tlopex opened a new pull request, #20006:
URL: https://github.com/apache/tvm/pull/20006

   This PR reduces Jenkins CI latency by removing redundant infrastructure and 
test work.
   
   This pr
   - Keep current ARM/CPU/GPU/WASM Docker images warm on shared Jenkins nodes 
instead of having jobs evict one another's images. Old images are still removed 
based on age and disk usage.
   - Move JVM packaging from the GPU test stage to the CPU build stage, where 
it runs in parallel with artifact upload and additional runtime builds.
   - Expose 336 LLVM division cases and 156 CUDA vectorized-cast cases as 
pytest parameters so existing xdist workers can compile them in parallel.
   - Reuse compiled paged-attention kernels while continuing to create fresh 
runtime KV caches for every test.
   - Replace redundant ONNX Conv and Reduce executions with exhaustive Relax IR 
validation plus representative numerical tests.
   - Register only runnable ONNX backend cases instead of collecting thousands 
of tests that are immediately skipped.
   - Regenerate all affected Jenkinsfiles from their templates.
   
   
   LLVM and CUDA case coverage is unchanged. Paged-attention tests still use 
independent runtime state. The same 388 ONNX backend tests remain runnable.
   
   ONNX Conv and Reduce retain their complete configuration matrices as 
structural importer tests, with representative end-to-end numerical canaries.
   
   Targeted measurements showed:
   
   - LLVM division: 25.34s → 19.87s
   - CUDA vectorized casts: 142.85s → 108.40s
   - Paged-attention CPU: 316.32s → 192.50s
   - ONNX Conv: 25.84s → 1.81s
   - ONNX Reduce: 20.66s → 4.04s
   
   Besides, Jenkins log measurements show that retaining an existing CPU image 
reduces pull time from 85.034s to 0.826s, saving 84.208s per cache hit. For the 
GPU image, pull time drops from 270.640s to 0.562s, saving approximately 4 
minutes 30 seconds per cache hit.


-- 
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.

To unsubscribe, e-mail: [email protected]

For queries about this service, please contact Infrastructure at:
[email protected]


---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]

Reply via email to