SUMIN/HuggingFace_transformer
1
0
Fork 0
Code Issues Pull Requests Actions 7 Packages Projects Releases Wiki Activity

Support input_embeds in torch exportable decoders (#39836)

Browse Source 
* Support input_embeds in torch exportable decoders

* Hybrid cache update

* Manually change some callsites

* AI changes the rest of the call sites

* Make either input_ids/inputs_embeds mandatory

* Clean up

* Ruff check --fix

* Fix test

* pr review

* Revert config/generation_config changes

* Ruff check
This commit is contained in:
Jack
2025-08-07 01:51:31 -07:00
committed by GitHub
parent cdeaad96b7
commit 6121e9e46c
11 changed files with 325 additions and 85 deletions
Download Patch File Download Diff File Expand all files Collapse all files

5
tests/models/gemma/test_modeling_gemma.py
View File

@@ -460,7 +460,10 @@ class GemmaIntegrationTest(unittest.TestCase):
from transformers.integrations.executorch import TorchExportableModuleForDecoderOnlyLM
exportable_module = TorchExportableModuleForDecoderOnlyLM(model)
exported_program = exportable_module.export()
exported_program = exportable_module.export(
input_ids=prompt_token_ids,
cache_position=torch.arange(prompt_token_ids.shape[-1], dtype=torch.long, device=model.device),
)
ep_generated_ids = TorchExportableModuleWithStaticCache.generate(
exported_program=exported_program, prompt_token_ids=prompt_token_ids, max_new_tokens=max_new_tokens
)