[daichu-nv]

Automatically pack eligible BF16 Q/K/V tensors directly into cuDNN's MXFP8 payload and scale layouts. The CuTe kernels remove intermediate layout copies while preserving the native quantization arithmetic. Dispatch selects the fast path on SM100 for supported shapes and strides; unsupported inputs keep the native implementation. No configuration flag is required.

The full-model comparison uses Wan2.2 A14B with the official static-NVFP4 checkpoint on one B200: 480×832, 33 frames, 40 steps, seed 42, three prompts, CUDNN MXFP8 attention, compilation on and CUDA graphs off. Both arms use TensorRT-LLM 1.3.0rc27.dev202609150000 in the same process.

Full-generation latency	Native	Candidate	Reduction
Mean of prompt medians	21.57 s	20.05 s	7.06%

Validation: byte-level payload/scale and cuDNN consumer checks passed; production GPU tests: 34 passed and one expected SM103 hardware skip. Worst-prompt LPIPS across all 33 frames: 0.0 (limit 0.25).

[YihuiLu512]

Heads-up: #20214 moved visual_gen/attention_backend/ to visual_gen/attention/backends/; the old package is now a deprecated shim that only forwards package-root exports.

This PR merges cleanly, but the result breaks: your cudnn.py edits follow the rename to attention/backends/cudnn.py, while the new mxfp8_pack.py lands at the old attention_backend/mxfp8_pack.py. As a result, from .mxfp8_pack import ... raises ImportError, and pre-merge CI won't catch it because it only tests the PR head.

Could you rebase and move mxfp8_pack.py to tensorrt_llm/_torch/visual_gen/attention/backends/?

[daichu-nv]

Let me close this PR first.