Repository navigation

All pull requests

fix: restore caller CUDA stream when dispatch or combine throws
#825anranxia opened Nov 12, 2026
Keep deterministic dispatch indices graph-capturable
#824ZenAlexa opened Nov 11, 2026
[HybridEP] Support MXFP8 low latency mode dispatch
#823Aneureka opened Nov 10, 2026
1
fix: validate routing-weight width before dispatch
#822anranxia opened Nov 9, 2026
feat: add PCIe support with optimized performance
#821lus94 opened Nov 9, 2026
docs: add Simplified Chinese translation (README.zh.md)
#819JasonYeYuhe opened Nov 5, 2026
Probe NVLink bandwidth for the current CUDA device
#8180z5a opened Nov 4, 2026
Release weak LRU entries with their receiver
#8160z5a opened Nov 4, 2026
Keep ibstat parsing within the requested CA
#8140z5a opened Nov 4, 2026
fix: validate group-consistent EPBuffer settings at construction
#811YYuJLiu opened Nov 3, 2026
8
docs: qualify bf16 reduce-scatter support and fix the Engram NCCL note
#809stwilliam503388-creator opened Oct 30, 2026
fix: compute SM/QP counts after handle unpacking in dispatch()
#808stwilliam503388-creator opened Oct 30, 2026
Fix shared event completion after deferred epilogues
#806ZenAlexa opened Oct 28, 2026
test: support PyTorch 2.10 collective APIs
#805bvolpato opened Oct 28, 2026
1
test(engram): release gathered references before profiling
#804bvolpato opened Oct 28, 2026
test: trim redundant profiler warmup callbacks
#803bvolpato opened Oct 28, 2026
test(ep): reuse invariant combine reference data
#802bvolpato opened Oct 28, 2026
test(pp): drop redundant 16 GiB benchmark allocation
#801bvolpato opened Oct 28, 2026
Support optional topk_weights in the reference dispatch
#800lsyx6399 opened Oct 27, 2026
Add EP_RDMA_GBS env to declare RDMA link bandwidth when ibstat under-reports it (fixes under-provisioned num_sms on CX8 800G)
#797sssssakuraaaaa opened Oct 26, 2026
fix(utils): enhance NCCL library directory resolution and environment robustness
#796reginaldalfret opened Oct 25, 2026
fix(build): stop helper scripts after failed commands
#795bvolpato opened Oct 25, 2026
Pack channel tail signals by runtime rank count (fix ~2x v2.5 hybrid dispatch regression)
#794sssssakuraaaaa opened Oct 23, 2026
Fix HybridEP metadata scan forward progress with cooperative launch
#792Autumn1998 opened Oct 23, 2026Collaborator
Make symmetric-window ordering configurable: opt into relaxed ordering to recover ~6x hybrid inter-node EP bandwidth
#791sssssakuraaaaa opened Oct 23, 2026

Navigation Menu

Repository navigation

All pull requests

fix: restore caller CUDA stream when dispatch or combine throws

Keep deterministic dispatch indices graph-capturable

[HybridEP] Support MXFP8 low latency mode dispatch

fix: validate routing-weight width before dispatch

feat: add PCIe support with optimized performance

docs: add Simplified Chinese translation (README.zh.md)

Probe NVLink bandwidth for the current CUDA device

Release weak LRU entries with their receiver

Keep ibstat parsing within the requested CA

fix: validate group-consistent EPBuffer settings at construction

docs: qualify bf16 reduce-scatter support and fix the Engram NCCL note

fix: compute SM/QP counts after handle unpacking in dispatch()

Fix shared event completion after deferred epilogues

test: support PyTorch 2.10 collective APIs

test(engram): release gathered references before profiling

test: trim redundant profiler warmup callbacks

test(ep): reuse invariant combine reference data

test(pp): drop redundant 16 GiB benchmark allocation

Support optional `topk_weights` in the reference dispatch

Add EP_RDMA_GBS env to declare RDMA link bandwidth when ibstat under-reports it (fixes under-provisioned num_sms on CX8 800G)

fix(utils): enhance NCCL library directory resolution and environment robustness

fix(build): stop helper scripts after failed commands

Pack channel tail signals by runtime rank count (fix ~2x v2.5 hybrid dispatch regression)

Fix HybridEP metadata scan forward progress with cooperative launch

Make symmetric-window ordering configurable: opt into relaxed ordering to recover ~6x hybrid inter-node EP bandwidth

Pull requests