digital-garbageandClaude Opus 5
c454e3b2d9
Locating a phrase's token range runs the text encoder once per phrase — on H3 that is the same multi-billion-parameter encoder the prompt goes through, up to eight times. The ranges exist only to feed the attn2 direction patch, and H3 has no cross-attention for it to hook, so the patch declined on the very next line and the results were dropped. The same question is now asked before the work instead of after it. LTX is unaffected: it reads attn2, so the ranges are still computed. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
47 lines
1.5 KiB
Python
47 lines
1.5 KiB
Python
"""Work whose only consumer has already declined must not be done.
|
|
|
|
Locating a phrase's token range costs one full pass of the text encoder per phrase — on
|
|
MiniMax H3 that is the same multi-billion-parameter encoder the prompt itself goes through.
|
|
The ranges feed the attn2 direction patch, which H3 refuses (no cross-attention). The refusal
|
|
used to happen after the ranges were computed, so every H3 run paid for up to eight encodes
|
|
whose result was dropped on the next line.
|
|
"""
|
|
import sys
|
|
import types
|
|
|
|
import pytest
|
|
|
|
sys.path.insert(0, ".")
|
|
|
|
|
|
@pytest.fixture
|
|
def refiner():
|
|
from conditioning import FunPackVideoRefinerV2
|
|
return FunPackVideoRefinerV2()
|
|
|
|
|
|
def _model(image_model):
|
|
unet = {"image_model": image_model}
|
|
cfg = types.SimpleNamespace(unet_config=unet)
|
|
return types.SimpleNamespace(model=types.SimpleNamespace(model_config=cfg, diffusion_model=None))
|
|
|
|
|
|
def test_h3_does_not_read_attn2(refiner):
|
|
assert refiner._v2_model_reads_attn2(_model("minimax_h3")) is False
|
|
|
|
|
|
def test_an_ltx_model_still_does(refiner):
|
|
assert refiner._v2_model_reads_attn2(_model("ltxav")) is True
|
|
|
|
|
|
def test_no_model_reads_nothing(refiner):
|
|
assert refiner._v2_model_reads_attn2(None) is False
|
|
|
|
|
|
def test_the_range_finder_and_the_patch_agree(refiner):
|
|
"""The early check and the patch's own guard must answer the same question, or the
|
|
ranges get skipped for a model that would have used them."""
|
|
h3 = _model("minimax_h3")
|
|
assert refiner._v2_model_reads_attn2(h3) is False
|
|
assert refiner._v2_apply_model_patches(h3, {}, {}, {}, [], 0.05) is None
|