Merge branch 'master' into mrwyattii/pydantic-2-support

microsoft · Mar 2, 2024 · 1e8ba21 · 1e8ba21
2 parents 43e6367 + 4578c24
commit 1e8ba21
Show file tree

Hide file tree

Showing 2 changed files with 21 additions and 13 deletions.
diff --git a/.github/workflows/nv-mii.yml b/.github/workflows/nv-mii.yml
@@ -2,6 +2,12 @@ name: nv-mii
 
 on:
   workflow_dispatch:
+    inputs:
+      mii_branch:
+        description: 'DeepSpeed-MII Branch'
+        required: false
+        default: 'main'
+        type: string
   pull_request:
     paths:
       - '.github/workflows/nv-mii.yml'
@@ -55,7 +61,12 @@ jobs:
 
       - name: MII unit tests
         run: |
-          git clone https://github.com/microsoft/DeepSpeed-MII.git
+          BRANCH="main"
+          if [[ ! -z "${{ github.event.inputs.mii_branch }}" ]]; then
+              BRANCH="${{ github.event.inputs.mii_branch }}"
+          fi
+          echo "Cloning DeepSpeed-MII branch: $BRANCH"
+          git clone -b $BRANCH --depth=1 https://github.com/microsoft/DeepSpeed-MII.git
           cd DeepSpeed-MII
           pip install .[dev]
           unset TORCH_CUDA_ARCH_LIST # only jit compile for current arch

diff --git a/deepspeed/runtime/zero/partition_parameters.py b/deepspeed/runtime/zero/partition_parameters.py
@@ -1635,19 +1635,16 @@ def _partition_param_sec(self, param, buffer=None, has_been_updated=False):
             secondary_end = secondary_start + secondary_partition_size
 
             one_dim_param = param.contiguous().view(-1)
-            start = partition_size * self.rank
-            end = start + partition_size
-            if start < param.ds_numel and end <= param.ds_numel:
-                if secondary_start < param.ds_numel and secondary_end <= param.ds_numel:
-                    sec_src_tensor = one_dim_param.narrow(0, secondary_start, secondary_partition_size)
-                    param.ds_secondary_tensor.copy_(sec_src_tensor)
 
-            else:
-                if start < param.ds_numel:
-                    elements_to_copy = param.ds_numel - start
-                    elements_to_copy_sec = elements_to_copy * param.ds_secondary_tensor_num_of_groups
-                    param.ds_secondary_tensor.narrow(0, 0, elements_to_copy_sec).copy_(
-                        one_dim_param.narrow(0, secondary_start, elements_to_copy_sec))
+            # ds_numel is unpadded, so the last chunk of the secondary tensor might not be secondary_partition_size
+            sec_numel = param.ds_numel - secondary_start if secondary_end > param.ds_numel else secondary_partition_size
+
+            # copy from full tensor to secondary tensor
+            param.ds_secondary_tensor.narrow(0, 0,
+                                             sec_numel).copy_(one_dim_param.narrow(0, secondary_start, sec_numel))
+
+            # TODO: This is a temporary fix to avoid the issue that 2nd tensor all-gather happens before 2nd tensor partition is done
+            get_accelerator().current_stream().synchronize()
 
             print_rank_0(f"{param.ds_id} partitioned type {param.dtype} dev {param.device} shape {param.shape}",
                          force=False)