Remove legacy triton operators (fairinternal/xformers#1140)

* Remove legacy triton operators Those haven't been maintained in a while. Following #848 * Remove docs __original_commit__ = fairinternal/xformers@5d7c0de
facebookresearch · Jun 24, 2024 · 9ed9561 · 9ed9561
1 parent b1b69fa
commit 9ed9561
Show file tree

Hide file tree

Showing 95 changed files with 11 additions and 3,497 deletions.
diff --git a/docs/plots/fused_dropout/Dropout_Bias_False_FW+BW_torch.float16_Act_None.png b/docs/plots/fused_dropout/Dropout_Bias_False_FW+BW_torch.float16_Act_None.png
diff --git a/docs/plots/fused_dropout/Dropout_Bias_False_FW+BW_torch.float16_Act_gelu.png b/docs/plots/fused_dropout/Dropout_Bias_False_FW+BW_torch.float16_Act_gelu.png
diff --git a/...plots/fused_dropout/Dropout_Bias_False_FW+BW_torch.float16_Act_squared_relu.png b/...plots/fused_dropout/Dropout_Bias_False_FW+BW_torch.float16_Act_squared_relu.png
diff --git a/docs/plots/fused_dropout/Dropout_Bias_False_FW+BW_torch.float32_Act_None.png b/docs/plots/fused_dropout/Dropout_Bias_False_FW+BW_torch.float32_Act_None.png
diff --git a/docs/plots/fused_dropout/Dropout_Bias_False_FW+BW_torch.float32_Act_gelu.png b/docs/plots/fused_dropout/Dropout_Bias_False_FW+BW_torch.float32_Act_gelu.png
diff --git a/...plots/fused_dropout/Dropout_Bias_False_FW+BW_torch.float32_Act_squared_relu.png b/...plots/fused_dropout/Dropout_Bias_False_FW+BW_torch.float32_Act_squared_relu.png
diff --git a/docs/plots/fused_dropout/Dropout_Bias_False_FW_torch.float16_Act_None.png b/docs/plots/fused_dropout/Dropout_Bias_False_FW_torch.float16_Act_None.png
diff --git a/docs/plots/fused_dropout/Dropout_Bias_False_FW_torch.float16_Act_gelu.png b/docs/plots/fused_dropout/Dropout_Bias_False_FW_torch.float16_Act_gelu.png
diff --git a/docs/plots/fused_dropout/Dropout_Bias_False_FW_torch.float16_Act_squared_relu.png b/docs/plots/fused_dropout/Dropout_Bias_False_FW_torch.float16_Act_squared_relu.png
diff --git a/docs/plots/fused_dropout/Dropout_Bias_False_FW_torch.float32_Act_None.png b/docs/plots/fused_dropout/Dropout_Bias_False_FW_torch.float32_Act_None.png
diff --git a/docs/plots/fused_dropout/Dropout_Bias_False_FW_torch.float32_Act_gelu.png b/docs/plots/fused_dropout/Dropout_Bias_False_FW_torch.float32_Act_gelu.png
diff --git a/docs/plots/fused_dropout/Dropout_Bias_False_FW_torch.float32_Act_squared_relu.png b/docs/plots/fused_dropout/Dropout_Bias_False_FW_torch.float32_Act_squared_relu.png
diff --git a/docs/plots/fused_dropout/Dropout_Bias_True_FW+BW_torch.float16_Act_None.png b/docs/plots/fused_dropout/Dropout_Bias_True_FW+BW_torch.float16_Act_None.png
diff --git a/docs/plots/fused_dropout/Dropout_Bias_True_FW+BW_torch.float16_Act_gelu.png b/docs/plots/fused_dropout/Dropout_Bias_True_FW+BW_torch.float16_Act_gelu.png
diff --git a/.../plots/fused_dropout/Dropout_Bias_True_FW+BW_torch.float16_Act_squared_relu.png b/.../plots/fused_dropout/Dropout_Bias_True_FW+BW_torch.float16_Act_squared_relu.png
diff --git a/docs/plots/fused_dropout/Dropout_Bias_True_FW+BW_torch.float32_Act_None.png b/docs/plots/fused_dropout/Dropout_Bias_True_FW+BW_torch.float32_Act_None.png
diff --git a/docs/plots/fused_dropout/Dropout_Bias_True_FW+BW_torch.float32_Act_gelu.png b/docs/plots/fused_dropout/Dropout_Bias_True_FW+BW_torch.float32_Act_gelu.png
diff --git a/.../plots/fused_dropout/Dropout_Bias_True_FW+BW_torch.float32_Act_squared_relu.png b/.../plots/fused_dropout/Dropout_Bias_True_FW+BW_torch.float32_Act_squared_relu.png
diff --git a/docs/plots/fused_dropout/Dropout_Bias_True_FW_torch.float16_Act_None.png b/docs/plots/fused_dropout/Dropout_Bias_True_FW_torch.float16_Act_None.png
diff --git a/docs/plots/fused_dropout/Dropout_Bias_True_FW_torch.float16_Act_gelu.png b/docs/plots/fused_dropout/Dropout_Bias_True_FW_torch.float16_Act_gelu.png
diff --git a/docs/plots/fused_dropout/Dropout_Bias_True_FW_torch.float16_Act_squared_relu.png b/docs/plots/fused_dropout/Dropout_Bias_True_FW_torch.float16_Act_squared_relu.png
diff --git a/docs/plots/fused_dropout/Dropout_Bias_True_FW_torch.float32_Act_None.png b/docs/plots/fused_dropout/Dropout_Bias_True_FW_torch.float32_Act_None.png
diff --git a/docs/plots/fused_dropout/Dropout_Bias_True_FW_torch.float32_Act_gelu.png b/docs/plots/fused_dropout/Dropout_Bias_True_FW_torch.float32_Act_gelu.png
diff --git a/docs/plots/fused_dropout/Dropout_Bias_True_FW_torch.float32_Act_squared_relu.png b/docs/plots/fused_dropout/Dropout_Bias_True_FW_torch.float32_Act_squared_relu.png
diff --git a/docs/plots/fused_linear/FusedLinear_fp16_FW_BW_gelu.png b/docs/plots/fused_linear/FusedLinear_fp16_FW_BW_gelu.png
diff --git a/docs/plots/fused_linear/FusedLinear_fp16_FW_BW_leaky_relu.png b/docs/plots/fused_linear/FusedLinear_fp16_FW_BW_leaky_relu.png
diff --git a/docs/plots/fused_linear/FusedLinear_fp16_FW_BW_none.png b/docs/plots/fused_linear/FusedLinear_fp16_FW_BW_none.png
diff --git a/docs/plots/fused_linear/FusedLinear_fp16_FW_BW_relu.png b/docs/plots/fused_linear/FusedLinear_fp16_FW_BW_relu.png
diff --git a/docs/plots/fused_linear/FusedLinear_fp16_FW_BW_smelu.png b/docs/plots/fused_linear/FusedLinear_fp16_FW_BW_smelu.png
diff --git a/docs/plots/fused_linear/FusedLinear_fp16_FW_BW_squared_relu.png b/docs/plots/fused_linear/FusedLinear_fp16_FW_BW_squared_relu.png
diff --git a/docs/plots/fused_linear/FusedLinear_fp16_FW_BW_star_relu.png b/docs/plots/fused_linear/FusedLinear_fp16_FW_BW_star_relu.png
diff --git a/docs/plots/fused_linear/FusedLinear_fp16_FW_gelu.png b/docs/plots/fused_linear/FusedLinear_fp16_FW_gelu.png
diff --git a/docs/plots/fused_linear/FusedLinear_fp16_FW_leaky_relu.png b/docs/plots/fused_linear/FusedLinear_fp16_FW_leaky_relu.png
diff --git a/docs/plots/fused_linear/FusedLinear_fp16_FW_none.png b/docs/plots/fused_linear/FusedLinear_fp16_FW_none.png
diff --git a/docs/plots/fused_linear/FusedLinear_fp16_FW_relu.png b/docs/plots/fused_linear/FusedLinear_fp16_FW_relu.png
diff --git a/docs/plots/fused_linear/FusedLinear_fp16_FW_smelu.png b/docs/plots/fused_linear/FusedLinear_fp16_FW_smelu.png
diff --git a/docs/plots/fused_linear/FusedLinear_fp16_FW_squared_relu.png b/docs/plots/fused_linear/FusedLinear_fp16_FW_squared_relu.png
diff --git a/docs/plots/fused_linear/FusedLinear_fp16_FW_star_relu.png b/docs/plots/fused_linear/FusedLinear_fp16_FW_star_relu.png
diff --git a/docs/plots/fused_linear/FusedLinear_fp32_FW_BW_gelu.png b/docs/plots/fused_linear/FusedLinear_fp32_FW_BW_gelu.png
diff --git a/docs/plots/fused_linear/FusedLinear_fp32_FW_BW_leaky_relu.png b/docs/plots/fused_linear/FusedLinear_fp32_FW_BW_leaky_relu.png
diff --git a/docs/plots/fused_linear/FusedLinear_fp32_FW_BW_none.png b/docs/plots/fused_linear/FusedLinear_fp32_FW_BW_none.png
diff --git a/docs/plots/fused_linear/FusedLinear_fp32_FW_BW_relu.png b/docs/plots/fused_linear/FusedLinear_fp32_FW_BW_relu.png
diff --git a/docs/plots/fused_linear/FusedLinear_fp32_FW_BW_smelu.png b/docs/plots/fused_linear/FusedLinear_fp32_FW_BW_smelu.png
diff --git a/docs/plots/fused_linear/FusedLinear_fp32_FW_BW_squared_relu.png b/docs/plots/fused_linear/FusedLinear_fp32_FW_BW_squared_relu.png
diff --git a/docs/plots/fused_linear/FusedLinear_fp32_FW_BW_star_relu.png b/docs/plots/fused_linear/FusedLinear_fp32_FW_BW_star_relu.png
diff --git a/docs/plots/fused_linear/FusedLinear_fp32_FW_gelu.png b/docs/plots/fused_linear/FusedLinear_fp32_FW_gelu.png
diff --git a/docs/plots/fused_linear/FusedLinear_fp32_FW_leaky_relu.png b/docs/plots/fused_linear/FusedLinear_fp32_FW_leaky_relu.png
diff --git a/docs/plots/fused_linear/FusedLinear_fp32_FW_none.png b/docs/plots/fused_linear/FusedLinear_fp32_FW_none.png
diff --git a/docs/plots/fused_linear/FusedLinear_fp32_FW_relu.png b/docs/plots/fused_linear/FusedLinear_fp32_FW_relu.png
diff --git a/docs/plots/fused_linear/FusedLinear_fp32_FW_smelu.png b/docs/plots/fused_linear/FusedLinear_fp32_FW_smelu.png
diff --git a/docs/plots/fused_linear/FusedLinear_fp32_FW_squared_relu.png b/docs/plots/fused_linear/FusedLinear_fp32_FW_squared_relu.png
diff --git a/docs/plots/fused_linear/FusedLinear_fp32_FW_star_relu.png b/docs/plots/fused_linear/FusedLinear_fp32_FW_star_relu.png
diff --git a/docs/plots/fused_softmax/Softmax_Bandwidth_FW_BW_fp16.png b/docs/plots/fused_softmax/Softmax_Bandwidth_FW_BW_fp16.png
diff --git a/docs/plots/fused_softmax/Softmax_Bandwidth_FW_BW_fp32.png b/docs/plots/fused_softmax/Softmax_Bandwidth_FW_BW_fp32.png
diff --git a/docs/plots/fused_softmax/Softmax_Bandwidth_FW_fp16.png b/docs/plots/fused_softmax/Softmax_Bandwidth_FW_fp16.png
diff --git a/docs/plots/fused_softmax/Softmax_Bandwidth_FW_fp32.png b/docs/plots/fused_softmax/Softmax_Bandwidth_FW_fp32.png
diff --git a/docs/plots/layer_norm/LayerNorm_FW+BW_torch.float16.png b/docs/plots/layer_norm/LayerNorm_FW+BW_torch.float16.png
diff --git a/docs/plots/layer_norm/LayerNorm_FW+BW_torch.float32.png b/docs/plots/layer_norm/LayerNorm_FW+BW_torch.float32.png
diff --git a/docs/plots/layer_norm/LayerNorm_FW_torch.float16.png b/docs/plots/layer_norm/LayerNorm_FW_torch.float16.png
diff --git a/docs/plots/layer_norm/LayerNorm_FW_torch.float32.png b/docs/plots/layer_norm/LayerNorm_FW_torch.float32.png
diff --git a/docs/plots/strided_sum/Strided_sum_fp16.png b/docs/plots/strided_sum/Strided_sum_fp16.png
diff --git a/docs/plots/strided_sum/Strided_sum_fp32.png b/docs/plots/strided_sum/Strided_sum_fp32.png
diff --git a/docs/source/custom_parts/index.rst b/docs/source/custom_parts/index.rst
@@ -35,31 +35,6 @@ The sparse attention computation is automatically triggered when using the **sca
 There is nothing specific to do, and a couple of examples are provided in the tutorials.
 
 
-
-
-Triton parts
-############
-
-1. Requirements
-***************
-
-We use Triton_ to implement the following parts.
-These parts will only be visible on a CUDA-enabled machine, and Triton needs to be installed (`pip install triton`),
-if any of these conditions are not met a warning is issued.
-
-
-2. Possible usage
-*****************
-
-The following parts are independent and can be used as-is in any model,
-provided the above limitations (Triton is installed, and there is a CUDA GPU present) are fullfilled.
-They are used by default, when possible, in some of the xFormers building blocks.
-
-.. automodule:: xformers.triton
-    :members:
-    :undoc-members:
-
-
 .. _Triton: https://triton-lang.org/
 .. _Sputnik: https://github.com/google-research/sputnik
 .. _see: https://github.com/facebookresearch/xformers/blob/main/xformers/components/attention/scaled_dot_product.py
diff --git a/docs/source/tutorials/index.rst b/docs/source/tutorials/index.rst
@@ -9,4 +9,3 @@ Tutorials
    extend_attentions
    use_attention
    reversible
-   triton
diff --git a/docs/source/tutorials/triton.rst b/docs/source/tutorials/triton.rst
diff --git a/examples/cifar_ViT.py b/examples/cifar_ViT.py
@@ -79,7 +79,7 @@ def __init__(
                     },
                 },
                 "feedforward_config": {
-                    "name": "FusedMLP",
+                    "name": "MLP",
                     "dropout": mlp_pdrop,
                     "activation": "gelu",
                     "hidden_layer_multiplier": hidden_layer_multiplier,

diff --git a/examples/microGPT.py b/examples/microGPT.py
@@ -72,7 +72,7 @@ def __init__(
                     },
                 },
                 "feedforward_config": {
-                    "name": "FusedMLP",  # Use MLP if Triton is not available
+                    "name": "MLP",
                     "dropout": self.hparams.mlp_pdrop,
                     "activation": "gelu",
                     "hidden_layer_multiplier": self.hparams.hidden_layer_multiplier,

diff --git a/tests/test_pickling.py b/tests/test_pickling.py
@@ -12,7 +12,6 @@
 import pytest
 from torch import nn
 
-from xformers import _is_triton_available
 from xformers.factory import xFormer, xFormerConfig
 
 test_config = [
@@ -33,7 +32,7 @@
             },
         },
         "feedforward_config": {
-            "name": "FusedMLP",
+            "name": "MLP",
             "dropout": 0.1,
             "activation": "gelu",
             "hidden_layer_multiplier": 4,
@@ -51,8 +50,6 @@ def __init__(self, mlp):
 
 
 MLPs = ["MLP"]
-if _is_triton_available():
-    MLPs.append("FusedMLP")
 
 
 @pytest.mark.parametrize("mlp", MLPs)