已合并
docs: restore WeightNz FP8 support notes #5874
Hu1L1创建于 6月9日
docs: restore WeightNz FP8 support notes #5874
已合并
共 2 个文件变更+4-0
| @@ -537,6 +537,7 @@ aclnnStatus aclnnQuantMatmulWeightNz( | |||
| 537 | | INT8 | INT8 | null | FLOAT32/BFLOAT16 | null | null | null/INT32/FLOAT32/BFLOAT16| BFLOAT16 | | 537 | | INT8 | INT8 | null | FLOAT32/BFLOAT16 | null | null | null/INT32/FLOAT32/BFLOAT16| BFLOAT16 | |
| 538 | | INT8 | INT8 | null | FLOAT32/BFLOAT16 | null | null | null/INT32 | INT32 | | 538 | | INT8 | INT8 | null | FLOAT32/BFLOAT16 | null | null | null/INT32 | INT32 | |
| 539 | | HIFLOAT8 | HIFLOAT8 | null | UINT64/INT64 | null | null | null/FLOAT32 | FLOAT16/BFLOAT16/FLOAT32 | | 539 | | HIFLOAT8 | HIFLOAT8 | null | UINT64/INT64 | null | null | null/FLOAT32 | FLOAT16/BFLOAT16/FLOAT32 | |
| 540 | + | FLOAT8_E4M3FN | FLOAT8_E4M3FN | null | UINT64/INT64 | null | null | null/FLOAT32 | FLOAT16/BFLOAT16/FLOAT32 | | ||
| 540 | 541 | ||
| 541 | - T-T量化场景下，x1Scale传入nullptr，x2Scale的shape为(1,)。 | 542 | - T-T量化场景下，x1Scale传入nullptr，x2Scale的shape为(1,)。 |
| 542 | - T-C量化场景下，x1Scale传入nullptr，x2Scale的shape为(n,)，其中n与x2的n一致。 | 543 | - T-C量化场景下，x1Scale传入nullptr，x2Scale的shape为(n,)，其中n与x2的n一致。 |
| @@ -60,6 +60,8 @@ namespace ge { | |||
| 60 | scale is (batch, ceil(k / 128), ceil(n / 128)), where 128 is group_size_k, group_size_n(refer to group_size). \n | 60 | scale is (batch, ceil(k / 128), ceil(n / 128)), where 128 is group_size_k, group_size_n(refer to group_size). \n |
| 61 | - When x2 format is NZ and the data type of x1 and x2 is hifloat8, | 61 | - When x2 format is NZ and the data type of x1 and x2 is hifloat8, |
| 62 | scale must be uint64 or int64. \n | 62 | scale must be uint64 or int64. \n |
| 63 | + - When x2 format is NZ, the data type of x1 and x2 is float8_e4m3fn | ||
| 64 | + and quantization mode is pertensor-perchannel/pertensor-pertensor, scale must be uint64 or int64. \n | ||
| 63 | * @li offset: An optional matrix tensor, quantization parameter. Must be one of the following types: float32. | 65 | * @li offset: An optional matrix tensor, quantization parameter. Must be one of the following types: float32. |
| 64 | supports ND format. The shape is 1D (t,), with t equal to 1 or n, where n is the same as that of x2. | 66 | supports ND format. The shape is 1D (t,), with t equal to 1 or n, where n is the same as that of x2. |
| 65 | * @li bias: An optional matrix tensor. Must be one of the following types: int32, bfloat16, float16, float32, supports ND format. | 67 | * @li bias: An optional matrix tensor. Must be one of the following types: int32, bfloat16, float16, float32, supports ND format. |
| @@ -131,6 +133,7 @@ namespace ge { | |||
| 131 | * - when transpose_x2 is true, n must be greater than 1. | 133 | * - when transpose_x2 is true, n must be greater than 1. |
| 132 | * @li Only weight supports ND and NZ format on Ascend 950 AI Processor. All other inputs and outputs only support ND format. | 134 | * @li Only weight supports ND and NZ format on Ascend 950 AI Processor. All other inputs and outputs only support ND format. |
| 133 | * @li When x2 is NZ format and input type of x1 and x2 is hifloat8, scale must be uint64/int64. | 135 | * @li When x2 is NZ format and input type of x1 and x2 is hifloat8, scale must be uint64/int64. |
| 136 | +* @li When x2 format is NZ, the data type of x1 and x2 is float8_e4m3fn and quantization mode is pertensor-perchannel/pertensor-pertensor, scale must be uint64 or int64. | ||
| 134 | * @li The following are the supported data type combinations by platform. | 137 | * @li The following are the supported data type combinations by platform. |
| 135 | 138 | ||
| 136 | * - Atlas Inference Series Product: | 139 | * - Atlas Inference Series Product: |