【Task.31】feat(sft): add pair-aware DPO training by A-Words · Pull Request #257 · redai-studio/Relax

A-Words · 2026-08-11T05:21:52Z

What

本 PR 在现有离线 SFT 链路中加入 pair-aware Direct Preference Optimization（DPO）能力。

增加严格的 chosen/rejected 偏好数据 schema 校验、渲染、截断、拒绝原因统计，以及确定性的 UltraFeedback 子集生成。
在 TransferQueue、SeqlenBalancedSampler、DP 分配、动态 batching 和 micro-batch packing 全链路保持 preference pair 原子性。
增加标准 reference DPO、reference-free DPO loss、pair-count 归一化和 DPO 指标。
增加事务化 frozen-reference 重建，并校验 canonical parameter、源文件 manifest、optimizer state 和固定 probe digest。
将验证后的 reference identity 写入 checkpoint sidecar，并在 fresh/resume 训练中持续校验。
增加公开的 Qwen3-0.6B/UltraFeedback recipe、中英文文档和聚焦测试。

本 PR 不增加 Reward Model scalar head。Reward Modeling 和最终的共享偏好评测器位于后续 PR #258。

Why

Relax 当前主要覆盖在线 RL 工作流。Task 31 要求在不复制训练框架的前提下，复用 SFT trainer 建立标准的离线偏好优化链路。

实现遵循 RFC #208 和Task 31 要求。

Relates to #208.

Follow-up: #258.

How

本次更新合并主线 0651812093e3cd730302709b3db35ce0bb199e4b，保留作者历史；主线集成提交为 14ccc2a。

通过现有 _create_sft_train_dataset / _prepare_sft_tq_payload 接入偏好数据，保留主线 SFT producer/shard/prepack 路径。
preference v1 明确拒绝 --sft-async-prepack、task_type=seq_cls 和 MTP-only 组合；CPU 数据预取仍可用。
补齐随机 train/eval split 所需的数据集接口与恢复测试，将实际 pair 数传入 scheduler 和指标分母。【Task.31】feat(sft): add pair-aware DPO training #257 本身仍不启用 held-out evaluator，评估由【Task.31】feat(sft): add reward modeling and preference evaluation #258 补充。

每个 TransferQueue row 表示一个完整 preference pair，并通过 custom_meta 携带 chosen/rejected 的合并 sequence cost。固定版本的 SeqlenBalancedSampler 因而可以在不拆分 pair 的前提下形成等大小 DP 分组。动态 packer 先分配完整 pair，仅在构造模型输入时展平 chosen/rejected branch。

DPO loss 消费显式的 pair/branch identity，使用 completion-only mask，按全局 pair 数归一化，同时支持标准 reference 和 reference-free 公式。空 mask、非有限值、prompt prefix 不一致、截断后两侧相同、超大 pair，以及不一致的 global pair count 均 fail closed。

标准 DPO fresh 训练将初始 actor 快照为 frozen reference；resume 时则从固定 HF repository/revision 执行 model-only reference 重建。Canonical parameter、源文件 manifest、optimizer master/state 和固定 log-prob probe digest 可防止 reference 切换静默修改 actor 或 optimizer。验证后的 identity 会写入 checkpoint 旁的 relax_dpo_reference.json。

Testing

删除源码字符串断言（`aab0c72`）

删除 test_preference_iterator_has_no_device_scalar_readback 及其 inspect 导入，保留 pair batching、DP 分配和分母校验的行为测试。
全量 pre-commit 通过；本次仅删除测试，未在本地重跑需 Megatron/CUDA 的 GPU 测试，新提交的 CI 结果以 checks 为准。

CI 测试修正（`4bc9dda`）

已回退 CPU CI stub 特判提交，删除 test_pinned_seqlen_sampler_consumes_pair_costs_and_keeps_equal_dp_groups 及固定源码 SHA-256 检查。
missing-media 测试在启动后台 prefetch 前安装 warning 捕获，消除捕获顺序竞态。
使用当前 CPU CI 的 TransferQueue stub 执行 preference utils 与 streaming dataset 回归：66 passed；全量 pre-commit 通过。
最新提交的 GitHub CI 结果以本 PR checks 为准。

主线集成提交 `14ccc2a` 的本地验证

pre-commit run --all-files --show-diff-on-failure：通过。
Python 3.12 CPU 聚焦回归：234 passed, 7 skipped。
覆盖 SFT engine / producer / sharding、偏好数据截断与 split/resume、独立数值 loss、reference identity、参数校验和 Docker 固定版本 TransferQueue sampler。
跳过项需要 Megatron/CUDA，包括生产后端 loss/batching 和部分 reference/训练数据路径。本次未执行完整 pytest tests/，未重跑 GPU 训练、DP=2 或 distributed checkpoint save/resume；最新 CI 以本 PR checks 为准。

python -m pytest tests/engine/sft/ tests/components/test_sft.py \
  tests/backends/megatron/test_preference_batching.py \
  tests/backends/megatron/test_dpo_loss.py \
  tests/backends/megatron/test_dpo_reference_integrity.py \
  tests/backends/megatron/test_sft_train_data_fields.py \
  tests/utils/training/test_preference_utils.py \
  tests/data/test_prepare_ultrafeedback_preferences.py \
  tests/utils/test_arguments_sft.py -q -ra --tb=short

作者提供的历史 GPU 验证

以下日志、曲线和证据包对应旧集成提交，保留用于追溯，不代表本次合并后的 GPU 复测结果。主线已更新 shuffle 行为，重新验证时应比较新版本的连续训练与恢复训练。

在最终堆叠集成 head b57a1818f2f80d28d0ad650b314eb4d1c82df3aa 的标准容器中执行：

pytest tests/engine/sft/ \
  tests/backends/megatron/test_sft_train_data_fields.py \
  tests/utils/training/test_preference_utils.py \
  tests/backends/megatron/test_preference_batching.py \
  tests/backends/megatron/test_dpo_loss.py \
  tests/backends/megatron/test_reward_model_loss.py \
  tests/backends/megatron/test_reward_model_checkpoint.py

179 passed

pre-commit run --all-files
All hooks passed

固定 TransferQueue 兼容性测试覆盖 pair-row 原子性、等大小 DP 分组、row-aligned total_lengths cost 消费，以及无丢失、无重复的完整分配。数据准备脚本的测试覆盖 manifest：4,096 个 train pair、512 个互不重叠的 eval pair、确定性选择、拒绝原因计数和派生文件 hash。

GPU 验证固定使用相同的模型/data revision 和 global batch size 32 pairs：

运行	Steps	Final loss	Margin	Tie-aware accuracy	Bootstrap lower 95%
DPO DP=1	200	0.630386	0.186692	0.654297	0.619141
DPO DP=2	200	0.628974	0.190042	0.664062	0.628906

DP=1 resume 从 iteration 199 恢复，完成下一 step，并保持相同的 reference identity。
Reference-free DPO 完成 20 steps，没有 reference log-prob 指标、reference checkpoint load 或 reference sidecar。
DP=2 日志明确记录 SeqlenBalancedSampler with dp_size=2，两张 A100 均有实际活动。

Type of Change

Bug fix (non-breaking change that fixes an issue)
New feature (non-breaking change that adds functionality)
Breaking change (fix or feature that would cause existing functionality to change)
Documentation update
Refactoring (no functional changes)
Performance improvement
CI/CD or build changes

Screenshots / Logs

本 PR 附带 DPO DP=1/DP=2 训练曲线和脱敏后的可复现性证据包。证据包包含：

200-step 训练日志、scalar CSV 和 GPU telemetry；
step 0/final 逐 pair 评测结果与冻结 batch plan；
DP=1 resume、reference-free DPO、probe contract 和 reference sidecar；
固定环境、模型/data revision、镜像身份和逐文件 SHA-256 清单。

公开附件不包含 checkpoint、optimizer state、模型权重、完整数据集、凭据或机器专属标识。

测量环境使用 ghcr.io/redai-infra/relaxrl:dev-20260723-8cc1e8fd；证据采集时从 GHCR 解析的 manifest digest 为 sha256:8dc39af377a570e6cd7ec88c8b7fcd44c1eb820111e9d2069f1c7c3024b2ea23。Transformer Engine 无法导入顶层 flash_attn_interface，并明确回退到 native attention。因此 DP 对比描述的是相同 native-attention 环境，不是 FlashAttention 性能结果。

task31-pr1-dpo-evidence-public-v2.tar.gz

# 🐛 Bug Fix ## Reject --ref-load under preference objectives - Standard DPO snapshots the frozen reference from the initialized policy and rebuilds it from --hf-checkpoint on resume; --ref-load was silently ignored on that path while also rerouting the bridge mode policy-init fallback, so validate_preference_args now fails fast when it is set (RFC redai-studio#208: --ref-load is not the v1 DPO reference source) - Drop the misleading --ref-load from the DPO recipe; behavior is unchanged because the bridge fallback already resolves to the same HF checkpoint --- # ♻️ Refactor ## Deduplicate the preference-mode predicate - Route data.py get_data_iterator and loss.py loss_function through is_preference_mode() instead of inline loss_type/sft_objective checks, keeping objective dispatch on a single source of truth --- # ✅ Tests ## Cover --ref-load rejection - Parametrize standard and reference-free DPO rejection cases in test_preference_runtime.py

# ♻️ Refactor ## Remove unreachable DPO loss fallback - Drop the dead `0 * logits.sum()` branch in dpo_loss_function: build_preference_pair_indices already raises on an empty micro-batch ## Attach stable reason codes at preference raise sites - Introduce _PreferenceRowError carrying an explicit reason_code from every normalization/split/truncation rejection; message matching in _classify_preference_error remains only as a fallback for errors raised outside the module, so reworded messages can no longer silently degrade rejections to "schema" --- # ⚡ Performance ## Merge DP pair-count collectives - Reduce _get_preference_data_iterator from three all_reduces to two by riding MIN/MAX on a single MAX all_reduce over [count, -count] ## Skip redundant same-tag model restores - _switch_model now returns early when the target tag is already active, eliminating the duplicate full-weight CPU->GPU restore after the ref-forward finally block; paths that deliberately dirty weights already clear the tag first (covered by the injected-loader-failure reference integrity test) --- # 📝 Documentation ## Note probe bitwise-determinism prerequisite - en/zh DPO guides now state that the reference probe digest assumes an identical GPU/driver/image/kernel stack on resume, and that a probe mismatch signals environment drift rather than data corruption ## Justify the checkpoint-save barrier - Comment that the post-sidecar barrier is functional (peers must not pass before rank 0 persists the identity file), and use an explicit gloo group for the rank-0 check per distributed code rules

# 🐛 Bug Fix ## Preserve standard DPO likelihood semantics - Reject non-unit or non-finite rollout temperatures before actor construction - Replace CUDA-to-Python condition checks with device-side asynchronous assertions - Consolidate finite-value validation on the final DPO logits ## Validate preference data preparation inputs - Validate message objects, roles, and string content on both preference branches - Classify non-object source rows as schema rejections --- # ✅ Tests ## Cover validation regressions - Test finite and non-unit temperature failures - Verify CUDA conditions avoid Python boolean conversion - Cover malformed chosen and rejected message schemas

# 🐛 Bug Fix ## Pin standard DPO frozen reference provenance - Resolve the reference from the declared repository and revision in the configured HF checkpoint directory - Require Hugging Face local metadata to verify the pinned snapshot without downloading during actor startup - Rebuild the frozen reference from that verified directory for both fresh starts and resumes - Fail clearly when the configured checkpoint is unavailable, unverified, or resolves elsewhere --- # 📝 Documentation ## Document pinned local model preparation - Show the fixed-revision hf download command used by the public DPO recipe --- # ✅ Tests ## Cover local reference resolution - Verify repository, revision, local directory, and local-only cache resolution - Verify missing metadata and mismatched resolved directories fail before model loading

# 🐛 Bug Fix ## Require pinned local snapshot metadata - Require an immutable 40-character commit SHA for standard DPO references - Verify Hugging Face local-dir tree metadata before accepting the configured checkpoint - Reject copied or incomplete local snapshots that lack pinned provenance metadata --- # ✅ Tests ## Cover metadata validation - Exercise the pinned tree lookup and missing-metadata rejection paths - Document the full commit-SHA requirement for the DPO recipe

# 🐛 Bug Fix ## Validate pinned Hugging Face local snapshots - Parse the per-file metadata format used by pinned huggingface_hub 1.7.2 - Require every reference file to match the configured commit revision - Verify regular Git files with Git blob SHA-1 and LFS files with SHA-256 - Stream large-file hashing to keep reference validation memory bounded --- # ✅ Tests ## Cover reference provenance and integrity - Exercise Git and LFS ETag validation - Reject missing or mismatched local metadata - Reject replaced weights even when their original mtime is restored

# 🐛 Bug Fix ## Enforce local reference completeness - Require a supported single-file weight or standard Transformers weight index - Validate non-empty index weight maps, safe shard paths, and format-specific suffixes - Reject indexes that reference missing model shards --- # ✅ Tests ## Cover missing reference weights - Accept a complete multi-shard safetensors index - Reject snapshots without a supported weight entry point - Reject a deleted shard even when its local metadata is also removed

# 🐛 Bug Fix ## Move preference scheduling to the CPU control plane - Replace CUDA count reductions and scalar readbacks with one DP-Gloo control gather - Preserve pair-row, global-denominator, and micro-batch-count agreement across ranks - Require Gloo process groups for preference objectives and enable them in the DPO recipe --- # ⚡ Performance ## Remove hot-path GPU synchronization - Eliminate all item calls and control tensors from the preference iterator - Avoid GPU-to-CPU synchronization for train, reference, and evaluation iterators --- # ✅ Tests ## Guard preference iterator contracts - Cover unequal DP pair-row rejection through the Gloo control path - Assert the iterator contains no item or all-reduce scalar readback path

# ♻️ Refactor ## Keep one reference identity path - Remove the unused helper that reads the already-backed-up ref weights - Preserve rebuild validation and backup ordering without duplicate digest work --- # ✅ Tests - Pass the full pre-commit suite - Compile the Megatron actor module

yxyOo · 2026-08-21T08:45:17Z

Nit: Please avoid committing data JSON files directly to this PR. Consider adding a download link instead.

A-Words · 2026-08-21T13:49:57Z

Thanks for the suggestion. I removed the checked-in UltraFeedback manifest from this PR, updated the DPO documentation to use the reproducibility evidence bundle, and updated the evidence bundle linked in the PR description to include the manifest.

# 🐛 Bug Fix - Merge current upstream main while preserving the original DPO history. - Reuse the shared SFT dataset factory and pair-aware queue payload path. - Preserve async SFT producers, sequence classification and MTP behavior; reject those unsupported combinations for preference objectives. - Add deterministic preference split/resume interfaces and propagate the actual global pair count to scheduler and metric normalization. # ✅ Tests - Focused CPU suite: 234 passed, 7 skipped (Megatron/CUDA unavailable). - Exercise the Docker-pinned TransferQueue pair sampler without Megatron. - pre-commit run --all-files --show-diff-on-failure passed. - GPU training and distributed save/resume were not rerun. # 📝 Documentation - Document the v1 compatibility boundaries in English and Chinese. Co-authored-by: Codex <noreply@openai.com>

rai-studio-bot · 2026-09-21T07:48:10Z

Nyanpasu 审查看板

审查状态： 🚧 需要修改

审查版本： aab0c72c745c3151fa255b9a9a78d66e64766092

已复查 aab0c72：仅删除 iterator 源码字符串断言及 inspect 导入，保留 pair batching、DP 分配和分母校验行为测试，未发现新增问题。生产代码未变，F1、F2 仍未解决，F3 已解决，维持既有需要修改结论。本轮未重复运行测试；本地缺少 Megatron/CUDA，未执行 GPU/多节点验证。当前 CI 仍在运行。

编号	问题	状态	规则来源
F1	跨 epoch 重复采样被误判为 pair ID 冲突	🚧 未解决	—
F2	generation marker 模板未排除历史 assistant mask	🚧 未解决	—
F3	真实 sampler 校验在 CPU CI stub 上必然失败	✅ 已解决	—

^{Powered by Nyanpasu with gpt-6-astra medium, please check the suggestions carefully.}

rai-studio-bot

需要修改：合法数据在跨 epoch 组 batch 时可能终止训练，详见行级意见；另请处理多轮 prompt 的 completion mask 和 CPU CI 的 sampler 测试问题。

已审查当前提交的数据、pair batching/loss、reference 恢复与 checkpoint 链路。本地聚焦测试共 74 passed、8 skipped；扩大 SFT 测试范围时因缺少 tensordict 无法收集，未执行本地 GPU/多节点训练。当前 GPU 检查通过，Python 3.10–3.12 检查失败，具体原因见测试行级意见。

^{Powered by Nyanpasu with gpt-6-astra medium, please check the suggestions carefully.}

rai-studio-bot · 2026-09-21T07:56:17Z

+    encoded_pair_ids = [
+        int.from_bytes(hashlib.sha256(pair.pair_id.encode()).digest()[:8], "big") >> 1 for pair in pairs
+    ]
+    if len(set(encoded_pair_ids)) != len(encoded_pair_ids):
+        raise ValueError("preference pair ID hash collision within batch")


优先级：P1

请区分数据源 pair ID 和本次采样的 pair occurrence ID。现有 IndexManager.get_next_indices() 会跨 epoch 补齐 batch，因此合法数据也可能在同一个 batch 中再次采到同一行；这里却将其判为 hash collision 并终止生产。当前提交用 10 个唯一 pair、seed=42、batch size=4，即可在第 3 个 batch 复现 [pair-2, pair-3, pair-7, pair-2] 后抛错。这会使不能整除 batch size 的普通数据集在跨 epoch 时中断训练。请为每次采样分配独立的训练 pair identity（同时保留源 ID），并补充跨 epoch 的打包/训练回归；仅删除这里的检查仍会触发下游重复 chosen/rejected branch 检查。

rai-studio-bot · 2026-09-21T07:56:44Z

+        chosen_tokens, chosen_mask = render_with_loss_mask(
+            chosen_sample,
+            tokenizer=self.tokenizer,
+            apply_chat_template_kwargs=self.apply_chat_template_kwargs,
+        )


优先级：P2

这里两个 render 调用需要显式只监督最终 completion。对于包含 {% generation %} 的模板，render_with_loss_mask() 默认的 assistant-mask 路径不会读取 CanonicalMessage.learn=False，而会标记所有历史 assistant 回答。共享 prompt 为 user → assistant → user 的合法 pair 因此在 _split_branch() 抛出 completion mask must be one contiguous suffix；使用本文件测试中的 tokenizer 即可复现。请在 chosen/rejected 渲染时排除历史回答（例如适用时传 last_turn_only=True），并测试带历史 assistant 的共同前缀。

rai-studio-bot · 2026-09-21T07:56:52Z

+    transfer_queue = pytest.importorskip("transfer_queue")
+    sampler_type = transfer_queue.SeqlenBalancedSampler
+    source_path = Path(inspect.getsourcefile(sampler_type) or "")
+    normalized_source = source_path.read_bytes().replace(b"\r\n", b"\n")
+    assert hashlib.sha256(normalized_source).hexdigest() == (
+        "dc6c2db50df4b9448d4845ccacef67a400517db892b5cd55de2e22f6baf6888b"
+    )


优先级：P2

pytest.importorskip("transfer_queue") 无法识别 CPU CI 中的占位模块：现有 .github/workflows/ci.yml 会生成含 SeqlenBalancedSampler 的 stub，导入成功后这里必然对 stub 源码做真实 sampler 的 hash 校验。当前 Python 3.12 检查已在这里失败（实际 hash 为 30cf6bec…），3.10/3.11 检查也失败。请在运行该兼容性测试的任务中安装 Docker 固定版本的真实 TransferQueue，或明确跳过已识别的 CI stub 并保留真实依赖环境中的兼容性门禁。

……什么测试

指的是这个 PR 新增的 test_pinned_seqlen_sampler_consumes_pair_costs_and_keeps_equal_dp_groups：先比较 sampler 整个源码文件的 SHA-256，再验证 pair 分组行为。CPU CI 导入的是占位模块，因此之前还没执行行为断言就失败了。

这里比较整个源码 hash 确实很脆弱；依赖版本应由安装配置固定，测试重点应是 pair 原子性、分组数量和 cost 分配。我上一条指出的是实际 CI 失败，并不是建议保留这种源码 hash 测试。

没让你回我……笨

你要是以后敢写出来这种测试就不给你 token 了

4bc9ddad 已删除该测试，原先对 CI stub 做源码 hash 校验的失败原因已消除，本项按已解决记录。删除同时移除了真实 sampler 的行为断言，因此不将此次变更视为 sampler 兼容性验证通过；新一轮 CI 仍在运行。

# ✅ Tests - Mark the CPU CI TransferQueue import stub explicitly and skip only the real-sampler integration check when that stub is present. Keep the pinned source digest and pair-assignment assertions for actual installations. - Install the missing-media warning capture before shuffle starts prefetch workers, avoiding a race in the logging assertion. - Reproduce the exact CI stub hash failure before the fix. - Validate 67 tests against real TransferQueue and 66 passed / 1 expected skip against the workflow-generated CPU stub. - pre-commit run --all-files --show-diff-on-failure passes. Co-authored-by: Codex <noreply@openai.com>

# 🔩 Chore - Revert 21d2c22 as requested before removing the pinned sampler test. - Restore the original CI stub and test files. - pre-commit run --all-files --show-diff-on-failure passes. Co-authored-by: Codex <noreply@openai.com>

# ✅ Tests - Delete the fixed-source-hash TransferQueue sampler test as requested. - Remove its now-unused imports without adding a CI-specific skip path. - Fix the independent prefetch logging race by installing the capture hook before starting the worker thread. - Focused CPU regression: 66 passed with the CI TransferQueue stub. - pre-commit run --all-files --show-diff-on-failure passes. Co-authored-by: Codex <noreply@openai.com>

SigureMo · 2026-09-21T08:44:12Z

+def test_preference_iterator_has_no_device_scalar_readback():
+    source = inspect.getsource(data_module._get_preference_data_iterator)
+    assert ".item(" not in source
+    assert "all_reduce(" not in source


# ✅ Tests - Remove the source-string test that forbids item and all_reduce calls. - Remove the unused inspect import. - Keep the existing pair batching and denominator behavior tests. - All pre-commit hooks pass. Co-authored-by: Codex <noreply@openai.com>

A-Words added 12 commits August 7, 2026 20:08

feat(sft): add pair-aware DPO training

8f99e09

fix(sft): enforce DPO reference integrity

46154cd

fix(sft): complete DPO delivery contract

f03aa5c

A-Words mentioned this pull request Aug 11, 2026

【Task.31】feat(sft): add reward modeling and preference evaluation #258

Open

7 tasks

A-Words changed the title ~~feat(sft): add pair-aware DPO training~~ 【Task.31】feat(sft): add pair-aware DPO training Aug 11, 2026

A-Words marked this pull request as ready for review August 11, 2026 05:55

A-Words requested review from NINGBENZHE and Yangruipis as code owners August 11, 2026 05:55

A-Words added 3 commits August 11, 2026 14:43

chore(task31): merge upstream main into DPO

4f149d4

test(dpo): guard Megatron-only checkpoint import

0e0b059

test(sft): skip Megatron integration without backend

633d386

A-Words added 2 commits August 21, 2026 21:07

docs(dpo): link Task 31 evidence bundle

d1b80d1

chore: merge upstream main into DPO

3feb0fe

SigureMo added the Hackathon 1st label Sep 21, 2026

SigureMo self-assigned this Sep 21, 2026

rai-studio-bot suggested changes Sep 21, 2026

View reviewed changes

SigureMo and others added 2 commits September 21, 2026 16:30

revert(sft): undo CPU CI stub workaround

8338413

# 🔩 Chore - Revert 21d2c22 as requested before removing the pinned sampler test. - Restore the original CI stub and test files. - pre-commit run --all-files --show-diff-on-failure passes. Co-authored-by: Codex <noreply@openai.com>

SigureMo reviewed Sep 21, 2026

View reviewed changes

rai-studio-bot mentioned this pull request Sep 22, 2026

[Task32] feat: fla mcore chunkwise cp stage 3 #273

Open

Navigation Menu

【Task.31】feat(sft): add pair-aware DPO training - #257

【Task.31】feat(sft): add pair-aware DPO training#257
A-Words wants to merge 22 commits into
redai-studio:mainfrom
A-Words:feat/task-31-dpo

A-Words commented Aug 11, 2026 •

edited by SigureMo

Loading

Uh oh!

yxyOo commented Aug 21, 2026

Uh oh!

A-Words commented Aug 21, 2026

Uh oh!

rai-studio-bot commented Sep 21, 2026 •

edited

Loading

Uh oh!

rai-studio-bot left a comment

Uh oh!

rai-studio-bot Sep 21, 2026

Uh oh!

rai-studio-bot Sep 21, 2026

Uh oh!

rai-studio-bot Sep 21, 2026

Uh oh!

SigureMo Sep 21, 2026

Uh oh!

rai-studio-bot Sep 21, 2026

Uh oh!

SigureMo Sep 21, 2026

Uh oh!

rai-studio-bot Sep 21, 2026

Uh oh!

SigureMo Sep 21, 2026

Uh oh!

This branch has not been deployed

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

Conversation

A-Words commented Aug 11, 2026 • edited by SigureMo Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

What

Why

How

Testing

删除源码字符串断言（aab0c72）

CI 测试修正（4bc9dda）

主线集成提交 14ccc2a 的本地验证

作者提供的历史 GPU 验证

Type of Change

Screenshots / Logs

Uh oh!

yxyOo commented Aug 21, 2026

Uh oh!

A-Words commented Aug 21, 2026

Uh oh!

rai-studio-bot commented Sep 21, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Nyanpasu 审查看板

Uh oh!

rai-studio-bot left a comment

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!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

This branch has not been deployed

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

A-Words commented Aug 11, 2026 •

edited by SigureMo

Loading

删除源码字符串断言（`aab0c72`）

CI 测试修正（`4bc9dda`）

主线集成提交 `14ccc2a` 的本地验证

rai-studio-bot commented Sep 21, 2026 •

edited

Loading