Read an expert weight, handling GPTQ dequantization if needed.

Set the conv state for a linear attention layer.

Directly set the KV cache for a layer (used for voice prompt injection).

Set the recurrent state for a linear attention layer.

Return the cached sine value for the given position and sequence length.

SwooshL(x) = log(1 + exp(x - 4)) - 0.08*x - 0.035 Matches Python's SwooshLForward with offset=4.

Create a WorkerInfo structure to be sent to the master.

Encode as WAV bytes (16-bit PCM).

Convert text to a sequence of token IDs.

Map a u8 wire tag back to a candle DType.

Write f32 data into the buffer via mapped pointer.

Apply partial RoPE: rotate only the first rotary_dim dimensions.

Apply rotary embeddings matching diffusers apply_rotary_emb exactly. x: [B, H, S, D], cos/sin: [S, D]

Map a config.json `architectures` string to our `TextModelArch` enum. Feature-gated variants are only matched when the corresponding feature is enable

Build a synthetic [`discovery::GpuInfo`] list from this node's VRAM/TFLOPS. Used when a topology file specifies nodes without layer assignments, so t

Apply causal depthwise conv1d + SiLU on a single new token, updating conv_state in-place. x: (batch, 1, channels), conv_state: (batch, channels, kerne

Check basic auth if `--ui-auth` is configured. Returns true if OK.

Compute STFT magnitude squared, returned as flat Vec [n_frames * n_freq].

Return the default guidance scale for a given SD version.

Return the default number of sampling steps for a given SD version.

Dequantize a packed 4-bit weight tensor with per-group scales and biases. Used for embeddings in some GPTQ models where the format is: - `weight`: ui

called only with feature = "metal"

Detect the compute backend description for this node.

Detect the CUDA toolkit version by reading version files from the CUDA install directory. Checks `$CUDA_HOME`, `$CUDA_PATH`, and `/usr/local/cuda` fo

Detect total system memory in bytes.

Auto-detect text model architecture from config.json's "architectures" field.

Return a human-readable device type string for a candle Device.

Async download using pinned host memory for faster DMA, then sync.

Encode the dialog to llama3 prompt format.

Estimate TFLOPS from a list of GPUs. Falls back to a VRAM-based heuristic when workers report 0 (old binaries).

Resolve the file path, downloading from HuggingFace if needed. Returns the list of files to download for this component.

Find a model by name. Supports exact match (`org/model`) or suffix match (`model`). Returns `Err` if the suffix match is ambiguous (multiple models ma

First-order DPM-Solver++ update (from timestep s to t, s > t). x_t = (sigma_t/sigma_s) * x - alpha_t * expm1(-h) * x0

Public forward for testing.

Forward with pre-computed silu(cond).

Create an Fp8Linear from a VarBuilder (no bias).

Create an Fp8Linear from a VarBuilder (with bias).

Generate audio from text with a voice prompt.

Get all downloaded file paths for this component.

Create a copy of self with new kv-cache.

Internal: load expert weights without cache lookup.

Generate a Hann window of the given size.

Return the HuggingFace hub cache directory if it exists.

Consume self and create a Tensor, avoiding a copy on CPU by reusing the data buffer.

Return true if this node hosts the specified layer.

Derive the layer name prefix from config.json. Returns e.g. "model.language_model.layers" for Qwen3.5, "model.layers" otherwise.