From 722509e57f9ce1c564e759fba0ed326ff7c1ce35 Mon Sep 17 00:00:00 2001 From: Eve <139727413+netrunnereve@users.noreply.github.com> Date: Wed, 31 Dec 2025 16:48:36 -0500 Subject: [PATCH] q5_k --- .../ggml-vulkan/vulkan-shaders/mul_mm_funcs.glsl | 14 +++++++------- .../vulkan-shaders/vulkan-shaders-gen.cpp | 2 +- 2 files changed, 8 insertions(+), 8 deletions(-) diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_funcs.glsl b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_funcs.glsl index b97749cd9e..c91626a178 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_funcs.glsl +++ b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_funcs.glsl @@ -208,8 +208,8 @@ void load_a_to_shmem(const uint pos_a, const uint row, const uint col, const uin const uint idx = pos_a + col * p.stride_a / LOAD_VEC_A + row; const uint buf_idx = col * SHMEM_STRIDE + row * LOAD_VEC_A / 2; - const uint ib = idx / 128; // 2 values per idx - const uint iqs = idx % 128; // 0..127 + const uint ib = idx / 64; // 4 values per idx + const uint iqs = (idx % 64) * 2; // 0,2,4..126 const uint n = iqs / 32; // 0,1,2,3 const uint b = (iqs % 32) / 16; // 0,1 @@ -236,12 +236,12 @@ void load_a_to_shmem(const uint pos_a, const uint row, const uint col, const uin const float d = loadd.x * sc; const float m = -loadd.y * mbyte; - const uint qs = (uint(data_a_packed16[ib].qs[qsi / 2]) >> (b * 4)) & 0x0F0F; - const uint qh = ((uint(data_a_packed16[ib].qh[qhi / 2]) >> (iqs / 16)) & 0x0101) << 4; - const vec2 q = vec2(unpack8(qs | qh).xy); + const uint qs = (data_a_packed32[ib].qs[qsi / 4] >> (b * 4)) & 0x0F0F0F0F; + const uint qh = ((data_a_packed32[ib].qh[qhi / 4] >> (iqs / 16)) & 0x01010101) << 4; + const vec4 q = vec4(unpack8(qs | qh)); - buf_a[buf_idx] = FLOAT_TYPE_VEC2(fma(d, q.x, m), - fma(d, q.y, m)); + buf_a[buf_idx ] = FLOAT_TYPE_VEC2(fma(d, q.x, m), fma(d, q.y, m)); + buf_a[buf_idx + 1] = FLOAT_TYPE_VEC2(fma(d, q.z, m), fma(d, q.w, m)); #elif defined(DATA_A_Q6_K) const uint idx = pos_a + col * p.stride_a / LOAD_VEC_A + row; const uint buf_idx = col * SHMEM_STRIDE + row * LOAD_VEC_A / 2; diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp b/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp index 44836a3bd2..2dded1e6a5 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp +++ b/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp @@ -554,7 +554,7 @@ void matmul_shaders(bool fp16, MatMulIdType matmul_id_type, bool coopmat, bool c std::string load_vec_quant = "2"; if ((tname == "q4_0") || (tname == "q4_1") || (tname == "iq1_s") || (tname == "iq1_m") || (tname == "iq2_xxs") || (tname == "iq2_xs") || (tname == "iq2_s")) load_vec_quant = "8"; - else if ((tname == "q5_0") || (tname == "q5_1") || (tname == "q8_0") || (tname == "q4_k") || (tname == "iq3_xxs") || (tname == "iq3_s") || (tname == "iq4_nl") || (tname == "mxfp4")) + else if ((tname == "q5_0") || (tname == "q5_1") || (tname == "q8_0") || (tname == "q4_k") || (tname == "q5_k") || (tname == "iq3_xxs") || (tname == "iq3_s") || (tname == "iq4_nl") || (tname == "mxfp4")) load_vec_quant = "4"; if (tname == "bf16") {