deepgemm jit过程分析

youzjuer

已于 2025-08-01 11:04:01 修改

阅读量14

点赞数

CC 4.0 BY-SA版权

分类专栏： CUDA和LLM通俗易懂技术站文章标签：深度学习 pytorch 人工智能

于 2025-07-31 10:27:44 首次发布

本文链接：https://blog.youkuaiyun.com/youzjuer/article/details/149799079

CUDA和LLM通俗易懂技术站专栏收录该内容

110 篇文章 ¥59.90 ¥99.00

订阅专栏

超级会员免费看

一流程

1.1 runtime = jit_tuner.compile_and_tune

runtime = jit_tuner.compile_and_tune(
        name='gemm_fp8_fp8_bf16_nt',
        keys={'N': n, 'K': k, 'BLOCK_M': block_m, 'BLOCK_N': block_n,
              'SWIZZLE_D_MODE': smem_config[1],
              'BLOCK_N_PADDING': smem_config[2],
              'NUM_STAGES': num_stages,
              'NUM_TMA_MULTICAST': tma_multicast_config[0],
              'IS_TMA_MULTICAST_ON_A': tma_multicast_config[1]},
        space=(),
        includes=includes,
        arg_defs=(('lhs', torch.float8_e4m3fn), ('lhs_scales', torch.float),
                  ('rhs', torch.float8_e4m3fn), ('rhs_scales', torch.float),
                  ('out', torch.bfloat16), ('m', int),
                  ('stream', torch.c

了解本专栏