[Examples] appy flash attention op for decode phrase

Author: GuoningHuang

examples/BuddyDeepSeekR1/CMakeLists.txt

@@ -264,7 +264,9 @@ add_custom_command(
 
 add_custom_command(
     OUTPUT subgraph-f16.o
-    COMMAND ${LLVM_TOOLS_BINARY_DIR}/mlir-opt ${CMAKE_CURRENT_BINARY_DIR}/subgraph0-f16.mlir
+    COMMAND ${BUDDY_BINARY_DIR}/buddy-opt ${CMAKE_CURRENT_BINARY_DIR}/subgraph0-f16.mlir
+              -simplify-tosa-reshape |
+            ${LLVM_TOOLS_BINARY_DIR}/mlir-opt
               -pass-pipeline ${TOSA_PIPELINE} |
             ${BUDDY_BINARY_DIR}/buddy-opt
             -eliminate-empty-tensors
@@ -277,7 +279,8 @@ add_custom_command(
             -ownership-based-buffer-deallocation
             -buffer-deallocation-simplification
             -bufferization-lower-deallocations
-            -matmul-parallel-vectorization-optimize
+            -assume-tight-memref-layout
+            -matmul-vectorization-blis
             -batchmatmul-optimize
             -canonicalize
             -cse
@@ -292,7 +295,7 @@ add_custom_command(
             -canonicalize
             -cse
             -sccp
-            -remove-dead-values
+            # -remove-dead-values
             -memref-expand
             -arith-expand
             -convert-vector-to-llvm

examples/BuddyDeepSeekR1/import-deepseek-r1.py

@@ -38,6 +38,7 @@
     simply_fuse,
     apply_classic_fusion,
     eliminate_transpose,
+    flash_attention,
 )
 from buddy.compiler.graph.type import DeviceType
 from buddy.compiler.graph.operation import *
@@ -175,10 +176,11 @@
     params = dynamo_compiler_prefill.imported_params[graph_prefill]
     graphs_prefill[0].perform([eliminate_transpose])
     graphs_decode[0].perform([eliminate_transpose])
-    pattern_list = [simply_fuse]
+    pattern_list_prefill = [simply_fuse]
+    pattern_list_decode = [simply_fuse,flash_attention]
 
-    graphs_prefill[0].fuse_ops(pattern_list)
-    graphs_decode[0].fuse_ops(pattern_list)
+    graphs_prefill[0].fuse_ops(pattern_list_prefill)
+    graphs_decode[0].fuse_ops(pattern_list_decode)
 
     graph_prefill.op_groups["subgraph0_prefill"] = graph_prefill.op_groups.pop(
         "subgraph0"