FlagOS 2.1 — FlagGems v5.3.0

Release v5.3.0

Changes since v5.0.2

New Features

• Add simple extensible FlagTune API for selected matmul ops (#3462) by @Fattysand
• Flagtune mm with GROUP_M (#3713) by @Fattysand

Performance

• Add more tests case for to_copy, fix#3339 (#3420) by @bin913
• [Performance Optimization] Reduce test cases in QUICK_MODE (#3423) by @git-flyer
• [Performance Optimization] Add QUICK_MODE support to reduce test suite runtime (#3932) by @git-flyer

Documentation

• Fix links to supported platforms (#3322) by @tengqm
• Fix operator-list shortcode (#3323) by @tengqm
• Add zh-cn version of operator list (#3324) by @tengqm
• Add page generation logic for publishing test results. (#3451) by @tengqm
• Make sure the website renders well under dark theme (#3452) by @tengqm
• Feat: add precision log (#3377) by @YY4994
• Add benchmark data (#3543) by @tengqm
• Update gh-pages (#3763) by @huangyiqun
• Integrate backend code (#3945) by @Galaxy1458
• Integrate backend code (#3945) (#4050) by @0x45f

Testing

• Case add dtypes (#3338) by @tianxiao-baai

Dependencies

• Bump tspascoal/get-user-teams-membership from 4.0.1 to 4.0.2 in the actions-minor group (#3415) by @app/dependabot
• Bump ad-m/github-push-action from 1.2.0 to 1.3.0 in the actions-minor group (#3563) by @app/dependabot

Other

• Refactor import organization and setup logic for spacemit (#3112) by @alex-spacemit
• Improve zero_out test coverage (#3217) by @Schopenhauer-loves-Hegel
• Improve where_self_out test coverage (#3216) by @Schopenhauer-loves-Hegel
• Improve weight_norm_interface_backward test coverage (#3215) by @Schopenhauer-loves-Hegel
• Improve weight_norm test coverage (#3214) by @Schopenhauer-loves-Hegel
• Improve var_correction test coverage (#3213) by @Schopenhauer-loves-Hegel
• Improve trunc_divide test coverage (#3211) by @Schopenhauer-loves-Hegel
• Improve triton_lighting_indexer_k_tiled_interface test coverage (#3210) by @Schopenhauer-loves-Hegel
• Improve tril_ test coverage (#3209) by @Schopenhauer-loves-Hegel
• Revert "Improve triton_lighting_indexer_k_tiled_interface test coverage" (#3263) by @tengqm
• Fix vendor error (#3270) by @0x45f
• Improve threshold_backward test coverage (#3208) by @Schopenhauer-loves-Hegel
• Improve t_copy_out test coverage (#3207) by @Schopenhauer-loves-Hegel
• Improve tanh_backward test coverage (#3206) by @Schopenhauer-loves-Hegel
• Revert "[SiliconFlow] Op: Flash attention backward" (#3310) by @tengqm
• Attach test report to on-demand test (#3311) by @tengqm
• Fix addmm sqmma epilogue precision for bf16 (#3295) by @Oslomayor
• Improve enflame setup (#3249) by @tengqm
• Use triton.testing.do_bench for all devices (#3289) by @Oslomayor
• Improve lerp test coverage (#3286) by @Schopenhauer-loves-Hegel
• Improve addmm test coverage (#3277) by @Schopenhauer-loves-Hegel
• Fix import error and set device query command for spacemit (#3275) by @alex-spacemit
• Improve arange test coverage (#3274) by @Schopenhauer-loves-Hegel
• Improve clamp test coverage (#3271) by @Schopenhauer-loves-Hegel
• Improve copysign test coverage (#3269) by @Schopenhauer-loves-Hegel
• Improve floor_divide test coverage (#3268) by @Schopenhauer-loves-Hegel
• Improve bitwise_and test coverage (#3260) by @Schopenhauer-loves-Hegel
• Add dtypes in consts (#3221) by @tianxiao-baai
• Improve conv3d_padding test coverage (#3116) by @Schopenhauer-loves-Hegel
• Improve constant_pad_nd test coverage (#3113) by @Schopenhauer-loves-Hegel
• Improve conv1d_padding test coverage (#3114) by @Schopenhauer-loves-Hegel
• Improve conv2d_padding test coverage (#3115) by @Schopenhauer-loves-Hegel
• Improve copy test coverage (#3117) by @Schopenhauer-loves-Hegel
• Improve einsum test coverage (#3132) by @Schopenhauer-loves-Hegel
• Improve dropout_backward test coverage (#3129) by @Schopenhauer-loves-Hegel
• Improve gcd_out test coverage (#3142) by @Schopenhauer-loves-Hegel
• Improve eye_m test coverage (#3134) by @Schopenhauer-loves-Hegel
• Improve fill_scalar test coverage (#3135) by @Schopenhauer-loves-Hegel
• Improve dunder_ior_scalar test coverage (#3130) by @Schopenhauer-loves-Hegel
• Improve dunder_or_scalar test coverage (#3131) by @Schopenhauer-loves-Hegel
• Improve eq_scalar test coverage (#3133) by @Schopenhauer-loves-Hegel
• Improve fused_recurrent_gated_delta_rule_fwd test coverage (#3141) by @Schopenhauer-loves-Hegel
• Improve fmin_out test coverage (#3140) by @Schopenhauer-loves-Hegel
• Improve ge_scalar test coverage (#3143) by @Schopenhauer-loves-Hegel
• Improve gt_scalar test coverage (#3148) by @Schopenhauer-loves-Hegel
• Improve inplace_fused_experts test coverage (#3150) by @Schopenhauer-loves-Hegel
• Improve hypot_out test coverage (#3149) by @Schopenhauer-loves-Hegel
• Improve layer_norm_backward test coverage (#3153) by @Schopenhauer-loves-Hegel
• Improve leaky_relu_out test coverage (#3154) by @Schopenhauer-loves-Hegel
• Improve le_scalar test coverage (#3157) by @Schopenhauer-loves-Hegel
• Improve lt_scalar test coverage (#3162) by @Schopenhauer-loves-Hegel
• Improve logaddexp_out test coverage (#3158) by @Schopenhauer-loves-Hegel
• Improve mean_dim test coverage (#3168) by @Schopenhauer-loves-Hegel
• Improve max_dim test coverage (#3165) by @Schopenhauer-loves-Hegel
• Improve min_dim test coverage (#3169) by @Schopenhauer-loves-Hegel
• Improve reflection_pad1d_out test coverage (#3186) by @Schopenhauer-loves-Hegel
• Improve reflection_pad2d_out test coverage (#3187) by @Schopenhauer-loves-Hegel
• Improve select_backward test coverage (#3194) by @Schopenhauer-loves-Hegel
• Improve mm_out test coverage (#3170) by @Schopenhauer-loves-Hegel
• Improve silu_and_mul_out test coverage (#3196) by @Schopenhauer-loves-Hegel
• Improve silu_backward test coverage (#3197) by @Schopenhauer-loves-Hegel
• Improve special_i0e_out test coverage (#3202) by @Schopenhauer-loves-Hegel
• Improve isin test coverage (#3288) by @Schopenhauer-loves-Hegel
• Improve log_softmax test coverage (#3291) by @Schopenhauer-loves-Hegel
• Improve masked_fill test coverage (#3293) by @Schopenhauer-loves-Hegel
• Improve normal test coverage (#3294) by @Schopenhauer-loves-Hegel
• Improve greater test coverage (#3281) by @Schopenhauer-loves-Hegel
• Improve sort_stable test coverage (#3200) by @Schopenhauer-loves-Hegel
• Improve softmax test coverage (#3292) by @Schopenhauer-loves-Hegel
• Tune pyproject setup (#3317) by @tengqm
• Amend labeller for new vendors (#3316) by @tengqm
• Fix setup script for vendor list and some typo (#3318) by @tengqm
• Tweak setup process for enflame backend (#3319) by @tengqm
• Add colorama to kunlunxin dependency (#3320) by @tengqm
• Tweak project setup for missing entries (#3321) by @tengqm
• Add router_gemm op (#3220) by @Jennie1004
• Fix benchmark parse and opt psummary html style (#3283) by @tianxiao-baai
• Case opname fix (#3333) by @tianxiao-baai
• Split the psummary tool (#3331) by @tengqm
• Bump flagtree version for mthreads (#3341) by @tengqm
• Import arch_version directly from the parent package (#3345) by @Galaxy1458
• Improve sum test coverage (#3343) by @Schopenhauer-loves-Hegel
• Improve sigmoid_backward test coverage (#3195) by @Schopenhauer-loves-Hegel
• Migrate SQMMA descriptor to triton 3.6 TensorDescriptor API (#3351) by @Oslomayor
• Add and optimize FP8 block-wise BMM kernel and fp8_einsum (bhr,hdr->bhd) (#3297) by @monellz
• Optimize fused MoE kernel (#3340) by @KingsleyLiu-NV
• Add setup support for spacemit (#3355) by @tengqm
• Support performance comparison for psum_text (#3357) by @tengqm
• Fix test result summary typo (#3360) by @tengqm
• Improve nll_loss_backward test coverage (#3173) by @Schopenhauer-loves-Hegel
• Improve nll_loss2d_backward test coverage (#3172) by @Schopenhauer-loves-Hegel
• Improve ne_scalar test coverage (#3171) by @Schopenhauer-loves-Hegel
• Add fused marlin moe (#3222) by @Ankaluoer
• Fix the summarization tool (#3374) by @tengqm
• Add fused_deepseek_v4_qnorm_rope_kv_rope_quant_insert (#3327) by @liuxiao0909c
• Fix(setup): downgrade spacemit supported Python version to 3.12 (#3385) by @alex-spacemit
• Bump Triton to 3.6 for the NVIDIA backend (#3383) by @tengqm
• Fix errors in psum_text script (#3388) by @tengqm
• Improve any test coverage (#3276) by @Schopenhauer-loves-Hegel
• Bump Pytorch for the NVIDIA backend (#3383) (#3391) by @tengqm
• Fix on-demand test workflow (#3394) by @tengqm
• Fix the error that the span tag was not closed (#3396) by @tengqm
• Improve bitwise_or test coverage (#3262) by @Schopenhauer-loves-Hegel
• Pin github action tags for speedup (#3406) by @tengqm
• Remove deprecated MUSA_ENABLE_SQMMA env references (#3378) by @Oslomayor
• Remove deprecated MUSA_ENABLE_SQMMA monkeypatch in benchmark (#3381) by @Oslomayor
• Fix cpp-ops CI: add missing IS_FP64 arg and update runner label (#3382) by @douxetpur
• Perf(fused_marlin_moe): transposed B layout + tunable BLOCK_SIZE_K (#3375) by @Schopenhauer-loves-Hegel
• Remove deprecated MUSA_ENABLE_SQMMA monkeypatch from tests (#3380) by @Oslomayor
• [Nvidia] Add affine_grid_generator operator with Triton kernel (#3373) by @XDYuanzhuLee
• Improve psum_text tool (#3416) by @tengqm
• Add missing operators.yaml entries for merged KernelGen PRs (#3407) by @factnn
• Miragtion ops to enflame backend (#3335) by @baoqiliu
• Fix deepseekv4 op bug (#3419) by @huangyiqun
• [Nvidia] Add _euclidean_dist operator with Triton kernel (#3371) by @XDYuanzhuLee
• [FlagGems Operator Development Competition] Add conv_transpose2d Triton implementation (#3405) by @lcthej
• Improve remainder test coverage (#3342) by @Schopenhauer-loves-Hegel
• Add Triton top_k_per_row_prefill for DeepSeek V4 (#3279) by @yzw1128
• Fix installation script to support C++ wrapped operators (#3430) by @tengqm
• Skip test_scatter_reduce_two_ test (#3436) by @tengqm
• Use git commit hash ID rather than version tag (#3440) by @tengqm
• Skip accuracy test for fused_marlin_moe (#3442) by @tengqm
• [Nvidia] Fix topk_softplus_sqrt compliance: logger, yaml registration, test naming (#3428) by @yzw1128
• Add benchmark data (#3453) by @tengqm
• Fix grid_sample test (#3457) by @tengqm
• [Nvidia] Add Triton top_k_per_row_decode for DeepSeek V4 (#3431) by @yzw1128
• Optimize flashmla sparse with TLE (#3459) by @huangyiqun
• Add fused_inv_rope_fp8_quant (#3332) by @ly2zyl
• Added the pack_seq_triton op (#3433) by @hellojack163
• Added the unpack_seq_triton op for Deepseek V4 (#3437) by @hellojack163
• Fix test of fused_qnorm_rope_kv_insert (#3461) by @liuxiao0909c
• Add cp gather indexer k quant cache kernel (#3438) by @chenkc2025
• Add indexer k quant and cache kernel (#3439) by @chenkc2025
• Support int64 ops (#3471) by @chongzhouyang
• Revert scatter.py (#3472) by @chongzhouyang
• Update index_add.py (#3473) by @chongzhouyang
• Change tle to tl (#3474) by @chongzhouyang
• Add .claude in .gitignore (#3468) by @0x45f
• Add benchmarks for router_gemm op (#3464) by @Jennie1004
• Fix backend test to ensure that environment are set (#3482) by @tengqm
• Add DeepSeek V4 attention fused_q_kv_rmsnorm op (#3478) by @huangyiqun
• Add DeepSeek V4 attention compute_global_topk_indices_and_lens op (#3480) by @huangyiqun
• Add DeepSeek V4 attention dequantize_and_gather_k_cache op (#3479) by @huangyiqun
• Add DeepSeek V4 attention combine_topk_swa_indices op (#3481) by @huangyiqun
• Benchmark DeepSeekV4 ops against vLLM APIs (#3494) by @huangyiqun
• Add autotune to mthreads mm op (#3493) by @zhzhcookie
• Add vLLM accuracy tests for DeepSeekV4 attention ops (#3500) by @huangyiqun
• Fix/for hygon adap (#3477) by @creaspoggg
• [Nvidia] Add rad2deg operator with Triton kernel (#3445) by @XDYuanzhuLee
• [Nvidia] Add randint operator with Triton kernel (#3446) by @XDYuanzhuLee
• Sort lists (#3502) by @huangyiqun
• Add DeepSeekV4 attention ops to operators yaml (#3505) by @huangyiqun
• Add cauchy operators (#3496) by @fpzh2011
• Fix flash attn varlen error when key cache non-contiguous (#3410) by @0x45f
• Fix copy fp8 (#3509) by @huangyiqun
• Fix typo for kunlunxin setup (#3527) by @tengqm
• Solve elu_backward ops (#3484) by @llaboon
• Add Containerfile for NVIDIA cu128 (#3528) by @tengqm
• [Nvidia] Add reflection_pad1d_backward operator with Triton kernel (#3448) by @XDYuanzhuLee
• Improve nll_loss test coverage (#3296) by @Schopenhauer-loves-Hegel
• [Advanced Compiler] Repeat interleave (#3328) by @AdvancedCompiler
• Fix uv condition check in setup script (#3541) by @tengqm
• Add as_strided_copy operators (#3501) by @winfan1314
• Fix safe_softmax: replace is_cuda check with device.type (#3499) by @Oslomayor
• Add CI support for Sunrise (#3544) by @tengqm
• Fix torch init error (#3551) by @0x45f
• Fix flash varlen C++ launcher arguments (#3560) by @yysheng26
• Fix issues#521 remainder,floor_divide test bug (#3559) by @w1120029931-bit
• Fix infinite recursion of upsample_nearest_exact1d (#3495) by @Oslomayor
• Fix index_add config (#3498) by @Oslomayor
• Drop mkdocs.yml and related workflow (#3580) by @tengqm
• Skip histc accuracy test (#3586) by @tengqm
• Refactor: extract remainder operator from div.py (#3429) by @Schopenhauer-loves-Hegel
• Add KernelGen label to one_hot operator (#3587) by @Schopenhauer-loves-Hegel
• Add KernelGen label to fill_scalar_out and fill_tensor_out (#3588) by @Schopenhauer-loves-Hegel
• Add KernelGen label to unique_consecutive (#3589) by @Schopenhauer-loves-Hegel
• Add KernelGen label to bernoulli_ (#3590) by @Schopenhauer-loves-Hegel
• Add KernelGen label to bincount (#3591) by @Schopenhauer-loves-Hegel
• Add KernelGen label to roll (#3592) by @Schopenhauer-loves-Hegel
• Add KernelGen label to scatter_reduce_ (#3593) by @Schopenhauer-loves-Hegel
• Add KernelGen label to var (#3594) by @Schopenhauer-loves-Hegel
• Bump torch and triton for NVIDIA backend (#3601) by @tengqm
• Add containerfile for NVIDIA CU130 (#3600) by @tengqm
• Change vendor detection logic in device.py (#3595) by @Galaxy1458
• Use tl.constexpr for stride args in block_ptr kernel (#3565) by @Oslomayor
• Add exception handling for CUDA device check (#3603) by @Galaxy1458
• Add W8A16 support and tune W4A16 configs for fused_marlin_moe (#3609) by @Ankaluoer
• Fix issues#595 KeyError (#3606) by @w1120029931-bit
• Update device.py (#3613) by @Galaxy1458
• Fix trunc_divide integer dispatch and floor_div dtype check (#3608) by @stringl1l1l1l
• Fix runner label for mthreads backend (#3617) by @tengqm
• Bump package version for mthreads (#3621) by @tengqm
• Re-enable mthreads backend tests (#3623) by @tengqm
• Add back MKL for mthreads backend (#3631) by @tengqm
• Downgrade MKL on mthreads to 2024.0.0 (#3632) by @tengqm
• Improve concurrent data writing logic for accuracy tests (#3633) by @tengqm
• Improve run_tests script (#3634) by @tengqm
• Fix operator inventory by removing non-existent operators (#3635) by @tengqm
• Optimize the run_tests script (#3639) by @tengqm
• Fix nits in run_tests.py script (#3640) by @tengqm
• Bump torch version on H20 again (#3641) by @tengqm
• Polish run_tests script for time info and status report (#3643) by @tengqm
• Fix nit introduced in last commit (#3644) by @tengqm
• Fix nits in run tests (#3646) by @tengqm
• Use shared queues for batch testing (#3647) by @tengqm
• Fix vender problem (#3660) by @nianqi-tian
• [Nvidia] Add add_rms_norm operator with Triton kernel (#3515) by @LoserCheems
• [Nvidia] Add clamp_max operator with Triton kernel (#3516) by @LoserCheems
• Add flash_mla_with_kvcache op (#3659) by @Jennie1004
• Prefer Triton over FlagTree on Iluvatar (#3666) by @tengqm
• Fix environment setting for iluvatar backend (#3667) by @tengqm
• Fix patch vllm util (#3657) by @0x45f
• Skip flash_mla_sparse_fwd operator (#3692) by @tengqm
• Be tolerant when running batch/daily tests (#3696) by @tengqm
• Enflame to flagos 20260529 (#3602) by @chongzhouyang
• Fix parameters used for benchmarking kunlunxin (#3690) by @tengqm
• Perf/cp gather indexer k quant cache (#3686) by @chenkc2025
• Perf/indexer k quant and cache (#3688) by @chenkc2025
• Restore pre_hook and libtuner for SQMMA (#3364) by @Oslomayor
• Improve atan2 test coverage (#3278) by @Schopenhauer-loves-Hegel
• Improve sparse_mla_fwd_interface test coverage (#3201) by @Schopenhauer-loves-Hegel
• Standardize MetaX backend debug log format (#3712) by @Schopenhauer-loves-Hegel
• Standardize Iluvatar backend debug log format (#3711) by @Schopenhauer-loves-Hegel
• Initial version of container file for mthreads (#3697) by @tengqm
• Add_triton_version_event (#3336) by @Galaxy1458
• Skip the fused_marlin_moe accuracy test due to assertion failures (#3734) by @tengqm
• Scaled_dot_product_attention_backward support non-square tuning, non-causal, and unequal lengths of q and kv sequences (#3607) by @103yiran
• Fix cumprod_ boolean tensor in-place operation (#3747) by @sh1653487844
• Improve run_tests script for progress bar update (#3754) by @tengqm
• Improve run_tests script for progress bar update (#3757) by @tengqm
• Pin triton to 3.6.0 in setup script (#3758) by @tengqm
• Fix assert_async (#3504) by @Oslomayor
• : add ARM64 CPU backend (NEON/SVE2) with INT8 quant and fuse… (#3616) by @kevinzs2048
• Fix unhashable error for benchmakr (#3760) by @0x45f
• Fix cpp test (#3761) by @tengqm
• [Nvidia] Add randint_like operator with Triton kernel (#3447) by @XDYuanzhuLee
• [Nvidia] Add renorm operator with Triton kernel (#3508) by @XDYuanzhuLee
• Bump torch version to 2.9.0 on Kunlunxin backends (#3773) by @tengqm
• [Nvidia] Add remainder operator with Triton kernel (#3513) by @XDYuanzhuLee
• Fix addmm_dtype and addmm_dtype_out accuracy tests for --ref cpu mode (#3762) by @Schopenhauer-loves-Hegel
• Add benchmark test for cumulative sum operation (#3748) by @tianxiao-baai
• Standardize Cambricon backend debug log format (#3737) by @Schopenhauer-loves-Hegel
• Standardize Aipu backend debug log format (#3740) by @Schopenhauer-loves-Hegel
• Fix --ref cpu cross-device tensor comparison (#3750) by @Schopenhauer-loves-Hegel
• [Backend ] Fix_replace_op_bugs (#3785) by @Galaxy1458
• [Nvidia] Add atanh operator with Triton kernel (#3669) by @LoserCheems
• [Nvidia] Add view_copy operator with Triton kernel (#3670) by @LoserCheems
• [Nvidia] Add tensor_split operator with Triton kernel (#3671) by @LoserCheems
• [Nvidia] Add rot90 operator with Triton kernel (#3556) by @XDYuanzhuLee
• [Nvidia] Add upsample_trilinear3d operator with Triton kernel (#3555) by @XDYuanzhuLee
• Optimize fused_marlin_moe (w4a16) (#3778) by @monellz
• [Nvidia] Add split_with_sizes_copy operator with Triton kernel (#3672) by @LoserCheems
• Add searchsorted operator (#3658) by @winfan1314
• Fix github CI for script rename (#3795) by @tengqm
• Update _sunrise backend (#3791) by @Dayrker
• Container file for the iluvatar backend (#3808) by @tengqm
• Fix ops of gcu300 some bugs to enflame backend (#3814) by @chaaa-a
• Revert arm vendor cmd (#3820) by @0x45f
• Fix issue 3733 (#3815) by @huangyiqun
• : add ARM64 CPU backend (NEON/SVE2) with INT8 quant and fuse… (#3775) by @kevinzs2048
• Fix svd & pointwise_dynamic (#3824) by @Dayrker
• All_dim (#3781) by @sh1653487844
• Downgrade torch from 2.12 to 2.11 for Nvidia (#3825) by @tengqm
• Update Triton to 3.6.0+spacemit.a5 and fix mm, argmax, gelu, bmm (#3793) by @alex-spacemit
• Re-enable tsingmicro backend test (#3827) by @tengqm
• Enable backend tests for enflame, spacemit and sunrise (#3830) by @tengqm
• Opname fix (#3780) by @tianxiao-baai
• [Nvidia] Add fmod_ operator with Triton kernel (#3694) by @XDYuanzhuLee
• Add index_reduce_ operator (#3709) by @winfan1314
• Bump metax versions (#3842) by @tengqm
• Fix ix containerfile (#3841) by @tengqm
• [Sunrise backend] fix scatter_reduce & mse_loss (#3829) by @Dayrker
• Simplify pyproject.toml for extra dependency specification (#3843) by @tengqm
• Fix env tsingmicro (#3844) by @tengqm
• Allowing for specifying all GPUs when testing (#3845) by @tengqm
• Bump package versions for tsingmicro backend (#3846) by @tengqm
• Try skip vendor.sh from the setup.sh script (#3848) by @tengqm
• Containerfile for MetaX 3.7.2.1 (#3849) by @tengqm
• Rename some containerfiles (#3850) by @tengqm
• Update pyproject.toml (#3851) by @tengqm
• Bump enflame package dependencies (#3853) by @tengqm
• Containerfile for Hygon 26.04 (#3858) by @tengqm
• Dockerfile for enflame (#3855) by @tengqm
• Fix replace op bugs (#3863) by @Galaxy1458
• Dev sunrise fix v4 (#3865) by @Dayrker
• Fix fused_recurrent (#3878) by @Dayrker
• Tsingmicro chip count (#3887) by @tengqm
• Fix cumprod_ (#3562) by @sh1653487844
• Update containerfile for mthreads backend (#3903) by @tengqm
• Add containerfile for tsingmicro (#3901) by @tengqm
• Add containerfile for Kunlunxin backend (#3914) by @tengqm
• Standardize Tsingmicro backend debug log format (#3743) by @Schopenhauer-loves-Hegel
• Standardize Kunlunxin backend debug log format (#3739) by @Schopenhauer-loves-Hegel
• Update operator registry (#3925) by @tengqm
• [Advanced Compiler] add fp64 dtype test for polar operator (#3746) by @AdvancedCompiler
• Fix package dependency on klx (#3934) by @tengqm
• No gpu op fix (#3935) by @tianxiao-baai
• Add nanmedian operators (#3337) by @winfan1314
• Add unique_dim operator (#3642) by @ShawnXuan
• Fix operator inventory (#3937) by @tengqm
• Add sunrise's constraint (#3822) by @Dayrker
• Sort operators by name (#3938) by @tengqm
• Fix scatter_reduce_two kernel (#3458) by @tengqm
• [Nvidia] Add trunc_ operator with Triton kernel (#3729) by @XDYuanzhuLee
• Regist vllm ops in torch (#3497) by @0x45f
• Standardize Mthreads backend debug log format (#3738) by @Schopenhauer-loves-Hegel
• Containerfile for Ascend backend (#3974) by @tengqm
• Fix has_triton_tle (#3975) by @0x45f
• Fix issue 3585 (#3973) by @huangyiqun
• Add mode operation (#3921) by @fpzh2011
• [c++ wrapper] Add gemma_rms_norm (#3942) by @chen98038
• Clarify operator stages (#3979) by @tengqm
• Update package/environment setting for tsingmicro (#3982) by @tengqm
• [Nvidia] Add _cdist_backward operator with Triton kernel (#3507) by @LoserCheems
• [Fused][Nvidia] Add mrope operator with Triton kernel (#3976) by @LoserCheems
• [Hygon] Add top_k_per_row_prefill fused kernel for sparse attention (#3981) by @LoserCheems
• [Nvidia] Add _unsafe_masked_index operator with Triton kernel (#3731) by @XDYuanzhuLee
• [KernelGen][Iluvatar] Add top_k_per_row_prefill fused kernel for sparse attention (#3999) by @LoserCheems
• [KernelGen][MetaX] Add top_k_per_row_prefill fused kernel for sparse attention (#4000) by @LoserCheems
• Fix tle_enabled (#3990) by @zhzhcookie
• Standardize Hygon backend debug log format (#3742) by @Schopenhauer-loves-Hegel
• Update triton version for sunrise backend (#3989) by @tengqm
• Fix issue 3007 (#3972) by @huangyiqun
• [MetaX] Upgrade sparse_attention with multi-tier dispatch (#3998) by @LoserCheems
• Fix TsingMicro CI (#4010) by @tengqm
• [Nvidia] Add _resize_output operator with Triton kernel (#3860) by @Dingxingdi
• Update tsingmicro container file to support flagtree (#4017) by @tengqm
• [Nvidia] Migrate logical_xor_ inplace operator from experimental to ops (#3745) by @XDYuanzhuLee
• Fix floor_divide for mixed int/float types and fp16/bf16 dtypes (#3751) by @Schopenhauer-loves-Hegel
• Update Triton to 3.6.0+spacemit.a5 and fix mm, argmax, gel… (#3828) by @alex-spacemit
• Fix grouped_mm accuracy test and benchmark configuration (#3752) by @Schopenhauer-loves-Hegel
• Fix sparse_mla TLE kernel group/head-block index decomposition (#3759) by @Schopenhauer-loves-Hegel
• [Nvidia] Migrate deg2rad, fix, negative from experimental to ops (#3764) by @XDYuanzhuLee
• [Nvidia] Migrate addcmul_ operator from experimental to ops (#3766) by @XDYuanzhuLee
• Simplify ascend container file (#4021) by @tengqm
• Update containerfile for tsingmicro backend (#4035) by @tengqm
• Update CI setup for tsingmicro backend (#4044) by @tengqm
• Fix GPU check script for Ascend (#4045) by @tengqm
• Fix version string for 5.3.0 release (#4037) by @tengqm
• Update version string for the master branch (#4038) by @tengqm
• Fix new_full: accuracy_fail (#3962) by @Schopenhauer-loves-Hegel
• [Nvidia] Migrate atanh_operator from experimental to ops (#3768) by @XDYuanzhuLee
• Fix issues:1012 (#3831) by @w1120029931-bit
• [Nvidia] Add unfold_copy operator with Triton kernel (#3788) by @XDYuanzhuLee
• [Nvidia] Add linear operator with Triton kernel (#3797) by @XDYuanzhuLee
• Add setup support for CANN9.0.0 for Ascend backend (#4046) by @tengqm
• Fix on-demand workflow for GPU check (#4047) by @tengqm
• Fix operator inventory (#4048) by @tengqm
• Fix issue 2896 (#4051) by @huangyiqun
• [Nvidia] Add ilshift operator with Triton kernel (#3730) by @XDYuanzhuLee
• Fix Kunlunxin setup (#4082) by @tengqm
• Debug Ascend chip check (#4062) by @tengqm
• Fix unhashable error for benchmakr (#3760) (#4084) by @0x45f
• Fix incomplete operator entry (#4049) by @tengqm
• Add and optimize the Ascend implementation of upsample_linear1d_backward operator (#3944) by @winfan1314
• Fix pow, mul, per_token_group_quant_fp8, allclose, isclose and reshape_and_cache (#3476) by @fangzexian
• Remove buggy plugin dragged in by Triton on kunlunxin (#4090) by @tengqm
• Fix issue 2850 (#4053) by @huangyiqun
• Fix issue 2829 (#4052) by @huangyiqun
• [Nvidia] Add not_equal operator with Triton kernel (#3424) by @XDYuanzhuLee
• [Nvidia] Add reflection_pad3d_backward operator with Triton kernel (#3470) by @XDYuanzhuLee
• Add a new env shell script to source for Ascend (#4094) by @tengqm
• Add script for checking GPU availability on Enflame (#4078) by @tengqm
• [Nvidia] Add bernoulli operator with Triton kernel (#3417) by @XDYuanzhuLee
• Fix thead setup for CI (#4103) by @tengqm
• Flaggems.ops to flaggems. (#4085) by @chongzhouyang
• Fix: fill_ ZeroDivisionError on empty tensor (#4096) by @JosephNew
• Fix weight_norm, aminmax,l eaky_relu, mm, bmm, copysign, signbit, logsumexp,nonzero_numpy and tril (#3953) by @RobertLuobo
• Fix: Ascend fill_ ZeroDivisionError on empty tensor (#4105) by @tengqm
• Add script for GPU availability check on Sunrise (#4102) by @tengqm
• Fix var accuracy failure on Iluvatar BI-V150 (Triton specialization bug) (#4099) by @Schopenhauer-loves-Hegel
• Fix any_dim accuracy failure on Iluvatar BI-V150 (float16 input) (#4098) by @Schopenhauer-loves-Hegel
• Op: flash_attention_backward (#3386) by @LiangJiabaoY
• Amend the env.sh for environment setup on THead backend (#4107) by @tengqm
• Ci(setup.sh): specify source when installing uv (#4106) by @103yiran
• Add code owners (#4109) by @0x45f
• [Advanced Compiler] topk (#3329) by @AdvancedCompiler
• [Nvidia] Add fp8_fp4_mqa_logits Triton kernel (#4092) by @yzw1128
• Improve pixel test coverage (#3344) by @Schopenhauer-loves-Hegel
• [Nvidia] Add special_chebyshev_polynomial_v operator (#3539) by @XDYuanzhuLee
• [Nvidia] Add logical_not_ operator with Triton kernel (#3561) by @XDYuanzhuLee
• Skip segment_reduce benchmark on MUSA (#4115) by @winfan1314
• Update from enflame until 20260616 (#4111) by @chongzhouyang
• Optimize upsample_bicubic2d_aa_backward on Ascend NPU and Fix Tests (#4101) by @winfan1314
• Optimize unpack_seq (#4126) by @chenkc2025
• Optimize pack_seq (#4125) by @chenkc2025
• Optimize mxfp4 w4a16 fused_moe (#4117) by @monellz
• Optimize compute_global_topk_indices_and_lens for DeepSeek V4 attention (#4113) by @LittleShun1214
• Fix VendorDescriptor bug (#4134) by @Galaxy1458
• Upgrade containerfile for nvidia 13.3 (#4135) by @tengqm
• Fix count_nonzero: replace deprecated logical operators and add sparse/negative dim support (#3816) by @yzw1128
• Install vllm=0.21.0 on H20 for operator testing (#4136) by @tengqm
• [Nvidia] Add addcdiv_ operator with Triton kernel (#3857) by @bwbwzzz
• Add fused_indexer_q_rope_quant op (#4081) by @huangyiqun
• Fix containerfile for Hygon (#4138) by @tengqm
• Fix container file for KLX (#4139) by @tengqm
• Update containerfile for MetaX (#4140) by @tengqm
• Fix PyYAML version in pyproject.toml (#4143) by @tengqm
• Update container file for mthreads (#4149) by @tengqm
• Update containerfile for Iluvatar (#4150) by @tengqm
• Update container file for Ascend 8.5.0 (#4151) by @tengqm
• Update project setting for CANN 9.0.0 (#4152) by @tengqm
• Optimize command workflow (#4161) by @tengqm
• Cleanse the operator registry again (#4162) by @tengqm
• Rename 'ilshift__' to 'ilshift' (#4163) by @tengqm
• Fix the vllm installation cache problem (#4164) by @tengqm
• Fix ascend backend (#4165) by @tengqm
• Allow cann850 and cann9 to be specified for workflow (#4166) by @tengqm
• Improve the setup.sh script (#4168) by @tengqm
• Improve mark collection logic (#4169) by @tengqm
• Update masked_fill_ op (#4170) by @Alvin-YCHEN
• Speedup batch test run for non-existent test cases (#4171) by @tengqm
• Reorg CI workflows (#4172) by @tengqm
• Refactor workflows for a bootstrap checkout (#4174) by @tengqm
• Drop some useless tools that are no longer useful or maintained (#4173) by @tengqm
• Update containerfile for tsingmicro (#4197) by @tengqm
• Remove skip annotation from trace benchmark (#4230) by @tengqm
• Use correct vendor string for GPU check (#4235) by @tengqm
• Cache gh CLI install on self-hosted runners (#4241) by @tengqm
• Consolidate backend dispatch using matrix strategy (#4242) by @tengqm
• Fix baddbmm.out: missing_test (#3958) by @Schopenhauer-loves-Hegel
• Optimized baddbmm & w8a8_block_fp8_matmul with FlagOSTune (#4095) by @Fattysand
• [Advanced Compiler] fix dgeglu geglu dreglu reglu skipped test caused by invalid zero-dimension test shape (#3755) by @AdvancedCompiler
• [Advanced Compiler] Fix conv_depthwise2d test with --ref cpu (#3859) by @AdvancedCompiler
• Change logging warning to info (#4246) by @0x45f
• Fix metadata for the tan_ operator (#3983) by @XDYuanzhuLee
• Fix mode operator on Mthreads backend (#4248) by @winfan1314
• Fix(grid_sample): enable large output sizes (#3421) by @mygitljf
• [Nvidia] Add _upsample_nearest_exact2d_backward operator with Triton kernel (#3787) by @XDYuanzhuLee
• Fix nanmedian cross-backend accuracy (#4116) by @winfan1314
• Fix renorm fp16 precision: allocate norms buffer in fp32 (#4237) by @tengqm
• Add benchmark test for addmv_out (#4240) by @tengqm
• Drop experimental operator 'upsample_nearest1d' (#4233) by @tengqm
• [Nvidia] Add kthvalue operator with Triton kernel (#3928) by @XDYuanzhuLee
• Drop the experimental reciprocal operator (#4266) by @tengqm
• Add Iluvatar-specific CodeGenConfig for pointwise_dynamic ops (#4265) by @tengqm
• Drop experimental masked_select operator (#4264) by @tengqm
• Drop experimental operator atanh_ (#4260) by @tengqm
• Drop experimental operator addcmul_ (#4259) by @tengqm
• Drop experimental masked_scatter implementation (#4257) by @tengqm
• Drop experimental operator fix (#4249) by @tengqm
• [Nvidia] Add _linalg_eigvals operator with Triton kernel (#4217) by @XDYuanzhuLee
• Drop experimental operator 'trace' (#4229) by @tengqm
• Drop experimental operator 'diag' (#4221) by @tengqm
• Drop experimental operator 'pixel_shuffle' (#4225) by @tengqm
• Add gather backward tests (#4269) by @huangyiqun
• Add i0 out tests (#4268) by @huangyiqun
• Change to debug level (#4273) by @0x45f
• Add hardsigmoid out tests (#4272) by @huangyiqun
• Fix hardsigmoid_out benchmark (#4271) by @huangyiqun
• [Nvidia] Add _thnn_fused_lstm_cell_backward_impl operator with Triton kernel (#3907) by @XDYuanzhuLee
• Add div_out tests (#4274) by @huangyiqun
• [Nvidia] Add adaptive_max_pool3d_backward operator with Triton kernel (#3906) by @XDYuanzhuLee
• [Nvidia] Add _upsample_bilinear2d_aa operator with Triton kernel (#3813) by @XDYuanzhuLee
• [Nvidia] Add arcsin operator with Triton kernel (#3912) by @XDYuanzhuLee
• [Nvidia] Add expand operator with Triton kernel (#3885) by @XDYuanzhuLee
• [Nvidia] Add unbind_copy operator with Triton kernel (#3811) by @XDYuanzhuLee
• [Nvidia] Add special_log_softmax operator with Triton kernel (#3803) by @XDYuanzhuLee
• [Nvidia] Add special_gammainc operator with Triton kernel (#3872) by @XDYuanzhuLee
• [Nvidia] Add rrelu_with_noise_functional operator with Triton kernel (#3557) by @XDYuanzhuLee
• [Nvidia] Add channel_shuffle operator with Triton kernel (#3800) by @XDYuanzhuLee
• [Nvidia] Add log_normal_ operator with Triton kernel (#3789) by @XDYuanzhuLee
• [Nvidia] Add empty operator with Triton kernel (#3693) by @XDYuanzhuLee
• [Nvidia] Add soft_margin_loss_backward operator with Triton kernel (#3726) by @XDYuanzhuLee
• [Nvidia] Add addmm_ operator with Triton kernel (#3786) by @XDYuanzhuLee
• [Nvidia] Add im2col operator with Triton kernel (#3698) by @XDYuanzhuLee
• [Nvidia] Add amp_foreach_non_finite_check_and_unscale operator with Triton kernel (#3799) by @XDYuanzhuLee
• [Nvidia] Add _jagged_to_padded_dense_forward operator with Triton kernel (#3809) by @XDYuanzhuLee
• Enable_nvidia_unused_ops (#4285) by @Galaxy1458
• [Nvidia] Add BatchNorm operator with Muxi and Tianshu domestic backends (#3664) by @XDYuanzhuLee
• Optimize dequantize_and_gather_k_cache for DeepSeek V4 attention (#4267) by @littlekid-zxt
• [Nvidia] Add _sparse_semi_structured_mm operator with Triton kernel (#3893) by @XDYuanzhuLee
• [Nvidia] Add _prelu_kernel_backward operator with Triton kernel (#3798) by @XDYuanzhuLee
• [Nvidia] Add special_shifted_chebyshev_polynomial_u operator with Triton kernel (#3801) by @XDYuanzhuLee
• [Nvidia] Add special_hermite_polynomial_h operator with Triton kernel (#3807) by @XDYuanzhuLee
• [Nvidia] Add subtract_ operator with Triton kernel (#3870) by @XDYuanzhuLee
• [Nvidia] Add broadcast_to operator with Triton kernel (#3882) by @XDYuanzhuLee
• [Nvidia] Add diagonal_copy operator with Triton kernel (#3889) by @XDYuanzhuLee
• [Nvidia] Add alpha_dropout operator with Triton kernel (#3908) by @XDYuanzhuLee
• [Nvidia] Add erfinv_ operator with Triton kernel (#3919) by @XDYuanzhuLee
• [Nvidia] Add clone operator with Triton kernel (#3917) by @XDYuanzhuLee
• [Nvidia] Add asin operator with Triton kernel (#3915) by @XDYuanzhuLee

Contributors

Thanks to all contributors who made this release possible:

• @tengqm (163 PRs)
• @Schopenhauer-loves-Hegel (98 PRs)
• @XDYuanzhuLee (58 PRs)
• @huangyiqun (24 PRs)
• @0x45f (14 PRs)
• @LoserCheems (12 PRs)
• @Oslomayor (12 PRs)
• @Galaxy1458 (10 PRs)
• @winfan1314 (9 PRs)
• @chongzhouyang (7 PRs)
• @tianxiao-baai (7 PRs)
• @Dayrker (6 PRs)
• @chenkc2025 (6 PRs)
• @AdvancedCompiler (5 PRs)
• @alex-spacemit (5 PRs)
• @yzw1128 (5 PRs)
• @Fattysand (3 PRs)
• @Jennie1004 (3 PRs)
• @monellz (3 PRs)
• @sh1653487844 (3 PRs)
• @w1120029931-bit (3 PRs)
• @103yiran (2 PRs)
• @Ankaluoer (2 PRs)
• @app/dependabot (2 PRs)
• @fpzh2011 (2 PRs)
• @git-flyer (2 PRs)
• @hellojack163 (2 PRs)
• @kevinzs2048 (2 PRs)
• @liuxiao0909c (2 PRs)
• @zhzhcookie (2 PRs)
• @Alvin-YCHEN
• @Dingxingdi
• @JosephNew
• @KingsleyLiu-NV
• @LiangJiabaoY
• @LittleShun1214
• @RobertLuobo
• @ShawnXuan
• @YY4994
• @baoqiliu
• @bin913
• @bwbwzzz
• @chaaa-a
• @chen98038
• @creaspoggg
• @douxetpur
• @factnn
• @fangzexian
• @lcthej
• @littlekid-zxt
• @llaboon
• @ly2zyl
• @mygitljf
• @nianqi-tian
• @stringl1l1l1l
• @yysheng26

---

Generated by Release Notes Generator