Skip to content

Add Deepseek V4 REAP recipes with tuned GPU memory settings - #318

Open
akashdeepjassal wants to merge 3 commits into
eugr:mainfrom
akashdeepjassal:feature/DS4-REAP
Open

Add Deepseek V4 REAP recipes with tuned GPU memory settings#318
akashdeepjassal wants to merge 3 commits into
eugr:mainfrom
akashdeepjassal:feature/DS4-REAP

Conversation

@akashdeepjassal

Copy link
Copy Markdown

Deepseek V4 REAP recipes for single Spark.
Tested on Asus Ascent GX10

  • Add deepseek-v4-flash-162b-reap-gb10-tuned.yaml (gpu_memory_utilization: 0.89)
  • Add deepseek-v4-flash-180b-reap-gb10-tuned.yaml (gpu_memory_utilization: 0.91)
  • Add mods/deepseek-v4-reap-gb10/ directory
  • Add Dockerfile.deepseek-v4-reap-gb10

@akashdeepjassal
akashdeepjassal force-pushed the feature/DS4-REAP branch 2 times, most recently from b8feb14 to 3b81247 Compare July 12, 2026 22:02
- Add deepseek-v4-flash-162b-reap-gb10-tuned.yaml (gpu_memory_utilization: 0.89)
- Add deepseek-v4-flash-180b-reap-gb10-tuned.yaml (gpu_memory_utilization: 0.91)
- Add mods/deepseek-v4-reap-gb10/ patch script
- Apply reap patch inline in main Dockerfile (same flow as base, no variant Dockerfile)
@akashdeepjassal

akashdeepjassal commented Jul 15, 2026

Copy link
Copy Markdown
Author

@eugr I've opened this PR to add Deepseek V4 Flash REAP support for single Spark (tested on Asus Ascent GX10).

Benchmarks:

Quick question on REAP:
Since this is the first REAP integration in the repo, does this recipe structure and Dockerfile setup work as a good baseline template for future REAP support?

Let me know if you'd like any adjustments!

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