From e913376d8aacd05c7122f3bc12d19138420cfa99 Mon Sep 17 00:00:00 2001 From: "Yu, Zijun" Date: Fri, 23 Jan 2026 15:49:01 +0800 Subject: [PATCH 1/2] Remove hardcode names --- ggml/src/ggml-openvino/ggml-decoder.cpp | 63 +++++++++++++------------ ggml/src/ggml-openvino/ggml-decoder.h | 8 ++-- ggml/src/ggml-openvino/utils.cpp | 4 +- 3 files changed, 39 insertions(+), 36 deletions(-) diff --git a/ggml/src/ggml-openvino/ggml-decoder.cpp b/ggml/src/ggml-openvino/ggml-decoder.cpp index b8fe6358c8d..01e2c2ff193 100644 --- a/ggml/src/ggml-openvino/ggml-decoder.cpp +++ b/ggml/src/ggml-openvino/ggml-decoder.cpp @@ -169,9 +169,11 @@ void GgmlOvDecoder::set_input_output(ggml_tensor * node, bool naive) { // TODO: The shape modification for stateful model below is not validated for all supported models yet. More generic solution might be needed // to enable additional cases. Ideally, this could be removed from decoder and done as part of a transformation later. auto stateless_kv_shape = get_graph_input_shape(node, src); - assert(stateless_kv_shape.size() == 4 && stateless_kv_shape[0] == 1 && stateless_kv_shape[1] == 1 - && stateless_kv_shape[2].is_dynamic() && stateless_kv_shape[3] == (m_model_params.n_heads_kv*m_model_params.head_size)); - stateful_kv_shape = {stateless_kv_shape[0], ov::Dimension::dynamic(), m_model_params.n_heads_kv, m_model_params.head_size}; + assert(stateless_kv_shape.size() == 4 && stateless_kv_shape[0] == 1 && + stateless_kv_shape[1] == 1 && stateless_kv_shape[2].is_dynamic() && + stateless_kv_shape[3] == (m_model_params.n_heads_kv * m_model_params.head_size)); + stateful_kv_shape = {stateless_kv_shape[0], ov::Dimension::dynamic(), + m_model_params.n_heads_kv, m_model_params.head_size}; } } } @@ -180,9 +182,8 @@ void GgmlOvDecoder::set_input_output(ggml_tensor * node, bool naive) { } m_inputs[src_name] = src; assert(stateful_kv_shape.rank().is_static()); - ov::PartialShape param_shape = (stateful_kv_shape.rank().get_length() != 0) - ? stateful_kv_shape - : get_graph_input_shape(node, src); + ov::PartialShape param_shape = + (stateful_kv_shape.rank().get_length() != 0) ? stateful_kv_shape : get_graph_input_shape(node, src); auto param_node = std::make_shared(get_ov_type(src), param_shape); param_node->set_friendly_name(src_name); param_node->output(0).get_tensor().set_names({src_name}); @@ -197,7 +198,7 @@ void GgmlOvDecoder::set_input_output(ggml_tensor * node, bool naive) { static std::set debug_output_names = {}; // Workaround: the final tensor "result_output" does not have GGML_TENSOR_FLAG_OUTPUT flag set in cgraph if (node->op == GGML_OP_SET_ROWS || node->flags & GGML_TENSOR_FLAG_OUTPUT || - node_output_name.find("output") != std::string::npos || debug_output_names.count(node_output_name)) { + debug_output_names.count(node_output_name)) { if (m_model_outputs.find(node_output_name) == m_model_outputs.end()) { m_model_outputs[node_output_name] = node_output; } @@ -312,6 +313,11 @@ std::pair GgmlOvDecoder::compute_llm_params(ggml_cgr auto * node = cgraph->nodes[i]; std::string name = std::string(node->name); if (node->op == GGML_OP_FLASH_ATTN_EXT) { + model_params.n_heads = node->src[0]->ne[2]; + model_params.n_heads_kv = node->src[1]->ne[2]; + model_params.head_size = node->src[0]->ne[0]; + compute_params.input_len = node->src[0]->ne[1]; + auto * cache_k_perm = node->src[1]; if (cache_k_perm->op == GGML_OP_CPY) { cache_k_perm = cache_k_perm->src[0]; @@ -324,9 +330,8 @@ std::pair GgmlOvDecoder::compute_llm_params(ggml_cgr int layer = extract_layer_from_name(cache_k->name); auto * mask = node->src[3]; std::string mask_name(mask->name); - assert(mask_name.find("self_kq_mask") == 0); - if (std::string(node->src[3]->name).find("swa") != std::string::npos) { + if (mask_name.find("swa") != std::string::npos) { model_params.swa_layers.push_back(layer); model_params.ctx_per_seq_swa = cache_k->ne[1]; } else { @@ -351,25 +356,18 @@ std::pair GgmlOvDecoder::compute_llm_params(ggml_cgr compute_params.attention_size_swa = model_params.ctx_per_seq_swa; compute_params.token_len_per_seq = 1; } - - } else if (node->op == GGML_OP_ROPE) { - if (name.find("Qcur-0") == 0 || std::string(node->src[0]->name).find("Qcur-0") == 0) { - model_params.head_size = node->ne[0]; - model_params.n_heads = node->ne[1]; - model_params.rope_params = node->op_params; - auto * inp_pos = node->src[1]; - compute_params.input_len = inp_pos->ne[0]; - } else if (name.find("Kcur-0") == 0 || std::string(node->src[0]->name).find("Kcur-0") == 0) { - model_params.n_heads_kv = node->ne[1]; - } - } else if (node->op == GGML_OP_GET_ROWS && std::string(node->src[1]->name) == "inp_out_ids") { - // for static case, output_len is always 1 except for llama-perplexity - compute_params.output_len = node->src[1]->ne[0]; - if (is_static && compute_params.output_len == 0) { - compute_params.output_len = 1; - } + break; + } + if (node->op == GGML_OP_ROPE) { + model_params.rope_params = node->op_params; } } + auto * output_tensor = cgraph->nodes[cgraph->n_nodes - 1]; + compute_params.output_len = output_tensor->ne[1]; + // for NPU, output_len is always 1 except for llama-perplexity + if (is_static && compute_params.output_len == 0) { + compute_params.output_len = 1; + } model_params.ctx = model_params.ctx_per_seq * model_params.n_seq; model_params.ctx_swa = model_params.ctx_per_seq_swa * model_params.n_seq; return {model_params, compute_params}; @@ -385,14 +383,17 @@ ov::PartialShape GgmlOvDecoder::get_graph_input_shape(const ggml_tensor * op, co auto name = std::string(input->name); ov::PartialShape input_shape; - if (name == "inp_tokens" || name == "inp_pos") { + if ((op->op == GGML_OP_GET_ROWS && op->src[0]->op == GGML_OP_NONE) || op->op == GGML_OP_ROPE) { + // tokens or positions int len = m_is_static ? (m_is_prefill ? m_prefill_chunk_size : 1) : -1; input_shape = ov::PartialShape{1, 1, 1, len}; - } else if (name == "inp_out_ids") { + } else if (op->op == GGML_OP_GET_ROWS) { + // output index input_shape = ov::PartialShape{1, 1, 1, m_is_static ? m_compute_params.output_len : -1}; - } else if (name.find("self_kq_mask") == 0) { + } else if (op->op == GGML_OP_CPY || op->op == GGML_OP_FLASH_ATTN_EXT) { + // mask if (m_is_static) { input_shape = ov::PartialShape{1, 1, m_is_prefill ? m_prefill_chunk_size : 1, m_model_params.ctx}; } else if (m_is_stateful) { @@ -401,7 +402,8 @@ ov::PartialShape GgmlOvDecoder::get_graph_input_shape(const ggml_tensor * op, co input_shape = ov::PartialShape{-1, 1, -1, -1}; } - } else if (name.find("cache_") == 0) { + } else if (op && op->op == GGML_OP_SET_ROWS && op->src[2] == input) { + // kvcache input_shape = ov::PartialShape{get_shape(input)}; if (!m_is_static) { // do not fix ctx size to make llama-bench work @@ -409,6 +411,7 @@ ov::PartialShape GgmlOvDecoder::get_graph_input_shape(const ggml_tensor * op, co } } else if (op && op->op == GGML_OP_SET_ROWS && op->src[1] == input) { + // kv update index int len = m_is_static ? (m_is_prefill ? m_prefill_chunk_size : 1) : -1; input_shape = ov::PartialShape{1, 1, 1, len}; diff --git a/ggml/src/ggml-openvino/ggml-decoder.h b/ggml/src/ggml-openvino/ggml-decoder.h index 4afec272e1a..c0d18b7512e 100644 --- a/ggml/src/ggml-openvino/ggml-decoder.h +++ b/ggml/src/ggml-openvino/ggml-decoder.h @@ -16,7 +16,7 @@ struct ModelParams { int ctx_swa = -1; int ctx_per_seq = -1; int ctx_per_seq_swa = -1; - int n_seq = -1; + int n_seq = 1; int n_heads = -1; int n_heads_kv = -1; int head_size = -1; @@ -37,14 +37,14 @@ struct ModelParams { }; struct ComputeParams { - int n_seq_active = -1; - int seq_active_start = -1; + int n_seq_active = 1; + int seq_active_start = 0; int attention_size = -1; int attention_size_swa = -1; int input_len = -1; int token_len_per_seq = -1; int past_kv_len = -1; - int output_len = -1; + int output_len = 1; }; class GgmlOvDecoder : public ov::frontend::ggml::GgmlDecoder { diff --git a/ggml/src/ggml-openvino/utils.cpp b/ggml/src/ggml-openvino/utils.cpp index 2d30eef941f..8c3717472b4 100644 --- a/ggml/src/ggml-openvino/utils.cpp +++ b/ggml/src/ggml-openvino/utils.cpp @@ -614,10 +614,10 @@ ov::Tensor get_ov_output_tensor(std::shared_ptr ggml_decoder, con auto output_type = ggml_decoder->get_ov_type(ggml_tensor); auto output_shape = ggml_decoder->get_shape(ggml_tensor); - if (ggml_decoder->is_static() && result_name == "result_output" && output_shape[2] == 0) { + if (ggml_decoder->is_static() && output_shape[2] == 0) { output_shape[2] = 1; } - if (ggml_decoder->is_stateful() && result_name == "result_output") { + if (ggml_decoder->is_stateful() && ggml_tensor->flags & GGML_TENSOR_FLAG_OUTPUT) { std::vector output_shape_3d; for (size_t i=1; i Date: Fri, 23 Jan 2026 15:49:36 +0800 Subject: [PATCH 2/2] Fix stateful shapes --- .../ggml-openvino/openvino/op/glu_geglu.cpp | 2 +- .../ggml-openvino/openvino/op/glu_swiglu.cpp | 2 +- ggml/src/ggml-openvino/openvino/op/rope.cpp | 22 +++++-------------- ggml/src/ggml-openvino/openvino/utils.cpp | 2 +- ggml/src/ggml-openvino/utils.cpp | 2 ++ 5 files changed, 11 insertions(+), 19 deletions(-) diff --git a/ggml/src/ggml-openvino/openvino/op/glu_geglu.cpp b/ggml/src/ggml-openvino/openvino/op/glu_geglu.cpp index ad5cd3f6ba5..8be9e8deb06 100644 --- a/ggml/src/ggml-openvino/openvino/op/glu_geglu.cpp +++ b/ggml/src/ggml-openvino/openvino/op/glu_geglu.cpp @@ -26,7 +26,7 @@ OutputVector translate_glu_geglu(const NodeContext & context) { src1 = context.get_input(1); } else { auto combined = context.get_input(0); - auto split_axis = ov::op::v0::Constant::create(ov::element::i64, {}, {3}); + auto split_axis = ov::op::v0::Constant::create(ov::element::i64, {}, {-1}); auto split = std::make_shared(combined, split_axis, 2); src0 = split->output(0); src1 = split->output(1); diff --git a/ggml/src/ggml-openvino/openvino/op/glu_swiglu.cpp b/ggml/src/ggml-openvino/openvino/op/glu_swiglu.cpp index 2b7f13629f2..6e0b85517e6 100644 --- a/ggml/src/ggml-openvino/openvino/op/glu_swiglu.cpp +++ b/ggml/src/ggml-openvino/openvino/op/glu_swiglu.cpp @@ -26,7 +26,7 @@ OutputVector translate_glu_swiglu(const NodeContext & context) { src1 = context.get_input(1); } else { auto combined = context.get_input(0); - auto split_axis = ov::op::v0::Constant::create(ov::element::i64, {}, {3}); + auto split_axis = ov::op::v0::Constant::create(ov::element::i64, {}, {-1}); auto split = std::make_shared(combined, split_axis, 2); src0 = split->output(0); src1 = split->output(1); diff --git a/ggml/src/ggml-openvino/openvino/op/rope.cpp b/ggml/src/ggml-openvino/openvino/op/rope.cpp index 01bc46131e1..44e3368217e 100644 --- a/ggml/src/ggml-openvino/openvino/op/rope.cpp +++ b/ggml/src/ggml-openvino/openvino/op/rope.cpp @@ -70,22 +70,16 @@ OutputVector translate_rope(const NodeContext & context) { constexpr int ROPE_TYPE_NORM = 0; if (mode == ROPE_TYPE_NORM) { + auto neg_one = ov::op::v0::Constant::create(ov::element::i64, {1}, {-1}); auto zero = ov::op::v0::Constant::create(ov::element::i64, {1}, {0}); auto one = ov::op::v0::Constant::create(ov::element::i64, {1}, {1}); auto two = ov::op::v0::Constant::create(ov::element::i64, {1}, {2}); auto end = ov::op::v0::Constant::create(ov::element::i64, {1}, {output_shape[3]}); Output even_slice; Output odd_slice; - int32_t unsqueeze_dim = 4; - if (context.is_stateful()) { - unsqueeze_dim = 3; - even_slice = std::make_shared(data_node, zero, end, two, two); - odd_slice = std::make_shared(data_node, one, end, two, two); - } else { - auto three = ov::op::v0::Constant::create(ov::element::i64, {1}, {3}); - even_slice = std::make_shared(data_node, zero, end, two, three); - odd_slice = std::make_shared(data_node, one, end, two, three); - } + int32_t unsqueeze_dim = context.is_stateful() ? 3 : 4; + even_slice = std::make_shared(data_node, zero, end, two, neg_one); + odd_slice = std::make_shared(data_node, one, end, two, neg_one); Output first_half = std::make_shared(std::make_shared(even_slice, cos_theta_node), @@ -105,7 +99,7 @@ OutputVector translate_rope(const NodeContext & context) { res = std::make_shared(stack, data_shape, false); } else if (mode == ROPE_TYPE_NEOX) { auto data_split = std::make_shared( - data_node, ov::op::v0::Constant::create(ov::element::i64, ov::Shape{}, {3}), 2); + data_node, ov::op::v0::Constant::create(ov::element::i64, ov::Shape{}, {-1}), 2); Output slice_data_node_0 = data_split->outputs()[0]; Output slice_data_node_1 = data_split->outputs()[1]; @@ -117,11 +111,7 @@ OutputVector translate_rope(const NodeContext & context) { std::make_shared(slice_data_node_0, sin_theta_node), std::make_shared(slice_data_node_1, cos_theta_node)); - int32_t concat_dim = 3; - if (context.is_stateful()) { - concat_dim = 2; - } - res = std::make_shared(ov::OutputVector{first_half_node, second_half_node}, concat_dim); + res = std::make_shared(ov::OutputVector{first_half_node, second_half_node}, -1); } return rename_outputs_with_suffix({res}, context.get_name()); diff --git a/ggml/src/ggml-openvino/openvino/utils.cpp b/ggml/src/ggml-openvino/openvino/utils.cpp index b7553f99c86..a0215b97b11 100644 --- a/ggml/src/ggml-openvino/openvino/utils.cpp +++ b/ggml/src/ggml-openvino/openvino/utils.cpp @@ -216,7 +216,7 @@ ov::Output process_view_input(const NodeContext & context, int input_i auto begin = ov::op::v0::Constant::create(ov::element::i64, {1}, {split_addr}); auto end = ov::op::v0::Constant::create(ov::element::i64, {1}, {slice_end}); auto stride = ov::op::v0::Constant::create(ov::element::i64, {1}, {1}); - auto axes = ov::op::v0::Constant::create(ov::element::i64, {1}, {3}); + auto axes = ov::op::v0::Constant::create(ov::element::i64, {1}, {context.is_stateful() ? 2 : 3}); auto sliced = std::make_shared(input, begin, end, stride, axes); return sliced; } diff --git a/ggml/src/ggml-openvino/utils.cpp b/ggml/src/ggml-openvino/utils.cpp index 8c3717472b4..edf42cd9854 100644 --- a/ggml/src/ggml-openvino/utils.cpp +++ b/ggml/src/ggml-openvino/utils.cpp @@ -497,6 +497,7 @@ ov::Tensor get_ov_input_tensor(std::shared_ptr ggml_decoder, cons ov::Tensor get_ov_input_tensor_static_decode(std::shared_ptr ggml_decoder, const std::string & param_name) { + // NPU decoding stage const auto * ggml_tensor = ggml_decoder->get_input_ggml_tensor(param_name); const auto * op = ggml_decoder->get_tensor_used_op(ggml_tensor); @@ -540,6 +541,7 @@ ov::Tensor get_ov_input_tensor_static_decode(std::shared_ptr ggml ov::Tensor get_ov_input_tensor_static_prefill(std::shared_ptr ggml_decoder, const std::string & param_name, int chunk_index) { + // NPU prompt processing stage const auto * ggml_tensor = ggml_decoder->get_input_ggml_tensor(param_name); const auto * op = ggml_decoder->get_tensor_used_op(ggml_tensor);