Encode audio waveform to latent representation. Input: (batch, 1, samples) or (batch, samples) Output: (batch, frames, vae_dim)

Load and run the CLIP-L encoder from a safetensors file. Returns the pooled output (batch_size, 768).

Encode the dialog to ChatML prompt format.

Streaming encode: uses cache for correct context between frames.

Load and run the T5-XXL encoder from a safetensors file. Returns hidden states (batch_size, seq_len, 4096). When `device` is a CUDA GPU with enough V

Encode voice reference from f32 PCM samples (24kHz mono). Returns: (acoustic_features, connected_embeds)

Encode voice reference audio tensor to scaled acoustic features. Returns: (acoustic_features, connected_embeds)

Estimate average layer size in bytes by summing safetensors shard file sizes and dividing by number of layers.

Extract the text inside an open (unclosed) `<think>` block, if any. Returns empty string if there's no open think block.

Remap fm_decoder.encoders.{stack}.layers.{layer}.* to fm_decoder.layers.{flat}.* Also handle fm_decoder.encoders.{stack}.encoder.layers.{layer}.*

FLUX.2-klein transformer configuration. Values derived from `transformer/config.json` of `black-forest-labs/FLUX.2-klein-4B`.

FLUX.2-klein VAE configuration. Values from `vae/config.json` — uses KL autoencoder with 32 latent channels.

Compute (cos, sin) PE for given position IDs [S, num_axes].

Map acoustic VAE latent (batch, vae_dim) → LLM hidden (batch, hidden).

Forward pass through all blocks.

Only used when the provider is StackedResidentProvider (all experts in RAM).

Forward with streaming cache: uses cached context instead of zero-padding.

Optimized forward using pre-computed timestep embedding and projected condition. Caches silu(cond) across blocks (saves 4 silu kernels per step).

attn_mask: optional (1, seq) tensor with 1 for real tokens, 0 for padding

Fused gate+up: (num_experts, 2*intermediate_size, hidden_size)

Run the full TTS pipeline: text -> phonemes -> mel -> waveform.

Map a GGUF tensor name to the HuggingFace convention. GGUF uses `blk.{N}.attn_q.weight` style; HF uses `model.layers.{N}.self_attn.q_proj.weight` sty

clear worker kv cache

Read the GPTQ group_size from config.json (defaults to 128).

Check whether a cache directory contains valid model data for the given layers. For sharded models, verifies that the cached index's weight_map refer

Return the HuggingFace hub cache directory if it exists.

Return a generalized Config object.

Return a generalized Config object.

Return a generalized Config object.

Convert the LLM backbone config to cake's common Config.

Check if the given token ID is an EOS token.

Check whether a model uses FP8 block-wise quantization by looking at its config.

Check whether a path points to a GGUF file.

Returns true if `layer_idx` is a global (no-RoPE, full context) layer. Default: every 4th layer (0-indexed: 3, 7, 11, ...) is global.

Returns true if `layer_idx` is a global (full attention) layer.

Check whether a model uses 4-bit quantization by inspecting its config.json. Detects both standard GPTQ (`quant_method: "gptq"`) and affine 4-bit (`mo

Inverse STFT matching Vocos's ISTFT implementation. Uses irfft semantics (half-spectrum input, norm="backward") with "same" padding.

Manual layer norm (no learned params, eps=1e-6). Always computes and returns in F32.

Compute logits from hidden states. Handles both 2D and 3D inputs.

Load a GGUF file and create a VarBuilder that serves dequantized tensors. All quantized tensors are dequantized to F32 on CPU at load time, then cast

Create a VarBuilder that transparently dequantizes GPTQ 4-bit weights. # Safety Inherits the mmap safety requirements from `MmapedSafetensors`.

Load T5 tokenizer — downloads tokenizer.json from HuggingFace.

Load the tokenizer and resolve EOS token ID(s). `default_eos_token` is the model-specific fallback (e.g. "<|eot_id|>" for LLaMA, "<|endoftext|>" for Q

Create GPTQ-quantized expert tensors in safetensors format. Each expert has qweight (int32 packed), scales (f16), qzeros (int32 packed).

Create CompactRelPositionalEncoding [1, 2*seq-1, pos_dim].

Create CompactRelPositionalEncoding [1, 2*seq-1, pos_dim]. Matches the Python implementation: log compression -> atan -> Fourier.

Create a multi-shard safetensors model directory.

Build VarBuilder for MoE (Qwen3 MoE style: gate + stacked expert weights).

Build VarBuilder for a Qwen3_5 full attention block.

Build VarBuilder for Qwen3.5 MoE (per-expert linear + shared expert + sigmoid gate).