transformers

361 issues found

[New Model] Add Microsoft Samba - Hybrid SSM + Sliding Window Attention

5/25/2026

[New Model] Add TIGER — Recommender Systems with Generative Retrieval (Google DeepMind, NeurIPS 2023)

5/25/2026

cohere2_moe fails training + tensor parallel tests

5/25/2026

Qwen3.5 GatedDeltaNet: Large logit divergence between full-sequence forward and prefill+decode with cache

5/25/2026

return_tensors is silently ignored when text_kwargs is explicitly passed

5/25/2026

HIGH: Pickle model loading (torch.load) enabled by default — no warning when loading untrusted weights

5/25/2026

MEDIUM: All CI jobs use secrets:inherit — full org secrets passed to every reusable workflow

5/25/2026

Roundtrip Failure for Gemma Pipeline on "▁"

5/25/2026

ValueError: Backend should be defined in the BACKENDS_MAPPING. Offending backend: tensorflow_text

5/24/2026

[New model] Add Fun-ASR-Nano (FunAudioLLM/Fun-ASR-Nano-2512)

5/24/2026

ImportError / TypeError on Windows with AMD ROCm PyTorch due to torch.distributed dependency

5/23/2026

[BUG] deepseek-v4 `comb.to(dtype).transpose(-1, -2)`

5/23/2026

[Gemma4] `Gemma4VisionPatchEmbedder._position_embeddings` materializes a ~19 GiB one-hot tensor that's mathematically a 2-row embedding lookup

5/23/2026

[Feature Proposal] Cognitive Weighting Protocol: Hierarchical context management to mitigate "Contextual Drift" in long-running sessions.

5/22/2026

Regression in v5.x: `ProcessorMixin._load_tokenizer_from_pretrained` forces subfolder for non-primary sub-tokenizers, breaking repos that put tokenizer files at root

5/22/2026

[RT-DETRv2] MPS crash: build_2d_sinusoidal_position_embedding hardcodes torch.float64, breaking Apple Silicon / MPS inference

5/22/2026

Unable to resume model training with a locally downloaded model and trust remote code

5/22/2026

[deepseek_v4] save_pretrained silently downcasts FP32 tensors to BF16 (hc_*, attn_sink, ffn.gate.bias, compressor.ape, indexer.compressor.ape)

5/23/2026

Discussion: optional RankSEG-style decoding for Transformers semantic segmentation post-processing

5/21/2026

`**kwargs` not passed through methods of RoFormer models

5/21/2026

Data

Security

Network

Code

UI/UX

Text

System

Multimedia

Protocol

API

Engineering

transformers

[New Model] Add Microsoft Samba - Hybrid SSM + Sliding Window Attention

[New Model] Add TIGER — Recommender Systems with Generative Retrieval (Google DeepMind, NeurIPS 2023)

cohere2_moe fails training + tensor parallel tests

Qwen3.5 GatedDeltaNet: Large logit divergence between full-sequence forward and prefill+decode with cache

return_tensors is silently ignored when text_kwargs is explicitly passed

HIGH: Pickle model loading (torch.load) enabled by default — no warning when loading untrusted weights

MEDIUM: All CI jobs use secrets:inherit — full org secrets passed to every reusable workflow

Roundtrip Failure for Gemma Pipeline on "▁"

ValueError: Backend should be defined in the BACKENDS_MAPPING. Offending backend: tensorflow_text

[New model] Add Fun-ASR-Nano (FunAudioLLM/Fun-ASR-Nano-2512)

ImportError / TypeError on Windows with AMD ROCm PyTorch due to torch.distributed dependency

[BUG] deepseek-v4 `comb.to(dtype).transpose(-1, -2)`

[Gemma4] `Gemma4VisionPatchEmbedder._position_embeddings` materializes a ~19 GiB one-hot tensor that's mathematically a 2-row embedding lookup

[Feature Proposal] Cognitive Weighting Protocol: Hierarchical context management to mitigate "Contextual Drift" in long-running sessions.

Regression in v5.x: `ProcessorMixin._load_tokenizer_from_pretrained` forces subfolder for non-primary sub-tokenizers, breaking repos that put tokenizer files at root

[RT-DETRv2] MPS crash: build_2d_sinusoidal_position_embedding hardcodes torch.float64, breaking Apple Silicon / MPS inference

Unable to resume model training with a locally downloaded model and trust remote code

[deepseek_v4] save_pretrained silently downcasts FP32 tensors to BF16 (hc_*, attn_sink, ffn.gate.bias, compressor.ape, indexer.compressor.ape)

Discussion: optional RankSEG-style decoding for Transformers semantic segmentation post-processing

`**kwargs` not passed through methods of RoFormer models

transformers

[New Model] Add Microsoft Samba - Hybrid SSM + Sliding Window Attention

[New Model] Add TIGER — Recommender Systems with Generative Retrieval (Google DeepMind, NeurIPS 2023)

cohere2_moe fails training + tensor parallel tests

Qwen3.5 GatedDeltaNet: Large logit divergence between full-sequence forward and prefill+decode with cache

return_tensors is silently ignored when text_kwargs is explicitly passed

HIGH: Pickle model loading (torch.load) enabled by default — no warning when loading untrusted weights

MEDIUM: All CI jobs use secrets:inherit — full org secrets passed to every reusable workflow

Roundtrip Failure for Gemma Pipeline on "▁"

ValueError: Backend should be defined in the BACKENDS_MAPPING. Offending backend: tensorflow_text

[New model] Add Fun-ASR-Nano (FunAudioLLM/Fun-ASR-Nano-2512)

ImportError / TypeError on Windows with AMD ROCm PyTorch due to torch.distributed dependency

[BUG] deepseek-v4 `comb.to(dtype).transpose(-1, -2)`

[Gemma4] `Gemma4VisionPatchEmbedder._position_embeddings` materializes a ~19 GiB one-hot tensor that's mathematically a 2-row embedding lookup

​[Feature Proposal] Cognitive Weighting Protocol: Hierarchical context management to mitigate "Contextual Drift" in long-running sessions.

Regression in v5.x: `ProcessorMixin._load_tokenizer_from_pretrained` forces subfolder for non-primary sub-tokenizers, breaking repos that put tokenizer files at root

[RT-DETRv2] MPS crash: build_2d_sinusoidal_position_embedding hardcodes torch.float64, breaking Apple Silicon / MPS inference

Unable to resume model training with a locally downloaded model and trust remote code

[deepseek_v4] save_pretrained silently downcasts FP32 tensors to BF16 (hc_*, attn_sink, ffn.gate.bias, compressor.ape, indexer.compressor.ape)

Discussion: optional RankSEG-style decoding for Transformers semantic segmentation post-processing

`**kwargs` not passed through methods of RoFormer models

[Feature Proposal] Cognitive Weighting Protocol: Hierarchical context management to mitigate "Contextual Drift" in long-running sessions.