SolitaryThinker
6096b00aeb
The full joint text->video+audio LTX-2.3 path now generates on the real 18.99B model: - GPU audio components: build_torch_audio_vae (AudioDecoderLoader -> LTX2AudioDecoder, chains the vocoder) + build_torch_vocoder (VocoderLoader -> LTX2Vocoder); registered the 'audio_vae'/'vocoder' cuda component kinds; stamped their checkpoint subfolders (_WAN21_SUBFOLDERS). - Fix: TorchGemma.encode_av must pass output_hidden_states=True — the 2.3 connector's SEPARATE audio projection lives in hidden_states[0] only then (gemma.py:703); without it the audio text fell back to the video embedding (4096 vs 2048 -> audio cross-attn shape mismatch). - GPU-verified T2VS: video (3,33,256,384, std 0.68) + audio (stereo 2x61920 @24kHz, nonzero, std 0.059). - test_torch_backend: cuda component kinds now include audio_vae + vocoder. CPU suite green (202+2).
110 lines
5.2 KiB
Python
110 lines
5.2 KiB
Python
"""GPU torch backend — the CPU-verifiable tier (design_v3 §17).
|
|
|
|
The torch adapters/kernels themselves need a GPU box (see GPU_BRINGUP.md). What IS verifiable on a
|
|
torchless laptop is the wiring that must be right for the GPU path to even resolve correctly: the
|
|
cells are registered with honest sources, importing the backends never imports torch, a cuda
|
|
platform resolves the REAL cells (not a silent toy fallback), and building without torch fails loudly
|
|
rather than quietly returning a toy. Those are exactly the bugs that would otherwise surface only on
|
|
the box — so they're worth pinning here.
|
|
"""
|
|
from __future__ import annotations
|
|
|
|
import importlib.util
|
|
import sys
|
|
|
|
import pytest
|
|
|
|
from v2.card import load_card
|
|
from v2.models.wan21 import build_wan21_card
|
|
from v2.platform import (
|
|
COMPONENTS,
|
|
FLOW_MATCH_STEP,
|
|
FLOW_SDE_STEP,
|
|
KERNELS,
|
|
Platform,
|
|
component_matrix,
|
|
ensure_backends_loaded,
|
|
kernel_matrix,
|
|
)
|
|
|
|
|
|
def _make_cuda_available(monkeypatch):
|
|
"""Flip every registered cuda cell to 'available' (auto-restored), so resolution can be tested
|
|
without torch actually present."""
|
|
ensure_backends_loaded()
|
|
for reg in (*COMPONENTS._reg.values(), *KERNELS._reg.values()):
|
|
if reg.key[1] == "cuda":
|
|
monkeypatch.setattr(reg, "available", lambda: True)
|
|
|
|
|
|
# --------------------------------------------------------------------------- #
|
|
# Registration + honest sources #
|
|
# --------------------------------------------------------------------------- #
|
|
def test_cuda_has_all_components_and_two_solver_ops():
|
|
comps = sorted(r["kind"] for r in component_matrix() if r["device"] == "cuda")
|
|
# + upsampler (LTX-2 spatial), audio_vae + vocoder (LTX-2.3 T2VS audio branch)
|
|
assert comps == ["audio_vae", "dit", "text_encoder", "upsampler", "vae", "vocoder"]
|
|
ops = sorted(r["op"] for r in kernel_matrix() if r["device"] == "cuda")
|
|
assert ops == ["flow_match_step", "flow_sde_step"]
|
|
|
|
|
|
def test_cuda_solver_source_is_honest_no_fake_kernel():
|
|
"""fastvideo-kernel ships NO fused solver kernel — the cuda solver is plain torch. The source
|
|
must say so, not claim a 'fastvideo-kernel:flow_*' that does not exist."""
|
|
for r in kernel_matrix():
|
|
if r["device"] == "cuda":
|
|
assert "torch elementwise" in r["source"]
|
|
assert "fastvideo-kernel" not in r["source"]
|
|
assert r["arch"] == "generic" # elementwise: arch-agnostic, no static scratch
|
|
|
|
|
|
def test_cuda_cells_unavailable_on_this_box():
|
|
assert all(not r["available"] for r in component_matrix() if r["device"] == "cuda")
|
|
assert all(not r["available"] for r in kernel_matrix() if r["device"] == "cuda")
|
|
|
|
|
|
# --------------------------------------------------------------------------- #
|
|
# The load-bearing guard: importing the backends never imports torch #
|
|
# --------------------------------------------------------------------------- #
|
|
@pytest.mark.skipif(
|
|
importlib.util.find_spec("torch") is not None,
|
|
reason="torch installed (GPU box): the cuda-availability probe imports torch by design to call "
|
|
"torch.cuda.is_available(); this no-torch-import invariant is only verifiable without torch.")
|
|
def test_loading_backends_does_not_import_torch():
|
|
ensure_backends_loaded()
|
|
component_matrix(); kernel_matrix()
|
|
assert "torch" not in sys.modules
|
|
assert "v2.platform.backends.torch_adapters" not in sys.modules # imported only inside builders
|
|
assert "v2.platform.backends.torch_kernels" not in sys.modules
|
|
|
|
|
|
# --------------------------------------------------------------------------- #
|
|
# Resolution: when cuda IS available, the REAL cells win (no silent toy) #
|
|
# --------------------------------------------------------------------------- #
|
|
def test_cuda_platform_resolves_real_cells_not_toy(monkeypatch):
|
|
_make_cuda_available(monkeypatch)
|
|
p = Platform.cuda("sm90")
|
|
for kind in ("dit", "vae", "text_encoder"):
|
|
reg = COMPONENTS.resolve_first(p._component_keys(kind, "default"))
|
|
assert reg is not None and reg.key[1] == "cuda", kind # the cuda adapter, not the cpu toy
|
|
reg = p.resolve_kernel(FLOW_MATCH_STEP)
|
|
assert reg.key[1] == "cuda" and "torch elementwise" in reg.source
|
|
assert p.resolve_kernel(FLOW_SDE_STEP).key[1] == "cuda"
|
|
|
|
|
|
def test_cuda_build_fails_loudly_without_torch(monkeypatch):
|
|
"""With cuda 'available' but torch absent, building a component must RAISE (the lazy import of the
|
|
torch adapter fails) — never silently hand back a numpy toy decoding real latents."""
|
|
_make_cuda_available(monkeypatch)
|
|
inst = load_card(build_wan21_card(), cache_manager=None, platform=Platform.cuda("sm90"))
|
|
with pytest.raises((ImportError, ModuleNotFoundError, RuntimeError)):
|
|
inst.component("transformer")
|
|
|
|
|
|
# --------------------------------------------------------------------------- #
|
|
# The checkpoint seam (risk A) is additive — toy cards are unaffected #
|
|
# --------------------------------------------------------------------------- #
|
|
def test_checkpoint_field_defaults_empty_on_toy_cards():
|
|
card = build_wan21_card()
|
|
assert all(spec.checkpoint == "" for spec in card.components.values()) # CPU toys need no weights
|