GitLab

Co-authored-by: zhuyue <zhuyue@qiyuanlab.com>

* issue/472: p800 ccl

* issue/472: 删掉无用操作

* issue/472: fix format

* issue/472: memcpy h2h case

issue/477 - Cambricon MLU NeoX

Added NeoX support to Cambricon RoPE; Added a missing argument in the profiling script;

issue/474: rename Dequantize to DequantizeAWQ in nvidia gpu

issue/469: disable NVIDIA-dequantize on Iluvatar GPU via ENABLE_NVIDIA_API marco

* issue/459 - Support more data type combinations

* issue/459 - added test cases for 9G7B and 9G70B

* issue/459 - modified rms kernel to support larger tensors

issue/466: 昆仑平台rope关于NEOX算法的实现

issue/434 - added bf16 support for Cambricon MLU

* issue/436: support kunlun rope U32

* issue/436: 支持9g7b 4b模型

---------
Co-authored-by: zhangyue <zhangyue@qiyuanlab.com>

issue/434 hccl support bf16

Signed-off-by: Ceng <441651826@qq.com>

Signed-off-by: Ceng <441651826@qq.com>

Issue/428: Merge `rope_v2` into `rope`

issue/428: update the rope implementation on Ascend, Cambricon, and Kunlun to use the refactored interface and return unimplemented error for NEOX-style algorithm

* issue/450: change indexToReducedOffset() to indexToOffset in elementwise framework on CPU, NVIDIA, Cambricon, Metax, Moore, and Kunlun

* issue/450: remove indexToReducedOffset() in all platforms

* issue/450: add the testcases that pinpoint the issue in infiniop-test

issue/450: change indexToReducedOffset() to indexToOffset in elementwise framework on CPU, NVIDIA, Cambricon, Metax, Moore, and Kunlun

issue/434 nccl support bf16

issue/425: implement GEMM with MUBLAS and MUDNN backends in moore gpu