Class Fp8BlockDequant

java.lang.Object

smile.llm.quant.Fp8BlockDequant

public final class Fp8BlockDequant
extends Object

Dequantizes DeepSeek / Qwen-style fine-grained FP8 weights (weight_block_size = [128, 128] + weight_scale_inv).

Official Qwen3.8-FP8 checkpoints use block scales; Fp8BlockLinear runs them via at::_scaled_mm_v2. This helper dequantizes for reference tests and debug only.

Field Summary

Fields

Modifier and Type	Field	Description
static final int	BLOCK	Qwen / DeepSeek fine-grained FP8 block size.

Method Summary

Modifier and Type	Method	Description
static Tensor	dequant(Tensor weightFp8, Tensor scaleInv, ScalarType outDtype)	Dequantizes weight_fp8 * broadcast(scale_inv) to outDtype.
static boolean	isBlockScale(Tensor weight, Tensor scale)	Returns true when scale is a block layout for weight.
static boolean	isTensorScale(Tensor scale)	Returns true when scale is a scalar / length-1 tensor scale.

Methods inherited from class Object

clone, equals, finalize, getClass, hashCode, notify, notifyAll, toString, wait, wait, wait

Field Details

BLOCK

public static final int BLOCK

Qwen / DeepSeek fine-grained FP8 block size.

See Also:

• Constant Field Values

Method Details

isBlockScale

public static boolean isBlockScale(Tensor weight,
 Tensor scale)

Returns true when scale is a block layout for weight.

Parameters:

weight - FP8 weight [N, K].

scale - candidate scale tensor.

Returns:

true when scale is a [ceil(N/128), ceil(K/128)] block layout for weight shaped [N, K].

isTensorScale

public static boolean isTensorScale(Tensor scale)

Returns true when scale is a scalar / length-1 tensor scale.

Parameters:

scale - candidate scale tensor.

Returns:

true when scale is a scalar / length-1 tensor scale.

dequant

public static Tensor dequant(Tensor weightFp8,
 Tensor scaleInv,
 ScalarType outDtype)

Dequantizes weight_fp8 * broadcast(scale_inv) to outDtype.

Parameters:

weightFp8 - FP8 weight [N, K].

scaleInv - block inverse scales [ceil(N/128), ceil(K/128)].

outDtype - compute dtype (BF16 or FP16).

Returns:

dense weight on the same device as weightFp8 (caller owns).