GSPO sequence-parallel (SDP/SP) paths are untested (loss + metrics)
还没有人认领这个 Issue。
评估
- 难度
- 4/5
- 预计耗时
- 3-5 天
- 新手友好度
- 25/100
- Issue 类型
- 缺陷
- 描述清晰度
- 基本清楚
- 活跃度
- 冷清
- 技术栈
- distributed-systems, python
- 领域
- testing-qa
调研方向
Start in fast_llm/layers/language_model/loss/policy_gradient.py, where fused_gspo_loss_forward_backward and compute_gspo_metrics all-reduce over sdp_group and sp_group. Then read _run_lm_loss_distributed and _test_gspo_metrics in tests/layers/test_lm_losses.py, which only exercise the vocab-parallel group. Done means a sequence-split test path that checks the reduced segment aggregates against a single-rank reference, plus a gspo loss subtest.
由索引模型根据 Issue 内容生成。
描述
Claude Opus 4.8 note: surfaced during review of #555.
Both the GSPO loss kernel (fused_gspo_loss_forward_backward) and the GSPO diagnostic metrics (compute_gspo_metrics) in fast_llm/layers/language_model/loss/policy_gradient.py all-reduce their per-segment buffers over the sequence-data-parallel (sdp_group) and sequence-parallel (sp_group) groups, to reconstruct per-document/segment aggregates when the sequence is split across those ranks.
Gap: the distributed test harness (_run_lm_loss_distributed / _test_gspo_metrics in tests/layers/test_lm_losses.py) only ever drives the vocab-parallel group (it splits the vocab dimension). sdp_group/sp_group stay at their None defaults, so those all_reduce branches never execute under test. There is also no gspo loss subtest in _run_lm_loss_distributed at all — only grpo, grpo_metrics, and gspo_metrics.
Why it needs its own effort: exercising the SDP/SP reduction requires a sequence-split test path (partition tokens/documents across ranks along the sequence dimension, then verify the reduced segment aggregates match a single-rank reference), which is beyond the current 2-rank vocab-split harness.
Single-rank and vocab-parallel paths are tested and verified; this issue tracks the remaining SDP/SP coverage for both the GSPO loss and its metrics.
- 主要语言
- Python
- 星标
- 337
- 派生
- 52
- PR 合并指标
- 30 天内没有已合并 PR
环境准备
- 提供 Dockerfile 或 Docker Compose 文件
- 有 Pull Request 模板
- 阅读贡献指南
从这里开始
- 先读完整个 Issue，再读项目的贡献指南。
- 在 Issue 下留言说明你要接手 —— 这能避免两个人做同样的事。
- Fork 仓库，在一个分支上完成修改。
- 提交 Pull Request，并在描述里引用这个 Issue 编号。
ServiceNow/Fast-LLM 的其他 Issue
-
难度 4/5 3-5 天 新手友好度 30/100
ServiceNow/Fast-LLM#538 ·
-
enhancement
难度 5/5 一周以上 新手友好度 12/100
ServiceNow/Fast-LLM#533 ·
-
难度 5/5 一周以上 新手友好度 15/100
ServiceNow/Fast-LLM#510 · 1 条评论 ·
-
难度 5/5 一周以上 新手友好度 12/100
ServiceNow/Fast-LLM#507 ·
-
难度 4/5 3-5 天 新手友好度 25/100
ServiceNow/Fast-LLM#500 ·
查看 ServiceNow/Fast-LLM 的全部 Issue
相似的 Issue
-
enhancement
难度 2/5 1-3 小时 新手友好度 78/100
glossyio/traffic-monitor#264 ·
维护者通常 1 天内回复
-
Add self-host未关闭
难度 1/5 1 小时以内 新手友好度 75/100
Atarity/deploy-your-own-saas#102 ·
-
[Bug]: RDFlex silently drops the covariate adjustment when the outcome or treatment column is integer-typed可能已有人在做 @vrjdev 今天认领。 未关闭
难度 2/5 1-3 小时 新手友好度 78/100
DoubleML/doubleml-for-py#424 ·
维护者通常 1 天内回复
-
agentic-dmx docs enhancement
难度 1/5 1-3 小时 新手友好度 86/100
-
难度 1/5 1 小时以内 新手友好度 72/100
521xueweihan/HelloGitHub#4150 ·