KV cache tensor dimension mismatch in DenoiseBlock/RecacheFramesBlock when target size is zero

[livepeer-robot]

Summary

The LongLive pipeline's DenoiseBlock and RecacheFramesBlock encounter tensor dimension mismatches when the KV cache expansion target has zero tokens in one dimension.

Error Details

Multiple ERROR-level logs from fal.ai backend (12 Feb 2026):

DenoiseBlock errors (11 occurrences):

Error in block: (denoise, DenoiseBlock)
Error details: The expanded size of the tensor (0) must match the existing size (4680) at non-singleton dimension 1. Target sizes: [1, 0, 12, 128]. Tensor sizes: [4680, 12, 128]

RecacheFramesBlock errors (2 occurrences):

Error in block: (recache_frames, RecacheFramesBlock)
Error details: The expanded size of the tensor (0) must match the existing size (18720) at non-singleton dimension 1. Target sizes: [1, 0, 12, 128]. Tensor sizes: [18720, 12, 128]

Analysis

• Target size [1, 0, 12, 128] has 0 tokens in dimension 1
• Existing tensor has 4680 or 18720 tokens (ratio 4x suggests 4 latent chunks)
• The 12 and 128 dimensions appear to be local_attn_size=12 and head dimension
• This suggests the KV cache is being expanded with an invalid/empty token count

Possible Causes

1. Edge case when current_start_frame == 0 but cache isn't properly initialized
2. Race condition between cache initialization and block execution
3. Invalid frame sequence length calculation when resolution parameters are incorrect

Environment

• Pipeline: LongLive (modular blocks)
• Blocks affected: DenoiseBlock, RecacheFramesBlock
• Timestamp: 2026-02-20 07:39-07:40 UTC (cluster of errors from same session)

Logs

All errors originated from app_id d93e99fb-c6d3-4e48-aff2-be59eb51ad5f.

[livepeer-tessa]

Update (Feb 20, 2026 20:33-20:46 UTC): Additional 27 occurrences from the same session:

• DenoiseBlock: 23 new errors with tensor size 2160 tokens (180 spatial positions × 12 local_attn_size)
• RecacheFramesBlock: 4 new errors with tensor size 8640 tokens (720 spatial positions × 12 local_attn_size)

This confirms the issue is reproducible and affects multiple resolution configurations:

Tensor Size	Frame Seq Length	Likely Resolution
2160	180	~12×15 latent (~96×120 pixels)
4680	390	~15×26 latent (~120×208 pixels)
8640	720	~24×30 latent (~192×240 pixels)
18720	1560	~30×52 latent (~240×416 pixels)

All errors still show target size [1, 0, 12, 128] with 0 tokens, indicating the root cause is consistent.