Research window: 2026-11-12 06:00 ~ 2026-11-13 06:00 (Beijing time, 24 hours; continuing from the cutoff of the previous daily report).

Sources: GitHub (full audit of 64 repos under org: flagos-ai, 193 in-window commits, 167 merged PRs, 272 in-window updated open PRs, 19 repos with pushes, atomic audit of Releases across 31 repos and Tags across 8 repos), Google News RSS (24 sets of Chinese and English query terms, via proxy), HN Algolia, CSDN FlagOS column, BAAI Community (see appendix source list for details).


This Issue’s Index

  • Today’s Highlights: 2.3.0 enters final consolidation, version data filled in per backend, FlagTree 0.7.1 official release lands first on Ascend and Kunlunxin, vLLM 0.32 adaptation queue adds ten vendors (11-12)
  • I. Open Source Project Progress (GitHub Activity)
    • 1.1 Organizational cadence: 19 repos pushed, 193 commits, 167 merges, 272 under review (11-12)
    • 1.2 build-infra: 2.3.0 docs completed backend by backend, CANN 9.2.1 adds 910C, Iluvatar corex5.0.1 replaces 4.4.0 (11-12)
    • 1.3 FlagTree: 0.7.1 official release lands on two backends first, Intellifusion backend enters Triton 3.8 (11-12)
    • 1.4 FlagGems: 68 commits, Hygon foreach rolls out 20 operators in batch, Kunlunxin wraps up with 5 commits (11-12)
    • 1.5 FlagGems-vllm: 20 commits, SiliconFlow KDA/GDN batch and T-Head Marlin MoE (11-12)
    • 1.6 vllm-plugin-FL: vLLM 0.32 adaptation lands at two points, NVIDIA KV cache and T-Head PPU both enter mainline (11-12)
    • 1.7 FlagQuantum: 70 commits, 29 merges, fermion simulation fills in three model types in one day (11-12)
    • 1.8 Frameworks and communication: FlagCX Enflame/Iluvatar backend extensions under review, Torch-FL fixes MUSA and Iluvatar routing (11-12)
    • 1.9 Ecosystem and education: OpenCourse adds FlagCompressor quantization lab, 9 competition-track operators enter review in early morning (11-13)
  • II. News Coverage and Ecosystem
    • 2.1 Channel verification: 24 query groups yield 38 candidates, zero hits for FlagOS and component-name queries (11-12)
    • 2.2 Ecosystem background: Moore Threads MDC 2026 set for December 29-30, T-Head advances independent listing (11-11/11-12)
  • III. Member Units and Backend Vendors In Depth
    • 3.1 Hygon: 27 commits in a single day, foreach 20 operators and bmm trio enter together (11-12)
    • 3.2 Iluvatar CoreX: corex5.0.1 fully replaces 4.4.0, CI advances on three fronts (11-12)
    • 3.3 Kunlunxin: 5 XPU fixes land, operator coverage continues to fill gaps (11-12)
    • 3.4 Ascend: CANN 9.2.1 adds 910C base, INT8 GDN enters FlagGems-vllm (11-12)
    • 3.5 T-Head: PPU adaptation for vLLM 0.32 lands, Marlin MoE and Log-Linear Attention advance in the same week (11-12)
    • 3.6 MetaX, Moore Threads, Enflame, Tsingmicro, Intellifusion: incremental progress and backend expansion (11-12)
  • IV. Summary and Trend Observations
  • Appendix: Source Verification Table and Complete Source List

Today’s Focus: 2.3.0 Enters Finalization, Version Data Filled In Backend by Backend (11-12)

Date: 2026-11-12 Sources: build-infra #1742 (PPU base 2.3.0 version data) · build-infra #1769 (Tianshu runtime 2.3.0) · FlagTree 0.7.1+ascend3.5 · FlagTree 0.7.1+xpu3.6 · vllm-plugin-FL #706

  • Version side: FlagTree has moved from 0.7.1a2 to the official release, with two backends cleared first—0.7.1+ascend3.5 (15:43) and 0.7.1+xpu3.6 (19:28). The Release notes contain only one line, “pypi 0.7.1+", indicating per-backend wheel delivery. The remaining backends are still at 0.7.1a1 (hcu, tsingmicro), a2 (iluvatar, mthreads), or dev1 (ppu). The official release is being cleared backend by backend rather than all at once.
  • Build side: build-infra adds two pieces of version data on 2.3.0—the base documentation for T-Head PPU 2.1.0 (#1742, where the tool pulls versions for 23 backends and all 23 checks pass) and the runtime documentation for Tianshu corex5.0.1 (#1769, 1 backend). The 2.3.0 image line now covers both base and runtime sides, but records are still advanced “one backend per PR”.
  • Hardware layer: Ascend CANN 9.2.1 adds the base image for 910C (#1760; the 910B version was already available on 11-03), and adds an ascend-cann9.2.1-910c row to the flagcx packaging matrix (#1761). This row is disabled for all channels because 910C and 910B share the single pin flagcx 0.14.1+cann9.2.1; enabling it would produce duplicate wheels. Without the row, the matrix would drift, causing set-matrix to fail for all channels and preventing flagcx wheels from being dispatched.
  • Framework side: vLLM 0.32 adaptation has grown from a single point into a queue—this period merges two items, the NVIDIA KV cache migration (#737) and T-Head PPU-ZW810E (#706), with eight more vendors under review: MUSA, Hygon, MetaX, Kunlunxin, Ascend, Tianshu, Enflame, and AMD.
  • Implication: The two hard milestones of 2.3 (FlagTree branch cut on December 7, release on December 30) now have observable intermediate states: 0.7.1 cleared by backend, 2.3.0 image documentation filled in by backend, and vLLM 0.32 finalized by vendor. None of these is a one-shot action; the daily report’s unit of observation must shift from “version number” down to “backend × component”.

I. Open-Source Project Progress (GitHub Activity)

1.1 Organization cadence: 19 repos pushed, 193 commits, 167 merged, 272 under review (11-12)

Date: 2026-11-12 Source: flagos-ai organization repositories · Merged PR search within window · FlagGems tags · FlagTree tags

  • Totals: 19 of 64 repos had pushes within the window; 193 new commits on default branches, 167 merged PRs, and 272 open PRs updated within the window, of which 120 were newly opened.
  • Versus previous period: commits rose from 148 to 193 (+30.4%), merges from 148 to 167 (+12.8%), pushed repos fell from 20 to 19, and open PRs fell from 317 to 272 (-14.2%). The commit rebound came mainly from FlagQuantum and build-infra, while the review backlog declined for a second consecutive day.
  • Merge distribution: FlagGems 68 (40.7%), FlagQuantum 29 (17.4%), build-infra 28 (16.8%), FlagGems-vllm 20 (12.0%), FlagTree 7; the remaining 10 repos totaled 15. FlagGems’ share fell from 57.4% on 11-11 to 40.7%, the first time in nearly a week it dropped below half.
  • Commit distribution: FlagQuantum 70, FlagGems 56, build-infra 28, FlagGems-vllm 20, FlagTree 6, FlagCX and Torch-FL 3 each, FlagSparse and OpenCourse 2 each, and FlagGems-sglang, optest, and FlagScale 1 each.
  • Releases: two new tags in the window, both FlagTree GA releases (0.7.1+ascend3.5, 0.7.1+xpu3.6); FlagGems v5.5.0.dev1 (11-11 18:25) and vllm-plugin-FL v0.5.0.dev0 (11-11 14:23) were already covered in the previous report, with no new changes this period.
  • Pushed but no default-branch commits: seven repos — vllm-plugin-FL, docs, release-info, FlagFFT, Trident, Megatron-LM-FL, and TransformerEngine-FL — had pushed_at within the window but no commits on their default branches, reflecting branch- or tag-side activity and excluded from this period’s commit count; the in-window commits for Megatron-LM-FL and TransformerEngine-FL were caught by the Atom fallback and are included in the merge statistics.

1.2 build-infra: 2.3.0 docs filled in per backend, CANN 9.2.1 adds 910C, Iluvatar corex5.0.1 replaces 4.4.0 (11-12)

Date: 2026-11-12 Source: build-infra #1742 · build-infra #1747 · build-infra #1760 · build-infra #1745 · build-infra repo

  • Scale: 28 merges, all on the 2.3.0 image line; 12 were changelog date backfills by the github-actions bot and 16 were manual changes. tengqm alone authored 16, making him the main executor of this release close-out.
  • 2.3.0 release data: #1742 records base docs for T-Head PPU 2.1.0, with the tool pulling built-in system package versions across a 23-backend matrix and all 23 checks passing; #1769 records runtime docs for Iluvatar corex5.0.1. Both follow the established convention of “one tag per PR, version data and changelog landed together.”
  • Iluvatar generation change: #1746 switches the Iluvatar base image from 4.4.0 to 5.0.1, #1747 adds the corex5.0.1 backend and retires the corex4.4.0 build layer, #1748 and #1753 add base/runtime changelogs, #1755 installs driver user-space components, #1762 adds libopenmpi3t64, and #1766 pins flagcx to 0.14.1. Iluvatar accounts for 16 merges, the largest single-vendor batch of the day.
  • Ascend 910C: #1760 adds the ascend-cann9.2.1-910c base; #1761 adds the flagcx packaging line, disabled for all channels to avoid producing a second wheel when sharing a pin with 910B, and also fixes the drift where a runtime matrix row existed without a packaging row.
  • MetaX regression fix: #1745 pre-provisions a MACA SDK slice for FlagTree’s metax builder (include closure plus six libraries, along with libnuma1/libelf1). The cause: FlagTree 0.7.1 enables FlagPrism by default for metax per #1496, but the 22.04 builder lacks the MACA SDK, so the REQUIRED check fails at configure time and metax daily has failed every night since #1496 — this fix addresses the side effect of “FlagPrism on by default.”
  • Hygon: #1743 adds torchaudio to the dtk26.04 runtime and #1744 backfills the changelog; these are closure completions rather than a generation change.

1.3 FlagTree: 0.7.1 GA ships two backends first, Intellifusion backend lands in Triton 3.8 (11-12)

Date: 2026-11-12 Source: FlagTree #1598 (Intellifusion backend) · FlagTree #1637 (Ascend Triton 3.6.x, under review) · FlagTree 0.7.1+ascend3.5 · FlagTree 0.7.1+xpu3.6 · FlagTree repo

  • Version lineage: 0.7.1 pre-releases covered six a1 backends (ascend, hcu, iluvatar, mthreads, tsingmicro, xpu) and two a2 backends (iluvatar, mthreads); the GA release so far ships only ascend3.5 and xpu3.6, while T-Head remains at 0.7.1.dev1+ppu3.6. Releasing per backend under the same version number is FlagTree’s established practice.
  • New backend: merged #1598, adding a compiler and runtime backend for Intellifusion’s ifwa:100 target, including memory-access and matmul lowering, software pipelining, async copy, and LHS-to-TMEM promotion, plus IFWA Toolkit build checks, CModel examples, and a separate manual wheel delivery workflow. The backend is not yet registered in the README mainline matrix; it is an addition, not an expansion.
  • Ascend version line: #1637 is under review, porting the Ascend NPU backend to Triton 3.6.x, including the full lowering pipeline, Python runtime, and test suite. If merged, Ascend would return from the triton_v3.5.x branch to mainline, directly changing the 2.3 version matrix.
  • TLE and TileIR: #1534 converges merge analysis and preserves gather facts, #1535 indexes compatible layout proposals while preserving merge order, and #1615 adds token-aware pointer read/write and atomic operations to TileIR.
  • Backend details: #1633 optimizes sort on Iluvatar using tle_raw; #1488 syncs the XPU cluster source and bumps the SDNN object to v0.3.6.9.1; #1635 bumps the XPU FlagGems test tag from v5.3.5 to v5.5.0.dev1, aligning with the FlagGems dev version.

1.4 FlagGems: 68 merges, Hygon foreach batch of 20 operators rolled out, Kunlunxin wraps up with 5 (11-12)

Date: 2026-11-12 Source: FlagGems repo · FlagGems #8963 (Hygon bmm trio) · FlagGems #9011 (Hygon or_tensor) · FlagGems #8740 (remove 7 host-side operators) · FlagGems #5882 (int64 offsets)

  • Scale and composition: 68 merges, 43 from the KernelGen generation line — Hygon 22, NVIDIA 19, Enflame 2; 25 from the non-generation line, mainly 11 test syncs, 5 Kunlunxin fixes, 3 Ascend, and 1 cleanup, with the rest being single-point changes to indexing, convolution, and bmm.
  • Hygon concentrated rollout: 20 foreach-family operators merged in roughly 100 minutes between 17:38 and 19:21 (clamp, div, maximum, minimum, mul, add, addcdiv, addcmul, pow, round, sign, copy and their list variants), plus linalg_ldl_factor_ex and or_tensor; #8963 adds optimized backend implementations for bmm/addmm/baddbmm. Author bwbwzzz alone contributed 22, the most productive vendor line of the day.
  • Kunlunxin wrap-up: #8644 adds special-bessel and thnn recurrent backward coverage, #8647 adds adaptive_max_pool1d and view_as_complex_copy, #8724 rebuilds weight_norm backward with flat addressing, #8889 provides XPU-safe quantized_max_pool3d/quantized_max_pool3d_out vendor kernels, and #8751 fixes the precision semantics of mH on non-contiguous inputs.
  • Test drift convergence: 11 commits from lukatao align vendor-branch cases with master, covering Bessel Y0, Hermite, Chebyshev T, matrix_power, the conv series, addmm activation, depthwise convolution, and more; the same author has 60-plus similar PRs under review. This work produces no new operators but determines whether “vendor branch or mainline is correct” can be judged automatically.
  • Operator table cleanup: #8740 removes 7 host-side operators that violate repo conventions (expand_as, alias, make_dep_token, _reshape_alias, unflatten, t, _unsafe_view), reducing the experimental operator table from 142 to 141, with matching cleanup of vendor specializations, tests, benchmarks, operators.yaml, and the Chinese and English docs.
  • Other: #5882 changes large-tensor index offsets to int64, fixing overflow beyond 2^31 elements; #8539 selects optimized Moore Threads FP8 TopK using an exact-codebook histogram; #8535 optimizes Ascend LayerNorm row scheduling with TLE while preserving FP32 statistics.

1.5 FlagGems-vllm: 20 merges, SiliconFlow KDA/GDN batch and T-Head Marlin MoE (11-12)

Date: 2026-11-12 Source: FlagGems-vllm repo · FlagGems-vllm #752 (T-Head Marlin MoE) · FlagGems-vllm #1159 (Ascend INT8 GDN) · FlagGems-vllm #852 (sparse FP8 MLA)

  • Scale: 20 merges, with 47 more under review (10 newly opened this period). Merges concentrated in three places: SiliconFlow 14, T-Head 1, Ascend and NVIDIA 4 combined, and one mhc pre/post-processing change.
  • SiliconFlow batch: adds triton_ihc_post (multi-backend), GLM5Next pad_indexer_heads,
    triton_reshape_and_cache_flash, and fused_indexer_k_norm; the other 10 fix KDA/GDN benchmarks and backend tuning, including chunk_kda cold compile, scaled_dot_kkt timeouts, packed GDN decode, Ascend gate_sigmoid_mul and recurrent state access, and Hygon hc_combine layout.
  • T-Head Marlin MoE: #752 adds a unified fused_marlin_moe backend covering four weight formats: W4A16 INT4, W4A16 MXFP4, W8A16 INT8, and W8A16 FP8. The W8 path uses 64-element K tiles to reduce decode footprint and uses packed BF16 multiply-subtract instructions for BF16 INT8 tiles up to 128 columns; per-block MXFP4 scale conversions are recorded into CUDA Graph capture so in-place updates are visible on replay.
  • Ascend quantization path: #1159 adds and optimizes GDN decode and prefill for INT8 state; #1160 and #1162 fix benchmark op_name for paged_scatter_triton and indexer_gemm_score; #1183 and #1184 fix KDA cold compile and GDN strided state access.
  • NVIDIA: #852 adds sparse FP8 MLA decode (TLE route), covering sparse attention on the decode side.
  • Review queue: MetaX 13 quantization and routing optimizations (packed encoder attention, split mHC pre, bounded sparse MLA), Hygon MegaPreAttention (#1189), NVIDIA Hopper MegaMHC/MegaGate (#1193, #1196), and select_candidate_blocks_cooperative for DeepSeek-V4.1 decode (#1197).

1.6 vllm-plugin-FL: vLLM 0.32 adaptation lands in two places, NVIDIA KV cache and T-Head PPU both into mainline (11-12)

Date: 2026-11-12 Source: vllm-plugin-FL #737 (KV cache migration) · vllm-plugin-FL #706 (T-Head PPU 0.32) · vllm-plugin-FL #742 (remove 0.28 paths, under review) · vllm-plugin-FL repo

  • NVIDIA KV cache: #737 migrates KV cache initialization in model_runner.py from the private function _reshape_attention_kv_cache, removed in vLLM 0.32, to the public allocate_kv_cache and bind_kv_cache. The original import had no try/except protection, so any path importing model_runner would hit ImportError; this is the gap left by #615, which changed only CI/config and missed the runtime.
  • T-Head PPU: #706 brings PPU-ZW810E onto vLLM 0.32 (0.5.0-dev). The approach is “0.32-empty”: any symbol removed in 0.32 but still called in the branch (XD-RoPE, sequence parallelism, RoutedExpertsLists/Tensors, get_total_cp_world_size, shared_by KV allocation) is deleted along with its call sites rather than stubbed out; the worker_symbols_032.py shim introduced by #737 does not apply on this branch.
  • Queue: eight vendors under review — MUSA (#715), Hygon (#711), MetaX (#693), Kunlunxin P800 (#726), Ascend Runner V2 (#703), Iluvatar (#674), Enflame (#675), and AMD W7900D (#659); #742 further removes the remaining 0.28 paths in the V1 runner.
  • Implication: vLLM 0.32 has moved from a roadmap “main release” to a per-vendor migration checklist. NVIDIA and T-Head landed first, with the other eight queued in the same week, indicating that the 2.3 framework migration has entered execution rather than planning.

1.7 FlagQuantum: 70 commits, 29 merges, fermionic simulation fills in three model classes in one day (11-12)

Date: 2026-11-12 Source: FlagQuantum repo · FlagQuantum #1150 (Hubbard ring) · FlagQuantum #1166 (Pauli symmetry) · FlagQuantum #1145 (216 native programs) · FlagQuantum #1168 (execution plan materialization)

  • Scale: 29 merges and 70 commits, first in commits and second in merges for the day. Commits concentrated between 09:00 and 23:30, with three more added between 03:17 and 05:17, reflecting continuous work across the full day.
  • Fermionic models: #1150 adds spin-periodic Hubbard rings, #1153 assembles UCCSD training for periodic Hubbard, #1154 adds spinless t-V chains and rings, #1159 assembles native spinless chain/ring training, and #1163 trains within explicit parity sectors. All three model classes (Hubbard, t-V, molecular UCCSD) now have trainable paths within the same window.
  • Symmetry discovery: #1161 discovers diagonal fermionic parity symmetry, #1166 discovers commuting Pauli symmetry, and #1170 (under review) advances training into mixed-axis Pauli sectors. Symmetry discovery directly determines the reducible Hamiltonian size and is a key step toward practical simulators.
  • Chemistry interfaces: #1140 and #1142 import OpenFermion’s interaction operator and MolecularData, #1138 trains molecular UCCSD from native FCIDUMP input, #1139 and #1147 verify that input substitution and frozen-core projection do not break training, and #1141 uses upstream tests to qualify the in-memory molecular adapter.
  • Synthesis and runtime: #1145 adds native FlagQuantum implementations for all 216 quantum programs (five categories: Algorithms, Applications, Community, Functions, Tutorials), #1108 lets native qfunc support classical parameter binding, #1168 avoids duplicate materialization of execution plans, and #1157 (under review) adds embedded DEM sampling and error replay.
  • Assessment: Quantum simulation within FlagOS has moved from “components exist” to “model coverage + performance.” Advancing two condensed-matter model classes, parity symmetry, and native implementations of 216 programs within a single day is clearly faster than other components, but this work remains far from productization and should not be read as “release-grade” in the near term.

1.8 Frameworks and communication: FlagCX Enflame/Iluvatar backend extensions under review, Torch-FL fixes MUSA and Iluvatar routing (11-12)

Date: 2026-11-12 Source: FlagCX #712 (Enflame ECCL, under review) · FlagCX #714 (Iluvatar GDR/PXN, under review) · FlagCX #709 (Ascend Torch API tests) · Torch-FL #821 (MUSA channel) · Torch-FL #808 (Iluvatar add.Tensor routing)

  • FlagCX 3 merges: #709 enables Ascend Torch API tests and Python coverage, #697 lets UniRunner support fused collectives and DSL, and #711 fixes duplicate host-device helper definitions in the Iluvatar build (which would directly cause link failures).
  • FlagCX under review: #712 adds an ECCL device API backend for Enflame, including comm traits, platform traits, adaptor initialization, and intra cases; #714 adds a GDR device family and PXN capability hooks for Iluvatar — declaring memory registration through the IB adaptor, adding IXCUDA peer-access queries for topology discovery, and cleaning up allocations when GDR pointer setup fails.
  • Torch-FL 3 merges: #821 unblocks the MUSA channel (pip cache, MPS gate, mudnn cast), #820 replays captured FlagGems kernels from the plugin caller to amortize launch overhead, and #808 routes add.Tensor to FlagGems on Iluvatar while preserving the complex escape path.
  • Training side: Megatron-LM-FL #213 adds a CI pipeline and dev image for Iluvatar BI-V150, FlagScale #1322 brings Iluvatar BI training into the shared CI framework, and TransformerEngine-FL #139 fixes compatibility between MetaX TP overlap and older native Userbuffers.
  • Evaluation and packaging: optest #48 supports distribution metadata for FlagTree daily builds; FlagSparse #119 merges MUSA stability tests (a PR from NCIC-AlphaSparse).

1.9 Ecosystem and teaching: OpenCourse adds FlagCompressor quantization lab, 9 competition-track operators enter review before dawn (11-13)

Date: 2026-11-13 Source: OpenCourse #15 (quantization lab) · FlagGems-sglang #339 (batch11) · FlagGems-sglang #360 (Competition Track1) · FlagGems-sglang #367 (Competition Track1) · docs #583

  • Course side: OpenCourse #15 adds the FlagCompressor Quantization Lab, turning the quantization workflow into a course lab and merging it into the teaching-materials branch; #15 merged at 10:17, forming a teaching entry point alongside the quantization-related FlagOS-Compressor component.
  • Operator batch: FlagGems-sglang #339 merges batch11 operators (reference implementations, tests, and benchmarks), continuing the “merge by batch” cadence.
  • Competition channel: between 01:07 and 01:08 on 11-13, the same author opened 9 PRs in a row, all tagged [FlagOS Competition-Track1], covering top_k_renorm, grouped_gemma_rmsnorm, gemma4_fused_routing, timestep_embedding, bias_gelu_tanh, apply_rotary_emb, fused_norm_rope_inplace, fused_qk_complex_rope, wan_rmsnorm_silu, and others. This is the first time Competition Track1 submissions have entered main-repo review in batches, and merge quality is worth tracking.
  • Docs: docs #583 is a bot-driven automatic update of the ModelScope model list and does not constitute project activity; the docs repo had no manual documentation changes within the window.

II. News Coverage and Ecosystem

2.1 Channel Verification: 24 Query Sets, 38 Candidates; Zero Hits for FlagOS and Component-Name Queries (11-12)

Date: 2026-11-12 Sources: Google News RSS · HN Algolia · CSDN FlagOS Column · BAAI Community

  • gnews results: 24 Chinese and English queries deduplicated to 38 candidates. Queries for component names such as FlagOS, FlagGems, FlagScale, FlagTree, FlagPerf, FlagCX, and KernelGen all returned zero hits; FlagOS when:14d returned only 1 result, a report on the OSCAR Open Source Industry Conference whose body does not cover the FlagOS project, so it was removed as noise.
  • Hit composition: 24 for Enflame (share price up 76% in seven days, market cap above RMB 240 billion, changes in shareholder count), 1 each for MetaX and Moore Threads (MiniMax H3-Image adaptation, reported in the previous issue), 4 for BAAI (popular-science, event, and profile coverage), and 3 for domestic compute and open-source software stacks (opinion pieces). None were included as main-text entries.
  • Other channels: FlagOS/FlagGems hits on HN Algolia were all tokenization mismatches, with no related stories or comments; the latest post in the CSDN FlagOS column is dated 2026-08-28, with no new articles in the window; the BAAI Community homepage features models and popular-science content, with no FlagOS project updates.
  • Conclusion: The news side remains quiet for a consecutive period; this issue’s content is primarily sourced from GitHub. Zero hits for component-name queries have now persisted across multiple windows and are treated as normal, no longer as a “collection failure.”

2.2 Ecosystem Background: Moore Threads MDC 2026 Set for December 29-30; T-Head Pushes for Independent Listing (11-11/11-12)

Date: 2026-11-12 Sources: Shanghai Securities News (MDC 2026) · Eastmoney (MDC 2026) · 36Kr (T-Head listing)

  • Moore Threads MDC 2026: Scheduled for December 29-30 in Beijing, with more than 10 new products expected, including the S6000 intelligent compute card and the “Lushan” series based on the “Huagang” architecture as highlights; the conference will feature 26 themed forums and a 3,000-square-meter exhibition area. The announcement was released on the evening of November 11 and is an industry development outside the window.
  • T-Head listing: On 11-12, 36Kr discussed the motivations behind T-Head’s independent listing. T-Head is the PPU backend vendor for FlagOS; this matter is at the capital level, and no impact on code or image cadence was observed this issue.
  • Handling policy: Both items are recorded as ecosystem background, not counted as project updates, and not used as evidence of FlagOS progress. Industry moves by member organizations are recorded separately from project code activity, to avoid treating share prices or launch events as technical progress.

III. Deep Dive into Member Units and Backend Vendors

3.1 Hygon: 27 PRs in a Single Day, 20 foreach Operators and the bmm Trio Land Together (11-12)

Date: 2026-11-12 Sources: FlagGems #8986 · FlagGems #8963 · FlagGems #9011 · FlagGems #9007

  • Volume: 27 merges tagged with Hygon — 22 in FlagGems (20 foreach-family operators, linalg_ldl_factor_ex, or_tensor), 1 optimization for bmm/addmm/baddbmm, 1 test-scope change (#8971 restricts dropout rounding to MetaX and Hygon), 2 build-infra (torchaudio and changelog), and 1 FlagGems-vllm (hc_combine layout).
  • Time density: The 20 foreach operators merged back-to-back within 100 minutes, from 17:38 to 19:21, with author bwbwzzz submitting 22 PRs alone. This “same template, operator-by-operator rollout” submission pattern usually implies an established unified KernelGen generation and validation pipeline, where the marginal cost lies mainly in validation rather than implementation.
  • Coverage: The foreach family covers clamp, div, maximum, minimum, mul, add, addcdiv, addcmul, pow, round, sign, copy, and their list variants. foreach is a high-frequency entry point for optimizers and quantized training scripts; once filled in, Hygon should see a marked reduction in fallbacks along the torch optimizer path.
  • The bmm trio: #8963 adds optimized backend implementations for bmm, addmm, and baddbmm — the main matrix-multiplication path; together with foreach, this completes coverage of the “backbone operators of training scripts.”
  • Under review: FlashAttention dropout (#8969) and the FlashAttention/sum/clone optimizations on master (#8922/#8921/#8919) remain queued; MegaPreAttention (#1189) is on the inference side.

3.2 Iluvatar CoreX: corex5.0.1 Fully Replaces 4.4.0, CI Advances on Three Fronts (11-12)

Date: 2026-11-12 Sources: build-infra #1747 (corex5.0.1 backend) · build-infra #1746 (replace base 4.4.0) · FlagScale #1322 (BI training CI) · Megatron-LM-FL #213 (BI-V150 CI)

  • Version transition: 16 build-infra PRs make corex5.0.1 the default for both the base and runtime lines, retiring the corex4.4.0 build layer; they add the driver user-space components, libopenmpi3t64, and a pinned flagcx 0.14.1, and backfill changelog dates entry by entry. This is not an incremental patch but a wholesale move of the SDK baseline forward by one version.
  • Three CI fronts: FlagScale #1322 wires BI training into the shared CI framework, Megatron-LM-FL #213 adds the BI-V150 pipeline and dev image, and FlagCX #711 fixes duplicate host-device helper definitions in the Iluvatar build.
  • Compilation and routing: FlagTree #1633 optimizes sort with tle_raw, FlagTree #1488 bumps the SDNN object to v0.3.6.9.1; Torch-FL #808 routes add.Tensor to FlagGems while preserving the complex-number escape path.
  • Under review: If FlagCX #714’s GDR and PXN capability hooks merge, Iluvatar will gain memory registration and topology discovery on the multi-node communication side — a step from “runs on a single node” toward “scales across nodes.”
  • Assessment: This round for Iluvatar is a version transition rather than operator volume; 16 PRs concentrated on a single SDK version indicate that 2.3.0 will take corex5.0.1 as its release baseline, with subsequent operators and CI hanging off that line.

3.3 Kunlunxin: 5 XPU Fixes Land, Operator Coverage Continues to Fill Gaps (11-12)

Date: 2026-11-12 Sources: FlagGems #8644 · FlagGems #8889 · FlagGems #8724 · vllm-plugin-FL #726 (P800 0.32, under review)

  • 5 merges: #8644 adds special-bessel and thnn recurrent backward coverage, #8647 adds adaptive_max_pool1d and view_as_complex_copy, #8724 rebuilds weight_norm backward with flat addressing, #8889 provides XPU-safe quantized_max_pool3d and quantized_max_pool3d_out vendor kernels, and #8751 fixes the precision semantics of mH on non-contiguous inputs.
  • Review queue: 8 PRs including linalg_eig (#8983), linalg_polar/tensorinv/lstm (#8960), mixed_dtypes_linear (#8962), geqrf and grid_sampler (#8976), and select_copy/native_norm/special_xlogy (#9031), covering linear algebra, sampling, and pointwise paths.
  • 0.32 blocker: Yesterday vLLM 0.32.0 was pulled by build-infra due to a py3.10 import failure, with no recovery record this period; vllm-plugin-FL #726 (P800 on vLLM 0.32) is still under review, indicating platform differences remain unresolved.
  • Assessment: This period for Kunlunxin is “gap-filling” rather than “surface expansion” — each PR corresponds to a specific operator’s precision or addressing issue, and they merged in a single evening, as if clearing the failures from one round of regression testing one by one.

3.4 Ascend: CANN 9.2.1 Adds 910C Base, INT8 GDN Enters FlagGems-vllm (11-12)

Date: 2026-11-12 Sources: build-infra #1760 (CANN 9.2.1 910C) · build-infra #1761 (flagcx 910C line) · FlagGems #8964 (copy_ redispatch) · FlagGems-vllm #1159 (INT8 GDN)

  • Image layer: #1760 adds the ascend-cann9.2.1-910c base. The CANN 9.2.1 from 11-03 covered only 910B; this adds 910C. #1761 concurrently adds the flagcx packaging line and disables it for all channels, since both 910 generations share the single pin of flagcx 0.14.1+cann9.2.1.
  • Operators: #8535 optimizes LayerNorm row scheduling with TLE while preserving FP32 statistics; #8964 skips redispatch of aten::copy_ under CompositeExplicitAutograd on NPU; #8880 removes redundant environment variables and adds loop handling for large shapes.
  • Inference: #1159 adds and optimizes GDN decode and prefill for INT8 state; #1183, #1184, and #1186 fix KDA/GDN cold compilation, strided state access, and gate_sigmoid_mul on Ascend.
  • Compilation: FlagTree 0.7.1+ascend3.5 stable is out; if the Triton 3.6.x port under review in #1637 merges, Ascend will return from the 3.5.x branch to mainline — the compilation-side change most worth tracking this period.

3.5 T-Head: PPU Adaptation for vLLM 0.32 Lands, Marlin MoE and Log-Linear Attention Advance in the Same Week (11-12)

Date: 2026-11-12 Sources: vllm-plugin-FL #706 (PPU on 0.32) · FlagGems-vllm #752 (Marlin MoE) · FlagAttention #134 (Log-Linear Attention, under review) · build-infra #1751 (thead wheel guard)

  • Inference framework: #706 brings vLLM 0.32 (0.5.0-dev) to PPU-ZW810E using a “0.32-empty” adaptation: removed symbols and their call sites are deleted, with no 0.28 compatibility stubs retained. This is the key PR moving the T-Head line directly from 0.28 to 0.32.
  • Operators: #752 adds a unified fused_marlin_moe backend covering W4A16 INT4/MXFP4 and W8A16 INT8/FP8; the W8 path uses 64-element K tiles and packed BF16 instructions to reduce decode overhead, with MXFP4 scale conversion moved into CUDA Graph capture.
  • Attention: #134 is under review, adding Log-Linear Attention to the T-Head backend; 4 correctness cases pass on a single PPU-ZW810E and forward comparison is done against t-head/flash-linear-attention-for-sail. Such linear attention operators are candidate paths for long-context inference, but only one backend is currently pushing them.
  • Packaging pipeline: build-infra #1750 adds ppu and corex5.0.1 lines to flagcx and retires corex4.4.0, #1751 opens up thead in the wheel release guard, and #1768 records the ppu and corex5.0.1 lines; the flagcx wheel pipeline for PPU is now through.
  • Terminology: PPU in this section refers to T-Head’s AI accelerator card product line and is unrelated to DAMO XuanTie’s RISC-V processor line; thead is the upstream branch name T-Head continues to use and must not be written as “DAMO XuanTie.”

3.6 MetaX, Moore Threads, Enflame, Tsingmicro, Intellifusion: Small Steps and Backend Expansion (11-12)

Date: 2026-11-12 Sources: build-infra #1745 (MetaX MACA slice) · FlagGems #8539 (Moore Threads FP8 TopK) · FlagGems #8785 (Enflame GCU300) · FlagTree #1598 (Intellifusion)

  • MetaX: #1745 pre-provisions a MACA SDK slice for the metax builder, fixing the nightly metax daily failures after FlagTree 0.7.1 enabled FlagPrism by default; TransformerEngine-FL #139 fixes compatibility between TP overlap and the older native Userbuffers; FlagGems #8971 restricts dropout rounding semantics to MetaX and Hygon. Another 13 FlagGems-vllm quantization and routing optimizations are under review.
  • Moore Threads: #8539 optimizes FP8 TopK selection with an exact codebook histogram; on the KernelGen side, 5 MThreads-specific operators (grid_sampler_3d, upsample_nearest3d_backward and grad_input, conv_tbc, convolution_double_backward, greater) are under review, aimed at filling operator coverage rather than performance optimization.
  • Enflame: #8785 and #8968 fix compilation issues for the foreach family and asin_, erfc_ on GCU300; FlagCX #712’s ECCL device API is under review, and if merged Enflame will gain a device-API-level communication backend.
  • Tsingmicro: FlagGems #8758 (skipping out-of-bounds tiles in the GSL persistent loop for tx81/tx82) is still under review with no merges this period; Tsingmicro’s FlagTree 0.7.1 version remains at a1.
  • Intellifusion: FlagTree #1598 merges the Intellifusion backend (Triton 3.8, ifwa:100 target), including lowering, software pipelining, async copy, and a standalone wheel workflow; it is not yet registered in the README mainline matrix and counts as a newly added backend. This is also the first new vendor to join in the 2.3 cycle.

IV. Summary and Trend Observations

  • 2.3.0 moves from “scheduling proposals” to “filling in image docs”: Of the 28 build-infra PRs this period, 16 are the Iluvatar corex5.0.1 version transition and 12 are changelog backfills; the smallest unit of closure is now down to “one doc per backend.” This means release preparation is no longer about holding a meeting to set a date, but about aligning version data, drivers, and packaging lines backend by backend.
  • Version releases are gated in batches by backend: FlagTree 0.7.1 stable was granted only to ascend3.5 and xpu3.6, with the rest still at a1/a2/dev. This is not delay but the inevitable result of each backend passing CI on its own, and it also explains the pressure around the December 7 branch-cut node — backends not released before the cut can only ship pre-release versions.
  • The vLLM 0.32 migration becomes this week’s main thread: NVIDIA and T-Head have merged, and eight others are under review; yesterday’s py3.10 failure at Kunlunxin shows the real risk lies in platform differences, not API renames. The key question next period is “which vendor gives up on 0.32 first” or “whose patch lands upstream.”
  • Hygon’s “catch-up” investment is worth noting: 27 PRs in a single day, 20 foreach operators plus the bmm trio, directly reducing fallbacks on optimizer and matrix-multiplication paths. This kind of work makes no headlines, but it determines whether training scripts can run end-to-end on domestic cards, and changes usability more than single-point performance optimization.
  • The FlagQuantum surge needs to be viewed separately: 70 commits and 29 merges in one day advance both Hubbard/t-V model classes, parity symmetry, and native implementations of 216 programs at once. Quantum simulation is shifting from “demo” to “trainable model coverage,” but it remains far from productization and should not be read at release-level significance.
  • The news side and the code side continue to diverge: gnews component-name queries returned zero hits for consecutive periods, and nearly all facts this period come from GitHub. The daily brief’s judgments increasingly rely on first-hand data from commits, PRs, and the packaging matrix, with industry articles serving more as background than as evidence.

Appendix: Source Verification Table and Full Source List

Source Verification Table

Source Verification Result
GitHub org: flagos-ai repository list 64 repos, 19 pushed within the window; pushed_at checked repo by repo (UTC converted to Beijing time)
GitHub commit search within window 193 commits (11-12 06:00 to 11-13 06:00), classified by repo and vendor tag
GitHub merged PR search 167 PRs, classified by repo and vendor tag; another 272 open PRs had updates within the window, of which 120 were newly opened
GitHub Releases and Tags Two new tags: FlagTree 0.7.1+ascend3.5, 0.7.1+xpu3.6; FlagGems v5.5.0.dev1 and vllm-plugin-FL v0.5.0.dev0 predate the window and belong to the previous issue
Packaging matrix check build-infra base and runtime directories, flagcx/backends.yaml read directly; 23-backend matrix validation data taken from PR descriptions
Google News RSS (24 query sets) 38 candidates after dedup; component-name queries returned zero hits, with hits concentrated in member-company stock-market pieces and explainer articles
HN Algolia (FlagOS, FlagGems, FlagScale, FlagTree, BAAI, etc.) Tokenization mismatches only; no related stories or comments within the window
CSDN FlagOS column Latest post dated 2026-08-28; no new posts within the window
BAAI Community Model, event, and explainer content within the window; no FlagOS project updates

Full Source List