#12330: Removed the deprecated flash decode gqa op as its functionali…

…ty is covered in the generic flash decode op
tenstorrent · Oct 22, 2024 · a7ba1ad · a7ba1ad
1 parent 044a345
commit a7ba1ad
Show file tree

Hide file tree

Showing 10 changed files with 1 addition and 821 deletions.
diff --git a/models/demos/llama3/tt/llama_attention.py b/models/demos/llama3/tt/llama_attention.py
@@ -279,19 +279,17 @@ def forward_decode(
                 values,
                 cur_pos_tensor=current_pos,
                 page_table_tensor=page_table,
-                transpose_q=False,
                 scale=self.scale,
                 program_config=self.model_config["SDPA_DECODE_PROGCFG"],
                 compute_kernel_config=self.model_config["SDPA_DECODE_COMPUTE_PROGCFG"],
                 memory_config=ttnn.DRAM_MEMORY_CONFIG,
             )
         else:
-            attn_output_1G4D = ttnn.transformer.scaled_dot_product_attention_decode_gqa(
+            attn_output_1G4D = ttnn.transformer.scaled_dot_product_attention_decode(
                 q_heads_1BQD,
                 keys,
                 values,
                 cur_pos_tensor=current_pos,
-                transpose_q=False,
                 scale=self.scale,
                 program_config=self.model_config["SDPA_DECODE_PROGCFG"],
                 compute_kernel_config=self.model_config["SDPA_DECODE_COMPUTE_PROGCFG"],

diff --git a/...ager/python_api_testing/unit_testing/misc/test_scaled_dot_product_attention_decode_gqa.py b/...ager/python_api_testing/unit_testing/misc/test_scaled_dot_product_attention_decode_gqa.py
diff --git a/ttnn/CMakeLists.txt b/ttnn/CMakeLists.txt
@@ -348,9 +348,6 @@ set(ALL_TTNN_SRCS
     ${CMAKE_CURRENT_SOURCE_DIR}/cpp/ttnn/operations/transformer/sdpa_decode/sdpa_decode_pybind.cpp
     ${CMAKE_CURRENT_SOURCE_DIR}/cpp/ttnn/operations/transformer/sdpa_decode/device/sdpa_decode_op.cpp
     ${CMAKE_CURRENT_SOURCE_DIR}/cpp/ttnn/operations/transformer/sdpa_decode/device/sdpa_decode_program_factory.cpp
-    ${CMAKE_CURRENT_SOURCE_DIR}/cpp/ttnn/operations/transformer/sdpa_decode_gqa/sdpa_decode_gqa.cpp
-    ${CMAKE_CURRENT_SOURCE_DIR}/cpp/ttnn/operations/transformer/sdpa_decode_gqa/sdpa_decode_gqa_pybind.cpp
-    ${CMAKE_CURRENT_SOURCE_DIR}/cpp/ttnn/operations/transformer/sdpa_decode_gqa/device/sdpa_decode_gqa_op.cpp
     ${CMAKE_CURRENT_SOURCE_DIR}/cpp/ttnn/operations/data_movement/sharded/interleaved_to_sharded/interleaved_to_sharded.cpp
     ${CMAKE_CURRENT_SOURCE_DIR}/cpp/ttnn/operations/data_movement/sharded/interleaved_to_sharded/interleaved_to_sharded_pybind.cpp
     ${CMAKE_CURRENT_SOURCE_DIR}/cpp/ttnn/operations/data_movement/sharded/interleaved_to_sharded/device/interleaved_to_sharded_op.cpp