Add Jev + decision core #50054
ArthurZucker
wants to merge
48 commits into
main
from
jev-1-core
Add Jev + decision core
e2cf9479
Use config is_causal=False instead of use_bidirectional_attention for…
00a241fb
Rename *DecisionHead to *ForDecision; move clef-only target capture o…
d0eb7d6a
Decision heads read the model passed to them; the pipeline loads deci…
cdd1a389
Revert the base_model_prefixes call sites in accelerate/deepspeed wit…
76babab3
Simplify the shared code of the decision PR
f88f5717
Simplify the decision code and Jev
0c7f8ef3
encode_marked_chat computes its own marker pattern
df19144a
Merge branch 'jev-1-core-b' into jev-1-core
c5d687d0
Fix test_serve and the question_type_ids doc after the cleanup
09810e8d
Chart of where each head reads
b98199e3
Decision docstring: real <|mark|> tokens in the chart, drop the clef-…
2648ad54
update doc
9d1d0a88
style
ca64b0ae
Model card date
b805e9f5
Tests: one decider per prompt/attention mode, revert unrelated colqwe…
c1b168ce
Decision files next to the model: decision.json, decision_head.json n…
a531266c
JevForDecision: the model and its JevDecisionHead from one repo
15d8f3bd
ForDecision is a PreTrainedModel: config of its head, save_pretrained…
2fa772a9
ForDecision: attention is the model's and the head's
c12199e8
JevForDecision like a ForCausalLM: the model of its repo and the head…
ec0b3fcb
Decision models support the attention implementations of the model th…
175b8eb2
Decision test mixin: the backbone file has every key of the backbone,…
fb82f63d
Decision test mixin: backbone keys under the names of the backbone class
27d48668
No DecisionPreTrainedModel: config_file_name loads decision_head.json…
8a728eab
Merge remote-tracking branch 'origin/main' into jev-1-core
d18d7185
One model_inputs per prompt, batched: ids, marks, questions and the i…
7121b57f
serve reads the decision config of the repo it loads
70cde08b
Jev docs: the model with its files next to the checkpoint, and its us…
e7d4597e
Merge branch 'jev-1-core-fix-serve' into jev-1-core
b2a06557
AutoModelForDecision
50b883ce
Merge branch 'jev-1-core-fix-auto' into jev-1-core
068cb6c9
Decision models read the files of the checkpoint they ship next to: J…
c76f595a
head_files → extra_files
28c4aaa1
Stateless decision template and pipeline: each request carries its qu…
19dbefd7
Decision tests read hf-internal-testing
b182f723
Repo -> JevConfig in __post_init__: extra_config_files read next to t…
e2f8fac7
Decision models like generate: chat_template_kwargs + apply_chat_temp…
e7fd386a
DecisionMixin.from_pretrained reads the checkpoint's head configs; co…
1549208c
Docstring: decide keeps the token ids and marks
f331cd85
decision_config from the head configs already read by checkpoint_conf…
a164abac
DecisionConfig.logits_processors: built once per question type
23104d0f
extra_files: from the config class only, safetensors file names in th…
bfa93e2c
Ponytail: drop dead auto_docstring entries and imports, one process_s…
d2b6d43a
One sanitize check, in _mark_chat_inputs
0a08f615
Ponytail: jev docs on the decide API, one tiny_qwen2 test helper, dro…
f4a96a95
Merge remote-tracking branch 'origin/main' into jev-1-core
1ddd4302
Serve loads decision checkpoints like the pipeline, decision config s…
e25282cb
Login to write a write a comment.
Login via GitHub