Research Window: 2026-10-07 10:00 ~ 2026-10-08 10:00 (24 hours; first issue after the National Day holiday) Sources: GitHub (org: flagos-ai, full check of pushed_at across 54 repos, 5 repos with pushes in-window, full verification of 77 commit search hits, item-by-item comparison of the last 40 PRs and org event stream, atomic verification of tags and Releases across 6 repos), Google News RSS (24 sets of Chinese and English query terms, via proxy), HN Algolia, BAAI Community and official CSDN column verification (see appendix source checklist for details)

Index for This Issue

  • Today’s Highlights: FlagQuantum lands 23 merges in-window — distributed MPS training promoted to release certification (10-07 ~ 10-08)
  • I. Open-Source Project Progress (GitHub Activity)
    • 1.1 FlagQuantum: Distributed MPS release certification lands — two-node training promoted to release_certified (10-08)
    • 1.2 FlagQuantum: statevector and kernel dispatch performance batch — SV-010 / 011 / 014 and SWAP fusion (10-07)
    • 1.3 FlagQuantum: Gradients and geometric quantities — jacobian / jvp / vjp, Fubini-Study metric and second-order derivatives (10-07)
    • 1.4 FlagQuantum: QEC and native entropy export — sliding-window decoder and von Neumann entropy (10-08)
    • 1.5 FlagQuantum: Evidence and CI wrap-up — dispatch revision backfill and concurrency defect fixes (10-07 ~ 10-08)
    • 1.6 Post-holiday morning: FlagTree / vllm-plugin-FL / docs / build-infra new PRs and merges (10-08)
    • 1.7 Other activity: FlagGems-sglang competition kernels and FlagGems-Experimental KernelGen batch (10-08)
  • II. News Coverage and Ecosystem
    • 2.1 Component-level search remains quiet — 3 hits out of 24 query sets, all excluded (10-07 ~ 10-08)
    • 2.2 Official channel verification: CSDN column and BAAI Community (10-08)
  • III. Member Company Deep Dive
    • 3.1 Hygon: FlagGems performance batch — bmm / baddbmm / addmm and index_copy refactor (10-07 ~ 10-08)
    • 3.2 Kunlunxin: P800 upgrade to vLLM 0.24 advances and XPU3 fix batch (10-08)
    • 3.3 DAMO XuanTie: PPU backend FlagGems baseline workflow enters FlagTree (10-08)
    • 3.4 Moore Threads: mthreads benchmark TP adjustments and tle.gpu.alloc alias (10-08)
    • 3.5 Ascend: W8A8 INT8 matrix multiplication and other highlights (10-07 ~ 10-08)
    • 3.6 Other members: Tsingmicro / Iluvatar / Enflame / Cambricon / Horizon (no new activity in-window)
  • IV. Summary and Trend Observations
  • Appendix: Source Verification Table
  • Complete Source List

Today’s Highlight: FlagQuantum Lands 23 Merges in This Window — Distributed MPS Training Promoted to Release Certified (10-07 ~ 10-08)

Date: 2026-10-07 ~ 2026-10-08　Source: FlagQuantum #587, #590, #591, #595

In the first window after the holiday, FlagQuantum (the FlagOS quantum simulation and training component) became the largest increment within the organization: 76 commits and 23 merged PRs in the window, advancing continuously from the morning of 10-07 to the morning of 10-08, covering five tracks — release certification, performance dispatch, gradients and geometric quantities, error-correction decoding, and native entropy export:

  • Release certification landed (#587, merged 10-08 07:16): The distributed MPS release campaign concluded — the two_node_mps_training capability row was promoted from production_supported to release_certified (the only capability row promoted in this window), with certification built on two signed, measured production run payloads that passed the release gate; at the same time, the unsatisfiability of the tensor network (TN) capacity precondition on the current revision was recorded with measured justification, and all output-side defects found by the inspection process were fixed (28 non-merge commits since the self-merge baseline c7bba053). This campaign continues the distributed MPS / tensor network release work opened by #535.
  • Performance dispatch batch (#576 / #581 / #582 / #583): Public statevector dispatch was extended to SV-014 (SWAP), SV-010 (diagonal), and bounded SV-011 (Pauli rotations), and bounded Pauli rotation dispatch recertification was completed — continuing the SV series’ item-by-item “implementation + evidence” approach.
  • Gradients and geometric quantities (#562 / #568 / #575 / #585 / #564 / #565): Added fq.jacobian / fq.jvp / fq.vjp (multi-output programs); second-order derivatives (Hessian) reuse first-order rules via parameter shifting; added the Fubini-Study metric tensor (central-difference implementation, with a counterexample table explaining why operator-shift rules are not reused); N3-9 finalized the 30-pair matrix of “gradient method × execution mode” into a contract and re-tested it at the gate.
  • QEC and entropy export (#590 / #591 / #592 / #589): Added a sliding-window decoder (flagquantum/qec/sliding_window.py, 772 new lines, reading the detector error model per window and decoding window by window); fq.vn_entropy promoted von Neumann entropy to a native Stable Core export (root export budget 37 → 38); the four-row QEC alignment items converged; Quafu remote submission now fails closed when credentials are missing.
  • Quality closeout (#595 / #597): #595 backfilled 16 undeclared dispatch revisions; #597 (open) fixes a CI concurrency group defect — two consecutive merges cancel the entire previous validation batch (the cancellation of #590’s 23/23 checks was an instance of this mechanism).

Assessment: FlagQuantum’s dual track of “release certification + contractualized governance” is taking shape — capabilities are documented as contract files and re-verified through gates and measured payloads, and its engineering discipline is isomorphic to the FlagOS mainline release system; the window straddles the tail of the holiday and the first day back, making this repo a development line that maintained high intensity throughout the long break.

I. Open-Source Project Progress (GitHub Activity)

1.1 FlagQuantum: Distributed MPS Release Certification Lands — Two-Node Training Promoted to release_certified (10-08)

Date: 2026-10-08　Source: FlagQuantum #587, #535

  • Certification content: two_node_mps_training was promoted from production_supported to release_certified, the only changed row in the capability table; certification is based on two signed measured_production_run measured production run payloads accepted by the release gate.
  • Capacity record: The tensor-network capacity precondition is recorded as “unsatisfiable on the current revision” with measured justification — making the capacity boundary explicitly documented rather than left as an implicit constraint.
  • Engineering scope: Since merge baseline c7bba053, includes 28 non-merge commits and 2 mainline merges; all output-side defects found by the repo’s self-check process were fixed in the same pass.
  • Lineage: The campaign was opened by #535 (merged 10-06); this completes it and delivers the first distributed capability row to the repository’s release certification system (release_certified tier).

1.2 FlagQuantum: statevector and Kernel Dispatch Performance Batch — SV-010 / 011 / 014 and SWAP Fusion (10-07)

Date: 2026-10-07　Source: #576, #581, #582, #583, #566, #571, #574, #572

  • Public dispatch (SV series): #576 enables SV-014 public SWAP dispatch; #581 enables SV-010 diagonal dispatch; #582 adds bounded SV-011 Pauli rotations; #583 re-certifies the existing bounded Pauli rotation dispatch.
  • Gate batch (#566 / #571): Bounded reversible three-qubit gate dispatch (#566), dense local two-qubit matrix dispatch (#571) — local gate dispatch coverage continues to densify.
  • Kernel fusion and sampling (#574 / #572): Bounded SWAP sequence fusion (#574); MPS sampling collapse kernel dispatch (#572) — performance work now covers both the statevector and tensor-network simulation paths.
  • Implication: The performance batch advances from “in-kernel optimization” to “dispatch-layer semantics” — the same operation is routed to specialized kernels by matrix form and boundedness, with evidence recorded item by item.

1.3 FlagQuantum: Gradients and Geometric Quantities — jacobian / jvp / vjp, Fubini-Study Metric, and Second-Order Derivatives (10-07)

Date: 2026-10-07　Source: #562, #568, #564, #565, #575, #585

  • Differentiation interface surface (#562 / #568): Adds fq.jacobian / fq.jvp / fq.vjp for multi-output programs; native gradient capabilities are now centrally registered (single-point registration).
  • Contractualized measurement (#564 / #565): N3-9 codifies the 30-pair matrix of “gradient method × execution mode” as contracts/gradient-methods-contract.toml (30 unit cells + 30 noisy cells) and re-tests it in the gate; N3-11 measures every differentiable instruction × execution mode item by item.
  • Geometric quantities and second-order derivatives (#575 / #585): #575 makes second-order derivatives (parameter-shift Hessian) reuse first-order rules; #585 adds the Fubini-Study metric tensor (fq.gradients.metric_tensor) — implemented via central differences, with a counterexample table explaining why operator-shift rules are not reused (differentiating with respect to a state and with respect to an expectation value are different objects).

1.4 FlagQuantum: QEC and Native Entropy Export — Sliding-Window Decoder and von Neumann Entropy (10-08)

Date: 2026-10-08　Source: #590, #591, #592, #589

  • Sliding-window decoder (#590, merged 09:06): Adds flagquantum/qec/sliding_window.py (772 lines) — builds the decoding graph by window, decodes the detection event stream window by window, commits corrections, and advances by a fixed stride; also unblocks three block-level projections previously reserved for it.
  • Native entropy export (#591, merged 09:06): fq.vn_entropy becomes a native Stable Core export (root export budget 37 → 38) — fq.run(circuit, outputs=fq.vn_entropy([0])) returns the entropy value directly and supports log_base; the numerical implementation lives in simulation/entropy.py, with the runtime responsible only for dispatch.
  • Alignment convergence and fail-closed (#592 / #589): The four QEC alignment items converge on their respective targets (#592); Quafu remote submission switches to fail-closed when credentials are missing (#589, the earliest merge in the window).

1.5 FlagQuantum: Evidence and CI Closeout — Dispatch Revision Backfill and Concurrency Defect Fix (10-07 ~ 10-08)

Date: 2026-10-07 ~ 2026-10-08　Source: #595, #593, #573, #597, #579

  • Evidence backfill (#595): Declares 16 previously undeclared dispatch revisions in one pass, fixing the quality census accounting.
  • Docs and benchmarks (#593 / #573): Records CircuitIR.global_phase approval and version migration cost (#593); benchmark anchors now disclose “which channel produced this reading” and read via the certified channel (#573).
  • CI concurrency defect (#597, open): Identifies the “green PR merges into red main” mechanism — ci.yml / pre-commit.yaml use github.ref as the concurrency group with cancel-in-progress, so on two consecutive merges the earlier validation is cancelled wholesale (this is why #590’s 23/23 checks were cancelled); the fix reads by revision and enforces it.
  • Cleanup: Another 13 stale PRs superseded by later merges were closed in a batch on the afternoon of 10-07, along with 16 post-merge branch deletions.
  • Vocabulary migration: The qubit vocabulary migration batch (WQ-4 runtime segment, WQ-5 algorithm segment) advanced with this window, accompanied by api-changes records.

1.6 Back-to-Work Morning: New PRs and Merges in FlagTree / vllm-plugin-FL / docs / build-infra (10-08)

Date: 2026-10-08　Source: FlagTree #1324, #1325, #1316, vllm-plugin-FL #559, #569, #562, #561, docs #543, build-infra #1322

  • vllm-plugin-FL #559 (merged 09:44): CI self-healing upgrade — /rerun-failed-ci precisely reruns failed/cancelled jobs (successful jobs are no longer rerun); Enflame and MUSA images pre-bake the plugin into the image, so it runs without a source checkout.
  • vllm-plugin-FL morning opens and updates: #569 enables non-eager PIECEWISE graph mode for the Sunrise PTPU backend and fixes a GDN core-attn-out buffer reuse zeroing issue (aligning with the precision hazard in vLLM upstream #28182); #562 adds GPU container test support on the k8s runner pool; #561 (Kunlunxin P800 upgrade to vLLM 0.24.0) remains under review, with new updates on the morning of 10-08.
  • FlagTree morning opens (2): #1324 adds FlagGems baseline and test workflows for ppu3.6 (see 3.3); #1325 adjusts the tensor-parallel-size of the mthreads vllm benchmark to 2 (see 3.4); #1316 advances the tle.gpu.alloc alias to the mthreads and hcu backends.
  • docs / build-infra: docs #543 is a bot-submitted HuggingFace model list update (opened 08:01); build-infra #1322 is a dependabot version bump for the download-artifact action.
  • Additional note: All in-window pushes to FlagTree / docs / build-infra consist of newly opened PR branches; the main branches of all three repositories had no commits in the window.

1.7 Other Activity: FlagGems-sglang Competition Kernel and FlagGems-Experimental KernelGen Batch (10-08)

Date: 2026-10-08　Source: FlagGems-sglang #134, FlagGems-Experimental #694, #661, #647

  • FlagGems-sglang #134: The add3 Triton kernel submission for FlagOS Competition Track1, with review advancing on the morning of 10-08.
  • FlagGems-Experimental: Review of KernelGen’s NVIDIA operator PRs continues: 7 open PRs including slow_conv_dilated2d, sparse_mask, and dstack remain under review (covering convolution, sparse, and stacking scenarios); the dependabot dependency batch (#724 ~ #729) is under re-review.
  • FlagGems main repo: The window was dominated by open-PR review (see 3.1 / 3.2 / 3.5), with no merges.

II. News Coverage and Ecosystem

2.1 Component-Level Search Remains Quiet — 3 Hits Out of 24 Query Sets, All Filtered Out (10-07 ~ 10-08)

Date: 2026-10-07 ~ 2026-10-08　Source: Google News RSS (24 Chinese and English query terms, via proxy), HN Algolia

  • Component-term searches (FlagOS / FlagGems / FlagScale / FlagTree / FlagPerf / FlagCX / KernelGen / FlagQuantum, etc.) returned zero hits in the window, continuing the quiet streak (the earlier 09-30 daily report marked the 23rd consecutive quiet window).
  • Ecosystem-term searches returned 3 hits, all filtered out by topic: AMD meeting with South Korean chip companies (unrelated to FlagOS); an Enflame stock-price piece (a market piece, filtered out per established criteria); a BAAI Community repost introducing Microsoft Research’s Agent Lightning v1.0 (community-side content, not a FlagOS topic).
  • Six HN Algolia queries returned no direct hits (fuzzy matches were all unrelated topics).

2.2 Official Channel Check: CSDN Column and BAAI Community (10-08)

Date: 2026-10-08　Source: flagos.csdn.net, BAAI Community

  • Official CSDN column: No new articles published in the window (page metadata updated 10-08, content dates no later than 09-24).
  • BAAI Community: No FlagOS-themed entries on the homepage; the day’s additions were routine community content.

III. Deep Dive into Member Organizations

3.1 Hygon: FlagGems Performance Batch — bmm / baddbmm / addmm and index_copy Refactor (10-07 ~ 10-08)

Date: 2026-10-07 ~ 2026-10-08　Source: FlagGems #6865, #6866, #6696, #6714

  • Triple performance implementation (#6865 / #6866 / #6696): bmm (batched matrix multiply) adds a new Hygon-optimized backend implementation; baddbmm adds an optimized implementation; addmm is tuned with the FlagTune cost model — all three continuously updated from the evening of 10-07 to the morning of 10-08.
  • Kernel refactor (#6714): KMCompiler’s index_copy is refactored to a “clone without generic copy_ dispatch” path, removing one layer of dispatch overhead.
  • Implication: The Hygon line extends performance coverage using its established working pattern of “multi-platform parallel implementations of the same operator”; all four PRs in the window are densely updated open PRs.

3.2 Kunlunxin: P800 Upgrade to vLLM 0.24 Advances and XPU3 Fix Batch (10-08)

Date: 2026-10-08　Source: vllm-plugin-FL #561, FlagGems #6867

  • Inference plugin: #561 “kunlunxin upgrade P800 to vLLM 0.24.0” updated on the morning of 10-08 (entered review around 09-30, continuing to advance this round).
  • Operator fixes (#6867): XPU3 backend SDNN 46-item marked fix batch (rnn_relu, cholesky_solve, matmuladd, etc.), newly opened in the early hours of 10-08.

3.3 DAMO XuanTie: PPU Backend FlagGems Baseline Workflow Enters FlagTree (10-08)

Date: 2026-10-08　Source: FlagTree #1324, FlagTree PPU Manual

  • New workflows: ppu3.6-gems-baseline.yml and ppu3.6-gems-test.yml — generate FlagGems baselines and run FlagGems tests for the DAMO XuanTie PPU (ppu3.6) backend (newly opened on the morning of 10-08).
  • Background: PPU has entered the FlagTree 0.7.0 platform variant system (0.7.0+ppu3.6) and application matrix; this brings its FlagGems baseline into CI — a routine pairing of “compiler platform tag + operator library baseline.”

3.4 Moore Threads: mthreads Benchmark TP Adjustment and tle.gpu.alloc Alias (10-08)

Date: 2026-10-08　Source: FlagTree #1325, #1316

  • Benchmark adjustment (#1325): mthreads vllm benchmark tensor-parallel-size adjusted to 2 (newly opened on the morning of 10-08).
  • Alias support (#1316): tle.gpu.alloc kernel-level alias support extended to the mthreads and hcu backends (updated on the morning of 10-08).

3.5 Ascend: W8A8 INT8 Matrix Multiply and Other Highlights (10-07 ~ 10-08)

Date: 2026-10-07 ~ 2026-10-08　Source: FlagGems #6060

  • Quantized kernel: #6060 adds a W8A8 INT8 matrix multiply implementation for the Ascend backend, updated on the evening of 10-07, review advancing.
  • Other: The Ascend line in this window is mainly open PR reviews, with no merges within the window.

3.6 Other Members: Tsingmicro / Iluvatar / Enflame / Cambricon / Horizon (No New Activity in Window)

Date: 2026-10-07 ~ 2026-10-08

  • Tsingmicro, Iluvatar, Cambricon, Horizon: no new public activity in the window.
  • Enflame: only seen in vllm-plugin-FL #559’s Enflame image pre-baking (build infrastructure side, see 1.6).
  • BAAI (lead organization): no new public activity in the window (no new PRs in the community repo, release governance quiet).

IV. Summary and Trend Observations

  1. First post-holiday day restarts from the code side, with FlagQuantum carrying the bulk of incremental activity in a single repo — the organization merged 24 PRs in the window (23 from FlagQuantum + 1 from vllm-plugin-FL) and 77 commits, with FlagQuantum advancing continuously across the holiday tail and the return-to-work morning; other repos restarted with morning-opened PRs (FlagTree 2, vllm-plugin-FL 2, docs 1, build-infra 1).
  2. FlagQuantum’s dual track of “release certification + contract governance” takes shape — two_node_mps_training is promoted to release_certified, with certification built on signed measured payloads; root export budget 37 → 38, 30 “gradient method × execution mode” contract pairs, metric tensor counterexample tables, etc. — capabilities documented as contract files and re-verified via gates.
  3. Performance work advances from “kernel optimization” to “dispatch-layer semantics” — the SV series public dispatch extends to SV-010 / 011 / 014, with MPS sampling collapse, SWAP fusion, local two-bit matrix and other kernels entering the same evidence system.
  4. Multi-chip adaptation shows quiet movement — Hygon (bmm / baddbmm / addmm / index_copy), Kunlunxin (P800 upgrade and XPU3 fixes), DAMO XuanTie (PPU baseline workflow), Moore Threads (benchmark and alias), Ascend (W8A8 INT8) each have batches; new partner Sunrise’s PTPU graph mode wiring enters vllm-plugin-FL (#569).
  5. Build and quality infrastructure fixes in parallel — vllm-plugin-FL CI self-healing mechanism (#559: precise re-runs of failed/cancelled jobs, vendor image pre-baking) and FlagQuantum CI concurrency group defect (#597 pending merge) appear together, giving the first return-to-work day an “infrastructure day” character.
  6. Next observation points: FlagQuantum #597 merge and the follow-up to the distributed release campaign (#535 → #587 line); review outcomes of vllm-plugin-FL #561 / #569; FlagTree PPU / hcu workflow merges; pace recovery across organization lines in the first full post-holiday work week.

Appendix: Source Verification Table

Category Source Verification Method Result
GitHub org: flagos-ai repos API Full verification of pushed_at across 54 repos 5 repos had pushes within the window (FlagTree / FlagQuantum / vllm-plugin-FL / docs / build-infra), plus FlagGems-Experimental near the window edge (10-07 06:56); no new repos
GitHub Commit search (committer-date basis) Item-by-item verification of all 77 commits FlagQuantum 76 / vllm-plugin-FL 1; FlagQuantum includes 2 scheduled batches with timestamps labeled up to 10-09 (the repo’s latest push was 10-08 09:46)
GitHub PR and org event stream Item-by-item comparison of the status of the last 40 PRs + 98 org events 24 merged within the window (FlagQuantum 23 / vllm-plugin-FL 1); additionally 13 residual PRs closed in a batch and 16 branch deletions
Releases Dual check of tags / Release atoms and API 6 repos’ atoms + 5 repos’ APIs No new tags / new Releases within the window (latest is the 09-28 batch)
News Google News RSS (24 sets of Chinese and English query terms, via proxy) 24-hour window filtering + item-by-item elimination 3 hits, all eliminated; the component line remains quiet
Community HN Algolia Review of 6 query sets No direct hits
Official channels CSDN column / BAAI Community Page-scrape review No new articles within the window on CSDN; no FlagOS entries seen on the BAAI Community homepage
Historical comparison 09-30 daily report (last issue before the holiday) Deduplication verification No historical duplicates among items within the window; 10/01 ~ 10/07 was the National Day holiday

Complete Source List