Summary. On GPUs below SM90, infer_helios.py:258 and app.py:34 select diffusers'
flash_hub backend (per PR #28's hardware detection). Since ~June 2026 that path crashes at
the first attention call: the kernels-community/flash-attn2 hub artifact renamed its private
wrappers, and diffusers' registry (stable 0.39.0 and current main) still references the old
_wrapped_flash_attn_* names — tracked upstream as huggingface/diffusers#14012 (open; partial
fix PR #14015 covers only varlen). Net effect: the documented recipes crash on
4090/A6000/A40/L40S-class hardware unless the user knows about --disable_flash_attention.
Environment. Helios @ 8f2a2fa; torch 2.10.0+cu128; diffusers 0.39.0 (also
reproduced on git-main); kernels 0.13.0 (repo pin; also 0.16.0); RTX 4090 (SM89), driver 580.
Reproduction. Run any official SM<90 recipe, e.g.
bash scripts/inference/helios-distilled_t2v.sh on a 4090 → after model load, first
denoising step raises
AttributeError: Kernel module 'flash_attn2_…' does not define attribute path 'flash_attn_interface._wrapped_flash_attn_forward'.
Expected. Generation proceeds (as it did when PR #28 landed in March).
Suggested fixes (either/both, minimal):
- README/FAQ note: on SM<90 add
--disable_flash_attention until diffusers#14012 is fixed
(PyTorch SDPA path works; we measured ~7.3 FPS vs 10.7 FPS compiled-FA3 on H100 for scale).
- In
infer_helios.py, wrap the set_attention_backend("flash_hub") call in try/except with
a logged fallback to SDPA — cheap resilience against exactly this class of hub/registry
drift (the FA3 branch already has a try/except precedent at infer_helios.py:252-255).
Impact. All consumer/workstation GPUs (the community PR #28 explicitly targeted 4090)
crash on default settings; the failure message doesn't mention the existing escape hatch.
Summary. On GPUs below SM90,
infer_helios.py:258andapp.py:34select diffusers'flash_hubbackend (per PR #28's hardware detection). Since ~June 2026 that path crashes atthe first attention call: the
kernels-community/flash-attn2hub artifact renamed its privatewrappers, and diffusers' registry (stable 0.39.0 and current main) still references the old
_wrapped_flash_attn_*names — tracked upstream as huggingface/diffusers#14012 (open; partialfix PR #14015 covers only varlen). Net effect: the documented recipes crash on
4090/A6000/A40/L40S-class hardware unless the user knows about
--disable_flash_attention.Environment. Helios @
8f2a2fa; torch 2.10.0+cu128; diffusers 0.39.0 (alsoreproduced on git-main); kernels 0.13.0 (repo pin; also 0.16.0); RTX 4090 (SM89), driver 580.
Reproduction. Run any official SM<90 recipe, e.g.
bash scripts/inference/helios-distilled_t2v.shon a 4090 → after model load, firstdenoising step raises
AttributeError: Kernel module 'flash_attn2_…' does not define attribute path 'flash_attn_interface._wrapped_flash_attn_forward'.Expected. Generation proceeds (as it did when PR #28 landed in March).
Suggested fixes (either/both, minimal):
--disable_flash_attentionuntil diffusers#14012 is fixed(PyTorch SDPA path works; we measured ~7.3 FPS vs 10.7 FPS compiled-FA3 on H100 for scale).
infer_helios.py, wrap theset_attention_backend("flash_hub")call in try/except witha logged fallback to SDPA — cheap resilience against exactly this class of hub/registry
drift (the FA3 branch already has a try/except precedent at
infer_helios.py:252-255).Impact. All consumer/workstation GPUs (the community PR #28 explicitly targeted 4090)
crash on default settings; the failure message doesn't mention the existing escape hatch.