nikola/llama_cpp
1
0
Fork 0
Code Issues Pull Requests Actions 19 Packages Projects Releases Wiki Activity

ggml-webgpu: add adapter toggles for F16 on Vulkan + NVIDIA

Browse Source
This commit is contained in:
Masashi Yoshimura
2026-06-20 08:12:32 +09:00
committed by GitHub
parent 2b686a9120
commit f449e05537
1 changed files with 16 additions and 26 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-webgpu/ggml-webgpu.cpp
+9 -19
View File
@@ -3788,7 +3788,7 @@ static void ggml_webgpu_init_memset_pipeline(webgpu_global_context & ctx) {
ctx->memset_pipeline = ggml_webgpu_create_pipeline(ctx->device, wgsl_memset, "memset", constants); ctx->memset_pipeline = ggml_webgpu_create_pipeline(ctx->device, wgsl_memset, "memset", constants);
} }
static void create_webgpu_device(ggml_backend_webgpu_reg_context * ctx) { static void ggml_backend_webgpu_request_adapter(wgpu::Instance & instance, wgpu::Adapter & adapter) {
wgpu::RequestAdapterOptions options = {}; wgpu::RequestAdapterOptions options = {};
#ifndef __EMSCRIPTEN__ #ifndef __EMSCRIPTEN__
@@ -3800,17 +3800,20 @@ static void create_webgpu_device(ggml_backend_webgpu_reg_context * ctx) {
options.nextInChain = &adapterTogglesDesc; options.nextInChain = &adapterTogglesDesc;
#endif #endif
ctx->webgpu_global_ctx->instance.WaitAny( instance.WaitAny(instance.RequestAdapter(
ctx->webgpu_global_ctx->instance.RequestAdapter(
&options, wgpu::CallbackMode::AllowSpontaneous, &options, wgpu::CallbackMode::AllowSpontaneous,
[&ctx](wgpu::RequestAdapterStatus status, wgpu::Adapter adapter, const char * message) { [&adapter](wgpu::RequestAdapterStatus status, wgpu::Adapter _adapter, const char * message) {
if (status != wgpu::RequestAdapterStatus::Success) { if (status != wgpu::RequestAdapterStatus::Success) {
GGML_LOG_ERROR("ggml_webgpu: Failed to get an adapter: %s\n", message); GGML_LOG_ERROR("ggml_webgpu: Failed to get an adapter: %s\n", message);
return; return;
} }
ctx->webgpu_global_ctx->adapter = std::move(adapter); adapter = std::move(_adapter);
}), }),
UINT64_MAX); UINT64_MAX);
}
static void create_webgpu_device(ggml_backend_webgpu_reg_context * ctx) {
ggml_backend_webgpu_request_adapter(ctx->webgpu_global_ctx->instance, ctx->webgpu_global_ctx->adapter);
GGML_ASSERT(ctx->webgpu_global_ctx->adapter != nullptr); GGML_ASSERT(ctx->webgpu_global_ctx->adapter != nullptr);
ctx->webgpu_global_ctx->adapter.GetLimits(&ctx->webgpu_global_ctx->capabilities.limits); ctx->webgpu_global_ctx->adapter.GetLimits(&ctx->webgpu_global_ctx->capabilities.limits);
@@ -4543,20 +4546,7 @@ ggml_backend_reg_t ggml_backend_webgpu_reg() {
// Probe for adapter support // Probe for adapter support
wgpu::Adapter adapter; wgpu::Adapter adapter;
if (ctx->webgpu_global_ctx->instance != nullptr) { if (ctx->webgpu_global_ctx->instance != nullptr) {
wgpu::RequestAdapterOptions options = {}; ggml_backend_webgpu_request_adapter(ctx->webgpu_global_ctx->instance, adapter);
// probe for adapter support
ctx->webgpu_global_ctx->instance.WaitAny(
ctx->webgpu_global_ctx->instance.RequestAdapter(
&options, wgpu::CallbackMode::AllowSpontaneous,
[&adapter](wgpu::RequestAdapterStatus status, wgpu::Adapter _adapter, const char * message) {
if (status != wgpu::RequestAdapterStatus::Success) {
GGML_LOG_ERROR("ggml_webgpu: Failed to get an adapter: %s\n", message);
return;
}
adapter = std::move(_adapter);
}),
UINT64_MAX);
} }
// WebGPU backend requires f16 support and, on native, implicit device synchronization. // WebGPU backend requires f16 support and, on native, implicit device synchronization.