Default max_seq_length to 128 for ExecuTorch export

[swolchok]

Stack from ghstack (oldest at bottom):

• export.py: fix custom SDPA type conversion logic & re-enable for bfloat16 #1171
• -> Default max_seq_length to 128 for ExecuTorch export #1170
• Also use default max_seq_length of 128 for ExecuTorch native runner #1186
• Use default max_seq_length of 128 when loading ExecuTorch models #1184
• add ConvertToLinear, disable custom SDPA for bfloat16 #1169
• parallelize ExecuTorch build #1168
• Update ExecuTorch pin to pick up bfloat16 fixes #1181

With the current default behavior, performance for e.g. stories110Mwithout custom SDPA is bad because the QKV tensors are long (8192 in the last dim). Limiting the max sequence length remedies this.

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/torchchat/1170

• 📄 Preview Python docs built from this PR

Note: Links to docs will display an error until the docs builds have been completed.

✅ No Failures

As of commit c7efb72 with merge base 2cf4016 ():
💚 Looks good so far! There are no failures yet. 💚

This comment was automatically generated by Dr. CI and updates every 15 minutes.

[swolchok]

looks like we need to fix torchchat not respecting max_seq_length from PTEs in order to land this