ggml: enable PowerPC backend variants on AIX (#25983)

* ggml: enable PowerPC backend variants on AIX

Allow the PowerPC CPU backend variants to be built on AIX by extending the platform check in the CMake configuration. This reuses the existing PowerPC backend implementations without changing their behavior.

Also fix a missing semicolon in the PowerPC Q0 matmul implementation.

* Fix missing semicolon in sgemm.cpp

.github/workflows/release.yml

@@ -1109,6 +1109,8 @@ jobs:
             -DGGML_SYCL=ON \
             -DCMAKE_C_COMPILER=icx \
             -DCMAKE_CXX_COMPILER=icpx \
+            -DCMAKE_INSTALL_RPATH='$ORIGIN' \
+            -DCMAKE_BUILD_WITH_INSTALL_RPATH=ON \
             -DLLAMA_OPENSSL=OFF \
             -DGGML_NATIVE=OFF \
             -DGGML_SYCL_F16=${{ matrix.fp16 }}

common/arg.cpp

@@ -527,8 +527,43 @@ void common_models_handler_apply(common_models_handler & handler, common_params
         }
     };
 
+    // when a sidecar type is requested, the draft repo resolves to its sidecar instead of a full model
+    const bool spec_sidecar_found = !plan_spec.mtp.local_path.empty() ||
+                                    !plan_spec.dflash.local_path.empty() ||
+                                    !plan_spec.eagle3.local_path.empty();
+    if (!plan_spec.mtp.local_path.empty() && !had_spec_url) {
+        tasks.emplace_back(plan_spec.mtp, opts, [&]() {
+            // only use the discovered MTP head when no draft path is set yet
+            if (params.speculative.draft.mparams.path.empty()) {
+                params.speculative.draft.mparams.path = hf_cache::finalize_file(plan_spec.mtp);
+            } else {
+                hf_cache::finalize_file(plan_spec.mtp);
+            }
+        });
+    }
+    if (!plan_spec.dflash.local_path.empty() && !had_spec_url) {
+        tasks.emplace_back(plan_spec.dflash, opts, [&]() {
+            // only use the discovered DFlash sidecar when no draft path is set yet
+            if (params.speculative.draft.mparams.path.empty()) {
+                params.speculative.draft.mparams.path = hf_cache::finalize_file(plan_spec.dflash);
+            } else {
+                hf_cache::finalize_file(plan_spec.dflash);
+            }
+        });
+    }
+    if (!plan_spec.eagle3.local_path.empty() && !had_spec_url) {
+        tasks.emplace_back(plan_spec.eagle3, opts, [&]() {
+            // only use the discovered Eagle3 sidecar when no draft path is set yet
+            if (params.speculative.draft.mparams.path.empty()) {
+                params.speculative.draft.mparams.path = hf_cache::finalize_file(plan_spec.eagle3);
+            } else {
+                hf_cache::finalize_file(plan_spec.eagle3);
+            }
+        });
+    }
+
     // handle plan_spec (e.g. --spec-draft-hf)
-    if (!plan_spec.model_files.empty() && !had_spec_url) {
+    if (!plan_spec.model_files.empty() && !had_spec_url && !spec_sidecar_found) {
         add_tasks(plan_spec.model_files, plan_spec.primary, params.speculative.draft.mparams);
         had_spec_url = true;
     }

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/hunyuan.py

@@ -338,6 +338,12 @@ class HunyuanVLTextModel(HunYuanModel):
 
     def __init__(self, dir_model: Path, *args, **kwargs):
         super().__init__(dir_model, *args, **kwargs)
+        # transformers 5.13.0 encodes HunyuanVL XD-RoPE as dynamic + mrope_section.
+        # Normalize it to avoid the HunYuan dynamic-RoPE context assertion.
+        if self.rope_parameters.get("rope_type") == "dynamic" and "mrope_section" in self.rope_parameters:
+            self.rope_parameters["rope_type"] = "xdrope"
+            self.rope_parameters["type"] = "xdrope"
+            self.rope_parameters["xdrope_section"] = list(self.rope_parameters["mrope_section"])
 
     def set_gguf_parameters(self):
         super().set_gguf_parameters()

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

docs/ops.md

@@ -25,10 +25,10 @@ Legend:
 |                             CEIL | ❌ | ❌ | ✅ | 🟡 | 🟡 | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                            CLAMP | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | 🟡 | ✅ | 🟡 | ✅ | ❌ | ❌ |
 |                        COL2IM_1D | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ |
-|                           CONCAT | ❌ | ✅ | ✅ | 🟡 | 🟡 | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                           CONCAT | ❌ | ✅ | ✅ | 🟡 | 🟡 | ✅ | 🟡 | ✅ | ✅ | 🟡 | ❌ | ❌ |
 |                             CONT | ❌ | 🟡 | ✅ | ✅ | 🟡 | ✅ | 🟡 | ✅ | ✅ | 🟡 | ❌ | ❌ |
 |                          CONV_2D | ❌ | ❌ | ✅ | ✅ | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
-|                       CONV_2D_DW | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
+|                       CONV_2D_DW | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                          CONV_3D | ❌ | ❌ | ✅ | ❌ | ❌ | ✅ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ |
 |                CONV_TRANSPOSE_1D | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                CONV_TRANSPOSE_2D | ❌ | ❌ | ✅ | ✅ | ❌ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
@@ -41,6 +41,9 @@ Legend:
 |                             DIAG | ❌ | ❌ | ✅ | ✅ | 🟡 | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                    DIAG_MASK_INF | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | 🟡 | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                              DIV | ❌ | ✅ | ✅ | ✅ | ❌ | 🟡 | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                     DSV4_HC_COMB | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
+|                     DSV4_HC_POST | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
+|                      DSV4_HC_PRE | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
 |                              DUP | ❌ | ✅ | ✅ | 🟡 | ❌ | 🟡 | 🟡 | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                              ELU | ❌ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                              EXP | ❌ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
@@ -63,16 +66,17 @@ Legend:
 |                        HARDSWISH | ❌ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                           IM2COL | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                        IM2COL_3D | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
-|                          L2_NORM | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                          L2_NORM | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ❌ | ✅ | ✅ | 🟡 | ❌ | ❌ |
 |                       LEAKY_RELU | ❌ | ✅ | ✅ | ✅ | ❌ | 🟡 | ❌ | ✅ | 🟡 | ❌ | ❌ | ❌ |
+|                LIGHTNING_INDEXER | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
 |                              LOG | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                             MEAN | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                              MUL | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                          MUL_MAT | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 |
-|                 MUL_MAT_HADAMARD | ❌ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
+|                 MUL_MAT_HADAMARD | ❌ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                       MUL_MAT_ID | ❌ | 🟡 | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | ✅ | 🟡 | 🟡 | ❌ |
 |                              NEG | ❌ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
-|                             NORM | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | ✅ | 🟡 | ✅ | ❌ | ❌ |
+|                             NORM | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | ✅ | 🟡 | 🟡 | ❌ | ❌ |
 |                   OPT_STEP_ADAMW | ❌ | ❌ | ✅ | ✅ | ❌ | ✅ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ |
 |                     OPT_STEP_SGD | ❌ | ❌ | ✅ | ✅ | ❌ | ✅ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ |
 |                         OUT_PROD | 🟡 | 🟡 | 🟡 | 🟡 | ❌ | ❌ | ❌ | 🟡 | ❌ | ❌ | ❌ | 🟡 |
@@ -82,7 +86,7 @@ Legend:
 |                          POOL_2D | ❌ | 🟡 | ✅ | ✅ | ❌ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                            REGLU | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                             RELU | ❌ | ✅ | ✅ | 🟡 | 🟡 | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
-|                           REPEAT | ❌ | ✅ | ✅ | 🟡 | 🟡 | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                           REPEAT | ❌ | ✅ | ✅ | 🟡 | 🟡 | ✅ | 🟡 | ✅ | ✅ | 🟡 | ❌ | ❌ |
 |                      REPEAT_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                         RMS_NORM | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                    RMS_NORM_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |

ggml/src/CMakeLists.txt

@@ -430,7 +430,7 @@ if (GGML_CPU_ALL_VARIANTS)
             message(FATAL_ERROR "Unsupported ARM target OS: ${CMAKE_SYSTEM_NAME}")
         endif()
     elseif (GGML_SYSTEM_ARCH STREQUAL "PowerPC")
-        if (CMAKE_SYSTEM_NAME MATCHES "Linux")
+        if (CMAKE_SYSTEM_NAME MATCHES "Linux|AIX")
             ggml_add_cpu_backend_variant(power0)
             ggml_add_cpu_backend_variant(power7_1       POWER7)
             ggml_add_cpu_backend_variant(power7_2       POWER7  VSX)

ggml/src/ggml-cpu/kleidiai/kleidiai.cpp

@@ -1719,6 +1719,7 @@ class extra_buffer_type : ggml::cpu::extra_buffer_type {
 
             return true;
         }
+
         return false;
     }
 
@@ -1727,6 +1728,20 @@ class extra_buffer_type : ggml::cpu::extra_buffer_type {
             if (op->src[0]->buffer && op->src[0]->buffer->buft == ggml_backend_cpu_kleidiai_buffer_type()) {
                 return (ggml::cpu::tensor_traits *) op->src[0]->extra;
             } else {
+                // KleidiAI only has kernels for Q4_0 and Q8_0. For a quantized weight of any
+                // other type (K-quants, IQ) it declines the op and returns nullptr below, so
+                // KleidiAI does not accelerate it. Another CPU backend may still take the op,
+                // and this can run during graph planning, so the message says what KleidiAI
+                // did rather than what ends up executing. Warn once per process.
+                if (ggml_is_quantized(op->src[0]->type) &&
+                    op->src[0]->type != GGML_TYPE_Q4_0 && op->src[0]->type != GGML_TYPE_Q8_0) {
+                    static std::atomic<bool> warned(false);
+                    if (!warned.exchange(true)) {
+                        GGML_LOG_WARN("kleidiai: no kernel for tensor type %s, not accelerated by KleidiAI "
+                                      "(kernels available for Q4_0 and Q8_0)\n",
+                                      ggml_type_name(op->src[0]->type));
+                    }
+                }
                 if (op->src[0]->type != GGML_TYPE_F16) {
                     return nullptr;
                 }

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

@@ -2329,7 +2329,7 @@ class tinyBLAS_Q0_PPC {
         mc = 32;
         nc = 32;
         kc = 32;
-        n_chunk = 32
+        n_chunk = 32;
     #endif
         int64_t n_aligned = 0;
         if (n % n_chunk == 0) {

ggml/src/ggml-cuda/common.cuh

@@ -937,6 +937,9 @@ static __device__ __forceinline__ uint2 fast_div_modulo(uint32_t n, const uint3
 
 typedef void (*dequantize_kernel_t)(const void * vx, const int64_t ib, const int iqs, float2 & v);
 
+template<typename dst_t>
+using dequantize_kq_t = void (*)(const void * vx, const int64_t ib, dst_t * y, const int tid);
+
 static __device__ __forceinline__ float get_alibi_slope(
     const float max_bias, const uint32_t h, const uint32_t n_head_log2, const float m0, const float m1
 ) {

ggml/src/ggml-cuda/convert.cu

@@ -140,358 +140,107 @@ static __global__ void dequantize_block_q4_1(const void * __restrict__ vx, dst_t
 
 template<typename dst_t>
 static __global__ void dequantize_block_q2_K(const void * __restrict__ vx, dst_t * __restrict__ yy) {
+    const int64_t i = blockIdx.x;
 
-    const int64_t i   = blockIdx.x;
-    const block_q2_K * x = (const block_q2_K *) vx;
-
-    const int64_t tid = threadIdx.x;
-    const int64_t n   = tid/32;
-    const int64_t l   = tid - 32*n;
-    const int64_t is  = 8*n + l/16;
-
-    const uint8_t q = x[i].qs[32*n + l];
-    dst_t * y = yy + i*QK_K + 128*n;
-
-    float dall = __low2half(x[i].dm);
-    float dmin = __high2half(x[i].dm);
-    y[l+ 0] = ggml_cuda_cast<dst_t>(dall * (x[i].scales[is+0] & 0xF) * ((q >> 0) & 3) - dmin * (x[i].scales[is+0] >> 4));
-    y[l+32] = ggml_cuda_cast<dst_t>(dall * (x[i].scales[is+2] & 0xF) * ((q >> 2) & 3) - dmin * (x[i].scales[is+2] >> 4));
-    y[l+64] = ggml_cuda_cast<dst_t>(dall * (x[i].scales[is+4] & 0xF) * ((q >> 4) & 3) - dmin * (x[i].scales[is+4] >> 4));
-    y[l+96] = ggml_cuda_cast<dst_t>(dall * (x[i].scales[is+6] & 0xF) * ((q >> 6) & 3) - dmin * (x[i].scales[is+6] >> 4));
+    dequantize_q2_K(vx, i, yy + i*QK_K, threadIdx.x);
 }
 
 template<typename dst_t>
 static __global__ void dequantize_block_q3_K(const void * __restrict__ vx, dst_t * __restrict__ yy) {
-
     const int64_t i = blockIdx.x;
-    const block_q3_K * x = (const block_q3_K *) vx;
 
-    const int64_t r = threadIdx.x/4;
-    const int64_t tid = r/2;
-    const int64_t is0 = r%2;
-    const int64_t l0 = 16*is0 + 4*(threadIdx.x%4);
-    const int64_t n = tid / 4;
-    const int64_t j = tid - 4*n;
-
-    uint8_t m = 1 << (4*n + j);
-    int64_t is = 8*n + 2*j + is0;
-    int shift = 2*j;
-
-    int8_t us = is <  4 ? (x[i].scales[is-0] & 0xF) | (((x[i].scales[is+8] >> 0) & 3) << 4) :
-                is <  8 ? (x[i].scales[is-0] & 0xF) | (((x[i].scales[is+4] >> 2) & 3) << 4) :
-                is < 12 ? (x[i].scales[is-8] >>  4) | (((x[i].scales[is+0] >> 4) & 3) << 4) :
-                          (x[i].scales[is-8] >>  4) | (((x[i].scales[is-4] >> 6) & 3) << 4);
-    float d_all = x[i].d;
-    float dl = d_all * (us - 32);
-
-    dst_t * y = yy + i*QK_K + 128*n + 32*j;
-    const uint8_t * q = x[i].qs + 32*n;
-    const uint8_t * hm = x[i].hmask;
-
-    for (int l = l0; l < l0+4; ++l) {
-        y[l] = ggml_cuda_cast<dst_t>(dl * ((int8_t)((q[l] >> shift) & 3) - ((hm[l] & m) ? 0 : 4)));
-    }
-}
-
-static inline __device__ void get_scale_min_k4(int j, const uint8_t * q, uint8_t & d, uint8_t & m) {
-    if (j < 4) {
-        d = q[j] & 63; m = q[j + 4] & 63;
-    } else {
-        d = (q[j+4] & 0xF) | ((q[j-4] >> 6) << 4);
-        m = (q[j+4] >>  4) | ((q[j-0] >> 6) << 4);
-    }
+    dequantize_q3_K(vx, i, yy + i*QK_K, threadIdx.x);
 }
 
 template<typename dst_t>
 static __global__ void dequantize_block_q4_K(const void * __restrict__ vx, dst_t * __restrict__ yy) {
-    const block_q4_K * x = (const block_q4_K *) vx;
-
     const int64_t i = blockIdx.x;
 
-    // assume 32 threads
-    const int64_t tid = threadIdx.x;
-    const int64_t il  = tid/8;
-    const int64_t ir  = tid%8;
-    const int64_t is  = 2*il;
-    const int64_t n   = 4;
-
-    dst_t * y = yy + i*QK_K + 64*il + n*ir;
-
-    const float dall = __low2half(x[i].dm);
-    const float dmin = __high2half(x[i].dm);
-
-    const uint8_t * q = x[i].qs + 32*il + n*ir;
-
-    uint8_t sc, m;
-    get_scale_min_k4(is + 0, x[i].scales, sc, m);
-    const float d1 = dall * sc; const float m1 = dmin * m;
-    get_scale_min_k4(is + 1, x[i].scales, sc, m);
-    const float d2 = dall * sc; const float m2 = dmin * m;
-    for (int l = 0; l < n; ++l) {
-        y[l + 0] = ggml_cuda_cast<dst_t>(d1 * (q[l] & 0xF) - m1);
-        y[l +32] = ggml_cuda_cast<dst_t>(d2 * (q[l] >>  4) - m2);
-    }
+    dequantize_q4_K(vx, i, yy + i*QK_K, threadIdx.x);
 }
 
 template<typename dst_t>
 static __global__ void dequantize_block_q5_K(const void * __restrict__ vx, dst_t * __restrict__ yy) {
-    const block_q5_K * x = (const block_q5_K *) vx;
-
     const int64_t i = blockIdx.x;
 
-    // assume 64 threads - this is very slightly better than the one below
-    const int64_t tid = threadIdx.x;
-    const int64_t il  = tid/16;   // il is in 0...3
-    const int64_t ir  = tid%16;   // ir is in 0...15
-    const int64_t is  = 2*il;     // is is in 0...6
-
-    dst_t * y = yy + i*QK_K + 64*il + 2*ir;
-
-    const float dall = __low2half(x[i].dm);
-    const float dmin = __high2half(x[i].dm);
-
-    const uint8_t * ql = x[i].qs + 32*il + 2*ir;
-    const uint8_t * qh = x[i].qh + 2*ir;
-
-    uint8_t sc, m;
-    get_scale_min_k4(is + 0, x[i].scales, sc, m);
-    const float d1 = dall * sc; const float m1 = dmin * m;
-    get_scale_min_k4(is + 1, x[i].scales, sc, m);
-    const float d2 = dall * sc; const float m2 = dmin * m;
-
-    uint8_t   hm  = 1 << (2*il);
-    y[ 0] = ggml_cuda_cast<dst_t>(d1 * ((ql[ 0] & 0xF) + (qh[ 0] & hm ? 16 : 0)) - m1);
-    y[ 1] = ggml_cuda_cast<dst_t>(d1 * ((ql[ 1] & 0xF) + (qh[ 1] & hm ? 16 : 0)) - m1);
-    hm <<= 1;
-    y[32] = ggml_cuda_cast<dst_t>(d2 * ((ql[ 0] >>  4) + (qh[ 0] & hm ? 16 : 0)) - m2);
-    y[33] = ggml_cuda_cast<dst_t>(d2 * ((ql[ 1] >>  4) + (qh[ 1] & hm ? 16 : 0)) - m2);
+    dequantize_q5_K(vx, i, yy + i*QK_K, threadIdx.x);
 }
 
 template<typename dst_t>
 static __global__ void dequantize_block_q6_K(const void * __restrict__ vx, dst_t * __restrict__ yy) {
-    const block_q6_K * x = (const block_q6_K *) vx;
-
     const int64_t i = blockIdx.x;
 
-    // assume 64 threads - this is very slightly better than the one below
-    const int64_t tid = threadIdx.x;
-    const int64_t ip  = tid/32;   // ip is 0 or 1
-    const int64_t il  = tid - 32*ip; // 0...32
-    const int64_t is  = 8*ip + il/16;
-
-    dst_t * y = yy + i*QK_K + 128*ip + il;
-
-    const float d = x[i].d;
-
-    const uint8_t * ql = x[i].ql + 64*ip + il;
-    const uint8_t   qh = x[i].qh[32*ip + il];
-    const int8_t  * sc = x[i].scales + is;
-
-    y[ 0] = ggml_cuda_cast<dst_t>(d * sc[0] * ((int8_t)((ql[ 0] & 0xF) | (((qh >> 0) & 3) << 4)) - 32));
-    y[32] = ggml_cuda_cast<dst_t>(d * sc[2] * ((int8_t)((ql[32] & 0xF) | (((qh >> 2) & 3) << 4)) - 32));
-    y[64] = ggml_cuda_cast<dst_t>(d * sc[4] * ((int8_t)((ql[ 0]  >> 4) | (((qh >> 4) & 3) << 4)) - 32));
-    y[96] = ggml_cuda_cast<dst_t>(d * sc[6] * ((int8_t)((ql[32]  >> 4) | (((qh >> 6) & 3) << 4)) - 32));
+    dequantize_q6_K(vx, i, yy + i*QK_K, threadIdx.x);
 }
 
 template<typename dst_t>
 static __global__ void dequantize_block_iq2_xxs(const void * __restrict__ vx, dst_t * __restrict__ yy) {
+    const int64_t i = blockIdx.x;
 
-    const int64_t i   = blockIdx.x;
-    const block_iq2_xxs * x = (const block_iq2_xxs  *) vx;
-
-    const int64_t tid = threadIdx.x;
-    const int64_t il = tid/8; // 0...3
-    const int64_t ib = tid%8; // 0...7
-    dst_t * y = yy + i*QK_K + 32*ib + 8*il;
-    const uint16_t * q2 = x[i].qs + 4*ib;
-    const uint8_t  * aux8 = (const uint8_t *)q2;
-    const uint8_t  * grid = (const uint8_t *)(iq2xxs_grid + aux8[il]);
-    const uint32_t aux32 = q2[2] | (q2[3] << 16);
-    const float d = (float)x[i].d * (0.5f + (aux32 >> 28)) * 0.25f;
-    const uint8_t signs = ksigns_iq2xs[(aux32 >> 7*il) & 127];
-    for (int j = 0; j < 8; ++j) {
-        y[j] = ggml_cuda_cast<dst_t>(d * grid[j] * (signs & kmask_iq2xs[j] ? -1.f : 1.f));
-    }
+    dequantize_iq2_xxs(vx, i, yy + i*QK_K, threadIdx.x);
 }
 
 template<typename dst_t>
 static __global__ void dequantize_block_iq2_xs(const void * __restrict__ vx, dst_t * __restrict__ yy) {
+    const int64_t i = blockIdx.x;
 
-    const int64_t i   = blockIdx.x;
-    const block_iq2_xs * x = (const block_iq2_xs *) vx;
-
-    const int64_t tid = threadIdx.x;
-    const int64_t il = tid/8; // 0...3
-    const int64_t ib = tid%8; // 0...7
-    dst_t * y = yy + i*QK_K + 32*ib + 8*il;
-    const uint16_t * q2 = x[i].qs + 4*ib;
-    const uint8_t  * grid = (const uint8_t *)(iq2xs_grid + (q2[il] & 511));
-    const float d = (float)x[i].d * (0.5f + ((x[i].scales[ib] >> 4*(il/2)) & 0xf)) * 0.25f;
-    const uint8_t signs = ksigns_iq2xs[q2[il] >> 9];
-    for (int j = 0; j < 8; ++j) {
-        y[j] = ggml_cuda_cast<dst_t>(d * grid[j] * (signs & kmask_iq2xs[j] ? -1.f : 1.f));
-    }
+    dequantize_iq2_xs(vx, i, yy + i*QK_K, threadIdx.x);
 }
 
 template<typename dst_t>
 static __global__ void dequantize_block_iq2_s(const void * __restrict__ vx, dst_t * __restrict__ yy) {
+    const int64_t i = blockIdx.x;
 
-    const int64_t i   = blockIdx.x;
-    const block_iq2_s * x = (const block_iq2_s *) vx;
-
-    const int64_t tid = threadIdx.x;
-    const int64_t il = tid/8; // 0...3
-    const int64_t ib = tid%8; // 0...7
-    dst_t * y = yy + i*QK_K + 32*ib + 8*il;
-    const uint8_t * grid = (const uint8_t *)(iq2s_grid + (x[i].qs[4*ib+il] | ((x[i].qh[ib] << (8-2*il)) & 0x300)));
-    const float d = (float)x[i].d * (0.5f + ((x[i].scales[ib] >> 4*(il/2)) & 0xf)) * 0.25f;
-    const uint8_t signs = x[i].qs[QK_K/8+4*ib+il];
-    for (int j = 0; j < 8; ++j) {
-        y[j] = ggml_cuda_cast<dst_t>(d * grid[j] * (signs & kmask_iq2xs[j] ? -1.f : 1.f));
-    }
+    dequantize_iq2_s(vx, i, yy + i*QK_K, threadIdx.x);
 }
 
 template<typename dst_t>
 static __global__ void dequantize_block_iq3_xxs(const void * __restrict__ vx, dst_t * __restrict__ yy) {
+    const int64_t i = blockIdx.x;
 
-    const int64_t i   = blockIdx.x;
-    const block_iq3_xxs * x = (const block_iq3_xxs  *) vx;
-
-    const int64_t tid = threadIdx.x;
-    const int64_t il = tid/8; // 0...3
-    const int64_t ib = tid%8; // 0...7
-    dst_t * y = yy + i*QK_K + 32*ib + 8*il;
-    const uint8_t  * q3 = x[i].qs + 8*ib;
-    const uint16_t * gas = (const uint16_t *)(x[i].qs + QK_K/4) + 2*ib;
-    const uint8_t  * grid1 = (const uint8_t *)(iq3xxs_grid + q3[2*il+0]);
-    const uint8_t  * grid2 = (const uint8_t *)(iq3xxs_grid + q3[2*il+1]);
-    const uint32_t aux32 = gas[0] | (gas[1] << 16);
-    const float d = (float)x[i].d * (0.5f + (aux32 >> 28)) * 0.5f;
-    const uint8_t signs = ksigns_iq2xs[(aux32 >> 7*il) & 127];
-    for (int j = 0; j < 4; ++j) {
-        y[j+0] = ggml_cuda_cast<dst_t>(d * grid1[j] * (signs & kmask_iq2xs[j+0] ? -1.f : 1.f));
-        y[j+4] = ggml_cuda_cast<dst_t>(d * grid2[j] * (signs & kmask_iq2xs[j+4] ? -1.f : 1.f));
-    }
+    dequantize_iq3_xxs(vx, i, yy + i*QK_K, threadIdx.x);
 }
 
 template<typename dst_t>
 static __global__ void dequantize_block_iq3_s(const void * __restrict__ vx, dst_t * __restrict__ yy) {
+    const int64_t i = blockIdx.x;
 
-    const int64_t i   = blockIdx.x;
-    const block_iq3_s * x = (const block_iq3_s *) vx;
-
-    const int64_t tid = threadIdx.x;
-    const int64_t il = tid/8; // 0...3
-    const int64_t ib = tid%8; // 0...7
-    dst_t * y = yy + i*QK_K + 32*ib + 8*il;
-    const uint8_t * qs = x[i].qs + 8*ib;
-    const uint8_t * grid1 = (const uint8_t *)(iq3s_grid + (qs[2*il+0] | ((x[i].qh[ib] << (8-2*il)) & 256)));
-    const uint8_t * grid2 = (const uint8_t *)(iq3s_grid + (qs[2*il+1] | ((x[i].qh[ib] << (7-2*il)) & 256)));
-    const float d = (float)x[i].d * (1 + 2*((x[i].scales[ib/2] >> 4*(ib%2)) & 0xf));
-    const uint8_t signs = x[i].signs[4*ib + il];
-    for (int j = 0; j < 4; ++j) {
-        y[j+0] = ggml_cuda_cast<dst_t>(d * grid1[j] * (signs & kmask_iq2xs[j+0] ? -1.f : 1.f));
-        y[j+4] = ggml_cuda_cast<dst_t>(d * grid2[j] * (signs & kmask_iq2xs[j+4] ? -1.f : 1.f));
-    }
+    dequantize_iq3_s(vx, i, yy + i*QK_K, threadIdx.x);
 }
 
 template<typename dst_t>
 static __global__ void dequantize_block_iq1_s(const void * __restrict__ vx, dst_t * __restrict__ yy) {
+    const int64_t i = blockIdx.x;
 
-    const int64_t i   = blockIdx.x;
-    const block_iq1_s * x = (const block_iq1_s  *) vx;
-
-    const int64_t tid = threadIdx.x;
-    const int64_t il = tid/8; // 0...3
-    const int64_t ib = tid%8; // 0...7
-    dst_t * y = yy + i*QK_K + 32*ib + 8*il;
-    const float delta = x[i].qh[ib] & 0x8000 ? -1 - IQ1S_DELTA : -1 + IQ1S_DELTA;
-    const float d = (float)x[i].d * (2*((x[i].qh[ib] >> 12) & 7) + 1);
-    uint32_t grid32[2]; const int8_t * q = (const int8_t *)grid32;
-    grid32[0] = iq1s_grid_gpu[x[i].qs[4*ib+il] | (((x[i].qh[ib] >> 3*il) & 7) << 8)];
-    grid32[1] = (grid32[0] >> 4) & 0x0f0f0f0f;
-    grid32[0] &= 0x0f0f0f0f;
-    for (int j = 0; j < 8; ++j) {
-        y[j] = ggml_cuda_cast<dst_t>(d * (q[j] + delta));
-    }
+    dequantize_iq1_s(vx, i, yy + i*QK_K, threadIdx.x);
 }
 
 template<typename dst_t>
 static __global__ void dequantize_block_iq1_m(const void * __restrict__ vx, dst_t * __restrict__ yy) {
+    const int64_t i = blockIdx.x;
 
-    const int64_t i   = blockIdx.x;
-    const block_iq1_m * x = (const block_iq1_m  *) vx;
-
-    const int64_t tid = threadIdx.x;
-    const int64_t il = tid/8; // 0...3
-    const int64_t ib = tid%8; // 0...7
-    dst_t * y = yy + i*QK_K + 32*ib + 8*il;
-    const uint16_t * sc = (const uint16_t *)x[i].scales;
-    iq1m_scale_t scale;
-    scale.u16 = (sc[0] >> 12) | ((sc[1] >> 8) & 0x00f0) | ((sc[2] >> 4) & 0x0f00) | (sc[3] & 0xf000);
-    const int64_t ib16 = 2*ib + il/2; // sc[ib16/4] >> 3*(ib16%4) -> sc[ib/2] >> 3*((2*ib+il/2)%4);
-    const float d = (float)scale.f16 * (2*((sc[ib16/4] >> 3*(ib16%4)) & 0x7) + 1);
-    const float delta = x[i].qh[2*ib+il/2] & (0x08 << 4*(il%2)) ? -1 - IQ1M_DELTA : -1 + IQ1M_DELTA;
-    uint32_t grid32[2]; const int8_t * q = (const int8_t *)grid32;
-    grid32[0] = iq1s_grid_gpu[x[i].qs[4*ib+il] | (((x[i].qh[2*ib+il/2] >> 4*(il%2)) & 7) << 8)];
-    grid32[1] = (grid32[0] >> 4) & 0x0f0f0f0f;
-    grid32[0] &= 0x0f0f0f0f;
-    for (int j = 0; j < 8; ++j) {
-        y[j] = ggml_cuda_cast<dst_t>(d * (q[j] + delta));
-    }
+    dequantize_iq1_m(vx, i, yy + i*QK_K, threadIdx.x);
 }
 
 template<typename dst_t>
 static __global__ void dequantize_block_iq4_nl(const void * __restrict__ vx, dst_t * __restrict__ yy) {
+    const int64_t i = blockIdx.x;
 
-    const int64_t i   = blockIdx.x;
-    const block_iq4_nl * x = (const block_iq4_nl *) vx + i*(QK_K/QK4_NL);
-
-    const int64_t tid = threadIdx.x;
-    const int64_t il = tid/8; // 0...3
-    const int64_t ib = tid%8; // 0...7
-    dst_t * y = yy + i*QK_K + 32*ib + 4*il;
-    const uint8_t  * q4 = x[ib].qs + 4*il;
-    const float d = (float)x[ib].d;
-    for (int j = 0; j < 4; ++j) {
-        y[j+ 0] = ggml_cuda_cast<dst_t>(d * kvalues_iq4nl[q4[j] & 0xf]);
-        y[j+16] = ggml_cuda_cast<dst_t>(d * kvalues_iq4nl[q4[j] >>  4]);
-    }
+    dequantize_iq4_nl(vx, i, yy + i*QK_K, threadIdx.x);
 }
 
 template<typename dst_t>
 static __global__ void dequantize_block_iq4_xs(const void * __restrict__ vx, dst_t * __restrict__ yy) {
-    const int64_t i   = blockIdx.x;
-    const block_iq4_xs * x = (const block_iq4_xs *)vx;
+    const int64_t i = blockIdx.x;
 
-    const int64_t tid = threadIdx.x;
-    const int64_t il = tid/8; // 0...3
-    const int64_t ib = tid%8; // 0...7
-    dst_t * y = yy + i*QK_K + 32*ib + 4*il;
-    const uint8_t  * q4 = x[i].qs + 16*ib + 4*il;
-    const float d = (float)x[i].d * ((((x[i].scales_l[ib/2] >> 4*(ib%2)) & 0xf) | (((x[i].scales_h >> 2*ib) & 3) << 4)) - 32);
-    for (int j = 0; j < 4; ++j) {
-        y[j+ 0] = ggml_cuda_cast<dst_t>(d * kvalues_iq4nl[q4[j] & 0xf]);
-        y[j+16] = ggml_cuda_cast<dst_t>(d * kvalues_iq4nl[q4[j] >>  4]);
-    }
+    dequantize_iq4_xs(vx, i, yy + i*QK_K, threadIdx.x);
 }
 
 template<typename dst_t>
 static __global__ void dequantize_block_mxfp4(const void * __restrict__ vx, dst_t * __restrict__ yy) {
+    const int64_t i = blockIdx.x;
 
-    const int64_t i   = blockIdx.x;
-    const block_mxfp4 * x = (const block_mxfp4 *) vx + i*(QK_K/QK_MXFP4);
-
-    const int64_t tid = threadIdx.x;
-    const int64_t il = tid/8; // 0...3
-    const int64_t ib = tid%8; // 0...7
-    dst_t * y = yy + i*QK_K + 32*ib + 4*il;
-    const uint8_t  * q4 = x[ib].qs + 4*il;
-    const float d = ggml_cuda_e8m0_to_fp32(x[ib].e);
-    for (int j = 0; j < 4; ++j) {
-        y[j+ 0] = ggml_cuda_cast<dst_t>(d * kvalues_mxfp4[q4[j] & 0xf]*0.5f);
-        y[j+16] = ggml_cuda_cast<dst_t>(d * kvalues_mxfp4[q4[j] >>  4]*0.5f);
-    }
+    dequantize_mxfp4(vx, i, yy + i*QK_K, threadIdx.x);
 }
 
 template <int qk, int qr, dequantize_kernel_t dequantize_kernel, typename dst_t>

ggml/src/ggml-cuda/dequantize.cuh

@@ -1,4 +1,5 @@
 #include "common.cuh"
+#include "convert.cuh"
 
 static __device__ __forceinline__ void dequantize_q1_0(const void * vx, const int64_t ib, const int iqs, float2 & v){
     const block_q1_0 * x = (const block_q1_0 *) vx;
@@ -97,3 +98,335 @@ static __device__ __forceinline__ void dequantize_q8_0(const void * vx, const in
     v.x *= d;
     v.y *= d;
 }
+
+//================================== k-quants
+
+// Each call dequantizes one super-block of QK_K values into y using the
+// thread layout of the caller: 32 threads for q4_K, 64 threads otherwise.
+
+template<typename dst_t>
+static __device__ __forceinline__ void dequantize_q2_K(const void * vx, const int64_t ib, dst_t * yy, const int tid) {
+    const block_q2_K * x = (const block_q2_K *) vx;
+
+    const int64_t n   = tid/32;
+    const int64_t l   = tid - 32*n;
+    const int64_t is  = 8*n + l/16;
+
+    const uint8_t q = x[ib].qs[32*n + l];
+    dst_t * y = yy + 128*n;
+
+    float dall = __low2half(x[ib].dm);
+    float dmin = __high2half(x[ib].dm);
+    y[l+ 0] = ggml_cuda_cast<dst_t>(dall * (x[ib].scales[is+0] & 0xF) * ((q >> 0) & 3) - dmin * (x[ib].scales[is+0] >> 4));
+    y[l+32] = ggml_cuda_cast<dst_t>(dall * (x[ib].scales[is+2] & 0xF) * ((q >> 2) & 3) - dmin * (x[ib].scales[is+2] >> 4));
+    y[l+64] = ggml_cuda_cast<dst_t>(dall * (x[ib].scales[is+4] & 0xF) * ((q >> 4) & 3) - dmin * (x[ib].scales[is+4] >> 4));
+    y[l+96] = ggml_cuda_cast<dst_t>(dall * (x[ib].scales[is+6] & 0xF) * ((q >> 6) & 3) - dmin * (x[ib].scales[is+6] >> 4));
+}
+
+template<typename dst_t>
+static __device__ __forceinline__ void dequantize_q3_K(const void * vx, const int64_t ib, dst_t * yy, const int tid) {
+    const block_q3_K * x = (const block_q3_K *) vx;
+
+    const int64_t r = tid/4;
+    const int64_t t = r/2;
+    const int64_t is0 = r%2;
+    const int64_t l0 = 16*is0 + 4*(tid%4);
+    const int64_t n = t / 4;
+    const int64_t j = t - 4*n;
+
+    uint8_t m = 1 << (4*n + j);
+    int64_t is = 8*n + 2*j + is0;
+    int shift = 2*j;
+
+    int8_t us = is <  4 ? (x[ib].scales[is-0] & 0xF) | (((x[ib].scales[is+8] >> 0) & 3) << 4) :
+                is <  8 ? (x[ib].scales[is-0] & 0xF) | (((x[ib].scales[is+4] >> 2) & 3) << 4) :
+                is < 12 ? (x[ib].scales[is-8] >>  4) | (((x[ib].scales[is+0] >> 4) & 3) << 4) :
+                          (x[ib].scales[is-8] >>  4) | (((x[ib].scales[is-4] >> 6) & 3) << 4);
+    float d_all = x[ib].d;
+    float dl = d_all * (us - 32);
+
+    dst_t * y = yy + 128*n + 32*j;
+    const uint8_t * q = x[ib].qs + 32*n;
+    const uint8_t * hm = x[ib].hmask;
+
+    for (int l = l0; l < l0+4; ++l) {
+        y[l] = ggml_cuda_cast<dst_t>(dl * ((int8_t)((q[l] >> shift) & 3) - ((hm[l] & m) ? 0 : 4)));
+    }
+}
+
+static inline __device__ void get_scale_min_k4(int j, const uint8_t * q, uint8_t & d, uint8_t & m) {
+    if (j < 4) {
+        d = q[j] & 63; m = q[j + 4] & 63;
+    } else {
+        d = (q[j+4] & 0xF) | ((q[j-4] >> 6) << 4);
+        m = (q[j+4] >>  4) | ((q[j-0] >> 6) << 4);
+    }
+}
+
+template<typename dst_t>
+static __device__ __forceinline__ void dequantize_q4_K(const void * vx, const int64_t ib, dst_t * yy, const int tid) {
+    const block_q4_K * x = (const block_q4_K *) vx;
+
+    // assume 32 threads
+    const int64_t il  = tid/8;
+    const int64_t ir  = tid%8;
+    const int64_t is  = 2*il;
+    const int64_t n   = 4;
+
+    dst_t * y = yy + 64*il + n*ir;
+
+    const float dall = __low2half(x[ib].dm);
+    const float dmin = __high2half(x[ib].dm);
+
+    const uint8_t * q = x[ib].qs + 32*il + n*ir;
+
+    uint8_t sc, m;
+    get_scale_min_k4(is + 0, x[ib].scales, sc, m);
+    const float d1 = dall * sc; const float m1 = dmin * m;
+    get_scale_min_k4(is + 1, x[ib].scales, sc, m);
+    const float d2 = dall * sc; const float m2 = dmin * m;
+    for (int l = 0; l < n; ++l) {
+        y[l + 0] = ggml_cuda_cast<dst_t>(d1 * (q[l] & 0xF) - m1);
+        y[l +32] = ggml_cuda_cast<dst_t>(d2 * (q[l] >>  4) - m2);
+    }
+}
+
+template<typename dst_t>
+static __device__ __forceinline__ void dequantize_q5_K(const void * vx, const int64_t ib, dst_t * yy, const int tid) {
+    const block_q5_K * x = (const block_q5_K *) vx;
+
+    // assume 64 threads - this is very slightly better than the one below
+    const int64_t il  = tid/16;   // il is in 0...3
+    const int64_t ir  = tid%16;   // ir is in 0...15
+    const int64_t is  = 2*il;     // is is in 0...6
+
+    dst_t * y = yy + 64*il + 2*ir;
+
+    const float dall = __low2half(x[ib].dm);
+    const float dmin = __high2half(x[ib].dm);
+
+    const uint8_t * ql = x[ib].qs + 32*il + 2*ir;
+    const uint8_t * qh = x[ib].qh + 2*ir;
+
+    uint8_t sc, m;
+    get_scale_min_k4(is + 0, x[ib].scales, sc, m);
+    const float d1 = dall * sc; const float m1 = dmin * m;
+    get_scale_min_k4(is + 1, x[ib].scales, sc, m);
+    const float d2 = dall * sc; const float m2 = dmin * m;
+
+    uint8_t   hm  = 1 << (2*il);
+    y[ 0] = ggml_cuda_cast<dst_t>(d1 * ((ql[ 0] & 0xF) + (qh[ 0] & hm ? 16 : 0)) - m1);
+    y[ 1] = ggml_cuda_cast<dst_t>(d1 * ((ql[ 1] & 0xF) + (qh[ 1] & hm ? 16 : 0)) - m1);
+    hm <<= 1;
+    y[32] = ggml_cuda_cast<dst_t>(d2 * ((ql[ 0] >>  4) + (qh[ 0] & hm ? 16 : 0)) - m2);
+    y[33] = ggml_cuda_cast<dst_t>(d2 * ((ql[ 1] >>  4) + (qh[ 1] & hm ? 16 : 0)) - m2);
+}
+
+template<typename dst_t>
+static __device__ __forceinline__ void dequantize_q6_K(const void * vx, const int64_t ib, dst_t * yy, const int tid) {
+    const block_q6_K * x = (const block_q6_K *) vx;
+
+    // assume 64 threads - this is very slightly better than the one below
+    const int64_t ip  = tid/32;   // ip is 0 or 1
+    const int64_t il  = tid - 32*ip; // 0...32
+    const int64_t is  = 8*ip + il/16;
+
+    dst_t * y = yy + 128*ip + il;
+
+    const float d = x[ib].d;
+
+    const uint8_t * ql = x[ib].ql + 64*ip + il;
+    const uint8_t   qh = x[ib].qh[32*ip + il];
+    const int8_t  * sc = x[ib].scales + is;
+
+    y[ 0] = ggml_cuda_cast<dst_t>(d * sc[0] * ((int8_t)((ql[ 0] & 0xF) | (((qh >> 0) & 3) << 4)) - 32));
+    y[32] = ggml_cuda_cast<dst_t>(d * sc[2] * ((int8_t)((ql[32] & 0xF) | (((qh >> 2) & 3) << 4)) - 32));
+    y[64] = ggml_cuda_cast<dst_t>(d * sc[4] * ((int8_t)((ql[ 0]  >> 4) | (((qh >> 4) & 3) << 4)) - 32));
+    y[96] = ggml_cuda_cast<dst_t>(d * sc[6] * ((int8_t)((ql[32]  >> 4) | (((qh >> 6) & 3) << 4)) - 32));
+}
+
+//================================== i-quants
+
+// Each call dequantizes one super-block of QK_K values into y with 32
+// threads; iq4_nl packs QK_K/QK4_NL sub-blocks per super-block.
+
+template<typename dst_t>
+static __device__ __forceinline__ void dequantize_iq2_xxs(const void * vx, const int64_t ibs, dst_t * yy, const int tid) {
+
+    const block_iq2_xxs * x = (const block_iq2_xxs  *) vx;
+
+    const int64_t il = tid/8; // 0...3
+    const int64_t ib = tid%8; // 0...7
+    dst_t * y = yy + 32*ib + 8*il;
+    const uint16_t * q2 = x[ibs].qs + 4*ib;
+    const uint8_t  * aux8 = (const uint8_t *)q2;
+    const uint8_t  * grid = (const uint8_t *)(iq2xxs_grid + aux8[il]);
+    const uint32_t aux32 = q2[2] | (q2[3] << 16);
+    const float d = (float)x[ibs].d * (0.5f + (aux32 >> 28)) * 0.25f;
+    const uint8_t signs = ksigns_iq2xs[(aux32 >> 7*il) & 127];
+    for (int j = 0; j < 8; ++j) {
+        y[j] = ggml_cuda_cast<dst_t>(d * grid[j] * (signs & kmask_iq2xs[j] ? -1.f : 1.f));
+    }
+}
+
+template<typename dst_t>
+static __device__ __forceinline__ void dequantize_iq2_xs(const void * vx, const int64_t ibs, dst_t * yy, const int tid) {
+
+    const block_iq2_xs * x = (const block_iq2_xs *) vx;
+
+    const int64_t il = tid/8; // 0...3
+    const int64_t ib = tid%8; // 0...7
+    dst_t * y = yy + 32*ib + 8*il;
+    const uint16_t * q2 = x[ibs].qs + 4*ib;
+    const uint8_t  * grid = (const uint8_t *)(iq2xs_grid + (q2[il] & 511));
+    const float d = (float)x[ibs].d * (0.5f + ((x[ibs].scales[ib] >> 4*(il/2)) & 0xf)) * 0.25f;
+    const uint8_t signs = ksigns_iq2xs[q2[il] >> 9];
+    for (int j = 0; j < 8; ++j) {
+        y[j] = ggml_cuda_cast<dst_t>(d * grid[j] * (signs & kmask_iq2xs[j] ? -1.f : 1.f));
+    }
+}
+
+template<typename dst_t>
+static __device__ __forceinline__ void dequantize_iq2_s(const void * vx, const int64_t ibs, dst_t * yy, const int tid) {
+
+    const block_iq2_s * x = (const block_iq2_s *) vx;
+
+    const int64_t il = tid/8; // 0...3
+    const int64_t ib = tid%8; // 0...7
+    dst_t * y = yy + 32*ib + 8*il;
+    const uint8_t * grid = (const uint8_t *)(iq2s_grid + (x[ibs].qs[4*ib+il] | ((x[ibs].qh[ib] << (8-2*il)) & 0x300)));
+    const float d = (float)x[ibs].d * (0.5f + ((x[ibs].scales[ib] >> 4*(il/2)) & 0xf)) * 0.25f;
+    const uint8_t signs = x[ibs].qs[QK_K/8+4*ib+il];
+    for (int j = 0; j < 8; ++j) {
+        y[j] = ggml_cuda_cast<dst_t>(d * grid[j] * (signs & kmask_iq2xs[j] ? -1.f : 1.f));
+    }
+}
+
+template<typename dst_t>
+static __device__ __forceinline__ void dequantize_iq3_xxs(const void * vx, const int64_t ibs, dst_t * yy, const int tid) {
+
+    const block_iq3_xxs * x = (const block_iq3_xxs  *) vx;
+
+    const int64_t il = tid/8; // 0...3
+    const int64_t ib = tid%8; // 0...7
+    dst_t * y = yy + 32*ib + 8*il;
+    const uint8_t  * q3 = x[ibs].qs + 8*ib;
+    const uint16_t * gas = (const uint16_t *)(x[ibs].qs + QK_K/4) + 2*ib;
+    const uint8_t  * grid1 = (const uint8_t *)(iq3xxs_grid + q3[2*il+0]);
+    const uint8_t  * grid2 = (const uint8_t *)(iq3xxs_grid + q3[2*il+1]);
+    const uint32_t aux32 = gas[0] | (gas[1] << 16);
+    const float d = (float)x[ibs].d * (0.5f + (aux32 >> 28)) * 0.5f;
+    const uint8_t signs = ksigns_iq2xs[(aux32 >> 7*il) & 127];
+    for (int j = 0; j < 4; ++j) {
+        y[j+0] = ggml_cuda_cast<dst_t>(d * grid1[j] * (signs & kmask_iq2xs[j+0] ? -1.f : 1.f));
+        y[j+4] = ggml_cuda_cast<dst_t>(d * grid2[j] * (signs & kmask_iq2xs[j+4] ? -1.f : 1.f));
+    }
+}
+
+template<typename dst_t>
+static __device__ __forceinline__ void dequantize_iq3_s(const void * vx, const int64_t ibs, dst_t * yy, const int tid) {
+
+    const block_iq3_s * x = (const block_iq3_s *) vx;
+
+    const int64_t il = tid/8; // 0...3
+    const int64_t ib = tid%8; // 0...7
+    dst_t * y = yy + 32*ib + 8*il;
+    const uint8_t * qs = x[ibs].qs + 8*ib;
+    const uint8_t * grid1 = (const uint8_t *)(iq3s_grid + (qs[2*il+0] | ((x[ibs].qh[ib] << (8-2*il)) & 256)));
+    const uint8_t * grid2 = (const uint8_t *)(iq3s_grid + (qs[2*il+1] | ((x[ibs].qh[ib] << (7-2*il)) & 256)));
+    const float d = (float)x[ibs].d * (1 + 2*((x[ibs].scales[ib/2] >> 4*(ib%2)) & 0xf));
+    const uint8_t signs = x[ibs].signs[4*ib + il];
+    for (int j = 0; j < 4; ++j) {
+        y[j+0] = ggml_cuda_cast<dst_t>(d * grid1[j] * (signs & kmask_iq2xs[j+0] ? -1.f : 1.f));
+        y[j+4] = ggml_cuda_cast<dst_t>(d * grid2[j] * (signs & kmask_iq2xs[j+4] ? -1.f : 1.f));
+    }
+}
+
+template<typename dst_t>
+static __device__ __forceinline__ void dequantize_iq1_s(const void * vx, const int64_t ibs, dst_t * yy, const int tid) {
+
+    const block_iq1_s * x = (const block_iq1_s  *) vx;
+
+    const int64_t il = tid/8; // 0...3
+    const int64_t ib = tid%8; // 0...7
+    dst_t * y = yy + 32*ib + 8*il;
+    const float delta = x[ibs].qh[ib] & 0x8000 ? -1 - IQ1S_DELTA : -1 + IQ1S_DELTA;
+    const float d = (float)x[ibs].d * (2*((x[ibs].qh[ib] >> 12) & 7) + 1);
+    uint32_t grid32[2]; const int8_t * q = (const int8_t *)grid32;
+    grid32[0] = iq1s_grid_gpu[x[ibs].qs[4*ib+il] | (((x[ibs].qh[ib] >> 3*il) & 7) << 8)];
+    grid32[1] = (grid32[0] >> 4) & 0x0f0f0f0f;
+    grid32[0] &= 0x0f0f0f0f;
+    for (int j = 0; j < 8; ++j) {
+        y[j] = ggml_cuda_cast<dst_t>(d * (q[j] + delta));
+    }
+}
+
+template<typename dst_t>
+static __device__ __forceinline__ void dequantize_iq1_m(const void * vx, const int64_t ibs, dst_t * yy, const int tid) {
+
+    const block_iq1_m * x = (const block_iq1_m  *) vx;
+
+    const int64_t il = tid/8; // 0...3
+    const int64_t ib = tid%8; // 0...7
+    dst_t * y = yy + 32*ib + 8*il;
+    const uint16_t * sc = (const uint16_t *)x[ibs].scales;
+    iq1m_scale_t scale;
+    scale.u16 = (sc[0] >> 12) | ((sc[1] >> 8) & 0x00f0) | ((sc[2] >> 4) & 0x0f00) | (sc[3] & 0xf000);
+    const int64_t ib16 = 2*ib + il/2; // sc[ib16/4] >> 3*(ib16%4) -> sc[ib/2] >> 3*((2*ib+il/2)%4);
+    const float d = (float)scale.f16 * (2*((sc[ib16/4] >> 3*(ib16%4)) & 0x7) + 1);
+    const float delta = x[ibs].qh[2*ib+il/2] & (0x08 << 4*(il%2)) ? -1 - IQ1M_DELTA : -1 + IQ1M_DELTA;
+    uint32_t grid32[2]; const int8_t * q = (const int8_t *)grid32;
+    grid32[0] = iq1s_grid_gpu[x[ibs].qs[4*ib+il] | (((x[ibs].qh[2*ib+il/2] >> 4*(il%2)) & 7) << 8)];
+    grid32[1] = (grid32[0] >> 4) & 0x0f0f0f0f;
+    grid32[0] &= 0x0f0f0f0f;
+    for (int j = 0; j < 8; ++j) {
+        y[j] = ggml_cuda_cast<dst_t>(d * (q[j] + delta));
+    }
+}
+
+template<typename dst_t>
+static __device__ __forceinline__ void dequantize_iq4_nl(const void * vx, const int64_t ibs, dst_t * yy, const int tid) {
+
+    const block_iq4_nl * x = (const block_iq4_nl *) vx + ibs*(QK_K/QK4_NL);
+
+    const int64_t il = tid/8; // 0...3
+    const int64_t ib = tid%8; // 0...7
+    dst_t * y = yy + 32*ib + 4*il;
+    const uint8_t  * q4 = x[ib].qs + 4*il;
+    const float d = (float)x[ib].d;
+    for (int j = 0; j < 4; ++j) {
+        y[j+ 0] = ggml_cuda_cast<dst_t>(d * kvalues_iq4nl[q4[j] & 0xf]);
+        y[j+16] = ggml_cuda_cast<dst_t>(d * kvalues_iq4nl[q4[j] >>  4]);
+    }
+}
+
+template<typename dst_t>
+static __device__ __forceinline__ void dequantize_iq4_xs(const void * vx, const int64_t ibs, dst_t * yy, const int tid) {
+    const block_iq4_xs * x = (const block_iq4_xs *)vx;
+
+    const int64_t il = tid/8; // 0...3
+    const int64_t ib = tid%8; // 0...7
+    dst_t * y = yy + 32*ib + 4*il;
+    const uint8_t  * q4 = x[ibs].qs + 16*ib + 4*il;
+    const float d = (float)x[ibs].d * ((((x[ibs].scales_l[ib/2] >> 4*(ib%2)) & 0xf) | (((x[ibs].scales_h >> 2*ib) & 3) << 4)) - 32);
+    for (int j = 0; j < 4; ++j) {
+        y[j+ 0] = ggml_cuda_cast<dst_t>(d * kvalues_iq4nl[q4[j] & 0xf]);
+        y[j+16] = ggml_cuda_cast<dst_t>(d * kvalues_iq4nl[q4[j] >>  4]);
+    }
+}
+
+template<typename dst_t>
+static __device__ __forceinline__ void dequantize_mxfp4(const void * vx, const int64_t ibs, dst_t * yy, const int tid) {
+
+    const block_mxfp4 * x = (const block_mxfp4 *) vx + ibs*(QK_K/QK_MXFP4);
+
+    const int64_t il = tid/8; // 0...3
+    const int64_t ib = tid%8; // 0...7
+    dst_t * y = yy + 32*ib + 4*il;
+    const uint8_t  * q4 = x[ib].qs + 4*il;
+    const float d = ggml_cuda_e8m0_to_fp32(x[ib].e);
+    for (int j = 0; j < 4; ++j) {
+        y[j+ 0] = ggml_cuda_cast<dst_t>(d * kvalues_mxfp4[q4[j] & 0xf]*0.5f);
+        y[j+16] = ggml_cuda_cast<dst_t>(d * kvalues_mxfp4[q4[j] >>  4]*0.5f);
+    }
+}

ggml/src/ggml-cuda/getrows.cu

@@ -40,6 +40,35 @@ static __global__ void k_get_rows(
     }
 }
 
+template<typename dst_t, dequantize_kq_t<dst_t> dequantize_kq>
+static __global__ void k_get_rows_kq(
+        const void * __restrict__ src0, const int32_t * __restrict__ src1, dst_t * __restrict__ dst,
+        const int64_t ne00, /*const int64_t ne01, const int64_t ne02, const int64_t ne03,*/
+        /*const int64_t ne10,*/ const int64_t ne11, const uint3 ne12_fdv, /*const int64_t ne13,*/
+        /*const size_t s0,*/ const size_t s1, const size_t s2, const size_t s3,
+        /*const size_t nb00,*/ const size_t nb01, const size_t nb02, const size_t nb03,
+        const size_t s10, const size_t s11, const size_t s12/*, const size_t s13*/) {
+
+    ggml_cuda_pdl_sync();
+    const int64_t nsb = ne00/QK_K; // super-blocks per row
+    for (int64_t z = blockIdx.z; z < ne11*(int64_t)ne12_fdv.z; z += gridDim.z) {
+        // The x and y dimensions of the grid are swapped because the maximum allowed grid size for x is higher.
+        const int i10 = blockIdx.x;
+        const uint2 dm  = fast_div_modulo((uint32_t)z, ne12_fdv);
+        const int i11 = dm.x;
+        const int i12 = dm.y;
+
+        const int i01 = src1[i10*s10 + i11*s11 + i12*s12];
+
+        dst_t * dst_row = dst + i10*s1 + i11*s2 + i12*s3;
+        const void * src0_row = (const char *) src0 + i01*nb01 + i11*nb02 + i12*nb03;
+
+        for (int64_t ib = blockIdx.y; ib < nsb; ib += gridDim.y) {
+            dequantize_kq(src0_row, ib, dst_row + ib*QK_K, threadIdx.x);
+        }
+    }
+}
+
 template<typename src0_t, typename dst_t>
 static __global__ void k_get_rows_float(
         const src0_t * src0_ptr, const int32_t * src1_ptr, dst_t * dst_ptr,
@@ -55,27 +84,51 @@ static __global__ void k_get_rows_float(
     dst_t         * GGML_CUDA_RESTRICT dst  = dst_ptr;
     ggml_cuda_pdl_sync();
     for (int64_t z = blockIdx.z; z < ne11*(int64_t)ne12_fdv.z; z += gridDim.z) {
+        // The x and y dimensions of the grid are swapped because the maximum allowed grid size for x is higher.
+        const int i10 = blockIdx.x;
+        const uint2 dm = fast_div_modulo((uint32_t)z, ne12_fdv);
+        const int i11 = dm.x;
+        const int i12 = dm.y;
+
+        const int i01 = src1[i10*s10 + i11*s11 + i12*s12];
+
+        dst_t * GGML_CUDA_RESTRICT dst_row = dst + i10*s1 + i11*s2 + i12*s3;
+        const src0_t * GGML_CUDA_RESTRICT src0_row = (const src0_t *)((const char *) src0 + i01*nb01 + i11*nb02 + i12*nb03);
+
         for (int64_t i00 = blockIdx.y*blockDim.x + threadIdx.x; i00 < ne00; i00 += gridDim.y*blockDim.x) {
-            // The x and y dimensions of the grid are swapped because the maximum allowed grid size for x is higher.
-            const int i10 = blockIdx.x;
-            const uint2 dm = fast_div_modulo((uint32_t)z, ne12_fdv);
-            const int i11 = dm.x;
-            const int i12 = dm.y;
-
-            if (i00 >= ne00) {
-                return;
-            }
-
-            const int i01 = src1[i10*s10 + i11*s11 + i12*s12];
-
-            dst_t * dst_row = dst + i10*s1 + i11*s2 + i12*s3;
-            const src0_t * src0_row = (const src0_t *)((const char *) src0 + i01*nb01 + i11*nb02 + i12*nb03);
-
             dst_row[i00] = ggml_cuda_cast<dst_t>(src0_row[i00]);
         }
     }
 }
 
+template<typename dst_t>
+static __global__ void k_get_rows_float_vec(
+        const dst_t * src0_ptr, const int32_t * src1_ptr, dst_t * dst_ptr,
+        const int64_t ne00v,
+        const int64_t ne11, const uint3 ne12_fdv,
+        const size_t s1, const size_t s2, const size_t s3,
+        const size_t nb01, const size_t nb02, const size_t nb03,
+        const size_t s10, const size_t s11, const size_t s12) {
+
+    ggml_cuda_pdl_lc();
+    ggml_cuda_pdl_sync();
+    for (int64_t z = blockIdx.z; z < ne11*(int64_t)ne12_fdv.z; z += gridDim.z) {
+        const int i10 = blockIdx.x;
+        const uint2 dm = fast_div_modulo((uint32_t)z, ne12_fdv);
+        const int i11 = dm.x;
+        const int i12 = dm.y;
+
+        const int i01 = src1_ptr[i10*s10 + i11*s11 + i12*s12];
+
+        int4       * GGML_CUDA_RESTRICT dst_row  = (int4 *)      (dst_ptr + i10*s1 + i11*s2 + i12*s3);
+        const int4 * GGML_CUDA_RESTRICT src0_row = (const int4 *)((const char *) src0_ptr + i01*nb01 + i11*nb02 + i12*nb03);
+
+        for (int64_t i = blockIdx.y*blockDim.x + threadIdx.x; i < ne00v; i += gridDim.y*blockDim.x) {
+            dst_row[i] = src0_row[i];
+        }
+    }
+}
+
 template<typename grad_t, typename dst_t>
 static __global__ void k_get_rows_back_float(
         const grad_t * __restrict__ grad, const int32_t * __restrict__ rows, dst_t * __restrict__ dst,
@@ -140,16 +193,18 @@ static void get_rows_cuda_q(
         s10, s11, s12/*, s13*/);
 }
 
-template<typename src0_t, typename dst_t>
-static void get_rows_cuda_float(
-        const src0_t * src0_d, const int32_t * src1_d, dst_t * dst_d,
+template<int block_dim, typename dst_t, dequantize_kq_t<dst_t> dequantize_kq>
+static void get_rows_cuda_kq(
+        const void * src0_d, const int32_t * src1_d, dst_t * dst_d,
         const int64_t ne00, const size_t nb01, const size_t nb02, const size_t nb03,
         const int64_t ne10, const int64_t ne11, const int64_t ne12, const size_t nb10, const size_t nb11, const size_t nb12,
         const size_t nb1, const size_t nb2, const size_t nb3,
         cudaStream_t stream) {
-    const dim3 block_dims(CUDA_GET_ROWS_BLOCK_SIZE, 1, 1);
-    const int block_num_y = (ne00 + CUDA_GET_ROWS_BLOCK_SIZE - 1) / CUDA_GET_ROWS_BLOCK_SIZE;
-    const dim3 block_nums(ne10, MIN(block_num_y, UINT16_MAX), MIN(ne11*ne12, UINT16_MAX));
+    GGML_ASSERT(ne00 % QK_K == 0);
+    const int64_t nsb = ne00/QK_K;
+
+    const dim3 block_dims(block_dim, 1, 1);
+    const dim3 block_nums(ne10, MIN(nsb, UINT16_MAX), MIN(ne11*ne12, UINT16_MAX));
 
     // strides in elements
     // const size_t s0 = nb0 / sizeof(dst_t);
@@ -166,6 +221,67 @@ static void get_rows_cuda_float(
     GGML_ASSERT(ne11 <= std::numeric_limits<uint32_t>::max() / ne12);
     const uint3 ne12_fdv = init_fastdiv_values(ne12);
 
+    k_get_rows_kq<dst_t, dequantize_kq><<<block_nums, block_dims, 0, stream>>>(
+        src0_d, src1_d, dst_d,
+        ne00, /*ne01, ne02, ne03,*/
+        /*ne10,*/ ne11, ne12_fdv, /*ne13,*/
+        /* s0,*/ s1, s2, s3,
+        /* nb00,*/ nb01, nb02, nb03,
+        s10, s11, s12/*, s13*/);
+}
+
+template<typename src0_t, typename dst_t>
+static void get_rows_cuda_float(
+        const src0_t * src0_d, const int32_t * src1_d, dst_t * dst_d,
+        const int64_t ne00, const size_t nb01, const size_t nb02, const size_t nb03,
+        const int64_t ne10, const int64_t ne11, const int64_t ne12, const size_t nb10, const size_t nb11, const size_t nb12,
+        const size_t nb1, const size_t nb2, const size_t nb3,
+        cudaStream_t stream) {
+    const dim3 block_dims(CUDA_GET_ROWS_BLOCK_SIZE, 1, 1);
+
+    // strides in elements
+    // const size_t s0 = nb0 / sizeof(dst_t);
+    const size_t s1 = nb1 / sizeof(dst_t);
+    const size_t s2 = nb2 / sizeof(dst_t);
+    const size_t s3 = nb3 / sizeof(dst_t);
+
+    const size_t s10 = nb10 / sizeof(int32_t);
+    const size_t s11 = nb11 / sizeof(int32_t);
+    const size_t s12 = nb12 / sizeof(int32_t);
+    // const size_t s13 = nb13 / sizeof(int32_t);
+
+    GGML_ASSERT(ne12 > 0);
+    GGML_ASSERT(ne11 <= std::numeric_limits<uint32_t>::max() / ne12);
+    const uint3 ne12_fdv = init_fastdiv_values(ne12);
+
+    if constexpr (std::is_same<src0_t, dst_t>::value) {
+        constexpr int VEC = 16 / sizeof(dst_t);
+        const int64_t ne00v = ne00 / VEC;
+        const int64_t vec_block_num_y = (ne00v + CUDA_GET_ROWS_BLOCK_SIZE - 1) / CUDA_GET_ROWS_BLOCK_SIZE;
+        const bool enough_blocks = vec_block_num_y * ne10 * ne11 * ne12 >= 128;
+        const bool can_vec = VEC > 1 && enough_blocks &&
+            (ne00 % VEC == 0) &&
+            (nb01 % 16 == 0) && (nb02 % 16 == 0) && (nb03 % 16 == 0) &&
+            (nb1  % 16 == 0) && (nb2  % 16 == 0) && (nb3  % 16 == 0) &&
+            (((uintptr_t) src0_d) % 16 == 0) && (((uintptr_t) dst_d) % 16 == 0);
+
+        if (can_vec) {
+            const int block_num_y = vec_block_num_y;
+            const dim3 block_nums(ne10, MIN(block_num_y, UINT16_MAX), MIN(ne11*ne12, UINT16_MAX));
+            const ggml_cuda_kernel_launch_params launch_params = ggml_cuda_kernel_launch_params{block_nums, block_dims, 0, stream};
+            ggml_cuda_kernel_launch(k_get_rows_float_vec<dst_t>, launch_params,
+                (const dst_t *) src0_d, src1_d, dst_d,
+                ne00v, ne11, ne12_fdv,
+                s1, s2, s3,
+                nb01, nb02, nb03,
+                s10, s11, s12);
+            return;
+        }
+    }
+
+    const int block_num_y = (ne00 + CUDA_GET_ROWS_BLOCK_SIZE - 1) / CUDA_GET_ROWS_BLOCK_SIZE;
+    const dim3 block_nums(ne10, MIN(block_num_y, UINT16_MAX), MIN(ne11*ne12, UINT16_MAX));
+
     const ggml_cuda_kernel_launch_params launch_params = ggml_cuda_kernel_launch_params{block_nums, block_dims, 0, stream};
     ggml_cuda_kernel_launch(k_get_rows_float<src0_t, dst_t>, launch_params,
         src0_d, src1_d, dst_d,
@@ -224,8 +340,67 @@ static void ggml_cuda_get_rows_switch_src0_type(
             get_rows_cuda_q<QK8_0, QR8_0, dequantize_q8_0>(src0_d, src1_d, dst_d,
                 ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream);
             break;
+        case GGML_TYPE_Q2_K:
+            get_rows_cuda_kq<64, dst_t, dequantize_q2_K<dst_t>>(src0_d, src1_d, dst_d,
+                ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream);
+            break;
+        case GGML_TYPE_Q3_K:
+            get_rows_cuda_kq<64, dst_t, dequantize_q3_K<dst_t>>(src0_d, src1_d, dst_d,
+                ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream);
+            break;
+        case GGML_TYPE_Q4_K:
+            get_rows_cuda_kq<32, dst_t, dequantize_q4_K<dst_t>>(src0_d, src1_d, dst_d,
+                ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream);
+            break;
+        case GGML_TYPE_Q5_K:
+            get_rows_cuda_kq<64, dst_t, dequantize_q5_K<dst_t>>(src0_d, src1_d, dst_d,
+                ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream);
+            break;
+        case GGML_TYPE_Q6_K:
+            get_rows_cuda_kq<64, dst_t, dequantize_q6_K<dst_t>>(src0_d, src1_d, dst_d,
+                ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream);
+            break;
+        case GGML_TYPE_IQ2_XXS:
+            get_rows_cuda_kq<32, dst_t, dequantize_iq2_xxs<dst_t>>(src0_d, src1_d, dst_d,
+                ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream);
+            break;
+        case GGML_TYPE_IQ2_XS:
+            get_rows_cuda_kq<32, dst_t, dequantize_iq2_xs<dst_t>>(src0_d, src1_d, dst_d,
+                ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream);
+            break;
+        case GGML_TYPE_IQ2_S:
+            get_rows_cuda_kq<32, dst_t, dequantize_iq2_s<dst_t>>(src0_d, src1_d, dst_d,
+                ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream);
+            break;
+        case GGML_TYPE_IQ3_XXS:
+            get_rows_cuda_kq<32, dst_t, dequantize_iq3_xxs<dst_t>>(src0_d, src1_d, dst_d,
+                ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream);
+            break;
+        case GGML_TYPE_IQ3_S:
+            get_rows_cuda_kq<32, dst_t, dequantize_iq3_s<dst_t>>(src0_d, src1_d, dst_d,
+                ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream);
+            break;
+        case GGML_TYPE_IQ1_S:
+            get_rows_cuda_kq<32, dst_t, dequantize_iq1_s<dst_t>>(src0_d, src1_d, dst_d,
+                ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream);
+            break;
+        case GGML_TYPE_IQ1_M:
+            get_rows_cuda_kq<32, dst_t, dequantize_iq1_m<dst_t>>(src0_d, src1_d, dst_d,
+                ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream);
+            break;
+        case GGML_TYPE_IQ4_NL:
+            get_rows_cuda_kq<32, dst_t, dequantize_iq4_nl<dst_t>>(src0_d, src1_d, dst_d,
+                ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream);
+            break;
+        case GGML_TYPE_IQ4_XS:
+            get_rows_cuda_kq<32, dst_t, dequantize_iq4_xs<dst_t>>(src0_d, src1_d, dst_d,
+                ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream);
+            break;
+        case GGML_TYPE_MXFP4:
+            get_rows_cuda_kq<32, dst_t, dequantize_mxfp4<dst_t>>(src0_d, src1_d, dst_d,
+                ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream);
+            break;
         default:
-            // TODO: k-quants
             GGML_ABORT("%s: unsupported src0 type: %s\n", __func__, ggml_type_name(src0_type));
             break;
     }

ggml/src/ggml-cuda/ggml-cuda.cu

@@ -2703,6 +2703,7 @@ static int ggml_cuda_try_gdn_cache_fusion(
 
 static bool ggml_cuda_topk_moe_fusion(const struct ggml_cgraph * cgraph, int node_idx, ggml_cuda_topk_moe_args & args) {
     args.sigmoid         = false;
+    args.sqrt_softplus   = false;
     args.softmax         = false;
     args.delayed_softmax = false;
     args.prob_bias       = false;
@@ -2716,10 +2717,17 @@ static bool ggml_cuda_topk_moe_fusion(const struct ggml_cgraph * cgraph, int nod
     }
 
     if (nodes[node_idx]->op == GGML_OP_UNARY) {
-        if (ggml_get_unary_op(nodes[node_idx]) != GGML_UNARY_OP_SIGMOID) {
+        const ggml_unary_op unary_op = ggml_get_unary_op(nodes[node_idx]);
+        if (unary_op == GGML_UNARY_OP_SIGMOID) {
+            args.sigmoid = true;
+        } else if (unary_op == GGML_UNARY_OP_SOFTPLUS && node_idx + 1 < n_nodes &&
+                   nodes[node_idx + 1]->op == GGML_OP_SQRT && nodes[node_idx + 1]->src[0] == nodes[node_idx]) {
+            // sqrt(softplus(x)) scoring (DeepSeek-V4)
+            args.sqrt_softplus = true;
+            node_idx++;
+        } else {
             return false;
         }
-        args.sigmoid = true;
     }
 
     if (nodes[node_idx]->op == GGML_OP_ARGSORT) {
@@ -2728,7 +2736,7 @@ static bool ggml_cuda_topk_moe_fusion(const struct ggml_cgraph * cgraph, int nod
 
     node_idx++;
 
-    if (args.sigmoid || args.softmax) {
+    if (args.sigmoid || args.sqrt_softplus || args.softmax) {
         // SOFTMAX -> RESHAPE
         if (node_idx >= n_nodes || nodes[node_idx]->op != GGML_OP_RESHAPE ||
                 nodes[node_idx]->src[0] != nodes[node_idx - 1]) {
@@ -3172,21 +3180,27 @@ static int ggml_cuda_try_fuse(ggml_backend_cuda_context * cuda_ctx, ggml_cgraph
             const ggml_tensor * scale   = nullptr;
 
             if (!args.delayed_softmax) {
-                ggml_op gating_op = args.sigmoid ? GGML_OP_UNARY : GGML_OP_SOFT_MAX;
-                int     out_nodes[2];  // nodes which can't be elided
+                int out_nodes[2];  // nodes which can't be elided
+
+                if (args.sigmoid) {
+                    ops.insert(ops.end(), { GGML_OP_UNARY });
+                } else if (args.sqrt_softplus) {
+                    ops.insert(ops.end(), { GGML_OP_UNARY, GGML_OP_SQRT });
+                } else {
+                    ops.insert(ops.end(), { GGML_OP_SOFT_MAX });
+                }
+                const int i_probs = i + (int) ops.size() - 1;  // last node of the gating activation
 
                 if (args.prob_bias) {
-                    bias = cgraph->nodes[i + 2]->src[1];
-                    ops.insert(ops.end(), { gating_op, GGML_OP_RESHAPE, GGML_OP_ADD, GGML_OP_ARGSORT, GGML_OP_VIEW,
+                    bias = cgraph->nodes[i_probs + 2]->src[1];
+                    ops.insert(ops.end(), { GGML_OP_RESHAPE, GGML_OP_ADD, GGML_OP_ARGSORT, GGML_OP_VIEW,
                                             GGML_OP_GET_ROWS });
-                    out_nodes[0] = i + 4;
-                    ids          = cgraph->nodes[i + 4];
+                    out_nodes[0] = i_probs + 4;
                 } else {
-                    ops.insert(ops.end(),
-                               { gating_op, GGML_OP_RESHAPE, GGML_OP_ARGSORT, GGML_OP_VIEW, GGML_OP_GET_ROWS });
-                    out_nodes[0] = i + 3;
-                    ids          = cgraph->nodes[i + 3];
+                    ops.insert(ops.end(), { GGML_OP_RESHAPE, GGML_OP_ARGSORT, GGML_OP_VIEW, GGML_OP_GET_ROWS });
+                    out_nodes[0] = i_probs + 3;
                 }
+                ids = cgraph->nodes[out_nodes[0]];
 
                 if (args.norm) {
                     ops.insert(ops.end(),
@@ -4831,7 +4845,25 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
                     case GGML_TYPE_Q5_0:
                     case GGML_TYPE_Q5_1:
                     case GGML_TYPE_Q8_0:
+                    case GGML_TYPE_Q2_K:
+                    case GGML_TYPE_Q3_K:
+                    case GGML_TYPE_Q4_K:
+                    case GGML_TYPE_Q5_K:
+                    case GGML_TYPE_Q6_K:
+                    case GGML_TYPE_IQ2_XXS:
+                    case GGML_TYPE_IQ2_XS:
+                    case GGML_TYPE_IQ2_S:
+                    case GGML_TYPE_IQ3_XXS:
+                    case GGML_TYPE_IQ3_S:
+                    case GGML_TYPE_IQ1_S:
+                    case GGML_TYPE_IQ1_M:
+                    case GGML_TYPE_IQ4_XS:
                         return true;
+                    case GGML_TYPE_IQ4_NL:
+                    case GGML_TYPE_MXFP4:
+                        // 32-value sub-blocks, the row size does not guarantee
+                        // the QK_K super-blocks the get_rows kernel iterates on
+                        return op->src[0]->ne[0] % QK_K == 0;
                     default:
                         return false;
                 }

ggml/src/ggml-cuda/topk-moe.cu

@@ -8,6 +8,7 @@
 // Kernel config struct - passed by value to CUDA kernel
 struct topk_moe_config {
     bool use_sigmoid;
+    bool use_sqrt_softplus;
     bool with_norm;
     bool delayed_softmax;
 };
@@ -67,6 +68,16 @@ __device__ void sigmoid_warp_inplace(float (&vals)[experts_per_thread], const in
     }
 }
 
+template <int experts_per_thread, bool use_limit>
+__device__ void sqrt_softplus_warp_inplace(float (&vals)[experts_per_thread], const int limit, const int lane) {
+#pragma unroll
+    for (int i = 0; i < experts_per_thread; i++) {
+        const int  idx    = lane + i * WARP_SIZE;
+        const bool active = !use_limit || (idx < limit);
+        vals[i]           = active ? sqrtf(vals[i] > 20.0f ? vals[i] : logf(1.0f + expf(vals[i]))) : -INFINITY;
+    }
+}
+
 /*
     This kernel does the following:
     1. optionally softmax over the logits per token [n_experts, n_tokens]
@@ -115,6 +126,8 @@ __launch_bounds__(4 * WARP_SIZE, 1) __global__ void topk_moe_cuda(const float *
     if (!config.delayed_softmax) {
         if (config.use_sigmoid) {
            sigmoid_warp_inplace<experts_per_thread, false>(wt, n_experts, threadIdx.x);
+        } else if (config.use_sqrt_softplus) {
+           sqrt_softplus_warp_inplace<experts_per_thread, false>(wt, n_experts, threadIdx.x);
         } else {
            softmax_warp_inplace<experts_per_thread, false>(wt, n_experts, threadIdx.x);
         }
@@ -364,9 +377,10 @@ void ggml_cuda_op_topk_moe(ggml_backend_cuda_context &     ctx,
     }
 
     topk_moe_config config;
-    config.use_sigmoid     = args.sigmoid;
-    config.with_norm       = with_norm;
-    config.delayed_softmax = args.delayed_softmax;
+    config.use_sigmoid       = args.sigmoid;
+    config.use_sqrt_softplus = args.sqrt_softplus;
+    config.with_norm         = with_norm;
+    config.delayed_softmax   = args.delayed_softmax;
 
     if (bias) {
         launch_topk_moe_cuda<true>(ctx, logits_d, weights_d, ids_d, bias_d, n_rows, n_experts, n_expert_used, clamp_val,
@@ -415,7 +429,7 @@ bool ggml_cuda_should_use_topk_moe(const ggml_tensor * gating_op,
     } else if (gating_op->op == GGML_OP_UNARY) {
         ggml_unary_op op = ggml_get_unary_op(gating_op);
 
-        if (op != GGML_UNARY_OP_SIGMOID) {
+        if (op != GGML_UNARY_OP_SIGMOID && op != GGML_UNARY_OP_SOFTPLUS) {
             return false;
         }
     }

ggml/src/ggml-cuda/topk-moe.cuh

@@ -5,6 +5,7 @@
 
 struct ggml_cuda_topk_moe_args {
     bool sigmoid{};
+    bool sqrt_softplus{};
     bool softmax{};
     bool delayed_softmax{};
     bool prob_bias{};

ggml/src/ggml-hexagon/ggml-hexagon.cpp

@@ -1286,7 +1286,8 @@ struct ggml_hexagon_opbatch {
         int64_t nb2 = is_repack ? nb1 * ne1 : t->nb[2];
         int64_t nb3 = is_repack ? nb2 * t->ne[2] : t->nb[3];
 
-        return (h->ne[0] == ne0) && (h->ne[1] == ne1) && (h->ne[2] == t->ne[2]) && (h->ne[3] == t->ne[3]) &&
+        return (h->type == t->type) &&
+               (h->ne[0] == ne0) && (h->ne[1] == ne1) && (h->ne[2] == t->ne[2]) && (h->ne[3] == t->ne[3]) &&
                (h->nb[0] == t->nb[0]) && (h->nb[1] == nb1) && (h->nb[2] == nb2) && (h->nb[3] == nb3);
     }
 
@@ -3476,6 +3477,7 @@ static htp_op_code op_remap_to_htp(const ggml_tensor * t) {
         case GGML_OP_RMS_NORM:        return HTP_OP_RMS_NORM;
         case GGML_OP_CONCAT:          return HTP_OP_CONCAT;
         case GGML_OP_SCALE:           return HTP_OP_SCALE;
+        case GGML_OP_CLAMP:           return HTP_OP_CLAMP;
         case GGML_OP_SQR:             return HTP_OP_SQR;
         case GGML_OP_SQRT:            return HTP_OP_SQRT;
         case GGML_OP_SOFT_MAX:        return HTP_OP_SOFTMAX;
@@ -4126,6 +4128,7 @@ static bool ggml_backend_hexagon_device_supports_op(ggml_backend_dev_t dev, cons
         case GGML_OP_L2_NORM:
         case GGML_OP_RMS_NORM:
         case GGML_OP_SCALE:
+        case GGML_OP_CLAMP:
             supp = ggml_hexagon_supported_unary(sess, op);
             break;
 

ggml/src/ggml-hexagon/htp/htp-ops.h

@@ -97,6 +97,7 @@ enum htp_op_code {
     HTP_OP_PAD,
     HTP_OP_NORM,
     HTP_OP_CONCAT,
+    HTP_OP_CLAMP,
 
     HTP_OP_INVALID
 };

ggml/src/ggml-hexagon/htp/main.c

@@ -718,6 +718,7 @@ static int execute_op(struct htp_ops_context * octx) {
         case HTP_OP_RMS_NORM:
         case HTP_OP_RMS_NORM_MUL:
         case HTP_OP_SCALE:
+        case HTP_OP_CLAMP:
         case HTP_OP_SQR:
         case HTP_OP_SQRT:
         case HTP_OP_UNARY_SOFTPLUS:

ggml/src/ggml-hexagon/htp/unary-ops.c

@@ -138,6 +138,24 @@ static void scale_f32(const float * restrict src,
     }
 }
 
+static void clamp_f32(const float * restrict src,
+                      float * restrict dst,
+                      const uint32_t num_rows,
+                      const struct htp_unary_context * uctx) {
+    htp_unary_op_preamble;
+    float min = 0.f;
+    float max = 0.f;
+    memcpy(&min, &op_params[0], sizeof(float));
+    memcpy(&max, &op_params[1], sizeof(float));
+
+    for (uint32_t ir = 0; ir < num_rows; ir++) {
+        const uint8_t * restrict src_local = (const uint8_t *)src + (ir * src0_row_size_aligned);
+        uint8_t * restrict dst_local       = (uint8_t *)dst + (ir * dst_row_size_aligned);
+
+        hvx_clamp_scalar_f32(dst_local, src_local, min, max, ne0);
+    }
+}
+
 static void rms_norm_f32(const float * restrict src,
                          float * restrict dst,
                          const uint32_t num_rows,
@@ -542,6 +560,7 @@ DEFINE_UNARY_TASK(norm,           false, false, norm_f32(src0_vtcm, dst_vtcm, bl
 DEFINE_UNARY_TASK(rms_norm,       false, false, rms_norm_f32(src0_vtcm, dst_vtcm, block_size, uctx))
 DEFINE_UNARY_TASK(rms_norm_mul,   true,  false, rms_norm_mul_f32(src0_vtcm, uctx->broadcast_weight ? (const float *) src1_vtcm_data : src1_vtcm, dst_vtcm, block_size, uctx))
 DEFINE_UNARY_TASK(scale,          false, false, scale_f32(src0_vtcm, dst_vtcm, block_size, uctx))
+DEFINE_UNARY_TASK(clamp,          false, false, clamp_f32(src0_vtcm, dst_vtcm, block_size, uctx))
 DEFINE_UNARY_TASK(sqr,            false, false, sqr_f32(src0_vtcm, dst_vtcm, block_size, uctx))
 DEFINE_UNARY_TASK(sqrt,           false, false, sqrt_f32(src0_vtcm, dst_vtcm, block_size, uctx))
 DEFINE_UNARY_TASK(unary_neg,      false, false, neg_f32(src0_vtcm, dst_vtcm, block_size, uctx))
@@ -681,6 +700,14 @@ static inline void tile_scale_f32(uint8_t * dst_vtcm, const uint8_t * src_vtcm,
     hvx_scale_offset_f32_aa(dst_vtcm, src_vtcm, tw, scale, bias);
 }
 
+static inline void tile_clamp_f32(uint8_t * dst_vtcm, const uint8_t * src_vtcm, uint32_t tw, const int32_t * op_params) {
+    float min = 0.f;
+    float max = 0.f;
+    memcpy(&min, &op_params[0], sizeof(float));
+    memcpy(&max, &op_params[1], sizeof(float));
+    hvx_clamp_scalar_f32(dst_vtcm, src_vtcm, min, max, tw);
+}
+
 static inline void tile_unary_softplus_f32(uint8_t * dst_vtcm, const uint8_t * src_vtcm, uint32_t tw) {
     const float * restrict sf = (const float *) src_vtcm;
     float * restrict df       = (float *) dst_vtcm;
@@ -765,6 +792,7 @@ static inline void tri_apply_tile_f32(const uint8_t * restrict src, uint8_t * re
 }
 
 DEFINE_UNARY_TILED_TASK(scale,          false, tile_scale_f32(dst_vtcm, src_vtcm, tw, op_params))
+DEFINE_UNARY_TILED_TASK(clamp,          false, tile_clamp_f32(dst_vtcm, src_vtcm, tw, op_params))
 DEFINE_UNARY_TILED_TASK(sqr,            false, hvx_sqr_f32_aa(dst_vtcm, src_vtcm, tw))
 DEFINE_UNARY_TILED_TASK(sqrt,           false, hvx_sqrt_f32_aa(dst_vtcm, src_vtcm, tw))
 DEFINE_UNARY_TILED_TASK(unary_neg,      false, hvx_scale_f32_aa(dst_vtcm, src_vtcm, tw, -1.0f))
@@ -787,6 +815,7 @@ static int execute_op_unary_f32(struct htp_ops_context * octx) {
         case HTP_OP_RMS_NORM:        op_type = "rmsnorm-f32";      break;
         case HTP_OP_RMS_NORM_MUL:    op_type = "rmsnorm-mul-f32";  break;
         case HTP_OP_SCALE:           op_type = "scale-f32";        break;
+        case HTP_OP_CLAMP:           op_type = "clamp-f32";        break;
         case HTP_OP_SQR:             op_type = "sqr-f32";          break;
         case HTP_OP_SQRT:            op_type = "sqrt-f32";         break;
         case HTP_OP_UNARY_NEG:       op_type = "neg-f32";          break;
@@ -882,6 +911,7 @@ static int execute_op_unary_f32(struct htp_ops_context * octx) {
         if (col_tile) {
             switch (octx->op) {
                 case HTP_OP_SCALE:           task_func = unary_task_f32_tiled_scale;          break;
+                case HTP_OP_CLAMP:           task_func = unary_task_f32_tiled_clamp;          break;
                 case HTP_OP_SQR:             task_func = unary_task_f32_tiled_sqr;            break;
                 case HTP_OP_SQRT:            task_func = unary_task_f32_tiled_sqrt;           break;
                 case HTP_OP_UNARY_NEG:       task_func = unary_task_f32_tiled_unary_neg;      break;
@@ -898,6 +928,7 @@ static int execute_op_unary_f32(struct htp_ops_context * octx) {
                 case HTP_OP_RMS_NORM:        task_func = unary_task_f32_rms_norm;             break;
                 case HTP_OP_RMS_NORM_MUL:    task_func = unary_task_f32_rms_norm_mul;         break;
                 case HTP_OP_SCALE:           task_func = unary_task_f32_scale;                break;
+                case HTP_OP_CLAMP:           task_func = unary_task_f32_clamp;                break;
                 case HTP_OP_SQR:             task_func = unary_task_f32_sqr;                  break;
                 case HTP_OP_SQRT:            task_func = unary_task_f32_sqrt;                 break;
                 case HTP_OP_UNARY_NEG:       task_func = unary_task_f32_unary_neg;            break;

ggml/src/ggml-hexagon/htp/unary-ops.h

@@ -41,6 +41,7 @@ _Static_assert(sizeof(struct htp_unary_kernel_params) <= 128, "htp_unary_kernel_
 
 static inline bool htp_op_is_unary(uint32_t opcode) {
     switch (opcode) {
+        case HTP_OP_CLAMP:
         case HTP_OP_NORM:
         case HTP_OP_RMS_NORM:
         case HTP_OP_RMS_NORM_MUL:

ggml/src/ggml-opencl/ggml-opencl.cpp

@@ -17085,9 +17085,6 @@ static void ggml_cl_mul_mat_q8_0_f32_adreno(ggml_backend_t backend, const ggml_t
     cl_ulong offset1 = extra1->offset + src1->view_offs;
     cl_ulong offsetd = extrad->offset + dst->view_offs;
 
-    GGML_ASSERT(src1->view_offs == 0);
-    GGML_ASSERT(dst->view_offs == 0);
-
     const int  ne00 = src0->ne[0];
     const int  ne01 = src0->ne[1];
     const int  ne02 = src0->ne[2];
@@ -17148,9 +17145,9 @@ static void ggml_cl_mul_mat_q8_0_f32_adreno(ggml_backend_t backend, const ggml_t
         CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),   &q_img));
         CL_CHECK(clSetKernelArg(kernel,  1, sizeof(cl_mem),   &extra0_q8_0->d));
         CL_CHECK(clSetKernelArg(kernel,  2, sizeof(cl_mem),   &b_img));
-        CL_CHECK(clSetKernelArg(kernel,  3, sizeof(cl_ulong), &extra1->offset));
+        CL_CHECK(clSetKernelArg(kernel,  3, sizeof(cl_ulong), &offset1));
         CL_CHECK(clSetKernelArg(kernel,  4, sizeof(cl_mem),   &extrad->data_device));
-        CL_CHECK(clSetKernelArg(kernel,  5, sizeof(cl_ulong), &extrad->offset));
+        CL_CHECK(clSetKernelArg(kernel,  5, sizeof(cl_ulong), &offsetd));
         CL_CHECK(clSetKernelArg(kernel,  6, sizeof(int),      &ne00));
         CL_CHECK(clSetKernelArg(kernel,  7, sizeof(int),      &ne01));
         CL_CHECK(clSetKernelArg(kernel,  8, sizeof(int),      &ne02));
@@ -18545,6 +18542,26 @@ static void ggml_cl_mul_mat(ggml_backend_t backend, const ggml_tensor * src0, co
 
     GGML_ASSERT(ne00 == ne10);
 
+#ifdef GGML_OPENCL_USE_ADRENO_KERNELS
+    // adreno GEMM/GEMV kernels do not support broadcast, assuming ne2 and ne3 are 1 for src1
+    // so we handle broadcast here
+    if ((ne12 > 1 || ne13 > 1) && ne02 == 1 && ne03 == 1 &&
+        src0t != GGML_TYPE_F16 && src0t != GGML_TYPE_F32) {
+        for (int i13 = 0; i13 < ne13; ++i13) {
+            for (int i12 = 0; i12 < ne12; ++i12) {
+                ggml_tensor s1 = *src1;
+                s1.ne[2] = 1; s1.ne[3] = 1;
+                s1.view_offs = src1->view_offs + (size_t)i12*nb12 + (size_t)i13*nb13;
+                ggml_tensor d = *dst;
+                d.ne[2] = 1; d.ne[3] = 1;
+                d.view_offs = dst->view_offs + (size_t)i12*nb2 + (size_t)i13*nb3;
+                ggml_cl_mul_mat(backend, src0, &s1, &d);
+            }
+        }
+        return;
+    }
+#endif
+
     int nth0 = 32;
     int nth1 = 1;
     int nrows = 1;

ggml/src/ggml-openvino/ggml-openvino.cpp

@@ -1378,3 +1378,5 @@ GGML_BACKEND_API ggml_backend_reg_t ggml_backend_openvino_reg(void) {
 
     return ®
 }
+
+GGML_BACKEND_DL_IMPL(ggml_backend_openvino_reg)

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

@@ -156,6 +156,24 @@ typedef struct VkPhysicalDeviceShaderFloat8FeaturesEXT {
 } VkPhysicalDeviceShaderFloat8FeaturesEXT;
 #endif
 
+#ifndef VK_KHR_INTERNALLY_SYNCHRONIZED_QUEUES_EXTENSION_NAME
+#define VK_KHR_INTERNALLY_SYNCHRONIZED_QUEUES_EXTENSION_NAME "VK_KHR_internally_synchronized_queues"
+#define VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_INTERNALLY_SYNCHRONIZED_QUEUES_FEATURES_KHR ((VkStructureType)1000504000)
+#define VK_DEVICE_QUEUE_CREATE_INTERNALLY_SYNCHRONIZED_BIT_KHR ((VkDeviceQueueCreateFlagBits)0x00000004)
+
+// Compile-time constant guaranteed; no runtime initialization overhead
+static constexpr vk::DeviceQueueCreateFlagBits eInternallySynchronizedKHR =
+    static_cast<vk::DeviceQueueCreateFlagBits>(0x00000004);
+
+typedef struct VkPhysicalDeviceInternallySynchronizedQueuesFeaturesKHR {
+    VkStructureType    sType;
+    void* pNext;
+    VkBool32           internallySynchronizedQueues;
+} VkPhysicalDeviceInternallySynchronizedQueuesFeaturesKHR;
+#else
+static constexpr vk::DeviceQueueCreateFlagBits eInternallySynchronizedKHR = vk::DeviceQueueCreateFlagBits::eInternallySynchronizedKHR;
+#endif
+
 #define ROUNDUP_POW2(M, N) (((M) + (N) - 1) & ~((N) - 1))
 #define CEIL_DIV(M, N) (((M) / (N)) + (((M) % (N)) != 0))
 static bool is_pow2(uint32_t x) { return x > 1 && (x & (x-1)) == 0; }
@@ -285,27 +303,41 @@ struct vk_command_pool {
 };
 
 // Prevent simultaneous submissions to the same queue.
-// This could be per vk_queue if we stopped having two vk_queue structures
-// sharing the same vk::Queue.
-static std::mutex queue_mutex;
+struct vk_queue_handle {
+    vk::Queue queue;
+    virtual void submit(vk::ArrayProxy<const vk::SubmitInfo> submits, vk::Fence fence) = 0;
+    virtual void lock()   {}   // no-op by default (internally synchronized case)
+    virtual void unlock() {}
+    virtual ~vk_queue_handle() = default;
+};
+
+struct vk_queue_handle_synchronized : vk_queue_handle {
+    std::mutex mutex;
+    void submit(vk::ArrayProxy<const vk::SubmitInfo> submits, vk::Fence fence) override {
+        std::lock_guard<std::mutex> guard(mutex);
+        queue.submit(submits, fence);
+    }
+    void lock()   override { mutex.lock(); }
+    void unlock() override { mutex.unlock(); }
+};
+
+struct vk_queue_handle_unsynchronized : vk_queue_handle {
+    void submit(vk::ArrayProxy<const vk::SubmitInfo> submits, vk::Fence fence) override {
+        // Driver guarantees internal synchronization via VK_KHR_internally_synchronized_queues
+        queue.submit(submits, fence);
+    }
+    // lock()/unlock() inherited no-ops
+};
 
 struct vk_queue {
     uint32_t queue_family_index;
-    vk::Queue queue;
+    std::shared_ptr<vk_queue_handle> handle;
 
     vk_command_pool cmd_pool;
 
     vk::PipelineStageFlags stage_flags;
 
     bool transfer_only;
-
-    // copy everything except the cmd_pool
-    void copyFrom(vk_queue &other) {
-        queue_family_index = other.queue_family_index;
-        queue = other.queue;
-        stage_flags = other.stage_flags;
-        transfer_only = other.transfer_only;
-    }
 };
 
 static const char * ggml_backend_vk_buffer_type_name(ggml_backend_buffer_type_t buft);
@@ -712,11 +744,12 @@ struct vk_device_struct {
     uint32_t vendor_id;
     vk::DriverId driver_id;
     vk_device_architecture architecture;
-    vk_queue compute_queue;
-    vk_queue transfer_queue;
+    std::unique_ptr<vk_queue> compute_queue;
+    std::unique_ptr<vk_queue> transfer_queue;
     bool single_queue;
     bool support_async;
     bool async_use_transfer_queue;
+    bool has_internally_synchronized_queues = false;
     uint32_t subgroup_size;
     uint32_t subgroup_size_log2;
     uint32_t shader_core_count;
@@ -1019,8 +1052,13 @@ struct vk_device_struct {
 
         ggml_vk_destroy_buffer(sync_staging);
 
-        compute_queue.cmd_pool.destroy(device);
-        transfer_queue.cmd_pool.destroy(device);
+        if (compute_queue) compute_queue->cmd_pool.destroy(device);
+        if (transfer_queue) transfer_queue->cmd_pool.destroy(device);
+
+        // Explicitly clear to ensure queues drop their shared_ptrs to handles
+        // before the Vulkan logical device instance is destroyed
+        compute_queue.reset();
+        transfer_queue.reset();
 
         for (auto& pipeline : all_pipelines) {
             if (pipeline.expired()) {
@@ -2909,8 +2947,7 @@ static vk_command_buffer* ggml_vk_create_cmd_buffer(vk_device& device, vk_comman
 static void ggml_vk_submit(vk_context& ctx, vk::Fence fence) {
     if (ctx->seqs.empty()) {
         if (fence) {
-            std::lock_guard<std::mutex> guard(queue_mutex);
-            ctx->p->q->queue.submit({}, fence);
+            ctx->p->q->handle->submit({}, fence);
         }
         return;
     }
@@ -2979,8 +3016,7 @@ static void ggml_vk_submit(vk_context& ctx, vk::Fence fence) {
         }
     }
 
-    std::lock_guard<std::mutex> guard(queue_mutex);
-    ctx->p->q->queue.submit(submit_infos, fence);
+    ctx->p->q->handle->submit(submit_infos, fence);
 
     ctx->seqs.clear();
 }
@@ -3031,18 +3067,44 @@ static uint32_t ggml_vk_find_queue_family_index(std::vector<vk::QueueFamilyPrope
     abort();
 }
 
-static void ggml_vk_create_queue(vk_device& device, vk_queue& q, uint32_t queue_family_index, uint32_t queue_index, vk::PipelineStageFlags&& stage_flags, bool transfer_only) {
+static std::unique_ptr<vk_queue> ggml_vk_create_queue(vk_device& device, uint32_t queue_family_index, uint32_t queue_index, vk::PipelineStageFlags&& stage_flags, bool transfer_only) {
     VK_LOG_DEBUG("ggml_vk_create_queue()");
     std::lock_guard<std::recursive_mutex> guard(device->mutex);
 
-    q.queue_family_index = queue_family_index;
-    q.transfer_only = transfer_only;
+    auto q = std::make_unique<vk_queue>();
+    q->queue_family_index = queue_family_index;
+    q->transfer_only = transfer_only;
 
-    q.cmd_pool.init(device, &q);
+    std::shared_ptr<vk_queue_handle> h;
+    vk::DeviceQueueInfo2 queue_info2{};
+    queue_info2.queueFamilyIndex = queue_family_index;
+    queue_info2.queueIndex = queue_index;
 
-    q.queue = device->device.getQueue(queue_family_index, queue_index);
+    if (device->has_internally_synchronized_queues) {
+        h = std::make_shared<vk_queue_handle_unsynchronized>();
+        queue_info2.flags = eInternallySynchronizedKHR;
+    } else {
+        h = std::make_shared<vk_queue_handle_synchronized>();
+    }
 
-    q.stage_flags = stage_flags;
+    h->queue = device->device.getQueue2(queue_info2);
+    q->handle = h;
+
+    q->cmd_pool.init(device, q.get());
+
+    q->stage_flags = stage_flags;
+    return q;
+}
+
+static std::unique_ptr<vk_queue> ggml_vk_create_aliased_queue(vk_device& device, const std::unique_ptr<vk_queue>& source) {
+    std::lock_guard<std::recursive_mutex> guard(device->mutex);
+    auto q = std::make_unique<vk_queue>();
+    q->handle = source->handle;
+    q->queue_family_index = source->queue_family_index;
+    q->stage_flags = source->stage_flags;
+    q->transfer_only = source->transfer_only;
+    q->cmd_pool.init(device, q.get());
+    return q;
 }
 
 static vk_context ggml_vk_create_context(ggml_backend_vk_context * ctx, vk_command_pool& p) {
@@ -3107,11 +3169,11 @@ static void ggml_vk_queue_command_pools_cleanup(vk_device& device) {
     // Arbitrary frequency to cleanup/reuse command buffers
     static constexpr uint32_t cleanup_frequency = 10;
 
-    if (device->compute_queue.cmd_pool.buffers_in_use() >= cleanup_frequency) {
-        ggml_vk_command_pool_cleanup(device, device->compute_queue.cmd_pool);
+    if (device->compute_queue->cmd_pool.buffers_in_use() >= cleanup_frequency) {
+        ggml_vk_command_pool_cleanup(device, device->compute_queue->cmd_pool);
     }
-    if (device->transfer_queue.cmd_pool.buffers_in_use() >= cleanup_frequency) {
-        ggml_vk_command_pool_cleanup(device, device->transfer_queue.cmd_pool);
+    if (device->transfer_queue->cmd_pool.buffers_in_use() >= cleanup_frequency) {
+        ggml_vk_command_pool_cleanup(device, device->transfer_queue->cmd_pool);
     }
 }
 
@@ -5886,6 +5948,7 @@ static vk_device ggml_vk_get_device(size_t idx) {
         bool coopmat2_support = false;
         bool coopmat2_decode_vector_support = false;
         bool pipeline_executable_properties_support = false;
+        bool internally_sync_support = false;
         device->coopmat_support = false;
         device->integer_dot_product = false;
         device->shader_64b_indexing = false;
@@ -5957,6 +6020,8 @@ static vk_device ggml_vk_get_device(size_t idx) {
             } else if (strcmp("VK_EXT_shader_64bit_indexing", properties.extensionName) == 0) {
                 device->shader_64b_indexing = true;
 #endif
+            } else if (strcmp(VK_KHR_INTERNALLY_SYNCHRONIZED_QUEUES_EXTENSION_NAME, properties.extensionName) == 0) {
+                internally_sync_support = true;
             }
         }
 
@@ -6143,14 +6208,6 @@ static vk_device ggml_vk_get_device(size_t idx) {
         device->single_queue = compute_queue_family_index == transfer_queue_family_index && queue_family_props[compute_queue_family_index].queueCount == 1;
 
         std::vector<vk::DeviceQueueCreateInfo> device_queue_create_infos;
-        if (compute_queue_family_index != transfer_queue_family_index) {
-            device_queue_create_infos.push_back({vk::DeviceQueueCreateFlags(), compute_queue_family_index, 1, priorities});
-            device_queue_create_infos.push_back({vk::DeviceQueueCreateFlags(), transfer_queue_family_index, 1, priorities + 1});
-        } else if(!device->single_queue) {
-            device_queue_create_infos.push_back({vk::DeviceQueueCreateFlags(), compute_queue_family_index, 2, priorities});
-        } else {
-            device_queue_create_infos.push_back({vk::DeviceQueueCreateFlags(), compute_queue_family_index, 1, priorities});
-        }
         vk::DeviceCreateInfo device_create_info{};
         std::vector<const char *> device_extensions;
         vk::PhysicalDeviceFeatures device_features = device->physical_device.getFeatures();
@@ -6172,6 +6229,17 @@ static vk_device ggml_vk_get_device(size_t idx) {
 
         last_struct = (VkBaseOutStructure *)&vk12_features;
 
+        VkPhysicalDeviceInternallySynchronizedQueuesFeaturesKHR internally_synchronized_queues_features{};
+        internally_synchronized_queues_features.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_INTERNALLY_SYNCHRONIZED_QUEUES_FEATURES_KHR;
+        internally_synchronized_queues_features.pNext = nullptr;
+        internally_synchronized_queues_features.internallySynchronizedQueues = VK_FALSE;
+
+        if (internally_sync_support) {
+            last_struct->pNext = (VkBaseOutStructure *)&internally_synchronized_queues_features;
+            last_struct = (VkBaseOutStructure *)&internally_synchronized_queues_features;
+            device_extensions.push_back(VK_KHR_INTERNALLY_SYNCHRONIZED_QUEUES_EXTENSION_NAME);
+        }
+
         VkPhysicalDevicePipelineRobustnessFeaturesEXT pl_robustness_features;
         pl_robustness_features.pNext = nullptr;
         pl_robustness_features.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PIPELINE_ROBUSTNESS_FEATURES_EXT;
@@ -6310,6 +6378,23 @@ static vk_device ggml_vk_get_device(size_t idx) {
 
         vkGetPhysicalDeviceFeatures2(device->physical_device, &device_features2);
 
+        device->has_internally_synchronized_queues = internally_synchronized_queues_features.internallySynchronizedQueues;
+
+        // Build queue create infos only after querying whether internally synchronized queues are enabled.
+        // getQueue2() later uses the same flag, so creation/retrieval must stay consistent.
+        vk::DeviceQueueCreateFlags queue_flags = device->has_internally_synchronized_queues ?
+                                                eInternallySynchronizedKHR :
+                                                vk::DeviceQueueCreateFlags();
+
+        if (compute_queue_family_index != transfer_queue_family_index) {
+            device_queue_create_infos.push_back({queue_flags, compute_queue_family_index, 1, priorities});
+            device_queue_create_infos.push_back({queue_flags, transfer_queue_family_index, 1, priorities + 1});
+        } else if(!device->single_queue) {
+            device_queue_create_infos.push_back({queue_flags, compute_queue_family_index, 2, priorities});
+        } else {
+            device_queue_create_infos.push_back({queue_flags, compute_queue_family_index, 1, priorities});
+        }
+
         device->pipeline_executable_properties_support = pipeline_executable_properties_support;
 
         device->fp16 = device->fp16 && vk12_features.shaderFloat16;
@@ -6592,7 +6677,7 @@ static vk_device ggml_vk_get_device(size_t idx) {
         device->device = device->physical_device.createDevice(device_create_info);
 
         // Queues
-        ggml_vk_create_queue(device, device->compute_queue, compute_queue_family_index, 0, { vk::PipelineStageFlagBits::eComputeShader | vk::PipelineStageFlagBits::eTransfer }, false);
+        device->compute_queue = ggml_vk_create_queue(device, compute_queue_family_index, 0, { vk::PipelineStageFlagBits::eComputeShader | vk::PipelineStageFlagBits::eTransfer }, false);
 
         // Shaders
         // Disable matmul tile sizes early if performance low or not supported
@@ -6694,13 +6779,11 @@ static vk_device ggml_vk_get_device(size_t idx) {
 
         if (!device->single_queue) {
             const uint32_t transfer_queue_index = compute_queue_family_index == transfer_queue_family_index ? 1 : 0;
-            ggml_vk_create_queue(device, device->transfer_queue, transfer_queue_family_index, transfer_queue_index, { vk::PipelineStageFlagBits::eTransfer }, true);
+            device->transfer_queue = ggml_vk_create_queue(device, transfer_queue_family_index, transfer_queue_index, { vk::PipelineStageFlagBits::eTransfer }, true);
 
             device->async_use_transfer_queue = prefers_transfer_queue || (getenv("GGML_VK_ASYNC_USE_TRANSFER_QUEUE") != nullptr);
         } else {
-            // TODO: Use pointer or reference to avoid copy
-            device->transfer_queue.copyFrom(device->compute_queue);
-            device->transfer_queue.cmd_pool.init(device, &device->transfer_queue);
+            device->transfer_queue = ggml_vk_create_aliased_queue(device, device->compute_queue);
 
             device->async_use_transfer_queue = false;
         }
@@ -7263,7 +7346,7 @@ static void ggml_vk_init(ggml_backend_vk_context * ctx, size_t idx) {
     ctx->fence = ctx->device->device.createFence({});
     ctx->almost_ready_fence = ctx->device->device.createFence({});
 
-    ctx->compute_cmd_pool.init(ctx->device, &ctx->device->compute_queue);
+    ctx->compute_cmd_pool.init(ctx->device, ctx->device->compute_queue.get());
     if (ctx->device->async_use_transfer_queue) {
         vk::SemaphoreTypeCreateInfo tci{ vk::SemaphoreType::eTimeline, 0 };
         vk::SemaphoreCreateInfo ci{};
@@ -7271,7 +7354,7 @@ static void ggml_vk_init(ggml_backend_vk_context * ctx, size_t idx) {
         ctx->transfer_semaphore.s = ctx->device->device.createSemaphore(ci);
         ctx->transfer_semaphore.value = 0;
 
-        ctx->transfer_cmd_pool.init(ctx->device, &ctx->device->transfer_queue);
+        ctx->transfer_cmd_pool.init(ctx->device, ctx->device->transfer_queue.get());
     }
 
     if (vk_perf_logger_enabled) {
@@ -8126,7 +8209,7 @@ static void ggml_vk_buffer_write_2d(vk_buffer& dst, size_t offset, const void *
     } else {
         std::lock_guard<std::recursive_mutex> guard(dst->device->mutex);
 
-        vk_context subctx = ggml_vk_create_temporary_context(dst->device->transfer_queue.cmd_pool);
+        vk_context subctx = ggml_vk_create_temporary_context(dst->device->transfer_queue->cmd_pool);
         ggml_vk_ctx_begin(dst->device, subctx);
         bool ret = ggml_vk_buffer_write_2d_async(subctx, dst, offset, src, spitch, dpitch, width, height, true);
         GGML_ASSERT(ret);
@@ -8241,7 +8324,7 @@ static void ggml_vk_buffer_read_2d(vk_buffer& src, size_t offset, void * dst, si
         GGML_ASSERT(src->memory_property_flags & vk::MemoryPropertyFlagBits::eHostCoherent);
 
         std::lock_guard<std::recursive_mutex> guard(src->device->mutex);
-        vk_context subctx = ggml_vk_create_temporary_context(src->device->compute_queue.cmd_pool);
+        vk_context subctx = ggml_vk_create_temporary_context(src->device->compute_queue->cmd_pool);
         ggml_vk_ctx_begin(src->device, subctx);
         subctx->s->buffer->buf.pipelineBarrier(
             vk::PipelineStageFlagBits::eComputeShader | vk::PipelineStageFlagBits::eTransfer,
@@ -8267,7 +8350,7 @@ static void ggml_vk_buffer_read_2d(vk_buffer& src, size_t offset, void * dst, si
     } else {
         std::lock_guard<std::recursive_mutex> guard(src->device->mutex);
 
-        vk_context subctx = ggml_vk_create_temporary_context(src->device->transfer_queue.cmd_pool);
+        vk_context subctx = ggml_vk_create_temporary_context(src->device->transfer_queue->cmd_pool);
         ggml_vk_ctx_begin(src->device, subctx);
         bool ret = ggml_vk_buffer_read_2d_async(subctx, src, offset, dst, spitch, dpitch, width, height, true);
         GGML_ASSERT(ret);
@@ -8304,7 +8387,7 @@ static void ggml_vk_buffer_copy(vk_buffer& dst, size_t dst_offset, vk_buffer& sr
         std::lock_guard<std::recursive_mutex> guard(src->device->mutex);
         VK_LOG_DEBUG("ggml_vk_buffer_copy(SINGLE_DEVICE, " << size << ")");
         // Copy within the device
-        vk_context subctx = ggml_vk_create_temporary_context(src->device->transfer_queue.cmd_pool);
+        vk_context subctx = ggml_vk_create_temporary_context(src->device->transfer_queue->cmd_pool);
         ggml_vk_ctx_begin(src->device, subctx);
         ggml_vk_buffer_copy_async(subctx, dst, dst_offset, src, src_offset, size);
         ggml_vk_ctx_end(subctx);
@@ -8347,7 +8430,7 @@ static void ggml_vk_buffer_memset(vk_buffer& dst, size_t offset, uint32_t c, siz
     }
 
     std::lock_guard<std::recursive_mutex> guard(dst->device->mutex);
-    vk_context subctx = ggml_vk_create_temporary_context(dst->device->transfer_queue.cmd_pool);
+    vk_context subctx = ggml_vk_create_temporary_context(dst->device->transfer_queue->cmd_pool);
     ggml_vk_ctx_begin(dst->device, subctx);
     subctx->s->buffer->buf.fillBuffer(dst->buffer, offset, size, c);
     ggml_vk_ctx_end(subctx);
@@ -15875,19 +15958,17 @@ static void ggml_vk_synchronize(ggml_backend_vk_context * ctx) {
                 1, &ctx->transfer_semaphore.value,
                 0, nullptr,
             };
-            vk::PipelineStageFlags stage = ctx->device->transfer_queue.stage_flags;
+            vk::PipelineStageFlags stage = ctx->device->transfer_queue->stage_flags;
             vk::SubmitInfo si{
                 1, &ctx->transfer_semaphore.s, &stage,
                 0, nullptr,
                 0, nullptr,
             };
             si.setPNext(&tl_info);
-            std::lock_guard<std::mutex> guard(queue_mutex);
-            ctx->device->compute_queue.queue.submit({ si }, ctx->fence);
+            ctx->device->compute_queue->handle->submit({ si }, ctx->fence);
             ctx->transfer_semaphore_last_submitted = ctx->transfer_semaphore.value;
         } else {
-            std::lock_guard<std::mutex> guard(queue_mutex);
-            ctx->device->compute_queue.queue.submit({}, ctx->fence);
+            ctx->device->compute_queue->handle->submit({}, ctx->fence);
         }
         ggml_vk_wait_for_fence(ctx);
         ctx->submit_pending = false;
@@ -16449,7 +16530,9 @@ static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cg
         vk::DebugUtilsLabelEXT dul = {};
         dul.pLabelName = "ggml_backend_vk_graph_compute";
         dul.color = std::array<float,4>{1.0f, 1.0f, 1.0f, 1.0f};
-        vk_instance.pfn_vkQueueBeginDebugUtilsLabelEXT(ctx->device->compute_queue.queue, reinterpret_cast<VkDebugUtilsLabelEXT*>(&dul));
+
+        std::lock_guard<vk_queue_handle> guard(*ctx->device->compute_queue->handle);
+        vk_instance.pfn_vkQueueBeginDebugUtilsLabelEXT(ctx->device->compute_queue->handle->queue, reinterpret_cast<VkDebugUtilsLabelEXT*>(&dul));
     }
 
     ctx->prealloc_size_add_rms_partials_offset = 0;

ggml/src/ggml-webgpu/ggml-webgpu-shader-lib.hpp

@@ -355,6 +355,30 @@ struct ggml_webgpu_conv2d_pipeline_key_hash {
     }
 };
 
+// Same type fields as conv2d plus the input layout (WHCN vs CWHN).
+struct ggml_webgpu_conv2d_dw_pipeline_key {
+    ggml_type weight_type;
+    ggml_type input_type;
+    ggml_type output_type;
+    bool      whcn;
+
+    bool operator==(const ggml_webgpu_conv2d_dw_pipeline_key & other) const {
+        return weight_type == other.weight_type && input_type == other.input_type && output_type == other.output_type &&
+               whcn == other.whcn;
+    }
+};
+
+struct ggml_webgpu_conv2d_dw_pipeline_key_hash {
+    size_t operator()(const ggml_webgpu_conv2d_dw_pipeline_key & key) const {
+        size_t seed = 0;
+        ggml_webgpu_hash_combine(seed, key.weight_type);
+        ggml_webgpu_hash_combine(seed, key.input_type);
+        ggml_webgpu_hash_combine(seed, key.output_type);
+        ggml_webgpu_hash_combine(seed, key.whcn);
+        return seed;
+    }
+};
+
 /** Im2Col **/
 struct ggml_webgpu_im2col_pipeline_key {
     ggml_type input_type;
@@ -1210,6 +1234,8 @@ class ggml_webgpu_shader_lib {
         soft_max_pipelines;
     std::unordered_map<ggml_webgpu_conv2d_pipeline_key, webgpu_pipeline, ggml_webgpu_conv2d_pipeline_key_hash>
         conv2d_pipelines;
+    std::unordered_map<ggml_webgpu_conv2d_dw_pipeline_key, webgpu_pipeline, ggml_webgpu_conv2d_dw_pipeline_key_hash>
+        conv2d_dw_pipelines;
     std::unordered_map<ggml_webgpu_im2col_pipeline_key, webgpu_pipeline, ggml_webgpu_im2col_pipeline_key_hash>
         im2col_pipelines;
 
@@ -3172,6 +3198,50 @@ class ggml_webgpu_shader_lib {
         return conv2d_pipelines[key];
     }
 
+    // whcn selects the input layout: contiguous WHCN vs contiguous-channels CWHN
+    webgpu_pipeline get_conv2d_dw_pipeline(const ggml_webgpu_shader_lib_context & context, bool whcn) {
+        ggml_webgpu_conv2d_dw_pipeline_key key = {};
+        key.weight_type                        = context.src0->type;
+        key.input_type                         = context.src1->type;
+        key.output_type                        = context.dst->type;
+        key.whcn                               = whcn;
+
+        auto it = conv2d_dw_pipelines.find(key);
+        if (it != conv2d_dw_pipelines.end()) {
+            return it->second;
+        }
+
+        std::vector<std::string> defines;
+        std::string              variant = whcn ? "conv_2d_dw_whcn" : "conv_2d_dw_cwhn";
+
+        auto push_type_defines = [&](const char * prefix, ggml_type type) {
+            std::string s_prefix = prefix;
+            if (type == GGML_TYPE_F32) {
+                defines.push_back(s_prefix + "_F32");
+            } else if (type == GGML_TYPE_F16) {
+                defines.push_back(s_prefix + "_F16");
+            } else {
+                GGML_ABORT("Unsupported type for CONV_2D_DW shader");
+            }
+        };
+
+        push_type_defines("WEIGHT", key.weight_type);
+        push_type_defines("INPUT", key.input_type);
+        push_type_defines("OUTPUT", key.output_type);
+        if (whcn) {
+            defines.push_back("WHCN");
+        }
+        defines.push_back(std::string("WG_SIZE=") + std::to_string(context.max_wg_size));
+
+        auto processed           = preprocessor.preprocess(wgsl_conv2d_dw, defines);
+        auto decisions           = std::make_shared<ggml_webgpu_generic_shader_decisions>();
+        decisions->wg_size       = context.max_wg_size;
+        webgpu_pipeline pipeline = ggml_webgpu_create_pipeline(device, processed, variant);
+        pipeline.context         = decisions;
+        conv2d_dw_pipelines[key] = pipeline;
+        return conv2d_dw_pipelines[key];
+    }
+
     webgpu_pipeline get_im2col_pipeline(const ggml_webgpu_shader_lib_context & context) {
         ggml_webgpu_im2col_pipeline_key key = {};
         key.input_type                      = context.src1->type;

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

@@ -978,6 +978,67 @@ static webgpu_encoded_op ggml_webgpu_conv_2d(webgpu_context & ctx,
     return ggml_backend_webgpu_build(ctx, pipeline, params, entries, wg_x, wg_y);
 }
 
+// Same param/binding layout as conv_2d; the shader differs
+static webgpu_encoded_op ggml_webgpu_conv_2d_dw(webgpu_context & ctx,
+                                                ggml_tensor *    src0,
+                                                ggml_tensor *    src1,
+                                                ggml_tensor *    dst) {
+    const int32_t s0 = ggml_get_op_params_i32(dst, 0);
+    const int32_t s1 = ggml_get_op_params_i32(dst, 1);
+    const int32_t p0 = ggml_get_op_params_i32(dst, 2);
+    const int32_t p1 = ggml_get_op_params_i32(dst, 3);
+    const int32_t d0 = ggml_get_op_params_i32(dst, 4);
+    const int32_t d1 = ggml_get_op_params_i32(dst, 5);
+
+    // Scalar params matching conv2d_dw.wgsl (weight src0 [KW,KH,1,C], input src1, output dst).
+    std::vector<uint32_t> params = {
+        (uint32_t) (ggml_webgpu_tensor_misalignment(ctx, src0) / ggml_type_size(src0->type)),
+        (uint32_t) (ggml_webgpu_tensor_misalignment(ctx, src1) / ggml_type_size(src1->type)),
+        (uint32_t) (ggml_webgpu_tensor_misalignment(ctx, dst) / ggml_type_size(dst->type)),
+
+        (uint32_t) ggml_nelements(dst),
+        (uint32_t) dst->ne[2],
+        (uint32_t) dst->ne[3],
+        (uint32_t) dst->ne[0],
+        (uint32_t) dst->ne[1],
+        (uint32_t) src1->ne[0],
+        (uint32_t) src1->ne[1],
+        (uint32_t) src0->ne[0],
+        (uint32_t) src0->ne[1],
+
+        (uint32_t) s0,
+        (uint32_t) s1,
+        (uint32_t) p0,
+        (uint32_t) p1,
+        (uint32_t) d0,
+        (uint32_t) d1,
+    };
+
+    std::vector<wgpu::BindGroupEntry> entries = {
+        ggml_webgpu_make_tensor_bind_group_entry(ctx, 0, src0),
+        ggml_webgpu_make_tensor_bind_group_entry(ctx, 1, src1),
+        ggml_webgpu_make_tensor_bind_group_entry(ctx, 2, dst),
+    };
+
+    ggml_webgpu_shader_lib_context shader_lib_ctx = {};
+    shader_lib_ctx.src0                           = src0;
+    shader_lib_ctx.src1                           = src1;
+    shader_lib_ctx.dst                            = dst;
+    shader_lib_ctx.max_wg_size = ctx->global_ctx->capabilities.limits.maxComputeInvocationsPerWorkgroup;
+
+    // Input layout: contiguous -> WHCN, contiguous-channels -> CWHN
+    const bool      whcn      = ggml_is_contiguous(src1);
+    webgpu_pipeline pipeline  = ctx->shader_lib->get_conv2d_dw_pipeline(shader_lib_ctx, whcn);
+    auto *          decisions = static_cast<ggml_webgpu_generic_shader_decisions *>(pipeline.context.get());
+
+    uint32_t wg_x;
+    uint32_t wg_y;
+    uint32_t total_wg = CEIL_DIV((uint32_t) ggml_nelements(dst), decisions->wg_size);
+    compute_2d_workgroups(total_wg, ctx->global_ctx->capabilities.limits.maxComputeWorkgroupsPerDimension, wg_x, wg_y);
+
+    return ggml_backend_webgpu_build(ctx, pipeline, params, entries, wg_x, wg_y);
+}
+
 static webgpu_encoded_op ggml_webgpu_im2col(webgpu_context & ctx,
                                             ggml_tensor *    src0,
                                             ggml_tensor *    src1,
@@ -3164,6 +3225,8 @@ static std::optional<webgpu_encoded_op> ggml_webgpu_encode(webgpu_context ctx,
             return ggml_webgpu_sum_rows(ctx, src0, node);
         case GGML_OP_CONV_2D:
             return ggml_webgpu_conv_2d(ctx, src0, src1, node);
+        case GGML_OP_CONV_2D_DW:
+            return ggml_webgpu_conv_2d_dw(ctx, src0, src1, node);
         case GGML_OP_IM2COL:
             return ggml_webgpu_im2col(ctx, src0, src1, node);
         case GGML_OP_UPSCALE:
@@ -4349,6 +4412,12 @@ static bool ggml_backend_webgpu_device_supports_op(ggml_backend_dev_t dev, const
                           (src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16) &&
                           (src1->type == GGML_TYPE_F32 || src1->type == GGML_TYPE_F16);
             break;
+        case GGML_OP_CONV_2D_DW:
+            supports_op = (op->type == GGML_TYPE_F32 || op->type == GGML_TYPE_F16) &&
+                          (src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16) &&
+                          (src1->type == GGML_TYPE_F32 || src1->type == GGML_TYPE_F16) &&
+                          (ggml_is_contiguous(src1) || ggml_is_contiguous_channels(src1));
+            break;
         case GGML_OP_IM2COL:
             supports_op = (op->type == GGML_TYPE_F32 || op->type == GGML_TYPE_F16) &&
                           (src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16);

ggml/src/ggml-webgpu/wgsl-shaders/conv2d_dw.wgsl

@@ -0,0 +1,137 @@
+#include "common_decls.tmpl"
+enable f16;
+
+// Ported from the Vulkan backend's conv2d_dw.comp. Two variants (based on WHCN)
+// selected by the input (src1) layout: contiguous -> WHCN, else CWHN.
+// weight (src0) is [KW,KH,1,C]; output matches the input layout.
+
+@group(0) @binding(0)
+#if defined(WEIGHT_F32)
+var<storage, read_write> weights: array<f32>;
+#elif defined(WEIGHT_F16)
+var<storage, read_write> weights: array<f16>;
+#endif
+
+@group(0) @binding(1)
+#if defined(INPUT_F32)
+var<storage, read_write> input: array<f32>;
+#elif defined(INPUT_F16)
+var<storage, read_write> input: array<f16>;
+#endif
+
+@group(0) @binding(2)
+#if defined(OUTPUT_F32)
+var<storage, read_write> output: array<f32>;
+#elif defined(OUTPUT_F16)
+var<storage, read_write> output: array<f16>;
+#endif
+
+struct Params {
+    offset_w: u32,
+    offset_i: u32,
+    offset_o: u32,
+
+    ne: u32,
+    channels: u32,
+    batches: u32,
+    dst_w: u32, dst_h: u32,
+    src_w: u32, src_h: u32,
+    knl_w: u32, knl_h: u32,
+
+    stride_x: i32, stride_y: i32,
+    pad_x: i32, pad_y: i32,
+    dilation_x: i32, dilation_y: i32,
+};
+
+@group(0) @binding(3)
+var<uniform> params: Params;
+
+fn load_weight(idx: u32) -> f32 {
+    #if defined(WEIGHT_F32)
+        return weights[idx];
+    #elif defined(WEIGHT_F16)
+        return f32(weights[idx]);
+    #endif
+}
+fn load_input(idx: u32) -> f32 {
+    #if defined(INPUT_F32)
+        return input[idx];
+    #elif defined(INPUT_F16)
+        return f32(input[idx]);
+    #endif
+}
+fn store_output(idx: u32, val: f32) {
+    #if defined(OUTPUT_F32)
+        output[idx] = val;
+    #elif defined(OUTPUT_F16)
+        output[idx] = f16(val);
+    #endif
+}
+
+#if defined(WHCN)
+// Input/output/kernel contiguous in [W, H, C, N] order (kernel [KW,KH,C]).
+fn conv_2d_dw(idx: u32) -> f32 {
+    let i0    = idx / params.dst_w;
+    let dst_x = idx - i0 * params.dst_w;
+    let i1    = i0 / params.dst_h;
+    let dst_y = i0 - i1 * params.dst_h;
+    let n     = i1 / params.channels;
+    let c     = i1 - n * params.channels;
+
+    let src_i = params.offset_i + n * params.channels * params.src_h * params.src_w
+                                + c * params.src_h * params.src_w;
+    let knl_i = params.offset_w + c * params.knl_h * params.knl_w;
+
+    var sum: f32 = 0.0;
+    for (var ky: u32 = 0u; ky < params.knl_h; ky += 1u) {
+        let src_y = i32(dst_y) * params.stride_y + i32(ky) * params.dilation_y - params.pad_y;
+        if (src_y < 0 || src_y >= i32(params.src_h)) { continue; }
+        for (var kx: u32 = 0u; kx < params.knl_w; kx += 1u) {
+            let src_x = i32(dst_x) * params.stride_x + i32(kx) * params.dilation_x - params.pad_x;
+            if (src_x < 0 || src_x >= i32(params.src_w)) { continue; }
+            let v = load_input(src_i + u32(src_y) * params.src_w + u32(src_x));
+            let k = load_weight(knl_i + ky * params.knl_w + kx);
+            sum += v * k;
+        }
+    }
+    return sum;
+}
+#else
+// Channels contiguous (CWHN): channel is the innermost axis.
+fn conv_2d_dw(idx: u32) -> f32 {
+    let i0    = idx / params.channels;
+    let c     = idx - i0 * params.channels;
+    let i1    = i0 / params.dst_w;
+    let dst_x = i0 - i1 * params.dst_w;
+    let n     = i1 / params.dst_h;
+    let dst_y = i1 - n * params.dst_h;
+
+    let src_i   = params.offset_i + n * params.channels * params.src_h * params.src_w;
+    let src_row = params.src_w * params.channels;
+    let knl_row = params.knl_w * params.channels;
+
+    var sum: f32 = 0.0;
+    for (var ky: u32 = 0u; ky < params.knl_h; ky += 1u) {
+        let src_y = i32(dst_y) * params.stride_y + i32(ky) * params.dilation_y - params.pad_y;
+        if (src_y < 0 || src_y >= i32(params.src_h)) { continue; }
+        for (var kx: u32 = 0u; kx < params.knl_w; kx += 1u) {
+            let src_x = i32(dst_x) * params.stride_x + i32(kx) * params.dilation_x - params.pad_x;
+            if (src_x < 0 || src_x >= i32(params.src_w)) { continue; }
+            let v = load_input(src_i + u32(src_y) * src_row + u32(src_x) * params.channels + c);
+            let k = load_weight(params.offset_w + ky * knl_row + kx * params.channels + c);
+            sum += v * k;
+        }
+    }
+    return sum;
+}
+#endif
+
+@compute @workgroup_size(WG_SIZE)
+fn main(
+    @builtin(global_invocation_id) gid: vec3<u32>,
+    @builtin(num_workgroups) num_wg: vec3<u32>
+) {
+    let idx = gid.x + (num_wg.x * u32(WG_SIZE)) * gid.y;
+    if (idx >= params.ne) { return; }
+    store_output(params.offset_o + idx, conv_2d_dw(idx));
+}

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"      },
@@ -665,7 +666,7 @@ static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = {
     {LLM_TENSOR_HC_FFN_SCALE,               {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
     {LLM_TENSOR_ATTN_COMPRESSOR_WKV,        {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
     {LLM_TENSOR_ATTN_COMPRESSOR_WGATE,      {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
-    {LLM_TENSOR_ATTN_COMPRESSOR_APE,        {LLM_TENSOR_LAYER_REPEATING, GGML_OP_ADD}},
+    {LLM_TENSOR_ATTN_COMPRESSOR_APE,        {LLM_TENSOR_LAYER_REPEATING, GGML_OP_GET_ROWS}},
     {LLM_TENSOR_ATTN_COMPRESSOR_NORM,       {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
     {LLM_TENSOR_ATTN_K_B,                   {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
     {LLM_TENSOR_ATTN_V_B,                   {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
@@ -832,7 +833,7 @@ static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = {
     {LLM_TENSOR_INDEXER_ATTN_Q_B,           {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
     {LLM_TENSOR_INDEXER_COMPRESSOR_WKV,     {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
     {LLM_TENSOR_INDEXER_COMPRESSOR_WGATE,   {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
-    {LLM_TENSOR_INDEXER_COMPRESSOR_APE,     {LLM_TENSOR_LAYER_REPEATING, GGML_OP_ADD}},
+    {LLM_TENSOR_INDEXER_COMPRESSOR_APE,     {LLM_TENSOR_LAYER_REPEATING, GGML_OP_GET_ROWS}},
     {LLM_TENSOR_INDEXER_COMPRESSOR_NORM,    {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
     {LLM_TENSOR_FFN_GATE_TID2EID,           {LLM_TENSOR_LAYER_REPEATING, GGML_OP_GET_ROWS}},
     {LLM_TENSOR_NEXTN_PROJ_PRE,             {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},

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-kv-cache-dsv4.cpp

@@ -22,7 +22,7 @@ static constexpr uint32_t DSV4_STATE_MAGIC         = 0x34565344; // DSV4
 static constexpr uint32_t DSV4_STATE_VERSION       = 1;
 static constexpr uint32_t DSV4_STATE_MODE_FULL     = 0;
 static constexpr uint32_t DSV4_STATE_MODE_PARTIAL  = 1;
-static constexpr uint32_t DSV4_K_CACHE_STATE_VER   = 1;
+static constexpr uint32_t DSV4_K_CACHE_STATE_VER   = 2;
 static constexpr uint32_t DSV4_COMP_STATE_VER      = 1;
 
 static uint32_t dsv4_comp_size(uint32_t kv_size, uint32_t ratio) {
@@ -38,6 +38,16 @@ static void dsv4_clear_tensor_stream(ggml_tensor * tensor, uint32_t stream) {
     ggml_backend_tensor_memset(tensor, 0, stream*stream_size, stream_size);
 }
 
+static uint32_t dsv4_state_n_used_k_rows(llama_pos pos_max, uint32_t ratio, uint32_t kv_size) {
+    if (pos_max < 0) {
+        return 0;
+    }
+
+    const uint64_t n_rows = ((uint64_t) pos_max + 1)/ratio;
+
+    return (uint32_t) std::min<uint64_t>(kv_size, n_rows);
+}
+
 static int64_t dsv4_stream_offset(uint32_t n_stream, llama_seq_id seq_id, uint32_t size) {
     if (n_stream <= 1) {
         return 0;
@@ -239,6 +249,7 @@ static void dsv4_state_dst_stream_range(
 static void dsv4_state_write_tensor_streams(
         llama_io_write_i & io,
         ggml_tensor      * tensor,
+        uint32_t           tensor_rows,
         uint32_t           n_rows,
         uint32_t           s0,
         uint32_t           ns) {
@@ -247,20 +258,31 @@ static void dsv4_state_write_tensor_streams(
     const uint64_t rows     = n_rows;
     const uint64_t row_size = ggml_row_size(tensor->type, tensor->ne[0]);
 
+    if (n_rows > tensor_rows) {
+        throw std::runtime_error("DSV4 state tensor row count exceeds storage");
+    }
+
     io.write(&type_i,   sizeof(type_i));
     io.write(&ne0,      sizeof(ne0));
     io.write(&rows,     sizeof(rows));
     io.write(&row_size, sizeof(row_size));
 
-    const size_t offset = (size_t) s0*n_rows*row_size;
-    const size_t size   = (size_t) ns*n_rows*row_size;
+    const size_t stream_stride = (size_t) tensor_rows*row_size;
+    const size_t size          = (size_t) n_rows*row_size;
+    if (size == 0) {
+        return;
+    }
 
-    io.write_tensor(tensor, offset, size);
+    for (uint32_t s = 0; s < ns; ++s) {
+        const size_t offset = (size_t) (s0 + s)*stream_stride;
+        io.write_tensor(tensor, offset, size);
+    }
 }
 
 static void dsv4_state_read_tensor_streams(
         llama_io_read_i & io,
         ggml_tensor     * tensor,
+        uint32_t          tensor_rows,
         uint32_t          n_rows,
         uint32_t          s0,
         uint32_t          ns) {
@@ -282,18 +304,28 @@ static void dsv4_state_read_tensor_streams(
     if (type_i != type_i_ref || ne0 != ne0_ref || rows != rows_ref || row_size != row_size_ref) {
         throw std::runtime_error("DSV4 state tensor metadata mismatch");
     }
+    if (n_rows > tensor_rows) {
+        throw std::runtime_error("DSV4 state tensor row count exceeds storage");
+    }
 
-    const size_t offset = (size_t) s0*n_rows*row_size;
-    const size_t size   = (size_t) ns*n_rows*row_size;
+    const size_t stream_stride = (size_t) tensor_rows*row_size;
+    const size_t size          = (size_t) n_rows*row_size;
+    if (size == 0) {
+        return;
+    }
 
-    io.read_tensor(tensor, offset, size);
+    for (uint32_t s = 0; s < ns; ++s) {
+        const size_t offset = (size_t) (s0 + s)*stream_stride;
+        io.read_tensor(tensor, offset, size);
+    }
 }
 
 static void dsv4_state_write_k_cache(
         llama_io_write_i    & io,
         const llama_kv_cache * kv,
         llama_seq_id          seq_id,
-        llama_state_seq_flags flags) {
+        llama_state_seq_flags flags,
+        uint32_t              n_rows) {
     GGML_UNUSED(flags);
 
     uint32_t s0;
@@ -305,14 +337,18 @@ static void dsv4_state_write_k_cache(
     const auto layer_ids = kv->get_layer_ids();
     const uint32_t n_layer = layer_ids.size();
 
+    if (n_rows > kv_size) {
+        throw std::runtime_error("DSV4 K-cache state row count exceeds cache size");
+    }
+
     io.write(&version, sizeof(version));
-    io.write(&kv_size, sizeof(kv_size));
+    io.write(&n_rows,  sizeof(n_rows));
     io.write(&ns,      sizeof(ns));
     io.write(&n_layer, sizeof(n_layer));
 
     for (uint32_t il : layer_ids) {
         io.write(&il, sizeof(il));
-        dsv4_state_write_tensor_streams(io, kv->get_k_storage(il), kv_size, s0, ns);
+        dsv4_state_write_tensor_streams(io, kv->get_k_storage(il), kv_size, n_rows, s0, ns);
     }
 }
 
@@ -324,19 +360,26 @@ static void dsv4_state_read_k_cache(
     GGML_UNUSED(flags);
 
     uint32_t version;
-    uint32_t kv_size_ref;
+    uint32_t n_rows_ref;
     uint32_t ns;
     uint32_t n_layer_ref;
 
     io.read(&version,     sizeof(version));
-    io.read(&kv_size_ref, sizeof(kv_size_ref));
+    io.read(&n_rows_ref,  sizeof(n_rows_ref));
     io.read(&ns,          sizeof(ns));
     io.read(&n_layer_ref, sizeof(n_layer_ref));
 
-    if (version != DSV4_K_CACHE_STATE_VER) {
+    if (version != 1 && version != DSV4_K_CACHE_STATE_VER) {
         throw std::runtime_error("DSV4 K-cache state version mismatch");
     }
-    if (kv_size_ref != kv->get_size()) {
+
+    const uint32_t kv_size = kv->get_size();
+    if (version == 1 && n_rows_ref != kv_size) {
+        LLAMA_LOG_INFO("kv size ref %d kv %d\n", n_rows_ref, kv_size);
+        throw std::runtime_error("DSV4 K-cache state size mismatch");
+    }
+    if (n_rows_ref > kv_size) {
+        LLAMA_LOG_INFO("kv rows ref %d kv %d\n", n_rows_ref, kv_size);
         throw std::runtime_error("DSV4 K-cache state size mismatch");
     }
 
@@ -355,7 +398,7 @@ static void dsv4_state_read_k_cache(
             throw std::runtime_error("DSV4 K-cache layer id mismatch");
         }
 
-        dsv4_state_read_tensor_streams(io, kv->get_k_storage(il), kv->get_size(), s0, ns);
+        dsv4_state_read_tensor_streams(io, kv->get_k_storage(il), kv_size, n_rows_ref, s0, ns);
     }
 }
 
@@ -882,8 +925,8 @@ void llama_dsv4_comp_state::state_write(llama_io_write_i & io, llama_seq_id seq_
     for (const auto & layer : layers) {
         io.write(&layer.il, sizeof(layer.il));
 
-        dsv4_state_write_tensor_streams(io, layer.kv,    state_size, s0, ns);
-        dsv4_state_write_tensor_streams(io, layer.score, state_size, s0, ns);
+        dsv4_state_write_tensor_streams(io, layer.kv,    state_size, state_size, s0, ns);
+        dsv4_state_write_tensor_streams(io, layer.score, state_size, state_size, s0, ns);
     }
 }
 
@@ -924,8 +967,8 @@ void llama_dsv4_comp_state::state_read(llama_io_read_i & io, llama_seq_id seq_id
             throw std::runtime_error("DSV4 compressor state layer id mismatch");
         }
 
-        dsv4_state_read_tensor_streams(io, layer.kv,    state_size, s0, ns);
-        dsv4_state_read_tensor_streams(io, layer.score, state_size, s0, ns);
+        dsv4_state_read_tensor_streams(io, layer.kv,    state_size, state_size, s0, ns);
+        dsv4_state_read_tensor_streams(io, layer.score, state_size, state_size, s0, ns);
     }
 }
 
@@ -1328,9 +1371,19 @@ void llama_kv_cache_dsv4::state_write(llama_io_write_i & io, llama_seq_id seq_id
     kv_raw->state_write(io, seq_id, flags);
 
     if (!partial_only) {
-        dsv4_state_write_k_cache(io, kv_csa.get(), seq_id, flags);
-        dsv4_state_write_k_cache(io, kv_hca.get(), seq_id, flags);
-        dsv4_state_write_k_cache(io, kv_lid.get(), seq_id, flags);
+        const llama_pos pos_max = seq_id >= 0 ? kv_raw->seq_pos_max(seq_id) : -1;
+
+        //FIXME : note that we conflate token positions with rows, which is not true for multi-modal case.
+        const uint32_t n_rows_csa = seq_id >= 0 ?
+            dsv4_state_n_used_k_rows(pos_max, DSV4_CSA_RATIO, kv_csa->get_size()) : kv_csa->get_size();
+        const uint32_t n_rows_hca = seq_id >= 0 ?
+            dsv4_state_n_used_k_rows(pos_max, DSV4_HCA_RATIO, kv_hca->get_size()) : kv_hca->get_size();
+        const uint32_t n_rows_lid = seq_id >= 0 ?
+            dsv4_state_n_used_k_rows(pos_max, DSV4_CSA_RATIO, kv_lid->get_size()) : kv_lid->get_size();
+
+        dsv4_state_write_k_cache(io, kv_csa.get(), seq_id, flags, n_rows_csa);
+        dsv4_state_write_k_cache(io, kv_hca.get(), seq_id, flags, n_rows_hca);
+        dsv4_state_write_k_cache(io, kv_lid.get(), seq_id, flags, n_rows_lid);
     }
 
     csa_state->state_write(io, seq_id, flags);
@@ -1366,6 +1419,10 @@ void llama_kv_cache_dsv4::state_read(llama_io_read_i & io, llama_seq_id seq_id,
     kv_raw->state_read(io, seq_id, flags);
 
     if (!partial_only) {
+        kv_csa->clear(true);
+        kv_hca->clear(true);
+        kv_lid->clear(true);
+
         dsv4_state_read_k_cache(io, kv_csa.get(), seq_id, flags);
         dsv4_state_read_k_cache(io, kv_hca.get(), seq_id, flags);
         dsv4_state_read_k_cache(io, kv_lid.get(), seq_id, flags);

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-backend-ops.cpp

@@ -5960,6 +5960,7 @@ enum MoeGatingFunc {
     GATING_FUNC_SOFTMAX,
     GATING_FUNC_SIGMOID,
     GATING_FUNC_SOFTMAX_WEIGHT,
+    GATING_FUNC_SQRT_SOFTPLUS,
 };
 
 struct test_topk_moe : public test_case {
@@ -6003,7 +6004,8 @@ struct test_topk_moe : public test_case {
         ggml_tensor * logits = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, ne.data());
         ggml_tensor * probs            =
             (gating_func == GATING_FUNC_SOFTMAX) ? ggml_soft_max(ctx, logits) :
-            (gating_func == GATING_FUNC_SIGMOID) ? ggml_sigmoid(ctx, logits) : logits;
+            (gating_func == GATING_FUNC_SIGMOID) ? ggml_sigmoid(ctx, logits) :
+            (gating_func == GATING_FUNC_SQRT_SOFTPLUS) ? ggml_sqrt(ctx, ggml_softplus(ctx, logits)) : logits;
         ggml_set_name(probs, "probs");
 
         ggml_tensor * selection_probs = probs;
@@ -9584,7 +9586,7 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
         }
     }
 
-    for (auto gate : {GATING_FUNC_SOFTMAX, GATING_FUNC_SIGMOID, GATING_FUNC_SOFTMAX_WEIGHT}) {
+    for (auto gate : {GATING_FUNC_SOFTMAX, GATING_FUNC_SIGMOID, GATING_FUNC_SOFTMAX_WEIGHT, GATING_FUNC_SQRT_SOFTPLUS}) {
         for (bool with_norm : {false, true}) {
             for (bool bias_probs : {false, true}) {
                 for (float scale_w : {0.0f, 2.0f}) {
@@ -9596,6 +9598,7 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
                     test_cases.emplace_back(new test_topk_moe({128, 1, 1, 1}, 128, with_norm, bias_probs, gate, scale_w));
                     test_cases.emplace_back(new test_topk_moe({129, 1, 1, 1}, 128, with_norm, bias_probs, gate, scale_w));
                     test_cases.emplace_back(new test_topk_moe({160, 4, 1, 1}, 160, with_norm, bias_probs, gate, scale_w));
+                    test_cases.emplace_back(new test_topk_moe({256, 22, 1, 1}, 6, with_norm, bias_probs, gate, scale_w)); // Used by DeepSeek-V4
                     test_cases.emplace_back(new test_topk_moe({288, 22, 1, 1}, 8, with_norm, bias_probs, gate, scale_w)); // Used by StepFun 3.7
                 }
             }

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;

tools/mtmd/models/qwen3vl.cpp

@@ -37,7 +37,7 @@ ggml_cgraph * clip_graph_qwen3vl::build() {
     }
 
     // calculate absolute position embedding and apply
-    ggml_tensor * learned_pos_embd = resize_position_embeddings();
+    ggml_tensor * learned_pos_embd = resize_position_embeddings(GGML_SCALE_MODE_BILINEAR | GGML_SCALE_FLAG_ALIGN_CORNERS);
     learned_pos_embd = ggml_cont_4d(
         ctx0, learned_pos_embd,
         n_embd * 2, n_patches_x / 2, n_patches_y, batch_size);

tools/server/server-context.cpp

@@ -1152,6 +1152,11 @@ private:
             return false;
         }
 
+        if (ctx_tgt == nullptr) {
+            SRV_ERR("failed to create_context with model '%s'\n", params_base.model.path.c_str());
+            return false;
+        }
+
         vocab = llama_model_get_vocab(model_tgt);
 
         n_ctx = llama_n_ctx(ctx_tgt);

tools/server/server-stream.cpp

@@ -632,6 +632,13 @@ void server_res_spipe::on_complete() {
     if (!spipe || next_finished) {
         return;
     }
+    // an empty next_orig means set_next() never ran: the request failed before streaming
+    // started, typically a params validation throw. evict the session installed by set_req()
+    // so the failed request leaves nothing behind for discovery or replay
+    if (!next_orig) {
+        g_stream_sessions.evict(server_stream_conv_id_from_headers(req->headers));
+        return;
+    }
     std::string chunk;
     while (!spipe->is_cancelled()) {
         chunk.clear();

tools/ui/package-lock.json

@@ -12,7 +12,7 @@
 				"@eslint/compat": "1.4.1",
 				"@eslint/js": "9.39.2",
 				"@internationalized/date": "3.12.2",
-				"@lucide/svelte": "0.515.0",
+				"@lucide/svelte": "1.25.0",
 				"@modelcontextprotocol/sdk": "1.26.0",
 				"@playwright/test": "1.56.1",
 				"@storybook/addon-a11y": "10.2.4",
@@ -3065,9 +3065,9 @@
 			}
 		},
 		"node_modules/@lucide/svelte": {
-			"version": "0.515.0",
-			"resolved": "https://registry.npmjs.org/@lucide/svelte/-/svelte-0.515.0.tgz",
-			"integrity": "sha512-CEAyqcZmNBfYzVgaRmK2RFJP5tnbXxekRyDk0XX/eZQRfsJmkDvmQwXNX8C869BgNeryzmrRyjHhUL6g9ZOHNA==",
+			"version": "1.25.0",
+			"resolved": "https://registry.npmjs.org/@lucide/svelte/-/svelte-1.25.0.tgz",
+			"integrity": "sha512-v9m+dD68jxVnqkU3K59mG/RSRFlPGzmKCGSyMfnXcaGv9jODDQMyQkcp1CGvk3Y/cUj9v7f8rw1n//K0B53xGQ==",
 			"dev": true,
 			"license": "ISC",
 			"peerDependencies": {

tools/ui/package.json

@@ -31,7 +31,7 @@
 		"@eslint/compat": "1.4.1",
 		"@eslint/js": "9.39.2",
 		"@internationalized/date": "3.12.2",
-		"@lucide/svelte": "0.515.0",
+		"@lucide/svelte": "1.25.0",
 		"@modelcontextprotocol/sdk": "1.26.0",
 		"@playwright/test": "1.56.1",
 		"@storybook/addon-a11y": "10.2.4",

tools/ui/src/lib/components/app/actions/ActionIcon.svelte

@@ -66,7 +66,14 @@
 		<Tooltip.Trigger>
 			<!-- prevent another nested button element -->
 			{#snippet child({ props })}
-				{@render button(props)}
+				{#if disabled}
+					<!-- disabled buttons have pointer-events:none; wrap in a span so the tooltip hover surface stays alive -->
+					<span {...props}>
+						{@render button({})}
+					</span>
+				{:else}
+					{@render button(props)}
+				{/if}
 			{/snippet}
 		</Tooltip.Trigger>
 

tools/ui/src/lib/components/app/chat/ChatMessages/ChatMessage/ChatMessageSystem/ChatMessageSystem.svelte

@@ -157,10 +157,7 @@
 						>
 							{#if currentConfig.renderUserContentAsMarkdown}
 								<div bind:this={messageElement} class={isExpanded ? 'cursor-text' : ''}>
-									<MarkdownContent
-										class="markdown-system-content -my-4"
-										content={message.content}
-									/>
+									<MarkdownContent class="markdown-system-content" content={message.content} />
 								</div>
 							{:else}
 								<span

tools/ui/src/lib/components/app/chat/ChatMessages/ChatMessage/ChatMessageUser/ChatMessageUserBubble.svelte

@@ -65,7 +65,7 @@
 	>
 		{#if renderMarkdown && currentConfig.renderUserContentAsMarkdown}
 			<div bind:this={messageElement}>
-				<MarkdownContent class="markdown-user-content -my-4" {content} />
+				<MarkdownContent class="markdown-user-content" {content} />
 			</div>
 		{:else}
 			<span bind:this={messageElement} class="text-md whitespace-pre-wrap">

tools/ui/src/lib/components/app/chat/ChatMessages/ChatMessageAgenticContent.svelte

@@ -43,6 +43,7 @@
 
 	const renderThinkingAsMarkdown = $derived(config().renderThinkingAsMarkdown as boolean);
 	const showThoughtInProgress = $derived(Boolean(config().showThoughtInProgress));
+	const alwaysShowToolCallContent = $derived(Boolean(config().alwaysShowToolCallContent));
 	const showMessageStats = $derived(Boolean(config().showMessageStats));
 	const showAgenticTurnStats = $derived(showMessageStats && Boolean(config().showAgenticTurnStats));
 
@@ -98,19 +99,6 @@
 		isStreaming ? agenticExecutingToolCallId(message.convId) : null
 	);
 
-	// Skip sections the user manually collapsed - we never override an explicit false.
-	let lastSeenExecutingToolCallId: string | null = null;
-	$effect(() => {
-		const current = currentlyExecutingToolCallId;
-		const previous = lastSeenExecutingToolCallId;
-		lastSeenExecutingToolCallId = current;
-		if (!current || current === previous) return;
-		const idx = sections.findIndex((s) => s.toolCallId === current);
-		if (idx >= 0 && expandedStates[idx] === undefined) {
-			expandedStates[idx] = true;
-		}
-	});
-
 	type TurnGroup = {
 		sections: AgenticSection[];
 		flatIndices: number[];
@@ -149,10 +137,11 @@
 
 	function getDefaultExpanded(section: AgenticSection): boolean {
 		if (
+			section.type === AgenticSectionType.TOOL_CALL ||
 			section.type === AgenticSectionType.TOOL_CALL_PENDING ||
 			section.type === AgenticSectionType.TOOL_CALL_STREAMING
 		) {
-			return false;
+			return alwaysShowToolCallContent;
 		}
 
 		if (section.type === AgenticSectionType.REASONING_PENDING) {

tools/ui/src/lib/components/app/dialogs/DialogConversationRename.svelte

@@ -0,0 +1,85 @@
+<script lang="ts">
+	import * as AlertDialog from '$lib/components/ui/alert-dialog';
+	import { Button } from '$lib/components/ui/button';
+	import { Input } from '$lib/components/ui/input';
+	import { Pencil } from '@lucide/svelte';
+
+	interface Props {
+		open: boolean;
+		currentTitle: string;
+		value: string;
+		onConfirm: () => void;
+		onCancel: () => void;
+	}
+
+	let {
+		open = $bindable(),
+		currentTitle,
+		value = $bindable(''),
+		onConfirm,
+		onCancel
+	}: Props = $props();
+
+	let inputRef = $state<HTMLInputElement | null>(null);
+
+	const canSubmit = $derived(value.trim().length > 0 && value.trim() !== currentTitle.trim());
+
+	$effect(() => {
+		if (open) {
+			value = currentTitle;
+			queueMicrotask(() => {
+				inputRef?.focus();
+				inputRef?.select();
+			});
+		}
+	});
+
+	function handleOpenChange(newOpen: boolean) {
+		if (!newOpen) {
+			onCancel();
+		}
+	}
+
+	function handleSubmit(event: Event) {
+		event.preventDefault();
+		if (!canSubmit) return;
+		value = value.trim();
+		onConfirm();
+	}
+</script>
+
+<AlertDialog.Root bind:open onOpenChange={handleOpenChange}>
+	<AlertDialog.Content>
+		<AlertDialog.Header>
+			<AlertDialog.Title class="flex items-center gap-2">
+				<Pencil class="h-5 w-5" />
+				Rename conversation
+			</AlertDialog.Title>
+
+			<AlertDialog.Description>Choose a new title for this conversation.</AlertDialog.Description>
+		</AlertDialog.Header>
+
+		<form onsubmit={handleSubmit} class="space-y-2 pt-2 pb-4">
+			<label for="conversation-rename-input" class="text-sm font-medium text-muted-foreground">
+				Conversation title
+			</label>
+
+			<Input
+				id="conversation-rename-input"
+				bind:ref={inputRef}
+				bind:value
+				placeholder="Conversation title"
+				maxlength={200}
+				autocomplete="off"
+				autocorrect="off"
+				spellcheck={false}
+			/>
+		</form>
+
+		<AlertDialog.Footer>
+			<AlertDialog.Cancel>Cancel</AlertDialog.Cancel>
+
+			<Button type="button" onclick={handleSubmit} disabled={!canSubmit}>Save</Button>
+		</AlertDialog.Footer>
+	</AlertDialog.Content>
+</AlertDialog.Root>

tools/ui/src/lib/components/app/dialogs/DialogConversationSelection.svelte

@@ -37,9 +37,9 @@
 
 <Dialog.Root bind:open>
 	<Dialog.Portal>
-		<Dialog.Overlay class="z-[1000000]" />
+		<Dialog.Overlay class="z-1000000" />
 
-		<Dialog.Content class="z-[1000001] max-w-2xl">
+		<Dialog.Content class="z-1000001 max-w-2xl">
 			<Dialog.Header>
 				<Dialog.Title>
 					Select Conversations to {mode === 'export' ? 'Export' : 'Import'}
@@ -58,6 +58,7 @@
 
 			<ConversationSelection
 				bind:this={conversationSelectionRef}
+				isOpen={open}
 				{conversations}
 				{messageCountMap}
 				{mode}

tools/ui/src/lib/components/app/dialogs/DialogConversationTitleUpdate.svelte

@@ -1,46 +0,0 @@
-<script lang="ts">
-	import * as AlertDialog from '$lib/components/ui/alert-dialog';
-	import { Button } from '$lib/components/ui/button';
-
-	interface Props {
-		open: boolean;
-		currentTitle: string;
-		newTitle: string;
-		onConfirm: () => void;
-		onCancel: () => void;
-	}
-
-	let { open = $bindable(), currentTitle, newTitle, onConfirm, onCancel }: Props = $props();
-</script>
-
-<AlertDialog.Root bind:open>
-	<AlertDialog.Content>
-		<AlertDialog.Header>
-			<AlertDialog.Title>Update Conversation Title?</AlertDialog.Title>
-
-			<AlertDialog.Description>
-				Do you want to update the conversation title to match the first message content?
-			</AlertDialog.Description>
-		</AlertDialog.Header>
-
-		<div class="space-y-4 pt-2 pb-6">
-			<div class="space-y-2">
-				<p class="text-sm font-medium text-muted-foreground">Current title:</p>
-
-				<p class="rounded-md bg-muted/50 p-3 text-sm font-medium">{currentTitle}</p>
-			</div>
-
-			<div class="space-y-2">
-				<p class="text-sm font-medium text-muted-foreground">New title would be:</p>
-
-				<p class="rounded-md bg-muted/50 p-3 text-sm font-medium">{newTitle}</p>
-			</div>
-		</div>
-
-		<AlertDialog.Footer>
-			<Button variant="outline" onclick={onCancel}>Keep Current Title</Button>
-
-			<Button onclick={onConfirm}>Update Title</Button>
-		</AlertDialog.Footer>
-	</AlertDialog.Content>
-</AlertDialog.Root>

tools/ui/src/lib/components/app/dialogs/DialogExportSettings.svelte

@@ -68,6 +68,7 @@
 
 		<AlertDialog.Footer>
 			<AlertDialog.Cancel onclick={onCancel}>Cancel</AlertDialog.Cancel>
+
 			<AlertDialog.Action
 				onclick={onConfirm}
 				class="bg-destructive text-white hover:bg-destructive/80"

tools/ui/src/lib/components/app/dialogs/index.ts

@@ -92,33 +92,36 @@ export { default as DialogExportSettings } from './DialogExportSettings.svelte';
 export { default as DialogConfirmation } from './DialogConfirmation.svelte';
 
 /**
- * **DialogConversationTitleUpdate** - Conversation rename confirmation
+ * **DialogConversationRename** - Rename a conversation
  *
- * Confirmation dialog shown when editing the first user message in a conversation.
- * Asks user whether to update the conversation title to match the new message content.
+ * Modal dialog for renaming a conversation. Replaces the prior
+ * `window.prompt()`-based flow with a styled, accessible AlertDialog
+ * containing an editable input. Triggered from the sidebar conversation
+ * item's "Edit" action.
  *
  * **Architecture:**
  * - Uses ShadCN AlertDialog
- * - Shows current vs proposed title comparison
- * - Triggered by ChatMessages when first message is edited
+ * - Bindable `value` keeps the new title in sync with parent state
+ * - Submit is gated on a non-empty trimmed value that differs from the current title
  *
  * **Features:**
- * - Side-by-side display of current and new title
- * - "Keep Current Title" and "Update Title" action buttons
- * - Styled title previews in muted background boxes
+ * - Autofocus on open with text selected for quick overwrite
+ * - Disabled Save button when value is empty or unchanged
+ * - Trim-on-submit normalization
+ * - Cancel via AlertDialog.Cancel or `onOpenChange(false)`
  *
  * @example
  * ```svelte
- * <DialogConversationTitleUpdate
- *   bind:open={showTitleUpdate}
+ * <DialogConversationRename
+ *   bind:open={showRename}
  *   currentTitle={conversation.name}
- *   newTitle={truncatedMessageContent}
- *   onConfirm={updateTitle}
- *   onCancel={() => showTitleUpdate = false}
+ *   bind:value={renameDraft}
+ *   onConfirm={handleRenameConfirm}
+ *   onCancel={() => (showRename = false)}
  * />
  * ```
  */
-export { default as DialogConversationTitleUpdate } from './DialogConversationTitleUpdate.svelte';
+export { default as DialogConversationRename } from './DialogConversationRename.svelte';
 
 /**
  *

tools/ui/src/lib/components/app/misc/ConversationSelection.svelte

@@ -4,6 +4,7 @@
 	import SearchInput from '$lib/components/app/forms/SearchInput.svelte';
 	import { ScrollArea } from '$lib/components/ui/scroll-area';
 	import { SvelteSet } from 'svelte/reactivity';
+	import { useMarqueeSelection } from '$lib/hooks/use-marquee-selection.svelte';
 
 	interface Props {
 		conversations: DatabaseConversation[];
@@ -11,13 +12,20 @@
 		mode: 'export' | 'import';
 		onCancel: () => void;
 		onConfirm: (selectedConversations: DatabaseConversation[]) => void;
+		isOpen?: boolean;
 	}
 
-	let { conversations, messageCountMap = new Map(), mode, onCancel, onConfirm }: Props = $props();
+	let {
+		conversations,
+		messageCountMap = new Map(),
+		mode,
+		onCancel,
+		onConfirm,
+		isOpen = true
+	}: Props = $props();
 
 	let searchQuery = $state('');
 	let selectedIds = $state.raw<SvelteSet<string>>(getInitialSelectedIds());
-	let lastClickedId = $state<string | null>(null);
 
 	function getInitialSelectedIds(): SvelteSet<string> {
 		return new SvelteSet(conversations.map((c) => c.id));
@@ -30,6 +38,8 @@
 		})
 	);
 
+	let orderedIds = $derived(filteredConversations.map((c) => c.id));
+
 	let allSelected = $derived(
 		filteredConversations.length > 0 &&
 			filteredConversations.every((conv) => selectedIds.has(conv.id))
@@ -39,54 +49,20 @@
 		filteredConversations.some((conv) => selectedIds.has(conv.id)) && !allSelected
 	);
 
-	function toggleConversation(id: string, shiftKey: boolean = false) {
-		const newSet = new SvelteSet(selectedIds);
-
-		if (shiftKey && lastClickedId !== null) {
-			const lastIndex = filteredConversations.findIndex((c) => c.id === lastClickedId);
-			const currentIndex = filteredConversations.findIndex((c) => c.id === id);
-
-			if (lastIndex !== -1 && currentIndex !== -1) {
-				const start = Math.min(lastIndex, currentIndex);
-				const end = Math.max(lastIndex, currentIndex);
-
-				const shouldSelect = !newSet.has(id);
-
-				for (let i = start; i <= end; i++) {
-					if (shouldSelect) {
-						newSet.add(filteredConversations[i].id);
-					} else {
-						newSet.delete(filteredConversations[i].id);
-					}
-				}
-
-				selectedIds = newSet;
-				return;
-			}
-		}
-
-		if (newSet.has(id)) {
-			newSet.delete(id);
-		} else {
-			newSet.add(id);
-		}
-
-		selectedIds = newSet;
-		lastClickedId = id;
-	}
+	const marquee = useMarqueeSelection({
+		selectedIds: () => selectedIds,
+		orderedIds: () => orderedIds,
+		enabled: () => isOpen
+	});
 
 	function toggleAll() {
+		const newSet = new SvelteSet(selectedIds);
 		if (allSelected) {
-			const newSet = new SvelteSet(selectedIds);
-
 			filteredConversations.forEach((conv) => newSet.delete(conv.id));
-			selectedIds = newSet;
 		} else {
-			const newSet = new SvelteSet(selectedIds);
-
 			filteredConversations.forEach((conv) => newSet.add(conv.id));
-			selectedIds = newSet;
 		}
+		selectedIds = newSet;
 	}
 
 	function handleConfirm() {
@@ -97,7 +73,7 @@
 	function handleCancel() {
 		selectedIds = getInitialSelectedIds();
 		searchQuery = '';
-		lastClickedId = null;
+		marquee.reset();
 
 		onCancel();
 	}
@@ -105,7 +81,7 @@
 	export function reset() {
 		selectedIds = getInitialSelectedIds();
 		searchQuery = '';
-		lastClickedId = null;
+		marquee.reset();
 	}
 </script>
 
@@ -122,7 +98,7 @@
 	</div>
 
 	<div class="overflow-hidden rounded-md border">
-		<ScrollArea class="h-[400px]">
+		<ScrollArea class="h-100">
 			<table class="w-full">
 				<thead class="sticky top-0 z-10 bg-muted">
 					<tr class="border-b">
@@ -139,6 +115,7 @@
 						<th class="w-32 p-3 text-left text-sm font-medium">Messages</th>
 					</tr>
 				</thead>
+
 				<tbody>
 					{#if filteredConversations.length === 0}
 						<tr>
@@ -152,23 +129,28 @@
 						</tr>
 					{:else}
 						{#each filteredConversations as conv (conv.id)}
+							{@const checked = selectedIds.has(conv.id)}
 							<tr
-								class="cursor-pointer border-b transition-colors hover:bg-muted/50"
-								onclick={(event) => toggleConversation(conv.id, event.shiftKey)}
+								class="cursor-pointer border-b transition-colors hover:bg-muted/50 {checked
+									? 'bg-muted/75'
+									: ''}"
+								data-conversation-row={conv.id}
+								onmousedown={(event) => marquee.rowMouseDown(conv.id, event)}
+								onclick={(event) => marquee.rowClick(conv.id, event.shiftKey)}
 							>
 								<td class="p-3">
 									<Checkbox
-										checked={selectedIds.has(conv.id)}
+										{checked}
 										onclick={(event) => {
 											event.preventDefault();
 											event.stopPropagation();
-											toggleConversation(conv.id, event.shiftKey);
+											marquee.rowClick(conv.id, event.shiftKey);
 										}}
 									/>
 								</td>
 
 								<td class="p-3 text-sm">
-									<div class="max-w-[17rem] truncate" title={conv.name || 'Untitled conversation'}>
+									<div class="max-w-68 truncate" title={conv.name || 'Untitled conversation'}>
 										{conv.name || 'Untitled conversation'}
 									</div>
 								</td>

tools/ui/src/lib/components/app/navigation/SidebarNavigation/SidebarNavigation.svelte

@@ -4,15 +4,22 @@
 	import { PanelLeftClose, PanelLeftOpen, X } from '@lucide/svelte';
 	import {
 		ActionIcon,
+		DialogConversationRename,
 		Logo,
 		SidebarNavigationConversationList,
 		SidebarNavigationActions
 	} from '$lib/components/app';
 	import { ROUTES } from '$lib/constants';
 	import { fade } from 'svelte/transition';
+	import { SvelteSet } from 'svelte/reactivity';
+	import { useMarqueeSelection } from '$lib/hooks/use-marquee-selection.svelte';
 
 	import { useKeyboardShortcuts } from '$lib/hooks/use-keyboard-shortcuts.svelte';
-	import { conversationsStore, conversations } from '$lib/stores/conversations.svelte';
+	import {
+		buildConversationTree,
+		conversationsStore,
+		conversations
+	} from '$lib/stores/conversations.svelte';
 	import { chatStore } from '$lib/stores/chat.svelte';
 	import { config } from '$lib/stores/settings.svelte';
 	import { RouterService } from '$lib/services/router.service';
@@ -40,7 +47,6 @@
 	const isOnMobile = $derived(isMobile.current);
 	const alwaysShowOnDesktop = $derived(config().alwaysShowSidebarOnDesktop as boolean);
 
-	// Keep the sidebar expanded on desktop when the user pins it open
 	$effect(() => {
 		if (alwaysShowOnDesktop && !isOnMobile) {
 			isExpandedMode = true;
@@ -58,13 +64,11 @@
 		if (!isExpandedMode) {
 			isSearchModeActive = false;
 			searchQuery = '';
+			if (isSelectionMode) exitSelectionMode();
 			cancelMobileCollapse();
 		}
 	});
 
-	// On mobile the dedicated /search route hides the sidebar (see the aside
-	// render guard below). Collapse it as we enter /search so it doesn't
-	// reappear expanded when the user navigates back via the back button.
 	$effect(() => {
 		if (isMobile.current && page.url.hash.includes(ROUTES.SEARCH)) {
 			isExpandedMode = false;
@@ -89,6 +93,121 @@
 		return conversations();
 	});
 
+	let isSelectionMode = $state(false);
+	let selectedIds = new SvelteSet<string>();
+
+	let renameDialogOpen = $state(false);
+	let renameTargetConversationId = $state<string | null>(null);
+	let renameDraft = $state('');
+	let renameOriginalTitle = $state('');
+
+	const renderedOrderIds = $derived(
+		buildConversationTree(filteredConversations).map((t) => t.conversation.id)
+	);
+
+	const allSelectedArePinned = $derived.by(() => {
+		if (selectedIds.size === 0) return false;
+		const convs = conversations();
+		for (const id of selectedIds) {
+			const c = convs.find((conv) => conv.id === id);
+			if (c && !c.pinned) return false;
+		}
+		return true;
+	});
+
+	const pinStateIsMixed = $derived.by(() => {
+		if (selectedIds.size === 0) return false;
+		const convs = conversations();
+		let anyPinned = false;
+		let anyUnpinned = false;
+		for (const id of selectedIds) {
+			const c = convs.find((conv) => conv.id === id);
+			if (!c) continue;
+			if (c.pinned) anyPinned = true;
+			else anyUnpinned = true;
+			if (anyPinned && anyUnpinned) return true;
+		}
+		return false;
+	});
+
+	const visibleSelectionStats = $derived.by(() => {
+		const visibleIds = filteredConversations.map((c) => c.id);
+		let selectedVisible = 0;
+		for (const id of visibleIds) {
+			if (selectedIds.has(id)) selectedVisible++;
+		}
+		return {
+			visibleCount: visibleIds.length,
+			selectedVisibleCount: selectedVisible
+		};
+	});
+
+	function enterSelectionMode(id?: string) {
+		isSelectionMode = true;
+		if (id !== undefined) {
+			selectedIds.add(id);
+		}
+	}
+
+	function exitSelectionMode() {
+		isSelectionMode = false;
+		selectedIds.clear();
+	}
+
+	function toggleSelected(id: string) {
+		if (selectedIds.has(id)) {
+			selectedIds.delete(id);
+		} else {
+			selectedIds.add(id);
+		}
+	}
+
+	function toggleSelectAllVisible() {
+		const visibleIds = filteredConversations.map((c) => c.id);
+		const allSelected = visibleIds.length > 0 && visibleIds.every((id) => selectedIds.has(id));
+
+		if (allSelected) {
+			for (const id of visibleIds) selectedIds.delete(id);
+		} else {
+			for (const id of visibleIds) selectedIds.add(id);
+		}
+	}
+
+	async function handleBulkDelete() {
+		const ids = Array.from(selectedIds);
+		if (ids.length === 0) return;
+		await conversationsStore.bulkDeleteConversations(ids);
+		exitSelectionMode();
+	}
+
+	async function handleBulkPinToggle() {
+		const ids = Array.from(selectedIds);
+		if (ids.length === 0) return;
+		await conversationsStore.bulkToggleConversationPin(ids);
+	}
+
+	async function handleBulkExport() {
+		const ids = Array.from(selectedIds);
+		if (ids.length === 0) return;
+		await conversationsStore.bulkExportConversations(ids);
+	}
+
+	const marquee = useMarqueeSelection({
+		selectedIds: () => selectedIds,
+		orderedIds: () => renderedOrderIds,
+		enabled: () => isSelectionMode
+	});
+
+	function handleRowMouseDown(id: string, event: MouseEvent) {
+		if (!isSelectionMode) return;
+		marquee.rowMouseDown(id, event);
+	}
+
+	function handleSelectionClick(id: string, options: { shiftKey: boolean }): void {
+		if (!isSelectionMode) return;
+		marquee.rowClick(id, options.shiftKey);
+	}
+
 	async function selectConversation(id: string) {
 		if (isMobile.current) {
 			scheduleMobileCollapse();
@@ -100,10 +219,30 @@
 		const conversation = conversations().find((conv) => conv.id === id);
 		if (!conversation) return;
 
-		const newName = window.prompt('Rename conversation', conversation.name);
-		if (newName && newName.trim()) {
-			await conversationsStore.updateConversationName(id, newName.trim());
-		}
+		renameTargetConversationId = id;
+		renameOriginalTitle = conversation.name;
+		renameDraft = conversation.name;
+		renameDialogOpen = true;
+	}
+
+	async function handleRenameConfirm() {
+		const id = renameTargetConversationId;
+		if (!id) return;
+
+		const nextName = renameDraft.trim();
+		if (!nextName || nextName === renameOriginalTitle.trim()) return;
+
+		await conversationsStore.updateConversationName(id, nextName);
+
+		renameDialogOpen = false;
+		renameTargetConversationId = null;
+	}
+
+	function handleRenameCancel() {
+		renameDialogOpen = false;
+		renameTargetConversationId = null;
+		renameDraft = '';
+		renameOriginalTitle = '';
 	}
 
 	async function handleDeleteConversation(id: string) {
@@ -148,9 +287,7 @@
 {#if innerWidth > 768 || (!page.url.hash.includes(ROUTES.SETTINGS) && !page.url.hash.includes(ROUTES.MCP_SERVERS) && !page.url.hash.includes(ROUTES.SEARCH))}
 	<aside
 		class={[
-			// Layout & positioning
 			'fixed md:sticky top-2 left-2 md:left-0 md:ml-2 md:mt-2 pt-2 z-10 w-[calc(100dvw-1rem)]',
-			// Dimensions & overflow
 			'md:h-[calc(100dvh-1.125rem)]',
 			isExpandedMode &&
 				(device.isStandalone
@@ -158,17 +295,11 @@
 					: device.isIOSDevice
 						? 'h-[calc(100dvh-0.5rem)]'
 						: 'h-[calc(100dvh-1rem)]'),
-			// Shape & depth
 			'rounded-3xl md:rounded-2xl',
-			// Flex layout
 			'flex flex-col justify-between',
-			// Transition
 			'md:transition-[width,padding] duration-200 ease-out',
-			// Expanded state: width, surface, depth
 			isStripExpanded && 'md:w-72 md:bg-muted/60 md:backdrop-blur-xl border-border shadow-md',
-			// Collapsed state
 			!isStripExpanded && 'md:w-12',
-			// Expanded mode flag (for mobile ::before overlay)
 			isExpandedMode && 'is-expanded'
 		]}
 	>
@@ -218,52 +349,78 @@
 			{/if}
 		</div>
 
-		<div class="mt-2 flex min-h-0 flex-1 flex-col gap-4 md:gap-1 overflow-y-auto">
-			<div
-				class="flex min-h-0 flex-1 flex-col gap-4 md:gap-1 {isMobile.current
-					? 'transition-[opacity,height] duration-200 ease-out'
-					: ''} {isMobile.current && !isExpandedMode ? 'opacity-0 !h-0' : ''}"
-				in:fade={{ duration: 200 }}
-				out:fade={{ duration: 200 }}
-			>
-				<SidebarNavigationActions
-					isExpandedMode={innerWidth > 768 ? isExpandedMode : true}
-					class="px-2"
-					bind:isSearchModeActive
-					bind:searchQuery
-					onSearchDeactivated={() => {
-						isSearchModeActive = false;
-						searchQuery = '';
-					}}
-					onSearchClick={() => {
-						isExpandedMode = true;
-						isSearchModeActive = true;
-					}}
-					onNewChat={() => {
-						if (isMobile.current) {
-							scheduleMobileCollapse();
-						}
-					}}
-				/>
+		<div
+			class="mt-2 flex min-h-0 flex-1 flex-col gap-4 md:gap-1 {isMobile.current
+				? 'transition-[opacity,height] duration-200 ease-out'
+				: ''} {isMobile.current && !isExpandedMode ? 'opacity-0 !h-0' : ''}"
+			in:fade={{ duration: 200 }}
+			out:fade={{ duration: 200 }}
+		>
+			<SidebarNavigationActions
+				isExpandedMode={innerWidth > 768 ? isExpandedMode : true}
+				class="px-2"
+				bind:isSearchModeActive
+				bind:searchQuery
+				onSearchDeactivated={() => {
+					isSearchModeActive = false;
+					searchQuery = '';
+				}}
+				onSearchClick={() => {
+					isExpandedMode = true;
+					isSearchModeActive = true;
+				}}
+				onNewChat={() => {
+					if (isMobile.current) {
+						scheduleMobileCollapse();
+					}
+				}}
+			/>
 
-				{#if isExpandedMode || isOnMobile}
+			{#if isExpandedMode || isOnMobile}
+				<div class="flex min-h-0 flex-1 flex-col overflow-y-auto">
 					<SidebarNavigationConversationList
 						class="px-2"
 						{filteredConversations}
 						{currentChatId}
 						{isSearchModeActive}
 						{searchQuery}
+						{isSelectionMode}
+						{selectedIds}
 						onSelect={selectConversation}
 						onEdit={handleEditConversation}
 						onDelete={handleDeleteConversation}
 						onStop={handleStopGeneration}
+						onToggleSelect={toggleSelected}
+						onEnterSelectionMode={enterSelectionMode}
+						onSelectionClick={handleSelectionClick}
+						onRowMouseDown={handleRowMouseDown}
+						visibleCount={visibleSelectionStats.visibleCount}
+						allVisibleSelected={visibleSelectionStats.visibleCount > 0 &&
+							visibleSelectionStats.selectedVisibleCount === visibleSelectionStats.visibleCount}
+						someVisibleSelected={visibleSelectionStats.selectedVisibleCount > 0 &&
+							visibleSelectionStats.selectedVisibleCount < visibleSelectionStats.visibleCount}
+						{allSelectedArePinned}
+						{pinStateIsMixed}
+						onSelectAllToggle={toggleSelectAllVisible}
+						onBulkPinToggle={handleBulkPinToggle}
+						onBulkExport={handleBulkExport}
+						onBulkDelete={handleBulkDelete}
+						onCloseSelection={exitSelectionMode}
 					/>
-				{/if}
-			</div>
+				</div>
+			{/if}
 		</div>
 	</aside>
 {/if}
 
+<DialogConversationRename
+	bind:open={renameDialogOpen}
+	currentTitle={renameOriginalTitle}
+	bind:value={renameDraft}
+	onConfirm={handleRenameConfirm}
+	onCancel={handleRenameCancel}
+/>
+
 <style>
 	aside {
 		@media (max-width: 768px) {

tools/ui/src/lib/components/app/navigation/SidebarNavigation/SidebarNavigationActions.svelte

@@ -119,9 +119,7 @@
 					: onSearchClick}
 			{@const itemTransition = {
 				duration: ICON_STRIP_TRANSITION_DURATION,
-				delay: !initialized
-					? ICON_STRIP_TRANSITION_DELAY_MULTIPLIER + i * ICON_STRIP_TRANSITION_DELAY_MULTIPLIER
-					: 0,
+				delay: !initialized ? i * ICON_STRIP_TRANSITION_DELAY_MULTIPLIER : 0,
 				easing: circIn
 			}}
 
@@ -140,10 +138,8 @@
 							{@render itemIcon(item.icon)}
 
 							{#if showIcons}
-								<span
-									in:fade={{ duration: 150, easing: circIn, delay: 50 }}
-									out:fade={{ duration: 100 }}
-									class="min-w-0 truncate">{item.tooltip}</span
+								<span in:fade={itemTransition} out:fade={itemTransition} class="min-w-0 truncate"
+									>{item.tooltip}</span
 								>
 							{/if}
 						</span>
@@ -171,9 +167,7 @@
 					: onSearchClick}
 			{@const itemTransition = {
 				duration: ICON_STRIP_TRANSITION_DURATION,
-				delay: !initialized
-					? ICON_STRIP_TRANSITION_DELAY_MULTIPLIER + i * ICON_STRIP_TRANSITION_DELAY_MULTIPLIER
-					: 0,
+				delay: !initialized ? i * ICON_STRIP_TRANSITION_DELAY_MULTIPLIER : 0,
 				easing: circIn
 			}}
 

tools/ui/src/lib/components/app/navigation/SidebarNavigation/SidebarNavigationConversationItem.svelte

@@ -9,10 +9,12 @@
 		Square,
 		GitBranch,
 		Pin,
-		PinOff
+		PinOff,
+		ListChecks
 	} from '@lucide/svelte';
 	import { DropdownMenuActions } from '$lib/components/app';
 	import * as Tooltip from '$lib/components/ui/tooltip';
+	import { Checkbox } from '$lib/components/ui/checkbox';
 	import { FORK_TREE_DEPTH_PADDING } from '$lib/constants';
 	import { RouterService } from '$lib/services/router.service';
 	import { getAllLoadingChats } from '$lib/stores/chat.svelte';
@@ -24,10 +26,16 @@
 		isActive?: boolean;
 		depth?: number;
 		conversation: DatabaseConversation;
+		isSelectionMode?: boolean;
+		isSelected?: boolean;
 		onDelete?: (id: string) => void;
 		onEdit?: (id: string) => void;
 		onSelect?: (id: string) => void;
 		onStop?: (id: string) => void;
+		onToggleSelect?: (id: string) => void;
+		onEnterSelectionMode?: (id: string) => void;
+		onSelectionClick?: (id: string, options: { shiftKey: boolean }) => void;
+		onRowMouseDown?: (id: string, event: MouseEvent) => void;
 	}
 
 	let {
@@ -36,7 +44,13 @@
 		onEdit,
 		onSelect,
 		onStop,
+		onToggleSelect,
+		onEnterSelectionMode,
+		onSelectionClick,
+		onRowMouseDown,
 		isActive = false,
+		isSelectionMode = false,
+		isSelected = false,
 		depth = 0
 	}: Props = $props();
 
@@ -64,6 +78,11 @@
 		conversationsStore.toggleConversationPin(conversation.id);
 	}
 
+	function handleEnterSelectionMode(event: Event) {
+		event.stopPropagation();
+		onEnterSelectionMode?.(conversation.id);
+	}
+
 	function handleGlobalEditEvent(event: Event) {
 		const customEvent = event as CustomEvent<{ conversationId: string }>;
 
@@ -79,11 +98,40 @@
 	}
 
 	function handleMouseOver() {
+		if (isSelectionMode) return;
 		renderActionsDropdown = true;
 	}
 
-	function handleSelect() {
-		onSelect?.(conversation.id);
+	function handleSelect(event: MouseEvent) {
+		if (isSelectionMode) {
+			onSelectionClick?.(conversation.id, { shiftKey: event.shiftKey });
+		} else {
+			onSelect?.(conversation.id);
+		}
+	}
+
+	function handleCheckboxClick(event: MouseEvent) {
+		event.stopPropagation();
+		if (isSelectionMode) {
+			onSelectionClick?.(conversation.id, { shiftKey: event.shiftKey });
+		} else {
+			onToggleSelect?.(conversation.id);
+		}
+	}
+
+	function handleRowMouseDown(event: MouseEvent) {
+		onRowMouseDown?.(conversation.id, event);
+	}
+
+	function handleCheckboxKeydown(event: KeyboardEvent) {
+		if (event.key !== ' ' && event.key !== 'Enter') return;
+		event.stopPropagation();
+		event.preventDefault();
+		if (isSelectionMode) {
+			onSelectionClick?.(conversation.id, { shiftKey: event.shiftKey });
+		} else {
+			onToggleSelect?.(conversation.id);
+		}
 	}
 
 	$effect(() => {
@@ -108,10 +156,14 @@
 <button
 	class="group flex min-h-9 w-full cursor-pointer items-center justify-between space-x-3 rounded-lg py-1.5 text-left transition-colors hover:bg-foreground/10 {isActive
 		? 'bg-foreground/5 text-accent-foreground'
-		: ''} px-3"
-	onclick={handleSelect}
+		: ''} {isSelected ? 'bg-primary/10 hover:bg-primary/15' : ''} {isSelectionMode
+		? 'is-selection-mode'
+		: ''} px-2"
+	data-conversation-row={conversation.id}
+	onclick={(e) => handleSelect(e)}
 	onmouseover={handleMouseOver}
 	onmouseleave={handleMouseLeave}
+	onmousedown={(e) => handleRowMouseDown(e)}
 	onfocusin={handleMouseOver}
 	onfocusout={(e) => {
 		if (!e.currentTarget.contains(e.relatedTarget as Node | null)) {
@@ -123,6 +175,23 @@
 		class="flex min-w-0 flex-1 items-center gap-2"
 		style:padding-left="{depth * FORK_TREE_DEPTH_PADDING}px"
 	>
+		{#if isSelectionMode}
+			<div
+				class="shrink-0"
+				onclick={(e) => handleCheckboxClick(e)}
+				onkeydown={handleCheckboxKeydown}
+				role="checkbox"
+				aria-checked={isSelected}
+				aria-label={isSelected ? `Deselect ${conversation.name}` : `Select ${conversation.name}`}
+				tabindex="-1"
+			>
+				<Checkbox
+					checked={isSelected}
+					aria-label={isSelected ? `Deselect ${conversation.name}` : `Select ${conversation.name}`}
+				/>
+			</div>
+		{/if}
+
 		{#if depth > 0}
 			<Tooltip.Root>
 				<Tooltip.Trigger>
@@ -170,7 +239,7 @@
 		<TruncatedText text={conversation.name} class="text-sm font-medium" showTooltip={false} />
 	</div>
 
-	{#if renderActionsDropdown}
+	{#if !isSelectionMode && renderActionsDropdown}
 		<div class="actions flex items-center">
 			<DropdownMenuActions
 				triggerIcon={MoreHorizontal}
@@ -200,6 +269,11 @@
 						},
 						shortcut: ['shift', 'cmd', 's']
 					},
+					{
+						icon: ListChecks,
+						label: 'Select',
+						onclick: handleEnterSelectionMode
+					},
 					{
 						icon: Trash2,
 						label: 'Delete',
@@ -230,6 +304,10 @@
 			}
 		}
 
+		&.is-selection-mode :global([data-slot='dropdown-menu-trigger']) {
+			display: none !important;
+		}
+
 		.stop-button {
 			:global(.stop-icon) {
 				display: none;

tools/ui/src/lib/components/app/navigation/SidebarNavigation/SidebarNavigationConversationList.svelte

@@ -3,6 +3,7 @@
 	import { buildConversationTree } from '$lib/stores/conversations.svelte';
 	import SidebarNavigationConversationItem from './SidebarNavigationConversationItem.svelte';
 	import SidebarNavigationSearchResults from './SidebarNavigationSearchResults.svelte';
+	import SidebarNavigationSelectionBar from './SidebarNavigationSelectionBar.svelte';
 
 	interface Props {
 		class: string;
@@ -10,10 +11,26 @@
 		currentChatId: string | undefined;
 		isSearchModeActive: boolean;
 		searchQuery: string;
+		isSelectionMode?: boolean;
+		selectedIds?: Set<string>;
 		onSelect: (id: string) => void;
 		onEdit: (id: string) => void;
 		onDelete: (id: string) => void;
 		onStop: (id: string) => void;
+		onToggleSelect?: (id: string) => void;
+		onEnterSelectionMode?: (id: string) => void;
+		onSelectionClick?: (id: string, options: { shiftKey: boolean }) => void;
+		onRowMouseDown?: (id: string, event: MouseEvent) => void;
+		visibleCount: number;
+		allVisibleSelected: boolean;
+		someVisibleSelected: boolean;
+		allSelectedArePinned: boolean;
+		pinStateIsMixed: boolean;
+		onSelectAllToggle: () => void;
+		onBulkPinToggle: () => void;
+		onBulkExport: () => void;
+		onBulkDelete: () => void;
+		onCloseSelection: () => void;
 	}
 
 	let {
@@ -22,10 +39,26 @@
 		currentChatId,
 		isSearchModeActive,
 		searchQuery,
+		isSelectionMode = false,
+		selectedIds = new Set<string>(),
 		onSelect,
 		onEdit,
 		onDelete,
-		onStop
+		onStop,
+		onToggleSelect,
+		onEnterSelectionMode,
+		onSelectionClick,
+		onRowMouseDown,
+		visibleCount,
+		allVisibleSelected,
+		someVisibleSelected,
+		allSelectedArePinned,
+		pinStateIsMixed,
+		onSelectAllToggle,
+		onBulkPinToggle,
+		onBulkExport,
+		onBulkDelete,
+		onCloseSelection
 	}: Props = $props();
 
 	let conversationTree = $derived(buildConversationTree(filteredConversations));
@@ -43,65 +76,38 @@
 	);
 </script>
 
-{#if isSearchModeActive}
-	<SidebarNavigationSearchResults
-		class={className}
-		{searchQuery}
-		{filteredConversations}
-		{currentChatId}
-		{onSelect}
-		{onEdit}
-		{onDelete}
-		{onStop}
-	/>
-{:else}
-	{#if pinnedConversations.length > 0}
-		<div class="py-2 flex whitespace-nowrap {className}">
-			<div
-				class="text-muted-foreground inline-flex h-8 shrink-0 items-center rounded-md px-2 text-xs font-medium gap-1"
-			>
-				<Pin class="h-3.5 w-3.5" />
+<div class="flex min-h-0 flex-1 flex-col">
+	{#if isSearchModeActive}
+		<SidebarNavigationSearchResults
+			class={className}
+			{searchQuery}
+			{filteredConversations}
+			{currentChatId}
+			{onSelect}
+			{onEdit}
+			{onDelete}
+			{onStop}
+			{isSelectionMode}
+			{selectedIds}
+			{onToggleSelect}
+			{onEnterSelectionMode}
+			{onSelectionClick}
+			{onRowMouseDown}
+		/>
+	{:else}
+		{#if pinnedConversations.length > 0}
+			<div class="py-2 flex whitespace-nowrap {className}">
+				<div
+					class="text-muted-foreground inline-flex h-8 shrink-0 items-center rounded-md px-2 text-xs font-medium gap-1"
+				>
+					<Pin class="h-3.5 w-3.5" />
 
-				<span>Pinned</span>
+					<span>Pinned</span>
+				</div>
 			</div>
-		</div>
 
-		<ul class="flex w-full min-w-0 flex-col gap-4 md:gap-1 {className}">
-			{#each pinnedConversations as { conversation, depth } (conversation.id)}
-				<li class="group/item relative mb-1 p-0">
-					<SidebarNavigationConversationItem
-						conversation={{
-							id: conversation.id,
-							name: conversation.name,
-							lastModified: conversation.lastModified,
-							currNode: conversation.currNode,
-							forkedFromConversationId: conversation.forkedFromConversationId,
-							pinned: conversation.pinned
-						}}
-						{depth}
-						isActive={currentChatId === conversation.id}
-						{onSelect}
-						{onEdit}
-						{onDelete}
-						{onStop}
-					/>
-				</li>
-			{/each}
-		</ul>
-	{/if}
-
-	<div class="mt-2 flex min-h-0 flex-1 flex-col gap-4 md:gap-2 whitespace-nowrap {className}">
-		{#if filteredConversations.length > 0}
-			<div
-				class="text-muted-foreground flex h-8 shrink-0 items-center rounded-md px-2 text-xs font-medium"
-			>
-				Recent conversations
-			</div>
-		{/if}
-
-		<div class="min-h-0 flex-1 md:overflow-y-auto">
-			<ul class="flex w-full min-w-0 flex-col gap-4 md:gap-1">
-				{#each unpinnedConversations as { conversation, depth } (conversation.id)}
+			<ul class="flex w-full min-w-0 flex-col gap-4 md:gap-1 {className}">
+				{#each pinnedConversations as { conversation, depth } (conversation.id)}
 					<li class="group/item relative mb-1 p-0">
 						<SidebarNavigationConversationItem
 							conversation={{
@@ -114,22 +120,86 @@
 							}}
 							{depth}
 							isActive={currentChatId === conversation.id}
+							{isSelectionMode}
+							isSelected={selectedIds.has(conversation.id)}
 							{onSelect}
 							{onEdit}
 							{onDelete}
 							{onStop}
+							{onToggleSelect}
+							{onEnterSelectionMode}
+							{onSelectionClick}
+							{onRowMouseDown}
 						/>
 					</li>
 				{/each}
-
-				{#if unpinnedConversations.length === 0}
-					<li class="px-2 py-4 text-center">
-						<p class="mb-4 p-4 text-sm text-muted-foreground">
-							{recentEmptyMessage}
-						</p>
-					</li>
-				{/if}
 			</ul>
+		{/if}
+
+		<div class="mt-2 flex min-h-0 flex-1 flex-col gap-4 md:gap-0 whitespace-nowrap {className}">
+			{#if filteredConversations.length > 0}
+				<div
+					class="text-muted-foreground flex h-8 shrink-0 items-center rounded-md px-2 text-xs font-medium"
+				>
+					Recent conversations
+				</div>
+			{/if}
+
+			<div class="min-h-0 flex-1 md:overflow-y-auto">
+				<ul class="flex w-full min-w-0 flex-col gap-4 md:gap-0">
+					{#each unpinnedConversations as { conversation, depth } (conversation.id)}
+						<li class="group/item relative mb-1 p-0">
+							<SidebarNavigationConversationItem
+								conversation={{
+									id: conversation.id,
+									name: conversation.name,
+									lastModified: conversation.lastModified,
+									currNode: conversation.currNode,
+									forkedFromConversationId: conversation.forkedFromConversationId,
+									pinned: conversation.pinned
+								}}
+								{depth}
+								isActive={currentChatId === conversation.id}
+								{isSelectionMode}
+								isSelected={selectedIds.has(conversation.id)}
+								{onSelect}
+								{onEdit}
+								{onDelete}
+								{onStop}
+								{onToggleSelect}
+								{onEnterSelectionMode}
+								{onSelectionClick}
+								{onRowMouseDown}
+							/>
+						</li>
+					{/each}
+
+					{#if unpinnedConversations.length === 0}
+						<li class="px-2 py-4 text-center">
+							<p class="mb-4 p-4 text-sm text-muted-foreground">
+								{recentEmptyMessage}
+							</p>
+						</li>
+					{/if}
+				</ul>
+			</div>
 		</div>
-	</div>
-{/if}
+
+		{#if isSelectionMode}
+			<SidebarNavigationSelectionBar
+				class="sticky top-0 z-10 m-2 mt-0"
+				selectedCount={selectedIds.size}
+				{visibleCount}
+				{allVisibleSelected}
+				{someVisibleSelected}
+				someSelectedPinned={allSelectedArePinned}
+				{pinStateIsMixed}
+				{onSelectAllToggle}
+				{onBulkPinToggle}
+				{onBulkExport}
+				{onBulkDelete}
+				onClose={onCloseSelection}
+			/>
+		{/if}
+	{/if}
+</div>

tools/ui/src/lib/components/app/navigation/SidebarNavigation/SidebarNavigationSearchResults.svelte

@@ -7,10 +7,16 @@
 		searchQuery: string;
 		filteredConversations: DatabaseConversation[];
 		currentChatId: string | undefined;
+		isSelectionMode?: boolean;
+		selectedIds?: Set<string>;
 		onSelect: (id: string) => void;
 		onEdit: (id: string) => void;
 		onDelete: (id: string) => void;
 		onStop: (id: string) => void;
+		onToggleSelect?: (id: string) => void;
+		onEnterSelectionMode?: (id: string) => void;
+		onSelectionClick?: (id: string, options: { shiftKey: boolean }) => void;
+		onRowMouseDown?: (id: string, event: MouseEvent) => void;
 	}
 
 	let {
@@ -18,10 +24,16 @@
 		searchQuery,
 		filteredConversations,
 		currentChatId,
+		isSelectionMode = false,
+		selectedIds = new Set<string>(),
 		onSelect,
 		onEdit,
 		onDelete,
-		onStop
+		onStop,
+		onToggleSelect,
+		onEnterSelectionMode,
+		onSelectionClick,
+		onRowMouseDown
 	}: Props = $props();
 
 	let tree = $derived(buildConversationTree(filteredConversations));
@@ -56,10 +68,16 @@
 						}}
 						{depth}
 						isActive={currentChatId === conversation.id}
+						{isSelectionMode}
+						isSelected={selectedIds.has(conversation.id)}
 						{onSelect}
 						{onEdit}
 						{onDelete}
 						{onStop}
+						{onToggleSelect}
+						{onEnterSelectionMode}
+						{onSelectionClick}
+						{onRowMouseDown}
 					/>
 				</li>
 			{/each}

tools/ui/src/lib/components/app/navigation/SidebarNavigation/SidebarNavigationSelectionBar.svelte

@@ -0,0 +1,163 @@
+<script lang="ts">
+	import { Download, Pin, PinOff, Trash2, X } from '@lucide/svelte';
+	import { ActionIcon, DialogConfirmation } from '$lib/components/app';
+	import { Checkbox } from '$lib/components/ui/checkbox';
+	import { TooltipSide } from '$lib/enums';
+
+	interface Props {
+		class?: string;
+		selectedCount: number;
+		visibleCount: number;
+		allVisibleSelected: boolean;
+		someVisibleSelected: boolean;
+		someSelectedPinned: boolean;
+		pinStateIsMixed: boolean;
+		onSelectAllToggle: () => void;
+		onBulkPinToggle: () => void;
+		onBulkExport: () => void;
+		onBulkDelete: () => void;
+		onClose: () => void;
+	}
+
+	let {
+		class: className = '',
+		selectedCount,
+		visibleCount,
+		allVisibleSelected,
+		someVisibleSelected,
+		someSelectedPinned,
+		pinStateIsMixed,
+		onSelectAllToggle,
+		onBulkPinToggle,
+		onBulkExport,
+		onBulkDelete,
+		onClose
+	}: Props = $props();
+
+	let showDeleteDialog = $state(false);
+
+	function handleDeleteClick() {
+		showDeleteDialog = true;
+	}
+
+	function handleDeleteConfirm() {
+		showDeleteDialog = false;
+		onBulkDelete();
+	}
+
+	function handleDeleteCancel() {
+		showDeleteDialog = false;
+	}
+
+	const hasSelection = $derived(selectedCount > 0);
+	const isMasterChecked = $derived(allVisibleSelected);
+	const isMasterIndeterminate = $derived(!allVisibleSelected && someVisibleSelected);
+
+	const pinTooltip = $derived(
+		hasSelection
+			? pinStateIsMixed
+				? 'Unavailable for mixed state selection'
+				: someSelectedPinned
+					? selectedCount === 1
+						? 'Unpin'
+						: 'Unpin all'
+					: selectedCount === 1
+						? 'Pin'
+						: 'Pin all'
+			: 'Pin'
+	);
+
+	const pinDisabled = $derived(!hasSelection || pinStateIsMixed);
+</script>
+
+<div
+	role="toolbar"
+	aria-label="Bulk actions for selected conversations"
+	class="flex items-center gap-1.5 rounded-xl border border-border/50 bg-background/50 px-2 py-1.5 shadow-sm backdrop-blur-xl {className}"
+>
+	<label class="flex min-w-0 cursor-pointer items-center gap-2">
+		<Checkbox
+			checked={isMasterChecked}
+			indeterminate={isMasterIndeterminate}
+			onCheckedChange={onSelectAllToggle}
+			aria-label={isMasterChecked ? 'Deselect all' : 'Select all'}
+		/>
+
+		<span class="truncate text-xs font-medium text-muted-foreground">
+			{selectedCount} / {visibleCount} selected
+		</span>
+	</label>
+
+	<div class="ml-auto flex items-center gap-0.75">
+		<ActionIcon
+			icon={someSelectedPinned ? PinOff : Pin}
+			tooltip={pinTooltip}
+			tooltipSide={TooltipSide.TOP}
+			disabled={pinDisabled}
+			ariaLabel={pinTooltip}
+			size="sm"
+			iconSize="h-3.5 w-3.5"
+			class="h-7 w-7 rounded-md bg-transparent backdrop-blur-none hover:bg-accent! {pinDisabled
+				? 'cursor-not-allowed'
+				: ''} {!pinDisabled ? 'opacity-100' : 'opacity-40'}"
+			onclick={onBulkPinToggle}
+		/>
+
+		<ActionIcon
+			icon={Download}
+			tooltip={hasSelection ? 'Export' : 'Export'}
+			tooltipSide={TooltipSide.TOP}
+			disabled={!hasSelection}
+			ariaLabel="Export selected"
+			size="sm"
+			iconSize="h-3.5 w-3.5"
+			class="h-7 w-7 rounded-md bg-transparent backdrop-blur-none hover:bg-accent! {hasSelection
+				? 'opacity-100'
+				: 'opacity-40'}"
+			onclick={onBulkExport}
+		/>
+
+		<ActionIcon
+			icon={Trash2}
+			tooltip="Delete selected"
+			tooltipSide={TooltipSide.TOP}
+			disabled={!hasSelection}
+			ariaLabel="Delete selected"
+			size="sm"
+			iconSize="h-3.5 w-3.5 text-destructive"
+			class="h-7 w-7 rounded-md bg-transparent backdrop-blur-none hover:bg-destructive/10! dark:hover:bg-destructive/20! disabled:hover:bg-transparent {hasSelection
+				? 'opacity-100'
+				: 'opacity-40'}"
+			onclick={handleDeleteClick}
+		/>
+
+		<div class="mx-1 h-4 w-px bg-border" aria-hidden="true"></div>
+
+		<ActionIcon
+			icon={X}
+			tooltip="Exit bulk selection mode"
+			tooltipSide={TooltipSide.TOP}
+			ariaLabel="Exit bulk selection mode"
+			size="sm"
+			iconSize="h-3.5 w-3.5"
+			class="h-7 w-7 rounded-md bg-transparent backdrop-blur-none hover:bg-accent!"
+			onclick={onClose}
+		/>
+	</div>
+</div>
+
+<DialogConfirmation
+	bind:open={showDeleteDialog}
+	title="Delete {selectedCount} conversation{selectedCount === 1 ? '' : 's'}"
+	description="This action cannot be undone. The selected conversation{selectedCount === 1
+		? ''
+		: 's'} and {selectedCount === 1
+		? 'its'
+		: 'their'} messages will be permanently removed, including any forks."
+	confirmText={selectedCount === 1 ? 'Delete' : `Delete ${selectedCount}`}
+	cancelText="Cancel"
+	variant="destructive"
+	icon={Trash2}
+	onConfirm={handleDeleteConfirm}
+	onCancel={handleDeleteCancel}
+/>

tools/ui/src/lib/components/app/navigation/index.ts

@@ -114,6 +114,36 @@ export { default as SidebarNavigation } from './SidebarNavigation/SidebarNavigat
  */
 export { default as SidebarNavigationConversationItem } from './SidebarNavigation/SidebarNavigationConversationItem.svelte';
 
+/**
+ * **SidebarNavigationSelectionBar** - Bulk action toolbar for selection mode
+ *
+ * Rendered above the conversation list when the sidebar enters selection mode.
+ * Hosts a master checkbox (with select-all / clear-all semantics over the
+ * currently-visible items), a selected-count caption, and bulk actions for
+ * pin/unpin, export, and delete. Delete uses
+ * {@link DialogConfirmation} before invoking the bulk store method.
+ *
+ * Pure-presentational; all operations are delegated via callbacks so the
+ * sidebar owns selection state and persistence.
+ *
+ * @example
+ * ```svelte
+ * <SidebarNavigationSelectionBar
+ *   selectedCount={selectedIds.size}
+ *   visibleCount={visibleConversations.length}
+ *   allVisibleSelected={...}
+ *   someVisibleSelected={...}
+ *   someSelectedPinned={...}
+ *   onSelectAllToggle={toggleSelectAll}
+ *   onBulkPinToggle={handleBulkPin}
+ *   onBulkExport={handleBulkExport}
+ *   onBulkDelete={handleBulkDelete}
+ *   onClose={exitSelectionMode}
+ * />
+ * ```
+ */
+export { default as SidebarNavigationSelectionBar } from './SidebarNavigation/SidebarNavigationSelectionBar.svelte';
+
 /**
  * **SidebarNavigationConversationList** - Grouped conversation list
  *

tools/ui/src/lib/components/app/settings/SettingsChat/SettingsChatFields.svelte

@@ -4,6 +4,7 @@
 	import { Checkbox } from '$lib/components/ui/checkbox';
 	import { Input } from '$lib/components/ui/input';
 	import Label from '$lib/components/ui/label/label.svelte';
+	import * as RadioGroup from '$lib/components/ui/radio-group';
 	import * as Select from '$lib/components/ui/select';
 	import { Textarea } from '$lib/components/ui/textarea';
 	import { SETTING_CONFIG_INFO, SETTINGS_KEYS } from '$lib/constants';
@@ -84,10 +85,7 @@
 						type={field.isPositiveInteger ? 'number' : 'text'}
 						{...field.isPositiveInteger ? { min: '1', step: '1' } : {}}
 						value={currentValue}
-						oninput={(e) => {
-							// Update local config immediately for real-time badge feedback
-							onConfigChange(field.key, e.currentTarget.value);
-						}}
+						oninput={(e) => onConfigChange(field.key, e.currentTarget.value)}
 						placeholder={currentModelParams[field.key] != null
 							? `Default: ${normalizeFloatingPoint(currentModelParams[field.key])}`
 							: ''}
@@ -236,6 +234,52 @@
 						{field.help || SETTING_CONFIG_INFO[field.key]}
 					</p>
 				{/if}
+			{:else if field.type === SettingsFieldType.RADIO && field.radioOptions}
+				{@const radioOptions = field.radioOptions}
+				{@const currentMode =
+					radioOptions.find((o: { key: string }) => Boolean(localConfig[o.key]))?.value ??
+					radioOptions[0].value}
+
+				<Label class="flex items-center gap-1.5 text-sm font-medium mb-4">
+					{field.label}
+
+					{#if field.isExperimental}
+						<FlaskConical class="h-3.5 w-3.5 text-muted-foreground" />
+					{/if}
+				</Label>
+
+				<RadioGroup.Root
+					class="gap-4"
+					value={currentMode}
+					onValueChange={(value) => {
+						for (const opt of radioOptions) {
+							onConfigChange(opt.key, opt.value === value);
+						}
+					}}
+				>
+					{#each radioOptions as opt (opt.value)}
+						{@const itemId = `${field.key}-${opt.value}`}
+						<div class="flex items-center gap-2">
+							<RadioGroup.Item value={opt.value} id={itemId} />
+							<Label
+								for={itemId}
+								class="flex cursor-pointer items-center gap-1.5 text-sm font-normal"
+							>
+								{opt.label}
+
+								{#if opt.isExperimental}
+									<FlaskConical class="h-3.5 w-3.5 text-muted-foreground" />
+								{/if}
+							</Label>
+						</div>
+					{/each}
+				</RadioGroup.Root>
+
+				{#if field.help || SETTING_CONFIG_INFO[field.key]}
+					<p class="text-xs text-muted-foreground">
+						{field.help || SETTING_CONFIG_INFO[field.key]}
+					</p>
+				{/if}
 			{:else if field.type === SettingsFieldType.CHECKBOX}
 				<div class="flex items-start space-x-3">
 					<Checkbox

tools/ui/src/lib/components/app/settings/SettingsChat/SettingsChatToolsTab.svelte

@@ -7,6 +7,8 @@
 	import { toolsStore } from '$lib/stores/tools.svelte';
 	import { permissionsStore } from '$lib/stores/permissions.svelte';
 	import { mcpStore } from '$lib/stores/mcp.svelte';
+	import { getBuiltinToolUi } from '$lib/constants/built-in-tools';
+	import { ToolSource } from '$lib/enums/tools.enums';
 	import { SvelteSet } from 'svelte/reactivity';
 
 	let expandedGroups = new SvelteSet<string>();
@@ -69,12 +71,23 @@
 
 						{#each group.tools as entry (entry.key)}
 							{@const toolName = entry.definition.function.name}
+							{@const builtinUi =
+								entry.source === ToolSource.BUILTIN || entry.source === ToolSource.FRONTEND
+									? getBuiltinToolUi(toolName)
+									: null}
+							{@const displayLabel = builtinUi?.label ?? toolName}
+							{@const IconComponent = builtinUi?.icon ?? null}
 							{@const isEnabled = toolsStore.isToolEnabled(entry.key)}
 							{@const permissionKey = entry.key}
 							{@const isAlwaysAllowed = permissionsStore.hasTool(permissionKey)}
 
 							<div class="flex items-center gap-2 rounded px-2 py-1.5 text-sm hover:bg-muted/50">
-								<TruncatedText text={toolName} class="flex-1" showTooltip={true} />
+								<span class="flex min-w-0 flex-1 items-center gap-1.5">
+									{#if IconComponent}
+										<IconComponent class={ICON_CLASS_DEFAULT} />
+									{/if}
+									<TruncatedText text={displayLabel} class="min-w-0" showTooltip={true} />
+								</span>
 
 								<div class="flex w-16 shrink-0 justify-center">
 									<Checkbox

tools/ui/src/lib/components/ui/checkbox/checkbox.svelte

@@ -17,7 +17,7 @@
 	bind:ref
 	data-slot="checkbox"
 	class={cn(
-		'peer flex size-4 shrink-0 items-center justify-center rounded-[4px] border border-input shadow-xs transition-shadow outline-none focus-visible:border-ring focus-visible:ring-[3px] focus-visible:ring-ring/50 disabled:cursor-not-allowed disabled:opacity-50 aria-invalid:border-destructive aria-invalid:ring-destructive/20 data-[state=checked]:border-primary data-[state=checked]:bg-primary data-[state=checked]:text-primary-foreground dark:bg-input/30 dark:aria-invalid:ring-destructive/40 dark:data-[state=checked]:bg-primary',
+		'peer flex size-4 shrink-0 items-center justify-center rounded-[4px] border border-input bg-background shadow-xs transition-shadow outline-none focus-visible:border-ring focus-visible:ring-[3px] focus-visible:ring-ring/50 disabled:cursor-not-allowed disabled:opacity-50 aria-invalid:border-destructive aria-invalid:ring-destructive/20 data-[state=checked]:border-primary data-[state=checked]:bg-primary data-[state=checked]:text-primary-foreground dark:bg-input/30 dark:aria-invalid:ring-destructive/40 dark:data-[state=checked]:bg-primary',
 		className
 	)}
 	bind:checked

tools/ui/src/lib/components/ui/radio-group/index.ts

@@ -0,0 +1,10 @@
+import Root from './radio-group.svelte';
+import Item from './radio-group-item.svelte';
+
+export {
+	Root,
+	Item,
+	//
+	Root as RadioGroup,
+	Item as RadioGroupItem
+};

tools/ui/src/lib/components/ui/radio-group/radio-group-item.svelte

@@ -0,0 +1,31 @@
+<script lang="ts">
+	import { RadioGroup as RadioGroupPrimitive } from 'bits-ui';
+	import CircleIcon from '@lucide/svelte/icons/circle';
+	import { cn, type WithoutChildrenOrChild } from '$lib/components/ui/utils.js';
+
+	let {
+		ref = $bindable(null),
+		class: className,
+		...restProps
+	}: WithoutChildrenOrChild<RadioGroupPrimitive.ItemProps> = $props();
+</script>
+
+<RadioGroupPrimitive.Item
+	bind:ref
+	data-slot="radio-group-item"
+	class={cn(
+		'border-input dark:bg-input/30 data-checked:bg-primary data-checked:text-primary-foreground dark:data-checked:bg-primary data-checked:border-primary aria-invalid:aria-checked:border-primary aria-invalid:border-destructive focus-visible:border-ring focus-visible:ring-ring/50 aria-invalid:ring-destructive/20 dark:aria-invalid:ring-destructive/40 dark:aria-invalid:border-destructive/50 flex size-4 rounded-full focus-visible:ring-3 aria-invalid:ring-3 group/radio-group-item peer relative aspect-square shrink-0 border outline-none after:absolute after:-inset-x-3 after:-inset-y-2 disabled:cursor-not-allowed disabled:opacity-50',
+		className
+	)}
+	{...restProps}
+>
+	{#snippet children({ checked })}
+		<div data-slot="radio-group-indicator" class="flex size-4 items-center justify-center">
+			{#if checked}
+				<CircleIcon
+					class="bg-primary-foreground absolute top-1/2 left-1/2 size-2 -translate-x-1/2 -translate-y-1/2 rounded-full"
+				/>
+			{/if}
+		</div>
+	{/snippet}
+</RadioGroupPrimitive.Item>

tools/ui/src/lib/components/ui/radio-group/radio-group.svelte

@@ -0,0 +1,19 @@
+<script lang="ts">
+	import { RadioGroup as RadioGroupPrimitive } from 'bits-ui';
+	import { cn } from '$lib/components/ui/utils.js';
+
+	let {
+		ref = $bindable(null),
+		class: className,
+		value = $bindable(''),
+		...restProps
+	}: RadioGroupPrimitive.RootProps = $props();
+</script>
+
+<RadioGroupPrimitive.Root
+	bind:ref
+	bind:value
+	data-slot="radio-group"
+	class={cn('grid gap-2 w-full', className)}
+	{...restProps}
+/>

tools/ui/src/lib/constants/settings-keys.ts

@@ -14,7 +14,6 @@ export const SETTINGS_KEYS = {
 	SEND_ON_ENTER: 'sendOnEnter',
 	ENABLE_CONTINUE_GENERATION: 'enableContinueGeneration',
 	PDF_AS_IMAGE: 'pdfAsImage',
-	ASK_FOR_TITLE_CONFIRMATION: 'askForTitleConfirmation',
 	TITLE_GENERATION_USE_FIRST_LINE: 'titleGenerationUseFirstLine',
 	TITLE_GENERATION_USE_LLM: 'titleGenerationUseLLM',
 	TITLE_GENERATION_PROMPT: 'titleGenerationPrompt',
@@ -60,15 +59,15 @@ export const SETTINGS_KEYS = {
 	MCP_SERVERS: 'mcpServers',
 	MCP_REQUEST_TIMEOUT_SECONDS: 'mcpRequestTimeoutSeconds',
 	AGENTIC_MAX_TURNS: 'agenticMaxTurns',
-	SHOW_TOOL_CALL_IN_PROGRESS: 'showToolCallInProgress',
+	ALWAYS_SHOW_TOOL_CALL_CONTENT: 'alwaysShowToolCallContent',
 	// Performance
 	PRE_ENCODE_CONVERSATION: 'preEncodeConversation',
 	// Developer
 	DISABLE_REASONING_PARSING: 'disableReasoningParsing',
 	EXCLUDE_REASONING_FROM_CONTEXT: 'excludeReasoningFromContext',
 	SHOW_RAW_OUTPUT_SWITCH: 'showRawOutputSwitch',
-	// PY_INTERPRETER_ENABLED: 'pyInterpreterEnabled',
 	JS_SANDBOX_ENABLED: 'jsSandboxEnabled',
+	// PY_INTERPRETER_ENABLED: 'pyInterpreterEnabled',
 	CUSTOM_JSON: 'customJson',
 	CUSTOM_CSS: 'customCss'
 } as const;

tools/ui/src/lib/constants/settings-registry.ts

@@ -54,6 +54,34 @@ const COLOR_MODE_OPTIONS: Array<{ value: string; label: string; icon: Component
 	{ value: ColorMode.DARK, label: 'Dark', icon: Moon }
 ];
 
+// Shared options for the title-generation radio group. Both paired registry entries
+// (USE_FIRST_LINE, USE_LLM) reference this list so labels stay in lockstep.
+const TITLE_GENERATION_RADIO_OPTIONS: Array<{
+	value: string;
+	label: string;
+	key: string;
+	isExperimental?: boolean;
+}> = [
+	{
+		value: 'firstLine',
+		label: 'Use first non-empty line for the conversation title',
+		key: SETTINGS_KEYS.TITLE_GENERATION_USE_FIRST_LINE
+	},
+	{
+		value: 'llm',
+		label: 'Generate title with LLM',
+		key: SETTINGS_KEYS.TITLE_GENERATION_USE_LLM,
+		isExperimental: true
+	}
+];
+
+// Common shape for the conversation title radio entry.
+const TITLE_GENERATION_BASE = {
+	type: SettingsFieldType.RADIO,
+	section: SETTINGS_SECTION_SLUGS.GENERAL,
+	radioOptions: TITLE_GENERATION_RADIO_OPTIONS
+} as const;
+
 const SETTINGS_REGISTRY: Record<string, SettingsSectionEntry> = {
 	[SETTINGS_SECTION_SLUGS.GENERAL]: {
 		title: SETTINGS_SECTION_TITLES.GENERAL,
@@ -115,14 +143,15 @@ const SETTINGS_REGISTRY: Record<string, SettingsSectionEntry> = {
 				}
 			},
 			{
-				key: SETTINGS_KEYS.COPY_TEXT_ATTACHMENTS_AS_PLAIN_TEXT,
-				label: 'Copy text attachments as plain text',
-				help: 'When copying a message with text attachments, combine them into a single plain text string instead of a special format that can be pasted back as attachments.',
+				key: SETTINGS_KEYS.AUTO_MIC_ON_EMPTY,
+				label: 'Show microphone on empty input',
+				help: 'Automatically show microphone button instead of send button when textarea is empty for models with audio modality support.',
 				defaultValue: false,
 				type: SettingsFieldType.CHECKBOX,
 				section: SETTINGS_SECTION_SLUGS.GENERAL,
+				isExperimental: true,
 				sync: {
-					serverKey: SETTINGS_KEYS.COPY_TEXT_ATTACHMENTS_AS_PLAIN_TEXT,
+					serverKey: SETTINGS_KEYS.AUTO_MIC_ON_EMPTY,
 					paramType: SyncableParameterType.BOOLEAN
 				}
 			},
@@ -140,54 +169,16 @@ const SETTINGS_REGISTRY: Record<string, SettingsSectionEntry> = {
 				}
 			},
 			{
-				key: SETTINGS_KEYS.PDF_AS_IMAGE,
-				label: 'Parse PDF as image',
-				help: 'Parse PDF as image instead of text. Automatically falls back to text processing for non-vision models.',
-				defaultValue: false,
-				type: SettingsFieldType.CHECKBOX,
-				section: SETTINGS_SECTION_SLUGS.GENERAL,
-				sync: {
-					serverKey: SETTINGS_KEYS.PDF_AS_IMAGE,
-					paramType: SyncableParameterType.BOOLEAN
-				}
-			},
-			{
-				key: SETTINGS_KEYS.ASK_FOR_TITLE_CONFIRMATION,
-				label: 'Ask for confirmation before changing conversation title',
-				help: 'Ask for confirmation before automatically changing conversation title when editing the first message.',
-				defaultValue: false,
-				type: SettingsFieldType.CHECKBOX,
-				section: SETTINGS_SECTION_SLUGS.GENERAL,
-				sync: {
-					serverKey: SETTINGS_KEYS.ASK_FOR_TITLE_CONFIRMATION,
-					paramType: SyncableParameterType.BOOLEAN
-				}
-			},
-			{
+				...TITLE_GENERATION_BASE,
 				key: SETTINGS_KEYS.TITLE_GENERATION_USE_FIRST_LINE,
-				label: 'Use first non-empty line for conversation title',
-				help: 'Use only the first non-empty line of the prompt to generate the conversation title.',
-				defaultValue: false,
-				type: SettingsFieldType.CHECKBOX,
-				section: SETTINGS_SECTION_SLUGS.GENERAL,
+				label: 'Conversation title',
+				help: 'Choose how conversation titles are generated. The first non-empty line uses a fast deterministic rule; the LLM option uses a model-generated title from the first message exchange.',
+				defaultValue: true,
 				sync: {
 					serverKey: SETTINGS_KEYS.TITLE_GENERATION_USE_FIRST_LINE,
 					paramType: SyncableParameterType.BOOLEAN
 				}
 			},
-			{
-				key: SETTINGS_KEYS.TITLE_GENERATION_USE_LLM,
-				label: 'Use LLM to generate conversation title',
-				help: 'Use the LLM to automatically generate conversation titles based on the first message exchange.',
-				defaultValue: false,
-				type: SettingsFieldType.CHECKBOX,
-				section: SETTINGS_SECTION_SLUGS.GENERAL,
-				isExperimental: true,
-				sync: {
-					serverKey: SETTINGS_KEYS.TITLE_GENERATION_USE_LLM,
-					paramType: SyncableParameterType.BOOLEAN
-				}
-			},
 			{
 				key: SETTINGS_KEYS.TITLE_GENERATION_PROMPT,
 				label: 'LLM title generation prompt',
@@ -195,11 +186,36 @@ const SETTINGS_REGISTRY: Record<string, SettingsSectionEntry> = {
 				defaultValue: TITLE_GENERATION.DEFAULT_PROMPT,
 				type: SettingsFieldType.TEXTAREA,
 				section: SETTINGS_SECTION_SLUGS.GENERAL,
+				dependsOn: SETTINGS_KEYS.TITLE_GENERATION_USE_LLM,
 				sync: {
 					serverKey: SETTINGS_KEYS.TITLE_GENERATION_PROMPT,
 					paramType: SyncableParameterType.STRING
 				}
 			},
+			{
+				key: SETTINGS_KEYS.COPY_TEXT_ATTACHMENTS_AS_PLAIN_TEXT,
+				label: 'Copy text attachments as plain text',
+				help: 'When copying a message with text attachments, combine them into a single plain text string instead of a special format that can be pasted back as attachments.',
+				defaultValue: false,
+				type: SettingsFieldType.CHECKBOX,
+				section: SETTINGS_SECTION_SLUGS.GENERAL,
+				sync: {
+					serverKey: SETTINGS_KEYS.COPY_TEXT_ATTACHMENTS_AS_PLAIN_TEXT,
+					paramType: SyncableParameterType.BOOLEAN
+				}
+			},
+			{
+				key: SETTINGS_KEYS.PDF_AS_IMAGE,
+				label: 'Parse PDF as image',
+				help: 'Parse PDF as image instead of text. Automatically falls back to text processing for non-vision models.',
+				defaultValue: false,
+				type: SettingsFieldType.CHECKBOX,
+				section: SETTINGS_SECTION_SLUGS.GENERAL,
+				sync: {
+					serverKey: SETTINGS_KEYS.PDF_AS_IMAGE,
+					paramType: SyncableParameterType.BOOLEAN
+				}
+			},
 			{
 				key: SETTINGS_KEYS.MAX_IMAGE_RESOLUTION,
 				label: 'Maximum image resolution (megapixels)',
@@ -253,27 +269,14 @@ const SETTINGS_REGISTRY: Record<string, SettingsSectionEntry> = {
 				}
 			},
 			{
-				key: SETTINGS_KEYS.SHOW_TOOL_CALL_IN_PROGRESS,
-				label: 'Show tool call in progress',
+				key: SETTINGS_KEYS.ALWAYS_SHOW_TOOL_CALL_CONTENT,
+				label: 'Always show tool call content',
 				help: 'Automatically expand tool call details while executing and keep them expanded after completion.',
 				defaultValue: false,
 				type: SettingsFieldType.CHECKBOX,
 				section: SETTINGS_SECTION_SLUGS.DISPLAY,
 				sync: {
-					serverKey: SETTINGS_KEYS.SHOW_TOOL_CALL_IN_PROGRESS,
-					paramType: SyncableParameterType.BOOLEAN
-				}
-			},
-			{
-				key: SETTINGS_KEYS.AUTO_MIC_ON_EMPTY,
-				label: 'Show microphone on empty input',
-				help: 'Automatically show microphone button instead of send button when textarea is empty for models with audio modality support.',
-				defaultValue: false,
-				type: SettingsFieldType.CHECKBOX,
-				section: SETTINGS_SECTION_SLUGS.DISPLAY,
-				isExperimental: true,
-				sync: {
-					serverKey: SETTINGS_KEYS.AUTO_MIC_ON_EMPTY,
+					serverKey: SETTINGS_KEYS.ALWAYS_SHOW_TOOL_CALL_CONTENT,
 					paramType: SyncableParameterType.BOOLEAN
 				}
 			},
@@ -764,6 +767,17 @@ const NON_UI_SETTINGS: SettingsEntry[] = [
 		defaultValue: '[]',
 		type: SettingsFieldType.INPUT,
 		sync: { serverKey: SETTINGS_KEYS.MCP_SERVERS, paramType: SyncableParameterType.STRING }
+	},
+	{
+		key: SETTINGS_KEYS.TITLE_GENERATION_USE_LLM,
+		label: 'Generate title with LLM',
+		help: 'Counterpart of the conversation title radio; stored and synced without a dedicated UI field.',
+		defaultValue: false,
+		type: SettingsFieldType.CHECKBOX,
+		sync: {
+			serverKey: SETTINGS_KEYS.TITLE_GENERATION_USE_LLM,
+			paramType: SyncableParameterType.BOOLEAN
+		}
 	}
 	// {
 	// 	key: SETTINGS_KEYS.PY_INTERPRETER_ENABLED,
@@ -812,7 +826,8 @@ export const SETTINGS_CHAT_SECTIONS: SettingsSection[] = [
 			isPositiveInteger: s.isPositiveInteger,
 			dependsOn: s.dependsOn,
 			help: s.help,
-			options: s.options
+			options: s.options,
+			radioOptions: s.radioOptions
 		}))
 	})),
 	...STANDALONE_SECTIONS

tools/ui/src/lib/enums/settings.enums.ts

@@ -22,5 +22,6 @@ export enum SettingsFieldType {
 	INPUT = 'input',
 	TEXTAREA = 'textarea',
 	CHECKBOX = 'checkbox',
-	SELECT = 'select'
+	SELECT = 'select',
+	RADIO = 'radio'
 }

tools/ui/src/lib/hooks/use-marquee-selection.svelte.ts

@@ -0,0 +1,239 @@
+/**
+ * Reusable selection state-machine: shift+click/shift+drag range select plus
+ * rubber-band marquee drag, anchored on the last clicked row. Both the sidebar
+ * conversation list and the dialog conversations table share this.
+ *
+ * The hook mutates the consumer's SvelteSet<string> directly; the consumer
+ * owns the source of truth and reads it like any other $state. orderedIds
+ * must reflect the current visual order of selectable rows so the range
+ * matches what the user sees on screen.
+ */
+
+import { SvelteSet } from 'svelte/reactivity';
+
+interface UseMarqueeSelectionOptions {
+	/** Latest selected-IDs set. Re-read per selection event so consumer-side reassignment works. */
+	selectedIds: () => SvelteSet<string>;
+	/** IDs in the current rendered order; used to compute shift+click ranges and gate marquee visibility. */
+	orderedIds: () => string[];
+	/** Document listeners attach only while the getter returns true. */
+	enabled: () => boolean;
+	/** DOM attribute key (after the `data-` prefix) that marks selectable rows. */
+	attributeName?: () => string;
+	/** Minimum pixel distance before a press becomes a marquee drag. */
+	dragThresholdPx?: number;
+}
+
+export function useMarqueeSelection(options: UseMarqueeSelectionOptions) {
+	const dragThresholdPx = options.dragThresholdPx ?? 5;
+
+	let dragAnchorId = $state<string | null>(null);
+	let isMarqueeDragging = $state(false);
+	let mouseDownActive = false;
+	let dragStartX = 0;
+	let dragStartY = 0;
+	let mousedownRowId: string | null = null;
+	let dragMode: 'add' | 'remove' | null = null;
+	let suppressNextClick = false;
+
+	function resolveAttributeName(): string {
+		return options.attributeName?.() ?? 'conversation-row';
+	}
+
+	/**
+	 * `dataset` keys are camelCased. `data-conversation-row` -> `conversationRow`.
+	 * We resolve the attribute name once per call and read via the camelCase key.
+	 */
+	function datasetKey(key: string = resolveAttributeName()): string {
+		return key.replace(/-([a-z])/g, (_, c) => c.toUpperCase());
+	}
+
+	function decideDragMode(startingRowId: string | null, currentlySelected: ReadonlySet<string>) {
+		return startingRowId !== null && currentlySelected.has(startingRowId) ? 'remove' : 'add';
+	}
+
+	/**
+	 * Range-select uses Finder-style toggle-by-target semantics: if the target
+	 * row is currently selected the range becomes deselected, otherwise it
+	 * becomes selected. Anchor moves to `toId` so chained shift+clicks keep
+	 * extending from the previous endpoint.
+	 */
+	function rangeSelect(fromId: string, toId: string) {
+		const selected = options.selectedIds();
+		const order = options.orderedIds();
+		const fromIdx = order.indexOf(fromId);
+		const toIdx = order.indexOf(toId);
+		if (fromIdx === -1 || toIdx === -1) return;
+		const [lo, hi] = fromIdx < toIdx ? [fromIdx, toIdx] : [toIdx, fromIdx];
+		const shouldSelect = !selected.has(toId);
+		for (let i = lo; i <= hi; i++) {
+			const id = order[i];
+			if (shouldSelect) selected.add(id);
+			else selected.delete(id);
+		}
+	}
+
+	function findRowAtPoint(x: number, y: number): string | null {
+		const attr = resolveAttributeName();
+		const selector = `[data-${attr}]`;
+		const key = datasetKey(attr);
+		let bestMatch: HTMLElement | null = null;
+		let bestCenterDistance = Infinity;
+
+		for (const row of document.querySelectorAll<HTMLElement>(selector)) {
+			const rect = row.getBoundingClientRect();
+			if (y >= rect.top && y <= rect.bottom && x >= rect.left && x <= rect.right) {
+				return row.dataset[key] ?? null;
+			}
+			if (x >= rect.left && x <= rect.right) {
+				const centerDistance = Math.abs(y - (rect.top + rect.height / 2));
+				if (centerDistance < bestCenterDistance) {
+					bestCenterDistance = centerDistance;
+					bestMatch = row;
+				}
+			}
+		}
+		return bestMatch ? (bestMatch.dataset[key] ?? null) : null;
+	}
+
+	function updateMarqueeRect(currentX: number, currentY: number) {
+		const attr = resolveAttributeName();
+		const selector = `[data-${attr}]`;
+		const key = datasetKey(attr);
+		const selected = options.selectedIds();
+		const left = Math.min(dragStartX, currentX);
+		const top = Math.min(dragStartY, currentY);
+		const right = Math.max(dragStartX, currentX);
+		const bottom = Math.max(dragStartY, currentY);
+		const visibleIds = new SvelteSet(options.orderedIds());
+
+		for (const row of document.querySelectorAll<HTMLElement>(selector)) {
+			const id = row.dataset[key];
+			if (!id || !visibleIds.has(id)) continue;
+
+			const rect = row.getBoundingClientRect();
+			const intersects = !(
+				rect.right < left ||
+				rect.left > right ||
+				rect.bottom < top ||
+				rect.top > bottom
+			);
+
+			if (dragMode === 'add') {
+				if (intersects) selected.add(id);
+			} else if (dragMode === 'remove') {
+				if (intersects && selected.has(id)) selected.delete(id);
+			}
+		}
+	}
+
+	function handleDocumentMouseMove(event: MouseEvent) {
+		if (!mouseDownActive) return;
+
+		if (event.shiftKey && dragAnchorId !== null) {
+			const target = findRowAtPoint(event.clientX, event.clientY);
+			if (target && target !== mousedownRowId) rangeSelect(dragAnchorId, target);
+			return;
+		}
+
+		if (!isMarqueeDragging) {
+			const dx = event.clientX - dragStartX;
+			const dy = event.clientY - dragStartY;
+			if (Math.hypot(dx, dy) < dragThresholdPx) return;
+			isMarqueeDragging = true;
+			dragMode = decideDragMode(mousedownRowId, options.selectedIds());
+		}
+		updateMarqueeRect(event.clientX, event.clientY);
+	}
+
+	function handleDocumentMouseUp(event: MouseEvent) {
+		if (isMarqueeDragging) {
+			suppressNextClick = true;
+			const target = findRowAtPoint(event.clientX, event.clientY);
+			if (target) dragAnchorId = target;
+		}
+		isMarqueeDragging = false;
+		mouseDownActive = false;
+		mousedownRowId = null;
+		dragMode = null;
+		dragStartX = 0;
+		dragStartY = 0;
+	}
+
+	function handleClickCapture(event: MouseEvent) {
+		if (suppressNextClick) {
+			event.stopPropagation();
+			event.preventDefault();
+			suppressNextClick = false;
+		}
+	}
+
+	$effect(() => {
+		if (!options.enabled()) {
+			reset();
+			return;
+		}
+		document.addEventListener('mousemove', handleDocumentMouseMove);
+		document.addEventListener('mouseup', handleDocumentMouseUp);
+		document.addEventListener('click', handleClickCapture, { capture: true });
+		return () => {
+			document.removeEventListener('mousemove', handleDocumentMouseMove);
+			document.removeEventListener('mouseup', handleDocumentMouseUp);
+			document.removeEventListener('click', handleClickCapture, { capture: true });
+		};
+	});
+
+	function rowMouseDown(id: string, event: MouseEvent) {
+		if (!options.enabled()) return;
+		if (event.button !== 0) return;
+		event.preventDefault();
+		mouseDownActive = true;
+		mousedownRowId = id;
+		dragStartX = event.clientX;
+		dragStartY = event.clientY;
+		isMarqueeDragging = false;
+		dragMode = null;
+	}
+
+	function rowClick(id: string, shiftKey: boolean) {
+		if (!options.enabled()) return;
+		const selected = options.selectedIds();
+
+		if (shiftKey) {
+			const anchor = dragAnchorId;
+			if (anchor !== null && anchor !== id) {
+				rangeSelect(anchor, id);
+			} else if (selected.has(id)) {
+				selected.delete(id);
+			} else {
+				selected.add(id);
+			}
+			dragAnchorId = id;
+			return;
+		}
+
+		if (selected.has(id)) selected.delete(id);
+		else selected.add(id);
+		dragAnchorId = id;
+	}
+
+	function reset() {
+		dragAnchorId = null;
+		isMarqueeDragging = false;
+		mouseDownActive = false;
+		suppressNextClick = false;
+		mousedownRowId = null;
+		dragMode = null;
+		dragStartX = 0;
+		dragStartY = 0;
+	}
+
+	return {
+		rowMouseDown,
+		rowClick,
+		reset,
+		get dragAnchorId() {
+			return dragAnchorId;
+		}
+	};
+}

tools/ui/src/lib/services/database.service.ts

@@ -3,6 +3,7 @@ import { findDescendantMessages, uuid, filterByLeafNodeId } from '$lib/utils';
 import { IDXDB_TABLES, IDXDB_STORES, STORAGE_APP_NAME } from '$lib/constants';
 import { MessageRole } from '$lib/enums';
 import type { McpServerOverride } from '$lib/types/database';
+import type { ExportedConversation } from '$lib/types/database';
 
 class LlamaUiDatabase extends Dexie {
 	[IDXDB_TABLES.conversations]!: EntityTable<DatabaseConversation, string>;
@@ -206,18 +207,7 @@ export class DatabaseService {
 						await db[IDXDB_TABLES.messages].where('convId').equals(forkId).delete();
 					}
 				} else {
-					// Reparent direct children to deleted conv's parent
-					const conv = await db[IDXDB_TABLES.conversations].get(id);
-					const newParent = conv?.forkedFromConversationId;
-					const directChildren = await db[IDXDB_TABLES.conversations]
-						.filter((c) => c.forkedFromConversationId === id)
-						.toArray();
-
-					for (const child of directChildren) {
-						await db[IDXDB_TABLES.conversations].update(child.id, {
-							forkedFromConversationId: newParent ?? undefined
-						});
-					}
+					await this.reparentDirectChildren(id);
 				}
 
 				await db[IDXDB_TABLES.conversations].delete(id);
@@ -226,6 +216,101 @@ export class DatabaseService {
 		);
 	}
 
+	/**
+	 * Reparents direct children of `parentId` to the nearest surviving
+	 * ancestor (or promotes them to top-level when the immediate parent was
+	 * top-level). Walking skips any ancestor listed in `excludeIds`, since
+	 * those will be deleted in the same batch — leaving a grandchild pointing
+	 * at an `excludeIds` entry would orphan it. Children whose own id is in
+	 * `excludeIds` are dropped from the updates (the bulk-delete pass will
+	 * remove them). `prefetched` may carry a pre-fetched ancestor map to
+	 * avoid repeat reads inside a bulk transaction.
+	 */
+	private static async reparentDirectChildren(
+		parentId: string,
+		excludeIds: ReadonlySet<string> = new Set(),
+		prefetched?: ReadonlyMap<string, DatabaseConversation>
+	): Promise<void> {
+		const conv = prefetched?.get(parentId) ?? (await db[IDXDB_TABLES.conversations].get(parentId));
+		if (!conv) return;
+
+		let newParent = conv.forkedFromConversationId;
+		const visited = new Set<string>([parentId]);
+		while (newParent && excludeIds.has(newParent)) {
+			if (visited.has(newParent)) {
+				newParent = undefined;
+				break;
+			}
+			visited.add(newParent);
+			const next =
+				prefetched?.get(newParent) ?? (await db[IDXDB_TABLES.conversations].get(newParent));
+			if (!next) {
+				newParent = undefined;
+				break;
+			}
+			newParent = next.forkedFromConversationId;
+		}
+
+		const directChildren = await db[IDXDB_TABLES.conversations]
+			.filter((c) => c.forkedFromConversationId === parentId)
+			.toArray();
+
+		const updates: DatabaseConversation[] = [];
+		for (const child of directChildren) {
+			if (excludeIds.has(child.id)) continue;
+			updates.push({ ...child, forkedFromConversationId: newParent });
+		}
+		if (updates.length === 0) return;
+		await db[IDXDB_TABLES.conversations].bulkPut(updates);
+	}
+
+	/**
+	 * Deletes multiple conversations in a single transaction. Each deleted
+	 * conversation has its direct children reparented to the nearest surviving
+	 * ancestor (or promoted to top-level). Children also in `ids` are dropped
+	 * entirely rather than reparented.
+	 *
+	 * @param ids - Conversation IDs to delete
+	 */
+	static async bulkDeleteConversations(ids: string[]): Promise<void> {
+		const cleanIds = ids.filter((id): id is string => typeof id === 'string' && id.length > 0);
+		if (cleanIds.length === 0) return;
+		const idSet = new Set(cleanIds);
+
+		await db.transaction(
+			'rw',
+			[db[IDXDB_TABLES.conversations], db[IDXDB_TABLES.messages]],
+			async () => {
+				// Pre-load each to-delete conversation so the per-id reparent
+				// walk-up doesn't ping-pong the same ancestry chain.
+				const prefetched = new Map<string, DatabaseConversation>();
+				let frontier = [...cleanIds];
+				const requested = new Set<string>(frontier);
+				while (frontier.length > 0) {
+					const fetched = await db[IDXDB_TABLES.conversations].bulkGet(frontier);
+					frontier = [];
+					for (let i = 0; i < fetched.length; i++) {
+						const conv = fetched[i];
+						if (!conv || !conv.id) continue;
+						prefetched.set(conv.id, conv);
+						const ancestor = conv.forkedFromConversationId;
+						if (ancestor && !prefetched.has(ancestor) && !requested.has(ancestor)) {
+							frontier.push(ancestor);
+							requested.add(ancestor);
+						}
+					}
+				}
+
+				for (const id of cleanIds) {
+					await this.reparentDirectChildren(id, idSet, prefetched);
+				}
+
+				await db[IDXDB_TABLES.conversations].bulkDelete(cleanIds);
+				await db[IDXDB_TABLES.messages].where('convId').anyOf(cleanIds).delete();
+			}
+		);
+	}
+
 	/**
 	 * Deletes a message and removes it from its parent's children array.
 	 *
@@ -319,6 +404,43 @@ export class DatabaseService {
 		return await db[IDXDB_TABLES.messages].where('convId').equals(convId).sortBy('timestamp');
 	}
 
+	/**
+	 * Loads multiple conversations with all of their messages in two bulk
+	 * reads. Missing conversations are silently omitted from the result.
+	 *
+	 * @param convIds - Conversation IDs to load
+	 * @returns Map of id -> { conv, messages }. Messages are sorted ascending by timestamp.
+	 */
+	static async getConversationsWithMessages(
+		convIds: string[]
+	): Promise<Map<string, ExportedConversation>> {
+		const result = new Map<string, ExportedConversation>();
+		const cleanIds = convIds.filter((id): id is string => typeof id === 'string' && id.length > 0);
+		if (cleanIds.length === 0) return result;
+
+		const [convs, allMessages] = await Promise.all([
+			db[IDXDB_TABLES.conversations].bulkGet(cleanIds),
+			db[IDXDB_TABLES.messages].where('convId').anyOf(cleanIds).toArray()
+		]);
+
+		const messagesByConv = new Map<string, DatabaseMessage[]>();
+		for (const msg of allMessages) {
+			const bucket = messagesByConv.get(msg.convId);
+			if (bucket) bucket.push(msg);
+			else messagesByConv.set(msg.convId, [msg]);
+		}
+
+		for (let i = 0; i < cleanIds.length; i++) {
+			const conv = convs[i];
+			if (!conv) continue;
+			const messages = (messagesByConv.get(conv.id) ?? []).sort(
+				(a, b) => a.timestamp - b.timestamp
+			);
+			result.set(conv.id, { conv, messages });
+		}
+		return result;
+	}
+
 	/**
 	 * Updates a conversation.
 	 *
@@ -360,6 +482,38 @@ export class DatabaseService {
 		return newPinnedState;
 	}
 
+	/**
+	 * Toggles the pinned status of each conversation in `ids` inside a single
+	 * transaction. Treats `pinned === undefined` as `false`, matching the
+	 * semantics of {@link toggleConversationPin} where `!undefined` evaluates
+	 * to `true`. Returns the resulting pinned state for every id that was
+	 * updated; missing ids are omitted from the map.
+	 *
+	 * @param ids - Conversation IDs to toggle
+	 * @returns Map of id -> new pinned state
+	 */
+	static async bulkToggleConversationPins(ids: string[]): Promise<Map<string, boolean>> {
+		const cleanIds = ids.filter((id): id is string => typeof id === 'string' && id.length > 0);
+		const result = new Map<string, boolean>();
+		if (cleanIds.length === 0) return result;
+
+		const now = Date.now();
+		await db.transaction('rw', db[IDXDB_TABLES.conversations], async () => {
+			const convs = await db[IDXDB_TABLES.conversations].bulkGet(cleanIds);
+			const updates: DatabaseConversation[] = [];
+			for (let i = 0; i < cleanIds.length; i++) {
+				const conv = convs[i];
+				if (!conv) continue;
+				const newPinned = !conv.pinned;
+				updates.push({ ...conv, pinned: newPinned, lastModified: now });
+				result.set(cleanIds[i], newPinned);
+			}
+			if (updates.length === 0) return;
+			await db[IDXDB_TABLES.conversations].bulkPut(updates);
+		});
+		return result;
+	}
+
 	/**
 	 * Updates the conversation's current node (active branch).
 	 * This determines which conversation path is currently being viewed.

tools/ui/src/lib/stores/agentic.svelte.ts

@@ -288,6 +288,7 @@ class AgenticStore {
 		const hasTools =
 			mcpStore.hasEnabledServers(perChatOverrides) ||
 			toolsStore.builtinTools.length > 0 ||
+			toolsStore.frontendTools.length > 0 ||
 			toolsStore.customTools.length > 0;
 		return {
 			enabled: hasTools && DEFAULT_AGENTIC_CONFIG.enabled,

tools/ui/src/lib/stores/chat.svelte.ts

@@ -1596,7 +1596,7 @@ class ChatStore {
 			conversationsStore.updateMessageAtIndex(messageIndex, { content: newContent });
 			await DatabaseService.updateMessage(messageId, { content: newContent });
 			if (isFirstUserMessage && newContent.trim())
-				await conversationsStore.updateConversationTitleWithConfirmation(
+				await conversationsStore.updateConversationName(
 					activeConv.id,
 					generateConversationTitle(newContent, Boolean(config().titleGenerationUseFirstLine))
 				);
@@ -2113,7 +2113,7 @@ class ChatStore {
 			const rootMessage = allMessages.find((m) => m.type === 'root' && m.parent === null);
 
 			if (rootMessage && msg.parent === rootMessage.id && newContent.trim()) {
-				await conversationsStore.updateConversationTitleWithConfirmation(
+				await conversationsStore.updateConversationName(
 					activeConv.id,
 					generateConversationTitle(newContent, Boolean(config().titleGenerationUseFirstLine))
 				);
@@ -2187,7 +2187,7 @@ class ChatStore {
 
 			conversationsStore.updateConversationTimestamp();
 			if (isFirstUserMessage && newContent.trim())
-				await conversationsStore.updateConversationTitleWithConfirmation(
+				await conversationsStore.updateConversationName(
 					activeConv.id,
 					generateConversationTitle(newContent, Boolean(config().titleGenerationUseFirstLine))
 				);
@@ -2428,7 +2428,8 @@ class ChatStore {
 
 		if (currentConfig.samplers) apiOptions.samplers = currentConfig.samplers;
 
-		apiOptions.backend_sampling = currentConfig.backend_sampling;
+		if (hasValue(currentConfig.backend_sampling))
+			apiOptions.backend_sampling = currentConfig.backend_sampling;
 
 		if (currentConfig.customJson) apiOptions.custom = currentConfig.customJson;
 

tools/ui/src/lib/stores/conversations.svelte.ts

@@ -108,9 +108,6 @@ class ConversationsStore {
 		localStorage.setItem(REASONING_EFFORT_DEFAULT_LOCALSTORAGE_KEY, this.pendingReasoningEffort);
 	}
 
-	/** Callback for title update confirmation dialog */
-	titleUpdateConfirmationCallback?: (currentTitle: string, newTitle: string) => Promise<boolean>;
-
 	/**
 	 * Callback for updating message content in chatStore.
 	 * Registered by chatStore to enable cross-store updates without circular dependency.
@@ -209,15 +206,6 @@ class ConversationsStore {
 		return undefined;
 	}
 
-	/**
-	 * Sets the callback function for title update confirmations
-	 */
-	setTitleUpdateConfirmationCallback(
-		callback: (currentTitle: string, newTitle: string) => Promise<boolean>
-	): void {
-		this.titleUpdateConfirmationCallback = callback;
-	}
-
 	/**
 	 *
 	 *
@@ -386,6 +374,123 @@ class ConversationsStore {
 		}
 	}
 
+	/**
+	 * Deletes multiple conversations in sequence.
+	 * Mirrors deleteConversation() per-id; navigates to NEW_CHAT only if the
+	 * currently-open chat was among the deleted ones.
+	 * @param convIds - Conversation IDs to delete
+	 */
+	async bulkDeleteConversations(convIds: string[]): Promise<void> {
+		if (convIds.length === 0) return;
+
+		try {
+			const idsToRemove = new SvelteSet(convIds);
+			// Collect all descendants recursively so the local cache stays consistent
+			// even when deleteWithForks is omitted.
+			const queue = [...convIds];
+			while (queue.length > 0) {
+				const parentId = queue.pop()!;
+				for (const c of this.conversations) {
+					if (c.forkedFromConversationId === parentId && !idsToRemove.has(c.id)) {
+						idsToRemove.add(c.id);
+						queue.push(c.id);
+					}
+				}
+			}
+
+			const activeWasDeleted =
+				this.activeConversation !== null && idsToRemove.has(this.activeConversation.id);
+
+			await DatabaseService.bulkDeleteConversations([...idsToRemove]);
+
+			this.conversations = this.conversations.filter((c) => !idsToRemove.has(c.id));
+
+			if (activeWasDeleted) {
+				this.clearActiveConversation();
+				await goto(ROUTES.NEW_CHAT);
+			}
+
+			toast.success(
+				convIds.length === 1 ? 'Conversation deleted' : `${convIds.length} conversations deleted`
+			);
+		} catch (error) {
+			console.error('Failed to bulk delete conversations:', error);
+			toast.error('Failed to delete conversations');
+		}
+	}
+
+	/**
+	 * Toggles the pinned state of each conversation individually.
+	 * Mixed-pin selections are intentionally not normalised here; the bulk
+	 * action UI surfaces them as a disabled mixed-state instead.
+	 * @param convIds - Conversation IDs to toggle
+	 */
+	async bulkToggleConversationPin(convIds: string[]): Promise<void> {
+		if (convIds.length === 0) return;
+
+		try {
+			const updates = await DatabaseService.bulkToggleConversationPins(convIds);
+
+			const activeId = this.activeConversation?.id;
+			if (activeId && updates.has(activeId)) {
+				this.activeConversation = {
+					...this.activeConversation!,
+					pinned: updates.get(activeId)!
+				};
+			}
+			for (let i = 0; i < this.conversations.length; i++) {
+				const newPinned = updates.get(this.conversations[i].id);
+				if (newPinned !== undefined) this.conversations[i].pinned = newPinned;
+			}
+			this.conversations = [...this.conversations];
+
+			toast.success(
+				convIds.length === 1
+					? 'Conversation pin toggled'
+					: `Updated pin state for ${convIds.length} conversations`
+			);
+		} catch (error) {
+			console.error('Failed to bulk toggle pin:', error);
+			toast.error('Failed to update pin state');
+		}
+	}
+
+	/**
+	 * Bundles the given conversations into a single zip archive and triggers a
+	 * browser download (one JSONL file per conversation).
+	 * @param convIds - Conversation IDs to export
+	 */
+	async bulkExportConversations(convIds: string[]): Promise<void> {
+		if (convIds.length === 0) return;
+
+		try {
+			const fetched = await DatabaseService.getConversationsWithMessages(convIds);
+
+			const activeId = this.activeConversation?.id;
+			const overridden = fetched.get(activeId ?? '');
+			if (overridden && activeId) {
+				overridden.conv = { ...this.activeConversation! };
+			}
+
+			const exported = [...fetched.values()];
+			if (exported.length === 0) {
+				toast.error('No conversations to export');
+				return;
+			}
+
+			this.downloadConversationsArchive(exported);
+
+			toast.success(
+				exported.length === 1
+					? 'Conversation exported'
+					: `${exported.length} conversations exported`
+			);
+		} catch (error) {
+			console.error('Failed to bulk export conversations:', error);
+			toast.error('Failed to export conversations');
+		}
+	}
+
 	/**
 	 *
 	 *
@@ -484,38 +589,6 @@ class ConversationsStore {
 		}
 	}
 
-	/**
-	 * Updates conversation title with optional confirmation dialog based on settings
-	 * @param convId - The conversation ID to update
-	 * @param newTitle - The new title content
-	 * @returns True if title was updated, false if cancelled
-	 */
-	async updateConversationTitleWithConfirmation(
-		convId: string,
-		newTitle: string
-	): Promise<boolean> {
-		try {
-			const currentConfig = config();
-
-			if (currentConfig.askForTitleConfirmation && this.titleUpdateConfirmationCallback) {
-				const conversation = await DatabaseService.getConversation(convId);
-				if (!conversation) return false;
-
-				const shouldUpdate = await this.titleUpdateConfirmationCallback(
-					conversation.name,
-					newTitle
-				);
-				if (!shouldUpdate) return false;
-			}
-
-			await this.updateConversationName(convId, newTitle);
-			return true;
-		} catch (error) {
-			console.error('Failed to update conversation title with confirmation:', error);
-			return false;
-		}
-	}
-
 	/**
 	 * Updates conversation lastModified timestamp and moves it to top of list
 	 */
@@ -581,7 +654,7 @@ class ConversationsStore {
 					newFirstUserMessage.id !== currentFirstUserMessage.id ||
 					newFirstUserMessage.content.trim() !== currentFirstUserMessage.content.trim())
 			) {
-				await this.updateConversationTitleWithConfirmation(
+				await this.updateConversationName(
 					this.activeConversation.id,
 					generateConversationTitle(
 						newFirstUserMessage.content,
@@ -1162,6 +1235,10 @@ export const isConversationsInitialized = () => conversationsStore.isInitialized
 /**
  * Builds a flat tree of conversations with depth levels for nested forks.
  * Accepts a pre-filtered list so search filtering stays in the component.
+ *
+ * Output order matches the sidebar render exactly: pinned first, then
+ * unpinned by lastModified desc, with forks interleaved under their parents.
+ * Range-select / marquee in the sidebar rely on this alignment.
  */
 
 // Pinned conversations first, then by lastModified descending

tools/ui/src/lib/types/settings.d.ts

@@ -27,6 +27,8 @@ export interface SettingsEntry {
 	type: SettingsFieldType;
 	section?: string;
 	options?: Array<{ value: string; label: string; icon: Component }>;
+	/** Options rendered for RADIO fields. Each entry maps a `value` (the radio's selected value) to the underlying config `key` whose boolean state mirrors it. */
+	radioOptions?: Array<{ value: string; label: string; key: string; isExperimental?: boolean }>;
 	isExperimental?: boolean;
 	isPositiveInteger?: boolean;
 	dependsOn?: string;
@@ -53,6 +55,8 @@ export interface SettingsFieldConfig {
 	dependsOn?: string;
 	help?: string;
 	options?: Array<{ value: string; label: string; icon?: typeof Icon }>;
+	/** Options rendered for RADIO fields. Each entry maps a `value` (the radio's selected value) to the underlying config `key` whose boolean state mirrors it. */
+	radioOptions?: Array<{ value: string; label: string; key: string; isExperimental?: boolean }>;
 }
 
 /** Re-exported for backward compatibility. */

tools/ui/src/routes/+layout.svelte

@@ -7,12 +7,11 @@
 	import { untrack } from 'svelte';
 	import { onMount } from 'svelte';
 
-	import { SidebarNavigation, DialogConversationTitleUpdate } from '$lib/components/app';
+	import { SidebarNavigation } from '$lib/components/app';
 	import { PwaMetaTags, PwaRefreshAlert } from '$lib/components/pwa';
 	import { pwaAssetsHead } from 'virtual:pwa-assets/head';
 
 	import { chatStore } from '$lib/stores/chat.svelte';
-	import { conversationsStore } from '$lib/stores/conversations.svelte';
 	import * as Tooltip from '$lib/components/ui/tooltip';
 	import { isRouterMode, serverStore } from '$lib/stores/server.svelte';
 	import { config, settingsStore } from '$lib/stores/settings.svelte';
@@ -46,11 +45,6 @@
 
 	let showBuildVersion = $derived(config()[SETTINGS_KEYS.SHOW_BUILD_VERSION] as boolean);
 
-	let titleUpdateDialogOpen = $state(false);
-	let titleUpdateCurrentTitle = $state('');
-	let titleUpdateNewTitle = $state('');
-	let titleUpdateResolve: ((value: boolean) => void) | null = null;
-
 	// Keep the hook object intact: destructuring needRefreshByStorage reads the getter once and freezes it
 	const pwa = usePwa();
 	const { needRefresh, updateServiceWorker } = pwa;
@@ -135,24 +129,6 @@
 		});
 	}
 
-	function handleTitleUpdateCancel() {
-		titleUpdateDialogOpen = false;
-
-		if (titleUpdateResolve) {
-			titleUpdateResolve(false);
-			titleUpdateResolve = null;
-		}
-	}
-
-	function handleTitleUpdateConfirm() {
-		titleUpdateDialogOpen = false;
-
-		if (titleUpdateResolve) {
-			titleUpdateResolve(true);
-			titleUpdateResolve = null;
-		}
-	}
-
 	onMount(() => {
 		updateFavicon();
 		// snapshot of every backend running stream on first load, populates the sidebar spinners
@@ -264,20 +240,6 @@
 	$effect(() => {
 		checkApiKey();
 	});
-
-	// Set up title update confirmation callback
-	$effect(() => {
-		conversationsStore.setTitleUpdateConfirmationCallback(
-			async (currentTitle: string, newTitle: string) => {
-				return new Promise<boolean>((resolve) => {
-					titleUpdateCurrentTitle = currentTitle;
-					titleUpdateNewTitle = newTitle;
-					titleUpdateResolve = resolve;
-					titleUpdateDialogOpen = true;
-				});
-			}
-		);
-	});
 </script>
 
 <svelte:head>
@@ -319,14 +281,6 @@
 	<ModeWatcher />
 
 	<Toaster richColors />
-
-	<DialogConversationTitleUpdate
-		bind:open={titleUpdateDialogOpen}
-		currentTitle={titleUpdateCurrentTitle}
-		newTitle={titleUpdateNewTitle}
-		onConfirm={handleTitleUpdateConfirm}
-		onCancel={handleTitleUpdateCancel}
-	/>
 </Tooltip.Provider>
 
 <!-- PWA update prompt + version -->