diff --git a/.nojekyll b/.nojekyll index 50e4ddf8e..7052830ab 100644 --- a/.nojekyll +++ b/.nojekyll @@ -1 +1 @@ -d0e306c9 \ No newline at end of file +463de0cb \ No newline at end of file diff --git a/docs/api/kernels.lora.html b/docs/api/kernels.lora.html index e858b9c75..afeaeefac 100644 --- a/docs/api/kernels.lora.html +++ b/docs/api/kernels.lora.html @@ -639,11 +639,11 @@ gtag('config', 'G-9KYCVJBNMQ', { 'anonymize_ip': true}); None -- None for weights/quantization states +- None for weights/biases/quantization states -torch.Tensor | None +None - LoRA A/B matrix gradients (or None) @@ -653,7 +653,7 @@ gtag('config', 'G-9KYCVJBNMQ', { 'anonymize_ip': true}); -None +torch.Tensor | None - None for activation functions and flags @@ -666,24 +666,27 @@ gtag('config', 'G-9KYCVJBNMQ', { 'anonymize_ip': true}); ctx, X, gate_weight, - gate_quant, - gate_A, - gate_B, - gate_scale, - up_weight, - up_quant, - up_A, - up_B, - up_scale, - down_weight, - down_quant, - down_A, - down_B, - down_scale, - activation_fn, - activation_fn_backward, - inplace=True, -) + gate_bias, + gate_quant, + gate_A, + gate_B, + gate_scale, + up_weight, + up_bias, + up_quant, + up_A, + up_B, + up_scale, + down_weight, + down_bias, + down_quant, + down_A, + down_B, + down_scale, + activation_fn, + activation_fn_backward, + inplace=True, +)

Forward pass for LoRA MLP.

Parameters
@@ -722,87 +725,99 @@ gtag('config', 'G-9KYCVJBNMQ', { 'anonymize_ip': true}); required -gate_quant -object | None -Gate quantization state +gate_bias +torch.Tensor | None +Gate projection bias required +gate_quant +QuantState | None +Gate quantization state +required + + gate_A torch.Tensor | None Gate LoRA A matrix required - + gate_B torch.Tensor | None Gate LoRA B matrix required - + gate_scale float Gate LoRA scale required - + up_weight torch.Tensor -Up-projection weight -required - - -up_quant -object | None -Up-projection quantization state +Up projection weight required +up_quant +QuantState | None +Up projection quantization state +required + + up_A torch.Tensor | None -Up-projection LoRA A matrix -required - - -up_B -torch.Tensor | None -Up-projection LoRA B matrix +Up projection LoRA A matrix required -up_scale -float -Up-projection LoRA scale +up_B +torch.Tensor | None +Up projection LoRA B matrix required +up_scale +float +Up projection LoRA scale +required + + down_weight torch.Tensor -Down-projection weight +Down projection weight +required + + +down_bias +torch.Tensor | None +Down projection bias required down_quant -object | None -Down-projection quantization state +QuantState | None +Down projection quantization state required down_A torch.Tensor | None -Down-projection LoRA A matrix +Down projection LoRA A matrix required down_B torch.Tensor | None -Down-projection LoRA B matrix +Down projection LoRA B matrix required down_scale float -Down-projection LoRA scale +Down projection LoRA scale required @@ -919,8 +934,8 @@ gtag('config', 'G-9KYCVJBNMQ', { 'anonymize_ip': true}); --++ @@ -932,7 +947,7 @@ gtag('config', 'G-9KYCVJBNMQ', { 'anonymize_ip': true}); - + @@ -941,7 +956,7 @@ gtag('config', 'G-9KYCVJBNMQ', { 'anonymize_ip': true});
forward
-
kernels.lora.LoRA_O.forward(ctx, X, W, W_quant, A, B, S)
+
kernels.lora.LoRA_O.forward(ctx, X, W, b, W_quant, A, B, s)

Forward pass for output projection with LoRA.

Parameters
@@ -980,25 +995,31 @@ gtag('config', 'G-9KYCVJBNMQ', { 'anonymize_ip': true});
+ + + + + + - + - + - + - + - - + + @@ -1020,7 +1041,7 @@ gtag('config', 'G-9KYCVJBNMQ', { 'anonymize_ip': true}); - +
tuple[torch.Tensor, None, None, torch.Tensor | None, torch.Tensor | None, None]tuple[torch.Tensor, None, None, None, torch.Tensor, torch.Tensor, None] Tuple containing gradients for all forward inputs
required
btorch.TensorOutput projection biasrequired
W_quant QuantState | None Weight quantization state required
Atorch.Tensor | Nonetorch.Tensor LoRA A matrix required
Btorch.Tensor | Nonetorch.Tensor LoRA B matrix required
S
s float LoRA scaling factor required
torch.TensorOutput projection tensorOutput projection result
@@ -1108,8 +1129,8 @@ supporting quantization and memory optimization.

--++ @@ -1121,7 +1142,7 @@ supporting quantization and memory optimization.

- + @@ -1134,22 +1155,25 @@ supporting quantization and memory optimization.

ctx, X, q_weight, - q_quant, - q_A, - q_B, - q_scale, - k_weight, - k_quant, - k_A, - k_B, - k_scale, - v_weight, - v_quant, - v_A, - v_B, - v_scale, - inplace=True, -) + q_bias, + q_quant, + q_A, + q_B, + q_scale, + k_weight, + k_bias, + k_quant, + k_A, + k_B, + k_scale, + v_weight, + v_bias, + v_quant, + v_A, + v_B, + v_scale, + inplace=True, +)

Forward pass computing Q, K, V projections with LoRA.

Parameters
@@ -1188,35 +1212,47 @@ supporting quantization and memory optimization.

+ + + + + + - + - + - + - + + + + + + + @@ -1248,30 +1284,36 @@ supporting quantization and memory optimization.

+ + + + + + - + - + - + - + @@ -1625,17 +1667,17 @@ supporting quantization and memory optimization.

- + - - + + - - + +
tuple[torch.Tensor, None, None, torch.Tensor | None, torch.Tensor | None, None, None, None, torch.Tensor | None, torch.Tensor | None, None, None, None, torch.Tensor | None, torch.Tensor | None, None, None]tuple[torch.Tensor, None, None, None, torch.Tensor | None, torch.Tensor | None, None, None, None, None, torch.Tensor | None, torch.Tensor | None, None, None, None, None, torch.Tensor | None, torch.Tensor | None, None, None] Tuple containing gradients for all forward inputs
required
q_biastorch.Tensor | NoneQuery projection biasrequired
q_quant QuantState | None Query quantization state required
q_A torch.Tensor | None Query LoRA A matrix required
q_B torch.Tensor | None Query LoRA B matrix required
q_scale float Query LoRA scale required
k_weight torch.Tensor Key projection weight required
k_biastorch.Tensor | NoneKey projection biasrequired
k_quant QuantState | None required
v_biastorch.Tensor | NoneValue projection biasrequired
v_quant QuantState | None Value quantization state required
v_A torch.Tensor | None Value LoRA A matrix required
v_B torch.Tensor | None Value LoRA B matrix required
v_scale float Value LoRA scale required
inplace bool Whether to perform operations in-place
torch.TensorA tuple containing the base weight matrix, quantization state, LoRA A matrix,A tuple containing the base weights, quantization state, LoRA A and B weights,
QuantState | NoneLoRA B matrix, and scaling factor. States and matrices may be None if nottorch.Tensor | Nonescaling factor, and base layer bias. Quant state, weights, and bias may be
torch.Tensor | Noneavailable.QuantState | NoneNone if not available.
@@ -1643,7 +1685,7 @@ supporting quantization and memory optimization.

matmul_lora

-
kernels.lora.matmul_lora(X, W, W_quant, A, B, s, out=None)
+
kernels.lora.matmul_lora(X, W, b, W_quant, A, B, s, out=None)

Efficient fused matmul + LoRA computation.

Parameters

@@ -1677,25 +1719,25 @@ supporting quantization and memory optimization.

W_quant -QuantState +QuantState | None Quantization state for W required A -torch.Tensor +torch.Tensor | None LoRA A matrix [rank, in_features] required B -torch.Tensor +torch.Tensor | None LoRA B matrix [out_features, rank] required s -float +float | None LoRA scaling factor required diff --git a/search.json b/search.json index 5e8dd582c..1b1fa6afc 100644 --- a/search.json +++ b/search.json @@ -3424,21 +3424,21 @@ "href": "docs/api/kernels.lora.html", "title": "kernels.lora", "section": "", - "text": "kernels.lora\nModule for definition of Low-Rank Adaptation (LoRA) Triton kernels.\nSee “LoRA: Low-Rank Adaptation of Large Language Models”\n(https://arxiv.org/abs/2106.09685).\nCredit to unsloth (https://unsloth.ai/) for inspiration for this implementation.\n\n\n\n\n\nName\nDescription\n\n\n\n\nLoRA_MLP\nOptimized LoRA MLP implementation.\n\n\nLoRA_O\nOptimized LoRA implementation for output projection.\n\n\nLoRA_QKV\nOptimized LoRA QKV implementation with quantization support.\n\n\n\n\n\nkernels.lora.LoRA_MLP()\nOptimized LoRA MLP implementation.\n\n\n\n\n\nName\nDescription\n\n\n\n\nbackward\nPerforms backward pass computation for LoRA MLP.\n\n\nforward\nForward pass for LoRA MLP.\n\n\n\n\n\nkernels.lora.LoRA_MLP.backward(ctx, grad_output)\nPerforms backward pass computation for LoRA MLP.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nctx\ntorch.autograd.function.FunctionCtx\nContext object storing tensors saved during forward pass\nrequired\n\n\ngrad_output\ntorch.Tensor\nGradient of loss with respect to layer output\nrequired\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\ntorch.Tensor | None\nTuple containing gradients for all inputs from forward pass:\n\n\n\nNone\n- Input gradient tensor (or None)\n\n\n\nNone\n- None for weights/quantization states\n\n\n\ntorch.Tensor | None\n- LoRA A/B matrix gradients (or None)\n\n\n\ntorch.Tensor | None\n- None for scaling factors\n\n\n\nNone\n- None for activation functions and flags\n\n\n\n\n\n\n\nkernels.lora.LoRA_MLP.forward(\n ctx,\n X,\n gate_weight,\n gate_quant,\n gate_A,\n gate_B,\n gate_scale,\n up_weight,\n up_quant,\n up_A,\n up_B,\n up_scale,\n down_weight,\n down_quant,\n down_A,\n down_B,\n down_scale,\n activation_fn,\n activation_fn_backward,\n inplace=True,\n)\nForward pass for LoRA MLP.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nctx\n\nAutograd context\nrequired\n\n\nX\ntorch.Tensor\nInput features\nrequired\n\n\ngate_weight\ntorch.Tensor\nGate projection weight\nrequired\n\n\ngate_quant\nobject | None\nGate quantization state\nrequired\n\n\ngate_A\ntorch.Tensor | None\nGate LoRA A matrix\nrequired\n\n\ngate_B\ntorch.Tensor | None\nGate LoRA B matrix\nrequired\n\n\ngate_scale\nfloat\nGate LoRA scale\nrequired\n\n\nup_weight\ntorch.Tensor\nUp-projection weight\nrequired\n\n\nup_quant\nobject | None\nUp-projection quantization state\nrequired\n\n\nup_A\ntorch.Tensor | None\nUp-projection LoRA A matrix\nrequired\n\n\nup_B\ntorch.Tensor | None\nUp-projection LoRA B matrix\nrequired\n\n\nup_scale\nfloat\nUp-projection LoRA scale\nrequired\n\n\ndown_weight\ntorch.Tensor\nDown-projection weight\nrequired\n\n\ndown_quant\nobject | None\nDown-projection quantization state\nrequired\n\n\ndown_A\ntorch.Tensor | None\nDown-projection LoRA A matrix\nrequired\n\n\ndown_B\ntorch.Tensor | None\nDown-projection LoRA B matrix\nrequired\n\n\ndown_scale\nfloat\nDown-projection LoRA scale\nrequired\n\n\nactivation_fn\nCallable\nForward activation function\nrequired\n\n\nactivation_fn_backward\nCallable\nBackward activation function\nrequired\n\n\ninplace\nbool | None\nWhether to perform operations in-place\nTrue\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\ntorch.Tensor\nOutput transformed by multi-layer perceptron and activation function\n\n\n\n\n\n\n\n\n\nkernels.lora.LoRA_O()\nOptimized LoRA implementation for output projection.\n\n\n\n\n\nName\nDescription\n\n\n\n\nbackward\nBackward pass computing gradients for LoRA output projection.\n\n\nforward\nForward pass for output projection with LoRA.\n\n\n\n\n\nkernels.lora.LoRA_O.backward(ctx, dY)\nBackward pass computing gradients for LoRA output projection.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nctx\ntorch.autograd.function.FunctionCtx\nAutograd context\nrequired\n\n\ndY\ntorch.Tensor\nGradient of loss with respect to output\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, None, None, torch.Tensor | None, torch.Tensor | None, None]\nTuple containing gradients for all forward inputs\n\n\n\n\n\n\n\nkernels.lora.LoRA_O.forward(ctx, X, W, W_quant, A, B, S)\nForward pass for output projection with LoRA.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nctx\ntorch.autograd.function.FunctionCtx\nAutograd context\nrequired\n\n\nX\ntorch.Tensor\nInput tensor\nrequired\n\n\nW\ntorch.Tensor\nOutput projection weight\nrequired\n\n\nW_quant\nQuantState | None\nWeight quantization state\nrequired\n\n\nA\ntorch.Tensor | None\nLoRA A matrix\nrequired\n\n\nB\ntorch.Tensor | None\nLoRA B matrix\nrequired\n\n\nS\nfloat\nLoRA scaling factor\nrequired\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\ntorch.Tensor\nOutput projection tensor\n\n\n\n\n\n\n\n\n\nkernels.lora.LoRA_QKV()\nOptimized LoRA QKV implementation with quantization support.\nImplements efficient computation of query, key, value projections with LoRA,\nsupporting quantization and memory optimization.\n\n\n\n\n\nName\nDescription\n\n\n\n\nbackward\nBackward pass computing gradients for LoRA QKV.\n\n\nforward\nForward pass computing Q, K, V projections with LoRA.\n\n\n\n\n\nkernels.lora.LoRA_QKV.backward(ctx, q_grad, k_grad, v_grad)\nBackward pass computing gradients for LoRA QKV.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nctx\ntorch.autograd.function.FunctionCtx\nAutograd context\nrequired\n\n\nq_grad\ntorch.Tensor\nGradient for query projection\nrequired\n\n\nk_grad\ntorch.Tensor\nGradient for key projection\nrequired\n\n\nv_grad\ntorch.Tensor\nGradient for value projection\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, None, None, torch.Tensor | None, torch.Tensor | None, None, None, None, torch.Tensor | None, torch.Tensor | None, None, None, None, torch.Tensor | None, torch.Tensor | None, None, None]\nTuple containing gradients for all forward inputs\n\n\n\n\n\n\n\nkernels.lora.LoRA_QKV.forward(\n ctx,\n X,\n q_weight,\n q_quant,\n q_A,\n q_B,\n q_scale,\n k_weight,\n k_quant,\n k_A,\n k_B,\n k_scale,\n v_weight,\n v_quant,\n v_A,\n v_B,\n v_scale,\n inplace=True,\n)\nForward pass computing Q, K, V projections with LoRA.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nctx\ntorch.autograd.function.FunctionCtx\nAutograd context\nrequired\n\n\nX\ntorch.Tensor\nInput tensor\nrequired\n\n\nq_weight\ntorch.Tensor\nQuery projection weight\nrequired\n\n\nq_quant\nQuantState | None\nQuery quantization state\nrequired\n\n\nq_A\ntorch.Tensor | None\nQuery LoRA A matrix\nrequired\n\n\nq_B\ntorch.Tensor | None\nQuery LoRA B matrix\nrequired\n\n\nq_scale\nfloat\nQuery LoRA scale\nrequired\n\n\nk_weight\ntorch.Tensor\nKey projection weight\nrequired\n\n\nk_quant\nQuantState | None\nKey quantization state\nrequired\n\n\nk_A\ntorch.Tensor | None\nKey LoRA A matrix\nrequired\n\n\nk_B\ntorch.Tensor | None\nKey LoRA B matrix\nrequired\n\n\nk_scale\nfloat\nKey LoRA scale\nrequired\n\n\nv_weight\ntorch.Tensor\nValue projection weight\nrequired\n\n\nv_quant\nQuantState | None\nValue quantization state\nrequired\n\n\nv_A\ntorch.Tensor | None\nValue LoRA A matrix\nrequired\n\n\nv_B\ntorch.Tensor | None\nValue LoRA B matrix\nrequired\n\n\nv_scale\nfloat\nValue LoRA scale\nrequired\n\n\ninplace\nbool\nWhether to perform operations in-place\nTrue\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]\nTuple of (Query, Key, Value) projection tensors\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nDescription\n\n\n\n\napply_lora_mlp_geglu\nApplies LoRA to MLP layer with GEGLU activation.\n\n\napply_lora_mlp_swiglu\nApplies LoRA to MLP layer with SwiGLU activation.\n\n\napply_lora_o\nApplies LoRA to output projection layer.\n\n\napply_lora_qkv\nApplies LoRA to compute Query, Key, Value projections.\n\n\nget_lora_parameters\nGets LoRA parameters from a projection module.\n\n\nmatmul_lora\nEfficient fused matmul + LoRA computation.\n\n\n\n\n\nkernels.lora.apply_lora_mlp_geglu(self, X, inplace=True)\nApplies LoRA to MLP layer with GEGLU activation.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nX\ntorch.Tensor\nInput tensor for the MLP layer\nrequired\n\n\ninplace\nbool\nWhether to perform operations in-place to save memory\nTrue\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\ntorch.Tensor\nOutput tensor after applying LoRA-adapted MLP with GEGLU activation\n\n\n\n\n\n\n\nkernels.lora.apply_lora_mlp_swiglu(self, X, inplace=True)\nApplies LoRA to MLP layer with SwiGLU activation.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nX\ntorch.Tensor\nInput tensor for the MLP layer\nrequired\n\n\ninplace\nbool\nWhether to perform operations in-place to save memory\nTrue\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\ntorch.Tensor\nOutput tensor after applying LoRA-adapted MLP with SwiGLU activation\n\n\n\n\n\n\n\nkernels.lora.apply_lora_o(self, X)\nApplies LoRA to output projection layer.\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nX\ntorch.Tensor\nInput tensor\nrequired\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\ntorch.Tensor\nTransformed output tensor\n\n\n\n\n\n\n\nkernels.lora.apply_lora_qkv(self, X, inplace=True)\nApplies LoRA to compute Query, Key, Value projections.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nX\ntorch.Tensor\nInput tensor\nrequired\n\n\ninplace\nbool\nWhether to perform operations in-place\nTrue\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]\nTuple of (Query, Key, Value) projection tensors\n\n\n\n\n\n\n\nkernels.lora.get_lora_parameters(proj)\nGets LoRA parameters from a projection module.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nproj\nnn.Module\nThe projection module to extract parameters from.\nrequired\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\ntorch.Tensor\nA tuple containing the base weight matrix, quantization state, LoRA A matrix,\n\n\n\nQuantState | None\nLoRA B matrix, and scaling factor. States and matrices may be None if not\n\n\n\ntorch.Tensor | None\navailable.\n\n\n\n\n\n\n\nkernels.lora.matmul_lora(X, W, W_quant, A, B, s, out=None)\nEfficient fused matmul + LoRA computation.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nX\ntorch.Tensor\nInput tensor [*, in_features]\nrequired\n\n\nW\ntorch.Tensor\nBase weight matrix [out_features, in_features]\nrequired\n\n\nW_quant\nQuantState\nQuantization state for W\nrequired\n\n\nA\ntorch.Tensor\nLoRA A matrix [rank, in_features]\nrequired\n\n\nB\ntorch.Tensor\nLoRA B matrix [out_features, rank]\nrequired\n\n\ns\nfloat\nLoRA scaling factor\nrequired\n\n\nout\ntorch.Tensor | None\nOptional output tensor for inplace operations\nNone\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\ntorch.Tensor\nResult of X @ W + X @ A @ B" + "text": "kernels.lora\nModule for definition of Low-Rank Adaptation (LoRA) Triton kernels.\nSee “LoRA: Low-Rank Adaptation of Large Language Models”\n(https://arxiv.org/abs/2106.09685).\nCredit to unsloth (https://unsloth.ai/) for inspiration for this implementation.\n\n\n\n\n\nName\nDescription\n\n\n\n\nLoRA_MLP\nOptimized LoRA MLP implementation.\n\n\nLoRA_O\nOptimized LoRA implementation for output projection.\n\n\nLoRA_QKV\nOptimized LoRA QKV implementation with quantization support.\n\n\n\n\n\nkernels.lora.LoRA_MLP()\nOptimized LoRA MLP implementation.\n\n\n\n\n\nName\nDescription\n\n\n\n\nbackward\nPerforms backward pass computation for LoRA MLP.\n\n\nforward\nForward pass for LoRA MLP.\n\n\n\n\n\nkernels.lora.LoRA_MLP.backward(ctx, grad_output)\nPerforms backward pass computation for LoRA MLP.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nctx\ntorch.autograd.function.FunctionCtx\nContext object storing tensors saved during forward pass\nrequired\n\n\ngrad_output\ntorch.Tensor\nGradient of loss with respect to layer output\nrequired\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\ntorch.Tensor | None\nTuple containing gradients for all inputs from forward pass:\n\n\n\nNone\n- Input gradient tensor (or None)\n\n\n\nNone\n- None for weights/biases/quantization states\n\n\n\nNone\n- LoRA A/B matrix gradients (or None)\n\n\n\ntorch.Tensor | None\n- None for scaling factors\n\n\n\ntorch.Tensor | None\n- None for activation functions and flags\n\n\n\n\n\n\n\nkernels.lora.LoRA_MLP.forward(\n ctx,\n X,\n gate_weight,\n gate_bias,\n gate_quant,\n gate_A,\n gate_B,\n gate_scale,\n up_weight,\n up_bias,\n up_quant,\n up_A,\n up_B,\n up_scale,\n down_weight,\n down_bias,\n down_quant,\n down_A,\n down_B,\n down_scale,\n activation_fn,\n activation_fn_backward,\n inplace=True,\n)\nForward pass for LoRA MLP.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nctx\n\nAutograd context\nrequired\n\n\nX\ntorch.Tensor\nInput features\nrequired\n\n\ngate_weight\ntorch.Tensor\nGate projection weight\nrequired\n\n\ngate_bias\ntorch.Tensor | None\nGate projection bias\nrequired\n\n\ngate_quant\nQuantState | None\nGate quantization state\nrequired\n\n\ngate_A\ntorch.Tensor | None\nGate LoRA A matrix\nrequired\n\n\ngate_B\ntorch.Tensor | None\nGate LoRA B matrix\nrequired\n\n\ngate_scale\nfloat\nGate LoRA scale\nrequired\n\n\nup_weight\ntorch.Tensor\nUp projection weight\nrequired\n\n\nup_quant\nQuantState | None\nUp projection quantization state\nrequired\n\n\nup_A\ntorch.Tensor | None\nUp projection LoRA A matrix\nrequired\n\n\nup_B\ntorch.Tensor | None\nUp projection LoRA B matrix\nrequired\n\n\nup_scale\nfloat\nUp projection LoRA scale\nrequired\n\n\ndown_weight\ntorch.Tensor\nDown projection weight\nrequired\n\n\ndown_bias\ntorch.Tensor | None\nDown projection bias\nrequired\n\n\ndown_quant\nQuantState | None\nDown projection quantization state\nrequired\n\n\ndown_A\ntorch.Tensor | None\nDown projection LoRA A matrix\nrequired\n\n\ndown_B\ntorch.Tensor | None\nDown projection LoRA B matrix\nrequired\n\n\ndown_scale\nfloat\nDown projection LoRA scale\nrequired\n\n\nactivation_fn\nCallable\nForward activation function\nrequired\n\n\nactivation_fn_backward\nCallable\nBackward activation function\nrequired\n\n\ninplace\nbool | None\nWhether to perform operations in-place\nTrue\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\ntorch.Tensor\nOutput transformed by multi-layer perceptron and activation function\n\n\n\n\n\n\n\n\n\nkernels.lora.LoRA_O()\nOptimized LoRA implementation for output projection.\n\n\n\n\n\nName\nDescription\n\n\n\n\nbackward\nBackward pass computing gradients for LoRA output projection.\n\n\nforward\nForward pass for output projection with LoRA.\n\n\n\n\n\nkernels.lora.LoRA_O.backward(ctx, dY)\nBackward pass computing gradients for LoRA output projection.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nctx\ntorch.autograd.function.FunctionCtx\nAutograd context\nrequired\n\n\ndY\ntorch.Tensor\nGradient of loss with respect to output\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, None, None, None, torch.Tensor, torch.Tensor, None]\nTuple containing gradients for all forward inputs\n\n\n\n\n\n\n\nkernels.lora.LoRA_O.forward(ctx, X, W, b, W_quant, A, B, s)\nForward pass for output projection with LoRA.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nctx\ntorch.autograd.function.FunctionCtx\nAutograd context\nrequired\n\n\nX\ntorch.Tensor\nInput tensor\nrequired\n\n\nW\ntorch.Tensor\nOutput projection weight\nrequired\n\n\nb\ntorch.Tensor\nOutput projection bias\nrequired\n\n\nW_quant\nQuantState | None\nWeight quantization state\nrequired\n\n\nA\ntorch.Tensor\nLoRA A matrix\nrequired\n\n\nB\ntorch.Tensor\nLoRA B matrix\nrequired\n\n\ns\nfloat\nLoRA scaling factor\nrequired\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\ntorch.Tensor\nOutput projection result\n\n\n\n\n\n\n\n\n\nkernels.lora.LoRA_QKV()\nOptimized LoRA QKV implementation with quantization support.\nImplements efficient computation of query, key, value projections with LoRA,\nsupporting quantization and memory optimization.\n\n\n\n\n\nName\nDescription\n\n\n\n\nbackward\nBackward pass computing gradients for LoRA QKV.\n\n\nforward\nForward pass computing Q, K, V projections with LoRA.\n\n\n\n\n\nkernels.lora.LoRA_QKV.backward(ctx, q_grad, k_grad, v_grad)\nBackward pass computing gradients for LoRA QKV.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nctx\ntorch.autograd.function.FunctionCtx\nAutograd context\nrequired\n\n\nq_grad\ntorch.Tensor\nGradient for query projection\nrequired\n\n\nk_grad\ntorch.Tensor\nGradient for key projection\nrequired\n\n\nv_grad\ntorch.Tensor\nGradient for value projection\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, None, None, None, torch.Tensor | None, torch.Tensor | None, None, None, None, None, torch.Tensor | None, torch.Tensor | None, None, None, None, None, torch.Tensor | None, torch.Tensor | None, None, None]\nTuple containing gradients for all forward inputs\n\n\n\n\n\n\n\nkernels.lora.LoRA_QKV.forward(\n ctx,\n X,\n q_weight,\n q_bias,\n q_quant,\n q_A,\n q_B,\n q_scale,\n k_weight,\n k_bias,\n k_quant,\n k_A,\n k_B,\n k_scale,\n v_weight,\n v_bias,\n v_quant,\n v_A,\n v_B,\n v_scale,\n inplace=True,\n)\nForward pass computing Q, K, V projections with LoRA.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nctx\ntorch.autograd.function.FunctionCtx\nAutograd context\nrequired\n\n\nX\ntorch.Tensor\nInput tensor\nrequired\n\n\nq_weight\ntorch.Tensor\nQuery projection weight\nrequired\n\n\nq_bias\ntorch.Tensor | None\nQuery projection bias\nrequired\n\n\nq_quant\nQuantState | None\nQuery quantization state\nrequired\n\n\nq_A\ntorch.Tensor | None\nQuery LoRA A matrix\nrequired\n\n\nq_B\ntorch.Tensor | None\nQuery LoRA B matrix\nrequired\n\n\nq_scale\nfloat\nQuery LoRA scale\nrequired\n\n\nk_weight\ntorch.Tensor\nKey projection weight\nrequired\n\n\nk_bias\ntorch.Tensor | None\nKey projection bias\nrequired\n\n\nk_quant\nQuantState | None\nKey quantization state\nrequired\n\n\nk_A\ntorch.Tensor | None\nKey LoRA A matrix\nrequired\n\n\nk_B\ntorch.Tensor | None\nKey LoRA B matrix\nrequired\n\n\nk_scale\nfloat\nKey LoRA scale\nrequired\n\n\nv_weight\ntorch.Tensor\nValue projection weight\nrequired\n\n\nv_bias\ntorch.Tensor | None\nValue projection bias\nrequired\n\n\nv_quant\nQuantState | None\nValue quantization state\nrequired\n\n\nv_A\ntorch.Tensor | None\nValue LoRA A matrix\nrequired\n\n\nv_B\ntorch.Tensor | None\nValue LoRA B matrix\nrequired\n\n\nv_scale\nfloat\nValue LoRA scale\nrequired\n\n\ninplace\nbool\nWhether to perform operations in-place\nTrue\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]\nTuple of (Query, Key, Value) projection tensors\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nDescription\n\n\n\n\napply_lora_mlp_geglu\nApplies LoRA to MLP layer with GEGLU activation.\n\n\napply_lora_mlp_swiglu\nApplies LoRA to MLP layer with SwiGLU activation.\n\n\napply_lora_o\nApplies LoRA to output projection layer.\n\n\napply_lora_qkv\nApplies LoRA to compute Query, Key, Value projections.\n\n\nget_lora_parameters\nGets LoRA parameters from a projection module.\n\n\nmatmul_lora\nEfficient fused matmul + LoRA computation.\n\n\n\n\n\nkernels.lora.apply_lora_mlp_geglu(self, X, inplace=True)\nApplies LoRA to MLP layer with GEGLU activation.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nX\ntorch.Tensor\nInput tensor for the MLP layer\nrequired\n\n\ninplace\nbool\nWhether to perform operations in-place to save memory\nTrue\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\ntorch.Tensor\nOutput tensor after applying LoRA-adapted MLP with GEGLU activation\n\n\n\n\n\n\n\nkernels.lora.apply_lora_mlp_swiglu(self, X, inplace=True)\nApplies LoRA to MLP layer with SwiGLU activation.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nX\ntorch.Tensor\nInput tensor for the MLP layer\nrequired\n\n\ninplace\nbool\nWhether to perform operations in-place to save memory\nTrue\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\ntorch.Tensor\nOutput tensor after applying LoRA-adapted MLP with SwiGLU activation\n\n\n\n\n\n\n\nkernels.lora.apply_lora_o(self, X)\nApplies LoRA to output projection layer.\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nX\ntorch.Tensor\nInput tensor\nrequired\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\ntorch.Tensor\nTransformed output tensor\n\n\n\n\n\n\n\nkernels.lora.apply_lora_qkv(self, X, inplace=True)\nApplies LoRA to compute Query, Key, Value projections.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nX\ntorch.Tensor\nInput tensor\nrequired\n\n\ninplace\nbool\nWhether to perform operations in-place\nTrue\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]\nTuple of (Query, Key, Value) projection tensors\n\n\n\n\n\n\n\nkernels.lora.get_lora_parameters(proj)\nGets LoRA parameters from a projection module.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nproj\nnn.Module\nThe projection module to extract parameters from.\nrequired\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\ntorch.Tensor\nA tuple containing the base weights, quantization state, LoRA A and B weights,\n\n\n\ntorch.Tensor | None\nscaling factor, and base layer bias. Quant state, weights, and bias may be\n\n\n\nQuantState | None\nNone if not available.\n\n\n\n\n\n\n\nkernels.lora.matmul_lora(X, W, b, W_quant, A, B, s, out=None)\nEfficient fused matmul + LoRA computation.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nX\ntorch.Tensor\nInput tensor [*, in_features]\nrequired\n\n\nW\ntorch.Tensor\nBase weight matrix [out_features, in_features]\nrequired\n\n\nW_quant\nQuantState | None\nQuantization state for W\nrequired\n\n\nA\ntorch.Tensor | None\nLoRA A matrix [rank, in_features]\nrequired\n\n\nB\ntorch.Tensor | None\nLoRA B matrix [out_features, rank]\nrequired\n\n\ns\nfloat | None\nLoRA scaling factor\nrequired\n\n\nout\ntorch.Tensor | None\nOptional output tensor for inplace operations\nNone\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\ntorch.Tensor\nResult of X @ W + X @ A @ B" }, { "objectID": "docs/api/kernels.lora.html#classes", "href": "docs/api/kernels.lora.html#classes", "title": "kernels.lora", "section": "", - "text": "Name\nDescription\n\n\n\n\nLoRA_MLP\nOptimized LoRA MLP implementation.\n\n\nLoRA_O\nOptimized LoRA implementation for output projection.\n\n\nLoRA_QKV\nOptimized LoRA QKV implementation with quantization support.\n\n\n\n\n\nkernels.lora.LoRA_MLP()\nOptimized LoRA MLP implementation.\n\n\n\n\n\nName\nDescription\n\n\n\n\nbackward\nPerforms backward pass computation for LoRA MLP.\n\n\nforward\nForward pass for LoRA MLP.\n\n\n\n\n\nkernels.lora.LoRA_MLP.backward(ctx, grad_output)\nPerforms backward pass computation for LoRA MLP.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nctx\ntorch.autograd.function.FunctionCtx\nContext object storing tensors saved during forward pass\nrequired\n\n\ngrad_output\ntorch.Tensor\nGradient of loss with respect to layer output\nrequired\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\ntorch.Tensor | None\nTuple containing gradients for all inputs from forward pass:\n\n\n\nNone\n- Input gradient tensor (or None)\n\n\n\nNone\n- None for weights/quantization states\n\n\n\ntorch.Tensor | None\n- LoRA A/B matrix gradients (or None)\n\n\n\ntorch.Tensor | None\n- None for scaling factors\n\n\n\nNone\n- None for activation functions and flags\n\n\n\n\n\n\n\nkernels.lora.LoRA_MLP.forward(\n ctx,\n X,\n gate_weight,\n gate_quant,\n gate_A,\n gate_B,\n gate_scale,\n up_weight,\n up_quant,\n up_A,\n up_B,\n up_scale,\n down_weight,\n down_quant,\n down_A,\n down_B,\n down_scale,\n activation_fn,\n activation_fn_backward,\n inplace=True,\n)\nForward pass for LoRA MLP.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nctx\n\nAutograd context\nrequired\n\n\nX\ntorch.Tensor\nInput features\nrequired\n\n\ngate_weight\ntorch.Tensor\nGate projection weight\nrequired\n\n\ngate_quant\nobject | None\nGate quantization state\nrequired\n\n\ngate_A\ntorch.Tensor | None\nGate LoRA A matrix\nrequired\n\n\ngate_B\ntorch.Tensor | None\nGate LoRA B matrix\nrequired\n\n\ngate_scale\nfloat\nGate LoRA scale\nrequired\n\n\nup_weight\ntorch.Tensor\nUp-projection weight\nrequired\n\n\nup_quant\nobject | None\nUp-projection quantization state\nrequired\n\n\nup_A\ntorch.Tensor | None\nUp-projection LoRA A matrix\nrequired\n\n\nup_B\ntorch.Tensor | None\nUp-projection LoRA B matrix\nrequired\n\n\nup_scale\nfloat\nUp-projection LoRA scale\nrequired\n\n\ndown_weight\ntorch.Tensor\nDown-projection weight\nrequired\n\n\ndown_quant\nobject | None\nDown-projection quantization state\nrequired\n\n\ndown_A\ntorch.Tensor | None\nDown-projection LoRA A matrix\nrequired\n\n\ndown_B\ntorch.Tensor | None\nDown-projection LoRA B matrix\nrequired\n\n\ndown_scale\nfloat\nDown-projection LoRA scale\nrequired\n\n\nactivation_fn\nCallable\nForward activation function\nrequired\n\n\nactivation_fn_backward\nCallable\nBackward activation function\nrequired\n\n\ninplace\nbool | None\nWhether to perform operations in-place\nTrue\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\ntorch.Tensor\nOutput transformed by multi-layer perceptron and activation function\n\n\n\n\n\n\n\n\n\nkernels.lora.LoRA_O()\nOptimized LoRA implementation for output projection.\n\n\n\n\n\nName\nDescription\n\n\n\n\nbackward\nBackward pass computing gradients for LoRA output projection.\n\n\nforward\nForward pass for output projection with LoRA.\n\n\n\n\n\nkernels.lora.LoRA_O.backward(ctx, dY)\nBackward pass computing gradients for LoRA output projection.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nctx\ntorch.autograd.function.FunctionCtx\nAutograd context\nrequired\n\n\ndY\ntorch.Tensor\nGradient of loss with respect to output\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, None, None, torch.Tensor | None, torch.Tensor | None, None]\nTuple containing gradients for all forward inputs\n\n\n\n\n\n\n\nkernels.lora.LoRA_O.forward(ctx, X, W, W_quant, A, B, S)\nForward pass for output projection with LoRA.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nctx\ntorch.autograd.function.FunctionCtx\nAutograd context\nrequired\n\n\nX\ntorch.Tensor\nInput tensor\nrequired\n\n\nW\ntorch.Tensor\nOutput projection weight\nrequired\n\n\nW_quant\nQuantState | None\nWeight quantization state\nrequired\n\n\nA\ntorch.Tensor | None\nLoRA A matrix\nrequired\n\n\nB\ntorch.Tensor | None\nLoRA B matrix\nrequired\n\n\nS\nfloat\nLoRA scaling factor\nrequired\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\ntorch.Tensor\nOutput projection tensor\n\n\n\n\n\n\n\n\n\nkernels.lora.LoRA_QKV()\nOptimized LoRA QKV implementation with quantization support.\nImplements efficient computation of query, key, value projections with LoRA,\nsupporting quantization and memory optimization.\n\n\n\n\n\nName\nDescription\n\n\n\n\nbackward\nBackward pass computing gradients for LoRA QKV.\n\n\nforward\nForward pass computing Q, K, V projections with LoRA.\n\n\n\n\n\nkernels.lora.LoRA_QKV.backward(ctx, q_grad, k_grad, v_grad)\nBackward pass computing gradients for LoRA QKV.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nctx\ntorch.autograd.function.FunctionCtx\nAutograd context\nrequired\n\n\nq_grad\ntorch.Tensor\nGradient for query projection\nrequired\n\n\nk_grad\ntorch.Tensor\nGradient for key projection\nrequired\n\n\nv_grad\ntorch.Tensor\nGradient for value projection\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, None, None, torch.Tensor | None, torch.Tensor | None, None, None, None, torch.Tensor | None, torch.Tensor | None, None, None, None, torch.Tensor | None, torch.Tensor | None, None, None]\nTuple containing gradients for all forward inputs\n\n\n\n\n\n\n\nkernels.lora.LoRA_QKV.forward(\n ctx,\n X,\n q_weight,\n q_quant,\n q_A,\n q_B,\n q_scale,\n k_weight,\n k_quant,\n k_A,\n k_B,\n k_scale,\n v_weight,\n v_quant,\n v_A,\n v_B,\n v_scale,\n inplace=True,\n)\nForward pass computing Q, K, V projections with LoRA.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nctx\ntorch.autograd.function.FunctionCtx\nAutograd context\nrequired\n\n\nX\ntorch.Tensor\nInput tensor\nrequired\n\n\nq_weight\ntorch.Tensor\nQuery projection weight\nrequired\n\n\nq_quant\nQuantState | None\nQuery quantization state\nrequired\n\n\nq_A\ntorch.Tensor | None\nQuery LoRA A matrix\nrequired\n\n\nq_B\ntorch.Tensor | None\nQuery LoRA B matrix\nrequired\n\n\nq_scale\nfloat\nQuery LoRA scale\nrequired\n\n\nk_weight\ntorch.Tensor\nKey projection weight\nrequired\n\n\nk_quant\nQuantState | None\nKey quantization state\nrequired\n\n\nk_A\ntorch.Tensor | None\nKey LoRA A matrix\nrequired\n\n\nk_B\ntorch.Tensor | None\nKey LoRA B matrix\nrequired\n\n\nk_scale\nfloat\nKey LoRA scale\nrequired\n\n\nv_weight\ntorch.Tensor\nValue projection weight\nrequired\n\n\nv_quant\nQuantState | None\nValue quantization state\nrequired\n\n\nv_A\ntorch.Tensor | None\nValue LoRA A matrix\nrequired\n\n\nv_B\ntorch.Tensor | None\nValue LoRA B matrix\nrequired\n\n\nv_scale\nfloat\nValue LoRA scale\nrequired\n\n\ninplace\nbool\nWhether to perform operations in-place\nTrue\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]\nTuple of (Query, Key, Value) projection tensors" + "text": "Name\nDescription\n\n\n\n\nLoRA_MLP\nOptimized LoRA MLP implementation.\n\n\nLoRA_O\nOptimized LoRA implementation for output projection.\n\n\nLoRA_QKV\nOptimized LoRA QKV implementation with quantization support.\n\n\n\n\n\nkernels.lora.LoRA_MLP()\nOptimized LoRA MLP implementation.\n\n\n\n\n\nName\nDescription\n\n\n\n\nbackward\nPerforms backward pass computation for LoRA MLP.\n\n\nforward\nForward pass for LoRA MLP.\n\n\n\n\n\nkernels.lora.LoRA_MLP.backward(ctx, grad_output)\nPerforms backward pass computation for LoRA MLP.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nctx\ntorch.autograd.function.FunctionCtx\nContext object storing tensors saved during forward pass\nrequired\n\n\ngrad_output\ntorch.Tensor\nGradient of loss with respect to layer output\nrequired\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\ntorch.Tensor | None\nTuple containing gradients for all inputs from forward pass:\n\n\n\nNone\n- Input gradient tensor (or None)\n\n\n\nNone\n- None for weights/biases/quantization states\n\n\n\nNone\n- LoRA A/B matrix gradients (or None)\n\n\n\ntorch.Tensor | None\n- None for scaling factors\n\n\n\ntorch.Tensor | None\n- None for activation functions and flags\n\n\n\n\n\n\n\nkernels.lora.LoRA_MLP.forward(\n ctx,\n X,\n gate_weight,\n gate_bias,\n gate_quant,\n gate_A,\n gate_B,\n gate_scale,\n up_weight,\n up_bias,\n up_quant,\n up_A,\n up_B,\n up_scale,\n down_weight,\n down_bias,\n down_quant,\n down_A,\n down_B,\n down_scale,\n activation_fn,\n activation_fn_backward,\n inplace=True,\n)\nForward pass for LoRA MLP.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nctx\n\nAutograd context\nrequired\n\n\nX\ntorch.Tensor\nInput features\nrequired\n\n\ngate_weight\ntorch.Tensor\nGate projection weight\nrequired\n\n\ngate_bias\ntorch.Tensor | None\nGate projection bias\nrequired\n\n\ngate_quant\nQuantState | None\nGate quantization state\nrequired\n\n\ngate_A\ntorch.Tensor | None\nGate LoRA A matrix\nrequired\n\n\ngate_B\ntorch.Tensor | None\nGate LoRA B matrix\nrequired\n\n\ngate_scale\nfloat\nGate LoRA scale\nrequired\n\n\nup_weight\ntorch.Tensor\nUp projection weight\nrequired\n\n\nup_quant\nQuantState | None\nUp projection quantization state\nrequired\n\n\nup_A\ntorch.Tensor | None\nUp projection LoRA A matrix\nrequired\n\n\nup_B\ntorch.Tensor | None\nUp projection LoRA B matrix\nrequired\n\n\nup_scale\nfloat\nUp projection LoRA scale\nrequired\n\n\ndown_weight\ntorch.Tensor\nDown projection weight\nrequired\n\n\ndown_bias\ntorch.Tensor | None\nDown projection bias\nrequired\n\n\ndown_quant\nQuantState | None\nDown projection quantization state\nrequired\n\n\ndown_A\ntorch.Tensor | None\nDown projection LoRA A matrix\nrequired\n\n\ndown_B\ntorch.Tensor | None\nDown projection LoRA B matrix\nrequired\n\n\ndown_scale\nfloat\nDown projection LoRA scale\nrequired\n\n\nactivation_fn\nCallable\nForward activation function\nrequired\n\n\nactivation_fn_backward\nCallable\nBackward activation function\nrequired\n\n\ninplace\nbool | None\nWhether to perform operations in-place\nTrue\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\ntorch.Tensor\nOutput transformed by multi-layer perceptron and activation function\n\n\n\n\n\n\n\n\n\nkernels.lora.LoRA_O()\nOptimized LoRA implementation for output projection.\n\n\n\n\n\nName\nDescription\n\n\n\n\nbackward\nBackward pass computing gradients for LoRA output projection.\n\n\nforward\nForward pass for output projection with LoRA.\n\n\n\n\n\nkernels.lora.LoRA_O.backward(ctx, dY)\nBackward pass computing gradients for LoRA output projection.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nctx\ntorch.autograd.function.FunctionCtx\nAutograd context\nrequired\n\n\ndY\ntorch.Tensor\nGradient of loss with respect to output\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, None, None, None, torch.Tensor, torch.Tensor, None]\nTuple containing gradients for all forward inputs\n\n\n\n\n\n\n\nkernels.lora.LoRA_O.forward(ctx, X, W, b, W_quant, A, B, s)\nForward pass for output projection with LoRA.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nctx\ntorch.autograd.function.FunctionCtx\nAutograd context\nrequired\n\n\nX\ntorch.Tensor\nInput tensor\nrequired\n\n\nW\ntorch.Tensor\nOutput projection weight\nrequired\n\n\nb\ntorch.Tensor\nOutput projection bias\nrequired\n\n\nW_quant\nQuantState | None\nWeight quantization state\nrequired\n\n\nA\ntorch.Tensor\nLoRA A matrix\nrequired\n\n\nB\ntorch.Tensor\nLoRA B matrix\nrequired\n\n\ns\nfloat\nLoRA scaling factor\nrequired\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\ntorch.Tensor\nOutput projection result\n\n\n\n\n\n\n\n\n\nkernels.lora.LoRA_QKV()\nOptimized LoRA QKV implementation with quantization support.\nImplements efficient computation of query, key, value projections with LoRA,\nsupporting quantization and memory optimization.\n\n\n\n\n\nName\nDescription\n\n\n\n\nbackward\nBackward pass computing gradients for LoRA QKV.\n\n\nforward\nForward pass computing Q, K, V projections with LoRA.\n\n\n\n\n\nkernels.lora.LoRA_QKV.backward(ctx, q_grad, k_grad, v_grad)\nBackward pass computing gradients for LoRA QKV.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nctx\ntorch.autograd.function.FunctionCtx\nAutograd context\nrequired\n\n\nq_grad\ntorch.Tensor\nGradient for query projection\nrequired\n\n\nk_grad\ntorch.Tensor\nGradient for key projection\nrequired\n\n\nv_grad\ntorch.Tensor\nGradient for value projection\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, None, None, None, torch.Tensor | None, torch.Tensor | None, None, None, None, None, torch.Tensor | None, torch.Tensor | None, None, None, None, None, torch.Tensor | None, torch.Tensor | None, None, None]\nTuple containing gradients for all forward inputs\n\n\n\n\n\n\n\nkernels.lora.LoRA_QKV.forward(\n ctx,\n X,\n q_weight,\n q_bias,\n q_quant,\n q_A,\n q_B,\n q_scale,\n k_weight,\n k_bias,\n k_quant,\n k_A,\n k_B,\n k_scale,\n v_weight,\n v_bias,\n v_quant,\n v_A,\n v_B,\n v_scale,\n inplace=True,\n)\nForward pass computing Q, K, V projections with LoRA.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nctx\ntorch.autograd.function.FunctionCtx\nAutograd context\nrequired\n\n\nX\ntorch.Tensor\nInput tensor\nrequired\n\n\nq_weight\ntorch.Tensor\nQuery projection weight\nrequired\n\n\nq_bias\ntorch.Tensor | None\nQuery projection bias\nrequired\n\n\nq_quant\nQuantState | None\nQuery quantization state\nrequired\n\n\nq_A\ntorch.Tensor | None\nQuery LoRA A matrix\nrequired\n\n\nq_B\ntorch.Tensor | None\nQuery LoRA B matrix\nrequired\n\n\nq_scale\nfloat\nQuery LoRA scale\nrequired\n\n\nk_weight\ntorch.Tensor\nKey projection weight\nrequired\n\n\nk_bias\ntorch.Tensor | None\nKey projection bias\nrequired\n\n\nk_quant\nQuantState | None\nKey quantization state\nrequired\n\n\nk_A\ntorch.Tensor | None\nKey LoRA A matrix\nrequired\n\n\nk_B\ntorch.Tensor | None\nKey LoRA B matrix\nrequired\n\n\nk_scale\nfloat\nKey LoRA scale\nrequired\n\n\nv_weight\ntorch.Tensor\nValue projection weight\nrequired\n\n\nv_bias\ntorch.Tensor | None\nValue projection bias\nrequired\n\n\nv_quant\nQuantState | None\nValue quantization state\nrequired\n\n\nv_A\ntorch.Tensor | None\nValue LoRA A matrix\nrequired\n\n\nv_B\ntorch.Tensor | None\nValue LoRA B matrix\nrequired\n\n\nv_scale\nfloat\nValue LoRA scale\nrequired\n\n\ninplace\nbool\nWhether to perform operations in-place\nTrue\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]\nTuple of (Query, Key, Value) projection tensors" }, { "objectID": "docs/api/kernels.lora.html#functions", "href": "docs/api/kernels.lora.html#functions", "title": "kernels.lora", "section": "", - "text": "Name\nDescription\n\n\n\n\napply_lora_mlp_geglu\nApplies LoRA to MLP layer with GEGLU activation.\n\n\napply_lora_mlp_swiglu\nApplies LoRA to MLP layer with SwiGLU activation.\n\n\napply_lora_o\nApplies LoRA to output projection layer.\n\n\napply_lora_qkv\nApplies LoRA to compute Query, Key, Value projections.\n\n\nget_lora_parameters\nGets LoRA parameters from a projection module.\n\n\nmatmul_lora\nEfficient fused matmul + LoRA computation.\n\n\n\n\n\nkernels.lora.apply_lora_mlp_geglu(self, X, inplace=True)\nApplies LoRA to MLP layer with GEGLU activation.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nX\ntorch.Tensor\nInput tensor for the MLP layer\nrequired\n\n\ninplace\nbool\nWhether to perform operations in-place to save memory\nTrue\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\ntorch.Tensor\nOutput tensor after applying LoRA-adapted MLP with GEGLU activation\n\n\n\n\n\n\n\nkernels.lora.apply_lora_mlp_swiglu(self, X, inplace=True)\nApplies LoRA to MLP layer with SwiGLU activation.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nX\ntorch.Tensor\nInput tensor for the MLP layer\nrequired\n\n\ninplace\nbool\nWhether to perform operations in-place to save memory\nTrue\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\ntorch.Tensor\nOutput tensor after applying LoRA-adapted MLP with SwiGLU activation\n\n\n\n\n\n\n\nkernels.lora.apply_lora_o(self, X)\nApplies LoRA to output projection layer.\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nX\ntorch.Tensor\nInput tensor\nrequired\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\ntorch.Tensor\nTransformed output tensor\n\n\n\n\n\n\n\nkernels.lora.apply_lora_qkv(self, X, inplace=True)\nApplies LoRA to compute Query, Key, Value projections.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nX\ntorch.Tensor\nInput tensor\nrequired\n\n\ninplace\nbool\nWhether to perform operations in-place\nTrue\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]\nTuple of (Query, Key, Value) projection tensors\n\n\n\n\n\n\n\nkernels.lora.get_lora_parameters(proj)\nGets LoRA parameters from a projection module.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nproj\nnn.Module\nThe projection module to extract parameters from.\nrequired\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\ntorch.Tensor\nA tuple containing the base weight matrix, quantization state, LoRA A matrix,\n\n\n\nQuantState | None\nLoRA B matrix, and scaling factor. States and matrices may be None if not\n\n\n\ntorch.Tensor | None\navailable.\n\n\n\n\n\n\n\nkernels.lora.matmul_lora(X, W, W_quant, A, B, s, out=None)\nEfficient fused matmul + LoRA computation.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nX\ntorch.Tensor\nInput tensor [*, in_features]\nrequired\n\n\nW\ntorch.Tensor\nBase weight matrix [out_features, in_features]\nrequired\n\n\nW_quant\nQuantState\nQuantization state for W\nrequired\n\n\nA\ntorch.Tensor\nLoRA A matrix [rank, in_features]\nrequired\n\n\nB\ntorch.Tensor\nLoRA B matrix [out_features, rank]\nrequired\n\n\ns\nfloat\nLoRA scaling factor\nrequired\n\n\nout\ntorch.Tensor | None\nOptional output tensor for inplace operations\nNone\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\ntorch.Tensor\nResult of X @ W + X @ A @ B" + "text": "Name\nDescription\n\n\n\n\napply_lora_mlp_geglu\nApplies LoRA to MLP layer with GEGLU activation.\n\n\napply_lora_mlp_swiglu\nApplies LoRA to MLP layer with SwiGLU activation.\n\n\napply_lora_o\nApplies LoRA to output projection layer.\n\n\napply_lora_qkv\nApplies LoRA to compute Query, Key, Value projections.\n\n\nget_lora_parameters\nGets LoRA parameters from a projection module.\n\n\nmatmul_lora\nEfficient fused matmul + LoRA computation.\n\n\n\n\n\nkernels.lora.apply_lora_mlp_geglu(self, X, inplace=True)\nApplies LoRA to MLP layer with GEGLU activation.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nX\ntorch.Tensor\nInput tensor for the MLP layer\nrequired\n\n\ninplace\nbool\nWhether to perform operations in-place to save memory\nTrue\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\ntorch.Tensor\nOutput tensor after applying LoRA-adapted MLP with GEGLU activation\n\n\n\n\n\n\n\nkernels.lora.apply_lora_mlp_swiglu(self, X, inplace=True)\nApplies LoRA to MLP layer with SwiGLU activation.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nX\ntorch.Tensor\nInput tensor for the MLP layer\nrequired\n\n\ninplace\nbool\nWhether to perform operations in-place to save memory\nTrue\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\ntorch.Tensor\nOutput tensor after applying LoRA-adapted MLP with SwiGLU activation\n\n\n\n\n\n\n\nkernels.lora.apply_lora_o(self, X)\nApplies LoRA to output projection layer.\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nX\ntorch.Tensor\nInput tensor\nrequired\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\ntorch.Tensor\nTransformed output tensor\n\n\n\n\n\n\n\nkernels.lora.apply_lora_qkv(self, X, inplace=True)\nApplies LoRA to compute Query, Key, Value projections.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nX\ntorch.Tensor\nInput tensor\nrequired\n\n\ninplace\nbool\nWhether to perform operations in-place\nTrue\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]\nTuple of (Query, Key, Value) projection tensors\n\n\n\n\n\n\n\nkernels.lora.get_lora_parameters(proj)\nGets LoRA parameters from a projection module.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nproj\nnn.Module\nThe projection module to extract parameters from.\nrequired\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\ntorch.Tensor\nA tuple containing the base weights, quantization state, LoRA A and B weights,\n\n\n\ntorch.Tensor | None\nscaling factor, and base layer bias. Quant state, weights, and bias may be\n\n\n\nQuantState | None\nNone if not available.\n\n\n\n\n\n\n\nkernels.lora.matmul_lora(X, W, b, W_quant, A, B, s, out=None)\nEfficient fused matmul + LoRA computation.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nX\ntorch.Tensor\nInput tensor [*, in_features]\nrequired\n\n\nW\ntorch.Tensor\nBase weight matrix [out_features, in_features]\nrequired\n\n\nW_quant\nQuantState | None\nQuantization state for W\nrequired\n\n\nA\ntorch.Tensor | None\nLoRA A matrix [rank, in_features]\nrequired\n\n\nB\ntorch.Tensor | None\nLoRA B matrix [out_features, rank]\nrequired\n\n\ns\nfloat | None\nLoRA scaling factor\nrequired\n\n\nout\ntorch.Tensor | None\nOptional output tensor for inplace operations\nNone\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\ntorch.Tensor\nResult of X @ W + X @ A @ B" }, { "objectID": "docs/api/cli.vllm_serve.html", diff --git a/sitemap.xml b/sitemap.xml index 96f9fb626..5c5839075 100644 --- a/sitemap.xml +++ b/sitemap.xml @@ -2,794 +2,794 @@ https://docs.axolotl.ai/TODO.html - 2025-08-06T18:29:01.128Z + 2025-08-07T00:20:15.795Z https://docs.axolotl.ai/index.html - 2025-08-06T18:29:01.149Z + 2025-08-07T00:20:15.816Z https://docs.axolotl.ai/docs/debugging.html - 2025-08-06T18:29:01.130Z + 2025-08-07T00:20:15.797Z https://docs.axolotl.ai/docs/amd_hpc.html - 2025-08-06T18:29:01.129Z + 2025-08-07T00:20:15.796Z https://docs.axolotl.ai/docs/api/utils.callbacks.mlflow_.html - 2025-08-06T18:32:31.939Z + 2025-08-07T00:23:32.931Z https://docs.axolotl.ai/docs/api/monkeypatch.llama_expand_mask.html - 2025-08-06T18:32:31.354Z + 2025-08-07T00:23:32.357Z https://docs.axolotl.ai/docs/api/loaders.patch_manager.html - 2025-08-06T18:32:30.993Z + 2025-08-07T00:23:32.000Z https://docs.axolotl.ai/docs/api/core.chat.format.llama3x.html - 2025-08-06T18:32:30.668Z + 2025-08-07T00:23:31.687Z https://docs.axolotl.ai/docs/api/cli.train.html - 2025-08-06T18:32:30.727Z + 2025-08-07T00:23:31.743Z https://docs.axolotl.ai/docs/api/utils.callbacks.perplexity.html - 2025-08-06T18:32:31.930Z + 2025-08-07T00:23:32.922Z https://docs.axolotl.ai/docs/api/core.chat.messages.html - 2025-08-06T18:32:30.665Z + 2025-08-07T00:23:31.684Z https://docs.axolotl.ai/docs/api/utils.callbacks.lisa.html - 2025-08-06T18:32:31.935Z + 2025-08-07T00:23:32.928Z https://docs.axolotl.ai/docs/api/cli.merge_sharded_fsdp_weights.html - 2025-08-06T18:32:30.823Z + 2025-08-07T00:23:31.836Z https://docs.axolotl.ai/docs/api/monkeypatch.mixtral.html - 2025-08-06T18:32:31.414Z + 2025-08-07T00:23:32.415Z https://docs.axolotl.ai/docs/api/utils.chat_templates.html - 2025-08-06T18:32:31.452Z + 2025-08-07T00:23:32.452Z https://docs.axolotl.ai/docs/api/core.chat.format.shared.html - 2025-08-06T18:32:30.670Z + 2025-08-07T00:23:31.689Z https://docs.axolotl.ai/docs/api/core.trainers.mixins.optimizer.html - 2025-08-06T18:32:31.000Z + 2025-08-07T00:23:32.007Z https://docs.axolotl.ai/docs/api/utils.collators.mamba.html - 2025-08-06T18:32:31.878Z + 2025-08-07T00:23:32.871Z https://docs.axolotl.ai/docs/api/logging_config.html - 2025-08-06T18:32:30.613Z + 2025-08-07T00:23:31.633Z https://docs.axolotl.ai/docs/api/utils.collators.mm_chat.html - 2025-08-06T18:32:31.882Z + 2025-08-07T00:23:32.875Z https://docs.axolotl.ai/docs/api/prompt_strategies.completion.html - 2025-08-06T18:32:31.122Z + 2025-08-07T00:23:32.126Z https://docs.axolotl.ai/docs/api/kernels.utils.html - 2025-08-06T18:32:31.339Z + 2025-08-07T00:23:32.342Z https://docs.axolotl.ai/docs/api/prompt_strategies.dpo.chat_template.html - 2025-08-06T18:32:31.156Z + 2025-08-07T00:23:32.160Z https://docs.axolotl.ai/docs/api/kernels.swiglu.html - 2025-08-06T18:32:31.330Z + 2025-08-07T00:23:32.333Z https://docs.axolotl.ai/docs/api/common.const.html - 2025-08-06T18:32:31.838Z + 2025-08-07T00:23:32.831Z https://docs.axolotl.ai/docs/api/cli.cloud.base.html - 2025-08-06T18:32:30.847Z + 2025-08-07T00:23:31.859Z https://docs.axolotl.ai/docs/api/prompt_strategies.orpo.chat_template.html - 2025-08-06T18:32:31.220Z + 2025-08-07T00:23:32.222Z https://docs.axolotl.ai/docs/api/core.builders.rl.html - 2025-08-06T18:32:30.629Z + 2025-08-07T00:23:31.649Z https://docs.axolotl.ai/docs/api/utils.dict.html - 2025-08-06T18:32:31.544Z + 2025-08-07T00:23:32.543Z https://docs.axolotl.ai/docs/api/utils.schemas.integrations.html - 2025-08-06T18:32:31.657Z + 2025-08-07T00:23:32.654Z https://docs.axolotl.ai/docs/api/core.trainers.utils.html - 2025-08-06T18:32:30.957Z + 2025-08-07T00:23:31.966Z https://docs.axolotl.ai/docs/api/monkeypatch.trainer_fsdp_optim.html - 2025-08-06T18:32:31.403Z + 2025-08-07T00:23:32.405Z https://docs.axolotl.ai/docs/api/cli.evaluate.html - 2025-08-06T18:32:30.735Z + 2025-08-07T00:23:31.751Z https://docs.axolotl.ai/docs/api/core.builders.causal.html - 2025-08-06T18:32:30.625Z + 2025-08-07T00:23:31.644Z https://docs.axolotl.ai/docs/api/monkeypatch.multipack.html - 2025-08-06T18:32:31.349Z + 2025-08-07T00:23:32.352Z https://docs.axolotl.ai/docs/api/monkeypatch.llama_patch_multipack.html - 2025-08-06T18:32:31.394Z + 2025-08-07T00:23:32.396Z https://docs.axolotl.ai/docs/api/cli.delinearize_llama4.html - 2025-08-06T18:32:30.788Z + 2025-08-07T00:23:31.802Z https://docs.axolotl.ai/docs/api/utils.schemas.trl.html - 2025-08-06T18:32:31.640Z + 2025-08-07T00:23:32.637Z https://docs.axolotl.ai/docs/api/prompt_strategies.dpo.zephyr.html - 2025-08-06T18:32:31.178Z + 2025-08-07T00:23:32.181Z https://docs.axolotl.ai/docs/api/integrations.kd.trainer.html - 2025-08-06T18:32:31.825Z + 2025-08-07T00:23:32.818Z https://docs.axolotl.ai/docs/api/monkeypatch.gradient_checkpointing.offload_disk.html - 2025-08-06T18:32:31.443Z + 2025-08-07T00:23:32.444Z https://docs.axolotl.ai/docs/api/utils.optimizers.adopt.html - 2025-08-06T18:32:31.552Z + 2025-08-07T00:23:32.551Z https://docs.axolotl.ai/docs/api/monkeypatch.data.batch_dataset_fetcher.html - 2025-08-06T18:32:31.412Z + 2025-08-07T00:23:32.414Z https://docs.axolotl.ai/docs/api/cli.cloud.modal_.html - 2025-08-06T18:32:30.854Z + 2025-08-07T00:23:31.865Z https://docs.axolotl.ai/docs/api/prompt_strategies.alpaca_chat.html - 2025-08-06T18:32:31.082Z + 2025-08-07T00:23:32.087Z https://docs.axolotl.ai/docs/api/utils.freeze.html - 2025-08-06T18:32:31.474Z + 2025-08-07T00:23:32.474Z https://docs.axolotl.ai/docs/api/prompt_strategies.bradley_terry.llama3.html - 2025-08-06T18:32:31.224Z + 2025-08-07T00:23:32.225Z https://docs.axolotl.ai/docs/api/integrations.base.html - 2025-08-06T18:32:31.812Z + 2025-08-07T00:23:32.806Z https://docs.axolotl.ai/docs/api/monkeypatch.unsloth_.html - 2025-08-06T18:32:31.411Z + 2025-08-07T00:23:32.413Z https://docs.axolotl.ai/docs/api/prompt_strategies.kto.chatml.html - 2025-08-06T18:32:31.198Z + 2025-08-07T00:23:32.200Z https://docs.axolotl.ai/docs/api/cli.main.html - 2025-08-06T18:32:30.718Z + 2025-08-07T00:23:31.735Z https://docs.axolotl.ai/docs/api/common.datasets.html - 2025-08-06T18:32:31.853Z + 2025-08-07T00:23:32.846Z https://docs.axolotl.ai/docs/api/train.html - 2025-08-06T18:32:30.527Z + 2025-08-07T00:23:31.544Z https://docs.axolotl.ai/docs/api/utils.trainer.html - 2025-08-06T18:32:31.491Z + 2025-08-07T00:23:32.491Z https://docs.axolotl.ai/docs/api/prompt_strategies.llama2_chat.html - 2025-08-06T18:32:31.116Z + 2025-08-07T00:23:32.121Z https://docs.axolotl.ai/docs/api/index.html - 2025-08-06T18:32:30.464Z + 2025-08-07T00:23:31.482Z https://docs.axolotl.ai/docs/api/prompt_strategies.chat_template.html - 2025-08-06T18:32:31.068Z + 2025-08-07T00:23:32.073Z https://docs.axolotl.ai/docs/api/core.training_args.html - 2025-08-06T18:32:30.642Z + 2025-08-07T00:23:31.662Z https://docs.axolotl.ai/docs/api/kernels.quantize.html - 2025-08-06T18:32:31.337Z + 2025-08-07T00:23:32.340Z https://docs.axolotl.ai/docs/api/convert.html - 2025-08-06T18:32:30.562Z + 2025-08-07T00:23:31.579Z https://docs.axolotl.ai/docs/api/integrations.grokfast.optimizer.html - 2025-08-06T18:32:31.817Z + 2025-08-07T00:23:32.811Z https://docs.axolotl.ai/docs/api/prompt_strategies.stepwise_supervised.html - 2025-08-06T18:32:31.133Z + 2025-08-07T00:23:32.137Z https://docs.axolotl.ai/docs/api/utils.schemas.model.html - 2025-08-06T18:32:31.603Z + 2025-08-07T00:23:32.600Z https://docs.axolotl.ai/docs/api/utils.callbacks.qat.html - 2025-08-06T18:32:31.949Z + 2025-08-07T00:23:32.941Z https://docs.axolotl.ai/docs/api/loaders.constants.html - 2025-08-06T18:32:30.994Z + 2025-08-07T00:23:32.002Z https://docs.axolotl.ai/docs/api/cli.utils.sweeps.html - 2025-08-06T18:32:30.884Z + 2025-08-07T00:23:31.894Z https://docs.axolotl.ai/docs/api/prompt_strategies.dpo.llama3.html - 2025-08-06T18:32:31.167Z + 2025-08-07T00:23:32.170Z https://docs.axolotl.ai/docs/api/core.datasets.transforms.chat_builder.html - 2025-08-06T18:32:30.683Z + 2025-08-07T00:23:31.701Z https://docs.axolotl.ai/docs/api/cli.utils.fetch.html - 2025-08-06T18:32:30.872Z + 2025-08-07T00:23:31.883Z https://docs.axolotl.ai/docs/api/core.trainers.mamba.html - 2025-08-06T18:32:30.926Z + 2025-08-07T00:23:31.935Z https://docs.axolotl.ai/docs/api/utils.schemas.enums.html - 2025-08-06T18:32:31.668Z + 2025-08-07T00:23:32.664Z https://docs.axolotl.ai/docs/api/utils.callbacks.profiler.html - 2025-08-06T18:32:31.934Z + 2025-08-07T00:23:32.926Z https://docs.axolotl.ai/docs/api/prompt_strategies.metharme.html - 2025-08-06T18:32:31.140Z + 2025-08-07T00:23:32.144Z https://docs.axolotl.ai/docs/api/core.trainers.trl.html - 2025-08-06T18:32:30.920Z + 2025-08-07T00:23:31.930Z https://docs.axolotl.ai/docs/api/prompt_strategies.orcamini.html - 2025-08-06T18:32:31.144Z + 2025-08-07T00:23:32.147Z https://docs.axolotl.ai/docs/api/utils.samplers.multipack.html - 2025-08-06T18:32:31.924Z + 2025-08-07T00:23:32.916Z https://docs.axolotl.ai/docs/api/utils.schedulers.html - 2025-08-06T18:32:31.518Z + 2025-08-07T00:23:32.518Z https://docs.axolotl.ai/docs/api/core.trainers.grpo.trainer.html - 2025-08-06T18:32:30.943Z + 2025-08-07T00:23:31.953Z https://docs.axolotl.ai/docs/api/prompt_tokenizers.html - 2025-08-06T18:32:30.604Z + 2025-08-07T00:23:31.624Z https://docs.axolotl.ai/docs/config-reference.html - 2025-08-06T18:32:46.257Z + 2025-08-07T00:23:46.075Z https://docs.axolotl.ai/docs/multimodal.html - 2025-08-06T18:29:01.133Z + 2025-08-07T00:20:15.800Z https://docs.axolotl.ai/docs/mixed_precision.html - 2025-08-06T18:29:01.133Z + 2025-08-07T00:20:15.800Z https://docs.axolotl.ai/docs/unsloth.html - 2025-08-06T18:29:01.134Z + 2025-08-07T00:20:15.801Z https://docs.axolotl.ai/docs/ray-integration.html - 2025-08-06T18:29:01.134Z + 2025-08-07T00:20:15.801Z https://docs.axolotl.ai/docs/dataset-formats/stepwise_supervised.html - 2025-08-06T18:29:01.130Z + 2025-08-07T00:20:15.797Z https://docs.axolotl.ai/docs/dataset-formats/template_free.html - 2025-08-06T18:29:01.130Z + 2025-08-07T00:20:15.797Z https://docs.axolotl.ai/docs/dataset-formats/index.html - 2025-08-06T18:29:01.130Z + 2025-08-07T00:20:15.797Z https://docs.axolotl.ai/docs/dataset-formats/pretraining.html - 2025-08-06T18:29:01.130Z + 2025-08-07T00:20:15.797Z https://docs.axolotl.ai/docs/nd_parallelism.html - 2025-08-06T18:29:01.133Z + 2025-08-07T00:20:15.801Z https://docs.axolotl.ai/docs/sequence_parallelism.html - 2025-08-06T18:29:01.134Z + 2025-08-07T00:20:15.801Z https://docs.axolotl.ai/docs/inference.html - 2025-08-06T18:29:01.133Z + 2025-08-07T00:20:15.800Z https://docs.axolotl.ai/docs/fsdp_qlora.html - 2025-08-06T18:29:01.130Z + 2025-08-07T00:20:15.797Z https://docs.axolotl.ai/docs/multi-node.html - 2025-08-06T18:29:01.133Z + 2025-08-07T00:20:15.800Z https://docs.axolotl.ai/docs/lora_optims.html - 2025-08-06T18:29:01.133Z + 2025-08-07T00:20:15.800Z https://docs.axolotl.ai/docs/getting-started.html - 2025-08-06T18:29:01.130Z + 2025-08-07T00:20:15.797Z https://docs.axolotl.ai/docs/dataset_loading.html - 2025-08-06T18:29:01.130Z + 2025-08-07T00:20:15.797Z https://docs.axolotl.ai/docs/lr_groups.html - 2025-08-06T18:29:01.133Z + 2025-08-07T00:20:15.800Z https://docs.axolotl.ai/docs/input_output.html - 2025-08-06T18:29:01.133Z + 2025-08-07T00:20:15.800Z https://docs.axolotl.ai/src/axolotl/integrations/LICENSE.html - 2025-08-06T18:29:01.154Z + 2025-08-07T00:20:15.820Z https://docs.axolotl.ai/src/axolotl/integrations/cut_cross_entropy/ACKNOWLEDGEMENTS.html - 2025-08-06T18:29:01.154Z + 2025-08-07T00:20:15.821Z https://docs.axolotl.ai/docs/mac.html - 2025-08-06T18:29:01.133Z + 2025-08-07T00:20:15.800Z https://docs.axolotl.ai/docs/optimizers.html - 2025-08-06T18:29:01.133Z + 2025-08-07T00:20:15.801Z https://docs.axolotl.ai/docs/gradient_checkpointing.html - 2025-08-06T18:29:01.130Z + 2025-08-07T00:20:15.797Z https://docs.axolotl.ai/docs/qat.html - 2025-08-06T18:29:01.134Z + 2025-08-07T00:20:15.801Z https://docs.axolotl.ai/docs/faq.html - 2025-08-06T18:29:01.130Z + 2025-08-07T00:20:15.797Z https://docs.axolotl.ai/docs/dataset_preprocessing.html - 2025-08-06T18:29:01.130Z + 2025-08-07T00:20:15.797Z https://docs.axolotl.ai/docs/nccl.html - 2025-08-06T18:29:01.133Z + 2025-08-07T00:20:15.801Z https://docs.axolotl.ai/docs/cli.html - 2025-08-06T18:29:01.129Z + 2025-08-07T00:20:15.797Z https://docs.axolotl.ai/docs/torchao.html - 2025-08-06T18:29:01.134Z + 2025-08-07T00:20:15.801Z https://docs.axolotl.ai/docs/multi-gpu.html - 2025-08-06T18:29:01.133Z + 2025-08-07T00:20:15.800Z https://docs.axolotl.ai/docs/rlhf.html - 2025-08-06T18:29:01.134Z + 2025-08-07T00:20:15.801Z https://docs.axolotl.ai/docs/dataset-formats/tokenized.html - 2025-08-06T18:29:01.130Z + 2025-08-07T00:20:15.797Z https://docs.axolotl.ai/docs/dataset-formats/conversation.html - 2025-08-06T18:29:01.130Z + 2025-08-07T00:20:15.797Z https://docs.axolotl.ai/docs/dataset-formats/inst_tune.html - 2025-08-06T18:29:01.130Z + 2025-08-07T00:20:15.797Z https://docs.axolotl.ai/docs/reward_modelling.html - 2025-08-06T18:29:01.134Z + 2025-08-07T00:20:15.801Z https://docs.axolotl.ai/docs/docker.html - 2025-08-06T18:29:01.130Z + 2025-08-07T00:20:15.797Z https://docs.axolotl.ai/docs/installation.html - 2025-08-06T18:29:01.133Z + 2025-08-07T00:20:15.800Z https://docs.axolotl.ai/docs/quantize.html - 2025-08-06T18:29:01.134Z + 2025-08-07T00:20:15.801Z https://docs.axolotl.ai/docs/custom_integrations.html - 2025-08-06T18:29:01.129Z + 2025-08-07T00:20:15.797Z https://docs.axolotl.ai/docs/batch_vs_grad.html - 2025-08-06T18:29:01.129Z + 2025-08-07T00:20:15.796Z https://docs.axolotl.ai/docs/api/cli.utils.train.html - 2025-08-06T18:32:30.894Z + 2025-08-07T00:23:31.905Z https://docs.axolotl.ai/docs/api/cli.art.html - 2025-08-06T18:32:30.758Z + 2025-08-07T00:23:31.773Z https://docs.axolotl.ai/docs/api/core.trainers.grpo.sampler.html - 2025-08-06T18:32:30.956Z + 2025-08-07T00:23:31.965Z https://docs.axolotl.ai/docs/api/loaders.model.html - 2025-08-06T18:32:30.967Z + 2025-08-07T00:23:31.976Z https://docs.axolotl.ai/docs/api/cli.preprocess.html - 2025-08-06T18:32:30.832Z + 2025-08-07T00:23:31.844Z https://docs.axolotl.ai/docs/api/cli.utils.html - 2025-08-06T18:32:30.855Z + 2025-08-07T00:23:31.866Z https://docs.axolotl.ai/docs/api/cli.inference.html - 2025-08-06T18:32:30.802Z + 2025-08-07T00:23:31.816Z https://docs.axolotl.ai/docs/api/monkeypatch.btlm_attn_hijack_flash.html - 2025-08-06T18:32:31.392Z + 2025-08-07T00:23:32.394Z https://docs.axolotl.ai/docs/api/datasets.html - 2025-08-06T18:32:30.549Z + 2025-08-07T00:23:31.565Z https://docs.axolotl.ai/docs/api/monkeypatch.transformers_fa_utils.html - 2025-08-06T18:32:31.409Z + 2025-08-07T00:23:32.411Z https://docs.axolotl.ai/docs/api/monkeypatch.llama_attn_hijack_flash.html - 2025-08-06T18:32:31.344Z + 2025-08-07T00:23:32.347Z https://docs.axolotl.ai/docs/api/monkeypatch.relora.html - 2025-08-06T18:32:31.352Z + 2025-08-07T00:23:32.355Z https://docs.axolotl.ai/docs/api/monkeypatch.stablelm_attn_hijack_flash.html - 2025-08-06T18:32:31.399Z + 2025-08-07T00:23:32.401Z https://docs.axolotl.ai/docs/api/loaders.adapter.html - 2025-08-06T18:32:30.983Z + 2025-08-07T00:23:31.990Z https://docs.axolotl.ai/docs/api/core.trainers.dpo.trainer.html - 2025-08-06T18:32:30.932Z + 2025-08-07T00:23:31.942Z https://docs.axolotl.ai/docs/api/integrations.cut_cross_entropy.args.html - 2025-08-06T18:32:31.816Z + 2025-08-07T00:23:32.810Z https://docs.axolotl.ai/docs/api/monkeypatch.utils.html - 2025-08-06T18:32:31.391Z + 2025-08-07T00:23:32.393Z https://docs.axolotl.ai/docs/api/loaders.processor.html - 2025-08-06T18:32:30.977Z + 2025-08-07T00:23:31.985Z https://docs.axolotl.ai/docs/api/cli.config.html - 2025-08-06T18:32:30.783Z + 2025-08-07T00:23:31.797Z https://docs.axolotl.ai/docs/api/integrations.liger.args.html - 2025-08-06T18:32:31.828Z + 2025-08-07T00:23:32.822Z https://docs.axolotl.ai/docs/api/loaders.tokenizer.html - 2025-08-06T18:32:30.976Z + 2025-08-07T00:23:31.984Z https://docs.axolotl.ai/docs/api/utils.schemas.config.html - 2025-08-06T18:32:31.596Z + 2025-08-07T00:23:32.593Z https://docs.axolotl.ai/docs/api/utils.ctx_managers.sequence_parallel.html - 2025-08-06T18:32:31.034Z + 2025-08-07T00:23:32.040Z https://docs.axolotl.ai/docs/api/core.trainers.mixins.scheduler.html - 2025-08-06T18:32:31.010Z + 2025-08-07T00:23:32.017Z https://docs.axolotl.ai/docs/api/core.trainers.base.html - 2025-08-06T18:32:30.905Z + 2025-08-07T00:23:31.915Z https://docs.axolotl.ai/docs/api/cli.utils.args.html - 2025-08-06T18:32:30.867Z + 2025-08-07T00:23:31.878Z https://docs.axolotl.ai/docs/api/prompt_strategies.messages.chat.html - 2025-08-06T18:32:31.155Z + 2025-08-07T00:23:32.158Z https://docs.axolotl.ai/docs/api/monkeypatch.lora_kernels.html - 2025-08-06T18:32:31.383Z + 2025-08-07T00:23:32.385Z https://docs.axolotl.ai/docs/api/kernels.lora.html - 2025-08-06T18:32:31.309Z + 2025-08-07T00:23:32.312Z https://docs.axolotl.ai/docs/api/cli.vllm_serve.html - 2025-08-06T18:32:30.844Z + 2025-08-07T00:23:31.856Z https://docs.axolotl.ai/docs/api/utils.schemas.multimodal.html - 2025-08-06T18:32:31.645Z + 2025-08-07T00:23:32.642Z https://docs.axolotl.ai/docs/api/utils.schemas.utils.html - 2025-08-06T18:32:31.674Z + 2025-08-07T00:23:32.670Z https://docs.axolotl.ai/docs/api/monkeypatch.llama_attn_hijack_xformers.html - 2025-08-06T18:32:31.346Z + 2025-08-07T00:23:32.349Z https://docs.axolotl.ai/docs/api/integrations.lm_eval.args.html - 2025-08-06T18:32:31.831Z + 2025-08-07T00:23:32.825Z https://docs.axolotl.ai/docs/api/monkeypatch.mistral_attn_hijack_flash.html - 2025-08-06T18:32:31.347Z + 2025-08-07T00:23:32.350Z https://docs.axolotl.ai/docs/api/utils.collators.core.html - 2025-08-06T18:32:31.855Z + 2025-08-07T00:23:32.849Z https://docs.axolotl.ai/docs/api/core.chat.format.chatml.html - 2025-08-06T18:32:30.667Z + 2025-08-07T00:23:31.686Z https://docs.axolotl.ai/docs/api/prompt_strategies.dpo.passthrough.html - 2025-08-06T18:32:31.181Z + 2025-08-07T00:23:32.184Z https://docs.axolotl.ai/docs/api/core.datasets.chat.html - 2025-08-06T18:32:30.675Z + 2025-08-07T00:23:31.694Z https://docs.axolotl.ai/docs/api/utils.bench.html - 2025-08-06T18:32:31.466Z + 2025-08-07T00:23:32.466Z https://docs.axolotl.ai/docs/api/utils.schemas.training.html - 2025-08-06T18:32:31.610Z + 2025-08-07T00:23:32.607Z https://docs.axolotl.ai/docs/api/utils.collators.batching.html - 2025-08-06T18:32:31.874Z + 2025-08-07T00:23:32.867Z https://docs.axolotl.ai/docs/api/prompt_strategies.input_output.html - 2025-08-06T18:32:31.129Z + 2025-08-07T00:23:32.132Z https://docs.axolotl.ai/docs/api/utils.lora.html - 2025-08-06T18:32:31.457Z + 2025-08-07T00:23:32.457Z https://docs.axolotl.ai/docs/api/prompt_strategies.base.html - 2025-08-06T18:32:31.035Z + 2025-08-07T00:23:32.042Z https://docs.axolotl.ai/docs/api/prompt_strategies.alpaca_w_system.html - 2025-08-06T18:32:31.095Z + 2025-08-07T00:23:32.100Z https://docs.axolotl.ai/docs/api/utils.schemas.datasets.html - 2025-08-06T18:32:31.628Z + 2025-08-07T00:23:32.625Z https://docs.axolotl.ai/docs/api/prompt_strategies.dpo.user_defined.html - 2025-08-06T18:32:31.180Z + 2025-08-07T00:23:32.183Z https://docs.axolotl.ai/docs/api/utils.schemas.peft.html - 2025-08-06T18:32:31.637Z + 2025-08-07T00:23:32.633Z https://docs.axolotl.ai/docs/api/prompt_strategies.pygmalion.html - 2025-08-06T18:32:31.150Z + 2025-08-07T00:23:32.154Z https://docs.axolotl.ai/docs/api/common.architectures.html - 2025-08-06T18:32:31.836Z + 2025-08-07T00:23:32.830Z https://docs.axolotl.ai/docs/api/monkeypatch.gradient_checkpointing.offload_cpu.html - 2025-08-06T18:32:31.417Z + 2025-08-07T00:23:32.419Z https://docs.axolotl.ai/docs/api/utils.callbacks.comet_.html - 2025-08-06T18:32:31.942Z + 2025-08-07T00:23:32.935Z https://docs.axolotl.ai/docs/api/integrations.spectrum.args.html - 2025-08-06T18:32:31.835Z + 2025-08-07T00:23:32.828Z https://docs.axolotl.ai/docs/api/cli.quantize.html - 2025-08-06T18:32:30.837Z + 2025-08-07T00:23:31.849Z https://docs.axolotl.ai/docs/api/cli.checks.html - 2025-08-06T18:32:30.764Z + 2025-08-07T00:23:31.780Z https://docs.axolotl.ai/docs/api/prompt_strategies.kto.llama3.html - 2025-08-06T18:32:31.190Z + 2025-08-07T00:23:32.192Z https://docs.axolotl.ai/docs/api/utils.model_shard_quant.html - 2025-08-06T18:32:31.462Z + 2025-08-07T00:23:32.463Z https://docs.axolotl.ai/docs/api/utils.quantization.html - 2025-08-06T18:32:31.581Z + 2025-08-07T00:23:32.579Z https://docs.axolotl.ai/docs/api/core.trainers.mixins.rng_state_loader.html - 2025-08-06T18:32:31.003Z + 2025-08-07T00:23:32.010Z https://docs.axolotl.ai/docs/api/kernels.geglu.html - 2025-08-06T18:32:31.319Z + 2025-08-07T00:23:32.323Z https://docs.axolotl.ai/docs/api/utils.data.pretraining.html - 2025-08-06T18:32:31.553Z + 2025-08-07T00:23:32.552Z https://docs.axolotl.ai/docs/api/prompt_strategies.kto.user_defined.html - 2025-08-06T18:32:31.199Z + 2025-08-07T00:23:32.202Z https://docs.axolotl.ai/docs/api/core.builders.base.html - 2025-08-06T18:32:30.620Z + 2025-08-07T00:23:31.640Z https://docs.axolotl.ai/docs/api/cli.merge_lora.html - 2025-08-06T18:32:30.811Z + 2025-08-07T00:23:31.824Z https://docs.axolotl.ai/docs/api/cli.utils.load.html - 2025-08-06T18:32:30.878Z + 2025-08-07T00:23:31.889Z https://docs.axolotl.ai/docs/api/utils.data.sft.html - 2025-08-06T18:32:31.560Z + 2025-08-07T00:23:32.559Z https://docs.axolotl.ai/docs/api/prompt_strategies.user_defined.html - 2025-08-06T18:32:31.103Z + 2025-08-07T00:23:32.108Z https://docs.axolotl.ai/docs/api/utils.tokenization.html - 2025-08-06T18:32:31.450Z + 2025-08-07T00:23:32.451Z https://docs.axolotl.ai/docs/api/prompt_strategies.dpo.chatml.html - 2025-08-06T18:32:31.177Z + 2025-08-07T00:23:32.180Z https://docs.axolotl.ai/docs/api/models.mamba.modeling_mamba.html - 2025-08-06T18:32:31.854Z + 2025-08-07T00:23:32.847Z https://docs.axolotl.ai/docs/api/cli.args.html - 2025-08-06T18:32:30.755Z + 2025-08-07T00:23:31.770Z https://docs.axolotl.ai/docs/api/evaluate.html - 2025-08-06T18:32:30.537Z + 2025-08-07T00:23:31.554Z https://docs.axolotl.ai/docs/api/prompt_strategies.alpaca_instruct.html - 2025-08-06T18:32:31.083Z + 2025-08-07T00:23:32.088Z https://docs.axolotl.ai/docs/api/utils.distributed.html - 2025-08-06T18:32:31.539Z + 2025-08-07T00:23:32.538Z https://docs.axolotl.ai/docs/multipack.html - 2025-08-06T18:29:01.133Z + 2025-08-07T00:20:15.801Z https://docs.axolotl.ai/examples/colab-notebooks/colab-axolotl-example.html - 2025-08-06T18:29:01.138Z + 2025-08-07T00:20:15.805Z https://docs.axolotl.ai/FAQS.html - 2025-08-06T18:29:01.128Z + 2025-08-07T00:20:15.795Z