Added Pairwise Lp Distance

kernel_overview.md

@@ -16,4 +16,6 @@
 | 12 | Batched Patchng | Rearranging | [link](kernels/0012_Patching.md) |
 | 13 | Block-Sparse Matrix Multiplication | Sparsity, Matmul | [link](kernels/0013_Sparse_Matmul.md) |
 | 14 | Fused Linear Cross Entropy | Linear Cross Entropy | [link](kernels/0014_Fused_Linear_Cross_Entropy.md) |
+| 15 | GPTQ Inference | GPTQ | [link](kernels/0015_GPTQ_Triton.md) |
+| 16 | Pairwise Lp Distance | Distance | [link](kernels/0016_Pairwise_Lp_Distance.md) |
 | n+1 | your contribution | its tags |  |

kernels/0016_Pairwise_Lp_Distance.md

@@ -0,0 +1,25 @@
+# Pairwise Lp Distance
+
+**Link:** https://github.com/jinensetpal/triton_cdist/blob/3305c5592c51b51f1080933e58ab66c1fbaa620d/triton_cdist/lp_reduce.py
+
+**Author:** Jinen Setpal
+
+**Tags:** Distance
+
+**Description:** A direct replacement of torch's [cdist](https://docs.pytorch.org/docs/stable/generated/torch.cdist.html#torch-cdist), including support for backprop. Doesn't take torch's `compute_mode` argument.
+
+**Minimal Usage:**
+```py
+import triton_cdist  # registers operator
+
+# from here you can use it as a stand-in replacement of `torch.cdist`.
+
+x1 = ...
+x2 = ...
+p = ...
+torch.ops.triton_cdist.opt_cdist(x1, x2, p=p)  #  previously, `torch.cdist(x1, x1, p=p)`
+```
+
+**Triton Version:** v3.3.1+
+
+**Id in triton index:** 0016