Support intranode DeepEP without IMEX fabric by rchalamala · Pull Request #7 · modal-projects/foundry

rchalamala · 2026-07-23T06:45:47Z

Summary

define the approved first-class deepep_transport contract for vLLM and SGLang
design cross-process VMM IPC v2 using authenticated SCM_RIGHTS descriptor transfer
preserve fabric as the backward-compatible default
provide a test-driven implementation plan covering native, standalone DeepEP, and full-engine validation

Current state

The approved design and detailed implementation plan are committed. Implementation follows docs/superpowers/plans/2026-07-23-no-imex-deepep.md.

Co-authored-by: Rahul Chalamala <rchalamala@users.noreply.github.com>

Signed-off-by: Cursor Agent <cursoragent@cursor.com> Co-authored-by: Rahul Chalamala <rchalamala@users.noreply.github.com>

cursoragent and others added 30 commits July 23, 2026 06:45

docs: design no-IMEX DeepEP transport

d1a7237

Co-authored-by: Rahul Chalamala <rchalamala@users.noreply.github.com>

docs: plan no-IMEX DeepEP support

6d8ae26

Co-authored-by: Rahul Chalamala <rchalamala@users.noreply.github.com>

fix: pass default CUDA graph pool

785aadf

Co-authored-by: Rahul Chalamala <rchalamala@users.noreply.github.com>

test: cover cross-process VMM IPC

7ed212d

Co-authored-by: Rahul Chalamala <rchalamala@users.noreply.github.com>

fix: support cross-process VMM IPC

c7fd766

Co-authored-by: Rahul Chalamala <rchalamala@users.noreply.github.com>

test: cover DeepEP transport configuration

ccda6c5

Co-authored-by: Rahul Chalamala <rchalamala@users.noreply.github.com>

feat: configure DeepEP transport in TOML

4287591

Co-authored-by: Rahul Chalamala <rchalamala@users.noreply.github.com>

test: cover archived DeepEP transport profiles

d6165b7

Signed-off-by: Cursor Agent <cursoragent@cursor.com> Co-authored-by: Rahul Chalamala <rchalamala@users.noreply.github.com>

feat: validate archived DeepEP transport

d74f290

Signed-off-by: Cursor Agent <cursoragent@cursor.com> Co-authored-by: Rahul Chalamala <rchalamala@users.noreply.github.com>

test: cover DeepEP transport engine hooks

b0234e5

Signed-off-by: Cursor Agent <cursoragent@cursor.com> Co-authored-by: Rahul Chalamala <rchalamala@users.noreply.github.com>

test: preserve vLLM first-save profiling

60477ff

Signed-off-by: Cursor Agent <cursoragent@cursor.com> Co-authored-by: Rahul Chalamala <rchalamala@users.noreply.github.com>

feat: apply DeepEP transport in engine hooks

db3555c

Signed-off-by: Cursor Agent <cursoragent@cursor.com> Co-authored-by: Rahul Chalamala <rchalamala@users.noreply.github.com>

style: format DeepEP hook changes

d816458

Signed-off-by: Cursor Agent <cursoragent@cursor.com> Co-authored-by: Rahul Chalamala <rchalamala@users.noreply.github.com>

test: preserve vLLM first-save metadata

efc8728

Signed-off-by: Cursor Agent <cursoragent@cursor.com> Co-authored-by: Rahul Chalamala <rchalamala@users.noreply.github.com>

fix: preserve vLLM first-save metadata

69efcd0

Signed-off-by: Cursor Agent <cursoragent@cursor.com> Co-authored-by: Rahul Chalamala <rchalamala@users.noreply.github.com>

test: cover no-IMEX experimental recipes

694f22f

Signed-off-by: Cursor Agent <cursoragent@cursor.com> Co-authored-by: Rahul Chalamala <rchalamala@users.noreply.github.com>

test: cover no-IMEX recipe documentation

2ec7250

Signed-off-by: Cursor Agent <cursoragent@cursor.com> Co-authored-by: Rahul Chalamala <rchalamala@users.noreply.github.com>

docs: add no-IMEX DeepEP recipes

f743855

Signed-off-by: Cursor Agent <cursoragent@cursor.com> Co-authored-by: Rahul Chalamala <rchalamala@users.noreply.github.com>

test: cover modern SGLang attention metadata API

0a5d36b

Signed-off-by: Cursor Agent <cursoragent@cursor.com> Co-authored-by: Rahul Chalamala <rchalamala@users.noreply.github.com>

fix: support modern SGLang attention metadata API

4c30d5e

Signed-off-by: Cursor Agent <cursoragent@cursor.com> Co-authored-by: Rahul Chalamala <rchalamala@users.noreply.github.com>

test: cover vLLM NVSHMEM host auto-detection

f48ef04

Signed-off-by: Cursor Agent <cursoragent@cursor.com> Co-authored-by: Rahul Chalamala <rchalamala@users.noreply.github.com>

fix: detect installed NVSHMEM host library for vLLM

384017d

Signed-off-by: Cursor Agent <cursoragent@cursor.com> Co-authored-by: Rahul Chalamala <rchalamala@users.noreply.github.com>

test: cover CUDA graph dependency edge metadata

8a4a0bd

Signed-off-by: Cursor Agent <cursoragent@cursor.com> Co-authored-by: Rahul Chalamala <rchalamala@users.noreply.github.com>

test: link edge metadata helper to CUDA driver

721f783

Signed-off-by: Cursor Agent <cursoragent@cursor.com> Co-authored-by: Rahul Chalamala <rchalamala@users.noreply.github.com>

fix: preserve CUDA graph dependency edge metadata

d81340e

Signed-off-by: Cursor Agent <cursoragent@cursor.com> Co-authored-by: Rahul Chalamala <rchalamala@users.noreply.github.com>

test: use canonical graph archive names

52ca812

Signed-off-by: Cursor Agent <cursoragent@cursor.com> Co-authored-by: Rahul Chalamala <rchalamala@users.noreply.github.com>

test: compare edge replay output on CPU

d98f9b3

Signed-off-by: Cursor Agent <cursoragent@cursor.com> Co-authored-by: Rahul Chalamala <rchalamala@users.noreply.github.com>

test: isolate legacy parallel graph fixture

aa6a920

Signed-off-by: Cursor Agent <cursoragent@cursor.com> Co-authored-by: Rahul Chalamala <rchalamala@users.noreply.github.com>

style: format edge metadata CUDA helper

60ac2a6

Signed-off-by: Cursor Agent <cursoragent@cursor.com> Co-authored-by: Rahul Chalamala <rchalamala@users.noreply.github.com>

test: cover mixed CUDA graph edge metadata

fc173a3

Signed-off-by: Cursor Agent <cursoragent@cursor.com> Co-authored-by: Rahul Chalamala <rchalamala@users.noreply.github.com>

cursoragent and others added 8 commits July 24, 2026 02:07

fix: batch CUDA dependencies by edge metadata

f7f27a4

Signed-off-by: Cursor Agent <cursoragent@cursor.com> Co-authored-by: Rahul Chalamala <rchalamala@users.noreply.github.com>

style: format dependency batching helper

622e49b

Signed-off-by: Cursor Agent <cursoragent@cursor.com> Co-authored-by: Rahul Chalamala <rchalamala@users.noreply.github.com>

test: cover SGLang LOAD logits shard pre-copy

550b9d5

Signed-off-by: Cursor Agent <cursoragent@cursor.com> Co-authored-by: Rahul Chalamala <rchalamala@users.noreply.github.com>

fix: pre-copy SGLang LOAD local logits shard

7c8323a

Signed-off-by: Cursor Agent <cursoragent@cursor.com> Co-authored-by: Rahul Chalamala <rchalamala@users.noreply.github.com>

test: cover SGLang regular TP logits shard pre-copy

5cf1906

Signed-off-by: Cursor Agent <cursoragent@cursor.com> Co-authored-by: Rahul Chalamala <rchalamala@users.noreply.github.com>

test: complete regular TP pre-copy coverage

c39b595

Signed-off-by: Cursor Agent <cursoragent@cursor.com> Co-authored-by: Rahul Chalamala <rchalamala@users.noreply.github.com>

fix: pre-copy SGLang regular TP local shard

b2d4079

Signed-off-by: Cursor Agent <cursoragent@cursor.com> Co-authored-by: Rahul Chalamala <rchalamala@users.noreply.github.com>

test: complete fake SGLang gather modules

3e32b73

Signed-off-by: Cursor Agent <cursoragent@cursor.com> Co-authored-by: Rahul Chalamala <rchalamala@users.noreply.github.com>

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

Support intranode DeepEP without IMEX fabric#7

Support intranode DeepEP without IMEX fabric#7
rchalamala wants to merge 38 commits into
mainfrom
cursor/no-imex-deepep-f10b

rchalamala commented Jul 23, 2026 •

edited by cursor Bot

Loading

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Uh oh!

Conversation

rchalamala commented Jul 23, 2026 • edited by cursor Bot Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary

Current state

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

rchalamala commented Jul 23, 2026 •

edited by cursor Bot

Loading