add structure probe for the browser export

kaggle/describe_model.py

@@ -0,0 +1,112 @@
+"""Print the model's internal structure so the ONNX export can be written
+against real APIs instead of guesses.
+
+The original export notebook is gone, and the graphs it produced died with it.
+Everything downstream (browser bundle, quantization) hangs off being able to
+re-export, so this dumps the exact callable signatures, module classes and
+parameter shapes needed to write that exporter once and correctly.
+
+Writes /kaggle/working/structure.json and prints a readable summary.
+"""
+import inspect
+import json
+import sys
+from pathlib import Path
+
+ROOT = Path("/kaggle/working")
+REPO = ROOT / "Ming-3.0-tts"
+MODEL = ROOT / "Ming-3.0-tts-0.5b"
+OUT = ROOT / "structure.json"
+
+
+def describe(name: str, fn) -> dict:
+    try:
+        sig = str(inspect.signature(fn))
+    except Exception:
+        sig = "<no signature>"
+    return {"name": name, "type": type(fn).__name__, "signature": sig,
+            "doc": (inspect.getdoc(fn) or "")[:400]}
+
+
+def main() -> None:
+    import torch
+
+    sys.path.insert(0, str(REPO))
+    for name in ("ming_tts.py", "modeling_bailingmm.py"):
+        p = REPO / name
+        s = p.read_text()
+        s = s.replace("torch_dtype=torch.bfloat16", "torch_dtype=torch.float32")
+        s = s.replace(".eval().to(torch.bfloat16).to(device)", ".eval().to(torch.float32).to(device)")
+        s = s.replace("dtype=torch.bfloat16", "dtype=torch.float32")
+        p.write_text(s)
+
+    from ming_tts import MingTTS
+    import tts_tasks
+
+    tts = MingTTS(str(MODEL), device="cuda", text_normalization=False)
+    mm = tts.model
+
+    info = {"mm_class": type(mm).__name__, "attrs": [a for a in dir(mm) if not a.startswith("_")][:80]}
+
+    lm = getattr(mm, "model", None)
+    info["lm_class"] = type(lm).__name__ if lm is not None else None
+    if lm is not None:
+        cfg = getattr(lm, "config", None)
+        info["lm_config"] = {k: v for k, v in vars(cfg).items()
+                             if isinstance(v, (int, float, str, bool, list))} if cfg else None
+        info["lm_forward"] = describe("lm.forward", lm.forward)
+
+    for attr in ("linear_proj_audio", "prepare_input_embed", "prepare_prompt", "synthesize"):
+        if hasattr(mm, attr):
+            info[attr] = describe(attr, getattr(mm, attr))
+        if hasattr(tts, attr):
+            info[f"tts.{attr}"] = describe(attr, getattr(tts, attr))
+
+    audio = getattr(mm, "audio", None)
+    info["audio_class"] = type(audio).__name__ if audio is not None else None
+    for m in ("encode", "decode", "quantize", "dequantize"):
+        if audio is not None and hasattr(audio, m):
+            info[f"audio.{m}"] = describe(m, getattr(audio, m))
+    if audio is not None:
+        cfg = getattr(audio, "config", None)
+        info["audio_config"] = {k: v for k, v in vars(cfg).items()
+                                if isinstance(v, (int, float, str, bool))} if cfg else None
+
+    flow = getattr(mm, "flowloss", None)
+    info["flow_class"] = type(flow).__name__ if flow is not None else None
+    if flow is not None:
+        info["flow_attrs"] = [a for a in dir(flow) if not a.startswith("_")]
+        dit = getattr(getattr(flow, "cfm", None), "model", None)
+        if dit is not None:
+            info["dit_forward"] = describe("dit.forward", dit.forward)
+
+    totals = {}
+    for name, mod in (("lm", lm), ("audio", audio), ("flow", flow)):
+        if mod is not None and hasattr(mod, "parameters"):
+            totals[name] = sum(p.numel() for p in mod.parameters())
+    if hasattr(mm, "linear_proj_audio") and hasattr(mm.linear_proj_audio, "parameters"):
+        totals["linear_proj_audio"] = sum(p.numel() for p in mm.linear_proj_audio.parameters())
+    info["param_counts"] = totals
+
+    heads = {}
+    for h in ("stop_head", "spk_head", "fc1"):
+        for owner, label in ((mm, f"mm.{h}"), (audio, f"audio.{h}")):
+            if owner is not None and hasattr(owner, h):
+                mod = getattr(owner, h)
+                n = sum(p.numel() for p in mod.parameters()) if hasattr(mod, "parameters") else None
+                heads[label] = n
+    info["heads"] = heads
+
+    OUT.write_text(json.dumps(info, ensure_ascii=False, indent=1))
+    print(json.dumps({k: v for k, v in info.items() if k in
+                      ("mm_class", "lm_class", "audio_class", "flow_class",
+                       "param_counts", "heads", "lm_config", "audio_config")},
+                     ensure_ascii=False, indent=1)[:4000])
+    print("\nsignatures:")
+    for k, v in info.items():
+        if isinstance(v, dict) and "signature" in v:
+            print(f"  {k}{v['signature']}")
+
+
+if __name__ == "__main__":
+    main()