port k-quants to new matvec

Author: reeselevine

ggml/src/ggml-webgpu/ggml-webgpu-shader-lib.hpp

@@ -1338,7 +1338,9 @@ class ggml_webgpu_shader_lib {
         const bool use_row_tiled =
             context.src0->type == GGML_TYPE_F32 || context.src0->type == GGML_TYPE_F16 || context.src0->type == GGML_TYPE_Q4_0 ||
             context.src0->type == GGML_TYPE_Q4_1 || context.src0->type == GGML_TYPE_Q5_0 || context.src0->type == GGML_TYPE_Q5_1 ||
-            context.src0->type == GGML_TYPE_Q8_0 || context.src0->type == GGML_TYPE_Q8_1;
+            context.src0->type == GGML_TYPE_Q8_0 || context.src0->type == GGML_TYPE_Q8_1 || context.src0->type == GGML_TYPE_Q6_K ||
+            context.src0->type == GGML_TYPE_Q4_K || context.src0->type == GGML_TYPE_Q5_K || context.src0->type == GGML_TYPE_Q3_K ||
+            context.src0->type == GGML_TYPE_Q2_K;
         ggml_webgpu_mul_mat_vec_pipeline_key key = {
             .src0_type  = context.src0->type,
             .src1_type  = context.src1->type,

ggml/src/ggml-webgpu/ggml-webgpu.cpp

@@ -1329,17 +1329,12 @@ static webgpu_encoded_op ggml_webgpu_mul_mat(webgpu_context &       ctx,
                 case GGML_TYPE_Q5_1:
                 case GGML_TYPE_Q8_0:
                 case GGML_TYPE_Q8_1:
-                    use_fast = true;
-                    break;
                 case GGML_TYPE_Q6_K:
-                    use_fast = !is_vec || ctx->global_ctx->capabilities.supports_subgroups;
-                    break;
-                case GGML_TYPE_Q2_K:
-                case GGML_TYPE_Q3_K:
                 case GGML_TYPE_Q4_K:
                 case GGML_TYPE_Q5_K:
-                    // we don't have fast mat-vec for these types, but we do have (semi) fast mat-mat
-                    use_fast = !is_vec;
+                case GGML_TYPE_Q3_K:
+                case GGML_TYPE_Q2_K:
+                    use_fast = true;
                     break;
                 default:
                     break;