vulkan: fix non-contig rope (#19299)

* add missing llama_add_compile_flags

* disable all warnings for ssl, crypto and fipsmodule

.github/workflows/build.yml

@@ -293,6 +293,7 @@ jobs:
           cmake -B build \
             -DLLAMA_FATAL_WARNINGS=ON \
             -DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON \
+            -DGGML_SANITIZE_${{ matrix.sanitizer }}=ON \
             -DCMAKE_BUILD_TYPE=${{ matrix.build_type }}
           cmake --build build --config ${{ matrix.build_type }} -j $(nproc)
 
@@ -303,6 +304,7 @@ jobs:
           cmake -B build \
             -DLLAMA_FATAL_WARNINGS=ON \
             -DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON \
+            -DGGML_SANITIZE_${{ matrix.sanitizer }}=ON \
             -DCMAKE_BUILD_TYPE=${{ matrix.build_type }} \
             -DGGML_OPENMP=OFF
           cmake --build build --config ${{ matrix.build_type }} -j $(nproc)

CODEOWNERS

@@ -27,6 +27,7 @@
 /examples/batched.swift/                @ggerganov
 /examples/batched/                      @ggerganov
 /examples/convert-llama2c-to-ggml/      @ggerganov
+/examples/debug/                        @danbev @pwilkin
 /examples/deprecation-warning/          @ggerganov
 /examples/diffusion/                    @am17an
 /examples/embedding/                    @ggerganov

common/debug.cpp

@@ -45,6 +45,8 @@ static float common_ggml_get_float_value(const uint8_t * data,
     return v;
 }
 
+#define INDENT "    "
+
 template <bool abort>
 void common_debug_print_tensor(uint8_t * data, ggml_type type, const int64_t * ne, const size_t * nb, int64_t n) {
     GGML_ASSERT(n > 0);
@@ -60,41 +62,41 @@ void common_debug_print_tensor(uint8_t * data, ggml_type type, const int64_t * n
         }
     }
     for (int64_t i3 = 0; i3 < ne[3]; i3++) {
-        LOG_ERR("                                     [\n");
+        LOG(INDENT "[\n");
         for (int64_t i2 = 0; i2 < ne[2]; i2++) {
             if (i2 == n && ne[2] > 2 * n) {
-                LOG_ERR("                                      ..., \n");
+                LOG(INDENT INDENT "..., \n");
                 i2 = ne[2] - n;
             }
-            LOG_ERR("                                      [\n");
+            LOG(INDENT INDENT "[\n");
             for (int64_t i1 = 0; i1 < ne[1]; i1++) {
                 if (i1 == n && ne[1] > 2 * n) {
-                    LOG_ERR("                                       ..., \n");
+                    LOG(INDENT INDENT INDENT "..., \n");
                     i1 = ne[1] - n;
                 }
-                LOG_ERR("                                       [");
+                LOG(INDENT INDENT INDENT "[");
                 for (int64_t i0 = 0; i0 < ne[0]; i0++) {
                     if (i0 == n && ne[0] > 2 * n) {
-                        LOG_ERR("..., ");
+                        LOG("   ..., ");
                         i0 = ne[0] - n;
                     }
                     const float v = common_ggml_get_float_value(data, type, nb, i0, i1, i2, i3);
-                    LOG_ERR("%12.4f", v);
+                    LOG("%12.4f", v);
                     if (i0 < ne[0] - 1) {
-                        LOG_ERR(", ");
+                        LOG(", ");
                     }
                 }
-                LOG_ERR("],\n");
+                LOG("  ],\n");
             }
-            LOG_ERR("                                      ],\n");
+            LOG(INDENT INDENT "],\n");
         }
-        LOG_ERR("                                     ]\n");
-        LOG_ERR("                                     sum = %f\n", sum);
+        LOG(INDENT "]\n");
+        LOG(INDENT "sum = %f\n", sum);
     }
 
     if constexpr (abort) {
         if (std::isnan(sum)) {
-            LOG_ERR("encountered NaN - aborting\n");
+            LOG("encountered NaN - aborting\n");
             exit(0);
         }
     }
@@ -137,9 +139,9 @@ template <bool abort_on_nan> bool common_debug_cb_eval(struct ggml_tensor * t, b
     }
 
     if (matches_filter) {
-        LOG_ERR("%s: %24s = (%s) %10s(%s{%s}, %s}) = {%s}\n", __func__, t->name, ggml_type_name(t->type),
-                ggml_op_desc(t), src0->name, common_ggml_ne_string(src0).c_str(), src1 ? src1_str : "",
-                common_ggml_ne_string(t).c_str());
+        LOG("%s: %24s = (%s) %10s(%s{%s}, %s}) = {%s}\n", __func__, t->name, ggml_type_name(t->type),
+            ggml_op_desc(t), src0->name, common_ggml_ne_string(src0).c_str(), src1 ? src1_str : "",
+            common_ggml_ne_string(t).c_str());
     }
 
     const bool is_host = ggml_backend_buffer_is_host(t->buffer);

common/ngram-map.cpp

@@ -47,21 +47,15 @@ static std::string common_tokens_to_str(const llama_tokens & inp, size_t start,
  * @return Vector of draft tokens, empty if no matching pattern is found
  */
 llama_tokens common_ngram_simple_draft(
-        common_ngram_simple_state & state,
+        const common_ngram_simple_config & config,
         const llama_tokens & tokens, llama_token sampled) {
 
     // Simple implementation of self-speculative decoding without a draft model.
     //
     const size_t cur_len = tokens.size();
-    // Only check every check_rate tokens to save compute
-    // i.e., perform check if (cur_len - idx_last_check) >= check_rate
-    if (state.idx_last_check + state.config.check_rate > cur_len) {
-        llama_tokens draft_tokens;
-        return draft_tokens;
-    }
 
-    size_t n_draft_min = state.config.size_ngram; // size of n-gram to lookup in token history
-    size_t n_draft_max = state.config.size_mgram; // the m-gram following the found n-gram is used for draft
+    const size_t n_draft_min = config.size_ngram; // size of n-gram to lookup in token history
+    const size_t n_draft_max = config.size_mgram; // the m-gram following the found n-gram is used for draft
 
     // vector for tokens we want to verify.
     // return empty vector if there is no match.
@@ -80,9 +74,6 @@ llama_tokens common_ngram_simple_draft(
     }
     pattern.push_back(sampled); // add the last token to the pattern
 
-    // We do a search in the token history.
-    state.idx_last_check = cur_len;
-
     size_t match_pos = 0; // we ignore position 0, position 0 == no match
                           // search backwards, but skip the current match (we are currently there)
     for (size_t j = cur_len - n_draft_min - 1; j > 0; --j) {

common/ngram-map.h

@@ -27,23 +27,9 @@ struct common_ngram_simple_config {
     uint16_t   check_rate;      // check for speculative decoding without draft model for each check_rate token
 };
 
-// current state (and config) of n-gram simple.
-struct common_ngram_simple_state {
-    common_ngram_simple_config config;
-
-    size_t idx_last_check = 0; // index of last check in context history (mutable)
-
-    common_ngram_simple_state(const common_ngram_simple_config & config)
-        : config(config) {}
-};
-
 // Searches for a n-gram in the history and checks whether a draft sequence should be generated.
-// state:              the ngram simple state to search in.
-// inp:                the tokens generated so far.
-// sampled:            the token that was just sampled.
-// draft:              vector to store the draft tokens, initially empty.
 llama_tokens common_ngram_simple_draft(
-        common_ngram_simple_state & state,
+        const common_ngram_simple_config & config,
         const llama_tokens & tokens, llama_token sampled);
 
 

common/speculative.cpp

@@ -463,12 +463,14 @@ struct common_speculative_state_eagle3 : public common_speculative_state {
 
 // state of self-speculation (simple implementation, not ngram-map)
 struct common_speculative_state_ngram_simple : public common_speculative_state {
-    common_ngram_simple_state state;
+    common_ngram_simple_config config;
+
+    uint16_t check_id = 0; // used to control the frequency of generating drafts
 
     common_speculative_state_ngram_simple(
             enum common_speculative_type type,
-            common_ngram_simple_state state)
-        : common_speculative_state(type), state(state) {}
+            common_ngram_simple_config config)
+        : common_speculative_state(type), config(config) {}
 
     void begin(const llama_tokens & prompt) override {
         GGML_UNUSED(prompt);
@@ -479,7 +481,13 @@ struct common_speculative_state_ngram_simple : public common_speculative_state {
             const llama_tokens & prompt_tgt,
             llama_token id_last,
             llama_tokens & result) override {
-        result = common_ngram_simple_draft(state, prompt_tgt, id_last);
+        ++check_id;
+        if (check_id < config.check_rate) {
+            return;
+        }
+        check_id = 0;
+
+        result = common_ngram_simple_draft(config, prompt_tgt, id_last);
         GGML_UNUSED(params);
     }
 
@@ -889,14 +897,14 @@ common_speculative * common_speculative_init(
                 uint16_t mgram_size_value = ngram_map.size_value;
                 uint16_t check_rate       = ngram_map.check_rate;
 
-                auto config_simple = common_ngram_simple_config{
+                auto config_simple = common_ngram_simple_config {
                     /* .size_ngram      = */ ngram_size_key,
                     /* .size_mgram      = */ mgram_size_value,
                     /* .check_rate      = */ check_rate
                 };
                 auto state = std::make_unique<common_speculative_state_ngram_simple>(
                     /* .type            = */ config.type,
-                    /* .state           = */ common_ngram_simple_state(config_simple)
+                    /* .state           = */ config_simple
                 );
                 impls.push_back(std::move(state));
                 break;

examples/model-conversion/scripts/utils/tensor-info.py

@@ -0,0 +1,159 @@
+#!/usr/bin/env python3
+
+import argparse
+import json
+import os
+import re
+import sys
+from pathlib import Path
+from typing import Optional
+from safetensors import safe_open
+
+
+MODEL_SAFETENSORS_FILE = "model.safetensors"
+MODEL_SAFETENSORS_INDEX = "model.safetensors.index.json"
+
+
+def get_weight_map(model_path: Path) -> Optional[dict[str, str]]:
+    index_file = model_path / MODEL_SAFETENSORS_INDEX
+
+    if index_file.exists():
+        with open(index_file, 'r') as f:
+            index = json.load(f)
+            return index.get("weight_map", {})
+
+    return None
+
+
+def get_all_tensor_names(model_path: Path) -> list[str]:
+    weight_map = get_weight_map(model_path)
+
+    if weight_map is not None:
+        return list(weight_map.keys())
+
+    single_file = model_path / MODEL_SAFETENSORS_FILE
+    if single_file.exists():
+        try:
+            with safe_open(single_file, framework="pt", device="cpu") as f:
+                return list(f.keys())
+        except Exception as e:
+            print(f"Error reading {single_file}: {e}")
+            sys.exit(1)
+
+    print(f"Error: No safetensors files found in {model_path}")
+    sys.exit(1)
+
+
+def find_tensor_file(model_path: Path, tensor_name: str) -> Optional[str]:
+    weight_map = get_weight_map(model_path)
+
+    if weight_map is not None:
+        return weight_map.get(tensor_name)
+
+    single_file = model_path / MODEL_SAFETENSORS_FILE
+    if single_file.exists():
+        return single_file.name
+
+    return None
+
+
+def normalize_tensor_name(tensor_name: str) -> str:
+    normalized = re.sub(r'\.\d+\.', '.#.', tensor_name)
+    normalized = re.sub(r'\.\d+$', '.#', normalized)
+    return normalized
+
+
+def list_all_tensors(model_path: Path, unique: bool = False):
+    tensor_names = get_all_tensor_names(model_path)
+
+    if unique:
+        seen = set()
+        for tensor_name in sorted(tensor_names):
+            normalized = normalize_tensor_name(tensor_name)
+            if normalized not in seen:
+                seen.add(normalized)
+                print(normalized)
+    else:
+        for tensor_name in sorted(tensor_names):
+            print(tensor_name)
+
+
+def print_tensor_info(model_path: Path, tensor_name: str):
+    tensor_file = find_tensor_file(model_path, tensor_name)
+
+    if tensor_file is None:
+        print(f"Error: Could not find tensor '{tensor_name}' in model index")
+        print(f"Model path: {model_path}")
+        sys.exit(1)
+
+    file_path = model_path / tensor_file
+
+    try:
+        with safe_open(file_path, framework="pt", device="cpu") as f:
+            if tensor_name in f.keys():
+                tensor_slice = f.get_slice(tensor_name)
+                shape = tensor_slice.get_shape()
+                print(f"Tensor: {tensor_name}")
+                print(f"File:   {tensor_file}")
+                print(f"Shape:  {shape}")
+            else:
+                print(f"Error: Tensor '{tensor_name}' not found in {tensor_file}")
+                sys.exit(1)
+
+    except FileNotFoundError:
+        print(f"Error: The file '{file_path}' was not found.")
+        sys.exit(1)
+    except Exception as e:
+        print(f"An error occurred: {e}")
+        sys.exit(1)
+
+
+def main():
+    parser = argparse.ArgumentParser(
+        description="Print tensor information from a safetensors model"
+    )
+    parser.add_argument(
+        "tensor_name",
+        nargs="?",  # optional (if --list is used for example)
+        help="Name of the tensor to inspect"
+    )
+    parser.add_argument(
+        "-m", "--model-path",
+        type=Path,
+        help="Path to the model directory (default: MODEL_PATH environment variable)"
+    )
+    parser.add_argument(
+        "-l", "--list",
+        action="store_true",
+        help="List unique tensor patterns in the model (layer numbers replaced with #)"
+    )
+
+    args = parser.parse_args()
+
+    model_path = args.model_path
+    if model_path is None:
+        model_path_str = os.environ.get("MODEL_PATH")
+        if model_path_str is None:
+            print("Error: --model-path not provided and MODEL_PATH environment variable not set")
+            sys.exit(1)
+        model_path = Path(model_path_str)
+
+    if not model_path.exists():
+        print(f"Error: Model path does not exist: {model_path}")
+        sys.exit(1)
+
+    if not model_path.is_dir():
+        print(f"Error: Model path is not a directory: {model_path}")
+        sys.exit(1)
+
+    if args.list:
+        list_all_tensors(model_path, unique=True)
+    else:
+        if args.tensor_name is None:
+            print("Error: tensor_name is required when not using --list")
+            sys.exit(1)
+        print_tensor_info(model_path, args.tensor_name)
+
+
+if __name__ == "__main__":
+    main()

ggml/src/ggml-metal/ggml-metal.cpp

@@ -7,6 +7,9 @@
 #include "ggml-metal-context.h"
 #include "ggml-metal-ops.h"
 
+#include <mutex>
+#include <string>
+
 #define GGML_METAL_NAME "MTL"
 #define GGML_METAL_MAX_DEVICES 16
 

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

@@ -1263,25 +1263,30 @@ struct vk_op_diag_mask_push_constants {
 
 struct vk_op_rope_push_constants {
     uint32_t rope_mode;
-    uint32_t ncols;
     uint32_t nrows;
     uint32_t n_dims;
     float freq_scale;
-    uint32_t p_delta_rows;
     float freq_base;
     float ext_factor;
     float attn_factor;
     float corr_dims[2];
     float theta_scale;
     uint32_t has_ff;
-    uint32_t ne02;
-    uint32_t s1;
-    uint32_t s2;
     int32_t sections[4];
     uint32_t is_imrope;
     uint32_t is_back;
     uint32_t set_rows_stride;
+    uint32_t ne00;
+    uint32_t ne01;
+    uint32_t ne02;
+    uint32_t nb01;
+    uint32_t nb02;
+    uint32_t nb03;
+    uint32_t nb11;
+    uint32_t nb12;
+    uint32_t nb13;
 };
+static_assert(sizeof(vk_op_rope_push_constants) <= 128, "sizeof(vk_op_rope_push_constants) must be <= 128");
 
 // For fused rms_norm+mul+rope(+view+set_rows)
 struct vk_op_rms_norm_mul_rope_push_constants {
@@ -10405,12 +10410,22 @@ static vk_op_rope_push_constants ggml_vk_make_rope_constants(const ggml_tensor *
 
     uint32_t nb01 = src0->nb[1] / ggml_type_size(src0->type);
     uint32_t nb02 = src0->nb[2] / ggml_type_size(src0->type);
+    uint32_t nb03 = src0->nb[3] / ggml_type_size(src0->type);
+
+    uint32_t nb11 = dst->nb[1] / ggml_type_size(dst->type);
+    uint32_t nb12 = dst->nb[2] / ggml_type_size(dst->type);
+    uint32_t nb13 = dst->nb[3] / ggml_type_size(dst->type);
 
     vk_op_rope_push_constants rope {
-        (uint32_t)mode, (uint32_t)src0->ne[0], (uint32_t)ggml_nrows(src0), (uint32_t)n_dims, freq_scale, (uint32_t)src0->ne[1],
-        freq_base, ext_factor, attn_factor, {corr_dims[0], corr_dims[1]}, theta_scale,
-        has_ff, (uint32_t)src0->ne[2], nb01, nb02,
+        (uint32_t)mode, (uint32_t)ggml_nrows(src0), (uint32_t)n_dims, freq_scale,
+        freq_base, ext_factor, attn_factor, {corr_dims[0], corr_dims[1]}, theta_scale, has_ff,
         { sections[0], sections[1], sections[2], sections[3] }, is_imrope, backprop, set_rows_stride,
+
+        (uint32_t)src0->ne[0],
+        (uint32_t)src0->ne[1],
+        (uint32_t)src0->ne[2],
+        nb01, nb02, nb03,
+        nb11, nb12, nb13,
     };
 
     return rope;
@@ -14798,6 +14813,7 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
         case GGML_OP_REPEAT_BACK:
             return op->type == GGML_TYPE_F32 && op->src[0]->type == GGML_TYPE_F32;
         case GGML_OP_ROPE:
+            return ggml_is_contiguous_rows(op) && ggml_is_contiguous_rows(op->src[0]);
         case GGML_OP_ROPE_BACK:
         case GGML_OP_NONE:
         case GGML_OP_RESHAPE:

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

@@ -112,12 +112,11 @@ void rms_norm(uint num_iters) {
 #if RMS_NORM_ROPE_FUSION
     barrier();
     rope_params rp = p.rope;
-    uint rope_row = (samp*nchannels + channel)*nrows + row;
     for (uint t = 2*tid; t < ncols; t += 2*BLOCK_SIZE) {
         if (rp.rope_mode == GGML_ROPE_TYPE_NEOX) {
-            rope_neox(t, rope_row, rp);
+            rope_neox(t, row, channel, samp, rp);
         } else if (rp.rope_mode == GGML_ROPE_TYPE_NORMAL) {
-            rope_norm(t, rope_row, rp);
+            rope_norm(t, row, channel, samp, rp);
         }
     }
 #endif

ggml/src/ggml-vulkan/vulkan-shaders/rope_funcs.glsl

@@ -4,12 +4,12 @@ float rope_yarn_ramp(const float low, const float high, const uint i0) {
     return 1.0f - min(1.0f, max(0.0f, y));
 }
 
-uint rope_a_coord(const uint i0, const uint i01, const uint i02, rope_params p) {
+uint rope_a_coord(const uint i0, const uint i01, const uint i02, const uint i03, rope_params p) {
 #if RMS_NORM_ROPE_FUSION
     // Per-row offset in shared memory
     const uint ix = i0;
 #else
-    const uint ix = i02*p.nb02 + i01*p.nb01 + i0;
+    const uint ix = i03*p.nb03 + i02*p.nb02 + i01*p.nb01 + i0;
 #endif
     return ix;
 }
@@ -34,26 +34,19 @@ void rope_yarn(const float theta_extrap, const uint i0, out float cos_theta, out
     sin_theta = sin(theta) * mscale;
 }
 
-void rope_norm(const uint i0, const uint i1, rope_params p) {
-    uint ne0 = p.ncols;
-    uint ne1 = p.p_delta_rows;
-
-    if (i0 >= ne0) {
+void rope_norm(const uint i0, const uint i1, const uint i2, const uint i3, rope_params p) {
+    if (i0 >= p.ne00) {
         return;
     }
 
-    // i1 is actually i2*nb2+i1, but the rows are contiguous
-    const uint i01 = i1 % ne1;
-    const uint i02 = i1 / ne1;
-
-    uint idst = i1*ne0 + i0;
-    const uint ix = rope_a_coord(i0, i01, i02, p);
+    uint idst = i0 + i1 * p.nb11 + i2 * p.nb12 + i3 * p.nb13;
+    const uint ix = rope_a_coord(i0, i1, i2, i3, p);
 
     // Fusion optimization: ROPE + VIEW + SET_ROWS.
     // The rope output is viewed as a 1D tensor and offset based on a row index in rope_data_i.
     if (p.set_rows_stride != 0) {
-        idst = i01*ne0 + i0;
-        idst += rope_data_i[i02].x * p.set_rows_stride;
+        idst = i1*p.nb11 + i0;
+        idst += rope_data_i[i2].x * p.set_rows_stride;
     }
 
     if (i0 >= p.n_dims) {
@@ -63,7 +56,7 @@ void rope_norm(const uint i0, const uint i1, rope_params p) {
         return;
     }
 
-    const float theta_base = rope_data_pos[i02] * pow(p.theta_scale, i0/2.0f);
+    const float theta_base = rope_data_pos[i2] * pow(p.theta_scale, i0/2.0f);
 
     const float freq_factor = p.has_ff != 0 ? rope_data_ff[i0/2] : 1.0f;
 
@@ -77,25 +70,19 @@ void rope_norm(const uint i0, const uint i1, rope_params p) {
     rope_data_d[idst + 1] = ROPE_D_TYPE(x0*sin_theta + x1*cos_theta);
 }
 
-void rope_neox(const uint i0, const uint i1, rope_params p) {
-    uint ne0 = p.ncols;
-    uint ne1 = p.p_delta_rows;
-
-    if (i0 >= ne0) {
+void rope_neox(const uint i0, const uint i1, const uint i2, const uint i3, rope_params p) {
+    if (i0 >= p.ne00) {
         return;
     }
 
-    const uint i01 = i1 % ne1;
-    const uint i02 = i1 / ne1;
-
-    uint idst = i1*ne0 + i0/2;
-    const uint ix = rope_a_coord(i0/2, i01, i02, p);
+    uint idst = i0/2 + i1 * p.nb11 + i2 * p.nb12 + i3 * p.nb13;
+    const uint ix = rope_a_coord(i0/2, i1, i2, i3, p);
 
     // Fusion optimization: ROPE + VIEW + SET_ROWS.
     // The rope output is viewed as a 1D tensor and offset based on a row index in rope_data_i.
     if (p.set_rows_stride != 0) {
-        idst = i01*ne0 + i0/2;
-        idst += rope_data_i[i02].x * p.set_rows_stride;
+        idst = i1*p.nb11 + i0/2;
+        idst += rope_data_i[i2].x * p.set_rows_stride;
     }
 
     if (i0 >= p.n_dims) {
@@ -105,7 +92,7 @@ void rope_neox(const uint i0, const uint i1, rope_params p) {
         return;
     }
 
-    const float theta_base = rope_data_pos[i02] * pow(p.theta_scale, i0/2.0f);
+    const float theta_base = rope_data_pos[i2] * pow(p.theta_scale, i0/2.0f);
 
     const float freq_factor = p.has_ff != 0 ? rope_data_ff[i0/2] : 1.0f;
 
@@ -120,26 +107,19 @@ void rope_neox(const uint i0, const uint i1, rope_params p) {
 }
 
 
-void rope_multi(const uint i0, const uint i1, rope_params p) {
-    uint ne0 = p.ncols;
-    uint ne1 = p.p_delta_rows;
-    uint ne2 = p.ne02;
-
-    if (i0 >= ne0) {
+void rope_multi(const uint i0, const uint i1, const uint i2, const uint i3, rope_params p) {
+    if (i0 >= p.ne00) {
         return;
     }
 
-    const uint i01 = i1 % ne1;
-    const uint i02 = i1 / ne1;
-
-    uint idst = i1*ne0 + i0/2;
-    const uint ix = rope_a_coord(i0/2, i01, i02, p);
+    uint idst = i0/2 + i1 * p.nb11 + i2 * p.nb12 + i3 * p.nb13;
+    const uint ix = rope_a_coord(i0/2, i1, i2, i3, p);
 
     // Fusion optimization: ROPE + VIEW + SET_ROWS.
     // The rope output is viewed as a 1D tensor and offset based on a row index in rope_data_i.
     if (p.set_rows_stride != 0) {
-        idst = i01*ne0 + i0/2;
-        idst += rope_data_i[i02].x * p.set_rows_stride;
+        idst = i1*p.nb11 + i0/2;
+        idst += rope_data_i[i2].x * p.set_rows_stride;
     }
 
     if (i0 >= p.n_dims) {
@@ -156,26 +136,26 @@ void rope_multi(const uint i0, const uint i1, rope_params p) {
     float theta_base = 0.0;
     if (p.is_imrope != 0) {
         if (sector % 3 == 1 && sector < 3 * p.sections[1]) {
-            theta_base = rope_data_pos[i02 + ne2 * 1]*pow(p.theta_scale, i0/2.0f);
+            theta_base = rope_data_pos[i2 + p.ne02 * 1]*pow(p.theta_scale, i0/2.0f);
         } else if (sector % 3 == 2 && sector < 3 * p.sections[2]) {
-            theta_base = rope_data_pos[i02 + ne2 * 2]*pow(p.theta_scale, i0/2.0f);
+            theta_base = rope_data_pos[i2 + p.ne02 * 2]*pow(p.theta_scale, i0/2.0f);
         } else if (sector % 3 == 0 && sector < 3 * p.sections[0]) {
-            theta_base = rope_data_pos[i02]*pow(p.theta_scale, i0/2.0f);
+            theta_base = rope_data_pos[i2]*pow(p.theta_scale, i0/2.0f);
         } else {
-            theta_base = rope_data_pos[i02 + ne2 * 3]*pow(p.theta_scale, i0/2.0f);
+            theta_base = rope_data_pos[i2 + p.ne02 * 3]*pow(p.theta_scale, i0/2.0f);
         }
     } else {
         if (sector < p.sections[0]) {
-            theta_base = rope_data_pos[i02]*pow(p.theta_scale, i0/2.0f);
+            theta_base = rope_data_pos[i2]*pow(p.theta_scale, i0/2.0f);
         }
         else if (sector >= p.sections[0] && sector < sec_w) {
-            theta_base = rope_data_pos[i02 + ne2 * 1]*pow(p.theta_scale, i0/2.0f);
+            theta_base = rope_data_pos[i2 + p.ne02 * 1]*pow(p.theta_scale, i0/2.0f);
         }
         else if (sector >= sec_w && sector < sec_w + p.sections[2]) {
-            theta_base = rope_data_pos[i02 + ne2 * 2]*pow(p.theta_scale, i0/2.0f);
+            theta_base = rope_data_pos[i2 + p.ne02 * 2]*pow(p.theta_scale, i0/2.0f);
         }
         else if (sector >= sec_w + p.sections[2]) {
-            theta_base = rope_data_pos[i02 + ne2 * 3]*pow(p.theta_scale, i0/2.0f);
+            theta_base = rope_data_pos[i2 + p.ne02 * 3]*pow(p.theta_scale, i0/2.0f);
         }
     }
 
@@ -191,20 +171,13 @@ void rope_multi(const uint i0, const uint i1, rope_params p) {
     rope_data_d[idst + p.n_dims/2] = ROPE_D_TYPE(x0*sin_theta + x1*cos_theta);
 }
 
-void rope_vision(const uint i0, const uint i1, rope_params p) {
-    uint ne0 = p.ncols;
-    uint ne1 = p.p_delta_rows;
-    uint ne2 = p.ne02;
-
-    if (i0 >= ne0) {
+void rope_vision(const uint i0, const uint i1, const uint i2, const uint i3, rope_params p) {
+    if (i0 >= p.ne00) {
         return;
     }
 
-    const uint i01 = i1 % ne1;
-    const uint i02 = i1 / ne1;
-
-    const uint idst = i1*ne0 + i0/2;
-    const uint ix = rope_a_coord(i0/2, i01, i02, p);
+    const uint idst = i0/2 + i1 * p.nb11 + i2 * p.nb12 + i3 * p.nb13;
+    const uint ix = rope_a_coord(i0/2, i1, i2, i3, p);
 
     const int sect_dims = p.sections[0] + p.sections[1];
     const int sec_w = p.sections[1] + p.sections[0];
@@ -213,11 +186,11 @@ void rope_vision(const uint i0, const uint i1, rope_params p) {
     float theta_base = 0.0;
     if (sector < p.sections[0]) {
         const uint p0 = sector;
-        theta_base = rope_data_pos[i02]*pow(p.theta_scale, p0);
+        theta_base = rope_data_pos[i2]*pow(p.theta_scale, p0);
     }
     else if (sector >= p.sections[0] && sector < sec_w) {
         const uint p0 = sector - p.sections[0];
-        theta_base = rope_data_pos[i02 + ne2]*pow(p.theta_scale, p0);
+        theta_base = rope_data_pos[i2 + p.ne02]*pow(p.theta_scale, p0);
     }
 
     const float freq_factor = p.has_ff != 0 ? rope_data_ff[i0/2] : 1.0f;

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

@@ -5,10 +5,13 @@
 
 void main() {
     const uint i0 = 2*gl_GlobalInvocationID.y;
-    // i1 is actually i2*nb2+i1, but the rows are contiguous
-    const uint i1 = gl_GlobalInvocationID.x + 32768 * gl_GlobalInvocationID.z;
-    if (i1 >= pc.nrows) {
+    const uint row = gl_GlobalInvocationID.x + 32768 * gl_GlobalInvocationID.z;
+    if (row >= pc.nrows) {
         return;
     }
-    rope_multi(i0, i1, pc);
+    const uint i3 = row / (pc.ne01*pc.ne02);
+    const uint i2 = (row - i3 * pc.ne01*pc.ne02) / pc.ne01;
+    const uint i1 = (row - i3 * pc.ne01*pc.ne02 - i2 * pc.ne01);
+
+    rope_multi(i0, i1, i2, i3, pc);
 }

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

@@ -5,10 +5,13 @@
 
 void main() {
     const uint i0 = 2*gl_GlobalInvocationID.y;
-    // i1 is actually i2*nb2+i1, but the rows are contiguous
-    const uint i1 = gl_GlobalInvocationID.x + 32768 * gl_GlobalInvocationID.z;
-    if (i1 >= pc.nrows) {
+    const uint row = gl_GlobalInvocationID.x + 32768 * gl_GlobalInvocationID.z;
+    if (row >= pc.nrows) {
         return;
     }
-    rope_neox(i0, i1, pc);
+    const uint i3 = row / (pc.ne01*pc.ne02);
+    const uint i2 = (row - i3 * pc.ne01*pc.ne02) / pc.ne01;
+    const uint i1 = (row - i3 * pc.ne01*pc.ne02 - i2 * pc.ne01);
+
+    rope_neox(i0, i1, i2, i3, pc);
 }

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

@@ -5,10 +5,13 @@
 
 void main() {
     const uint i0 = 2*gl_GlobalInvocationID.y;
-    // i1 is actually i2*nb2+i1, but the rows are contiguous
-    const uint i1 = gl_GlobalInvocationID.x + 32768 * gl_GlobalInvocationID.z;
-    if (i1 >= pc.nrows) {
+    const uint row = gl_GlobalInvocationID.x + 32768 * gl_GlobalInvocationID.z;
+    if (row >= pc.nrows) {
         return;
     }
-    rope_norm(i0, i1, pc);
+    const uint i3 = row / (pc.ne01*pc.ne02);
+    const uint i2 = (row - i3 * pc.ne01*pc.ne02) / pc.ne01;
+    const uint i1 = (row - i3 * pc.ne01*pc.ne02 - i2 * pc.ne01);
+
+    rope_norm(i0, i1, i2, i3, pc);
 }

ggml/src/ggml-vulkan/vulkan-shaders/rope_params.glsl

@@ -5,24 +5,29 @@
 
 struct rope_params {
     uint rope_mode;
-    uint ncols;
     uint nrows;
     uint n_dims;
     float freq_scale;
-    uint p_delta_rows;
     float freq_base;
     float ext_factor;
     float attn_factor;
     float corr_dims[2];
     float theta_scale;
     uint has_ff;
-    uint ne02;
-    uint nb01;
-    uint nb02;
     int sections[4];
     uint is_imrope;
     uint is_back;
     uint set_rows_stride;
+
+    uint ne00;
+    uint ne01;
+    uint ne02;
+    uint nb01;
+    uint nb02;
+    uint nb03;
+    uint nb11;
+    uint nb12;
+    uint nb13;
 };
 
 #endif // !defined(GGML_ROPE_PARAMS)

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

@@ -5,10 +5,13 @@
 
 void main() {
     const uint i0 = 2*gl_GlobalInvocationID.y;
-    // i1 is actually i2*nb2+i1, but the rows are contiguous
-    const uint i1 = gl_GlobalInvocationID.x + 32768 * gl_GlobalInvocationID.z;
-    if (i1 >= pc.nrows) {
+    const uint row = gl_GlobalInvocationID.x + 32768 * gl_GlobalInvocationID.z;
+    if (row >= pc.nrows) {
         return;
     }
-    rope_vision(i0, i1, pc);
+    const uint i3 = row / (pc.ne01*pc.ne02);
+    const uint i2 = (row - i3 * pc.ne01*pc.ne02) / pc.ne01;
+    const uint i1 = (row - i3 * pc.ne01*pc.ne02 - i2 * pc.ne01);
+
+    rope_vision(i0, i1, i2, i3, pc);
 }

scripts/sync_vendor.py

@@ -12,8 +12,8 @@ vendor = {
     # "https://github.com/mackron/miniaudio/raw/refs/tags/0.11.23/miniaudio.h": "vendor/miniaudio/miniaudio.h",
     "https://github.com/mackron/miniaudio/raw/669ed3e844524fcd883231b13095baee9f6de304/miniaudio.h": "vendor/miniaudio/miniaudio.h",
 
-    "https://raw.githubusercontent.com/yhirose/cpp-httplib/refs/tags/v0.30.1/httplib.h": "vendor/cpp-httplib/httplib.h",
-    "https://raw.githubusercontent.com/yhirose/cpp-httplib/refs/tags/v0.30.1/LICENSE":   "vendor/cpp-httplib/LICENSE",
+    "https://raw.githubusercontent.com/yhirose/cpp-httplib/refs/tags/v0.30.2/httplib.h": "vendor/cpp-httplib/httplib.h",
+    "https://raw.githubusercontent.com/yhirose/cpp-httplib/refs/tags/v0.30.2/LICENSE":   "vendor/cpp-httplib/LICENSE",
 
     "https://raw.githubusercontent.com/sheredom/subprocess.h/b49c56e9fe214488493021017bf3954b91c7c1f5/subprocess.h": "vendor/sheredom/subprocess.h",
 }

src/models/qwen3next.cpp

@@ -265,9 +265,15 @@ std::pair<ggml_tensor *, ggml_tensor *> llm_build_qwen3next::build_delta_net_chu
     cb(g_diff, "g_diff", il); // shape: (chunk_size, 1, n_chunks, H_v * n_seqs)
 
     ggml_tensor * g_diff_exp = ggml_exp(ctx0, g_diff);
-    ggml_tensor * key_gdiff = ggml_mul(ctx0, k, g_diff_exp);
+    ggml_tensor * g_diff_exp_t = ggml_reshape_4d(ctx0, g_diff_exp,
+                                                 1, chunk_size, n_chunks, g_diff_exp->ne[3]);
+
+    ggml_tensor * key_gdiff = ggml_mul(ctx0, k, g_diff_exp_t);
     cb(key_gdiff, "key_gdiff", il); // shape: (S_k, chunk_size, n_chunks, H_v * n_seqs)
 
+    ggml_tensor * key_gdiff_t = ggml_cont(ctx0, ggml_transpose(ctx0, key_gdiff));
+    cb(key_gdiff_t, "key_gdiff_t", il); // shape: (chunk_size, S_k, n_chunks, H_v * n_seqs)
+
 
     // state to be updated per chunk
     ggml_tensor * new_state = state; // ggml_dup(ctx0, state);
@@ -322,9 +328,9 @@ std::pair<ggml_tensor *, ggml_tensor *> llm_build_qwen3next::build_delta_net_chu
             : ggml_concat(ctx0, core_attn_out, core_attn_out_chunk, 2);
 
         // kgdmulvnew = (key_gdiff).transpose(-1, -2) @ v_new
-        ggml_tensor * k_gdiff = ggml_cont(ctx0, get_slice_2d(ctx0, key_gdiff, chunk));
+        ggml_tensor * k_gdiff_t = get_slice_2d(ctx0, key_gdiff_t, chunk);
         //ggml_tensor * kgdmulvnew = ggml_mul_mat(ctx0, k_gdiff, v_new); // this is slower on metal, why?
-        ggml_tensor * kgdmulvnew = ggml_mul_mat(ctx0, v_new_t, ggml_cont(ctx0, ggml_transpose(ctx0, k_gdiff)));
+        ggml_tensor * kgdmulvnew = ggml_mul_mat(ctx0, v_new_t, k_gdiff_t);
 
         // last_recurrent_state = last_recurrent_state * g_last + kgdmulvnew
         ggml_tensor * gexp_last_chunk = ggml_cont(ctx0, get_slice_2d(ctx0, g_last_exp, chunk));

tests/test-backend-ops.cpp

@@ -8032,6 +8032,8 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
         for (int mode : {GGML_ROPE_TYPE_NORMAL, GGML_ROPE_TYPE_NEOX, GGML_ROPE_TYPE_MROPE, GGML_ROPE_TYPE_IMROPE, GGML_ROPE_TYPE_VISION}) {
             for (bool ff : {false, true}) {
                 test_cases.emplace_back(new test_rope(type, {128,  32, 2, 1}, 128, mode, 512, 1.4245f, 0.7465f, 1.4245f, ff, 0, true, true));
+                test_cases.emplace_back(new test_rope(type, {128,  32, 2, 1}, 128, mode, 512, 1.4245f, 0.7465f, 1.4245f, ff, 1, true, true));
+                test_cases.emplace_back(new test_rope(type, {128,  32, 2, 3}, 128, mode, 512, 1.4245f, 0.7465f, 1.4245f, ff, 1, true, true));
             }
         }
     }

tools/completion/completion.cpp

@@ -674,15 +674,12 @@ int main(int argc, char ** argv) {
                 }
             }
 
-            for (int i = 0; i < (int) embd.size(); i += params.n_batch) {
-                int n_eval = (int) embd.size() - i;
-                if (n_eval > params.n_batch) {
-                    n_eval = params.n_batch;
-                }
-
+            if (!embd.empty()) {
+                int n_eval = (int) embd.size();
                 LOG_DBG("eval: %s\n", string_from(ctx, embd).c_str());
 
-                if (llama_decode(ctx, llama_batch_get_one(&embd[i], n_eval))) {
+                GGML_ASSERT(n_eval <= params.n_batch);
+                if (llama_decode(ctx, llama_batch_get_one(embd.data(), n_eval))) {
                     LOG_ERR("%s : failed to eval\n", __func__);
                     return 1;
                 }
@@ -743,7 +740,7 @@ int main(int argc, char ** argv) {
                 common_sampler_accept(smpl, embd_inp[n_consumed], /* accept_grammar= */ false);
 
                 ++n_consumed;
-                if ((int) embd.size() >= params.n_batch) {
+                if ((int) embd.size() == params.n_batch) {
                     break;
                 }
             }

vendor/cpp-httplib/CMakeLists.txt

@@ -3,9 +3,14 @@ license_add_file("cpp-httplib" "LICENSE")
 
 find_package(Threads REQUIRED)
 
+llama_add_compile_flags()
+
 add_library(${TARGET} STATIC httplib.cpp httplib.h)
-if (NOT MSVC)
-    # disable warnings in 3rd party code
+
+# disable warnings in 3rd party code
+if (CMAKE_CXX_COMPILER_ID STREQUAL "MSVC")
+    target_compile_options(${TARGET} PRIVATE /w)
+else()
     target_compile_options(${TARGET} PRIVATE -w)
 endif()
 
@@ -146,6 +151,23 @@ elseif (LLAMA_OPENSSL)
     endif()
 endif()
 
+# disable warnings in 3rd party code
+if(LLAMA_BUILD_BORINGSSL OR LLAMA_BUILD_LIBRESSL)
+    if (CMAKE_CXX_COMPILER_ID STREQUAL "MSVC")
+        target_compile_options(ssl PRIVATE /w)
+        target_compile_options(crypto PRIVATE /w)
+        if(LLAMA_BUILD_BORINGSSL)
+            target_compile_options(fipsmodule PRIVATE /w)
+        endif()
+    else()
+        target_compile_options(ssl PRIVATE -w)
+        target_compile_options(crypto PRIVATE -w)
+        if(LLAMA_BUILD_BORINGSSL)
+            target_compile_options(fipsmodule PRIVATE -w)
+        endif()
+    endif()
+endif()
+
 if (CPPHTTPLIB_OPENSSL_SUPPORT)
     target_compile_definitions(${TARGET} PUBLIC CPPHTTPLIB_OPENSSL_SUPPORT) # used in server.cpp
     if (APPLE AND CMAKE_SYSTEM_NAME STREQUAL "Darwin")

vendor/cpp-httplib/httplib.cpp

@@ -117,6 +117,8 @@ time_t parse_http_date(const std::string &date_str) {
 
 #ifdef _WIN32
   return _mkgmtime(&tm_buf);
+#elif defined _AIX
+  return mktime(&tm_buf);
 #else
   return timegm(&tm_buf);
 #endif
@@ -1376,7 +1378,7 @@ int getaddrinfo_with_timeout(const char *node, const char *service,
 
   // Allocate on the heap, so the resolver thread can keep using the data.
   auto state = std::make_shared<GetAddrInfoState>();
-  state->node = node;
+  if (node) { state->node = node; }
   state->service = service;
   state->hints = *hints;
 
@@ -2896,10 +2898,20 @@ bool parse_range_header(const std::string &s, Ranges &ranges) try {
         return;
       }
 
-      const auto first =
-          static_cast<ssize_t>(lhs.empty() ? -1 : std::stoll(lhs));
-      const auto last =
-          static_cast<ssize_t>(rhs.empty() ? -1 : std::stoll(rhs));
+      ssize_t first = -1;
+      if (!lhs.empty()) {
+        ssize_t v;
+        auto res = detail::from_chars(lhs.data(), lhs.data() + lhs.size(), v);
+        if (res.ec == std::errc{}) { first = v; }
+      }
+
+      ssize_t last = -1;
+      if (!rhs.empty()) {
+        ssize_t v;
+        auto res = detail::from_chars(rhs.data(), rhs.data() + rhs.size(), v);
+        if (res.ec == std::errc{}) { last = v; }
+      }
+
       if ((first == -1 && last == -1) ||
           (first != -1 && last != -1 && first > last)) {
         all_valid_ranges = false;
@@ -2974,25 +2986,17 @@ bool parse_accept_header(const std::string &s,
         return;
       }
 
-#ifdef CPPHTTPLIB_NO_EXCEPTIONS
       {
-        std::istringstream iss(quality_str);
-        iss >> accept_entry.quality;
-
-        // Check if conversion was successful and entire string was consumed
-        if (iss.fail() || !iss.eof()) {
+        double v = 0.0;
+        auto res = detail::from_chars(
+            quality_str.data(), quality_str.data() + quality_str.size(), v);
+        if (res.ec == std::errc{}) {
+          accept_entry.quality = v;
+        } else {
           has_invalid_entry = true;
           return;
         }
       }
-#else
-      try {
-        accept_entry.quality = std::stod(quality_str);
-      } catch (...) {
-        has_invalid_entry = true;
-        return;
-      }
-#endif
       // Check if quality is in valid range [0.0, 1.0]
       if (accept_entry.quality < 0.0 || accept_entry.quality > 1.0) {
         has_invalid_entry = true;
@@ -5570,13 +5574,26 @@ bool Server::read_content(Stream &strm, Request &req, Response &res) {
           strm, req, res,
           // Regular
           [&](const char *buf, size_t n) {
+            // Prevent arithmetic overflow when checking sizes.
+            // Avoid computing (req.body.size() + n) directly because
+            // adding two unsigned `size_t` values can wrap around and
+            // produce a small result instead of indicating overflow.
+            // Instead, check using subtraction: ensure `n` does not
+            // exceed the remaining capacity `max_size() - size()`.
+            if (req.body.size() >= req.body.max_size() ||
+                n > req.body.max_size() - req.body.size()) {
+              return false;
+            }
+
             // Limit decompressed body size to payload_max_length_ to protect
             // against "zip bomb" attacks where a small compressed payload
             // decompresses to a massive size.
-            if (req.body.size() + n > payload_max_length_ ||
-                req.body.size() + n > req.body.max_size()) {
+            if (payload_max_length_ > 0 &&
+                (req.body.size() >= payload_max_length_ ||
+                 n > payload_max_length_ - req.body.size())) {
               return false;
             }
+
             req.body.append(buf, n);
             return true;
           },
@@ -5666,22 +5683,29 @@ bool Server::read_content_core(
   // oversized request and fail early (causing connection close). For SSL
   // builds we cannot reliably peek the decrypted application bytes, so keep
   // the original behaviour.
-#if !defined(CPPHTTPLIB_OPENSSL_SUPPORT) && !defined(_WIN32)
+#if !defined(CPPHTTPLIB_OPENSSL_SUPPORT)
   if (!req.has_header("Content-Length") &&
       !detail::is_chunked_transfer_encoding(req.headers)) {
-    socket_t s = strm.socket();
-    if (s != INVALID_SOCKET) {
-      // Peek up to payload_max_length_ + 1 bytes. If more than
-      // payload_max_length_ bytes are pending, reject the request.
-      size_t to_peek =
-          (payload_max_length_ > 0)
-              ? (std::min)(payload_max_length_ + 1, static_cast<size_t>(4096))
-              : 1;
-      std::vector<char> peekbuf(to_peek);
-      ssize_t n = ::recv(s, peekbuf.data(), to_peek, MSG_PEEK);
-      if (n > 0 && static_cast<size_t>(n) > payload_max_length_) {
-        // Indicate failure so connection will be closed.
-        return false;
+    // Only peek if payload_max_length is set to a finite value
+    if (payload_max_length_ > 0 &&
+        payload_max_length_ < (std::numeric_limits<size_t>::max)()) {
+      socket_t s = strm.socket();
+      if (s != INVALID_SOCKET) {
+        // Peek to check if there is any pending data
+        char peekbuf[1];
+        ssize_t n = ::recv(s, peekbuf, 1, MSG_PEEK);
+        if (n > 0) {
+          // There is data, so read it with payload limit enforcement
+          auto result = detail::read_content_without_length(
+              strm, payload_max_length_, out);
+          if (result == detail::ReadContentResult::PayloadTooLarge) {
+            res.status = StatusCode::PayloadTooLarge_413;
+            return false;
+          } else if (result != detail::ReadContentResult::Success) {
+            return false;
+          }
+          return true;
+        }
       }
     }
     return true;
@@ -6656,7 +6680,8 @@ void ClientImpl::close_socket(Socket &socket) {
 }
 
 bool ClientImpl::read_response_line(Stream &strm, const Request &req,
-                                           Response &res) const {
+                                           Response &res,
+                                           bool skip_100_continue) const {
   std::array<char, 2048> buf{};
 
   detail::stream_line_reader line_reader(strm, buf.data(), buf.size());
@@ -6677,8 +6702,8 @@ bool ClientImpl::read_response_line(Stream &strm, const Request &req,
   res.status = std::stoi(std::string(m[2]));
   res.reason = std::string(m[3]);
 
-  // Ignore '100 Continue'
-  while (res.status == StatusCode::Continue_100) {
+  // Ignore '100 Continue' (only when not using Expect: 100-continue explicitly)
+  while (skip_100_continue && res.status == StatusCode::Continue_100) {
     if (!line_reader.getline()) { return false; } // CRLF
     if (!line_reader.getline()) { return false; } // next response line
 
@@ -7463,7 +7488,8 @@ bool ClientImpl::write_content_with_provider(Stream &strm,
 }
 
 bool ClientImpl::write_request(Stream &strm, Request &req,
-                                      bool close_connection, Error &error) {
+                                      bool close_connection, Error &error,
+                                      bool skip_body) {
   // Prepare additional headers
   if (close_connection) {
     if (!req.has_header("Connection")) {
@@ -7582,7 +7608,59 @@ bool ClientImpl::write_request(Stream &strm, Request &req,
     }
   }
 
+  // After sending request line and headers, wait briefly for an early server
+  // response (e.g. 4xx) and avoid sending a potentially large request body
+  // unnecessarily. This workaround is only enabled on Windows because Unix
+  // platforms surface write errors (EPIPE) earlier; on Windows kernel send
+  // buffering can accept large writes even when the peer already responded.
+  // Check the stream first (which covers SSL via `is_readable()`), then
+  // fall back to select on the socket. Only perform the wait for very large
+  // request bodies to avoid interfering with normal small requests and
+  // reduce side-effects. Poll briefly (up to 50ms as default) for an early
+  // response. Skip this check when using Expect: 100-continue, as the protocol
+  // handles early responses properly.
+#if defined(_WIN32)
+  if (!skip_body &&
+      req.body.size() > CPPHTTPLIB_WAIT_EARLY_SERVER_RESPONSE_THRESHOLD &&
+      req.path.size() > CPPHTTPLIB_REQUEST_URI_MAX_LENGTH) {
+    auto start = std::chrono::high_resolution_clock::now();
+
+    for (;;) {
+      // Prefer socket-level readiness to avoid SSL_pending() false-positives
+      // from SSL internals. If the underlying socket is readable, assume an
+      // early response may be present.
+      auto sock = strm.socket();
+      if (sock != INVALID_SOCKET && detail::select_read(sock, 0, 0) > 0) {
+        return false;
+      }
+
+      // Fallback to stream-level check for non-socket streams or when the
+      // socket isn't reporting readable. Avoid using `is_readable()` for
+      // SSL, since `SSL_pending()` may report buffered records that do not
+      // indicate a complete application-level response yet.
+      if (!is_ssl() && strm.is_readable()) { return false; }
+
+      auto now = std::chrono::high_resolution_clock::now();
+      auto elapsed =
+          std::chrono::duration_cast<std::chrono::milliseconds>(now - start)
+              .count();
+      if (elapsed >= CPPHTTPLIB_WAIT_EARLY_SERVER_RESPONSE_TIMEOUT_MSECOND) {
+        break;
+      }
+
+      std::this_thread::sleep_for(std::chrono::milliseconds(1));
+    }
+  }
+#endif
+
   // Body
+  if (skip_body) { return true; }
+
+  return write_request_body(strm, req, error);
+}
+
+bool ClientImpl::write_request_body(Stream &strm, Request &req,
+                                           Error &error) {
   if (req.body.empty()) {
     return write_content_with_provider(strm, req, error);
   }
@@ -7758,8 +7836,20 @@ void ClientImpl::output_error_log(const Error &err,
 bool ClientImpl::process_request(Stream &strm, Request &req,
                                         Response &res, bool close_connection,
                                         Error &error) {
-  // Send request
-  if (!write_request(strm, req, close_connection, error)) { return false; }
+  // Auto-add Expect: 100-continue for large bodies
+  if (CPPHTTPLIB_EXPECT_100_THRESHOLD > 0 && !req.has_header("Expect")) {
+    auto body_size = req.body.empty() ? req.content_length_ : req.body.size();
+    if (body_size >= CPPHTTPLIB_EXPECT_100_THRESHOLD) {
+      req.set_header("Expect", "100-continue");
+    }
+  }
+
+  // Check for Expect: 100-continue
+  auto expect_100_continue = req.get_header_value("Expect") == "100-continue";
+
+  // Send request (skip body if using Expect: 100-continue)
+  auto write_request_success =
+      write_request(strm, req, close_connection, error, expect_100_continue);
 
 #ifdef CPPHTTPLIB_OPENSSL_SUPPORT
   if (is_ssl()) {
@@ -7774,14 +7864,48 @@ bool ClientImpl::process_request(Stream &strm, Request &req,
   }
 #endif
 
+  // Handle Expect: 100-continue with timeout
+  if (expect_100_continue && CPPHTTPLIB_EXPECT_100_TIMEOUT_MSECOND > 0) {
+    time_t sec = CPPHTTPLIB_EXPECT_100_TIMEOUT_MSECOND / 1000;
+    time_t usec = (CPPHTTPLIB_EXPECT_100_TIMEOUT_MSECOND % 1000) * 1000;
+    auto ret = detail::select_read(strm.socket(), sec, usec);
+    if (ret <= 0) {
+      // Timeout or error: send body anyway (server didn't respond in time)
+      if (!write_request_body(strm, req, error)) { return false; }
+      expect_100_continue = false; // Switch to normal response handling
+    }
+  }
+
   // Receive response and headers
-  if (!read_response_line(strm, req, res) ||
+  // When using Expect: 100-continue, don't auto-skip `100 Continue` response
+  if (!read_response_line(strm, req, res, !expect_100_continue) ||
       !detail::read_headers(strm, res.headers)) {
-    error = Error::Read;
+    if (write_request_success) { error = Error::Read; }
     output_error_log(error, &req);
     return false;
   }
 
+  if (!write_request_success) { return false; }
+
+  // Handle Expect: 100-continue response
+  if (expect_100_continue) {
+    if (res.status == StatusCode::Continue_100) {
+      // Server accepted, send the body
+      if (!write_request_body(strm, req, error)) { return false; }
+
+      // Read the actual response
+      res.headers.clear();
+      res.body.clear();
+      if (!read_response_line(strm, req, res) ||
+          !detail::read_headers(strm, res.headers)) {
+        error = Error::Read;
+        output_error_log(error, &req);
+        return false;
+      }
+    }
+    // If not 100 Continue, server returned an error; proceed with that response
+  }
+
   // Body
   if ((res.status != StatusCode::NoContent_204) && req.method != "HEAD" &&
       req.method != "CONNECT") {
@@ -9543,7 +9667,7 @@ bool SSLClient::load_certs() {
         last_openssl_error_ = ERR_get_error();
         ret = false;
       }
-    } else {
+    } else if (!ca_cert_store_) {
       auto loaded = false;
 #ifdef _WIN32
       loaded =
@@ -9790,7 +9914,11 @@ bool SSLClient::verify_host_with_common_name(X509 *server_cert) const {
 
 bool SSLClient::check_host_name(const char *pattern,
                                        size_t pattern_len) const {
-  if (host_.size() == pattern_len && host_ == pattern) { return true; }
+  // Exact match (case-insensitive)
+  if (host_.size() == pattern_len &&
+      detail::case_ignore::equal(host_, std::string(pattern, pattern_len))) {
+    return true;
+  }
 
   // Wildcard match
   // https://bugs.launchpad.net/ubuntu/+source/firefox-3.0/+bug/376484
@@ -9805,9 +9933,23 @@ bool SSLClient::check_host_name(const char *pattern,
   auto itr = pattern_components.begin();
   for (const auto &h : host_components_) {
     auto &p = *itr;
-    if (p != h && p != "*") {
-      auto partial_match = (p.size() > 0 && p[p.size() - 1] == '*' &&
-                            !p.compare(0, p.size() - 1, h));
+    if (!httplib::detail::case_ignore::equal(p, h) && p != "*") {
+      bool partial_match = false;
+      if (!p.empty() && p[p.size() - 1] == '*') {
+        const auto prefix_length = p.size() - 1;
+        if (prefix_length == 0) {
+          partial_match = true;
+        } else if (h.size() >= prefix_length) {
+          partial_match =
+              std::equal(p.begin(),
+                         p.begin() + static_cast<std::string::difference_type>(
+                                         prefix_length),
+                         h.begin(), [](const char ca, const char cb) {
+                           return httplib::detail::case_ignore::to_lower(ca) ==
+                                  httplib::detail::case_ignore::to_lower(cb);
+                         });
+        }
+      }
       if (!partial_match) { return false; }
     }
     ++itr;

vendor/cpp-httplib/httplib.h

@@ -8,8 +8,8 @@
 #ifndef CPPHTTPLIB_HTTPLIB_H
 #define CPPHTTPLIB_HTTPLIB_H
 
-#define CPPHTTPLIB_VERSION "0.30.1"
-#define CPPHTTPLIB_VERSION_NUM "0x001E01"
+#define CPPHTTPLIB_VERSION "0.30.2"
+#define CPPHTTPLIB_VERSION_NUM "0x001E02"
 
 /*
  * Platform compatibility check
@@ -98,6 +98,22 @@
 #define CPPHTTPLIB_CLIENT_MAX_TIMEOUT_MSECOND 0
 #endif
 
+#ifndef CPPHTTPLIB_EXPECT_100_THRESHOLD
+#define CPPHTTPLIB_EXPECT_100_THRESHOLD 1024
+#endif
+
+#ifndef CPPHTTPLIB_EXPECT_100_TIMEOUT_MSECOND
+#define CPPHTTPLIB_EXPECT_100_TIMEOUT_MSECOND 1000
+#endif
+
+#ifndef CPPHTTPLIB_WAIT_EARLY_SERVER_RESPONSE_THRESHOLD
+#define CPPHTTPLIB_WAIT_EARLY_SERVER_RESPONSE_THRESHOLD (1024 * 1024)
+#endif
+
+#ifndef CPPHTTPLIB_WAIT_EARLY_SERVER_RESPONSE_TIMEOUT_MSECOND
+#define CPPHTTPLIB_WAIT_EARLY_SERVER_RESPONSE_TIMEOUT_MSECOND 50
+#endif
+
 #ifndef CPPHTTPLIB_IDLE_INTERVAL_SECOND
 #define CPPHTTPLIB_IDLE_INTERVAL_SECOND 0
 #endif
@@ -286,8 +302,10 @@ using socket_t = int;
 #include <atomic>
 #include <cassert>
 #include <cctype>
+#include <chrono>
 #include <climits>
 #include <condition_variable>
+#include <cstdlib>
 #include <cstring>
 #include <errno.h>
 #include <exception>
@@ -305,6 +323,7 @@ using socket_t = int;
 #include <sstream>
 #include <string>
 #include <sys/stat.h>
+#include <system_error>
 #include <thread>
 #include <unordered_map>
 #include <unordered_set>
@@ -494,6 +513,69 @@ private:
   bool execute_on_destruction;
 };
 
+// Simple from_chars implementation for integer and double types (C++17
+// substitute)
+template <typename T> struct from_chars_result {
+  const char *ptr;
+  std::errc ec;
+};
+
+template <typename T>
+inline from_chars_result<T> from_chars(const char *first, const char *last,
+                                       T &value, int base = 10) {
+  value = 0;
+  const char *p = first;
+  bool negative = false;
+
+  if (p != last && *p == '-') {
+    negative = true;
+    ++p;
+  }
+  if (p == last) { return {first, std::errc::invalid_argument}; }
+
+  T result = 0;
+  for (; p != last; ++p) {
+    char c = *p;
+    int digit = -1;
+    if ('0' <= c && c <= '9') {
+      digit = c - '0';
+    } else if ('a' <= c && c <= 'z') {
+      digit = c - 'a' + 10;
+    } else if ('A' <= c && c <= 'Z') {
+      digit = c - 'A' + 10;
+    } else {
+      break;
+    }
+
+    if (digit < 0 || digit >= base) { break; }
+    if (result > ((std::numeric_limits<T>::max)() - digit) / base) {
+      return {p, std::errc::result_out_of_range};
+    }
+    result = result * base + digit;
+  }
+
+  if (p == first || (negative && p == first + 1)) {
+    return {first, std::errc::invalid_argument};
+  }
+
+  value = negative ? -result : result;
+  return {p, std::errc{}};
+}
+
+// from_chars for double (simple wrapper for strtod)
+inline from_chars_result<double> from_chars(const char *first, const char *last,
+                                            double &value) {
+  std::string s(first, last);
+  char *endptr = nullptr;
+  errno = 0;
+  value = std::strtod(s.c_str(), &endptr);
+  if (endptr == s.c_str()) { return {first, std::errc::invalid_argument}; }
+  if (errno == ERANGE) {
+    return {first + (endptr - s.c_str()), std::errc::result_out_of_range};
+  }
+  return {first + (endptr - s.c_str()), std::errc{}};
+}
+
 } // namespace detail
 
 enum SSLVerifierResponse {
@@ -1848,10 +1930,11 @@ private:
   Result send_(Request &&req);
 
   socket_t create_client_socket(Error &error) const;
-  bool read_response_line(Stream &strm, const Request &req,
-                          Response &res) const;
+  bool read_response_line(Stream &strm, const Request &req, Response &res,
+                          bool skip_100_continue = true) const;
   bool write_request(Stream &strm, Request &req, bool close_connection,
-                     Error &error);
+                     Error &error, bool skip_body = false);
+  bool write_request_body(Stream &strm, Request &req, Error &error);
   void prepare_default_headers(Request &r, bool for_stream,
                                const std::string &ct);
   bool redirect(Request &req, Response &res, Error &error);
@@ -3243,10 +3326,11 @@ private:
       msg.id = value;
     } else if (field == "retry") {
       // Parse retry interval in milliseconds
-      try {
-        retry_ms = std::stoi(value);
-      } catch (...) {
-        // Invalid retry value, ignore
+      {
+        int v = 0;
+        auto res =
+            detail::from_chars(value.data(), value.data() + value.size(), v);
+        if (res.ec == std::errc{}) { retry_ms = v; }
       }
     }
     // Unknown fields are ignored per SSE spec