🚨 Exporters: decompose generative models into one decode graph #50252
Cache: a cross-attention layer type, and shaped empty tensors on lazy…
858e915b
Generation and masking: export-safe causal skip and padding, a non-sl…
53207635
Keep VibeVoice's own static-cache override; no attribute-name paramet…
2f4ff6e1
T5Gemma2: drop its cross-attention cache override for the base one
661f914b
Merge remote-tracking branch 'origin/cache-cross-attention-layer' int…
d8b5652a
Merge remote-tracking branch 'origin/generation-masking-export' into …
d1dbc5d0
Exporters: take the generation decomposition's cache, masking and vis…
3c7b654d
Exporters: decompose generative models into one decode graph
733203f0
Exporters: format the ExecuTorch runner
7cfeaefb
Merge branch 'exporters-multimodal-runtime' into exporters-generation…
39cf5d5b
Merge branch 'exporters-multimodal-runtime' into exporters-generation…
ac8749f9
Exporters: keep a greedy generation config saveable after completing …
dc359c68
Merge branch 'exporters-multimodal-runtime' into exporters-generation…
4b8a9387
Merge branch 'exporters-multimodal-runtime' into exporters-generation…
c387d668
Merge branch 'exporters-multimodal-runtime' into exporters-generation…
427f7408
Exporters: let the generation decode graph take deepstack features
289d20a5
Merge branch 'exporters-multimodal-runtime' into exporters-generation…
d1332663
Merge branch 'exporters-multimodal-runtime' into exporters-generation…
3d9cbc60
Merge branch 'exporters-multimodal-runtime' into exporters-generation…
831e5c03
Merge branch 'exporters-multimodal-runtime' into exporters-generation…
810d75d3
Merge branch 'exporters-multimodal-runtime' into exporters-generation…
b62e4908
Exporters: decode a token at a time for linear attention that switche…
78d52c58
Exporters: size a captured static cache for the generation its config…
26d9a4cf
Merge branch 'exporters-multimodal-runtime' into exporters-generation…
93c1f7f9
Merge branch 'exporters-multimodal-runtime' into exporters-generation…
2b8d4295
Assignees
No one assigned
Login to write a write a comment.
Login via GitHub