Remove some tailing spaces

Author: sphish

deep_ep/buffers/elastic.py

@@ -275,7 +275,7 @@ def __init__(self,
         self.allow_multiple_reduction = allow_multiple_reduction
         self.prefer_overlap_with_compute = prefer_overlap_with_compute
         self.deterministic = deterministic
-        
+
         if os.environ.get('NCCL_GIN_CROSS_NIC') == '0':
             # TODO: move this variable into NCCL runtime
             # Multi-plane: all ranks share CPU segments, skip proxy re-export for sysmem handles
@@ -502,7 +502,7 @@ def barrier(self, use_comm_stream: bool = True, with_cpu_sync: bool = False, seq
             use_comm_stream: whether to use the communication stream (otherwise uses the current compute stream).
             with_cpu_sync: whether to also call `cudaDeviceSynchronize` before and after the barrier.
             sequential: whether to run the scaleout and scaleup barriers sequentially (on a single SM) instead of
-                in parallel across SMs. Sequential mode provides better synchronization guarantees, 
+                in parallel across SMs. Sequential mode provides better synchronization guarantees,
                 mainly used for test synchronization.
         """
         self.runtime.barrier(use_comm_stream, with_cpu_sync, sequential)

deep_ep/include/deep_ep/impls/engram_fetch.cuh

@@ -72,7 +72,7 @@ engram_fetch_impl(const ncclDevComm_t nccl_dev_comm, const ncclWindow_t nccl_win
                 // NOTES: requests may exceed the queue depth, flush if needed
                 (request_idx % kGinQPFlushDepth == (kGinQPFlushDepth - 1)) ? 0 : ncclGinOptFlagsAggregateRequests
             );
-            
+
             // TODO: once NCCL supports ncclCoopWarp gin.get, drop the elect_one_sync and let the whole warp
             // gather SF packs in parallel.
             if constexpr (kNumSFPacks > 0) {

tests/elastic/test_ep.py

@@ -103,7 +103,7 @@ def get_recv_x_bf16(recv_x) -> torch.Tensor:
                 return per_token_cast_back(recv_x[0], recv_x[1])
             else:
                 return recv_x
-            
+
         # Test correctness with NCCL reference
         if not args.skip_check:
             ref_recv_x, ref_recv_topk_idx, ref_recv_topk_weights, \
@@ -175,7 +175,7 @@ def get_recv_x_bf16(recv_x) -> torch.Tensor:
             do_zero_padding=True, handle=expanded_handle)
         cached_expanded_recv_x, _, cached_expanded_recv_topk_weights, _, _ = \
             launch(buffer, 'dispatch', with_previous_event, async_with_compute_stream, cached_expanded_dispatch_args)
-        
+
         # Count the number of received tokens
         num_recv_tokens = handle.psum_num_recv_tokens_per_scaleup_rank[-1].item()
         assert num_recv_tokens == expanded_handle.psum_num_recv_tokens_per_scaleup_rank[-1].item(), \
@@ -306,7 +306,7 @@ def get_unique_and_valid_dst_count(dst_idx: torch.Tensor,
                     dst_idx = dst_idx + torch.arange(0, dst_idx.shape[0], dtype=dst_idx.dtype, device=dst_idx.device).unsqueeze(-1) * (max_num_in_dst_idx + 1)  # So that different rows will have different values
                     dst_idx[ignore_mask] = dst_idx[0][0].item()  # So that these `-1`s won't affect the count of unique numbers
                     return torch.unique(dst_idx, sorted=False).numel()
-                
+
                 if not args.allow_multiple_reduction:
                     # No multiple reduction
                     if not is_expand_mode: