Add test for deepseek_vl #1136
Add test for deepseek_vl #1136
Conversation
meenakshiramanathan1
requested review from
vkovinicTT,
mstojkovicTT,
nvukobratTT,
pilkicTT and
dgolubovicTT
as code owners
meenakshiramanathan1
force-pushed
the
mramanathan/deepseek
branch
from
d680aab to
07bb2f6
Compare
|
||||||||||||||||||
|
||||||||||||||||||
meenakshiramanathan1
force-pushed
the
mramanathan/deepseek
branch
from
07bb2f6 to
b774e54
Compare
|
||||||||||||||||||
1 similar comment
|
||||||||||||||||||
meenakshiramanathan1
force-pushed
the
mramanathan/deepseek
branch
from
b774e54 to
42bf67b
Compare
|
||||||||||||||||||
3 similar comments
|
||||||||||||||||||
|
||||||||||||||||||
|
||||||||||||||||||
meenakshiramanathan1
force-pushed
the
mramanathan/deepseek
branch
4 times, most recently
from
eccc9ef to
9baabd6
Compare
|
||||||||||||||||||
1 similar comment
|
||||||||||||||||||
meenakshiramanathan1
force-pushed
the
mramanathan/deepseek
branch
3 times, most recently
from
a9a7143 to
563b03e
Compare
|
||||||||||||||||||
2 similar comments
|
||||||||||||||||||
|
||||||||||||||||||
|
||||||||||||||
meenakshiramanathan1
force-pushed
the
mramanathan/deepseek
branch
3 times, most recently
from
adfe853 to
273ee1b
Compare
| return framework_model, vl_gpt, tokenizer, inputs_embeds | ||
|
|
||
|
|
||
| def generation(max_new_tokens, model, inputs_embeds, tokenizer, vl_gpt): |
There was a problem hiding this comment.
Great work @meenakshiramanathan1 . Can we reuse this script for other generation models like t5, llama, and more? BTW, we can expect a conflict related to vl_gpt used here.
cc: @nvukobratTT May I know your thoughts in this?
meenakshiramanathan1
force-pushed
the
mramanathan/deepseek
branch
from
273ee1b to
d75cca4
Compare
meenakshiramanathan1
force-pushed
the
mramanathan/deepseek
branch
from
d75cca4 to
8373b96
Compare
Model doesn't have huggingface direct version, hence model implementation is from corresponding github repository. |
meenakshiramanathan1
force-pushed
the
mramanathan/deepseek
branch
2 times, most recently
from
b3431b2 to
20c7194
Compare
meenakshiramanathan1
requested review from
nvukobratTT
and removed request for
pilkicTT and
dgolubovicTT
|
||||||||||||||||||
3 similar comments
|
||||||||||||||||||
|
||||||||||||||||||
|
||||||||||||||||||
…ps config (#1231) The [generate model ops test pipeline](https://github.com/tenstorrent/tt-forge-fe/actions/runs/13328380954/job/37226649520) is currently freezing during the unique ops configuration extraction phase Error: `Failed on "DecomposeEinsum" TVM callback` This error is encountered in the test case: `forge/test/models/pytorch/vision/detr/test_detr.py::test_detr_segmentation[facebook/detr-resnet-50-panoptic]` test cases. To prevent the extraction process from hanging indefinitely, a timeout of 1200 seconds (20 minutes) has been added. This ensures that if the unique ops configuration extraction takes too long, the test will be terminated.
meenakshiramanathan1
force-pushed
the
mramanathan/deepseek
branch
from
20c7194 to
76b1781
Compare
|
||||||||||||||
|
||||||||||||||
1 similar comment
|
||||||||||||||
|
||||||||||||||
There was a problem hiding this comment.
As this is a lot of new code that is mostly copied from GitHub, let me see can we get new repository for this purpose, and then link it to ours.
E.g. have a repo that will contain all third-party GitHub repositories as submodules, from which we can load this and similar models to run compile against them. That way, in FFE we'll only store model tests + wrappers if some modification is needed.
What do you think @meenakshiramanathan1 ?
Sure @nvukobratTT , I will keep this in draft for now. |
That sound good @nvukobratTT . We can do that. As we are going to focus on P1 models and models supported in tt-torch which is not yet supported in tt-forge-fe. What will be the priority of moving all the GitHub copied code to a new repo? Because there are some other models also fall in this case as you know. Could you please clarify this? |
Synced offline |
Model doesn't have huggingface direct version, hence model implementation is from corresponding github repository.
Model requires timm version>=0.9.16 which was conflicting with segmentation_models_pytorch package so both versions have been upgraded to resolve conflicts.
The custom generate function generates text by iteratively predicting tokens while maintaining a fixed sequence length using padding.
The model input shape is kept static by preallocating padded_inputs_embeds, which is updated each step
with new token embeddings corresponding to next_token_id until the sequence is complete.
Model is currently failing in verify with pcc drop of 0.96.