[GH-ISSUE #173] Attention not available on the HIP backend #53

New issue

Closed

opened 2026-10-26 22:19:44 +00:00 by bsakaguchi · 1 comment

bsakaguchi commented

2026-10-26 22:19:44 +00:00

Owner

Originally created by @agronholm on GitHub (Sep 12, 2026).
Original GitHub issue: https://github.com/Comfy-Org/comfy-kitchen/issues/173

Building the wheel on Ubuntu 26.04 resulted in the lack of HIP support. I've now spent many hours trying to get this to work. The installation instructions are seemingly simple: just do pip install . – but for whatever reason, all the attention capabilities seem to be absent from the HIP back-end:

 'hip': {'available': True,
         'capabilities': ['adaln',
                          'apply_rope',
                          'apply_rope1',
                          'apply_rope1_',
                          'apply_rope_',
                          'apply_rope_split_half',
                          'apply_rope_split_half1',
                          'apply_rope_split_half1_',
                          'apply_rope_split_half_',
                          'dequantize_convrot_w4a4_weight',
                          'dequantize_int8_convrot_weight_dtype',
                          'dequantize_int8_simple_dtype',
                          'dequantize_per_tensor_fp8',
                          'dequantize_w4a8_int8_weight',
                          'gemv_awq_w4a16',
                          'quantize_and_rotate_rowwise',
                          'quantize_convrot_w4a4_weight',
                          'quantize_int8_convrot_weight',
                          'quantize_int8_rowwise',
                          'quantize_int8_tensorwise',
                          'quantize_per_tensor_fp8',
                          'quantize_svdquant_w4a4',
                          'quantize_w4a8_int8_weight',
                          'rms_adaln',
                          'rms_rope',
                          'rms_rope1',
                          'rms_rope1_',
                          'rms_rope_',
                          'rms_rope_split_half',
                          'rms_rope_split_half1',
                          'rms_rope_split_half1_',
                          'rms_rope_split_half_',
                          'stochastic_rounding_fp8'],
         'disabled': False,
         'unavailable_reason': None},

What's going on? Am I installing this wrong somehow? This is what I have in the virtualenv:

>>> import torch
>>> torch.version.__version__
'2.14.0+rocm7.14'
>>> torch.version.rocm
'7.14.0'

I have ROCm 7.14.1 installed on the system level. Is the patch level discrepancy somehow messing with things? I tried installing both from the PyPI and locally, no change. My hardware is gfx1201, for the record.

Originally created by @agronholm on GitHub (Sep 12, 2026). Original GitHub issue: https://github.com/Comfy-Org/comfy-kitchen/issues/173 Building the wheel on Ubuntu 26.04 resulted in the lack of HIP support. I've now spent many hours trying to get this to work. The installation instructions are seemingly simple: just do `pip install .` – but for whatever reason, all the attention capabilities seem to be absent from the HIP back-end: ``` 'hip': {'available': True, 'capabilities': ['adaln', 'apply_rope', 'apply_rope1', 'apply_rope1_', 'apply_rope_', 'apply_rope_split_half', 'apply_rope_split_half1', 'apply_rope_split_half1_', 'apply_rope_split_half_', 'dequantize_convrot_w4a4_weight', 'dequantize_int8_convrot_weight_dtype', 'dequantize_int8_simple_dtype', 'dequantize_per_tensor_fp8', 'dequantize_w4a8_int8_weight', 'gemv_awq_w4a16', 'quantize_and_rotate_rowwise', 'quantize_convrot_w4a4_weight', 'quantize_int8_convrot_weight', 'quantize_int8_rowwise', 'quantize_int8_tensorwise', 'quantize_per_tensor_fp8', 'quantize_svdquant_w4a4', 'quantize_w4a8_int8_weight', 'rms_adaln', 'rms_rope', 'rms_rope1', 'rms_rope1_', 'rms_rope_', 'rms_rope_split_half', 'rms_rope_split_half1', 'rms_rope_split_half1_', 'rms_rope_split_half_', 'stochastic_rounding_fp8'], 'disabled': False, 'unavailable_reason': None}, ``` What's going on? Am I installing this wrong somehow? This is what I have in the virtualenv: ```pycon >>> import torch >>> torch.version.__version__ '2.14.0+rocm7.14' >>> torch.version.rocm '7.14.0' ``` I have ROCm 7.14.1 installed on the system level. Is the patch level discrepancy somehow messing with things? I tried installing both from the PyPI and locally, no change. My hardware is gfx1201, for the record.

bsakaguchi closed this issue

2026-10-26 22:19:44 +00:00

bsakaguchi commented

2026-10-26 22:19:46 +00:00

Author

Owner

@agronholm commented on GitHub (Sep 12, 2026):

It looks like it was picking up the APU instead, and export HIP_VISIBLE_DEVICES=0 solved the problem. I can now see sol_attn et al. Sorry for the noise!

@agronholm commented on GitHub (Sep 12, 2026): It looks like it was picking up the APU instead, and `export HIP_VISIBLE_DEVICES=0` solved the problem. I can now see `sol_attn` et al. Sorry for the noise!

bsakaguchi referenced this issue

2026-10-26 22:20:28 +00:00

[PR #53] [MERGED] Add cutlass int8 kernel #108