Feat/batch nodes

[lukiod]

No description provided.

[eliteprox]

Can you also add an engine build command to entrypoint.sh? This worked for me:

  # Build Static Engine for Dreamshaper - Square (512x512) - Batch Size 2
  python src/comfystream/scripts/build_trt.py --model /workspace/ComfyUI/models/unet/dreamshaper-8-dmd-1kstep.safetensors --out-engine /workspace/ComfyUI/output/tensorrt/static-dreamshaper8_SD15_\$stat-b-2-h-512-w-512_00001_.engine --width 512 --height 512 --batch-size 2

You can ignore (or even remove) the multi-resolution builds as they are not used anymore (only slows down the script)

comfystream/docker/entrypoint.sh

Lines 96 to 101 in f13bd81

	# Build Static Engine for Dreamshaper - Portrait (384x704)
	python src/comfystream/scripts/build_trt.py --model /workspace/ComfyUI/models/unet/dreamshaper-8-dmd-1kstep.safetensors --out-engine /workspace/ComfyUI/output/tensorrt/static-dreamshaper8_SD15_\$stat-b-1-h-704-w-384_00001_.engine --width 384 --height 704
	# Build Static Engine for Dreamshaper - Landscape (704x384)
	python src/comfystream/scripts/build_trt.py --model /workspace/ComfyUI/models/unet/dreamshaper-8-dmd-1kstep.safetensors --out-engine /workspace/ComfyUI/output/tensorrt/static-dreamshaper8_SD15_\$stat-b-1-h-384-w-704_00001_.engine --width 704 --height 384

[lukiod]

@eliteprox which changes request is this i cant see it

[eliteprox]

Hi @lukiod I apologize for the delay getting back to you. I've tested the latest changes and still had issues with flickering frames. We recently had to fix some linting issues in the main branch, so please rebase and I will take a look.

Again, I apologize for the delay, this has been a lower priority feature since we have been working to ship BYOC pipelines and there were some overlapping changes in this PR.

Thank you again for this contribution, I've added the PR to our current planning cycle to maintain more visibility on this feature, as it is important to our roadmap. If it remains stale, we will pick up the feature in a coming sprint

[lukiod]

rebased. found the likely cause of the flickering while doing it: short batches were padding with a duplicated last frame instead of blocking for a real one. loader now blocks per-frame with the timeout and raises instead of padding, so a batch is always N real frames. also dropped the queue-maxsize bump and a later pipeline-side commit that was fighting your state-manager refactor — kept the PR to LoadTensor's batch_size + SaveTensor's split_batch, no pipeline changes.

heads up: i don't have a GPU streaming setup here so i couldn't actually re-test for flickering, just fixed the mechanism i think caused it. let me know if it's still there.