Repository navigation
Conversation
This was referenced Oct 21, 2026
Contributor
✅ Dynamo PR CI passed — run 41409732522 (attempt 1) on
|
Aphoh
force-pushed
the
warnold/sglang-generate-migration
branch
3 times, most recently
from
October 21, 2026 06:49
14d65fb to
1b66ee6
Compare
Aphoh
force-pushed
the
warnold/sglang-generate-migration
branch
from
October 21, 2026 08:46
1b66ee6 to
7621e81
Compare
Aphoh
force-pushed
the
warnold/sglang-generate-migration
branch
from
October 22, 2026 03:01
7621e81 to
5e0d8ee
Compare
Aphoh
added a commit
that referenced
this pull request
Oct 22, 2026
Unify native SGLang and vLLM generate routes through backend-owned format plugins and the shared HTTP lifecycle. Include SGLang migration, the external Base64+zstd codec example, generic agent tracing, and the one-token P/D expert preservation fix. Squash the complete #16983 -> #16977 -> #16978 -> #16997 stack without changing its tested source tree. Original stack head: c6c1c80. Refs: #16675, #16674, #16640 Fixes: #16762 Co-authored-by: gongshaotian <gstian5555@outlook.com> Signed-off-by: William Arnold <7565007+Aphoh@users.noreply.github.com>
Aphoh
force-pushed
the
warnold/sglang-generate-migration
branch
from
October 22, 2026 04:51
5e0d8ee to
780fdff
Compare
Aphoh
added a commit
that referenced
this pull request
Oct 22, 2026
Unify native SGLang and vLLM generate routes through backend-owned format plugins and the shared HTTP lifecycle. Include SGLang migration, the external Base64+zstd codec example, generic agent tracing, and the one-token P/D expert preservation fix. Reuse SSE serialization buffers and move consumed prompt snapshots to remove measured frontend CPU overhead. Squash the complete #16983 -> #16977 -> #16978 -> #16997 stack without changing its tested source tree. Stack head: d382dd1. Refs: #16675, #16674, #16640 Fixes: #16762 Co-authored-by: gongshaotian <gstian5555@outlook.com> Signed-off-by: William Arnold <7565007+Aphoh@users.noreply.github.com>
Aphoh
added a commit
that referenced
this pull request
Oct 22, 2026
Unify native SGLang and vLLM generate routes through backend-owned format plugins and the shared HTTP lifecycle. Include SGLang migration, the external Base64+zstd codec example, generic agent tracing, and the one-token P/D expert preservation fix. Reuse SSE serialization buffers and move consumed prompt snapshots to remove measured frontend CPU overhead. Preserve one-token vLLM prompt logprobs and fill missing native token labels. Squash the complete #16983 -> #16977 -> #16978 -> #16997 stack without changing its tested source tree. Stack head: 5e09cf5. Refs: #16675, #16674, #16640 Fixes: #16762, #17088 Co-authored-by: gongshaotian <gstian5555@outlook.com> Signed-off-by: William Arnold <7565007+Aphoh@users.noreply.github.com>
Aphoh
force-pushed
the
warnold/sglang-generate-migration
branch
2 times, most recently
from
October 22, 2026 09:23
3869390 to
154e998
Compare
Aphoh
added a commit
that referenced
this pull request
Oct 22, 2026
Unify native SGLang and vLLM generate routes through backend-owned format plugins and the shared HTTP lifecycle. Include SGLang migration, the external Base64+zstd codec example, generic agent tracing, and the one-token P/D expert preservation fix. Reuse SSE serialization buffers and move consumed prompt snapshots to remove measured frontend CPU overhead. Preserve one-token vLLM prompt logprobs and fill missing native token labels. Squash the complete #16983 -> #16977 -> #16978 -> #16997 stack without changing its tested source tree. Stack head: c541550. Remove duplicate vLLM sampling declarations and response-finalization plumbing. Refs: #16675, #16674, #16640 Fixes: #16762, #17088 Co-authored-by: gongshaotian <gstian5555@outlook.com> Signed-off-by: William Arnold <7565007+Aphoh@users.noreply.github.com>
Aphoh
force-pushed
the
warnold/sglang-generate-migration
branch
from
October 22, 2026 10:07
154e998 to
f62c303
Compare
Aphoh
added a commit
that referenced
this pull request
Oct 22, 2026
Unify native SGLang and vLLM generate routes through backend-owned format plugins and the shared HTTP lifecycle. Include SGLang migration, the external Base64+zstd codec example, generic agent tracing, and the one-token P/D expert preservation fix. Reuse SSE serialization buffers and move consumed prompt snapshots to remove measured frontend CPU overhead. Preserve one-token vLLM prompt logprobs and fill missing native token labels. Squash the complete #16983 -> #16977 -> #16978 -> #16997 stack without changing its tested source tree. Stack head: b2de7a8. Remove duplicate vLLM sampling declarations and response-finalization plumbing. Refs: #16675, #16674, #16640 Fixes: #16762, #17088 Remove RL-only sidecar rolling-upgrade shims; use current plugin projections and transparent sampling serialization. CPU validation: 338 tests, strict Clippy and pre-commit. Co-authored-by: gongshaotian <gstian5555@outlook.com> Signed-off-by: William Arnold <7565007+Aphoh@users.noreply.github.com>
Aphoh
added a commit
that referenced
this pull request
Oct 23, 2026
Unify native SGLang and vLLM generate routes through backend-owned plugins and the shared HTTP lifecycle. Include SGLang migration, the external Base64+zstd codec example, agent tracing, and vLLM prompt-logprob fixes. Remove duplicate sampling/finalization code and RL-only rolling-upgrade shims. Include the review fixes for route-test isolation and multimodal fallback fixtures. Preserve upstream live request/token metrics and SGLang HTTP connection retirement while removing the superseded generate handlers. Single-commit snapshot of #16983 -> #16977 -> #16978 -> #16997. Stack head: 1d59b49. Upstream parent: fc9b996. Refs: #16675, #16674, #16640 Fixes: #16762, #17088 Co-authored-by: gongshaotian <gstian5555@outlook.com> Signed-off-by: William Arnold <7565007+Aphoh@users.noreply.github.com>
Aphoh
force-pushed
the
warnold/sglang-generate-migration
branch
2 times, most recently
from
October 23, 2026 07:48
4cb9de3 to
870a83a
Compare
Aphoh
force-pushed
the
warnold/generate-format-plugins-base
branch
from
October 26, 2026 08:26
07c4505 to
e15018c
Compare
Aphoh
force-pushed
the
warnold/sglang-generate-migration
branch
from
October 26, 2026 08:29
870a83a to
e4a0219
Compare
Aphoh
force-pushed
the
warnold/generate-format-plugins-base
branch
from
October 28, 2026 06:34
e15018c to
9dd347e
Compare
Aphoh
force-pushed
the
warnold/sglang-generate-migration
branch
from
October 28, 2026 06:36
f7cd044 to
13d540f
Compare
Aphoh
force-pushed
the
warnold/generate-format-plugins-base
branch
from
October 29, 2026 04:31
9dd347e to
e721abe
Compare
Aphoh
force-pushed
the
warnold/sglang-generate-migration
branch
from
October 29, 2026 04:31
13d540f to
f046df7
Compare
Aphoh
force-pushed
the
warnold/generate-format-plugins-base
branch
from
October 30, 2026 04:18
e721abe to
e5b66e4
Compare
Aphoh
force-pushed
the
warnold/sglang-generate-migration
branch
from
October 30, 2026 04:41
f046df7 to
d2f3969
Compare
Contributor
Author
|
/ok to test d2f3969 |
Aphoh
force-pushed
the
warnold/sglang-generate-migration
branch
from
November 1, 2026 03:45
d2f3969 to
3c596fd
Compare
Contributor
Author
|
/ok to test 3c596fd |
Signed-off-by: William Arnold <7565007+Aphoh@users.noreply.github.com>
Retain migration snapshots after rendering the first aggregated attempt, moving disjoint fields out of the discarded chunk. Keep copy semantics for overlapping custom fields, P/D, retries, and unary assembly. Add a property law for nested custom snapshots. No plugin API changes. Refs: #16675, #16674 Signed-off-by: William Arnold <7565007+Aphoh@users.noreply.github.com>
Keep the vLLM projection fixture aligned with the Worker field added by this PR. The base PR remains independently buildable. Refs #16675. Signed-off-by: William Arnold <7565007+Aphoh@users.noreply.github.com>
Signed-off-by: William Arnold <warnold@nvidia.com> Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
System One binds through the SGLang format with no engine info and never migrates a scoring branch; the crate's test format implements migrate. Signed-off-by: William Arnold <warnold@nvidia.com> Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
Aphoh
force-pushed
the
warnold/sglang-generate-migration
branch
from
November 1, 2026 04:34
3c596fd to
488723f
Compare
Contributor
Author
|
/ok to test 488723f |
…enerate-migration Signed-off-by: William Arnold <warnold@nvidia.com> Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
Contributor
Author
|
/ok to test 14c8b2e |
1 task done
This branch has not been deployed
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
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.
Stacked on #17802, which adds the backend-neutral migration hooks; this PR holds only SGLang migration. It does not depend on #16978, #16997, or vLLM migration (#17766).
Summary
Enable migration for eligible native SGLang
/generaterequests through the plugin interface introduced by #16983. A format-owned continuation rebuilds the opaque engine request, reduces remaining token budgets, preserves prompt metadata, and rebases response counters across failed attempts.Worker-advertised sampling defaults participate in admission. Stateful sampling, unresolved token budgets, unbounded budgets, and unsupported response modes remain non-migratable. Retries respect the configured sequence limit; exhausted budgets finish without another worker call. vLLM
/generatemigration remains disabled.The property harness adds injected failures, repeated retries, migration-equivalence and admission laws, and an independent saturation oracle. HTTP integration tests exercise unary/SSE retries, sequence limits, and logical-request cache metrics.
Part of #16675. Part of #16674.
Validation
Incremental diff: 22 files, +1,344 / -163. No build-manifest or ownership-policy changes.
Live complete-stack qualification also passed on B200 with the rebuilt frontend wheel and Rust sidecar, pinned SGLang v0.5.21/cu130, and
PrimeIntellect/qwen3-moe-tiny. Two campaigns killed one of two token-only sidecars after eight streams started: 16/16 streams completed, eight actually migrated from the killed worker to the survivor. Captured engine outputs verify exact token order, prompt/output logprobs (up to two ULPs for JSON round trips), routed-expert bytes, replay prefixes, remaining budgets, and exactly-once trace rows. All migrated requests returned 256 tokens. The tracing follow-up #16997 fixes a per-attempt counter race found by this campaign.Migration uses
--skip-tokenizer-init, matching its existing token-only admission rule. This is correctness qualification; no throughput or latency claim.Frontend performance follow-up
Move disjoint prompt snapshots out of consumed first-attempt streaming chunks after rendering. Keep copy semantics for overlapping custom fields, P/D, retries, and unary assembly. No plugin API changes. The 512-case snapshot law, native merge/migration laws, HTTP tests, strict Clippy, and changed-file hooks pass.
Combined with #16983's reusable SSE buffer, the complete-stack logprob-heavy benchmark drops from 2.55 to 1.98 ms of frontend CPU/request (about 5% above the original baseline, down from 34%). This is a four-core release frontend with fixed mock workers, not an isolated measurement of this child PR or a GPU benchmark. Residual low-load short-request CPU cost remains (+61 µs/request, or 14%).