npm - @fugood/llama.node - Versions diffs - 0.3.16 → 0.4.0 - Mend

@fugood/llama.node 0.3.16 → 0.4.0

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.

Files changed (281) hide show

package/CMakeLists.txt +6 -1
package/bin/darwin/arm64/llama-node.node +0 -0
package/bin/darwin/x64/llama-node.node +0 -0
package/bin/linux/arm64/llama-node.node +0 -0
package/bin/linux/x64/llama-node.node +0 -0
package/bin/linux-cuda/arm64/llama-node.node +0 -0
package/bin/linux-cuda/x64/llama-node.node +0 -0
package/bin/linux-vulkan/arm64/llama-node.node +0 -0
package/bin/linux-vulkan/x64/llama-node.node +0 -0
package/bin/win32/arm64/llama-node.node +0 -0
package/bin/win32/arm64/node.lib +0 -0
package/bin/win32/x64/llama-node.node +0 -0
package/bin/win32/x64/node.lib +0 -0
package/bin/win32-vulkan/arm64/llama-node.node +0 -0
package/bin/win32-vulkan/arm64/node.lib +0 -0
package/bin/win32-vulkan/x64/llama-node.node +0 -0
package/bin/win32-vulkan/x64/node.lib +0 -0
package/lib/binding.ts +44 -2
package/lib/index.js +132 -1
package/lib/index.ts +203 -3
package/package.json +2 -1
package/src/EmbeddingWorker.cpp +1 -1
package/src/LlamaCompletionWorker.cpp +374 -19
package/src/LlamaCompletionWorker.h +31 -10
package/src/LlamaContext.cpp +216 -7
package/src/LlamaContext.h +12 -0
package/src/common.hpp +15 -0
package/src/llama.cpp/.github/workflows/build-linux-cross.yml +233 -0
package/src/llama.cpp/.github/workflows/build.yml +89 -767
package/src/llama.cpp/.github/workflows/docker.yml +9 -6
package/src/llama.cpp/.github/workflows/release.yml +716 -0
package/src/llama.cpp/.github/workflows/server.yml +19 -23
package/src/llama.cpp/CMakeLists.txt +11 -1
package/src/llama.cpp/cmake/build-info.cmake +8 -2
package/src/llama.cpp/cmake/x64-windows-llvm.cmake +0 -6
package/src/llama.cpp/common/CMakeLists.txt +35 -4
package/src/llama.cpp/common/arg.cpp +844 -121
package/src/llama.cpp/common/arg.h +9 -0
package/src/llama.cpp/common/chat.cpp +129 -107
package/src/llama.cpp/common/chat.h +2 -0
package/src/llama.cpp/common/common.cpp +64 -518
package/src/llama.cpp/common/common.h +35 -45
package/src/llama.cpp/common/json-schema-to-grammar.cpp +3 -0
package/src/llama.cpp/common/llguidance.cpp +31 -47
package/src/llama.cpp/common/minja/chat-template.hpp +23 -11
package/src/llama.cpp/common/minja/minja.hpp +186 -127
package/src/llama.cpp/common/regex-partial.cpp +204 -0
package/src/llama.cpp/common/regex-partial.h +56 -0
package/src/llama.cpp/common/sampling.cpp +60 -50
package/src/llama.cpp/docs/build.md +122 -7
package/src/llama.cpp/examples/CMakeLists.txt +2 -32
package/src/llama.cpp/examples/batched/batched.cpp +1 -1
package/src/llama.cpp/examples/embedding/embedding.cpp +9 -12
package/src/llama.cpp/examples/gritlm/gritlm.cpp +1 -1
package/src/llama.cpp/examples/llama.android/llama/build.gradle.kts +1 -0
package/src/llama.cpp/examples/parallel/parallel.cpp +89 -15
package/src/llama.cpp/examples/passkey/passkey.cpp +1 -1
package/src/llama.cpp/examples/speculative/speculative.cpp +1 -1
package/src/llama.cpp/examples/speculative-simple/speculative-simple.cpp +1 -1
package/src/llama.cpp/examples/sycl/build.sh +2 -2
package/src/llama.cpp/examples/sycl/win-build-sycl.bat +2 -2
package/src/llama.cpp/examples/training/CMakeLists.txt +5 -0
package/src/llama.cpp/examples/training/finetune.cpp +96 -0
package/src/llama.cpp/ggml/CMakeLists.txt +35 -2
package/src/llama.cpp/ggml/cmake/GitVars.cmake +22 -0
package/src/llama.cpp/ggml/include/ggml-backend.h +4 -4
package/src/llama.cpp/ggml/include/ggml-cpp.h +1 -1
package/src/llama.cpp/ggml/include/ggml-cpu.h +5 -0
package/src/llama.cpp/ggml/include/ggml-opt.h +47 -28
package/src/llama.cpp/ggml/include/ggml-rpc.h +6 -1
package/src/llama.cpp/ggml/include/ggml.h +76 -106
package/src/llama.cpp/ggml/src/CMakeLists.txt +11 -8
package/src/llama.cpp/ggml/src/ggml-alloc.c +4 -1
package/src/llama.cpp/ggml/src/ggml-backend.cpp +9 -5
package/src/llama.cpp/ggml/src/ggml-cann/CMakeLists.txt +0 -2
package/src/llama.cpp/ggml/src/ggml-cann/acl_tensor.cpp +8 -4
package/src/llama.cpp/ggml/src/ggml-cann/acl_tensor.h +5 -5
package/src/llama.cpp/ggml/src/ggml-cann/aclnn_ops.cpp +692 -1534
package/src/llama.cpp/ggml/src/ggml-cann/aclnn_ops.h +613 -122
package/src/llama.cpp/ggml/src/ggml-cann/common.h +135 -1
package/src/llama.cpp/ggml/src/ggml-cann/ggml-cann.cpp +507 -137
package/src/llama.cpp/ggml/src/ggml-common.h +12 -6
package/src/llama.cpp/ggml/src/ggml-cpu/CMakeLists.txt +66 -33
package/src/llama.cpp/ggml/src/ggml-cpu/binary-ops.cpp +158 -0
package/src/llama.cpp/ggml/src/ggml-cpu/binary-ops.h +16 -0
package/src/llama.cpp/ggml/src/ggml-cpu/common.h +72 -0
package/src/llama.cpp/ggml/src/ggml-cpu/cpu-feats-x86.cpp +1 -1
package/src/llama.cpp/ggml/src/ggml-cpu/ggml-cpu-aarch64.cpp +896 -194
package/src/llama.cpp/ggml/src/ggml-cpu/ggml-cpu-impl.h +2 -21
package/src/llama.cpp/ggml/src/ggml-cpu/ggml-cpu-quants.c +1060 -410
package/src/llama.cpp/ggml/src/ggml-cpu/ggml-cpu.c +1008 -13533
package/src/llama.cpp/ggml/src/ggml-cpu/ggml-cpu.cpp +31 -16
package/src/llama.cpp/ggml/src/ggml-cpu/kleidiai/kernels.cpp +90 -12
package/src/llama.cpp/ggml/src/ggml-cpu/kleidiai/kernels.h +47 -13
package/src/llama.cpp/ggml/src/ggml-cpu/kleidiai/kleidiai.cpp +266 -72
package/src/llama.cpp/ggml/src/ggml-cpu/llamafile/sgemm.cpp +1034 -88
package/src/llama.cpp/ggml/src/ggml-cpu/ops.cpp +8796 -0
package/src/llama.cpp/ggml/src/ggml-cpu/ops.h +110 -0
package/src/llama.cpp/ggml/src/ggml-cpu/simd-mappings.h +892 -0
package/src/llama.cpp/ggml/src/ggml-cpu/unary-ops.cpp +186 -0
package/src/llama.cpp/ggml/src/ggml-cpu/unary-ops.h +28 -0
package/src/llama.cpp/ggml/src/ggml-cpu/vec.cpp +252 -0
package/src/llama.cpp/ggml/src/ggml-cpu/vec.h +802 -0
package/src/llama.cpp/ggml/src/ggml-cuda/CMakeLists.txt +23 -4
package/src/llama.cpp/ggml/src/ggml-cuda/vendors/hip.h +7 -0
package/src/llama.cpp/ggml/src/ggml-cuda/vendors/musa.h +1 -0
package/src/llama.cpp/ggml/src/ggml-hip/CMakeLists.txt +0 -4
package/src/llama.cpp/ggml/src/ggml-impl.h +52 -18
package/src/llama.cpp/ggml/src/ggml-metal/ggml-metal-impl.h +106 -14
package/src/llama.cpp/ggml/src/ggml-opencl/CMakeLists.txt +67 -119
package/src/llama.cpp/ggml/src/ggml-opencl/ggml-opencl.cpp +1023 -262
package/src/llama.cpp/ggml/src/ggml-opt.cpp +368 -190
package/src/llama.cpp/ggml/src/ggml-quants.c +0 -6
package/src/llama.cpp/ggml/src/ggml-rpc/ggml-rpc.cpp +307 -40
package/src/llama.cpp/ggml/src/ggml-sycl/CMakeLists.txt +125 -45
package/src/llama.cpp/ggml/src/ggml-sycl/backend.hpp +10 -8
package/src/llama.cpp/ggml/src/ggml-sycl/binbcast.cpp +239 -0
package/src/llama.cpp/ggml/src/ggml-sycl/binbcast.hpp +39 -0
package/src/llama.cpp/ggml/src/ggml-sycl/common.cpp +0 -35
package/src/llama.cpp/ggml/src/ggml-sycl/common.hpp +9 -307
package/src/llama.cpp/ggml/src/ggml-sycl/convert.cpp +72 -25
package/src/llama.cpp/ggml/src/ggml-sycl/convert.hpp +14 -7
package/src/llama.cpp/ggml/src/ggml-sycl/dequantize.hpp +59 -21
package/src/llama.cpp/ggml/src/ggml-sycl/dmmv.cpp +7 -1
package/src/llama.cpp/ggml/src/ggml-sycl/dpct/helper.hpp +79 -90
package/src/llama.cpp/ggml/src/ggml-sycl/element_wise.cpp +944 -438
package/src/llama.cpp/ggml/src/ggml-sycl/element_wise.hpp +22 -23
package/src/llama.cpp/ggml/src/ggml-sycl/gemm.hpp +37 -8
package/src/llama.cpp/ggml/src/ggml-sycl/getrows.cpp +24 -20
package/src/llama.cpp/ggml/src/ggml-sycl/getrows.hpp +1 -4
package/src/llama.cpp/ggml/src/ggml-sycl/ggml-sycl.cpp +507 -411
package/src/llama.cpp/ggml/src/ggml-sycl/im2col.cpp +84 -74
package/src/llama.cpp/ggml/src/ggml-sycl/im2col.hpp +1 -3
package/src/llama.cpp/ggml/src/ggml-sycl/mmvq.cpp +185 -89
package/src/llama.cpp/ggml/src/ggml-sycl/norm.cpp +37 -49
package/src/llama.cpp/ggml/src/ggml-sycl/norm.hpp +7 -22
package/src/llama.cpp/ggml/src/ggml-sycl/outprod.cpp +4 -14
package/src/llama.cpp/ggml/src/ggml-sycl/quants.hpp +83 -0
package/src/llama.cpp/ggml/src/ggml-sycl/rope.cpp +204 -118
package/src/llama.cpp/ggml/src/ggml-sycl/rope.hpp +1 -3
package/src/llama.cpp/ggml/src/ggml-sycl/vecdotq.hpp +128 -53
package/src/llama.cpp/ggml/src/ggml-vulkan/CMakeLists.txt +83 -49
package/src/llama.cpp/ggml/src/ggml-vulkan/ggml-vulkan.cpp +1278 -282
package/src/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/CMakeLists.txt +32 -0
package/src/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp +133 -30
package/src/llama.cpp/ggml/src/ggml.c +170 -265
package/src/llama.cpp/ggml/src/gguf.cpp +34 -33
package/src/llama.cpp/include/llama.h +82 -22
package/src/llama.cpp/models/ggml-vocab-llama4.gguf.inp +112 -0
package/src/llama.cpp/models/ggml-vocab-llama4.gguf.out +46 -0
package/src/llama.cpp/models/ggml-vocab-pixtral.gguf.inp +112 -0
package/src/llama.cpp/models/ggml-vocab-pixtral.gguf.out +46 -0
package/src/llama.cpp/requirements/requirements-all.txt +5 -3
package/src/llama.cpp/requirements/requirements-gguf_editor_gui.txt +3 -0
package/src/llama.cpp/scripts/xxd.cmake +1 -1
package/src/llama.cpp/src/CMakeLists.txt +4 -2
package/src/llama.cpp/src/llama-adapter.cpp +43 -1
package/src/llama.cpp/src/llama-arch.cpp +163 -17
package/src/llama.cpp/src/llama-arch.h +16 -0
package/src/llama.cpp/src/llama-batch.cpp +5 -1
package/src/llama.cpp/src/llama-batch.h +2 -1
package/src/llama.cpp/src/llama-chat.cpp +91 -16
package/src/llama.cpp/src/llama-chat.h +7 -2
package/src/llama.cpp/src/llama-context.cpp +479 -575
package/src/llama.cpp/src/llama-context.h +44 -33
package/src/llama.cpp/src/llama-cparams.h +1 -0
package/src/llama.cpp/src/llama-graph.cpp +209 -157
package/src/llama.cpp/src/llama-graph.h +38 -14
package/src/llama.cpp/src/llama-hparams.h +13 -0
package/src/llama.cpp/src/llama-kv-cache.cpp +1604 -543
package/src/llama.cpp/src/llama-kv-cache.h +283 -171
package/src/llama.cpp/src/llama-memory.h +12 -2
package/src/llama.cpp/src/llama-mmap.cpp +1 -1
package/src/llama.cpp/src/llama-model-loader.cpp +34 -20
package/src/llama.cpp/src/llama-model-loader.h +5 -3
package/src/llama.cpp/src/llama-model-saver.cpp +281 -0
package/src/llama.cpp/src/llama-model-saver.h +37 -0
package/src/llama.cpp/src/llama-model.cpp +1803 -330
package/src/llama.cpp/src/llama-model.h +21 -2
package/src/llama.cpp/src/llama-quant.cpp +33 -10
package/src/llama.cpp/src/llama-sampling.cpp +25 -7
package/src/llama.cpp/src/llama-vocab.cpp +86 -10
package/src/llama.cpp/src/llama-vocab.h +6 -0
package/src/llama.cpp/src/llama.cpp +15 -1
package/src/llama.cpp/tests/CMakeLists.txt +52 -31
package/src/llama.cpp/tests/test-arg-parser.cpp +51 -4
package/src/llama.cpp/tests/test-backend-ops.cpp +189 -90
package/src/llama.cpp/tests/test-chat-template.cpp +26 -6
package/src/llama.cpp/tests/test-chat.cpp +15 -3
package/src/llama.cpp/{examples/gbnf-validator/gbnf-validator.cpp → tests/test-gbnf-validator.cpp} +2 -2
package/src/llama.cpp/tests/test-grammar-integration.cpp +3 -2
package/src/llama.cpp/tests/test-grammar-llguidance.cpp +63 -2
package/src/llama.cpp/tests/test-grammar-parser.cpp +3 -1
package/src/llama.cpp/tests/test-json-schema-to-grammar.cpp +17 -1
package/src/llama.cpp/tests/test-llama-grammar.cpp +2 -1
package/src/llama.cpp/tests/test-mtmd-c-api.c +63 -0
package/src/llama.cpp/tests/test-opt.cpp +33 -21
package/src/llama.cpp/{examples/quantize-stats/quantize-stats.cpp → tests/test-quantize-stats.cpp} +3 -1
package/src/llama.cpp/tests/test-regex-partial.cpp +288 -0
package/src/llama.cpp/tests/test-sampling.cpp +1 -1
package/src/llama.cpp/tests/test-tokenizer-1-bpe.cpp +2 -1
package/src/llama.cpp/tests/test-tokenizer-1-spm.cpp +2 -1
package/src/llama.cpp/tools/CMakeLists.txt +39 -0
package/src/llama.cpp/{examples → tools}/batched-bench/batched-bench.cpp +3 -3
package/src/llama.cpp/{examples → tools}/export-lora/export-lora.cpp +1 -1
package/src/llama.cpp/{examples → tools}/gguf-split/gguf-split.cpp +15 -16
package/src/llama.cpp/{examples → tools}/imatrix/imatrix.cpp +11 -9
package/src/llama.cpp/{examples → tools}/llama-bench/llama-bench.cpp +623 -274
package/src/llama.cpp/{examples → tools}/main/main.cpp +22 -14
package/src/llama.cpp/tools/mtmd/CMakeLists.txt +47 -0
package/src/llama.cpp/tools/mtmd/clip-impl.h +365 -0
package/src/llama.cpp/tools/mtmd/clip.cpp +3646 -0
package/src/llama.cpp/tools/mtmd/clip.h +99 -0
package/src/llama.cpp/tools/mtmd/deprecation-warning.cpp +22 -0
package/src/llama.cpp/tools/mtmd/mtmd-cli.cpp +370 -0
package/src/llama.cpp/tools/mtmd/mtmd-helper.cpp +310 -0
package/src/llama.cpp/tools/mtmd/mtmd.cpp +678 -0
package/src/llama.cpp/tools/mtmd/mtmd.h +331 -0
package/src/llama.cpp/{examples → tools}/perplexity/perplexity.cpp +21 -5
package/src/llama.cpp/{examples → tools}/quantize/quantize.cpp +53 -3
package/src/llama.cpp/tools/rpc/CMakeLists.txt +4 -0
package/src/llama.cpp/tools/rpc/rpc-server.cpp +322 -0
package/src/llama.cpp/tools/run/CMakeLists.txt +16 -0
package/src/llama.cpp/{examples → tools}/run/run.cpp +30 -30
package/src/llama.cpp/{examples → tools}/server/CMakeLists.txt +2 -1
package/src/llama.cpp/{examples → tools}/server/httplib.h +313 -247
package/src/llama.cpp/{examples → tools}/server/server.cpp +529 -215
package/src/llama.cpp/{examples → tools}/server/utils.hpp +427 -6
package/src/llama.cpp/{examples → tools}/tts/tts.cpp +6 -9
package/src/llama.cpp/cmake/arm64-windows-msvc.cmake +0 -6
package/src/llama.cpp/examples/gbnf-validator/CMakeLists.txt +0 -5
package/src/llama.cpp/examples/infill/CMakeLists.txt +0 -5
package/src/llama.cpp/examples/infill/infill.cpp +0 -590
package/src/llama.cpp/examples/llava/CMakeLists.txt +0 -66
package/src/llama.cpp/examples/llava/android/build_64.sh +0 -8
package/src/llama.cpp/examples/llava/clip-quantize-cli.cpp +0 -59
package/src/llama.cpp/examples/llava/clip.cpp +0 -3206
package/src/llama.cpp/examples/llava/clip.h +0 -118
package/src/llama.cpp/examples/llava/gemma3-cli.cpp +0 -341
package/src/llama.cpp/examples/llava/llava-cli.cpp +0 -332
package/src/llama.cpp/examples/llava/llava.cpp +0 -574
package/src/llama.cpp/examples/llava/llava.h +0 -49
package/src/llama.cpp/examples/llava/minicpmv-cli.cpp +0 -354
package/src/llama.cpp/examples/llava/qwen2vl-cli.cpp +0 -584
package/src/llama.cpp/examples/quantize-stats/CMakeLists.txt +0 -6
package/src/llama.cpp/examples/rpc/CMakeLists.txt +0 -2
package/src/llama.cpp/examples/rpc/rpc-server.cpp +0 -171
package/src/llama.cpp/examples/run/CMakeLists.txt +0 -5
package/src/llama.cpp/ggml/src/ggml-cann/kernels/CMakeLists.txt +0 -30
package/src/llama.cpp/ggml/src/ggml-cann/kernels/ascendc_kernels.h +0 -19
package/src/llama.cpp/ggml/src/ggml-cann/kernels/dup.cpp +0 -234
package/src/llama.cpp/ggml/src/ggml-cann/kernels/get_row_f16.cpp +0 -197
package/src/llama.cpp/ggml/src/ggml-cann/kernels/get_row_f32.cpp +0 -190
package/src/llama.cpp/ggml/src/ggml-cann/kernels/get_row_q4_0.cpp +0 -204
package/src/llama.cpp/ggml/src/ggml-cann/kernels/get_row_q8_0.cpp +0 -191
package/src/llama.cpp/ggml/src/ggml-cann/kernels/quantize_f16_q8_0.cpp +0 -218
package/src/llama.cpp/ggml/src/ggml-cann/kernels/quantize_f32_q8_0.cpp +0 -216
package/src/llama.cpp/ggml/src/ggml-cann/kernels/quantize_float_to_q4_0.cpp +0 -295
/package/src/llama.cpp/{examples → tools}/batched-bench/CMakeLists.txt +0 -0
/package/src/llama.cpp/{examples → tools}/cvector-generator/CMakeLists.txt +0 -0
/package/src/llama.cpp/{examples → tools}/cvector-generator/completions.txt +0 -0
/package/src/llama.cpp/{examples → tools}/cvector-generator/cvector-generator.cpp +0 -0
/package/src/llama.cpp/{examples → tools}/cvector-generator/mean.hpp +0 -0
/package/src/llama.cpp/{examples → tools}/cvector-generator/negative.txt +0 -0
/package/src/llama.cpp/{examples → tools}/cvector-generator/pca.hpp +0 -0
/package/src/llama.cpp/{examples → tools}/cvector-generator/positive.txt +0 -0
/package/src/llama.cpp/{examples → tools}/export-lora/CMakeLists.txt +0 -0
/package/src/llama.cpp/{examples → tools}/gguf-split/CMakeLists.txt +0 -0
/package/src/llama.cpp/{examples → tools}/imatrix/CMakeLists.txt +0 -0
/package/src/llama.cpp/{examples → tools}/llama-bench/CMakeLists.txt +0 -0
/package/src/llama.cpp/{examples → tools}/main/CMakeLists.txt +0 -0
/package/src/llama.cpp/{examples/llava → tools/mtmd}/requirements.txt +0 -0
/package/src/llama.cpp/{examples → tools}/perplexity/CMakeLists.txt +0 -0
/package/src/llama.cpp/{examples → tools}/quantize/CMakeLists.txt +0 -0
/package/src/llama.cpp/{examples → tools}/run/linenoise.cpp/linenoise.cpp +0 -0
/package/src/llama.cpp/{examples → tools}/run/linenoise.cpp/linenoise.h +0 -0
/package/src/llama.cpp/{examples → tools}/server/bench/requirements.txt +0 -0
/package/src/llama.cpp/{examples → tools}/server/tests/requirements.txt +0 -0
/package/src/llama.cpp/{examples → tools}/tokenize/CMakeLists.txt +0 -0
/package/src/llama.cpp/{examples → tools}/tokenize/tokenize.cpp +0 -0
/package/src/llama.cpp/{examples → tools}/tts/CMakeLists.txt +0 -0

package/src/llama.cpp/tests/test-backend-ops.cpp CHANGED Viewed

@@ -271,6 +271,14 @@ static std::string var_to_str(ggml_op_pool pool) {
     }
 }
+static std::string var_to_str(ggml_scale_mode mode) {
+    switch (mode) {
+        case GGML_SCALE_MODE_NEAREST:  return "nearest";
+        case GGML_SCALE_MODE_BILINEAR: return "bilinear";
+        default:                      return std::to_string(mode);
+    }
+}
 #define VAR_TO_STR(x) (#x "=" + var_to_str(x))
 #define VARS_TO_STR1(a) VAR_TO_STR(a)
@@ -815,7 +823,7 @@ struct test_case {
         ggml_build_forward_expand(gf, out);
         ggml_graph_cpy(gf, gb);
-        ggml_build_backward_expand(ctx.get(), ctx.get(), gb, false);
+        ggml_build_backward_expand(ctx.get(), gb, nullptr);
         if (expect.size() != 1 || expect[0] != 0.0f) {
             GGML_ASSERT(ggml_graph_n_nodes(gb) > ggml_graph_n_nodes(gf));
             for (ggml_tensor * t = ggml_get_first_tensor(ctx.get()); t != NULL; t = ggml_get_next_tensor(ctx.get(), t)) {
@@ -1018,7 +1026,7 @@ struct test_example : public test_case {
         // Step 3: return the output tensor.
         return out;
     }
-    // In order to also check the gradients for your op, add calls like ggml_set_param(ctx, a)
+    // In order to also check the gradients for your op, add calls like ggml_set_param(a)
     // immediately after you create the tensors.
     // This is optional and only makes sense if a backward pass has actually been implemented for the new op.
 };
@@ -1050,7 +1058,7 @@ struct test_unary : public test_case {
             auto ne = ne_a; ne[0] *= 3;
             a = ggml_new_tensor(ctx, type, 4, ne.data());
             if (grad_supported) {
-                ggml_set_param(ctx, a);
+                ggml_set_param(a);
             }
             ggml_set_name(a, "a");
@@ -1059,7 +1067,7 @@ struct test_unary : public test_case {
         } else {
             a = ggml_new_tensor(ctx, type, 4, ne_a.data());
             if (grad_supported) {
-                ggml_set_param(ctx, a);
+                ggml_set_param(a);
             }
             ggml_set_name(a, "a");
         }
@@ -1125,7 +1133,7 @@ struct test_get_rows : public test_case {
         const bool grad_supported = ggml_is_matrix(in) && ggml_is_vector(rows);
         if (grad_supported) {
-            ggml_set_param(ctx, in);
+            ggml_set_param(in);
             // rows is a constant input -> no gradients
         }
@@ -1314,7 +1322,7 @@ struct test_repeat : public test_case {
         ggml_set_name(target, "target");
         ggml_tensor * src = ggml_new_tensor(ctx, type, 4, ne.data());
-        ggml_set_param(ctx, src);
+        ggml_set_param(src);
         ggml_set_name(src, "src");
         ggml_tensor * out = ggml_repeat(ctx, src, target);
@@ -1398,7 +1406,7 @@ struct test_dup : public test_case {
     ggml_tensor * build_graph(ggml_context * ctx) override {
         ggml_tensor * src = ggml_new_tensor(ctx, type, 4, ne.data());
-        ggml_set_param(ctx, src);
+        ggml_set_param(src);
         ggml_set_name(src, "src");
         if (_use_permute) {
@@ -1434,7 +1442,7 @@ struct test_set : public test_case {
     ggml_tensor * build_graph(ggml_context * ctx) override {
         ggml_tensor * src = ggml_new_tensor(ctx, type_src, 4, ne.data());
-        ggml_set_param(ctx, src);
+        ggml_set_param(src);
         ggml_set_name(src, "src");
         auto ne_dst = ne;
@@ -1442,7 +1450,7 @@ struct test_set : public test_case {
             ne_dst[i] *= 2;
         }
         ggml_tensor* dst = ggml_new_tensor(ctx, type_dst, 4, ne_dst.data());
-        ggml_set_param(ctx, dst);
+        ggml_set_param(dst);
         ggml_set_name(dst, "dst");
         size_t offset = 0;
@@ -1490,7 +1498,7 @@ struct test_cpy : public test_case {
     ggml_tensor * build_graph(ggml_context * ctx) override {
         ggml_tensor * src = ggml_new_tensor(ctx, type_src, 4, ne.data());
-        ggml_set_param(ctx, src);
+        ggml_set_param(src);
         ggml_set_name(src, "src");
         if (_src_use_permute) {
@@ -1528,7 +1536,7 @@ struct test_cont : public test_case {
     ggml_tensor * build_graph(ggml_context * ctx) override {
         ggml_tensor * src = ggml_new_tensor(ctx, type, 4, ne.data());
-        ggml_set_param(ctx, src);
+        ggml_set_param(src);
         ggml_set_name(src, "src");
         src = ggml_transpose(ctx, src);
@@ -1575,8 +1583,8 @@ struct test_bin_bcast : public test_case {
         // The backward pass supports broadcasting only for GGML_ADD:
         const bool grad_supported = op == ggml_add || ggml_are_same_shape(a, b);
         if (grad_supported) {
-            ggml_set_param(ctx, a);
-            ggml_set_param(ctx, b);
+            ggml_set_param(a);
+            ggml_set_param(b);
         }
         ggml_tensor * out = op(ctx, a, b);
@@ -1624,11 +1632,11 @@ struct test_add1 : public test_case {
     ggml_tensor * build_graph(ggml_context * ctx) override {
         ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data());
-        ggml_set_param(ctx, a);
+        ggml_set_param(a);
         ggml_set_name(a, "a");
         ggml_tensor * b = ggml_new_tensor_1d(ctx, type, 1);
-        // ggml_set_param(ctx, b); // TODO: implement
+        // ggml_set_param(b); // TODO: implement
         ggml_set_name(b, "b");
         ggml_tensor * out = ggml_add1(ctx, a, b);
@@ -1659,7 +1667,7 @@ struct test_scale : public test_case {
     ggml_tensor * build_graph(ggml_context * ctx) override {
         ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data());
-        ggml_set_param(ctx, a);
+        ggml_set_param(a);
         ggml_set_name(a, "a");
         ggml_tensor * out = ggml_scale(ctx, a, scale);
@@ -1754,7 +1762,7 @@ struct test_rms_norm : public test_case {
     ggml_tensor * build_graph(ggml_context * ctx) override {
         ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data());
-        ggml_set_param(ctx, a);
+        ggml_set_param(a);
         ggml_set_name(a, "a");
         if (v) {
@@ -1973,7 +1981,7 @@ struct test_mul_mat : public test_case {
     const std::array<int64_t, 2> bs;  // dims 3 and 4
     const std::array<int64_t, 2> nr;  // repeat in dims 3 and 4
     const std::array<int64_t, 4> per; // permutation of dimensions
-    const bool v; // whether a is a non-contiguous view
+    const bool v; // whether a and b are non-contiguous views
     std::string vars() override {
         return VARS_TO_STR9(type_a, type_b, m, n, k, bs, nr, per, v);
@@ -2020,9 +2028,9 @@ struct test_mul_mat : public test_case {
             b = ggml_new_tensor_4d(ctx, type_b, ne_b[per[0]], ne_b[per[1]], ne_b[per[2]], ne_b[per[3]]);
             if (!ggml_is_quantized(type_a)) {
                 if (bs[1] == 1 && nr[1] == 1) {
-                    ggml_set_param(ctx, a);
+                    ggml_set_param(a);
                 }
-                ggml_set_param(ctx, b);
+                ggml_set_param(b);
             }
             ggml_set_name(a, "a");
             ggml_set_name(b, "b");
@@ -2032,19 +2040,29 @@ struct test_mul_mat : public test_case {
             ggml_set_name(a, "a_permuted");
             ggml_set_name(b, "b_permuted");
         } else {
             if (v) {
-                a = ggml_new_tensor_4d(ctx, type_a, k*2, m, bs[0], bs[1]);
-                a = ggml_view_4d(ctx, a, k, m, bs[0], bs[1], a->nb[1], a->nb[2], a->nb[3], 0);
+                a = ggml_new_tensor_4d(ctx, type_a, k*2, m, bs[0],       bs[1]);
+                b = ggml_new_tensor_4d(ctx, type_b, k*2, n, bs[0]*nr[0], bs[1]*nr[1]);
+                if (!ggml_is_quantized(type_a)) {
+                    if (bs[1] == 1 && nr[1] == 1) {
+                        ggml_set_param(a);
+                    }
+                    ggml_set_param(b);
+                }
+                a = ggml_view_4d(ctx, a, k, m, bs[0],       bs[1],       a->nb[1], a->nb[2], a->nb[3], 0);
+                b = ggml_view_4d(ctx, b, k, n, bs[0]*nr[0], bs[1]*nr[1], b->nb[1], b->nb[2], b->nb[3], 0);
             } else {
                 a = ggml_new_tensor_4d(ctx, type_a, k, m, bs[0],       bs[1]);
-            }
-            b = ggml_new_tensor_4d(ctx, type_b, k, n, bs[0]*nr[0], bs[1]*nr[1]);
-            if (!ggml_is_quantized(type_a)) {
-                if (bs[1] == 1 && nr[1] == 1) {
-                    ggml_set_param(ctx, a);
+                b = ggml_new_tensor_4d(ctx, type_b, k, n, bs[0]*nr[0], bs[1]*nr[1]);
+                if (!ggml_is_quantized(type_a)) {
+                    if (bs[1] == 1 && nr[1] == 1) {
+                        ggml_set_param(a);
+                    }
+                    ggml_set_param(b);
                 }
-                ggml_set_param(ctx, b);
             }
             ggml_set_name(a, "a");
             ggml_set_name(b, "b");
@@ -2063,7 +2081,7 @@ struct test_mul_mat_id : public test_case {
     const ggml_type type_b;
     const int n_mats;
     const int n_used;
-    const bool b; // brodcast b matrix
+    const bool b; // broadcast b matrix
     const int64_t m;
     const int64_t n;
     const int64_t k;
@@ -2193,7 +2211,7 @@ struct test_sqr : public test_case {
     ggml_tensor * build_graph(ggml_context * ctx) override {
         ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data());
-        ggml_set_param(ctx, a);
+        ggml_set_param(a);
         ggml_set_name(a, "a");
         ggml_tensor * out = ggml_sqr(ctx, a);
@@ -2222,7 +2240,7 @@ struct test_sqrt : public test_case {
     ggml_tensor * build_graph(ggml_context * ctx) override {
         ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data());
-        ggml_set_param(ctx, a);
+        ggml_set_param(a);
         ggml_set_name(a, "a");
         ggml_tensor * out = ggml_sqrt(ctx, a);
@@ -2262,7 +2280,7 @@ struct test_log : public test_case {
     ggml_tensor * build_graph(ggml_context * ctx) override {
         ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data());
-        ggml_set_param(ctx, a);
+        ggml_set_param(a);
         ggml_set_name(a, "a");
         ggml_tensor * out = ggml_log(ctx, a);
@@ -2298,7 +2316,7 @@ struct test_sin : public test_case {
     ggml_tensor * build_graph(ggml_context * ctx) override {
         ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data());
-        ggml_set_param(ctx, a);
+        ggml_set_param(a);
         ggml_set_name(a, "a");
         ggml_tensor * out = ggml_sin(ctx, a);
@@ -2341,7 +2359,7 @@ struct test_cos : public test_case {
     ggml_tensor * build_graph(ggml_context * ctx) override {
         ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data());
-        ggml_set_param(ctx, a);
+        ggml_set_param(a);
         ggml_set_name(a, "a");
         ggml_tensor * out = ggml_cos(ctx, a);
@@ -2421,7 +2439,7 @@ struct test_diag_mask_inf : public test_case {
     ggml_tensor * build_graph(ggml_context * ctx) override {
         ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data());
-        ggml_set_param(ctx, a);
+        ggml_set_param(a);
         ggml_set_name(a, "a");
         ggml_tensor * out = ggml_diag_mask_inf(ctx, a, n_past);
@@ -2460,7 +2478,7 @@ struct test_soft_max : public test_case {
     ggml_tensor * build_graph(ggml_context * ctx) override {
         ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data());
-        ggml_set_param(ctx, a);
+        ggml_set_param(a);
         ggml_set_name(a, "a");
         ggml_tensor * mask = nullptr;
@@ -2542,7 +2560,7 @@ struct test_rope : public test_case {
             auto ne = ne_a; ne[0] *= 2; ne[1] *= 4; ne[2] *= 3;
             a = ggml_new_tensor(ctx, type, 4, ne.data());
             if (forward) {
-                ggml_set_param(ctx, a);
+                ggml_set_param(a);
             }
             ggml_set_name(a, "a");
@@ -2551,7 +2569,7 @@ struct test_rope : public test_case {
         } else {
             a = ggml_new_tensor(ctx, type, 4, ne_a.data());
             if (forward) {
-                ggml_set_param(ctx, a);
+                ggml_set_param(a);
             }
             ggml_set_name(a, "a");
         }
@@ -2598,6 +2616,8 @@ struct test_rope : public test_case {
             } else {
                 out = ggml_rope_ext_back(ctx, a, pos, freq, n_dims, mode, 0, 10000.0f, fs, ef, af, 1.0f, 1.0f);
             }
+            // TODO: add test with a non-contiguous view as input ; this case is needed for build_rope_2d in clip.cpp
         }
         ggml_set_name(out, "out");
@@ -2663,7 +2683,7 @@ struct test_pool2d : public test_case {
     ggml_tensor * build_graph(ggml_context * ctx) override {
         ggml_tensor * input = ggml_new_tensor(ctx, type_input, 4, ne_input.data());
-        ggml_set_param(ctx, input);
+        ggml_set_param(input);
         ggml_set_name(input, "input");
         ggml_tensor * out = ggml_pool_2d(ctx, input, pool_type, k0, k1, s0, s1, p0, p1);
@@ -2739,7 +2759,7 @@ struct test_im2col : public test_case {
     ggml_tensor * build_graph(ggml_context * ctx) override {
         ggml_tensor * input = ggml_new_tensor(ctx, type_input, 4, ne_input.data());
-        ggml_set_param(ctx, input);
+        ggml_set_param(input);
         ggml_set_name(input, "input");
         ggml_tensor * kernel = ggml_new_tensor(ctx, type_kernel, 4, ne_kernel.data());
@@ -2752,6 +2772,48 @@ struct test_im2col : public test_case {
     }
 };
+// GGML_OP_CONV_2D_DW
+struct test_conv_2d_dw : public test_case {
+    const std::array<int64_t, 4> ne_input;
+    const std::array<int64_t, 4> ne_kernel;
+    const int stride;
+    const int padding;
+    const int dilation;
+    const bool cwhn;
+    std::string vars() override {
+        return VARS_TO_STR6(ne_input, ne_kernel, stride, padding, dilation, cwhn);
+    }
+    test_conv_2d_dw(std::array<int64_t, 4> ne_input = {64, 64, 16, 1},
+            std::array<int64_t, 4> ne_kernel = {3, 3, 1, 16},
+            int stride = 1, int padding = 0, int dilation = 1, bool cwhn = false)
+        : ne_input(ne_input), ne_kernel(ne_kernel), stride(stride), padding(padding), dilation(dilation), cwhn(cwhn) {}
+    ggml_tensor * build_graph(ggml_context * ctx) override {
+        ggml_tensor * input = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, ne_input.data());
+        ggml_set_name(input, "input");
+        ggml_tensor * kernel = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, ne_kernel.data());
+        ggml_set_name(kernel, "kernel");
+        if (cwhn) {
+            // change memory layout to channel-most-contiguous (CWHN),
+            // then permute it back so NE matches the original input
+            input = ggml_cont(ctx, ggml_permute(ctx, input, 1, 2, 0, 3));
+            input = ggml_permute(ctx, input, 2, 0, 1, 3);
+            kernel = ggml_cont(ctx, ggml_permute(ctx, kernel, 2, 3, 1, 0));
+            kernel = ggml_permute(ctx, kernel, 3, 2, 0, 1);
+        }
+        ggml_tensor * out = ggml_conv_2d_dw_direct(
+            ctx, kernel, input,
+            stride, stride, padding, padding, dilation, dilation);
+        ggml_set_name(out, "out");
+        return out;
+    }
+};
 // GGML_OP_CONCAT
 struct test_concat : public test_case {
     const ggml_type type;
@@ -2874,7 +2936,7 @@ struct test_sum : public test_case {
     ggml_tensor * build_graph(ggml_context * ctx) override {
         ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data());
-        ggml_set_param(ctx, a);
+        ggml_set_param(a);
         ggml_set_name(a, "a");
         ggml_tensor * out = ggml_sum(ctx, a);
@@ -2903,7 +2965,7 @@ struct test_sum_rows : public test_case {
     ggml_tensor * build_graph(ggml_context * ctx) override {
         ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data());
-        ggml_set_param(ctx, a);
+        ggml_set_param(a);
         ggml_set_name(a, "a");
         ggml_tensor * out = ggml_sum_rows(ctx, a);
@@ -2928,7 +2990,7 @@ struct test_mean : public test_case {
     ggml_tensor * build_graph(ggml_context * ctx) override {
         ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data());
-        ggml_set_param(ctx, a);
+        ggml_set_param(a);
         ggml_set_name(a, "a");
         ggml_tensor * out = ggml_mean(ctx, a);
@@ -2948,15 +3010,16 @@ struct test_upscale : public test_case {
     const std::array<int64_t, 4> ne;
     const int32_t scale_factor;
     const bool transpose;
+    const ggml_scale_mode mode;
     std::string vars() override {
-        return VARS_TO_STR4(type, ne, scale_factor, transpose);
+        return VARS_TO_STR5(type, ne, scale_factor, mode, transpose);
     }
     test_upscale(ggml_type type = GGML_TYPE_F32,
             std::array<int64_t, 4> ne = {512, 512, 3, 1},
-            int32_t scale_factor = 2, bool transpose = false)
-        : type(type), ne(ne), scale_factor(scale_factor), transpose(transpose) {}
+            int32_t scale_factor = 2, ggml_scale_mode mode = GGML_SCALE_MODE_NEAREST, bool transpose = false)
+        : type(type), ne(ne), scale_factor(scale_factor), transpose(transpose), mode(mode) {}
     ggml_tensor * build_graph(ggml_context * ctx) override {
         ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data());
@@ -2967,7 +3030,7 @@ struct test_upscale : public test_case {
             ggml_set_name(a, "a_transposed");
         }
-        ggml_tensor * out = ggml_upscale(ctx, a, scale_factor);
+        ggml_tensor * out = ggml_upscale(ctx, a, scale_factor, mode);
         ggml_set_name(out, "out");
         return out;
@@ -2979,21 +3042,23 @@ struct test_upscale_ext : public test_case {
     const ggml_type type;
     const std::array<int64_t, 4> ne;
     const std::array<int64_t, 4> ne_tgt;
+    const ggml_scale_mode mode = GGML_SCALE_MODE_NEAREST;
     std::string vars() override {
-        return VARS_TO_STR3(type, ne, ne_tgt);
+        return VARS_TO_STR4(type, ne, ne_tgt, mode);
     }
     test_upscale_ext(ggml_type type = GGML_TYPE_F32,
             std::array<int64_t, 4> ne     = {2, 5,  7, 11},
-            std::array<int64_t, 4> ne_tgt = {5, 7, 11, 13})
-        : type(type), ne(ne), ne_tgt(ne_tgt) {}
+            std::array<int64_t, 4> ne_tgt = {5, 7, 11, 13},
+            ggml_scale_mode mode = GGML_SCALE_MODE_NEAREST)
+        : type(type), ne(ne), ne_tgt(ne_tgt), mode(mode) {}
     ggml_tensor * build_graph(ggml_context * ctx) override {
         ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data());
         ggml_set_name(a, "a");
-        ggml_tensor * out = ggml_upscale_ext(ctx, a, ne_tgt[0], ne_tgt[1],ne_tgt[2], ne_tgt[3]);
+        ggml_tensor * out = ggml_upscale_ext(ctx, a, ne_tgt[0], ne_tgt[1],ne_tgt[2], ne_tgt[3], mode);
         ggml_set_name(out, "out");
         return out;
@@ -3071,11 +3136,11 @@ struct test_acc : public test_case {
     ggml_tensor * build_graph(ggml_context * ctx) override {
         ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne_a.data());
-        ggml_set_param(ctx, a);
+        ggml_set_param(a);
         ggml_set_name(a, "a");
         ggml_tensor * b = ggml_new_tensor(ctx, type, 4, ne_b.data());
-        ggml_set_param(ctx, b);
+        ggml_set_param(b);
         ggml_set_name(b, "b");
         ggml_tensor * out = ggml_acc(ctx, a, b, a->nb[1], a->nb[2], a->nb[3], b->nb[1]);
@@ -3217,7 +3282,8 @@ struct test_leaky_relu : public test_case {
 // GGML_OP_FLASH_ATTN_EXT
 struct test_flash_attn_ext : public test_case {
-    const int64_t hs; // head size
+    const int64_t hsk; // K head size
+    const int64_t hsv; // V head size
     const int64_t nh; // num heads
     const int64_t nr; // repeat in Q, tests for grouped-query attention
     const int64_t kv; // kv size
@@ -3233,7 +3299,7 @@ struct test_flash_attn_ext : public test_case {
     std::array<int32_t, 4> permute;
     std::string vars() override {
-        return VARS_TO_STR11(hs, nh, nr, kv, nb, mask, max_bias, logit_softcap, prec, type_KV, permute);
+        return VARS_TO_STR12(hsk, hsv, nh, nr, kv, nb, mask, max_bias, logit_softcap, prec, type_KV, permute);
     }
     double max_nmse_err() override {
@@ -3243,17 +3309,18 @@ struct test_flash_attn_ext : public test_case {
     uint64_t op_flops(ggml_tensor * t) override {
         GGML_UNUSED(t);
         // Just counting matmul costs:
-        // Q*K^T is nb x hs x kv, P*V is nb x kv x hs, per head
-        return 2 * 2 * nh*nr * nb * hs * kv;
+        // Q*K^T is nb x hsk x kv, P*V is nb x kv x hsv, per head
+        return 2 * nh*nr * nb * (hsk + hsv) * kv;
     }
-    test_flash_attn_ext(int64_t hs = 128, int64_t nh = 32, int64_t nr = 1, int64_t kv = 96, int64_t nb = 8,
+    test_flash_attn_ext(int64_t hsk = 128, int64_t hsv = 128, int64_t nh = 32, int64_t nr = 1, int64_t kv = 96, int64_t nb = 8,
                         bool mask = true, float max_bias = 0.0f, float logit_softcap = 0.0f, ggml_prec prec = GGML_PREC_F32,
                         ggml_type type_KV = GGML_TYPE_F16, std::array<int32_t, 4> permute = {0, 1, 2, 3})
-        : hs(hs), nh(nh), nr(nr), kv(kv), nb(nb), mask(mask), max_bias(max_bias), logit_softcap(logit_softcap), prec(prec), type_KV(type_KV), permute(permute) {}
+        : hsk(hsk), hsv(hsv), nh(nh), nr(nr), kv(kv), nb(nb), mask(mask), max_bias(max_bias), logit_softcap(logit_softcap), prec(prec), type_KV(type_KV), permute(permute) {}
     ggml_tensor * build_graph(ggml_context * ctx) override {
-        const int64_t hs_padded = GGML_PAD(hs, ggml_blck_size(type_KV));
+        const int64_t hsk_padded = GGML_PAD(hsk, ggml_blck_size(type_KV));
+        const int64_t hsv_padded = GGML_PAD(hsv, ggml_blck_size(type_KV));
         auto const &create_permuted = [&](ggml_type type, int64_t ne0, int64_t ne1, int64_t ne2, int64_t ne3) -> ggml_tensor * {
             int64_t ne[4] = {ne0, ne1, ne2, ne3};
@@ -3268,13 +3335,13 @@ struct test_flash_attn_ext : public test_case {
             return t;
         };
-        ggml_tensor * q = create_permuted(GGML_TYPE_F32, hs_padded, nb, nh*nr, 1);
+        ggml_tensor * q = create_permuted(GGML_TYPE_F32, hsk_padded, nb, nh*nr, 1);
         ggml_set_name(q, "q");
-        ggml_tensor * k = create_permuted(type_KV,       hs_padded, kv, nh,    1);
+        ggml_tensor * k = create_permuted(type_KV,       hsk_padded, kv, nh,    1);
         ggml_set_name(k, "k");
-        ggml_tensor * v = create_permuted(type_KV,       hs_padded, kv, nh,    1);
+        ggml_tensor * v = create_permuted(type_KV,       hsv_padded, kv, nh,    1);
         ggml_set_name(v, "v");
         ggml_tensor * m = nullptr;
@@ -3283,7 +3350,7 @@ struct test_flash_attn_ext : public test_case {
             ggml_set_name(m, "m");
         }
-        ggml_tensor * out = ggml_flash_attn_ext(ctx, q, k, v, m, 1.0f/sqrtf(hs), max_bias, logit_softcap);
+        ggml_tensor * out = ggml_flash_attn_ext(ctx, q, k, v, m, 1.0f/sqrtf(hsk), max_bias, logit_softcap);
         ggml_flash_attn_ext_set_prec(out, prec);
         ggml_set_name(out, "out");
@@ -3310,7 +3377,7 @@ struct test_cross_entropy_loss : public test_case {
     ggml_tensor * build_graph(ggml_context * ctx) override {
         ggml_tensor * logits = ggml_new_tensor(ctx, type, 4, ne.data());
-        ggml_set_param(ctx, logits);
+        ggml_set_param(logits);
         ggml_set_name(logits, "logits");
         ggml_tensor * labels = ggml_new_tensor(ctx, type, 4, ne.data());
@@ -3392,7 +3459,7 @@ struct test_opt_step_adamw : public test_case {
     ggml_tensor * build_graph(ggml_context * ctx) override {
         ggml_tensor * a = ggml_new_tensor_4d(ctx, type, ne[0], ne[1], ne[2], ne[3]);
-        ggml_set_param(ctx, a); // Despite tensor a having gradients the output tensor will not.
+        ggml_set_param(a); // Despite tensor a having gradients the output tensor will not.
         ggml_set_name(a, "a");
         ggml_tensor * grad = ggml_new_tensor_4d(ctx, type, ne[0], ne[1], ne[2], ne[3]);
@@ -3957,6 +4024,11 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
     // test_cases.emplace_back(new test_im2col(GGML_TYPE_F32, GGML_TYPE_F16, GGML_TYPE_F16, {1024, 1024, 256, 1}, {3, 3, 256, 1}, 1, 1, 1, 1, 1, 1, true));
     // test_cases.emplace_back(new test_im2col(GGML_TYPE_F32, GGML_TYPE_F16, GGML_TYPE_F32, {1024, 1024, 256, 1}, {3, 3, 256, 1}, 1, 1, 1, 1, 1, 1, true));
+    test_cases.emplace_back(new test_conv_2d_dw({17, 34, 9, 1}, {3, 3, 1, 9}, 1, 0, 1, false));
+    test_cases.emplace_back(new test_conv_2d_dw({17, 34, 9, 1}, {3, 3, 1, 9}, 1, 0, 1, true));
+    test_cases.emplace_back(new test_conv_2d_dw({32, 8, 64, 1}, {3, 3, 1, 64}, 2, 1, 1, false));
+    test_cases.emplace_back(new test_conv_2d_dw({32, 8, 64, 1}, {3, 3, 1, 64}, 2, 1, 1, true));
     test_cases.emplace_back(new test_conv_transpose_1d());
     test_cases.emplace_back(new test_conv_transpose_1d({3,2,1,1}, {2,3,2,1}, 3, 0, 1));
     test_cases.emplace_back(new test_conv_transpose_1d({3,2,1,1}, {2,3,2,1}, 2, 0, 1));
@@ -4169,6 +4241,11 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
             test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 16, 256, {2, 3}, {1, 1}, {0, 2, 1, 3}));
             test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 16, 256, {2, 3}, {1, 1}, {0, 1, 3, 2}));
             test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 16, 256, {2, 3}, {1, 1}, {0, 3, 2, 1}));
+            // test cases with large ne00/ne10 to cover stream-k fixup
+            test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16,  1, 1024, {3, 2}, {1, 1}));
+            test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16,  8, 1024, {3, 2}, {1, 1}));
+            test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 16, 1024, {3, 2}, {1, 1}));
         }
     }
     for (ggml_type type_a : other_types) {
@@ -4204,6 +4281,8 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
     test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F16, GGML_TYPE_F32,  83, 2,   64, { 8,  1}, {4, 1}));
     test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F16, GGML_TYPE_F32,  64, 45, 128, { 8,  1}, {4, 1}));
     test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F16, GGML_TYPE_F32, 128, 45,  64, { 8,  1}, {4, 1}));
+    test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F16, GGML_TYPE_F32, 1056, 1, 193, {1,  1}, {4, 1}, {0, 2, 1, 3}));
+    test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F16, GGML_TYPE_F32, 1056, 1, 67,  {1,  1}, {4, 1}, {0, 2, 1, 3}));
     for (auto bs : {1,2,4,8}) {
         for (auto nr : {1,4}) {
@@ -4395,12 +4474,15 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
         test_cases.emplace_back(new test_argsort(GGML_TYPE_F32, {60, 10, 10, 10}, order)); // qwen
     }
+    for (ggml_scale_mode mode : {GGML_SCALE_MODE_NEAREST, GGML_SCALE_MODE_BILINEAR}) {
+        test_cases.emplace_back(new test_upscale(GGML_TYPE_F32, {512, 512, 3, 2}, 2, mode));
+        test_cases.emplace_back(new test_upscale(GGML_TYPE_F32, {512, 512, 3, 2}, 2, mode, true));
+        test_cases.emplace_back(new test_upscale_ext(GGML_TYPE_F32, {2, 5,  7, 11}, {5, 7, 11, 13}, mode));
+    }
     test_cases.emplace_back(new test_sum());
     test_cases.emplace_back(new test_sum_rows());
     test_cases.emplace_back(new test_mean());
-    test_cases.emplace_back(new test_upscale());
-    test_cases.emplace_back(new test_upscale(GGML_TYPE_F32, { 512, 512, 3, 1 }, 2, true));
-    test_cases.emplace_back(new test_upscale_ext());
     test_cases.emplace_back(new test_group_norm(GGML_TYPE_F32, {64, 64, 320, 1}));
     test_cases.emplace_back(new test_group_norm(GGML_TYPE_F32, {9, 9, 1280, 1}));
     test_cases.emplace_back(new test_acc());
@@ -4410,27 +4492,33 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
     test_cases.emplace_back(new test_timestep_embedding());
     test_cases.emplace_back(new test_leaky_relu());
-    for (int hs : { 64, 80, 128, 256, }) {
-        for (bool mask : { true, false } ) {
-            for (float max_bias : { 0.0f, 8.0f }) {
-                if (!mask && max_bias > 0.0f) continue;
-                for (float logit_softcap : {0.0f, 10.0f}) {
-                    if (hs != 128 && logit_softcap != 0.0f) continue;
-                    for (int nh : { 4, }) {
-                        for (int nr : { 1, 4, 16 }) {
-                            if (nr == 16 && hs != 128) continue;
-                            for (int kv : { 512, 1024, }) {
-                                if (nr != 1 && kv != 512) continue;
-                                for (int nb : { 1, 3, 32, 35, }) {
-                                    for (ggml_prec prec : {GGML_PREC_F32, GGML_PREC_DEFAULT}) {
-                                        if (hs != 128 && prec == GGML_PREC_DEFAULT) continue;
-                                        for (ggml_type type_KV : {GGML_TYPE_F16, GGML_TYPE_BF16, GGML_TYPE_Q8_0, GGML_TYPE_Q4_0}) {
-                                            test_cases.emplace_back(new test_flash_attn_ext(
-                                                hs, nh, nr, kv, nb, mask, max_bias, logit_softcap, prec, type_KV));
-                                            // run fewer test cases permuted
-                                            if (mask == true && max_bias == 0.0f && logit_softcap == 0 && kv == 512) {
+    for (int hsk : { 64, 80, 128, 192, 256, 576 }) {
+        for (int hsv : { 64, 80, 128, 192, 256, 512 }) {
+            if (hsk != 192 && hsk != 576 && hsk != hsv) continue;
+            if (hsk == 192 && (hsv != 128 && hsv != 192)) continue;
+            if (hsk == 576 && hsv != 512) continue; // DeepSeek MLA
+            for (bool mask : { true, false } ) {
+                for (float max_bias : { 0.0f, 8.0f }) {
+                    if (!mask && max_bias > 0.0f) continue;
+                    for (float logit_softcap : {0.0f, 10.0f}) {
+                        if (hsk != 128 && logit_softcap != 0.0f) continue;
+                        for (int nh : { 4, }) {
+                            for (int nr : { 1, 4, 16 }) {
+                                if (nr == 16 && hsk != 128) continue;
+                                for (int kv : { 512, 1024, }) {
+                                    if (nr != 1 && kv != 512) continue;
+                                    for (int nb : { 1, 3, 32, 35, }) {
+                                        for (ggml_prec prec : {GGML_PREC_F32, GGML_PREC_DEFAULT}) {
+                                            if (hsk != 128 && prec == GGML_PREC_DEFAULT) continue;
+                                            for (ggml_type type_KV : {GGML_TYPE_F16, GGML_TYPE_BF16, GGML_TYPE_Q8_0, GGML_TYPE_Q4_0}) {
                                                 test_cases.emplace_back(new test_flash_attn_ext(
-                                                    hs, nh, nr, kv, nb, mask, max_bias, logit_softcap, prec, type_KV, {0, 2, 1, 3}));
+                                                    hsk, hsv, nh, nr, kv, nb, mask, max_bias, logit_softcap, prec, type_KV));
+                                                // run fewer test cases permuted
+                                                if (mask == true && max_bias == 0.0f && logit_softcap == 0 && kv == 512) {
+                                                    test_cases.emplace_back(new test_flash_attn_ext(
+                                                        hsk, hsv, nh, nr, kv, nb, mask, max_bias, logit_softcap, prec, type_KV, {0, 2, 1, 3}));
+                                                }
                                             }
                                         }
                                     }
@@ -4507,6 +4595,17 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_perf() {
         }
     }
+    for (int kv : { 4096, 8192, 16384, }) {
+        for (int hs : { 64, 128, }) {
+            for (int nr : { 1, 4, }) {
+                test_cases.emplace_back(new test_flash_attn_ext(hs, hs, 8, nr, kv, 1, true, 0, 0, GGML_PREC_F32, GGML_TYPE_F16));
+            }
+        }
+    }
+    test_cases.emplace_back(new test_conv_2d_dw({512, 512, 256, 1}, {3, 3, 1, 256}, 1, 1, 1, false));
+    test_cases.emplace_back(new test_conv_2d_dw({512, 512, 256, 1}, {3, 3, 1, 256}, 1, 1, 1, true));
     return test_cases;
 }