diff --git a/tests/test-hrx-ops.cpp b/tests/test-hrx-ops.cpp index c9afe389c98a..a6719e2a8b97 100644 --- a/tests/test-hrx-ops.cpp +++ b/tests/test-hrx-ops.cpp @@ -2344,8 +2344,6 @@ static void run_qwen_full_cache_prefill_flash_attention_scheduling_case(int64_t constexpr int64_t kQueryHeadCount = 32; constexpr int64_t kKeyValueHeadCount = 4; constexpr int64_t kFullCacheRowCount = 40960; - const size_t active_mask_byte_count = - static_cast(token_count) * static_cast(token_count) * sizeof(ggml_fp16_t); ggml_init_params params = {}; params.mem_size = 128 * 1024 * 1024; @@ -2387,38 +2385,40 @@ static void run_qwen_full_cache_prefill_flash_attention_scheduling_case(int64_t match_dispatch_at_index(imported.graph, *registry, plan, covered_nodes, producer_index_for_tensor(imported.graph, flash_output), flash_match); + // engine#2374 item 3: since the engine#2138 merge (#327) the postprocess match no longer publishes a compact + // [tokens x tokens] F16 mask through a qwen_attention_metadata initialisation dispatch. The flash attention reads + // the graph's own mask over the whole key/value view, one query chunk at a time, each chunk binding its own rows. + // A real graph's view is llama_kv_cache::get_n_kv() (the used cells padded to 256), so the full-capacity view + // this case builds only occurs in the reserve graph or a full cache, where every row is needed anyway. const ggml::hrx::Value * mask_value = imported.graph.values().find_tensor(attention.attention_mask); REQUIRE(mask_value != nullptr); - const ggml::hrx::CommandPlanAlternateValue * compact_mask = - ggml::hrx::find_alternate_value(plan, mask_value->id, GGML_TYPE_F16, active_mask_byte_count); - REQUIRE(compact_mask != nullptr); + REQUIRE(mask_value->ne[0] == kFullCacheRowCount); + REQUIRE(mask_value->ne[1] == token_count); + const size_t mask_row_bytes = static_cast(kFullCacheRowCount) * sizeof(ggml_fp16_t); - const ggml::hrx::Dispatch * metadata_dispatch = nullptr; for (const ggml::hrx::Dispatch & dispatch : plan.initialization_dispatches) { - if (kernel_name_for_id(dispatch.kernel.kernel_id) == "qwen3_moe:qwen_attention_metadata") { - metadata_dispatch = &dispatch; - } + REQUIRE(kernel_name_for_id(dispatch.kernel.kernel_id) != "qwen3_moe:qwen_attention_metadata"); } - REQUIRE(metadata_dispatch != nullptr); - REQUIRE(metadata_dispatch->kernel.integer_parameters.at("token_count") == token_count); - REQUIRE(metadata_dispatch->kernel.integer_parameters.at("context_capacity") == token_count); - REQUIRE(metadata_dispatch->bindings.size() == 5); - REQUIRE(metadata_dispatch->bindings[4].value == compact_mask->alternate_value); - REQUIRE(metadata_dispatch->bindings[4].length == active_mask_byte_count); - const ggml::hrx::Dispatch * flash_dispatch = nullptr; + int64_t covered_query_tokens = 0; + size_t flash_dispatch_count = 0; for (const ggml::hrx::Dispatch & dispatch : plan.dispatches) { - if (kernel_name_for_id(dispatch.kernel.kernel_id) == "loom_libs:ggml_flash_attention_f32_f16_wmma") { - flash_dispatch = &dispatch; + if (kernel_name_for_id(dispatch.kernel.kernel_id) != "loom_libs:ggml_flash_attention_f32_f16_wmma") { + continue; } + ++flash_dispatch_count; + const int64_t chunk_tokens = dispatch.kernel.integer_parameters.at("query_token_count"); + REQUIRE(chunk_tokens > 0); + REQUIRE(dispatch.kernel.integer_parameters.at("key_value_token_count") == kFullCacheRowCount); + REQUIRE(dispatch.bindings.size() == 6); + REQUIRE(dispatch.kernel.compile_parameters.at("ggml.flash_attention.apply_gate") == "0"); + REQUIRE(dispatch.bindings[3].value == mask_value->id); + REQUIRE(dispatch.bindings[3].offset == static_cast(covered_query_tokens) * mask_row_bytes); + REQUIRE(dispatch.bindings[3].length == static_cast(chunk_tokens) * mask_row_bytes); + covered_query_tokens += chunk_tokens; } - REQUIRE(flash_dispatch != nullptr); - REQUIRE(flash_dispatch->kernel.integer_parameters.at("query_token_count") == token_count); - REQUIRE(flash_dispatch->kernel.integer_parameters.at("key_value_token_count") == token_count); - REQUIRE(flash_dispatch->bindings.size() == 6); - REQUIRE(flash_dispatch->kernel.compile_parameters.at("ggml.flash_attention.apply_gate") == "0"); - REQUIRE(flash_dispatch->bindings[3].value == compact_mask->alternate_value); - REQUIRE(flash_dispatch->bindings[3].length == active_mask_byte_count); + REQUIRE(flash_dispatch_count > 0); + REQUIRE(covered_query_tokens == token_count); ggml_free(ctx); }