{"id":78332,"date":"2026-02-27T03:34:11","date_gmt":"2026-02-26T19:34:11","guid":{"rendered":"https:\/\/www.wsisp.com\/helps\/78332.html"},"modified":"2026-02-27T03:34:11","modified_gmt":"2026-02-26T19:34:11","slug":"%e5%9f%ba%e4%ba%8everl%e6%a1%86%e6%9e%b6%e7%9a%84gspo%e7%ae%97%e6%b3%95%e5%9c%a8atlas-800t-a2%e6%9c%8d%e5%8a%a1%e5%99%a8%e4%b8%8a%e5%ae%9e%e8%b7%b5","status":"publish","type":"post","link":"https:\/\/www.wsisp.com\/helps\/78332.html","title":{"rendered":"\u57fa\u4e8eVeRL\u6846\u67b6\u7684GSPO\u7b97\u6cd5\u5728Atlas 800T A2\u670d\u52a1\u5668\u4e0a\u5b9e\u8df5"},"content":{"rendered":"

\u4f5c\u8005：\u6607\u817e\u5b9e\u6218\u6d3e \u77e5\u8bc6\u5730\u56fe：\u5f3a\u5316\u5b66\u4e60\u77e5\u8bc6\u5730\u56fe<\/p>\n

\u80cc\u666f\u4e0e\u610f\u4e49<\/h2>\n

\u672c\u7bc7\u6587\u7ae0\u4e3b\u8981\u57fa\u4e8eVeRL\u6846\u67b6\u4e0a\u63d0\u51fa\u7684GSPO\u7b97\u6cd5\u5728\u6607\u817eNPU\u4e0a\u8fdb\u884c\u5b9e\u8df5\u90e8\u7f72，\u5e76\u4e3a\u5927\u5bb6\u7b80\u5355\u4ecb\u7ecdGRPO\u7b97\u6cd5\u601d\u60f3\u4ee5\u53ca\u5176\u548cGSPO\u7b97\u6cd5\u7279\u6027\u5dee\u5f02\u3002<\/p>\n

\u7b97\u6cd5\u539f\u7406<\/h3>\n

\u8bba\u6587\u5730\u5740 GRPO：https:\/\/arxiv.org\/abs\/2402.03300 GSPO：https:\/\/arxiv.org\/abs\/2507.18071<\/p>\n

GRPO：\u7fa4\u7ec4\u76f8\u5bf9\u7b56\u7565\u4f18\u5316<\/h4>\n

\u7fa4\u7ec4\u76f8\u5bf9\u7b56\u7565\u4f18\u5316（GRPO）\u7684\u6838\u5fc3\u521b\u65b0\u5728\u4e8e\u6d88\u9664\u4e86\u5bf9\u8ba1\u7b97\u5bc6\u96c6\u578b\u4ef7\u503c\u6a21\u578b\u7684\u4f9d\u8d56，\u8be5\u7b97\u6cd5\u91c7\u7528\u4e86\u4e00\u79cd\u521b\u65b0\u7684\u7fa4\u7ec4\u751f\u6210\u548c\u76f8\u5bf9\u8bc4\u4f30\u673a\u5236：\u5bf9\u4e8e\u7ed9\u5b9a\u7684\u8f93\u5165\u63d0\u793a，\u7cfb\u7edf\u751f\u6210G\u4e2a\u4e0d\u540c\u7684\u54cd\u5e94\u6784\u6210\u4e00\u4e2a\u7fa4\u7ec4，\u968f\u540e\u5956\u52b1\u6a21\u578b\u5bf9\u7fa4\u7ec4\u5185\u6240\u6709\u54cd\u5e94\u8fdb\u884c\u8bc4\u5206\u3002\u901a\u8fc7\u8ba1\u7b97\u7fa4\u7ec4\u5185\u5206\u6570\u7684\u5747\u503c\u548c\u6807\u51c6\u5dee，\u7b97\u6cd5\u4e3a\u6bcf\u4e2a\u54cd\u5e94\u8ba1\u7b97\u76f8\u5bf9\u4f18\u52bf\u503c<\/p>\n

A<\/p>\n

^<\/p>\n

i<\/p>\n

\\\\hat{A}_i<\/p>\n

<\/span><\/span><\/span>A<\/span><\/span><\/span>^<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>i<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>\u3002\u4f18\u4e8e\u7fa4\u7ec4\u5e73\u5747\u6c34\u5e73\u7684\u54cd\u5e94\u83b7\u5f97\u6b63\u5411\u4f18\u52bf，\u53cd\u4e4b\u5219\u83b7\u5f97\u8d1f\u5411\u4f18\u52bf\u3002\u8fd9\u79cd\u8bbe\u8ba1\u663e\u8457\u964d\u4f4e\u4e86\u5f3a\u5316\u5b66\u4e60\u8bad\u7ec3\u7684\u5185\u5b58\u5360\u7528\u548c\u8ba1\u7b97\u590d\u6742\u5ea6，\u4f7f\u5927\u89c4\u6a21\u6a21\u578b\u7684\u8bad\u7ec3\u53d8\u5f97\u66f4\u52a0\u9ad8\u6548\u548c\u53ef\u884c，\u4f46\u5176\u5e95\u5c42\u5b9e\u73b0\u5b58\u5728\u4e00\u4e2a\u5173\u952e\u7684\u8bbe\u8ba1\u7f3a\u9677，\u8be5\u7f3a\u9677\u5728\u5927\u89c4\u6a21\u6a21\u578b\u8bad\u7ec3\u4e2d\u4f1a\u5bfc\u81f4\u4e25\u91cd\u7684\u7a33\u5b9a\u6027\u95ee\u9898\u3002<\/p>\n

\u95ee\u9898\u7684\u6839\u6e90\u5728\u4e8e\u5956\u52b1\u5206\u914d\u4e0e\u4f18\u5316\u66f4\u65b0\u4e4b\u95f4\u7684\u7c92\u5ea6\u4e0d\u5339\u914d：\u5956\u52b1\u503c<\/p>\n

A<\/p>\n

^<\/p>\n

i<\/p>\n

\\\\hat{A}_i<\/p>\n

<\/span><\/span><\/span>A<\/span><\/span><\/span>^<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>i<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>\u662f\u57fa\u4e8e\u5b8c\u6574\u5e8f\u5217\u8ba1\u7b97\u5f97\u51fa，\u800cGRPO\u7684\u4f18\u5316\u66f4\u65b0\u5374\u5728token\u7ea7\u522b\u6267\u884c\u3002\u4e3a\u4e86\u5c06\u5e8f\u5217\u7ea7\u5956\u52b1\u5e94\u7528\u4e8e\u6bcf\u4e2atoken，GRPO\u5f15\u5165\u4e86token\u7ea7\u522b\u7684\u91cd\u8981\u6027\u6743\u91cd<\/p>\n

w<\/p>\n

i<\/p>\n

,<\/p>\n

t<\/p>\n

(<\/p>\n

\u03b8<\/p>\n

)<\/p>\n

w_{i,t}(\\\\theta)<\/p>\n

<\/span><\/span>w<\/span><\/span>i<\/span>,<\/span>t<\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>(<\/span>\u03b8<\/span>)<\/span><\/span><\/span><\/span><\/span>，\u6743\u91cd<\/p>\n

w<\/p>\n

i<\/p>\n

,<\/p>\n

t<\/p>\n

(<\/p>\n

\u03b8<\/p>\n

)<\/p>\n

w_{i,t}(\\\\theta)<\/p>\n

<\/span><\/span>w<\/span><\/span>i<\/span>,<\/span>t<\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>(<\/span>\u03b8<\/span>)<\/span><\/span><\/span><\/span><\/span>\u8868\u793a\u5f53\u524d\u6a21\u578b\u751f\u6210 token \u7684\u6982\u7387 \u00f7 \u65e7\u6a21\u578b\u751f\u6210\u76f8\u540c token \u7684\u6982\u7387\u6bd4\u503c，\u5982\u4e0b\u56fe\u4e2d\u7684<\/p>\n

w<\/p>\n

i<\/p>\n

,<\/p>\n

t<\/p>\n

(<\/p>\n

\u03b8<\/p>\n

)<\/p>\n

w_{i,t}(\\\\theta)<\/p>\n

<\/span><\/span>w<\/span><\/span>i<\/span>,<\/span>t<\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>(<\/span>\u03b8<\/span>)<\/span><\/span><\/span><\/span><\/span> \"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\"<\/p>\n

\u5e8f\u5217\u5185\u5404token\u7684\u91cd\u8981\u6027\u6743\u91cd\u53ef\u80fd\u51fa\u73b0\u663e\u8457\u5dee\u5f02，\u5bfc\u81f4\u5b66\u4e60\u4fe1\u53f7\u7684\u566a\u58f0\u5316\u548c\u4e0d\u4e00\u81f4\u6027\u3002\u968f\u7740\u8bad\u7ec3\u5e8f\u5217\u957f\u5ea6\u7684\u589e\u52a0，\u8fd9\u79cd\u566a\u58f0\u6548\u5e94\u7d2f\u79ef\u5e76\u53ef\u80fd\u89e6\u53d1\u6574\u4e2a\u8bad\u7ec3\u8fc7\u7a0b\u7684\u5931\u7a33，\u6700\u7ec8\u5bfc\u81f4\u6a21\u578b\u5d29\u6e83\u3002\u8be5\u95ee\u9898\u5728\u7a00\u758f\u4e13\u5bb6\u6df7\u5408（Mixture-of-Experts, MoE）\u6a21\u578b\u4e2d\u5c24\u4e3a\u4e25\u91cd，\u5728MoE\u67b6\u6784\u7684\u6a21\u578b\u8bad\u7ec3\u8fc7\u7a0b\u4e2d，\u7531\u4e8e\u6a21\u578b\u66f4\u65b0\u540e\u6bcf\u6b21\u6fc0\u6d3b\u7684\u4e13\u5bb6\u53ef\u80fd\u4f1a\u53d1\u751f\u53d8\u5316，off-policy\u504f\u5dee\u4f1a\u53d8\u5f97\u66f4\u4e25\u91cd\u3002\u56e0\u6b64，GRPO\u7684\u8fd9\u79cd\u91cd\u8981\u6027\u91c7\u6837\u7684\u65b9\u5f0f\u6709\u53ef\u80fd\u4f1a\u5bfc\u81f4\u66f4\u5927\u7684\u504f\u5dee，\u81f4\u4f7f\u8bad\u7ec3\u5d29\u6e83\u3002<\/p>\n

GSPO：\u7fa4\u7ec4\u5e8f\u5217\u7b56\u7565\u4f18\u5316<\/h4>\n

\u4e0a\u8ff0\u95ee\u9898\u53ef\u4ee5\u53d1\u73b0，\u5176\u6839\u6e90\u5728\u4e8e\u4f18\u5316\u7c92\u5ea6\u548c\u5956\u52b1\u7c92\u5ea6\u4e0d\u5728\u540c\u4e00\u5355\u4f4d\u3002\u6240\u4ee5GSPO\u7b97\u6cd5\u63d0\u51fa\u5e8f\u5217\u7ea7\u522b\u7684\u91cd\u8981\u6027\u91c7\u6837，\u5e76\u5f15\u5165<\/p>\n

y<\/p>\n

i<\/p>\n

y_i<\/p>\n

<\/span><\/span>y<\/span><\/span>i<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>\u4fdd\u8bc1\u5728\u6570\u503c\u4e0a\u4fdd\u6301\u7a33\u5b9a，\u65e0\u8bba\u5e8f\u5217\u957f\u5ea6\u4e3a10\u4e2atoken\u8fd8\u662f1000\u4e2atoken\u3002\u8be5\u7b97\u6cd5\u4f7f\u7528\u7a33\u5b9a\u7684\u5e8f\u5217\u7ea7\u91cd\u8981\u6027\u6bd4\u7387<\/p>\n

s<\/p>\n

i<\/p>\n

(<\/p>\n

\u03b8<\/p>\n

)<\/p>\n

s_i(\\\\theta)<\/p>\n

<\/span><\/span>s<\/span><\/span>i<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>(<\/span>\u03b8<\/span>)<\/span><\/span><\/span><\/span><\/span>\u66ff\u4ee3\u4e86\u566a\u58f0\u8f83\u5927\u7684token\u7ea7\u522b\u6743\u91cd，\u7ed9\u5b9a\u5e8f\u5217\u5185\u7684\u6240\u6709token\u63a5\u6536\u5b8c\u5168\u4e00\u81f4\u7684\u66f4\u65b0，\u8be5\u66f4\u65b0\u7531<\/p>\n

s<\/p>\n

i<\/p>\n

(<\/p>\n

\u03b8<\/p>\n

)<\/p>\n

A<\/p>\n

^<\/p>\n

i<\/p>\n

s_i(\\\\theta)\\\\hat{A}_i<\/p>\n

<\/span><\/span>s<\/span><\/span>i<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>(<\/span>\u03b8<\/span>)<\/span><\/span>A<\/span><\/span><\/span>^<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>i<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>\u786e\u5b9a\u3002token\u7ea7\u522b\u7684\u4e0d\u4e00\u81f4\u53cd\u9988\u88ab\u6d88\u9664，\u53d6\u800c\u4ee3\u4e4b\u7684\u662f\u57fa\u4e8e\u5b8c\u6574\u5e8f\u5217\u5956\u52b1\u7684\u7edf\u4e00\u66f4\u65b0\u673a\u5236\u3002 \"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\"<\/p>\n

\u9488\u5bf9\u591a\u8f6e\u573a\u666f，GSPO\u63d0\u4f9btoken-level\u7684\u53d8\u4f53，\u4ee5\u8fdb\u884c\u66f4\u7cbe\u7ec6\u7684\u5e94\u7528 \"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" \u8bba\u6587\u4e2d\u5b9e\u9a8c\u6570\u636e\u53ca\u7b97\u6cd5\u8868\u660e，\u5bf9\u4e8eGRPO\u7b97\u6cd5\u6765\u8bf4，\u5728MoE\u6a21\u578b\u4e0a\u8fdb\u884c\u8bad\u7ec3，\u7531\u4e8e\u6a21\u578b\u66f4\u65b0\u540e\u6bcf\u6b21\u6fc0\u6d3b\u7684\u4e13\u5bb6\u53ef\u80fd\u4f1a\u53d1\u751f\u53d8\u5316，off-policy\u504f\u5dee\u4f1a\u53d8\u5f97\u66f4\u4e25\u91cd，\u800c\u8fd9\u79cd\u6df7\u5408\u4e13\u5bb6\u6fc0\u6d3b\u5bfc\u81f4\u7684\u6ce2\u52a8\u4f1a\u4e25\u91cd\u5f71\u54cd\u6a21\u578b\u6536\u655b\u3002\u5728\u56fa\u5b9a\u91c7\u6837\u548c\u8bad\u7ec3\u6fc0\u6d3b\u76f8\u540c\u7684\u4e13\u5bb6\u540e（Routing Replay），\u8bad\u7ec3reward\u53ef\u4ee5\u6b63\u5e38\u4e0a\u6da8，\u4f46\u4f1a\u5f15\u5165\u989d\u5916\u7684\u5185\u5b58\u548c\u901a\u4fe1\u5f00\u9500，\u9650\u5236MOE\u7684\u5b9e\u9645\u5bb9\u91cf\u3002\u800cGSPO\u7528sequence-level\u7684clip\u8fdb\u884c\u4f18\u5316，\u5bf9\u7cbe\u5ea6\u5dee\u5f02\u7684\u5bb9\u5fcd\u5ea6\u8981\u66f4\u9ad8，\u4ece\u6839\u672c\u4e0a\u89e3\u51b3\u4e86MoE\u6a21\u578b\u4e2d\u7684\u4e13\u5bb6\u6fc0\u6d3b\u6ce2\u52a8\u95ee\u9898，\u7b80\u5316\u548c\u7a33\u5b9a\u4e86\u8bad\u7ec3\u8fc7\u7a0b，\u8bad\u7ec3\u6548\u679c\u4e5f\u66f4\u597d<\/p>\n

\u7279\u6027\u9002\u914d\u5de5\u4f5c\u5206\u6790<\/h3>\n

\u6574\u4f53\u7b97\u6cd5\u6d41\u7a0b<\/h4>\n

Verl\u7684GRPO\u6574\u4f53\u7b97\u6cd5\u6d41\u7a0b\u5982\u4e0b\u56fe\u6240\u793a，GSPO\u4e0e\u5176\u7c7b\u4f3c，\u5dee\u5f02\u70b9\u5728\u4e8e\u91cd\u8981\u6027\u91c7\u6837\u8ba1\u7b97\u548closs\u8ba1\u7b97\u6a21\u5757<\/p>\n

\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\"<\/p>\n

\u6a21\u5757\u5206\u6790<\/h4>\n

\u7b97\u6cd5pr: https:\/\/github.com\/volcengine\/verl\/pull\/2775<\/p>\n

\u57fa\u4e8everl\u4ee3\u7801\u5206\u6790，\u53ef\u4ee5\u53d1\u73b0\u5176\u91cd\u8981\u6027\u91c7\u6837\u8ba1\u7b97\u76f8\u5173\u4ee3\u7801\u4f4d\u7f6e\u4f4d\u4e8e：verl\/trainer\/ppo\/core_algos.py<\/p>\n

@register_policy_loss<\/span>(<\/span>"gspo"<\/span>)<\/span>
\ndef<\/span> compute_policy_loss_gspo<\/span>(<\/span>
\n old_log_prob:<\/span> torch.<\/span>Tensor,<\/span>
\n log_prob:<\/span> torch.<\/span>Tensor,<\/span>
\n advantages:<\/span> torch.<\/span>Tensor,<\/span>
\n response_mask:<\/span> torch.<\/span>Tensor,<\/span>
\n loss_agg_mode:<\/span> str<\/span> =<\/span> "seq-mean-token-mean"<\/span>,<\/span>
\n config:<\/span> Optional[<\/span>ActorConfig]<\/span> =<\/span> None<\/span>,<\/span>
\n rollout_is_weights:<\/span> torch.<\/span>Tensor |<\/span> None<\/span> =<\/span> None<\/span>,<\/span>
\n)<\/span> –<\/span>><\/span> tuple<\/span>[<\/span>torch.<\/span>Tensor,<\/span> dict<\/span>[<\/span>str<\/span>,<\/span> Any]<\/span>]<\/span>:<\/span>
\n """
\n Compute the clipped policy objective and related metrics for GSPO.<\/p>\n

See https:\/\/arxiv.org\/pdf\/2507.18071 for more details.<\/p>\n

Args:
\n old_log_prob (torch.Tensor):
\n Log-probabilities of actions under the old policy, shape (batch_size, response_length).
\n log_prob (torch.Tensor):
\n Log-probabilities of actions under the current policy, shape (batch_size, response_length).
\n advantages (torch.Tensor):
\n Advantage estimates for each action, shape (batch_size, response_length).
\n response_mask (torch.Tensor):
\n Mask indicating which tokens to include in the loss, shape (batch_size, response_length).
\n loss_agg_mode (str, optional):
\n Aggregation mode for `agg_loss`. For GSPO, it is recommended to use "seq-mean-token-mean".
\n """<\/span><\/p>\n

assert<\/span> config is<\/span> not<\/span> None<\/span>
\n assert<\/span> isinstance<\/span>(<\/span>config,<\/span> ActorConfig)<\/span>
\n clip_ratio_low =<\/span> config.<\/span>clip_ratio_low if<\/span> config.<\/span>clip_ratio_low is<\/span> not<\/span> None<\/span> else<\/span> config.<\/span>clip_ratio
\n clip_ratio_high =<\/span> config.<\/span>clip_ratio_high if<\/span> config.<\/span>clip_ratio_high is<\/span> not<\/span> None<\/span> else<\/span> config.<\/span>clip_ratio<\/p>\n

negative_approx_kl =<\/span> log_prob –<\/span> old_log_prob<\/p>\n

# compute sequence-level importance ratio:<\/span>
\n # si(\u03b8) = (\u03c0_\u03b8(yi|x)\/\u03c0_\u03b8old(yi|x))^(1\/|yi|) =<\/span>
\n # exp [(1\/|y_i|) * \u03a3_t log(\u03c0_\u03b8(y_i,t|x,y_i,<t)\/\u03c0_\u03b8old(y_i,t|x,y_i,<t))]<\/span>
\n seq_lengths =<\/span> torch.<\/span>sum<\/span>(<\/span>response_mask,<\/span> dim=<\/span>–<\/span>1<\/span>)<\/span>.<\/span>clamp(<\/span>min<\/span>=<\/span>1<\/span>)<\/span>
\n negative_approx_kl_seq =<\/span> torch.<\/span>sum<\/span>(<\/span>negative_approx_kl *<\/span> response_mask,<\/span> dim=<\/span>–<\/span>1<\/span>)<\/span> \/<\/span> seq_lengths<\/p>\n

# Combined ratio at token level:<\/span>
\n # s_i,t(\u03b8) = sg[s_i(\u03b8)] \u00b7 \u03c0_\u03b8(y_i,t|x, y_i,<t) \/ sg[\u03c0_\u03b8(y_i,t|x, y_i,<t)]<\/span>
\n # In log space: log(s_i,t(\u03b8)) = sg[log(s_i(\u03b8))] + log_prob – sg[log_prob]<\/span>
\n log_seq_importance_ratio =<\/span> log_prob –<\/span> log_prob.<\/span>detach(<\/span>)<\/span> +<\/span> negative_approx_kl_seq.<\/span>detach(<\/span>)<\/span>.<\/span>unsqueeze(<\/span>–<\/span>1<\/span>)<\/span>
\n log_seq_importance_ratio =<\/span> torch.<\/span>clamp(<\/span>log_seq_importance_ratio,<\/span> max<\/span>=<\/span>10.0<\/span>)<\/span> # clamp for numerical stability<\/span><\/p>\n

# finaly exp() to remove log<\/span>
\n seq_importance_ratio =<\/span> torch.<\/span>exp(<\/span>log_seq_importance_ratio)<\/span><\/p>\n

pg_losses1 =<\/span> –<\/span>advantages *<\/span> seq_importance_ratio
\n pg_losses2 =<\/span> –<\/span>advantages *<\/span> torch.<\/span>clamp(<\/span>seq_importance_ratio,<\/span> 1<\/span> –<\/span> clip_ratio_low,<\/span> 1<\/span> +<\/span> clip_ratio_high)<\/span>
\n pg_losses =<\/span> torch.<\/span>maximum(<\/span>pg_losses1,<\/span> pg_losses2)<\/span><\/p>\n

# Apply rollout correction weights if provided<\/span>
\n if<\/span> rollout_is_weights is<\/span> not<\/span> None<\/span>:<\/span>
\n pg_losses =<\/span> pg_losses *<\/span> rollout_is_weights<\/p>\n

# for GSPO, we need to aggregate the loss at the sequence level (seq-mean-token-mean)<\/span>
\n pg_loss =<\/span> agg_loss(<\/span>
\n loss_mat=<\/span>pg_losses,<\/span> loss_mask=<\/span>response_mask,<\/span> loss_agg_mode=<\/span>"seq-mean-token-mean"<\/span>,<\/span> **<\/span>config.<\/span>global_batch_info
\n )<\/span><\/p>\n

# For compatibility, return zero for pg_clipfrac_lower (not used in standard GSPO)<\/span>
\n pg_clipfrac =<\/span> verl_F.<\/span>masked_mean(<\/span>torch.<\/span>gt(<\/span>pg_losses2,<\/span> pg_losses1)<\/span>.<\/span>float<\/span>(<\/span>)<\/span>,<\/span> response_mask)<\/span>
\n pg_clipfrac_lower =<\/span> torch.<\/span>tensor(<\/span>0.0<\/span>,<\/span> device=<\/span>pg_loss.<\/span>device)<\/span><\/p>\n

ppo_kl =<\/span> verl_F.<\/span>masked_mean(<\/span>–<\/span>negative_approx_kl,<\/span> response_mask)<\/span>
\n pg_metrics =<\/span> {<\/span>
\n "actor\/pg_clipfrac"<\/span>:<\/span> pg_clipfrac.<\/span>detach(<\/span>)<\/span>.<\/span>item(<\/span>)<\/span>,<\/span>
\n "actor\/ppo_kl"<\/span>:<\/span> ppo_kl.<\/span>detach(<\/span>)<\/span>.<\/span>item(<\/span>)<\/span>,<\/span>
\n "actor\/pg_clipfrac_lower"<\/span>:<\/span> pg_clipfrac_lower.<\/span>detach(<\/span>)<\/span>.<\/span>item(<\/span>)<\/span>,<\/span>
\n }<\/span>
\n return<\/span> pg_loss,<\/span> pg_metrics<\/p>\n

GRPO\u57fa\u4e8etoken\u7ea7\u522b\u8ba1\u7b97\u91cd\u8981\u6027ratio,GSPO\u57fa\u4e8esequence\u7ea7\u522b\u8ba1\u7b97\u91cd\u8981\u6027ratio，\u76f8\u5173\u4ee3\u7801\u5dee\u5f02\u5982\u4e0b\u56fe\u6240\u793a<\/p>\n

\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\"<\/p>\n

\u7ec4\u4ef6\u5206\u6790<\/h4>\n

\u80cc\u666f：GRPO\u5df2\u5728NPU\u652f\u6301<\/p>\n

\u89c2\u6d4b\u73b0\u6709GRPO\u548cGSPO\u7684\u4ee3\u7801\u5dee\u5f02\u5206\u6790，\u4ec5\u6d89\u53ca\u5e38\u89c4torch\u7f16\u7801，\u4e0d\u6d89\u53ca\u5176\u4f59\u7ec4\u4ef6，\u4e0d\u6d89\u53caNPU\u9002\u914d\u5de5\u4f5c<\/p>\n

\u8c03\u8bd5\u7528\u4f8b<\/h2>\n

\u8c03\u8bd5\u76ee\u6807<\/p>\n

\u914d\u7f6eGSPO\u76f8\u5173\u53c2\u6570，\u62c9\u8d77\u8bad\u7ec3，reward\u66f2\u7ebf\u6b63\u5e38\u4e0a\u5347<\/p>\n

\u57fa\u4e8eGRPO\u811a\u672c，\u53c2\u6570\u4f7f\u80fd<\/p>\n

loss_mode=<\/span>gspo
\nloss_agg_mode=<\/span>"seq-mean-token-mean"<\/span>
\nactor_rollout_ref.<\/span>actor.<\/span>policy_loss.<\/span>loss_mode=<\/span>${<\/span>loss_mode}<\/span> \\\\
\nactor_rollout_ref.<\/span>actor.<\/span>loss_agg_mode=<\/span>${<\/span>loss_agg_mode}<\/span> \\\\<\/p>\n

\u8c03\u8bd5\u7528\u4f8b1：Qwen25-3B<\/p>\n

\u7a20\u5bc6\u6a21\u578b\u8c03\u8bd5FSDP\u540e\u7aef，\u4ee5\u5b98\u65b9\u811a\u672c\u4e3a\u51c6，\u89c2\u6d4breward\u66f2\u7ebf\u662f\u5426\u4e0a\u5347<\/p>\n

\u8c03\u8bd5\u7528\u4f8b2：Qwen3-30B-A3B<\/p>\n

MOE\u6a21\u578b\u8c03\u8bd5Megatron\u540e\u7aef，\u89c2\u6d4bGRPO\u548cGSPO\u4e24\u79cd\u7b97\u6cd5\u7684reward\u66f2\u7ebf<\/p>\n

\u8c03\u8bd5\u5b9e\u8df5<\/h2>\n

\u8c03\u8bd5\u73af\u5883<\/h3>\n\n\u914d\u7f6e\u9879\u7248\u672c\u4fe1\u606f<\/tr>\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n
AI\u670d\u52a1\u5668<\/td>\nAtlas 800T A2 64G<\/td>\n<\/tr>\n
\u9a71\u52a8\u3001\u56fa\u4ef6<\/td>\n24.1.0.3<\/td>\n<\/tr>\n
Python<\/td>\n3.10.12<\/td>\n<\/tr>\n
CANN<\/td>\n8.2.RC2<\/td>\n<\/tr>\n
torch<\/td>\n2.7.1<\/td>\n<\/tr>\n
torch_npu<\/td>\n2.7.1<\/td>\n<\/tr>\n
transformer<\/td>\n4.53.3<\/td>\n<\/tr>\n
vllm<\/td>\n0.10.0<\/td>\n<\/tr>\n
vllm-ascend<\/td>\n0.10.0rc1<\/td>\n<\/tr>\n
verl<\/td>\n0.7.0.dev0<\/td>\n<\/tr>\n
Megatron-core<\/td>\n0.12.1<\/td>\n<\/tr>\n
MindSpeed<\/td>\n2.2.0_core_r0.12.1<\/td>\n<\/tr>\n
Mbridge<\/td>\n0.13.1<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

\u8c03\u8bd5\u7528\u4f8b1：Qwen25-3B<\/h3>\n

\u8c03\u8bd5\u811a\u672c<\/h4>\n

set<\/span> -x
\npkill<\/span> -9 python
\nps<\/span> -ef |<\/span> grep<\/span> "python"<\/span> |<\/span>grep<\/span> -v grep<\/span> |<\/span>awk<\/span> '{print $2}'<\/span> |<\/span>xargs<\/span> -t -i kill<\/span> -9 {<\/span>}<\/span>
\nray stop –force
\npkill<\/span> -9 python
\npkill<\/span> -9 torchrun
\nps<\/span> -ef |<\/span> grep<\/span> "defaunct"<\/span> |<\/span>grep<\/span> python |<\/span>awk<\/span> '{print $3}'<\/span> |<\/span>xargs<\/span> -t -i kill<\/span> -9 {<\/span>}<\/span>
\nps<\/span> -ef |<\/span> grep<\/span> "defaunct"<\/span> |<\/span>grep<\/span> torchrun |<\/span>awk<\/span> '{print $3}'<\/span> |<\/span>xargs<\/span> -t -i kill<\/span> -9 {<\/span>}<\/span>
\nps<\/span> -ef |<\/span> grep<\/span> -i python |<\/span>grep<\/span> -i [<\/span>name]<\/span> |<\/span>grep<\/span> -v grep<\/span> |<\/span>awk<\/span> '{print $2}'<\/span> |<\/span>xargs<\/span> -t -I {<\/span>}<\/span> kill<\/span> -9 {<\/span>}<\/span>
\nps<\/span> -ef |<\/span> grep<\/span> -i torchrun |<\/span>grep<\/span> -i [<\/span>name]<\/span> |<\/span>grep<\/span> -v grep<\/span> |<\/span>awk<\/span> '{print $2}'<\/span> |<\/span>xargs<\/span> -t -I {<\/span>}<\/span> kill<\/span> -9 {<\/span>}<\/span>
\nps<\/span> -ef |<\/span> grep<\/span> "python"<\/span> |<\/span>grep<\/span> -v grep<\/span> |<\/span>awk<\/span> '{print $2}'<\/span> |<\/span>xargs<\/span> -t -i kill<\/span> -9 {<\/span>}<\/span>
\n# Set how many GPUs we actually have on this node.<\/span>
\nexport<\/span> GPUS_PER_NODE<\/span>=<\/span>8<\/span>
\nexport<\/span> NNODES<\/span>=<\/span>1<\/span>
\necho<\/span> "Using $NNODES<\/span> nodes for training…"<\/span>
\n# ————————————- Setup xp params —————————————<\/span>
\nproject_name<\/span>=<\/span>'RL-GSPO'<\/span>
\nadv_estimator<\/span>=<\/span>grpo
\nloss_mode<\/span>=<\/span>gspo
\nloss_agg_mode<\/span>=<\/span>"seq-mean-token-mean"<\/span>
\nMODEL_PATH<\/span>=<\/span>XX\/Qwen25-3B-Instruct
\noffload<\/span>=<\/span>false # it's a small model, offloading will just slow-down training<\/span>
\nrollout_engine<\/span>=<\/span>vllm
\nrollout_mode<\/span>=<\/span>sync # can be async to speedup large scale xps<\/span>
\ngpu_memory_utilization<\/span>=<\/span>0.6<\/span>
\nreward_manager<\/span>=<\/span>dapo
\nadv_estimator<\/span>=<\/span>grpo
\nshuffle_dataset<\/span>=<\/span>true
\nfirst_time_dataset_prep<\/span>=<\/span>true # prepare dataset<\/span>
\ntest_freq<\/span>=<\/span>10<\/span>
\nsave_freq<\/span>=<\/span>10<\/span>
\ntotal_epochs<\/span>=<\/span>10<\/span>
\ntotal_training_steps<\/span>=<\/span>500<\/span>
\nval_before_train<\/span>=<\/span>false
\nuse_kl_in_reward<\/span>=<\/span>false
\nkl_coef<\/span>=<\/span>0.0<\/span>
\nuse_kl_loss<\/span>=<\/span>false
\nkl_loss_coef<\/span>=<\/span>0.0<\/span>
\nclip_ratio_low<\/span>=<\/span>0.0003<\/span> # as recommended by the paper, see Sec. 5.1<\/span>
\nclip_ratio_high<\/span>=<\/span>0.0004<\/span> # as recommended by the paper, see Sec. 5.1<\/span>
\ntrain_batch_size<\/span>=<\/span>512<\/span>
\nppo_mini_batch_size<\/span>=<\/span>128<\/span> # maintain 4 mini-batches as recommended by the paper, see Sec. 5.1<\/span>
\nppo_micro_batch_size_per_gpu<\/span>=<\/span>8<\/span> # setup depending on your GPU memory<\/span>
\nn_resp_per_prompt<\/span>=<\/span>16<\/span>
\nmax_prompt_length<\/span>=<\/span>$((<\/span>1024<\/span> *<\/span> 2<\/span>))<\/span><\/span>
\nmax_response_length<\/span>=<\/span>$((<\/span>1024<\/span> *<\/span> 8<\/span>))<\/span><\/span>
\n# dapo reward manager params<\/span>
\nenable_overlong_buffer<\/span>=<\/span>false # true<\/span>
\noverlong_buffer_len<\/span>=<\/span>$((<\/span>1024<\/span> *<\/span> 4<\/span>))<\/span><\/span>
\noverlong_penalty_factor<\/span>=<\/span>1.0<\/span>
\n# Paths and namings<\/span>
\nSFT_MODEL<\/span>=<\/span>$(<\/span>basename<\/span> $MODEL_PATH)<\/span><\/span>
\nexp_name<\/span>=<\/span>"${loss_mode}<\/span>-epslow-${clip_ratio_low}<\/span>-epshigh-${clip_ratio_high}<\/span>–${SFT_MODEL}<\/span>-RL"<\/span>
\nCKPTS_DIR<\/span>=<\/span>\/rl\/checkpoints\/experimental\/4b\/${loss_mode}<\/span>\/${exp_name}<\/span>
\n# Sampling params at rollouts<\/span>
\ntemperature<\/span>=<\/span>1.0<\/span>
\ntop_p<\/span>=<\/span>1.0<\/span>
\ntop_k<\/span>=<\/span>-1 # 0 for HF rollout, -1 for vLLM rollout<\/span>
\nval_top_p<\/span>=<\/span>0.7<\/span>
\n# Performance Related Parameter<\/span>
\nsp_size<\/span>=<\/span>4<\/span>
\nuse_dynamic_bsz<\/span>=<\/span>true
\nactor_ppo_max_token_len<\/span>=<\/span>$((<\/span>(<\/span>max_prompt_length +<\/span> max_response_length)<\/span> *<\/span> 1<\/span>))<\/span><\/span>
\ninfer_ppo_max_token_len<\/span>=<\/span>$((<\/span>(<\/span>max_prompt_length +<\/span> max_response_length)<\/span> *<\/span> 1<\/span>))<\/span><\/span>
\noffload<\/span>=<\/span>true
\ngen_tp<\/span>=<\/span>2<\/span>
\nentropy_checkpointing<\/span>=<\/span>true # This enables entropy recomputation specifically for the entropy calculation, lowering memory usage during training.<\/span>
\ngsm8k_train_path<\/span>=<\/span>xx\/gsm8k\/post_data\/gsm8k\/train.parquet
\ngsm8k_test_path<\/span>=<\/span>xx\/gsm8k\/post_data\/gsm8k\/test.parquet
\n# set the paths<\/span>
\ntrain_files<\/span>=<\/span>"['$gsm8k_train_path<\/span>']"<\/span>
\ntest_files<\/span>=<\/span>"['$gsm8k_test_path<\/span>']"<\/span>
\n#! \u4fee\u6539filter_overlong_prompts false<\/span>
\npython3 -m verl.trainer.main_ppo \\\\<\/span>
\n algorithm.adv_estimator=<\/span>${adv_estimator}<\/span> \\\\<\/span>
\n actor_rollout_ref.actor.policy_loss.loss_mode=<\/span>${loss_mode}<\/span> \\\\<\/span>
\n data.train_files=<\/span>"${train_files}<\/span>"<\/span> \\\\<\/span>
\n data.val_files=<\/span>"${test_files}<\/span>"<\/span> \\\\<\/span>
\n data.shuffle=<\/span>$shuffle_dataset<\/span> \\\\<\/span>
\n data.prompt_key=<\/span>prompt \\\\<\/span>
\n data.truncation=<\/span>'error'<\/span> \\\\<\/span>
\n data.filter_overlong_prompts=<\/span>true \\\\<\/span>
\n data.train_batch_size=<\/span>${train_batch_size}<\/span> \\\\<\/span>
\n data.max_prompt_length=<\/span>${max_prompt_length}<\/span> \\\\<\/span>
\n data.max_response_length=<\/span>${max_response_length}<\/span> \\\\<\/span>
\n actor_rollout_ref.rollout.n=<\/span>${n_resp_per_prompt}<\/span> \\\\<\/span>
\n algorithm.use_kl_in_reward=<\/span>${use_kl_in_reward}<\/span> \\\\<\/span>
\n algorithm.kl_ctrl.kl_coef=<\/span>${kl_coef}<\/span> \\\\<\/span>
\n actor_rollout_ref.actor.use_kl_loss=<\/span>${use_kl_loss}<\/span> \\\\<\/span>
\n actor_rollout_ref.actor.kl_loss_coef=<\/span>${kl_loss_coef}<\/span> \\\\<\/span>
\n actor_rollout_ref.actor.clip_ratio_low=<\/span>${clip_ratio_low}<\/span> \\\\<\/span>
\n actor_rollout_ref.actor.clip_ratio_high=<\/span>${clip_ratio_high}<\/span> \\\\<\/span>
\n actor_rollout_ref.model.use_remove_padding=<\/span>true \\\\<\/span>
\n actor_rollout_ref.actor.use_dynamic_bsz=<\/span>${use_dynamic_bsz}<\/span> \\\\<\/span>
\n actor_rollout_ref.ref.log_prob_use_dynamic_bsz=<\/span>${use_dynamic_bsz}<\/span> \\\\<\/span>
\n actor_rollout_ref.rollout.log_prob_use_dynamic_bsz=<\/span>${use_dynamic_bsz}<\/span> \\\\<\/span>
\n actor_rollout_ref.actor.ppo_max_token_len_per_gpu=<\/span>${actor_ppo_max_token_len}<\/span> \\\\<\/span>
\n actor_rollout_ref.ref.log_prob_max_token_len_per_gpu=<\/span>${infer_ppo_max_token_len}<\/span> \\\\<\/span>
\n actor_rollout_ref.rollout.log_prob_max_token_len_per_gpu=<\/span>${infer_ppo_max_token_len}<\/span> \\\\<\/span>
\n actor_rollout_ref.rollout.name=<\/span>vllm \\\\<\/span>
\n actor_rollout_ref.rollout.name=<\/span>${rollout_engine}<\/span> \\\\<\/span>
\n actor_rollout_ref.rollout.mode=<\/span>${rollout_mode}<\/span> \\\\<\/span>
\n actor_rollout_ref.model.path=<\/span>"${MODEL_PATH}<\/span>"<\/span> \\\\<\/span>
\n actor_rollout_ref.model.enable_gradient_checkpointing=<\/span>true \\\\<\/span>
\n actor_rollout_ref.actor.optim.lr=<\/span>1e-6 \\\\<\/span>
\n actor_rollout_ref.actor.optim.lr_warmup_steps_ratio=<\/span>0.05<\/span> \\\\<\/span>
\n actor_rollout_ref.actor.optim.weight_decay=<\/span>0.1<\/span> \\\\<\/span>
\n actor_rollout_ref.actor.ppo_mini_batch_size=<\/span>${ppo_mini_batch_size}<\/span> \\\\<\/span>
\n actor_rollout_ref.actor.ppo_micro_batch_size_per_gpu=<\/span>${ppo_micro_batch_size_per_gpu}<\/span> \\\\<\/span>
\n actor_rollout_ref.actor.fsdp_config.param_offload=<\/span>${offload}<\/span> \\\\<\/span>
\n actor_rollout_ref.actor.fsdp_config.optimizer_offload=<\/span>${offload}<\/span> \\\\<\/span>
\n actor_rollout_ref.actor.entropy_coeff=<\/span>0<\/span> \\\\<\/span>
\n actor_rollout_ref.actor.grad_clip=<\/span>1.0<\/span> \\\\<\/span>
\n actor_rollout_ref.actor.loss_agg_mode=<\/span>${loss_agg_mode}<\/span> \\\\<\/span>
\n actor_rollout_ref.actor.ulysses_sequence_parallel_size=<\/span>${sp_size}<\/span> \\\\<\/span>
\n actor_rollout_ref.rollout.gpu_memory_utilization=<\/span>${gpu_memory_utilization}<\/span> \\\\<\/span>
\n actor_rollout_ref.rollout.tensor_model_parallel_size=<\/span>${gen_tp}<\/span> \\\\<\/span>
\n actor_rollout_ref.rollout.enable_chunked_prefill=<\/span>true \\\\<\/span>
\n actor_rollout_ref.rollout.max_num_batched_tokens=<\/span>$((<\/span>max_prompt_length +<\/span> max_response_length))<\/span><\/span> \\\\<\/span>
\n actor_rollout_ref.rollout.temperature=<\/span>${temperature}<\/span> \\\\<\/span>
\n actor_rollout_ref.rollout.top_p=<\/span>${top_p}<\/span> \\\\<\/span>
\n actor_rollout_ref.rollout.top_k=<\/span>${top_k}<\/span> \\\\<\/span>
\n actor_rollout_ref.rollout.val_kwargs.temperature=<\/span>${temperature}<\/span> \\\\<\/span>
\n actor_rollout_ref.rollout.val_kwargs.top_p=<\/span>${val_top_p}<\/span> \\\\<\/span>
\n actor_rollout_ref.rollout.val_kwargs.top_k=<\/span>${top_k}<\/span> \\\\<\/span>
\n actor_rollout_ref.rollout.val_kwargs.do_sample=<\/span>true \\\\<\/span>
\n actor_rollout_ref.rollout.val_kwargs.n=<\/span>1<\/span> \\\\<\/span>
\n actor_rollout_ref.ref.fsdp_config.param_offload=<\/span>${offload}<\/span> \\\\<\/span>
\n actor_rollout_ref.ref.ulysses_sequence_parallel_size=<\/span>${sp_size}<\/span> \\\\<\/span>
\n actor_rollout_ref.actor.entropy_checkpointing=<\/span>${entropy_checkpointing}<\/span> \\\\<\/span>
\n reward_model.reward_manager=<\/span>${reward_manager}<\/span> \\\\<\/span>
\n +reward_model.reward_kwargs.overlong_buffer_cfg.enable=<\/span>${enable_overlong_buffer}<\/span> \\\\<\/span>
\n +reward_model.reward_kwargs.overlong_buffer_cfg.len=<\/span>${overlong_buffer_len}<\/span> \\\\<\/span>
\n +reward_model.reward_kwargs.overlong_buffer_cfg.penalty_factor=<\/span>${overlong_penalty_factor}<\/span> \\\\<\/span>
\n +reward_model.reward_kwargs.overlong_buffer_cfg.log=<\/span>false \\\\<\/span>
\n +reward_model.reward_kwargs.max_resp_len=<\/span>${max_response_length}<\/span> \\\\<\/span>
\n trainer.logger=<\/span>'["console"]'<\/span> \\\\<\/span>
\n actor_rollout_ref.rollout.enforce_eager=<\/span>True \\\\<\/span>
\n actor_rollout_ref.actor.use_torch_compile=<\/span>False \\\\<\/span>
\n trainer.project_name=<\/span>"${project_name}<\/span>"<\/span> \\\\<\/span>
\n trainer.experiment_name=<\/span>"${exp_name}<\/span>"<\/span> \\\\<\/span>
\n trainer.n_gpus_per_node=<\/span>16<\/span> \\\\<\/span>
\n trainer.nnodes=<\/span>1<\/span> \\\\<\/span>
\n trainer.val_before_train=<\/span>${val_before_train}<\/span> \\\\<\/span>
\n trainer.test_freq=<\/span>${test_freq}<\/span> \\\\<\/span>
\n trainer.save_freq=<\/span>${save_freq}<\/span> \\\\<\/span>
\n trainer.total_epochs=<\/span>${total_epochs}<\/span> \\\\<\/span>
\n trainer.total_training_steps=<\/span>${total_training_steps}<\/span> \\\\<\/span>
\n trainer.default_local_dir=<\/span>"${CKPTS_DIR}<\/span>"<\/span> \\\\<\/span>
\n trainer.resume_mode=<\/span>auto \\\\<\/span>
\n trainer.device=<\/span>npu \\\\<\/span>
\n trainer.log_val_generations=<\/span>2<\/span> \\\\<\/span>
\n $@<\/span><\/p>\n

\u5b98\u65b9\u8c03\u8bd5\u6570\u636e<\/h4>\n

\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\"<\/p>\n

\u8c03\u8bd5\u7ed3\u679c<\/h4>\n

\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\"<\/p>\n

\u7ed3\u8bba：\u66f2\u7ebf\u6b63\u5e38\u4e0a\u5347，\u4e0a\u5347\u8d8b\u52bf\u4e00\u81f4，\u7b26\u5408\u9884\u671f\u7ed3\u679c<\/p>\n

\u8c03\u8bd5\u7528\u4f8b2：Qwen3-30B-A3B<\/h3>\n

\u8c03\u8bd5\u811a\u672c<\/h4>\n

project_name=<\/span>'DAPO'<\/span>
\nexp_name=<\/span>'GSPO-Qwen3-30B-A3B-4nodes'<\/span>
\nadv_estimator=<\/span>grpo
\nuse_kl_in_reward=<\/span>False<\/span>
\nkl_coef=<\/span>0.0<\/span>
\nuse_kl_loss=<\/span>False<\/span>
\nkl_loss_coef=<\/span>0.0<\/span>
\nclip_ratio_low=<\/span>3e-4<\/span>
\nclip_ratio_high=<\/span>4e-4<\/span>
\nmax_prompt_length=<\/span>$(<\/span>(<\/span>1024<\/span> *<\/span> 2<\/span>)<\/span>)<\/span>
\nmax_response_length=<\/span>$(<\/span>(<\/span>1024<\/span> *<\/span> 8<\/span>)<\/span>)<\/span>
\nenable_overlong_buffer=<\/span>True<\/span>
\noverlong_buffer_len=<\/span>$(<\/span>(<\/span>1024<\/span> *<\/span> 4<\/span>)<\/span>)<\/span>
\noverlong_penalty_factor=<\/span>1.0<\/span>
\nloss_agg_mode=<\/span>"token-mean"<\/span>
\nloss_mode=<\/span>gspo
\ntrain_prompt_bsz=<\/span>32<\/span>
\nn_resp_per_prompt=<\/span>2<\/span>
\ntrain_prompt_mini_bsz=<\/span>4<\/span>
\nNNODES=<\/span>${<\/span>NNODES:<\/span>–<\/span>2<\/span>}<\/span>
\nNGPUS_PER_NODE=<\/span>${<\/span>NGPUS_PER_NODE:<\/span>–<\/span>8<\/span>}<\/span>
\nMODEL_PATH=<\/span>xx\/<\/span>weight\/<\/span>Qwen3_30B\/<\/span>Qwen3_30B
\nCKPTS_DIR=<\/span>$DATA_ROOT\/<\/span>checkpoint\/<\/span>${<\/span>project_name}<\/span>\/<\/span>${<\/span>exp_name}<\/span>
\nTRAIN_FILE=<\/span>xx\/<\/span>rl_data\/<\/span>dapo–<\/span>math\/<\/span>dapo–<\/span>math–<\/span>17k.<\/span>parquet
\n# aime24_test_path=\/data01\/huawei-2025\/rl_data\/aime-2024\/aime-2024.parquet<\/span>
\naime24_test_path=<\/span>xx\/<\/span>rl_data\/<\/span>dapo–<\/span>math\/<\/span>dapo–<\/span>math–<\/span>17k.<\/span>parquet
\nTEST_FILE=<\/span>"['$aime24_test_path']"<\/span>
\n# Algorithm<\/span>
\ntemperature=<\/span>1.0<\/span>
\ntop_p=<\/span>1.0<\/span>
\ntop_k=<\/span>–<\/span>1<\/span> # 0 for HF rollout, -1 for vLLM rollout<\/span>
\nval_top_p=<\/span>0.7<\/span>
\n# Performance Related Parameter<\/span>
\nuse_dynamic_bsz=<\/span>True<\/span>
\nactor_ppo_max_token_len=<\/span>$(<\/span>(<\/span>(<\/span>max_prompt_length +<\/span> max_response_length)<\/span> *<\/span> 1<\/span>)<\/span>)<\/span>
\ninfer_ppo_max_token_len=<\/span>$(<\/span>(<\/span>(<\/span>max_prompt_length +<\/span> max_response_length)<\/span> *<\/span> 1<\/span>)<\/span>)<\/span>
\noffload=<\/span>True<\/span>
\n# gen<\/span>
\nrollout_name=<\/span>vllm # vllm or sglang<\/span>
\ngen_tp=<\/span>4<\/span>
\ngen_dp=<\/span>8<\/span>
\n# gen_ep=4<\/span>
\n# train<\/span>
\ntrain_tp=<\/span>4<\/span>
\ntrain_pp=<\/span>4<\/span>
\nEP=<\/span>8<\/span>
\nETP=<\/span>1<\/span>
\nRUNTIME_ENV=<\/span>verl\/<\/span>trainer\/<\/span>mc2_env.<\/span>yaml
\ncd \/<\/span>opt\/<\/span>verl
\nray job submit –<\/span>–<\/span>runtime–<\/span>env=<\/span>"${RUNTIME_ENV}"<\/span> \\\\
\n –<\/span>–<\/span> python3 –<\/span>m verl.<\/span>trainer.<\/span>main_ppo \\\\
\n –<\/span>–<\/span>config–<\/span>path=<\/span>config \\\\
\n –<\/span>–<\/span>config–<\/span>name=<\/span>'ppo_megatron_trainer.yaml'<\/span> \\\\
\n data.<\/span>train_files=<\/span>"${TRAIN_FILE}"<\/span> \\\\
\n data.<\/span>val_files=<\/span>"${TEST_FILE}"<\/span> \\\\
\n data.<\/span>prompt_key=<\/span>prompt \\\\
\n data.<\/span>return_raw_chat=<\/span>True<\/span> \\\\
\n data.<\/span>truncation=<\/span>'left'<\/span> \\\\
\n data.<\/span>max_prompt_length=<\/span>${<\/span>max_prompt_length}<\/span> \\\\
\n data.<\/span>max_response_length=<\/span>${<\/span>max_response_length}<\/span> \\\\
\n data.<\/span>train_batch_size=<\/span>${<\/span>train_prompt_bsz}<\/span> \\\\
\n actor_rollout_ref.<\/span>rollout.<\/span>n=<\/span>${<\/span>n_resp_per_prompt}<\/span> \\\\
\n actor_rollout_ref.<\/span>actor.<\/span>policy_loss.<\/span>loss_mode=<\/span>${<\/span>loss_mode}<\/span> \\\\
\n algorithm.<\/span>adv_estimator=<\/span>${<\/span>adv_estimator}<\/span> \\\\
\n algorithm.<\/span>use_kl_in_reward=<\/span>${<\/span>use_kl_in_reward}<\/span> \\\\
\n algorithm.<\/span>kl_ctrl.<\/span>kl_coef=<\/span>${<\/span>kl_coef}<\/span> \\\\
\n actor_rollout_ref.<\/span>actor.<\/span>use_kl_loss=<\/span>${<\/span>use_kl_loss}<\/span> \\\\
\n actor_rollout_ref.<\/span>actor.<\/span>kl_loss_coef=<\/span>${<\/span>kl_loss_coef}<\/span> \\\\
\n actor_rollout_ref.<\/span>actor.<\/span>clip_ratio_low=<\/span>${<\/span>clip_ratio_low}<\/span> \\\\
\n actor_rollout_ref.<\/span>actor.<\/span>clip_ratio_high=<\/span>${<\/span>clip_ratio_high}<\/span> \\\\
\n actor_rollout_ref.<\/span>actor.<\/span>clip_ratio_c=<\/span>10.0<\/span> \\\\
\n actor_rollout_ref.<\/span>actor.<\/span>use_dynamic_bsz=<\/span>${<\/span>use_dynamic_bsz}<\/span> \\\\
\n actor_rollout_ref.<\/span>ref.<\/span>log_prob_use_dynamic_bsz=<\/span>${<\/span>use_dynamic_bsz}<\/span> \\\\
\n actor_rollout_ref.<\/span>rollout.<\/span>log_prob_use_dynamic_bsz=<\/span>${<\/span>use_dynamic_bsz}<\/span> \\\\
\n actor_rollout_ref.<\/span>actor.<\/span>ppo_max_token_len_per_gpu=<\/span>${<\/span>actor_ppo_max_token_len}<\/span> \\\\
\n actor_rollout_ref.<\/span>ref.<\/span>log_prob_max_token_len_per_gpu=<\/span>${<\/span>infer_ppo_max_token_len}<\/span> \\\\
\n actor_rollout_ref.<\/span>rollout.<\/span>log_prob_max_token_len_per_gpu=<\/span>${<\/span>infer_ppo_max_token_len}<\/span> \\\\
\n actor_rollout_ref.<\/span>model.<\/span>path=<\/span>"${MODEL_PATH}"<\/span> \\\\
\n actor_rollout_ref.<\/span>actor.<\/span>optim.<\/span>lr=<\/span>1e-6<\/span> \\\\
\n actor_rollout_ref.<\/span>actor.<\/span>optim.<\/span>lr_warmup_steps=<\/span>10<\/span> \\\\
\n actor_rollout_ref.<\/span>actor.<\/span>optim.<\/span>weight_decay=<\/span>0.1<\/span> \\\\
\n actor_rollout_ref.<\/span>actor.<\/span>ppo_mini_batch_size=<\/span>${<\/span>train_prompt_mini_bsz}<\/span> \\\\
\n actor_rollout_ref.<\/span>actor.<\/span>entropy_coeff=<\/span>0<\/span> \\\\
\n actor_rollout_ref.<\/span>actor.<\/span>optim.<\/span>clip_grad=<\/span>1.0<\/span> \\\\
\n actor_rollout_ref.<\/span>actor.<\/span>loss_agg_mode=<\/span>${<\/span>loss_agg_mode}<\/span> \\\\
\n actor_rollout_ref.<\/span>actor.<\/span>megatron.<\/span>param_offload=<\/span>${<\/span>offload}<\/span> \\\\
\n actor_rollout_ref.<\/span>actor.<\/span>megatron.<\/span>optimizer_offload=<\/span>${<\/span>offload}<\/span> \\\\
\n actor_rollout_ref.<\/span>actor.<\/span>megatron.<\/span>grad_offload=<\/span>${<\/span>offload}<\/span> \\\\
\n actor_rollout_ref.<\/span>actor.<\/span>megatron.<\/span>pipeline_model_parallel_size=<\/span>${<\/span>train_pp}<\/span> \\\\
\n actor_rollout_ref.<\/span>actor.<\/span>megatron.<\/span>tensor_model_parallel_size=<\/span>${<\/span>train_tp}<\/span> \\\\
\n actor_rollout_ref.<\/span>actor.<\/span>megatron.<\/span>expert_model_parallel_size=<\/span>$EP \\\\
\n actor_rollout_ref.<\/span>actor.<\/span>megatron.<\/span>expert_tensor_parallel_size=<\/span>$ETP \\\\
\n actor_rollout_ref.<\/span>rollout.<\/span>gpu_memory_utilization=<\/span>0.80<\/span> \\\\
\n actor_rollout_ref.<\/span>rollout.<\/span>enable_chunked_prefill=<\/span>True<\/span> \\\\
\n actor_rollout_ref.<\/span>rollout.<\/span>max_num_batched_tokens=<\/span>$(<\/span>(<\/span>max_prompt_length +<\/span> max_response_length)<\/span>)<\/span> \\\\
\n actor_rollout_ref.<\/span>rollout.<\/span>temperature=<\/span>${<\/span>temperature}<\/span> \\\\
\n actor_rollout_ref.<\/span>rollout.<\/span>top_p=<\/span>${<\/span>top_p}<\/span> \\\\
\n actor_rollout_ref.<\/span>rollout.<\/span>top_k=<\/span>${<\/span>top_k}<\/span> \\\\
\n actor_rollout_ref.<\/span>rollout.<\/span>val_kwargs.<\/span>temperature=<\/span>${<\/span>temperature}<\/span> \\\\
\n actor_rollout_ref.<\/span>rollout.<\/span>val_kwargs.<\/span>top_p=<\/span>${<\/span>val_top_p}<\/span> \\\\
\n actor_rollout_ref.<\/span>rollout.<\/span>val_kwargs.<\/span>top_k=<\/span>${<\/span>top_k}<\/span> \\\\
\n actor_rollout_ref.<\/span>rollout.<\/span>val_kwargs.<\/span>do_sample=<\/span>True<\/span> \\\\
\n actor_rollout_ref.<\/span>rollout.<\/span>val_kwargs.<\/span>n=<\/span>1<\/span> \\\\
\n actor_rollout_ref.<\/span>rollout.<\/span>name=<\/span>${<\/span>rollout_name}<\/span> \\\\
\n actor_rollout_ref.<\/span>rollout.<\/span>mode=<\/span>sync \\\\
\n actor_rollout_ref.<\/span>rollout.<\/span>calculate_log_probs=<\/span>True<\/span> \\\\
\n actor_rollout_ref.<\/span>rollout.<\/span>tensor_model_parallel_size=<\/span>${<\/span>gen_tp}<\/span> \\\\
\n actor_rollout_ref.<\/span>rollout.<\/span>data_parallel_size=<\/span>${<\/span>gen_dp}<\/span> \\\\
\n actor_rollout_ref.<\/span>ref.<\/span>megatron.<\/span>pipeline_model_parallel_size=<\/span>${<\/span>train_pp}<\/span> \\\\
\n actor_rollout_ref.<\/span>ref.<\/span>megatron.<\/span>tensor_model_parallel_size=<\/span>${<\/span>train_tp}<\/span> \\\\
\n actor_rollout_ref.<\/span>ref.<\/span>megatron.<\/span>expert_model_parallel_size=<\/span>$EP \\\\
\n actor_rollout_ref.<\/span>ref.<\/span>megatron.<\/span>expert_tensor_parallel_size=<\/span>$ETP \\\\
\n actor_rollout_ref.<\/span>ref.<\/span>megatron.<\/span>param_offload=<\/span>${<\/span>offload}<\/span> \\\\
\n actor_rollout_ref.<\/span>actor.<\/span>megatron.<\/span>use_mbridge=<\/span>True<\/span> \\\\
\n +<\/span>actor_rollout_ref.<\/span>actor.<\/span>megatron.<\/span>override_transformer_config.<\/span>moe_router_dtype=<\/span>fp32 \\\\
\n +<\/span>actor_rollout_ref.<\/span>actor.<\/span>megatron.<\/span>override_transformer_config.<\/span>recompute_method=<\/span>uniform \\\\
\n +<\/span>actor_rollout_ref.<\/span>actor.<\/span>megatron.<\/span>override_transformer_config.<\/span>recompute_granularity=<\/span>full \\\\
\n +<\/span>actor_rollout_ref.<\/span>actor.<\/span>megatron.<\/span>override_transformer_config.<\/span>recompute_num_layers=<\/span>1<\/span> \\\\
\n reward_model.<\/span>reward_manager=<\/span>dapo \\\\
\n +<\/span>reward_model.<\/span>reward_kwargs.<\/span>overlong_buffer_cfg.<\/span>enable=<\/span>${<\/span>enable_overlong_buffer}<\/span> \\\\
\n +<\/span>reward_model.<\/span>reward_kwargs.<\/span>overlong_buffer_cfg.<\/span>len<\/span>=<\/span>${<\/span>overlong_buffer_len}<\/span> \\\\
\n +<\/span>reward_model.<\/span>reward_kwargs.<\/span>overlong_buffer_cfg.<\/span>penalty_factor=<\/span>${<\/span>overlong_penalty_factor}<\/span> \\\\
\n +<\/span>reward_model.<\/span>reward_kwargs.<\/span>overlong_buffer_cfg.<\/span>log=<\/span>False<\/span> \\\\
\n +<\/span>reward_model.<\/span>reward_kwargs.<\/span>max_resp_len=<\/span>${<\/span>max_response_length}<\/span> \\\\
\n trainer.<\/span>logger=<\/span>"console"<\/span> \\\\
\n trainer.<\/span>project_name=<\/span>"${project_name}"<\/span> \\\\
\n trainer.<\/span>experiment_name=<\/span>"${exp_name}-tp${gen_tp}-ep${gen_ep}"<\/span> \\\\
\n trainer.<\/span>n_gpus_per_node=<\/span>"${NGPUS_PER_NODE}"<\/span> \\\\
\n trainer.<\/span>nnodes=<\/span>"${NNODES}"<\/span> \\\\
\n trainer.<\/span>val_before_train=<\/span>False<\/span> \\\\
\n trainer.<\/span>test_freq=<\/span>–<\/span>1<\/span> \\\\
\n trainer.<\/span>save_freq=<\/span>–<\/span>1<\/span> \\\\
\n trainer.<\/span>total_epochs=<\/span>10<\/span> \\\\
\n trainer.<\/span>total_training_steps=<\/span>300<\/span> \\\\
\n trainer.<\/span>default_local_dir=<\/span>"${CKPTS_DIR}"<\/span> \\\\
\n trainer.<\/span>resume_mode=<\/span>auto \\\\
\n trainer.<\/span>log_val_generations=<\/span>10<\/span> \\\\
\n +<\/span>actor_rollout_ref.<\/span>actor.<\/span>megatron.<\/span>override_transformer_config.<\/span>use_flash_attn=<\/span>True<\/span> \\\\
\n trainer.<\/span>device=<\/span>"npu"<\/span> $@<\/p>\n

GRPO vs GSPO \u8c03\u8bd5\u7ed3\u679c<\/h4>\n

\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\"<\/p>\n

\u200b\u7ed3\u8bba\u200b：GSPO\u5728MOE\u6a21\u578b\u4e0a\u5c55\u793a\u51fa\u4e86\u66f4\u597d\u7684\u8bad\u7ec3\u6548\u679c<\/p>\n

\u6700\u65b0\u7248\u672c\u8c03\u8bd5<\/h3>\n

\u57fa\u4e8e\u6700\u65b0\u7248\u672cverl\u8c03\u8bd5\u670d\u52a1\u5316vllm\u540e\u7aef\u548cengineworker\u529f\u80fd<\/p>\n

\u73af\u5883<\/h4>\n\n\u914d\u7f6e\u9879\u7248\u672c\u4fe1\u606f\u5907\u6ce8<\/tr>\n\n\n\n\n\n\n\n\n
CANN<\/td>\n8.3.RC1<\/td>\n<\/td>\n<\/tr>\n
torch<\/td>\n2.7.1<\/td>\n<\/td>\n<\/tr>\n
torch_npu<\/td>\n2.7.1<\/td>\n<\/td>\n<\/tr>\n
transformer<\/td>\n4.57.3<\/td>\n<\/td>\n<\/tr>\n
vllm<\/td>\nv0.11.0<\/td>\n<\/td>\n<\/tr>\n
vllm-ascend<\/td>\nv0.11.0rc1<\/td>\n<\/td>\n<\/tr>\n
verl<\/td>\ncommit:5a2e0b1c272b33<\/td>\n10\u670810\u65e5\u4ee3\u7801<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

\u8c03\u8bd5\u7528\u4f8b3：\u670d\u52a1\u5316vllm\u540e\u7aefGSPO\u529f\u80fd\u8c03\u8bd5<\/h4>\n

\u8c03\u8bd5\u811a\u672c<\/h5>\n

#!\/usr\/bin\/env bash<\/span>
\nset<\/span> -x<\/p>\n

pkill<\/span> -9 python
\nps<\/span> -ef |<\/span> grep<\/span> "python"<\/span> |<\/span>grep<\/span> -v grep<\/span> |<\/span>awk<\/span> '{print $2}'<\/span> |<\/span>xargs<\/span> -t -i kill<\/span> -9 {<\/span>}<\/span>
\nray stop –force
\npkill<\/span> -9 python
\npkill<\/span> -9 torchrun
\nps<\/span> -ef |<\/span> grep<\/span> "defaunct"<\/span> |<\/span>grep<\/span> python |<\/span>awk<\/span> '{print $3}'<\/span> |<\/span>xargs<\/span> -t -i kill<\/span> -9 {<\/span>}<\/span>
\nps<\/span> -ef |<\/span> grep<\/span> "defaunct"<\/span> |<\/span>grep<\/span> torchrun |<\/span>awk<\/span> '{print $3}'<\/span> |<\/span>xargs<\/span> -t -i kill<\/span> -9 {<\/span>}<\/span>
\nps<\/span> -ef |<\/span> grep<\/span> -i python |<\/span>grep<\/span> -i [<\/span>name]<\/span> |<\/span>grep<\/span> -v grep<\/span> |<\/span>awk<\/span> '{print $2}'<\/span> |<\/span>xargs<\/span> -t -I {<\/span>}<\/span> kill<\/span> -9 {<\/span>}<\/span>
\nps<\/span> -ef |<\/span> grep<\/span> -i torchrun |<\/span>grep<\/span> -i [<\/span>name]<\/span> |<\/span>grep<\/span> -v grep<\/span> |<\/span>awk<\/span> '{print $2}'<\/span> |<\/span>xargs<\/span> -t -I {<\/span>}<\/span> kill<\/span> -9 {<\/span>}<\/span>
\nps<\/span> -ef |<\/span> grep<\/span> "python"<\/span> |<\/span>grep<\/span> -v grep<\/span> |<\/span>awk<\/span> '{print $2}'<\/span> |<\/span>xargs<\/span> -t -i kill<\/span> -9 {<\/span>}<\/span>
\n# Set how many GPUs we actually have on this node.<\/span>
\nexport<\/span> GPUS_PER_NODE<\/span>=<\/span>8<\/span><\/p>\n

export<\/span> NNODES<\/span>=<\/span>1<\/span>
\nexport<\/span> VLLM_ASCEND_ENABLE_NZ<\/span>=<\/span>0<\/span><\/p>\n

echo<\/span> "Using $NNODES<\/span> nodes for training…"<\/span><\/p>\n

#export ASCEND_LAUNCH_BLOCKING=1 <\/span>
\n# ————————————- Setup xp params —————————————<\/span>
\nproject_name<\/span>=<\/span>'RL-GSPO'<\/span><\/p>\n

adv_estimator<\/span>=<\/span>grpo
\nloss_mode<\/span>=<\/span>gspo
\nloss_agg_mode<\/span>=<\/span>"seq-mean-token-mean"<\/span>
\nMODEL_PATH<\/span>=<\/span>xx\/weights\/Qwen2.5-3B-Instruct
\noffload<\/span>=<\/span>false # it's a small model, offloading will just slow-down training<\/span>
\nrollout_engine<\/span>=<\/span>vllm
\nrollout_mode<\/span>=<\/span>async
\nreturn_raw_chat<\/span>=<\/span>"True"<\/span>
\nif<\/span> [<\/span> "$rollout_engine<\/span>"<\/span> =<\/span> "vllm"<\/span> ]<\/span>;<\/span> then<\/span>
\n export<\/span> VLLM_USE_V1<\/span>=<\/span>1<\/span>
\nfi<\/span>
\ngpu_memory_utilization<\/span>=<\/span>0.6<\/span>
\nreward_manager<\/span>=<\/span>dapo
\nadv_estimator<\/span>=<\/span>grpo
\nshuffle_dataset<\/span>=<\/span>true
\nfirst_time_dataset_prep<\/span>=<\/span>true # prepare dataset<\/span><\/p>\n

test_freq<\/span>=<\/span>10<\/span>
\nsave_freq<\/span>=<\/span>10<\/span>
\ntotal_epochs<\/span>=<\/span>10<\/span>
\ntotal_training_steps<\/span>=<\/span>500<\/span>
\nval_before_train<\/span>=<\/span>false<\/p>\n

use_kl_in_reward<\/span>=<\/span>false
\nkl_coef<\/span>=<\/span>0.0<\/span>
\nuse_kl_loss<\/span>=<\/span>false
\nkl_loss_coef<\/span>=<\/span>0.0<\/span><\/p>\n

clip_ratio_low<\/span>=<\/span>0.0003<\/span> # as recommended by the paper, see Sec. 5.1<\/span>
\nclip_ratio_high<\/span>=<\/span>0.0004<\/span> # as recommended by the paper, see Sec. 5.1<\/span>
\ntrain_batch_size<\/span>=<\/span>512<\/span>
\nppo_mini_batch_size<\/span>=<\/span>128<\/span> # maintain 4 mini-batches as recommended by the paper, see Sec. 5.1<\/span>
\nppo_micro_batch_size_per_gpu<\/span>=<\/span>8<\/span> # setup depending on your GPU memory<\/span>
\nn_resp_per_prompt<\/span>=<\/span>16<\/span><\/p>\n

max_prompt_length<\/span>=<\/span>$((<\/span>1024<\/span> *<\/span> 2<\/span>))<\/span><\/span>
\nmax_response_length<\/span>=<\/span>$((<\/span>1024<\/span> *<\/span> 8<\/span>))<\/span><\/span>
\n# dapo reward manager params<\/span>
\nenable_overlong_buffer<\/span>=<\/span>false # true<\/span>
\noverlong_buffer_len<\/span>=<\/span>$((<\/span>1024<\/span> *<\/span> 4<\/span>))<\/span><\/span>
\noverlong_penalty_factor<\/span>=<\/span>1.0<\/span><\/p>\n

# Paths and namings<\/span>
\nSFT_MODEL<\/span>=<\/span>$(<\/span>basename<\/span> $MODEL_PATH)<\/span><\/span>
\nexp_name<\/span>=<\/span>"${loss_mode}<\/span>-epslow-${clip_ratio_low}<\/span>-epshigh-${clip_ratio_high}<\/span>–${SFT_MODEL}<\/span>-RL"<\/span>
\nCKPTS_DIR<\/span>=<\/span>\/rl\/checkpoints\/experimental\/4b\/${loss_mode}<\/span>\/${exp_name}<\/span><\/p>\n

# Sampling params at rollouts<\/span>
\ntemperature<\/span>=<\/span>1.0<\/span>
\ntop_p<\/span>=<\/span>1.0<\/span>
\ntop_k<\/span>=<\/span>-1 # 0 for HF rollout, -1 for vLLM rollout<\/span>
\nval_top_p<\/span>=<\/span>0.7<\/span><\/p>\n

# Performance Related Parameter<\/span>
\nsp_size<\/span>=<\/span>4<\/span>
\nuse_dynamic_bsz<\/span>=<\/span>true
\nactor_ppo_max_token_len<\/span>=<\/span>$((<\/span>(<\/span>max_prompt_length +<\/span> max_response_length)<\/span> *<\/span> 1<\/span>))<\/span><\/span>
\ninfer_ppo_max_token_len<\/span>=<\/span>$((<\/span>(<\/span>max_prompt_length +<\/span> max_response_length)<\/span> *<\/span> 1<\/span>))<\/span><\/span>
\noffload<\/span>=<\/span>true
\ngen_tp<\/span>=<\/span>2<\/span>
\nentropy_checkpointing<\/span>=<\/span>true # This enables entropy recomputation specifically for the entropy calculation, lowering memory usage during training.<\/span><\/p>\n

# ————————————- train\/val data preparation —————————————<\/span>
\n# if [ "$first_time_dataset_prep" = true ]; then<\/span>
\n# echo "Preprocessing GSM8K dataset…"<\/span>
\n# python examples\/data_preprocess\/gsm8k.py –local_save_dir \/data01\/huawei-2025\/rl_data\/gsm8k\/data_later –local_dataset_path \/data01\/huawei-2025\/rl_data\/gsm8k\/<\/span>
\n# fi<\/span><\/p>\n

gsm8k_train_path<\/span>=<\/span>xx\/data\/post_gsm8k\/train.parquet
\ngsm8k_test_path<\/span>=<\/span>xx\/data\/post_gsm8k\/test.parquet<\/p>\n

# set the paths<\/span>
\ntrain_files<\/span>=<\/span>"['$gsm8k_train_path<\/span>']"<\/span>
\ntest_files<\/span>=<\/span>"['$gsm8k_test_path<\/span>']"<\/span>
\n#! \u4fee\u6539filter_overlong_prompts false<\/span>
\npython3 -m verl.trainer.main_ppo \\\\<\/span>
\n algorithm.adv_estimator=<\/span>${adv_estimator}<\/span> \\\\<\/span>
\n actor_rollout_ref.actor.policy_loss.loss_mode=<\/span>${loss_mode}<\/span> \\\\<\/span>
\n data.train_files=<\/span>"${train_files}<\/span>"<\/span> \\\\<\/span>
\n data.val_files=<\/span>"${test_files}<\/span>"<\/span> \\\\<\/span>
\n data.shuffle=<\/span>$shuffle_dataset<\/span> \\\\<\/span>
\n data.prompt_key=<\/span>prompt \\\\<\/span>
\n data.truncation=<\/span>'error'<\/span> \\\\<\/span>
\n data.filter_overlong_prompts=<\/span>true \\\\<\/span>
\n data.return_raw_chat=<\/span>${return_raw_chat}<\/span> \\\\<\/span>
\n data.train_batch_size=<\/span>${train_batch_size}<\/span> \\\\<\/span>
\n data.max_prompt_length=<\/span>${max_prompt_length}<\/span> \\\\<\/span>
\n data.max_response_length=<\/span>${max_response_length}<\/span> \\\\<\/span>
\n actor_rollout_ref.rollout.n=<\/span>${n_resp_per_prompt}<\/span> \\\\<\/span>
\n algorithm.use_kl_in_reward=<\/span>${use_kl_in_reward}<\/span> \\\\<\/span>
\n algorithm.kl_ctrl.kl_coef=<\/span>${kl_coef}<\/span> \\\\<\/span>
\n actor_rollout_ref.actor.use_kl_loss=<\/span>${use_kl_loss}<\/span> \\\\<\/span>
\n actor_rollout_ref.actor.kl_loss_coef=<\/span>${kl_loss_coef}<\/span> \\\\<\/span>
\n actor_rollout_ref.actor.clip_ratio_low=<\/span>${clip_ratio_low}<\/span> \\\\<\/span>
\n actor_rollout_ref.actor.clip_ratio_high=<\/span>${clip_ratio_high}<\/span> \\\\<\/span>
\n actor_rollout_ref.model.use_remove_padding=<\/span>true \\\\<\/span>
\n actor_rollout_ref.actor.use_dynamic_bsz=<\/span>${use_dynamic_bsz}<\/span> \\\\<\/span>
\n actor_rollout_ref.ref.log_prob_use_dynamic_bsz=<\/span>${use_dynamic_bsz}<\/span> \\\\<\/span>
\n actor_rollout_ref.rollout.log_prob_use_dynamic_bsz=<\/span>${use_dynamic_bsz}<\/span> \\\\<\/span>
\n actor_rollout_ref.actor.ppo_max_token_len_per_gpu=<\/span>${actor_ppo_max_token_len}<\/span> \\\\<\/span>
\n actor_rollout_ref.ref.log_prob_max_token_len_per_gpu=<\/span>${infer_ppo_max_token_len}<\/span> \\\\<\/span>
\n actor_rollout_ref.rollout.log_prob_max_token_len_per_gpu=<\/span>${infer_ppo_max_token_len}<\/span> \\\\<\/span>
\n actor_rollout_ref.rollout.name=<\/span>${rollout_engine}<\/span> \\\\<\/span>
\n actor_rollout_ref.rollout.mode=<\/span>${rollout_mode}<\/span> \\\\<\/span>
\n actor_rollout_ref.model.path=<\/span>"${MODEL_PATH}<\/span>"<\/span> \\\\<\/span>
\n actor_rollout_ref.model.enable_gradient_checkpointing=<\/span>true \\\\<\/span>
\n actor_rollout_ref.actor.optim.lr=<\/span>1e-6 \\\\<\/span>
\n actor_rollout_ref.actor.optim.lr_warmup_steps_ratio=<\/span>0.05<\/span> \\\\<\/span>
\n actor_rollout_ref.actor.optim.weight_decay=<\/span>0.1<\/span> \\\\<\/span>
\n actor_rollout_ref.actor.ppo_mini_batch_size=<\/span>${ppo_mini_batch_size}<\/span> \\\\<\/span>
\n actor_rollout_ref.actor.ppo_micro_batch_size_per_gpu=<\/span>${ppo_micro_batch_size_per_gpu}<\/span> \\\\<\/span>
\n actor_rollout_ref.actor.fsdp_config.param_offload=<\/span>${offload}<\/span> \\\\<\/span>
\n actor_rollout_ref.actor.fsdp_config.optimizer_offload=<\/span>${offload}<\/span> \\\\<\/span>
\n actor_rollout_ref.actor.entropy_coeff=<\/span>0<\/span> \\\\<\/span>
\n actor_rollout_ref.actor.grad_clip=<\/span>1.0<\/span> \\\\<\/span>
\n actor_rollout_ref.actor.loss_agg_mode=<\/span>${loss_agg_mode}<\/span> \\\\<\/span>
\n actor_rollout_ref.actor.ulysses_sequence_parallel_size=<\/span>${sp_size}<\/span> \\\\<\/span>
\n actor_rollout_ref.rollout.gpu_memory_utilization=<\/span>${gpu_memory_utilization}<\/span> \\\\<\/span>
\n actor_rollout_ref.rollout.tensor_model_parallel_size=<\/span>${gen_tp}<\/span> \\\\<\/span>
\n actor_rollout_ref.rollout.enable_chunked_prefill=<\/span>true \\\\<\/span>
\n actor_rollout_ref.rollout.max_num_batched_tokens=<\/span>$((<\/span>max_prompt_length +<\/span> max_response_length))<\/span><\/span> \\\\<\/span>
\n actor_rollout_ref.rollout.temperature=<\/span>${temperature}<\/span> \\\\<\/span>
\n actor_rollout_ref.rollout.top_p=<\/span>${top_p}<\/span> \\\\<\/span>
\n actor_rollout_ref.rollout.top_k=<\/span>${top_k}<\/span> \\\\<\/span>
\n actor_rollout_ref.rollout.val_kwargs.temperature=<\/span>${temperature}<\/span> \\\\<\/span>
\n actor_rollout_ref.rollout.val_kwargs.top_p=<\/span>${val_top_p}<\/span> \\\\<\/span>
\n actor_rollout_ref.rollout.val_kwargs.top_k=<\/span>${top_k}<\/span> \\\\<\/span>
\n actor_rollout_ref.rollout.val_kwargs.do_sample=<\/span>true \\\\<\/span>
\n actor_rollout_ref.rollout.val_kwargs.n=<\/span>1<\/span> \\\\<\/span>
\n actor_rollout_ref.ref.fsdp_config.param_offload=<\/span>${offload}<\/span> \\\\<\/span>
\n actor_rollout_ref.ref.ulysses_sequence_parallel_size=<\/span>${sp_size}<\/span> \\\\<\/span>
\n actor_rollout_ref.actor.entropy_checkpointing=<\/span>${entropy_checkpointing}<\/span> \\\\<\/span>
\n reward_model.reward_manager=<\/span>${reward_manager}<\/span> \\\\<\/span>
\n +reward_model.reward_kwargs.overlong_buffer_cfg.enable=<\/span>${enable_overlong_buffer}<\/span> \\\\<\/span>
\n +reward_model.reward_kwargs.overlong_buffer_cfg.len=<\/span>${overlong_buffer_len}<\/span> \\\\<\/span>
\n +reward_model.reward_kwargs.overlong_buffer_cfg.penalty_factor=<\/span>${overlong_penalty_factor}<\/span> \\\\<\/span>
\n +reward_model.reward_kwargs.overlong_buffer_cfg.log=<\/span>false \\\\<\/span>
\n +reward_model.reward_kwargs.max_resp_len=<\/span>${max_response_length}<\/span> \\\\<\/span>
\n trainer.logger=<\/span>'["console"]'<\/span> \\\\<\/span>
\n actor_rollout_ref.rollout.enforce_eager=<\/span>True \\\\<\/span>
\n actor_rollout_ref.actor.use_torch_compile=<\/span>False \\\\<\/span>
\n trainer.project_name=<\/span>"${project_name}<\/span>"<\/span> \\\\<\/span>
\n trainer.experiment_name=<\/span>"${exp_name}<\/span>"<\/span> \\\\<\/span>
\n trainer.n_gpus_per_node=<\/span>8<\/span> \\\\<\/span>
\n trainer.nnodes=<\/span>1<\/span> \\\\<\/span>
\n trainer.val_before_train=<\/span>${val_before_train}<\/span> \\\\<\/span>
\n trainer.test_freq=<\/span>${test_freq}<\/span> \\\\<\/span>
\n trainer.save_freq=<\/span>-1 \\\\<\/span>
\n trainer.total_epochs=<\/span>${total_epochs}<\/span> \\\\<\/span>
\n trainer.total_training_steps=<\/span>${total_training_steps}<\/span> \\\\<\/span>
\n trainer.default_local_dir=<\/span>"${CKPTS_DIR}<\/span>"<\/span> \\\\<\/span>
\n trainer.resume_mode=<\/span>auto \\\\<\/span>
\n trainer.device=<\/span>npu \\\\<\/span>
\n trainer.log_val_generations=<\/span>2<\/span> \\\\<\/span>
\n $@<\/span> <\/p>\n

\u8c03\u8bd5\u7ed3\u679c<\/h5>\n

\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\"<\/p>\n

\u7ed3\u8bba：\u66f2\u7ebf\u6b63\u5e38\u4e0a\u5347，\u4e0a\u5347\u8d8b\u52bf\u4e0e\u5b98\u65b9\u811a\u672c\u4e00\u81f4，\u8c03\u8bd5\u7ed3\u679c\u7b26\u5408\u8981\u6c42<\/p>\n

\u8c03\u8bd5\u7528\u4f8b4：engineworker\u540e\u7aef\u8c03\u8bd5<\/h4>\n

engineworker\u4f7f\u80fd\u65b9\u6cd5：<\/h5>\n

trainer.<\/span>use_legacy_worker_impl=<\/span>disable \\\\<\/p>\n

\u8c03\u8bd5\u811a\u672c<\/h5>\n

#!\/usr\/bin\/env bash<\/span>
\nset<\/span> -x<\/p>\n

pkill<\/span> -9 python
\nps<\/span> -ef |<\/span> grep<\/span> "python"<\/span> |<\/span>grep<\/span> -v grep<\/span> |<\/span>awk<\/span> '{print $2}'<\/span> |<\/span>xargs<\/span> -t -i kill<\/span> -9 {<\/span>}<\/span>
\nray stop –force
\npkill<\/span> -9 python
\npkill<\/span> -9 torchrun
\nps<\/span> -ef |<\/span> grep<\/span> "defaunct"<\/span> |<\/span>grep<\/span> python |<\/span>awk<\/span> '{print $3}'<\/span> |<\/span>xargs<\/span> -t -i kill<\/span> -9 {<\/span>}<\/span>
\nps<\/span> -ef |<\/span> grep<\/span> "defaunct"<\/span> |<\/span>grep<\/span> torchrun |<\/span>awk<\/span> '{print $3}'<\/span> |<\/span>xargs<\/span> -t -i kill<\/span> -9 {<\/span>}<\/span>
\nps<\/span> -ef |<\/span> grep<\/span> -i python |<\/span>grep<\/span> -i [<\/span>name]<\/span> |<\/span>grep<\/span> -v grep<\/span> |<\/span>awk<\/span> '{print $2}'<\/span> |<\/span>xargs<\/span> -t -I {<\/span>}<\/span> kill<\/span> -9 {<\/span>}<\/span>
\nps<\/span> -ef |<\/span> grep<\/span> -i torchrun |<\/span>grep<\/span> -i [<\/span>name]<\/span> |<\/span>grep<\/span> -v grep<\/span> |<\/span>awk<\/span> '{print $2}'<\/span> |<\/span>xargs<\/span> -t -I {<\/span>}<\/span> kill<\/span> -9 {<\/span>}<\/span>
\nps<\/span> -ef |<\/span> grep<\/span> "python"<\/span> |<\/span>grep<\/span> -v grep<\/span> |<\/span>awk<\/span> '{print $2}'<\/span> |<\/span>xargs<\/span> -t -i kill<\/span> -9 {<\/span>}<\/span>
\n# Set how many GPUs we actually have on this node.<\/span>
\nexport<\/span> GPUS_PER_NODE<\/span>=<\/span>8<\/span><\/p>\n

export<\/span> NNODES<\/span>=<\/span>1<\/span>
\nexport<\/span> VLLM_ASCEND_ENABLE_NZ<\/span>=<\/span>0<\/span><\/p>\n

echo<\/span> "Using $NNODES<\/span> nodes for training…"<\/span><\/p>\n

#export ASCEND_LAUNCH_BLOCKING=1 <\/span>
\n# ————————————- Setup xp params —————————————<\/span>
\nproject_name<\/span>=<\/span>'RL-GSPO'<\/span><\/p>\n

adv_estimator<\/span>=<\/span>grpo
\nloss_mode<\/span>=<\/span>gspo
\nloss_agg_mode<\/span>=<\/span>"seq-mean-token-mean"<\/span>
\nMODEL_PATH<\/span>=<\/span>xx\/weights\/Qwen2.5-3B-Instruct
\noffload<\/span>=<\/span>false # it's a small model, offloading will just slow-down training<\/span>
\nrollout_engine<\/span>=<\/span>vllm
\nrollout_mode<\/span>=<\/span>async
\nreturn_raw_chat<\/span>=<\/span>"True"<\/span>
\nif<\/span> [<\/span> "$rollout_engine<\/span>"<\/span> =<\/span> "vllm"<\/span> ]<\/span>;<\/span> then<\/span>
\n export<\/span> VLLM_USE_V1<\/span>=<\/span>1<\/span>
\nfi<\/span>
\ngpu_memory_utilization<\/span>=<\/span>0.6<\/span>
\nreward_manager<\/span>=<\/span>dapo
\nadv_estimator<\/span>=<\/span>grpo
\nshuffle_dataset<\/span>=<\/span>true
\nfirst_time_dataset_prep<\/span>=<\/span>true # prepare dataset<\/span><\/p>\n

test_freq<\/span>=<\/span>10<\/span>
\nsave_freq<\/span>=<\/span>10<\/span>
\ntotal_epochs<\/span>=<\/span>10<\/span>
\ntotal_training_steps<\/span>=<\/span>500<\/span>
\nval_before_train<\/span>=<\/span>false<\/p>\n

use_kl_in_reward<\/span>=<\/span>false
\nkl_coef<\/span>=<\/span>0.0<\/span>
\nuse_kl_loss<\/span>=<\/span>false
\nkl_loss_coef<\/span>=<\/span>0.0<\/span><\/p>\n

clip_ratio_low<\/span>=<\/span>0.0003<\/span> # as recommended by the paper, see Sec. 5.1<\/span>
\nclip_ratio_high<\/span>=<\/span>0.0004<\/span> # as recommended by the paper, see Sec. 5.1<\/span>
\ntrain_batch_size<\/span>=<\/span>512<\/span>
\nppo_mini_batch_size<\/span>=<\/span>128<\/span> # maintain 4 mini-batches as recommended by the paper, see Sec. 5.1<\/span>
\nppo_micro_batch_size_per_gpu<\/span>=<\/span>8<\/span> # setup depending on your GPU memory<\/span>
\nn_resp_per_prompt<\/span>=<\/span>16<\/span><\/p>\n

max_prompt_length<\/span>=<\/span>$((<\/span>1024<\/span> *<\/span> 2<\/span>))<\/span><\/span>
\nmax_response_length<\/span>=<\/span>$((<\/span>1024<\/span> *<\/span> 8<\/span>))<\/span><\/span>
\n# dapo reward manager params<\/span>
\nenable_overlong_buffer<\/span>=<\/span>false # true<\/span>
\noverlong_buffer_len<\/span>=<\/span>$((<\/span>1024<\/span> *<\/span> 4<\/span>))<\/span><\/span>
\noverlong_penalty_factor<\/span>=<\/span>1.0<\/span><\/p>\n

# Paths and namings<\/span>
\nSFT_MODEL<\/span>=<\/span>$(<\/span>basename<\/span> $MODEL_PATH)<\/span><\/span>
\nexp_name<\/span>=<\/span>"${loss_mode}<\/span>-epslow-${clip_ratio_low}<\/span>-epshigh-${clip_ratio_high}<\/span>–${SFT_MODEL}<\/span>-RL"<\/span>
\nCKPTS_DIR<\/span>=<\/span>\/rl\/checkpoints\/experimental\/4b\/${loss_mode}<\/span>\/${exp_name}<\/span><\/p>\n

# Sampling params at rollouts<\/span>
\ntemperature<\/span>=<\/span>1.0<\/span>
\ntop_p<\/span>=<\/span>1.0<\/span>
\ntop_k<\/span>=<\/span>-1 # 0 for HF rollout, -1 for vLLM rollout<\/span>
\nval_top_p<\/span>=<\/span>0.7<\/span><\/p>\n

# Performance Related Parameter<\/span>
\nsp_size<\/span>=<\/span>4<\/span>
\nuse_dynamic_bsz<\/span>=<\/span>true
\nactor_ppo_max_token_len<\/span>=<\/span>$((<\/span>(<\/span>max_prompt_length +<\/span> max_response_length)<\/span> *<\/span> 1<\/span>))<\/span><\/span>
\ninfer_ppo_max_token_len<\/span>=<\/span>$((<\/span>(<\/span>max_prompt_length +<\/span> max_response_length)<\/span> *<\/span> 1<\/span>))<\/span><\/span>
\noffload<\/span>=<\/span>true
\ngen_tp<\/span>=<\/span>2<\/span>
\nentropy_checkpointing<\/span>=<\/span>true # This enables entropy recomputation specifically for the entropy calculation, lowering memory usage during training.<\/span><\/p>\n

# ————————————- train\/val data preparation —————————————<\/span>
\n# if [ "$first_time_dataset_prep" = true ]; then<\/span>
\n# echo "Preprocessing GSM8K dataset…"<\/span>
\n# python examples\/data_preprocess\/gsm8k.py –local_save_dir \/xx\/rl_data\/gsm8k\/data_later –local_dataset_path xx\/rl_data\/gsm8k\/<\/span>
\n# fi<\/span><\/p>\n

gsm8k_train_path<\/span>=<\/span>xx\/data\/post_gsm8k\/train.parquet
\ngsm8k_test_path<\/span>=<\/span>xx\/data\/post_gsm8k\/test.parquet<\/p>\n

# set the paths<\/span>
\ntrain_files<\/span>=<\/span>"['$gsm8k_train_path<\/span>']"<\/span>
\ntest_files<\/span>=<\/span>"['$gsm8k_test_path<\/span>']"<\/span>
\n#! \u4fee\u6539filter_overlong_prompts false<\/span>
\npython3 -m verl.trainer.main_ppo \\\\<\/span>
\n algorithm.adv_estimator=<\/span>${adv_estimator}<\/span> \\\\<\/span>
\n actor_rollout_ref.actor.policy_loss.loss_mode=<\/span>${loss_mode}<\/span> \\\\<\/span>
\n data.train_files=<\/span>"${train_files}<\/span>"<\/span> \\\\<\/span>
\n data.val_files=<\/span>"${test_files}<\/span>"<\/span> \\\\<\/span>
\n data.shuffle=<\/span>$shuffle_dataset<\/span> \\\\<\/span>
\n data.prompt_key=<\/span>prompt \\\\<\/span>
\n data.truncation=<\/span>'error'<\/span> \\\\<\/span>
\n data.filter_overlong_prompts=<\/span>false \\\\<\/span>
\n data.return_raw_chat=<\/span>${return_raw_chat}<\/span> \\\\<\/span>
\n data.train_batch_size=<\/span>${train_batch_size}<\/span> \\\\<\/span>
\n data.max_prompt_length=<\/span>${max_prompt_length}<\/span> \\\\<\/span>
\n data.max_response_length=<\/span>${max_response_length}<\/span> \\\\<\/span>
\n actor_rollout_ref.rollout.n=<\/span>${n_resp_per_prompt}<\/span> \\\\<\/span>
\n algorithm.use_kl_in_reward=<\/span>${use_kl_in_reward}<\/span> \\\\<\/span>
\n algorithm.kl_ctrl.kl_coef=<\/span>${kl_coef}<\/span> \\\\<\/span>
\n actor_rollout_ref.actor.use_kl_loss=<\/span>${use_kl_loss}<\/span> \\\\<\/span>
\n actor_rollout_ref.actor.kl_loss_coef=<\/span>${kl_loss_coef}<\/span> \\\\<\/span>
\n actor_rollout_ref.actor.clip_ratio_low=<\/span>${clip_ratio_low}<\/span> \\\\<\/span>
\n actor_rollout_ref.actor.clip_ratio_high=<\/span>${clip_ratio_high}<\/span> \\\\<\/span>
\n actor_rollout_ref.model.use_remove_padding=<\/span>true \\\\<\/span>
\n actor_rollout_ref.actor.use_dynamic_bsz=<\/span>${use_dynamic_bsz}<\/span> \\\\<\/span>
\n actor_rollout_ref.ref.log_prob_use_dynamic_bsz=<\/span>${use_dynamic_bsz}<\/span> \\\\<\/span>
\n actor_rollout_ref.rollout.log_prob_use_dynamic_bsz=<\/span>${use_dynamic_bsz}<\/span> \\\\<\/span>
\n actor_rollout_ref.actor.ppo_max_token_len_per_gpu=<\/span>${actor_ppo_max_token_len}<\/span> \\\\<\/span>
\n actor_rollout_ref.ref.log_prob_max_token_len_per_gpu=<\/span>${infer_ppo_max_token_len}<\/span> \\\\<\/span>
\n actor_rollout_ref.rollout.log_prob_max_token_len_per_gpu=<\/span>${infer_ppo_max_token_len}<\/span> \\\\<\/span>
\n actor_rollout_ref.rollout.name=<\/span>${rollout_engine}<\/span> \\\\<\/span>
\n actor_rollout_ref.rollout.mode=<\/span>${rollout_mode}<\/span> \\\\<\/span>
\n actor_rollout_ref.model.path=<\/span>"${MODEL_PATH}<\/span>"<\/span> \\\\<\/span>
\n actor_rollout_ref.model.enable_gradient_checkpointing=<\/span>true \\\\<\/span>
\n actor_rollout_ref.actor.optim.lr=<\/span>1e-6 \\\\<\/span>
\n actor_rollout_ref.actor.optim.lr_warmup_steps_ratio=<\/span>0.05<\/span> \\\\<\/span>
\n actor_rollout_ref.actor.optim.weight_decay=<\/span>0.1<\/span> \\\\<\/span>
\n actor_rollout_ref.actor.ppo_mini_batch_size=<\/span>${ppo_mini_batch_size}<\/span> \\\\<\/span>
\n actor_rollout_ref.actor.ppo_micro_batch_size_per_gpu=<\/span>${ppo_micro_batch_size_per_gpu}<\/span> \\\\<\/span>
\n actor_rollout_ref.actor.fsdp_config.param_offload=<\/span>${offload}<\/span> \\\\<\/span>
\n actor_rollout_ref.actor.fsdp_config.optimizer_offload=<\/span>${offload}<\/span> \\\\<\/span>
\n actor_rollout_ref.actor.entropy_coeff=<\/span>0<\/span> \\\\<\/span>
\n actor_rollout_ref.actor.grad_clip=<\/span>1.0<\/span> \\\\<\/span>
\n actor_rollout_ref.actor.loss_agg_mode=<\/span>${loss_agg_mode}<\/span> \\\\<\/span>
\n actor_rollout_ref.actor.ulysses_sequence_parallel_size=<\/span>${sp_size}<\/span> \\\\<\/span>
\n actor_rollout_ref.rollout.gpu_memory_utilization=<\/span>${gpu_memory_utilization}<\/span> \\\\<\/span>
\n actor_rollout_ref.rollout.tensor_model_parallel_size=<\/span>${gen_tp}<\/span> \\\\<\/span>
\n actor_rollout_ref.rollout.enable_chunked_prefill=<\/span>true \\\\<\/span>
\n actor_rollout_ref.rollout.max_num_batched_tokens=<\/span>$((<\/span>max_prompt_length +<\/span> max_response_length))<\/span><\/span> \\\\<\/span>
\n actor_rollout_ref.rollout.temperature=<\/span>${temperature}<\/span> \\\\<\/span>
\n actor_rollout_ref.rollout.top_p=<\/span>${top_p}<\/span> \\\\<\/span>
\n actor_rollout_ref.rollout.top_k=<\/span>${top_k}<\/span> \\\\<\/span>
\n actor_rollout_ref.rollout.val_kwargs.temperature=<\/span>${temperature}<\/span> \\\\<\/span>
\n actor_rollout_ref.rollout.val_kwargs.top_p=<\/span>${val_top_p}<\/span> \\\\<\/span>
\n actor_rollout_ref.rollout.val_kwargs.top_k=<\/span>${top_k}<\/span> \\\\<\/span>
\n actor_rollout_ref.rollout.val_kwargs.do_sample=<\/span>true \\\\<\/span>
\n actor_rollout_ref.rollout.val_kwargs.n=<\/span>1<\/span> \\\\<\/span>
\n actor_rollout_ref.ref.fsdp_config.param_offload=<\/span>${offload}<\/span> \\\\<\/span>
\n actor_rollout_ref.ref.ulysses_sequence_parallel_size=<\/span>${sp_size}<\/span> \\\\<\/span>
\n actor_rollout_ref.actor.entropy_checkpointing=<\/span>${entropy_checkpointing}<\/span> \\\\<\/span>
\n reward_model.reward_manager=<\/span>${reward_manager}<\/span> \\\\<\/span>
\n +reward_model.reward_kwargs.overlong_buffer_cfg.enable=<\/span>${enable_overlong_buffer}<\/span> \\\\<\/span>
\n +reward_model.reward_kwargs.overlong_buffer_cfg.len=<\/span>${overlong_buffer_len}<\/span> \\\\<\/span>
\n +reward_model.reward_kwargs.overlong_buffer_cfg.penalty_factor=<\/span>${overlong_penalty_factor}<\/span> \\\\<\/span>
\n +reward_model.reward_kwargs.overlong_buffer_cfg.log=<\/span>false \\\\<\/span>
\n +reward_model.reward_kwargs.max_resp_len=<\/span>${max_response_length}<\/span> \\\\<\/span>
\n trainer.logger=<\/span>'["console"]'<\/span> \\\\<\/span>
\n actor_rollout_ref.rollout.enforce_eager=<\/span>True \\\\<\/span>
\n actor_rollout_ref.actor.use_torch_compile=<\/span>False \\\\<\/span>
\n trainer.project_name=<\/span>"${project_name}<\/span>"<\/span> \\\\<\/span>
\n trainer.experiment_name=<\/span>"${exp_name}<\/span>"<\/span> \\\\<\/span>
\n trainer.n_gpus_per_node=<\/span>8<\/span> \\\\<\/span>
\n trainer.nnodes=<\/span>1<\/span> \\\\<\/span>
\n trainer.val_before_train=<\/span>${val_before_train}<\/span> \\\\<\/span>
\n trainer.test_freq=<\/span>${test_freq}<\/span> \\\\<\/span>
\n trainer.save_freq=<\/span>-1 \\\\<\/span>
\n trainer.total_epochs=<\/span>${total_epochs}<\/span> \\\\<\/span>
\n trainer.total_training_steps=<\/span>${total_training_steps}<\/span> \\\\<\/span>
\n trainer.use_legacy_worker_impl=<\/span>disable \\\\<\/span>
\n trainer.default_local_dir=<\/span>"${CKPTS_DIR}<\/span>"<\/span> \\\\<\/span>
\n trainer.resume_mode=<\/span>auto \\\\<\/span>
\n trainer.device=<\/span>npu \\\\<\/span>
\n trainer.log_val_generations=<\/span>2<\/span> \\\\<\/span>
\n $@<\/span> <\/p>\n

\u8c03\u8bd5\u7ed3\u679c<\/h5>\n

\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\"<\/p>\n

\u7ed3\u8bba：reward\u66f2\u7ebf\u6b63\u5e38\u4e0a\u5347，\u8c03\u8bd5\u7ed3\u679c\u7b26\u5408\u8981\u6c42<\/p>\n

\u603b\u7ed3<\/h2>\n
  • GSPO\u57fa\u4e8eGRPO\u7b97\u6cd5，\u4ec5\u4fee\u6539\u635f\u5931\u51fd\u6570\u8ba1\u7b97\u90e8\u5206，\u8be5\u7b97\u6cd5\u5b9e\u73b0\u4e0d\u4f9d\u8d56\u786c\u4ef6\u5e95\u5c42<\/li>\n
  • \u4e0d\u9002\u7528dynamic_bsz\u7684\u65f6\u5019，\u9700\u8981\u914d\u7f6emicro_batch_size or use_flash_attn=True in config<\/li>\n
  • \u539f\u59cb\u811a\u672c\u57fa\u4e8eGPU\u8bad\u7ec3，\u5176\u5185\u5b58\u5f80\u5f80\u6bd4NPU\u66f4\u5927，\u4e3a\u907f\u514dOOM，\u53ef\u4ee5\u8c03\u6574gbs\u914d\u7f6e，\u5e76\u6839\u636eseq，\u6a21\u578b\u7b49\u52a8\u6001\u8c03\u6574gpu_utils_memory\u503c<\/li>\n
  • \u5bf9\u5e94GSPO\u7684\u751f\u6001pr：https:\/\/github.com\/volcengine\/verl\/pull\/4405<\/li>\n

    \u5e38\u89c1\u95ee\u9898\u4e0e\u89e3\u51b3\u65b9\u6848<\/h2>\n

    \u95ee\u98981：importlib.metadata.PackageNotFoundError:Npackage metadata was found for flash attnith overrides<\/h3>\n

    \u200b\u5b9a\u4f4d：transformer\u7684npu\u7248\u672cpatch\u672a\u6b63\u786e\u4f7f\u80fd，\u53ef\u4fee\u6539\u5982\u4e0b\u90e8\u5206\u4ee3\u7801\/\u200b\u200b\u4f7f\u80fd4.52.4\u62164.57.3\u7684transformer\u7248\u672c\u4e5f\u53ef\u4ee5\u89e3\u51b3**** \"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\"<\/p>\n

    \u5728\u4ee3\u7801\u5b9a\u4f4d\u8fc7\u7a0b\u4e2d\u53d1\u73b0\u5b98\u65b9bug：https:\/\/github.com\/volcengine\/verl\/pull\/3978，\u73b0\u5df2\u5408\u5165<\/p>\n

    \u95ee\u98982：raise e.remove_dynamo_frames() from None # see TORCHDYNAMO_VERBOSE=1<\/h3>\n

    File \u201c\/opt\/pyvenv\/lib\/python3.10\/site-packages\/torch\/_inductor\/compile_fx.py\u201d, line 760, in _compile_fx_inner<\/p>\n

    \u5b9a\u4f4d：\u8d70\u5230\u4e86\u56fe\u6a21\u5f0f \"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" \"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\"<\/p>\n

    \u6240\u4ee5\u589e\u52a0\u53c2\u6570 actor_rollout_ref.actor.use_torch_compile=False \\\\<\/p>\n

    \u95ee\u9898\u89e3\u51b3\u540e，\u62c9\u8d77\u8bad\u7ec3 \"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\"<\/p>\n

    \u95ee\u98983：\u5185\u5b58\u5206\u914d\u5931\u8d25<\/h3>\n

    \"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\"<\/p>\n

    \u8c03\u5c0fgbs\u914d\u7f6e，\u4fee\u6539gpu_utils_memory\u503c，\u8fd9\u4e2a\u8981\u6839\u636eseq，\u6a21\u578b\u52a8\u6001\u8c03\u6574，\u5426\u5219\u5bb9\u6613\u5b58\u5728\u788e\u7247\u5f20\u91cfOOM<\/p>\n

    \u95ee\u98984：RuntimeError: Please set micro_batch_size or set use_flash_attn=True in config.<\/h3>\n

    \"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\"<\/p>\n

    \u89e3\u51b3\u65b9\u6848：<\/p>\n

    rain_prompt_bsz=32
    \nn_resp_per_prompt=8
    \ntrain_prompt_mini_bsz=8<\/p>\n

    \u95ee\u98985 \u4f18\u5316\u5668\u66f4\u65b0\u9636\u6bb5OOM<\/h3>\n

    \"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\"<\/p>\n

    \u4fee\u6539\u8bad\u7ec3\u9636\u6bb5\u7684\u7b56\u7565，\u4f7f\u7528\u5e8f\u5217\u5e76\u884c<\/p>\n

    +actor_rollout_ref.actor.megatron.override_transformer_config.use_flash_attn=True \\\\<\/p>\n

    \u95ee\u98986.ValueError: num_query_groups (4) must be a multiple of tensor_model_parallel_size (8).<\/h3>\n

    \"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\"<\/p>\n

    \u6a21\u578b\u53c2\u6570\u9650\u5236<\/p>\n

    num_query_groups (4)\u9700\u8981\u80fd\u88abTP\u5207\u5206，\u5207\u6362\u8bad\u7ec3\u5207\u5206\u7b56\u7565<\/p>\n

    \u4fee\u6539\u540e，\u62c9\u8d77megaton\u540e\u7aef\u7684\u4ee3\u7801<\/p>\n

    \"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\"<\/p>\n

    \u95ee\u98987：\u57fa\u4e8eDAPO\u811a\u672c\u4fee\u6539GSPO\u7b97\u6cd5\u7684\u65f6\u5019，\u53d1\u751f\u62a5\u9519Could not override \u2018algorithm. filter_groups. enable\u2019.<\/h3>\n

    \u539f\u56e0：GSPO\u6ca1\u6709\u8fd9\u4e2afilter_groups\u5c5e\u6027，\u811a\u672c\u6539\u52a8\u758f\u5ffd\u4fee\u6539\u5373\u53ef \"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\"<\/p>\n","protected":false},"excerpt":{"rendered":"

    \u4f5c\u8005：\u6607\u817e\u5b9e\u6218\u6d3e \u77e5\u8bc6\u5730\u56fe：\u5f3a\u5316\u5b66\u4e60\u77e5\u8bc6\u5730\u56fe
    \n\u80cc\u666f\u4e0e\u610f\u4e49
    \n\u672c\u7bc7\u6587\u7ae0\u4e3b\u8981\u57fa\u4e8eVeRL\u6846\u67b6\u4e0a\u63d0\u51fa\u7684GSPO\u7b97\u6cd5\u5728\u6607\u817eNPU\u4e0a\u8fdb\u884c\u5b9e\u8df5\u90e8\u7f72，\u5e76\u4e3a\u5927\u5bb6\u7b80\u5355\u4ecb\u7ecdGRPO\u7b97\u6cd5\u601d\u60f3\u4ee5\u53ca\u5176\u548cGSPO\u7b97\u6cd5\u7279\u6027\u5dee\u5f02\u3002
    \n\u7b97\u6cd5\u539f\u7406
    \n\u8bba\u6587\u5730\u5740 GRPO：https:\/\/arxiv.org\/abs\/2402.03300 GSPO：https:\/\/arxiv.org\/abs\/2507.18071
    \nGRPO：\u7fa4\u7ec4\u76f8<\/p>\n","protected":false},"author":2,"featured_media":78312,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[81,50,207,427],"topic":[],"class_list":["post-78332","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-server","tag-python","tag-50","tag-207","tag-427"],"yoast_head":"\n\u57fa\u4e8eVeRL\u6846\u67b6\u7684GSPO\u7b97\u6cd5\u5728Atlas 800T A2\u670d\u52a1\u5668\u4e0a\u5b9e\u8df5 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3<\/title>\n<meta name=\"robots\" content=\"index, follow, max-snippet:-1, max-image-preview:large, max-video-preview:-1\" \/>\n<link rel=\"canonical\" href=\"https:\/\/www.wsisp.com\/helps\/78332.html\" \/>\n<meta property=\"og:locale\" content=\"zh_CN\" \/>\n<meta property=\"og:type\" content=\"article\" \/>\n<meta property=\"og:title\" content=\"\u57fa\u4e8eVeRL\u6846\u67b6\u7684GSPO\u7b97\u6cd5\u5728Atlas 800T A2\u670d\u52a1\u5668\u4e0a\u5b9e\u8df5 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\" \/>\n<meta property=\"og:description\" content=\"\u4f5c\u8005：\u6607\u817e\u5b9e\u6218\u6d3e \u77e5\u8bc6\u5730\u56fe：\u5f3a\u5316\u5b66\u4e60\u77e5\u8bc6\u5730\u56fe \u80cc\u666f\u4e0e\u610f\u4e49 \u672c\u7bc7\u6587\u7ae0\u4e3b\u8981\u57fa\u4e8eVeRL\u6846\u67b6\u4e0a\u63d0\u51fa\u7684GSPO\u7b97\u6cd5\u5728\u6607\u817eNPU\u4e0a\u8fdb\u884c\u5b9e\u8df5\u90e8\u7f72，\u5e76\u4e3a\u5927\u5bb6\u7b80\u5355\u4ecb\u7ecdGRPO\u7b97\u6cd5\u601d\u60f3\u4ee5\u53ca\u5176\u548cGSPO\u7b97\u6cd5\u7279\u6027\u5dee\u5f02\u3002 \u7b97\u6cd5\u539f\u7406 \u8bba\u6587\u5730\u5740 GRPO：https:\/\/arxiv.org\/abs\/2402.03300 GSPO：https:\/\/arxiv.org\/abs\/2507.18071 GRPO：\u7fa4\u7ec4\u76f8\" \/>\n<meta property=\"og:url\" content=\"https:\/\/www.wsisp.com\/helps\/78332.html\" \/>\n<meta property=\"og:site_name\" content=\"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\" \/>\n<meta property=\"article:published_time\" content=\"2026-02-26T19:34:11+00:00\" \/>\n<meta property=\"og:image\" content=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/02\/20260226193350-69a0a01ed9355.png\" \/>\n<meta name=\"author\" content=\"admin\" \/>\n<meta name=\"twitter:card\" content=\"summary_large_image\" \/>\n<meta name=\"twitter:label1\" content=\"\u4f5c\u8005\" \/>\n\t<meta name=\"twitter:data1\" content=\"admin\" \/>\n\t<meta name=\"twitter:label2\" content=\"\u9884\u8ba1\u9605\u8bfb\u65f6\u95f4\" \/>\n\t<meta name=\"twitter:data2\" content=\"27 \u5206\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\/\/schema.org\",\"@graph\":[{\"@type\":\"WebPage\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/78332.html\",\"url\":\"https:\/\/www.wsisp.com\/helps\/78332.html\",\"name\":\"\u57fa\u4e8eVeRL\u6846\u67b6\u7684GSPO\u7b97\u6cd5\u5728Atlas 800T A2\u670d\u52a1\u5668\u4e0a\u5b9e\u8df5 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\",\"isPartOf\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/#website\"},\"datePublished\":\"2026-02-26T19:34:11+00:00\",\"dateModified\":\"2026-02-26T19:34:11+00:00\",\"author\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41\"},\"breadcrumb\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/78332.html#breadcrumb\"},\"inLanguage\":\"zh-Hans\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\/\/www.wsisp.com\/helps\/78332.html\"]}]},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/78332.html#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"\u9996\u9875\",\"item\":\"https:\/\/www.wsisp.com\/helps\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"\u57fa\u4e8eVeRL\u6846\u67b6\u7684GSPO\u7b97\u6cd5\u5728Atlas 800T A2\u670d\u52a1\u5668\u4e0a\u5b9e\u8df5\"}]},{\"@type\":\"WebSite\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/#website\",\"url\":\"https:\/\/www.wsisp.com\/helps\/\",\"name\":\"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\",\"description\":\"\u9999\u6e2f\u670d\u52a1\u5668_\u9999\u6e2f\u4e91\u670d\u52a1\u5668\u8d44\u8baf_\u670d\u52a1\u5668\u5e2e\u52a9\u6587\u6863_\u670d\u52a1\u5668\u6559\u7a0b\",\"potentialAction\":[{\"@type\":\"SearchAction\",\"target\":{\"@type\":\"EntryPoint\",\"urlTemplate\":\"https:\/\/www.wsisp.com\/helps\/?s={search_term_string}\"},\"query-input\":\"required name=search_term_string\"}],\"inLanguage\":\"zh-Hans\"},{\"@type\":\"Person\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41\",\"name\":\"admin\",\"image\":{\"@type\":\"ImageObject\",\"inLanguage\":\"zh-Hans\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/image\/\",\"url\":\"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery\",\"contentUrl\":\"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery\",\"caption\":\"admin\"},\"sameAs\":[\"http:\/\/wp.wsisp.com\"],\"url\":\"https:\/\/www.wsisp.com\/helps\/author\/admin\"}]}<\/script>\n<!-- \/ Yoast SEO plugin. -->","yoast_head_json":{"title":"\u57fa\u4e8eVeRL\u6846\u67b6\u7684GSPO\u7b97\u6cd5\u5728Atlas 800T A2\u670d\u52a1\u5668\u4e0a\u5b9e\u8df5 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","robots":{"index":"index","follow":"follow","max-snippet":"max-snippet:-1","max-image-preview":"max-image-preview:large","max-video-preview":"max-video-preview:-1"},"canonical":"https:\/\/www.wsisp.com\/helps\/78332.html","og_locale":"zh_CN","og_type":"article","og_title":"\u57fa\u4e8eVeRL\u6846\u67b6\u7684GSPO\u7b97\u6cd5\u5728Atlas 800T A2\u670d\u52a1\u5668\u4e0a\u5b9e\u8df5 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","og_description":"\u4f5c\u8005：\u6607\u817e\u5b9e\u6218\u6d3e \u77e5\u8bc6\u5730\u56fe：\u5f3a\u5316\u5b66\u4e60\u77e5\u8bc6\u5730\u56fe \u80cc\u666f\u4e0e\u610f\u4e49 \u672c\u7bc7\u6587\u7ae0\u4e3b\u8981\u57fa\u4e8eVeRL\u6846\u67b6\u4e0a\u63d0\u51fa\u7684GSPO\u7b97\u6cd5\u5728\u6607\u817eNPU\u4e0a\u8fdb\u884c\u5b9e\u8df5\u90e8\u7f72，\u5e76\u4e3a\u5927\u5bb6\u7b80\u5355\u4ecb\u7ecdGRPO\u7b97\u6cd5\u601d\u60f3\u4ee5\u53ca\u5176\u548cGSPO\u7b97\u6cd5\u7279\u6027\u5dee\u5f02\u3002 \u7b97\u6cd5\u539f\u7406 \u8bba\u6587\u5730\u5740 GRPO：https:\/\/arxiv.org\/abs\/2402.03300 GSPO：https:\/\/arxiv.org\/abs\/2507.18071 GRPO：\u7fa4\u7ec4\u76f8","og_url":"https:\/\/www.wsisp.com\/helps\/78332.html","og_site_name":"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","article_published_time":"2026-02-26T19:34:11+00:00","og_image":[{"url":"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/02\/20260226193350-69a0a01ed9355.png"}],"author":"admin","twitter_card":"summary_large_image","twitter_misc":{"\u4f5c\u8005":"admin","\u9884\u8ba1\u9605\u8bfb\u65f6\u95f4":"27 \u5206"},"schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"WebPage","@id":"https:\/\/www.wsisp.com\/helps\/78332.html","url":"https:\/\/www.wsisp.com\/helps\/78332.html","name":"\u57fa\u4e8eVeRL\u6846\u67b6\u7684GSPO\u7b97\u6cd5\u5728Atlas 800T A2\u670d\u52a1\u5668\u4e0a\u5b9e\u8df5 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","isPartOf":{"@id":"https:\/\/www.wsisp.com\/helps\/#website"},"datePublished":"2026-02-26T19:34:11+00:00","dateModified":"2026-02-26T19:34:11+00:00","author":{"@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41"},"breadcrumb":{"@id":"https:\/\/www.wsisp.com\/helps\/78332.html#breadcrumb"},"inLanguage":"zh-Hans","potentialAction":[{"@type":"ReadAction","target":["https:\/\/www.wsisp.com\/helps\/78332.html"]}]},{"@type":"BreadcrumbList","@id":"https:\/\/www.wsisp.com\/helps\/78332.html#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"\u9996\u9875","item":"https:\/\/www.wsisp.com\/helps"},{"@type":"ListItem","position":2,"name":"\u57fa\u4e8eVeRL\u6846\u67b6\u7684GSPO\u7b97\u6cd5\u5728Atlas 800T A2\u670d\u52a1\u5668\u4e0a\u5b9e\u8df5"}]},{"@type":"WebSite","@id":"https:\/\/www.wsisp.com\/helps\/#website","url":"https:\/\/www.wsisp.com\/helps\/","name":"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","description":"\u9999\u6e2f\u670d\u52a1\u5668_\u9999\u6e2f\u4e91\u670d\u52a1\u5668\u8d44\u8baf_\u670d\u52a1\u5668\u5e2e\u52a9\u6587\u6863_\u670d\u52a1\u5668\u6559\u7a0b","potentialAction":[{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https:\/\/www.wsisp.com\/helps\/?s={search_term_string}"},"query-input":"required name=search_term_string"}],"inLanguage":"zh-Hans"},{"@type":"Person","@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41","name":"admin","image":{"@type":"ImageObject","inLanguage":"zh-Hans","@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/image\/","url":"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery","contentUrl":"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery","caption":"admin"},"sameAs":["http:\/\/wp.wsisp.com"],"url":"https:\/\/www.wsisp.com\/helps\/author\/admin"}]}},"_links":{"self":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts\/78332","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/comments?post=78332"}],"version-history":[{"count":0,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts\/78332\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/media\/78312"}],"wp:attachment":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/media?parent=78332"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/categories?post=78332"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/tags?post=78332"},{"taxonomy":"topic","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/topic?post=78332"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}