Qualcomm AI Engine Direct - CI for QNN Static Stories Llama (#7884)

* Add Static Stories Llama CI

* Enable x86 runner for static llama, create a script for static llama ci

Author: winskuo-quic

.ci/scripts/test_qnn_static_llama.sh

@@ -0,0 +1,59 @@
+#!/bin/bash
+# Copyright (c) Qualcomm Innovation Center, Inc.
+# All rights reserved
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
+set -exu
+
+source "$(dirname "${BASH_SOURCE[0]}")/utils.sh"
+
+export EXECUTORCH_ROOT="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")/.." && pwd)"
+export QNN_SDK_ROOT=/tmp/qnn/2.28.0.241029
+export LD_LIBRARY_PATH="${QNN_SDK_ROOT}/lib/x86_64-linux-clang"
+export PYTHONPATH=".."
+cp schema/program.fbs exir/_serialize/program.fbs
+cp schema/scalar_type.fbs exir/_serialize/scalar_type.fbs
+cp -f build-x86/backends/qualcomm/PyQnnManagerAdaptor.cpython-310-x86_64-linux-gnu.so backends/qualcomm/python
+cp -f build-x86/backends/qualcomm/PyQnnWrapperAdaptor.cpython-310-x86_64-linux-gnu.so backends/qualcomm/python
+
+if [[ -z "${PYTHON_EXECUTABLE:-}" ]]; then
+  PYTHON_EXECUTABLE=python3
+fi
+
+which "${PYTHON_EXECUTABLE}"
+
+# Although static llama CI does not require graphviz, it is required by test_qnn_delegate.py
+pip install graphviz
+
+# Download stories llama110m artifacts
+download_stories_model_artifacts
+echo "Creating tokenizer.bin"
+$PYTHON_EXECUTABLE -m extension.llm.tokenizer.tokenizer -t tokenizer.model -o tokenizer.bin
+
+set +e
+# Compile only as weight sharing is not applicable on x86
+$PYTHON_EXECUTABLE backends/qualcomm/tests/test_qnn_delegate.py -k TestExampleScript.test_stories_single_llama --model SM8650 --build_folder build-android/ --executorch_root . --artifact_dir . --compile_only
+exit_code1=$?
+
+# Checks accuracy with weight sharing disabled since x86 does not support weight sharing.
+$PYTHON_EXECUTABLE backends/qualcomm/tests/test_qnn_delegate.py -k TestExampleScript.test_stories_single_llama --model SM8650 --build_folder build-x86/ --executorch_root . --artifact_dir . --enable_x86_64
+exit_code2=$?
+
+# Check the exit codes and print messages
+if [ $exit_code1 -ne 0 ]; then
+    echo "Static Llama compile only with weight sharing test failed. $exit_code1."
+fi
+
+if [ $exit_code2 -ne 0 ]; then
+    echo "Static Llama accuracy test failed. $exit_code2."
+fi
+
+# Return failure if either program failed
+if [ $exit_code1 -ne 0 ] || [ $exit_code2 -ne 0 ]; then
+    exit 1
+else
+    exit 0
+fi
+set -e

.github/workflows/pull.yml

@@ -437,6 +437,39 @@ jobs:
         # Test llama2
         PYTHON_EXECUTABLE=python bash .ci/scripts/test_llama.sh -model stories110M -build_tool "${BUILD_TOOL}" -mode "${MODE}" -dtype "${DTYPE}" -pt2e_quantize "${PT2E_QUANTIZE}"
 
+  test-static-llama-qnn-linux:
+    name: test-static-llama-qnn-linux
+    uses: pytorch/test-infra/.github/workflows/linux_job_v2.yml@main
+    permissions:
+      id-token: write
+      contents: read
+    strategy:
+      fail-fast: false
+    with:
+      runner: linux.2xlarge
+      docker-image: executorch-ubuntu-22.04-qnn-sdk
+      submodules: 'true'
+      ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
+      timeout: 180
+      script: |
+        # The generic Linux job chooses to use base env, not the one setup by the image
+        CONDA_ENV=$(conda env list --json | jq -r ".envs | .[-1]")
+        conda activate "${CONDA_ENV}"
+
+        BUILD_TOOL="cmake"
+
+        PYTHON_EXECUTABLE=python bash .ci/scripts/setup-qnn-deps.sh
+        PYTHON_EXECUTABLE=python bash .ci/scripts/build-qnn-sdk.sh
+
+        # Setup executorch
+        PYTHON_EXECUTABLE=python bash .ci/scripts/setup-linux.sh "${BUILD_TOOL}"
+
+        # Setup install_requirements for llama
+        PYTHON_EXECUTABLE=python bash examples/models/llama/install_requirements.sh
+        
+        # Test static llama weight sharing and accuracy
+        PYTHON_EXECUTABLE=python bash .ci/scripts/test_qnn_static_llama.sh
+
   test-qnn-models-linux:
     name: test-qnn-models-linux
     uses: pytorch/test-infra/.github/workflows/linux_job_v2.yml@main

backends/qualcomm/tests/test_qnn_delegate.py

@@ -2014,6 +2014,7 @@ def test_qnn_backend_multi_graphs(self):
                 soc_model=self.chipset_table[TestQNN.model],
                 backend_options=backend_options,
                 multiple_graphs=True,
+                weight_sharing=True,
                 graph_name=graph_name,
             )
             for graph_name in graph_names
@@ -2577,6 +2578,7 @@ def test_qnn_backend_multi_graphs(self):
                 soc_model=self.chipset_table[TestQNN.model],
                 backend_options=backend_options,
                 multiple_graphs=True,
+                weight_sharing=True,
                 graph_name=graph_name,
             )
             for graph_name in graph_names
@@ -3822,8 +3824,6 @@ def test_stories_single_llama(self):
             self.artifact_dir,
             "--build_folder",
             self.build_folder,
-            "--device",
-            self.device,
             "--model",
             self.model,
             "--checkpoint",
@@ -3846,9 +3846,21 @@ def test_stories_single_llama(self):
             "0",
             "--llama_model",
             "stories110m",
+            "--model_mode",
+            "hybrid",
+            "--prefill_seq_len",
+            "32",
+            "--kv_seq_len",
+            "128",
         ]
+        if self.compile_only:
+            cmds.extend(["--compile_only"])
+        elif self.device:
+            cmds.extend(["--device", self.device])
         if self.host:
             cmds.extend(["--host", self.host])
+        elif self.enable_x86_64:
+            cmds.extend(["--enable_x86_64"])
 
         golden_start_with = "Once upon a time,"
         p = subprocess.Popen(cmds, stdout=subprocess.DEVNULL)
@@ -3859,8 +3871,13 @@ def test_stories_single_llama(self):
             if "Error" in msg:
                 self.fail(msg["Error"])
             else:
-                model_out = msg["result"][0]
-                self.assertTrue(model_out.startswith(golden_start_with))
+                if not self.compile_only:
+                    model_out = msg["result"][0]
+                    self.assertTrue(model_out.startswith(golden_start_with))
+                # x86 does not allow weight sharing, so we don't check pte size
+                if not self.enable_x86_64:
+                    pte_size = msg["pte_size"]
+                    self.assertLessEqual(pte_size, 130000000)
 
     @unittest.skip("dynamic shape inputs appear in recent torch.export.export")
     def test_mobilebert(self):
@@ -4065,12 +4082,6 @@ def setup_environment():
         help="Path to open source software model repository",
         type=str,
     )
-    parser.add_argument(
-        "-x",
-        "--enable_x86_64",
-        help="Enable unittest to be executed on x86_64 platform",
-        action="store_true",
-    )
 
     args, ns_args = parser.parse_known_args(namespace=unittest)
     TestQNN.host = args.host
@@ -4089,6 +4100,8 @@ def setup_environment():
     TestQNN.shared_buffer = args.shared_buffer
     TestQNN.enable_x86_64 = args.enable_x86_64
     TestQNN.dump_intermediate_outputs = args.dump_intermediate_outputs
+    TestQNN.compile_only = args.compile_only
+
     return sys.argv[:1] + ns_args
 
 

backends/qualcomm/tests/utils.py

@@ -182,6 +182,7 @@ class TestQNN(unittest.TestCase):
     use_16a4w: str = "16a4w"
     shared_buffer: bool = False
     enable_x86_64: bool = False
+    compile_only: bool = False
 
     def _assert_outputs_equal(self, model_output, ref_output):
         self.assertTrue(len(ref_output) == len(model_output))

backends/qualcomm/utils/utils.py

@@ -1166,6 +1166,7 @@ def generate_qnn_executorch_compiler_spec(
     shared_buffer: bool = False,
     is_from_context_binary: bool = False,
     multiple_graphs: bool = False,
+    weight_sharing: bool = False,
     graph_name: str = "forward",
 ) -> List[CompileSpec]:
     """
@@ -1196,6 +1197,7 @@ def generate_qnn_executorch_compiler_spec(
         is_from_context_binary: True if current graph comes from pre-built context binary.
         multiple_graphs: True if multiple methods are expected to have in single .pte file.
             Please see test cases for post-processing example.
+        weight_sharing: Used with multiple_graphs, where model size will be reduced when operations have the same weights across multiple graphs.
         graph_name: Assign unique graph name if 'multiple_graphs' is used.
 
     Returns:
@@ -1216,6 +1218,12 @@ def generate_qnn_executorch_compiler_spec(
             stacklevel=1,
         )
 
+    if weight_sharing and not multiple_graphs:
+        warnings.warn(
+            "Weight sharing is intended for multiple graphs scenario, please ensure if there are multiple graphs",
+            stacklevel=1,
+        )
+
     qnn_executorch_options = QnnExecuTorchOptions(
         _soc_info_table[soc_model], backend_options
     )
@@ -1257,7 +1265,10 @@ def generate_qnn_executorch_compiler_spec(
 
     if multiple_graphs:
         # enable weight sharing mechanism if multiple graphs appear
-        if backend_options.backend_type == QnnExecuTorchBackendType.kHtpBackend:
+        if (
+            backend_options.backend_type == QnnExecuTorchBackendType.kHtpBackend
+            and weight_sharing
+        ):
             backend_options.htp_options.use_weight_sharing = True
 
     return [