Constraints

  • ntokens = sum(seqlen[i]), i=0, 1...batch-1.
  • There are the following restrictions on the Atlas inference products:
    • The compression function is not supported.
    • If the data format is marked with NZ, the data format must be NZ.
    • The dimensions of keyCache and valueCache must be [num_blocks, num_head*head_size/16, block_size, 16]. The last dimension must be 16, and block_size must be 16-aligned.
    • When kvCacheCfg is set to K_CACHE_V_CACHE or K_CACHE_V_CACHE_NZ, the data format is NZ.
    • head_size must be 32-byte aligned.
  • If the last dimensions of the key and value are different, only Atlas A2 training products/Atlas A2 inference products and Atlas A3 inference products/Atlas A3 training products are supported. In this scenario, keyCache, valueCache, keyCacheOut, and valueCacheOut support only the ND format.
  • The data types of key, value, keyCache, valueCache, keyCacheOut, and valueCacheOut must be the same.
  • The data formats of keyCache, valueCache, keyCacheOut, and valueCacheOut must be the same.
  • The single-input single-output function is supported only by the Atlas A2 training products/Atlas A2 inference products and Atlas A3 inference products/Atlas A3 training products and does not support compression.
  • In the compression scenario, the cache size must be greater than the size after compression.
  • If multi-head compression is disabled, the NZ format supports only the float16 data type on the Atlas 200I/500 A2 inference products.
  • When the function of converting the format of the input to NZ is enabled, the following restrictions apply:
    • Only Atlas A2 training products/Atlas A2 inference products and Atlas A3 inference products/Atlas A3 training products are supported.
    • The length of key can be different from that of value. That is, the last dimension of key can be different from that of value.
    • The data type of key can be different from that of value.
  • When the multi-head compression function is enabled in the rope scenario, the value of head_size cannot exceed 16 × 1024.