From e5b996affaa024754bc10e1c33d85dff3a3628ca Mon Sep 17 00:00:00 2001
From: Tarun Karuturi <tkaruturi@fb.com>
Date: Wed, 9 Apr 2025 17:14:16 -0700
Subject: [PATCH] Adding qualcomm recipes

Differential Revision: [D72410492](https://our.internmc.facebook.com/intern/diff/D72410492/)

[ghstack-poisoned]
---
 backends/qualcomm/recipes/TARGETS    | 12 +++++
 backends/qualcomm/recipes/recipes.py | 71 ++++++++++++++++++++++++++++
 extension/llm/export/TARGETS         |  1 +
 3 files changed, 84 insertions(+)
 create mode 100644 backends/qualcomm/recipes/TARGETS
 create mode 100644 backends/qualcomm/recipes/recipes.py

diff --git a/backends/qualcomm/recipes/TARGETS b/backends/qualcomm/recipes/TARGETS
new file mode 100644
index 00000000000..d2ee8b1e684
--- /dev/null
+++ b/backends/qualcomm/recipes/TARGETS
@@ -0,0 +1,12 @@
+load("@fbcode_macros//build_defs:python_library.bzl", "python_library")
+
+python_library(
+    name = "recipes",
+    srcs = ["recipes.py"],
+    deps = [
+        "//caffe2:torch",
+        "//executorch/exir:lib",
+        "//executorch/backends/qualcomm/quantizer:quantizer",
+        "//executorch/examples/qualcomm:utils"
+    ]
+)
diff --git a/backends/qualcomm/recipes/recipes.py b/backends/qualcomm/recipes/recipes.py
new file mode 100644
index 00000000000..0e5540455b3
--- /dev/null
+++ b/backends/qualcomm/recipes/recipes.py
@@ -0,0 +1,71 @@
+# (c) Meta Platforms, Inc. and affiliates. Confidential and proprietary.
+
+# pyre-strict
+from typing import Optional
+
+from executorch import exir
+from executorch.backends.qualcomm.quantizer.quantizer import QuantDtype
+
+from executorch.examples.qualcomm.utils import make_quantizer
+from executorch.exir import EdgeCompileConfig, ExecutorchBackendConfig, ExportRecipe
+from executorch.exir.passes import MemoryPlanningPass
+from torch.ao.quantization.quantizer import Quantizer
+
+def get_qualcomm_htp_et_recipe(
+    name: str,
+    soc_model: str = "SM8650",
+    qnn_version: str = "2.25",
+    quant_dtype: Optional[QuantDtype] = None,
+    shared_buffer=False,
+    skip_node_id_set: set[int] = set(),
+    skip_node_op_set: set[str] = set(),
+) -> ExportRecipe:
+    from executorch.backends.qualcomm.partition.qnn_partitioner import QnnPartitioner
+    from executorch.backends.qualcomm.serialization.qc_schema import QcomChipset
+
+    from executorch.backends.qualcomm.utils.utils import (
+        _transform,
+        generate_htp_compiler_spec,
+        generate_qnn_executorch_compiler_spec,
+    )
+
+    if quant_dtype:
+        qnn_quantizer: Quantizer  = make_quantizer(quant_dtype=quant_dtype)
+    else:
+        qnn_quantizer = make_quantizer()
+
+    qnn_partitioner = QnnPartitioner(
+        generate_qnn_executorch_compiler_spec(
+            soc_model=getattr(QcomChipset, soc_model),
+            backend_options=generate_htp_compiler_spec(
+                use_fp16=False if quant_dtype else True
+            ),
+        ),
+        skip_node_id_set=skip_node_id_set,
+        skip_node_op_set=skip_node_op_set,
+    )
+    
+    executorch_config = ExecutorchBackendConfig(
+        # For shared buffer, user must pass the memory address
+        # which is allocated by RPC memory to executor runner.
+        # Therefore, won't want to pre-allocate
+        # by memory manager in runtime.
+        memory_planning_pass=MemoryPlanningPass(
+            alloc_graph_input=not shared_buffer,
+            alloc_graph_output=not shared_buffer,
+        ),
+    )
+
+    return ExportRecipe(
+        name,
+        quantizer=qnn_quantizer,
+        partitioners=[qnn_partitioner],
+        pre_edge_transform_passes=_transform,
+        edge_compile_config=EdgeCompileConfig(
+            _check_ir_validity=False,
+            _skip_dim_order=True,
+        ),
+        edge_transform_passes=[],
+        transform_check_ir_validity=True,
+        executorch_backend_config=executorch_config,
+    )
diff --git a/extension/llm/export/TARGETS b/extension/llm/export/TARGETS
index 40f8599e9e0..35ff63ac3a6 100644
--- a/extension/llm/export/TARGETS
+++ b/extension/llm/export/TARGETS
@@ -31,6 +31,7 @@ runtime.python_library(
         "//executorch/backends/apple/mps:partitioner",
         "//executorch/backends/qualcomm/partition:partition",
         "//executorch/backends/qualcomm/quantizer:quantizer",
+        "//executorch/backends/qualcomm/recipes:recipes",
         "//executorch/backends/transforms:duplicate_dynamic_quant_chain",
         "//executorch/backends/vulkan/partitioner:vulkan_partitioner",
         "//executorch/backends/vulkan/quantizer:vulkan_quantizer",