Levi 9862a23985 【0.11.0-dev】optimization of kimi-k2 in cann8.3 (#4555) ...

### What this PR does / why we need it?
In cann8.3， npu_moe_gating_top_k operator can support expert nums with
384, so kimi can use the operator to get better preformance.
---------

Signed-off-by: Levi-JQ <yujinqi2@huawei.com>
Co-authored-by: Levi-JQ <yujinqi2@huawei.com>

2025-12-09 08:49:15 +08:00

..

__init__.py

[1/N][refactor] torchair fused_moe refactor (#2438)

2025-08-25 15:46:10 +08:00

sequence_parallel.py

[Refactor] [SP]The sequence parallelism characteristics in the MoE and Dense models are integrated into a single solution. (#3085)

2025-09-24 11:29:59 +08:00

shared_weight_layer.py

[1/N][Feat] Cut down memory usage for o_proj in DeepSeek (#2931)

2025-09-24 17:16:41 +08:00

torchair_activation.py

[main] mlp weight prefetch in Qwen Dense Models (#2816)

2025-09-11 21:20:09 +08:00

torchair_fused_moe.py

【0.11.0-dev】optimization of kimi-k2 in cann8.3 (#4555)

2025-12-09 08:49:15 +08:00

torchair_layernorm.py

fix deepseek torchair recompile (#3679)

2025-10-23 22:53:13 +08:00

torchair_rotary_embedding.py

Fix the bugs about operator registration by PyTorch Dispatcher (#2786)

2025-09-13 11:58:52 +08:00

torchair_vocab_parallel_embedding.py

[Feature] optimize sp & qwen3 next support sp. (#3225)

2025-10-13 23:02:12 +08:00