Change History#

v5.0#

Release date: TBD

  • Added math operators: absolute (generated), acos, arcsinh (generated), arcsinh_ (generated), arcsinh.out (generated), arctanh_ (generated), asinh_ (generated), ceil, ceil_, ceil.out, diagmma_ (generated), equal, floor_ (generated), fmin (generated), fmin.out (generated) hardswish_ (generated), hypot (generated), i0 (generated), i0_ (generated), i0.out (generated), log1p_ (generated), logaddexp (generated), logaddexp.out (generated), logical_and_, logical_or_ logit (generated), logit_ (generated), logit.out (generated), sgn_ (generated), sinh_ (generated), special_i1 (generated), special_i1.out (generated)
  • Added BLAS operator: bmm.out, cutlass_scaled_mm_sm_90, tril (generated)
  • Added MoE operators: dispatch_fused_moe_kernel, grouped_topk, inplace_fused_experts, outplace_fused_experts
  • Added distribution operator: normal_,
  • Added neural network operators: _upsample_nearest_exact1d, apply_repetition_penalties (generated), chunk_gated_delta_rule_fwd, dswiglu, embedding_dense_backward, fused_recurrent_gated_delta_rule_fwd, hardsigmoid (generated), hardsigmoid.out (generated), nll_loss_nd_backward, nll_loss_nd_forward, one_hot, pixel_unshuffle (generated), pixel_unshuffle.out (generated), prelu (generated), reflection_pad1d (generated), reflection_pad1d.out (generated), reflection_pad2d (generated), reflection_pad2d.out (generated), relu6 (generated), swiglu, triu_, unfold_backward, upsample_bicubic2d, upsample_linear1d, upsample_nearest1d, upsample_nearest3d
  • Added tensor operators: _functional_sym_constrain_range_for_size (generated), alias_copy (generated), alias_copy.out (generated), fill.Scalar_out, fill.Tensor_out, lift_fresh_copy (generated), replication_pad1d (generated), replication_pad1d.out (generated), replication_pad3d, rrelu_with_noise_backward (generated), selu (generated), selu_ (generated), slice_backward (generated), softshrink (generated), softshrink.out (generated), t_copy (generated), t_copy.out (generated), unfold_backward, zero (generated), zero_, zero.out (generated)
  • Removed reduction operator moe_sum.
  • Added reduction operators: bincount
  • Added DSA operators: spare_mla_fwd

v4.2#

Release date: 2026-01-04

  • This release contains 216 operators, see Operator List
  • Added neural network operators: avg_pool2d, avg_pool2d_backward, dgeglu (alpha), dreglu (alpha), geglu (alpha), reglu (alpha),
  • Added BLAS operators: baddbmm,
  • Added convolution operators: conv1d, conv2d, conv3d,
  • Added tensor operator: copy_, masked_scatter, masked_scatter_, per_token_group_quant_fp8 (alpha), scatter_add_, to_copy.
  • Added reduction operator: moe_sum, scaled_softmax_backward, scaled_softmax_forward,
  • Added math operator: exp_out, tan, tan_, true_divide_out,
  • Promoted tensor operator continuous to beta stage.
  • Promoted reduction operator index to stable stage.
  • The previous upsample operator is now split into upsample_nearest2d and upsample_bicubic2d_aa

v4.1#

Release date: 2025-11-01

v4.0#

Release date: 2025-10-31

  • Supports 202 operators in total.

  • Added linear algebra operators: addcdiv, addcmul, addmv, addmv_out, addr

  • Added BLAS operators: addmm_out

  • Added math operators: atan, atan_, bitwise_left_shift, bitwise_right_shift, clamp_min, clamp_min_, exp2, exp2_, sqrt_

  • Added neural network operators: celu, celu_, elu_, elu_backward, get_scheduler_metadata, glu_backward, moe_align_block_size, softplus

  • Added reduction operators: index (beta), std, trace

  • Added tensor operators: index_add_, logspace, max_pool2d_backward, max_pool2d_with_indices, topk_softmax

  • Triton JIT C++ runtime now ships precompiled kernels for: add, addmm, argmax, bmm, bmm_out, cat, contiguous, embedding, exponential_, fill, flash_attn_varlen_func, fused_add_rms_norm, max, mm, nonzero, reshape_and_cache, reshape_and_cache_flash, rms_norm, rms_norm_backward, rotary_embedding, softmax, sum, topk, zeros

v3.0#

Release date: 2025-07-14

  • Support 184 operators in total, including custom operators used in large model inference
  • New hardware platforms supported: Ascend, AIPU, etc.
  • Added compatibility with the vLLM framework, with the inference verification of DeepSeek model passed
  • Added BLAS operator: dot, mm_out
  • Added tensor operators: index_put_, scatter_, sort_stable
  • Added linear algebra operators: lerp, sum_dim_out, sum_out
  • Added math operators: angle, nan_to_num, polar, tanh_backward
  • Added reduction operators: cummax, cumsum_out, eye, index (alpha), layer_norm_backward, softmax_backward
  • Added neural network operators: batch_norm, batch_norm_backward, dropout_backward, embedding_backward, flash_attention_forward, gelu_backward, glu, group_norm_backward, log_softmax (stable), log_softmax_backward, sigmoid_backward, silu_backward, threshold, threshold_backward, weight_norm_interface_backward
  • Removed neural network operator: cross_entropy_loss, instance_norm
  • Removed linear algebra operators: outer (fused),

v2.2#

Release date: 2025-04-17

  • Added BLAS operators: vdot
  • Added tensor operators: cat, constant_pad_nd, contiguous (alpha), diag, diag_embed, fill, fill_, hstack, index_add, index_put, isin, kron, linspace, masked_fill, masked_fill_, quantile, repeat_interleave_self_tensor, repeat_interleave_tensor, scatter, select_scatter, slice_scatter, sort, stack, vstack, where_out (stable),
  • Added linear algebra operators: diagonal_backward
  • Added convolution operators: _conv_depthwise2d,
  • Added neural network operators: _upsample_bicubic2d_aa, elu, gelu_, instance_norm (fused), mse_loss, nll_loss_backward, nll_loss_forward, nll_loss2d_backward, nll_loss2d_forward, relu_, scaled_dot_product_attention, scaled_dot_product_attention_backward, scaled_dot_product_attention_forward, sigmoid_, silu_, upsample_nearest2d, weight_norm_interface,
  • Added math operators: abs_, add_, bitwise_and_, bitwise_not_, bitwise_or_, clamp_, cos_, div_mode_, exp_, floor_divide_, log, log_sigmoid, logical_and, logical_not, logical_or, logical_xor, mul_, pow_, reciprocal_, remainder, remainder_, rsqrt_, sin_, sub_, tanh_,
  • Added reduction operators: argmin, count_nonzero, cummin, gather, gather_backward,
  • Added science operators: erf_,
  • Added distribution operators: randperm,

v2.1#

Release date: 2024-09-05

  • Added Tensor operators: _unique2, arange, flip, full, full_like, index_select, masked_fill (alpha), masked_select, ones, ones_like, pad, repeat, tile, unique, where, zeros, zeros_like
  • support neural network operator: embedding
  • support basic math operators: allclose, floor_divide, isclose, isfinite, maximum, minimum, true_divide, trunc_divide
  • support distribution operators: exponential_, multinomial, nonzero, normal, topk, rand, rand_like, randn, randn_like, topk, uniform_
  • Added science operators: erf, resolve_conj, resolve_neg

v2.0#

Release date: 2024-05-31

  • Added BLAS operators: mv, outer
  • Added math operators: bitwise_and, bitwise_not, bitwise_or, clamp, cos, eq, ge, gt, isinf, isnan, le, lt, ne, neg, or, sigmoid sin, tanh,
  • Added reduction operators: all, amax, any, argmax, cross_entropy_loss, group_norm, log_softmax (alpha), max, min, prod, rms_norm, rms_norm_backward, rms_norm_forward, sum, var_mean, vector_norm,
  • Added neural network operators: apply_rotary_position_embedding, fused_add_rms_norm, gelu_and_mul, outer (alpha), silu_and_mul, silu_and_mul_out, skip_layer_norm

v1.0#

Release date: 2024-05-10

  • Added BLAS operators: addmm, bmm, mm
  • Added math operators: abs, add, div, dropout, exp, gelu, mul, pow, reciprocal, relu, rsqrt, silu, sub, triu
  • Added reduction operators: cumsum, layer_norm, mean, softmax
Calendar March 26, 2026
Backward Packaging and Release 20260314 Summary Forward