add support SA8397 SoC (automotive) in executorch Qualcomm backend

[ares89]

🚀 The feature, motivation and pitch

Request: add support for SA8397 in executorch's Qualcomm backend so scripts accept -m SA8397 and generate correct QNN/HTP artifacts.
Current problem: using -m SA8397 fails.
Desired: add SA8397 to QcomChipset enum and to the soc->chipset map, and update any chipset-specific compiler/runtime options (can mirror SA8295 defaults initially). Add a basic CI/integration test that compiles a small model with -m SA8397.

Alternatives

No response

Additional context

No response

RFC (Optional)

No response

cc @cccclai @winskuo-quic @shewu-quic @haowhsu-quic @DannyYuyang-quic @cbilgin

[cccclai]

That should be easy - will you be willing to add the SoC? There are some other similar PRs contributed from external users like #16331 and #16332

[ares89]

I'll give it a try. If it works, I'll submit a pull request.

[cccclai]

Sounds good! Once you have it, feel free to tag me to review. If you run into issues, let us know

[ares89]

I made some change like these (SA8397 is the same as SA8797):

diff --git a/backends/qualcomm/serialization/qc_compiler_spec.fbs b/backends/qualcomm/serialization/qc_compiler_spec.fbs
index 7b81711b58..0d66d7d683 100644
--- a/backends/qualcomm/serialization/qc_compiler_spec.fbs
+++ b/backends/qualcomm/serialization/qc_compiler_spec.fbs
@@ -18,6 +18,7 @@ enum HtpArch: int {
   V73 = 73,
   V75 = 75,
   V79 = 79,
+  V81 = 81,
 }
 
 table HtpInfo {
@@ -42,7 +43,7 @@ enum QcomChipset: int {
   SSG2125P = 58,
   SXR1230P = 45,
   SXR2230P = 53,
-  SA8797 =72,
+  SA8797 = 72,
   SXR2330P = 75,
 }
 

commit 5d48672a7ee956315e9e32569538fe59fa4b7923
Author: ares89 <ares89@126.com>
Date:   Tue Jan 13 03:41:50 2026 +0000

    add support for SA8797(SA8397)

diff --git a/backends/qualcomm/serialization/qc_compiler_spec.fbs b/backends/qualcomm/serialization/qc_compiler_spec.fbs
index 8aeaa060a5..7b81711b58 100644
--- a/backends/qualcomm/serialization/qc_compiler_spec.fbs
+++ b/backends/qualcomm/serialization/qc_compiler_spec.fbs
@@ -42,6 +42,7 @@ enum QcomChipset: int {
   SSG2125P = 58,
   SXR1230P = 45,
   SXR2230P = 53,
+  SA8797 =72,
   SXR2330P = 75,
 }
 
diff --git a/backends/qualcomm/serialization/qc_schema.py b/backends/qualcomm/serialization/qc_schema.py
index f3b9e2cc1a..fabee6f6d5 100644
--- a/backends/qualcomm/serialization/qc_schema.py
+++ b/backends/qualcomm/serialization/qc_schema.py
@@ -27,6 +27,7 @@ class HtpArch(IntEnum):
     V73 = 73
     V75 = 75
     V79 = 79
+    V81 = 81
 
 
 @dataclass
@@ -48,6 +49,7 @@ class QcomChipset(IntEnum):
     SSG2125P = 58  # v73
     SXR1230P = 45  # v73
     SXR2230P = 53  # v69
+    SA8797 = 72  # v81
     SXR2330P = 75  # v79
 
 
@@ -69,6 +71,7 @@ _soc_info_table = {
     QcomChipset.SXR1230P: SocInfo(QcomChipset.SXR1230P, HtpInfo(HtpArch.V73, 2)),
     QcomChipset.SXR2230P: SocInfo(QcomChipset.SXR2230P, HtpInfo(HtpArch.V69, 8)),
     QcomChipset.SXR2330P: SocInfo(QcomChipset.SXR2330P, HtpInfo(HtpArch.V79, 8)),
+    QcomChipset.SA8797: SocInfo(QcomChipset.SA8797, HtpInfo(HtpArch.V81, 16)),
 }
 
 
diff --git a/backends/qualcomm/utils/utils.py b/backends/qualcomm/utils/utils.py
index be4e86de50..0204fab178 100644
--- a/backends/qualcomm/utils/utils.py
+++ b/backends/qualcomm/utils/utils.py
@@ -1089,6 +1089,7 @@ def generate_qnn_executorch_compiler_spec(
 def get_soc_to_arch_map():
     return {
         "SA8295": HtpArch.V68,
+        "SA8797": HtpArch.V81,
         "SM8450": HtpArch.V69,
         "SM8475": HtpArch.V69,
         "SM8550": HtpArch.V73,
@@ -1105,6 +1106,7 @@ def get_soc_to_arch_map():
 def get_soc_to_chipset_map():
     return {
         "SA8295": QcomChipset.SA8295,
+        "SA8797": QcomChipset.SA8797,
         "SM8450": QcomChipset.SM8450,
         "SM8475": QcomChipset.SM8475,
         "SM8550": QcomChipset.SM8550,

when i run like this:
nohup python -m examples.qualcomm.oss_scripts.qwen2_5.qwen2_5 -b build-android -m SA8797 --compile_only --decoder_model qwen2.5_0.5B --prompt 你好啊 1>logs/qwen25_8797.log 2>&1 &

the log is full of this:

[ERROR] [Qnn ExecuTorch]: The SocModel doesn't support FP16
[ERROR] [Qnn ExecuTorch]: QnnBackend_validateOpConfig failed 3110
[ERROR] [Qnn ExecuTorch]: Failed to validate op aten_rms_norm_default_48 with error 0xc26
[WARNING] [Qnn ExecuTorch]: Qnn Backend op validation failed with error: 3110

qwen25_8797.log

[ares89]

Also in this file，

executorch/backends/qualcomm/README.md

Line 38 in 064ef7c

Currently, users cannot add additional chipset models because the chipset ID is not accessible to community users. If you have specific chipset models you wish to add, please contact one of the authors in the `Code Reviews` section at the bottom of this page.

Are there differences between QNN(for mobile phone) and QNN(for autoware)?

[shewu-quic]

qwen2_5

Hi @ares89 ,
This device seems not support FP16 graph based on the log. Could you please try to add --ptq 16a8w in your command?
BTW, if you would like to get the better performance, you could try static llama. https://github.com/pytorch/executorch/tree/064ef7ca0049d483429412e55d1be53e1835d525/examples/qualcomm/oss_scripts/llama

End to End Commanad:

python examples/qualcomm/oss_scripts/llama/llama.py -b build-android -s ${SERIAL_NUM} -m ${SOC_MODEL} --temperature 0 --model_mode hybrid --max_seq_len 1024 --prefill_ar_len 128 --decoder_model qwen2_5-0_5b --prompt "I would like to learn python, could you teach me with a simple example?" --tasks wikitext --limit 1

[ares89]

Hi @shewu-quic,
I tried this command,
python examples/qualcomm/oss_scripts/llama/llama.py -b build-android -m **SA8797** --temperature 0 --model_mode hybrid --max_seq_len 1024 --prefill_ar_len 128 --decoder_model qwen3-0_6b --compile_only --prompt "你好啊"
and there were some errors in the log.

qwen3_8797_0115_04.log

[WARNING] [Qnn ExecuTorch]: Qnn Backend op validation failed with error: 3110
[ERROR] [Qnn ExecuTorch]: [4294967295] has incorrect Value 0, expected >= 69.
[ERROR] [Qnn ExecuTorch]: QnnBackend_validateOpConfig failed 3110
[ERROR] [Qnn ExecuTorch]: Failed to validate op aten_convolution_default_1005 with error 0xc26

[WARNING] [Qnn ExecuTorch]: Qnn Backend op validation failed with error: 3110
[ERROR] [Qnn ExecuTorch]: [4294967295] has incorrect Value 0, expected >= 73.
[ERROR] [Qnn ExecuTorch]: QnnBackend_validateOpConfig failed 3110
[ERROR] [Qnn ExecuTorch]: Failed to validate op aten_matmul_default_894 with error 0xc26

compare with the log from this command SA8295
python examples/qualcomm/oss_scripts/llama/llama.py -b build-android -m **SA8295** --temperature 0 --model_mode hybrid --max_seq_len 1024 --prefill_ar_len 128 --decoder_model qwen3-0_6b --compile_only --prompt "你好啊"

qwen3_8295_0115_03.log

[WARNING] [Qnn ExecuTorch]: Qnn Backend op validation failed with error: 3110
[ERROR] [Qnn ExecuTorch]: [4294967295] has incorrect Value 68, expected >= 69.
[ERROR] [Qnn ExecuTorch]: QnnBackend_validateOpConfig failed 3110
[ERROR] [Qnn ExecuTorch]: Failed to validate op aten_convolution_default_11 with error 0xc26

[WARNING] [Qnn ExecuTorch]: Qnn Backend op validation failed with error: 3110
[ERROR] [Qnn ExecuTorch]: [4294967295] has incorrect Value 68, expected >= 73.
[ERROR] [Qnn ExecuTorch]: QnnBackend_validateOpConfig failed 3110
[ERROR] [Qnn ExecuTorch]: Failed to validate op aten_matmul_default_479 with error 0xc26

I had set the soc info to
QcomChipset.SA8797: SocInfo(QcomChipset.SA8797, HtpInfo(HtpArch.V81, 16))

[shewu-quic]

Hi @ares89 ,
Based on the below error message, it can not identify htp_arch of SA8797 in QAIRT SDK. Could you please use QAIRT SDK v2.42.0?
You can set QNN_SDK_ROOT to change the sdk be used.
https://softwarecenter.qualcomm.com/api/download/software/sdks/Qualcomm_AI_Runtime_Community/All/2.42.0.251225/v2.42.0.251225.zip

[ERROR] [Qnn ExecuTorch]: [4294967295] has incorrect Value 0 (should be 81), expected >= 73 (htp_arch).

[ares89]

Hi @shewu-quic , I successfully​ compiled qwen3-0.6b for​ SA8797 with QAIRT SDK v2.42.0, but when I run it on that device, the following error occurs.

the compile command:
python examples/qualcomm/oss_scripts/llama/llama.py -b build-android -m SA8797 --temperature 0 --model_mode hybrid --max_seq_len 1024 --prefill_ar_len 128 --decoder_model qwen3-0_6b --compile_only --prompt "你好啊,1+1等于几" --checkpoint /home/xxx/.cache/meta_checkpoints/Qwen_Qwen3-0.6B.pth --artifact ./llama_qnn/qwen3/8797_04

I checked the logs with adb, and here they are.

# make sure you have write-permission on below path.
DEVICE_DIR=/data/local/tmp/executorch_qualcomm_tutorial/
DEVICE_NAME=localhost:5555
adb -s ${DEVICE_NAME} shell "mkdir -p ${DEVICE_DIR}"
adb -s ${DEVICE_NAME} push ${QNN_SDK_ROOT}/lib/aarch64-android/libQnnHtp.so ${DEVICE_DIR}
adb -s ${DEVICE_NAME} push ${QNN_SDK_ROOT}/lib/aarch64-android/libQnnSystem.so ${DEVICE_DIR}
adb -s ${DEVICE_NAME} push ${QNN_SDK_ROOT}/lib/aarch64-android/libQnnHtpV69Stub.so ${DEVICE_DIR}
adb -s ${DEVICE_NAME} push ${QNN_SDK_ROOT}/lib/aarch64-android/libQnnHtpV73Stub.so ${DEVICE_DIR}
adb -s ${DEVICE_NAME} push ${QNN_SDK_ROOT}/lib/aarch64-android/libQnnHtpV75Stub.so ${DEVICE_DIR}
adb -s ${DEVICE_NAME} push ${QNN_SDK_ROOT}/lib/aarch64-android/libQnnHtpV79Stub.so ${DEVICE_DIR}
adb -s ${DEVICE_NAME} push ${QNN_SDK_ROOT}/lib/hexagon-v69/unsigned/libQnnHtpV69Skel.so ${DEVICE_DIR}
adb -s ${DEVICE_NAME} push ${QNN_SDK_ROOT}/lib/hexagon-v73/unsigned/libQnnHtpV73Skel.so ${DEVICE_DIR}
adb -s ${DEVICE_NAME} push ${QNN_SDK_ROOT}/lib/hexagon-v75/unsigned/libQnnHtpV75Skel.so ${DEVICE_DIR}
adb -s ${DEVICE_NAME} push ${QNN_SDK_ROOT}/lib/hexagon-v79/unsigned/libQnnHtpV79Skel.so ${DEVICE_DIR}
adb -s ${DEVICE_NAME} push ${QNN_SDK_ROOT}/lib/hexagon-v81/unsigned/libQnnHtpV81Skel.so ${DEVICE_DIR}

MODEL_DIR=/mnt/d/tmp/llama_qnn_qwen3_06b/8797_04
MODEL_NAME=hybrid_llama_qnn.pte
adb -s ${DEVICE_NAME} push ${MODEL_DIR}/${MODEL_NAME} ${DEVICE_DIR}
adb -s ${DEVICE_NAME} push ${EXECUTORCH_ROOT}/build-android/examples/qualcomm/executor_runner/qnn_executor_runner ${DEVICE_DIR}
adb -s ${DEVICE_NAME} push ${EXECUTORCH_ROOT}/build-android/lib/libqnn_executorch_backend.so ${DEVICE_DIR}
adb -s ${DEVICE_NAME} shell "cd ${DEVICE_DIR} \
           && export LD_LIBRARY_PATH=${DEVICE_DIR} \
           && export ADSP_LIBRARY_PATH=${DEVICE_DIR} \
           && ./qnn_executor_runner --model_path ./${MODEL_NAME}"

I 00:00:00.000575 executorch:qnn_executor_runner.cpp:232] Model file hybrid_llama_qnn.pte is loaded.
I 00:00:00.000589 executorch:qnn_executor_runner.cpp:242] Using method kv_forward
I 00:00:00.000592 executorch:qnn_executor_runner.cpp:289] Setting up planned buffer 0, size 0.
[INFO] [Qnn ExecuTorch]: Deserializing processed data using QnnContextCustomProtocol
[INFO] [Qnn ExecuTorch]: create QNN Logger with log_level 1
[INFO] [Qnn ExecuTorch]: Initialize Qnn backend parameters for Qnn executorch backend type 2
[INFO] [Qnn ExecuTorch]: Caching: Caching is in RESTORE MODE.
[INFO] [Qnn ExecuTorch]: QnnContextCustomProtocol expected magic number: 0x5678abcd but get: 0x2000000
[ERROR] [Qnn ExecuTorch]: QnnDsp Failed to create transport for device, error: 4000

[ERROR] [Qnn ExecuTorch]: QnnDsp Failed to load skel, error: 4000

[ERROR] [Qnn ExecuTorch]: QnnDsp Transport layer setup failed: 14001

[ERROR] [Qnn ExecuTorch]: QnnDsp Failed to parse default platform info: 14001

[ERROR] [Qnn ExecuTorch]: QnnDsp Failed to load default platform info: 14001

[ERROR] [Qnn ExecuTorch]: QnnDsp Failed to parse platform config: 14001

[ERROR] [Qnn ExecuTorch]: Failed to create device_handle for Backend ID 6, error=14001
E 00:00:00.817679 executorch:QnnManager.cpp:340] Fail to configure Qnn device
E 00:00:00.817690 executorch:QnnExecuTorchBackend.cpp:95] Fail to initialize Qnn Manager
E 00:00:00.817700 executorch:method.cpp:114] Init failed for backend QnnBackend: 0x1
F 00:00:00.871813 executorch:qnn_executor_runner.cpp:312] In function main(), assert failed (method.ok()): Loading of method kv_forward failed with status 0x1
Aborted

[ares89]

After i add adb -s ${DEVICE_NAME} push ${QNN_SDK_ROOT}/lib/aarch64-android/libQnnHtpV81Stub.so ${DEVICE_DIR}.
It seems ok now:

I 00:00:00.000609 executorch:qnn_executor_runner.cpp:232] Model file ./hybrid_llama_qnn.pte is loaded.
I 00:00:00.000624 executorch:qnn_executor_runner.cpp:242] Using method kv_forward
I 00:00:00.000627 executorch:qnn_executor_runner.cpp:289] Setting up planned buffer 0, size 0.
[INFO] [Qnn ExecuTorch]: Deserializing processed data using QnnContextCustomProtocol
[INFO] [Qnn ExecuTorch]: create QNN Logger with log_level 1
[INFO] [Qnn ExecuTorch]: Initialize Qnn backend parameters for Qnn executorch backend type 2
[INFO] [Qnn ExecuTorch]: Caching: Caching is in RESTORE MODE.
[INFO] [Qnn ExecuTorch]: QnnContextCustomProtocol expected magic number: 0x5678abcd but get: 0x2000000
[INFO] [Qnn ExecuTorch]: Running level=1 optimization.
[INFO] [Qnn ExecuTorch]: Running level=1 optimization.
I 00:00:01.503306 executorch:qnn_executor_runner.cpp:313] Method loaded.
I 00:00:01.577709 executorch:qnn_executor_runner.cpp:376] Inputs prepared.
I 00:00:01.703052 executorch:qnn_executor_runner.cpp:570] Input list not provided. Inputs prepared with default values set.
I 00:00:01.732098 executorch:qnn_executor_runner.cpp:579] Model executed successfully.
I 00:00:01.732120 executorch:qnn_executor_runner.cpp:582] Perform 0 inferences for warming up
I 00:00:01.759329 executorch:qnn_executor_runner.cpp:604] 1 inferences took 27.204000 ms, avg 27.204000 ms
I 00:00:01.782321 executorch:qnn_executor_runner.cpp:615] Write etdump to etdump.etdp, Size = 576
[INFO] [Qnn ExecuTorch]: Destroy Qnn backend parameters
[INFO] [Qnn ExecuTorch]: Destroy Qnn context
[INFO] [Qnn ExecuTorch]: Destroy Qnn device
[INFO] [Qnn ExecuTorch]: Destroy Qnn backend

[ares89]

With everyone's help, I finally successfully ran the quantized Qwen3-0.6B model on SA8797(SA8397).

python examples/qualcomm/oss_scripts/llama/llama.py -b build-android -m SA8797 -s localhost:5555 --decoder_model qwen3-0_6b --model_mode hybrid --prefill_ar_len 128 --max_seq_len 1024 --prompt "qnn是高通的工具，干嘛的" --pre_gen_pte D:\\tmp\\llama_qnn_qwen3_06b\\8797_04

Skipping import of cpp extensions due to incompatible torch version 2.9.0+cpu for torchao version 0.14.0 Please see GitHub issue #2919 for more info
I tokenizers:regex.cpp:27] Registering override fallback regex
NOTE: Using slow Hadamard transform for SpinQuant. For better performance on GPU, install fast_hadamard_transform: pip install git+https://github.com/Dao-AILab/fast-hadamard-transform.git
QNN_SDK_ROOT=/home/sunlit/executorch/backends/qualcomm/sdk/qnn
INFO:root:*** qwen3-0_6b ***
+-----------------------------+-------------------------------+
| Config | Value |
+=============================+===============================+
| custom_annotation | (partial(annotate_down_proj)) |
+-----------------------------+-------------------------------+
| decoder_model_version | qwen3 |
+-----------------------------+-------------------------------+
| get_kv_io_bit_width | 16 |
+-----------------------------+-------------------------------+
| get_logits_output_bit_width | 16 |
+-----------------------------+-------------------------------+
| group_size | 32 |
+-----------------------------+-------------------------------+
| instruct_model | True |
+-----------------------------+-------------------------------+
| masked_softmax | True |
+-----------------------------+-------------------------------+
| num_sharding | 1 |
+-----------------------------+-------------------------------+
| ptq | QuantDtype.use_16a4w_block |
+-----------------------------+-------------------------------+
| r1 | False |
+-----------------------------+-------------------------------+
| r2 | False |
+-----------------------------+-------------------------------+
| r3 | False |
+-----------------------------+-------------------------------+
| repo_id | Qwen/Qwen3-0.6B |
+-----------------------------+-------------------------------+
| seq_mse_candidates | 1000 |
+-----------------------------+-------------------------------+
| transform_weight | False |
+-----------------------------+-------------------------------+
D:\tmp\llama_qnn_qwen3_06b\8797_04/hybrid_llama...skipped. 26.8 MB/s (716363284 bytes in 25.479s)
/home/sunlit/executorch/backends/qualcomm/sdk/q... 0 skipped. 38.4 MB/s (2641976 bytes in 0.066s)
/home/sunlit/executorch/backends/qualcomm/sdk/q...0 skipped. 49.4 MB/s (11271968 bytes in 0.217s)
/home/sunlit/executorch/backends/qualcomm/sdk/q..., 0 skipped. 17.3 MB/s (714000 bytes in 0.039s)
/home/sunlit/executorch/backends/qualcomm/sdk/q...0 skipped. 46.8 MB/s (83685856 bytes in 1.706s)
/home/sunlit/executorch/backends/qualcomm/sdk/q... 0 skipped. 39.1 MB/s (2689096 bytes in 0.066s)
build-android/examples/qualcomm/oss_scripts/lla...0 skipped. 66.1 MB/s (84027576 bytes in 1.212s)
build-android/backends/qualcomm/libqnn_executor..., 0 skipped. 17.3 MB/s (646112 bytes in 0.036s)
/home/sunlit/executorch/backends/qualcomm/sdk/q... 0 skipped. 40.2 MB/s (2527744 bytes in 0.060s)
./llama_qnn/tokenizer.json: 1 file pushed, 0 skipped. 68.1 MB/s (11422654 bytes in 0.160s)
I tokenizers:regex.cpp:27] Registering override fallback regex
I 00:00:00.000515 executorch:runner.cpp:142] creating module: model_path=hybrid_llama_qnn.pte
I 00:00:00.000521 executorch:runner.cpp:143] creating runner: tokenizer_path=tokenizer.json
I 00:00:00.000523 executorch:runner.cpp:144] eval mode=1
I 00:00:00.000525 executorch:runner.cpp:145] kv updater=SmartMask
I tokenizers:hf_tokenizer.cpp:142] Setting up normalizer...
I tokenizers:normalizer.cpp:89] Using NFC normalizer. Please notice that our implementation may not handle all edge cases.
I tokenizers:hf_tokenizer.cpp:146] Normalizer set up
I tokenizers:hf_tokenizer.cpp:160] Setting up pretokenizer...
WARNING: All log messages before absl::InitializeLog() is called are written to STDERR
E0000 00:00:1767735957.296788 4419 re2.cc:237] Error parsing '((?i:'s|'t|'re|'ve|'m|'ll|'d)|[^\r\n\p{L}\p{N}]?\p{L}+|\p{N}| ?[^\s\p{L}\p{N}]+[\r\n]|\s[\r\n]+|\s...': invalid perl operator: (?!
I tokenizers:re2_regex.cpp:27] Re2 failed to compile regex: ((?i:'s|'t|'re|'ve|'m|'ll|'d)|[^\r\n\p{L}\p{N}]?\p{L}+|\p{N}| ?[^\s\p{L}\p{N}]+[\r\n]|\s[\r\n]+|\s+(?!\S)|\s+), error: invalid perl operator: (?!
This may be ok if a fallback regex is used.
I tokenizers:regex_lookahead.cpp:27] Creating PCRE2 regex
I tokenizers:hf_tokenizer.cpp:164] Pretokenizer set up
I tokenizers:hf_tokenizer.cpp:180] Loading BPE merges...
I tokenizers:hf_tokenizer.cpp:240] Loaded 151387 BPE merge rules
I tokenizers:hf_tokenizer.cpp:252] Built merge ranks map with 151387 entries
I 00:00:01.312683 executorch:llm_runner_helper.cpp:54] Loaded json tokenizer
I tokenizers:hf_tokenizer.cpp:415] normalized input: '' -> ''
[INFO] [Qnn ExecuTorch]: Deserializing processed data using QnnContextCustomProtocol
[INFO] [Qnn ExecuTorch]: create QNN Logger with log_level 1
[INFO] [Qnn ExecuTorch]: Initialize Qnn backend parameters for Qnn executorch backend type 2
[INFO] [Qnn ExecuTorch]: Caching: Caching is in RESTORE MODE.
[INFO] [Qnn ExecuTorch]: QnnContextCustomProtocol expected magic number: 0x5678abcd but get: 0x2000000
[INFO] [Qnn ExecuTorch]: Running level=1 optimization.
[INFO] [Qnn ExecuTorch]: Running level=1 optimization.
[INFO] [Qnn ExecuTorch]: Deserializing processed data using QnnContextCustomProtocol
[INFO] [Qnn ExecuTorch]: Use cached delegate handle for current method: kv_forward
I 00:00:02.158452 executorch:runner.cpp:219] Reading metadata from model
I 00:00:02.158694 executorch:runner.cpp:340] creating io_memory
I tokenizers:hf_tokenizer.cpp:415] normalized input: '' -> ''
I tokenizers:hf_tokenizer.cpp:415] normalized input: 'user
qnn是高通的工具，干嘛的' -> 'user
qnn是高通的工具，干嘛的'
I tokenizers:hf_tokenizer.cpp:415] normalized input: '
' -> '
'
I tokenizers:hf_tokenizer.cpp:415] normalized input: 'assistant' -> 'assistant'
I 00:00:02.164968 executorch:prompt_processor.cpp:265] Prompt Processor: total 18 prompt tokens (AR-128 * 1 iters)
I 00:00:02.268086 executorch:runner.cpp:436] RSS after prompt prefill: 790.593750 MiB (0 if unsupported)
I 00:00:07.393589 executorch:token_generator.cpp:322]
Reached to the end of generation
I 00:00:07.393610 executorch:runner.cpp:446] RSS after finishing text generation: 790.593750 MiB (0 if unsupported)
I 00:00:07.393641 executorch:stats.h:108] Prompt Tokens: 18 Generated Tokens: 248
I 00:00:07.393644 executorch:stats.h:114] Model Load Time: 2.159000 (seconds)
I 00:00:07.393646 executorch:stats.h:124] Total inference time: 5.234000 (seconds) Rate: 47.382499 (tokens/second)
I 00:00:07.393649 executorch:stats.h:132] Prompt evaluation: 0.108000 (seconds) Rate: 166.666667 (tokens/second)
I 00:00:07.393651 executorch:stats.h:143] Generated 248 tokens: 5.126000 (seconds) Rate: 48.380804 (tokens/second)
I 00:00:07.393653 executorch:stats.h:151] Time to first generated token: 0.108000 (seconds)
I 00:00:07.393655 executorch:stats.h:158] Sampling time over 266 tokens: 0.512000 (seconds)
[INFO] [Qnn ExecuTorch]: Destroy Qnn backend parameters
[INFO] [Qnn ExecuTorch]: Destroy Qnn context
[INFO] [Qnn ExecuTorch]: Destroy Qnn device
[INFO] [Qnn ExecuTorch]: Destroy Qnn backend

PyTorchObserver {"prompt_tokens":18,"generated_tokens":248,"model_load_start_ms":1767735956629,"model_load_end_ms":1767735958788,"inference_start_ms":1767735958788,"inference_end_ms":1767735964022,"prompt_eval_end_ms":1767735958896,"first_token_ms":1767735958896,"aggregate_sampling_time_ms":512,"SCALING_FACTOR_UNITS_PER_SECOND":1000}
/data/local/tmp/sunlit/executorch/single_llama/...led, 0 skipped. 0.0 MB/s (1331 bytes in 0.273s)
INFO:root:Results[0]:
<|im_start|>user
qnn是高通的工具，干嘛的<|im_end|>
<|im_start|>assistant

好的，我现在需要回答用户关于“qnn是高通的工具，干嘛的”的问题。首先，我需要确认用户的问题意图。用户可能对高通公司的产品或服务感兴趣，或者想了解qnn的具体用途。

接下来，我需要考虑如何回答这个问题。回答应该明确指出qnn的用途，同时可能提供一些背景信息，比如qnn是什么工具，以及它的功能。

然后，我需要检查是否有任何可能的误解或错误。例如，是否有可能将qnn与某个特定的工具混淆？或者是否需要进一步澄清问题。

最后，我需要组织回答，确保清晰、准确，并且符合用户的需求。

QNN是高通（Qualcomm）公司开发的一款智能手机操作系统，主要用于以下用途：

1. 增强用户体验：提供丰富的功能模块，如媒体、游戏、社交等，支持多设备互联。
2. 智能设备管理：通过QNN平台实现设备间的通信与数据同步。
3. 应用程序开发：提供开发者工具，支持开发者与用户之间的无缝交互。

QNN是高通公司为智能手机生态系统设计的集成平台，旨在优化用户与设备之间的交互体验。<|im_end|>
Finish the running pre_gen_pte from D:\tmp\llama_qnn_qwen3_06b\8797_04

[ares89]

base on qnn v2.42.0.251225.zip
Implementation is in PR #16694.