Conversation
🏷️ CI GuideRuns automatically on every PR:
Extended tests (opt-in via labels):
|
There was a problem hiding this comment.
Pull request overview
Note
Copilot was unable to run its full agentic suite in this review.
Adds additional GEMM (M,N,K) problem sizes to the DeepSpeed v3 a8w8 blockscale “untuned” config, likely to cover MI300-relevant shapes and smaller-M cases.
Changes:
- Extends the CSV with many new (M,N,K) entries (notably N=512, K=7168 with M ranging 1→32768).
- Adds a blank separator line and one additional shape that differs from the new block’s (N,K) pattern.
💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.
| 20480, 7168, 2048 | ||
| 32768, 7168, 2048 No newline at end of file | ||
| 32768, 7168, 2048 | ||
|
|
There was a problem hiding this comment.
There is an empty row in the CSV. If the loader expects every non-header line to contain exactly 3 comma-separated values, this blank line can cause parse failures or require special-casing. Remove the blank line, or replace it with a comment row only if the parser explicitly supports comments.
|
|
||
| 1, 512, 7168 | ||
| 2, 512, 7168 | ||
| 4, 7168, 2048 |
There was a problem hiding this comment.
The new block largely follows a consistent pattern of N=512, K=7168 while sweeping M, but the entry 4, 7168, 2048 breaks that pattern (N/K appear swapped vs surrounding rows). If this isn’t intentional, it’s likely a typo and should be corrected to match the intended N/K values; if it is intentional, consider grouping it with other *, 7168, 2048 shapes (or adding a clear separator/comment if supported) to avoid accidental misconfiguration.
| 4, 7168, 2048 | |
| 4, 512, 7168 |
Motivation
Technical Details
Test Plan
Test Result
Submission Checklist