address review item and linting

Author: shewu-quic

backends/vulkan/runtime/graph/ops/impl/BinaryOp.cpp

@@ -132,18 +132,21 @@ void add_binary_op_buffer_node(
       // Shader params buffers
       {},
       // Specialization Constants
-      {graph.packed_dim_of(out), graph.packed_dim_of(in1), graph.packed_dim_of(in2)},
+      {graph.packed_dim_of(out),
+       graph.packed_dim_of(in1),
+       graph.packed_dim_of(in2)},
       // Resizing Logic
       resize_binary_op_node,
       {},
-      {{graph.sizes_pc_of(in1),
-        graph.sizes_pc_of(in2),
-        graph.strides_pc_of(out),
-        graph.strides_pc_of(in1),
-        graph.strides_pc_of(in2),
-        graph.numel_pc_of(out),
-        PushConstantDataInfo(&alpha_val, sizeof(float)),
-        }}));
+      {{
+          graph.sizes_pc_of(in1),
+          graph.sizes_pc_of(in2),
+          graph.strides_pc_of(out),
+          graph.strides_pc_of(in1),
+          graph.strides_pc_of(in2),
+          graph.numel_pc_of(out),
+          PushConstantDataInfo(&alpha_val, sizeof(float)),
+      }}));
 }
 
 void add_binary_op_node(

examples/qualcomm/oss_scripts/llama/model/static_llama.py

@@ -19,9 +19,11 @@
 def apply_rotary_emb_single(
     x: torch.Tensor, freqs_cos: torch.Tensor, freqs_sin: torch.Tensor
 ) -> torch.Tensor:
-    # Change to RoPE of huggingface version
+    # The implementation of RoPE in HuggingFace processes query and key with two half instead of interleaved way.
+    # The main difference is stride in StrideSlice op. For interleaved way, stride is two which is not friendly for HTP backend.
+    # Ref: https://github.com/huggingface/transformers/issues/25199
     x_r, x_i = x[..., : x.shape[-1] // 2], x[..., x.shape[-1] // 2 :]
-    # brodcast for batch_prefill mode input x
+    # broadcast for batch_prefill mode input x
     if x.dim() == 4:
         freqs_cos = freqs_cos[None, None, :, :]
         freqs_sin = freqs_sin[None, None, :, :]