Skip to content

[Feat] xpu: Consolidate simple XPU enablements - #45201

Closed
airMeng wants to merge 4 commits into
sgl-project:mainfrom
airMeng:xpu/weekly-simple-model-enablement-2026-10-19
Closed

airMeng wants to merge 4 commits into
sgl-project:mainfrom
airMeng:xpu/weekly-simple-model-enablement-2026-10-19

Conversation

@airMeng

@airMeng airMeng commented Oct 19, 2026 •

Copy link
Copy Markdown
Collaborator

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+xpu from current python/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.

  • #32136 by @cyxlily, source head 5d3455e1261a6aeffcfb6f9c59768b9e564eb75a — Focused XPU top-k transform coverage. Kernel openai chat speculative execution #248 (3cfe90204f9367820c7c2230dfae72194097fd76) is an ancestor of release R. Inspected python/sgl_kernel/top_k.py, src/sycl/TopKTransform.cpp, operator registrations and tests/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.
  • #44561 by @hotschmoe, source head 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.
  • #44851 by @alexnails, source head 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.
  • #44863 by @alexnails, source head 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 to airMeng; nothing is merged.

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:

For #43989’s VERIFY CONTENT gate, kernel #537 merge c63d14b414bdff1ba505f7745f2c5f592c764dc0 is ten commits beyond R. At exact R, Dockerfile.xpu_kernel retains oneAPI 2026.1.2 and src/sycl/Device.cpp retains the old __SYCL_COMPILER_VERSION guard; the required compiler-guard backport is absent.

For #41068, current main’s rotary_embedding/base.py:574-586 reshapes partial-RoPE inputs to 3D at the full head width. Exact R’s src/sycl/Rope.cpp:427-428 requires 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/main at a3d00eff58b6c22ea7539f882cc10c537634200b and 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 AST parsing — all 11 changed Python files 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

@github-actions github-actions Bot added documentation Improvements or additions to documentation Multi-modal multi-modal language model jit-kernel labels Oct 19, 2026
@airMeng
airMeng marked this pull request as ready for review October 19, 2026 18:06
@airMeng

airMeng commented Oct 19, 2026

Copy link
Copy Markdown
Collaborator Author

/tag-and-run-ci

@github-actions

This comment has been minimized.

@github-actions

Copy link
Copy Markdown
Contributor

⛔ /tag-and-run-ci isn't a recognized CI command. Did you mean /tag-and-rerun-ci?

See the command reference.

@airMeng airMeng changed the title [XPU Weekly] Simple enablement consolidation — 2026-10-19 [Feat] xpu: Consolidate simple XPU enablements Oct 19, 2026
@airMeng

airMeng commented Oct 19, 2026

Copy link
Copy Markdown
Collaborator Author

/tag-and-rerun-ci

cyxlily and others added 4 commits October 21, 2026 12:43
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.
@airMeng
airMeng force-pushed the xpu/weekly-simple-model-enablement-2026-10-19 branch from c6f612d to 49f6b6c Compare October 21, 2026 19:48
@mintlify

mintlify Bot commented Oct 21, 2026

Copy link
Copy Markdown
Contributor

Preview approval required

This pull request comes from a fork, so Mintlify won't build a preview until someone with write access approves commit 49f6b6c.

Approve preview

@desmond-intel

Copy link
Copy Markdown
Contributor

@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?

@airMeng

airMeng commented Oct 26, 2026

Copy link
Copy Markdown
Collaborator Author

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.

@airMeng airMeng closed this Oct 26, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

documentation Improvements or additions to documentation jit-kernel Multi-modal multi-modal language model run-ci CI: run the baseline test suite on this PR

Projects

None yet

Development

Successfully merging this pull request may close these issues.

5 participants