Build from kernel-builder job

build/torch26-cxx11-cu118-x86_64-linux/megablocks/_megablocks_76c7de7.abi3.so

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3c5605ba50f2661b9dc4c5609572323fb4f52787181109c5900c261c5e2bf602
+size 10517576

build/torch26-cxx11-cu118-x86_64-linux/megablocks/_ops.py

@@ -1,9 +1,9 @@
 import torch
-from . import _megablocks_9a1816c
-ops = torch.ops._megablocks_9a1816c
+from . import _megablocks_76c7de7
+ops = torch.ops._megablocks_76c7de7
 
 def add_op_namespace_prefix(op_name: str):
     """
     Prefix op by namespace.
     """
-    return f"_megablocks_9a1816c::{op_name}"
+    return f"_megablocks_76c7de7::{op_name}"

build/torch26-cxx11-cu118-x86_64-linux/megablocks/layers.py

@@ -692,8 +692,8 @@ class MegaBlocksMoeMLP(torch.nn.Module):
         moe_normalize_expert_weights = getattr(self.experts, "normalize_expert_weights", None)
         uniform_expert_assignment = getattr(self, "uniform_expert_assignment", False)
         
-        has_parallel = hasattr(self, "expert_parallel_group")
-        expert_parallel_group = torch.distributed.group.WORLD
+        expert_parallel_group = getattr(self, "expert_parallel_group", None)
+        has_parallel = expert_parallel_group is not None and dist.is_initialized() and dist.get_world_size(expert_parallel_group) > 1
         forward_fn = parallel_forward_once if has_parallel else forward_once
         
         sort_end_bit = max(int(torch.ceil(torch.log2(torch.tensor(moe_num_experts)))), 1)

build/torch26-cxx11-cu124-x86_64-linux/megablocks/_megablocks_76c7de7.abi3.so

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:12c21d6f72f90950adbda156534691dd753476a18719b416541e8d6920a173b4
+size 11869392

build/torch26-cxx11-cu124-x86_64-linux/megablocks/_ops.py

@@ -1,9 +1,9 @@
 import torch
-from . import _megablocks_9a1816c
-ops = torch.ops._megablocks_9a1816c
+from . import _megablocks_76c7de7
+ops = torch.ops._megablocks_76c7de7
 
 def add_op_namespace_prefix(op_name: str):
     """
     Prefix op by namespace.
     """
-    return f"_megablocks_9a1816c::{op_name}"
+    return f"_megablocks_76c7de7::{op_name}"

build/torch26-cxx11-cu124-x86_64-linux/megablocks/layers.py

@@ -692,8 +692,8 @@ class MegaBlocksMoeMLP(torch.nn.Module):
         moe_normalize_expert_weights = getattr(self.experts, "normalize_expert_weights", None)
         uniform_expert_assignment = getattr(self, "uniform_expert_assignment", False)
         
-        has_parallel = hasattr(self, "expert_parallel_group")
-        expert_parallel_group = torch.distributed.group.WORLD
+        expert_parallel_group = getattr(self, "expert_parallel_group", None)
+        has_parallel = expert_parallel_group is not None and dist.is_initialized() and dist.get_world_size(expert_parallel_group) > 1
         forward_fn = parallel_forward_once if has_parallel else forward_once
         
         sort_end_bit = max(int(torch.ceil(torch.log2(torch.tensor(moe_num_experts)))), 1)

build/torch26-cxx11-cu126-x86_64-linux/megablocks/_megablocks_76c7de7.abi3.so

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:db7e3b7c3c15af78fe9ef0ba50c33cb2cb988bdf5dfb1f46807b7871e7c8e70e
+size 11931048

build/torch26-cxx11-cu126-x86_64-linux/megablocks/_ops.py

@@ -1,9 +1,9 @@
 import torch
-from . import _megablocks_9a1816c
-ops = torch.ops._megablocks_9a1816c
+from . import _megablocks_76c7de7
+ops = torch.ops._megablocks_76c7de7
 
 def add_op_namespace_prefix(op_name: str):
     """
     Prefix op by namespace.
     """
-    return f"_megablocks_9a1816c::{op_name}"
+    return f"_megablocks_76c7de7::{op_name}"

build/torch26-cxx11-cu126-x86_64-linux/megablocks/layers.py

@@ -692,8 +692,8 @@ class MegaBlocksMoeMLP(torch.nn.Module):
         moe_normalize_expert_weights = getattr(self.experts, "normalize_expert_weights", None)
         uniform_expert_assignment = getattr(self, "uniform_expert_assignment", False)
         
-        has_parallel = hasattr(self, "expert_parallel_group")
-        expert_parallel_group = torch.distributed.group.WORLD
+        expert_parallel_group = getattr(self, "expert_parallel_group", None)
+        has_parallel = expert_parallel_group is not None and dist.is_initialized() and dist.get_world_size(expert_parallel_group) > 1
         forward_fn = parallel_forward_once if has_parallel else forward_once
         
         sort_end_bit = max(int(torch.ceil(torch.log2(torch.tensor(moe_num_experts)))), 1)

build/torch26-cxx98-cu118-x86_64-linux/megablocks/_megablocks_76c7de7.abi3.so

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d9971a30d397598ee0a58118b8cca337d142de1ca34404532dfda6328122ab11
+size 10510040

build/torch26-cxx98-cu118-x86_64-linux/megablocks/_ops.py

@@ -1,9 +1,9 @@
 import torch
-from . import _megablocks_9a1816c
-ops = torch.ops._megablocks_9a1816c
+from . import _megablocks_76c7de7
+ops = torch.ops._megablocks_76c7de7
 
 def add_op_namespace_prefix(op_name: str):
     """
     Prefix op by namespace.
     """
-    return f"_megablocks_9a1816c::{op_name}"
+    return f"_megablocks_76c7de7::{op_name}"

build/torch26-cxx98-cu118-x86_64-linux/megablocks/layers.py

@@ -692,8 +692,8 @@ class MegaBlocksMoeMLP(torch.nn.Module):
         moe_normalize_expert_weights = getattr(self.experts, "normalize_expert_weights", None)
         uniform_expert_assignment = getattr(self, "uniform_expert_assignment", False)
         
-        has_parallel = hasattr(self, "expert_parallel_group")
-        expert_parallel_group = torch.distributed.group.WORLD
+        expert_parallel_group = getattr(self, "expert_parallel_group", None)
+        has_parallel = expert_parallel_group is not None and dist.is_initialized() and dist.get_world_size(expert_parallel_group) > 1
         forward_fn = parallel_forward_once if has_parallel else forward_once
         
         sort_end_bit = max(int(torch.ceil(torch.log2(torch.tensor(moe_num_experts)))), 1)

build/torch26-cxx98-cu124-x86_64-linux/megablocks/_megablocks_76c7de7.abi3.so

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c136e90b35e7fd43fcc4d987588f68b3f4cfea295a00f1fda343acc9c8848577
+size 11857920

build/torch26-cxx98-cu124-x86_64-linux/megablocks/_ops.py

@@ -1,9 +1,9 @@
 import torch
-from . import _megablocks_9a1816c
-ops = torch.ops._megablocks_9a1816c
+from . import _megablocks_76c7de7
+ops = torch.ops._megablocks_76c7de7
 
 def add_op_namespace_prefix(op_name: str):
     """
     Prefix op by namespace.
     """
-    return f"_megablocks_9a1816c::{op_name}"
+    return f"_megablocks_76c7de7::{op_name}"

build/torch26-cxx98-cu124-x86_64-linux/megablocks/layers.py

@@ -692,8 +692,8 @@ class MegaBlocksMoeMLP(torch.nn.Module):
         moe_normalize_expert_weights = getattr(self.experts, "normalize_expert_weights", None)
         uniform_expert_assignment = getattr(self, "uniform_expert_assignment", False)
         
-        has_parallel = hasattr(self, "expert_parallel_group")
-        expert_parallel_group = torch.distributed.group.WORLD
+        expert_parallel_group = getattr(self, "expert_parallel_group", None)
+        has_parallel = expert_parallel_group is not None and dist.is_initialized() and dist.get_world_size(expert_parallel_group) > 1
         forward_fn = parallel_forward_once if has_parallel else forward_once
         
         sort_end_bit = max(int(torch.ceil(torch.log2(torch.tensor(moe_num_experts)))), 1)

build/torch26-cxx98-cu126-x86_64-linux/megablocks/_megablocks_76c7de7.abi3.so

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a3f893773ec7b8157a4531a57821807f5f27ac48ceaa695c342cc7a39ad318dc
+size 11927768

build/torch26-cxx98-cu126-x86_64-linux/megablocks/_ops.py

@@ -1,9 +1,9 @@
 import torch
-from . import _megablocks_9a1816c
-ops = torch.ops._megablocks_9a1816c
+from . import _megablocks_76c7de7
+ops = torch.ops._megablocks_76c7de7
 
 def add_op_namespace_prefix(op_name: str):
     """
     Prefix op by namespace.
     """
-    return f"_megablocks_9a1816c::{op_name}"
+    return f"_megablocks_76c7de7::{op_name}"

build/torch26-cxx98-cu126-x86_64-linux/megablocks/layers.py

@@ -692,8 +692,8 @@ class MegaBlocksMoeMLP(torch.nn.Module):
         moe_normalize_expert_weights = getattr(self.experts, "normalize_expert_weights", None)
         uniform_expert_assignment = getattr(self, "uniform_expert_assignment", False)
         
-        has_parallel = hasattr(self, "expert_parallel_group")
-        expert_parallel_group = torch.distributed.group.WORLD
+        expert_parallel_group = getattr(self, "expert_parallel_group", None)
+        has_parallel = expert_parallel_group is not None and dist.is_initialized() and dist.get_world_size(expert_parallel_group) > 1
         forward_fn = parallel_forward_once if has_parallel else forward_once
         
         sort_end_bit = max(int(torch.ceil(torch.log2(torch.tensor(moe_num_experts)))), 1)

build/torch27-cxx11-cu118-x86_64-linux/megablocks/_megablocks_76c7de7.abi3.so

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:aff7108245384777d22e9023ae3fd4cf2bcb0015a0938e314d556dbd3e59fe00
+size 10517816

build/torch27-cxx11-cu118-x86_64-linux/megablocks/_ops.py

@@ -1,9 +1,9 @@
 import torch
-from . import _megablocks_9a1816c
-ops = torch.ops._megablocks_9a1816c
+from . import _megablocks_76c7de7
+ops = torch.ops._megablocks_76c7de7
 
 def add_op_namespace_prefix(op_name: str):
     """
     Prefix op by namespace.
     """
-    return f"_megablocks_9a1816c::{op_name}"
+    return f"_megablocks_76c7de7::{op_name}"

build/torch27-cxx11-cu118-x86_64-linux/megablocks/layers.py

@@ -692,8 +692,8 @@ class MegaBlocksMoeMLP(torch.nn.Module):
         moe_normalize_expert_weights = getattr(self.experts, "normalize_expert_weights", None)
         uniform_expert_assignment = getattr(self, "uniform_expert_assignment", False)
         
-        has_parallel = hasattr(self, "expert_parallel_group")
-        expert_parallel_group = torch.distributed.group.WORLD
+        expert_parallel_group = getattr(self, "expert_parallel_group", None)
+        has_parallel = expert_parallel_group is not None and dist.is_initialized() and dist.get_world_size(expert_parallel_group) > 1
         forward_fn = parallel_forward_once if has_parallel else forward_once
         
         sort_end_bit = max(int(torch.ceil(torch.log2(torch.tensor(moe_num_experts)))), 1)

build/torch27-cxx11-cu126-x86_64-linux/megablocks/_megablocks_76c7de7.abi3.so

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:aa1eeccba0a3a26435538a2aa87bc22a40c0201a79979872f6296af984e7bf1e
+size 11931080

build/torch27-cxx11-cu126-x86_64-linux/megablocks/_ops.py

@@ -1,9 +1,9 @@
 import torch
-from . import _megablocks_9a1816c
-ops = torch.ops._megablocks_9a1816c
+from . import _megablocks_76c7de7
+ops = torch.ops._megablocks_76c7de7
 
 def add_op_namespace_prefix(op_name: str):
     """
     Prefix op by namespace.
     """
-    return f"_megablocks_9a1816c::{op_name}"
+    return f"_megablocks_76c7de7::{op_name}"

build/torch27-cxx11-cu126-x86_64-linux/megablocks/layers.py

@@ -692,8 +692,8 @@ class MegaBlocksMoeMLP(torch.nn.Module):
         moe_normalize_expert_weights = getattr(self.experts, "normalize_expert_weights", None)
         uniform_expert_assignment = getattr(self, "uniform_expert_assignment", False)
         
-        has_parallel = hasattr(self, "expert_parallel_group")
-        expert_parallel_group = torch.distributed.group.WORLD
+        expert_parallel_group = getattr(self, "expert_parallel_group", None)
+        has_parallel = expert_parallel_group is not None and dist.is_initialized() and dist.get_world_size(expert_parallel_group) > 1
         forward_fn = parallel_forward_once if has_parallel else forward_once
         
         sort_end_bit = max(int(torch.ceil(torch.log2(torch.tensor(moe_num_experts)))), 1)

build/torch27-cxx11-cu128-x86_64-linux/megablocks/_megablocks_76c7de7.abi3.so

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:64e2fd33ed4a5e9497ad304763c3c174ade26702a8e43fe8e7b3d3e79eb1e021
+size 17892624

build/torch27-cxx11-cu128-x86_64-linux/megablocks/_ops.py

@@ -1,9 +1,9 @@
 import torch
-from . import _megablocks_9a1816c
-ops = torch.ops._megablocks_9a1816c
+from . import _megablocks_76c7de7
+ops = torch.ops._megablocks_76c7de7
 
 def add_op_namespace_prefix(op_name: str):
     """
     Prefix op by namespace.
     """
-    return f"_megablocks_9a1816c::{op_name}"
+    return f"_megablocks_76c7de7::{op_name}"

build/torch27-cxx11-cu128-x86_64-linux/megablocks/layers.py

@@ -692,8 +692,8 @@ class MegaBlocksMoeMLP(torch.nn.Module):
         moe_normalize_expert_weights = getattr(self.experts, "normalize_expert_weights", None)
         uniform_expert_assignment = getattr(self, "uniform_expert_assignment", False)
         
-        has_parallel = hasattr(self, "expert_parallel_group")
-        expert_parallel_group = torch.distributed.group.WORLD
+        expert_parallel_group = getattr(self, "expert_parallel_group", None)
+        has_parallel = expert_parallel_group is not None and dist.is_initialized() and dist.get_world_size(expert_parallel_group) > 1
         forward_fn = parallel_forward_once if has_parallel else forward_once
         
         sort_end_bit = max(int(torch.ceil(torch.log2(torch.tensor(moe_num_experts)))), 1)