[tp] Consolidate branches and validate vLLM/SGLang by rchalamala · Pull Request #6 · modal-projects/foundry

rchalamala · 2026-07-23T02:43:12Z

Summary

consolidate active Foundry branches directly onto main
harden shared VMM/IPC and graph restoration
port SGLang to clean upstream main through a lightweight plugin/HookRegistry integration with zero upstream edits
reject disproven plain-NCCL SGLang restore and use stable torch symmetric-memory collectives with one padded decode graph
retain the pinned vLLM V1/PyNCCL TP profile while final archive/restart hardening is completed
ground scope in upstream Foundry/SGLang/vLLM issues, PRs, and discussions

SGLang main — VERIFIED

Candidate 435518206f501b2f6b20080ee53375d0d6a3ff76

Run: https://modal.com/apps/modal-labs/rahul-dev/ap-RHFcD96CAsGSqO7Z236yzF

SGLang a25164bda355250015c0ba6d8821f0d4b37abe7d
Qwen3-8B b968826d9c46dd6066d109eabc6255188de91218
baseline → SAVE → SAVE2 → LOAD → LOAD2 on 2×H100
exact sequential and native bs=8 outputs in SAVE2/LOAD/LOAD2 versus baseline
one graph saved/restored per rank; offsets 68908220416 in every Foundry phase
73 symmetric all-reduce + one symmetric all-gather node and zero NCCL collective nodes per graph/rank
immutable LOAD/LOAD2 archive bytes, no graph re-save, bs=1/bs=8 replay
all process/error/cleanup checks passed

Focused treatment: https://modal.com/apps/modal-labs/rahul-dev/ap-SwjRFbh7zqLfmKlGGkDuTM

Plain-NCCL restore is explicitly rejected: it restored structurally but produced deterministic corrupt decode output.

Remaining before final handoff

The final branch audit found selective late deltas from the advanced vLLM/no-IMEX branches: native IPC fork/pitched/peer-zone hardening, archive-defined DeepEP transport, strict DeepEP replay output checks, and a complete vLLM SAVE2/LOAD/LOAD2 immutable-archive matrix. These are being ported selectively without merging obsolete SGLang experiments.

Supported scope

Pinned experimental profiles only. SGLang: single-host dense Qwen3-8B TP=2, torch symmetric memory, one full-decode graph. vLLM: single-host dense Qwen3-8B TP=2, V1/full-decode PyNCCL P2P/IPC. Broader topology/backends remain unsupported.

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

Signed-off-by: xenshinu <liuxs@umich.edu>

Co-Authored-By: Rahul Chalamala <22563365+rchalamala@users.noreply.github.com> Signed-off-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>

Signed-off-by: xenshinu <liuxs@umich.edu>

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

The strongest remaining divergence from the proven PR4 backend: our LOAD short-circuited DecodeCudaGraphRunner.capture and manually rebuilt a SimpleNamespace metadata wrapper + pool + seq fill + load_all_graphs. That built the LOAD-time allocator trajectory and FlashInfer metadata wrappers differently from SAVE, so captured kernels referenced wrong workspace addresses and LOAD produced corrupt output despite one graph/rank, positive symmetric AR/AG kernels, zero NCCL nodes, and a stable symmetric buffer. Port PR4's design into the current FullCudaGraphBackend subclass: * backend.py: extend the subclass with LOAD state (_pending / _graph_files / _load_index / _closed) and a capture_session override. On LOAD it establishes the inherited shared pool + stream, preallocates, scans the archive, issues exactly one start_graph_builds(paths, pool=self._pool), yields to the upstream per-shape loop, and on exit fails closed unless every archived file was consumed (then logs the final offset/count). capture_one on LOAD validates the requested dense ShapeKey against the archive order and calls finish_one_graph_load per index — no forward, native capture, or warmup. cleanup() latches _closed to block recapture. SAVE capture path and its skip of the two warmups are unchanged. * hooks.py: the DecodeCudaGraphRunner.capture AROUND LOAD branch now calls original, so the pinned upstream executes graph_capture(), the seq_lens fill, capture_prepare, ForwardContext, real init_forward_metadata_out_graph, and per-shape dispatch — the identical lifecycle SAVE runs. The capture_bs==[256] fail-closed guard and SAVE finalization are unchanged. * graph_ops.py: remove the now-dead manual LOAD path (load_all_graphs, get_or_create_graph_pool) and the SimpleNamespace metadata helpers (build_capture_fb_view, initialize_attention_metadata_for_bs); upstream now owns metadata via a real ForwardBatch. Tests: repurpose test_sglang_graph_load.py to drive the backend LOAD lifecycle through a fake upstream capture() (one-start / per-index finish / exact map+output / order-mismatch / extra / missing-on-exit / no-forward / recapture guard); update test_sglang_capture_hooks.py so LOAD proves it calls original with no manual orchestration and the backend __init__/capture_session/SAVE contracts hold; delete the obsolete attention-metadata helper test. Co-authored-by: Rahul Chalamala <rchalamala@users.noreply.github.com>

…NONE-transparent cleanup Finding 1: the pinned LOAD lifecycle runs the real upstream capture() traversal, so rank0 emits "Capturing batches" progress on LOAD even though the Foundry backend restores archived graphs without any CUDA capture. Using that progress count as a no-recapture gate now falsely fails LOAD/focused/full acceptance. Keep native_capture_progress_count diagnostic and prove no-recapture via positive archived-load counts + zero Foundry graph-save logs instead: rename the focused check (sglang_tp_validation.focused_treatment_checks) and the full check (modal_sglang_tp) to load_restored_without_resave with predicate saved_graph_log_count == 0 and sum(loaded_graphs) > 0. The backend LOAD no-forward/no-CUDAGraph/no-context + exact-restored-maps proof remains in the CPU backend contracts (test_sglang_graph_load.py). Focused contract now also asserts a high native_capture_progress_count still VERIFIES (lifecycle traversal), and flips load_restored_without_resave on a save log or zero restored counts. Finding 2: make subclass cleanup() transparent in NONE mode by delegating to super().cleanup() and not latching _closed, so the runner's legitimate hidden-mode recapture keeps working in stock operation. Active SAVE/LOAD retains the fail-closed recapture latch and its extended cleanup. Adds NONE-transparency and active-latch contracts. Co-authored-by: Rahul Chalamala <rchalamala@users.noreply.github.com>

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

decode_log_interval defaults to 40, so the short concurrent bs>1 decode in the TP acceptance harness emitted no '#running-req ... cuda graph: True' metrics line and the restored_graph_replay_observed evidence silently disappeared. Add --decode-log-interval 1 to the TP recipe and require the resolved value 1 in the supported-profile guard (fails closed on missing/other values), so every phase logs every decode batch. No generation/output logic changed; the bs>1 gate is unchanged. Co-authored-by: Rahul Chalamala <rchalamala@users.noreply.github.com>

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

Address Task 1 review findings 1, 2, 5, 6 (plus policy/source contracts for finding 3): - Claim process ownership at the very start of cuIpcGetMemHandle, cuIpcOpenMemHandle, and cuIpcCloseMemHandle, before any allocation metadata, chunk map, or magic-dependent state is touched. Applied uniformly to plain and VMM handles (conservative all-IPC claim) so fork behavior is consistent; server init is also covered. - Export invalidation now returns before any map access unless the caller is the exact owner: an unclaimed process (-1) has nothing exported and a fork child must not touch the inherited concurrent map. - Fork-child rejection is async-signal-safe: a fixed compile-time message (including 'use spawn/exec') via write(2), plus close(2) of the inherited listener fd; no fprintf/snprintf/malloc/locks. The parked-fd registry is intentionally left untouched (documented) since its concurrent map cannot be locked safely in a fork child. - require_exact_ipc_reservation cleanup callbacks now return CUresult; both run exactly once on the wrong-VA path (a failed address-free never skips the handle release -> no leak; neither is retried -> no double release). The owner process logs cleanup failures. - Factor the ownership decision into a pure, CUDA-free policy header (decide_ipc_ownership / owns_ipc_state / reject message) with a C++ unit test and a hook.cpp source contract for the guards that a CUDA fork child cannot be driven to assert natively. - Mark both host-policy test executables EXCLUDE_FROM_ALL so editable builds skip them; explicit target + ctest paths remain. Co-authored-by: Rahul Chalamala <rchalamala@users.noreply.github.com>

Address Task 1 review findings 3 (native) and 4: - Expand test_vmm_ipc_rejects_use_after_fork so the post-claim fork child exercises cuIpcGetMemHandle, cuIpcOpenMemHandle, and cuIpcCloseMemHandle (each must return CUDA_ERROR_NOT_SUPPORTED) and assert the exact 'use spawn/exec' rejection message on captured stderr; the owner remains usable afterwards. - Add durable native lifecycle regressions that assert observable usability/failure (device memset + memcpy round trips), not prints: - stale-export-fetch: a live same-process import reads the exported bytes; after cuMemFree the parked fd is invalidated and a later open of the same handle fails. - duplicate-interior-open: opening the same imported interior twice is refcounted (same VA, usable after the first close, torn down only after the second, re-openable afterward). - generation-recreate: freeing and recreating a preallocation at the same base bumps the chunk generation and maps fresh data. These reuse the existing LD_PRELOAD spawn helpers and run under the Modal focused suite (local VM has no CUDA device). Co-authored-by: Rahul Chalamala <rchalamala@users.noreply.github.com>

Address remaining Task 1 review findings 5 and 6. Finding 5 (claim ordering + lock-free owner pid): - Move vmm_ipc_claim_process() to be the first executable statement in cuIpcGetMemHandle, cuIpcOpenMemHandle, and cuIpcCloseMemHandle, ahead of the cached CUDA_DRIVER_CALL symbol lookup (which takes a lock a post-claim fork child must not touch). A leading typedef is not executable, so the real entry-point resolution now happens only after the claim/reject decision. - Add static_assert(std::atomic<pid_t>::is_always_lock_free): the owner pid is CAS'd on the async-signal-safe fork-child reject path, where a lock-backed atomic could deadlock. Finding 6 (fallback free + parked-fd docs): - Check and log the initial driver-fallback reservation free in the owning importer before re-reserving in the dedicated import zone; a silent failure would leak that reservation. The wrong-zone cleanup callbacks already fail closed and log. - Document the parked-fd limitation accurately: a rejected fork child cannot lock the export-fd map, so it keeps inherited duplicate fds that pin the exporter's physical allocations until the child exits or execs (FD_CLOEXEC closes them on exec). Source contracts in test_vmm_ipc_ownership_contract.py cover the new claim-before-driver-lookup ordering, the lock-free static_assert, the fallback-free logging, and the parked-fd pinning documentation. Co-authored-by: Rahul Chalamala <rchalamala@users.noreply.github.com>

…oofs Address remaining Task 1 review findings 1-4 in the native lifecycle suite. Every lifecycle subprocess and protocol wait is now bounded (_LIFECYCLE_TIMEOUT_S); hangs fail with captured stdout/stderr and kill the child (finding 4). Helpers _run_capture and _wait_for_file centralize the timeout + diagnostics; the fork child keeps its signal.alarm guard. Finding 1 (generation recreate truly reuses the same VA): fully free the preallocation and stop, then set_allocation_region + preallocate at the same base in the same owner process. Assert the recreated source pointer and chunk base are equal and the generation differs. A first-generation mapping is deliberately kept open so that, were the generation dropped from the mapping key, the second open would alias that stale mapping and read the old bytes; with the generation in the key it maps fresh 0xB2. The retained mapping is never dereferenced (only closed at teardown). Finding 2 (cross-process stale exporter over socket + SCM_RIGHTS + token): replace the same-process stale test with a real exporter/importer protocol. The exporter allocates, exports (starting the abstract-socket server thread), publishes the serialized 64-byte handle blob, and stays alive. A fresh importer process opens that blob, fetching the fd across processes via SCM_RIGHTS (token verified server-side) and reads the sentinel. After the exporter frees and exits, a fresh importer using the same serialized handle fails promptly because the socket is gone. Finding 3 (duplicate open/close proves the VA is unreserved): after the second close of a refcounted interior mapping, reserve exactly the former whole-chunk range [local_base, local_base+chunk_size) with local_base as the hint (computed from the interior pointer and the handle's chunk base). A still-reserved range would relocate the reservation; requiring the hint be honored proves the VA is free, with no dereference of the freed range. Then a fresh open remaps and reads current data. These run under the Modal focused suite (the local VM has no CUDA device). Co-authored-by: Rahul Chalamala <rchalamala@users.noreply.github.com>

…free Address Task 1 review findings 1 (pure part), 2, and 4. Finding 1 (generation-sensitive key as a pure, testable contract): - Add include/foundry/vmm_ipc_chunk_key.h: a CUDA-free ChunkKey value type (exporter pid, token, chunk base, generation) with generation-inclusive equality/ordering/hash and a make_chunk_key factory. The hook now keys vmm_ipc_chunk_mappings on foundry::vmm_ipc::ChunkKey via make_chunk_key instead of an ad-hoc std::make_tuple, so the generation can never be dropped by accident. - Add tests/vmm_ipc_chunk_key_test.cpp (CPU-only, EXCLUDE_FROM_ALL + ctest): proves same pid/token/base + different generation are DISTINCT map/hash keys, and that a generation-less identity collapses two recreated chunks onto one key (the exact stale-mapping-reuse bug the field prevents). This replaces the fragile native same-VA discriminator with a hermetic contract; no unsafe production rewind/unset API is introduced. Finding 2 (fail closed on a failed fallback reservation free): - If freeing the driver's fallback reservation fails, release the imported handle exactly once, log both cleanup statuses, and return immediately with CUDA_ERROR_OUT_OF_MEMORY - never continue to the dedicated-zone mapping. The decision is factored into the pure require_fallback_reservation_freed helper (vmm_ipc_reservation.h) with cases 6-7 in the reservation policy test. No double release (the handle is released on the failure branch only, which returns before the zone path). Finding 4 (parked-fd limitation kept explicit and fail-fast): - Guard contract: the fork-child reject path closes the inherited listener fd, clears it, and does not access the inherited concurrent export-fd map. The limitation (inherited duplicate fds pin physical memory until the child exits/execs) stays documented; no complete child-FD reclamation is claimed. Source contracts in test_vmm_ipc_ownership_contract.py cover the pure key type, the fail-closed-before-zone ordering with both statuses logged, and the listener-close / no-map-access reject invariant. Co-authored-by: Rahul Chalamala <rchalamala@users.noreply.github.com>

…aits Address Task 1 review findings 1 (native part) and 3. Finding 1 (native complement, no impossible same-VA re-reservation): - Rewrite the generation-recreate native test to assert only OBSERVABLE behavior: the exported chunk generation increments across a full free + stop + recreate cycle (owner process stays), and a fresh import reads the recreated (fresh) bytes, never stale data. Drop the assertions that the recreated source pointer / chunk base are identical - the peer-import zone hint advances monotonically and is never rewound, so the local import VA legitimately differs; this is named and documented. The generation-sensitive key identity itself is proven hermetically by the pure CPU test tests/vmm_ipc_chunk_key_test.cpp. Finding 3 (every wait bounded, including post-kill cleanup): - Add _kill_and_collect: kills a process and collects its output under a bounded secondary wait; if the process will not terminate after the kill it raises with diagnostics rather than blocking forever. - Route _run_capture (now Popen + communicate(timeout)), _wait_for_file, the exporter signal-exit path, and the orchestrator teardown through it, so no post-kill communicate()/wait() can hang indefinitely. Secondary timeout is _KILL_TIMEOUT_S. These run under the Modal focused suite (the local VM has no CUDA device). Co-authored-by: Rahul Chalamala <rchalamala@users.noreply.github.com>

The two new C++ files briefly carried a stray hash-style SPDX header (the Python-only SPDX fixer was run against them by mistake) on top of the correct C++ // header. Remove the hash lines so the files carry a single, valid //-comment SPDX header and pass clang-format. Co-authored-by: Rahul Chalamala <rchalamala@users.noreply.github.com>

…eration Address the sole remaining Task 1 review finding for the native generation-recreate test. - Keep a SINGLE allocation region for both generations: set the region once and stop it once. Between generations only the preallocation is freed and recreated. Because the region cursor advances monotonically and is never rewound (cuMemFree_v2 / free_preallocated_region do not rewind it), generation 2's preallocation lands at a DIFFERENT chunk base within the same region. Assert the observable behavior: the chunk base differs (cb2 != cb1), the global generation counter increments (gen2 > gen1), and a fresh import reads the recreated bytes (0xB2), never stale generation-1 data. No stop/set between generations and no same-VA claim. - Add tests/test_generation_recreate_contract.py: a pure text source contract (no CUDA/torch/foundry import) that guards the invariant - set/stop each appear exactly once in the generation test, only the preallocation is recreated, and no identical-VA/chunk-base claim can reappear. The generation-sensitivity of the mapping key itself remains proven hermetically by tests/vmm_ipc_chunk_key_test.cpp; this native test complements it with a real device round trip. Runs under the Modal focused suite (the local VM has no CUDA device). Co-authored-by: Rahul Chalamala <rchalamala@users.noreply.github.com>

Address the sole remaining Task 1 review finding for the native generation-recreate test. After closing the importer and deleting each generation's tensor, synchronize and call torch.cuda.empty_cache() BEFORE free_preallocated_region so the caching allocator cannot retain a cached block pointing into that generation's soon-to-be-unmapped backing. empty_cache() releases blocks via cuMemFree_v2, which does not rewind the region cursor, so the monotonic cursor (and thus the distinct generation-2 chunk base) is preserved. Add a source contract asserting the del tensor -> empty_cache -> free_preallocated_region ordering in the generation loop. Co-authored-by: Rahul Chalamala <rchalamala@users.noreply.github.com>

cursoragent and others added 30 commits July 23, 2026 02:42

docs: design TP upstream consolidation

59fa558

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

docs: plan TP upstream consolidation

7f6662d

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

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

84d2b95

Signed-off-by: xenshinu <liuxs@umich.edu>

[sglang] add DeepEP no-fabric IPC policy

afa9d36

Co-Authored-By: Rahul Chalamala <22563365+rchalamala@users.noreply.github.com> Signed-off-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>

[sglang] add no-fabric DeepEP IPC recipe

403fed2

Co-Authored-By: Rahul Chalamala <22563365+rchalamala@users.noreply.github.com> Signed-off-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>

[sglang] fail clearly on old DeepEP API

b18b186

Co-Authored-By: Rahul Chalamala <22563365+rchalamala@users.noreply.github.com> Signed-off-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>

[sglang] move inspect import to module top

2328691

Co-Authored-By: Rahul Chalamala <22563365+rchalamala@users.noreply.github.com> Signed-off-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>

[sglang] enforce DeepEP use_fabric requirement at Buffer construction

05df555

Co-Authored-By: Rahul Chalamala <22563365+rchalamala@users.noreply.github.com> Signed-off-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>

[sglang] add experimental dense tensor-parallel recipe

b681569

Co-Authored-By: Rahul Chalamala <22563365+rchalamala@users.noreply.github.com> Signed-off-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>

[fix] fix workspace mismatch when using fa2

310cb49

Signed-off-by: xenshinu <liuxs@umich.edu>

test: add reproducible SGLang TP validation

89f43fa

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

fix: keep TP NCCL transport symmetric

0ccb9e9

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

fix: pin TP validation to immutable revisions

98df069

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

fix: align TP graph mode across phases

f23e245

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

docs: mark SGLang tensor parallel validated

8da5295

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

test: strengthen tensor parallel verification

af40052

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

fix: harden graph and IPC restoration

eaf3fea

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

fix: version preallocated IPC chunks

ab22a21

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

test: require complete TP replay evidence

a7a2dd2

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

fix: pass default graph pool explicitly

57e1dba

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

test: compare semantic TP graph archives

fb86222

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

docs: make TP validation invocation portable

c85cd3b

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

test: unpack loaded graph results

335d135

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

fix: pin TP validation seed

6d98241

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

docs: explain deterministic TP validation

c109bb2

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

test: compare deterministic TP graph structure

b89d114

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

docs: align TP scope with upstream roadmap

4480305

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

fix: preserve Python CUDAGraph public API

7c9f355

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

test: require exact RNG replay restoration

eee0131

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

fix: make experimental recipe defaults portable

eb7a0d2

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

cursoragent and others added 11 commits July 23, 2026 16:49

fix(sglang): preserve focused startup causes

05442fb

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

fix(sglang): classify nested startup exceptions

5a3a35a

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

fix(sglang): close traceback at log boundaries

3c34442

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

fix(sglang): recognize CR-prefixed markers

02923d1

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

fix(sglang): verify archived symmetric collectives

67d18cf

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

fix(sglang): make cleanup part of verification

6a61b4b

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

fix(sglang): fail closed on cleanup errors

733f766

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

test(sglang): require symmetric evidence in full TP matrix

94a4d00

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

fix(sglang): stabilize concurrent TP oracle

7302233

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

cursor Bot mentioned this pull request Jul 23, 2026

Consolidate TP support and validate vLLM and SGLang #5

Draft

cursoragent and others added 18 commits July 23, 2026 21:20

fix(sglang): force native concurrent batch

7304e81

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

fix(sglang): guarantee native scheduler batching

176e1fe

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

test(sglang): align native batch decode lifetimes

c2b10d6

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

docs(sglang): record verified symmetric TP evidence

0437586

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

docs(sglang): address symmetric TP review findings

c127642

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

docs(sglang): clarify LOAD metadata planning

c237554

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

docs: plan final branch deltas

cfec7f2

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

fix: harden VMM IPC process and reservation ownership

5609bad

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!

[tp] Consolidate branches and validate vLLM/SGLang#6

[tp] Consolidate branches and validate vLLM/SGLang#6
rchalamala wants to merge 125 commits into
mainfrom
cursor/tp-upstream-consolidation-fe33

rchalamala commented Jul 23, 2026 •

edited by cursor Bot

Loading

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 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

SGLang main — VERIFIED

Remaining before final handoff

Supported scope

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

rchalamala commented Jul 23, 2026 •

edited by cursor Bot

Loading