[kyo-zzz]

Fixes #1322.

Problem

F5TTS.infer() with the default seed=None draws random.randint(0, sys.maxsize) — a 64-bit value — and seed_everything writes it verbatim into os.environ["PYTHONHASHSEED"]. CPython only accepts "random" or an integer in [0, 2**32 - 1] for that variable, so on every 64-bit platform the environment is polluted with an invalid value (~99.99999995% probability), and any child Python process inheriting it aborts at pre-initialization:

Fatal Python error: config_init_hash_seed: PYTHONHASHSEED must be "random" or an integer in range [0; 4294967295]

This bites users who spawn subprocesses after inference (batch drivers, job schedulers, notebook %run flows).

Fix

Reduce the environment value modulo 2**32 in seed_everything:

os.environ["PYTHONHASHSEED"] = str(seed % 2**32)

random.seed() and torch.manual_seed() keep the original seed, so reproducibility semantics are unchanged — only the environment variable is made legal. This also covers any other caller passing a large seed, and the finetune_gradio.py site is unaffected (its constant seed 666 is already valid).

Verification

Standalone reproduction against src/f5_tts/model/utils.py (before → after):

seed_everything(2**40 + 123)
PYTHONHASHSEED in env: 1099511627899        -> child: Fatal Python error: config_init_hash_seed
after fix: PYTHONHASHSEED in env: 123       -> child: ok