Repository navigation
Conversation
d1c5593 to
b08f81b
Compare
46a5d08 to
ce98541
Compare
ce98541 to
23280cc
Compare
|
/check-pr |
|
/retest |
|
🔵 The pipeline #1820 is running. Please wait a moment... (Link) |
|
|
@uhuuh, your gate pipeline failed at the code check stage. You can configure some code check tools locally following the (pre-push quick guide), to enable automatic checking when you push your code every time. Tools included are: Check_CMakelint, Check_Codespell, Check_Cpplint, Check_Lizard, Check_Pylint, Check_ClangFormat, Check_ShellCheck. |
23280cc to
c7c522e
Compare
|
/retest |
|
🔵 The pipeline #1824 is running. Please wait a moment... (Link) |
|
|
@uhuuh, your gate pipeline failed at the code check stage. You can configure some code check tools locally following the (pre-push quick guide), to enable automatic checking when you push your code every time. Tools included are: Check_CMakelint, Check_Codespell, Check_Cpplint, Check_Lizard, Check_Pylint, Check_ClangFormat, Check_ShellCheck. |
c3f5e7809eaa485549a45626184f50888f694459 feat: integrate VeRL with native Hyper training 137cadb3a17a34333d090cb41dc596ee3d70bd33 fix: align VeRL Hyper configuration and document integration usage 3ce7798fb17b7f8afa460b5766873af1af481816 test: cover Hyper engine validation and optional entropy chunking 54607e7517d2671a31257bb4b1c34f081f44d85b fix: quote VeRL checkpoint paths for Hydra overrides e60f573effc8e2093be3a4a2f1ef95297871ccb3 Revert "fix: quote VeRL checkpoint paths for Hydra overrides" 67a18cbfd0a1a5226900d6eb8096be7b77b1e4ce docs: update the VeRL integration guides 1d940b814a91c0a0b0272c2cf0cabe69de17e764 fix: align VeRL actor progress with optimizer steps
c7c522e to
bc708a2
Compare
|
/retest |
|
🔵 The pipeline #1838 is running. Please wait a moment... (Link) |
|
|
@uhuuh, your gate pipeline failed at the code check stage. You can configure some code check tools locally following the (pre-push quick guide), to enable automatic checking when you push your code every time. Tools included are: Check_CMakelint, Check_Codespell, Check_Cpplint, Check_Lizard, Check_Pylint, Check_ClangFormat, Check_ShellCheck. |
Paired: GitHub #1609 ↔ GitCode !1983
What type of PR is this?
/kind feature
What does this PR do / why do we need it:
新增 VeRL 的 Hyper 后端,使 VeRL 能使用 Hyper 的模型构建、并行执行和优化器能力完成训练。
train_batch()中统一管理训练步生命周期,仅训练阶段触发训练回调。HyperEngineConfig加载原生 recipe,支持使用trainer_config覆盖配置;在patch包中注册后端及 VeRL 兼容修复,无需修改 VeRL 源码。实验信息
对比汇总
全 10 步中,
training/rollout_actor_probs_pearson_corr的范围为 MindSpeed 0.999066234–0.999530435、Hyper 0.998790264–0.999674618,波动幅度分别为 0.000464201 和 0.000884354。actor/grad_norm的范围为 MindSpeed 0.000000–0.239585、Hyper 0.000000–0.515615,波动幅度分别为 0.239585 和 0.515615。波动幅度按最大值减最小值计算。各后端的 Pearson 相关系数比较其 actor 与自身 rollout 概率。training/rollout_actor_probs_pearson_corractor/grad_normgrad norm 任一侧为 0 时,相对变化记为「—」。
首步与后续各步差异较大,下表为第 2 步至第 10 步的汇总数据对比。
actor/perf/max_memory_allocated_gbmaxactor/perf/max_memory_reserved_gbmaxactor/perf/cpu_memory_used_gbmeanactor/perf/cpu_memory_used_gbmaxtiming_s/old_log_probmeantiming_s/update_actormeantiming_s/genmeantiming_s/update_weightsmeantiming_s/stepmeanperf/total_num_tokensmeantiming_per_token_ms/update_actormean相对变化以 MindSpeed 为基准。
逐步明细
以下列出第 2–10 步。M = MindSpeed,H = Hyper。
输入与优势
perf/total_num_tokensM / Hcritic/advantages/maxM / H资源占用
前两列为 Device 峰值显存,最后一列为 Host 整机已用内存,单位均为 GiB。
actor/perf/max_memory_allocated_gbM / Hactor/perf/max_memory_reserved_gbM / Hactor/perf/cpu_memory_used_gbM / H阶段性能
timing_s/old_log_probM / Htiming_s/update_actorM / Htiming_s/genM / Htiming_s/update_weightsM / Htiming_s/stepM / Htiming_per_token_ms/update_actorM / HWhich issue(s) this PR fixes:
无。
Code review checklist [illustration]: