fix(ltx2): wire conv_in to correct widths in LTX2VideoUpBlock3d

[Solaris-star]

Summary

Fixes #14307

Two independent defects in LTX2VideoUpBlock3d.__init__ made any non-nominal decoder_block_out_channels unloadable:

1. conv_in projected onto out_channels, but the upsampler expects out_channels * upscale_factor.

The upsampler is built with in_channels=out_channels * upscale_factor, so conv_in must project onto that same width. The guard should compare in_channels != out_channels * upscale_factor.

2. LTX2VideoDecoder3d divided output_channel by the current block's upsample_factor to compute input_channel.

The tensor arriving from the previous block is output_channel wide — no division needed. The division produced a fictitious width corresponding to no tensor in the graph.

Both defects are invisible on released LTX-2 checkpoints because the nominal decoder_block_out_channels happen to make conv_in unnecessary and the division a no-op.

Tests

Added tests/models/autoencoders/test_autoencoder_kl_ltx2.py with 3 tests:

• test_conv_in_projects_to_upsampler_input_width — non-nominal widths trigger conv_in and forward pass succeeds
• test_no_conv_in_when_widths_match — nominal case still works without conv_in
• test_conv_in_with_no_upsampler — no-upsampler case works correctly

$ pytest tests/models/autoencoders/test_autoencoder_kl_ltx2.py -v
3 passed in 0.89s

[ErenAta16]

Reproduced #14307 and confirmed both halves of the fix on CPU with a tiny AutoencoderKLLTX2Video, plus checked the nominal path is byte-for-byte unchanged.

Three decoder width patterns, latent_channels=4, upsample_factor=(2,2,2), decode() on a (1, 4, 3, 8, 8) latent:

decoder_block_out_channels	conv_in count	before	after
(16, 32, 64) — nominal, each block 2× the last	0	decode OK (1, 3, 17, 64, 64)	same
(32, 32, 32) — constant width	2	RuntimeError weight of size [16, 8, 3, 3, 3], expected input[1, 16, ...]	decode OK (1, 3, 17, 64, 64)
(64, 32, 16) — decreasing	2	RuntimeError weight of size [16, 4, 3, 3, 3], expected input[1, 8, ...]	decode OK (1, 3, 17, 64, 64)

So the conv_in branch is unreachable on the nominal pattern (0 built) and broken on every other one, exactly as the issue describes.

On the regression risk, since the second hunk changes input_channel for all blocks rather than just the conv_in ones: I diffed every decoder.up_blocks.* parameter shape in the nominal configuration before and after. 18 parameters, all identical. That's the part I'd have wanted checked before merging, and it holds.

The three new tests in tests/models/autoencoders/test_autoencoder_kl_ltx2.py pass (3 passed).

One note on the second hunk. The old input_channel = output_channel // upsample_factor[i] is only equal to the incoming width when block_out_channels[i-1] // upsample_factor[i-1] == block_out_channels[i] // upsample_factor[i] * upsample_factor[i] — i.e. exactly the nominal doubling pattern with a constant factor. Worth saying in the PR body, because it explains why the released checkpoints load: they sit on the one pattern where the wrong expression coincides with the right one.

Measured on main @ 9f6fc2c with the PR applied, torch CPU, Python 3.10.12.