From 5245729e3317064959faefc5e5cbc63f4e9cfbea Mon Sep 17 00:00:00 2001 From: Jeff Bolz Date: Thu, 23 Jan 2025 01:01:17 -0600 Subject: [PATCH 1/8] vulkan: fix diag_mask_inf (#11323) With robustbufferaccess disabled, this shader was showing OOB stores. There is a bounds check in the code, but the workgrouop dimensions were reversed vs CUDA and it was running the wrong number of threads. So fix the workgroup dimensions and disable robustness for this pipeline. --- ggml/src/ggml-vulkan/ggml-vulkan.cpp | 2 +- ggml/src/ggml-vulkan/vulkan-shaders/diag_mask_inf.comp | 2 +- 2 files changed, 2 insertions(+), 2 deletions(-) diff --git a/ggml/src/ggml-vulkan/ggml-vulkan.cpp b/ggml/src/ggml-vulkan/ggml-vulkan.cpp index 78c2f5c45..eae4f0739 100644 --- a/ggml/src/ggml-vulkan/ggml-vulkan.cpp +++ b/ggml/src/ggml-vulkan/ggml-vulkan.cpp @@ -2012,7 +2012,7 @@ static void ggml_vk_load_shaders(vk_device& device) { ggml_vk_create_pipeline(device, device->pipeline_leaky_relu_f32, "leaky_relu_f32", leaky_relu_f32_len, leaky_relu_f32_data, "main", 2, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1); ggml_vk_create_pipeline(device, device->pipeline_tanh_f32, "tanh_f32", tanh_f32_len, tanh_f32_data, "main", 2, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1); - ggml_vk_create_pipeline(device, device->pipeline_diag_mask_inf_f32, "diag_mask_inf_f32", diag_mask_inf_f32_len, diag_mask_inf_f32_data, "main", 2, sizeof(vk_op_diag_mask_push_constants), {512, 1, 1}, {}, 1); + ggml_vk_create_pipeline(device, device->pipeline_diag_mask_inf_f32, "diag_mask_inf_f32", diag_mask_inf_f32_len, diag_mask_inf_f32_data, "main", 2, sizeof(vk_op_diag_mask_push_constants), {1, 512, 1}, {}, 1, true); ggml_vk_create_pipeline(device, device->pipeline_soft_max_f32, "soft_max_f32", soft_max_f32_len, soft_max_f32_data, "main", 3, sizeof(vk_op_soft_max_push_constants), {1, 1, 1}, { device->subgroup_size }, 1); ggml_vk_create_pipeline(device, device->pipeline_soft_max_f32_wg512, "soft_max_f32_wg512", soft_max_f32_len, soft_max_f32_data, "main", 3, sizeof(vk_op_soft_max_push_constants), {1, 1, 1}, { 512 }, 1); diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/diag_mask_inf.comp b/ggml/src/ggml-vulkan/vulkan-shaders/diag_mask_inf.comp index 4e68742b5..26d8bc22a 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/diag_mask_inf.comp +++ b/ggml/src/ggml-vulkan/vulkan-shaders/diag_mask_inf.comp @@ -12,7 +12,7 @@ layout (push_constant) uniform parameter #include "types.comp" -layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in; +layout(local_size_x = 1, local_size_y = 512, local_size_z = 1) in; layout (binding = 0) readonly buffer X {A_TYPE data_a[];}; layout (binding = 1) writeonly buffer D {D_TYPE data_d[];}; From 1971adf55e3ebbfab83771d6174d37b77c352c71 Mon Sep 17 00:00:00 2001 From: Jeff Bolz Date: Thu, 23 Jan 2025 01:07:50 -0600 Subject: [PATCH 2/8] vulkan: sort shaders for more deterministic binary (#11315) Fixes #11306. --- ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) 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 8bcb64101..e9c6cb9d4 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp +++ b/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp @@ -17,13 +17,13 @@ #include #include #include +#include #include #include #ifdef _WIN32 #include #include // For _mkdir on Windows - #include // For std::replace on w64devkit #else #include #include @@ -502,6 +502,7 @@ void write_output_files() { fprintf(hdr, "#include \n\n"); fprintf(src, "#include \"%s\"\n\n", basename(target_hpp).c_str()); + std::sort(shader_fnames.begin(), shader_fnames.end()); for (const auto& pair : shader_fnames) { const std::string& name = pair.first; #ifdef _WIN32 From 955a6c2d91480954e90469517c4b91c4052c6a59 Mon Sep 17 00:00:00 2001 From: amd-dwang Date: Thu, 23 Jan 2025 15:14:28 +0800 Subject: [PATCH 3/8] Vulkan-run-test: fix mmq_wg_denoms (#11343) There should be a copy-and-paste error here. *mmq_wg_denoms should be used together with *warptile_mmq, instead of wg_denoms. --- ggml/src/ggml-vulkan/ggml-vulkan.cpp | 88 ++++++++++++++-------------- 1 file changed, 44 insertions(+), 44 deletions(-) diff --git a/ggml/src/ggml-vulkan/ggml-vulkan.cpp b/ggml/src/ggml-vulkan/ggml-vulkan.cpp index eae4f0739..c325416d1 100644 --- a/ggml/src/ggml-vulkan/ggml-vulkan.cpp +++ b/ggml/src/ggml-vulkan/ggml-vulkan.cpp @@ -1673,31 +1673,31 @@ static void ggml_vk_load_shaders(vk_device& device) { CREATE_MM2(pipeline_matmul_f16_f32, matmul_f16_f32, wg_denoms, warptile, vk_mat_mat_push_constants, 3, ); if (device->coopmat_acc_f16_support) { - CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0].f16acc, matmul_q4_0_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); - CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1].f16acc, matmul_q4_1_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); - CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0].f16acc, matmul_q5_0_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); - CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1].f16acc, matmul_q5_1_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); - CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0].f16acc, matmul_q8_0_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); + CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0].f16acc, matmul_q4_0_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); + CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1].f16acc, matmul_q4_1_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); + CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0].f16acc, matmul_q5_0_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); + CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1].f16acc, matmul_q5_1_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); + CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0].f16acc, matmul_q8_0_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); - CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K].f16acc, matmul_q2_k_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); - CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K].f16acc, matmul_q3_k_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); - CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K].f16acc, matmul_q4_k_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); - CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K].f16acc, matmul_q5_k_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); - CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K].f16acc, matmul_q6_k_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); - CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL].f16acc, matmul_iq4_nl_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); + CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K].f16acc, matmul_q2_k_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); + CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K].f16acc, matmul_q3_k_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); + CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K].f16acc, matmul_q4_k_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); + CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K].f16acc, matmul_q5_k_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); + CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K].f16acc, matmul_q6_k_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); + CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL].f16acc, matmul_iq4_nl_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); } else { - CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0].f16acc, matmul_q4_0_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); - CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1].f16acc, matmul_q4_1_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); - CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0].f16acc, matmul_q5_0_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); - CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1].f16acc, matmul_q5_1_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); - CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0].f16acc, matmul_q8_0_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); + CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0].f16acc, matmul_q4_0_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); + CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1].f16acc, matmul_q4_1_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); + CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0].f16acc, matmul_q5_0_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); + CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1].f16acc, matmul_q5_1_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); + CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0].f16acc, matmul_q8_0_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); - CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K].f16acc, matmul_q2_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); - CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K].f16acc, matmul_q3_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); - CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K].f16acc, matmul_q4_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); - CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K].f16acc, matmul_q5_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); - CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K].f16acc, matmul_q6_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); - CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL].f16acc, matmul_iq4_nl_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); + CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K].f16acc, matmul_q2_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); + CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K].f16acc, matmul_q3_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); + CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K].f16acc, matmul_q4_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); + CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K].f16acc, matmul_q5_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); + CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K].f16acc, matmul_q6_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); + CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL].f16acc, matmul_iq4_nl_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); } // If there's not enough shared memory for row_ids and the result tile, don't create these pipelines. @@ -1707,31 +1707,31 @@ static void ggml_vk_load_shaders(vk_device& device) { CREATE_MM2(pipeline_matmul_id_f16_f32, matmul_id_f16_f32, wg_denoms, warptile, vk_mat_mat_push_constants, 4, _id); if (device->coopmat_acc_f16_support) { - CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0].f16acc, matmul_id_q4_0_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); - CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1].f16acc, matmul_id_q4_1_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); - CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0].f16acc, matmul_id_q5_0_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); - CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1].f16acc, matmul_id_q5_1_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); - CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0].f16acc, matmul_id_q8_0_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0].f16acc, matmul_id_q4_0_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1].f16acc, matmul_id_q4_1_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0].f16acc, matmul_id_q5_0_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1].f16acc, matmul_id_q5_1_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0].f16acc, matmul_id_q8_0_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); - CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K].f16acc, matmul_id_q2_k_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); - CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K].f16acc, matmul_id_q3_k_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); - CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K].f16acc, matmul_id_q4_k_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); - CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K].f16acc, matmul_id_q5_k_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); - CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K].f16acc, matmul_id_q6_k_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); - CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL].f16acc, matmul_id_iq4_nl_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K].f16acc, matmul_id_q2_k_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K].f16acc, matmul_id_q3_k_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K].f16acc, matmul_id_q4_k_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K].f16acc, matmul_id_q5_k_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K].f16acc, matmul_id_q6_k_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL].f16acc, matmul_id_iq4_nl_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); } else { - CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0].f16acc, matmul_id_q4_0_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); - CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1].f16acc, matmul_id_q4_1_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); - CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0].f16acc, matmul_id_q5_0_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); - CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1].f16acc, matmul_id_q5_1_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); - CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0].f16acc, matmul_id_q8_0_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0].f16acc, matmul_id_q4_0_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1].f16acc, matmul_id_q4_1_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0].f16acc, matmul_id_q5_0_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1].f16acc, matmul_id_q5_1_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0].f16acc, matmul_id_q8_0_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); - CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K].f16acc, matmul_id_q2_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); - CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K].f16acc, matmul_id_q3_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); - CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K].f16acc, matmul_id_q4_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); - CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K].f16acc, matmul_id_q5_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); - CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K].f16acc, matmul_id_q6_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); - CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL].f16acc, matmul_id_iq4_nl_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K].f16acc, matmul_id_q2_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K].f16acc, matmul_id_q3_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K].f16acc, matmul_id_q4_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K].f16acc, matmul_id_q5_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K].f16acc, matmul_id_q6_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL].f16acc, matmul_id_iq4_nl_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); } } #undef CREATE_MM2 From f211d1dc10332b7e89a4abd1041903fa63bfed27 Mon Sep 17 00:00:00 2001 From: Eric Curtin Date: Thu, 23 Jan 2025 10:38:20 +0000 Subject: [PATCH 4/8] Treat hf.co/ prefix the same as hf:// (#11350) ollama uses hf.co/ to specify huggingface prefix, like RamaLama uses hf:// Treat them similarly. Signed-off-by: Eric Curtin --- examples/run/run.cpp | 17 ++++++++++------- 1 file changed, 10 insertions(+), 7 deletions(-) diff --git a/examples/run/run.cpp b/examples/run/run.cpp index e567ad716..c710d4326 100644 --- a/examples/run/run.cpp +++ b/examples/run/run.cpp @@ -634,20 +634,20 @@ class LlamaData { return path.substr(pos + 1); } - int remove_proto(std::string & model_) { - const std::string::size_type pos = model_.find("://"); + int rm_until_substring(std::string & model_, const std::string & substring) { + const std::string::size_type pos = model_.find(substring); if (pos == std::string::npos) { return 1; } - model_ = model_.substr(pos + 3); // Skip past "://" + model_ = model_.substr(pos + substring.size()); // Skip past the substring return 0; } int resolve_model(std::string & model_) { int ret = 0; if (string_starts_with(model_, "file://") || std::filesystem::exists(model_)) { - remove_proto(model_); + rm_until_substring(model_, "://"); return ret; } @@ -656,13 +656,16 @@ class LlamaData { const std::vector headers = { "--header", "Accept: application/vnd.docker.distribution.manifest.v2+json" }; if (string_starts_with(model_, "hf://") || string_starts_with(model_, "huggingface://")) { - remove_proto(model_); + rm_until_substring(model_, "://"); + ret = huggingface_dl(model_, headers, bn); + } else if (string_starts_with(model_, "hf.co/")) { + rm_until_substring(model_, "hf.co/"); ret = huggingface_dl(model_, headers, bn); } else if (string_starts_with(model_, "ollama://")) { - remove_proto(model_); + rm_until_substring(model_, "://"); ret = ollama_dl(model_, headers, bn); } else if (string_starts_with(model_, "https://")) { - download(model_, headers, bn, true); + ret = download(model_, headers, bn, true); } else { ret = ollama_dl(model_, headers, bn); } From 58456616408c828a1ce6d2481940fd1c97bce3b5 Mon Sep 17 00:00:00 2001 From: Xuan Son Nguyen Date: Thu, 23 Jan 2025 13:56:05 +0100 Subject: [PATCH 5/8] server : add more clean up when cancel_tasks is called (#11340) * server : add more clean up when cancel_tasks is called * fix recv_with_timeout * std::remove_if * fix std::remove_if --- examples/server/server.cpp | 43 +++++++++++++++++++++++++++++++------- 1 file changed, 35 insertions(+), 8 deletions(-) diff --git a/examples/server/server.cpp b/examples/server/server.cpp index 4cfb3c9bb..a94c3822c 100644 --- a/examples/server/server.cpp +++ b/examples/server/server.cpp @@ -1433,6 +1433,10 @@ struct server_queue { } else { queue_tasks.push_back(std::move(task)); } + // if this is cancel task make sure to clean up pending tasks + if (task.type == SERVER_TASK_TYPE_CANCEL) { + cleanup_pending_task(task.id_target); + } condition_tasks.notify_one(); return task.id; } @@ -1450,6 +1454,10 @@ struct server_queue { } else { queue_tasks.push_back(std::move(task)); } + // if this is cancel task make sure to clean up pending tasks + if (task.type == SERVER_TASK_TYPE_CANCEL) { + cleanup_pending_task(task.id_target); + } } condition_tasks.notify_one(); return 0; @@ -1544,6 +1552,20 @@ struct server_queue { } } } + +private: + void cleanup_pending_task(int id_task) { + // no need lock because this is called exclusively by post() + auto rm_func = [id_task](const server_task & task) { + return task.id_target == id_task; + }; + queue_tasks.erase( + std::remove_if(queue_tasks.begin(), queue_tasks.end(), rm_func), + queue_tasks.end()); + queue_tasks_deferred.erase( + std::remove_if(queue_tasks_deferred.begin(), queue_tasks_deferred.end(), rm_func), + queue_tasks_deferred.end()); + } }; struct server_response { @@ -1579,6 +1601,12 @@ struct server_response { std::unique_lock lock(mutex_results); waiting_task_ids.erase(id_task); + // make sure to clean up all pending results + queue_results.erase( + std::remove_if(queue_results.begin(), queue_results.end(), [id_task](const server_task_result_ptr & res) { + return res->id == id_task; + }), + queue_results.end()); } void remove_waiting_task_ids(const std::unordered_set & id_tasks) { @@ -1598,7 +1626,7 @@ struct server_response { return !queue_results.empty(); }); - for (int i = 0; i < (int) queue_results.size(); i++) { + for (size_t i = 0; i < queue_results.size(); i++) { if (id_tasks.find(queue_results[i]->id) != id_tasks.end()) { server_task_result_ptr res = std::move(queue_results[i]); queue_results.erase(queue_results.begin() + i); @@ -1615,12 +1643,6 @@ struct server_response { server_task_result_ptr recv_with_timeout(const std::unordered_set & id_tasks, int timeout) { while (true) { std::unique_lock lock(mutex_results); - bool cr_res = condition_results.wait_for(lock, std::chrono::seconds(timeout), [&]{ - return !queue_results.empty(); - }); - if (!cr_res) { - return nullptr; - } for (int i = 0; i < (int) queue_results.size(); i++) { if (id_tasks.find(queue_results[i]->id) != id_tasks.end()) { @@ -1629,6 +1651,11 @@ struct server_response { return res; } } + + std::cv_status cr_res = condition_results.wait_for(lock, std::chrono::seconds(timeout)); + if (cr_res == std::cv_status::timeout) { + return nullptr; + } } // should never reach here @@ -2376,8 +2403,8 @@ struct server_context { server_task task(SERVER_TASK_TYPE_CANCEL); task.id_target = id_task; - cancel_tasks.push_back(task); queue_results.remove_waiting_task_id(id_task); + cancel_tasks.push_back(task); } // push to beginning of the queue, so it has highest priority queue_tasks.post(cancel_tasks, true); From f7fb43cd0b84280c261f440dc8e85eafad4a0ca6 Mon Sep 17 00:00:00 2001 From: Eric Curtin Date: Thu, 23 Jan 2025 16:16:18 +0000 Subject: [PATCH 6/8] Add -ngl (#11372) Most other llama.cpp cli tools accept -ngl with a single dash. Signed-off-by: Eric Curtin --- examples/run/run.cpp | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) diff --git a/examples/run/run.cpp b/examples/run/run.cpp index c710d4326..715f21b5d 100644 --- a/examples/run/run.cpp +++ b/examples/run/run.cpp @@ -147,7 +147,8 @@ class Opt { if (handle_option_with_value(argc, argv, i, context_size) == 1) { return 1; } - } else if (options_parsing && (strcmp(argv[i], "-n") == 0 || strcmp(argv[i], "--ngl") == 0)) { + } else if (options_parsing && + (strcmp(argv[i], "-n") == 0 || strcmp(argv[i], "-ngl") == 0 || strcmp(argv[i], "--ngl") == 0)) { if (handle_option_with_value(argc, argv, i, ngl) == 1) { return 1; } From 05f63cc9ee859de07f585f7b12939345f39ada8b Mon Sep 17 00:00:00 2001 From: Eric Curtin Date: Thu, 23 Jan 2025 20:04:31 +0000 Subject: [PATCH 7/8] Update documentation (#11373) To show -n, -ngl, --ngl is acceptable. Signed-off-by: Eric Curtin --- examples/run/run.cpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/examples/run/run.cpp b/examples/run/run.cpp index 715f21b5d..92a49eb74 100644 --- a/examples/run/run.cpp +++ b/examples/run/run.cpp @@ -195,7 +195,7 @@ class Opt { "Options:\n" " -c, --context-size \n" " Context size (default: %d)\n" - " -n, --ngl \n" + " -n, -ngl, --ngl \n" " Number of GPU layers (default: %d)\n" " --temp \n" " Temperature (default: %.1f)\n" From 564804b79b78df1469ec8646869972de5e885ec4 Mon Sep 17 00:00:00 2001 From: Jeff Bolz Date: Thu, 23 Jan 2025 14:51:24 -0600 Subject: [PATCH 8/8] tests: fix some mul_mat test gaps (#11375) Now that we have batched mat-vec mul Vulkan shaders for up to n==8, these tests weren't actually exercising the mat-mat mul path. Test n==9 as well. Also, change to use all_types. --- tests/test-backend-ops.cpp | 15 ++++----------- 1 file changed, 4 insertions(+), 11 deletions(-) diff --git a/tests/test-backend-ops.cpp b/tests/test-backend-ops.cpp index 74d1bee39..381956a04 100644 --- a/tests/test-backend-ops.cpp +++ b/tests/test-backend-ops.cpp @@ -3909,17 +3909,10 @@ static std::vector> make_test_cases_eval() { test_cases.emplace_back(new test_gla(GGML_TYPE_F32, 32, 64, 32, 4)); test_cases.emplace_back(new test_gla(GGML_TYPE_F32, 32, 64, 128, 4)); - for (int i = 1; i < 9; ++i) { - test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F16, GGML_TYPE_F32, 16, i, 256, { 1, 1}, {1, 1})); - test_cases.emplace_back(new test_mul_mat(GGML_TYPE_Q4_0, GGML_TYPE_F32, 16, i, 256, { 1, 1}, {1, 1})); - test_cases.emplace_back(new test_mul_mat(GGML_TYPE_Q4_1, GGML_TYPE_F32, 16, i, 256, { 1, 1}, {1, 1})); - test_cases.emplace_back(new test_mul_mat(GGML_TYPE_Q5_0, GGML_TYPE_F32, 16, i, 256, { 1, 1}, {1, 1})); - test_cases.emplace_back(new test_mul_mat(GGML_TYPE_Q5_1, GGML_TYPE_F32, 16, i, 256, { 1, 1}, {1, 1})); - test_cases.emplace_back(new test_mul_mat(GGML_TYPE_Q8_0, GGML_TYPE_F32, 16, i, 256, { 1, 1}, {1, 1})); - test_cases.emplace_back(new test_mul_mat(GGML_TYPE_Q4_K, GGML_TYPE_F32, 16, i, 256, { 1, 1}, {1, 1})); - test_cases.emplace_back(new test_mul_mat(GGML_TYPE_Q5_K, GGML_TYPE_F32, 16, i, 256, { 1, 1}, {1, 1})); - test_cases.emplace_back(new test_mul_mat(GGML_TYPE_Q6_K, GGML_TYPE_F32, 16, i, 256, { 1, 1}, {1, 1})); - test_cases.emplace_back(new test_mul_mat(GGML_TYPE_IQ4_NL, GGML_TYPE_F32, 16, i, 256, { 1, 1}, {1, 1})); + for (ggml_type type_a : all_types) { + for (int i = 1; i < 10; ++i) { + test_cases.emplace_back(new test_mul_mat(type_a, GGML_TYPE_F32, 16, i, 256, { 1, 1}, {1, 1})); + } } #if 1