diff --git a/common/chat-auto-parser-generator.cpp b/common/chat-auto-parser-generator.cpp
index bf44091d6..3f036bb5b 100644
--- a/common/chat-auto-parser-generator.cpp
+++ b/common/chat-auto-parser-generator.cpp
@@ -65,7 +65,7 @@ common_chat_params peg_generator::generate_parser(const common_chat_template &
         data.grammar      = build_grammar([&](const common_grammar_builder & builder) {
             foreach_function(inputs.tools, [&](const json & tool) {
                 const auto & function = tool.at("function");
-                auto         schema   = function.at("parameters");
+                auto         schema   = function.contains("parameters") ? function.at("parameters") : json::object();
                 builder.resolve_refs(schema);
             });
             parser.build_grammar(builder, data.grammar_lazy);
@@ -221,7 +221,7 @@ common_peg_parser analyze_tools::build_tool_parser_tag_json(parser_build_context
     foreach_function(inputs.tools, [&](const json & tool) {
         const auto & func   = tool.at("function");
         std::string  name   = func.at("name");
-        const auto & schema = func.at("parameters");
+        const auto & schema = func.contains("parameters") ? func.at("parameters") : json::object();
 
         // Build call_id parser based on position (if supported)
         common_peg_parser call_id_section = p.eps();
@@ -282,19 +282,11 @@ common_peg_parser analyze_tools::build_tool_parser_tag_tagged(parser_build_conte
     common_peg_parser tool_choice = p.choice();
 
     foreach_function(inputs.tools, [&](const json & tool) {
-        const auto & func   = tool.at("function");
-        std::string  name   = func.at("name");
-        const auto & params = func.at("parameters");
-
-        if (!params.contains("properties") || !params.at("properties").is_object()) {
-            return;
-        }
-
-        const auto &          properties = params.at("properties");
+        const auto &          func       = tool.at("function");
+        std::string           name       = func.at("name");
+        const auto &          params     = func.contains("parameters") ? func.at("parameters") : json::object();
+        const auto &          properties = params.contains("properties") ? params.at("properties") : json::object();
         std::set<std::string> required;
-        if (params.contains("required") && params.at("required").is_array()) {
-            params.at("required").get_to(required);
-        }
 
         // Build parser for each argument, separating required and optional
         std::vector<common_peg_parser> required_parsers;
@@ -311,17 +303,18 @@ common_peg_parser analyze_tools::build_tool_parser_tag_tagged(parser_build_conte
                 }
             }
 
-            auto arg = p.tool_arg(
-                p.tool_arg_open(arguments.name_prefix + p.tool_arg_name(p.literal(param_name)) +
-                                arguments.name_suffix) +
-                arguments.value_prefix +
-                (type == "string" ? p.tool_arg_string_value(p.schema(p.until(arguments.value_suffix),
-                                                                     "tool-" + name + "-arg-" + param_name + "-schema",
-                                                                     param_schema, true)) :
-                                    p.tool_arg_json_value(p.schema(
-                                        p.json(), "tool-" + name + "-arg-" + param_name + "-schema", param_schema, false)) +
-                                        p.space()) +
-                p.tool_arg_close(p.literal(arguments.value_suffix)));
+            auto arg =
+                p.tool_arg(p.tool_arg_open(arguments.name_prefix + p.tool_arg_name(p.literal(param_name)) +
+                                           arguments.name_suffix) +
+                           arguments.value_prefix +
+                           (type == "string" ?
+                                p.tool_arg_string_value(p.schema(p.until(arguments.value_suffix),
+                                                                 "tool-" + name + "-arg-" + param_name + "-schema",
+                                                                 param_schema, true)) :
+                                p.tool_arg_json_value(p.schema(
+                                    p.json(), "tool-" + name + "-arg-" + param_name + "-schema", param_schema, false)) +
+                                    p.space()) +
+                           p.tool_arg_close(p.literal(arguments.value_suffix)));
 
             auto named_arg = p.rule("tool-" + name + "-arg-" + param_name, arg);
             if (is_required) {
diff --git a/common/common.cpp b/common/common.cpp
index 4ea4eaea1..f807f3b04 100644
--- a/common/common.cpp
+++ b/common/common.cpp
@@ -663,14 +663,53 @@ bool string_parse_kv_override(const char * data, std::vector<llama_model_kv_over
     return true;
 }
 
-// simple glob: * matches non-/ chars, ** matches anything including /
+static inline bool glob_class_match(const char c, const char * pattern, const char * class_end) {
+    const char * class_start = pattern;
+    bool negated = false;
+
+    if (*class_start == '!') {
+        negated = true;
+        class_start++;
+    }
+
+    // If first character after negation is ']' or '-', treat it as literal
+    if (*class_start == ']' || *class_start == '-') {
+        if (class_start < class_end && *class_start == c) {
+            return !negated;
+        }
+        class_start++;
+    }
+
+    bool matched = false;
+
+    while (class_start < class_end) {
+        if (class_start + 2 < class_end && class_start[1] == '-' && class_start[2] != ']') {
+            char start_char = *class_start;
+            char end_char = class_start[2];
+            if (c >= start_char && c <= end_char) {
+                matched = true;
+                break;
+            }
+            class_start += 3;
+        } else {
+            if (*class_start == c) {
+                matched = true;
+                break;
+            }
+            class_start++;
+        }
+    }
+
+    return negated ? !matched : matched;
+}
+
+// simple glob: * matches non-/ chars, ** matches anything including /, [] matches character class
 static inline bool glob_match(const char * pattern, const char * str) {
     if (*pattern == '\0') {
         return *str == '\0';
     }
     if (pattern[0] == '*' && pattern[1] == '*') {
         const char * p = pattern + 2;
-        if (*p == '/') p++;
         if (glob_match(p, str)) return true;
         if (*str != '\0') return glob_match(pattern, str + 1);
         return false;
@@ -685,6 +724,26 @@ static inline bool glob_match(const char * pattern, const char * str) {
     if (*pattern == '?' && *str != '\0' && *str != '/') {
         return glob_match(pattern + 1, str + 1);
     }
+    if (*pattern == '[') {
+        const char * class_end = pattern + 1;
+        // If first character after '[' is ']' or '-', treat it as literal
+        if (*class_end == ']' || *class_end == '-') {
+            class_end++;
+        }
+        while (*class_end != '\0' && *class_end != ']') {
+            class_end++;
+        }
+        if (*class_end == ']') {
+            if (*str == '\0') return false;
+            bool matched = glob_class_match(*str, pattern + 1, class_end);
+            return matched && glob_match(class_end + 1, str + 1);
+        } else {
+            if (*str == '[') {
+                return glob_match(pattern + 1, str + 1);
+            }
+            return false;
+        }
+    }
     if (*pattern == *str) {
         return glob_match(pattern + 1, str + 1);
     }
diff --git a/convert_hf_to_gguf.py b/convert_hf_to_gguf.py
index bcf98cfae..51f0d1ab1 100755
--- a/convert_hf_to_gguf.py
+++ b/convert_hf_to_gguf.py
@@ -31,10 +31,10 @@ import gguf
 from gguf.vocab import MistralTokenizerType, MistralVocab
 
 try:
-    from mistral_common.tokens.tokenizers.base import TokenizerVersion # type: ignore[import-not-found]
-    from mistral_common.tokens.tokenizers.multimodal import DATASET_MEAN as _MISTRAL_COMMON_DATASET_MEAN, DATASET_STD as _MISTRAL_COMMON_DATASET_STD # type: ignore[import-not-found]
-    from mistral_common.tokens.tokenizers.tekken import Tekkenizer # type: ignore[import-not-found]
-    from mistral_common.tokens.tokenizers.sentencepiece import ( # type: ignore[import-not-found]
+    from mistral_common.tokens.tokenizers.base import TokenizerVersion # type: ignore[import-not-found, ty:unresolved-import]
+    from mistral_common.tokens.tokenizers.multimodal import DATASET_MEAN as _MISTRAL_COMMON_DATASET_MEAN, DATASET_STD as _MISTRAL_COMMON_DATASET_STD # type: ignore[import-not-found, ty:unresolved-import]
+    from mistral_common.tokens.tokenizers.tekken import Tekkenizer # type: ignore[import-not-found, ty:unresolved-import]
+    from mistral_common.tokens.tokenizers.sentencepiece import ( # type: ignore[import-not-found, ty:unresolved-import]
         SentencePieceTokenizer,
     )
 
diff --git a/ggml/src/ggml-cuda/ggml-cuda.cu b/ggml/src/ggml-cuda/ggml-cuda.cu
index a389767fb..6d4ea5d07 100644
--- a/ggml/src/ggml-cuda/ggml-cuda.cu
+++ b/ggml/src/ggml-cuda/ggml-cuda.cu
@@ -2351,7 +2351,8 @@ static void ggml_cuda_mul_mat_id(ggml_backend_cuda_context & ctx, ggml_tensor *
         static_assert(MMVQ_MAX_BATCH_SIZE == MMVF_MAX_BATCH_SIZE);
         if (ne2 <= MMVQ_MAX_BATCH_SIZE) {
             if (ggml_is_quantized(src0->type)) {
-                if (ne2 <= MMVQ_MMID_MAX_BATCH_SIZE) {
+                const int mmvq_mmid_max = get_mmvq_mmid_max_batch(src0->type, cc);
+                if (ne2 <= mmvq_mmid_max) {
                     ggml_cuda_mul_mat_vec_q(ctx, src0, src1, ids, dst);
                     return;
                 }
@@ -2958,14 +2959,18 @@ static bool ggml_cuda_graph_check_compability(ggml_cgraph * cgraph) {
         }
 
         // [TAG_MUL_MAT_ID_CUDA_GRAPHS]
-        if (node->op == GGML_OP_MUL_MAT_ID && (!ggml_is_quantized(node->src[0]->type) || node->ne[2] > MMVQ_MMID_MAX_BATCH_SIZE)) {
-            // under these conditions, the mul_mat_id operation will need to synchronize the stream, so we cannot use CUDA graphs
-            // TODO: figure out a way to enable for larger batch sizes, without hurting performance
-            // ref: https://github.com/ggml-org/llama.cpp/pull/18958
-            use_cuda_graph = false;
+        if (node->op == GGML_OP_MUL_MAT_ID) {
+            const int cc = ggml_cuda_info().devices[ggml_cuda_get_device()].cc;
+            const int mmvq_mmid_max = get_mmvq_mmid_max_batch(node->src[0]->type, cc);
+            if (!ggml_is_quantized(node->src[0]->type) || node->ne[2] > mmvq_mmid_max) {
+                // under these conditions, the mul_mat_id operation will need to synchronize the stream, so we cannot use CUDA graphs
+                // TODO: figure out a way to enable for larger batch sizes, without hurting performance
+                // ref: https://github.com/ggml-org/llama.cpp/pull/18958
+                use_cuda_graph = false;
 #ifndef NDEBUG
-            GGML_LOG_DEBUG("%s: disabling CUDA graphs due to unsupported node type\n", __func__);
+                GGML_LOG_DEBUG("%s: disabling CUDA graphs due to unsupported node type\n", __func__);
 #endif
+            }
         }
 
         if (node->op == GGML_OP_ADD && node->src[1] && node->src[1]->ne[1] > 1) {
diff --git a/ggml/src/ggml-cuda/mmvq.cu b/ggml/src/ggml-cuda/mmvq.cu
index 66bd8beea..8d80d1dd9 100644
--- a/ggml/src/ggml-cuda/mmvq.cu
+++ b/ggml/src/ggml-cuda/mmvq.cu
@@ -97,6 +97,194 @@ static __host__ mmvq_parameter_table_id get_device_table_id(int cc) {
     return MMVQ_PARAMETERS_GENERIC;
 }
 
+// Per-architecture maximum batch size for which MMVQ should be used for MUL_MAT_ID.
+// Returns a value <= MMVQ_MAX_BATCH_SIZE. Default is MMVQ_MAX_BATCH_SIZE.
+// Check https://github.com/ggml-org/llama.cpp/pull/20905#issuecomment-4145835627 for details
+
+static constexpr __host__ __device__ int get_mmvq_mmid_max_batch_pascal_older(ggml_type type) {
+    switch (type) {
+        case GGML_TYPE_IQ1_S:   return 6;
+        case GGML_TYPE_IQ1_M:   return 6;
+        case GGML_TYPE_IQ2_S:   return 4;
+        case GGML_TYPE_IQ2_XS:  return 5;
+        case GGML_TYPE_IQ2_XXS: return 5;
+        case GGML_TYPE_IQ3_S:   return 4;
+        case GGML_TYPE_IQ3_XXS: return 4;
+        case GGML_TYPE_IQ4_NL:  return 6;
+        case GGML_TYPE_IQ4_XS:  return 5;
+        case GGML_TYPE_MXFP4:   return 4;
+        case GGML_TYPE_Q2_K:    return 4;
+        case GGML_TYPE_Q3_K:    return 4;
+        case GGML_TYPE_Q4_0:    return 6;
+        case GGML_TYPE_Q4_1:    return 6;
+        case GGML_TYPE_Q4_K:    return 5;
+        case GGML_TYPE_Q5_0:    return 6;
+        case GGML_TYPE_Q5_1:    return 6;
+        case GGML_TYPE_Q5_K:    return 5;
+        case GGML_TYPE_Q6_K:    return 4;
+        case GGML_TYPE_Q8_0:    return 4;
+        default:                return MMVQ_MAX_BATCH_SIZE;
+    }
+}
+
+static constexpr __host__ __device__ int get_mmvq_mmid_max_batch_turing_plus(ggml_type type) {
+    switch (type) {
+        case GGML_TYPE_IQ2_S:   return 7;
+        case GGML_TYPE_IQ3_S:   return 6;
+        case GGML_TYPE_IQ3_XXS: return 7;
+        case GGML_TYPE_MXFP4:   return 7;
+        case GGML_TYPE_Q2_K:    return 7;
+        case GGML_TYPE_Q3_K:    return 5;
+        default:                return MMVQ_MAX_BATCH_SIZE;
+    }
+}
+
+static constexpr __host__ __device__ int get_mmvq_mmid_max_batch_gcn(ggml_type type) {
+    switch (type) {
+        case GGML_TYPE_IQ1_S:   return 5;
+        case GGML_TYPE_IQ1_M:   return 5;
+        case GGML_TYPE_IQ2_S:   return 4;
+        case GGML_TYPE_IQ2_XS:  return 4;
+        case GGML_TYPE_IQ2_XXS: return 4;
+        case GGML_TYPE_IQ3_S:   return 4;
+        case GGML_TYPE_IQ3_XXS: return 4;
+        case GGML_TYPE_IQ4_NL:  return 6;
+        case GGML_TYPE_IQ4_XS:  return 4;
+        case GGML_TYPE_Q2_K:    return 4;
+        case GGML_TYPE_Q3_K:    return 4;
+        case GGML_TYPE_Q4_0:    return 5;
+        case GGML_TYPE_Q4_1:    return 5;
+        case GGML_TYPE_Q4_K:    return 4;
+        case GGML_TYPE_Q5_K:    return 4;
+        case GGML_TYPE_Q6_K:    return 4;
+        case GGML_TYPE_Q8_0:    return 4;
+        default:                return MMVQ_MAX_BATCH_SIZE;
+    }
+}
+
+static constexpr __host__ __device__ int get_mmvq_mmid_max_batch_cdna(ggml_type type) {
+    switch (type) {
+        case GGML_TYPE_IQ2_S:   return 5;
+        case GGML_TYPE_IQ2_XS:  return 5;
+        case GGML_TYPE_IQ2_XXS: return 5;
+        case GGML_TYPE_IQ3_S:   return 4;
+        case GGML_TYPE_IQ3_XXS: return 5;
+        default:                return MMVQ_MAX_BATCH_SIZE;
+    }
+}
+
+static constexpr __host__ __device__ int get_mmvq_mmid_max_batch_rdna1_rdna2(ggml_type type) {
+    switch (type) {
+        case GGML_TYPE_IQ2_S:   return 4;
+        case GGML_TYPE_IQ2_XS:  return 4;
+        case GGML_TYPE_IQ2_XXS: return 4;
+        case GGML_TYPE_IQ3_S:   return 4;
+        case GGML_TYPE_IQ3_XXS: return 4;
+        case GGML_TYPE_Q2_K:    return 7;
+        case GGML_TYPE_Q3_K:    return 4;
+        case GGML_TYPE_Q4_K:    return 5;
+        case GGML_TYPE_Q5_K:    return 6;
+        case GGML_TYPE_Q6_K:    return 5;
+        default:                return MMVQ_MAX_BATCH_SIZE;
+    }
+}
+
+static constexpr __host__ __device__ int get_mmvq_mmid_max_batch_rdna3(ggml_type type) {
+    switch (type) {
+        case GGML_TYPE_IQ1_S:   return 6;
+        case GGML_TYPE_IQ1_M:   return 6;
+        case GGML_TYPE_IQ2_S:   return 4;
+        case GGML_TYPE_IQ2_XS:  return 4;
+        case GGML_TYPE_IQ2_XXS: return 4;
+        case GGML_TYPE_IQ3_S:   return 4;
+        case GGML_TYPE_IQ3_XXS: return 4;
+        case GGML_TYPE_IQ4_NL:  return 6;
+        case GGML_TYPE_IQ4_XS:  return 6;
+        case GGML_TYPE_Q4_K:    return 4;
+        case GGML_TYPE_Q5_K:    return 4;
+        case GGML_TYPE_Q6_K:    return 4;
+        default:                return MMVQ_MAX_BATCH_SIZE;
+    }
+}
+
+static constexpr __host__ __device__ int get_mmvq_mmid_max_batch_rdna4(ggml_type type) {
+    switch (type) {
+        case GGML_TYPE_IQ1_S:   return 7;
+        case GGML_TYPE_IQ1_M:   return 7;
+        case GGML_TYPE_IQ2_S:   return 4;
+        case GGML_TYPE_IQ2_XS:  return 4;
+        case GGML_TYPE_IQ2_XXS: return 4;
+        case GGML_TYPE_IQ3_S:   return 4;
+        case GGML_TYPE_IQ3_XXS: return 4;
+        case GGML_TYPE_IQ4_NL:  return 7;
+        case GGML_TYPE_IQ4_XS:  return 5;
+        case GGML_TYPE_MXFP4:   return 5;
+        case GGML_TYPE_Q3_K:    return 4;
+        case GGML_TYPE_Q4_0:    return 7;
+        case GGML_TYPE_Q4_1:    return 7;
+        case GGML_TYPE_Q4_K:    return 4;
+        case GGML_TYPE_Q5_0:    return 7;
+        case GGML_TYPE_Q5_1:    return 7;
+        case GGML_TYPE_Q5_K:    return 5;
+        case GGML_TYPE_Q6_K:    return 5;
+        case GGML_TYPE_Q8_0:    return 7;
+        default:                return MMVQ_MAX_BATCH_SIZE;
+    }
+}
+
+// Host function: returns the max batch size for the current arch+type at runtime.
+int get_mmvq_mmid_max_batch(ggml_type type, int cc) {
+    // NVIDIA: Volta, Ada Lovelace, and Blackwell always use MMVQ for MUL_MAT_ID.
+    if (cc == GGML_CUDA_CC_VOLTA || cc >= GGML_CUDA_CC_ADA_LOVELACE) {
+        return MMVQ_MAX_BATCH_SIZE;
+    }
+    if (cc >= GGML_CUDA_CC_TURING) {
+        return get_mmvq_mmid_max_batch_turing_plus(type);
+    }
+    if (GGML_CUDA_CC_IS_NVIDIA(cc)) {
+        return get_mmvq_mmid_max_batch_pascal_older(type);
+    }
+    // AMD
+    if (GGML_CUDA_CC_IS_RDNA4(cc)) {
+        return get_mmvq_mmid_max_batch_rdna4(type);
+    }
+    if (GGML_CUDA_CC_IS_RDNA3(cc)) {
+        return get_mmvq_mmid_max_batch_rdna3(type);
+    }
+    if (GGML_CUDA_CC_IS_RDNA1(cc) || GGML_CUDA_CC_IS_RDNA2(cc)) {
+        return get_mmvq_mmid_max_batch_rdna1_rdna2(type);
+    }
+    if (GGML_CUDA_CC_IS_CDNA(cc)) {
+        return get_mmvq_mmid_max_batch_cdna(type);
+    }
+    if (GGML_CUDA_CC_IS_GCN(cc)) {
+        return get_mmvq_mmid_max_batch_gcn(type);
+    }
+    return MMVQ_MAX_BATCH_SIZE;
+}
+
+// Device constexpr: returns the max batch size for the current arch+type at compile time.
+template <ggml_type type>
+static constexpr __device__ int get_mmvq_mmid_max_batch_for_device() {
+#if defined(RDNA4)
+    return get_mmvq_mmid_max_batch_rdna4(type);
+#elif defined(RDNA3)
+    return get_mmvq_mmid_max_batch_rdna3(type);
+#elif defined(RDNA2) || defined(RDNA1)
+    return get_mmvq_mmid_max_batch_rdna1_rdna2(type);
+#elif defined(CDNA)
+    return get_mmvq_mmid_max_batch_cdna(type);
+#elif defined(GCN)
+    return get_mmvq_mmid_max_batch_gcn(type);
+#elif defined(__CUDA_ARCH__) && (__CUDA_ARCH__ == GGML_CUDA_CC_VOLTA || __CUDA_ARCH__ >= GGML_CUDA_CC_ADA_LOVELACE)
+    return MMVQ_MAX_BATCH_SIZE;
+#elif defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= GGML_CUDA_CC_TURING
+    return get_mmvq_mmid_max_batch_turing_plus(type);
+#else
+    return get_mmvq_mmid_max_batch_pascal_older(type);
+#endif
+}
+
 static constexpr __host__ __device__ int calc_nwarps(ggml_type type, int ncols_dst, mmvq_parameter_table_id table_id) {
     if (table_id == MMVQ_PARAMETERS_GENERIC) {
         switch (ncols_dst) {
@@ -195,7 +383,7 @@ static constexpr __host__ __device__ int calc_rows_per_block(int ncols_dst, int
     return 1;
 }
 
-template <ggml_type type, int ncols_dst, bool has_fusion, bool is_multi_token_id = false, bool small_k = false>
+template <ggml_type type, int ncols_dst, bool has_fusion, bool small_k = false>
 __launch_bounds__(calc_nwarps(type, ncols_dst, get_device_table_id())*ggml_cuda_get_physical_warp_size(), 1)
 static __global__ void mul_mat_vec_q(
         const void * __restrict__ vx, const void * __restrict__ vy, const int32_t * __restrict__ ids, const ggml_cuda_mm_fusion_args_device fusion, float * __restrict__ dst,
@@ -222,22 +410,13 @@ static __global__ void mul_mat_vec_q(
 
     const uint32_t channel_dst = blockIdx.y;
 
-    uint32_t token_idx = 0;
     uint32_t channel_x;
     uint32_t channel_y;
     uint32_t sample_dst;
 
-    if constexpr (is_multi_token_id) {
-        // Multi-token MUL_MAT_ID path, adding these in the normal path causes a perf regression for n_tokens=1 case
-        token_idx  = blockIdx.z;
-        channel_x  = ids[channel_dst + token_idx * ids_stride];
-        channel_y  = fastmodulo(channel_dst, nchannels_y);
-        sample_dst = 0;
-    } else {
-        channel_x  = ncols_dst == 1 && ids ? ids[channel_dst]                     : fastdiv(channel_dst, channel_ratio);
-        channel_y  = ncols_dst == 1 && ids ? fastmodulo(channel_dst, nchannels_y) : channel_dst;
-        sample_dst = blockIdx.z;
-    }
+    channel_x  = ncols_dst == 1 && ids ? ids[channel_dst]                     : fastdiv(channel_dst, channel_ratio);
+    channel_y  = ncols_dst == 1 && ids ? fastmodulo(channel_dst, nchannels_y) : channel_dst;
+    sample_dst = blockIdx.z;
 
     const uint32_t sample_x    = fastdiv(sample_dst, sample_ratio);
     const uint32_t sample_y    = sample_dst;
@@ -294,9 +473,6 @@ static __global__ void mul_mat_vec_q(
     float tmp_gate[ncols_dst][rows_per_cuda_block] = {{0.0f}};
 
     const block_q8_1 * y = ((const block_q8_1 *) vy) + sample_y*stride_sample_y + channel_y*stride_channel_y;
-    if constexpr (is_multi_token_id) {
-        y += token_idx*stride_col_y;
-    }
     const int kbx_offset = sample_x*stride_sample_x + channel_x*stride_channel_x + row0*stride_row_x;
 
     for (int kbx = tid / (qi/vdr); kbx < blocks_per_row_x; kbx += blocks_per_iter) {
@@ -350,10 +526,6 @@ static __global__ void mul_mat_vec_q(
 
     dst += sample_dst*stride_sample_dst + channel_dst*stride_channel_dst + row0;
 
-    if constexpr (is_multi_token_id) {
-        dst += token_idx*stride_col_dst;
-    }
-
     // sum up partial sums and write back result
 #pragma unroll
     for (int j = 0; j < ncols_dst; ++j) {
@@ -413,6 +585,69 @@ static __global__ void mul_mat_vec_q(
     }
 }
 
+// Dedicated MoE multi-token kernel.
+// Grid: (ceil(nrows_x / c_rows_per_block), nchannels_dst)
+// Block: (warp_size, ncols_dst) - each warp handles one token independently.
+// No shared memory reduction needed since each warp works alone.
+template <ggml_type type, int c_rows_per_block>
+__launch_bounds__(get_mmvq_mmid_max_batch_for_device<type>()*ggml_cuda_get_physical_warp_size(), 1)
+static __global__ void mul_mat_vec_q_moe(
+        const void * __restrict__ vx, const void * __restrict__ vy, const int32_t * __restrict__ ids,
+        float * __restrict__ dst,
+        const uint32_t ncols_x, const uint3 nchannels_y, const uint32_t nrows_x,
+        const uint32_t stride_row_x, const uint32_t stride_col_y, const uint32_t stride_col_dst,
+        const uint32_t stride_channel_x, const uint32_t stride_channel_y, const uint32_t stride_channel_dst,
+        const uint32_t ncols_dst, const uint32_t ids_stride) {
+
+    constexpr int qk  = ggml_cuda_type_traits<type>::qk;
+    constexpr int qi  = ggml_cuda_type_traits<type>::qi;
+    constexpr int vdr = get_vdr_mmvq(type);
+    constexpr int warp_size = ggml_cuda_get_physical_warp_size();
+
+    constexpr vec_dot_q_cuda_t vec_dot_q_cuda = get_vec_dot_q_cuda(type);
+
+    const uint32_t token_idx   = threadIdx.y;
+    const int      row0        = c_rows_per_block*blockIdx.x;
+    const int      blocks_per_row_x = ncols_x / qk;
+    constexpr int  blocks_per_iter  = vdr * warp_size / qi;
+
+    const uint32_t channel_dst = blockIdx.y;
+
+    if (token_idx >= ncols_dst) {
+        return;
+    }
+
+    const uint32_t channel_x = ids[channel_dst + token_idx * ids_stride];
+    const uint32_t channel_y = fastmodulo(channel_dst, nchannels_y);
+
+    const block_q8_1 * y = ((const block_q8_1 *) vy) + channel_y*stride_channel_y + token_idx*stride_col_y;
+    const int kbx_offset  = channel_x*stride_channel_x + row0*stride_row_x;
+
+    // partial sum for each thread
+    float tmp[c_rows_per_block] = {0.0f};
+
+    for (int kbx = threadIdx.x / (qi/vdr); kbx < blocks_per_row_x; kbx += blocks_per_iter) {
+        const int kby = kbx * (qk/QK8_1);
+        const int kqs = vdr * (threadIdx.x % (qi/vdr));
+
+#pragma unroll
+        for (int i = 0; i < c_rows_per_block; ++i) {
+            tmp[i] += vec_dot_q_cuda(vx, &y[kby], kbx_offset + i*stride_row_x + kbx, kqs);
+        }
+    }
+
+    // Warp-level reduction only - no shared memory needed
+#pragma unroll
+    for (int i = 0; i < c_rows_per_block; ++i) {
+        tmp[i] = warp_reduce_sum<warp_size>(tmp[i]);
+    }
+
+    // Write results
+    if (threadIdx.x < c_rows_per_block && (c_rows_per_block == 1 || uint32_t(row0 + threadIdx.x) < nrows_x)) {
+        dst[channel_dst*stride_channel_dst + token_idx*stride_col_dst + row0 + threadIdx.x] = tmp[threadIdx.x];
+    }
+}
+
 template<ggml_type type>
 static std::pair<dim3, dim3> calc_launch_params(
         const int ncols_dst, const int nrows_x, const int nchannels_dst, const int nsamples_or_ntokens,
@@ -425,7 +660,7 @@ static std::pair<dim3, dim3> calc_launch_params(
     return {block_nums, block_dims};
 }
 
-template<ggml_type type, int c_ncols_dst, bool is_multi_token_id = false, bool small_k = false>
+template<ggml_type type, int c_ncols_dst, bool small_k = false>
 static void mul_mat_vec_q_switch_fusion(
         const void * vx, const void * vy, const int32_t * ids, const ggml_cuda_mm_fusion_args_device fusion, float * dst,
         const uint32_t ncols_x, const uint3 nchannels_y, const uint32_t stride_row_x, const uint32_t stride_col_y,
@@ -438,7 +673,7 @@ static void mul_mat_vec_q_switch_fusion(
     const bool has_fusion = fusion.gate != nullptr || fusion.x_bias != nullptr || fusion.gate_bias != nullptr;
     if constexpr (c_ncols_dst == 1) {
         if (has_fusion) {
-            mul_mat_vec_q<type, c_ncols_dst, true, is_multi_token_id, small_k><<<block_nums, block_dims, nbytes_shared, stream>>>
+            mul_mat_vec_q<type, c_ncols_dst, true, small_k><<<block_nums, block_dims, nbytes_shared, stream>>>
                 (vx, vy, ids, fusion, dst, ncols_x, nchannels_y, stride_row_x, stride_col_y, stride_col_dst,
                  channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
                  sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst, ids_stride);
@@ -448,12 +683,33 @@ static void mul_mat_vec_q_switch_fusion(
 
     GGML_ASSERT(!has_fusion && "fusion only supported for ncols_dst=1");
 
-    mul_mat_vec_q<type, c_ncols_dst, false, is_multi_token_id, small_k><<<block_nums, block_dims, nbytes_shared, stream>>>
+    mul_mat_vec_q<type, c_ncols_dst, false, small_k><<<block_nums, block_dims, nbytes_shared, stream>>>
         (vx, vy, ids, fusion, dst, ncols_x, nchannels_y, stride_row_x, stride_col_y, stride_col_dst,
         channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
         sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst, ids_stride);
 }
 
+template <ggml_type type>
+static void mul_mat_vec_q_moe_launch(
+        const void * vx, const void * vy, const int32_t * ids, float * dst,
+        const uint32_t ncols_x, const uint3 nchannels_y, const uint32_t nrows_x,
+        const uint32_t stride_row_x, const uint32_t stride_col_y, const uint32_t stride_col_dst,
+        const uint32_t stride_channel_x, const uint32_t stride_channel_y, const uint32_t stride_channel_dst,
+        const uint32_t ncols_dst, const uint32_t ids_stride,
+        const int warp_size, const int nchannels_dst, cudaStream_t stream) {
+
+    constexpr int rows_per_block = 2; // 2 gives best perf based on tuning
+    const int64_t nblocks_rows = (nrows_x + rows_per_block - 1) / rows_per_block;
+    const dim3 block_nums(nblocks_rows, nchannels_dst);
+    const dim3 block_dims(warp_size, ncols_dst);
+
+    mul_mat_vec_q_moe<type, rows_per_block><<<block_nums, block_dims, 0, stream>>>(
+        vx, vy, ids, dst, ncols_x, nchannels_y, nrows_x,
+        stride_row_x, stride_col_y, stride_col_dst,
+        stride_channel_x, stride_channel_y, stride_channel_dst,
+        ncols_dst, ids_stride);
+}
+
 template <ggml_type type>
 static void mul_mat_vec_q_switch_ncols_dst(
         const void * vx, const void * vy, const int32_t * ids, const ggml_cuda_mm_fusion_args_device fusion, float * dst,
@@ -472,20 +728,62 @@ static void mul_mat_vec_q_switch_ncols_dst(
     const uint3 sample_ratio_fd  = init_fastdiv_values(nsamples_dst  / nsamples_x);
 
     const int device = ggml_cuda_get_device();
+    const int                     cc        = ggml_cuda_info().devices[device].cc;
     const int warp_size = ggml_cuda_info().devices[device].warp_size;
-    const mmvq_parameter_table_id table_id = get_device_table_id(ggml_cuda_info().devices[device].cc);
+    const mmvq_parameter_table_id table_id  = get_device_table_id(cc);
 
     const bool has_fusion = fusion.gate != nullptr || fusion.x_bias != nullptr || fusion.gate_bias != nullptr;
     const bool has_ids = ids != nullptr;
 
+    const auto should_use_small_k = [&](int c_ncols_dst) {
+        // When K is small, increase rows_per_block to match nwarps so each warp has more work to do
+        // Trigger when the full thread block covers all K blocks in a single loop iteration and few threads remain idle.
+        constexpr int qk                    = ggml_cuda_type_traits<type>::qk;
+        constexpr int qi                    = ggml_cuda_type_traits<type>::qi;
+        constexpr int vdr                   = get_vdr_mmvq(type);
+        const int     blocks_per_row_x      = ncols_x / qk;
+        const int     blocks_per_iter_1warp = vdr * warp_size / qi;
+        const int     nwarps                = calc_nwarps(type, c_ncols_dst, table_id);
+        bool          use                   = nwarps > 1 && blocks_per_row_x < nwarps * blocks_per_iter_1warp;
+
+        constexpr std::array<ggml_type, 2> iq_slow_turing = {
+            GGML_TYPE_IQ3_XXS,
+            GGML_TYPE_IQ3_S,
+        };
+        constexpr std::array<ggml_type, 8> iq_slow_other = {
+            GGML_TYPE_IQ1_S, GGML_TYPE_IQ1_M,   GGML_TYPE_IQ2_XXS, GGML_TYPE_IQ2_XS,
+            GGML_TYPE_IQ2_S, GGML_TYPE_IQ3_XXS, GGML_TYPE_IQ3_S,   GGML_TYPE_IQ4_XS,
+        };
+        constexpr std::array<ggml_type, 3> slow_pascal = {
+            GGML_TYPE_IQ3_S,
+            GGML_TYPE_Q2_K,
+            GGML_TYPE_Q3_K,
+        };
+
+        const bool is_nvidia_turing_plus  = GGML_CUDA_CC_IS_NVIDIA(cc) && cc >= GGML_CUDA_CC_TURING;
+        const bool is_nvidia_pascal_older = GGML_CUDA_CC_IS_NVIDIA(cc) && cc < GGML_CUDA_CC_VOLTA;
+
+        if (is_nvidia_turing_plus) {
+            if (ncols_dst == 1 &&
+                    std::find(iq_slow_turing.begin(), iq_slow_turing.end(), type) != iq_slow_turing.end()) {
+                use = false;
+            }
+        } else if ((ncols_dst == 1 && std::find(iq_slow_other.begin(), iq_slow_other.end(), type) != iq_slow_other.end()) ||
+                (is_nvidia_pascal_older && std::find(slow_pascal.begin(), slow_pascal.end(), type) != slow_pascal.end()) ||
+                GGML_CUDA_CC_IS_RDNA(cc)) {
+            use = false;
+        }
+
+        return use;
+    };
+
     if (has_ids && ncols_dst > 1) {
-        // Multi-token MUL_MAT_ID path only - single-token goes through regular path below
-        constexpr int c_ncols_dst = 1;
-        std::pair<dim3, dim3> dims = calc_launch_params<type>(c_ncols_dst, nrows_x, nchannels_dst, ncols_dst, warp_size, table_id);
-        mul_mat_vec_q_switch_fusion<type, c_ncols_dst, true>(vx, vy, ids, fusion, dst, ncols_x, nchannels_y_fd, stride_row_x, stride_col_y, stride_col_dst,
-             channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
-             sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst,
-             dims.first, dims.second, 0, ids_stride, stream);
+        // Multi-token MUL_MAT_ID path - dedicated MoE kernel
+        mul_mat_vec_q_moe_launch<type>(
+            vx, vy, ids, dst, ncols_x, nchannels_y_fd, nrows_x,
+            stride_row_x, stride_col_y, stride_col_dst,
+            stride_channel_x, stride_channel_y, stride_channel_dst,
+            ncols_dst, ids_stride, warp_size, nchannels_dst, stream);
         return;
     }
 
@@ -493,31 +791,24 @@ static void mul_mat_vec_q_switch_ncols_dst(
         case 1: {
             constexpr int c_ncols_dst = 1;
 
-            // When K is small, increase rows_per_block to match nwarps so each warp has more work to do
-            // Trigger when the full thread block covers all K blocks in a single loop iteration and few threads remain idle.
-            constexpr int qk  = ggml_cuda_type_traits<type>::qk;
-            constexpr int qi  = ggml_cuda_type_traits<type>::qi;
-            constexpr int vdr = get_vdr_mmvq(type);
-            const int blocks_per_row_x = ncols_x / qk;
-            const int blocks_per_iter_1warp = vdr * warp_size / qi;
-            const int nwarps = calc_nwarps(type, c_ncols_dst, table_id);
-            const bool use_small_k = nwarps > 1 && blocks_per_row_x < nwarps * blocks_per_iter_1warp;
+            bool use_small_k = should_use_small_k(c_ncols_dst);
+
             if (use_small_k) {
-                std::pair<dim3, dim3> dims = calc_launch_params<type>(c_ncols_dst, nrows_x, nchannels_dst, nsamples_dst,
-                                                                    warp_size, table_id, true);
-                mul_mat_vec_q_switch_fusion<type, c_ncols_dst, false, true>(
+                std::pair<dim3, dim3> dims = calc_launch_params<type>(c_ncols_dst, nrows_x, nchannels_dst,
+                                                                        nsamples_dst, warp_size, table_id, true);
+                mul_mat_vec_q_switch_fusion<type, c_ncols_dst, true>(
                     vx, vy, ids, fusion, dst, ncols_x, nchannels_y_fd, stride_row_x, stride_col_y, stride_col_dst,
-                    channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
-                    sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst,
-                    dims.first, dims.second, 0, ids_stride, stream);
+                    channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst, sample_ratio_fd,
+                    stride_sample_x, stride_sample_y, stride_sample_dst, dims.first, dims.second, 0, ids_stride,
+                    stream);
             } else {
-                std::pair<dim3, dim3> dims = calc_launch_params<type>(c_ncols_dst, nrows_x, nchannels_dst, nsamples_dst,
-                                                                    warp_size, table_id);
+                std::pair<dim3, dim3> dims = calc_launch_params<type>(c_ncols_dst, nrows_x, nchannels_dst,
+                                                                        nsamples_dst, warp_size, table_id);
                 mul_mat_vec_q_switch_fusion<type, c_ncols_dst>(
                     vx, vy, ids, fusion, dst, ncols_x, nchannels_y_fd, stride_row_x, stride_col_y, stride_col_dst,
-                    channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
-                    sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst,
-                    dims.first, dims.second, 0, ids_stride, stream);
+                    channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst, sample_ratio_fd,
+                    stride_sample_x, stride_sample_y, stride_sample_dst, dims.first, dims.second, 0, ids_stride,
+                    stream);
             }
         } break;
         case 2: {
diff --git a/ggml/src/ggml-cuda/mmvq.cuh b/ggml/src/ggml-cuda/mmvq.cuh
index 8a154631f..6bf0a8e86 100644
--- a/ggml/src/ggml-cuda/mmvq.cuh
+++ b/ggml/src/ggml-cuda/mmvq.cuh
@@ -1,7 +1,10 @@
 #include "common.cuh"
 
 #define MMVQ_MAX_BATCH_SIZE 8 // Max. batch size for which to use MMVQ kernels.
-#define MMVQ_MMID_MAX_BATCH_SIZE 4 // Max. batch size for which to use MMVQ kernels for MUL_MAT_ID
+
+// Returns the maximum batch size for which MMVQ should be used for MUL_MAT_ID,
+// based on the quantization type and GPU architecture (compute capability).
+int get_mmvq_mmid_max_batch(ggml_type type, int cc);
 
 void ggml_cuda_mul_mat_vec_q(ggml_backend_cuda_context & ctx,
     const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * ids, ggml_tensor * dst, const ggml_cuda_mm_fusion_args_host * fusion = nullptr);
diff --git a/gguf-py/gguf/vocab.py b/gguf-py/gguf/vocab.py
index e4ab5e1e4..5cd729dfa 100644
--- a/gguf-py/gguf/vocab.py
+++ b/gguf-py/gguf/vocab.py
@@ -14,12 +14,12 @@ except ImportError:
     SentencePieceProcessor: Any = None
 
 try:
-    from mistral_common.tokens.tokenizers.mistral import MistralTokenizer # type: ignore[import-not-found]
-    from mistral_common.tokens.tokenizers.tekken import Tekkenizer # type: ignore[import-not-found]
-    from mistral_common.tokens.tokenizers.utils import ( # type: ignore[import-not-found]
+    from mistral_common.tokens.tokenizers.mistral import MistralTokenizer # type: ignore[import-not-found, ty:unresolved-import]
+    from mistral_common.tokens.tokenizers.tekken import Tekkenizer # type: ignore[import-not-found, ty:unresolved-import]
+    from mistral_common.tokens.tokenizers.utils import ( # type: ignore[import-not-found, ty:unresolved-import]
         _filter_valid_tokenizer_files,
     )
-    from mistral_common.tokens.tokenizers.sentencepiece import ( # type: ignore[import-not-found]
+    from mistral_common.tokens.tokenizers.sentencepiece import ( # type: ignore[import-not-found, ty:unresolved-import]
         SentencePieceTokenizer,
     )
 except ImportError:
@@ -32,7 +32,7 @@ else:
     _mistral_common_installed = True
 
 try:
-    from mistral_common.tokens.tokenizers.utils import ( # type: ignore[import-not-found]
+    from mistral_common.tokens.tokenizers.utils import ( # type: ignore[import-not-found, ty:unresolved-import]
         get_one_valid_tokenizer_file,
     )
 except ImportError:
diff --git a/src/llama-arch.cpp b/src/llama-arch.cpp
index a90518763..bd6d28e5f 100644
--- a/src/llama-arch.cpp
+++ b/src/llama-arch.cpp
@@ -557,6 +557,8 @@ static std::set<llm_tensor> llm_get_tensor_names(llm_arch arch) {
                 LLM_TENSOR_OUTPUT_NORM,
                 LLM_TENSOR_OUTPUT,
                 LLM_TENSOR_ROPE_FREQS,
+                LLM_TENSOR_ROPE_FACTORS_LONG,
+                LLM_TENSOR_ROPE_FACTORS_SHORT,
                 LLM_TENSOR_ATTN_NORM,
                 LLM_TENSOR_ATTN_Q,
                 LLM_TENSOR_ATTN_K,
diff --git a/src/llama-model-loader.cpp b/src/llama-model-loader.cpp
index 13212b10d..a8c2675a3 100644
--- a/src/llama-model-loader.cpp
+++ b/src/llama-model-loader.cpp
@@ -1162,6 +1162,12 @@ struct ggml_tensor * llama_model_loader::create_tensor(
                     if (overrides->buft == ggml_backend_cpu_buffer_type()) {
                         // when overriding to a CPU buffer, consider the extra buffer types
                         buft = select_weight_buft(hparams, t_meta, op, buft_list_cpu);
+                        if (use_mmap) {
+                            static std::once_flag once;
+                            std::call_once(once, [] {
+                                LLAMA_LOG_WARN("llama_model_loader: tensor overrides to CPU are used with mmap enabled - consider using --no-mmap for better performance\n");
+                            });
+                        }
                     } else {
                         buft = overrides->buft;
                     }
diff --git a/tools/server/public/index.html.gz b/tools/server/public/index.html.gz
index 5e7326a58..1a69dde32 100644
Binary files a/tools/server/public/index.html.gz and b/tools/server/public/index.html.gz differ
diff --git a/tools/server/server-cors-proxy.h b/tools/server/server-cors-proxy.h
index c412d4c25..ffdd073b9 100644
--- a/tools/server/server-cors-proxy.h
+++ b/tools/server/server-cors-proxy.h
@@ -32,13 +32,22 @@ static server_http_res_ptr proxy_request(const server_http_req & req, std::strin
 
     SRV_INF("proxying %s request to %s://%s:%i%s\n", method.c_str(), parsed_url.scheme.c_str(), parsed_url.host.c_str(), parsed_url.port, parsed_url.path.c_str());
 
+    std::map<std::string, std::string> headers;
+    for (auto [key, value] : req.headers) {
+        auto new_key = key;
+        if (string_starts_with(new_key, "X-Proxy-Header-")) {
+            string_replace_all(new_key, "X-Proxy-Header-", "");
+        }
+        headers[new_key] = value;
+    }
+
     auto proxy = std::make_unique<server_http_proxy>(
             method,
             parsed_url.scheme,
             parsed_url.host,
             parsed_url.port,
             parsed_url.path,
-            req.headers,
+            headers,
             req.body,
             req.should_stop,
             600, // timeout_read (default to 10 minutes)
diff --git a/tools/server/server-http.h b/tools/server/server-http.h
index 3621064cd..f8a174c44 100644
--- a/tools/server/server-http.h
+++ b/tools/server/server-http.h
@@ -35,7 +35,7 @@ using server_http_res_ptr = std::unique_ptr<server_http_res>;
 
 struct server_http_req {
     std::map<std::string, std::string> params; // path_params + query_params
-    std::map<std::string, std::string> headers; // reserved for future use
+    std::map<std::string, std::string> headers; // used by MCP proxy
     std::string path;
     std::string query_string; // query parameters string (e.g. "action=save")
     std::string body;
diff --git a/tools/server/tests/utils.py b/tools/server/tests/utils.py
index 8f62210bc..a9a7e3c4f 100644
--- a/tools/server/tests/utils.py
+++ b/tools/server/tests/utils.py
@@ -116,7 +116,7 @@ class ServerProcess:
             self.server_port = int(os.environ["PORT"])
         self.external_server = "DEBUG_EXTERNAL" in os.environ
 
-    def start(self, timeout_seconds: int | None = DEFAULT_HTTP_TIMEOUT) -> None:
+    def start(self, timeout_seconds: int = DEFAULT_HTTP_TIMEOUT) -> None:
         if self.external_server:
             print(f"[external_server]: Assuming external server running on {self.server_host}:{self.server_port}")
             return
diff --git a/tools/server/webui/src/lib/components/app/models/ModelId.svelte b/tools/server/webui/src/lib/components/app/models/ModelId.svelte
index 5fda49342..272b697c4 100644
--- a/tools/server/webui/src/lib/components/app/models/ModelId.svelte
+++ b/tools/server/webui/src/lib/components/app/models/ModelId.svelte
@@ -18,7 +18,8 @@
 		showRaw = undefined,
 		aliases,
 		tags,
-		class: className = ''
+		class: className = '',
+		...rest
 	}: Props = $props();
 
 	const badgeClass =
@@ -36,9 +37,9 @@
 </script>
 
 {#if resolvedShowRaw}
-	<TruncatedText class="font-medium {className}" showTooltip={false} text={modelId} />
+	<TruncatedText class="font-medium {className}" showTooltip={false} text={modelId} {...rest} />
 {:else}
-	<span class="flex min-w-0 flex-wrap items-center gap-1 {className}">
+	<span class="flex min-w-0 flex-wrap items-center gap-1 {className}" {...rest}>
 		<span class="min-w-0 truncate font-medium">
 			{#if showOrgName && parsed.orgName && !(aliases && aliases.length > 0)}{parsed.orgName}/{/if}{displayName}
 		</span>
diff --git a/tools/server/webui/src/lib/components/app/models/ModelsSelector.svelte b/tools/server/webui/src/lib/components/app/models/ModelsSelector.svelte
index 0ed6e5c1d..ccda78690 100644
--- a/tools/server/webui/src/lib/components/app/models/ModelsSelector.svelte
+++ b/tools/server/webui/src/lib/components/app/models/ModelsSelector.svelte
@@ -271,50 +271,49 @@
 		{#if isRouter}
 			<DropdownMenu.Root bind:open={isOpen} onOpenChange={handleOpenChange}>
 				<DropdownMenu.Trigger
-					disabled={disabled || updating}
-					onclick={(e) => {
-						e.preventDefault();
-						e.stopPropagation();
-					}}
-				>
-					<button
-						type="button"
-						class={cn(
-							`inline-grid cursor-pointer grid-cols-[1fr_auto_1fr] items-center gap-1.5 rounded-sm bg-muted-foreground/10 px-1.5 py-1 text-xs transition hover:text-foreground focus:outline-none focus-visible:ring-2 focus-visible:ring-ring focus-visible:ring-offset-2 disabled:cursor-not-allowed disabled:opacity-60`,
-							!isCurrentModelInCache
-								? 'bg-red-400/10 !text-red-400 hover:bg-red-400/20 hover:text-red-400'
-								: forceForegroundText
+					class={cn(
+						`inline-grid cursor-pointer grid-cols-[1fr_auto_1fr] items-center gap-1.5 rounded-sm bg-muted-foreground/10 px-1.5 py-1 text-xs transition hover:text-foreground focus:outline-none focus-visible:ring-2 focus-visible:ring-ring focus-visible:ring-offset-2 disabled:cursor-not-allowed disabled:opacity-60`,
+						!isCurrentModelInCache
+							? 'bg-red-400/10 !text-red-400 hover:bg-red-400/20 hover:text-red-400'
+							: forceForegroundText
+								? 'text-foreground'
+								: isHighlightedCurrentModelActive
 									? 'text-foreground'
-									: isHighlightedCurrentModelActive
-										? 'text-foreground'
-										: 'text-muted-foreground',
-							isOpen ? 'text-foreground' : ''
-						)}
-						style="max-width: min(calc(100cqw - 9rem), 20rem)"
-						disabled={disabled || updating}
-					>
-						<Package class="h-3.5 w-3.5" />
+									: 'text-muted-foreground',
+						isOpen ? 'text-foreground' : ''
+					)}
+					style="max-width: min(calc(100cqw - 9rem), 20rem)"
+					disabled={disabled || updating}
+				>
+					<Package class="h-3.5 w-3.5" />
 
-						{#if selectedOption}
-							<Tooltip.Root>
-								<Tooltip.Trigger class="min-w-0 overflow-hidden">
-									<ModelId modelId={selectedOption.model} class="min-w-0" showOrgName />
-								</Tooltip.Trigger>
+					{#if selectedOption}
+						<Tooltip.Root>
+							<Tooltip.Trigger>
+								<!-- prevent another nested button element -->
+								{#snippet child({ props })}
+									<ModelId
+										modelId={selectedOption.model}
+										class="min-w-0 overflow-hidden"
+										showOrgName
+										{...props}
+									/>
+								{/snippet}
+							</Tooltip.Trigger>
 
-								<Tooltip.Content>
-									<p class="font-mono">{selectedOption.model}</p>
-								</Tooltip.Content>
-							</Tooltip.Root>
-						{:else}
-							<span class="min-w-0 font-medium">Select model</span>
-						{/if}
+							<Tooltip.Content>
+								<p class="font-mono">{selectedOption.model}</p>
+							</Tooltip.Content>
+						</Tooltip.Root>
+					{:else}
+						<span class="min-w-0 font-medium">Select model</span>
+					{/if}
 
-						{#if updating || isLoadingModel}
-							<Loader2 class="h-3 w-3.5 animate-spin" />
-						{:else}
-							<ChevronDown class="h-3 w-3.5" />
-						{/if}
-					</button>
+					{#if updating || isLoadingModel}
+						<Loader2 class="h-3 w-3.5 animate-spin" />
+					{:else}
+						<ChevronDown class="h-3 w-3.5" />
+					{/if}
 				</DropdownMenu.Trigger>
 
 				<DropdownMenu.Content
@@ -407,8 +406,16 @@
 
 				{#if selectedOption}
 					<Tooltip.Root>
-						<Tooltip.Trigger class="min-w-0 overflow-hidden">
-							<ModelId modelId={selectedOption.model} class="min-w-0" showOrgName />
+						<Tooltip.Trigger>
+							<!-- prevent another nested button element -->
+							{#snippet child({ props })}
+								<ModelId
+									modelId={selectedOption.model}
+									class="min-w-0 overflow-hidden"
+									showOrgName
+									{...props}
+								/>
+							{/snippet}
 						</Tooltip.Trigger>
 
 						<Tooltip.Content>
diff --git a/tools/server/webui/src/lib/services/mcp.service.ts b/tools/server/webui/src/lib/services/mcp.service.ts
index b38d18676..cf401bc30 100644
--- a/tools/server/webui/src/lib/services/mcp.service.ts
+++ b/tools/server/webui/src/lib/services/mcp.service.ts
@@ -39,7 +39,13 @@ import type {
 	MCPResourceContent,
 	MCPReadResourceResult
 } from '$lib/types';
-import { buildProxiedUrl, throwIfAborted, isAbortError, createBase64DataUrl } from '$lib/utils';
+import {
+	buildProxiedUrl,
+	buildProxiedHeaders,
+	throwIfAborted,
+	isAbortError,
+	createBase64DataUrl
+} from '$lib/utils';
 
 interface ToolResultContentItem {
 	type: string;
@@ -118,7 +124,7 @@ export class MCPService {
 		const requestInit: RequestInit = {};
 
 		if (config.headers) {
-			requestInit.headers = config.headers;
+			requestInit.headers = buildProxiedHeaders(config.headers);
 		}
 
 		if (config.credentials) {
diff --git a/tools/server/webui/src/lib/stores/chat.svelte.ts b/tools/server/webui/src/lib/stores/chat.svelte.ts
index 3af91be83..5f3812ed3 100644
--- a/tools/server/webui/src/lib/stores/chat.svelte.ts
+++ b/tools/server/webui/src/lib/stores/chat.svelte.ts
@@ -28,6 +28,7 @@ import {
 	filterByLeafNodeId,
 	findDescendantMessages,
 	findLeafNode,
+	findMessageById,
 	isAbortError
 } from '$lib/utils';
 import {
@@ -416,7 +417,7 @@ class ChatStore {
 		if (!activeConv) return false;
 		try {
 			const allMessages = await conversationsStore.getConversationMessages(activeConv.id);
-			const systemMessage = allMessages.find((m) => m.id === messageId);
+			const systemMessage = findMessageById(allMessages, messageId);
 			if (!systemMessage || systemMessage.role !== MessageRole.SYSTEM) return false;
 			const rootMessage = allMessages.find((m) => m.type === 'root' && m.parent === null);
 			if (!rootMessage) return false;
@@ -878,7 +879,7 @@ class ChatStore {
 			const msg = conversationsStore.activeMessages[idx];
 			if (msg.role !== MessageRole.ASSISTANT) return;
 			const allMessages = await conversationsStore.getConversationMessages(activeConv.id);
-			const parentMessage = allMessages.find((m) => m.id === msg.parent);
+			const parentMessage = findMessageById(allMessages, msg.parent);
 			if (!parentMessage) return;
 			this.setChatLoading(activeConv.id, true);
 			this.clearChatStreaming(activeConv.id);
@@ -928,7 +929,7 @@ class ChatStore {
 		if (!activeConv)
 			return { totalCount: 0, userMessages: 0, assistantMessages: 0, messageTypes: [] };
 		const allMessages = await conversationsStore.getConversationMessages(activeConv.id);
-		const messageToDelete = allMessages.find((m) => m.id === messageId);
+		const messageToDelete = findMessageById(allMessages, messageId);
 
 		// For system messages, don't count descendants as they will be preserved (reparented to root)
 		if (messageToDelete?.role === MessageRole.SYSTEM) {
@@ -975,7 +976,7 @@ class ChatStore {
 		if (!activeConv) return;
 		try {
 			const allMessages = await conversationsStore.getConversationMessages(activeConv.id);
-			const messageToDelete = allMessages.find((m) => m.id === messageId);
+			const messageToDelete = findMessageById(allMessages, messageId);
 
 			if (!messageToDelete) return;
 
@@ -1024,7 +1025,7 @@ class ChatStore {
 			this.clearChatStreaming(activeConv.id);
 
 			const allMessages = await conversationsStore.getConversationMessages(activeConv.id);
-			const dbMessage = allMessages.find((m) => m.id === messageId);
+			const dbMessage = findMessageById(allMessages, messageId);
 
 			if (!dbMessage) {
 				this.setChatLoading(activeConv.id, false);
@@ -1280,7 +1281,10 @@ class ChatStore {
 
 			let messageIdForResponse: string;
 
-			if (msg.children.length === 0) {
+			const dbMsg = findMessageById(allMessages, msg.id);
+			const hasChildren = dbMsg ? dbMsg.children.length > 0 : msg.children.length > 0;
+
+			if (!hasChildren) {
 				// No responses after this message — update in place instead of branching
 				const updates: Partial<DatabaseMessage> = {
 					content: newContent,
diff --git a/tools/server/webui/src/lib/utils/branching.ts b/tools/server/webui/src/lib/utils/branching.ts
index e60fb206f..4e117b3c2 100644
--- a/tools/server/webui/src/lib/utils/branching.ts
+++ b/tools/server/webui/src/lib/utils/branching.ts
@@ -17,6 +17,17 @@
 
 import { MessageRole } from '$lib/enums';
 
+/**
+ * Finds a message by its ID in the given messages array.
+ */
+export function findMessageById(
+	messages: readonly DatabaseMessage[],
+	id: string | null | undefined
+): DatabaseMessage | undefined {
+	if (!id) return undefined;
+	return messages.find((m) => m.id === id);
+}
+
 /**
  * Filters messages to get the conversation path from root to a specific leaf node.
  * If the leafNodeId doesn't exist, returns the path with the latest timestamp.
diff --git a/tools/server/webui/src/lib/utils/cors-proxy.ts b/tools/server/webui/src/lib/utils/cors-proxy.ts
index 39c368ba0..2eb10b16d 100644
--- a/tools/server/webui/src/lib/utils/cors-proxy.ts
+++ b/tools/server/webui/src/lib/utils/cors-proxy.ts
@@ -19,6 +19,21 @@ export function buildProxiedUrl(targetUrl: string): URL {
 	return proxyUrl;
 }
 
+/**
+ * Wrap original headers for proxying through the CORS proxy. This avoids issues with duplicated llama.cpp-specific and target headers when using the CORS proxy.
+ * @param headers - The original headers to be proxied to target
+ * @returns List of "wrapped" headers to be sent to the CORS proxy
+ */
+export function buildProxiedHeaders(headers: Record<string, string>): Record<string, string> {
+	const proxiedHeaders: Record<string, string> = {};
+
+	for (const [key, value] of Object.entries(headers)) {
+		proxiedHeaders[`X-Proxy-Header-${key}`] = value;
+	}
+
+	return proxiedHeaders;
+}
+
 /**
  * Get a proxied URL string for use in fetch requests.
  * @param targetUrl - The original URL to proxy
diff --git a/tools/server/webui/src/lib/utils/index.ts b/tools/server/webui/src/lib/utils/index.ts
index 2caaf9ac3..455d4f2c3 100644
--- a/tools/server/webui/src/lib/utils/index.ts
+++ b/tools/server/webui/src/lib/utils/index.ts
@@ -22,6 +22,7 @@ export { default as autoResizeTextarea } from './autoresize-textarea';
 // Branching utilities
 export {
 	filterByLeafNodeId,
+	findMessageById,
 	findLeafNode,
 	findDescendantMessages,
 	getMessageSiblings,
@@ -38,7 +39,7 @@ export { highlightCode, detectIncompleteCodeBlock, type IncompleteCodeBlock } fr
 export { setConfigValue, getConfigValue, configToParameterRecord } from './config-helpers';
 
 // CORS Proxy
-export { buildProxiedUrl, getProxiedUrlString } from './cors-proxy';
+export { buildProxiedUrl, getProxiedUrlString, buildProxiedHeaders } from './cors-proxy';
 
 // Conversation utilities
 export { createMessageCountMap, getMessageCount } from './conversation-utils';