Update AMD ROCm install to 7.2 for RDNA 3.5/4 support

[Kaihui-AMD]

Summary

The README recommends torch==2.5.1+rocm6.2 for AMD GPUs, but ROCm 6.x does not include gfx1151 or gfx1201 targets. Users with RDNA 3.5/4 GPUs (Radeon 8050S/8060S, RX 9060/9070) hit HIP error: invalid device function on first model load (#1236).

Changes

• Default PyTorch recommendation → torch==2.9.1+rocm7.2
• Old ROCm 6.2 command preserved as comment for RDNA1/2/3 users
• Added note with links to official ROCm 6.2 vs 7.2 compatibility matrices confirming the architecture gap

References

• Issue: RuntimeError: HIP error: invalid device function #1236
• ROCm 6.2 compat matrix (no gfx1201): https://rocm.docs.amd.com/en/docs-6.2.4/compatibility/compatibility-matrix.html
• ROCm 7.2 compat matrix (has gfx1201): https://rocm.docs.amd.com/en/docs-7.2.3/compatibility/compatibility-matrix.html

[Kaihui-AMD]

test code:

import torch
print('device:', torch.cuda.get_device_name(0))
print('hip:', torch.version.hip)
from f5_tts.infer.utils_infer import load_model
from f5_tts.model import DiT
from huggingface_hub import hf_hub_download
ckpt_path = hf_hub_download(repo_id='SWivid/F5-TTS', filename='F5TTS_v1_Base/model_1250000.safetensors')
F5TTS_model_cfg = dict(dim=1024, depth=22, heads=16, ff_mult=2, text_dim=512, conv_layers=4)
model = load_model(DiT, F5TTS_model_cfg, ckpt_path)
print(f'Model loaded! dtype={next(model.parameters()).dtype}')

output:

device: AMD Radeon Graphics
hip: 7.2.53211-c2d9476115
Model loaded! dtype=torch.float16