{"id":10021566,"date":"2026-11-17T10:06:42","date_gmt":"2026-11-17T02:06:42","guid":{"rendered":"https:\/\/itmax.cn\/?p=10021566"},"modified":"2026-11-17T10:06:55","modified_gmt":"2026-11-17T02:06:55","slug":"liger-kernel-20261117","status":"publish","type":"post","link":"https:\/\/itmax.cn\/?p=10021566","title":{"rendered":"\u5fae\u8c03\u4e00\u5f00\u957f\u5e8f\u5217\u5c31\u7206\u663e\u5b58\uff5cLiger-Kernel \u6211\u5728 Mac \u4e0a\u6ca1\u88c5\u6210\uff0c\u8d26\u5012\u662f\u7b97\u660e\u767d\u4e86"},"content":{"rendered":"<h1 id=\"liger-kernel-mac\">\u5fae\u8c03\u4e00\u5f00\u957f\u5e8f\u5217\u5c31\u7206\u663e\u5b58\uff5cLiger-Kernel \u6211\u5728 Mac \u4e0a\u6ca1\u88c5\u6210\uff0c\u8d26\u5012\u662f\u7b97\u660e\u767d\u4e86<\/h1>\n<h2 id=\"_1\">\u663e\u5b58\u5230\u5e95\u88ab\u8c01\u5403\u6389\u7684\uff0c\u6211\u7b97\u8fc7\u4e00\u7b14\u5f88\u618b\u5c48\u7684\u8d26<\/h2>\n<p>\u4e0a\u4e2a\u6708\u5e2e\u540c\u4e8b\u8c03\u4e00\u4e2a 8B \u6a21\u578b\u7684\u5fae\u8c03\u4efb\u52a1\uff0c\u5e8f\u5217\u957f\u5ea6\u4ece 4K \u62c9\u5230 8K\uff0c\u7b2c\u4e8c\u4e2a step \u5c31 CUDA out of memory\u3002\u4ed6\u7684\u7b2c\u4e00\u53cd\u5e94\u662f\u719f\u6089\u7684\u5957\u9910\uff1a\u964d batch size\u3001\u5f00 gradient checkpointing\u3001\u518d\u628a LoRA rank \u780d\u534a\u3002\u6298\u817e\u5230\u51cc\u6668\uff0c\u663e\u5b58\u662f\u4e0b\u6765\u4e86\uff0c\u541e\u5410\u4e5f\u8ddf\u7740\u8170\u65a9\u2014\u2014gradient checkpointing \u672c\u8d28\u662f\u62ff\u65f6\u95f4\u6362\u663e\u5b58\uff0c\u53cd\u5411\u65f6\u628a\u4e22\u6389\u7684\u6fc0\u6d3b\u91cd\u7b97\u4e00\u904d\uff0c\u591a\u51fa\u4e24\u6210\u5de6\u53f3\u7684 step \u65f6\u95f4\uff0c\u8fd9\u7b14\u8d26\u600e\u4e48\u7b97\u90fd\u4e0d\u8212\u670d\u3002\u6211\u5f53\u65f6\u5c31\u7eb3\u95f7\uff1a\u6a21\u578b\u6743\u91cd\u5341\u51e0 GB \u6709\u529e\u6cd5\u91cf\u5316\uff0c\u4f18\u5316\u5668\u72b6\u6001\u6709 ZeRO \u5206\u644a\uff0c\u600e\u4e48\u8fd8\u6709\u4e00\u5927\u5757\u663e\u5b58\u8c01\u90fd\u5543\u4e0d\u52a8\uff1f<\/p>\n<p>\u540e\u6765\u6211\u628a\u8d26\u644a\u5f00\u7b97\u4e86\u4e00\u904d\uff0c\u5143\u51f6\u5c45\u7136\u662f\u6700\u4e0d\u8d77\u773c\u7684 loss \u5c42\u30028B \u7ea7\u6a21\u578b\u7684\u8bcd\u8868\u666e\u904d\u5728 128K \u4e0a\u4e0b\uff0cHugging Face \u90a3\u5957\u6807\u51c6\u5b9e\u73b0\u8981\u5148\u628a\u6574\u4e2a logits \u5f20\u91cf\u7269\u5316\u51fa\u6765\u518d\u7b97\u4ea4\u53c9\u71b5\uff1abatch size 8\u3001\u5e8f\u5217 4096\u3001\u8bcd\u8868 128256\u3001bf16 \u5b58\u4e0b\u6765\uff0c\u5149 logits \u5c31\u662f 8 \u00d7 4096 \u00d7 128256 \u00d7 2 \u5b57\u8282\uff0c\u5dee\u4e0d\u591a 8.4 GB\u3002\u8fd9\u8fd8\u53ea\u662f\u524d\u5411\uff0c\u53cd\u5411\u8981\u4fdd\u7559\u7684\u4e1c\u897f\u66f4\u591a\u3002\u5e8f\u5217\u7ffb\u500d\u8fd9\u5757\u5c31\u7ffb\u500d\uff0c\u8ddf\u6a21\u578b\u5927\u5c0f\u65e0\u5173\uff0c\u7eaf\u7cb9\u662f\u8bcd\u8868\u548c\u4e0a\u4e0b\u6587\u957f\u5ea6\u5728\u4e58\u3002\u5f53\u5e74\u8bad BERT \u8bcd\u8868\u624d\u4e09\u4e07\uff0c\u8fd9\u5757\u6839\u672c\u4e0d\u53eb\u4e8b\uff1b\u73b0\u5728\u5927\u8bcd\u8868\u52a0\u957f\u4e0a\u4e0b\u6587\uff0closs \u5c42\u53cd\u5012\u6210\u4e86\u663e\u5b58\u7684\u5934\u53f7\u6740\u624b\u3002<\/p>\n<p>\u66f4\u6c14\u4eba\u7684\u662f\u8fd9\u7b14\u5f00\u9500\u7684\u6027\u8d28\uff1a\u5b83\u4e0d\u662f\u6743\u91cd\u3001\u4e0d\u662f\u4f18\u5316\u5668\u72b6\u6001\uff0c\u7eaf\u7cb9\u662f\u300c\u7b97 loss \u7684\u4e2d\u95f4\u8fc7\u7a0b\u300d\uff0c\u7b97\u5b8c\u5c31\u6254\uff0c\u5374\u8981\u6c42\u5168\u7a0b\u9a7b\u7559\u663e\u5b58\u3002\u6743\u91cd\u53ef\u4ee5\u91cf\u5316\u3001\u4f18\u5316\u5668\u53ef\u4ee5\u5206\u7247\u3001\u6fc0\u6d3b\u53ef\u4ee5\u91cd\u7b97\uff0c\u552f\u72ec\u8fd9\u4e00\u5757\u5728\u6807\u51c6\u5b9e\u73b0\u91cc\u6ca1\u6709\u4efb\u4f55\u56de\u65cb\u4f59\u5730\u2014\u2014\u9664\u975e\u4f60\u628a\u300c\u5148\u7269\u5316 logits \u518d\u7b97 loss\u300d\u8fd9\u4e2a\u987a\u5e8f\u672c\u8eab\u6539\u6389\u3002<\/p>\n<p>\u987a\u7740\u8fd9\u4e2a\u7ed3\u8bba\u53bb\u627e\u89e3\u6cd5\uff0c\u5c31\u649e\u89c1\u4e86 LinkedIn \u5f00\u6e90\u7684 Liger-Kernel\u2014\u2014\u4e00\u5957\u4e13\u95e8\u7ed9 LLM \u8bad\u7ec3\u5199\u7684 Triton \u7b97\u5b50\u5e93\uff0c\u4e3b\u6253\u7684\u5c31\u662f\u628a RMSNorm\u3001RoPE\u3001SwiGLU\u3001\u4ea4\u53c9\u71b5\u8fd9\u51e0\u4e2a\u9ad8\u9891\u5c42\u6362\u6389\uff0c\u5176\u4e2d\u6700\u72e0\u7684\u4e00\u62db\u53eb FusedLinearCrossEntropy\uff1a\u4e0d\u8ba9 logits \u6574\u4e2a\u843d\u5730\uff0c\u6309 chunk \u5206\u5757\u7b97\u5b8c\u5c31\u4e22\uff0c\u5b98\u65b9\u53e3\u5f84\u4e0b\u591a\u5361\u8bad\u7ec3\u541e\u5410\u63d0\u5347 20%\u3001\u663e\u5b58\u7701 60%\uff0c\u539f\u672c 4K \u4e0a\u4e0b\u6587\u5c31 OOM \u7684 LLaMA 3-8B \u80fd\u4e00\u8def\u6491\u5230 16K\u3002\u6570\u5b57\u5148\u653e\u5728\u8fd9\uff0c\u4e0b\u9762\u6211\u4f1a\u62c6\u5f00\u8bb2\u54ea\u4e9b\u662f\u771f\u91d1\u3001\u54ea\u4e9b\u8981\u6253\u6298\u3002<\/p>\n<h2 id=\"patch\">\u4e00\u884c patch \u80cc\u540e\uff0c\u5176\u5b9e\u6709\u4e09\u6863\u7528\u6cd5<\/h2>\n<p>Liger-Kernel \u7684\u63a5\u5165\u65b9\u5f0f\u5206\u4e09\u6863\uff0c\u590d\u6742\u5ea6\u548c\u7075\u6d3b\u6027\u9012\u589e\u3002<\/p>\n<p>\u7b2c\u4e00\u6863\u6700\u7701\u4e8b\uff0c\u6362\u4e00\u4e2a\u6a21\u578b\u52a0\u8f7d\u7c7b\uff1a<\/p>\n<pre><code class=\"language-python\">from liger_kernel.transformers import AutoLigerKernelForCausalLM\n\nmodel = AutoLigerKernelForCausalLM.from_pretrained(&quot;meta-llama\/Meta-Llama-3-8B&quot;)\n<\/code><\/pre>\n<p>\u6a21\u578b\u7c7b\u578b\u5728\u652f\u6301\u5217\u8868\u91cc\u5c31\u81ea\u52a8 monkey-patch\uff0c\u9ed8\u8ba4\u5168\u5f00\u3002\u7b2c\u4e8c\u6863\u662f\u663e\u5f0f\u6307\u5b9a\u7b97\u5b50\uff0c\u6bd4\u5982\u53ea\u8981\u878d\u5408\u4ea4\u53c9\u71b5\u3001\u4e0d\u8981\u52a8 RoPE\uff1a<\/p>\n<pre><code class=\"language-python\">from liger_kernel.transformers import apply_liger_kernel_to_qwen3\n\napply_liger_kernel_to_qwen3(\n    rope=True,\n    swiglu=True,\n    cross_entropy=True,\n    fused_linear_cross_entropy=True,\n    rms_norm=False,\n)\n<\/code><\/pre>\n<p>\u7b2c\u4e09\u6863\u662f\u628a\u5355\u4e2a\u7b97\u5b50\u5f53\u666e\u901a\u6a21\u5757\u7528\uff0c\u9002\u5408\u81ea\u5df1\u6512\u6a21\u578b\u7684\u4eba\uff1a<\/p>\n<pre><code class=\"language-python\">from liger_kernel.transformers import LigerFusedLinearCrossEntropyLoss\n\nloss_fn = LigerFusedLinearCrossEntropyLoss()\nloss = loss_fn(lm_head.weight, hidden_states, targets)\n<\/code><\/pre>\n<p>\u6309\u80fd\u529b\u5206\u7ec4\u8bf4\u6211\u7684\u5224\u65ad\u3002<strong>\u6a21\u578b patching \u8fd9\u4e00\u6863\u662f\u6210\u719f\u4e3b\u529b<\/strong>\uff1aREADME \u7684\u652f\u6301\u8868\u4ece Llama\u3001Qwen2\/2.5\/3\/3.5\u3001Gemma2\/3\/4\u3001Mistral\u3001DeepSeek-V3\/V4 \u4e00\u76f4\u5230 GLM-4\u3001InternVL3\uff0c\u51e0\u5341\u4e2a\u6a21\u578b\u6709\u4e13\u5c5e patch API\uff0cHugging Face Trainer\uff08PR #32860 \u8d77\uff09\u548c TRL \u4e5f\u5185\u7f6e\u4e86\u5f00\u5173\uff0cAxolotl\u3001LLaMA-Factory\u3001ms-swift\u3001oumi \u8fd9\u4e9b\u6846\u67b6\u5168\u90e8\u63a5\u4e86\u8fdb\u53bb\u3002<strong>\u540e\u8bad\u7ec3 loss \u662f\u7b2c\u4e8c\u589e\u957f\u66f2\u7ebf<\/strong>\uff1aDPO\u3001ORPO\u3001SimPO\u3001KTO\u3001JSD\u3001GRPO \u90fd\u6709\u5206\u5757\u5b9e\u73b0\uff0cGRPO \u8fd8\u5e26 bnpo\u3001dr_grpo\u3001dapo\u3001cispo \u7b49\u4e00\u4e32\u53d8\u4f53\uff0c\u5b98\u65b9\u79f0\u5bf9\u9f50\u7c7b\u4efb\u52a1\u6700\u9ad8\u7701 80% \u663e\u5b58\u2014\u2014v0.5.0 \u53d1\u5e03\u65f6\u7684\u5356\u70b9\u5c31\u662f\u8fd9\u7ec4\u3002\u7528\u6cd5\u662f\u666e\u901a\u6a21\u5757\uff0c\u76f4\u63a5\u5582 lm_head \u6743\u91cd\uff1a<\/p>\n<pre><code class=\"language-python\">from liger_kernel.chunked_loss import LigerFusedLinearORPOLoss\n\norpo_loss = LigerFusedLinearORPOLoss()\ny = orpo_loss(lm_head.weight, x, target)\n<\/code><\/pre>\n<p>\u8fd9\u7c7b\u5bf9\u9f50 loss \u6bd4\u4ea4\u53c9\u71b5\u66f4\u5403\u663e\u5b58\uff0c\u56e0\u4e3a\u504f\u597d\u8bad\u7ec3\u8981\u540c\u65f6\u8fc7 chosen \u548c rejected \u4e24\u6761\u5e8f\u5217\uff0clogits \u7ffb\u7740\u500d\u5730\u957f\uff0c\u5206\u5757\u5316\u7684\u6536\u76ca\u4e5f\u8ddf\u7740\u7ffb\u500d\u3002<strong>\u65b0\u540e\u7aef\u5c5e\u4e8e\u6fc0\u8fdb\u533a<\/strong>\uff1acuTile \u548c CuTe DSL \u9762\u5411 Hopper\uff08SM90\uff09\u548c Blackwell\uff08SM100\/110\uff09\uff0cREADME \u91cc H100 \u4e0a\u90a3\u4efd Fused Scaled Cross Entropy \u8868\u683c\u5f88\u6f02\u4eae\uff0cCuTe SM90 \u8def\u5f84 3.12 \u6beb\u79d2 706 TFLOPS\uff0c\u5bf9\u6bd4 Verl \u7684 Torch fallback 11.36 \u6beb\u79d2 194 TFLOPS\uff0c\u5dee\u7740\u4e09\u500d\u591a\uff1bB200 \u4e0a cuTile \u524d\u53cd\u5411\u5168\u7a0b 11.36 \u6beb\u79d2\u3001\u5cf0\u503c 2.64 GiB\uff0c\u5bf9 Torch \u7684 7.22 GiB\u3002\u4f46\u8fd9\u5957\u4e1c\u897f\u9700\u8981 CUDA Toolkit 13.1 \u8fd9\u7c7b\u73af\u5883\uff0c\u666e\u901a\u7528\u6237\u6682\u65f6\u5f53\u65b0\u95fb\u770b\u5c31\u884c\u3002\u53e6\u5916\u5b83\u5bf9\u6607\u817e NPU \u7684\u652f\u6301\u5728 2026 \u5e74\u63a8\u8fdb\u5f97\u5f88\u731b\uff0cissue \u533a\u751a\u81f3\u6302\u7740\u4e13\u95e8\u7684 NPU \u8def\u7ebf\u56fe\uff08#969\uff0c33 \u6761\u8bc4\u8bba\uff09\uff0c\u63d0\u4ea4\u8bb0\u5f55\u91cc\u80fd\u770b\u5230\u4e00\u6279 <code>fix(npu)<\/code> \u8fdb main\u3002<\/p>\n<p>\u54ea\u4e9b\u53ea\u662f\u5671\u5934\uff1f\u6211\u73b0\u5728\u6ca1\u6709\u53d1\u73b0\u660e\u663e\u7684\u2014\u2014\u8fd9\u4e2a\u9879\u76ee\u4e1a\u52a1\u4e0a\u5f88\u514b\u5236\uff0c\u53ea\u505a\u7b97\u5b50\u5c42\uff0c\u4e0d\u78b0\u5206\u5e03\u5f0f\u901a\u4fe1\u548c\u663e\u5b58\u7ba1\u7406\uff0c\u6ca1\u6709\u90a3\u79cd\u300c\u4e00\u7ad9\u5f0f\u8bad\u7ec3\u5e73\u53f0\u300d\u7684\u81a8\u80c0\u53d9\u4e8b\u3002\u8981\u8bf4\u6c34\u5206\uff0c\u662f\u7b2c\u4e09\u65b9\u8f6c\u8ff0\u91cc\u7684\uff1a\u7f51\u4e0a\u4e0d\u5c11\u6587\u7ae0\u628a\u5b83\u8bf4\u6210\u300cLightning AI \u5f00\u6e90\u300d\u300c\u5b57\u8282\u8df3\u52a8\u51fa\u54c1\u300d\uff0c\u90fd\u4e0d\u662f\uff0c\u5b83\u4ece 2024 \u5e74 8 \u6708\u8d77\u5c31\u5728 LinkedIn \u7684 GitHub \u7ec4\u7ec7\u4e0b\uff0c\u5b98\u65b9 engineering blog \u5199\u5f97\u6e05\u6e05\u695a\u695a\u3002<\/p>\n<h2 id=\"logits\">\u62ff logits \u6362\u663e\u5b58\uff0c\u8fd9\u7b14\u4ea4\u6613\u5728\u4e09\u5904\u4f1a\u9732\u9985<\/h2>\n<p>\u8bbe\u8ba1\u4e0a\u5b83\u548c Flash-Attention \u662f\u540c\u4e00\u4e2a\u601d\u8def\u7684\u5ef6\u4f38\uff1a\u4e0d\u8fd1\u4f3c\u3001\u7b97\u5f97\u7cbe\u786e\uff0c\u9760 kernel fusion\u3001\u539f\u5730\u66ff\u6362\u548c\u5206\u5757\u7701\u663e\u5b58\u7701\u5e26\u5bbd\u3002\u62ff\u6700\u6838\u5fc3\u7684 FusedLinearCrossEntropy \u8bf4\u4eba\u8bdd\uff1a\u6807\u51c6\u6d41\u7a0b\u662f\u300clm_head \u5168\u91cf\u6295\u5f71\u51fa logits \u2192 \u7b97\u4ea4\u53c9\u71b5 \u2192 \u53cd\u5411\u300d\uff0c\u5b83\u6539\u6210\u628a token \u6309 chunk \u5207\u5f00\uff0c\u6bcf\u4e2a chunk \u6295\u5f71\u5b8c\u7acb\u523b\u7b97 loss\u3001\u7acb\u523b\u628a\u68af\u5ea6\u6512\u56de lm_head\uff0c\u4e2d\u95f4 logits \u7528\u5b8c\u5373\u5f03\uff0c\u5168\u7a0b\u4e0d\u843d\u5730\u3002\u4ee3\u4ef7\u662f\u90a3\u4e2a\u300c\u7acb\u523b\u300d\u5fc5\u987b\u5728\u540c\u4e00\u4e2a Triton kernel \u91cc\u5b8c\u6210\uff0c\u4e8e\u662f\u4f60\u62ff\u5230\u7684\u4e0d\u518d\u662f\u5f20\u91cf\u800c\u662f\u6807\u91cf loss\uff0c\u51e1\u662f\u8981\u300c\u770b\u770b logits \u957f\u4ec0\u4e48\u6837\u300d\u7684\u4ee3\u7801\u90fd\u5f97\u6539\u9053\u3002\u6211\u8ba4\u540c\u8fd9\u4e2a\u53d6\u820d\u2014\u2014\u6570\u503c\u4e0a\u5b83\u6709\u5bf9\u9f50\u6d4b\u8bd5\u515c\u5e95\uff0cREADME \u660e\u786e\u8bf4\u53cd\u5411\u4e5f\u505a\u4e86\u5355\u5143\u6d4b\u8bd5\u548c\u6536\u655b\u6bd4\u5bf9\uff0c\u7701\u4e0b\u7684\u53ef\u662f\u5b9e\u6253\u5b9e\u7684 8 GB \u7ea7\u5cf0\u503c\u3002\u4f46\u5929\u4e0b\u6ca1\u6709\u767d\u7701\u7684\u663e\u5b58\uff0c\u4ee3\u4ef7\u5728\u4e09\u4e2a\u5730\u65b9\uff0c\u90fd\u5f88\u5177\u4f53\u3002<\/p>\n<p>\u7b2c\u4e00\u5904\uff0c<strong>logits \u6ca1\u4e86<\/strong>\u3002FusedLinearCrossEntropy \u4e0d\u7269\u5316 logits\uff0c\u8fd9\u662f\u5b83\u7701 8 GB \u7684\u5168\u90e8\u539f\u7406\uff0c\u526f\u4f5c\u7528\u662f\u4efb\u4f55\u4f9d\u8d56\u5b8c\u6574 logits \u7684\u4e0b\u6e38\u7edf\u8ba1\u90fd\u65ad\u4e86\u3002TRL \u96c6\u6210\u6587\u6863\u91cc\u5199\u5f97\u76f4\u767d\uff1aLiger \u6a21\u5f0f\u4e0b\u4e0d\u8fd4\u56de logits\uff0c\u624b\u52a8\u7b97\u71b5\u4e4b\u7c7b\u7684\u903b\u8f91\u5f97\u6539\u8d70\u5185\u6838\u8fd4\u56de\u7684 token_accuracy\uff1b<code>loss_type='chunked_nll'<\/code> \u548c\u5b83\u76f4\u63a5\u51b2\u7a81\uff0c\u542f\u7528\u5c31\u629b ValueError\u3002\u5982\u679c\u4f60\u7684\u8bad\u7ec3\u811a\u672c\u91cc\u585e\u4e86\u4e00\u5806\u81ea\u5b9a\u4e49\u7684 logits \u540e\u5904\u7406\uff0c\u5148\u505a\u597d\u91cd\u6784\u7684\u51c6\u5907\u3002<\/p>\n<p>\u7b2c\u4e8c\u5904\uff0c<strong>\u7f16\u8bd1\u6469\u64e6<\/strong>\u3002issue #174\u300ctorch.compile() throws exception when LigerKernel is used\u300d\u4ece 2024 \u5e74 8 \u6708\u6302\u5230\u73b0\u5728\u8fd8\u662f open\uff0c24 \u6761\u8bc4\u8bba\u3002\u6709\u610f\u601d\u7684\u662f\u7ffb\u6700\u8fd1\u7684\u63d0\u4ea4\u8bb0\u5f55\uff1a11 \u6708 10 \u65e5\u7684 <code>fix(compile): use tl.cast for scalar kernel params so torch.compile does not fall back<\/code>\uff08#1544\uff09\u548c 11 \u6708 14 \u65e5\u7684 <code>fix(utils): skip device_context guard under torch.compile<\/code>\uff08#1590\uff09\u90fd\u5728\u4fee\u8fd9\u4e2a\u65b9\u5411\u7684\u95ee\u9898\u3002\u7ef4\u62a4\u6001\u5ea6\u6ca1\u95ee\u9898\uff0c\u4f46\u5982\u679c\u4f60\u6307\u671b Liger \u548c torch.compile \u65e0\u75db\u53e0\u52a0\uff0c\u73b0\u9636\u6bb5\u8981\u7559\u4e00\u624b\uff0c\u9010\u7248\u672c\u9a8c\u8bc1\u3002<\/p>\n<p>\u7b2c\u4e09\u5904\uff0c<strong>\u5e73\u53f0\u9501\u5b9a<\/strong>\u3002Triton \u540e\u7aef\u610f\u5473\u7740 Linux + NVIDIA \/ AMD ROCm \/ \u6607\u817e\uff0cMac \u662f\u5f7b\u5e95\u51fa\u5c40\u3002\u8fd9\u4ef6\u4e8b\u6211\u4e0d\u662f\u731c\u7684\uff0c\u4e0b\u4e00\u6bb5\u6709\u6211\u7684\u5b89\u88c5\u5b9e\u5f55\u3002<\/p>\n<p>\u8fd9\u4e09\u5904\u91cc\uff0c\u300c\u4ec0\u4e48\u60c5\u51b5\u4e0b\u4e0d\u53ef\u63a5\u53d7\u300d\u6211\u8bf4\u5f97\u5f88\u660e\u786e\uff1a\u8bad\u7ec3\u7ba1\u7ebf\u91cc\u5927\u91cf\u4f9d\u8d56 logits \u81ea\u5b9a\u4e49\u7edf\u8ba1\u3001\u53c8\u6ca1\u4eba\u529b\u91cd\u6784\u7684\u56e2\u961f\uff0c\u4ee5\u53ca\u751f\u4ea7\u73af\u5883\u8981\u6c42 torch.compile \u5168\u5f00\u7684\u6d41\u6c34\u7ebf\uff0c\u73b0\u5728\u4e0a Liger \u662f\u7ed9\u81ea\u5df1\u627e\u6d3b\u5e72\u3002\u53cd\u8fc7\u6765\uff0c\u6807\u51c6 SFT\/DPO \u6d41\u7a0b\u52a0 HF Trainer \u7684\u7ec4\u5408\uff0c\u6469\u64e6\u6700\u5c0f\u3002<\/p>\n<h2 id=\"mac\">\u5b89\u88c5\u8fd9\u4ef6\u4e8b\uff0c\u6211\u5728 Mac \u4e0a\u6298\u621f\uff0c\u4f46\u6b7b\u5f97\u660e\u660e\u767d\u767d<\/h2>\n<p>\u5148\u4ea4\u4ee3\u7acb\u573a\uff1a\u6211\u624b\u4e0a\u53ea\u6709\u4e00\u53f0 M2 \u7684 MacBook\uff0c\u6ca1\u6709 N \u5361\u3002\u6211\u8fd8\u662f\u51b3\u5b9a\u771f\u88c5\u4e00\u904d\u8bd5\u8bd5\uff0c\u56e0\u4e3a\u300c\u652f\u4e0d\u652f\u6301 Mac\u300d\u8fd9\u4ef6\u4e8b\u7f51\u4e0a\u8bf4\u6cd5\u4e0d\u4e00\uff0c\u503c\u5f97\u4eb2\u624b\u9a8c\u3002<\/p>\n<p>\u865a\u62df\u73af\u5883\u5efa\u597d\uff0c<code>pip install liger-kernel<\/code>\uff0c\u4f9d\u8d56\u89e3\u6790\u76f4\u63a5\u62a5 ResolutionImpossible\u3002\u62c6\u5f00\u770b\uff1atorch \u672c\u8eab\u6ca1\u610f\u89c1\uff0c2.15.0 \u987a\u5229\u88c5\u4e0a\uff1b\u5361\u4f4f\u7684\u662f triton\u2014\u2014PyPI \u4e0a\u7ffb\u904d\u4e86\u5b83\u7684\u53d1\u5e03\u8bb0\u5f55\uff0c\u538b\u6839\u6ca1\u6709 macOS \u7684 wheel\uff0cLinux \u4e13\u5c5e\u3002Triton \u88c5\u4e0d\u4e0a\uff0c<code>triton&gt;=2.3.1<\/code> \u8fd9\u6761\u4f9d\u8d56\u5c31\u6b7b\u9501\u4e86\uff0c\u4e00\u884c kernel \u90fd\u522b\u60f3\u8dd1\u3002\u6240\u4ee5\u7ed3\u8bba\u5f88\u5e72\u51c0\uff1aMac \u7528\u6237\u4e0d\u7528\u6323\u624e\uff0c\u8fd9\u4e0d\u662f\u88c5\u4e0d\u88c5\u5f97\u8d39\u52b2\u7684\u95ee\u9898\uff0c\u662f\u95e8\u90fd\u6ca1\u6709\uff1bApple Silicon \u60f3\u8e6d\u8fd9\u7c7b\u7b97\u5b50\u4f18\u5316\uff0c\u76ee\u524d\u53ea\u80fd\u7b49 MLX \u751f\u6001\u81ea\u5df1\u957f\u3002<\/p>\n<p>\u6b63\u7ecf\u7684\u5b89\u88c5\u53d1\u751f\u5728 Linux \u670d\u52a1\u5668\u4e0a\uff0c\u59ff\u52bf\u5c31\u4e24\u6761\uff1a<\/p>\n<pre><code class=\"language-bash\"># \u7a33\u5b9a\u7248\npip install liger-kernel\n\n# \u9700\u8981\u7528\u6a21\u578b patching API \u7684\u8bdd\uff0ctransformers \u662f\u53ef\u9009\u4f9d\u8d56\npip install &quot;liger-kernel[dev]&quot;  # \u6216\u81ea\u884c\u786e\u4fdd transformers &gt;= 5.0.0\n<\/code><\/pre>\n<p>\u4f9d\u8d56\u95e8\u69db\u5b98\u65b9\u5199\u5f97\u6e05\u695a\uff1a<code>torch &gt;= 2.1.2<\/code>\u3001<code>triton &gt;= 2.3.1<\/code>\u3002AMD \u5361\u8d70 ROCm\uff0c\u8981\u6c42 torch &gt;= 2.5.0\u3001triton &gt;= 3.0.0\uff1b\u6607\u817e\u8981\u6c42 torch 2.10.0 \u914d CANN 9.1.0 \u548c triton-ascend\u3002\u8ffd\u65b0\u7684\u4eba\u8fd8\u6709 nightly \u6e20\u9053\uff08<code>pip install liger-kernel-nightly<\/code>\uff09\uff0c\u8981\u6e90\u7801\u88c5\u5c31 clone \u4ed3\u5e93\u8dd1 <code>pip install -e .<\/code>\uff0csetup.py \u4f1a\u81ea\u52a8\u63a2\u6d4b\u672c\u5730\u540e\u7aef\u9009\u4f9d\u8d56\uff1b\u8ffd H100\/B200 \u6781\u81f4\u6570\u5b57\u7684\u518d\u8865 <code>liger-kernel[lck]<\/code> \u90a3\u7ec4\u539f\u751f\u901a\u4fe1 kernel extra\uff0c\u5b98\u65b9\u8fde\u300cCUDA 12.9 \u7684\u539f\u751f wheel\u3001CUDA 13 \u8d70\u6e90\u7801\u7f16\u8bd1\u3001Blackwell MoE \u53cd\u5411\u8fd8\u6709\u963b\u585e\u300d\u8fd9\u79cd\u7ec6\u8282\u90fd\u5199\u5728 README \u91cc\u4e86\uff0c\u5766\u8bda\u5f97\u5c11\u89c1\u3002\u4f9d\u8d56\u8fd9\u4e48\u5c11\uff08\u53ea\u6709 Torch \u548c Triton \u4e24\u4e2a\u786c\u4f9d\u8d56\uff09\u662f\u5b83\u7684\u53e6\u4e00\u4e2a\u5356\u70b9\u2014\u2014\u8bad\u7ec3\u73af\u5883\u88ab CUDA \u7248\u672c\u5730\u72f1\u6298\u78e8\u8fc7\u7684\u4eba\u90fd\u61c2\u8fd9\u610f\u5473\u7740\u4ec0\u4e48\u3002<\/p>\n<p>\u786c\u4ef6\u95e8\u69db\u8981\u5355\u72ec\u8bf4\uff1a<strong>\u5b98\u65b9\u6ca1\u7ed9\u6574\u673a\u914d\u7f6e\u8868<\/strong>\u2014\u2014\u4e00\u4e2a\u7b97\u5b50\u5e93\u786e\u5b9e\u4e5f\u4e0d\u8be5\u7ed9\uff0c\u4f60\u7684\u663e\u5b58\u9700\u6c42\u7531\u8bad\u7ec3\u7684\u6a21\u578b\u51b3\u5b9a\uff0cLiger \u53ea\u8d1f\u8d23\u628a\u5cf0\u503c\u5f80\u4e0b\u538b\u3002\u6211\u6309\u5b98\u65b9 benchmark \u6761\u4ef6\u53cd\u63a8\u4e24\u6863\u4f9b\u53c2\u8003\uff1a\u5b98\u65b9\u90a3\u5f20\u300c\u541e\u5410 +20% \/ \u663e\u5b58 -60%\u300d\u56fe\u662f\u5728 8 \u5f20 A100 \u4e0a\u5168\u53c2\u5fae\u8c03 LLaMA 3-8B \u8dd1\u7684\uff0c\u8fd9\u662f\u5b83\u7684\u300c\u4e3b\u573a\u300d\uff1b\u5355\u5361\u573a\u666f\u4e0b\u7b2c\u4e09\u65b9\u5b9e\u6d4b\uff08RTX 4090\u3001Llama-3-8B QLoRA\uff09\u662f\u663e\u5b58 14.8 GB \u964d\u5230 13.2 GB\u3001\u541e\u5410 42 \u5230 51 samples\/s\uff0c\u63d0\u5347 21%\u3002\u7c97\u7565\u6362\u7b97\uff0c8 GB \u7ea7\u5355\u5361\u505a 7B QLoRA \u53ef\u4ee5\u8fdb\u95e8\uff0c\u8fd9\u662f\u7ecf\u9a8c\u63a8\u7b97\u4e0d\u662f\u5b98\u65b9\u6570\u5b57\uff1b\u60f3\u590d\u73b0\u5b98\u65b9\u90a3\u4e2a 60% \u7684\u663e\u5b58\u6536\u76ca\uff0c\u8bf7\u6309 8\u00d7A100 \u7684\u5168\u53c2\u573a\u666f\u5bf9\u53f7\u5165\u5ea7\uff0c\u6d88\u8d39\u7ea7\u5355\u5361\u62ff\u4e0d\u5230\u8fd9\u4e2a\u6570\u3002<\/p>\n<h2 id=\"gpu\">\u4e0a\u624b\u8dd1\u4e00\u904d\uff1a\u6211\u6ca1\u6709 GPU\uff0c\u4f46\u6d41\u7a0b\u6838\u5230\u4e86\u6bcf\u4e00\u884c<\/h2>\n<p>\u8001\u5b9e\u4ea4\u4ee3\uff1a\u6211\u672c\u673a\u8fde\u88c5\u90fd\u88c5\u4e0d\u4e0a\uff0c\u7aef\u5230\u7aef\u793a\u4f8b\u6ca1\u6cd5\u4eb2\u624b\u8dd1\u3002\u4e0b\u9762\u8fd9\u5957\u6d41\u7a0b\u662f\u6211\u5bf9\u7740\u5b98\u65b9 examples\u3001Hugging Face Trainer \u6587\u6863\u548c TRL \u96c6\u6210\u8bf4\u660e\u9010\u884c\u6838\u8fc7\u7684\uff0c\u547d\u4ee4\u3001\u53c2\u6570\u540d\u3001\u9884\u671f\u884c\u4e3a\u90fd\u6709\u51fa\u5904\uff0c\u6709\u4e00\u6761\u5bf9\u4e0d\u4e0a\u6211\u90fd\u4f1a\u5199\u51fa\u6765\u3002<\/p>\n<p>\u6700\u5c0f\u9a8c\u8bc1\u8def\u5f84\u7528 HF Trainer \u5185\u7f6e\u5f00\u5173\uff0c\u8fde patch API \u90fd\u4e0d\u7528\u81ea\u5df1\u8c03\uff1a<\/p>\n<pre><code class=\"language-python\">from transformers import AutoModelForCausalLM, TrainingArguments\n\nmodel = AutoModelForCausalLM.from_pretrained(&quot;Qwen\/Qwen3-8B&quot;)\n\nargs = TrainingArguments(\n    output_dir=&quot;.\/out&quot;,\n    use_liger_kernel=True,       # HF Trainer \u5185\u7f6e\u5f00\u5173\uff0cPR #32860 \u5f15\u5165\n    per_device_train_batch_size=8,\n)\n\n# \u6b63\u5e38\u7684 Trainer \u6d41\u7a0b\uff0c\u6570\u636e\u3001\u635f\u5931\u90fd\u4e0d\u7528\u6539\n<\/code><\/pre>\n<p>\u9884\u671f\u884c\u4e3a\uff1a\u8bad\u7ec3\u542f\u52a8\u65e5\u5fd7\u91cc Liger \u76f8\u5173 patch \u751f\u6548\uff0closs \u66f2\u7ebf\u4e0e\u4e0d\u5f00\u5f00\u5173\u65f6\u5bf9\u9f50\uff08\u5b98\u65b9\u505a\u8fc7\u6536\u655b\u6bd4\u5bf9\uff09\uff1b\u663e\u5b58\u5cf0\u503c\u663e\u8457\u4f4e\u4e8e\u4e0d\u5f00\u7684\u5bf9\u7167\u7ec4\uff0c4K \u4e0a\u4e0b\u6587\u7684 8B \u6a21\u578b\u4e0d\u518d\u8d77\u6b65 OOM\u3002\u5982\u679c\u60f3\u786e\u8ba4\u54ea\u4e9b\u7b97\u5b50\u771f\u7684\u88ab\u6362\u6389\u4e86\uff0c\u7528\u663e\u5f0f patch \u518d\u9a8c\u4e00\u6b21\uff1a<\/p>\n<pre><code class=\"language-python\">import transformers\nfrom liger_kernel.transformers import apply_liger_kernel_to_qwen3\n\napply_liger_kernel_to_qwen3()\nmodel = transformers.AutoModelForCausalLM.from_pretrained(&quot;Qwen\/Qwen3-8B&quot;)\n# patch \u53d1\u751f\u5728\u5b9e\u4f8b\u5316\u65f6\uff1bRMSNorm\/SwiGLU\/\u4ea4\u53c9\u71b5\u88ab\u66ff\u6362\u4e3a Liger \u5b9e\u73b0\n<\/code><\/pre>\n<p>\u8fd9\u4e00\u6b65\u6709\u4e2a\u5b98\u65b9\u6587\u6863\u91cc\u5199\u660e\u7684\u524d\u63d0\uff1apatching API \u8981 <code>transformers &gt;= 5.0.0<\/code>\uff0c\u6a21\u578b\u672c\u8eab\u4e5f\u5f97\u5728\u652f\u6301\u5217\u8868\u91cc\u3002\u5982\u679c\u4f60\u7528\u7684\u662f Axolotl \u6216 ms-swift \u8fd9\u7c7b\u6846\u67b6\uff0c\u4e8b\u60c5\u66f4\u7b80\u5355\uff1a\u4e24\u8fb9\u90fd\u662f\u540c\u4e00\u4e2a\u5f00\u5173\u540d\uff0cAxolotl \u914d\u7f6e\u5199 <code>use_liger_kernel: true<\/code>\uff0cms-swift \u547d\u4ee4\u884c\u6302 <code>--use_liger_kernel true<\/code>\uff0c\u6846\u67b6\u81ea\u5df1\u5904\u7406\u7248\u672c\u517c\u5bb9\uff0c\u8fd9\u4e5f\u662f\u6211\u63a8\u8350\u666e\u901a\u7528\u6237\u8d70\u7684\u8def\u2014\u2014\u522b\u81ea\u5df1 patch\uff0c\u8ba9\u6846\u67b6\u80cc\u9505\u3002ms-swift \u4ed3\u5e93\u91cc\u90a3\u4e2a\u5b98\u65b9\u793a\u4f8b\u811a\u672c\u8fde\u8d26\u90fd\u66ff\u4f60\u7b97\u597d\u4e86\uff1a4 \u5f20 A100 \u5168\u53c2\u5fae\u8c03 Qwen2.5-7B \u52a0 packing\uff0c\u5f00 Liger \u6bcf\u5361 42 GB\u30011 \u5c0f\u65f6 35 \u5206\uff0c\u4e0d\u5f00 54 GB\u30011 \u5c0f\u65f6 40 \u5206\u2014\u2014\u7701 12 GB \u663e\u5b58\uff0c\u591a\u51fa\u6765\u7684\u663e\u5b58\u76f4\u63a5\u6362\u6210\u66f4\u5927\u7684 batch \u6216\u66f4\u957f\u7684\u5e8f\u5217\u3002\u6211\u6838\u4e0b\u6765\u7684\u552f\u4e00\u4fdd\u7559\u610f\u89c1\u662f\u300c\u9884\u671f\u8f93\u51fa\u300d\u91cc\u7684\u663e\u5b58\u6570\u5b57\u2014\u2014\u5b83\u4f9d\u8d56\u4f60\u7684 batch\u3001\u5e8f\u5217\u957f\u5ea6\u548c\u662f\u5426\u5f00 gradient checkpointing\uff0c\u5b98\u65b9\u53e3\u5f84\u53ea\u5728\u7279\u5b9a\u914d\u7f6e\u4e0b\u6210\u7acb\uff0c\u522b\u62ff\u81ea\u5df1\u7684\u5355\u5361\u6570\u5b57\u53bb\u5bf9\u6807\u90a3 60%\uff0c\u5bf9\u4e0d\u4e0a\u4e0d\u4ee3\u8868\u88c5\u9519\u4e86\u3002<\/p>\n<h2 id=\"_2\">\u522b\u4eba\u8e29\u8fc7\u7684\u5751\uff0c\u6211\u6311\u4e86\u4e94\u4e2a\u6709\u51fa\u5904\u7684<\/h2>\n<p>\u9ad8\u9891 issue \u6211\u6309\u8bc4\u8bba\u6570\u7ffb\u4e86\uff0c\u5f00\u6e90 issue \u603b\u91cf 250 \u4e0a\u4e0b\uff08\u5f53\u524d open 251\uff0c\u542b\u5927\u91cf\u5df2\u5408\u5e76\u7684 RFC\uff09\uff0c\u4e0b\u9762\u4e94\u6761\u6700\u503c\u5f97\u63d0\u524d\u77e5\u9053\uff1a<\/p>\n<p><strong>\u4e00\uff0ctorch.compile \u6469\u64e6\u662f\u8001\u5927\u96be\u3002<\/strong> issue #174\uff0c2024 \u5e74 8 \u6708\u5f00\u3001\u81f3\u4eca open\u300124 \u6761\u8bc4\u8bba\uff0c\u73b0\u8c61\u662f\u53e0\u52a0 compile \u76f4\u63a5\u629b\u5f02\u5e38\u3002\u597d\u6d88\u606f\u662f 11 \u6708\u8fd8\u6709\u4e24\u4e2a fix \u5728\u8fd9\u4e2a\u65b9\u5411\u5408\u5e76\uff0c\u574f\u6d88\u606f\u662f\u300c\u8001\u5927\u96be\u300d\u4e09\u4e2a\u5b57\u5c31\u662f\u73b0\u72b6\uff0c\u7528\u524d\u5148\u5728\u81ea\u5df1\u7684 torch \u7248\u672c\u4e0a\u9a8c\u3002<\/p>\n<p><strong>\u4e8c\uff0c\u7279\u5b9a\u6a21\u578b loss \u4e0d\u6536\u655b\u3002<\/strong> issue #257\u300cLoss does not drop when using Liger Kernel at Qwen2.5\u300d\uff0c14 \u6761\u8bc4\u8bba\uff0c\u8fd8\u662f open\u3002\u867d\u7136\u5927\u6982\u7387\u662f\u914d\u7f6e\u7ec4\u5408\u95ee\u9898\uff0c\u4f46\u300closs \u4e0d\u964d\u300d\u8fd9\u79cd\u5751\u4e00\u65e6\u649e\u4e0a\uff0c\u4f60\u7b2c\u4e00\u53cd\u5e94\u6c38\u8fdc\u662f\u6000\u7591\u81ea\u5df1\u7684\u6570\u636e\u2014\u2014\u8bb0\u5f97\u5148\u5173 Liger \u5bf9\u7167\u8dd1\u4e00\u8f6e\uff0c\u6392\u9664\u5b83\u518d\u6392\u67e5\u522b\u7684\u3002<\/p>\n<p><strong>\u4e09\uff0cGRPO \u7684 grad_norm \u5dee\u4e86\u4e24\u4e2a\u6570\u91cf\u7ea7\u3002<\/strong> issue #1082\uff0c\u6807\u9898\u76f4\u63a5\u5199\u660e Liger \u7684 GRPO loss \u4ea7\u51fa\u7684 grad_norm \u6bd4 TRL \u975e Liger \u8def\u5f84\u5927 100 \u500d\u5de6\u53f3\uff0c\u5f52\u56e0\u6307\u5411\u7f3a vLLM \u91cd\u8981\u6027\u91c7\u6837\u4fee\u6b63\u7b49\u5dee\u5f02\u3002\u505a\u5f3a\u5316\u5bf9\u9f50\u7684\u4eba\u8981\u76ef\u4f4f\u8fd9\u6761\uff0c\u68af\u5ea6\u8303\u6570\u5931\u771f\u4f1a\u5f71\u54cd\u88c1\u526a\u7b56\u7565\u3002<\/p>\n<p><strong>\u56db\uff0c\u5c0f\u6a21\u578b\u5c0f batch \u51e0\u4e4e\u767d\u5f00\u3002<\/strong> \u6709\u7bc7 GitCode \u4e0a\u7684\u5b9e\u8df5\u6587\uff08Qwen2-0.5B \u5728 Colab T4 \u4e0a\u7684\u5bf9\u7167\uff09\u6d4b\u5f97\u5f88\u624e\u5b9e\uff1abatch 2\u3001\u5e8f\u5217 128 \u65f6\u5f00\u4e0d\u5f00 Liger \u663e\u5b58\u7eb9\u4e1d\u4e0d\u52a8\uff0c\u62c9\u5230 batch 8\u3001\u5e8f\u5217 512 \u540e\u5dee\u8ddd\u624d\u51fa\u6765\u2014\u2014\u4e0d\u5f00\u76f4\u63a5 OOM\uff0c\u5f00\u4e86\u7a33\u7a33\u8dd1\u5b8c\u3002\u5b83\u7701\u7684\u662f logits \u8fd9\u7c7b\u968f token \u6570\u7ebf\u6027\u6da8\u7684\u4e2d\u95f4\u91cf\uff0ctoken \u603b\u91cf\u4e0a\u4e0d\u53bb\u5c31\u65e0\u8d26\u53ef\u7701\u3002<\/p>\n<p><strong>\u4e94\uff0c\u6b63\u9762\u6848\u4f8b\u4e5f\u662f\u771f\u7684\u3002<\/strong> Reddit \u8ba8\u8bba\u91cc\u6709\u7528\u6237\u62a5\u544a 4 \u5f20 3090 \u505a 4B \u5168\u53c2\u5fae\u8c03\uff0c\u603b\u65f6\u957f\u4ece 15 \u5c0f\u65f6\u964d\u5230 9.5 \u5c0f\u65f6\uff1b51CTO \u90a3\u7bc7 4090 QLoRA \u5b9e\u6d4b\u7684 +21% \u541e\u5410\u548c\u524d\u6587\u5f15\u7528\u4e00\u81f4\u3002\u65b9\u5411\u5bf9\u3001\u573a\u666f\u5bf9\uff0c\u6536\u76ca\u662f\u771f\u91d1\u3002\u8865\u5145\u4e00\u6761\u517c\u5bb9\u6027\u65e7\u8d26\uff1aPEFT \u52a0 BitsAndBytes \u91cf\u5316\u4e0e Liger \u7684\u7ec4\u5408\u65e9\u671f\u649e\u8fc7\u517c\u5bb9\u95ee\u9898\uff08issue #235\uff0c\u5df2\u5173\uff09\uff0c\u73b0\u5728\u8d70 QLoRA \u524d\u8fd8\u662f\u5efa\u8bae\u62ff\u5c0f\u4efb\u52a1\u9a8c\u4e00\u904d\u7248\u672c\u7ec4\u5408\uff0c\u91cf\u5316\u5e93\u7684 patch \u987a\u5e8f\u548c\u7b97\u5b50\u66ff\u6362\u662f\u4e24\u5957 monkey-patch \u5728\u62a2\u540c\u4e00\u6279\u6a21\u5757\u3002<\/p>\n<p>\u516c\u5f00\u8ba8\u8bba\u6574\u4f53\u4e0d\u7b97\u55a7\u95f9\uff0cReddit \u548c HN \u4e0a\u4e3b\u8981\u662f\u53d1\u5e03\u671f\u7684\u8ba8\u8bba\uff08\u8fd8\u6709\u6761\u795e\u8bc4\u8bba\uff1a\u300cLinkedIn \u5c45\u7136\u641e\u5f00\u6e90 LLM \u8bad\u7ec3\u7ec4\u4ef6\uff0c\u8fd9\u53ef\u4e0d\u5728\u6211\u7684\u5bbe\u679c\u5361\u4e0a\u300d\uff09\uff0c\u4e2d\u6587\u5708\u7684\u6df1\u5ea6\u6587\u7ae0\u4ee5\u96c6\u6210\u5b9e\u8df5\u4e3a\u4e3b\u3002\u4e0a\u9762\u7684\u5751\u5355\u57fa\u672c\u5c31\u662f\u793e\u533a\u53cd\u9988\u7684\u5168\u90e8\u91cd\u5fc3\u4e86\u3002<\/p>\n<h2 id=\"lora\">\u6211\u7684\u5224\u65ad\uff1a\u5168\u53c2\u548c\u5bf9\u9f50\u8bad\u7ec3\u7684\u663e\u5b58\u6551\u5175\uff0c\u5355\u5361 LoRA \u4e0d\u5fc5\u6fc0\u52a8<\/h2>\n<p>\u503c\u4e0d\u503c\u5f97\u4e0a\u624b\uff0c\u6211\u7684\u7ed3\u8bba\u5206\u4e24\u534a\u3002<\/p>\n<p><strong>\u503c\u5f97<\/strong>\uff1a\u4f60\u5728\u505a\u591a\u5361\u5168\u53c2\u5fae\u8c03\u3001\u957f\u4e0a\u4e0b\u6587\uff084K \u5f80\u4e0a\uff09\u3001\u6216 DPO\/GRPO \u8fd9\u7c7b\u5bf9\u9f50\u8bad\u7ec3\uff0c\u663e\u5b58\u662f\u5f53\u524d\u74f6\u9888\uff0c\u90a3\u4e48 Liger-Kernel \u662f\u76ee\u524d\u6469\u64e6\u6700\u5c0f\u7684\u65b9\u6848\u2014\u2014\u4e00\u884c\u5f00\u5173\uff0c\u6570\u503c\u6709\u6536\u655b\u6d4b\u8bd5\u80cc\u4e66\uff0cAxolotl\u3001LLaMA-Factory\u3001ms-swift \u5168\u90fd\u5185\u7f6e\u4e86\u3002\u8fd9\u51e0\u4e2a\u6846\u67b6\u6211\u90fd\u6478\u8fc7\uff0c\u5b83\u4eec\u5f00 Liger \u7684\u59ff\u52bf\u57fa\u672c\u5c31\u662f\u914d\u7f6e\u6587\u4ef6\u91cc\u6539\u4e00\u4e2a\u5e03\u5c14\u503c\u3002<\/p>\n<p><strong>\u4e0d\u5fc5\u6fc0\u52a8<\/strong>\uff1a\u4f60\u7684\u4e3b\u6218\u573a\u662f 8 GB \u7ea7\u5355\u5361\u7684 LoRA\/QLoRA\uff0c\u5e8f\u5217\u77ed\u3001batch \u5c0f\u3002\u8fd9\u6761\u8d5b\u9053\u4e0a unsloth \u662f\u66f4\u5bf9\u53e3\u7684\u9009\u62e9\u2014\u2014\u5b83\u628a LoRA \u7b97\u5b50\u878d\u5408\u3001\u6570\u636e\u9884\u5904\u7406\u3001\u8bad\u7ec3\u6d41\u7a0b\u6253\u5305\u6210\u4e00\u7ad9\u5f0f\uff0c\u5355\u5361\u6536\u76ca\u8986\u76d6\u66f4\u5168\uff0c\u4ee3\u4ef7\u662f\u6d41\u7a0b\u88ab\u5b83\u63a5\u7ba1\u3001\u591a\u5361 FSDP \u573a\u666f\u4e0d\u662f\u5b83\u7684\u4e3b\u573a\u3002\u7b2c\u4e09\u4e2a\u53ef\u9009\u9879\u662f\u5e72\u8106\u53ea\u7528 torch.compile\uff1a\u4e0d\u7528\u5f15\u7b2c\u4e09\u65b9\u3001\u6ca1\u6709 monkey-patch \u6469\u64e6\uff0c\u4f46\u5b83\u7701\u7684\u662f\u53e6\u4e00\u79cd\u8d26\uff08\u7b97\u5b50\u7f16\u8bd1\u4f18\u5316\uff09\uff0c\u5bf9 logits \u7269\u5316\u8fd9\u4e2a\u5177\u4f53\u75db\u70b9\u65e0\u6548\u3002\u5bf9\u6bd4\u7ef4\u5ea6\u5c31\u4e09\u6761\uff1a\u5355\u5361\u8fd8\u662f\u591a\u5361\u3001LoRA \u8fd8\u662f\u5168\u53c2\u3001\u8981\u4e0d\u8981\u4fdd\u7559\u8bad\u7ec3\u6d41\u7a0b\u63a7\u5236\u6743\u3002\u5355\u5361 LoRA \u9009 unsloth\uff0c\u591a\u5361\u5168\u53c2\u9009 Liger\uff0c\u60f3\u8981\u96f6\u4f9d\u8d56\u5c31\u7528 compile \u52a0\u81ea\u5df1\u8c03 batch\u3002<\/p>\n<p>\u6700\u540e\u8bf4\u53e5\u5b9e\u5728\u7684\u3002\u8fd9\u4e2a\u9879\u76ee\u6211\u4ece\u663e\u5b58\u90a3\u7b14\u618b\u5c48\u7684\u8d26\u8ffd\u8fc7\u6765\uff0c\u672c\u6765\u4ee5\u4e3a\u662f\u4e2a\u5c0f\u5de5\u5177\uff0c\u7ffb\u5b8c\u53d1\u73b0\u5b83\u662f ms-swift\u3001Axolotl \u8fd9\u4e9b\u8bad\u7ec3\u6846\u67b6\u811a\u4e0b\u90a3\u5c42\u516c\u5171\u5730\u57fa\u2014\u2014\u987a\u624b\u62c9\u4e86\u4e0b\uff0c\u73b0\u5728\u662f 6,704 \u4e2a star\uff0cBSD-2 \u8bb8\u53ef\uff0c\u6700\u65b0\u7248 0.8.5 \u5728 10 \u6708 21 \u65e5\u53d1\u7684\uff0c\u6700\u8fd1\u56db\u5468 26 \u6b21\u63d0\u4ea4\uff0c\u8282\u594f\u4e0d\u7b97\u731b\u4f46\u6bcf\u4e00\u7b14\u90fd\u662f\u5b9e\u6253\u5b9e\u7684 fix\u3002\u6211\u81ea\u5df1\u7684 Mac \u88c5\u4e0d\u4e0a\u5b83\uff0c\u53cd\u5012\u628a\u300c\u5e73\u53f0\u9501\u5b9a\u300d\u8fd9\u4e2a\u4ee3\u4ef7\u9a8c\u5f97\u660e\u660e\u767d\u767d\uff1b\u7b49\u54ea\u5929\u516c\u53f8\u5728 Linux \u673a\u5668\u4e0a\u6392\u5230\u6211\u8dd1\u5168\u53c2\uff0c\u6211\u4f1a\u7b2c\u4e00\u65f6\u95f4\u628a <code>use_liger_kernel=True<\/code> \u6253\u5f00\uff0c\u7136\u540e\u518d\u6765\u8fd8\u8fd9\u7bc7\u6b20\u4e0b\u7684\u5b9e\u6d4b\u8d26\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"<p>LinkedIn \u5f00\u6e90\u7684 Liger-Kernel \u7528\u5206\u5757\u7b97\u5b50\u628a\u5fae\u8c03\u65f6\u6700\u5360\u663e\u5b58\u7684 logits \u5c42\u5e72\u6389\uff0c\u5b98\u65b9\u53e3\u5f84\u541e\u5410\u52a0\u4e24\u6210\u3001\u663e\u5b58\u7701\u516d\u6210\u3002\u6211\u5728 Mac \u4e0a\u8fde\u88c5\u90fd\u6ca1\u88c5\u6210\uff0c\u4f46\u628a\u5b83\u7684\u4e09\u6863\u7528\u6cd5\u3001\u7f16\u8bd1\u6469\u64e6\u3001GRPO \u68af\u5ea6\u5931\u771f\u8fd9\u4e9b\u5751\u548c\u66ff\u4ee3\u65b9\u6848\u5148\u7b97\u4e86\u4e00\u904d\u8d26\u3002<\/p>\n","protected":false},"author":1,"featured_media":10021565,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"_jetpack_memberships_contains_paid_content":false,"footnotes":""},"categories":[4],"tags":[],"class_list":["post-10021566","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-share"],"jetpack_sharing_enabled":true,"jetpack_featured_media_url":"https:\/\/itmax.cn\/wp-content\/uploads\/2026\/11\/liger-kernel-cover.png","_links":{"self":[{"href":"https:\/\/itmax.cn\/index.php?rest_route=\/wp\/v2\/posts\/10021566","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/itmax.cn\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/itmax.cn\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/itmax.cn\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/itmax.cn\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=10021566"}],"version-history":[{"count":1,"href":"https:\/\/itmax.cn\/index.php?rest_route=\/wp\/v2\/posts\/10021566\/revisions"}],"predecessor-version":[{"id":10021567,"href":"https:\/\/itmax.cn\/index.php?rest_route=\/wp\/v2\/posts\/10021566\/revisions\/10021567"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/itmax.cn\/index.php?rest_route=\/wp\/v2\/media\/10021565"}],"wp:attachment":[{"href":"https:\/\/itmax.cn\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=10021566"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/itmax.cn\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=10021566"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/itmax.cn\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=10021566"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}