Skip to content

Remove residual XPK references across training recipes - #394

Open
cj-klein wants to merge 1 commit into
AI-Hypercomputer:mainfrom
cj-klein:remove-residual-xpk-references
Open

cj-klein wants to merge 1 commit into
AI-Hypercomputer:mainfrom
cj-klein:remove-residual-xpk-references

Conversation

@cj-klein

Copy link
Copy Markdown
Contributor

As a follow-up to the XPK deprecation cleanups and Cluster Toolkit migrations (#389, #390, #391, #392), this PR removes the remaining residual XPK references across 11 training recipe files (0 remaining xpk matches in the repository):

  • Fix broken container-image link (1 file):
    • training/ironwood/deepseek3-671b/4k-bf16-tpu7x-4x4x8/k8s/README.md: update the WORKLOAD_IMAGE link from deleted ../xpk/README.md#docker-container-image to [Docker container image](../cluster_toolkit/README.md#docker-container-image).
  • Remove legacy XPK log banners (1 file):
    • training/ironwood/wan2.1-14b/bf16-tpu7x-4x4x4/k8s/k8s_manifest.yaml: remove echo XPK Start: $(date); and echo XPK End: $(date); from the JobSet container launch command.
  • Update KUEUE_QUEUE LocalQueue prose (7 files):
    • Across training/ironwood/{llama3.1-405b/8k-fp8-tpu7x-4x8x8,llama3.1-70b/8k-fp8-tpu7x-4x4x4,llama3.1-70b/8k-bf16-tpu7x-4x8x8,llama3.1-70b/8k-fp8-tpu7x-4x8x8,llama3.1-70b/128k-fp8-tpu7x-4x8x8,qwen3-235b-a22b/4k-fp8-tpu7x-4x8x8,wan2.2-t2v-14b/bf16-tpu7x-2x4x4}/cluster_toolkit/README.md, reword the KUEUE_QUEUE example from `multislice-queue` on a cluster provisioned with XPK to on a custom or shared GKE cluster.
  • Clean up Long-Context benchmark tables (2 files):
    • training/ironwood/Llama3.1-8B-LongContext-Maxtext/cluster_toolkit/README.md and training/v5p/Llama3.1-8B-LongContext-Maxtext/cluster_toolkit/README.md: remove the internal-only UBench Run Name & BigQuery `run_id` column (which contained internal LDAP-prefixed run names and deprecated maxtext_training_with_xpk-* IDs rather than external links) while keeping all context lengths, parallelism/offload configurations, step times, throughputs, and MFU numbers intact.

Follows up on the XPK recipe removals and Cluster Toolkit migrations (AI-Hypercomputer#389-AI-Hypercomputer#392) to clean up the remaining 20 residual XPK references across 11 files:

- Fix broken relative link in training/ironwood/deepseek3-671b/4k-bf16-tpu7x-4x4x8/k8s/README.md pointing to deleted ../xpk/README.md#docker-container-image -> ../cluster_toolkit/README.md#docker-container-image
- Remove legacy 'echo XPK Start/End' log lines from training/ironwood/wan2.1-14b/bf16-tpu7x-4x4x4/k8s/k8s_manifest.yaml
- Reword KUEUE_QUEUE LocalQueue guidance across 7 Ironwood cluster_toolkit/README.md files to drop 'multislice-queue on a cluster provisioned with XPK'
- Remove the internal-only 'UBench Run Name & BigQuery run_id' column (containing deprecated maxtext_training_with_xpk-* IDs and an internal LDAP) from the verified benchmark tables in the Ironwood and v5p Llama3.1-8B-LongContext-Maxtext cluster_toolkit/README.md files while preserving all configuration and performance metrics

@gemini-code-assist gemini-code-assist Bot left a comment

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Code Review

This pull request updates documentation and configuration files across several model training directories. Key changes include removing the 'UBench Run Name & BigQuery run_id' column from performance tables in the READMEs, updating documentation links to point to the cluster toolkit instead of XPK, modifying the description of LocalQueues to refer to custom or shared GKE clusters, and removing XPK start/end timestamp logging from a Kubernetes manifest. There are no review comments to evaluate, and I have no additional feedback to provide.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant