASSERT(grammar && "Failed to build Step-3.5-Flash grammar");
+        }
     }
 }
 
@@ -3643,6 +3143,135 @@ static void test_template_output_peg_parsers() {
         });
     }
 
+    {
+        // Qwen3-Coder
+        auto tmpls = read_templates("models/templates/Qwen3-Coder.jinja");
+
+        // Test basic message
+        test_peg_parser(tmpls.get(), [&](auto & t) {
+            t.input = "Hello, world!\nWhat's up?";
+            t.expect = message_assist;
+        });
+
+        // Test tool call
+        test_peg_parser(tmpls.get(), [&](auto & t) {
+            t.input =
+                "<tool_call>\n"
+                "<function=special_function>\n"
+                "<parameter=arg1>\n"
+                "1\n"
+                "</parameter>\n"
+                "</function>\n"
+                "</tool_call>";
+            t.params.tools = {special_function_tool};
+            t.expect = message_assist_call;
+        });
+
+        // Test parallel tool calls
+        test_peg_parser(tmpls.get(), [&](auto & t) {
+            t.input =
+                "<tool_call>\n"
+                "<function=special_function>\n"
+                "<parameter=arg1>\n"
+                "1\n"
+                "</parameter>\n"
+                "</function>\n"
+                "</tool_call>\n"
+                "<tool_call>\n"
+                "<function=special_function_with_opt>\n"
+                "<parameter=arg1>\n"
+                "1\n"
+                "</parameter>\n"
+                "<parameter=arg2>\n"
+                "2\n"
+                "</parameter>\n"
+                "</function>\n"
+                "</tool_call>";
+            t.params.parallel_tool_calls = true;
+            t.params.tools = {special_function_tool, special_function_tool_with_optional_param};
+
+            t.expect.tool_calls = {{
+                /* .name = */      "special_function",
+                /* .arguments = */ R"({"arg1": 1})",
+                /* .id = */        {},
+            }, {
+                /* .name = */      "special_function_with_opt",
+                /* .arguments = */ R"({"arg1": 1, "arg2": 2})",
+                /* .id = */        {},
+            }};
+        });
+
+        // Test tool call with string parameter
+        test_peg_parser(tmpls.get(), [&](auto & t) {
+            t.input =
+                "<tool_call>\n"
+                "<function=python>\n"
+                "<parameter=code>\n"
+                "def hello():\n"
+                "    print(\"Hello, world!\")\n"
+                "\n"
+                "hello()\n"
+                "</parameter>\n"
+                "</function>\n"
+                "</tool_call>";
+            t.params.tools = {python_tool};
+
+            t.expect.tool_calls = {{
+                /* .name = */      "python",
+                /* .arguments = */ "{\"code\": \"def hello():\\n    print(\\\"Hello, world!\\\")\\n\\nhello()\"}",
+                /* .id = */        {},
+            }};
+        });
+
+        // Test tool call with JSON parameter
+        test_peg_parser(tmpls.get(), [&](auto & t) {
+            t.input =
+                "<tool_call>\n"
+                "<function=todo_list>\n"
+                "<parameter=todos>\n"
+                "[{\"item\": \"Check stuff\", \"selected\": false}, {\"item\": \"Prepare stuff\", \"selected\": true}]\n"
+                "</parameter>\n"
+                "</function>\n"
+                "</tool_call>";
+            t.params.tools = {todo_list_tool};
+
+            t.expect.tool_calls = {{
+                /* .name = */      "todo_list",
+                /* .arguments = */ "{\"todos\": [{\"item\": \"Check stuff\", \"selected\": false}, {\"item\": \"Prepare stuff\", \"selected\": true}]}",
+                /* .id = */        {},
+            }};
+        });
+
+        // Test tool call with string parameter and no closing </parameter> tag
+        test_peg_parser(tmpls.get(), [&](auto & t) {
+            t.input =
+                "<tool_call>\n"
+                "<function=python>\n"
+                "<parameter=code>\n"
+                "def hello():\n"
+                "    print(\"Hello, world!\")\n"
+                "\n"
+                "hello()\n"
+                "</function>\n"
+                "</tool_call>";
+            t.params.tools = {python_tool};
+
+            t.expect.tool_calls = {{
+                /* .name = */      "python",
+                /* .arguments = */ "{\"code\": \"def hello():\\n    print(\\\"Hello, world!\\\")\\n\\nhello()\"}",
+                /* .id = */        {},
+            }};
+        });
+
+        // Test response format
+        test_peg_parser(tmpls.get(), [&](auto & t) {
+            t.input = R"({"amount": 123.45, "date": "2025-12-03"})";
+            t.params.json_schema = invoice_schema;
+
+            t.expect.content = R"({"amount": 123.45, "date": "2025-12-03"})";
+        });
+    }
+
     {
         // NVIDIA Nemotron-3 Nano
         auto tmpls = read_templates("models/templates/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16.jinja");
@@ -3799,6 +3428,196 @@ static void test_template_output_peg_parsers() {
         });
     }
 
+    {
+        // Step-3.5-Flash (uses Nemotron v3 PEG parser with thinking_forced_open)
+        // Unlike Nemotron, Step-3.5-Flash always emits <think> regardless of enable_thinking,
+        // so all inputs must include a </think> delimiter.
+        auto tmpls = read_templates("models/templates/stepfun-ai-Step-3.5-Flash.jinja");
+
+        // Test basic message with reasoning
+        test_peg_parser(tmpls.get(), [&](auto & t) {
+            t.input = "I'm\nthinking\n</think>\nHello, world!\nWhat's up?";
+            t.params.reasoning_format = COMMON_REASONING_FORMAT_AUTO;
+
+            t.expect = message_assist_thoughts;
+        });
+
+        // Test basic message without thinking content
+        test_peg_parser(tmpls.get(), [&](auto & t) {
+            t.input = "</think>\nHello, world!\nWhat's up?";
+            t.params.reasoning_format = COMMON_REASONING_FORMAT_AUTO;
+
+            t.expect = message_assist;
+        });
+
+        // Test tool call without thinking content
+        test_peg_parser(tmpls.get(), [&](auto & t) {
+            t.input =
+                "</think>\n"
+                "<tool_call>\n"
+                "<function=special_function>\n"
+                "<parameter=arg1>\n"
+                "1\n"
+                "</parameter>\n"
+                "</function>\n"
+                "</tool_call>";
+            t.params.reasoning_format = COMMON_REASONING_FORMAT_AUTO;
+            t.params.tools = {special_function_tool};
+
+            t.expect = message_assist_call;
+        });
+
+        // Test tool call with thinking
+        test_peg_parser(tmpls.get(), [&](auto & t) {
+            t.input =
+                "I'm\nthinking\n</think>\n"
+                "<tool_call>\n"
+                "<function=special_function>\n"
+                "<parameter=arg1>\n"
+                "1\n"
+                "</parameter>\n"
+                "</function>\n"
+                "</tool_call>";
+            t.params.reasoning_format = COMMON_REASONING_FORMAT_AUTO;
+            t.params.tools = {special_function_tool};
+
+            t.expect = message_assist_call_thoughts;
+        });
+
+        // Test parallel tool calls with thinking
+        test_peg_parser(tmpls.get(), [&](auto & t) {
+            t.input =
+                "I'm\nthinking\n</think>\n"
+                "<tool_call>\n"
+                "<function=special_function>\n"
+                "<parameter=arg1>\n"
+                "1\n"
+                "</parameter>\n"
+                "</function>\n"
+                "</tool_call>\n"
+                "<tool_call>\n"
+                "<function=special_function_with_opt>\n"
+                "<parameter=arg1>\n"
+                "1\n"
+                "</parameter>\n"
+                "<parameter=arg2>\n"
+                "2\n"
+                "</parameter>\n"
+                "</function>\n"
+                "</tool_call>";
+            t.params.reasoning_format = COMMON_REASONING_FORMAT_AUTO;
+            t.params.parallel_tool_calls = true;
+            t.params.tools = {special_function_tool, special_function_tool_with_optional_param};
+
+            t.expect.reasoning_content = "I'm\nthinking";
+            t.expect.tool_calls = {{
+                /* .name = */      "special_function",
+                /* .arguments = */ R"({"arg1": 1})",
+                /* .id = */        {},
+            }, {
+                /* .name = */      "special_function_with_opt",
+                /* .arguments = */ R"({"arg1": 1, "arg2": 2})",
+                /* .id = */        {},
+            }};
+        });
+
+        // Test parallel tool calls without thinking content
+        test_peg_parser(tmpls.get(), [&](auto & t) {
+            t.input =
+                "</think>\n"
+                "<tool_call>\n"
+                "<function=special_function>\n"
+                "<parameter=arg1>\n"
+                "1\n"
+                "</parameter>\n"
+                "</function>\n"
+                "</tool_call>\n"
+                "<tool_call>\n"
+                "<function=special_function_with_opt>\n"
+                "<parameter=arg1>\n"
+                "1\n"
+                "</parameter>\n"
+                "<parameter=arg2>\n"
+                "2\n"
+                "</parameter>\n"
+                "</function>\n"
+                "</tool_call>";
+            t.params.reasoning_format = COMMON_REASONING_FORMAT_AUTO;
+            t.params.parallel_tool_calls = true;
+            t.params.tools = {special_function_tool, special_function_tool_with_optional_param};
+
+            t.expect.tool_calls = {{
+                /* .name = */      "special_function",
+                /* .arguments = */ R"({"arg1": 1})",
+                /* .id = */        {},
+            }, {
+                /* .name = */      "special_function_with_opt",
+                /* .arguments = */ R"({"arg1": 1, "arg2": 2})",
+                /* .id = */        {},
+            }};
+        });
+
+        // Test tool call with code string parameter
+        test_peg_parser(tmpls.get(), [&](auto & t) {
+            t.input =
+                "</think>\n"
+                "<tool_call>\n"
+                "<function=python>\n"
+                "<parameter=code>\n"
+                "def hello():\n"
+                "    print(\"Hello, world!\")\n"
+                "\n"
+                "hello()\n"
+                "</parameter>\n"
+                "</function>\n"
+                "</tool_call>";
+            t.params.reasoning_format = COMMON_REASONING_FORMAT_AUTO;
+            t.params.tools = {python_tool};
+
+            t.expect.tool_calls = {{
+                /* .name = */      "python",
+                /* .arguments = */ "{\"code\": \"def hello():\\n    print(\\\"Hello, world!\\\")\\n\\nhello()\"}",
+                /* .id = */        {},
+            }};
+        });
+
+        // Test tool call with string parameter and no closing </parameter> tag
+        test_peg_parser(tmpls.get(), [&](auto & t) {
+            t.input =
+                "</think>\n"
+                "<tool_call>\n"
+                "<function=python>\n"
+                "<parameter=code>\n"
+                "def hello():\n"
+                "    print(\"Hello, world!\")\n"
+                "\n"
+                "hello()\n"
+                "</function>\n"
+                "</tool_call>";
+            t.params.reasoning_format = COMMON_REASONING_FORMAT_AUTO;
+            t.params.tools = {python_tool};
+
+            t.expect.tool_calls = {{
+                /* .name = */      "python",
+                /* .arguments = */ "{\"code\": \"def hello():\\n    print(\\\"Hello, world!\\\")\\n\\nhello()\"}",
+                /* .id = */        {},
+            }};
+        });
+
+        // Test response format (JSON schema with thinking)
+        test_peg_parser(tmpls.get(), [&](auto & t) {
+            t.input =
+              "I need to output the invoice details in JSON\n"
+              "</think>\n"
+              R"({"amount": 123.45, "date": "2025-12-03"})";
+            t.params.reasoning_format = COMMON_REASONING_FORMAT_AUTO;
+            t.params.json_schema = invoice_schema;
+
+            t.expect.reasoning_content = "I need to output the invoice details in JSON";
+            t.expect.content = R"({"amount": 123.45, "date": "2025-12-03"})";
+        });
+    }
+
     {
         // Solar-Open-100B
         auto tmpls = read_templates("models/templates/upstage-Solar-Open-100B.jinja");

tests/test-jinja.cpp

@@ -691,6 +691,48 @@ static void test_filters(testing & t) {
         "{\n  \"a\": 1,\n  \"b\": [\n    1,\n    2\n  ]\n}"
     );
 
+    test_template(t, "indent",
+        "{{ data|indent(2) }}",
+        {{ "data", "foo\nbar" }},
+        "foo\n  bar"
+    );
+
+    test_template(t, "indent first only",
+        "{{ data|indent(width=3,first=true) }}",
+        {{ "data", "foo\nbar" }},
+        "   foo\n   bar"
+    );
+
+    test_template(t, "indent blank lines and first line",
+        "{{ data|indent(width=5,blank=true,first=true) }}",
+        {{ "data", "foo\n\nbar" }},
+        "     foo\n     \n     bar"
+    );
+
+    test_template(t, "indent with default width",
+        "{{ data|indent() }}",
+        {{ "data", "foo\nbar" }},
+        "foo\n    bar"
+    );
+
+    test_template(t, "indent with no newline",
+        "{{ data|indent }}",
+        {{ "data", "foo" }},
+        "foo"
+    );
+
+    test_template(t, "indent with trailing newline",
+        "{{ data|indent(blank=true) }}",
+        {{ "data", "foo\n" }},
+        "foo\n    "
+    );
+
+    test_template(t, "indent with string",
+        "{{ data|indent(width='>>>>') }}",
+        {{ "data", "foo\nbar" }},
+        "foo\n>>>>bar"
+    );
+
     test_template(t, "chained filters",
         "{{ '  HELLO  '|trim|lower }}",
         json::object(),

tools/mtmd/CMakeLists.txt

@@ -24,6 +24,7 @@ add_library(mtmd
             models/llama4.cpp
             models/llava.cpp
             models/minicpmv.cpp
+            models/paddleocr.cpp
             models/pixtral.cpp
             models/qwen2vl.cpp
             models/qwen3vl.cpp

tools/mtmd/clip-impl.h

@@ -229,6 +229,7 @@ enum projector_type {
     PROJECTOR_TYPE_MUSIC_FLAMINGO,
     PROJECTOR_TYPE_LFM2,
     PROJECTOR_TYPE_KIMIVL,
+    PROJECTOR_TYPE_PADDLEOCR,
     PROJECTOR_TYPE_LIGHTONOCR,
     PROJECTOR_TYPE_COGVLM,
     PROJECTOR_TYPE_JANUS_PRO,
@@ -264,6 +265,7 @@ static std::map<projector_type, std::string> PROJECTOR_TYPE_NAMES = {
     { PROJECTOR_TYPE_MUSIC_FLAMINGO, "musicflamingo"},
     { PROJECTOR_TYPE_LFM2,      "lfm2"},
     { PROJECTOR_TYPE_KIMIVL,    "kimivl"},
+    { PROJECTOR_TYPE_PADDLEOCR, "paddleocr"},
     { PROJECTOR_TYPE_LIGHTONOCR,"lightonocr"},
     { PROJECTOR_TYPE_COGVLM,    "cogvlm"},
     { PROJECTOR_TYPE_JANUS_PRO, "janus_pro"},

tools/mtmd/clip.cpp

@@ -628,9 +628,6 @@ ggml_tensor * clip_graph::build_attn(
         ggml_tensor * v = ggml_permute(ctx0, v_cur, 1, 2, 0, 3);
         v = ggml_cont(ctx0, v);
 
-        const auto n_tokens = q->ne[1];
-        const auto n_head   = q->ne[2];
-
         ggml_tensor * kq = ggml_mul_mat(ctx0, k, q);
         // F32 may not needed for vision encoders?
         // ggml_mul_mat_set_prec(kq, GGML_PREC_F32);
@@ -639,7 +636,7 @@ ggml_tensor * clip_graph::build_attn(
 
         ggml_tensor * kqv = ggml_mul_mat(ctx0, v, kq);
         cur = ggml_permute(ctx0, kqv, 0, 2, 1, 3);
-        cur = ggml_cont_2d(ctx0, cur, cur->ne[0]*n_head, n_tokens);
+        cur = ggml_cont_2d(ctx0, cur, cur->ne[0] * cur->ne[1], cur->ne[2] * cur->ne[3]);
     }
 
     cb(cur, "kqv_out", il);
@@ -844,6 +841,10 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
             {
                 builder = std::make_unique<clip_graph_kimivl>(ctx, img);
             } break;
+        case PROJECTOR_TYPE_PADDLEOCR:
+            {
+                builder = std::make_unique<clip_graph_paddleocr>(ctx, img);
+            } break;
         case PROJECTOR_TYPE_KIMIK25:
             {
                 builder = std::make_unique<clip_graph_kimik25>(ctx, img);
@@ -1259,6 +1260,14 @@ struct clip_model_loader {
                         hparams.audio_window_len   = 400;
                         hparams.audio_hop_len      = 160;
                     } break;
+                case PROJECTOR_TYPE_PADDLEOCR:
+                    {
+                        hparams.n_merge = 2;
+                        get_u32(KEY_IMAGE_MIN_PIXELS, hparams.image_min_pixels);
+                        get_u32(KEY_IMAGE_MAX_PIXELS, hparams.image_max_pixels);
+
+                        hparams.set_warmup_n_tokens(28*28); // avoid OOM on warmup
+                    } break;
                 case PROJECTOR_TYPE_LFM2A:
                     {
                         // audio preprocessing params
@@ -1707,6 +1716,7 @@ struct clip_model_loader {
                     model.mm_2_b = get_tensor(string_format(TN_LLAVA_PROJ, 2, "bias"));
                 } break;
             case PROJECTOR_TYPE_KIMIVL:
+            case PROJECTOR_TYPE_PADDLEOCR:
             case PROJECTOR_TYPE_KIMIK25:
                 {
                     model.mm_input_norm_w = get_tensor(TN_MM_INP_NORM);
@@ -2993,6 +3003,7 @@ bool clip_image_preprocess(struct clip_ctx * ctx, const clip_image_u8 * img, str
         case PROJECTOR_TYPE_QWEN25VL:
         case PROJECTOR_TYPE_QWEN3VL:
         case PROJECTOR_TYPE_GLM4V:
+        case PROJECTOR_TYPE_PADDLEOCR:
             {
                 GGML_ASSERT(params.image_min_pixels > 0 && params.image_max_pixels > 0);
                 clip_image_u8 resized;
@@ -3333,6 +3344,7 @@ int clip_n_output_tokens_x(const struct clip_ctx * ctx, struct clip_image_f32 *
         case PROJECTOR_TYPE_QWEN25VL:
         case PROJECTOR_TYPE_QWEN3VL:
         case PROJECTOR_TYPE_GLM4V:
+        case PROJECTOR_TYPE_PADDLEOCR:
         case PROJECTOR_TYPE_YOUTUVL:
             return (img->nx / params.patch_size) / 2;
         default:
@@ -3349,6 +3361,7 @@ int clip_n_output_tokens_y(const struct clip_ctx * ctx, struct clip_image_f32 *
         case PROJECTOR_TYPE_QWEN25VL:
         case PROJECTOR_TYPE_QWEN3VL:
         case PROJECTOR_TYPE_GLM4V:
+        case PROJECTOR_TYPE_PADDLEOCR:
         case PROJECTOR_TYPE_YOUTUVL:
             return (img->ny / params.patch_size) / 2;
         default:
@@ -3446,6 +3459,13 @@ int clip_n_output_tokens(const struct clip_ctx * ctx, struct clip_image_f32 * im
                 int y_patch = CLIP_ALIGN(img->ny, out_patch_size) / out_patch_size;
                 n_patches = x_patch * y_patch;
             } break;
+        case PROJECTOR_TYPE_PADDLEOCR:
+            {
+                // dynamic size
+                int n_merge = ctx->model.hparams.n_merge;
+                int stride = n_merge * n_merge;
+                n_patches = CLIP_ALIGN(n_patches, stride) / stride;
+            } break;
         case PROJECTOR_TYPE_PIXTRAL:
         case PROJECTOR_TYPE_LIGHTONOCR:
             {
@@ -3693,6 +3713,30 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
                     }
                 }
 
+                set_input_i32("positions", positions);
+            } break;
+        case PROJECTOR_TYPE_PADDLEOCR:
+            {
+                const int merge_ratio = hparams.n_merge;
+                const int pw = image_size_width  / patch_size;
+                const int ph = image_size_height / patch_size;
+                std::vector<int> positions(n_pos * 4);
+                int ptr = 0;
+                // NOTE: same as Qwen-VL, but x and y are swapped
+                for (int y = 0; y < ph; y += merge_ratio) {
+                    for (int dy = 0; dy < 2; dy++) {
+                        for (int x = 0; x < pw; x += merge_ratio) {
+                            for (int dx = 0; dx < 2; dx++) {
+                                positions[                  ptr] = y + dy;
+                                positions[    num_patches + ptr] = x + dx;
+                                positions[2 * num_patches + ptr] = y + dy;
+                                positions[3 * num_patches + ptr] = x + dx;
+                                ptr++;
+                            }
+                        }
+                    }
+                }
+
                 set_input_i32("positions", positions);
             } break;
         case PROJECTOR_TYPE_QWEN25VL:
@@ -4006,6 +4050,7 @@ int clip_n_mmproj_embd(const struct clip_ctx * ctx) {
             return ctx->model.mm_2_w->ne[1];
         case PROJECTOR_TYPE_LFM2:
         case PROJECTOR_TYPE_KIMIVL:
+        case PROJECTOR_TYPE_PADDLEOCR:
         case PROJECTOR_TYPE_KIMIK25:
             return ctx->model.mm_2_w->ne[1];
         case PROJECTOR_TYPE_COGVLM:

tools/mtmd/models/models.h

@@ -57,6 +57,11 @@ struct clip_graph_kimivl : clip_graph {
     ggml_cgraph * build() override;
 };
 
+struct clip_graph_paddleocr : clip_graph {
+    clip_graph_paddleocr(clip_ctx * ctx, const clip_image_f32 & img) : clip_graph(ctx, img) {}
+    ggml_cgraph * build() override;
+};
+
 struct clip_graph_cogvlm : clip_graph {
     clip_graph_cogvlm(clip_ctx * ctx, const clip_image_f32 & img) : clip_graph(ctx, img) {}
     ggml_cgraph * build() override;

tools/mtmd/models/paddleocr.cpp

@@ -0,0 +1,52 @@
+#include "models.h"
+
+ggml_cgraph * clip_graph_paddleocr::build() {
+    const int n_pos            = n_patches;
+    const int num_position_ids = n_pos * 4; // m-rope requires 4 dim per position
+
+    int mrope_sections[4] = {d_head/4, d_head/4, d_head/4, d_head/4};
+
+    ggml_tensor * positions = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, num_position_ids);
+    ggml_set_name(positions, "positions");
+    ggml_set_input(positions);
+
+    auto add_pos = [&](ggml_tensor * cur, const clip_layer &) {
+        return ggml_rope_multi(
+                    ctx0, cur, positions, nullptr,
+                    d_head/2, mrope_sections, GGML_ROPE_TYPE_VISION,
+                    32768, 10000, 1, 0, 1, 32, 1);
+    };
+
+    ggml_tensor * learned_pos_embd = resize_position_embeddings();
+    ggml_tensor * inp = build_inp();
+    ggml_tensor * cur = build_vit(
+                            inp, n_patches,
+                            NORM_TYPE_NORMAL,
+                            hparams.ffn_op,
+                            learned_pos_embd,
+                            add_pos);
+
+    cb(cur, "vit_out", -1);
+
+    {
+        // mlp_AR paddleocr projector
+        float proj_norm_eps = 1e-5;
+        cur = build_norm(cur,
+                    model.mm_input_norm_w, model.mm_input_norm_b,
+                    NORM_TYPE_NORMAL, proj_norm_eps, -1);
+
+        const int scale_factor = model.hparams.n_merge;
+        cur = build_patch_merge_permute(cur, scale_factor);
+        cur = build_ffn(cur,
+                    model.mm_1_w, model.mm_1_b,
+                    nullptr, nullptr,
+                    model.mm_2_w, model.mm_2_b,
+                    hparams.ffn_op, -1);
+        cb(cur, "mlp_out", -1);
+    }
+
+    // build the graph
+    ggml_build_forward_expand(gf, cur);
+
+    return gf;
+}

tools/mtmd/mtmd.cpp

@@ -175,7 +175,7 @@ struct mtmd_context {
 
         clip_context_params ctx_clip_params {
             /* use_gpu           */ ctx_params.use_gpu,
-            /* flash_attn_type   */ CLIP_FLASH_ATTN_TYPE_AUTO,
+            /* flash_attn_type   */ mtmd_get_clip_flash_attn_type(ctx_params.flash_attn_type),
             /* image_min_tokens  */ ctx_params.image_min_tokens,
             /* image_max_tokens  */ ctx_params.image_max_tokens,
             /* warmup            */ ctx_params.warmup,
@@ -325,6 +325,10 @@ struct mtmd_context {
             img_beg = "<|begin_of_image|>";
             img_end = "<|end_of_image|>";
 
+        } else if (proj == PROJECTOR_TYPE_PADDLEOCR) {
+            // <|IMAGE_START|> ... (image embeddings) ... <|IMAGE_END|>
+            img_beg = "<|IMAGE_START|>";
+            img_end = "<|IMAGE_END|>";
         }
     }
 
@@ -890,6 +894,7 @@ bool mtmd_decode_use_mrope(mtmd_context * ctx) {
         case PROJECTOR_TYPE_QWEN25VL:
         case PROJECTOR_TYPE_QWEN3VL:
         case PROJECTOR_TYPE_GLM4V:
+        case PROJECTOR_TYPE_PADDLEOCR:
             return true;
         default:
             return false;

tools/mtmd/tests.sh

@@ -28,6 +28,14 @@ if [ "${1:-}" = "huge" ]; then
     echo "Include BIG and HUGE models..."
 fi
 
+# Check if the second argument is "flash", then enable flash attention
+# This is useful to test if flash attention off works correctly
+FLASH_ATTN="on"
+if [ "${2:-}" = "flash_off" ] || [ "${1:-}" = "flash_off" ]; then
+    FLASH_ATTN="off"
+    echo "Flash attention disabled..."
+fi
+
 ###############
 
 arr_prefix=()
@@ -143,6 +151,7 @@ for i in "${!arr_hf[@]}"; do
         -hf $(printf %q "$hf") \
         --image $(printf %q "$SCRIPT_DIR/$inp_file") \
         --temp 0 -n 128 \
+        --flash-attn $(printf %q "$FLASH_ATTN") \
         ${extra_args}"
 
     # if extra_args does not contain -p, we add a default prompt

tools/quantize/quantize.cpp

@@ -120,7 +120,7 @@ static bool try_parse_ftype(const std::string & ftype_str_in, llama_ftype & ftyp
 static void usage(const char * executable) {
     printf("usage: %s [--help] [--allow-requantize] [--leave-output-tensor] [--pure] [--imatrix] [--include-weights]\n", executable);
     printf("       [--exclude-weights] [--output-tensor-type] [--token-embedding-type] [--tensor-type] [--tensor-type-file]\n");
-    printf("       [--prune-layers] [--keep-split] [--override-kv]\n");
+    printf("       [--prune-layers] [--keep-split] [--override-kv] [--dry-run]\n");
     printf("       model-f32.gguf [model-quant.gguf] type [nthreads]\n\n");
     printf("  --allow-requantize\n");
     printf("                                      allow requantizing tensors that have already been quantized\n");
@@ -156,7 +156,10 @@ static void usage(const char * executable) {
     printf("                                      generate quantized model in the same shards as input\n");
     printf("  --override-kv KEY=TYPE:VALUE\n");
     printf("                                      override model metadata by key in the quantized model. may be specified multiple times.\n");
-    printf("                                      WARNING: this is an advanced option, use with care.\n\n");
+    printf("                                      WARNING: this is an advanced option, use with care.\n");
+    printf("  --dry-run\n");
+    printf("                                      calculate and show the final quantization size without performing quantization\n");
+    printf("                                      example: llama-quantize --dry-run model-f32.gguf Q4_K\n\n");
     printf("note: --include-weights and --exclude-weights cannot be used together\n\n");
     printf("-----------------------------------------------------------------------------\n");
     printf(" allowed quantization types\n");
@@ -532,6 +535,8 @@ int main(int argc, char ** argv) {
             if (arg_idx == argc-1 || !string_parse_kv_override(argv[++arg_idx], kv_overrides)) {
                 usage(argv[0]);
             }
+        } else if (strcmp(argv[arg_idx], "--dry-run") == 0) {
+            params.dry_run = true;
         } else if (strcmp(argv[arg_idx], "--allow-requantize") == 0) {
             params.allow_requantize = true;
         } else if (strcmp(argv[arg_idx], "--pure") == 0) {
@@ -630,22 +635,26 @@ int main(int argc, char ** argv) {
     std::string ftype_str;
     std::string suffix = ".gguf";
     if (try_parse_ftype(argv[arg_idx], params.ftype, ftype_str)) {
-        std::string fpath;
-        const size_t pos = fname_inp.find_last_of("/\\");
-        if (pos != std::string::npos) {
-            fpath = fname_inp.substr(0, pos + 1);
-        }
+        // argv[arg_idx] is the ftype directly: <input> <ftype>
+        if (!params.dry_run) {
+            std::string fpath;
+            const size_t pos = fname_inp.find_last_of("/\\");
+            if (pos != std::string::npos) {
+                fpath = fname_inp.substr(0, pos + 1);
+            }
 
-        // export as [inp path]/ggml-model-[ftype]. Only add extension if there is no splitting
-        fname_out = fpath + "ggml-model-" + ftype_str;
-        if (!params.keep_split) {
-            fname_out += suffix;
+            // export as [inp path]/ggml-model-[ftype]. Only add extension if there is no splitting
+            fname_out = fpath + "ggml-model-" + ftype_str;
+            if (!params.keep_split) {
+                fname_out += suffix;
+            }
         }
         arg_idx++;
         if (ftype_str == "COPY") {
             params.only_copy = true;
         }
     } else {
+        // argv[arg_idx] is not a valid ftype, so treat it as output path: <input> <output> <ftype>
         fname_out = argv[arg_idx];
         if (params.keep_split && fname_out.find(suffix) != std::string::npos) {
             fname_out = fname_out.substr(0, fname_out.length() - suffix.length());
@@ -677,25 +686,33 @@ int main(int argc, char ** argv) {
         }
     }
 
-    if ((params.ftype == LLAMA_FTYPE_MOSTLY_IQ2_XS || params.ftype == LLAMA_FTYPE_MOSTLY_IQ2_XXS ||
-         params.ftype == LLAMA_FTYPE_MOSTLY_IQ2_S  ||
-         params.ftype == LLAMA_FTYPE_MOSTLY_Q2_K_S ||
-         params.ftype == LLAMA_FTYPE_MOSTLY_IQ1_S  ||
-         params.ftype == LLAMA_FTYPE_MOSTLY_IQ1_M) && imatrix_data.empty()) {
+    if (!params.dry_run &&
+        (
+            params.ftype == LLAMA_FTYPE_MOSTLY_IQ2_XS  || params.ftype == LLAMA_FTYPE_MOSTLY_IQ2_XXS ||
+            params.ftype == LLAMA_FTYPE_MOSTLY_IQ2_S   || params.ftype == LLAMA_FTYPE_MOSTLY_Q2_K_S  ||
+            params.ftype == LLAMA_FTYPE_MOSTLY_IQ1_S   || params.ftype == LLAMA_FTYPE_MOSTLY_IQ1_M
+        ) && imatrix_data.empty()) {
         fprintf(stderr, "\n==========================================================================================================\n");
         fprintf(stderr, "Please do not use IQ1_S, IQ1_M, IQ2_S, IQ2_XXS, IQ2_XS or Q2_K_S quantization without an importance matrix\n");
         fprintf(stderr, "==========================================================================================================\n\n\n");
         return 1;
     }
 
-    if (std::error_code ec; std::filesystem::equivalent(fname_inp, fname_out, ec)) {
-        fprintf(stderr, "%s: error: input and output files are the same: '%s'\n", __func__, fname_inp.c_str());
-        return 1;
+    if (!params.dry_run) {
+        if (std::error_code ec; std::filesystem::equivalent(fname_inp, fname_out, ec)) {
+            fprintf(stderr, "%s: error: input and output files are the same: '%s'\n", __func__, fname_inp.c_str());
+            return 1;
+        }
     }
 
     print_build_info();
 
-    fprintf(stderr, "%s: quantizing '%s' to '%s' as %s", __func__, fname_inp.c_str(), fname_out.c_str(), ftype_str.c_str());
+    if (params.dry_run) {
+        fprintf(stderr, "%s: calculating quantization size for '%s' as %s", __func__, fname_inp.c_str(), ftype_str.c_str());
+    } else {
+        fprintf(stderr, "%s: quantizing '%s' to '%s' as %s", __func__, fname_inp.c_str(), fname_out.c_str(), ftype_str.c_str());
+    }
+
     if (params.nthread > 0) {
         fprintf(stderr, " using %d threads", params.nthread);
     }

tools/server/server-common.cpp

@@ -916,8 +916,7 @@ json oaicompat_chat_params_parse(
                 json image_url = json_value(p, "image_url", json::object());
                 handle_media(out_files, image_url, opt.media_path);
 
-                // replace this chunk with a marker
-                p["type"] = "text";
+                p["type"] = "media_marker";
                 p["text"] = mtmd_default_marker();
                 p.erase("image_url");
 
@@ -938,8 +937,7 @@ json oaicompat_chat_params_parse(
 
                 // TODO: add audio_url support by reusing handle_media()
 
-                // replace this chunk with a marker
-                p["type"] = "text";
+                p["type"] = "media_marker";
                 p["text"] = mtmd_default_marker();
                 p.erase("input_audio");
 

tools/server/webui/src/lib/components/app/models/ModelsSelector.svelte

@@ -42,7 +42,13 @@
 		useGlobalSelection = false
 	}: Props = $props();
 
-	let options = $derived(modelOptions());
+	let options = $derived(
+		modelOptions().filter((option) => {
+			const modelProps = modelsStore.getModelProps(option.model);
+
+			return modelProps?.webui !== false;
+		})
+	);
 	let loading = $derived(modelsLoading());
 	let updating = $derived(modelsUpdating());
 	let activeId = $derived(selectedModelId());
@@ -245,6 +251,9 @@
 			return options.find((option) => option.id === activeId);
 		}
 
+		if (options.length === 1) {
+			return options[0];
+		}
 		// No selection - return undefined to show "Select model"
 		return undefined;
 	}

tools/server/webui/src/lib/stores/chat.svelte.ts

@@ -498,7 +498,8 @@ class ChatStore {
 				MessageRole.USER,
 				content,
 				MessageType.TEXT,
-				parentIdForUserMessage ?? '-1'
+				parentIdForUserMessage ?? '-1',
+				extras
 			);
 			if (isNewConversation && content)
 				await conversationsStore.updateConversationName(currentConv.id, content.trim());