"""量清楚 condition_proj 的 fp32 真值幅度：决定是局部修还是整条路径抬 fp32。"""
import os, torch, torch.nn.functional as F
BASE="/data/models/minimax-h3"
os.environ.setdefault("DIFFSYNTH_ATTENTION_IMPLEMENTATION","torch")
os.environ.setdefault("DIFFSYNTH_MODEL_BASE_PATH", f"{BASE}/models")
os.environ.setdefault("MODELSCOPE_CACHE", f"{BASE}/.mscache")
from diffsynth.pipelines.minimax_h3_audio_video import MiniMaxH3Pipeline, ModelConfig
DT=torch.float16
free=torch.cuda.mem_get_info("cuda")[0]/1024**3
vc=dict(offload_dtype=DT,offload_device="cpu",onload_dtype=DT,onload_device="cpu",
        preparing_dtype=DT,preparing_device="cuda",computation_dtype=DT,computation_device="cuda")
NF4="DiffSynth-Studio/MiniMax-H3-NF4"
pipe=MiniMaxH3Pipeline.from_pretrained(torch_dtype=DT,device="cuda",
  model_configs=[ModelConfig(model_id=NF4,origin_file_pattern=p,**vc) for p in
    ["minimax-h3-fl2va-nf4.safetensors","minimax-h3-text-encoder-nf4.safetensors",
     "video_vae_nf4.safetensors","audio_vae_nf4.safetensors"]],
  processor_config=ModelConfig(model_id="MiniMax/MiniMax-H3",origin_file_pattern="FL2VA/processor/"),
  vram_limit=max(6.0,free-2.0))
pipe.load_lora(pipe.dit, ModelConfig(model_id="lightx2v/Minimax-h3-Turbo",
  origin_file_pattern="minimax_h3_fl2v_turbo_4step_v1.0_768p_bf16.safetensors"))

cp = pipe.dit.condition_proj
class Stop(Exception): pass
res={}
orig_lf = cp.linear_forward
orig_lora = cp.lora_forward
def lf(x, weight, bias):
    o32 = F.linear(x.float(), weight.float(), bias.float() if bias is not None else None)
    res["x_absmax"]=float(x.abs().max()); res["x_dtype"]=str(x.dtype)
    res["lin32_absmax"]=float(o32.abs().max())
    res["lin32_over_fp16"]=float((o32.abs()>65504).float().mean())*100
    res["lin32"]=o32
    return o32
def lora(x, out):
    o = out
    for A,B in zip(cp.lora_A_weights, cp.lora_B_weights):
        o = o + x.float() @ A.T.to(device=x.device,dtype=torch.float32) @ B.T.to(device=x.device,dtype=torch.float32)
    res["n_lora"]=len(cp.lora_A_weights)
    res["final32_absmax"]=float(o.abs().max())
    res["final32_over_fp16"]=float((o.abs()>65504).float().mean())*100
    res["p999"]=float(torch.quantile(o.abs().flatten().float(), 0.999))
    raise Stop()
cp.linear_forward=lf; cp.lora_forward=lora
try:
    pipe(prompt="一只橘猫在窗台上打哈欠", height=480,width=832,num_frames=124,
         num_inference_steps=1, seed=42, flow_shift=6.0)
except Stop:
    pass
print("\n===== condition_proj fp32 真值 =====")
for k in ["x_dtype","x_absmax","lin32_absmax","lin32_over_fp16","n_lora","final32_absmax","final32_over_fp16","p999"]:
    if k in res: print(f"  {k:20s} = {res[k]}")
print(f"  fp16 上限 = 65504")