Search results

Silent conversion skip on model_type mismatch: loading a converted checkpoint via the wrong Auto class yields randomly-initialized weights that generate fluently

bug

Status: Open.

#47405 In huggingface/transformers;

· SelfSL opened on Jul 18, 2026

device_map loading OOMs on conversion-op transients: WeightConverter contiguous() runs on destination GPU, unbudgeted by max_memory packing

bug

Status: Open.

#47404 In huggingface/transformers;

· SelfSL opened on Jul 18, 2026

[i18n-<languageCode>] Translating docs to <languageName>

WIPLabel your PR/Issue with WIP for some long outstanding Issues/PRs that are work in progressLabel your PR/Issue with WIP for some long outstanding Issues/PRs that are work in progress

Status: Open.

#47402 In huggingface/transformers;

· diplrcdi-cmd opened on Jul 18, 2026

GitForCausalLM double-shifts the training loss and leaks targets across batch rows (regression from #35875)

Status: Open.

#47394 In huggingface/transformers;

· Yigtwxx opened on Jul 18, 2026

[serge] integration failure triage - 2026-07-17

Status: Open.

#47390 In huggingface/transformers;

· github-actions opened on Jul 18, 2026

[New model] PersonaPlex (full-duplex speech model, based on Moshi)

New model

Status: Open.

#47388 In huggingface/transformers;

· un1876 opened on Jul 17, 2026

Question about overriding bnb_4bit_compute_dtype for pre-quantized checkpoints

Status: Open.

#47378 In huggingface/transformers;

· John6666cat opened on Jul 17, 2026

Two systematic bugs in examples/pytorch/*_no_trainer.py: broken --resume_from_checkpoint auto-discover branch, and wrong eval loss/perplexity on partial batches

Status: Open.

#47375 In huggingface/transformers;

· berkgungor opened on Jul 16, 2026

GlmMoeDsaConfig: legacy num_experts unconditionally overrides explicit n_routed_experts (sibling of #46739)

Status: Open.

#47355 In huggingface/transformers;

· hermia-ai opened on Jul 16, 2026

[Offloading] Cannot save disk-offloaded Qwen/Qwen3-30B-A3B

bug

Status: Open.

#47333 In huggingface/transformers;

· kylesayrs opened on Jul 15, 2026

MetalConfig quantization: batched generate silently corrupts all rows after the first — affine_qmm_t computes only batch element 0 of 3D inputs (verified, one-line fix)

Status: Open.

#47331 In huggingface/transformers;

· fenilsonani opened on Jul 15, 2026

Qwen2.5-Omni Token2Wav DiT uses mismatched RoPE layout (half-split cos/sin with interleaved rotate) — regression from #39847

Status: Open.

#47328 In huggingface/transformers;

· HenryVarro666 opened on Jul 14, 2026