tocmo0nlord/axolotl
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

Built site for gh-pages

Browse Source
This commit is contained in:
Quarto GHA Workflow Runner
2025-06-14 18:56:28 +00:00
parent 84db47f3c0
commit 5b66b8e86c
6 changed files with 393 additions and 339 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
search.json
View File

@@ -455,7 +455,7 @@
"href": "docs/custom_integrations.html#cut-cross-entropy",
"title": "Custom Integrations",
"section": "Cut Cross Entropy",
"text": "Cut Cross Entropy\nCut Cross Entropy (CCE) reduces VRAM usage through optimization on the cross-entropy operation during loss calculation.\nSee https://github.com/apple/ml-cross-entropy\n\nRequirements\n\nPyTorch 2.4.0 or higher\n\n\n\nInstallation\nRun the following command to install cut_cross_entropy[transformers] if you don’t have it already.\n\nIf you are in dev environment\n\npython scripts/cutcrossentropy_install.py | sh\n\nIf you are installing from pip\n\npip3 uninstall -y cut-cross-entropy && pip3 install \"cut-cross-entropy[transformers] @ git+https://github.com/apple/ml-cross-entropy.git@bad6f7b49c75fdec69471abb71b4cddd0f0c6438\"\n\n\nUsage\nplugins:\n - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin\n\n\nSupported Models\n\nllama\nllama4\nllama4_text\nmllama\nphi3\ngemma\ngemma2\ngemma3\ngemma3_text\nmistral\nmistral3\nqwen2\nqwen2_moe\nqwen2_vl\nqwen2_5_vl\nqwen3\nqwen3_moe\ncohere\ncohere2\nglm\nglm4\n\n\n\nCitation\n@article{wijmans2024cut,\n author = {Erik Wijmans and\n Brody Huval and\n Alexander Hertzberg and\n Vladlen Koltun and\n Philipp Kr\\\"ahenb\\\"uhl},\n title = {Cut Your Losses in Large-Vocabulary Language Models},\n journal = {arXiv},\n year = {2024},\n url = {https://arxiv.org/abs/2411.09009},\n}\nPlease see reference here",
"text": "Cut Cross Entropy\nCut Cross Entropy (CCE) reduces VRAM usage through optimization on the cross-entropy operation during loss calculation.\nSee https://github.com/apple/ml-cross-entropy\n\nRequirements\n\nPyTorch 2.4.0 or higher\n\n\n\nInstallation\nRun the following command to install cut_cross_entropy[transformers] if you don’t have it already.\n\nIf you are in dev environment\n\npython scripts/cutcrossentropy_install.py | sh\n\nIf you are installing from pip\n\npip3 uninstall -y cut-cross-entropy && pip3 install \"cut-cross-entropy[transformers] @ git+https://github.com/apple/ml-cross-entropy.git@bad6f7b49c75fdec69471abb71b4cddd0f0c6438\"\n\n\nUsage\nNOTE: If you are training a VLM model, please use older version of Axolotl as upstream has applied a major VLM refactor, and our patches have not been updated yet.\ngit checkout 787880215b3ab32ccaf81c1b2e9588c6f3e6e764\n\npip3 install --no-build-isolation -e .\nplugins:\n - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin\n\n\nSupported Models\n\nllama\nllama4\nllama4_text\nmllama\nphi3\ngemma\ngemma2\ngemma3\ngemma3_text\nmistral\nmistral3\nqwen2\nqwen2_moe\nqwen2_vl\nqwen2_5_vl\nqwen3\nqwen3_moe\ncohere\ncohere2\nglm\nglm4\n\n\n\nCitation\n@article{wijmans2024cut,\n author = {Erik Wijmans and\n Brody Huval and\n Alexander Hertzberg and\n Vladlen Koltun and\n Philipp Kr\\\"ahenb\\\"uhl},\n title = {Cut Your Losses in Large-Vocabulary Language Models},\n journal = {arXiv},\n year = {2024},\n url = {https://arxiv.org/abs/2411.09009},\n}\nPlease see reference here",
"crumbs": [
"Advanced Features",
"Custom Integrations"
@@ -1407,7 +1407,7 @@
"href": "docs/api/monkeypatch.lora_kernels.html",
"title": "monkeypatch.lora_kernels",
"section": "",
"text": "monkeypatch.lora_kernels\nModule for patching custom LoRA Triton kernels and torch.autograd functions.\n\n\n\n\n\nName\nDescription\n\n\n\n\nFakeMLP\nplaceholder MLP for triton patching\n\n\n\n\n\nmonkeypatch.lora_kernels.FakeMLP(gate_proj, up_proj, down_proj)\nplaceholder MLP for triton patching\n\n\n\n\n\n\n\nName\nDescription\n\n\n\n\napply_lora_kernel_patches\nApplies optimized Triton kernel patches to a PEFT model.\n\n\nget_attention_cls_from_config\nGet the appropriate attention class by inspecting the model config.\n\n\noriginal_apply_o\nOriginal implementation of output projection without optimizations.\n\n\noriginal_apply_qkv\nOriginal implementation of QKV projection without optimizations.\n\n\npatch_self_attn_lora\nGiven an axolotl config, this method patches the inferred attention class forward\n\n\n\n\n\nmonkeypatch.lora_kernels.apply_lora_kernel_patches(model, cfg)\nApplies optimized Triton kernel patches to a PEFT model.\nPatches a PEFT model with optimized implementations for MLP and attention\ncomputations. The optimizations include custom Triton kernels for activation\nfunctions and specialized autograd functions for LoRA computations.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nmodel\nPeftModelForCausalLM\nA PEFT model to be patched with optimized kernels.\nrequired\n\n\ncfg\nDictDefault\nDictionary mapping axolotl config keys to values.\nrequired\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\nPeftModelForCausalLM\nPeftModelForCausalLM\nThe patched model with optimized kernels.\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\nTypeError\nIf the provided model is not a PeftModelForCausalLM.\n\n\n\nNotImplementedError\nIf the model type is not supported.\n\n\n\nAssertionError\nIf multiple adapters are active (currently unsupported).\n\n\n\n\n\n\nThe optimizations require LoRA adapters with no dropout and no bias terms. The\nfunction will skip patching if these conditions aren’t met.\n\n\n\n\nmonkeypatch.lora_kernels.get_attention_cls_from_config(cfg)\nGet the appropriate attention class by inspecting the model config.\nUses dynamic import to support any model architecture that follows\nthe standard transformers naming convention.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ncfg\nDictDefault\nDictionary mapping axolotl config keys to values.\nrequired\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\nType[nn.Module]\nThe appropriate attention class for the model.\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\nValueError\nIf base_model not specified or attention class cannot be imported\n\n\n\nImportError\nIf the model module or attention class doesn’t exist\n\n\n\n\n\n\n\nmonkeypatch.lora_kernels.original_apply_o(self, hidden_states)\nOriginal implementation of output projection without optimizations.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nself\nnn.Module\nThe attention module instance.\nrequired\n\n\nhidden_states\ntorch.Tensor\nInput tensor of shape [batch_size, seq_len, hidden_dim]`.\nrequired\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\ntorch.Tensor\nThe output projection result.\n\n\n\n\n\n\n\nmonkeypatch.lora_kernels.original_apply_qkv(self, hidden_states)\nOriginal implementation of QKV projection without optimizations.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nself\nnn.Module\nThe attention module instance.\nrequired\n\n\nhidden_states\ntorch.Tensor\nInput tensor of shape [batch_size, seq_len, hidden_dim].\nrequired\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\ntuple[torch.Tensor, torch.Tensor, torch.Tensor]\nA tuple (query_states, key_states, value_states) containing the projected states for query, key, and value.\n\n\n\n\n\n\n\nmonkeypatch.lora_kernels.patch_self_attn_lora(cfg)\nGiven an axolotl config, this method patches the inferred attention class forward\npass with optimized LoRA implementations.\nIt modifies the attention class to use optimized QKV and output projections. The\noriginal implementation is preserved and can be restored if needed.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ncfg\nDictDefault\nDictionary mapping axolotl config keys to values.\nrequired\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\nAssertionError\nIf the required code blocks are not found in the attention implementation."
"text": "monkeypatch.lora_kernels\nModule for patching custom LoRA Triton kernels and torch.autograd functions.\n\n\n\n\n\nName\nDescription\n\n\n\n\nFakeMLP\nplaceholder MLP for triton patching\n\n\n\n\n\nmonkeypatch.lora_kernels.FakeMLP(gate_proj, up_proj, down_proj)\nplaceholder MLP for triton patching\n\n\n\n\n\n\n\nName\nDescription\n\n\n\n\napply_lora_kernel_patches\nApplies optimized Triton kernel patches to a PEFT model.\n\n\nget_attention_cls_from_config\nGet the appropriate attention class by inspecting the model config.\n\n\nget_layers\nGet the layers of the model. Handles text-only and multimodal models.\n\n\noriginal_apply_o\nOriginal implementation of output projection without optimizations.\n\n\noriginal_apply_qkv\nOriginal implementation of QKV projection without optimizations.\n\n\npatch_self_attn_lora\nGiven an axolotl config, this method patches the inferred attention class forward\n\n\n\n\n\nmonkeypatch.lora_kernels.apply_lora_kernel_patches(model, cfg)\nApplies optimized Triton kernel patches to a PEFT model.\nPatches a PEFT model with optimized implementations for MLP and attention\ncomputations. The optimizations include custom Triton kernels for activation\nfunctions and specialized autograd functions for LoRA computations.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nmodel\nPeftModelForCausalLM\nA PEFT model to be patched with optimized kernels.\nrequired\n\n\ncfg\nDictDefault\nDictionary mapping axolotl config keys to values.\nrequired\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\nPeftModelForCausalLM\nPeftModelForCausalLM\nThe patched model with optimized kernels.\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\nTypeError\nIf the provided model is not a PeftModelForCausalLM.\n\n\n\nNotImplementedError\nIf the model type is not supported.\n\n\n\nAssertionError\nIf multiple adapters are active (currently unsupported).\n\n\n\n\n\n\nThe optimizations require LoRA adapters with no dropout and no bias terms. The\nfunction will skip patching if these conditions aren’t met.\n\n\n\n\nmonkeypatch.lora_kernels.get_attention_cls_from_config(cfg)\nGet the appropriate attention class by inspecting the model config.\nUses dynamic import to support any model architecture that follows\nthe standard transformers naming convention.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ncfg\nDictDefault\nDictionary mapping axolotl config keys to values.\nrequired\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\nType[nn.Module]\nThe appropriate attention class for the model.\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\nValueError\nIf base_model not specified or attention class cannot be imported\n\n\n\nImportError\nIf the model module or attention class doesn’t exist\n\n\n\n\n\n\n\nmonkeypatch.lora_kernels.get_layers(model)\nGet the layers of the model. Handles text-only and multimodal models.\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nmodel\nPeftModelForCausalLM\nA PEFT model.\nrequired\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\nlist[nn.Module]\nA list of layers.\n\n\n\n\n\n\n\nmonkeypatch.lora_kernels.original_apply_o(self, hidden_states)\nOriginal implementation of output projection without optimizations.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nself\nnn.Module\nThe attention module instance.\nrequired\n\n\nhidden_states\ntorch.Tensor\nInput tensor of shape [batch_size, seq_len, hidden_dim]`.\nrequired\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\ntorch.Tensor\nThe output projection result.\n\n\n\n\n\n\n\nmonkeypatch.lora_kernels.original_apply_qkv(self, hidden_states)\nOriginal implementation of QKV projection without optimizations.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nself\nnn.Module\nThe attention module instance.\nrequired\n\n\nhidden_states\ntorch.Tensor\nInput tensor of shape [batch_size, seq_len, hidden_dim].\nrequired\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\ntuple[torch.Tensor, torch.Tensor, torch.Tensor]\nA tuple (query_states, key_states, value_states) containing the projected states for query, key, and value.\n\n\n\n\n\n\n\nmonkeypatch.lora_kernels.patch_self_attn_lora(cfg)\nGiven an axolotl config, this method patches the inferred attention class forward\npass with optimized LoRA implementations.\nIt modifies the attention class to use optimized QKV and output projections. The\noriginal implementation is preserved and can be restored if needed.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ncfg\nDictDefault\nDictionary mapping axolotl config keys to values.\nrequired\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\nAssertionError\nIf the required code blocks are not found in the attention implementation."
},
{
"objectID": "docs/api/monkeypatch.lora_kernels.html#classes",
@@ -1421,7 +1421,7 @@
"href": "docs/api/monkeypatch.lora_kernels.html#functions",
"title": "monkeypatch.lora_kernels",
"section": "",
"text": "Name\nDescription\n\n\n\n\napply_lora_kernel_patches\nApplies optimized Triton kernel patches to a PEFT model.\n\n\nget_attention_cls_from_config\nGet the appropriate attention class by inspecting the model config.\n\n\noriginal_apply_o\nOriginal implementation of output projection without optimizations.\n\n\noriginal_apply_qkv\nOriginal implementation of QKV projection without optimizations.\n\n\npatch_self_attn_lora\nGiven an axolotl config, this method patches the inferred attention class forward\n\n\n\n\n\nmonkeypatch.lora_kernels.apply_lora_kernel_patches(model, cfg)\nApplies optimized Triton kernel patches to a PEFT model.\nPatches a PEFT model with optimized implementations for MLP and attention\ncomputations. The optimizations include custom Triton kernels for activation\nfunctions and specialized autograd functions for LoRA computations.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nmodel\nPeftModelForCausalLM\nA PEFT model to be patched with optimized kernels.\nrequired\n\n\ncfg\nDictDefault\nDictionary mapping axolotl config keys to values.\nrequired\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\nPeftModelForCausalLM\nPeftModelForCausalLM\nThe patched model with optimized kernels.\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\nTypeError\nIf the provided model is not a PeftModelForCausalLM.\n\n\n\nNotImplementedError\nIf the model type is not supported.\n\n\n\nAssertionError\nIf multiple adapters are active (currently unsupported).\n\n\n\n\n\n\nThe optimizations require LoRA adapters with no dropout and no bias terms. The\nfunction will skip patching if these conditions aren’t met.\n\n\n\n\nmonkeypatch.lora_kernels.get_attention_cls_from_config(cfg)\nGet the appropriate attention class by inspecting the model config.\nUses dynamic import to support any model architecture that follows\nthe standard transformers naming convention.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ncfg\nDictDefault\nDictionary mapping axolotl config keys to values.\nrequired\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\nType[nn.Module]\nThe appropriate attention class for the model.\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\nValueError\nIf base_model not specified or attention class cannot be imported\n\n\n\nImportError\nIf the model module or attention class doesn’t exist\n\n\n\n\n\n\n\nmonkeypatch.lora_kernels.original_apply_o(self, hidden_states)\nOriginal implementation of output projection without optimizations.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nself\nnn.Module\nThe attention module instance.\nrequired\n\n\nhidden_states\ntorch.Tensor\nInput tensor of shape [batch_size, seq_len, hidden_dim]`.\nrequired\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\ntorch.Tensor\nThe output projection result.\n\n\n\n\n\n\n\nmonkeypatch.lora_kernels.original_apply_qkv(self, hidden_states)\nOriginal implementation of QKV projection without optimizations.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nself\nnn.Module\nThe attention module instance.\nrequired\n\n\nhidden_states\ntorch.Tensor\nInput tensor of shape [batch_size, seq_len, hidden_dim].\nrequired\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\ntuple[torch.Tensor, torch.Tensor, torch.Tensor]\nA tuple (query_states, key_states, value_states) containing the projected states for query, key, and value.\n\n\n\n\n\n\n\nmonkeypatch.lora_kernels.patch_self_attn_lora(cfg)\nGiven an axolotl config, this method patches the inferred attention class forward\npass with optimized LoRA implementations.\nIt modifies the attention class to use optimized QKV and output projections. The\noriginal implementation is preserved and can be restored if needed.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ncfg\nDictDefault\nDictionary mapping axolotl config keys to values.\nrequired\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\nAssertionError\nIf the required code blocks are not found in the attention implementation."
"text": "Name\nDescription\n\n\n\n\napply_lora_kernel_patches\nApplies optimized Triton kernel patches to a PEFT model.\n\n\nget_attention_cls_from_config\nGet the appropriate attention class by inspecting the model config.\n\n\nget_layers\nGet the layers of the model. Handles text-only and multimodal models.\n\n\noriginal_apply_o\nOriginal implementation of output projection without optimizations.\n\n\noriginal_apply_qkv\nOriginal implementation of QKV projection without optimizations.\n\n\npatch_self_attn_lora\nGiven an axolotl config, this method patches the inferred attention class forward\n\n\n\n\n\nmonkeypatch.lora_kernels.apply_lora_kernel_patches(model, cfg)\nApplies optimized Triton kernel patches to a PEFT model.\nPatches a PEFT model with optimized implementations for MLP and attention\ncomputations. The optimizations include custom Triton kernels for activation\nfunctions and specialized autograd functions for LoRA computations.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nmodel\nPeftModelForCausalLM\nA PEFT model to be patched with optimized kernels.\nrequired\n\n\ncfg\nDictDefault\nDictionary mapping axolotl config keys to values.\nrequired\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\nPeftModelForCausalLM\nPeftModelForCausalLM\nThe patched model with optimized kernels.\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\nTypeError\nIf the provided model is not a PeftModelForCausalLM.\n\n\n\nNotImplementedError\nIf the model type is not supported.\n\n\n\nAssertionError\nIf multiple adapters are active (currently unsupported).\n\n\n\n\n\n\nThe optimizations require LoRA adapters with no dropout and no bias terms. The\nfunction will skip patching if these conditions aren’t met.\n\n\n\n\nmonkeypatch.lora_kernels.get_attention_cls_from_config(cfg)\nGet the appropriate attention class by inspecting the model config.\nUses dynamic import to support any model architecture that follows\nthe standard transformers naming convention.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ncfg\nDictDefault\nDictionary mapping axolotl config keys to values.\nrequired\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\nType[nn.Module]\nThe appropriate attention class for the model.\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\nValueError\nIf base_model not specified or attention class cannot be imported\n\n\n\nImportError\nIf the model module or attention class doesn’t exist\n\n\n\n\n\n\n\nmonkeypatch.lora_kernels.get_layers(model)\nGet the layers of the model. Handles text-only and multimodal models.\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nmodel\nPeftModelForCausalLM\nA PEFT model.\nrequired\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\nlist[nn.Module]\nA list of layers.\n\n\n\n\n\n\n\nmonkeypatch.lora_kernels.original_apply_o(self, hidden_states)\nOriginal implementation of output projection without optimizations.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nself\nnn.Module\nThe attention module instance.\nrequired\n\n\nhidden_states\ntorch.Tensor\nInput tensor of shape [batch_size, seq_len, hidden_dim]`.\nrequired\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\ntorch.Tensor\nThe output projection result.\n\n\n\n\n\n\n\nmonkeypatch.lora_kernels.original_apply_qkv(self, hidden_states)\nOriginal implementation of QKV projection without optimizations.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nself\nnn.Module\nThe attention module instance.\nrequired\n\n\nhidden_states\ntorch.Tensor\nInput tensor of shape [batch_size, seq_len, hidden_dim].\nrequired\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\ntuple[torch.Tensor, torch.Tensor, torch.Tensor]\nA tuple (query_states, key_states, value_states) containing the projected states for query, key, and value.\n\n\n\n\n\n\n\nmonkeypatch.lora_kernels.patch_self_attn_lora(cfg)\nGiven an axolotl config, this method patches the inferred attention class forward\npass with optimized LoRA implementations.\nIt modifies the attention class to use optimized QKV and output projections. The\noriginal implementation is preserved and can be restored if needed.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ncfg\nDictDefault\nDictionary mapping axolotl config keys to values.\nrequired\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\nAssertionError\nIf the required code blocks are not found in the attention implementation."
},
{
"objectID": "docs/api/integrations.kd.trainer.html",