Research window: 2026-11-19 07:00 ~ 2026-11-20 07:00 (Beijing time, 24 hours). The previous issue’s cutoff was 11-19 07:00.

Sources: Repository-by-repository review of the GitHub tile-ai organization; 8 of 30 repositories had pushes within the window. Organization-level ground-truth search: 23 merged, 20 updated in review, 7 newly opened issues. Per-commit main-branch review: TileOPs 16, tilelang 2, TileFoundry 1, Hygon 1. The 3 MLIR Ascend commits landed on the mlir-dialect branch. Release check: tilelang remains at v0.1.15, with no new tags in the window. Also includes five groups of Google News queries in Chinese and English, plus Hacker News and a four-source sweep of TileLang×RISC-V.


This Issue’s Index

  • Today’s Highlights: Three mHC operators merged into TileOPs on the same day, sequence modeling interface handover complete (11-19)
  • Today’s Highlights: One rule change in the GEMM selector touches 302 shapes, MXFP4 MoE decoding reaches acceptance threshold (11-19)

  • I. Core Project Progress
    • 1.1 Layout placeholder variables stay bound across serialization, spawn subprocess offload no longer reports non-injective (11-20)
    • 1.2 Dev version number drops toolchain tag, same wheel no longer forks by CUDA version (11-19)
    • 1.3 Newly reported defect: MLA decode example missing write-after-read synchronization between two warpgroups (11-19)
    • 1.4 Fix under review: skip TMA promotion for incompatible layouts during warp specialization (11-19)
  • II. Multi-Backend Adaptation (Ascend / Hygon / Moore Threads / MetaX)
    • 2.1 MLIR Ascend merges tl.or, tl.xor, tl.pow in one day, vsigmoid and sort under review (11-19)
    • 2.2 Ascend main repo review shifts to release: OBS on shared wheels, nightly NPU smoke tests, tuning skills (11-19)
    • 2.3 Atlas A3 misidentified as 910A, carver’s L0C capacity undercounted by half (11-19)
    • 2.4 Hygon adds PR quality gate workflow, no kernel changes in window (11-19)
    • 2.5 No commits from MetaX or Moore Threads in window (11-19)
  • II·V. RISC-V Backend and Community (Special Focus: TVM route / MLIR route / SOPHGO TPU / community index page)
    • 2.5.1 Four-source inspection shows no change, both routes and index page unchanged from last issue (11-20)
  • III. Ecosystem and Adopters (TileOPs operator library / TileFoundry / TileKernels)
    • 3.1 Attention sink completed: DSADense and DSAPaged merged same day (11-19)
    • 3.2 FP4 grouped quantization lands, MoE gating adds hash routing and sqrt-softplus (11-19)
    • 3.3 GatedPoolCompress completes per-request ring-partitioned KV pooling in a single launch (11-19)
    • 3.4 TileFoundry incorporates sliced mesh and Varying values into parallel plan analysis (11-19)
    • 3.5 Four new targets opened: DSAPaged bf16, AvgPool1d, DeltaNet, Mamba (11-19)
    • 3.6 No commits from TileKernels or FlashQLA in window (11-19)
  • IV. Community, Tutorials and Events
    • 4.1 News side quiet: five query sets in Chinese and English return no relevant coverage (11-20)
    • 4.2 Organization activity: 23 merges in a single day, main repo at 9006 stars (11-19)
  • V. Trend Observations

  • Appendix: Sources and Verification Notes
    • Source verification table
    • Full source list

Today’s Highlights: Three mHC operators merged into TileOPs on the same day, sequence modeling interface handover complete

On 11-19, TileOPs merged all three HyperConnection operators — pre, post, and head — in a single day. The three PRs are #3032, #3035, and #3039. Earlier, #3021 had removed the two old mHC operators on 11-18 and was registered as a reviewed public API break. One day apart, the replacement path is complete.

The pre step was merged via #3032, touching 18 files, +1106/-21. It runs the mHC front segment for DeepSeek-V4-Flash and GLM-5.3-Flash. Contents include fn projection and squared norm, pre and post weights, Sinkhorn, and weighted stream summation.

The projection guarantees float32 precision. fn enters the tensor cores as three bfloat16 terms, one 64×72 WGMMA per 16 channels. Each 64-channel slot then accumulates into float32. The token axis is a runtime length, so a single launch covers any token count. The comparison targets are vLLM’s TileLang mhc_pre and the torch.compile reference.

The post step was merged via #3035, touching 16 files, +592/-21. It mixes block outputs back into n residual streams. Element-wise it is written as post_layer_mix[j] * x + sum_i comb_res_mix[i,j] * residual[i], computed in float32 and rounded once to bfloat16.

Threads mix x and each stream in 16-byte segments. Small calls use one segment, large calls use four. Output is bit-identical to TileKernels’ mhc_post_fwd.

The head step was merged via #3039, touching 15 files, +778/-9. It folds n residual streams into one before the LM head, corresponding to DeepSeek-V4-Flash’s ParallelHead.hc_head. At most two tokens per SM; beyond that, one CTA streams one to four tokens, with each hc_fn vector serving all tokens.

The manifest entry is marked implemented, with a 3000-token row for GLM-5.3-Flash. The comparison targets are vLLM’s hc_head fused kernel and the torch reference.

All three PRs closed their corresponding proposals: #2921, #2970, #2971. The intermediate state between old operator removal and new operator merge lasted only one day; the interface replacement left no gap version.

Sources: https://github.com/tile-ai/TileOPs/pull/3032 https://github.com/tile-ai/TileOPs/pull/3035 https://github.com/tile-ai/TileOPs/pull/3039

Today’s Highlight: One Rule Change in the GEMM Selector Affects 302 Shapes; MXFP4 MoE Decoding Reaches Acceptance Threshold

The 11-19 TileOPs performance line saw six merges in total. Two self-assessments reached the acceptance threshold (#3044, #3041), while four remain marked below it (#3045, #3053, #3050, #3040). All four supporting notes state that each row is faster than the main branch.

The GEMM selector changed one rule (#3045). For NT-class GEMM, if a 128-row grid can fit into a single wave at coop2 width, the narrowest coop2 tile is selected instead of splitting into narrower multi-wave tiles.

There are two criteria. Both 64-row consumer groups of the last row tile must be assigned rows, and the grid must fill 15/16 of a wave.

The rule affects 302 model shapes. On a single H200, none is slower than the main branch’s selection, and 286 shapes are 1.03 to 2.6 times faster. GLM-5.3-Flash’s routed prefill (N=288) drops from 82.24 microseconds to 32.13 microseconds.

This PR also fixes a latent issue. When a float32 coop2 tile is tiled as a whole block, accumulator copies on threads 128 to 383 are discarded, leaving uninitialized content in shared memory.

Despite these numbers, the PR’s self-assessment conclusion remains below the acceptance threshold.

MXFP4’s MoE grouped GEMM reaches the acceptance threshold (#3044). A new shared decoding kernel is added, with the conditions that row alignment is a multiple of 128, scales are e8m0, and K is a multiple of 2048. 128 tokens serve as the WGMMA M operand, and 128 rows of weights as the N operand.

Weights are decoded once per CTA by seven production warps and written into fp8 shared memory tiles. The root cause is that the old kernel decoded weights in math warps, with one FMUL plus one FFMA per element. Each math warp issued about 363 instructions per K step, corresponding to 512 tensor cycles, leaving the tensor pipeline only 36% busy.

GQA variable-length attention switches to three consumer warpgroups for head dimensions 56 to 80 (#3041). With a 192-row tile, three 64-row consumer groups take turns occupying the tensor cores. This PR’s self-assessment reaches the acceptance threshold.

The root cause is that at 72 channels, exp2 takes 64/80 of its product tensor time. With two consumer groups, one group’s softmax drags past the other group’s product. ncu shows tensor at 54.7% and XU at 45.2%, the two being nearly serial.

The PR notes one limitation: TileFoundry’s schedule finalize cannot yet lower the delivered schedule to TIR on the current build (see tile-ai/topa#42). The kernel is hand-written from the measured prototype, and the schedule lives in final_hir.py.

KDA’s bf16 given-gating prefill switches to a single centered launch (#3053, 9 files, +749/-225). On SM90 it previously took two paths depending on sequence length and value-head log count. The new path first prepares all chunks via a joint centered tensor core program, then runs the state traversal. fp16 given-gating and raw gating still take the old path.

The root cause given in the PR is that the old path was timed by the longest sequence. On glm-5.3-flash-pf-pool, the 64 CTAs of a 3000-token sequence must traverse 47 chunks, doing full chunk preparation at every step. This PR’s self-assessment is likewise below the acceptance threshold.

The other two self-assessments are also below the acceptance threshold. W4A16’s K tile metadata switches to asynchronous copies (#3050), over-fetching sub-4-byte per-row zero points into aligned four-group windows and moving them with 4- and 8-byte cp.async. The producer can therefore issue the K tile’s TMA immediately.

RoPE adjusts its load order (#3040). Heads with column replication load the input vector before reading positions, avoiding ptxas waiting at the position clamp. Small GPT-J calls thus pay one less DRAM latency. The same PR also fixes the failure of ragged keys at key.view(-1, 0), which now returns an empty result.

Sources: https://github.com/tile-ai/TileOPs/pull/3045 https://github.com/tile-ai/TileOPs/pull/3044 https://github.com/tile-ai/TileOPs/pull/3041


I. Core Project Progress

1.1 Layout placeholder variables remain bound across serialization; spawn subprocess offloading no longer reports non-injectivity

Two PRs were merged within the window on the main tilelang repo. The first is #4410, which touches 3 files, +301/-2. It fixes layout objects becoming invalid after round-tripping across processes and across disk.

tl.Layout, tl.Fragment, and tl.PartialFragment use in-process singleton tirx::Var placeholders to express index and thread mappings. _i, _j, etc. use InputPlaceholder(i), and _rep uses ReplicationPlaceholder().

Forward, OutputShape, DetectInjective, and ForwardThread base their decisions on Var object identity, not name_hint.

Once a layout or a PrimFunc carrying T.annotate_layout goes through pickle or tvm.ir.save_json and load_json, deserialization rebuilds the placeholders. The newly created VarNode is no longer identical to the singleton in the receiving process.

Three errors reproduce directly on main. After a pickle round-trip, map_forward_index([7]) returns [_i] instead of [7]. OutputShape() computes (32,) as (_i + 1,). Offloading an annotated PrimFunc in a spawned worker process fails with the forward map must be injective.

The fix routes the bound placeholder Var references through the same ObjectGraphSerializer channel. Two fields are involved: _forward_vars on LayoutNode and _replicate_var on FragmentNode. Both are serialized before forward_index and forward_thread. These fields do not participate in repr, init, compare, hash, or structural equality, and default to nullptr.

The practical effect is that the multi-process compilation path now works.

Source: https://github.com/tile-ai/tilelang/pull/4410

1.2 Dev version number drops toolchain tag; the same wheel no longer forks by CUDA version

#4409 touches 7 files, +16/-76. It cleans up the backend suffix in the version number.

The cause: TileLang used to build separate wheels for CUDA 12.8 and 13.0. Since #1826, a single wheel serves multiple CUDA toolchain versions. The old suffix was therefore misleading: a checkout built in a CUDA 13.0 image would report 0.1.15+cu130.gitabcdef12, yet the build machine’s toolchain version says nothing about which backends the wheel supports.

The change removes the backend suffix and the NO_TOOLCHAIN_VERSION switch. The same checkout now uniformly reports 0.1.15+gitabcdef12 across configurations. Setting
TILELANG_BUILD_WHEEL_WITH_DATE=ON reports 0.1.15+d20261119.gitabcdef12. PyPI releases still use the base version number.

tilelang.build_date() is updated to accept both positions. The date can follow + directly, and the old wheel form such as 0.1.15+cuda.d20261119.gitabcdef12 is also supported.

The author lists the verification scope. Three packaging tests pass; PEP 517 metadata was checked for CUDA 12.8, CUDA 13.0, ROCm, CPU, dated nightly builds, and release builds; pre-commit passes in full.

Source: https://github.com/tile-ai/tilelang/pull/4409

1.3 Newly reported defect: MLA decode example lacks write-after-read synchronization between two warpgroups

#4412 is a newly filed defect report pointing at examples/deepseek_mla/example_mla_decode_ws.py. The report says the right warpgroup reads sum_exp_shared without waiting for the left warpgroup’s write.

The file is the warp-specialized MLA decode example. The report argues that the shared-memory handoff between the two halves relies only on return values or implicit ordering, which is insufficient for synchronization.

The report has no fix attached and no reproduction environment noted. This issue is recorded as “reported, pending” and is not cited as a fixed issue.

Source: https://github.com/tile-ai/tilelang/issues/4412

1.4 Fix under review: skip TMA promotion for layout-incompatible copies under warp specialization

#4413 fixes a lowering failure. On b1367a59, a pipelined GEMM targeting sm_90a fails to lower. The trigger is that its 8-bit B tile uses a padded layout via make_swizzled_layout(Bs, k_major=False).

Loading A makes the loop eligible for warp specialization. The pass then also promotes B’s ordinary T.copy to T.tma_copy, and LowerBulk rejects the padded layout (corresponding to issue #4053).

The fix checks the destination layout annotation of each candidate TMA load, covering the pipeline and its prelude. Copies with incompatible layouts stay on the consumer side, while compatible ones are still promoted to TMA within the same pipeline.

The author reports that all eight cases in #4053 now lower for sm_90a. A new FP8 regression on H200 with CUDA 13.0 confirms A goes through TMA and B through an ordinary copy. The PR is still under review.

Source: https://github.com/tile-ai/tilelang/pull/4413

II. Multi-Backend Adaptation (Ascend / Hygon / Moore Threads / MetaX)

2.1 MLIR Ascend merges tl.or, tl.xor, tl.pow in one day; vsigmoid and sort under review

Merges in the MLIR Ascend repo continue to land on the mlir-dialect branch. That branch’s main has not moved since 11-04. Three TLIR operators were merged on 11-19, all between UTC 14:08 and 14:37.

#225 adds tl.or, #228 adds tl.xor, #229 adds tl.pow. All three follow the same flow: add the dialect operator, lower from TIR, then convert to HIVM, each with tests. The first two were still listed as under review in the previous report and are now merged.

On the review side, #264 was newly opened to add tl.vsigmoid to TLIR. It was opened 11-20 01:35 and updated once the same day. Separately, #232 lowers the TLIR sort operation to HIVM; it was updated again within the window but not merged.

The three merged PRs and the two under review form one batch of operator-by-operator groundwork.

Source: https://github.com/tile-ai/tilelang-mlir-ascend/pull/229

2.2 Ascend main repo shifts from kernels to release: shared wheels on OBS, nightly NPU smoke tests, tuning skill

The Ascend main repo had no merges in the window. Five PRs under review were updated, with the focus shifting from kernels to release and toolchain.

#4411 was newly opened 11-19 19:15 to publish standard Linux wheels to the Ascend community OBS. The premise is that these wheels already bundle Ascend, CUDA and ROCm support, whereas the existing Dist workflow only builds for x86_64 and aarch64 and does not upload to OBS.

The approach reuses the abi3 wheels from build-wheels. The abi3 tag must pass abi3audit --strict. Once the build passes, both architectures are uploaded on the existing daily schedule and on manual trigger, with filenames carrying a date and commit suffix so nightly builds can coexist. PyPI release version numbers are unchanged. The PR also requires serializing OBS uploads and index updates to prevent concurrent runs from overwriting the index with a stale manifest.

#3924 is the nightly Ascend wheel pipeline with NPU smoke tests and rolling releases, opened 11-11 and updated again in the window. #4016 is a TileLang performance tuning skill written for Ascend, and #4012 changes the max_vector_bits default to 128 and adds support for 16-lane fp8 SIMT vectors. Both were opened 11-13.

Separately, #1760 adds FP32 GEMM support to the AscendC backend. It was opened 9-08 and updated again at 22:36 within the window.

Source: https://github.com/tile-ai/tilelang/pull/4411

2.3 Atlas A3 misidentified as 910A; carver undercounts L0C capacity by half

#2023 fixes carver’s model detection. Ascend.__init__ identifies chips by substring matching: it checks 910B first, then 310P, and the fallback 910 is always treated as Ascend910A.

The Atlas A3 device name is Ascend910_9362, as written in examples/situ_glu/benchmark.md in the repo. It matches neither 910B nor 310P, so it falls into the fallback bucket.

The result is three mismatched hardware parameters. ub_cap is set to 262144 bytes, but the actual value is 196608 bytes. l0c_cap is set to 262144 bytes, but the actual value is 524288 bytes — exactly half. compute_max_core takes the 910A default of 32, whereas the actual value should come from a cube_core_num query.

Two contradictory mappings coexist in the same codebase. The compiler’s main path, tilelang/utils/target.py, classifies the same device as 910_93/910C → A3; carver uses the wrong one.

carver uses these three values for feasibility screening and cost modeling of candidate configurations (tilelang/carver/roller/policy/ascend.py). Undercounting L0C by half will screen out configurations with large C tiles outright. The PR was opened 11-19 08:14 and is still under review.

Source: https://github.com/tile-ai/tilelang-ascend/pull/2023

2.4 Hygon adds PR quality-gate workflow; no kernel changes in the window

The Hygon repo had only one merge in the window (#42, 1 file, +19/-0). It adds a quality-gate workflow for pull requests. The PR has no commit message body, and the change touches only CI configuration. #43 and #44 are versions of the same workflow under review.

After the twelve intrinsic fixes in the previous report (#27), the Hygon side had no kernel, precision or codegen changes this period.

Source: https://github.com/tile-ai/tilelang-hygon/pull/42

2.5 No commits from MetaX or Moore Threads in the window

The MetaX repo tilelang-metax last pushed on 11-06. The Moore Threads repo tilelang-musa last pushed 11-18 03:47 (UTC). Both fall outside the window. MetaX’s release/v0.1.12-musa.5-source branch and release proposal saw no new activity in the window.

II-b. RISC-V Backend and Community (Special Focus: TVM Route / MLIR Route / SOPHGO TPU / Community Index Page)

2.5.1 Four-source scan shows no change; both routes and the index page match the previous period

Four sources were scanned on 11-20. The first is the kubuds- prefixed branch of kubuds/tilelang-riscv-kubuds, i.e. the TVM backend route. The second is main of RuyiAI-Stack/tilelang-riscv, i.e. the MLIR rewrite route. The third is the community index page at TongmingLAIC/tilelang-riscv. The fourth is main and new branches of TongmingLAIC/tilelang-tpu, i.e. the SOPHGO TPU extension.

References for all four sources match the previous period’s record: no new commits, no new PRs, and no additions or removals in the index page’s project table. These conclusions cover only the four sources above and do not extend to other RISC-V work. Neither the TVM route nor the MLIR route saw new activity this period, and no further interpretation is offered.

Source: No new commits.

III. Ecosystem and Adopters (TileOPs Operator Library / TileFoundry / TileKernels)

3.1 Attention sink completed: DSADense and DSAPaged merged on the same day

Two attention merges complete the sink and bf16 cache rows used by DeepSeek-V4 and GLM-5.3.

On the DSADense side is #3038, touching 5 files, +228/-82. Its SM90 kernel now supports attn_sink and is_causal=False. In this case keys are selected by index alone, where -1 means skip that slot, and each head’s sink logit is counted into its softmax denominator.

topk is no longer required to be a multiple of 128. The excess becomes dead slots; in the last pair of blocks, if the odd block lies entirely beyond topk, only its even block is loaded and scored. The FlashMLA baseline now runs all kv rows of the batch in a single launch, with its indices padded to a multiple of 128 using -1.

On the DSAPaged side is #3048, touching 6 files, +411/-79. It supports bf16 cache rows with 512 value columns and 0 or 64 key columns, covering GLM-5.3 and DeepSeek-V4.

The single-CTA kernel gathers with a merged cp.async. It supports arbitrary topk by masking the tail of the last block, and adds attn_sink into the softmax denominator, also counted when merging blocks.

The scope is stated clearly. The FP8 656-byte row generation code is unchanged, and that row still does not support attn_sink or a topk that is not a multiple of 64. SM89 does not support bf16 rows.

Sources: https://github.com/tile-ai/TileOPs/pull/3038 https://github.com/tile-ai/TileOPs/pull/3048

3.2 FP4 grouped quantization lands; MoE gating gains hash routing and sqrt-softplus

On the quantization side, FP4QuantPerGroupFwdOp is added (#3027, 15 files, +810/-41). Its signature is block_size=32, inplace=False, quantizing bfloat16 into e2m1 codes, two codes per byte, with block scales as powers of two in float8_e8m0fnu. The output is bit-for-bit identical to DeepSeek-V4-Flash’s fp4_act_quant.

Rounding does not use the FP4 conversion instruction. Instead it does integer arithmetic on the float32 bits, followed by one float32 addition. The kernel shares group_quant.py with
FP8QuantPerTokenGroupFwdOp.

write_in_place lets strided rows whose group count exceeds the grid’s y range go through a contiguous copy first. This incidentally fixes the inplace=True path of FP8 per-token grouped quantization.

On the MoE side is #3036, touching 4 files, +360/-66. It adds scoring_func="sqrtsoftplus" to FusedTopKFwdOp, with or without correction_bias and grouping. It also supports caller-supplied expert_ids, covering DeepSeek-V4-Flash’s gating and the first layer of its hash routing.

When expert_ids is given, the kernel reads only the logits of those experts. The un-renormalized softmax reads the whole row and does not read the bias. The combination still rejected is “softmax selection with bias.”

Sources: https://github.com/tile-ai/TileOPs/pull/3027 https://github.com/tile-ai/TileOPs/pull/3036

3.3 GatedPoolCompress completes per-request ring KV pooling in a single launch

#3033 touches 12 files, +1363/-3. It implements GatedPoolCompressFwdOp, corresponding to DeepSeek-V4-Flash’s compressor and GLM-5.3-Flash indexer key pooling.

Requests are processed packed. Positions before a request’s start come from that request’s ring, whose size is C * ratio rows. The call writes out the rows of its last several positions and allows rolling back the last token of the previous call.

The implementation uses a single launch with no host synchronization. With one block per request, it first pools the windows that read its ring, then writes back the ring and the valid flag. The remaining windows locate their owning request via a warp-level ballot search over cu_seqlens.

This change also adds a batched torch_program torch baseline for benchmarking, and declares kv_state and score_state as contiguous.

Source: https://github.com/tile-ai/TileOPs/pull/3033

3.4 TileFoundry incorporates sliced meshes and Varying values into parallel-plan analysis

One merge within the window for TileFoundry (#356, 123 files, +2752/-557). It changes memory and traffic analysis.

The trigger was discussion #350. The analyzer needs to compare PP, TP, EP, and HSDP parallel plans. The old implementation summed per-GPU peaks by pipeline stage, recorded Reshard between sliced meshes as a zero-byte view, and dropped operands carrying Varying from operators that take only one operand type.

After the change, type inference reads operands’ sharding attributes through a single rule (operand_types, combine_operand_attrs). Broadcast meeting Varying yields Varying’s per-cell storage, while Partial meeting Varying is rejected.

Reshard between sliced meshes counts one copy per “number of cells that obtain the value.” Split or Partial movement between partially overlapping slices is rejected. Peaks are now counted by owning cell.

There are three supporting changes. JSON call lines state the callee (callee, callee_ref) and the Module that defines it. A call yields one output boundary per tensor leaf, nested tuples included, so the generated Module can be analyzed. repeat_interleave preserves the input’s Broadcast mesh.

This change adds CUDA cases comparing PP, TP, EP, HSDP, and FSDP. The generated modules are written out by the CLI, reloaded, checked against autodiff results, and measured.

Source: https://github.com/tile-ai/TileFoundry/pull/356

3.5 Four new goals opened: DSAPaged bf16, AvgPool1d, DeltaNet, Mamba

Within the window, TileOPs opens four work items with shapes and target multipliers.

#3049 requires DSAPaged’s bf16 row to match FlashMLA on DeepSeek-V4-Flash. This corresponds to #3048 merged in this issue, following the order of “set the goal first, then merge the implementation.”

#3051 handles edge-truncated AvgPool1d windows. The approach reads data into registers and issues only one block per SM when blocks can be saved.

#3054 requires DeltaNet’s chunk input to accept strided tensors. #3013 adds dilated convolution and float32 filters to CausalConv1dFwdOp and CausalConv1dUpdateFwdOp.

There is also #3052, a maintenance merge that only adds comments and docs. It explains why W4A16’s metadata window only applies to a single consuming warpgroup, and writes the selector’s 128, 256, and 512 wide K tiles and explicit 384 wide K tile into the docstring.

This change also adds a regression test locking down the group-count check. The author gives the consequence of dropping that check: the explicit 384 wide configuration returns wrong results, with a relative error of 0.457 on 1x128x36864, split 24.

Source: https://github.com/tile-ai/TileOPs/issues/3049

3.6 No commits for TileKernels and FlashQLA within the window

Neither adopter repo falls within the window. deepseek-ai/TileKernels last pushed on 10-14, and QwenLM/FlashQLA on 10-30.

On the docs site side, TileOPs.github.io has 0 commits in the window. tilelang.github.io has one Update docs (11-19 13:45). Neither constitutes technical progress; they are recorded only as site maintenance.

Source: https://github.com/tile-ai/tilelang.github.io/commit/dc1d73c9

IV. Community, Tutorials, and Events

4.1 News-side gap: five sets of Chinese and English queries find no relevant coverage

Automated collection pulled 4 candidates within the window. After checking each, none is usable. Three are the repos’ own page entries (tilelang, TileFoundry, tilelang-hygon), a case of the crawler treating repo metadata as news. The fourth is a riscv.org Q&A about Infineon AURIX RC1, unrelated to TileLang.

Five supplementary Google News queries were run. The English TileLang when:2d, "TileLang" compiler when:3d, and tile-ai TileOPs when:3d all returned empty. The Chinese TileLang when:2d was empty, and TileLang 算子 when:3d hit only a university Challenge Cup award writeup, unrelated to the topic. Hacker News has no entries within the window.

Source: No relevant coverage.

4.2 Organization activity: 23 merges in a single day, main repo at 9,006 stars

Of the organization’s 30 repos, 8 were pushed within the window. The top five are tilelang, TileOPs, tilelang-mlir-ascend, TileFoundry, and tilelang-hygon. The last three are tilelang.github.io, TileOPs.github.io, and TileOPs-nightly.

23 merges, by repo: TileOPs 16, tilelang 2, MLIR Ascend 3, TileFoundry 1, Hygon 1. 20 in review and updated, concentrated in newly opened bug fixes and the Ascend release pipeline.

7 new issues. Of these, TileOPs has 2 (one performance goal, one MoE defect), TileFoundry 3, topa 1, tilelang 1. The main repo has 9,006 stars, up 5 from 9,001 in the previous issue.

Source: https://github.com/tile-ai/tilelang

V. Trend Observations

  • Interface breaks come with a handoff: on 11-18 the two old mHC operators were deleted and logged as an API break, and on 11-19 all three replacement operators were merged. The replacement PR was opened before the deletion, so the intermediate state lasted only one day.
  • Quantization alignment is measured by bit-for-bit identity: FP4 grouped quantization requires bit-for-bit equality with DeepSeek-V4-Flash’s fp4_act_quant. The FP8 path shares the same group_quant.py. The acceptance criterion is bit-level equality, not an error band.
  • Performance optimization shifts from per-shape tuning to selector rules: #3045 covers 302 model shapes with a single rule and requires that none slow down. That change self-reports as still short of target, showing the acceptance line is the target multiplier, not the speedup relative to main.
  • The two Ascend lines have a clear division of labor: the MLIR route keeps laying down dialects operator by operator, three in a day; the main repo turns to release and CI engineering, with OBS staging, nightly NPU smoke tests, and tuning skills advancing in the same week. The former expands the expression surface, the latter lowers the barrier to use.
  • TileFoundry moves from “can compile” to “can quantify”: PP, TP, EP, and HSDP enter the same analyzer for comparison, and Reshard between sliced meshes starts counting copy cost. Once the cost model is trustworthy, it will in turn influence front-end choices.

Appendix: Sources and Verification Notes

Source Verification Table

Source Verification Result
GitHub tile-ai per-repo check 30 repos, 8 pushed within window; main branch actual additions: TileOPs 16, tilelang 2, TileFoundry 1, Hygon 1
MLIR Ascend branch check 3 merges into mlir-dialect; its main untouched since 11-04
Org-level merge ground-truth search 23 merges: TileOPs 16, tilelang 2, MLIR Ascend 3, TileFoundry 1, Hygon 1
Org-level open-PR search 20 open and updated, concentrated in Ascend release pipeline, layout and TMA fixes, new operator proposals
New issues 7: TileFoundry 3, TileOPs 2, topa 1, tilelang 1
Releases and tags tilelang v0.1.15, TileOPs v0.1.0, MLIR Ascend v0.1.2.020; no new tags within window
News side All 4 candidates unusable; five Google News queries (Chinese and English) returned no coverage; no HN entries
Adopter repos TileKernels last push 10-14, FlashQLA 10-30, both outside window
Docs sites tilelang.github.io one “Update docs” commit; TileOPs.github.io 0 commits within window
TileLang×RISC-V four sources No change, consistent with prior issue’s archive

Full Source List

Verification Notes

Scope: the window is Beijing time 11-19 07:00 to 11-20 07:00. Repo commits are counted by committer time, compared against the prior issue’s cutoff before being judged as new. MLIR Ascend merges land on the mlir-dialect branch; main has been untouched since 11-04, so the body names the branch explicitly and does not call it the main branch.

TileOPs’ 16 merges align item-by-item with the 23 total from the org-level search, with no double counting. Open items are annotated with both their open date and latest update date, to avoid presenting old proposals as newly opened that day. #1760, #3924, #4012, and #4016 were all opened before this window; the body states their open dates and counts only in-window progress.

Collection hit unauthenticated GitHub API rate limits late in the window. Branch and tag information was re-verified via Atom feeds and the search API; conclusions were unaffected. The RISC-V four-source sweep returned empty output, treated as no change rather than a collection failure. The news side was confirmed as a blank window after five query groups; the empty result was not inferred to be a collection failure.