tree 143d7fd2c5edbd5a9b572df9d67b7c558afd703f
parent c62c3d00583a91d82f4f15c89d97d525b2d2c529
author zjgarvey <47986913+zjgarvey@users.noreply.github.com> 1733782033 -0600
committer GitHub <noreply@github.com> 1733782033 -0600
gpgsig -----BEGIN PGP SIGNATURE-----
 
 wsFcBAABCAAQBQJnV2oRCRC1aQ7uu5UhlAAAvdsQAJvUpSA1M8MPHx9lCED6TrpP
 gr7nnG4OWH9v0iAZHs989WULIFnj+Doq0lIngY+YFMUc3q4dTxJq93RxB530za29
 l2c2d6kfc9AjGy19hHW6fivp4R8mmD+yStKF7yDq/lm4vdT1GQ2FY8CwFOl/CNxw
 eD2pz74MjQ5i4zixTwqqzDIZPPhnLEw2X3RO79qVq7ZRS6M6CeUxwQ4Y/ioCD4cN
 YSUfT92HWvXmOxf1OEMNjY6xnytaQWb3tFLrMIps3ky6dBjDY58eSF6ckVvMccJV
 U+2WaPGrDIzv6pDxve/ej2N25L4RIXkyxn6PsE2dUKJvfur1ke14niFQPeRmNmuj
 /EebTYZSA2r73ylbTgCOenjJOs1VPa5saff/rtqYBzLb56y9gwEU7Y6FdM9HkPyL
 R/TsL8ERXEN2irlgR/7DqHtjs+pQQ9zkHslUwhJ2cQizk77iuJ2ZlFYIjWTnT4+9
 XzeSJz81Ob3SzKHqnIdNSGnuJjf7Jhm8dO2emZVTxW8cvMTelTdKNzUXYvxNO1ZM
 iE9HCiiSv1pdP9ccOHBmoLLYrFKLPbCytog13hC7dNF6JzdI2+KZmN8GXyBUCzvy
 XQ6s1y6wYh7H9eJV4pq+9p0Wib4EbC55KGqLSuzOrGBSkh4v7bcp8QyNS69CT6uy
 UXIky4ODeHiEpNFXtAGL
 =QIhw
 -----END PGP SIGNATURE-----
 

[iree-import-onnx] improve handling of large models (#19217)

This pr adds a few options:

1. `--large-model` allows disabling the onnx model checker if a user
knows ahead of time that the model is too large. It will also not load
the external weights in memory unless saving the parameters.
2. `--num-initializers-threshold` allows storing initializers to the
irpa file in batches with a specified number of entries. This can reduce
the memory overhead of first gathering all of the initializers, then
saving them to the irpa at once.
3. `--externalize-inputs-threshold` allows converting inputs to
externalized weights. This is useful for the following workflow:
exporting a HF pytorch model with safetensors, saving a `.irpa` from the
safetensor weights directly, and exporting to onnx with
`export_params=False` and `do_constant_folding=False` (which converts
weights to inputs and avoids folding weights with things like
transposes). When importing to mlir, you can set
`externalize-inputs-threshold=<num_original_inputs>` and it will convert
the inputs from and beyond that threshold to `util.global` ops.
4. `--save-params`/`--no-save-params` factors saving parameters out of
`import_initializer`, and one can avoid saving parameters with
`--no-save-params`. Useful for debugging compilation failures.

## TODO:

Figure out what to do about loading the onnx model and updating opset
version. It's possible to do opset version updating without weights in a
somewhat hacky way, since models > 2GB fail on opset version updating.

Add documentation

---------

Signed-off-by: zjgarvey <zjgarvey@gmail.com>