[Feat] beam: Add Beam model, runtime support and checkpoints (4/7) by Jiminator · Pull Request #45256 · sgl-project/sglang

Jiminator · 2026-10-19T23:54:57Z

Motivation

Split #44747 into reviewable parts. This part adds the runtime support Beam relies on and the Beam model, position 4 of 7 in the merge order below. It folds in #45255, which is closed in its favor, the way Muse Spark's runtime and model landed together. It is based on #45254's branch (beam-stack-02-kernels), so its diff shows only this part.

Modifications

Runtime support:

TRT-LLM MHA. An extend batch whose trailing requests extend by one token runs that tail on the decode kernel. This now also covers FP8 KV and batches made only of one-token requests. It extends [Perf] Route BF16 mixed-batch decode tails through decode attention #43238 and runs on trtllm-gen (SM100) only; SM90 and SM120 take FMHA-v2 or the context-parallel path and never reach it.
FlashInfer MegaMoE. Beam is added to the validated MegaMoE architectures. The startup autotune fix that the high-throughput recipe needs with default autotune, and the optional decode workspace, are in [Fix] flashinfer-megamoe: Fix startup autotune on FlashInfer 0.7.2 and add an optional decode workspace #46213.
PD decode. A token-ID logprob row is reserved for the handoff token, so streamed rows stay aligned. That row is empty, because prefill does not send token-ID logprobs.
Forward-pass metrics. DFlash is added to the spec decode pre-settlement snapshot.
Pool configurator. The SWA request-cap budget now includes the DFlash draft cell.

Beam model and config:

Hybrid sliding-window/full attention, SSMax query scaling, attention output gating, sandwich norms, depth-scaled residuals and grouped MoE routing.
BF16, ModelOpt NVFP4 and compressed-tensors FP8 checkpoints. Compressed-tensors NVFP4 is rejected.
DFlash integration. Beam uses the general CUDA attention-DP implementation from merged [Feat] spec: Support CUDA attention DP for DFlash and DFlash2 #46330 and draft placement from merged [Refactor] spec: Resolve draft parallel placement from model declarations #46290. This PR adds the DFlash2BeamDraftModel alias, Beam's embedding/lm-head DP mapping, and hidden-state capture. The generic DFlash admission hook, worker, embedding helper and verify-metadata path match the parent branch.
GB300 Triton MoE configs for E=256,N=640, used by the BF16 Triton MoE path.
Opt-in single-token MoE GEMV (SGLANG_OPT_BEAM_DECODE_GEMV).
- It uses the unquantized Triton MoE path and requires SGLANG_OPT_FUSE_SWIGLU_INTERLEAVED=1, BF16 experts of the Beam shape, MoE TP4/EP1 and SM103. Each invocation also needs one contiguous BF16 token.
- Other setups, and forwards with more than one token, keep their normal MoE path. Only a failed static check logs that the switch is unused.
- It is not bit-exact with the Triton path.

Beam's router uses the tiny FP32 GEMM up to 16 tokens for BF16 checkpoints, then the GEMV, then the cuBLAS BF16x3 GEMM. The cuTile router and SGLANG_BEAM_ROUTER_BACKEND are in #45361.

General fixes found while adding Beam are in their own PRs:

[Fix] flashinfer-megamoe: Fix startup autotune on FlashInfer 0.7.2 and add an optional decode workspace #46213: MegaMoE startup autotune on FlashInfer 0.7.2, and an optional decode workspace;
[Fix] flashinfer: Apply the sliding window in ragged prefill for hybrid-SWA models #46214: the FlashInfer ragged prefill sliding window;
[Fix] swa: Start the SWA KV pool at its first layer under pipeline parallelism #46200: the SWA KV pool start layer under pipeline parallelism.

The only test change is DFlash in test_forward_pass_metrics.py's snapshot loop. Without the metrics change, Beam with DFlash and --enable-forward-pass-metrics fails during startup warmup with TypeError: 'NoneType' object is not iterable.

The Beam E2E test is in #45257. The unchanged upstream DFlash admission, embedding and metadata tests come from #46330.

LoRA serving for Beam is not part of this split. Beam raises NotImplementedError for --enable-lora and --lora-paths when the model is built. LoRA serving will follow separately.

The split commits carry the Co-authored-by trailers of the Beam commits that touched these files.

Validation

Current DFlash integration cleanup (903d45e17d): 58 tests passed across seven existing DFlash/draft-placement files, with one existing CUDA-only test skipped in the CPU-only run. The nine admission/embedding/metadata cases pass unchanged. No test code was added or modified. The two remaining generic DFlash overrides were removed so this branch uses #46330's implementation.

Historical GPU/model results: the runs below predate this cleanup and were not repeated for it.

All runs are on 4×GB300 with FlashInfer 0.7.2. The tree is this PR plus #43216, which has since merged.

Checkpoints: the public reflection/Beam-1.0-501B-A23B-NVFP4, reflection/Beam-1.0-501B-A23B-FP8 and reflection/Beam-1.0-501B-A23B checkpoints, plus the DFlash2 draft.
Evaluation: GSM8K with sgl-eval 0.1.3, 1,319 questions, temperature 0.
Revisions: all runs are on earlier commits of this branch.
- Those commits also carried the FlashInfer ragged sliding-window fix (now [Fix] flashinfer: Apply the sliding window in ragged prefill for hybrid-SWA models #46214), the MegaMoE changes (now [Fix] flashinfer-megamoe: Fix startup autotune on FlashInfer 0.7.2 and add an optional decode workspace #46213) and the SWA pool start-layer change (now [Fix] swa: Start the SWA KV pool at its first layer under pipeline parallelism #46200). None of the runs uses the FlashInfer attention backend, the decode workspace or pipeline parallelism.
- The high-throughput rows with default autotune need [Fix] flashinfer-megamoe: Fix startup autotune on FlashInfer 0.7.2 and add an optional decode workspace #46213's autotune fix.
- Since these runs, the branch has merged newer main changes through [Feat] kernels: Beam model support #45254, including [Feat] spec: Support CUDA attention DP for DFlash and DFlash2 #46330 and [Refactor] spec: Resolve draft parallel placement from model declarations #46290. The DFlash cleanup removes the earlier Beam-only admission check and the broader selector-warmup override.
Cold-JIT row: this run deliberately hid the AOT top-k artifact and removed its cached build.

Configuration	GSM8K	Notes
Cookbook low-latency NVFP4 (TP4, DFlash, without parser flags)	96.36%	zero request errors; the DFlash2 selector is folded into the draft CUDA graph
Cookbook low-latency FP8	96.66%	zero errors
Cookbook high-throughput NVFP4 (MegaMoE TP4/DP4/EP4), cold autotune cache	96.44%	with #46213's autotune fix; MegaMoE autotune completes
High-throughput NVFP4 plus DFlash, default autotune	96.44%	with #46213's autotune fix; zero errors
The same with autotune off	96.44%	zero errors
BF16 TP4, attention DP2, DFlash	96.44%	concurrent outputs identical to isolated in the probe
BF16 TP4 with the decode GEMV	96.21%	the GEMV kernels run at batch size 1
Cold FlashInfer JIT (no AOT top-k), DFlash with attention DP4	96.51%	the selector's top-k builds during startup; without the warm-up, the first decode builds it inside the scheduler watchdog and the server is killed

Verifier. With [Feat] parser: Add Beam reasoning, tool calls and response templates (5/7) #45257's parser changes on top of this branch, the pinned Beam verifier dfa85fe771b3bcf014f5af5913f2ed363d31dfdf passes all 133 required e2e cases at temperature 0, with its source and fixture integrity check passing and no skips.
- The run uses the cookbook low-latency recipe with parser flags and scores 96.59% on GSM8K with zero request errors.
- Chat reasoning, streaming, a chat tool call, Responses text, a Responses function call, a Responses shell round trip and /tokenize all pass.
- No optional verifier suite or third-party evaluation was run.
TRT-LLM mixed tail. The mixed-tail cases from [Feat] beam: Add native Beam serving support #44747 pass (16 passed, 2 skipped on SM100). They include batches of only one-token requests, split-versus-unsplit parity, long-context FP64 references and CUDA-graph replay.
GEMV numerics. Both kernels were compared with an FP64 reference that rounds to BF16 where the kernels do: after the gate/up projection, after silu(gate) * up, and at the routing-weighted output.
- On 8 real checkpoint experts, the GEMV's BF16 output matches the reference on nearly every element, with relative L2 below 2e-9.
- In the tested random-input cases, the GEMV is within relative L2 3.4e-4.
- The Triton deferred-sum path is within 4.2e-4 to 6.8e-4 of the same reference, and the GEMV versus Triton is within 6.3e-4.
- With a fixed kernel config, the shared Triton MoE path is bitwise unchanged from the base.
Existing GPU tests. Main's tests for the touched code pass unedited: TRT-LLM MHA dense, graph metadata, encoder-only and stride parity; Triton MoE padding and fused MoE; and the DFlash buffer, Domino and graph-pool tests.
CPU tests. The CPU tests that import the touched modules pass, except three files that fail the same way on main in this environment. pre-commit passes.
Known behavior. Under attention DP4, a request's greedy output can change with other ranks' load.
- This was observed with MegaMoE, with the TRT-LLM MoE runner without MegaMoE, without DFlash, and on the previously published head.
- In the probe's valid comparisons, the isolated run's top-2 logprob margin at the first differing token ranged from 0 to 0.262 nat.
- Isolated requests were deterministic in repeat probes, and these runs showed no GSM8K accuracy regression and zero request errors.
- This does not establish universal batch invariance or a specific kernel-level cause. BF16 with attention DP2 matched isolated outputs exactly.
Not validated here:
- NPU;
- Hopper and SM120 (all runs are on GB300);
- BF16 on two nodes (TP8), including the TP8 cluster path of [Feat] kernels: Beam model support #45254's all-reduce boundary;
- serving with PD token-ID logprobs (that fix has CPU checks only).

Merge order

Each part is based on the one before it, so its diff shows only its own changes. Merge bottom-up; after each part merges, the next part is retargeted to main and main is merged into it. #44747 stays open, unchanged, as the full reference. Merge in this order:

[Chore] docker: Add Beam CUDA image (1/8) #45253 — Beam CUDA Docker image (merged)
[Chore] deps: Bump the OpenAI SDK to 2.8.0 for Responses shell tools (2/8) #45291 — OpenAI SDK 2.8.0, needed by the Responses shell tools in [Feat] parser: Add Beam reasoning, tool calls and response templates (5/7) #45257 (merged)
[Feat] kernels: Beam model support #45254 — Kernels
[Feat] beam: Add Beam model, runtime support and checkpoints (4/7) #45256 — Beam model and runtime integration (using upstream [Feat] spec: Support CUDA attention DP for DFlash and DFlash2 #46330 for DFlash CUDA DP) ([Feat] srt: Add attention, memory and DFlash DP support for Beam (4/8) #45255 folded in; also after [Fix] flashinfer-megamoe: Fix startup autotune on FlashInfer 0.7.2 and add an optional decode workspace #46213 for the high-throughput recipe with default autotune; [Fix] nccl: Disable graph buffer registration whenever a captured graph can hold a collective #43216 is merged)
[Feat] parser: Add Beam reasoning, tool calls and response templates (5/7) #45257 — Beam reasoning, tool calls, response templates and the Beam E2E test ([Feat] parser: Enrich chat_parsing streaming events #40478, [Fix] parser: Accept whole-number floats for integer tool arguments #43941 and [Feat] parser: Add checkpoint-driven response-template adapters #40479 are merged)
[Chore] deps: Add nvidia-cuda-tileiras for the optional cuTile router #45290 — nvidia-cuda-tileiras, needed by [Perf] beam: Add optional cuTile BF16x3 router GEMM #45361
[Perf] beam: Add optional cuTile BF16x3 router GEMM #45361 — cuTile BF16x3 router GEMM (optional: without it Beam uses the cuBLAS router)

#45292 (Dynamo and grpcio-tools pins for the Beam image) is not in the merge order for now.

CI States

Latest PR Test (Base): 🚫 Run #40237989506
Latest PR Test (Extra): 🚫 Run #40237989484
Latest PR Test (AMD ROCm 10): 🚫 Run #40237989472

Add the shared runtime support Beam relies on: TRT-LLM MHA and FlashInfer attention fixes, unified hybrid SWA memory, DFlash with attention DP, FlashInfer MegaMoE all-to-all under attention DP, and their tests. Split from #44747. Co-authored-by: BBuf <1182563586@qq.com> Co-authored-by: Caihua Li <caihua.li@bytedance.com> Co-authored-by: Florent Gbelidji <florent@reflection.ai> Co-authored-by: Florent Gbelidji <florentgbelidji@Florents-MacBook-Pro.local> Co-authored-by: Florent Gbelidji <florentgbelidji@florents-macbook-pro1.home> Co-authored-by: Florent Gbelidji <florentgbelidji@Host-002.lan> Co-authored-by: iaroslav-reflection <iaroslav@reflection.ai> Co-authored-by: Jiajun Li <jiajun.li@radixark.ai> Co-authored-by: Jinze Xue <jinzex@nvidia.com> Co-authored-by: Kaiwu <kai@reflection.ai> Co-authored-by: mloh <mloh@nvidia.com> Co-authored-by: Neil Gollapudi <neil@reflection.ai> Co-authored-by: neilkg <33635204+neilkg@users.noreply.github.com> Co-authored-by: Qiaolin-Yu <liin1211@outlook.com> Co-authored-by: Renault Young <renault@reflection.ai> Co-authored-by: thibautbar <t.barroyer@icloud.com> Co-authored-by: Yoni Gozlan <yoni@reflection.ai> Co-authored-by: zhihengy <andy.ye@radixark.ai>

Add the Beam model and config, BF16/ModelOpt NVFP4/compressed-tensors FP8 checkpoint support, the DFlash2 Beam draft alias, GB300 MoE tuning, and the registered unit and end-to-end tests. Beam rejects LoRA adapters here. Split from #44747. Co-authored-by: AymenKallala <aymen.kallala@gmail.com> Co-authored-by: BBuf <1182563586@qq.com> Co-authored-by: Florent GBELIDJI <florent@reflection.ai> Co-authored-by: Florent Gbelidji <florent@reflection.ai> Co-authored-by: Florent Gbelidji <florentgbelidji@Florents-MacBook-Pro.local> Co-authored-by: Florent Gbelidji <florentgbelidji@florents-macbook-pro1.home> Co-authored-by: Florent Gbelidji <florentgbelidji@Host-002.lan> Co-authored-by: Florent Gbelidji <florentgbelidji@Mac.lan> Co-authored-by: iaroslav-reflection <iaroslav@reflection.ai> Co-authored-by: Jinze Xue <155670984+jinzex@users.noreply.github.com> Co-authored-by: Jinze Xue <jinzex@nvidia.com> Co-authored-by: kai <kai@reflection.ai> Co-authored-by: mloh <mloh@nvidia.com> Co-authored-by: mreso <13337103+mreso@users.noreply.github.com> Co-authored-by: Qiaolin-Yu <liin1211@outlook.com> Co-authored-by: Renault Young <renault@reflection.ai> Co-authored-by: thibautbar <t.barroyer@icloud.com> Co-authored-by: Yanbin Jiang <jybsuper@gmail.com> Co-authored-by: Yoni Gozlan <yoni@reflection.ai>

Restore the router dispatch from before the cuTile BF16x3 GEMM: the tiny FP32 GEMM up to 16 tokens for BF16 checkpoints, then the GEMV, then the cuBLAS BF16x3 GEMM. SGLANG_BEAM_ROUTER_BACKEND and the router dispatch tests move with the cuTile router.

This reverts commit 472393e.

This reverts commit 3ca2455.

Fridge003 · 2026-10-23T22:50:49Z

    }
    fake_modules["sglang.srt.environ"].envs = types.SimpleNamespace(
        SGLANG_FLASHINFER_CUTEDSL_NVFP4_W4A16=types.SimpleNamespace(get=lambda: False),
        SGLANG_FLASHINFER_MEGAMOE_MAX_TOKENS_PER_RANK=types.SimpleNamespace(


For the extra unit tests, please remove all the unnecessary ones.
We need to add an e2e test for Beam (maybe at extra groups or nightly groups)

Removed the extra unit tests.

The one left adds DFlash to the snapshot loop in test_forward_pass_metrics.py. It guards a real crash where Beam’s low-latency recipe with --enable-forward-pass-metrics dies in startup warmup because trtllm_mha doesn’t keep CPU seq lens on spec decode batches.

The Beam e2e test is test_beam_10_b200_pd.py in #45257

Fridge003 · 2026-10-23T23:11:25Z

        )

+    if attn_dp_enabled_of(cfg) and cfg.device != "npu":
+        if not cfg.device.startswith("cuda"):


Seems dp attention + dflash is open to every model. Can we add a guard that throws error when detecting the model is not Beam here

Fridge003 · 2026-10-23T23:15:51Z

            full_cell_size = (
                self._full_cell_bytes()
                + self._swa_per_token * self._draft_swa_full_layers_num
+                + self._draft_cell_size


Is it correct to add draft size here?

Should be correct, the DFlash draft pool reuses the target’s token count, so each full-pool token also costs _draft_cell_size. HybridSWAPoolConfigurator already counts it; the request-cap override dropped it, so the full pool takes the whole budget and the draft pool overflows it.

Fridge003 · 2026-10-23T23:18:26Z

        self.layer_num = self.full_layer_nums + self.swa_layer_nums
-        self.start_layer = 0
+        # PP stages keep global layer ids; HiCache load events count from the first.
+        self.start_layer = min(


Why do this change?

Removed, previously on an old version of the fork, PP=2 + HIcache was broken without it, but its no longer needed. There is a very related fix opened in #46200 for a seperate bug with PP

Fridge003 · 2026-10-23T23:21:34Z

 communication internally via the deep_gemm symmetric buffer, so the dispatcher
 and combine stay pure no-ops; this module owns the layer build + forward.
 """



Can you split the part of megamoe autotune optimziation into a standalone PR, since it's not specific to Beam

Done and split into #46213

Fridge003 · 2026-10-23T23:57:58Z

            )
            if (
                head_dims in cutlass_supported_head_dims
                and not check_cuda_graph_backend(Phase.PREFILL, Backend.TC_PIECEWISE)


Can we revert the changes in this PR? Since beam only applies trtllm mha backend in cookbook.

Done and split into #46214

Beam selects trtllm_mha when no attention backend is set. The fix applies to every hybrid-SWA model on --attention-backend flashinfer, so it moves to a standalone PR.

Neither is specific to Beam: the startup autotune output buffer fixes every NVFP4 MegaMoE model on FlashInfer 0.7.2, and the decode workspace is an opt-in profile for any ModelOpt NVFP4 MoE. Beam stays in the validated MegaMoE architectures.

Only Beam is validated on this path; other CUDA targets are rejected at startup as before. The mRoPE check it subsumes goes away.

…(5/7) (#45257) Signed-off-by: Yoni Gozlan <yonigozlan@users.noreply.github.com> Co-authored-by: Yoni Gozlan <yonigozlan@users.noreply.github.com> Co-authored-by: Cursor <cursoragent@cursor.com> Co-authored-by: Xinyuan Tong <xinyuantong.cs@gmail.com> Co-authored-by: Yoni Gozlan <74535834+yonigozlan@users.noreply.github.com> Co-authored-by: BBuf <1182563586@qq.com> Co-authored-by: Jinze Xue <jinzex@nvidia.com> Co-authored-by: mloh <mloh@nvidia.com> Co-authored-by: Neil <neil@reflection.ai> Co-authored-by: neilkg <33635204+neilkg@users.noreply.github.com> Co-authored-by: thibautbar <t.barroyer@icloud.com> Co-authored-by: Yoni Gozlan <yoni@reflection.ai>

Jiminator and others added 2 commits October 19, 2026 16:50

This was referenced Oct 19, 2026

[Chore] docker: Add Beam CUDA image (1/8) #45253

Merged

[Feat] kernels: Beam model support #45254

Merged

[Feat] srt: Add attention, memory and DFlash DP support for Beam (4/8) #45255

Closed

github-actions Bot added quant LLM Quantization dependencies Pull requests that update a dependency file speculative-decoding hicache Hierarchical Caching for SGLang blackwell SM100/SM120 jit-kernel unified-radix-cache memory-pool labels Oct 19, 2026

Jiminator mentioned this pull request Oct 19, 2026

[Feat] parser: Add Beam reasoning, tool calls and response templates (5/7) #45257

Merged

zhihengy mentioned this pull request Oct 20, 2026

[sglang-miles] [Feat] beam: Port Beam BF16 serving and LoRA #45273

Draft

Jiminator added 2 commits October 19, 2026 19:44

Merge branch 'beam-stack-03-srt' into beam-stack-04-model

a80a6ab

Merge branch 'beam-stack-02-kernels' into beam-stack-03-srt

48c5264

This was referenced Oct 20, 2026

[Chore] deps: Add nvidia-cuda-tileiras for the optional cuTile router #45290

Draft

[Chore] deps: Bump the OpenAI SDK to 2.8.0 for Responses shell tools (2/8) #45291

Merged

[Chore] docker: Pin Dynamo and grpcio-tools in the Beam image #45292

Draft

Jiminator added 3 commits October 20, 2026 01:53

Merge branch 'beam-stack-03-srt' into beam-stack-04-model

2a2b7e3

Merge branch 'beam-stack-02-kernels' into beam-stack-03-srt

422389f

Jiminator mentioned this pull request Oct 20, 2026

[Perf] beam: Add optional cuTile BF16x3 router GEMM #45361

Draft

Jiminator changed the title ~~[Feat] beam: Add Beam model, checkpoints and tests (4/5)~~ [Feat] beam: Add Beam model, checkpoints and tests (5/8) Oct 20, 2026

Jiminator mentioned this pull request Oct 20, 2026

[Roadmap] Beam Development & Optimization #45370

Open

Jiminator added 4 commits October 20, 2026 11:40

Merge branch 'beam-stack-02-kernels' into beam-stack-03-srt

1f36b57

Merge branch 'beam-stack-03-srt' into beam-stack-04-model

cbf000c

Revert "[Test] Set finite capacity in allocation eviction fixture"

fca4b46

This reverts commit 472393e.

Revert "[unified-memory] Free SWA segments without a host sync"

8392f58

This reverts commit 3ca2455.

Jiminator requested review from ByronHsu, Duyi-Wang, ShangmingCai, hebiao064, sogalin, wisclmy0611 and zijiexia as code owners October 22, 2026 06:41

Jiminator added 2 commits October 22, 2026 15:54

Merge branch 'beam-stack-02-kernels' into beam-stack-04-model

3818ff6

Merge branch 'beam-stack-02-kernels' into beam-stack-04-model

f698850

Fridge003 reviewed Oct 24, 2026

View reviewed changes

Jiminator added 3 commits October 23, 2026 17:26

Leave the FlashInfer ragged sliding-window fix to its own PR

8d276ab

Beam selects trtllm_mha when no attention backend is set. The fix applies to every hybrid-SWA model on --attention-backend flashinfer, so it moves to a standalone PR.

Allow DFlash with CUDA dp attention only for Beam targets

d365acd

Only Beam is validated on this path; other CUDA targets are rejected at startup as before. The mRoPE check it subsumes goes away.

Jiminator and others added 4 commits October 23, 2026 19:37

Drop the pool-configurator and Beam override unit tests

6757701

Leave the SWA pool start layer fix to its own PR

1fc8fcd

Merge branch 'beam-stack-02-kernels' into beam-stack-04-model

b8c1956

Fridge003 requested a review from DarkSharpness as a code owner October 25, 2026 05:25

Jiminator mentioned this pull request Oct 25, 2026

[Perf] beam: Add the SM103 single-token BF16 MoE GEMV follow-up #46401

Draft

Jiminator added run-ci CI: run the baseline test suite on this PR run-ci-extra CI: also run the extra suite (requires run-ci) bypass-fail-fast CI: a failing job no longer aborts its siblings (lint still gates) labels Oct 25, 2026

Jiminator added 3 commits October 25, 2026 03:32

Merge beam-stack-02-kernels into beam-stack-04-model

1c9663c

Reuse upstream DFlash CUDA DP support for Beam

903d45e

Merge branch 'beam-stack-02-kernels' into beam-stack-04-model

46196e5

Jiminator merged commit e58899d into beam-stack-02-kernels Oct 25, 2026
34 of 87 checks passed

Jiminator deleted the beam-stack-04-model branch October 25, 2026 22:00

Navigation Menu

Repository navigation

[Feat] beam: Add Beam model, runtime support and checkpoints (4/7) - #45256

[Feat] beam: Add Beam model, runtime support and checkpoints (4/7)#45256
Jiminator merged 66 commits into
beam-stack-02-kernelsfrom
beam-stack-04-model

Jiminator commented Oct 19, 2026 •

edited by github-actions Bot

Loading

Uh oh!

Fridge003 Oct 23, 2026

Uh oh!

Jiminator Oct 24, 2026

Uh oh!

Fridge003 Oct 23, 2026

Uh oh!

Jiminator Oct 24, 2026

Uh oh!

Fridge003 Oct 23, 2026

Uh oh!

Jiminator Oct 24, 2026

Uh oh!

Fridge003 Oct 23, 2026

Uh oh!

Jiminator Oct 24, 2026 •

edited

Loading

Uh oh!

Fridge003 Oct 23, 2026

Uh oh!

Jiminator Oct 24, 2026

Uh oh!

Fridge003 Oct 23, 2026

Uh oh!

Jiminator Oct 24, 2026

Uh oh!

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Conversation

Jiminator commented Oct 19, 2026 • edited by github-actions Bot Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Motivation

Modifications

Validation

Merge order

CI States

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Jiminator Oct 24, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Jiminator commented Oct 19, 2026 •

edited by github-actions Bot

Loading

Jiminator Oct 24, 2026 •

edited

Loading