multi_view/mytest.py

import torch
import os

path = "/root/paddlejob/workspace/qizipeng/baidu/personal-code/Multi-view/multi_view/ckpts/Wan2.2_5B-Multi_view-normal_rope_384_640-3ref_bak/checkpoint-step-1050-epoch-13/pytorch_model.bin"   # 改成你的路径

print(f"\n=== Loading {path} ===")

state = torch.load(path, map_location="cpu")

import pdb; pdb.set_trace()

print(f"\nTotal parameters: {len(state)}")

import torch
from safetensors.torch import load_file

# ====== 1. 加载 DeepSpeed 的完整模型参数 ======
mp_path = "/root/paddlejob/workspace/qizipeng/baidu/personal-code/Multi-view/multi_view/ckpts/Wan2.2_5B-Multi_view-normal_rope_384_640-3ref/checkpoint-step-2-epoch-1/pytorch_model/mp_rank_00_model_states.pt"
mp_sd = torch.load(mp_path, map_location="cpu")

# DS 的 state_dict 存在 mp_sd["module"] 或 mp_sd，本项目一般是 "module"
if "module" in mp_sd:
    mp_sd = mp_sd["module"]

print(f"[INFO] Loaded DeepSpeed full model: {len(mp_sd)} params")

# ====== 2. 加载你保存的可训练权重 ======
safe_path = "/root/paddlejob/workspace/qizipeng/baidu/personal-code/Multi-view/multi_view/ckpts/Wan2.2_5B-Multi_view-normal_rope_384_640-3ref/checkpoint-step-2-epoch-1/weights.safetensors"
safe_sd = load_file(safe_path)
print(f"[INFO] Loaded safetensors trainable params: {len(safe_sd)} params")

# ====== 3. key 集合对比 ======
mp_keys = set(mp_sd.keys())
safe_keys = set(safe_sd.keys())
for key in state.keys():
    if not "pipe." + key in mp_sd.keys():
        print(key)

import pdb; pdb.set_trace()

common_keys = mp_keys & safe_keys
only_mp = mp_keys - safe_keys
only_safe = safe_keys - mp_keys

print("\n===== KEY COMPARISON =====")
print(f"Common keys: {len(common_keys)}")
print(f"Only in mp_rank_00: {len(only_mp)}")
print(f"Only in weights.safetensors: {len(only_safe)}")

if only_mp:
    print("\n-- Keys ONLY in mp_rank_00_model_states.pt (show first 10):")
    for k in list(only_mp)[:10]:
        print(" ", k)

if only_safe:
    print("\n-- Keys ONLY in weights.safetensors (show first 10):")
    for k in list(only_safe)[:10]:
        print(" ", k)

# ====== 4. 查看共同 keys 的 shape 差异 ======
print("\n===== SHAPE CHECK (common keys) =====")
shape_mismatch = []
for k in list(common_keys)[:50]:  # 前50个
    if mp_sd[k].shape != safe_sd[k].shape:
        shape_mismatch.append((k, mp_sd[k].shape, safe_sd[k].shape))

if shape_mismatch:
    print("Shape mismatch found:")
    for k, s1, s2 in shape_mismatch[:20]:
        print(f"  {k}: mp={s1}, safe={s2}")
else:
    print("All shapes match ✓")

# ====== 5. 可选：检查参数差异（可能很大，默认只检查前 20 个 key）=====
print("\n===== VALUE DIFFERENCE CHECK (first 20 keys) =====")
for k in list(common_keys)[:20]:
    diff = torch.abs(mp_sd[k] - safe_sd[k]).mean().item()
    print(f"{k}: mean(abs(diff)) = {diff:.6f}")