Address PR comments: clean up comments

Author: s-Nick

ggml/src/ggml-sycl/ggml-sycl.cpp

@@ -3106,7 +3106,6 @@ static void reorder_qw_q6_k(uint8_t * data_device, size_t size, size_t offset, d
 
     auto *       ql_ptr     = data_device;
     auto *       qh_ptr     = ql_ptr + (QK_K / 2) * nblocks;
-    // scales are after all quants' bits so adding both to get correct offset
     auto *       scales_ptr = qh_ptr + (QK_K / 4) * nblocks;
     sycl::half * dm_ptr     = (sycl::half *) (scales_ptr + (QK_K / 16) * nblocks);
 

ggml/src/ggml-sycl/mmvq.cpp

@@ -32,7 +32,7 @@ static void mul_mat_vec_q_reorder(const void * __restrict__ vx, const void * __r
     float partial_sum = 0.0f;
     for (int i = sg.get_local_linear_id() / block_elements_per_subgroup; i < blocks_per_row; i += blocks_per_subgroup) {
         const int ibx       = row * blocks_per_row + i;  // x block index
-        // TODO: Generalize offsets, right now only works for quantizations that don't split high and low bits
+
         const auto         bx_offset      = block_type::get_block_offset(ibx, nblocks);
         const auto         d_offset       = block_type::get_d_offset(nrows, ncols, ibx);
         // Y block index that aligns with ibx

ggml/src/ggml-sycl/vecdotq.hpp

@@ -403,7 +403,6 @@ template <> struct reorder_vec_dot_q_sycl<GGML_TYPE_Q6_K> {
     using q6_k_block  = ggml_sycl_reordered::block_q_t<GGML_TYPE_Q6_K>;
     using q6_k_traits = typename q6_k_block::traits;
 
-    // contiguous v/x values
     __dpct_inline__ float vec_dot_q6_K_q8_1_impl_mmvq(const int & vl, const int & vh, const int * __restrict__ u,
                                                       const int8_t * __restrict__ scales, const float d,
                                                       const float * __restrict__ d8) {