Add DeepSeek V4 #45643
ArthurZucker
merged
44 commits into
main
from
add-deepseek-v4
ArthurZucker
commented
on 2026-04-27
ArthurZucker
commented
on 2026-04-28
Add DeepSeek V4 (modular)
26c62d03
Split V4 HCA / CSA caches and compressors into independent classes
8b3c91ab
Fix tests_generate / tests_tensor_parallel CI failures
b4b3a202
0hujun
commented
on 2026-04-29
Address PR review feedback batch (comments 2-24)
6baa6534
Address PR review feedback batch (comments 25-42)
838cc0cd
Fix Fp8Dequantize.reverse_op to actually re-quantize on save
f18a6b8e
Address Arthur's review batch + revisit two of vasqu's comments
921a8dc0
ArthurZucker
commented
on 2026-04-29
Apply suggestions from code review
6d82332c
Move DeepseekV4Config out of modular + simplify __post_init__
4ee8e479
Fix V4 TP failures: dynamic num_key_value_groups + FP8-safe GroupedLi…
ea29922b
Revert GroupedLinear F.linear loop, keep bmm
092dcd62
Fix V4 GroupedLinear comment with real V4-Flash / V4-Pro config values
def7199e
up
631151d4
up
fa0369fd
small cleanup
5f711e01
repo fixes
b5998a0f
nits
dcbe66bf
more nits
aecf3192
nits, small thing left to do
4fc49a03
update
02356fec
update DeepseekV4HCACache
8124d3e6
more update
45c43230
nits
a48b51b6
update
46bc5642
Merge branch 'add-deepseek-v4' of github.com:huggingface/transformers…
240647ad
Merge branch 'add-deepseek-v4' of github.com:huggingface/transformers…
5eb7bec3
update
8f807c19
nits
0c62b587
update
e32e9a84
fixes
5b125c67
nits
07aca9ff
Merge branch 'main' of https://github.com/huggingface/transformers in…
913adcc4
Fix CI: \N>1 backrefs in conversion mapping + drop irrelevant drift
7b5b54a2
Drop TP plan from V4 attention path; keep EP-only on experts
7feb6a15
Address vasqu review batch + skip quantized-TP test
d515ec5d
Drop model. prefix from V4 conversion mapping + tester / parity cleanup
65e564f4
Silence Unrecognized rope_parameters keys warning
a021e849
Disable flex_attention on V4; guard tensor-only mask pad
0e35cfd4
Switch V4 to EP-only plan (gpt-oss style)
a55ffee8
Fix V4 yarn rope validation under nested rope_parameters
d86d46d1
Trim V4 config-attribute allowlist from 24 → 6
ff720b31
Inline V4 config legacy-kwarg + resolve_* helpers into __post_init__
c9460879
Disable FlashAttention on V4: head_dim=512 exceeds the 256 cap
62a75c01
update
08e4cf82
yiakwy-xpu-ml-framework-team
commented
on 2026-05-08
yiakwy-xpu-ml-framework-team
commented
on 2026-05-08
yiakwy-xpu-ml-framework-team
commented
on 2026-05-08
Reviewers
Assignees
No one assigned
Labels
New model
Milestone
No milestone
Login to write a write a comment.
Login via GitHub