Skip to content

Navigation Menu

Appearance settings

View all features
- BY COMPANY SIZE
  Enterprises
  Small and medium teams
  Startups
  Nonprofits
- BY USE CASE
  App Modernization
  DevSecOps
  DevOps
  CI/CD
  View all use cases
- BY INDUSTRY
  Healthcare
  Financial services
  Manufacturing
  Government
  View all industries
View all solutions
- EXPLORE BY TOPIC
  AI
  Software Development
  DevOps
  Security
  View all topics
- EXPLORE BY TYPE
  Customer stories
  Events & webinars
  Ebooks & reports
  Business insights
  GitHub Skills
- SUPPORT & SERVICES
  Documentation
  Customer support
  Community forum
  Trust center
  Partners
View all resources
- COMMUNITY
  GitHub SponsorsFund open source developers
- PROGRAMS
  Security Lab
  Maintainer Community
  Accelerator
  GitHub Stars
  Archive Program
- REPOSITORIES
  Topics
  Trending
  Collections
- ENTERPRISE SOLUTIONS
  Enterprise platformAI-powered developer platform
- AVAILABLE ADD-ONS
  GitHub Advanced SecurityEnterprise-grade security features
  Copilot for BusinessEnterprise-grade AI features
  Premium SupportEnterprise-grade 24/7 support
Pricing

Search code, repositories, users, issues, pull requests...

Search

Clear

Search syntax tips

Provide feedback

We read every piece of feedback, and take your input very seriously.

Include my email address so I can be contacted

Saved searches

Use saved searches to filter your results more quickly

Name

Query

To see all available qualifiers, see our documentation.

Appearance settings

You signed in with another tab or window. Reload to refresh your session. You signed out in another tab or window. Reload to refresh your session. You switched accounts on another tab or window. Reload to refresh your session.

Dismiss alert

Uh oh!

There was an error while loading. Please reload this page.

modal-projects / foundry Public

forked from foundry-org/foundry

Notifications You must be signed in to change notification settings
Fork 0
Star 0

Code
Pull requests 7
Actions
Projects
Security and quality
Insights

Additional navigation options

Code
Pull requests
Actions
Projects
Security and quality
Insights

[sglang] Persist TP, prefill, NEXTN, and DeepEP graphs on main - #4

#4

Draft

rchalamala wants to merge 161 commits into

mainmodal-projects/foundry:mainfrom

cursor/tp-consolidation-2e2cmodal-projects/foundry:cursor/tp-consolidation-2e2cCopy head branch name to clipboard

Conversation Commits161 (161)Checks Files changed

Draft

[sglang] Persist TP, prefill, NEXTN, and DeepEP graphs on main#4
rchalamala wants to merge 161 commits into
mainfrom
cursor/tp-consolidation-2e2c

Commits

Commits on Jul 23, 2026

[fix] add ep ipc fix, deepep works without nvl_bytes=0

xenshinu
authored and
cursoragent
committed
[sglang] add DeepEP no-fabric IPC policy

devin-ai-integration[bot]
authored and
cursoragent
committed
[sglang] add no-fabric DeepEP IPC recipe

devin-ai-integration[bot]
authored and
cursoragent
committed
[sglang] fail clearly on old DeepEP API

devin-ai-integration[bot]
authored and
cursoragent
committed
[sglang] move inspect import to module top

devin-ai-integration[bot]
authored and
cursoragent
committed
[sglang] enforce DeepEP use_fabric requirement at Buffer construction

devin-ai-integration[bot]
authored and
cursoragent
committed
[sglang] add experimental dense tensor-parallel recipe

devin-ai-integration[bot]
authored and
cursoragent
committed
[fix] fix workspace mismatch when using fa2

xenshinu
authored and
cursoragent
committed
test: add reproducible SGLang TP validation

cursoragent
and
rchalamala
committed
fix: keep TP NCCL transport symmetric

cursoragent
and
rchalamala
committed
fix: pin TP validation to immutable revisions

cursoragent
and
rchalamala
committed
fix: align TP graph mode across phases

cursoragent
and
rchalamala
committed
docs: mark SGLang tensor parallel validated

cursoragent
and
rchalamala
committed
test: strengthen tensor parallel verification

cursoragent
and
rchalamala
committed
fix: harden graph and IPC restoration

cursoragent
and
rchalamala
committed
fix: version preallocated IPC chunks

cursoragent
and
rchalamala
committed
test: require complete TP replay evidence

cursoragent
and
rchalamala
committed
fix: pass default graph pool explicitly

cursoragent
and
rchalamala
committed
test: compare semantic TP graph archives

cursoragent
and
rchalamala
committed
docs: make TP validation invocation portable

cursoragent
and
rchalamala
committed
test: unpack loaded graph results

cursoragent
and
rchalamala
committed
fix: pin TP validation seed

cursoragent
and
rchalamala
committed
docs: explain deterministic TP validation

cursoragent
and
rchalamala
committed
test: compare deterministic TP graph structure

cursoragent
and
rchalamala
committed
docs: align TP scope with upstream roadmap

cursoragent
and
rchalamala
committed
feat: port SGLang integration to main runner API

cursoragent
and
rchalamala
committed
test: add SGLang main SAVE LOAD smoke

cursoragent
and
rchalamala
committed
fix: fail closed on unsupported SGLang main modes

cursoragent
and
rchalamala
committed
fix: propagate immutable refs into Modal smoke

cursoragent
and
rchalamala
committed
test: run TP validation on SGLang main

cursoragent
and
rchalamala
committed
fix: disable multimem gather under Foundry

cursoragent
and
rchalamala
committed
fix: mirror SGLang main graph warmups on load

cursoragent
and
rchalamala
committed
test: parameterize TP validation models

cursoragent
and
rchalamala
committed
test: exercise SGLang torch symmetric memory TP

cursoragent
and
rchalamala
committed
fix: isolate torch symmetric memory from Foundry VMM

cursoragent
and
rchalamala
committed
test: verify outputs in every TP phase

cursoragent
and
rchalamala
committed
fix: load SGLang main graphs without shared executors

cursoragent
and
rchalamala
committed
test: require deterministic TP all-reduce path

cursoragent
and
rchalamala
committed
fix: disable auto-fused TP all-reduce

cursoragent
and
rchalamala
committed
test: ignore expected optional codec tracebacks

cursoragent
and
rchalamala
committed
fix: ignore optional TorchCodec import traces

cursoragent
and
rchalamala
committed
Revert "fix: load SGLang main graphs without shared executors"
cursoragent
committed
fix: register graph pool during parallel load

cursoragent
and
rchalamala
committed
Revert "fix: register graph pool during parallel load"
cursoragent
committed
test: preserve dependencies for independent graph loads

cursoragent
and
rchalamala
committed
fix: load SGLang graphs independently

cursoragent
and
rchalamala
committed
test: exclude manifest from graph member checks

cursoragent
and
rchalamala
committed
fix: share SGLang graph pool during restore

cursoragent
and
rchalamala
committed
fix: replay SGLang per-shape warmup state

cursoragent
and
rchalamala
committed
Revert "fix: replay SGLang per-shape warmup state"
cursoragent
committed
fix: persist symmetric TP collectives on SGLang main

cursoragent
and
rchalamala
committed
fix: reuse stable symmetric buffer for TP gather

cursoragent
and
rchalamala
committed
fix: match symmetric communicators across process groups

cursoragent
and
rchalamala
committed
fix: rendezvous TP gather on device group

cursoragent
and
rchalamala
committed
fix: capture TP gather with stable symmetric group

cursoragent
and
rchalamala
committed
fix: restore TP graphs through shared pool loader

cursoragent
and
rchalamala
committed
fix: use one explicit graph shape for SGLang TP

cursoragent
and
rchalamala
committed
test: pin validation to current SGLang main

cursoragent
and
rchalamala
committed
fix: restore hybrid Mamba pool sizing

cursoragent
and
rchalamala
committed
test: compare concurrent TP outputs across restore

cursoragent
and
rchalamala
committed
test: verify TP collectives in every rank archive

cursoragent
and
rchalamala
committed
fix: harden SGLang TP restore and documentation

cursoragent
and
rchalamala
committed
test: assert plugin-based TP recipe activation

cursoragent
and
rchalamala
committed
test: use deterministic batched TP requests

cursoragent
and
rchalamala
committed
test: verify DeepEP and RNG replay data

cursoragent
and
rchalamala
committed
style: apply branch lint formatting

cursoragent
and
rchalamala
committed
test: validate DeepEP dispatch payloads semantically

cursoragent
and
rchalamala
committed
test: compare DeepEP payloads on CPU

cursoragent
and
rchalamala
committed
test: compare first batched TP decode token

cursoragent
and
rchalamala
committed
docs: describe fixed-batch TP verification

cursoragent
and
rchalamala
committed
style: format DeepEP semantic checks

cursoragent
and
rchalamala
committed
test: strengthen TP4 and DeepEP correctness oracles

cursoragent
and
rchalamala
committed
test: compare stable eight-token TP batch prefix

cursoragent
and
rchalamala
committed
test: decouple SGLang log scanning from Modal

cursoragent
and
rchalamala
committed
test: load TP log scanner without package import

cursoragent
and
rchalamala
committed
style: format log scanning helpers

cursoragent
and
rchalamala
committed
docs: design SGLang graph mode persistence

cursoragent
and
rchalamala
committed
docs: plan SGLang graph mode support

cursoragent
and
rchalamala
committed
test: define SGLang graph output codec

cursoragent
and
rchalamala
committed
feat: serialize SGLang graph output trees

cursoragent
and
rchalamala
committed
test: cover output codec error paths with importlib load

cursoragent
and
rchalamala
committed
test: define SGLang graph session catalog

cursoragent
and
rchalamala
committed
feat: catalog SGLang graph sessions

cursoragent
and
rchalamala
committed
test: cover catalog RuntimeError guard paths

cursoragent
and
rchalamala
committed
test: define generalized SGLang backend contracts

cursoragent
and
rchalamala
committed
feat: persist multiple SGLang graph sessions

cursoragent
and
rchalamala
committed
test: cover generalized SGLang backend behavior

cursoragent
and
rchalamala
committed
fix: refresh SGLang graph catalogs before operations

cursoragent
and
rchalamala
committed

Commits on Jul 24, 2026

test: define SGLang full graph routing contracts

cursoragent
and
rchalamala
committed
feat: route SGLang full prefill graphs through Foundry

cursoragent
and
rchalamala
committed
test: cover speculative SGLang resolver imports

cursoragent
and
rchalamala
committed
fix: patch speculative SGLang backend resolvers

cursoragent
and
rchalamala
committed
style: organize SGLang resolver imports

cursoragent
and
rchalamala
committed
test: define SGLang-main DeepEP bootstrap contracts

cursoragent
and
rchalamala
committed
test: define SGLang-main DeepEP lifecycle ordering

cursoragent
and
rchalamala
committed
feat: bootstrap DeepEP for SGLang graph restore

cursoragent
and
rchalamala
committed
test: add SGLang graph-mode recipe argument contracts

cursoragent
and
rchalamala
committed
feat: configure SGLang persistent graph modes

cursoragent
and
rchalamala
committed
test: require nargs+ prefill buckets and sanitize graph-mode env

cursoragent
and
rchalamala
committed
fix: emit prefill buckets as one nargs+ CLI invocation

cursoragent
and
rchalamala
committed
test: validate SGLang full prefill restore

cursoragent
and
rchalamala
committed
fix: rebuild SGLang manifest across graph sessions

cursoragent
and
rchalamala
committed
fix: preallocate SGLang load region once

cursoragent
and
rchalamala
committed
test: stabilize SGLang smoke run evidence

cursoragent
and
rchalamala
committed
test: run smoke cleanup in Foundry image

cursoragent
and
rchalamala
committed
test: validate Qwen3.5 NEXTN graph restore

cursoragent
and
rchalamala
committed
fix: preserve process-local SGLang graph state

cursoragent
and
rchalamala
committed
fix: order speculative prefill after verify capture

cursoragent
and
rchalamala
committed
fix: stabilize hybrid prefill graph metadata

cursoragent
and
rchalamala
committed
fix: preserve SGLang full graph warmups

cursoragent
and
rchalamala
committed
fix: mask padded Mamba prefill slots

cursoragent
and
rchalamala
committed
fix: isolate Mamba prefill metadata by shape

cursoragent
and
rchalamala
committed
test: isolate TP phase compiler caches

cursoragent
and
rchalamala
committed
test: validate Qwen3.5 NEXTN graph restore

cursoragent
and
rchalamala
committed
test: package TP cleanup with Foundry image

cursoragent
and
rchalamala
committed
docs: report Qwen3.5 NEXTN acceptance

cursoragent
and
rchalamala
committed
test: validate SGLang DeepEP graph restore

cursoragent
and
rchalamala
committed
fix: package standalone DeepEP harness

cursoragent
and
rchalamala
committed
fix: preserve Modal GPU access after DeepEP install

cursoragent
and
rchalamala
committed
test: retain failed SGLang phase artifacts

cursoragent
and
rchalamala
committed
docs: report SGLang DeepEP acceptance blockers

cursoragent
and
rchalamala
committed
fix: support pinned DeepEP IPC transport

cursoragent
and
rchalamala
committed
test: initialize DeepEP NCCL before capture region

cursoragent
and
rchalamala
committed
test: support pinned DeepEP top-k dtype

cursoragent
and
rchalamala
committed
test: preload pinned DeepEP NVSHMEM host

cursoragent
and
rchalamala
committed
fix: report DeepEP transport after buffer bootstrap

cursoragent
and
rchalamala
committed
test: make DeepEP acceptance transport-aware

cursoragent
and
rchalamala
committed
test: forward supported NVSHMEM IBGDA control

cursoragent
and
rchalamala
committed
fix: honor local-only transport in pinned DeepEP

cursoragent
and
rchalamala
committed
fix: preserve pinned DeepEP source indentation

cursoragent
and
rchalamala
committed
revert: preserve required DeepEP IBGDA state

cursoragent
and
rchalamala
committed
test: classify local NVSHMEM transport probes

cursoragent
and
rchalamala
committed
test: handle interleaved IBGDA probe logs

cursoragent
and
rchalamala
committed
test: separate DeepEP configuration evidence

cursoragent
and
rchalamala
committed
docs: report Task 9 DeepEP acceptance

cursoragent
and
rchalamala
committed
chore: address Task 9 review cleanup

cursoragent
and
rchalamala
committed
test: compose SGLang graph mode acceptance

cursoragent
and
rchalamala
committed
test: keep full-prefill baseline eager

cursoragent
and
rchalamala
committed
fix: pad hybrid prefill replay metadata

cursoragent
and
rchalamala
committed
test: bound combined SGLang request capacity

cursoragent
and
rchalamala
committed
fix: defer full prefill capture until decode

cursoragent
and
rchalamala
committed
test: cap combined SGLang token capacity

cursoragent
and
rchalamala
committed
style: format SGLang integration files

cursoragent
and
rchalamala
committed
style: format SGLang smoke harness

cursoragent
and
rchalamala
committed
test: retain semantic graph diagnostics

cursoragent
and
rchalamala
committed
debug: trace Mamba prefill metadata buffers

cursoragent
and
rchalamala
committed
fix: stabilize Mamba prefill metadata addresses

cursoragent
and
rchalamala
committed
debug: trace hybrid prefill metadata pointers

cursoragent
and
rchalamala
committed
fix: register graph pool during parallel load

cursoragent
and
rchalamala
committed
Revert "fix: register graph pool during parallel load"
cursoragent
committed
fix: load full prefill graphs independently

cursoragent
and
rchalamala
committed
fix: preallocate shape-specific Mamba metadata

cursoragent
and
rchalamala
committed
Revert "fix: load full prefill graphs independently"
cursoragent
committed
fix: prime shape-specific FLA metadata caches

cursoragent
and
rchalamala
committed
fix: preserve Eagle prefill allocation order

cursoragent
and
rchalamala
committed
fix: format Mamba metadata diagnostics

cursoragent
and
rchalamala
committed
test: retain remote graph diagnostics

cursoragent
and
rchalamala
committed
docs: record SGLang graph mode validation

cursoragent
and
rchalamala
committed
fix: prime GDN metadata per prefill shape

cursoragent
and
rchalamala
committed
docs: refresh SGLang graph validation evidence

cursoragent
and
rchalamala
committed
fix: harden SGLang full prefill replay

cursoragent
and
rchalamala
committed

Footer

© 2026 GitHub, Inc.

Footer navigation

Terms
Privacy
Security
Status
Community
Docs
Contact

You can’t perform that action at this time.