FlagOS Daily Intelligence Report (2026-11-20)
Research window: 2026-11-19 06:00 ~ 2026-11-20 06:00 (Beijing time, 24 hours; continuing from the cutoff of the previous daily report).
Sources: On the GitHub side, all 65 flagos-ai repositories were checked, including 196 commits within the window, 167 merged PRs, 261 in-review updates within the window, and pushes to 19 repositories. In addition, Releases and Tags of 31 repositories, 24 sets of Google News queries, HN Algolia, the CSDN FlagOS column, and member organizations’ official sites were checked; see the appendix source list for details.
This Issue’s Index
- Today’s Focus: Cambricon Neuware 4.7.6 connects all three layers in one day; 2.3.0 base adds the latest SDK line (11-19)
- Today’s Focus: New BIS rules take effect, triggering a broad selloff in domestic compute stocks; Enflame issues announcement of “abundant orders on hand” after limit-down (11-19)
- I. Open Source Project Progress (GitHub Activity)
- 1.1 Organizational cadence: 19 repos pushed, 196 commits, 167 merges, peak at 12:00 (11-19)
- 1.2 build-infra: Cambricon 4.7.6 lands across three layers, Hygon adds amdsmi (11-19)
- 1.3 build-infra: 910C three-line change routed to FILE_STORE via dispatch, hardcoding reverted (11-19)
- 1.4 Enflame vendor docs correct three statements at once; ECCL’s 64-bit limit recorded separately from operator limits (11-19)
- 1.5 FlagGems: 83 commits, KernelGen runs three lines in parallel covering Moore Threads, Nvidia, and Ascend (11-19)
- 1.6 FlagGems: fused/ renamed to deprecated/, 343 files with compatibility shim (11-19)
- 1.7 FlagGems: Ascend A5 architecture detection and 950DT/950PR scaffolding merged (11-19)
- 1.8 FlagGems-vllm: v0.2.1.dev1 tagged, first A5 operator speeds up ~4.5x (11-19)
- 1.9 Torch-FL: MetaX, Hygon, and Enflame native hosts merged same day (11-19)
- 1.10 sglang-plugin-FL: SGLang 0.5.18 adaptation, T-Head and Enflame close out same day (11-19)
- 1.11 FlagCX: Kunlunxin hardware CI merged, all 260 test cases pass (11-19)
- 1.12 FlagQuantum: 52 commits, focus shifts from frontend construction to reuse and evidence chain (11-19)
- 1.13 Other repos: FlagTrain, FlagTree, FlagAttention, Compressor, FlagScale (11-19)
- 1.14 Review queue: 261 updates, vLLM 0.32 adaptation wave covers six backends (11-19)
- II. News Coverage and Ecosystem
- 2.1 Channel verification: 24 query groups, 37 candidates, still zero hits on component names (11-19)
- 2.2 Phoenix News recaps Ascend PyTorch Meetup: FlagOS presents TLE three layers and Recurrent KDA (11-19)
- 2.3 Pandaily follows up in English on Token Factory, no new details beyond 11-17 (11-19)
- 2.4 CSDN column and BAAI Community: no new technical articles in window (11-19)
- III. Member Company Deep Dive
- 3.1 Enflame: the announcement on limit-down day covered only three things (11-19)
- 3.2 Cambricon: the trade-off of 4.7.6 and 4.7.2 coexisting (11-19)
- 3.3 Intellifusion: IFWA integrated explicitly via GEMS_VENDOR, still under review (11-19)
- IV. Summary
Today’s Focus: Cambricon Neuware 4.7.6 lands all three layers in one day; 2.3.0 base adds the latest SDK line (11-19)
Date: 2026-11-19 Source: #1906 · #1907 · #1909 · cambricon-neuware4.7.6 changelog
- Three layers same day: #1906 adds the base layer (ubuntu:24.04), #1907 adds the FlagCX build line, #1909 adds the runtime layer and changelog. All three merged in sequence between 06:02 and 06:47 Beijing time.
- Version details: The toolchain is cntoolkit 4.7.6, cnnl 2.2.18, cnnlextra 2.5.0, and mluops 1.13.0. On the communication and video side: cncl 1.30.10, cncv 2.12.1, cnclep 1.5.0; monitoring is cnmon 6.5.52. The changelog states this line is the vendor’s latest SDK line, with 4.7.2 remaining in service.
- Runtime stack: Python 3.12, torch 2.13.0+cpu, torch-mlu 1.34.1, torch-mlu-ops 1.13.0, with torchvision 0.28.0 and torchaudio 2.11.0. The compiler is vendor Triton 3.4.0+mlu2.1.3. This line does not install FlagTree; Triton is the sole compiler.
- Why split into three PRs: The three pip dependencies for 4.7.6 are not yet ready: torch-mlu 1.34.1, torch-mlu-ops 1.13.0, and flagcx 0.14.0+neuware4.7.6. They are waiting on flagos-pypi-cambricon to be completed. Base lands first, with the FlagCX build line and runtime each in their own PR.
- An easy-to-miss fix: The new line reuses the
cambriconlabel. Without this override, the matrix would fall back to the defaulth20runner, and since Cambricon nodes don’t carry that label, jobs would queue indefinitely. - Why FlagCX has a separate recipe: The deb filenames for 4.7.6 differ from 4.7.2 — cncl is 1.30.10 and cnnl is 2.2.18 — neither derivable from the line key. The recipe follows the self-contained ubuntu:22.04 form of 4.7.2, with wheels linking at most glibc 2.35;
default_for_vendoris left false for now, to be switched after node validation. - Observation: Same playbook as the 11-18 MetaX line switch — base first, runtime follows, index completed, then wheels. MetaX took six PRs that day; Cambricon took three this time. With ten days left until rc0 on 11-30, the base layer is still aligning the manifest item by item.
Today’s Focus: New BIS Rule Takes Effect, Triggering Broad Selloff in Domestic Compute Stocks; Enflame Hits Limit Down, Then Announces “Robust Order Backlog” (11-19)
Date: 2026-11-19 Sources: Enflame Technology Abnormal Fluctuation Announcement · National Business Daily · EE Times China · 21jingji.com reposting Shanghai Securities News
- The rule itself: On November 16, the U.S. Bureau of Industry and Security (BIS) issued an interim final rule, “Addressing Remote Access to Advanced Semiconductors” (RIN 0694-AK50), which was published in the Federal Register and took effect on November 19. Existing compute services have a 90-day license transition period.
- Three tightenings: Overseas data centers providing compute exceeding a cumulative 12 million TPP to companies in the D:5 country group must apply for a license; Indonesia, Malaysia, Singapore, Thailand, and Brazil are added to the licensing scope; for 16/14 nm and below FinFET logic chips, foundries can no longer override the controlled presumption under 3A090.a by self-verifying via transistor count.
- The third point is key: Design companies must be placed on BIS’s “Approved IC Designer” list before foundries can accept their orders. This was not fully priced in during the final two trading days before the rule took effect.
- Market action: On November 19, Enflame Technology sealed a 20% limit-down, closing at RMB 364.33, down RMB 91.08, with turnover of RMB 1.962 billion and a turnover rate of 28.35%. MetaX, Moore Threads, Hygon, Centec, and Hong Kong-listed Iluvatar CoreX fell the same day, most down more than 10% intraday.
- The other side: On the same day, SMIC’s A-shares rose 4.38% and Hua Hong Semiconductor rose 7.47%. Capital did not leave domestic compute; it merely shifted from the design side to the manufacturing side.
- Rumor and debunking: Starting on the evening of November 18, rumors circulated that Samsung was conducting strict, penetrating scrutiny of foundry arrangements; on the 19th, a screenshot naming each company’s foundry dependencies spread, claiming Iluvatar CoreX, ASR Microelectronics, and Enflame Technology were most dependent on TSMC. After seeking confirmation from multiple parties, Shanghai Securities News reported that Samsung has no formal document on strict foundry scrutiny and the rumor is inaccurate; approval certainty for 14/16 nm orders is relatively high, while 5 nm and below are reviewed case by case. ASR Microelectronics responded the same day that chips which have obtained an ECCN classification code and are confirmed to fall outside restricted scope can proceed with tape-out normally.
- Relationship to FlagOS: Five of the companies named are FlagOS member organizations, including Enflame, MetaX, Moore Threads, Hygon, and Iluvatar CoreX. This fluctuation does not involve any component code, but the capital-market environment for member organizations changed together within a single day.
- Observation: The market traded the “licensing regime” as a “blanket ban.” What has actually changed is the commercial certainty of overseas advanced-process channels, not same-day capacity. The structural fact that domestic compute has “achieved design localization but not yet manufacturing localization” was brought to the surface by a rumor for the first time.
I. Open-Source Project Progress (GitHub Activity)
1.1 Organizational Rhythm: 19 Repos Pushed, 196 Commits, 167 Merges, Peak at 12:00 (11–19)
Date: 2026-11-19 Source: flagos-ai organization · FlagGems · FlagQuantum
- Totals: 19 of 65 repos had pushes in the window. Commit search returned 196 hits, 167 merged PRs, and 261 open PRs updated in the window.
- Commit distribution: FlagGems 83, FlagQuantum 52, build-infra 17, FlagGems-vllm 15. FlagQAI and FlagTrain 6 each, Torch-FL 5, FlagTree 4, and the remaining six repos 8 combined.
- Time distribution: first commit at 08:44, last at 03:57 the next day. 12:00 peaked at 33 commits, 14:00 at 32, 11:00 at 24, 15:00 at 17, and 16:00 and 09:00 at 15 each. After 22:00 it dropped to single digits.
- One overnight stretch: the five commits from 01:57 to 03:57 all came from FlagQuantum — CZ block compression, coverage environment reuse, Cirq layer lookahead, and runtime slicing. Other repos had largely wrapped up after 21:00.
- What’s absent: no new repos and no new formal Releases in the window, only one pre-release tag for FlagGems-vllm.
- Observation: the commit-to-merge ratio went from 148 to 149 yesterday to 196 to 167 today. Merges haven’t kept pace with commits, indicating a batch of large changes under review that day, especially FlagGems’ KernelGen batch and the rename.
1.2 build-infra: Cambricon 4.7.6 Lands Across Three Layers, Hygon Adds amdsmi (11–19)
Date: 2026-11-19 Source: #1906 · #1909 · #1917 · #1918
- Cambricon line: 5 merges covering base, the FlagCX build line, changelog dates, and runtime. Details in today’s first highlight.
- Hygon line: 3 merges. #1917 adds
amdsmi==7.0.2to runtime dependencies so torch’s cuda monitoring calls work; #1918 records the pending changelog; #1919 backfills the 2.3.0 date. - Why it was missing:
hy-smidoesn’t expose a Python API, and the amdsmi source tree under/opt/hyhal/share/smiis a host mount that the image never installs, so the module was entirely absent before. - Ascend line: 3 merges. #1921 changes how FILE_STORE is passed; #1920 and #1923 backfill the 2.3.0 date and refresh image descriptions.
- Bookkeeping: #1911 records the flagcx wheel release, #1914 and #1910 refresh base and runtime image descriptions, and #1922 records the PYTHONPATH rebuild in vendor.sh.
- Total: of 17 merges, 11 are line switches and fixes, 6 are changelog and description refreshes. None are new features.
- Observation: yesterday was a three-vendor swap; today is cleanup and gap-filling. Base-layer work has shifted from front-line switching to completing changelogs — a shape approaching feature freeze.
1.3 build-infra: Three 910C Lines Now Dispatch FILE_STORE, Hardcoding Reverted (11–19)
Date: 2026-11-19 Source: #1921 · #1925
- Problem:
resource.flagos.netdoesn’t resolve on hw114 and hw115, the two machines that build the 910C lines. Both base rebuilds hung on the first SDK download withCould not resolve host. - Why only one survived: 9.2.1-910c didn’t hang because it had earlier hardcoded the intranet endpoint into its Containerfile (#1770).
- Fix:
FILE_STOREbecomes anARGwith a public default, andbuild-config.ymlgainsbuild_argsso only lines needing the intranet use the intranet endpoint. The three 910C lines now share one public default, with the difference confined to dispatch. - Rationale in the commit: the address is a site fact, not part of the image recipe, so hardcoding isn’t a good fix.
- One under review: #1925 points the runtime rebuilds of the 8.5.0 and 9.0.0 lines at the intranet PyPI for the same reason — uv downloads hang on
resource.flagos.net. The 910B line still uses the public index. - Observation: the same class of problem was handled at two granularities in one day — structural change at the base layer, spot patches at the runtime layer. That fits the tradeoff near rc: structural changes only once, everything else just to get builds passing.
1.4 Enflame Vendor Doc Corrects Three Claims at Once, ECCL’s 64-bit Limit Separated from Operator Limits (11–19)
Date: 2026-11-19 Source: #1916 · #1915
- Stale claims: the vendor page still listed
enflame-tops1.9.10andtops1.10.6, while the matrix had long since moved totops1.9.29andtops4.1.6. Both are corrected, and the L600 chip corresponding to 4.1.6 is added. - ECCL versions unrelated: the two lines’ ECCL versions are 3.6.3.13 and 4.1.4, independent of each other. The doc previously implied they shared a source.
- Evidence in #1915: fp64 and int64
all_reducefail on both MIN and SUM. This is an ECCL dtype boundary, not a regression in a particular SDK version. - Two different walls:
TORCH_GCU_ENABLE_INT64_AND_UINT64andENABLE_I64_CHECKonly govern kernel code generation. Enabling them won’t make collective communication accept 64-bit types, and the doc now states this explicitly. - Evidence grading: 1.9.10 and 1.10.6 were tested in this repo, 1.9.29 comes from an issue and wasn’t reproduced here, and 4.1.6 is unverified. The three levels are labeled separately rather than mixed together.
- Observation: this is claim governance. The same doc had both stale version numbers and an inference error conflating two walls into one; the fix also labels the evidence sources.
1.5 FlagGems: 83 Commits, KernelGen Runs Three Lines in Parallel Across Moore Threads, Nvidia, and Ascend (11–19)
Date: 2026-11-19 Source: FlagGems commits · #9467 · #9533 · #9499
- Counts by category: 43 KernelGen-generated operators (Moore Threads 21, Nvidia 16, Ascend 4, other 2); the remaining 40 are precision fixes, CI, and refactors.
- Moore Threads: a batch of 14
_foreachoperators from 09:36 to 09:57, plus 7 more from 15:08 to 15:13. Covers scalar and list variants of addcmul, asin, atan, ceil, clamp, div, erf, erfc, exp, and more. - Nvidia: two sparse parameter-validation operators on the sparse side (bsc and compressed), plus
_sparse_addmmandmatrix_H. The backward side adds autograd for logcumsumexp, log1p, log2, lu_unpack, max, min, median, logsumexp, and more. - Ascend: four operators — divide, true_divide, conv_tbc,
_adaptive_avg_pool3d_backward— merged within the same minute. - Precision and timeouts: one concentrated batch of 13 merges covering both Ascend and MetaX. On Ascend: FP8 input construction, grid upper bounds for cumprod and unique_dim, and a lower-triangular failure in triangular_solve. On MetaX C550: multi_margin_loss validation and 4-bit matmul precision. On T-Head PPU: a nanmedian cold-compile timeout.
- Kunlunxin: 4 commits, including ldexp precision, a batch_norm fix, a sobol_engine fast-forward optimization, and routing GRU to the vendor kernel.
- Observation: KernelGen has turned operator completion into a pipeline, so 43 in a day isn’t unusual. What’s genuinely scarce is the acceptance side — nearly all of the day’s precision fixes landed on individual backends, meaning operators are being added faster than they’re being aligned.
1.6 FlagGems: fused/ Renamed to deprecated/, 343 Files with a Compatibility Shim (11–19)
Date: 2026-11-19 Source: #9462
- Scope:
src/flag_gems/fused/becomessrc/flag_gems/deprecated/, with the same-named directories in 18 vendor backends renamed too. 343 files, 2020 insertions, 1468 deletions. - Rationale: these operators have been superseded by implementations in FlagGems-vllm and other modules, and the directory name should reflect that.
- Compatibility:
src/flag_gems/fused/__init__.pykeeps a shim that points the old package name at the new one via a singlesys.modulesassignment.from flag_gems.fused import Xin tests and benchmarks needs no changes. - Synchronized updates: paths in
AGENTS.md,CLAUDE.md, and the Chinese and English docs are all changed todeprecated/. - Leftover: after the rename, the
check-no-torch-fallbackhook errors on several vendor files; the author notes these violations predate the rename and weren’t introduced by it, and separate PRs will be filed per vendor. - Observation: a pure rename, but it touches 343 files. This kind of change must be done before feature freeze, or backend adaptation gets dragged down by old paths.
1.7 FlagGems: Ascend A5 Architecture Detection and 950DT/950PR Scaffolding Merged (11–19)
Date: 2026-11-19 Source: #9543
- Problem: Ascend generations can’t be distinguished by a generic lookup. On NPU,
props.majoris None, and both 910B and 950 SoC names start with 9. - First change: the backend framework exposes a
get_arch_keyhook so vendors can take over architecture determination. The returned key indexesARCH_MAP; returning None falls through to the generic kernel. - Second change: the Ascend vendor layer declares
ARCH_MAP = {"a5": "a5"}; the 910B series remains the baseline and needs no entry. TheARCHenvironment variable takes precedence, with aliases includinga5,ascend_a5,950,ascend_950; otherwise it substring-matches the driver-reported SoC name, with950covering all SKUs. - Fallback: returns None for CPU-only builds or when the driver is unavailable, leaving the generic path unaffected.
- Scaffolding: new empty
a5/anda5/ops/packages as the landing spot for future A5 operator specializations, with reserved per-architecture tune_configs and enable_configs. - Observation: this commit adds no kernels, only detection and a place to put them. The same day, FlagGems-vllm merged its first A5 operator on the
ascend-a5-opsbranch — two lines making the same preparation. - Note: A5 corresponds to Ascend950DT and Ascend950PR, with NpuArch DAV_3510.
1.8 FlagGems-vllm: v0.2.1.dev1 Tagged, First A5 Operator ~4.5× Faster (11–19)
Date: 2026-11-19 Source: v0.2.1.dev1 · #1338 · #1321
- Tagging: v0.2.1.dev1 was released at 10:27 Beijing time, with a one-line release note: “Release version 0.2.1.dev1”. The same day, FlagTree’s CI bumped the HCU gems-vllm baseline to that version.
- First A5 operator: #1338 adds a persistent tiled implementation of
topk_softplus_sqrtfor 950DT, landing inascend/a5/ops/. With 256 experts, bf16, topk=6, 4096 tokens drops from ~250 µs to 55 µs. - Why it’s fast: each program handles
BLOCK_Mtokens and fetches the whole block of scores with one widetl.load; the grid is capped at the AIV core count (950DT falls back to 96), so launch overhead no longer grows with token count. - Two write-back paths: wide blocks write directly to HBM by rank; narrow blocks assemble
[BLOCK_M, BLOCK_K]first and write once contiguously. Indices take the smallest id among the max-value experts, preserving the original tie semantics. - 910 side aligned the same day: #1321 ports the A5 KDA kernel back to 910 and changes the benchmark to take values against the vendor AscendC operator. Two constants changed in the port:
_ASCEND_BTfrom 64 to 32 (910B’s AIV core UB is only 192 KB, and 64 needs 224 KB), and the Neumann expansion from 0 to 3. - Rejected approach: a tiled port of A5’s
causal_conv1d_fnran at 63.9 ms on (10, 4096, 4096) versus 1.20 ms for the current 910 version. The cause: channel-first tiling pinsBLOCK_Nat 32, producing 1664 small programs. This was empirically rejected and documented in the PR. - Observation: this is day one of A5 moving from scaffolding to measurement. The same batch also fixed the benchmark methodology — earlier speedups were inflated by taking values against torch loops or vllm-ascend’s Triton fallback.
1.9 Torch-FL: MetaX, Hygon, and Enflame Native Hosts Merged the Same Day (11–19)
Date: 2026-11-19 Source: #849 MetaX · #850 Enflame · #851 Hygon · #875 refactor
- Shared goal: get the full Qwen3-0.6B BF16 eager pipeline running on each vendor’s native host, with logits aligned to the vendor PyTorch and no numerical fallback to CPU.
- MetaX: all 86,907,392 logits match, and all 572 top-1 positions align. Median over three full passes is 14,066.858 ms versus a vendor baseline of 11,771.663 ms, reaching 83.68% of vendor throughput.
- Hygon: 28 layers, 1,080 checkpoint comparisons, 190,831,616 logit elements bit-identical. But the two short-sequence groups remain at 1.264× and 1.260×, below the 1.25× threshold, and the PR states the performance gate is still open.
- Enflame: the reverse-order control group reaches 81.85% prefill and 79.40% decode; combining both rounds gives 81.38% and 78.60%, missing the 80% threshold on the decode side. The slower first round (69.43% and 67.01%) is retained and explicitly reported.
- Refactor: #875 centralizes native-host builds and reorganizes the C++ sources, paving the way for a shared generation pipeline across the three vendors.
- One process detail: all three PRs carry AI-agent information, naming the tool and the assigned reviewers, and noting that “requesting review is not approval”.
- Observation: three vendors merging the same day shows the generative-host route has gone from pilot to pipeline. All three PRs leave the missed numbers as-is, which is more useful than reporting a flattering speedup.
1.10 sglang-plugin-FL: SGLang 0.5.18 Adaptation, T-Head and Enflame Close Out the Same Day (11–19)
Date: 2026-11-19 Source: #160 T-Head · #168 Enflame
- Shared premise: SGLang 0.5.18 is installed as-is in empty mode, with all adaptation kept in the plugin and no upstream changes.
- T-Head: PPU switches to its own dispatch config
thead.yamlinstead of inheritingnvidia.yaml; the default attention backend is triton. Single-node, cross-node TP4×PP2, and MTP configurations are verified with Qwen3.6-27B and 35B-A3B. - Stub mechanism: 0.5.18 imports flashinfer and deep_gemm at module level under
is_cuda(), but the empty runtime doesn’t ship these packages. The plugin substitutes placeholder modules that only fail when called, with real installed packages taking precedence. - Enflame: 1106 unit tests pass, 50 skipped, 0 failed. E2E runs on GCU with Qwen3.6-27B and 35B-A3B at TP4, covering both concurrent and serving modes.
- One real fix: the GCU MoE path previously enabled in-place unconditionally and returned
hidden_states, ignoring the caller’s config. The new version respectsmoe_runner_config.inplace, supports out-of-place, and returns the correct output buffer. - Two under review: #167 MetaX and #181 Ascend 0.5.18 adaptations were updated the same day but not yet merged.
- Observation: the SGLang 0.5.18 adaptation is a synchronized move across five vendors. Upstream 0.5.18’s empty mode opened up room for plugins, at the cost of every vendor having to handle the same batch of module-level imports.
1.11 FlagCX: Kunlunxin Hardware CI Merged, All 260 Cases Pass (11–19)
Date: 2026-11-19 Source: #671 · #726 · #725
- New CI: #671 adds a
kl-8g-cicd-flagcxrunner for Kunlunxin, plus runtime environment and MPICH test scripts, with both core and service suites entering the matrix. All 203 core cases and 57 service cases pass in-container. - Resource ownership: #726 changes Window deregistration to a local operation and unifies Device API resource ownership. A small change, but it determines whether destruction order is symmetric with creation order.
- Platform extensions under review: #719 T-Head PPU, #705 Ascend, #722 MUSA, #725 IXCUDA VMM/LSA — all four add platforms to the Device API.
- Another under review: #672 adds unit-test CI for Enflame.
- Observation: the communication library is doing two things — rolling hardware CI out to each vendor, and expanding the Device API from single-platform to multi-platform. The former determines whether regressions run automatically; the latter determines whether new backends require core changes.
1.12 FlagQuantum: 52 Commits, Focus Shifts from Front-End Construction to Reuse and Evidence Chains (11–19)
Date: 2026-11-19 Source: #1226 · #1229 · #1220 · #1232
- Volume: of 52 commits, 25 are compiler-related, 6 synth, 6 CI and coverage, 6 merge commits, with the rest spread across qec, deployment, runtime, and evidence.
- CZ block optimization: #1226 wires automatic CZ block compilation into the main flow, and #1229 further compresses native CZ blocks. A 50-seed development probe cuts emitted operands from 1859 to 1299 with the CZ count unchanged and a full-operator residual ceiling of 1.39e-14.
- Upstream reuse: #1232, #1233, and #1234 bring Qiskit’s frontier-relative scoring, bounded routing trials, and Cirq’s layer lookahead into the FlagQuantum IR respectively. Two rejected Cirq interaction-kernel adaptations also keep their numerical records.
- Error-correction side: #1220 pushes the exact toric code distance proof to d=15.
- One under review: #1084 wires native fault-tolerant compilation and reversible phase execution into the compiler, not yet merged.
- Observation: yesterday’s quantum line was filling in error-correction decoding and the research workbench; today it moved to the compiler front end. Acceptance for this kind of work rests on residuals and operand counts, not throughput, so the numerical standards are stricter than a performance report.
1.13 Other Repos: FlagTrain, FlagTree, FlagAttention, Compressor, FlagScale (11–19)
Date: 2026-11-19 Source: FlagTrain #92 · FlagTree #1653 · FlagAttention #113 · Compressor #13
- FlagTrain: 6 commits. Adds Ascend
multi_tensor_lionandmulti_tensor_adamoperators, alignsevoformer_attnwith DeepSpeed’s public interface, and adds single-case benchmark profiling plus ascaled_upper_triang_masked_softmax_backwardwith a triton kernel. - FlagTree: 4 commits. HCU’s gems-vllm baseline is bumped to v0.2.1.dev1; TLE raw’s
topsandefshmemsupport Enflame; T-Head PPU’s ldmatrix now preserves broadcast registers; and a PyTorch/vLLM operator timing script is added. - FlagAttention: #113 optimizes GDN’s prefill and recurrent decode.
chunk_gated_delta_rule_fwdreaches a 1.996× geometric mean andpacked_decode1.102×, against a vLLM 0.19.1 baseline on an H100 80GB test card. - FlagOS-Compressor: #13 adds
--devicestoquantizeandconvertfor per-device parallel processing of safetensors shards. GPTQ, AWQ, and native AutoRound are explicitly rejected and still use serial calibration. - FlagScale: 1 commit fixing an aiohttp security vulnerability and pinning the version to one without it.
- FlagGems-sglang: 1 commit adding 15 public operator interfaces — HC, QSA, and PLE — for Qwen3.8-Flash-Next, with sparse attention parameters on MUSA measured on S5000.
- Observation: these smaller repos shared one thing that day — each was filling in the last piece for a specific model or a specific backend. The division of labor has shifted from “who maintains which library” to “who owns which pipeline”.
1.14 Review Queue: 261 Updates, vLLM 0.32 Adaptation Wave Covers Six Backends (11–19)
Date: 2026-11-19 Source: vllm-plugin-FL #711 · #693 · #726 · FlagGems #9545
- Totals: 261 open PRs updated in the window — FlagGems 43, FlagGems-vllm 21, vllm-plugin-FL 9, FlagTree 5, FlagCX 5, FlagQuantum 4.
- vLLM 0.32 wave: MetaX C550 (#693), Kunlunxin P800 (#726), and Hygon DCU (#711) are under review. Moore Threads MUSA (#715), Tsingmicro txda (#766), and Sunrise (#697) are doing the same thing, all against the 0.32-empty wheel.
- Hygon’s approach: instead of maintaining a private runner fork, it provides the vLLM 0.32 symbol shims from its own patches module. Both V2 and V1 runners are verified: eager 51/52, graph 52/52.
- New face: FlagGems #9545 adds explicit backend selection for Intellifusion IFWA, registering the
ifwadevice viaGEMS_VENDOR=intellifusion, starting with generic heuristics and tuning data to follow. - Another main review thread: FlagGems’ Nvidia backward operators still have 20-plus queued, with KernelGen generating clearly faster than review proceeds.
- Observation: the most concentrated item among the 261 open PRs is vLLM 0.32 adaptation. Six vendors doing the same thing at once shows upstream release cadence has bound downstream into synchronized action. That also explains why the day’s merge count was 167 rather than 196.
II. News Coverage and Ecosystem
2.1 Channel Check: 24 Query Sets, 37 Candidates, Still Zero Hits on Component Names (11-19)
Date: 2026-11-19 Source: Google News FlagOS query
- Query scope: 24 sets in Chinese and English combined. Component names covered FlagOS, FlagGems, FlagScale, FlagTree, FlagPerf, FlagCX, and KernelGen; organization terms covered BAAI, Moore Threads, MetaX, Enflame, and Qingwei; HN Algolia was also run.
- Component names: FlagGems, FlagScale, FlagTree, FlagPerf, FlagCX, and KernelGen continued to return zero hits within the 14-day window. FlagOS itself had 2 hits, both reprints of the same event.
- Valid candidates: Only about 6 of the 37 made it into screening; the rest were stock-price pieces. Nearly all 30 Enflame-related items were capital-flow and top-trader-list reports.
- Excluded: One gambling-SEO spam result failed at the parsing stage; BAAI model-related news was excluded per the rules.
- HN: None of the three queries returned relevant hits; all results were generic AI-industry news.
- Observation: FlagOS’s news surface is still dominated by capital-market news from member organizations, while component versions only appear on GitHub. This pattern has held for several consecutive issues.
2.2 Ifeng Recaps Ascend PyTorch Meetup: FlagOS Covers TLE’s Three Layers and Recurrent KDA (11-19)
Date: 2026-11-19 Source: Ifeng Tech · PyTorch official post
- Occasion: The Beijing stop of the Ascend for PyTorch Meetup on November 7, jointly hosted by communities including PyTorch, vLLM, OpenBMB, LLaMA-Factory, FlagOS, and PenguinHarness. Ifeng published its recap on November 19.
- FlagOS’s framing: BAAI engineer Wang Kaize described FlagOS as a cross-chip unified operator layer centered on FlagGems, analogous to PyTorch’s ATen. Integration takes one line of code, and more than 1,400 operators can run on over 20 AI chips.
- Ascend specifics: TLE’s light, strict, and raw layers are written into the same Triton function, controlling all the way down to NPU IR.
- KDA optimization: Recurrent KDA (Kimi Delta Attention) splits 40 cores into 8 producers and 32 iterators, cutting TPOT by 3% to 4%.
- Other topics: TorchNPU distributed optimization, vLLM-Omni’s continuation and editing pipeline, ModelBest’s expert weight swap-in/swap-out, and 30 community tasks opened on site at the PyTorch contribution workshop.
- Observation: This is a late on-site recap, but it contains two figures that had not appeared in the daily brief before: the operator scale of over 1,400 and coverage of over 20 chips, plus KDA’s 3% to 4% TPOT gain.
2.3 Pandaily Follows Up in English on the Token Factory, Without Going Beyond the 11-17 Line (11-19)
Date: 2026-11-19 Source: Pandaily
- Timing: Pandaily published an English report on November 19 covering the domestic Token Factory end-to-end inference optimization solution jointly released by BAAI and Qingmao Intelligence on November 17.
- Figures carried over: 21 chip companies, 86 ecosystem partners, 35 AI chips; the 9 major operator libraries hold more than 1,454 Triton operators cumulatively, of which nearly 800 reach over 80% of native-operator performance on at least 5 domestic chips; a single code repository covers backends from 20 vendors.
- New phrasing: Qingmao Intelligence founder and CEO Guan Chaoyu described the demand as “hardware bought by card and by scale, while the application side wants tokens by the second,” requiring an intermediate conversion layer that organizes heterogeneous compute into metered capacity.
- What’s missing: No new performance figures, and no disclosure of the chip and model list for the commercial service.
- Corroboration: Two other English aggregators, Tiwen and ByteBrief, also followed up on the same event on November 19, both citing the November 17 launch event.
- Observation: This is international pickup, not a new event. Its value lies in spelling out FlagOS’s positioning in English for the first time: no upstream fork; plugins connect vLLM, SGLang, and Megatron-LM.
2.4 CSDN Column and BAAI Community: No New Technical Posts in the Window (11-19)
Date: 2026-11-19 Source: FlagOS CSDN column · BAAI Community
- CSDN column: The latest batch of articles visible on the homepage is still three: GLM-5.3-Flash Day0 adaptation for 9 chips, Qwen3.8-Flash-Next adaptation for 8 chips, and LightX2V’s integration with FlagOS, all already covered in previous daily briefs.
- Timestamps: The time field on the list page is uniform site-wide, so publication times cannot be distinguished per article. After fetching each article individually, no new titles were found within the window.
- BAAI Community: No FlagOS-related items on the homepage; pushes within the window were mainly model and leaderboard content.
- Observation: The cadence on the technical-communications side is out of sync with the release cadence. FlagGems-vllm shipped a release, but the column did not follow up; when the column does have posts, they usually correspond to external milestones such as model Day0 adaptation.
III. Deep Dive on Member Units
3.1 Enflame: The Announcement on the Limit-Down Day Covered Only Three Things (11-19)
Date: 2026-11-19 Sources: Abnormal Fluctuation Announcement · Securities Times · NBD
- First, the characterization of the abnormal fluctuation: Over the three trading days of November 17, 18, and 19, the cumulative deviation of closing price declines exceeded 30%, which constitutes abnormal fluctuation under Shanghai Stock Exchange trading rules.
- Second, operations and orders: The company has ample orders in hand and is working with multiple domestic semiconductor industry partners on cloud AI chip products based on a domestic supply chain. Production and operations are proceeding normally as planned.
- Third, the shareholding structure: 17.900325 million unrestricted tradable shares, accounting for 4.1595% of total share capital. The announcement listed this ratio separately as a risk warning, pointing to the low free float amplifying volatility.
- What it did not say: The announcement did not disclose the amount, customer names, or contract terms of orders in hand, nor did it respond to which specific industry news item had drawn attention.
- Listing background: The company listed on the STAR Market on September 11, 2026, at an issue price of 142.18 yuan, closing at 397 yuan on its first trading day. On November 13 it hit an intraday high of 613 yuan; on November 19 it closed at 364.33 yuan, still 156.25% above the issue price.
- Market data: On November 19 total market capitalization was 156.790 billion yuan, evaporating 39.2 billion yuan in a single day; from the intraday high on the 13th, it has shrunk by about 107 billion yuan. The four-trading-day cumulative drawdown was 36.16%, with a turnover rate of 28.35% that day and net selling of over 500 million yuan by four institutional seats.
- Interface with FlagOS: Enflame’s work on the FlagOS side consists of the tops1.9.29 and tops4.1.6 base lines in build-infra, plus a fix to the device attribute type in FlagCX. None of this changed due to the stock price fluctuation.
- Observation: The announcement is a standard fluctuation statement, and its informational value lies in what it chose to say. Orders, supply chain, and free float correspond exactly to the three areas the market was worried about that day.
3.2 Cambricon: The Trade-off of 4.7.6 and 4.7.2 Coexisting (11-19)
Date: 2026-11-19 Sources: #1906 · #1907 · Changelog
- Three lines coexisting: 4.4.3, 4.7.2, and the newly added 4.7.6. The changelog explicitly states that 4.7.6 is the vendor’s latest SDK line, while 4.7.2 remains in service.
- Not switching the default: FlagCX’s
default_for_vendoris set to false for the new line; 4.7.2 continues to providelibflagcx-cambriconuntil the new line is validated on nodes. - Build-side differences: The 22.04 deb for 4.7.6 sits in the
cambricon/root directory of the filestore, whereas 4.7.2’s is underu2204/. The recipe’s FILE_STORE therefore points to different locations. - Dependency gaps: torch-mlu 1.34.1 and torch-mlu-ops 1.13.0 are not yet complete on Cambricon’s private index; the wheels must wait for the FlagCX build image to finish before they can be compiled.
- Why it’s worth waiting: 4.7.6 corresponds to torch 2.13.0 and Triton 3.4.0+mlu2.1.3, one step newer than the 4.7.2 set. If 2.3.0 can use it, operator coverage for the Cambricon backend will move up accordingly.
- Observation: This is a deliberately slowed line switch. Not switching the default, not deleting the old line, and landing only the layers that can land—all three steps leave room for rollback. With only ten days to rc0, this pace means 4.7.6 will most likely miss the 2.3.0 default matrix.
3.3 Intellifusion: IFWA Explicitly Integrated via GEMS_VENDOR, Still Under Review (11-19)
Date: 2026-11-19 Source: FlagGems #9545
- Change: Adds explicit IFWA backend selection to FlagGems, registering the
ifwadevice viaGEMS_VENDOR=intellifusion. Generic heuristics are used until tuning data is available. - Bug fixed along the way: The shared pointwise generator used
out_bptr.dtype.element_tyto get the storage type on Triton 3.8. It previously usedout_bptr.type, which on 3.8 is an aggregate type withoutelement_ty, causing generated kernels to throw AttributeError. - Verification scope: Compared 135 generated kernel bodies; only the storage type expression changed.
- Status: Still under review, performance data pending, unit tests not yet covering it.
- Observation: Intellifusion was not previously on the member unit list. If this merges, FlagOS will have one more explicit backend. It also shows that backend integration is becoming standardized—one environment variable plus one device registration, no longer requiring core changes.
IV. Summary
- The base layer is entering its final stretch: Cambricon’s 4.7.6 landed across three layers, Hygon added amdsmi, and Ascend 910C switched to dispatch. Of the 17 merges, 11 were line switches and fixes, with no new features. Ten days remain until rc0 on 11-30.
- The operator layer is doing two things: KernelGen generated 43 operators in a day, while 40 commits fixed precision and timeouts across various backends. Coverage is being added quickly, but alignment is slow, making acceptance the bottleneck.
- A5 has moved from scaffolding to real testing: FlagGems added architecture detection and an empty directory, and FlagGems-vllm merged its first 950DT operator on the same branch with a 4.5x result. Both lines are making the same preparation.
- Generative hosts have become a pipeline: Native hosts from MetaX, Hygon, and Enflame merged on the same day, and all three PRs left subpar numbers in place as-is. This approach supports subsequent decisions better than a pretty speedup ratio.
- vLLM 0.32 ties six vendors into a synchronized action: This wave is the most concentrated part of the 261 under review, with MetaX, Kunlunxin, Hygon, Moore Threads, VeriSilicon, and Sunrise all doing the same thing. The merge count of 167 being lower than the submission count of 196 is explained here.
- Only one hard news item on the news front: On the day the new BIS rules took effect, the share prices of five member units fell sharply together, and Enflame announced ample orders after hitting limit-down. Nothing changed on the component side, but the capital market environment for member units changed together on the same day.
- To watch: Whether Cambricon 4.7.6 can make it into the 2.3.0 default matrix; whether A5 operators can be backfilled on the 910; whether the six vLLM 0.32 items can all merge this week; and when Intellifusion’s IFWA tuning data will be ready.
Appendix: Source Verification Table and Complete Source List
Source Verification Table
| Source | Verification Result |
|---|---|
| GitHub org: flagos-ai | Full verification of 65 repos, 19 repos pushed within window, 196 commits, 167 merged PRs, 261 in-review updates |
| GitHub Release / Tag | Atomic verification of 31 repos, only one new tag FlagGems-vllm v0.2.1.dev1, no formal Release |
| Google News RSS | 24 groups of Chinese and English queries, 37 candidates, about 31 excluded, 6 entered screening |
| HN Algolia | Four queries for FlagOS, FlagGems, FlagScale, BAAI returned no relevant hits |
| CSDN FlagOS Column | Latest batch is the two 11-17 posts on GLM-5.3-Flash and Qwen3.8-Flash-Next, no new additions within window |
| BAAI Community | No FlagOS entries on homepage, pushes within window are model and leaderboard content |
| Member company original sites | Enflame abnormal fluctuation announcement (CNINFO), National Business Daily, Securities Times, EET-China, 21st Century Business Herald |
| Regulatory scope | BIS interim final rule RIN 0694-AK50, published in the Federal Register and effective 11-19 |
| community 2.3 roadmap | community #176, target GA 12-30, no updates that day |
Complete Source List
- [1] build-infra #1906 Cambricon Neuware 4.7.6 base layer — https://github.com/flagos-ai/build-infra/pull/1906
- [2] build-infra #1907 FlagCX build line — https://github.com/flagos-ai/build-infra/pull/1907
- [3] build-infra #1909 runtime layer — https://github.com/flagos-ai/build-infra/pull/1909
- [4] cambricon-neuware4.7.6 changelog — https://github.com/flagos-ai/build-infra/blob/main/base/changelogs/cambricon-neuware4.7.6.yaml
- [5] build-infra #1917 Hygon amdsmi — https://github.com/flagos-ai/build-infra/pull/1917
- [6] build-infra #1921 910C FILE_STORE — https://github.com/flagos-ai/build-infra/pull/1921
- [7] build-infra #1925 910C intranet PyPI (in review) — https://github.com/flagos-ai/build-infra/pull/1925
- [8] build-infra #1916 Enflame vendor page wording — https://github.com/flagos-ai/build-infra/pull/1916
- [9] build-infra #1915 ECCL 64-bit issue — https://github.com/flagos-ai/build-infra/issues/1915
- [10] FlagGems #9462 fused renamed deprecated — https://github.com/flagos-ai/FlagGems/pull/9462
- [11] FlagGems #9543 Ascend A5 scaffolding — https://github.com/flagos-ai/FlagGems/pull/9543
- [12] FlagGems #9467 Moore Threads _foreach series — https://github.com/flagos-ai/FlagGems/pull/9467
- [13] FlagGems #9499 Ascend divide operator — https://github.com/flagos-ai/FlagGems/pull/9499
- [14] FlagGems #8966 19 low-precision pointwise — https://github.com/flagos-ai/FlagGems/pull/8966
- [15] FlagGems #9545 Intellifusion IFWA (in review) — https://github.com/flagos-ai/FlagGems/pull/9545
- [16] FlagGems-vllm v0.2.1.dev1 — https://github.com/flagos-ai/FlagGems-vllm/releases/tag/v0.2.1.dev1
- [17] FlagGems-vllm #1338 A5 topk_softplus_sqrt — https://github.com/flagos-ai/FlagGems-vllm/pull/1338
- [18] FlagGems-vllm #1321 A5 KDA ported back to 910 — https://github.com/flagos-ai/FlagGems-vllm/pull/1321
- [19] Torch-FL #849 MetaX native host — https://github.com/flagos-ai/Torch-FL/pull/849
- [20] Torch-FL #850 Enflame native host — https://github.com/flagos-ai/Torch-FL/pull/850
- [21] Torch-FL #851 Hygon native host — https://github.com/flagos-ai/Torch-FL/pull/851
- [22] Torch-FL #875 native host build centralization — https://github.com/flagos-ai/Torch-FL/pull/875
- [23] sglang-plugin-FL #160 T-Head 0.5.18 — https://github.com/flagos-ai/sglang-plugin-FL/pull/160
- [24] sglang-plugin-FL #168 Enflame 0.5.18 — https://github.com/flagos-ai/sglang-plugin-FL/pull/168
- [25] FlagCX #671 Kunlunxin hardware CI — https://github.com/flagos-ai/FlagCX/pull/671
- [26] FlagCX #726 Window deregistration and resource ownership — https://github.com/flagos-ai/FlagCX/pull/726
- [27] FlagQuantum #1229 CZ block compression — https://github.com/flagos-ai/FlagQuantum/pull/1229
- [28] FlagQuantum #1226 automatic CZ block compilation — https://github.com/flagos-ai/FlagQuantum/pull/1226
- [29] FlagQuantum #1220 toric code distance d=15 — https://github.com/flagos-ai/FlagQuantum/pull/1220
- [30] FlagAttention #113 GDN optimization — https://github.com/flagos-ai/FlagAttention/pull/113
- [31] FlagOS-Compressor #13 multi-device quantization — https://github.com/flagos-ai/FlagOS-Compressor/pull/13
- [32] vllm-plugin-FL #711 Hygon vLLM 0.32 (in review) — https://github.com/flagos-ai/vllm-plugin-FL/pull/711
- [33] Enflame Technology announcement on abnormal stock trading fluctuations — https://static.cninfo.com.cn/finalpage/2026-11-20/1225680125.PDF
- [34] National Business Daily on Enflame market cap shrinkage and announcement — https://www.nbd.com.cn/articles/2026-11-19/4652923.html
- [35] Securities Times: Enflame urgently warns of low free float — https://finance.sina.cn/2026-11-19/detail-inizhrry7716375.d.html
- [36] EET-China on supply cutoff rumors and multi-party responses — https://www.eet-china.com/mp/a540540.html
- [37] 21st Century Business Herald (via Shanghai Securities News) rumor verification — https://www.21jingji.com/article/20261119/herald/a58a9d1e732ccd62ff402540ae5e363e.html
- [38] Phoenix Tech recap of Ascend PyTorch Meetup — https://tech.ifeng.com/c/8yBATVZVmuN
- [39] Pandaily English report on FlagOS Token factory — https://pandaily.com/baai-flagos-qingmao-end-to-end-inference-optimization
- [40] FlagOS CSDN Column — https://flagos.csdn.net/