AI Repos Weekly — May 20–27, 2026

Quick upgrade verdicts

pytorch/pytorch

• Deprecation landed (merged May 26): PR #184984 adds deprecation warnings to torch.quantize_per_tensor, torch.quantize_per_channel, and related functions producing quint8/qint8/qint32 tensors. PR author vkuzo states: "deprecated and will be removed in a future PyTorch release." Migration path tracked at issue #184982. torchvision already dropped the affected calls. 2
• Most-discussed PR (merged May 25): PR #184592 by ezyang replaced all direct sympy.Min/Max construction sites in Inductor with PyTorch's custom torch.utils._sympy.functions.Min/Max classes, eliminating expensive is-connected simplification checks on the hot path. Marked "not user facing" — transparent to end users. 1,696 comments over five days. 3
• Active development areas: Inductor CPP tail-loop vectorization, ROCm MI355 dashboard job, MPS native_group_norm_backward Metal implementation, CUDA sdpa edge-case fix.

langchain-ai/langchain

langchain 1.3.2

langchain-fireworks 1.4.0 / 1.4.1

langchain-openai 1.2.2

vllm-project/vllm

• Deprecation sweep (merged May 27): Commit c02c758 by yewentao256 completed the v0.21.0 deprecation schedule — logit_bias and logit_scale aliases removed from PoolerConfig (use logit_mean / logit_sigma directly), and vllm/utils/profiling.py deleted entirely (66 lines gone). The cprofile decorator and cprofile_context context manager no longer exist; switch to Python's standard cProfile. 10
• Active bug fix (PR #43691, opened May 26): SvenLorenz opened a fix for a race condition in speculative decoding + reasoning models. When the reasoning-end token appears as a rejected draft token, the old code "unconditionally set reasoning_ended=True and force-fed the unconstrained bonus token to the grammar, corrupting its state." Fix adds mid-batch detection and a suppress_accept_errors flag. Under review by WoosukKwon, njhill, mgoin, and others. 11
• Open issues to watch: Prefix-caching + MTP causes ~20% accuracy drop (issue #43559); streaming reasoning tokens truncated when end-of-thinking token coincides with a content delta (issue #43221).

ggml-org/llama.cpp

• Nemotron perf restored (b9330): ffn_latent MUL_MAT flag fix by ServeurpersoCom recovered Nemotron 3 Super 120B Q5_K_M throughput from 64.9 t/s back to 103.22 t/s. 13
• CUDA fast Walsh-Hadamard transform (b9329, PR #23615): Adds GPU-accelerated FWHT for Hadamard-transform-dependent models. b9334 followed up with a PDL sync fix.
• Gemma 4 model support (PR #23682): Gemma4ForCausalLM architecture conversion added.
• WebGPU MMVQ paths (PR #23594): Mixed matrix-vector quantized (MMVQ) paths for Q4/Q8/Q2_K/Q4_K quantization levels added for WebGPU, replacing the older MUL_MAT pipeline.
• ggml bumped to v0.13.0 (ggerganov, May 25). No changelog attached — treat as routine internal bump unless you build against ggml directly.
• Open issues: VRAM growth per run until OOM (issue #23446), Qwen3.6-35B-A3B KV cache losing ~4k tokens per round since b9235 (issue #23589), MTP + quantized KV cache memory leak (issue #23635).

huggingface/transformers

• Cohere2Moe: MoE LLM with sliding-window and full-attention hybrid, combining shared and routed experts.
• HRM-Text: Hierarchical recurrent forward pass with dual-Transformer stack and PrefixLM attention — a new base LLM architecture.
• Parakeet tdt: Speech model.
• Expanded audio support: AudioFlamingoNext checkpoints, audio/visual encoder compilability improvements.
• Fix for lru decorator-induced memory leak in vision models.

• deepseek_v4: hc_head, sinks, and position_bias tensors kept in fp32 (PR #46198) — addresses a silent precision downgrade. Issue #46129 (MTP keys silently random-initialized due to missing conversion_mapping entries) remains open as of May 27. 14
• FSDP2 initialization via from_pretrained now supported (PR #46102).
• GLM-4.6V with GLM-GA Processor merged (PR #46184).
• Issue #46153: ProcessorMixin._load_tokenizer_from_pretrained forces subfolder lookup, breaking root-level tokenizer files — open, tagged deprecation, watch if you use non-standard repo layouts.