Check if libnuma is available and functional.

Parse a Krasis .conf file into a dict.

Parse user@host or user@host:port into an SSH destination and port.

Construct A Single Stage's Accumulator Shape

After reduce call, before smem async fence. Smem stores usually performed here. Upon exit, all smem stores for TMA must have been issued

Issue Tma Descriptor Prefetch -- ideally from a single thread for best performance

Before visit callback. Smem broadcasts usually performed here. Upon entry, all producer loads for this subtile are completed and visible.

Accesses raw internal state

Render messages with the chat template and show the result.

Validate an explicit HQQ sidecar manifest without treating it as a cache manifest.

Select HQQ promotions by budgeted normalized-error reduction. Uses a bounded-resolution 0/1 knapsack. The unit size is derived from the reque

Free server request state (KV cache pages, etc.). Also handles single-slot AWQ: restore Marlin data into GPU slots for instant prefil

Stores a fragment

After smem async fence, before TMA store commit. Aux stores usually performed here Upon exit, all TMA stores for this subtile must have been issued Be

Get the total VRAM on a device (in MB).

Computes SM number the thread is running on

Compute VRAM/RAM budget from current config.

Dense MLP: gate_proj + up_proj → SiLU(gate) * up → down_proj.

Recurrent decode: process tokens using recurrent formulation. Matches HF torch_recurrent_gated_delta_rule exactly.

Gated RMSNorm: rmsnorm(x) * silu(gate). Matches HF Qwen3NextRMSNormGated: norm first, then multiply by silu(gate).

Load a prompt file bundled with the package.

Apply +1 only to norm roles that use the delta-RMS contract.

Resolve selected GPU indices to GPU dicts from hardware info.

Re-register LA state pointers after prefill (states may have been reallocated). Prefill may reassign _conv_state and _recurrent_state tensors,

Adds a pointer offset in units of Element

Adds a pointer offset in units of Element

Adds a tile offset

Advances an iterator along logical dimensions of matrix in units of whole tiles

Record that num_tokens were appended to the cache.

Case mapping for trace-gated layer-0 input-row captures.

Resolve the cache directory for a model: `~/.krasis/cache/<model_folder_name>/`. Falls back to `<model_dir>/.krasis_cache/` if HOME is not set.

< Efficiently disables all accesses guarded by mask

Returns whether the current work_tile_info passed in should continue to be used. This occurs only in the stream-K decomposition with stream-K work uni

Detect expert sublayer: "mlp" (standard) or "mixer" (Nemotron).

Detect the native group_size from a pre-quantized model's weight_scale dimensions.

Detect shared expert naming: "shared_experts" (DeepSeek) vs "shared_expert" (QCN). Returns the substring to use in weight name construction.

Byte offset of expert data in the mmap'd file.

Returns the biased exponent

Generate the Marlin weight permutation table for INT4. Returns a 1024-element array mapping destination → source index within a 16×64 tile. Matches v

Generate the Marlin weight permutation table for INT8. Returns a 1024-element array mapping destination → source index within a 16×64 tile. Matches v

Returns a pointer

Return if the address in on the diagonal

Get expert weights for a given MoE layer index and expert index. moe_layer_idx is 0-based within MoE layers (not absolute layer index).

Get A extents. fprop: A extents array contains [N,D,H,W,C]. Turn that into ((W,H,D,N), (C)) dgrad: A extents array contains [N,Z,P,Q,K]. Turn that int

Get B extents. fprop: B extents array contains [K,T,R,S,C]. Turn that into ((K), (C,S,R,T)) dgrad: B extents array contains [K,T,R,S,C]. Turn that int

Computes the grid size given a chosen threadblock shape