Encode prompts in BF16 without stalling on references, and cut references to the generated length (0.2.4)

[KBRASK]

Prompts with reference images or videos could stop responding during encoding on 16 GB cards. ComfyUI's text-encoder base ran the H3 Qwen3-VL encoder in FP32, and on Windows a long reference pushed the encode into shared GPU memory, where it ran many times slower. This runs the encoder in BF16 like the official pipeline, keeps it in dedicated GPU memory, and cuts references to the generated length as the official pipeline does. Release 0.2.4.

BF16, as in the official pipeline

• The language model gets BF16 embeddings, DeepStack features and masks instead of FP32, so the NVFP4 weights no longer run as FP32 matmuls.
• The BF16 output is kept as it is; FreeVideo stores conditioning in BF16 anyway.

Memory planned from the real input

• Room is planned from the sequence the language model really sees: each image or video block expands to its patches / 4.
• It is checked against the dedicated memory actually free after loading. On Windows this is also bounded by the DXGI budget, which cudaMemGetInfo overstates.
• Encoder weights move back to host memory until the forward fits.
• Inputs too long for one forward run the language model in blocks, without a dense T×T mask.
• DeepStack features are joined once instead of once per image, and each vision block returns its cache. Without expandable segments, 33 blocks had grown reserved memory from 12.7 to 26.5 GiB.

Never in shared GPU memory

• On Windows a watcher stops a forward whose allocations outgrow the dedicated room, within seconds.
• The encode retries by itself with more room or in blocks.
• What each forward used is remembered per machine and per kind of input, from the latest forwards only.

References cut to the generated length, and said so

• A reference video or audio clip contributes as much as the video being generated, at most 15 s, as in the official pipeline and ComfyUI's native H3 node.
• Before, clips over 15 s were refused and shorter ones were used whole. A 15 s reference on a 5 s video gave the model 2.5× the vision tokens and 2.8× the VAE frames and transformer reference rows the official pipeline would.
• The media card, the generation progress and the result say which clip was cut and how much was used. A reused input cache reports it too.

Measured on an RTX 5060 Ti 16 GB (Windows, 832×480, 5 s, the same long prompt; prompt encoding and the encoder's PyTorch peak)

Input	0.2.3	This PR
Text only	22.7 s, 13.9 GiB	20.0–21.6 s, 13.4 GiB
1 reference image	34.5 s, 15.0 GiB	22.8 s, 12.6 GiB
2 reference images	37.0 s, 15.2 GiB	22.3 s, 12.4 GiB
3 reference images	41.2 s, 15.4 GiB	24.0 s, 12.3 GiB
14.5 s reference video + 3 images	Stopped responding for over 10 min	30.2 s, 11.5 GiB
Two 15 s videos + 3 images, 1344×768	Out of memory 3 times	Completes