Class FlashCce

constructor

• new FlashCce(options?: FlashCceOptions): FlashCce

  Parameters

  • options: FlashCceOptions = DEFAULT_FLASH_CCE

  Returns FlashCce

Readonlyoptions

backward

• backward(
      hResp: MlxArray,
      head: FlashCceHead,
      targets: number[],
      lse: MlxArray,
      cot: number[],
      filterEps?: string,
      blockMax?: MlxArray,
      blockEps?: string,
  ): MlxArray

  Flash-CCE backward: dh [M, hidden] (f32) = ∂(Σ_t cotangent_t · logp_t)/∂h. lse is the forward's saved log-sum-exp [M]; cot is the per-token cotangent ∂loss/∂logp_t [M]. Recomputes logits in-kernel (no [M,V]). filterEps and blockEps default to the composed epsilons; pass "0" for guaranteed-exact gradients (the parity gate does). The block skip fires only with the forward's blockMax. Caller owns the result.

  Parameters

  • hResp: MlxArray
  • head: FlashCceHead
  • targets: number[]
  • lse: MlxArray
  • cot: number[]
  • filterEps: string = ...
  • OptionalblockMax: MlxArray
  • blockEps: string = ...

  Returns MlxArray

forward

• forward(
      hResp: MlxArray,
      head: FlashCceHead,
      targets: number[],
  ): { blockMax: MlxArray; logp: MlxArray; lse: MlxArray }

  Flash-CCE forward over hResp [M, hidden]: logp [M], the log-sum-exp [M] the backward reuses, and the per-vocab-block maxima [M, NBLK] that drive its block skip. The caller owns all three.

  Parameters

  • hResp: MlxArray
  • head: FlashCceHead
  • targets: number[]

  Returns { blockMax: MlxArray; logp: MlxArray; lse: MlxArray }