Repository navigation

Issues

New issue

Transformers' version 5 is out!
#43489 · LysandreJik opened on Jan 26, 2026
8
Tell Us: What Would Make Trainer Better?
#43595 · SunMarc opened on Jan 29, 2026
16

[serge] integration failure triage - 2026-10-14
huggingface/transformers#49620 · opened on Oct 14, 2026
SAM 3: text-only prompts drop the geometry CLS token that Meta's implementation always adds
huggingface/transformers#49592 · opened on Oct 14, 2026
2
import transformers spends 1-3 s in an eager importlib.metadata.packages_distributions() when torch >= 2.14 is installed (Python 3.12)
huggingface/transformers#49591 · opened on Oct 14, 2026
3
Tokenizer loses its normalizer from tokenizer.json in 5.18
huggingface/transformers#49578 · opened on Oct 13, 2026
SDPA: enable_gqa=True silently falls back to the math kernel at batch 1 on torch < 2.14 without Flash Attention
huggingface/transformers#49564 · opened on Oct 13, 2026
2
[Bug] WhisperTimeStampLogitsProcessor crashes with TypeError when eos_token_id is a single-element list
huggingface/transformers#49556 · opened on Oct 13, 2026
7
/v1/systemone: 500s for chat_template_kwargs overrides and for models without a chat template
huggingface/transformers#49553 · opened on Oct 13, 2026
1
CPU SDPA could use native GQA with a mask on torch >= 2.9 (1.2x decode, no repeat_kv)
huggingface/transformers#49546 · opened on Oct 12, 2026
3
Gemma 3/4 multimodal models fail with tp_plan="auto" because of mixed Tensor/DTensor usage
huggingface/transformers#49540 · opened on Oct 12, 2026
4
Mamba2: non-fused gated RMSNorm ignores n_groups and disagrees with fused training
huggingface/transformers#49536 · opened on Oct 12, 2026
MM Grounding DINO and LLMDet: decoder box and class heads are tied to layer 0 although the config says they are not shared
huggingface/transformers#49529 · opened on Oct 12, 2026
Assigning rope_scaling drops existing RoPE base parameters
huggingface/transformers#49519 · opened on Oct 12, 2026
HunYuan-VL beam search gives returned sequences the wrong image for batched prompts
huggingface/transformers#49499 · opened on Oct 11, 2026
9
SynthID-Text: CPU and MPS produce different sampling tables with the same seed
huggingface/transformers#49498 · opened on Oct 11, 2026
4
ZeRO-3: from_pretrained reads and converts the full checkpoint on every rank
huggingface/transformers#49497 · opened on Oct 11, 2026
1
Discussion: configurable decoder loops with shared parameters and independent execution caches
huggingface/transformers#49488 · opened on Oct 10, 2026
1
Chunked prefill mishandles prefix-cache continuation
huggingface/transformers#49483 · opened on Oct 10, 2026
4
Torch fallbacks of the linear-attention ops drop seq_idx / cu_seqlens, so packed documents leak without the optional packages
huggingface/transformers#49472 · opened on Oct 9, 2026
4
AutoVideoProcessor.register treats positional exist_ok as a backend dictionary
huggingface/transformers#49469 · opened on Oct 9, 2026
Mamba2 / Zamba2: cached decoding does not respect the time-step limits used by sequence scan
huggingface/transformers#49465 · opened on Oct 9, 2026
3
Trainer checkpoint rotation can delete valid checkpoints when backup directories match the checkpoint prefix
huggingface/transformers#49448 · opened on Oct 9, 2026
1
Trainer disables find_unused_parameters under any gradient checkpointing, not just reentrant
huggingface/transformers#49443 · opened on Oct 8, 2026
Gemma 4 response_template fails on tool names containing - or .
huggingface/transformers#49422 · opened on Oct 8, 2026
2
Parakeet TDT greedy decoding ignores max_symbols_per_step: unbounded token repeats on one encoder frame
huggingface/transformers#49388 · opened on Oct 7, 2026
2
Wav2Vec2 and 6 other audio models leave masked_spec_embed uninitialized when the checkpoint lacks it
huggingface/transformers#49385 · opened on Oct 7, 2026

Navigation Menu

Repository navigation

Issues

Issues

[serge] integration failure triage - 2026-10-14

SAM 3: text-only prompts drop the geometry CLS token that Meta's implementation always adds

import transformers spends 1-3 s in an eager importlib.metadata.packages_distributions() when torch >= 2.14 is installed (Python 3.12)

Tokenizer loses its normalizer from tokenizer.json in 5.18

SDPA: enable_gqa=True silently falls back to the math kernel at batch 1 on torch < 2.14 without Flash Attention

[Bug] WhisperTimeStampLogitsProcessor crashes with TypeError when eos_token_id is a single-element list

`/v1/systemone`: 500s for `chat_template_kwargs` overrides and for models without a chat template

CPU SDPA could use native GQA with a mask on torch >= 2.9 (1.2x decode, no repeat_kv)

Gemma 3/4 multimodal models fail with tp_plan="auto" because of mixed Tensor/DTensor usage

Mamba2: non-fused gated RMSNorm ignores n_groups and disagrees with fused training

MM Grounding DINO and LLMDet: decoder box and class heads are tied to layer 0 although the config says they are not shared

Assigning rope_scaling drops existing RoPE base parameters

HunYuan-VL beam search gives returned sequences the wrong image for batched prompts

SynthID-Text: CPU and MPS produce different sampling tables with the same seed

ZeRO-3: from_pretrained reads and converts the full checkpoint on every rank

Discussion: configurable decoder loops with shared parameters and independent execution caches

Chunked prefill mishandles prefix-cache continuation

Torch fallbacks of the linear-attention ops drop `seq_idx` / `cu_seqlens`, so packed documents leak without the optional packages

AutoVideoProcessor.register treats positional exist_ok as a backend dictionary

Mamba2 / Zamba2: cached decoding does not respect the time-step limits used by sequence scan

Trainer checkpoint rotation can delete valid checkpoints when backup directories match the checkpoint prefix

Trainer disables find_unused_parameters under any gradient checkpointing, not just reentrant

Gemma 4 `response_template` fails on tool names containing `-` or `.`

Parakeet TDT greedy decoding ignores `max_symbols_per_step`: unbounded token repeats on one encoder frame

Wav2Vec2 and 6 other audio models leave masked_spec_embed uninitialized when the checkpoint lacks it