Add support for Laguna XS.2 & M.1 (#25165)

common/chat-auto-parser-generator.cpp

@@ -47,6 +47,8 @@ common_chat_params peg_generator::generate_parser(const common_chat_template &
     data.generation_prompt = common_chat_template_generation_prompt(tmpl, inputs);
     data.format            = COMMON_CHAT_FORMAT_PEG_NATIVE;
     data.preserved_tokens  = autoparser.preserved_tokens;
+    data.additional_stops.insert(data.additional_stops.end(),
+        autoparser.additional_stops.begin(), autoparser.additional_stops.end());
 
     std::string parser_generation_prompt = data.generation_prompt;
 
@@ -286,7 +288,13 @@ common_peg_parser analyze_tools::build_func_parser(common_chat_peg_builder & p,
         // we only emit tool_close when we can actually see the closing marker. This prevents
         // premature closing during partial parsing when we've seen e.g. "</" which could be
         // either "</tool_call>" (end) or "<arg_key>" prefix that failed to match.
-        func_parser = func_parser + p.tool_close(p.peek(p.literal(format.per_call_end)));
+        // Laguna (v4): the model may emit whitespace between the last </arg_value> and
+        // </tool_call> even though the template renders them tight. Tolerate optional
+        // leading space in the close lookahead so the tool call still closes.
+        auto close_peek = arguments.tolerate_intertag_whitespace
+                              ? p.peek(p.space() + p.literal(format.per_call_end))
+                              : p.peek(p.literal(format.per_call_end));
+        func_parser = func_parser + p.tool_close(close_peek);
     } else {
         func_parser = func_parser + p.tool_close(p.space());  // force this to process tool closing callbacks in mapper
     }

common/chat-auto-parser.h

@@ -206,6 +206,7 @@ struct tool_arguments_analysis {
     std::string value_prefix;  // e.g., "", "<arg_value>", ""
     std::string value_suffix;  // e.g., "</param>", "</arg_value>", ""
     std::string separator;     // e.g., "", "\n", ","
+    bool tolerate_intertag_whitespace = false; // Laguna: accept optional whitespace between arg tags
 };
 
 struct tool_id_analysis {
@@ -388,6 +389,7 @@ struct autoparser {
 
     // Preserved tokens for tokenizer (union of all non-empty markers)
     std::vector<std::string> preserved_tokens;
+    std::vector<std::string> additional_stops;  // literal stop strings (e.g. Laguna </assistant>) caught however tokenized
 
     autoparser() = default;
 

common/chat-diff-analyzer.cpp

@@ -173,6 +173,26 @@ static std::vector<std::function<void(const common_chat_template & tmpl, autopar
               LOG_DBG(ANSI_ORANGE "[Patch: JSON name/parameters tool instruction]\n" ANSI_RESET);
           }
       },
+      // Laguna (poolside) - the v4 chat template renders reasoning and tool-arg
+      // delimiters with formatting whitespace ("<think>\n", "</arg_value>\n") that
+      // the model does not emit, so the inferred delimiters carry a spurious
+      // newline and never match the model output. Trim to the bare tag. (v8
+      // renders without the whitespace, so this is a no-op there.)
+      [](const common_chat_template & tmpl, autoparser & analysis) -> void {
+          if (tmpl.src.find("laguna_glm_thinking") != std::string::npos) {
+              analysis.reasoning.start              = trim_whitespace(analysis.reasoning.start);
+              analysis.reasoning.end                = trim_whitespace(analysis.reasoning.end);
+              analysis.tools.arguments.value_prefix = trim_whitespace(analysis.tools.arguments.value_prefix);
+              analysis.tools.arguments.value_suffix = trim_whitespace(analysis.tools.arguments.value_suffix);
+              analysis.tools.arguments.separator    = trim_whitespace(analysis.tools.arguments.separator);
+              analysis.tools.arguments.tolerate_intertag_whitespace = true;
+              // The CONTROL/eot </assistant> token only halts generation when emitted as the
+              // single token; after tool calls the model can spell it out as text tokens.
+              // A literal stop string catches it either way.
+              analysis.additional_stops.push_back("</assistant>");
+              LOG_DBG(ANSI_ORANGE "[Patch: Laguna]\n" ANSI_RESET);
+          }
+      },
 
     });
 

conversion/__init__.py

@@ -18,6 +18,7 @@ __all__ = [
 
 TEXT_MODEL_MAP: dict[str, str] = {
     "AfmoeForCausalLM": "afmoe",
+    "LagunaForCausalLM": "laguna",
     "ApertusForCausalLM": "llama",
     "ArceeForCausalLM": "llama",
     "ArcticForCausalLM": "arctic",

conversion/base.py

@@ -1682,6 +1682,9 @@ class TextModel(ModelBase):
         if chkhsh == "9dcf830ee9990cdbf78cc523a5f7bd9ad8f3f9890c2d3581d2785ad10f07049d":
             # ref: https://huggingface.co/JetBrains/Mellum2-12B-A2.5B-Base
             res = "mellum2"
+        if chkhsh == "972da7b59cec44d1f0a490a86c96df53859e486e481563e5dddac155013d87ac":
+            # ref: https://huggingface.co/poolside/Laguna-XS.2
+            res = "laguna"
 
         if res is None:
             logger.warning("\n")

conversion/laguna.py

@@ -0,0 +1,207 @@
+from __future__ import annotations
+
+import re
+from collections.abc import Iterable
+from typing import TYPE_CHECKING
+
+import torch
+
+if TYPE_CHECKING:
+    from torch import Tensor
+
+from .base import ModelBase, TextModel, gguf, logger
+
+
+@ModelBase.register("LagunaForCausalLM")
+class LagunaModel(TextModel):
+    model_arch = gguf.MODEL_ARCH.LAGUNA
+    _experts: list[dict] | None = None
+    _gate_types: list[str] | None = None
+
+    # --- vocab ---------------------------------------------------------------
+
+    def set_vocab(self) -> None:
+        self._set_vocab_gpt2()
+
+        # Some Laguna releases wrap the chat template in tokenizer_config.json as
+        # "{% include 'chat_template.jinja' %}", which SpecialVocab embeds verbatim
+        # and llama.cpp's jinja engine cannot process. Prefer the resolved template
+        # from the chat_template.jinja file so the GGUF is self-contained.
+        tmpl_file = self.dir_model / "chat_template.jinja"
+        if tmpl_file.is_file():
+            self.gguf_writer.add_chat_template(tmpl_file.read_text(encoding="utf-8"))
+            logger.info("gguf: embedded resolved chat_template.jinja (overriding include directive)")
+
+        # eos_token_id is a list [2, 24]: token 2 (EOS, also BOS) and token 24
+        # (</assistant>, the turn-end). _set_vocab_gpt2 only records the scalar
+        # eos, so register the extra id as eot; llama.cpp folds eot into its EOG
+        # set, so the model halts on </assistant> natively.
+        eos_ids = self.hparams.get("eos_token_id")
+        if isinstance(eos_ids, list):
+            bos_id = self.hparams.get("bos_token_id")
+            extra = [e for e in eos_ids if e != bos_id]
+            if extra:
+                self.gguf_writer.add_eot_token_id(extra[0])
+                logger.info(f"gguf: registered eot_token_id={extra[0]} from eos list {eos_ids}")
+
+    def get_vocab_base(self) -> tuple[list[str], list[int], str]:
+        # </assistant> is the assistant turn-end (registered as eot below). The
+        # HF tokenizer flags it special=false, so the base classifies it as
+        # USER_DEFINED and llama.cpp renders its text into generated content,
+        # leaking "</assistant>" and breaking response parsing. It is a control
+        # marker, so promote it to CONTROL: llama.cpp then treats it as
+        # end-of-generation and suppresses its text.
+        tokens, toktypes, tokpre = super().get_vocab_base()
+        for i, tok in enumerate(tokens):
+            if tok == "</assistant>":
+                toktypes[i] = gguf.TokenType.CONTROL
+                logger.info(f"gguf: marked </assistant> (id {i}) as CONTROL token")
+        return tokens, toktypes, tokpre
+
+    # --- hparams -------------------------------------------------------------
+
+    def set_gguf_parameters(self) -> None:
+        super().set_gguf_parameters()
+        hparams = self.hparams
+
+        # super() does not emit vocab_size for the gpt2 vocab path; head_count is
+        # overridden with a per-layer array (XS.2 varies heads per layer via
+        # num_attention_heads_per_layer; M.1 is uniform and omits it).
+        self.gguf_writer.add_vocab_size(hparams["vocab_size"])
+
+        per_layer_heads = hparams.get("num_attention_heads_per_layer")
+        if not per_layer_heads:
+            per_layer_heads = [hparams["num_attention_heads"]] * hparams["num_hidden_layers"]
+        assert len(per_layer_heads) == hparams["num_hidden_layers"], (
+            f"num_attention_heads_per_layer length {len(per_layer_heads)} != "
+            f"num_hidden_layers {hparams['num_hidden_layers']}"
+        )
+        self.gguf_writer.add_head_count(per_layer_heads)
+
+        # Resolve + validate the attention gate type now so an inconsistent
+        # `gating` field fails at conversion time. See _attn_gate_types.
+        self._attn_gate_types()
+
+        # SWA window size (M.1 has none -> key omitted, swa_type stays NONE).
+        sliding_window = hparams.get("sliding_window") or 0
+        if sliding_window > 0:
+            self.gguf_writer.add_sliding_window(sliding_window)
+
+        # MoE (expert_count / expert_used_count come from super().set_gguf_parameters())
+        self.gguf_writer.add_expert_feed_forward_length(hparams["moe_intermediate_size"])
+        self.gguf_writer.add_expert_shared_feed_forward_length(hparams["shared_expert_intermediate_size"])
+        self.gguf_writer.add_expert_weights_norm(True)  # HF reference always sum-normalises after top-k
+        self.gguf_writer.add_expert_weights_scale(float(hparams["moe_routed_scaling_factor"]))
+        self.gguf_writer.add_expert_gating_func(gguf.ExpertGatingFuncType.SIGMOID)
+
+        # Leading dense layers (XS.2 has 1, M.1 has 3) before the MoE layers.
+        mlp_layer_types: list[str] = hparams["mlp_layer_types"]
+        leading_dense = 0
+        for t in mlp_layer_types:
+            if t == "dense":
+                leading_dense += 1
+            else:
+                break
+        self.gguf_writer.add_leading_dense_block_count(leading_dense)
+
+        # Per-layer-type RoPE dimension count (partial rotary). base emits
+        # rope_freq_base(_swa) and the YaRN params from self.rope_parameters.
+        head_dim = hparams["head_dim"]
+        full_rope = self.rope_parameters["full_attention"]
+        self.gguf_writer.add_rope_dimension_count(
+            int(head_dim * float(full_rope.get("partial_rotary_factor", 1.0))))
+        swa_rope = self.rope_parameters.get("sliding_attention")
+        if swa_rope is not None:
+            self.gguf_writer.add_rope_dimension_count_swa(
+                int(head_dim * float(swa_rope.get("partial_rotary_factor", 1.0))))
+
+    def _attn_gate_types(self) -> list[str]:
+        """Per-layer attention output gate type: "per_head" or "per_element".
+
+        `gating_types` (per layer) is authoritative when present; otherwise the
+        scalar `gating` field is used (the "per-element"/"per-head" string, or
+        the legacy boolean True == per-head, as in Laguna-XS.2).
+
+        Fails loudly when the model is per-element but the `gating` field does
+        not declare that as a string: runtimes that key off `gating` (vLLM,
+        transformers) ignore gating_types and read a bare boolean True as
+        per-head, silently corrupting the model. Surfacing it here keeps a
+        broken checkpoint from being packaged as if it were fine.
+        """
+        if self._gate_types is not None:
+            return self._gate_types
+        hparams = self.hparams
+        n_layer = hparams["num_hidden_layers"]
+        gating = hparams.get("gating")
+        gating_types = hparams.get("gating_types")
+
+        def _norm(t: object) -> str:
+            sval = str(t).replace("-", "_")
+            if sval in ("per_element", "per_head"):
+                return sval
+            raise ValueError(f"Laguna: unrecognised attention gate type {t!r}")
+
+        if gating_types:
+            assert len(gating_types) == n_layer, (
+                f"gating_types length {len(gating_types)} != num_hidden_layers {n_layer}")
+            types = [_norm(t) for t in gating_types]
+        elif isinstance(gating, str):
+            types = [_norm(gating)] * n_layer
+        elif gating is True:
+            types = ["per_head"] * n_layer
+        else:
+            raise ValueError(
+                f"Laguna: cannot determine attention gate type "
+                f"(gating={gating!r}, gating_types={gating_types!r})")
+
+        if any(t == "per_element" for t in types) and not (
+                isinstance(gating, str) and _norm(gating) == "per_element"):
+            raise ValueError(
+                f"Laguna config declares a per-element attention gate but "
+                f"`gating`={gating!r} is not the string \"per-element\". Runtimes that "
+                f"read `gating` (vLLM, transformers) will mis-handle this checkpoint as "
+                f"per-head. Set gating=\"per-element\" in the source config.")
+
+        self._gate_types = types
+        return types
+
+    # --- tensor handling -----------------------------------------------------
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        # Per-expert MoE weights: model.layers.{bid}.mlp.experts.{xid}.{w}.weight.
+        # Only the NUMBERED per-expert weights are stacked; the router bias
+        # (mlp.experts.e_score_correction_bias) takes the normal mapping path.
+        if re.search(r"mlp\.experts\.\d+\.", name):
+            n_experts = self.find_hparam(["num_local_experts", "num_experts"])
+            assert bid is not None
+            if self._experts is None:
+                self._experts = [{} for _ in range(self.block_count)]
+            self._experts[bid][name] = data_torch
+            needed = [f"model.layers.{bid}.mlp.experts.{x}.{w}.weight"
+                      for x in range(n_experts) for w in ("gate_proj", "up_proj", "down_proj")]
+            if all(e in self._experts[bid] for e in needed):
+                for w_name in ["gate_proj", "up_proj", "down_proj"]:
+                    datas = [self._experts[bid][f"model.layers.{bid}.mlp.experts.{x}.{w_name}.weight"]
+                             for x in range(n_experts)]
+                    stacked = torch.stack(datas, dim=0)
+                    merged = f"model.layers.{bid}.mlp.experts.{w_name}.weight"
+                    yield from TextModel.modify_tensors(self, stacked, merged, bid)
+                self._experts[bid].clear()
+                return
+            return
+        # Cross-check the gate projection width against the declared gate type;
+        # a mismatch means the weights and config disagree -> fail, do not guess.
+        if bid is not None and name.endswith("self_attn.g_proj.weight"):
+            heads = (self.hparams.get("num_attention_heads_per_layer")
+                     or [self.hparams["num_attention_heads"]] * self.hparams["num_hidden_layers"])
+            n_head = heads[bid]
+            head_dim = self.hparams["head_dim"]
+            gate_type = self._attn_gate_types()[bid]
+            expected = n_head * head_dim if gate_type == "per_element" else n_head
+            out_features = int(data_torch.shape[0])
+            if out_features != expected:
+                raise ValueError(
+                    f"Laguna layer {bid}: g_proj output width {out_features} contradicts the "
+                    f"declared {gate_type} gate (expected {expected}); weights and config disagree.")
+
+        yield from TextModel.modify_tensors(self, data_torch, name, bid)

convert_hf_to_gguf_update.py

@@ -162,6 +162,7 @@ models = [
     {"name": "granite-embed-multi-97m", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/ibm-granite/granite-embedding-97m-multilingual-r2", },
     {"name": "granite-embed-multi-311m", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/ibm-granite/granite-embedding-311m-multilingual-r2", },
     {"name": "mellum2",          "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/JetBrains/Mellum2-12B-A2.5B-Base"},
+    {"name": "laguna",           "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/poolside/Laguna-XS.2", },
 ]
 
 # some models are known to be broken upstream, so we will skip them as exceptions

gguf-py/gguf/constants.py

@@ -507,6 +507,7 @@ class MODEL_ARCH(IntEnum):
     DOTS1            = auto()
     ARCEE            = auto()
     AFMOE            = auto()
+    LAGUNA           = auto()
     ERNIE4_5         = auto()
     ERNIE4_5_MOE     = auto()
     HUNYUAN_MOE      = auto()
@@ -1088,6 +1089,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
     MODEL_ARCH.DOTS1:            "dots1",
     MODEL_ARCH.ARCEE:            "arcee",
     MODEL_ARCH.AFMOE:            "afmoe",
+    MODEL_ARCH.LAGUNA:           "laguna",
     MODEL_ARCH.ERNIE4_5:         "ernie4_5",
     MODEL_ARCH.ERNIE4_5_MOE:     "ernie4_5-moe",
     MODEL_ARCH.FALCON_H1:        "falcon-h1",
@@ -3823,6 +3825,31 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.FFN_POST_NORM,
         MODEL_TENSOR.FFN_EXP_PROBS_B,
     ],
+    MODEL_ARCH.LAGUNA: [
+        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.ATTN_Q_NORM,
+        MODEL_TENSOR.ATTN_K_NORM,
+        MODEL_TENSOR.ATTN_GATE,
+        MODEL_TENSOR.FFN_NORM,
+        MODEL_TENSOR.FFN_GATE,
+        MODEL_TENSOR.FFN_DOWN,
+        MODEL_TENSOR.FFN_UP,
+        MODEL_TENSOR.FFN_GATE_INP,
+        MODEL_TENSOR.FFN_EXP_PROBS_B,
+        MODEL_TENSOR.FFN_GATE_EXP,
+        MODEL_TENSOR.FFN_DOWN_EXP,
+        MODEL_TENSOR.FFN_UP_EXP,
+        MODEL_TENSOR.FFN_GATE_SHEXP,
+        MODEL_TENSOR.FFN_UP_SHEXP,
+        MODEL_TENSOR.FFN_DOWN_SHEXP,
+    ],
     MODEL_ARCH.ERNIE4_5: [
         MODEL_TENSOR.TOKEN_EMBD,
         MODEL_TENSOR.OUTPUT_NORM,

gguf-py/gguf/tensor_mapping.py

@@ -479,6 +479,7 @@ class TensorNameMap:
             "model.layers.{bid}.mlp.e_score_correction",                    # exaone-moe
             "model.layers.{bid}.block_sparse_moe.gate.e_score_correction",  # kimi
             "model.layers.{bid}.moe.router_bias",                           # step3.5 expert selection bias
+            "model.layers.{bid}.mlp.experts.e_score_correction",       # laguna
         ),
 
         # Feed-forward up

models/templates/poolside-Laguna-S-2.1.jinja

@@ -0,0 +1,93 @@
+{#- Iteration on laguna_glm_thinking_v8/chat_template.jinja -#}
+{#- No formatting instructions -#}
+{{- "〈|EOS|〉" -}}
+{%- set enable_thinking = enable_thinking | default(false) -%}
+{%- set add_generation_prompt = add_generation_prompt | default(false) -%}
+
+{#- ───── header (system message) ───── -#}
+{#- A caller-supplied system message with empty content opts out of the default below, producing no <system> block — used to train without a system message. -#}
+{%- set system_message = "You are a helpful, conversationally-fluent assistant made by Poolside. You are here to be helpful to users through natural language conversations." -%}
+{%- if messages and messages[0].role == "system" -%}
+  {%- set system_message = messages[0].content -%}
+  {%- set messages = messages[1:] -%}
+{%- endif -%}
+
+{%- set has_sys = system_message and system_message.strip() -%}
+{%- if has_sys or tools or enable_thinking -%}
+  {{- "<system>" -}}
+
+  {%- if has_sys -%}
+    {{- system_message.rstrip() -}}
+    {%- if tools -%}{{- "\n\n" -}}{%- endif -%}
+  {%- endif -%}
+
+  {%- if tools -%}
+    {{- "### Tools\n\n" -}}
+    {{- "You may call functions to assist with the user query.\n" -}}
+    {{- "All available function signatures are listed below:\n" -}}
+    {{- "<available_tools>\n" -}}
+    {%- for tool in tools -%}
+      {{- (tool | tojson) ~ "\n" -}}
+    {%- endfor -%}
+    {{- "</available_tools>" -}}
+  {%- endif -%}
+
+  {{- "</system>\n" -}}
+{%- endif -%}
+
+{#- ───── main loop ───── -#}
+{%- for message in messages -%}
+  {%- set content = message.content if message.content is string else "" -%}
+  {%- if message.role == "user" -%}
+    {{- "<user>" + content + "</user>\n" -}}
+  {%- elif message.role == "assistant" -%}
+    {%- generation -%}
+      {{- "<assistant>" -}}
+      {#- Extract reasoning content from message.reasoning (vLLM field name) or message.reasoning_content -#}
+      {%- set reasoning_content = '' -%}
+      {%- if message.reasoning is string -%}
+        {%- set reasoning_content = message.reasoning -%}
+      {%- elif message.reasoning_content is string -%}
+        {%- set reasoning_content = message.reasoning_content -%}
+      {%- endif -%}
+      {#- Display reasoning content for all messages if enable_thinking -#}
+      {%- if enable_thinking -%}
+        {{- '<think>' + reasoning_content + '</think>' -}}
+      {%- else -%}
+        {{- '</think>' -}}
+      {%- endif -%}
+      {#- Display main content (trailing newline only when no tool_calls follow) -#}
+      {%- if content -%}
+        {{- content -}}
+      {%- endif -%}
+      {%- if message.tool_calls -%}
+        {%- for tool_call in message.tool_calls -%}
+          {%- set function_data = tool_call.function -%}
+          {{- '<tool_call>' + function_data.name -}}
+          {%- set _args = function_data.arguments -%}
+          {%- for k, v in _args.items() -%}
+            {{- "<arg_key>" ~ k ~ "</arg_key>" -}}
+            {{- "<arg_value>" -}}{{- v | tojson(ensure_ascii=False) if v is not string else v -}}{{- "</arg_value>" -}}
+          {%- endfor -%}
+          {{- "</tool_call>" -}}
+        {%- endfor -%}
+      {%- endif -%}
+      {{- "</assistant>\n" -}}
+    {%- endgeneration -%}
+  {%- elif message.role == "tool" -%}
+    {{- "<tool_response>" + content + "</tool_response>\n" -}}
+  {%- elif message.role == "system" -%}
+    {#- Render additional system messages (the first one, if any, is handled separately in the header and was sliced off above) -#}
+    {{- "<system>" + content + "</system>\n" -}}
+  {%- endif -%}
+{%- endfor -%}
+{#- ───── generation prompt ───── -#}
+{%- if add_generation_prompt -%}
+  {{- "<assistant>" -}}
+  {#- ───── Include reasoning mode directive ───── -#}
+  {%- if enable_thinking -%}
+    {{- '<think>' -}}
+  {%- else -%}
+    {{- '</think>' -}}
+  {%- endif -%}
+{%- endif -%}

models/templates/poolside-Laguna-XS-2.1.jinja

@@ -0,0 +1,132 @@
+{#- Copied from laguna_glm_thinking_v4/chat_template.jinja -#}
+{#- Removes prefix that references <think> token, and replaces message.reasoning_content reference with message.reasoning -#}
+{{- "〈|EOS|〉" -}}
+{%- set enable_thinking = enable_thinking | default(false) -%}
+{%- set render_assistant_messages_raw = render_assistant_messages_raw | default(false) -%}
+{%- set add_generation_prompt = add_generation_prompt | default(false) -%}
+
+{#- ───── header (system message) ───── -#}
+{%- set system_message = "" -%}
+{%- if messages and messages[0].role == "system" -%}
+  {%- set system_message = messages[0].content -%}
+{%- endif -%}
+
+{%- if (system_message and system_message.strip()) or tools -%}
+  {{- "<system>\n" -}}
+
+  {%- if system_message and system_message.strip() -%}
+    {{- "\n" -}}
+    {{- system_message.rstrip() -}}
+  {%- endif -%}
+
+  {%- if tools -%}
+    {{- "\n\n### Tools\n\n" -}}
+    {%- set ns = namespace(tool_string="You may call functions to assist with the user query.\n"
+          ~ "All available function signatures are listed below:\n"
+          ~ "<available_tools>\n") -%}
+    {%- for tool in tools -%}
+      {%- set ns.tool_string = ns.tool_string ~ (tool | tojson) ~ "\n" -%}
+    {%- endfor -%}
+    {%- if enable_thinking -%}
+      {%- set tool_string = ns.tool_string + "</available_tools>\n\n" ~
+            "Wrap your thinking in '<think>', '</think>' tags, followed by a function call. For each function call, return an unescaped XML-like object with function name and arguments within '<tool_call>' and '</tool_call>' tags, like here:\n" ~
+            "<think> your thoughts here </think>\n" ~
+            "<tool_call>function-name\n<arg_key>argument-key</arg_key>\n<arg_value>value-of-argument-key</arg_value>\n" ~
+            "</tool_call>" -%}
+    {%- else -%}
+      {%- set tool_string = ns.tool_string + "</available_tools>\n\n" ~
+            "For each function call, return an unescaped XML-like object " ~
+            "with function name and arguments within '<tool_call>' and '</tool_call>' tags, like here:\n" ~
+            "<tool_call>function-name\n<arg_key>argument-key</arg_key>\n<arg_value>value-of-argument-key</arg_value>\n" ~
+            "</tool_call>" -%}
+    {%- endif -%}
+    {{- tool_string -}}
+  {%- endif -%}
+
+  {{- "\n</system>\n" -}}
+{%- endif -%}
+
+{#- ───── main loop ───── -#}
+{%- for message in messages -%}
+  {%- set content = message.content if message.content is string else "" -%}
+  {%- if message.role == "user" -%}
+    {{- "<user>\n" + content + "\n</user>\n" -}}
+  {%- elif message.role == "assistant" -%}
+    {%- generation -%}
+      {{- "<assistant>\n" -}}
+      {%- if render_assistant_messages_raw -%}
+        {#- Raw mode: prepend the generation prompt token, then dump content verbatim. -#}
+        {#- The generation prompt is <think> when enable_thinking, </think> otherwise. -#}
+        {#- Only prepend if content doesn't already start with it. -#}
+        {%- if enable_thinking -%}
+          {%- if not content.startswith('<think>') -%}
+            {{- '<think>' -}}
+          {%- endif -%}
+        {%- else -%}
+          {%- if not content.startswith('</think>') -%}
+            {{- '</think>' -}}
+          {%- endif -%}
+        {%- endif -%}
+        {{- content -}}
+        {#- Append closing tag if content doesn't already end with it. -#}
+        {%- if not content.endswith('</assistant>\n') and not content.endswith('</assistant>') -%}
+          {{- '\n</assistant>' -}}
+        {%- endif -%}
+        {{- "\n" -}}
+      {%- else -%}
+        {#- Extract reasoning content from message.reasoning (vLLM field name) or message.reasoning_content, or from <think> tags -#}
+        {%- set reasoning_content = '' %}
+        {%- if message.reasoning is string %}
+          {%- set reasoning_content = message.reasoning %}
+        {%- elif message.reasoning_content is string %}
+          {%- set reasoning_content = message.reasoning_content %}
+        {%- endif %}
+        {#- Always strip <think> tags from content if present to avoid duplication -#}
+        {%- if '</think>' in content %}
+          {%- if not reasoning_content %}
+            {%- set reasoning_content = content.split('</think>')[0].rstrip('\n').split('<think>')[-1].lstrip('\n') %}
+          {%- endif %}
+          {%- set content = content.split('</think>')[-1].lstrip('\n') %}
+        {%- endif %}
+        {#- Display reasoning content for all messages -#}
+        {%- if reasoning_content -%}
+          {{- '<think>\n' + reasoning_content.strip() + '\n</think>\n' -}}
+        {%- else -%}
+          {{- '</think>\n' -}}
+        {%- endif -%}
+        {#- Display main content -#}
+        {%- if content.strip() -%}
+          {{- content.strip() ~ "\n" -}}
+        {%- endif -%}
+        {%- if message.tool_calls -%}
+          {%- for tool_call in message.tool_calls -%}
+            {%- set function_data = tool_call.function -%}
+            {{- '<tool_call>' + function_data.name }}
+            {% set _args = function_data.arguments %}
+            {%- for k, v in _args.items() -%}
+              {{- "<arg_key>" ~ k ~ "</arg_key>\n" -}}
+              {{- "<arg_value>"}}{{ v | tojson(ensure_ascii=False) if v is not string else v }}{{ "</arg_value>\n" -}}
+            {%- endfor -%}
+            {{- "</tool_call>\n" -}}
+          {%- endfor -%}
+        {%- endif -%}
+        {{- "</assistant>\n" -}}
+      {%- endif -%}
+    {%- endgeneration -%}
+  {%- elif message.role == "tool" -%}
+    {{- "<tool_response>\n" + content + "\n</tool_response>\n" -}}
+  {%- elif message.role == "system" and loop.index0 != 0 -%}
+    {#- Render additional system messages (skip the first one which is handled separately in the header) -#}
+    {{- "<system>\n" + content + "\n</system>\n" -}}
+  {%- endif -%}
+{%- endfor -%}
+{#- ───── generation prompt ───── -#}
+{%- if add_generation_prompt -%}
+  {{- "<assistant>\n" -}}
+  {#- ───── Include reasoning mode directive ───── -#}
+  {%- if not enable_thinking %}
+    {{- '</think>' -}}
+  {%- else %}
+    {{- '<think>' -}}
+  {%- endif %}
+{%- endif -%}

models/templates/poolside-Laguna-XS.2.jinja

@@ -0,0 +1,132 @@
+{#- Iteration on laguna_glm_thinking_v5/chat_template.jinja -#}
+{#- Adds a default system message (used when no system message is provided in `messages`). -#}
+{{- "〈|EOS|〉" -}}
+{%- set enable_thinking = enable_thinking | default(false) -%}
+{%- set render_assistant_messages_raw = render_assistant_messages_raw | default(false) -%}
+{%- set add_generation_prompt = add_generation_prompt | default(false) -%}
+
+{#- ───── header (system message) ───── -#}
+{%- set system_message = "You are a helpful, conversationally-fluent assistant made by Poolside. You are here to be helpful to users through natural language conversations." -%}
+{%- if messages and messages[0].role == "system" -%}
+  {%- set system_message = messages[0].content -%}
+{%- endif -%}
+
+{%- if (system_message and system_message.strip()) or tools -%}
+  {{- "<system>\n" -}}
+
+  {%- if system_message and system_message.strip() -%}
+    {{- "\n" -}}
+    {{- system_message.rstrip() -}}
+  {%- endif -%}
+
+  {%- if tools -%}
+    {{- "\n\n### Tools\n\n" -}}
+    {%- set ns = namespace(tool_string="You may call functions to assist with the user query.\n"
+          ~ "All available function signatures are listed below:\n"
+          ~ "<available_tools>\n") -%}
+    {%- for tool in tools -%}
+      {%- set ns.tool_string = ns.tool_string ~ (tool | tojson) ~ "\n" -%}
+    {%- endfor -%}
+    {%- if enable_thinking -%}
+      {%- set tool_string = ns.tool_string + "</available_tools>\n\n" ~
+            "Wrap your thinking in '<think>', '</think>' tags, followed by a function call. For each function call, return an unescaped XML-like object with function name and arguments within '<tool_call>' and '</tool_call>' tags, like here:\n" ~
+            "<think> your thoughts here </think>\n" ~
+            "<tool_call>function-name\n<arg_key>argument-key</arg_key>\n<arg_value>value-of-argument-key</arg_value>\n" ~
+            "</tool_call>" -%}
+    {%- else -%}
+      {%- set tool_string = ns.tool_string + "</available_tools>\n\n" ~
+            "For each function call, return an unescaped XML-like object " ~
+            "with function name and arguments within '<tool_call>' and '</tool_call>' tags, like here:\n" ~
+            "<tool_call>function-name\n<arg_key>argument-key</arg_key>\n<arg_value>value-of-argument-key</arg_value>\n" ~
+            "</tool_call>" -%}
+    {%- endif -%}
+    {{- tool_string -}}
+  {%- endif -%}
+
+  {{- "\n</system>\n" -}}
+{%- endif -%}
+
+{#- ───── main loop ───── -#}
+{%- for message in messages -%}
+  {%- set content = message.content if message.content is string else "" -%}
+  {%- if message.role == "user" -%}
+    {{- "<user>\n" + content + "\n</user>\n" -}}
+  {%- elif message.role == "assistant" -%}
+    {%- generation -%}
+      {{- "<assistant>\n" -}}
+      {%- if render_assistant_messages_raw -%}
+        {#- Raw mode: prepend the generation prompt token, then dump content verbatim. -#}
+        {#- The generation prompt is <think> when enable_thinking, </think> otherwise. -#}
+        {#- Only prepend if content doesn't already start with it. -#}
+        {%- if enable_thinking -%}
+          {%- if not content.startswith('<think>') -%}
+            {{- '<think>' -}}
+          {%- endif -%}
+        {%- else -%}
+          {%- if not content.startswith('</think>') -%}
+            {{- '</think>' -}}
+          {%- endif -%}
+        {%- endif -%}
+        {{- content -}}
+        {#- Append closing tag if content doesn't already end with it. -#}
+        {%- if not content.endswith('</assistant>\n') and not content.endswith('</assistant>') -%}
+          {{- '\n</assistant>' -}}
+        {%- endif -%}
+        {{- "\n" -}}
+      {%- else -%}
+        {#- Extract reasoning content from message.reasoning (vLLM field name) or message.reasoning_content, or from <think> tags -#}
+        {%- set reasoning_content = '' %}
+        {%- if message.reasoning is string %}
+          {%- set reasoning_content = message.reasoning %}
+        {%- elif message.reasoning_content is string %}
+          {%- set reasoning_content = message.reasoning_content %}
+        {%- endif %}
+        {#- Always strip <think> tags from content if present to avoid duplication -#}
+        {%- if '</think>' in content %}
+          {%- if not reasoning_content %}
+            {%- set reasoning_content = content.split('</think>')[0].rstrip('\n').split('<think>')[-1].lstrip('\n') %}
+          {%- endif %}
+          {%- set content = content.split('</think>')[-1].lstrip('\n') %}
+        {%- endif %}
+        {#- Display reasoning content for all messages -#}
+        {%- if reasoning_content -%}
+          {{- '<think>\n' + reasoning_content.strip() + '\n</think>\n' -}}
+        {%- else -%}
+          {{- '</think>\n' -}}
+        {%- endif -%}
+        {#- Display main content -#}
+        {%- if content.strip() -%}
+          {{- content.strip() ~ "\n" -}}
+        {%- endif -%}
+        {%- if message.tool_calls -%}
+          {%- for tool_call in message.tool_calls -%}
+            {%- set function_data = tool_call.function -%}
+            {{- '<tool_call>' + function_data.name }}
+            {% set _args = function_data.arguments %}
+            {%- for k, v in _args.items() -%}
+              {{- "<arg_key>" ~ k ~ "</arg_key>\n" -}}
+              {{- "<arg_value>"}}{{ v | tojson(ensure_ascii=False) if v is not string else v }}{{ "</arg_value>\n" -}}
+            {%- endfor -%}
+            {{- "</tool_call>\n" -}}
+          {%- endfor -%}
+        {%- endif -%}
+        {{- "</assistant>\n" -}}
+      {%- endif -%}
+    {%- endgeneration -%}
+  {%- elif message.role == "tool" -%}
+    {{- "<tool_response>\n" + content + "\n</tool_response>\n" -}}
+  {%- elif message.role == "system" and loop.index0 != 0 -%}
+    {#- Render additional system messages (skip the first one which is handled separately in the header) -#}
+    {{- "<system>\n" + content + "\n</system>\n" -}}
+  {%- endif -%}
+{%- endfor -%}
+{#- ───── generation prompt ───── -#}
+{%- if add_generation_prompt -%}
+  {{- "<assistant>\n" -}}
+  {#- ───── Include reasoning mode directive ───── -#}
+  {%- if not enable_thinking %}
+    {{- '</think>' -}}
+  {%- else %}
+    {{- '<think>' -}}
+  {%- endif %}
+{%- endif -%}

src/llama-arch.cpp

@@ -108,6 +108,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_DOTS1,            "dots1"            },
     { LLM_ARCH_ARCEE,            "arcee"            },
     { LLM_ARCH_AFMOE,            "afmoe"            },
+    { LLM_ARCH_LAGUNA,           "laguna"           },
     { LLM_ARCH_ERNIE4_5,         "ernie4_5"         },
     { LLM_ARCH_ERNIE4_5_MOE,     "ernie4_5-moe"     },
     { LLM_ARCH_HUNYUAN_MOE,      "hunyuan-moe"      },

src/llama-arch.h

@@ -113,6 +113,7 @@ enum llm_arch {
     LLM_ARCH_DOTS1,
     LLM_ARCH_ARCEE,
     LLM_ARCH_AFMOE,
+    LLM_ARCH_LAGUNA,
     LLM_ARCH_ERNIE4_5,
     LLM_ARCH_ERNIE4_5_MOE,
     LLM_ARCH_HUNYUAN_MOE,

src/llama-model-saver.cpp

@@ -28,6 +28,7 @@ bool llama_model_saver_supports_arch(llm_arch arch) {
         case LLM_ARCH_MIMO2:
         case LLM_ARCH_STEP35:
         case LLM_ARCH_MELLUM:
+        case LLM_ARCH_LAGUNA:
             return false;
         default:
             return true;

src/llama-model.cpp

@@ -250,6 +250,8 @@ static llama_model * llama_model_mapping(llm_arch arch, const llama_model_params
             return new llama_model_arcee(params);
         case LLM_ARCH_AFMOE:
             return new llama_model_afmoe(params);
+        case LLM_ARCH_LAGUNA:
+            return new llama_model_laguna(params);
         case LLM_ARCH_ERNIE4_5:
             return new llama_model_ernie4_5(params);
         case LLM_ARCH_ERNIE4_5_MOE:
@@ -2549,6 +2551,7 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
         case LLM_ARCH_COGVLM:
         case LLM_ARCH_PANGU_EMBED:
         case LLM_ARCH_AFMOE:
+        case LLM_ARCH_LAGUNA:
         case LLM_ARCH_QWEN3NEXT:
         case LLM_ARCH_MIMO2:
         case LLM_ARCH_STEP35:

src/llama-vocab.cpp

@@ -496,6 +496,12 @@ struct llm_tokenizer_bpe : llm_tokenizer {
                     "[!\"#$%&'()*+,\\-./:;<=>?@\\[\\\\\\]^_`{|}~][A-Za-z]+|[^\\r\\n\\p{L}\\p{P}\\p{S}]?[\\p{L}\\p{M}]+| ?[\\p{P}\\p{S}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+",
                 };
                 break;
+            case LLAMA_VOCAB_PRE_TYPE_LAGUNA:
+                regex_exprs = {
+                    "[^\\n]+|[\\n]+",
+                    "(?:'[sS]|'[tT]|'[rR][eE]|'[vV][eE]|'[mM]|'[lL][lL]|'[dD])|[^\\r\\n\\p{L}\\p{N}]?\\p{L}+|\\p{N}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+",
+                };
+                break;
             case LLAMA_VOCAB_PRE_TYPE_EXAONE_MOE:
                 regex_exprs = {
                     // original regex from tokenizer.json
@@ -2342,6 +2348,10 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
                 tokenizer_pre == "afmoe") {
                 pre_type = LLAMA_VOCAB_PRE_TYPE_AFMOE;
                 clean_spaces = false;
+            } else if (
+                tokenizer_pre == "laguna") {
+                pre_type = LLAMA_VOCAB_PRE_TYPE_LAGUNA;
+                clean_spaces = false;
             } else if (
                 tokenizer_pre == "minimax-m2") {
                 pre_type = LLAMA_VOCAB_PRE_TYPE_MINIMAX_M2;

src/llama-vocab.h

@@ -64,6 +64,7 @@ enum llama_vocab_pre_type {
     LLAMA_VOCAB_PRE_TYPE_WHITESPACE        = 53,
     LLAMA_VOCAB_PRE_TYPE_GRANITE_EMB_MULTI = 54,
     LLAMA_VOCAB_PRE_TYPE_MELLUM2           = 55,
+    LLAMA_VOCAB_PRE_TYPE_LAGUNA            = 56,
 };
 
 struct LLM_KV;

src/models/laguna.cpp

@@ -0,0 +1,332 @@
+// Laguna (poolside): sigmoid-routed MoE with a score-correction bias, one shared
+// expert, a softplus attention output gate, QK-norm, and per-layer-type RoPE
+// (YaRN on full-attention layers, plain RoPE on sliding-window layers). XS.2 is
+// hybrid full/SWA with a per-head gate; M.1 is full-attention with a per-element
+// gate. Shares the MoE/gate structure with afmoe.
+
+#include "models.h"
+
+void llama_model_laguna::load_arch_hparams(llama_model_loader & ml) {
+    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+    ml.get_key(LLM_KV_LEADING_DENSE_BLOCK_COUNT,   hparams.n_layer_dense_lead);
+    ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH,  hparams.n_ff_exp);
+    ml.get_key(LLM_KV_EXPERT_GATING_FUNC,          hparams.expert_gating_func, false);
+    ml.get_key(LLM_KV_EXPERT_WEIGHTS_SCALE,        hparams.expert_weights_scale, false);
+    ml.get_key(LLM_KV_EXPERT_WEIGHTS_NORM,         hparams.expert_weights_norm, false);
+
+    // Laguna ships one shared expert and stores its size directly (routed and
+    // shared experts may differ), so read the size from expert_shared_feed_forward_length.
+    // The count is not in the config; default to 1 but read the key if present.
+    hparams.n_expert_shared = 1;
+    ml.get_key(LLM_KV_EXPERT_SHARED_COUNT,               hparams.n_expert_shared, false);
+    ml.get_key(LLM_KV_EXPERT_SHARED_FEED_FORWARD_LENGTH, hparams.n_ff_shexp, false);
+    if (hparams.n_ff_shexp == 0) {
+        // Weightless fixtures (test-llama-archs) omit this key; derive a nonzero
+        // size so the shared expert is still built. Real GGUFs always carry the
+        // exact value (routed and shared FF lengths may differ).
+        hparams.n_ff_shexp = hparams.n_ff_exp * hparams.n_expert_shared;
+    }
+
+    // Sliding-window attention is OPTIONAL. XS.2 is hybrid (full / SWA / SWA /
+    // SWA repeating, period 4 starting with full); M.1 has no sliding window
+    // (all layers full attention). When sliding_window is absent or zero we
+    // leave swa_type = NONE and skip the SWA-specific per-layer-type RoPE.
+    hparams.n_swa = 0;
+    ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW, hparams.n_swa, false);
+    if (hparams.n_swa > 0) {
+        hparams.swa_type = LLAMA_SWA_TYPE_STANDARD;
+
+        uint32_t swa_period = 4;
+        ml.get_key_or_arr(LLM_KV_ATTENTION_SLIDING_WINDOW_PATTERN, swa_period, false);
+        hparams.set_swa_pattern(swa_period, /*dense_first=*/true);  // XS.2: FULL at il%4==0
+
+        // Per-layer-type RoPE: full layers use YaRN θ=500000 over 64 dims;
+        // SWA layers use default RoPE θ=10000 over 128 dims. Base load_hparams
+        // already reads ROPE_FREQ_BASE and ROPE_DIMENSION_COUNT into the
+        // non-SWA fields; we explicitly pull the SWA mirrors here.
+        hparams.rope_freq_base_train_swa  = hparams.rope_freq_base_train;
+        hparams.rope_freq_scale_train_swa = 1.0f;  // SWA uses plain RoPE (no YaRN scaling); do NOT inherit full layers 1/factor
+        ml.get_key(LLM_KV_ROPE_FREQ_BASE_SWA, hparams.rope_freq_base_train_swa, false);
+        ml.get_key(LLM_KV_ROPE_DIMENSION_COUNT_SWA, hparams.n_rot_swa, false);
+    }
+
+    // Default the expert gating function to SIGMOID when the key is absent
+    // (matches the HF reference).
+    if (hparams.expert_gating_func == LLAMA_EXPERT_GATING_FUNC_TYPE_NONE) {
+        hparams.expert_gating_func = LLAMA_EXPERT_GATING_FUNC_TYPE_SIGMOID;
+    }
+
+    switch (hparams.n_layer()) {
+        case 40: type = LLM_TYPE_30B_A3B;   break;  // Laguna-XS.2
+        case 70: type = LLM_TYPE_230B_A10B; break;  // Laguna-M.1
+        default: type = LLM_TYPE_UNKNOWN;
+    }
+}
+
+void llama_model_laguna::load_arch_tensors(llama_model_loader & ml) {
+    LLAMA_LOAD_LOCALS;
+
+    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
+
+    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
+    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
+    if (output == NULL) {
+        // tied embeddings fallback
+        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
+    }
+
+    const int64_t n_ff_exp   = hparams.n_ff_exp;
+    const int64_t n_ff_shexp = hparams.n_ff_shexp;
+
+    for (int i = 0; i < n_layer; ++i) {
+        auto & layer = layers[i];
+
+        // Per-layer head count — Laguna varies n_head between full and SWA
+        // layers (48 vs 64 in XS.2). KV head count is uniform.
+        const int64_t n_head_il    = hparams.n_head(i);
+        const int64_t n_head_kv_il = hparams.n_head_kv(i);
+        const int64_t n_embd_q_il  = n_embd_head_k * n_head_il;
+        const int64_t n_embd_k_il  = n_embd_head_k * n_head_kv_il;
+        const int64_t n_embd_v_il  = n_embd_head_v * n_head_kv_il;
+
+        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
+
+        create_tensor_qkv(layer, i, n_embd, n_embd_q_il, n_embd_k_il, n_embd_v_il, 0);
+        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_q_il, n_embd}, 0);
+
+        layer.attn_q_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {n_embd_head_k}, 0);
+        layer.attn_k_norm = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_embd_head_k}, 0);
+
+        // Attention output gate. XS.2 is per-head (g_proj -> n_head, one scalar
+        // per head broadcast over head_dim at multiply time); M.1 is per-element
+        // (g_proj -> n_head*head_dim, like afmoe). Detect from the stored tensor
+        // shape so a single arch handles both; the graph mirrors this check.
+        // Gate width selects per-head vs per-element. Real GGUFs always carry the
+        // gate tensor, so read the width from it and require EXACTLY one of the two
+        // valid widths -- never guess between them. Weightless fixtures
+        // (test-llama-archs) have no gate tensor; fall back to the per-head layout so
+        // the per-head reshape path is still exercised.
+        const int64_t n_gate_per_head = n_head_il;
+        const int64_t n_gate_per_elem = n_embd_head_k * n_head_il;
+        const ggml_tensor * gate_meta = ml.get_tensor_meta(tn(LLM_TENSOR_ATTN_GATE, "weight", i).str().c_str());
+        int64_t n_gate_out;
+        if (gate_meta != nullptr) {
+            n_gate_out = gate_meta->ne[1];
+            if (n_gate_out != n_gate_per_head && n_gate_out != n_gate_per_elem) {
+                GGML_ABORT("Laguna: unexpected attention gate width %lld at layer %d "
+                           "(expected %lld per-head or %lld per-element)",
+                           (long long) n_gate_out, i, (long long) n_gate_per_head, (long long) n_gate_per_elem);
+            }
+        } else {
+            n_gate_out = n_gate_per_head;
+        }
+        layer.wqkv_gate = create_tensor(tn(LLM_TENSOR_ATTN_GATE, "weight", i), {n_embd, n_gate_out}, 0);
+
+        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
+
+        if ((uint32_t)i >= hparams.n_layer_dense_lead) {
+            // MoE layer
+            layer.ffn_gate_inp    = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP,    "weight", i), {n_embd, n_expert}, 0);
+            layer.ffn_exp_probs_b = create_tensor(tn(LLM_TENSOR_FFN_EXP_PROBS_B, "bias",   i), {n_expert}, 0);
+
+            layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {n_embd,   n_ff_exp, n_expert}, 0);
+            layer.ffn_up_exps   = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i), {n_embd,   n_ff_exp, n_expert}, 0);
+            layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {n_ff_exp, n_embd,   n_expert}, 0);
+
+            // Always-on shared expert.
+            layer.ffn_gate_shexp = create_tensor(tn(LLM_TENSOR_FFN_GATE_SHEXP, "weight", i), {n_embd,    n_ff_shexp}, 0);
+            layer.ffn_up_shexp   = create_tensor(tn(LLM_TENSOR_FFN_UP_SHEXP,   "weight", i), {n_embd,    n_ff_shexp}, 0);
+            layer.ffn_down_shexp = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), {n_ff_shexp, n_embd},    0);
+        } else {
+            // Dense layer (the leading n_layer_dense_lead layers)
+            layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, 0);
+            layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd, n_ff}, 0);
+            layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff,   n_embd}, 0);
+        }
+    }
+}
+
+std::unique_ptr<llm_graph_context> llama_model_laguna::build_arch_graph(const llm_graph_params & params) const {
+    return std::make_unique<graph>(*this, params);
+}
+
+llama_model_laguna::graph::graph(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_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+    // No MuP embedding scale (laguna omits this; afmoe scales by sqrt(hidden)).
+
+    ggml_tensor * inp_pos = build_inp_pos();
+    // XS.2 is hybrid SWA -> interleaved-SWA KV input; M.1 is all-full -> plain
+    // KV input. Pick the matching input (and build_attn overload) per swa_type.
+    const bool has_swa = hparams.swa_type != LLAMA_SWA_TYPE_NONE;
+    llm_graph_input_attn_kv      * inp_attn_kv   = has_swa ? nullptr : build_attn_inp_kv();
+    llm_graph_input_attn_kv_iswa * inp_attn_iswa = has_swa ? build_attn_inp_kv_iswa() : nullptr;
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    const float kq_scale = 1.0f / sqrtf(float(n_embd_head));
+
+    for (int il = 0; il < n_layer; ++il) {
+        const bool    is_swa_il   = hparams.is_swa(il);
+        const int64_t n_head_il   = hparams.n_head(il);
+        const int64_t n_head_kv_il = hparams.n_head_kv(il);
+
+        // Per-layer-type RoPE config. SWA layers run plain rope (no YaRN),
+        // achieved by zeroing the YaRN ext/beta params for those layers.
+        const int   n_rot_l       = is_swa_il ? hparams.n_rot_swa : n_rot;
+        const float freq_base_l   = is_swa_il ? hparams.rope_freq_base_train_swa : freq_base;
+        const float freq_scale_l  = is_swa_il ? hparams.rope_freq_scale_train_swa : freq_scale;
+        const float ext_factor_l  = is_swa_il ? 0.0f : ext_factor;
+        // YaRN magnitude scaling (mscale) is already handled by the framework:
+        // llama_context pre-divides cparams.yarn_attn_factor by (1 + 0.1*ln(factor))
+        // to cancel ggml rope_yarn's internal mscale *= 1 + 0.1*ln(1/freq_scale).
+        // Pass attn_factor straight through (like every other arch); SWA layers run
+        // plain RoPE (ext_factor 0, no mscale) so force 1.0 there.
+        const float attn_factor_l = is_swa_il ? 1.0f : attn_factor;
+        const float beta_fast_l   = is_swa_il ? 0.0f : beta_fast;
+        const float beta_slow_l   = is_swa_il ? 0.0f : beta_slow;
+        const int   n_ctx_orig_l  = is_swa_il ? hparams.n_ctx_train : n_ctx_orig;
+
+        ggml_tensor * inpSA = inpL;
+
+        // Pre-norm
+        cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        // Self-attention
+        {
+            ggml_tensor * attn_inp = cur;  // saved for the gate projection
+
+            auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur,
+                    n_embd_head, n_head_il, n_head_kv_il, il);
+
+            // g_proj on the *pre-attention* hidden state (matches HF
+            // reference: gate is computed from the same `hidden_states`
+            // input as q/k/v, not from the attn output).
+            ggml_tensor * gate = build_lora_mm(model.layers[il].wqkv_gate, attn_inp);
+            cb(gate, "attn_gate_proj", il);
+
+            // QK RMSNorm at head_dim level (Qwen3 style)
+            Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);
+            Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il);
+            cb(Qcur, "Qcur_normed", il);
+            cb(Kcur, "Kcur_normed", il);
+
+            Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr,
+                    n_rot_l, rope_type, n_ctx_orig_l, freq_base_l, freq_scale_l,
+                    ext_factor_l, attn_factor_l, beta_fast_l, beta_slow_l);
+            Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr,
+                    n_rot_l, rope_type, n_ctx_orig_l, freq_base_l, freq_scale_l,
+                    ext_factor_l, attn_factor_l, beta_fast_l, beta_slow_l);
+            cb(Qcur, "Qcur_rope", il);
+            cb(Kcur, "Kcur_rope", il);
+
+            cur = has_swa
+                ? build_attn(inp_attn_iswa,
+                        NULL, NULL, NULL,    // o_proj deferred until after gating
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il)
+                : build_attn(inp_attn_kv,
+                        NULL, NULL, NULL,
+                        Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
+            cb(cur, "attn_out", il);
+
+            // Softplus output gate (the unary kernel computes softplus in fp32
+            // and casts back). Two shapes, distinguished by the g_proj output
+            // dim (matching the load-time detection):
+            //   XS.2 per-head     : gate [n_head_il, n_tokens] -> reshape to
+            //                       [1, n_head_il, n_tokens] and broadcast over
+            //                       head_dim against cur [head_dim, n_head, T].
+            //   M.1  per-element  : gate [n_head_il*head_dim, n_tokens] spans the
+            //                       full attention output -> direct ggml_mul.
+            gate = ggml_softplus(ctx0, gate);
+            cb(gate, "attn_gate_softplus", il);
+
+            const int64_t n_tokens = cur->ne[1];
+            if (model.layers[il].wqkv_gate->ne[1] == n_head_il) {
+                cur  = ggml_reshape_3d(ctx0, cur,  n_embd_head, n_head_il, n_tokens);
+                gate = ggml_reshape_3d(ctx0, gate, 1,           n_head_il, n_tokens);
+                cur  = ggml_mul(ctx0, cur, gate);
+                cur  = ggml_reshape_2d(ctx0, cur, n_embd_head * n_head_il, n_tokens);
+            } else {
+                cur = ggml_mul(ctx0, cur, gate);
+            }
+            cb(cur, "attn_gated", il);
+
+            cur = build_lora_mm(model.layers[il].wo, cur, model.layers[il].wo_s);
+            cb(cur, "attn_o_proj", il);
+        }
+
+        if (il == n_layer - 1 && inp_out_ids) {
+            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);
+
+        // Pre-norm only (no post-attn norm)
+        cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
+        cb(cur, "ffn_norm", il);
+
+        if ((uint32_t)il >= hparams.n_layer_dense_lead) {
+            // MoE: sigmoid routing + score-correction bias + sum-norm +
+            // routed_scaling_factor (all handled by build_moe_ffn).
+            ggml_tensor * moe_out = 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,
+                    hparams.expert_weights_norm,
+                    hparams.expert_weights_scale,
+                    (llama_expert_gating_func_type) hparams.expert_gating_func,
+                    il);
+            cb(moe_out, "ffn_moe_out", il);
+
+            // Always-on shared expert, summed in parallel.
+            ggml_tensor * ffn_shexp = build_ffn(cur,
+                    model.layers[il].ffn_up_shexp,   NULL, NULL,
+                    model.layers[il].ffn_gate_shexp, NULL, NULL,
+                    model.layers[il].ffn_down_shexp, NULL, NULL,
+                    NULL,
+                    LLM_FFN_SILU, LLM_FFN_PAR, il);
+            cb(ffn_shexp, "ffn_shexp", il);
+
+            cur = ggml_add(ctx0, moe_out, ffn_shexp);
+            cb(cur, "ffn_out", il);
+        } else {
+            // Dense FFN for the leading n_layer_dense_lead layers (XS.2: 1, M.1: 3)
+            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);
+        }
+
+        // No post-ffn norm
+        cur = ggml_add(ctx0, cur, ffn_inp);
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        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;
+
+    cur = build_lora_mm(model.output, cur);
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}

src/models/models.h

@@ -1681,6 +1681,19 @@ struct llama_model_afmoe : public llama_model_base {
 };
 
 
+struct llama_model_laguna : public llama_model_base {
+    llama_model_laguna(const struct llama_model_params & params) : llama_model_base(params) {}
+    void load_arch_hparams(llama_model_loader & ml) override;
+    void load_arch_tensors(llama_model_loader & ml) override;
+
+    struct graph : public llm_graph_context {
+        graph(const llama_model & model, const llm_graph_params & params);
+    };
+
+    std::unique_ptr<llm_graph_context> build_arch_graph(const llm_graph_params & params) const override;
+};
+
+
 struct llama_model_ernie4_5 : public llama_model_base {
     llama_model_ernie4_5(const struct llama_model_params & params) : llama_model_base(params) {}
     void load_arch_hparams(llama_model_loader & ml) override;

tests/test-chat-auto-parser.cpp

@@ -57,6 +57,15 @@ static void test_seed_oss_tool_with_reasoning(testing & t);
 static void test_nemotron_analysis(testing & t);
 static void test_nemotron_reasoning_detection(testing & t);
 static void test_nemotron_tool_format(testing & t);
+static void test_laguna_analysis(testing & t);
+static void test_laguna_reasoning_detection(testing & t);
+static void test_laguna_tool_format(testing & t);
+static void test_laguna_s_analysis(testing & t);
+static void test_laguna_s_reasoning_detection(testing & t);
+static void test_laguna_s_tool_format(testing & t);
+static void test_laguna_xs2_analysis(testing & t);
+static void test_laguna_xs2_reasoning_detection(testing & t);
+static void test_laguna_xs2_tool_format(testing & t);
 
 // CohereForAI template analysis tests
 static void test_cohere_reasoning_detection(testing & t);
@@ -101,6 +110,9 @@ int main(int argc, char * argv[]) {
     t.test("seed_oss_diffs", test_seed_oss_tool_analysis);
     t.test("cohere", test_cohere_analysis);
     t.test("nemotron", test_nemotron_analysis);
+    t.test("laguna", test_laguna_analysis);
+    t.test("laguna-s", test_laguna_s_analysis);
+    t.test("laguna-xs2", test_laguna_xs2_analysis);
     t.test("smollm3", test_smollm3_analysis);
     t.test("standard_json_tools", test_standard_json_tools_formats);
     t.test("normalize_quotes_to_json", test_normalize_quotes_to_json);
@@ -1378,6 +1390,94 @@ static void test_nemotron_tool_format(testing & t) {
     t.assert_true("should support tools", analysis.jinja_caps.supports_tools);
 }
 
+// ============================================================================
+// Laguna Template Analysis Tests
+// ============================================================================
+static common_chat_template load_laguna_template(testing & t) {
+    return load_template(t, "models/templates/poolside-Laguna-XS-2.1.jinja");
+}
+
+static void test_laguna_reasoning_detection(testing & t) {
+    common_chat_template tmpl = load_laguna_template(t);
+    struct autoparser analysis;
+    analysis.analyze_template(tmpl);
+    // Laguna's template renders reasoning delimiters with formatting whitespace
+    // ("<think>\n") that the model does not emit; the Laguna patch trims them.
+    t.assert_equal("reasoning_start should be '<think>'", "<think>", analysis.reasoning.start);
+    t.assert_equal("reasoning_end should be '</think>'", "</think>", analysis.reasoning.end);
+    t.assert_equal("reasoning should be TAG_BASED", reasoning_mode::TAG_BASED, analysis.reasoning.mode);
+}
+
+static void test_laguna_tool_format(testing & t) {
+    common_chat_template tmpl = load_laguna_template(t);
+    struct autoparser analysis;
+    analysis.analyze_template(tmpl);
+    t.assert_equal("arg_value_suffix should be '</arg_value>'", "</arg_value>", analysis.tools.arguments.value_suffix);
+}
+
+static void test_laguna_stop_string(testing & t) {
+    // The </assistant> turn terminator can be emitted as ordinary text tokens
+    // (not the single eot token), so it must also be a literal stop string.
+    common_chat_template tmpl = load_laguna_template(t);
+    struct autoparser analysis;
+    analysis.analyze_template(tmpl);
+    bool has_stop = false;
+    for (const auto & stop : analysis.additional_stops) {
+        if (stop == "</assistant>") { has_stop = true; break; }
+    }
+    t.assert_true("Laguna additional_stops contains </assistant>", has_stop);
+}
+
+static void test_laguna_analysis(testing & t) {
+    t.test("Laguna reasoning detection", test_laguna_reasoning_detection);
+    t.test("Laguna tool format", test_laguna_tool_format);
+    t.test("Laguna stop string", test_laguna_stop_string);
+}
+
+static common_chat_template load_laguna_s_template(testing & t) {
+    return load_template(t, "models/templates/poolside-Laguna-S-2.1.jinja");
+}
+static void test_laguna_s_reasoning_detection(testing & t) {
+    common_chat_template tmpl = load_laguna_s_template(t);
+    struct autoparser analysis;
+    analysis.analyze_template(tmpl);
+    t.assert_equal("Laguna-S(v8) reasoning_start should be '<think>'", "<think>", analysis.reasoning.start);
+    t.assert_equal("Laguna-S(v8) reasoning_end should be '</think>'", "</think>", analysis.reasoning.end);
+    t.assert_equal("Laguna-S(v8) reasoning should be TAG_BASED", reasoning_mode::TAG_BASED, analysis.reasoning.mode);
+}
+static void test_laguna_s_tool_format(testing & t) {
+    common_chat_template tmpl = load_laguna_s_template(t);
+    struct autoparser analysis;
+    analysis.analyze_template(tmpl);
+    t.assert_equal("Laguna-S(v8) arg_value_suffix should be '</arg_value>'", "</arg_value>", analysis.tools.arguments.value_suffix);
+}
+static void test_laguna_s_analysis(testing & t) {
+    t.test("Laguna-S(v8) reasoning detection", test_laguna_s_reasoning_detection);
+    t.test("Laguna-S(v8) tool format", test_laguna_s_tool_format);
+}
+
+static common_chat_template load_laguna_xs2_template(testing & t) {
+    return load_template(t, "models/templates/poolside-Laguna-XS.2.jinja");
+}
+static void test_laguna_xs2_reasoning_detection(testing & t) {
+    common_chat_template tmpl = load_laguna_xs2_template(t);
+    struct autoparser analysis;
+    analysis.analyze_template(tmpl);
+    t.assert_equal("Laguna-XS.2(v5) reasoning_start should be '<think>'", "<think>", analysis.reasoning.start);
+    t.assert_equal("Laguna-XS.2(v5) reasoning_end should be '</think>'", "</think>", analysis.reasoning.end);
+    t.assert_equal("Laguna-XS.2(v5) reasoning should be TAG_BASED", reasoning_mode::TAG_BASED, analysis.reasoning.mode);
+}
+static void test_laguna_xs2_tool_format(testing & t) {
+    common_chat_template tmpl = load_laguna_xs2_template(t);
+    struct autoparser analysis;
+    analysis.analyze_template(tmpl);
+    t.assert_equal("Laguna-XS.2(v5) arg_value_suffix should be '</arg_value>'", "</arg_value>", analysis.tools.arguments.value_suffix);
+}
+static void test_laguna_xs2_analysis(testing & t) {
+    t.test("Laguna-XS.2(v5) reasoning detection", test_laguna_xs2_reasoning_detection);
+    t.test("Laguna-XS.2(v5) tool format", test_laguna_xs2_tool_format);
+}
+
 static common_chat_template load_cohere_template(testing & t) {
     return load_template(t, "models/templates/CohereForAI-c4ai-command-r7b-12-2024-tool_use.jinja");
 }

tests/test-llama-archs.cpp

@@ -362,6 +362,7 @@ static bool moe_mandatory(const llm_arch arch) {
         case LLM_ARCH_STEP35:
         case LLM_ARCH_MISTRAL4:
         case LLM_ARCH_MELLUM:
+        case LLM_ARCH_LAGUNA:
             return true;
         default:
             return false;