transformers
Add Qwen4Exp model
#48337
Merged

Add Qwen4Exp model #48337

Cyrilvallez merged 90 commits into main from qwen4-exp
Cyrilvallez
support Qwen4-Exp.
30ea9842
JJJYmmm Fix placement of concatenated checkpoint tensors
a514ed86
JJJYmmm Align Qwen4-Exp checkpoint conversion
2dbfa108
JJJYmmm Simplify Qwen4-Exp model integration
5c92d61b
JJJYmmm Add Qwen4-Exp parallel coverage and documentation
afed628f
JJJYmmm revert some equivalent operation
11284d0a
update qwen4_exp doc.
23ea7559
fix
2b8d53b0
Cyrilvallez Merge remote-tracking branch 'origin/main' into qwen4-exp-clean
e676759a
Cyrilvallez no need to change qwen3.5
6d09b103
Cyrilvallez sets activation directly inside norm
63fe9993
Cyrilvallez remove specific head models
12bb7259
Cyrilvallez native rotary
77937b37
Cyrilvallez simplify
120d28cd
Cyrilvallez remove overwrite
2bf8ff4c
Cyrilvallez remove non compile criteria
13eb28a2
Cyrilvallez move outside the class
84c52bee
Cyrilvallez remove constants
e7d0d009
Cyrilvallez Merge remote-tracking branch 'origin/main' into qwen4-exp-clean
1a62d729
Cyrilvallez simplify config
e9ac87df
Cyrilvallez indexer is always present
33d97e0a
Cyrilvallez indexer needs to train
901a6e13
Cyrilvallez fix and simplify the selection of valid indices in indexer
a2f6038d
Cyrilvallez remove external functions
3abe2c5e
Cyrilvallez simplify indexer further
aaf467ac
Cyrilvallez replace einsum with matmul
412231b7
Cyrilvallez start vectorizing
5642e57d
Cyrilvallez simplify
80f6918f
Cyrilvallez always use a mask for simplicity
f52cbc48
Cyrilvallez create and combine mask much more easily in indexer
49e31770
Cyrilvallez simplify hyper connections
18d9aad7
Cyrilvallez fix all prefixes
ba91022b
Cyrilvallez fix ple input by reversing embeddings instead
891e5443
Cyrilvallez capture hidden states
d18f17e2
Cyrilvallez follow signature order
6c17711b
Cyrilvallez always forward both masks
01377e76
Cyrilvallez fix and simplify past context in ngram embedding
ca261705
Cyrilvallez fix
1c189717
JJJYmmm make do_natural_key more readable
a8114b29
Cyrilvallez combine norm and grouped norm
075d0aed
Cyrilvallez fix shapes
a454b4ac
JJJYmmm fix qsa check
e6db072e
Cyrilvallez fix mask
e9810b4f
Cyrilvallez style
164eabc9
Cyrilvallez add a few decorators
b199d7f3
JJJYmmm update minimal test
75e9095e
JJJYmmm adapt to newest config
1d30f205
JJJYmmm remove unnecessary test
e42de784
JJJYmmm combine related test
ebea6955
JJJYmmm fix index mask broadcast error
c837cc41
JJJYmmm fix gdn previous state check when ple is enable
05bc9504
JJJYmmm avoid update conv state in place
47a975c2
JJJYmmm confirmed hybrid_indexed was not used
6daa4a70
JJJYmmm fix rope test
f6936677
JJJYmmm use vlm tester to pass ci
e3bca939
Cyrilvallez rewrite layer type in config
6163b28e
Cyrilvallez no need to reslice
afc9c2d8
Cyrilvallez short conv up to latest standards
878a16d4
Cyrilvallez dilution not supported in kernels 😭
50aefd63
Cyrilvallez remove unused attr
884d32d3
JJJYmmm vectorize qsa indexer!!!
931ee4ea
JJJYmmm qsa can use a separate rotary for simply
89fe6529
JJJYmmm add qwen_sparse_attention to support block key store!!!
19c8b624
JJJYmmm update test
4888fd1c
Cyrilvallez Merge remote-tracking branch 'origin/main' into qwen4-exp-clean
eedb5c60
Cyrilvallez modular based on latest main
a6d31667
Cyrilvallez make quality happy
6f22d1a7
Cyrilvallez keep simple for now, generalize all library indexers later
807cb38f
Cyrilvallez do not keep cos/sin in the cache for memory
5cd919bd
Cyrilvallez a bit cleaner
1dda1695
Cyrilvallez extend 2d position_ids
b7db0596
Cyrilvallez ignore rule
99250c57
Cyrilvallez simplify conversion Ops
b3bd1cff
Cyrilvallez fix loading and tests
02bcc13f
Cyrilvallez do not use attn in tp_plan cause num_kv_head=2
369c6cee
Cyrilvallez remove created file
1ddbe200
Cyrilvallez oupsi
68559baa
Cyrilvallez fix chunked prefill
4d4d8f25
Cyrilvallez fix positions
b8135c9c
JJJYmmm fix reverse loading
04305920
Cyrilvallez improve conversion
92227c0a
Cyrilvallez slightly clearer key ordering
d807e71a
Cyrilvallez shard the embedding on dim1
59cb4883
Cyrilvallez revert test
72de5567
Cyrilvallez force conversion on cpu for device_map
83b4888b
Cyrilvallez make things work on simple hardware
0b696f40
Cyrilvallez small fix
0682e8ec
Cyrilvallez oupsi
2ff8a4b2
github-actions
Cyrilvallez Merge branch 'main' into qwen4-exp
0cdf3e47
Cyrilvallez latest upstream modular
b61b98be
ExtReMLapin
github-actions
HuggingFaceDocBuilderDev
vasqu vasqu added New model
Cyrilvallez Cyrilvallez merged fc5c5bde into main 2 days ago
Cyrilvallez Cyrilvallez deleted the qwen4-exp branch 2 days ago

Login to write a write a comment.

Login via GitHub

Reviewers
No reviews
Assignees
No one assigned
Labels
Milestone