Skip to content

[ROCm][CI] Make /ci rocm cache a cache-only CI build that also copies the runtime - #66032

Merged
AndreasKaratzas merged 6 commits into
vllm-project:mainfrom
AndreasKaratzas:akaratza_rocm_cache_on_ci
Nov 16, 2026
Merged

AndreasKaratzas merged 6 commits into
vllm-project:mainfrom
AndreasKaratzas:akaratza_rocm_cache_on_ci

Conversation

@AndreasKaratzas

@AndreasKaratzas AndreasKaratzas commented Nov 6, 2026 •

Copy link
Copy Markdown
Member

Overview

/ci rocm cache now runs as a CI build that only copies images, and it copies the PR's runtime and test-deps images along with the stack, so main stops rebuilding them after the merge. Rebased onto current main; supersedes this PR's earlier version.

Claims

  • The cache build is a main build of the ci pipeline that runs only the new rocm-base-image-cache step (VLLM_CI_ONLY_STEP_KEYS), instead of an AMD CI build. No ci-infra change is needed; ci-infra's AMD CI promote mode (ci-infra#886) can be removed after this merges.
  • build.sh promote copies the PR's runtime and test-deps images too, so main reuses them when it has the same runtime key after the merge. After [ROCm] Bump AITER to v0.1.28.post1 #65937, main's first build spent 31 min in "Ensure ROCm stack + runtime" (ci build 100464) rebuilding a runtime and test-deps that the PR's CI had already built under those keys.
  • Only images the PR's CI build pushed on top of a parent it also pushed are recorded and copied, so a retried job or another PR overwriting this PR's tags can't pair the copied stack with a runtime built on a different one.

Validation

Check Result
pytest tests/tools/test_rocm_build.py 53 passed; the new recording, copy-chain and no-request tests fail on main's build.sh
cd .github/workflows/scripts && pytest test_run_ci_command.py 74 passed, 117 subtests
ci-infra's pipeline_generator run on this branch, main build with VLLM_CI_ONLY_STEP_KEYS=["rocm-base-image-cache"] exactly one step, queue amd-cpu, ROCM_CI_REPO: rocm/vllm-ci, no image builds
same, nightly (NIGHTLY=1) step included; promote exits 0 without ROCM_PROMOTE_PR
same, normal main build step only offered under "Run more jobs"
pre-commit on the changed files clean

Not run end to end: an actual cache build on the ci pipeline (it needs a PR that changes the ROCm stack and a maintainer's /ci rocm cache).

Details

  • build.sh base mode records rocm-rocm-{runtime,test-deps}-built like rocm-rocm-stack-built, but only when the same job also pushed the parent (a BUILT_HERE file, since each ensure runs in a subshell); building a stack resets both to none, because metadata outlives a retried job. promote copies stack-wheels, stack, runtime, test-deps in order, each child only over a parent this run copied, never over an image main built; a missing child record stops the chain with exit 0.
  • The step lives in .buildkite/image_build/amd_rocm_cache.yaml. ci-infra's AMD CI pipeline renders every file in hardware_tests/, but only the main pipeline lists image_build/. It is optional. Nightlies still run optional steps, so without a request promote exits 0 right after its arguments, before the builder, build context or tag fetch. It also validates the PR number and commit format that ci-infra's bootstrap-amd.sh used to check.
  • The handler still looks for the source build in both pipelines, since PRs still run AMD CI ([ROCm] AITER bump to v0.1.29 #66596's cache came from an AMD CI build).
  • Limits: if main changes a runtime-key input (requirements, Dockerfile.rocm above the marker) between the PR's CI build and the merge, the copied runtime is unused and main rebuilds it as before ([ROCm] AITER bump to v0.1.29 #66596: 46 min). A retried promote finishes the chain: an existing tag holding the digest the PR's CI recorded is an earlier attempt's copy, and only a different digest is treated as main's.

AI assistance (Claude Code) was used to investigate, implement and validate this change.


Pull Request Checklist
  • I used vLLM's /pr-checklist skill. (Mandatory for agents, optional for humans).

  • AI assistance was used during the creation of this PR.

  • Design Fit: Minimizes impact on core components, reuses existing functionality, and justifies added complexity.

  • Testing and Validation: Validates the change and ensures any added tests are meaningful and reliable, with CI coverage or documented CI resource constraints and validation performed outside CI.

  • Code Quality and Style: Keeps code and comments clear and concise, and updates relevant documentation and examples.

  • Pull Request Contents: Includes a brief summary and relevant links, supports claims with evidence, explains root causes and implementation trade-offs, and follows the contributing guide.

BEFORE SUBMITTING, PLEASE READ https://docs.vllm.ai/en/latest/contributing (anything written below this line will be removed by GitHub Actions)

@mergify

mergify Bot commented Nov 7, 2026

Copy link
Copy Markdown
Contributor

This pull request has merge conflicts that must be resolved before it can be
merged. Please rebase the PR, @AndreasKaratzas.

https://docs.github.com/en/pull-requests/collaborating-with-pull-requests/working-with-forks/syncing-a-fork

@mergify mergify Bot added the needs-rebase label Nov 7, 2026
… the runtime

/ci rocm cache copied only the stack pair, so main's first build after the
merge still rebuilt the runtime and test-deps images from source: 31 min
after vllm-project#65937, whose CI had already built both under the same keys. It
also ran as an AMD CI build.

build.sh base mode now records the runtime and test-deps it pushed itself
(rocm-rocm-{runtime,test-deps}-built), as it already did for the stack, but
only when the same job also pushed the parent it was built on: metadata
outlives a retried job, and another PR can replace this PR's stack tag
between attempts. A new stack resets them to none. promote copies them
under the runtime key of the PR's commit, each only over the parent the
same run copied, never over an image main built; main uses them when it
still has that key after the merge.

The cache build is now a CI build of main that runs only the new
rocm-base-image-cache step (VLLM_CI_ONLY_STEP_KEYS), in its own file
because ci-infra's AMD CI pipeline renders hardware_tests/amd.yaml. The
step is optional; nightlies run optional steps, so promote exits 0
without a request. It also checks the PR number and commit format, which
ci-infra's AMD CI bootstrap did.

Co-Authored-By: Claude <noreply@anthropic.com>
Signed-off-by: Andreas Karatzas <akaratza@amd.com>
@AndreasKaratzas AndreasKaratzas changed the title [ROCm][CI] Run /ci rocm cache on the ci pipeline instead of AMD CI [ROCm][CI] Make /ci rocm cache a cache-only CI build that also copies the runtime Nov 11, 2026
@AndreasKaratzas AndreasKaratzas changed the title [ROCm][CI] Make /ci rocm cache a cache-only CI build that also copies the runtime [ROCm][CI] Run /ci rocm cache on the ci pipeline instead of AMD CI Nov 11, 2026
@AndreasKaratzas
AndreasKaratzas force-pushed the akaratza_rocm_cache_on_ci branch from 89af346 to 4adf949 Compare November 11, 2026 04:07
@AndreasKaratzas AndreasKaratzas changed the title [ROCm][CI] Run /ci rocm cache on the ci pipeline instead of AMD CI [ROCm][CI] Make /ci rocm cache a cache-only CI build that also copies the runtime Nov 11, 2026
@mergify

mergify Bot commented Nov 11, 2026

Copy link
Copy Markdown
Contributor

Documentation preview: https://vllm--66032.org.readthedocs.build/en/66032/

@mergify mergify Bot added documentation Improvements or additions to documentation and removed needs-rebase labels Nov 11, 2026
@AndreasKaratzas
AndreasKaratzas marked this pull request as ready for review November 11, 2026 07:19
@AndreasKaratzas
AndreasKaratzas requested a review from a team November 11, 2026 07:19

@claude claude Bot left a comment

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Claude Code Review

This pull request is from a fork — automated review is disabled. A repository maintainer can comment @claude review to run a one-time review.

@AndreasKaratzas

Copy link
Copy Markdown
Member Author

/ci run

@github-actions

Copy link
Copy Markdown

✅ Triggered Buildkite CI #101715 for commit dc092b3ce519.

ℹ️ This PR is 29 commits behind upstream main. Merged upstream main into this branch as dc092b3ce519 before starting CI. Pull the branch before pushing more changes.

@AndreasKaratzas

Copy link
Copy Markdown
Member Author

/ci retry

@github-actions

Copy link
Copy Markdown

✅ Triggered Buildkite CI #101890 for commit 7ad9a8e9cfb3, running 2 failed step(s) from Buildkite CI #101715.

@divakar-amd divakar-amd left a comment

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Looking good overall. The chain of base->runtime-> test-deps is reasonable. Tested locally with retries, passes. Added a few comments for you to look at. The main concern being copied=0 one and the "nothing to promote" guard sitting at a later stage.

Comment thread .buildkite/scripts/rocm/build.sh Outdated
# ROCM_PROMOTE_SHA pushed. Copy them here under the keys main's keys_at gives that commit, if
# it was built with main's recipe; main reuses each once it gets that key. The runtime and
# test-deps are copied only over the parent this run copied, never over one main built.
[[ -n ${ROCM_PROMOTE_PR:-} ]] || { echo "No /ci rocm cache request: nothing to promote" >&2; exit 0; }

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Could this guard move to the top of the script?
ci_config_rocm.yaml lists .buildkite/hardware_tests as a job_dir, so this step renders in vllm_rocm_ci too, and nightlies run optional steps. By the time we reach here we've already created a buildx builder, run git archive HEAD into .rocm-src, and done git fetch --force --tags (possibly --unshallow) - all to print "nothing to promote", on an amd_cpu agent, every nightly.

Copy link
Copy Markdown
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Done in e2a98ee. The three request checks (no request, malformed, wrong repo) now run right after the arguments, before the builder, git archive and the tag fetch, so a nightly without a request exits at once. promote never bakes, so with a request it also skips the context export and tag fetch. test_promote_without_a_request_does_nothing now asserts that no builder, context or registry call happens.

timeout_in_minutes: 30
retry:
automatic:
- exit_status: 75 # Registry error (build.sh)

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This could be a misleady green signal. Let's fix this or drop the 75 retry.
Given the copied reset in build.sh, this retry can't actually recover a partial copy - it re-runs into the break and goes green.

@divakar-amd divakar-amd Nov 12, 2026 •

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Taking some of it back... retry does earn its place, so don't simply drop it.
The narrower point still stands though: a failure mid-chain retries into the break and
goes green having not finished. That's the misleading-green bit.

Copy link
Copy Markdown
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Agreed, kept the retry and fixed the chain instead (e2a98ee, see the thread on build.sh). A retry after a mid-chain registry error now finishes the runtime and test-deps rather than going green.

Comment thread .buildkite/scripts/rocm/build.sh Outdated
"runtime-$STACK-$pKr" "test-deps-$STACK-$pKr"; do # the stack tag completes the pair
stacked=0; [[ $tag == stack-* ]] && stacked=1
if digest "$TRUSTED:$tag" >/dev/null; then
echo "$TRUSTED:$tag already exists" >&2; copied=0; continue

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

the copying here is a concern. Please check the detailed comments below:

copied=0 fires on any "already exists" — including tags this job's own earlier attempt copied. So if attempt 1 copies the two stack tags and the runtime copy then hits a registry error (exit 75), the retry finds both stack tags present, resets copied, and breaks before the runtime. Green build,
nothing copied, main still rebuilds. I reproduced it against this script: STATUS 0, COPIED []. Same for the agent-lost retries, the 30-min timeout, and a re-issued /ci rocm cache after a cancelled build.

Worth distinguishing "exists because main built it" from "exists because we copied it last attempt" — the digest in ROCM_PROMOTE_IMAGES tells you which.

One trap if you do: digest() ends in exit 75, so calling it as ed=$(digest ...) puts it in a subshell and a registry error silently becomes "tag absent" — we'd overwrite main's image. ensure()'s idiom handles it: rc=$?; ((rc == 1)) || exit $rc.

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I see you've already called this out in the limits section of the PR ("A promote retried after copying the stack skips the runtime (safe, main builds it)").

Reason I'd still like it looked at: the exit-75 retry on the cache step makes that path routine rather than rare, and ROCM_PROMOTE_IMAGES already carries the digest that tells you whether an existing tag is yours or main's, so the distinction looks cheap. Happy for it to be a follow-up if you'd rather.

Copy link
Copy Markdown
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Fixed in e2a98ee, as you suggested. An existing tag whose digest matches the one ROCM_PROMOTE_IMAGES records for it is an earlier attempt's copy, so the chain continues. A different digest is main's, so it stops as before. On the trap: the lookup is if d=$(digest ...); then ...; else rc=$?; ((rc == 1)) || exit $rc; fi, so a registry error stays exit 75 and never reads as absent. Tests: test_promote_finishes_a_chain_an_earlier_attempt_left (your scenario; it fails on the previous build.sh) and test_promote_a_registry_error_on_an_existing_tag_fails_the_step. The fake registry now keeps per-ref digests.

Comment on lines +4 to +6
# copies the ROCm images a PR's CI built into rocm/vllm-ci. Not in hardware_tests/amd.yaml, which
# ci-infra's AMD CI pipeline also renders. Optional, so other builds skip it; nightlies run optional
# steps, and without a request it does nothing.

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I'm not sure on this part but could you double check w.r.t ci-infra pipeline?
ci_config_rocm.yaml (vllm_rocm_ci) declares the whole "hardware_tests" directory, so this might get rendered?

Copy link
Copy Markdown
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

You were right: ci-infra's read_steps_from_job_dir (step.py) walks every .yaml in each job dir, and ci_config_rocm.yaml lists all of hardware_tests/, so vllm_rocm_ci rendered it. e2a98ee moves it to .buildkite/image_build/, which only ci_config.yaml lists. ci-infra's image and always-run handling match on step keys, not on the directory, and /ci rocm cache selects the step by key, so nothing else changes. With the early guard it would have been harmless there anyway.

…a request

- An existing tag holding the digest #ROCM_PROMOTE_PR's CI recorded is an
  earlier attempt's copy, so a retried promote goes on to the runtime and
  test-deps instead of stopping there and going green. A different digest
  is main's, as before. The lookup keeps a registry error an error (exit 75).
- Check the promote request right after the arguments, before the builder,
  the build context and the tag fetch: nightlies run the step without one.
  promote never bakes, so it skips the context export and tag fetch.
- Move the step to image_build/: ci-infra's AMD pipeline renders all of
  hardware_tests/, image_build/ only the main pipeline.

Co-authored-by: Claude <noreply@anthropic.com>
Signed-off-by: Andreas Karatzas <akaratza@amd.com>
@AndreasKaratzas

Copy link
Copy Markdown
Member Author

/ci run

@github-actions

Copy link
Copy Markdown

✅ Triggered Buildkite CI #102410 for commit 08717c3a45b4.

ℹ️ This PR is 156 commits behind upstream main. Merged upstream main into this branch as 08717c3a45b4 before starting CI. Pull the branch before pushing more changes.

@AndreasKaratzas

Copy link
Copy Markdown
Member Author

/ci retry

@github-actions

Copy link
Copy Markdown

✅ Triggered Buildkite CI #102651 for commit 3c1bd6e1538d, running 3 failed step(s) from Buildkite CI #102410.

@divakar-amd divakar-amd left a comment

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Thanks for addressing the suggestions above. Re-checked the PR with the latest commit. Looks good. Approving.

Comment on lines 114 to +133
@@ -120,6 +130,7 @@ git describe --tags --match 'v[0-9]*' >/dev/null 2>&1 \
printf 'node: %s\nnode-date: %s\ndescribe-name: %s\n' "$(git rev-parse HEAD)" \
"$(git log -1 --format=%cI)" "$(git describe --tags --long --abbrev=40 --match 'v[0-9]*')" \
> .rocm-src/.git_archival.txt
fi

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

NIT: (Cosmetic change) missing indentation under the if-else block

@AndreasKaratzas
AndreasKaratzas merged commit ce88d40 into vllm-project:main Nov 16, 2026
25 checks passed
@AndreasKaratzas
AndreasKaratzas deleted the akaratza_rocm_cache_on_ci branch November 16, 2026 18:25
Csrayz pushed a commit to Csrayz/vllm that referenced this pull request Nov 19, 2026
)

* [Security] Fall back to eager when encoder CUDA graph replay exceeds capture (#60112)

Signed-off-by: Juan Pérez de Algaba <jperezde@redhat.com>

* [CPU][Perf] Split oversized final attention spans (#66855)

Signed-off-by: Li, Tianmu <tianmu.li@intel.com>
Co-authored-by: OpenAI Codex <noreply@openai.com>
Co-authored-by: vllm-agent <claw@inferact.ai>

* [CI][Bugfix] Pin the GLM5-next shared-expert fusion test to the stock decode (#67369)

Signed-off-by: Lucas Wilkinson <lwilkins@redhat.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: vllm-agent <claw@inferact.ai>

* [Rust Frontend][gRPC] Add the Render service with RenderChat (#65994)

Signed-off-by: Bugen Zhao <i@bugenzhao.com>
Signed-off-by: Dao Le <Dao007forever@gmail.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: Bugen Zhao <i@bugenzhao.com>

* [Feature] add the option to continue serving unfinished requests after the recovery (#50265)

Signed-off-by: TianZhuo <2770730562@qq.com>
Signed-off-by: fangyuchu <fangyuchu@qq.com>
Signed-off-by: Tyler Michael Smith <tlrmchlsmth@gmail.com>
Co-authored-by: tianzhuo <2770730562@qq.com>
Co-authored-by: Tyler Michael Smith <tlrmchlsmth@gmail.com>
Co-authored-by: Nicolò Lucchesi <nicolo.lucchesi@mistral.ai>

* [Test][Model Loader] Leave reload headroom in the LoRA-MoE and online-quantize reload tests (#67210)

Signed-off-by: swappy <59965507+rycerzes@users.noreply.github.com>

* [ROCm][Perf][GLM-5.3-Flash] Keep the fused gated RMSNorm for the KDA o_norm (#66219)

Signed-off-by: Zheng Gong <zgong@amd.com>
Co-authored-by: vllm-agent <claw@inferact.ai>

* [Spec Decode] Keep dynamic speculative decoding for dense models with DP (#67005)

Signed-off-by: Lantao Jin <ltjin@amazon.com>
Co-authored-by: Claude <noreply@anthropic.com>

* [CI] Let ci-selector decide CMakeLists.txt changes on CUDA (#65881)

Signed-off-by: khluu <khluu000@gmail.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>

* [Misc] Fix override and context manager type annotations (#67295)

Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>

* [Rust Frontend] Keep the mock engine alive in the test_app fixture (#67409)

Signed-off-by: Bugen Zhao <i@bugenzhao.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: vllm-agent <claw@inferact.ai>

* [CI] Retry the CUDA image bake when the registry rate-limits the push (#67437)

Signed-off-by: Kevin Luu <51931015+khluu@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>

* [Bugfix][Kernel] Bump DeepGEMM for the MegaMoE combine-readiness fix under CUDA graph replay (#67181)

Signed-off-by: Thien Tran <gau.nernst@yahoo.com.sg>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: vllm-agent <claw@inferact.ai>

* [Perf][Kernel][Quantization] Tensor-descriptor path for int4 fused_moe_kernel_gptq_awq (XPU default, sm100 opt-in) (#52652)

Signed-off-by: Artur Fierka <artur.fierka@intel.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: Kunshang Ji <kunshang.ji@intel.com>

* [Bugfix][Spec Decode] Allow min_p and logit_bias with speculative decoding (#67061)

Signed-off-by: Anh Tran <anhtra@nvidia.com>

* [CI][Bugfix] Migrate UMBP lazy offload tests to the ManagerConfig KV cache API (#67420)

Signed-off-by: PerryZhang01 <Perry.Zhang@amd.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>

* [Perf][GLM-5.3] build eager attention metadata on the host for ViT (#67194)

Signed-off-by: Canlin <canlinguosdu@gmail.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: vllm-agent <claw@inferact.ai>

* [Bugfix][Benchmark] Fix prefix_repetition crash and custom_audio --output-len in bench throughput (#67288)

Signed-off-by: Aditya Ganesh Kumar <adty910@gmail.com>

* [Bugfix][Multimodal] Preserve values from non-native-endian NumPy embeddings (#67479)

Signed-off-by: Afloat <ShuffleMonkey@163.com>

* [ROCm][Docs] Add DeepSeek-V4.1-Flash MI355X DP2 AgentX preset with UMBP KV offload (#67378)

Signed-off-by: shen-shanshan <467638484@qq.com>
Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

* [Model] Add Mistral Large 4 (#64534)

Signed-off-by: NickLucche <nicolo.lucchesi@mistral.ai>
Co-authored-by: Lucas Wilkinson <lwilkins@redhat.com>
Co-authored-by: Matthew Bonanni <mbonanni@redhat.com>
Co-authored-by: Felix Marty <Felix.Marty@amd.com>
Co-authored-by: Hongxia Yang <hongxia.yang@amd.com>
Co-authored-by: Andy Luo <andy.luo@amd.com>
Co-authored-by: Julien Denize <julien.denize@mistral.ai>
Co-authored-by: Dimitrios Bariamis <12195802+dbari@users.noreply.github.com>
Co-authored-by: Julien Debache <jdebache@nvidia.com>
Co-authored-by: vanshil shah <49670736+vanshilshah97@users.noreply.github.com>

* [Rust Frontend] Render tool-result media in the Inkling renderer (#67398)

Signed-off-by: Bugen Zhao <i@bugenzhao.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>

* [Model] Use FlashInfer for Gemma 4 with an NVFP4 KV cache when FA4 is unavailable (#47547)

Signed-off-by: lesj0610 <lesj0610@users.noreply.github.com>
Signed-off-by: lesj0610 <lesj0610@gmail.com>
Co-authored-by: lesj0610 <lesj0610@users.noreply.github.com>
Co-authored-by: vllm-agent <claw@inferact.ai>
Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>

* [Spec Decode] Consistent drafted, verified and accepted token counts (#64156)

Signed-off-by: Yiliu Dong <91178480+qianlihuang@users.noreply.github.com>
Signed-off-by: Tomas Ruiz <tomas.ruiz.te@gmail.com>
Co-authored-by: Yiliu Dong <91178480+qianlihuang@users.noreply.github.com>
Co-authored-by: GLM <noreply@z.ai>
Co-authored-by: Codex <noreply@openai.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>

* [CI][Bugfix] Read sliding_window from ManagerConfig.params in voxtral realtime test (#67476)

Signed-off-by: khluu <khluu000@gmail.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>

* [Bugfix][Pooling] Honor the CrossEncoder activation saved by Sentence Transformers v6 (#62785)

Signed-off-by: Cy <15118312+cycancode@users.noreply.github.com>
Co-authored-by: Taneem Ibrahim <taneem.ibrahim@gmail.com>

* [Quantization] Make activation-quantization fallbacks explicit, add VLLM_STRICT_QUANT_SCHEME (#59460)

Signed-off-by: mgoin <mgoin64@gmail.com>
Co-authored-by: Claude <noreply@anthropic.com>
Co-authored-by: vllm-agent <claw@inferact.ai>

* [Model] Load q/k/v scales for the quantized Kimi-K2.6 (#63097)

Signed-off-by: yiliu30 <yi4.liu@intel.com>
Co-authored-by: vllm-agent <claw@inferact.ai>

* [Model] Extend device-side mm normalization to DeepSeekV4 Vision / DeepSeekV41 (#57985)

Signed-off-by: cjackal <44624812+cjackal@users.noreply.github.com>
Co-authored-by: vllm-agent <claw@inferact.ai>

* [Perf][Model] Muse vision encoders: no GPU<->CPU syncs, build Glimmer's vision prep once per call (#64822)

Signed-off-by: subhash polisetti <subhashr161347@gmail.com>
Co-authored-by: vllm-agent <claw@inferact.ai>
Co-authored-by: Isotr0py <Isotr0py@outlook.com>

* [Spec Decode] Enable adaptive verification for `FlashInferMLASparseSM90Backend` (#62515)

Signed-off-by: cjackal <44624812+cjackal@users.noreply.github.com>
Co-authored-by: vllm-agent <claw@inferact.ai>

* [ROCm][gfx1151] ROCM_RDNA_ATTN: prefer 16-token KV blocks (#67344)

Signed-off-by: Zohaib Moti <zohaib.moti@amd.com>
Signed-off-by: Robert Esclapez Garcia <robert.garcia@amd.com>
Co-authored-by: Robert Esclapez <Robert.Garcia@amd.com>
Co-authored-by: Claude <noreply@anthropic.com>

* [ROCm][DI-CI] Enable MiniMax-M3.1-Flash in AMD CI pipeline (#67466)

Signed-off-by: lcskrishna <lollachaitanya@gmail.com>
Co-authored-by: vllm-agent <claw@inferact.ai>

* [Quantization] Skip kernel dispatch for online requantization sources (#66803)

Signed-off-by: Felix Marty <Felix.Marty@amd.com>
Co-authored-by: vllm-agent <claw@inferact.ai>

* [Quantization] Allow online per-channel fp8 quantization of an unquantized lm_head (#61815)

Signed-off-by: riceharvest <myappleiddarimaan@gmail.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: vllm-agent <claw@inferact.ai>

* [HiSparse] Revert the FlashInfer all-reduce backend override (#65201, #65522) (#67041)

Signed-off-by: Matthew Bonanni <mbonanni@redhat.com>

* [Bugfix][HiSparse] Fix GLM-5.3-Flash startup: derive resident/hot specs without grouping on workers (#67516)

Signed-off-by: Matthew Bonanni <mbonanni@redhat.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: vllm-agent <claw@inferact.ai>

* [Test] Restore PaddleOCR-VL generation coverage on Transformers v5 (#66267)

Signed-off-by: YuZheJin <1370324293@qq.com>
Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
Co-authored-by: YuZheJin <1370324293@qq.com>
Co-authored-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>

* [Bugfix][Model] DiffusionGemma: stop sampler recompiles during serving (#61669)

Signed-off-by: Pranav Abbott <abbott@meta.com>

* [Bugfix] Apply `WeightsMapper` stacked rules last, and fix the `--mm-encoder-only` shared root check (#65965)

Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: vllm-agent <claw@inferact.ai>

* [Bugfix][V2 runner] Keep idle-DP dummy requests within max_model_len (#64923)

Signed-off-by: aoshen02 <aoshen@inferact.ai>
Signed-off-by: 0z5a <0z5a@users.noreply.github.com>
Signed-off-by: 0z5a <dezhen.lu@student.uni-tuebingen.de>
Signed-off-by: Nick Hill <nickhill123@gmail.com>
Co-authored-by: aoshen02 <aoshen@inferact.ai>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: Nick Hill <nickhill123@gmail.com>
Co-authored-by: vllm-agent <claw@inferact.ai>

* [Bugfix] Reject prompt embeddings for BERT/RoBERTa models (#55233)

Signed-off-by: iharshlalakiya <harshlalakiya.me@gmail.com>
Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
Co-authored-by: Taneem Ibrahim <taneem.ibrahim@gmail.com>

* [Bugfix][Quantization] Handle dict-valued Quark `algo_config` in `maybe_update_config` (#65951)

Signed-off-by: Andy Friedrich <afriedri@amd.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: vllm-agent <claw@inferact.ai>

* [Bugfix] Allow query parameters in benchmark endpoint URLs (#55719)

Signed-off-by: Cenab Batu Bora <40250311+cenab@users.noreply.github.com>
Co-authored-by: Codex <noreply@openai.com>
Co-authored-by: Taneem Ibrahim <taneem.ibrahim@gmail.com>

* [Bugfix][Multimodal] Domain-separate content digests and caller uuids in mm hashes (#66473)

Signed-off-by: shadowmodder <sudhirvissa@gmail.com>
Co-authored-by: vllm-agent <claw@inferact.ai>

* [Bugfix][MoE] Preserve invalid expert IDs in NIXL round-robin mapping (#64086)

Signed-off-by: Chenxi Ma <114001934+algebra-MCX@users.noreply.github.com>
Co-authored-by: vllm-agent <claw@inferact.ai>
Co-authored-by: Lucas Wilkinson <LucasWilkinson@users.noreply.github.com>

* [XPU][Perf][Spec Decode] Enable triton greedy_argmax on XPU (#67196)

Signed-off-by: RyanMa29 <ziyang.ma@intel.com>
Co-authored-by: vllm-agent <claw@inferact.ai>

* Delete tools/flashinfer-build.sh (#67304)

Signed-off-by: Misha Goin <mgoin64@gmail.com>

* [Feature][DSv4.1] Support arbitrary pipeline stage partitions (#63106)

Signed-off-by: wangyicong <wangyicong@bytedance.com>
Signed-off-by: Yifei Wang <yifei.w@inferact.ai>
Co-authored-by: Codex <noreply@openai.com>
Co-authored-by: saipakalapati <79748893+saichowdary007@users.noreply.github.com>
Co-authored-by: Yifei Wang <yifei.w@inferact.ai>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>

* [PP] Require exact intermediate-tensor match on every PP transport (#67093)

Signed-off-by: Nick Hill <nickhill123@gmail.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>

* [Chore] Remove deprecated `rope_scaling` usage and forbid it in pre-commit (#67533)

Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>

* [Build] Bump vllm-flash-attn to 8dd9eaea (flash-attention#218, #208) (#67280)

Signed-off-by: Jhao-Ting Chen <jhaotingc@nvidia.com>
Signed-off-by: Matthew Bonanni <mbonanni@redhat.com>
Co-authored-by: vllm-agent <claw@inferact.ai>
Co-authored-by: Matthew Bonanni <mbonanni@redhat.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>

* [CI/Build] Fix docstring Markdown style in test_quark.py (#67549)

Signed-off-by: Madhav Jivrajani <madhav.jiv@gmail.com>

* [Bugfix] Keep extensible KV cache headroom for measured allocator fragmentation (#67296)

Signed-off-by: Nick Hill <nickhill123@gmail.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: vllm-agent <claw@inferact.ai>
Co-authored-by: Mirza Halilčević <mhalilce@amd.com>

* [Bugfix] Fix startup OOM in KV cache sizing warmup by releasing kernel warmup's KV blocks (#67485)

Signed-off-by: Vadim Gimpelson <vadim.gimpelson@gmail.com>
Signed-off-by: Nick Hill <nickhill123@gmail.com>
Co-authored-by: Claude <noreply@anthropic.com>
Co-authored-by: vllm-agent <claw@inferact.ai>
Co-authored-by: Nick Hill <nickhill123@gmail.com>

* [Transformers backend] Support DeepStack vision features (#67312)

Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: vllm-agent <claw@inferact.ai>

* [CI] Split ReplaySSM E2E into 1-GPU, TP2 and nightly H100 steps (#67553)

Signed-off-by: Kevin Luu <51931015+khluu@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: vllm-agent <claw@inferact.ai>

* [Bugfix][Spec Decode] Use the target's kernel block sizes for EAGLE draft attention (#64777)

Signed-off-by: Ankur-singh <ankusingh@nvidia.com>
Signed-off-by: wzhao18 <wzhao18.sz@gmail.com>
Co-authored-by: Claude <noreply@anthropic.com>
Co-authored-by: wzhao18 <wzhao18.sz@gmail.com>

* [ROCm][Muse Spark] DSpark Markov sampling with one launch a round (--monokernel) (#66051)

Signed-off-by: Aliasger Zaidy <aliasger.zaidy@amd.com>
Signed-off-by: Matt <156021403+mawong-amd@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: Matt <156021403+mawong-amd@users.noreply.github.com>
Co-authored-by: vllm-agent <claw@inferact.ai>
Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
Co-authored-by: Fangzhou Ai <31551580+Fangzhou-Ai@users.noreply.github.com>

* [CI] Make DSpark sharded-sampling test deterministic under torch.compile (#65733)

Signed-off-by: Aarushi Jain <aarushi.jain2@amd.com>
Co-authored-by: Andreas Karatzas <akaratza@amd.com>

* [ROCm][CI] Fix Sharded RDT bake test replay under persisted ROCm MoE padding (#67394)

Signed-off-by: PerryZhang01 <Perry.Zhang@amd.com>
Signed-off-by: Andreas Karatzas <akaratza@amd.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: vllm-agent <claw@inferact.ai>
Co-authored-by: Andreas Karatzas <akaratza@amd.com>

* [Bugfix] Fix KV cache pages for hybrid models with draft layers (#62800)

Signed-off-by: Seyfal Sultanov <seyfal.s@outlook.com>
Co-authored-by: vllm-agent <claw@inferact.ai>

* [EPLB] Support heterogeneous DeepSeek V4/V4.1 DSpark draft experts (#61395)

Signed-off-by: Markov Ilya <markovilya197@gmail.com>
Signed-off-by: ilmarkov <markovilya197@gmail.com>

* [Bugfix][Test] Avoid capfd in test_pp_rejects_split_kv_sharing (#67321)

Signed-off-by: aarushjain29 <aarushi.jain2@amd.com>

* [ROCm][CI] Run the nightly Python 3.14 smoke test on AMD (#66314)

Signed-off-by: Andreas Karatzas <akaratza@amd.com>
Co-authored-by: Claude <noreply@anthropic.com>
Co-authored-by: vllm-agent <claw@inferact.ai>

* [ROCm][CI] Cover test_areas gaps on AMD mirrors (#67317)

Signed-off-by: aarushjain29 <aarushi.jain2@amd.com>
Co-authored-by: Andreas Karatzas <akaratza@amd.com>

* [Transformers backend] Default decoder pooling to LAST and honour `is_causal=False` (#67455)

Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>

* [DeepSeek-V4] Warm the DeepGEMM kernels DeepSeek-V4/V4.1 JIT-compile at runtime (#64827)

Signed-off-by: hangrui cao <hangrui.cao@zoom.us>
Signed-off-by: HangruiCao2000 <caohangrui2000@gmail.com>
Co-authored-by: hangrui cao <hangrui.cao@zoom.us>
Co-authored-by: wangyicong <wangyicong@bytedance.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: Woosuk Kwon <woosuk@inferact.ai>
Co-authored-by: vllm-agent <claw@inferact.ai>

* [KV Connector][NIXL] Pull KV from pipeline-parallel prefill stages, with the hybrid KV cache manager (#65336)

Signed-off-by: liusy58 <mg21330037@smail.nju.edu.cn>
Signed-off-by: Yuhan Zhou <zhouyuhan_@outlook.com>
Signed-off-by: Nick Hill <nickhill123@gmail.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: Yuhan Zhou <zhouyuhan_@outlook.com>
Co-authored-by: OpenAI Codex <noreply@openai.com>
Co-authored-by: Nick Hill <nickhill123@gmail.com>
Co-authored-by: vllm-agent <claw@inferact.ai>

* [Perf][MoE] Mask DP padding rows in grouped top-k routing by default (#66799)

Signed-off-by: Bo Pang <bopang314@gmail.com>
Co-authored-by: vllm-agent <claw@inferact.ai>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
Co-authored-by: aoshen02 <aoshen@inferact.ai>

* [Recipe] Kimi-K3.1-Pro: MI355X kimi-k31-6d74a15 image, concurrency-48 with 3 draft tokens, align with InferenceX#5043 (#67487)

* [CI][Bugfix] Fix fake backend in drafter kernel block size test (#67579)

Signed-off-by: PerryZhang01 <Perry.Zhang@amd.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>

* [Bugfix][Multimodal] Repeat-last default padding for `cu_seqlens` in multimodal CG (#65453)

Signed-off-by: cjackal <44624812+cjackal@users.noreply.github.com>

* [Feature][EPD] Support Kimi-K3 image metadata-only inputs (#66925)

Signed-off-by: Tianyu Guo <guoty@inferact.ai>
Co-authored-by: vllm-agent <claw@inferact.ai>

* [CI][Multimodal] Use `uuid:`-tagged cache key for features in /generate test and docs (#67564)

Signed-off-by: PerryZhang01 <Perry.Zhang@amd.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: vllm-agent <claw@inferact.ai>

* [ROCm][Disagg][CI] Give Wide-EP headless nodes the KV config (#65098)

Signed-off-by: avpaul <avpaul@amd.com>
Signed-off-by: avininjamay8 <avpaul@amd.com>
Co-authored-by: avpaul <avpaul@amd.com>
Co-authored-by: Chaitanya Sri Krishna Lolla <lollachaitanya@gmail.com>
Co-authored-by: Shanshan Shen <467638484@qq.com>
Co-authored-by: Andreas Karatzas <akaratza@amd.com>
Co-authored-by: vllm-agent <claw@inferact.ai>

* [Perf] Optimize Garbage Collection at Init Time (#64542)

Signed-off-by: Alex Brooks <albrooks@redhat.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: vllm-agent <claw@inferact.ai>

* [ROCm][CI] Stub dcp_world_size in the merged-cache sparse MLA test (#67555)

Signed-off-by: aarushjain29 <aarushi.jain2@amd.com>
Signed-off-by: Andreas Karatzas <akaratza@amd.com>
Co-authored-by: Andreas Karatzas <akaratza@amd.com>
Co-authored-by: Claude <noreply@anthropic.com>
Co-authored-by: vllm-agent <claw@inferact.ai>

* [CI][ROCm][Kimi-K3] Compile each TP8 mono build and line the ranks up before launching it (#67439)

Signed-off-by: Kevin Luu <51931015+khluu@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: vllm-agent <claw@inferact.ai>

* [ROCm][CI] Retarget 36 MI300 1 GPU jobs to the mi300_dpx queue (#67308)

Signed-off-by: Sheral Kumar <shekumar@amd.com>
Co-authored-by: vllm-agent <claw@inferact.ai>

* [BugFix] Fix Mistral3 multimodal offline example (#43638)

Signed-off-by: FED4 <feiyyfrank@outlook.com>

* [ROCm] Keep the extensible KV cache for fd-exporting KV connectors (#67577)

Signed-off-by: Lin Sun <lin.sun@amd.com>
Co-authored-by: Cursor Agent <cursoragent@cursor.com>

* [Spec Decode][MRV2] Build no draft attention metadata on a context-K/V-only prefill (#67291)

Signed-off-by: zixi-qi <zixi@inferact.ai>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: vllm-agent <claw@inferact.ai>

* [Bugfix][Rust Frontend] SIGTERM only the managed engine's launcher so in-flight requests drain (#67282)

Signed-off-by: Liangjun Feng <lfeng@inferact.ai>
Signed-off-by: Bugen Zhao <i@bugenzhao.com>
Co-authored-by: Liangjun Feng <lfeng@inferact.ai>
Co-authored-by: Bugen Zhao <i@bugenzhao.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: vllm-agent <claw@inferact.ai>

* [Bugfix][Audio] Fall back to soundfile when torchcodec fails to decode (#67105)

Signed-off-by: Ekagra Ranjan <3116519+ekagra-ranjan@users.noreply.github.com>

* [Bugfix][Rust Frontend] Return 400 instead of 500 when tool parsing is disabled (#63580)

Signed-off-by: vito-lin-dev <ac1246897@gmail.com>
Co-authored-by: Claude <noreply@anthropic.com>
Co-authored-by: vllm-agent <claw@inferact.ai>

* [Bugfix][Multimodal] Size the MOSS-Audio profiling dummy to its audio budget (#67474)

Signed-off-by: subhash polisetti <subhashr161347@gmail.com>
Co-authored-by: vllm-agent <claw@inferact.ai>

* [Recipe] Kimi-K3.1-Pro: MI355X AgentX preset covering all six InferenceX#5043 concurrencies (#67600)

* [ROCm] Minimax M3.1 attn indexer XCD balancing (#66934)

Signed-off-by: Vinayak Gokhale <vinayak.gokhale@amd.com>
Co-authored-by: Micah Williamson <micah.williamson@amd.com>

* [ROCm] M3.1 GEMM improvements (#66933)

Signed-off-by: Vinayak Gokhale <vinayak.gokhale@amd.com>
Co-authored-by: Micah Williamson <micah.williamson@amd.com>

* [Perf] Use `slots=True` on all internal dataclasses (#67297)

Signed-off-by: Nick Hill <nickhill123@gmail.com>

* [CI][ROCm] Bake torchao into the ROCm test image off the default path and re-enable MI300/MI355 Quantization (#66672)

Signed-off-by: Thang Nguyen <thangnguyenvn647@gmail.com>
Signed-off-by: Andreas Karatzas <akaratza@amd.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: Andreas Karatzas <akaratza@amd.com>
Co-authored-by: vllm-agent <claw@inferact.ai>

* [Bugfix][XPU] Use int64 token index in the DeepSeek-V4 qnorm/RoPE kernel (#64515)

Signed-off-by: Artur Fierka <artur.fierka@intel.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: vllm-agent <claw@inferact.ai>

* [Bugfix] Fix EAGLE-3 on Transformers backend VLMs (#67519)

Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: vllm-agent <claw@inferact.ai>

* [Quantization][MoE] Keep NVFP4 expert scales CSF-compressed on the GPU with B12X (#64753)

Signed-off-by: Stefano Castagnetta <scastagnetta@nvidia.com>
Signed-off-by: lesj0610 <lesj0610@gmail.com>
Signed-off-by: Kam Basra <kameldipbasra@gmail.com>
Co-authored-by: lesj0610 <lesj0610@gmail.com>
Co-authored-by: Kam Basra <kameldipbasra@gmail.com>
Co-authored-by: Codex <noreply@openai.com>
Co-authored-by: vllm-agent <claw@inferact.ai>

* [Perf][Bugfix][RL] Fix sparse NCCL update OOM and improve performance (#66375)

Signed-off-by: Shuolei Wang <shuoleiwang123@gmail.com>
Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
Co-authored-by: vllm-agent <claw@inferact.ai>
Co-authored-by: aoshen02 <aoshen@inferact.ai>

* [Model] Remove OpenPangu-VL (#67632)

Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>

* [Bugfix][AuxOutput] Size the routed-experts row cache from one source (#67599)

Signed-off-by: Ao Shen <aoshen524@gmail.com>
Co-authored-by: Ao Shen <aoshen524@gmail.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>

* [Perf] Allow EPLB with modular TRT-LLM FP8 block-scale MoE (#67591)

Signed-off-by: aoshen524 <aoshen524@gmail.com>
Co-authored-by: aoshen524 <aoshen524@gmail.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>

* [CI][XPU] Bound DeepStack test profiling to avoid OOM (#67601)

Signed-off-by: Tony Lin <tony.lin@intel.com>

* [CI][Test] Run test_mm_prefix_lm_pipeline_parallel with spawned workers (#67637)

Signed-off-by: khluu <khluu000@gmail.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>

* [MM] Support mm_device_do_normalize for Qwen3-Omni (#67426)

Signed-off-by: gordan12 <gordan941222@gmail.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>

* [Chore] Update agent skill for vLLM (#67273)

Signed-off-by: yewentao256 <zhyanwentao@126.com>
Signed-off-by: Wentao Ye <44945378+yewentao256@users.noreply.github.com>
Co-authored-by: vllm-agent <claw@inferact.ai>
Co-authored-by: Benjamin Chislett <bchislett@nvidia.com>

* [MM] Support mm_device_do_normalize for LLaVA-OneVision (#67505)

Signed-off-by: Xinsheng Ooi <ooixinsheng@gmail.com>

* [CI] Bump Transformers version to 5.23.0 (#67502)

Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>

* [Transformers backend] Serve Qwen2 with the Transformers modeling backend (#66821)

Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: vllm-agent <claw@inferact.ai>

* [Bugfix][MRV2] Allow non 8-bytes aligned `--max-num-batched-tokens` for DeepSeek-v4 (#66401)

Signed-off-by: lazymio <mio@lazym.io>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: Shengqi Chen <harry-chen@outlook.com>
Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>

* [XPU][MLA] Register XPUFlashAttnMLABackend for FLASH_ATTN_MLA in XPU (#64255)

Signed-off-by: Zhouxiaoya25 <xiaoya.zhou@intel.com>
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Co-authored-by: vllm-agent <claw@inferact.ai>

* [Bugfix][Core] Keep dummy-run requests within max_model_len (#67545)

Signed-off-by: Simon Veitner <sveitner@redhat.com>
Co-authored-by: Claude Fable 5.2 <noreply@anthropic.com>
Co-authored-by: Lucas Wilkinson <LucasWilkinson@users.noreply.github.com>
Co-authored-by: vllm-agent <claw@inferact.ai>

* [CI][Test] Stub pipeline_relay in the engram DP-shard model forward test (#67636)

Signed-off-by: khluu <khluu000@gmail.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>

* [Core][DBO] Dual batch overlap under pure tensor parallelism (split PR of M3.1) (#63483)

Signed-off-by: Taoyu Zhu <zhutaoyu97@gmail.com>
Signed-off-by: benzhuzhu <zhutaoyu97@gmail.com>
Signed-off-by: Lucas Wilkinson <lwilkins@redhat.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: Lucas Wilkinson <lwilkins@redhat.com>
Co-authored-by: vllm-agent <claw@inferact.ai>

* [Model Loader] Fill dummy FP8 weights with random codes across the range (#67693)

Signed-off-by: Woosuk Kwon <woosuk@inferact.ai>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>

* [Perf][ROCm][DSv4.1] DeepSeek-V4.1-Flash: dense MXFP8 GEMMs and wo_a on AITER's tuned FlyDSL kernels, and a faster MXFP8 activation quant (MI355X) (#67377)

Signed-off-by: LI MOU <limou102@amd.com>
Co-authored-by: vllm-agent <claw@inferact.ai>
Co-authored-by: Fangzhou Ai <31551580+Fangzhou-Ai@users.noreply.github.com>
Co-authored-by: Shanshan Shen <467638484@qq.com>

* [CI] Start multi_process_parallel's Ray cluster with one CPU per rank (#67391)

Signed-off-by: wzhao18 <wzhao18.sz@gmail.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: vllm-agent <claw@inferact.ai>
Co-authored-by: Taneem Ibrahim <taneem.ibrahim@gmail.com>

* [Rust Frontend] Bump llm-multimodal version (#67670)

Signed-off-by: Isotr0py <Isotr0py@outlook.com>
Co-authored-by: vllm-agent <claw@inferact.ai>

* [Bugfix][Audio] Treat zero decode limits as disabled (#67611)

Signed-off-by: Ylean <61880605+Woolgathererer@users.noreply.github.com>

* [CI] Load gte-Qwen2 without remote code in test_non_causal_models (#67716)

Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>

* [XPU] Flatten N-D input in XPU MXFP8 linear kernel (#66494)

Signed-off-by: Ma, Liangliang <liangliang.ma@intel.com>
Co-authored-by: vllm-agent <claw@inferact.ai>
Co-authored-by: Kunshang Ji <kunshang.ji@intel.com>

* [Rust Frontend] Support n > 1 on /inference/v1/generate (#66878)

Signed-off-by: Dev Shah <dev.shah@cohere.com>
Signed-off-by: Bugen Zhao <i@bugenzhao.com>
Co-authored-by: Bugen Zhao <i@bugenzhao.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: vllm-agent <claw@inferact.ai>

* [Bugfix][KV Offload] Report only written blocks from fs tier store jobs (#63524)

Signed-off-by: Advay Monga <advaymonga@gmail.com>
Co-authored-by: Claude <noreply@anthropic.com>
Co-authored-by: vllm-agent <claw@inferact.ai>

* [CI] Fix NIXL EP import test (#67574)

Signed-off-by: Itay Alroy <ialroy@nvidia.com>
Co-authored-by: vllm-agent <claw@inferact.ai>

* [ROCm][PP] Stage hand-off on SDMA copy engines (--pp-sdma-handoff) (#65658)

Signed-off-by: Taoyu Zhu <zhutaoyu97@gmail.com>
Signed-off-by: Nick Hill <nickhill123@gmail.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: vllm-agent <claw@inferact.ai>
Co-authored-by: Nick Hill <nickhill123@gmail.com>

* [Bugfix][Muse Spark] Load the model classes lazily so importing the package doesn't initialize CUDA (#67691)

Signed-off-by: Nick Hill <nickhill123@gmail.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>

* [CI] Require batch invariance for derender parity tests (#67742)

Signed-off-by: Tianyu Guo <guoty9@mail2.sysu.edu.cn>

* [KV Offload][Bugfix] Hold P2P e2e test ports until the tiers bind them (#67607)

Signed-off-by: Andreas Karatzas <akaratza@amd.com>
Co-authored-by: Claude <noreply@anthropic.com>

* [CI][Test] Keep dummy FP8 weights within +-1 in the HiSparse e2e tests (#67753)

Signed-off-by: khluu <khluu000@gmail.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>

* [Core][MM] Make the multimodal encoder cache size configurable  (#65534)

Signed-off-by: HangruiCao2000 <caohangrui2000@gmail.com>

* [Perf] Allow DeepGEMM/Triton block-FP8 MoE with FlashInfer NVLink all2all (#67592)

Signed-off-by: aoshen524 <aoshen524@gmail.com>
Co-authored-by: aoshen524 <aoshen524@gmail.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>

* [ROCm][DSv4.1][Perf] Run the fused mHC seams as AITER's single-launch HIP Mega-MHC (#65992)

Signed-off-by: David Zhu <david.zhu@amd.com>
Co-authored-by: Fangzhou Ai <31551580+Fangzhou-Ai@users.noreply.github.com>

* [Bugfix][DSv4.1] Use int64 row offsets in the K-cache gather kernels (#65926)

Signed-off-by: Tuukka Sarvi <tuukka.sarvi@amd.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: Douglas Lehr <91553416+dllehr-amd@users.noreply.github.com>

* [CI][Test] Patch the platform class, not the instance, in the humming quantization tests (#67752)

Signed-off-by: khluu <khluu000@gmail.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>

* [Fast Start] Support compressed-tensors CUTLASS block FP8 cache restarts (#63976)

Signed-off-by: liusy58 <liusy58@smail.nju.edu.cn>
Signed-off-by: Isotr0py <Isotr0py@outlook.com>
Co-authored-by: liusy58 <liusy58@smail.nju.edu.cn>
Co-authored-by: Isotr0py <Isotr0py@outlook.com>

* [ROCm][DeepSeek-V4.1] Enable eager prefill context parallel for DeepSeek-V4.1-Flash (#66842)

Signed-off-by: Andreas Karatzas <akaratza@amd.com>
Co-authored-by: haic0 <149741444+haic0@users.noreply.github.com>
Co-authored-by: Claude <noreply@anthropic.com>
Co-authored-by: vllm-agent <claw@inferact.ai>
Co-authored-by: Divakar Verma <137818590+divakar-amd@users.noreply.github.com>
Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>

* [Bugfix] Cap warmup and CUDA graph sizes to max_num_seqs * max_model_len (#67469)

Signed-off-by: Yibo Liu <yibol9768@gmail.com>
Signed-off-by: Nick Hill <nickhill123@gmail.com>
Co-authored-by: vllm-agent <claw@inferact.ai>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>

* [MoE][NVIDIA] Support SM107 in flashinfer_moe_ep_cutedsl for NVFP4 experts for FI 0.7.3 (#65974)

Signed-off-by: Mingyuan Ma <mingyuanm@nvidia.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>

* [Bugfix][Mamba] Don't cache a block that a mid-block prefill resume ran past (#67717)

Signed-off-by: wzhao18 <wzhao18.sz@gmail.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>

* [XPU][LoRA] Support unquantized MoE LoRA (#62807)

Signed-off-by: Chaojun Zhang <chaojun.zhang@intel.com>
Co-authored-by: Kunshang Ji <kunshang.ji@intel.com>

* [ROCm][Perf][DSv4.1] Sort Engram host lookups by table offset on MI355X (#67805)

Signed-off-by: Fangzhou Ai <fangzhou.ai@amd.com>
Co-authored-by: Claude Sonnet 5.5 <noreply@anthropic.com>

* [Recipe] Kimi-K3.1-Pro: MI355X kimi-k31-5c14fe8 image and SP prefill, align with InferenceX#5089 (#67817)

* [Mooncake] Route KV transfers through a backend interface (#64363)

Signed-off-by: CAICAIIs <3360776475@qq.com>
Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
Co-authored-by: vllm-agent <claw@inferact.ai>

* [Feature][AuxOutput] Support routed experts with P/D disaggregation (#62056)

Signed-off-by: Ao Shen <aoshen524@gmail.com>
Co-authored-by: Ao Shen <aoshen524@gmail.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: aoshen02 <aoshen@inferact.ai>

* [ROCm][Muse Spark] DSpark draft layers as one persistent FlyDSL launch (--monokernel) (#65985)

Signed-off-by: Aliasger Zaidy <aliasger.zaidy@amd.com>
Signed-off-by: Matt <156021403+mawong-amd@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: Matt <156021403+mawong-amd@users.noreply.github.com>
Co-authored-by: vllm-agent <claw@inferact.ai>
Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
Co-authored-by: Fangzhou Ai <31551580+Fangzhou-Ai@users.noreply.github.com>

* [ROCm][Bugfix] Run the Canary WER correctness test in fp32 (#67606)

Signed-off-by: Andreas Karatzas <akaratza@amd.com>
Co-authored-by: Claude <noreply@anthropic.com>

* [Perf][Qwen3.5] Fuse dynamic per-token FP8 input quant into the Gemma add-RMSNorm (#67393)

Signed-off-by: Lantao Jin <ltjin@amazon.com>
Signed-off-by: Thien Tran <gau.nernst@yahoo.com.sg>
Co-authored-by: Claude <noreply@anthropic.com>
Co-authored-by: Thien Tran <gau.nernst@yahoo.com.sg>
Co-authored-by: vllm-agent <claw@inferact.ai>

* [ROCm][Quantization] Add load-time int4 lm_head quantization for RDNA… (#62708)

Signed-off-by: riceharvest <myappleiddarimaan@gmail.com>
Signed-off-by: Xiao Yu <xiao.yu.dc@outlook.com>
Co-authored-by: riceharvest <myappleiddarimaan@gmail.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
Co-authored-by: vllm-agent <claw@inferact.ai>

* [ROCm][Docs] Add DeepSeek-V4.1-Flash MI300X AgentX preset (#67586)

Signed-off-by: lishuoshuo-amd <shuoli@amd.com>
Co-authored-by: Claude <noreply@anthropic.com>
Co-authored-by: Shanshan Shen <467638484@qq.com>

* [Bugfix][EPD] Clear ECMooncakeConnector local cache on reset_encoder_cache (#66846)

Signed-off-by: grYe99 <guorongye99@gmail.com>
Co-authored-by: Claude Opus 5 <noreply@anthropic.com>
Co-authored-by: vllm-agent <claw@inferact.ai>

* [Bugfix] Reject bad_words on generate servers without a tokenizer (#67408)

Signed-off-by: paul <paulqwer.cs13@nycu.edu.tw>
Co-authored-by: Claude Fable 5.2 <noreply@anthropic.com>

* [Bugfix][XPU] Apply attn_sink in the DeepSeek-V4 sparse MLA Triton kernel (#64252)

Signed-off-by: Artur Fierka <artur.fierka@intel.com>
Co-authored-by: Claude Sonnet 5.5 <noreply@anthropic.com>
Co-authored-by: vllm-agent <claw@inferact.ai>
Co-authored-by: Kunshang Ji <kunshang.ji@intel.com>

* [Perf][Qwen4Exp] Drop the defensive zero fill from the QSA output path (#67728)

Signed-off-by: Thien Tran <gau.nernst@yahoo.com.sg>
Co-authored-by: vllm-agent <claw@inferact.ai>

* [Perf][Mamba] Optimize prefill splits for mid-block resume (3 forward passes to 1) (#66938)

Signed-off-by: wzhao18 <wzhao18.sz@gmail.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: vllm-agent <claw@inferact.ai>

* [Bugfix][Security] Check the uncompressed size of embedding archives before torch.load (#67316)

Signed-off-by: LeizhenZhang <zhangleizhen645@gmail.com>

* [Bugfix][Multimodal] Profile ERNIE-4.5-VL with the largest admitted video (#66937)

Signed-off-by: subhash polisetti <subhashr161347@gmail.com>

* [Model] Add EAGLE-3 / DSpark support (aux hidden states) for Lfm3ForCausalLM (#66323)

Signed-off-by: Yosef W. Alemneh <37704356+rasyosef@users.noreply.github.com>

* [Bugfix][Multimodal] Llama 4: honor per-request max_patches (#67813)

Signed-off-by: Yibo Liu <yibol9768@gmail.com>

* [XPU] quick fix for lora test workspace (#67824)

Signed-off-by: Yan Ma <yan.ma@intel.com>
Co-authored-by: vllm-agent <claw@inferact.ai>

* [Perf][Qwen4Exp] Fuse the MTP drafter's pre-FC Gemma RMSNorms into one Triton launch (#67725)

Signed-off-by: Thien Tran <gau.nernst@yahoo.com.sg>
Co-authored-by: vllm-agent <claw@inferact.ai>

* [Perf][GDN] Drop the prefill copies of the fused spec-decode path (#67733)

Signed-off-by: Thien Tran <gau.nernst@yahoo.com.sg>
Co-authored-by: vllm-agent <claw@inferact.ai>

* [KV Offload] Rename the filtered group index to config_group_idx (#65197)

Signed-off-by: Amir Friedman <Amir.Friedman1@ibm.com>
Co-authored-by: vllm-agent <claw@inferact.ai>
Co-authored-by: Or Ozeri <oro@il.ibm.com>

* [CI][ROCm] Fix the DSv4.1 sparse MLA forward_mqa test after #66842 (#67845)

Signed-off-by: Lantao Jin <ltjin@amazon.com>
Co-authored-by: Claude <noreply@anthropic.com>

* [Perf][Hy4] Enable sink-aware ShardQ prefill (#67490)

Signed-off-by: Yiliu Dong <91178480+qianlihuang@users.noreply.github.com>
Co-authored-by: OpenAI Codex <noreply@openai.com>
Co-authored-by: vllm-agent <claw@inferact.ai>

* [KV Offload] Add mmap_path to SharedOffloadRegion (#67258)

Signed-off-by: omerpaz95 <omerpaz95@gmail.com>
Co-authored-by: vllm-agent <claw@inferact.ai>

* [CI] Align CPU offload latency-test prefixes with cache blocks (#67761)

Signed-off-by: SyaOtiLan <954239196@qq.com>
Co-authored-by: Codex <noreply@openai.com>
Co-authored-by: vllm-agent <claw@inferact.ai>

* [DeepSeek-V3.2] Support sequence parallelism with pipeline parallelism (#65947)

Signed-off-by: Yifei Wang <yifei.w@inferact.ai>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: vllm-agent <claw@inferact.ai>

* [Perf][MiniMax-M3.1] MiniMax-M3.1-Flash Specialized Decode: CuTe DSL Optimized GB300 (up to 1.45x) (#66324)

Signed-off-by: Xin Li <xinli@nvidia.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>

* [Model] Support sequence parallelism with pipeline parallelism for Qwen3-Next, Qwen3.5 and DeepSeek-V2/V3 (#66866)

Signed-off-by: Yifei Wang <yifei.w@inferact.ai>
Co-authored-by: AceCoder0 <104333925+AceCoder0@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>

* [Multimodal] Require Rust preprocessing for GLM-5.3-Flash (#67237)

Signed-off-by: Yuxuan Zhang <Yuxuan.Zhang2@liverpool.ac.uk>
Signed-off-by: Isotr0py <Isotr0py@outlook.com>
Co-authored-by: OpenAI Codex <codex@openai.com>
Co-authored-by: vllm-agent <claw@inferact.ai>
Co-authored-by: Isotr0py <Isotr0py@outlook.com>

* [Fast Start] Add weights mirror support for daemon2daemon (#64641)

Signed-off-by: Isotr0py <Isotr0py@outlook.com>
Co-authored-by: liusy58 <mg21330037@smail.nju.edu.cn>
Co-authored-by: Xun Sun <UNIDY2002@outlook.com>
Co-authored-by: Kimi Code <noreply@moonshot.cn>
Co-authored-by: vllm-agent <claw@inferact.ai>

* [Kernel][XPU] Mamba2 SSD: tensor-descriptor path for _chunk_scan_fwd_kernel and _chunk_state_fwd_kernel (#60801)

Signed-off-by: Marceli Fylcek <marceli.fylcek@intel.com>
Signed-off-by: Artur Fierka <artur.fierka@intel.com>
Co-authored-by: Marceli Fylcek <marceli.fylcek@intel.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: vllm-agent <claw@inferact.ai>
Co-authored-by: Kunshang Ji <kunshang.ji@intel.com>

* [Bugfix][Distributed] Order packed NCCL loads after the caller stream (#65487)

Signed-off-by: Aaron Hao <hao-aaron@users.noreply.github.com>
Signed-off-by: aoshen02 <aoshen@inferact.ai>
Co-authored-by: Aaron Hao <hao-aaron@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: vllm-agent <claw@inferact.ai>
Co-authored-by: Nick Hill <nickhill123@gmail.com>
Co-authored-by: aoshen02 <aoshen@inferact.ai>

* [Rust Frontend] Pass pulled object storage model dir to the Rust frontend (#61788)

Signed-off-by: Lucas-Fernandes-Martins <lucasfmartins16@gmail.com>
Signed-off-by: Bugen Zhao <i@bugenzhao.com>
Co-authored-by: vllm-agent <claw@inferact.ai>
Co-authored-by: Bugen Zhao <i@bugenzhao.com>

* [Bugfix][Multimodal] Fix LLaVA-OneVision embeddings for unequal video frame counts (#67744)

Signed-off-by: chuncaohenli <604265882@qq.com>
Co-authored-by: OpenAI Codex <noreply@openai.com>

* [Bugfix][MRV2] Apply EVS-recomputed M-RoPE positions on every PP stage (#66632)

Signed-off-by: Yifei Wang <yifei.w@inferact.ai>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: vllm-agent <claw@inferact.ai>

* [Bugfix][KV Offload] Report transferred bytes in tiering metrics (#67788)

Signed-off-by: waizuichougou <2082431897@qq.com>
Co-authored-by: vllm-agent <claw@inferact.ai>

* [Bugfix][ROCm][MoRIIO] Slice WRITE heads by each layer's head count (#62474)

Signed-off-by: tjtanaa <tunjian.tan@embeddedllm.com>
Signed-off-by: kliuae <kuanfu.liu@embeddedllm.com>
Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Co-authored-by: kliuae <kuanfu.liu@embeddedllm.com>
Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>

* [Bugfix][Rust Frontend] Return 400 for media the request got wrong (#63749)

Signed-off-by: Tianyao Wu <rayroy31@gmail.com>
Signed-off-by: Bugen Zhao <i@bugenzhao.com>
Co-authored-by: Claude <noreply@anthropic.com>
Co-authored-by: Bugen Zhao <i@bugenzhao.com>
Co-authored-by: vllm-agent <claw@inferact.ai>

* [Rust Frontend] Return 500 for engine errors on non-streaming generate (#67869)

Signed-off-by: Bugen Zhao <i@bugenzhao.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: vllm-agent <claw@inferact.ai>

* [Bugfix][Rust Frontend] Composite transparent images over white like Python (#63912)

Signed-off-by: Tianyao Wu <rayroy31@gmail.com>
Signed-off-by: Bugen Zhao <i@bugenzhao.com>
Co-authored-by: Claude <noreply@anthropic.com>
Co-authored-by: Bugen Zhao <i@bugenzhao.com>
Co-authored-by: vllm-agent <claw@inferact.ai>

* [Rust Frontend] Domain-tag multimodal cache keys like Python (#67870)

Signed-off-by: Bugen Zhao <i@bugenzhao.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: vllm-agent <claw@inferact.ai>

* [Rust Frontend] Support --default-mm-loras (#67873)

Signed-off-by: Bugen Zhao <i@bugenzhao.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: vllm-agent <claw@inferact.ai>

* [Rust Frontend] Support --hf-token (#67871)

Signed-off-by: Bugen Zhao <i@bugenzhao.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: vllm-agent <claw@inferact.ai>

* [Test][Model Loader] Direct weight reload equals a cold start, bitwise (#65347)

Signed-off-by: Aaron Hao <hao-aaron@users.noreply.github.com>
Co-authored-by: Aaron Hao <hao-aaron@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: vllm-agent <claw@inferact.ai>

* [Bugfix][Attention] Reject TORCH_SDPA for decoder attention (#64965)

Signed-off-by: Aaron Deng <84045863+aarongdeng@users.noreply.github.com>
Co-authored-by: ChatGPT <noreply@openai.com>

* [ROCm][CI] Keep the CI kernel recorder out of the Python 3.14 smoke venv (#67694)

Signed-off-by: PerryZhang01 <Perry.Zhang@amd.com>
Co-authored-by: vllm-agent <claw@inferact.ai>

* [XPU][CI]Disable BMG Rust Frontend Core Correctness due to random accuracy issue (#67867)

Signed-off-by: zengxian <xiangdong.zeng@intel.com>

* [Rust Frontend] Support --kv-sharing-fast-prefill (#67872)

Signed-off-by: Bugen Zhao <i@bugenzhao.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: vllm-agent <claw@inferact.ai>

* [Rust Frontend] Fix tests that fail when sharing one test process (#67729)

Signed-off-by: Bugen Zhao <i@bugenzhao.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: vllm-agent <claw@inferact.ai>

* [Refactor][Worker] Expose platform initialization hooks (#65565)

Signed-off-by: Ronald1995 <ronaldautomobile@163.com>
Co-authored-by: vllm-agent <claw@inferact.ai>

* [Bugfix][Rust Frontend] Preserve SentencePiece leading spaces in logprobs (#63773)

Signed-off-by: reidliu41 <reid201711@gmail.com>
Signed-off-by: Bugen Zhao <i@bugenzhao.com>
Co-authored-by: Bugen Zhao <i@bugenzhao.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: vllm-agent <claw@inferact.ai>

* [Rust Frontend] Accept features together with content_parts on /generate (#67400)

Signed-off-by: Bugen Zhao <i@bugenzhao.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: vllm-agent <claw@inferact.ai>

* [MoE] enable all2all fi_one_sided by default only for Blackwell and Rubin (#54268)

Signed-off-by: Artem Perevedentsev <aperevedents@nvidia.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: vllm-agent <claw@inferact.ai>
Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>

* [CI] Fix SPLADE late-interaction reset test after ec_connector.reset() (#66846) (#67888)

Signed-off-by: PerryZhang01 <Perry.Zhang@amd.com>

* [Rust Frontend] Honor X-Vllm-Priority on generate and reject unparsable request headers (#67399)

Signed-off-by: Bugen Zhao <i@bugenzhao.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: vllm-agent <claw@inferact.ai>

* [CI] Sample GPU activity on Kubernetes jobs (L4, B200, H100) (#67910)

Signed-off-by: Kevin Luu <51931015+khluu@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>

* [XPU][UT] Enable v1/attention tests on XPU (#65788)

Signed-off-by: Yizhou Wang <yizhou.wang@intel.com>
Co-authored-by: vllm-agent <claw@inferact.ai>

* [Fast Start] Support compressed-tensors CUTLASS FP8 weight-cache restarts (#63673)

Signed-off-by: liusy58 <liusy58@smail.nju.edu.cn>
Signed-off-by: Isotr0py <Isotr0py@outlook.com>
Co-authored-by: liusy58 <liusy58@smail.nju.edu.cn>
Co-authored-by: Isotr0py <Isotr0py@outlook.com>

* [Perf][KV Connector][MooncakeConnector] Submit Mooncake transfer descriptors with array API (#67656)

Signed-off-by: Yuhan Zhou <zhouyuhan_@outlook.com>
Co-authored-by: Codex <noreply@openai.com>
Co-authored-by: vllm-agent <claw@inferact.ai>

* [Perf][KDA] Export Triton KDA prefill checkpoints via a per-sequence buffer (#67866)

Signed-off-by: Jared Wen <jaredwen@inferact.ai>
Co-authored-by: zjy0516 <riverclouds.zhu@qq.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: vllm-agent <claw@inferact.ai>

* [ROCm][DSv4.1][Perf] Always use the unfolded layer-0 mHC entry on gfx942 (#67853)

Signed-off-by: lishuoshuo-amd <shuoli@amd.com>
Co-authored-by: Claude <noreply@anthropic.com>
Co-authored-by: Shanshan Shen <467638484@qq.com>

* [RL][Model Loader] Load FP8 indexer wk and MLA projections independently, as FP8 or BF16, in any scale layout (#66118)

Signed-off-by: aoshen02 <aoshen@inferact.ai>
Signed-off-by: Aaron Hao <hao-aaron@users.noreply.github.com>
Co-authored-by: aoshen02 <aoshen@inferact.ai>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: Aaron Hao <hao-aaron@users.noreply.github.com>
Co-authored-by: vllm-agent <claw@inferact.ai>
Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>

* [Model Loader] Direct reload of online-quantized modules (#65802)

Signed-off-by: Aaron Hao <hao-aaron@users.noreply.github.com>
Signed-off-by: aoshen02 <aoshen@inferact.ai>
Co-authored-by: Aaron Hao <hao-aaron@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: vllm-agent <claw@inferact.ai>
Co-authored-by: aoshen02 <aoshen@inferact.ai>

* [Bugfix][Model] Gemma4/Gemma3n: allocate the PLE cache on the input device (#62659)

Signed-off-by: Yannick Schnider <Yannick.Schnider1@ibm.com>
Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

* [KV Offload] Explain EFAULT from a full /dev/shm in _madvise_populate_write (#62058)

Signed-off-by: Andrew Goh <andrew.goh@zillion.network>
Co-authored-by: Itay Etelis <92247226+Etelis@users.noreply.github.com>
Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>

* [Bugfix][KV Offload] Avoid false flushes for shared Full Attention prefixes (#67486)

Signed-off-by: Alex <alex.tech.lab@outlook.com>
Co-authored-by: OpenAI Codex <noreply@openai.com>
Co-authored-by: vllm-agent <claw@inferact.ai>

* [Perf] Add opt-in FlashInfer multimodal encoder autotuning (#65901)

Signed-off-by: Jakub Zakrzewski <jzakrzewski@nvidia.com>
Co-authored-by: Codex <noreply@openai.com>
Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>

* [ROCm][CI] Widen the bound of test_quark_online_rotation_wikitext (#67938)

Signed-off-by: bogdan-petkovic <bpetkovi@amd.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: vllm-agent <claw@inferact.ai>

* [Quantization] Online weight-only NVFP4 for dense linear layers (#67006)

Signed-off-by: Stefano Castagnetta <scastagnetta@nvidia.com>
Co-authored-by: vllm-agent <claw@inferact.ai>

* [Rust Frontend] Serve /inference/v1/abort_requests (#67893)

Signed-off-by: Bugen Zhao <i@bugenzhao.com>
Co-authored-by: weihantu <das41086@gmail.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: vllm-agent <claw@inferact.ai>

* [Bugfix][Mooncake] Make early prefill abort cleanup idempotent (#65672)

Signed-off-by: staryxchen <staryxchen@tencent.com>

* [Rust Frontend] Support --tokens-only (#67708)

Signed-off-by: Dev Shah <dev.shah@cohere.com>
Signed-off-by: Bugen Zhao <i@bugenzhao.com>
Co-authored-by: Bugen Zhao <i@bugenzhao.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>

* [Model] Add opt-in PrimTS attention for Step5 on Blackwell (#62550)

Signed-off-by: Perkz Zheng <67892460+PerkzZheng@users.noreply.github.com>
Signed-off-by: <>
Co-authored-by: Codex <noreply@openai.com>
Co-authored-by: Claude Sonnet 5.5 <noreply@anthropic.com>
Co-authored-by: Perkz Zheng <perkzz@oci-jhb-slurm-1-vscode-01.cm.cluster>
Co-authored-by: Jee Jee Li <pandaleefree@gmail.com>

* [MRV2][Spec Decode] Cap dummy-run requests by their draft slots (#67571)

Signed-off-by: JonSnow1807 <JonSnow1807@users.noreply.github.com>
Signed-off-by: Nick Hill <nickhill123@gmail.com>
Co-authored-by: JonSnow1807 <JonSnow1807@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>

* [Model] Add LFM3-Omni (#65829)

Signed-off-by: paulpak58 <52512091+paulpak58@users.noreply.github.com>

* [Bugfix][Distributed] Fall back to torch.distributed when pynccl is disabled (#67882)

Signed-off-by: zjy0516 <riverclouds.zhu@qq.com>
Co-authored-by: zjy0516 <riverclouds.zhu@qq.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>

* [CI][Test] Compare logprobs in the Machete bf16 repeat-run test (#67917)

Signed-off-by: khluu <khluu000@gmail.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>

* [Model][DSv4.1] DeepSeek-V4.1 on A100 (SM80) and H100 (SM90) (#66497)

Signed-off-by: Yaoyao Ding <yaoyaod@nvidia.com>
Signed-off-by: Misha Goin <mgoin64@gmail.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: Misha Goin <mgoin64@gmail.com>
Co-authored-by: vllm-agent <claw@inferact.ai>
Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>

* [ROCm][CI] Make /ci rocm cache a cache-only CI build that also copies the runtime (#66032)

Signed-off-by: Andreas Karatzas <akaratza@amd.com>
Co-authored-by: Claude <noreply@anthropic.com>
Co-authored-by: vllm-agent <claw@inferact.ai>

* [ROCm][Model] Run Qwen3.5 on the eager breakable-graph model on MI355X (#66320)

Signed-off-by: Andreas Karatzas <akaratza@amd.com>
Co-authored-by: Claude <noreply@anthropic.com>
Co-authored-by: vllm-agent <claw@inferact.ai>

* [ROCm][CI] Re-enable VLLM_GPU_SYNC_CHECK=error in the ROCm test image (#64176)

Signed-off-by: Andreas Karatzas <akaratza@amd.com>
Co-authored-by: Claude <noreply@anthropic.com>
Co-authored-by: vllm-agent <claw@inferact.ai>

* [ROCm][CI] Dump child process stacks when an AMD CI test times out (#67605)

Signed-off-by: Andreas Karatzas <akaratza@amd.com>
Co-authored-by: Claude <noreply@anthropic.com>

* [Bugfix][ROCm] Fix ROCM_AITER_FA context loss on out-of-order batches (#59098)

Signed-off-by: Pham Binh <194774975+phambinhfin@users.noreply.github.com>
Co-authored-by: Pham Binh <194774975+phambinhfin@users.noreply.github.com>
Co-authored-by: Qiang Li <qiang.li2@amd.com>
Co-authored-by: vllm-agent <claw@inferact.ai>
Co-authored-by: Douglas Lehr <91553416+dllehr-amd@users.noreply.github.com>
Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
Co-authored-by: Simon Danielsson <70206058+simondanielsson@users.noreply.github.com>

* [Recipe][ROCm] ripple-65b on mi350x and mi355x (#67055)

Signed-off-by: jamesETsmith <james.smith9113@gmail.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
Co-authored-by: Hongxia Yang <62075498+hongxiayang@users.noreply.github.com>

* [Bugfix] Select only batch-invariant MXFP8 linear kernels under VLLM_BATCH_INVARIANT (#67806)

Signed-off-by: Flora Feng <4florafeng@gmail.com>

* [Config] Make --monokernel and decode_backend=latency_specialized one switch (#67315)

Signed-off-by: Hongxia Yang <hongxia.yang@amd.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>

* [Kernel][Perf] Add H100 BF16 batch-invariant matmul configs for gpt-oss (#65931)

Signed-off-by: imthomp <it238@byu.edu>
Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>

* [KV Offload] Set the CPU tier up in the background under async_init (#66583)

Signed-off-by: Itay Etelis <92247226+Etelis@users.noreply.github.com>
Co-authored-by: vllm-agent <claw@inferact.ai>

* [KV Connector][Mooncake] Fail pulls from a lost prefill worker (#65575)

Signed-off-by: CJJ1008 <843515223@qq.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>

* [Doc] Correct the batch-invariance comment on Hopper/Blackwell split-k (#67179)

Signed-off-by: imthomp <it238@byu.edu>

* [CI][Test] Prevent NVSHMEM IB transport setup hang in DBO GSM8K tests (#67852)

Signed-off-by: Manikanta Bandham <bandhammanikanta@gmail.com>
Co-authored-by: Wei Zhao <51183510+wzhao18@users.noreply.github.com>

* [ROCm][Quantization][Perf] Fuse the block Hadamard input rotation into AITER's MXFP4 activation quant (#63272)

Signed-off-by: Mikko Tukiainen <Mikko.Tukiainen@amd.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
Co-authored-by: vllm-agent <claw@inferact.ai>
Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>

* [Perf][MoE] OCP MX emulation: dequantize only the routed experts (#64406)

Signed-off-by: Jon Belof <jonathan.belof@amd.com>
Co-authored-by: Claude <noreply@anthropic.com>
Co-authored-by: Eugene Kuznetsov <eugene.kuznetsov@amd.com>
Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
Co-authored-by: vllm-agent <claw@inferact.ai>

* [ROCm][Bugfix] Port Mistral Large 4 AMD indexer cache to get_kv_cache_config (#67984)

Signed-off-by: Xavier Aguilar <xavier.aguilarfruto@amd.com>
Co-authored-by: vllm-agent <claw@inferact.ai>

* [MoE] Disable expanded DeepEP v2 dispatch in all execution modes (#66653)

Signed-off-by: Tyler Michael Smith <tlrmchlsmth@gmail.com>
Signed-off-by: Bill Nell <bnell@redhat.com>
Signed-off-by: bnellnm <49004751+bnellnm@users.noreply.github.com>
Signed-off-by: mgoin <mgoin64@gmail.com>
Co-authored-by: Tyler Michael Smith <tlrmchlsmth@gmail.com>
Co-authored-by: Codex <noreply@openai.com>
Co-authored-by: vllm-agent <claw@inferact.ai>
Co-authored-by: Misha Goin <mgoin64@gmail.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>

* [Bugfix][MiniMax-M3] Drop the stale config read from the FlashInfer indexer builder (#67215)

Signed-off-by: 최재영 <144660874+Jaeyeong-CHOI@users.noreply.github.com>
Co-authored-by: Claude Fable 5.2 <noreply@anthropic.com>

* [ROCm][Kimi-K3] Enable encoder graph capture for Kimi-K3 Vision Tower (#63086)

Signed-off-by: Samu Tamminen <samu.tamminen@amd.com>
Co-authored-by: Claude Code <noreply@anthropic.com>
Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
Co-authored-by: vllm-agent <claw@inferact.ai>

* [Bugfix][Distributed] Guard cross-node symmetric-memory rendezvous (#59208)

Signed-off-by: Mickael Seznec <mickael@mistral.ai>
Co-authored-by: Wei Zhao <51183510+wzhao18@users.noreply.github.com>

* [ROCm][Perf] Use tuned AITER Triton `gemm_a8w8` for per-tensor FP8 decode GEMMs on gfx950 (#65711)

Signed-off-by: amd-jrosas <joshua.rosas@amd.com>
Co-authored-by: vllm-agent <claw@inferact.ai>

* [CI][Test] Share the narrow dummy FP8 fixture with the NVFP4 fusion test (#67809)

Signed-off-by: wzhao18 <wzhao18.sz@gmail.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>

* [Spec Decode] Add DSpine support (#67094)

Signed-off-by: shanjiaz <zsjwpianpian@gmail.com>
Signed-off-by: Tomas Ruiz <tomas.ruiz.te@gmail.com>
Co-authored-by: Codex <codex@openai.com>
Co-authored-by: Tomas Ruiz <tomas.ruiz.te@gmail.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: Codex <noreply@openai.com>

* [XPU][batch invariant] Optimize batch-invariant collectives with sharded reductions (#67182)

Signed-off-by: Tony Lin <tony.lin@intel.com>
Co-authored-by: Flora Feng <4florafeng@gmail.com>

* [ROCm][Kernel] Extend skinny gemm N=5 to N=8 cases on GFX12 (RDNA4) using  SWMMAC optimization (#45559)

Signed-off-by: keneoneth <keneoneth18@gmail.com>
Co-authored-by: vllm-agent <claw@inferact.ai>
Co-authored-by: Douglas Lehr <91553416+dllehr-amd@users.noreply.github.com>
Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>

* [Perf] Use value-only reduction for FP8 group quantization (#59564)

* [Bugfix][DSv4.1] Import the SM80/SM90 metadata builders lazily (#68005)

Signed-off-by: Thang Nguyen <thangnguyenvn647@gmail.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>

* [Quant] LUT-B: fit per-channel block32 weight scales (#67267)

Signed-off-by: mgoin <mgoin64@gmail.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: vllm-agent <claw@inferact.ai>

* [CI][ROCm] Keep HIP uninitialized in the GPU memory wait (#67821)

Signed-off-by: Andreas Karatzas <akaratza@amd.com>
Co-authored-by: Claude <noreply@anthropic.com>
Co-authored-by: vllm-agent <claw@inferact.ai>

* [Spec Decode][MRV2] Pass NewRequestData to speculator add_request; add shutdown hook (#64306)

Signed-off-by: Chenyu Zhang <chenyzhang@nvidia.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: vllm-agent <claw@inferact.ai>

* [ROCm][CI] Retarget 8 MI300 1 GPU jobs to the mi300_dpx queue (#67993)

Signed-off-by: Sheral Kumar <shekumar@amd.com>
Co-authored-by: vllm-agent <claw@inferact.ai>

* [Tracing] Test the GenAI span attributes without a GPU; cache_read is standard

- Add a CPU-only unit test of OutputProcessor.do_tracing covering
  gen_ai.usage.input_tokens, output_tokens, cache_read.input_tokens and
  gen_ai.response.finish_reasons (stop, length, none), plus the deprecated
  prompt/completion names that stay for existing consumers.
- Assert finish_reasons in the GPU test too.
- gen_ai.usage.cache_read.input_tokens is in the OTel GenAI registry now,
  so list it with the standard attributes, not the custom ones.

Signed-off-by: Sugi Venugeethan <email@cognilogue.com>

* [Tracing] Keep finish_reasons for requests that stop normally

FinishReason is an IntEnum and STOP is 0, so "if finish_reason" dropped
the most common reason and the span carried an empty list. Compare with
None, and use str(finish_reason) (FINISH_REASON_STRINGS) rather than the
enum name. The unit test covers stop, length, abort and none.

Signed-off-by: Sugi Venugeethan <email@cognilogue.com>

---------

Signed-off-by: Juan Pérez de Algaba <jperezde@redhat.com>
Signed-off-by: Li, Tianmu <tianmu.li@intel.com>
Signed-off-by: Lucas Wilkinson <lwilkins@redhat.com>
Signed-off-by: Bugen Zhao <i@bugenzhao.com>
Signed-off-by: Dao Le <Dao007forever@gmail.com>
Signed-off-by: TianZhuo <2770730562@qq.com>
Signed-off-by: fangyuchu <fangyuchu@qq.com>
Signed-off-by: Tyler Michael Smith <tlrmchlsmth@gmail.com>
Signed-off-by: swappy <59965507+rycerzes@users.noreply.github.com>
Signed-off-by: Zheng Gong <zgong@amd.com>
Signed-off-by: Lantao Jin <ltjin@amazon.com>
Signed-off-by: khluu <khluu000@gmail.com>
Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
Signed-off-by: Kevin Luu <51931015+khluu@users.noreply.github.com>
Signed-off-by: Thien Tran <gau.nernst@yahoo.com.sg>
Signed-off-by: Artur Fierka <artur.fierka@intel.com>
Signed-off-by: Anh Tran <anhtra@nvidia.com>
Signed-off-by: PerryZhang01 <Perry.Zhang@amd.com>
Signed-off-by: Canlin <canlinguosdu@gmail.com>
Signed-off-by: Aditya Ganesh Kumar <adty910@gmail.com>
Signed-off-by: Afloat <ShuffleMonkey@163.com>
Signed-off-by: shen-shanshan <467638484@qq.com>
Signed-off-by: NickLucche <nicolo.lucchesi@mistral.ai>
Signed-off-by: lesj0610 <lesj0610@users.noreply.github.com>
Signed-off-by: lesj0610 <lesj0610@gmail.com>
Signed-off-by: Yiliu Dong <91178480+qianlihuang@users.noreply.github.com>
Signed-off-by: Tomas Ruiz <tomas.ruiz.te@gmail.com>
Signed-off-by: Cy <15118312+cycancode@users.noreply.github.com>
Signed-off-by: mgoin <mgoin64@gmail.com>
Signed-off-by: yiliu30 <yi4.liu@intel.com>
Signed-off-by: cjackal <44624812+cjackal@users.noreply.github.com>
Signed-off-by: subhash polisetti <subhashr161347@gmail.com>
Signed-off-by: Zohaib Moti <zohaib.moti@amd.com>
Signed-off-by: Robert Esclapez Garcia <robert.garcia@amd.com>
Signed-off-by: lcskrishna <lollachaitanya@gmail.com>
Signed-off-by: Felix Marty <Felix.Marty@amd.com>
Signed-off-by: riceharvest <myappleiddarimaan@gmail.com>
Signed-off-by: Matthew Bonanni <mbonanni@redhat.com>
Signed-off-by: YuZheJin <1370324293@qq.com>
Signed-off-by: Pranav Abbott <abbott@meta.com>
Signed-off-by: aoshen02 <aoshen@inferact.ai>
Signed-off-by: 0z5a <0z5a@users.noreply.github.com>
Signed-off-by: 0z5a <dezhen.lu@student.uni-tuebingen.de>
Signed-off-by: Nick Hill <nickhill123@gmail.com>
Signed-off-by: iharshlalakiya <harshlalakiya.me@gmail.com>
Signed-off-by: Andy Friedrich <afriedri@amd.com>
Signed-off-by: Cenab Batu Bora <40250311+cenab@users.noreply.github.com>
Signed-off-by: shadowmodder <sudhirvissa@gmail.com>
Signed-off-by: Chenxi Ma <114001934+algebra-MCX@users.noreply.github.com>
Signed-off-by: RyanMa29 <ziyang.ma@intel.com>
Signed-off-by: Misha Goin <mgoin64@gmail.com>
Signed-off-by: wangyicong <wangyicong@bytedance.com>
Signed-off-by: Yifei Wang <yifei.w@inferact.ai>
Signed-off-by: Jhao-Ting Chen <jhaotingc@nvidia.com>
Signed-off-by: Madhav Jivrajani <madhav.jiv@gmail.com>
Signed-off-by: Vadim Gimpelson <vadim.gimpelson@gmail.com>
Signed-off-by: Ankur-singh <ankusingh@nvidia.com>
Signed-off-by: wzhao18 <wzhao18.sz@gmail.com>
Signed-off-by: Aliasger Zaidy <aliasger.zaidy@amd.com>
Signed-off-by: Matt <156021403+mawong-amd@users.noreply.github.com>
Signed-off-by: Aarushi Jain <aarushi.jain2@amd.com>
Signed-off-by: Andreas Karatzas <akaratza@amd.com>
Signed-off-by: Seyfal Sultanov <seyfal.s@outlook.com>
Signed-off-by: Markov Ilya <markovilya197@gmail.com>
Signed-off-by: ilmarkov <markovilya197@gmail.com>
Signed-off-by: aarushjain29 <aarushi.jain2@amd.com>
Signed-off-by: hangrui cao <hangrui.cao@zoom.us>
Signed-off-by: HangruiCao2000 <caohangrui2000@gmail.com>
Signed-off-by: liusy58 <mg21330037@smail.nju.edu.cn>
Signed-off-by: Yuhan Zhou <zhouyuhan_@outlook.com>
Signed-off-by: Bo Pang <bopang314@gmail.com>
Signed-off-by: Tianyu Guo <guoty@inferact.ai>
Signed-off-by: avpaul <avpaul@amd.com>
Signed-off-by: avininjamay8 <avpaul@amd.com>
Signed-off-by: Alex Brooks <albrooks@redhat.com>
Signed-off-by: Sheral Kumar <shekumar@amd.com>
Signed-off-by: FED4 <feiyyfrank@outlook.com>
Signed-off-by: Lin Sun <lin.sun@amd.com>
Signed-off-by: zixi-qi <zixi@inferact.ai>
Signed-off-by: Liangjun Feng <lfeng@inferact.ai>
Signed-off-by: Ekagra Ranjan <3116519+ekagra-ranjan@users.noreply.github.com>
Signed-off-by: vito-lin-dev <ac1246897@gmail.com>
Signed-off-by: Vinayak Gokhale <vinayak.gokhale@amd.com>
Signed-off-by: Thang Nguyen <thangnguyenvn647@gmail.com>
Signed-off-by: Stefano Castagnetta <scastagnetta@nvidia.com>
Signed-off-by: Kam Basra <kameldipbasra@gmail.com>
Signed-off-by: Shuolei Wang <shuoleiwang123@gmail.com>
Signed-off-by: Ao Shen…
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

ci/build documentation Improvements or additions to documentation rocm Related to AMD ROCm

Projects

Status: Todo

Development

Successfully merging this pull request may close these issues.

4 participants