Upload folder using huggingface_hub
Browse files- .gitattributes +3 -0
- Qwen3-Embedding-0.6B/added_tokens.json +28 -0
- Qwen3-Embedding-0.6B/chat_template.jinja +85 -0
- Qwen3-Embedding-0.6B/config.json +60 -0
- Qwen3-Embedding-0.6B/merges.txt +0 -0
- Qwen3-Embedding-0.6B/model.safetensors +3 -0
- Qwen3-Embedding-0.6B/special_tokens_map.json +31 -0
- Qwen3-Embedding-0.6B/tokenizer.json +3 -0
- Qwen3-Embedding-0.6B/tokenizer_config.json +239 -0
- Qwen3-Embedding-0.6B/vocab.json +0 -0
- acestep-5Hz-lm-1.7B/added_tokens.json +0 -0
- acestep-5Hz-lm-1.7B/chat_template.jinja +89 -0
- acestep-5Hz-lm-1.7B/config.json +61 -0
- acestep-5Hz-lm-1.7B/merges.txt +0 -0
- acestep-5Hz-lm-1.7B/model.safetensors +3 -0
- acestep-5Hz-lm-1.7B/special_tokens_map.json +0 -0
- acestep-5Hz-lm-1.7B/tokenizer.json +3 -0
- acestep-5Hz-lm-1.7B/tokenizer_config.json +3 -0
- acestep-5Hz-lm-1.7B/vocab.json +0 -0
- acestep-v15-turbo/config.json +82 -0
- acestep-v15-turbo/configuration_acestep_v15.py +263 -0
- acestep-v15-turbo/model.safetensors +3 -0
- acestep-v15-turbo/modeling_acestep_v15_turbo.py +2136 -0
- acestep-v15-turbo/silence_latent.pt +3 -0
- config.json +82 -0
- vae/config.json +24 -0
- vae/diffusion_pytorch_model.safetensors +3 -0
.gitattributes
CHANGED
|
@@ -33,3 +33,6 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
|
|
| 33 |
*.zip filter=lfs diff=lfs merge=lfs -text
|
| 34 |
*.zst filter=lfs diff=lfs merge=lfs -text
|
| 35 |
*tfevents* filter=lfs diff=lfs merge=lfs -text
|
|
|
|
|
|
|
|
|
|
|
|
| 33 |
*.zip filter=lfs diff=lfs merge=lfs -text
|
| 34 |
*.zst filter=lfs diff=lfs merge=lfs -text
|
| 35 |
*tfevents* filter=lfs diff=lfs merge=lfs -text
|
| 36 |
+
acestep-5Hz-lm-1.7B/tokenizer.json filter=lfs diff=lfs merge=lfs -text
|
| 37 |
+
acestep-5Hz-lm-1.7B/tokenizer_config.json filter=lfs diff=lfs merge=lfs -text
|
| 38 |
+
Qwen3-Embedding-0.6B/tokenizer.json filter=lfs diff=lfs merge=lfs -text
|
Qwen3-Embedding-0.6B/added_tokens.json
ADDED
|
@@ -0,0 +1,28 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"</think>": 151668,
|
| 3 |
+
"</tool_call>": 151658,
|
| 4 |
+
"</tool_response>": 151666,
|
| 5 |
+
"<think>": 151667,
|
| 6 |
+
"<tool_call>": 151657,
|
| 7 |
+
"<tool_response>": 151665,
|
| 8 |
+
"<|box_end|>": 151649,
|
| 9 |
+
"<|box_start|>": 151648,
|
| 10 |
+
"<|endoftext|>": 151643,
|
| 11 |
+
"<|file_sep|>": 151664,
|
| 12 |
+
"<|fim_middle|>": 151660,
|
| 13 |
+
"<|fim_pad|>": 151662,
|
| 14 |
+
"<|fim_prefix|>": 151659,
|
| 15 |
+
"<|fim_suffix|>": 151661,
|
| 16 |
+
"<|im_end|>": 151645,
|
| 17 |
+
"<|im_start|>": 151644,
|
| 18 |
+
"<|image_pad|>": 151655,
|
| 19 |
+
"<|object_ref_end|>": 151647,
|
| 20 |
+
"<|object_ref_start|>": 151646,
|
| 21 |
+
"<|quad_end|>": 151651,
|
| 22 |
+
"<|quad_start|>": 151650,
|
| 23 |
+
"<|repo_name|>": 151663,
|
| 24 |
+
"<|video_pad|>": 151656,
|
| 25 |
+
"<|vision_end|>": 151653,
|
| 26 |
+
"<|vision_pad|>": 151654,
|
| 27 |
+
"<|vision_start|>": 151652
|
| 28 |
+
}
|
Qwen3-Embedding-0.6B/chat_template.jinja
ADDED
|
@@ -0,0 +1,85 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{%- if tools %}
|
| 2 |
+
{{- '<|im_start|>system\n' }}
|
| 3 |
+
{%- if messages[0].role == 'system' %}
|
| 4 |
+
{{- messages[0].content + '\n\n' }}
|
| 5 |
+
{%- endif %}
|
| 6 |
+
{{- "# Tools\n\nYou may call one or more functions to assist with the user query.\n\nYou are provided with function signatures within <tools></tools> XML tags:\n<tools>" }}
|
| 7 |
+
{%- for tool in tools %}
|
| 8 |
+
{{- "\n" }}
|
| 9 |
+
{{- tool | tojson }}
|
| 10 |
+
{%- endfor %}
|
| 11 |
+
{{- "\n</tools>\n\nFor each function call, return a json object with function name and arguments within <tool_call></tool_call> XML tags:\n<tool_call>\n{\"name\": <function-name>, \"arguments\": <args-json-object>}\n</tool_call><|im_end|>\n" }}
|
| 12 |
+
{%- else %}
|
| 13 |
+
{%- if messages[0].role == 'system' %}
|
| 14 |
+
{{- '<|im_start|>system\n' + messages[0].content + '<|im_end|>\n' }}
|
| 15 |
+
{%- endif %}
|
| 16 |
+
{%- endif %}
|
| 17 |
+
{%- set ns = namespace(multi_step_tool=true, last_query_index=messages|length - 1) %}
|
| 18 |
+
{%- for message in messages[::-1] %}
|
| 19 |
+
{%- set index = (messages|length - 1) - loop.index0 %}
|
| 20 |
+
{%- if ns.multi_step_tool and message.role == "user" and not(message.content.startswith('<tool_response>') and message.content.endswith('</tool_response>')) %}
|
| 21 |
+
{%- set ns.multi_step_tool = false %}
|
| 22 |
+
{%- set ns.last_query_index = index %}
|
| 23 |
+
{%- endif %}
|
| 24 |
+
{%- endfor %}
|
| 25 |
+
{%- for message in messages %}
|
| 26 |
+
{%- if (message.role == "user") or (message.role == "system" and not loop.first) %}
|
| 27 |
+
{{- '<|im_start|>' + message.role + '\n' + message.content + '<|im_end|>' + '\n' }}
|
| 28 |
+
{%- elif message.role == "assistant" %}
|
| 29 |
+
{%- set content = message.content %}
|
| 30 |
+
{%- set reasoning_content = '' %}
|
| 31 |
+
{%- if message.reasoning_content is defined and message.reasoning_content is not none %}
|
| 32 |
+
{%- set reasoning_content = message.reasoning_content %}
|
| 33 |
+
{%- else %}
|
| 34 |
+
{%- if '</think>' in message.content %}
|
| 35 |
+
{%- set content = message.content.split('</think>')[-1].lstrip('\n') %}
|
| 36 |
+
{%- set reasoning_content = message.content.split('</think>')[0].rstrip('\n').split('<think>')[-1].lstrip('\n') %}
|
| 37 |
+
{%- endif %}
|
| 38 |
+
{%- endif %}
|
| 39 |
+
{%- if loop.index0 > ns.last_query_index %}
|
| 40 |
+
{%- if loop.last or (not loop.last and reasoning_content) %}
|
| 41 |
+
{{- '<|im_start|>' + message.role + '\n<think>\n' + reasoning_content.strip('\n') + '\n</think>\n\n' + content.lstrip('\n') }}
|
| 42 |
+
{%- else %}
|
| 43 |
+
{{- '<|im_start|>' + message.role + '\n' + content }}
|
| 44 |
+
{%- endif %}
|
| 45 |
+
{%- else %}
|
| 46 |
+
{{- '<|im_start|>' + message.role + '\n' + content }}
|
| 47 |
+
{%- endif %}
|
| 48 |
+
{%- if message.tool_calls %}
|
| 49 |
+
{%- for tool_call in message.tool_calls %}
|
| 50 |
+
{%- if (loop.first and content) or (not loop.first) %}
|
| 51 |
+
{{- '\n' }}
|
| 52 |
+
{%- endif %}
|
| 53 |
+
{%- if tool_call.function %}
|
| 54 |
+
{%- set tool_call = tool_call.function %}
|
| 55 |
+
{%- endif %}
|
| 56 |
+
{{- '<tool_call>\n{"name": "' }}
|
| 57 |
+
{{- tool_call.name }}
|
| 58 |
+
{{- '", "arguments": ' }}
|
| 59 |
+
{%- if tool_call.arguments is string %}
|
| 60 |
+
{{- tool_call.arguments }}
|
| 61 |
+
{%- else %}
|
| 62 |
+
{{- tool_call.arguments | tojson }}
|
| 63 |
+
{%- endif %}
|
| 64 |
+
{{- '}\n</tool_call>' }}
|
| 65 |
+
{%- endfor %}
|
| 66 |
+
{%- endif %}
|
| 67 |
+
{{- '<|im_end|>\n' }}
|
| 68 |
+
{%- elif message.role == "tool" %}
|
| 69 |
+
{%- if loop.first or (messages[loop.index0 - 1].role != "tool") %}
|
| 70 |
+
{{- '<|im_start|>user' }}
|
| 71 |
+
{%- endif %}
|
| 72 |
+
{{- '\n<tool_response>\n' }}
|
| 73 |
+
{{- message.content }}
|
| 74 |
+
{{- '\n</tool_response>' }}
|
| 75 |
+
{%- if loop.last or (messages[loop.index0 + 1].role != "tool") %}
|
| 76 |
+
{{- '<|im_end|>\n' }}
|
| 77 |
+
{%- endif %}
|
| 78 |
+
{%- endif %}
|
| 79 |
+
{%- endfor %}
|
| 80 |
+
{%- if add_generation_prompt %}
|
| 81 |
+
{{- '<|im_start|>assistant\n' }}
|
| 82 |
+
{%- if enable_thinking is defined and enable_thinking is false %}
|
| 83 |
+
{{- '<think>\n\n</think>\n\n' }}
|
| 84 |
+
{%- endif %}
|
| 85 |
+
{%- endif %}
|
Qwen3-Embedding-0.6B/config.json
ADDED
|
@@ -0,0 +1,60 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"architectures": [
|
| 3 |
+
"Qwen3Model"
|
| 4 |
+
],
|
| 5 |
+
"attention_bias": false,
|
| 6 |
+
"attention_dropout": 0.0,
|
| 7 |
+
"bos_token_id": 151643,
|
| 8 |
+
"dtype": "bfloat16",
|
| 9 |
+
"eos_token_id": 151643,
|
| 10 |
+
"head_dim": 128,
|
| 11 |
+
"hidden_act": "silu",
|
| 12 |
+
"hidden_size": 1024,
|
| 13 |
+
"initializer_range": 0.02,
|
| 14 |
+
"intermediate_size": 3072,
|
| 15 |
+
"layer_types": [
|
| 16 |
+
"full_attention",
|
| 17 |
+
"full_attention",
|
| 18 |
+
"full_attention",
|
| 19 |
+
"full_attention",
|
| 20 |
+
"full_attention",
|
| 21 |
+
"full_attention",
|
| 22 |
+
"full_attention",
|
| 23 |
+
"full_attention",
|
| 24 |
+
"full_attention",
|
| 25 |
+
"full_attention",
|
| 26 |
+
"full_attention",
|
| 27 |
+
"full_attention",
|
| 28 |
+
"full_attention",
|
| 29 |
+
"full_attention",
|
| 30 |
+
"full_attention",
|
| 31 |
+
"full_attention",
|
| 32 |
+
"full_attention",
|
| 33 |
+
"full_attention",
|
| 34 |
+
"full_attention",
|
| 35 |
+
"full_attention",
|
| 36 |
+
"full_attention",
|
| 37 |
+
"full_attention",
|
| 38 |
+
"full_attention",
|
| 39 |
+
"full_attention",
|
| 40 |
+
"full_attention",
|
| 41 |
+
"full_attention",
|
| 42 |
+
"full_attention",
|
| 43 |
+
"full_attention"
|
| 44 |
+
],
|
| 45 |
+
"max_position_embeddings": 32768,
|
| 46 |
+
"max_window_layers": 28,
|
| 47 |
+
"model_type": "qwen3",
|
| 48 |
+
"num_attention_heads": 16,
|
| 49 |
+
"num_hidden_layers": 28,
|
| 50 |
+
"num_key_value_heads": 8,
|
| 51 |
+
"rms_norm_eps": 1e-06,
|
| 52 |
+
"rope_scaling": null,
|
| 53 |
+
"rope_theta": 1000000,
|
| 54 |
+
"sliding_window": null,
|
| 55 |
+
"tie_word_embeddings": true,
|
| 56 |
+
"transformers_version": "4.57.0.dev0",
|
| 57 |
+
"use_cache": true,
|
| 58 |
+
"use_sliding_window": false,
|
| 59 |
+
"vocab_size": 151669
|
| 60 |
+
}
|
Qwen3-Embedding-0.6B/merges.txt
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
Qwen3-Embedding-0.6B/model.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:0437e45c94563b09e13cb7a64478fc406947a93cb34a7e05870fc8dcd48e23fd
|
| 3 |
+
size 1191586416
|
Qwen3-Embedding-0.6B/special_tokens_map.json
ADDED
|
@@ -0,0 +1,31 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"additional_special_tokens": [
|
| 3 |
+
"<|im_start|>",
|
| 4 |
+
"<|im_end|>",
|
| 5 |
+
"<|object_ref_start|>",
|
| 6 |
+
"<|object_ref_end|>",
|
| 7 |
+
"<|box_start|>",
|
| 8 |
+
"<|box_end|>",
|
| 9 |
+
"<|quad_start|>",
|
| 10 |
+
"<|quad_end|>",
|
| 11 |
+
"<|vision_start|>",
|
| 12 |
+
"<|vision_end|>",
|
| 13 |
+
"<|vision_pad|>",
|
| 14 |
+
"<|image_pad|>",
|
| 15 |
+
"<|video_pad|>"
|
| 16 |
+
],
|
| 17 |
+
"eos_token": {
|
| 18 |
+
"content": "<|im_end|>",
|
| 19 |
+
"lstrip": false,
|
| 20 |
+
"normalized": false,
|
| 21 |
+
"rstrip": false,
|
| 22 |
+
"single_word": false
|
| 23 |
+
},
|
| 24 |
+
"pad_token": {
|
| 25 |
+
"content": "<|endoftext|>",
|
| 26 |
+
"lstrip": false,
|
| 27 |
+
"normalized": false,
|
| 28 |
+
"rstrip": false,
|
| 29 |
+
"single_word": false
|
| 30 |
+
}
|
| 31 |
+
}
|
Qwen3-Embedding-0.6B/tokenizer.json
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:def76fb086971c7867b829c23a26261e38d9d74e02139253b38aeb9df8b4b50a
|
| 3 |
+
size 11423705
|
Qwen3-Embedding-0.6B/tokenizer_config.json
ADDED
|
@@ -0,0 +1,239 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"add_bos_token": false,
|
| 3 |
+
"add_prefix_space": false,
|
| 4 |
+
"added_tokens_decoder": {
|
| 5 |
+
"151643": {
|
| 6 |
+
"content": "<|endoftext|>",
|
| 7 |
+
"lstrip": false,
|
| 8 |
+
"normalized": false,
|
| 9 |
+
"rstrip": false,
|
| 10 |
+
"single_word": false,
|
| 11 |
+
"special": true
|
| 12 |
+
},
|
| 13 |
+
"151644": {
|
| 14 |
+
"content": "<|im_start|>",
|
| 15 |
+
"lstrip": false,
|
| 16 |
+
"normalized": false,
|
| 17 |
+
"rstrip": false,
|
| 18 |
+
"single_word": false,
|
| 19 |
+
"special": true
|
| 20 |
+
},
|
| 21 |
+
"151645": {
|
| 22 |
+
"content": "<|im_end|>",
|
| 23 |
+
"lstrip": false,
|
| 24 |
+
"normalized": false,
|
| 25 |
+
"rstrip": false,
|
| 26 |
+
"single_word": false,
|
| 27 |
+
"special": true
|
| 28 |
+
},
|
| 29 |
+
"151646": {
|
| 30 |
+
"content": "<|object_ref_start|>",
|
| 31 |
+
"lstrip": false,
|
| 32 |
+
"normalized": false,
|
| 33 |
+
"rstrip": false,
|
| 34 |
+
"single_word": false,
|
| 35 |
+
"special": true
|
| 36 |
+
},
|
| 37 |
+
"151647": {
|
| 38 |
+
"content": "<|object_ref_end|>",
|
| 39 |
+
"lstrip": false,
|
| 40 |
+
"normalized": false,
|
| 41 |
+
"rstrip": false,
|
| 42 |
+
"single_word": false,
|
| 43 |
+
"special": true
|
| 44 |
+
},
|
| 45 |
+
"151648": {
|
| 46 |
+
"content": "<|box_start|>",
|
| 47 |
+
"lstrip": false,
|
| 48 |
+
"normalized": false,
|
| 49 |
+
"rstrip": false,
|
| 50 |
+
"single_word": false,
|
| 51 |
+
"special": true
|
| 52 |
+
},
|
| 53 |
+
"151649": {
|
| 54 |
+
"content": "<|box_end|>",
|
| 55 |
+
"lstrip": false,
|
| 56 |
+
"normalized": false,
|
| 57 |
+
"rstrip": false,
|
| 58 |
+
"single_word": false,
|
| 59 |
+
"special": true
|
| 60 |
+
},
|
| 61 |
+
"151650": {
|
| 62 |
+
"content": "<|quad_start|>",
|
| 63 |
+
"lstrip": false,
|
| 64 |
+
"normalized": false,
|
| 65 |
+
"rstrip": false,
|
| 66 |
+
"single_word": false,
|
| 67 |
+
"special": true
|
| 68 |
+
},
|
| 69 |
+
"151651": {
|
| 70 |
+
"content": "<|quad_end|>",
|
| 71 |
+
"lstrip": false,
|
| 72 |
+
"normalized": false,
|
| 73 |
+
"rstrip": false,
|
| 74 |
+
"single_word": false,
|
| 75 |
+
"special": true
|
| 76 |
+
},
|
| 77 |
+
"151652": {
|
| 78 |
+
"content": "<|vision_start|>",
|
| 79 |
+
"lstrip": false,
|
| 80 |
+
"normalized": false,
|
| 81 |
+
"rstrip": false,
|
| 82 |
+
"single_word": false,
|
| 83 |
+
"special": true
|
| 84 |
+
},
|
| 85 |
+
"151653": {
|
| 86 |
+
"content": "<|vision_end|>",
|
| 87 |
+
"lstrip": false,
|
| 88 |
+
"normalized": false,
|
| 89 |
+
"rstrip": false,
|
| 90 |
+
"single_word": false,
|
| 91 |
+
"special": true
|
| 92 |
+
},
|
| 93 |
+
"151654": {
|
| 94 |
+
"content": "<|vision_pad|>",
|
| 95 |
+
"lstrip": false,
|
| 96 |
+
"normalized": false,
|
| 97 |
+
"rstrip": false,
|
| 98 |
+
"single_word": false,
|
| 99 |
+
"special": true
|
| 100 |
+
},
|
| 101 |
+
"151655": {
|
| 102 |
+
"content": "<|image_pad|>",
|
| 103 |
+
"lstrip": false,
|
| 104 |
+
"normalized": false,
|
| 105 |
+
"rstrip": false,
|
| 106 |
+
"single_word": false,
|
| 107 |
+
"special": true
|
| 108 |
+
},
|
| 109 |
+
"151656": {
|
| 110 |
+
"content": "<|video_pad|>",
|
| 111 |
+
"lstrip": false,
|
| 112 |
+
"normalized": false,
|
| 113 |
+
"rstrip": false,
|
| 114 |
+
"single_word": false,
|
| 115 |
+
"special": true
|
| 116 |
+
},
|
| 117 |
+
"151657": {
|
| 118 |
+
"content": "<tool_call>",
|
| 119 |
+
"lstrip": false,
|
| 120 |
+
"normalized": false,
|
| 121 |
+
"rstrip": false,
|
| 122 |
+
"single_word": false,
|
| 123 |
+
"special": false
|
| 124 |
+
},
|
| 125 |
+
"151658": {
|
| 126 |
+
"content": "</tool_call>",
|
| 127 |
+
"lstrip": false,
|
| 128 |
+
"normalized": false,
|
| 129 |
+
"rstrip": false,
|
| 130 |
+
"single_word": false,
|
| 131 |
+
"special": false
|
| 132 |
+
},
|
| 133 |
+
"151659": {
|
| 134 |
+
"content": "<|fim_prefix|>",
|
| 135 |
+
"lstrip": false,
|
| 136 |
+
"normalized": false,
|
| 137 |
+
"rstrip": false,
|
| 138 |
+
"single_word": false,
|
| 139 |
+
"special": false
|
| 140 |
+
},
|
| 141 |
+
"151660": {
|
| 142 |
+
"content": "<|fim_middle|>",
|
| 143 |
+
"lstrip": false,
|
| 144 |
+
"normalized": false,
|
| 145 |
+
"rstrip": false,
|
| 146 |
+
"single_word": false,
|
| 147 |
+
"special": false
|
| 148 |
+
},
|
| 149 |
+
"151661": {
|
| 150 |
+
"content": "<|fim_suffix|>",
|
| 151 |
+
"lstrip": false,
|
| 152 |
+
"normalized": false,
|
| 153 |
+
"rstrip": false,
|
| 154 |
+
"single_word": false,
|
| 155 |
+
"special": false
|
| 156 |
+
},
|
| 157 |
+
"151662": {
|
| 158 |
+
"content": "<|fim_pad|>",
|
| 159 |
+
"lstrip": false,
|
| 160 |
+
"normalized": false,
|
| 161 |
+
"rstrip": false,
|
| 162 |
+
"single_word": false,
|
| 163 |
+
"special": false
|
| 164 |
+
},
|
| 165 |
+
"151663": {
|
| 166 |
+
"content": "<|repo_name|>",
|
| 167 |
+
"lstrip": false,
|
| 168 |
+
"normalized": false,
|
| 169 |
+
"rstrip": false,
|
| 170 |
+
"single_word": false,
|
| 171 |
+
"special": false
|
| 172 |
+
},
|
| 173 |
+
"151664": {
|
| 174 |
+
"content": "<|file_sep|>",
|
| 175 |
+
"lstrip": false,
|
| 176 |
+
"normalized": false,
|
| 177 |
+
"rstrip": false,
|
| 178 |
+
"single_word": false,
|
| 179 |
+
"special": false
|
| 180 |
+
},
|
| 181 |
+
"151665": {
|
| 182 |
+
"content": "<tool_response>",
|
| 183 |
+
"lstrip": false,
|
| 184 |
+
"normalized": false,
|
| 185 |
+
"rstrip": false,
|
| 186 |
+
"single_word": false,
|
| 187 |
+
"special": false
|
| 188 |
+
},
|
| 189 |
+
"151666": {
|
| 190 |
+
"content": "</tool_response>",
|
| 191 |
+
"lstrip": false,
|
| 192 |
+
"normalized": false,
|
| 193 |
+
"rstrip": false,
|
| 194 |
+
"single_word": false,
|
| 195 |
+
"special": false
|
| 196 |
+
},
|
| 197 |
+
"151667": {
|
| 198 |
+
"content": "<think>",
|
| 199 |
+
"lstrip": false,
|
| 200 |
+
"normalized": false,
|
| 201 |
+
"rstrip": false,
|
| 202 |
+
"single_word": false,
|
| 203 |
+
"special": false
|
| 204 |
+
},
|
| 205 |
+
"151668": {
|
| 206 |
+
"content": "</think>",
|
| 207 |
+
"lstrip": false,
|
| 208 |
+
"normalized": false,
|
| 209 |
+
"rstrip": false,
|
| 210 |
+
"single_word": false,
|
| 211 |
+
"special": false
|
| 212 |
+
}
|
| 213 |
+
},
|
| 214 |
+
"additional_special_tokens": [
|
| 215 |
+
"<|im_start|>",
|
| 216 |
+
"<|im_end|>",
|
| 217 |
+
"<|object_ref_start|>",
|
| 218 |
+
"<|object_ref_end|>",
|
| 219 |
+
"<|box_start|>",
|
| 220 |
+
"<|box_end|>",
|
| 221 |
+
"<|quad_start|>",
|
| 222 |
+
"<|quad_end|>",
|
| 223 |
+
"<|vision_start|>",
|
| 224 |
+
"<|vision_end|>",
|
| 225 |
+
"<|vision_pad|>",
|
| 226 |
+
"<|image_pad|>",
|
| 227 |
+
"<|video_pad|>"
|
| 228 |
+
],
|
| 229 |
+
"bos_token": null,
|
| 230 |
+
"clean_up_tokenization_spaces": false,
|
| 231 |
+
"eos_token": "<|im_end|>",
|
| 232 |
+
"errors": "replace",
|
| 233 |
+
"extra_special_tokens": {},
|
| 234 |
+
"model_max_length": 131072,
|
| 235 |
+
"pad_token": "<|endoftext|>",
|
| 236 |
+
"split_special_tokens": false,
|
| 237 |
+
"tokenizer_class": "Qwen2Tokenizer",
|
| 238 |
+
"unk_token": null
|
| 239 |
+
}
|
Qwen3-Embedding-0.6B/vocab.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
acestep-5Hz-lm-1.7B/added_tokens.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
acestep-5Hz-lm-1.7B/chat_template.jinja
ADDED
|
@@ -0,0 +1,89 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{%- if tools %}
|
| 2 |
+
{{- '<|im_start|>system\n' }}
|
| 3 |
+
{%- if messages[0].role == 'system' %}
|
| 4 |
+
{{- messages[0].content + '\n\n' }}
|
| 5 |
+
{%- endif %}
|
| 6 |
+
{{- "# Tools\n\nYou may call one or more functions to assist with the user query.\n\nYou are provided with function signatures within <tools></tools> XML tags:\n<tools>" }}
|
| 7 |
+
{%- for tool in tools %}
|
| 8 |
+
{{- "\n" }}
|
| 9 |
+
{{- tool | tojson }}
|
| 10 |
+
{%- endfor %}
|
| 11 |
+
{{- "\n</tools>\n\nFor each function call, return a json object with function name and arguments within <tool_call></tool_call> XML tags:\n<tool_call>\n{\"name\": <function-name>, \"arguments\": <args-json-object>}\n</tool_call><|im_end|>\n" }}
|
| 12 |
+
{%- else %}
|
| 13 |
+
{%- if messages[0].role == 'system' %}
|
| 14 |
+
{{- '<|im_start|>system\n' + messages[0].content + '<|im_end|>\n' }}
|
| 15 |
+
{%- endif %}
|
| 16 |
+
{%- endif %}
|
| 17 |
+
{%- set ns = namespace(multi_step_tool=true, last_query_index=messages|length - 1) %}
|
| 18 |
+
{%- for message in messages[::-1] %}
|
| 19 |
+
{%- set index = (messages|length - 1) - loop.index0 %}
|
| 20 |
+
{%- if ns.multi_step_tool and message.role == "user" and message.content is string and not(message.content.startswith('<tool_response>') and message.content.endswith('</tool_response>')) %}
|
| 21 |
+
{%- set ns.multi_step_tool = false %}
|
| 22 |
+
{%- set ns.last_query_index = index %}
|
| 23 |
+
{%- endif %}
|
| 24 |
+
{%- endfor %}
|
| 25 |
+
{%- for message in messages %}
|
| 26 |
+
{%- if message.content is string %}
|
| 27 |
+
{%- set content = message.content %}
|
| 28 |
+
{%- else %}
|
| 29 |
+
{%- set content = '' %}
|
| 30 |
+
{%- endif %}
|
| 31 |
+
{%- if (message.role == "user") or (message.role == "system" and not loop.first) %}
|
| 32 |
+
{{- '<|im_start|>' + message.role + '\n' + content + '<|im_end|>' + '\n' }}
|
| 33 |
+
{%- elif message.role == "assistant" %}
|
| 34 |
+
{%- set reasoning_content = '' %}
|
| 35 |
+
{%- if message.reasoning_content is string %}
|
| 36 |
+
{%- set reasoning_content = message.reasoning_content %}
|
| 37 |
+
{%- else %}
|
| 38 |
+
{%- if '</think>' in content %}
|
| 39 |
+
{%- set reasoning_content = content.split('</think>')[0].rstrip('\n').split('<think>')[-1].lstrip('\n') %}
|
| 40 |
+
{%- set content = content.split('</think>')[-1].lstrip('\n') %}
|
| 41 |
+
{%- endif %}
|
| 42 |
+
{%- endif %}
|
| 43 |
+
{%- if loop.index0 > ns.last_query_index %}
|
| 44 |
+
{%- if loop.last or (not loop.last and reasoning_content) %}
|
| 45 |
+
{{- '<|im_start|>' + message.role + '\n<think>\n' + reasoning_content.strip('\n') + '\n</think>\n\n' + content.lstrip('\n') }}
|
| 46 |
+
{%- else %}
|
| 47 |
+
{{- '<|im_start|>' + message.role + '\n' + content }}
|
| 48 |
+
{%- endif %}
|
| 49 |
+
{%- else %}
|
| 50 |
+
{{- '<|im_start|>' + message.role + '\n' + content }}
|
| 51 |
+
{%- endif %}
|
| 52 |
+
{%- if message.tool_calls %}
|
| 53 |
+
{%- for tool_call in message.tool_calls %}
|
| 54 |
+
{%- if (loop.first and content) or (not loop.first) %}
|
| 55 |
+
{{- '\n' }}
|
| 56 |
+
{%- endif %}
|
| 57 |
+
{%- if tool_call.function %}
|
| 58 |
+
{%- set tool_call = tool_call.function %}
|
| 59 |
+
{%- endif %}
|
| 60 |
+
{{- '<tool_call>\n{"name": "' }}
|
| 61 |
+
{{- tool_call.name }}
|
| 62 |
+
{{- '", "arguments": ' }}
|
| 63 |
+
{%- if tool_call.arguments is string %}
|
| 64 |
+
{{- tool_call.arguments }}
|
| 65 |
+
{%- else %}
|
| 66 |
+
{{- tool_call.arguments | tojson }}
|
| 67 |
+
{%- endif %}
|
| 68 |
+
{{- '}\n</tool_call>' }}
|
| 69 |
+
{%- endfor %}
|
| 70 |
+
{%- endif %}
|
| 71 |
+
{{- '<|im_end|>\n' }}
|
| 72 |
+
{%- elif message.role == "tool" %}
|
| 73 |
+
{%- if loop.first or (messages[loop.index0 - 1].role != "tool") %}
|
| 74 |
+
{{- '<|im_start|>user' }}
|
| 75 |
+
{%- endif %}
|
| 76 |
+
{{- '\n<tool_response>\n' }}
|
| 77 |
+
{{- content }}
|
| 78 |
+
{{- '\n</tool_response>' }}
|
| 79 |
+
{%- if loop.last or (messages[loop.index0 + 1].role != "tool") %}
|
| 80 |
+
{{- '<|im_end|>\n' }}
|
| 81 |
+
{%- endif %}
|
| 82 |
+
{%- endif %}
|
| 83 |
+
{%- endfor %}
|
| 84 |
+
{%- if add_generation_prompt %}
|
| 85 |
+
{{- '<|im_start|>assistant\n' }}
|
| 86 |
+
{%- if enable_thinking is defined and enable_thinking is false %}
|
| 87 |
+
{{- '<think>\n\n</think>\n\n' }}
|
| 88 |
+
{%- endif %}
|
| 89 |
+
{%- endif %}
|
acestep-5Hz-lm-1.7B/config.json
ADDED
|
@@ -0,0 +1,61 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"architectures": [
|
| 3 |
+
"Qwen3Model"
|
| 4 |
+
],
|
| 5 |
+
"attention_bias": false,
|
| 6 |
+
"attention_dropout": 0.0,
|
| 7 |
+
"bos_token_id": 151643,
|
| 8 |
+
"dtype": "bfloat16",
|
| 9 |
+
"eos_token_id": 151645,
|
| 10 |
+
"head_dim": 128,
|
| 11 |
+
"hidden_act": "silu",
|
| 12 |
+
"hidden_size": 2048,
|
| 13 |
+
"initializer_range": 0.02,
|
| 14 |
+
"intermediate_size": 6144,
|
| 15 |
+
"layer_types": [
|
| 16 |
+
"full_attention",
|
| 17 |
+
"full_attention",
|
| 18 |
+
"full_attention",
|
| 19 |
+
"full_attention",
|
| 20 |
+
"full_attention",
|
| 21 |
+
"full_attention",
|
| 22 |
+
"full_attention",
|
| 23 |
+
"full_attention",
|
| 24 |
+
"full_attention",
|
| 25 |
+
"full_attention",
|
| 26 |
+
"full_attention",
|
| 27 |
+
"full_attention",
|
| 28 |
+
"full_attention",
|
| 29 |
+
"full_attention",
|
| 30 |
+
"full_attention",
|
| 31 |
+
"full_attention",
|
| 32 |
+
"full_attention",
|
| 33 |
+
"full_attention",
|
| 34 |
+
"full_attention",
|
| 35 |
+
"full_attention",
|
| 36 |
+
"full_attention",
|
| 37 |
+
"full_attention",
|
| 38 |
+
"full_attention",
|
| 39 |
+
"full_attention",
|
| 40 |
+
"full_attention",
|
| 41 |
+
"full_attention",
|
| 42 |
+
"full_attention",
|
| 43 |
+
"full_attention"
|
| 44 |
+
],
|
| 45 |
+
"max_position_embeddings": 40960,
|
| 46 |
+
"max_window_layers": 28,
|
| 47 |
+
"model_type": "qwen3",
|
| 48 |
+
"num_attention_heads": 16,
|
| 49 |
+
"num_hidden_layers": 28,
|
| 50 |
+
"num_key_value_heads": 8,
|
| 51 |
+
"pad_token_id": 151643,
|
| 52 |
+
"rms_norm_eps": 1e-06,
|
| 53 |
+
"rope_scaling": null,
|
| 54 |
+
"rope_theta": 1000000,
|
| 55 |
+
"sliding_window": null,
|
| 56 |
+
"tie_word_embeddings": true,
|
| 57 |
+
"transformers_version": "4.57.0.dev0",
|
| 58 |
+
"use_cache": true,
|
| 59 |
+
"use_sliding_window": false,
|
| 60 |
+
"vocab_size": 217204
|
| 61 |
+
}
|
acestep-5Hz-lm-1.7B/merges.txt
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
acestep-5Hz-lm-1.7B/model.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:f161689da73e5ecefa28ff780d51c2d92a00f056d021d7933c779ed5c6cd7db8
|
| 3 |
+
size 3708521528
|
acestep-5Hz-lm-1.7B/special_tokens_map.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
acestep-5Hz-lm-1.7B/tokenizer.json
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:35af56c3f5cb3ea2cc578aa28a8937770981d504f183ac5c8c38baf4bbd4af4d
|
| 3 |
+
size 24321939
|
acestep-5Hz-lm-1.7B/tokenizer_config.json
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:6cd70cdd89425971794f5235562edcc608b0629a6c4686ae51a8b8c8b8ba5e95
|
| 3 |
+
size 14072925
|
acestep-5Hz-lm-1.7B/vocab.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
acestep-v15-turbo/config.json
ADDED
|
@@ -0,0 +1,82 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"architectures": [
|
| 3 |
+
"AceStepConditionGenerationModel"
|
| 4 |
+
],
|
| 5 |
+
"attention_bias": false,
|
| 6 |
+
"attention_dropout": 0.0,
|
| 7 |
+
"audio_acoustic_hidden_dim": 64,
|
| 8 |
+
"auto_map": {
|
| 9 |
+
"AutoConfig": "configuration_acestep_v15.AceStepConfig",
|
| 10 |
+
"AutoModel": "modeling_acestep_v15_turbo.AceStepConditionGenerationModel"
|
| 11 |
+
},
|
| 12 |
+
"data_proportion": 0.5,
|
| 13 |
+
"dtype": "bfloat16",
|
| 14 |
+
"fsq_dim": 2048,
|
| 15 |
+
"fsq_input_levels": [
|
| 16 |
+
8,
|
| 17 |
+
8,
|
| 18 |
+
8,
|
| 19 |
+
5,
|
| 20 |
+
5,
|
| 21 |
+
5
|
| 22 |
+
],
|
| 23 |
+
"fsq_input_num_quantizers": 1,
|
| 24 |
+
"head_dim": 128,
|
| 25 |
+
"hidden_act": "silu",
|
| 26 |
+
"hidden_size": 2048,
|
| 27 |
+
"in_channels": 192,
|
| 28 |
+
"initializer_range": 0.02,
|
| 29 |
+
"intermediate_size": 6144,
|
| 30 |
+
"is_turbo": true,
|
| 31 |
+
"layer_types": [
|
| 32 |
+
"sliding_attention",
|
| 33 |
+
"full_attention",
|
| 34 |
+
"sliding_attention",
|
| 35 |
+
"full_attention",
|
| 36 |
+
"sliding_attention",
|
| 37 |
+
"full_attention",
|
| 38 |
+
"sliding_attention",
|
| 39 |
+
"full_attention",
|
| 40 |
+
"sliding_attention",
|
| 41 |
+
"full_attention",
|
| 42 |
+
"sliding_attention",
|
| 43 |
+
"full_attention",
|
| 44 |
+
"sliding_attention",
|
| 45 |
+
"full_attention",
|
| 46 |
+
"sliding_attention",
|
| 47 |
+
"full_attention",
|
| 48 |
+
"sliding_attention",
|
| 49 |
+
"full_attention",
|
| 50 |
+
"sliding_attention",
|
| 51 |
+
"full_attention",
|
| 52 |
+
"sliding_attention",
|
| 53 |
+
"full_attention",
|
| 54 |
+
"sliding_attention",
|
| 55 |
+
"full_attention"
|
| 56 |
+
],
|
| 57 |
+
"max_position_embeddings": 32768,
|
| 58 |
+
"model_type": "acestep",
|
| 59 |
+
"model_version": "turbo",
|
| 60 |
+
"num_attention_heads": 16,
|
| 61 |
+
"num_attention_pooler_hidden_layers": 2,
|
| 62 |
+
"num_audio_decoder_hidden_layers": 24,
|
| 63 |
+
"num_hidden_layers": 24,
|
| 64 |
+
"num_key_value_heads": 8,
|
| 65 |
+
"num_lyric_encoder_hidden_layers": 8,
|
| 66 |
+
"num_timbre_encoder_hidden_layers": 4,
|
| 67 |
+
"patch_size": 2,
|
| 68 |
+
"pool_window_size": 5,
|
| 69 |
+
"rms_norm_eps": 1e-06,
|
| 70 |
+
"rope_scaling": null,
|
| 71 |
+
"rope_theta": 1000000,
|
| 72 |
+
"sliding_window": 128,
|
| 73 |
+
"text_hidden_dim": 1024,
|
| 74 |
+
"timbre_fix_frame": 750,
|
| 75 |
+
"timbre_hidden_dim": 64,
|
| 76 |
+
"timestep_mu": -0.4,
|
| 77 |
+
"timestep_sigma": 1.0,
|
| 78 |
+
"transformers_version": "4.57.0.dev0",
|
| 79 |
+
"use_cache": true,
|
| 80 |
+
"use_sliding_window": true,
|
| 81 |
+
"vocab_size": 64003
|
| 82 |
+
}
|
acestep-v15-turbo/configuration_acestep_v15.py
ADDED
|
@@ -0,0 +1,263 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# coding=utf-8
|
| 2 |
+
# Copyright 2024 The Qwen team, Alibaba Group and the HuggingFace Inc. team. All rights reserved.
|
| 3 |
+
#
|
| 4 |
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
| 5 |
+
# you may not use this file except in compliance with the License.
|
| 6 |
+
# You may obtain a copy of the License at
|
| 7 |
+
#
|
| 8 |
+
# http://www.apache.org/licenses/LICENSE-2.0
|
| 9 |
+
#
|
| 10 |
+
# Unless required by applicable law or agreed to in writing, software
|
| 11 |
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
| 12 |
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
| 13 |
+
# See the License for the specific language governing permissions and
|
| 14 |
+
# limitations under the License.
|
| 15 |
+
"""AceStep model configuration"""
|
| 16 |
+
|
| 17 |
+
from transformers.configuration_utils import PretrainedConfig, layer_type_validation
|
| 18 |
+
from transformers.modeling_rope_utils import rope_config_validation
|
| 19 |
+
from transformers.utils import logging
|
| 20 |
+
|
| 21 |
+
|
| 22 |
+
logger = logging.get_logger(__name__)
|
| 23 |
+
|
| 24 |
+
|
| 25 |
+
class AceStepConfig(PretrainedConfig):
|
| 26 |
+
r"""
|
| 27 |
+
This is the configuration class to store the configuration of a [`AceStepModel`]. It is used to instantiate an
|
| 28 |
+
AceStep model according to the specified arguments, defining the model architecture.
|
| 29 |
+
|
| 30 |
+
Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
|
| 31 |
+
documentation from [`PretrainedConfig`] for more information.
|
| 32 |
+
|
| 33 |
+
Args:
|
| 34 |
+
vocab_size (`int`, *optional*, defaults to 64003):
|
| 35 |
+
Vocabulary size of the AceStep model. Defines the number of different tokens that can be represented by the
|
| 36 |
+
`inputs_ids` passed when calling the model.
|
| 37 |
+
hidden_size (`int`, *optional*, defaults to 4096):
|
| 38 |
+
Dimension of the hidden representations.
|
| 39 |
+
intermediate_size (`int`, *optional*, defaults to 22016):
|
| 40 |
+
Dimension of the MLP representations.
|
| 41 |
+
num_hidden_layers (`int`, *optional*, defaults to 32):
|
| 42 |
+
Number of hidden layers in the Transformer encoder.
|
| 43 |
+
num_attention_heads (`int`, *optional*, defaults to 32):
|
| 44 |
+
Number of attention heads for each attention layer in the Transformer encoder.
|
| 45 |
+
num_key_value_heads (`int`, *optional*, defaults to 32):
|
| 46 |
+
This is the number of key_value heads that should be used to implement Grouped Query Attention. If
|
| 47 |
+
`num_key_value_heads=num_attention_heads`, the model will use Multi Head Attention (MHA), if
|
| 48 |
+
`num_key_value_heads=1` the model will use Multi Query Attention (MQA) otherwise GQA is used. When
|
| 49 |
+
converting a multi-head checkpoint to a GQA checkpoint, each group key and value head should be constructed
|
| 50 |
+
by meanpooling all the original heads within that group. For more details, check out [this
|
| 51 |
+
paper](https://huggingface.co/papers/2305.13245). If it is not specified, will default to `32`.
|
| 52 |
+
head_dim (`int`, *optional*, defaults to 128):
|
| 53 |
+
The attention head dimension.
|
| 54 |
+
hidden_act (`str` or `function`, *optional*, defaults to `"silu"`):
|
| 55 |
+
The non-linear activation function (function or string) in the decoder.
|
| 56 |
+
max_position_embeddings (`int`, *optional*, defaults to 32768):
|
| 57 |
+
The maximum sequence length that this model might ever be used with.
|
| 58 |
+
initializer_range (`float`, *optional*, defaults to 0.02):
|
| 59 |
+
The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
|
| 60 |
+
rms_norm_eps (`float`, *optional*, defaults to 1e-06):
|
| 61 |
+
The epsilon used by the rms normalization layers.
|
| 62 |
+
use_cache (`bool`, *optional*, defaults to `True`):
|
| 63 |
+
Whether or not the model should return the last key/values attentions (not used by all models). Only
|
| 64 |
+
relevant if `config.is_decoder=True`.
|
| 65 |
+
tie_word_embeddings (`bool`, *optional*, defaults to `False`):
|
| 66 |
+
Whether the model's input and output word embeddings should be tied.
|
| 67 |
+
rope_theta (`float`, *optional*, defaults to 10000.0):
|
| 68 |
+
The base period of the RoPE embeddings.
|
| 69 |
+
rope_scaling (`Dict`, *optional*):
|
| 70 |
+
Dictionary containing the scaling configuration for the RoPE embeddings. NOTE: if you apply new rope type
|
| 71 |
+
and you expect the model to work on longer `max_position_embeddings`, we recommend you to update this value
|
| 72 |
+
accordingly.
|
| 73 |
+
Expected contents:
|
| 74 |
+
`rope_type` (`str`):
|
| 75 |
+
The sub-variant of RoPE to use. Can be one of ['default', 'linear', 'dynamic', 'yarn', 'longrope',
|
| 76 |
+
'llama3'], with 'default' being the original RoPE implementation.
|
| 77 |
+
`factor` (`float`, *optional*):
|
| 78 |
+
Used with all rope types except 'default'. The scaling factor to apply to the RoPE embeddings. In
|
| 79 |
+
most scaling types, a `factor` of x will enable the model to handle sequences of length x *
|
| 80 |
+
original maximum pre-trained length.
|
| 81 |
+
`original_max_position_embeddings` (`int`, *optional*):
|
| 82 |
+
Used with 'dynamic', 'longrope' and 'llama3'. The original max position embeddings used during
|
| 83 |
+
pretraining.
|
| 84 |
+
`attention_factor` (`float`, *optional*):
|
| 85 |
+
Used with 'yarn' and 'longrope'. The scaling factor to be applied on the attention
|
| 86 |
+
computation. If unspecified, it defaults to value recommended by the implementation, using the
|
| 87 |
+
`factor` field to infer the suggested value.
|
| 88 |
+
`beta_fast` (`float`, *optional*):
|
| 89 |
+
Only used with 'yarn'. Parameter to set the boundary for extrapolation (only) in the linear
|
| 90 |
+
ramp function. If unspecified, it defaults to 32.
|
| 91 |
+
`beta_slow` (`float`, *optional*):
|
| 92 |
+
Only used with 'yarn'. Parameter to set the boundary for interpolation (only) in the linear
|
| 93 |
+
ramp function. If unspecified, it defaults to 1.
|
| 94 |
+
`short_factor` (`list[float]`, *optional*):
|
| 95 |
+
Only used with 'longrope'. The scaling factor to be applied to short contexts (<
|
| 96 |
+
`original_max_position_embeddings`). Must be a list of numbers with the same length as the hidden
|
| 97 |
+
size divided by the number of attention heads divided by 2
|
| 98 |
+
`long_factor` (`list[float]`, *optional*):
|
| 99 |
+
Only used with 'longrope'. The scaling factor to be applied to long contexts (<
|
| 100 |
+
`original_max_position_embeddings`). Must be a list of numbers with the same length as the hidden
|
| 101 |
+
size divided by the number of attention heads divided by 2
|
| 102 |
+
`low_freq_factor` (`float`, *optional*):
|
| 103 |
+
Only used with 'llama3'. Scaling factor applied to low frequency components of the RoPE
|
| 104 |
+
`high_freq_factor` (`float`, *optional*):
|
| 105 |
+
Only used with 'llama3'. Scaling factor applied to high frequency components of the RoPE
|
| 106 |
+
attention_bias (`bool`, defaults to `False`, *optional*, defaults to `False`):
|
| 107 |
+
Whether to use a bias in the query, key, value and output projection layers during self-attention.
|
| 108 |
+
use_sliding_window (`bool`, *optional*, defaults to `False`):
|
| 109 |
+
Whether to use sliding window attention.
|
| 110 |
+
sliding_window (`int`, *optional*, defaults to 4096):
|
| 111 |
+
Sliding window attention (SWA) window size. If not specified, will default to `4096`.
|
| 112 |
+
layer_types (`list`, *optional*):
|
| 113 |
+
Attention pattern for each layer.
|
| 114 |
+
attention_dropout (`float`, *optional*, defaults to 0.0):
|
| 115 |
+
The dropout ratio for the attention probabilities.
|
| 116 |
+
|
| 117 |
+
```python
|
| 118 |
+
>>> from acestep.models import AceStepConfig
|
| 119 |
+
|
| 120 |
+
>>> # Initializing an AceStep configuration
|
| 121 |
+
>>> configuration = AceStepConfig()
|
| 122 |
+
|
| 123 |
+
>>> # Initializing a model from the configuration
|
| 124 |
+
>>> model = AceStepConditionGenerationModel(configuration)
|
| 125 |
+
|
| 126 |
+
>>> # Accessing the model configuration
|
| 127 |
+
>>> configuration = model.config
|
| 128 |
+
```"""
|
| 129 |
+
|
| 130 |
+
model_type = "acestep"
|
| 131 |
+
keys_to_ignore_at_inference = ["past_key_values"]
|
| 132 |
+
|
| 133 |
+
# Default tensor parallel plan for the base model
|
| 134 |
+
base_model_tp_plan = {
|
| 135 |
+
"layers.*.self_attn.q_proj": "colwise",
|
| 136 |
+
"layers.*.self_attn.k_proj": "colwise",
|
| 137 |
+
"layers.*.self_attn.v_proj": "colwise",
|
| 138 |
+
"layers.*.self_attn.o_proj": "rowwise",
|
| 139 |
+
"layers.*.mlp.gate_proj": "colwise",
|
| 140 |
+
"layers.*.mlp.up_proj": "colwise",
|
| 141 |
+
"layers.*.mlp.down_proj": "rowwise",
|
| 142 |
+
}
|
| 143 |
+
base_model_pp_plan = {
|
| 144 |
+
"embed_tokens": (["input_ids"], ["inputs_embeds"]),
|
| 145 |
+
"layers": (["hidden_states", "attention_mask"], ["hidden_states"]),
|
| 146 |
+
"norm": (["hidden_states"], ["hidden_states"]),
|
| 147 |
+
}
|
| 148 |
+
def __init__(
|
| 149 |
+
self,
|
| 150 |
+
vocab_size=64003,
|
| 151 |
+
fsq_dim=2048,
|
| 152 |
+
fsq_input_levels=[8, 8, 8, 5, 5, 5],
|
| 153 |
+
fsq_input_num_quantizers=1,
|
| 154 |
+
hidden_size=2048,
|
| 155 |
+
intermediate_size=6144,
|
| 156 |
+
num_hidden_layers=24,
|
| 157 |
+
num_attention_heads=16,
|
| 158 |
+
num_key_value_heads=8,
|
| 159 |
+
head_dim=128,
|
| 160 |
+
hidden_act="silu",
|
| 161 |
+
max_position_embeddings=32768,
|
| 162 |
+
initializer_range=0.02,
|
| 163 |
+
rms_norm_eps=1e-6,
|
| 164 |
+
use_cache=True,
|
| 165 |
+
tie_word_embeddings=True,
|
| 166 |
+
rope_theta=1000000,
|
| 167 |
+
rope_scaling=None,
|
| 168 |
+
attention_bias=False,
|
| 169 |
+
use_sliding_window=True,
|
| 170 |
+
sliding_window=128,
|
| 171 |
+
layer_types=None,
|
| 172 |
+
attention_dropout=0.0,
|
| 173 |
+
num_lyric_encoder_hidden_layers=8,
|
| 174 |
+
audio_acoustic_hidden_dim=64,
|
| 175 |
+
pool_window_size=5,
|
| 176 |
+
text_hidden_dim=1024,
|
| 177 |
+
in_channels=192,
|
| 178 |
+
data_proportion=0.5,
|
| 179 |
+
timestep_mu=-0.4,
|
| 180 |
+
timestep_sigma=1.0,
|
| 181 |
+
timbre_hidden_dim=64,
|
| 182 |
+
num_timbre_encoder_hidden_layers=4,
|
| 183 |
+
timbre_fix_frame=750,
|
| 184 |
+
patch_size=2,
|
| 185 |
+
num_attention_pooler_hidden_layers=2,
|
| 186 |
+
num_audio_decoder_hidden_layers=24,
|
| 187 |
+
model_version="turbo",
|
| 188 |
+
**kwargs,
|
| 189 |
+
):
|
| 190 |
+
self.max_position_embeddings = max_position_embeddings
|
| 191 |
+
self.hidden_size = hidden_size
|
| 192 |
+
self.intermediate_size = intermediate_size
|
| 193 |
+
self.num_hidden_layers = num_hidden_layers
|
| 194 |
+
self.num_attention_heads = num_attention_heads
|
| 195 |
+
self.use_sliding_window = use_sliding_window
|
| 196 |
+
self.sliding_window = sliding_window if self.use_sliding_window else None
|
| 197 |
+
|
| 198 |
+
# Text encoder configuration
|
| 199 |
+
self.text_hidden_dim = text_hidden_dim
|
| 200 |
+
|
| 201 |
+
# Lyric encoder configuration
|
| 202 |
+
self.num_lyric_encoder_hidden_layers = num_lyric_encoder_hidden_layers
|
| 203 |
+
self.patch_size = patch_size
|
| 204 |
+
|
| 205 |
+
# Audio semantic token generation configuration
|
| 206 |
+
self.audio_acoustic_hidden_dim = audio_acoustic_hidden_dim
|
| 207 |
+
self.pool_window_size = pool_window_size
|
| 208 |
+
self.in_channels = in_channels
|
| 209 |
+
self.data_proportion = data_proportion
|
| 210 |
+
self.timestep_mu = timestep_mu
|
| 211 |
+
self.timestep_sigma = timestep_sigma
|
| 212 |
+
|
| 213 |
+
# FSQ (Finite Scalar Quantization) configuration
|
| 214 |
+
self.fsq_dim = fsq_dim
|
| 215 |
+
self.fsq_input_levels = fsq_input_levels
|
| 216 |
+
self.fsq_input_num_quantizers = fsq_input_num_quantizers
|
| 217 |
+
|
| 218 |
+
# Timbre encoder configuration
|
| 219 |
+
self.timbre_hidden_dim = timbre_hidden_dim
|
| 220 |
+
self.num_timbre_encoder_hidden_layers = num_timbre_encoder_hidden_layers
|
| 221 |
+
self.timbre_fix_frame = timbre_fix_frame
|
| 222 |
+
self.num_attention_pooler_hidden_layers = num_attention_pooler_hidden_layers
|
| 223 |
+
self.num_audio_decoder_hidden_layers = num_audio_decoder_hidden_layers
|
| 224 |
+
self.vocab_size = vocab_size
|
| 225 |
+
|
| 226 |
+
# Backward compatibility: ensure num_key_value_heads is set
|
| 227 |
+
if num_key_value_heads is None:
|
| 228 |
+
num_key_value_heads = num_attention_heads
|
| 229 |
+
|
| 230 |
+
self.num_key_value_heads = num_key_value_heads
|
| 231 |
+
self.head_dim = head_dim
|
| 232 |
+
self.hidden_act = hidden_act
|
| 233 |
+
self.initializer_range = initializer_range
|
| 234 |
+
self.rms_norm_eps = rms_norm_eps
|
| 235 |
+
self.use_cache = use_cache
|
| 236 |
+
self.rope_theta = rope_theta
|
| 237 |
+
self.rope_scaling = rope_scaling
|
| 238 |
+
self.attention_bias = attention_bias
|
| 239 |
+
self.attention_dropout = attention_dropout
|
| 240 |
+
self.model_version = model_version
|
| 241 |
+
|
| 242 |
+
# Validate rotary position embeddings parameters
|
| 243 |
+
# Backward compatibility: if there is a 'type' field, move it to 'rope_type'
|
| 244 |
+
if self.rope_scaling is not None and "type" in self.rope_scaling:
|
| 245 |
+
self.rope_scaling["rope_type"] = self.rope_scaling["type"]
|
| 246 |
+
rope_config_validation(self)
|
| 247 |
+
|
| 248 |
+
self.layer_types = layer_types
|
| 249 |
+
|
| 250 |
+
# Set default layer types if not specified
|
| 251 |
+
if self.layer_types is None:
|
| 252 |
+
self.layer_types = [
|
| 253 |
+
"sliding_attention" if bool((i + 1) % 2) else "full_attention" for i in range(self.num_hidden_layers)
|
| 254 |
+
]
|
| 255 |
+
layer_type_validation(self.layer_types)
|
| 256 |
+
|
| 257 |
+
super().__init__(
|
| 258 |
+
tie_word_embeddings=tie_word_embeddings,
|
| 259 |
+
**kwargs,
|
| 260 |
+
)
|
| 261 |
+
|
| 262 |
+
|
| 263 |
+
__all__ = ["AceStepConfig"]
|
acestep-v15-turbo/model.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:3f6e0797fad420a39bd33979eb6e840e30989e34a3794e843d23b60ec6e422d7
|
| 3 |
+
size 4787825604
|
acestep-v15-turbo/modeling_acestep_v15_turbo.py
ADDED
|
@@ -0,0 +1,2136 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Copyright 2025 The ACESTEO Team. All rights reserved.
|
| 2 |
+
#
|
| 3 |
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
| 4 |
+
# you may not use this file except in compliance with the License.
|
| 5 |
+
# You may obtain a copy of the License at
|
| 6 |
+
#
|
| 7 |
+
# http://www.apache.org/licenses/LICENSE-2.0
|
| 8 |
+
#
|
| 9 |
+
# Unless required by applicable law or agreed to in writing, software
|
| 10 |
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
| 11 |
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
| 12 |
+
# See the License for the specific language governing permissions and
|
| 13 |
+
# limitations under the License.
|
| 14 |
+
import math
|
| 15 |
+
import time
|
| 16 |
+
from typing import Callable, List, Optional, Union
|
| 17 |
+
|
| 18 |
+
import torch
|
| 19 |
+
import torch.nn.functional as F
|
| 20 |
+
from torch import nn
|
| 21 |
+
|
| 22 |
+
from einops import rearrange
|
| 23 |
+
|
| 24 |
+
# Transformers imports (sorted by submodule, then alphabetically)
|
| 25 |
+
from transformers.cache_utils import Cache, DynamicCache, EncoderDecoderCache
|
| 26 |
+
from transformers.modeling_attn_mask_utils import _prepare_4d_causal_attention_mask
|
| 27 |
+
from transformers.modeling_flash_attention_utils import FlashAttentionKwargs
|
| 28 |
+
from transformers.modeling_layers import GradientCheckpointingLayer
|
| 29 |
+
from transformers.modeling_outputs import BaseModelOutput
|
| 30 |
+
from transformers.modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
|
| 31 |
+
from transformers.processing_utils import Unpack
|
| 32 |
+
from transformers.utils import auto_docstring, can_return_tuple, logging
|
| 33 |
+
from transformers.models.qwen3.modeling_qwen3 import (
|
| 34 |
+
Qwen3MLP,
|
| 35 |
+
Qwen3RMSNorm,
|
| 36 |
+
Qwen3RotaryEmbedding,
|
| 37 |
+
apply_rotary_pos_emb,
|
| 38 |
+
eager_attention_forward,
|
| 39 |
+
)
|
| 40 |
+
|
| 41 |
+
from vector_quantize_pytorch import ResidualFSQ
|
| 42 |
+
|
| 43 |
+
# Local config import with fallback
|
| 44 |
+
try:
|
| 45 |
+
from .configuration_acestep_v15 import AceStepConfig
|
| 46 |
+
except ImportError:
|
| 47 |
+
from configuration_acestep_v15 import AceStepConfig
|
| 48 |
+
|
| 49 |
+
|
| 50 |
+
logger = logging.get_logger(__name__)
|
| 51 |
+
|
| 52 |
+
|
| 53 |
+
def create_4d_mask(
|
| 54 |
+
seq_len: int,
|
| 55 |
+
dtype: torch.dtype,
|
| 56 |
+
device: torch.device,
|
| 57 |
+
attention_mask: Optional[torch.Tensor] = None, # [Batch, Seq_Len]
|
| 58 |
+
sliding_window: Optional[int] = None,
|
| 59 |
+
is_sliding_window: bool = False,
|
| 60 |
+
is_causal: bool = True,
|
| 61 |
+
) -> torch.Tensor:
|
| 62 |
+
"""
|
| 63 |
+
General 4D Attention Mask generator compatible with CPU/Mac/SDPA and Eager mode.
|
| 64 |
+
Supports use cases:
|
| 65 |
+
1. Causal Full: is_causal=True, is_sliding_window=False (standard GPT)
|
| 66 |
+
2. Causal Sliding: is_causal=True, is_sliding_window=True (Mistral/Qwen local window)
|
| 67 |
+
3. Bidirectional Full: is_causal=False, is_sliding_window=False (BERT/Encoder)
|
| 68 |
+
4. Bidirectional Sliding: is_causal=False, is_sliding_window=True (Longformer local)
|
| 69 |
+
|
| 70 |
+
Returns:
|
| 71 |
+
[Batch, 1, Seq_Len, Seq_Len] additive mask (0.0 for keep, -inf for mask)
|
| 72 |
+
"""
|
| 73 |
+
# ------------------------------------------------------
|
| 74 |
+
# 1. Construct basic geometry mask [Seq_Len, Seq_Len]
|
| 75 |
+
# ------------------------------------------------------
|
| 76 |
+
|
| 77 |
+
# Build index matrices
|
| 78 |
+
# i (Query): [0, 1, ..., L-1]
|
| 79 |
+
# j (Key): [0, 1, ..., L-1]
|
| 80 |
+
indices = torch.arange(seq_len, device=device)
|
| 81 |
+
# diff = i - j
|
| 82 |
+
diff = indices.unsqueeze(1) - indices.unsqueeze(0)
|
| 83 |
+
|
| 84 |
+
# Initialize all True (all positions visible)
|
| 85 |
+
valid_mask = torch.ones((seq_len, seq_len), device=device, dtype=torch.bool)
|
| 86 |
+
|
| 87 |
+
# (A) Handle causality (Causal)
|
| 88 |
+
if is_causal:
|
| 89 |
+
# i >= j => diff >= 0
|
| 90 |
+
valid_mask = valid_mask & (diff >= 0)
|
| 91 |
+
|
| 92 |
+
# (B) Handle sliding window
|
| 93 |
+
if is_sliding_window and sliding_window is not None:
|
| 94 |
+
if is_causal:
|
| 95 |
+
# Causal sliding: only attend to past window steps
|
| 96 |
+
# i - j <= window => diff <= window
|
| 97 |
+
# (diff >= 0 already handled above)
|
| 98 |
+
valid_mask = valid_mask & (diff <= sliding_window)
|
| 99 |
+
else:
|
| 100 |
+
# Bidirectional sliding: attend past and future window steps
|
| 101 |
+
# |i - j| <= window => abs(diff) <= sliding_window
|
| 102 |
+
valid_mask = valid_mask & (torch.abs(diff) <= sliding_window)
|
| 103 |
+
|
| 104 |
+
# Expand dimensions to [1, 1, Seq_Len, Seq_Len] for broadcasting
|
| 105 |
+
valid_mask = valid_mask.unsqueeze(0).unsqueeze(0)
|
| 106 |
+
|
| 107 |
+
# ------------------------------------------------------
|
| 108 |
+
# 2. Apply padding mask (Key Masking)
|
| 109 |
+
# ------------------------------------------------------
|
| 110 |
+
if attention_mask is not None:
|
| 111 |
+
# attention_mask shape: [Batch, Seq_Len] (1=valid, 0=padding)
|
| 112 |
+
# We want to mask out invalid keys (columns)
|
| 113 |
+
# Expand shape: [Batch, 1, 1, Seq_Len]
|
| 114 |
+
padding_mask_4d = attention_mask.view(attention_mask.shape[0], 1, 1, seq_len).to(torch.bool)
|
| 115 |
+
|
| 116 |
+
# Broadcasting: Geometry Mask [1, 1, L, L] & Padding Mask [B, 1, 1, L]
|
| 117 |
+
# Result shape: [B, 1, L, L]
|
| 118 |
+
valid_mask = valid_mask & padding_mask_4d
|
| 119 |
+
|
| 120 |
+
# ------------------------------------------------------
|
| 121 |
+
# 3. Convert to additive mask
|
| 122 |
+
# ------------------------------------------------------
|
| 123 |
+
# Get the minimal value for current dtype
|
| 124 |
+
min_dtype = torch.finfo(dtype).min
|
| 125 |
+
|
| 126 |
+
# Create result tensor filled with -inf by default
|
| 127 |
+
mask_tensor = torch.full(valid_mask.shape, min_dtype, dtype=dtype, device=device)
|
| 128 |
+
|
| 129 |
+
# Set valid positions to 0.0
|
| 130 |
+
mask_tensor.masked_fill_(valid_mask, 0.0)
|
| 131 |
+
|
| 132 |
+
return mask_tensor
|
| 133 |
+
|
| 134 |
+
|
| 135 |
+
def pack_sequences(hidden1: torch.Tensor, hidden2: torch.Tensor, mask1: torch.Tensor, mask2: torch.Tensor):
|
| 136 |
+
"""
|
| 137 |
+
Pack two sequences by concatenating and sorting them based on mask values.
|
| 138 |
+
|
| 139 |
+
Args:
|
| 140 |
+
hidden1: First hidden states tensor of shape [B, L1, D]
|
| 141 |
+
hidden2: Second hidden states tensor of shape [B, L2, D]
|
| 142 |
+
mask1: First mask tensor of shape [B, L1]
|
| 143 |
+
mask2: Second mask tensor of shape [B, L2]
|
| 144 |
+
|
| 145 |
+
Returns:
|
| 146 |
+
Tuple of (packed_hidden_states, new_mask) where:
|
| 147 |
+
- packed_hidden_states: Packed hidden states with valid tokens (mask=1) first, shape [B, L1+L2, D]
|
| 148 |
+
- new_mask: New mask tensor indicating valid positions, shape [B, L1+L2]
|
| 149 |
+
"""
|
| 150 |
+
# Step 1: Concatenate hidden states and masks along sequence dimension
|
| 151 |
+
hidden_cat = torch.cat([hidden1, hidden2], dim=1) # [B, L, D]
|
| 152 |
+
mask_cat = torch.cat([mask1, mask2], dim=1) # [B, L]
|
| 153 |
+
|
| 154 |
+
B, L, D = hidden_cat.shape
|
| 155 |
+
|
| 156 |
+
# Step 2: Sort indices so that mask values of 1 come before 0
|
| 157 |
+
sort_idx = mask_cat.argsort(dim=1, descending=True, stable=True) # [B, L]
|
| 158 |
+
|
| 159 |
+
# Step 3: Reorder hidden states using sorted indices
|
| 160 |
+
hidden_left = torch.gather(hidden_cat, 1, sort_idx.unsqueeze(-1).expand(B, L, D))
|
| 161 |
+
|
| 162 |
+
# Step 4: Create new mask based on valid sequence lengths
|
| 163 |
+
lengths = mask_cat.sum(dim=1) # [B]
|
| 164 |
+
new_mask = (torch.arange(L, dtype=torch.long, device=hidden_cat.device).unsqueeze(0) < lengths.unsqueeze(1))
|
| 165 |
+
|
| 166 |
+
return hidden_left, new_mask
|
| 167 |
+
|
| 168 |
+
|
| 169 |
+
def sample_t_r(batch_size, device, dtype, data_proportion=0.0, timestep_mu=-0.4, timestep_sigma=1.0, use_meanflow=True):
|
| 170 |
+
"""
|
| 171 |
+
Sample timestep t and r for flow matching training.
|
| 172 |
+
|
| 173 |
+
Args:
|
| 174 |
+
batch_size: Batch size
|
| 175 |
+
device: Device to create tensors on
|
| 176 |
+
dtype: Data type for tensors
|
| 177 |
+
data_proportion: Proportion of data samples (0.0 to 1.0)
|
| 178 |
+
timestep_mu: Mean for timestep sampling
|
| 179 |
+
timestep_sigma: Standard deviation for timestep sampling
|
| 180 |
+
use_meanflow: Whether to use meanflow (if False, data_proportion is set to 1.0)
|
| 181 |
+
|
| 182 |
+
Returns:
|
| 183 |
+
Tuple of (t, r) tensors, each of shape [batch_size]
|
| 184 |
+
"""
|
| 185 |
+
t = torch.sigmoid(torch.randn((batch_size,), device=device, dtype=dtype) * timestep_sigma + timestep_mu)
|
| 186 |
+
r = torch.sigmoid(torch.randn((batch_size,), device=device, dtype=dtype) * timestep_sigma + timestep_mu)
|
| 187 |
+
# Assign t = max, r = min, for each pair
|
| 188 |
+
t, r = torch.maximum(t, r), torch.minimum(t, r)
|
| 189 |
+
if not use_meanflow:
|
| 190 |
+
data_proportion = 1.0
|
| 191 |
+
data_size = int(batch_size * data_proportion)
|
| 192 |
+
zero_mask = torch.arange(batch_size, device=device) < data_size
|
| 193 |
+
r = torch.where(zero_mask, t, r)
|
| 194 |
+
return t, r
|
| 195 |
+
|
| 196 |
+
|
| 197 |
+
class TimestepEmbedding(nn.Module):
|
| 198 |
+
"""
|
| 199 |
+
Timestep embedding module for diffusion models.
|
| 200 |
+
|
| 201 |
+
Converts timestep values into high-dimensional embeddings using sinusoidal
|
| 202 |
+
positional encoding, followed by MLP layers. Used for conditioning diffusion
|
| 203 |
+
models on timestep information.
|
| 204 |
+
"""
|
| 205 |
+
def __init__(
|
| 206 |
+
self,
|
| 207 |
+
in_channels: int,
|
| 208 |
+
time_embed_dim: int,
|
| 209 |
+
scale: float = 1000,
|
| 210 |
+
):
|
| 211 |
+
super().__init__()
|
| 212 |
+
|
| 213 |
+
self.linear_1 = nn.Linear(in_channels, time_embed_dim, bias=True)
|
| 214 |
+
self.act1 = nn.SiLU()
|
| 215 |
+
self.linear_2 = nn.Linear(time_embed_dim, time_embed_dim, bias=True)
|
| 216 |
+
self.in_channels = in_channels
|
| 217 |
+
|
| 218 |
+
self.act2 = nn.SiLU()
|
| 219 |
+
self.time_proj = nn.Linear(time_embed_dim, time_embed_dim * 6)
|
| 220 |
+
self.scale = scale
|
| 221 |
+
|
| 222 |
+
def timestep_embedding(self, t, dim, max_period=10000):
|
| 223 |
+
"""
|
| 224 |
+
Create sinusoidal timestep embeddings.
|
| 225 |
+
|
| 226 |
+
Args:
|
| 227 |
+
t: A 1-D tensor of N indices, one per batch element. These may be fractional.
|
| 228 |
+
dim: The dimension of the output embeddings.
|
| 229 |
+
max_period: Controls the minimum frequency of the embeddings.
|
| 230 |
+
|
| 231 |
+
Returns:
|
| 232 |
+
An (N, D) tensor of positional embeddings.
|
| 233 |
+
"""
|
| 234 |
+
t = t * self.scale
|
| 235 |
+
half = dim // 2
|
| 236 |
+
freqs = torch.exp(
|
| 237 |
+
-math.log(max_period) * torch.arange(start=0, end=half, dtype=torch.float32) / half
|
| 238 |
+
).to(device=t.device)
|
| 239 |
+
args = t[:, None].float() * freqs[None]
|
| 240 |
+
embedding = torch.cat([torch.cos(args), torch.sin(args)], dim=-1)
|
| 241 |
+
if dim % 2:
|
| 242 |
+
embedding = torch.cat([embedding, torch.zeros_like(embedding[:, :1])], dim=-1)
|
| 243 |
+
return embedding
|
| 244 |
+
|
| 245 |
+
def forward(self, t):
|
| 246 |
+
t_freq = self.timestep_embedding(t, self.in_channels)
|
| 247 |
+
temb = self.linear_1(t_freq.to(t.dtype))
|
| 248 |
+
temb = self.act1(temb)
|
| 249 |
+
temb = self.linear_2(temb)
|
| 250 |
+
timestep_proj = self.time_proj(self.act2(temb)).unflatten(1, (6, -1))
|
| 251 |
+
return temb, timestep_proj
|
| 252 |
+
|
| 253 |
+
class AceStepAttention(nn.Module):
|
| 254 |
+
"""
|
| 255 |
+
Multi-headed attention module for AceStep model.
|
| 256 |
+
|
| 257 |
+
Implements the attention mechanism from 'Attention Is All You Need' paper,
|
| 258 |
+
with support for both self-attention and cross-attention modes. Uses RMSNorm
|
| 259 |
+
for query and key normalization, and supports sliding window attention for
|
| 260 |
+
efficient long-sequence processing.
|
| 261 |
+
"""
|
| 262 |
+
|
| 263 |
+
def __init__(self, config: AceStepConfig, layer_idx: int, is_cross_attention: bool = False, is_causal: bool = False):
|
| 264 |
+
super().__init__()
|
| 265 |
+
self.config = config
|
| 266 |
+
self.layer_idx = layer_idx
|
| 267 |
+
self.head_dim = getattr(config, "head_dim", config.hidden_size // config.num_attention_heads)
|
| 268 |
+
self.num_key_value_groups = config.num_attention_heads // config.num_key_value_heads
|
| 269 |
+
self.scaling = self.head_dim**-0.5
|
| 270 |
+
self.attention_dropout = config.attention_dropout
|
| 271 |
+
if is_cross_attention:
|
| 272 |
+
is_causal = False
|
| 273 |
+
self.is_causal = is_causal
|
| 274 |
+
self.is_cross_attention = is_cross_attention
|
| 275 |
+
|
| 276 |
+
self.q_proj = nn.Linear(config.hidden_size, config.num_attention_heads * self.head_dim, bias=config.attention_bias)
|
| 277 |
+
self.k_proj = nn.Linear(config.hidden_size, config.num_key_value_heads * self.head_dim, bias=config.attention_bias)
|
| 278 |
+
self.v_proj = nn.Linear(config.hidden_size, config.num_key_value_heads * self.head_dim, bias=config.attention_bias)
|
| 279 |
+
self.o_proj = nn.Linear(config.num_attention_heads * self.head_dim, config.hidden_size, bias=config.attention_bias)
|
| 280 |
+
# Apply RMS normalization only on the head dimension (unlike OLMo)
|
| 281 |
+
self.q_norm = Qwen3RMSNorm(self.head_dim, eps=config.rms_norm_eps)
|
| 282 |
+
self.k_norm = Qwen3RMSNorm(self.head_dim, eps=config.rms_norm_eps)
|
| 283 |
+
self.attention_type = config.layer_types[layer_idx]
|
| 284 |
+
self.sliding_window = config.sliding_window if config.layer_types[layer_idx] == "sliding_attention" else None
|
| 285 |
+
|
| 286 |
+
def forward(
|
| 287 |
+
self,
|
| 288 |
+
hidden_states: torch.Tensor,
|
| 289 |
+
attention_mask: Optional[torch.Tensor],
|
| 290 |
+
past_key_value: Optional[Cache] = None,
|
| 291 |
+
cache_position: Optional[torch.LongTensor] = None,
|
| 292 |
+
encoder_hidden_states: Optional[torch.Tensor] = None,
|
| 293 |
+
position_embeddings: tuple[torch.Tensor, torch.Tensor] = None,
|
| 294 |
+
output_attentions: Optional[bool] = False,
|
| 295 |
+
**kwargs: Unpack[FlashAttentionKwargs],
|
| 296 |
+
) -> tuple[torch.Tensor, Optional[torch.Tensor], Optional[tuple[torch.Tensor]]]:
|
| 297 |
+
input_shape = hidden_states.shape[:-1]
|
| 298 |
+
hidden_shape = (*input_shape, -1, self.head_dim)
|
| 299 |
+
|
| 300 |
+
# Project and normalize query states
|
| 301 |
+
query_states = self.q_norm(self.q_proj(hidden_states).view(hidden_shape)).transpose(1, 2)
|
| 302 |
+
|
| 303 |
+
# Determine if this is cross-attention (requires encoder_hidden_states)
|
| 304 |
+
is_cross_attention = self.is_cross_attention and encoder_hidden_states is not None
|
| 305 |
+
|
| 306 |
+
# Cross-attention path: attend to encoder hidden states
|
| 307 |
+
if is_cross_attention:
|
| 308 |
+
encoder_hidden_shape = (*encoder_hidden_states.shape[:-1], -1, self.head_dim)
|
| 309 |
+
if past_key_value is not None:
|
| 310 |
+
is_updated = past_key_value.is_updated.get(self.layer_idx)
|
| 311 |
+
# After the first generated token, we can reuse all key/value states from cache
|
| 312 |
+
curr_past_key_value = past_key_value.cross_attention_cache
|
| 313 |
+
|
| 314 |
+
# Conditions for calculating key and value states
|
| 315 |
+
if not is_updated:
|
| 316 |
+
# Compute and cache K/V for the first time
|
| 317 |
+
key_states = self.k_norm(self.k_proj(encoder_hidden_states).view(encoder_hidden_shape)).transpose(1, 2)
|
| 318 |
+
value_states = self.v_proj(encoder_hidden_states).view(encoder_hidden_shape).transpose(1, 2)
|
| 319 |
+
# Update cache: save all key/value states to cache for fast auto-regressive generation
|
| 320 |
+
key_states, value_states = curr_past_key_value.update(key_states, value_states, self.layer_idx)
|
| 321 |
+
# Set flag that this layer's cross-attention cache is updated
|
| 322 |
+
past_key_value.is_updated[self.layer_idx] = True
|
| 323 |
+
else:
|
| 324 |
+
# Reuse cached key/value states for subsequent tokens
|
| 325 |
+
key_states = curr_past_key_value.layers[self.layer_idx].keys
|
| 326 |
+
value_states = curr_past_key_value.layers[self.layer_idx].values
|
| 327 |
+
else:
|
| 328 |
+
# No cache used, compute K/V directly
|
| 329 |
+
key_states = self.k_norm(self.k_proj(encoder_hidden_states).view(encoder_hidden_shape)).transpose(1, 2)
|
| 330 |
+
value_states = self.v_proj(encoder_hidden_states).view(encoder_hidden_shape).transpose(1, 2)
|
| 331 |
+
|
| 332 |
+
# Self-attention path: attend to the same sequence
|
| 333 |
+
else:
|
| 334 |
+
# Project and normalize key/value states for self-attention
|
| 335 |
+
key_states = self.k_norm(self.k_proj(hidden_states).view(hidden_shape)).transpose(1, 2)
|
| 336 |
+
value_states = self.v_proj(hidden_states).view(hidden_shape).transpose(1, 2)
|
| 337 |
+
# Apply rotary position embeddings (RoPE) if provided
|
| 338 |
+
if position_embeddings is not None:
|
| 339 |
+
cos, sin = position_embeddings
|
| 340 |
+
query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)
|
| 341 |
+
|
| 342 |
+
# Update cache for auto-regressive generation
|
| 343 |
+
if past_key_value is not None:
|
| 344 |
+
# Sin and cos are specific to RoPE models; cache_position needed for the static cache
|
| 345 |
+
cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}
|
| 346 |
+
key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
|
| 347 |
+
|
| 348 |
+
attention_interface: Callable = eager_attention_forward
|
| 349 |
+
if is_cross_attention and output_attentions:
|
| 350 |
+
attention_interface: Callable = eager_attention_forward
|
| 351 |
+
elif self.config._attn_implementation != "eager":
|
| 352 |
+
attention_interface = ALL_ATTENTION_FUNCTIONS[self.config._attn_implementation]
|
| 353 |
+
|
| 354 |
+
attn_output, attn_weights = attention_interface(
|
| 355 |
+
self,
|
| 356 |
+
query_states,
|
| 357 |
+
key_states,
|
| 358 |
+
value_states,
|
| 359 |
+
attention_mask,
|
| 360 |
+
dropout=self.attention_dropout if self.training else 0.0,
|
| 361 |
+
scaling=self.scaling,
|
| 362 |
+
sliding_window=self.sliding_window if not self.is_cross_attention else None,
|
| 363 |
+
**kwargs,
|
| 364 |
+
)
|
| 365 |
+
|
| 366 |
+
attn_output = attn_output.reshape(*input_shape, -1).contiguous()
|
| 367 |
+
attn_output = self.o_proj(attn_output)
|
| 368 |
+
return attn_output, attn_weights
|
| 369 |
+
|
| 370 |
+
|
| 371 |
+
class AceStepEncoderLayer(GradientCheckpointingLayer):
|
| 372 |
+
"""
|
| 373 |
+
Encoder layer for AceStep model.
|
| 374 |
+
|
| 375 |
+
Consists of self-attention and MLP (feed-forward) sub-layers with residual connections.
|
| 376 |
+
"""
|
| 377 |
+
|
| 378 |
+
def __init__(self, config, layer_idx: int):
|
| 379 |
+
super().__init__()
|
| 380 |
+
self.hidden_size = config.hidden_size
|
| 381 |
+
self.config = config
|
| 382 |
+
self.layer_idx = layer_idx
|
| 383 |
+
|
| 384 |
+
# Self-attention sub-layer
|
| 385 |
+
self.self_attn = AceStepAttention(
|
| 386 |
+
config=config,
|
| 387 |
+
layer_idx=layer_idx,
|
| 388 |
+
is_cross_attention=False,
|
| 389 |
+
is_causal=False,
|
| 390 |
+
)
|
| 391 |
+
self.input_layernorm = Qwen3RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
|
| 392 |
+
self.post_attention_layernorm = Qwen3RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
|
| 393 |
+
|
| 394 |
+
# MLP (feed-forward) sub-layer
|
| 395 |
+
self.mlp = Qwen3MLP(config)
|
| 396 |
+
self.attention_type = config.layer_types[layer_idx]
|
| 397 |
+
|
| 398 |
+
def forward(
|
| 399 |
+
self,
|
| 400 |
+
hidden_states: torch.Tensor,
|
| 401 |
+
position_embeddings: tuple[torch.Tensor, torch.Tensor],
|
| 402 |
+
attention_mask: Optional[torch.Tensor] = None,
|
| 403 |
+
position_ids: Optional[torch.LongTensor] = None,
|
| 404 |
+
output_attentions: Optional[bool] = False,
|
| 405 |
+
**kwargs,
|
| 406 |
+
) -> tuple[
|
| 407 |
+
torch.FloatTensor,
|
| 408 |
+
Optional[tuple[torch.FloatTensor, torch.FloatTensor]],
|
| 409 |
+
]:
|
| 410 |
+
# Self-attention with residual connection
|
| 411 |
+
residual = hidden_states
|
| 412 |
+
hidden_states = self.input_layernorm(hidden_states)
|
| 413 |
+
hidden_states, self_attn_weights = self.self_attn(
|
| 414 |
+
hidden_states=hidden_states,
|
| 415 |
+
position_embeddings=position_embeddings,
|
| 416 |
+
attention_mask=attention_mask,
|
| 417 |
+
position_ids=position_ids,
|
| 418 |
+
output_attentions=output_attentions,
|
| 419 |
+
# Encoders don't use cache
|
| 420 |
+
use_cache=False,
|
| 421 |
+
past_key_value=None,
|
| 422 |
+
**kwargs,
|
| 423 |
+
)
|
| 424 |
+
hidden_states = residual + hidden_states
|
| 425 |
+
|
| 426 |
+
# MLP with residual connection
|
| 427 |
+
residual = hidden_states
|
| 428 |
+
hidden_states = self.post_attention_layernorm(hidden_states)
|
| 429 |
+
hidden_states = self.mlp(hidden_states)
|
| 430 |
+
hidden_states = residual + hidden_states
|
| 431 |
+
|
| 432 |
+
outputs = (hidden_states,)
|
| 433 |
+
|
| 434 |
+
if output_attentions:
|
| 435 |
+
outputs += (self_attn_weights,)
|
| 436 |
+
|
| 437 |
+
return outputs
|
| 438 |
+
|
| 439 |
+
|
| 440 |
+
class AceStepDiTLayer(GradientCheckpointingLayer):
|
| 441 |
+
"""
|
| 442 |
+
DiT (Diffusion Transformer) layer for AceStep model.
|
| 443 |
+
|
| 444 |
+
Implements a transformer layer with three main components:
|
| 445 |
+
1. Self-attention with adaptive layer norm (AdaLN)
|
| 446 |
+
2. Cross-attention (optional) for conditioning on encoder outputs
|
| 447 |
+
3. Feed-forward MLP with adaptive layer norm
|
| 448 |
+
|
| 449 |
+
Uses scale-shift modulation from timestep embeddings for adaptive normalization.
|
| 450 |
+
"""
|
| 451 |
+
def __init__(self, config: AceStepConfig, layer_idx: int, use_cross_attention: bool = True):
|
| 452 |
+
super().__init__()
|
| 453 |
+
|
| 454 |
+
# 1. Self-attention sub-layer with adaptive normalization
|
| 455 |
+
self.self_attn_norm = Qwen3RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
|
| 456 |
+
self.self_attn = AceStepAttention(config=config, layer_idx=layer_idx)
|
| 457 |
+
|
| 458 |
+
# 2. Cross-attention sub-layer (optional, for encoder conditioning)
|
| 459 |
+
self.use_cross_attention = use_cross_attention
|
| 460 |
+
if self.use_cross_attention:
|
| 461 |
+
self.cross_attn_norm = Qwen3RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
|
| 462 |
+
self.cross_attn = AceStepAttention(config=config, layer_idx=layer_idx, is_cross_attention=True)
|
| 463 |
+
|
| 464 |
+
# 3. Feed-forward MLP sub-layer with adaptive normalization
|
| 465 |
+
self.mlp_norm = Qwen3RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
|
| 466 |
+
self.mlp = Qwen3MLP(config)
|
| 467 |
+
|
| 468 |
+
# Scale-shift table for adaptive layer norm modulation (6 values: 3 for self-attn, 3 for MLP)
|
| 469 |
+
self.scale_shift_table = nn.Parameter(torch.randn(1, 6, config.hidden_size) / config.hidden_size**0.5)
|
| 470 |
+
self.attention_type = config.layer_types[layer_idx]
|
| 471 |
+
|
| 472 |
+
def forward(
|
| 473 |
+
self,
|
| 474 |
+
hidden_states: torch.Tensor,
|
| 475 |
+
position_embeddings: tuple[torch.Tensor, torch.Tensor],
|
| 476 |
+
temb: torch.Tensor,
|
| 477 |
+
attention_mask: Optional[torch.Tensor] = None,
|
| 478 |
+
position_ids: Optional[torch.LongTensor] = None,
|
| 479 |
+
past_key_value: Optional[EncoderDecoderCache] = None,
|
| 480 |
+
output_attentions: Optional[bool] = False,
|
| 481 |
+
use_cache: Optional[bool] = False,
|
| 482 |
+
cache_position: Optional[torch.LongTensor] = None,
|
| 483 |
+
encoder_hidden_states: Optional[torch.Tensor] = None,
|
| 484 |
+
encoder_attention_mask: Optional[torch.Tensor] = None,
|
| 485 |
+
**kwargs,
|
| 486 |
+
) -> torch.Tensor:
|
| 487 |
+
|
| 488 |
+
# Extract scale-shift parameters for adaptive layer norm from timestep embeddings
|
| 489 |
+
# 6 values: (shift_msa, scale_msa, gate_msa, c_shift_msa, c_scale_msa, c_gate_msa)
|
| 490 |
+
shift_msa, scale_msa, gate_msa, c_shift_msa, c_scale_msa, c_gate_msa = (
|
| 491 |
+
self.scale_shift_table + temb
|
| 492 |
+
).chunk(6, dim=1)
|
| 493 |
+
|
| 494 |
+
# Step 1: Self-attention with adaptive layer norm (AdaLN)
|
| 495 |
+
# Apply adaptive normalization: norm(x) * (1 + scale) + shift
|
| 496 |
+
norm_hidden_states = (self.self_attn_norm(hidden_states) * (1 + scale_msa) + shift_msa).type_as(hidden_states)
|
| 497 |
+
attn_output, self_attn_weights = self.self_attn(
|
| 498 |
+
hidden_states=norm_hidden_states,
|
| 499 |
+
position_embeddings=position_embeddings,
|
| 500 |
+
attention_mask=attention_mask,
|
| 501 |
+
position_ids=position_ids,
|
| 502 |
+
output_attentions=output_attentions,
|
| 503 |
+
use_cache=False,
|
| 504 |
+
past_key_value=None,
|
| 505 |
+
**kwargs,
|
| 506 |
+
)
|
| 507 |
+
# Apply gated residual connection: x = x + attn_output * gate
|
| 508 |
+
hidden_states = (hidden_states + attn_output * gate_msa).type_as(hidden_states)
|
| 509 |
+
|
| 510 |
+
# Step 2: Cross-attention (if enabled) for conditioning on encoder outputs
|
| 511 |
+
if self.use_cross_attention:
|
| 512 |
+
norm_hidden_states = self.cross_attn_norm(hidden_states).type_as(hidden_states)
|
| 513 |
+
attn_output, cross_attn_weights = self.cross_attn(
|
| 514 |
+
hidden_states=norm_hidden_states,
|
| 515 |
+
encoder_hidden_states=encoder_hidden_states,
|
| 516 |
+
attention_mask=encoder_attention_mask,
|
| 517 |
+
past_key_value=past_key_value,
|
| 518 |
+
output_attentions=output_attentions,
|
| 519 |
+
use_cache=use_cache,
|
| 520 |
+
**kwargs,
|
| 521 |
+
)
|
| 522 |
+
# Standard residual connection for cross-attention
|
| 523 |
+
hidden_states = hidden_states + attn_output
|
| 524 |
+
|
| 525 |
+
# Step 3: Feed-forward (MLP) with adaptive layer norm
|
| 526 |
+
# Apply adaptive normalization for MLP: norm(x) * (1 + scale) + shift
|
| 527 |
+
norm_hidden_states = (self.mlp_norm(hidden_states) * (1 + c_scale_msa) + c_shift_msa).type_as(hidden_states)
|
| 528 |
+
ff_output = self.mlp(norm_hidden_states)
|
| 529 |
+
# Apply gated residual connection: x = x + mlp_output * gate
|
| 530 |
+
hidden_states = (hidden_states + ff_output * c_gate_msa).type_as(hidden_states)
|
| 531 |
+
|
| 532 |
+
outputs = (hidden_states,)
|
| 533 |
+
if output_attentions:
|
| 534 |
+
outputs += (self_attn_weights, cross_attn_weights)
|
| 535 |
+
|
| 536 |
+
return outputs
|
| 537 |
+
|
| 538 |
+
|
| 539 |
+
@auto_docstring
|
| 540 |
+
class AceStepPreTrainedModel(PreTrainedModel):
|
| 541 |
+
config_class = AceStepConfig
|
| 542 |
+
base_model_prefix = "model"
|
| 543 |
+
supports_gradient_checkpointing = True
|
| 544 |
+
_no_split_modules = ["AceStepEncoderLayer", "AceStepDiTLayer"]
|
| 545 |
+
_skip_keys_device_placement = ["past_key_values"]
|
| 546 |
+
_supports_flash_attn_3 = True
|
| 547 |
+
_supports_flash_attn_2 = True
|
| 548 |
+
_supports_sdpa = True
|
| 549 |
+
_supports_flex_attn = True
|
| 550 |
+
_supports_cache_class = True
|
| 551 |
+
_supports_quantized_cache = True
|
| 552 |
+
_supports_static_cache = True
|
| 553 |
+
_supports_attention_backend = True
|
| 554 |
+
|
| 555 |
+
def _init_weights(self, module):
|
| 556 |
+
"""
|
| 557 |
+
Initialize weights for different module types.
|
| 558 |
+
|
| 559 |
+
TODO: Support separate initialization for encoders and decoders.
|
| 560 |
+
"""
|
| 561 |
+
std = self.config.initializer_range
|
| 562 |
+
if isinstance(module, nn.Linear):
|
| 563 |
+
module.weight.data.normal_(mean=0.0, std=std)
|
| 564 |
+
if module.bias is not None:
|
| 565 |
+
module.bias.data.zero_()
|
| 566 |
+
elif isinstance(module, nn.Embedding):
|
| 567 |
+
module.weight.data.normal_(mean=0.0, std=std)
|
| 568 |
+
if module.padding_idx is not None:
|
| 569 |
+
module.weight.data[module.padding_idx].zero_()
|
| 570 |
+
elif isinstance(module, Qwen3RMSNorm):
|
| 571 |
+
module.weight.data.fill_(1.0)
|
| 572 |
+
|
| 573 |
+
|
| 574 |
+
class AceStepLyricEncoder(AceStepPreTrainedModel):
|
| 575 |
+
"""
|
| 576 |
+
Encoder for processing lyric text embeddings.
|
| 577 |
+
|
| 578 |
+
Encodes lyric text hidden states using a transformer encoder architecture
|
| 579 |
+
with bidirectional attention. Projects text embeddings to model hidden size
|
| 580 |
+
and processes them through multiple encoder layers.
|
| 581 |
+
"""
|
| 582 |
+
def __init__(self, config):
|
| 583 |
+
super().__init__(config)
|
| 584 |
+
|
| 585 |
+
# Project text embeddings to model hidden size
|
| 586 |
+
self.embed_tokens = nn.Linear(config.text_hidden_dim, config.hidden_size)
|
| 587 |
+
self.norm = Qwen3RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
|
| 588 |
+
self.rotary_emb = Qwen3RotaryEmbedding(config=config)
|
| 589 |
+
self.gradient_checkpointing = False
|
| 590 |
+
|
| 591 |
+
# Stack of encoder layers
|
| 592 |
+
self.layers = nn.ModuleList(
|
| 593 |
+
[AceStepEncoderLayer(config, layer_idx) for layer_idx in range(config.num_lyric_encoder_hidden_layers)]
|
| 594 |
+
)
|
| 595 |
+
|
| 596 |
+
# Initialize weights and apply final processing
|
| 597 |
+
self.post_init()
|
| 598 |
+
|
| 599 |
+
@can_return_tuple
|
| 600 |
+
def forward(
|
| 601 |
+
self,
|
| 602 |
+
input_ids: Optional[torch.LongTensor] = None,
|
| 603 |
+
attention_mask: Optional[torch.Tensor] = None,
|
| 604 |
+
position_ids: Optional[torch.LongTensor] = None,
|
| 605 |
+
inputs_embeds: Optional[torch.FloatTensor] = None,
|
| 606 |
+
output_attentions: Optional[bool] = None,
|
| 607 |
+
output_hidden_states: Optional[bool] = None,
|
| 608 |
+
**flash_attn_kwargs: Unpack[FlashAttentionKwargs],
|
| 609 |
+
) -> BaseModelOutput:
|
| 610 |
+
output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
|
| 611 |
+
output_hidden_states = (
|
| 612 |
+
output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
|
| 613 |
+
)
|
| 614 |
+
|
| 615 |
+
assert input_ids is None, "Only `input_ids` is supported for the lyric encoder."
|
| 616 |
+
assert attention_mask is not None, "Attention mask must be provided for the lyric encoder."
|
| 617 |
+
assert inputs_embeds is not None, "Inputs embeddings must be provided for the lyric encoder."
|
| 618 |
+
|
| 619 |
+
# Project input embeddings: N x T x text_hidden_dim -> N x T x hidden_size
|
| 620 |
+
inputs_embeds = self.embed_tokens(inputs_embeds)
|
| 621 |
+
# Cache position: only used for mask construction (not for actual caching)
|
| 622 |
+
cache_position = torch.arange(0, inputs_embeds.shape[1], device=inputs_embeds.device)
|
| 623 |
+
|
| 624 |
+
# Positional IDs
|
| 625 |
+
if position_ids is None:
|
| 626 |
+
position_ids = cache_position.unsqueeze(0)
|
| 627 |
+
|
| 628 |
+
# Attention masks
|
| 629 |
+
seq_len = inputs_embeds.shape[1]
|
| 630 |
+
dtype = inputs_embeds.dtype
|
| 631 |
+
device = inputs_embeds.device
|
| 632 |
+
|
| 633 |
+
# 判断是否使用 Flash Attention 2
|
| 634 |
+
is_flash_attn = (self.config._attn_implementation == "flash_attention_2")
|
| 635 |
+
|
| 636 |
+
# 初始化 Mask 变量
|
| 637 |
+
full_attn_mask = None
|
| 638 |
+
sliding_attn_mask = None
|
| 639 |
+
|
| 640 |
+
if is_flash_attn:
|
| 641 |
+
# -------------------------------------------------------
|
| 642 |
+
# 场景 A: Flash Attention 模式
|
| 643 |
+
# -------------------------------------------------------
|
| 644 |
+
# FA 不需要 4D Mask。
|
| 645 |
+
# 如果有 padding mask (attention_mask [B, L])，直接传给它即可。
|
| 646 |
+
# 如果没有 padding mask，传 None。
|
| 647 |
+
# 滑动窗口逻辑由 Layer 内部传给 FA kernel 的 sliding_window 参数控制。
|
| 648 |
+
full_attn_mask = attention_mask
|
| 649 |
+
|
| 650 |
+
# 这里的逻辑是：如果配置启用了滑动窗口，FA 模式下我们也只需要传基础的 padding mask
|
| 651 |
+
# Layer 会自己决定是否调用带 sliding window 的 kernel
|
| 652 |
+
sliding_attn_mask = attention_mask if self.config.use_sliding_window else None
|
| 653 |
+
|
| 654 |
+
else:
|
| 655 |
+
# -------------------------------------------------------
|
| 656 |
+
# 场景 B: CPU / Mac / SDPA (Eager 模式)
|
| 657 |
+
# -------------------------------------------------------
|
| 658 |
+
# 必须手动生成 4D Mask [B, 1, L, L]
|
| 659 |
+
|
| 660 |
+
# 1. Full Attention (Bidirectional, Global)
|
| 661 |
+
# 对应原来的 create_causal_mask + bidirectional
|
| 662 |
+
full_attn_mask = create_4d_mask(
|
| 663 |
+
seq_len=seq_len,
|
| 664 |
+
dtype=dtype,
|
| 665 |
+
device=device,
|
| 666 |
+
attention_mask=attention_mask, # [B, L]
|
| 667 |
+
sliding_window=None,
|
| 668 |
+
is_sliding_window=False,
|
| 669 |
+
is_causal=False # <--- 关键：双向注意力
|
| 670 |
+
)
|
| 671 |
+
|
| 672 |
+
# 2. Sliding Attention (Bidirectional, Local)
|
| 673 |
+
# 对应原来的 create_sliding_window... + bidirectional
|
| 674 |
+
if self.config.use_sliding_window:
|
| 675 |
+
sliding_attn_mask = create_4d_mask(
|
| 676 |
+
seq_len=seq_len,
|
| 677 |
+
dtype=dtype,
|
| 678 |
+
device=device,
|
| 679 |
+
attention_mask=attention_mask, # [B, L]
|
| 680 |
+
sliding_window=self.config.sliding_window,
|
| 681 |
+
is_sliding_window=True, # <--- 开启滑动窗口
|
| 682 |
+
is_causal=False # <--- 关键：双向注意力
|
| 683 |
+
)
|
| 684 |
+
|
| 685 |
+
# 构建 Mapping
|
| 686 |
+
self_attn_mask_mapping = {
|
| 687 |
+
"full_attention": full_attn_mask,
|
| 688 |
+
"sliding_attention": sliding_attn_mask,
|
| 689 |
+
}
|
| 690 |
+
|
| 691 |
+
# Initialize hidden states with input embeddings
|
| 692 |
+
hidden_states = inputs_embeds
|
| 693 |
+
|
| 694 |
+
# Create position embeddings to be shared across all layers
|
| 695 |
+
position_embeddings = self.rotary_emb(hidden_states, position_ids)
|
| 696 |
+
|
| 697 |
+
# Pass through transformer layers
|
| 698 |
+
all_hidden_states = () if output_hidden_states else None
|
| 699 |
+
all_self_attns = () if output_attentions else None
|
| 700 |
+
|
| 701 |
+
for layer_module in self.layers[: self.config.num_hidden_layers]:
|
| 702 |
+
if output_hidden_states:
|
| 703 |
+
all_hidden_states += (hidden_states,)
|
| 704 |
+
|
| 705 |
+
layer_outputs = layer_module(
|
| 706 |
+
hidden_states,
|
| 707 |
+
position_embeddings,
|
| 708 |
+
self_attn_mask_mapping[layer_module.attention_type],
|
| 709 |
+
position_ids,
|
| 710 |
+
output_attentions,
|
| 711 |
+
**flash_attn_kwargs,
|
| 712 |
+
)
|
| 713 |
+
|
| 714 |
+
hidden_states = layer_outputs[0]
|
| 715 |
+
|
| 716 |
+
if output_attentions:
|
| 717 |
+
all_self_attns += (layer_outputs[1],)
|
| 718 |
+
|
| 719 |
+
hidden_states = self.norm(hidden_states)
|
| 720 |
+
|
| 721 |
+
if output_hidden_states:
|
| 722 |
+
all_hidden_states += (hidden_states,)
|
| 723 |
+
|
| 724 |
+
return BaseModelOutput(
|
| 725 |
+
last_hidden_state=hidden_states,
|
| 726 |
+
hidden_states=all_hidden_states,
|
| 727 |
+
attentions=all_self_attns,
|
| 728 |
+
)
|
| 729 |
+
|
| 730 |
+
|
| 731 |
+
class AttentionPooler(AceStepPreTrainedModel):
|
| 732 |
+
"""
|
| 733 |
+
Attention-based pooling module.
|
| 734 |
+
|
| 735 |
+
Pools sequences of patches using a special token and attention mechanism.
|
| 736 |
+
The special token attends to all patches and its output is used as the
|
| 737 |
+
pooled representation. Used for aggregating patch-level features into
|
| 738 |
+
sequence-level representations.
|
| 739 |
+
"""
|
| 740 |
+
def __init__(self, config):
|
| 741 |
+
super().__init__(config)
|
| 742 |
+
self.config = config
|
| 743 |
+
self.embed_tokens = nn.Linear(config.hidden_size, config.hidden_size)
|
| 744 |
+
self.norm = Qwen3RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
|
| 745 |
+
self.rotary_emb = Qwen3RotaryEmbedding(config=config)
|
| 746 |
+
self.gradient_checkpointing = False
|
| 747 |
+
# Special token used for pooling (CLS-like token)
|
| 748 |
+
self.special_token = nn.Parameter(torch.randn(1, 1, config.hidden_size) * 0.02)
|
| 749 |
+
self.layers = nn.ModuleList(
|
| 750 |
+
[AceStepEncoderLayer(config, layer_idx) for layer_idx in range(config.num_attention_pooler_hidden_layers)]
|
| 751 |
+
)
|
| 752 |
+
|
| 753 |
+
# Initialize weights and apply final processing
|
| 754 |
+
self.post_init()
|
| 755 |
+
|
| 756 |
+
@can_return_tuple
|
| 757 |
+
def forward(self,
|
| 758 |
+
x,
|
| 759 |
+
attention_mask: Optional[torch.Tensor] = None,
|
| 760 |
+
**flash_attn_kwargs: Unpack[FlashAttentionKwargs],
|
| 761 |
+
) -> BaseModelOutput:
|
| 762 |
+
B, T, P, D = x.shape
|
| 763 |
+
x = self.embed_tokens(x)
|
| 764 |
+
special_tokens = self.special_token.expand(B, T, 1, -1)
|
| 765 |
+
x = torch.cat([special_tokens, x], dim=2)
|
| 766 |
+
x = rearrange(x, "b t p c -> (b t) p c")
|
| 767 |
+
|
| 768 |
+
# Cache position: only used for mask construction.
|
| 769 |
+
cache_position = torch.arange(0, x.shape[1], device=x.device)
|
| 770 |
+
# Postional ids.
|
| 771 |
+
position_ids = cache_position.unsqueeze(0)
|
| 772 |
+
|
| 773 |
+
# embed positions
|
| 774 |
+
hidden_states = x
|
| 775 |
+
|
| 776 |
+
# create position embeddings to be shared across the decoder layers
|
| 777 |
+
position_embeddings = self.rotary_emb(hidden_states, position_ids)
|
| 778 |
+
|
| 779 |
+
seq_len = x.shape[1]
|
| 780 |
+
dtype = x.dtype
|
| 781 |
+
device = x.device
|
| 782 |
+
|
| 783 |
+
# 判断是否使用 Flash Attention 2
|
| 784 |
+
is_flash_attn = (self.config._attn_implementation == "flash_attention_2")
|
| 785 |
+
|
| 786 |
+
# 初始化 Mask 变量
|
| 787 |
+
full_attn_mask = None
|
| 788 |
+
sliding_attn_mask = None
|
| 789 |
+
|
| 790 |
+
if is_flash_attn:
|
| 791 |
+
# -------------------------------------------------------
|
| 792 |
+
# 场景 A: Flash Attention 模式
|
| 793 |
+
# -------------------------------------------------------
|
| 794 |
+
# FA 不需要 4D Mask。
|
| 795 |
+
# 如果有 padding mask (attention_mask [B, L])，直接传给它即可。
|
| 796 |
+
# 如果没有 padding mask，传 None。
|
| 797 |
+
# 滑动窗口逻辑由 Layer 内部传给 FA kernel 的 sliding_window 参数控制。
|
| 798 |
+
full_attn_mask = attention_mask
|
| 799 |
+
|
| 800 |
+
# 这里的逻辑是：如果配置启用了滑动窗口，FA 模式下我们也只需要传基础的 padding mask
|
| 801 |
+
# Layer 会自己决定是否调用带 sliding window 的 kernel
|
| 802 |
+
sliding_attn_mask = attention_mask if self.config.use_sliding_window else None
|
| 803 |
+
|
| 804 |
+
else:
|
| 805 |
+
# -------------------------------------------------------
|
| 806 |
+
# 场景 B: CPU / Mac / SDPA (Eager 模式)
|
| 807 |
+
# -------------------------------------------------------
|
| 808 |
+
# 必须手动生成 4D Mask [B, 1, L, L]
|
| 809 |
+
|
| 810 |
+
# 1. Full Attention (Bidirectional, Global)
|
| 811 |
+
# 对应原来的 create_causal_mask + bidirectional
|
| 812 |
+
full_attn_mask = create_4d_mask(
|
| 813 |
+
seq_len=seq_len,
|
| 814 |
+
dtype=dtype,
|
| 815 |
+
device=device,
|
| 816 |
+
attention_mask=attention_mask, # [B, L]
|
| 817 |
+
sliding_window=None,
|
| 818 |
+
is_sliding_window=False,
|
| 819 |
+
is_causal=False # <--- 关键：双向注意力
|
| 820 |
+
)
|
| 821 |
+
|
| 822 |
+
# 2. Sliding Attention (Bidirectional, Local)
|
| 823 |
+
# 对应原来的 create_sliding_window... + bidirectional
|
| 824 |
+
if self.config.use_sliding_window:
|
| 825 |
+
sliding_attn_mask = create_4d_mask(
|
| 826 |
+
seq_len=seq_len,
|
| 827 |
+
dtype=dtype,
|
| 828 |
+
device=device,
|
| 829 |
+
attention_mask=attention_mask, # [B, L]
|
| 830 |
+
sliding_window=self.config.sliding_window,
|
| 831 |
+
is_sliding_window=True, # <--- 开启滑动窗口
|
| 832 |
+
is_causal=False # <--- 关键：双向注意力
|
| 833 |
+
)
|
| 834 |
+
|
| 835 |
+
# 构建 Mapping
|
| 836 |
+
self_attn_mask_mapping = {
|
| 837 |
+
"full_attention": full_attn_mask,
|
| 838 |
+
"sliding_attention": sliding_attn_mask,
|
| 839 |
+
}
|
| 840 |
+
|
| 841 |
+
for layer_module in self.layers:
|
| 842 |
+
layer_outputs = layer_module(
|
| 843 |
+
hidden_states,
|
| 844 |
+
position_embeddings,
|
| 845 |
+
attention_mask=self_attn_mask_mapping[layer_module.attention_type],
|
| 846 |
+
**flash_attn_kwargs,
|
| 847 |
+
)
|
| 848 |
+
|
| 849 |
+
hidden_states = layer_outputs[0]
|
| 850 |
+
|
| 851 |
+
hidden_states = self.norm(hidden_states)
|
| 852 |
+
|
| 853 |
+
# Extract the special token output (first position) as pooled representation
|
| 854 |
+
cls_output = hidden_states[:, 0, :]
|
| 855 |
+
cls_output = rearrange(cls_output, "(b t) c -> b t c", b=B)
|
| 856 |
+
return cls_output
|
| 857 |
+
|
| 858 |
+
|
| 859 |
+
class AudioTokenDetokenizer(AceStepPreTrainedModel):
|
| 860 |
+
"""
|
| 861 |
+
Audio token detokenizer module.
|
| 862 |
+
|
| 863 |
+
Converts quantized audio tokens back to continuous acoustic representations.
|
| 864 |
+
Expands each token into multiple patches using special tokens, processes them
|
| 865 |
+
through encoder layers, and projects to acoustic hidden dimension.
|
| 866 |
+
"""
|
| 867 |
+
def __init__(self, config):
|
| 868 |
+
super().__init__(config)
|
| 869 |
+
self.config = config
|
| 870 |
+
self.embed_tokens = nn.Linear(config.hidden_size, config.hidden_size)
|
| 871 |
+
self.norm = Qwen3RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
|
| 872 |
+
self.rotary_emb = Qwen3RotaryEmbedding(config=config)
|
| 873 |
+
self.gradient_checkpointing = False
|
| 874 |
+
# Special tokens for expanding each quantized token into patches
|
| 875 |
+
self.special_tokens = nn.Parameter(torch.randn(1, config.pool_window_size, config.hidden_size) * 0.02)
|
| 876 |
+
self.layers = nn.ModuleList(
|
| 877 |
+
[AceStepEncoderLayer(config, layer_idx) for layer_idx in range(config.num_attention_pooler_hidden_layers)]
|
| 878 |
+
)
|
| 879 |
+
# Project back to acoustic hidden dimension
|
| 880 |
+
self.proj_out = nn.Linear(config.hidden_size, config.audio_acoustic_hidden_dim)
|
| 881 |
+
|
| 882 |
+
# Initialize weights and apply final processing
|
| 883 |
+
self.post_init()
|
| 884 |
+
|
| 885 |
+
@can_return_tuple
|
| 886 |
+
def forward(self,
|
| 887 |
+
x,
|
| 888 |
+
attention_mask: Optional[torch.Tensor] = None,
|
| 889 |
+
**flash_attn_kwargs: Unpack[FlashAttentionKwargs],
|
| 890 |
+
) -> BaseModelOutput:
|
| 891 |
+
B, T, D = x.shape
|
| 892 |
+
x = self.embed_tokens(x)
|
| 893 |
+
# Expand and add special tokens: N x T x D -> N x T x P x D
|
| 894 |
+
# Each token is expanded into pool_window_size patches
|
| 895 |
+
x = x.unsqueeze(2) # N x T x 1 x D
|
| 896 |
+
x = x.repeat(1, 1, self.config.pool_window_size, 1) # N x T x P x D
|
| 897 |
+
# Add learnable special tokens to each patch
|
| 898 |
+
special_tokens = self.special_tokens.expand(B, T, -1, -1)
|
| 899 |
+
x = x + special_tokens
|
| 900 |
+
# Reshape for processing: (batch * time) x patches x hidden
|
| 901 |
+
x = rearrange(x, "b t p c -> (b t) p c")
|
| 902 |
+
|
| 903 |
+
# Cache position: only used for mask construction
|
| 904 |
+
cache_position = torch.arange(0, x.shape[1], device=x.device)
|
| 905 |
+
# Positional IDs
|
| 906 |
+
position_ids = cache_position.unsqueeze(0)
|
| 907 |
+
|
| 908 |
+
# Initialize hidden states
|
| 909 |
+
hidden_states = x
|
| 910 |
+
|
| 911 |
+
# Create position embeddings to be shared across all layers
|
| 912 |
+
position_embeddings = self.rotary_emb(hidden_states, position_ids)
|
| 913 |
+
|
| 914 |
+
seq_len = x.shape[1]
|
| 915 |
+
dtype = x.dtype
|
| 916 |
+
device = x.device
|
| 917 |
+
|
| 918 |
+
# 判断是否使用 Flash Attention 2
|
| 919 |
+
is_flash_attn = (self.config._attn_implementation == "flash_attention_2")
|
| 920 |
+
|
| 921 |
+
# 初始化 Mask 变量
|
| 922 |
+
full_attn_mask = None
|
| 923 |
+
sliding_attn_mask = None
|
| 924 |
+
|
| 925 |
+
if is_flash_attn:
|
| 926 |
+
# -------------------------------------------------------
|
| 927 |
+
# 场景 A: Flash Attention 模式
|
| 928 |
+
# -------------------------------------------------------
|
| 929 |
+
# FA 不需要 4D Mask。
|
| 930 |
+
# 如果有 padding mask (attention_mask [B, L])，直接传给它即可。
|
| 931 |
+
# 如果没有 padding mask，传 None。
|
| 932 |
+
# 滑动窗口逻辑由 Layer 内部传给 FA kernel 的 sliding_window 参数控制。
|
| 933 |
+
full_attn_mask = attention_mask
|
| 934 |
+
|
| 935 |
+
# 这里的逻辑是：如果配置启用了滑动窗口，FA 模式下我们也只需要传基础的 padding mask
|
| 936 |
+
# Layer 会自己决定是否调用带 sliding window 的 kernel
|
| 937 |
+
sliding_attn_mask = attention_mask if self.config.use_sliding_window else None
|
| 938 |
+
|
| 939 |
+
else:
|
| 940 |
+
# -------------------------------------------------------
|
| 941 |
+
# 场景 B: CPU / Mac / SDPA (Eager 模式)
|
| 942 |
+
# -------------------------------------------------------
|
| 943 |
+
# 必须手动生成 4D Mask [B, 1, L, L]
|
| 944 |
+
|
| 945 |
+
# 1. Full Attention (Bidirectional, Global)
|
| 946 |
+
# 对应原来的 create_causal_mask + bidirectional
|
| 947 |
+
full_attn_mask = create_4d_mask(
|
| 948 |
+
seq_len=seq_len,
|
| 949 |
+
dtype=dtype,
|
| 950 |
+
device=device,
|
| 951 |
+
attention_mask=attention_mask, # [B, L]
|
| 952 |
+
sliding_window=None,
|
| 953 |
+
is_sliding_window=False,
|
| 954 |
+
is_causal=False # <--- 关键：双向注意力
|
| 955 |
+
)
|
| 956 |
+
|
| 957 |
+
# 2. Sliding Attention (Bidirectional, Local)
|
| 958 |
+
# 对应原来的 create_sliding_window... + bidirectional
|
| 959 |
+
if self.config.use_sliding_window:
|
| 960 |
+
sliding_attn_mask = create_4d_mask(
|
| 961 |
+
seq_len=seq_len,
|
| 962 |
+
dtype=dtype,
|
| 963 |
+
device=device,
|
| 964 |
+
attention_mask=attention_mask, # [B, L]
|
| 965 |
+
sliding_window=self.config.sliding_window,
|
| 966 |
+
is_sliding_window=True, # <--- 开启滑动窗口
|
| 967 |
+
is_causal=False # <--- 关键：双向注意力
|
| 968 |
+
)
|
| 969 |
+
|
| 970 |
+
# 构建 Mapping
|
| 971 |
+
self_attn_mask_mapping = {
|
| 972 |
+
"full_attention": full_attn_mask,
|
| 973 |
+
"sliding_attention": sliding_attn_mask,
|
| 974 |
+
}
|
| 975 |
+
|
| 976 |
+
for layer_module in self.layers:
|
| 977 |
+
layer_outputs = layer_module(
|
| 978 |
+
hidden_states,
|
| 979 |
+
position_embeddings,
|
| 980 |
+
attention_mask=self_attn_mask_mapping[layer_module.attention_type],
|
| 981 |
+
**flash_attn_kwargs,
|
| 982 |
+
)
|
| 983 |
+
|
| 984 |
+
hidden_states = layer_outputs[0]
|
| 985 |
+
|
| 986 |
+
hidden_states = self.norm(hidden_states)
|
| 987 |
+
|
| 988 |
+
hidden_states = self.proj_out(hidden_states)
|
| 989 |
+
|
| 990 |
+
hidden_states = rearrange(hidden_states, "(b t) p c -> b (t p) c", b=B, p=self.config.pool_window_size)
|
| 991 |
+
return hidden_states
|
| 992 |
+
|
| 993 |
+
|
| 994 |
+
class AceStepTimbreEncoder(AceStepPreTrainedModel):
|
| 995 |
+
"""
|
| 996 |
+
Encoder for extracting timbre embeddings from reference audio.
|
| 997 |
+
|
| 998 |
+
Processes packed reference audio acoustic features to extract timbre
|
| 999 |
+
representations. Uses a special token (CLS-like) to aggregate information
|
| 1000 |
+
from the entire reference audio sequence. Outputs are unpacked back to
|
| 1001 |
+
batch format for use in conditioning.
|
| 1002 |
+
"""
|
| 1003 |
+
def __init__(self, config):
|
| 1004 |
+
super().__init__(config)
|
| 1005 |
+
|
| 1006 |
+
# Project acoustic features to model hidden size
|
| 1007 |
+
self.embed_tokens = nn.Linear(config.timbre_hidden_dim, config.hidden_size)
|
| 1008 |
+
self.norm = Qwen3RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
|
| 1009 |
+
self.rotary_emb = Qwen3RotaryEmbedding(config=config)
|
| 1010 |
+
self.gradient_checkpointing = False
|
| 1011 |
+
# Special token for aggregating timbre information (prepended to sequence)
|
| 1012 |
+
self.special_token = nn.Parameter(torch.randn(1, 1, config.hidden_size))
|
| 1013 |
+
self.layers = nn.ModuleList(
|
| 1014 |
+
[AceStepEncoderLayer(config, layer_idx) for layer_idx in range(config.num_timbre_encoder_hidden_layers)]
|
| 1015 |
+
)
|
| 1016 |
+
|
| 1017 |
+
# Initialize weights and apply final processing
|
| 1018 |
+
self.post_init()
|
| 1019 |
+
|
| 1020 |
+
def unpack_timbre_embeddings(self, timbre_embs_packed, refer_audio_order_mask):
|
| 1021 |
+
"""
|
| 1022 |
+
Unpack packed timbre embeddings into batch format.
|
| 1023 |
+
|
| 1024 |
+
Args:
|
| 1025 |
+
timbre_embs_packed: Packed timbre embeddings of shape [N, d]
|
| 1026 |
+
refer_audio_order_mask: Order mask indicating batch assignment for each packed embedding
|
| 1027 |
+
|
| 1028 |
+
Returns:
|
| 1029 |
+
Tuple of (unpacked_embeddings, mask):
|
| 1030 |
+
- unpacked_embeddings: Unpacked embeddings of shape [B, max_count, d]
|
| 1031 |
+
- new_mask: Mask indicating valid positions, shape [B, max_count]
|
| 1032 |
+
"""
|
| 1033 |
+
N, d = timbre_embs_packed.shape
|
| 1034 |
+
device = timbre_embs_packed.device
|
| 1035 |
+
dtype = timbre_embs_packed.dtype
|
| 1036 |
+
|
| 1037 |
+
# Get batch size
|
| 1038 |
+
B = int(refer_audio_order_mask.max().item() + 1)
|
| 1039 |
+
|
| 1040 |
+
# Calculate element count and positions for each batch
|
| 1041 |
+
counts = torch.bincount(refer_audio_order_mask, minlength=B)
|
| 1042 |
+
max_count = counts.max().item()
|
| 1043 |
+
|
| 1044 |
+
# Calculate positions within batch
|
| 1045 |
+
sorted_indices = torch.argsort(refer_audio_order_mask * N + torch.arange(N, device=device), stable=True)
|
| 1046 |
+
sorted_batch_ids = refer_audio_order_mask[sorted_indices]
|
| 1047 |
+
|
| 1048 |
+
positions = torch.arange(N, device=device)
|
| 1049 |
+
batch_starts = torch.cat([torch.tensor([0], device=device),
|
| 1050 |
+
torch.cumsum(counts, dim=0)[:-1]])
|
| 1051 |
+
positions_in_sorted = positions - batch_starts[sorted_batch_ids]
|
| 1052 |
+
|
| 1053 |
+
inverse_indices = torch.empty_like(sorted_indices)
|
| 1054 |
+
inverse_indices[sorted_indices] = torch.arange(N, device=device)
|
| 1055 |
+
positions_in_batch = positions_in_sorted[inverse_indices]
|
| 1056 |
+
|
| 1057 |
+
# Use one-hot encoding and matrix multiplication (gradient-friendly approach)
|
| 1058 |
+
# Create one-hot encoding
|
| 1059 |
+
indices_2d = refer_audio_order_mask * max_count + positions_in_batch # (N,)
|
| 1060 |
+
one_hot = F.one_hot(indices_2d, num_classes=B * max_count).to(dtype) # (N, B*max_count)
|
| 1061 |
+
|
| 1062 |
+
# Rearrange using matrix multiplication
|
| 1063 |
+
timbre_embs_flat = one_hot.t() @ timbre_embs_packed # (B*max_count, d)
|
| 1064 |
+
timbre_embs_unpack = timbre_embs_flat.reshape(B, max_count, d)
|
| 1065 |
+
|
| 1066 |
+
# Create mask indicating valid positions
|
| 1067 |
+
mask_flat = (one_hot.sum(dim=0) > 0).long() # (B*max_count,)
|
| 1068 |
+
new_mask = mask_flat.reshape(B, max_count)
|
| 1069 |
+
|
| 1070 |
+
return timbre_embs_unpack, new_mask
|
| 1071 |
+
|
| 1072 |
+
@can_return_tuple
|
| 1073 |
+
def forward(
|
| 1074 |
+
self,
|
| 1075 |
+
refer_audio_acoustic_hidden_states_packed: Optional[torch.FloatTensor] = None,
|
| 1076 |
+
refer_audio_order_mask: Optional[torch.LongTensor] = None,
|
| 1077 |
+
attention_mask: Optional[torch.Tensor] = None,
|
| 1078 |
+
**flash_attn_kwargs: Unpack[FlashAttentionKwargs],
|
| 1079 |
+
) -> BaseModelOutput:
|
| 1080 |
+
inputs_embeds = refer_audio_acoustic_hidden_states_packed
|
| 1081 |
+
# Project embeddings: N x T x timbre_hidden_dim -> N x T x hidden_size
|
| 1082 |
+
inputs_embeds = self.embed_tokens(inputs_embeds)
|
| 1083 |
+
# Prepend special token for timbre aggregation (CLS-like token)
|
| 1084 |
+
# inputs_embeds = torch.cat([self.special_token.expand(inputs_embeds.shape[0], 1, -1), inputs_embeds], dim=1)
|
| 1085 |
+
# Cache position: only used for mask construction (not for actual caching)
|
| 1086 |
+
cache_position = torch.arange(0, inputs_embeds.shape[1], device=inputs_embeds.device)
|
| 1087 |
+
# Positional IDs
|
| 1088 |
+
position_ids = cache_position.unsqueeze(0)
|
| 1089 |
+
|
| 1090 |
+
seq_len = inputs_embeds.shape[1]
|
| 1091 |
+
dtype = inputs_embeds.dtype
|
| 1092 |
+
device = inputs_embeds.device
|
| 1093 |
+
|
| 1094 |
+
# 判断是否使用 Flash Attention 2
|
| 1095 |
+
is_flash_attn = (self.config._attn_implementation == "flash_attention_2")
|
| 1096 |
+
|
| 1097 |
+
# 初始化 Mask 变量
|
| 1098 |
+
full_attn_mask = None
|
| 1099 |
+
sliding_attn_mask = None
|
| 1100 |
+
|
| 1101 |
+
if is_flash_attn:
|
| 1102 |
+
# -------------------------------------------------------
|
| 1103 |
+
# 场景 A: Flash Attention 模式
|
| 1104 |
+
# -------------------------------------------------------
|
| 1105 |
+
# FA 不需要 4D Mask。
|
| 1106 |
+
# 如果有 padding mask (attention_mask [B, L])，直接传给它即可。
|
| 1107 |
+
# 如果没有 padding mask，传 None。
|
| 1108 |
+
# 滑动窗口逻辑由 Layer 内部传给 FA kernel 的 sliding_window 参数控制。
|
| 1109 |
+
full_attn_mask = attention_mask
|
| 1110 |
+
|
| 1111 |
+
# 这里的逻辑是：如果配置启用了滑动窗口，FA 模式下我们也只需要传基础的 padding mask
|
| 1112 |
+
# Layer 会自己决定是否调用带 sliding window 的 kernel
|
| 1113 |
+
sliding_attn_mask = attention_mask if self.config.use_sliding_window else None
|
| 1114 |
+
|
| 1115 |
+
else:
|
| 1116 |
+
# -------------------------------------------------------
|
| 1117 |
+
# 场景 B: CPU / Mac / SDPA (Eager 模式)
|
| 1118 |
+
# -------------------------------------------------------
|
| 1119 |
+
# 必须手动生成 4D Mask [B, 1, L, L]
|
| 1120 |
+
|
| 1121 |
+
# 1. Full Attention (Bidirectional, Global)
|
| 1122 |
+
# 对应原来的 create_causal_mask + bidirectional
|
| 1123 |
+
full_attn_mask = create_4d_mask(
|
| 1124 |
+
seq_len=seq_len,
|
| 1125 |
+
dtype=dtype,
|
| 1126 |
+
device=device,
|
| 1127 |
+
attention_mask=attention_mask, # [B, L]
|
| 1128 |
+
sliding_window=None,
|
| 1129 |
+
is_sliding_window=False,
|
| 1130 |
+
is_causal=False # <--- 关键：双向注意力
|
| 1131 |
+
)
|
| 1132 |
+
|
| 1133 |
+
# 2. Sliding Attention (Bidirectional, Local)
|
| 1134 |
+
# 对应原来的 create_sliding_window... + bidirectional
|
| 1135 |
+
if self.config.use_sliding_window:
|
| 1136 |
+
sliding_attn_mask = create_4d_mask(
|
| 1137 |
+
seq_len=seq_len,
|
| 1138 |
+
dtype=dtype,
|
| 1139 |
+
device=device,
|
| 1140 |
+
attention_mask=attention_mask, # [B, L]
|
| 1141 |
+
sliding_window=self.config.sliding_window,
|
| 1142 |
+
is_sliding_window=True, # <--- 开启滑动窗口
|
| 1143 |
+
is_causal=False # <--- 关键：双向注意力
|
| 1144 |
+
)
|
| 1145 |
+
|
| 1146 |
+
# 构建 Mapping
|
| 1147 |
+
self_attn_mask_mapping = {
|
| 1148 |
+
"full_attention": full_attn_mask,
|
| 1149 |
+
"sliding_attention": sliding_attn_mask,
|
| 1150 |
+
}
|
| 1151 |
+
|
| 1152 |
+
# Initialize hidden states
|
| 1153 |
+
hidden_states = inputs_embeds
|
| 1154 |
+
|
| 1155 |
+
# Create position embeddings to be shared across all layers
|
| 1156 |
+
position_embeddings = self.rotary_emb(hidden_states, position_ids)
|
| 1157 |
+
|
| 1158 |
+
# Pass through transformer layers
|
| 1159 |
+
for layer_module in self.layers[: self.config.num_hidden_layers]:
|
| 1160 |
+
layer_outputs = layer_module(
|
| 1161 |
+
hidden_states,
|
| 1162 |
+
position_embeddings,
|
| 1163 |
+
self_attn_mask_mapping[layer_module.attention_type],
|
| 1164 |
+
position_ids,
|
| 1165 |
+
**flash_attn_kwargs,
|
| 1166 |
+
)
|
| 1167 |
+
|
| 1168 |
+
hidden_states = layer_outputs[0]
|
| 1169 |
+
|
| 1170 |
+
hidden_states = self.norm(hidden_states)
|
| 1171 |
+
# Extract special token output (first position) as timbre embedding: N x T x D -> N x D
|
| 1172 |
+
hidden_states = hidden_states[:, 0, :]
|
| 1173 |
+
# Unpack packed embeddings back to batch format
|
| 1174 |
+
timbre_embs_unpack, timbre_embs_mask = self.unpack_timbre_embeddings(hidden_states, refer_audio_order_mask)
|
| 1175 |
+
return timbre_embs_unpack, timbre_embs_mask
|
| 1176 |
+
|
| 1177 |
+
|
| 1178 |
+
class AceStepAudioTokenizer(AceStepPreTrainedModel):
|
| 1179 |
+
"""
|
| 1180 |
+
Audio tokenizer module.
|
| 1181 |
+
|
| 1182 |
+
Converts continuous acoustic features into discrete quantized tokens.
|
| 1183 |
+
Process: project -> pool patches -> quantize. Used for converting audio
|
| 1184 |
+
representations into discrete tokens for processing by the diffusion model.
|
| 1185 |
+
"""
|
| 1186 |
+
def __init__(self, config):
|
| 1187 |
+
super().__init__(config)
|
| 1188 |
+
# Project acoustic features to hidden size
|
| 1189 |
+
self.audio_acoustic_proj = nn.Linear(config.audio_acoustic_hidden_dim, config.hidden_size)
|
| 1190 |
+
# Pool patches into sequence-level representations
|
| 1191 |
+
self.attention_pooler = AttentionPooler(config)
|
| 1192 |
+
# Quantize continuous representations into discrete tokens
|
| 1193 |
+
self.quantizer = ResidualFSQ(
|
| 1194 |
+
dim=config.fsq_dim,
|
| 1195 |
+
levels=config.fsq_input_levels,
|
| 1196 |
+
num_quantizers=config.fsq_input_num_quantizers
|
| 1197 |
+
)
|
| 1198 |
+
self.pool_window_size = config.pool_window_size
|
| 1199 |
+
# Initialize weights and apply final processing
|
| 1200 |
+
self.post_init()
|
| 1201 |
+
|
| 1202 |
+
@can_return_tuple
|
| 1203 |
+
def forward(
|
| 1204 |
+
self,
|
| 1205 |
+
hidden_states: Optional[torch.FloatTensor] = None,
|
| 1206 |
+
**flash_attn_kwargs: Unpack[FlashAttentionKwargs],
|
| 1207 |
+
) -> BaseModelOutput:
|
| 1208 |
+
|
| 1209 |
+
# Project acoustic features to hidden size
|
| 1210 |
+
hidden_states = self.audio_acoustic_proj(hidden_states)
|
| 1211 |
+
# Pool sequences: N x T//pool_window_size x pool_window_size x d -> N x T//pool_window_size x d
|
| 1212 |
+
hidden_states = self.attention_pooler(hidden_states)
|
| 1213 |
+
# Quantize continuous representations into discrete tokens: N x T//pool_window_size x d
|
| 1214 |
+
quantized, indices = self.quantizer(hidden_states)
|
| 1215 |
+
return quantized, indices
|
| 1216 |
+
|
| 1217 |
+
def tokenize(self, x):
|
| 1218 |
+
x = rearrange(x, 'n (t_patch p) d -> n t_patch p d', p=self.pool_window_size)
|
| 1219 |
+
quantized, indices = self.forward(x)
|
| 1220 |
+
return quantized, indices
|
| 1221 |
+
|
| 1222 |
+
class Lambda(nn.Module):
|
| 1223 |
+
"""
|
| 1224 |
+
Wrapper module for arbitrary lambda functions.
|
| 1225 |
+
|
| 1226 |
+
Allows using lambda functions in nn.Sequential by wrapping them in a Module.
|
| 1227 |
+
Useful for simple transformations like transpose operations.
|
| 1228 |
+
"""
|
| 1229 |
+
def __init__(self, func):
|
| 1230 |
+
super().__init__()
|
| 1231 |
+
self.func = func
|
| 1232 |
+
|
| 1233 |
+
def forward(self, x):
|
| 1234 |
+
return self.func(x)
|
| 1235 |
+
|
| 1236 |
+
|
| 1237 |
+
class AceStepDiTModel(AceStepPreTrainedModel):
|
| 1238 |
+
"""
|
| 1239 |
+
DiT (Diffusion Transformer) model for AceStep.
|
| 1240 |
+
|
| 1241 |
+
Main diffusion model that generates audio latents conditioned on text, lyrics,
|
| 1242 |
+
and timbre. Uses patch-based processing with transformer layers, timestep
|
| 1243 |
+
conditioning, and cross-attention to encoder outputs.
|
| 1244 |
+
"""
|
| 1245 |
+
def __init__(self, config: AceStepConfig):
|
| 1246 |
+
super().__init__(config)
|
| 1247 |
+
# Rotary position embeddings for transformer layers
|
| 1248 |
+
self.rotary_emb = Qwen3RotaryEmbedding(config)
|
| 1249 |
+
# Stack of DiT transformer layers
|
| 1250 |
+
self.layers = nn.ModuleList(
|
| 1251 |
+
[AceStepDiTLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]
|
| 1252 |
+
)
|
| 1253 |
+
|
| 1254 |
+
in_channels = config.in_channels
|
| 1255 |
+
inner_dim = config.hidden_size
|
| 1256 |
+
patch_size = config.patch_size
|
| 1257 |
+
self.patch_size = patch_size
|
| 1258 |
+
|
| 1259 |
+
# Input projection: patch embedding using 1D convolution
|
| 1260 |
+
# Converts sequence into patches for efficient processing
|
| 1261 |
+
self.proj_in = nn.Sequential(
|
| 1262 |
+
Lambda(lambda x: x.transpose(1, 2)), # [B, T, C] -> [B, C, T]
|
| 1263 |
+
nn.Conv1d(
|
| 1264 |
+
in_channels=in_channels,
|
| 1265 |
+
out_channels=inner_dim,
|
| 1266 |
+
kernel_size=patch_size,
|
| 1267 |
+
stride=patch_size,
|
| 1268 |
+
padding=0,
|
| 1269 |
+
),
|
| 1270 |
+
Lambda(lambda x: x.transpose(1, 2)), # [B, C, T//patch_size] -> [B, T//patch_size, C]
|
| 1271 |
+
)
|
| 1272 |
+
|
| 1273 |
+
# Timestep embeddings for diffusion conditioning
|
| 1274 |
+
# Two embeddings: one for timestep t, one for timestep difference (t - r)
|
| 1275 |
+
self.time_embed = TimestepEmbedding(in_channels=256, time_embed_dim=inner_dim)
|
| 1276 |
+
self.time_embed_r = TimestepEmbedding(in_channels=256, time_embed_dim=inner_dim)
|
| 1277 |
+
|
| 1278 |
+
# Project encoder hidden states to model dimension
|
| 1279 |
+
self.condition_embedder = nn.Linear(inner_dim, inner_dim, bias=True)
|
| 1280 |
+
|
| 1281 |
+
# Output normalization and projection
|
| 1282 |
+
# Adaptive layer norm with scale-shift modulation, then de-patchify
|
| 1283 |
+
self.norm_out = Qwen3RMSNorm(inner_dim, eps=config.rms_norm_eps)
|
| 1284 |
+
self.proj_out = nn.Sequential(
|
| 1285 |
+
Lambda(lambda x: x.transpose(1, 2)), # [B, T//patch_size, inner_dim] -> [B, inner_dim, T//patch_size]
|
| 1286 |
+
nn.ConvTranspose1d(
|
| 1287 |
+
in_channels=inner_dim,
|
| 1288 |
+
out_channels=config.audio_acoustic_hidden_dim,
|
| 1289 |
+
kernel_size=patch_size,
|
| 1290 |
+
stride=patch_size,
|
| 1291 |
+
padding=0,
|
| 1292 |
+
),
|
| 1293 |
+
Lambda(lambda x: x.transpose(1, 2)), # [B, out_channels, T] -> [B, T, out_channels]
|
| 1294 |
+
)
|
| 1295 |
+
# Scale-shift table for adaptive output normalization (2 values: shift, scale)
|
| 1296 |
+
self.scale_shift_table = nn.Parameter(torch.randn(1, 2, inner_dim) / inner_dim**0.5)
|
| 1297 |
+
|
| 1298 |
+
self.gradient_checkpointing = False
|
| 1299 |
+
|
| 1300 |
+
def forward(
|
| 1301 |
+
self,
|
| 1302 |
+
hidden_states: torch.Tensor,
|
| 1303 |
+
timestep: torch.Tensor,
|
| 1304 |
+
timestep_r: torch.Tensor,
|
| 1305 |
+
attention_mask: torch.Tensor,
|
| 1306 |
+
encoder_hidden_states: torch.Tensor,
|
| 1307 |
+
encoder_attention_mask: torch.Tensor,
|
| 1308 |
+
context_latents: torch.Tensor,
|
| 1309 |
+
use_cache: Optional[bool] = None,
|
| 1310 |
+
past_key_values: Optional[EncoderDecoderCache] = None,
|
| 1311 |
+
cache_position: Optional[torch.LongTensor] = None,
|
| 1312 |
+
position_ids: Optional[torch.LongTensor] = None,
|
| 1313 |
+
output_attentions: Optional[bool] = False,
|
| 1314 |
+
return_hidden_states: int = None,
|
| 1315 |
+
custom_layers_config: Optional[dict] = None,
|
| 1316 |
+
enable_early_exit: bool = False,
|
| 1317 |
+
**flash_attn_kwargs: Unpack[FlashAttentionKwargs],
|
| 1318 |
+
):
|
| 1319 |
+
|
| 1320 |
+
use_cache = use_cache if use_cache is not None else self.config.use_cache
|
| 1321 |
+
|
| 1322 |
+
# Disable cache during training or when gradient checkpointing is enabled
|
| 1323 |
+
if self.gradient_checkpointing and self.training and use_cache:
|
| 1324 |
+
logger.warning_once(
|
| 1325 |
+
"`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`."
|
| 1326 |
+
)
|
| 1327 |
+
use_cache = False
|
| 1328 |
+
if self.training:
|
| 1329 |
+
use_cache = False
|
| 1330 |
+
|
| 1331 |
+
# Initialize cache if needed (only during inference for auto-regressive generation)
|
| 1332 |
+
if not self.training and use_cache and past_key_values is None:
|
| 1333 |
+
past_key_values = EncoderDecoderCache(DynamicCache(), DynamicCache())
|
| 1334 |
+
|
| 1335 |
+
# Compute timestep embeddings for diffusion conditioning
|
| 1336 |
+
# Two embeddings: one for timestep t, one for timestep difference (t - r)
|
| 1337 |
+
temb_t, timestep_proj_t = self.time_embed(timestep)
|
| 1338 |
+
temb_r, timestep_proj_r = self.time_embed_r(timestep - timestep_r)
|
| 1339 |
+
# Combine embeddings
|
| 1340 |
+
temb = temb_t + temb_r
|
| 1341 |
+
timestep_proj = timestep_proj_t + timestep_proj_r
|
| 1342 |
+
|
| 1343 |
+
# Concatenate context latents (source latents + chunk masks) with hidden states
|
| 1344 |
+
hidden_states = torch.cat([context_latents, hidden_states], dim=-1)
|
| 1345 |
+
# Record original sequence length for later restoration after padding
|
| 1346 |
+
original_seq_len = hidden_states.shape[1]
|
| 1347 |
+
# Apply padding if sequence length is not divisible by patch_size
|
| 1348 |
+
# This ensures proper patch extraction
|
| 1349 |
+
pad_length = 0
|
| 1350 |
+
if hidden_states.shape[1] % self.patch_size != 0:
|
| 1351 |
+
pad_length = self.patch_size - (hidden_states.shape[1] % self.patch_size)
|
| 1352 |
+
hidden_states = F.pad(hidden_states, (0, 0, 0, pad_length), mode='constant', value=0)
|
| 1353 |
+
|
| 1354 |
+
# Project input to patches and project encoder states
|
| 1355 |
+
hidden_states = self.proj_in(hidden_states)
|
| 1356 |
+
encoder_hidden_states = self.condition_embedder(encoder_hidden_states)
|
| 1357 |
+
|
| 1358 |
+
# Cache positions
|
| 1359 |
+
if cache_position is None:
|
| 1360 |
+
past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
|
| 1361 |
+
cache_position = torch.arange(
|
| 1362 |
+
past_seen_tokens, past_seen_tokens + hidden_states.shape[1], device=hidden_states.device
|
| 1363 |
+
)
|
| 1364 |
+
|
| 1365 |
+
# Position IDs
|
| 1366 |
+
if position_ids is None:
|
| 1367 |
+
position_ids = cache_position.unsqueeze(0)
|
| 1368 |
+
|
| 1369 |
+
|
| 1370 |
+
seq_len = hidden_states.shape[1]
|
| 1371 |
+
encoder_seq_len = encoder_hidden_states.shape[1]
|
| 1372 |
+
dtype = hidden_states.dtype
|
| 1373 |
+
device = hidden_states.device
|
| 1374 |
+
|
| 1375 |
+
# 判断是否使用 Flash Attention 2
|
| 1376 |
+
is_flash_attn = (self.config._attn_implementation == "flash_attention_2")
|
| 1377 |
+
|
| 1378 |
+
# 初始化 Mask 变量
|
| 1379 |
+
full_attn_mask = None
|
| 1380 |
+
sliding_attn_mask = None
|
| 1381 |
+
encoder_attention_mask = None
|
| 1382 |
+
attention_mask = None
|
| 1383 |
+
if is_flash_attn:
|
| 1384 |
+
# -------------------------------------------------------
|
| 1385 |
+
# 场景 A: Flash Attention 模式
|
| 1386 |
+
# -------------------------------------------------------
|
| 1387 |
+
# FA 不需要 4D Mask。
|
| 1388 |
+
# 如果有 padding mask (attention_mask [B, L])，直接传给它即可。
|
| 1389 |
+
# 如果没有 padding mask，传 None。
|
| 1390 |
+
# 滑动窗口逻辑由 Layer 内部传给 FA kernel 的 sliding_window 参数控制。
|
| 1391 |
+
full_attn_mask = attention_mask
|
| 1392 |
+
|
| 1393 |
+
# 这里的逻辑是：如果配置启用了滑动窗口，FA 模式下我们也只需要传基础的 padding mask
|
| 1394 |
+
# Layer 会自己决定是否调用带 sliding window 的 kernel
|
| 1395 |
+
sliding_attn_mask = attention_mask if self.config.use_sliding_window else None
|
| 1396 |
+
|
| 1397 |
+
else:
|
| 1398 |
+
# -------------------------------------------------------
|
| 1399 |
+
# 场景 B: CPU / Mac / SDPA (Eager 模式)
|
| 1400 |
+
# -------------------------------------------------------
|
| 1401 |
+
# 必须手动生成 4D Mask [B, 1, L, L]
|
| 1402 |
+
|
| 1403 |
+
# 1. Full Attention (Bidirectional, Global)
|
| 1404 |
+
# 对应原来的 create_causal_mask + bidirectional
|
| 1405 |
+
full_attn_mask = create_4d_mask(
|
| 1406 |
+
seq_len=seq_len,
|
| 1407 |
+
dtype=dtype,
|
| 1408 |
+
device=device,
|
| 1409 |
+
attention_mask=attention_mask, # [B, L]
|
| 1410 |
+
sliding_window=None,
|
| 1411 |
+
is_sliding_window=False,
|
| 1412 |
+
is_causal=False # <--- 关键：双向注意力
|
| 1413 |
+
)
|
| 1414 |
+
max_len = max(seq_len, encoder_seq_len)
|
| 1415 |
+
|
| 1416 |
+
encoder_attention_mask = create_4d_mask(
|
| 1417 |
+
seq_len=max_len,
|
| 1418 |
+
dtype=dtype,
|
| 1419 |
+
device=device,
|
| 1420 |
+
attention_mask=attention_mask, # [B, L]
|
| 1421 |
+
sliding_window=None,
|
| 1422 |
+
is_sliding_window=False,
|
| 1423 |
+
is_causal=False # <--- 关键：双向注意力
|
| 1424 |
+
)
|
| 1425 |
+
encoder_attention_mask = encoder_attention_mask[:, :, :seq_len, :encoder_seq_len]
|
| 1426 |
+
# 2. Sliding Attention (Bidirectional, Local)
|
| 1427 |
+
# 对应原来的 create_sliding_window... + bidirectional
|
| 1428 |
+
if self.config.use_sliding_window:
|
| 1429 |
+
sliding_attn_mask = create_4d_mask(
|
| 1430 |
+
seq_len=seq_len,
|
| 1431 |
+
dtype=dtype,
|
| 1432 |
+
device=device,
|
| 1433 |
+
attention_mask=attention_mask, # [B, L]
|
| 1434 |
+
sliding_window=self.config.sliding_window,
|
| 1435 |
+
is_sliding_window=True, # <--- 开启滑动窗口
|
| 1436 |
+
is_causal=False # <--- 关键：双向注意力
|
| 1437 |
+
)
|
| 1438 |
+
|
| 1439 |
+
# 构建 Mapping
|
| 1440 |
+
self_attn_mask_mapping = {
|
| 1441 |
+
"full_attention": full_attn_mask,
|
| 1442 |
+
"sliding_attention": sliding_attn_mask,
|
| 1443 |
+
"encoder_attention_mask": encoder_attention_mask,
|
| 1444 |
+
}
|
| 1445 |
+
|
| 1446 |
+
# Create position embeddings to be shared across all decoder layers
|
| 1447 |
+
position_embeddings = self.rotary_emb(hidden_states, position_ids)
|
| 1448 |
+
all_cross_attentions = () if output_attentions else None
|
| 1449 |
+
|
| 1450 |
+
# Handle early exit for custom layer configurations
|
| 1451 |
+
max_needed_layer = float('inf')
|
| 1452 |
+
if custom_layers_config is not None and enable_early_exit:
|
| 1453 |
+
max_needed_layer = max(custom_layers_config.keys())
|
| 1454 |
+
# Force output_attentions to True when early exit is enabled for attention extraction
|
| 1455 |
+
output_attentions = True
|
| 1456 |
+
if all_cross_attentions is None:
|
| 1457 |
+
all_cross_attentions = ()
|
| 1458 |
+
|
| 1459 |
+
# Process through transformer layers
|
| 1460 |
+
for index_block, layer_module in enumerate(self.layers):
|
| 1461 |
+
|
| 1462 |
+
layer_outputs = layer_module(
|
| 1463 |
+
hidden_states,
|
| 1464 |
+
position_embeddings,
|
| 1465 |
+
timestep_proj,
|
| 1466 |
+
self_attn_mask_mapping[layer_module.attention_type],
|
| 1467 |
+
position_ids,
|
| 1468 |
+
past_key_values,
|
| 1469 |
+
output_attentions,
|
| 1470 |
+
use_cache,
|
| 1471 |
+
cache_position,
|
| 1472 |
+
encoder_hidden_states,
|
| 1473 |
+
self_attn_mask_mapping["encoder_attention_mask"],
|
| 1474 |
+
**flash_attn_kwargs,
|
| 1475 |
+
)
|
| 1476 |
+
hidden_states = layer_outputs[0]
|
| 1477 |
+
|
| 1478 |
+
if output_attentions and self.layers[index_block].use_cross_attention:
|
| 1479 |
+
# layer_outputs structure: (hidden_states, self_attn_weights, cross_attn_weights)
|
| 1480 |
+
# Extract the last element which is cross_attn_weights
|
| 1481 |
+
if len(layer_outputs) >= 3:
|
| 1482 |
+
all_cross_attentions += (layer_outputs[2],)
|
| 1483 |
+
|
| 1484 |
+
if return_hidden_states:
|
| 1485 |
+
return hidden_states
|
| 1486 |
+
|
| 1487 |
+
# Extract scale-shift parameters for adaptive output normalization
|
| 1488 |
+
shift, scale = (self.scale_shift_table + temb.unsqueeze(1)).chunk(2, dim=1)
|
| 1489 |
+
shift = shift.to(hidden_states.device)
|
| 1490 |
+
scale = scale.to(hidden_states.device)
|
| 1491 |
+
|
| 1492 |
+
# Apply adaptive layer norm: norm(x) * (1 + scale) + shift
|
| 1493 |
+
hidden_states = (self.norm_out(hidden_states) * (1 + scale) + shift).type_as(hidden_states)
|
| 1494 |
+
# Project output: de-patchify back to original sequence format
|
| 1495 |
+
hidden_states = self.proj_out(hidden_states)
|
| 1496 |
+
|
| 1497 |
+
# Crop back to original sequence length to ensure exact length match (remove padding)
|
| 1498 |
+
hidden_states = hidden_states[:, :original_seq_len, :]
|
| 1499 |
+
|
| 1500 |
+
outputs = (hidden_states, past_key_values)
|
| 1501 |
+
|
| 1502 |
+
if output_attentions:
|
| 1503 |
+
outputs += (all_cross_attentions,)
|
| 1504 |
+
return outputs
|
| 1505 |
+
|
| 1506 |
+
class AceStepConditionEncoder(AceStepPreTrainedModel):
|
| 1507 |
+
"""
|
| 1508 |
+
Condition encoder for AceStep model.
|
| 1509 |
+
|
| 1510 |
+
Encodes multiple conditioning inputs (text, lyrics, timbre) and packs them
|
| 1511 |
+
into a single sequence for cross-attention in the diffusion model. Handles
|
| 1512 |
+
projection, encoding, and sequence packing.
|
| 1513 |
+
"""
|
| 1514 |
+
def __init__(self, config: AceStepConfig):
|
| 1515 |
+
super().__init__(config)
|
| 1516 |
+
self.config = config
|
| 1517 |
+
# Project text embeddings to model hidden size
|
| 1518 |
+
self.text_projector = nn.Linear(config.text_hidden_dim, config.hidden_size, bias=False)
|
| 1519 |
+
# Encoder for lyric text
|
| 1520 |
+
self.lyric_encoder = AceStepLyricEncoder(config)
|
| 1521 |
+
# Encoder for timbre from reference audio
|
| 1522 |
+
self.timbre_encoder = AceStepTimbreEncoder(config)
|
| 1523 |
+
|
| 1524 |
+
def forward(
|
| 1525 |
+
self,
|
| 1526 |
+
# Text inputs
|
| 1527 |
+
text_hidden_states: Optional[torch.FloatTensor] = None,
|
| 1528 |
+
text_attention_mask: Optional[torch.Tensor] = None,
|
| 1529 |
+
# Lyric inputs
|
| 1530 |
+
lyric_hidden_states: Optional[torch.LongTensor] = None,
|
| 1531 |
+
lyric_attention_mask: Optional[torch.Tensor] = None,
|
| 1532 |
+
# Reference audio for timbre
|
| 1533 |
+
refer_audio_acoustic_hidden_states_packed: Optional[torch.Tensor] = None,
|
| 1534 |
+
refer_audio_order_mask: Optional[torch.LongTensor] = None,
|
| 1535 |
+
):
|
| 1536 |
+
# Project and encode text
|
| 1537 |
+
text_hidden_states = self.text_projector(text_hidden_states)
|
| 1538 |
+
# Encode lyrics
|
| 1539 |
+
lyric_encoder_outputs = self.lyric_encoder(
|
| 1540 |
+
inputs_embeds=lyric_hidden_states,
|
| 1541 |
+
attention_mask=lyric_attention_mask,
|
| 1542 |
+
)
|
| 1543 |
+
lyric_hidden_states = lyric_encoder_outputs.last_hidden_state
|
| 1544 |
+
# Encode timbre from reference audio
|
| 1545 |
+
timbre_embs_unpack, timbre_embs_mask = self.timbre_encoder(refer_audio_acoustic_hidden_states_packed, refer_audio_order_mask)
|
| 1546 |
+
|
| 1547 |
+
# Pack sequences: combine lyrics and timbre, then add text
|
| 1548 |
+
# This creates a single sequence with all conditioning information
|
| 1549 |
+
encoder_hidden_states, encoder_attention_mask = pack_sequences(lyric_hidden_states, timbre_embs_unpack, lyric_attention_mask, timbre_embs_mask)
|
| 1550 |
+
encoder_hidden_states, encoder_attention_mask = pack_sequences(encoder_hidden_states, text_hidden_states, encoder_attention_mask, text_attention_mask)
|
| 1551 |
+
return encoder_hidden_states, encoder_attention_mask
|
| 1552 |
+
|
| 1553 |
+
|
| 1554 |
+
class AceStepConditionGenerationModel(AceStepPreTrainedModel):
|
| 1555 |
+
"""
|
| 1556 |
+
Main conditional generation model for AceStep.
|
| 1557 |
+
|
| 1558 |
+
End-to-end model for generating audio conditioned on text, lyrics, and timbre.
|
| 1559 |
+
Combines encoder (for conditioning), decoder (diffusion model), tokenizer
|
| 1560 |
+
(for discrete tokenization), and detokenizer (for reconstruction).
|
| 1561 |
+
Supports flow matching training and inference with various sampling methods.
|
| 1562 |
+
"""
|
| 1563 |
+
def __init__(self, config: AceStepConfig):
|
| 1564 |
+
super().__init__(config)
|
| 1565 |
+
self.config = config
|
| 1566 |
+
# Diffusion model components
|
| 1567 |
+
self.decoder = AceStepDiTModel(config) # Main diffusion transformer
|
| 1568 |
+
self.encoder = AceStepConditionEncoder(config) # Condition encoder
|
| 1569 |
+
self.tokenizer = AceStepAudioTokenizer(config) # Audio tokenizer
|
| 1570 |
+
self.detokenizer = AudioTokenDetokenizer(config) # Audio detokenizer
|
| 1571 |
+
# Null condition embedding for classifier-free guidance
|
| 1572 |
+
self.null_condition_emb = nn.Parameter(torch.randn(1, 1, config.hidden_size))
|
| 1573 |
+
|
| 1574 |
+
# Initialize weights and apply final processing
|
| 1575 |
+
self.post_init()
|
| 1576 |
+
|
| 1577 |
+
def tokenize(self, x, silence_latent, attention_mask):
|
| 1578 |
+
if x.shape[1] % self.config.pool_window_size != 0:
|
| 1579 |
+
pad_len = self.config.pool_window_size - (x.shape[1] % self.config.pool_window_size)
|
| 1580 |
+
x = torch.cat([x, silence_latent[:1,:pad_len].repeat(x.shape[0],1,1)], dim=1)
|
| 1581 |
+
attention_mask = F.pad(attention_mask, (0, pad_len), mode='constant', value=0)
|
| 1582 |
+
x = rearrange(x, 'n (t_patch p) d -> n t_patch p d', p=self.config.pool_window_size)
|
| 1583 |
+
seq_len = x.shape[1]
|
| 1584 |
+
chunk = math.ceil(attention_mask.shape[1] / seq_len)
|
| 1585 |
+
attention_mask = attention_mask.to(x.dtype)
|
| 1586 |
+
attention_mask = F.max_pool1d(attention_mask.unsqueeze(1), kernel_size=chunk, stride=chunk, ceil_mode=True).squeeze(1)
|
| 1587 |
+
quantized, indices = self.tokenizer(x)
|
| 1588 |
+
return quantized, indices, attention_mask
|
| 1589 |
+
|
| 1590 |
+
def detokenize(self, quantized):
|
| 1591 |
+
"""
|
| 1592 |
+
Detokenize quantized audio tokens back to continuous representations.
|
| 1593 |
+
|
| 1594 |
+
Args:
|
| 1595 |
+
quantized: Quantized tokens of shape [N, T//pool_window_size, d]
|
| 1596 |
+
|
| 1597 |
+
Returns:
|
| 1598 |
+
Detokenized hidden states of shape [N, T, d]
|
| 1599 |
+
"""
|
| 1600 |
+
hidden_states = self.detokenizer(quantized)
|
| 1601 |
+
return hidden_states
|
| 1602 |
+
|
| 1603 |
+
@torch.no_grad()
|
| 1604 |
+
def prepare_condition(
|
| 1605 |
+
self,
|
| 1606 |
+
text_hidden_states: torch.FloatTensor,
|
| 1607 |
+
text_attention_mask: torch.Tensor,
|
| 1608 |
+
lyric_hidden_states: torch.FloatTensor,
|
| 1609 |
+
lyric_attention_mask: torch.Tensor,
|
| 1610 |
+
refer_audio_acoustic_hidden_states_packed: torch.FloatTensor,
|
| 1611 |
+
refer_audio_order_mask: torch.Tensor,
|
| 1612 |
+
hidden_states: torch.FloatTensor,
|
| 1613 |
+
attention_mask: torch.Tensor,
|
| 1614 |
+
silence_latent: torch.FloatTensor,
|
| 1615 |
+
src_latents: torch.FloatTensor,
|
| 1616 |
+
chunk_masks: torch.Tensor,
|
| 1617 |
+
is_covers: torch.Tensor,
|
| 1618 |
+
precomputed_lm_hints_25Hz: Optional[torch.FloatTensor] = None,
|
| 1619 |
+
audio_codes: torch.FloatTensor = None,
|
| 1620 |
+
):
|
| 1621 |
+
|
| 1622 |
+
dtype = hidden_states.dtype
|
| 1623 |
+
encoder_hidden_states, encoder_attention_mask = self.encoder(
|
| 1624 |
+
text_hidden_states=text_hidden_states,
|
| 1625 |
+
text_attention_mask=text_attention_mask,
|
| 1626 |
+
lyric_hidden_states=lyric_hidden_states,
|
| 1627 |
+
lyric_attention_mask=lyric_attention_mask,
|
| 1628 |
+
refer_audio_acoustic_hidden_states_packed=refer_audio_acoustic_hidden_states_packed,
|
| 1629 |
+
refer_audio_order_mask=refer_audio_order_mask,
|
| 1630 |
+
)
|
| 1631 |
+
|
| 1632 |
+
# N x T x d -> N x T//pool_window_size x pool_window_size x d
|
| 1633 |
+
# tokenize and detokenize to get LM hints for cover songs (when is_covers=True)
|
| 1634 |
+
# Use precomputed hints if provided (e.g., from audio codes), otherwise tokenize hidden_states
|
| 1635 |
+
if precomputed_lm_hints_25Hz is not None:
|
| 1636 |
+
print("Using precomputed LM hints")
|
| 1637 |
+
lm_hints_25Hz = precomputed_lm_hints_25Hz[:, :src_latents.shape[1], :]
|
| 1638 |
+
else:
|
| 1639 |
+
if audio_codes is not None:
|
| 1640 |
+
lm_hints_5Hz = self.tokenize.quantizer.get_output_from_indices(audio_codes)
|
| 1641 |
+
else:
|
| 1642 |
+
lm_hints_5Hz, indices, llm_mask = self.tokenize(hidden_states, silence_latent, attention_mask)
|
| 1643 |
+
lm_hints_25Hz = self.detokenize(lm_hints_5Hz)
|
| 1644 |
+
# Crop lm_hints_25Hz to match src_latents length (tokenize may have added padding)
|
| 1645 |
+
lm_hints_25Hz = lm_hints_25Hz[:, :src_latents.shape[1], :]
|
| 1646 |
+
src_latents = torch.where(is_covers.unsqueeze(-1).unsqueeze(-1) > 0, lm_hints_25Hz, src_latents)
|
| 1647 |
+
# Concatenate source latents with chunk masks as context
|
| 1648 |
+
context_latents = torch.cat([src_latents, chunk_masks.to(dtype)], dim=-1)
|
| 1649 |
+
return encoder_hidden_states, encoder_attention_mask, context_latents
|
| 1650 |
+
|
| 1651 |
+
def forward(
|
| 1652 |
+
self,
|
| 1653 |
+
# Diffusion inputs
|
| 1654 |
+
hidden_states: torch.FloatTensor,
|
| 1655 |
+
attention_mask: torch.Tensor,
|
| 1656 |
+
# Encoder inputs
|
| 1657 |
+
# Text
|
| 1658 |
+
text_hidden_states: Optional[torch.FloatTensor] = None,
|
| 1659 |
+
text_attention_mask: Optional[torch.Tensor] = None,
|
| 1660 |
+
# Lyric
|
| 1661 |
+
lyric_hidden_states: Optional[torch.LongTensor] = None,
|
| 1662 |
+
lyric_attention_mask: Optional[torch.Tensor] = None,
|
| 1663 |
+
# Reference audio for timbre
|
| 1664 |
+
refer_audio_acoustic_hidden_states_packed: Optional[torch.Tensor] = None,
|
| 1665 |
+
refer_audio_order_mask: Optional[torch.LongTensor] = None,
|
| 1666 |
+
src_latents: torch.FloatTensor = None,
|
| 1667 |
+
chunk_masks: torch.FloatTensor = None,
|
| 1668 |
+
is_covers: torch.Tensor = None,
|
| 1669 |
+
silence_latent: torch.FloatTensor = None,
|
| 1670 |
+
cfg_ratio: float = 0.15,
|
| 1671 |
+
):
|
| 1672 |
+
"""
|
| 1673 |
+
Forward pass for training (computes training losses).
|
| 1674 |
+
"""
|
| 1675 |
+
# Prepare conditioning inputs (encoder states, context latents)
|
| 1676 |
+
encoder_hidden_states, encoder_attention_mask, context_latents = self.prepare_condition(
|
| 1677 |
+
text_hidden_states=text_hidden_states,
|
| 1678 |
+
text_attention_mask=text_attention_mask,
|
| 1679 |
+
lyric_hidden_states=lyric_hidden_states,
|
| 1680 |
+
lyric_attention_mask=lyric_attention_mask,
|
| 1681 |
+
refer_audio_acoustic_hidden_states_packed=refer_audio_acoustic_hidden_states_packed,
|
| 1682 |
+
refer_audio_order_mask=refer_audio_order_mask,
|
| 1683 |
+
hidden_states=src_latents,
|
| 1684 |
+
attention_mask=attention_mask,
|
| 1685 |
+
silence_latent=silence_latent,
|
| 1686 |
+
src_latents=src_latents,
|
| 1687 |
+
chunk_masks=chunk_masks,
|
| 1688 |
+
is_covers=is_covers,
|
| 1689 |
+
)
|
| 1690 |
+
bsz, device, dtype = hidden_states.shape[0], hidden_states.device, hidden_states.dtype
|
| 1691 |
+
# Classifier-free guidance: randomly drop conditions with probability cfg_ratio
|
| 1692 |
+
# This helps the model learn to work with and without conditions
|
| 1693 |
+
full_cfg_condition_mask = torch.where(
|
| 1694 |
+
(torch.rand(size=(bsz,), device=device, dtype=dtype) < cfg_ratio),
|
| 1695 |
+
torch.zeros(size=(bsz,), device=device, dtype=dtype),
|
| 1696 |
+
torch.ones(size=(bsz,), device=device, dtype=dtype)
|
| 1697 |
+
).view(-1, 1, 1)
|
| 1698 |
+
# Replace dropped conditions with null condition embedding
|
| 1699 |
+
encoder_hidden_states = torch.where(full_cfg_condition_mask > 0, encoder_hidden_states, self.null_condition_emb.expand_as(encoder_hidden_states))
|
| 1700 |
+
|
| 1701 |
+
# Flow matching setup: sample noise x1 and interpolate with data x0
|
| 1702 |
+
x1 = torch.randn_like(hidden_states) # Noise
|
| 1703 |
+
x0 = hidden_states # Data
|
| 1704 |
+
# Sample timesteps t and r for flow matching
|
| 1705 |
+
t, r = sample_t_r(bsz, device, dtype, self.config.data_proportion, self.config.timestep_mu, self.config.timestep_sigma, use_meanflow=False)
|
| 1706 |
+
t_ = t.unsqueeze(-1).unsqueeze(-1)
|
| 1707 |
+
# Interpolate: x_t = t * x1 + (1 - t) * x0
|
| 1708 |
+
xt = t_ * x1 + (1.0 - t_) * x0
|
| 1709 |
+
|
| 1710 |
+
# Predict flow (velocity) from diffusion model
|
| 1711 |
+
decoder_outputs = self.decoder(
|
| 1712 |
+
hidden_states=xt,
|
| 1713 |
+
timestep=t,
|
| 1714 |
+
timestep_r=t,
|
| 1715 |
+
attention_mask=attention_mask,
|
| 1716 |
+
encoder_hidden_states=encoder_hidden_states,
|
| 1717 |
+
encoder_attention_mask=encoder_attention_mask,
|
| 1718 |
+
context_latents=context_latents,
|
| 1719 |
+
)
|
| 1720 |
+
# Flow matching loss: predict the flow field v = x1 - x0
|
| 1721 |
+
flow = x1 - x0
|
| 1722 |
+
diffusion_loss = F.mse_loss(decoder_outputs[0], flow)
|
| 1723 |
+
return {
|
| 1724 |
+
"diffusion_loss": diffusion_loss,
|
| 1725 |
+
}
|
| 1726 |
+
|
| 1727 |
+
def training_losses(self, **kwargs):
|
| 1728 |
+
return self.forward(**kwargs)
|
| 1729 |
+
|
| 1730 |
+
def prepare_noise(self, context_latents: torch.FloatTensor, seed: Union[int, List[int], None] = None):
|
| 1731 |
+
"""
|
| 1732 |
+
Prepare noise tensor for generation with optional seeding.
|
| 1733 |
+
|
| 1734 |
+
Args:
|
| 1735 |
+
context_latents: Context latents to determine noise shape
|
| 1736 |
+
seed: Can be int, List[int], or None. If None, uses random noise.
|
| 1737 |
+
|
| 1738 |
+
Returns:
|
| 1739 |
+
Noise tensor of appropriate shape
|
| 1740 |
+
"""
|
| 1741 |
+
bsz = context_latents.shape[0]
|
| 1742 |
+
device = context_latents.device
|
| 1743 |
+
dtype = context_latents.dtype
|
| 1744 |
+
# Handle seed: can be int, List[int], or None
|
| 1745 |
+
src_latents_shape = (context_latents.shape[0], context_latents.shape[1], context_latents.shape[-1] // 2)
|
| 1746 |
+
if seed is None:
|
| 1747 |
+
# No seed provided - use random
|
| 1748 |
+
noise = torch.randn(src_latents_shape, device=device, dtype=dtype)
|
| 1749 |
+
elif isinstance(seed, list):
|
| 1750 |
+
# List of seeds - generate noise for each sample separately
|
| 1751 |
+
noise_list = []
|
| 1752 |
+
for i, s in enumerate(seed):
|
| 1753 |
+
if s is None or s < 0:
|
| 1754 |
+
# Random seed for this sample
|
| 1755 |
+
noise_i = torch.randn(1, src_latents_shape[1], src_latents_shape[2], device=device, dtype=dtype)
|
| 1756 |
+
else:
|
| 1757 |
+
# Use specific seed for this sample
|
| 1758 |
+
generator = torch.Generator(device=device).manual_seed(int(s))
|
| 1759 |
+
noise_i = torch.randn(1, src_latents_shape[1], src_latents_shape[2], generator=generator, device=device, dtype=dtype)
|
| 1760 |
+
noise_list.append(noise_i)
|
| 1761 |
+
noise = torch.cat(noise_list, dim=0)
|
| 1762 |
+
else:
|
| 1763 |
+
# Single seed for all samples
|
| 1764 |
+
generator = torch.Generator(device=device).manual_seed(int(seed))
|
| 1765 |
+
noise = torch.randn(src_latents_shape, generator=generator, device=device, dtype=dtype)
|
| 1766 |
+
|
| 1767 |
+
return noise
|
| 1768 |
+
|
| 1769 |
+
def get_x0_from_noise(self, zt, vt, t):
|
| 1770 |
+
return zt - vt * t.unsqueeze(-1).unsqueeze(-1)
|
| 1771 |
+
|
| 1772 |
+
def renoise(self, x, t, noise=None):
|
| 1773 |
+
if noise is None:
|
| 1774 |
+
noise = torch.randn_like(x)
|
| 1775 |
+
if isinstance(t, torch.Tensor) and t.ndim != x.ndim:
|
| 1776 |
+
t = t.unsqueeze(-1).unsqueeze(-1)
|
| 1777 |
+
xt = t * noise + (1 - t) * x
|
| 1778 |
+
return xt
|
| 1779 |
+
|
| 1780 |
+
def generate_audio(
|
| 1781 |
+
self,
|
| 1782 |
+
text_hidden_states: torch.FloatTensor,
|
| 1783 |
+
text_attention_mask: torch.FloatTensor,
|
| 1784 |
+
lyric_hidden_states: torch.FloatTensor,
|
| 1785 |
+
lyric_attention_mask: torch.FloatTensor,
|
| 1786 |
+
refer_audio_acoustic_hidden_states_packed: torch.FloatTensor,
|
| 1787 |
+
refer_audio_order_mask: torch.LongTensor,
|
| 1788 |
+
src_latents: torch.FloatTensor,
|
| 1789 |
+
chunk_masks: torch.FloatTensor,
|
| 1790 |
+
is_covers: torch.Tensor,
|
| 1791 |
+
silence_latent: Optional[torch.FloatTensor] = None,
|
| 1792 |
+
attention_mask: torch.Tensor = None,
|
| 1793 |
+
seed: int = None,
|
| 1794 |
+
fix_nfe: int = 8,
|
| 1795 |
+
infer_method: str = "ode",
|
| 1796 |
+
use_cache: bool = True,
|
| 1797 |
+
audio_cover_strength: float = 1.0,
|
| 1798 |
+
non_cover_text_hidden_states: Optional[torch.FloatTensor] = None,
|
| 1799 |
+
non_cover_text_attention_mask: Optional[torch.FloatTensor] = None,
|
| 1800 |
+
precomputed_lm_hints_25Hz: Optional[torch.FloatTensor] = None,
|
| 1801 |
+
audio_codes: Optional[torch.FloatTensor] = None,
|
| 1802 |
+
shift: float = 3.0,
|
| 1803 |
+
timesteps: Optional[torch.Tensor] = None,
|
| 1804 |
+
**kwargs,
|
| 1805 |
+
):
|
| 1806 |
+
# Valid shifts: only discrete values 1, 2, 3 are supported
|
| 1807 |
+
VALID_SHIFTS = [1.0, 2.0, 3.0]
|
| 1808 |
+
|
| 1809 |
+
# Valid timesteps: all unique timesteps from shift=1,2,3 with fix_nfe=8 (total 20 values)
|
| 1810 |
+
VALID_TIMESTEPS = [
|
| 1811 |
+
1.0, 0.9545454545454546, 0.9333333333333333, 0.9, 0.875,
|
| 1812 |
+
0.8571428571428571, 0.8333333333333334, 0.7692307692307693, 0.75,
|
| 1813 |
+
0.6666666666666666, 0.6428571428571429, 0.625, 0.5454545454545454,
|
| 1814 |
+
0.5, 0.4, 0.375, 0.3, 0.25, 0.2222222222222222, 0.125
|
| 1815 |
+
]
|
| 1816 |
+
|
| 1817 |
+
# Pre-defined timestep schedules for each valid shift (fix_nfe=8, excluding final 0)
|
| 1818 |
+
SHIFT_TIMESTEPS = {
|
| 1819 |
+
1.0: [1.0, 0.875, 0.75, 0.625, 0.5, 0.375, 0.25, 0.125],
|
| 1820 |
+
2.0: [1.0, 0.9333333333333333, 0.8571428571428571, 0.7692307692307693, 0.6666666666666666, 0.5454545454545454, 0.4, 0.2222222222222222],
|
| 1821 |
+
3.0: [1.0, 0.9545454545454546, 0.9, 0.8333333333333334, 0.75, 0.6428571428571429, 0.5, 0.3],
|
| 1822 |
+
}
|
| 1823 |
+
|
| 1824 |
+
# Determine the timestep schedule to use
|
| 1825 |
+
t_schedule_list = None
|
| 1826 |
+
|
| 1827 |
+
if timesteps is not None:
|
| 1828 |
+
# Process custom timesteps: map each value to nearest valid timestep
|
| 1829 |
+
timesteps_list = timesteps.tolist() if isinstance(timesteps, torch.Tensor) else list(timesteps)
|
| 1830 |
+
|
| 1831 |
+
# Remove trailing zeros
|
| 1832 |
+
while len(timesteps_list) > 0 and timesteps_list[-1] == 0:
|
| 1833 |
+
timesteps_list.pop()
|
| 1834 |
+
|
| 1835 |
+
# Validate length: 1-20
|
| 1836 |
+
if len(timesteps_list) < 1:
|
| 1837 |
+
logger.warning(f"timesteps length is too short after removing trailing zeros, using default shift={shift}")
|
| 1838 |
+
elif len(timesteps_list) > 20:
|
| 1839 |
+
logger.warning(f"timesteps length={len(timesteps_list)} exceeds maximum 20, truncating to 20")
|
| 1840 |
+
timesteps_list = timesteps_list[:20]
|
| 1841 |
+
t_schedule_list = timesteps_list
|
| 1842 |
+
else:
|
| 1843 |
+
t_schedule_list = timesteps_list
|
| 1844 |
+
|
| 1845 |
+
if t_schedule_list is not None:
|
| 1846 |
+
# Map each timestep to nearest valid timestep
|
| 1847 |
+
original_timesteps = t_schedule_list.copy()
|
| 1848 |
+
mapped_timesteps = []
|
| 1849 |
+
for t in t_schedule_list:
|
| 1850 |
+
nearest = min(VALID_TIMESTEPS, key=lambda x: abs(x - t))
|
| 1851 |
+
mapped_timesteps.append(nearest)
|
| 1852 |
+
|
| 1853 |
+
if original_timesteps != mapped_timesteps:
|
| 1854 |
+
logger.warning(f"timesteps mapped to nearest valid values: {original_timesteps} -> {mapped_timesteps}")
|
| 1855 |
+
|
| 1856 |
+
t_schedule_list = mapped_timesteps
|
| 1857 |
+
|
| 1858 |
+
if t_schedule_list is None:
|
| 1859 |
+
# Use shift-based schedule: round to nearest valid shift
|
| 1860 |
+
original_shift = shift
|
| 1861 |
+
shift = min(VALID_SHIFTS, key=lambda x: abs(x - shift))
|
| 1862 |
+
if original_shift != shift:
|
| 1863 |
+
logger.warning(f"shift={original_shift} not supported, rounded to nearest valid shift={shift}")
|
| 1864 |
+
t_schedule_list = SHIFT_TIMESTEPS[shift]
|
| 1865 |
+
|
| 1866 |
+
if attention_mask is None:
|
| 1867 |
+
latent_length = src_latents.shape[1]
|
| 1868 |
+
attention_mask = torch.ones(src_latents.shape[0], latent_length, device=src_latents.device, dtype=src_latents.dtype)
|
| 1869 |
+
|
| 1870 |
+
time_costs = {}
|
| 1871 |
+
start_time = time.time()
|
| 1872 |
+
total_start_time = start_time
|
| 1873 |
+
encoder_hidden_states, encoder_attention_mask, context_latents = self.prepare_condition(
|
| 1874 |
+
text_hidden_states=text_hidden_states,
|
| 1875 |
+
text_attention_mask=text_attention_mask,
|
| 1876 |
+
lyric_hidden_states=lyric_hidden_states,
|
| 1877 |
+
lyric_attention_mask=lyric_attention_mask,
|
| 1878 |
+
refer_audio_acoustic_hidden_states_packed=refer_audio_acoustic_hidden_states_packed,
|
| 1879 |
+
refer_audio_order_mask=refer_audio_order_mask,
|
| 1880 |
+
hidden_states=src_latents,
|
| 1881 |
+
attention_mask=attention_mask,
|
| 1882 |
+
silence_latent=silence_latent,
|
| 1883 |
+
src_latents=src_latents,
|
| 1884 |
+
chunk_masks=chunk_masks,
|
| 1885 |
+
is_covers=is_covers,
|
| 1886 |
+
precomputed_lm_hints_25Hz=precomputed_lm_hints_25Hz,
|
| 1887 |
+
audio_codes=audio_codes,
|
| 1888 |
+
)
|
| 1889 |
+
|
| 1890 |
+
encoder_hidden_states_non_cover, encoder_attention_mask_non_cover, context_latents_non_cover = None, None, None
|
| 1891 |
+
if audio_cover_strength < 1.0:
|
| 1892 |
+
non_is_covers = torch.zeros_like(is_covers, device=is_covers.device, dtype=is_covers.dtype)
|
| 1893 |
+
# Use silence_latent for non-cover condition to simulate text2music mode (no reference audio)
|
| 1894 |
+
silence_latent_expanded = silence_latent[:, :src_latents.shape[1], :].expand(src_latents.shape[0], -1, -1)
|
| 1895 |
+
encoder_hidden_states_non_cover, encoder_attention_mask_non_cover, context_latents_non_cover = self.prepare_condition(
|
| 1896 |
+
text_hidden_states=non_cover_text_hidden_states,
|
| 1897 |
+
text_attention_mask=non_cover_text_attention_mask,
|
| 1898 |
+
lyric_hidden_states=lyric_hidden_states,
|
| 1899 |
+
lyric_attention_mask=lyric_attention_mask,
|
| 1900 |
+
refer_audio_acoustic_hidden_states_packed=refer_audio_acoustic_hidden_states_packed,
|
| 1901 |
+
refer_audio_order_mask=refer_audio_order_mask,
|
| 1902 |
+
hidden_states=silence_latent_expanded,
|
| 1903 |
+
attention_mask=attention_mask,
|
| 1904 |
+
silence_latent=silence_latent,
|
| 1905 |
+
src_latents=silence_latent_expanded,
|
| 1906 |
+
chunk_masks=chunk_masks,
|
| 1907 |
+
is_covers=non_is_covers,
|
| 1908 |
+
precomputed_lm_hints_25Hz=None,
|
| 1909 |
+
audio_codes=None,
|
| 1910 |
+
)
|
| 1911 |
+
|
| 1912 |
+
end_time = time.time()
|
| 1913 |
+
time_costs["encoder_time_cost"] = end_time - start_time
|
| 1914 |
+
start_time = end_time
|
| 1915 |
+
|
| 1916 |
+
noise = self.prepare_noise(context_latents, seed)
|
| 1917 |
+
bsz, device, dtype = context_latents.shape[0], context_latents.device, context_latents.dtype
|
| 1918 |
+
past_key_values = EncoderDecoderCache(DynamicCache(), DynamicCache())
|
| 1919 |
+
|
| 1920 |
+
# Use pre-computed t_schedule_list (already validated and mapped to valid timesteps)
|
| 1921 |
+
t_schedule = torch.tensor(t_schedule_list, device=device, dtype=dtype)
|
| 1922 |
+
num_steps = len(t_schedule)
|
| 1923 |
+
|
| 1924 |
+
# Recalculate cover_steps based on actual num_steps
|
| 1925 |
+
cover_steps = int(num_steps * audio_cover_strength)
|
| 1926 |
+
|
| 1927 |
+
xt = noise
|
| 1928 |
+
for step_idx in range(num_steps):
|
| 1929 |
+
current_timestep = t_schedule[step_idx].item()
|
| 1930 |
+
t_curr_tensor = current_timestep * torch.ones((bsz,), device=device, dtype=dtype)
|
| 1931 |
+
|
| 1932 |
+
if step_idx >= cover_steps:
|
| 1933 |
+
encoder_hidden_states = encoder_hidden_states_non_cover
|
| 1934 |
+
encoder_attention_mask = encoder_attention_mask_non_cover
|
| 1935 |
+
context_latents = context_latents_non_cover
|
| 1936 |
+
past_key_values = EncoderDecoderCache(DynamicCache(), DynamicCache())
|
| 1937 |
+
|
| 1938 |
+
with torch.no_grad():
|
| 1939 |
+
decoder_outputs = self.decoder(
|
| 1940 |
+
hidden_states=xt,
|
| 1941 |
+
timestep=t_curr_tensor,
|
| 1942 |
+
timestep_r=t_curr_tensor,
|
| 1943 |
+
attention_mask=attention_mask,
|
| 1944 |
+
encoder_hidden_states=encoder_hidden_states,
|
| 1945 |
+
encoder_attention_mask=encoder_attention_mask,
|
| 1946 |
+
context_latents=context_latents,
|
| 1947 |
+
use_cache=True,
|
| 1948 |
+
past_key_values=past_key_values,
|
| 1949 |
+
)
|
| 1950 |
+
|
| 1951 |
+
vt = decoder_outputs[0]
|
| 1952 |
+
past_key_values = decoder_outputs[1]
|
| 1953 |
+
|
| 1954 |
+
# On final step, directly compute x0 from noise
|
| 1955 |
+
if step_idx == num_steps - 1:
|
| 1956 |
+
xt = self.get_x0_from_noise(xt, vt, t_curr_tensor)
|
| 1957 |
+
break
|
| 1958 |
+
|
| 1959 |
+
# Update x_t based on inference method
|
| 1960 |
+
if infer_method == "sde":
|
| 1961 |
+
# Stochastic Differential Equation: predict clean, then re-add noise
|
| 1962 |
+
pred_clean = self.get_x0_from_noise(xt, vt, t_curr_tensor)
|
| 1963 |
+
next_timestep = t_schedule[step_idx + 1].item()
|
| 1964 |
+
xt = self.renoise(pred_clean, next_timestep)
|
| 1965 |
+
elif infer_method == "ode":
|
| 1966 |
+
# Ordinary Differential Equation: Euler method
|
| 1967 |
+
# dx/dt = -v, so x_{t+1} = x_t - v_t * dt
|
| 1968 |
+
next_timestep = t_schedule[step_idx + 1].item()
|
| 1969 |
+
dt = current_timestep - next_timestep
|
| 1970 |
+
dt_tensor = dt * torch.ones((bsz,), device=device, dtype=dtype).unsqueeze(-1).unsqueeze(-1)
|
| 1971 |
+
xt = xt - vt * dt_tensor
|
| 1972 |
+
|
| 1973 |
+
x_gen = xt
|
| 1974 |
+
end_time = time.time()
|
| 1975 |
+
time_costs["diffusion_time_cost"] = end_time - start_time
|
| 1976 |
+
time_costs["diffusion_per_step_time_cost"] = time_costs["diffusion_time_cost"] / num_steps
|
| 1977 |
+
time_costs["total_time_cost"] = end_time - total_start_time
|
| 1978 |
+
return {
|
| 1979 |
+
"target_latents": x_gen,
|
| 1980 |
+
"time_costs": time_costs,
|
| 1981 |
+
}
|
| 1982 |
+
|
| 1983 |
+
|
| 1984 |
+
def test_forward(model, seed=42):
|
| 1985 |
+
# Fix random seed for reproducibility
|
| 1986 |
+
import random
|
| 1987 |
+
import numpy as np
|
| 1988 |
+
random.seed(seed)
|
| 1989 |
+
np.random.seed(seed)
|
| 1990 |
+
torch.manual_seed(seed)
|
| 1991 |
+
if torch.cuda.is_available():
|
| 1992 |
+
torch.cuda.manual_seed(seed)
|
| 1993 |
+
torch.cuda.manual_seed_all(seed)
|
| 1994 |
+
torch.backends.cudnn.deterministic = True
|
| 1995 |
+
torch.backends.cudnn.benchmark = False
|
| 1996 |
+
|
| 1997 |
+
# Get model dtype and device
|
| 1998 |
+
model_dtype = next(model.parameters()).dtype
|
| 1999 |
+
device = next(model.parameters()).device
|
| 2000 |
+
|
| 2001 |
+
print(f"Testing with dtype: {model_dtype}, device: {device}, seed: {seed}")
|
| 2002 |
+
|
| 2003 |
+
# Test data preparation with matching dtype
|
| 2004 |
+
text_hidden_states = torch.randn(2, 77, 1024, dtype=model_dtype, device=device)
|
| 2005 |
+
text_attention_mask = torch.ones(2, 77, dtype=model_dtype, device=device)
|
| 2006 |
+
lyric_hidden_states = torch.randn(2, 123, 1024, dtype=model_dtype, device=device)
|
| 2007 |
+
lyric_attention_mask = torch.ones(2, 123, dtype=model_dtype, device=device)
|
| 2008 |
+
refer_audio_acoustic_hidden_states_packed = torch.randn(3, 750, 64, dtype=model_dtype, device=device)
|
| 2009 |
+
refer_audio_order_mask = torch.LongTensor([0, 0, 1]).to(device)
|
| 2010 |
+
|
| 2011 |
+
# Base config: 25 Hz hidden states → 10 s = 250 frames (round to int)
|
| 2012 |
+
base_seconds = 10
|
| 2013 |
+
frames_per_second = 25
|
| 2014 |
+
base_seq_len = base_seconds * frames_per_second
|
| 2015 |
+
|
| 2016 |
+
hidden_states = torch.randn(2, base_seq_len, 64, dtype=model_dtype, device=device)
|
| 2017 |
+
attention_mask = torch.ones(2, base_seq_len, dtype=model_dtype, device=device)
|
| 2018 |
+
# Add some padding to test mask behavior
|
| 2019 |
+
pad_start = max(base_seq_len // 2, 1)
|
| 2020 |
+
attention_mask[0, pad_start:] = 0
|
| 2021 |
+
chunk_mask = torch.ones(2, base_seq_len, 64, dtype=model_dtype, device=device)
|
| 2022 |
+
chunk_mask[0, pad_start:] = 0
|
| 2023 |
+
|
| 2024 |
+
silence_latent = torch.randn(2, base_seq_len, 64, dtype=model_dtype, device=device)
|
| 2025 |
+
# New required parameters for updated training logic
|
| 2026 |
+
src_latents = torch.randn(2, base_seq_len, 64, dtype=model_dtype, device=device) # Source latents for context
|
| 2027 |
+
is_covers = torch.tensor([0, 1], dtype=torch.long, device=device) # Cover song indicators (0=original, 1=cover)
|
| 2028 |
+
|
| 2029 |
+
# Test 1: Flow matching training (using 10s sequence for sanity check by default)
|
| 2030 |
+
print(f"Testing flow matching training with {base_seconds}s sequence ({base_seq_len} frames @ {frames_per_second}Hz)...")
|
| 2031 |
+
outputs = model.training_losses(
|
| 2032 |
+
hidden_states=hidden_states,
|
| 2033 |
+
attention_mask=attention_mask,
|
| 2034 |
+
chunk_masks=chunk_mask,
|
| 2035 |
+
text_hidden_states=text_hidden_states,
|
| 2036 |
+
text_attention_mask=text_attention_mask,
|
| 2037 |
+
lyric_hidden_states=lyric_hidden_states,
|
| 2038 |
+
lyric_attention_mask=lyric_attention_mask,
|
| 2039 |
+
refer_audio_acoustic_hidden_states_packed=refer_audio_acoustic_hidden_states_packed,
|
| 2040 |
+
refer_audio_order_mask=refer_audio_order_mask,
|
| 2041 |
+
silence_latent=silence_latent,
|
| 2042 |
+
src_latents=src_latents,
|
| 2043 |
+
is_covers=is_covers,
|
| 2044 |
+
cfg_ratio=0.15,
|
| 2045 |
+
)
|
| 2046 |
+
loss = outputs['diffusion_loss']
|
| 2047 |
+
print(f"Flow matching loss: {loss.item():.6f}")
|
| 2048 |
+
print(f" Loss stats - min: {loss.min().item():.6f}, max: {loss.max().item():.6f}, mean: {loss.mean().item():.6f}, std: {loss.std().item() if loss.numel() > 1 else 0:.6f}")
|
| 2049 |
+
|
| 2050 |
+
# Test 2: Generation with flow matching, testing throughput for different sequence lengths
|
| 2051 |
+
lengths_seconds = [10, 30, 60, 120, 180, 240]
|
| 2052 |
+
infer_steps = 2 # Can be increased as needed (e.g., 50/100) to better approximate real inference
|
| 2053 |
+
|
| 2054 |
+
print("\n===== Throughput benchmark (25Hz hidden states) =====")
|
| 2055 |
+
for seconds in lengths_seconds:
|
| 2056 |
+
seq_len = seconds * frames_per_second
|
| 2057 |
+
|
| 2058 |
+
# Reconstruct inputs for current sequence length
|
| 2059 |
+
cur_hidden_states = torch.randn(2, seq_len, 64, dtype=model_dtype, device=device)
|
| 2060 |
+
cur_attention_mask = torch.ones(2, seq_len, dtype=model_dtype, device=device)
|
| 2061 |
+
cur_chunk_mask = torch.ones(2, seq_len, 64, dtype=model_dtype, device=device)
|
| 2062 |
+
cur_silence_latent = torch.randn(2, seq_len, 64, dtype=model_dtype, device=device)
|
| 2063 |
+
cur_src_latents = torch.randn(2, seq_len, 64, dtype=model_dtype, device=device)
|
| 2064 |
+
|
| 2065 |
+
print(f"\n--- {seconds}s input ({seq_len} frames @ {frames_per_second}Hz) ---")
|
| 2066 |
+
outputs = model.generate_audio(
|
| 2067 |
+
text_hidden_states=text_hidden_states,
|
| 2068 |
+
text_attention_mask=text_attention_mask,
|
| 2069 |
+
lyric_hidden_states=lyric_hidden_states,
|
| 2070 |
+
lyric_attention_mask=lyric_attention_mask,
|
| 2071 |
+
refer_audio_acoustic_hidden_states_packed=refer_audio_acoustic_hidden_states_packed,
|
| 2072 |
+
refer_audio_order_mask=refer_audio_order_mask,
|
| 2073 |
+
src_latents=cur_src_latents,
|
| 2074 |
+
chunk_masks=cur_chunk_mask,
|
| 2075 |
+
silence_latent=cur_silence_latent,
|
| 2076 |
+
infer_steps=infer_steps,
|
| 2077 |
+
is_covers=is_covers,
|
| 2078 |
+
seed=1234,
|
| 2079 |
+
)
|
| 2080 |
+
|
| 2081 |
+
target_latents = outputs["target_latents"]
|
| 2082 |
+
time_costs = outputs.get("time_costs", {})
|
| 2083 |
+
|
| 2084 |
+
total_time = time_costs.get("total_time_cost", None)
|
| 2085 |
+
diffusion_time = time_costs.get("diffusion_time_cost", None)
|
| 2086 |
+
|
| 2087 |
+
# Output shape and statistics
|
| 2088 |
+
print(f"Generated latents shape: {target_latents.shape}")
|
| 2089 |
+
print(
|
| 2090 |
+
f"Stats - min: {target_latents.min().item():.4f}, "
|
| 2091 |
+
f"max: {target_latents.max().item():.4f}, "
|
| 2092 |
+
f"mean: {target_latents.mean().item():.4f}, "
|
| 2093 |
+
f"std: {target_latents.std().item():.4f}"
|
| 2094 |
+
)
|
| 2095 |
+
|
| 2096 |
+
# Calculate throughput: statistics by frame count and audio seconds
|
| 2097 |
+
bsz, t_len = target_latents.shape[0], target_latents.shape[1]
|
| 2098 |
+
audio_seconds = t_len / frames_per_second
|
| 2099 |
+
|
| 2100 |
+
if total_time is not None:
|
| 2101 |
+
frames_throughput = (bsz * t_len) / total_time
|
| 2102 |
+
seconds_throughput = (bsz * audio_seconds) / total_time
|
| 2103 |
+
print(
|
| 2104 |
+
f"Time costs: total={total_time:.4f}s, diffusion={diffusion_time:.4f}s "
|
| 2105 |
+
f"({infer_steps} steps)"
|
| 2106 |
+
if diffusion_time is not None
|
| 2107 |
+
else f"Time costs: total={total_time:.4f}s"
|
| 2108 |
+
)
|
| 2109 |
+
print(
|
| 2110 |
+
f"Throughput (based on total_time): "
|
| 2111 |
+
f"{frames_throughput:.2f} frames/s, "
|
| 2112 |
+
f"{seconds_throughput:.2f} audio-seconds/s (batch={bsz})"
|
| 2113 |
+
)
|
| 2114 |
+
else:
|
| 2115 |
+
print("Time costs not available in outputs['time_costs']; only basic stats printed.")
|
| 2116 |
+
|
| 2117 |
+
|
| 2118 |
+
if __name__ == "__main__":
|
| 2119 |
+
from torch.profiler import profile, record_function, ProfilerActivity
|
| 2120 |
+
import os, torch
|
| 2121 |
+
import time
|
| 2122 |
+
from transformers import AutoModel
|
| 2123 |
+
config = AceStepConfig()
|
| 2124 |
+
start = time.time()
|
| 2125 |
+
import os
|
| 2126 |
+
model_dir = os.path.dirname(os.path.abspath(__file__))
|
| 2127 |
+
model = AceStepConditionGenerationModel.from_pretrained(model_dir)
|
| 2128 |
+
end = time.time()
|
| 2129 |
+
# model.config._attn_implementation = "sdpa"
|
| 2130 |
+
model.config._attn_implementation = "flash_attention_2"
|
| 2131 |
+
model.eval()
|
| 2132 |
+
# model = model.to("cpu")
|
| 2133 |
+
# model = model.float()
|
| 2134 |
+
model = model.to("cuda")
|
| 2135 |
+
model = model.bfloat16()
|
| 2136 |
+
test_forward(model)
|
acestep-v15-turbo/silence_latent.pt
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:a778e9dd942f5e8b2c09c55370782d318834432b03dabbcdf70e6ed49ad6358b
|
| 3 |
+
size 3841215
|
config.json
ADDED
|
@@ -0,0 +1,82 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"architectures": [
|
| 3 |
+
"AceStepConditionGenerationModel"
|
| 4 |
+
],
|
| 5 |
+
"attention_bias": false,
|
| 6 |
+
"attention_dropout": 0.0,
|
| 7 |
+
"audio_acoustic_hidden_dim": 64,
|
| 8 |
+
"auto_map": {
|
| 9 |
+
"AutoConfig": "configuration_acestep_v15.AceStepConfig",
|
| 10 |
+
"AutoModel": "modeling_acestep_v15_turbo.AceStepConditionGenerationModel"
|
| 11 |
+
},
|
| 12 |
+
"data_proportion": 0.5,
|
| 13 |
+
"dtype": "bfloat16",
|
| 14 |
+
"fsq_dim": 2048,
|
| 15 |
+
"fsq_input_levels": [
|
| 16 |
+
8,
|
| 17 |
+
8,
|
| 18 |
+
8,
|
| 19 |
+
5,
|
| 20 |
+
5,
|
| 21 |
+
5
|
| 22 |
+
],
|
| 23 |
+
"fsq_input_num_quantizers": 1,
|
| 24 |
+
"head_dim": 128,
|
| 25 |
+
"hidden_act": "silu",
|
| 26 |
+
"hidden_size": 2048,
|
| 27 |
+
"in_channels": 192,
|
| 28 |
+
"initializer_range": 0.02,
|
| 29 |
+
"intermediate_size": 6144,
|
| 30 |
+
"is_turbo": true,
|
| 31 |
+
"layer_types": [
|
| 32 |
+
"sliding_attention",
|
| 33 |
+
"full_attention",
|
| 34 |
+
"sliding_attention",
|
| 35 |
+
"full_attention",
|
| 36 |
+
"sliding_attention",
|
| 37 |
+
"full_attention",
|
| 38 |
+
"sliding_attention",
|
| 39 |
+
"full_attention",
|
| 40 |
+
"sliding_attention",
|
| 41 |
+
"full_attention",
|
| 42 |
+
"sliding_attention",
|
| 43 |
+
"full_attention",
|
| 44 |
+
"sliding_attention",
|
| 45 |
+
"full_attention",
|
| 46 |
+
"sliding_attention",
|
| 47 |
+
"full_attention",
|
| 48 |
+
"sliding_attention",
|
| 49 |
+
"full_attention",
|
| 50 |
+
"sliding_attention",
|
| 51 |
+
"full_attention",
|
| 52 |
+
"sliding_attention",
|
| 53 |
+
"full_attention",
|
| 54 |
+
"sliding_attention",
|
| 55 |
+
"full_attention"
|
| 56 |
+
],
|
| 57 |
+
"max_position_embeddings": 32768,
|
| 58 |
+
"model_type": "acestep",
|
| 59 |
+
"model_version": "turbo",
|
| 60 |
+
"num_attention_heads": 16,
|
| 61 |
+
"num_attention_pooler_hidden_layers": 2,
|
| 62 |
+
"num_audio_decoder_hidden_layers": 24,
|
| 63 |
+
"num_hidden_layers": 24,
|
| 64 |
+
"num_key_value_heads": 8,
|
| 65 |
+
"num_lyric_encoder_hidden_layers": 8,
|
| 66 |
+
"num_timbre_encoder_hidden_layers": 4,
|
| 67 |
+
"patch_size": 2,
|
| 68 |
+
"pool_window_size": 5,
|
| 69 |
+
"rms_norm_eps": 1e-06,
|
| 70 |
+
"rope_scaling": null,
|
| 71 |
+
"rope_theta": 1000000,
|
| 72 |
+
"sliding_window": 128,
|
| 73 |
+
"text_hidden_dim": 1024,
|
| 74 |
+
"timbre_fix_frame": 750,
|
| 75 |
+
"timbre_hidden_dim": 64,
|
| 76 |
+
"timestep_mu": -0.4,
|
| 77 |
+
"timestep_sigma": 1.0,
|
| 78 |
+
"transformers_version": "4.57.0.dev0",
|
| 79 |
+
"use_cache": true,
|
| 80 |
+
"use_sliding_window": true,
|
| 81 |
+
"vocab_size": 64003
|
| 82 |
+
}
|
vae/config.json
ADDED
|
@@ -0,0 +1,24 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"_class_name": "AutoencoderOobleck",
|
| 3 |
+
"_diffusers_version": "0.34.0",
|
| 4 |
+
"_name_or_path": "/root/data/repo/gongjunmin/ACE-Step-1.5/checkpoints/vae/",
|
| 5 |
+
"audio_channels": 2,
|
| 6 |
+
"channel_multiples": [
|
| 7 |
+
1,
|
| 8 |
+
2,
|
| 9 |
+
4,
|
| 10 |
+
8,
|
| 11 |
+
16
|
| 12 |
+
],
|
| 13 |
+
"decoder_channels": 128,
|
| 14 |
+
"decoder_input_channels": 64,
|
| 15 |
+
"downsampling_ratios": [
|
| 16 |
+
2,
|
| 17 |
+
4,
|
| 18 |
+
4,
|
| 19 |
+
6,
|
| 20 |
+
10
|
| 21 |
+
],
|
| 22 |
+
"encoder_hidden_size": 128,
|
| 23 |
+
"sampling_rate": 48000
|
| 24 |
+
}
|
vae/diffusion_pytorch_model.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:da17edb604c40deaf09e9b24974e590d1ca83a374070e5d0884cfa4bed9a99b0
|
| 3 |
+
size 337431388
|