ggml-hip : add -fno-finite-math-only alongside -ffast-math (#25373)

-ffast-math implies -ffinite-math-only under ROCm/clang 22, which
disables INFINITY/NaN and triggers -Wnan-infinity-disabled (errors
under -Werror in CI). Re-enable infinity handling without dropping
the rest of fast-math.

Fixes #25361

common/common.cpp

@@ -55,6 +55,10 @@
 #include <pwd.h>
 #endif
 
+#if defined(_AIX)
+#include <sys/systemcfg.h>
+#endif
+
 #if defined(_MSC_VER)
 #pragma warning(disable: 4244 4267) // possible loss of data
 #endif
@@ -72,7 +76,16 @@ common_time_meas::~common_time_meas() {
 //
 
 int32_t common_cpu_get_num_physical_cores() {
-#ifdef __linux__
+#if defined(_AIX)
+    int32_t logical_cpus = _system_configuration.ncpus;
+    int32_t smt_threads = _system_configuration.smt_threads;
+    if (smt_threads > 0) {
+        return static_cast<int32_t>(logical_cpus / smt_threads);
+    }
+    if (logical_cpus > 0) {
+        return static_cast<int32_t>(logical_cpus);
+    }
+#elif defined(__linux__)
     // enumerate the set of thread siblings, num entries is num cores
     std::unordered_set<std::string> siblings;
     for (uint32_t cpu=0; cpu < UINT32_MAX; ++cpu) {
@@ -202,6 +215,14 @@ int32_t common_cpu_get_num_math() {
             }
         }
     }
+#elif defined(__powerpc64__) || defined(__powerpc__)
+    int32_t smt_factor = 1;
+    int phy_cpus = common_cpu_get_num_physical_cores();
+    int logical_cpus = sysconf(_SC_NPROCESSORS_ONLN);
+    if (phy_cpus > 0 && logical_cpus > phy_cpus) {
+        smt_factor = logical_cpus / phy_cpus;
+    }
+    return phy_cpus * std::min(smt_factor, 2);
 #endif
     return common_cpu_get_num_physical_cores();
 }

common/ngram-map.cpp

@@ -125,6 +125,16 @@ void common_ngram_map_begin(
     LOG_DBG("%s: begin, idx_last_draft=%zu, new begin=%zu, #keys=%zu\n", __func__,
             map.idx_last_check, size_begin, map.keys.size());
 
+    size_t idx_begin_cleanup = map.size_last_begin;
+    if (idx_begin_cleanup > size_begin) {
+        if (size_begin > (size_t) map.size_key + map.size_value) {
+            idx_begin_cleanup = size_begin - map.size_key - map.size_value;
+        } else {
+            idx_begin_cleanup = 0;
+        }
+        LOG_INF("%s: shrink cleanup begin: %zu -> %zu\n", __func__, map.size_last_begin, idx_begin_cleanup);
+    }
+
     size_t count_map_entries_upd = 0;
     if (!map.key_map.empty() && size_begin < map.idx_last_check) {
         if (map.show_key_map_stats) {
@@ -150,27 +160,23 @@ void common_ngram_map_begin(
         // Update the map from hash to key index (clear outdated entries).
         for (size_t i = 0; i < map.key_map.size(); ++i) {
             uint32_t key_idx = map.key_map[i];
-            if (key_idx >= map.size_last_begin) {
+            if (key_idx != 0 && key_idx >= idx_begin_cleanup) {
                 map.key_map[i] = 0;
                 count_map_entries_upd++;
             }
         }
-        map.key_map_last_idx = (map.size_last_begin > 0) ? map.size_last_begin - 1 : 0;
+        map.key_map_last_idx = (idx_begin_cleanup > 0) ? (uint32_t) (idx_begin_cleanup - 1) : 0;
     }
 
     if (size_begin < map.idx_last_check && !map.keys.empty()) {
-        // The next token generation will start at index size_begin.
-        // The tokens between map.size_last_begin and size_begin are no longer valid.
-        //
-        // Refresh map: Remove all entries with index >= map.size_last_begin.
         size_t count_keys = map.keys.size();
         size_t count_keys_del = 0;
         size_t count_values_del = 0;
         for (int32_t i = map.keys.size() - 1; i >= 0; --i) {
             common_ngram_map_key & key = map.keys[i];
-            if (key.key_idx >= map.size_last_begin) {
+            if (key.key_idx >= idx_begin_cleanup) {
                 // Delete the key.
-                LOG_DBG("%s: delete key %d at index %zu (>= size_last_begin=%zu)\n", __func__, i, key.key_idx, map.size_last_begin);
+                LOG_DBG("%s: delete key %d at index %zu (>= idx_begin_cleanup=%zu)\n", __func__, i, key.key_idx, idx_begin_cleanup);
                 map.keys.erase(map.keys.begin() + i);
                 count_keys_del++;
                 continue;
@@ -182,7 +188,7 @@ void common_ngram_map_begin(
             // Check the indices of the values.
             for (int16_t j = COMMON_NGRAM_MAX_VALUES - 1; j >= 0; --j) {
                 common_ngram_map_value & value = key.values[j];
-                if (value.value_idx >= map.size_last_begin) {
+                if (value.value_idx != 0 && value.value_idx >= idx_begin_cleanup) {
                     // Delete the value.
                     count_values_del++;
 

docs/backend/SYCL.md

@@ -790,10 +790,10 @@ use 1 SYCL GPUs: [0] with Max compute units:512
 | GGML_SYCL_DEBUG   | 0 (default) or 1 | Enable log function by macro: GGML_SYCL_DEBUG                                                                             |
 | GGML_SYCL_DEV2DEV_MEMCPY | 0 (default) or 1 | Choose the SYCL or L0 API in dev2dev memory copy.<br>Value: <br>*  0: SYCL API (default)<br>* 1: L0 API -- L0 API is found to lead to abnormal crash in some case. This debug flag is used to check the issue.|
 | GGML_SYCL_ENABLE_FLASH_ATTN | 1 (default) or 0| Enable Flash-Attention. It can reduce memory usage. The performance impact depends on the LLM.|
-| GGML_SYCL_DISABLE_OPT | 0 (default) or 1 | Disable optimize features for Intel GPUs. (Recommended to 1 for Intel devices older than Gen 10) |
-| GGML_SYCL_DISABLE_GRAPH | 0 or 1 (default) | Disable running computations through SYCL Graphs feature. Disabled by default because SYCL Graph is still on development, no better performance. |
+| GGML_SYCL_ENABLE_OPT | 0 or 1 (default)| Enable optimize features for Intel GPUs. (Recommended to 0 for Intel devices older than Gen 10) |
+| GGML_SYCL_ENABLE_GRAPH | 0 (default) or 1 | Enable running computations through SYCL Graphs feature. Disabled by default because SYCL Graph is still on development, no better performance. |
 | GGML_SYCL_USE_LEVEL_ZERO_API | 1 (default) or 0 | Use Level Zero API for device memory allocation instead of SYCL. Reduces system RAM usage on Intel dGPUs by avoiding DMA-buf/TTM host memory staging. Requires GGML_SYCL_SUPPORT_LEVEL_ZERO_API=ON at build time. SYCL backend always runs on Level Zero running time even if it's set as OFF (The SYCL api will be usage for memory allocation).|
-| GGML_SYCL_DISABLE_DNN | 0 (default) or 1 | Disable running computations through oneDNN and always use oneMKL. |
+| GGML_SYCL_ENABLE_DNN | 0 or 1 (default)| Enable running computations through oneDNN and always use oneMKL. |
 | GGML_SYCL_ENABLE_VMM | 0 or 1 (default) | Enable the virtual-memory device pool. |
 | ZES_ENABLE_SYSMAN | 0 (default) or 1 | Support to get free memory of GPU by sycl::aspect::ext_intel_free_memory.<br>Recommended to use when --split-mode = layer |
 | UR_L0_ENABLE_RELAXED_ALLOCATION_LIMITS | 0 (default) or 1 | Allow SYCL/Unified Runtime Level Zero device allocations larger than 4 GiB. llama.cpp's direct Level Zero allocation path requests the relaxed maximum-size limit itself when GGML_SYCL_ENABLE_LEVEL_ZERO=1. |
@@ -807,7 +807,7 @@ Pass these via `CXXFLAGS` or add a one-off `#define` to enable a flag on the spo
 |-----------------|----------------------------------------------------------------------------------|
 | DEBUG_SYCL_POOL | Enable device memory pool logging on teardown. Useful for profiling allocations. |
 | DEBUG_SYCL_MALLOC | Enable verbose per-call logging of device pool alloc/free operations. |
-
+| GGML_SYCL_SUPPORT_VMM | Support to building with VMM code. Default is Yes. |
 
 ## Design Rule
 

docs/build.md

@@ -270,13 +270,10 @@ The environment variable [`CUDA_SCALE_LAUNCH_QUEUES`](https://docs.nvidia.com/cu
 
 Consider setting `CUDA_SCALE_LAUNCH_QUEUES=4x`, which increases the CUDA command buffer to 4 times its default size. This optimization is particularly beneficial for **Multi-GPU setups with pipeline parallelism**, where it significantly improves prompt processing throughput by allowing more operations to be enqueued across GPUs.
 
-#### GGML_CUDA_FORCE_CUBLAS_COMPUTE_32F
+#### GGML_CUDA_CUBLAS_COMPUTE_TYPE
 
-Use `GGML_CUDA_FORCE_CUBLAS_COMPUTE_32F` environment variable to use FP32 compute type on all GPUs in FP16 cuBLAS for preventing possible numerical overflows in exchange for slower prompt processing (small impact on RTX PRO/Datacenter products and significant on GeForce products).
-
-#### GGML_CUDA_FORCE_CUBLAS_COMPUTE_16F
-
-Use `GGML_CUDA_FORCE_CUBLAS_COMPUTE_16F` environment variable to force use FP16 compute type (instead of default FP32) in FP16 cuBLAS for V100, CDNA and RDNA4.
+Override default, speed-optimized compute types for cuBLAS matrix multiplications.
+Legal values: `auto`, `f16`, `fp16`, `bf16`, `f32`, `fp32`.
 
 ### Unified Memory
 

docs/ops.md

@@ -21,12 +21,12 @@ Legend:
 |                           ADD_ID | ❌ | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                           ARANGE | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                           ARGMAX | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
-|                          ARGSORT | ❌ | ✅ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ✅ | ❌ | ❌ |
+|                          ARGSORT | ❌ | ✅ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                             CEIL | ❌ | ❌ | ✅ | 🟡 | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                            CLAMP | ❌ | ✅ | ✅ | ✅ | ✅ | 🟡 | ✅ | 🟡 | ✅ | ❌ | ❌ |
-|                        COL2IM_1D | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
+|                        COL2IM_1D | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ |
 |                           CONCAT | ❌ | ✅ | ✅ | 🟡 | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
-|                             CONT | ❌ | 🟡 | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | 🟡 | ❌ | ❌ |
+|                             CONT | ❌ | 🟡 | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 | ❌ | ❌ |
 |                          CONV_2D | ❌ | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                       CONV_2D_DW | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                          CONV_3D | ❌ | ❌ | ✅ | ❌ | ✅ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ |
@@ -35,8 +35,8 @@ Legend:
 |                              COS | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | 🟡 | ✅ | ❌ | ❌ |
 |                      COUNT_EQUAL | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                              CPY | ❌ | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | ❌ | ❌ |
-|               CROSS_ENTROPY_LOSS | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
-|          CROSS_ENTROPY_LOSS_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
+|               CROSS_ENTROPY_LOSS | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ |
+|          CROSS_ENTROPY_LOSS_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ |
 |                           CUMSUM | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                             DIAG | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                    DIAG_MASK_INF | ❌ | ✅ | ✅ | ✅ | ❌ | 🟡 | ✅ | ✅ | ❌ | ❌ | ❌ |
@@ -70,7 +70,7 @@ Legend:
 |                              MUL | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                          MUL_MAT | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 |
 |                 MUL_MAT_HADAMARD | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
-|                       MUL_MAT_ID | ❌ | 🟡 | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | 🟡 | 🟡 | ❌ |
+|                       MUL_MAT_ID | ❌ | 🟡 | ✅ | ✅ | 🟡 | 🟡 | ✅ | ✅ | 🟡 | 🟡 | ❌ |
 |                              NEG | ❌ | ✅ | ✅ | 🟡 | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                             NORM | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | 🟡 | ✅ | ❌ | ❌ |
 |                   OPT_STEP_ADAMW | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ |

ggml/include/ggml-cuda.h

@@ -30,9 +30,6 @@ GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_cuda_buffer_type(int de
 // conduct allreduce operation between devices
 GGML_BACKEND_API bool ggml_backend_cuda_allreduce_tensor(ggml_backend_t * backends, struct ggml_tensor ** tensors, size_t n_backends);
 
-// split tensor buffer that splits matrices by rows across multiple devices
-GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_cuda_split_buffer_type(int main_device, const float * tensor_split);
-
 // pinned host buffer for use with the CPU backend for faster copies between CPU and GPU
 GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_cuda_host_buffer_type(void);
 

ggml/src/ggml-backend-meta.cpp

@@ -1144,6 +1144,11 @@ static enum ggml_status ggml_backend_meta_buffer_init_tensor_impl(ggml_backend_m
         ggml_context          * simple_ctx = stc.ctxs[j].get();
         ggml_backend_buffer_t   simple_buf = buf_ctx->bufs[j].get();
 
+        if ((simple_buf != nullptr) && ggml_backend_buffer_is_multi_buffer(simple_buf)) {
+            // see https://github.com/ggml-org/llama.cpp/issues/22197
+            GGML_ABORT("multi buffers are not supported by the meta backend");
+        }
+
         if (split_dim >= 0 && split_dim < GGML_MAX_DIMS) {
             // TODO: the following assert fails for llama-parallel even though the results are correct:
             // GGML_ASSERT(ggml_is_contiguously_allocated(tensor));
@@ -1245,9 +1250,8 @@ static enum ggml_status ggml_backend_meta_buffer_init_tensor(ggml_backend_buffer
 
 static void ggml_backend_meta_buffer_set_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
     const size_t n_bufs = ggml_backend_meta_buffer_n_bufs(buffer);
-    GGML_ASSERT(ggml_is_contiguous(tensor));
-
     const ggml_backend_meta_split_state split_state = ggml_backend_meta_get_split_state(tensor, /*assume_sync =*/ false);
+    GGML_ASSERT(ggml_is_contiguous(tensor) || split_state.axis == GGML_BACKEND_SPLIT_AXIS_MIRRORED);
 
     if (split_state.n_segments != 1 || split_state.nr[0] != 1) {
         GGML_ASSERT(split_state.axis >= 0 && split_state.axis < GGML_MAX_DIMS);
@@ -1360,9 +1364,8 @@ static void ggml_backend_meta_buffer_set_tensor(ggml_backend_buffer_t buffer, gg
 
 static void ggml_backend_meta_buffer_get_tensor(ggml_backend_buffer_t buffer, const ggml_tensor * tensor, void * data, size_t offset, size_t size) {
     const size_t n_bufs = ggml_backend_meta_buffer_n_bufs(buffer);
-    GGML_ASSERT(ggml_is_contiguous(tensor));
-
     const ggml_backend_meta_split_state split_state = ggml_backend_meta_get_split_state(tensor, /*assume_sync =*/ false);
+    GGML_ASSERT(ggml_is_contiguous(tensor) || split_state.axis == GGML_BACKEND_SPLIT_AXIS_MIRRORED);
 
     if (split_state.n_segments != 1 || split_state.nr[0] != 1) {
         GGML_ASSERT(split_state.axis >= 0 && split_state.axis < GGML_MAX_DIMS);

ggml/src/ggml-cpu/arch/arm/quants.c

@@ -812,10 +812,10 @@ void ggml_vec_dot_nvfp4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo
         const float dy0 = GGML_CPU_FP16_TO_FP32(y[2*ib].d);
         const float dy1 = GGML_CPU_FP16_TO_FP32(y[2*ib+1].d);
         const float32x4_t nvsc = {
-            ggml_ue4m3_to_fp32(x[ib].d[0]),
-            ggml_ue4m3_to_fp32(x[ib].d[1]),
-            ggml_ue4m3_to_fp32(x[ib].d[2]),
-            ggml_ue4m3_to_fp32(x[ib].d[3])
+            GGML_CPU_UE4M3_TO_FP32(x[ib].d[0]),
+            GGML_CPU_UE4M3_TO_FP32(x[ib].d[1]),
+            GGML_CPU_UE4M3_TO_FP32(x[ib].d[2]),
+            GGML_CPU_UE4M3_TO_FP32(x[ib].d[3])
         };
         const float32x4_t scales = vmulq_f32(nvsc, (float32x4_t){dy0, dy0, dy1, dy1});
 

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

@@ -2321,24 +2321,28 @@ class tinyBLAS_Q0_PPC {
     }
 
     void matmul(int64_t m, int64_t n) {
-    #if defined(_AIX) || defined(__BIG_ENDIAN__)
-        mnpack(0, m, 0, n);
-    #else
-        const int64_t mc = 64;
-        const int64_t kc = 64;
+        int64_t mc = 64;
         int64_t nc = 64;
+        int64_t kc = 64;
+        int64_t n_chunk = 64;
+    #if defined(_AIX) || defined(__BIG_ENDIAN__)
+        mc = 32;
+        nc = 32;
+        kc = 32;
+        n_chunk = 32
+    #endif
         int64_t n_aligned = 0;
-        if (n % 64 == 0) {
+        if (n % n_chunk == 0) {
             n_aligned = n;
         } else if (n == 4) {
             n_aligned = 4;
-        } else if (n < 64) {
+        } else if (n < n_chunk) {
             n_aligned = (n / 8) * 8;
         } else {
-            n_aligned = (n / 64) * 64;
+            n_aligned = (n / n_chunk) * n_chunk;
         }
         if (n_aligned > 0) {
-            if (n_aligned % 64 == 0)      nc = 64;
+            if (n_aligned % n_chunk == 0) nc = n_chunk;
             else if (n_aligned == n)      nc = n;
             else if (n_aligned % 32 == 0) nc = 32;
             else if (n_aligned % 24 == 0) nc = 24;
@@ -2354,7 +2358,6 @@ class tinyBLAS_Q0_PPC {
         } else {
             mnpack(0, m, 0, n);
         }
-    #endif
     }
 
   private:
@@ -3195,16 +3198,19 @@ class tinyBLAS_PPC {
     }
 
     void matmul(int64_t m, int64_t n) {
+        int64_t mc = 256;
+        int64_t nc = 256;
+        int64_t kc = 256;
     #if defined(_AIX) || defined(__BIG_ENDIAN__)
-        mnpack(0, m, 0, n);
-    #else
-        int64_t mc = 256; int64_t nc = 256; int64_t kc = 256;
+        mc = 128;
+        nc = 128;
+        kc = 128;
+    #endif
         if (m % mc == 0 && n % nc == 0 && k % kc == 0) {
             matmul_tiled(m, n, mc, nc, kc);
         } else {
             mnpack(0, m, 0, n);
         }
-    #endif
     }
 
   private:

ggml/src/ggml-cpu/simd-mappings.h

@@ -131,8 +131,8 @@ extern float ggml_table_f32_ue4m3[1 << 8];
 #define GGML_CPU_E8M0_TO_FP32_HALF(x) GGML_E8M0_TO_FP32_HALF(x)
 #endif
 
-// Use lookup table for UE4M3 on x86 (faster than bit manipulation)
-#if defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__)
+// Use lookup table for UE4M3 on x86 and ARM (faster than bit manipulation)
+#if defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__) || defined(__ARM_NEON)
 #define GGML_CPU_UE4M3_TO_FP32(x) ggml_table_f32_ue4m3[(uint8_t)(x)]
 #else
 #define GGML_CPU_UE4M3_TO_FP32(x) ggml_ue4m3_to_fp32(x)

ggml/src/ggml-cuda/conv-transpose-1d.cu

@@ -11,30 +11,32 @@ static  __global__ void conv_transpose_1d_kernel(
         return;
     }
 
-    int out_index = global_index / dst_ne0;
+    int out_t = global_index % dst_ne0;
+    int out_ch = (global_index / dst_ne0) % dst_ne1;
+    int plane = global_index / (dst_ne0 * dst_ne1);
 
     float accumulator = 0;
 
     for (int c = 0; c < src0_ne2; c++) {
-        int idx = global_index % dst_ne0;
+        int kernel_offset = src0_ne0 * (out_ch + src0_ne1 * c);
+        int input_offset = src1_ne0 * (c + src1_ne1 * plane);
 
-        int kernel_offset = (src0_ne0 * src0_ne1 * c) + (out_index * src0_ne0);
-        int input_offset = src1_ne0 * c;
-
-        for (int i = 0; i < src1_ne0; i++) {
-            if (!(idx >= i*s0 && idx < i*s0 + src0_ne0)) {
+        for (int k = 0; k < src0_ne0; k++) {
+            int input_numer = out_t + p0 - k*d0;
+            if (input_numer < 0 || input_numer % s0 != 0) {
                 continue;
             }
-            int weight_idx = idx - i*s0;
 
-            float kernel_weight = src0[kernel_offset + weight_idx];
-            float input_value =  src1[input_offset+i];
+            int input_t = input_numer / s0;
+            if (input_t >= src1_ne0) {
+                continue;
+            }
 
-            accumulator += kernel_weight * input_value;
+            accumulator += src0[kernel_offset + k] * src1[input_offset + input_t];
         }
     }
     dst[global_index] = accumulator;
-    GGML_UNUSED_VARS(p0, d0, src0_ne3, src1_ne3, dst_ne3, src1_ne1, dst_ne1, src1_ne2, dst_ne2);
+    GGML_UNUSED_VARS(src0_ne3, src1_ne2, src1_ne3, dst_ne2, dst_ne3);
 }
 
 static void conv_transpose_1d_f32_f32_cuda(

ggml/src/ggml-cuda/convert.cu

@@ -104,8 +104,8 @@ static __global__ void dequantize_block_q4_0(const void * __restrict__ vx, dst_t
     const uint8_t * q = x->qs + 4*il;
 
     for (int l = 0; l < 4; ++l) {
-        y[l+ 0] = d * (q[l] & 0xF) + dm;
-        y[l+16] = d * (q[l] >>  4) + dm;
+        y[l+ 0] = ggml_cuda_cast<dst_t>(d * (q[l] & 0xF) + dm);
+        y[l+16] = ggml_cuda_cast<dst_t>(d * (q[l] >>  4) + dm);
     }
 }
 
@@ -131,8 +131,8 @@ static __global__ void dequantize_block_q4_1(const void * __restrict__ vx, dst_t
     const uint8_t * q = x->qs + 4*il;
 
     for (int l = 0; l < 4; ++l) {
-        y[l+ 0] = d.x * (q[l] & 0xF) + d.y;
-        y[l+16] = d.x * (q[l] >>  4) + d.y;
+        y[l+ 0] = ggml_cuda_cast<dst_t>(d.x * (q[l] & 0xF) + d.y);
+        y[l+16] = ggml_cuda_cast<dst_t>(d.x * (q[l] >>  4) + d.y);
     }
 }
 
@@ -154,10 +154,10 @@ static __global__ void dequantize_block_q2_K(const void * __restrict__ vx, dst_t
 
     float dall = __low2half(x[i].dm);
     float dmin = __high2half(x[i].dm);
-    y[l+ 0] = dall * (x[i].scales[is+0] & 0xF) * ((q >> 0) & 3) - dmin * (x[i].scales[is+0] >> 4);
-    y[l+32] = dall * (x[i].scales[is+2] & 0xF) * ((q >> 2) & 3) - dmin * (x[i].scales[is+2] >> 4);
-    y[l+64] = dall * (x[i].scales[is+4] & 0xF) * ((q >> 4) & 3) - dmin * (x[i].scales[is+4] >> 4);
-    y[l+96] = dall * (x[i].scales[is+6] & 0xF) * ((q >> 6) & 3) - dmin * (x[i].scales[is+6] >> 4);
+    y[l+ 0] = ggml_cuda_cast<dst_t>(dall * (x[i].scales[is+0] & 0xF) * ((q >> 0) & 3) - dmin * (x[i].scales[is+0] >> 4));
+    y[l+32] = ggml_cuda_cast<dst_t>(dall * (x[i].scales[is+2] & 0xF) * ((q >> 2) & 3) - dmin * (x[i].scales[is+2] >> 4));
+    y[l+64] = ggml_cuda_cast<dst_t>(dall * (x[i].scales[is+4] & 0xF) * ((q >> 4) & 3) - dmin * (x[i].scales[is+4] >> 4));
+    y[l+96] = ggml_cuda_cast<dst_t>(dall * (x[i].scales[is+6] & 0xF) * ((q >> 6) & 3) - dmin * (x[i].scales[is+6] >> 4));
 }
 
 template<typename dst_t>
@@ -188,7 +188,9 @@ static __global__ void dequantize_block_q3_K(const void * __restrict__ vx, dst_t
     const uint8_t * q = x[i].qs + 32*n;
     const uint8_t * hm = x[i].hmask;
 
-    for (int l = l0; l < l0+4; ++l) y[l] = dl * ((int8_t)((q[l] >> shift) & 3) - ((hm[l] & m) ? 0 : 4));
+    for (int l = l0; l < l0+4; ++l) {
+        y[l] = ggml_cuda_cast<dst_t>(dl * ((int8_t)((q[l] >> shift) & 3) - ((hm[l] & m) ? 0 : 4)));
+    }
 }
 
 static inline __device__ void get_scale_min_k4(int j, const uint8_t * q, uint8_t & d, uint8_t & m) {
@@ -226,8 +228,8 @@ static __global__ void dequantize_block_q4_K(const void * __restrict__ vx, dst_t
     get_scale_min_k4(is + 1, x[i].scales, sc, m);
     const float d2 = dall * sc; const float m2 = dmin * m;
     for (int l = 0; l < n; ++l) {
-        y[l + 0] = d1 * (q[l] & 0xF) - m1;
-        y[l +32] = d2 * (q[l] >>  4) - m2;
+        y[l + 0] = ggml_cuda_cast<dst_t>(d1 * (q[l] & 0xF) - m1);
+        y[l +32] = ggml_cuda_cast<dst_t>(d2 * (q[l] >>  4) - m2);
     }
 }
 
@@ -258,11 +260,11 @@ static __global__ void dequantize_block_q5_K(const void * __restrict__ vx, dst_t
     const float d2 = dall * sc; const float m2 = dmin * m;
 
     uint8_t   hm  = 1 << (2*il);
-    y[ 0] = d1 * ((ql[ 0] & 0xF) + (qh[ 0] & hm ? 16 : 0)) - m1;
-    y[ 1] = d1 * ((ql[ 1] & 0xF) + (qh[ 1] & hm ? 16 : 0)) - m1;
+    y[ 0] = ggml_cuda_cast<dst_t>(d1 * ((ql[ 0] & 0xF) + (qh[ 0] & hm ? 16 : 0)) - m1);
+    y[ 1] = ggml_cuda_cast<dst_t>(d1 * ((ql[ 1] & 0xF) + (qh[ 1] & hm ? 16 : 0)) - m1);
     hm <<= 1;
-    y[32] = d2 * ((ql[ 0] >>  4) + (qh[ 0] & hm ? 16 : 0)) - m2;
-    y[33] = d2 * ((ql[ 1] >>  4) + (qh[ 1] & hm ? 16 : 0)) - m2;
+    y[32] = ggml_cuda_cast<dst_t>(d2 * ((ql[ 0] >>  4) + (qh[ 0] & hm ? 16 : 0)) - m2);
+    y[33] = ggml_cuda_cast<dst_t>(d2 * ((ql[ 1] >>  4) + (qh[ 1] & hm ? 16 : 0)) - m2);
 }
 
 template<typename dst_t>
@@ -285,10 +287,10 @@ static __global__ void dequantize_block_q6_K(const void * __restrict__ vx, dst_t
     const uint8_t   qh = x[i].qh[32*ip + il];
     const int8_t  * sc = x[i].scales + is;
 
-    y[ 0] = d * sc[0] * ((int8_t)((ql[ 0] & 0xF) | (((qh >> 0) & 3) << 4)) - 32);
-    y[32] = d * sc[2] * ((int8_t)((ql[32] & 0xF) | (((qh >> 2) & 3) << 4)) - 32);
-    y[64] = d * sc[4] * ((int8_t)((ql[ 0]  >> 4) | (((qh >> 4) & 3) << 4)) - 32);
-    y[96] = d * sc[6] * ((int8_t)((ql[32]  >> 4) | (((qh >> 6) & 3) << 4)) - 32);
+    y[ 0] = ggml_cuda_cast<dst_t>(d * sc[0] * ((int8_t)((ql[ 0] & 0xF) | (((qh >> 0) & 3) << 4)) - 32));
+    y[32] = ggml_cuda_cast<dst_t>(d * sc[2] * ((int8_t)((ql[32] & 0xF) | (((qh >> 2) & 3) << 4)) - 32));
+    y[64] = ggml_cuda_cast<dst_t>(d * sc[4] * ((int8_t)((ql[ 0]  >> 4) | (((qh >> 4) & 3) << 4)) - 32));
+    y[96] = ggml_cuda_cast<dst_t>(d * sc[6] * ((int8_t)((ql[32]  >> 4) | (((qh >> 6) & 3) << 4)) - 32));
 }
 
 template<typename dst_t>
@@ -307,7 +309,9 @@ static __global__ void dequantize_block_iq2_xxs(const void * __restrict__ vx, ds
     const uint32_t aux32 = q2[2] | (q2[3] << 16);
     const float d = (float)x[i].d * (0.5f + (aux32 >> 28)) * 0.25f;
     const uint8_t signs = ksigns_iq2xs[(aux32 >> 7*il) & 127];
-    for (int j = 0; j < 8; ++j) y[j] = d * grid[j] * (signs & kmask_iq2xs[j] ? -1.f : 1.f);
+    for (int j = 0; j < 8; ++j) {
+        y[j] = ggml_cuda_cast<dst_t>(d * grid[j] * (signs & kmask_iq2xs[j] ? -1.f : 1.f));
+    }
 }
 
 template<typename dst_t>
@@ -324,7 +328,9 @@ static __global__ void dequantize_block_iq2_xs(const void * __restrict__ vx, dst
     const uint8_t  * grid = (const uint8_t *)(iq2xs_grid + (q2[il] & 511));
     const float d = (float)x[i].d * (0.5f + ((x[i].scales[ib] >> 4*(il/2)) & 0xf)) * 0.25f;
     const uint8_t signs = ksigns_iq2xs[q2[il] >> 9];
-    for (int j = 0; j < 8; ++j) y[j] = d * grid[j] * (signs & kmask_iq2xs[j] ? -1.f : 1.f);
+    for (int j = 0; j < 8; ++j) {
+        y[j] = ggml_cuda_cast<dst_t>(d * grid[j] * (signs & kmask_iq2xs[j] ? -1.f : 1.f));
+    }
 }
 
 template<typename dst_t>
@@ -340,7 +346,9 @@ static __global__ void dequantize_block_iq2_s(const void * __restrict__ vx, dst_
     const uint8_t * grid = (const uint8_t *)(iq2s_grid + (x[i].qs[4*ib+il] | ((x[i].qh[ib] << (8-2*il)) & 0x300)));
     const float d = (float)x[i].d * (0.5f + ((x[i].scales[ib] >> 4*(il/2)) & 0xf)) * 0.25f;
     const uint8_t signs = x[i].qs[QK_K/8+4*ib+il];
-    for (int j = 0; j < 8; ++j) y[j] = d * grid[j] * (signs & kmask_iq2xs[j] ? -1.f : 1.f);
+    for (int j = 0; j < 8; ++j) {
+        y[j] = ggml_cuda_cast<dst_t>(d * grid[j] * (signs & kmask_iq2xs[j] ? -1.f : 1.f));
+    }
 }
 
 template<typename dst_t>
@@ -361,8 +369,8 @@ static __global__ void dequantize_block_iq3_xxs(const void * __restrict__ vx, ds
     const float d = (float)x[i].d * (0.5f + (aux32 >> 28)) * 0.5f;
     const uint8_t signs = ksigns_iq2xs[(aux32 >> 7*il) & 127];
     for (int j = 0; j < 4; ++j) {
-        y[j+0] = d * grid1[j] * (signs & kmask_iq2xs[j+0] ? -1.f : 1.f);
-        y[j+4] = d * grid2[j] * (signs & kmask_iq2xs[j+4] ? -1.f : 1.f);
+        y[j+0] = ggml_cuda_cast<dst_t>(d * grid1[j] * (signs & kmask_iq2xs[j+0] ? -1.f : 1.f));
+        y[j+4] = ggml_cuda_cast<dst_t>(d * grid2[j] * (signs & kmask_iq2xs[j+4] ? -1.f : 1.f));
     }
 }
 
@@ -382,8 +390,8 @@ static __global__ void dequantize_block_iq3_s(const void * __restrict__ vx, dst_
     const float d = (float)x[i].d * (1 + 2*((x[i].scales[ib/2] >> 4*(ib%2)) & 0xf));
     const uint8_t signs = x[i].signs[4*ib + il];
     for (int j = 0; j < 4; ++j) {
-        y[j+0] = d * grid1[j] * (signs & kmask_iq2xs[j+0] ? -1.f : 1.f);
-        y[j+4] = d * grid2[j] * (signs & kmask_iq2xs[j+4] ? -1.f : 1.f);
+        y[j+0] = ggml_cuda_cast<dst_t>(d * grid1[j] * (signs & kmask_iq2xs[j+0] ? -1.f : 1.f));
+        y[j+4] = ggml_cuda_cast<dst_t>(d * grid2[j] * (signs & kmask_iq2xs[j+4] ? -1.f : 1.f));
     }
 }
 
@@ -404,7 +412,7 @@ static __global__ void dequantize_block_iq1_s(const void * __restrict__ vx, dst_
     grid32[1] = (grid32[0] >> 4) & 0x0f0f0f0f;
     grid32[0] &= 0x0f0f0f0f;
     for (int j = 0; j < 8; ++j) {
-        y[j] = d * (q[j] + delta);
+        y[j] = ggml_cuda_cast<dst_t>(d * (q[j] + delta));
     }
 }
 
@@ -429,7 +437,7 @@ static __global__ void dequantize_block_iq1_m(const void * __restrict__ vx, dst_
     grid32[1] = (grid32[0] >> 4) & 0x0f0f0f0f;
     grid32[0] &= 0x0f0f0f0f;
     for (int j = 0; j < 8; ++j) {
-        y[j] = d * (q[j] + delta);
+        y[j] = ggml_cuda_cast<dst_t>(d * (q[j] + delta));
     }
 }
 
@@ -446,8 +454,8 @@ static __global__ void dequantize_block_iq4_nl(const void * __restrict__ vx, dst
     const uint8_t  * q4 = x[ib].qs + 4*il;
     const float d = (float)x[ib].d;
     for (int j = 0; j < 4; ++j) {
-        y[j+ 0] = d * kvalues_iq4nl[q4[j] & 0xf];
-        y[j+16] = d * kvalues_iq4nl[q4[j] >>  4];
+        y[j+ 0] = ggml_cuda_cast<dst_t>(d * kvalues_iq4nl[q4[j] & 0xf]);
+        y[j+16] = ggml_cuda_cast<dst_t>(d * kvalues_iq4nl[q4[j] >>  4]);
     }
 }
 
@@ -463,8 +471,8 @@ static __global__ void dequantize_block_iq4_xs(const void * __restrict__ vx, dst
     const uint8_t  * q4 = x[i].qs + 16*ib + 4*il;
     const float d = (float)x[i].d * ((((x[i].scales_l[ib/2] >> 4*(ib%2)) & 0xf) | (((x[i].scales_h >> 2*ib) & 3) << 4)) - 32);
     for (int j = 0; j < 4; ++j) {
-        y[j+ 0] = d * kvalues_iq4nl[q4[j] & 0xf];
-        y[j+16] = d * kvalues_iq4nl[q4[j] >>  4];
+        y[j+ 0] = ggml_cuda_cast<dst_t>(d * kvalues_iq4nl[q4[j] & 0xf]);
+        y[j+16] = ggml_cuda_cast<dst_t>(d * kvalues_iq4nl[q4[j] >>  4]);
     }
 }
 
@@ -481,8 +489,8 @@ static __global__ void dequantize_block_mxfp4(const void * __restrict__ vx, dst_
     const uint8_t  * q4 = x[ib].qs + 4*il;
     const float d = ggml_cuda_e8m0_to_fp32(x[ib].e);
     for (int j = 0; j < 4; ++j) {
-        y[j+ 0] = d * kvalues_mxfp4[q4[j] & 0xf]*0.5f;
-        y[j+16] = d * kvalues_mxfp4[q4[j] >>  4]*0.5f;
+        y[j+ 0] = ggml_cuda_cast<dst_t>(d * kvalues_mxfp4[q4[j] & 0xf]*0.5f);
+        y[j+16] = ggml_cuda_cast<dst_t>(d * kvalues_mxfp4[q4[j] >>  4]*0.5f);
     }
 }
 
@@ -700,6 +708,50 @@ static void convert_unary_cont_cuda(const void * vx, dst_t * y, const int64_t k,
 
 to_bf16_cuda_t ggml_get_to_bf16_cuda(ggml_type type) {
     switch (type) {
+        case GGML_TYPE_Q1_0:
+            return dequantize_block_cont_cuda<QK1_0, QR1_0, dequantize_q1_0>;
+        case GGML_TYPE_Q4_0:
+            return dequantize_row_q4_0_cuda;
+        case GGML_TYPE_Q4_1:
+            return dequantize_row_q4_1_cuda;
+        case GGML_TYPE_Q5_0:
+            return dequantize_block_cont_cuda<QK5_0, QR5_0, dequantize_q5_0>;
+        case GGML_TYPE_Q5_1:
+            return dequantize_block_cont_cuda<QK5_1, QR5_1, dequantize_q5_1>;
+        case GGML_TYPE_Q8_0:
+            return dequantize_block_cont_cuda<QK8_0, QR8_0, dequantize_q8_0>;
+        case GGML_TYPE_Q2_K:
+            return dequantize_row_q2_K_cuda;
+        case GGML_TYPE_Q3_K:
+            return dequantize_row_q3_K_cuda;
+        case GGML_TYPE_Q4_K:
+            return dequantize_row_q4_K_cuda;
+        case GGML_TYPE_Q5_K:
+            return dequantize_row_q5_K_cuda;
+        case GGML_TYPE_Q6_K:
+            return dequantize_row_q6_K_cuda;
+        case GGML_TYPE_IQ2_XXS:
+            return dequantize_row_iq2_xxs_cuda;
+        case GGML_TYPE_IQ2_XS:
+            return dequantize_row_iq2_xs_cuda;
+        case GGML_TYPE_IQ2_S:
+            return dequantize_row_iq2_s_cuda;
+        case GGML_TYPE_IQ3_XXS:
+            return dequantize_row_iq3_xxs_cuda;
+        case GGML_TYPE_IQ1_S:
+            return dequantize_row_iq1_s_cuda;
+        case GGML_TYPE_IQ1_M:
+            return dequantize_row_iq1_m_cuda;
+        case GGML_TYPE_IQ4_NL:
+            return dequantize_row_iq4_nl_cuda;
+        case GGML_TYPE_IQ4_XS:
+            return dequantize_row_iq4_xs_cuda;
+        case GGML_TYPE_IQ3_S:
+            return dequantize_row_iq3_s_cuda;
+        case GGML_TYPE_MXFP4:
+            return dequantize_row_mxfp4_cuda;
+        case GGML_TYPE_NVFP4:
+            return dequantize_row_nvfp4_cuda;
         case GGML_TYPE_F32:
             return convert_unary_cont_cuda<float>;
         case GGML_TYPE_F16:

ggml/src/ggml-cuda/fattn.cu

@@ -337,6 +337,26 @@ enum best_fattn_kernel {
     BEST_FATTN_KERNEL_MMA_F16  = 400,
 };
 
+static bool ggml_cuda_fattn_kv_type_supported(ggml_type type) {
+    switch (type) {
+        case GGML_TYPE_F32:
+        case GGML_TYPE_F16:
+            return true;
+        case GGML_TYPE_Q4_1:
+        case GGML_TYPE_Q5_0:
+        case GGML_TYPE_Q5_1:
+#ifndef GGML_CUDA_FA_ALL_QUANTS
+            return false;
+#endif // GGML_CUDA_FA_ALL_QUANTS
+        case GGML_TYPE_Q4_0:
+        case GGML_TYPE_Q8_0:
+        case GGML_TYPE_BF16:
+            return true;
+        default:
+            return false;
+    }
+}
+
 static best_fattn_kernel ggml_cuda_get_best_fattn_kernel(const int device, const ggml_tensor * dst) {
 #ifndef FLASH_ATTN_AVAILABLE
     GGML_UNUSED(device); GGML_UNUSED(dst);
@@ -427,22 +447,8 @@ static best_fattn_kernel ggml_cuda_get_best_fattn_kernel(const int device, const
     }
 #endif // GGML_CUDA_FA_ALL_QUANTS
 
-    switch (K->type) {
-        case GGML_TYPE_F32:
-        case GGML_TYPE_F16:
-            break;
-        case GGML_TYPE_Q4_1:
-        case GGML_TYPE_Q5_0:
-        case GGML_TYPE_Q5_1:
-#ifndef GGML_CUDA_FA_ALL_QUANTS
-            return BEST_FATTN_KERNEL_NONE;
-#endif // GGML_CUDA_FA_ALL_QUANTS
-        case GGML_TYPE_Q4_0:
-        case GGML_TYPE_Q8_0:
-        case GGML_TYPE_BF16:
-            break;
-        default:
-            return BEST_FATTN_KERNEL_NONE;
+    if (!ggml_cuda_fattn_kv_type_supported(K->type) || !ggml_cuda_fattn_kv_type_supported(V->type)) {
+        return BEST_FATTN_KERNEL_NONE;
     }
 
     if (mask && mask->ne[2] != 1) {

ggml/src/ggml-cuda/mmvq.cu

@@ -278,6 +278,9 @@ int get_mmvq_mmid_max_batch(ggml_type type, int cc) {
 }
 
 bool ggml_cuda_should_use_mmvq(enum ggml_type type, int cc, int64_t ne11) {
+    if (!ggml_is_quantized(type)) {
+        return false;
+    }
     if (GGML_CUDA_CC_IS_CDNA(cc)) {
         if (GGML_CUDA_CC_IS_CDNA1(cc)) {
             switch (type) {

ggml/src/ggml-hip/CMakeLists.txt

@@ -155,3 +155,5 @@ if (GGML_HIP_RCCL)
 endif()
 
 target_link_libraries(ggml-hip PRIVATE ggml-base hip::host roc::rocblas roc::hipblas)
+
+target_compile_options(ggml-hip PRIVATE "$<$<COMPILE_LANGUAGE:HIP>:-ffast-math;-fno-finite-math-only>")

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

@@ -1800,6 +1800,26 @@ ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_conv_transpose_1
     return res;
 }
 
+ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_col2im_1d(ggml_metal_library_t lib, const ggml_tensor * op) {
+    assert(op->op == GGML_OP_COL2IM_1D);
+
+    GGML_ASSERT(ggml_is_contiguous(op->src[0]));
+    GGML_ASSERT(op->src[0]->type == GGML_TYPE_F32 || op->src[0]->type == GGML_TYPE_F16 || op->src[0]->type == GGML_TYPE_BF16);
+
+    char base[256];
+    char name[256];
+
+    snprintf(base, 256, "kernel_col2im_1d_%s", ggml_type_name(op->src[0]->type));
+    snprintf(name, 256, "%s", base);
+
+    ggml_metal_pipeline_with_params res = ggml_metal_library_get_pipeline(lib, name);
+    if (!res.pipeline) {
+        res = ggml_metal_library_compile_pipeline(lib, base, name, nullptr);
+    }
+
+    return res;
+}
+
 ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_conv_transpose_2d(ggml_metal_library_t lib, const ggml_tensor * op) {
     assert(op->op == GGML_OP_CONV_TRANSPOSE_2D);
 

ggml/src/ggml-metal/ggml-metal-device.h

@@ -150,6 +150,7 @@ struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_rope
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_im2col            (ggml_metal_library_t lib, const struct ggml_tensor * op);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_conv_transpose_1d (ggml_metal_library_t lib, const struct ggml_tensor * op);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_conv_transpose_2d (ggml_metal_library_t lib, const struct ggml_tensor * op);
+struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_col2im_1d         (ggml_metal_library_t lib, const struct ggml_tensor * op);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_conv_2d           (ggml_metal_library_t lib, const struct ggml_tensor * op);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_conv_3d           (ggml_metal_library_t lib, const struct ggml_tensor * op);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_upscale           (ggml_metal_library_t lib, const struct ggml_tensor * op);

ggml/src/ggml-metal/ggml-metal-device.m

@@ -1157,6 +1157,11 @@ bool ggml_metal_device_supports_op(ggml_metal_device_t dev, const struct ggml_te
                 (op->src[0]->type == GGML_TYPE_F16 || op->src[0]->type == GGML_TYPE_F32) &&
                 op->src[1]->type == GGML_TYPE_F32 &&
                 op->type == GGML_TYPE_F32;
+        case GGML_OP_COL2IM_1D:
+            return (op->src[0]->type == GGML_TYPE_F32 || op->src[0]->type == GGML_TYPE_F16 || op->src[0]->type == GGML_TYPE_BF16) &&
+                op->type == op->src[0]->type &&
+                ggml_is_contiguous(op->src[0]) &&
+                ggml_is_contiguous(op);
         case GGML_OP_CONV_3D:
             return ggml_is_contiguous(op->src[0]) &&
                    ggml_is_contiguous(op->src[1]) &&

ggml/src/ggml-metal/ggml-metal-impl.h

@@ -603,6 +603,16 @@ typedef struct {
     uint64_t nb1;
 } ggml_metal_kargs_conv_transpose_1d;
 
+typedef struct {
+    int32_t  T_in;
+    int32_t  T_out;
+    int32_t  OC;
+    int32_t  K;
+    int32_t  K_OC;
+    int32_t  s0;
+    int32_t  p0;
+} ggml_metal_kargs_col2im_1d;
+
 typedef struct {
     int32_t  IC;
     int32_t  IH;

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

@@ -395,6 +395,10 @@ static int ggml_metal_op_encode_impl(ggml_metal_op_t ctx, int idx) {
             {
                 n_fuse = ggml_metal_op_conv_transpose_2d(ctx, idx);
             } break;
+        case GGML_OP_COL2IM_1D:
+            {
+                n_fuse = ggml_metal_op_col2im_1d(ctx, idx);
+            } break;
         case GGML_OP_CONV_3D:
             {
                 n_fuse = ggml_metal_op_conv_3d(ctx, idx);
@@ -3854,6 +3858,47 @@ int ggml_metal_op_conv_transpose_1d(ggml_metal_op_t ctx, int idx) {
     return 1;
 }
 
+int ggml_metal_op_col2im_1d(ggml_metal_op_t ctx, int idx) {
+    ggml_tensor * op = ctx->node(idx);
+
+    ggml_metal_library_t lib = ctx->lib;
+    ggml_metal_encoder_t enc = ctx->enc;
+
+    const int32_t s0 = ((const int32_t *)(op->op_params))[0];
+    const int32_t OC = ((const int32_t *)(op->op_params))[1];
+    const int32_t p0 = ((const int32_t *)(op->op_params))[2];
+
+    const int32_t K_OC  = (int32_t) op->src[0]->ne[0];
+    const int32_t T_in  = (int32_t) op->src[0]->ne[1];
+    const int32_t K     = K_OC / OC;
+    const int32_t T_out = (int32_t) op->ne[0];
+
+    ggml_metal_kargs_col2im_1d args = {
+        /*.T_in  =*/ T_in,
+        /*.T_out =*/ T_out,
+        /*.OC    =*/ OC,
+        /*.K     =*/ K,
+        /*.K_OC  =*/ K_OC,
+        /*.s0    =*/ s0,
+        /*.p0    =*/ p0,
+    };
+
+    auto pipeline = ggml_metal_library_get_pipeline_col2im_1d(lib, op);
+
+    const int total = T_out * OC;
+    const int nth   = 256;
+    const int ntg   = (total + nth - 1) / nth;
+
+    ggml_metal_encoder_set_pipeline(enc, pipeline);
+    ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), 0);
+    ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[0]), 1);
+    ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op),         2);
+
+    ggml_metal_encoder_dispatch_threadgroups(enc, ntg, 1, 1, nth, 1, 1);
+
+    return 1;
+}
+
 int ggml_metal_op_conv_transpose_2d(ggml_metal_op_t ctx, int idx) {
     ggml_tensor * op = ctx->node(idx);
 

ggml/src/ggml-metal/ggml-metal-ops.h

@@ -78,6 +78,7 @@ int ggml_metal_op_conv_2d           (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_conv_3d           (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_conv_transpose_1d (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_conv_transpose_2d (ggml_metal_op_t ctx, int idx);
+int ggml_metal_op_col2im_1d         (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_upscale           (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_pad               (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_pad_reflect_1d    (ggml_metal_op_t ctx, int idx);

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

@@ -4977,6 +4977,49 @@ kernel void kernel_conv_transpose_1d<half>(
     uint3    tgpg[[threadgroups_per_grid]]);
 
 
+template <typename T>
+kernel void kernel_col2im_1d(
+        constant ggml_metal_kargs_col2im_1d & args,
+        device const T * col,
+        device       T * dst,
+        uint         tgpig [[threadgroup_position_in_grid]],
+        uint         tpitg [[thread_position_in_threadgroup]],
+        uint         ntg   [[threads_per_threadgroup]]) {
+
+    const int idx = tgpig * ntg + tpitg;
+    if (idx >= args.T_out * args.OC) {
+        return;
+    }
+
+    const int t_out = idx % args.T_out;
+    const int oc    = idx / args.T_out;
+    const int t_abs = t_out + args.p0;  // absolute position in uncropped signal
+
+    int t_in_min = (t_abs - args.K + args.s0) / args.s0;  // ceil((t_abs - K + 1) / s0)
+    if (t_in_min < 0) {
+        t_in_min = 0;
+    }
+    int t_in_max = t_abs / args.s0;
+    if (t_in_max >= args.T_in) {
+        t_in_max = args.T_in - 1;
+    }
+
+    float sum = 0.0f;
+    for (int t_in = t_in_min; t_in <= t_in_max; t_in++) {
+        const int k = t_abs - t_in * args.s0;
+        sum += float(col[(oc * args.K + k) + t_in * args.K_OC]);
+    }
+
+    dst[t_out + oc * args.T_out] = T(sum);
+}
+
+template [[host_name("kernel_col2im_1d_f32")]]  kernel void kernel_col2im_1d<float>(constant ggml_metal_kargs_col2im_1d &, device const float *, device float *, uint, uint, uint);
+template [[host_name("kernel_col2im_1d_f16")]]  kernel void kernel_col2im_1d<half>(constant ggml_metal_kargs_col2im_1d &, device const half *, device half *, uint, uint, uint);
+#if defined(GGML_METAL_HAS_BF16)
+template [[host_name("kernel_col2im_1d_bf16")]] kernel void kernel_col2im_1d<bfloat>(constant ggml_metal_kargs_col2im_1d &, device const bfloat *, device bfloat *, uint, uint, uint);
+#endif
+
+
 typedef void (conv_transpose_2d_t)(
         constant ggml_metal_kargs_conv_transpose_2d & args,
         device const float * src0,

ggml/src/ggml-opencl/fa_tune.h

@@ -20,6 +20,7 @@ static const ggml_opencl_fa_dim g_fa_dims_adreno_default[] = {
     {192, 128, 16, 16, 1, 0},
     {192, 192, 16, 16, 1, 0},
     {256, 256, 16, 16, 16, 0},
+    {512, 512,  8, 16, 64, 0},
 };
 
 struct ggml_opencl_fa_dim_table {

ggml/src/ggml-opencl/kernels/flash_attn_f16.cl

@@ -10,7 +10,12 @@
 #define DK_VEC (DK/4)
 #define DV_VEC (DV/4)
 #define WG_SIZE (BLOCK_M)
-#define Q1_WG_SIZE 64
+// q1 reduces over a Q1_WG_SIZE-wide WG via work-group barriers; the launch WG
+// must match. Defaults to the Adreno sg (64); host passes -D FA_SG=32 on Intel.
+#ifndef FA_SG
+#define FA_SG 64
+#endif
+#define Q1_WG_SIZE FA_SG
 
 // The kernels are built with -cl-finite-math-only. On some older Adreno GPUs,
 // infinite operand can cause undefined behavior and miscompilation for exp.

ggml/src/ggml-opencl/kernels/flash_attn_f32.cl

@@ -11,7 +11,12 @@
 #define DK_VEC (DK/4)
 #define DV_VEC (DV/4)
 #define WG_SIZE (BLOCK_M)
-#define Q1_WG_SIZE 64
+// q1 reduces over a Q1_WG_SIZE-wide WG via work-group barriers; the launch WG
+// must match. Defaults to the Adreno sg (64); host passes -D FA_SG=32 on Intel.
+#ifndef FA_SG
+#define FA_SG 64
+#endif
+#define Q1_WG_SIZE FA_SG
 
 // The kernels are built with -cl-finite-math-only. On some older Adreno GPUs,
 // infinite operand can cause undefined behavior and miscompilation for exp.
@@ -114,6 +119,15 @@ __kernel void flash_attn_f32(
     __local DATA_TYPE4 l_v[BLOCK_N][DV_VEC];
 
     for (int k_start = 0; k_start < n_kv; k_start += BLOCK_N) {
+#if FA_SG < 64
+        // WAR on l_k/l_v: threads with my_query_row >= n_q skip the compute below
+        // (continue) and would race ahead to reload the tiles while active threads
+        // still read them. A single 64-wide Adreno subgroup (WG == sg) runs lockstep
+        // and hides this; a WG that spans multiple narrower subgroups (Intel sg=32)
+        // corrupts the result. All threads reach this each iteration (no-op on the
+        // first), so it does not diverge with the continue. Compiled out at sg=64.
+        barrier(CLK_LOCAL_MEM_FENCE);
+#endif
         for (int i = tid; i < BLOCK_N * DK_VEC; i += WG_SIZE) {
             const int row = i / DK_VEC;
             const int col = i % DK_VEC;

ggml/src/ggml-opencl/kernels/flash_attn_f32_q4_0.cl

@@ -27,7 +27,11 @@
 
 #define DK_VEC (DK/4)
 #define DV_VEC (DV/4)
-#define Q1_WG_SIZE 64
+
+#ifndef FA_SG
+#define FA_SG 64
+#endif
+#define Q1_WG_SIZE FA_SG
 
 // The kernels are built with -cl-finite-math-only. On some older Adreno GPUs,
 // infinite operand can cause undefined behavior and miscompilation for exp.
@@ -365,6 +369,263 @@ __kernel void flash_attn_f32_q4_0_q1(
     }
 }
 
+#ifdef cl_intel_subgroups
+#pragma OPENCL EXTENSION cl_intel_subgroups : enable
+#else
+#pragma OPENCL EXTENSION cl_khr_subgroups : enable
+#endif
+
+#ifdef cl_qcom_reqd_sub_group_size
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define REQD_SUBGROUP_SIZE_64 __attribute__((qcom_reqd_sub_group_size("half")))
+#else
+#define REQD_SUBGROUP_SIZE_64
+#endif
+
+#define VEC_NSG          4
+#define VEC_WG_SIZE      (Q1_WG_SIZE * VEC_NSG)
+#define Q1V_DV_PER_THREAD ((DV_VEC + Q1_WG_SIZE - 1) / Q1_WG_SIZE)
+
+// Dequant one float4 lane (0..7) from a q4_0 block.
+// Lanes 0..3 → low nibbles of qs[0..15], lanes 4..7 → high nibbles.
+inline float4 dequant_q4_0_lane(const global char * block_ptr, int lane) {
+    const float d = vload_half(0, (const global half *)block_ptr);
+    const global uchar * qs = (const global uchar *)(block_ptr + 2);
+    const int g     = lane & 3;
+    const int shift = (lane < 4) ? 0 : 4;
+    return d * (float4)((float)((qs[g*4+0] >> shift) & 0x0F) - 8.0f,
+                        (float)((qs[g*4+1] >> shift) & 0x0F) - 8.0f,
+                        (float)((qs[g*4+2] >> shift) & 0x0F) - 8.0f,
+                        (float)((qs[g*4+3] >> shift) & 0x0F) - 8.0f);
+}
+
+REQD_SUBGROUP_SIZE_64
+__kernel void flash_attn_f32_q4_0_q1_vec(
+    const global void * q_void, ulong q_offset,
+    const global void * k_void, ulong k_offset,
+    const global void * v_void, ulong v_offset,
+    global void * o_void, ulong o_offset,
+    const float scale,
+    const int n_q,
+    const int n_kv,
+    const int is_causal,
+    const int n_head,
+    const ulong q_nb1, const ulong q_nb2, const ulong q_nb3,
+    const ulong k_nb1, const ulong k_nb2, const ulong k_nb3,
+    const ulong v_nb1, const ulong v_nb2, const ulong v_nb3,
+    const ulong o_nb1, const ulong o_nb2, const ulong o_nb3,
+    const float max_bias,
+    const float m0,
+    const float m1,
+    const int n_head_log2,
+    const float logit_softcap,
+    const int n_head_kv,
+    const global void* mask_void,
+    const ulong mask_offset,
+    const ulong mask_nb1,
+    const ulong mask_nb2,
+    const ulong mask_nb3,
+    const int mask_ne2,
+    const int mask_ne3,
+    const global void* sinks_void,
+    const ulong sinks_offset
+) {
+    const int tid             = get_local_id(0);
+    const int sgid            = tid / Q1_WG_SIZE;
+    const int tid_sg          = tid % Q1_WG_SIZE;
+    const int head_batch_idx  = get_global_id(1);
+
+    const int batch_idx = head_batch_idx / n_head;
+    const int head_idx  = head_batch_idx % n_head;
+
+    const int gqa_ratio   = n_head / n_head_kv;
+    const int head_kv_idx = head_idx / gqa_ratio;
+
+    const global char * q_base = (const global char *) q_void + q_offset;
+    const global char * k_base = (const global char *) k_void + k_offset;
+    const global char * v_base = (const global char *) v_void + v_offset;
+    global       char * o_base = (global       char *) o_void + o_offset;
+
+    const global char * mask_base = NULL;
+    if (mask_void != NULL) {
+        const int mask_head_idx  = head_idx  % mask_ne2;
+        const int mask_batch_idx = batch_idx % mask_ne3;
+        mask_base = (const global char *) mask_void + mask_offset +
+                    mask_batch_idx * mask_nb3 + mask_head_idx * mask_nb2;
+    }
+
+    __local ACC_TYPE4 q_shared[DK_VEC];
+    {
+        const ulong q_row_offset = batch_idx * q_nb3 + head_idx * q_nb2;
+        const global Q_DATA_TYPE4 * q_ptr = (const global Q_DATA_TYPE4 *) (q_base + q_row_offset);
+        for (int i = tid; i < DK_VEC; i += VEC_WG_SIZE) {
+            q_shared[i] = CONVERT_Q_ACC4(q_ptr[i]);
+        }
+    }
+    barrier(CLK_LOCAL_MEM_FENCE);
+
+#ifdef FA_HAVE_INT_DOT
+    // quantize Q to int8-packed uints + per-block (qd, q_sum) once per WG for dp4a
+    // one thread per Q block, remaining threads idle this step
+    __local uint  q_packed_shared[DK_Q4_BLOCKS * 8];
+    __local float q_d_shared[DK_Q4_BLOCKS];
+    __local int   q_sum_shared[DK_Q4_BLOCKS];
+    if (tid < DK_Q4_BLOCKS) {
+        ACC_TYPE4 q_block[8];
+        #pragma unroll
+        for (int i = 0; i < 8; ++i) q_block[i] = q_shared[tid * 8 + i];
+        uint packed[8];
+        q4_q_block_info info = quant_q_block_int8_packed_q4(q_block, packed);
+        #pragma unroll
+        for (int i = 0; i < 8; ++i) q_packed_shared[tid * 8 + i] = packed[i];
+        q_d_shared[tid]   = info.qd;
+        q_sum_shared[tid] = info.q_sum;
+    }
+    barrier(CLK_LOCAL_MEM_FENCE);
+#endif
+
+    const float slope = get_alibi_slope(max_bias, head_idx, n_head_log2, m0, m1);
+
+    const global ACC_TYPE * sinks_ptr = NULL;
+    if (sinks_void != NULL) {
+        sinks_ptr = (const global ACC_TYPE *) ((const global char *) sinks_void + sinks_offset);
+    }
+
+    ACC_TYPE4 o_acc[Q1V_DV_PER_THREAD];
+    #pragma unroll
+    for (int i = 0; i < Q1V_DV_PER_THREAD; ++i) o_acc[i] = (ACC_TYPE4)(0.0f);
+
+    ACC_TYPE m_i = FA_M_INIT;
+    ACC_TYPE l_i = 0.0f;
+
+    const int kv_per_sg = (n_kv + VEC_NSG - 1) / VEC_NSG;
+    const int kv_start  = sgid * kv_per_sg;
+    const int kv_end    = min(n_kv, kv_start + kv_per_sg);
+
+    for (int k_idx = kv_start; k_idx < kv_end; ++k_idx) {
+        const global char * k_row = k_base + batch_idx * k_nb3 + head_kv_idx * k_nb2 + k_idx * k_nb1;
+        const global char * v_row = v_base + batch_idx * v_nb3 + head_kv_idx * v_nb2 + k_idx * v_nb1;
+
+#ifdef FA_HAVE_INT_DOT
+        // per-lane dp4a: each lane packs 4 raw q4_0 nibbles into a uint,
+        // then dot_acc_sat_4x8packed_ss_int against the matching uint.
+        ACC_TYPE lane_contrib = 0.0f;
+        for (int qk = tid_sg; qk < DK_VEC; qk += Q1_WG_SIZE) {
+            const int block_idx     = qk / 8;
+            const int lane_in_block = qk % 8;
+            const int g             = lane_in_block & 3;
+            const int shift         = (lane_in_block < 4) ? 0 : 4;
+            const global char *  k_block = k_row + block_idx * Q4_0_BLOCK_SIZE;
+            const float kd = vload_half(0, (const global half *)k_block);
+            const global uchar * k_qs = (const global uchar *)(k_block + 2);
+            const uchar b0 = k_qs[g*4 + 0];
+            const uchar b1 = k_qs[g*4 + 1];
+            const uchar b2 = k_qs[g*4 + 2];
+            const uchar b3 = k_qs[g*4 + 3];
+            const uint k_packed = ((uint)((b0 >> shift) & 0x0F))       |
+                                  ((uint)((b1 >> shift) & 0x0F)) <<  8 |
+                                  ((uint)((b2 >> shift) & 0x0F)) << 16 |
+                                  ((uint)((b3 >> shift) & 0x0F)) << 24;
+            const uint q_packed_lane = q_packed_shared[block_idx * 8 + lane_in_block];
+            const int  raw_dot = dot_acc_sat_4x8packed_ss_int(q_packed_lane, k_packed, 0);
+            const float qd          = q_d_shared[block_idx];
+            const float block_scale = qd * kd;
+            float contrib = (float)raw_dot * block_scale;
+            if (lane_in_block == 0) {
+                // block bias correction is per-block
+                const int q_sum_b = q_sum_shared[block_idx];
+                contrib -= 8.0f * block_scale * (float)q_sum_b;
+            }
+            lane_contrib += contrib;
+        }
+        ACC_TYPE score = sub_group_reduce_add(lane_contrib) * scale;
+#else
+        ACC_TYPE4 dot4 = (ACC_TYPE4)(0.0f);
+        for (int qk = tid_sg; qk < DK_VEC; qk += Q1_WG_SIZE) {
+            const int block_idx = qk / 8;
+            const int lane      = qk % 8;
+            const float4 k_v = dequant_q4_0_lane(k_row + block_idx * Q4_0_BLOCK_SIZE, lane);
+            dot4 = mad(q_shared[qk], k_v, dot4);
+        }
+        ACC_TYPE dot_partial = dot4.s0 + dot4.s1 + dot4.s2 + dot4.s3;
+        ACC_TYPE score = sub_group_reduce_add(dot_partial) * scale;
+#endif
+
+        if (mask_base != NULL) {
+            const global MASK_DATA_TYPE * mask_ptr = (const global MASK_DATA_TYPE *) mask_base;
+            score += slope * (ACC_TYPE) mask_ptr[k_idx];
+        }
+        if (logit_softcap > 0.0f) {
+            score = logit_softcap * tanh(score / logit_softcap);
+        }
+
+        const ACC_TYPE m_new      = max(m_i, score);
+        const ACC_TYPE scale_prev = native_exp(m_i - m_new);
+        const ACC_TYPE p          = native_exp(score - m_new);
+
+        int idx = 0;
+        for (int dv = tid_sg; dv < DV_VEC; dv += Q1_WG_SIZE, ++idx) {
+            const int block_idx = dv / 8;
+            const int lane      = dv % 8;
+            const float4 v_v = dequant_q4_0_lane(v_row + block_idx * Q4_0_BLOCK_SIZE, lane);
+            o_acc[idx] = mad(p, v_v, o_acc[idx] * scale_prev);
+        }
+        l_i = l_i * scale_prev + p;
+        m_i = m_new;
+    }
+
+    __local ACC_TYPE  sg_m[VEC_NSG];
+    __local ACC_TYPE  sg_l[VEC_NSG];
+    __local ACC_TYPE4 sg_o[VEC_NSG][DV_VEC];
+
+    if (tid_sg == 0) {
+        sg_m[sgid] = m_i;
+        sg_l[sgid] = l_i;
+    }
+    {
+        int idx = 0;
+        for (int dv = tid_sg; dv < DV_VEC; dv += Q1_WG_SIZE, ++idx) {
+            sg_o[sgid][dv] = o_acc[idx];
+        }
+    }
+    barrier(CLK_LOCAL_MEM_FENCE);
+
+    if (sgid == 0) {
+        ACC_TYPE m_final = sg_m[0];
+        #pragma unroll
+        for (int s = 1; s < VEC_NSG; ++s) {
+            m_final = max(m_final, sg_m[s]);
+        }
+        if (sinks_ptr != NULL) {
+            m_final = max(m_final, sinks_ptr[head_idx]);
+        }
+
+        ACC_TYPE l_final = 0.0f;
+        #pragma unroll
+        for (int s = 0; s < VEC_NSG; ++s) {
+            l_final += sg_l[s] * native_exp(sg_m[s] - m_final);
+        }
+        if (sinks_ptr != NULL) {
+            l_final += native_exp(sinks_ptr[head_idx] - m_final);
+        }
+        const ACC_TYPE l_inv = (l_final > 0.0f) ? (1.0f / l_final) : 0.0f;
+
+        const ulong o_row_offset = batch_idx * o_nb3 + head_idx * o_nb1;
+        global O_DATA_TYPE4 * o_row = (global O_DATA_TYPE4 *) (o_base + o_row_offset);
+
+        int idx = 0;
+        for (int dv = tid_sg; dv < DV_VEC; dv += Q1_WG_SIZE, ++idx) {
+            ACC_TYPE4 o_merged = (ACC_TYPE4)(0.0f);
+            #pragma unroll
+            for (int s = 0; s < VEC_NSG; ++s) {
+                const ACC_TYPE alpha = native_exp(sg_m[s] - m_final);
+                o_merged = mad((ACC_TYPE4)(alpha), sg_o[s][dv], o_merged);
+            }
+            o_row[dv] = CONVERT_O_DATA4(o_merged * l_inv);
+        }
+    }
+}
+
 // Flash-decoding split pass for q4_0 KV. Merge kernel is type-agnostic and
 // shared with the f16/q8_0 FA kernels.
 #define FA_PARTIAL_FLOATS (2 + DV)
@@ -583,6 +844,319 @@ __kernel void flash_attn_f32_q4_0_q1_split(
 #define WG_SIZE             BLOCK_M
 #endif
 
+#ifndef MQ_GQA
+#define MQ_GQA 4
+#endif
+#ifndef MQ_NSG_SPLIT
+#define MQ_NSG_SPLIT 4
+#endif
+#define MQ_SPLIT_WG_SIZE_Q4 (Q1_WG_SIZE * MQ_NSG_SPLIT)
+
+REQD_SUBGROUP_SIZE_64
+__kernel void flash_attn_f32_q4_0_q1_vec_mq_split(
+    const global void * q_void, ulong q_offset,
+    const global void * k_void, ulong k_offset,
+    const global void * v_void, ulong v_offset,
+    const float scale,
+    const int n_q,
+    const int n_kv,
+    const int n_head,
+    const ulong q_nb1, const ulong q_nb2, const ulong q_nb3,
+    const ulong k_nb1, const ulong k_nb2, const ulong k_nb3,
+    const ulong v_nb1, const ulong v_nb2, const ulong v_nb3,
+    const float max_bias,
+    const float m0,
+    const float m1,
+    const int n_head_log2,
+    const float logit_softcap,
+    const int n_head_kv,
+    const global void * mask_void,
+    const ulong mask_offset,
+    const ulong mask_nb1,
+    const ulong mask_nb2,
+    const ulong mask_nb3,
+    const int mask_ne2,
+    const int mask_ne3,
+    global float * partial_void,
+    const int n_splits,
+    const int kv_per_split
+) {
+    const int tid              = get_local_id(0);
+    const int sgid             = tid / Q1_WG_SIZE;
+    const int tid_sg           = tid % Q1_WG_SIZE;
+    const int kvhead_batch_idx = get_global_id(1);
+    const int split_q_idx      = get_global_id(2);
+    const int split_idx        = split_q_idx % n_splits;
+    const int q_idx            = split_q_idx / n_splits;
+
+    const int batch_idx   = kvhead_batch_idx / n_head_kv;
+    const int head_kv_idx = kvhead_batch_idx % n_head_kv;
+
+    const int kv_start = split_idx * kv_per_split;
+    const int kv_end   = min(kv_start + kv_per_split, n_kv);
+
+    const ulong record_stride = (ulong) FA_PARTIAL_FLOATS;
+
+    if (kv_start >= kv_end) {
+        if (tid == 0) {
+            #pragma unroll
+            for (int h = 0; h < MQ_GQA; ++h) {
+                const int head_idx = head_kv_idx * MQ_GQA + h;
+                const ulong rec_idx = ((((ulong) batch_idx * n_head + head_idx) * n_q + q_idx)
+                                       * n_splits + split_idx);
+                global float * rec = partial_void + rec_idx * record_stride;
+                rec[0] = FA_M_INIT;
+                rec[1] = 0.0f;
+            }
+        }
+        return;
+    }
+
+    const global char * q_base = (const global char *) q_void + q_offset;
+    const global char * k_base = (const global char *) k_void + k_offset;
+    const global char * v_base = (const global char *) v_void + v_offset;
+
+    __local ACC_TYPE4 q_shared[MQ_GQA * DK_VEC];
+    for (int i = tid; i < MQ_GQA * DK_VEC; i += MQ_SPLIT_WG_SIZE_Q4) {
+        const int h        = i / DK_VEC;
+        const int k        = i % DK_VEC;
+        const int head_idx = head_kv_idx * MQ_GQA + h;
+        const ulong q_row_offset = batch_idx * q_nb3 + head_idx * q_nb2 + (ulong) q_idx * q_nb1;
+        const global Q_DATA_TYPE4 * q_ptr = (const global Q_DATA_TYPE4 *) (q_base + q_row_offset);
+        q_shared[h * DK_VEC + k] = CONVERT_Q_ACC4(q_ptr[k]);
+    }
+    barrier(CLK_LOCAL_MEM_FENCE);
+
+#ifdef FA_HAVE_INT_DOT
+    __local uint  q_packed_shared[MQ_GQA * DK_Q4_BLOCKS * 8];
+    __local float q_d_shared[MQ_GQA * DK_Q4_BLOCKS];
+    __local int   q_sum_shared[MQ_GQA * DK_Q4_BLOCKS];
+    {
+        const int active = MQ_GQA * DK_Q4_BLOCKS;
+        if (tid < active) {
+            const int h        = tid / DK_Q4_BLOCKS;
+            const int block_id = tid % DK_Q4_BLOCKS;
+            ACC_TYPE4 q_block[8];
+            #pragma unroll
+            for (int i = 0; i < 8; ++i) q_block[i] = q_shared[h * DK_VEC + block_id * 8 + i];
+            uint packed[8];
+            q4_q_block_info info = quant_q_block_int8_packed_q4(q_block, packed);
+            #pragma unroll
+            for (int i = 0; i < 8; ++i) q_packed_shared[(h * DK_Q4_BLOCKS + block_id) * 8 + i] = packed[i];
+            q_d_shared[h * DK_Q4_BLOCKS + block_id]   = info.qd;
+            q_sum_shared[h * DK_Q4_BLOCKS + block_id] = info.q_sum;
+        }
+    }
+    barrier(CLK_LOCAL_MEM_FENCE);
+#endif
+
+    float slope[MQ_GQA];
+    #pragma unroll
+    for (int h = 0; h < MQ_GQA; ++h) {
+        slope[h] = get_alibi_slope(max_bias, head_kv_idx * MQ_GQA + h, n_head_log2, m0, m1);
+    }
+
+    const global char * mask_base[MQ_GQA];
+    if (mask_void != NULL) {
+        const int mask_batch_idx = batch_idx % mask_ne3;
+        const global char * mask_base_b = (const global char *) mask_void + mask_offset +
+                                          mask_batch_idx * mask_nb3 +
+                                          (ulong) q_idx * mask_nb1;
+        #pragma unroll
+        for (int h = 0; h < MQ_GQA; ++h) {
+            const int head_idx      = head_kv_idx * MQ_GQA + h;
+            const int mask_head_idx = head_idx % mask_ne2;
+            mask_base[h] = mask_base_b + mask_head_idx * mask_nb2;
+        }
+    } else {
+        #pragma unroll
+        for (int h = 0; h < MQ_GQA; ++h) mask_base[h] = NULL;
+    }
+
+    ACC_TYPE4 o_acc[MQ_GQA][Q1V_DV_PER_THREAD];
+    ACC_TYPE  m_i[MQ_GQA];
+    ACC_TYPE  l_i[MQ_GQA];
+    #pragma unroll
+    for (int h = 0; h < MQ_GQA; ++h) {
+        m_i[h] = FA_M_INIT;
+        l_i[h] = 0.0f;
+        #pragma unroll
+        for (int i = 0; i < Q1V_DV_PER_THREAD; ++i) o_acc[h][i] = (ACC_TYPE4)(0.0f);
+    }
+
+    const int kv_len    = kv_end - kv_start;
+    const int kv_per_sg = (kv_len + MQ_NSG_SPLIT - 1) / MQ_NSG_SPLIT;
+    const int kv_lo     = kv_start + sgid * kv_per_sg;
+    const int kv_hi     = min(kv_end, kv_lo + kv_per_sg);
+
+    for (int k_idx = kv_lo; k_idx < kv_hi; ++k_idx) {
+        const global char * k_row = k_base + batch_idx * k_nb3 + head_kv_idx * k_nb2 + k_idx * k_nb1;
+        const global char * v_row = v_base + batch_idx * v_nb3 + head_kv_idx * v_nb2 + k_idx * v_nb1;
+
+#ifdef FA_HAVE_INT_DOT
+        ACC_TYPE lane_contrib[MQ_GQA];
+        #pragma unroll
+        for (int h = 0; h < MQ_GQA; ++h) lane_contrib[h] = 0.0f;
+
+        for (int qk = tid_sg; qk < DK_VEC; qk += Q1_WG_SIZE) {
+            const int block_idx     = qk / 8;
+            const int lane_in_block = qk % 8;
+            const int g             = lane_in_block & 3;
+            const int shift         = (lane_in_block < 4) ? 0 : 4;
+            const global char *  k_block = k_row + block_idx * Q4_0_BLOCK_SIZE;
+            const float kd = vload_half(0, (const global half *)k_block);
+            const global uchar * k_qs = (const global uchar *)(k_block + 2);
+            const uchar b0 = k_qs[g*4 + 0];
+            const uchar b1 = k_qs[g*4 + 1];
+            const uchar b2 = k_qs[g*4 + 2];
+            const uchar b3 = k_qs[g*4 + 3];
+            const uint k_packed = ((uint)((b0 >> shift) & 0x0F))       |
+                                  ((uint)((b1 >> shift) & 0x0F)) <<  8 |
+                                  ((uint)((b2 >> shift) & 0x0F)) << 16 |
+                                  ((uint)((b3 >> shift) & 0x0F)) << 24;
+            #pragma unroll
+            for (int h = 0; h < MQ_GQA; ++h) {
+                const uint  q_packed_lane = q_packed_shared[(h * DK_Q4_BLOCKS + block_idx) * 8 + lane_in_block];
+                const int   raw_dot       = dot_acc_sat_4x8packed_ss_int(q_packed_lane, k_packed, 0);
+                const float qd            = q_d_shared[h * DK_Q4_BLOCKS + block_idx];
+                const float block_scale   = qd * kd;
+                float contrib = (float) raw_dot * block_scale;
+                if (lane_in_block == 0) {
+                    const int q_sum_b = q_sum_shared[h * DK_Q4_BLOCKS + block_idx];
+                    contrib -= 8.0f * block_scale * (float) q_sum_b;
+                }
+                lane_contrib[h] += contrib;
+            }
+        }
+
+        ACC_TYPE score[MQ_GQA];
+        #pragma unroll
+        for (int h = 0; h < MQ_GQA; ++h) {
+            ACC_TYPE s = sub_group_reduce_add(lane_contrib[h]) * scale;
+            if (mask_base[h] != NULL) {
+                const global MASK_DATA_TYPE * mask_ptr = (const global MASK_DATA_TYPE *) mask_base[h];
+                s += slope[h] * (ACC_TYPE) mask_ptr[k_idx];
+            }
+            if (logit_softcap > 0.0f) {
+                s = logit_softcap * tanh(s / logit_softcap);
+            }
+            score[h] = s;
+        }
+#else
+        // fallback float-dequant K dot
+        ACC_TYPE4 dot4[MQ_GQA];
+        #pragma unroll
+        for (int h = 0; h < MQ_GQA; ++h) dot4[h] = (ACC_TYPE4)(0.0f);
+
+        for (int qk = tid_sg; qk < DK_VEC; qk += Q1_WG_SIZE) {
+            const int block_idx = qk / 8;
+            const int lane      = qk % 8;
+            const float4 k_v = dequant_q4_0_lane(k_row + block_idx * Q4_0_BLOCK_SIZE, lane);
+            #pragma unroll
+            for (int h = 0; h < MQ_GQA; ++h) {
+                dot4[h] = mad(q_shared[h * DK_VEC + qk], k_v, dot4[h]);
+            }
+        }
+
+        ACC_TYPE score[MQ_GQA];
+        #pragma unroll
+        for (int h = 0; h < MQ_GQA; ++h) {
+            const ACC_TYPE dot_partial = dot4[h].s0 + dot4[h].s1 + dot4[h].s2 + dot4[h].s3;
+            ACC_TYPE s = sub_group_reduce_add(dot_partial) * scale;
+            if (mask_base[h] != NULL) {
+                const global MASK_DATA_TYPE * mask_ptr = (const global MASK_DATA_TYPE *) mask_base[h];
+                s += slope[h] * (ACC_TYPE) mask_ptr[k_idx];
+            }
+            if (logit_softcap > 0.0f) {
+                s = logit_softcap * tanh(s / logit_softcap);
+            }
+            score[h] = s;
+        }
+#endif
+
+        ACC_TYPE p_h[MQ_GQA];
+        ACC_TYPE sp_h[MQ_GQA];
+        #pragma unroll
+        for (int h = 0; h < MQ_GQA; ++h) {
+            const ACC_TYPE m_new = max(m_i[h], score[h]);
+            sp_h[h] = native_exp(m_i[h] - m_new);
+            p_h[h]  = native_exp(score[h] - m_new);
+            l_i[h]  = l_i[h] * sp_h[h] + p_h[h];
+            m_i[h]  = m_new;
+        }
+
+        int idx = 0;
+        for (int dv = tid_sg; dv < DV_VEC; dv += Q1_WG_SIZE, ++idx) {
+            const int block_idx = dv / 8;
+            const int lane      = dv % 8;
+            const float4 v_v = dequant_q4_0_lane(v_row + block_idx * Q4_0_BLOCK_SIZE, lane);
+            #pragma unroll
+            for (int h = 0; h < MQ_GQA; ++h) {
+                o_acc[h][idx] = mad(p_h[h], v_v, o_acc[h][idx] * sp_h[h]);
+            }
+        }
+    }
+
+    // per-h cross-subgroup merge
+    __local ACC_TYPE  sg_m[MQ_GQA][MQ_NSG_SPLIT];
+    __local ACC_TYPE  sg_l[MQ_GQA][MQ_NSG_SPLIT];
+    __local ACC_TYPE4 sg_o[MQ_NSG_SPLIT][DV_VEC];
+
+    if (tid_sg == 0) {
+        #pragma unroll
+        for (int h = 0; h < MQ_GQA; ++h) {
+            sg_m[h][sgid] = m_i[h];
+            sg_l[h][sgid] = l_i[h];
+        }
+    }
+
+    #pragma unroll
+    for (int h = 0; h < MQ_GQA; ++h) {
+        {
+            int idx = 0;
+            for (int dv_idx = tid_sg; dv_idx < DV_VEC; dv_idx += Q1_WG_SIZE, ++idx) {
+                sg_o[sgid][dv_idx] = o_acc[h][idx];
+            }
+        }
+        barrier(CLK_LOCAL_MEM_FENCE);
+
+        if (sgid == 0) {
+            const int head_idx = head_kv_idx * MQ_GQA + h;
+
+            ACC_TYPE m_c = sg_m[h][0];
+            #pragma unroll
+            for (int s = 1; s < MQ_NSG_SPLIT; ++s) {
+                m_c = max(m_c, sg_m[h][s]);
+            }
+            ACC_TYPE l_c = 0.0f;
+            #pragma unroll
+            for (int s = 0; s < MQ_NSG_SPLIT; ++s) {
+                l_c += sg_l[h][s] * native_exp(sg_m[h][s] - m_c);
+            }
+
+            const ulong rec_idx = ((((ulong) batch_idx * n_head + head_idx) * n_q + q_idx)
+                                   * n_splits + split_idx);
+            global float  * rec   = partial_void + rec_idx * record_stride;
+            global float4 * rec_o = (global float4 *) (rec + 2);
+
+            if (tid_sg == 0) {
+                rec[0] = (float) m_c;
+                rec[1] = (float) l_c;
+            }
+            for (int dv_idx = tid_sg; dv_idx < DV_VEC; dv_idx += Q1_WG_SIZE) {
+                ACC_TYPE4 o_merged = (ACC_TYPE4)(0.0f);
+                #pragma unroll
+                for (int s = 0; s < MQ_NSG_SPLIT; ++s) {
+                    const ACC_TYPE alpha = native_exp(sg_m[h][s] - m_c);
+                    o_merged = mad((ACC_TYPE4)(alpha), sg_o[s][dv_idx], o_merged);
+                }
+                rec_o[dv_idx] = o_merged;
+            }
+        }
+        barrier(CLK_LOCAL_MEM_FENCE);
+    }
+}
+
 __kernel void flash_attn_f32_q4_0(
     const global void * q_void, ulong q_offset,
     const global void * k_void, ulong k_offset,

ggml/src/ggml-opencl/kernels/flash_attn_f32_q8_0.cl

@@ -24,7 +24,11 @@
 
 #define DK_VEC (DK/4)
 #define DV_VEC (DV/4)
-#define Q1_WG_SIZE 64
+
+#ifndef FA_SG
+#define FA_SG 64
+#endif
+#define Q1_WG_SIZE FA_SG
 
 // The kernels are built with -cl-finite-math-only. On some older Adreno GPUs,
 // infinite operand can cause undefined behavior and miscompilation for exp.
@@ -310,6 +314,201 @@ __kernel void flash_attn_f32_q8_0_q1(
     }
 }
 
+#ifdef cl_intel_subgroups
+#pragma OPENCL EXTENSION cl_intel_subgroups : enable
+#else
+#pragma OPENCL EXTENSION cl_khr_subgroups : enable
+#endif
+
+#ifdef cl_qcom_reqd_sub_group_size
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define REQD_SUBGROUP_SIZE_64 __attribute__((qcom_reqd_sub_group_size("half")))
+#else
+#define REQD_SUBGROUP_SIZE_64
+#endif
+
+#define VEC_NSG          4
+#define VEC_WG_SIZE      (Q1_WG_SIZE * VEC_NSG)
+#define Q1V_DV_PER_THREAD ((DV_VEC + Q1_WG_SIZE - 1) / Q1_WG_SIZE)
+
+inline float4 dequant_q8_0_lane(const global char * block_ptr, int lane) {
+    const float d = vload_half(0, (const global half *)block_ptr);
+    const global char * qs = block_ptr + 2 + lane * 4;
+    return d * (float4)((float)qs[0], (float)qs[1], (float)qs[2], (float)qs[3]);
+}
+
+REQD_SUBGROUP_SIZE_64
+__kernel void flash_attn_f32_q8_0_q1_vec(
+    const global void * q_void, ulong q_offset,
+    const global void * k_void, ulong k_offset,
+    const global void * v_void, ulong v_offset,
+    global void * o_void, ulong o_offset,
+    const float scale,
+    const int n_q,
+    const int n_kv,
+    const int is_causal,
+    const int n_head,
+    const ulong q_nb1, const ulong q_nb2, const ulong q_nb3,
+    const ulong k_nb1, const ulong k_nb2, const ulong k_nb3,
+    const ulong v_nb1, const ulong v_nb2, const ulong v_nb3,
+    const ulong o_nb1, const ulong o_nb2, const ulong o_nb3,
+    const float max_bias,
+    const float m0,
+    const float m1,
+    const int n_head_log2,
+    const float logit_softcap,
+    const int n_head_kv,
+    const global void* mask_void,
+    const ulong mask_offset,
+    const ulong mask_nb1,
+    const ulong mask_nb2,
+    const ulong mask_nb3,
+    const int mask_ne2,
+    const int mask_ne3,
+    const global void* sinks_void,
+    const ulong sinks_offset
+) {
+    const int tid             = get_local_id(0);
+    const int sgid            = tid / Q1_WG_SIZE;
+    const int tid_sg          = tid % Q1_WG_SIZE;
+    const int head_batch_idx  = get_global_id(1);
+
+    const int batch_idx = head_batch_idx / n_head;
+    const int head_idx  = head_batch_idx % n_head;
+
+    const int gqa_ratio   = n_head / n_head_kv;
+    const int head_kv_idx = head_idx / gqa_ratio;
+
+    const global char * q_base = (const global char *) q_void + q_offset;
+    const global char * k_base = (const global char *) k_void + k_offset;
+    const global char * v_base = (const global char *) v_void + v_offset;
+    global       char * o_base = (global       char *) o_void + o_offset;
+
+    const global char * mask_base = NULL;
+    if (mask_void != NULL) {
+        const int mask_head_idx  = head_idx  % mask_ne2;
+        const int mask_batch_idx = batch_idx % mask_ne3;
+        mask_base = (const global char *) mask_void + mask_offset +
+                    mask_batch_idx * mask_nb3 + mask_head_idx * mask_nb2;
+    }
+
+    __local ACC_TYPE4 q_shared[DK_VEC];
+    {
+        const ulong q_row_offset = batch_idx * q_nb3 + head_idx * q_nb2;
+        const global Q_DATA_TYPE4 * q_ptr = (const global Q_DATA_TYPE4 *) (q_base + q_row_offset);
+        for (int i = tid; i < DK_VEC; i += VEC_WG_SIZE) {
+            q_shared[i] = CONVERT_Q_ACC4(q_ptr[i]);
+        }
+    }
+    barrier(CLK_LOCAL_MEM_FENCE);
+
+    const float slope = get_alibi_slope(max_bias, head_idx, n_head_log2, m0, m1);
+
+    const global ACC_TYPE * sinks_ptr = NULL;
+    if (sinks_void != NULL) {
+        sinks_ptr = (const global ACC_TYPE *) ((const global char *) sinks_void + sinks_offset);
+    }
+
+    ACC_TYPE4 o_acc[Q1V_DV_PER_THREAD];
+    #pragma unroll
+    for (int i = 0; i < Q1V_DV_PER_THREAD; ++i) o_acc[i] = (ACC_TYPE4)(0.0f);
+
+    ACC_TYPE m_i = FA_M_INIT;
+    ACC_TYPE l_i = 0.0f;
+
+    const int kv_per_sg = (n_kv + VEC_NSG - 1) / VEC_NSG;
+    const int kv_start  = sgid * kv_per_sg;
+    const int kv_end    = min(n_kv, kv_start + kv_per_sg);
+
+    for (int k_idx = kv_start; k_idx < kv_end; ++k_idx) {
+        const global char * k_row = k_base + batch_idx * k_nb3 + head_kv_idx * k_nb2 + k_idx * k_nb1;
+        const global char * v_row = v_base + batch_idx * v_nb3 + head_kv_idx * v_nb2 + k_idx * v_nb1;
+
+        ACC_TYPE4 dot4 = (ACC_TYPE4)(0.0f);
+        for (int qk = tid_sg; qk < DK_VEC; qk += Q1_WG_SIZE) {
+            const int block_idx = qk / 8;
+            const int lane      = qk % 8;
+            const float4 k_v = dequant_q8_0_lane(k_row + block_idx * Q8_0_BLOCK_SIZE, lane);
+            dot4 = mad(q_shared[qk], k_v, dot4);
+        }
+        ACC_TYPE dot_partial = dot4.s0 + dot4.s1 + dot4.s2 + dot4.s3;
+        ACC_TYPE score = sub_group_reduce_add(dot_partial) * scale;
+
+        if (mask_base != NULL) {
+            const global MASK_DATA_TYPE * mask_ptr = (const global MASK_DATA_TYPE *) mask_base;
+            score += slope * (ACC_TYPE) mask_ptr[k_idx];
+        }
+        if (logit_softcap > 0.0f) {
+            score = logit_softcap * tanh(score / logit_softcap);
+        }
+
+        const ACC_TYPE m_new      = max(m_i, score);
+        const ACC_TYPE scale_prev = native_exp(m_i - m_new);
+        const ACC_TYPE p          = native_exp(score - m_new);
+
+        int idx = 0;
+        for (int dv = tid_sg; dv < DV_VEC; dv += Q1_WG_SIZE, ++idx) {
+            const int block_idx = dv / 8;
+            const int lane      = dv % 8;
+            const float4 v_v = dequant_q8_0_lane(v_row + block_idx * Q8_0_BLOCK_SIZE, lane);
+            o_acc[idx] = mad(p, v_v, o_acc[idx] * scale_prev);
+        }
+        l_i = l_i * scale_prev + p;
+        m_i = m_new;
+    }
+
+    __local ACC_TYPE  sg_m[VEC_NSG];
+    __local ACC_TYPE  sg_l[VEC_NSG];
+    __local ACC_TYPE4 sg_o[VEC_NSG][DV_VEC];
+
+    if (tid_sg == 0) {
+        sg_m[sgid] = m_i;
+        sg_l[sgid] = l_i;
+    }
+    {
+        int idx = 0;
+        for (int dv = tid_sg; dv < DV_VEC; dv += Q1_WG_SIZE, ++idx) {
+            sg_o[sgid][dv] = o_acc[idx];
+        }
+    }
+    barrier(CLK_LOCAL_MEM_FENCE);
+
+    if (sgid == 0) {
+        ACC_TYPE m_final = sg_m[0];
+        #pragma unroll
+        for (int s = 1; s < VEC_NSG; ++s) {
+            m_final = max(m_final, sg_m[s]);
+        }
+        if (sinks_ptr != NULL) {
+            m_final = max(m_final, sinks_ptr[head_idx]);
+        }
+
+        ACC_TYPE l_final = 0.0f;
+        #pragma unroll
+        for (int s = 0; s < VEC_NSG; ++s) {
+            l_final += sg_l[s] * native_exp(sg_m[s] - m_final);
+        }
+        if (sinks_ptr != NULL) {
+            l_final += native_exp(sinks_ptr[head_idx] - m_final);
+        }
+        const ACC_TYPE l_inv = (l_final > 0.0f) ? (1.0f / l_final) : 0.0f;
+
+        const ulong o_row_offset = batch_idx * o_nb3 + head_idx * o_nb1;
+        global O_DATA_TYPE4 * o_row = (global O_DATA_TYPE4 *) (o_base + o_row_offset);
+
+        int idx = 0;
+        for (int dv = tid_sg; dv < DV_VEC; dv += Q1_WG_SIZE, ++idx) {
+            ACC_TYPE4 o_merged = (ACC_TYPE4)(0.0f);
+            #pragma unroll
+            for (int s = 0; s < VEC_NSG; ++s) {
+                const ACC_TYPE alpha = native_exp(sg_m[s] - m_final);
+                o_merged = mad((ACC_TYPE4)(alpha), sg_o[s][dv], o_merged);
+            }
+            o_row[dv] = CONVERT_O_DATA4(o_merged * l_inv);
+        }
+    }
+}
+
 // Flash-decoding split pass for q8_0 KV. Partial record: [m, l, O[DV]].
 // Merge kernel from flash_attn_f32_f16.cl is type-agnostic and reused.
 #define FA_PARTIAL_FLOATS (2 + DV)
@@ -533,6 +732,244 @@ __kernel void flash_attn_f32_q8_0_q1_split(
 #define FA_V_STRATEGY 0
 #endif
 
+#ifndef MQ_GQA
+#define MQ_GQA 4
+#endif
+#ifndef MQ_NSG_SPLIT
+#define MQ_NSG_SPLIT 4
+#endif
+#define MQ_SPLIT_WG_SIZE_Q8 (Q1_WG_SIZE * MQ_NSG_SPLIT)
+
+REQD_SUBGROUP_SIZE_64
+__kernel void flash_attn_f32_q8_0_q1_vec_mq_split(
+    const global void * q_void, ulong q_offset,
+    const global void * k_void, ulong k_offset,
+    const global void * v_void, ulong v_offset,
+    const float scale,
+    const int n_q,
+    const int n_kv,
+    const int n_head,
+    const ulong q_nb1, const ulong q_nb2, const ulong q_nb3,
+    const ulong k_nb1, const ulong k_nb2, const ulong k_nb3,
+    const ulong v_nb1, const ulong v_nb2, const ulong v_nb3,
+    const float max_bias,
+    const float m0,
+    const float m1,
+    const int n_head_log2,
+    const float logit_softcap,
+    const int n_head_kv,
+    const global void * mask_void,
+    const ulong mask_offset,
+    const ulong mask_nb1,
+    const ulong mask_nb2,
+    const ulong mask_nb3,
+    const int mask_ne2,
+    const int mask_ne3,
+    global float * partial_void,
+    const int n_splits,
+    const int kv_per_split
+) {
+    const int tid              = get_local_id(0);
+    const int sgid             = tid / Q1_WG_SIZE;
+    const int tid_sg           = tid % Q1_WG_SIZE;
+    const int kvhead_batch_idx = get_global_id(1);
+    const int split_q_idx      = get_global_id(2);
+    const int split_idx        = split_q_idx % n_splits;
+    const int q_idx            = split_q_idx / n_splits;
+
+    const int batch_idx   = kvhead_batch_idx / n_head_kv;
+    const int head_kv_idx = kvhead_batch_idx % n_head_kv;
+
+    const int kv_start = split_idx * kv_per_split;
+    const int kv_end   = min(kv_start + kv_per_split, n_kv);
+
+    const ulong record_stride = (ulong) FA_PARTIAL_FLOATS;
+
+    if (kv_start >= kv_end) {
+        // Empty split — write sentinel for each of the MQ_GQA Q-heads.
+        if (tid == 0) {
+            #pragma unroll
+            for (int h = 0; h < MQ_GQA; ++h) {
+                const int head_idx = head_kv_idx * MQ_GQA + h;
+                const ulong rec_idx = ((((ulong) batch_idx * n_head + head_idx) * n_q + q_idx)
+                                       * n_splits + split_idx);
+                global float * rec = partial_void + rec_idx * record_stride;
+                rec[0] = FA_M_INIT;
+                rec[1] = 0.0f;
+            }
+        }
+        return;
+    }
+
+    const global char * q_base = (const global char *) q_void + q_offset;
+    const global char * k_base = (const global char *) k_void + k_offset;
+    const global char * v_base = (const global char *) v_void + v_offset;
+
+    __local ACC_TYPE4 q_shared[MQ_GQA * DK_VEC];
+    for (int i = tid; i < MQ_GQA * DK_VEC; i += MQ_SPLIT_WG_SIZE_Q8) {
+        const int h        = i / DK_VEC;
+        const int k        = i % DK_VEC;
+        const int head_idx = head_kv_idx * MQ_GQA + h;
+        const ulong q_row_offset = batch_idx * q_nb3 + head_idx * q_nb2 + (ulong) q_idx * q_nb1;
+        const global Q_DATA_TYPE4 * q_ptr = (const global Q_DATA_TYPE4 *) (q_base + q_row_offset);
+        q_shared[h * DK_VEC + k] = CONVERT_Q_ACC4(q_ptr[k]);
+    }
+    barrier(CLK_LOCAL_MEM_FENCE);
+
+    float slope[MQ_GQA];
+    #pragma unroll
+    for (int h = 0; h < MQ_GQA; ++h) {
+        slope[h] = get_alibi_slope(max_bias, head_kv_idx * MQ_GQA + h, n_head_log2, m0, m1);
+    }
+
+    const global char * mask_base[MQ_GQA];
+    if (mask_void != NULL) {
+        const int mask_batch_idx = batch_idx % mask_ne3;
+        const global char * mask_base_b = (const global char *) mask_void + mask_offset +
+                                          mask_batch_idx * mask_nb3 +
+                                          (ulong) q_idx * mask_nb1;
+        #pragma unroll
+        for (int h = 0; h < MQ_GQA; ++h) {
+            const int head_idx      = head_kv_idx * MQ_GQA + h;
+            const int mask_head_idx = head_idx % mask_ne2;
+            mask_base[h] = mask_base_b + mask_head_idx * mask_nb2;
+        }
+    } else {
+        #pragma unroll
+        for (int h = 0; h < MQ_GQA; ++h) mask_base[h] = NULL;
+    }
+
+    ACC_TYPE4 o_acc[MQ_GQA][Q1V_DV_PER_THREAD];
+    ACC_TYPE  m_i[MQ_GQA];
+    ACC_TYPE  l_i[MQ_GQA];
+    #pragma unroll
+    for (int h = 0; h < MQ_GQA; ++h) {
+        m_i[h] = FA_M_INIT;
+        l_i[h] = 0.0f;
+        #pragma unroll
+        for (int i = 0; i < Q1V_DV_PER_THREAD; ++i) o_acc[h][i] = (ACC_TYPE4)(0.0f);
+    }
+
+    const int kv_len    = kv_end - kv_start;
+    const int kv_per_sg = (kv_len + MQ_NSG_SPLIT - 1) / MQ_NSG_SPLIT;
+    const int kv_lo     = kv_start + sgid * kv_per_sg;
+    const int kv_hi     = min(kv_end, kv_lo + kv_per_sg);
+
+    for (int k_idx = kv_lo; k_idx < kv_hi; ++k_idx) {
+        const global char * k_row = k_base + batch_idx * k_nb3 + head_kv_idx * k_nb2 + k_idx * k_nb1;
+        const global char * v_row = v_base + batch_idx * v_nb3 + head_kv_idx * v_nb2 + k_idx * v_nb1;
+
+        ACC_TYPE4 dot4[MQ_GQA];
+        #pragma unroll
+        for (int h = 0; h < MQ_GQA; ++h) dot4[h] = (ACC_TYPE4)(0.0f);
+
+        for (int qk = tid_sg; qk < DK_VEC; qk += Q1_WG_SIZE) {
+            const int block_idx = qk / 8;
+            const int lane      = qk % 8;
+            const float4 k_v = dequant_q8_0_lane(k_row + block_idx * Q8_0_BLOCK_SIZE, lane);
+            #pragma unroll
+            for (int h = 0; h < MQ_GQA; ++h) {
+                dot4[h] = mad(q_shared[h * DK_VEC + qk], k_v, dot4[h]);
+            }
+        }
+
+        ACC_TYPE score[MQ_GQA];
+        #pragma unroll
+        for (int h = 0; h < MQ_GQA; ++h) {
+            const ACC_TYPE dot_partial = dot4[h].s0 + dot4[h].s1 + dot4[h].s2 + dot4[h].s3;
+            ACC_TYPE s = sub_group_reduce_add(dot_partial) * scale;
+            if (mask_base[h] != NULL) {
+                const global MASK_DATA_TYPE * mask_ptr = (const global MASK_DATA_TYPE *) mask_base[h];
+                s += slope[h] * (ACC_TYPE) mask_ptr[k_idx];
+            }
+            if (logit_softcap > 0.0f) {
+                s = logit_softcap * tanh(s / logit_softcap);
+            }
+            score[h] = s;
+        }
+
+        ACC_TYPE p_h[MQ_GQA];
+        ACC_TYPE sp_h[MQ_GQA];
+        #pragma unroll
+        for (int h = 0; h < MQ_GQA; ++h) {
+            const ACC_TYPE m_new = max(m_i[h], score[h]);
+            sp_h[h] = native_exp(m_i[h] - m_new);
+            p_h[h]  = native_exp(score[h] - m_new);
+            l_i[h]  = l_i[h] * sp_h[h] + p_h[h];
+            m_i[h]  = m_new;
+        }
+
+        int idx = 0;
+        for (int dv = tid_sg; dv < DV_VEC; dv += Q1_WG_SIZE, ++idx) {
+            const int block_idx = dv / 8;
+            const int lane      = dv % 8;
+            const float4 v_v = dequant_q8_0_lane(v_row + block_idx * Q8_0_BLOCK_SIZE, lane);
+            #pragma unroll
+            for (int h = 0; h < MQ_GQA; ++h) {
+                o_acc[h][idx] = mad(p_h[h], v_v, o_acc[h][idx] * sp_h[h]);
+            }
+        }
+    }
+
+    __local ACC_TYPE  sg_m[MQ_GQA][MQ_NSG_SPLIT];
+    __local ACC_TYPE  sg_l[MQ_GQA][MQ_NSG_SPLIT];
+    __local ACC_TYPE4 sg_o[MQ_NSG_SPLIT][DV_VEC];
+
+    if (tid_sg == 0) {
+        #pragma unroll
+        for (int h = 0; h < MQ_GQA; ++h) {
+            sg_m[h][sgid] = m_i[h];
+            sg_l[h][sgid] = l_i[h];
+        }
+    }
+
+    #pragma unroll
+    for (int h = 0; h < MQ_GQA; ++h) {
+        {
+            int idx = 0;
+            for (int dv_idx = tid_sg; dv_idx < DV_VEC; dv_idx += Q1_WG_SIZE, ++idx) {
+                sg_o[sgid][dv_idx] = o_acc[h][idx];
+            }
+        }
+        barrier(CLK_LOCAL_MEM_FENCE);
+
+        if (sgid == 0) {
+            const int head_idx = head_kv_idx * MQ_GQA + h;
+
+            ACC_TYPE m_c = sg_m[h][0];
+            #pragma unroll
+            for (int s = 1; s < MQ_NSG_SPLIT; ++s) {
+                m_c = max(m_c, sg_m[h][s]);
+            }
+            ACC_TYPE l_c = 0.0f;
+            #pragma unroll
+            for (int s = 0; s < MQ_NSG_SPLIT; ++s) {
+                l_c += sg_l[h][s] * native_exp(sg_m[h][s] - m_c);
+            }
+
+            const ulong rec_idx = ((((ulong) batch_idx * n_head + head_idx) * n_q + q_idx)
+                                   * n_splits + split_idx);
+            global float  * rec   = partial_void + rec_idx * record_stride;
+            global float4 * rec_o = (global float4 *) (rec + 2);
+
+            if (tid_sg == 0) {
+                rec[0] = (float) m_c;
+                rec[1] = (float) l_c;
+            }
+            for (int dv_idx = tid_sg; dv_idx < DV_VEC; dv_idx += Q1_WG_SIZE) {
+                ACC_TYPE4 o_merged = (ACC_TYPE4)(0.0f);
+                #pragma unroll
+                for (int s = 0; s < MQ_NSG_SPLIT; ++s) {
+                    const ACC_TYPE alpha = native_exp(sg_m[h][s] - m_c);
+                    o_merged = mad((ACC_TYPE4)(alpha), sg_o[s][dv_idx], o_merged);
+                }
+                rec_o[dv_idx] = o_merged;
+            }
+        }
+        barrier(CLK_LOCAL_MEM_FENCE);
+    }
+}
+
 __kernel void flash_attn_f32_q8_0(
     const global void * q_void, ulong q_offset,
     const global void * k_void, ulong k_offset,

ggml/src/ggml-opencl/kernels/mul_mv_f16_f32_l4.cl

@@ -18,6 +18,14 @@
 #define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
 #endif
 
+#ifdef cl_khr_subgroup_shuffle
+#pragma OPENCL EXTENSION cl_khr_subgroup_shuffle : enable
+#define HAS_SUBGROUP_SHUFFLE 1
+#elif defined(cl_qcom_subgroup_shuffle)
+#pragma OPENCL EXTENSION cl_qcom_subgroup_shuffle : enable
+#define HAS_SUBGROUP_SHUFFLE 1
+#endif
+
 // Assumes row size (ne00) is a multiple of 4
 #ifdef ADRENO_GPU
 REQD_SUBGROUP_SIZE_64
@@ -378,3 +386,848 @@ kernel void kernel_mul_mat_f16_f32_l4_dr_lq(
     }
 }
 #endif // ADRENO_GPU
+
+#define N_ROWS_PER_WG 8
+#define N_OUTS_PER_WG 8
+
+#ifdef ADRENO_GPU
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_mul_mat_f16_f32_l4_x8(
+        global char * src0,
+        ulong offset0,
+        global char * src1,
+        ulong offset1,
+        global float * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne02,
+        ulong nb00,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int ne10,
+        int ne11,
+        int ne12,
+        ulong nb10,
+        ulong nb11,
+        ulong nb12,
+        ulong nb13,
+        int ne0,
+        int ne1,
+        int r2,
+        int r3
+) {
+    src0 = (global char *)((global char *)src0 + offset0);
+    src1 = (global char *)((global char *)src1 + offset1);
+    dst  = (global float*)((global char *)dst  + offsetd);
+
+    const int sgs_lid = get_sub_group_local_id();
+    const int sgs_sz  = get_max_sub_group_size();
+
+    const int r0_base = get_group_id(0) * N_ROWS_PER_WG;
+    const int im      = get_group_id(2);
+
+    const int i12 = im % ne12;
+    const int i13 = im / ne12;
+
+    const ulong offset_src1 = (i12) * nb12 + (i13) * nb13;
+    global float4 * y4 = (global float4 *)(src1 + offset_src1);
+
+    __local float4 q_loc[64];   // ne00/4 max for sub_group_size 64
+    if (sgs_lid < ne00 / 4) {
+        q_loc[sgs_lid] = y4[sgs_lid];
+    }
+    barrier(CLK_LOCAL_MEM_FENCE);
+
+    #pragma unroll
+    for (int dr = 0; dr < N_ROWS_PER_WG; ++dr) {
+        const int r0 = r0_base + dr;
+        if (r0 >= ne01) return;
+
+        const ulong offset_src0 = r0 * nb01 + (i12 / r2) * nb02 + (i13 / r3) * nb03;
+        global half4 * x4 = (global half4 *)(src0 + offset_src0);
+
+        float sumf = 0.0f;
+        for (int i = sgs_lid; i < ne00 / 4; i += sgs_sz) {
+            const half4   k4 = x4[i];
+            const float4  q  = q_loc[i];
+            sumf += convert_float(k4.s0) * q.s0
+                  + convert_float(k4.s1) * q.s1
+                  + convert_float(k4.s2) * q.s2
+                  + convert_float(k4.s3) * q.s3;
+        }
+
+        const float all_sum = sub_group_reduce_add(sumf);
+        if (sgs_lid == 0) {
+            dst[im * ne1 * ne0 + r0] = all_sum;  // ne11 == 1, so r1==0
+        }
+    }
+}
+
+#ifdef ADRENO_GPU
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_mul_mat_f16_f32_l4_y8(
+        global char * src0,
+        ulong offset0,
+        global char * src1,
+        ulong offset1,
+        global float * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne02,
+        ulong nb00,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int ne10,
+        int ne11,
+        int ne12,
+        ulong nb10,
+        ulong nb11,
+        ulong nb12,
+        ulong nb13,
+        int ne0,
+        int ne1,
+        int r2,
+        int r3
+) {
+    src0 = (global char *)((global char *)src0 + offset0);
+    src1 = (global char *)((global char *)src1 + offset1);
+    dst  = (global float*)((global char *)dst  + offsetd);
+
+    const int sgs_lid = get_sub_group_local_id();
+    const int sgs_sz  = get_max_sub_group_size();
+
+    const int r0_base = get_group_id(0) * N_OUTS_PER_WG;
+    const int im      = get_group_id(2);
+
+    const int i12 = im % ne12;
+    const int i13 = im / ne12;
+
+    const ulong offset_src1 = (i12) * nb12 + (i13) * nb13;
+    global float4 * y4 = (global float4 *)(src1 + offset_src1);
+
+    global half4 * x4_o[N_OUTS_PER_WG];
+    #pragma unroll
+    for (int o = 0; o < N_OUTS_PER_WG; ++o) {
+        const int r0 = r0_base + o;
+        const int r0c = (r0 < ne01) ? r0 : 0;
+        const ulong off = r0c * nb01 + (i12 / r2) * nb02 + (i13 / r3) * nb03;
+        x4_o[o] = (global half4 *)(src0 + off);
+    }
+
+    float sum[N_OUTS_PER_WG] = { 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f };
+
+    for (int i = sgs_lid; i < ne00 / 4; i += sgs_sz) {
+        const float4 q4 = y4[i];
+        #pragma unroll
+        for (int o = 0; o < N_OUTS_PER_WG; ++o) {
+            const half4 v4 = x4_o[o][i];
+            sum[o] += convert_float(v4.s0) * q4.s0
+                    + convert_float(v4.s1) * q4.s1
+                    + convert_float(v4.s2) * q4.s2
+                    + convert_float(v4.s3) * q4.s3;
+        }
+    }
+
+    #pragma unroll
+    for (int o = 0; o < N_OUTS_PER_WG; ++o) {
+        const int r0 = r0_base + o;
+        const float s = sub_group_reduce_add(sum[o]);
+        if (sgs_lid == 0 && r0 < ne01) {
+            dst[im * ne1 * ne0 + r0] = s;
+        }
+    }
+}
+
+#define N_OUTS_PAIR  8
+#define N_PAIRS_PAIR (N_OUTS_PAIR / 2)
+
+#ifdef ADRENO_GPU
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_mul_mat_f16_f32_l4_x8_pair(
+        global char * src0,
+        ulong offset0,
+        global char * src1,
+        ulong offset1,
+        global float * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne02,
+        ulong nb00,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int ne10,
+        int ne11,
+        int ne12,
+        ulong nb10,
+        ulong nb11,
+        ulong nb12,
+        ulong nb13,
+        int ne0,
+        int ne1,
+        int r2,
+        int r3
+) {
+    src0 = (global char *)((global char *)src0 + offset0);
+    src1 = (global char *)((global char *)src1 + offset1);
+    dst  = (global float*)((global char *)dst  + offsetd);
+
+    const int sgs_lid = get_sub_group_local_id();
+    const int half_id = sgs_lid >> 5;     // 0 = lower half, 1 = upper half
+    const int lane_h  = sgs_lid & 31;     // lane 0..31 within half
+
+    const int r0_base = get_group_id(0) * N_OUTS_PAIR;
+    const int im      = get_group_id(2);
+
+    const int i12 = im % ne12;
+    const int i13 = im / ne12;
+
+    const ulong offset_src1 = (i12) * nb12 + (i13) * nb13;
+    global float4 * y4 = (global float4 *)(src1 + offset_src1);
+
+    __local float4 q_loc[64];   // ne00/4 max for sub_group_size 64
+    if (sgs_lid < ne00 / 4) {
+        q_loc[sgs_lid] = y4[sgs_lid];
+    }
+    barrier(CLK_LOCAL_MEM_FENCE);
+
+    const int dk_vec = ne00 / 4;
+
+    #pragma unroll
+    for (int p = 0; p < N_PAIRS_PAIR; ++p) {
+        const int r0 = r0_base + 2 * p + half_id;
+
+        const ulong offset_src0 = r0 * nb01 + (i12 / r2) * nb02 + (i13 / r3) * nb03;
+        global half4 * x4 = (global half4 *)(src0 + offset_src0);
+
+        float sumf = 0.0f;
+        for (int i = lane_h; i < dk_vec; i += 32) {
+            const half4  k4 = x4[i];
+            const float4 q  = q_loc[i];
+            sumf += convert_float(k4.s0) * q.s0
+                  + convert_float(k4.s1) * q.s1
+                  + convert_float(k4.s2) * q.s2
+                  + convert_float(k4.s3) * q.s3;
+        }
+
+        sumf += sub_group_shuffle_xor(sumf, 16);
+        sumf += sub_group_shuffle_xor(sumf, 8);
+        sumf += sub_group_shuffle_xor(sumf, 4);
+        sumf += sub_group_shuffle_xor(sumf, 2);
+        sumf += sub_group_shuffle_xor(sumf, 1);
+
+        if (lane_h == 0) {
+            dst[im * ne1 * ne0 + r0] = sumf;
+        }
+    }
+}
+
+#define N_K_ROWS_GQA   16
+#define GQA_RATIO_GQA  8
+#define LANES_PER_QH   8    // 64 / GQA_RATIO_GQA
+#define DK_VEC_GQA     32   // DK / 4 for DK=128
+
+#ifdef ADRENO_GPU
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_mul_mat_f16_f32_l4_x8_gqa4(
+        global char * src0,
+        ulong offset0,
+        global char * src1,
+        ulong offset1,
+        global float * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne02,
+        ulong nb00,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int ne10,
+        int ne11,
+        int ne12,
+        ulong nb10,
+        ulong nb11,
+        ulong nb12,
+        ulong nb13,
+        int ne0,
+        int ne1,
+        int r2,
+        int r3
+) {
+    src0 = (global char *)((global char *)src0 + offset0);
+    src1 = (global char *)((global char *)src1 + offset1);
+    dst  = (global float*)((global char *)dst  + offsetd);
+
+    const int sgs_lid = get_sub_group_local_id();
+    const int q_id    = sgs_lid >> 3;       // 0..7: which Q-head (8 per WG)
+    const int lane_q  = sgs_lid & 7;        // 0..7: lane within Q-head partition
+
+    const int r0_base = get_group_id(0) * N_K_ROWS_GQA;
+    const int im_kv   = get_group_id(2);
+
+    const int i02 = im_kv % ne02;           // K-head index (also K2 batch)
+    const int i03 = im_kv / ne02;           // n13 batch index
+
+    const int q_head_lo = i02 * GQA_RATIO_GQA;
+
+    __local float4 q_loc[GQA_RATIO_GQA * DK_VEC_GQA];   // 4 × 32 = 128 float4
+    #pragma unroll
+    for (int qh = 0; qh < GQA_RATIO_GQA; ++qh) {
+        const int qh_idx = q_head_lo + qh;
+        global float4 * y4 = (global float4 *)(src1 + qh_idx * nb12 + i03 * nb13);
+
+        if (sgs_lid < DK_VEC_GQA) {
+            q_loc[qh * DK_VEC_GQA + sgs_lid] = y4[sgs_lid];
+        }
+    }
+    barrier(CLK_LOCAL_MEM_FENCE);
+
+    // K base offset for this WG. All 8 K-rows × 4 Q-heads share this K-head.
+    const ulong offset_src0_base = (i02) * nb02 + (i03 / r3) * nb03;
+
+    #pragma unroll
+    for (int dr = 0; dr < N_K_ROWS_GQA; ++dr) {
+        const int r0 = r0_base + dr;
+
+        const ulong offset_src0 = r0 * nb01 + offset_src0_base;
+        global half4 * x4 = (global half4 *)(src0 + offset_src0);
+
+        float sumf = 0.0f;
+        #pragma unroll
+        for (int t = 0; t < 4; ++t) {
+            const int i = lane_q + t * LANES_PER_QH;   // 8, 16, 24-step
+            const half4  k4 = x4[i];
+            const float4 q  = q_loc[q_id * DK_VEC_GQA + i];
+            sumf += convert_float(k4.s0) * q.s0
+                  + convert_float(k4.s1) * q.s1
+                  + convert_float(k4.s2) * q.s2
+                  + convert_float(k4.s3) * q.s3;
+        }
+
+        sumf += sub_group_shuffle_xor(sumf, 4);
+        sumf += sub_group_shuffle_xor(sumf, 2);
+        sumf += sub_group_shuffle_xor(sumf, 1);
+
+        if (lane_q == 0) {
+            const int im_out = i03 * ne12 + (q_head_lo + q_id);
+            dst[im_out * ne1 * ne0 + r0] = sumf;
+        }
+    }
+}
+
+#define N_DV_ROWS_Y8GQA  8
+#define GQA_RATIO_Y8GQA  8
+
+#ifdef ADRENO_GPU
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_mul_mat_f16_f32_l4_y8_gqa(
+        global char * src0,
+        ulong offset0,
+        global char * src1,
+        ulong offset1,
+        global float * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne02,
+        ulong nb00,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int ne10,
+        int ne11,
+        int ne12,
+        ulong nb10,
+        ulong nb11,
+        ulong nb12,
+        ulong nb13,
+        int ne0,
+        int ne1,
+        int r2,
+        int r3
+) {
+    src0 = (global char *)((global char *)src0 + offset0);
+    src1 = (global char *)((global char *)src1 + offset1);
+    dst  = (global float*)((global char *)dst  + offsetd);
+
+    const int sgs_lid = get_sub_group_local_id();
+    const int sgs_sz  = get_max_sub_group_size();
+
+    const int r0_base = get_group_id(0) * N_DV_ROWS_Y8GQA;
+    const int im_kv   = get_group_id(2);
+
+    const int i02 = im_kv % ne02;           // K-head index
+    const int i03 = im_kv / ne02;           // n13 batch index
+
+    // GQA Q-heads sharing this K-head.
+    const int q_head_lo = i02 * GQA_RATIO_Y8GQA;
+
+    global float4 * y4_q[GQA_RATIO_Y8GQA];
+    #pragma unroll
+    for (int qh = 0; qh < GQA_RATIO_Y8GQA; ++qh) {
+        const int qh_idx = q_head_lo + qh;
+        y4_q[qh] = (global float4 *)(src1 + qh_idx * nb12 + i03 * nb13);
+    }
+
+    global half4 * x4_o[N_DV_ROWS_Y8GQA];
+    #pragma unroll
+    for (int o = 0; o < N_DV_ROWS_Y8GQA; ++o) {
+        const int r0 = r0_base + o;
+        const int r0c = (r0 < ne01) ? r0 : 0;
+        const ulong off = r0c * nb01 + (i02) * nb02 + (i03 / r3) * nb03;
+        x4_o[o] = (global half4 *)(src0 + off);
+    }
+
+    float sum[N_DV_ROWS_Y8GQA][GQA_RATIO_Y8GQA] = { {0.0f} };
+
+    for (int i = sgs_lid; i < ne00 / 4; i += sgs_sz) {
+        // load 8 V values (one per DV row), same K-head, K-pos = i.
+        half4 v[N_DV_ROWS_Y8GQA];
+        #pragma unroll
+        for (int o = 0; o < N_DV_ROWS_Y8GQA; ++o) {
+            v[o] = x4_o[o][i];
+        }
+
+        // load 8 softmax values (one per Q-head).
+        float4 q[GQA_RATIO_Y8GQA];
+        #pragma unroll
+        for (int qh = 0; qh < GQA_RATIO_Y8GQA; ++qh) {
+            q[qh] = y4_q[qh][i];
+        }
+
+        #pragma unroll
+        for (int o = 0; o < N_DV_ROWS_Y8GQA; ++o) {
+            const float4 vf = (float4)(convert_float(v[o].s0),
+                                       convert_float(v[o].s1),
+                                       convert_float(v[o].s2),
+                                       convert_float(v[o].s3));
+            #pragma unroll
+            for (int qh = 0; qh < GQA_RATIO_Y8GQA; ++qh) {
+                sum[o][qh] += vf.s0 * q[qh].s0
+                            + vf.s1 * q[qh].s1
+                            + vf.s2 * q[qh].s2
+                            + vf.s3 * q[qh].s3;
+            }
+        }
+    }
+
+    #pragma unroll
+    for (int o = 0; o < N_DV_ROWS_Y8GQA; ++o) {
+        const int r0 = r0_base + o;
+        #pragma unroll
+        for (int qh = 0; qh < GQA_RATIO_Y8GQA; ++qh) {
+            const float s = sub_group_reduce_add(sum[o][qh]);
+            if (sgs_lid == 0 && r0 < ne01) {
+                const int im_out = i03 * ne12 + (q_head_lo + qh);
+                dst[im_out * ne1 * ne0 + r0] = s;
+            }
+        }
+    }
+}
+
+#ifdef ADRENO_GPU
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_mul_mat_f16_f32_l4_x8_gqa4_img(
+        __read_only image1d_buffer_t src0_img,
+        global char * src1,
+        ulong offset1,
+        global float * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne02,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int ne10,
+        int ne11,
+        int ne12,
+        ulong nb10,
+        ulong nb11,
+        ulong nb12,
+        ulong nb13,
+        int ne0,
+        int ne1,
+        int r2,
+        int r3
+) {
+    src1 = (global char *)((global char *)src1 + offset1);
+    dst  = (global float*)((global char *)dst  + offsetd);
+
+    const int sgs_lid = get_sub_group_local_id();
+    const int q_id    = sgs_lid >> 3;       // 0..7: which Q-head (8 per WG)
+    const int lane_q  = sgs_lid & 7;        // 0..7: lane within Q-head partition
+
+    const int r0_base = get_group_id(0) * N_K_ROWS_GQA;
+    const int im_kv   = get_group_id(2);
+
+    const int i02 = im_kv % ne02;
+    const int i03 = im_kv / ne02;
+
+    const int q_head_lo = i02 * GQA_RATIO_GQA;
+
+    __local float4 q_loc[GQA_RATIO_GQA * DK_VEC_GQA];
+    #pragma unroll
+    for (int qh = 0; qh < GQA_RATIO_GQA; ++qh) {
+        const int qh_idx = q_head_lo + qh;
+        global float4 * y4 = (global float4 *)(src1 + qh_idx * nb12 + i03 * nb13);
+        if (sgs_lid < DK_VEC_GQA) {
+            q_loc[qh * DK_VEC_GQA + sgs_lid] = y4[sgs_lid];
+        }
+    }
+    barrier(CLK_LOCAL_MEM_FENCE);
+
+    const int pitch_px_row  = (int)(nb01 >> 4);
+    const int pitch_px_head = (int)(nb02 >> 4);
+    const int pitch_px_n13  = (int)(nb03 >> 4);
+
+    const int head_px_base = i02 * pitch_px_head + (i03 / r3) * pitch_px_n13;
+
+    #pragma unroll
+    for (int dr = 0; dr < N_K_ROWS_GQA; ++dr) {
+        const int r0 = r0_base + dr;
+        const int row_px_base = r0 * pitch_px_row + head_px_base;
+
+        float sumf = 0.0f;
+        #pragma unroll
+        for (int t = 0; t < 2; ++t) {
+            const int p = lane_q + t * LANES_PER_QH;          // pixel idx in row, 0..15
+            const half8 k8 = as_half8(read_imagef(src0_img, row_px_base + p));
+            const int   i0 = 2 * p;                            // first half4 idx
+            const float4 qa = q_loc[q_id * DK_VEC_GQA + i0    ];
+            const float4 qb = q_loc[q_id * DK_VEC_GQA + i0 + 1];
+            sumf += convert_float(k8.s0) * qa.s0
+                  + convert_float(k8.s1) * qa.s1
+                  + convert_float(k8.s2) * qa.s2
+                  + convert_float(k8.s3) * qa.s3
+                  + convert_float(k8.s4) * qb.s0
+                  + convert_float(k8.s5) * qb.s1
+                  + convert_float(k8.s6) * qb.s2
+                  + convert_float(k8.s7) * qb.s3;
+        }
+
+        sumf += sub_group_shuffle_xor(sumf, 4);
+        sumf += sub_group_shuffle_xor(sumf, 2);
+        sumf += sub_group_shuffle_xor(sumf, 1);
+
+        if (lane_q == 0) {
+            const int im_out = i03 * ne12 + (q_head_lo + q_id);
+            dst[im_out * ne1 * ne0 + r0] = sumf;
+        }
+    }
+}
+
+#ifdef ADRENO_GPU
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_mul_mat_f16_f32_l4_y8_gqa_img(
+        __read_only image1d_buffer_t src0_img,
+        global char * src1,
+        ulong offset1,
+        global float * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne02,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int ne10,
+        int ne11,
+        int ne12,
+        ulong nb10,
+        ulong nb11,
+        ulong nb12,
+        ulong nb13,
+        int ne0,
+        int ne1,
+        int r2,
+        int r3
+) {
+    src1 = (global char *)((global char *)src1 + offset1);
+    dst  = (global float*)((global char *)dst  + offsetd);
+
+    const int sgs_lid = get_sub_group_local_id();
+    const int sgs_sz  = get_max_sub_group_size();
+
+    const int r0_base = get_group_id(0) * N_DV_ROWS_Y8GQA;
+    const int im_kv   = get_group_id(2);
+
+    const int i02 = im_kv % ne02;
+    const int i03 = im_kv / ne02;
+
+    const int q_head_lo = i02 * GQA_RATIO_Y8GQA;
+
+    // Q (= softmax(KQ)) base pointers per Q-head
+    global float4 * y4_q[GQA_RATIO_Y8GQA];
+    #pragma unroll
+    for (int qh = 0; qh < GQA_RATIO_Y8GQA; ++qh) {
+        const int qh_idx = q_head_lo + qh;
+        y4_q[qh] = (global float4 *)(src1 + qh_idx * nb12 + i03 * nb13);
+    }
+
+    const int pitch_px_row  = (int)(nb01 >> 3);
+    const int pitch_px_head = (int)(nb02 >> 3);
+    const int pitch_px_n13  = (int)(nb03 >> 3);
+
+    const int head_px_base = i02 * pitch_px_head + (i03 / r3) * pitch_px_n13;
+
+    // per-DV-row pixel base
+    int row_px_base[N_DV_ROWS_Y8GQA];
+    #pragma unroll
+    for (int o = 0; o < N_DV_ROWS_Y8GQA; ++o) {
+        const int r0  = r0_base + o;
+        const int r0c = (r0 < ne01) ? r0 : 0;
+        row_px_base[o] = r0c * pitch_px_row + head_px_base;
+    }
+
+    float sum[N_DV_ROWS_Y8GQA][GQA_RATIO_Y8GQA] = { {0.0f} };
+
+    for (int i = sgs_lid; i < ne00 / 4; i += sgs_sz) {
+        half4 v[N_DV_ROWS_Y8GQA];
+
+        #pragma unroll
+        for (int o = 0; o < N_DV_ROWS_Y8GQA; ++o) {
+            v[o] = read_imageh(src0_img, row_px_base[o] + i);
+        }
+
+        float4 q[GQA_RATIO_Y8GQA];
+        #pragma unroll
+        for (int qh = 0; qh < GQA_RATIO_Y8GQA; ++qh) {
+            q[qh] = y4_q[qh][i];
+        }
+        // 64 mads.
+        #pragma unroll
+        for (int o = 0; o < N_DV_ROWS_Y8GQA; ++o) {
+            const float4 vf = (float4)(convert_float(v[o].s0),
+                                       convert_float(v[o].s1),
+                                       convert_float(v[o].s2),
+                                       convert_float(v[o].s3));
+            #pragma unroll
+            for (int qh = 0; qh < GQA_RATIO_Y8GQA; ++qh) {
+                sum[o][qh] += vf.s0 * q[qh].s0
+                            + vf.s1 * q[qh].s1
+                            + vf.s2 * q[qh].s2
+                            + vf.s3 * q[qh].s3;
+            }
+        }
+    }
+
+    #pragma unroll
+    for (int o = 0; o < N_DV_ROWS_Y8GQA; ++o) {
+        const int r0 = r0_base + o;
+        #pragma unroll
+        for (int qh = 0; qh < GQA_RATIO_Y8GQA; ++qh) {
+            const float s = sub_group_reduce_add(sum[o][qh]);
+            if (sgs_lid == 0 && r0 < ne01) {
+                const int im_out = i03 * ne12 + (q_head_lo + qh);
+                dst[im_out * ne1 * ne0 + r0] = s;
+            }
+        }
+    }
+}
+
+#define N_K_ROWS_GQA_R4   16
+#define GQA_RATIO_R4      4
+#define LANES_PER_QH_R4   16    // = 64 / GQA_RATIO_R4
+#define DK_VEC_R4         32    // DK / 4 for DK=128
+
+#ifdef ADRENO_GPU
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_mul_mat_f16_f32_l4_x8_gqa_r4_img(
+        __read_only image1d_buffer_t src0_img,
+        global char * src1,
+        ulong offset1,
+        global float * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne02,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int ne10,
+        int ne11,
+        int ne12,
+        ulong nb10,
+        ulong nb11,
+        ulong nb12,
+        ulong nb13,
+        int ne0,
+        int ne1,
+        int r2,
+        int r3
+) {
+    src1 = (global char *)((global char *)src1 + offset1);
+    dst  = (global float*)((global char *)dst  + offsetd);
+
+    const int sgs_lid = get_sub_group_local_id();
+    const int q_id    = sgs_lid >> 4;       // 0..3
+    const int lane_q  = sgs_lid & 15;       // 0..15
+
+    const int r0_base = get_group_id(0) * N_K_ROWS_GQA_R4;
+    const int im_kv   = get_group_id(2);
+
+    const int i02 = im_kv % ne02;
+    const int i03 = im_kv / ne02;
+
+    const int q_head_lo = i02 * GQA_RATIO_R4;
+
+    __local float4 q_loc[GQA_RATIO_R4 * DK_VEC_R4];
+    #pragma unroll
+    for (int qh = 0; qh < GQA_RATIO_R4; ++qh) {
+        const int qh_idx = q_head_lo + qh;
+        global float4 * y4 = (global float4 *)(src1 + qh_idx * nb12 + i03 * nb13);
+        if (sgs_lid < DK_VEC_R4) {
+            q_loc[qh * DK_VEC_R4 + sgs_lid] = y4[sgs_lid];
+        }
+    }
+    barrier(CLK_LOCAL_MEM_FENCE);
+
+    const int pitch_px_row  = (int)(nb01 >> 4);
+    const int pitch_px_head = (int)(nb02 >> 4);
+    const int pitch_px_n13  = (int)(nb03 >> 4);
+
+    const int head_px_base = i02 * pitch_px_head + (i03 / r3) * pitch_px_n13;
+
+    #pragma unroll
+    for (int dr = 0; dr < N_K_ROWS_GQA_R4; ++dr) {
+        const int r0 = r0_base + dr;
+        const int row_px_base = r0 * pitch_px_row + head_px_base;
+
+        const int p = lane_q;
+        const half8 k8 = as_half8(read_imagef(src0_img, row_px_base + p));
+        const int   i0 = 2 * p;
+        const float4 qa = q_loc[q_id * DK_VEC_R4 + i0    ];
+        const float4 qb = q_loc[q_id * DK_VEC_R4 + i0 + 1];
+
+        float sumf =
+              convert_float(k8.s0) * qa.s0
+            + convert_float(k8.s1) * qa.s1
+            + convert_float(k8.s2) * qa.s2
+            + convert_float(k8.s3) * qa.s3
+            + convert_float(k8.s4) * qb.s0
+            + convert_float(k8.s5) * qb.s1
+            + convert_float(k8.s6) * qb.s2
+            + convert_float(k8.s7) * qb.s3;
+
+        sumf += sub_group_shuffle_xor(sumf, 8);
+        sumf += sub_group_shuffle_xor(sumf, 4);
+        sumf += sub_group_shuffle_xor(sumf, 2);
+        sumf += sub_group_shuffle_xor(sumf, 1);
+
+        if (lane_q == 0) {
+            const int im_out = i03 * ne12 + (q_head_lo + q_id);
+            dst[im_out * ne1 * ne0 + r0] = sumf;
+        }
+    }
+}
+
+#define N_K_ROWS_GQA_R2_DK256   16
+#define GQA_RATIO_R2            2
+#define LANES_PER_QH_R2         32    // = 64 / GQA_RATIO_R2
+#define DK_VEC_DK256            64    // DK / 4 for DK=256
+
+#ifdef ADRENO_GPU
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_mul_mat_f16_f32_l4_x8_gqa_r2_dk256_img(
+        __read_only image1d_buffer_t src0_img,
+        global char * src1,
+        ulong offset1,
+        global float * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne02,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int ne10,
+        int ne11,
+        int ne12,
+        ulong nb10,
+        ulong nb11,
+        ulong nb12,
+        ulong nb13,
+        int ne0,
+        int ne1,
+        int r2,
+        int r3
+) {
+    src1 = (global char *)((global char *)src1 + offset1);
+    dst  = (global float*)((global char *)dst  + offsetd);
+
+    const int sgs_lid = get_sub_group_local_id();
+    const int q_id    = sgs_lid >> 5;       // 0..1
+    const int lane_q  = sgs_lid & 31;       // 0..31
+
+    const int r0_base = get_group_id(0) * N_K_ROWS_GQA_R2_DK256;
+    const int im_kv   = get_group_id(2);
+
+    const int i02 = im_kv % ne02;
+    const int i03 = im_kv / ne02;
+
+    const int q_head_lo = i02 * GQA_RATIO_R2;
+
+    __local float4 q_loc[GQA_RATIO_R2 * DK_VEC_DK256];
+    #pragma unroll
+    for (int qh = 0; qh < GQA_RATIO_R2; ++qh) {
+        const int qh_idx = q_head_lo + qh;
+        global float4 * y4 = (global float4 *)(src1 + qh_idx * nb12 + i03 * nb13);
+        q_loc[qh * DK_VEC_DK256 + sgs_lid] = y4[sgs_lid];
+    }
+    barrier(CLK_LOCAL_MEM_FENCE);
+
+    const int pitch_px_row  = (int)(nb01 >> 4);
+    const int pitch_px_head = (int)(nb02 >> 4);
+    const int pitch_px_n13  = (int)(nb03 >> 4);
+
+    const int head_px_base = i02 * pitch_px_head + (i03 / r3) * pitch_px_n13;
+
+    #pragma unroll
+    for (int dr = 0; dr < N_K_ROWS_GQA_R2_DK256; ++dr) {
+        const int r0 = r0_base + dr;
+        const int row_px_base = r0 * pitch_px_row + head_px_base;
+
+        const int p = lane_q;
+        const half8 k8 = as_half8(read_imagef(src0_img, row_px_base + p));
+        const int   i0 = 2 * p;
+        const float4 qa = q_loc[q_id * DK_VEC_DK256 + i0    ];
+        const float4 qb = q_loc[q_id * DK_VEC_DK256 + i0 + 1];
+
+        float sumf =
+              convert_float(k8.s0) * qa.s0
+            + convert_float(k8.s1) * qa.s1
+            + convert_float(k8.s2) * qa.s2
+            + convert_float(k8.s3) * qa.s3
+            + convert_float(k8.s4) * qb.s0
+            + convert_float(k8.s5) * qb.s1
+            + convert_float(k8.s6) * qb.s2
+            + convert_float(k8.s7) * qb.s3;
+
+        sumf += sub_group_shuffle_xor(sumf, 16);
+        sumf += sub_group_shuffle_xor(sumf, 8);
+        sumf += sub_group_shuffle_xor(sumf, 4);
+        sumf += sub_group_shuffle_xor(sumf, 2);
+        sumf += sub_group_shuffle_xor(sumf, 1);
+
+        if (lane_q == 0) {
+            const int im_out = i03 * ne12 + (q_head_lo + q_id);
+            dst[im_out * ne1 * ne0 + r0] = sumf;
+        }
+    }
+}

ggml/src/ggml-sycl/backend.hpp

@@ -14,6 +14,7 @@
 #define GGML_SYCL_BACKEND_HPP
 
 #include "binbcast.hpp"
+#include "col2im-1d.hpp"
 #include "common.hpp"
 #include "concat.hpp"
 #include "conv.hpp"

ggml/src/ggml-sycl/col2im-1d.cpp

@@ -0,0 +1,102 @@
+#include "col2im-1d.hpp"
+
+template <typename T>
+static void col2im_1d_sycl(
+        const T * col,
+        T * dst,
+        const int T_in,
+        const sycl::uint3 T_out_fd,
+        const int K,
+        const int K_OC,
+        const int32_t s0,
+        const int32_t p0,
+        const int total,
+        dpct::queue_ptr stream) {
+
+    const uint32_t block_size = SYCL_COL2IM_1D_BLOCK_SIZE;
+    const uint32_t num_blocks = (uint32_t) ((total + block_size - 1) / block_size);
+
+    stream->parallel_for(
+        sycl::nd_range<3>(
+            sycl::range<3>(1, 1, num_blocks * block_size),
+            sycl::range<3>(1, 1, block_size)),
+        [=](sycl::nd_item<3> item_ct1) {
+            const int idx = (int) item_ct1.get_global_id(2);
+            if (idx >= total) {
+                return;
+            }
+
+            const sycl::uint2 qr = fast_div_modulo((uint32_t) idx, T_out_fd);
+            const int oc    = (int) qr.x();
+            const int t_out = (int) qr.y();
+            const int t_abs = t_out + p0;
+
+            int t_in_min = (t_abs - K + s0) / s0;
+            if (t_in_min < 0) {
+                t_in_min = 0;
+            }
+            int t_in_max = t_abs / s0;
+            if (t_in_max >= T_in) {
+                t_in_max = T_in - 1;
+            }
+
+            float sum = 0.0f;
+            for (int t_in = t_in_min; t_in <= t_in_max; ++t_in) {
+                const int k = t_abs - t_in * s0;
+                sum += static_cast<float>(col[(oc * K + k) + t_in * K_OC]);
+            }
+
+            dst[idx] = static_cast<T>(sum);
+        });
+}
+
+void ggml_sycl_op_col2im_1d(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+
+    GGML_ASSERT(src0 != nullptr);
+    GGML_ASSERT(ggml_is_contiguous(src0));
+    GGML_ASSERT(src0->type == dst->type);
+
+    const int32_t s0 = ((const int32_t *) dst->op_params)[0];
+    const int32_t OC = ((const int32_t *) dst->op_params)[1];
+    const int32_t p0 = ((const int32_t *) dst->op_params)[2];
+
+    const int K_OC  = (int) src0->ne[0];
+    const int T_in  = (int) src0->ne[1];
+    const int K     = K_OC / OC;
+    const int T_out = (int) dst->ne[0];
+
+    GGML_ASSERT(OC > 0);
+    GGML_ASSERT(K_OC % OC == 0);
+
+    const sycl::uint3 T_out_fd = init_fastdiv_values((uint32_t) T_out);
+
+    const int total = T_out * OC;
+
+    dpct::queue_ptr stream = ctx.stream();
+
+    switch (src0->type) {
+        case GGML_TYPE_F32:
+            col2im_1d_sycl<float>(
+                (const float *) src0->data,
+                (float *) dst->data,
+                T_in, T_out_fd, K, K_OC, s0, p0, total, stream);
+            break;
+        case GGML_TYPE_F16:
+            col2im_1d_sycl<sycl::half>(
+                (const sycl::half *) src0->data,
+                (sycl::half *) dst->data,
+                T_in, T_out_fd, K, K_OC, s0, p0, total, stream);
+            break;
+#ifdef GGML_SYCL_HAS_BF16
+        case GGML_TYPE_BF16:
+            col2im_1d_sycl<sycl::ext::oneapi::bfloat16>(
+                (const sycl::ext::oneapi::bfloat16 *) src0->data,
+                (sycl::ext::oneapi::bfloat16 *) dst->data,
+                T_in, T_out_fd, K, K_OC, s0, p0, total, stream);
+            break;
+#endif
+        default:
+            GGML_ABORT("col2im_1d: unsupported type %d", src0->type);
+    }
+}

ggml/src/ggml-sycl/col2im-1d.hpp

@@ -0,0 +1,8 @@
+#ifndef GGML_SYCL_COL2IM_1D_HPP
+#define GGML_SYCL_COL2IM_1D_HPP
+
+#include "common.hpp"
+
+void ggml_sycl_op_col2im_1d(ggml_backend_sycl_context & ctx, ggml_tensor * dst);
+
+#endif // GGML_SYCL_COL2IM_1D_HPP

ggml/src/ggml-sycl/common.hpp

@@ -59,7 +59,7 @@ void ggml_sycl_host_free(void* ptr);
 
 
 extern int g_ggml_sycl_debug;
-extern int g_ggml_sycl_disable_optimize;
+extern int g_ggml_sycl_enable_optimize;
 extern int g_ggml_sycl_prioritize_dmmv;
 extern int g_ggml_sycl_enable_flash_attention;
 extern int g_ggml_sycl_dev2dev_memcpy;

ggml/src/ggml-sycl/cpy.cpp

@@ -1,6 +1,7 @@
 #include "cpy.hpp"
 
 #include <float.h>
+#include <vector>
 
 #include "dequantize.hpp"
 #include "ggml-sycl/common.hpp"
@@ -50,6 +51,57 @@ static void cpy_1_i32_i32(const char * cxi, char * cdsti) {
     *dsti = *xi;
 }
 
+static void cpy_1_f32_i32(const char * cxi, char * cdsti) {
+    const float * xi     = (const float *) cxi;
+    int32_t *     dsti   = (int32_t *) cdsti;
+
+    *dsti = (int32_t) *xi;
+}
+
+static void cpy_1_i32_f32(const char * cxi, char * cdsti) {
+    const int32_t * xi   = (const int32_t *) cxi;
+    float *         dsti = (float *) cdsti;
+
+    *dsti = (float) *xi;
+}
+
+#ifdef GGML_SYCL_HAS_BF16
+static void cpy_1_f32_bf16(const char * cxi, char * cdsti) {
+    const float * xi   = (const float *) cxi;
+    sycl::ext::oneapi::bfloat16 *  dsti = (sycl::ext::oneapi::bfloat16 *) cdsti;
+
+    *dsti = sycl::ext::oneapi::bfloat16(*xi);
+}
+
+static void cpy_1_bf16_f32(const char * cxi, char * cdsti) {
+    const sycl::ext::oneapi::bfloat16 * xi   = (const sycl::ext::oneapi::bfloat16 *) cxi;
+    float *            dsti = (float *) cdsti;
+
+    *dsti = static_cast<float>(*xi);
+}
+
+static void cpy_1_bf16_bf16(const char * cxi, char * cdsti) {
+    const sycl::ext::oneapi::bfloat16 * xi   = (const sycl::ext::oneapi::bfloat16 *) cxi;
+    sycl::ext::oneapi::bfloat16 *       dsti = (sycl::ext::oneapi::bfloat16 *) cdsti;
+
+    *dsti = *xi;
+}
+
+static void cpy_1_f16_bf16(const char * cxi, char * cdsti) {
+    const sycl::half * xi   = (const sycl::half *) cxi;
+    sycl::ext::oneapi::bfloat16 *  dsti = (sycl::ext::oneapi::bfloat16 *) cdsti;
+
+    *dsti = sycl::ext::oneapi::bfloat16(static_cast<float>(*xi));
+}
+
+static void cpy_1_bf16_f16(const char * cxi, char * cdsti) {
+    const sycl::ext::oneapi::bfloat16 * xi   = (const sycl::ext::oneapi::bfloat16 *) cxi;
+    sycl::half *            dsti = (sycl::half *) cdsti;
+
+    *dsti = sycl::half(static_cast<float>(*xi));
+}
+#endif
+
 template <cpy_kernel_t cpy_1>
 static void cpy_f32_f16(const char * cx, char * cdst, const int ne, const int ne00, const int ne01, const int ne02,
                         const int nb00, const int nb01, const int nb02, const int nb03, const int ne10, const int ne11,
@@ -247,6 +299,38 @@ static void ggml_cpy_f32_f16_sycl(const char * cx, char * cdst, const int ne, co
     }
 }
 
+static void ggml_cpy_f32_i32_sycl(const char * cx, char * cdst, const int ne, const int ne00, const int ne01,
+                                  const int ne02, const int nb00, const int nb01, const int nb02, const int nb03,
+                                  const int ne10, const int ne11, const int ne12, const int nb10, const int nb11,
+                                  const int nb12, const int nb13, queue_ptr stream) {
+    const int num_blocks = (ne + SYCL_CPY_BLOCK_SIZE - 1) / SYCL_CPY_BLOCK_SIZE;
+    {
+        stream->parallel_for(
+            sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE),
+                              sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)),
+            [=](sycl::nd_item<3> item_ct1) {
+                cpy_f32_f16<cpy_1_f32_i32>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12,
+                                           nb10, nb11, nb12, nb13, item_ct1);
+            });
+    }
+}
+
+static void ggml_cpy_i32_f32_sycl(const char * cx, char * cdst, const int ne, const int ne00, const int ne01,
+                                  const int ne02, const int nb00, const int nb01, const int nb02, const int nb03,
+                                  const int ne10, const int ne11, const int ne12, const int nb10, const int nb11,
+                                  const int nb12, const int nb13, queue_ptr stream) {
+    const int num_blocks = (ne + SYCL_CPY_BLOCK_SIZE - 1) / SYCL_CPY_BLOCK_SIZE;
+    {
+        stream->parallel_for(
+            sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE),
+                              sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)),
+            [=](sycl::nd_item<3> item_ct1) {
+                cpy_f32_f16<cpy_1_i32_f32>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12,
+                                           nb10, nb11, nb12, nb13, item_ct1);
+            });
+    }
+}
+
 static void ggml_cpy_f32_q8_0_sycl(const char * cx, char * cdst, const int ne, const int ne00, const int ne01,
                                    const int ne02, const int nb00, const int nb01, const int nb02, const int nb03,
                                    const int ne10, const int ne11, const int ne12, const int nb10, const int nb11,
@@ -376,6 +460,19 @@ static void ggml_cpy_q5_1_f32_sycl(const char * cx, char * cdst, const int ne, c
         });
 }
 
+static void ggml_cpy_mxfp4_f32_sycl(const char * cx, char * cdst, const int ne, const int ne00, const int ne01,
+                                    const int ne02, const int nb00, const int nb01, const int nb02, const int nb03,
+                                    const int ne10, const int ne11, const int ne12, const int nb10, const int nb11,
+                                    const int nb12, const int nb13, queue_ptr stream) {
+    const int num_blocks = ne;
+    stream->parallel_for(
+        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), sycl::range<3>(1, 1, 1)), [=](sycl::nd_item<3> item_ct1) {
+            cpy_q_f32<cpy_blck_q_f32<dequantize_mxfp4, QK_MXFP4>, QK_MXFP4>(cx, cdst, ne, ne00, ne01, ne02, nb00,
+                                                                             nb01, nb02, nb03, ne10, ne11, ne12,
+                                                                             nb10, nb11, nb12, nb13, item_ct1);
+        });
+}
+
 static void ggml_cpy_f32_iq4_nl_sycl(const char * cx, char * cdst, const int ne, const int ne00, const int ne01,
                                      const int ne02, const int nb00, const int nb01, const int nb02, const int nb03,
                                      const int ne10, const int ne11, const int ne12, const int nb10, const int nb11,
@@ -389,6 +486,269 @@ static void ggml_cpy_f32_iq4_nl_sycl(const char * cx, char * cdst, const int ne,
         });
 }
 
+static void cpy_blck_f16_q4_0(const char * cxi, char * cdsti) {
+    const sycl::half * xi = (const sycl::half *) cxi;
+    float xf[QK4_0];
+
+    for (int j = 0; j < QK4_0; ++j) {
+        xf[j] = (float) xi[j];
+    }
+
+    cpy_blck_f32_q4_0((const char *) xf, cdsti);
+}
+
+static void cpy_blck_f16_q4_1(const char * cxi, char * cdsti) {
+    const sycl::half * xi = (const sycl::half *) cxi;
+    float xf[QK4_1];
+
+    for (int j = 0; j < QK4_1; ++j) {
+        xf[j] = (float) xi[j];
+    }
+
+    cpy_blck_f32_q4_1((const char *) xf, cdsti);
+}
+
+static void cpy_blck_f16_q5_0(const char * cxi, char * cdsti) {
+    const sycl::half * xi = (const sycl::half *) cxi;
+    float xf[QK5_0];
+
+    for (int j = 0; j < QK5_0; ++j) {
+        xf[j] = (float) xi[j];
+    }
+
+    cpy_blck_f32_q5_0((const char *) xf, cdsti);
+}
+
+static void ggml_cpy_f16_q4_0_sycl(const char * cx, char * cdst, const int ne, const int ne00, const int ne01,
+                                   const int ne02, const int nb00, const int nb01, const int nb02, const int nb03,
+                                   const int ne10, const int ne11, const int ne12, const int nb10, const int nb11,
+                                   const int nb12, const int nb13, queue_ptr stream) {
+    GGML_ASSERT(ne % QK4_0 == 0);
+    const int num_blocks = ne / QK4_0;
+    stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), sycl::range<3>(1, 1, 1)),
+                         [=](sycl::nd_item<3> item_ct1) {
+                             cpy_f32_q<cpy_blck_f16_q4_0, QK4_0>(cx, cdst, ne, ne00, ne01, ne02,
+                                                                 nb00, nb01, nb02, nb03,
+                                                                 ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1);
+                         });
+}
+
+static void ggml_cpy_f16_q4_1_sycl(const char * cx, char * cdst, const int ne, const int ne00, const int ne01,
+                                   const int ne02, const int nb00, const int nb01, const int nb02, const int nb03,
+                                   const int ne10, const int ne11, const int ne12, const int nb10, const int nb11,
+                                   const int nb12, const int nb13, queue_ptr stream) {
+    GGML_ASSERT(ne % QK4_1 == 0);
+    const int num_blocks = ne / QK4_1;
+    stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), sycl::range<3>(1, 1, 1)),
+                         [=](sycl::nd_item<3> item_ct1) {
+                             cpy_f32_q<cpy_blck_f16_q4_1, QK4_1>(cx, cdst, ne, ne00, ne01, ne02,
+                                                                 nb00, nb01, nb02, nb03,
+                                                                 ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1);
+                         });
+}
+
+static void ggml_cpy_f16_q5_0_sycl(const char * cx, char * cdst, const int ne, const int ne00, const int ne01,
+                                   const int ne02, const int nb00, const int nb01, const int nb02, const int nb03,
+                                   const int ne10, const int ne11, const int ne12, const int nb10, const int nb11,
+                                   const int nb12, const int nb13, queue_ptr stream) {
+    GGML_ASSERT(ne % QK5_0 == 0);
+    const int num_blocks = ne / QK5_0;
+    stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), sycl::range<3>(1, 1, 1)),
+                         [=](sycl::nd_item<3> item_ct1) {
+                             cpy_f32_q<cpy_blck_f16_q5_0, QK5_0>(cx, cdst, ne, ne00, ne01, ne02,
+                                                                 nb00, nb01, nb02, nb03,
+                                                                 ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1);
+                         });
+}
+
+static bool ggml_sycl_is_quantized_type(enum ggml_type type) {
+    switch (type) {
+        case GGML_TYPE_Q1_0:
+        case GGML_TYPE_Q4_0:
+        case GGML_TYPE_Q4_1:
+        case GGML_TYPE_Q5_0:
+        case GGML_TYPE_Q5_1:
+        case GGML_TYPE_Q8_0:
+        case GGML_TYPE_MXFP4:
+        case GGML_TYPE_NVFP4:
+        case GGML_TYPE_Q2_K:
+        case GGML_TYPE_Q3_K:
+        case GGML_TYPE_Q4_K:
+        case GGML_TYPE_Q5_K:
+        case GGML_TYPE_Q6_K:
+        case GGML_TYPE_IQ2_XXS:
+        case GGML_TYPE_IQ2_XS:
+        case GGML_TYPE_IQ2_S:
+        case GGML_TYPE_IQ3_XXS:
+        case GGML_TYPE_IQ3_S:
+        case GGML_TYPE_IQ1_S:
+        case GGML_TYPE_IQ1_M:
+        case GGML_TYPE_IQ4_NL:
+        case GGML_TYPE_IQ4_XS:
+            return true;
+        default:
+            return false;
+    }
+}
+
+static bool ggml_sycl_can_quantize_rows_sycl(enum ggml_type type) {
+    switch (type) {
+        case GGML_TYPE_Q1_0:
+        case GGML_TYPE_Q4_0:
+        case GGML_TYPE_Q4_1:
+        case GGML_TYPE_Q5_0:
+        case GGML_TYPE_Q5_1:
+        case GGML_TYPE_Q8_0:
+        case GGML_TYPE_MXFP4:
+        case GGML_TYPE_NVFP4:
+        case GGML_TYPE_Q2_K:
+        case GGML_TYPE_Q3_K:
+        case GGML_TYPE_Q4_K:
+        case GGML_TYPE_Q5_K:
+        case GGML_TYPE_Q6_K:
+        case GGML_TYPE_IQ4_NL:
+        case GGML_TYPE_IQ4_XS:
+            return true;
+        default:
+            return false;
+    }
+}
+
+template <typename SrcScalar>
+static inline float ggml_sycl_src_to_f32(const SrcScalar & x) {
+    return (float) x;
+}
+
+#ifdef GGML_SYCL_HAS_BF16
+template <>
+inline float ggml_sycl_src_to_f32<sycl::ext::oneapi::bfloat16>(const sycl::ext::oneapi::bfloat16 & x) {
+    return static_cast<float>(x);
+}
+
+template <>
+inline float ggml_sycl_src_to_f32<ggml_bf16_t>(const ggml_bf16_t & x) {
+    union {
+        uint32_t u32;
+        float f32;
+    } value;
+
+    value.u32 = (uint32_t) x.bits << 16;
+    return value.f32;
+}
+#endif
+
+template <typename SrcScalar, cpy_kernel_t quantize_block, int qk>
+static void ggml_sycl_quantize_rows_q(const char * cx, char * cdst, const int64_t ne,
+                                      const int64_t ne00, const int64_t ne01, const int64_t ne02,
+                                      const size_t nb00, const size_t nb01, const size_t nb02, const size_t nb03,
+                                      const int64_t ne10, const int64_t ne11, const int64_t ne12,
+                                      const size_t nb10, const size_t nb11, const size_t nb12, const size_t nb13,
+                                      queue_ptr stream) {
+    GGML_ASSERT(ne % qk == 0);
+    GGML_ASSERT(ne00 % qk == 0);
+
+    const int64_t total_blocks = ne / qk;
+    constexpr int block_size = 256;
+    const int64_t grid_size = ceil_div(total_blocks, (int64_t) block_size);
+
+    stream->parallel_for(sycl::nd_range<1>(grid_size * block_size, block_size), [=](sycl::nd_item<1> item_ct1) {
+        const int64_t block_idx = item_ct1.get_global_linear_id();
+        if (block_idx >= total_blocks) {
+            return;
+        }
+
+        const int64_t i = block_idx * qk;
+
+        const int64_t i03 = i / (ne00 * ne01 * ne02);
+        const int64_t i02 = (i - i03 * ne00 * ne01 * ne02) / (ne00 * ne01);
+        const int64_t i01 = (i - i03 * ne00 * ne01 * ne02 - i02 * ne01 * ne00) / ne00;
+        const int64_t i00 = i - i03 * ne00 * ne01 * ne02 - i02 * ne01 * ne00 - i01 * ne00;
+        const size_t x_offset = i00 * nb00 + i01 * nb01 + i02 * nb02 + i03 * nb03;
+
+        const int64_t i13 = i / (ne10 * ne11 * ne12);
+        const int64_t i12 = (i - i13 * ne10 * ne11 * ne12) / (ne10 * ne11);
+        const int64_t i11 = (i - i13 * ne10 * ne11 * ne12 - i12 * ne10 * ne11) / ne10;
+        const int64_t i10 = i - i13 * ne10 * ne11 * ne12 - i12 * ne10 * ne11 - i11 * ne10;
+        const size_t dst_offset = (i10 / qk) * nb10 + i11 * nb11 + i12 * nb12 + i13 * nb13;
+
+        float xf[qk];
+        if (nb00 == sizeof(SrcScalar)) {
+            const SrcScalar * src_row = (const SrcScalar *) (cx + x_offset);
+            for (int j = 0; j < qk; ++j) {
+                xf[j] = ggml_sycl_src_to_f32(src_row[j]);
+            }
+        } else {
+            for (int j = 0; j < qk; ++j) {
+                const SrcScalar * src_val = (const SrcScalar *) (cx + x_offset + j * nb00);
+                xf[j] = ggml_sycl_src_to_f32(*src_val);
+            }
+        }
+
+        quantize_block((const char *) xf, cdst + dst_offset);
+    });
+}
+
+template <typename SrcScalar>
+static void ggml_sycl_quantize_rows_sycl(const char * cx, char * cdst, const ggml_tensor * src0, const ggml_tensor * src1,
+                                         const int64_t ne, const int64_t ne00, const int64_t ne01, const int64_t ne02,
+                                         const size_t nb00, const size_t nb01, const size_t nb02, const size_t nb03,
+                                         const int64_t ne10, const int64_t ne11, const int64_t ne12, const size_t nb10,
+                                         const size_t nb11, const size_t nb12, const size_t nb13, queue_ptr stream) {
+    GGML_UNUSED(src0);
+    GGML_UNUSED(src1);
+
+    switch (src1->type) {
+        case GGML_TYPE_Q8_0:
+            ggml_sycl_quantize_rows_q<SrcScalar, cpy_blck_f32_q8_0, QK8_0>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01,
+                                                                            nb02, nb03, ne10, ne11, ne12, nb10, nb11,
+                                                                            nb12, nb13, stream);
+            break;
+        case GGML_TYPE_Q1_0:
+            ggml_sycl_quantize_rows_q<SrcScalar, cpy_blck_f32_q1_0, QK1_0>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01,
+                                                                            nb02, nb03, ne10, ne11, ne12, nb10, nb11,
+                                                                            nb12, nb13, stream);
+            break;
+        case GGML_TYPE_Q5_1:
+            ggml_sycl_quantize_rows_q<SrcScalar, cpy_blck_f32_q5_1, QK5_1>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01,
+                                                                            nb02, nb03, ne10, ne11, ne12, nb10, nb11,
+                                                                            nb12, nb13, stream);
+            break;
+        case GGML_TYPE_Q5_0:
+            ggml_sycl_quantize_rows_q<SrcScalar, cpy_blck_f32_q5_0, QK5_0>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01,
+                                                                            nb02, nb03, ne10, ne11, ne12, nb10, nb11,
+                                                                            nb12, nb13, stream);
+            break;
+        case GGML_TYPE_Q4_1:
+            ggml_sycl_quantize_rows_q<SrcScalar, cpy_blck_f32_q4_1, QK4_1>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01,
+                                                                            nb02, nb03, ne10, ne11, ne12, nb10, nb11,
+                                                                            nb12, nb13, stream);
+            break;
+        case GGML_TYPE_Q4_0:
+            ggml_sycl_quantize_rows_q<SrcScalar, cpy_blck_f32_q4_0, QK4_0>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01,
+                                                                            nb02, nb03, ne10, ne11, ne12, nb10, nb11,
+                                                                            nb12, nb13, stream);
+            break;
+        case GGML_TYPE_IQ4_NL:
+            ggml_sycl_quantize_rows_q<SrcScalar, cpy_blck_f32_iq4_nl, QK4_NL>(cx, cdst, ne, ne00, ne01, ne02, nb00,
+                                                                               nb01, nb02, nb03, ne10, ne11, ne12,
+                                                                               nb10, nb11, nb12, nb13, stream);
+            break;
+        case GGML_TYPE_MXFP4:
+            ggml_sycl_quantize_rows_q<SrcScalar, cpy_blck_f32_mxfp4, QK_MXFP4>(cx, cdst, ne, ne00, ne01, ne02, nb00,
+                                                                                nb01, nb02, nb03, ne10, ne11, ne12,
+                                                                                nb10, nb11, nb12, nb13, stream);
+            break;
+        case GGML_TYPE_NVFP4:
+            ggml_sycl_quantize_rows_q<SrcScalar, cpy_blck_f32_nvfp4, QK_NVFP4>(cx, cdst, ne, ne00, ne01, ne02, nb00,
+                                                                                nb01, nb02, nb03, ne10, ne11, ne12,
+                                                                                nb10, nb11, nb12, nb13, stream);
+            break;
+        default:
+            GGML_ABORT("unsupported quantized target type in sycl quantizer src1->type=%s\n",
+                       ggml_type_name(src1->type));
+    }
+}
+
 static void ggml_cpy_f16_f16_sycl(const char * cx, char * cdst, const int ne, const int ne00, const int ne01,
                                   const int ne02, const int nb00, const int nb01, const int nb02, const int nb03,
                                   const int ne10, const int ne11, const int ne12, const int nb10, const int nb11,
@@ -509,8 +869,269 @@ static void ggml_cpy_q4_1_q4_1(const char * cx, char * cdst, const int ne, const
         });
 }
 
+static void ggml_cpy_q1_0_q1_0(const char * cx, char * cdst, const int ne, const int ne00, const int ne01,
+                                   const int ne02, const int nb00, const int nb01, const int nb02, const int nb03,
+                                   const int ne10, const int ne11, const int ne12, const int nb10, const int nb11,
+                                   const int nb12, const int nb13, queue_ptr stream) {
+    const int num_blocks = ceil_div(ne, SYCL_CPY_BLOCK_SIZE);
+    stream->parallel_for(
+        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE), sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)), [=](sycl::nd_item<3> item_ct1) {
+            cpy_q_q<block_q1_0, QK1_0>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1);
+        });
+}
+
+static void ggml_cpy_mxfp4_mxfp4(const char * cx, char * cdst, const int ne, const int ne00, const int ne01,
+                                   const int ne02, const int nb00, const int nb01, const int nb02, const int nb03,
+                                   const int ne10, const int ne11, const int ne12, const int nb10, const int nb11,
+                                   const int nb12, const int nb13, queue_ptr stream) {
+    const int num_blocks = ceil_div(ne, SYCL_CPY_BLOCK_SIZE);
+    stream->parallel_for(
+        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE), sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)), [=](sycl::nd_item<3> item_ct1) {
+            cpy_q_q<block_mxfp4, QK_MXFP4>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1);
+        });
+}
+
+static void ggml_cpy_nvfp4_nvfp4(const char * cx, char * cdst, const int ne, const int ne00, const int ne01,
+                                   const int ne02, const int nb00, const int nb01, const int nb02, const int nb03,
+                                   const int ne10, const int ne11, const int ne12, const int nb10, const int nb11,
+                                   const int nb12, const int nb13, queue_ptr stream) {
+    const int num_blocks = ceil_div(ne, SYCL_CPY_BLOCK_SIZE);
+    stream->parallel_for(
+        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE), sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)), [=](sycl::nd_item<3> item_ct1) {
+            cpy_q_q<block_nvfp4, QK_NVFP4>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1);
+        });
+}
+
+static void ggml_cpy_q2_K_q2_K(const char * cx, char * cdst, const int ne, const int ne00, const int ne01,
+                                   const int ne02, const int nb00, const int nb01, const int nb02, const int nb03,
+                                   const int ne10, const int ne11, const int ne12, const int nb10, const int nb11,
+                                   const int nb12, const int nb13, queue_ptr stream) {
+    const int num_blocks = ceil_div(ne, SYCL_CPY_BLOCK_SIZE);
+    stream->parallel_for(
+        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE), sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)), [=](sycl::nd_item<3> item_ct1) {
+            cpy_q_q<block_q2_K, QK_K>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1);
+        });
+}
+
+static void ggml_cpy_q3_K_q3_K(const char * cx, char * cdst, const int ne, const int ne00, const int ne01,
+                                   const int ne02, const int nb00, const int nb01, const int nb02, const int nb03,
+                                   const int ne10, const int ne11, const int ne12, const int nb10, const int nb11,
+                                   const int nb12, const int nb13, queue_ptr stream) {
+    const int num_blocks = ceil_div(ne, SYCL_CPY_BLOCK_SIZE);
+    stream->parallel_for(
+        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE), sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)), [=](sycl::nd_item<3> item_ct1) {
+            cpy_q_q<block_q3_K, QK_K>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1);
+        });
+}
+
+static void ggml_cpy_q4_K_q4_K(const char * cx, char * cdst, const int ne, const int ne00, const int ne01,
+                                   const int ne02, const int nb00, const int nb01, const int nb02, const int nb03,
+                                   const int ne10, const int ne11, const int ne12, const int nb10, const int nb11,
+                                   const int nb12, const int nb13, queue_ptr stream) {
+    const int num_blocks = ceil_div(ne, SYCL_CPY_BLOCK_SIZE);
+    stream->parallel_for(
+        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE), sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)), [=](sycl::nd_item<3> item_ct1) {
+            cpy_q_q<block_q4_K, QK_K>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1);
+        });
+}
+
+static void ggml_cpy_q5_K_q5_K(const char * cx, char * cdst, const int ne, const int ne00, const int ne01,
+                                   const int ne02, const int nb00, const int nb01, const int nb02, const int nb03,
+                                   const int ne10, const int ne11, const int ne12, const int nb10, const int nb11,
+                                   const int nb12, const int nb13, queue_ptr stream) {
+    const int num_blocks = ceil_div(ne, SYCL_CPY_BLOCK_SIZE);
+    stream->parallel_for(
+        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE), sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)), [=](sycl::nd_item<3> item_ct1) {
+            cpy_q_q<block_q5_K, QK_K>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1);
+        });
+}
+
+static void ggml_cpy_q6_K_q6_K(const char * cx, char * cdst, const int ne, const int ne00, const int ne01,
+                                   const int ne02, const int nb00, const int nb01, const int nb02, const int nb03,
+                                   const int ne10, const int ne11, const int ne12, const int nb10, const int nb11,
+                                   const int nb12, const int nb13, queue_ptr stream) {
+    const int num_blocks = ceil_div(ne, SYCL_CPY_BLOCK_SIZE);
+    stream->parallel_for(
+        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE), sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)), [=](sycl::nd_item<3> item_ct1) {
+            cpy_q_q<block_q6_K, QK_K>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1);
+        });
+}
+
+static void ggml_cpy_iq2_xxs_iq2_xxs(const char * cx, char * cdst, const int ne, const int ne00, const int ne01,
+                                   const int ne02, const int nb00, const int nb01, const int nb02, const int nb03,
+                                   const int ne10, const int ne11, const int ne12, const int nb10, const int nb11,
+                                   const int nb12, const int nb13, queue_ptr stream) {
+    const int num_blocks = ceil_div(ne, SYCL_CPY_BLOCK_SIZE);
+    stream->parallel_for(
+        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE), sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)), [=](sycl::nd_item<3> item_ct1) {
+            cpy_q_q<block_iq2_xxs, QK_K>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1);
+        });
+}
+
+static void ggml_cpy_iq2_xs_iq2_xs(const char * cx, char * cdst, const int ne, const int ne00, const int ne01,
+                                   const int ne02, const int nb00, const int nb01, const int nb02, const int nb03,
+                                   const int ne10, const int ne11, const int ne12, const int nb10, const int nb11,
+                                   const int nb12, const int nb13, queue_ptr stream) {
+    const int num_blocks = ceil_div(ne, SYCL_CPY_BLOCK_SIZE);
+    stream->parallel_for(
+        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE), sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)), [=](sycl::nd_item<3> item_ct1) {
+            cpy_q_q<block_iq2_xs, QK_K>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1);
+        });
+}
+
+static void ggml_cpy_iq2_s_iq2_s(const char * cx, char * cdst, const int ne, const int ne00, const int ne01,
+                                   const int ne02, const int nb00, const int nb01, const int nb02, const int nb03,
+                                   const int ne10, const int ne11, const int ne12, const int nb10, const int nb11,
+                                   const int nb12, const int nb13, queue_ptr stream) {
+    const int num_blocks = ceil_div(ne, SYCL_CPY_BLOCK_SIZE);
+    stream->parallel_for(
+        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE), sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)), [=](sycl::nd_item<3> item_ct1) {
+            cpy_q_q<block_iq2_s, QK_K>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1);
+        });
+}
+
+static void ggml_cpy_iq3_xxs_iq3_xxs(const char * cx, char * cdst, const int ne, const int ne00, const int ne01,
+                                   const int ne02, const int nb00, const int nb01, const int nb02, const int nb03,
+                                   const int ne10, const int ne11, const int ne12, const int nb10, const int nb11,
+                                   const int nb12, const int nb13, queue_ptr stream) {
+    const int num_blocks = ceil_div(ne, SYCL_CPY_BLOCK_SIZE);
+    stream->parallel_for(
+        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE), sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)), [=](sycl::nd_item<3> item_ct1) {
+            cpy_q_q<block_iq3_xxs, QK_K>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1);
+        });
+}
+
+static void ggml_cpy_iq1_s_iq1_s(const char * cx, char * cdst, const int ne, const int ne00, const int ne01,
+                                   const int ne02, const int nb00, const int nb01, const int nb02, const int nb03,
+                                   const int ne10, const int ne11, const int ne12, const int nb10, const int nb11,
+                                   const int nb12, const int nb13, queue_ptr stream) {
+    const int num_blocks = ceil_div(ne, SYCL_CPY_BLOCK_SIZE);
+    stream->parallel_for(
+        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE), sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)), [=](sycl::nd_item<3> item_ct1) {
+            cpy_q_q<block_iq1_s, QK_K>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1);
+        });
+}
+
+static void ggml_cpy_iq1_m_iq1_m(const char * cx, char * cdst, const int ne, const int ne00, const int ne01,
+                                   const int ne02, const int nb00, const int nb01, const int nb02, const int nb03,
+                                   const int ne10, const int ne11, const int ne12, const int nb10, const int nb11,
+                                   const int nb12, const int nb13, queue_ptr stream) {
+    const int num_blocks = ceil_div(ne, SYCL_CPY_BLOCK_SIZE);
+    stream->parallel_for(
+        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE), sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)), [=](sycl::nd_item<3> item_ct1) {
+            cpy_q_q<block_iq1_m, QK_K>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1);
+        });
+}
+
+static void ggml_cpy_iq4_nl_iq4_nl(const char * cx, char * cdst, const int ne, const int ne00, const int ne01,
+                                   const int ne02, const int nb00, const int nb01, const int nb02, const int nb03,
+                                   const int ne10, const int ne11, const int ne12, const int nb10, const int nb11,
+                                   const int nb12, const int nb13, queue_ptr stream) {
+    const int num_blocks = ceil_div(ne, SYCL_CPY_BLOCK_SIZE);
+    stream->parallel_for(
+        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE), sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)), [=](sycl::nd_item<3> item_ct1) {
+            cpy_q_q<block_iq4_nl, QK4_NL>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1);
+        });
+}
+
+static void ggml_cpy_iq3_s_iq3_s(const char * cx, char * cdst, const int ne, const int ne00, const int ne01,
+                                   const int ne02, const int nb00, const int nb01, const int nb02, const int nb03,
+                                   const int ne10, const int ne11, const int ne12, const int nb10, const int nb11,
+                                   const int nb12, const int nb13, queue_ptr stream) {
+    const int num_blocks = ceil_div(ne, SYCL_CPY_BLOCK_SIZE);
+    stream->parallel_for(
+        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE), sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)), [=](sycl::nd_item<3> item_ct1) {
+            cpy_q_q<block_iq3_s, QK_K>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1);
+        });
+}
+
+static void ggml_cpy_iq4_xs_iq4_xs(const char * cx, char * cdst, const int ne, const int ne00, const int ne01,
+                                   const int ne02, const int nb00, const int nb01, const int nb02, const int nb03,
+                                   const int ne10, const int ne11, const int ne12, const int nb10, const int nb11,
+                                   const int nb12, const int nb13, queue_ptr stream) {
+    const int num_blocks = ceil_div(ne, SYCL_CPY_BLOCK_SIZE);
+    stream->parallel_for(
+        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE), sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)), [=](sycl::nd_item<3> item_ct1) {
+            cpy_q_q<block_iq4_xs, QK_K>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1);
+        });
+}
+
+#ifdef GGML_SYCL_HAS_BF16
+static void ggml_cpy_f32_bf16_sycl(const char * cx, char * cdst, const int ne, const int ne00, const int ne01,
+                                  const int ne02, const int nb00, const int nb01, const int nb02, const int nb03,
+                                  const int ne10, const int ne11, const int ne12, const int nb10, const int nb11,
+                                  const int nb12, const int nb13, queue_ptr stream) {
+    const int num_blocks = (ne + SYCL_CPY_BLOCK_SIZE - 1) / SYCL_CPY_BLOCK_SIZE;
+    stream->parallel_for(
+        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE),
+                          sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)),
+        [=](sycl::nd_item<3> item_ct1) {
+            cpy_f32_f16<cpy_1_f32_bf16>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12,
+                                           nb10, nb11, nb12, nb13, item_ct1);
+        });
+}
+
+static void ggml_cpy_bf16_f32_sycl(const char * cx, char * cdst, const int ne, const int ne00, const int ne01,
+                                  const int ne02, const int nb00, const int nb01, const int nb02, const int nb03,
+                                  const int ne10, const int ne11, const int ne12, const int nb10, const int nb11,
+                                  const int nb12, const int nb13, queue_ptr stream) {
+    const int num_blocks = (ne + SYCL_CPY_BLOCK_SIZE - 1) / SYCL_CPY_BLOCK_SIZE;
+    stream->parallel_for(
+        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE),
+                          sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)),
+        [=](sycl::nd_item<3> item_ct1) {
+            cpy_f32_f16<cpy_1_bf16_f32>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12,
+                                           nb10, nb11, nb12, nb13, item_ct1);
+        });
+}
+
+static void ggml_cpy_bf16_bf16_sycl(const char * cx, char * cdst, const int ne, const int ne00, const int ne01,
+                                  const int ne02, const int nb00, const int nb01, const int nb02, const int nb03,
+                                  const int ne10, const int ne11, const int ne12, const int nb10, const int nb11,
+                                  const int nb12, const int nb13, queue_ptr stream) {
+    const int num_blocks = (ne + SYCL_CPY_BLOCK_SIZE - 1) / SYCL_CPY_BLOCK_SIZE;
+    stream->parallel_for(
+        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE),
+                          sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)),
+        [=](sycl::nd_item<3> item_ct1) {
+            cpy_f32_f16<cpy_1_bf16_bf16>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12,
+                                            nb10, nb11, nb12, nb13, item_ct1);
+        });
+}
+
+static void ggml_cpy_f16_bf16_sycl(const char * cx, char * cdst, const int ne, const int ne00, const int ne01,
+                                  const int ne02, const int nb00, const int nb01, const int nb02, const int nb03,
+                                  const int ne10, const int ne11, const int ne12, const int nb10, const int nb11,
+                                  const int nb12, const int nb13, queue_ptr stream) {
+    const int num_blocks = (ne + SYCL_CPY_BLOCK_SIZE - 1) / SYCL_CPY_BLOCK_SIZE;
+    stream->parallel_for(
+        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE),
+                          sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)),
+        [=](sycl::nd_item<3> item_ct1) {
+            cpy_f32_f16<cpy_1_f16_bf16>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12,
+                                           nb10, nb11, nb12, nb13, item_ct1);
+        });
+}
+
+static void ggml_cpy_bf16_f16_sycl(const char * cx, char * cdst, const int ne, const int ne00, const int ne01,
+                                  const int ne02, const int nb00, const int nb01, const int nb02, const int nb03,
+                                  const int ne10, const int ne11, const int ne12, const int nb10, const int nb11,
+                                  const int nb12, const int nb13, queue_ptr stream) {
+    const int num_blocks = (ne + SYCL_CPY_BLOCK_SIZE - 1) / SYCL_CPY_BLOCK_SIZE;
+    stream->parallel_for(
+        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE),
+                          sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)),
+        [=](sycl::nd_item<3> item_ct1) {
+            cpy_f32_f16<cpy_1_bf16_f16>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12,
+                                           nb10, nb11, nb12, nb13, item_ct1);
+        });
+}
+#endif
+
 void ggml_sycl_cpy(ggml_backend_sycl_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1) try {
     // Unlike other operators ggml_sycl_cpy takes 2 distinct tensors instead of a dst ggml_tensor and rely on its src field
+    GGML_SYCL_DEBUG("ggml_sycl_cpy: src0->type=%s, src1->type=%s\n",
+        ggml_type_name(src0->type), ggml_type_name(src1->type));
     scope_op_debug_print scope_dbg_print(__func__, src1, /*num_src=*/0, debug_get_tensor_str("\tsrc0", src0));
     const int64_t ne = ggml_nelements(src0);
     GGML_ASSERT(ne == ggml_nelements(src1));
@@ -525,12 +1146,31 @@ void ggml_sycl_cpy(ggml_backend_sycl_context & ctx, const ggml_tensor * src0, co
     if ((src0->type == src1->type) && (ggml_is_contiguous(src0) && ggml_is_contiguous(src1))) {
         GGML_SYCL_DEBUG("%s: memcpy path\n", __func__);
         main_stream->memcpy(src1_ddc, src0_ddc, ggml_nbytes(src0));
+    } else if (src0->type == GGML_TYPE_F32 && ggml_sycl_is_quantized_type(src1->type)) {
+        GGML_ASSERT(ggml_sycl_can_quantize_rows_sycl(src1->type));
+        ggml_sycl_quantize_rows_sycl<float>(src0_ddc, src1_ddc, src0, src1, ne, ne00, ne01, ne02, nb00, nb01,
+                                            nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+    } else if (src0->type == GGML_TYPE_F16 && ggml_sycl_is_quantized_type(src1->type)) {
+        GGML_ASSERT(ggml_sycl_can_quantize_rows_sycl(src1->type));
+        ggml_sycl_quantize_rows_sycl<sycl::half>(src0_ddc, src1_ddc, src0, src1, ne, ne00, ne01, ne02, nb00,
+                                                 nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13,
+                                                 main_stream);
+#ifdef GGML_SYCL_HAS_BF16
+    } else if (src0->type == GGML_TYPE_BF16 && ggml_sycl_is_quantized_type(src1->type)) {
+        GGML_ASSERT(ggml_sycl_can_quantize_rows_sycl(src1->type));
+        ggml_sycl_quantize_rows_sycl<ggml_bf16_t>(src0_ddc, src1_ddc, src0, src1, ne, ne00, ne01, ne02,
+                                                  nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11,
+                                                  nb12, nb13, main_stream);
+#endif
     } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32) {
         ggml_cpy_f32_f32_sycl(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10,
                               nb11, nb12, nb13, main_stream);
     } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F16) {
         ggml_cpy_f32_f16_sycl(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10,
                               nb11, nb12, nb13, main_stream);
+    } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_I32) {
+        ggml_cpy_f32_i32_sycl(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10,
+                              nb11, nb12, nb13, main_stream);
     } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q8_0) {
         ggml_cpy_f32_q8_0_sycl(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10,
                                nb11, nb12, nb13, main_stream);
@@ -546,12 +1186,24 @@ void ggml_sycl_cpy(ggml_backend_sycl_context & ctx, const ggml_tensor * src0, co
     } else if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F16) {
         ggml_cpy_f16_f16_sycl(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10,
                               nb11, nb12, nb13, main_stream);
+    } else if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_Q4_0) {
+        ggml_cpy_f16_q4_0_sycl(src0_ddc, src1_ddc, ne, ne00, ne01, ne02,
+                               nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+    } else if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_Q4_1) {
+        ggml_cpy_f16_q4_1_sycl(src0_ddc, src1_ddc, ne, ne00, ne01, ne02,
+                               nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+    } else if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_Q5_0) {
+        ggml_cpy_f16_q5_0_sycl(src0_ddc, src1_ddc, ne, ne00, ne01, ne02,
+                               nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
     } else if (src0->type == GGML_TYPE_I16 && src1->type == GGML_TYPE_I16) {
         ggml_cpy_i16_i16_sycl(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10,
                               nb11, nb12, nb13, main_stream);
     } else if (src0->type == GGML_TYPE_I32 && src1->type == GGML_TYPE_I32) {
         ggml_cpy_i32_i32_sycl(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10,
                               nb11, nb12, nb13, main_stream);
+    } else if (src0->type == GGML_TYPE_I32 && src1->type == GGML_TYPE_F32) {
+        ggml_cpy_i32_f32_sycl(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10,
+                              nb11, nb12, nb13, main_stream);
     } else if (src0->type == GGML_TYPE_Q4_0 && src1->type == GGML_TYPE_F32) {
         ggml_cpy_q4_0_f32_sycl(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10,
                                nb11, nb12, nb13, main_stream);
@@ -573,6 +1225,9 @@ void ggml_sycl_cpy(ggml_backend_sycl_context & ctx, const ggml_tensor * src0, co
     } else if (src0->type == GGML_TYPE_Q5_1 && src1->type == GGML_TYPE_F32) {
         ggml_cpy_q5_1_f32_sycl(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10,
                                nb11, nb12, nb13, main_stream);
+    } else if (src0->type == GGML_TYPE_MXFP4 && src1->type == GGML_TYPE_F32) {
+        ggml_cpy_mxfp4_f32_sycl(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12,
+                                nb10, nb11, nb12, nb13, main_stream);
     } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_IQ4_NL) {
         ggml_cpy_f32_iq4_nl_sycl(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12,
                                  nb10, nb11, nb12, nb13, main_stream);
@@ -586,6 +1241,57 @@ void ggml_sycl_cpy(ggml_backend_sycl_context & ctx, const ggml_tensor * src0, co
         ggml_cpy_q4_0_q4_0(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
     } else if (src0->type == GGML_TYPE_Q4_1 && src1->type == GGML_TYPE_Q4_1) {
         ggml_cpy_q4_1_q4_1(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+    } else if (src0->type == GGML_TYPE_Q1_0 && src1->type == GGML_TYPE_Q1_0) {
+        ggml_cpy_q1_0_q1_0(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+    } else if (src0->type == GGML_TYPE_MXFP4 && src1->type == GGML_TYPE_MXFP4) {
+        ggml_cpy_mxfp4_mxfp4(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+    } else if (src0->type == GGML_TYPE_NVFP4 && src1->type == GGML_TYPE_NVFP4) {
+        ggml_cpy_nvfp4_nvfp4(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+    } else if (src0->type == GGML_TYPE_Q2_K && src1->type == GGML_TYPE_Q2_K) {
+        ggml_cpy_q2_K_q2_K(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+    } else if (src0->type == GGML_TYPE_Q3_K && src1->type == GGML_TYPE_Q3_K) {
+        ggml_cpy_q3_K_q3_K(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+    } else if (src0->type == GGML_TYPE_Q4_K && src1->type == GGML_TYPE_Q4_K) {
+        ggml_cpy_q4_K_q4_K(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+    } else if (src0->type == GGML_TYPE_Q5_K && src1->type == GGML_TYPE_Q5_K) {
+        ggml_cpy_q5_K_q5_K(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+    } else if (src0->type == GGML_TYPE_Q6_K && src1->type == GGML_TYPE_Q6_K) {
+        ggml_cpy_q6_K_q6_K(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+    } else if (src0->type == GGML_TYPE_IQ2_XXS && src1->type == GGML_TYPE_IQ2_XXS) {
+        ggml_cpy_iq2_xxs_iq2_xxs(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+    } else if (src0->type == GGML_TYPE_IQ2_XS && src1->type == GGML_TYPE_IQ2_XS) {
+        ggml_cpy_iq2_xs_iq2_xs(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+    } else if (src0->type == GGML_TYPE_IQ2_S && src1->type == GGML_TYPE_IQ2_S) {
+        ggml_cpy_iq2_s_iq2_s(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+    } else if (src0->type == GGML_TYPE_IQ3_XXS && src1->type == GGML_TYPE_IQ3_XXS) {
+        ggml_cpy_iq3_xxs_iq3_xxs(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+    } else if (src0->type == GGML_TYPE_IQ1_S && src1->type == GGML_TYPE_IQ1_S) {
+        ggml_cpy_iq1_s_iq1_s(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+    } else if (src0->type == GGML_TYPE_IQ1_M && src1->type == GGML_TYPE_IQ1_M) {
+        ggml_cpy_iq1_m_iq1_m(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+    } else if (src0->type == GGML_TYPE_IQ4_NL && src1->type == GGML_TYPE_IQ4_NL) {
+        ggml_cpy_iq4_nl_iq4_nl(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+    } else if (src0->type == GGML_TYPE_IQ3_S && src1->type == GGML_TYPE_IQ3_S) {
+        ggml_cpy_iq3_s_iq3_s(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+    } else if (src0->type == GGML_TYPE_IQ4_XS && src1->type == GGML_TYPE_IQ4_XS) {
+        ggml_cpy_iq4_xs_iq4_xs(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+#ifdef GGML_SYCL_HAS_BF16
+    } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_BF16) {
+        ggml_cpy_f32_bf16_sycl(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10,
+                              nb11, nb12, nb13, main_stream);
+    } else if (src0->type == GGML_TYPE_BF16 && src1->type == GGML_TYPE_F32) {
+        ggml_cpy_bf16_f32_sycl(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10,
+                              nb11, nb12, nb13, main_stream);
+    } else if (src0->type == GGML_TYPE_BF16 && src1->type == GGML_TYPE_BF16) {
+        ggml_cpy_bf16_bf16_sycl(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10,
+                               nb11, nb12, nb13, main_stream);
+    } else if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_BF16) {
+        ggml_cpy_f16_bf16_sycl(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10,
+                              nb11, nb12, nb13, main_stream);
+    } else if (src0->type == GGML_TYPE_BF16 && src1->type == GGML_TYPE_F16) {
+        ggml_cpy_bf16_f16_sycl(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10,
+                              nb11, nb12, nb13, main_stream);
+#endif
     } else {
         GGML_LOG_ERROR("%s: unsupported type combination (%s to %s)\n", __func__, ggml_type_name(src0->type),
                        ggml_type_name(src1->type));

ggml/src/ggml-sycl/cpy.hpp

@@ -317,7 +317,7 @@ inline void cpy_blck_f32_nvfp4(const char * cxi, char * cdsti) {
 
         const uint8_t ue = ggml_fp32_to_ue4m3(amax / 6.0f);
         dsti->d[s] = ue;
-        const float d = ggml_ue4m3_to_fp32(ue);
+        const float d = ggml_sycl_ue4m3_to_fp32(ue);
 
         for (int j = 0; j < QK_NVFP4_SUB / 2; ++j) {
             const uint8_t x0 = best_index_mxfp4(xb[0 + j], d);

ggml/src/ggml-sycl/cross_entropy_loss.cpp

@@ -0,0 +1,255 @@
+#include "cross_entropy_loss.hpp"
+
+#include <cstdint>
+#include <cmath>
+
+template <bool has_shared>
+static __dpct_inline__ void cross_entropy_loss_f32_kernel(
+        const float * __restrict__ logits,
+        const float * __restrict__ labels,
+        float * __restrict__ row_loss,
+        const int nclasses,
+        const int nrows,
+        float * __restrict__ smem,
+        const sycl::nd_item<3> & item) {
+
+    const int row = item.get_group(2);
+    const int tid = item.get_local_id(2);
+
+    logits += (int64_t) row * nclasses;
+    labels += (int64_t) row * nclasses;
+
+    float max_logit = -INFINITY;
+    for (int i = tid; i < nclasses; i += WARP_SIZE) {
+        const float v = logits[i];
+        max_logit = sycl::fmax(max_logit, v);
+        if (has_shared) {
+            smem[i] = v;
+        }
+    }
+    max_logit = warp_reduce_max<WARP_SIZE>(max_logit);
+
+    float sum_exp = 0.0f;
+    for (int i = tid; i < nclasses; i += WARP_SIZE) {
+        const float v = has_shared ? smem[i] : logits[i];
+        sum_exp += sycl::exp(v - max_logit);
+    }
+    sum_exp = warp_reduce_sum<WARP_SIZE>(sum_exp);
+    const float log_sum = sycl::log(sum_exp);
+
+    float loss = 0.0f;
+    for (int i = tid; i < nclasses; i += WARP_SIZE) {
+        const float v = has_shared ? smem[i] : logits[i];
+        loss += (v - max_logit - log_sum) * labels[i];
+    }
+    loss = -warp_reduce_sum<WARP_SIZE>(loss) / (float) nrows;
+
+    if (tid == 0) {
+        row_loss[row] = loss;
+    }
+}
+
+template <bool has_shared>
+static __dpct_inline__ void cross_entropy_loss_back_f32_kernel(
+        const float * __restrict__ grad,
+        const float * __restrict__ logits,
+        const float * __restrict__ labels,
+        float * __restrict__ dst,
+        const int nclasses,
+        const int nrows,
+        float * __restrict__ smem,
+        const sycl::nd_item<3> & item) {
+
+    const int row = item.get_group(2);
+    const int tid = item.get_local_id(2);
+
+    logits += (int64_t) row * nclasses;
+    labels += (int64_t) row * nclasses;
+    dst    += (int64_t) row * nclasses;
+
+    float max_logit = -INFINITY;
+    for (int i = tid; i < nclasses; i += WARP_SIZE) {
+        const float v = logits[i];
+        max_logit = sycl::fmax(max_logit, v);
+        if (has_shared) {
+            smem[i] = v;
+        }
+    }
+    max_logit = warp_reduce_max<WARP_SIZE>(max_logit);
+
+    float sum_exp = 0.0f;
+    for (int i = tid; i < nclasses; i += WARP_SIZE) {
+        const float v = sycl::exp((has_shared ? smem[i] : logits[i]) - max_logit);
+        sum_exp += v;
+        if (has_shared) {
+            smem[i] = v;
+        } else {
+            dst[i] = v;
+        }
+    }
+    sum_exp = warp_reduce_sum<WARP_SIZE>(sum_exp);
+    const float inv_sum = 1.0f / sum_exp;
+
+    const float d_by_nrows = grad[0] / (float) nrows;
+    for (int i = tid; i < nclasses; i += WARP_SIZE) {
+        const float sm_num = has_shared ? smem[i] : dst[i];
+        dst[i] = (sm_num * inv_sum - labels[i]) * d_by_nrows;
+    }
+}
+
+static void cross_entropy_reduce_rows(
+        ggml_backend_sycl_context & ctx,
+        const float * row_loss,
+        float * dst,
+        const int64_t nrows) {
+    if (nrows == 1) {
+        SYCL_CHECK(CHECK_TRY_ERROR(
+            ctx.stream()->memcpy(dst, row_loss, sizeof(float))));
+        return;
+    }
+
+    ggml_sycl_pool_alloc<float> tmp_alloc(ctx.pool(), nrows);
+    float * tmp = tmp_alloc.get();
+    SYCL_CHECK(CHECK_TRY_ERROR(
+        ctx.stream()->memcpy(tmp, row_loss, nrows * sizeof(float))));
+
+    int64_t cur = nrows;
+    while (cur > 1) {
+        const int64_t out = (cur + WARP_SIZE - 1) / WARP_SIZE;
+        const sycl::range<3> block(1, 1, WARP_SIZE);
+        const sycl::range<3> grid(1, 1, out);
+        ctx.stream()->parallel_for(
+            sycl::nd_range<3>(grid * block, block),
+            [=](sycl::nd_item<3> item) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                const int row = item.get_group(2);
+                const int tid = item.get_local_id(2);
+                const int64_t i = (int64_t) row * WARP_SIZE + tid;
+                float v = i < cur ? tmp[i] : 0.0f;
+                v = warp_reduce_sum<WARP_SIZE>(v);
+                if (tid == 0) {
+                    tmp[row] = v;
+                }
+            });
+        cur = out;
+    }
+
+    SYCL_CHECK(CHECK_TRY_ERROR(
+        ctx.stream()->memcpy(dst, tmp, sizeof(float))));
+}
+
+void ggml_sycl_cross_entropy_loss(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/2);
+
+    const ggml_tensor * src0 = dst->src[0];
+    const ggml_tensor * src1 = dst->src[1];
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT(src1->type == GGML_TYPE_F32);
+    GGML_ASSERT(dst->type  == GGML_TYPE_F32);
+    GGML_ASSERT(ggml_is_contiguous(src0));
+    GGML_ASSERT(ggml_is_contiguous(src1));
+    GGML_ASSERT(ggml_is_contiguous(dst));
+    GGML_ASSERT(ggml_are_same_shape(src0, src1));
+    GGML_ASSERT(ggml_is_scalar(dst));
+
+    SYCL_CHECK(ggml_sycl_set_device(ctx.device));
+
+    const int64_t nclasses = src0->ne[0];
+    const int64_t nrows = ggml_nrows(src0);
+
+    const float * logits_d = (const float *) src0->data;
+    const float * labels_d = (const float *) src1->data;
+    float * dst_d = (float *) dst->data;
+
+    ggml_sycl_pool_alloc<float> row_loss_alloc(ctx.pool(), nrows);
+    float * row_loss = row_loss_alloc.get();
+
+    const sycl::range<3> block(1, 1, WARP_SIZE);
+    const sycl::range<3> grid(1, 1, nrows);
+    const size_t nbytes_shared = (size_t) nclasses * sizeof(float);
+    const size_t smpbo = ggml_sycl_info().devices[ctx.device].smpbo;
+
+    if (nbytes_shared <= smpbo) {
+        ctx.stream()->submit([&](sycl::handler & cgh) {
+            sycl::local_accessor<float, 1> smem(sycl::range<1>(nclasses), cgh);
+            cgh.parallel_for(
+                sycl::nd_range<3>(grid * block, block),
+                [=](sycl::nd_item<3> item) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                    cross_entropy_loss_f32_kernel<true>(
+                        logits_d, labels_d, row_loss,
+                        (int) nclasses, (int) nrows,
+                        get_pointer(smem), item);
+                });
+        });
+    } else {
+        ctx.stream()->parallel_for(
+            sycl::nd_range<3>(grid * block, block),
+            [=](sycl::nd_item<3> item) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                cross_entropy_loss_f32_kernel<false>(
+                    logits_d, labels_d, row_loss,
+                    (int) nclasses, (int) nrows,
+                    nullptr, item);
+            });
+    }
+
+    cross_entropy_reduce_rows(ctx, row_loss, dst_d, nrows);
+}
+
+void ggml_sycl_cross_entropy_loss_back(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/3);
+
+    const ggml_tensor * grad  = dst->src[0];
+    const ggml_tensor * src0f = dst->src[1];
+    const ggml_tensor * src1f = dst->src[2];
+
+    GGML_ASSERT(grad->type == GGML_TYPE_F32);
+    GGML_ASSERT(src0f->type == GGML_TYPE_F32);
+    GGML_ASSERT(src1f->type == GGML_TYPE_F32);
+    GGML_ASSERT(dst->type  == GGML_TYPE_F32);
+
+    GGML_ASSERT(ggml_is_scalar(grad));
+    GGML_ASSERT(ggml_is_contiguous(grad));
+    GGML_ASSERT(ggml_is_contiguous(src0f));
+    GGML_ASSERT(ggml_is_contiguous(src1f));
+    GGML_ASSERT(ggml_is_contiguous(dst));
+    GGML_ASSERT(ggml_are_same_shape(src0f, src1f));
+    GGML_ASSERT(ggml_are_same_shape(src0f, dst));
+
+    SYCL_CHECK(ggml_sycl_set_device(ctx.device));
+
+    const int64_t nclasses = src0f->ne[0];
+    const int64_t nrows = ggml_nrows(src0f);
+
+    const float * grad_d  = (const float *) grad->data;
+    const float * logits_d = (const float *) src0f->data;
+    const float * labels_d = (const float *) src1f->data;
+    float * dst_d = (float *) dst->data;
+
+    const sycl::range<3> block(1, 1, WARP_SIZE);
+    const sycl::range<3> grid(1, 1, nrows);
+    const size_t nbytes_shared = (size_t) nclasses * sizeof(float);
+    const size_t smpbo = ggml_sycl_info().devices[ctx.device].smpbo;
+
+    if (nbytes_shared <= smpbo) {
+        ctx.stream()->submit([&](sycl::handler & cgh) {
+            sycl::local_accessor<float, 1> smem(sycl::range<1>(nclasses), cgh);
+            cgh.parallel_for(
+                sycl::nd_range<3>(grid * block, block),
+                [=](sycl::nd_item<3> item) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                    cross_entropy_loss_back_f32_kernel<true>(
+                        grad_d, logits_d, labels_d, dst_d,
+                        (int) nclasses, (int) nrows,
+                        get_pointer(smem), item);
+                });
+        });
+    } else {
+        ctx.stream()->parallel_for(
+            sycl::nd_range<3>(grid * block, block),
+            [=](sycl::nd_item<3> item) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                cross_entropy_loss_back_f32_kernel<false>(
+                    grad_d, logits_d, labels_d, dst_d,
+                    (int) nclasses, (int) nrows,
+                    nullptr, item);
+            });
+    }
+}

ggml/src/ggml-sycl/cross_entropy_loss.hpp

@@ -0,0 +1,7 @@
+#pragma once
+
+#include "common.hpp"
+
+void ggml_sycl_cross_entropy_loss(ggml_backend_sycl_context & ctx, ggml_tensor * dst);
+
+void ggml_sycl_cross_entropy_loss_back(ggml_backend_sycl_context & ctx, ggml_tensor * dst);

ggml/src/ggml-sycl/dmmv.cpp

@@ -680,14 +680,14 @@ static void dequantize_mul_mat_vec_q4_k(const void *__restrict__ vx,
             q16[2] = q2[0] & 0x0f0f;
             q16[3] = q2[0] & 0xf0f0;
 
-            float4 s = {0.f, 0.f, 0.f, 0.f};
+            sycl::float4 s = {0.f, 0.f, 0.f, 0.f};
             float smin = 0;
             for (int l = 0; l < 2; ++l) {
-                s.x += y1[l] * q4[l+0]; s.y += y1[l+32] * q4[l+2];
-                s.z += y2[l] * q4[l+4]; s.w += y2[l+32] * q4[l+6];
+                s.x() += y1[l] * q4[l+0]; s.y() += y1[l+32] * q4[l+2];
+                s.z() += y2[l] * q4[l+4]; s.w() += y2[l+32] * q4[l+6];
                 smin += y1[l] * sc[2] + y1[l+32] * sc[3] + y2[l] * sc[6] + y2[l+32] * sc[7];
             }
-            tmp += dall * (s.x * sc[0] + s.y * sc[1] * 1.f/16.f + s.z * sc[4] + s.w * sc[5] * 1.f/16.f) - dmin * smin;
+            tmp += dall * (s.x() * sc[0] + s.y() * sc[1] * 1.f/16.f + s.z() * sc[4] + s.w() * sc[5] * 1.f/16.f) - dmin * smin;
 #endif
         }
 
@@ -835,14 +835,14 @@ static void dequantize_mul_mat_vec_q4_k_reorder(const void *__restrict__ vx,
             q16[2] = q2[0] & 0x0f0f;
             q16[3] = q2[0] & 0xf0f0;
 
-            float4 s = {0.f, 0.f, 0.f, 0.f};
+            sycl::float4 s = {0.f, 0.f, 0.f, 0.f};
             float smin = 0;
             for (int l = 0; l < 2; ++l) {
-                s.x += y1[l] * q4[l+0]; s.y += y1[l+32] * q4[l+2];
-                s.z += y2[l] * q4[l+4]; s.w += y2[l+32] * q4[l+6];
+                s.x() += y1[l] * q4[l+0]; s.y() += y1[l+32] * q4[l+2];
+                s.z() += y2[l] * q4[l+4]; s.w() += y2[l+32] * q4[l+6];
                 smin += y1[l] * sc[2] + y1[l+32] * sc[3] + y2[l] * sc[6] + y2[l+32] * sc[7];
             }
-            tmp += dall * (s.x * sc[0] + s.y * sc[1] * 1.f/16.f + s.z * sc[4] + s.w * sc[5] * 1.f/16.f) - dmin * smin;
+            tmp += dall * (s.x() * sc[0] + s.y() * sc[1] * 1.f/16.f + s.z() * sc[4] + s.w() * sc[5] * 1.f/16.f) - dmin * smin;
 #endif
         }
 
@@ -1126,7 +1126,7 @@ static void dequantize_mul_mat_vec_q5_k_reorder(const void *__restrict__ vx,
 
     // sum up partial sums and write back result
 #pragma unroll
-    for (int mask = QK_WARP_SIZE / 2; mask > 0; mask >>= 1) {
+    for (int mask = WARP_SIZE / 2; mask > 0; mask >>= 1) {
         tmp +=
             dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask);
     }
@@ -1762,10 +1762,13 @@ static void dequantize_mul_mat_vec_q5_K_sycl_reorder(const void *vx, const float
                                                      const int nrows,
                                                      dpct::queue_ptr stream) {
     GGML_ASSERT(ncols % QK_K == 0);
-    const sycl::range<3> block_dims(1, 1, QK_WARP_SIZE);
+    const int ny = 2 / K_QUANTS_PER_ITERATION;
+    const int block_num_y = (nrows + ny - 1) / ny;
+    const sycl::range<3> block_nums(1, 1, block_num_y);
+    const sycl::range<3> block_dims(1, ny, WARP_SIZE);
     stream->parallel_for(
-        sycl::nd_range<3>(sycl::range<3>(1, 1, nrows) * block_dims, block_dims),
-        [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(QK_WARP_SIZE)]] {
+        sycl::nd_range<3>(block_nums * block_dims, block_dims),
+        [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
             dequantize_mul_mat_vec_q5_k_reorder(vx, y, dst, ncols, nrows, item_ct1);
         });
 }

ggml/src/ggml-sycl/element_wise.cpp

@@ -9,9 +9,12 @@
 #define SYCL_LOCAL_ID_CALC(ITEM, IDX) \
     (ITEM.get_local_range(IDX) * ITEM.get_group(IDX) + ITEM.get_local_id(IDX))
 
-static void acc_f32(const float * x, const float * y, float * dst, const int64_t ne,
-        const int64_t ne10, const int64_t ne11, const int64_t ne12, const int64_t ne13,
-        const int64_t s11, const int64_t s12, const int64_t s13, const int64_t offset) {
+static void acc_f32(const char * x, const char * y, float * dst, const int64_t ne,
+    const int64_t ne0, const int64_t ne1, const int64_t ne2, const int64_t ne3,
+    const int64_t nb00, const int64_t nb01, const int64_t nb02, const int64_t nb03,
+    const int64_t ne10, const int64_t ne11, const int64_t ne12, const int64_t ne13,
+    const int64_t nb10, const int64_t nb11, const int64_t nb12, const int64_t nb13,
+    const int64_t s11, const int64_t s12, const int64_t s13, const int64_t offset) {
     auto item_ct1 = sycl::ext::oneapi::this_work_item::get_nd_item<3>();
     const int64_t i = SYCL_LOCAL_ID_CALC(item_ct1, 2);
 
@@ -30,9 +33,18 @@ static void acc_f32(const float * x, const float * y, float * dst, const int64_t
     tmp -= i11 * s11;
     const int64_t i10 = tmp;
 
-    float val = x[i];
+    int64_t tmp_dst = i;
+    const int64_t i3 = tmp_dst / (ne2*ne1*ne0);
+    tmp_dst -= i3 * (ne2*ne1*ne0);
+    const int64_t i2 = tmp_dst / (ne1*ne0);
+    tmp_dst -= i2 * (ne1*ne0);
+    const int64_t i1 = tmp_dst / ne0;
+    tmp_dst -= i1 * ne0;
+    const int64_t i0 = tmp_dst;
+
+    float val = *(const float *) (x + i0*nb00 + i1*nb01 + i2*nb02 + i3*nb03);
     if (src1_idx >= 0 && i10 < ne10 && i11 < ne11 && i12 < ne12 && i13 < ne13) {
-        val += y[((i13*ne12 + i12) * ne11 + i11) * ne10 + i10];
+        val += *(const float *) (y + i10*nb10 + i11*nb11 + i12*nb12 + i13*nb13);
     }
     dst[i] = val;
 }
@@ -422,15 +434,24 @@ static void gated_op_fused_geglu_quick(const T * x, const T * g, T * dst, const
 }
 
 namespace ggml_sycl_detail {
-static void acc_f32_sycl(const float *x, const float *y, float *dst,
-                         const int64_t n_elements, const int64_t ne10, const int64_t ne11,
-                         const int64_t ne12, const int64_t ne13, const int64_t s1, const int64_t s2, const int64_t s3,
+static void acc_f32_sycl(const char *x, const char *y, float *dst,
+                         const int64_t n_elements,
+                         const int64_t ne0, const int64_t ne1, const int64_t ne2, const int64_t ne3,
+                         const int64_t nb00, const int64_t nb01, const int64_t nb02, const int64_t nb03,
+                         const int64_t ne10, const int64_t ne11, const int64_t ne12, const int64_t ne13,
+                         const int64_t nb10, const int64_t nb11, const int64_t nb12, const int64_t nb13,
+                         const int64_t s1, const int64_t s2, const int64_t s3,
                          const int64_t offset, queue_ptr stream) {
     const int num_blocks = (n_elements + SYCL_ACC_BLOCK_SIZE - 1) / SYCL_ACC_BLOCK_SIZE;
     stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_ACC_BLOCK_SIZE),
                                            sycl::range<3>(1, 1, SYCL_ACC_BLOCK_SIZE)),
                          [=](sycl::nd_item<3> /*item_ct1*/) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-                             acc_f32(x, y, dst, n_elements, ne10, ne11, ne12, ne13, s1, s2, s3, offset);
+                            acc_f32(x, y, dst, n_elements,
+                                ne0, ne1, ne2, ne3,
+                                nb00, nb01, nb02, nb03,
+                                ne10, ne11, ne12, ne13,
+                                nb10, nb11, nb12, nb13,
+                                s1, s2, s3, offset);
                          });
 }
 
@@ -843,8 +864,8 @@ static inline void ggml_sycl_op_acc(ggml_backend_sycl_context & ctx, ggml_tensor
     const ggml_tensor * src0 = dst->src[0];
     const ggml_tensor * src1 = dst->src[1];
 
-    const float * src0_d = (const float *) src0->data;
-    const float * src1_d = (const float *) src1->data;
+    const char  * src0_d = (const char  *) src0->data;
+    const char  * src1_d = (const char  *) src1->data;
     float       * dst_d  = (float       *)  dst->data;
 
     dpct::queue_ptr stream = ctx.stream();
@@ -853,17 +874,20 @@ static inline void ggml_sycl_op_acc(ggml_backend_sycl_context & ctx, ggml_tensor
     GGML_ASSERT(src1->type == GGML_TYPE_F32);
     GGML_ASSERT( dst->type == GGML_TYPE_F32);
 
-    GGML_ASSERT(ggml_is_contiguous(src1));
     GGML_ASSERT(dst->nb[0] == ggml_element_size(dst));
     GGML_ASSERT(ggml_is_contiguously_allocated(dst));
+    GGML_ASSERT(ggml_are_same_shape(src0, dst));
 
-    const int64_t s1     = dst->op_params[0] / sizeof(float);
-    const int64_t s2     = dst->op_params[1] / sizeof(float);
-    const int64_t s3     = dst->op_params[2] / sizeof(float);
-    const int64_t offset = dst->op_params[3] / sizeof(float);
+    const int64_t s1     = (int64_t) ((const int32_t *) dst->op_params)[0] / (int64_t) sizeof(float);
+    const int64_t s2     = (int64_t) ((const int32_t *) dst->op_params)[1] / (int64_t) sizeof(float);
+    const int64_t s3     = (int64_t) ((const int32_t *) dst->op_params)[2] / (int64_t) sizeof(float);
+    const int64_t offset = (int64_t) ((const int32_t *) dst->op_params)[3] / (int64_t) sizeof(float);
 
     ggml_sycl_detail::acc_f32_sycl(src0_d, src1_d, dst_d, ggml_nelements(dst),
+        dst->ne[0], dst->ne[1], dst->ne[2], dst->ne[3],
+        src0->nb[0], src0->nb[1], src0->nb[2], src0->nb[3],
         src1->ne[0], src1->ne[1], src1->ne[2], src1->ne[3],
+        src1->nb[0], src1->nb[1], src1->nb[2], src1->nb[3],
         s1, s2, s3, offset, stream);
 }
 

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

@@ -41,7 +41,7 @@
 #if SYCL_EXT_ONEAPI_VIRTUAL_MEM
 #    include <sycl/ext/oneapi/virtual_mem/physical_mem.hpp>
 #    include <sycl/ext/oneapi/virtual_mem/virtual_mem.hpp>
-#    define GGML_SYCL_USE_VMM
+#    define GGML_SYCL_SUPPORT_VMM
 #endif
 #include <sycl/half_type.hpp>
 
@@ -74,15 +74,16 @@
 #include "ggml-sycl/solve_tri.hpp"
 #include "ggml-sycl/gated_delta_net.hpp"
 #include "ggml-sycl/pool.hpp"
+#include "ggml-sycl/cross_entropy_loss.hpp"
 
 #define MEM_SIZE_2M	0x00200000
 #define MEM_SIZE_1G	0x40000000
 
 static bool g_sycl_loaded = false;
 int g_ggml_sycl_debug = 0;
-int g_ggml_sycl_disable_optimize = 0;
-int g_ggml_sycl_disable_graph = 0;
-int g_ggml_sycl_disable_dnn = 0;
+int g_ggml_sycl_enable_optimize = 1;
+int g_ggml_sycl_enable_graph = 0;
+int g_ggml_sycl_enable_dnn = 1;
 int g_ggml_sycl_enable_vmm = 1;
 int g_ggml_sycl_prioritize_dmmv = 0;
 int g_ggml_sycl_use_async_mem_op = 0;
@@ -117,7 +118,7 @@ static ggml_sycl_device_info ggml_sycl_init() {
         SYCL_CHECK(CHECK_TRY_ERROR(dpct::get_device_info(
             prop, device)));
 
-#if !defined(GGML_SYCL_USE_VMM)
+#if !defined(GGML_SYCL_SUPPORT_VMM)
         info.devices[i].vmm = 0;
 #else
         info.devices[i].vmm = device.has(sycl::aspect::ext_oneapi_virtual_mem);
@@ -265,14 +266,24 @@ void ggml_backend_sycl_print_sycl_devices() {
     print_device_opt_feature(device_count);
 }
 
+static const char* dev2dev_int2str(int dev2dev) {
+    if (dev2dev == DEV2DEV_MEMCPY_SYCL) {
+        return "SYCL API";
+    } else if (dev2dev == DEV2DEV_MEMCPY_L0) {
+        return "Level Zero API";
+    } else {
+        return "Unknown";
+    }
+}
+
 static void ggml_check_sycl() try {
     static bool initialized = false;
 
     if (!initialized) {
         g_ggml_sycl_debug = ggml_sycl_get_env("GGML_SYCL_DEBUG", 0);
-        g_ggml_sycl_disable_optimize = ggml_sycl_get_env("GGML_SYCL_DISABLE_OPT", 0);
-        g_ggml_sycl_disable_graph = ggml_sycl_get_env("GGML_SYCL_DISABLE_GRAPH", 1);
-        g_ggml_sycl_disable_dnn = ggml_sycl_get_env("GGML_SYCL_DISABLE_DNN", 0);
+        g_ggml_sycl_enable_optimize = ggml_sycl_get_env("GGML_SYCL_ENABLE_OPT", 1);
+        g_ggml_sycl_enable_graph = ggml_sycl_get_env("GGML_SYCL_ENABLE_GRAPH", 0);
+        g_ggml_sycl_enable_dnn = ggml_sycl_get_env("GGML_SYCL_ENABLE_DNN", 1);
         g_ggml_sycl_enable_vmm = ggml_sycl_get_env("GGML_SYCL_ENABLE_VMM", 1);
         g_ggml_sycl_prioritize_dmmv = ggml_sycl_get_env("GGML_SYCL_PRIORITIZE_DMMV", 0);
 
@@ -292,66 +303,56 @@ static void ggml_check_sycl() try {
         GGML_SYCL_DEBUG("[SYCL] call ggml_check_sycl\n");
 
         GGML_LOG_INFO("Build with Macros:\n");
-#if defined(GGML_SYCL_FORCE_MMQ)
-        GGML_LOG_INFO("  GGML_SYCL_FORCE_MMQ: yes\n");
-#else
-        GGML_LOG_INFO("  GGML_SYCL_FORCE_MMQ: no\n");
-#endif
-#if defined(GGML_SYCL_F16)
-        GGML_LOG_INFO("  GGML_SYCL_F16: yes\n");
-#else
-        GGML_LOG_INFO("  GGML_SYCL_F16: no\n");
-#endif
-#if defined(GGML_SYCL_GRAPH)
-        GGML_LOG_INFO("  GGML_SYCL_GRAPH: yes\n");
-#else
-        GGML_LOG_INFO("  GGML_SYCL_GRAPH: no\n");
-#endif
 #if defined(GGML_SYCL_DNNL)
         GGML_LOG_INFO("  GGML_SYCL_DNNL: yes\n");
 #else
         GGML_LOG_INFO("  GGML_SYCL_DNNL: no\n");
 #endif
+
+#if defined(GGML_SYCL_F16)
+        GGML_LOG_INFO("  GGML_SYCL_F16: yes\n");
+#else
+        GGML_LOG_INFO("  GGML_SYCL_F16: no\n");
+#endif
+
+#if defined(GGML_SYCL_FORCE_MMQ)
+        GGML_LOG_INFO("  GGML_SYCL_FORCE_MMQ: yes\n");
+#else
+        GGML_LOG_INFO("  GGML_SYCL_FORCE_MMQ: no\n");
+#endif
+
+#if defined(GGML_SYCL_GRAPH)
+        GGML_LOG_INFO("  GGML_SYCL_GRAPH: yes\n");
+#else
+        GGML_LOG_INFO("  GGML_SYCL_GRAPH: no\n");
+#endif
+
 #if defined(GGML_SYCL_SUPPORT_LEVEL_ZERO_API)
         GGML_LOG_INFO("  GGML_SYCL_SUPPORT_LEVEL_ZERO_API: yes\n");
 #else
         GGML_LOG_INFO("  GGML_SYCL_SUPPORT_LEVEL_ZERO_API: no\n");
 #endif
-#if defined(GGML_SYCL_USE_VMM)
-        GGML_LOG_INFO("  GGML_SYCL_USE_VMM: yes\n");
+#if defined(GGML_SYCL_SUPPORT_VMM)
+        GGML_LOG_INFO("  GGML_SYCL_SUPPORT_VMM: yes\n");
 #else
-        GGML_LOG_INFO("  GGML_SYCL_USE_VMM: no\n");
+        GGML_LOG_INFO("  GGML_SYCL_SUPPORT_VMM: no\n");
 #endif
 
         GGML_LOG_INFO("Running with Environment Variables:\n");
         GGML_LOG_INFO("  GGML_SYCL_DEBUG: %d\n", g_ggml_sycl_debug);
-        GGML_LOG_INFO("  GGML_SYCL_DISABLE_OPT: %d\n", g_ggml_sycl_disable_optimize);
-#ifdef GGML_SYCL_GRAPH
-        GGML_LOG_INFO("  GGML_SYCL_DISABLE_GRAPH: %d\n", g_ggml_sycl_disable_graph);
-#else
-        GGML_LOG_INFO("  GGML_SYCL_DISABLE_GRAPH: graph disabled by compile flag\n");
-#endif
+
 #ifdef GGML_SYCL_SUPPORT_LEVEL_ZERO_API
-        GGML_LOG_INFO("  GGML_SYCL_USE_LEVEL_ZERO_API: %d\n", g_ggml_sycl_use_level_zero_api);
-        GGML_LOG_INFO("  GGML_SYCL_DEV2DEV_MEMCPY: %d\n", g_ggml_sycl_dev2dev_memcpy);
+        GGML_LOG_INFO("  GGML_SYCL_DEV2DEV_MEMCPY: %d (%s)\n", g_ggml_sycl_dev2dev_memcpy, dev2dev_int2str(g_ggml_sycl_dev2dev_memcpy));
 #else
-        GGML_LOG_INFO("  GGML_SYCL_USE_LEVEL_ZERO_API: Disable Level Zero API usage by compile flag\n");
-        GGML_LOG_INFO("  GGML_SYCL_DEV2DEV_MEMCPY: %d, enable to SYCL API since missing GGML_SYCL_SUPPORT_LEVEL_ZERO_API\n",
-                      g_ggml_sycl_dev2dev_memcpy);
+        GGML_LOG_INFO("  GGML_SYCL_DEV2DEV_MEMCPY: %d (%s), enable to SYCL API since missing GGML_SYCL_SUPPORT_LEVEL_ZERO_API\n",
+                      g_ggml_sycl_dev2dev_memcpy, dev2dev_int2str(g_ggml_sycl_dev2dev_memcpy));
 #endif
-#if GGML_SYCL_DNNL
-        GGML_LOG_INFO("  GGML_SYCL_DISABLE_DNN: %d\n", g_ggml_sycl_disable_dnn);
+
+#if defined(GGML_SYCL_DNNL)
+        GGML_LOG_INFO("  GGML_SYCL_ENABLE_DNN: %d\n", g_ggml_sycl_enable_dnn);
 #else
-        GGML_LOG_INFO("  GGML_SYCL_DISABLE_DNN: DNN disabled by compile flag\n");
+        GGML_LOG_INFO("  GGML_SYCL_ENABLE_DNN: DNN disabled by compile flag\n");
 #endif
-#if defined(GGML_SYCL_USE_VMM)
-        GGML_LOG_INFO("  GGML_SYCL_ENABLE_VMM: %d\n", g_ggml_sycl_enable_vmm);
-#else
-        GGML_LOG_INFO("  GGML_SYCL_ENABLE_VMM: virtual memory extension is not available\n");
-#endif
-        GGML_LOG_INFO("  GGML_SYCL_PRIORITIZE_DMMV: %d\n", g_ggml_sycl_prioritize_dmmv);
-        g_ggml_sycl_use_async_mem_op_requested = ggml_sycl_get_env("GGML_SYCL_USE_ASYNC_MEM_OP", 1);
-        GGML_LOG_INFO("  GGML_SYCL_USE_ASYNC_MEM_OP: %d\n", g_ggml_sycl_use_async_mem_op_requested);
 
 #ifdef SYCL_FLASH_ATTN
         GGML_LOG_INFO("  GGML_SYCL_ENABLE_FLASH_ATTN: %d\n", g_ggml_sycl_enable_flash_attention);
@@ -360,6 +361,31 @@ static void ggml_check_sycl() try {
             g_ggml_sycl_enable_flash_attention);
 #endif
 
+#ifdef GGML_SYCL_GRAPH
+        GGML_LOG_INFO("  GGML_SYCL_ENABLE_GRAPH: %d\n", g_ggml_sycl_enable_graph);
+#else
+        GGML_LOG_INFO("  GGML_SYCL_ENABLE_GRAPH: graph disabled by compile flag\n");
+#endif
+
+        GGML_LOG_INFO("  GGML_SYCL_ENABLE_OPT: %d\n", g_ggml_sycl_enable_optimize);
+
+#if defined(GGML_SYCL_SUPPORT_VMM)
+        GGML_LOG_INFO("  GGML_SYCL_ENABLE_VMM: %d\n", g_ggml_sycl_enable_vmm);
+#else
+        GGML_LOG_INFO("  GGML_SYCL_ENABLE_VMM: virtual memory extension is not available\n");
+#endif
+
+        GGML_LOG_INFO("  GGML_SYCL_PRIORITIZE_DMMV: %d\n", g_ggml_sycl_prioritize_dmmv);
+
+        g_ggml_sycl_use_async_mem_op_requested = ggml_sycl_get_env("GGML_SYCL_USE_ASYNC_MEM_OP", 1);
+        GGML_LOG_INFO("  GGML_SYCL_USE_ASYNC_MEM_OP: %d\n", g_ggml_sycl_use_async_mem_op_requested);
+
+#ifdef GGML_SYCL_SUPPORT_LEVEL_ZERO_API
+        GGML_LOG_INFO("  GGML_SYCL_USE_LEVEL_ZERO_API: %d\n", g_ggml_sycl_use_level_zero_api);
+#else
+        GGML_LOG_INFO("  GGML_SYCL_USE_LEVEL_ZERO_API: Disable Level Zero API usage by compile flag\n");
+#endif
+
         GGML_LOG_INFO("  GGML_SYCL_USM_SYSTEM: %d\n", g_ggml_sycl_usm_system);
 
 /* NOT REMOVE, keep it for next optimize for XMX.
@@ -373,7 +399,7 @@ static void ggml_check_sycl() try {
         // staging path while preserving queue ordering semantics. Graph support still depends on the extension being
         // available, but it no longer needs to control the non-graph fast path.
 #if defined(GGML_SYCL_GRAPH) && SYCL_EXT_ONEAPI_ASYNC_MEMORY_ALLOC
-        g_ggml_sycl_use_async_mem_op = g_ggml_sycl_use_async_mem_op_requested || !g_ggml_sycl_disable_graph;
+        g_ggml_sycl_use_async_mem_op = g_ggml_sycl_use_async_mem_op_requested || g_ggml_sycl_enable_graph;
         if (g_ggml_sycl_use_async_mem_op) {
             for (unsigned int i = 0; i < dpct::dev_mgr::instance().device_count(); ++i) {
                 if (!dpct::dev_mgr::instance().get_device(i).has(sycl::aspect::ext_oneapi_async_memory_alloc)) {
@@ -516,12 +542,14 @@ ggml_backend_sycl_buffer_init_tensor(ggml_backend_buffer_t buffer,
         return GGML_STATUS_SUCCESS;
     }
 
-    if (!g_ggml_sycl_disable_optimize) {
+    if (g_ggml_sycl_enable_optimize) {
         // set reorder extra buffer based on supported type
         switch (tensor->type) {
             case GGML_TYPE_Q4_0:
             case GGML_TYPE_Q8_0:
+            case GGML_TYPE_Q3_K:
             case GGML_TYPE_Q4_K:
+            case GGML_TYPE_Q5_K:
             case GGML_TYPE_Q6_K:{
                 ggml_tensor_extra_gpu * extra = new ggml_tensor_extra_gpu{};
                 tensor->extra                 = extra;
@@ -1562,7 +1590,7 @@ struct ggml_sycl_pool_leg : public ggml_sycl_pool {
 };
 
 // pool with virtual memory management
-#if defined(GGML_SYCL_USE_VMM)
+#if defined(GGML_SYCL_SUPPORT_VMM)
 struct ggml_sycl_pool_vmm : public ggml_sycl_pool {
     static const size_t SYCL_POOL_VMM_MAX_SIZE = 1ull << 35; // 32 GB
 
@@ -1674,7 +1702,7 @@ struct ggml_sycl_pool_vmm : public ggml_sycl_pool {
         GGML_ASSERT(ptr == reinterpret_cast<void *>(pool_addr + pool_used));
     }
 };
-#endif // defined(GGML_SYCL_USE_VMM)
+#endif // defined(GGML_SYCL_SUPPORT_VMM)
 
 struct ggml_sycl_pool_host : public ggml_sycl_pool {
     queue_ptr qptr;
@@ -1756,11 +1784,11 @@ std::unique_ptr<ggml_sycl_pool> ggml_backend_sycl_context::new_pool_for_host(que
 }
 
 std::unique_ptr<ggml_sycl_pool> ggml_backend_sycl_context::new_pool_for_device(queue_ptr qptr, int device) {
-#if defined(GGML_SYCL_USE_VMM)
+#if defined(GGML_SYCL_SUPPORT_VMM)
     if (g_ggml_sycl_enable_vmm && ggml_sycl_info().devices[device].vmm) {
         return std::unique_ptr<ggml_sycl_pool>(new ggml_sycl_pool_vmm(qptr, device));
     }
-#endif // defined(GGML_SYCL_USE_VMM)
+#endif // defined(GGML_SYCL_SUPPORT_VMM)
     return std::unique_ptr<ggml_sycl_pool>(new ggml_sycl_pool_leg(qptr, device));
 }
 
@@ -2088,11 +2116,148 @@ static int next_power_of_2(int x) {
     return n;
 }
 
+static void init_argsort_indices_padded(
+        int * idx,
+        const int nrows,
+        const int ncols_pad,
+        const sycl::nd_item<1> & item_ct1) {
+    const size_t gid = item_ct1.get_local_range(0) * item_ct1.get_group(0) + item_ct1.get_local_id(0);
+    const size_t total = (size_t) nrows * (size_t) ncols_pad;
+
+    if (gid >= total) {
+        return;
+    }
+
+    idx[gid] = (int) (gid % (size_t) ncols_pad);
+}
+
+template <ggml_sort_order order>
+static void argsort_f32_i32_global_pass(const float *            x,
+                                        int *                    idx,
+                                        const int                ncols,
+                                        const int                nrows,
+                                        const int                ncols_pad,
+                                        const int                j,
+                                        const int                k,
+                                        const sycl::nd_item<1> & item_ct1) {
+    const size_t gid   = item_ct1.get_local_range(0) * item_ct1.get_group(0) + item_ct1.get_local_id(0);
+    const size_t total = (size_t) nrows * (size_t) ncols_pad;
+
+    if (gid >= total) {
+        return;
+    }
+
+    const int row = (int) (gid / (size_t) ncols_pad);
+    const int col = (int) (gid % (size_t) ncols_pad);
+    const int ixj = col ^ j;
+
+    if (ixj <= col || ixj >= ncols_pad) {
+        return;
+    }
+
+    const size_t base  = (size_t) row * (size_t) ncols_pad;
+    const size_t pos_a = base + (size_t) col;
+    const size_t pos_b = base + (size_t) ixj;
+
+    const int a = idx[pos_a];
+    const int b = idx[pos_b];
+
+    bool do_swap = false;
+
+    if ((col & k) == 0) {
+        if (a >= ncols ||
+            (b < ncols &&
+             (order == GGML_SORT_ORDER_ASC ?
+                  x[(size_t) row * (size_t) ncols + (size_t) a] > x[(size_t) row * (size_t) ncols + (size_t) b] :
+                  x[(size_t) row * (size_t) ncols + (size_t) a] < x[(size_t) row * (size_t) ncols + (size_t) b]))) {
+            do_swap = true;
+        }
+    } else {
+        if (b >= ncols ||
+            (a < ncols &&
+             (order == GGML_SORT_ORDER_ASC ?
+                  x[(size_t) row * (size_t) ncols + (size_t) a] < x[(size_t) row * (size_t) ncols + (size_t) b] :
+                  x[(size_t) row * (size_t) ncols + (size_t) a] > x[(size_t) row * (size_t) ncols + (size_t) b]))) {
+            do_swap = true;
+        }
+    }
+
+    if (do_swap) {
+        idx[pos_a] = b;
+        idx[pos_b] = a;
+    }
+}
+
+static void copy_argsort_indices_unpadded(const int *              idx_padded,
+                                          int *                    dst,
+                                          const int                nrows,
+                                          const int                ncols,
+                                          const int                ncols_pad,
+                                          const sycl::nd_item<1> & item_ct1) {
+    const size_t gid   = item_ct1.get_local_range(0) * item_ct1.get_group(0) + item_ct1.get_local_id(0);
+    const size_t total = (size_t) nrows * (size_t) ncols;
+
+    if (gid >= total) {
+        return;
+    }
+
+    const int row = (int) (gid / (size_t) ncols);
+    const int col = (int) (gid % (size_t) ncols);
+
+    dst[(size_t) row * (size_t) ncols + (size_t) col] = idx_padded[(size_t) row * (size_t) ncols_pad + (size_t) col];
+}
+
 static void argsort_f32_i32_sycl(const float *x, int *dst, const int ncols,
                                  const int nrows, ggml_sort_order order,
-                                 queue_ptr stream, int device) {
+                                 queue_ptr stream, int device, ggml_sycl_pool & pool) {
     // bitonic sort requires ncols to be power of 2
     const int ncols_pad = next_power_of_2(ncols);
+    const size_t shared_mem = (size_t) ncols_pad * sizeof(int);
+    const size_t smpbo = ggml_sycl_info().devices[device].smpbo;
+
+    if (shared_mem > smpbo) {
+        ggml_sycl_pool_alloc<int> idx_padded_alloc(pool, (size_t) nrows * (size_t) ncols_pad);
+        int *                     idx_padded = idx_padded_alloc.get();
+
+        constexpr size_t block_size     = 256;
+        const size_t     total_padded   = (size_t) nrows * (size_t) ncols_pad;
+        const size_t     nblocks_padded = (total_padded + block_size - 1) / block_size;
+
+        stream->parallel_for(
+            sycl::nd_range<1>(sycl::range<1>(nblocks_padded * block_size), sycl::range<1>(block_size)),
+            [=](sycl::nd_item<1> item_ct1) { init_argsort_indices_padded(idx_padded, nrows, ncols_pad, item_ct1); });
+
+        for (int k = 2; k <= ncols_pad; k *= 2) {
+            for (int j = k / 2; j > 0; j /= 2) {
+                if (order == GGML_SORT_ORDER_ASC) {
+                    stream->parallel_for(
+                        sycl::nd_range<1>(sycl::range<1>(nblocks_padded * block_size), sycl::range<1>(block_size)),
+                        [=](sycl::nd_item<1> item_ct1) {
+                            argsort_f32_i32_global_pass<GGML_SORT_ORDER_ASC>(x, idx_padded, ncols, nrows, ncols_pad, j,
+                                                                             k, item_ct1);
+                        });
+                } else if (order == GGML_SORT_ORDER_DESC) {
+                    stream->parallel_for(
+                        sycl::nd_range<1>(sycl::range<1>(nblocks_padded * block_size), sycl::range<1>(block_size)),
+                        [=](sycl::nd_item<1> item_ct1) {
+                            argsort_f32_i32_global_pass<GGML_SORT_ORDER_DESC>(x, idx_padded, ncols, nrows, ncols_pad, j,
+                                                                              k, item_ct1);
+                        });
+                } else {
+                    GGML_ABORT("invalid sort order");
+                }
+            }
+        }
+
+        const size_t total   = (size_t) nrows * (size_t) ncols;
+        const size_t nblocks = (total + block_size - 1) / block_size;
+        stream->parallel_for(sycl::nd_range<1>(sycl::range<1>(nblocks * block_size), sycl::range<1>(block_size)),
+                             [=](sycl::nd_item<1> item_ct1) {
+                                 copy_argsort_indices_unpadded(idx_padded, dst, nrows, ncols, ncols_pad, item_ct1);
+                             });
+
+        return;
+    }
 
     int nth = 1;
     int max_block_size = ggml_sycl_info().max_work_group_sizes[device];
@@ -2105,8 +2270,6 @@ static void argsort_f32_i32_sycl(const float *x, int *dst, const int ncols,
 
     const sycl::range<3> block_dims(1, 1, nth);
     const sycl::range<3> block_nums(1, nrows, 1);
-    const size_t shared_mem = ncols_pad * sizeof(int);
-    GGML_ASSERT(shared_mem<=ggml_sycl_info().devices[device].smpbo);
 
     if (order == GGML_SORT_ORDER_ASC) {
         stream->submit([&](sycl::handler &cgh) {
@@ -2429,7 +2592,7 @@ inline void ggml_sycl_op_mul_mat_sycl(
 
 #if GGML_SYCL_DNNL && defined(GGML_SYCL_HAS_BF16)
     // Fast path for bf16 src0
-    if (src0->type == GGML_TYPE_BF16 && !g_ggml_sycl_disable_dnn && ggml_is_contiguous(src0) &&
+    if (src0->type == GGML_TYPE_BF16 && g_ggml_sycl_enable_dnn && ggml_is_contiguous(src0) &&
         row_diff == src0->ne[1]) {
         using bf16_t = sycl::ext::oneapi::bfloat16;
         ggml_sycl_pool_alloc<bf16_t> src1_as_bf16(ctx.pool(), src1_ncols*ne10);
@@ -2482,7 +2645,7 @@ inline void ggml_sycl_op_mul_mat_sycl(
                                          : src1_as_f16.get();
 
 #if GGML_SYCL_DNNL
-        if (!g_ggml_sycl_disable_dnn) {
+        if (g_ggml_sycl_enable_dnn) {
                 DnnlGemmWrapper::row_gemm(ctx,row_diff, src1_ncols , ne10, src0_ptr,
                                      DnnlGemmWrapper::to_dt<sycl::half>(), src1_ptr, DnnlGemmWrapper::to_dt<sycl::half>(),
                                       dst_dd_i, DnnlGemmWrapper::to_dt<float>(), stream);
@@ -2532,7 +2695,7 @@ inline void ggml_sycl_op_mul_mat_sycl(
             const int64_t gemm_flops = (int64_t)row_diff * src1_ncols * ne10;
             const bool use_mkl_direct = gemm_flops < 256 * 256 * 256;
 #if GGML_SYCL_DNNL
-            if (!g_ggml_sycl_disable_dnn && !use_mkl_direct) {
+            if (g_ggml_sycl_enable_dnn && !use_mkl_direct) {
                 DnnlGemmWrapper::row_gemm(ctx, row_diff, src1_ncols, ne10, src0_ddf_i,
                                           DnnlGemmWrapper::to_dt<float>(), src1_ddf1_i, DnnlGemmWrapper::to_dt<float>(),
                                           dst_dd_i, DnnlGemmWrapper::to_dt<float>(), stream);
@@ -2625,7 +2788,7 @@ inline void ggml_sycl_op_argsort(ggml_backend_sycl_context & ctx, ggml_tensor *
     enum ggml_sort_order order = (enum ggml_sort_order) dst->op_params[0];
 
     argsort_f32_i32_sycl(src0_dd, (int *)dst_dd, ncols, nrows, order,
-                         main_stream, ctx.device);
+                         main_stream, ctx.device, ctx.pool());
 }
 
 static void ggml_sycl_op_top_k(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
@@ -3352,7 +3515,7 @@ static void ggml_sycl_mul_mat_batched_sycl(ggml_backend_sycl_context & ctx, cons
     const int64_t r3 = ne13 / ne03;
 
 #if GGML_SYCL_DNNL
-    if (!g_ggml_sycl_disable_dnn) {
+    if (g_ggml_sycl_enable_dnn) {
             int64_t str_a0 = nb00 / type_size_src0;
             int64_t str_a1 = nb01 / type_size_src0;
             int64_t str_a2 = nb02 / type_size_src0;
@@ -3527,6 +3690,10 @@ inline bool ggml_sycl_supports_reorder_dmmv(enum ggml_type type) {
         case GGML_TYPE_Q1_0:
         case GGML_TYPE_Q4_0:
         case GGML_TYPE_Q8_0:
+        case GGML_TYPE_Q3_K:
+        case GGML_TYPE_Q4_K:
+        case GGML_TYPE_Q5_K:
+        case GGML_TYPE_Q6_K:
             return true;
         default:
             return false;
@@ -4092,12 +4259,12 @@ static bool reorder_qw(const ggml_tensor * src0, dpct::queue_ptr stream) {
 }
 
 static bool should_reorder_tensor(ggml_backend_sycl_context& ctx, const ggml_tensor * dst) {
-    return !g_ggml_sycl_disable_optimize && //allow optimize, controlled by $GGML_SYCL_DISABLE_OPT
-            ctx.opt_feature.reorder &&      //allow this device due to good perf, skip the devices with bad perf.
-            dst->op == GGML_OP_MUL_MAT &&   //limit to some supported cases of Q4_0, to do for more cases.
-            // ne[1] <= 8 so multi-column decode (spec / MTP verify) also bootstraps the reorder;
-            // all reorderable types have a _switch_ncols kernel.
-            dst->src[1]->ne[1] <= 8 && dst->src[1]->ne[2]==1 && dst->src[1]->ne[3]==1;
+    return g_ggml_sycl_enable_optimize && //allow optimize, controlled by $GGML_SYCL_ENABLE_OPT
+           ctx.opt_feature.reorder &&      //allow this device due to good perf, skip the devices with bad perf.
+           dst->op == GGML_OP_MUL_MAT &&   //limit to some supported cases of Q4_0, to do for more cases.
+           // ne[1] <= 8 so multi-column decode (spec / MTP verify) also bootstraps the reorder;
+           // all reorderable types have a _switch_ncols kernel.
+           dst->src[1]->ne[1] <= 8 && dst->src[1]->ne[2]==1 && dst->src[1]->ne[3]==1;
 }
 
 static void opt_for_reorder(ggml_backend_sycl_context * ctx, const ggml_tensor * src0, const ggml_tensor * /* src1 */,
@@ -4136,7 +4303,7 @@ static void opt_for_reorder(ggml_backend_sycl_context * ctx, const ggml_tensor *
 
 // Lazily reorder supported MoE expert weights once their fused path is used.
 static void opt_for_reorder_id(ggml_backend_sycl_context * ctx, const ggml_tensor * src0) {
-    if (g_ggml_sycl_disable_optimize || !ctx->opt_feature.reorder) {
+    if (!g_ggml_sycl_enable_optimize || !ctx->opt_feature.reorder) {
         return;
     }
     if (src0->type != GGML_TYPE_Q4_K && src0->type != GGML_TYPE_Q5_K && src0->type != GGML_TYPE_Q6_K) {
@@ -4604,6 +4771,11 @@ static void ggml_sycl_im2col_3d(ggml_backend_sycl_context & ctx, ggml_tensor * d
     ggml_sycl_op_im2col_3d(ctx, dst);
 }
 
+static void ggml_sycl_col2im_1d(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/1);
+    ggml_sycl_op_col2im_1d(ctx, dst);
+}
+
 static void ggml_sycl_conv_3d(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
     scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/2);
     ggml_sycl_op_conv_3d(ctx, dst);
@@ -4912,6 +5084,12 @@ static bool ggml_sycl_compute_forward(ggml_backend_sycl_context & ctx, struct gg
         case GGML_OP_SOFT_MAX_BACK:
             ggml_sycl_op_soft_max_back(ctx, dst);
             break;
+        case GGML_OP_CROSS_ENTROPY_LOSS:
+            ggml_sycl_cross_entropy_loss(ctx, dst);
+            break;
+        case GGML_OP_CROSS_ENTROPY_LOSS_BACK:
+            ggml_sycl_cross_entropy_loss_back(ctx, dst);
+            break;
         case GGML_OP_ROPE:
             ggml_sycl_rope(ctx, dst);
             break;
@@ -4924,6 +5102,9 @@ static bool ggml_sycl_compute_forward(ggml_backend_sycl_context & ctx, struct gg
         case GGML_OP_IM2COL_3D:
             ggml_sycl_im2col_3d(ctx, dst);
             break;
+        case GGML_OP_COL2IM_1D:
+            ggml_sycl_col2im_1d(ctx, dst);
+            break;
         case GGML_OP_POOL_2D:
             ggml_sycl_pool2d(ctx, dst);
             break;
@@ -5204,7 +5385,10 @@ static ggml_status ggml_backend_sycl_graph_compute(ggml_backend_t backend, ggml_
     auto * sycl_ctx = static_cast<ggml_backend_sycl_context *>(backend->context);
 
 #ifdef GGML_SYCL_GRAPH
-    bool use_sycl_graph = !g_ggml_sycl_disable_graph && check_graph_compatibility(cgraph);
+    bool use_sycl_graph = false;
+    if (g_ggml_sycl_enable_graph) {
+        use_sycl_graph = check_graph_compatibility(cgraph);
+    }
     if (use_sycl_graph) {
         const bool graph_support = dpct::get_device(sycl_ctx->device).has(sycl::aspect::ext_oneapi_limited_graph);
         if (!graph_support) {
@@ -5470,7 +5654,6 @@ static bool do_ggml_backend_sycl_device_supports_op(ggml_backend_dev_t dev, cons
                 // TODO: This specific configuration can fail with oneDNN and needs more debugging
                 if (!ggml_is_permuted(a) && ggml_is_permuted(b) && b->ne[2] > 1 && b->ne[3] > 1 &&
                     a->ne[0] > 128 && a->ne[2] == 1 && src0_type == GGML_TYPE_F16) {
-                        printf("zjy 2\n");
                     return false;
                 }
                 return true;
@@ -5538,70 +5721,99 @@ static bool do_ggml_backend_sycl_device_supports_op(ggml_backend_dev_t dev, cons
             {
                 ggml_type src0_type = op->src[0]->type;
                 ggml_type src1_type = op->src[1]->type;
-                if (src0_type == src1_type && (ggml_is_contiguous(op->src[0]) && ggml_is_contiguous(op->src[1])) && src0_type != GGML_TYPE_BF16) {
-                    return true;
+
+                if (src0_type == GGML_TYPE_F16) {
+                    if (src1_type == GGML_TYPE_Q2_K ||
+                        src1_type == GGML_TYPE_Q3_K ||
+                        src1_type == GGML_TYPE_Q4_K ||
+                        src1_type == GGML_TYPE_Q5_K ||
+                        src1_type == GGML_TYPE_Q6_K ||
+                        src1_type == GGML_TYPE_IQ2_XXS ||
+                        src1_type == GGML_TYPE_IQ2_XS ||
+                        src1_type == GGML_TYPE_IQ2_S ||
+                        src1_type == GGML_TYPE_IQ3_XXS ||
+                        src1_type == GGML_TYPE_IQ1_S ||
+                        src1_type == GGML_TYPE_IQ1_M ||
+                        src1_type == GGML_TYPE_IQ3_S ||
+                        src1_type == GGML_TYPE_IQ4_XS) {
+                        return false;
+                    }
                 }
-                if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_F32) {
-                    return true;
+
+                if (src0_type == GGML_TYPE_BF16) {
+                    if (src1_type == GGML_TYPE_Q4_0 || //big error in ut
+                        src1_type == GGML_TYPE_Q4_1 || //big error in ut
+                        src1_type == GGML_TYPE_Q8_0 || //big error in ut
+                        src1_type == GGML_TYPE_Q2_K ||
+                        src1_type == GGML_TYPE_Q3_K ||
+                        src1_type == GGML_TYPE_Q4_K ||
+                        src1_type == GGML_TYPE_Q5_K ||
+                        src1_type == GGML_TYPE_Q6_K ||
+                        src1_type == GGML_TYPE_IQ2_XXS ||
+                        src1_type == GGML_TYPE_IQ2_XS ||
+                        src1_type == GGML_TYPE_IQ2_S ||
+                        src1_type == GGML_TYPE_IQ3_XXS ||
+                        src1_type == GGML_TYPE_IQ1_S ||
+                        src1_type == GGML_TYPE_IQ1_M ||
+                        src1_type == GGML_TYPE_IQ3_S ||
+                        src1_type == GGML_TYPE_IQ4_XS) {
+                        return false;
+                    }
                 }
-                if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_F16) {
-                    return true;
+
+                if (src0_type == GGML_TYPE_F32) {
+                    if (src1_type == GGML_TYPE_Q2_K ||
+                        src1_type == GGML_TYPE_Q3_K ||
+                        src1_type == GGML_TYPE_Q4_K ||
+                        src1_type == GGML_TYPE_Q5_K ||
+                        src1_type == GGML_TYPE_Q6_K ||
+                        src1_type == GGML_TYPE_IQ2_XXS ||
+                        src1_type == GGML_TYPE_IQ2_XS ||
+                        src1_type == GGML_TYPE_IQ2_S ||
+                        src1_type == GGML_TYPE_IQ3_XXS ||
+                        src1_type == GGML_TYPE_IQ1_S ||
+                        src1_type == GGML_TYPE_IQ1_M ||
+                        src1_type == GGML_TYPE_IQ3_S ||
+                        src1_type == GGML_TYPE_IQ4_XS) {
+                        return false;
+                    }
                 }
-                if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_Q8_0) {
-                    return true;
+
+                if (src1_type == GGML_TYPE_F32) {
+                    if (src0_type == GGML_TYPE_Q1_0 ||
+                        src0_type == GGML_TYPE_NVFP4 ||
+                        src0_type == GGML_TYPE_Q2_K ||
+                        src0_type == GGML_TYPE_Q3_K ||
+                        src0_type == GGML_TYPE_Q4_K ||
+                        src0_type == GGML_TYPE_Q5_K ||
+                        src0_type == GGML_TYPE_Q6_K ||
+                        src0_type == GGML_TYPE_IQ2_XXS ||
+                        src0_type == GGML_TYPE_IQ2_XS ||
+                        src0_type == GGML_TYPE_IQ2_S ||
+                        src0_type == GGML_TYPE_IQ3_XXS ||
+                        src0_type == GGML_TYPE_IQ1_S ||
+                        src0_type == GGML_TYPE_IQ1_M ||
+                        src0_type == GGML_TYPE_IQ3_S ||
+                        src0_type == GGML_TYPE_IQ4_NL ||
+                        src0_type == GGML_TYPE_IQ4_XS
+                    ) {
+                        return false;
+                    }
                 }
-                if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_Q4_0) {
-                    return true;
+
+                if (src0_type == src1_type) {
+                    if (src1_type == GGML_TYPE_IQ2_XXS ||
+                        src1_type == GGML_TYPE_IQ2_XS ||
+                        src1_type == GGML_TYPE_IQ2_S ||
+                        src1_type == GGML_TYPE_IQ3_XXS ||
+                        src1_type == GGML_TYPE_IQ3_S ||
+                        src1_type == GGML_TYPE_IQ1_S ||
+                        src1_type == GGML_TYPE_IQ1_M) {
+                        return false;
+                    }
                 }
-                if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_Q4_1) {
-                    return true;
-                }
-                if (src0_type == GGML_TYPE_F16 && src1_type == GGML_TYPE_F16) {
-                    return true;
-                }
-                if (src0_type == GGML_TYPE_F16 && src1_type == GGML_TYPE_F32) {
-                    return true;
-                }
-                if (src0_type == GGML_TYPE_Q8_0 && src1_type == GGML_TYPE_F32) {
-                    return true;
-                }
-                if (src0_type == GGML_TYPE_Q4_0 && src1_type == GGML_TYPE_F32) {
-                    return true;
-                }
-                if (src0_type == GGML_TYPE_Q4_1 && src1_type == GGML_TYPE_F32) {
-                    return true;
-                }
-                if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_Q5_0) {
-                    return true;
-                }
-                if (src0_type == GGML_TYPE_Q5_0 && src1_type == GGML_TYPE_F32) {
-                    return true;
-                }
-                if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_Q5_1) {
-                    return true;
-                }
-                if (src0_type == GGML_TYPE_Q5_1 && src1_type == GGML_TYPE_F32) {
-                    return true;
-                }
-                if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_IQ4_NL) {
-                    return true;
-                }
-                if(src0_type == GGML_TYPE_Q8_0 && src1_type == GGML_TYPE_Q8_0) {
-                    return true;
-                }
-                if(src0_type == GGML_TYPE_Q5_0 && src1_type == GGML_TYPE_Q5_0) {
-                    return true;
-                }
-                if(src0_type == GGML_TYPE_Q5_1 && src1_type == GGML_TYPE_Q5_1) {
-                    return true;
-                }
-                if(src0_type == GGML_TYPE_Q4_0 && src1_type == GGML_TYPE_Q4_0) {
-                    return true;
-                }
-                if(src0_type == GGML_TYPE_Q4_1 && src1_type == GGML_TYPE_Q4_1) {
-                    return true;
-                }
-                return false;
+
+                return true;
             }
         case GGML_OP_REPEAT_BACK:
             {
@@ -5643,7 +5855,7 @@ static bool do_ggml_backend_sycl_device_supports_op(ggml_backend_dev_t dev, cons
         case GGML_OP_SCALE:
             return true;
         case GGML_OP_CONT:
-            return op->src[0]->type != GGML_TYPE_BF16;
+            return true;
         case GGML_OP_TRI:
             {
                 const ggml_tensor * src0 = op->src[0];
@@ -5666,6 +5878,14 @@ static bool do_ggml_backend_sycl_device_supports_op(ggml_backend_dev_t dev, cons
         case GGML_OP_IM2COL_3D:
         case GGML_OP_UPSCALE:
             return true;
+        case GGML_OP_COL2IM_1D:
+            return ggml_is_contiguous(op->src[0]) &&
+                   (op->type == GGML_TYPE_F32 || op->type == GGML_TYPE_F16
+#ifdef GGML_SYCL_HAS_BF16
+                    || op->type == GGML_TYPE_BF16
+#endif
+                   ) &&
+                   op->src[0]->type == op->type;
         case GGML_OP_CONV_3D:
             return op->type == GGML_TYPE_F32 &&
                    (op->src[0]->type == GGML_TYPE_F32 || op->src[0]->type == GGML_TYPE_F16) &&
@@ -5677,8 +5897,7 @@ static bool do_ggml_backend_sycl_device_supports_op(ggml_backend_dev_t dev, cons
         case GGML_OP_MEAN:
             return ggml_is_contiguous(op->src[0]);
         case GGML_OP_ARGSORT:
-            return op->src[0]->ne[0] * sizeof(int) <=
-                   ggml_sycl_info().devices[device].smpbo;
+            return true;
         case GGML_OP_TOP_K: {
             const ggml_tensor * src0 = op->src[0];
             const int k = op->ne[0];
@@ -5690,9 +5909,8 @@ static bool do_ggml_backend_sycl_device_supports_op(ggml_backend_dev_t dev, cons
         }
         case GGML_OP_POOL_2D:
         case GGML_OP_POOL_1D:
-            return true;
         case GGML_OP_ACC:
-            return ggml_is_contiguous(op->src[0]) && ggml_is_contiguous(op->src[1]);
+            return true;
         case GGML_OP_PAD:
             if (ggml_get_op_params_i32(op, 8) != 0) {
                 return false;
@@ -5725,6 +5943,8 @@ static bool do_ggml_backend_sycl_device_supports_op(ggml_backend_dev_t dev, cons
         case GGML_OP_FILL:
         case GGML_OP_CUMSUM:
         case GGML_OP_DIAG:
+        case GGML_OP_CROSS_ENTROPY_LOSS:
+        case GGML_OP_CROSS_ENTROPY_LOSS_BACK:
             return true;
         case GGML_OP_SOLVE_TRI:
             return op->src[0]->ne[0] <= SYCL_SOLVE_TRI_MAX_N && op->src[1]->ne[0] <= SYCL_SOLVE_TRI_MAX_K;

ggml/src/ggml-sycl/presets.hpp

@@ -19,6 +19,7 @@
 #define WARP_SIZE GGML_SYCL_WARP_SIZE
 #define MATRIX_ROW_PADDING 512 // last row of quant. matrices is a multiple of this to avoid out-of-bounds memory accesses
 
+#define SYCL_COL2IM_1D_BLOCK_SIZE 256
 #define SYCL_GELU_BLOCK_SIZE 256
 #define SYCL_SILU_BLOCK_SIZE 256
 #define SYCL_TANH_BLOCK_SIZE 256
@@ -62,7 +63,7 @@
 #endif
 
 #ifndef K_QUANTS_PER_ITERATION
-#define K_QUANTS_PER_ITERATION 2
+#define K_QUANTS_PER_ITERATION 1
 #else
 static_assert(K_QUANTS_PER_ITERATION == 1 || K_QUANTS_PER_ITERATION == 2, "K_QUANTS_PER_ITERATION must be 1 or 2");
 #endif

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

@@ -129,7 +129,7 @@ typedef struct VkPhysicalDeviceShaderMixedFloatDotProductFeaturesVALVE {
 #endif
 
 #define ROUNDUP_POW2(M, N) (((M) + (N) - 1) & ~((N) - 1))
-#define CEIL_DIV(M, N) (((M) + (N)-1) / (N))
+#define CEIL_DIV(M, N) (((M) / (N)) + (((M) % (N)) != 0))
 static bool is_pow2(uint32_t x) { return x > 1 && (x & (x-1)) == 0; }
 
 #define VK_VENDOR_ID_AMD 0x1002
@@ -17370,21 +17370,24 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
             return op->type == GGML_TYPE_F32 && op->src[0]->type == GGML_TYPE_F32;
         case GGML_OP_SET_ROWS:
             {
-                switch (op->type) {
-                    case GGML_TYPE_F32:
-                    case GGML_TYPE_F16:
-                    case GGML_TYPE_BF16:
-                    case GGML_TYPE_Q1_0:
-                    case GGML_TYPE_Q4_0:
-                    case GGML_TYPE_Q4_1:
-                    case GGML_TYPE_Q5_0:
-                    case GGML_TYPE_Q5_1:
-                    case GGML_TYPE_Q8_0:
-                    case GGML_TYPE_IQ4_NL:
-                        return true;
-                    default:
-                        return false;
+                if (op->src[0]->type == GGML_TYPE_F32) {
+                    switch (op->type) {
+                        case GGML_TYPE_F32:
+                        case GGML_TYPE_F16:
+                        case GGML_TYPE_BF16:
+                        case GGML_TYPE_Q1_0:
+                        case GGML_TYPE_Q4_0:
+                        case GGML_TYPE_Q4_1:
+                        case GGML_TYPE_Q5_0:
+                        case GGML_TYPE_Q5_1:
+                        case GGML_TYPE_Q8_0:
+                        case GGML_TYPE_IQ4_NL:
+                            return true;
+                        default:
+                            return false;
+                    }
                 }
+                return false;
             }
         case GGML_OP_CONT:
         case GGML_OP_CPY:

ggml/src/ggml.c

@@ -525,7 +525,11 @@ const char * ggml_commit(void) {
 
 #if defined(_MSC_VER) || defined(__MINGW32__)
 static int64_t timer_freq, timer_start;
-void ggml_time_init(void) {
+static BOOL CALLBACK ggml_time_init_once(PINIT_ONCE once, PVOID param, PVOID *ctx) {
+    UNUSED(once);
+    UNUSED(param);
+    UNUSED(ctx);
+
     LARGE_INTEGER t;
     QueryPerformanceFrequency(&t);
     timer_freq = t.QuadPart;
@@ -535,6 +539,12 @@ void ggml_time_init(void) {
     // We subtract the program start time to reduce the likelihood of that happening.
     QueryPerformanceCounter(&t);
     timer_start = t.QuadPart;
+
+    return TRUE;
+}
+void ggml_time_init(void) {
+    static INIT_ONCE once = INIT_ONCE_STATIC_INIT;
+    InitOnceExecuteOnce(&once, ggml_time_init_once, NULL, NULL);
 }
 int64_t ggml_time_ms(void) {
     LARGE_INTEGER t;

scripts/ui-assets.cmake

@@ -186,6 +186,12 @@ function(hf_download version out_var out_resolved)
 
     set(archive "${UI_BINARY_DIR}/dist.tar.gz")
 
+    # Use HF_TOKEN to benefit from higher rate limits
+    set(auth_headers "")
+    if(DEFINED ENV{HF_TOKEN} AND NOT "$ENV{HF_TOKEN}" STREQUAL "")
+        list(APPEND auth_headers "HTTPHEADER" "Authorization: Bearer $ENV{HF_TOKEN}")
+    endif()
+
     set(candidates "")
     if(NOT "${version}" STREQUAL "")
         list(APPEND candidates "${version}")
@@ -198,7 +204,7 @@ function(hf_download version out_var out_resolved)
         message(STATUS "UI: downloading from ${resolved}: ${base}/dist.tar.gz")
 
         file(DOWNLOAD "${base}/dist.tar.gz?download=true" "${archive}"
-            STATUS status TIMEOUT 300
+            STATUS status TIMEOUT 300 ${auth_headers}
         )
         list(GET status 0 rc)
         if(NOT rc EQUAL 0)
@@ -208,7 +214,7 @@ function(hf_download version out_var out_resolved)
         endif()
 
         file(DOWNLOAD "${base}/dist.tar.gz.sha256?download=true" "${archive}.sha256"
-            STATUS status TIMEOUT 30
+            STATUS status TIMEOUT 30 ${auth_headers}
         )
         list(GET status 0 rc)
         if(NOT rc EQUAL 0)

src/llama-graph.cpp

@@ -63,26 +63,6 @@ static bool can_reuse_kq_mask(
 
 // impl
 
-static ggml_tensor * ggml_mul_mat_aux(
-        ggml_context * ctx,
-        ggml_tensor * cur,
-        ggml_tensor * rot) {
-    const auto n = rot->ne[0];
-
-    ggml_tensor * res;
-
-    if (!ggml_is_contiguous(cur)) {
-        res = ggml_cont_2d   (ctx, cur, n, ggml_nelements(cur)/n);
-    } else {
-        res = ggml_reshape_2d(ctx, cur, n, ggml_nelements(cur)/n);
-    }
-    res = ggml_mul_mat   (ctx, rot, res);
-    ggml_mul_mat_set_hint(res, GGML_HINT_SRC0_IS_HADAMARD);
-    res = ggml_reshape_4d(ctx, res, cur->ne[0], cur->ne[1], cur->ne[2], cur->ne[3]);
-
-    return res;
-}
-
 void llm_graph_input_embd::set_input(const llama_ubatch * ubatch) {
     if (ubatch->token) {
         const int64_t n_tokens = ubatch->n_tokens;
@@ -881,6 +861,14 @@ void llm_graph_input_dsv4::set_input(const llama_ubatch * ubatch) {
     dsv4_set_comp_inputs(inp_hca, plan_hca, "hca", debug > 0, ubatch->n_tokens, n_stream);
     dsv4_set_comp_inputs(inp_lid, plan_lid, "lid", debug > 0, ubatch->n_tokens, n_stream);
 
+    if (inp_csa.k_rot && inp_csa.k_rot->buffer) {
+        mctx->get_csa()->set_input_k_rot(inp_csa.k_rot);
+    }
+
+    if (inp_hca.k_rot && inp_hca.k_rot->buffer) {
+        mctx->get_hca()->set_input_k_rot(inp_hca.k_rot);
+    }
+
     if (inp_lid.k_rot && inp_lid.k_rot->buffer) {
         mctx->get_lid()->set_input_k_rot(inp_lid.k_rot);
     }
@@ -2633,12 +2621,12 @@ ggml_tensor * llm_graph_context::build_attn(
     GGML_ASSERT(v_mla == nullptr);
 
     if (inp->self_k_rot) {
-        q_cur = ggml_mul_mat_aux(ctx0, q_cur, inp->self_k_rot);
-        k_cur = ggml_mul_mat_aux(ctx0, k_cur, inp->self_k_rot);
+        q_cur = llama_mul_mat_hadamard(ctx0, q_cur, inp->self_k_rot);
+        k_cur = llama_mul_mat_hadamard(ctx0, k_cur, inp->self_k_rot);
     }
 
     if (inp->self_v_rot) {
-        v_cur = ggml_mul_mat_aux(ctx0, v_cur, inp->self_v_rot);
+        v_cur = llama_mul_mat_hadamard(ctx0, v_cur, inp->self_v_rot);
     }
 
     // these nodes are added to the graph together so that they are not reordered
@@ -2669,7 +2657,7 @@ ggml_tensor * llm_graph_context::build_attn(
     cb(cur, "kqv_out", il);
 
     if (inp->self_v_rot) {
-        cur = ggml_mul_mat_aux(ctx0, cur, inp->self_v_rot);
+        cur = llama_mul_mat_hadamard(ctx0, cur, inp->self_v_rot);
     }
 
     if (wo) {
@@ -2874,14 +2862,14 @@ ggml_tensor * llm_graph_context::build_attn(
     auto * v_rot = is_swa ? inp->self_v_rot_swa : inp->self_v_rot;
 
     if (k_rot) {
-        q_cur = ggml_mul_mat_aux(ctx0, q_cur, k_rot);
+        q_cur = llama_mul_mat_hadamard(ctx0, q_cur, k_rot);
         if (k_cur) {
-            k_cur = ggml_mul_mat_aux(ctx0, k_cur, k_rot);
+            k_cur = llama_mul_mat_hadamard(ctx0, k_cur, k_rot);
         }
     }
     if (v_rot) {
         if (v_cur) {
-            v_cur = ggml_mul_mat_aux(ctx0, v_cur, v_rot);
+            v_cur = llama_mul_mat_hadamard(ctx0, v_cur, v_rot);
         }
     }
 
@@ -2924,7 +2912,7 @@ ggml_tensor * llm_graph_context::build_attn(
     cb(cur, "kqv_out", il);
 
     if (v_rot) {
-        cur = ggml_mul_mat_aux(ctx0, cur, v_rot);
+        cur = llama_mul_mat_hadamard(ctx0, cur, v_rot);
     }
 
     if (wo) {
@@ -3084,6 +3072,8 @@ llm_graph_input_dsv4 * llm_graph_context::build_inp_dsv4() const {
     dsv4_build_comp_inputs(ctx0, inp->inp_csa, mctx_cur->get_csa_plan(ubatch), "csa", n_stream);
     dsv4_build_comp_inputs(ctx0, inp->inp_hca, mctx_cur->get_hca_plan(ubatch), "hca", n_stream);
     dsv4_build_comp_inputs(ctx0, inp->inp_lid, mctx_cur->get_lid_plan(ubatch), "lid", n_stream);
+    inp->inp_csa.k_rot = mctx_cur->get_csa()->build_input_k_rot(ctx0);
+    inp->inp_hca.k_rot = mctx_cur->get_hca()->build_input_k_rot(ctx0);
     inp->inp_lid.k_rot = mctx_cur->get_lid()->build_input_k_rot(ctx0);
 
     return (llm_graph_input_dsv4 *) res->add_input(std::move(inp));

src/llama-impl.h

@@ -54,6 +54,26 @@ static inline dst_t llama_cast(src_t v) {
     }
 }
 
+static inline ggml_tensor * llama_mul_mat_hadamard(
+        ggml_context * ctx,
+        ggml_tensor * cur,
+        ggml_tensor * rot) {
+    const auto n = rot->ne[0];
+
+    ggml_tensor * res;
+
+    if (!ggml_is_contiguous(cur)) {
+        res = ggml_cont_2d(ctx, cur, n, ggml_nelements(cur)/n);
+    } else {
+        res = ggml_reshape_2d(ctx, cur, n, ggml_nelements(cur)/n);
+    }
+    res = ggml_mul_mat(ctx, rot, res);
+    ggml_mul_mat_set_hint(res, GGML_HINT_SRC0_IS_HADAMARD);
+    res = ggml_reshape_4d(ctx, res, cur->ne[0], cur->ne[1], cur->ne[2], cur->ne[3]);
+
+    return res;
+}
+
 struct time_meas {
     time_meas(int64_t & t_acc, bool disable = false);
     ~time_meas();

src/llama-kv-cache.cpp

@@ -57,22 +57,6 @@ static void ggml_gen_hadamard(ggml_tensor * tensor) {
     }
 }
 
-static ggml_tensor * ggml_mul_mat_aux(
-        ggml_context * ctx,
-        ggml_tensor * cur,
-        ggml_tensor * rot) {
-    const auto n = rot->ne[0];
-
-    ggml_tensor * res;
-
-    res = ggml_reshape_2d(ctx, cur, n, ggml_nelements(cur)/n);
-    res = ggml_mul_mat   (ctx, rot, res);
-    ggml_mul_mat_set_hint(res, GGML_HINT_SRC0_IS_HADAMARD);
-    res = ggml_reshape_4d(ctx, res, cur->ne[0], cur->ne[1], cur->ne[2], cur->ne[3]);
-
-    return res;
-}
-
 //
 // llama_kv_cache
 //
@@ -1875,14 +1859,14 @@ ggml_tensor * llama_kv_cache::build_rope_shift(
         tmp = ggml_cast(ctx, cur, GGML_TYPE_F32);
 
         // rotate back
-        tmp = ggml_mul_mat_aux(ctx, tmp, rot);
+        tmp = llama_mul_mat_hadamard(ctx, tmp, rot);
 
         tmp = ggml_rope_ext(ctx, tmp,
                 shift, factors, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
                 yarn_ext_factor, yarn_attn_factor, yarn_beta_fast, yarn_beta_slow);
 
         // rotate fwd
-        tmp = ggml_mul_mat_aux(ctx, tmp, rot);
+        tmp = llama_mul_mat_hadamard(ctx, tmp, rot);
 
         tmp = ggml_cpy(ctx, tmp, cur);
     } else {

src/llama-model.cpp

@@ -953,6 +953,8 @@ static buft_list_t make_gpu_buft_list(ggml_backend_dev_t dev, llama_split_mode s
             if (buft != nullptr) {
                 buft_list.emplace_back(dev, buft);
             }
+        } else {
+            throw std::runtime_error(format("device %s does not support split buffers", ggml_backend_dev_name(dev)));
         }
     }
 
@@ -1012,9 +1014,17 @@ struct llama_model::impl {
     std::vector<layer_dev> dev_layer;
 
     bool has_tensor_overrides;
+
+    std::vector<float> tensor_split_owned;
 };
 
 llama_model::llama_model(const llama_model_params & params) : params(params), pimpl(std::make_unique<impl>()) {
+    if (params.tensor_split != nullptr) {
+        // llama_model_params stores tensor_split as a borrowed pointer, but the model
+        // may need it later for tensor-parallel KV-cache split metadata.
+        pimpl->tensor_split_owned.assign(params.tensor_split, params.tensor_split + llama_max_devices());
+        this->params.tensor_split = pimpl->tensor_split_owned.data();
+    }
     pimpl->has_tensor_overrides = params.tensor_buft_overrides && params.tensor_buft_overrides[0].pattern;
 }
 

src/models/deepseek4.cpp

@@ -557,7 +557,7 @@ ggml_tensor * llama_model_deepseek4::graph::build_lid_top_k(
     cb(indexer_q_pe, "lid_q_pe", il);
 
     indexer_q = ggml_concat(ctx0, indexer_q_nope, indexer_q_pe, 0);
-    indexer_q = ggml_mul_mat(ctx0, inp_lid.k_rot, indexer_q);
+    indexer_q = llama_mul_mat_hadamard(ctx0, indexer_q, inp_lid.k_rot);
     cb(indexer_q, "lid_q_rot", il);
 
     ggml_tensor * indexer_weights = build_lora_mm(layer.indexer_proj, cur);
@@ -652,10 +652,15 @@ ggml_tensor * llama_model_deepseek4::graph::build_csa_lid_attention(
         int il) const {
     const auto & inp_csa = inp_dsv4->get_csa();
     GGML_ASSERT(inp_csa.kq_mask);
-    GGML_ASSERT(inp_attn->self_k_rot == nullptr);
 
     ggml_tensor * top_k = build_lid_top_k(model, inp_dsv4, qr, cur, inp_pos, il);
 
+    ggml_tensor * k_rot = inp_attn->self_k_rot;
+    if (k_rot) {
+        q  = llama_mul_mat_hadamard(ctx0, q, k_rot);
+        kv = llama_mul_mat_hadamard(ctx0, kv, k_rot);
+    }
+
     ggml_build_forward_expand(gf, q);
     ggml_build_forward_expand(gf, kv);
 
@@ -696,6 +701,9 @@ ggml_tensor * llama_model_deepseek4::graph::build_csa_lid_attention(
 
     ggml_tensor * kq_b = dsv4_build_kq_zero_bias(ctx0, cparams, kq_mask, q->ne[1]);
     ggml_tensor * out = build_attn_mha(q, k_all, k_all, kq_b, kq_mask, sinks, nullptr, kq_scale, il);
+    if (k_rot) {
+        out = llama_mul_mat_hadamard(ctx0, out, k_rot);
+    }
     cb(out, "attn_csa_lid", il);
 
     return out;
@@ -711,7 +719,12 @@ ggml_tensor * llama_model_deepseek4::graph::build_hca_attention(
         int il) const {
     const auto & inp_hca = inp_dsv4->get_hca();
     GGML_ASSERT(inp_hca.kq_mask);
-    GGML_ASSERT(inp_attn->self_k_rot == nullptr);
+
+    ggml_tensor * k_rot = inp_attn->self_k_rot;
+    if (k_rot) {
+        q  = llama_mul_mat_hadamard(ctx0, q, k_rot);
+        kv = llama_mul_mat_hadamard(ctx0, kv, k_rot);
+    }
 
     ggml_build_forward_expand(gf, q);
     ggml_build_forward_expand(gf, kv);
@@ -753,6 +766,9 @@ ggml_tensor * llama_model_deepseek4::graph::build_hca_attention(
 
     ggml_tensor * kq_b = dsv4_build_kq_zero_bias(ctx0, cparams, kq_mask, q->ne[1]);
     ggml_tensor * out = build_attn_mha(q, k_all, k_all, kq_b, kq_mask, sinks, nullptr, kq_scale, il);
+    if (k_rot) {
+        out = llama_mul_mat_hadamard(ctx0, out, k_rot);
+    }
     cb(out, "attn_hca", il);
 
     return out;
@@ -770,8 +786,8 @@ ggml_tensor * llama_model_deepseek4::graph::build_raw_attention(
     ggml_tensor * k_rot = inp_attn->self_k_rot;
 
     if (k_rot) {
-        q  = ggml_mul_mat(ctx0, k_rot, q);
-        kv = ggml_mul_mat(ctx0, k_rot, kv);
+        q  = llama_mul_mat_hadamard(ctx0, q, k_rot);
+        kv = llama_mul_mat_hadamard(ctx0, kv, k_rot);
     }
 
     ggml_build_forward_expand(gf, q);
@@ -788,6 +804,9 @@ ggml_tensor * llama_model_deepseek4::graph::build_raw_attention(
 
     ggml_tensor * kq_b = dsv4_build_kq_zero_bias(ctx0, cparams, kq_mask, q->ne[1]);
     ggml_tensor * out = build_attn_mha(q, k, k, kq_b, kq_mask, sinks, nullptr, kq_scale, il);
+    if (k_rot) {
+        out = llama_mul_mat_hadamard(ctx0, out, k_rot);
+    }
     cb(out, "attn_raw", il);
 
     return out;
@@ -917,6 +936,11 @@ ggml_tensor * llama_model_deepseek4::graph::build_attention(
                 "csa_state_compress",
                 il);
 
+        if (inp_dsv4->get_csa().k_rot) {
+            kv_comp_csa_state = llama_mul_mat_hadamard(ctx0, kv_comp_csa_state, inp_dsv4->get_csa().k_rot);
+            cb(kv_comp_csa_state, "csa_state_compress_rot", il);
+        }
+
         ggml_build_forward_expand(gf, inp_dsv4->mctx->get_csa()->cpy_k(ctx0,
                     kv_comp_csa_state, inp_dsv4->get_csa().state_write_idxs, il));
 
@@ -965,7 +989,7 @@ ggml_tensor * llama_model_deepseek4::graph::build_attention(
                 il);
 
         if (inp_dsv4->get_lid().k_rot) {
-            kv_comp_lid_state = ggml_mul_mat(ctx0, inp_dsv4->get_lid().k_rot, kv_comp_lid_state);
+            kv_comp_lid_state = llama_mul_mat_hadamard(ctx0, kv_comp_lid_state, inp_dsv4->get_lid().k_rot);
             cb(kv_comp_lid_state, "lid_state_compress_rot", il);
         }
 
@@ -1007,6 +1031,11 @@ ggml_tensor * llama_model_deepseek4::graph::build_attention(
                 "hca_state_compress",
                 il);
 
+        if (inp_dsv4->get_hca().k_rot) {
+            kv_comp_hca = llama_mul_mat_hadamard(ctx0, kv_comp_hca, inp_dsv4->get_hca().k_rot);
+            cb(kv_comp_hca, "hca_state_compress_rot", il);
+        }
+
         ggml_build_forward_expand(gf, inp_dsv4->mctx->get_hca()->cpy_k(ctx0,
                     kv_comp_hca, inp_dsv4->get_hca().state_write_idxs, il));
         hca_state_dep = kv_comp_hca;
@@ -1035,13 +1064,11 @@ ggml_tensor * llama_model_deepseek4::graph::build_attention(
     if (ratio == DSV4_CSA_RATIO &&
             inp_dsv4->get_csa().kq_mask &&
             inp_dsv4->get_lid().kq_mask &&
-            inp_dsv4->get_lid().k_rot &&
-            inp_attn->self_k_rot == nullptr) {
+            inp_dsv4->get_lid().k_rot) {
         out = build_csa_lid_attention(model, inp_dsv4, inp_attn, q, kv, qr, cur, inp_pos, layer.attn_sinks,
                 1.0f/sqrtf(float(n_embd_head)), il);
     } else if (ratio == DSV4_HCA_RATIO &&
-            inp_dsv4->get_hca().kq_mask &&
-            inp_attn->self_k_rot == nullptr) {
+            inp_dsv4->get_hca().kq_mask) {
         out = build_hca_attention(inp_dsv4, inp_attn, q, kv, layer.attn_sinks,
                 1.0f/sqrtf(float(n_embd_head)), il);
     } else {

tools/server/server-models.cpp

@@ -523,6 +523,7 @@ void server_models::load_models() {
 
         // collect all threads to join in one pass while the lock is held:
         // - monitoring threads from just-unloaded models (to_unload)
+        // - threads of finished downloads (DOWNLOADED), they acquire the mutex on exit
         // - threads of already-UNLOADED models that are being removed from source
         std::vector<std::thread> threads_to_join;
         for (const auto & name : to_unload) {
@@ -535,6 +536,13 @@ void server_models::load_models() {
             if (inst.meta.status == SERVER_MODEL_STATUS_DOWNLOADING) {
                 continue; // downloading models are not from config sources, leave them alone
             }
+            if (inst.meta.status == SERVER_MODEL_STATUS_DOWNLOADED) {
+                // joining this thread under the lock deadlocks: it locks the mutex on its way out
+                if (inst.th.joinable()) {
+                    threads_to_join.push_back(std::move(inst.th));
+                }
+                continue;
+            }
             if (final_presets.find(name) == final_presets.end() && !inst.meta.is_running() && inst.th.joinable()) {
                 threads_to_join.push_back(std::move(inst.th));
             }
@@ -550,10 +558,8 @@ void server_models::load_models() {
             if (it->second.meta.status == SERVER_MODEL_STATUS_DOWNLOADING) {
                 ++it; // download thread is still busy, skip
             } else if (it->second.meta.status == SERVER_MODEL_STATUS_DOWNLOADED) {
-                // download finished, safe to erase
-                if (it->second.th.joinable()) {
-                    it->second.th.join();
-                }
+                // download finished, thread is joined above, safe to erase
+                GGML_ASSERT(!it->second.th.joinable());
                 it = mapping.erase(it);
             } else if (final_presets.find(it->first) == final_presets.end()) {
                 SRV_INF("(reload) removing model name=%s (no longer in source)\n", it->first.c_str());

tools/server/tests/utils.py

@@ -31,6 +31,9 @@ import wget
 
 DEFAULT_HTTP_TIMEOUT = 60
 
+# per-request timeout, a hung server fails the test instead of stalling the CI for hours
+DEFAULT_REQUEST_TIMEOUT = 600
+
 
 class ServerResponse:
     headers: dict
@@ -330,7 +333,7 @@ class ServerProcess:
         path: str,
         data: dict | Any | None = None,
         headers: dict | None = None,
-        timeout: float | None = None,
+        timeout: float | None = DEFAULT_REQUEST_TIMEOUT,
     ) -> ServerResponse:
         url = f"http://{self.server_host}:{self.server_port}{path}"
         parse_body = False
@@ -389,7 +392,7 @@ class ServerProcess:
         path: str,
         data: dict | None = None,
         headers: dict | None = None,
-        timeout: float | None = None,
+        timeout: float | None = DEFAULT_REQUEST_TIMEOUT,
     ) -> dict:
         stream = data.get('stream', False)
         if stream:

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

@@ -27,7 +27,10 @@
 
 	let { onSearchClick = () => {} }: Props = $props();
 
-	const { handleKeydown } = useKeyboardShortcuts({ activateSearchMode: () => onSearchClick() });
+	const { handleKeydown } = useKeyboardShortcuts({
+		activateSearchMode: () => onSearchClick(),
+		toggleSidebar: () => toggleExpandedMode()
+	});
 
 	let isExpandedMode = $state(false);
 	let hoveredTooltip = $state<string | null>(null);

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

@@ -9,6 +9,7 @@ export enum KeyboardKey {
 	ARROW_LEFT = 'ArrowLeft',
 	ARROW_RIGHT = 'ArrowRight',
 	TAB = 'Tab',
+	B_LOWER = 'b',
 	D_LOWER = 'd',
 	D_UPPER = 'D',
 	E_UPPER = 'E',

tools/ui/src/lib/hooks/use-keyboard-shortcuts.svelte.ts

@@ -9,6 +9,7 @@ interface KeyboardShortcutsCallbacks {
 	deleteActiveConversation?: () => void;
 	navigateToPrevConversation?: () => void;
 	navigateToNextConversation?: () => void;
+	toggleSidebar?: () => void;
 }
 
 export function useKeyboardShortcuts(callbacks: KeyboardShortcutsCallbacks) {
@@ -21,6 +22,11 @@ export function useKeyboardShortcuts(callbacks: KeyboardShortcutsCallbacks) {
 			callbacks.onSearchActivated?.();
 		}
 
+		if (isCmdOrCtrl && event.key === KeyboardKey.B_LOWER) {
+			event.preventDefault();
+			callbacks.toggleSidebar?.();
+		}
+
 		if (
 			isCmdOrCtrl &&
 			event.shiftKey &&

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

@@ -314,6 +314,30 @@ export class MCPService {
 					)
 				);
 
+				if (method === 'DELETE' && url.includes(CORS_PROXY_ENDPOINT)) {
+					const response = new Response(null, { status: 200, statusText: 'OK' });
+
+					logIfEnabled(
+						this.createLog(
+							MCPConnectionPhase.INITIALIZING,
+							`HTTP 200 ${method} ${url} (fake response)`,
+							MCPLogLevel.INFO,
+							{
+								response: {
+									url,
+									status: response.status,
+									statusText: response.statusText,
+									durationMs: 0,
+									isFake: true
+								}
+							}
+						)
+					);
+
+					// fake response, bypass real fetch()
+					return response;
+				}
+
 				try {
 					const response = await fetch(input, {
 						...baseInit,

tools/ui/tests/unit/mcp-service.test.ts

@@ -154,6 +154,32 @@ describe('MCPService', () => {
 		});
 	});
 
+	it('DELETE request with CORS proxy should return a fake 200 response', async () => {
+		const logs: MCPConnectionLog[] = [];
+		const fetchMock = vi.fn();
+
+		vi.stubGlobal('fetch', fetchMock);
+
+		const config: MCPServerConfig = {
+			url: 'https://example.com/mcp',
+			transport: MCPTransportType.STREAMABLE_HTTP,
+			useProxy: true
+		};
+
+		const controller = createDiagnosticFetch(config, (log) => logs.push(log), {}, true);
+
+		const response = await controller.fetch(
+			'http://localhost:8080/cors-proxy?url=https%3A%2F%2Fexample.com%2Fmcp',
+			{ method: 'DELETE' }
+		);
+
+		expect(fetchMock).not.toHaveBeenCalled();
+		expect(response.status).toBe(200);
+		expect(logs.at(-1)?.details).toMatchObject({
+			response: { status: 200, isFake: true }
+		});
+	});
+
 	it('partially redacts mcp-session-id in diagnostic request and response logs', async () => {
 		const logs: MCPConnectionLog[] = [];
 		const response = new Response('{}', {