diff --git a/conditioner.hpp b/conditioner.hpp index fbf1325..0d58ffd 100644 --- a/conditioner.hpp +++ b/conditioner.hpp @@ -1708,6 +1708,9 @@ struct LLMEmbedder : public Conditioner { int prompt_template_encode_start_idx = 34; int max_length = 0; std::set out_layers; + std::vector tokens; + std::vector weights; + std::vector mask; if (llm->enable_vision && conditioner_params.ref_images.size() > 0) { LOG_INFO("QwenImageEditPlusPipeline"); prompt_template_encode_start_idx = 64; @@ -1795,6 +1798,7 @@ struct LLMEmbedder : public Conditioner { prompt += "<|im_end|>\n<|im_start|>assistant\n"; } else if (version == VERSION_FLUX2_KLEIN) { prompt_template_encode_start_idx = 0; + max_length = 512; out_layers = {9, 18, 27}; prompt = "<|im_start|>user\n"; @@ -1804,6 +1808,16 @@ struct LLMEmbedder : public Conditioner { prompt_attn_range.second = static_cast(prompt.size()); prompt += "<|im_end|>\n<|im_start|>assistant\n\n\n\n\n"; + + auto tokens_and_weights = tokenize(prompt, prompt_attn_range, 0, false); + tokens = std::get<0>(tokens_and_weights); + weights = std::get<1>(tokens_and_weights); + + mask.insert(mask.end(), tokens.size(), 1.f); + if (tokens.size() < max_length) { + mask.insert(mask.end(), max_length - tokens.size(), 0.f); + tokenizer->pad_tokens(tokens, weights, max_length, true); + } } else if (version == VERSION_OVIS_IMAGE) { prompt_template_encode_start_idx = 28; max_length = prompt_template_encode_start_idx + 256; @@ -1827,17 +1841,34 @@ struct LLMEmbedder : public Conditioner { prompt += "<|im_end|>\n<|im_start|>assistant\n"; } - auto tokens_and_weights = tokenize(prompt, prompt_attn_range, max_length, max_length > 0); - auto& tokens = std::get<0>(tokens_and_weights); - auto& weights = std::get<1>(tokens_and_weights); + if (tokens.empty()) { + auto tokens_and_weights = tokenize(prompt, prompt_attn_range, max_length, max_length > 0); + tokens = std::get<0>(tokens_and_weights); + weights = std::get<1>(tokens_and_weights); + } int64_t t0 = ggml_time_ms(); struct ggml_tensor* hidden_states = nullptr; // [N, n_token, 3584] auto input_ids = vector_to_ggml_tensor_i32(work_ctx, tokens); + ggml_tensor* attention_mask = nullptr; + if (!mask.empty()) { + attention_mask = ggml_new_tensor_2d(work_ctx, GGML_TYPE_F32, mask.size(), mask.size()); + ggml_ext_tensor_iter(attention_mask, [&](ggml_tensor* attention_mask, int64_t i0, int64_t i1, int64_t i2, int64_t i3) { + float value = 0.f; + if (mask[i0] == 0.f) { + value = -INFINITY; + } else if (i0 > i1) { + value = -INFINITY; + } + ggml_ext_tensor_set_f32(attention_mask, value, i0, i1, i2, i3); + }); + } + llm->compute(n_threads, input_ids, + attention_mask, image_embeds, out_layers, &hidden_states, @@ -1861,7 +1892,7 @@ struct LLMEmbedder : public Conditioner { GGML_ASSERT(hidden_states->ne[1] > prompt_template_encode_start_idx); int64_t min_length = 0; - if (sd_version_is_flux2(version)) { + if (version == VERSION_FLUX2) { min_length = 512; } diff --git a/flux.hpp b/flux.hpp index 6b43940..6ac4731 100644 --- a/flux.hpp +++ b/flux.hpp @@ -1354,8 +1354,8 @@ namespace Flux { flux_params.num_heads = static_cast(flux_params.hidden_size / head_dim); - LOG_INFO("double blocks: %d, single blocks: %d, guidance_embed: %s, context_in_dim: %" PRId64 - ", hidden_size: %" PRId64 ", num_heads: %d", + LOG_INFO("flux: depth = %d, depth_single_blocks = %d, guidance_embed = %s, context_in_dim = %" PRId64 + ", hidden_size = %" PRId64 ", num_heads = %d", flux_params.depth, flux_params.depth_single_blocks, flux_params.guidance_embed ? "true" : "false", diff --git a/ggml_extend.hpp b/ggml_extend.hpp index 6f498ff..24b3831 100644 --- a/ggml_extend.hpp +++ b/ggml_extend.hpp @@ -1228,7 +1228,7 @@ __STATIC_INLINE__ struct ggml_tensor* ggml_ext_attention_ext(struct ggml_context bool diag_mask_inf = false, bool skip_reshape = false, bool flash_attn = false, - float kv_scale = 1.0f) { // avoid overflow + float kv_scale = 1.0f / 128.f) { // avoid overflow int64_t L_q; int64_t L_k; int64_t C; @@ -1348,6 +1348,7 @@ __STATIC_INLINE__ struct ggml_tensor* ggml_ext_attention_ext(struct ggml_context v = ggml_reshape_3d(ctx, v, L_k, d_head, n_kv_head * N); // [N * n_kv_head, d_head, L_k] auto kq = ggml_mul_mat(ctx, k, q); // [N * n_head, L_q, L_k] + ggml_mul_mat_set_prec(kq, GGML_PREC_F32); kq = ggml_scale_inplace(ctx, kq, scale); if (mask) { kq = ggml_add_inplace(ctx, kq, mask); @@ -2183,7 +2184,7 @@ public: bool bias = true, bool force_f32 = false, bool force_prec_f32 = false, - float scale = 1.f) + float scale = 1.f / 128.f) : in_features(in_features), out_features(out_features), bias(bias), diff --git a/llm.hpp b/llm.hpp index 67b1ea1..781774d 100644 --- a/llm.hpp +++ b/llm.hpp @@ -837,7 +837,8 @@ namespace LLM { struct ggml_tensor* forward(GGMLRunnerContext* ctx, struct ggml_tensor* x, - struct ggml_tensor* input_pos) { + struct ggml_tensor* input_pos, + struct ggml_tensor* attention_mask = nullptr) { // x: [N, n_token, hidden_size] int64_t n_token = x->ne[1]; int64_t N = x->ne[2]; @@ -880,7 +881,7 @@ namespace LLM { k = ggml_cont(ctx->ggml_ctx, ggml_ext_torch_permute(ctx->ggml_ctx, k, 0, 2, 1, 3)); // [N, num_kv_heads, n_token, head_dim] k = ggml_reshape_3d(ctx->ggml_ctx, k, k->ne[0], k->ne[1], k->ne[2] * k->ne[3]); // [N*num_kv_heads, n_token, head_dim] - x = ggml_ext_attention_ext(ctx->ggml_ctx, ctx->backend, q, k, v, num_heads, nullptr, true, true, false); // [N, n_token, hidden_size] + x = ggml_ext_attention_ext(ctx->ggml_ctx, ctx->backend, q, k, v, num_heads, attention_mask, false, true, false); // [N, n_token, hidden_size] x = out_proj->forward(ctx, x); // [N, n_token, hidden_size] return x; @@ -898,7 +899,8 @@ namespace LLM { struct ggml_tensor* forward(GGMLRunnerContext* ctx, struct ggml_tensor* x, - struct ggml_tensor* input_pos) { + struct ggml_tensor* input_pos, + struct ggml_tensor* attention_mask = nullptr) { // x: [N, n_token, hidden_size] auto self_attn = std::dynamic_pointer_cast(blocks["self_attn"]); auto mlp = std::dynamic_pointer_cast(blocks["mlp"]); @@ -907,7 +909,7 @@ namespace LLM { auto residual = x; x = input_layernorm->forward(ctx, x); - x = self_attn->forward(ctx, x, input_pos); + x = self_attn->forward(ctx, x, input_pos, attention_mask); x = ggml_add_inplace(ctx->ggml_ctx, x, residual); residual = x; @@ -936,6 +938,7 @@ namespace LLM { struct ggml_tensor* forward(GGMLRunnerContext* ctx, struct ggml_tensor* input_ids, struct ggml_tensor* input_pos, + struct ggml_tensor* attention_mask, std::vector> image_embeds, std::set out_layers) { // input_ids: [N, n_token] @@ -990,7 +993,7 @@ namespace LLM { for (int i = 0; i < num_layers; i++) { auto block = std::dynamic_pointer_cast(blocks["layers." + std::to_string(i)]); - x = block->forward(ctx, x, input_pos); + x = block->forward(ctx, x, input_pos, attention_mask); if (out_layers.find(i + 1) != out_layers.end()) { intermediate_outputs.push_back(x); } @@ -1036,12 +1039,13 @@ namespace LLM { struct ggml_tensor* forward(GGMLRunnerContext* ctx, struct ggml_tensor* input_ids, struct ggml_tensor* input_pos, + struct ggml_tensor* attention_mask, std::vector> image_embeds, std::set out_layers) { // input_ids: [N, n_token] auto model = std::dynamic_pointer_cast(blocks["model"]); - auto x = model->forward(ctx, input_ids, input_pos, image_embeds, out_layers); + auto x = model->forward(ctx, input_ids, input_pos, attention_mask, image_embeds, out_layers); return x; } @@ -1063,6 +1067,7 @@ namespace LLM { LLM model; std::vector input_pos_vec; + std::vector attention_mask_vec; std::vector window_mask_vec; std::vector window_index_vec; std::vector window_inverse_index_vec; @@ -1157,9 +1162,10 @@ namespace LLM { struct ggml_tensor* forward(GGMLRunnerContext* ctx, struct ggml_tensor* input_ids, struct ggml_tensor* input_pos, + struct ggml_tensor* attention_mask, std::vector> image_embeds, std::set out_layers) { - auto hidden_states = model.forward(ctx, input_ids, input_pos, image_embeds, out_layers); // [N, n_token, hidden_size] + auto hidden_states = model.forward(ctx, input_ids, input_pos, attention_mask, image_embeds, out_layers); // [N, n_token, hidden_size] return hidden_states; } @@ -1174,6 +1180,7 @@ namespace LLM { } struct ggml_cgraph* build_graph(struct ggml_tensor* input_ids, + struct ggml_tensor* attention_mask, std::vector> image_embeds, std::set out_layers) { struct ggml_cgraph* gf = ggml_new_graph(compute_ctx); @@ -1205,9 +1212,26 @@ namespace LLM { input_pos_vec.size()); set_backend_tensor_data(input_pos, input_pos_vec.data()); + if (attention_mask != nullptr) { + attention_mask = to_backend(attention_mask); + } else { + attention_mask_vec.resize(n_tokens * n_tokens); + for (int i0 = 0; i0 < n_tokens; i0++) { + for (int i1 = 0; i1 < n_tokens; i1++) { + float value = 0.f; + if (i0 > i1) { + value = -INFINITY; + } + attention_mask_vec[i1 * n_tokens + i0] = value; + } + } + attention_mask = ggml_new_tensor_2d(compute_ctx, GGML_TYPE_F32, n_tokens, n_tokens); + set_backend_tensor_data(attention_mask, attention_mask_vec.data()); + } + auto runner_ctx = get_context(); - struct ggml_tensor* hidden_states = forward(&runner_ctx, input_ids, input_pos, image_embeds, out_layers); + struct ggml_tensor* hidden_states = forward(&runner_ctx, input_ids, input_pos, attention_mask, image_embeds, out_layers); ggml_build_forward_expand(gf, hidden_states); @@ -1216,12 +1240,13 @@ namespace LLM { bool compute(const int n_threads, struct ggml_tensor* input_ids, + struct ggml_tensor* attention_mask, std::vector> image_embeds, std::set out_layers, ggml_tensor** output, ggml_context* output_ctx = nullptr) { auto get_graph = [&]() -> struct ggml_cgraph* { - return build_graph(input_ids, image_embeds, out_layers); + return build_graph(input_ids, attention_mask, image_embeds, out_layers); }; return GGMLRunner::compute(get_graph, n_threads, true, output, output_ctx); } @@ -1525,7 +1550,7 @@ namespace LLM { struct ggml_tensor* out = nullptr; int64_t t0 = ggml_time_ms(); - model.compute(8, input_ids, image_embeds, {}, &out, work_ctx); + model.compute(8, input_ids, nullptr, image_embeds, {}, &out, work_ctx); int64_t t1 = ggml_time_ms(); print_ggml_tensor(out); @@ -1565,7 +1590,7 @@ namespace LLM { struct ggml_tensor* out = nullptr; int64_t t0 = ggml_time_ms(); - model.compute(8, input_ids, {}, {10, 20, 30}, &out, work_ctx); + model.compute(8, input_ids, nullptr, {}, {10, 20, 30}, &out, work_ctx); int64_t t1 = ggml_time_ms(); print_ggml_tensor(out); @@ -1588,7 +1613,7 @@ namespace LLM { struct ggml_tensor* out = nullptr; int64_t t0 = ggml_time_ms(); - model.compute(8, input_ids, {}, {35}, &out, work_ctx); + model.compute(8, input_ids, nullptr, {}, {35}, &out, work_ctx); int64_t t1 = ggml_time_ms(); print_ggml_tensor(out); @@ -1611,7 +1636,7 @@ namespace LLM { struct ggml_tensor* out = nullptr; int64_t t0 = ggml_time_ms(); - model.compute(8, input_ids, {}, {}, &out, work_ctx); + model.compute(8, input_ids, nullptr, {}, {}, &out, work_ctx); int64_t t1 = ggml_time_ms(); print_ggml_tensor(out);