Coverage for src/flag_gems/fused/__init__.py: 100%

1from flag_gems.fused.apply_repetition_penalties import apply_repetition_penalties 

2from flag_gems.fused.concat_and_cache_mla import concat_and_cache_mla 

3from flag_gems.fused.cross_entropy_loss import cross_entropy_loss 

4from flag_gems.fused.cutlass_scaled_mm import cutlass_scaled_mm 

5from flag_gems.fused.FLA import ( 

6 chunk_gated_delta_rule_fwd, 

7 fused_recurrent_gated_delta_rule_fwd, 

8) 

9from flag_gems.fused.flash_mla import flash_mla 

10from flag_gems.fused.fused_add_rms_norm import fused_add_rms_norm 

11from flag_gems.fused.fused_moe import ( 

12 dispatch_fused_moe_kernel, 

13 fused_experts_impl, 

14 inplace_fused_experts, 

15 invoke_fused_moe_triton_kernel, 

16 outplace_fused_experts, 

17) 

18from flag_gems.fused.geglu import dgeglu, geglu 

19from flag_gems.fused.gelu_and_mul import gelu_and_mul 

20from flag_gems.fused.grouped_topk import grouped_topk 

21from flag_gems.fused.instance_norm import instance_norm 

22from flag_gems.fused.moe_align_block_size import ( 

23 moe_align_block_size, 

24 moe_align_block_size_triton, 

25) 

26from flag_gems.fused.moe_sum import moe_sum 

27from flag_gems.fused.outer import outer 

28from flag_gems.fused.reglu import dreglu, reglu 

29from flag_gems.fused.reshape_and_cache import reshape_and_cache 

30from flag_gems.fused.reshape_and_cache_flash import reshape_and_cache_flash 

31from flag_gems.fused.rotary_embedding import apply_rotary_pos_emb 

32from flag_gems.fused.rwkv_ka_fusion import rwkv_ka_fusion 

33from flag_gems.fused.rwkv_mm_sparsity import rwkv_mm_sparsity 

34from flag_gems.fused.silu_and_mul import silu_and_mul, silu_and_mul_out 

35from flag_gems.fused.skip_layernorm import skip_layer_norm 

36from flag_gems.fused.swiglu import dswiglu, swiglu 

37from flag_gems.fused.topk_softmax import topk_softmax 

38from flag_gems.fused.weight_norm import weight_norm 

39 

40__all__ = [ 

41 "apply_repetition_penalties", 

42 "apply_rotary_pos_emb", 

43 "chunk_gated_delta_rule_fwd", 

44 "concat_and_cache_mla", 

45 "cutlass_scaled_mm", 

46 "cross_entropy_loss", 

47 "dispatch_fused_moe_kernel", 

48 "dgeglu", 

49 "dreglu", 

50 "dswiglu", 

51 "flash_mla", 

52 "fused_add_rms_norm", 

53 "fused_experts_impl", 

54 "fused_recurrent_gated_delta_rule_fwd", 

55 "geglu", 

56 "gelu_and_mul", 

57 "grouped_topk", 

58 "inplace_fused_experts", 

59 "instance_norm", 

60 "invoke_fused_moe_triton_kernel", 

61 "moe_sum", 

62 "moe_align_block_size", 

63 "moe_align_block_size_triton", 

64 "outer", 

65 "outplace_fused_experts", 

66 "reglu", 

67 "reshape_and_cache", 

68 "reshape_and_cache_flash", 

69 "rwkv_ka_fusion", 

70 "rwkv_mm_sparsity", 

71 "silu_and_mul", 

72 "silu_and_mul_out", 

73 "skip_layer_norm", 

74 "swiglu", 

75 "topk_softmax", 

76 "weight_norm", 

77]