GitLab

* independent logfmt_simulate function

* draft: logfmt low latency combine

* Minor bug fixes

* Fix non-logfmt bugs

* Fix logfmt bugs

* Fix logfmt bugs

* Minor fix

* Minor fix

* Clean code

* Clean code

* Use fewer regs

* Use two warp groups

* Correct shared memory size

* Minor fix

* Minor fix

* More rigorous tests

* Clean code

* Use more SMs

* Use different unroll factor for send & recv

* Update csrc/kernels/internode_ll.cu
Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com>

* Update csrc/kernels/internode_ll.cu
Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com>

* Some renaming

* Some comments of tests

* Format `logfmt_encode`

* More lints

* Some refactors on sends

* Fix testing

* Fix bugs

* Renaming

* Use the full warp

* Unify combine recv

* Lint

* Lint

* Support 2560

* Fix meta buffer dtype

* Better encode calls

* Better amin/max writes

* Extra sync

* Read `topk_idx` by once

* Better specialization

* Read weights by once

* Rename

* Bug fixed

* Some renaming

* Fix local memory usage for sending

* Fix local memory usage for receiving

* Less writes

* Optimize performance

* Optimize performance

* Better performance

* Optimize performance

* Fix rounding

* Manually unroll

* Fix bench

---------
Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com>
Co-authored-by: Chenggang Zhao <chenggangz@deepseek.com>

* Add diagnosis module for precise identification of slow ranks
Signed-off-by: wangfakang <fakangwang@gmail.com>

* Add tests for the slow diagnosis module
Signed-off-by: wangfakang <fakangwang@gmail.com>

* Update some comments for diagnose
Signed-off-by: wangfakang <fakangwang@gmail.com>

* Update test case for diagnose
Signed-off-by: wangfakang <fakangwang@gmail.com>

* Strip the diagnose module, thx LyricZhao and sphish.
Signed-off-by: wangfakang <fakangwang@gmail.com>

* update variable name and cumulative wait recv cost, thx sphish.
Signed-off-by: wangfakang <fakangwang@gmail.com>

* remove invalid comments.
Signed-off-by: wangfakang <fakangwang@gmail.com>

---------
Signed-off-by: wangfakang <fakangwang@gmail.com>

* Add arg --pressure-test for test_low_latency.py

Add arg --pressure-test for test_low_latency.py
Signed-off-by: Guangguan Wang <guangguan.wang@linux.alibaba.com>

* Export NVSHMEM_QP_DEPTH

Export NVSHMEM_QP_DEPTH
Signed-off-by: Guangguan Wang <guangguan.wang@linux.alibaba.com>

---------
Signed-off-by: Guangguan Wang <guangguan.wang@linux.alibaba.com>

* Strengthen the barrier in `cached_notify`

* lint

* Change the timing method

* lint

* Explicitly destroy the C++ runtime and release resources.

* Small fix

* fix typo

* Add a flag to control whether explicit `destroy()` is required.

* Let forwarders use a dedicated SM

* Shuffle rdma idx

* Sender use TMA.

* Adjust the tuning chunk size.

* Modify NVL chunk layout.

* Update some combine config.

* Small lint

* Minor fix

* Overlap TMA store

---------
Co-authored-by: Chenggang Zhao <chenggangz@deepseek.com>

* Add LogFMT interface

* Update comments

* Add simulated code

* Fix comments

* Change to 128 channels

* Add notes

* Optimize performance

* optimize simulate logfmt 10bit

* Minor fix

* Stronger low latency tests

* Minor fix

* Stronger low latency tests for logfmt

* Optimize logfmt simulate: lg2/ex2 ptx, step_inv

* Minor fix

* Minor fix

* Add non-logfmt bench

* Fix value=0 for logfmt

* Optimize performance

* Refactor tests

---------
Co-authored-by: Zhean Xu <xza@deepseek.com>

* Add TMA buffer allocation

* Use TMA for forwarders and NVL receivers

* Use lane 31 to operate TMA.

* Change rdma buffer layout.

* Use TMA to transfer scales also.

* Increase the NVL recv buffer size.

* Disable early stopping.

* Apply similar optimizations on receiver warps.

* Prevent warp divergence.

* Disable aggressive ptx by default.

* Revert using TMA to transfer scales.

* Format.

* Change the layout of dispatch NVL buffer.

* Move topk transformation to recv warps.

* Use TMA to transfer all data in foward warps

* Use TMA to store scales.

* Code lint

---------
Co-authored-by: Chenggang Zhao <chenggangz@deepseek.com>

* use cli arg for num_processes
Signed-off-by: youkaichao <youkaichao@gmail.com>

* update low-latency
Signed-off-by: youkaichao <youkaichao@gmail.com>

* update intranode
Signed-off-by: youkaichao <youkaichao@gmail.com>

* update internode
Signed-off-by: youkaichao <youkaichao@gmail.com>

---------
Signed-off-by: youkaichao <youkaichao@gmail.com>

* more

* more

* more

* more

* more

* more

* Support bias.

* Fix.

* Fix style.

* Add draft

* Add fast-debugging flags

* Fix several bugs

* Add sender timeout checks

* Fix stuck

* Fix bugs

* Fix bugs

* Increase the test round.

* Add warp synchronization.

* Shuffle the send warps.

* Add time elapsed into bench result.

let the sender SM use the channel_id, and the receiver SM use channel_id + num_channels

* Update README

* Update `setup.py`

* Fix headers

* Add `DISABLE_NVSHMEM` for APIs

* Fix launch

* Fix TMA settings

* Fix TMA usages

* Fix dlink

* Separate layout kernels

* Update version

* Add `is_sm90_compiled`

* Fix tests

* Add NVLink connection checks

* Update README

* Fix tests

* Add some comments

* Minor fix

* Minor fix

* Fix bugs

* Add low-latency kernel usage flag

* Update comments

* Update CMake files

* Use TMA instead of LD/ST for intranode dispatch

* Use TMA instead of LD/ST for intranode combine

* Adjust configs

* Test default configs as well

* More warps for combine

* Add inter-thread fence

* Enable more warps

* Do not use TMA for senders

* Update configs

* Remove useless wait

In the Internode Normal Kernel, when using nvshmem ibrc for RDMA data transmission, a single QP is used for data transfer between two GPUs, which limits kernel performance in network card dual-port and RoCE network scenarios.

In our optimized Internode Normal Kernel, we implemented multiple QPs for data transmission between two GPUs, setting a different QP for each channel. Additionally, we modified the transmission method from IBRC to IBGAD.

Through these optimizations, the Internode Normal Kernel achieves optimal performance in both H800 and H20 environments, with RDMA transmission performance nearly reaching the physical network performance limit. Using the current default statistical method, in 4-node H800 and H20 environments, RDMA performance can reach 60GB/s+.