longcat_next_visual_codec / code_provenance.json

Extract pinned native components; verify source bytes, strict loading and native encode/decode interfaces

f69e160 verified 8 days ago

545 Bytes

{
  "source_repo": "meituan-longcat/LongCat-Next",
  "source_revision": "0cf0631862402ff36366e513e4023d22e7e5c84c",
  "compatibility_base": "Qwen2.5-VL window-attention implementation from the previously pinned/tested Lance release; original notices retained.",
  "changes": [
    "CPU selects eager attention instead of forced FlashAttention",
    "Lazy import of optional image-refiner pipeline",
    "PyTorch segmented-attention fallback for CPU",
    "No tensor mutation during extraction; CPU decoder/VAE runtime casts are explicit."
  ]
}

	{
	"source_repo": "meituan-longcat/LongCat-Next",
	"source_revision": "0cf0631862402ff36366e513e4023d22e7e5c84c",
	"compatibility_base": "Qwen2.5-VL window-attention implementation from the previously pinned/tested Lance release; original notices retained.",
	"changes": [
	"CPU selects eager attention instead of forced FlashAttention",
	"Lazy import of optional image-refiner pipeline",
	"PyTorch segmented-attention fallback for CPU",
	"No tensor mutation during extraction; CPU decoder/VAE runtime casts are explicit."
	]
	}