draft
40d43055
simplify processor a bit
f3dd8930
reduce modular, 4d pixel values and fix mappings
6f3b9fb8
small iteration
35541104
ternary
b93a00ce
refactor decoder layer forward
d66306fd
refactor block_attention_residual
1c3719bc
add Inkling2BlockAttentionResidual
16676a3b
keep vision encoder in block order
61557170
Merge pull request #3 from huggingface/update-modular
b793ec0a
remove copy?
6ae08c0b
docs + checkerboard testing
82caa46a
fix router bias precision, parallel plans and log-scaling positions
578c68cb
remove shared experts and unused config options
15786c62
remove host syncs from the vision path
55f23276
address review comments on config and modules
8348c659
share test overrides and move the vision parity check out
7066cea0
add a processor test
9257f858
better naming
602c1e1c
Merge branch 'main' into add-inkling2
09c75518
torch.Tensor
dddb06d5
Merge branch 'main' into fast-load
43e78e6c
[inkling2] load the NVFP4 / MXFP8 experts through the quantization plan
fccda543
add Inkling2ScaledL2Clip
cb959ef5
add linear
74cf4ba2
add comment
8aec11b0
update router/moe
29be87a3
fix is_concatenated
88e42122
add Inkling2Softcap
d8ee70b2
remove n_shared_experts
f16c9602
inherit vision config
dd0fa4fd
cleanup image processor
636cecad
processor image token handling
2db59081
split attention classes
b4b6d69f
use norm layer
6be2a57f
update docs
74573ed2
update docstrings
1e767543
enable sdpa for vision tower
3df17228
small improvements
f8bcef31
update config
a97a1515
hardcode missing config values
f461941d
keep block states as tensors
e800588e
update tests
2ed6bd03
[inkling2] keep the convolution weight renaming off `weight_proj`
76c5fd94
Merge branch 'main' into add-inkling2
617ac971
Merge branch 'add-inkling2' into fast-load
968c8791
Merge pull request #4 from huggingface/fast-load
2ad0e4b4
[inkling2] quantization config as glob patterns
6a1e78e3
fix export
05ca9c44
fix bos token
b0580b7f
add tokenizer
a91fc50d
[inkling2] move block_states to each layer's device (device_map)
cc86252e
Merge remote-tracking branch 'upstream/main' into add-inkling2
a38bf7c6
[inkling2] default pad_token to <|endoftext|> (checkpoints don't set …
4d0429be
[inkling2] small integration test: pin logits measured on CUDA with t…
325bf24b
Revert "[inkling2] default pad_token to <|endoftext|> (checkpoints do…
6b00849a
remove tokenizer
cb8e5fa4
small improvement
d687f4be
Merge remote-tracking branch 'upstream/main' into add-inkling2
719dbaa6
update docs
c7f60cf9
Fix onnx export
c1bbc7c3
add CI expectations
832708f7
Merge branch 'main' into add-inkling2
0237409d
small fixes
8e79c4e3
update docs
92b568c9
make it work
226521c5
update docs
1b2f6259
vasqu
approved these changes
on 2026-11-12
vasqu
commented
on 2026-11-12
vasqu
commented
on 2026-11-12
vasqu
commented
on 2026-11-12
cleanup
64ccab4b
rename
a8741344
guarin
merged
109a7d19
into main 58 minutes ago
guarin
deleted the add-inkling2 branch 58 minutes ago
Assignees
No one assigned
Login to write a write a comment.
Login via GitHub