diff --git a/tests/hrx-loom-jit-test.cpp b/tests/hrx-loom-jit-test.cpp index fcbb81ae56d1..23e5373433e1 100644 --- a/tests/hrx-loom-jit-test.cpp +++ b/tests/hrx-loom-jit-test.cpp @@ -239,6 +239,10 @@ static std::map flash_attention_config(const char * qu { "ggml.flash_attention.key_value_head_count", key_value_head_count }, { "ggml.flash_attention.qk_head_size", head_size }, { "ggml.flash_attention.value_head_size", head_size }, + // engine#2138: this line's prefill kernel takes the V row stride as a config (#95, MLA shapes); + // dispatch-flash-attention.cpp emits it, and test-hrx-loom-jit.cpp passes the same value. + { "ggml.flash_attention.value_stride", + std::to_string(std::stoll(key_value_head_count) * std::stoll(head_size)) }, { "ggml.flash_attention.attention_scale", attention_scale }, { "ggml.flash_attention.apply_gate", "0" }, { "ggml.flash_attention.gate_stride_head", "1" }, @@ -829,6 +833,12 @@ static void run_loom_jit_compile_checks() { { "column_count", 1024 }, }, { + // engine#2138: this line's transpose is specialized on its shape + // (6c4a83f26); the values dispatch-flash-attention.cpp emits for a + // 8192-token V with the default spread pitch (8192 % 64 == 0 -> +32) + { "ggml.copy_transpose_f16.row_count", "8192" }, + { "ggml.copy_transpose_f16.column_count", "1024" }, + { "ggml.copy_transpose_f16.output_pitch", "8224" }, }); require_compiled_kernel(transpose_ref); auto transposed_attention_config = flash_attention_config("24", "4", "256", "0.0625"); @@ -1369,6 +1379,10 @@ static void run_binary_fusion_compile_checks() { { { "llm.ssm_conv.generic.d_conv", "4" }, { "llm.ssm_conv.generic.d_inner", "64" }, + // engine#2138: this line's generic SSM_CONV specializes on the token and + // sequence counts too (dispatch-ssm-conv.cpp sets both) + { "llm.ssm_conv.generic.n_t", "8" }, + { "llm.ssm_conv.generic.n_s", "1" }, { "llm.ssm_conv.generic.unary_op", "23" }, { "llm.ssm_conv.generic.binary_op", "3" }, { "llm.ssm_conv.generic.binary_lhs", "0" },