cuda: make im2col a little faster (#15025)

This commit removes the right alignment the `n_stream` value in the
log message in the `llama_kv_cache_unified` constructor.

The motivation for this change is to enhance the readability of log
message. Currently the output looks like this:
```console
llama_kv_cache_unified: size = 2048.00 MiB (  4096 cells,  32 layers,  1/ 1 seqs), K (f16): 1024.00 MiB, V (f16): 1024.00 MiB
```
Notice that the `n_stream` value is right aligned, which makes it a
little harder to read.

With the change in this commit the output will look like
```console
llama_kv_cache_unified: size = 2048.00 MiB (  4096 cells,  32 layers, 1/1 seqs), K (f16): 1024.00 MiB, V (f16): 1024.00 MiB
```

.github/workflows/pre-tokenizer-hashes.yml

@@ -0,0 +1,45 @@
+name: Check Pre-Tokenizer Hashes
+
+on:
+    push:
+        paths:
+            - 'convert_hf_to_gguf.py'
+            - 'convert_hf_to_gguf_update.py'
+    pull_request:
+        paths:
+            - 'convert_hf_to_gguf.py'
+            - 'convert_hf_to_gguf_update.py'
+
+jobs:
+    pre-tokenizer-hashes:
+        runs-on: ubuntu-latest
+
+        steps:
+        - name: Checkout repository
+          uses: actions/checkout@v4
+
+        - name: Set up Python
+          uses: actions/setup-python@v5
+          with:
+              python-version: '3.11'
+
+        - name: Install Python dependencies
+          run: |
+              python3 -m venv .venv
+              .venv/bin/pip install -r requirements/requirements-convert_hf_to_gguf_update.txt
+
+        - name: Update pre-tokenizer hashes
+          run: |
+              cp convert_hf_to_gguf.py /tmp
+              .venv/bin/python convert_hf_to_gguf_update.py --check-missing
+
+        - name: Check if committed pre-tokenizer hashes matches generated version
+          run: |
+              if ! diff -q convert_hf_to_gguf.py /tmp/convert_hf_to_gguf.py; then
+                  echo "Model pre-tokenizer hashes (in convert_hf_to_gguf.py) do not match generated hashes (from convert_hf_to_gguf_update.py)."
+                  echo "To fix: run ./convert_hf_to_gguf_update.py and commit the updated convert_hf_to_gguf.py along with your changes"
+                  echo "Differences found:"
+                  diff convert_hf_to_gguf.py /tmp/convert_hf_to_gguf.py || true
+                  exit 1
+              fi
+              echo "Model pre-tokenizer hashes are up to date."

common/chat.cpp

@@ -1646,7 +1646,7 @@ static void common_chat_parse_hermes_2_pro(common_chat_msg_parser & builder) {
         "|<function name=\"([^\"]+)\">"  // match 5 (function name again)
     );
 
-    if (auto res = builder.try_find_regex(open_regex)) {
+    while (auto res = builder.try_find_regex(open_regex)) {
         const auto & block_start = res->groups[1];
         std::string block_end = block_start.empty() ? "" : "```";
 
@@ -1668,7 +1668,6 @@ static void common_chat_parse_hermes_2_pro(common_chat_msg_parser & builder) {
                     builder.consume_literal(block_end);
                     builder.consume_spaces();
                 }
-                builder.add_content(builder.consume_rest());
             } else {
                 throw common_chat_msg_partial_exception("failed to parse tool call");
             }
@@ -1693,11 +1692,10 @@ static void common_chat_parse_hermes_2_pro(common_chat_msg_parser & builder) {
                     builder.consume_spaces();
                 }
             }
-            builder.add_content(builder.consume_rest());
         }
-    } else {
-        builder.add_content(builder.consume_rest());
     }
+
+    builder.add_content(builder.consume_rest());
 }
 
 static common_chat_params common_chat_params_init_without_tools(const common_chat_template & tmpl, const struct templates_params & inputs) {

convert_hf_to_gguf.py

@@ -702,6 +702,9 @@ class TextModel(ModelBase):
         if chkhsh == "81212dc7cdb7e0c1074ca62c5aeab0d43c9f52b8a737be7b12a777c953027890":
             # ref: https://huggingface.co/moonshotai/Kimi-K2-Base
             res = "kimi-k2"
+        if chkhsh == "d4540891389ea895b53b399da6ac824becc30f2fba0e9ddbb98f92e55ca0e97c":
+            # ref: https://huggingface.co/Qwen/Qwen3-Embedding-0.6B
+            res = "qwen2"
         if chkhsh == "0ef9807a4087ebef797fc749390439009c3b9eda9ad1a097abbe738f486c01e5":
             # ref: https://huggingface.co/meta-llama/Meta-Llama-3-8B
             res = "llama-bpe"

convert_hf_to_gguf_update.py

@@ -59,6 +59,10 @@ parser.add_argument(
     "--full", action="store_true",
     help="download full list of models - make sure you have access to all of them",
 )
+parser.add_argument(
+    "--check-missing", action="store_true",
+    help="only check for missing pre-tokenizer hashes",
+)
 parser.add_argument(
     "hf_token",
     help="optional HF token",
@@ -70,6 +74,10 @@ hf_token = args.hf_token if args.hf_token is not None else hf_token
 if hf_token is None:
     logger.warning("HF token not found. You can provide it as an argument or set it in ~/.cache/huggingface/token")
 
+if args.check_missing and args.full:
+    logger.warning("Downloading full list of models requested, ignoring --check-missing!")
+    args.check_missing = False
+
 # TODO: this string has to exercise as much pre-tokenizer functionality as possible
 #       will be updated with time - contributions welcome
 CHK_TXT = '\n \n\n \n\n\n \t \t\t \t\n  \n   \n    \n     \n🚀 (normal) 😶‍🌫️ (multiple emojis concatenated) ✅ 🦙🦙 3 33 333 3333 33333 333333 3333333 33333333 3.3 3..3 3...3 កាន់តែពិសេសអាច😁 ?我想在apple工作1314151天～ ------======= нещо на Български \'\'\'\'\'\'```````\"\"\"\"......!!!!!!?????? I\'ve been \'told he\'s there, \'RE you sure? \'M not sure I\'ll make it, \'D you like some tea? We\'Ve a\'lL'
@@ -147,6 +155,7 @@ pre_computed_hashes = [
     {"name": "falcon-h1", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/tiiuae/Falcon-H1-7B-Base", "chkhsh": "3eda48b4c4dc7de733d1a8b3e3b4a85243dbbf704da2ee9d42c6beced8897896"},
     {"name": "falcon-h1", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/tiiuae/Falcon-H1-34B-Base", "chkhsh": "48f8e02c0359c0bbdd82f26909171fac1c18a457bb47573ed1fe3bbb2c1cfd4b"},
     {"name": "kimi-k2",   "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/moonshotai/Kimi-K2-Base",   "chkhsh": "81212dc7cdb7e0c1074ca62c5aeab0d43c9f52b8a737be7b12a777c953027890"},
+    {"name": "qwen2",     "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/Qwen/Qwen3-Embedding-0.6B", "chkhsh": "d4540891389ea895b53b399da6ac824becc30f2fba0e9ddbb98f92e55ca0e97c"},
 ]
 
 
@@ -221,12 +230,13 @@ if not args.full:
     all_models = models.copy()
     models = [model for model in all_models if model["name"] not in existing_models]
 
-logging.info(f"Downloading {len(models)} models...")
-for model in models:
-    try:
-        download_model(model)
-    except Exception as e:
-        logger.error(f"Failed to download model {model['name']}. Error: {e}")
+if not args.check_missing:
+    logging.info(f"Downloading {len(models)} models...")
+    for model in models:
+        try:
+            download_model(model)
+        except Exception as e:
+            logger.error(f"Failed to download model {model['name']}. Error: {e}")
 
 
 # generate the source code for the convert_hf_to_gguf.py:get_vocab_base_pre() function:

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

@@ -1852,6 +1852,9 @@ static void ggml_cuda_mul_mat_batched_cublas_impl(ggml_backend_cuda_context & ct
     ggml_cuda_pool_alloc<cuda_t> src0_alloc(ctx.pool());
     ggml_cuda_pool_alloc<cuda_t> src1_alloc(ctx.pool());
 
+    bool is_src0_cont_2 = ggml_is_contiguous_2(src0);
+    bool is_src1_cont_2 = ggml_is_contiguous_2(src1);
+
     // Handle src0
     src0_ptr = (const cuda_t *) src0->data;
 
@@ -1870,6 +1873,8 @@ static void ggml_cuda_mul_mat_batched_cublas_impl(ggml_backend_cuda_context & ct
         s11 = ne10;
         s12 = ne11*s11;
         s13 = ne12*s12;
+
+        is_src1_cont_2 = true;
     }
 
     // Setup destination buffer
@@ -1918,15 +1923,19 @@ static void ggml_cuda_mul_mat_batched_cublas_impl(ggml_backend_cuda_context & ct
     const int64_t r2 = ne12/ne02;
     const int64_t r3 = ne13/ne03;
 
-    if (r2 == 1 && r3 == 1 && ggml_is_contiguous_2(src0) && ggml_is_contiguous_2(src1)) {
+    if (r2 == 1 && r3 == 1 && is_src0_cont_2 && is_src1_cont_2) {
+        // with a [0, 2, 1, 3] perm. and ne02==1 the matrix strides need to be determined from dim 3:
+        const int64_t sma = ne02 == 1 ? nb03/nb00 : nb02/nb00;
+        const int64_t smb = ne12 == 1 ? s13       : s12;
+
         // there is no broadcast and src0, src1 are contiguous across dims 2, 3
         // use cublasGemmStridedBatchedEx
         CUBLAS_CHECK(
         cublasGemmStridedBatchedEx(ctx.cublas_handle(), CUBLAS_OP_T, CUBLAS_OP_N,
                 ne01, ne11, ne10,
-                alpha, src0_ptr, cu_data_type_a, nb01/nb00, nb02/nb00, // strideA
-                       src1_ptr, cu_data_type_b, s11,       s12,       // strideB
-                beta,     dst_t, cu_data_type,   ne0,       ne1*ne0,   // strideC
+                alpha, src0_ptr, cu_data_type_a, nb01/nb00, sma,     // strideA
+                       src1_ptr, cu_data_type_b, s11,       smb,     // strideB
+                beta,     dst_t, cu_data_type,   ne0,       ne1*ne0, // strideC
                 ne12*ne13,
                 cu_compute_type,
                 CUBLAS_GEMM_DEFAULT_TENSOR_OP));

ggml/src/ggml-cuda/im2col.cu

@@ -1,65 +1,75 @@
 #include "im2col.cuh"
 
+#define MIN(a, b) (a) < (b) ? (a) : (b)
+
+#define MAX_GRIDDIM_Z 65535
+
 template <typename T>
 static  __global__ void im2col_kernel(
-        const float * x, T * dst, int64_t batch_offset,
-        int64_t offset_delta, int64_t IC, int64_t IW, int64_t IH, int64_t OH, int64_t OW, int64_t KW, int64_t KH, int64_t pelements, int64_t CHW,
+        const float * x, T * dst,
+        int64_t IC, int64_t IW, int64_t IH, int64_t OH, int64_t OW, int64_t KW, int64_t KH,
+        int64_t IC_IH_IW, int64_t IH_IW, int64_t N_OH, int64_t KH_KW, int64_t IC_KH_KW,
         int s0, int s1, int p0, int p1, int d0, int d1) {
     const int64_t i = threadIdx.x + blockIdx.x * blockDim.x;
-    if (i >= pelements) {
+    if (i >= IC_KH_KW) {
         return;
     }
 
-    const int64_t  ksize = OW * KH;
-    const int64_t  kx = i / ksize;
-    const int64_t  kd = kx * ksize;
-    const int64_t  ky = (i - kd) / OW;
-    const int64_t  ix = i % OW;
+    const int64_t iic = i / (KH_KW);
+    const int64_t rem = i - iic * KH_KW;
+    const int64_t ikh = rem / KW;
+    const int64_t ikw = rem - ikh * KW;
 
-    const int64_t  oh = blockIdx.y;
-    const int64_t  batch = blockIdx.z / IC;
-    const int64_t  ic = blockIdx.z % IC;
+    const int64_t  iow = blockIdx.y;
+    for (int64_t iz = blockIdx.z; iz < N_OH; iz+=MAX_GRIDDIM_Z) {
+        const int64_t  in = iz / OH;
+        const int64_t  ioh = iz - in * OH;
 
-    const int64_t iiw = ix * s0 + kx * d0 - p0;
-    const int64_t iih = oh * s1 + ky * d1 - p1;
+        const int64_t iiw = iow * s0 + ikw * d0 - p0;
+        const int64_t iih = ioh * s1 + ikh * d1 - p1;
 
-    const int64_t offset_dst =
-        ((batch * OH + oh) * OW + ix) * CHW +
-        (ic * (KW * KH) + ky * KW + kx);
+        const int64_t offset_dst =
+            ((in * OH + ioh) * OW + iow) * IC_KH_KW + iic * KH_KW + ikh * KW + ikw;
 
-    if (iih < 0 || iih >= IH || iiw < 0 || iiw >= IW) {
-        dst[offset_dst] = 0.0f;
-    } else {
-        const int64_t offset_src = ic * offset_delta + batch * batch_offset;
-        dst[offset_dst] = x[offset_src + iih * IW + iiw];
+        if (iih < 0 || iih >= IH || iiw < 0 || iiw >= IW) {
+            dst[offset_dst] = 0.0f;
+        } else {
+            const int64_t offset_src = iic * IC_IH_IW + in * IH_IW;
+            dst[offset_dst] = x[offset_src + iih * IW + iiw];
+        }
     }
 }
 
+// im2col: [N, IC, IH, IW] => [N, OH, OW, IC*KH*KW]
 template <typename T>
 static void im2col_cuda(const float * x, T* dst,
     int64_t IW, int64_t IH, int64_t OW, int64_t OH, int64_t KW, int64_t KH, int64_t IC,
-    int64_t batch, int64_t batch_offset, int64_t offset_delta,
+    int64_t N, int64_t IC_IH_IW, int64_t IH_IW,
     int s0,int s1,int p0,int p1,int d0,int d1, cudaStream_t stream) {
-    const int parallel_elements = OW * KW * KH;
-    const int num_blocks = (parallel_elements + CUDA_IM2COL_BLOCK_SIZE - 1) / CUDA_IM2COL_BLOCK_SIZE;
-    dim3 block_nums(num_blocks, OH, batch * IC);
-    im2col_kernel<<<block_nums, CUDA_IM2COL_BLOCK_SIZE, 0, stream>>>(x, dst, batch_offset, offset_delta, IC, IW, IH, OH, OW, KW, KH, parallel_elements, (IC * KH * KW), s0, s1, p0, p1, d0, d1);
+    const int64_t IC_KH_KW = IC * KH * KW;
+    const int64_t num_blocks = (IC_KH_KW + CUDA_IM2COL_BLOCK_SIZE - 1) / CUDA_IM2COL_BLOCK_SIZE;
+    const int64_t N_OH = N * OH;
+    const int64_t KH_KW = KW*KH;
+    dim3 block_nums(num_blocks, OW, MIN(N_OH, MAX_GRIDDIM_Z));
+    im2col_kernel<<<block_nums, MIN(IC_KH_KW, CUDA_IM2COL_BLOCK_SIZE) , 0, stream>>>(x, dst, IC, IW, IH, OH, OW, KW, KH,
+                                                                                     IC_IH_IW, IH_IW, N_OH, KH_KW, IC_KH_KW,
+                                                                                     s0, s1, p0, p1, d0, d1);
 }
 
 static void im2col_cuda_f16(const float * x, half * dst,
     int64_t IW, int64_t IH, int64_t OW, int64_t OH, int64_t KW, int64_t KH, int64_t IC,
-    int64_t batch, int64_t batch_offset, int64_t offset_delta,
+    int64_t N, int64_t IC_IH_IW, int64_t IH_IW,
     int s0,int s1,int p0,int p1,int d0,int d1, cudaStream_t stream) {
 
-    im2col_cuda<half>(x, dst, IW, IH, OW, OH, KW, KH, IC, batch, batch_offset, offset_delta, s0, s1, p0, p1, d0, d1, stream);
+    im2col_cuda<half>(x, dst, IW, IH, OW, OH, KW, KH, IC, N, IC_IH_IW, IH_IW, s0, s1, p0, p1, d0, d1, stream);
 }
 
 static void im2col_cuda_f32(const float * x, float * dst,
     int64_t IW, int64_t IH, int64_t OW, int64_t OH, int64_t KW, int64_t KH, int64_t IC,
-    int64_t batch, int64_t batch_offset, int64_t offset_delta,
+    int64_t N, int64_t IC_IH_IW, int64_t IH_IW,
     int s0,int s1,int p0,int p1,int d0,int d1, cudaStream_t stream) {
 
-    im2col_cuda<float>(x, dst, IW, IH, OW, OH, KW, KH, IC, batch, batch_offset, offset_delta, s0, s1, p0, p1, d0, d1, stream);
+    im2col_cuda<float>(x, dst, IW, IH, OW, OH, KW, KH, IC, N, IC_IH_IW, IH_IW, s0, s1, p0, p1, d0, d1, stream);
 }
 
 void ggml_cuda_op_im2col(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
@@ -91,13 +101,13 @@ void ggml_cuda_op_im2col(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     const int64_t OH = is_2D ? dst->ne[2] : 1;
     const int64_t OW =         dst->ne[1];
 
-    const size_t  delta_offset = src1->nb[is_2D ? 2 : 1] / 4; // nb is byte offset, src is type float32
-    const int64_t batch        = src1->ne[is_2D ? 3 : 2];
-    const size_t  batch_offset = src1->nb[is_2D ? 3 : 2] / 4; // nb is byte offset, src is type float32
+    const int64_t IC_IH_IW = src1->nb[is_2D ? 2 : 1] / 4; // nb is byte offset, src is type float32
+    const int64_t N        = src1->ne[is_2D ? 3 : 2];
+    const int64_t IH_IW    = src1->nb[is_2D ? 3 : 2] / 4; // nb is byte offset, src is type float32
 
     if(dst->type == GGML_TYPE_F16) {
-        im2col_cuda_f16(src1_d, (half *) dst_d, IW, IH, OW, OH, KW, KH, IC, batch, batch_offset, delta_offset, s0, s1, p0, p1, d0, d1, stream);
+        im2col_cuda_f16(src1_d, (half *) dst_d, IW, IH, OW, OH, KW, KH, IC, N, IC_IH_IW, IH_IW, s0, s1, p0, p1, d0, d1, stream);
     } else {
-        im2col_cuda_f32(src1_d, (float *) dst_d, IW, IH, OW, OH, KW, KH, IC, batch, batch_offset, delta_offset, s0, s1, p0, p1, d0, d1, stream);
+        im2col_cuda_f32(src1_d, (float *) dst_d, IW, IH, OW, OH, KW, KH, IC, N, IC_IH_IW, IH_IW, s0, s1, p0, p1, d0, d1, stream);
     }
 }

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

@@ -2688,6 +2688,9 @@ static void ggml_sycl_mul_mat_batched_sycl(ggml_backend_sycl_context & ctx, cons
     const size_t       type_size_src0 = ggml_type_size(src0->type);
     const size_t       type_size_src1 = ggml_type_size(src1->type);
 
+    bool is_src0_cont_2 = ggml_is_contiguous_2(src0);
+    bool is_src1_cont_2 = ggml_is_contiguous_2(src1);
+
     // SRC1 strides
     int64_t                          s11 = nb11 / type_size_src1;
     int64_t                          s12 = nb12 / type_size_src1;
@@ -2737,6 +2740,8 @@ static void ggml_sycl_mul_mat_batched_sycl(ggml_backend_sycl_context & ctx, cons
         s11      = ne10;
         s12      = ne11 * s11;
         s13      = ne12 * s12;
+
+        is_src1_cont_2 = true;
     }
 
     ggml_sycl_pool_alloc<sycl::half> dst_f16(ctx.pool());
@@ -2852,12 +2857,16 @@ static void ggml_sycl_mul_mat_batched_sycl(ggml_backend_sycl_context & ctx, cons
     else
 #endif
     {
-        if (r2 == 1 && r3 == 1 && ggml_is_contiguous_2(src0) && ggml_is_contiguous_2(src1)) {
+        if (r2 == 1 && r3 == 1 && is_src0_cont_2 && is_src1_cont_2) {
+            // with a [0, 2, 1, 3] perm. and ne02==1 the matrix strides need to be determined from dim 3:
+            const int64_t sma = ne02 == 1 ? nb03/nb00 : nb02/nb00;
+            const int64_t smb = ne12 == 1 ? s13       : s12;
+
             // there is no broadcast and src0, src1 are contiguous across dims 2, 3
             SYCL_CHECK(CHECK_TRY_ERROR(dpct::gemm_batch(*queue, oneapi::math::transpose::trans,
                                                         oneapi::math::transpose::nontrans, ne01, ne11, ne10, alpha,
-                                                        src0_f16, dpct::library_data_t::real_half, nb01 / nb00, nb02 / nb00,
-                                                        src1_f16, dpct::library_data_t::real_half, s11, s12, beta, dst_ddf,
+                                                        src0_f16, dpct::library_data_t::real_half, nb01 / nb00, sma,
+                                                        src1_f16, dpct::library_data_t::real_half, s11, smb, beta, dst_ddf,
                                                         mkl_data_type, ne0, ne1 * ne0, ne12 * ne13, mkl_compute_type)));
         } else {
             const int ne23 = ne12 * ne13;

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

@@ -222,6 +222,7 @@ enum vk_device_architecture {
     AMD_RDNA2,
     AMD_RDNA3,
     INTEL_XE2,
+    NVIDIA_PRE_TURING,
 };
 
 // HSK x HSV
@@ -315,10 +316,33 @@ static vk_device_architecture get_device_architecture(const vk::PhysicalDevice&
             // https://www.intel.com/content/www/us/en/docs/oneapi/optimization-guide-gpu/2025-0/intel-xe-gpu-architecture.html
             return vk_device_architecture::INTEL_XE2;
         }
+    } else if (props.vendorID == VK_VENDOR_ID_NVIDIA) {
+        const std::vector<vk::ExtensionProperties> ext_props = device.enumerateDeviceExtensionProperties();
+
+        bool cooperative_matrix = false;
+
+        // Detect "pre-turing" based on lack of coopmat support.
+        for (const auto& properties : ext_props) {
+            if (strcmp("VK_KHR_cooperative_matrix", properties.extensionName) == 0) {
+                cooperative_matrix = true;
+                break;
+            }
+        }
+
+        if (!cooperative_matrix) {
+            return vk_device_architecture::NVIDIA_PRE_TURING;
+        }
     }
     return vk_device_architecture::OTHER;
 }
 
+enum vk_conv_shapes {
+    CONV_SHAPE_128x128,
+    CONV_SHAPE_64x32,
+    CONV_SHAPE_32x256,
+    CONV_SHAPE_COUNT,
+};
+
 struct vk_device_struct {
     std::recursive_mutex mutex;
 
@@ -483,8 +507,8 @@ struct vk_device_struct {
     vk_pipeline pipeline_rwkv_wkv6_f32;
     vk_pipeline pipeline_rwkv_wkv7_f32;
     vk_pipeline pipeline_opt_step_adamw_f32;
-    vk_pipeline pipeline_conv2d_f32;
-    vk_pipeline pipeline_conv2d_f16_f32;
+    vk_pipeline pipeline_conv2d_f32[CONV_SHAPE_COUNT];
+    vk_pipeline pipeline_conv2d_f16_f32[CONV_SHAPE_COUNT];
     vk_pipeline pipeline_conv2d_dw_whcn_f32;
     vk_pipeline pipeline_conv2d_dw_cwhn_f32;
 
@@ -908,8 +932,22 @@ struct vk_op_conv2d_push_constants {
     uint32_t nb1;
     uint32_t nb2;
     uint32_t nb3;
+
+    // init_fastdiv_values constants for dividing by KW, KW*KH, OW, OW*OH
+    uint32_t KWmp;   uint32_t KWL;
+    uint32_t KWKHmp; uint32_t KWKHL;
+    uint32_t OWmp;   uint32_t OWL;
+    uint32_t OWOHmp; uint32_t OWOHL;
 };
 
+template <> void init_pushconst_fastdiv(vk_op_conv2d_push_constants &p) {
+    // Compute magic values to divide by KW, KW*KH, OW, OW*OH
+    init_fastdiv_values(p.KW,       p.KWmp,    p.KWL);
+    init_fastdiv_values(p.KW*p.KH,  p.KWKHmp,  p.KWKHL);
+    init_fastdiv_values(p.OW,       p.OWmp,    p.OWL);
+    init_fastdiv_values(p.OW*p.OH,  p.OWOHmp,  p.OWOHL);
+}
+
 struct vk_op_conv2d_dw_push_constants {
     uint32_t ne;
     uint32_t batches;
@@ -2068,12 +2106,12 @@ static void ggml_vk_load_shaders(vk_device& device) {
         s_mmq_wg_denoms = { 32,  64,  1 };
 
         // spec constants and tile sizes for quant matmul (Qi_K)
-        l_warptile_mmq_k = { 256, 64, 128, 64,  1 };
-        m_warptile_mmq_k = { 256, 32,  64, 64,  0 };
-        s_warptile_mmq_k = { 256, 32,  32, 128, 0 };
-        l_mmq_wg_denoms_k = { 64, 128, 1 };
-        m_mmq_wg_denoms_k = { 32,  64, 1 };
-        s_mmq_wg_denoms_k = { 32,  32, 1 };
+        l_warptile_mmq_k = { 256, 128, 256, 64, 1 };
+        m_warptile_mmq_k = { 256, 128, 128, 64, 1 };
+        s_warptile_mmq_k = { 256, 32,  64, 128, 0 };
+        l_mmq_wg_denoms_k = { 128, 256, 1 };
+        m_mmq_wg_denoms_k = { 128, 128, 1 };
+        s_mmq_wg_denoms_k = { 32,  64,  1 };
 
         // spec constants and tile sizes for quant matmul_id
         l_warptile_mmqid = { 256, 128, 128, 16, 0 };
@@ -2847,7 +2885,7 @@ static void ggml_vk_load_shaders(vk_device& device) {
             ggml_vk_create_pipeline(device, device->pipeline_mul_mat_vec_p021_f16_f32[i], "mul_mat_vec_p021_f16_f32"+std::to_string(i+1), mul_mat_vec_p021_f16_f32_len,              mul_mat_vec_p021_f16_f32_data,              "main", 3, 6 * sizeof(uint32_t), {1, 1, 1}, {device->subgroup_size, i + 1}, 1, true);
         }
     }
-    ggml_vk_create_pipeline(device, device->pipeline_mul_mat_vec_nc_f16_f32, "mul_mat_vec_nc_f16_f32", mul_mat_vec_nc_f16_f32_len, mul_mat_vec_nc_f16_f32_data, "main", 3, 9 * sizeof(uint32_t), {1, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_mul_mat_vec_nc_f16_f32, "mul_mat_vec_nc_f16_f32", mul_mat_vec_nc_f16_f32_len, mul_mat_vec_nc_f16_f32_data, "main", 3, 12 * sizeof(uint32_t), {1, 1, 1}, {}, 1);
 
     ggml_vk_create_pipeline(device, device->pipeline_norm_f32, "norm_f32", norm_f32_len, norm_f32_data, "main", 2, sizeof(vk_op_push_constants), {1, 1, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_group_norm_f32, "group_norm_f32", group_norm_f32_len, group_norm_f32_data, "main", 2, sizeof(vk_op_push_constants), {1, 1, 1}, {}, 1);
@@ -3048,48 +3086,89 @@ static void ggml_vk_load_shaders(vk_device& device) {
     ggml_vk_create_pipeline(device, device->pipeline_opt_step_adamw_f32, "opt_step_adamw_f32", opt_step_adamw_f32_len, opt_step_adamw_f32_data, "main", 5, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1);
 
     // conv2d
-    uint32_t conv2d_WG_SIZE  = 256;
-    uint32_t conv2d_BS_K     = 128;
-    uint32_t conv2d_BS_CRS   = 16;
-    uint32_t use_collectives = 0;  // Enables subgroup ops for preventing the re-calculation of indices.
-    if (device->subgroup_shuffle &&
-        device->vendor_id != VK_VENDOR_ID_INTEL) {  // Do not enable collectives on Intel, see PR 14316
-        use_collectives = 1;
-        conv2d_BS_CRS   = std::min(
-            device->subgroup_size,
-            conv2d_BS_CRS);  // CRS block size should be capped at sugroup size for correctness when shuffle is used.
-    }
-    uint32_t conv2d_BS_NPQ = 128;
-    uint32_t conv2d_TS_K   = 8;
-    uint32_t conv2d_shmem_req =
-        (conv2d_BS_K * (conv2d_BS_CRS + 1) + conv2d_BS_CRS * (conv2d_BS_NPQ + 1)) * sizeof(float);
-    if (device->properties.limits.maxComputeSharedMemorySize < conv2d_shmem_req) {
-        conv2d_BS_CRS = 8;
-        if (use_collectives) {
-            conv2d_BS_CRS = std::min(device->subgroup_size, conv2d_BS_CRS);
-        }
-    }
+    for (uint32_t s = 0; s < CONV_SHAPE_COUNT; ++s) {
+        uint32_t conv2d_WG_SIZE  = 256;
+        uint32_t conv2d_BS_K     = 128;
+        uint32_t conv2d_BS_CRS   = 16;
+        uint32_t use_collectives = 0;  // Enables subgroup ops for preventing the re-calculation of indices.
+        uint32_t conv2d_BS_NPQ = 128;
+        uint32_t conv2d_TS_K   = 8;
+        uint32_t conv2d_SHMEM_PAD = 4;
+        bool conv2d_UNROLL = true;
 
-    if (use_collectives) {
-        ggml_vk_create_pipeline(
-            device, device->pipeline_conv2d_f32, "conv2d_f32", conv2d_f32_len, conv2d_f32_data, "main", 3,
-            sizeof(vk_op_conv2d_push_constants), { conv2d_BS_K, conv2d_BS_NPQ, 1 },
-            { conv2d_WG_SIZE, conv2d_BS_K, conv2d_BS_CRS, conv2d_BS_NPQ, conv2d_TS_K, use_collectives }, 1, true, true);
-        ggml_vk_create_pipeline(
-            device, device->pipeline_conv2d_f16_f32, "conv2d_f16_f32", conv2d_f16_f32_len, conv2d_f16_f32_data, "main", 3,
-            sizeof(vk_op_conv2d_push_constants), { conv2d_BS_K, conv2d_BS_NPQ, 1 },
-            { conv2d_WG_SIZE, conv2d_BS_K, conv2d_BS_CRS, conv2d_BS_NPQ, conv2d_TS_K, use_collectives }, 1, true, true);
-    } else {
-        ggml_vk_create_pipeline(
-            device, device->pipeline_conv2d_f32, "conv2d_f32", conv2d_f32_len, conv2d_f32_data, "main", 3,
-            sizeof(vk_op_conv2d_push_constants), { conv2d_BS_K, conv2d_BS_NPQ, 1 },
-            { conv2d_WG_SIZE, conv2d_BS_K, conv2d_BS_CRS, conv2d_BS_NPQ, conv2d_TS_K, use_collectives }, 1, true,
-            false);
-        ggml_vk_create_pipeline(
-            device, device->pipeline_conv2d_f16_f32, "conv2d_f16_f32", conv2d_f16_f32_len, conv2d_f16_f32_data, "main", 3,
-            sizeof(vk_op_conv2d_push_constants), { conv2d_BS_K, conv2d_BS_NPQ, 1 },
-            { conv2d_WG_SIZE, conv2d_BS_K, conv2d_BS_CRS, conv2d_BS_NPQ, conv2d_TS_K, use_collectives }, 1, true,
-            false);
+        if (device->vendor_id == VK_VENDOR_ID_INTEL) {
+            conv2d_SHMEM_PAD = 0;
+            conv2d_UNROLL = false;
+        } else if (device->vendor_id == VK_VENDOR_ID_AMD) {
+            conv2d_SHMEM_PAD = device->architecture == vk_device_architecture::AMD_GCN ? 1 : 4;
+        }
+
+        switch (s) {
+        default:
+        case CONV_SHAPE_128x128:
+            conv2d_BS_K = 128;
+            conv2d_BS_NPQ = 128;
+            conv2d_BS_CRS = 16;
+            if (device->vendor_id == VK_VENDOR_ID_AMD && device->architecture != vk_device_architecture::AMD_GCN) {
+                conv2d_UNROLL = false;
+            }
+            break;
+        case CONV_SHAPE_64x32:
+            conv2d_BS_K = 64;
+            conv2d_BS_NPQ = 32;
+            conv2d_BS_CRS = 32;
+            conv2d_TS_K   = 4;
+            break;
+        case CONV_SHAPE_32x256:
+            conv2d_BS_K = 32;
+            conv2d_BS_NPQ = 256;
+            conv2d_BS_CRS = 16;
+            break;
+        }
+
+        // Use collectives on pre-Turing NVIDIA GPUs and GCN AMD cards, which had slower integer math.
+        bool allow_collectives_nv = device->vendor_id != VK_VENDOR_ID_NVIDIA ||
+                                    device->architecture == vk_device_architecture::NVIDIA_PRE_TURING;
+        bool allow_collectives_amd = device->vendor_id != VK_VENDOR_ID_AMD ||
+                                     device->architecture == vk_device_architecture::AMD_GCN;
+
+        if (device->subgroup_shuffle &&
+            device->vendor_id != VK_VENDOR_ID_INTEL &&   // Do not enable collectives on Intel, see PR 14316.
+            allow_collectives_nv &&
+            allow_collectives_amd) {
+            use_collectives = 1;
+            conv2d_BS_CRS   = std::min(
+                device->subgroup_size,
+                conv2d_BS_CRS);  // CRS block size should be capped at subgroup size for correctness when shuffle is used.
+        }
+
+        uint32_t conv2d_shmem_req =
+            (conv2d_BS_K * (conv2d_BS_CRS + conv2d_SHMEM_PAD) + conv2d_BS_CRS * (conv2d_BS_NPQ + conv2d_SHMEM_PAD)) * sizeof(float);
+        if (device->properties.limits.maxComputeSharedMemorySize < conv2d_shmem_req) {
+            conv2d_BS_CRS = 8;
+            if (use_collectives) {
+                conv2d_BS_CRS = std::min(device->subgroup_size, conv2d_BS_CRS);
+            }
+        }
+
+        std::array<uint32_t, 3> wg_denoms = { conv2d_BS_K, conv2d_BS_NPQ, 1 };
+        std::vector<uint32_t> spec_constants = { conv2d_WG_SIZE, conv2d_BS_K, conv2d_BS_CRS, conv2d_BS_NPQ, conv2d_TS_K, use_collectives, conv2d_SHMEM_PAD };
+
+        if (conv2d_UNROLL) {
+            ggml_vk_create_pipeline(
+                device, device->pipeline_conv2d_f32[s], "conv2d_f32", conv2d_f32_unroll_len, conv2d_f32_unroll_data, "main", 3,
+                sizeof(vk_op_conv2d_push_constants), wg_denoms, spec_constants, 1, true, use_collectives);
+            ggml_vk_create_pipeline(
+                device, device->pipeline_conv2d_f16_f32[s], "conv2d_f16_f32", conv2d_f16_f32_unroll_len, conv2d_f16_f32_unroll_data, "main", 3,
+                sizeof(vk_op_conv2d_push_constants), wg_denoms, spec_constants, 1, true, use_collectives);
+        } else {
+            ggml_vk_create_pipeline(
+                device, device->pipeline_conv2d_f32[s], "conv2d_f32", conv2d_f32_len, conv2d_f32_data, "main", 3,
+                sizeof(vk_op_conv2d_push_constants), wg_denoms, spec_constants, 1, true, use_collectives);
+            ggml_vk_create_pipeline(
+                device, device->pipeline_conv2d_f16_f32[s], "conv2d_f16_f32", conv2d_f16_f32_len, conv2d_f16_f32_data, "main", 3,
+                sizeof(vk_op_conv2d_push_constants), wg_denoms, spec_constants, 1, true, use_collectives);
+        }
     }
 
     ggml_vk_create_pipeline(device, device->pipeline_conv2d_dw_whcn_f32, "conv2d_dw_whcn_f32", conv2d_dw_whcn_f32_len, conv2d_dw_whcn_f32_data, "main", 3, sizeof(vk_op_conv2d_dw_push_constants), {512, 1, 1}, {}, 1);
@@ -4943,26 +5022,37 @@ static void ggml_vk_buffer_memset(vk_buffer& dst, size_t offset, uint32_t c, siz
     ggml_vk_queue_command_pools_cleanup(dst->device);
 }
 
-static uint32_t ggml_vk_guess_split_k(ggml_backend_vk_context * ctx, int m, int n, int k, const vk_pipeline& pipeline) {
+static uint32_t ggml_vk_guess_split_k(ggml_backend_vk_context * ctx, uint32_t m, uint32_t n, uint32_t k, const vk_pipeline& pipeline) {
     VK_LOG_DEBUG("ggml_vk_guess_split_k(" << m << ", " << n << ", " << k << ")");
 
     uint32_t split_k = 1;
-    if (ctx->device->shader_core_count != 0 && m >= (int)pipeline->wg_denoms[0] && n >= (int)pipeline->wg_denoms[1]) {
+    if (ctx->device->shader_core_count != 0 && m >= pipeline->wg_denoms[0] && n >= pipeline->wg_denoms[1]) {
         // If k is 'large' and the SMs will fill less than halfway, use split_k.
         uint32_t m_tiles = CEIL_DIV(m, pipeline->wg_denoms[0]);
         uint32_t n_tiles = CEIL_DIV(n, pipeline->wg_denoms[1]);
-        if (k >= 2048 && m_tiles * n_tiles < ctx->device->shader_core_count / 2) {
-            split_k = ctx->device->shader_core_count / (m_tiles * n_tiles);
-            // Clamp to 2 or 4
-            split_k = std::min(split_k, 4u);
-            if (split_k == 3) {
-                split_k = 2;
+
+        if (k >= 2048) {
+            if (m_tiles * n_tiles <= ctx->device->shader_core_count / 2) {
+                split_k = ctx->device->shader_core_count / (m_tiles * n_tiles);
+            } else if (m_tiles * n_tiles <= ctx->device->shader_core_count * 2 / 3) {
+                split_k = 3;
             }
-            if (ctx->device->coopmat2) {
-                // coopmat2 shader expects splits to be aligned to 256
-                while (split_k > 1 && ((k / split_k) % 256) != 0) {
-                    split_k /= 2;
+            // Cap the split at 8x. Unless k is huge this is a lot of overhead.
+            split_k = std::min(split_k, 8u);
+
+            // ggml_vk_matmul will align the splits to be a multiple of 256.
+            // If this rounded up size would cause the last split to be empty,
+            // then reduce the split count.
+            while (true) {
+                if (split_k == 1) {
+                    break;
                 }
+                uint32_t k_split = CEIL_DIV(k, split_k);
+                k_split = ROUNDUP_POW2(k_split, 256);
+                if (k_split * (split_k - 1) < k) {
+                    break;
+                }
+                split_k--;
             }
         }
     }
@@ -4974,9 +5064,22 @@ static vk_pipeline ggml_vk_guess_matmul_pipeline(ggml_backend_vk_context * ctx,
     VK_LOG_DEBUG("ggml_vk_guess_matmul_pipeline(" << m << ", " << n << ", " << aligned << ", " << ggml_type_name(src0_type) << ", " << ggml_type_name(src1_type) << ")");
 
     if (ctx->device->coopmat2) {
+        const uint32_t shader_core_count = ctx->device->shader_core_count;
+        const uint32_t tiles_l = CEIL_DIV(m, mmp->a_l->wg_denoms[0]) * CEIL_DIV(n, mmp->a_l->wg_denoms[1]);
+        const uint32_t tiles_m = CEIL_DIV(m, mmp->a_m->wg_denoms[0]) * CEIL_DIV(n, mmp->a_m->wg_denoms[1]);
+
         // Use large shader when the N dimension is greater than the medium shader's tile size
         uint32_t crossover_large = mmp->m->wg_denoms[1];
-        if ((ctx->device->mul_mat_l[src0_type] && (n > crossover_large)) || (!ctx->device->mul_mat_m[src0_type] && !ctx->device->mul_mat_s[src0_type])) {
+
+        // Prefer large over medium if either:
+        // - medium or large tiles would overfill the GPU
+        // - large tiles with a split_k==3 fits in the GPU and medium tiles with split_k==2 does not
+        //   (medium with split_k==2 is probably better if it fits - more workgroups running and less split_k overhead)
+        bool prefer_large = tiles_m > shader_core_count || tiles_l > shader_core_count ||
+                            // split_k==3 with large tiles likely better than medium tiles with no split_k.
+                            (tiles_l <= shader_core_count / 3 && tiles_m > shader_core_count / 2);
+
+        if ((ctx->device->mul_mat_l[src0_type] && (n > crossover_large && prefer_large)) || (!ctx->device->mul_mat_m[src0_type] && !ctx->device->mul_mat_s[src0_type])) {
             return aligned ? mmp->a_l : mmp->l;
         }
         // Use medium shader when the N dimension is greater than the small shader's tile size
@@ -5020,7 +5123,11 @@ static void ggml_vk_matmul(
 
     GGML_ASSERT(batch_stride_d == m * n);
 
-    const vk_mat_mat_push_constants pc1 = { m, n, k, stride_a, stride_b, stride_d, batch_stride_a, batch_stride_b, batch_stride_d, CEIL_DIV(k, split_k), ne02, ne12, broadcast2, broadcast3, padded_n };
+    // Round the split size up to a multiple of 256 (k-quant alignment)
+    uint32_t k_split = CEIL_DIV(k, split_k);
+    k_split = ROUNDUP_POW2(k_split, 256);
+
+    const vk_mat_mat_push_constants pc1 = { m, n, k, stride_a, stride_b, stride_d, batch_stride_a, batch_stride_b, batch_stride_d, k_split, ne02, ne12, broadcast2, broadcast3, padded_n };
     // Make sure enough workgroups get assigned for split k to work
     ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { a, b, split_k_buffer }, pc1, { (CEIL_DIV(m, pipeline->wg_denoms[0]) * pipeline->wg_denoms[0]) * split_k, n, batch });
     ggml_vk_sync_buffers(subctx);
@@ -5742,7 +5849,7 @@ static void ggml_vk_mul_mat_vec_nc_f16_f32(ggml_backend_vk_context * ctx, vk_con
     const uint64_t ne00 = src0->ne[0];
     const uint64_t ne01 = src0->ne[1];
     const uint64_t ne02 = src0->ne[2];
-    // const uint64_t ne03 = src0->ne[3];
+    const uint64_t ne03 = src0->ne[3];
 
     const uint64_t nb01 = src0->nb[1];
     const uint64_t nb02 = src0->nb[2];
@@ -5754,7 +5861,12 @@ static void ggml_vk_mul_mat_vec_nc_f16_f32(ggml_backend_vk_context * ctx, vk_con
     const uint64_t ne12 = src1->ne[2];
     // const uint64_t ne13 = src1->ne[3];
 
+    const uint32_t nb03 = (uint32_t)(src0->nb[3] / sizeof(ggml_fp16_t));
+    const uint32_t nb13 = (uint32_t)(src1->nb[3] / sizeof(float));
+    const uint32_t nb23 = (uint32_t)(dst->nb[3] / sizeof(float));
+
     GGML_ASSERT(ne11 == 1);
+    GGML_ASSERT(src0->ne[3] == src1->ne[3]); // checked in supports_op
 
     ggml_backend_vk_buffer_context * dst_buf_ctx = (ggml_backend_vk_buffer_context *)dst->buffer->context;
     ggml_backend_vk_buffer_context * src0_buf_ctx = (ggml_backend_vk_buffer_context *)src0->buffer->context;
@@ -5770,7 +5882,7 @@ static void ggml_vk_mul_mat_vec_nc_f16_f32(ggml_backend_vk_context * ctx, vk_con
         src1_uma = d_Qy != nullptr;
     }
 
-    const uint64_t d_ne = ne01 * ne11 * ne12;
+    const uint64_t d_ne = ne01 * ne11 * ne12 * ne03;
 
     const uint32_t row_stride_x = nb01 / sizeof(ggml_fp16_t);
     const uint32_t channel_stride_x = nb02 / sizeof(ggml_fp16_t);
@@ -5805,10 +5917,10 @@ static void ggml_vk_mul_mat_vec_nc_f16_f32(ggml_backend_vk_context * ctx, vk_con
     const uint64_t d_shader_offset = d_buf_offset - d_buffer_offset;
 
     // compute
-    const std::array<uint32_t, 9> pc = { (uint32_t)ne00, (uint32_t)ne01, row_stride_x, channel_stride_x, channel_stride_y, (uint32_t)(ne12 / ne02), (uint32_t)ne12, (uint32_t)(qy_shader_offset / ggml_type_size(src1->type)), (uint32_t)(d_shader_offset / ggml_type_size(dst->type)) };
+    const std::array<uint32_t, 12> pc = { (uint32_t)ne00, (uint32_t)ne01, row_stride_x, channel_stride_x, channel_stride_y, (uint32_t)(ne12 / ne02), (uint32_t)ne12, (uint32_t)(qy_shader_offset / ggml_type_size(src1->type)), (uint32_t)(d_shader_offset / ggml_type_size(dst->type)), nb03, nb13, nb23 };
     ggml_vk_sync_buffers(subctx);
     ggml_vk_dispatch_pipeline(ctx, subctx, ctx->device->pipeline_mul_mat_vec_nc_f16_f32,
-        { vk_subbuffer{ d_Qx, qx_buf_offset, qx_sz }, vk_subbuffer{ d_Qy, qy_buffer_offset, qy_sz + qy_shader_offset }, vk_subbuffer{ d_D, d_buffer_offset, d_sz + d_shader_offset } }, pc, { 1, (uint32_t)ne01, (uint32_t)ne12 });
+        { vk_subbuffer{ d_Qx, qx_buf_offset, qx_sz }, vk_subbuffer{ d_Qy, qy_buffer_offset, qy_sz + qy_shader_offset }, vk_subbuffer{ d_D, d_buffer_offset, d_sz + d_shader_offset } }, pc, { (uint32_t)ne03, (uint32_t)ne01, (uint32_t)ne12 });
 }
 
 static void ggml_vk_mul_mat(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) {
@@ -6641,6 +6753,34 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
     }
 }
 
+static std::array<uint32_t, 3> ggml_vk_get_conv_elements(const ggml_tensor *dst) {
+    const ggml_tensor *src0 = dst->src[0];
+    const ggml_tensor *src1 = dst->src[1];
+
+    // src0 - kernel:   [KW, KH, Cin, Cout]
+    // src1 - input:    [W, H, Cin, N]
+    // dst - result:    [OW, OH, Cout, N]
+
+    // Copied from ggml.c: int64_t ggml_calc_conv_output_size(int64_t ins, int64_t ks, int s, int p, int d)
+    auto calc_conv_output_size = [](int64_t ins, int64_t ks, int s, int p, int d) -> int64_t {
+        return (ins + 2 * p - d * (ks - 1) - 1) / s + 1;
+    };
+    // parallelize in {OW/BS_K, OH/BS_NPQ, 1}
+    int64_t W    = src1->ne[0];
+    int64_t H    = src1->ne[1];
+    int64_t KW   = src0->ne[0];
+    int64_t KH   = src0->ne[1];
+    int64_t Cout = src0->ne[3];
+    int64_t N    = src1->ne[3];
+    int64_t OH   = calc_conv_output_size(H, KH, dst->op_params[1], dst->op_params[3], dst->op_params[5]);
+    int64_t OW   = calc_conv_output_size(W, KW, dst->op_params[0], dst->op_params[2], dst->op_params[4]);
+    int64_t NPQ  = N * OW * OH;
+
+    // Tile output matrix to (K/NB_K, NPQ/NB_NPQ, 1) workgroups
+    std::array<uint32_t, 3> elements = { static_cast<uint32_t>(Cout), static_cast<uint32_t>(NPQ), 1 };
+    return elements;
+}
+
 static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, ggml_tensor * dst, ggml_op op) {
     switch (op) {
     case GGML_OP_GET_ROWS:
@@ -6970,10 +7110,30 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
     case GGML_OP_CONV_2D:
         if (src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32 &&
             ggml_is_contiguous(src0) && ggml_is_contiguous(src1) && ggml_is_contiguous(dst)) {
+            auto elements = ggml_vk_get_conv_elements(dst);
+            vk_conv_shapes shape;
+
+            uint32_t tiles[CONV_SHAPE_COUNT];
+            for (uint32_t i = 0; i < CONV_SHAPE_COUNT; ++i) {
+                tiles[i] = CEIL_DIV(elements[0], ctx->device->pipeline_conv2d_f32[i]->wg_denoms[0]) * CEIL_DIV(elements[1], ctx->device->pipeline_conv2d_f32[i]->wg_denoms[1]);
+            }
+
+            // We can't query number of shader cores on Intel, use 32 as a placeholder
+            // so small convolutions will still choose a smaller tile.
+            const uint32_t shader_core_count = ctx->device->shader_core_count > 0 ? ctx->device->shader_core_count : 32;
+
+            if (elements[0] > 64 && tiles[CONV_SHAPE_128x128] >= shader_core_count * 2) {
+                shape = CONV_SHAPE_128x128;
+            } else if (elements[0] <= 32 && tiles[CONV_SHAPE_32x256] >= shader_core_count * 2) {
+                shape = CONV_SHAPE_32x256;
+            } else {
+                shape = CONV_SHAPE_64x32;
+            }
+
             if (src0->type == GGML_TYPE_F32) {
-                return ctx->device->pipeline_conv2d_f32;
+                return ctx->device->pipeline_conv2d_f32[shape];
             } else if (src0->type == GGML_TYPE_F16) {
-                return ctx->device->pipeline_conv2d_f16_f32;
+                return ctx->device->pipeline_conv2d_f16_f32[shape];
             }
         }
         return nullptr;
@@ -7301,29 +7461,8 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
         } break;
     case GGML_OP_CONV_2D:
         {
-            // src0 - kernel:   [KW, KH, Cin, Cout]
-            // src1 - input:    [W, H, Cin, N]
-            // dst - result:    [OW, OH, Cout, N]
-
-            // Copied from ggml.c: int64_t ggml_calc_conv_output_size(int64_t ins, int64_t ks, int s, int p, int d)
-            auto calc_conv_output_size = [](int64_t ins, int64_t ks, int s, int p, int d) -> int64_t {
-                return (ins + 2 * p - d * (ks - 1) - 1) / s + 1;
-            };
-            // parallelize in {OW/BS_K, OH/BS_NPQ, 1}
-            int64_t W    = src1->ne[0];
-            int64_t H    = src1->ne[1];
-            int64_t KW   = src0->ne[0];
-            int64_t KH   = src0->ne[1];
-            int64_t Cout = src0->ne[3];
-            int64_t N    = src1->ne[3];
-            int64_t OH   = calc_conv_output_size(H, KH, dst->op_params[1], dst->op_params[3], dst->op_params[5]);
-            int64_t OW   = calc_conv_output_size(W, KW, dst->op_params[0], dst->op_params[2], dst->op_params[4]);
-            int64_t NPQ  = N * OW * OH;
-
-            // Tile output matrix to (K/NB_K, NPQ/NB_NPQ, 1) workgroups
-            elements = { static_cast<uint32_t>(Cout), static_cast<uint32_t>(NPQ), 1 };
-        }
-        break;
+            elements = ggml_vk_get_conv_elements(dst);
+        } break;
     case GGML_OP_ADD:
     case GGML_OP_SUB:
     case GGML_OP_DIV:

ggml/src/ggml-vulkan/vulkan-shaders/conv2d_mm.comp

@@ -1,14 +1,13 @@
 #version 450
 
+#extension GL_EXT_control_flow_attributes : enable
+
 #ifdef USE_COLLECTIVES
 #    extension GL_KHR_shader_subgroup_shuffle : enable
 #endif
 
 #include "types.comp"
 
-// Make spec constant
-#define SHMEM_PAD 0
-
 // shape notation: [dim(N), ..., dim(0)] -- stride(dim(j)) >= stride(dim(i)) if i > j
 layout(binding = 0) readonly buffer A {
     A_TYPE knl_data[];
@@ -56,6 +55,12 @@ layout(push_constant) uniform parameter {
     uint32_t nb1;
     uint32_t nb2;
     uint32_t nb3;
+
+    // fastdiv helper values
+    uint32_t KWmp;   uint32_t KWL;
+    uint32_t KWKHmp; uint32_t KWKHL;
+    uint32_t OWmp;   uint32_t OWL;
+    uint32_t OWOHmp; uint32_t OWOHL;
 }
 
 p;
@@ -68,6 +73,7 @@ layout(constant_id = 3) const uint BS_NPQ          = 128;
 // Thread-tile sizes
 layout(constant_id = 4) const uint TS_K            = 8;
 layout(constant_id = 5) const uint use_collectives = 1;
+layout(constant_id = 6) const uint SHMEM_PAD       = 4;
 
 uint32_t       tid     = gl_LocalInvocationID.x;
 const uint32_t WG_SIZE = gl_WorkGroupSize.x;
@@ -131,6 +137,14 @@ uint32_t       Br    = tid / BS_NPQ;
 uint32_t       Bc    = tid % BS_NPQ;
 const uint32_t BrpWg = WG_SIZE / BS_NPQ;
 
+// see init_fastdiv_values in ggml-vulkan.cpp
+uint fastdiv(uint n, uint mp, uint L) {
+    uint msbs, lsbs;
+    // msbs = mulhi(n, mp)
+    umulExtended(n, mp, msbs, lsbs);
+    return (msbs + n) >> L;
+}
+
 void main() {
     for (uint32_t T_ly = 0; T_ly < TS_K; T_ly++) {
         for (uint32_t T_lx = 0; T_lx < TS_NPQ; T_lx++) {
@@ -151,9 +165,9 @@ void main() {
         uint32_t cached_KW_idx;
         if (use_collectives == 1) {
             cached_CRS_idx                = B_idx_CRS * BS_CRS + gl_SubgroupInvocationID;
-            cached_Cin_idx                = cached_CRS_idx / (p.KW * p.KH);
+            cached_Cin_idx                = fastdiv(cached_CRS_idx, p.KWKHmp, p.KWKHL); // divide by (p.KW * p.KH);
             uint32_t cached_CRS_remainder = (cached_CRS_idx - cached_Cin_idx * p.KW * p.KH);
-            cached_KH_idx                 = cached_CRS_remainder / p.KW;
+            cached_KH_idx                 = fastdiv(cached_CRS_remainder, p.KWmp, p.KWL); // divide by p.KW;
             cached_KW_idx                 = cached_CRS_remainder - cached_KH_idx * p.KW;
 
             CRS_idx_a = subgroupShuffle(cached_CRS_idx, Ac);
@@ -162,16 +176,16 @@ void main() {
             KW_idx_a  = subgroupShuffle(cached_KW_idx, Ac);
         } else {
             CRS_idx_a              = B_idx_CRS * BS_CRS + Ac;  // Global CRS_idx_a (column index of A)
-            Cin_idx_a              = CRS_idx_a / (p.KW * p.KH);
+            Cin_idx_a              = fastdiv(CRS_idx_a, p.KWKHmp, p.KWKHL); // divide by (p.KW * p.KH);
             uint32_t CRS_remainder = CRS_idx_a - Cin_idx_a * p.KW * p.KH;
-            KH_idx_a               = CRS_remainder / p.KW;
+            KH_idx_a               = fastdiv(CRS_remainder, p.KWmp, p.KWL); // divide by p.KW;
             KW_idx_a               = CRS_remainder - KH_idx_a * p.KW;
         }
 #else
         CRS_idx_a     = B_idx_CRS * BS_CRS + Ac;  // Global CRS_idx_a (column index of A)
-        Cin_idx_a     = CRS_idx_a / (p.KW * p.KH);
+        Cin_idx_a     = fastdiv(CRS_idx_a, p.KWKHmp, p.KWKHL); // divide by (p.KW * p.KH); / (p.KW * p.KH);
         CRS_remainder = CRS_idx_a - Cin_idx_a * p.KW * p.KH;
-        KH_idx_a      = CRS_remainder / p.KW;
+        KH_idx_a      = fastdiv(CRS_remainder, p.KWmp, p.KWL); // divide by p.KW;
         KW_idx_a      = CRS_remainder - KH_idx_a * p.KW;
 #endif
 
@@ -188,13 +202,13 @@ void main() {
             Ash[B_ly * Ash_stride + B_lx] = val;
         }
         /* Load input to B_block: (BS_CRS x BS_NPQ) */
-        for (uint32_t r_offset = 0; r_offset < BS_CRS; r_offset += BrpWg) {
+        UNROLL for (uint32_t r_offset = 0; r_offset < BS_CRS; r_offset += BrpWg) {
             uint32_t B_ly          = r_offset + Br;             /* Row index of B block */
             uint32_t B_lx          = Bc;
             uint32_t NPQ_idx       = B_idx_NPQ * BS_NPQ + B_lx; /* Global NPQ index (column index of B) */
-            uint32_t N_idx         = NPQ_idx / (p.OH * p.OW);
+            uint32_t N_idx         = fastdiv(NPQ_idx, p.OWOHmp, p.OWOHL); // divide by p.OH * p.OW;
             uint32_t NPQ_remainder = NPQ_idx - N_idx * p.OH * p.OW;
-            uint32_t OH_idx        = NPQ_remainder / p.OW;
+            uint32_t OH_idx        = fastdiv(NPQ_remainder, p.OWmp, p.OWL); // divide by p.OW;
             uint32_t OW_idx        = NPQ_remainder - OH_idx * p.OW;
 
             uint32_t CRS_idx_b;
@@ -209,16 +223,16 @@ void main() {
                 KW_idx_b  = subgroupShuffle(cached_KW_idx, r_offset + Br);
             } else {
                 CRS_idx_b              = B_idx_CRS * BS_CRS + B_ly; /* Global CRS index (row index of B) */
-                Cin_idx_b              = CRS_idx_b / (p.KW * p.KH);
+                Cin_idx_b              = fastdiv(CRS_idx_b, p.KWKHmp, p.KWKHL); // divide by (p.KW * p.KH);
                 uint32_t CRS_remainder = CRS_idx_b - Cin_idx_b * p.KW * p.KH;
-                KH_idx_b               = CRS_remainder / p.KW;
+                KH_idx_b               = fastdiv(CRS_remainder, p.KWmp, p.KWL); // divide by p.KW;
                 KW_idx_b               = CRS_remainder - KH_idx_b * p.KW;
             }
 #else
             CRS_idx_b              = B_idx_CRS * BS_CRS + B_ly; /* Global CRS index (row index of B) */
-            Cin_idx_b              = CRS_idx_b / (p.KW * p.KH);
+            Cin_idx_b              = fastdiv(CRS_idx_b, p.KWKHmp, p.KWKHL); // divide by (p.KW * p.KH);
             uint32_t CRS_remainder = CRS_idx_b - Cin_idx_b * p.KW * p.KH;
-            KH_idx_b               = CRS_remainder / p.KW;
+            KH_idx_b               = fastdiv(CRS_remainder, p.KWmp, p.KWL); // divide by p.KW;
             KW_idx_b               = CRS_remainder - KH_idx_b * p.KW;
 #endif
 
@@ -233,32 +247,36 @@ void main() {
             Bsh[B_ly * Bsh_stride + B_lx] = val;
         }
         barrier();
-        for (uint32_t CRS_lidx = 0; CRS_lidx < BS_CRS; CRS_lidx++) {
-            for (uint32_t T_ly = 0; T_ly < TS_K; T_ly++) {
-                regA[T_ly] = Ash[(T_y * TS_K + T_ly) * Ash_stride + CRS_lidx];
-            }
-            for (uint32_t T_lx = 0; T_lx < TS_NPQ; T_lx++) {
-                regB[T_lx] = Bsh[CRS_lidx * Bsh_stride + T_x * TS_NPQ + T_lx];
-            }
-            for (uint32_t T_ly = 0; T_ly < TS_K; T_ly++) {
+        if (T_y * TS_K < K) {
+            UNROLL for (uint32_t CRS_lidx = 0; CRS_lidx < BS_CRS; CRS_lidx++) {
+                for (uint32_t T_ly = 0; T_ly < TS_K; T_ly++) {
+                    regA[T_ly] = Ash[(T_y * TS_K + T_ly) * Ash_stride + CRS_lidx];
+                }
                 for (uint32_t T_lx = 0; T_lx < TS_NPQ; T_lx++) {
-                    regC[T_ly][T_lx] = fma(regA[T_ly], regB[T_lx], regC[T_ly][T_lx]);
+                    regB[T_lx] = Bsh[CRS_lidx * Bsh_stride + T_x * TS_NPQ + T_lx];
+                }
+                for (uint32_t T_ly = 0; T_ly < TS_K; T_ly++) {
+                    for (uint32_t T_lx = 0; T_lx < TS_NPQ; T_lx++) {
+                        regC[T_ly][T_lx] = fma(regA[T_ly], regB[T_lx], regC[T_ly][T_lx]);
+                    }
                 }
             }
         }
         barrier();
     }
     /* Save C* */
-    for (uint32_t T_ly = 0; T_ly < TS_K; T_ly++) {
-        for (uint32_t T_lx = 0; T_lx < TS_NPQ; T_lx++) {
-            uint32_t K_idx   = B_idx_K * BS_K + T_y * TS_K + T_ly;
-            uint32_t NPQ_idx = B_idx_NPQ * BS_NPQ + T_x * TS_NPQ + T_lx;
-            uint32_t N_idx   = NPQ_idx / (p.OH * p.OW);
-            uint32_t OH_idx  = (NPQ_idx - N_idx * p.OH * p.OW) / p.OW;
-            uint32_t OW_idx  = NPQ_idx - N_idx * p.OH * p.OW - OH_idx * p.OW;
-            uint32_t dst_idx = OW_idx + OH_idx * p.nb1 + K_idx * p.nb2 + N_idx * p.nb3;
-            if (K_idx < K && NPQ_idx < NPQ) {
-                dst_data[dst_idx] = regC[T_ly][T_lx];
+    if (T_y * TS_K < K) {
+        for (uint32_t T_ly = 0; T_ly < TS_K; T_ly++) {
+            for (uint32_t T_lx = 0; T_lx < TS_NPQ; T_lx++) {
+                uint32_t K_idx   = B_idx_K * BS_K + T_y * TS_K + T_ly;
+                uint32_t NPQ_idx = B_idx_NPQ * BS_NPQ + T_x * TS_NPQ + T_lx;
+                uint32_t N_idx   = fastdiv(NPQ_idx, p.OWOHmp, p.OWOHL); // divide by p.OH * p.OW;
+                uint32_t OH_idx  = fastdiv(NPQ_idx - N_idx * p.OH * p.OW, p.OWmp, p.OWL); // divide by p.OW;
+                uint32_t OW_idx  = NPQ_idx - N_idx * p.OH * p.OW - OH_idx * p.OW;
+                uint32_t dst_idx = OW_idx + OH_idx * p.nb1 + K_idx * p.nb2 + N_idx * p.nb3;
+                if (K_idx < K && NPQ_idx < NPQ) {
+                    dst_data[dst_idx] = regC[T_ly][T_lx];
+                }
             }
         }
     }

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

@@ -26,6 +26,9 @@ layout (push_constant) uniform parameter
     uint ne12;
     uint b_offset;
     uint d_offset;
+    uint nb03;
+    uint nb13;
+    uint nb23;
 } p;
 
 shared FLOAT_TYPE tmp[BLOCK_SIZE];
@@ -34,6 +37,7 @@ void main() {
     const uint tid       = gl_LocalInvocationID.x;
     const uint row_x     = gl_GlobalInvocationID.y;
     const uint channel   = gl_GlobalInvocationID.z;
+    const uint i3        = gl_WorkGroupID.x;
     const uint channel_x = channel / p.channel_x_divisor;
     const uint channel_y = channel % p.ne12;
 
@@ -41,7 +45,7 @@ void main() {
     const uint nrows_dst = p.nrows_x;
     const uint row_dst   = row_x;
 
-    const uint idst = channel*nrows_dst + row_dst;
+    const uint idst = i3*p.nb23 + channel*nrows_dst + row_dst;
 
     FLOAT_TYPE temp = 0.0f;
 
@@ -58,8 +62,8 @@ void main() {
 
                 const uint row_y = col_x;
 
-                const uint ix = channel_x*p.channel_stride_x + row_x*p.row_stride_x + col_x;
-                const uint iy = channel_y*p.channel_stride_y + row_y;
+                const uint ix = i3*p.nb03 + channel_x*p.channel_stride_x + row_x*p.row_stride_x + col_x;
+                const uint iy = i3*p.nb13 + channel_y*p.channel_stride_y + row_y;
 
                 const vec4 av4 = vec4(data_a_v4[ix / 4]);
                 const vec4 bv4 = vec4(data_b_v4[iy / 4]);
@@ -74,8 +78,8 @@ void main() {
 
             const uint row_y = col_x;
 
-            const uint ix = channel_x*p.channel_stride_x + row_x*p.row_stride_x + col_x;
-            const uint iy = channel_y*p.channel_stride_y + row_y;
+            const uint ix = i3*p.nb03 + channel_x*p.channel_stride_x + row_x*p.row_stride_x + col_x;
+            const uint iy = i3*p.nb13 + channel_y*p.channel_stride_y + row_y;
 
             const vec4 av4 = vec4(data_a_v4[ix / 4]);
             const vec4 bv4 = vec4(data_b_v4[iy / 4]);
@@ -91,8 +95,8 @@ void main() {
 
             const uint row_y = col_x;
 
-            const uint ix = channel_x*p.channel_stride_x + row_x*p.row_stride_x + col_x;
-            const uint iy = channel_y*p.channel_stride_y + row_y;
+            const uint ix = i3*p.nb03 + channel_x*p.channel_stride_x + row_x*p.row_stride_x + col_x;
+            const uint iy = i3*p.nb13 + channel_y*p.channel_stride_y + row_y;
 
             const FLOAT_TYPE xi = FLOAT_TYPE(data_a[ix]);
 

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

@@ -655,8 +655,11 @@ void process_shaders() {
 
     string_to_spv("opt_step_adamw_f32", "opt_step_adamw.comp", merge_maps(base_dict, {{"A_TYPE", "float"}}));
 
-    string_to_spv("conv2d_f32", "conv2d_mm.comp", {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"USE_COLLECTIVES", "1"}});
-    string_to_spv("conv2d_f16_f32", "conv2d_mm.comp", {{"A_TYPE", "float16_t"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"USE_COLLECTIVES", "1"}});
+    string_to_spv("conv2d_f32_unroll", "conv2d_mm.comp", {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"USE_COLLECTIVES", "1"}, {"UNROLL", "[[unroll]]"}});
+    string_to_spv("conv2d_f16_f32_unroll", "conv2d_mm.comp", {{"A_TYPE", "float16_t"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"USE_COLLECTIVES", "1"}, {"UNROLL", "[[unroll]]"}});
+
+    string_to_spv("conv2d_f32", "conv2d_mm.comp", {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"USE_COLLECTIVES", "1"}, {"UNROLL", ""}});
+    string_to_spv("conv2d_f16_f32", "conv2d_mm.comp", {{"A_TYPE", "float16_t"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"USE_COLLECTIVES", "1"}, {"UNROLL", ""}});
 
     string_to_spv("conv2d_dw_whcn_f32", "conv2d_dw.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"WHCN", "1"}}));
     string_to_spv("conv2d_dw_cwhn_f32", "conv2d_dw.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"CWHN", "1"}}));

gguf-py/gguf/tensor_mapping.py

@@ -33,6 +33,7 @@ class TensorNameMap:
             "language_model.model.embed_tokens",         # llama4
             "encoder",                                   # neobert
             "model.transformer.wte",                     # llada
+            "embed_tokens",                              # qwen3-embedding
         ),
 
         # Token type embeddings
@@ -143,6 +144,7 @@ class TensorNameMap:
             "transformer_encoder.{bid}.attention_norm",             # neobert
             "model.layers.{bid}.operator_norm",                     # lfm2
             "model.transformer.blocks.{bid}.attn_norm",             # llada
+            "layers.{bid}.input_layernorm",                         # qwen3-embedding
         ),
 
         # Attention norm 2
@@ -188,6 +190,7 @@ class TensorNameMap:
             "transformer.h.{bid}.attn.attention.q_proj",                 # exaone
             "model.layers.{bid}.self_attn.q_proj",                       # llama4
             "model.transformer.blocks.{bid}.q_proj",                     # llada
+            "layers.{bid}.self_attn.q_proj",                             # qwen3-embedding
         ),
 
         # Attention key
@@ -205,6 +208,7 @@ class TensorNameMap:
             "transformer.h.{bid}.attn.attention.k_proj",               # exaone
             "model.layers.{bid}.self_attn.k_proj",                     # llama4
             "model.transformer.blocks.{bid}.k_proj",                   # llada
+            "layers.{bid}.self_attn.k_proj",                           # qwen3-embedding
         ),
 
         # Attention value
@@ -221,6 +225,7 @@ class TensorNameMap:
             "transformer.h.{bid}.attn.attention.v_proj",                 # exaone
             "model.layers.{bid}.self_attn.v_proj",                       # llama4
             "model.transformer.blocks.{bid}.v_proj",                     # llada
+            "layers.{bid}.self_attn.v_proj",                             # qwen3-embedding
         ),
 
         # Attention output
@@ -254,6 +259,7 @@ class TensorNameMap:
             "model.layers.{bid}.self_attn.o_proj",                          # llama4
             "transformer_encoder.{bid}.wo",                                 # neobert
             "model.transformer.blocks.{bid}.attn_out",                      # llada
+            "layers.{bid}.self_attn.o_proj",                                # qwen3-embedding
         ),
 
         # Attention output norm
@@ -300,6 +306,7 @@ class TensorNameMap:
             "transformer_encoder.{bid}.ffn_norm",                            # neobert
             "model.layers.layers.{bid}.pre_mlp_norm",                        # plamo2
             "model.transformer.blocks.{bid}.ff_norm",                        # llada
+            "layers.{bid}.post_attention_layernorm",                         # qwen3-embedding
         ),
 
         # Post feed-forward norm
@@ -373,7 +380,8 @@ class TensorNameMap:
             "model.layers.{bid}.feed_forward.up_proj",                # llama4 jamba granite-hybrid
             "transformer_encoder.{bid}.ffn.w12",                      # neobert
             "model.layers.{bid}.block_sparse_moe.up",                 # smallthinker
-            "model.transformer.blocks.{bid}.up_proj",                  # llada
+            "model.transformer.blocks.{bid}.up_proj",                 # llada
+            "layers.{bid}.mlp.up_proj",                               # qwen3-embedding
         ),
 
         MODEL_TENSOR.FFN_UP_EXP: (
@@ -416,6 +424,7 @@ class TensorNameMap:
             "model.layers.{bid}.feed_forward.gate_proj",  # llama4 jamba granite-hybrid
             "model.layers.{bid}.block_sparse_moe.gate",   # smallthinker
             "model.transformer.blocks.{bid}.ff_proj",     # llada
+            "layers.{bid}.mlp.gate_proj",                 # qwen3-embedding
         ),
 
         MODEL_TENSOR.FFN_GATE_EXP: (
@@ -465,7 +474,8 @@ class TensorNameMap:
             "model.layers.{bid}.feed_forward.down_proj",              # llama4 jamba granite-hybrid
             "transformer_encoder.{bid}.ffn.w3",                       # neobert
             "model.layers.{bid}.block_sparse_moe.down",               # smallthinker
-            "model.transformer.blocks.{bid}.ff_out",                   # llada
+            "model.transformer.blocks.{bid}.ff_out",                  # llada
+            "layers.{bid}.mlp.down_proj",                             # qwen3-embedding
         ),
 
         MODEL_TENSOR.FFN_DOWN_EXP: (
@@ -497,6 +507,7 @@ class TensorNameMap:
             "encoder.layer.{bid}.attention.self.layer_norm_q",                # jina-bert-v2
             "transformer.layers.{bid}.attn.q_norm",                           # openelm
             "model.layers.layers.{bid}.mixer.q",                              # plamo2
+            "layers.{bid}.self_attn.q_norm",                                  # qwen3-embedding
         ),
 
         MODEL_TENSOR.ATTN_K_NORM: (
@@ -508,6 +519,7 @@ class TensorNameMap:
             "encoder.layer.{bid}.attention.self.layer_norm_k",                # jina-bert-v2
             "transformer.layers.{bid}.attn.k_norm",                           # openelm
             "model.layers.layers.{bid}.mixer.k",                              # plamo2
+            "layers.{bid}.self_attn.k_norm",                                  # qwen3-embedding
         ),
 
         MODEL_TENSOR.ROPE_FREQS: (

requirements/requirements-convert_hf_to_gguf_update.txt

@@ -1,7 +1 @@
 -r ./requirements-convert_legacy_llama.txt
---extra-index-url https://download.pytorch.org/whl/cpu
-torch~=2.2.1; platform_machine != "s390x"
-
-# torch s390x packages can only be found from nightly builds
---extra-index-url https://download.pytorch.org/whl/nightly
-torch>=0.0.0.dev0; platform_machine == "s390x"

scripts/compare-llama-bench.py

@@ -326,7 +326,7 @@ class LlamaBenchDataSQLite3(LlamaBenchData):
 
         # Set table name and schema based on tool
         if self.tool == "llama-bench":
-            self.table_name = "test"
+            self.table_name = "llama_bench"
             db_fields = LLAMA_BENCH_DB_FIELDS
             db_types = LLAMA_BENCH_DB_TYPES
         elif self.tool == "test-backend-ops":
@@ -409,8 +409,8 @@ class LlamaBenchDataSQLite3File(LlamaBenchDataSQLite3):
 
         # Tool selection logic
         if tool is None:
-            if "test" in table_names:
-                self.table_name = "test"
+            if "llama_bench" in table_names:
+                self.table_name = "llama_bench"
                 self.tool = "llama-bench"
             elif "test_backend_ops" in table_names:
                 self.table_name = "test_backend_ops"
@@ -418,8 +418,8 @@ class LlamaBenchDataSQLite3File(LlamaBenchDataSQLite3):
             else:
                 raise RuntimeError(f"No suitable table found in database. Available tables: {table_names}")
         elif tool == "llama-bench":
-            if "test" in table_names:
-                self.table_name = "test"
+            if "llama_bench" in table_names:
+                self.table_name = "llama_bench"
                 self.tool = "llama-bench"
             else:
                 raise RuntimeError(f"Table 'test' not found for tool 'llama-bench'. Available tables: {table_names}")

src/llama-context.cpp

@@ -105,7 +105,7 @@ llama_context::llama_context(
 
     {
         const char * LLAMA_SET_ROWS = getenv("LLAMA_SET_ROWS");
-        supports_set_rows = LLAMA_SET_ROWS ? (atoi(LLAMA_SET_ROWS) != 0) : false;
+        supports_set_rows = LLAMA_SET_ROWS ? (atoi(LLAMA_SET_ROWS) != 0) : supports_set_rows;
 
         if (!supports_set_rows && !cparams.kv_unified) {
             LLAMA_LOG_WARN("%s: non-unified KV cache requires ggml_set_rows() - forcing unified KV cache\n", __func__);

src/llama-context.h

@@ -289,7 +289,7 @@ private:
 
     // env: LLAMA_SET_ROWS (temporary)
     // ref: https://github.com/ggml-org/llama.cpp/pull/14285
-    bool supports_set_rows = false;
+    bool supports_set_rows = true;
 
     // env: LLAMA_GRAPH_REUSE_DISABLE
     bool graph_reuse_disable = false;

src/llama-kv-cache-unified.cpp

@@ -183,7 +183,7 @@ llama_kv_cache_unified::llama_kv_cache_unified(
         const size_t memory_size_k = size_k_bytes();
         const size_t memory_size_v = size_v_bytes();
 
-        LLAMA_LOG_INFO("%s: size = %7.2f MiB (%6u cells, %3d layers, %2u/%2u seqs), K (%s): %7.2f MiB, V (%s): %7.2f MiB\n", __func__,
+        LLAMA_LOG_INFO("%s: size = %7.2f MiB (%6u cells, %3d layers, %2u/%u seqs), K (%s): %7.2f MiB, V (%s): %7.2f MiB\n", __func__,
                 (float)(memory_size_k + memory_size_v) / (1024.0f * 1024.0f), kv_size, (int) layers.size(), n_seq_max, n_stream,
                 ggml_type_name(type_k), (float)memory_size_k / (1024.0f * 1024.0f),
                 ggml_type_name(type_v), (float)memory_size_v / (1024.0f * 1024.0f));
@@ -193,7 +193,7 @@ llama_kv_cache_unified::llama_kv_cache_unified(
     debug = LLAMA_KV_CACHE_DEBUG ? atoi(LLAMA_KV_CACHE_DEBUG) : 0;
 
     const char * LLAMA_SET_ROWS = getenv("LLAMA_SET_ROWS");
-    supports_set_rows = LLAMA_SET_ROWS ? atoi(LLAMA_SET_ROWS) != 0 : 0;
+    supports_set_rows = LLAMA_SET_ROWS ? atoi(LLAMA_SET_ROWS) != 0 : supports_set_rows;
 
     if (!supports_set_rows) {
         // ref: https://github.com/ggml-org/llama.cpp/pull/14363

src/llama-kv-cache-unified.h

@@ -230,7 +230,7 @@ private:
 
     // env: LLAMA_SET_ROWS (temporary)
     // ref: https://github.com/ggml-org/llama.cpp/pull/14285
-    bool supports_set_rows = false;
+    bool supports_set_rows = true;
 
     const llama_swa_type swa_type = LLAMA_SWA_TYPE_NONE;
 

src/llama-model.cpp

@@ -899,6 +899,7 @@ void llama_model::load_hparams(llama_model_loader & ml) {
             } break;
         case LLM_ARCH_QWEN3:
             {
+                ml.get_key(LLM_KV_POOLING_TYPE, hparams.pooling_type, false);
                 ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
                 switch (hparams.n_layer) {
                     case 28: type = hparams.n_embd == 1024 ? LLM_TYPE_0_6B : LLM_TYPE_1_7B; break;

tests/test-backend-ops.cpp

@@ -5592,13 +5592,15 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
     test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F16, GGML_TYPE_F32, 1056, 1, 193, {1,  1}, {4, 1}, {0, 2, 1, 3}));
     test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F16, GGML_TYPE_F32, 1056, 1, 67,  {1,  1}, {4, 1}, {0, 2, 1, 3}));
 
-    for (auto bs : {1,2,4,8}) {
-        for (auto nr : {1,4}) {
-            for (uint32_t m = 0; m < 2; ++m) {
-                for (uint32_t k = 0; k < 2; ++k) {
-                    for (ggml_type type: {GGML_TYPE_F16, GGML_TYPE_BF16, GGML_TYPE_F32}) {
-                        test_cases.emplace_back(new test_mul_mat(type, GGML_TYPE_F32, 1056 + m, 1, 128 + k,  {bs,  1}, {nr, 1}, {0, 2, 1, 3}));
-                        test_cases.emplace_back(new test_mul_mat(type, GGML_TYPE_F32, 128 + m,  1, 1056 + k, {bs,  1}, {nr, 1}, {0, 1, 2, 3}, true));
+    for (auto bs2 : {1,3}) {
+        for (auto bs : {1,2,4,8}) {
+            for (auto nr : {1,4}) {
+                for (uint32_t m = 0; m < 2; ++m) {
+                    for (uint32_t k = 0; k < 2; ++k) {
+                        for (ggml_type type: {GGML_TYPE_F16, GGML_TYPE_BF16, GGML_TYPE_F32}) {
+                            test_cases.emplace_back(new test_mul_mat(type, GGML_TYPE_F32, 1056 + m, 1, 128 + k,  {bs,  bs2}, {nr, 1}, {0, 2, 1, 3}));
+                            test_cases.emplace_back(new test_mul_mat(type, GGML_TYPE_F32, 128 + m,  1, 1056 + k, {bs,  bs2}, {nr, 1}, {0, 1, 2, 3}, true));
+                        }
                     }
                 }
             }

tests/test-chat.cpp

@@ -953,6 +953,33 @@ static void test_template_output_parsers() {
                 /* is_partial= */ false,
                 {COMMON_CHAT_FORMAT_HERMES_2_PRO}));
 
+        // Test multiple tool calls
+        common_chat_msg message_assist_multiple_calls;
+        message_assist_multiple_calls.role = "assistant";
+        message_assist_multiple_calls.content = "";
+        message_assist_multiple_calls.tool_calls.push_back({"special_function", "{\"arg1\": 1}", ""});
+        message_assist_multiple_calls.tool_calls.push_back({"python", "{\"code\":\"print('hello')\"}", ""});
+
+        assert_msg_equals(
+            message_assist_multiple_calls,
+            common_chat_parse(
+                "<tool_call>\n"
+                "{\"name\": \"special_function\", \"arguments\": {\"arg1\": 1}}\n"
+                "</tool_call>\n"
+                "<tool_call>\n"
+                "{\"name\": \"python\", \"arguments\": {\"code\":\"print('hello')\"}}\n"
+                "</tool_call>",
+                /* is_partial= */ false,
+                {COMMON_CHAT_FORMAT_HERMES_2_PRO}));
+
+        assert_msg_equals(
+            message_assist_multiple_calls,
+            common_chat_parse(
+                "<function=special_function>{\"arg1\": 1}</function>\n"
+                "<function=python>{\"code\":\"print('hello')\"}</function>",
+                /* is_partial= */ false,
+                {COMMON_CHAT_FORMAT_HERMES_2_PRO}));
+
         assert_msg_equals(
             simple_assist_msg(
                 "This is not a tool call:",
@@ -1039,6 +1066,22 @@ static void test_template_output_parsers() {
                       "<tool_call>\n"
                       "{\"name\": \"special_function\", \"arguments\": {\"arg1\": 1}}\n"
                       "</tool_call>");
+
+        // Test multiple tool calls with template
+        common_chat_msg message_assist_multiple_calls_template;
+        message_assist_multiple_calls_template.role = "assistant";
+        message_assist_multiple_calls_template.content = "";
+        message_assist_multiple_calls_template.tool_calls.push_back({"special_function", "{\"arg1\": 1}", ""});
+        message_assist_multiple_calls_template.tool_calls.push_back({"python", "{\"code\":\"print('test')\"}", ""});
+
+        test_templates(tmpls.get(), end_tokens, message_assist_multiple_calls_template, tools,
+                      "<tool_call>\n"
+                      "{\"name\": \"special_function\", \"arguments\": {\"arg1\": 1}}\n"
+                      "</tool_call>\n"
+                      "<tool_call>\n"
+                      "{\"name\": \"python\", \"arguments\": {\"code\":\"print('test')\"}}\n"
+                      "</tool_call>");
+
         test_templates(tmpls.get(), end_tokens, message_assist_call_python_lines, tools,
                       "<tool_call>\n"
                       "{\"name\": \"python\", \"arguments\": {\"code\":\"# This is a program:\\nprint('hey')\"}}\n"

tools/llama-bench/llama-bench.cpp

@@ -1738,7 +1738,7 @@ struct sql_printer : public printer {
 
     void print_header(const cmd_params & params) override {
         std::vector<std::string> fields = test::get_fields();
-        fprintf(fout, "CREATE TABLE IF NOT EXISTS test (\n");
+        fprintf(fout, "CREATE TABLE IF NOT EXISTS llama_bench (\n");
         for (size_t i = 0; i < fields.size(); i++) {
             fprintf(fout, "  %s %s%s\n", fields.at(i).c_str(), get_sql_field_type(fields.at(i)).c_str(),
                     i < fields.size() - 1 ? "," : "");
@@ -1749,7 +1749,7 @@ struct sql_printer : public printer {
     }
 
     void print_test(const test & t) override {
-        fprintf(fout, "INSERT INTO test (%s) ", join(test::get_fields(), ", ").c_str());
+        fprintf(fout, "INSERT INTO llama_bench (%s) ", join(test::get_fields(), ", ").c_str());
         fprintf(fout, "VALUES (");
         std::vector<std::string> values = t.get_values();
         for (size_t i = 0; i < values.size(); i++) {

tools/server/server.cpp

@@ -4249,9 +4249,6 @@ int main(int argc, char ** argv) {
 
             // process prompt
             std::vector<server_tokens> inputs;
-            if (oaicompat && !prompt.is_string()) {
-                throw std::runtime_error("prompt must be a string");
-            }
 
             if (oaicompat && has_mtmd) {
                 // multimodal