[object Object]

Note: This API will be deprecated in later versions. Use the latest API instead.

[object Object][object Object]undefined[object Object]

• Adapts to the [object Object] operator in the decode ([object Object]) and prefill ([object Object]) inference scenarios.

  Compared with FusedInferAttentionScore, this API adds the keyAntiquantScaleOptional, keyAntiquantOffsetOptional, valueAntiquantScaleOptional, valueAntiquantOffsetOptional, keySharedPrefixOptional, valueSharedPrefixOptional, actualSharedPrefixLenOptional, keyAntiquantMode, and valueAntiquantMode parameters.

  Note: KV cache specific to the decode scenario: KV cache is a common technology for optimizing the inference performance of foundation models. During sampling, the transformer model uses the given prompt/context as the initial input for inference (parallel processing supported), and then generates additional tokens one by one to improve the generated sequence (reflecting the auto-regressive property of the model). The transformer performs the self-attention operation during sampling. Therefore, KV vectors need to be extracted for each item (regardless of the prompt/context or generated token) in the current sequence. These vectors are stored in a matrix called KV cache.

• Formula

  Self-attention constructs an attention model by leveraging the relationships within the input samples. The principle assumes an input sample sequence $x$ of length $n$, where each element of $x$ is a $d$-dimensional vector. Each $d$-dimensional vector can be regarded as a token embedding. Such a sequence is transformed by three weight matrices to produce three matrices of $n*d$.

  The calculation formula for self-attention is generally defined as follows, where $Q$, $K$, and $V$ are key attribute elements of the input sample, obtained through spatial transformation and unified into a single feature space. "Attention" in the formula and operator name is an abbreviation for "self-attention."

  $$Attention(Q,K,V)=Score(Q,K)V$$

  The score function in this operator employs the softmax function. The self-attention calculation formula is as follows:

  $$Attention(Q,K,V)=Softmax(\frac{QK^T}{\sqrt{d}})V$$

  The product of $Q$ and $K^T$ represents the attention to the input $x$. To prevent this value from becoming excessively large, it is typically scaled by dividing by the square root of $d$, followed by row-wise softmax normalization. The result is then multiplied by $V$ to produce an $n*d$ matrix.

[object Object]

Each operator has calls. First, [object Object] is called to obtain the input parameters and compute the required workspace size based on the process. Then, [object Object] is called to perform computation.

[object Object]

• Parameters

  [object Object]

• Returns

  [object Object]: status code. For details, see .

  The first-phase API implements input parameter verification. The following errors may be thrown.

  [object Object]

[object Object]

• Parameters

  [object Object]

• Returns

  [object Object]: status code. For details, see .

[object Object]

• Deterministic computation

  • [object Object] defaults to a deterministic implementation.

• When this API is used together with PyTorch, ensure that the CANN package versions match the PyTorch package versions.

• Processing logic for a null input parameter: The operator checks whether [object Object] is a null pointer. If so, an error is reported. If [object Object] is not an empty tensor but [object Object] and [object Object] are empty tensors (that is, S2 is 0), [object Object] is filled with all zeros. When attentionOut is an empty tensor, the AscendCLNN framework will process it. For other input parameters which support the passing of null pointers as described in the preceding parameter description, no processing is performed when they are null pointers.

• The shapes of the tensors corresponding to the [object Object] and [object Object] parameters must be identical. In non-contiguous scenarios, the batch size in the tensor lists of [object Object] and [object Object] can only be 1, and the number of elements must be equal to the batch size (B) of [object Object]. The N and D dimensions must be the same. Due to the tensor list restrictions, B cannot be greater than 256 in non-contiguous scenarios.

• Restrictions on [object Object] and [object Object]

  • [object Object]Atlas A2 training products/Atlas A2 inference products[object Object]: The data type can be FLOAT16, BFLOAT16, INT8, or INT4 (INT32).
  • Ascend 950PR/Ascend 950DT: The data type can be FLOAT16, BFLOAT16, INT8, HIFLOAT8, FLOAT8_E4M3FN, INT4(INT32) or FLOAT4_E2M1.

• Restrictions on [object Object]

  • If [object Object] is not 1, [object Object] must be of type FLOAT16 or INT8 when [object Object] is of type FLOAT16, and [object Object] must be of type BFLOAT16 when [object Object] is of type BFLOAT16. The input shape must be (B, N, Q_S, KV_S) or (1, N, Q_S, KV_S), where [object Object] is the S dimension in the shape of [object Object], and [object Object] is the S dimension in the shapes of [object Object] and [object Object]. For scenarios where [object Object] of [object Object] is not 32-byte aligned, it is recommended that the value be padded to 32 bytes to improve performance. There is no requirement on the padding value of the extra part.
  • If [object Object] is 1, [object Object] must be of type FLOAT16 when [object Object] is of type FLOAT16, and [object Object] must be of type BFLOAT16 when [object Object] is of type BFLOAT16. The input shape must be (B, N, 1, KV_S) or (1, N, 1, KV_S), where [object Object] is the S dimension in the shapes of [object Object] and [object Object]. For scenarios where [object Object] of [object Object] is not 32-byte aligned, it is recommended that the value be padded to 32 bytes to improve performance. There is no requirement on the padding value of the extra part.

• [object Object]: masks the QK product to define the attention visibility between tokens.

  • If [object Object] is not 1, the recommended input shapes are (Q_S, KV_S), (B, Q_S, KV_S), (1, Q_S, KV_S), (B, 1, Q_S, KV_S), and (1, 1, Q_S, KV_S).
  • If [object Object] is 1, the recommended input shapes are (B, KV_S), (B, 1, KV_S), and (B, 1, 1, KV_S).
  • [object Object] refers to the S dimension in the shape of [object Object], while [object Object] refers to the S dimension in the shapes of [object Object] and [object Object]. If [object Object] or [object Object] is not 16- or 32-byte aligned, it can be rounded up to the nearest aligned value.

• When the data type of [object Object] is INT8 or UINT8, the value in the tensor must be 0 or 1.

• [object Object]: aclIntArray on the host, indicating the valid sequence lengths of [object Object] in different batches. The data type can be INT64. If you do not specify [object Object], [object Object] is passed, indicating that the length is the same as that of [object Object] in the shape of [object Object]. Restriction: The valid sequence length of each batch in the input parameter must be less than or equal to that of the corresponding batch in [object Object]. This parameter is invalid when [object Object] is [object Object]. If the input length of [object Object] is 1, all batches use the same [object Object]. If the input length is greater than or equal to the batch size, the first N elements (where N equals the batch size) of [object Object] are used. Other lengths are not supported.

• [object Object]: aclIntArray on the host. It can be set to [object Object], indicating the valid sequence lengths of [object Object]/[object Object] in different batches. The data type can be INT64. If you do not specify [object Object], [object Object] is passed, indicating that the length is the same as that of [object Object] in the shape of [object Object]/[object Object]. Restriction: The valid sequence length of each batch in the input parameter must be less than or equal to that of the corresponding batch in [object Object]/[object Object]. If the input length of [object Object] is 1, all batches use the same [object Object]. If the input length is greater than or equal to the batch size, the first N elements (where N equals the batch size) of [object Object] are used. Other lengths are not supported.

• Constraints on the number of input parameters and input and output data formats related to INT8 quantization:

  • If both the input and output are of the INT8 type, the input parameters [object Object], [object Object], [object Object], and [object Object] must exist at the same time. [object Object] is optional and defaults to 0 if not passed.
  • If the input is of the INT8 type and the output is of the FLOAT16 type, the input parameters [object Object], [object Object], and [object Object] must exist at the same time. If the input parameter [object Object] or [object Object] exists (not [object Object]), an error is reported and returned.
  • When the input is of the FLOAT16 or BFLOAT16 type and the output is of the INT8 type, the input parameter [object Object] must exist, and [object Object] is optional (defaults to 0 if not passed). If the input parameter [object Object], [object Object], or [object Object] exists (not [object Object]), an error is reported and returned.
  • The input parameters [object Object] and [object Object] support both the per-tensor and per-channel formats and the FLOAT32 and BFLOAT16 data types. If [object Object] is passed, ensure that its type and shape are consistent with those of [object Object]. When the input is of the BFLOAT16 type, both FLOAT32 and BFLOAT16 are supported. Otherwise, only FLOAT32 is supported. In per-channel format, when the output layout is BSH, the product of all dimensions of [object Object] must be equal to H. For other layouts, the product must be equal to N × D. (When the output layout is BSH, it is recommended that the shape of [object Object] be set to [1,1,H] or [H]. When the output layout is BNSD, it is recommended that the shape of [object Object] be set to [1,N,1,D] or [N,D]. When the output layout is BSND, it is recommended that the shape of [object Object] be set to [1,1,N,D] or [N,D].)
  • inputLayout supports only BSH, BNSD, BSND, and BNSD_BSND.

• Constraints on the fake-quantization parameters [object Object] and [object Object]:

  • Only the fake-quantization scenario where [object Object] is INT8 is supported.
  • Per-channel mode: The shapes of the two parameters can be (2, N, 1, D), (2, N, D), (2, H), where N is [object Object]. The data type is the same as that of [object Object], and [object Object] is set to [object Object].
  • Per-tensor mode: The shapes of the two parameters are (2), the data type is the same as that of [object Object], and [object Object] is set to [object Object].
  • Per-token mode: The shapes of the two parameters are (2, B, S), the data type is fixed at FLOAT32, and [object Object] is set to [object Object].
  • In asymmetric quantization mode, both [object Object] and [object Object] must be present.
  • In symmetric quantization mode, [object Object] can be [object Object]. If [object Object] is [object Object], symmetric quantization is performed. Otherwise, asymmetric quantization is performed.
  • [object Object]Atlas A2 training products/Atlas A2 inference products[object Object]: When [object Object] is 1, the data type can be FLOAT16, BFLOAT16, or FLOAT32. If [object Object] is greater than or equal to 2, only FLOAT16 is supported.
  • Ascend 950PR/Ascend 950DT: The data type can be FLOAT16, BFLOAT16, or FLOAT32.

• keyAntiquantScaleOptional: aclTensor on the device. The data format supports ND. When the kv fake-quantization parameters are separated, this parameter indicates the key dequantization factor. If this function is not used, pass [object Object].

  • [object Object]Atlas A2 training products/Atlas A2 inference products[object Object]: The data type can be FLOAT16, BFLOAT16, or FLOAT32. When Q_S is greater than or equal to 2, only the per-token and per-channel modes are supported. When [object Object] is 1, only the per-tensor, per-channel, and per-token modes are supported. The per-tensor mode can be combined with the per-head mode, the per-token mode can be combined with the per-head mode, the per-token mode can be combined with the paged attention mode to manage scale/offset, and the per-token mode can be combined with the per-head mode and the paged attention mode to manage scale/offset.
  • Ascend 950PR/Ascend 950DT: The data type can be FLOAT16, BFLOAT16, FLOAT32, or FLOAT8_E8M0. The following modes are supported: per-tensor, per-channel, per-token, per-token with per-head, per-token with per-head, per-token with paged attention mode for managing scale/offset, and per-token with per-head and paged attention mode for managing scale/offset and per-token-group.

• [object Object]: aclTensor on the device. The data type can be FLOAT16, BFLOAT16, or FLOAT32. The data format can be ND. When the KV fake-quantization parameters are separated, this parameter indicates the dequantization offset of [object Object]. If this function is used, the data type and shape must be the same as those of [object Object]. If this function is not used, pass [object Object].

  • [object Object]Atlas A2 training products/Atlas A2 inference products[object Object]: When [object Object] is greater than or equal to 2, only the per-token and per-channel modes are supported. When [object Object] is 1, only the per-tensor, per-channel, and per-token modes are supported. The per-tensor mode can be combined with the per-head mode, the per-token mode can be combined with the per-head mode, the per-token mode can be combined with the paged attention mode to manage scale/offset, and the per-token mode can be combined with the per-head mode and the paged attention mode to manage scale/offset.
  • Ascend 950PR/Ascend 950DT: The following modes are supported: per-tensor, per-channel, per-token, per-token with per-head, per-token with per-head, per-token with paged attention mode for managing scale/offset, and per-token with per-head and paged attention mode for managing scale/offset.

• [object Object]: aclTensor on the device. The data format can be ND. When the KV fake-quantization parameters are separated, this parameter indicates the dequantization factor of [object Object]. If this function is not used, pass [object Object].

  • [object Object]Atlas A2 training products/Atlas A2 inference products[object Object]: The data type can be FLOAT16, BFLOAT16, or FLOAT32. When Q_S is greater than or equal to 2, only the per-token and per-channel modes are supported. When [object Object] is 1, only the per-tensor, per-channel, and per-token modes are supported. The per-tensor mode can be combined with the per-head mode, the per-token mode can be combined with the per-head mode, the per-token mode can be combined with the paged attention mode to manage scale/offset, and the per-token mode can be combined with the per-head mode and the paged attention mode to manage scale/offset.
  • Ascend 950PR/Ascend 950DT: The data type can be FLOAT16, BFLOAT16, FLOAT32, or FLOAT8_E8M0. The following modes are supported: per-tensor, per-channel, per-token, per-tensor with per-head, per-token with per-head, per-token with paged attention mode for managing scale/offset, and per-token with per-head and paged attention mode for managing scale/offset and per-token-group.

• valueAntiquantOffsetOptional: aclTensor on the device. The data type can be FLOAT16, BFLOAT16, or FLOAT32. The data format can be ND. When the fake-quantization parameters of the key-value pair are separated, this parameter indicates the dequantization offset of the value. If this function is used, the data type and shape of this parameter must be the same as those of valueAntiquantScaleOptional. If this function is not used, pass [object Object].

  • [object Object]Atlas A2 training products/Atlas A2 inference products[object Object]: When [object Object] is greater than or equal to 2, only the per-token and per-channel modes are supported. When [object Object] is 1, only the per-tensor, per-channel, and per-token modes are supported. The per-tensor mode can be combined with the per-head mode, the per-token mode can be combined with the per-head mode, the per-token mode can be combined with the paged attention mode to manage scale/offset, and the per-token mode can be combined with the per-head mode and the paged attention mode to manage scale/offset.
  • Ascend 950PR/Ascend 950DT: supports per-tensor, per-channel, and per-token. per-tensor and per-head are supported, per-token and per-head are supported, per-token and per-token are supported, and the paged attention mode is used to manage scale/offset. per-token and per-head are supported, and the paged attention mode is used to manage scale/offset.

• [object Object]: aclIntArray on the host. [object Object] can be passed, indicating the valid sequence length of [object Object]/[object Object]. The data type can be INT64. If [object Object] is not specified, [object Object] can be passed, indicating that the length is the same as that of [object Object] in [object Object]/[object Object]. Restriction: The valid sequence length in this input parameter must be less than or equal to that in [object Object]/[object Object].

• inputLayout: indicates the data layout format of the input query, key, and value. Currently, BSH, BSND, BNSD, BNSD_BSND (when the input is BNSD, the output format is BSND, and only Q_S greater than 1 is supported), and TND are supported. If no specific layout is required, BSH is recommended.

  • Note: The data layout formats of the query, key, and value can be interpreted from multiple dimensions. B (Batch) indicates the batch size of input samples, S (Seq-Length) indicates the sequence length of input samples, H (Hidden-Size) indicates the size of the hidden layer, N (Head-Num) indicates the number of heads, and D (Head-Dim) indicates the minimum unit size of the hidden layer, and D=H/N.

• Restrictions on [object Object]: [object Object] must be exactly divided by [object Object]. In the BSND, BNSD, BNSD_BSND, and TND scenarios, the value of numKeyValueHeads must be the same as the shape value of the N axis of the key/value in the shape. Otherwise, an exception occurs.

• Restrictions on [object Object]

  [object Object]

• Restrictions on [object Object]

  • There are four modes (0, 1, 2, and 3) in total, represented by 2-bit combinations. Bit 0 indicates whether to use the high-precision or high-performance mode, and bit 1 indicates whether to perform invalid row correction.

    [object Object]

  • Note: The high-precision and high-performance modes are applicable to both BFLOAT16 and INT8. Invalid row correction takes effect for FLOAT16, BFLOAT16, and INT8. The values [object Object] and [object Object] are reserved. If the masks involved in the computation are all 1s, the precision may be affected. In this case, you can set this parameter to [object Object] or [object Object] to enable invalid row correction to improve the precision. However, this configuration deteriorates the performance. If the operator can determine that invalid rows exist, the invalid row correction is automatically enabled, such as in scenarios where [object Object] is set to [object Object] and Sq is greater than Skv.

• Restrictions on [object Object]

  [object Object]
  • [object Object]Atlas A2 training products/Atlas A2 inference products[object Object]: When [object Object] is 1, passing a value other than 0, 1, 2, 3, 4, or 5 will result in an execution error. If [object Object] is greater than or equal to 2, only the values 0 and 1 are supported. Other values will result in an execution error.
  • Ascend 950PR/Ascend 950DT: If a value other than 0, 1, 2, 3, 4, 5, and 6 is passed, an exception occurs.

• Restrictions on [object Object]

  • Except for the scenario where [object Object] is 0 and [object Object] is 1, the value must be the same as that of [object Object].
  • [object Object]Atlas A2 training products/Atlas A2 inference products[object Object]: When [object Object] is 1, passing a value other than 0, 1, 2, 3, 4, or 5 will result in an execution error. If [object Object] is greater than or equal to 2, only the values 0 and 1 are supported. Other values will result in an execution error.
  • Ascend 950PR/Ascend 950DT: If a value other than 0, 1, 2, 3, 4, 5, and 6 is passed, an exception occurs.

• Restrictions on [object Object]

  • In the ring attention algorithm, the product of query and key is first processed to obtain softmax_max. This max value is subtracted from the product before calculating the exponential, which is then summed to yield softmax_sum. Finally, the log of softmax_sum is added back to softmax_max to obtain the final result.
  • When [object Object] is [object Object], the shape must be [B,N,Q_S,1], where inf values represent invalid data.
  • When [object Object] is [object Object], if the [object Object] tensor is not [object Object], the tensor data is returned directly. If [object Object] is [object Object], a tensor of shape {1} filled with zeros is returned.

• When Q_S is greater than 1

  • Restrictions on [object Object], [object Object], and [object Object]:

    • B-axis restriction

      • The B axis must be less than or equal to 65536.
      • In the discontinuous scenario, the batch size in the tensor list of key and value can only be 1, and the number of key and value is equal to the B axis of query. The N and D axes of query and value must be the same. Due to the tensor list restrictions, B cannot be greater than 256 in non-contiguous scenarios.
      • [object Object]Atlas A2 training products/Atlas A2 inference products[object Object]
        • If the input type is INT8 and the D axis is not 32-byte aligned, the maximum value of the B axis is 128. If the input type is FLOAT16 or BFLOAT16 and the D axis is not 16-byte aligned, the maximum value of the B axis is also 128.

    • The N axis can be less than or equal to 256, and the D axis can be less than or equal to 512. If [object Object] is BSH or BSND, it is recommended that N × D be less than 65535.

    • The S axis must be less than or equal to 20971520 (20M). In some long sequence scenarios, if the computation load is too large, the PFA operator execution may time out (an AI Core error is reported, and [object Object] is [object Object]). In this case, S axis splitting is recommended. Note: The computation load is affected by parameters such as B, S, N, and D. Larger values indicate larger computation loads. The following lists some typical scenarios with long sequences (that is, the product of B, S, N, and D is large).

      [object Object]

    • Constraints on the [object Object] dimension:

      • [object Object]Atlas A2 training products/Atlas A2 inference products[object Object]

        • If the data type of query, key, value, or attentionOut is INT8, the D axis must be 32-pixel aligned. If the data type of query, key, value, or attentionOut is INT4, the D axis must be 64-pixel aligned. If their data types are all FLOAT16 or BFLOAT16, the D axis must be 16-pixel aligned.

      • Ascend 950PR/Ascend 950DT:

        • Non-quantization scenario: The types of query, key, and value are all FLOAT16 or BFLOAT16. The D axis ranges from 1 to 512.
        • Full quantization scenario: The types of query, key, and value are all INT8. The D axis ranges from 1 to 512.
        • Fake-quantization scenario: The query type is FLOAT16 or BFLOAT16, and the key and value types are INT8/HIFLOAT8/FLOAT8_E4M3FN/FLOAT4_E2M1/INT4(INT32). When the key and value types are FLOAT4_E2M1/INT4(INT32), the D axis of the query and the D axis of the key and value support only 64-byte alignment. (For INT32, only the D axis of the key and value supports 8-byte alignment.)

  • Currently, [object Object] can only be set to 0, 1, 2, 3, or 4. An error will be reported if it is set to other values.

    • When [object Object] is set to 0, if [object Object] is a null pointer or is passed in the left padding scenario, the input parameters [object Object] and [object Object] are ignored.
    • When sparseMode is set to 2, 3, or 4, the shape of attenMask must be S, S, 1, S, S, or 1, 1, S, S. The value of S must be fixed to 2048. In addition, you need to ensure that the input attenMask is a lower triangle. If attenMask is nullptr or the input shape is incorrect, an error is reported.
    • When [object Object] is set to 1, 2, or 3, the input parameters [object Object] and [object Object] are ignored, and their values are assigned based on related rules.

  • In the synthesis parameter scenario of KV cache dequantization, only when [object Object] is of the FLOAT16 type, [object Object] and [object Object] of the INT8 type can be dequantized to FLOAT16. If the product of the data ranges of the input [object Object] and [object Object] and the data range of the input [object Object] must be within the range of (–1, 1), the high-performance mode can ensure precision. Otherwise, the high-precision mode needs to be enabled to ensure precision.

  • Paged attention scenario

    • The prerequisite for enabling paged attention is that [object Object] exists and is valid, and [object Object] and [object Object] are arranged in a continuous memory based on the indexes in [object Object]. In this scenario, [object Object] of [object Object] and [object Object] is invalid. [object Object] is filled with block IDs. Currently, the validity of block IDs is not verified. You need to ensure the validity of block IDs.
      • [object Object]Atlas A2 training products/Atlas A2 inference products[object Object]: The data types of [object Object] and [object Object] can be FLOAT16 or BFLOAT16.
      • Ascend 950PR/Ascend 950DT: The key and value data types can be FLOAT16/BFLOAT16/INT8/HIFLOAT8/FLOAT8_E4M3FN/FLOAT4_E2M1/INT4(INT32).
    • [object Object] is a user-defined parameter. Its value affects the paged attention performance. When paged attention is enabled, the value of [object Object] must be a multiple of 128, ranging from 128 to 512. Generally, paged attention can improve the throughput but deteriorate the performance.
    • In the paged attention scenario, if the input KV cache layout is BnBsH (blocknum, blocksize, H) and the product of [object Object] multiplied by [object Object] exceeds 65535, an error will be reported due to hardware instruction restrictions. This problem can be solved by enabling GQA (decreasing [object Object]) or adjusting the KV cache layout to BnNBsD (blocknum, KV_N, blocksize, D). When [object Object] of [object Object] is BNSD, the KV cache layout can beBnBsH or BnNBsD. When [object Object] of [object Object] is BSH or BSND, the KV cache layout can only be BnBsH. The value of [object Object] cannot be less than the sum of blocks in each batch calculated based on [object Object] and [object Object]. The shapes of [object Object] and [object Object] must be the same.
    • Paged attention fake-quantization scenario
      • [object Object]Atlas A2 training products/Atlas A2 inference products[object Object]: The data type of [object Object] can be FLOAT16 or BFLOAT16, and the data types of [object Object] and [object Object] can be INT8.
      • Ascend 950PR/Ascend 950DT: The dtype of query can be FLOAT16 or BFLOAT16, and the dtype of key and value can be INT8/HIFLOAT8/FLOAT8_E4M3FN/FLOAT4_E2M1/INT4(INT32).
    • Paged attention full-quantization scenario
      • [object Object]Atlas A2 training products/Atlas A2 inference products[object Object]: The data type of [object Object] cannot be INT8.
      • Ascend 950PR/Ascend 950DT: The dtype of query and kv cache can be INT8.
    • Paged attention does not support the tensor list or left padding.
    • In the paged attention scenario, [object Object] must be passed.
    • In the paged attention scenario, [object Object] must be two-dimensional. The length of the first dimension must be equal to B, and the length of the second dimension must be greater than or equal to [object Object] (the maximum number of blocks corresponding to [object Object] in different batches).
    • When paged attention is enabled, the input [object Object] must be greater than or equal to [object Object] × [object Object] in the following scenarios:
      • When [object Object] is passed, the mask shape is (B, 1, Q_S, KV_S).
      • When [object Object] is passed, the [object Object] shape is (B, N, Q_S, KV_S).
      • Enabling the fake-quantization per-token mode: The shapes of the input parameters antiquantScale and antiquantOffset are both (2, B, S).
      • Enabling the fake-quantization per-token mode with the per-head mode: The shapes of the input parameters antiquantScale and antiquantOffset are both (B, N, S), and the data type is fixed to FLOAT32. This mode is supported when the data type of key and value is INT8 or INT4 (INT32).
      • Per-token-group mode: The shape of [object Object] is [object Object], and its data type is fixed at FLOAT8_E8M0. [object Object] is not supported. This parameter is supported when the key and value data types are FLOAT4_E2M1.

  • Left padding for [object Object]

    • The transfer start point of [object Object] is calculated as follows: [object Object] – [object Object] – [object Object]. The transfer end point of [object Object] is calculated as follows: [object Object] – [object Object]. The transfer start point of [object Object] cannot be less than 0, while the end point cannot be greater than [object Object]. Otherwise, the result will not meet the expectation.
    • If [object Object] is less than 0, it will be set to 0.
    • It must be enabled together with [object Object]. Otherwise, the default scenario is right padding for [object Object].
    • It does not support paged attention and cannot be enabled together with [object Object].
    • [object Object]Atlas A2 training products/Atlas A2 inference products[object Object]: The scenario where Q is BF16/FP16 and KV is INT4 is not supported.

  • Left padding for [object Object]:

    • The transfer start point of [object Object] and [object Object] is calculated as follows: [object Object] – [object Object] – [object Object]. The transfer end point of [object Object] and [object Object] is calculated as follows: [object Object] – [object Object]. The transfer start point of [object Object] and [object Object] cannot be less than 0, while the end point cannot be greater than [object Object]. Otherwise, the result will not meet the expectation.
    • If [object Object] is less than 0, it will be set to 0.
    • It must be enabled together with [object Object]. Otherwise, the default scenario is right padding for [object Object].
    • It does not support paged attention and cannot be enabled together with [object Object].
    • [object Object]Atlas A2 training products/Atlas A2 inference products[object Object]: The scenario where Q is BF16/FP16 and KV is INT4 is not supported.

  • When the output is of type INT8 and [object Object] and [object Object] are per-channel, left padding, ring attention, or non-32-byte alignment of D is not supported.

  • When the output is of type INT8, [object Object] cannot be [object Object] and [object Object] or [object Object] cannot be negative.

  • When the output of type INT8, if the input parameter[object Object] is a non-null pointer and a non-null tensor, and [object Object], [object Object], and [object Object] meet the following conditions, certain rows of the matrix will not be involved in computation, resulting in a computation result error. In this scenario, the computation will be intercepted. (Solution: To prevent interception, perform post-quantization outside the FIA interface.)

    • When [object Object] is [object Object] and [object Object] is a non-null pointer, interception occurs if for any batch: [object Object] – [object Object] – [object Object] – [object Object] > [object Object], or [object Object] < [object Object].
    • When [object Object] is 1 or 2, interception does not occur.
    • When [object Object] is 3, interception occurs if [object Object] + [object Object] – [object Object] < 0 for any batch.
    • When [object Object] is 4, interception occurs if [object Object] < 0 or [object Object] + [object Object] + [object Object] – [object Object] < 0 for any batch.

  • Restrictions on [object Object]

    • This function is supported when the data type of [object Object] is FLOAT16, BFLOAT16, or INT8.
    • When the data type of [object Object] is FLOAT16 and [object Object] exists, the high-precision mode is forcibly used. The corresponding restrictions are the same as those of the high-precision mode.
    • [object Object] must be greater than or equal to [object Object] of [object Object], and [object Object] must be greater than or equal to [object Object] of [object Object]. In the prefix scenario, [object Object] must be greater than or equal to the sum of [object Object] and [object Object] of [object Object].
    • Ascend 950PR/Ascend 950DT: In non-quantization and full quantization scenarios, there is no alignment restriction.

  • Restrictions on prefix parameters

    • Both [object Object] and [object Object] must be either null or non-null.
    • If neither [object Object] nor [object Object] is null, the dimensions and data types of [object Object], [object Object], [object Object], and [object Object] must be the same.
    • If neither [object Object] nor [object Object] is null, the first dimension (batch) of the shape of [object Object] must be 1. When the layout is BNSD or BSND, the N and D axes must be the same as those of [object Object]. When the layout is BSH, the H axis must be the same as that of [object Object]. The same rules apply to [object Object]. [object Object] of [object Object] and [object Object] must be the same.
    • When [object Object] exists, its shape must be [1], and its value cannot be greater than [object Object] of [object Object] and [object Object].
    • The sum of [object Object] of the public prefix and [object Object] of [object Object] or [object Object] must meet the original restriction on [object Object] of [object Object] or [object Object].
    • The prefix does not support paged attention, left padding, or tensor list.
    • In the prefix scenario, when [object Object] is 0 or 1 and [object Object] is passed, [object Object] must be greater than or equal to the sum of [object Object] and [object Object] of [object Object].
    • In the prefix scenario, the input [object Object] cannot be all INT8.

  • KV fake-quantization parameter separation

    • [object Object]Atlas A2 training products/Atlas A2 inference products[object Object]
      • [object Object] and [object Object] must be the same.
      • Both [object Object] and [object Object] must be either null or non-null. Both [object Object] and [object Object] must be either null or non-null.
      • If neither [object Object] nor [object Object] is null, their shapes must be the same. If neither [object Object] nor [object Object] is null, their shapes must be the same.
      • Only the per-token and per-channel modes are supported. In per-token mode, the shapes of the two parameters must be both (B, S), and the data type is fixed at FLOAT32. In per-channel mode, the shapes of the two parameters must be (N, D), (N, 1, D), or (H), and the data type is fixed at BF16.
      • When both fake-quantization parameters and KV separation quantization parameters are passed, the KV separation quantization parameters take effect.
      • When [object Object] and [object Object] are non-null, [object Object] of [object Object] must be less than or equal to 16.
      • When [object Object] and [object Object] are non-null, the data type of [object Object] must be BFLOAT16, the data type of [object Object] and [object Object] must be INT8, and the data type of the output must be BFLOAT16.
      • When [object Object] and [object Object] are non-null, the tensor list, left padding, and paged attention functions are not supported.
    • Ascend 950PR/Ascend 950DT:
      • Except when [object Object] is 0 and [object Object] is 1, the values of [object Object] and [object Object] must be the same.
      • Both [object Object] and [object Object] must be either null or non-null. Both [object Object] and [object Object] must be either null or non-null.
      • If neither [object Object] nor [object Object] is null, their shapes must be the same, except when [object Object] is 0 and [object Object] is 1. If neither [object Object] nor [object Object] is null, their shapes must be the same, except when [object Object] is 0 and [object Object] is 1.
      • The following modes are supported: per-channel, per-tensor, per-token, per-tensor + per-head, per-token + per-head, per-token + paged attention to manage scale/offset, per-token + per-head + paged attention to manage scale/offset, per-channel for [object Object] + per-token for [object Object], and [object Object]. In the following description, [object Object] indicates [object Object].
        • Per-channel mode: The shapes of the two parameters can be (1, N, 1, D), (1, N, D), (1, H), (N, 1, D), (N, D), or (H). The data type of the parameter is the same as that of the query. This mode is supported when the key and value data types are INT8, INT4 (INT32), HIFLOAT8, or FLOAT8_E4M3FN. When the key and value data types are HIFLOAT8 or FLOAT8_E4M3FN, antiquantOffset is not supported.
        • Per-tensor mode: The shapes of the two parameters are both (1), and the data type is the same as that of the query. This mode is supported when the key and value data types are INT8.
        • Per-token mode: The shapes of the two parameters can be (1, B, S) or (B, S), and the data type is fixed to FLOAT32. This mode is supported when the key and value data types are INT8 or INT4 (INT32).
        • Per-tensor and per-head mode: The shapes of the two parameters are both (N), and the data type is the same as that of the query. This mode is supported when the key and value data types are INT8.
        • Per-channel key and per-token value: For the key, the per-channel mode is supported. The shapes of the two parameters can be (1, N, 1, D), (1, N, D), (1, H), (N, 1, D), (N, D), or (H), and the data type of the parameters is the same as that of the query. For the value, the per-token mode is supported. The shapes of the two parameters are both (1, B, S), and the data type is fixed to FLOAT32. This mode is supported when the key and value data types are INT8 or INT4 (INT32). When the key and value data types are INT8, the query and output support only FLOAT16.
        • Per-token-group mode: The shape of antiquantScale is (1, B, N, S, D/32), and the data type is fixed to FLOAT8_E8M0. antiquantOffset is not supported. This mode is supported when the key and value data types are FLOAT4_E2M1.
        • Per-token + per-head mode: The shapes of the two parameters are both (B, N, S), and their data types are fixed at FLOAT32. This mode is supported when the data types of [object Object] and [object Object] are INT8 or INT4 (INT32).
        • Per-token + paged attention to manage scale/offset: The shapes of the two parameters are both (blocknum, blocksize), and their data types are fixed at FLOAT32. This mode is supported when the data types of [object Object] and [object Object] are INT8.
        • Per-token + per-head + paged attention to manage scale/offset: The shapes of the two parameters are both (blocknum, N, blocksize), and their data types are fixed at FLOAT32. This mode is supported when the data types of [object Object] and [object Object] are INT8.
      • When both fake-quantization parameters and KV separation quantization parameters are passed, the KV separation quantization parameters take effect.
      • In the INT4 (INT32) fake-quantization scenario, only separate quantization of keys and values is supported. The details are as follows:
        • Per-channel mode
        • Per-token mode
        • Per-token + per-head mode
        • Per-channel for [object Object] + per-token for [object Object] mode
      • Post-quantization is not supported in some fake-quantization scenarios.
        • [object Object]Atlas A2 training products/Atlas A2 inference products[object Object]: Post-quantization is not supported in the INT4 (INT32) fake-quantization scenario.
        • Ascend 950PR/Ascend 950DT: Post-quantization is not supported in the INT4 (INT32) and FLOAT4_E2M1 fake-quantization scenarios.

• When Q_S is equal to 1

  • Restrictions on [object Object], [object Object], and [object Object]:
    • The B axis can be less than or equal to 65536, the N axis can be less than or equal to 256, and the D axis can be less than or equal to 512.
    • The input types of [object Object], [object Object], and [object Object] cannot be all INT8.
    • In the INT4 (INT32) fake-quantization scenario, the aclnn single-operator calling supports the INT4 input of keys and values or the INT32 input formed by combining INT4 inputs. (You are advised to use dynamicQuant to generate INT4 data because dynamicQuant is an INT32 data that contains eight INT4 data.)
    • In the INT4 (INT32) fake-quantization scenario, if the INT4 inputs of keys and values are combined into an INT32 input, the N, D, or H value of the keys and values is 1/8 of the actual value. (The same applies to the prefix.)
    • Restrictions on the D axis for [object Object] and [object Object] in specific data types
      • [object Object]Atlas A2 training products/Atlas A2 inference products[object Object]: When the key and value inputs are of type INT4 (INT32), the D axis must be 64-byte aligned. (INT32 supports only 8-byte alignment for the D axis.)
      • Ascend 950PR/Ascend 950DT: When the key and value inputs are of type FLOAT4_E2M1/INT4(INT32), the D axis of the query and the D axis of the key and value must be 64-byte aligned. (INT32 supports only 8-byte alignment for the D axis of the key and value.)
  • Paged attention scenario
    • The prerequisite for enabling paged attention is that [object Object] exists and is valid, and [object Object] and [object Object] are arranged in a continuous memory based on the indexes in [object Object]. In this scenario, [object Object] of [object Object] and [object Object] is invalid.
    • [object Object] is a user-defined parameter. Its value affects the paged attention performance. When paged attention is enabled, a non-zero value must be passed for [object Object], and the maximum value is [object Object]. Generally, paged attention can improve the throughput but deteriorate the performance.
    • In the paged attention scenario, when the inputLayout of the query is BNSD, the KV cache layout supports two formats: BnBsH (blocknum, blocksize, H) and BnNBsD (blocknum, KV_N, blocksize, D). When the inputLayout of the query is BSH or BSND, the KV cache layout supports only the BnBsH format. The value of [object Object] cannot be less than the sum of blocks in each batch calculated based on [object Object] and [object Object]. The shapes of [object Object] and [object Object] must be the same.
    • In the paged attention scenario, the performance is generally better when the KV cache layout is BnNBsD than when it is BnBsH. Therefore, BnNBsD is recommended.
    • In the paged attention scenario, if the input KV cache layout is BnBsH and numKvHeads × headDim exceeds 64 KB, an error will be reported due to hardware instruction constraints. This problem can be solved by enabling GQA (decreasing [object Object]) or adjusting the KV cache layout to BnNBsD.
    • Paged attention does not support the tensor list or left padding:
    • In the paged attention scenario, [object Object] must be passed.
    • In the paged attention scenario, [object Object] must be two-dimensional. The length of the first dimension must be equal to B, and the length of the second dimension must be greater than or equal to [object Object] (the maximum number of blocks corresponding to [object Object] in each batch).
    • When paged attention is enabled, the input [object Object] must be greater than or equal to [object Object] × [object Object] in the following scenarios:
      • When the attention mask is enabled, the mask shape is (B, 1, 1, S).
      • When [object Object] is enabled, the [object Object] shape is (B, N, 1, S).
      • When the fake-quantization per-token mode is enabled, the shape of the input parameters [object Object] and [object Object] is both (2, B, S).
      • Per-token + per-head mode: The shapes of the two parameters are both [object Object], their data type is fixed at FLOAT32, and the data types of [object Object] and [object Object] are INT8 or INT4 (INT32).
      • Per-token-group mode: The shape of [object Object] is [object Object], and its data type is fixed at FLOAT8_E8M0. [object Object] is not supported. This mode is supported when the key and value data types are FLOAT4_E2M1.
  • Left padding for [object Object]:
    • [object Object]Atlas A2 training products/Atlas A2 inference products[object Object]: The scenario where Q is BF16/FP16 and KV is INT4 (INT32) is not supported.
    • Ascend 950PR/Ascend 950DT: The scenario where Q is BF16/FP16 and KV is INT4 (INT32) is supported. There is no restriction on the QKV data type.
    • The formula for calculating the start point of kvCache transfer is as follows: KV_S – kvPaddingSize – actualSeqLengthsKv. The transfer end point of [object Object] is calculated as follows: [object Object] – [object Object] If the transfer start point or end point is less than 0, the returned data is all 0s.
    • If [object Object] is less than 0, it will be set to 0.
    • It must be enabled together with [object Object]. Otherwise, the default scenario is right padding for [object Object].
    • Paged attention and tensor list are not supported. Otherwise, the default scenario is right padding for [object Object].
    • When it is enabled together with [object Object], ensure that the meaning of [object Object] is correct, that is, invalid data can be correctly masked. Otherwise, precision issues may occur.
  • Restrictions on [object Object]
    • The data types of [object Object] and [object Object] must be consistent.
  • KV fake-quantization parameter separation
    • Except when [object Object] is 0 and [object Object] is 1, the values of [object Object] and [object Object] must be the same.
    • Both [object Object] and [object Object] must be either null or non-null. Both [object Object] and [object Object] must be either null or non-null.
    • If neither [object Object] nor [object Object] is null, their shapes must be the same, except when [object Object] is 0 and [object Object] is 1. If neither [object Object] nor [object Object] is null, their shapes must be the same, except when [object Object] is 0 and [object Object] is 1.
    • [object Object]Atlas A2 training products/Atlas A2 inference products[object Object]: The following eight modes are supported: per-channel, per-tensor, per-token, per-tensor + per-head, per-token + per-head, per-token + paged attention to manage scale/offset, per-token + per-head + paged attention to manage scale/offset, and [object Object] in per-channel + [object Object] in per-token. In the following description, [object Object] indicates [object Object].
      • Per-channel mode: The shapes of the two parameters can be (1, N, 1, D), (1, N, D), or (1, H). Their data types are the same as that of [object Object]. This mode is supported when the data types of [object Object] and [object Object] are INT8 or INT4 (INT32).
      • Per-tensor mode: The shapes of the two parameters are both (1), and their data types are the same as that of [object Object]. This mode is supported when the data types of [object Object] and [object Object] are INT8.
      • Per-token mode: The shapes of the two parameters are both (1, B, S), and their data type is fixed at FLOAT32. This mode is supported when the data types of [object Object] and [object Object] are INT8 or INT4 (INT32).
      • Per-tensor + per-head mode: The shapes of the two parameters are both (N), and their data types are the same as that of [object Object]. This mode is supported when the data types of [object Object] and [object Object] are INT8.
      • Per-channel for [object Object] + per-token for [object Object] mode: For [object Object] in per-channel, the shapes of the two parameters can be (1, N, 1, D), (1, N, D), or (1, H), and their data types are the same as that of [object Object]. For [object Object] in per-token, the shapes of the two parameters are both (1, B, S), and their data types are fixed at FLOAT32. This mode is supported when the data types of [object Object] and [object Object] are INT8 or INT4 (INT32). When the data types of [object Object] and [object Object] are INT8, only the data types of [object Object] and [object Object] can be FLOAT16.
    • Ascend 950PR/Ascend 950DT: The following nine modes are supported: per-channel, per-tensor, per-token, per-tensor and per-head, per-token and per-head, per-token using the paged attention mode to manage scale/offset, per-token and per-head using the paged attention mode to manage scale/offset, per-channel for key and per-token and per-token-group for value. In the following description, N indicates numKeyValueHeads.
      • Per-channel mode: The shapes of the two parameters can be (1, N, 1, D), (1, N, D), or (1, H). The data type of the parameter is the same as that of the query. This mode is supported when the data type of key and value is INT8, INT4 (INT32), HIFLOAT8, or FLOAT8_E4M3FN. When the data type of key and value is HIFLOAT8 or FLOAT8_E4M3FN, antiquantOffset is not supported.
      • Per-tensor mode: The shapes of the two parameters are both (1), and the data type is the same as that of the query. This mode is supported when the data type of key and value is INT8 or INT4 (INT32).
      • Per-token mode: The shapes of the two parameters can be (1, B, S) or (B, S), and the data type is fixed to FLOAT32. This mode is supported when the data type of key and value is INT8 or INT4 (INT32).
      • Per-tensor and per-head mode: The shapes of the two parameters are both (N), and the data type is the same as that of the query. This mode is supported when the data type of key and value is INT8 or INT4 (INT32).
      • Per-channel key and per-token value: For the key, the per-channel mode is supported. The shapes of the two parameters can be (1, N, 1, D), (1, N, D), or (1, H), and the data type is the same as that of the query. For the value, the per-token mode is supported. The shapes of the two parameters are both (1, B, S), and the data type is fixed to FLOAT32. This mode is supported when the data type of key and value is INT8 or INT4 (INT32). When the data type of key and value is INT8, the query and output support only FLOAT16.
      • Per-token-group mode: The shape of antiquantScale is (1, B, N, S, D/32), and the data type is fixed to FLOAT8_E8M0. antiquantOffset is not supported. This mode is supported when the data type of key and value is FLOAT4_E2M1.
      • Per-token + per-head mode: The shapes of the two parameters are both (B, N, S), and their data types are fixed at FLOAT32. This mode is supported when the data types of [object Object] and [object Object] are INT8 or INT4 (INT32).
      • Per-token + paged attention to manage scale/offset: The shapes of the two parameters are both (blocknum, blocksize), and their data types are fixed at FLOAT32. This mode is supported when the data types of [object Object] and [object Object] are INT8.
      • Per-token + per-head + paged attention to manage scale/offset: The shapes of the two parameters are both (blocknum, N, blocksize), and their data types are fixed at FLOAT32. This mode is supported when the data types of [object Object] and [object Object] are INT8.
    • When both fake-quantization parameters and KV separation quantization parameters are passed, the KV separation quantization parameters take effect.
    • In the INT4 (INT32) fake-quantization scenario, only the KV fake-quantization parameters can be separated. The details are as follows:
      • Per-tensor mode
      • Per-channel mode
      • Per-token mode
      • Per-tensor + per-head mode
      • Per-token + per-head mode
      • Per-channel for [object Object] + per-token for [object Object] mode
    • Post-quantization is not supported in some fake-quantization scenarios.
      • [object Object]Atlas A2 training products/Atlas A2 inference products[object Object]: Post-quantization is not supported in the INT4 (INT32) fake-quantization scenario.
      • Ascend 950PR/Ascend 950DT: Post-quantization is not supported in the INT4 (INT32) and FLOAT4_E2M1 fake-quantization scenarios.
  • Restrictions on prefix parameters
    • Both [object Object] and [object Object] must be either null or non-null.
    • If neither [object Object] nor [object Object] is null, the dimensions and data types of [object Object], [object Object], [object Object], and [object Object] must be the same.
    • If neither [object Object] nor [object Object] is null, the first dimension (batch) of the shape of [object Object] must be 1. When the layout is BNSD or BSND, the N and D axes must be the same as those of [object Object]. When the layout is BSH, the H axis must be the same as that of [object Object]. The same rules apply to [object Object]. [object Object] of [object Object] and [object Object] must be the same.
    • When [object Object] exists, its shape must be [1], and its value cannot be greater than [object Object] of [object Object] and [object Object].
    • The sum of [object Object] of the public prefix and [object Object] of [object Object] or [object Object] must meet the original restriction on [object Object] of [object Object] or [object Object].

[object Object]

The following example is for reference only. For details, see .