NVFP4 vision temporal patch merge with projection
FP4-quantized temporal patch merging with linear projection for video processing in Qwen3-VL. Merges temporal patches from video frames by reshaping (B, seq_len, C) -> (B, seq_len//temporal_merge_size, C*temporal_merge_size) and applies FP4-quantized linear projection to hidden_size using block-wise scaling (16 elements per block). The merged_seq_len (M dimension) must be a multiple of 128 for FP4 GEMM compatibility.
Reported evidence · last observed 2026-08-01. Reported by source; not independently reproduced. No public source. License unknown.
Current records
Workloadsuite of 16 cases · mean latencyhidden_size = 3584 · in_channels = 1792 · merged_channels = 3584 · hidden_size_scale = 224 · temporal_merge_size = 2 · merged_channels_half = 1792 · merged_channels_scale = 22416 cases
Not measured on H100 for this workload. Challenges →
131.2µs96% SOL1.00×Reported · license unknown · no source2026-08-01
Measured exactly what you asked. Reported by source; not independently reproduced.
238.2µs95% SOL1.23×Reported · license unknown · no source2026-07-18
Measured exactly what you asked. Reported by source; not independently reproduced.
338.3µs95% SOL1.23×Reported · license unknown · no source2026-08-11
Measured exactly what you asked. Reported by source; not independently reproduced.
438.9µs95% SOL1.25×Reported · license unknown · no source2026-07-13
Measured exactly what you asked. Reported by source; not independently reproduced.
539.0µs95% SOL1.25×Reported · license unknown · no source2026-08-17
Measured exactly what you asked. Reported by source; not independently reproduced.
641.5µs94% SOL1.33×Reported · license unknown · no source2026-08-01
Measured exactly what you asked. Reported by source; not independently reproduced.
742.2µs94% SOL1.35×Reported · license unknown · no source2026-08-03
Measured exactly what you asked. Reported by source; not independently reproduced.
842.4µs94% SOL1.36×Reported · license unknown · no source2026-08-05
Measured exactly what you asked. Reported by source; not independently reproduced.
942.8µs94% SOL1.37×Reported · license unknown · no source2026-07-13
Measured exactly what you asked. Reported by source; not independently reproduced.
1044.2µs94% SOL1.42×Reported · license unknown · no source2026-08-04
Measured exactly what you asked. Reported by source; not independently reproduced.
1146.1µs94% SOL1.48×Reported · license unknown · no source2026-08-03
Measured exactly what you asked. Reported by source; not independently reproduced.
1269.0µs90% SOL2.21×Reported · license unknown · no source2026-07-13
Measured exactly what you asked. Reported by source; not independently reproduced.
13114.3µs84% SOL3.66×Reported · license unknown · no source2026-07-23
Measured exactly what you asked. Reported by source; not independently reproduced.
14117.2µs83% SOL3.76×Reported · license unknown · no source2026-07-13
Measured exactly what you asked. Reported by source; not independently reproduced.
151.20ms32% SOL38.57×Reported · license unknown · no source2026-07-13
Measured exactly what you asked. Reported by source; not independently reproduced.