Research Stage
Previous existing literature and research
I don't know if it can be useful but I wanted to report that while with version v0.4.2 the V100 arch 70 GPU did not work, now with version v0.4.3 preview it works without errors, but very slowly. A test on 128k ctx that with turboquant turbo3 takes 253 seconds with bellama kwarn3 takes 11480 seconds.
Hypothesis
No response
Implementation
No response
Analysis
No response
Relevant log output
Research Stage
Previous existing literature and research
I don't know if it can be useful but I wanted to report that while with version v0.4.2 the V100 arch 70 GPU did not work, now with version v0.4.3 preview it works without errors, but very slowly. A test on 128k ctx that with turboquant turbo3 takes 253 seconds with bellama kwarn3 takes 11480 seconds.
Hypothesis
No response
Implementation
No response
Analysis
No response
Relevant log output