● Sovereign Hard-Fork · Apple Silicon Native · 1-Bit Ternary
Ultra-Fast Intelligence.
Zero Gatekeeping.
A sovereign downstream distribution of Transformers engineered for extreme inference speed, BitNet ternary quantization, MEM8 wave-interference associative memory, and Python 3.11+ zero-overhead execution.
$ pip install transformers-ultra
SOTA Classroom Training Swarm · Real-Time WebGPU/Canvas Flow
Γ Math (650Hz)
B Code (350Hz)
A Logic (200Hz)
Θ Narrative (50Hz)
"Peter opened three PRs to fix critical bugs in 1-bit ternary models, BitNet quantization, and Apple Silicon MPS memory leaks. Instead of reviewing the math or benchmarking the speedup, an upstream gatekeeper remarked: 'lol this idiot is spamming MY repo' and banned him. Open source belongs to no single gatekeeper. We didn't wait for permission—we took the starting gun, fixed every hot path, integrated native Apple Silicon Metal & BitNet kernels, ported MEM8 wave-interference associative memory, and shipped transformers-ultra."
⚡ Architectural Innovations
Engineered from ground up for maximum throughput and mathematical purity.
Apple Silicon MPS Native
Single-device unindexed Metal Performance Shaders integration eliminating mps:0 crashes, with resilient safetensors buffer fallback staging on macOS unified memory.
MEM8 Wave Interference
Associative recall across 4 cognitive frequency bands (Math Γ, Code B, Logic A, General Θ) operating in $O(1)$ memory without re-evaluating deep transformer layers.
BitNet 1.58-Bit Ternary Packing
Hardware-accelerated bitmask packing and unpacking routines ported from MLX-QUANT and BitNet b1.58 for ultra-compact LLM weights and integer matrix multiplication.
>2.2× Hot Path Acceleration
Eliminated per-token inspect.signature reflection in LogitsProcessorList, dimension-guarded RoPE embeddings, and bypassed torch.cat on empty prefill caches.
Python 3.11+ & Free-Threading
C-slot memory layouts via __slots__, native StrEnum, structural match/case, and atomic synchronization locks prepared for Python 3.13 nogil runtimes.
Hardened Security & Integrity
Automated boundary containment on checkpoint shard weight maps blocking path traversal attacks (#47176) and deterministic generation config precedence (#47752).
📜 Append-Only Sovereign Commit Ledger
Chronological record of every architectural enhancement and upstream bugfix merged into 8b-is/transformers:main.
release(transformers-ultra): v4.58.0.post9 — Nemotron-H use_mamba_kernels flag preservation (#47577) & README / documentation sync
Gated Hub kernel fetches in `NemotronHMamba2Mixer` behind `config.use_mamba_kernels` (#47577), preventing unconditional kernel fetches on CPU/CI runners, updated sovereign `README.md` manifest with Pydantic v2 schemas and prompt injection defense, and verified 30/30 unit tests green.
release(transformers-ultra): v4.58.0.post8 — Pipeline GenerationConfig model precedence (#47752), _is_hf_initialized zero-redundancy init (#47427), CompressedTensors MoE dequant (#47407) & 6 architecture aliases (RWKV-7, MiniCPM4, Chronos-2, Nanbeige4, Talkie, Kimi Linear)
Fixed pipeline `generation_config` parameter resolution to strictly prioritize model modifications (#47752), generalized `_is_hf_initialized` flag checking to avoid redundant normal_ init on non-CUDA accelerators (#47427), added AutoConfig recognition for RWKV-7, MiniCPM4, Chronos-2, Nanbeige4, Talkie, and Kimi Linear, and verified 29/29 tests green.
release(transformers-ultra): v4.58.0.post7 — Special tokens preservation under extra_special_tokens merge (#47838) & chat template prompt injection mitigation (#47822)
Hardened `PreTrainedTokenizerBase` against silent discarding of `additional_special_tokens` when `extra_special_tokens` is also present (#47838). Added `escape_chat_special_tokens` filter to neutralize turn delimiter injections like `<turn|>` and `[INST]` in chat templates (#47822).
release(transformers-ultra): v4.58.0.post6 — Forensic cleanup of dead compatibility shims, modernized strtobool boolean parser with pattern matching
Pruned legacy Python <3.11 version checks and typing fallbacks across `processing_utils.py`, modernized vendored `strtobool` with zero-overhead `match/case` multi-branch boolean resolution in `generic.py`, and verified 23/23 tests passing.
release(transformers-ultra): v4.58.0.post5 — Ultramarathon architecture: Pydantic v2 config ingress & zero-overhead slotted dataclass forward loop
Shipped `UltraGenerationConfigSchema` and `UltraQuantizationConfigSchema` powered by Rust-backed Pydantic v2 for high-speed edge validation, combined with `UltraFastCausalLMOutput` using `@dataclass(slots=True)` for zero `__dict__` overhead and instant structural pattern matching in forward passes.
release(transformers-ultra): v4.58.0.post4 — Apple Silicon first-class MPS cache management, bitsandbytes PEFT allocator warmup resilience & PyPI release
Hardened `caching_allocator_warmup` and `get_total_byte_count` to safely skip non-registered quantizer state (SCB) and PEFT adapter attributes (#47914). Added native Apple Silicon `get_optimal_mac_device()` and `clear_mps_memory_cache()` APIs for zero-overhead Unified Memory tensor management.
release(transformers-ultra): v4.58.0.post3 — Speculative decoding proposal distribution alignment, DistilBERT tied weights torch.export mapping & Apple Silicon Metal low-bit integration
Hardened speculative candidate generation with logits processing to preserve exact proposal distribution `q` (#47932), explicit tied weight mapping for `torch.export` on DistilBERT (#47979), automated base SHA resolution on branch secret scans (#47973), Whisper large-v3 float16 auto-cast (#47805), HuBERT zero-padding conv normalization mask (#47739), and Continuous Batching CUDA SM 8.0+ capability gating (#47926).
fix(whisper,hubert,generation): auto-align whisper-large-v3 input dtype, preserve zero-padding in HuBERT conv embeddings & gate FlashAttention on CUDA SM capability
Auto-aligned float32 input features to float16 conv1 weights in WhisperEncoder (#47805). Propagated padding masks in HubertPositionalConvEmbedding with batch norm to prevent batch cross-contamination (#47739). Gated continuous batching FlashAttention on GPU SM capability (≥ 8.0) to prevent crashes on T4/V100 (#47926).
release(transformers-ultra): v4.58.0.post2 — 14 hardened upstream bugfixes, ASR audio preservation, DynamicCache rollback clamp & PyPI distribution
Fixed stereo ASR audio collapse (#47886), Whisper NaN/Inf audio protection (#47885), early pipeline device validation (#47869), DynamicCache crop rollback clamping (#47433), SentencePiece byte-fallback (#47473), added-tokens encoder cache sync (#47439), tokenizers <0.24.0 security patch (#47429), assisted decoding mid-block EOS termination (#47912), BitNet use_sub_norms opt-out (#47957), Grounding DINO sinusoidal position embeddings (#47674), SigLIP2 vocab size sync (#47612), gauge-independent repetition penalty (#47595), ESMFold fp16 argmax (#47470), and Gemma-4 auto-config architecture mappings (#47448).
feat(runtime): modernize minimum Python baseline to 3.11+ and adopt modern stdlib features
Bumped minimum Python requirement to 3.11+ across setup.py and pyproject.toml. Integrated
__slots__ memory layouts, StrEnum frequency bands, pattern matching (match/case), @ wave interference operator, and thread safety for Python 3.13 free-threading (nogil).
fix: propagate loss kwargs across classification heads and harden autograd imports (#47688)
Forwarded
**kwargs (including num_items_in_batch) to self.loss_function across Generic classification heads, Swinv2, ResNet, CLIP, SigLIP, and SigLIP2, preventing inflated loss during gradient accumulation. Added autograd fallback shims in tensor parallel modules.
fix(mps): add safe fallback for safetensors loading on Apple Silicon MPS
Ensured safe_open falls back to mmap/cpu buffer staging if direct pread/mps mapping is unsupported, preserving zero-copy memory efficiency across unified memory architecture.
fix(mps): correctly instantiate unindexed mps device on Apple Silicon
Fixed
torch.device("mps") in pipelines/base.py to prevent RuntimeError: Invalid device string 'mps:0' when running on macOS Metal architectures.
fix: patch sharded checkpoint path traversal (#47176) & generation config precedence (#47752)
Secured
get_checkpoint_shard_files against path traversal attacks attempting to escape model directories. Fixed generation config priority so explicit model attributes take precedence over pipeline defaults.
fix(modeling,tokenization): protect get_total_byte_count against un-registered quantized attrs (#47914) & preserve additional_special_tokens (#47838)
Fixed caching allocator warmup AttributeError during PEFT quantized loading. Prevented silent erasure of additional special tokens when extra special tokens are present.
feat(transformers-ultra): add device-native PyTorch BitNet packing and RoPE dimension guards across Granite, OLMo, Bamba, Qwen3, Solar, EXAONE, and Nomic
Integrated device-native BitNet ternary packing and extended RoPE dimension guards to 7 additional model families for accelerated forward passes.
perf(core): eliminate inspect.signature reflection in LogitsProcessorList, add RoPE unsqueeze guards to Gemma/Starcoder, and optimize DynamicIndexedLayer
Replaced runtime signature reflection with cached dispatch in token generation loops and integrated MEM8 wave memory engine.