Unable to use indictrans2-en-indic-1B for a translation project

#18
by SivanKTR - opened

I am trying to develop a translation toolkit using free tools. During this, uncovered these issue. We ran the code provided in the model usage section and it didn't work either.
indic-issue

Environment: Kaggle notebook, Python 3.12, transformers (current release as of Sept 2026 β€” appears to be past the v5 line based on symptoms below), tokenizers==0.22.2, IndicTransToolkit (latest pip package), model: ai4bharat/indictrans2-en-indic-1B.

Summary: Six distinct incompatibilities between IndicTrans2's vendored trust_remote_code files / IndicTransToolkit, and current transformers internals. Individually patchable, but the last two (#5, #6) reveal an unresolvable version conflict: the model code needs an older transformers (pre-Cache-refactor), while the toolkit code needs a newer one β€” no single pinned version satisfies both without patching.

Issue 1 β€” IndicTransToolkit import fails on current transformers

ImportError: cannot import name 'PreTrainedTokenizerBase' from 'transformers.tokenization_utils'

IndicTransToolkit/collator.py does from transformers.tokenization_utils import PreTrainedTokenizerBase. That module used to re-export the class (real home: transformers.tokenization_utils_base) for backward compatibility; current transformers no longer does.

Issue 2 β€” configuration_indictrans.py imports a removed module

ModuleNotFoundError: No module named 'transformers.onnx'

from transformers.onnx import OnnxConfig, OnnxSeq2SeqConfigWithPast β€” transformers.onnx has been removed from current transformers (deprecated long ago in favor of optimum).

Issue 3 β€” Tokenizer constructor ordering breaks under current transformers

AttributeError: IndicTransTokenizer has no attribute _special_tokens_map

tokenization_indictrans.py's IndicTransTokenizer.init sets self.unk_token = ... (and other special tokens) before calling super().init(). Current PreTrainedTokenizerBase.setattr requires self._special_tokens_map to already exist before any special-token attribute assignment β€” it doesn't yet at that point in the custom init.

Issue 4 β€” tie_weights() signature mismatch

TypeError: IndicTransForConditionalGeneration.tie_weights() got an unexpected keyword argument 'recompute_mapping'

Current transformers' PreTrainedModel.init_weights() calls self.tie_weights(recompute_mapping=False). The vendored modeling_indictrans.py overrides tie_weights(self): with no parameters, written before this kwarg existed.

Issue 5 β€” KV-cache format incompatibility (reproduces with AI4Bharat's own official example, unmodified)

TypeError: 'EncoderDecoderCache' object is not subscriptable

Ran the exact generate() example from the ai4bharat/indictrans2-en-indic-1B model card as-is. Current transformers wraps past_key_values in an EncoderDecoderCache object (introduced 4.36+) rather than a plain tuple; the vendored model code still indexes it like a tuple/list, written before this refactor.

Issue 6 β€” No single transformers version satisfies both the model code and the toolkit
Attempted downgrading to transformers==4.35.2 (predates the EncoderDecoderCache introduction in 4.36, which should fix #5). This required also downgrading tokenizers to <0.19 and pulled in huggingface-hub==0.16.4 as a dependency. But then:

ImportError: cannot import name 'pad_without_fast_tokenizer_warning' from 'transformers.data.data_collator'

IndicTransToolkit/collator.py needs pad_without_fast_tokenizer_warning, which doesn't exist in transformers 4.35.2 (added later). So the model code (modeling_indictrans.py) wants an older transformers, while the toolkit code (collator.py) wants a newer one β€” these two pieces of AI4Bharat's own released code have contradictory minimum-version requirements, with no overlapping compatible version.

Suggested fix for maintainers: Either update the vendored modeling_indictrans.py/tokenization_indictrans.py to target current transformers' Cache API and post-4.51 tokenizer internals, or pin/document a specific known-working transformers+tokenizers combination in the model card and IndicTransToolkit's requirements.txt, since right now neither the toolkit nor the model's own trust_remote_code files state a compatible upper/lower bound clearly enough to avoid this.

Sign up or log in to comment