Primus v26.3

[gargrahul]

Primus v26.3 release introduces the following new models and upgrades previously supported models

Qwen 3 30B BF16/FP8
Qwen 3 235B BF16/FP8
GPT OSS 20B BF16/FP8
GPT OSS 120B BF16/FP8

[Copilot]

Pull request overview

Updates MAD’s Primus integration to align with the Primus v26.3 release, adding new Megatron-LM training targets (Qwen3 30B/235B, GPT-OSS 20B/120B) and bumping Docker base images/documentation to the newer container stack.

Changes:

• Add new Primus Megatron-LM model repos + datatype support logic (BF16/FP8 where applicable), and extend benchmark parsing to recognize these models.
• Introduce a setup-time patch step to add training-log metrics summarization to Primus’ primus-cli-direct.sh.
• Bump training Docker base images to rocm/primus:v26.3 and refresh benchmark README component versions.

Reviewed changes

Copilot reviewed 12 out of 12 changed files in this pull request and generated 6 comments.

Show a summary per file

File	Description
scripts/pytorch_train/run.sh	Ensure certain post-train model repos explicitly run with BF16.
scripts/primus/pytorch_train/primus_pytorch_benchmark_report.sh	Adjust device→config mapping logic for pretrain benchmarks.
scripts/primus/megatron-lm/run.sh	Add new model repo selectors + datatype support matrix updates.
scripts/primus/megatron-lm/primus_megatron-lm_benchmark_setup.sh	Apply an inline patch to Primus to parse/append training metrics summaries.
scripts/primus/megatron-lm/primus_megatron-lm_benchmark_report.sh	Add new models and refine benchmark execution behaviors.
scripts/primus/megatron-lm/primus_megatron-lm_benchmark_report.py	Extend log parsing eligibility list for new models.
models.json	Register new model repos for MAD runs (including skip-arch for GPT-OSS-120B).
docker/pytorch_train.ubuntu.amd.Dockerfile	Bump base image to rocm/primus:v26.3.
docker/primus_pytorch_train.ubuntu.amd.Dockerfile	Bump base image to rocm/primus:v26.3.
docker/primus_megatron_train.ubuntu.amd.Dockerfile	Bump base image to rocm/primus:v26.3.
benchmark/pytorch_train/README.md	Update component versions listed for the training container.
benchmark/megatron_lm/README.md	Update component versions and document some new supported models + examples.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

Copilot was unable to review this pull request because the user who requested the review is ineligible. To be eligible to request a review, you need a paid Copilot license, or your organization must enable Copilot code review.

[Copilot]

Copilot was unable to review this pull request because the user who requested the review is ineligible. To be eligible to request a review, you need a paid Copilot license, or your organization must enable Copilot code review.

[Copilot]

Copilot was unable to review this pull request because the user who requested the review is ineligible. To be eligible to request a review, you need a paid Copilot license, or your organization must enable Copilot code review.

[vidushi8]

LGTM

[Copilot]

Pull request overview

Copilot reviewed 12 out of 12 changed files in this pull request and generated 8 comments.

[amd-fuyuajin]

Made Copilot suggested changes. Added multi-node training examples. Ready to merge.

[Copilot]

Copilot was unable to review this pull request because the user who requested the review is ineligible. To be eligible to request a review, you need a paid Copilot license, or your organization must enable Copilot code review.

[Copilot]

Copilot was unable to review this pull request because the user who requested the review is ineligible. To be eligible to request a review, you need a paid Copilot license, or your organization must enable Copilot code review.

[peterjunpark]

Suggested change

	The docker container hosts verified commit `e16b27b` from [Primus repository](https://github.com/AMD-AGI/Primus/tree/e16b27bf6c1b2798f38848fc574fee60d9a9b902).
	The docker container hosts verified commit `43a6e0` from [Primus repository](https://github.com/AMD-AGI/Primus/tree/release/v26.3).

Should this be 43a6e006c419697208295c5523b99070e8198ad9? That's the head of the release branch https://github.com/AMD-AGI/Primus/commits/release/v26.3/

[peterjunpark]

Suggested change

Primus defines training yaml for each model inside [‎examples/megatron/configs/](https://github.com/AMD-AGI/Primus/tree/e16b27bf6c1b2798f38848fc574fee60d9a9b902/examples/megatron/configs) repository. For example, use `examples/megatron/configs/llama3.1_8B-pretrain.yaml` for updating llama3.1_8B training parameters. Other yaml for the supported model can be found with `examples/megatron/configs/${MODEL_NAME}-pretrain.yaml` naming convention in this repository.

Primus defines training yaml for each model inside [‎examples/megatron/configs/](https://github.com/AMD-AGI/Primus/tree/release/v26.3) repository. For example, use `examples/megatron/configs/llama3.1_8B-pretrain.yaml` for updating llama3.1_8B training parameters. Other yaml for the supported model can be found with `examples/megatron/configs/${MODEL_NAME}-pretrain.yaml` naming convention in this repository.