[Bugfix] Make torch registration of punica ops optional (vllm-project…

…#7970)
Hyper-Accel · Aug 28, 2024 · 3cdfe1f · 3cdfe1f
1 parent fdd9daa
commit 3cdfe1f
Show file tree

Hide file tree

Showing 7 changed files with 37 additions and 21 deletions.
diff --git a/vllm/lora/ops/bgmv_expand.py b/vllm/lora/ops/bgmv_expand.py
@@ -160,6 +160,9 @@ def _bgmv_expand(
     return
 
 
-bgmv_expand = torch.library.custom_op("lora::bgmv_expand",
-                                      _bgmv_expand,
-                                      mutates_args=["output_tensor"])
+try:
+    bgmv_expand = torch.library.custom_op("lora::bgmv_expand",
+                                          _bgmv_expand,
+                                          mutates_args=["output_tensor"])
+except AttributeError:
+    bgmv_expand = _bgmv_expand
diff --git a/vllm/lora/ops/bgmv_expand_slice.py b/vllm/lora/ops/bgmv_expand_slice.py
@@ -173,6 +173,9 @@ def _bgmv_expand_slice(
     return
 
 
-bgmv_expand_slice = torch.library.custom_op("lora::bgmv_expand_slice",
-                                            _bgmv_expand_slice,
-                                            mutates_args=["output_tensor"])
+try:
+    bgmv_expand_slice = torch.library.custom_op("lora::bgmv_expand_slice",
+                                                _bgmv_expand_slice,
+                                                mutates_args=["output_tensor"])
+except AttributeError:
+    bgmv_expand_slice = _bgmv_expand_slice
diff --git a/vllm/lora/ops/bgmv_shrink.py b/vllm/lora/ops/bgmv_shrink.py
@@ -142,6 +142,9 @@ def _bgmv_shrink(
     return
 
 
-bgmv_shrink = torch.library.custom_op("lora::bgmv_shrink",
-                                      _bgmv_shrink,
-                                      mutates_args=["output_tensor"])
+try:
+    bgmv_shrink = torch.library.custom_op("lora::bgmv_shrink",
+                                          _bgmv_shrink,
+                                          mutates_args=["output_tensor"])
+except AttributeError:
+    bgmv_shrink = _bgmv_shrink
diff --git a/vllm/lora/ops/sgmv_expand.py b/vllm/lora/ops/sgmv_expand.py
@@ -192,6 +192,9 @@ def _sgmv_expand(
     return
 
 
-sgmv_expand = torch.library.custom_op("lora::sgmv_expand",
-                                      _sgmv_expand,
-                                      mutates_args=["output_tensor"])
+try:
+    sgmv_expand = torch.library.custom_op("lora::sgmv_expand",
+                                          _sgmv_expand,
+                                          mutates_args=["output_tensor"])
+except AttributeError:
+    sgmv_expand = _sgmv_expand
diff --git a/vllm/lora/ops/sgmv_expand_slice.py b/vllm/lora/ops/sgmv_expand_slice.py
@@ -205,6 +205,9 @@ def _sgmv_expand_slice(
     return
 
 
-sgmv_expand_slice = torch.library.custom_op("lora::sgmv_expand_slice",
-                                            _sgmv_expand_slice,
-                                            mutates_args=["output_tensor"])
+try:
+    sgmv_expand_slice = torch.library.custom_op("lora::sgmv_expand_slice",
+                                                _sgmv_expand_slice,
+                                                mutates_args=["output_tensor"])
+except AttributeError:
+    sgmv_expand_slice = _sgmv_expand_slice
diff --git a/vllm/lora/ops/sgmv_shrink.py b/vllm/lora/ops/sgmv_shrink.py
@@ -189,6 +189,9 @@ def _sgmv_shrink(
     return
 
 
-sgmv_shrink = torch.library.custom_op("lora::sgmv_shrink",
-                                      _sgmv_shrink,
-                                      mutates_args=["output_tensor"])
+try:
+    sgmv_shrink = torch.library.custom_op("lora::sgmv_shrink",
+                                          _sgmv_shrink,
+                                          mutates_args=["output_tensor"])
+except AttributeError:
+    sgmv_shrink = _sgmv_shrink
diff --git a/vllm/lora/punica.py b/vllm/lora/punica.py
@@ -10,10 +10,8 @@
 import torch
 
 from vllm.triton_utils import HAS_TRITON
-from vllm.utils import is_xpu
 
-# FIXME: xpu path doesn't support torch.library.custom_op
-if HAS_TRITON and not is_xpu():
+if HAS_TRITON:
     from vllm.lora.ops.bgmv_expand import bgmv_expand
     from vllm.lora.ops.bgmv_expand_slice import bgmv_expand_slice
     from vllm.lora.ops.bgmv_shrink import bgmv_shrink