24 Pull requests merged by 8 people

• Instructions for using cpuprofile

  #4089 merged Mar 18, 2025

• Update SDPA flash attention API

  #4065 merged Mar 17, 2025

• Mark supports_segmentation=False in test_issue1273

  #4088 merged Mar 17, 2025

• Move debugging instructions from Wiki to main

  #4084 merged Mar 17, 2025

• Mark is_clonable=False to work around a memory corruption

  #4085 merged Mar 17, 2025

• Rename TorchGather to Gather

  #4082 merged Mar 15, 2025

• Another test cleanup

  #4081 merged Mar 15, 2025

• Safe resize vectorization

  #3906 merged Mar 15, 2025

• Cleaning up test_gpu_indexing_ops.cpp

  #4080 merged Mar 14, 2025

• Run a DeepSeek-V3 transformer block from Hugging Face

  #4009 merged Mar 14, 2025

• add register count checks for warp specialization with register sharing

  #4061 merged Mar 14, 2025

• Add warning about kir::IfThenElse

  #4073 merged Mar 14, 2025

• Make Hopper mma tests sparse

  #4060 merged Mar 13, 2025

• Add missing block sync for TMem

  #4026 merged Mar 13, 2025

• Fix C++23 backport of zip and enumerate

  #4068 merged Mar 13, 2025

• Enable TMABankConflictFreeTranspose

  #4027 merged Mar 13, 2025

• [CudaIpc 3/3]: p2p get-Zcopy

  #3911 merged Mar 12, 2025

• Expose MultiDeviceExecutor and overlapped AG+matmul to python API

  #3923 merged Mar 12, 2025

• Globally use TensorIndexer if TMem is used

  #4030 merged Mar 12, 2025

• Translate MatmulOp and LinearOp on Hopper without AxisMapping

  #3986 merged Mar 12, 2025

• register sharing, add launch bound, disable tests with illegal paras

  #4059 merged Mar 12, 2025

• Indexing for TMem ld and st

  #4017 merged Mar 11, 2025

• Tensor memory 32x32b data path pattern matching

  #4015 merged Mar 11, 2025

• Enable hard-coded index for LdMatrix and create basic copy tutorial

  #4039 merged Mar 11, 2025

18 Pull requests opened by 7 people

• Tensor-parallelize the DeepSeek V3 transformer layer

  #4062 opened Mar 12, 2025

• Adding IndexPutAccumulateOp

  #4063 opened Mar 12, 2025

• Add a backprop test

  #4064 opened Mar 12, 2025

• indexAccumulate python api

  #4066 opened Mar 12, 2025

• Print local sizes and strides of each nn.Linear

  #4067 opened Mar 12, 2025

• Load Epilogue Inputs with LdMatrix in Hopper Matmul Scheduler

  #4069 opened Mar 13, 2025

• TMem check the stride of outer dims

  #4070 opened Mar 13, 2025

• Align smem buffer for TMA store at 128B

  #4071 opened Mar 13, 2025

• [WIP] Translate MmaOp patterns properly on Hopper

  #4072 opened Mar 13, 2025

• fixing max_vect_factor initialization

  #4076 opened Mar 14, 2025

• [RFC] Create a basic binding for CPP Fusion in python frontend using Gemini

  #4077 opened Mar 14, 2025

• [WIP] Recompute non-cacheable unmappable tensors

  #4078 opened Mar 14, 2025

• Add Blackwell MMA macros

  #4079 opened Mar 14, 2025

• WAR for false ComputeAtLogicalDomainMap results

  #4083 opened Mar 15, 2025

• Issue 4086

  #4087 opened Mar 17, 2025

• cache unmappable buffer in smem

  #4090 opened Mar 17, 2025

• Update asan instructions to work with .so

  #4092 opened Mar 18, 2025

• Revert "add register count checks for warp specialization with register sharing"

  #4093 opened Mar 18, 2025

1 Issue closed by 1 person

• Vectorization analysis returns wrong Vectorization Factor

  #3640 closed Mar 15, 2025

5 Issues opened by 4 people

• INTERNAL ASSERT FAILED: Inconsistent parallelization found between `TV59` and `TV2`

  #4091 opened Mar 18, 2025

• TestNvFuserFrontend::test_broadcast_mixing corrupts memory.

  #4086 opened Mar 17, 2025

• max_persistent_buffer_size may be smaller than total_reduction_numel

  #4075 opened Mar 13, 2025

• Persistent buffer with broadcast results in inconsistent parallelization

  #4074 opened Mar 13, 2025

• Refactor `IndexLowering::handle(const LoadStoreOp* ldst)`

  #4058 opened Mar 11, 2025

25 Unresolved conversations

Sometimes conversations happen on old items that aren’t yet closed. Here is a list of all the Issues and Pull Requests with unresolved conversations.

• supporting vectorized load on IndexSelectOp

  #4048 commented on Mar 17, 2025 • 11 new comments

• Add benchmarks for cross entropy loss

  #3924 commented on Mar 18, 2025 • 7 new comments

• Check that warps are only accessing the subpartition of TMem that it can access

  #4016 commented on Mar 14, 2025 • 6 new comments

• Concretization replays loop transforms

  #3950 commented on Mar 14, 2025 • 6 new comments

• project persistent buffer to it producer if the buffer is the output of an upcast

  #4051 commented on Mar 14, 2025 • 5 new comments

• DID loop split for reshape without pre-sharding reshape propagation

  #3953 commented on Mar 17, 2025 • 5 new comments

• Reimplement isResharding using index calculation

  #3482 commented on Mar 18, 2025 • 3 new comments

• s/reshape/set when no transforms are applied

  #4056 commented on Mar 11, 2025 • 2 new comments

• Simplify selfAllocationReplay

  #4057 commented on Mar 13, 2025 • 0 new comments

• getOutputShardings checks all TVs to decide single-GPU vs multi-GPU

  #4046 commented on Mar 12, 2025 • 0 new comments

• [WIP] simple L2 model for setting grid swizzle and cta order

  #4044 commented on Mar 12, 2025 • 0 new comments

• Don't revert upcast for persistent schedulers to avoid increasing persistent buffer sizes

  #4040 commented on Mar 12, 2025 • 0 new comments

• revise smem usage estimation

  #4033 commented on Mar 11, 2025 • 0 new comments

• WIP

  #4006 commented on Mar 16, 2025 • 0 new comments

• Remove MmaOp::AxisMapping

  #3995 commented on Mar 12, 2025 • 0 new comments

• redo register sharing PR-3972

  #3993 commented on Mar 17, 2025 • 0 new comments

• Expose backend type to python

  #3928 commented on Mar 12, 2025 • 0 new comments

• [CudaIpc 2/3]: Ipc handle exchange

  #3910 commented on Mar 12, 2025 • 0 new comments

• Enable resize scheduler by default

  #3848 commented on Mar 17, 2025 • 0 new comments

• [WIP] update propagateSharding preseg pass for DID loop split

  #3838 commented on Mar 14, 2025 • 0 new comments

• [WIP] IdModel-based indexing

  #2238 commented on Mar 14, 2025 • 0 new comments

• test_python_frontend.py::test_issue1273 corrupts memory

  #3856 commented on Mar 17, 2025 • 0 new comments

• RuntimeError from normalization_utils.cpp, could not resolve persistent buffer

  #4020 commented on Mar 15, 2025 • 0 new comments

• Allow allocation to be a split and different from loop.

  #3479 commented on Mar 13, 2025 • 0 new comments

• internal assert failure from sync_information.cpp

  #4052 commented on Mar 13, 2025 • 0 new comments