Search results

Performance Improvement: Implement custom op for MoE token combination

CLA SignedThis label is managed by the Meta Open Source bot.This label is managed by the Meta Open Source bot.

high priority

Status: Open (in progress).

pytorch/torchtitannumber 2466

#2466 In pytorch/torchtitan;

· hirschsn opened on Mar 2, 2026

FlexAttention with wrap_inductor_compiled_regions doesn't work

bugSomething isn't workingSomething isn't working

high priority

triage review

Status: Open.

#2463 In pytorch/torchtitan;

· fegin opened on Feb 28, 2026

qwen3 TP + CP bug

high priority

triage review

Status: Open.

Bug

#2446 In pytorch/torchtitan;

· tianyu-l opened on Feb 27, 2026

·  

Bugs, Composability

Execute moe.gate in float32

CLA SignedThis label is managed by the Meta Open Source bot.This label is managed by the Meta Open Source bot.

high priority

Status: Open (in progress).

pytorch/torchtitannumber 2389

#2389 In pytorch/torchtitan;

· chelsea0x3b opened on Feb 17, 2026

·  

New Feature, Model, Misc

MoE router replication is broken w/ TP

high priority

triage review

Status: Open.

Bug

#2387 In pytorch/torchtitan;

· volcacius opened on Feb 17, 2026

·  

Bugs, Composability

Feature Request: Async activation offloading to CPU with prefetch

high priority

triage review

Status: Open.

Feature

#2379 In pytorch/torchtitan;

· niyunsheng opened on Feb 13, 2026

·  

New Feature, Model, Misc

Functional issues with TorchTitan when TorchComms NCCL backend is used for running Llama3-8B model

high priority

triage review

Status: Open.

Bug

#2365 In pytorch/torchtitan;

· newtdms opened on Feb 10, 2026

·  

Bugs, Composability

Support for VLMs on Roadmap

high priority

triage review

Status: Open.

Feature

#2320 In pytorch/torchtitan;

· Jeremalloch opened on Feb 3, 2026

·  

New Feature, Model, Misc

Flux attention is_causal need be False + Flex Attention support with is_causal=False

high priority

triage review

Status: Open.

#2297 In pytorch/torchtitan;

· yh8899 opened on Jan 29, 2026

·  

New Feature, Model, Misc

Remove unnecessary token padding for MoE in BF16 mode

CLA SignedThis label is managed by the Meta Open Source bot.This label is managed by the Meta Open Source bot.

high priority

Status: Open (in progress).

pytorch/torchtitannumber 2255

#2255 In pytorch/torchtitan;

· rakkit opened on Jan 20, 2026

·  

Bugs, Composability

[mxfp8 moe training] mxfp8 all to all

ciflow/8gpu

CLA SignedThis label is managed by the Meta Open Source bot.This label is managed by the Meta Open Source bot.

high priority

Status: Open (in progress).

pytorch/torchtitannumber 2250

#2250 In pytorch/torchtitan;

· danielvegamyhre opened on Jan 17, 2026

Expert Parallel Fast Path and Router Dtype Restoration (from solar-open 102B)

high priority

triage review

Status: Open.

Bug

#2225 In pytorch/torchtitan;

· rakkit opened on Jan 13, 2026

·  

New Feature, Model, Misc