minimal_async_ep: split token exchange launches and waits - #3954
Draft
sanketpurandare wants to merge 5 commits into
Draft
minimal_async_ep: split token exchange launches and waits#3954sanketpurandare wants to merge 5 commits into
sanketpurandare wants to merge 5 commits into
Conversation
[ghstack-poisoned]
meta-cla
Bot
added
the
CLA Signed
This was referenced Jul 21, 2026
[ghstack-poisoned]
…aunches and waits" ## Summary Split MinimalAsyncEP dispatch and combine into graph-visible launch and wait operations. Launches perform metadata preparation and start symmetric-memory row copies; waits only join the current stream to completion and return aliases of the pending outputs. Keepalive inputs describe storage read asynchronously by the communication stream. Routing metadata consumed by later operations remains ordinary graph dataflow rather than an artificial wait dependency. ## Why A monolithic custom operator hides the useful scheduling boundary and forces communication to complete before independent expert work can run. Thin aliasing waits expose the boundary while remaining compatible with autograd, functionalization, full Inductor, and CUDA graph wrapping. ## Test Plan - Trace the registered production APIs with fake tensors and verify launch/wait dependency edges. - Validate wait aliasing through structured operator schema metadata. - Compile wait forward and backward paths with full Inductor. - Compare two-rank outputs and gradients with a local reference. - Trace the production APIs through the CUDA graph pass. [ghstack-poisoned]
This was referenced Jul 29, 2026
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Stack from ghstack (oldest at bottom):
Summary
Split MinimalAsyncEP dispatch and combine into graph-visible launch and wait
operations. Launches perform metadata preparation and start symmetric-memory
row copies; waits only join the current stream to completion and return aliases
of the pending outputs.
Keepalive inputs describe storage read asynchronously by the communication
stream. Routing metadata consumed by later operations remains ordinary graph
dataflow rather than an artificial wait dependency.
Why
A monolithic custom operator hides the useful scheduling boundary and forces
communication to complete before independent expert work can run. Thin aliasing
waits expose the boundary while remaining compatible with autograd,
functionalization, full Inductor, and CUDA graph wrapping.
Test Plan
dependency edges.