LLM inference in C/C++
C++ ggml
ggml-cuda : support ssm_conv kernel size 7 (#30873)
Tarek Dakhran39da130fcd1b
diff --git a/ggml/src/ggml-cuda/ssm-conv.cu b/ggml/src/ggml-cuda/ssm-conv.cuindex 1463169cf7..f41438b318 100644--- a/ggml/src/ggml-cuda/ssm-conv.cu+++ b/ggml/src/ggml-cuda/ssm-conv.cu@@ -151,9 +151,10 @@ static void ssm_conv_f32_cuda(const float * src0, const float * src1, const floacase 3: launch_kernel(std::integral_constant<int, 3 >{}); break;case 4: launch_kernel(std::integral_constant<int, 4 >{}); break;case 5: launch_kernel(std::integral_constant<int, 5 >{}); break;+ case 7: launch_kernel(std::integral_constant<int, 7 >{}); break;case 9: launch_kernel(std::integral_constant<int, 9 >{}); break;case 15: launch_kernel(std::integral_constant<int, 15>{}); break;- default: GGML_ABORT("Only support kernel sizes 3, 4, 5, 9, 15 right now.");+ default: GGML_ABORT("Only support kernel sizes 3, 4, 5, 7, 9, 15 right now.");}}diff --git a/tests/test-backend-ops.cpp b/tests/test-backend-ops.cppindex 485629cf64..d3194bf47f 100644--- a/tests/test-backend-ops.cpp+++ b/tests/test-backend-ops.cpp@@ -10659,7 +10659,7 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {// ne[2] > 65535 exceeds the CUDA gridDim.y limit (#27901)test_cases.emplace_back(new test_rms_norm_mul_rope({4, 1, 65536, 1}, 1e-6f, false, false, false, GGML_ROPE_TYPE_NORMAL));test_cases.emplace_back(new test_rms_norm_mul_rope({4, 1, 65536, 1}, 1e-6f, false, true, false, GGML_ROPE_TYPE_NEOX));- for (int64_t d_conv : {3, 4, 9}) {+ for (int64_t d_conv : {3, 4, 7, 9}) {for (int64_t d_inner: {1024, 1536, 2048}) {test_cases.emplace_back(new test_ssm_conv(GGML_TYPE_F32, {d_conv, d_inner, 1, 1}, {d_conv, d_inner, 1, 1}));test_cases.emplace_back(new test_ssm_conv(GGML_TYPE_F32, {2 * d_conv, d_inner, 1, 1}, {d_conv, d_inner, 1, 1}));@@ -10672,7 +10672,7 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {// fused ssm_conv + (optional) bias_add + silu. The bias-only graph (no silu) is intentionally// not tested since there's no fusion for that pattern in ggml_cuda_can_fuse.- for (int64_t d_conv : {3, 4, 9}) {+ for (int64_t d_conv : {3, 4, 7, 9}) {for (int64_t d_inner : {1024, 1536, 2048}) {for (bool fuse_bias : {false, true}) {// short token path (n_t <= 32)