ui : models discover dialog

Add the Models Discover explorer behind a new sidebar action and full-screen
dialog: a searchable HuggingFace GGUF list (org avatars, capability icons,
catalog size ranges) and a detail pane with header, metadata chips, the
sanitized model-card readme and the download area - per-quant action chips
grouped by bit depth, sidecar queuing and a scrollable serve-command preview
with dynamic quant selects. Searchable dropdowns gain sticky headers/footers.

Assisted-by: pi:GLM-5.3-Flash

.devops/nix/package.nix

@@ -31,7 +31,7 @@
     ]
     && blas.meta.available,
   useCuda ? config.cudaSupport,
-  useMetalKit ? stdenv.isAarch64 && stdenv.isDarwin,
+  useMetalKit ? stdenv.hostPlatform.isAarch64 && stdenv.hostPlatform.isDarwin,
   # Increases the runtime closure size by ~700M
   useMpi ? false,
   useRocm ? config.rocmSupport,
@@ -92,7 +92,7 @@ let
 
   cudaBuildInputs = with cudaPackages; [
     cuda_cudart
-    cuda_cccl # <nv/target>
+    cccl # <nv/target>
     libcublas
   ];
 
@@ -166,7 +166,7 @@ effectiveStdenv.mkDerivation (finalAttrs: {
   # `xcrun` is used find the path of the Metal compiler, which is varible
   # and not on $PATH
   # see https://github.com/ggml-org/llama.cpp/pull/6118 for discussion
-  __noChroot = effectiveStdenv.isDarwin && useMetalKit && precompileMetalShaders;
+  __noChroot = effectiveStdenv.hostPlatform.isDarwin && useMetalKit && precompileMetalShaders;
 
   nativeBuildInputs =
     [
@@ -181,10 +181,10 @@ effectiveStdenv.mkDerivation (finalAttrs: {
       autoAddDriverRunpath
     ]
     ++ optionals (effectiveStdenv.hostPlatform.isGnu && enableStatic) [ glibc.static ]
-    ++ optionals (effectiveStdenv.isDarwin && useMetalKit && precompileMetalShaders) [ xcrunHost ];
+    ++ optionals (effectiveStdenv.hostPlatform.isDarwin && useMetalKit && precompileMetalShaders) [ xcrunHost ];
 
   buildInputs =
-    optionals effectiveStdenv.isDarwin darwinBuildInputs
+    optionals effectiveStdenv.hostPlatform.isDarwin darwinBuildInputs
     ++ optionals useCuda cudaBuildInputs
     ++ optionals useMpi [ mpi ]
     ++ optionals useRocm rocmBuildInputs
@@ -245,7 +245,7 @@ effectiveStdenv.mkDerivation (finalAttrs: {
 
     # Configurations that are known to result in build failures. Can be
     # overridden by importing Nixpkgs with `allowBroken = true`.
-    broken = (useMetalKit && !effectiveStdenv.isDarwin);
+    broken = (useMetalKit && !effectiveStdenv.hostPlatform.isDarwin);
 
     description = "Inference of LLaMA model in pure C/C++${descriptionSuffix}";
     homepage = "https://github.com/ggml-org/llama.cpp/";

common/download.cpp

@@ -8,6 +8,7 @@
 #include "json.h"
 
 #include <algorithm>
+#include <cctype>
 #include <filesystem>
 #include <fstream>
 #include <future>
@@ -534,15 +535,18 @@ static gguf_split_info get_gguf_split_info(const std::string & path) {
 }
 
 // Q4_0 -> 4, F16 -> 16, NVFP4 -> 4, Q8_K_M -> 8, etc
-static int extract_quant_bits(const std::string & filename) {
-    auto split = get_gguf_split_info(filename);
+static int quant_bits_from_tag(const std::string & tag) {
+    auto pos = tag.find_first_of("0123456789");
 
-    auto pos = split.tag.find_first_of("0123456789");
     if (pos == std::string::npos) {
         return 0;
     }
 
-    return std::stoi(split.tag.substr(pos));
+    return std::stoi(tag.substr(pos));
+}
+
+static int extract_quant_bits(const std::string & filename) {
+    return quant_bits_from_tag(get_gguf_split_info(filename).tag);
 }
 
 static hf_cache::hf_files get_split_files(const hf_cache::hf_files & files,
@@ -563,12 +567,127 @@ static hf_cache::hf_files get_split_files(const hf_cache::hf_files & files,
     return result;
 }
 
-// pick the best sibling GGUF whose filename contains `keyword` (e.g. "mmproj" / "mtp"),
+// sidecar filename tokens, as used in `<quant>-<sidecar>` download tags,
+// e.g. `Q4_0-mtp` for `mtp-Model-Q4_0.gguf`, `BF16-mmproj` for `mmproj-BF16.gguf`
+static const std::vector<std::string> sidecar_tokens = {
+    "mtp", "eagle3", "dflash", "dspark", "mmproj", "imatrix",
+};
+
+static bool iequals(const std::string & a, const std::string & b) {
+    return a.size() == b.size() &&
+        std::equal(a.begin(), a.end(), b.begin(), [](char x, char y) {
+            return std::tolower((unsigned char) x) == std::tolower((unsigned char) y);
+        });
+}
+
+// split a `<quant>-<sidecar>` tag into its parts, e.g. `Q4_0-mtp` -> {`Q4_0`, `mtp`};
+// a bare sidecar tag (`mtp`) yields an empty quant; no sidecar yields an empty token
+static std::pair<std::string, std::string> split_sidecar_tag(const std::string & tag) {
+    for (const auto & t : sidecar_tokens) {
+        if (tag.size() > t.size() + 1 && iequals(tag.substr(tag.size() - t.size() - 1), "-" + t)) {
+            return { tag.substr(0, tag.size() - t.size() - 1), t };
+        }
+        if (iequals(tag, t)) {
+            return { "", t };
+        }
+    }
+    return { tag, "" };
+}
+
+// filename with directory and extension removed, e.g. `sub/mtp-Model-Q4_0.gguf` -> `mtp-Model-Q4_0`
+static std::string stem_of(const std::string & path) {
+    std::string base = path;
+    if (auto pos = base.rfind('/'); pos != std::string::npos) {
+        base = base.substr(pos + 1);
+    }
+    string_remove_suffix(base, ".gguf");
+    return base;
+}
+
+// a sidecar file carries its token as a name segment in any position and case,
+// e.g. `mmproj-Model-F16.gguf`, `Model-mtp-Q4_0.gguf`, `Model-Q4_0-mtp.gguf`
+// or the short `mmproj-F16.gguf`; the token must be a whole segment, so an
+// unrelated name that merely contains it (`smtp-Model.gguf`) is a plain model
+static std::string sidecar_token_of(const std::string & path) {
+    std::string base = stem_of(path);
+
+    for (char & c : base) {
+        c = (char) std::tolower((unsigned char) c);
+    }
+    if (base.empty()) {
+        return {};
+    }
+
+    for (const auto & t : sidecar_tokens) {
+        if (base == t) {
+            return t; // the sidecar file itself, e.g. `imatrix.gguf`
+        }
+        if (base.rfind(t + "-", 0) == 0) {
+            return t; // `mtp-Model-Q4_0.gguf`
+        }
+        if (base.find("-" + t + "-") != std::string::npos) {
+            return t; // `Model-mtp-Q4_0.gguf`
+        }
+        // `Model-Q4_0-mtp.gguf`, optionally with a `-draft` tail
+        if (string_ends_with(base, "-" + t) || string_ends_with(base, "-" + t + "-draft")) {
+            return t;
+        }
+    }
+    return {};
+}
+
+// name with the sidecar token segment removed, lowercased for tag parsing,
+// e.g. `Model-MTP-Q4_0` -> `model-q4_0`
+static std::string strip_sidecar_token(const std::string & base, const std::string & token) {
+    std::string lower = base;
+
+    for (char & c : lower) {
+        c = (char) std::tolower((unsigned char) c);
+    }
+
+    if (lower == token) {
+        return {};
+    }
+    if (lower.rfind(token + "-", 0) == 0) {
+        return lower.substr(token.size() + 1);
+    }
+    const std::string seg = "-" + token + "-";
+    if (auto pos = lower.find(seg); pos != std::string::npos) {
+        return lower.substr(0, pos) + "-" + lower.substr(pos + seg.size());
+    }
+    if (string_ends_with(lower, "-" + token + "-draft")) {
+        return lower.substr(0, lower.size() - token.size() - 7);
+    }
+    if (string_ends_with(lower, "-" + token)) {
+        return lower.substr(0, lower.size() - token.size() - 1);
+    }
+    return lower;
+}
+
+// the quant a sidecar file belongs to, from its name with the token stripped;
+// a short-form name (`mmproj-F16.gguf`) leaves the bare quant, which has no
+// `-` separator for the tag regex, so it becomes the tag directly
+static std::string sidecar_quant(const std::string & path, const std::string & token) {
+    std::string name = strip_sidecar_token(stem_of(path), token);
+    std::string tag  = get_gguf_split_info(name + ".gguf").tag;
+
+    if (tag.empty() && name.find('-') == std::string::npos) {
+        for (char & c : name) {
+            c = (char) std::toupper((unsigned char) c);
+        }
+        tag = name;
+    }
+
+    return tag;
+}
+
+// pick the best sibling GGUF carrying the sidecar `token` (e.g. "mmproj" / "mtp"),
 // preferring deeper shared directory prefix with the model, then exact `tag` match,
 // then closest quantization to the tag when given, or to the model otherwise
+// an empty `model` skips the directory constraint: the sidecar is matched by tag alone
 static hf_cache::hf_file find_best_sibling(const hf_cache::hf_files & files,
                                            const std::string        & model,
-                                           const std::string        & keyword,
+                                           const std::string        & token,
                                            const std::string        & tag = "") {
     hf_cache::hf_file best;
     size_t best_depth = 0;
@@ -589,32 +708,32 @@ static hf_cache::hf_file find_best_sibling(const hf_cache::hf_files & files,
         model_bits = extract_quant_bits(model);
     }
     auto model_parts = string_split<std::string>(model, '/');
-    auto model_dir = model_parts.end() - 1;
 
     for (const auto & f : files) {
-        if (!string_ends_with(f.path, ".gguf") ||
-            f.path.find(keyword) == std::string::npos) {
+        if (sidecar_token_of(f.path) != token) {
             continue;
         }
 
         auto sib_parts = string_split<std::string>(f.path, '/');
         auto sib_dir = sib_parts.end() - 1;
 
-        auto [_, dir] = std::mismatch(model_parts.begin(), model_dir,
-                                      sib_parts.begin(), sib_dir);
-        if (dir != sib_dir) {
-            continue;
+        size_t depth = 0;
+        if (!model.empty()) {
+            auto model_dir = model_parts.end() - 1;
+            auto [_, dir] = std::mismatch(model_parts.begin(), model_dir,
+                                          sib_parts.begin(), sib_dir);
+            if (dir != sib_dir) {
+                continue;
+            }
+            depth = dir - sib_parts.begin();
         }
 
-        size_t depth = dir - sib_parts.begin();
-        auto bits = extract_quant_bits(f.path);
-        auto diff = std::abs(bits - model_bits);
-
-        std::string path_upper = f.path;
-        for (char & c : path_upper) {
-            c = (char) std::toupper((unsigned char) c);
-        }
-        bool exact = !tag_upper.empty() && path_upper.find("-" + tag_upper + ".") != std::string::npos;
+        // rank by the quant the sidecar belongs to, with the token segment
+        // stripped from its name
+        auto tag   = sidecar_quant(f.path, token);
+        auto bits  = quant_bits_from_tag(tag);
+        auto diff  = std::abs(bits - model_bits);
+        bool exact = !tag_upper.empty() && tag == tag_upper;
 
         if (!found || depth > best_depth ||
             (depth == best_depth && exact && !best_exact) ||
@@ -637,43 +756,31 @@ static hf_cache::hf_file find_best_mmproj(const hf_cache::hf_files & files,
 static hf_cache::hf_file find_best_mtp(const hf_cache::hf_files & files,
                                        const std::string        & model,
                                        const std::string        & tag = "") {
-    return find_best_sibling(files, model, "mtp-", tag);
+    return find_best_sibling(files, model, "mtp", tag);
 }
 
 static hf_cache::hf_file find_best_eagle3(const hf_cache::hf_files & files,
                                           const std::string        & model,
                                           const std::string        & tag = "") {
-    return find_best_sibling(files, model, "eagle3-", tag);
+    return find_best_sibling(files, model, "eagle3", tag);
 }
 
 static hf_cache::hf_file find_best_dflash(const hf_cache::hf_files & files,
                                           const std::string        & model,
                                           const std::string        & tag = "") {
-    return find_best_sibling(files, model, "dflash-", tag);
+    return find_best_sibling(files, model, "dflash", tag);
 }
 
 static hf_cache::hf_file find_best_dspark(const hf_cache::hf_files & files,
                                           const std::string        & model,
                                           const std::string        & tag = "") {
-    return find_best_sibling(files, model, "dspark-", tag);
+    return find_best_sibling(files, model, "dspark", tag);
 }
 
+// a plain model file: a GGUF whose name carries no sidecar token segment,
+// so `smtp-Model.gguf` counts and every sidecar form does not
 static bool gguf_filename_is_model(const std::string & filepath) {
-    if (!string_ends_with(filepath, ".gguf")) {
-        return false;
-    }
-
-    std::string filename = filepath;
-    if (auto pos = filename.rfind('/'); pos != std::string::npos) {
-        filename = filename.substr(pos + 1);
-    }
-
-    return filename.find("mmproj")  == std::string::npos &&
-           filename.find("imatrix") == std::string::npos &&
-           filename.find("mtp-")    == std::string::npos &&
-           filename.find("eagle3-") == std::string::npos &&
-           filename.find("dflash-") == std::string::npos &&
-           filename.find("dspark-") == std::string::npos;
+    return string_ends_with(filepath, ".gguf") && sidecar_token_of(filepath).empty();
 }
 
 static hf_cache::hf_file find_best_model(const hf_cache::hf_files & files,
@@ -765,8 +872,27 @@ common_download_hf_plan common_download_get_hf_plan(const common_params_model &
         }
     } else {
         primary = find_best_model(all, tag);
+
+        // a `<quant>-<sidecar>` tag (e.g. `Q4_0-mtp`) requests that sidecar alone;
+        // every token-bearing file is a sidecar (find_best_model skips them),
+        // so the sidecar resolves here whenever the tag carries one
+        auto [base_tag, sidecar] = split_sidecar_tag(tag);
+
+        if (primary.path.empty() && !sidecar.empty()) {
+            auto found = find_best_sibling(all, "", sidecar, base_tag);
+
+            if (!found.path.empty()) {
+                if      (sidecar == "mtp")    plan.mtp    = found;
+                else if (sidecar == "eagle3") plan.eagle3 = found;
+                else if (sidecar == "dflash") plan.dflash = found;
+                else if (sidecar == "dspark") plan.dspark = found;
+                else                          plan.mmproj = found;
+            }
+        }
+
         // a requested sidecar can resolve on its own, without a full model of the same tag
-        if (primary.path.empty() && !opts.download_mtp && !opts.download_dflash && !opts.download_eagle3 && !opts.download_dspark) {
+        if (primary.path.empty() && sidecar.empty() &&
+            !opts.download_mtp && !opts.download_dflash && !opts.download_eagle3 && !opts.download_dspark) {
             LOG_ERR("%s: no GGUF files found in repository %s\n", __func__, repo.c_str());
             list_available_gguf_files(all);
             return plan;
@@ -794,7 +920,7 @@ common_download_hf_plan common_download_get_hf_plan(const common_params_model &
         plan.dspark = find_best_dspark(all, primary.path, tag);
     }
 
-    if (primary.path.empty() &&
+    if (primary.path.empty() && plan.mmproj.local_path.empty() &&
         plan.mtp.local_path.empty() && plan.dflash.local_path.empty() && plan.eagle3.local_path.empty() && plan.dspark.local_path.empty()) {
         LOG_ERR("%s: no GGUF files found in repository %s\n", __func__, repo.c_str());
         list_available_gguf_files(all);
@@ -968,17 +1094,34 @@ std::vector<common_cached_model_info> common_list_cached_models() {
     auto files = hf_cache::get_cached_files();
 
     for (const auto & f : files) {
-        auto split = get_gguf_split_info(f.path);
-        if (split.index != 1 || split.tag.empty() ||
-            split.prefix.find("mmproj")  != std::string::npos ||
-            split.prefix.find("mtp-")    != std::string::npos ||
-            split.prefix.find("eagle3-") != std::string::npos ||
-            split.prefix.find("dflash-") != std::string::npos ||
-            split.prefix.find("dspark-") != std::string::npos) {
-            continue;
+        // a sidecar file is listed under its own `<quant>-<sidecar>` tag, so a
+        // cached `mtp-Model-Q4_0.gguf`, `Model-mtp-Q4_0.gguf`, `Model-Q4_0-mtp.gguf`
+        // or short `mmproj-F16.gguf` shows up as `<repo>:Q4_0-mtp` / `<repo>:F16-mmproj`;
+        // files whose name carries no token stay loadable models
+        auto token = sidecar_token_of(f.path);
+
+        std::string tag;
+        if (token.empty()) {
+            auto split = get_gguf_split_info(f.path);
+
+            if (split.index != 1 || split.tag.empty()) {
+                continue;
+            }
+
+            tag = split.tag;
+        } else {
+            tag = sidecar_quant(f.path, token);
+
+            // a bare sidecar file has no tag to request it by, stay hidden
+            if (tag.empty()) {
+                continue;
+            }
+
+            tag += "-" + token;
         }
-        if (seen.insert(f.repo_id + ":" + split.tag).second) {
-            result.push_back({f.repo_id, split.tag});
+
+        if (seen.insert(f.repo_id + ":" + tag).second) {
+            result.push_back({f.repo_id, tag});
         }
     }
 
@@ -1014,7 +1157,15 @@ bool common_download_remove(const std::string & hf_repo_with_tag) {
         return hf_cache::remove_cached_repo(repo_id);
     }
 
-    std::string tag_upper = tag;
+    // a `<quant>-<sidecar>` tag (`Q4_0-mtp`) targets that sidecar alone; a bare
+    // sidecar tag (`mtp`) is ambiguous across quants and is rejected
+    auto [base_tag, sidecar] = split_sidecar_tag(tag);
+    if (!sidecar.empty() && base_tag.empty()) {
+        LOG_ERR("%s: bare sidecar tag '%s': use `<quant>-<sidecar>`\n", __func__, tag.c_str());
+        return false;
+    }
+
+    std::string tag_upper = sidecar.empty() ? tag : base_tag;
     for (char & c : tag_upper) {
         c = (char) std::toupper((unsigned char) c);
     }
@@ -1024,13 +1175,25 @@ bool common_download_remove(const std::string & hf_repo_with_tag) {
         return false;
     }
 
-    // collect snapshot entries whose tag matches
+    // collect the snapshot entries the tag selects; sidecar files keep their
+    // own tags, so a plain quant tag never removes them
     std::vector<fs::path> to_remove;
     for (const auto & f : files) {
-        auto split = get_gguf_split_info(f.path);
-        if (split.tag == tag_upper) {
-            to_remove.emplace_back(f.local_path);
+        auto token = sidecar_token_of(f.path);
+
+        if (sidecar.empty()) {
+            if (!token.empty()) {
+                continue;
+            }
+            if (get_gguf_split_info(f.path).tag != tag_upper) {
+                continue;
+            }
+        } else {
+            if (token != sidecar || sidecar_quant(f.path, sidecar) != tag_upper) {
+                continue;
+            }
         }
+        to_remove.emplace_back(f.local_path);
     }
 
     if (to_remove.empty()) {

common/jinja/caps.cpp

@@ -117,6 +117,7 @@ caps caps_get(jinja::program & prog) {
 
     JJ_DEBUG("%s\n", ">>> Running capability check: typed content");
 
+    bool checks_for_string = false;
     static const std::string content_marker = "STRING_MARKER";
 
     // case: typed content support
@@ -136,6 +137,10 @@ caps caps_get(jinja::program & prog) {
         [&](context &, bool success, value & messages, value &, const std::string & rendered) {
             auto & content = messages->at(0)->at("content");
             caps_print_stats(content, "messages[0].content");
+            if (has_op(content, "test_is_string")) {
+                // checked if content is string
+                checks_for_string = true;
+            }
             bool used_as_array = has_op(content, "selectattr") || has_op(content, "array_access");
             if (used_as_array) {
                 // accessed as an array
@@ -151,6 +156,33 @@ caps caps_get(jinja::program & prog) {
         }
     );
 
+    if (checks_for_string) {
+        caps_try_execute(
+            prog,
+            [&]() {
+                // messages
+                return json::array({
+                    {
+                        {"role", "user"},
+                        {"content", json::array({
+                        })}
+                    }
+                });
+            },
+            nullptr, // ctx_fn
+            nullptr, // tools_fn
+            [&](context &, bool success, value & messages, value &, const std::string &) {
+                auto & content = messages->at(0)->at("content");
+                caps_print_stats(content, "messages[0].content");
+                bool used_as_array = has_op(content, "selectattr") || has_op(content, "array_access");
+                if (used_as_array && success) {
+                    // accessed as an array
+                    result.supports_typed_content = true;
+                }
+            }
+        );
+    }
+
     JJ_DEBUG("%s\n", ">>> Running capability check: system prompt");
 
     // case: system prompt support

common/jinja/runtime.cpp

@@ -412,12 +412,18 @@ value test_expression::execute_impl(context & ctx) {
         throw std::runtime_error("Invalid test expression");
     }
 
-    auto it = builtins.find("test_is_" + test_id);
-    JJ_DEBUG("Test expression %s '%s' %s (using function 'test_is_%s')", operand->type().c_str(), test_id.c_str(), negate ? "(negate)" : "", test_id.c_str());
+    const std::string test_name = "test_is_" + test_id;
+    auto it = builtins.find(test_name);
+    JJ_DEBUG("Test expression %s '%s' %s (using function '%s')", operand->type().c_str(), test_id.c_str(), negate ? "(negate)" : "", test_name.c_str());
     if (it == builtins.end()) {
         throw std::runtime_error("Unknown test '" + test_id + "'");
     }
 
+    if (ctx.is_get_stats) {
+        value_t::stats_t::mark_used(input);
+        input->stats.ops.insert(test_name);
+    }
+
     auto res = it->second(args);
 
     if (negate) {

conversion/hy_v4.py

@@ -9,20 +9,6 @@ from .base import ModelBase, gguf, logger
 from .deepseek import DeepseekV2Model
 
 
-def split_kv_b_proj(weight: torch.Tensor, n_head: int, qk_nope: int, v_head_dim: int):
-    """Split kv_b_proj into k_b (transposed) and v_b, matching DeepSeek MLA absorption.
-
-    weight: [n_head*(qk_nope+v_head_dim), kv_lora_rank].
-    Returns (k_b, v_b): k_b [n_head, kv_lora_rank, qk_nope], v_b [n_head, v_head_dim, kv_lora_rank].
-    """
-    kv_lora = weight.shape[-1]
-    assert weight.shape[0] == n_head * (qk_nope + v_head_dim)
-    kv_b = weight.view(n_head, qk_nope + v_head_dim, kv_lora)
-    k_b, v_b = torch.split(kv_b, [qk_nope, v_head_dim], dim=1)
-    k_b = k_b.transpose(1, 2).contiguous()  # [n_head, kv_lora, qk_nope]
-    return k_b, v_b.contiguous()
-
-
 def split_gate_up(weight: torch.Tensor, moe_intermediate_size: int):
     """Split a fused stacked gate_up expert tensor into (gate, up).
 
@@ -36,6 +22,7 @@ def split_gate_up(weight: torch.Tensor, moe_intermediate_size: int):
 
 
 @ModelBase.register("HYV4ForCausalLM")
+@ModelBase.example("tencent/Hy4-preview")
 class HYV4Model(DeepseekV2Model):
     """HY_V4: DeepSeek-V3 style MLA + MoE with iHC, a gated MLA output and a learnable sink.
 
@@ -54,6 +41,8 @@ class HYV4Model(DeepseekV2Model):
 
     model_arch = gguf.MODEL_ARCH.HY_V4
 
+    merge_expert = False
+
     # tensors a "full" indexer layer must carry
     INDEXER_SUFFIXES = frozenset({
         "self_attn.indexer.wq_b.weight",
@@ -186,6 +175,10 @@ class HYV4Model(DeepseekV2Model):
             )
 
     def prepare_tensors(self):
+        # Hy4-preview for some reason has num_key_value_heads equal to 8, so override it here
+        # without this conversion/deepseek.py fails on assert
+        self.hparams["num_key_value_heads"] = self.hparams["num_attention_heads"]
+
         # validate before the base materializes tensors, so a mismatch fails early
         is_full = self.indexer_is_full()
         if is_full is not None:
@@ -227,85 +220,25 @@ class HYV4Model(DeepseekV2Model):
 
     def modify_tensors(self, data_torch: torch.Tensor, name: str, bid: int | None) -> Iterable[tuple[str, torch.Tensor]]:
         hparams = self.hparams
-        n_head = hparams["num_attention_heads"]
-        qk_nope = hparams["qk_nope_head_dim"]
-        v_head_dim = hparams["v_head_dim"]
         moe_inter = hparams["moe_intermediate_size"]
 
         tn = self.format_tensor_name
 
-        # ---- global (non per-layer) ----
-        if name == "model.embed_tokens.weight":
-            return [(tn(gguf.MODEL_TENSOR.TOKEN_EMBD), data_torch)]
-        if name == "model.norm.weight":
-            return [(tn(gguf.MODEL_TENSOR.OUTPUT_NORM), data_torch)]
-        if name == "lm_head.weight":
-            return [(tn(gguf.MODEL_TENSOR.OUTPUT), data_torch)]
-        if name == "model.hc_head.hc_head_fn":
-            return [(tn(gguf.MODEL_TENSOR.HC_HEAD_FN), data_torch)]
-        if name == "model.hc_head.hc_head_base":
-            return [(tn(gguf.MODEL_TENSOR.HC_HEAD_BASE), data_torch)]
-        if name == "model.hc_head.hc_head_scale":
-            return [(tn(gguf.MODEL_TENSOR.HC_HEAD_SCALE), data_torch)]
-
-        assert bid is not None, f"expected a per-layer tensor, got {name!r}"
-
-        # ---- per-layer, keyed by suffix after 'model.layers.{bid}.' ----
-        suffix = name.split(f"model.layers.{bid}.", 1)[-1]
-
-        # note: q_b_proj and kv_a_proj_with_mqa are mapped straight through (no RoPE permute),
-        # the graph rotates consecutive pairs so the rows need no reordering
-        simple = {
-            "input_layernorm.weight":          (gguf.MODEL_TENSOR.ATTN_NORM, ".weight"),
-            "post_attention_layernorm.weight": (gguf.MODEL_TENSOR.FFN_NORM,  ".weight"),
-            "self_attn.q_a_proj.weight":       (gguf.MODEL_TENSOR.ATTN_Q_A, ".weight"),
-            "self_attn.q_a_layernorm.weight":  (gguf.MODEL_TENSOR.ATTN_Q_A_NORM, ".weight"),
-            "self_attn.q_b_proj.weight":       (gguf.MODEL_TENSOR.ATTN_Q_B, ".weight"),
-            "self_attn.kv_a_proj_with_mqa.weight": (gguf.MODEL_TENSOR.ATTN_KV_A_MQA, ".weight"),
-            "self_attn.kv_a_layernorm.weight": (gguf.MODEL_TENSOR.ATTN_KV_A_NORM, ".weight"),
-            "self_attn.o_proj.weight":         (gguf.MODEL_TENSOR.ATTN_OUT, ".weight"),
-            "self_attn.linear_gate.weight":    (gguf.MODEL_TENSOR.ATTN_GATE, ".weight"),
-            "self_attn.learnable_sink_param":  (gguf.MODEL_TENSOR.ATTN_SINKS, ".weight"),
-            "self_attn.indexer.wq_b.weight":   (gguf.MODEL_TENSOR.INDEXER_ATTN_Q_B, ".weight"),
-            "self_attn.indexer.wk.weight":     (gguf.MODEL_TENSOR.INDEXER_ATTN_K, ".weight"),
-            "self_attn.indexer.k_norm.weight": (gguf.MODEL_TENSOR.INDEXER_K_NORM, ".weight"),
-            "self_attn.indexer.k_norm.bias":   (gguf.MODEL_TENSOR.INDEXER_K_NORM, ".bias"),
-            "self_attn.indexer.weights_proj.weight": (gguf.MODEL_TENSOR.INDEXER_PROJ, ".weight"),
-            "hc_attn_layer.hc_pre.hc_fn":      (gguf.MODEL_TENSOR.HC_ATTN_FN, ".weight"),
-            "hc_attn_layer.hc_pre.hc_base":    (gguf.MODEL_TENSOR.HC_ATTN_BASE, ".weight"),
-            "hc_attn_layer.hc_pre.hc_scale":   (gguf.MODEL_TENSOR.HC_ATTN_SCALE, ".weight"),
-            "hc_mlp_layer.hc_pre.hc_fn":       (gguf.MODEL_TENSOR.HC_FFN_FN, ".weight"),
-            "hc_mlp_layer.hc_pre.hc_base":     (gguf.MODEL_TENSOR.HC_FFN_BASE, ".weight"),
-            "hc_mlp_layer.hc_pre.hc_scale":    (gguf.MODEL_TENSOR.HC_FFN_SCALE, ".weight"),
-            "mlp.gate.weight":                 (gguf.MODEL_TENSOR.FFN_GATE_INP, ".weight"),
-            "mlp.gate.e_score_correction.bias":(gguf.MODEL_TENSOR.FFN_EXP_PROBS_B, ".bias"),
-            "mlp.gate_proj.weight":            (gguf.MODEL_TENSOR.FFN_GATE, ".weight"),
-            "mlp.up_proj.weight":              (gguf.MODEL_TENSOR.FFN_UP, ".weight"),
-            "mlp.down_proj.weight":            (gguf.MODEL_TENSOR.FFN_DOWN, ".weight"),
-            "mlp.shared_experts.gate_proj.weight": (gguf.MODEL_TENSOR.FFN_GATE_SHEXP, ".weight"),
-            "mlp.shared_experts.up_proj.weight":   (gguf.MODEL_TENSOR.FFN_UP_SHEXP, ".weight"),
-            "mlp.shared_experts.down_proj.weight": (gguf.MODEL_TENSOR.FFN_DOWN_SHEXP, ".weight"),
-        }
-        if suffix in simple:
-            key, sfx = simple[suffix]
-            return [(tn(key, bid, sfx), data_torch)]
-
-        # kv_b_proj: split into k_b (transposed) and v_b
-        if suffix == "self_attn.kv_b_proj.weight":
-            k_b, v_b = split_kv_b_proj(data_torch, n_head, qk_nope, v_head_dim)
-            return [
-                (tn(gguf.MODEL_TENSOR.ATTN_K_B, bid), k_b),
-                (tn(gguf.MODEL_TENSOR.ATTN_V_B, bid), v_b),
-            ]
-
         # fused stacked experts: split gate_up into gate/up
-        if suffix == "mlp.experts.gate_up_proj":
+        if name.endswith("mlp.experts.gate_up_proj"):
             gate, up = split_gate_up(data_torch, moe_inter)
-            return [
-                (tn(gguf.MODEL_TENSOR.FFN_GATE_EXP, bid), gate),
-                (tn(gguf.MODEL_TENSOR.FFN_UP_EXP, bid), up),
-            ]
-        if suffix == "mlp.experts.down_proj":
-            return [(tn(gguf.MODEL_TENSOR.FFN_DOWN_EXP, bid), data_torch)]
+            yield from super().modify_tensors(gate, tn(gguf.MODEL_TENSOR.FFN_GATE_EXP, bid), bid)
+            yield from super().modify_tensors(up,   tn(gguf.MODEL_TENSOR.FFN_UP_EXP,   bid), bid)
+            return
 
-        raise ValueError(f"Unsupported HY_V4 tensor {name!r} (suffix {suffix!r})")
+        # add .weight suffixes
+        if name.endswith("mlp.experts.down_proj") or name.endswith(".self_attn.learnable_sink_param"):
+            name += ".weight"
+
+        if re.search(r"\.hc_head\.hc_head_(?:fn|base|scale)$", name):
+            name += ".weight"
+
+        if re.search(r"\.hc_(?:attn|mlp)_layer\.hc_pre\.hc_(?:fn|base|scale)$", name):
+            name += ".weight"
+
+        yield from super().modify_tensors(data_torch, name, bid)

conversion/qwen.py

@@ -379,6 +379,13 @@ class Qwen3NextModel(_QwenMtpMixin, Qwen2MoeModel):
         self.gguf_writer.add_ssm_group_count(self.hparams["linear_num_key_heads"])
         self.gguf_writer.add_ssm_time_step_rank(self.hparams["linear_num_value_heads"])
         self.gguf_writer.add_ssm_inner_size(self.hparams["linear_value_head_dim"] * self.hparams["linear_num_value_heads"])
+        if (layer_types := self.hparams.get("layer_types")) is not None:
+            n_layer = self.hparams["num_hidden_layers"]
+            if len(layer_types) != n_layer:
+                raise ValueError(f"layer_types has {len(layer_types)} entries, expected num_hidden_layers ({n_layer})")
+            recurrent = [t == "linear_attention" for t in layer_types]
+            recurrent += [False] * (self.block_count - n_layer)
+            self.gguf_writer.add_recurrent_layers(recurrent)
         self.gguf_writer.add_full_attention_interval(self.hparams.get("full_attention_interval", 4))
         if (rope_dim := self.hparams.get("head_dim")) is None:
             rope_dim = self.hparams["hidden_size"] // self.hparams["num_attention_heads"]

examples/test-cmake/test-cmake.cpp

@@ -2,8 +2,9 @@
 #include <cstdio>
 
 int main(void) {
-    printf("[test-cmake] version: %s, build: %d (%s)\n",
+    printf("[test-cmake] llama.cpp version: %s, build: %d (%s)\n",
            llama_version(), LLAMA_BUILD_NUMBER, LLAMA_BUILD_COMMIT);
+    printf("[test-cmake] ggml version: %s, commit: %s\n", ggml_version(), ggml_commit());
     printf("[test-cmake] Initializing backend...\n");
     llama_backend_init();
     printf("[test-cmake] Backend initialized.\n");

flake.nix

@@ -128,7 +128,7 @@
           }:
           {
             # For standardised reproducible formatting with `nix fmt`
-            formatter = pkgs.nixfmt-rfc-style;
+            formatter = pkgs.nixfmt;
 
             # Unlike `.#packages`, legacyPackages may contain values of
             # arbitrary types (including nested attrsets) and may even throw
@@ -156,7 +156,7 @@
                 windows = config.legacyPackages.llamaPackagesWindows.llama-cpp;
                 python-scripts = config.legacyPackages.llamaPackages.python-scripts;
               }
-              // lib.optionalAttrs pkgs.stdenv.isLinux {
+              // lib.optionalAttrs pkgs.stdenv.hostPlatform.isLinux {
                 cuda = config.legacyPackages.llamaPackagesCuda.llama-cpp;
 
                 mpi-cpu = config.packages.default.override { useMpi = true; };

ggml/src/ggml-backend.cpp

@@ -849,7 +849,7 @@ static void ggml_backend_sched_split_inputs_grow(struct ggml_backend_sched_split
     int new_cap = GGML_SCHED_MAX_SPLIT_INPUTS;
     if (split->inputs_capacity > 0) {
         new_cap = 2*split->inputs_capacity;
-        GGML_LOG_WARN("%s: increasing split inputs capacity from %d to %d\n", __func__, split->inputs_capacity, new_cap);
+        GGML_LOG_DEBUG("%s: increasing split inputs capacity from %d to %d\n", __func__, split->inputs_capacity, new_cap);
     }
     auto * pnew = (struct ggml_tensor **) realloc((void *) split->inputs, new_cap * sizeof(struct ggml_tensor *));
     if (pnew == NULL) {
@@ -864,7 +864,7 @@ static void ggml_backend_sched_graph_inputs_grow(ggml_backend_sched_t sched) {
     int new_cap = GGML_SCHED_MAX_SPLIT_INPUTS;
     if (sched->graph_inputs_capacity > 0) {
         new_cap = 2*sched->graph_inputs_capacity;
-        GGML_LOG_WARN("%s: increasing graph inputs capacity from %d to %d\n", __func__, sched->graph_inputs_capacity, new_cap);
+        GGML_LOG_DEBUG("%s: increasing graph inputs capacity from %d to %d\n", __func__, sched->graph_inputs_capacity, new_cap);
     }
     auto * pnew = (struct ggml_tensor **) realloc((void *) sched->graph_inputs, new_cap * sizeof(struct ggml_tensor *));
     if (pnew == NULL) {
@@ -1338,17 +1338,6 @@ void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct ggml_cgra
                             break;
                         }
                     }
-                    // check if the split has too many inputs
-                    // FIXME: count the number of inputs instead of only checking when full
-                    if (split->n_inputs >= split->inputs_capacity) {
-                        const size_t id = hash_id(src);
-                        int src_backend_id = sched->hv_tensor_backend_ids[id];
-                        bool supported = ggml_backend_sched_buffer_supported(sched, src, cur_backend_id);
-                        if (src_backend_id != cur_backend_id && tensor_id_copy(id, cur_backend_id, 0) == NULL && !supported) {
-                            need_new_split = true;
-                            break;
-                        }
-                    }
                 }
             }
 

ggml/src/ggml-cuda/common.cuh

@@ -69,6 +69,8 @@
 #define GGML_CUDA_CC_GCN4       (GGML_CUDA_CC_OFFSET_AMD + 0x803)  // Tonga, Fiji, Polaris, minimum for fast fp16
 #define GGML_CUDA_CC_VEGA       (GGML_CUDA_CC_OFFSET_AMD + 0x900)  // Vega56/64, minimum for fp16 dual issue
 #define GGML_CUDA_CC_VEGA20     (GGML_CUDA_CC_OFFSET_AMD + 0x906)  // MI50/Radeon VII, minimum for dp4a
+#define GGML_CUDA_CC_GFX909     (GGML_CUDA_CC_OFFSET_AMD + 0x909)  // GCN APU
+#define GGML_CUDA_CC_GFX90C     (GGML_CUDA_CC_OFFSET_AMD + 0x90c)  // GCN APU
 #define GGML_CUDA_CC_CDNA1      (GGML_CUDA_CC_OFFSET_AMD + 0x908)  // MI100, minimum for MFMA, acc registers
 #define GGML_CUDA_CC_CDNA2      (GGML_CUDA_CC_OFFSET_AMD + 0x90a)  // MI210 (gfx90a), minimum acc register renaming
 #define GGML_CUDA_CC_CDNA3      (GGML_CUDA_CC_OFFSET_AMD + 0x942)  // MI300
@@ -89,12 +91,13 @@
 #define GGML_CUDA_CC_IS_RDNA3_5(cc) (cc >= GGML_CUDA_CC_RDNA3_5 && cc < GGML_CUDA_CC_RDNA4)
 #define GGML_CUDA_CC_IS_RDNA3(cc)   (GGML_CUDA_CC_IS_RDNA3_0(cc) || GGML_CUDA_CC_IS_RDNA3_5(cc))
 #define GGML_CUDA_CC_IS_RDNA4(cc)   (cc >= GGML_CUDA_CC_RDNA4)
-#define GGML_CUDA_CC_IS_GCN(cc)     (cc > GGML_CUDA_CC_OFFSET_AMD && cc < GGML_CUDA_CC_CDNA1)
-#define GGML_CUDA_CC_IS_CDNA(cc)    (cc >= GGML_CUDA_CC_CDNA1 && cc < GGML_CUDA_CC_RDNA1)
-#define GGML_CUDA_CC_IS_CDNA1(cc)   (cc >= GGML_CUDA_CC_CDNA1 && cc < GGML_CUDA_CC_CDNA2)
-#define GGML_CUDA_CC_IS_CDNA2(cc)   (cc >= GGML_CUDA_CC_CDNA2 && cc < GGML_CUDA_CC_CDNA3)
-#define GGML_CUDA_CC_IS_CDNA3(cc)   (cc >= GGML_CUDA_CC_CDNA3 && cc < GGML_CUDA_CC_CDNA4)
-#define GGML_CUDA_CC_IS_CDNA4(cc)   (cc >= GGML_CUDA_CC_CDNA4 && cc < GGML_CUDA_CC_RDNA1)
+#define GGML_CUDA_CC_IS_GCN_APU(cc) ((cc) == GGML_CUDA_CC_GFX909 || (cc) == GGML_CUDA_CC_GFX90C)
+#define GGML_CUDA_CC_IS_GCN(cc)     ((cc > GGML_CUDA_CC_OFFSET_AMD && cc < GGML_CUDA_CC_CDNA1) || GGML_CUDA_CC_IS_GCN_APU(cc))
+#define GGML_CUDA_CC_IS_CDNA(cc)    (!GGML_CUDA_CC_IS_GCN_APU(cc) && cc >= GGML_CUDA_CC_CDNA1 && cc < GGML_CUDA_CC_RDNA1)
+#define GGML_CUDA_CC_IS_CDNA1(cc)   (GGML_CUDA_CC_IS_CDNA(cc) && cc >= GGML_CUDA_CC_CDNA1 && cc < GGML_CUDA_CC_CDNA2)
+#define GGML_CUDA_CC_IS_CDNA2(cc)   (GGML_CUDA_CC_IS_CDNA(cc) && cc >= GGML_CUDA_CC_CDNA2 && cc < GGML_CUDA_CC_CDNA3)
+#define GGML_CUDA_CC_IS_CDNA3(cc)   (GGML_CUDA_CC_IS_CDNA(cc) && cc >= GGML_CUDA_CC_CDNA3 && cc < GGML_CUDA_CC_CDNA4)
+#define GGML_CUDA_CC_IS_CDNA4(cc)   (GGML_CUDA_CC_IS_CDNA(cc) && cc >= GGML_CUDA_CC_CDNA4 && cc < GGML_CUDA_CC_RDNA1)
 
 // Moore Threads
 #define MUSART_HMASK 40300 // MUSA rc4.3, min. ver. for half2 -> uint mask comparisons
@@ -976,6 +979,7 @@ template<>
 struct ggml_cuda_type_traits<GGML_TYPE_F16> {
     static constexpr int qk = 1;
     static constexpr int qr = 1;
+    static constexpr int bs = sizeof(ggml_half);
 };
 
 template<>
@@ -983,6 +987,7 @@ struct ggml_cuda_type_traits<GGML_TYPE_Q1_0> {
     static constexpr int qk = QK1_0;
     static constexpr int qr = QR1_0;
     static constexpr int qi = QI1_0;
+    static constexpr int bs = sizeof(block_q1_0);
 };
 
 template<>
@@ -990,6 +995,7 @@ struct ggml_cuda_type_traits<GGML_TYPE_Q2_0> {
     static constexpr int qk = QK2_0;
     static constexpr int qr = QR2_0;
     static constexpr int qi = QI2_0;
+    static constexpr int bs = sizeof(block_q2_0);
 };
 
 template<>
@@ -997,6 +1003,7 @@ struct ggml_cuda_type_traits<GGML_TYPE_Q4_0> {
     static constexpr int qk = QK4_0;
     static constexpr int qr = QR4_0;
     static constexpr int qi = QI4_0;
+    static constexpr int bs = sizeof(block_q4_0);
 };
 
 template<>
@@ -1004,6 +1011,7 @@ struct ggml_cuda_type_traits<GGML_TYPE_Q4_1> {
     static constexpr int qk = QK4_1;
     static constexpr int qr = QR4_1;
     static constexpr int qi = QI4_1;
+    static constexpr int bs = sizeof(block_q4_1);
 };
 
 template<>
@@ -1011,6 +1019,7 @@ struct ggml_cuda_type_traits<GGML_TYPE_Q5_0> {
     static constexpr int qk = QK5_0;
     static constexpr int qr = QR5_0;
     static constexpr int qi = QI5_0;
+    static constexpr int bs = sizeof(block_q5_0);
 };
 
 template<>
@@ -1018,6 +1027,7 @@ struct ggml_cuda_type_traits<GGML_TYPE_Q5_1> {
     static constexpr int qk = QK5_1;
     static constexpr int qr = QR5_1;
     static constexpr int qi = QI5_1;
+    static constexpr int bs = sizeof(block_q5_1);
 };
 
 template<>
@@ -1025,6 +1035,7 @@ struct ggml_cuda_type_traits<GGML_TYPE_Q8_0> {
     static constexpr int qk = QK8_0;
     static constexpr int qr = QR8_0;
     static constexpr int qi = QI8_0;
+    static constexpr int bs = sizeof(block_q8_0);
 };
 
 template<>
@@ -1032,6 +1043,7 @@ struct ggml_cuda_type_traits<GGML_TYPE_MXFP4> {
     static constexpr int qk = QK_MXFP4;
     static constexpr int qr = QR_MXFP4;
     static constexpr int qi = QI_MXFP4;
+    static constexpr int bs = sizeof(block_mxfp4);
 };
 
 template<>
@@ -1039,6 +1051,7 @@ struct ggml_cuda_type_traits<GGML_TYPE_NVFP4> {
     static constexpr int qk = QK_NVFP4;
     static constexpr int qr = QR_NVFP4;
     static constexpr int qi = QI_NVFP4;
+    static constexpr int bs = sizeof(block_nvfp4);
 };
 
 template<>
@@ -1046,6 +1059,7 @@ struct ggml_cuda_type_traits<GGML_TYPE_Q2_K> {
     static constexpr int qk = QK_K;
     static constexpr int qr = QR2_K;
     static constexpr int qi = QI2_K;
+    static constexpr int bs = sizeof(block_q2_K);
 };
 
 template<>
@@ -1053,6 +1067,7 @@ struct ggml_cuda_type_traits<GGML_TYPE_Q3_K> {
     static constexpr int qk = QK_K;
     static constexpr int qr = QR3_K;
     static constexpr int qi = QI3_K;
+    static constexpr int bs = sizeof(block_q3_K);
 };
 
 template<>
@@ -1060,6 +1075,7 @@ struct ggml_cuda_type_traits<GGML_TYPE_Q4_K> {
     static constexpr int qk = QK_K;
     static constexpr int qr = QR4_K;
     static constexpr int qi = QI4_K;
+    static constexpr int bs = sizeof(block_q4_K);
 };
 
 template<>
@@ -1067,6 +1083,7 @@ struct ggml_cuda_type_traits<GGML_TYPE_Q5_K> {
     static constexpr int qk = QK_K;
     static constexpr int qr = QR5_K;
     static constexpr int qi = QI5_K;
+    static constexpr int bs = sizeof(block_q5_K);
 };
 
 template<>
@@ -1074,6 +1091,7 @@ struct ggml_cuda_type_traits<GGML_TYPE_Q6_K> {
     static constexpr int qk = QK_K;
     static constexpr int qr = QR6_K;
     static constexpr int qi = QI6_K;
+    static constexpr int bs = sizeof(block_q6_K);
 };
 
 template<>
@@ -1081,6 +1099,7 @@ struct ggml_cuda_type_traits<GGML_TYPE_IQ2_XXS> {
     static constexpr int qk = QK_K;
     static constexpr int qr = QR2_XXS;
     static constexpr int qi = QI2_XXS;
+    static constexpr int bs = sizeof(block_iq2_xxs);
 };
 
 template<>
@@ -1088,6 +1107,7 @@ struct ggml_cuda_type_traits<GGML_TYPE_IQ2_XS> {
     static constexpr int qk = QK_K;
     static constexpr int qr = QR2_XS;
     static constexpr int qi = QI2_XS;
+    static constexpr int bs = sizeof(block_iq2_xs);
 };
 
 template<>
@@ -1095,6 +1115,7 @@ struct ggml_cuda_type_traits<GGML_TYPE_IQ2_S> {
     static constexpr int qk = QK_K;
     static constexpr int qr = QR2_S;
     static constexpr int qi = QI2_S;
+    static constexpr int bs = sizeof(block_iq2_s);
 };
 
 template<>
@@ -1102,6 +1123,7 @@ struct ggml_cuda_type_traits<GGML_TYPE_IQ3_XXS> {
     static constexpr int qk = QK_K;
     static constexpr int qr = QR3_XXS;
     static constexpr int qi = QI3_XXS;
+    static constexpr int bs = sizeof(block_iq3_xxs);
 };
 
 template<>
@@ -1109,6 +1131,7 @@ struct ggml_cuda_type_traits<GGML_TYPE_IQ1_S> {
     static constexpr int qk = QK_K;
     static constexpr int qr = QR1_S;
     static constexpr int qi = QI1_S;
+    static constexpr int bs = sizeof(block_iq1_s);
 };
 
 template<>
@@ -1116,6 +1139,7 @@ struct ggml_cuda_type_traits<GGML_TYPE_IQ1_M> {
     static constexpr int qk = QK_K;
     static constexpr int qr = QR1_M;
     static constexpr int qi = QI1_M;
+    static constexpr int bs = sizeof(block_iq1_m);
 };
 
 template<>
@@ -1123,6 +1147,7 @@ struct ggml_cuda_type_traits<GGML_TYPE_IQ4_NL> {
     static constexpr int qk = QK4_NL;
     static constexpr int qr = QR4_NL;
     static constexpr int qi = QI4_NL;
+    static constexpr int bs = sizeof(block_iq4_nl);
 };
 
 template<>
@@ -1130,6 +1155,7 @@ struct ggml_cuda_type_traits<GGML_TYPE_IQ4_XS> {
     static constexpr int qk = QK_K;
     static constexpr int qr = QR4_XS;
     static constexpr int qi = QI4_XS;
+    static constexpr int bs = sizeof(block_iq4_xs);
 };
 
 template<>
@@ -1137,6 +1163,7 @@ struct ggml_cuda_type_traits<GGML_TYPE_IQ3_S> {
     static constexpr int qk = QK_K;
     static constexpr int qr = QR3_S;
     static constexpr int qi = QI3_S;
+    static constexpr int bs = sizeof(block_iq3_s);
 };
 
 //////////////////////

ggml/src/ggml-cuda/fattn-mma-f16.cuh

@@ -1545,77 +1545,77 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile(
         }
     }
 
-    if (np > 1 && threadIdx.y % np == 0) {
-        // Combine the meta data for parallel warps via shared memory.
-        // Warps with threadIdx.y % np != 0 must NOT return early.
-        // All threads must return simultaneously to avoid race conditions with work on the next tile.
-
+    if (np > 1) {
         constexpr int nmeta = np*cols_per_warp >= warp_size ? np*cols_per_warp/warp_size : 1;
 
+        float KQ_cmn;
+        float KQ_cms[nmeta];
+        float KQ_crs;
+
         const int jc_meta = threadIdx.y*cols_per_warp + (np*cols_per_warp < warp_size ? threadIdx.x % (np*cols_per_warp) : threadIdx.x);
         float2 * const meta_ptr = ((float2 *) tile_Q) + jc_meta*(tile_stride/2) + nbatch_combine/2;
-        float2 meta[nmeta];
-#pragma unroll
-        for (int imeta = 0; imeta < nmeta; ++imeta) {
-            meta[imeta] = meta_ptr[imeta * warp_size * tile_stride/2];
-        }
 
-        float KQ_cmn = meta[0].x; // KQ combine max new, max between all parallel warps.
+        if (threadIdx.y % np == 0) {
+            // Combine the meta data for parallel warps via shared memory.
+            float2 meta[nmeta];
 #pragma unroll
-        for (int imeta = 1; imeta < nmeta; ++imeta) {
-            KQ_cmn = fmaxf(KQ_cmn, meta[imeta].x);
-        }
-#pragma unroll
-        for (int offset = np*cols_per_warp/2; offset >= cols_per_warp; offset >>= 1) {
-            if (offset < warp_size) {
-                KQ_cmn = fmaxf(KQ_cmn, __shfl_xor_sync(0xFFFFFFFF, KQ_cmn, offset, warp_size));
+            for (int imeta = 0; imeta < nmeta; ++imeta) {
+                meta[imeta] = meta_ptr[imeta * warp_size * tile_stride/2];
             }
-        }
 
-        float KQ_cms[nmeta]; // KQ combine max scale per warp.
+            KQ_cmn = meta[0].x; // KQ combine max new, max between all parallel warps.
 #pragma unroll
-        for (int imeta = 0; imeta < nmeta; ++imeta) {
-            KQ_cms[imeta] = expf(meta[imeta].x - KQ_cmn);
-        }
+            for (int imeta = 1; imeta < nmeta; ++imeta) {
+                KQ_cmn = fmaxf(KQ_cmn, meta[imeta].x);
+            }
+#pragma unroll
+            for (int offset = np*cols_per_warp/2; offset >= cols_per_warp; offset >>= 1) {
+                if (offset < warp_size) {
+                    KQ_cmn = fmaxf(KQ_cmn, __shfl_xor_sync(0xFFFFFFFF, KQ_cmn, offset, warp_size));
+                }
+            }
 
-        float KQ_crs = KQ_cms[0]*meta[0].y; // KQ combine rowsum, scaled sum of all parallel warps.
 #pragma unroll
-        for (int imeta = 1; imeta < nmeta; ++imeta) {
-            KQ_crs += KQ_cms[imeta]*meta[imeta].y;
-        }
+            for (int imeta = 0; imeta < nmeta; ++imeta) {
+                KQ_cms[imeta] = expf(meta[imeta].x - KQ_cmn);
+            }
+
+            KQ_crs = KQ_cms[0]*meta[0].y; // KQ combine rowsum, scaled sum of all parallel warps.
 #pragma unroll
-        for (int offset = np*cols_per_warp/2; offset >= cols_per_warp; offset >>= 1) {
-            if (offset < warp_size) {
-                KQ_crs += __shfl_xor_sync(0xFFFFFFFF, KQ_crs, offset, warp_size);
+            for (int imeta = 1; imeta < nmeta; ++imeta) {
+                KQ_crs += KQ_cms[imeta]*meta[imeta].y;
+            }
+#pragma unroll
+            for (int offset = np*cols_per_warp/2; offset >= cols_per_warp; offset >>= 1) {
+                if (offset < warp_size) {
+                    KQ_crs += __shfl_xor_sync(0xFFFFFFFF, KQ_crs, offset, warp_size);
+                }
             }
         }
 
         __syncthreads();
 
-        // Write back combined meta data:
+        if (threadIdx.y % np == 0) {
+            // Write back combined meta data:
 #pragma unroll
-        for (int imeta = 0; imeta < nmeta; ++imeta) {
-            if (np*cols_per_warp >= warp_size || threadIdx.x < np*cols_per_warp) {
-                // Combined KQ max scale + rowsum.
-                meta_ptr[imeta * warp_size * tile_stride/2] = make_float2(KQ_cms[imeta], KQ_crs);
+            for (int imeta = 0; imeta < nmeta; ++imeta) {
+                if (np*cols_per_warp >= warp_size || threadIdx.x < np*cols_per_warp) {
+                    // Combined KQ max scale + rowsum.
+                    meta_ptr[imeta * warp_size * tile_stride/2] = make_float2(KQ_cms[imeta], KQ_crs);
+                }
+            }
+
+            // Combined KQ max + rowsum.
+            static_assert(cols_per_warp <= warp_size);
+            if (needs_fixup && (cols_per_warp == warp_size || threadIdx.x < cols_per_warp)) {
+                float2 * dstk_fixup_meta = dstk_fixup + blockIdx.x*ncols;
+                dstk_fixup_meta[(threadIdx.y/np)*cols_per_warp + threadIdx.x] = make_float2(KQ_cmn, KQ_crs);
+            }
+            if (is_fixup && (cols_per_warp == warp_size || threadIdx.x < cols_per_warp)) {
+                float2 * dstk_fixup_meta = dstk_fixup + (gridDim.x + blockIdx.x)*ncols;
+                dstk_fixup_meta[(threadIdx.y/np)*cols_per_warp + threadIdx.x] = make_float2(KQ_cmn, KQ_crs);
             }
         }
-
-        // Combined KQ max + rowsum.
-        static_assert(cols_per_warp <= warp_size);
-        if (needs_fixup && (cols_per_warp == warp_size || threadIdx.x < cols_per_warp)) {
-            float2 * dstk_fixup_meta = dstk_fixup + blockIdx.x*ncols;
-            dstk_fixup_meta[(threadIdx.y/np)*cols_per_warp + threadIdx.x] = make_float2(KQ_cmn, KQ_crs);
-        }
-        if (is_fixup && (cols_per_warp == warp_size || threadIdx.x < cols_per_warp)) {
-            float2 * dstk_fixup_meta = dstk_fixup + (gridDim.x + blockIdx.x)*ncols;
-            dstk_fixup_meta[(threadIdx.y/np)*cols_per_warp + threadIdx.x] = make_float2(KQ_cmn, KQ_crs);
-        }
-    } else if (np > 1) {
-        // Warps with threadIdx.y % np == 0 execute a __syncthreads() in the if branch.
-        // Therefore, all other warps also need to execute a __syncthreads().
-        // Otherwise the points at which warps synchronize with each other would become misaligned.
-        __syncthreads();
     }
 
 #pragma unroll

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

@@ -212,6 +212,7 @@ static int ggml_cuda_parse_id(char devName[]) {
     }
     archNum += archMajor * 0x100;
     archNum += archMinor;
+
     return archNum;
 }
 #endif // defined(GGML_USE_HIP)

ggml/src/ggml-cuda/mmq-config-ampere.cuh

@@ -1,4 +1,5 @@
 static constexpr __host__ __device__ ggml_cuda_mmq_config ggml_cuda_mmq_get_config_ampere(ggml_type type, int J, bool fallback) {
+    constexpr bool use_typical_moe_ncols = false;
     CASE(GGML_TYPE_Q1_0, 256, 1, 128,   8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true);
     CASE(GGML_TYPE_Q1_0, 256, 1, 128,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true);
     CASE(GGML_TYPE_Q1_0, 256, 1, 128,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true);
@@ -379,5 +380,5 @@ static constexpr __host__ __device__ ggml_cuda_mmq_config ggml_cuda_mmq_get_conf
     CASE(GGML_TYPE_NVFP4, 256, 1, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, true, false);
     CASE(GGML_TYPE_NVFP4, 256, 1, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, true, false);
 
-    return ggml_cuda_mmq_config(GGML_TYPE_COUNT, 256, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, 256, false, true);
+    return ggml_cuda_mmq_config(GGML_TYPE_COUNT, 256, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, 256, use_typical_moe_ncols, false, true);
 }

ggml/src/ggml-cuda/mmq-config-blackwell.cuh

@@ -1,4 +1,5 @@
 static constexpr __host__ __device__ ggml_cuda_mmq_config ggml_cuda_mmq_get_config_blackwell(ggml_type type, int J, bool fallback) {
+    constexpr bool use_typical_moe_ncols = false;
     CASE(GGML_TYPE_MXFP4, 256, 1, 128,   8, GGML_CUDA_MMQ_SRAM_LAYOUT_FP4, MMQ_ITER_K_FP4, true, true);
     CASE(GGML_TYPE_MXFP4, 256, 1, 128,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_FP4, MMQ_ITER_K_FP4, true, true);
     CASE(GGML_TYPE_MXFP4, 256, 1, 128,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_FP4, MMQ_ITER_K_FP4, true, true);

ggml/src/ggml-cuda/mmq-config-cdna.cuh

@@ -1,4 +1,5 @@
 static constexpr __host__ __device__ ggml_cuda_mmq_config ggml_cuda_mmq_get_config_cdna(ggml_type type, int J, bool fallback) {
+    constexpr bool use_typical_moe_ncols = false;
     CASE(GGML_TYPE_Q1_0, 512, 1, 128,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true);
     CASE(GGML_TYPE_Q1_0, 512, 1, 128,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true);
     CASE(GGML_TYPE_Q1_0, 512, 1, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true);
@@ -181,5 +182,5 @@ static constexpr __host__ __device__ ggml_cuda_mmq_config ggml_cuda_mmq_get_conf
     CASE(GGML_TYPE_NVFP4, 512, 1, 128,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, true, false);
     CASE(GGML_TYPE_NVFP4, 512, 1, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, true, false);
 
-    return ggml_cuda_mmq_config(GGML_TYPE_COUNT, 512, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, 256, false, true);
+    return ggml_cuda_mmq_config(GGML_TYPE_COUNT, 512, 1, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, 256, use_typical_moe_ncols, false, true);
 }

ggml/src/ggml-cuda/mmq-config-pascal-dp4a.cuh

@@ -1,4 +1,5 @@
 static constexpr __host__ __device__ ggml_cuda_mmq_config ggml_cuda_mmq_get_config_pascal_dp4a(ggml_type type, int J, bool fallback) {
+    constexpr bool use_typical_moe_ncols = false;
     CASE(GGML_TYPE_Q1_0, 256, 2, 64,   8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
     CASE(GGML_TYPE_Q1_0, 256, 2, 64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
     CASE(GGML_TYPE_Q1_0, 256, 2, 64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
@@ -269,5 +270,5 @@ static constexpr __host__ __device__ ggml_cuda_mmq_config ggml_cuda_mmq_get_conf
     CASE(GGML_TYPE_NVFP4, 256, 2, 64,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, false);
     CASE(GGML_TYPE_NVFP4, 256, 2, 64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, false);
 
-    return ggml_cuda_mmq_config(GGML_TYPE_COUNT, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, 256, false, true);
+    return ggml_cuda_mmq_config(GGML_TYPE_COUNT, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, 256, use_typical_moe_ncols, false, true);
 }

ggml/src/ggml-cuda/mmq-config-pascal-older.cuh

@@ -1,4 +1,5 @@
 static constexpr __host__ __device__ ggml_cuda_mmq_config ggml_cuda_mmq_get_config_pascal_older(ggml_type type, int J, bool fallback) {
+    constexpr bool use_typical_moe_ncols = false;
     CASE(GGML_TYPE_Q1_0, 256, 2, 64,   8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
     CASE(GGML_TYPE_Q1_0, 256, 2, 64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
     CASE(GGML_TYPE_Q1_0, 256, 2, 64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
@@ -269,5 +270,5 @@ static constexpr __host__ __device__ ggml_cuda_mmq_config ggml_cuda_mmq_get_conf
     CASE(GGML_TYPE_NVFP4, 256, 2, 64,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, false);
     CASE(GGML_TYPE_NVFP4, 256, 2, 64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, false);
 
-    return ggml_cuda_mmq_config(GGML_TYPE_COUNT, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, 256, false, true);
+    return ggml_cuda_mmq_config(GGML_TYPE_COUNT, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, 256, use_typical_moe_ncols, false, true);
 }

ggml/src/ggml-cuda/mmq-config-rdna2.cuh

@@ -1,4 +1,5 @@
 static constexpr __host__ __device__ ggml_cuda_mmq_config ggml_cuda_mmq_get_config_rdna2(ggml_type type, int J, bool fallback) {
+    constexpr bool use_typical_moe_ncols = false;
     CASE(GGML_TYPE_Q1_0, 256, 2, 128,   8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
     CASE(GGML_TYPE_Q1_0, 256, 2, 128,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
     CASE(GGML_TYPE_Q1_0, 256, 2, 128,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
@@ -269,5 +270,5 @@ static constexpr __host__ __device__ ggml_cuda_mmq_config ggml_cuda_mmq_get_conf
     CASE(GGML_TYPE_NVFP4, 256, 2, 128,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, false);
     CASE(GGML_TYPE_NVFP4, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, false);
 
-    return ggml_cuda_mmq_config(GGML_TYPE_COUNT, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, 256, false, true);
+    return ggml_cuda_mmq_config(GGML_TYPE_COUNT, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, 256, use_typical_moe_ncols, false, true);
 }

ggml/src/ggml-cuda/mmq-config-rdna3-5.cuh

@@ -1,4 +1,5 @@
 static constexpr __host__ __device__ ggml_cuda_mmq_config ggml_cuda_mmq_get_config_rdna3_5(ggml_type type, int J, bool fallback) {
+    constexpr bool use_typical_moe_ncols = false;
     CASE(GGML_TYPE_Q1_0, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
     CASE(GGML_TYPE_Q1_0, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
     CASE(GGML_TYPE_Q1_0, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
@@ -286,5 +287,5 @@ static constexpr __host__ __device__ ggml_cuda_mmq_config ggml_cuda_mmq_get_conf
     CASE(GGML_TYPE_NVFP4, 256, 2, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, false);
     CASE(GGML_TYPE_NVFP4, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, false);
 
-    return ggml_cuda_mmq_config(GGML_TYPE_COUNT, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, 256, false, true);
+    return ggml_cuda_mmq_config(GGML_TYPE_COUNT, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, 256, use_typical_moe_ncols, false, true);
 }

ggml/src/ggml-cuda/mmq-config-rdna3.cuh

@@ -1,4 +1,5 @@
 static constexpr __host__ __device__ ggml_cuda_mmq_config ggml_cuda_mmq_get_config_rdna3(ggml_type type, int J, bool fallback) {
+    constexpr bool use_typical_moe_ncols = true;
     CASE(GGML_TYPE_Q1_0, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
     CASE(GGML_TYPE_Q1_0, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
     CASE(GGML_TYPE_Q1_0, 128, 2,  64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
@@ -270,5 +271,5 @@ static constexpr __host__ __device__ ggml_cuda_mmq_config ggml_cuda_mmq_get_conf
     CASE(GGML_TYPE_NVFP4, 256, 2, 128,  96, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, false);
     CASE(GGML_TYPE_NVFP4, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, false);
 
-    return ggml_cuda_mmq_config(GGML_TYPE_COUNT, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, 256, false, true);
+    return ggml_cuda_mmq_config(GGML_TYPE_COUNT, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, 256, use_typical_moe_ncols, false, true);
 }

ggml/src/ggml-cuda/mmq-config-rdna4.cuh

@@ -1,4 +1,5 @@
 static constexpr __host__ __device__ ggml_cuda_mmq_config ggml_cuda_mmq_get_config_rdna4(ggml_type type, int J, bool fallback) {
+    constexpr bool use_typical_moe_ncols = true;
     CASE(GGML_TYPE_Q1_0, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
     CASE(GGML_TYPE_Q1_0, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
     CASE(GGML_TYPE_Q1_0, 128, 2,  64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
@@ -286,5 +287,5 @@ static constexpr __host__ __device__ ggml_cuda_mmq_config ggml_cuda_mmq_get_conf
     CASE(GGML_TYPE_NVFP4, 256, 2, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, false);
     CASE(GGML_TYPE_NVFP4, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, false);
 
-    return ggml_cuda_mmq_config(GGML_TYPE_COUNT, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, 256, false, true);
+    return ggml_cuda_mmq_config(GGML_TYPE_COUNT, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, 256, use_typical_moe_ncols, false, true);
 }

ggml/src/ggml-cuda/mmq.cu

@@ -375,10 +375,10 @@ bool ggml_cuda_should_use_mmq(enum ggml_type type, int cc, int64_t ne11, int64_t
         return true;
     }
 
-    // gfx900 (Vega 10) lacks native dp4a, loses to dequant + hipBLAS
+    // gfx900 (Vega 10), gfx909, and gfx90c lack native dp4a, losing to dequant + hipBLAS
     // for dense matrices; keep MMQ only for MoE, where the
     // hipBLAS path is much slower.
-    if (cc == GGML_CUDA_CC_VEGA) {
+    if (cc == GGML_CUDA_CC_VEGA || GGML_CUDA_CC_IS_GCN_APU(cc)) {
         return n_experts > 0;
     }
 

ggml/src/ggml-cuda/mmq.cuh

@@ -170,12 +170,13 @@ struct ggml_cuda_mmq_config {
     int                       J;           // SRAM tile width in src1->ne[1]/dst->ne[1] direction.
     ggml_cuda_mmq_sram_layout sram_layout; // SRAM tile length in src0->ne[0]/src1->ne[0] direction (physical 32 bit elements).
     int                       K_vram;      // VRAM tile length in src0->ne[0]/src1->ne[0] direction (logical elements).
+    bool                      use_typical_moe_ncols;
     bool                      stream_k;    // Whether or not to use stream-k decomposition.
     bool                      fallback;    // Whether a fallback for out-of-bounds check in src0->ne[1] direction is needed.
 
     constexpr __host__ __device__ ggml_cuda_mmq_config(
-            ggml_type type, int nthreads, int occupancy, int I, int J, ggml_cuda_mmq_sram_layout sram_layout, int K_vram, bool stream_k, bool fallback) :
-        type(type), nthreads(nthreads), occupancy(occupancy), I(I), J(J), sram_layout(sram_layout), K_vram(K_vram), stream_k(stream_k), fallback(fallback) {}
+            ggml_type type, int nthreads, int occupancy, int I, int J, ggml_cuda_mmq_sram_layout sram_layout, int K_vram, bool use_typical_moe_ncols, bool stream_k, bool fallback) :
+        type(type), nthreads(nthreads), occupancy(occupancy), I(I), J(J), sram_layout(sram_layout), K_vram(K_vram), use_typical_moe_ncols(use_typical_moe_ncols), stream_k(stream_k), fallback(fallback) {}
 
     constexpr __device__ int rows_per_warp() const {
 #if defined(AMD_MFMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
@@ -210,7 +211,7 @@ struct ggml_cuda_mmq_config {
         static_assert((I_)        %  32 == 0,                             "bad I");                                                       \
         static_assert((J_)        %   8 == 0,                             "bad J");                                                       \
         static_assert((K_vram_)   % 256 == 0,                             "bad K_vram");                                                  \
-        return ggml_cuda_mmq_config((type_), (nthreads_), (occupancy_), (I_), (J_), (sram_layout_), (K_vram_), (stream_k_), (fallback_)); \
+        return ggml_cuda_mmq_config((type_), (nthreads_), (occupancy_), (I_), (J_), (sram_layout_), (K_vram_), use_typical_moe_ncols, (stream_k_), (fallback_)); \
     }                                                                                                                                     \
 
 #include "mmq-config-pascal-older.cuh"
@@ -1473,6 +1474,20 @@ void mul_mat_q_switch_J(ggml_backend_cuda_context & ctx, const mmq_args & args,
     const int    cc    = ggml_cuda_info().devices[id].cc;
     const size_t smpbo = ggml_cuda_info().devices[id].smpbo;
 
+    int64_t ncols_picker = args.ncols_max;
+    if (args.expert_bounds != nullptr && args.nchannels_x > 0) {
+        const int J_max = ggml_cuda_mmq_get_J_max(type, fallback, cc, 128);
+        const ggml_cuda_mmq_config config_max = ggml_cuda_mmq_get_config(type, J_max, fallback, cc);
+        if (config_max.use_typical_moe_ncols) {
+            // Use the typical expert width only for tile selection.
+            // The launch grid still uses args.ncols_max.
+            const int64_t ncols_typical = (args.ncols_dst + args.nchannels_x - 1) / args.nchannels_x;
+            if (ncols_typical >= 1 && ncols_typical < J_max && ncols_typical < ncols_picker) {
+                ncols_picker = ncols_typical;
+            }
+        }
+    }
+
     int J_best        = 0;
     int ntiles_J_best = INT_MAX;
 
@@ -1486,7 +1501,7 @@ void mul_mat_q_switch_J(ggml_backend_cuda_context & ctx, const mmq_args & args,
             continue;
         }
 
-        const int ntiles_x = (args.ncols_max + config.J - 1) / config.J;
+        const int ntiles_x = (ncols_picker + config.J - 1) / config.J;
 
         if (ntiles_x < ntiles_J_best) {
             J_best = J;

ggml/src/ggml-cuda/mmvq.cu

@@ -6,6 +6,35 @@
 #include <cstdint>
 #include <type_traits>
 
+// only enabled on DGX Spark, where it is a gain on every type below. On the higher-bandwidth parts the kernel
+// has little exposed latency left to hide and the extra requests cost more than they save.
+// For perf data, see https://github.com/ggml-org/llama.cpp/pull/26705#issuecomment-5569335031
+#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ == GGML_CUDA_CC_DGX_SPARK
+// returns true only for those quants that benefit from prefetch and false otherwise
+static constexpr __host__ __device__ bool mmvq_should_prefetch(ggml_type type) {
+    switch (type) {
+        case GGML_TYPE_Q4_0:
+        case GGML_TYPE_Q5_0:
+        case GGML_TYPE_Q8_0:
+        case GGML_TYPE_MXFP4:
+        case GGML_TYPE_Q3_K:
+        case GGML_TYPE_Q4_K:
+        case GGML_TYPE_Q5_K:
+        case GGML_TYPE_Q6_K:
+        case GGML_TYPE_IQ1_M:
+        case GGML_TYPE_IQ4_NL:
+        case GGML_TYPE_IQ4_XS:
+            return true;
+        default:
+            return false;
+    }
+}
+
+static __device__ __forceinline__ void mmvq_prefetch_l2(const void * p) {
+    asm volatile("prefetch.global.L2 [%0];" :: "l"(p));
+}
+#endif
+
 typedef float (*vec_dot_q_cuda_t)(const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & kbx, const int & iqs);
 
 static constexpr __device__ vec_dot_q_cuda_t get_vec_dot_q_cuda(ggml_type type) {
@@ -298,9 +327,6 @@ bool ggml_cuda_should_use_mmvq(enum ggml_type type, int cc, int64_t ne11) {
                 return ne11 <= 4;
             case GGML_TYPE_Q3_K:
                 return ne11 <= 6;
-            case GGML_TYPE_Q4_K:
-            case GGML_TYPE_Q5_K:
-                return ne11 <= 7;
             default:
                 return ne11 <= MMVQ_MAX_BATCH_SIZE;
         }
@@ -310,8 +336,9 @@ bool ggml_cuda_should_use_mmvq(enum ggml_type type, int cc, int64_t ne11) {
             case GGML_TYPE_Q2_K:
             case GGML_TYPE_Q3_K:
             case GGML_TYPE_Q4_K:
-            case GGML_TYPE_Q5_K:
                 return ne11 <= 5;
+            case GGML_TYPE_Q5_K:
+                return ne11 <= 6;
             case GGML_TYPE_Q6_K:
                 return ne11 <= 7;
             default:
@@ -675,6 +702,26 @@ static __global__ void mul_mat_vec_q(
         // x block quant index when casting the quants to int
         const int kqs = vdr * (tid % (qi/vdr));
 
+#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ == GGML_CUDA_CC_DGX_SPARK
+        // start the next iterations' weight loads early
+        if constexpr (mmvq_should_prefetch(type)) {
+            constexpr int pf_dist = 2; // loop iterations, not blocks
+            const int kbx_pf = kbx + pf_dist*blocks_per_iter;
+            if (kbx_pf < blocks_per_row_x) {
+#pragma unroll
+                for (int i = 0; i < rows_per_cuda_block; ++i) {
+                    const size_t off = (size_t)(kbx_offset + i*stride_row_x + kbx_pf) * ggml_cuda_type_traits<type>::bs;
+                    mmvq_prefetch_l2((const char *) vx + off);
+                    if constexpr (has_fusion) {
+                        if (use_gate) {
+                            mmvq_prefetch_l2((const char *) vgate + off);
+                        }
+                    }
+                }
+            }
+        }
+#endif
+
 #pragma unroll
         for (int j = 0; j < ncols_dst; ++j) {
 #pragma unroll

ggml/src/ggml-cuda/vecdotq.cuh

@@ -936,16 +936,20 @@ static __device__ __forceinline__ float vec_dot_q4_K_q8_1(
     v[0] = q4[0];
     v[1] = q4[4];
 
+    // branchless so nvcc can hoist this out of the ncols_dst loop
     const uint16_t * scales = (const uint16_t *)bq4_K->scales;
+    const int j  = bq8_offset/2;
+    const int jm = j & 1;
+
+    const uint32_t s0 = scales[jm + 0];
+    const uint32_t s2 = scales[jm + 2];
+    const uint32_t s4 = scales[jm + 4];
+
+    const uint32_t hi = (uint32_t) -(int32_t) (j >= 2);
+
     uint16_t aux[2];
-    const int j = bq8_offset/2;
-    if (j < 2) {
-        aux[0] = scales[j+0] & 0x3f3f;
-        aux[1] = scales[j+2] & 0x3f3f;
-    } else {
-        aux[0] = ((scales[j+2] >> 0) & 0x0f0f) | ((scales[j-2] & 0xc0c0) >> 2);
-        aux[1] = ((scales[j+2] >> 4) & 0x0f0f) | ((scales[j-0] & 0xc0c0) >> 2);
-    }
+    aux[0] = (uint16_t) (((s0 & 0x3f3f) & ~hi) | ((((s4 >> 0) & 0x0f0f) | ((s0 & 0xc0c0) >> 2)) & hi));
+    aux[1] = (uint16_t) (((s2 & 0x3f3f) & ~hi) | ((((s4 >> 4) & 0x0f0f) | ((s2 & 0xc0c0) >> 2)) & hi));
     const uint8_t * sc = (const uint8_t *)aux;
     const uint8_t * m  = sc + 2;
 
@@ -981,16 +985,21 @@ static __device__ __forceinline__ float vec_dot_q5_K_q8_1(
     vh[0] = qh[0] >> bq8_offset;
     vh[1] = qh[4] >> bq8_offset;
 
+    // same as q4_K
     const uint16_t * scales = (const uint16_t *)bq5_K->scales;
+    const int j  = bq8_offset/2;
+    const int jm = j & 1;
+
+    const uint32_t s0 = scales[jm + 0];
+    const uint32_t s2 = scales[jm + 2];
+    const uint32_t s4 = scales[jm + 4];
+
+    const uint32_t hi = (uint32_t) -(int32_t) (j >= 2);
+
     uint16_t aux[2];
-    const int j = bq8_offset/2;
-    if (j < 2) {
-        aux[0] = scales[j+0] & 0x3f3f;
-        aux[1] = scales[j+2] & 0x3f3f;
-    } else {
-        aux[0] = ((scales[j+2] >> 0) & 0x0f0f) | ((scales[j-2] & 0xc0c0) >> 2);
-        aux[1] = ((scales[j+2] >> 4) & 0x0f0f) | ((scales[j-0] & 0xc0c0) >> 2);
-    }
+    aux[0] = (uint16_t) (((s0 & 0x3f3f) & ~hi) | ((((s4 >> 0) & 0x0f0f) | ((s0 & 0xc0c0) >> 2)) & hi));
+    aux[1] = (uint16_t) (((s2 & 0x3f3f) & ~hi) | ((((s4 >> 4) & 0x0f0f) | ((s2 & 0xc0c0) >> 2)) & hi));
+
     const uint8_t * sc = (const uint8_t *)aux;
     const uint8_t * m  = sc + 2;
 

ggml/src/ggml-cuda/vendors/hip.h

@@ -176,9 +176,9 @@
 
 #define __CUDA_ARCH__ 1300
 
-#if defined(__gfx900__) || defined(__gfx906__)
+#if defined(__gfx900__) || defined(__gfx906__) || defined(__gfx909__) || defined(__gfx90c__)
 #define GCN5
-#endif // defined(__gfx900__) || defined(__gfx906__)
+#endif // defined(__gfx900__) || defined(__gfx906__) || defined(__gfx909__) || defined(__gfx90c__)
 
 #if defined(__gfx803__)
 #define GCN4

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

@@ -1091,6 +1091,10 @@ static ggml_openvino_op_support is_op_supported_case(const ggml_tensor * op) {
         if (op->ne[3] != 1) {
             return {false, "GET_ROWS/SET_ROWS with ne[3] != 1 (ne[3]=" + std::to_string(op->ne[3]) + ") is not supported"};
         }
+        if (op->op == GGML_OP_GET_ROWS && ggml_is_quantized(op->src[0]->type) &&
+            op->src[0]->view_src != nullptr && op->src[0]->view_offs != 0) {
+            return {false, "GET_ROWS with a nonzero quantized src0 view offset is not supported"};
+        }
         if (op->op == GGML_OP_GET_ROWS && ggml_openvino_get_device_name() == "GPU" &&
             op->src[0]->type == GGML_TYPE_BF16) {
             return {false, "GET_ROWS with BF16 src0 is not supported on GPU"};

ggml/src/ggml-sycl/ggml-sycl.cpp

@@ -4858,6 +4858,78 @@ static bool ggml_sycl_mul_mat_glu_mmvq_fused(ggml_backend_sycl_context & ctx, gg
                                                /*stride_col_dst=*/(int) glu->ne[0], stream);
 }
 
+// Batch the run of consecutive L2_NORM siblings starting at node_idx into one launch.
+// Returns the number of extra graph nodes consumed, or 0 if the run is shorter than two
+// (the caller then runs the norm through the per-tensor kernel).
+static int ggml_sycl_l2_norm_batch_fused(ggml_backend_sycl_context & ctx, ggml_cgraph * cgraph, int node_idx) {
+    const ggml_tensor * node = cgraph->nodes[node_idx];
+    if (ggml_sycl_info().device_count != 1 || node->type != GGML_TYPE_F32 ||
+        node->src[0]->type != GGML_TYPE_F32 || node->src[0]->ne[0] >= 1024) {
+        return 0;
+    }
+
+    ggml_tensor * batch[GGML_SYCL_L2_BATCH_MAX];
+    int           count = 0;
+    int           last  = node_idx;
+    float         eps0;
+    memcpy(&eps0, node->op_params, sizeof(float));
+
+    // Conservative aliasing test: the batched norms run concurrently in one kernel,
+    // so none may read what another writes, and none may write where another writes.
+    auto overlaps = [](const ggml_tensor * a, const ggml_tensor * b) {
+        const char * ab = (const char *) a->data;
+        const char * bb = (const char *) b->data;
+        return ab < bb + ggml_nbytes(b) && bb < ab + ggml_nbytes(a);
+    };
+
+    for (int j = node_idx; j < cgraph->n_nodes && count < GGML_SYCL_L2_BATCH_MAX; ++j) {
+        ggml_tensor * nj = cgraph->nodes[j];
+        if (ggml_is_empty(nj) || nj->op == GGML_OP_RESHAPE || nj->op == GGML_OP_TRANSPOSE ||
+            nj->op == GGML_OP_VIEW || nj->op == GGML_OP_PERMUTE || nj->op == GGML_OP_NONE ||
+            (nj->flags & GGML_TENSOR_FLAG_COMPUTE) == 0) {
+            continue;  // not a launch; cannot break a run of adjacent norms
+        }
+        if (nj->op != GGML_OP_L2_NORM || nj->type != GGML_TYPE_F32 ||
+            nj->src[0]->type != GGML_TYPE_F32 || !ggml_are_same_shape(nj, node) ||
+            !ggml_are_same_shape(nj->src[0], node->src[0])) {
+            break;  // any other launch ends the run
+        }
+        bool same_nb = true;
+        for (int d = 0; d < GGML_MAX_DIMS; ++d) {
+            if (nj->nb[d] != node->nb[d] || nj->src[0]->nb[d] != node->src[0]->nb[d]) {
+                same_nb = false;
+                break;
+            }
+        }
+        if (!same_nb) {
+            break;  // one nb[] stride set is shared by the whole batch
+        }
+        float epsj;
+        memcpy(&epsj, nj->op_params, sizeof(float));
+        if (epsj != eps0) {
+            break;  // eps mismatch ends the run
+        }
+        bool indep = true;
+        for (int k = 0; k < count; ++k) {
+            if (overlaps(nj->src[0], batch[k]) || overlaps(nj, batch[k])) {
+                indep = false;
+                break;
+            }
+        }
+        if (!indep) {
+            break;  // an overlapping tensor would race inside one launch
+        }
+        batch[count++] = nj;
+        last           = j;
+    }
+    if (count < 2) {
+        return 0;  // a lone norm falls through to the per-tensor kernel
+    }
+    ggml_sycl_l2_norm_batch(ctx, batch, count);
+    return last - node_idx;
+}
+
+
 __dpct_inline__ static void k_copy_src1_to_contiguous(
     const char *__restrict__ src1_original, char *__restrict__ src1_contiguous,
     const mmid_row_mapping *__restrict__ row_mapping,
@@ -5908,6 +5980,17 @@ static void ggml_backend_sycl_graph_compute_impl(ggml_backend_sycl_context * syc
             continue;
         }
 
+        // Batch consecutive independent same-shape F32 L2_NORM siblings (the GDN q/k
+        // norms) into one launch; sources are strided views of the fused qkv buffer, so
+        // the scan skips the interleaved view nodes instead of breaking on them.
+        if (node->op == GGML_OP_L2_NORM) {
+            const int l2_batch_skip = ggml_sycl_l2_norm_batch_fused(*sycl_ctx, cgraph, i);
+            if (l2_batch_skip > 0) {
+                i += l2_batch_skip;
+                continue;
+            }
+        }
+
         if (node->op == GGML_OP_MUL_MAT && ggml_sycl_mul_mat_glu_mmvq_fused(*sycl_ctx, cgraph, i)) {
             i += 2;
             continue;

ggml/src/ggml-sycl/norm.cpp

@@ -543,6 +543,62 @@ static void l2_norm_f32_sycl(const float *   x,
     }
 }
 
+// Batched L2 norm: N independent same-shape F32 tensors in one launch; the tensor
+// index is folded into grid dim0 and each row's reduction is identical to the
+// single-tensor kernel, so the result is bit-exact.
+struct l2_batch_ptrs {
+    const float * src[GGML_SYCL_L2_BATCH_MAX];
+    float *       dst[GGML_SYCL_L2_BATCH_MAX];
+};
+
+// One stride set shared by the whole batch: the caller only groups tensors whose nb[]
+// all match, so per-tensor state stays two pointers.
+struct l2_batch_strides {
+    int     ne1, ne2;
+    int64_t ss0, ss1, ss2, ss3;
+    int64_t ds0, ds1, ds2, ds3;
+};
+
+template <int warp_size>
+static void l2_norm_f32_batch(l2_batch_ptrs p, l2_batch_strides st, const int ncols, const float eps,
+                              const sycl::nd_item<3> & item_ct1) {
+    const int t   = item_ct1.get_group(0);  // tensor index
+    const int r   = item_ct1.get_group(2);  // flattened row over ne1*ne2*ne3
+    const int tid = item_ct1.get_local_id(2);
+
+    const int i1 = r % st.ne1;
+    const int i2 = (r / st.ne1) % st.ne2;
+    const int i3 = r / (st.ne1 * st.ne2);
+
+    const float * x   = p.src[t] + i3 * st.ss3 + i2 * st.ss2 + i1 * st.ss1;
+    float *       dst = p.dst[t] + i3 * st.ds3 + i2 * st.ds2 + i1 * st.ds1;
+
+    float tmp = 0.0f;
+    for (int col = tid; col < ncols; col += warp_size) {
+        const float xi = x[col * st.ss0];
+        tmp += xi * xi;
+    }
+    tmp = block_reduce<block_reduce_method::SUM, warp_size>(tmp, (float *) nullptr, warp_size);
+    const float scale = sycl::rsqrt(sycl::fmax(tmp, eps * eps));
+    for (int col = tid; col < ncols; col += warp_size) {
+        dst[col * st.ds0] = scale * x[col * st.ss0];
+    }
+}
+
+template <int warp_size>
+static void l2_norm_f32_batch_sycl(l2_batch_ptrs p, l2_batch_strides st, const int n_tensors,
+                                   const int ncols, const int nrows_total, const float eps,
+                                   queue_ptr stream) {
+    const dpct::dim3 blocks_num(nrows_total, 1, n_tensors);
+    const dpct::dim3 block_dims(warp_size, 1, 1);
+    stream->submit([&](sycl::handler & cgh) {
+        cgh.parallel_for(sycl::nd_range<3>(blocks_num * block_dims, block_dims),
+            [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(warp_size)]] {
+                l2_norm_f32_batch<warp_size>(p, st, ncols, eps, item_ct1);
+            });
+    });
+}
+
 void ggml_sycl_op_norm(ggml_backend_sycl_context& ctx, ggml_tensor* dst) {
     const ggml_tensor * src0 = dst->src[0];
 
@@ -961,3 +1017,30 @@ void ggml_sycl_op_l2_norm(ggml_backend_sycl_context& ctx, ggml_tensor* dst) {
     l2_norm_f32_sycl<WARP_SIZE>(src0_d, dst_d, ne00, ne01, ne02, ne03,
             ss0, ss1, ss2, ss3, ds0, ds1, ds2, ds3, eps, stream, ctx.device);
 }
+
+// nodes[0..count) are independent, same-shape, same-eps, same-nb L2_NORM ops validated
+// by the caller; requires ncols < 1024 (the warp reduction path).
+void ggml_sycl_l2_norm_batch(ggml_backend_sycl_context & ctx, ggml_tensor ** nodes, int count) {
+    const ggml_tensor * s0 = nodes[0]->src[0];
+    const int ncols       = (int) s0->ne[0];
+    const int nrows_total = (int) ggml_nrows(s0);
+    float eps;
+    memcpy(&eps, nodes[0]->op_params, sizeof(float));
+    GGML_ASSERT(eps >= 0.0f);
+
+    l2_batch_ptrs p{};
+    for (int t = 0; t < count; ++t) {
+        p.src[t] = (const float *) nodes[t]->src[0]->data;
+        p.dst[t] = (float *) nodes[t]->data;
+    }
+
+    const ggml_tensor * d0 = nodes[0];
+    const size_t        ts = ggml_type_size(GGML_TYPE_F32);
+    l2_batch_strides    st{};
+    st.ne1 = (int) s0->ne[1];
+    st.ne2 = (int) s0->ne[2];
+    st.ss0 = s0->nb[0] / ts; st.ss1 = s0->nb[1] / ts; st.ss2 = s0->nb[2] / ts; st.ss3 = s0->nb[3] / ts;
+    st.ds0 = d0->nb[0] / ts; st.ds1 = d0->nb[1] / ts; st.ds2 = d0->nb[2] / ts; st.ds3 = d0->nb[3] / ts;
+
+    l2_norm_f32_batch_sycl<WARP_SIZE>(p, st, count, ncols, nrows_total, eps, ctx.stream());
+}

ggml/src/ggml-sycl/norm.hpp

@@ -29,4 +29,7 @@ void ggml_sycl_op_group_norm(ggml_backend_sycl_context& ctx, ggml_tensor* dst);
 
 void ggml_sycl_op_l2_norm(ggml_backend_sycl_context& ctx, ggml_tensor* dst);
 
+#define GGML_SYCL_L2_BATCH_MAX 8
+void ggml_sycl_l2_norm_batch(ggml_backend_sycl_context & ctx, ggml_tensor ** nodes, int count);
+
 #endif // GGML_SYCL_NORM_HPP

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

@@ -1110,6 +1110,9 @@ struct vk_device_struct {
     vk_pipeline pipeline_cumsum_multipass2_f32;
     vk_pipeline pipeline_argmax_f32;
     vk_pipeline pipeline_count_equal_i32;
+    vk_pipeline pipeline_dsv4_hc_comb_f32;
+    vk_pipeline pipeline_dsv4_hc_pre_f32;
+    vk_pipeline pipeline_dsv4_hc_post_f32;
     std::map<vk_solve_tri_pipeline_state, vk_pipeline> pipeline_solve_tri_f32;
     vk_pipeline pipeline_im2col_f32, pipeline_im2col_f32_f16;
     vk_pipeline pipeline_im2col_3d_f32, pipeline_im2col_3d_f32_f16;
@@ -1467,6 +1470,53 @@ struct vk_op_fwht_push_constants {
     float scale;
 };
 
+struct vk_op_dsv4_hc_comb_push_constants {
+    uint32_t n_tokens;
+
+    uint32_t nbm0; uint32_t nbm1;
+    uint32_t nbs0;
+    uint32_t nbb0;
+    uint32_t nbd0; uint32_t nbd1; uint32_t nbd2;
+
+    uint32_t m_offset;
+    uint32_t s_offset;
+    uint32_t b_offset;
+    uint32_t d_offset;
+
+    float eps;
+    uint32_t n_iter;
+};
+
+struct vk_op_dsv4_hc_pre_push_constants {
+    uint32_t n_embd;
+    uint32_t n_tokens;
+
+    uint32_t nbx0; uint32_t nbx1; uint32_t nbx2;
+    uint32_t nbw0; uint32_t nbw1;
+    uint32_t nbd0; uint32_t nbd1;
+
+    uint32_t x_offset;
+    uint32_t w_offset;
+    uint32_t d_offset;
+};
+
+struct vk_op_dsv4_hc_post_push_constants {
+    uint32_t n_embd;
+    uint32_t n_tokens;
+
+    uint32_t nbx0; uint32_t nbx1;
+    uint32_t nbr0; uint32_t nbr1; uint32_t nbr2;
+    uint32_t nbp0; uint32_t nbp1;
+    uint32_t nbc0; uint32_t nbc1; uint32_t nbc2;
+    uint32_t nbd0; uint32_t nbd1; uint32_t nbd2;
+
+    uint32_t x_offset;
+    uint32_t r_offset;
+    uint32_t p_offset;
+    uint32_t c_offset;
+    uint32_t d_offset;
+};
+
 struct vk_op_count_experts_push_constants {
     uint32_t ne00;
     uint32_t ne01;
@@ -2515,9 +2565,38 @@ static uint64_t vk_tensor_offset(const ggml_tensor * tensor) {
     return (uint8_t *) tensor->data - (uint8_t *) vk_ptr_base;
 }
 
-static uint32_t get_misalign_bytes(const ggml_backend_vk_context * ctx, const ggml_tensor * t)
-{
-    return ((vk_tensor_offset(t) + t->view_offs) & (ctx->device->properties.limits.minStorageBufferOffsetAlignment - 1));;
+static void ggml_vk_host_get(const vk_device& device, const void * ptr, vk_buffer& buf, size_t& buf_offset);
+
+static size_t ggml_vk_tensor_buffer_offset(const ggml_backend_vk_context * ctx, const ggml_tensor * t) {
+    // vk_tensor_offset() is relative to vk_ptr_base, but mapped host tensors need an offset relative to their Vulkan buffer.
+    if (ctx->device->uma) {
+        vk_buffer buf = nullptr;
+        size_t off = 0;
+        ggml_vk_host_get(ctx->device, t->data, buf, off);
+        if (buf) {
+            return off;
+        }
+    }
+    return (size_t)(vk_tensor_offset(t) + t->view_offs);
+}
+
+static size_t ggml_vk_descriptor_offset(size_t tensor_offset, size_t alignment, size_t type_size) {
+    // Move the descriptor back until its distance to the tensor is divisible by the tensor type size.
+    size_t descriptor_offset = tensor_offset & ~(alignment - 1);
+    while ((tensor_offset - descriptor_offset) % type_size != 0) {
+        GGML_ASSERT(descriptor_offset >= alignment);
+        descriptor_offset -= alignment;
+    }
+
+    return descriptor_offset;
+}
+
+static uint32_t get_misalign_bytes(const ggml_backend_vk_context * ctx, const ggml_tensor * t) {
+    const size_t tensor_offset = ggml_vk_tensor_buffer_offset(ctx, t);
+    const size_t descriptor_offset = ggml_vk_descriptor_offset(
+        tensor_offset, ctx->device->properties.limits.minStorageBufferOffsetAlignment, ggml_type_size(t->type));
+    GGML_ASSERT(tensor_offset - descriptor_offset <= UINT32_MAX);
+    return tensor_offset - descriptor_offset;
 }
 
 static uint32_t ggml_vk_concat_unit_size(ggml_type type) {
@@ -2602,6 +2681,32 @@ template <> void init_pushconst_tensor_offsets(ggml_backend_vk_context * ctx, vk
     GGML_UNUSED(src3);
 }
 
+template <> void init_pushconst_tensor_offsets(ggml_backend_vk_context * ctx, vk_op_dsv4_hc_comb_push_constants &p, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, const ggml_tensor * src3, ggml_tensor * dst) {
+    p.m_offset = get_misalign_bytes(ctx, src0) / ggml_type_size(src0->type);
+    p.s_offset = get_misalign_bytes(ctx, src1) / ggml_type_size(src1->type);
+    p.b_offset = get_misalign_bytes(ctx, src2) / ggml_type_size(src2->type);
+    p.d_offset = get_misalign_bytes(ctx, dst)  / ggml_type_size(dst->type);
+
+    GGML_UNUSED(src3);
+}
+
+template <> void init_pushconst_tensor_offsets(ggml_backend_vk_context * ctx, vk_op_dsv4_hc_pre_push_constants &p, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, const ggml_tensor * src3, ggml_tensor * dst) {
+    p.x_offset = get_misalign_bytes(ctx, src0) / ggml_type_size(src0->type);
+    p.w_offset = get_misalign_bytes(ctx, src1) / ggml_type_size(src1->type);
+    p.d_offset = get_misalign_bytes(ctx, dst)  / ggml_type_size(dst->type);
+
+    GGML_UNUSED(src2);
+    GGML_UNUSED(src3);
+}
+
+template <> void init_pushconst_tensor_offsets(ggml_backend_vk_context * ctx, vk_op_dsv4_hc_post_push_constants &p, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, const ggml_tensor * src3, ggml_tensor * dst) {
+    p.x_offset = get_misalign_bytes(ctx, src0) / ggml_type_size(src0->type);
+    p.r_offset = get_misalign_bytes(ctx, src1) / ggml_type_size(src1->type);
+    p.p_offset = get_misalign_bytes(ctx, src2) / ggml_type_size(src2->type);
+    p.c_offset = get_misalign_bytes(ctx, src3) / ggml_type_size(src3->type);
+    p.d_offset = get_misalign_bytes(ctx, dst)  / ggml_type_size(dst->type);
+}
+
 struct ggml_backend_vk_buffer_context {
     vk_device_ref device;
     vk_buffer dev_buffer;
@@ -5948,6 +6053,16 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {
         ggml_vk_create_pipeline(device, device->pipeline_count_experts, "count_experts", count_experts_len, count_experts_data, "main", 2, sizeof(vk_op_count_experts_push_constants), {1, 1, 1}, {}, 1, true);
     }
 
+    // comb holds a token's 4x4 matrix in one 16-lane slice of a subgroup, so it
+    // needs at least 16 lanes, pinned to a known size.
+    if (device->subgroup_basic && device->subgroup_shuffle && device->subgroup_require_full_support && device->subgroup_size >= 16) {
+        const uint32_t tokens_per_workgroup = 4 * (device->subgroup_size / 16);
+        ggml_vk_create_pipeline(device, device->pipeline_dsv4_hc_comb_f32, "dsv4_hc_comb_f32", dsv4_hc_comb_f32_len, dsv4_hc_comb_f32_data, "main", 4, sizeof(vk_op_dsv4_hc_comb_push_constants), {tokens_per_workgroup, 1, 1}, { device->subgroup_size }, 1, true, true, device->subgroup_size);
+    }
+
+    ggml_vk_create_pipeline(device, device->pipeline_dsv4_hc_pre_f32,  "dsv4_hc_pre_f32",  dsv4_hc_pre_f32_len,  dsv4_hc_pre_f32_data,  "main", 3, sizeof(vk_op_dsv4_hc_pre_push_constants),  {256, 1, 1}, { 256 }, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dsv4_hc_post_f32, "dsv4_hc_post_f32", dsv4_hc_post_f32_len, dsv4_hc_post_f32_data, "main", 5, sizeof(vk_op_dsv4_hc_post_push_constants), {256, 1, 1}, { 256 }, 1);
+
     for (auto &s : device->pipeline_solve_tri_f32) {
         const vk_solve_tri_pipeline_state &state = s.first;
 
@@ -8265,10 +8380,12 @@ static vk_subbuffer ggml_vk_tensor_subbuffer(
 
     size_t size = ggml_nbytes(tensor);
 
-    size_t misalign_bytes = offset & (ctx->device->properties.limits.minStorageBufferOffsetAlignment - 1);
+    const size_t descriptor_offset = ggml_vk_descriptor_offset(
+        offset, ctx->device->properties.limits.minStorageBufferOffsetAlignment, ggml_type_size(tensor->type));
+    const size_t misalign_bytes = offset - descriptor_offset;
     // The shader must support misaligned offsets when indexing into the buffer
     GGML_ASSERT(allow_misalign || misalign_bytes == 0);
-    offset &= ~misalign_bytes;
+    offset = descriptor_offset;
     size += misalign_bytes;
 
     return vk_subbuffer{buffer, offset, size};
@@ -10173,6 +10290,98 @@ static void ggml_vk_fwht(ggml_backend_vk_context * ctx, vk_context& subctx, cons
     ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { src_buf, dst_buf }, pc, { workgroups_x, 1, 1 });
 }
 
+static uint32_t ggml_vk_nb_elem(const ggml_tensor * t, int i) {
+    return (uint32_t)(t->nb[i] / ggml_type_size(t->type));
+}
+
+static void ggml_vk_dsv4_hc_comb(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * mixes, const ggml_tensor * scale, const ggml_tensor * base, ggml_tensor * dst) {
+    VK_LOG_DEBUG("ggml_vk_dsv4_hc_comb(" << mixes << ", " << scale << ", " << base << ", " << dst << ")");
+
+    vk_pipeline pipeline = ctx->device->pipeline_dsv4_hc_comb_f32;
+    GGML_ASSERT(pipeline != nullptr);
+
+    const uint32_t n_tokens = (uint32_t)mixes->ne[1];
+
+    ggml_pipeline_request_descriptor_sets(ctx, pipeline, 1);
+
+    const vk_subbuffer mixes_buf = ggml_vk_tensor_subbuffer(ctx, mixes, true);
+    const vk_subbuffer scale_buf = ggml_vk_tensor_subbuffer(ctx, scale, true);
+    const vk_subbuffer base_buf  = ggml_vk_tensor_subbuffer(ctx, base,  true);
+    const vk_subbuffer dst_buf   = ggml_vk_tensor_subbuffer(ctx, dst,   true);
+
+    vk_op_dsv4_hc_comb_push_constants pc = {
+        n_tokens,
+        ggml_vk_nb_elem(mixes, 0), ggml_vk_nb_elem(mixes, 1),
+        ggml_vk_nb_elem(scale, 0),
+        ggml_vk_nb_elem(base,  0),
+        ggml_vk_nb_elem(dst,   0), ggml_vk_nb_elem(dst, 1), ggml_vk_nb_elem(dst, 2),
+        0, 0, 0, 0,
+        ggml_get_op_params_f32(dst, 0),
+        (uint32_t)ggml_get_op_params_i32(dst, 1),
+    };
+    init_pushconst_tensor_offsets(ctx, pc, mixes, scale, base, nullptr, dst);
+
+    ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { mixes_buf, scale_buf, base_buf, dst_buf }, pc, { n_tokens, 1, 1 });
+}
+
+static void ggml_vk_dsv4_hc_pre(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * x, const ggml_tensor * weights, ggml_tensor * dst) {
+    VK_LOG_DEBUG("ggml_vk_dsv4_hc_pre(" << x << ", " << weights << ", " << dst << ")");
+
+    vk_pipeline pipeline = ctx->device->pipeline_dsv4_hc_pre_f32;
+    GGML_ASSERT(pipeline != nullptr);
+
+    const uint32_t n_embd   = (uint32_t)x->ne[0];
+    const uint32_t n_tokens = (uint32_t)x->ne[2];
+
+    ggml_pipeline_request_descriptor_sets(ctx, pipeline, 1);
+
+    const vk_subbuffer x_buf = ggml_vk_tensor_subbuffer(ctx, x,       true);
+    const vk_subbuffer w_buf = ggml_vk_tensor_subbuffer(ctx, weights, true);
+    const vk_subbuffer d_buf = ggml_vk_tensor_subbuffer(ctx, dst,     true);
+
+    vk_op_dsv4_hc_pre_push_constants pc = {
+        n_embd, n_tokens,
+        ggml_vk_nb_elem(x, 0), ggml_vk_nb_elem(x, 1), ggml_vk_nb_elem(x, 2),
+        ggml_vk_nb_elem(weights, 0), ggml_vk_nb_elem(weights, 1),
+        ggml_vk_nb_elem(dst, 0), ggml_vk_nb_elem(dst, 1),
+        0, 0, 0,
+    };
+    init_pushconst_tensor_offsets(ctx, pc, x, weights, nullptr, nullptr, dst);
+
+    ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { x_buf, w_buf, d_buf }, pc, { n_embd, n_tokens, 1 });
+}
+
+static void ggml_vk_dsv4_hc_post(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * x, const ggml_tensor * residual, const ggml_tensor * post, const ggml_tensor * comb, ggml_tensor * dst) {
+    VK_LOG_DEBUG("ggml_vk_dsv4_hc_post(" << x << ", " << residual << ", " << post << ", " << comb << ", " << dst << ")");
+
+    vk_pipeline pipeline = ctx->device->pipeline_dsv4_hc_post_f32;
+    GGML_ASSERT(pipeline != nullptr);
+
+    const uint32_t n_embd   = (uint32_t)x->ne[0];
+    const uint32_t n_tokens = (uint32_t)x->ne[1];
+
+    ggml_pipeline_request_descriptor_sets(ctx, pipeline, 1);
+
+    const vk_subbuffer x_buf = ggml_vk_tensor_subbuffer(ctx, x,        true);
+    const vk_subbuffer r_buf = ggml_vk_tensor_subbuffer(ctx, residual, true);
+    const vk_subbuffer p_buf = ggml_vk_tensor_subbuffer(ctx, post,     true);
+    const vk_subbuffer c_buf = ggml_vk_tensor_subbuffer(ctx, comb,     true);
+    const vk_subbuffer d_buf = ggml_vk_tensor_subbuffer(ctx, dst,      true);
+
+    vk_op_dsv4_hc_post_push_constants pc = {
+        n_embd, n_tokens,
+        ggml_vk_nb_elem(x, 0), ggml_vk_nb_elem(x, 1),
+        ggml_vk_nb_elem(residual, 0), ggml_vk_nb_elem(residual, 1), ggml_vk_nb_elem(residual, 2),
+        ggml_vk_nb_elem(post, 0), ggml_vk_nb_elem(post, 1),
+        ggml_vk_nb_elem(comb, 0), ggml_vk_nb_elem(comb, 1), ggml_vk_nb_elem(comb, 2),
+        ggml_vk_nb_elem(dst,  0), ggml_vk_nb_elem(dst,  1), ggml_vk_nb_elem(dst,  2),
+        0, 0, 0, 0, 0,
+    };
+    init_pushconst_tensor_offsets(ctx, pc, x, residual, post, comb, dst);
+
+    ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { x_buf, r_buf, p_buf, c_buf, d_buf }, pc, { n_embd, n_tokens, 1 });
+}
+
 static void ggml_vk_mul_mat(ggml_backend_vk_context * ctx, vk_context& subctx, const struct ggml_cgraph * cgraph, int node_idx) {
     ggml_tensor * dst = cgraph->nodes[node_idx];
     ggml_tensor * src0 = dst->src[0];
@@ -12155,7 +12364,9 @@ template <> void init_pushconst_tensor_offsets(ggml_backend_vk_context * ctx, vk
     const uint32_t b_offset = get_misalign_bytes(ctx, src1) / ggml_type_size(src1->type);
     const uint32_t d_offset = get_misalign_bytes(ctx, dst) / ggml_type_size(dst->type);
 
-    GGML_ASSERT(dst->op != GGML_OP_GET_ROWS || (a_offset == 0 && b_offset == 0 && d_offset == 0));
+    GGML_ASSERT(a_offset <= 0xFFFF);
+    GGML_ASSERT(b_offset <= 0xFF);
+    GGML_ASSERT(d_offset <= 0xFF);
 
     p.misalign_offsets = (a_offset << 16) | (b_offset << 8) | d_offset;
 
@@ -16189,6 +16400,18 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_cgraph * cgr
     case GGML_OP_CUMSUM:
         ggml_vk_cumsum(ctx, compute_ctx, src0, node);
 
+        break;
+    case GGML_OP_DSV4_HC_COMB:
+        ggml_vk_dsv4_hc_comb(ctx, compute_ctx, src0, src1, src2, node);
+
+        break;
+    case GGML_OP_DSV4_HC_PRE:
+        ggml_vk_dsv4_hc_pre(ctx, compute_ctx, src0, src1, node);
+
+        break;
+    case GGML_OP_DSV4_HC_POST:
+        ggml_vk_dsv4_hc_post(ctx, compute_ctx, src0, src1, src2, src3, node);
+
         break;
     case GGML_OP_MEAN:
         ggml_vk_mean(ctx, compute_ctx, src0, node);
@@ -19256,6 +19479,31 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
                 }
                 return false;
             }
+        case GGML_OP_DSV4_HC_COMB:
+        case GGML_OP_DSV4_HC_PRE:
+        case GGML_OP_DSV4_HC_POST:
+            {
+                if (op->type != GGML_TYPE_F32) {
+                    return false;
+                }
+                for (uint32_t i = 0; i < GGML_MAX_SRC; ++i) {
+                    if (op->src[i] && op->src[i]->type != GGML_TYPE_F32) {
+                        return false;
+                    }
+                }
+                // hc is hardcoded to 4 in the shaders. ggml only constrains it
+                // to 4 for COMB, so PRE/POST have to be checked here.
+                if (op->op == GGML_OP_DSV4_HC_PRE && op->src[0]->ne[1] != 4) {
+                    return false;
+                }
+                if (op->op == GGML_OP_DSV4_HC_POST && op->src[1]->ne[1] != 4) {
+                    return false;
+                }
+                if (op->op == GGML_OP_DSV4_HC_COMB) {
+                    return device->pipeline_dsv4_hc_comb_f32 != nullptr;
+                }
+                return true;
+            }
         case GGML_OP_SOLVE_TRI:
             {
                 if (op->type != GGML_TYPE_F32 || op->src[0]->type != GGML_TYPE_F32) {
@@ -20244,6 +20492,13 @@ static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_cgraph *
             tensor_clone = ggml_sum_rows(ggml_ctx, src_clone[0]);
         } else if (tensor->op == GGML_OP_CUMSUM) {
             tensor_clone = ggml_cumsum(ggml_ctx, src_clone[0]);
+        } else if (tensor->op == GGML_OP_DSV4_HC_COMB) {
+            tensor_clone = ggml_dsv4_hc_comb(ggml_ctx, src_clone[0], src_clone[1], src_clone[2],
+                ggml_get_op_params_f32(tensor, 0), ggml_get_op_params_i32(tensor, 1));
+        } else if (tensor->op == GGML_OP_DSV4_HC_PRE) {
+            tensor_clone = ggml_dsv4_hc_pre(ggml_ctx, src_clone[0], src_clone[1]);
+        } else if (tensor->op == GGML_OP_DSV4_HC_POST) {
+            tensor_clone = ggml_dsv4_hc_post(ggml_ctx, src_clone[0], src_clone[1], src_clone[2], src_clone[3]);
         } else if (tensor->op == GGML_OP_MEAN) {
             tensor_clone = ggml_mean(ggml_ctx, src_clone[0]);
         } else if (tensor->op == GGML_OP_ARGMAX) {

ggml/src/ggml-vulkan/vulkan-shaders/dsv4_hc_comb.comp

@@ -0,0 +1,90 @@
+#version 450
+
+#extension GL_EXT_control_flow_attributes : require
+#extension GL_KHR_shader_subgroup_basic : require
+#extension GL_KHR_shader_subgroup_shuffle : require
+
+// 16 lanes per token, indexed idst + hc*isrc: idst in bits 0..1, isrc in bits 2..3,
+// so subgroupShuffleXor by 1|2 reduces a row and by 4|8 a column.
+
+layout(constant_id = 0) const uint SUBGROUP_SIZE = 32;
+
+layout(local_size_x_id = 0, local_size_y = 4, local_size_z = 1) in;
+
+layout(push_constant) uniform parameter
+{
+    uint n_tokens;
+
+    uint nbm0; uint nbm1;   // mixes
+    uint nbs0;              // scale
+    uint nbb0;              // base
+    uint nbd0; uint nbd1; uint nbd2;   // dst
+
+    uint m_offset;
+    uint s_offset;
+    uint b_offset;
+    uint d_offset;
+
+    float eps;
+    uint n_iter;
+};
+
+layout(binding = 0, std430) readonly buffer M { float data_m[]; };
+layout(binding = 1, std430) readonly buffer S { float data_s[]; };
+layout(binding = 2, std430) readonly buffer B { float data_b[]; };
+layout(binding = 3, std430) writeonly buffer D { float data_d[]; };
+
+const uint hc          = 4;
+const uint comb_offset = 2 * hc;
+
+const uint TOKENS_PER_SUBGROUP = SUBGROUP_SIZE / 16;
+
+void main() {
+    const uint lane = gl_SubgroupInvocationID;
+    const uint blk  = lane >> 4;    // which 16-lane block, i.e. which token
+    const uint idx  = lane & 15;    // idst + hc*isrc
+
+    const uint sg = gl_WorkGroupID.x * gl_WorkGroupSize.y + gl_SubgroupID;
+    const uint it = sg * TOKENS_PER_SUBGROUP + blk;
+
+    // no early return, the shuffles need every lane; out-of-range blocks compute a discarded value
+    const bool in_range = it < n_tokens;
+
+    const float scale_comb = data_s[s_offset + 2 * nbs0];
+
+    float v = 0.0f;
+    if (in_range) {
+        v = data_m[m_offset + (comb_offset + idx) * nbm0 + it * nbm1] * scale_comb
+          + data_b[b_offset + (comb_offset + idx) * nbb0];
+    }
+
+    // Softmax across destinations: the four lanes sharing an isrc.
+    float vmax = max(v, subgroupShuffleXor(v, 1));
+    vmax = max(vmax, subgroupShuffleXor(vmax, 2));
+    v = exp(v - vmax);
+
+    float sum = v + subgroupShuffleXor(v, 1);
+    sum += subgroupShuffleXor(sum, 2);
+    v = v / sum + eps;
+
+    // Normalize columns: equal destination indices are four lanes apart.
+    sum = v + subgroupShuffleXor(v, 4);
+    sum += subgroupShuffleXor(sum, 8);
+    v /= sum + eps;
+
+    for (uint i = 1; i < n_iter; ++i) {
+        sum = v + subgroupShuffleXor(v, 1);
+        sum += subgroupShuffleXor(sum, 2);
+        v /= sum + eps;
+
+        sum = v + subgroupShuffleXor(v, 4);
+        sum += subgroupShuffleXor(sum, 8);
+        v /= sum + eps;
+    }
+
+    if (in_range) {
+        const uint idst = idx & 3;
+        const uint isrc = idx >> 2;
+        data_d[d_offset + idst * nbd0 + isrc * nbd1 + it * nbd2] = v;
+    }
+}

ggml/src/ggml-vulkan/vulkan-shaders/dsv4_hc_post.comp

@@ -0,0 +1,83 @@
+#version 450
+
+#extension GL_EXT_control_flow_attributes : require
+
+// Fan one stream back out to hc streams and add the combination-weighted
+// residuals:
+//
+//   dst[i0, idst, it] = x[i0, it]*post[idst, it]
+//                     + sum_isrc residual[i0, isrc, it]*comb[idst, isrc, it]
+
+layout(constant_id = 0) const uint BLOCK_SIZE = 256;
+
+layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
+
+layout(push_constant) uniform parameter
+{
+    uint n_embd;
+    uint n_tokens;
+
+    uint nbx0; uint nbx1;              // x
+    uint nbr0; uint nbr1; uint nbr2;   // residual
+    uint nbp0; uint nbp1;              // post
+    uint nbc0; uint nbc1; uint nbc2;   // comb
+    uint nbd0; uint nbd1; uint nbd2;   // dst
+
+    uint x_offset;
+    uint r_offset;
+    uint p_offset;
+    uint c_offset;
+    uint d_offset;
+};
+
+layout(binding = 0, std430) readonly buffer X { float data_x[]; };
+layout(binding = 1, std430) readonly buffer R { float data_r[]; };
+layout(binding = 2, std430) readonly buffer P { float data_p[]; };
+layout(binding = 3, std430) readonly buffer C { float data_c[]; };
+layout(binding = 4, std430) writeonly buffer D { float data_d[]; };
+
+const uint hc = 4;
+
+shared float post_s[hc];
+shared float comb_s[hc * hc];
+
+void main() {
+    const uint tid = gl_LocalInvocationID.x;
+    const uint it  = gl_WorkGroupID.y;
+
+    if (tid < hc) {
+        post_s[tid] = data_p[p_offset + tid * nbp0 + it * nbp1];
+    }
+    if (tid < hc * hc) {
+        const uint idst = tid & 3;
+        const uint isrc = tid >> 2;
+        comb_s[tid] = data_c[c_offset + idst * nbc0 + isrc * nbc1 + it * nbc2];
+    }
+    barrier();
+
+    // After the barrier, so every invocation reaches it.
+    const uint i0 = gl_WorkGroupID.x * BLOCK_SIZE + tid;
+    if (i0 >= n_embd) {
+        return;
+    }
+
+    const float xv = data_x[x_offset + i0 * nbx0 + it * nbx1];
+
+    const uint rb = r_offset + i0 * nbr0 + it * nbr2;
+
+    float r[hc];
+    [[unroll]]
+    for (uint isrc = 0; isrc < hc; ++isrc) {
+        r[isrc] = data_r[rb + isrc * nbr1];
+    }
+
+    [[unroll]]
+    for (uint idst = 0; idst < hc; ++idst) {
+        float result = xv * post_s[idst];
+        [[unroll]]
+        for (uint isrc = 0; isrc < hc; ++isrc) {
+            result = fma(r[isrc], comb_s[idst + hc * isrc], result);
+        }
+        data_d[d_offset + i0 * nbd0 + idst * nbd1 + it * nbd2] = result;
+    }
+}

ggml/src/ggml-vulkan/vulkan-shaders/dsv4_hc_pre.comp

@@ -0,0 +1,59 @@
+#version 450
+
+#extension GL_EXT_control_flow_attributes : require
+
+// Collapse the hc residual streams of a token into one, weighted per stream:
+//
+//   dst[i0, it] = sum_ih x[i0, ih, it] * weights[ih, it]
+
+layout(constant_id = 0) const uint BLOCK_SIZE = 256;
+
+layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
+
+layout(push_constant) uniform parameter
+{
+    uint n_embd;
+    uint n_tokens;
+
+    uint nbx0; uint nbx1; uint nbx2;   // x
+    uint nbw0; uint nbw1;              // weights
+    uint nbd0; uint nbd1;              // dst
+
+    uint x_offset;
+    uint w_offset;
+    uint d_offset;
+};
+
+layout(binding = 0, std430) readonly buffer X { float data_x[]; };
+layout(binding = 1, std430) readonly buffer W { float data_w[]; };
+layout(binding = 2, std430) writeonly buffer D { float data_d[]; };
+
+const uint hc = 4;
+
+shared float w[hc];
+
+void main() {
+    const uint tid = gl_LocalInvocationID.x;
+    const uint it  = gl_WorkGroupID.y;
+
+    if (tid < hc) {
+        w[tid] = data_w[w_offset + tid * nbw0 + it * nbw1];
+    }
+    barrier();
+
+    // After the barrier, so every invocation reaches it.
+    const uint i0 = gl_WorkGroupID.x * BLOCK_SIZE + tid;
+    if (i0 >= n_embd) {
+        return;
+    }
+
+    const uint xb = x_offset + i0 * nbx0 + it * nbx2;
+
+    float result = 0.0f;
+    [[unroll]]
+    for (uint ih = 0; ih < hc; ++ih) {
+        result = fma(data_x[xb + ih * nbx1], w[ih], result);
+    }
+
+    data_d[d_offset + i0 * nbd0 + it * nbd1] = result;
+}

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

@@ -27,10 +27,10 @@ void main() {
             const uint i11 = gid_z / p.ne12;
             const uint i12 = gid_z % p.ne12;
 
-            const uint i01 = data_b[i10*p.nb10 + i11*p.nb11 + i12*p.nb12];
+            const uint i01 = data_b[get_boffset() + i10*p.nb10 + i11*p.nb11 + i12*p.nb12];
 
-            const uint a_offset = i01*p.nb01 + i11*p.nb02 + i12*p.nb03;
-            const uint d_offset = i10*p.nb21 + i11*p.nb22 + i12*p.nb23;
+            const uint a_offset = get_aoffset() + i01*p.nb01 + i11*p.nb02 + i12*p.nb03;
+            const uint d_offset = get_doffset() + i10*p.nb21 + i11*p.nb22 + i12*p.nb23;
 
             const uint ib = a_offset + i00/QUANT_K; // block index
             const uint iqs = (i00%QUANT_K)/QUANT_R; // quant index

ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp

@@ -1042,6 +1042,9 @@ void process_shaders() {
     string_to_spv("fwht_f32", "fwht.comp", {});
     string_to_spv("fwht_shmem_f32", "fwht.comp", {{"FWHT_SHMEM", "1"}});
     string_to_spv("count_equal_i32", "count_equal.comp", merge_maps(base_dict, {{"A_TYPE", "int"}, {"B_TYPE", "int"}, {"D_TYPE", "int"}}));
+    string_to_spv("dsv4_hc_comb_f32", "dsv4_hc_comb.comp", {});
+    string_to_spv("dsv4_hc_pre_f32",  "dsv4_hc_pre.comp",  {});
+    string_to_spv("dsv4_hc_post_f32", "dsv4_hc_post.comp", {});
     string_to_spv("cumsum_f32", "cumsum.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}}));
     string_to_spv("cumsum_multipass1_f32", "cumsum_multipass1.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}}));
     string_to_spv("cumsum_multipass2_f32", "cumsum_multipass2.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}}));

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

@@ -4324,12 +4324,22 @@ static bool ggml_backend_webgpu_device_supports_op(ggml_backend_dev_t dev, const
                            src0->type == GGML_TYPE_F32 && (src1->type == GGML_TYPE_I64 || src1->type == GGML_TYPE_I32));
             break;
         case GGML_OP_GET_ROWS:
-            if (src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16 || ggml_webgpu_supported_qtype(src0->type)) {
-                supports_op = (op->type == GGML_TYPE_F32);
-            } else if (src0->type == GGML_TYPE_I32) {
-                supports_op = op->type == GGML_TYPE_I32;
+            {
+                const size_t storage_alignment =
+                    ctx->webgpu_global_ctx->capabilities.limits.minStorageBufferOffsetAlignment;
+                const size_t src_address_unit =
+                    src0->type == GGML_TYPE_F32 && op->ne[0] % 4 == 0 ? 4 * sizeof(float) : ggml_type_size(src0->type);
+                if (ggml_webgpu_tensor_misalignment(src0, storage_alignment) % src_address_unit != 0) {
+                    break;
+                }
+                if (src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16 ||
+                    ggml_webgpu_supported_qtype(src0->type)) {
+                    supports_op = (op->type == GGML_TYPE_F32);
+                } else if (src0->type == GGML_TYPE_I32) {
+                    supports_op = op->type == GGML_TYPE_I32;
+                }
+                break;
             }
-            break;
         case GGML_OP_MUL_MAT:
             {
                 switch (src1->type) {

gguf-py/gguf/constants.py

@@ -215,6 +215,7 @@ class Keys:
         KV_LORA_RANK_SWA             = "{arch}.attention.kv_lora_rank_swa"
         SHARED_KV_LAYERS             = "{arch}.attention.shared_kv_layers"
         SLIDING_WINDOW_PATTERN       = "{arch}.attention.sliding_window_pattern"
+        RECURRENT_LAYERS             = "{arch}.attention.recurrent_layers"
         TEMPERATURE_SCALE            = "{arch}.attention.temperature_scale"
         ROPE_PATTERN                 = "{arch}.attention.rope_pattern"
 

gguf-py/gguf/gguf_writer.py

@@ -841,6 +841,9 @@ class GGUFWriter:
         else:
             self.add_array(key, value)
 
+    def add_recurrent_layers(self, value: Sequence[bool]) -> None:
+        self.add_array(Keys.Attention.RECURRENT_LAYERS.format(arch=self.arch), value)
+
     def add_rope_pattern(self, value: Sequence[bool]) -> None:
         self.add_array(Keys.Attention.ROPE_PATTERN.format(arch=self.arch), value)
 

gguf-py/gguf/tensor_mapping.py

@@ -385,6 +385,7 @@ class TensorNameMap:
         MODEL_TENSOR.ATTN_SINKS: (
             "model.layers.{bid}.self_attn.sinks", # openai-moe
             "model.layers.{bid}.self_attn.attention_sink_bias", # mimov2
+            "model.layers.{bid}.self_attn.learnable_sink_param", # hy-v4
         ),
 
         MODEL_TENSOR.ATTN_GATE: (
@@ -392,6 +393,7 @@ class TensorNameMap:
             "model.layers.{bid}.linear_attn.in_proj_z",  # qwen3.5
             "model.layers.{bid}.self_attn.g_proj",    # step3.5 head-wise attention gate
             "model.layers.{bid}.self_attn.output_gate",  # minimax-01
+            "model.layers.{bid}.self_attn.linear_gate",  # hy-v4
         ),
 
         # Feed-forward norm
@@ -1329,6 +1331,42 @@ class TensorNameMap:
             "model.layers.{bid}.self_attn.index_q_norm", # MSA
         ),
 
+        MODEL_TENSOR.HC_ATTN_FN: (
+            "model.layers.{bid}.hc_attn_layer.hc_pre.hc_fn", # hy-v4
+        ),
+
+        MODEL_TENSOR.HC_ATTN_BASE: (
+            "model.layers.{bid}.hc_attn_layer.hc_pre.hc_base", # hy-v4
+        ),
+
+        MODEL_TENSOR.HC_ATTN_SCALE: (
+            "model.layers.{bid}.hc_attn_layer.hc_pre.hc_scale", # hy-v4
+        ),
+
+        MODEL_TENSOR.HC_FFN_FN: (
+            "model.layers.{bid}.hc_mlp_layer.hc_pre.hc_fn", # hy-v4
+        ),
+
+        MODEL_TENSOR.HC_FFN_BASE: (
+            "model.layers.{bid}.hc_mlp_layer.hc_pre.hc_base", # hy-v4
+        ),
+
+        MODEL_TENSOR.HC_FFN_SCALE: (
+            "model.layers.{bid}.hc_mlp_layer.hc_pre.hc_scale", # hy-v4
+        ),
+
+        MODEL_TENSOR.HC_HEAD_FN: (
+            "model.hc_head.hc_head_fn",  # hy-v4
+        ),
+
+        MODEL_TENSOR.HC_HEAD_BASE: (
+            "model.hc_head.hc_head_base", # hy-v4
+        ),
+
+        MODEL_TENSOR.HC_HEAD_SCALE: (
+            "model.hc_head.hc_head_scale", # hy-v4
+        ),
+
         ############################################################################
         # TODO: these do not belong to block_mappings_cfg - move them to mappings_cfg
         MODEL_TENSOR.ENC_OUTPUT_NORM: (

tests/test-backend-ops.cpp

@@ -2336,27 +2336,40 @@ struct test_get_rows : public test_case {
     const int r; // rows to get
     const int be1; // batch size
     const int be2; // batch size
-    const bool v; // view (non-contiguous src1)
+    const bool v; // view src1
+    const bool vs0; // view src0
 
     std::string vars() override {
-        return VARS_TO_STR7(type, n, m, r, be1, be2, v);
+        return VARS_TO_STR8(type, n, m, r, be1, be2, v, vs0);
     }
 
-    test_get_rows(ggml_type type = GGML_TYPE_F32, int n = 10, int m = 5, int r = 3, int be1 = 1, int be2 = 1, bool v = false)
-        : type(type), n(n), m(m), r(r), be1(be1), be2(be2), v(v) {}
+    test_get_rows(ggml_type type = GGML_TYPE_F32, int n = 10, int m = 5, int r = 3, int be1 = 1, int be2 = 1, bool v = false, bool vs0 = false)
+        : type(type), n(n), m(m), r(r), be1(be1), be2(be2), v(v), vs0(vs0) {}
 
     ggml_tensor * build_graph(ggml_context * ctx) override {
-        ggml_tensor * in = ggml_new_tensor_4d(ctx, type, n, m, be1, be2);
-        ggml_set_name(in, "in");
+        ggml_tensor * in;
+        if (vs0) {
+            const int offset_rows = 3;
+            const int padded_m = m + offset_rows;
+            ggml_tensor * in_padded = ggml_new_tensor_4d(ctx, type, n, padded_m, be1, be2);
+            ggml_set_name(in_padded, "in_padded");
+            in = ggml_view_4d(ctx, in_padded, n, m, be1, be2,
+                              in_padded->nb[1], in_padded->nb[2], in_padded->nb[3],
+                              offset_rows * in_padded->nb[1]);
+            ggml_set_name(in, "in_view");
+        } else {
+            in = ggml_new_tensor_4d(ctx, type, n, m, be1, be2);
+            ggml_set_name(in, "in");
+        }
 
-        ggml_tensor * rows = ggml_new_tensor_3d(ctx, GGML_TYPE_I32, r, be1, be2);
+        ggml_tensor * rows = ggml_new_tensor_3d(ctx, GGML_TYPE_I32, v ? r + 1 : r, be1, be2);
         ggml_set_name(rows, "rows");
         if (v) {
-            rows = ggml_view_3d(ctx, rows, r/2, be1, be2, rows->nb[1], rows->nb[2], 0);
+            rows = ggml_view_3d(ctx, rows, r/2, be1, be2, rows->nb[1], rows->nb[2], rows->nb[0]);
             ggml_set_name(rows, "view_of_rows");
         }
 
-        const bool grad_supported = ggml_is_matrix(in) && ggml_is_vector(rows);
+        const bool grad_supported = !vs0 && ggml_is_matrix(in) && ggml_is_vector(rows);
         if (grad_supported) {
             ggml_set_param(in);
             // rows is a constant input -> no gradients
@@ -2370,14 +2383,16 @@ struct test_get_rows : public test_case {
 
     void initialize_tensors(ggml_context * ctx) override {
         for (ggml_tensor * t = ggml_get_first_tensor(ctx); t != NULL; t = ggml_get_next_tensor(ctx, t)) {
+            if (ggml_is_view_op(t->op)) {
+                continue;
+            }
             if (t->type == GGML_TYPE_I32) {
-                if (ggml_is_view_op(t->op)) { continue; }
                 // rows
-                std::vector<int> data(r*be1*be2);
-                for (int i = 0; i < r*be1*be2; i++) {
+                std::vector<int> data(ggml_nelements(t));
+                for (size_t i = 0; i < data.size(); i++) {
                     data[i] = rand() % m;
                 }
-                ggml_backend_tensor_set(t, data.data(), 0, r * be1 * be2 * sizeof(int));
+                ggml_backend_tensor_set(t, data.data(), 0, data.size() * sizeof(int));
             } else {
                 init_tensor_uniform(t);
             }
@@ -7191,6 +7206,49 @@ struct test_group_norm_mul_add : public test_case {
     }
 };
 
+// GGML_OP_L2_NORM x N: independent same-shape norms in one graph (strided qkv views or
+// contiguous), consuming adds nested so the norms stay adjacent in the graph.
+struct test_l2_norm_batch : public test_case {
+    const ggml_type              type;
+    const std::array<int64_t, 4> ne;
+    const int                    n_norms;
+    const float                  eps;
+    const bool                   strided;
+
+    std::string vars() override { return VARS_TO_STR5(type, ne, n_norms, eps, strided); }
+    std::string op_desc(ggml_tensor * t) override { GGML_UNUSED(t); return "L2_NORM_BATCH"; }
+    bool run_whole_graph() override { return true; }
+
+    test_l2_norm_batch(ggml_type type = GGML_TYPE_F32, std::array<int64_t, 4> ne = { 128, 16, 16, 1 },
+                       int n_norms = 4, float eps = 1e-12f, bool strided = true)
+        : type(type), ne(ne), n_norms(n_norms), eps(eps), strided(strided) {}
+
+    ggml_tensor * build_graph(ggml_context * ctx) override {
+        GGML_ASSERT(n_norms >= 2 && n_norms <= 8);
+        ggml_tensor * parent = nullptr;
+        if (strided) {
+            parent = ggml_new_tensor_4d(ctx, type, ne[0], ne[1] * n_norms, ne[2], ne[3]);  // qkv buffer
+        }
+        ggml_tensor * norms[8];
+        for (int t = 0; t < n_norms; ++t) {
+            ggml_tensor * src;
+            if (strided) {
+                src = ggml_view_4d(ctx, parent, ne[0], ne[1], ne[2], ne[3], parent->nb[1], parent->nb[2],
+                                   parent->nb[3], t * ne[1] * parent->nb[1]);
+            } else {
+                src = ggml_new_tensor(ctx, type, 4, ne.data());
+            }
+            norms[t] = ggml_l2_norm(ctx, src, eps);
+        }
+        ggml_tensor * out = norms[n_norms - 1];
+        for (int t = n_norms - 2; t >= 0; --t) {
+            out = ggml_add(ctx, norms[t], out);
+        }
+        ggml_set_name(out, "out");
+        return out;
+    }
+};
+
 // GGML_OP_L2_NORM
 struct test_l2_norm : public test_case {
     const ggml_type type;
@@ -8792,6 +8850,11 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
     test_cases.emplace_back(new test_dsv4_hc_comb(17, 4));
     test_cases.emplace_back(new test_dsv4_hc_comb(257, 8));
     test_cases.emplace_back(new test_dsv4_hc_comb(17, 20));
+    // production n_iter (DeepSeek-V4 uses 20) across batch sizes that cross
+    // subgroup and workgroup boundaries; 1 = single-token decode
+    for (int64_t n_tokens : {1, 256, 336, 512, 513, 1024, 2048}) {
+        test_cases.emplace_back(new test_dsv4_hc_comb(n_tokens, 20));
+    }
 
     test_cases.emplace_back(new test_dsv4_hc_pre(1, 1));
     test_cases.emplace_back(new test_dsv4_hc_pre(31, 17));
@@ -8848,13 +8911,17 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
     for (ggml_type type : all_types) {
         for (int b : {1, 7}) {
             for (bool v : {false, true}) {
-                test_cases.emplace_back(new test_get_rows(type, 256, 5, 4, b, 1, v));
+                for (bool vs0 : {false, true}) {
+                    test_cases.emplace_back(new test_get_rows(type, 256, 5, 4, b, 1, v, vs0));
+                }
             }
         }
     }
     for (int b : {1, 7}) {
         for (bool v : {false, true}) {
-            test_cases.emplace_back(new test_get_rows(GGML_TYPE_I32, 256, 5, 4, b, 1, v));
+            for (bool vs0 : {false, true}) {
+                test_cases.emplace_back(new test_get_rows(GGML_TYPE_I32, 256, 5, 4, b, 1, v, vs0));
+            }
         }
     }
 
@@ -9471,6 +9538,10 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
             test_cases.emplace_back(new test_l2_norm(GGML_TYPE_F32, { n, 5, 4, 3 }, eps, false));
             test_cases.emplace_back(new test_l2_norm(GGML_TYPE_F32, { n, 5, 4, 3 }, eps, true));
             test_cases.emplace_back(new test_l2_norm(GGML_TYPE_F32, { n, 5, 4, 3 }, eps, false, true));
+            // sibling batching: strided (production shape) and contiguous, 2 and 4 wide
+            test_cases.emplace_back(new test_l2_norm_batch(GGML_TYPE_F32, { n, 5, 4, 3 }, 2, eps, true));
+            test_cases.emplace_back(new test_l2_norm_batch(GGML_TYPE_F32, { n, 5, 4, 3 }, 4, eps, true));
+            test_cases.emplace_back(new test_l2_norm_batch(GGML_TYPE_F32, { n, 5, 4, 3 }, 4, eps, false));
         }
         // row lengths that are not a multiple of 32, for the scalar (33) and float4 (132, 260) paths
         for (uint32_t n : { 33, 132, 260 }) {
@@ -11157,6 +11228,16 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_perf() {
         }
     }
 
+    // launch-overhead isolation: single L2_NORM launch vs batched siblings at the GDN
+    // production shape (strided qkv views) -- perf-mode only, the eval list has its own
+    // 2/4-wide coverage
+    for (int n : { 128, 256 }) {
+        test_cases.emplace_back(new test_l2_norm(GGML_TYPE_F32, { n, 16, 16, 1 }, 1e-12f, false, false));
+        test_cases.emplace_back(new test_l2_norm_batch(GGML_TYPE_F32, { n, 16, 16, 1 }, 2, 1e-12f, true));
+        test_cases.emplace_back(new test_l2_norm_batch(GGML_TYPE_F32, { n, 16, 16, 1 }, 4, 1e-12f, true));
+    }
+
+
     return test_cases;
 }
 

tests/test-model-resolution.cpp

@@ -15,6 +15,7 @@
 #include <cstdio>
 #include <cstdlib>
 #include <filesystem>
+#include <fstream>
 #include <map>
 #include <thread>
 #include <string>
@@ -139,7 +140,7 @@ static const std::vector<std::string> hole = {
     "dflash-model-Q8_0.gguf",
 };
 
-// unsloth-style naming with UD quants and a suffix MTP file
+// unsloth-style naming with UD quants and an uppercase infix MTP head
 static const std::vector<std::string> unsloth = {
     "model-UD-Q8_K_XL.gguf",
     "mmproj-BF16.gguf",
@@ -180,6 +181,13 @@ static const std::vector<std::string> spark = {
     "dspark-model-MXFP4.gguf",
 };
 
+// main weights plus a trailing-form mtp head, as in unsloth
+// gemma-4-E2B-it-GGUF (Model-Q4_0-mtp.gguf)
+static const std::vector<std::string> trailing = {
+    "model-BF16.gguf",
+    "model-BF16-mtp.gguf",
+};
+
 // dspark outranks dflash in the type auto-selection
 static const std::vector<std::string> dspark_dflash = {
     "model-Q8_0.gguf",
@@ -223,7 +231,7 @@ static const plan_case plan_cases[] = {
     // no tag and no default match falls back to the first model in the listing
     {"unsloth fallback", unsloth, "test/repo", "", true, true,
      "model-UD-Q8_K_XL.gguf", {"model-UD-Q8_K_XL.gguf"},
-     "mmproj-BF16.gguf", "", "", "", ""},
+     "mmproj-BF16.gguf", "model-MTP-BF16.gguf", "", "", ""},
 
     // explicit hf_file picks that exact file
     {"flat hf_file", flat, "test/repo", "model-BF16.gguf", false, false,
@@ -258,10 +266,11 @@ static const plan_case plan_cases[] = {
      "model-Q4_K_M.gguf", {"model-Q4_K_M.gguf"},
      "", "mtp-model-Q4_0.gguf", "dflash-model-Q8_0.gguf", "", ""},
 
-    // the mtp- keyword is case sensitive, a suffix -MTP file is not discovered
+    // the uppercase infix MTP head is a sidecar too; it resolves at the
+    // nearest quant to the tag when the exact one does not exist
     {"unsloth suffix mtp", unsloth, "test/repo:Q8_K_XL", "", true, false,
      "model-UD-Q8_K_XL.gguf", {"model-UD-Q8_K_XL.gguf"},
-     "mmproj-BF16.gguf", "", "", "", ""},
+     "mmproj-BF16.gguf", "model-MTP-BF16.gguf", "", "", ""},
 
     // vendor prefixes and the dot quant convention both match the tag,
     // first match wins between two files at the same quant
@@ -283,6 +292,45 @@ static const plan_case plan_cases[] = {
     {"spark tag sidecar", spark, "test/repo:BF16", "", true, false,
      "", {},
      "", "", "", "", "dspark-model-BF16.gguf"},
+
+    // a `<quant>-<sidecar>` tag resolves that sidecar alone, without a primary
+    {"hole quant-sidecar tag", hole, "test/repo:Q4_0-mtp", "", false, false,
+     "", {},
+     "", "mtp-model-Q4_0.gguf", "", "", ""},
+
+    {"spark quant-sidecar tag", spark, "test/repo:BF16-dspark", "", false, false,
+     "", {},
+     "", "", "", "", "dspark-model-BF16.gguf"},
+
+    // a bare sidecar tag resolves the sidecar at any quant
+    {"hole bare sidecar tag", hole, "test/repo:mtp", "", false, false,
+     "", {},
+     "", "mtp-model-Q4_0.gguf", "", "", ""},
+
+    // a trailing-form sidecar resolves as the sidecar alone, never as a primary
+    {"trailing quant-sidecar tag", trailing, "test/repo:BF16-mtp", "", false, false,
+     "", {},
+     "", "model-BF16-mtp.gguf", "", "", ""},
+
+    // the plain tag resolves the plain model; the trailing-token file is skipped
+    {"trailing plain tag", trailing, "test/repo:BF16", "", false, false,
+     "model-BF16.gguf", {"model-BF16.gguf"},
+     "", "", "", "", ""},
+
+    // a short-form sidecar (`mmproj-F16.gguf`) resolves by its bare quant tag
+    {"unsloth quant-sidecar tag", unsloth, "test/repo:BF16-mmproj", "", false, false,
+     "", {},
+     "mmproj-BF16.gguf", "", "", "", ""},
+
+    // an uppercase infix sidecar resolves through the lowercase tag
+    {"unsloth uppercase infix tag", unsloth, "test/repo:BF16-mtp", "", false, false,
+     "", {},
+     "", "model-MTP-BF16.gguf", "", "", ""},
+
+    // a sidecar token in the middle of the name resolves as the sidecar alone
+    {"subdir quant-sidecar tag", subdir, "test/repo:Q8_0-mtp", "", false, false,
+     "", {},
+     "", "model-mtp-Q8_0.gguf", "", "", ""},
 };
 
 static void check_plan(const plan_case & c) {
@@ -468,6 +516,90 @@ static void test_task_assembly() {
     g_repos.clear();
 }
 
+//
+// cache listing and removal against the isolated LLAMA_CACHE, using the
+// same filename grammar the plan tests exercise above
+//
+
+static void cache_put(const std::string & repo, const std::string & path) {
+    namespace fs = std::filesystem;
+    auto local = fs::path(cached(repo, path));
+    fs::create_directories(local.parent_path());
+    { std::ofstream(local) << "gguf"; }
+    auto repo_dir = local.parent_path().parent_path().parent_path();
+    auto refs = repo_dir / "refs";
+    fs::create_directories(refs);
+    auto ref = refs / "main";
+    if (!fs::exists(ref)) {
+        std::ofstream(ref) << COMMIT << "\n";
+    }
+}
+
+static bool cache_lists(const std::string & repo_tag) {
+    for (const auto & e : common_list_cached_models()) {
+        if (e.to_string() == repo_tag) {
+            return true;
+        }
+    }
+    return false;
+}
+
+static void test_cache_listing_and_remove() {
+    namespace fs = std::filesystem;
+    const std::string repo = "test/eh";
+
+    printf("test-model-resolution: cache listing and removal\n");
+
+    g_context = "cache fixture";
+    cache_put(repo, "gemma-4-E2B-it-BF16.gguf");        // main weights
+    cache_put(repo, "gemma-4-E2B-it-BF16-mtp.gguf");    // trailing-form draft head
+    cache_put(repo, "gemma-4-31B-it-MTP-BF16.gguf");    // uppercase infix draft head
+    cache_put(repo, "mmproj-gemma-4-E2B-it-BF16.gguf");  // mmproj sidecar
+    cache_put(repo, "mmproj-F16.gguf");                 // short-form mmproj
+    cache_put(repo, "model-mtp-Q8_0.gguf");             // mid-name mtp sidecar
+
+    // a sidecar is listed under `<quant>-<sidecar>` in every form and case it
+    // can be named; only the token-less main weights stay a loadable model
+    REQUIRE(cache_lists("test/eh:BF16"));
+    REQUIRE(cache_lists("test/eh:BF16-mtp"));
+    REQUIRE(cache_lists("test/eh:BF16-mmproj"));
+    REQUIRE(cache_lists("test/eh:F16-mmproj"));
+    REQUIRE(cache_lists("test/eh:Q8_0-mtp"));
+    REQUIRE(!cache_lists("test/eh:MTP"));
+
+    // a plain quant tag removes the model files and leaves every sidecar,
+    // whatever form or case its name carries
+    g_context = "remove plain quant";
+    REQUIRE(common_download_remove("test/eh:BF16"));
+    REQUIRE(!fs::exists(cached(repo, "gemma-4-E2B-it-BF16.gguf")));
+    REQUIRE(fs::exists(cached(repo, "gemma-4-E2B-it-BF16-mtp.gguf")));
+    REQUIRE(fs::exists(cached(repo, "gemma-4-31B-it-MTP-BF16.gguf")));
+    REQUIRE(fs::exists(cached(repo, "mmproj-gemma-4-E2B-it-BF16.gguf")));
+    REQUIRE(fs::exists(cached(repo, "mmproj-F16.gguf")));
+    REQUIRE(fs::exists(cached(repo, "model-mtp-Q8_0.gguf")));
+
+    // a `<quant>-<sidecar>` tag removes every sidecar listed under it,
+    // across forms and cases
+    g_context = "remove quant-sidecar";
+    REQUIRE(common_download_remove("test/eh:BF16-mtp"));
+    REQUIRE(!fs::exists(cached(repo, "gemma-4-E2B-it-BF16-mtp.gguf")));
+    REQUIRE(!fs::exists(cached(repo, "gemma-4-31B-it-MTP-BF16.gguf")));
+    REQUIRE(fs::exists(cached(repo, "mmproj-gemma-4-E2B-it-BF16.gguf")));
+
+    // the short form removes by its bare quant tag
+    REQUIRE(common_download_remove("test/eh:F16-mmproj"));
+    REQUIRE(!fs::exists(cached(repo, "mmproj-F16.gguf")));
+
+    REQUIRE(common_download_remove("test/eh:Q8_0-mtp"));
+    REQUIRE(!fs::exists(cached(repo, "model-mtp-Q8_0.gguf")));
+
+    // a bare sidecar tag is ambiguous across quants and is rejected
+    g_context = "remove bare sidecar";
+    cache_put(repo, "mtp-Model-Q4_0.gguf");
+    REQUIRE(!common_download_remove("test/eh:mtp"));
+    REQUIRE(fs::exists(cached(repo, "mtp-Model-Q4_0.gguf")));
+}
+
 int main(void) {
     // unbuffered, so a crash cannot swallow the reports already printed
     setvbuf(stdout, nullptr, _IONBF, 0);
@@ -496,6 +628,7 @@ int main(void) {
 
     test_plan_resolution();
     test_task_assembly();
+    test_cache_listing_and_remove();
 
     server.stop();
     server_thread.join();

tools/server/server-models.cpp

@@ -80,18 +80,19 @@ struct server_lru_sched {
     }
 
     // returns "" if no model can be given up
-    std::string pick_victim(std::unique_lock<std::mutex> & lk, const std::string & exclude) {
+    std::string pick_victim(std::unique_lock<std::mutex> & lk) {
         check_lock(lk);
         std::string victim;
         int64_t victim_last_used = 0;
         for (const auto & m : models.mapping) {
-            if (m.first == exclude) {
-                continue;
-            }
             // a busy model is mid-request, one still coming up has no request to finish
             if (m.second.req_count != 0 || !m.second.meta.is_ready_or_sleep()) {
                 continue;
             }
+            // already on its way out, or a queued request wants it
+            if (models.stopping_models.count(m.first) || find(m.first)) {
+                continue;
+            }
             if (victim.empty() || m.second.meta.last_used < victim_last_used) {
                 victim           = m.first;
                 victim_last_used = m.second.meta.last_used;
@@ -109,7 +110,7 @@ struct server_lru_sched {
             SRV_INF("request for name=%s joined the queue, %d waiting\n", model_id.c_str(), e->n_waiters);
             return;
         }
-        queue.push_back({ model_id, 1, false, false });
+        queue.push_back({ model_id, 1, false });
         SRV_INF("models_max reached, request for name=%s queued at position %zu\n",
                 model_id.c_str(), queue.size());
     }
@@ -144,85 +145,67 @@ struct server_lru_sched {
         return true;
     }
 
-    // ok means the model is up: drop the entry, the other waiters just watch its status now
+    // on failure the entry is back in line; on success it stays until its waiters leave,
+    // so the model coming up is never picked as a victim before they use it
     void claim_done(std::unique_lock<std::mutex> & lk, const std::string & model_id, bool ok) {
         check_lock(lk);
+        if (ok) {
+            return;
+        }
         for (auto it = queue.begin(); it != queue.end(); ++it) {
             if (it->model_id == model_id) {
-                if (ok) {
-                    queue.erase(it);
-                } else {
-                    it->loading = false;
-                }
+                it->loading = false;
                 return;
             }
         }
     }
 
-    // a model is on its way out for this entry, so other requests do not also give up one
-    void mark_slot_pending(std::unique_lock<std::mutex> & lk, const std::string & model_id) {
+    // evict idle models while queued requests outnumber the slots that are free or being freed
+    // caller must hold models.mutex; never blocks, so it is safe from any thread
+    void tick(std::unique_lock<std::mutex> & lk) {
         check_lock(lk);
-        if (entry_t * e = find(model_id)) {
-            e->slot_pending = true;
+        if (models.base_params.models_max <= 0 || queue.empty()) {
+            return;
         }
-    }
-
-    // model_id went idle: give up its slot if a queued request needs one
-    // thread-safe, caller must NOT hold models.mutex
-    void on_model_idle(const std::string & model_id) {
-        if (models.base_params.models_max <= 0) {
-            return; // no limit, nothing is ever queued
-        }
-        {
-            std::unique_lock<std::mutex> lk(models.mutex);
-            if (queue.empty()) {
-                return;
-            }
-            size_t promised     = 0;
-            bool   has_unserved = false;
-            for (const auto & e : queue) {
-                if (e.needs_slot()) {
-                    has_unserved = true;
-                } else {
-                    promised++;
-                }
-            }
-            if (!has_unserved) {
-                return;
-            }
-            if ((int) count_running() - (int) promised < models.base_params.models_max) {
-                return; // a slot is already on its way
-            }
-            // never give up a model that a queued request wants
-            for (const auto & e : queue) {
-                if (e.model_id == model_id) {
-                    return;
-                }
-            }
-            auto it = models.mapping.find(model_id);
-            if (it == models.mapping.end() || it->second.req_count != 0 || !it->second.meta.is_ready_or_sleep()) {
-                return;
-            }
-            for (auto & e : queue) {
-                if (!e.slot_pending) {
-                    e.slot_pending = true;
-                    break;
+        int n_running  = 0;
+        int n_stopping = 0;
+        for (const auto & m : models.mapping) {
+            if (m.second.meta.is_running()) {
+                n_running++;
+                if (models.stopping_models.count(m.first)) {
+                    n_stopping++;
                 }
             }
         }
-        SRV_INF("model name=%s went idle, giving up its slot to a queued request\n", model_id.c_str());
-        models.unload(model_id);
+        int n_needed  = 0;
+        int n_claimed = 0; // claimed the slot, but load() has not spawned yet
+        for (const auto & e : queue) {
+            if (!e.loading) {
+                n_needed++;
+                continue;
+            }
+            auto it = models.mapping.find(e.model_id);
+            if (it != models.mapping.end() && !it->second.meta.is_running()) {
+                n_claimed++;
+            }
+        }
+        int n_free = models.base_params.models_max - n_running + n_stopping - n_claimed;
+        while (n_free < n_needed) {
+            std::string victim = pick_victim(lk);
+            if (victim.empty()) {
+                return; // all remaining models are busy, wait for a request to end
+            }
+            SRV_INF("evicting idle LRU name=%s for a queued request\n", victim.c_str());
+            models.request_stop(victim);
+            n_free++;
+        }
     }
 
   private:
     struct entry_t {
         std::string model_id;
-        int  n_waiters;    // requests waiting for this model
-        bool slot_pending; // a model is already being evicted for this entry
-        bool loading;      // one of the waiters is doing the load right now
-
-        // a slot is already coming, or already taken by the load in flight
-        bool needs_slot() const { return !slot_pending && !loading; }
+        int  n_waiters; // requests waiting for this model
+        bool loading;   // one of the waiters is doing the load right now
     };
 
     entry_t * find(const std::string & model_id) {
@@ -946,7 +929,7 @@ void server_models::unload_lru() {
         if (sched->has_capacity(lk)) {
             return;
         }
-        lru_model_name = sched->pick_victim(lk, "");
+        lru_model_name = sched->pick_victim(lk);
     }
     if (!lru_model_name.empty()) {
         SRV_INF("models_max limit reached, removing LRU name=%s\n", lru_model_name.c_str());
@@ -1169,6 +1152,11 @@ void server_models::load(const std::string & name, const load_options & opts) {
     cv.notify_all();
 }
 
+void server_models::request_stop(const std::string & name) {
+    stopping_models.insert(name);
+    cv_stop.notify_all();
+}
+
 void server_models::unload(const std::string & name) {
     std::unique_lock<std::mutex> lk(mutex);
     auto it = mapping.find(name);
@@ -1182,13 +1170,12 @@ void server_models::unload(const std::string & name) {
             });
         } else if (it->second.meta.is_running()) {
             SRV_INF("stopping model instance name=%s\n", name.c_str());
-            stopping_models.insert(name);
             if (it->second.meta.status == SERVER_MODEL_STATUS_LOADING) {
                 // special case: if model is in loading state, unloading means force-killing it
                 SRV_WRN("model name=%s is still loading, force-killing\n", name.c_str());
                 it->second.subproc->terminate();
             }
-            cv_stop.notify_all();
+            request_stop(name);
             // status change will be handled by the managing thread
         } else {
             SRV_WRN("model instance name=%s is not running\n", name.c_str());
@@ -1206,8 +1193,7 @@ void server_models::unload_all() {
                 inst.subproc->stopped.store(true, std::memory_order_relaxed);
             } else if (inst.meta.is_running()) {
                 SRV_INF("stopping model instance name=%s\n", name.c_str());
-                stopping_models.insert(name);
-                cv_stop.notify_all();
+                request_stop(name);
                 // status change will be handled by the managing thread
             }
             // moving the thread to join list to avoid deadlock
@@ -1234,6 +1220,8 @@ void server_models::update_status(const std::string & name, const update_status_
         if (!args.progress.is_null()) {
             meta.progress = args.progress;
         }
+        // a model that comes up idle or goes down changes the slot count for queued requests
+        sched->tick(lk);
     }
     // broadcast status change to SSE
     {
@@ -1332,14 +1320,22 @@ bool server_models::remove(const std::string & name) {
         return true;
     }
 
-    // join before erasing - thread no longer acquires this mutex
-    if (it->second.th.joinable()) {
-        it->second.th.join();
+    // on the cancelled-download path the status flips to DOWNLOADED while the
+    // monitoring thread still has a mutex-guarded step left, so joining under
+    // the lock would deadlock - join outside, as load_models() does
+    std::thread th = std::move(it->second.th);
+    mapping.erase(name);
+    lk.unlock();
+
+    // join first so the monitoring thread's final mutex-guarded cleanup cannot
+    // race the disk removal, then remove from disk without holding the lock
+    // (best-effort: cancelled downloads may have no cached files)
+    if (th.joinable()) {
+        th.join();
     }
 
-    // remove from disk (best-effort: cancelled downloads may have no cached files)
     bool ok = common_download_remove(name);
-    mapping.erase(name);
+
     if (!ok) {
         SRV_WRN("removing model name=%s from disk returned false (no cached files?)\n", name.c_str());
     }
@@ -1380,13 +1376,11 @@ bool server_models::ensure_model_ready(const std::string & name, const std::func
 
     bool queued   = false;
     bool did_load = false;
-    std::string victim;
     {
         std::unique_lock<std::mutex> lk(mutex);
         auto it = mapping.find(name);
         if (it != mapping.end() && it->second.meta.status == SERVER_MODEL_STATUS_UNLOADED) {
-            bool has_capacity = sched->has_capacity(lk);
-            if (has_capacity && sched->queue_empty(lk)) {
+            if (sched->has_capacity(lk) && sched->queue_empty(lk)) {
                 lk.unlock();
                 SRV_INF("model name=%s is not loaded, loading...\n", name.c_str());
                 load(name);
@@ -1394,21 +1388,11 @@ bool server_models::ensure_model_ready(const std::string & name, const std::func
             } else {
                 // also queue when a slot looks free but others wait already, else they starve
                 sched->join(lk, name);
+                sched->tick(lk);
                 queued = true;
-                if (!has_capacity) {
-                    // an idle model may sit here right now, do not wait for a request to end
-                    victim = sched->pick_victim(lk, name);
-                    if (!victim.empty()) {
-                        sched->mark_slot_pending(lk, name);
-                    }
-                }
             }
         }
     }
-    if (!victim.empty()) {
-        SRV_INF("evicting idle LRU name=%s to make room for name=%s\n", victim.c_str(), name.c_str());
-        unload(victim);
-    }
 
     // while queued, this is also where the load happens: the head of the queue does it
     SRV_INF("waiting until model name=%s is fully loaded...\n", name.c_str());
@@ -1470,9 +1454,7 @@ bool server_models::ensure_model_ready(const std::string & name, const std::func
                 }
                 lk.lock();
                 sched->claim_done(lk, name, ok);
-                if (ok) {
-                    queued = false; // entry is gone, the other waiters watch the status now
-                }
+                sched->tick(lk);
                 continue;
             }
 
@@ -1480,6 +1462,7 @@ bool server_models::ensure_model_ready(const std::string & name, const std::func
         }
     } catch (...) {
         leave_queue();
+        sched->tick(lk); // a slot freed for this waiter goes to the next one
         throw;
     }
     leave_queue();
@@ -1529,18 +1512,14 @@ server_http_res_ptr server_models::proxy_request(const server_http_req & req, co
             );
 
     proxy->cleanup = [this, name]() {
-        bool went_idle = false;
-        {
-            std::unique_lock<std::mutex> lk(mutex);
-            auto it = mapping.find(name);
-            if (it != mapping.end() && it->second.req_count > 0) {
-                it->second.req_count--;
-                went_idle = it->second.req_count == 0;
+        std::unique_lock<std::mutex> lk(mutex);
+        auto it = mapping.find(name);
+        if (it != mapping.end() && it->second.req_count > 0) {
+            it->second.req_count--;
+            if (it->second.req_count == 0) {
+                sched->tick(lk);
             }
         }
-        if (went_idle) {
-            sched->on_model_idle(name);
-        }
     };
 
     return proxy;

tools/server/server-models.h

@@ -216,6 +216,10 @@ private:
     // not thread-safe, caller must hold mutex
     void add_model(server_model_meta && meta);
 
+    // ask the monitoring thread to stop a running instance
+    // not thread-safe, caller must hold mutex
+    void request_stop(const std::string & name);
+
     // notify SSE clients
     void notify_sse(const std::string & event, const std::string & model_id, const json & data = nullptr);
 

tools/server/tests/unit/test_router.py

@@ -297,6 +297,26 @@ def test_router_queue_is_fifo():
     assert first.done_at < second.done_at, "queue was not served in arrival order"
 
 
+def test_router_queue_two_waiters_share_one_eviction():
+    """two requests that both find the same idle model must both be served in the end"""
+    global server
+    server.models_max = 1
+    server.start()
+
+    _load_model_and_wait(MODEL_A, timeout=120)
+
+    # both arrive while MODEL_A is idle, so both want its slot; only one eviction can happen
+    first = _Bg(lambda: _tokenize(MODEL_B)).start()
+    second = _Bg(lambda: _tokenize(MODEL_C)).start()
+
+    first.join(90)
+    second.join(90)
+
+    first.assert_ok("first queued request")
+    second.assert_ok("second queued request")
+    assert _get_model_status(MODEL_A) == "unloaded"
+
+
 def test_router_no_models_autoload():
     global server
     server.no_models_autoload = True

tools/ui/package-lock.json

@@ -71,7 +71,7 @@
 				"svelte-check": "4.6.0",
 				"svelte-sonner": "1.1.1",
 				"tailwind-merge": "3.6.0",
-				"tailwind-variants": "3.2.2",
+				"tailwind-variants": "3.3.1",
 				"tailwindcss": "4.3.0",
 				"tw-animate-css": "1.4.0",
 				"typescript": "5.9.3",
@@ -16048,13 +16048,13 @@
 			}
 		},
 		"node_modules/tailwind-variants": {
-			"version": "3.2.2",
-			"resolved": "https://registry.npmjs.org/tailwind-variants/-/tailwind-variants-3.2.2.tgz",
-			"integrity": "sha512-Mi4kHeMTLvKlM98XPnK+7HoBPmf4gygdFmqQPaDivc3DpYS6aIY6KiG/PgThrGvii5YZJqRsPz0aPyhoFzmZgg==",
+			"version": "3.3.1",
+			"resolved": "https://registry.npmjs.org/tailwind-variants/-/tailwind-variants-3.3.1.tgz",
+			"integrity": "sha512-4pAvwUtM4HKBiRZftncAbpn6V9Hhwoa5Fl7O2u5zbp7Z5Cvu+/o/6+176WY3WCEES209543quG8zFIcXCsc5Jw==",
 			"dev": true,
 			"license": "MIT",
 			"engines": {
-				"node": ">=16.x",
+				"node": ">=16.9.x",
 				"pnpm": ">=7.x"
 			},
 			"peerDependencies": {
@@ -16064,6 +16064,9 @@
 			"peerDependenciesMeta": {
 				"tailwind-merge": {
 					"optional": true
+				},
+				"tailwindcss": {
+					"optional": true
 				}
 			}
 		},

tools/ui/package.json

@@ -13,6 +13,7 @@
 		"check:watch": "svelte-kit sync && svelte-check --tsconfig ./tsconfig.json --watch",
 		"reset": "rm -rf .svelte-kit node_modules",
 		"format": "eslint --fix . && prettier --write .",
+		"format:files": "sh -c 'eslint --fix \"$@\" && prettier --write \"$@\"' sh",
 		"lint": "prettier --check . && eslint .",
 		"test": "npm run test:ui -- --run && npm run test:client -- --run && npm run test:unit -- --run && npm run test:e2e",
 		"test:e2e": "playwright test",
@@ -90,7 +91,7 @@
 		"svelte-check": "4.6.0",
 		"svelte-sonner": "1.1.1",
 		"tailwind-merge": "3.6.0",
-		"tailwind-variants": "3.2.2",
+		"tailwind-variants": "3.3.1",
 		"tailwindcss": "4.3.0",
 		"tw-animate-css": "1.4.0",
 		"typescript": "5.9.3",

tools/ui/src/app.d.ts

@@ -18,20 +18,16 @@ import type {
 	ApiErrorResponse,
 	ApiLlamaCppServerProps,
 	ApiModelDataEntry,
-	ApiModelListResponse,
 	ApiModelLoadStage,
+	ApiModelsDownloadRequest,
+	ApiModelsDownloadResponse,
+	ApiModelsListResponse,
+	ApiModelsLoadResponse,
 	ApiModelsSseData,
 	ApiModelsSseEvent,
 	ApiModelsSseProgress,
+	ApiModelsUnloadResponse,
 	ApiProcessingState,
-	ApiRouterModelMeta,
-	ApiRouterModelsListResponse,
-	ApiRouterModelsLoadRequest,
-	ApiRouterModelsLoadResponse,
-	ApiRouterModelsStatusRequest,
-	ApiRouterModelsStatusResponse,
-	ApiRouterModelsUnloadRequest,
-	ApiRouterModelsUnloadResponse,
 	ChatAttachmentDisplayItem,
 	// Chat types
 	ChatMessagePromptProgress,
@@ -86,19 +82,17 @@ declare global {
 		ApiLlamaCppServerProps,
 		ApiModelDataEntry,
 		ApiModelLoadStage,
+		ModelDownloadProgress,
 		ApiModelsSseProgress,
 		ApiModelsSseData,
+		ApiModelsSseDownloadProgressData,
 		ApiModelsSseEvent,
-		ApiModelListResponse,
+		ApiModelsListResponse,
+		ApiModelsLoadResponse,
+		ApiModelsDownloadRequest,
+		ApiModelsDownloadResponse,
+		ApiModelsUnloadResponse,
 		ApiProcessingState,
-		ApiRouterModelMeta,
-		ApiRouterModelsLoadRequest,
-		ApiRouterModelsLoadResponse,
-		ApiRouterModelsStatusRequest,
-		ApiRouterModelsStatusResponse,
-		ApiRouterModelsListResponse,
-		ApiRouterModelsUnloadRequest,
-		ApiRouterModelsUnloadResponse,
 		// Chat types
 		ChatAttachmentDisplayItem,
 		ChatMessagePromptProgress,

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

@@ -220,8 +220,7 @@ export { default as ChatFormActionModels } from './ChatForm/ChatFormActions/Chat
  */
 export { default as ChatFormActionAddToolsSubmenu } from './ChatForm/ChatFormActions/ChatFormActionAdd/ChatFormActionAddToolsSubmenu.svelte';
 
-/**
- * Dropdown submenu for selecting reasoning effort level.
+/** Dropdown submenu for selecting reasoning effort level.
  *
  * Shows a "Reasoning" sub-menu item with a lightbulb icon indicating
  * thinking status, and a nested list of effort levels.

tools/ui/src/lib/components/app/content/MarkdownContent/MarkdownContent.svelte

@@ -7,6 +7,7 @@
 		getMdastNodeHash,
 		isAppendMode
 	} from './markdown-utils';
+	import { SAFE_HTML_CONFIG } from './safe-html-config';
 	import {
 		ActionIconCopyToClipboard,
 		CodeBlockActions,
@@ -44,6 +45,7 @@
 	import { detectIncompleteCodeBlock, highlightCode, type IncompleteCodeBlock } from '$lib/utils';
 	import { sanitizeSvg } from '$lib/utils/sanitize-svg';
 	import { mountSvgShadow } from '$lib/utils/svg-shadow';
+	import DOMPurify from 'dompurify';
 	import type { Root as HastRoot, RootContent as HastRootContent } from 'hast';
 	import githubLightCss from 'highlight.js/styles/github.css?inline';
 	import githubDarkCss from 'highlight.js/styles/github-dark.css?inline';
@@ -57,6 +59,8 @@
 		content: string;
 		class?: string;
 		disableMath?: boolean;
+		/** Render raw HTML found in the markdown (sanitized) instead of escaping it. */
+		allowHtml?: boolean;
 	}
 
 	interface MarkdownBlock {
@@ -65,7 +69,13 @@
 		contentHash?: string;
 	}
 
-	let { attachments, class: className = '', content, disableMath = false }: Props = $props();
+	let {
+		allowHtml = false,
+		attachments,
+		class: className = '',
+		content,
+		disableMath = false
+	}: Props = $props();
 
 	let containerRef = $state<HTMLDivElement>();
 	let renderedBlocks = $state<MarkdownBlock[]>([]);
@@ -119,6 +129,11 @@
 	let pendingMarkdown: string | null = null;
 	let isProcessing = false;
 
+	// Raw HTML in model cards renders after sanitization with an explicit allow
+	// list: no scripts, event handlers, forms, iframes or style blocks pass.
+	// Covers the tags model cards use plus the KaTeX output (spans with inline
+	// styles and MathML).
+
 	// Per-instance transform cache, avoids re-transforming stable blocks during streaming
 	// Garbage collected when component is destroyed (on conversation change)
 	const transformCache = new SvelteMap<string, string>();
@@ -183,7 +198,10 @@
 		index: number
 	): Promise<{ html: string; hash: string }> {
 		const hash = getMdastNodeHash(node, index);
-		const cached = transformCache.get(hash);
+		// the rendered HTML also depends on allowHtml (sanitized raw vs escaped),
+		// so the cache is keyed per mode
+		const cacheKey = `${allowHtml ? 'a' : 'p'}:${hash}`;
+		const cached = transformCache.get(cacheKey);
 
 		if (cached) {
 			return { hash, html: cached };
@@ -192,10 +210,13 @@
 		const singleNodeRoot = { children: [node], type: 'root' };
 		const transformedRoot = (await processorInstance.run(singleNodeRoot as MdastRoot)) as HastRoot;
 		const html = processorInstance.stringify(transformedRoot);
+		const safeHtml = allowHtml
+			? (DOMPurify.sanitize(html, SAFE_HTML_CONFIG) as unknown as string)
+			: html;
 
-		transformCache.set(hash, html);
+		transformCache.set(cacheKey, safeHtml);
 
-		return { hash, html };
+		return { hash, html: safeHtml };
 	}
 
 	/**
@@ -300,7 +321,7 @@
 
 			if (prefixMarkdown.trim()) {
 				const normalizedPrefix = preprocessLaTeX(prefixMarkdown);
-				const processorInstance = getMarkdownProcessor({ attachments, disableMath });
+				const processorInstance = getMarkdownProcessor({ allowHtml, attachments, disableMath });
 				const ast = processorInstance.parse(normalizedPrefix) as MdastRoot;
 				const mdastChildren = (ast as { children?: unknown[] }).children ?? [];
 				const nextBlocks: MarkdownBlock[] = [];
@@ -350,7 +371,7 @@
 		incompleteCodeBlock = null;
 
 		const normalized = preprocessLaTeX(markdown);
-		const processorInstance = getMarkdownProcessor({ attachments, disableMath });
+		const processorInstance = getMarkdownProcessor({ allowHtml, attachments, disableMath });
 		const ast = processorInstance.parse(normalized) as MdastRoot;
 		const mdastChildren = (ast as { children?: unknown[] }).children ?? [];
 		const stableCount = Math.max(mdastChildren.length - 1, 0);
@@ -394,6 +415,10 @@
 			)) as HastRoot;
 
 			unstableHtml = processorInstance.stringify(transformedRoot);
+
+			if (allowHtml) {
+				unstableHtml = DOMPurify.sanitize(unstableHtml, SAFE_HTML_CONFIG) as unknown as string;
+			}
 		}
 
 		renderedBlocks = nextBlocks;

tools/ui/src/lib/components/app/content/MarkdownContent/markdown-processor.ts

@@ -41,12 +41,15 @@ export interface MarkdownProcessor {
 export interface MarkdownProcessorOptions {
 	attachments?: DatabaseMessageExtra[];
 	disableMath?: boolean;
+	/** Render raw HTML found in the markdown instead of escaping it. */
+	allowHtml?: boolean;
 }
 
 const sharedPipelines = new Map<string, MarkdownProcessor>();
-const attachmentPipelines = new WeakMap<object, MarkdownProcessor>();
+const attachmentPipelines = new WeakMap<object, Map<string, MarkdownProcessor>>();
 
 function buildPipeline({
+	allowHtml = false,
 	attachments,
 	disableMath = false
 }: MarkdownProcessorOptions): MarkdownProcessor {
@@ -57,11 +60,15 @@ function buildPipeline({
 		proc = proc.use(remarkMath); // Parse $inline$ and $$block$$ math
 	}
 
-	proc = proc
-		.use(remarkBreaks) // Convert line breaks to <br>
+	proc = proc.use(remarkBreaks); // Convert line breaks to <br>
+
+	if (!allowHtml) {
 		// Treat raw HTML as literal text with preserved indentation
-		.use(remarkLiteralHtml)
-		.use(remarkRehype); // Convert Markdown AST to rehype
+		proc = proc.use(remarkLiteralHtml);
+	}
+
+	// Convert Markdown AST to rehype. Keep raw HTML as-is when allowHtml is set.
+	proc = proc.use(remarkRehype, allowHtml ? { allowDangerousHtml: true } : undefined);
 
 	if (!disableMath) {
 		proc = proc.use(rehypeKatex); // Render math using KaTeX
@@ -89,17 +96,26 @@ function buildPipeline({
 
 export function getMarkdownProcessor(options: MarkdownProcessorOptions): MarkdownProcessor {
 	if (options.attachments && options.attachments.length > 0) {
-		let cached = attachmentPipelines.get(options.attachments);
+		let byOptions = attachmentPipelines.get(options.attachments);
+
+		if (!byOptions) {
+			byOptions = new Map<string, MarkdownProcessor>();
+			attachmentPipelines.set(options.attachments, byOptions);
+		}
+
+		const key = `${Boolean(options.disableMath)}:${Boolean(options.allowHtml)}`;
+
+		let cached = byOptions.get(key);
 
 		if (!cached) {
 			cached = buildPipeline(options);
-			attachmentPipelines.set(options.attachments, cached);
+			byOptions.set(key, cached);
 		}
 
 		return cached;
 	}
 
-	const key = String(Boolean(options.disableMath));
+	const key = `${Boolean(options.disableMath)}:${Boolean(options.allowHtml)}`;
 
 	let cached = sharedPipelines.get(key);
 

tools/ui/src/lib/components/app/content/MarkdownContent/safe-html-config.ts

@@ -0,0 +1,121 @@
+// Sanitization rules for raw HTML in model cards (MarkdownContent's
+// allowHtml mode). Kept as its own module so the adversarial test in
+// tests/client can import the exact shipped configuration.
+
+/**
+ * Explicit DOMPurify allow list: the tags and attributes model cards may use
+ * plus the KaTeX output (spans with inline styles and MathML). Everything
+ * else - scripts, event handlers, forms, iframes, style tags - is stripped.
+ */
+export const SAFE_HTML_CONFIG = {
+		ALLOWED_ATTR: [
+			'class',
+			'style',
+			'id',
+			'title',
+			'role',
+			'aria-hidden',
+			'aria-label',
+			'aria-describedby',
+			'aria-expanded',
+			'href',
+			'target',
+			'rel',
+			'src',
+			'srcset',
+			'alt',
+			'width',
+			'height',
+			'align',
+			'valign',
+			'colspan',
+			'rowspan',
+			'controls',
+			'type',
+			// MathML
+			'xmlns',
+			'display',
+			'encoding',
+			'mathvariant'
+		],
+		ALLOWED_TAGS: [
+			// text and structure
+			'a',
+			'b',
+			'i',
+			'u',
+			's',
+			'em',
+			'strong',
+			'code',
+			'pre',
+			'kbd',
+			'sub',
+			'sup',
+			'br',
+			'hr',
+			'p',
+			'span',
+			'div',
+			'blockquote',
+			'center',
+			'ul',
+			'ol',
+			'li',
+			'dl',
+			'dt',
+			'dd',
+			'h1',
+			'h2',
+			'h3',
+			'h4',
+			'h5',
+			'h6',
+			'table',
+			'thead',
+			'tbody',
+			'tr',
+			'th',
+			'td',
+			'colgroup',
+			'col',
+			'caption',
+			// media
+			'img',
+			'picture',
+			'source',
+			'figure',
+			'figcaption',
+			'video',
+			'audio',
+			// MathML, for the KaTeX output
+			'math',
+			'mrow',
+			'mi',
+			'mo',
+			'mn',
+			'ms',
+			'mtext',
+			'mfrac',
+			'mroot',
+			'msqrt',
+			'msub',
+			'msup',
+			'msubsup',
+			'munder',
+			'mover',
+			'mmultiscripts',
+			'mtable',
+			'mtr',
+			'mtd',
+			'mspace',
+			'mglyph',
+			'maligngroup',
+			'malignmark',
+			'mpadded',
+			'mphantom',
+			'mstyle',
+			'semantics',
+			'annotation'
+		]
+	};

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

@@ -0,0 +1,60 @@
+<script lang="ts">
+	import DialogConfirmation from '$lib/components/app/dialogs/DialogConfirmation.svelte';
+	import { ModelDownloadConfirmAction } from '$lib/enums';
+	import { modelsStore } from '$lib/stores';
+
+	interface Props {
+		/** Action being confirmed; drives the wording. */
+		action: ModelDownloadConfirmAction;
+		/** `<repo>:<tag>` the action targets. */
+		repoWithTag: string;
+		onClose: () => void;
+		/** Overrides the default store removal; defaults to removing the entry. */
+		onConfirm?: (repoWithTag: string) => void;
+		open?: boolean;
+	}
+
+	let { action, onClose, onConfirm, open = true, repoWithTag }: Props = $props();
+
+	// Both actions resolve through the same store removal (cancelDownload drops a
+	// running download's partial files or a cached model's files); only the copy
+	// differs. One component so the discover chips and the selector rows word the
+	// destructive confirmations identically.
+	const COPY = {
+		[ModelDownloadConfirmAction.CANCEL]: {
+			cancelText: 'Keep downloading',
+			confirmText: 'Cancel download',
+			description: (name: string) =>
+				`This stops the download of ${name} and removes the partial files. Pause it instead to keep the progress.`,
+			title: 'Cancel download'
+		},
+		[ModelDownloadConfirmAction.DELETE]: {
+			cancelText: 'Keep model',
+			confirmText: 'Delete',
+			description: (name: string) =>
+				`This permanently removes ${name} from disk. You can download it again later.`,
+			title: 'Delete model'
+		}
+	} as const;
+
+	let copy = $derived(COPY[action]);
+	let displayName = $derived(modelsStore.toDisplayName(repoWithTag));
+
+	function confirm() {
+		if (onConfirm) onConfirm(repoWithTag);
+		else void modelsStore.status.cancelDownload(repoWithTag);
+
+		onClose();
+	}
+</script>
+
+<DialogConfirmation
+	cancelText={copy.cancelText}
+	confirmText={copy.confirmText}
+	description={copy.description(displayName)}
+	onCancel={onClose}
+	onConfirm={confirm}
+	{open}
+	title={copy.title}
+	variant="destructive"
+/>

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

@@ -0,0 +1,25 @@
+<script lang="ts">
+	import { ModelsDiscover } from '$lib/components/app/models/discover';
+	import * as Dialog from '$lib/components/ui/dialog';
+
+	interface Props {
+		open?: boolean;
+		onOpenChange?: (open: boolean) => void;
+	}
+
+	let { onOpenChange, open = $bindable(false) }: Props = $props();
+
+	function handleOpenChange(value: boolean) {
+		open = value;
+		onOpenChange?.(value);
+	}
+</script>
+
+<Dialog.Root onOpenChange={handleOpenChange} {open}>
+	<Dialog.Content
+		class="grid gap-0 p-0 md:h-[calc(100vh-4rem)]! md:max-h-240! md:w-[calc(100vw-4rem)]! md:max-w-380!"
+		style="grid-template-columns: auto 1fr;"
+	>
+		<ModelsDiscover />
+	</Dialog.Content>
+</Dialog.Root>

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

@@ -108,6 +108,17 @@ export { default as DialogExportSettings } from './DialogExportSettings.svelte';
  */
 export { default as DialogConfirmation } from './DialogConfirmation.svelte';
 
+/**
+ * **DialogConfirmDownload** - Confirm a destructive download action
+ *
+ * Shared confirmation for stopping/cancelling an in-flight download or deleting
+ * a downloaded model, used by the discover quant chips and the model selector's
+ * download rows so both word the action identically. Owns the copy and the
+ * default store removal; render one instance per surface keyed by the acted-on
+ * repo:tag.
+ */
+export { default as DialogConfirmDownload } from './DialogConfirmDownload.svelte';
+
 /**
  * **DialogConversationRename** - Rename a conversation
  *
@@ -526,3 +537,13 @@ export { default as DialogMcpResourcePreview } from './DialogMcpResourcePreview.
  * ```
  */
 export { default as DialogMermaidPreview } from './DialogMermaidPreview.svelte';
+
+/**
+ * **DialogModelsDiscover** - full-screen model discovery dialog.
+ *
+ * Two-pane layout: searchable model list (Hugging Face + llama.app catalog)
+ * on the left, model details with download options on the right.
+ *
+ * @see ModelsDiscover in $lib/components/app/models/discover
+ */
+export { default as DialogModelsDiscover } from './DialogModelsDiscover.svelte';

tools/ui/src/lib/components/app/models/ModelCapabilityIcons.svelte

@@ -0,0 +1,93 @@
+<script lang="ts">
+	import { Image, Lightbulb, Mic, Video, Wrench } from '@lucide/svelte';
+	import * as Tooltip from '$lib/components/ui/tooltip';
+	import type { ModelModalities } from '$lib/types/models';
+
+	interface Props {
+		modalities?: ModelModalities;
+		supportsThinking?: boolean;
+		supportsToolUse?: boolean;
+		hideCapabilities?: boolean;
+		hideModalities?: boolean;
+		iconSize?: string;
+		gapClass?: string;
+	}
+
+	let {
+		gapClass = 'gap-1.25',
+		hideCapabilities = false,
+		hideModalities = false,
+		iconSize = 'h-3 w-3',
+		modalities,
+		supportsThinking = false,
+		supportsToolUse = false
+	}: Props = $props();
+
+	let hasModalityIcons = $derived(modalities?.vision || modalities?.video || modalities?.audio);
+</script>
+
+<span class="inline-flex items-center {gapClass}">
+	{#if supportsToolUse && !hideCapabilities}
+		<Tooltip.Root>
+			<Tooltip.Trigger>
+				<Wrench class="{iconSize} text-muted-foreground" />
+			</Tooltip.Trigger>
+
+			<Tooltip.Content>
+				<p>Tool use</p>
+			</Tooltip.Content>
+		</Tooltip.Root>
+	{/if}
+
+	{#if supportsThinking && !hideCapabilities}
+		<Tooltip.Root>
+			<Tooltip.Trigger>
+				<Lightbulb class="{iconSize} text-muted-foreground" />
+			</Tooltip.Trigger>
+
+			<Tooltip.Content>
+				<p>Reasoning</p>
+			</Tooltip.Content>
+		</Tooltip.Root>
+	{/if}
+
+	{#if hasModalityIcons && !hideModalities}
+		<span class="inline-flex items-center text-muted-foreground">
+			{#if modalities?.vision}
+				<Tooltip.Root>
+					<Tooltip.Trigger>
+						<Image class={iconSize} />
+					</Tooltip.Trigger>
+
+					<Tooltip.Content>
+						<p>Vision</p>
+					</Tooltip.Content>
+				</Tooltip.Root>
+			{/if}
+
+			{#if modalities?.video}
+				<Tooltip.Root>
+					<Tooltip.Trigger>
+						<Video class={iconSize} />
+					</Tooltip.Trigger>
+
+					<Tooltip.Content>
+						<p>Video</p>
+					</Tooltip.Content>
+				</Tooltip.Root>
+			{/if}
+
+			{#if modalities?.audio}
+				<Tooltip.Root>
+					<Tooltip.Trigger>
+						<Mic class={iconSize} />
+					</Tooltip.Trigger>
+
+					<Tooltip.Content>
+						<p>Audio</p>
+					</Tooltip.Content>
+				</Tooltip.Root>
+			{/if}
+		</span>
+	{/if}
+</span>

tools/ui/src/lib/components/app/models/ModelId.svelte

@@ -1,45 +1,66 @@
 <script lang="ts">
+	import ModelCapabilityIcons from './ModelCapabilityIcons.svelte';
+	import { Database, ScrollText } from '@lucide/svelte';
 	import { TruncatedText } from '$lib/components/app';
 	import * as Tooltip from '$lib/components/ui/tooltip';
-	import {
-		CAPABILITY_FLAG_KEYS,
-		CAPABILITY_ICONS,
-		CAPABILITY_LABELS,
-		MODALITY_FLAG_KEYS,
-		MODALITY_ICONS,
-		MODALITY_LABELS
-	} from '$lib/constants';
-	import { ModelCapability, ModelModality } from '$lib/enums';
+	import { type ModelSidecar } from '$lib/constants';
+	import { HuggingFaceService } from '$lib/services';
 	import { ModelsService } from '$lib/services/models.service';
 	import { settingsStore } from '$lib/stores';
-	import type { ModelCapabilities, ModelModalities } from '$lib/types/models';
+	import type { ModelModalities } from '$lib/types/models';
+	import { isAuxSidecar } from '$lib/utils';
+	import { formatParameters } from '$lib/utils';
 
 	interface Props {
 		modelId: string;
 		hideOrgName?: boolean;
+		hideName?: boolean;
+		hideModalities?: boolean;
+		hideCapabilities?: boolean;
+		hideParameters?: boolean;
 		showRaw?: boolean;
 		showRawTooltip?: boolean;
 		hideQuantization?: boolean;
 		hideTags?: boolean;
 		aliases?: string[];
 		tags?: string[];
+		/** Render the capability/modality/context icons on a second row. */
+		iconsOnNewLine?: boolean;
 		modalities?: ModelModalities;
-		capabilities?: ModelCapabilities;
+		supportsThinking?: boolean;
+		supportsToolUse?: boolean;
+		/** Context length in tokens; renders a context icon when set. */
+		contextLength?: number;
+		/** Min/max GGUF file size (main + draft) across quants; renders a range when set. */
+		sizeRange?: { min: number; max: number } | null;
+		draftSidecars?: ModelSidecar[];
+		/** Allow badges to wrap onto new lines instead of truncating. */
+		wrap?: boolean;
 		class?: string;
 	}
 
 	let {
 		aliases,
-		capabilities,
 		class: className = '',
+		contextLength,
+		draftSidecars = [],
+		hideCapabilities = false,
+		hideModalities = false,
+		hideName = false,
 		hideOrgName = false,
+		hideParameters = false,
 		hideQuantization,
 		hideTags,
+		iconsOnNewLine = false,
 		modalities,
 		modelId,
 		showRaw = undefined,
 		showRawTooltip = false,
+		sizeRange,
+		supportsThinking = false,
+		supportsToolUse = false,
 		tags,
+		wrap = false,
 		...rest
 	}: Props = $props();
 
@@ -47,6 +68,11 @@
 		'inline-flex w-fit shrink-0 items-center justify-center whitespace-nowrap rounded-md border border-border/50 px-1 py-0 text-[10px] font-mono bg-foreground/15 dark:bg-foreground/10 text-foreground [a&]:hover:bg-foreground/25';
 	const tagBadgeClass =
 		'inline-flex w-fit shrink-0 items-center justify-center whitespace-nowrap rounded-md border border-border/50 px-1 py-0 text-[10px] font-mono text-foreground [a&]:hover:bg-accent [a&]:hover:text-accent-foreground';
+	const variantBadgeClass =
+		'inline-flex w-fit shrink-0 items-center justify-center whitespace-nowrap rounded-md bg-primary px-1.5 py-0 text-[10px] font-mono font-semibold uppercase tracking-wide text-primary-foreground';
+
+	/** Alias badges beyond this many collapse into a single `+x more` badge. */
+	const MAX_ALIAS_BADGES = 2;
 
 	let parsed = $derived(ModelsService.parseModelId(modelId));
 	let resolvedShowRaw = $derived(
@@ -59,104 +85,147 @@
 
 	let uniqueAliases = $derived([...new Set(aliases ?? [])]);
 	let uniqueTags = $derived([...new Set([...(parsed.tags ?? []), ...(tags ?? [])])]);
-
-	const allModalities = [ModelModality.VISION, ModelModality.VIDEO, ModelModality.AUDIO] as const;
-	const allCapabilities: ModelCapability[] = [ModelCapability.REASONING];
-
-	let activeModalities = $derived(
-		allModalities.filter((modality) => modalities?.[MODALITY_FLAG_KEYS[modality]])
-	);
-	let activeCapabilities = $derived(
-		allCapabilities.filter((capability) => capabilities?.[CAPABILITY_FLAG_KEYS[capability]])
-	);
+	let uniqueDraftSidecars = $derived([...new Set(draftSidecars)].filter((s) => !isAuxSidecar(s)));
 
 	let primaryAlias = $derived(uniqueAliases.length === 1 ? uniqueAliases[0] : null);
 	let displayName = $derived(primaryAlias ?? parsed.modelName ?? modelId);
+
+	let hasBadges = $derived(
+		parsed.sidecar ||
+			(parsed.params && !hideParameters) ||
+			(parsed.quantization && !resolvedHideQuantization) ||
+			primaryAlias ||
+			uniqueAliases.length > 1 ||
+			(uniqueTags.length > 0 && !resolvedHideTags)
+	);
 </script>
 
 {#if resolvedShowRaw}
 	<TruncatedText class="font-medium {className}" showTooltip={false} text={modelId} {...rest} />
 {:else}
 	{#snippet nameAndBadges()}
-		<span class="min-w-0 truncate font-medium">
-			{#if !hideOrgName && parsed.orgName}{parsed.orgName}/{/if}{displayName}
-		</span>
-
-		<span class="inline-flex items-center gap-1">
-			{#if parsed.params}
-				<span class={badgeClass}>
-					{parsed.params}{parsed.activatedParams ? `-${parsed.activatedParams}` : ''}
-				</span>
-			{/if}
-
-			{#if parsed.quantization && !resolvedHideQuantization}
-				<span class={badgeClass}>
-					{parsed.quantization}
-				</span>
-			{/if}
-
-			{#if primaryAlias}
-				{#if primaryAlias !== parsed.modelName}
-					<span class={badgeClass}>{parsed.modelName ?? modelId}</span>
-				{/if}
-			{:else if uniqueAliases.length > 1}
-				{#each uniqueAliases as alias (alias)}
-					<span class={badgeClass}>{alias}</span>
-				{/each}
-			{/if}
-
-			{#if uniqueTags.length > 0 && !resolvedHideTags}
-				{#each uniqueTags as tag (tag)}
-					<span class={tagBadgeClass}>{tag}</span>
-				{/each}
-			{/if}
-		</span>
-	{/snippet}
-
-	<span class="flex min-w-0 items-center gap-1.5 {className}" {...rest}>
-		{#if showRawTooltip}
-			<Tooltip.Root>
-				<Tooltip.Trigger class="flex min-w-0 items-center gap-1.5">
-					{@render nameAndBadges()}
-				</Tooltip.Trigger>
-
-				<Tooltip.Content>
-					<p>{modelId}</p>
-				</Tooltip.Content>
-			</Tooltip.Root>
-		{:else}
-			{@render nameAndBadges()}
+		{#if !hideName}
+			<span class="min-w-0 truncate font-medium">
+				{#if !hideOrgName && parsed.orgName}{parsed.orgName}/{/if}{displayName}
+			</span>
 		{/if}
 
-		{#if activeCapabilities.length > 0 || activeModalities.length > 0}
-			<span class="inline-flex items-center gap-1.25 text-muted-foreground">
-				{#each activeCapabilities as capability (capability)}
-					{@const CapabilityIcon = CAPABILITY_ICONS[capability]}
+		{#if hasBadges}
+			<span class="inline-flex items-center gap-1 {wrap ? 'flex-wrap' : ''}">
+				{#if parsed.sidecar}
+					<span class={variantBadgeClass} title={`${parsed.sidecar.toUpperCase()} draft model`}>
+						{parsed.sidecar}
+					</span>
+				{/if}
 
-					<Tooltip.Root>
-						<Tooltip.Trigger>
-							<CapabilityIcon class="h-3 w-3 text-muted-foreground" />
-						</Tooltip.Trigger>
+				{#if parsed.params && !hideParameters}
+					<span class={badgeClass}>
+						{parsed.params}{parsed.activatedParams ? `-${parsed.activatedParams}` : ''}
+					</span>
+				{/if}
 
-						<Tooltip.Content>
-							<p>{CAPABILITY_LABELS[capability]}</p>
-						</Tooltip.Content>
-					</Tooltip.Root>
+				{#each uniqueDraftSidecars as sidecar (sidecar)}
+					<span class={variantBadgeClass} title={`${sidecar.toUpperCase()} draft model available`}>
+						{sidecar}
+					</span>
 				{/each}
 
-				{#each activeModalities as modality (modality)}
-					{@const ModalityIcon = MODALITY_ICONS[modality]}
+				{#if parsed.quantization && !resolvedHideQuantization}
+					<span class={badgeClass}>
+						{parsed.quantization}
+					</span>
+				{/if}
 
-					<Tooltip.Root>
-						<Tooltip.Trigger>
-							<ModalityIcon class="h-3 w-3 text-muted-foreground" />
-						</Tooltip.Trigger>
+				{#if primaryAlias}
+					{#if primaryAlias !== parsed.modelName}
+						<span class="{badgeClass} max-w-32 truncate" title={parsed.modelName ?? modelId}>
+							{parsed.modelName ?? modelId}
+						</span>
+					{/if}
+				{:else if uniqueAliases.length > 1}
+					{#each uniqueAliases.slice(0, MAX_ALIAS_BADGES) as alias (alias)}
+						<span class="{badgeClass} max-w-32 truncate" title={alias}>
+							{alias}
+						</span>
+					{/each}
 
-						<Tooltip.Content>
-							<p>{MODALITY_LABELS[modality]}</p>
-						</Tooltip.Content>
-					</Tooltip.Root>
-				{/each}
+					{#if uniqueAliases.length > MAX_ALIAS_BADGES}
+						<span class={badgeClass} title={uniqueAliases.slice(MAX_ALIAS_BADGES).join(', ')}>
+							+{uniqueAliases.length - MAX_ALIAS_BADGES} more
+						</span>
+					{/if}
+				{/if}
+
+				{#if uniqueTags.length > 0 && !resolvedHideTags}
+					{#each uniqueTags as tag (tag)}
+						<span class={tagBadgeClass}>{tag}</span>
+					{/each}
+				{/if}
+			</span>
+		{/if}
+	{/snippet}
+
+	<span
+		class="flex min-w-0 items-center gap-1.5 {wrap ? 'flex-wrap' : ''} {iconsOnNewLine
+			? 'flex-col items-start'
+			: ''} {className}"
+		{...rest}
+	>
+		<span class="flex min-w-0 items-center gap-1.5 {wrap ? 'flex-wrap' : ''}">
+			{#if showRawTooltip}
+				<Tooltip.Root>
+					<Tooltip.Trigger class="flex min-w-0 items-center gap-1.5">
+						{@render nameAndBadges()}
+					</Tooltip.Trigger>
+
+					<Tooltip.Content>
+						<p>{modelId}</p>
+					</Tooltip.Content>
+				</Tooltip.Root>
+			{:else}
+				{@render nameAndBadges()}
+			{/if}
+
+			{#if !iconsOnNewLine}
+				<ModelCapabilityIcons
+					{hideCapabilities}
+					{hideModalities}
+					{modalities}
+					{supportsThinking}
+					{supportsToolUse}
+				/>
+			{/if}
+		</span>
+
+		{#if iconsOnNewLine || contextLength || sizeRange}
+			<span class="inline-flex items-center gap-1.5">
+				{#if iconsOnNewLine}
+					<ModelCapabilityIcons
+						{hideCapabilities}
+						{hideModalities}
+						{modalities}
+						{supportsThinking}
+						{supportsToolUse}
+					/>
+				{/if}
+
+				{#if contextLength}
+					<span class="inline-flex items-center gap-1 text-muted-foreground">
+						<ScrollText class="h-3 w-3" />
+
+						<span class="text-xs">{formatParameters(contextLength)}</span>
+					</span>
+				{/if}
+
+				{#if sizeRange}
+					<span class="inline-flex items-center gap-1 text-muted-foreground">
+						<Database class="h-3 w-3" />
+
+						<span class="text-xs"
+							>{HuggingFaceService.formatSizeRange(sizeRange.min, sizeRange.max)}</span
+						>
+					</span>
+				{/if}
 			</span>
 		{/if}
 	</span>

tools/ui/src/lib/components/app/models/ModelsSelectorOption.svelte

@@ -12,7 +12,7 @@
 	} from '@lucide/svelte';
 	import { ActionIcon, ModelId } from '$lib/components/app';
 	import { ICON_CLASS_DEFAULT } from '$lib/constants';
-	import { ServerModelStatus } from '$lib/enums';
+	import { ModelCapability, ServerModelStatus } from '$lib/enums';
 	import { modelsStore } from '$lib/stores';
 	import type { ModelOption } from '$lib/types/models';
 	import { modelLoadFraction, modelLoadProgressText } from '$lib/utils';
@@ -60,7 +60,8 @@
 	let loadTitle = $derived(modelLoadProgressText(loadProgress));
 	let modalities = $derived(option.modalities);
 	let capabilities = $derived.by(() => ({
-		reasoning: modelsStore.props.checkModelSupportsThinking(option.model)
+		reasoning: modelsStore.props.checkModelSupportsThinking(option.model),
+		tools: option.capabilities.includes(ModelCapability.TOOL_USE)
 	}));
 </script>
 
@@ -85,12 +86,13 @@
 >
 	<ModelId
 		aliases={option.aliases}
-		{capabilities}
 		class="flex-1"
 		{hideOrgName}
 		{modalities}
 		modelId={option.model}
 		showRawTooltip
+		supportsThinking={capabilities.reasoning}
+		supportsToolUse={capabilities.tools}
 		tags={option.tags}
 	/>
 

tools/ui/src/lib/components/app/models/discover/ModelsDiscover.svelte

@@ -0,0 +1,138 @@
+<script lang="ts">
+	import {
+		ModelsDiscoverDetails,
+		ModelsDiscoverList,
+		ModelsDiscoverListSearch
+	} from '$lib/components/app/models/discover';
+	import { Button } from '$lib/components/ui/button';
+	import { HuggingFaceService } from '$lib/services';
+	import { modelsDiscoverStore } from '$lib/stores';
+	import type { HfModelDetailInfo, HfModelSibling } from '$lib/types';
+
+	let selectedId = $state<string | null>(null);
+	let searchQuery = $state('');
+	let searchTimeout: ReturnType<typeof setTimeout> | null = null;
+
+	// Detail pane state, reloaded when the selection changes.
+	let details = $state<HfModelDetailInfo | null>(null);
+	let files = $state<HfModelSibling[]>([]);
+	let readme = $state<string | null>(null);
+	let detailLoading = $state(false);
+	let detailError = $state<string | null>(null);
+
+	// Load the sidebar list on mount (the component is mounted when the dialog opens).
+	$effect(() => {
+		void modelsDiscoverStore.fetch();
+		void modelsDiscoverStore.search('');
+	});
+
+	// Auto-select the first model.
+	$effect(() => {
+		const first = modelsDiscoverStore.firstModel;
+
+		if (!selectedId && first) {
+			selectedId = first.id;
+		}
+	});
+
+	function handleSearchInput(value: string) {
+		searchQuery = value;
+
+		if (searchTimeout) clearTimeout(searchTimeout);
+
+		searchTimeout = setTimeout(() => {
+			void modelsDiscoverStore.search(value);
+		}, 300);
+	}
+
+	// Load the detail pane for the selected model (component is reused across
+	// selections, so this re-fetches on every change).
+	$effect(() => {
+		const id = selectedId;
+
+		if (!id) return;
+
+		let cancelled = false;
+
+		detailLoading = true;
+		detailError = null;
+
+		void (async () => {
+			try {
+				const [info, tree, readmeText] = await Promise.all([
+					HuggingFaceService.getDetails(id),
+					HuggingFaceService.getTree(id),
+					HuggingFaceService.getReadme(id)
+				]);
+
+				if (cancelled) return;
+
+				if (!info) {
+					detailError = 'Model not found';
+
+					return;
+				}
+
+				details = info;
+				files = HuggingFaceService.filterByExtension(
+					HuggingFaceService.collapseGgufShards(tree),
+					'.gguf'
+				);
+				readme = readmeText;
+			} catch (err) {
+				if (cancelled) return;
+
+				detailError = err instanceof Error ? err.message : 'Failed to load model';
+			} finally {
+				if (!cancelled) detailLoading = false;
+			}
+		})();
+
+		return () => {
+			cancelled = true;
+		};
+	});
+</script>
+
+<aside
+	class="w-md shrink-0 self-start border-r border-border/40 bg-background overflow-y-auto md:p-4 h-full space-y-1"
+>
+	<ModelsDiscoverListSearch bind:value={searchQuery} onSearch={handleSearchInput} />
+
+	<!-- One list instance, so the rows keep their state across search round trips;
+		 skeleton rows replace them while the initial catalog or a query loads. -->
+	<div>
+		{#if modelsDiscoverStore.error}
+			<div class="flex flex-col items-start gap-2 p-4">
+				<p class="text-sm text-destructive">{modelsDiscoverStore.error}</p>
+
+				<Button onclick={() => void modelsDiscoverStore.fetch()} size="sm" variant="outline">
+					Retry
+				</Button>
+			</div>
+		{:else if !modelsDiscoverStore.loading && !modelsDiscoverStore.searching && modelsDiscoverStore.models.length === 0}
+			<p class="p-4 text-sm text-muted-foreground">No models found</p>
+		{:else}
+			<ModelsDiscoverList
+				activeId={selectedId}
+				loading={modelsDiscoverStore.loading || modelsDiscoverStore.searching}
+				models={modelsDiscoverStore.models}
+				onSelect={(id) => (selectedId = id)}
+				showBaseModelAvatar
+			/>
+		{/if}
+	</div>
+</aside>
+
+<main class="overflow-y-auto">
+	{#if selectedId}
+		<ModelsDiscoverDetails
+			{details}
+			error={detailError}
+			{files}
+			loading={detailLoading}
+			modelId={selectedId}
+			{readme}
+		/>
+	{/if}
+</main>

tools/ui/src/lib/components/app/models/discover/ModelsDiscoverAvatar.svelte

@@ -0,0 +1,105 @@
+<script lang="ts">
+	import * as Tooltip from '$lib/components/ui/tooltip';
+	import { DARK_INVERT_AVATAR_ORGS } from '$lib/constants';
+	import { HuggingFaceService } from '$lib/services';
+
+	interface Props {
+		class?: string;
+		org: string;
+		quantOrg?: string;
+		size?: string;
+		baseImageClass?: string;
+		quantImageClass?: string;
+		quantPositionClass?: string;
+		quantSize?: string;
+	}
+
+	let {
+		baseImageClass = '',
+		class: className = '',
+		org,
+		quantImageClass = 'h-full w-full',
+		quantOrg,
+		quantPositionClass = '-bottom-0.75 -right-0.75',
+		quantSize = 'h-4.25 w-4.25',
+		size = 'h-9 w-9'
+	}: Props = $props();
+
+	let avatarError = $state(false);
+	let quantError = $state(false);
+
+	let invertAvatar = $derived(DARK_INVERT_AVATAR_ORGS.includes(org));
+	let invertQuant = $derived(DARK_INVERT_AVATAR_ORGS.includes(quantOrg ?? ''));
+
+	// Monogram fallback: org initial on a hue derived from its name, so each org
+	// gets a stable distinct color.
+	let hue = $derived.by(() => {
+		let h = 0;
+
+		for (let i = 0; i < org.length; i++) h = (h * 31 + org.charCodeAt(i)) >>> 0;
+
+		return h % 360;
+	});
+
+	let quantHue = $derived.by(() => {
+		const name = quantOrg ?? '';
+
+		let h = 0;
+
+		for (let i = 0; i < name.length; i++) h = (h * 31 + name.charCodeAt(i)) >>> 0;
+
+		return h % 360;
+	});
+</script>
+
+<span class="relative mt-0.5 inline-flex shrink-0 {className}">
+	{#if avatarError}
+		<span
+			aria-hidden="true"
+			class="flex {size} items-center justify-center rounded-md text-sm font-semibold text-white"
+			style="background-color: hsl({hue} 60% 45%)"
+		>
+			{org.charAt(0).toUpperCase()}
+		</span>
+	{:else}
+		<div class="rounded-md">
+			<img
+				alt=""
+				class="{size} rounded-md {invertAvatar ? 'dark:invert' : ''} {baseImageClass}"
+				loading="lazy"
+				onerror={() => (avatarError = true)}
+				src={HuggingFaceService.getAvatarUrl(org)}
+			/>
+		</div>
+	{/if}
+
+	{#if quantOrg && quantOrg !== org}
+		<Tooltip.Root>
+			<Tooltip.Trigger
+				class="absolute {quantPositionClass} {quantSize} overflow-hidden rounded-full border border-background bg-muted "
+			>
+				{#if quantError}
+					<span
+						aria-hidden="true"
+						class="flex h-full w-full items-center justify-center rounded-full text-[8px] font-semibold text-white"
+						style="background-color: hsl({quantHue} 60% 45%)"
+					>
+						{quantOrg.charAt(0).toUpperCase()}
+					</span>
+				{:else}
+					<img
+						alt=""
+						class="{quantImageClass} rounded-full {invertQuant ? 'dark:invert' : ''}"
+						loading="lazy"
+						onerror={() => (quantError = true)}
+						src={HuggingFaceService.getAvatarUrl(quantOrg)}
+					/>
+				{/if}
+			</Tooltip.Trigger>
+
+			<Tooltip.Content>
+				<p>{quantOrg}</p>
+			</Tooltip.Content>
+		</Tooltip.Root>
+	{/if}
+</span>

tools/ui/src/lib/components/app/models/discover/ModelsDiscoverDetails/ModelsDiscoverChatTemplateDialog.svelte

@@ -0,0 +1,67 @@
+<script lang="ts">
+	import { Check, Copy, X } from '@lucide/svelte';
+	import * as Dialog from '$lib/components/ui/dialog';
+	import { copyToClipboard } from '$lib/utils';
+
+	interface Props {
+		open?: boolean;
+		chatTemplate: string;
+		onOpenChange?: (open: boolean) => void;
+	}
+
+	let { chatTemplate, onOpenChange, open = $bindable(false) }: Props = $props();
+
+	function handleOpenChange(value: boolean) {
+		open = value;
+		onOpenChange?.(value);
+	}
+
+	let copied = $state(false);
+
+	async function copy() {
+		await copyToClipboard(chatTemplate);
+		copied = true;
+		setTimeout(() => (copied = false), 1500);
+	}
+</script>
+
+<Dialog.Root onOpenChange={handleOpenChange} {open}>
+	<Dialog.Content
+		class="flex max-h-[calc(100vh-4rem)] flex-col gap-0 p-0 md:w-[calc(100vw-4rem)]! md:max-w-4xl!"
+	>
+		<!-- The header's corner-pinned close X never lines up with a padded flex row,
+		     so it is replaced by one inside the row, aligned with the title -->
+		<Dialog.Header
+			class="flex-row items-center gap-2 border-b border-border/40 p-4"
+			showCloseButton={false}
+		>
+			<Dialog.Title class="text-sm font-semibold">Chat template</Dialog.Title>
+
+			<button
+				aria-label="Copy chat template"
+				class="inline-flex cursor-pointer items-center gap-1.5 rounded-md border px-2 py-1 text-xs font-medium transition-colors hover:bg-muted"
+				onclick={() => void copy()}
+				type="button"
+			>
+				{#if copied}
+					<Check class="h-3.5 w-3.5 text-green-500" />
+				{:else}
+					<Copy class="h-3.5 w-3.5" />
+				{/if}
+
+				Copy
+			</button>
+
+			<Dialog.Close
+				aria-label="Close"
+				class="ml-auto inline-flex cursor-pointer items-center justify-center rounded-md p-1.5 text-muted-foreground/70 transition-colors hover:bg-muted-foreground/10 hover:text-foreground"
+				type="button"
+			>
+				<X class="h-4 w-4" />
+			</Dialog.Close>
+		</Dialog.Header>
+
+		<pre
+			class="flex-1 overflow-auto p-4 font-mono text-xs break-all whitespace-pre-wrap text-muted-foreground">{chatTemplate}</pre>
+	</Dialog.Content>
+</Dialog.Root>

tools/ui/src/lib/components/app/models/discover/ModelsDiscoverDetails/ModelsDiscoverDetails.svelte

@@ -0,0 +1,96 @@
+<script lang="ts">
+	import { ModelsDiscoverDetailsDownloadOptions } from './ModelsDiscoverDetailsDownloadOptions';
+	import ModelsDiscoverDetailsHeader from './ModelsDiscoverDetailsHeader.svelte';
+	import ModelsDiscoverDetailsReadme from './ModelsDiscoverDetailsReadme.svelte';
+	import ModelsDiscoverDetailsSkeleton from './ModelsDiscoverDetailsSkeleton.svelte';
+	import { OTHER_BIT_DEPTH } from '$lib/constants';
+	import { ModelAuxSidecar } from '$lib/enums';
+	import { HuggingFaceService } from '$lib/services';
+	import type { HfModelDetailInfo, HfModelSibling, ModelBitDepthRow } from '$lib/types';
+	import { detectThinkingSupport, detectToolUseSupport } from '$lib/utils';
+	import { SvelteMap } from 'svelte/reactivity';
+
+	interface Props {
+		/** Full HuggingFace model id, e.g. `ggml-org/gemma-3-4b-it-GGUF`. */
+		modelId: string;
+		/** Model details from `/api/models/{id}?full=true`; null while loading. */
+		details: HfModelDetailInfo | null;
+		/** GGUF files of the repo, shards collapsed, sorted by size desc. */
+		files: HfModelSibling[];
+		/** README.md content, frontmatter stripped; null when unavailable. */
+		readme: string | null;
+		/** True while the model data is being fetched. */
+		loading?: boolean;
+		/** Error message when loading failed. */
+		error?: string | null;
+	}
+
+	let { details, error = null, files, loading = false, modelId, readme }: Props = $props();
+
+	let gguf = $derived(details?.gguf);
+	let baseModels = $derived(HuggingFaceService.getBaseModels(details));
+	let licenseTag = $derived.by(() => {
+		const tags = details?.tags ?? [];
+
+		return tags.find((t) => t.startsWith('license:'))?.replace('license:', '') ?? null;
+	});
+
+	// Capabilities derived from HF metadata. Vision comes from an mmproj sidecar
+	// or a multimodal pipeline tag; tool use / reasoning from the chat template.
+	let hasMmproj = $derived(
+		files.some(
+			(f) => HuggingFaceService.extractQuantMeta(f.path)?.sidecar === ModelAuxSidecar.MMPROJ
+		)
+	);
+	let hasVision = $derived(hasMmproj || details?.pipeline_tag === 'image-text-to-text');
+	let hasTools = $derived(detectToolUseSupport(gguf?.chat_template ?? ''));
+	let hasReasoning = $derived(detectThinkingSupport(gguf?.chat_template ?? ''));
+
+	let bitDepthRows = $derived.by<ModelBitDepthRow[]>(() => {
+		const rows = new SvelteMap<number, HfModelSibling[]>();
+
+		for (const file of files) {
+			const meta = HuggingFaceService.extractQuantMeta(file.path);
+
+			// mmproj sidecars are already conveyed by the Vision capability badge;
+			// imatrix ships as a normal chip with its own badge.
+			if (meta?.sidecar === ModelAuxSidecar.MMPROJ) continue;
+
+			const depth = meta?.quant ? HuggingFaceService.getBitDepth(meta.quant) : null;
+			const bucket = depth ?? OTHER_BIT_DEPTH;
+			const list = rows.get(bucket) ?? [];
+
+			list.push(file);
+			rows.set(bucket, list);
+		}
+
+		return Array.from(rows.entries())
+			.map(([bitDepth, rowFiles]) => ({ bitDepth, files: rowFiles }))
+			.sort((a, b) => a.bitDepth - b.bitDepth);
+	});
+</script>
+
+{#if loading}
+	<ModelsDiscoverDetailsSkeleton />
+{:else if error}
+	<div class="flex h-full items-center justify-center py-20">
+		<p class="text-sm text-destructive">{error}</p>
+	</div>
+{:else if details}
+	<div class="space-y-6 p-6">
+		<ModelsDiscoverDetailsHeader
+			{baseModels}
+			{details}
+			{gguf}
+			{hasReasoning}
+			{hasTools}
+			{hasVision}
+			{licenseTag}
+			{modelId}
+		/>
+
+		<ModelsDiscoverDetailsDownloadOptions {bitDepthRows} {modelId} />
+
+		<ModelsDiscoverDetailsReadme {readme} />
+	</div>
+{/if}

tools/ui/src/lib/components/app/models/discover/ModelsDiscoverDetails/ModelsDiscoverDetailsDownloadOptions/ModelsDiscoverDetailsDownloadOptions.svelte

@@ -0,0 +1,173 @@
+<script lang="ts">
+	import { classify, labelFor } from './download-options.utils';
+	import ModelsDiscoverDetailsDownloadOptionsDownloadCommand from './ModelsDiscoverDetailsDownloadOptionsDownloadCommand.svelte';
+	import ModelsDiscoverDetailsDownloadOptionsRow from './ModelsDiscoverDetailsDownloadOptionsRow.svelte';
+	import { DialogConfirmDownload } from '$lib/components/app/dialogs';
+	import { ModelDownloadConfirmAction, ModelSelectableFileKind } from '$lib/enums';
+	import { HuggingFaceService, ModelsService } from '$lib/services';
+	import { modelsStore } from '$lib/stores';
+	import type {
+		ModelBitDepthRow,
+		ModelDownloadEntryState,
+		ModelQuantOption,
+		ModelSelectableFile
+	} from '$lib/types';
+
+	interface Props {
+		/** Full HuggingFace repo id, e.g. `ggml-org/gemma-3-4b-it-GGUF`. */
+		modelId: string;
+		/** GGUF files grouped by bit depth. */
+		bitDepthRows: ModelBitDepthRow[];
+		/** Download state lookup; defaults to the models store status feed. */
+		getDownloadState?: (
+			repoWithTag: string,
+			filePath: string,
+			isSidecar: boolean
+		) => ModelDownloadEntryState;
+	}
+
+	let { bitDepthRows, getDownloadState, modelId }: Props = $props();
+
+	// Destructive chip actions (delete a downloaded model, cancel a download) are
+	// confirmed here rather than inside each chip: a single shared dialog owned by
+	// the options panel, keyed by the repo+tag the user acted on, so one dialog is
+	// mounted for the whole panel instead of one per chip.
+	// The acted-on target is kept after closing so the copy stays rendered through
+	// the dialog's close transition.
+	let pending: { action: ModelDownloadConfirmAction; repoWithTag: string } = $state({
+		action: ModelDownloadConfirmAction.CANCEL,
+		repoWithTag: ''
+	});
+	let confirmOpen = $state(false);
+
+	function requestCancel(repoWithTag: string) {
+		pending = { action: ModelDownloadConfirmAction.CANCEL, repoWithTag };
+		confirmOpen = true;
+	}
+
+	function requestDelete(repoWithTag: string) {
+		pending = { action: ModelDownloadConfirmAction.DELETE, repoWithTag };
+		confirmOpen = true;
+	}
+
+	function stateFor(
+		repoWithTag: string,
+		filePath: string,
+		isSidecar: boolean
+	): ModelDownloadEntryState {
+		if (getDownloadState) return getDownloadState(repoWithTag, filePath, isSidecar);
+
+		const isDownloading = modelsStore.status.isDownloadInProgress(repoWithTag);
+		const isPaused = modelsStore.status.isDownloadPaused(repoWithTag);
+
+		return {
+			// solo downloads register in /v1/models under the tag (args stay empty),
+			// while drafts pulled by a loaded model only show up as its --model-draft
+			isDownloaded:
+				!isDownloading &&
+				(modelsStore.status.isModelDownloaded(repoWithTag) ||
+					(isSidecar && modelsStore.status.isSidecarDownloaded(modelId, filePath))),
+			isDownloading,
+			isFailed: modelsStore.status.hasFailedDownload(repoWithTag),
+			isPaused,
+			// live progress while downloading, else the frozen snapshot of the pause
+			progress:
+				modelsStore.status.getDownloadProgress(repoWithTag) ??
+				modelsStore.status.getPausedDownloadProgress(repoWithTag),
+			repoWithTag
+		};
+	}
+
+	/**
+	 * Every selectable file with its kind and download state, in row order.
+	 * Single source of truth for the chip rows and the command selects.
+	 */
+	let selectableFiles = $derived.by(() => {
+		const files: (ModelSelectableFile & { state: ModelDownloadEntryState })[] = [];
+
+		for (const row of bitDepthRows) {
+			for (const file of row.files) {
+				const meta = HuggingFaceService.extractQuantMeta(file.path);
+				const tag = ModelsService.buildDownloadTag(
+					modelId,
+					meta?.quant ?? null,
+					meta?.sidecar ?? null
+				);
+
+				files.push({
+					...file,
+					kind: classify(file.path),
+					state: stateFor(tag, file.path, Boolean(meta?.sidecar))
+				});
+			}
+		}
+
+		return files;
+	});
+
+	/** Rows for the row component: per-bit-depth files with state attached. */
+	let rows = $derived.by(() =>
+		bitDepthRows.map((row) => {
+			const paths = new Set(row.files.map((f) => f.path));
+
+			return {
+				bitDepth: row.bitDepth,
+				files: selectableFiles.filter((f) => paths.has(f.path))
+			};
+		})
+	);
+
+	function optionFor(file: ModelSelectableFile): ModelQuantOption {
+		return {
+			label: labelFor(file.path),
+			path: file.path
+		};
+	}
+
+	/** Non-draft quants for the command's base select, in row order. */
+	let mainOptions = $derived(
+		selectableFiles.filter((f) => f.kind === ModelSelectableFileKind.MAIN).map(optionFor)
+	);
+
+	/**
+	 * Draft options for the command's draft select, with their sidecar type
+	 * (MTP, DFLASH...) since a repo can ship more than one draft flavour.
+	 */
+	let draftOptions = $derived(
+		selectableFiles
+			.filter((f) => f.kind === ModelSelectableFileKind.DRAFT)
+			.map((f) => ({
+				...optionFor(f),
+				badge: HuggingFaceService.extractQuantMeta(f.path)?.sidecar ?? null
+			}))
+	);
+</script>
+
+{#if bitDepthRows.length}
+	<section class="rounded-3xl border border-border/30 bg-muted/60 shadow-xs dark:border-border/20">
+		<!-- One chip per file, each an independent download action with its own
+			 lifecycle state; nothing here selects anything. -->
+		<div class="flex w-full flex-col divide-y divide-border/50 px-4 pb-1 dark:divide-border/35">
+			{#each rows as row (row.bitDepth)}
+				<ModelsDiscoverDetailsDownloadOptionsRow
+					bitDepth={row.bitDepth}
+					files={row.files}
+					onRequestCancel={requestCancel}
+					onRequestDelete={requestDelete}
+				/>
+			{/each}
+		</div>
+
+		<!-- Terminal command preview, standalone: its picks are not bound to the chips. -->
+		<div class="border-t border-border/50 px-4 pt-3.5 pb-4 dark:border-border/35">
+			<ModelsDiscoverDetailsDownloadOptionsDownloadCommand {draftOptions} {mainOptions} {modelId} />
+		</div>
+	</section>
+{/if}
+
+<DialogConfirmDownload
+	action={pending.action}
+	onClose={() => (confirmOpen = false)}
+	open={confirmOpen}
+	repoWithTag={pending.repoWithTag}
+/>

tools/ui/src/lib/components/app/models/discover/ModelsDiscoverDetails/ModelsDiscoverDetailsDownloadOptions/ModelsDiscoverDetailsDownloadOptionsDownloadCommand.svelte

@@ -0,0 +1,278 @@
+<script lang="ts">
+	import { Check, Copy, Plus, X } from '@lucide/svelte';
+	import * as Select from '$lib/components/ui/select';
+	import {
+		DEFAULT_BASE_BIT_DEPTH,
+		MODEL_ID,
+		type ModelSidecar,
+		OTHER_BIT_DEPTH,
+		SERVE_COMMAND,
+		SPEC_TYPE
+	} from '$lib/constants';
+	import { HuggingFaceService } from '$lib/services';
+	import type { ModelQuantOption } from '$lib/types';
+	import { copyToClipboard } from '$lib/utils';
+
+	interface Props {
+		modelId: string;
+		/** Non-draft quants of the repo, in bit-depth row order. */
+		mainOptions: ModelQuantOption[];
+		/** Draft sidecar files with their sidecar badge; empty when the repo ships none. */
+		draftOptions: (ModelQuantOption & { badge: ModelSidecar | null })[];
+	}
+
+	let { draftOptions, mainOptions, modelId }: Props = $props();
+
+	// Command picks, owned here: nothing two-way binds them to the quant chips.
+	let basePick = $state<string | null>(null);
+	let draftPick = $state<string | null>(null);
+	let draftTypePick = $state<ModelSidecar | null>(null);
+	let withDraft = $state(false);
+
+	function bitDepthOf(path: string): number {
+		const quant = HuggingFaceService.extractQuantMeta(path)?.quant;
+
+		return quant ? (HuggingFaceService.getBitDepth(quant) ?? OTHER_BIT_DEPTH) : OTHER_BIT_DEPTH;
+	}
+
+	/**
+	 * Base file the command points at: the user's pick while it still exists in
+	 * the options, else the 4-bit file, else the lowest bit depth available. A
+	 * stale pick (the details pane switched models) falls back on its own.
+	 */
+	let baseOption = $derived.by(() => {
+		const picked = mainOptions.find((option) => option.path === basePick);
+
+		if (picked) return picked;
+
+		const preferred = mainOptions.find(
+			(option) => bitDepthOf(option.path) === DEFAULT_BASE_BIT_DEPTH
+		);
+
+		if (preferred) return preferred;
+
+		const ranked = [...mainOptions].sort((a, b) => bitDepthOf(a.path) - bitDepthOf(b.path));
+
+		return ranked[0] ?? null;
+	});
+
+	/** Draft sidecar types the repo ships, in option order. */
+	let specTypes = $derived(
+		draftOptions
+			.map((option) => option.badge)
+			.filter((badge): badge is ModelSidecar => badge !== null)
+			.filter((badge, index, all) => all.indexOf(badge) === index)
+	);
+
+	/**
+	 * Draft type the --spec-type select points at: the user's pick while it
+	 * still exists, else the first type the repo ships.
+	 */
+	let draftType = $derived(
+		draftTypePick && specTypes.includes(draftTypePick) ? draftTypePick : (specTypes[0] ?? null)
+	);
+
+	/** Draft files of the picked type; the quant select only offers these. */
+	let typeDraftOptions = $derived(draftOptions.filter((option) => option.badge === draftType));
+
+	/** Draft file the -hfd tag points at: the user's pick, else the first of the type. */
+	let draftOption = $derived(
+		withDraft
+			? (typeDraftOptions.find((option) => option.path === draftPick) ??
+					typeDraftOptions[0] ??
+					null)
+			: null
+	);
+
+	/** Quant of the file the `-hf` tag points at; null when the file carries no quant. */
+	let mainQuant = $derived(
+		baseOption ? (HuggingFaceService.extractQuantMeta(baseOption.path)?.quant ?? null) : null
+	);
+
+	/** Quant of the file the `-hfd` tag points at. */
+	let draftQuant = $derived(
+		draftOption ? (HuggingFaceService.extractQuantMeta(draftOption.path)?.quant ?? null) : null
+	);
+
+	/** `--spec-type` value; null when no draft type resolved. */
+	let specType = $derived(draftType ? SPEC_TYPE[draftType] : null);
+
+	/** The llama serve command, composed from the inline picks. */
+	let command = $derived.by(() => {
+		const parts = [
+			SERVE_COMMAND.BIN,
+			SERVE_COMMAND.SUBCOMMAND,
+			SERVE_COMMAND.MODEL_FLAG,
+			mainQuant ? `${modelId}${MODEL_ID.QUANTIZATION_SEPARATOR}${mainQuant}` : modelId
+		];
+
+		if (draftOption && draftQuant) {
+			parts.push(
+				SERVE_COMMAND.DRAFT_FLAG,
+				`${modelId}${MODEL_ID.QUANTIZATION_SEPARATOR}${draftQuant}`,
+				SERVE_COMMAND.SPEC_TYPE_FLAG,
+				specType ?? ''
+			);
+		}
+
+		return parts.join(' ');
+	});
+
+	let copied = $state(false);
+
+	async function copy() {
+		await copyToClipboard(command);
+		copied = true;
+		setTimeout(() => (copied = false), 1500);
+	}
+</script>
+
+<!-- <div aria-hidden="true" class="flex items-center gap-3 mt-2 mb-4">
+	<span class="h-px flex-1 bg-border/50"></span>
+
+	<span class="text-xs whitespace-nowrap text-muted-foreground"> or run in your terminal </span>
+
+	<span class="h-px flex-1 bg-border/50"></span>
+</div> -->
+
+<div
+	class="relative flex items-center gap-2 overflow-hidden rounded-lg border border-border/40 bg-background py-2.5 pl-4 pr-10 shadow-xs dark:border-border/35 dark:bg-background/50"
+>
+	<!-- Single line: long commands scroll horizontally instead of wrapping. -->
+	<div
+		class="flex min-w-0 flex-1 items-center gap-x-2 overflow-x-auto py-0.5 font-mono text-xs whitespace-nowrap text-foreground/90"
+	>
+		<span class="shrink-0">{SERVE_COMMAND.BIN}</span>
+
+		<span class="shrink-0">{SERVE_COMMAND.SUBCOMMAND}</span>
+
+		<span class="shrink-0">{SERVE_COMMAND.MODEL_FLAG}</span>
+
+		<span class="shrink-0">
+			{modelId}{mainQuant ? MODEL_ID.QUANTIZATION_SEPARATOR : ''}
+		</span>
+
+		<!-- Base quant: always part of the command, the 4-bit file by default. -->
+		{#if baseOption}
+			<Select.Root onValueChange={(v) => v && (basePick = v)} type="single" value={baseOption.path}>
+				<Select.Trigger
+					aria-label="Base model quantization"
+					class="-ml-2 border-primary/15 bg-primary/[0.07] font-mono text-foreground hover:bg-primary/15 focus-visible:border-primary/40 focus-visible:ring-0"
+					size="xs"
+				>
+					{baseOption.label}
+				</Select.Trigger>
+
+				<Select.Content class="font-mono text-xs">
+					{#each mainOptions as option (option.path)}
+						<Select.Item class="text-xs" label={option.label} value={option.path}>
+							{option.label}
+						</Select.Item>
+					{/each}
+				</Select.Content>
+			</Select.Root>
+		{/if}
+
+		<!-- Draft add: a tiny dashed affordance right of the base part; gone once added. -->
+		{#if draftOptions.length && !withDraft}
+			<button
+				aria-label="Add draft model"
+				class="mx-1 inline-flex h-5 shrink-0 cursor-pointer items-center gap-1 rounded-md border border-dashed border-border/60 px-1.5 text-[10px] text-muted-foreground transition-colors hover:border-border hover:text-foreground"
+				onclick={() => (withDraft = true)}
+				type="button"
+			>
+				<Plus class="h-3 w-3" />
+
+				add draft model
+			</button>
+		{/if}
+
+		<!-- Draft segment: quant and spec type of the picked draft flavour. The X
+				 at the end drops the whole segment (the add button is gone once added);
+				 it only appears while hovering the segment, or directly on touch -->
+		{#if draftOption}
+			<span class="group/draft inline-flex shrink-0 items-center gap-x-2">
+				<span>{SERVE_COMMAND.DRAFT_FLAG}</span>
+
+				<span class="shrink-0">
+					{modelId}{draftQuant ? MODEL_ID.QUANTIZATION_SEPARATOR : ''}
+				</span>
+
+				<Select.Root
+					onValueChange={(v) => v && (draftPick = v)}
+					type="single"
+					value={draftOption.path}
+				>
+					<Select.Trigger
+						aria-label="Draft model quantization"
+						class="-ml-2 border-primary/15 bg-primary/[0.07] font-mono text-foreground hover:bg-primary/15 focus-visible:border-primary/40 focus-visible:ring-0"
+						size="xs"
+					>
+						{draftOption.label}
+					</Select.Trigger>
+
+					<Select.Content class="font-mono text-xs">
+						{#each typeDraftOptions as option (option.path)}
+							<Select.Item class="text-xs" label={option.label} value={option.path}>
+								{option.label}
+							</Select.Item>
+						{/each}
+					</Select.Content>
+				</Select.Root>
+
+				{#if draftType}
+					<span>{SERVE_COMMAND.SPEC_TYPE_FLAG}</span>
+
+					<!-- the select only earns its chrome when there is a real choice to make -->
+					{#if specTypes.length > 1}
+						<Select.Root
+							onValueChange={(v) => v && (draftTypePick = v as ModelSidecar)}
+							type="single"
+							value={draftType}
+						>
+							<Select.Trigger
+								aria-label="Draft type"
+								class="border-primary/15 bg-primary/[0.07] font-mono text-foreground hover:bg-primary/15 focus-visible:border-primary/40 focus-visible:ring-0"
+								size="xs"
+							>
+								{SPEC_TYPE[draftType]}
+							</Select.Trigger>
+
+							<Select.Content class="font-mono text-xs">
+								{#each specTypes as type (type)}
+									<Select.Item class="text-xs" label={SPEC_TYPE[type]} value={type}>
+										{SPEC_TYPE[type]}
+									</Select.Item>
+								{/each}
+							</Select.Content>
+						</Select.Root>
+					{:else}
+						<span>{SPEC_TYPE[draftType]}</span>
+					{/if}
+				{/if}
+
+				<button
+					aria-label="Remove draft model"
+					class="shrink-0 cursor-pointer text-muted-foreground/60 opacity-0 transition-[opacity,color] duration-150 ease-[cubic-bezier(0.23,1,0.32,1)] hover:text-destructive group-hover/draft:opacity-100 [@media(pointer:coarse)]:opacity-100"
+					onclick={() => (withDraft = false)}
+					type="button"
+				>
+					<X class="h-3 w-3" />
+				</button>
+			</span>
+		{/if}
+	</div>
+
+	<button
+		aria-label="Copy command"
+		class="absolute top-1/2 right-2 -translate-y-1/2 cursor-pointer rounded-md p-1.5 text-muted-foreground/70 transition-colors hover:bg-primary/10 hover:text-foreground"
+		onclick={copy}
+		type="button"
+	>
+		{#if copied}
+			<Check class="h-3.5 w-3.5 text-green-500" />
+		{:else}
+			<Copy class="h-3.5 w-3.5" />
+		{/if}
+	</button>
+</div>

tools/ui/src/lib/components/app/models/discover/ModelsDiscoverDetails/ModelsDiscoverDetailsDownloadOptions/ModelsDiscoverDetailsDownloadOptionsQuantDownloadButton.svelte

@@ -0,0 +1,244 @@
+<script lang="ts">
+	import ModelsDiscoverDownloadProgressBar from '../../ModelsDiscoverDownloadProgressBar.svelte';
+	import { labelFor } from './download-options.utils';
+	import { Check, Download, Loader2, Pause, Play, RotateCw, X } from '@lucide/svelte';
+	import * as Tooltip from '$lib/components/ui/tooltip';
+	import { HuggingFaceService } from '$lib/services';
+	import { modelsStore } from '$lib/stores';
+	import type { HfModelSibling, ModelDownloadEntryState } from '$lib/types';
+	import type { Snippet } from 'svelte';
+
+	interface Props {
+		/** GGUF file the chip stands for. */
+		file: HfModelSibling;
+		/** Download state of the file, from the parent's status feed. */
+		entry: ModelDownloadEntryState;
+		/**
+		 * Ask the parent to confirm deleting a downloaded model. The chip owns no
+		 * dialog; the parent renders the single confirmation and acts on confirm.
+		 */
+		onRequestDelete?: (repoWithTag: string) => void;
+		/** Ask the parent to confirm cancelling an in-flight download. */
+		onRequestCancel?: (repoWithTag: string) => void;
+	}
+
+	let { entry, file, onRequestCancel, onRequestDelete }: Props = $props();
+
+	// Sidecar kind (mtp, mmproj, ...) tag shown on every chip state; one source so
+	// the idle, in-flight and downloaded variants stay identical.
+	const SIDECAR_BADGE_CLASS =
+		'rounded-md bg-primary px-1 py-0.5 text-[10px] font-semibold tracking-wide text-primary-foreground uppercase';
+
+	/** Queue the download; a failed attempt leaves partial files, drop them first. */
+	async function startDownload() {
+		try {
+			if (entry.isFailed) await modelsStore.status.cancelDownload(entry.repoWithTag);
+
+			await modelsStore.status.downloadModel(entry.repoWithTag);
+		} catch {
+			// the store already toasted the failure
+		}
+	}
+
+	let meta = $derived(HuggingFaceService.extractQuantMeta(file.path));
+	let label = $derived(labelFor(file.path));
+
+	let percent = $derived(
+		entry.progress && entry.progress.totalBytes > 0
+			? Math.round((entry.progress.downloadedBytes / entry.progress.totalBytes) * 100)
+			: null
+	);
+
+	let tooltipText = $derived(
+		entry.isDownloading
+			? 'Pause downloading'
+			: entry.isPaused
+				? 'Resume downloading'
+				: entry.isDownloaded
+					? 'Delete model'
+					: entry.isFailed
+						? `Retry download: ${file.path}`
+						: `Download ${file.path}`
+	);
+</script>
+
+{#snippet chipBody(dividerClass: string)}
+	<!-- badge, label and divider: identical on every chip state -->
+	{#if meta?.sidecar}
+		<span class={SIDECAR_BADGE_CLASS}>
+			{meta.sidecar}
+		</span>
+	{/if}
+
+	<span class="font-medium">{label}</span>
+
+	<span class={dividerClass}></span>
+{/snippet}
+
+<!-- every chip is a tooltip trigger; the button renders as the trigger's child
+     so no nested button element is created -->
+{#snippet tooltipTrigger(tip: string, button: Snippet<[Record<string, unknown>]>)}
+	<Tooltip.Root>
+		<Tooltip.Trigger>
+			{#snippet child({ props })}
+				{@render button(props)}
+			{/snippet}
+		</Tooltip.Trigger>
+
+		<Tooltip.Content>
+			<p>{tip}</p>
+		</Tooltip.Content>
+	</Tooltip.Root>
+{/snippet}
+
+{#snippet deleteChip(props: Record<string, unknown>)}
+	<!-- downloaded chips are delete actions: green check by default, red X on hover -->
+	<button
+		{...props}
+		aria-label={tooltipText}
+		class="group relative inline-flex h-auto cursor-pointer items-center gap-1 rounded-md! border px-2 py-1 text-left font-mono text-xs shadow-xs transition-[background-color,border-color,transform] duration-200 ease-[cubic-bezier(0.23,1,0.32,1)] active:scale-[0.97]
+					border-green-600/25 bg-green-500/5 hover:border-destructive/50 hover:bg-destructive/10 dark:border-green-500/30 dark:bg-green-500/10 dark:hover:border-destructive/50 dark:hover:bg-destructive/15"
+		onclick={() => onRequestDelete?.(entry.repoWithTag)}
+		type="button"
+	>
+		{@render chipBody(
+			'-my-1 mx-0.75 w-px self-stretch bg-green-600/25 transition-colors duration-200 group-hover:bg-destructive/30 dark:bg-green-600/30 dark:group-hover:bg-destructive/30'
+		)}
+
+		<span>{HuggingFaceService.formatFileSize(file.size ?? 0)}</span>
+
+		<!-- icon slot: crossfade check -> x; touch devices show the delete affordance directly -->
+		<span class="relative inline-flex h-3.5 w-3.5 shrink-0 items-center justify-center">
+			<Check
+				class="absolute h-3.5 w-3.5 text-green-500 transition-[opacity,transform] duration-150 ease-[cubic-bezier(0.23,1,0.32,1)] group-hover:scale-75 group-hover:opacity-0 [@media(pointer:coarse)]:hidden"
+			/>
+
+			<X
+				class="absolute h-3.5 w-3.5 scale-75 text-destructive opacity-0 transition-[opacity,transform] duration-150 ease-[cubic-bezier(0.23,1,0.32,1)] group-hover:scale-100 group-hover:opacity-100 [@media(pointer:coarse)]:scale-100 [@media(pointer:coarse)]:opacity-100"
+			/>
+		</span>
+	</button>
+{/snippet}
+
+{#snippet pauseResumeChip(props: Record<string, unknown>)}
+	<button
+		{...props}
+		aria-label={tooltipText}
+		class="flex h-auto min-w-0 flex-1 cursor-pointer items-center gap-1 text-left"
+		onclick={() => {
+			if (entry.isDownloading) void modelsStore.status.pauseDownload(entry.repoWithTag);
+			else void modelsStore.status.downloadModel(entry.repoWithTag).catch(() => {});
+		}}
+		type="button"
+	>
+		{@render chipBody('-my-1 mx-0.75 w-px self-stretch bg-border')}
+
+		{#if percent !== null}
+			<span class="mr-1 tabular-nums">{percent}%</span>
+		{:else if entry.isPaused}
+			<span>Paused</span>
+		{:else}
+			<span>{HuggingFaceService.formatFileSize(file.size ?? 0)}</span>
+		{/if}
+
+		{#if entry.isDownloading}
+			<!-- spinner fades into the pause affordance on hover; opacity only, the
+			     spin keyframes own the transform so scale would fight them -->
+			<span class="relative inline-flex h-3.5 w-3.5 shrink-0 items-center justify-center">
+				<Loader2
+					class="absolute h-3.5 w-3.5 animate-spin text-muted-foreground transition-opacity duration-150 ease-[cubic-bezier(0.23,1,0.32,1)] group-hover:opacity-0 [@media(pointer:coarse)]:hidden"
+				/>
+
+				<Pause
+					class="absolute h-3.5 w-3.5 opacity-0 transition-opacity duration-150 ease-[cubic-bezier(0.23,1,0.32,1)] group-hover:opacity-100 [@media(pointer:coarse)]:opacity-100"
+				/>
+			</span>
+		{:else}
+			<!-- paused: the play affordance fades in on hover; visible directly on touch -->
+			<span class="relative inline-flex h-3.5 w-3.5 shrink-0 items-center justify-center">
+				<Play
+					class="absolute h-3.5 w-3.5 scale-75 opacity-0 transition-[opacity,transform] duration-150 ease-[cubic-bezier(0.23,1,0.32,1)] group-hover:scale-100 group-hover:opacity-100 [@media(pointer:coarse)]:scale-100 [@media(pointer:coarse)]:opacity-100"
+				/>
+			</span>
+		{/if}
+	</button>
+{/snippet}
+
+{#snippet cancelChip(props: Record<string, unknown>)}
+	<!-- cancel slot: same fixed slot and fade as the pause / play affordances
+	     so the two icons line up exactly; the X turns destructive on its hover -->
+	<button
+		{...props}
+		aria-label="Cancel downloading"
+		class="relative grid h-3.5 w-3.5 shrink-0 cursor-pointer items-center justify-center text-muted-foreground/70"
+		onclick={() => onRequestCancel?.(entry.repoWithTag)}
+		type="button"
+	>
+		<X
+			class="h-3.5 w-3.5 transition-[opacity,transform,color] duration-150 ease-[cubic-bezier(0.23,1,0.32,1)] text-destructive [@media(pointer:coarse)]:scale-100 [@media(pointer:coarse)]:opacity-100"
+		/>
+	</button>
+{/snippet}
+
+{#snippet downloadChip(props: Record<string, unknown>)}
+	<!-- idle chips download on click (retry when the last attempt failed) -->
+	<button
+		{...props}
+		aria-label={tooltipText}
+		class="group relative inline-flex h-auto cursor-pointer items-center gap-1 overflow-hidden rounded-md! border px-2 py-1 text-left font-mono text-xs shadow-xs transition-[background-color,border-color,transform] duration-200 ease-[cubic-bezier(0.23,1,0.32,1)] active:scale-[0.97]
+				border-border/30 bg-background hover:bg-muted-foreground/10 dark:border-border/20 dark:bg-muted-foreground/15 dark:text-secondary-foreground dark:hover:bg-muted-foreground/25
+				{entry.isFailed ? 'border-destructive!' : ''}"
+		onclick={() => void startDownload()}
+		type="button"
+	>
+		{#if entry.isFailed}
+			<span
+				class="rounded-md bg-destructive px-1 py-0.5 text-[10px] font-semibold tracking-wide text-destructive-foreground uppercase"
+			>
+				Failed
+			</span>
+		{/if}
+
+		{@render chipBody('-my-1 mx-0.75 w-px self-stretch bg-border')}
+
+		<span>{HuggingFaceService.formatFileSize(file.size ?? 0)}</span>
+
+		<span class="inline-flex h-3.5 w-3.5 shrink-0 items-center justify-center">
+			{#if entry.isFailed}
+				<RotateCw class="h-3.5 w-3.5 text-destructive" />
+			{:else}
+				<Download
+					class="h-3.5 w-3.5 text-muted-foreground transition-colors duration-150 group-hover:text-foreground"
+				/>
+			{/if}
+		</span>
+	</button>
+{/snippet}
+
+{#if entry.isDownloaded}
+	{@render tooltipTrigger(tooltipText, deleteChip)}
+{:else if entry.isDownloading || entry.isPaused}
+	<!-- in-flight / paused chips: the chip body pauses / resumes on click, the X
+	     inside the chip cancels (stops and discards the partial files). The X slot
+	     is reserved, so the chip never reflows when the affordance fades in -->
+	<div
+		class="group relative inline-flex h-auto items-center gap-1 overflow-hidden rounded-md! border px-2 py-1 text-left font-mono text-xs shadow-xs transition-[background-color,border-color,transform] duration-200 ease-[cubic-bezier(0.23,1,0.32,1)] active:scale-[0.97]
+			{entry.isPaused
+			? 'border-yellow-600/40 bg-yellow-500/10 hover:bg-yellow-500/20 dark:border-yellow-500/30 dark:bg-yellow-500/10'
+			: 'border-border/30 bg-background hover:bg-muted-foreground/10 dark:border-border/20 dark:bg-muted-foreground/15'}"
+	>
+		{@render tooltipTrigger(tooltipText, pauseResumeChip)}
+
+		{@render tooltipTrigger('Cancel downloading', cancelChip)}
+
+		{#if percent !== null}
+			<ModelsDiscoverDownloadProgressBar
+				downloadedBytes={entry.progress?.downloadedBytes ?? 0}
+				overlay
+				totalBytes={entry.progress?.totalBytes ?? 0}
+			/>
+		{/if}
+	</div>
+{:else}
+	{@render tooltipTrigger(tooltipText, downloadChip)}
+{/if}

tools/ui/src/lib/components/app/models/discover/ModelsDiscoverDetails/ModelsDiscoverDetailsDownloadOptions/ModelsDiscoverDetailsDownloadOptionsRow.svelte

@@ -0,0 +1,60 @@
+<script lang="ts">
+	import ModelsDiscoverDetailsDownloadOptionsQuantDownloadButton from './ModelsDiscoverDetailsDownloadOptionsQuantDownloadButton.svelte';
+	import {
+		BIT_DEPTH_LABEL_SUFFIX,
+		GIGABYTE_LABEL,
+		OTHER_BIT_DEPTH,
+		OTHER_BIT_DEPTH_LABEL
+	} from '$lib/constants';
+	import { ModelSelectableFileKind } from '$lib/enums';
+	import type { ModelDownloadEntryState, ModelSelectableFile } from '$lib/types';
+	import { minMemoryTierGb } from '$lib/utils';
+
+	interface Props {
+		/** Bit depth of the row; `99` renders as "Other". */
+		bitDepth: number;
+		/** Every GGUF of this bit depth, with download state attached. */
+		files: (ModelSelectableFile & { state: ModelDownloadEntryState })[];
+		/** Forwarded to each chip: ask the parent to confirm a cancel. */
+		onRequestCancel?: (repoWithTag: string) => void;
+		/** Forwarded to each chip: ask the parent to confirm a delete. */
+		onRequestDelete?: (repoWithTag: string) => void;
+	}
+
+	let { bitDepth, files, onRequestCancel, onRequestDelete }: Props = $props();
+
+	let mainFile = $derived(files.find((f) => f.kind === ModelSelectableFileKind.MAIN) ?? null);
+	let draftFile = $derived(files.find((f) => f.kind === ModelSelectableFileKind.DRAFT) ?? null);
+
+	let mainMemGb = $derived(mainFile ? minMemoryTierGb(mainFile.size ?? 0) : null);
+	let draftMemGb = $derived(draftFile ? minMemoryTierGb(draftFile.size ?? 0) : null);
+</script>
+
+<div class="grid grid-cols-[5rem_1fr] items-center gap-3 py-3">
+	<div class="pt-1 text-sm tabular-nums text-muted-foreground">
+		{#if bitDepth === OTHER_BIT_DEPTH}
+			{OTHER_BIT_DEPTH_LABEL}
+		{:else}
+			{bitDepth}{BIT_DEPTH_LABEL_SUFFIX}
+		{/if}
+
+		{#if mainMemGb}
+			<span class="block text-[10px] whitespace-nowrap text-muted-foreground/60">
+				needs at least {mainMemGb}{GIGABYTE_LABEL}{draftMemGb
+					? ` + ${draftMemGb}${GIGABYTE_LABEL}`
+					: ''} memory
+			</span>
+		{/if}
+	</div>
+
+	<div class="flex flex-wrap justify-end gap-1.5">
+		{#each files as file (file.path)}
+			<ModelsDiscoverDetailsDownloadOptionsQuantDownloadButton
+				entry={file.state}
+				{file}
+				{onRequestCancel}
+				{onRequestDelete}
+			/>
+		{/each}
+	</div>
+</div>

tools/ui/src/lib/components/app/models/discover/ModelsDiscoverDetails/ModelsDiscoverDetailsDownloadOptions/download-options.utils.ts

@@ -0,0 +1,28 @@
+import { DRAFT_FILE_LABEL, MODEL_ID, PATH_SEPARATOR } from '$lib/constants';
+import { ModelSelectableFileKind } from '$lib/enums';
+import { HuggingFaceService } from '$lib/services';
+import { isAuxSidecar, isDraftSidecar } from '$lib/utils';
+
+/** Kind of a file path: the main weights, a draft sidecar, or an aux sidecar (mmproj). */
+export function classify(path: string): ModelSelectableFileKind {
+	const sidecar = HuggingFaceService.extractQuantMeta(path)?.sidecar;
+
+	if (!sidecar) return ModelSelectableFileKind.MAIN;
+
+	return isAuxSidecar(sidecar) ? ModelSelectableFileKind.AUX : ModelSelectableFileKind.DRAFT;
+}
+
+/** Display label of a file: its quant, else the file name without the extension. */
+export function labelFor(path: string): string {
+	const meta = HuggingFaceService.extractQuantMeta(path);
+
+	if (meta?.quant) return meta.quant;
+
+	// Quantless sidecar files: the chip badge already carries the sidecar type,
+	// so the label only marks draft files; aux sidecars badge alone.
+	if (meta?.sidecar) return isDraftSidecar(meta.sidecar) ? DRAFT_FILE_LABEL : '';
+
+	const basename = path.split(PATH_SEPARATOR).pop() ?? path;
+
+	return basename.replace(MODEL_ID.WEIGHT_EXTENSION_REGEX, '');
+}

tools/ui/src/lib/components/app/models/discover/ModelsDiscoverDetails/ModelsDiscoverDetailsDownloadOptions/index.ts

@@ -0,0 +1,40 @@
+/**
+ *
+ * MODELS DISCOVER - DETAILS - DOWNLOAD OPTIONS
+ *
+ * The download area of the detail pane: GGUF files grouped by bit depth, each an
+ * independent download action chip, plus the standalone terminal command preview.
+ *
+ */
+
+/**
+ * **ModelsDiscoverDetailsDownloadOptions** - GGUF download options
+ *
+ * Groups GGUF files by bit depth and renders one independent download
+ * action chip per file, plus the standalone terminal command preview.
+ */
+export { default as ModelsDiscoverDetailsDownloadOptions } from './ModelsDiscoverDetailsDownloadOptions.svelte';
+
+/**
+ * **ModelsDiscoverDetailsDownloadOptionsRow** - One bit-depth group of quants
+ *
+ * A single bit-depth row inside ModelsDiscoverDetailsDownloadOptions: the
+ * depth label with its memory hint and the quant chips of that depth.
+ */
+export { default as ModelsDiscoverDetailsDownloadOptionsRow } from './ModelsDiscoverDetailsDownloadOptionsRow.svelte';
+
+/**
+ * **ModelsDiscoverDetailsDownloadOptionsQuantDownloadButton** - One quant chip
+ *
+ * A single GGUF file as an independent action chip: download / retry when
+ * idle, pause / resume / cancel while in flight, delete when downloaded.
+ */
+export { default as ModelsDiscoverDetailsDownloadOptionsQuantDownloadButton } from './ModelsDiscoverDetailsDownloadOptionsQuantDownloadButton.svelte';
+
+/**
+ * **ModelsDiscoverDetailsDownloadOptionsDownloadCommand** - Terminal command
+ *
+ * The `llama serve -hf ...` command box with inline quant selects and a copy
+ * button; owns its picks, nothing two-way binds them to the quant chips.
+ */
+export { default as ModelsDiscoverDetailsDownloadOptionsDownloadCommand } from './ModelsDiscoverDetailsDownloadOptionsDownloadCommand.svelte';

tools/ui/src/lib/components/app/models/discover/ModelsDiscoverDetails/ModelsDiscoverDetailsHeader.svelte

@@ -0,0 +1,93 @@
+<script lang="ts">
+	import ModelCapabilityIcons from '../../ModelCapabilityIcons.svelte';
+	import ModelsDiscoverAvatar from '../ModelsDiscoverAvatar.svelte';
+	import ModelsDiscoverDetailsHfHubStats from './ModelsDiscoverDetailsHfHubStats.svelte';
+	import ModelsDiscoverDetailsMetadata from './ModelsDiscoverDetailsMetadata.svelte';
+	import { ExternalLink } from '@lucide/svelte';
+	import { ICON_CLASS_SM } from '$lib/constants';
+	import { HuggingFaceService } from '$lib/services';
+	import type { HfModelDetailInfo, HfModelGguf } from '$lib/types/huggingface';
+	import { orgOf } from '$lib/utils';
+
+	interface Props {
+		modelId: string;
+		details: HfModelDetailInfo;
+		gguf?: HfModelGguf;
+		baseModels: string[];
+		licenseTag: string | null;
+		hasVision: boolean;
+		hasTools: boolean;
+		hasReasoning: boolean;
+	}
+
+	let { baseModels, details, gguf, hasReasoning, hasTools, hasVision, licenseTag, modelId }: Props =
+		$props();
+
+	// Avatar shows the base model's org (e.g. the Qwen logo for a ggml-org GGUF)
+	// with the quant org as a corner badge when they differ.
+	let repoOrg = $derived(orgOf(details.id) || orgOf(modelId));
+	let baseOrg = $derived(orgOf(baseModels[0]));
+	let avatarOrg = $derived(baseOrg || repoOrg);
+	let quantOrg = $derived(baseOrg && baseOrg !== repoOrg ? repoOrg : undefined);
+</script>
+
+<header class="space-y-3">
+	<div class="flex items-start justify-between gap-3">
+		<div class="flex min-w-0 items-center gap-2">
+			<ModelsDiscoverAvatar
+				org={avatarOrg}
+				{quantOrg}
+				quantPositionClass="-bottom-1.5 -right-1.5"
+				quantSize="h-6 w-6"
+				size="h-12 w-12"
+			/>
+
+			<div class="min-w-0">
+				<div class="flex items-center gap-2">
+					<h1 class="truncate text-lg font-semibold">{details.id ?? modelId}</h1>
+
+					<ModelCapabilityIcons
+						gapClass="gap-2"
+						iconSize="h-4 w-4"
+						modalities={{ audio: false, video: false, vision: hasVision }}
+						supportsThinking={hasReasoning}
+						supportsToolUse={hasTools}
+					/>
+				</div>
+
+				{#if baseModels.length}
+					<div class="flex items-center gap-1">
+						<span class="truncate text-xs text-muted-foreground">{baseModels.join(', ')}</span>
+
+						<a
+							aria-label="View base model on HuggingFace"
+							class="shrink-0 text-muted-foreground transition-colors hover:text-foreground"
+							href={HuggingFaceService.getModelUrl(baseModels[0])}
+							rel="noopener noreferrer"
+							target="_blank"
+						>
+							<ExternalLink class="h-3 w-3" />
+						</a>
+					</div>
+				{/if}
+			</div>
+		</div>
+
+		<a
+			class="inline-flex shrink-0 items-center gap-1.5 rounded-md border px-2.5 py-1.5 text-xs font-medium transition-colors hover:bg-muted"
+			href={HuggingFaceService.getModelUrl(modelId)}
+			rel="noopener noreferrer"
+			target="_blank"
+		>
+			<img alt="" class="h-3.5 w-3.5" src="/recommended-mcp/huggingface.ico" />
+
+			View on Hugging Face
+
+			<ExternalLink class={ICON_CLASS_SM} />
+		</a>
+	</div>
+
+	<ModelsDiscoverDetailsHfHubStats {details} />
+
+	<ModelsDiscoverDetailsMetadata {details} {gguf} {licenseTag} {modelId} />
+</header>

tools/ui/src/lib/components/app/models/discover/ModelsDiscoverDetails/ModelsDiscoverDetailsHfHubStats.svelte

@@ -0,0 +1,31 @@
+<script lang="ts">
+	import { Download, Heart } from '@lucide/svelte';
+	import { HuggingFaceService } from '$lib/services';
+	import type { HfModelDetailInfo } from '$lib/types/huggingface';
+
+	interface Props {
+		details: HfModelDetailInfo;
+	}
+
+	let { details }: Props = $props();
+</script>
+
+<div class="flex flex-wrap items-center gap-x-4 gap-y-1 text-sm text-muted-foreground">
+	{#if typeof details.downloads === 'number'}
+		<span class="inline-flex items-center gap-1.5">
+			<Download class="h-3.5 w-3.5" />
+			{HuggingFaceService.formatDownloads(details.downloads)}
+		</span>
+	{/if}
+
+	{#if typeof details.likes === 'number'}
+		<span class="inline-flex items-center gap-1.5">
+			<Heart class="h-3.5 w-3.5" />
+			{HuggingFaceService.formatLikes(details.likes)}
+		</span>
+	{/if}
+
+	{#if details.lastModified}
+		<span>Updated {HuggingFaceService.formatRelativeTime(details.lastModified)}</span>
+	{/if}
+</div>

tools/ui/src/lib/components/app/models/discover/ModelsDiscoverDetails/ModelsDiscoverDetailsMetadata.svelte

@@ -0,0 +1,80 @@
+<script lang="ts">
+	import ModelsDiscoverChatTemplateDialog from './ModelsDiscoverChatTemplateDialog.svelte';
+	import ModelsDiscoverDetailsMetadataItem from './ModelsDiscoverDetailsMetadataItem.svelte';
+	import { MessageSquareCode } from '@lucide/svelte';
+	import { modelsDiscoverStore } from '$lib/stores';
+	import type { HfModelDetailInfo, HfModelGguf } from '$lib/types/huggingface';
+	import { formatParameters } from '$lib/utils';
+
+	interface Props {
+		/** Full HuggingFace model id, e.g. `ggml-org/gemma-3-4b-it-GGUF`. */
+		modelId: string;
+		details: HfModelDetailInfo;
+		gguf?: HfModelGguf;
+		licenseTag: string | null;
+	}
+
+	let { details, gguf, licenseTag, modelId }: Props = $props();
+
+	// Catalog family description when curated, else the HF card description.
+	let description = $derived(
+		modelsDiscoverStore.descriptionFor(modelId) ?? details.cardData?.description
+	);
+
+	let chatTemplateOpen = $state(false);
+</script>
+
+{#if description}
+	<p class="text-sm text-muted-foreground">{description}</p>
+{/if}
+
+<!-- Metadata chips: label | value pairs, matching the HF model page style -->
+<div class="flex flex-wrap items-center gap-1.5">
+	{#if gguf?.total}
+		<ModelsDiscoverDetailsMetadataItem
+			label="Model size"
+			value="{formatParameters(gguf.total)} params"
+		/>
+	{/if}
+
+	{#if gguf?.context_length}
+		<ModelsDiscoverDetailsMetadataItem
+			label="Context"
+			value={gguf.context_length.toLocaleString()}
+		/>
+	{/if}
+
+	{#if gguf?.architecture}
+		<ModelsDiscoverDetailsMetadataItem label="Architecture" value={gguf.architecture} />
+	{/if}
+
+	{#if gguf?.chat_template}
+		<button
+			class="inline-flex items-center gap-1.5 rounded-md border px-2.5 py-1 text-xs font-medium transition-colors hover:bg-muted"
+			onclick={() => (chatTemplateOpen = true)}
+			type="button"
+		>
+			<MessageSquareCode class="h-3 w-3" />
+			Chat template
+		</button>
+	{/if}
+
+	{#if licenseTag}
+		<ModelsDiscoverDetailsMetadataItem label="License" value={licenseTag} />
+	{/if}
+
+	{#if details.gated === true}
+		<span
+			class="rounded bg-yellow-500/10 px-2 py-0.5 text-xs font-medium text-yellow-600 dark:text-yellow-400"
+		>
+			gated
+		</span>
+	{/if}
+</div>
+
+{#if gguf?.chat_template}
+	<ModelsDiscoverChatTemplateDialog
+		bind:open={chatTemplateOpen}
+		chatTemplate={gguf.chat_template}
+	/>
+{/if}

tools/ui/src/lib/components/app/models/discover/ModelsDiscoverDetails/ModelsDiscoverDetailsMetadataItem.svelte

@@ -0,0 +1,15 @@
+<script lang="ts">
+	interface Props {
+		label: string;
+		value: string;
+	}
+
+	let { label, value }: Props = $props();
+</script>
+
+<!-- Metadata chip: label | value pair, matching the HF model page style -->
+<span class="inline-flex items-center divide-x divide-border rounded-md border text-xs">
+	<span class="px-2.5 py-1 text-muted-foreground">{label}</span>
+
+	<span class="px-2.5 py-1 font-medium">{value}</span>
+</span>

tools/ui/src/lib/components/app/models/discover/ModelsDiscoverDetails/ModelsDiscoverDetailsReadme.svelte

@@ -0,0 +1,17 @@
+<script lang="ts">
+	import { MarkdownContent } from '$lib/components/app';
+
+	interface Props {
+		readme: string | null;
+	}
+
+	let { readme }: Props = $props();
+</script>
+
+{#if readme}
+	<section
+		class="p-3 rounded-3xl border border-border/30 bg-muted/60 shadow-xs dark:border-border/20"
+	>
+		<MarkdownContent allowHtml class="prose-sm max-w-none" content={readme} />
+	</section>
+{/if}

tools/ui/src/lib/components/app/models/discover/ModelsDiscoverDetails/ModelsDiscoverDetailsSkeleton.svelte

@@ -0,0 +1,93 @@
+<script lang="ts">
+	import { Skeleton } from '$lib/components/ui/skeleton';
+</script>
+
+<!-- Static skeleton of ModelsDiscoverDetails: header, metadata chips, download options and readme lines. -->
+<div class="space-y-6 p-6" data-slot="model-details-skeleton">
+	<!-- Header: avatar, name with capability icons, base model line, HF link -->
+	<div class="space-y-3">
+		<div class="flex items-start justify-between gap-3">
+			<div class="flex min-w-0 items-center gap-2">
+				<Skeleton class="h-12 w-12 rounded-md" />
+
+				<div class="min-w-0 space-y-1.5">
+					<Skeleton class="h-5 w-56 max-w-full" />
+
+					<Skeleton class="h-3 w-40 max-w-full" />
+				</div>
+			</div>
+
+			<Skeleton class="h-7.5 w-44 shrink-0 rounded-md" />
+		</div>
+
+		<!-- downloads / likes / last updated -->
+		<div class="flex items-center gap-4">
+			<Skeleton class="h-3.5 w-16" />
+
+			<Skeleton class="h-3.5 w-12" />
+
+			<Skeleton class="h-3.5 w-24" />
+		</div>
+
+		<!-- metadata chips -->
+		<div class="space-y-3">
+			<div class="flex flex-wrap items-center gap-1.5">
+				<Skeleton class="h-7 w-36 rounded-md" />
+
+				<Skeleton class="h-7 w-28 rounded-md" />
+
+				<Skeleton class="h-7 w-32 rounded-md" />
+
+				<Skeleton class="h-7 w-24 rounded-md" />
+			</div>
+		</div>
+	</div>
+
+	<!-- Download options: quant rows, CTA, terminal command -->
+	<div
+		class="rounded-3xl border border-border/30 bg-muted/40 p-4 shadow-xs dark:border-border/20 dark:bg-muted/50"
+	>
+		<div class="space-y-3 divide-y divide-border/50 dark:divide-border/35 pb-1">
+			{#each [0, 1, 2] as _, index (index)}
+				<div class="grid grid-cols-[5rem_1fr] items-start gap-3 py-3">
+					<div class="space-y-1.5 pt-1">
+						<Skeleton class="h-4 w-14" />
+
+						<Skeleton class="h-2.5 w-24" />
+					</div>
+
+					<div class="flex flex-wrap items-center justify-end gap-1.5">
+						<Skeleton class="h-6 w-28 rounded-md" />
+
+						<Skeleton class="h-6 w-20 rounded-md" />
+					</div>
+				</div>
+			{/each}
+		</div>
+
+		<div class="space-y-2.5 border-t border-border/50 pt-3.5 dark:border-border/35">
+			<Skeleton class="h-9 w-full rounded-md" />
+
+			<div class="flex items-center gap-3">
+				<span class="h-px flex-1 bg-border/50"></span>
+
+				<Skeleton class="h-3 w-44" />
+
+				<span class="h-px flex-1 bg-border/50"></span>
+			</div>
+
+			<Skeleton class="h-11 w-full rounded-md" />
+		</div>
+	</div>
+
+	<!-- Readme -->
+	<div class="space-y-2.5">
+		<Skeleton class="h-3.5 w-full" />
+
+		<Skeleton class="h-3.5 w-full" />
+
+		<Skeleton class="h-3.5 w-5/6" />
+
+		<Skeleton class="h-3.5 w-2/3" />
+	</div>
+</div>

tools/ui/src/lib/components/app/models/discover/ModelsDiscoverDetails/index.ts

@@ -0,0 +1,72 @@
+/**
+ *
+ * MODELS DISCOVER - DETAILS
+ *
+ * The right-hand detail pane of the discover view: header (avatar, name, stats, metadata
+ * chips, capability badges), the download options area and the model-card README.
+ *
+ */
+
+/**
+ * **ModelsDiscoverDetails** - Model detail view
+ *
+ * Detail pane for the selected model. Loads its own data (details + GGUF file
+ * list) from HuggingFaceService based on the `modelId` route param.
+ */
+export { default as ModelsDiscoverDetails } from './ModelsDiscoverDetails.svelte';
+
+/**
+ * **ModelsDiscoverDetailsHeader** - Detail view header
+ *
+ * Shows the model avatar (base org + quant org corner badge), name, base model
+ * info, stats, metadata chips and capability badges.
+ */
+export { default as ModelsDiscoverDetailsHeader } from './ModelsDiscoverDetailsHeader.svelte';
+
+/**
+ * **ModelsDiscoverDetailsHfHubStats** - HuggingFace Hub stats row
+ *
+ * Downloads, likes and last-updated for the viewed model, formatted the way
+ * the Hub shows them.
+ */
+export { default as ModelsDiscoverDetailsHfHubStats } from './ModelsDiscoverDetailsHfHubStats.svelte';
+
+/**
+ * **ModelsDiscoverDetailsSkeleton** - Detail view loading skeleton
+ *
+ * Static placeholder matching the detail layout: header with avatar and name,
+ * metadata chips, the download options box and readme text lines.
+ */
+export { default as ModelsDiscoverDetailsSkeleton } from './ModelsDiscoverDetailsSkeleton.svelte';
+
+/**
+ * **ModelsDiscoverDetailsReadme** - Detail view README
+ *
+ * Renders the model card README as markdown.
+ */
+export { default as ModelsDiscoverDetailsReadme } from './ModelsDiscoverDetailsReadme.svelte';
+
+/**
+ * **ModelsDiscoverDetailsMetadata** - Detail view metadata row
+ *
+ * The detail view's metadata chips (model size, context, architecture, license,
+ * chat template).
+ */
+export { default as ModelsDiscoverDetailsMetadata } from './ModelsDiscoverDetailsMetadata.svelte';
+
+/**
+ * **ModelsDiscoverDetailsMetadataItem** - Single metadata chip
+ *
+ * One label | value chip of the detail view's metadata row (model size,
+ * context, architecture, license).
+ */
+export { default as ModelsDiscoverDetailsMetadataItem } from './ModelsDiscoverDetailsMetadataItem.svelte';
+
+/**
+ * **ModelsDiscoverChatTemplateDialog** - Chat template viewer
+ *
+ * Shows the model's chat template in a scrollable dialog with a copy button.
+ */
+export { default as ModelsDiscoverChatTemplateDialog } from './ModelsDiscoverChatTemplateDialog.svelte';
+
+export * from './ModelsDiscoverDetailsDownloadOptions';

tools/ui/src/lib/components/app/models/discover/ModelsDiscoverDownloadProgressBar.svelte

@@ -0,0 +1,32 @@
+<script lang="ts">
+	interface Props {
+		downloadedBytes: number;
+		totalBytes: number;
+		overlay?: boolean;
+	}
+
+	let { downloadedBytes, overlay = false, totalBytes }: Props = $props();
+
+	let fraction = $derived.by(() => {
+		if (totalBytes <= 0) return 0;
+
+		return Math.min(Math.max(downloadedBytes / totalBytes, 0), 1);
+	});
+	let percent = $derived(Math.round(fraction * 100));
+</script>
+
+{#if overlay}
+	<div class="pointer-events-none absolute inset-x-0 bottom-0 h-0.5 overflow-hidden rounded-b-sm">
+		<div
+			class="h-full animate-pulse bg-primary transition-[width] duration-200 ease-out"
+			style="width: {percent}%"
+		></div>
+	</div>
+{:else}
+	<div class="h-1 w-full overflow-hidden rounded-full bg-muted">
+		<div
+			class="h-full animate-pulse bg-primary transition-[width] duration-200 ease-out"
+			style="width: {percent}%"
+		></div>
+	</div>
+{/if}

tools/ui/src/lib/components/app/models/discover/ModelsDiscoverList/ModelsDiscoverList.svelte

@@ -0,0 +1,41 @@
+<script lang="ts">
+	import ModelsDiscoverListItem from './ModelsDiscoverListItem.svelte';
+	import ModelsDiscoverListItemSkeleton from './ModelsDiscoverListItemSkeleton.svelte';
+	import type { HfModelInfo } from '$lib/types/huggingface';
+
+	interface Props {
+		models: HfModelInfo[];
+		activeId?: string | null;
+		loading?: boolean;
+		loadingSkeletonRowsCount?: number;
+		/** Show the original (base) model's org avatar instead of the repo's org. */
+		showBaseModelAvatar?: boolean;
+		onSelect?: (modelId: string) => void;
+	}
+
+	let {
+		activeId = null,
+		loading = false,
+		loadingSkeletonRowsCount = 8,
+		models,
+		onSelect,
+		showBaseModelAvatar = false
+	}: Props = $props();
+</script>
+
+<ul class="space-y-0.5 p-2">
+	{#if loading}
+		{#each Array(loadingSkeletonRowsCount) as _, index (index)}
+			<ModelsDiscoverListItemSkeleton {index} />
+		{/each}
+	{:else}
+		{#each models as model (model.id)}
+			<ModelsDiscoverListItem
+				active={model.id === activeId}
+				{model}
+				{onSelect}
+				{showBaseModelAvatar}
+			/>
+		{/each}
+	{/if}
+</ul>

tools/ui/src/lib/components/app/models/discover/ModelsDiscoverList/ModelsDiscoverListItem.svelte

@@ -0,0 +1,129 @@
+<script lang="ts">
+	import ModelId from '../../ModelId.svelte';
+	import ModelsDiscoverAvatar from '../ModelsDiscoverAvatar.svelte';
+	import {
+		HF_MMPROJ_FILENAME_TOKEN,
+		HF_MODALITY_PIPELINE_TAGS,
+		type ModelSidecar
+	} from '$lib/constants';
+	import { HuggingFaceService } from '$lib/services';
+	import { modelsDiscoverStore } from '$lib/stores';
+	import type { ModelsDiscoverSizeRange } from '$lib/stores/models-discover/index.svelte';
+	import type { HfModelInfo } from '$lib/types/huggingface';
+	import type { ModelModalities } from '$lib/types/models';
+	import { detectThinkingSupport, detectToolUseSupport, isAuxSidecar, orgOf } from '$lib/utils';
+	import { SvelteSet } from 'svelte/reactivity';
+
+	interface Props {
+		model: HfModelInfo;
+		active?: boolean;
+		/** Show the original (base) model's org avatar instead of the repo's org. */
+		showBaseModelAvatar?: boolean;
+		onSelect?: (modelId: string) => void;
+	}
+
+	let { active = false, model, onSelect, showBaseModelAvatar = false }: Props = $props();
+
+	let org = $derived(orgOf(model.id));
+
+	// Org whose avatar is shown: the base model's org when showBaseModelAvatar
+	// (e.g. the Qwen logo for ggml-org/Qwen3.8-27B-GGUF), else the repo's org.
+	let avatarOrg = $derived.by(() => {
+		if (!showBaseModelAvatar) return org;
+
+		return orgOf(HuggingFaceService.getBaseModels(model)[0]) || org;
+	});
+
+	let contextLength = $derived(model.gguf?.context_length);
+
+	// Reasoning support from the chat template, matching the details view.
+	let supportsThinking = $derived(detectThinkingSupport(model.gguf?.chat_template ?? ''));
+
+	// Tool use support from the chat template.
+	let supportsToolUse = $derived(detectToolUseSupport(model.gguf?.chat_template ?? ''));
+
+	// Modalities derived from HF metadata: vision from an mmproj sidecar or a
+	// multimodal pipeline tag, audio/video from their pipeline tags.
+	let modalities = $derived.by<ModelModalities>(() => {
+		const tag = model.pipeline_tag ?? '';
+		const vision =
+			HF_MODALITY_PIPELINE_TAGS.vision.includes(tag) ||
+			Boolean(
+				model.siblings?.some((s) => s.rfilename.toLowerCase().includes(HF_MMPROJ_FILENAME_TOKEN))
+			);
+		const audio = HF_MODALITY_PIPELINE_TAGS.audio.includes(tag);
+		const video = HF_MODALITY_PIPELINE_TAGS.video.includes(tag);
+
+		return { audio, video, vision };
+	});
+
+	// Draft sidecars (mtp, dflash, dspark, eagle3) present in the repo, e.g.
+	// speculative-decoding drafts. mmproj is excluded: it is vision, already
+	// conveyed by the modalities.
+	let draftSidecars = $derived.by<ModelSidecar[]>(() => {
+		const set = new SvelteSet<ModelSidecar>();
+
+		for (const sibling of model.siblings ?? []) {
+			const sidecar = HuggingFaceService.extractQuantMeta(sibling.rfilename)?.sidecar;
+
+			if (sidecar && !isAuxSidecar(sidecar)) set.add(sidecar);
+		}
+
+		return [...set];
+	});
+
+	// Min/max size across the repo's quants, draft sidecars included. The store
+	// has catalog rows covered already; any other row (a search hit) measures
+	// its repo once here and the result is cached per repo.
+	let measuredSize = $state<ModelsDiscoverSizeRange | null>(null);
+	let sizeRange = $derived(modelsDiscoverStore.cachedSizeRangeFor(model.id) ?? measuredSize);
+
+	$effect(() => {
+		const id = model.id;
+
+		if (modelsDiscoverStore.cachedSizeRangeFor(id)) return;
+
+		let cancelled = false;
+
+		void modelsDiscoverStore.sizeRange(id).then((range) => {
+			if (!cancelled) measuredSize = range ?? null;
+		});
+
+		return () => {
+			cancelled = true;
+		};
+	});
+</script>
+
+<li>
+	<button
+		aria-current={active ? 'page' : undefined}
+		class="flex w-full cursor-pointer items-start gap-2.5 rounded-lg p-2.5 text-left transition-colors {active
+			? 'bg-primary/10 hover:bg-primary/15'
+			: 'hover:bg-muted/60'}"
+		onclick={() => onSelect?.(model.id)}
+		type="button"
+	>
+		<ModelsDiscoverAvatar
+			class="mt-1"
+			org={avatarOrg}
+			quantOrg={showBaseModelAvatar ? org : undefined}
+		/>
+
+		<span class="min-w-0 flex-1">
+			<ModelId
+				class="min-w-0"
+				{contextLength}
+				{draftSidecars}
+				hideOrgName
+				iconsOnNewLine
+				{modalities}
+				modelId={model.id}
+				{sizeRange}
+				{supportsThinking}
+				{supportsToolUse}
+				wrap
+			/>
+		</span>
+	</button>
+</li>

tools/ui/src/lib/components/app/models/discover/ModelsDiscoverList/ModelsDiscoverListItemSkeleton.svelte

@@ -0,0 +1,39 @@
+<script lang="ts">
+	import { Skeleton } from '$lib/components/ui/skeleton';
+
+	interface Props {
+		/** Row index; varies skeleton widths so the list does not look mechanical. */
+		index?: number;
+	}
+
+	let { index = 0 }: Props = $props();
+
+	// Deterministic width variations keyed by row position.
+	const NAME_WIDTHS = ['w-40', 'w-52', 'w-44', 'w-56'];
+	const BADGE_WIDTHS = [
+		['w-12', 'w-14', 'w-10'],
+		['w-16', 'w-12', 'w-12'],
+		['w-10', 'w-16', 'w-10'],
+		['w-14', 'w-10', 'w-14']
+	];
+
+	let nameWidth = $derived(NAME_WIDTHS[index % NAME_WIDTHS.length]);
+	let badgeWidths = $derived(BADGE_WIDTHS[index % BADGE_WIDTHS.length]);
+</script>
+
+<!-- Static skeleton of ModelsDiscoverListItem: avatar, name and badge rows. -->
+<li>
+	<div class="flex w-full items-start gap-2.5 rounded-lg p-2.5 text-left">
+		<Skeleton class="h-9 w-9 shrink-0 rounded-md" />
+
+		<div class="min-w-0 flex-1 space-y-1.5">
+			<Skeleton class="{nameWidth} h-4 max-w-full" />
+
+			<div class="flex items-center gap-1">
+				{#each badgeWidths as width, i (i)}
+					<Skeleton class="{width} h-3.5 rounded" />
+				{/each}
+			</div>
+		</div>
+	</div>
+</li>

tools/ui/src/lib/components/app/models/discover/ModelsDiscoverList/ModelsDiscoverListSearch.svelte

@@ -0,0 +1,22 @@
+<script lang="ts">
+	import { SearchInput } from '$lib/components/app';
+
+	interface Props {
+		value?: string;
+		/** Search callback; the parent owns debouncing. */
+		onSearch?: (query: string) => void;
+		placeholder?: string;
+	}
+
+	let { onSearch, placeholder = 'Search models...', value = $bindable('') }: Props = $props();
+
+	function handleInput(next: string) {
+		value = next;
+
+		onSearch?.(value);
+	}
+</script>
+
+<div class="sticky top-0 z-99 p-2">
+	<SearchInput bind:value onInput={(v) => handleInput(v)} {placeholder} />
+</div>

tools/ui/src/lib/components/app/models/discover/ModelsDiscoverList/index.ts

@@ -0,0 +1,38 @@
+/**
+ *
+ * MODELS DISCOVER - LIST
+ *
+ * The discover sidebar column: a debounced search field above a navigable list of
+ * HuggingFace GGUF models, with skeleton rows while loading.
+ *
+ */
+
+/**
+ * **ModelsDiscoverList** - Sidebar model list
+ *
+ * Renders the discover model list as a navigable column. Each row links to the
+ * model's detail and highlights the active one.
+ */
+export { default as ModelsDiscoverList } from './ModelsDiscoverList.svelte';
+
+/**
+ * **ModelsDiscoverListSearch** - Sidebar search input
+ *
+ * Debounced search field for the model list.
+ */
+export { default as ModelsDiscoverListSearch } from './ModelsDiscoverListSearch.svelte';
+
+/**
+ * **ModelsDiscoverListItem** - Single sidebar row
+ *
+ * One model entry in the discover sidebar list, selectable via `onSelect`.
+ */
+export { default as ModelsDiscoverListItem } from './ModelsDiscoverListItem.svelte';
+
+/**
+ * **ModelsDiscoverListItemSkeleton** - Skeleton sidebar row
+ *
+ * Pulsing placeholder matching a ModelsDiscoverListItem row, shown while the
+ * list is loading.
+ */
+export { default as ModelsDiscoverListItemSkeleton } from './ModelsDiscoverListItemSkeleton.svelte';

tools/ui/src/lib/components/app/models/discover/index.ts

@@ -0,0 +1,39 @@
+/**
+ *
+ * MODELS DISCOVER
+ *
+ * Components for the Models Discover view: a sidebar search + list of
+ * HuggingFace GGUF models and a detail view for the selected model, used as the
+ * body of the discovery dialog. The list and detail trees live in their own
+ * subfolders; this barrel re-exports them alongside the shared leaves.
+ *
+ */
+
+/**
+ * **ModelsDiscover** - Models discover explorer
+ *
+ * The complete discovery layout: a sidebar search + model list on the left and a
+ * detail view for the selected model on the right. Used as the body of the
+ * discovery dialog.
+ */
+export { default as ModelsDiscover } from './ModelsDiscover.svelte';
+
+/**
+ * **ModelsDiscoverAvatar** - Org avatar for a model row
+ *
+ * Shows the org's avatar image, falling back to a monogram on a stable hue
+ * derived from the org name when the image fails to load. Shared by the list,
+ * the detail header and the model selector rows.
+ */
+export { default as ModelsDiscoverAvatar } from './ModelsDiscoverAvatar.svelte';
+
+/**
+ * **ModelsDiscoverDownloadProgressBar** - Thin download progress bar
+ *
+ * Normalizes bytes to a 0..100% bar; can pin to the bottom edge as an overlay.
+ * Shared by the quant chips and the model selector's download rows.
+ */
+export { default as ModelsDiscoverDownloadProgressBar } from './ModelsDiscoverDownloadProgressBar.svelte';
+
+export * from './ModelsDiscoverList';
+export * from './ModelsDiscoverDetails';

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

@@ -109,3 +109,12 @@ export { default as ModelBadge } from './ModelBadge.svelte';
  * Respects the user's `showRawModelNames` setting.
  */
 export { default as ModelId } from './ModelId.svelte';
+
+/**
+ * **ModelCapabilityIcons** - Capability and modality icon row
+ *
+ * The shared tool-use / reasoning / vision / video / audio icon cluster with
+ * tooltips, used by ModelId and the discover details header so the order and
+ * styling stay consistent across every model-id surface.
+ */
+export { default as ModelCapabilityIcons } from './ModelCapabilityIcons.svelte';

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

@@ -5,6 +5,7 @@
 	import {
 		ActionIcon,
 		DialogConversationRename,
+		DialogModelsDiscover,
 		DialogSettingsChat,
 		Logo,
 		SidebarNavigationActions,
@@ -93,6 +94,7 @@
 
 	let renameDialogOpen = $state(false);
 	let settingsDialogOpen = $state(false);
+	let modelsDiscoverOpen = $state(false);
 	let renameTargetConversationId = $state<string | null>(null);
 	let renameDraft = $state('');
 	let renameOriginalTitle = $state('');
@@ -389,6 +391,7 @@
 				bind:searchQuery
 				class="px-2"
 				isExpandedMode={innerWidth > 768 ? uiStore.isSidebarExpanded : true}
+				onDiscoverModelsClick={() => (modelsDiscoverOpen = true)}
 				onNewChat={() => {
 					if (deviceStore.isMobile) {
 						scheduleMobileCollapse();
@@ -452,6 +455,8 @@
 
 <DialogSettingsChat bind:open={settingsDialogOpen} />
 
+<DialogModelsDiscover bind:open={modelsDiscoverOpen} />
+
 <style>
 	aside {
 		@media (max-width: 768px) {

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

@@ -12,7 +12,7 @@
 		SIDEBAR_ACTIONS_ITEMS
 	} from '$lib/constants';
 	import { SidebarAction, TooltipSide } from '$lib/enums';
-	import { conversationsStore, deviceStore } from '$lib/stores';
+	import { conversationsStore, deviceStore, serverStore } from '$lib/stores';
 	import type { Component } from 'svelte';
 	import { onMount } from 'svelte';
 	import { circIn } from 'svelte/easing';
@@ -26,6 +26,7 @@
 		onSearchDeactivated?: () => void;
 		onSearchClick?: () => void;
 		onNewChat?: () => void;
+		onDiscoverModelsClick?: () => void;
 		onSettingsClick?: () => void;
 	}
 
@@ -33,6 +34,7 @@
 		class: className,
 		isExpandedMode = false,
 		isSearchModeActive = $bindable(false),
+		onDiscoverModelsClick,
 		onNewChat,
 		onSearchClick,
 		onSearchDeactivated,
@@ -46,6 +48,14 @@
 
 	const isOnMobile = $derived(deviceStore.isMobile);
 
+	// Discover models needs the router's download endpoints; hide it in
+	// single-model mode instead of opening a dialog that can only toast.
+	const actionsItems = $derived(
+		serverStore.isRouterMode
+			? SIDEBAR_ACTIONS_ITEMS
+			: SIDEBAR_ACTIONS_ITEMS.filter((item) => item.action !== SidebarAction.DISCOVER_MODELS)
+	);
+
 	$effect(() => {
 		if (isSearchModeActive && searchInputRef) {
 			searchInputRef.focus();
@@ -57,7 +67,7 @@
 
 		setTimeout(() => {
 			initialized = true;
-		}, ICON_STRIP_TRANSITION_DELAY_MULTIPLIER * SIDEBAR_ACTIONS_ITEMS.length);
+		}, ICON_STRIP_TRANSITION_DELAY_MULTIPLIER * actionsItems.length);
 	});
 
 	function handleSearchModeDeactivate() {
@@ -107,7 +117,7 @@
 			? 'hidden pointer-events-none'
 			: ''}"
 	>
-		{#each SIDEBAR_ACTIONS_ITEMS as item, i (item.tooltip)}
+		{#each actionsItems as item, i (item.tooltip)}
 			{@const isActive = isItemActive(item)}
 			{@const isSearchOnMobile = item.icon === Search && deviceStore.isMobile}
 			{@const itemHref = isSearchOnMobile ? ROUTES.SEARCH : item.route}
@@ -117,16 +127,18 @@
 							onNewChat?.();
 							void conversationsStore.openNewChat();
 						}
-					: item.action === SidebarAction.SETTINGS
-						? () => onSettingsClick?.()
-						: item.route
-							? () => {
-									onNewChat?.();
-									goto(item.route!);
-								}
-							: isSearchOnMobile
-								? undefined
-								: onSearchClick}
+					: item.action === SidebarAction.DISCOVER_MODELS
+						? () => onDiscoverModelsClick?.()
+						: item.action === SidebarAction.SETTINGS
+							? () => onSettingsClick?.()
+							: item.route
+								? () => {
+										onNewChat?.();
+										goto(item.route!);
+									}
+								: isSearchOnMobile
+									? undefined
+									: onSearchClick}
 			{@const itemTransition = {
 				delay: !initialized ? i * ICON_STRIP_TRANSITION_DELAY_MULTIPLIER : 0,
 				duration: ICON_STRIP_TRANSITION_DURATION,
@@ -164,7 +176,7 @@
 	</div>
 {:else}
 	<div class="{className} flex-col gap-1 hidden md:flex">
-		{#each SIDEBAR_ACTIONS_ITEMS as item, i (item.tooltip)}
+		{#each actionsItems as item, i (item.tooltip)}
 			{@const isActive = isItemActive(item)}
 			{@const isSearchOnMobile = item.icon === Search && deviceStore.isMobile}
 			{@const itemOnClick =
@@ -173,16 +185,18 @@
 							onNewChat?.();
 							void conversationsStore.openNewChat();
 						}
-					: item.action === SidebarAction.SETTINGS
-						? () => onSettingsClick?.()
-						: item.route
-							? () => {
-									onNewChat?.();
-									goto(item.route!);
-								}
-							: isSearchOnMobile
-								? undefined
-								: onSearchClick}
+					: item.action === SidebarAction.DISCOVER_MODELS
+						? () => onDiscoverModelsClick?.()
+						: item.action === SidebarAction.SETTINGS
+							? () => onSettingsClick?.()
+							: item.route
+								? () => {
+										onNewChat?.();
+										goto(item.route!);
+									}
+								: isSearchOnMobile
+									? undefined
+									: onSearchClick}
 			{@const itemTransition = {
 				delay: !initialized ? i * ICON_STRIP_TRANSITION_DELAY_MULTIPLIER : 0,
 				duration: ICON_STRIP_TRANSITION_DURATION,

tools/ui/src/lib/components/ui/select/select-trigger.svelte

@@ -11,18 +11,24 @@
 		variant = 'default',
 		...restProps
 	}: WithoutChild<SelectPrimitive.TriggerProps> & {
-		size?: 'sm' | 'default';
+		size?: 'xs' | 'sm' | 'default';
 		variant?: 'default' | 'plain';
 	} = $props();
 
+	// Super small trigger: fits its selected value, for dense inline use.
+	const xsClasses =
+		"group flex h-6 w-fit items-center justify-between gap-1 rounded-md border border-input bg-transparent px-2 py-0 text-xs whitespace-nowrap outline-none select-none transition-colors focus-visible:border-ring focus-visible:ring-[3px] focus-visible:ring-ring/50 disabled:cursor-not-allowed disabled:opacity-50 data-[placeholder]:text-muted-foreground dark:bg-input/30 dark:hover:bg-input/50 [&_svg]:pointer-events-none [&_svg]:shrink-0 [&_svg:not([class*='text-'])]:text-muted-foreground";
+
 	const baseClasses = $derived(
 		variant === 'plain'
 			? "group inline-flex w-full items-center justify-end gap-2 whitespace-nowrap px-0 py-0 text-sm font-medium text-muted-foreground transition-colors focus-visible:outline-none focus-visible:ring-0 focus-visible:ring-offset-0 disabled:cursor-not-allowed disabled:opacity-50 data-[placeholder]:text-muted-foreground data-[size=default]:h-9 data-[size=sm]:h-8 [&_svg]:pointer-events-none [&_svg]:shrink-0 [&_svg:not([class*='size-'])]:size-3 [&_svg:not([class*='text-'])]:text-muted-foreground"
-			: "flex w-fit items-center justify-between gap-2 rounded-md border border-input bg-transparent px-3 py-2 text-sm whitespace-nowrap shadow-xs transition-[color,box-shadow] outline-none select-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-[placeholder]:text-muted-foreground data-[size=default]:h-9 data-[size=sm]:h-8 *:data-[slot=select-value]:line-clamp-1 *:data-[slot=select-value]:flex *:data-[slot=select-value]:items-center *:data-[slot=select-value]:gap-2 dark:bg-input/30 dark:hover:bg-input/50 dark:aria-invalid:ring-destructive/40 [&_svg]:pointer-events-none [&_svg]:shrink-0 [&_svg:not([class*='size-'])]:size-4 [&_svg:not([class*='text-'])]:text-muted-foreground"
+			: size === 'xs'
+				? xsClasses
+				: "flex w-fit items-center justify-between gap-2 rounded-md border border-input bg-transparent px-3 py-2 text-sm whitespace-nowrap shadow-xs transition-[color,box-shadow] outline-none select-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-[placeholder]:text-muted-foreground data-[size=default]:h-9 data-[size=sm]:h-8 *:data-[slot=select-value]:line-clamp-1 *:data-[slot=select-value]:flex *:data-[slot=select-value]:items-center *:data-[slot=select-value]:gap-2 dark:bg-input/30 dark:hover:bg-input/50 dark:aria-invalid:ring-destructive/40 [&_svg]:pointer-events-none [&_svg]:shrink-0 [&_svg:not([class*='size-'])]:size-4 [&_svg:not([class*='text-'])]:text-muted-foreground"
 	);
 
 	const chevronClasses = $derived(
-		variant === 'plain'
+		variant === 'plain' || size === 'xs'
 			? 'size-3 opacity-60 transition-transform group-data-[state=open]:-rotate-180'
 			: 'size-4 opacity-50'
 	);

tools/ui/src/lib/constants/api-endpoints.constants.ts

@@ -1,4 +1,7 @@
 export const API_MODELS = {
+	/** Download a model from HuggingFace (ROUTER mode, POST) or cancel/remove it (DELETE) */
+	DELETE: '/models',
+	DOWNLOAD: '/models',
 	LIST: '/v1/models',
 	LOAD: '/models/load',
 	SSE: '/models/sse',

tools/ui/src/lib/constants/cli-flags.constants.ts

@@ -2,6 +2,12 @@ export const CLI_FLAGS = {
 	AGENT: '--agent',
 	API_KEY: '--api-key',
 	MCP_PROXY: '--ui-mcp-proxy',
+	/** Multimodal projector path; unlocks vision/audio for the model. */
+	MMPROJ: '--mmproj',
+	/** Draft model weights path (long form); the router records it per model. */
+	MODEL_DRAFT: '--model-draft',
+	/** Draft model weights path (short form). */
+	MODEL_DRAFT_SHORT: '-md',
 	SLOTS: '--slots',
 	TOOLS: '--tools'
 } as const;

tools/ui/src/lib/constants/huggingface.constants.ts

@@ -0,0 +1,232 @@
+/**
+ * HuggingFace Hub constants.
+ *
+ * URLs, parsing regexes and formatting units for the HuggingFaceService.
+ * Reference: https://huggingface.co/docs/huggingface_hub/package_reference/hf_api
+ */
+
+// API endpoints
+
+export const HF_BASE_URL = 'https://huggingface.co';
+export const HF_API_MODELS_URL = `${HF_BASE_URL}/api/models`;
+export const HF_AVATARS_URL = `${HF_BASE_URL}/api/avatars`;
+
+// Query params
+
+export const HF_FULL_DETAIL_PARAM = 'full=true';
+export const HF_RECURSIVE_TREE_PARAM = 'recursive=true';
+/** Search filter that restricts results to repos containing GGUF files. */
+export const HF_GGUF_FILTER = 'gguf';
+/** Repeatable `expand` query param selecting fields on the list endpoint. */
+export const HF_EXPAND_PARAM = 'expand';
+/**
+ * Fields the model list endpoint omits by default but the discover list rows
+ * render: `gguf` (chat template, context length, param count) drives the
+ * reasoning / tool-use icons and the context badge, `siblings` the vision and
+ * draft-sidecar badges. Without them those parts of a row stay empty.
+ */
+export const HF_MODEL_LIST_EXPAND: readonly string[] = [
+	'author',
+	'downloads',
+	'gguf',
+	'lastModified',
+	'likes',
+	'pipeline_tag',
+	'siblings',
+	// `base_model:` tags, so search rows can show the base org's avatar as the
+	// main avatar with the quant org as the corner badge, like catalog rows.
+	'tags'
+];
+
+// Repo file conventions
+
+export const HF_MAIN_BRANCH = 'main';
+export const HF_README_FILENAME = 'README.md';
+export const HF_RAW_PATH = 'raw';
+export const HF_TREE_PATH = 'tree';
+
+// Pagination
+
+export const HF_LINK_NEXT_REGEX = /<([^>]+)>;\s*rel="next"/;
+/** `Link` response header carrying the next page URL for cursor pagination. */
+export const HF_LINK_HEADER = 'Link';
+
+// Fetch retry
+
+export const HF_RETRY_ATTEMPTS = 3;
+export const HF_RETRY_DELAY_MS = 1000;
+export const HF_HTTP_NOT_FOUND = 404;
+export const HF_HTTP_SERVER_ERROR_MIN = 500;
+
+// Search limits
+
+export const HF_DEFAULT_LIMIT = 50;
+/** Safety cap on `/tree` pagination: more pages means a misbehaving endpoint. */
+export const HF_TREE_MAX_PAGES = 10;
+export const HF_MAX_LIMIT = 100;
+
+// GGUF shard files
+
+/** Matches a split-shard GGUF file name, e.g. `Model-00001-of-00015.gguf`. */
+export const HF_SHARD_REGEX = /-(\d{5})-of-(\d{5})\.gguf$/i;
+/** Index (1-based) of the first shard in a split-shard set. */
+export const HF_FIRST_SHARD = 1;
+/** Zero-padded width of the shard index in a split-shard file name. */
+export const HF_SHARD_PAD_WIDTH = 5;
+
+// Quantization tokens
+
+/** `UD-` (Unsloth Dynamic) custom quantization prefix, e.g. `UD-Q4_K_XL`. */
+export const HF_UD_QUANT_PREFIX = 'UD';
+export const HF_UD_QUANT_PREFIX_REGEX = /^UD-/i;
+/**
+ * Segment marking an Unsloth `shared-` draft head that borrows the target
+ * model's embedding/output weights, e.g. `...-shared-Q4_K_M.gguf`.
+ */
+export const HF_SHARED_DRAFT_TOKEN = 'shared';
+/**
+ * Extracts the leading precision digits from a quant token, e.g.
+ * `Q4_K_XL` -> 4, `IQ2_XXS` -> 2, `TQ1_0` -> 1, `BF16` -> 16.
+ */
+export const HF_QUANT_PRECISION_REGEX = /^(?:I?Q|TQ|BF|F|MXFP)?(\d+)/i;
+
+// Model card tags
+
+/** Matches the `base_model:` tag (plain or `quantized:`), capturing the repo id. */
+export const HF_BASE_MODEL_TAG_REGEX = /^base_model:(?:quantized:)?(.+)$/;
+export const HF_LICENSE_TAG_PREFIX = 'license:';
+export const HF_GATED_TAG = 'gated';
+export const HF_GGUF_TAG = 'gguf';
+export const HF_SAFETENSORS_TAG = 'safetensors';
+
+// Pipeline tasks (logic use only - matching `pipeline_tag` values against tags)
+
+/**
+ * `pipeline_tag` values grouped by the input/output modality they imply, used
+ * to derive a discover row's modality icons. A tag in more than one group (e.g.
+ * `image-to-video`) lights up each modality it belongs to.
+ */
+export const HF_MODALITY_PIPELINE_TAGS: Readonly<
+	Record<'audio' | 'video' | 'vision', readonly string[]>
+> = {
+	audio: [
+		'audio-classification',
+		'audio-to-audio',
+		'automatic-speech-recognition',
+		'text-to-speech',
+		'voice-activity-detection'
+	],
+	video: ['text-to-video', 'image-to-video', 'video-to-video'],
+	vision: ['image-text-to-text', 'image-to-text', 'text-to-image', 'image-to-video']
+};
+
+/** Filename token marking an mmproj sidecar sibling (unlocks vision / audio). */
+export const HF_MMPROJ_FILENAME_TOKEN = 'mmproj';
+
+export const HF_TASK_TAGS: readonly string[] = [
+	'audio-classification',
+	'audio-to-audio',
+	'automatic-speech-recognition',
+	'conversational',
+	'depth-estimation',
+	'feature-extraction',
+	'fill-mask',
+	'image-classification',
+	'image-feature-extraction',
+	'image-segmentation',
+	'image-text-to-text',
+	'image-to-text',
+	'image-to-video',
+	'object-detection',
+	'question-answering',
+	'reinforcement-learning',
+	'robotics',
+	'sentence-similarity',
+	'summarization',
+	'text2text-generation',
+	'text-classification',
+	'text-generation',
+	'text-to-image',
+	'text-to-speech',
+	'text-to-video',
+	'token-classification',
+	'translation',
+	'video-to-video',
+	'voice-activity-detection',
+	'zero-shot-classification'
+];
+
+// Formatting
+
+export const BYTE = 1;
+export const KILOBYTE = 1_000;
+export const MEGABYTE = 1_000_000;
+export const GIGABYTE = 1_000_000_000;
+export const TERABYTE = 1_000_000_000_000;
+
+/**
+ * Matches a human size string (`177GB`, `1.2 TB`, `500MB`), capturing the
+ * numeric value and its unit suffix. Used by `parseSizeBytes`.
+ */
+export const HF_SIZE_STRING_REGEX = /^\s*([\d.]+)\s*([a-z]+)\s*$/i;
+
+/**
+ * Byte multiplier for a size suffix (`k` kilobyte, `m` megabyte, ...) as used by
+ * the llama.app catalog `size` strings, whose suffix is lowercase.
+ */
+export const HF_SIZE_SUFFIX_BYTES: Readonly<Record<string, number>> = {
+	b: BYTE,
+	g: GIGABYTE,
+	k: KILOBYTE,
+	m: MEGABYTE,
+	t: TERABYTE
+};
+
+export const BYTE_LABEL = 'B';
+export const KILOBYTE_LABEL = 'KB';
+export const MEGABYTE_LABEL = 'MB';
+export const GIGABYTE_LABEL = 'GB';
+
+/** Count suffixes for compact number formatting, e.g. `1.5K`, `2.0M`. */
+export const KILO_LABEL = 'K';
+export const MEGA_LABEL = 'M';
+export const GIGA_LABEL = 'B';
+
+// Relative time
+
+export const MS_PER_DAY = 1000 * 60 * 60 * 24;
+export const DAYS_PER_WEEK = 7;
+/** Rough month length in days, used to bucket relative timestamps. */
+export const DAYS_PER_MONTH = 30;
+export const DAYS_PER_YEAR = 365;
+
+export const TODAY_LABEL = 'Today';
+export const YESTERDAY_LABEL = 'Yesterday';
+export const DAYS_AGO_LABEL = 'days ago';
+export const WEEKS_AGO_LABEL = 'weeks ago';
+export const MONTHS_AGO_LABEL = 'months ago';
+export const YEARS_AGO_LABEL = 'years ago';
+
+// Cache paths
+
+/**
+ * Matches a local HF cache file path
+ * (`.../models--<org>--<name>/snapshots/<sha>/<file>`), capturing the repo
+ * directory name and the repo-relative file path.
+ */
+export const HF_CACHE_PATH_REGEX = /models--(.+?)\/snapshots\/[^/]+\/(.+)$/;
+/** Separator between org and name segments in an HF cache directory name. */
+export const HF_CACHE_DIR_SEPARATOR = '--';
+
+// README
+
+/** Matches a leading YAML frontmatter block (--- ... ---) in a markdown document. */
+export const HF_FRONTMATTER_REGEX = /^---\r?\n[\s\S]*?\r?\n---\r?\n?/;
+
+// Param counts
+
+/**
+ * Best-effort parameter count token in a model id/name, e.g. `27B` from
+ * `Qwen3.8-27B-GGUF` or `300M` from `embeddinggemma-300M-GGUF`.
+ */
+export const HF_PARAM_COUNT_REGEX = /(?:^|[^a-z0-9])(\d+(?:[._]\d+)?)\s*([bm])(?![a-z0-9])/i;

tools/ui/src/lib/constants/icons.constants.ts

@@ -10,7 +10,8 @@ import {
 	Image as ImageIcon,
 	Lightbulb as ReasoningIcon,
 	Mic as AudioIcon,
-	Video as VideoIcon
+	Video as VideoIcon,
+	Wrench as ToolUseIcon
 } from '@lucide/svelte';
 import { FileTypeCategory, ModelCapability, ModelModality } from '$lib/enums';
 import type { ModelCapabilities, ModelModalities } from '$lib/types/models';
@@ -49,16 +50,19 @@ export const MODALITY_FLAG_KEYS: Record<
 };
 
 export const CAPABILITY_ICONS: Record<ModelCapability, Component> = {
-	[ModelCapability.REASONING]: ReasoningIcon
+	[ModelCapability.REASONING]: ReasoningIcon,
+	[ModelCapability.TOOL_USE]: ToolUseIcon
 } as const;
 
 export const CAPABILITY_LABELS: Record<ModelCapability, string> = {
-	[ModelCapability.REASONING]: 'Reasoning'
+	[ModelCapability.REASONING]: 'Reasoning',
+	[ModelCapability.TOOL_USE]: 'Tool use'
 } as const;
 
 /** Maps a ModelCapability to the boolean flag it drives on the ModelCapabilities type */
 export const CAPABILITY_FLAG_KEYS: Record<ModelCapability, keyof ModelCapabilities> = {
-	[ModelCapability.REASONING]: 'reasoning'
+	[ModelCapability.REASONING]: 'reasoning',
+	[ModelCapability.TOOL_USE]: 'tools'
 };
 
 // Shared SVG icon strings for copy and preview buttons

tools/ui/src/lib/constants/index.ts

@@ -45,6 +45,10 @@ export * from './message-export.constants';
 export * from './path-display.constants';
 export * from './model-id.constants';
 export * from './model-loading.constants';
+export * from './models-discover.constants';
+export * from './models-discover-download.constants';
+export * from './model-compatibility.constants';
+export * from './huggingface.constants';
 export * from './precision.constants';
 export * from './pwa.constants';
 export * from './routes.constants';

tools/ui/src/lib/constants/model-compatibility.constants.ts

@@ -0,0 +1,32 @@
+/**
+ * Model memory-fit constants.
+ *
+ * Mirrors the app's compatibility check (Model+Compatibility.swift):
+ *   budget      = RAM x RAM_BUDGET_RATIO - RAM_OVERHEAD_MB
+ *   weightBytes = fileBytes x QUANT_WEIGHT
+ * a file fits when weightBytes <= budget. Kept here so the estimation util and
+ * any caller share one source.
+ */
+
+/** Bytes in one mebibyte (MiB), used to convert a file size to MB. */
+export const MIB_BYTES = 1_048_576;
+
+/** MB in one GB. */
+export const MB_PER_GB = 1024;
+
+/** Overhead multiplier applied to the file size when estimating weight memory. */
+export const QUANT_WEIGHT = 1.05;
+
+/** Share of RAM the app allows the model to occupy. */
+export const RAM_BUDGET_RATIO = 0.75;
+
+/** Fixed RAM overhead (MB) reserved for the system and KV cache. */
+export const RAM_OVERHEAD_MB = 2048;
+
+/**
+ * Memory tiers (GB) covering the RAM sizes common machines ship with, in
+ * small enough steps that the requirement reads honestly. Device-agnostic on
+ * purpose: the server exposes no host RAM, so the UI presents the tier and
+ * lets the user judge.
+ */
+export const MEM_TIERS = [4, 6, 8, 12, 16, 24, 32, 48, 64, 96, 128, 192, 256, 384, 512, 768, 1024];

tools/ui/src/lib/constants/model-id.constants.ts

@@ -2,35 +2,48 @@
  * Parsing of `org/ModelName[-tag][:quant]` style model IDs.
  */
 
+import { ModelAuxSidecar, ModelDraftSidecar } from '$lib/enums';
+
+/** Any sidecar file type: a draft variant or an auxiliary sidecar like mmproj. */
+export type ModelSidecar = ModelDraftSidecar | ModelAuxSidecar;
+
+/** All sidecar filename tokens: the bare lowercase enum values, e.g. `mtp`, `mmproj`. */
+export const SIDECAR_TOKENS: string[] = [
+	...Object.values(ModelDraftSidecar),
+	...Object.values(ModelAuxSidecar)
+];
+
+/** Separator between token alternatives in the sidecar regexes. */
+const REGEX_ALTERNATION_SEPARATOR = '|';
+const SIDECAR_TOKEN_ALTERNATION = SIDECAR_TOKENS.join(REGEX_ALTERNATION_SEPARATOR);
+
 export const MODEL_ID = {
 	/**
 	 * Matches an activated-parameter-count segment, e.g. `A10B`, `a2.4b`.
 	 * The leading `A`/`a` distinguishes it from a regular params segment.
 	 */
-	ACTIVATED_PARAMS_RE: /^[Aa]\d+(\.\d+)?[BbMmKkTt]$/,
+	ACTIVATED_PARAMS_REGEX: /^[Aa]\d+(\.\d+)?[BbMmKkTt]$/,
 
 	/** Matches prefix for custom quantization types, e.g. `UD-Q8_K_XL`. */
-	CUSTOM_QUANTIZATION_PREFIX_RE: /^UD$/i,
+	CUSTOM_QUANTIZATION_PREFIX_REGEX: /^UD$/i,
 	/** Container format segments to exclude from tags (every model uses these). */
 	IGNORED_SEGMENTS: new Set(['GGUF', 'GGML']),
 	/** Sentinel value returned by `indexOf` when a substring is not found. */
 	NOT_FOUND: -1,
-
 	/** Separates `<org>` from `<model>` in a model ID, e.g. `org/ModelName`. */
 	ORG_SEPARATOR: '/',
-
 	/**
 	 * Matches a parameter-count segment, e.g. `7B`, `1.5b`, `120M`.
 	 * The optional leading `E` covers effective-parameter sizes, e.g. Gemma's
 	 * `E2B`/`E4B` (MatFormer models sized by resident params).
 	 */
-	PARAMS_RE: /^[Ee]?\d+(\.\d+)?[BbMmKkTt]$/,
+	PARAMS_REGEX: /^[Ee]?\d+(\.\d+)?[BbMmKkTt]$/,
 
 	/**
 	 * Matches a quantization/precision segment, e.g. `Q4_K_M`, `IQ4_XS`, `F16`, `BF16`, `MXFP4`.
 	 * Case-insensitive to handle both uppercase and lowercase inputs.
 	 */
-	QUANTIZATION_SEGMENT_RE: /^(I?Q\d+(_[A-Z0-9]+)*|F\d+|BF\d+|MXFP\d+(_[A-Z0-9]+)*)$/i,
+	QUANTIZATION_SEGMENT_REGEX: /^(I?Q\d+(_[A-Z0-9]+)*|F\d+|BF\d+|MXFP\d+(_[A-Z0-9]+)*)$/i,
 
 	/** Separates the model path from the quantization tag, e.g. `model:Q4_K_M`. */
 	QUANTIZATION_SEPARATOR: ':',
@@ -38,6 +51,36 @@ export const MODEL_ID = {
 	/** Separates named segments within the model path, e.g. `ModelName-7B-GGUF`. */
 	SEGMENT_SEPARATOR: '-',
 
+	/**
+	 * Sidecar token between name segments, e.g. `Model-mtp-Q4_0.gguf`,
+	 * `model-eagle3-BF16.gguf`. Captures the name head and tail around the
+	 * token; same case-insensitive rule as the prefix form.
+	 */
+	SIDECAR_INFIX_REGEX: new RegExp(`^(.*)-(${SIDECAR_TOKEN_ALTERNATION})-(.+)$`, 'i'),
+
+	/**
+	 * Sidecar prefix that wraps a model id with a sidecar type, e.g.
+	 * `mtp-<name>.gguf`, `dflash-<name>.gguf`, `dspark-<name>.gguf`,
+	 * `eagle3-<name>.gguf`, `mmproj-<name>.gguf`. Captures the bare type
+	 * token for typed lookup.
+	 *
+	 * The token matches case-insensitively (real repos ship uppercase
+	 * heads, e.g. `Model-MTP-BF16.gguf`) and is normalized through
+	 * `sidecarFromFileToken`; the server's filename grammar
+	 * (common/download.cpp) matches the same segments.
+	 */
+	SIDECAR_PREFIX_REGEX: new RegExp(`^(${SIDECAR_TOKEN_ALTERNATION})-(.*)$`, 'i'),
+
+	/**
+	 * Trailing `-<type>` suffix marking a GGUF with an embedded draft in the
+	 * same weight file (MTP) or a sidecar download entry, e.g.
+	 * `Hy3-IQ1_M-mtp.gguf`, `Q4_K_M-dspark`. An optional `-draft` tail covers
+	 * standalone sidecar files, e.g. `Model-mtp-draft.gguf`. The captured
+	 * prefix is the candidate model id; the caller decides whether it looks
+	 * quantized. Case-insensitive, like the prefix form.
+	 */
+	SIDECAR_SUFFIX_REGEX: new RegExp(`^(.*)-(${SIDECAR_TOKEN_ALTERNATION})(-draft)?$`, 'i'),
+
 	/** Matches a trailing weight file extension, e.g. `model.gguf` -> `model`. */
-	WEIGHT_EXTENSION_RE: /\.(gguf|ggml)$/i
+	WEIGHT_EXTENSION_REGEX: /\.(gguf|ggml)$/i
 };

tools/ui/src/lib/constants/models-discover-download.constants.ts

@@ -0,0 +1,54 @@
+/**
+ * Models Discover - download options constants.
+ *
+ * Shared values backing the download-options area of the details pane: the
+ * serve `--spec-type` mapping, the bit-depth buckets, and the standalone
+ * command builder. Kept here (not component-local) so they sit with the rest
+ * of the discover domain constants.
+ */
+
+import type { ModelSidecar } from '$lib/constants';
+import { ModelAuxSidecar, ModelDraftSidecar } from '$lib/enums';
+
+/**
+ * `--spec-type` value for each draft sidecar; aux sidecars (mmproj, imatrix)
+ * carry none and stay empty.
+ */
+export const SPEC_TYPE: Record<ModelSidecar, string> = {
+	[ModelAuxSidecar.IMATRIX]: '',
+	[ModelAuxSidecar.MMPROJ]: '',
+	[ModelDraftSidecar.DFLASH]: 'draft-dflash',
+	[ModelDraftSidecar.DSPARK]: 'draft-dspark',
+	[ModelDraftSidecar.EAGLE3]: 'draft-eagle3',
+	[ModelDraftSidecar.MTP]: 'draft-mtp'
+};
+
+/** Bit-depth bucket for files that carry no quant token; rendered as "Other". */
+export const OTHER_BIT_DEPTH = 99;
+
+/** Label for the OTHER_BIT_DEPTH bucket. */
+export const OTHER_BIT_DEPTH_LABEL = 'Other';
+
+/** Suffix appended after a bit-depth number to label a row, e.g. `4-bit`. */
+export const BIT_DEPTH_LABEL_SUFFIX = '-bit';
+
+/**
+ * Fixed tokens of the standalone `llama serve` command preview, so the command
+ * builder and its rendered spans share one source of the CLI spelling.
+ */
+export const SERVE_COMMAND = {
+	BIN: 'llama',
+	/** Draft model repo:tag argument (download sidecar / draft weights). */
+	DRAFT_FLAG: '-hfd',
+	/** Main model repo:tag argument. */
+	MODEL_FLAG: '-hf',
+	/** Speculative-decoding type argument for the draft. */
+	SPEC_TYPE_FLAG: '--spec-type',
+	SUBCOMMAND: 'serve'
+} as const;
+
+/** Bit depth preferred for the command's default base quant. */
+export const DEFAULT_BASE_BIT_DEPTH = 4;
+
+/** Label shown for a draft-sidecar chip whose quant is unknown. */
+export const DRAFT_FILE_LABEL = 'draft';

tools/ui/src/lib/constants/models-discover.constants.ts

@@ -0,0 +1,14 @@
+/**
+ * Models discover constants.
+ *
+ * Endpoints and settings for the Models Discover dialog.
+ */
+
+/** llama.app model catalog used as the default model list. Online-only source; the discover feature requires an internet connection anyway. */
+export const MODELS_DISCOVER_CATALOG_URL = 'https://llama.app/v1/catalog.json';
+
+/**
+ * Catalog repos fetched in parallel per batch; small on purpose so the HF API
+ * is not hit with the whole catalog at once.
+ */
+export const MODELS_DISCOVER_CATALOG_BATCH = 4;

tools/ui/src/lib/constants/storage.constants.ts

@@ -15,6 +15,9 @@ export const STORAGE_APP_NAME_DEPRECATED = 'LlamaCppWebui';
 export const DB_APP_NAME_DEPRECATED = 'LlamacppWebui';
 
 export const ALWAYS_ALLOWED_TOOLS_LOCALSTORAGE_KEY = `${STORAGE_APP_NAME}.alwaysAllowedTools`;
+
+/** Paused model download ids (`<repo>:<tag>`), restored on the next page load. */
+export const PAUSED_MODEL_DOWNLOADS_LOCALSTORAGE_KEY = `${STORAGE_APP_NAME}.pausedModelDownloads`;
 export const CONFIG_LOCALSTORAGE_KEY = `${STORAGE_APP_NAME}.config`;
 export const DISABLED_TOOLS_LOCALSTORAGE_KEY = `${STORAGE_APP_NAME}.disabledTools`;
 

tools/ui/src/lib/constants/ui.constants.ts

@@ -1,4 +1,4 @@
-import { Package, Search, Settings, SquarePen } from '@lucide/svelte';
+import { Package, PackageSearch, Search, Settings, SquarePen } from '@lucide/svelte';
 import { SidebarAction, ToolSource } from '$lib/enums';
 import type { DesktopIconStripItem } from '$lib/types';
 
@@ -44,6 +44,9 @@ export const STATS_UNITS = {
 
 export const DEFAULT_MOBILE_BREAKPOINT = 768;
 
+/** Orgs whose avatar is dark and needs inverting in dark mode. */
+export const DARK_INVERT_AVATAR_ORGS = ['openai'];
+
 /** Icon used for the model selector and the `/model` slash command. */
 export const MODEL_SELECTOR_ICON = Package;
 
@@ -61,6 +64,11 @@ export const SIDEBAR_ACTIONS_ITEMS: DesktopIconStripItem[] = [
 		tooltip: 'New chat'
 	},
 	{ icon: Search, keys: ['cmd', 'k'], tooltip: 'Search' },
+	{
+		action: SidebarAction.DISCOVER_MODELS,
+		icon: PackageSearch,
+		tooltip: 'Discover models'
+	},
 	{
 		action: SidebarAction.SETTINGS,
 		icon: Settings,

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

@@ -0,0 +1,35 @@
+/**
+ * HuggingFace Hub enums.
+ *
+ * Values mirror the strings used by the HF REST API
+ * (https://huggingface.co/docs/huggingface_hub/package_reference/hf_api)
+ * so they can be sent and compared directly.
+ */
+
+/** Sort field for /api/models search queries. */
+export enum HfModelSort {
+	CREATED_AT = 'createdAt',
+	DOWNLOADS = 'downloads',
+	LAST_MODIFIED = 'lastModified',
+	LIKES = 'likes',
+	TRENDING_SCORE = 'trendingScore'
+}
+
+/**
+ * Where the sidecar token (`mtp` / `dflash` / `mmproj` / ...) sits in the
+ * filename.
+ * - `prefix`  sidecar file that lives next to the main weights, e.g. `mtp-Q4_0.gguf`
+ * - `suffix`  embedded draft baked into the main weights, e.g. `Hy3-IQ1_M-mtp.gguf`
+ * - `infix`   standalone sidecar named between head and quant, e.g. `model-mtp-Q8_0.gguf`
+ */
+export enum SidecarForm {
+	INFIX = 'infix',
+	PREFIX = 'prefix',
+	SUFFIX = 'suffix'
+}
+
+/** Entry type in a model repository file tree (`/tree` responses). */
+export enum HfEntryType {
+	DIRECTORY = 'directory',
+	FILE = 'file'
+}

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

@@ -57,6 +57,8 @@ export {
 	SpecialFileType
 } from './files.enums';
 
+export { HfEntryType, HfModelSort, SidecarForm } from './huggingface.enums';
+
 export {
 	MCPConnectionPhase,
 	MCPLogLevel,
@@ -67,7 +69,15 @@ export {
 	JsonSchemaType
 } from './mcp.enums';
 
-export { ModelCapability, ModelModality } from './model.enums';
+export {
+	ModelAuxSidecar,
+	ModelCapability,
+	ModelDraftSidecar,
+	ModelModality,
+	ModelSelectableFileKind
+} from './model.enums';
+
+export { ModelDownloadConfirmAction, ModelDownloadStopRequest } from './model.enums';
 
 export { ServerRole, ServerModelStatus, ServerModelsSseEventType } from './server.enums';
 

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

@@ -6,5 +6,62 @@ export enum ModelModality {
 }
 
 export enum ModelCapability {
-	REASONING = 'REASONING'
+	REASONING = 'reasoning',
+	TOOL_USE = 'tools'
+}
+
+/**
+ * Speculative-decoding draft sidecars (server spec-type draft-*).
+ * Filenames use the lowercase token, e.g. `mtp-<name>.gguf` or `-mtp` suffix.
+ */
+export enum ModelDraftSidecar {
+	/** DFlash block-diffusion draft (spec-type draft-dflash). */
+	DFLASH = 'dflash',
+	/** DSpark block-diffusion draft (spec-type draft-dspark). */
+	DSPARK = 'dspark',
+	/** EAGLE-3 speculative draft (spec-type draft-eagle3). */
+	EAGLE3 = 'eagle3',
+	/** Multi-token-prediction draft head (spec-type draft-mtp). */
+	MTP = 'mtp'
+}
+
+/**
+ * Non-draft sidecar file types. A sidecar is any auxiliary GGUF file
+ * accompanying the main model weights.
+ */
+export enum ModelAuxSidecar {
+	/** Importance-matrix data used to build imatrix quants; not loaded at serve time. */
+	IMATRIX = 'imatrix',
+	/** Multimodal projector: unlocks vision and/or audio input modalities. */
+	MMPROJ = 'mmproj'
+}
+
+/**
+ * Role of a selectable GGUF in the download options: the main weights, a
+ * speculative-decoding draft sidecar, or an auxiliary sidecar (mmproj).
+ */
+export enum ModelSelectableFileKind {
+	AUX = 'aux',
+	DRAFT = 'draft',
+	MAIN = 'main'
+}
+
+/**
+ * Why an in-flight download is being stopped, so the terminal `download_failed`
+ * feed event can be attributed: a user pause (resumable) or a user cancel
+ * (discard). Distinguishes these from a genuine download failure.
+ */
+export enum ModelDownloadStopRequest {
+	CANCEL = 'cancel',
+	PAUSE = 'pause'
+}
+
+/**
+ * Destructive download action the user is asked to confirm: stop and discard an
+ * in-flight download, or delete an already-downloaded model from disk. Both
+ * resolve through the same store removal call, differing only in the copy.
+ */
+export enum ModelDownloadConfirmAction {
+	CANCEL = 'cancel',
+	DELETE = 'delete'
 }