LLM inference in C/C++
C++ ggml
ggml-webgpu: no dawn native features on wasi (#27069)
Mendy Bergerf498f864fbc0
diff --git a/ggml/src/ggml-webgpu/ggml-webgpu.cpp b/ggml/src/ggml-webgpu/ggml-webgpu.cppindex a7986b9c76..ebc8a5e782 100644--- a/ggml/src/ggml-webgpu/ggml-webgpu.cpp+++ b/ggml/src/ggml-webgpu/ggml-webgpu.cpp@@ -4104,7 +4104,7 @@ static void ggml_webgpu_init_memset_pipeline(webgpu_global_context & ctx) {static void ggml_backend_webgpu_request_adapter(wgpu::Instance & instance, wgpu::Adapter & adapter) {wgpu::RequestAdapterOptions options = {};-#ifndef __EMSCRIPTEN__+#if !defined(__EMSCRIPTEN__) && !defined(__wasi__)// TODO: track need for these toggles: https://issues.chromium.org/issues/42251215const char * const adapterEnabledToggles[] = { "vulkan_enable_f16_on_nvidia", "use_vulkan_memory_model" };wgpu::DawnTogglesDescriptor adapterTogglesDesc;@@ -4133,7 +4133,7 @@ static void create_webgpu_device(ggml_backend_webgpu_reg_context * ctx) {ctx->webgpu_global_ctx->adapter.GetLimits(&ctx->webgpu_global_ctx->capabilities.limits);wgpu::AdapterInfo info{};-#ifndef __EMSCRIPTEN__+#if !defined(__EMSCRIPTEN__) && !defined(__wasi__)wgpu::AdapterPropertiesSubgroupMatrixConfigs subgroup_matrix_configs{};if (ctx->webgpu_global_ctx->adapter.HasFeature(wgpu::FeatureName::ChromiumExperimentalSubgroupMatrix)) {info.nextInChain = &subgroup_matrix_configs;@@ -4150,7 +4150,7 @@ static void create_webgpu_device(ggml_backend_webgpu_reg_context * ctx) {wgpu::WGSLLanguageFeatureName::Packed4x8IntegerDotProduct);bool valid_subgroup_matrix_config = false;-#ifndef __EMSCRIPTEN__+#if !defined(__EMSCRIPTEN__) && !defined(__wasi__)// Accept f16 subgroup matrix configurations (square or non-square).// NVIDIA GPUs typically report square configs (e.g. 16x16x16),// while Intel Xe2 GPUs report non-square configs (e.g. 8x16x16).@@ -4178,7 +4178,7 @@ static void create_webgpu_device(ggml_backend_webgpu_reg_context * ctx) {// Initialize devicestd::vector<wgpu::FeatureName> required_features = { wgpu::FeatureName::ShaderF16 };-#ifndef __EMSCRIPTEN__+#if !defined(__EMSCRIPTEN__) && !defined(__wasi__)required_features.push_back(wgpu::FeatureName::ImplicitDeviceSynchronization);if (ctx->webgpu_global_ctx->capabilities.supports_subgroup_matrix) {required_features.push_back(wgpu::FeatureName::ChromiumExperimentalSubgroupMatrix);@@ -4214,7 +4214,7 @@ static void create_webgpu_device(ggml_backend_webgpu_reg_context * ctx) {std::string(message).c_str());});-#ifndef __EMSCRIPTEN__+#if !defined(__EMSCRIPTEN__) && !defined(__wasi__)// Enable Dawn-specific toggles to increase native performance// TODO: Maybe WebGPU needs a "fast" mode where you can request compilers skip adding checks like these,// only for native performance?@@ -4865,7 +4865,7 @@ ggml_backend_reg_t ggml_backend_webgpu_reg() {instance_descriptor.requiredFeatures = instance_features.data();instance_descriptor.requiredFeatureCount = instance_features.size();-#ifndef __EMSCRIPTEN__+#if !defined(__EMSCRIPTEN__) && !defined(__wasi__)const char * const instanceEnabledToggles[] = { "allow_unsafe_apis" };wgpu::DawnTogglesDescriptor instanceTogglesDesc;instanceTogglesDesc.enabledToggles = instanceEnabledToggles;@@ -4885,7 +4885,7 @@ ggml_backend_reg_t ggml_backend_webgpu_reg() {// WebGPU backend requires f16 support and, on native, implicit device synchronization.if (adapter != nullptr && adapter.HasFeature(wgpu::FeatureName::ShaderF16)-#ifndef __EMSCRIPTEN__+#if !defined(__EMSCRIPTEN__) && !defined(__wasi__)&& adapter.HasFeature(wgpu::FeatureName::ImplicitDeviceSynchronization)#endif) {