Fix BART processor compatibility with vLLM 0.18#20
Merged
Conversation
The multimodal processor in bart.py broke under vLLM 0.18 in three places: - TextDataParser relied on MultiModalDataParser._is_empty, which was removed in 0.18. Replaced with inline emptiness checks for str and list. - create_encoder_prompt previously tokenized `prompt` as the encoder text. In 0.18 `inputs.prompt` passed to this method is the DECODER prompt text, not the encoder text (the encoder content lives in mm_data). The method now returns a single [0] placeholder token; _get_prompt_updates replaces it with the correct number of encoder token slots during rendering. - _call_hf_processor is now sometimes called with an already-tokenized decoder prompt (list[int]) instead of a str. Handle both cases when building result["input_ids"]. Adds tests/test_vllm_018_compat.py with three narrow unit tests covering each of these paths; no GPU required. Signed-off-by: David Schulmeister <dschulmeist@users.noreply.github.com>
dschulmeist
force-pushed
the
v018-bart-fixes
branch
from
3ffb803 to
d3ef529
Compare
dschulmeist
added a commit
to dschulmeist/bart-plugin
that referenced
this pull request
Apr 16, 2026
Adds M2M100ForConditionalGeneration support for the three NLLB distilled translation models: facebook/nllb-200-distilled-600M, 1.3B, and 3.3B. All three share model_type=m2m_100. Architecture differences from BART implemented in nllb.py: - Sinusoidal (fixed) positional embeddings instead of learned - PRE-LayerNorm (norm before sublayer) instead of POST-LayerNorm - Additional layer_norm after all encoder/decoder layers - ReLU activation instead of GELU - No final_logits_bias Language routing: - Decoder starts with target language token via create_decoder_prompt, which resolves the FLORES-200 code (e.g. "fra_Latn") via convert_tokens_to_ids for reliable special-token handling. - Source language token is prepended to the encoder input via src_lang in mm_processor_kwargs (default "eng_Latn"); a make_nllb_prompt helper is provided. Depends on the BART processor vLLM 0.18 compatibility fix (PR vllm-project#20): M2M100MultiModalProcessor inherits create_encoder_prompt from BartMultiModalProcessor and needs the [0] placeholder behavior to function under vLLM >=0.18. Tests: - 12 unit tests (tests/test_nllb_model_structure.py), no GPU required - 13 integration tests (tests/test_nllb_inference.py) covering 4 target scripts, 3 non-English sources, batching, determinism, and max_tokens. All 13 pass on NVIDIA GB10 (DGX Spark) with vLLM 0.18.0. Signed-off-by: David Schulmeister <dschulmeist@users.noreply.github.com>
Signed-off-by: NickLucche <nlucches@redhat.com>
Merged
Member
|
I had to make some changes to your PR due to an output mismatch I was getting, caused by |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Fixes the existing BART multimodal processor for vLLM 0.18 compatibility.
Split out from #19 so it can be reviewed and released independently (as a patch release, per maintainer request).
Changes
Three generic compatibility fixes in
vllm_bart_plugin/bart.py:TextDataParser._parse_text_data: replace the removed_is_emptyhelper (deleted fromMultiModalDataParserin 0.18) with inline emptiness checks.BartMultiModalProcessor.create_encoder_prompt: in 0.18 thepromptargument is the decoder prompt text, not the encoder text. Always return[0]as a placeholder token;_get_prompt_updatesreplaces it with the correct number of encoder token slots.BartMultiModalProcessor._call_hf_processor: the 0.18 rendering pipeline may pass the decoder prompt as an already-tokenizedlist[int]instead of astr. Handle both.Tests
Adds
tests/test_vllm_018_compat.pywith three narrow unit tests covering each path. No GPU required.Notes
The NLLB/M2M100 feature PR (#19) depends on this compatibility fix, since
M2M100MultiModalProcessorinheritscreate_encoder_promptfromBartMultiModalProcessor.