[ROCm] Fix EAGLE spec-decode verify silently sampling greedy on HIP by xiaobochen-amd · Pull Request #37134 · sgl-project/sglang

xiaobochen-amd · 2026-08-30T13:10:46Z

Problem

EAGLE spec-decode verify has been committing argmax on ROCm, ignoring
temperature and top_p entirely, with nothing in the log saying so. At
temp > 0 it degenerates into repetition loops.

The cause is one line in eagle_sample:

if sampling_info.is_all_greedy or _is_cpu or _is_npu or _is_hip or _is_xpu:
    target_predict = torch.argmax(next_token_logits, dim=-1)

_is_hip sits next to is_all_greedy, so HIP takes the greedy path
unconditionally. It is there because the sampling branch needs
tree_speculative_sampling_target_only, top_k_renorm_prob and
top_p_renorm_prob, all of which are CUDA/MUSA-only in sgl_kernel.

Fix

Route HIP through the pure-Triton chain sampler in reject_sampling.py when
rejection sampling is active and the batch is not all-greedy, with a Triton
renorm (_renorm_top_k_top_p_hip) standing in for the CUDA-only renorm ops.

The gate becomes a pure helper, _verify_uses_greedy. On every non-HIP platform
it reduces to the predicate it replaces: with is_hip False the added term drops
out and the expression is the original one.

_renorm_top_k_top_p_hip follows the same shape as the op it replaces — a
threshold search, matching top_p_renorm_prob's ternary pivot search on
f(x) = sum(probs[probs > x]) — so it agrees with CUDA rather than with a
sort-based nucleus. top_p >= 1.0 short-circuits to an exact pass-through, to
stay bit-exact with CUDA's top_p_renorm_prob(p, 1.0) == p. top-k uses a sort,
skipped entirely when no row restricts it.

Defaulting the flag on matters: without it the fix only reaches users who pass
--speculative-use-rejection-sampling, and everyone else still silently gets
argmax verify. It flips only on HIP, for EAGLE/EAGLE3 with topk=1, default
accept thresholds and non-deterministic inference — the same conditions the
validation below raises on, so the flip can never turn a working config into an
error. Setting it also makes the draft worker emit the target-vocab proposal
(draft_probs) the Triton chain sampler consumes.

Two shapes that come from the current tree, not from the fix

Worth pointing out so they don't read as gratuitous:

sampling_fn dispatch. It used the eager A if cond else B form, which
resolves tree_speculative_sampling_target_only whether or not that branch is
taken. HIP must not import it at all, so the dispatch becomes an if/else
that imports inside the arm needing it. Non-HIP behaviour is unchanged.

One cfg read becomes resolved_view. _handle_eagle_family writes through
declare_resolution, whose contract is that a field it decides is not visible
to a later cfg. read. The rejection-sampling validation immediately below was
reading cfg, so it would have skipped every auto-enabled config — including its
own "rejection sampling is enabled" log. That read becomes
resolved_view(server_args), which is what the declare_resolution docstring
prescribes for exactly this case.

Test

test/registered/unit/spec/test_eagle_gate_routing.py — gate truth table, CPU.
Pins that the helper reduces to the original predicate off HIP, and that HIP
only leaves greedy when rejection sampling is on and the batch is not
all-greedy. 4 tests.
test/registered/spec/eagle/test_rocm_eagle_renorm.py — renorm numerical
equivalence against an independent torch reference, plus the top_p == 1.0
exact-no-op and top_p -> 0 edges. Registered for both CUDA and AMD runners;
the kernel is device-agnostic. 5 tests.

ruff, isort and black clean.

CI States

Latest PR Test (Base): ✅ Run #34937231207
Latest PR Test (Extra): ❌ Run #34937230795
Latest PR Test (AMD ROCm 10): ➖ No AMD PR run found for this commit.

1am9trash

Sorry, I mis-clicked approve. Please check the following comments.

1am9trash

Hi, @xiaobochen-amd

python/sglang/kernels/ops/sampling/renorm_triton.py already provides rocm triton replacements for top_k_renorm_prob / top_p_renorm_prob, and srt/speculative/dflash_utils.py already consumes them via a platform-conditional import that aliases them to the same names:

elif is_hip():
    from sglang.kernels.ops.sampling.renorm_triton import (
        top_k_renorm_probs_triton as top_k_renorm_prob,
    )
    from sglang.kernels.ops.sampling.renorm_triton import (
        top_p_renorm_probs_triton as top_p_renorm_prob,
    )

Should we reuse those instead of adding new implementations?

xiaobochen-amd · 2026-09-07T12:38:26Z

Hi, @xiaobochen-amd

python/sglang/kernels/ops/sampling/renorm_triton.py already provides rocm triton replacements for top_k_renorm_prob / top_p_renorm_prob, and srt/speculative/dflash_utils.py already consumes them via a platform-conditional import that aliases them to the same names:
elif is_hip():
    from sglang.kernels.ops.sampling.renorm_triton import (
        top_k_renorm_probs_triton as top_k_renorm_prob,
    )
    from sglang.kernels.ops.sampling.renorm_triton import (
        top_p_renorm_probs_triton as top_p_renorm_prob,
    )
Should we reuse those instead of adding new implementations?

@1am9trash Done — same signatures, so the alias import (as dflash_utils.py does) puts the
renorm back on one code path. Private kernel, _is_hip branch, and its test gone;
kernels/ops/moe/test_renorm.py covers what's left.

Also collapsed the gate predicate to is_hip and not use_rejection_sampling.

3 files, +171/−15, down from 4 and +409.

1am9trash

Comments resolved.
Thanks for the update.

1am9trash · 2026-09-09T01:52:20Z

Because of the new rebase commit pushed, the previous CI wae canceled, now I retriggered it (PR-test-base).

jiejingzhangamd · 2026-09-09T21:53:02Z

Pushed a follow-up: HIP no longer auto-enables --speculative-use-rejection-sampling for EAGLE3.

That was crashing AMD stage-a-test-1-gpu-small-amd / test_basic_sanity_eagle3.py (TestBasicSanityEagle3.setUpClass) with draft vocab 32000 vs target 128256. Same-vocab EAGLE on HIP is unchanged.

JohnQinAMD · 2026-09-13T03:48:21Z

I pushed four commits onto this branch. They fix bugs that only become reachable because of this PR — before it HIP never entered the sampling-verify path, so the code was dead. Landing without them ships the corruption in row 2.

commit	without it
`carry real top_k into the draft proposal`	the draft can't tell a greedy request from a `T=1` one, so it samples where it should take argmax
`copy draft_probs out of the CUDA graph memory pool`	the pool recycles that block before `eagle_sample` reads it; 2-4% of tokens come back as an untrained padding row, GPQA drops to 0.4646
`reject non-finite draft probabilities in the sampler`	the residual passes guard `q` with `q == q`, which infinities satisfy
`a malformed draft probability must reject, not accept`	the accept test has no guard, so a corrupt `q` accepts unconditionally and commits the draft head's token

GPQA-Diamond, GLM-5.2-FP8, 8x MI355X, TP8, NEXTN steps=1 draft=2, 198q, T=1.0 top_p=0.95 max_tokens=98304:

	base (`bb15be6d79`)	this branch
exact_match	0.7071 ± 0.0324	0.8889 ± 0.0224
requests reaching a final answer	62.1%	94.4%
accuracy on requests that do finish	0.9268	0.9358
wall clock	~60 min	~24 min

Accuracy on requests that finish is the same on both sides — the delta is entirely whether the model terminates. Without this PR, verify commits argmax whatever temperature was asked for, and greedy decoding on long chain-of-thought loops until the budget is gone.

GSM8K, all 1319 questions, 5-shot:

	base	this branch	speculation off
`T=0`	0.941	0.940	0.951
`T=1.0 top_p=0.95`	0.928	0.925	0.937
accept length `T=0` → `T=1.0`	1.932 → 1.931	1.931 → 1.886	—

No regression at either temperature — one question apart at T=0, four at T=1.0, against a 0.6pp sampling interval. The accept-length row is the one that matters: on base it does not move with temperature, because verify commits argmax whatever you ask for. This branch responds to it.

Performance

Same GSM8K sweep as above — 1319 questions, concurrency 16, so accuracy and throughput come from one run.

	base	this branch	speculation off
output tok/s, `T=0`	746.3	744.1	601.4
output tok/s, `T=1.0`	747.3	727.1	577.3
end-to-end latency, `T=0`	181.4 s	180.5 s	222.3 s
end-to-end latency, `T=1.0`	181.7 s	191.5 s	233.4 s

T=0 is free (-0.3%, inside noise): is_all_greedy sends verify down the argmax branch either way. T=1.0 costs -2.7%, from acceptance dropping 1.931 → 1.886 as rejection sampling starts rejecting. A single request at concurrency 1 shows the same thing: 125.98 vs 128.76 tok/s at T=1.0, 135.28 vs 135.66 at T=0.

Against that, speculative decoding is worth +24% at T=0 and +26% at T=1.0 over no speculation at all, and on the GPQA run above this branch finishes 198 questions in ~24 minutes where base takes ~60 — the runaways it prevents cost far more compute than the rejections it adds.

CI: stage-a-test-1-gpu-small-amd now passes. The remaining red is a canceled runner task on base-b-test-1-gpu-small (5) that fast-failed every other CUDA shard, plus a missing libavutil.so.60 on XPU.

xiaobochen-amd · 2026-09-14T05:12:56Z

/rerun-failed-ci

At1a8 · 2026-09-14T10:56:36Z

Hi @xiaobochen-amd, could you please test the benchmark listed in #39253?

xiaobochen-amd · 2026-09-14T12:58:34Z

Hi @xiaobochen-amd, could you please test the benchmark listed in #39253?

@At1a8 We haven't run that set.

@JohnQinAMD measured this branch on GLM-5.2-FP8 (NEXTN 1/2): GPQA-Diamond
0.7071 → 0.8889, requests reaching a final answer 62.1% → 94.4%, wall clock
60 → 24 min. Different model, but it looks like the same thing your Simple-QA
30.6% → 0.9% is measuring. GSM8K 1319 shows no regression.

xiaobochen-amd · 2026-09-14T14:41:10Z

Re-ran the failing tests on MI355X in the same image CI uses
(rocm/sgl-dev:v0.5.19-rocm10-mi35x-20260913), against clean upstream/main:

	main	this PR
`test_deterministic`	2 failed / 4 error	2 failed / 4 error
`test_sampling_mask`	8 passed / 17 error	8 passed / 17 error
`test_lean_attention`	11 passed	11 passed
`test_dsv4_indexer_quant`	6 passed	6 passed
`test_triton_mla_prefill_gfx950`	3 passed	3 passed
`test_triton_dense_prefill_gfx950`	6 passed	6 passed
`test_disaggregation_pp`	3 passed	3 passed

Identical on every file, so none of them come from this branch.
base-b-test-1-gpu-large (1) is test_inkling_unified, a CUDA KL check that runs
no speculative decoding, so eagle_sample is never called there either.

xiaobochen-amd · 2026-09-14T14:41:36Z

/rerun-failed-ci

The sgl_kernel sampling-verify ops (tree_speculative_sampling_target_only, top_k_renorm_prob, top_p_renorm_prob) are CUDA/MUSA-only, so HIP fell through eagle_sample's gate to argmax, silently ignoring temperature and top_p and degenerating into repetition loops at temp > 0. Route HIP through the pure-Triton chain sampler (reject_sampling.py) when rejection sampling is active and the batch is not all-greedy, with a Triton pivot-search renorm (_renorm_top_k_top_p_hip) standing in for the CUDA-only renorm ops. top_p >= 1.0 is an exact no-op there, to stay bit-exact with CUDA's top_p_renorm_prob(p, 1.0) == p. The gate is factored into a pure helper, _verify_uses_greedy, which reduces to the original predicate on every non-HIP platform: with is_hip False the added term drops out and the expression is the one it replaced. Two shapes here come from this branch rather than from the original fix, because the surrounding code has moved: - sampling_fn used the eager `A if cond else B` form, which resolves tree_speculative_sampling_target_only whether or not that branch is taken. Since HIP must not import it at all, the dispatch becomes an if/else that imports inside the arm that needs it. Non-HIP behaviour is unchanged. - _handle_eagle_family writes through declare_resolution, whose contract is that a field it decides is not visible to a later `cfg.` read. The rejection-sampling validation immediately below was reading cfg, so it would have skipped every auto-enabled config -- including its own "rejection sampling is enabled" log. That one read becomes resolved_view(server_args), which is what the declare_resolution docstring prescribes for exactly this case. Defaulting the flag on matters because without it the fix only reaches users who pass --speculative-use-rejection-sampling; everyone else still silently gets argmax verify. It flips only on HIP, for EAGLE/EAGLE3 with topk=1, default accept thresholds and non-deterministic inference -- the same conditions the validation below raises on, so the flip can never turn a working config into an error. Setting it also makes the draft worker emit the target-vocab proposal (draft_probs) that the Triton chain sampler consumes. Tests: a gate-routing truth table (CPU) and renorm-vs-nucleus numerical equivalence (GPU).

The blank line predates this PR, but pre-commit formats every file a change touches, so it has to go here.

check_registered_tests.py admits a newly added registered test only under test/registered/<kind>/<subsystem>/, and `spec` is not one of the kinds. This test drives a GPU and compares the Triton renorm against nucleus, so `kernel` is the kind that fits -- `unit` is restricted to CPU suites. Its suites move with it: base-b-kernel-unit on CUDA and jit-kernel-unit-test-amd on ROCm, both of which carry the *-kernel-* name that kind requires. test_eagle_gate_routing.py is a CPU truth table and already sat correctly under unit/spec.

…mentation renorm_triton.py already carries top_k_renorm_probs_triton and top_p_renorm_probs_triton, and dflash_utils.py already aliases them to the sgl_kernel names on HIP. Do the same here rather than keeping a private pivot kernel: the renorm block goes back to one code path for both platforms, and the ops that survive are the ones test/registered/kernels/ops/moe/test_renorm.py already covers. Drops _top_p_renorm_kernel, _renorm_top_k_top_p_hip, the _is_hip branch around the renorm calls, and the renorm equivalence test that only existed to guard the private implementation.

Co-authored-by: Cursor <cursoragent@cursor.com>

EAGLE3 hot-token vocabs cannot feed the rejection kernel yet, so defaulting the flag on crashed stage-a test_basic_sanity_eagle3. Leave that path greedy until the worker can scatter draft probs into the target vocab. Co-authored-by: Cursor <cursoragent@cursor.com>

sample_draft_proposal decides greedy-vs-sample from temperature alone, but SamplingParams rewrites temperature 0 to temperature=1.0 with top_k=1, so a greedy request is indistinguishable from a T=1 one there. It samples a sharp but non-degenerate distribution and proposes a non-argmax token often enough to break the draft chain, costing accept length. Pass the per-request top_ks through and let a greedy row propose its argmax. The CUDA graph runner has to carry top_ks in a device buffer the same way it already carries temperatures: its synthetic SamplingBatchInfo used a host-side placeholder, so without this the correction never sees a real top_k and only about a fifth of the loss comes back. TOP_K_ALL rather than -1 for the buffer fill: -1 is not a top_k this pipeline ever carries, and it reads as top_k <= 1, i.e. greedy, for the padded rows. GLM-5.2-MXFP4, MI355X, TP4/EP4, EAGLE steps=5 topk=1 draft=6, GSM8K 200q 5-shot temp=0 conc=8: baseline before after accept len 3.864 3.117 3.857 output tok/s 653.3 569.0 653.1 GSM8K accuracy 0.929 - 0.936 Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> (cherry picked from commit 72ea484)

draft_probs is the one draft-graph output that the target forward does not consume: it is read afterwards, by eagle_sample. The tensor the replay returns points into the graph's private memory pool, and that block is recycled before the read. What lands there is the DSA top-k mask, so eagle_sample sees -inf in draft_probs -- scattered runs whose position tracks the sequence at roughly the accept length, each ending on a 64-aligned column, always at the same draft step because the pool's allocation order is fixed at capture. Downstream that turns into corrupt output: the residual p - q goes infinite, norm_sum with it, the cumulative search never matches, and the sampler commits its degenerate fallback, an untrained embedding row. The model degenerates from there. Copy out at the graph boundary. Only rejection sampling produces draft_probs, so this is a no-op everywhere else. GLM-5.2-MXFP4, MI355X, TP4/EP4, EAGLE steps=5 topk=1 draft=6, rejection sampling on, temperature=1.0 top_p=0.95: before after non-finite entries in draft_probs 5 to 59 0 stray tokens per 1200 generated 25 to 81 0 The same three arms with CUDA graphs disabled show zero in both columns, which is what identified the pool rather than the producer. GSM8K 200q 5-shot at temperature=0: 0.940 / 0.000 invalid / accept len 3.035 / 512.8 tok/s, against 0.935 / 0.000 / 3.031 / 520.3 without the copy. The 1.4% throughput gap is inside this harness's noise: three repeats on one unchanged server span 582 to 646 tok/s. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> (cherry picked from commit 3e1fae0)

Defence in depth behind the previous commit. The residual guard tested q for NaN with `q_val == q_val`, which infinities satisfy, so a malformed draft distribution reaches `p - q` and drives norm_sum infinite. The cumulative search then never matches and the kernel commits its fallback, VOCAB_SIZE - 1 -- an id past the tokenizer's vocabulary whenever the embedding is padded, as it is for GLM-5.2 (154880 against 154856). Test q against the probability range instead. A comparison with NaN is false, so the range test rejects NaN along with the infinities and with negatives. A sampling kernel should not be able to emit an invalid token id because an input was malformed, whatever wrote that input. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> (cherry picked from commit dee891b)

) The previous commit guards q in the two residual passes but not in the accept test itself, where it does the most damage: `coin * q < p` is -inf < p for an -inf q and 0 < p for a zero one, so a corrupt row accepts unconditionally and the committed token comes from the draft head rather than the target. That is the direction that costs output quality, and it is invisible -- acceptance climbing to 1.0 reads as a good draft model. Zero also passes the range test the residual passes use, so widening that guard alone does not cover this site. X was sampled from q, so q(X) has to be strictly positive; anything else means the row is not the distribution X came from. Reject on that, which sends the step through the residual path and resamples from the target. Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com> (cherry picked from commit 8d63553)

At1a8 · 2026-09-16T00:56:50Z

Confirmed that this PR also fixes the accuracy issue(#39253) in DeepSeek-V4 MTP. Thanks @HaiShaw @xiaobochen-amd

dataset	offcial(https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro)	this PR
GSM8K	92.6	96.44
AA-LCR	66.3	76.00
GPQA-Diamond	90.1	89.39
LongBench v2	51.5	53.28
Simple-QA	57.9	57.33
LiveCodeBench	93.5	90.43
MMLU-Pro	87.5	87.35
HLE	37.7	35.35

xiaobochen-amd requested review from Qiaolin-Yu, Ying1123, hnyls2002 and merrymercy as code owners August 30, 2026 13:10

github-actions Bot added amd speculative-decoding labels Aug 30, 2026

JohnQinAMD mentioned this pull request Sep 7, 2026

[ROCm] Fix EAGLE spec-decode verify silently sampling greedy on HIP #31214

Closed

4 of 5 tasks

1am9trash added the run-ci CI: run the baseline test suite on this PR label Sep 7, 2026

1am9trash approved these changes Sep 7, 2026

View reviewed changes

1am9trash requested changes Sep 7, 2026

View reviewed changes

1am9trash approved these changes Sep 8, 2026

View reviewed changes

jiejingzhangamd force-pushed the rocm/eagle-verify-greedy-hip branch from 5c6e07c to acd4d47 Compare September 8, 2026 18:10

JohnQinAMD mentioned this pull request Sep 12, 2026

[ROCm] EAGLE: a malformed draft probability must reject, not accept xiaobochen-amd/sglang#59

Merged

JohnQinAMD requested review from BBuf, DarkSharpness, HaiShaw, HydraQYH, celve and yuan-luo as code owners September 12, 2026 21:57

github-actions Bot added the jit-kernel label Sep 12, 2026

bingxche mentioned this pull request Sep 13, 2026

[CI Monitor] Daily Report - 2026-09-13 bingxche/sglang-ci-bot#194

Open

JohnQinAMD force-pushed the rocm/eagle-verify-greedy-hip branch from 9eed096 to 2cd1fcb Compare September 13, 2026 22:02

HaiShaw approved these changes Sep 15, 2026

View reviewed changes

JohnQinAMD and others added 10 commits September 15, 2026 06:29

[Lint] ruff-format speculative_hook.py

bdad927

The blank line predates this PR, but pre-commit formats every file a change touches, so it has to go here.

[Rebase] Preserve NPU EAGLE sampling gate

1ea1c73

Co-authored-by: Cursor <cursoragent@cursor.com>

xiaobochen-amd force-pushed the rocm/eagle-verify-greedy-hip branch from 262741a to 69cf150 Compare September 15, 2026 06:30

HaiShaw mentioned this pull request Sep 15, 2026

[AMD][SPEC]Fix DeepSeek-V4 benchmark accuracy gap by supporting temperature sampling in EAGLE verify #39253

Open

4 of 5 tasks

HaiShaw merged commit 7eedd57 into sgl-project:main Sep 16, 2026
372 of 438 checks passed

This was referenced Sep 16, 2026

[CI Monitor] Daily Report - 2026-09-16 bingxche/sglang-ci-bot#197

Open

[CI Monitor] Daily Report - 2026-09-17 bingxche/sglang-ci-bot#198

Open

yichiche mentioned this pull request Sep 21, 2026

[AMD][Fix] AgentX HIP TPOT regression when SGLANG_SIMULATE_ACC_LEN is set #40598

Merged

4 of 5 tasks

This was referenced Sep 21, 2026

[CI Monitor] Daily Report - 2026-09-21 bingxche/sglang-ci-bot#202

Open

[CI Monitor] Daily Report - 2026-09-22 bingxche/sglang-ci-bot#203

Open

[CI Monitor] Daily Report - 2026-09-23 bingxche/sglang-ci-bot#204

Open

adi271213 mentioned this pull request Sep 24, 2026

[Feature][Spec][XPU] Honor rejection sampling in EAGLE chain verification #41070

Open

2 tasks done

rchalamala mentioned this pull request Sep 25, 2026

[Fix] Keep speculative chain samples on positive support modal-projects/sglang#35

Open

4 tasks done

hdt98 mentioned this pull request Sep 25, 2026

[AMD] [GLM-5.3-Flash] Enable FP8 and MXFP4 serving on gfx950 #39273

Merged

3 of 4 tasks

salexspb mentioned this pull request Sep 30, 2026

[PD][Spec] Give a PD decode's first EAGLE draft step a proposal distribution #41783

Open

nvpohanh mentioned this pull request Oct 1, 2026

[Speculative] Seed rejection-sampling draft proposals for deterministic inference #33395

Open

6 tasks done

dg1kjd mentioned this pull request Oct 5, 2026

[Bug] chain_speculative_sampling_triton rejects every draft whose q rounds just above 1.0 (since #37134) #42568

Open

ischencheng mentioned this pull request Oct 5, 2026

[Spec] Accept draft probabilities that round slightly above 1 in chain verify #42615

Open

4 of 5 tasks

Conversation

xiaobochen-amd commented Aug 30, 2026 • edited by github-actions Bot Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Problem

Fix

Two shapes that come from the current tree, not from the fix

Test

CI States

Uh oh!

1am9trash left a comment • 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!

1am9trash left a comment

Choose a reason for hiding this comment

Uh oh!

xiaobochen-amd commented Sep 7, 2026

Uh oh!

1am9trash left a comment

Choose a reason for hiding this comment

Uh oh!

1am9trash commented Sep 9, 2026

Uh oh!

jiejingzhangamd commented Sep 9, 2026

Uh oh!

JohnQinAMD commented Sep 13, 2026

Performance

Uh oh!

xiaobochen-amd commented Sep 14, 2026

Uh oh!

At1a8 commented Sep 14, 2026

Uh oh!

xiaobochen-amd commented Sep 14, 2026

Uh oh!

xiaobochen-amd commented Sep 14, 2026

Uh oh!

xiaobochen-amd commented Sep 14, 2026

Uh oh!

At1a8 commented Sep 16, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

6 participants

xiaobochen-amd commented Aug 30, 2026 •

edited by github-actions Bot

Loading

1am9trash left a comment •

edited

Loading

At1a8 commented Sep 16, 2026 •

edited

Loading