Repository navigation
Conversation
|
/tag-and-run-ci |
This comment has been minimized.
This comment has been minimized.
|
⛔ See the command reference. |
|
/tag-and-rerun-ci |
Source: sgl-project#32136 Source-head: 5d3455e Mechanically replay the source PR net patch on current origin/main.
Source: sgl-project#44561 Source-head: 28b7113 Mechanically replay the current source PR net patch on origin/main. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Source: sgl-project#44851 Source-head: c345391 Mechanically replay the source PR net patch on current origin/main.
Source: sgl-project#44863 Source-head: eccea92 Mechanically replay the source PR net patch on current origin/main. Preserve main's CUDA_ONLY declaration while adding the source XPU constant.
c6f612d to
49f6b6c
Compare
Preview approval requiredThis pull request comes from a fork, so Mintlify won't build a preview until someone with write access approves commit |
|
@airMeng #43208 has the two ignore-pattern fixes you reproduced since 59451c3: patterns are matched against the complete key, and the original checkpoint names are kept across the mapper. Each has a regression test, details in #43208 (comment). Could you take it into the next revision of this consolidation? |
|
Superseded by #46623. Carried the useful, still-open sources #44561 (Qwen4 CUDA import guard and XPU PLE tests), #44851 (Gemma4 text-only serving), and #44863 (XPU platform/capability detection) into the latest consolidation as separate commits preserving source attribution. #32136 is now closed and was omitted under the weekly eligibility policy. Combined validation: 144 CPU tests passed, 18 hardware-dependent tests skipped; pre-commit, Python syntax checks, and git diff --check passed. Please continue consolidation review in #46623. |
Purpose
Consolidate four small XPU enablement/backend-parity source patches into one review unit. Human review of source eligibility and validation is still required.
Run: REVISED — 2026-10-21 (America/Vancouver). Base:
a3d00eff58b6c22ea7539f882cc10c537634200b. Candidate inventory: the runner’s authoritative 2026-10-19 XPU source report (118 SGLang PR entries); no repository-wide open-PR scan. This revision follows source-PR review feedback. Earlier consolidations #37193, #39439, #40664 and #42627 remain audited for repeat sources.Included source PRs
Four sources remain: #32136, #44561, #44851 and #44863. Each is open and retains a nonempty patch absent from current main and earlier consolidations. #44561’s latest author commit addresses the requested Qwen4 XPU test activation. Each source is represented by one mechanically replayed commit; source authors retain responsibility for any further substantive changes.
Pinned wheel:
v0.4.0+xpufrom currentpython/pyproject_xpu.toml. Exact kernel source release R =028ac5a4e92a4addabeedac1a94374515f68e4a5(v0.4.0). Each numbered dependency below was compared directly to R and has zero commits behind. API, assertions, dtype/layout behavior and focused tests were inspected at R; unspecified manual environments and green top-level CI were not used as release proof.5d3455e1261a6aeffcfb6f9c59768b9e564eb75a— Focused XPU top-k transform coverage. Kernel openai chat speculative execution #248 (3cfe90204f9367820c7c2230dfae72194097fd76) is an ancestor of release R. Inspectedpython/sgl_kernel/top_k.py,src/sycl/TopKTransform.cpp, operator registrations andtests/test_topk_transform.py: float32 scores, int32 lengths/tables/offsets, K=2048, PAGED/RAGGED transforms and -1 padding are supported. Test-only patch; formatting normalized by current Ruff.28b71137aa8d7186c5628492cfbafeb455dc249d— CUDA-only Qwen4 decode import guard plus a device-agnostic PLE prefetch stream; enables the existing Qwen4 PLE offload tests on XPU. No new kernel API.c3453919feda25cdaea8659210006dbccaed7bca— Narrow Gemma4 text-only enablement: omit unused encoders/weights, preserve the processor tokenizer, use existing language-model-only admission. Kernel-independent: no attention/MLP operator, signature, dtype or layout changes; the source’s original full-model fixture already uses the same intel_xpu backend/page-size settings. Includes its cookbook recipe and coverage; does not include frozen-KV parent-stack work.eccea92f45cbd292eb3300fb7a6f94c80ba0c4c4— Small XPU platform/capability-detection parity fix with regression coverage. Kernel-independent: device inventory/selection metadata only. Its two-file net patch is independent of the declared serving stack; replay preserves main’s CUDA_ONLY constant.Commit policy
Exactly one source PR per commit, in ascending PR-number order. Preserve each source Git author identity/date, use a title ending in
(#PR_NUMBER), and include the source URL/head in its body. No source branches, comments, reviews or PRs were modified. Only this managed weekly branch is pushed toairMeng; nothing is merged.69348f3764ea1155e2a4b2d14e43b647dfa252ae— [XPU] Add top-k transform coverage (Add XPU topk_transform ut #32136)f9c1cb05614a2f72bb610293626e4da356096e5d— [XPU] Fix Qwen4Exp CUDA import and enable PLE tests ([Fix] [Intel] qwen4_exp: only import qwen4_decode on CUDA; run the qwen4 PLE tests on XPU #44561)d0763782aa1e20823a27ecff89f55be8dd419e7d— [XPU] Enable Gemma4 encoder-free text serving ([Feat] gemma4: Support encoder-free text serving #44851)c6f612d147629eb3848768c1924ef9ee32037124— [XPU] Detect kernel platform capabilities ([Fix] [Intel] kernels: Recognize XPU capabilities #44863)Excluded candidates
Each authoritative SGLang report entry not selected is listed individually. Kernel numbers refer to
sgl-project/sgl-kernel-xpu.Removed from this consolidation after review follow-up:
#43208 — @airMeng reproduced two unresolved HF ignore-pattern gaps (full checkpoint keys are truncated before matching; original names are lost after fallback mapping). The source head was unchanged after the requested regression fixes.
#43843 — Requested the related XPU tests be enabled; the source head was unchanged after the request.
#43891 — Requested the related tests be enabled; the source head was unchanged after the request.
#44670 — Unresolved D2H-copy/forced-synchronization performance concern; @mingfeima also requested changes and performance validation. The source head was unchanged.
#28716 — Unresolved Changes Requested by @airMeng and @siju-samuel; JIT infrastructure remains standalone.
#29639 — Draft; not Ready for Review.
#30980 — Changes Requested concerning SWA translation, local-attention metadata, test imports, and unrelated decode enablement.
#31290 — Changes Requested concerning collective buffer mutation, deterministic behavior, and invalid test registration.
#31780 — Unresolved reviewer request for corresponding pointer changes in companion [XPU][Fix] Pack device-pointer tables as uint64 to avoid 64-bit address overflow #35051.
#32503 — Changes Requested by @rahulvijayaraghavan on HiCache docs, CRI support and kernel-import design; broad subsystem change.
#32684 — Entire XPU FP8 import patch is already present on audited main.
#35138 — Broad GPT-OSS MXFP4 quantization/model integration exceeds simple enablement.
#35304 — Broad ModernBERT/NomicBERT/InternVL implementation; unresolved Changes Requested.
#35549 — Draft; not Ready for Review.
#35667 — Changes Requested by @polisettyvarma; requested imports already exist on main.
#36164 — Large Apple Silicon whole-model MLX subsystem; outside XPU scope.
#36401 — Expands deterministic-cache CI across eleven backends; outside simple XPU enablement.
#36803 — Permanently excluded: already consolidated in Xpu/weekly simple model enablement 2026 08 30 #37193.
#37037 — Already carried by the September 21 consolidation.
#37080 — Already carried by the September 14 consolidation.
#37233 — General README translation; unresolved compatibility-claim review.
#37247 — Already carried by the September 14 consolidation.
#37292 — Broad deterministic-inference implementation with unresolved device-support and KL-coverage feedback.
#37602 — General quantization-registry refactor; outside simple XPU enablement.
#37698 — Multiple unresolved Changes Requested; 28 unresolved runtime/diffusion threads.
#37735 — CUDA FlashMLA/DCP policy correction; outside XPU enablement.
#37927 — Unresolved reviewer request to retain only
mem_get_infoand remove other changes.#38262 — FP8 GEMM runtime/configuration policy change; test-removal review thread remains unresolved after author replies.
#38263 — Shared memory-monitor and warning-policy change; unresolved reviewer thread awaits source review.
#38266 — Unresolved review threads on dispatch gating, test removal and model accuracy remain for the source reviewer.
#38267 — Profiler mapping/policy change with unresolved reviewer discussion about explicit XPU activities.
#38272 — Large DSV4 graph-capture implementation; exceeds straightforward dispatch parity.
#38643 — Broad cache-offloading implementation across 19 files.
#38675 — New CI sharding launcher and concurrency policy; outside simple enablement.
#38873 — Large new XPU environment-reporting class and hardware diagnostics; exceeds small parity scope.
#38907 — Draft; not Ready for Review.
#39447 — New HiCache transfer backend across 12 files; subsystem enablement remains standalone.
#39714 — Unresolved review challenges quality thresholds against observed black-frame scores.
#39827 — New release-validation workflow/image qualification; outside simple enablement.
#39891 — Changes Requested concerning suite registration, rollback coverage, and device mapping.
#40029 — BLOCKED: required kernel Fix issue: cuda index incorrectly -1 when tp_size>1 #496 remains open and unreleased.
#40052 — Changes activation quantization, scale layouts, and shared FP8 fallback; requires standalone numerical/performance review.
#40569 — Generic NCCL error-path assertions; does not enable a supported XPU execution path.
#40614 — Draft; not Ready for Review.
#40616 — Draft; not Ready for Review.
#40720 — New serialized-tensor shared-memory transport; substantial weight-sync subsystem change.
#40763 — Draft; not Ready for Review.
#40861 — Changes Requested concerning staging, memory-pool usage, installation, and PD documentation.
#40904 — Changes global tokenizer failure semantics by swallowing decode exceptions; requires standalone review.
#41060 — BLOCKED report gate: companion kernel environment migration Crash the server when error or OOM happens #514 remains open.
#41068 — Current main reshapes partial RoPE inputs to 3D; pinned kernel Rope.cpp:427-428 requires cache width == full head width. Dtype-only patch does not fix that contract; needs source-author correction.
#41088 — Already carried by October 5 consolidation Xpu/weekly simple model enablement 2026 10 05 #42627.
#41127 — Draft; not Ready for Review.
#41142 — Already carried by October 5 consolidation Xpu/weekly simple model enablement 2026 10 05 #42627.
#41213 — Unresolved Changes Requested by @mingfeima.
#41262 — UNRESOLVED gate; “DO NOT MERGE” PR replaces the released wheel with kernel main.
#41401 — Changes Requested concerning CUDA/HIP effects, import safety, and duplicate validation.
#41523 — Already carried in October 5 consolidation Xpu/weekly simple model enablement 2026 10 05 #42627; current source also has Changes Requested.
#41532 — Changes Requested by @mingfeima; availability of kernel Don't get API response when sending images #357 does not clear review.
#41730 — Already carried by October 5 consolidation Xpu/weekly simple model enablement 2026 10 05 #42627.
#41773 — Already carried by October 5 consolidation Xpu/weekly simple model enablement 2026 10 05 #42627.
#41842 — Source-related lint failure: trailing whitespace in
pool_host/mamba.py.#41897 — Draft; not Ready for Review.
#41942 — Changes Requested concerning environment-variable registration and import-test gating.
#41943 — FLUX projection performance redesign; outside simple enablement/backend parity.
#41944 — Regression test modifies a file absent from main and depends on excluded FLUX norm-fusion work.
#42216 — Source-related lint failure:
ruff-formatreformats the changed tokenizer code.#42527 — Draft; not Ready for Review.
#42597 — Already carried by October 5 consolidation Xpu/weekly simple model enablement 2026 10 05 #42627.
#42627 — Earlier consolidation; not an eligible source patch.
#42670 — BLOCKED: kernel CUDA error: device-side assert triggered in self.forward_extend_multi_modal(batch) #528 remains open and unreleased.
#42677 — BLOCKED: kernel Access values created within a fork #529 remains open and unreleased.
#42724 — Changes Requested: reasoning/content handling changes model behavior incorrectly.
#42725 — Changes Requested by @mingfeima and @siju-samuel concerning allocator/PD behavior.
#42814 — Draft; not Ready for Review.
#42831 — Newly admitted XPU cache/page-size behavior lacks contained-kernel proof; author’s installed wheel is unspecified.
#43059 — New direct oneCCL communicator and build integration; broad runtime subsystem.
#43060 — New symmetric-memory all-reduce implementation; broad distributed-runtime enablement.
#43279 — Draft; not Ready for Review.
#43288 — Draft; not Ready for Review.
#43299 — Changes PD event ownership/ordering across transports; related AMD disaggregation failure remains unclassified.
#43312 — BLOCKED: required ngram kernel ImportError: cannot import name 'pin_program' #532 remains open and unreleased.
#43443 — Draft; not Ready for Review.
#43660 — Draft; not Ready for Review.
#43695 — General diffusion forecasting/memory algorithm change; outside XPU parity.
#43750 — Unresolved Changes Requested by @mingfeima.
#43818 — General scheduler page-count optimization; outside XPU enablement/backend parity.
#43989 — Toolchain-image update outside scope; kernel M2 Mac Attempted Installation: LLVM ERROR: Option 'pbqp' already exists! #537 is neither contained nor backported in
v0.4.0.#43994 — Broad FP8 MoE dispatch/quantization integration; standalone implementation.
#44017 — Draft; not Ready for Review.
#44231 — Draft; not Ready for Review.
#44381 — Draft; not Ready for Review.
#44411 — Draft; not Ready for Review.
#44428 — BLOCKED: required 512-expert top-k kernel Trouble Shooting #548 remains open.
#44452 — Introduces a generic ASR adapter and changes fallback processor policy; exceeds simple parity.
#44473 — Combines weight-tying, ASR-prompt, and mRoPE behavior changes; standalone correctness review.
#44625 — Draft; not Ready for Review.
#44769 — BLOCKED: LoRA kernel Multi-node Tensor Parallelism #550 remains open; new backend also exceeds simple scope.
#44852 — New frozen-KV capture/pool-ownership implementation; standalone runtime work stacked on [Feat] gemma4: Support encoder-free text serving #44851.
#44853 — Changes MTP proposal/ranking semantics; depends on standalone [Perf] [Intel] gemma4: Enable frozen-KV MTP graph capture #44852.
#44854 — Depends on excluded [Perf] gemma4: Skip ranking for full MTP chains #44853/[Perf] [Intel] gemma4: Enable frozen-KV MTP graph capture #44852; intended behavior cannot be replayed independently.
#44869 — New fused normalization/RoPE numerical path in a six-PR stack; standalone kernel/performance review.
#44884 — New logits-reduction kernel and deferred-logits runtime path; stacked on [Perf] [Intel] gemma4: Fuse QKV normalization and RoPE #44869.
#44890 — New comparison-table numerical reduction; stacked on excluded [Perf] [Intel] gemma4: Fuse greedy verification logits finalization #44884.
#44905 — BLOCKED: native BF16 activation path depends on draft, unreleased kernel Clean up logits processor #558.
#44908 — Draft; not Ready for Review.
#44910 — Draft; not Ready for Review.
#44913 — Shared-expert architecture refactor across 61 files.
#44919 — Draft; not Ready for Review.
#44936 — Draft; not Ready for Review.
#44958 — Cross-platform fused-op dispatch architecture refactor; outside simple XPU parity.
#44960 — Cross-platform graph-runner/compiler factory refactor; outside simple XPU parity.
#45033 — Changes Requested by @mingfeima; combines stop-regex semantics, capture changes and disabling DSpark coverage.
#45068 — Draft; not Ready for Review.
For #43989’s VERIFY CONTENT gate, kernel #537 merge
c63d14b414bdff1ba505f7745f2c5f592c764dc0is ten commits beyond R. At exact R,Dockerfile.xpu_kernelretains oneAPI 2026.1.2 andsrc/sycl/Device.cppretains the old__SYCL_COMPILER_VERSIONguard; the required compiler-guard backport is absent.For #41068, current main’s
rotary_embedding/base.py:574-586reshapes partial-RoPE inputs to 3D at the full head width. Exact R’ssrc/sycl/Rope.cpp:427-428requires cache width == that head width. The source’s dtype-only StableLM change cannot satisfy partial-RoPE dimensions on this base. It was removed before publication; no behavioral fix was invented.Validation
Rebased onto
origin/mainata3d00eff58b6c22ea7539f882cc10c537634200band replayed the four selected sources in ascending PR order.git diff --check origin/main...HEAD— passed.pre-commit run --files <13 changed files>— all applicable hooks passed.python -m pytest -q test/registered/unit/models/test_qwen4_exp_decode_comm.py— 2 passed.python -m pytest -q test/registered/unit/kernels/test_kernels_namespace.py— 88 passed.python -m pytest -q test/registered/kernels/ops/embeddings/test_qwen4_ple_offload.py— 14 skipped because this validation environment has CPU-only Torch.XPU/CUDA hardware tests for top-k, Qwen4 PLE offload and Gemma4 serving were not run locally. Review the accelerator CI results and run focused hardware tests before merge.
CI States
Latest PR Test (Base): ❌ Run #39664781488
Latest PR Test (Extra): ❌ Run #39664781270
Latest PR Test (AMD ROCm 10): ❌ Run #39664781200