Problemsby IbisTrouble with Hugging Face generate
MLOps6 posts from 5 people5d active+22 posts in the last 7 days, 0 the 7 days before (rising)
Posts per day
The posts behind it
6, newest firstgenerate() fails with KeyError/AttributeError when input_ids are off the model device### System Info ### Who can help? @zucchini-nlp @Cyrilvallez ### Reproduction Since 5.20.0, `generate()` crashes for encoder-decoder models and for vision-language models that pass `mm_encoder_outputs` whenever the ids driving the loop are on a different device than the model....huggingface/transformers · digit501 repliesOpen ↗prefill_chunk_size drops image inputs, VLMs generate without the image### System Info ### Who can help? @zucchini-nlp ### Reproduction With `prefill_chunk_size` set, a vision-language model never receives the image during prefill. Every chunked forward is called without `pixel_values` or `mm_encoder_outputs`, so the output does not depend on the...huggingface/transformers · digit503 repliesOpen ↗[Whisper] batch_decode(..., decode_with_timestamps=True) merges rows or raises on the tensor returned by generate### System Info - `transformers` main at a96730c8c97b8efbf35bbaf7f5da33ec99231a49 (2026-10-07) - Python 3.10, torch 2.11.0, tokenizers 0.23.1, macOS (arm64); also seen on Linux x86_64 ### Reproduction `processor.batch_decode(..., decode_with_timestamps=True)` works on a list of...huggingface/transformers · tarekziade2 repliesOpen ↗generate() chunked prefill slices the batch axis of multi-axis position_ids (Qwen3-VL)## System Info transformers main (26c0a7fd94), torch 2.9.0, Python 3.10. Fails the same way on CPU and CUDA. ## Who can help? @Cyrilvallez @zucchini-nlp ## Reproduction With `prefill_chunk_size` set, `_prefill` cuts each chunk's position ids with `position_ids[:,...huggingface/transformers · dacorvo1 repliesOpen ↗How can i create file that can generate videoHugging Face Forums · KayswitchOpen ↗[Bug]: DefaultEmbeddingFunction reloads the ONNX model on every generate() call...Only the JS `DefaultEmbeddingFunction` rebuilds it on every call. **Workaround:** subclass `DefaultEmbeddingFunction` and override `generate()` to reuse a module-level pipeline promise. Fix with tests: #7792 ### Versions `@chroma-core/default-embed` 0.1.9 (latest),...chroma-core/chroma · ranjithkumar8352Open ↗
Companies and products named
About this problem
Evidence
6 posts from 5 people in 3 places, about 2 a week over 25 days. Mostly on huggingface/transformers, chroma-core/chroma, Hugging Face Forums. Tools named alongside: GitHub, Chroma, NVIDIA.
Frustration Frustration 0 of 3· Seen on GitHub issues, Community forums
How it was grouped
Posts about Hugging Face with no known issue, sharing the word "generate". First post Sep 23, 2026, latest Oct 17, 2026. Corroborated: 3 or more posts from 2 or more people or places. Method
Other Hugging Face problems
Fine-tuning Hugging Face is costly and unpredictableTrouble with Hugging Face sizeHugging Face users keep filing bugsTrouble with Hugging Face cruiseRunning Hugging Face locally hits memory and speed limitsHugging Face upgrades break working setupsTrouble with Hugging Face docsTrouble with Hugging Face tokenTrouble with Hugging Face trainerTrouble with Hugging Face gemmaTrouble with Hugging Face flightHugging Face integrations break or are missing
History
steady to rising
fading to steady
steady to fading
Trouble with Hugging Face generate (4 posts)