ggml webgpu: Fix bug in dispatching large matrix-vector multiplication (llama/19535)

* Fix bug in dispatching large matrix-vector multiplication

diff --git a/ggml/src/ggml-webgpu/ggml-webgpu.cpp b/ggml/src/ggml-webgpu/ggml-webgpu.cpp
index 17bb2f47126247c228fd2b54cb41be9b011715fa..b5fee480562bd1bb3fdfdde10dc39d70eb362db3 100644 (file)
--- a/ggml/src/ggml-webgpu/ggml-webgpu.cpp
+++ b/ggml/src/ggml-webgpu/ggml-webgpu.cpp
@@ -1121,7 +1121,8 @@ static webgpu_command ggml_webgpu_mul_mat(webgpu_context & ctx,
         uint32_t batches       = dst->ne[2] * dst->ne[3];
         uint32_t output_groups = CEIL_DIV(dst->ne[0], decisions->outputs_per_wg);
         uint32_t total_wg      = output_groups * batches;
-        wg_x                   = total_wg % ctx->global_ctx->capabilities.limits.maxComputeWorkgroupsPerDimension;
+        // TODO: split large sizes into multiple batches to avoid way over-provisioning workgroups
+        wg_x = std::min(total_wg, ctx->global_ctx->capabilities.limits.maxComputeWorkgroupsPerDimension);
         wg_y = CEIL_DIV(total_wg, ctx->global_ctx->capabilities.limits.maxComputeWorkgroupsPerDimension);
     } else if (use_fast) {
         auto decisions = static_cast<ggml_webgpu_mul_mat_shader_decisions *>(pipeline.context.get());