From 37ae068deed9ad2dc0eefb605b664ee514fcf1bd Mon Sep 17 00:00:00 2001 From: Concedo <39025047+LostRuins@users.noreply.github.com> Date: Sun, 22 Feb 2026 17:03:43 +0800 Subject: [PATCH] set default to GPU test --- .../qwen3tts/audio_tokenizer_decoder.cpp | 306 +++++++++--------- .../qwen3tts/audio_tokenizer_encoder.cpp | 240 +++++++------- otherarch/qwen3tts/gguf_loader.cpp | 37 ++- otherarch/qwen3tts/gguf_loader.h | 30 +- otherarch/qwen3tts/tts_transformer.cpp | 4 +- 5 files changed, 310 insertions(+), 307 deletions(-) diff --git a/otherarch/qwen3tts/audio_tokenizer_decoder.cpp b/otherarch/qwen3tts/audio_tokenizer_decoder.cpp index 8129ab0cc..fa4bc5097 100644 --- a/otherarch/qwen3tts/audio_tokenizer_decoder.cpp +++ b/otherarch/qwen3tts/audio_tokenizer_decoder.cpp @@ -19,7 +19,7 @@ AudioTokenizerDecoder::~AudioTokenizerDecoder() { void AudioTokenizerDecoder::unload_model() { free_audio_decoder_model(model_); - + if (state_.sched) { ggml_backend_sched_free(state_.sched); state_.sched = nullptr; @@ -39,32 +39,32 @@ void AudioTokenizerDecoder::unload_model() { void AudioTokenizerDecoder::normalize_codebooks() { const float epsilon = 1e-5f; - + auto normalize_codebook = [epsilon](struct ggml_tensor * codebook, struct ggml_tensor * usage, const char *) { if (!codebook || !usage || !codebook->data || !usage->data) return; - + int64_t codebook_dim = codebook->ne[0]; int64_t codebook_size = codebook->ne[1]; - + ggml_fp16_t * cb_data = (ggml_fp16_t *)codebook->data; float * usage_data = (float *)usage->data; - + for (int64_t emb_idx = 0; emb_idx < codebook_size; ++emb_idx) { float u = usage_data[emb_idx]; if (u < epsilon) u = epsilon; float inv_u = 1.0f / u; - + for (int64_t dim_idx = 0; dim_idx < codebook_dim; ++dim_idx) { int64_t mem_idx = dim_idx + emb_idx * codebook_dim; float val = ggml_fp16_to_fp32(cb_data[mem_idx]); cb_data[mem_idx] = ggml_fp32_to_fp16(val * inv_u); } } - + }; - + normalize_codebook(model_.vq_first_codebook, model_.vq_first_usage, "first"); - + for (int i = 0; i < 15; ++i) { char name[16]; snprintf(name, sizeof(name), "rest%d", i); @@ -80,11 +80,11 @@ bool AudioTokenizerDecoder::load_model(const std::string & model_path) { error_msg_ = loader.get_error(); return false; } - + model_.config.sample_rate = loader.get_u32("qwen3-tts.tokenizer.sample_rate", 24000); model_.config.n_codebooks = loader.get_u32("qwen3-tts.tokenizer.num_codebooks", 16); model_.config.codebook_size = loader.get_u32("qwen3-tts.tokenizer.codebook_size", 2048); - + int64_t n_tensors = loader.get_n_tensors(); int dec_tensor_count = 0; for (int64_t i = 0; i < n_tensors; ++i) { @@ -93,48 +93,48 @@ bool AudioTokenizerDecoder::load_model(const std::string & model_path) { dec_tensor_count++; } } - + if (dec_tensor_count == 0) { error_msg_ = "No decoder tensors found in model"; return false; } - + size_t ctx_size = ggml_tensor_overhead() * dec_tensor_count; struct ggml_init_params params = { /*.mem_size =*/ ctx_size, /*.mem_buffer =*/ nullptr, /*.no_alloc =*/ true, }; - + model_.ctx = ggml_init(params); if (!model_.ctx) { error_msg_ = "Failed to initialize GGML context"; return false; } - + struct gguf_context * gguf_ctx = loader.get_ctx(); struct ggml_context * meta_ctx = loader.get_meta_ctx(); - + for (int64_t i = 0; i < n_tensors; ++i) { const char * name = loader.get_tensor_name(i); if (!name || strncmp(name, "tok_dec.", 8) != 0) { continue; } - - + + struct ggml_tensor * meta_tensor = ggml_get_tensor(meta_ctx, name); if (!meta_tensor) { continue; } - + struct ggml_tensor * tensor = ggml_dup_tensor(model_.ctx, meta_tensor); ggml_set_name(tensor, name); - + model_.tensors[name] = tensor; - + std::string sname(name); - + if (sname == "tok_dec.vq_first.input_proj.weight") model_.vq_first_input_proj = tensor; else if (sname == "tok_dec.vq_first.output_proj.weight") model_.vq_first_output_proj = tensor; else if (sname == "tok_dec.vq_first.0.codebook") model_.vq_first_codebook = tensor; @@ -174,13 +174,13 @@ bool AudioTokenizerDecoder::load_model(const std::string & model_path) { int blk_idx, res_idx, cb_idx, n = 0; char suffix[64]; size_t name_len = strlen(name); - - + + #define MATCH1(fmt, var) (sscanf(name, fmt "%n", &var, &n) == 1 && (size_t)n == name_len) #define MATCH2(fmt, v1, v2) (sscanf(name, fmt "%n", &v1, &v2, &n) == 2 && (size_t)n == name_len) #define MATCH1S(fmt, var, suf) (sscanf(name, fmt, &var, suf) == 2) - + if (MATCH1("tok_dec.vq_rest.%d.codebook", cb_idx)) { if (cb_idx >= 0 && cb_idx < 15) { model_.vq_rest_codebook[cb_idx] = tensor; @@ -314,19 +314,19 @@ bool AudioTokenizerDecoder::load_model(const std::string & model_path) { #undef MATCH1S } } - + if (!load_tensor_data_from_file(model_path, gguf_ctx, model_.ctx, model_.tensors, model_.buffer, error_msg_, GGML_BACKEND_DEVICE_TYPE_IGPU)) { return false; } - + for (int i = 0; i < 4; ++i) { model_.dec_blocks[i].res[0].dilation = 1; model_.dec_blocks[i].res[1].dilation = 3; model_.dec_blocks[i].res[2].dilation = 9; } - + normalize_codebooks(); // Codebooks are normalized in host memory; sync once to backend tensors. auto upload_if_present = [](struct ggml_tensor * t) { @@ -338,8 +338,8 @@ bool AudioTokenizerDecoder::load_model(const std::string & model_path) { for (int i = 0; i < 15; ++i) { upload_if_present(model_.vq_rest_codebook[i]); } - - state_.backend = init_preferred_backend("AudioTokenizerDecoder", &error_msg_); + + state_.backend = init_preferred_backend("AudioTokenizerDecoder", &error_msg_, true); if (!state_.backend) { return false; } @@ -347,7 +347,7 @@ bool AudioTokenizerDecoder::load_model(const std::string & model_path) { ggml_backend_dev_t device = ggml_backend_get_device(state_.backend); const char * device_name = device ? ggml_backend_dev_name(device) : "Unknown"; fprintf(stderr, " AudioTokenizerDecoder backend: %s\n", device_name); - + if (device && ggml_backend_dev_type(device) != GGML_BACKEND_DEVICE_TYPE_CPU) { state_.backend_cpu = ggml_backend_init_by_type(GGML_BACKEND_DEVICE_TYPE_CPU, nullptr); if (!state_.backend_cpu) { @@ -366,9 +366,9 @@ bool AudioTokenizerDecoder::load_model(const std::string & model_path) { error_msg_ = "Failed to create backend scheduler"; return false; } - + state_.compute_meta.resize(ggml_tensor_overhead() * QWEN3_TTS_DEC_MAX_NODES + ggml_graph_overhead()); - + return true; } @@ -379,25 +379,25 @@ struct ggml_tensor * AudioTokenizerDecoder::apply_snake(struct ggml_context * ct int64_t seq_len = x->ne[0]; int64_t channels = x->ne[1]; int64_t batch = x->ne[2]; - + struct ggml_tensor * alpha_exp = ggml_exp(ctx, alpha); - + struct ggml_tensor * alpha_3d = ggml_reshape_3d(ctx, alpha_exp, 1, channels, 1); - struct ggml_tensor * alpha_broad = ggml_repeat(ctx, alpha_3d, + struct ggml_tensor * alpha_broad = ggml_repeat(ctx, alpha_3d, ggml_new_tensor_3d(ctx, GGML_TYPE_F32, seq_len, channels, batch)); - + struct ggml_tensor * ax = ggml_mul(ctx, x, alpha_broad); struct ggml_tensor * sin_ax = ggml_sin(ctx, ax); struct ggml_tensor * sin_sq = ggml_sqr(ctx, sin_ax); - + struct ggml_tensor * neg_beta = ggml_scale(ctx, beta, -1.0f); struct ggml_tensor * inv_beta_exp = ggml_exp(ctx, neg_beta); struct ggml_tensor * inv_beta_3d = ggml_reshape_3d(ctx, inv_beta_exp, 1, channels, 1); - struct ggml_tensor * inv_beta = ggml_repeat(ctx, inv_beta_3d, + struct ggml_tensor * inv_beta = ggml_repeat(ctx, inv_beta_3d, ggml_new_tensor_3d(ctx, GGML_TYPE_F32, seq_len, channels, batch)); - + struct ggml_tensor * scaled_sin = ggml_mul(ctx, sin_sq, inv_beta); - + return ggml_add(ctx, x, scaled_sin); } @@ -418,72 +418,72 @@ struct ggml_tensor * AudioTokenizerDecoder::apply_pre_tfm_layer(struct ggml_cont const int n_heads = cfg.n_heads; const int qkv_dim = cfg.latent_dim; const int head_dim = qkv_dim / n_heads; - + if (!layer.attn_norm_w || !layer.attn_q_w || !layer.attn_k_w || !layer.attn_v_w || - !layer.attn_output_w || !layer.ffn_norm_w || !layer.ffn_gate_w || + !layer.attn_output_w || !layer.ffn_norm_w || !layer.ffn_gate_w || !layer.ffn_up_w || !layer.ffn_down_w) { return x; } - + struct ggml_tensor * residual = x; - + struct ggml_tensor * normed = apply_rms_norm(ctx, x, layer.attn_norm_w, cfg.rms_norm_eps); - + struct ggml_tensor * Qcur = ggml_mul_mat(ctx, layer.attn_q_w, normed); struct ggml_tensor * Kcur = ggml_mul_mat(ctx, layer.attn_k_w, normed); struct ggml_tensor * Vcur = ggml_mul_mat(ctx, layer.attn_v_w, normed); - + Qcur = ggml_reshape_3d(ctx, Qcur, head_dim, n_heads, n_frames); Kcur = ggml_reshape_3d(ctx, Kcur, head_dim, n_heads, n_frames); Vcur = ggml_reshape_3d(ctx, Vcur, head_dim, n_heads, n_frames); - + Qcur = ggml_rope_ext(ctx, Qcur, positions, nullptr, head_dim, GGML_ROPE_TYPE_NEOX, 0, cfg.rope_theta, 1.0f, 0.0f, 1.0f, 0.0f, 0.0f); - + Kcur = ggml_rope_ext(ctx, Kcur, positions, nullptr, head_dim, GGML_ROPE_TYPE_NEOX, 0, cfg.rope_theta, 1.0f, 0.0f, 1.0f, 0.0f, 0.0f); - + struct ggml_tensor * Q = ggml_permute(ctx, Qcur, 0, 2, 1, 3); struct ggml_tensor * K = ggml_permute(ctx, Kcur, 0, 2, 1, 3); struct ggml_tensor * V = ggml_permute(ctx, Vcur, 0, 2, 1, 3); - + struct ggml_tensor * KQ = ggml_mul_mat(ctx, K, Q); KQ = ggml_scale(ctx, KQ, 1.0f / sqrtf((float)head_dim)); // Apply causal mask (each position can only attend to itself and previous positions) KQ = ggml_diag_mask_inf(ctx, KQ, 0); KQ = ggml_soft_max(ctx, KQ); - + V = ggml_cont(ctx, ggml_transpose(ctx, V)); - + struct ggml_tensor * KQV = ggml_mul_mat(ctx, V, KQ); KQV = ggml_permute(ctx, KQV, 0, 2, 1, 3); struct ggml_tensor * attn_out = ggml_cont_2d(ctx, KQV, n_heads * head_dim, n_frames); - + attn_out = ggml_mul_mat(ctx, layer.attn_output_w, attn_out); - + if (layer.attn_scale) { attn_out = ggml_mul(ctx, attn_out, layer.attn_scale); } - + x = ggml_add(ctx, residual, attn_out); residual = x; - + normed = apply_rms_norm(ctx, x, layer.ffn_norm_w, cfg.rms_norm_eps); - + struct ggml_tensor * gate = ggml_mul_mat(ctx, layer.ffn_gate_w, normed); struct ggml_tensor * up = ggml_mul_mat(ctx, layer.ffn_up_w, normed); - + gate = ggml_silu(ctx, gate); struct ggml_tensor * ffn_out = ggml_mul(ctx, gate, up); - + ffn_out = ggml_mul_mat(ctx, layer.ffn_down_w, ffn_out); - + if (layer.ffn_scale) { ffn_out = ggml_mul(ctx, ffn_out, layer.ffn_scale); } - + return ggml_add(ctx, residual, ffn_out); } @@ -493,19 +493,19 @@ struct ggml_tensor * AudioTokenizerDecoder::apply_upsample_block(struct ggml_con int block_idx) { int64_t seq_len = x->ne[0]; int64_t channels = x->ne[1]; - + struct ggml_tensor * x_2d = ggml_reshape_2d(ctx, x, seq_len, channels); x_2d = ggml_conv_transpose_1d(ctx, block.conv_w, x_2d, 2, 0, 1); - + int64_t new_seq_len = x_2d->ne[0]; x = ggml_reshape_3d(ctx, x_2d, new_seq_len, channels, 1); - + if (block.conv_b) { x = ggml_add(ctx, x, ggml_reshape_3d(ctx, block.conv_b, 1, channels, 1)); } - + struct ggml_tensor * residual = x; - + if (block.dwconv_w) { // Causal padding: pad left with 6 zeros (kernel_size - 1 = 7 - 1 = 6) x = ggml_pad_ext(ctx, x, 6, 0, 0, 0, 0, 0, 0, 0); // left pad only @@ -514,37 +514,37 @@ struct ggml_tensor * AudioTokenizerDecoder::apply_upsample_block(struct ggml_con x = ggml_add(ctx, x, ggml_reshape_3d(ctx, block.dwconv_b, 1, channels, 1)); } } - + x = ggml_permute(ctx, x, 1, 0, 2, 3); x = ggml_cont(ctx, x); - + if (block.norm_w && block.norm_b) { x = ggml_norm(ctx, x, 1e-6f); x = ggml_mul(ctx, x, block.norm_w); x = ggml_add(ctx, x, block.norm_b); } - + x = ggml_mul_mat(ctx, block.pwconv1_w, x); if (block.pwconv1_b) { x = ggml_add(ctx, x, block.pwconv1_b); } - + x = ggml_gelu(ctx, x); - + x = ggml_mul_mat(ctx, block.pwconv2_w, x); if (block.pwconv2_b) { x = ggml_add(ctx, x, block.pwconv2_b); } - + x = ggml_permute(ctx, x, 1, 0, 2, 3); x = ggml_cont(ctx, x); - + if (block.gamma) { struct ggml_tensor * gamma_3d = ggml_reshape_3d(ctx, block.gamma, 1, channels, 1); - x = ggml_mul(ctx, x, ggml_repeat(ctx, gamma_3d, + x = ggml_mul(ctx, x, ggml_repeat(ctx, gamma_3d, ggml_new_tensor_3d(ctx, GGML_TYPE_F32, new_seq_len, channels, 1))); } - + return ggml_add(ctx, residual, x); } @@ -552,11 +552,11 @@ struct ggml_tensor * AudioTokenizerDecoder::apply_residual_block(struct ggml_con struct ggml_tensor * x, const residual_block & block) { struct ggml_tensor * residual = x; - + if (block.act1_alpha) { x = apply_snake(ctx, x, block.act1_alpha, block.act1_beta); } - + int64_t out_channels = block.conv1_w->ne[2]; int padding = 6 * block.dilation; x = ggml_pad_ext(ctx, x, padding, 0, 0, 0, 0, 0, 0, 0); @@ -564,17 +564,17 @@ struct ggml_tensor * AudioTokenizerDecoder::apply_residual_block(struct ggml_con if (block.conv1_b) { x = ggml_add(ctx, x, ggml_reshape_3d(ctx, block.conv1_b, 1, out_channels, 1)); } - + if (block.act2_alpha) { x = apply_snake(ctx, x, block.act2_alpha, block.act2_beta); } - + out_channels = block.conv2_w->ne[2]; x = ggml_conv_1d(ctx, block.conv2_w, x, 1, 0, 1); if (block.conv2_b) { x = ggml_add(ctx, x, ggml_reshape_3d(ctx, block.conv2_b, 1, out_channels, 1)); } - + return ggml_add(ctx, residual, x); } @@ -586,101 +586,101 @@ struct ggml_tensor * AudioTokenizerDecoder::apply_decoder_block(struct ggml_cont if (block.snake_alpha && block.snake_beta) { x = apply_snake(ctx, x, block.snake_alpha, block.snake_beta); } - + int64_t seq_len = x->ne[0]; int64_t in_channels = x->ne[1]; int64_t out_channels = block.conv_t_w->ne[1]; int kernel_size = block.conv_t_w->ne[0]; - + struct ggml_tensor * x_2d = ggml_reshape_2d(ctx, x, seq_len, in_channels); x_2d = ggml_conv_transpose_1d(ctx, block.conv_t_w, x_2d, upsample_rate, 0, 1); - + int64_t new_seq_len = x_2d->ne[0]; x = ggml_reshape_3d(ctx, x_2d, new_seq_len, out_channels, 1); - + // Python CausalTransConvNet: left_pad = right_pad = kernel_size - stride int pad = kernel_size - upsample_rate; int left_pad = pad; int right_pad = pad; int64_t out_seq_len = new_seq_len - left_pad - right_pad; - + x = ggml_view_3d(ctx, x, out_seq_len, out_channels, 1, x->nb[1], x->nb[2], left_pad * x->nb[0]); x = ggml_cont(ctx, x); - + if (block.conv_t_b) { x = ggml_add(ctx, x, ggml_reshape_3d(ctx, block.conv_t_b, 1, out_channels, 1)); } - + for (int i = 0; i < 3; ++i) { x = apply_residual_block(ctx, x, block.res[i]); } - + return x; } struct ggml_cgraph * AudioTokenizerDecoder::build_graph(int32_t n_frames) { const auto & cfg = model_.config; - + struct ggml_init_params params = { /*.mem_size =*/ state_.compute_meta.size(), /*.mem_buffer =*/ state_.compute_meta.data(), /*.no_alloc =*/ true, }; - + struct ggml_context * ctx0 = ggml_init(params); struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, QWEN3_TTS_DEC_MAX_NODES, false); - + static const char * cb_names[16] = { "codes_cb0", "codes_cb1", "codes_cb2", "codes_cb3", "codes_cb4", "codes_cb5", "codes_cb6", "codes_cb7", "codes_cb8", "codes_cb9", "codes_cb10", "codes_cb11", "codes_cb12", "codes_cb13", "codes_cb14", "codes_cb15" }; - + struct ggml_tensor * cb_codes_tensors[16]; for (int cb = 0; cb < 16; ++cb) { cb_codes_tensors[cb] = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_frames); ggml_set_name(cb_codes_tensors[cb], cb_names[cb]); ggml_set_input(cb_codes_tensors[cb]); } - + struct ggml_tensor * first_codes = cb_codes_tensors[0]; - + struct ggml_tensor * first_emb = ggml_get_rows(ctx0, model_.vq_first_codebook, first_codes); ggml_set_name(first_emb, "first_emb_raw"); - + struct ggml_tensor * rest_emb[15]; for (int cb = 0; cb < 15; ++cb) { struct ggml_tensor * cb_codes = cb_codes_tensors[cb + 1]; rest_emb[cb] = ggml_get_rows(ctx0, model_.vq_rest_codebook[cb], cb_codes); - + if (cb == 0) { ggml_set_name(rest_emb[cb], "rest_cb0_emb_raw"); } } - + struct ggml_tensor * first_emb_2d = ggml_reshape_2d(ctx0, first_emb, cfg.codebook_dim, n_frames); ggml_set_name(first_emb_2d, "first_emb_2d"); - - struct ggml_tensor * first_proj_weight_2d = ggml_reshape_2d(ctx0, model_.vq_first_output_proj, + + struct ggml_tensor * first_proj_weight_2d = ggml_reshape_2d(ctx0, model_.vq_first_output_proj, cfg.codebook_dim, cfg.hidden_dim); struct ggml_tensor * first_proj_2d = ggml_mul_mat(ctx0, first_proj_weight_2d, first_emb_2d); ggml_set_name(first_proj_2d, "first_proj_2d"); - + struct ggml_tensor * rest_proj_weight_2d = ggml_reshape_2d(ctx0, model_.vq_rest_output_proj, cfg.codebook_dim, cfg.hidden_dim); - + struct ggml_tensor * rest_proj_2d = nullptr; for (int cb = 0; cb < 15; ++cb) { struct ggml_tensor * cb_emb_2d = ggml_reshape_2d(ctx0, rest_emb[cb], cfg.codebook_dim, n_frames); - + if (cb == 0) { ggml_set_name(cb_emb_2d, "rest_cb0_emb_2d"); } - + struct ggml_tensor * cb_proj_2d = ggml_mul_mat(ctx0, rest_proj_weight_2d, cb_emb_2d); - + if (rest_proj_2d == nullptr) { rest_proj_2d = cb_proj_2d; } else { @@ -688,82 +688,82 @@ struct ggml_cgraph * AudioTokenizerDecoder::build_graph(int32_t n_frames) { } } ggml_set_name(rest_proj_2d, "rest_proj_2d"); - + struct ggml_tensor * latent_2d = ggml_add(ctx0, first_proj_2d, rest_proj_2d); ggml_set_name(latent_2d, "latent_2d"); - + struct ggml_tensor * latent_t = ggml_transpose(ctx0, latent_2d); ggml_set_name(latent_t, "latent_t"); - + struct ggml_tensor * latent_cont = ggml_cont(ctx0, latent_t); ggml_set_name(latent_cont, "latent_cont"); - + struct ggml_tensor * latent = ggml_reshape_3d(ctx0, latent_cont, n_frames, cfg.hidden_dim, 1); ggml_set_name(latent, "vq_output"); - + struct ggml_tensor * latent_for_conv = ggml_cont(ctx0, latent); struct ggml_tensor * latent_padded = ggml_pad_ext(ctx0, latent_for_conv, 2, 0, 0, 0, 0, 0, 0, 0); struct ggml_tensor * cur = ggml_conv_1d(ctx0, model_.pre_conv_w, latent_padded, 1, 0, 1); if (model_.pre_conv_b) { cur = ggml_add(ctx0, cur, ggml_reshape_3d(ctx0, model_.pre_conv_b, 1, cfg.latent_dim, 1)); } - + ggml_set_name(cur, "pre_conv_output"); - + struct ggml_tensor * cur_2d = ggml_reshape_2d(ctx0, cur, n_frames, cfg.latent_dim); struct ggml_tensor * cur_t = ggml_transpose(ctx0, cur_2d); cur = ggml_cont(ctx0, cur_t); - + ggml_set_name(cur, "pre_conv_reshaped"); - + cur = ggml_mul_mat(ctx0, model_.pre_tfm_input_proj_w, cur); if (model_.pre_tfm_input_proj_b) { cur = ggml_add(ctx0, cur, model_.pre_tfm_input_proj_b); } - + ggml_set_name(cur, "pre_tfm_input"); - + struct ggml_tensor * positions = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_frames); ggml_set_name(positions, "positions"); ggml_set_input(positions); - + for (int i = 0; i < cfg.n_pre_tfm_layers; ++i) { cur = apply_pre_tfm_layer(ctx0, cur, model_.pre_tfm_layers[i], n_frames, positions); } - + if (model_.pre_tfm_norm_w) { cur = apply_rms_norm(ctx0, cur, model_.pre_tfm_norm_w, cfg.rms_norm_eps); } - + cur = ggml_mul_mat(ctx0, model_.pre_tfm_output_proj_w, cur); if (model_.pre_tfm_output_proj_b) { cur = ggml_add(ctx0, cur, model_.pre_tfm_output_proj_b); } - + ggml_set_name(cur, "pre_tfm_output"); - + cur = ggml_permute(ctx0, cur, 1, 0, 2, 3); cur = ggml_cont(ctx0, cur); cur = ggml_reshape_3d(ctx0, cur, n_frames, cfg.latent_dim, 1); - + ggml_set_name(cur, "pre_tfm_reshaped"); - + for (int i = 0; i < 2; ++i) { cur = apply_upsample_block(ctx0, cur, model_.upsample[i], i); } - + ggml_set_name(cur, "upsample_output"); - + // Causal padding: left pad with 6 (kernel_size - 1 = 7 - 1 = 6) cur = ggml_pad_ext(ctx0, cur, 6, 0, 0, 0, 0, 0, 0, 0); cur = ggml_conv_1d(ctx0, model_.dec0_conv_w, cur, 1, 0, 1); if (model_.dec0_conv_b) { cur = ggml_add(ctx0, cur, ggml_reshape_3d(ctx0, model_.dec0_conv_b, 1, cfg.decoder_dim, 1)); } - + ggml_set_name(cur, "dec0_output"); - + int upsample_rates[4] = {8, 5, 4, 3}; for (int i = 0; i < 4; ++i) { cur = apply_decoder_block(ctx0, cur, model_.dec_blocks[i], upsample_rates[i], i); @@ -771,33 +771,33 @@ struct ggml_cgraph * AudioTokenizerDecoder::build_graph(int32_t n_frames) { snprintf(name, sizeof(name), "dec%d_output", i + 1); ggml_set_name(cur, name); } - + if (model_.dec5_snake_alpha) { cur = apply_snake(ctx0, cur, model_.dec5_snake_alpha, model_.dec5_snake_beta); } - + ggml_set_name(cur, "dec5_output"); - + // Causal padding: left pad with 6 (kernel_size - 1 = 7 - 1 = 6) cur = ggml_pad_ext(ctx0, cur, 6, 0, 0, 0, 0, 0, 0, 0); cur = ggml_conv_1d(ctx0, model_.dec6_conv_w, cur, 1, 0, 1); if (model_.dec6_conv_b) { cur = ggml_add(ctx0, cur, ggml_reshape_3d(ctx0, model_.dec6_conv_b, 1, 1, 1)); } - + ggml_set_name(cur, "dec6_output"); - + cur = ggml_tanh(ctx0, cur); - + cur = ggml_reshape_1d(ctx0, cur, cur->ne[0]); - + ggml_set_name(cur, "audio"); ggml_set_output(cur); - + ggml_build_forward_expand(gf, cur); - + ggml_free(ctx0); - + return gf; } @@ -807,23 +807,23 @@ bool AudioTokenizerDecoder::decode(const int32_t * codes, int32_t n_frames, error_msg_ = "Model not loaded"; return false; } - + const auto & cfg = model_.config; - + codes_buf_.resize(n_frames * cfg.n_codebooks); for (int f = 0; f < n_frames; ++f) { for (int cb = 0; cb < cfg.n_codebooks; ++cb) { codes_buf_[cb + f * cfg.n_codebooks] = codes[f * cfg.n_codebooks + cb]; } } - + struct ggml_cgraph * gf = build_graph(n_frames); - + if (!ggml_backend_sched_alloc_graph(state_.sched, gf)) { error_msg_ = "Failed to allocate graph"; return false; } - + std::vector cb_codes(n_frames); for (int cb = 0; cb < 16; ++cb) { char name[32]; @@ -834,47 +834,47 @@ bool AudioTokenizerDecoder::decode(const int32_t * codes, int32_t n_frames, ggml_backend_sched_reset(state_.sched); return false; } - + for (int f = 0; f < n_frames; ++f) { cb_codes[f] = codes_buf_[f * cfg.n_codebooks + cb]; } - + ggml_backend_tensor_set(cb_tensor, cb_codes.data(), 0, n_frames * sizeof(int32_t)); } - - + + struct ggml_tensor * positions_tensor = ggml_graph_get_tensor(gf, "positions"); if (positions_tensor) { std::vector positions(n_frames); for (int i = 0; i < n_frames; ++i) { positions[i] = i; } - ggml_backend_tensor_set(positions_tensor, positions.data(), 0, + ggml_backend_tensor_set(positions_tensor, positions.data(), 0, n_frames * sizeof(int32_t)); } - - + + if (ggml_backend_sched_graph_compute(state_.sched, gf) != GGML_STATUS_SUCCESS) { error_msg_ = "Failed to compute graph"; ggml_backend_sched_reset(state_.sched); return false; } - + struct ggml_tensor * audio_tensor = ggml_graph_get_tensor(gf, "audio"); if (!audio_tensor) { error_msg_ = "Failed to find audio tensor"; ggml_backend_sched_reset(state_.sched); return false; } - + int64_t n_samples = audio_tensor->ne[0]; samples.resize(n_samples); ggml_backend_tensor_get(audio_tensor, samples.data(), 0, n_samples * sizeof(float)); - + ggml_backend_sched_reset(state_.sched); - + return true; } diff --git a/otherarch/qwen3tts/audio_tokenizer_encoder.cpp b/otherarch/qwen3tts/audio_tokenizer_encoder.cpp index d2646737c..31b4eb31e 100644 --- a/otherarch/qwen3tts/audio_tokenizer_encoder.cpp +++ b/otherarch/qwen3tts/audio_tokenizer_encoder.cpp @@ -13,7 +13,7 @@ namespace qwen3_tts { // Mel filterbank computation using librosa slaney normalization // This matches librosa.filters.mel with norm='slaney' -static void compute_mel_filterbank_slaney(float * filterbank, int n_mels, int n_fft, +static void compute_mel_filterbank_slaney(float * filterbank, int n_mels, int n_fft, int sample_rate, float f_min, float f_max) { // Slaney-style mel scale (used by librosa default) auto hz_to_mel_slaney = [](float hz) -> float { @@ -22,64 +22,64 @@ static void compute_mel_filterbank_slaney(float * filterbank, int n_mels, int n_ const float min_log_hz = 1000.0f; const float min_log_mel = (min_log_hz - 0.0f) / f_sp; // 15 const float logstep = logf(6.4f) / 27.0f; // log(6400/1000) / 27 - + if (hz < min_log_hz) { return (hz - 0.0f) / f_sp; } else { return min_log_mel + logf(hz / min_log_hz) / logstep; } }; - + auto mel_to_hz_slaney = [](float mel) -> float { const float f_sp = 200.0f / 3.0f; const float min_log_hz = 1000.0f; const float min_log_mel = (min_log_hz - 0.0f) / f_sp; const float logstep = logf(6.4f) / 27.0f; - + if (mel < min_log_mel) { return 0.0f + f_sp * mel; } else { return min_log_hz * expf(logstep * (mel - min_log_mel)); } }; - + float mel_min = hz_to_mel_slaney(f_min); float mel_max = hz_to_mel_slaney(f_max); - + int n_fft_bins = n_fft / 2 + 1; - + // Compute mel center frequencies std::vector mel_points(n_mels + 2); for (int i = 0; i < n_mels + 2; ++i) { mel_points[i] = mel_min + (mel_max - mel_min) * i / (n_mels + 1); } - + // Convert to Hz and then to FFT bin indices std::vector hz_points(n_mels + 2); std::vector fft_freqs(n_fft_bins); - + for (int i = 0; i < n_mels + 2; ++i) { hz_points[i] = mel_to_hz_slaney(mel_points[i]); } - + for (int i = 0; i < n_fft_bins; ++i) { fft_freqs[i] = (float)i * sample_rate / n_fft; } - + memset(filterbank, 0, n_mels * n_fft_bins * sizeof(float)); - + // Create triangular filters with slaney normalization for (int m = 0; m < n_mels; ++m) { float f_left = hz_points[m]; float f_center = hz_points[m + 1]; float f_right = hz_points[m + 2]; - + // Slaney normalization: divide by bandwidth (area normalization) float enorm = 2.0f / (f_right - f_left); - + for (int k = 0; k < n_fft_bins; ++k) { float freq = fft_freqs[k]; - + if (freq >= f_left && freq <= f_center) { if (f_center > f_left) { filterbank[m * n_fft_bins + k] = enorm * (freq - f_left) / (f_center - f_left); @@ -116,7 +116,7 @@ static void compute_hann_window(float * window, int n) { static void compute_centered_window(float * window, int n_fft, int win_length) { // Zero-initialize memset(window, 0, n_fft * sizeof(float)); - + // Compute Hann window of win_length int offset = (n_fft - win_length) / 2; for (int i = 0; i < win_length; ++i) { @@ -128,7 +128,7 @@ AudioTokenizerEncoder::AudioTokenizerEncoder() = default; AudioTokenizerEncoder::~AudioTokenizerEncoder() { free_speaker_encoder_model(model_); - + if (state_.sched) { ggml_backend_sched_free(state_.sched); state_.sched = nullptr; @@ -149,10 +149,10 @@ bool AudioTokenizerEncoder::load_model(const std::string & model_path) { error_msg_ = loader.get_error(); return false; } - + model_.config.sample_rate = loader.get_u32("qwen3-tts.speaker_encoder.sample_rate", 24000); model_.config.embedding_dim = loader.get_u32("qwen3-tts.speaker_encoder.embedding_length", 1024); - + int64_t n_tensors = loader.get_n_tensors(); int spk_tensor_count = 0; for (int64_t i = 0; i < n_tensors; ++i) { @@ -161,46 +161,46 @@ bool AudioTokenizerEncoder::load_model(const std::string & model_path) { spk_tensor_count++; } } - + if (spk_tensor_count == 0) { error_msg_ = "No speaker encoder tensors found in model"; return false; } - + size_t ctx_size = ggml_tensor_overhead() * spk_tensor_count; struct ggml_init_params params = { /*.mem_size =*/ ctx_size, /*.mem_buffer =*/ nullptr, /*.no_alloc =*/ true, }; - + model_.ctx = ggml_init(params); if (!model_.ctx) { error_msg_ = "Failed to initialize GGML context"; return false; } - + struct gguf_context * gguf_ctx = loader.get_ctx(); struct ggml_context * meta_ctx = loader.get_meta_ctx(); - + for (int64_t i = 0; i < n_tensors; ++i) { const char * name = loader.get_tensor_name(i); if (!name || strncmp(name, "spk_enc.", 8) != 0) { continue; } - + struct ggml_tensor * meta_tensor = ggml_get_tensor(meta_ctx, name); if (!meta_tensor) { continue; } - + struct ggml_tensor * tensor = ggml_dup_tensor(model_.ctx, meta_tensor); ggml_set_name(tensor, name); - + model_.tensors[name] = tensor; - + std::string sname(name); - + if (sname == "spk_enc.conv0.weight") model_.conv0_w = tensor; else if (sname == "spk_enc.conv0.bias") model_.conv0_b = tensor; else if (sname == "spk_enc.mfa.weight") model_.mfa_w = tensor; @@ -214,7 +214,7 @@ bool AudioTokenizerEncoder::load_model(const std::string & model_path) { else { int blk_idx, res_idx; char suffix[64]; - + if (sscanf(name, "spk_enc.blk.%d.tdnn1.%s", &blk_idx, suffix) == 2) { if (blk_idx >= 1 && blk_idx <= 3) { if (strcmp(suffix, "weight") == 0) model_.blocks[blk_idx-1].tdnn1_w = tensor; @@ -247,20 +247,20 @@ bool AudioTokenizerEncoder::load_model(const std::string & model_path) { } } } - - if (!load_tensor_data_from_file(model_path, gguf_ctx, model_.ctx, - model_.tensors, model_.buffer, error_msg_)) { + + if (!load_tensor_data_from_file(model_path, gguf_ctx, model_.ctx, + model_.tensors, model_.buffer, error_msg_, GGML_BACKEND_DEVICE_TYPE_GPU)) { return false; } - - state_.backend = init_preferred_backend("AudioTokenizerEncoder", &error_msg_); + + state_.backend = init_preferred_backend("AudioTokenizerEncoder", &error_msg_, true); if (!state_.backend) { return false; } ggml_backend_dev_t device = ggml_backend_get_device(state_.backend); const char * device_name = device ? ggml_backend_dev_name(device) : "Unknown"; fprintf(stderr, " AudioTokenizerEncoder backend: %s\n", device_name); - + if (device && ggml_backend_dev_type(device) != GGML_BACKEND_DEVICE_TYPE_CPU) { state_.backend_cpu = ggml_backend_init_by_type(GGML_BACKEND_DEVICE_TYPE_CPU, nullptr); if (!state_.backend_cpu) { @@ -279,20 +279,20 @@ bool AudioTokenizerEncoder::load_model(const std::string & model_path) { error_msg_ = "Failed to create backend scheduler"; return false; } - + state_.compute_meta.resize(ggml_tensor_overhead() * QWEN3_TTS_MAX_NODES + ggml_graph_overhead()); - + return true; } bool AudioTokenizerEncoder::compute_mel_spectrogram(const float * samples, int32_t n_samples, std::vector & mel, int32_t & n_frames) { const auto & cfg = model_.config; - + // Match PyTorch STFT padding: (n_fft - hop_size) // 2 on each side with reflect int padding = (cfg.n_fft - cfg.hop_length) / 2; int padded_length = n_samples + 2 * padding; - + // Create padded signal with reflect padding std::vector padded(padded_length); for (int i = 0; i < padded_length; ++i) { @@ -310,50 +310,50 @@ bool AudioTokenizerEncoder::compute_mel_spectrogram(const float * samples, int32 src_idx = std::max(0, std::min(n_samples - 1, src_idx)); padded[i] = samples[src_idx]; } - + // With center=False, frames start at 0 and step by hop_length n_frames = (padded_length - cfg.n_fft) / cfg.hop_length + 1; if (n_frames <= 0) { error_msg_ = "Audio too short for mel spectrogram"; return false; } - + int n_fft_bins = cfg.n_fft / 2 + 1; - + std::vector filterbank(cfg.n_mels * n_fft_bins); - compute_mel_filterbank_slaney(filterbank.data(), cfg.n_mels, cfg.n_fft, + compute_mel_filterbank_slaney(filterbank.data(), cfg.n_mels, cfg.n_fft, cfg.sample_rate, cfg.f_min, cfg.f_max); - + // PyTorch STFT with win_length < n_fft centers the window in the n_fft frame // This is critical for matching PyTorch's output std::vector window(cfg.n_fft); compute_centered_window(window.data(), cfg.n_fft, cfg.win_length); - + // Output: [batch, n_mels, n_frames] but we store as [n_mels, n_frames] row-major // which means mel[m * n_frames + f] = value at mel bin m, frame f mel.resize(cfg.n_mels * n_frames); - + std::vector frame(cfg.n_fft, 0.0f); std::vector fft_real(cfg.n_fft); std::vector fft_imag(cfg.n_fft); std::vector magnitude(n_fft_bins); - + for (int32_t f = 0; f < n_frames; ++f) { int start = f * cfg.hop_length; - + // Apply centered window to n_fft samples for (int i = 0; i < cfg.n_fft; ++i) { frame[i] = padded[start + i] * window[i]; } - + compute_dft(frame.data(), fft_real.data(), fft_imag.data(), cfg.n_fft); - + // Compute magnitude (not power) - matches torch.stft with return_complex=True then abs() // spec = torch.sqrt(torch.view_as_real(spec).pow(2).sum(-1) + 1e-9) for (int k = 0; k < n_fft_bins; ++k) { magnitude[k] = sqrtf(fft_real[k] * fft_real[k] + fft_imag[k] * fft_imag[k] + 1e-9f); } - + // Apply mel filterbank and log compression // mel_spec = torch.matmul(mel_basis, spec) // mel_spec = dynamic_range_compression_torch(mel_spec) # log(clamp(x, min=1e-5) * 1) @@ -366,7 +366,7 @@ bool AudioTokenizerEncoder::compute_mel_spectrogram(const float * samples, int32 mel[m * n_frames + f] = logf(std::max(sum, 1e-5f)); } } - + return true; } @@ -376,41 +376,41 @@ static struct ggml_tensor * apply_reflect_pad_1d(struct ggml_context * ctx, if (pad == 0) { return x; } - + int64_t T = x->ne[0]; int64_t C = x->ne[1]; int64_t B = x->ne[2]; - + struct ggml_tensor * left_slices[16]; struct ggml_tensor * right_slices[16]; - + for (int i = 0; i < pad && i < 16; ++i) { int left_src_idx = pad - i; left_slices[i] = ggml_view_3d(ctx, x, 1, C, B, x->nb[1], x->nb[2], left_src_idx * x->nb[0]); left_slices[i] = ggml_cont(ctx, left_slices[i]); - + int right_src_idx = T - 2 - i; right_slices[i] = ggml_view_3d(ctx, x, 1, C, B, x->nb[1], x->nb[2], right_src_idx * x->nb[0]); right_slices[i] = ggml_cont(ctx, right_slices[i]); } - + struct ggml_tensor * left_pad = left_slices[0]; for (int i = 1; i < pad && i < 16; ++i) { left_pad = ggml_concat(ctx, left_pad, left_slices[i], 0); } - + struct ggml_tensor * right_pad = right_slices[0]; for (int i = 1; i < pad && i < 16; ++i) { right_pad = ggml_concat(ctx, right_pad, right_slices[i], 0); } - + struct ggml_tensor * padded = ggml_concat(ctx, left_pad, x, 0); padded = ggml_concat(ctx, padded, right_pad, 0); - + return padded; } @@ -423,12 +423,12 @@ static struct ggml_tensor * apply_conv1d(struct ggml_context * ctx, bool use_reflect_pad = true) { struct ggml_tensor * input = x; int actual_pad = pad; - + if (use_reflect_pad && pad > 0) { input = apply_reflect_pad_1d(ctx, x, pad); actual_pad = 0; } - + struct ggml_tensor * y = ggml_conv_1d(ctx, w, input, stride, actual_pad, dilation); if (debug_name) { char name[64]; @@ -447,32 +447,32 @@ struct ggml_cgraph * AudioTokenizerEncoder::build_graph(int32_t n_frames) { const int hidden_dim = cfg.hidden_dim; // 512 const int scale = cfg.res2net_scale; // 8 const int branch_dim = hidden_dim / scale; // 64 - + struct ggml_init_params params = { /*.mem_size =*/ state_.compute_meta.size(), /*.mem_buffer =*/ state_.compute_meta.data(), /*.no_alloc =*/ true, }; - + struct ggml_context * ctx0 = ggml_init(params); struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, QWEN3_TTS_MAX_NODES, false); - + // Input: mel spectrogram [n_mels, n_frames] - stored as [n_mels, n_frames] row-major // GGML uses column-major, so this is [n_frames, n_mels] in GGML notation struct ggml_tensor * mel = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_frames, cfg.n_mels); ggml_set_name(mel, "mel"); ggml_set_input(mel); - + // PyTorch: hidden_states = hidden_states.transpose(1, 2) # [B, T, C] -> [B, C, T] // Our mel is [n_frames, n_mels] in GGML = [n_mels, n_frames] row-major // For conv1d, we need [T, C, B] in GGML = [B, C, T] row-major // So reshape to [n_frames, n_mels, 1] struct ggml_tensor * cur = ggml_reshape_3d(ctx0, mel, n_frames, cfg.n_mels, 1); ggml_set_name(cur, "mel_3d"); - + struct ggml_tensor * mel_padded = apply_reflect_pad_1d(ctx0, cur, 2); ggml_set_name(mel_padded, "mel_padded"); - + cur = ggml_conv_1d(ctx0, model_.conv0_w, mel_padded, 1, 0, 1); ggml_set_name(cur, "conv0_conv"); @@ -483,37 +483,37 @@ struct ggml_cgraph * AudioTokenizerEncoder::build_graph(int32_t n_frames) { ggml_set_name(cur, "conv0_pre_relu"); cur = ggml_relu(ctx0, cur); ggml_set_name(cur, "conv0_out"); - + int64_t seq_len = cur->ne[0]; - + // Store block outputs for MFA (including block 0) struct ggml_tensor * block_outputs[4]; block_outputs[0] = cur; // Initial TDNN output - + // Blocks 1-3: SE-Res2Net blocks // Dilations: block1=2, block2=3, block3=4 int dilations[3] = {2, 3, 4}; - + for (int blk = 0; blk < 3; ++blk) { const auto & block = model_.blocks[blk]; int dilation = dilations[blk]; - + struct ggml_tensor * residual = cur; - + cur = apply_conv1d(ctx0, block.tdnn1_w, block.tdnn1_b, cur, 1, 0, 1); cur = ggml_relu(ctx0, cur); if (blk == 0) { ggml_set_name(cur, "blk1_tdnn1"); } - + // Res2Net: Split into 8 branches of 64 channels each // cur shape: [seq_len, 512, 1] // Branch 0: identity (no conv) // Branch i (1-7): conv(hidden_part + previous_output) for i >= 2, conv(hidden_part) for i == 1 - + // Split channels: view as [seq_len, 64, 8] then split struct ggml_tensor * branches[8]; - + // Extract each branch using view operations // cur is [seq_len, 512, 1], we want to split dim 1 into 8 parts of 64 for (int b = 0; b < scale; ++b) { @@ -522,17 +522,17 @@ struct ggml_cgraph * AudioTokenizerEncoder::build_graph(int32_t n_frames) { // nb1 = stride for dim 1 = cur->nb[1] (bytes to move from one channel to next) // nb2 = stride for dim 2 = cur->nb[2] (bytes to move from one batch to next) // offset = b * 64 * cur->nb[1] (skip b*64 channels) - branches[b] = ggml_view_3d(ctx0, cur, + branches[b] = ggml_view_3d(ctx0, cur, seq_len, branch_dim, 1, - cur->nb[1], cur->nb[2], + cur->nb[1], cur->nb[2], b * branch_dim * cur->nb[1]); branches[b] = ggml_cont(ctx0, branches[b]); } - + // Process branches according to Res2Net logic struct ggml_tensor * outputs[8]; outputs[0] = branches[0]; // Branch 0: identity - + for (int b = 1; b < scale; ++b) { struct ggml_tensor * input; if (b == 1) { @@ -541,18 +541,18 @@ struct ggml_cgraph * AudioTokenizerEncoder::build_graph(int32_t n_frames) { // Add previous output to current branch input = ggml_add(ctx0, branches[b], outputs[b - 1]); } - + // Apply conv with dilation (kernel=3) // Padding for kernel=3, dilation=d: pad = d * (3-1) / 2 = d if (block.res2net_w[b - 1]) { - outputs[b] = apply_conv1d(ctx0, block.res2net_w[b - 1], block.res2net_b[b - 1], + outputs[b] = apply_conv1d(ctx0, block.res2net_w[b - 1], block.res2net_b[b - 1], input, 1, dilation, dilation); outputs[b] = ggml_relu(ctx0, outputs[b]); } else { outputs[b] = input; // Fallback if weight missing } } - + cur = outputs[0]; for (int b = 1; b < scale; ++b) { cur = ggml_concat(ctx0, cur, outputs[b], 1); @@ -565,40 +565,40 @@ struct ggml_cgraph * AudioTokenizerEncoder::build_graph(int32_t n_frames) { ggml_set_name(outputs[b], name); } } - + cur = apply_conv1d(ctx0, block.tdnn2_w, block.tdnn2_b, cur, 1, 0, 1); cur = ggml_relu(ctx0, cur); if (blk == 0) { ggml_set_name(cur, "blk1_tdnn2"); } - + // SE (Squeeze-Excitation) // Global average pooling over time: mean(dim=2, keepdim=True) struct ggml_tensor * se = ggml_pool_1d(ctx0, cur, GGML_OP_POOL_AVG, seq_len, seq_len, 0); se = ggml_reshape_3d(ctx0, se, 1, hidden_dim, 1); - + // SE conv1: 512 -> 128 with ReLU se = apply_conv1d(ctx0, block.se_conv1_w, block.se_conv1_b, se, 1, 0, 1); se = ggml_relu(ctx0, se); - + // SE conv2: 128 -> 512 with Sigmoid se = apply_conv1d(ctx0, block.se_conv2_w, block.se_conv2_b, se, 1, 0, 1); se = ggml_sigmoid(ctx0, se); - + cur = ggml_mul(ctx0, cur, se); if (blk == 0) { ggml_set_name(cur, "blk1_se"); } cur = ggml_add(ctx0, cur, residual); - + char block_name[32]; snprintf(block_name, sizeof(block_name), "block_%d", blk + 1); ggml_set_name(cur, block_name); - + block_outputs[blk + 1] = cur; } - + // MFA: Concatenate block outputs [1:] (blocks 1, 2, 3 = indices 1, 2, 3) // hidden_states = torch.cat(hidden_states_list[1:], dim=1) // Each block output is [seq_len, 512, 1] @@ -606,20 +606,20 @@ struct ggml_cgraph * AudioTokenizerEncoder::build_graph(int32_t n_frames) { struct ggml_tensor * mfa_input = ggml_concat(ctx0, block_outputs[1], block_outputs[2], 1); mfa_input = ggml_concat(ctx0, mfa_input, block_outputs[3], 1); ggml_set_name(mfa_input, "mfa_input"); - + // MFA conv: 1536 -> 1536 with ReLU cur = apply_conv1d(ctx0, model_.mfa_w, model_.mfa_b, mfa_input, 1, 0, 1); cur = ggml_relu(ctx0, cur); ggml_set_name(cur, "mfa_out"); - + // ASP (Attentive Statistics Pooling) // cur shape: [seq_len, 1536, 1] - + // Step 1: Compute global mean and std over time // mean = hidden_states.mean(dim=2, keepdim=True) # [1, 1536, 1] struct ggml_tensor * global_mean = ggml_pool_1d(ctx0, cur, GGML_OP_POOL_AVG, seq_len, seq_len, 0); global_mean = ggml_reshape_3d(ctx0, global_mean, 1, 1536, 1); - + // std = sqrt(E[x^2] - E[x]^2) struct ggml_tensor * sq = ggml_sqr(ctx0, cur); struct ggml_tensor * mean_sq = ggml_pool_1d(ctx0, sq, GGML_OP_POOL_AVG, seq_len, seq_len, 0); @@ -627,20 +627,20 @@ struct ggml_cgraph * AudioTokenizerEncoder::build_graph(int32_t n_frames) { struct ggml_tensor * var = ggml_sub(ctx0, mean_sq, ggml_sqr(ctx0, global_mean)); var = ggml_clamp(ctx0, var, 1e-12f, 1e10f); struct ggml_tensor * global_std = ggml_sqrt(ctx0, var); - + // Step 2: Expand mean and std to full sequence length and concatenate with hidden_states // mean = mean.repeat(1, 1, seq_length) # [1, 1536, seq_len] // std = std.repeat(1, 1, seq_length) # [1, 1536, seq_len] // attention = torch.cat([hidden_states, mean, std], dim=1) # [1, 4608, seq_len] - struct ggml_tensor * mean_expanded = ggml_repeat(ctx0, global_mean, + struct ggml_tensor * mean_expanded = ggml_repeat(ctx0, global_mean, ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, seq_len, 1536, 1)); struct ggml_tensor * std_expanded = ggml_repeat(ctx0, global_std, ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, seq_len, 1536, 1)); - + struct ggml_tensor * attention = ggml_concat(ctx0, cur, mean_expanded, 1); attention = ggml_concat(ctx0, attention, std_expanded, 1); // attention shape: [seq_len, 4608, 1] - + // Step 3: TDNN (4608 -> 128) with ReLU, then Tanh // self.tdnn = TimeDelayNetBlock(channels * 3, attention_channels, 1, 1) # has ReLU // attention = self.conv(self.tanh(self.tdnn(attention))) @@ -648,17 +648,17 @@ struct ggml_cgraph * AudioTokenizerEncoder::build_graph(int32_t n_frames) { attention = ggml_relu(ctx0, attention); // TDNN has ReLU ggml_set_name(attention, "asp_tdnn"); attention = ggml_tanh(ctx0, attention); // Then tanh is applied - + // Step 4: Conv (128 -> 1536) for attention weights // self.conv = nn.Conv1d(attention_channels, channels, kernel_size=1) attention = apply_conv1d(ctx0, model_.asp_conv_w, model_.asp_conv_b, attention, 1, 0, 1); ggml_set_name(attention, "asp_conv"); // attention shape: [seq_len, 1536, 1] - + // Step 5: Softmax over time dimension attention = ggml_soft_max(ctx0, attention); ggml_set_name(attention, "asp_softmax"); - + // Step 6: Compute weighted mean and std // mean, std = self._compute_statistics(hidden_states, attention) // mean = (attention * hidden_states).sum(dim=2) @@ -666,7 +666,7 @@ struct ggml_cgraph * AudioTokenizerEncoder::build_graph(int32_t n_frames) { struct ggml_tensor * weighted_mean = ggml_pool_1d(ctx0, weighted, GGML_OP_POOL_AVG, seq_len, seq_len, 0); weighted_mean = ggml_scale(ctx0, weighted_mean, (float)seq_len); // Convert avg to sum weighted_mean = ggml_reshape_3d(ctx0, weighted_mean, 1, 1536, 1); - + // std = sqrt((attention * (hidden_states - mean)^2).sum(dim=2).clamp(eps)) struct ggml_tensor * mean_for_std = ggml_repeat(ctx0, weighted_mean, ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, seq_len, 1536, 1)); @@ -678,25 +678,25 @@ struct ggml_cgraph * AudioTokenizerEncoder::build_graph(int32_t n_frames) { var_sum = ggml_reshape_3d(ctx0, var_sum, 1, 1536, 1); var_sum = ggml_clamp(ctx0, var_sum, 1e-12f, 1e10f); struct ggml_tensor * weighted_std = ggml_sqrt(ctx0, var_sum); - + // Step 7: Concatenate mean and std: [1, 3072, 1] struct ggml_tensor * pooled = ggml_concat(ctx0, weighted_mean, weighted_std, 1); ggml_set_name(pooled, "asp_pooled"); - + // FC: 3072 -> 1024 cur = apply_conv1d(ctx0, model_.fc_w, model_.fc_b, pooled, 1, 0, 1); ggml_set_name(cur, "fc_out"); - + // Squeeze to 1D cur = ggml_reshape_1d(ctx0, cur, cfg.embedding_dim); - + ggml_set_name(cur, "embedding"); ggml_set_output(cur); - + ggml_build_forward_expand(gf, cur); - + ggml_free(ctx0); - + return gf; } @@ -706,27 +706,27 @@ bool AudioTokenizerEncoder::encode(const float * samples, int32_t n_samples, error_msg_ = "Model not loaded"; return false; } - + std::vector mel; int32_t n_frames; if (!compute_mel_spectrogram(samples, n_samples, mel, n_frames)) { return false; } - + struct ggml_cgraph * gf = build_graph(n_frames); if (!ggml_backend_sched_alloc_graph(state_.sched, gf)) { error_msg_ = "Failed to allocate graph"; return false; } - + struct ggml_tensor * mel_tensor = ggml_graph_get_tensor(gf, "mel"); if (!mel_tensor) { error_msg_ = "Failed to find mel tensor"; ggml_backend_sched_reset(state_.sched); return false; } - + // mel is stored as [n_mels, n_frames] row-major: mel[m * n_frames + f] = mel bin m at frame f // GGML tensor is [n_frames, n_mels] column-major: element (f, m) at memory[f + m * n_frames] // For GGML conv1d, we want input(t, c) = mel bin c at time t @@ -734,7 +734,7 @@ bool AudioTokenizerEncoder::encode(const float * samples, int32_t n_samples, // Since the memory layout matches (both are contiguous in frame order for each mel bin), // we can copy directly! ggml_backend_tensor_set(mel_tensor, mel.data(), 0, mel.size() * sizeof(float)); - + if (ggml_backend_sched_graph_compute(state_.sched, gf) != GGML_STATUS_SUCCESS) { error_msg_ = "Failed to compute graph"; ggml_backend_sched_reset(state_.sched); @@ -747,12 +747,12 @@ bool AudioTokenizerEncoder::encode(const float * samples, int32_t n_samples, ggml_backend_sched_reset(state_.sched); return false; } - + embedding.resize(model_.config.embedding_dim); ggml_backend_tensor_get(emb_tensor, embedding.data(), 0, embedding.size() * sizeof(float)); - + ggml_backend_sched_reset(state_.sched); - + return true; } diff --git a/otherarch/qwen3tts/gguf_loader.cpp b/otherarch/qwen3tts/gguf_loader.cpp index 649af520d..09b7a154f 100644 --- a/otherarch/qwen3tts/gguf_loader.cpp +++ b/otherarch/qwen3tts/gguf_loader.cpp @@ -24,7 +24,7 @@ GGUFLoader::~GGUFLoader() { close(); } -ggml_backend_t init_preferred_backend(const char * component_name, std::string * error_msg) { +ggml_backend_t init_preferred_backend(const char * component_name, std::string * error_msg, bool allow_gpu) { if (error_msg) error_msg->clear(); auto & shared = get_shared_backend_state(); @@ -34,9 +34,12 @@ ggml_backend_t init_preferred_backend(const char * component_name, std::string * } ggml_backend_t backend = ggml_backend_init_by_type(GGML_BACKEND_DEVICE_TYPE_IGPU, nullptr); + if(allow_gpu) + { if (!backend) { backend = ggml_backend_init_by_type(GGML_BACKEND_DEVICE_TYPE_GPU, nullptr); } + } if (!backend) { backend = ggml_backend_init_by_type(GGML_BACKEND_DEVICE_TYPE_ACCEL, nullptr); } @@ -78,20 +81,20 @@ void release_preferred_backend(ggml_backend_t backend) { bool GGUFLoader::open(const std::string & path) { close(); // Close any previously opened file - + file_path_ = path; - + struct gguf_init_params params = { /*.no_alloc =*/ true, /*.ctx =*/ &meta_ctx_, }; - + ctx_ = gguf_init_from_file(path.c_str(), params); if (!ctx_) { error_msg_ = "Failed to open GGUF file: " + path; return false; } - + return true; } @@ -168,7 +171,7 @@ bool load_tensor_data_from_file( error_msg = "Failed to initialize backend for GGUF tensor loader"; return false; } - + // Allocate buffer for all tensors buffer = ggml_backend_alloc_ctx_tensors(model_ctx, backend); if (!buffer) { @@ -176,7 +179,7 @@ bool load_tensor_data_from_file( ggml_backend_free(backend); return false; } - + // Open file for reading tensor data FILE * f = fopen(path.c_str(), "rb"); if (!f) { @@ -184,45 +187,45 @@ bool load_tensor_data_from_file( ggml_backend_free(backend); return false; } - + const size_t data_offset = gguf_get_data_offset(ctx); const int64_t n_tensors = gguf_get_n_tensors(ctx); std::vector read_buf; - + for (int64_t i = 0; i < n_tensors; ++i) { const char * name = gguf_get_tensor_name(ctx, i); size_t offset = gguf_get_tensor_offset(ctx, i); - + auto it = tensors.find(name); if (it == tensors.end()) { continue; // Skip tensors not in our map } - + struct ggml_tensor * tensor = it->second; size_t nbytes = ggml_nbytes(tensor); - + read_buf.resize(nbytes); - + if (fseek(f, data_offset + offset, SEEK_SET) != 0) { error_msg = "Failed to seek to tensor data: " + std::string(name); fclose(f); ggml_backend_free(backend); return false; } - + if (fread(read_buf.data(), 1, nbytes, f) != nbytes) { error_msg = "Failed to read tensor data: " + std::string(name); fclose(f); ggml_backend_free(backend); return false; } - + ggml_backend_tensor_set(tensor, read_buf.data(), 0, nbytes); } - + fclose(f); ggml_backend_free(backend); - + return true; } diff --git a/otherarch/qwen3tts/gguf_loader.h b/otherarch/qwen3tts/gguf_loader.h index d14d84f9c..fd949e2a7 100644 --- a/otherarch/qwen3tts/gguf_loader.h +++ b/otherarch/qwen3tts/gguf_loader.h @@ -17,44 +17,44 @@ class GGUFLoader { public: GGUFLoader(); ~GGUFLoader(); - + // Open GGUF file and parse metadata bool open(const std::string & path); - + // Close file and free resources void close(); - + // Get error message if operation failed const std::string & get_error() const { return error_msg_; } - + // Get number of tensors in file int64_t get_n_tensors() const; - + // Get tensor name by index const char * get_tensor_name(int64_t idx) const; - + // Get tensor type by index enum ggml_type get_tensor_type(int64_t idx) const; - + // Get tensor offset by index size_t get_tensor_offset(int64_t idx) const; - + // Get tensor size by index size_t get_tensor_size(int64_t idx) const; - + // Get metadata value (returns -1 if not found) int32_t get_u32(const char * key, int32_t default_val = 0) const; float get_f32(const char * key, float default_val = 0.0f) const; - + // Get data offset (start of tensor data in file) size_t get_data_offset() const; - + // Get GGUF context (for advanced usage) struct gguf_context * get_ctx() const { return ctx_; } - + // Get metadata context struct ggml_context * get_meta_ctx() const { return meta_ctx_; } - + protected: struct gguf_context * ctx_ = nullptr; struct ggml_context * meta_ctx_ = nullptr; @@ -70,11 +70,11 @@ bool load_tensor_data_from_file( const std::map & tensors, ggml_backend_buffer_t & buffer, std::string & error_msg, - enum ggml_backend_dev_type preferred_backend_type = GGML_BACKEND_DEVICE_TYPE_CPU + enum ggml_backend_dev_type preferred_backend_type = GGML_BACKEND_DEVICE_TYPE_GPU ); // Helper to initialize backend with GPU preference and CPU fallback -ggml_backend_t init_preferred_backend(const char * component_name, std::string * error_msg); +ggml_backend_t init_preferred_backend(const char * component_name, std::string * error_msg, bool allow_gpu); void release_preferred_backend(ggml_backend_t backend); // Helper function to free model resources diff --git a/otherarch/qwen3tts/tts_transformer.cpp b/otherarch/qwen3tts/tts_transformer.cpp index ffd947cf6..2bb0eb98b 100644 --- a/otherarch/qwen3tts/tts_transformer.cpp +++ b/otherarch/qwen3tts/tts_transformer.cpp @@ -97,7 +97,7 @@ bool TTSTransformer::load_model(const std::string & model_path) { gguf_free(ctx); if (meta_ctx) ggml_free(meta_ctx); - state_.backend = init_preferred_backend("TTSTransformer", &error_msg_); + state_.backend = init_preferred_backend("TTSTransformer", &error_msg_, true); if (!state_.backend) { return false; } @@ -553,7 +553,7 @@ bool TTSTransformer::create_tensors(struct gguf_context * ctx) { } bool TTSTransformer::load_tensor_data(const std::string & path, struct gguf_context * ctx) { - ggml_backend_t backend = init_preferred_backend("TTSTransformer", &error_msg_); + ggml_backend_t backend = init_preferred_backend("TTSTransformer", &error_msg_, true); if (!backend) { return false; }