| { | |
| "model": { | |
| "bos_token_id": 128000, | |
| "context_length": 4096, | |
| "decoder": { | |
| "session_options": { | |
| "log_id": "onnxruntime-genai", | |
| "provider_options": [] | |
| }, | |
| "head_size": 128, | |
| "hidden_size": 3072, | |
| "inputs": { | |
| "input_ids": "input_ids", | |
| "attention_mask": "attention_mask_dummy", | |
| "position_ids": "position_ids_dummy", | |
| "past_key_names": "past_key_%d_in", | |
| "past_value_names": "past_value_%d_in" | |
| }, | |
| "outputs": { | |
| "logits": "logits_dequantized", | |
| "present_key_names": "past_key_%d_out", | |
| "present_value_names": "past_value_%d_out" | |
| }, | |
| "num_attention_heads": 24, | |
| "num_hidden_layers": 28, | |
| "num_key_value_heads": 8, | |
| "sliding_window_key_value_cache": { | |
| "window_size": 128, | |
| "pad_value": 128 | |
| }, | |
| "pipeline": [ | |
| { | |
| "position-processor": { | |
| "filename": "position-processor.onnx", | |
| "inputs": [ | |
| "attention_mask_before_processor", | |
| "position_ids" | |
| ], | |
| "outputs": [ | |
| "attention_mask_before_quantizer", | |
| "position_ids_cos_before_quantizer", | |
| "position_ids_sin_before_quantizer" | |
| ], | |
| "session_options": { | |
| "log_id": "onnxruntime-genai.position_processor", | |
| "provider_options": [ | |
| {} | |
| ] | |
| }, | |
| "run_on_token_gen": false | |
| }, | |
| "position-shifter": { | |
| "filename": "position-shifter.onnx", | |
| "inputs": [ | |
| "attention_mask_before_processor", | |
| "position_ids" | |
| ], | |
| "outputs": [ | |
| "attention_mask_shifted", | |
| "position_ids_shifted", | |
| "attention_mask_before_quantizer", | |
| "position_ids_cos_before_quantizer", | |
| "position_ids_sin_before_quantizer" | |
| ], | |
| "output_names_forwarder": { | |
| "attention_mask_shifted": "attention_mask_before_processor", | |
| "position_ids_shifted": "position_ids" | |
| }, | |
| "session_options": { | |
| "log_id": "onnxruntime-genai.position_shifter", | |
| "provider_options": [ | |
| {} | |
| ] | |
| }, | |
| "run_on_prompt": false | |
| }, | |
| "quantizer": { | |
| "filename": "quantizer.onnx", | |
| "inputs": [ | |
| "attention_mask_before_quantizer", | |
| "position_ids_cos_before_quantizer", | |
| "position_ids_sin_before_quantizer" | |
| ], | |
| "outputs": [ | |
| "attention_mask", | |
| "position_ids_cos", | |
| "position_ids_sin" | |
| ], | |
| "session_options": { | |
| "log_id": "onnxruntime-genai.quantizer", | |
| "provider_options": [ | |
| {} | |
| ] | |
| } | |
| }, | |
| "prompt-processor-1": { | |
| "filename": "prompt_1_of_3_qnn_ctx.onnx", | |
| "inputs": [ | |
| "input_ids" | |
| ], | |
| "outputs": [ | |
| "_model_model_embed_tokens_Gather_output_0" | |
| ], | |
| "session_options": { | |
| "log_id": "onnxruntime-genai.pp1", | |
| "provider_options": [ | |
| { | |
| "qnn": { | |
| "backend_path": "libQnnHtp.so", | |
| "htp_performance_mode": "burst", | |
| "enable_htp_shared_memory_allocator": "1", | |
| "qnn_context_priority": "high" | |
| } | |
| } | |
| ] | |
| }, | |
| "run_on_token_gen": false | |
| }, | |
| "prompt-processor-2": { | |
| "filename": "prompt_2_of_3_qnn_ctx.onnx", | |
| "inputs": [ | |
| "_model_model_embed_tokens_Gather_output_0", | |
| "attention_mask", | |
| "position_ids_cos", | |
| "position_ids_sin", | |
| "past_key_0_in", | |
| "past_value_0_in", | |
| "past_key_1_in", | |
| "past_value_1_in", | |
| "past_key_2_in", | |
| "past_value_2_in", | |
| "past_key_3_in", | |
| "past_value_3_in", | |
| "past_key_4_in", | |
| "past_value_4_in", | |
| "past_key_5_in", | |
| "past_value_5_in", | |
| "past_key_6_in", | |
| "past_value_6_in", | |
| "past_key_7_in", | |
| "past_value_7_in", | |
| "past_key_8_in", | |
| "past_value_8_in", | |
| "past_key_9_in", | |
| "past_value_9_in", | |
| "past_key_10_in", | |
| "past_value_10_in", | |
| "past_key_11_in", | |
| "past_value_11_in", | |
| "past_key_12_in", | |
| "past_value_12_in", | |
| "past_key_13_in", | |
| "past_value_13_in" | |
| ], | |
| "outputs": [ | |
| "_model_model_layers_13_Add_1_output_0", | |
| "past_key_0_out", | |
| "past_value_0_out", | |
| "past_key_1_out", | |
| "past_value_1_out", | |
| "past_key_2_out", | |
| "past_value_2_out", | |
| "past_key_3_out", | |
| "past_value_3_out", | |
| "past_key_4_out", | |
| "past_value_4_out", | |
| "past_key_5_out", | |
| "past_value_5_out", | |
| "past_key_6_out", | |
| "past_value_6_out", | |
| "past_key_7_out", | |
| "past_value_7_out", | |
| "past_key_8_out", | |
| "past_value_8_out", | |
| "past_key_9_out", | |
| "past_value_9_out", | |
| "past_key_10_out", | |
| "past_value_10_out", | |
| "past_key_11_out", | |
| "past_value_11_out", | |
| "past_key_12_out", | |
| "past_value_12_out", | |
| "past_key_13_out", | |
| "past_value_13_out" | |
| ], | |
| "session_options": { | |
| "log_id": "onnxruntime-genai.pp2", | |
| "provider_options": [ | |
| { | |
| "qnn": { | |
| "backend_path": "libQnnHtp.so", | |
| "htp_performance_mode": "burst", | |
| "enable_htp_shared_memory_allocator": "1", | |
| "qnn_context_priority": "high" | |
| } | |
| } | |
| ] | |
| }, | |
| "run_on_token_gen": false | |
| }, | |
| "prompt-processor-3": { | |
| "filename": "prompt_3_of_3_qnn_ctx.onnx", | |
| "inputs": [ | |
| "_model_model_layers_13_Add_1_output_0", | |
| "attention_mask", | |
| "position_ids_cos", | |
| "position_ids_sin", | |
| "past_key_14_in", | |
| "past_value_14_in", | |
| "past_key_15_in", | |
| "past_value_15_in", | |
| "past_key_16_in", | |
| "past_value_16_in", | |
| "past_key_17_in", | |
| "past_value_17_in", | |
| "past_key_18_in", | |
| "past_value_18_in", | |
| "past_key_19_in", | |
| "past_value_19_in", | |
| "past_key_20_in", | |
| "past_value_20_in", | |
| "past_key_21_in", | |
| "past_value_21_in", | |
| "past_key_22_in", | |
| "past_value_22_in", | |
| "past_key_23_in", | |
| "past_value_23_in", | |
| "past_key_24_in", | |
| "past_value_24_in", | |
| "past_key_25_in", | |
| "past_value_25_in", | |
| "past_key_26_in", | |
| "past_value_26_in", | |
| "past_key_27_in", | |
| "past_value_27_in" | |
| ], | |
| "outputs": [ | |
| "logits", | |
| "past_key_14_out", | |
| "past_value_14_out", | |
| "past_key_15_out", | |
| "past_value_15_out", | |
| "past_key_16_out", | |
| "past_value_16_out", | |
| "past_key_17_out", | |
| "past_value_17_out", | |
| "past_key_18_out", | |
| "past_value_18_out", | |
| "past_key_19_out", | |
| "past_value_19_out", | |
| "past_key_20_out", | |
| "past_value_20_out", | |
| "past_key_21_out", | |
| "past_value_21_out", | |
| "past_key_22_out", | |
| "past_value_22_out", | |
| "past_key_23_out", | |
| "past_value_23_out", | |
| "past_key_24_out", | |
| "past_value_24_out", | |
| "past_key_25_out", | |
| "past_value_25_out", | |
| "past_key_26_out", | |
| "past_value_26_out", | |
| "past_key_27_out", | |
| "past_value_27_out" | |
| ], | |
| "session_options": { | |
| "log_id": "onnxruntime-genai.pp3", | |
| "provider_options": [ | |
| { | |
| "qnn": { | |
| "backend_path": "libQnnHtp.so", | |
| "htp_performance_mode": "burst", | |
| "enable_htp_shared_memory_allocator": "1", | |
| "qnn_context_priority": "high" | |
| } | |
| } | |
| ] | |
| }, | |
| "run_on_token_gen": false | |
| }, | |
| "token-generator-1": { | |
| "filename": "token_1_of_3_qnn_ctx.onnx", | |
| "inputs": [ | |
| "input_ids" | |
| ], | |
| "outputs": [ | |
| "_model_model_embed_tokens_Gather_output_0" | |
| ], | |
| "session_options": { | |
| "log_id": "onnxruntime-genai.tg1", | |
| "provider_options": [ | |
| { | |
| "qnn": { | |
| "backend_path": "libQnnHtp.so", | |
| "htp_performance_mode": "burst", | |
| "enable_htp_shared_memory_allocator": "1", | |
| "qnn_context_priority": "high" | |
| } | |
| } | |
| ] | |
| }, | |
| "run_on_prompt": false | |
| }, | |
| "token-generator-2": { | |
| "filename": "token_2_of_3_qnn_ctx.onnx", | |
| "inputs": [ | |
| "_model_model_embed_tokens_Gather_output_0", | |
| "attention_mask", | |
| "position_ids_cos", | |
| "position_ids_sin", | |
| "past_key_0_in", | |
| "past_value_0_in", | |
| "past_key_1_in", | |
| "past_value_1_in", | |
| "past_key_2_in", | |
| "past_value_2_in", | |
| "past_key_3_in", | |
| "past_value_3_in", | |
| "past_key_4_in", | |
| "past_value_4_in", | |
| "past_key_5_in", | |
| "past_value_5_in", | |
| "past_key_6_in", | |
| "past_value_6_in", | |
| "past_key_7_in", | |
| "past_value_7_in", | |
| "past_key_8_in", | |
| "past_value_8_in", | |
| "past_key_9_in", | |
| "past_value_9_in", | |
| "past_key_10_in", | |
| "past_value_10_in", | |
| "past_key_11_in", | |
| "past_value_11_in", | |
| "past_key_12_in", | |
| "past_value_12_in", | |
| "past_key_13_in", | |
| "past_value_13_in" | |
| ], | |
| "outputs": [ | |
| "_model_model_layers_13_Add_1_output_0", | |
| "past_key_0_out", | |
| "past_value_0_out", | |
| "past_key_1_out", | |
| "past_value_1_out", | |
| "past_key_2_out", | |
| "past_value_2_out", | |
| "past_key_3_out", | |
| "past_value_3_out", | |
| "past_key_4_out", | |
| "past_value_4_out", | |
| "past_key_5_out", | |
| "past_value_5_out", | |
| "past_key_6_out", | |
| "past_value_6_out", | |
| "past_key_7_out", | |
| "past_value_7_out", | |
| "past_key_8_out", | |
| "past_value_8_out", | |
| "past_key_9_out", | |
| "past_value_9_out", | |
| "past_key_10_out", | |
| "past_value_10_out", | |
| "past_key_11_out", | |
| "past_value_11_out", | |
| "past_key_12_out", | |
| "past_value_12_out", | |
| "past_key_13_out", | |
| "past_value_13_out" | |
| ], | |
| "session_options": { | |
| "log_id": "onnxruntime-genai.tg2", | |
| "provider_options": [ | |
| { | |
| "qnn": { | |
| "backend_path": "libQnnHtp.so", | |
| "htp_performance_mode": "burst", | |
| "enable_htp_shared_memory_allocator": "1", | |
| "qnn_context_priority": "high" | |
| } | |
| } | |
| ] | |
| }, | |
| "run_on_prompt": false | |
| }, | |
| "token-generator-3": { | |
| "filename": "token_3_of_3_qnn_ctx.onnx", | |
| "inputs": [ | |
| "_model_model_layers_13_Add_1_output_0", | |
| "attention_mask", | |
| "position_ids_cos", | |
| "position_ids_sin", | |
| "past_key_14_in", | |
| "past_value_14_in", | |
| "past_key_15_in", | |
| "past_value_15_in", | |
| "past_key_16_in", | |
| "past_value_16_in", | |
| "past_key_17_in", | |
| "past_value_17_in", | |
| "past_key_18_in", | |
| "past_value_18_in", | |
| "past_key_19_in", | |
| "past_value_19_in", | |
| "past_key_20_in", | |
| "past_value_20_in", | |
| "past_key_21_in", | |
| "past_value_21_in", | |
| "past_key_22_in", | |
| "past_value_22_in", | |
| "past_key_23_in", | |
| "past_value_23_in", | |
| "past_key_24_in", | |
| "past_value_24_in", | |
| "past_key_25_in", | |
| "past_value_25_in", | |
| "past_key_26_in", | |
| "past_value_26_in", | |
| "past_key_27_in", | |
| "past_value_27_in" | |
| ], | |
| "outputs": [ | |
| "logits", | |
| "past_key_14_out", | |
| "past_value_14_out", | |
| "past_key_15_out", | |
| "past_value_15_out", | |
| "past_key_16_out", | |
| "past_value_16_out", | |
| "past_key_17_out", | |
| "past_value_17_out", | |
| "past_key_18_out", | |
| "past_value_18_out", | |
| "past_key_19_out", | |
| "past_value_19_out", | |
| "past_key_20_out", | |
| "past_value_20_out", | |
| "past_key_21_out", | |
| "past_value_21_out", | |
| "past_key_22_out", | |
| "past_value_22_out", | |
| "past_key_23_out", | |
| "past_value_23_out", | |
| "past_key_24_out", | |
| "past_value_24_out", | |
| "past_key_25_out", | |
| "past_value_25_out", | |
| "past_key_26_out", | |
| "past_value_26_out", | |
| "past_key_27_out", | |
| "past_value_27_out" | |
| ], | |
| "session_options": { | |
| "log_id": "onnxruntime-genai.tg3", | |
| "provider_options": [ | |
| { | |
| "qnn": { | |
| "backend_path": "libQnnHtp.so", | |
| "htp_performance_mode": "burst", | |
| "enable_htp_shared_memory_allocator": "1", | |
| "qnn_context_priority": "high" | |
| } | |
| } | |
| ] | |
| }, | |
| "run_on_prompt": false | |
| }, | |
| "dequantizer": { | |
| "filename": "dequantizer.onnx", | |
| "inputs": [ | |
| "logits" | |
| ], | |
| "outputs": [ | |
| "logits_dequantized" | |
| ], | |
| "session_options": { | |
| "log_id": "onnxruntime-genai.dequantizer", | |
| "provider_options": [ | |
| {} | |
| ] | |
| } | |
| } | |
| } | |
| ] | |
| }, | |
| "eos_token_id": [ | |
| 128001, | |
| 128008, | |
| 128009 | |
| ], | |
| "pad_token_id": 128001, | |
| "type": "decoder-pipeline", | |
| "vocab_size": 128256 | |
| }, | |
| "search": { | |
| "diversity_penalty": 0.0, | |
| "do_sample": true, | |
| "early_stopping": true, | |
| "length_penalty": 1.0, | |
| "max_length": 131072, | |
| "min_length": 0, | |
| "no_repeat_ngram_size": 0, | |
| "num_beams": 1, | |
| "num_return_sequences": 1, | |
| "past_present_share_buffer": true, | |
| "repetition_penalty": 1.0, | |
| "temperature": 0.6, | |
| "top_k": 1, | |
| "top_p": 0.9 | |
| } | |
| } | |