[QNN:Bugfix] Raise stack limit for context generation

[LudovicoYIN]

Description

When converting Qwen3/Qwen3.5 4B LLM models to QNN context binaries on v73, the conversion may fail at the last large graph, which appears to include the lm_head part.

We debugged the failed qnn-context-binary-generator command with gdb and found that it crashes inside the generated QNN model shared library:

SIGSEGV in QnnModel_composeGraphs() from libgraph1_37.so

The generated C++ for the final graph contains very large local stack arrays for lm_head quantization metadata, for example:

Qnn_ScaleOffset_t tensor__lm_lm_head_Linear_bias_axis_scale_offset[] = { ... };

In our case, this array had about 151,936 entries. Qnn_ScaleOffset_t contains a float and an int32_t, so this single local array is roughly:

151,936 * 8 bytes = 1,215,488 bytes ~= 1.16 MiB

The generated function also contains many other local tensors and metadata arrays. With the default Linux stack limit:

ulimit -s
# 8192

the available stack is only 8192 KB, about 8 MiB. This can make qnn-context-binary-generator segfault while composing the graph.

After raising the stack limit:

ulimit -s unlimited

the same qnn-context-binary-generator command succeeds.

This PR raises the Python process stack limit before invoking QNN tools in npu_convert.py. The QNN subprocesses inherit this limit, so qnn-context-binary-generator can compose the large graph successfully.

No QNN graph options are changed in this PR.

Tested with Qwen3-4B QNN conversion on v73 using QNN SDK 2.42.

Module

QNN

Type

• Bugfix

Checklist

• Commit message follows [Module:Type] Description format
• Code compiles without errors
• Tested on relevant platform(s)
• No unrelated format or style changes included

fix: #4567