Skip to content

Commit a655269

Browse files
ggerganovggerganov
committed
cont
ggml-ci
1 parent 91a3530 commit a655269

File tree

2 files changed

+131
-140
lines changed

2 files changed

+131
-140
lines changed

‎examples/tts/convert_pt_to_hf.py

Lines changed: 0 additions & 4 deletions
Original file line numberDiff line numberDiff line change
@@ -172,10 +172,6 @@ def flatten_state_dict(state_dict, parent_key='', sep='.'):
172172
"n_embd": 768,
173173
"n_layer": 12
174174
},
175-
#"n_embd_posnet": 768,
176-
#"n_embd_convnext": 768,
177-
#"n_layer_posnet": 6,
178-
#"n_layer_convnext": 12
179175
}
180176

181177
with open(path_dst + '/config.json', 'w') as f:

‎src/llama.cpp

Lines changed: 131 additions & 136 deletions
Original file line numberDiff line numberDiff line change
@@ -2747,205 +2747,199 @@ struct llama_cparams {
27472747

27482748
struct llama_layer_posnet {
27492749
// resnet
2750-
struct ggml_tensor * norm1 = nullptr;
2750+
struct ggml_tensor * norm1 = nullptr;
27512751
struct ggml_tensor * norm1_b = nullptr;
27522752

2753-
struct ggml_tensor * conv1 = nullptr;
2753+
struct ggml_tensor * conv1 = nullptr;
27542754
struct ggml_tensor * conv1_b = nullptr;
27552755

2756-
struct ggml_tensor * norm2 = nullptr;
2756+
struct ggml_tensor * norm2 = nullptr;
27572757
struct ggml_tensor * norm2_b = nullptr;
27582758

2759-
struct ggml_tensor * conv2 = nullptr;
2759+
struct ggml_tensor * conv2 = nullptr;
27602760
struct ggml_tensor * conv2_b = nullptr;
27612761

27622762
// attention
2763-
struct ggml_tensor * attn_norm = nullptr;
2763+
struct ggml_tensor * attn_norm = nullptr;
27642764
struct ggml_tensor * attn_norm_b = nullptr;
27652765

2766-
struct ggml_tensor * attn_q = nullptr;
2766+
struct ggml_tensor * attn_q = nullptr;
27672767
struct ggml_tensor * attn_q_b = nullptr;
27682768

2769-
struct ggml_tensor * attn_k = nullptr;
2769+
struct ggml_tensor * attn_k = nullptr;
27702770
struct ggml_tensor * attn_k_b = nullptr;
27712771

2772-
struct ggml_tensor * attn_v = nullptr;
2772+
struct ggml_tensor * attn_v = nullptr;
27732773
struct ggml_tensor * attn_v_b = nullptr;
27742774

2775-
struct ggml_tensor * attn_o = nullptr;
2775+
struct ggml_tensor * attn_o = nullptr;
27762776
struct ggml_tensor * attn_o_b = nullptr;
27772777

27782778
// normalize
2779-
struct ggml_tensor * norm = nullptr;
2779+
struct ggml_tensor * norm = nullptr;
27802780
struct ggml_tensor * norm_b = nullptr;
27812781
};
27822782

27832783
struct llama_layer_convnext {
2784-
struct ggml_tensor * dw;
2785-
struct ggml_tensor * dw_b;
2784+
struct ggml_tensor * dw = nullptr;
2785+
struct ggml_tensor * dw_b = nullptr;
27862786

2787-
struct ggml_tensor * norm;
2788-
struct ggml_tensor * norm_b;
2787+
struct ggml_tensor * norm = nullptr;
2788+
struct ggml_tensor * norm_b = nullptr;
27892789

2790-
struct ggml_tensor * pw1;
2791-
struct ggml_tensor * pw1_b;
2790+
struct ggml_tensor * pw1 = nullptr;
2791+
struct ggml_tensor * pw1_b = nullptr;
27922792

2793-
struct ggml_tensor * pw2;
2794-
struct ggml_tensor * pw2_b;
2793+
struct ggml_tensor * pw2 = nullptr;
2794+
struct ggml_tensor * pw2_b = nullptr;
27952795

2796-
struct ggml_tensor * gamma;
2796+
struct ggml_tensor * gamma = nullptr;
27972797
};
27982798

2799-
// TODO: separate into "llama_layer_enc" and "llama_layer_dec"
28002799
struct llama_layer {
2801-
llama_layer() {
2802-
// initialize all pointers to NULL
2803-
std::memset(this, 0, sizeof(*this));
2804-
}
2805-
28062800
// normalization
2807-
struct ggml_tensor * attn_norm;
2808-
struct ggml_tensor * attn_norm_b;
2809-
struct ggml_tensor * attn_norm_2;
2810-
struct ggml_tensor * attn_norm_2_b;
2811-
struct ggml_tensor * attn_q_norm;
2812-
struct ggml_tensor * attn_q_norm_b;
2813-
struct ggml_tensor * attn_k_norm;
2814-
struct ggml_tensor * attn_k_norm_b;
2815-
struct ggml_tensor * attn_out_norm;
2816-
struct ggml_tensor * attn_out_norm_b;
2817-
struct ggml_tensor * attn_q_a_norm;
2818-
struct ggml_tensor * attn_kv_a_norm;
2819-
struct ggml_tensor * attn_sub_norm;
2820-
struct ggml_tensor * attn_post_norm;
2821-
struct ggml_tensor * ffn_sub_norm;
2822-
struct ggml_tensor * attn_norm_cross;
2823-
struct ggml_tensor * attn_norm_enc;
2801+
struct ggml_tensor * attn_norm = nullptr;
2802+
struct ggml_tensor * attn_norm_b = nullptr;
2803+
struct ggml_tensor * attn_norm_2 = nullptr;
2804+
struct ggml_tensor * attn_norm_2_b = nullptr;
2805+
struct ggml_tensor * attn_q_norm = nullptr;
2806+
struct ggml_tensor * attn_q_norm_b = nullptr;
2807+
struct ggml_tensor * attn_k_norm = nullptr;
2808+
struct ggml_tensor * attn_k_norm_b = nullptr;
2809+
struct ggml_tensor * attn_out_norm = nullptr;
2810+
struct ggml_tensor * attn_out_norm_b = nullptr;
2811+
struct ggml_tensor * attn_q_a_norm = nullptr;
2812+
struct ggml_tensor * attn_kv_a_norm = nullptr;
2813+
struct ggml_tensor * attn_sub_norm = nullptr;
2814+
struct ggml_tensor * attn_post_norm = nullptr;
2815+
struct ggml_tensor * ffn_sub_norm = nullptr;
2816+
struct ggml_tensor * attn_norm_cross = nullptr;
2817+
struct ggml_tensor * attn_norm_enc = nullptr;
28242818

28252819
// attention
2826-
struct ggml_tensor * wq;
2827-
struct ggml_tensor * wk;
2828-
struct ggml_tensor * wv;
2829-
struct ggml_tensor * wo;
2830-
struct ggml_tensor * wqkv;
2831-
struct ggml_tensor * wq_a;
2832-
struct ggml_tensor * wq_b;
2833-
struct ggml_tensor * wkv_a_mqa;
2834-
struct ggml_tensor * wkv_b;
2835-
struct ggml_tensor * wq_cross;
2836-
struct ggml_tensor * wk_cross;
2837-
struct ggml_tensor * wv_cross;
2838-
struct ggml_tensor * wo_cross;
2839-
struct ggml_tensor * wq_enc;
2840-
struct ggml_tensor * wk_enc;
2841-
struct ggml_tensor * wv_enc;
2842-
struct ggml_tensor * wo_enc;
2820+
struct ggml_tensor * wq = nullptr;
2821+
struct ggml_tensor * wk = nullptr;
2822+
struct ggml_tensor * wv = nullptr;
2823+
struct ggml_tensor * wo = nullptr;
2824+
struct ggml_tensor * wqkv = nullptr;
2825+
struct ggml_tensor * wq_a = nullptr;
2826+
struct ggml_tensor * wq_b = nullptr;
2827+
struct ggml_tensor * wkv_a_mqa = nullptr;
2828+
struct ggml_tensor * wkv_b = nullptr;
2829+
struct ggml_tensor * wq_cross = nullptr;
2830+
struct ggml_tensor * wk_cross = nullptr;
2831+
struct ggml_tensor * wv_cross = nullptr;
2832+
struct ggml_tensor * wo_cross = nullptr;
2833+
struct ggml_tensor * wq_enc = nullptr;
2834+
struct ggml_tensor * wk_enc = nullptr;
2835+
struct ggml_tensor * wv_enc = nullptr;
2836+
struct ggml_tensor * wo_enc = nullptr;
28432837

28442838
// attention bias
2845-
struct ggml_tensor * bq;
2846-
struct ggml_tensor * bk;
2847-
struct ggml_tensor * bv;
2848-
struct ggml_tensor * bo;
2849-
struct ggml_tensor * bqkv;
2839+
struct ggml_tensor * bq = nullptr;
2840+
struct ggml_tensor * bk = nullptr;
2841+
struct ggml_tensor * bv = nullptr;
2842+
struct ggml_tensor * bo = nullptr;
2843+
struct ggml_tensor * bqkv = nullptr;
28502844

28512845
// relative position bias
2852-
struct ggml_tensor * attn_rel_b;
2853-
struct ggml_tensor * attn_rel_b_enc;
2854-
struct ggml_tensor * attn_rel_b_cross;
2846+
struct ggml_tensor * attn_rel_b = nullptr;
2847+
struct ggml_tensor * attn_rel_b_enc = nullptr;
2848+
struct ggml_tensor * attn_rel_b_cross = nullptr;
28552849

28562850
// normalization
2857-
struct ggml_tensor * ffn_norm;
2858-
struct ggml_tensor * ffn_norm_b;
2859-
struct ggml_tensor * ffn_post_norm;
2860-
struct ggml_tensor * layer_out_norm;
2861-
struct ggml_tensor * layer_out_norm_b;
2862-
struct ggml_tensor * ffn_norm_exps;
2863-
struct ggml_tensor * ffn_norm_enc;
2851+
struct ggml_tensor * ffn_norm = nullptr;
2852+
struct ggml_tensor * ffn_norm_b = nullptr;
2853+
struct ggml_tensor * ffn_post_norm = nullptr;
2854+
struct ggml_tensor * layer_out_norm = nullptr;
2855+
struct ggml_tensor * layer_out_norm_b = nullptr;
2856+
struct ggml_tensor * ffn_norm_exps = nullptr;
2857+
struct ggml_tensor * ffn_norm_enc = nullptr;
28642858

28652859
// ff
2866-
struct ggml_tensor * ffn_gate; // w1
2867-
struct ggml_tensor * ffn_down; // w2
2868-
struct ggml_tensor * ffn_up; // w3
2869-
struct ggml_tensor * ffn_gate_enc;
2870-
struct ggml_tensor * ffn_down_enc;
2871-
struct ggml_tensor * ffn_up_enc;
2860+
struct ggml_tensor * ffn_gate = nullptr; // w1
2861+
struct ggml_tensor * ffn_down = nullptr; // w2
2862+
struct ggml_tensor * ffn_up = nullptr; // w3
2863+
struct ggml_tensor * ffn_gate_enc = nullptr;
2864+
struct ggml_tensor * ffn_down_enc = nullptr;
2865+
struct ggml_tensor * ffn_up_enc = nullptr;
28722866

28732867
// ff MoE
2874-
struct ggml_tensor * ffn_gate_inp;
2875-
struct ggml_tensor * ffn_gate_exps;
2876-
struct ggml_tensor * ffn_down_exps;
2877-
struct ggml_tensor * ffn_up_exps ;
2868+
struct ggml_tensor * ffn_gate_inp = nullptr;
2869+
struct ggml_tensor * ffn_gate_exps = nullptr;
2870+
struct ggml_tensor * ffn_down_exps = nullptr;
2871+
struct ggml_tensor * ffn_up_exps = nullptr;
28782872

28792873
// ff shared expert (shexp)
2880-
struct ggml_tensor * ffn_gate_inp_shexp;
2881-
struct ggml_tensor * ffn_gate_shexp;
2882-
struct ggml_tensor * ffn_down_shexp;
2883-
struct ggml_tensor * ffn_up_shexp;
2874+
struct ggml_tensor * ffn_gate_inp_shexp = nullptr;
2875+
struct ggml_tensor * ffn_gate_shexp = nullptr;
2876+
struct ggml_tensor * ffn_down_shexp = nullptr;
2877+
struct ggml_tensor * ffn_up_shexp = nullptr;
28842878

28852879
// ff bias
2886-
struct ggml_tensor * ffn_gate_b;
2887-
struct ggml_tensor * ffn_down_b; // b2
2888-
struct ggml_tensor * ffn_up_b; // b3
2889-
struct ggml_tensor * ffn_act;
2880+
struct ggml_tensor * ffn_gate_b = nullptr;
2881+
struct ggml_tensor * ffn_down_b = nullptr; // b2
2882+
struct ggml_tensor * ffn_up_b = nullptr; // b3
2883+
struct ggml_tensor * ffn_act = nullptr;
28902884

28912885
// mamba proj
2892-
struct ggml_tensor * ssm_in;
2893-
struct ggml_tensor * ssm_x;
2894-
struct ggml_tensor * ssm_dt;
2895-
struct ggml_tensor * ssm_out;
2886+
struct ggml_tensor * ssm_in = nullptr;
2887+
struct ggml_tensor * ssm_x = nullptr;
2888+
struct ggml_tensor * ssm_dt = nullptr;
2889+
struct ggml_tensor * ssm_out = nullptr;
28962890

28972891
// mamba
2898-
struct ggml_tensor * ssm_conv1d;
2899-
struct ggml_tensor * ssm_a;
2900-
struct ggml_tensor * ssm_d;
2892+
struct ggml_tensor * ssm_conv1d = nullptr;
2893+
struct ggml_tensor * ssm_a = nullptr;
2894+
struct ggml_tensor * ssm_d = nullptr;
29012895

29022896
// mamba bias
2903-
struct ggml_tensor * ssm_conv1d_b;
2904-
struct ggml_tensor * ssm_dt_b;
2897+
struct ggml_tensor * ssm_conv1d_b = nullptr;
2898+
struct ggml_tensor * ssm_dt_b = nullptr;
29052899

29062900
// rwkv
2907-
struct ggml_tensor * time_mix_w1;
2908-
struct ggml_tensor * time_mix_w2;
2909-
struct ggml_tensor * time_mix_lerp_x;
2910-
struct ggml_tensor * time_mix_lerp_w;
2911-
struct ggml_tensor * time_mix_lerp_k;
2912-
struct ggml_tensor * time_mix_lerp_v;
2913-
struct ggml_tensor * time_mix_lerp_r;
2914-
struct ggml_tensor * time_mix_lerp_g;
2915-
2916-
struct ggml_tensor * time_mix_first;
2917-
struct ggml_tensor * time_mix_decay;
2918-
struct ggml_tensor * time_mix_decay_w1;
2919-
struct ggml_tensor * time_mix_decay_w2;
2920-
struct ggml_tensor * time_mix_key;
2921-
struct ggml_tensor * time_mix_value;
2922-
struct ggml_tensor * time_mix_receptance;
2923-
struct ggml_tensor * time_mix_gate;
2924-
2925-
struct ggml_tensor * time_mix_ln;
2926-
struct ggml_tensor * time_mix_ln_b;
2927-
struct ggml_tensor * time_mix_output;
2928-
2929-
struct ggml_tensor * channel_mix_lerp_k;
2930-
struct ggml_tensor * channel_mix_lerp_r;
2931-
2932-
struct ggml_tensor * channel_mix_key;
2933-
struct ggml_tensor * channel_mix_receptance;
2934-
struct ggml_tensor * channel_mix_value;
2901+
struct ggml_tensor * time_mix_w1 = nullptr;
2902+
struct ggml_tensor * time_mix_w2 = nullptr;
2903+
struct ggml_tensor * time_mix_lerp_x = nullptr;
2904+
struct ggml_tensor * time_mix_lerp_w = nullptr;
2905+
struct ggml_tensor * time_mix_lerp_k = nullptr;
2906+
struct ggml_tensor * time_mix_lerp_v = nullptr;
2907+
struct ggml_tensor * time_mix_lerp_r = nullptr;
2908+
struct ggml_tensor * time_mix_lerp_g = nullptr;
2909+
2910+
struct ggml_tensor * time_mix_first = nullptr;
2911+
struct ggml_tensor * time_mix_decay = nullptr;
2912+
struct ggml_tensor * time_mix_decay_w1 = nullptr;
2913+
struct ggml_tensor * time_mix_decay_w2 = nullptr;
2914+
struct ggml_tensor * time_mix_key = nullptr;
2915+
struct ggml_tensor * time_mix_value = nullptr;
2916+
struct ggml_tensor * time_mix_receptance = nullptr;
2917+
struct ggml_tensor * time_mix_gate = nullptr;
2918+
2919+
struct ggml_tensor * time_mix_ln = nullptr;
2920+
struct ggml_tensor * time_mix_ln_b = nullptr;
2921+
struct ggml_tensor * time_mix_output = nullptr;
2922+
2923+
struct ggml_tensor * channel_mix_lerp_k = nullptr;
2924+
struct ggml_tensor * channel_mix_lerp_r = nullptr;
2925+
2926+
struct ggml_tensor * channel_mix_key = nullptr;
2927+
struct ggml_tensor * channel_mix_receptance = nullptr;
2928+
struct ggml_tensor * channel_mix_value = nullptr;
29352929

29362930
// long rope factors
29372931
struct ggml_tensor * rope_long = nullptr;
29382932
struct ggml_tensor * rope_short = nullptr;
29392933
struct ggml_tensor * rope_freqs = nullptr;
29402934

29412935
// bitnet scale
2942-
struct ggml_tensor * wq_scale;
2943-
struct ggml_tensor * wk_scale;
2944-
struct ggml_tensor * wv_scale;
2945-
struct ggml_tensor * wo_scale;
2946-
struct ggml_tensor * ffn_gate_scale;
2947-
struct ggml_tensor * ffn_up_scale;
2948-
struct ggml_tensor * ffn_down_scale;
2936+
struct ggml_tensor * wq_scale = nullptr;
2937+
struct ggml_tensor * wk_scale = nullptr;
2938+
struct ggml_tensor * wv_scale = nullptr;
2939+
struct ggml_tensor * wo_scale = nullptr;
2940+
struct ggml_tensor * ffn_gate_scale = nullptr;
2941+
struct ggml_tensor * ffn_up_scale = nullptr;
2942+
struct ggml_tensor * ffn_down_scale = nullptr;
29492943

29502944
struct llama_layer_posnet posnet;
29512945

@@ -3167,6 +3161,7 @@ struct llama_sbatch {
31673161
// batch indices of the output
31683162
std::vector<size_t> out_ids;
31693163
std::vector<llama_sbatch_seq> seq;
3164+
31703165
const llama_batch * batch = nullptr;
31713166

31723167
// buffers for the ubatch

0 commit comments

Comments
 (0)