Skip to content

Commit 09e65f5

Browse files
committed
Revert "Quacomm AI Engine Direct - Bump QNN SDK version from 2.37 to 2.50 (#2…"
This reverts commit adc0bcc.
1 parent 645387d commit 09e65f5

10 files changed

Lines changed: 80 additions & 76 deletions

File tree

‎.ci/scripts/test_llama.sh‎

Lines changed: 1 addition & 2 deletions
Original file line numberDiff line numberDiff line change
@@ -276,8 +276,7 @@ RESULT=$(cat result.txt)
276276
EXPECTED_PREFIX="Once upon a time,"
277277
# Expected result - may take too long to generate:
278278
# "Once upon a time, there was a little girl named Lily. She loved to play outside" ...
279-
# QNN libnative may print an initialization message to stdout before the generated text.
280-
if grep -qF -- "${EXPECTED_PREFIX}" result.txt; then
279+
if [[ "${RESULT}" == "${EXPECTED_PREFIX}"* ]]; then
281280
echo "Expected result prefix: ${EXPECTED_PREFIX}"
282281
echo "Actual result: ${RESULT}"
283282
echo "Success"

‎.github/workflows/android-release-artifacts.yml‎

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -168,7 +168,7 @@ jobs:
168168
source backends/qualcomm/scripts/qnn_config.sh
169169
export QNN_SDK_ROOT="/tmp/qnn/${QNN_VERSION}"
170170
export ANDROID_ABIS=arm64-v8a
171-
GRADLE_ARGS+=" -DqnnVersion=2.50.0"
171+
GRADLE_ARGS+=" -DqnnVersion=2.37.0"
172172
fi
173173
174174
# Build AAR Package

‎backends/qualcomm/scripts/download_qnn_sdk.py‎

Lines changed: 2 additions & 2 deletions
Original file line numberDiff line numberDiff line change
@@ -72,7 +72,7 @@ def _read_qnn_config() -> Dict[str, str]:
7272

7373

7474
_QNN_CONFIG = _read_qnn_config()
75-
QNN_VERSION = _QNN_CONFIG.get("QNN_VERSION", "2.50.0.260828")
75+
QNN_VERSION = _QNN_CONFIG.get("QNN_VERSION", "2.37.0.250724")
7676
QNN_ZIP_URL = _QNN_CONFIG.get(
7777
"QNN_ZIP_URL",
7878
f"https://softwarecenter.qualcomm.com/api/download/software/sdks/" # @lint-ignore only half of a URL, the rest is on the next line
@@ -81,7 +81,7 @@ def _read_qnn_config() -> Dict[str, str]:
8181

8282

8383
def _get_sdk_dir() -> pathlib.Path:
84-
"""Get the versioned SDK cache directory (e.g. ~/.cache/executorch/qnn/sdk-2.50.0.260828/)."""
84+
"""Get the versioned SDK cache directory (e.g. ~/.cache/executorch/qnn/sdk-2.37.0.250724/)."""
8585
try:
8686
return _get_staging_dir(f"sdk-{QNN_VERSION}")
8787
except ValueError:

‎backends/qualcomm/scripts/qnn_config.sh‎

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -6,7 +6,7 @@
66
# LICENSE file in the root directory of this source tree.
77

88
# QNN SDK Configuration
9-
QNN_VERSION="2.50.0.260828"
9+
QNN_VERSION="2.37.0.250724"
1010
QNN_ZIP_URL="https://softwarecenter.qualcomm.com/api/download/software/sdks/Qualcomm_AI_Runtime_Community/All/${QNN_VERSION}/v${QNN_VERSION}.zip"
1111

1212
# Hexagon SDK Configuration (used only by direct-mode CI build).

‎backends/qualcomm/tests/test_qnn_delegate.py‎

Lines changed: 51 additions & 43 deletions
Original file line numberDiff line numberDiff line change
@@ -1116,17 +1116,16 @@ def test_qnn_backend_expm1(self):
11161116
self.lower_module_and_test_output(module, sample_input)
11171117

11181118
def test_qnn_backend_fp16a8w_conv2d(self):
1119-
# fp16a8w: FP16 activation + INT8 weight; weight kernel must be [1,1],
1120-
# in channel must be multiple of 32/bw = 4
1119+
# fp16a8w: FP16 activation + INT8 weight; weight kernel must be [1,1]
11211120
modules = [
11221121
Conv2dSingle( # noqa: F405
1123-
in_channel=4, out_channel=4, kernel_size=1, padding=0
1122+
in_channel=2, out_channel=4, kernel_size=1, padding=0
11241123
),
11251124
Conv2dSingle( # noqa: F405
1126-
in_channel=4, out_channel=4, kernel_size=1, padding=0, bias=False
1125+
in_channel=2, out_channel=4, kernel_size=1, padding=0, bias=False
11271126
),
11281127
]
1129-
sample_input = (torch.randn([1, 4, 3, 3]),)
1128+
sample_input = (torch.randn([1, 2, 3, 3]),)
11301129
for i, module in enumerate(modules):
11311130
with self.subTest(i=i):
11321131
module = self.get_qdq_module(
@@ -1137,16 +1136,15 @@ def test_qnn_backend_fp16a8w_conv2d(self):
11371136
def test_qnn_backend_fp16a8w_conv2d_qat(self):
11381137
# fp16a8w QAT: FP16 activation + INT8 weight; weight kernel must be [1,1]
11391138
# QAT fake quantize (FusedMovingAvgObsFakeQuantize) requires float32 tensors,
1140-
# in channel must be multiple of 32/bw = 4
11411139
modules = [
11421140
Conv2dSingle( # noqa: F405
1143-
in_channel=4, out_channel=4, kernel_size=1, padding=0
1141+
in_channel=2, out_channel=4, kernel_size=1, padding=0
11441142
),
11451143
Conv2dSingle( # noqa: F405
1146-
in_channel=4, out_channel=4, kernel_size=1, padding=0, bias=False
1144+
in_channel=2, out_channel=4, kernel_size=1, padding=0, bias=False
11471145
),
11481146
]
1149-
sample_input = (torch.randn([1, 4, 3, 3]),)
1147+
sample_input = (torch.randn([1, 2, 3, 3]),)
11501148
for i, module in enumerate(modules):
11511149
with self.subTest(i=i):
11521150
# QAT in float32
@@ -1273,7 +1271,8 @@ def test_qnn_backend_gather(self):
12731271
Gather(), # noqa: F405
12741272
# TODO: resolve accuracy problem
12751273
# GatherArgmin(), # noqa: F405
1276-
GatherWhere(), # noqa: F405
1274+
# TODO: There is a accuracy regression after 2.37
1275+
# GatherWhere(), # noqa: F405
12771276
]
12781277
# shape = (2, 2, 3, 4)
12791278
sample_inputs = [
@@ -1685,10 +1684,10 @@ def test_qnn_backend_is_nan(self):
16851684
float("nan"),
16861685
-float("nan"),
16871686
0.2,
1688-
# float("inf"), # inf is treat as nan in QNN2.50
1687+
float("inf"),
16891688
3.2,
16901689
float("nan"),
1691-
# -float("inf"), # inf is treat as nan in QNN2.50
1690+
-float("inf"),
16921691
],
16931692
dtype=torch.float32,
16941693
),
@@ -2806,14 +2805,15 @@ def test_qnn_backend_where(self):
28062805
Where(), # noqa: F405
28072806
WhereConstant(torch.randn(3, 2), torch.randn(3, 2)), # noqa: F405
28082807
WhereConstantOther(), # noqa: F405
2809-
WhereConstantAll(), # noqa: F405
2808+
# TODO: There is a accuracy regression after 2.37
2809+
# WhereConstantAll(), # noqa: F405
28102810
WhereConstantInf(), # noqa: F405
28112811
]
28122812
sample_inputs = [
28132813
(torch.randn(3, 2), torch.randn(3, 2), torch.randn(3, 2)),
28142814
(torch.randn(3, 2),),
28152815
(torch.randn(3, 2),),
2816-
(torch.randn(3, 2),),
2816+
# (torch.randn(3, 2),),
28172817
(torch.randn(30, 20),),
28182818
]
28192819
for i, module in enumerate(modules):
@@ -2877,6 +2877,11 @@ def setUp(self):
28772877
shared_buffer=TestQNN.shared_buffer,
28782878
)
28792879

2880+
# TODO: Needs to be fixed in HTP
2881+
@unittest.skipIf(
2882+
is_qnn_sdk_version_greater_than("2.37"),
2883+
"Failed to prepare the graph because of an index operation with argmin output.",
2884+
)
28802885
def test_qnn_backend_argmin_view_squeeze_conv2d(self):
28812886
module = ArgminViewSqueezeConv2D() # noqa: F405
28822887
sample_input = (torch.randn(32), torch.randn(32, 3, 32, 32))
@@ -2916,6 +2921,11 @@ def test_qnn_backend_conv2d_avg_pool2d(self):
29162921
sample_input = (torch.randn(16, 3, 16, 16),)
29172922
self.lower_module_and_test_output(module, sample_input)
29182923

2924+
# TODO: Needs to be fixed in HTP
2925+
@unittest.skipIf(
2926+
is_qnn_sdk_version_greater_than("2.40"),
2927+
"UT did not pass because of aten.mean.dim when using keep_dim for some devices after QNN 2.41.",
2928+
)
29192929
def test_qnn_backend_conv2d_bn_hardtanh_mean(self):
29202930
module = Conv2dBnHardtanhMean() # noqa: F405
29212931
sample_input = (torch.randn(1, 1, 6, 6),)
@@ -3913,7 +3923,7 @@ def test_qnn_backend_conv_transpose2d(self):
39133923
gm = self.get_qdq_module(module, sample_input)
39143924
self.lower_module_and_test_output(gm, sample_input)
39153925

3916-
@unittest.skip("As of QNN 2.50, transpose conv block quant is not supported")
3926+
@unittest.skip("As of QNN 2.37, transpose conv block quant is not supported")
39173927
def test_qnn_backend_conv_transpose2d_block(self):
39183928
i_ch, o_ch, kernel, padding = 128, 32, (1, 1), 0
39193929
modules = [
@@ -4437,7 +4447,8 @@ def test_qnn_backend_gather(self):
44374447
Gather(), # noqa: F405
44384448
# TODO: resolve accuracy problem
44394449
# GatherArgmin(), # noqa: F405
4440-
GatherWhere(), # noqa: F405
4450+
# TODO: There is a accuracy regression after 2.37
4451+
# GatherWhere(), # noqa: F405
44414452
]
44424453
# shape = (2, 2, 3, 4)
44434454
sample_inputs = [
@@ -6429,14 +6440,15 @@ def test_qnn_backend_where(self):
64296440
Where(), # noqa: F405
64306441
WhereConstant(torch.randn(3, 2), torch.randn(3, 2)), # noqa: F405
64316442
WhereConstantOther(), # noqa: F405
6432-
WhereConstantAll(), # noqa: F405
6443+
# TODO: There is a accuracy regression after 2.37
6444+
# WhereConstantAll(), # noqa: F405
64336445
WhereConstantInf(), # noqa: F405
64346446
]
64356447
sample_inputs = [
64366448
(torch.randn(3, 2), torch.randn(3, 2), torch.randn(3, 2)),
64376449
(torch.randn(3, 2),),
64386450
(torch.randn(3, 2),),
6439-
(torch.randn(3, 2),),
6451+
# (torch.randn(3, 2),),
64406452
(torch.randn(30, 20),),
64416453
]
64426454
for i, module in enumerate(modules):
@@ -6830,6 +6842,7 @@ def test_qnn_backend_masked_softmax(self):
68306842
has_masked_softmax = True
68316843
self.assertTrue(has_masked_softmax)
68326844

6845+
@unittest.skip("UT pass before QNN 2.26, segfault during partitioner")
68336846
def test_qnn_backend_moe_feed_forward(self):
68346847
from executorch.examples.models.llama.llama_transformer import MOEFeedForward
68356848
from executorch.examples.models.llama.model_args import ModelArgs
@@ -9014,7 +9027,7 @@ def setUp(self):
90149027
SM8650=32,
90159028
SM8750=36,
90169029
pte_size=2_700_000_000, # 2.7 GB
9017-
wikitext_ppl=19,
9030+
wikitext_ppl=17,
90189031
hellaswag_acc_norm=None,
90199032
sqnr=27,
90209033
),
@@ -9024,13 +9037,13 @@ def setUp(self):
90249037
pte_size=2_860_000_000, # 2.86 GB
90259038
wikitext_ppl=14,
90269039
hellaswag_acc_norm=None,
9027-
sqnr=20,
9040+
sqnr=27,
90289041
),
90299042
"gemma3-1b": TestExampleLLMScript.LlmSpecs(
9030-
SM8650=68,
9031-
SM8750=72,
9043+
SM8650=70,
9044+
SM8750=100,
90329045
pte_size=1_200_000_000, # 1.2 GB
9033-
wikitext_ppl=24,
9046+
wikitext_ppl=23,
90349047
hellaswag_acc_norm=None,
90359048
sqnr=10,
90369049
),
@@ -9040,7 +9053,7 @@ def setUp(self):
90409053
pte_size=4_500_000_000, # 4.5 GB
90419054
wikitext_ppl=120,
90429055
hellaswag_acc_norm=None,
9043-
sqnr=9,
9056+
sqnr=10,
90449057
),
90459058
"glm-1_5b": TestExampleLLMScript.LlmSpecs(
90469059
SM8650=42,
@@ -9064,15 +9077,15 @@ def setUp(self):
90649077
pte_size=4_000_000_000, # 4GB
90659078
wikitext_ppl=14,
90669079
hellaswag_acc_norm=None,
9067-
sqnr=2,
9080+
sqnr=20,
90689081
),
90699082
"llama3_2-1b_instruct": TestExampleLLMScript.LlmSpecs(
90709083
SM8650=37,
90719084
SM8750=45,
90729085
pte_size=1_500_000_000, # 1.5 GB
90739086
wikitext_ppl=18,
90749087
hellaswag_acc_norm=None,
9075-
sqnr=13,
9088+
sqnr=15,
90769089
),
90779090
"llama3_2-3b_instruct": TestExampleLLMScript.LlmSpecs(
90789091
SM8650=21,
@@ -9083,20 +9096,20 @@ def setUp(self):
90839096
sqnr=14,
90849097
),
90859098
"qwen2_5-0_5b": TestExampleLLMScript.LlmSpecs(
9086-
SM8650=95,
9087-
SM8750=130,
9099+
SM8650=115,
9100+
SM8750=155,
90889101
pte_size=600_000_000, # 600 MB
90899102
wikitext_ppl=15,
90909103
hellaswag_acc_norm=None,
90919104
sqnr=8,
90929105
),
90939106
"qwen2_5-1_5b": TestExampleLLMScript.LlmSpecs(
90949107
SM8650=38,
9095-
SM8750=45,
9108+
SM8750=47,
90969109
pte_size=1_500_000_000, # 1.5 GB
90979110
wikitext_ppl=10,
90989111
hellaswag_acc_norm=None,
9099-
sqnr=9.5,
9112+
sqnr=10,
91009113
),
91019114
"qwen3-0_6b": TestExampleLLMScript.LlmSpecs(
91029115
SM8650=47,
@@ -9110,9 +9123,9 @@ def setUp(self):
91109123
SM8650=28,
91119124
SM8750=34,
91129125
pte_size=1_800_000_000, # 1.8 GB
9113-
wikitext_ppl=20,
9126+
wikitext_ppl=15,
91149127
hellaswag_acc_norm=None,
9115-
sqnr=11.5,
9128+
sqnr=12,
91169129
),
91179130
"smollm2_135m": TestExampleLLMScript.LlmSpecs(
91189131
SM8650=214,
@@ -9138,11 +9151,6 @@ def test_static_llm_model(self): # noqa: C901
91389151
assert (
91399152
self.model_name in self.llm_specs
91409153
), f"Unable to find {self.model_name} under model_specs."
9141-
if (
9142-
self.model_name == "granite_3_3-2b_instruct"
9143-
and is_qnn_sdk_version_greater_than("2.49")
9144-
):
9145-
self.skipTest("The model crush in dsp side since 2.50, skipped")
91469154

91479155
is_llama_model = self.model_name in {
91489156
"llama3_2-1b_instruct",
@@ -9441,7 +9449,7 @@ def test_codegen2_1b(self):
94419449
pte_size = msg["pte_size"]
94429450
self.assertLessEqual(pte_size, 1_200_000_000) # 1200MB
94439451
if not self.compile_only and not self.enable_x86_64:
9444-
self.assertGreaterEqual(msg["inference_speed"], 50) # Lanai
9452+
self.assertGreaterEqual(msg["inference_speed"], 60)
94459453

94469454
def test_llama_stories_260k(self):
94479455
if not self.required_envs():
@@ -9632,8 +9640,8 @@ def test_attention_sink(self):
96329640
else:
96339641
if not self.compile_only:
96349642
self.assertLessEqual(
9635-
msg["attention_sink_evictor_pte_size"], 1_850_000
9636-
) # 1.85 MB
9643+
msg["attention_sink_evictor_pte_size"], 1_700_000
9644+
) # 1.7 MB
96379645
self.assertLessEqual(
96389646
msg["wiki_ppl"], self.llm_specs[model_name].wikitext_ppl
96399647
)
@@ -9785,7 +9793,7 @@ def setUp(self):
97859793
self.alm_specs = {
97869794
"granite_speech_3_3-2b": TestExampleMultimodalityScript.ALMSpecs(
97879795
max_seq_len=1024,
9788-
sm8650_token_rate=4,
9796+
sm8650_token_rate=5,
97899797
sm8750_token_rate=8,
97909798
encoder_pte_size=900_000_000, # 900MB
97919799
tok_embedding_pte_size=240_000_000, # 240MB
@@ -9797,8 +9805,8 @@ def setUp(self):
97979805
self.vlm_specs = {
97989806
"smolvlm_500m_instruct": TestExampleMultimodalityScript.VLMSpecs(
97999807
max_seq_len=1024,
9800-
sm8650_token_rate=37,
9801-
sm8750_token_rate=40,
9808+
sm8650_token_rate=50,
9809+
sm8750_token_rate=55,
98029810
encoder_pte_size=110_000_000, # 110MB
98039811
tok_embedding_pte_size=100_000_000, # 100MB
98049812
decoder_pte_size=400_000_000, # 400MB

‎docs/source/backends-qualcomm.md‎

Lines changed: 19 additions & 22 deletions
Original file line numberDiff line numberDiff line change
@@ -79,30 +79,27 @@ The target SoC must be one of those listed in the `QcomChipset` enum; see [qc_sc
7979

8080
[Qualcomm AI Engine Direct SDK](https://developer.qualcomm.com/software/qualcomm-ai-engine-direct-sdk)
8181
- Click the "Get Software" button to download the latest version of the QNN SDK.
82-
- Although newer versions are available, we have verified and recommend using QNN 2.50.0 for stability.
83-
- You can download it directly from the following link: [QNN 2.50.0](https://softwarecenter.qualcomm.com/api/download/software/sdks/Qualcomm_AI_Runtime_Community/All/2.50.0.260828/v2.50.0.260828.zip)
82+
- Although newer versions are available, we have verified and recommend using QNN 2.37.0 for stability.
83+
- You can download it directly from the following link: [QNN 2.37.0](https://softwarecenter.qualcomm.com/api/download/software/sdks/Qualcomm_AI_Runtime_Community/All/2.37.0.250724/v2.37.0.250724.zip)
8484

8585
The directory with installed Qualcomm AI Engine Direct SDK looks like:
8686
```
87-
|-- GENIE_README.txt
88-
|-- LICENSE.pdf
89-
|-- NOTICE.txt
90-
|-- NOTICE_WINDOWS.txt
91-
|-- QAIRT_ReleaseNotes.txt
92-
|-- QNN_NOTICE.txt
93-
|-- QNN_README.txt
94-
|-- QNN_TFLITE_DELEGATE_NOTICE.txt
95-
|-- QNN_TFLITE_DELEGATE_README.txt
96-
|-- QNN_TFLITE_DELEGATE_ReleaseNotes.txt
97-
|-- benchmarks
98-
|-- bin
99-
|-- docs
100-
|-- examples
101-
|-- include
102-
|-- lib
103-
|-- lib-safe
104-
|-- sdk.yaml
105-
`-- share
87+
├── benchmarks
88+
├── bin
89+
├── docs
90+
├── examples
91+
├── include
92+
├── lib
93+
├── LICENSE.pdf
94+
├── NOTICE.txt
95+
├── NOTICE_WINDOWS.txt
96+
├── QNN_NOTICE.txt
97+
├── QNN_README.txt
98+
├── QNN_ReleaseNotes.txt
99+
├── ReleaseNotes.txt
100+
├── ReleaseNotesWindows.txt
101+
├── sdk.yaml
102+
└── share
106103
```
107104

108105
On Android / Linux devices:
@@ -495,7 +492,7 @@ cd $DEMO_APP
495492
unzip -l app/build/outputs/apk/debug/app-debug.apk | grep "libQnnHtp.so"
496493
```
497494

498-
Expected size for QNN 2.50.0: ~2,601,473,189 bytes
495+
Expected size for QNN 2.37.0: ~2,465,440 bytes
499496

500497
***Step 3***. Monitor Logs During Model Loading
501498

0 commit comments

Comments
 (0)