Accesses a bit within the predicate vector.

Called at the start of one step before starting accumulator exchange

Build a half activation tensor.

Build w1 (gate+up fused along N) and w2 (down) in fp8 + per-expert per-tensor scales. Layouts match Triton / CUTLASS / sglang convention:

Clears the predicates

Clears the predicate set efficiently

Performs an expected transaction bytes decrement without doing an arrive operation

Computes predicates based on internally tracked per-thread offset.

Computes integer division and modulus using precomputed values. This is computationally inexpensive. Simply returns the quotient

Clears the predicate set efficiently

Obtains calling threadblock's linear threadblock index

Obtains calling linear threadblock index of the first block to work on the given tile

Obtains iteration extends for the given SK block index

Gets the mask

Obtains tile index for the given sk iteration

Wait until we have at least one completed global fetch stage

Loads a fragment from memory at the location pointed to by the iterator.

Loads a fragment from memory at the location pointed to by the iterator.

Loads a fragment from memory at the location pointed to by the iterator.

Loads a fragment from memory at the location pointed to by the iterator.

Loads a fragment from memory with additional logical offset

Loads a fragment from memory

Load

Loads a fragment from memory with additional logical offset

Loads a fragment from memory with additional logical offset

Loads a fragment from memory with additional logical offset

Loads a fragment from memory

Gets the index of the problem

Set data SMEM ptr

Overrides the internal iteration index

Sets the predicate mask, overriding value stored in predicate iterator

Entry of the subtile load loop. Aux loads usually performed here Upon entry the producer acquire of the current subtile lock has completed. Upon exit

Stores a fragment to memory

Stores a fragment to memory at the location pointed to by the iterator

Store a fragment to memory

The entire warp must call this function collectively (that is, the instructions are aligned)

Whether this block will perform the first iteration of this tile

Whether this block will perform the first iteration of this tile

LUT x4 _e2m1_to_half_x4: Direct E2M1->FP16 (converts 4 E2M1 to 4 FP16 in one call)

Create random paged FP8 data with new-path K-scale (4D fp32) + V-scale (per-head).

Adds a pointer offset in units of Element

Adds a pointer offset in units of Element

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

Adds a tile offset

Performs an arrive operation + expected transaction bytes increment