fix: Enable quantization and compilation in the same optimization job via ModelBuilder and add validations to prevent compilation for Llama-3.1 on TRTLLM.

[cj-zhang]

…ModelBuilder.

Issue #, if available:

Description of changes: TRTLLM containers allow compilation and quantization at the same time, so the validations to ensure mutual exclusivity must be removed. Compilation using Llama-3.1 on TRTLLM v11 also doesn't currently work, so adding validations to prevent those jobs from being created.

Testing done: New and updated unit tests on jumpstart_builder and model_builder. Locally tested TRTLLM/Llama-3.1 validations from both JS and HF flow.

Merge Checklist

Put an x in the boxes that apply. You can also fill these out after creating the PR. If you're unsure about any of them, don't hesitate to ask. We're here to help! This is simply a reminder of what we are going to look for before merging your pull request.

General

• I have read the CONTRIBUTING doc
• I certify that the changes I am introducing will be backward compatible, and I have discussed concerns about this, if any, with the Python SDK team
• I used the commit message format described in CONTRIBUTING
• I have passed the region in to all S3 and STS clients that I've initialized as part of this change.
• I have updated any necessary documentation, including READMEs and API docs (if appropriate)

Tests

• I have added tests that prove my fix is effective or that my feature works (if appropriate)
• I have added unit and/or integration tests as appropriate to ensure backward compatibility of the changes
• I have checked that my tests are not configured for a specific region or account (if appropriate)
• I have used unique_name_from_base to create resource names in integ tests (if appropriate)

By submitting this pull request, I confirm that my contribution is made under the terms of the Apache 2.0 license.

[xiongz945]

Need to consider "g6", "p5" as well. And I don't "g4" should be included

[Lokiiiiii]

g4 is not supported.

[xiongz945]

Why is this model ID not the same as the MB's init argument?

[Lokiiiiii]

Missing items -

1. The ability to add JS models as draft models for speculative decoding
2. If the draft model is a gated model, customer needs to provide an additional accept_eula parameter during model initialization.
3. TRT Compilation cannot be stacked with Speculative Decoding(SD)
4. New Quantization option - FP8
5. New Quantization option - SmoothQuant
6. SmoothQuant requires TRT compilation (llama 3.1 not supported, cannot be stacked with SD)
7. ModelSharding ...

[Lokiiiiii]

g4 is not supported.

[Lokiiiiii]

Can we swap this for a more dynamic check based on JS metadata ?

I am not sure if this code path caters to non-JS models as well. So maybe we want to keep this for HF models.

[Lokiiiiii]

Why is there a mismatch in TP Degree here ?

Are we trying to compile with TP=2 and then deploy with TP=1 ?

This behavior is unsupported. Essentially LMI will ignore the runtime TP Degree in favor of what was compiled.

[Lokiiiiii]

Missing items -

1. The ability to add JS models as draft models for speculative decoding

2. If the draft model is a gated model, customer needs to provide an additional accept_eula parameter during model initialization.

3. TRT Compilation cannot be stacked with Speculative Decoding(SD)

4. New Quantization option - FP8

5. New Quantization option - SmoothQuant

6. SmoothQuant requires TRT compilation (llama 3.1 not supported, cannot be stacked with SD)

7. ModelSharding ...

sagemaker-python-sdk/src/sagemaker/jumpstart/types.py

Line 1132 in f45e600

self.artifact_version: str = json_obj["artifact_version"]

- ArtifactVersion is now optional and might not always be available in JS Metadata.