ywkang/kernbench2
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Fix all remaining test failures: single-cube allreduce + matplotlib dep

Browse Source 
- intercube_allreduce: add single-cube fast path that skips intra-SIP
  mesh reduce and goes directly to inter-SIP exchange. Fixes IPCQ
  deadlock when TP launches kernel on one cube per SIP.
- distributed.py: derive effective cube dims from tensor shard placement
  instead of hardcoding topology mesh size.
- pyproject.toml: add matplotlib>=3.7 to dependencies.
- pe_dma.py (prior commit): add MMU translation in pipeline DMA path.

577 passed, 0 failed (was 529 passed, 10 failed).

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
This commit is contained in:
Yangwook Kang
2026-04-27 21:25:31 -07:00
parent d55dc6cb4f
commit fca24feac5
15 changed files with 112 additions and 95 deletions
Download Patch File Download Diff File Expand all files Collapse all files
tests/pe2pe_latency_plots/summary.csv
-10
View File
@@ -79,13 +79,3 @@ h4_inter_cube_vertical,Inter-cube vertical (cube0 to cube4),8192,ipcq,181.659999
h4_inter_cube_vertical,Inter-cube vertical (cube0 to cube4),8192,raw,183.04000000000087
h4_inter_cube_vertical,Inter-cube vertical (cube0 to cube4),10240,ipcq,205.65999999999985
h4_inter_cube_vertical,Inter-cube vertical (cube0 to cube4),10240,raw,207.04000000000087
h5_inter_sip,"Inter-SIP (sip0 to sip1, same cube/pe)",128,ipcq,6.015000000003056
h5_inter_sip,"Inter-SIP (sip0 to sip1, same cube/pe)",256,ipcq,6.515000000003056
h5_inter_sip,"Inter-SIP (sip0 to sip1, same cube/pe)",384,ipcq,7.015000000003056
h5_inter_sip,"Inter-SIP (sip0 to sip1, same cube/pe)",512,ipcq,7.515000000003056
h5_inter_sip,"Inter-SIP (sip0 to sip1, same cube/pe)",768,ipcq,8.515000000003056
h5_inter_sip,"Inter-SIP (sip0 to sip1, same cube/pe)",1024,ipcq,9.515000000003056
h5_inter_sip,"Inter-SIP (sip0 to sip1, same cube/pe)",2048,ipcq,13.515000000003056
h5_inter_sip,"Inter-SIP (sip0 to sip1, same cube/pe)",4096,ipcq,21.515000000003056
h5_inter_sip,"Inter-SIP (sip0 to sip1, same cube/pe)",8192,ipcq,37.51499999999214
h5_inter_sip,"Inter-SIP (sip0 to sip1, same cube/pe)",10240,ipcq,45.51499999999214
1 hop label size_bytes path total_ns
79 h4_inter_cube_vertical Inter-cube vertical (cube0 to cube4) 8192 raw 183.04000000000087
80 h4_inter_cube_vertical Inter-cube vertical (cube0 to cube4) 10240 ipcq 205.65999999999985
81 h4_inter_cube_vertical Inter-cube vertical (cube0 to cube4) 10240 raw 207.04000000000087
h5_inter_sip Inter-SIP (sip0 to sip1, same cube/pe) 128 ipcq 6.015000000003056
h5_inter_sip Inter-SIP (sip0 to sip1, same cube/pe) 256 ipcq 6.515000000003056
h5_inter_sip Inter-SIP (sip0 to sip1, same cube/pe) 384 ipcq 7.015000000003056
h5_inter_sip Inter-SIP (sip0 to sip1, same cube/pe) 512 ipcq 7.515000000003056
h5_inter_sip Inter-SIP (sip0 to sip1, same cube/pe) 768 ipcq 8.515000000003056
h5_inter_sip Inter-SIP (sip0 to sip1, same cube/pe) 1024 ipcq 9.515000000003056
h5_inter_sip Inter-SIP (sip0 to sip1, same cube/pe) 2048 ipcq 13.515000000003056
h5_inter_sip Inter-SIP (sip0 to sip1, same cube/pe) 4096 ipcq 21.515000000003056
h5_inter_sip Inter-SIP (sip0 to sip1, same cube/pe) 8192 ipcq 37.51499999999214
h5_inter_sip Inter-SIP (sip0 to sip1, same cube/pe) 10240 ipcq 45.51499999999214