初始提交：3 款纯音频游戏 + 共享引擎 + 完整文档与资产管线

- 共享引擎（app/src/main）：GestureEngine 6 手势识别、AudioEngine 三层音频（ExoPlayer BGM 交叉淡入 / 对白 ducking / SoundPool 低延迟）、Game/GameEngine 状态机接口、MainActivity 全屏封面 + TalkBack - 《声命线》storyline：末世互动广播剧，3 幕 8 抉择点 4 结局 - 《声东击西》reflex：听声辨位反应，12 波战斗 + BGM 三档 BPM 变速 - 《暗室》escape：3x3 网格密室逃脱，钥匙→铁柜→保险丝→电箱→大门谜题链 - 文档：setup（从零搭建）/ architecture（架构）/ engine-api（引擎 API）/ asset-pipeline（资产流水线）/ games/*（三款游戏设计文档） - 资产管线：Qwen3-TTS 对白 + 程序合成音效/BGM + Mureka BGM 的 gen 脚本与台账，全部 ogg 资产入库可复现 - env.d 环境脚本：android/tts 入库；sfx/mureka 含密钥不入库，仅提交 .example.sh 模板
2026-08-29 13:00:34 +08:00
commit f0f0a9b50e
446 changed files with 2627836 additions and 0 deletions
@@ -0,0 +1,86 @@
+import json
+
+import torch
+from safetensors import safe_open
+from safetensors.torch import load_file
+from stable_audio_3.factory import (
+    create_autoencoder_from_config,
+    create_diffusion_cond_from_config,
+)
+
+
+def copy_state_dict(model, state_dict):
+    model_state_dict = model.state_dict()
+    state_dict = remap_state_dict_keys(state_dict, model_state_dict)
+    for key in state_dict:
+        if (
+            key in model_state_dict
+            and state_dict[key].shape == model_state_dict[key].shape
+        ):
+            model_state_dict[key] = state_dict[key]
+        else:
+            print(
+                f"Key {key} not found in target state_dict or shape mismatch. Skipping."
+            )
+
+    model.load_state_dict(model_state_dict, strict=False)
+
+
+def load_autoencoder(config_path: str, ckpt_path: str, device: str = "cpu"):
+    """Load only the autoencoder from a combined DiT+autoencoder checkpoint.
+
+    Only pretransform tensors are read from disk, directly onto the target device.
+    Standalone AE-only checkpoints (e.g. stabilityai/SAME-L / SAME-S) have no prefix.
+    """
+
+    with open(config_path) as f:
+        config = json.load(f)
+
+    autoencoder = create_autoencoder_from_config(config["model"], config["sample_rate"])
+
+    # Full DiT checkpoints nest the AE under pretransform.model.*;
+    # standalone AE-only checkpoints have no prefix.
+    nested_prefix = "pretransform.model."
+    with safe_open(ckpt_path, framework="pt", device=device) as f:
+        all_keys = list(f.keys())
+    if any(k.startswith(nested_prefix) for k in all_keys):
+        effective_prefix = nested_prefix
+    else:
+        effective_prefix = ""  # standalone AE — keys are already bare
+    with safe_open(ckpt_path, framework="pt", device=device) as f:
+        state_dict = {
+            k[len(effective_prefix) :]: f.get_tensor(k)
+            for k in all_keys
+            if k.startswith(effective_prefix)
+        }
+
+    copy_state_dict(autoencoder, state_dict)
+    return autoencoder.to(device)
+
+
+def load_diffusion_cond(
+    model_config,
+    ckpt_path: str,
+    device: str = "cuda",
+    model_half: bool = False,
+):
+    model = create_diffusion_cond_from_config(model_config)
+    copy_state_dict(model, load_file(ckpt_path))
+    model.to(device).eval().requires_grad_(False)
+    if model_half:
+        model.to(torch.float16)
+    return model
+
+
+def remap_state_dict_keys(state_dict, model_state_dict):
+    remapped = {}
+    for key, value in state_dict.items():
+        if key not in model_state_dict:
+            parts = key.split(".")
+            for i in range(1, len(parts)):
+                candidate = ".".join(parts[:i]) + "." + ".".join(parts[i + 1 :])
+                if candidate in model_state_dict:
+                    key = candidate
+                    break
+        remapped[key] = value
+    return remapped